test(heal): cover set bulkhead scheduler isolation

Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com>
ci(docker): move latest on every prerelease until first stable tag (#7341 )
2026-09-07 04:25:54 +00:00 · 2026-09-07 12:12:16 +08:00 · 2026-09-07 11:25:40 +08:00 · 2026-09-07 11:04:34 +08:00 · 2026-09-07 02:40:57 +00:00 · 2026-09-07 10:28:42 +08:00
153 changed files with 22934 additions and 5394 deletions
@@ -582,59 +582,13 @@ jobs:
          install-build-packaging-tools: 'false'

      - name: Build debug binary
-        run: |
-          python3 - <<'PYBUILD'
-          import hashlib
-          import json
-          import os
-          import pathlib
-          import subprocess
-
-          def git(*args):
-              return subprocess.check_output(["git", *args], text=True).strip()
-
-          def sha256(path):
-              digest = hashlib.sha256()
-              with pathlib.Path(path).open("rb") as source:
-                  for chunk in iter(lambda: source.read(1024 * 1024), b""):
-                      digest.update(chunk)
-              return digest.hexdigest()
-
-          argv = ["cargo", "build", "-p", "rustfs", "--bins", "--features", "e2e-test-hooks"]
-          commit, tree = git("rev-parse", "HEAD"), git("rev-parse", "HEAD^{tree}")
-          clean_before = not git("status", "--porcelain", "--untracked-files=normal")
-          if not clean_before:
-              raise SystemExit("hooks binary requires a clean build checkout")
-          lock_sha256 = sha256("Cargo.lock")
-          lock_git_blob = git("hash-object", "Cargo.lock")
-          rustc = subprocess.check_output(["rustc", "-vV"], text=True)
-          host = next(line.removeprefix("host: ") for line in rustc.splitlines() if line.startswith("host: "))
-          if os.environ.get("CARGO_BUILD_TARGET") or pathlib.Path(os.environ.get("CARGO_TARGET_DIR", "target")).resolve() != pathlib.Path("target").resolve():
-              raise SystemExit("this artifact requires the native target/debug output")
-          subprocess.run(argv, check=True)
-          clean_after = not git("status", "--porcelain", "--untracked-files=normal")
-          if not clean_after or commit != git("rev-parse", "HEAD") or tree != git("rev-parse", "HEAD^{tree}") or lock_sha256 != sha256("Cargo.lock"):
-              raise SystemExit("hooks binary source changed while building")
-          manifest = {
-              "schema": 1, "commit": commit, "tree": tree,
-              "clean_before": clean_before, "clean_after": clean_after,
-              "lock_sha256": lock_sha256, "lock_git_blob": lock_git_blob,
-              "argv": argv, "profile": "debug", "target": host,
-              "features": ["e2e-test-hooks"],
-              "rustc_verbose": rustc,
-              "build_flags": {key: os.environ[key] for key in ("RUSTFLAGS", "CARGO_ENCODED_RUSTFLAGS", "CARGO_BUILD_TARGET", "CARGO_TARGET_DIR", "RUSTUP_TOOLCHAIN") if key in os.environ},
-              "binary_sha256": sha256("target/debug/rustfs"),
-          }
-          pathlib.Path("target/debug/rustfs.e2e-startup-cas-build.json").write_text(json.dumps(manifest, indent=2) + "\n")
-          PYBUILD
+        run: cargo build -p rustfs --bins --features e2e-test-hooks

      - name: Upload debug binary
        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
        with:
          name: rustfs-debug-binary
-          path: |
-            target/debug/rustfs
-            target/debug/rustfs.e2e-startup-cas-build.json
+          path: target/debug/rustfs
          if-no-files-found: error
          retention-days: 1

@@ -900,6 +854,12 @@ jobs:
        run: |
          sudo apt-get install -y iptables
          sudo -n iptables --version
+          # The endpoint-blackhole heal scenario needs CAP_NET_ADMIN. Containerised
+          # runners can run iptables but not touch the rule set; the test then logs
+          # a skip instead of failing, so surface that here where it is visible.
+          if ! sudo -n iptables -w 5 -S OUTPUT >/dev/null 2>&1; then
+            echo "::warning::iptables cannot read the OUTPUT chain on this runner (no CAP_NET_ADMIN); the endpoint-blackhole heal scenario will be skipped"
+          fi

      - name: Set up Python
        uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6.3.0
@@ -952,36 +912,6 @@ jobs:
      - name: Make binary executable
        run: chmod +x ./target/debug/rustfs

-      - name: Preserve startup CAS binary input
-        env:
-          STARTUP_CAS_INPUT: ${{ runner.temp }}/rustfs-startup-cas-input
-        run: |
-          python3 - <<'PYINPUT'
-          import hashlib
-          import json
-          import os
-          import pathlib
-          import shutil
-          import subprocess
-
-          source = pathlib.Path("target/debug/rustfs")
-          manifest_path = source.with_name("rustfs.e2e-startup-cas-build.json")
-          manifest = json.loads(manifest_path.read_text())
-          target = pathlib.Path(os.environ["STARTUP_CAS_INPUT"])
-          target.mkdir(parents=True, exist_ok=True)
-          binary = target / "rustfs"
-          shutil.copy2(source, binary)
-          digest = hashlib.sha256()
-          with binary.open("rb") as stream:
-              for chunk in iter(lambda: stream.read(1024 * 1024), b""):
-                  digest.update(chunk)
-          commit = subprocess.check_output(["git", "rev-parse", "HEAD"], text=True).strip()
-          if manifest["binary_sha256"] != digest.hexdigest() or manifest["commit"] != commit:
-              raise SystemExit("downloaded hooks binary identity mismatch")
-          shutil.copy2(manifest_path, target / manifest_path.name)
-          binary.chmod(0o755)
-          PYINPUT
-
      - name: Verify e2e full membership
        env:
          NEXTEST_LISTING: ${{ runner.temp }}/rustfs-e2e-full-list.json
@@ -994,10 +924,6 @@ jobs:
      # extend that filter, never add ad-hoc e2e jobs here. Reuses the downloaded
      # debug binary; each test spawns its own rustfs server on a random port.
      - name: Run e2e full suite
-        env:
-          RUSTFS_E2E_STARTUP_CAS_BINARY: ${{ runner.temp }}/rustfs-startup-cas-input/rustfs
-          RUSTFS_E2E_STARTUP_CAS_BUILD_MANIFEST: ${{ runner.temp }}/rustfs-startup-cas-input/rustfs.e2e-startup-cas-build.json
-          RUSTFS_E2E_STARTUP_CAS_ARTIFACT_DIR: ${{ runner.temp }}/rustfs-startup-cas-evidence
        run: cargo nextest run --profile e2e-full -p e2e_test

      - name: Upload junit
@@ -1010,17 +936,6 @@ jobs:
            ${{ runner.temp }}/rustfs-e2e-full-list.json
          retention-days: 7

-      - name: Upload startup CAS evidence
-        if: always()
-        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
-        with:
-          name: fresh-startup-cas-evidence-${{ github.run_number }}
-          path: |
-            ${{ runner.temp }}/rustfs-startup-cas-evidence
-            ${{ runner.temp }}/rustfs-startup-cas-input/rustfs.e2e-startup-cas-build.json
-          if-no-files-found: warn
-          retention-days: 7
-
  e2e-tests-rio-v2:
    name: End-to-End Tests (rio-v2)
    # Inherits the schedule/dispatch-only gate through needs: on every other
@@ -121,6 +121,22 @@ jobs:
          create_latest=false
          source_ref="$GITHUB_SHA"

+          # Pre-GA policy: until the first stable (vX.Y.Z) tag exists, every
+          # prerelease (alpha/beta/rc) also moves `latest`, so users pulling
+          # `latest` get the newest test build. Once a stable tag is published
+          # this returns false and `latest` follows stable releases only.
+          prerelease_moves_latest() {
+            local stable_tags
+            stable_tags=$(git ls-remote --tags --refs origin 2>/dev/null \
+              | awk '{print $2}' \
+              | grep -E '^refs/tags/v?[0-9]+\.[0-9]+\.[0-9]+$' || true)
+            if [[ -z "$stable_tags" ]]; then
+              return 0
+            fi
+            echo "ℹ️  Stable release tag(s) already exist; prereleases no longer update latest"
+            return 1
+          }
+
          if [[ "${{ github.event_name }}" == "workflow_run" ]]; then
            # Triggered by build workflow completion
            echo "🔗 Triggered by build workflow completion"
@@ -184,8 +200,8 @@ jobs:
                if [[ "$version" == *"alpha"* ]] || [[ "$version" == *"beta"* ]] || [[ "$version" == *"rc"* ]]; then
                  build_type="prerelease"
                  is_prerelease=true
-                  # Current policy: create latest tags for stable releases and selected prereleases (alpha/beta).
-                  if [[ "$version" == *"alpha"* ]] || [[ "$version" == *"beta"* ]]; then
+                  # Pre-GA policy: prereleases update latest until the first stable tag exists.
+                  if prerelease_moves_latest; then
                    create_latest=true
                    echo "🧪 Building Docker image for prerelease: $version (creating latest tag)"
                  else
@@ -243,8 +259,8 @@ jobs:
              v*alpha*|v*beta*|v*rc*|*alpha*|*beta*|*rc*)
                build_type="prerelease"
                is_prerelease=true
-                # Current policy: create latest tags for stable releases and selected prereleases (alpha/beta).
-                if [[ "$version" == *"alpha"* ]] || [[ "$version" == *"beta"* ]]; then
+                # Pre-GA policy: prereleases update latest until the first stable tag exists.
+                if prerelease_moves_latest; then
                  create_latest=true
                  echo "🧪 Building with prerelease version: $input_version (creating latest tag)"
                else
@@ -394,11 +410,13 @@ jobs:
          TAG_BASE="${VERSION}${VARIANT_SUFFIX}"
          TAGS="${{ env.REGISTRY_DOCKERHUB }}:$TAG_BASE,${{ env.REGISTRY_GHCR }}:$TAG_BASE,${{ env.REGISTRY_QUAY }}:$TAG_BASE"

-          # Add channel tags for prereleases and latest for stable
+          # Add latest when requested (stable releases, and prereleases before GA)
          if [[ "$CREATE_LATEST" == "true" ]]; then
-            # Create latest tags for stable releases and selected prereleases when CREATE_LATEST=true.
            TAGS="$TAGS,${{ env.REGISTRY_DOCKERHUB }}:latest${VARIANT_SUFFIX},${{ env.REGISTRY_GHCR }}:latest${VARIANT_SUFFIX},${{ env.REGISTRY_QUAY }}:latest${VARIANT_SUFFIX}"
-          elif [[ "$BUILD_TYPE" == "prerelease" ]]; then
+          fi
+
+          # Always add the channel tag for prereleases, independent of latest
+          if [[ "$BUILD_TYPE" == "prerelease" ]]; then
            # Prerelease channel tags (alpha, beta, rc)
            if [[ "$VERSION" == *"alpha"* ]]; then
              CHANNEL="alpha"
@@ -555,7 +573,7 @@ jobs:
            "prerelease")
              echo "🧪 Prerelease Docker image has been built with ${VERSION} tags"
              echo "⚠️  This is a prerelease image - use with caution"
-              # Create latest tags for stable releases and selected prereleases when CREATE_LATEST=true.
+              # Prereleases move latest until the first stable tag exists (pre-GA policy).
              if [[ "$CREATE_LATEST" == "true" ]]; then
                echo "🏷️  Latest tag has been created for prerelease: $VERSION"
              else
@@ -2527,18 +2527,18 @@ checksum = "790eea4361631c5e7d22598ecd5723ff611904e3344ce8720784c93e3d83d40b"

 [[package]]
 name = "crossbeam-channel"
-version = "0.5.16"
+version = "0.5.17"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d85363c37faeca707aef026efa9f3b34d077bce547e48f770770625c6013679e"
+checksum = "98b0cc327b5bc766e7fda9c9260cc0fa81b43a8e240440422dff70788e3f9ef1"
 dependencies = [
 "crossbeam-utils",
 ]

 [[package]]
 name = "crossbeam-deque"
-version = "0.8.7"
+version = "0.8.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5181e0de7b61eb03a81e347d6dd8797bae9da5146707b51077e2d71a54ec0ceb"
+checksum = "622f3fc73690be383c7214310406f28a90e6edeadc3cea882f9d71e495b9711a"
 dependencies = [
 "crossbeam-epoch",
 "crossbeam-utils",
@@ -2546,27 +2546,27 @@ dependencies = [

 [[package]]
 name = "crossbeam-epoch"
-version = "0.9.20"
+version = "0.9.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2d6914041f254d6e9176c01941b21115dcfb7089e55135a35411081bd106ef3f"
+checksum = "dc74980687109a3b14c72fd458107bf0baa1da1a1a805e178d15501ba9b86d9d"
 dependencies = [
 "crossbeam-utils",
 ]

 [[package]]
 name = "crossbeam-queue"
-version = "0.3.13"
+version = "0.3.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "803d13fb3b09d88be9f4dbc29062c66b19bf7170867ceb746d2a8689bf6c7a26"
+checksum = "03e8bd762f7479489c70ed6c768ddca99d7296857de437a68dcb2a94365b3fae"
 dependencies = [
 "crossbeam-utils",
 ]

 [[package]]
 name = "crossbeam-utils"
-version = "0.8.22"
+version = "0.8.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "61803da095bee82a81bb1a452ecc25d3b2f1416d1897eb86430c6159ef717c17"
+checksum = "a31eee39dddec8330830986fcd7625edb5a24ec90ea038215273bbc3adb08ac6"

 [[package]]
 name = "crunchy"
@@ -3673,9 +3673,9 @@ dependencies = [

 [[package]]
 name = "der"
-version = "0.8.1"
+version = "0.8.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a69dedd701da44b0536442edf09c81a64b0ab97a7a4a5e3d1971f00027cbc63d"
+checksum = "a878c850e9e421b20262e9b41f9c860e4785fa07541c266b62ff9d1ef998a80a"
 dependencies = [
 "const-oid 0.10.2",
 "pem-rfc7468 1.0.0",
@@ -3946,7 +3946,7 @@ dependencies = [
 "libc",
 "option-ext",
 "redox_users 0.5.2",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -4057,7 +4057,6 @@ dependencies = [
 "sha1 0.11.0",
 "sha2 0.11.0",
 "suppaftp",
- "tempfile",
 "time",
 "tokio",
 "tokio-stream",
@@ -4091,7 +4090,7 @@ version = "0.17.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c0681a4fc24c767085329728d8dfba959af91228aa4610cca4f8ce317ba46ae0"
 dependencies = [
- "der 0.8.1",
+ "der 0.8.2",
 "digest 0.11.3",
 "elliptic-curve 0.14.1",
 "rfc6979 0.6.0",
@@ -4296,7 +4295,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
 dependencies = [
 "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -5694,7 +5693,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "20fd6de4ccfcc187e38bc21cfa543cb5a302cb86a8b114eb7f0bf0dc9f8ac00f"
 dependencies = [
 "io-lifetimes 3.0.1",
- "windows-sys 0.60.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -5735,9 +5734,9 @@ dependencies = [

 [[package]]
 name = "ipnet"
-version = "2.12.1"
+version = "2.12.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6a756c3fac73139e83f14c2d742155dd2b78d3ee56597b419a0579b7bdd6dd78"
+checksum = "791930b43c0d5973160d90a8f3894509f2b273430f5c5c73b668636d0287c5c0"
 dependencies = [
 "serde",
 ]
@@ -5759,7 +5758,7 @@ checksum = "3640c1c38b8e4e43584d8df18be5fc6b0aa314ce6ebf51b53313d4306cca8e46"
 dependencies = [
 "hermit-abi",
 "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -6140,9 +6139,9 @@ checksum = "3eaf3ede3fee6db1a4c2ee091bf8a8b4dccdc6d17f656fb07896ee72867612f2"

 [[package]]
 name = "libflate"
-version = "2.3.1"
+version = "2.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a4da9b700e758e57152a1fd1c52cbdc5727c1aa6d8743dc1acda917398f1d76c"
+checksum = "561a8da1a50e1428d3c51321dafeca849df992a5bb67720c386131234caba82e"
 dependencies = [
 "adler32",
 "crc32fast",
@@ -6956,7 +6955,7 @@ version = "0.50.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -7934,7 +7933,7 @@ version = "0.8.0-rc.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "986d2e952779af96ea048f160fd9194e1751b4faea78bcf3ceb456efe008088e"
 dependencies = [
- "der 0.8.1",
+ "der 0.8.2",
 "spki 0.8.0",
 ]

@@ -7977,7 +7976,7 @@ dependencies = [
 "aes 0.9.3",
 "aes-gcm",
 "cbc 0.2.1",
- "der 0.8.1",
+ "der 0.8.2",
 "pbkdf2 0.13.0",
 "rand_core 0.10.1",
 "scrypt 0.12.0",
@@ -8001,7 +8000,7 @@ version = "0.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "451913da69c775a56034ea8d9003d27ee8948e12443eae7c038ba100a4f21cb7"
 dependencies = [
- "der 0.8.1",
+ "der 0.8.2",
 "pkcs5 0.8.1",
 "rand_core 0.10.1",
 "spki 0.8.0",
@@ -8707,7 +8706,7 @@ dependencies = [
 "once_cell",
 "socket2",
 "tracing",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -8943,9 +8942,9 @@ dependencies = [

 [[package]]
 name = "redis"
-version = "1.6.0"
+version = "1.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e37a4ca5c6ca42aa3e6df2fd32b987a65d32a4c2159a6f3fe0fd1df306a2658f"
+checksum = "2acbc41a996f7652b2ddd9dfd98cc4ff602cfd742ae35382f07f608405ab50ed"
 dependencies = [
 "arc-swap",
 "arcstr",
@@ -9327,7 +9326,7 @@ dependencies = [
 "curve25519-dalek 5.0.0",
 "data-encoding",
 "delegate",
- "der 0.8.1",
+ "der 0.8.2",
 "digest 0.11.3",
 "ecdsa 0.17.0",
 "ed25519-dalek 3.0.0",
@@ -10746,6 +10745,7 @@ dependencies = [
 "rustfs-data-usage",
 "rustfs-ecstore",
 "rustfs-filemeta",
+ "rustfs-heal",
 "rustfs-heal-contracts",
 "rustfs-lifecycle",
 "rustfs-lock",
@@ -10943,9 +10943,9 @@ dependencies = [

 [[package]]
 name = "rustfs-uring"
-version = "0.2.1"
+version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0486e62d0efe25db95c00aeacb2da84368adcba299216cda99fcb11328061c84"
+checksum = "b29bc57b4bd62a73f4fae408b536adf578332e50e464797d09dc2382c7cb68c2"
 dependencies = [
 "io-uring",
 "libc",
@@ -11066,7 +11066,7 @@ dependencies = [
 "errno",
 "libc",
 "linux-raw-sys",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -11149,7 +11149,7 @@ dependencies = [
 "security-framework",
 "security-framework-sys",
 "webpki-root-certs",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -11420,7 +11420,7 @@ checksum = "d56d437c2f19203ce5f7122e507831de96f3d2d4d3be5af44a0b0a09d8a80e4d"
 dependencies = [
 "base16ct 1.0.0",
 "ctutils",
- "der 0.8.1",
+ "der 0.8.2",
 "hybrid-array",
 "subtle",
 "zeroize",
@@ -11994,7 +11994,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1d9efca8738c78ee9484207732f728b1ef517bbb1833d6fc0879ca898a522f6f"
 dependencies = [
 "base64ct",
- "der 0.8.1",
+ "der 0.8.2",
 ]

 [[package]]
@@ -12409,7 +12409,7 @@ dependencies = [
 "getrandom 0.4.3",
 "once_cell",
 "rustix",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -13528,7 +13528,7 @@ version = "0.1.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]

 [[package]]
@@ -256,10 +256,10 @@ clap = { version = "4.6.6" }
 const-str = { version = "1.1.0" }
 convert_case = "0.12.0"
 criterion = { version = "0.8" }
-crossbeam-queue = "0.3.13"
-crossbeam-channel = "0.5.16"
-crossbeam-deque = "0.8.7"
-crossbeam-utils = "0.8.22"
+crossbeam-queue = "0.3.14"
+crossbeam-channel = "0.5.17"
+crossbeam-deque = "0.8.8"
+crossbeam-utils = "0.8.23"
 datafusion = { default-features = false, version = "55.0.0" }
 derive_builder = "0.20.2"
 enumset = "1.1.14"
@@ -306,7 +306,7 @@ rustfs-erasure-codec = { version = "8.0.2" }
 reed-solomon-simd = "3.1.0"
 regex = { version = "1.13.1" }
 rumqttc = { package = "rumqttc-next", version = "0.34.0" }
-redis = { version = "1.6.0" }
+redis = { version = "1.7.0" }
 rustify = { version = "0.7", default-features = false }
 rustix = { version = "1.1.4" }
 rust-embed = { version = "8.12.0" }
@@ -422,9 +422,9 @@ fn unix_now_ms() -> u64 {
        .unwrap_or(0)
 }

-/// A repair the MRF consumer landed, fanned out so retry ledgers can drop
-/// entries the journal no longer tracks (backlog#1894 axis B). The payload
-/// mirrors the intent identity so consumers match without re-parsing.
+/// Legacy, unverified repair notice. Its identity lacks kind, set scope,
+/// bucket incarnation and responsibility generation. Consumers must not use
+/// it to discharge persisted repair responsibility.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct MrfRepairedEvent {
    pub bucket: Arc<str>,
@@ -439,8 +439,8 @@ const MRF_REPAIRED_EVENT_CAP: usize = 4096;

 static MRF_REPAIRED_EVENTS: OnceLock<std::sync::Mutex<std::collections::VecDeque<MrfRepairedEvent>>> = OnceLock::new();

-/// Record that the MRF consumer landed a repair. Never blocks: the critical
-/// section is a deque push under a std mutex.
+/// Record a legacy notification for compatibility. This is not an
+/// acknowledgement of storage verification or durable repair completion.
 pub fn note_mrf_repaired(bucket: &str, object: &str, version_id: Option<[u8; 16]>) {
    let registry = MRF_REPAIRED_EVENTS.get_or_init(|| std::sync::Mutex::new(std::collections::VecDeque::new()));
    let Ok(mut events) = registry.lock() else {
@@ -515,6 +515,9 @@ mod tests {
        }
        coalescer_release(&key, Some(lease));
        let retry_lease = coalescer_admit(key.clone()).expect("released identity must admit a retry");
+        assert_ne!(lease, retry_lease);
+        coalescer_release(&key, Some(lease));
+        assert_eq!(coalescer_admit(key.clone()), Err(MrfIngressResult::Coalesced));
        coalescer_release(&key, Some(retry_lease));
    }

@@ -144,6 +144,3 @@ russh = { workspace = true, features = ["serde"] }
 russh-sftp = { workspace = true }
 zip.workspace = true
 clap = { workspace = true, features = ["derive", "env"] }
-
-[dev-dependencies]
-tempfile.workspace = true
@@ -184,6 +184,8 @@ the wiring source of truth. Committed test-ID digests under

 ## Troubleshooting

+**Endpoint blackhole scenario skipped** — `heal_erasure_disk_rebuild_test::tests::test_cluster_root_heal_recovers_after_target_endpoint_blackhole` installs a loopback `iptables` DROP rule and therefore needs `CAP_NET_ADMIN` (root or passwordless `sudo -n iptables`). A host where `iptables` is missing or cannot read the OUTPUT chain (typical inside an unprivileged container, where the nf_tables backend reports "Permission denied" even under `sudo`) logs a `heal_interruption_skipped` warning and returns without exercising heal. Set `RUSTFS_E2E_REQUIRE_NET_FAULT_INJECTION=1` on lanes that do provision the capability so a broken runner fails instead of skipping.
+
 **Reproduce a CI failure locally** — run the exact profile/lane:

 ```bash
@@ -34,6 +34,8 @@ mod tests {
    use tokio::net::TcpStream;
    use tokio::time::{Duration, Instant, sleep, timeout};
    use tracing::info;
+    #[cfg(target_os = "linux")]
+    use tracing::warn;

    const POOL_METADATA_OBJECT: &str = "pool.bin";

@@ -115,6 +117,49 @@ mod tests {
    }

    impl TcpPortBlackhole {
+        /// Environment flag that turns an unusable fault-injection host into a
+        /// hard failure instead of a logged skip. Lanes that provision
+        /// `CAP_NET_ADMIN` set it so a broken runner cannot pass silently.
+        #[cfg(target_os = "linux")]
+        const REQUIRE_ENV: &str = "RUSTFS_E2E_REQUIRE_NET_FAULT_INJECTION";
+
+        /// Probe whether this host can manipulate the OUTPUT chain at all.
+        ///
+        /// Returns `Ok(Some(reason))` when `iptables` is missing or lacks
+        /// `CAP_NET_ADMIN` (the nf_tables backend reports "Permission denied"
+        /// even under `sudo` inside an unprivileged container) and the lane did
+        /// not demand fault injection; returns an error when the lane demands
+        /// it; returns `Ok(None)` when the blackhole can be installed.
+        #[cfg(target_os = "linux")]
+        fn unavailable_reason() -> Result<Option<String>, Box<dyn Error + Send + Sync>> {
+            let id = Command::new("id").arg("-u").output()?;
+            if !id.status.success() {
+                return Err(format!("failed to determine the test process uid: {}", String::from_utf8_lossy(&id.stderr)).into());
+            }
+            let use_sudo = String::from_utf8_lossy(&id.stdout).trim() != "0";
+            let mut command = if use_sudo {
+                let mut command = Command::new("sudo");
+                command.args(["-n", "iptables"]);
+                command
+            } else {
+                Command::new("iptables")
+            };
+            let probe = command.args(["-w", "5", "-S", "OUTPUT"]).output();
+            let reason = match probe {
+                Ok(output) if output.status.success() => return Ok(None),
+                Ok(output) => format!(
+                    "iptables cannot read the OUTPUT chain (status {}): {}",
+                    output.status,
+                    String::from_utf8_lossy(&output.stderr).trim()
+                ),
+                Err(err) => format!("iptables is not runnable: {err}"),
+            };
+            if std::env::var_os(Self::REQUIRE_ENV).is_some() {
+                return Err(format!("{} is set but network fault injection is unavailable: {reason}", Self::REQUIRE_ENV).into());
+            }
+            Ok(Some(reason))
+        }
+
        fn install(address: &str) -> Result<Self, Box<dyn Error + Send + Sync>> {
            let address = address.parse::<SocketAddr>()?;
            if !address.ip().is_loopback() {
@@ -199,6 +244,27 @@ mod tests {
        }
    }

+    /// Remove a disk directory underneath a running server. Background writers
+    /// (scanner, usage cache, heal markers) can recreate entries between the
+    /// recursive listing and the final `rmdir`, which surfaces as
+    /// `DirectoryNotEmpty` on macOS; retry briefly so the wipe reflects the
+    /// operator action rather than a listing race.
+    fn wipe_directory_while_server_runs(disk: &Path) -> std::io::Result<()> {
+        let mut last_err = None;
+        for _ in 0..20 {
+            match std::fs::remove_dir_all(disk) {
+                Ok(()) => return Ok(()),
+                Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ok(()),
+                Err(err) if err.kind() == std::io::ErrorKind::DirectoryNotEmpty => {
+                    last_err = Some(err);
+                    std::thread::sleep(std::time::Duration::from_millis(100));
+                }
+                Err(err) => return Err(err),
+            }
+        }
+        Err(last_err.expect("retry loop only exits without success after recording an error"))
+    }
+
    fn has_file_under(path: &Path) -> bool {
        let Ok(entries) = std::fs::read_dir(path) else {
            return false;
@@ -481,7 +547,7 @@ mod tests {
            );
        }

-        std::fs::remove_dir_all(&disk0).expect("disk0 wipe should succeed while server is running");
+        wipe_directory_while_server_runs(&disk0).expect("disk0 wipe should succeed while server is running");
        std::fs::create_dir_all(&disk0).expect("disk0 should be recreated empty while server is running");
        assert!(!has_file_under(&disk0), "disk0 must be empty immediately after runtime wipe");

@@ -868,6 +934,18 @@ mod tests {
    #[cfg(target_os = "linux")]
    #[tokio::test(flavor = "multi_thread")]
    async fn test_cluster_root_heal_recovers_after_target_endpoint_blackhole() -> Result<(), Box<dyn Error + Send + Sync>> {
+        if let Some(reason) = TcpPortBlackhole::unavailable_reason()? {
+            init_logging();
+            warn!(
+                event = "heal_interruption_skipped",
+                component = "e2e_test",
+                subsystem = "heal",
+                interruption_kind = "target_endpoint_blackhole",
+                reason,
+                "Skipping endpoint blackhole scenario: network fault injection is unavailable on this host"
+            );
+            return Ok(());
+        }
        timeout(
            Duration::from_secs(420),
            run_cluster_root_heal_interruption(InterruptionScenario::TargetEndpointBlackhole),
@@ -21,7 +21,7 @@
 use super::common::{BoxError, OdmSourceSpec, OdmTestEnv, SeedObject};
 use crate::fake_s3_target::{BucketMode, Operation};
 use aws_sdk_s3::error::ProvideErrorMetadata;
-use aws_sdk_s3::types::{BucketVersioningStatus, VersioningConfiguration};
+use aws_sdk_s3::types::{BucketVersioningStatus, ObjectAttributes, VersioningConfiguration};
 use bytes::Bytes;
 use std::time::Duration;

@@ -103,11 +103,19 @@ async fn get_miss_pulls_inline_and_serves_locally_afterwards() -> TestResult {

 #[tokio::test]
 async fn get_large_object_streams_through_and_backfills_in_background() -> TestResult {
+    const PART_SIZE: usize = 5 * 1024 * 1024;
    let bucket = "odm-get-large";
-    let env = configured_env(bucket, |spec| spec.policy.inline_max_bytes = 4096).await?;
+    let env = configured_env(bucket, |spec| {
+        spec.policy.inline_max_bytes = 4096;
+        spec.policy.multipart_part_size_bytes = u64::try_from(PART_SIZE).expect("part size fits in u64");
+    })
+    .await?;
    let key = "large/archive.bin";
-    let body = payload(512 * 1024);
-    env.seed_source(SOURCE_BUCKET, &[SeedObject::new(key, body.clone())]);
+    let body = payload(PART_SIZE + 4096);
+    let etag = env
+        .seed_source(SOURCE_BUCKET, &[SeedObject::new(key, body.clone())])
+        .remove(0);
+    assert_eq!(etag.len(), 32, "the source fixture has a plain MD5 ETag");

    let response = env.raw_get(bucket, key).await?;
    assert_eq!(response.status, 200, "{}", String::from_utf8_lossy(&response.body));
@@ -125,6 +133,68 @@ async fn get_large_object_streams_through_and_backfills_in_background() -> TestR
        vec![None, None],
        "one passthrough GET plus one background pull, both unranged"
    );
+
+    let source_requests = env.source.requests().len();
+    let second_part = env.client.get_object().bucket(bucket).key(key).part_number(2).send().await?;
+    assert_eq!(second_part.content_length(), Some(4096), "the completed second part is the tail");
+    assert_eq!(
+        second_part.content_range(),
+        Some(format!("bytes {PART_SIZE}-{}/{}", body.len() - 1, body.len()).as_str()),
+        "partNumber reads the stored multipart boundary"
+    );
+    assert_eq!(
+        second_part.body.collect().await?.into_bytes(),
+        body.slice(PART_SIZE..),
+        "the local second part contains the exact source tail"
+    );
+    let third_part = env
+        .client
+        .get_object()
+        .bucket(bucket)
+        .key(key)
+        .part_number(3)
+        .send()
+        .await
+        .expect_err("the completed object has exactly two parts");
+    assert_eq!(third_part.code(), Some("InvalidPart"));
+
+    let mut part_marker = None;
+    for (part_number, part_size) in [(1, PART_SIZE), (2, 4096)] {
+        let attributes = env
+            .client
+            .get_object_attributes()
+            .bucket(bucket)
+            .key(key)
+            .object_attributes(ObjectAttributes::ObjectParts)
+            .object_attributes(ObjectAttributes::Etag)
+            .max_parts(1)
+            .set_part_number_marker(part_marker.clone())
+            .send()
+            .await?;
+        assert_eq!(
+            attributes.e_tag().map(|value| value.trim_matches('"')),
+            Some(etag.as_str()),
+            "multipart write-back preserves the source MD5 ETag"
+        );
+        let parts = attributes
+            .object_parts()
+            .expect("RustFS must expose the stored multipart layout");
+        assert_eq!(parts.total_parts_count(), Some(2));
+        assert_eq!(parts.max_parts(), Some(1));
+        assert_eq!(parts.is_truncated(), Some(part_number == 1));
+        assert_eq!(parts.parts().len(), 1, "RustFS returns one stored part per requested page");
+        assert_eq!(parts.parts()[0].part_number(), Some(part_number));
+        assert_eq!(parts.parts()[0].size(), Some(i64::try_from(part_size).expect("part size fits in i64")));
+        part_marker = parts.next_part_number_marker().map(str::to_owned);
+        if part_number == 1 {
+            assert_eq!(part_marker.as_deref(), Some("1"), "the next request continues after the first part");
+        }
+    }
+    assert_eq!(
+        env.source.requests().len(),
+        source_requests,
+        "local part reads must not consult the source"
+    );
    Ok(())
 }

@@ -23,16 +23,20 @@

 use super::common::{
    ALLOW_LOOPBACK_SOURCE_ENV, AdminResponse, BackfillOp, BackfillRequest, BoxError, ODM_MODULE_SWITCH_ENV, ODM_SERVER_ENV,
-    OdmEnvOptions, OdmSourceSpec, OdmTestEnv, SeedObject, start_configured_env, start_configured_env_with,
+    OdmEnvOptions, OdmSourceSpec, OdmTestEnv, SeedObject, start_configured_env, start_configured_env_with, start_source_rustfs,
 };
 use crate::common::{RustFSTestEnvironment, replication_fast_env, signed_request};
 use crate::fake_s3_target::{BucketMode, FAKE_ACCESS_KEY, FAKE_SECRET_KEY, FakeS3Target, Operation};
 use crate::object_lock::common::put_object_lock_configuration;
+use crate::replication_extension_test::{
+    ReplicationTargetOptions, enable_bucket_versioning, set_replication_target_with_options,
+};
 use aws_sdk_s3::error::ProvideErrorMetadata;
 use aws_sdk_s3::types::{
    BucketVersioningStatus, Event, FilterRule, FilterRuleName, NotificationConfiguration, NotificationConfigurationFilter,
-    ObjectLockRetentionMode, QueueConfiguration, S3KeyFilter, ServerSideEncryption, ServerSideEncryptionByDefault,
-    ServerSideEncryptionConfiguration, ServerSideEncryptionRule, Tag, Tagging, VersioningConfiguration,
+    ObjectAttributes, ObjectLockRetentionMode, QueueConfiguration, S3KeyFilter, ServerSideEncryption,
+    ServerSideEncryptionByDefault, ServerSideEncryptionConfiguration, ServerSideEncryptionRule, Tag, Tagging,
+    VersioningConfiguration,
 };
 use bytes::Bytes;
 use local_ip_address::local_ip;
@@ -580,6 +584,130 @@ async fn test_odm_pulled_object_replicates_and_target_as_source_is_rejected() ->
        "a bucket may not migrate from its own replication target: {}",
        rejected.body
    );
+    Box::pin(assert_odm_multipart_replicates_to_rustfs(&env, bucket)).await?;
+    Ok(())
+}
+
+async fn assert_odm_multipart_replicates_to_rustfs(env: &OdmTestEnv, bucket: &str) -> TestResult {
+    const PART_SIZE: usize = 5 * 1024 * 1024;
+    let replica = start_source_rustfs().await?;
+    let replica_bucket = "odm-real-replica";
+    replica.create_test_bucket(replica_bucket).await?;
+    enable_bucket_versioning(&replica, replica_bucket).await?;
+    let arn = set_replication_target_with_options(
+        &env.rustfs,
+        bucket,
+        ReplicationTargetOptions {
+            endpoint: &replica.address,
+            access_key: &replica.access_key,
+            secret_key: &replica.secret_key,
+            target_bucket: replica_bucket,
+            secure: false,
+            skip_tls_verify: false,
+            ca_cert_pem: None,
+        },
+    )
+    .await?;
+    put_bucket_replication(&env.rustfs, bucket, &arn).await?;
+    let mut spec = env.fake_source_spec(SOURCE_BUCKET);
+    // Below the 16 MiB inline default the pull is one tee'd PUT with a single
+    // part; force the passthrough + background multipart write-back instead.
+    spec.policy.inline_max_bytes = 4096;
+    spec.policy.multipart_part_size_bytes = PART_SIZE as u64;
+    spec.policy.preserve_etag = true;
+    env.configure_and_wait(bucket, &spec).await?;
+
+    let key = "replicated/preserved-md5-multipart.bin";
+    let body = payload(PART_SIZE + 4096);
+    let source_put = env
+        .source_client()
+        .put_object()
+        .bucket(SOURCE_BUCKET)
+        .key(key)
+        .body(aws_sdk_s3::primitives::ByteStream::from(body.clone()))
+        .send()
+        .await?;
+    let etag = source_put.e_tag().ok_or("source PUT omitted its ETag")?.trim_matches('"');
+    assert_eq!(etag.len(), 32, "the source must retain a single-PUT MD5 ETag");
+    assert!(etag.bytes().all(|byte| byte.is_ascii_hexdigit()));
+    let pulled = env.raw_get(bucket, key).await?;
+    assert_eq!(pulled.status, 200, "{}", String::from_utf8_lossy(&pulled.body));
+    assert_eq!(pulled.body, body);
+    assert!(env.wait_local_listed(bucket, key, SETTLE).await?, "the multipart pull must persist");
+
+    let deadline = Instant::now() + SETTLE;
+    let source_head = loop {
+        let head = env.client.head_object().bucket(bucket).key(key).send().await?;
+        match head.replication_status().map(|status| status.as_str()) {
+            Some("COMPLETED") => break head,
+            Some("FAILED") => return Err("the ODM multipart copy failed replication to RustFS".into()),
+            _ => {
+                assert!(Instant::now() < deadline, "the ODM multipart copy never completed replication to RustFS");
+                tokio::time::sleep(Duration::from_millis(200)).await;
+            }
+        }
+    };
+    let version = source_head
+        .version_id()
+        .ok_or("the versioned ODM copy omitted its version id")?;
+    assert_ne!(version, "null");
+    let replica_client = replica.create_s3_client();
+    for (client, object_bucket) in [(&env.client, bucket), (&replica_client, replica_bucket)] {
+        let attributes = client
+            .get_object_attributes()
+            .bucket(object_bucket)
+            .key(key)
+            .version_id(version)
+            .object_attributes(ObjectAttributes::Etag)
+            .object_attributes(ObjectAttributes::ObjectParts)
+            .send()
+            .await?;
+        assert_eq!(attributes.e_tag().map(|value| value.trim_matches('"')), Some(etag));
+        let parts = attributes
+            .object_parts()
+            .ok_or("the local copy and replica must both expose two parts")?;
+        assert_eq!(parts.total_parts_count(), Some(2));
+        assert_eq!(
+            parts
+                .parts()
+                .iter()
+                .map(|part| (part.part_number(), part.size()))
+                .collect::<Vec<_>>(),
+            [(Some(1), Some(PART_SIZE as i64)), (Some(2), Some(4096))]
+        );
+    }
+    // REPLICA status surfaces on HEAD, like the other inbound-replica checks.
+    let replica_head = replica_client
+        .head_object()
+        .bucket(replica_bucket)
+        .key(key)
+        .version_id(version)
+        .send()
+        .await?;
+    assert_eq!(replica_head.replication_status().map(|status| status.as_str()), Some("REPLICA"));
+    let replica_get = replica_client
+        .get_object()
+        .bucket(replica_bucket)
+        .key(key)
+        .version_id(version)
+        .send()
+        .await?;
+    assert_eq!(replica_get.version_id(), Some(version));
+    assert_eq!(replica_get.body.collect().await?.into_bytes(), body);
+    let boundary = replica_client
+        .get_object()
+        .bucket(replica_bucket)
+        .key(key)
+        .version_id(version)
+        .range(format!("bytes={}-{}", PART_SIZE - 32, PART_SIZE + 31))
+        .send()
+        .await?;
+    assert_eq!(boundary.body.collect().await?.into_bytes(), body.slice(PART_SIZE - 32..PART_SIZE + 32));
+    assert_eq!(
+        env.source.count_requests(Operation::GetObject, key),
+        2,
+        "one passthrough GET plus one background pull; replication and local reads must not fetch the migration source again"
+    );
    Ok(())
 }

@@ -368,23 +368,23 @@ impl Drop for SlowReplicationTargetGuard {
 // Mirrors madmin-go `ResyncTargetsInfo`/`ResyncTarget` json tags — the same
 // shape `mc replicate resync status` decodes.
 #[derive(Debug, Clone, serde::Deserialize)]
-struct ReplicationResetStatusResponse {
+pub(crate) struct ReplicationResetStatusResponse {
    #[serde(rename = "target", default)]
-    targets: Vec<ReplicationResetStatusTarget>,
+    pub(crate) targets: Vec<ReplicationResetStatusTarget>,
 }

 #[derive(Debug, Clone, serde::Deserialize)]
-struct ReplicationResetStatusTarget {
+pub(crate) struct ReplicationResetStatusTarget {
    #[serde(rename = "arn", default)]
-    arn: String,
+    pub(crate) arn: String,
    #[serde(rename = "resetid", default)]
-    reset_id: String,
+    pub(crate) reset_id: String,
    #[serde(rename = "resyncStatus", default)]
-    status: String,
+    pub(crate) status: String,
    #[serde(rename = "replicationCount", default)]
-    replicated_count: i64,
+    pub(crate) replicated_count: i64,
    #[serde(rename = "object", default)]
-    object: String,
+    pub(crate) object: String,
 }

 fn extract_xml_tag(xml: &str, tag: &str) -> Option<String> {
@@ -2294,7 +2294,7 @@ async fn site_replication_state_edit(
 /// return the target `(arn, reset_id)`, asserting the response carries the
 /// madmin `ResyncTargetsInfo` shape (`target[0].arn` / `target[0].resetid`)
 /// that `mc replicate resync start` decodes.
-async fn start_bucket_replication_reset(
+pub(crate) async fn start_bucket_replication_reset(
    env: &RustFSTestEnvironment,
    bucket: &str,
 ) -> Result<(String, String), Box<dyn Error + Send + Sync>> {
@@ -2314,7 +2314,7 @@ async fn start_bucket_replication_reset(
    Ok((arn, reset_id))
 }

-async fn get_replication_reset_status(
+pub(crate) async fn get_replication_reset_status(
    env: &RustFSTestEnvironment,
    bucket: &str,
    arn: &str,
@@ -3837,6 +3837,244 @@ async fn test_bucket_replication_converges_delete_marker_and_version_purge() ->
    Ok(())
 }

+/// Regression for rustfs/backlog#2340 (not Wasabi specific): a directory
+/// marker (`prefix/` with a body) in a versioned bucket is stored as the null
+/// version, like MinIO (`putOpts`: "for directory objects skip creating new
+/// versions"), and must still replicate to completion instead of staying
+/// `PENDING`.
+#[tokio::test]
+async fn test_bucket_replication_replicates_directory_marker_in_versioned_bucket() -> TestResult {
+    init_logging();
+
+    let mut source_env = RustFSTestEnvironment::new().await?;
+    let mut source_env_vars = replication_fast_env();
+    source_env_vars.extend_from_slice(LOOPBACK_REPLICATION_TARGET_ENV);
+    source_env.start_rustfs_server_with_env(vec![], &source_env_vars).await?;
+
+    let mut target_env = RustFSTestEnvironment::new().await?;
+    target_env.start_rustfs_server_without_cleanup(vec![]).await?;
+
+    let source_bucket = "replication-dir-marker-src";
+    let target_bucket = "replication-dir-marker-dst";
+    let source_client = source_env.create_s3_client();
+    let target_client = target_env.create_s3_client();
+
+    source_client.create_bucket().bucket(source_bucket).send().await?;
+    target_client.create_bucket().bucket(target_bucket).send().await?;
+    enable_bucket_versioning(&source_env, source_bucket).await?;
+    enable_bucket_versioning(&target_env, target_bucket).await?;
+    let target_arn = set_replication_target(&source_env, source_bucket, &target_env, target_bucket).await?;
+    put_bucket_replication(&source_env, source_bucket, &target_arn).await?;
+
+    let marker_key = "dir/trailing/";
+    let body = b"directory marker body";
+    let put = source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(marker_key)
+        .body(ByteStream::from_static(body))
+        .send()
+        .await?;
+    assert!(
+        put.version_id()
+            .is_none_or(|id| id == "null" || id == uuid::Uuid::nil().to_string()),
+        "a directory marker is the null version even in a versioned bucket: {:?}",
+        put.version_id()
+    );
+
+    wait_for_source_replication_status(&source_client, source_bucket, marker_key, "COMPLETED", false).await?;
+
+    let replica = target_client
+        .get_object()
+        .bucket(target_bucket)
+        .key(marker_key)
+        .send()
+        .await?;
+    assert_eq!(replica.body.collect().await?.into_bytes().as_ref(), body);
+    let listed = target_client
+        .list_object_versions()
+        .bucket(target_bucket)
+        .prefix(marker_key)
+        .send()
+        .await?;
+    let marker_versions: Vec<_> = listed.versions().iter().filter(|v| v.key() == Some(marker_key)).collect();
+    assert_eq!(marker_versions.len(), 1, "the marker must land exactly once: {marker_versions:?}");
+    assert_eq!(
+        marker_versions[0].version_id(),
+        Some("null"),
+        "the replica keeps the null version identity"
+    );
+
+    Ok(())
+}
+
+/// Regression for rustfs/backlog#2340 (not Wasabi specific): permanently
+/// deleting a version whose payload lives in a data dir must leave the source
+/// clean once the purge replicates. Managed-SSE objects are never inlined and a
+/// plain object above the inline threshold takes the same layout. The version
+/// retained with a pending purge used to lose its data dir, so the purge state
+/// could never be applied (`VersionNotFound` on every retry) and the bucket
+/// stayed `BucketNotEmpty` while `ListObjectVersions` was already empty.
+#[tokio::test]
+async fn test_bucket_replication_version_purge_of_non_inline_object_releases_source_bucket() -> TestResult {
+    init_logging();
+
+    let (source_env, target_env, source_bucket, target_bucket) = build_sse_replication_pair("purge-datadir", true, true).await?;
+    let target_arn = wait_for_remote_target_arn(&source_env, &source_bucket).await?;
+    put_bucket_replication_with_delete_statuses(&source_env, &source_bucket, &target_arn, "Enabled", Some("Enabled")).await?;
+    let source_client = source_env.create_s3_client();
+    let target_client = target_env.create_s3_client();
+
+    let sse_key = "sse-object.bin";
+    let large_key = "large-object.bin";
+    let sse_put = source_client
+        .put_object()
+        .bucket(&source_bucket)
+        .key(sse_key)
+        .body(ByteStream::from_static(b"encrypted source payload"))
+        .server_side_encryption(ServerSideEncryption::Aes256)
+        .send()
+        .await?;
+    let large_put = source_client
+        .put_object()
+        .bucket(&source_bucket)
+        .key(large_key)
+        .body(ByteStream::from(vec![0x5a; 2 * 1024 * 1024]))
+        .send()
+        .await?;
+    let purged = [
+        (sse_key, sse_put.version_id().ok_or("SSE PUT omitted version ID")?.to_string()),
+        (large_key, large_put.version_id().ok_or("large PUT omitted version ID")?.to_string()),
+    ];
+    assert_replication_converged(&source_client, &source_bucket, &target_client, &target_bucket).await?;
+
+    for (key, version_id) in &purged {
+        source_client
+            .delete_object()
+            .bucket(&source_bucket)
+            .key(*key)
+            .version_id(version_id)
+            .send()
+            .await?;
+    }
+    assert_replication_converged(&source_client, &source_bucket, &target_client, &target_bucket).await?;
+    let target_state = list_replication_state(&target_client, &target_bucket).await?;
+    assert!(target_state.is_empty(), "target retained an explicitly purged version: {target_state:?}");
+
+    // The purge state is applied on the source asynchronously after the target
+    // acknowledges the delete; only then does the retained version go away and
+    // the bucket become deletable. A listing that is empty while DeleteBucket
+    // keeps answering BucketNotEmpty is exactly the regression.
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(60);
+    loop {
+        let listing = source_client.list_object_versions().bucket(&source_bucket).send().await?;
+        let listed = listing.versions().len() + listing.delete_markers().len();
+        match source_client.delete_bucket().bucket(&source_bucket).send().await {
+            Ok(_) => break,
+            Err(err) if err.code() == Some("BucketNotEmpty") => {
+                if tokio::time::Instant::now() >= deadline {
+                    return Err(format!(
+                        "source bucket stayed BucketNotEmpty after the version purge replicated; \
+                         ListObjectVersions shows {listed} entries"
+                    )
+                    .into());
+                }
+                sleep(Duration::from_millis(500)).await;
+            }
+            Err(err) => return Err(err.into()),
+        }
+    }
+
+    Ok(())
+}
+
+/// Regression for rustfs/backlog#2340 (not Wasabi specific): a single-part
+/// object uploaded with `x-amz-checksum-*` must reach the target with the same
+/// checksum. The outbound options keyed the stored record by algorithm name,
+/// which the target client sent as `x-amz-meta-*` user metadata, so a replica
+/// never carried a checksum although the source HEAD returned one.
+#[tokio::test]
+async fn test_bucket_replication_forwards_single_part_object_checksums() -> TestResult {
+    init_logging();
+
+    let mut source_env = RustFSTestEnvironment::new().await?;
+    let mut source_env_vars = replication_fast_env();
+    source_env_vars.extend_from_slice(LOOPBACK_REPLICATION_TARGET_ENV);
+    source_env.start_rustfs_server_with_env(vec![], &source_env_vars).await?;
+
+    let mut target_env = RustFSTestEnvironment::new().await?;
+    target_env.start_rustfs_server_without_cleanup(vec![]).await?;
+
+    let source_bucket = "replication-checksum-src";
+    let target_bucket = "replication-checksum-dst";
+    let source_client = source_env.create_s3_client();
+    let target_client = target_env.create_s3_client();
+
+    source_client.create_bucket().bucket(source_bucket).send().await?;
+    target_client.create_bucket().bucket(target_bucket).send().await?;
+    enable_bucket_versioning(&source_env, source_bucket).await?;
+    enable_bucket_versioning(&target_env, target_bucket).await?;
+    let target_arn = set_replication_target(&source_env, source_bucket, &target_env, target_bucket).await?;
+    put_bucket_replication(&source_env, source_bucket, &target_arn).await?;
+
+    let body = b"123456789";
+    let crc32_key = "checksum-crc32.txt";
+    let sha256_key = "checksum-sha256.txt";
+    let crc32_put = source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(crc32_key)
+        .body(ByteStream::from_static(body))
+        .checksum_algorithm(aws_sdk_s3::types::ChecksumAlgorithm::Crc32)
+        .send()
+        .await?;
+    let expected_crc32 = crc32_put.checksum_crc32().ok_or("source PUT omitted CRC32")?.to_string();
+    let sha256_put = source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(sha256_key)
+        .body(ByteStream::from_static(body))
+        .checksum_algorithm(aws_sdk_s3::types::ChecksumAlgorithm::Sha256)
+        .send()
+        .await?;
+    let expected_sha256 = sha256_put.checksum_sha256().ok_or("source PUT omitted SHA256")?.to_string();
+
+    for key in [crc32_key, sha256_key] {
+        wait_for_source_replication_status(&source_client, source_bucket, key, "COMPLETED", false).await?;
+    }
+
+    let replica = target_client
+        .head_object()
+        .bucket(target_bucket)
+        .key(crc32_key)
+        .checksum_mode(aws_sdk_s3::types::ChecksumMode::Enabled)
+        .send()
+        .await?;
+    assert_eq!(replica.checksum_crc32(), Some(expected_crc32.as_str()), "replica lost the CRC32 checksum");
+    let replica = target_client
+        .head_object()
+        .bucket(target_bucket)
+        .key(sha256_key)
+        .checksum_mode(aws_sdk_s3::types::ChecksumMode::Enabled)
+        .send()
+        .await?;
+    assert_eq!(
+        replica.checksum_sha256(),
+        Some(expected_sha256.as_str()),
+        "replica lost the SHA256 checksum"
+    );
+    // The bare algorithm name must not leak as user metadata either.
+    assert!(
+        replica
+            .metadata()
+            .is_none_or(|meta| !meta.keys().any(|k| k.eq_ignore_ascii_case("sha256"))),
+        "replica carries the checksum as user metadata: {:?}",
+        replica.metadata()
+    );
+
+    Ok(())
+}
+
 #[tokio::test]
 async fn test_bucket_replication_disabled_delete_marker_does_not_propagate() -> TestResult {
    init_logging();
@@ -31,17 +31,20 @@
 //! Adding a target behavior the fleet has shown: add the mode to the fake
 //! target, add a row here, and record any cell that is red before the fix.

-use crate::common::{RustFSTestEnvironment, init_logging, replication_fast_env};
-use crate::fake_s3_target::{FAKE_ACCESS_KEY, FAKE_SECRET_KEY};
-use crate::fake_s3_target::{FakeS3Target, Operation as FakeTargetOperation, RequestRecord};
-use crate::on_demand_migration::common::fake_source_client;
+use crate::common::{init_logging, replication_fast_env};
+use crate::fake_s3_target::{BucketMode, FAKE_ACCESS_KEY, FAKE_SECRET_KEY};
+use crate::fake_s3_target::{FakeS3Target, FaultAction as FakeTargetFault, Operation as FakeTargetOperation, RequestRecord};
+use crate::on_demand_migration::common::{OdmEnvOptions, OdmTestEnv, fake_source_client};
 use crate::replication_extension_test::{
-    LOOPBACK_REPLICATION_TARGET_ENV, ReplicationTargetOptions, enable_bucket_versioning, put_bucket_replication,
-    set_replication_target_with_options,
+    LOOPBACK_REPLICATION_TARGET_ENV, ReplicationTargetOptions, enable_bucket_versioning, get_replication_reset_status,
+    put_bucket_replication, set_replication_target_with_options, start_bucket_replication_reset,
 };
 use aws_sdk_s3::Client;
 use aws_sdk_s3::primitives::{ByteStream, DateTime};
-use aws_sdk_s3::types::{CompletedMultipartUpload, CompletedPart, ObjectLockLegalHoldStatus, ObjectLockMode};
+use aws_sdk_s3::types::{
+    Checksum, ChecksumAlgorithm, CompletedMultipartUpload, CompletedPart, ObjectAttributes, ObjectLockLegalHoldStatus,
+    ObjectLockMode,
+};
 use bytes::Bytes;
 use std::error::Error;
 use std::time::{SystemTime, UNIX_EPOCH};
@@ -110,16 +113,23 @@ enum ObjectShape {
    /// Two-part multipart upload with a GOVERNANCE retention period; the
    /// lock headers travel on CreateMultipartUpload, which has no body.
    LockedMultipart,
+    /// ODM stores two local parts while preserving a single-PUT source's MD5 ETag.
+    OdmPreservedMd5Multipart,
+    /// Single-part object uploaded with `x-amz-checksum-sha256`; the replica
+    /// must carry the same header (rustfs/backlog#2340).
+    Checksummed,
 }

 impl ObjectShape {
-    const ALL: [ObjectShape; 6] = [
+    const ALL: [ObjectShape; 8] = [
        ObjectShape::Empty,
        ObjectShape::Plain,
        ObjectShape::Retention,
        ObjectShape::LegalHold,
        ObjectShape::Multipart,
        ObjectShape::LockedMultipart,
+        ObjectShape::OdmPreservedMd5Multipart,
+        ObjectShape::Checksummed,
    ];

    fn key(self) -> &'static str {
@@ -130,6 +140,17 @@ impl ObjectShape {
            ObjectShape::LegalHold => "matrix/legal-hold.bin",
            ObjectShape::Multipart => "matrix/multipart.bin",
            ObjectShape::LockedMultipart => "matrix/locked-multipart.bin",
+            ObjectShape::OdmPreservedMd5Multipart => "matrix/odm-preserved-md5.bin",
+            ObjectShape::Checksummed => "matrix/checksummed.bin",
+        }
+    }
+
+    /// The `x-amz-checksum-*` header the source stored and every upload of
+    /// the replica must repeat.
+    fn forwarded_checksum_header(self) -> Option<&'static str> {
+        match self {
+            ObjectShape::Checksummed => Some("x-amz-checksum-sha256"),
+            _ => None,
        }
    }

@@ -139,7 +160,8 @@ impl ObjectShape {

    /// Upload the shape to the source and return the bytes the target must
    /// end up holding.
-    async fn put(self, client: &Client, bucket: &str) -> Result<Bytes, Box<dyn Error + Send + Sync>> {
+    async fn put(self, env: &OdmTestEnv, bucket: &str) -> Result<Bytes, Box<dyn Error + Send + Sync>> {
+        let client = &env.client;
        let key = self.key();
        match self {
            ObjectShape::Empty => {
@@ -190,6 +212,19 @@ impl ObjectShape {
            }
            ObjectShape::Multipart => multipart_put(client, bucket, key, 0x44, false).await,
            ObjectShape::LockedMultipart => multipart_put(client, bucket, key, 0x55, true).await,
+            ObjectShape::OdmPreservedMd5Multipart => odm_preserved_md5_multipart(env, bucket, key).await,
+            ObjectShape::Checksummed => {
+                let body = payload(40 * 1024, 0x66);
+                client
+                    .put_object()
+                    .bucket(bucket)
+                    .key(key)
+                    .body(ByteStream::from(body.clone()))
+                    .checksum_algorithm(ChecksumAlgorithm::Sha256)
+                    .send()
+                    .await?;
+                Ok(body)
+            }
        }
    }
 }
@@ -219,6 +254,170 @@ fn expectation(mode: TargetMode, shape: ObjectShape) -> Expectation {
        .unwrap_or(Expectation::Completed)
 }

+/// rustfs/backlog#2340: a target that mints its own version ids (Wasabi,
+/// AWS S3) answers 404 to a HEAD by the source uuid, which the worker used to
+/// read as "replica missing" and re-drive the PUT — one more target version
+/// per heal, MRF retry or resync. Two re-drive shapes, both must converge on
+/// the single version the first PUT created:
+/// - the first PUT lands but its response is lost, so the object is FAILED
+///   and the scanner heal pass re-drives it;
+/// - an existing-object resync re-drives a COMPLETED object unconditionally.
+#[tokio::test]
+async fn matrix_mint_own_version_ids_redrive_does_not_duplicate() -> TestResult {
+    init_logging();
+
+    let target = FakeS3Target::start().await?;
+    let target_bucket = "matrix-mint-own-redrive-dst".to_string();
+    target.create_bucket_with_object_lock(target_bucket.clone());
+    target.assign_own_version_ids(true);
+
+    let mut env_vars = replication_fast_env();
+    env_vars.extend_from_slice(LOOPBACK_REPLICATION_TARGET_ENV);
+    env_vars.extend_from_slice(&[
+        ("NO_PROXY", "127.0.0.1,localhost"),
+        ("HTTP_PROXY", ""),
+        ("HTTPS_PROXY", ""),
+        // The scanner heal pass is what re-drives a FAILED object.
+        ("RUSTFS_SCANNER_CYCLE", "1"),
+        ("RUSTFS_SCANNER_START_DELAY_SECS", "1"),
+    ]);
+    let env = OdmTestEnv::start_with(OdmEnvOptions {
+        env: env_vars,
+        ..OdmEnvOptions::default()
+    })
+    .await?;
+    let source_env = &env.rustfs;
+
+    let source_bucket = "matrix-mint-own-redrive-src";
+    let source_client = source_env.create_s3_client();
+    source_client
+        .create_bucket()
+        .bucket(source_bucket)
+        .object_lock_enabled_for_bucket(true)
+        .send()
+        .await?;
+    enable_bucket_versioning(source_env, source_bucket).await?;
+    let target_arn = set_replication_target_with_options(
+        source_env,
+        source_bucket,
+        ReplicationTargetOptions {
+            endpoint: &target.address(),
+            access_key: FAKE_ACCESS_KEY,
+            secret_key: FAKE_SECRET_KEY,
+            target_bucket: &target_bucket,
+            secure: false,
+            skip_tls_verify: false,
+            ca_cert_pem: None,
+        },
+    )
+    .await?;
+    put_bucket_replication(source_env, source_bucket, &target_arn).await?;
+
+    // Teach the worker the target's identity contract with one ordinary
+    // write, exactly as production learns it (the PUT response carries the
+    // minted id).
+    let probe_key = "redrive/identity-probe.bin";
+    source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(probe_key)
+        .body(ByteStream::from(payload(4 * 1024, 0x01)))
+        .send()
+        .await?;
+    assert_eq!(
+        wait_for_terminal_replication_status(&source_client, source_bucket, probe_key).await?,
+        "COMPLETED"
+    );
+
+    // Shape 1: the PUT is stored, its response never arrives, heal re-drives.
+    let heal_key = "redrive/heal.bin";
+    target.inject_for_key(FakeTargetOperation::PutObject, heal_key, FakeTargetFault::DisconnectAfterResponse, 1);
+    source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(heal_key)
+        .body(ByteStream::from(payload(8 * 1024, 0x02)))
+        .send()
+        .await?;
+    wait_for_replication_status_and_single_version(&source_client, source_bucket, &target, &target_bucket, heal_key).await?;
+
+    // Shape 2: an existing-object resync re-drives a COMPLETED object.
+    let resync_key = "redrive/resync.bin";
+    source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(resync_key)
+        .body(ByteStream::from(payload(8 * 1024, 0x03)))
+        .send()
+        .await?;
+    assert_eq!(
+        wait_for_terminal_replication_status(&source_client, source_bucket, resync_key).await?,
+        "COMPLETED"
+    );
+    let (reset_arn, _reset_id) = start_bucket_replication_reset(source_env, source_bucket).await?;
+    assert_eq!(reset_arn, target_arn);
+    let resync = async {
+        loop {
+            let status = get_replication_reset_status(source_env, source_bucket, &target_arn).await?;
+            if let Some(entry) = status.targets.iter().find(|entry| entry.arn == target_arn)
+                && entry.status == "Completed"
+            {
+                return Ok::<_, Box<dyn Error + Send + Sync>>(entry.replicated_count);
+            }
+            sleep(Duration::from_millis(250)).await;
+        }
+    };
+    let replicated = timeout(Duration::from_secs(90), resync)
+        .await
+        .map_err(|_| "existing-object resync did not complete within 90 seconds")??;
+    assert!(replicated >= 3, "resync must count the located replicas as replicated, got {replicated}");
+    for key in [probe_key, heal_key, resync_key] {
+        let versions = target.stored_versions(&target_bucket, key);
+        assert_eq!(
+            versions.len(),
+            1,
+            "{key}: a re-drive against a target that mints its own version ids must not mint another one: {versions:?}"
+        );
+    }
+
+    target.shutdown().await;
+    Ok(())
+}
+
+/// Wait until `key` is COMPLETED on the source and, for the observation
+/// window after that, the target still holds exactly one live version of it.
+async fn wait_for_replication_status_and_single_version(
+    source_client: &Client,
+    source_bucket: &str,
+    target: &FakeS3Target,
+    target_bucket: &str,
+    key: &str,
+) -> TestResult {
+    // The lost PUT response first settles the object FAILED; only the next
+    // scanner heal pass can turn that into COMPLETED, so FAILED is transient
+    // here and the wait is for COMPLETED alone.
+    let converged = async {
+        loop {
+            let head = source_client.head_object().bucket(source_bucket).key(key).send().await?;
+            if head.replication_status().is_some_and(|status| status.as_str() == "COMPLETED") {
+                return Ok::<_, Box<dyn Error + Send + Sync>>(());
+            }
+            sleep(Duration::from_millis(250)).await;
+        }
+    };
+    timeout(Duration::from_secs(90), converged)
+        .await
+        .map_err(|_| format!("{key}: heal re-drive did not converge to COMPLETED within 90 seconds"))??;
+    // The heal pass keeps visiting the key for a few scanner cycles; a
+    // duplicate would show up here as a second stored version.
+    for _ in 0..12 {
+        let versions = target.stored_versions(target_bucket, key);
+        assert_eq!(versions.len(), 1, "{key}: target minted another version on re-drive: {versions:?}");
+        sleep(Duration::from_millis(500)).await;
+    }
+    Ok(())
+}
+
 #[tokio::test]
 async fn matrix_baseline_target() -> TestResult {
    run_row(TargetMode::Baseline).await
@@ -270,11 +469,15 @@ async fn run_row(mode: TargetMode) -> TestResult {
    target.create_bucket_with_object_lock(target_bucket.clone());
    mode.apply(&target);

-    let mut source_env = RustFSTestEnvironment::new().await?;
    let mut env_vars = replication_fast_env();
    env_vars.extend_from_slice(LOOPBACK_REPLICATION_TARGET_ENV);
    env_vars.extend_from_slice(&[("NO_PROXY", "127.0.0.1,localhost"), ("HTTP_PROXY", ""), ("HTTPS_PROXY", "")]);
-    source_env.start_rustfs_server_with_env(vec![], &env_vars).await?;
+    let env = OdmTestEnv::start_with(OdmEnvOptions {
+        env: env_vars,
+        ..OdmEnvOptions::default()
+    })
+    .await?;
+    let source_env = &env.rustfs;

    let source_bucket = format!("matrix-{}-src", mode.slug());
    let source_client = source_env.create_s3_client();
@@ -284,9 +487,9 @@ async fn run_row(mode: TargetMode) -> TestResult {
        .object_lock_enabled_for_bucket(true)
        .send()
        .await?;
-    enable_bucket_versioning(&source_env, &source_bucket).await?;
+    enable_bucket_versioning(source_env, &source_bucket).await?;
    let target_arn = set_replication_target_with_options(
-        &source_env,
+        source_env,
        &source_bucket,
        ReplicationTargetOptions {
            endpoint: &target.address(),
@@ -299,14 +502,21 @@ async fn run_row(mode: TargetMode) -> TestResult {
        },
    )
    .await?;
-    put_bucket_replication(&source_env, &source_bucket, &target_arn).await?;
+    put_bucket_replication(source_env, &source_bucket, &target_arn).await?;

    let target_client = fake_source_client(&target);
    let mut failures = Vec::new();
    for shape in ObjectShape::ALL {
        let cell = format!("{}/{:?}", mode.slug(), shape);
-        let expected_body = shape.put(&source_client, &source_bucket).await?;
+        let expected_body = shape.put(&env, &source_bucket).await?;
        let status = wait_for_terminal_replication_status(&source_client, &source_bucket, shape.key()).await?;
+        if shape == ObjectShape::OdmPreservedMd5Multipart {
+            assert_eq!(
+                env.source.count_requests(FakeTargetOperation::GetObject, shape.key()),
+                2,
+                "one passthrough GET plus one background pull; replication must read the persisted local parts"
+            );
+        }
        let journal = target.requests();
        let outcome = match expectation(mode, shape) {
            Expectation::Completed => {
@@ -379,6 +589,36 @@ async fn check_completed_cell(
    if uploads.is_empty() {
        return Err("no upload reached the target although the source reports COMPLETED".into());
    }
+    if shape == ObjectShape::OdmPreservedMd5Multipart {
+        let key_requests: Vec<_> = journal
+            .iter()
+            .filter(|record| record.key.as_deref() == Some(shape.key()))
+            .collect();
+        for operation in [
+            FakeTargetOperation::CreateMultipartUpload,
+            FakeTargetOperation::CompleteMultipartUpload,
+        ] {
+            if !key_requests.iter().any(|record| record.operation == operation) {
+                return Err(format!("preserved-MD5 multipart object did not use {operation:?}").into());
+            }
+        }
+        if key_requests
+            .iter()
+            .any(|record| record.operation == FakeTargetOperation::PutObject)
+        {
+            return Err("preserved-MD5 multipart object used a single PutObject".into());
+        }
+        let mut part_numbers: Vec<_> = key_requests
+            .iter()
+            .filter(|record| record.operation == FakeTargetOperation::UploadPart)
+            .map(|record| record.part_number)
+            .collect();
+        part_numbers.sort_unstable();
+        part_numbers.dedup();
+        if part_numbers != [Some(1), Some(2)] {
+            return Err(format!("preserved-MD5 multipart object uploaded unexpected parts: {part_numbers:?}").into());
+        }
+    }
    if let Some(framed) = uploads.iter().find(|record| record.transport.aws_chunked) {
        return Err(format!("{cell}: an upload went out aws-chunked (rustfs#6853 framing): {framed:?}").into());
    }
@@ -401,6 +641,19 @@ async fn check_completed_cell(
    }) {
        return Err(format!("a locked PutObject went out without any integrity header (rustfs#7082): {bare:?}").into());
    }
+    // rustfs/backlog#2340 contract: a source checksum reaches the target as
+    // the `x-amz-checksum-*` header, not as user metadata; every PutObject of
+    // the shape carries it.
+    if let Some(header) = shape.forwarded_checksum_header()
+        && let Some(missing) = uploads.iter().find(|record| {
+            record.operation == FakeTargetOperation::PutObject
+                && !record.transport.checksum_headers.iter().any(|name| name == header)
+        })
+    {
+        return Err(
+            format!("a PutObject went out without the source's {header} header (rustfs/backlog#2340): {missing:?}").into(),
+        );
+    }
    Ok(())
 }

@@ -455,6 +708,71 @@ async fn wait_for_terminal_replication_status(
    }
 }

+async fn odm_preserved_md5_multipart(env: &OdmTestEnv, bucket: &str, key: &str) -> Result<Bytes, Box<dyn Error + Send + Sync>> {
+    const PART_SIZE: usize = 5 * 1024 * 1024;
+    let origin_bucket = format!("{bucket}-origin");
+    env.source.create_bucket_with_mode(&origin_bucket, BucketMode::Unversioned);
+    let mut spec = env.fake_source_spec(&origin_bucket);
+    // Below the 16 MiB inline default the pull is one tee'd PUT with a single
+    // part; force the passthrough + background multipart write-back instead.
+    spec.policy.inline_max_bytes = 4096;
+    spec.policy.multipart_part_size_bytes = PART_SIZE as u64;
+    spec.policy.preserve_etag = true;
+    env.configure_and_wait(bucket, &spec).await?;
+
+    // A normal source PUT produces the MD5 ETag; only ODM chooses the local parts.
+    let body = payload(PART_SIZE + 4096, 0x66);
+    let source_put = env
+        .source_client()
+        .put_object()
+        .bucket(&origin_bucket)
+        .key(key)
+        .body(ByteStream::from(body.clone()))
+        .send()
+        .await?;
+    let source_etag = source_put.e_tag().ok_or("source PUT omitted its ETag")?.trim_matches('"');
+    assert_eq!(source_etag.len(), 32, "source fixture must have a single-PUT MD5 ETag");
+    assert!(source_etag.bytes().all(|byte| byte.is_ascii_hexdigit()));
+
+    let pulled = env.raw_get(bucket, key).await?;
+    assert_eq!(pulled.status, 200, "{}", String::from_utf8_lossy(&pulled.body));
+    assert_eq!(pulled.body, body);
+    assert!(
+        env.wait_local_listed(bucket, key, Duration::from_secs(30)).await?,
+        "ODM must persist the object"
+    );
+    let attributes = env
+        .client
+        .get_object_attributes()
+        .bucket(bucket)
+        .key(key)
+        .object_attributes(ObjectAttributes::Etag)
+        .object_attributes(ObjectAttributes::ObjectParts)
+        .object_attributes(ObjectAttributes::Checksum)
+        .send()
+        .await?;
+    assert_eq!(attributes.e_tag().map(|etag| etag.trim_matches('"')), Some(source_etag));
+    let parts = attributes
+        .object_parts()
+        .ok_or("the ODM copy must expose its two local parts")?;
+    assert_eq!(parts.total_parts_count(), Some(2));
+    assert_eq!(
+        parts
+            .parts()
+            .iter()
+            .map(|part| (part.part_number(), part.size()))
+            .collect::<Vec<_>>(),
+        [(Some(1), Some(PART_SIZE as i64)), (Some(2), Some(4096))]
+    );
+    assert!(
+        attributes
+            .checksum()
+            .is_none_or(|checksum| checksum == &Checksum::builder().build()),
+        "multipart routing must work without an object checksum record"
+    );
+    Ok(body)
+}
+
 async fn multipart_put(
    client: &Client,
    bucket: &str,
@@ -14,9 +14,10 @@

 use crate::common::{
    RustFSTestClusterEnvironment, RustFSTestEnvironment, admin_request, init_logging, replication_fast_env, rustfs_binary_path,
+    signed_request,
 };
-use crate::fake_s3_target::{BucketMode, FAKE_ACCESS_KEY, FAKE_SECRET_KEY, FakeS3Target};
-use crate::on_demand_migration::common::{ODM_SERVER_ENV, OdmTestEnv, SeedObject};
+use crate::fake_s3_target::{BucketMode, FAKE_ACCESS_KEY, FAKE_SECRET_KEY, FakeS3Target, Operation as FakeTargetOperation};
+use crate::on_demand_migration::common::{ODM_SERVER_ENV, OdmTestEnv, SeedObject, fake_source_client};
 use crate::replication_extension_test::{
    LOOPBACK_REPLICATION_TARGET_ENV, ReplicationTargetOptions, put_bucket_replication, set_replication_target_with_options,
 };
@@ -25,9 +26,10 @@ use aws_sdk_s3::error::ProvideErrorMetadata;
 use aws_sdk_s3::primitives::ByteStream;
 use aws_sdk_s3::types::{
    BucketLifecycleConfiguration, BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, DefaultRetention,
-    ExpirationStatus, LifecycleExpiration, LifecycleRule, LifecycleRuleFilter, ObjectLockConfiguration, ObjectLockEnabled,
-    ObjectLockRetentionMode, ObjectLockRule, PublicAccessBlockConfiguration, ServerSideEncryption, ServerSideEncryptionByDefault,
-    ServerSideEncryptionConfiguration, ServerSideEncryptionRule, Tag, Tagging, VersioningConfiguration,
+    ExpirationStatus, LifecycleExpiration, LifecycleRule, LifecycleRuleFilter, ObjectAttributes, ObjectLockConfiguration,
+    ObjectLockEnabled, ObjectLockRetentionMode, ObjectLockRule, PublicAccessBlockConfiguration, ServerSideEncryption,
+    ServerSideEncryptionByDefault, ServerSideEncryptionConfiguration, ServerSideEncryptionRule, Tag, Tagging,
+    VersioningConfiguration,
 };
 use http::{Method, StatusCode};
 use std::path::{Path, PathBuf};
@@ -1205,3 +1207,724 @@ async fn rollback_to_previous_release_reads_current_bucket_metadata() -> TestRes
    replication_target.shutdown().await;
    Ok(())
 }
+
+// ---------------------------------------------------------------------------
+// rc.5 multipart layouts under the current build (backlog#2147 follow-up to
+// rustfs#7305)
+// ---------------------------------------------------------------------------
+//
+// rustfs#7305 changed `ObjectInfo::is_multipart` to consult the stored part
+// list before the ETag shape. Every earlier check of that change used
+// synthetic metadata; this scenario writes the layouts with the published
+// rc.5 binary and then reads, describes, and replicates them with the
+// current build on the same data directory.
+
+const LAYOUT_PLAIN_BUCKET: &str = "upgrade-layout-plain";
+const LAYOUT_ENCRYPTED_BUCKET: &str = "upgrade-layout-encrypted";
+const LAYOUT_REPLICA_BUCKET: &str = "upgrade-layout-replica";
+const LAYOUT_PART_SIZE: usize = 5 * 1024 * 1024;
+const LAYOUT_TAIL_SIZE: usize = 1024 * 1024 + 4096;
+const LAYOUT_SSEC_KEY: &str = "0123456789abcdef0123456789abcdef";
+const LAYOUT_REPLICATION_TIMEOUT: Duration = Duration::from_secs(180);
+
+struct LayoutCase {
+    bucket: &'static str,
+    key: &'static str,
+    /// Empty for a single PUT.
+    part_sizes: Vec<usize>,
+    body: Vec<u8>,
+    ssec: bool,
+    /// `false` for layouts whose replication is a known pre-existing failure;
+    /// their outcome is logged, not asserted.
+    assert_replication: bool,
+    /// Recorded from the rc.5 writer.
+    rc5_etag: String,
+    /// Whether rc.5 reported `ObjectParts` for the object.
+    rc5_reported_parts: Option<usize>,
+}
+
+impl LayoutCase {
+    fn is_multipart_layout(&self) -> bool {
+        self.part_sizes.len() > 1
+    }
+
+    fn label(&self) -> String {
+        format!("{}/{}", self.bucket, self.key)
+    }
+}
+
+fn layout_noise(len: usize, seed: u64) -> Vec<u8> {
+    let mut state = seed ^ 0x9E37_79B9_7F4A_7C15;
+    (0..len)
+        .map(|_| {
+            state ^= state << 13;
+            state ^= state >> 7;
+            state ^= state << 17;
+            (state >> 24) as u8
+        })
+        .collect()
+}
+
+fn layout_text(len: usize, seed: u64) -> Vec<u8> {
+    let mut out = Vec::with_capacity(len + 64);
+    let mut line = 0u64;
+    while out.len() < len {
+        out.extend_from_slice(format!("rc5 legacy layout seed={seed} line={line} lorem ipsum dolor sit amet\n").as_bytes());
+        line += 1;
+    }
+    out.truncate(len);
+    out
+}
+
+fn layout_ssec_key_md5() -> String {
+    use md5::{Digest as _, Md5};
+    let mut hasher = Md5::new();
+    hasher.update(LAYOUT_SSEC_KEY.as_bytes());
+    base64_simd::STANDARD.encode_to_string(hasher.finalize())
+}
+
+fn layout_ssec_key() -> String {
+    base64_simd::STANDARD.encode_to_string(LAYOUT_SSEC_KEY)
+}
+
+async fn layout_head(
+    client: &Client,
+    case: &LayoutCase,
+) -> Result<aws_sdk_s3::operation::head_object::HeadObjectOutput, BoxError> {
+    let request = client.head_object().bucket(case.bucket).key(case.key);
+    let request = if case.ssec {
+        request
+            .sse_customer_algorithm("AES256")
+            .sse_customer_key(layout_ssec_key())
+            .sse_customer_key_md5(layout_ssec_key_md5())
+    } else {
+        request
+    };
+    Ok(request.send().await?)
+}
+
+async fn layout_get(
+    client: &Client,
+    case: &LayoutCase,
+    range: Option<String>,
+    part_number: Option<i32>,
+) -> Result<aws_sdk_s3::operation::get_object::GetObjectOutput, BoxError> {
+    let request = client.get_object().bucket(case.bucket).key(case.key);
+    let request = if case.ssec {
+        request
+            .sse_customer_algorithm("AES256")
+            .sse_customer_key(layout_ssec_key())
+            .sse_customer_key_md5(layout_ssec_key_md5())
+    } else {
+        request
+    };
+    let request = request.set_range(range).set_part_number(part_number);
+    Ok(request.send().await?)
+}
+
+async fn layout_attributes(
+    client: &Client,
+    case: &LayoutCase,
+) -> Result<aws_sdk_s3::operation::get_object_attributes::GetObjectAttributesOutput, BoxError> {
+    let request = client
+        .get_object_attributes()
+        .bucket(case.bucket)
+        .key(case.key)
+        .object_attributes(ObjectAttributes::Etag)
+        .object_attributes(ObjectAttributes::ObjectParts)
+        .object_attributes(ObjectAttributes::ObjectSize)
+        .max_parts(100);
+    let request = if case.ssec {
+        request
+            .sse_customer_algorithm("AES256")
+            .sse_customer_key(layout_ssec_key())
+            .sse_customer_key_md5(layout_ssec_key_md5())
+    } else {
+        request
+    };
+    Ok(request.send().await?)
+}
+
+/// Write `case` with the rc.5 client; single PUT when `part_sizes` is empty.
+async fn layout_write(client: &Client, case: &LayoutCase) -> Result<(), BoxError> {
+    let content_type = "text/plain";
+    if case.part_sizes.is_empty() {
+        let request = client
+            .put_object()
+            .bucket(case.bucket)
+            .key(case.key)
+            .content_type(content_type)
+            .body(ByteStream::from(case.body.clone()));
+        let request = if case.ssec {
+            request
+                .sse_customer_algorithm("AES256")
+                .sse_customer_key(layout_ssec_key())
+                .sse_customer_key_md5(layout_ssec_key_md5())
+        } else {
+            request
+        };
+        request.send().await?;
+        return Ok(());
+    }
+    let create = client
+        .create_multipart_upload()
+        .bucket(case.bucket)
+        .key(case.key)
+        .content_type(content_type);
+    let create = if case.ssec {
+        create
+            .sse_customer_algorithm("AES256")
+            .sse_customer_key(layout_ssec_key())
+            .sse_customer_key_md5(layout_ssec_key_md5())
+    } else {
+        create
+    };
+    let created = create.send().await?;
+    let upload_id = created.upload_id().ok_or("CreateMultipartUpload omitted upload ID")?;
+    let mut completed = Vec::with_capacity(case.part_sizes.len());
+    let mut offset = 0usize;
+    for (index, size) in case.part_sizes.iter().enumerate() {
+        let part_number = i32::try_from(index + 1)?;
+        let chunk = case.body[offset..offset + size].to_vec();
+        offset += size;
+        let upload = client
+            .upload_part()
+            .bucket(case.bucket)
+            .key(case.key)
+            .upload_id(upload_id)
+            .part_number(part_number)
+            .body(ByteStream::from(chunk));
+        let upload = if case.ssec {
+            upload
+                .sse_customer_algorithm("AES256")
+                .sse_customer_key(layout_ssec_key())
+                .sse_customer_key_md5(layout_ssec_key_md5())
+        } else {
+            upload
+        };
+        let uploaded = upload.send().await?;
+        completed.push(
+            CompletedPart::builder()
+                .part_number(part_number)
+                .e_tag(uploaded.e_tag().ok_or("UploadPart omitted ETag")?)
+                .build(),
+        );
+    }
+    client
+        .complete_multipart_upload()
+        .bucket(case.bucket)
+        .key(case.key)
+        .upload_id(upload_id)
+        .multipart_upload(CompletedMultipartUpload::builder().set_parts(Some(completed)).build())
+        .send()
+        .await?;
+    Ok(())
+}
+
+fn layout_cases() -> Vec<LayoutCase> {
+    let two = vec![LAYOUT_PART_SIZE, LAYOUT_TAIL_SIZE];
+    let three = vec![LAYOUT_PART_SIZE, LAYOUT_PART_SIZE, 4096];
+    let total = |sizes: &[usize]| sizes.iter().sum::<usize>();
+    let case = |bucket, key, part_sizes: Vec<usize>, body: Vec<u8>, ssec| LayoutCase {
+        bucket,
+        key,
+        part_sizes,
+        body,
+        ssec,
+        assert_replication: true,
+        rc5_etag: String::new(),
+        rc5_reported_parts: None,
+    };
+    vec![
+        case(LAYOUT_PLAIN_BUCKET, "plain/single.bin", vec![], layout_noise(1024 * 1024 + 17, 1), false),
+        case(
+            LAYOUT_PLAIN_BUCKET,
+            "plain/multipart-2.bin",
+            two.clone(),
+            layout_noise(total(&two), 2),
+            false,
+        ),
+        case(
+            LAYOUT_PLAIN_BUCKET,
+            "plain/multipart-3.bin",
+            three.clone(),
+            layout_noise(total(&three), 3),
+            false,
+        ),
+        case(
+            LAYOUT_PLAIN_BUCKET,
+            "plain/compressed-single.txt",
+            vec![],
+            layout_text(1024 * 1024 + 17, 4),
+            false,
+        ),
+        case(
+            LAYOUT_PLAIN_BUCKET,
+            "plain/compressed-multipart-2.txt",
+            two.clone(),
+            layout_text(total(&two), 5),
+            false,
+        ),
+        case(
+            LAYOUT_PLAIN_BUCKET,
+            "plain/ssec-multipart-2.bin",
+            two.clone(),
+            layout_noise(total(&two), 6),
+            true,
+        ),
+        // SSE-C passthrough replicates the stored ciphertext part by part; a
+        // compressible first part is stored well below 5 MiB and a standard
+        // target rejects it with EntityTooSmall. rc.5 fails the same way (see
+        // `rc5_baseline_replicates_multipart_layouts`), so the outcome is
+        // recorded rather than asserted here; tracked as rustfs/backlog#2363.
+        LayoutCase {
+            assert_replication: false,
+            ..case(
+                LAYOUT_PLAIN_BUCKET,
+                "plain/ssec-compressed-multipart-2.txt",
+                two.clone(),
+                layout_text(total(&two), 7),
+                true,
+            )
+        },
+        case(
+            LAYOUT_ENCRYPTED_BUCKET,
+            "encrypted/single.bin",
+            vec![],
+            layout_noise(1024 * 1024 + 17, 8),
+            false,
+        ),
+        case(
+            LAYOUT_ENCRYPTED_BUCKET,
+            "encrypted/multipart-2.bin",
+            two.clone(),
+            layout_noise(total(&two), 9),
+            false,
+        ),
+        case(
+            LAYOUT_ENCRYPTED_BUCKET,
+            "encrypted/multipart-3.bin",
+            three.clone(),
+            layout_noise(total(&three), 10),
+            false,
+        ),
+        case(
+            LAYOUT_ENCRYPTED_BUCKET,
+            "encrypted/compressed-multipart-2.txt",
+            two.clone(),
+            layout_text(total(&two), 11),
+            false,
+        ),
+    ]
+}
+
+fn layout_server_env() -> Vec<(&'static str, &'static str)> {
+    let mut env = bucket_config_server_env();
+    env.push(("RUSTFS_COMPRESSION_ENABLED", "true"));
+    env.push(("RUSTFS_COMPRESSION_MULTIPART_ENABLED", "true"));
+    env
+}
+
+fn layout_reported_parts(attributes: &aws_sdk_s3::operation::get_object_attributes::GetObjectAttributesOutput) -> Option<usize> {
+    attributes.object_parts().map(|parts| parts.parts().len())
+}
+
+async fn assert_layout_readable(client: &Client, case: &LayoutCase, context: &str) -> TestResult {
+    let label = case.label();
+    let head = layout_head(client, case).await?;
+    assert_eq!(
+        head.e_tag().map(|etag| etag.trim_matches('"')),
+        Some(case.rc5_etag.as_str()),
+        "{context}: {label}: the ETag written by rc.5 must be reported unchanged"
+    );
+    assert_eq!(
+        head.content_length(),
+        Some(i64::try_from(case.body.len())?),
+        "{context}: {label}: HEAD content length"
+    );
+
+    let full = layout_get(client, case, None, None).await?.body.collect().await?.into_bytes();
+    assert_eq!(full.len(), case.body.len(), "{context}: {label}: full GET length");
+    assert!(full == case.body, "{context}: {label}: full GET body must equal the rc.5 upload");
+
+    if case.is_multipart_layout() {
+        let first = case.part_sizes[0];
+        let range = format!("bytes={}-{}", first - 32, first + 31);
+        let crossing = layout_get(client, case, Some(range), None)
+            .await?
+            .body
+            .collect()
+            .await?
+            .into_bytes();
+        assert!(
+            crossing == case.body[first - 32..first + 32],
+            "{context}: {label}: range across the first part boundary"
+        );
+        let tail_start: usize = case.part_sizes[..case.part_sizes.len() - 1].iter().sum();
+        let last_number = i32::try_from(case.part_sizes.len())?;
+        let last = layout_get(client, case, None, Some(last_number)).await?;
+        assert_eq!(
+            last.content_length(),
+            Some(i64::try_from(case.part_sizes[case.part_sizes.len() - 1])?),
+            "{context}: {label}: partNumber={last_number} length"
+        );
+        let last_body = last.body.collect().await?.into_bytes();
+        assert!(
+            last_body == case.body[tail_start..],
+            "{context}: {label}: partNumber={last_number} body must be the stored last part"
+        );
+    }
+    Ok(())
+}
+
+async fn assert_layout_attributes(client: &Client, case: &LayoutCase, context: &str) -> TestResult {
+    let label = case.label();
+    let attributes = layout_attributes(client, case).await?;
+    assert_eq!(
+        attributes.e_tag().map(|etag| etag.trim_matches('"')),
+        Some(case.rc5_etag.as_str()),
+        "{context}: {label}: attributes ETag"
+    );
+    assert_eq!(
+        attributes.object_size(),
+        Some(i64::try_from(case.body.len())?),
+        "{context}: {label}: attributes ObjectSize"
+    );
+    if case.is_multipart_layout() {
+        let parts = attributes
+            .object_parts()
+            .ok_or_else(|| format!("{context}: {label}: multipart layout must expose ObjectParts"))?;
+        assert_eq!(
+            parts.total_parts_count(),
+            Some(i32::try_from(case.part_sizes.len())?),
+            "{context}: {label}: TotalPartsCount"
+        );
+        let observed: Vec<(Option<i32>, Option<i64>)> =
+            parts.parts().iter().map(|part| (part.part_number(), part.size())).collect();
+        let expected: Vec<(Option<i32>, Option<i64>)> = case
+            .part_sizes
+            .iter()
+            .enumerate()
+            .map(|(index, size)| (Some(index as i32 + 1), Some(*size as i64)))
+            .collect();
+        assert_eq!(
+            observed, expected,
+            "{context}: {label}: ObjectParts must report the plaintext part layout"
+        );
+    } else {
+        assert!(
+            attributes.object_parts().is_none_or(|parts| parts.parts().is_empty()),
+            "{context}: {label}: a single PUT must not report stored parts"
+        );
+    }
+    Ok(())
+}
+
+async fn put_layout_replication_rule(env: &RustFSTestEnvironment, bucket: &str, arn: &str) -> TestResult {
+    let body = format!(
+        r#"<ReplicationConfiguration xmlns="http://s3.amazonaws.com/doc/2006-03-01/">
+  <Role></Role>
+  <Rule>
+    <ID>legacy-layouts</ID>
+    <Priority>1</Priority>
+    <Status>Enabled</Status>
+    <Filter><Prefix></Prefix></Filter>
+    <DeleteMarkerReplication><Status>Enabled</Status></DeleteMarkerReplication>
+    <DeleteReplication><Status>Enabled</Status></DeleteReplication>
+    <ExistingObjectReplication><Status>Enabled</Status></ExistingObjectReplication>
+    <Destination><Bucket>{arn}</Bucket></Destination>
+  </Rule>
+</ReplicationConfiguration>"#
+    );
+    let url = format!("{}/{bucket}?replication", env.url);
+    let response = signed_request(
+        Method::PUT,
+        &url,
+        &env.access_key,
+        &env.secret_key,
+        Some(body.into_bytes()),
+        Some("application/xml"),
+    )
+    .await?;
+    if response.status() != StatusCode::OK {
+        let status = response.status();
+        let body = response.text().await.unwrap_or_default();
+        return Err(format!("put replication rule on {bucket} failed: {status} {body}").into());
+    }
+    Ok(())
+}
+
+/// Wait for the existing-object replication of `case` to reach a terminal
+/// status and return it (`COMPLETED` or `FAILED`).
+async fn wait_layout_replication_terminal(client: &Client, case: &LayoutCase) -> Result<String, BoxError> {
+    let deadline = Instant::now() + LAYOUT_REPLICATION_TIMEOUT;
+    loop {
+        let head = layout_head(client, case).await?;
+        let status = head.replication_status().map(|status| status.as_str().to_string());
+        if matches!(status.as_deref(), Some("COMPLETED") | Some("FAILED")) {
+            return Ok(status.unwrap_or_default());
+        }
+        if Instant::now() >= deadline {
+            return Err(format!(
+                "{}: existing-object replication never reached a terminal status; last {status:?}",
+                case.label()
+            )
+            .into());
+        }
+        sleep(Duration::from_millis(500)).await;
+    }
+}
+
+#[derive(Debug)]
+struct LayoutTransport {
+    status: String,
+    uploaded_parts: Vec<i32>,
+    single_puts: usize,
+    completes: usize,
+    /// Raw per-key journal in target order: (sequence, operation, part number,
+    /// upload id), so duplicate drives can be told apart from retries.
+    journal: Vec<(u64, String, Option<i32>, Option<String>)>,
+}
+
+/// Configure every layout bucket to replicate its existing objects to a fresh
+/// fake target, wait for each case to settle, and report the transport the
+/// target observed per case.
+async fn replicate_layouts(
+    env: &RustFSTestEnvironment,
+    client: &Client,
+    cases: &[LayoutCase],
+) -> Result<(FakeS3Target, Vec<LayoutTransport>), BoxError> {
+    let target = FakeS3Target::start().await?;
+    target.create_bucket(LAYOUT_REPLICA_BUCKET);
+    for bucket in [LAYOUT_PLAIN_BUCKET, LAYOUT_ENCRYPTED_BUCKET] {
+        let arn = set_replication_target_with_options(
+            env,
+            bucket,
+            ReplicationTargetOptions {
+                endpoint: &target.address(),
+                access_key: FAKE_ACCESS_KEY,
+                secret_key: FAKE_SECRET_KEY,
+                target_bucket: LAYOUT_REPLICA_BUCKET,
+                secure: false,
+                skip_tls_verify: false,
+                ca_cert_pem: None,
+            },
+        )
+        .await?;
+        put_layout_replication_rule(env, bucket, &arn).await?;
+    }
+    let mut statuses = Vec::with_capacity(cases.len());
+    for case in cases {
+        statuses.push(wait_layout_replication_terminal(client, case).await?);
+    }
+    let journal = target.requests();
+    let mut transports = Vec::with_capacity(cases.len());
+    for (case, status) in cases.iter().zip(statuses) {
+        let key_requests: Vec<_> = journal
+            .iter()
+            .filter(|record| record.key.as_deref() == Some(case.key))
+            .collect();
+        let mut uploaded_parts: Vec<i32> = key_requests
+            .iter()
+            .filter(|record| record.operation == FakeTargetOperation::UploadPart)
+            .filter_map(|record| record.part_number)
+            .collect();
+        uploaded_parts.sort_unstable();
+        uploaded_parts.dedup();
+        let transport = LayoutTransport {
+            status,
+            uploaded_parts,
+            single_puts: key_requests
+                .iter()
+                .filter(|record| record.operation == FakeTargetOperation::PutObject)
+                .count(),
+            completes: key_requests
+                .iter()
+                .filter(|record| record.operation == FakeTargetOperation::CompleteMultipartUpload)
+                .count(),
+            journal: key_requests
+                .iter()
+                .map(|record| {
+                    (
+                        record.sequence,
+                        format!("{:?}", record.operation),
+                        record.part_number,
+                        record.upload_id.as_ref().map(|id| id.chars().take(12).collect()),
+                    )
+                })
+                .collect(),
+        };
+        tracing::info!(
+            target: "e2e_test::upgrade_compatibility_test",
+            object = %case.label(),
+            ?transport,
+            "replication transport observed on the target"
+        );
+        transports.push(transport);
+    }
+    Ok((target, transports))
+}
+
+/// rc.5 writes single-PUT, multipart, compressed, SSE-C and SSE-S3 layouts;
+/// the current build must read every byte, expose the stored part layout
+/// through GetObjectAttributes and partNumber reads, and replicate the objects
+/// with the transport that matches their stored parts.
+#[tokio::test]
+#[ignore = "requires the pinned 1.0.0-rc.5 release binary"]
+async fn direct_upgrade_from_rc5_preserves_multipart_layouts() -> TestResult {
+    init_logging();
+    let previous_binary = source_binary()?;
+    let server_env = layout_server_env();
+
+    let mut env = RustFSTestEnvironment::new().await?;
+    env.start_rustfs_server_from_binary(&previous_binary, vec![], &server_env)
+        .await?;
+    let old_client = env.create_s3_client();
+    env.create_test_bucket(LAYOUT_PLAIN_BUCKET).await?;
+    env.create_test_bucket(LAYOUT_ENCRYPTED_BUCKET).await?;
+    enable_versioning(&old_client, LAYOUT_PLAIN_BUCKET).await?;
+    enable_versioning(&old_client, LAYOUT_ENCRYPTED_BUCKET).await?;
+    put_default_sse_s3_encryption(&old_client, LAYOUT_ENCRYPTED_BUCKET).await?;
+    assert_default_sse_s3_encryption(&old_client, LAYOUT_ENCRYPTED_BUCKET, "rc.5").await?;
+
+    let mut cases = layout_cases();
+    for case in cases.iter_mut() {
+        layout_write(&old_client, case).await?;
+        let head = layout_head(&old_client, case).await?;
+        case.rc5_etag = head
+            .e_tag()
+            .ok_or_else(|| format!("{}: rc.5 HEAD omitted the ETag", case.label()))?
+            .trim_matches('"')
+            .to_string();
+        case.rc5_reported_parts = layout_attributes(&old_client, case)
+            .await
+            .ok()
+            .and_then(|a| layout_reported_parts(&a));
+        tracing::info!(
+            target: "e2e_test::upgrade_compatibility_test",
+            object = %case.label(),
+            parts = case.part_sizes.len(),
+            etag = %case.rc5_etag,
+            rc5_reported_parts = ?case.rc5_reported_parts,
+            "rc.5 wrote a legacy layout"
+        );
+    }
+    // The rc.5 writer must itself still read what it wrote, so a later
+    // failure is attributable to the upgrade rather than to the fixture.
+    for case in &cases {
+        assert_layout_readable(&old_client, case, "rc.5").await?;
+    }
+
+    // Upgrade in place.
+    env.restart_server_preserving_data(vec![], &server_env).await?;
+    let client = env.create_s3_client();
+    for case in &cases {
+        assert_layout_readable(&client, case, "upgraded").await?;
+        assert_layout_attributes(&client, case, "upgraded").await?;
+    }
+
+    // Replicate the pre-existing objects with the current build.
+    let (target, transports) = replicate_layouts(&env, &client, &cases).await?;
+    let replica_client = fake_source_client(&target);
+    for (case, transport) in cases.iter().zip(&transports) {
+        let label = case.label();
+        if !case.assert_replication {
+            continue;
+        }
+        assert_eq!(transport.status, "COMPLETED", "{label}: existing-object replication must complete");
+        if case.is_multipart_layout() {
+            let expected: Vec<i32> = (1..=i32::try_from(case.part_sizes.len())?).collect();
+            assert_eq!(
+                transport.uploaded_parts, expected,
+                "{label}: stored parts must replicate as the same multipart layout"
+            );
+            // The current build can drive an existing object twice (two
+            // full CreateMultipartUpload/UploadPart/Complete rounds with
+            // distinct upload ids) while its status is still PENDING; the
+            // rc.5 baseline drives once. That is a scheduling difference,
+            // not a layout one, tracked as rustfs/backlog#2362.
+            assert!(transport.completes >= 1, "{label}: at least one CompleteMultipartUpload");
+            if transport.completes > 1 {
+                tracing::warn!(
+                    target: "e2e_test::upgrade_compatibility_test",
+                    object = %label,
+                    completes = transport.completes,
+                    journal = ?transport.journal,
+                    "existing-object replication drove the same object more than once (rustfs/backlog#2362)"
+                );
+            }
+            assert_eq!(
+                transport.single_puts, 0,
+                "{label}: a multipart layout must not go out as a single PutObject"
+            );
+        } else {
+            assert!(transport.single_puts >= 1, "{label}: a single PUT replicates as PutObject");
+            assert!(transport.uploaded_parts.is_empty(), "{label}: a single PUT must not go out as multipart");
+        }
+        if !case.ssec {
+            let replica = replica_client
+                .get_object()
+                .bucket(LAYOUT_REPLICA_BUCKET)
+                .key(case.key)
+                .send()
+                .await
+                .map_err(|err| format!("{label}: replica missing on the target: {err}"))?
+                .body
+                .collect()
+                .await?
+                .into_bytes();
+            assert_eq!(replica.len(), case.body.len(), "{label}: replica length");
+            assert!(replica == case.body, "{label}: replica body must equal the rc.5 upload");
+        }
+    }
+    Ok(())
+}
+
+/// The same layouts replicated by rc.5 itself, without an upgrade. This is the
+/// baseline that tells a pre-existing transport failure apart from one the
+/// current build introduced; it records the outcome per layout and only fails
+/// when the fixture cannot run.
+#[tokio::test]
+#[ignore = "requires the pinned 1.0.0-rc.5 release binary"]
+async fn rc5_baseline_replicates_multipart_layouts() -> TestResult {
+    init_logging();
+    let previous_binary = source_binary()?;
+    let server_env = layout_server_env();
+
+    let mut env = RustFSTestEnvironment::new().await?;
+    env.start_rustfs_server_from_binary(&previous_binary, vec![], &server_env)
+        .await?;
+    let client = env.create_s3_client();
+    env.create_test_bucket(LAYOUT_PLAIN_BUCKET).await?;
+    env.create_test_bucket(LAYOUT_ENCRYPTED_BUCKET).await?;
+    enable_versioning(&client, LAYOUT_PLAIN_BUCKET).await?;
+    enable_versioning(&client, LAYOUT_ENCRYPTED_BUCKET).await?;
+    put_default_sse_s3_encryption(&client, LAYOUT_ENCRYPTED_BUCKET).await?;
+
+    let mut cases = layout_cases();
+    for case in cases.iter_mut() {
+        layout_write(&client, case).await?;
+        let head = layout_head(&client, case).await?;
+        case.rc5_etag = head
+            .e_tag()
+            .ok_or_else(|| format!("{}: rc.5 HEAD omitted the ETag", case.label()))?
+            .trim_matches('"')
+            .to_string();
+    }
+    let (_target, transports) = replicate_layouts(&env, &client, &cases).await?;
+    let summary: Vec<String> = cases
+        .iter()
+        .zip(&transports)
+        .map(|(case, transport)| {
+            format!(
+                "{}: {} parts={:?} puts={}",
+                case.label(),
+                transport.status,
+                transport.uploaded_parts,
+                transport.single_puts
+            )
+        })
+        .collect();
+    tracing::info!(target: "e2e_test::upgrade_compatibility_test", ?summary, "rc.5 baseline replication outcomes");
+    Ok(())
+}
@@ -118,8 +118,6 @@ hotpath-cpu = [
 # injection, xl.meta transition assertions) via `api::tier::test_util`.
 # Enable only from `[dev-dependencies]` (rustfs/backlog#1148 ilm-6).
 test-util = []
-# Observes real startup CAS only in the dedicated E2E binary.
-e2e-test-hooks = []

 [dependencies]
 hotpath.workspace = true
@@ -228,7 +226,7 @@ metrics = { workspace = true }
 # crates.io. The guard scripts/check_no_tokio_io_uring.sh allows an explicit
 # io-uring integration; only the tokio "io-uring" runtime feature is banned.
 [target.'cfg(target_os = "linux")'.dependencies]
-rustfs-uring = "0.2.1"
+rustfs-uring = "0.2.2"

 [target.'cfg(windows)'.dependencies]
 winapi-util.workspace = true
@@ -32,7 +32,7 @@ pub mod bucket {
    pub mod bucket_target_sys {
        pub use crate::bucket::bucket_target_sys::{
            AdvancedPutOptions, BucketTargetError, BucketTargetSys, PutObjectOptions, RemoveObjectOptions, S3ClientError,
-            SsecPassthroughCapability, TargetClient, append_version_id_query,
+            SsecPassthroughCapability, TargetClient, VersionIdentityCapability, append_version_id_query,
        };
    }

@@ -76,11 +76,26 @@ pub mod bucket {
            };
        }

+        pub mod recovery_disposition {
+            pub use crate::bucket::lifecycle::recovery_disposition::{
+                IlmRecoveryDispositionExecutionOutcome, IlmRecoveryDispositionReasonCode, IlmRecoveryDispositionState,
+                dry_run_recovery_disposition, execute_recovery_disposition,
+            };
+        }
+
+        pub mod recovery_export {
+            pub use crate::bucket::lifecycle::recovery_export::{
+                IlmRecoveryExportCreated, IlmRecoveryExportObservation, create_recovery_export,
+                inspect_recovery_export_observation, load_recovery_export,
+            };
+        }
+
        pub mod transition_transaction {
            pub use crate::bucket::lifecycle::transition_transaction::{
                TransitionOperatorDeleteResult, TransitionOperatorError, TransitionOperatorProbe, TransitionOperatorStatus,
-                delete_transition_candidate_for_operator, finalize_missing_transition_transaction_for_operator,
-                inspect_transition_transaction_for_operator,
+                TransitionRecoveryRetryResult, TransitionRecoveryRetryStatus, delete_transition_candidate_for_operator,
+                finalize_missing_transition_transaction_for_operator, inspect_transition_recovery_retry_for_operator,
+                inspect_transition_transaction_for_operator, retry_transition_recovery_for_operator,
            };
            #[cfg(feature = "test-util")]
            pub use crate::bucket::lifecycle::transition_transaction::{
@@ -293,7 +308,7 @@ pub mod cache {
 pub mod capacity {
    pub use crate::core::pools::{
        DecommissionUnresolvedEntry, PoolDecommissionInfo, PoolStatus, get_total_usable_capacity, get_total_usable_capacity_free,
-        path2_bucket_object, path2_bucket_object_with_base_path,
+        is_pool_activation_fleet_proof_error, path2_bucket_object, path2_bucket_object_with_base_path,
    };
    pub use crate::store::utils::is_reserved_or_invalid_bucket;
 }
@@ -368,8 +383,6 @@ pub mod data_usage {
 pub mod disk {
    pub use crate::disk::disk_store::get_object_disk_read_timeout;
    pub use crate::disk::local::ScanGuard;
-    #[cfg(all(feature = "test-util", not(windows)))]
-    pub use crate::disk::os::{LocalPublicationPause, LocalPublicationStage};
    pub use crate::disk::{
        BATCH_READ_VERSION_MAX_ITEMS, BUCKET_META_PREFIX, BatchReadVersionItem, BatchReadVersionReq, BatchReadVersionResp,
        CheckPartsResp, ConditionalFileUpdate, DeleteOptions, Disk, DiskAPI, DiskInfo, DiskInfoOptions, DiskLocation, DiskOption,
@@ -448,9 +461,12 @@ pub mod notification {
    #[cfg(any(test, feature = "test-util"))]
    pub use crate::services::notification_sys::rotate_cross_pool_fence_fleet_proof_for_test;
    pub use crate::services::notification_sys::{
-        ClusterTierDailyStats, CrossPoolFenceFleetProofToken, LegacyTransitionStateReconcileFleetProofToken, NotificationPeerErr,
-        NotificationSys, ScannerPublicationLeaseGrant, acquire_cross_pool_fence_fleet_proof,
+        ClusterTierDailyStats, CrossPoolFenceFleetProofToken, IlmRecoveryExportFleetProofToken,
+        LegacyTransitionStateReconcileFleetProofToken, NotificationPeerErr, NotificationSys, ScannerPublicationLeaseGrant,
+        acquire_cross_pool_fence_fleet_proof, acquire_ilm_recovery_export_fleet_proof,
        acquire_legacy_transition_state_reconcile_fleet_proof, cross_pool_fence_fleet_proof_matches, get_global_notification_sys,
+        ilm_recovery_export_fleet_proof_matches, ilm_recovery_export_local_process_epoch,
+        ilm_recovery_export_member_epochs_sha256, ilm_recovery_export_topology_generation,
        legacy_transition_state_reconcile_fleet_proof_matches, new_global_notification_sys,
        scanner_peer_transport_error_message_is_retryable, start_remote_version_state_fleet_probe,
    };
@@ -503,7 +519,8 @@ pub mod rpc {
    pub use crate::cluster::rpc::{
        AuthenticatedChannel, KMS_SIGNAL_SUBSYSTEM, LocalPeerS3Client, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS,
        PeerRestClient, PeerS3Client, S3PeerSys, SERVICE_SIGNAL_REFRESH_CONFIG, SERVICE_SIGNAL_RELOAD_DYNAMIC,
-        ScannerBucketListing, ScannerPeerActivity, ScannerPeerDirtyUsageSnapshot, ScannerPublicationLease, TONIC_RPC_PREFIX,
+        ScannerBucketListing, ScannerDirtyUsageAcknowledgement, ScannerPeerActivity, ScannerPeerDirtyUsageBucket,
+        ScannerPeerDirtyUsageSnapshot, ScannerPublicationLease, ScannerScopedDirtyUsageAckEntry, TONIC_RPC_PREFIX,
        TonicInterceptor, build_put_file_auth_trailer, check_and_record_signed_rpc_nonce, decode_heal_bucket_rpc_options,
        encode_heal_bucket_rpc_options, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
        gen_tonic_signature_interceptor, node_service_time_out_client, node_service_time_out_client_no_auth,
@@ -546,8 +563,8 @@ pub mod storage {
    pub use crate::core::pools::HealLifecycleExpiryContext;
    pub use crate::store::HealWalkVersion;
    pub use crate::store::{
-        BootstrapLocalTarget, ECStore, SCANNER_PUBLICATION_LEASE_TTL_MS, ScannerDataMovementPauseStatus, all_local_disk,
-        all_local_disk_path, find_local_disk_by_ref, init_local_disks, init_local_disks_with_instance_ctx, init_lock_clients,
+        ECStore, SCANNER_PUBLICATION_LEASE_TTL_MS, ScannerDataMovementPauseStatus, all_local_disk, all_local_disk_path,
+        find_local_disk_by_ref, init_local_disks, init_local_disks_with_instance_ctx, init_lock_clients,
        prewarm_local_disk_id_map, prewarm_local_disk_id_map_with_instance_ctx,
    };
 }
@@ -18,7 +18,7 @@ use crate::bucket::metadata_sys::get_replication_config;
 use crate::bucket::remote_s3_client::{
    PathStyle, REPLICATION_TARGET_RETRY_POLICY, RemoteCredentials, RemoteS3EndpointSpec, build_remote_s3_client,
 };
-use crate::bucket::replication::{ObjectLockIntegrity, object_lock_put_integrity};
+use crate::bucket::replication::{ObjectLockIntegrity, object_lock_put_integrity, replication_etags_match};
 use crate::bucket::replication::{ReplicationStatusType, ReplicationTargetConfigBridge};
 use crate::bucket::target::ARN;
 use crate::bucket::target::BucketTargetType;
@@ -126,6 +126,22 @@ impl From<&BucketTarget> for RemoteS3EndpointSpec {
 }

 pub type HeadObjectSdkError = Box<SdkError<HeadObjectError>>;
+
+/// Whether an edited bucket target still addresses the same remote service
+/// (endpoint, bucket, path style, TLS and identity), so a verdict learned
+/// about that service stays valid across the edit.
+fn same_replication_service(edited: &BucketTarget, previous: &BucketTarget) -> bool {
+    let access_key = |target: &BucketTarget| target.credentials.as_ref().map(|credentials| credentials.access_key.clone());
+    edited.endpoint == previous.endpoint
+        && edited.target_bucket == previous.target_bucket
+        && edited.secure == previous.secure
+        && edited.path == previous.path
+        && access_key(edited) == access_key(previous)
+}
+
+/// Page size and page budget for [`TargetClient::find_version_by_etag`].
+const FIND_VERSION_BY_ETAG_PAGE_SIZE: i32 = 1000;
+const FIND_VERSION_BY_ETAG_MAX_PAGES: usize = 8;
 pub type GetObjectSdkError = Box<SdkError<GetObjectError>>;
 pub type GetObjectTaggingSdkError = Box<SdkError<GetObjectTaggingError>>;
 pub type PutObjectTaggingSdkError = Box<SdkError<PutObjectTaggingError>>;
@@ -349,6 +365,13 @@ struct TargetClientBuildProbe {
 /// their import path while the verdict vocabulary lives with the
 /// replication decision logic.
 pub use crate::bucket::replication::SsecPassthroughCapability;
+/// Version-identity verdicts (see the enum's own docs in
+/// `rustfs-replication`) are cached here per target ARN and follow the same
+/// `arn_remotes_map` lifecycle. They carry no TTL: the verdict is refreshed
+/// by every replication write's response, so it can only go stale on a
+/// target that receives no writes — and a stale `MintsOwn` costs one extra
+/// content-identity lookup before a PUT, never a lost replica.
+pub use crate::bucket::replication::VersionIdentityCapability;

 /// How long an audited SSE-C passthrough verdict stays authoritative.
 ///
@@ -375,6 +398,11 @@ pub struct BucketTargetSys {
    /// SSE-C passthrough capability verdicts keyed by target ARN. See
    /// [`SsecPassthroughCapability`]; reset alongside `arn_remotes_map`.
    ssec_passthrough_map: Arc<RwLock<HashMap<String, SsecPassthroughRecord>>>,
+    /// Version-identity verdicts keyed by target ARN. See
+    /// [`VersionIdentityCapability`]; reset alongside `arn_remotes_map`. A std
+    /// lock (never held across an await) so the replication worker can record
+    /// a verdict from inside its synchronous PUT-response audit.
+    version_identity_map: Arc<std::sync::RwLock<HashMap<String, VersionIdentityCapability>>>,
    pub targets_map: Arc<RwLock<HashMap<String, Vec<BucketTarget>>>>,
    /// Buckets whose persisted `bucket-targets.json` exists but cannot be
    /// decoded (rustfs/backlog#2282). Written under the bucket's update mutex
@@ -423,6 +451,7 @@ impl BucketTargetSys {
        Self {
            arn_remotes_map: Arc::new(RwLock::new(HashMap::new())),
            ssec_passthrough_map: Arc::new(RwLock::new(HashMap::new())),
+            version_identity_map: Arc::new(std::sync::RwLock::new(HashMap::new())),
            targets_map: Arc::new(RwLock::new(HashMap::new())),
            unreadable_targets: Arc::new(RwLock::new(HashSet::new())),
            h_mutex: Arc::new(RwLock::new(HashMap::new())),
@@ -746,10 +775,40 @@ impl BucketTargetSys {
                arn_remotes_map.remove(&target.arn);
                health_map.remove(&target.arn);
                ssec_map.remove(&target.arn);
+                self.forget_version_identity_capability(&target.arn);
            }
        }
    }

+    /// Cached version-identity verdict for a target ARN; `Unknown` until a
+    /// replication write or a replication-check VersionFidelity probe judged
+    /// it since the target was built.
+    pub fn version_identity_capability(&self, arn: &str) -> VersionIdentityCapability {
+        self.version_identity_map
+            .read()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .get(arn)
+            .copied()
+            .unwrap_or_default()
+    }
+
+    /// Record a version-identity verdict for a target ARN. Written by the
+    /// replication worker after every PutObject / CompleteMultipartUpload
+    /// response and by the replication-check VersionFidelity phase.
+    pub fn record_version_identity_capability(&self, arn: &str, capability: VersionIdentityCapability) {
+        self.version_identity_map
+            .write()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .insert(arn.to_string(), capability);
+    }
+
+    fn forget_version_identity_capability(&self, arn: &str) {
+        self.version_identity_map
+            .write()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .remove(arn);
+    }
+
    /// Cached SSE-C passthrough capability for a target ARN, plus whether the
    /// verdict is older than [`SSEC_PASSTHROUGH_CAPABILITY_TTL`]. `(Unknown,
    /// false)` when no verdict has been recorded since the target was built.
@@ -1162,12 +1221,32 @@ impl BucketTargetSys {
        // Remove existing targets
        if let Some(existing_targets) = targets_map.remove(bucket) {
            let mut ssec_map = self.ssec_passthrough_map.write().await;
+            let unchanged_service: HashMap<&str, &BucketTarget> = targets
+                .map(|new_targets| {
+                    new_targets
+                        .targets
+                        .iter()
+                        .map(|target| (target.arn.as_str(), target))
+                        .collect()
+                })
+                .unwrap_or_default();
            for target in existing_targets {
                arn_remotes_map.remove(&target.arn);
                health_map.remove(&target.arn);
                // A rebuilt/edited target may point at a different service:
                // the SSE-C passthrough verdict must be re-audited from Unknown.
                ssec_map.remove(&target.arn);
+                // The version-identity verdict survives an edit that keeps the
+                // same remote service (a resync start or a bandwidth change
+                // rewrites the entry in place): forgetting it there would make
+                // the very resync that follows re-drive every object as a
+                // duplicate on a target that mints its own version ids.
+                if unchanged_service
+                    .get(target.arn.as_str())
+                    .is_none_or(|edited| !same_replication_service(edited, &target))
+                {
+                    self.forget_version_identity_capability(&target.arn);
+                }
                self.update_bandwidth_limit(bucket, &target.arn, 0);
            }
        }
@@ -1892,6 +1971,62 @@ impl TargetClient {
            .map_err(Box::new)
    }

+    /// Locate a replica by content identity on a target that mints its own
+    /// version ids: page `ListObjectVersions` under the exact key and return
+    /// the newest live version whose ETag matches `source_etag`. Delete
+    /// markers and prefix siblings never match. Bounded to
+    /// [`FIND_VERSION_BY_ETAG_MAX_PAGES`] pages so a key with a very deep
+    /// history cannot turn one convergence check into an unbounded scan; a
+    /// replica beyond that window reads as missing, which only costs a
+    /// re-PUT (today's behaviour), never a lost object.
+    pub async fn find_version_by_etag(
+        &self,
+        bucket: &str,
+        object: &str,
+        source_etag: &str,
+    ) -> Result<Option<String>, Box<SdkError<aws_sdk_s3::operation::list_object_versions::ListObjectVersionsError>>> {
+        let mut key_marker: Option<String> = None;
+        let mut version_id_marker: Option<String> = None;
+        for _ in 0..FIND_VERSION_BY_ETAG_MAX_PAGES {
+            let page = self
+                .client
+                .list_object_versions()
+                .bucket(bucket)
+                .prefix(object)
+                .max_keys(FIND_VERSION_BY_ETAG_PAGE_SIZE)
+                .set_key_marker(key_marker.take())
+                .set_version_id_marker(version_id_marker.take())
+                .send()
+                .await
+                .map_err(Box::new)?;
+            if let Some(version) = page.versions().iter().find(|version| {
+                version.key() == Some(object)
+                    && version.version_id().is_some_and(|id| !id.is_empty())
+                    && replication_etags_match(Some(source_etag), version.e_tag())
+            }) {
+                return Ok(version.version_id().map(str::to_string));
+            }
+            // Every listed key is >= the prefix; once the listing moved past
+            // the exact key there is nothing left to find.
+            if page
+                .versions()
+                .iter()
+                .any(|version| version.key().is_some_and(|key| key > object))
+            {
+                return Ok(None);
+            }
+            if !page.is_truncated().unwrap_or(false) {
+                return Ok(None);
+            }
+            key_marker = page.next_key_marker().map(str::to_string);
+            version_id_marker = page.next_version_id_marker().map(str::to_string);
+            if key_marker.is_none() {
+                return Ok(None);
+            }
+        }
+        Ok(None)
+    }
+
    /// HEAD used by the read-proxy path (GET/HEAD of an object not yet
    /// replicated locally, MinIO `proxyHeadToRepTarget`).
    ///
@@ -2064,7 +2199,15 @@ impl TargetClient {
            }
        }

-        match builder
+        // A forwarded source checksum is this PUT's integrity header. In
+        // streaming-checksum mode (`RUSTFS_REPLICATION_STREAMING_CHECKSUMS`)
+        // the SDK would still add its default CRC32 trailer, and a target that
+        // receives both keeps the trailer's algorithm: a forwarded SHA256
+        // vanished from the replica while the source reported COMPLETED. Pin
+        // this request to WhenRequired so nothing is sent beside the source's
+        // own checksum.
+        let forwards_source_checksum = headers.keys().any(|name| name.as_str().starts_with("x-amz-checksum-"));
+        let mut operation = builder
            .bucket(bucket)
            .key(object)
            .content_length(size)
@@ -2084,10 +2227,14 @@ impl TargetClient {
                }

                Result::<_, aws_smithy_types::error::operation::BuildError>::Ok(req)
-            })
-            .send()
-            .await
-        {
+            });
+        if forwards_source_checksum {
+            operation = operation.config_override(
+                aws_sdk_s3::config::Builder::new()
+                    .request_checksum_calculation(aws_sdk_s3::config::RequestChecksumCalculation::WhenRequired),
+            );
+        }
+        match operation.send().await {
            Ok(output) => {
                // Under SSE-KMS/DSSE or SSE-C the target's ETag is not the MD5
                // of the stored plaintext, so it cannot be compared against the
@@ -2507,13 +2654,21 @@ mod tests {
    }

    fn header_recording_target_client(response_headers: Vec<(String, String)>) -> (TargetClient, RecordedHeaders) {
+        header_recording_target_client_with_checksums(response_headers, replication_request_checksum_calculation())
+    }
+
+    fn header_recording_target_client_with_checksums(
+        response_headers: Vec<(String, String)>,
+        checksums: RequestChecksumCalculation,
+    ) -> (TargetClient, RecordedHeaders) {
        let request_headers: RecordedHeaders = Arc::new(std::sync::Mutex::new(Vec::new()));
        let connector = SharedHttpConnector::new(RecordingHeaderConnector {
            request_headers: Arc::clone(&request_headers),
            response_headers,
        });
        let http_client = http_client_fn(move |_settings, _components| connector.clone());
-        let client = s3_client_for_test(443, Some(http_client));
+        let client =
+            s3_client_for_endpoint_test_with_checksums("https://localhost:443".to_string(), Some(http_client), checksums);
        (
            TargetClient {
                endpoint: "https://localhost:443".to_string(),
@@ -2680,6 +2835,47 @@ mod tests {
        }
    }

+    /// With streaming checksums enabled the SDK adds a CRC32 trailer to every
+    /// upload. A PUT that forwards the source's checksum must not get that
+    /// second algorithm: a target that receives both keeps the trailer's and
+    /// the forwarded SHA256 never reaches the replica (rustfs/backlog#2340).
+    #[tokio::test]
+    async fn streaming_put_object_with_forwarded_checksum_sends_no_sdk_checksum() {
+        let (client, recorded) =
+            header_recording_target_client_with_checksums(Vec::new(), RequestChecksumCalculation::WhenSupported);
+        let mut forwarded = PutObjectOptions::default();
+        forwarded.user_metadata.insert(
+            "x-amz-checksum-sha256".to_string(),
+            "OoJ3yNhRwv3wwtZoGqEIPrPX9xwTnfLl+ka0wStN1g0=".to_string(),
+        );
+        client
+            .put_object("target-bucket", "object", 4, streaming_test_body(b"data"), &forwarded)
+            .await
+            .expect("recorded put_object should succeed");
+        client
+            .put_object("target-bucket", "object", 4, streaming_test_body(b"data"), &PutObjectOptions::default())
+            .await
+            .expect("recorded put_object should succeed");
+        let recorded = recorded.lock().expect("recorded header lock should not be poisoned");
+        let with_forwarded = &recorded[0];
+        assert_eq!(
+            recorded_header(with_forwarded, "x-amz-checksum-sha256"),
+            Some("OoJ3yNhRwv3wwtZoGqEIPrPX9xwTnfLl+ka0wStN1g0=")
+        );
+        assert_eq!(
+            recorded_header(with_forwarded, "x-amz-trailer"),
+            None,
+            "the SDK must not add a trailer checksum"
+        );
+        assert_eq!(recorded_header(with_forwarded, "x-amz-sdk-checksum-algorithm"), None);
+        // Control: the same client still streams a trailer when nothing is forwarded.
+        let without_forwarded = &recorded[1];
+        assert!(
+            recorded_header(without_forwarded, "x-amz-trailer").is_some(),
+            "streaming mode must still apply to uploads without a forwarded checksum: {without_forwarded:?}"
+        );
+    }
+
    /// A forwarded source checksum already satisfies the rule; nothing is added.
    #[tokio::test]
    async fn locked_put_object_keeps_a_forwarded_source_checksum() {
@@ -3045,6 +3241,14 @@ mod tests {
    }

    fn s3_client_for_endpoint_test(endpoint: String, http_client: Option<SharedHttpClient>) -> S3Client {
+        s3_client_for_endpoint_test_with_checksums(endpoint, http_client, replication_request_checksum_calculation())
+    }
+
+    fn s3_client_for_endpoint_test_with_checksums(
+        endpoint: String,
+        http_client: Option<SharedHttpClient>,
+        checksums: RequestChecksumCalculation,
+    ) -> S3Client {
        let credentials = SdkCredentials::builder()
            .access_key_id("test-access")
            .secret_access_key("test-secret")
@@ -3058,7 +3262,7 @@ mod tests {
            .behavior_version(aws_sdk_s3::config::BehaviorVersion::latest())
            // Mirror the production remote-target builder so recorded requests
            // exercise the same checksum/framing behavior (#6853).
-            .request_checksum_calculation(replication_request_checksum_calculation());
+            .request_checksum_calculation(checksums);
        if let Some(http_client) = http_client {
            config = config.http_client(http_client);
        }
@@ -3221,6 +3425,64 @@ mod tests {
        assert!(message.contains("connection refused"));
    }

+    #[test]
+    fn same_replication_service_ignores_resync_and_bandwidth_edits() {
+        let base = BucketTarget {
+            endpoint: "target.example:9000".to_string(),
+            target_bucket: "replica".to_string(),
+            secure: true,
+            path: "on".to_string(),
+            arn: "arn:rustfs:replication:us-east-1:bucket:same".to_string(),
+            credentials: Some(Credentials {
+                access_key: "access".to_string(),
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let resync_edit = BucketTarget {
+            reset_id: "reset-1".to_string(),
+            bandwidth_limit: 1024,
+            ..base.clone()
+        };
+        assert!(same_replication_service(&resync_edit, &base));
+        for moved in [
+            BucketTarget {
+                endpoint: "other.example:9000".to_string(),
+                ..base.clone()
+            },
+            BucketTarget {
+                target_bucket: "other".to_string(),
+                ..base.clone()
+            },
+            BucketTarget {
+                secure: false,
+                ..base.clone()
+            },
+            BucketTarget {
+                credentials: Some(Credentials {
+                    access_key: "rotated".to_string(),
+                    ..Default::default()
+                }),
+                ..base.clone()
+            },
+        ] {
+            assert!(!same_replication_service(&moved, &base));
+        }
+    }
+
+    #[test]
+    fn version_identity_verdict_is_per_arn_and_forgotten_with_the_target() {
+        let sys = BucketTargetSys::default();
+        let arn = "arn:rustfs:replication:us-east-1:bucket:identity";
+        assert_eq!(sys.version_identity_capability(arn), VersionIdentityCapability::Unknown);
+        sys.record_version_identity_capability(arn, VersionIdentityCapability::MintsOwn);
+        assert_eq!(sys.version_identity_capability(arn), VersionIdentityCapability::MintsOwn);
+        assert_eq!(sys.version_identity_capability("other"), VersionIdentityCapability::Unknown);
+        // A rebuilt target may point at a different service.
+        sys.forget_version_identity_capability(arn);
+        assert_eq!(sys.version_identity_capability(arn), VersionIdentityCapability::Unknown);
+    }
+
    #[test]
    fn endpoint_health_key_preserves_explicit_port() {
        let url = Url::parse("https://remote.example:9443").expect("url should parse");
@@ -32,6 +32,7 @@ use crate::bucket::lifecycle::manual_transition_job::{
    record_manual_transition_worker_result_with_reason, renew_manual_transition_job_lease_if_owned,
    save_manual_transition_job_record_if_current, save_manual_transition_task_if_absent, update_manual_transition_job_record,
 };
+use crate::bucket::lifecycle::recovery_disposition_runtime::run_recovery_disposition_maintenance_loop;
 use crate::bucket::lifecycle::replication_sink;
 use crate::bucket::lifecycle::replication_sink::{
    DeleteReplicationConfigSnapshot, ReplicationObjectBridge, ReplicationStatusType, replication_state_to_filemeta,
@@ -149,6 +150,7 @@ pub type ExpiryOpType = Box<dyn ExpiryOp + Send + Sync + 'static>;
 static XXHASH_SEED: u64 = 0;
 static TIER_FREE_VERSION_RECOVERY_STARTED: OnceLock<()> = OnceLock::new();
 static MANUAL_TRANSITION_JOB_RECOVERY_STARTED: OnceLock<()> = OnceLock::new();
+static RECOVERY_DISPOSITION_MAINTENANCE_STARTED: OnceLock<()> = OnceLock::new();

 #[cfg(test)]
 #[derive(Default)]
@@ -2398,9 +2400,20 @@ pub async fn init_background_expiry(api: Arc<ECStore>) {
    let _ = spawn_tier_free_version_recovery_once(api.clone(), &TIER_FREE_VERSION_RECOVERY_STARTED);
    spawn_tier_delete_journal_recovery_once(api.clone());
    spawn_transition_transaction_recovery_once(api.clone());
+    spawn_recovery_disposition_maintenance_once(api.clone());
    spawn_manual_transition_job_recovery_once(api);
 }

+fn spawn_recovery_disposition_maintenance_once(api: Arc<ECStore>) -> Option<JoinHandle<()>> {
+    let cancel_token = api.ctx.background_cancel_token()?;
+    if RECOVERY_DISPOSITION_MAINTENANCE_STARTED.set(()).is_err() {
+        return None;
+    }
+    Some(tokio::spawn(async move {
+        run_recovery_disposition_maintenance_loop(api, cancel_token).await;
+    }))
+}
+
 fn spawn_manual_transition_job_recovery_once(api: Arc<ECStore>) -> Option<JoinHandle<()>> {
    if MANUAL_TRANSITION_JOB_RECOVERY_STARTED.set(()).is_err() {
        return None;
@@ -41,6 +41,21 @@ where
    com::read_config(api, file).await
 }

+pub(crate) async fn read_config_limited_preserve_empty<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<Vec<u8>>
+where
+    S: ObjectIO<
+            Error = Error,
+            RangeSpec = HTTPRangeSpec,
+            HeaderMap = HeaderMap,
+            ObjectOptions = ObjectOptions,
+            ObjectInfo = ObjectInfo,
+            GetObjectReader = GetObjectReader,
+            PutObjectReader = PutObjReader,
+        >,
+{
+    com::read_config_limited_preserve_empty(api, file, max_bytes).await
+}
+
 pub(crate) async fn read_config_with_metadata<S>(api: Arc<S>, file: &str, opts: &ObjectOptions) -> Result<(Vec<u8>, ObjectInfo)>
 where
    S: ObjectIO<
@@ -56,6 +71,26 @@ where
    com::read_config_with_metadata(api, file, opts).await
 }

+pub(crate) async fn read_config_limited_preserve_empty_with_metadata<S>(
+    api: Arc<S>,
+    file: &str,
+    opts: &ObjectOptions,
+    max_bytes: usize,
+) -> Result<(Vec<u8>, ObjectInfo)>
+where
+    S: ObjectIO<
+            Error = Error,
+            RangeSpec = HTTPRangeSpec,
+            HeaderMap = HeaderMap,
+            ObjectOptions = ObjectOptions,
+            ObjectInfo = ObjectInfo,
+            GetObjectReader = GetObjectReader,
+            PutObjectReader = PutObjReader,
+        >,
+{
+    com::read_config_limited_preserve_empty_with_metadata_opts(api, file, opts, max_bytes).await
+}
+
 pub(crate) async fn save_config<S>(api: Arc<S>, file: &str, data: Vec<u8>) -> Result<()>
 where
    S: ObjectIO<
@@ -126,20 +161,30 @@ where
            DeletedObject = DeletedObject,
        >,
 {
-    match api
-        .delete_object(
-            RUSTFS_META_BUCKET,
-            file,
-            ObjectOptions {
-                http_preconditions: Some(HTTPPreconditions {
-                    if_match: Some(etag.to_string()),
-                    ..Default::default()
-                }),
-                ..Default::default()
-            },
-        )
-        .await
-    {
+    delete_config_if_match_with_opts(api, file, etag, ObjectOptions::default()).await
+}
+
+pub(crate) async fn delete_config_if_match_with_opts<S>(
+    api: Arc<S>,
+    file: &str,
+    etag: &str,
+    mut options: ObjectOptions,
+) -> Result<()>
+where
+    S: ObjectOperations<
+            Error = Error,
+            ObjectInfo = ObjectInfo,
+            ObjectOptions = ObjectOptions,
+            FileInfo = FileInfo,
+            ObjectToDelete = ObjectToDelete,
+            DeletedObject = DeletedObject,
+        >,
+{
+    options.http_preconditions = Some(HTTPPreconditions {
+        if_match: Some(etag.to_string()),
+        ..Default::default()
+    });
+    match api.delete_object(RUSTFS_META_BUCKET, file, options).await {
        Ok(_) => Ok(()),
        Err(err) => {
            if err == Error::FileNotFound || matches!(err, Error::ObjectNotFound(_, _)) {
@@ -22,7 +22,7 @@ use super::{
    bucket_lifecycle_ops::{
        ManualTransitionQueueSnapshot, ManualTransitionRunReport, decode_manual_transition_continuation_token,
    },
-    manual_transition_job, recovery_control, tier_delete_journal, transition_transaction,
+    manual_transition_job, recovery_control, recovery_disposition, recovery_export, tier_delete_journal, transition_transaction,
 };
 use crate::error::{Error, Result};
 use crate::services::tier::tier_probe_intent;
@@ -42,6 +42,8 @@ pub(crate) enum DurableIlmRecordKind {
    ManualTransitionTask,
    ManualTransitionWorkerResult,
    RecoveryControl,
+    RecoveryExport,
+    RecoveryDisposition,
 }

 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -112,8 +114,20 @@ pub(crate) const RECOVERY_CONTROL_NAMESPACE: DurableIlmNamespace = DurableIlmNam
    max_record_size: recovery_control::MAX_ILM_RECOVERY_CONTROL_SIZE,
    kind: DurableIlmRecordKind::RecoveryControl,
 };
+pub(crate) const RECOVERY_EXPORT_NAMESPACE: DurableIlmNamespace = DurableIlmNamespace {
+    name: "recovery-export",
+    prefix: recovery_export::ILM_RECOVERY_EXPORT_PREFIX,
+    max_record_size: recovery_export::MAX_ILM_RECOVERY_EXPORT_SIZE,
+    kind: DurableIlmRecordKind::RecoveryExport,
+};
+pub(crate) const RECOVERY_DISPOSITION_NAMESPACE: DurableIlmNamespace = DurableIlmNamespace {
+    name: "recovery-disposition",
+    prefix: recovery_disposition::ILM_RECOVERY_DISPOSITION_PREFIX,
+    max_record_size: recovery_disposition::MAX_ILM_RECOVERY_DISPOSITION_SIZE,
+    kind: DurableIlmRecordKind::RecoveryDisposition,
+};

-pub(crate) const DURABLE_ILM_NAMESPACES: [DurableIlmNamespace; 10] = [
+pub(crate) const DURABLE_ILM_NAMESPACES: [DurableIlmNamespace; 12] = [
    TIER_DELETE_JOURNAL_NAMESPACE,
    TIER_DELETE_JOURNAL_V6_NAMESPACE,
    TIER_DELETE_DISPATCH_MANIFEST_NAMESPACE,
@@ -124,6 +138,8 @@ pub(crate) const DURABLE_ILM_NAMESPACES: [DurableIlmNamespace; 10] = [
    MANUAL_TRANSITION_TASK_NAMESPACE,
    MANUAL_TRANSITION_WORKER_RESULT_NAMESPACE,
    RECOVERY_CONTROL_NAMESPACE,
+    RECOVERY_EXPORT_NAMESPACE,
+    RECOVERY_DISPOSITION_NAMESPACE,
 ];

 #[derive(Debug, Clone, PartialEq, Eq)]
@@ -261,6 +277,28 @@ pub(crate) enum DurableIlmRecordCheckpoint {
        #[serde(default, skip_serializing_if = "Option::is_none")]
        owner_fence_sha256: Option<String>,
    },
+    RecoveryExport {
+        content_sha256: String,
+        source_generation_sha256: String,
+        topology_generation: String,
+        member_epochs_sha256: String,
+        creator_sha256: String,
+        retain_until_unix_nanos: i64,
+    },
+    RecoveryDisposition {
+        content_sha256: String,
+        identity_sha256: String,
+        copy_manifest_sha256: String,
+        copy_manifest_count: usize,
+        created_at_unix_nanos: i64,
+        revision: u64,
+        state: recovery_disposition::IlmRecoveryDispositionState,
+        owner_fence_sha256: Option<String>,
+        owner_lease_acquired_at_unix_nanos: Option<i64>,
+        owner_lease_expires_at_unix_nanos: Option<i64>,
+        confirmed_absent_sha256: Vec<String>,
+        retain_until_unix_nanos: i64,
+    },
 }

 impl DurableIlmRecordCheckpoint {
@@ -275,7 +313,9 @@ impl DurableIlmRecordCheckpoint {
            | Self::ManualTransitionScope { content_sha256, .. }
            | Self::ManualTransitionTask { content_sha256 }
            | Self::ManualTransitionWorkerResult { content_sha256 }
-            | Self::RecoveryControl { content_sha256, .. } => content_sha256,
+            | Self::RecoveryControl { content_sha256, .. }
+            | Self::RecoveryExport { content_sha256, .. }
+            | Self::RecoveryDisposition { content_sha256, .. } => content_sha256,
        }
    }

@@ -316,6 +356,9 @@ impl DurableIlmRecordCheckpoint {
            {
                return Err(Error::other("durable ILM tier delete journal checkpoint is invalid"));
            }
+            if !recovery_disposition_checkpoint_is_valid(checkpoint) {
+                return Err(Error::other("durable ILM recovery disposition checkpoint is invalid"));
+            }
        }
        if self == next {
            if let Self::ManualTransitionJob {
@@ -456,9 +499,7 @@ impl DurableIlmRecordCheckpoint {
                },
            ) => {
                previous_identity == next_identity
-                    && transition_state_distance(*previous_state, *next_state)
-                        .and_then(|distance| previous_revision.checked_add(distance))
-                        .is_some_and(|expected_revision| *next_revision == expected_revision)
+                    && transition_state_revision_is_successor(*previous_state, *previous_revision, *next_state, *next_revision)
                    && (!previous_remote_version_known || previous_remote_version == next_remote_version)
            }
            (
@@ -594,6 +635,83 @@ impl DurableIlmRecordCheckpoint {
                    && previous_attempts == next_attempts;
                adjacent && (claim || source_refresh || completion)
            }
+            (
+                Self::RecoveryDisposition {
+                    identity_sha256: previous_identity,
+                    copy_manifest_sha256: previous_manifest,
+                    copy_manifest_count: previous_manifest_count,
+                    created_at_unix_nanos: previous_created_at,
+                    revision: previous_revision,
+                    state: previous_state,
+                    owner_fence_sha256: previous_owner,
+                    owner_lease_acquired_at_unix_nanos: previous_owner_acquired,
+                    owner_lease_expires_at_unix_nanos: previous_owner_expires,
+                    confirmed_absent_sha256: previous_confirmed,
+                    retain_until_unix_nanos: previous_retain_until,
+                    ..
+                },
+                Self::RecoveryDisposition {
+                    identity_sha256: next_identity,
+                    copy_manifest_sha256: next_manifest,
+                    copy_manifest_count: next_manifest_count,
+                    created_at_unix_nanos: next_created_at,
+                    revision: next_revision,
+                    state: next_state,
+                    owner_fence_sha256: next_owner,
+                    owner_lease_acquired_at_unix_nanos: next_owner_acquired,
+                    owner_lease_expires_at_unix_nanos: next_owner_expires,
+                    confirmed_absent_sha256: next_confirmed,
+                    retain_until_unix_nanos: next_retain_until,
+                    ..
+                },
+            ) => {
+                use recovery_disposition::IlmRecoveryDispositionState::{Applying, Completed, Prepared};
+
+                let immutable_identity_matches = previous_identity == next_identity
+                    && previous_manifest == next_manifest
+                    && previous_manifest_count == next_manifest_count
+                    && previous_created_at == next_created_at
+                    && previous_retain_until == next_retain_until;
+                let adjacent = previous_revision.checked_add(1) == Some(*next_revision);
+                let progress_is_monotonic = sorted_sha256_set_is_subset(previous_confirmed, next_confirmed);
+                let legal_edge = match (previous_state, next_state) {
+                    (Prepared, Prepared) => {
+                        let claim = previous_owner.is_none() && next_owner.is_some();
+                        let takeover = previous_owner.is_some()
+                            && previous_owner != next_owner
+                            && previous_owner_expires
+                                .zip(*next_owner_acquired)
+                                .is_some_and(|(expires, acquired)| acquired >= expires);
+                        previous_confirmed == next_confirmed && (claim || takeover)
+                    }
+                    (Prepared, Applying) => {
+                        previous_confirmed == next_confirmed
+                            && previous_owner.is_some()
+                            && previous_owner == next_owner
+                            && previous_owner_acquired == next_owner_acquired
+                            && previous_owner_expires == next_owner_expires
+                    }
+                    (Applying, Applying) => {
+                        let progress = previous_owner == next_owner
+                            && previous_owner_acquired == next_owner_acquired
+                            && previous_owner_expires == next_owner_expires
+                            && previous_confirmed.len().checked_add(1) == Some(next_confirmed.len());
+                        let takeover = previous_owner.is_some()
+                            && previous_owner != next_owner
+                            && previous_confirmed == next_confirmed
+                            && previous_owner_expires
+                                .zip(*next_owner_acquired)
+                                .is_some_and(|(expires, acquired)| acquired >= expires);
+                        progress || takeover
+                    }
+                    (Applying, Completed) => {
+                        previous_owner.is_some() && next_owner.is_none() && previous_confirmed == next_confirmed
+                    }
+                    _ => false,
+                };
+
+                immutable_identity_matches && adjacent && progress_is_monotonic && legal_edge
+            }
            _ => false,
        };

@@ -611,6 +729,11 @@ impl DurableIlmRecordCheckpoint {
    /// after the exact terminal ETag and terminal receipt were committed, to
    /// purge older object versions exposed by that deletion.
    pub(crate) fn is_predecessor_of_terminal(&self, terminal: &Self) -> bool {
+        for checkpoint in [self, terminal] {
+            if !recovery_disposition_checkpoint_is_valid(checkpoint) {
+                return false;
+            }
+        }
        if let Self::TierProbeIntent { state, .. } = terminal
            && !matches!(
                state,
@@ -627,6 +750,11 @@ impl DurableIlmRecordCheckpoint {
        {
            return false;
        }
+        if let Self::RecoveryDisposition { state, .. } = terminal
+            && state != &recovery_disposition::IlmRecoveryDispositionState::Completed
+        {
+            return false;
+        }
        if self == terminal || self.validate_successor(terminal).is_ok() {
            return true;
        }
@@ -752,11 +880,121 @@ impl DurableIlmRecordCheckpoint {
                    && terminal_revision > previous_revision
                    && terminal_attempts >= previous_attempts
            }
+            (
+                Self::RecoveryDisposition {
+                    identity_sha256: previous_identity,
+                    copy_manifest_sha256: previous_manifest,
+                    copy_manifest_count: previous_manifest_count,
+                    created_at_unix_nanos: previous_created_at,
+                    revision: previous_revision,
+                    state: previous_state,
+                    owner_fence_sha256: previous_owner,
+                    confirmed_absent_sha256: previous_confirmed,
+                    retain_until_unix_nanos: previous_retain_until,
+                    ..
+                },
+                Self::RecoveryDisposition {
+                    identity_sha256: terminal_identity,
+                    copy_manifest_sha256: terminal_manifest,
+                    copy_manifest_count: terminal_manifest_count,
+                    created_at_unix_nanos: terminal_created_at,
+                    revision: terminal_revision,
+                    state: recovery_disposition::IlmRecoveryDispositionState::Completed,
+                    confirmed_absent_sha256: terminal_confirmed,
+                    retain_until_unix_nanos: terminal_retain_until,
+                    ..
+                },
+            ) => {
+                matches!(
+                    previous_state,
+                    recovery_disposition::IlmRecoveryDispositionState::Prepared
+                        | recovery_disposition::IlmRecoveryDispositionState::Applying
+                ) && previous_identity == terminal_identity
+                    && previous_manifest == terminal_manifest
+                    && previous_manifest_count == terminal_manifest_count
+                    && previous_created_at == terminal_created_at
+                    && previous_retain_until == terminal_retain_until
+                    && terminal_revision.checked_sub(*previous_revision).is_some_and(|distance| {
+                        let minimum_distance = match previous_state {
+                            recovery_disposition::IlmRecoveryDispositionState::Prepared if previous_owner.is_some() => 3,
+                            recovery_disposition::IlmRecoveryDispositionState::Prepared => 4,
+                            recovery_disposition::IlmRecoveryDispositionState::Applying
+                                if previous_confirmed.len() == *previous_manifest_count =>
+                            {
+                                1
+                            }
+                            recovery_disposition::IlmRecoveryDispositionState::Applying => 2,
+                            recovery_disposition::IlmRecoveryDispositionState::Completed => u64::MAX,
+                        };
+                        distance >= minimum_distance
+                    })
+                    && sorted_sha256_set_is_subset(previous_confirmed, terminal_confirmed)
+            }
            _ => false,
        }
    }
 }

+fn recovery_disposition_checkpoint_is_valid(checkpoint: &DurableIlmRecordCheckpoint) -> bool {
+    use recovery_disposition::IlmRecoveryDispositionState::{Applying, Completed, Prepared};
+
+    let DurableIlmRecordCheckpoint::RecoveryDisposition {
+        content_sha256,
+        identity_sha256,
+        copy_manifest_sha256,
+        copy_manifest_count,
+        created_at_unix_nanos,
+        revision,
+        state,
+        owner_fence_sha256,
+        owner_lease_acquired_at_unix_nanos,
+        owner_lease_expires_at_unix_nanos,
+        confirmed_absent_sha256,
+        retain_until_unix_nanos,
+    } = checkpoint
+    else {
+        return true;
+    };
+    let owner_fence_sha256 = owner_fence_sha256.as_deref();
+
+    is_canonical_sha256(content_sha256)
+        && is_canonical_sha256(identity_sha256)
+        && is_canonical_sha256(copy_manifest_sha256)
+        && *copy_manifest_count > 0
+        && *created_at_unix_nanos > 0
+        && *revision > 0
+        && *retain_until_unix_nanos > 0
+        && owner_fence_sha256.is_none_or(is_canonical_sha256)
+        && match (
+            owner_fence_sha256,
+            *owner_lease_acquired_at_unix_nanos,
+            *owner_lease_expires_at_unix_nanos,
+        ) {
+            (None, None, None) => true,
+            (Some(_), Some(acquired), Some(expires)) => acquired > 0 && expires > acquired,
+            _ => false,
+        }
+        && confirmed_absent_sha256.len() <= *copy_manifest_count
+        && confirmed_absent_sha256.iter().all(|digest| is_canonical_sha256(digest))
+        && confirmed_absent_sha256.windows(2).all(|pair| pair[0] < pair[1])
+        && match *state {
+            Prepared => confirmed_absent_sha256.is_empty(),
+            Applying => owner_fence_sha256.is_some(),
+            Completed => owner_fence_sha256.is_none() && confirmed_absent_sha256.len() == *copy_manifest_count,
+        }
+}
+
+fn is_canonical_sha256(value: &str) -> bool {
+    is_sha256_checksum(value)
+        && !value
+            .bytes()
+            .any(|byte| byte.is_ascii_hexdigit() && byte.is_ascii_uppercase())
+}
+
+fn sorted_sha256_set_is_subset(subset: &[String], superset: &[String]) -> bool {
+    subset.iter().all(|candidate| superset.binary_search(candidate).is_ok())
+}
+
 fn tier_delete_dispatch_parent_progress_delta(
    previous_sequence: u64,
    previous_completed_journals: u64,
@@ -790,6 +1028,22 @@ fn transition_state_distance(
    }
 }

+fn transition_state_revision_is_successor(
+    from: transition_transaction::TransitionTransactionState,
+    from_revision: u64,
+    to: transition_transaction::TransitionTransactionState,
+    to_revision: u64,
+) -> bool {
+    use transition_transaction::TransitionTransactionState::{LocalCommitStarted, UploadOutcomeUnknown};
+
+    if from == UploadOutcomeUnknown && from_revision == 1 && to == LocalCommitStarted {
+        return to_revision == 2;
+    }
+    transition_state_distance(from, to)
+        .and_then(|distance| from_revision.checked_add(distance))
+        .is_some_and(|expected_revision| to_revision == expected_revision)
+}
+
 fn tier_probe_state_reaches(
    from: tier_probe_intent::TierProbeIntentState,
    to: tier_probe_intent::TierProbeIntentState,
@@ -1348,6 +1602,55 @@ pub(crate) fn validate_durable_ilm_record(path: &str, data: &[u8]) -> Result<Val
                },
            )
        }
+        DurableIlmRecordKind::RecoveryExport => {
+            let (protocol, export_id) = recovery_export::recovery_export_id_from_record_object_name(path)?;
+            let export = recovery_export::IlmRecoveryExport::decode(&export_id, data)?;
+            let canonical = recovery_export::recovery_export_record_object_name(protocol, &export_id)?;
+            if canonical != path || export.protocol != protocol {
+                return Err(Error::other("ILM recovery export path is not canonical"));
+            }
+            let source_generation_sha256 = checkpoint_hash(&export.source_generation)?;
+            (
+                "export_id",
+                export_id,
+                DurableIlmRecordCheckpoint::RecoveryExport {
+                    content_sha256,
+                    source_generation_sha256,
+                    topology_generation: export.topology_generation,
+                    member_epochs_sha256: export.member_epochs_sha256,
+                    creator_sha256: export.creator_sha256,
+                    retain_until_unix_nanos: export.retain_until_unix_nanos,
+                },
+            )
+        }
+        DurableIlmRecordKind::RecoveryDisposition => {
+            // The disposition module owns strict schema, checksum, canonical
+            // path, immutable-manifest, and state-specific validation. Keep
+            // this boundary limited to decommission identity/checkpoint
+            // projection so the two readers cannot accept different records.
+            let disposition = recovery_disposition::decode_recovery_disposition_checkpoint(path, data)?;
+            if disposition.content_sha256 != content_sha256 {
+                return Err(Error::other("ILM recovery disposition checkpoint content digest is invalid"));
+            }
+            (
+                "disposition_id",
+                disposition.disposition_id,
+                DurableIlmRecordCheckpoint::RecoveryDisposition {
+                    content_sha256: disposition.content_sha256,
+                    identity_sha256: disposition.identity_sha256,
+                    copy_manifest_sha256: disposition.copy_manifest_sha256,
+                    copy_manifest_count: disposition.copy_manifest_count,
+                    created_at_unix_nanos: disposition.created_at_unix_nanos,
+                    revision: disposition.revision,
+                    state: disposition.state,
+                    owner_fence_sha256: disposition.owner_fence_sha256,
+                    owner_lease_acquired_at_unix_nanos: disposition.owner_lease_acquired_at_unix_nanos,
+                    owner_lease_expires_at_unix_nanos: disposition.owner_lease_expires_at_unix_nanos,
+                    confirmed_absent_sha256: disposition.confirmed_absent_sha256,
+                    retain_until_unix_nanos: disposition.retain_until_unix_nanos,
+                },
+            )
+        }
        DurableIlmRecordKind::ManualTransitionJob => {
            let job_id = manual_transition_job::manual_transition_job_id_from_record_object_name(path)
                .map_err(|err| Error::other(err.to_string()))?;
@@ -1503,6 +1806,203 @@ mod tests {
        }
    }

+    fn recovery_disposition_checkpoint(
+        revision: u64,
+        state: recovery_disposition::IlmRecoveryDispositionState,
+        owner_fence: Option<&str>,
+        confirmed_absent_sha256: Vec<String>,
+    ) -> DurableIlmRecordCheckpoint {
+        let (owner_lease_acquired_at_unix_nanos, owner_lease_expires_at_unix_nanos) = match owner_fence {
+            Some("f") => (Some(10), Some(20)),
+            Some(_) => (Some(1), Some(10)),
+            None => (None, None),
+        };
+        DurableIlmRecordCheckpoint::RecoveryDisposition {
+            content_sha256: format!("{revision:064x}"),
+            identity_sha256: "a".repeat(64),
+            copy_manifest_sha256: "d".repeat(64),
+            copy_manifest_count: 2,
+            created_at_unix_nanos: 1_700_000_000_000_000_000,
+            revision,
+            state,
+            owner_fence_sha256: owner_fence.map(|digest| digest.repeat(64)),
+            owner_lease_acquired_at_unix_nanos,
+            owner_lease_expires_at_unix_nanos,
+            confirmed_absent_sha256,
+            retain_until_unix_nanos: 1_820_000_000_000_000_000,
+        }
+    }
+
+    #[test]
+    fn recovery_disposition_namespace_is_registered_without_shadowing_its_root() {
+        let disposition_id = "a".repeat(64);
+        let path = format!(
+            "{}/tier_delete_journal/{}/{}/{}.json",
+            recovery_disposition::ILM_RECOVERY_DISPOSITION_PREFIX,
+            &disposition_id[..2],
+            &disposition_id[2..4],
+            disposition_id
+        );
+        let namespace = classify_durable_ilm_record(&path)
+            .expect("recovery disposition path should classify")
+            .expect("recovery disposition should be durable");
+
+        assert_eq!(namespace, &RECOVERY_DISPOSITION_NAMESPACE);
+        assert!(classify_durable_ilm_record(recovery_disposition::ILM_RECOVERY_DISPOSITION_PREFIX).is_err());
+    }
+
+    #[test]
+    fn recovery_disposition_checkpoint_accepts_only_monotonic_progress() {
+        use recovery_disposition::IlmRecoveryDispositionState::{Applying, Completed, Prepared};
+
+        let first_copy = "b".repeat(64);
+        let second_copy = "c".repeat(64);
+        let prepared = recovery_disposition_checkpoint(1, Prepared, None, Vec::new());
+        let claimed = recovery_disposition_checkpoint(2, Prepared, Some("e"), Vec::new());
+        let applying = recovery_disposition_checkpoint(3, Applying, Some("e"), Vec::new());
+        let first_absent = recovery_disposition_checkpoint(4, Applying, Some("e"), vec![first_copy.clone()]);
+        let taken_over = recovery_disposition_checkpoint(5, Applying, Some("f"), vec![first_copy.clone()]);
+        let all_absent = recovery_disposition_checkpoint(6, Applying, Some("f"), vec![first_copy.clone(), second_copy.clone()]);
+        let completed = recovery_disposition_checkpoint(7, Completed, None, vec![first_copy.clone(), second_copy.clone()]);
+
+        prepared
+            .validate_successor(&claimed)
+            .expect("Prepared should record an owner claim without absence progress");
+        claimed
+            .validate_successor(&applying)
+            .expect("Prepared should advance to Applying without folding in deletion progress");
+        applying
+            .validate_successor(&first_absent)
+            .expect("Applying should append newly confirmed absent copies");
+        first_absent
+            .validate_successor(&taken_over)
+            .expect("Applying should record a fenced owner takeover without losing progress");
+        taken_over
+            .validate_successor(&all_absent)
+            .expect("Applying should preserve every earlier confirmation while making progress");
+        all_absent
+            .validate_successor(&completed)
+            .expect("a fully confirmed manifest should advance to Completed");
+
+        assert!(
+            prepared.validate_successor(&completed).is_err(),
+            "adjacent receipt updates must not skip Applying"
+        );
+        assert!(
+            first_absent
+                .validate_successor(&recovery_disposition_checkpoint(5, Applying, Some("e"), Vec::new()))
+                .is_err(),
+            "confirmed-absent progress must not move backwards"
+        );
+        assert!(
+            applying
+                .validate_successor(&recovery_disposition_checkpoint(4, Completed, None, vec![first_copy.clone()]))
+                .is_err(),
+            "Completed must cover the complete immutable copy manifest"
+        );
+        assert!(
+            completed
+                .validate_successor(&recovery_disposition_checkpoint(7, Applying, Some("e"), vec![second_copy]))
+                .is_err(),
+            "Completed is terminal"
+        );
+        assert!(
+            first_absent
+                .validate_successor(&recovery_disposition_checkpoint(5, Applying, Some("e"), vec![first_copy.clone()]))
+                .is_err(),
+            "a same-state revision bump must change the owner fence or absence progress"
+        );
+        assert!(
+            applying
+                .validate_successor(&recovery_disposition_checkpoint(4, Applying, None, vec![first_copy]))
+                .is_err(),
+            "Applying must retain a fenced owner"
+        );
+
+        let mut noncanonical_identity = claimed.clone();
+        if let DurableIlmRecordCheckpoint::RecoveryDisposition { identity_sha256, .. } = &mut noncanonical_identity {
+            *identity_sha256 = "A".repeat(64);
+        }
+        assert!(prepared.validate_successor(&noncanonical_identity).is_err());
+        let mut changed_created_at = claimed;
+        if let DurableIlmRecordCheckpoint::RecoveryDisposition {
+            created_at_unix_nanos, ..
+        } = &mut changed_created_at
+        {
+            *created_at_unix_nanos += 1;
+        }
+        assert!(prepared.validate_successor(&changed_created_at).is_err());
+
+        let mut early_takeover = taken_over;
+        if let DurableIlmRecordCheckpoint::RecoveryDisposition {
+            owner_lease_acquired_at_unix_nanos,
+            ..
+        } = &mut early_takeover
+        {
+            *owner_lease_acquired_at_unix_nanos = Some(9);
+        }
+        assert!(first_absent.validate_successor(&early_takeover).is_err());
+        assert!(
+            applying
+                .validate_successor(&recovery_disposition_checkpoint(
+                    4,
+                    Applying,
+                    Some("e"),
+                    vec!["c".repeat(64), "b".repeat(64)],
+                ))
+                .is_err(),
+            "confirmed-absent entries must be a canonical sorted set"
+        );
+    }
+
+    #[test]
+    fn recovery_disposition_terminal_predecessor_requires_exact_identity_and_full_manifest() {
+        use recovery_disposition::IlmRecoveryDispositionState::{Applying, Completed, Prepared};
+
+        let first_copy = "b".repeat(64);
+        let second_copy = "c".repeat(64);
+        let prepared = recovery_disposition_checkpoint(1, Prepared, None, Vec::new());
+        let applying = recovery_disposition_checkpoint(3, Applying, Some("e"), vec![first_copy.clone()]);
+        let completed = recovery_disposition_checkpoint(5, Completed, None, vec![first_copy.clone(), second_copy]);
+
+        assert!(prepared.is_predecessor_of_terminal(&completed));
+        assert!(applying.is_predecessor_of_terminal(&completed));
+        assert!(
+            !prepared.is_predecessor_of_terminal(&recovery_disposition_checkpoint(2, Applying, Some("e"), Vec::new())),
+            "a nonterminal disposition must not authorize terminal cleanup"
+        );
+        assert!(
+            !prepared.is_predecessor_of_terminal(&recovery_disposition_checkpoint(
+                4,
+                Completed,
+                None,
+                vec![first_copy.clone(), "c".repeat(64)],
+            )),
+            "terminal proof must leave enough revisions for claim, apply, progress, and completion"
+        );
+        assert!(
+            !applying.is_predecessor_of_terminal(&recovery_disposition_checkpoint(
+                4,
+                Completed,
+                None,
+                vec![first_copy.clone(), "c".repeat(64)],
+            )),
+            "an incomplete Applying checkpoint cannot complete without a progress generation"
+        );
+
+        let mut other_identity = completed;
+        if let DurableIlmRecordCheckpoint::RecoveryDisposition { identity_sha256, .. } = &mut other_identity {
+            *identity_sha256 = "e".repeat(64);
+        }
+        assert!(!prepared.is_predecessor_of_terminal(&other_identity));
+
+        let incomplete_terminal = recovery_disposition_checkpoint(4, Completed, None, vec![first_copy]);
+        assert!(
+            !prepared.is_predecessor_of_terminal(&incomplete_terminal),
+            "a partial confirmed-absent set must not become terminal proof"
+        );
+    }
+
    fn tier_probe_intent_fixture() -> tier_probe_intent::TierProbeIntent {
        let probe_id = Uuid::parse_str("36e2220e-9ad2-495b-b3bc-c4d2caf70a31").expect("fixture uuid should parse");
        tier_probe_intent::TierProbeIntent {
@@ -1687,6 +2187,64 @@ mod tests {
        );
    }

+    #[test]
+    fn transition_checkpoint_accepts_only_the_distinguishable_compact_edge() {
+        let identity_sha256 = "a".repeat(64);
+        let unknown_remote_sha256 = "b".repeat(64);
+        let known_remote_sha256 = "c".repeat(64);
+        let checkpoint = |revision, state, remote_version_sha256: String, remote_version_known| {
+            DurableIlmRecordCheckpoint::TransitionTransaction {
+                content_sha256: format!("{revision:064x}"),
+                identity_sha256: identity_sha256.clone(),
+                remote_version_sha256,
+                remote_version_known,
+                revision,
+                state,
+            }
+        };
+        let compact_unknown = checkpoint(
+            1,
+            transition_transaction::TransitionTransactionState::UploadOutcomeUnknown,
+            unknown_remote_sha256.clone(),
+            false,
+        );
+        let compact_local_commit = checkpoint(
+            2,
+            transition_transaction::TransitionTransactionState::LocalCommitStarted,
+            known_remote_sha256.clone(),
+            true,
+        );
+        compact_unknown
+            .validate_successor(&compact_local_commit)
+            .expect("compact pre-upload fence should advance directly to the exact local-commit fence");
+
+        let legacy_unknown = checkpoint(
+            2,
+            transition_transaction::TransitionTransactionState::UploadOutcomeUnknown,
+            unknown_remote_sha256,
+            false,
+        );
+        let invalid_legacy_skip = checkpoint(
+            3,
+            transition_transaction::TransitionTransactionState::LocalCommitStarted,
+            known_remote_sha256.clone(),
+            true,
+        );
+        assert!(
+            legacy_unknown.validate_successor(&invalid_legacy_skip).is_err(),
+            "legacy UploadOutcomeUnknown@2 must not masquerade as the compact edge"
+        );
+        let valid_legacy_skip = checkpoint(
+            4,
+            transition_transaction::TransitionTransactionState::LocalCommitStarted,
+            known_remote_sha256,
+            true,
+        );
+        legacy_unknown
+            .validate_successor(&valid_legacy_skip)
+            .expect("legacy receipts may still observe the existing two-edge state advance");
+    }
+
    #[test]
    fn tier_delete_dispatch_manifest_namespace_validates_monotonic_branches() {
        use tier_delete_journal::TierDeleteDispatchManifestState::{Aborted, Aborting, Completed, DispatchAuthorized, Preparing};
@@ -25,6 +25,9 @@ mod object_handlers_common;
 mod object_lock_boundary;
 pub use self::core as lifecycle;
 pub mod recovery_control;
+pub mod recovery_disposition;
+pub(crate) mod recovery_disposition_runtime;
+pub mod recovery_export;
 mod replication_sink;
 pub mod rule;
 mod runtime_boundary;
@@ -168,7 +168,7 @@ impl IlmRecoverySourceGeneration {
        Ok(generation)
    }

-    fn validate(&self) -> Result<()> {
+    pub(crate) fn validate(&self) -> Result<()> {
        if self.source_schema.trim().is_empty() {
            return Err(IlmRecoveryControlError::Corrupt("source schema is empty"));
        }
@@ -485,6 +485,40 @@ impl IlmRecoveryControl {
        self.validate()
    }

+    pub fn abandon_for_operator(&mut self, expected_source_generation: &IlmRecoverySourceGeneration) -> Result<()> {
+        if self.owner.is_some()
+            || self.classification != IlmRecoveryClassification::RetainedAmbiguous
+            || &self.observed_source_generation != expected_source_generation
+        {
+            return Err(IlmRecoveryControlError::InvalidSuccessor(
+                "operator abandonment requires the exact ownerless retained source generation",
+            ));
+        }
+        self.bump_revision()?;
+        self.classification = IlmRecoveryClassification::Abandoned;
+        self.validate()
+    }
+
+    pub fn retry_for_operator(&mut self, expected_source_generation: &IlmRecoverySourceGeneration) -> Result<()> {
+        if self.owner.is_some()
+            || !matches!(
+                self.classification,
+                IlmRecoveryClassification::RetainedAmbiguous | IlmRecoveryClassification::OperatorRequired
+            )
+            || self.attempt_count == u64::MAX
+            || &self.observed_source_generation != expected_source_generation
+        {
+            return Err(IlmRecoveryControlError::InvalidSuccessor(
+                "operator retry requires the exact ownerless retained source generation",
+            ));
+        }
+        self.bump_revision()?;
+        self.classification = IlmRecoveryClassification::Retrying;
+        self.consecutive_failure_count = 0;
+        self.next_attempt_at_unix_nanos = None;
+        self.validate()
+    }
+
    pub fn validate_successor(&self, next: &Self) -> Result<()> {
        self.validate()?;
        next.validate()?;
@@ -508,12 +542,58 @@ impl IlmRecoveryControl {
                self.validate_failure_successor(next)
            }
            (Some(_), None) => self.validate_finish_successor(next),
+            (None, None)
+                if self.classification == IlmRecoveryClassification::RetainedAmbiguous
+                    && next.classification == IlmRecoveryClassification::Abandoned =>
+            {
+                self.validate_operator_abandon_successor(next)
+            }
+            (None, None)
+                if matches!(
+                    self.classification,
+                    IlmRecoveryClassification::RetainedAmbiguous | IlmRecoveryClassification::OperatorRequired
+                ) && next.classification == IlmRecoveryClassification::Retrying =>
+            {
+                self.validate_operator_retry_successor(next)
+            }
            (None, None) => Err(IlmRecoveryControlError::InvalidSuccessor(
                "ownerless control cannot advance without a claim",
            )),
        }
    }

+    fn validate_operator_abandon_successor(&self, next: &Self) -> Result<()> {
+        if next.observed_source_generation != self.observed_source_generation
+            || next.attempt_count != self.attempt_count
+            || next.consecutive_failure_count != self.consecutive_failure_count
+            || next.first_failure_at_unix_nanos != self.first_failure_at_unix_nanos
+            || next.last_failure_at_unix_nanos != self.last_failure_at_unix_nanos
+            || next.next_attempt_at_unix_nanos != self.next_attempt_at_unix_nanos
+            || next.last_error_code != self.last_error_code
+        {
+            return Err(IlmRecoveryControlError::InvalidSuccessor(
+                "operator abandonment changed recovery history or source generation",
+            ));
+        }
+        Ok(())
+    }
+
+    fn validate_operator_retry_successor(&self, next: &Self) -> Result<()> {
+        if next.observed_source_generation != self.observed_source_generation
+            || next.attempt_count != self.attempt_count
+            || next.consecutive_failure_count != 0
+            || next.first_failure_at_unix_nanos != self.first_failure_at_unix_nanos
+            || next.last_failure_at_unix_nanos != self.last_failure_at_unix_nanos
+            || next.next_attempt_at_unix_nanos.is_some()
+            || next.last_error_code != self.last_error_code
+        {
+            return Err(IlmRecoveryControlError::InvalidSuccessor(
+                "operator retry changed recovery history or source generation",
+            ));
+        }
+        Ok(())
+    }
+
    fn validate_claim_successor(&self, next: &Self) -> Result<()> {
        if self.classification != IlmRecoveryClassification::Retrying
            || next.classification != IlmRecoveryClassification::Retrying
@@ -827,6 +907,23 @@ pub async fn observe_recovery_source(
    api: Arc<ECStore>,
    canonical_path: &str,
    source_schema: &str,
+) -> EcstoreResult<ObservedIlmRecoverySource> {
+    observe_recovery_source_with_options(api, canonical_path, source_schema, false).await
+}
+
+pub(crate) async fn observe_recovery_source_no_lock(
+    api: Arc<ECStore>,
+    canonical_path: &str,
+    source_schema: &str,
+) -> EcstoreResult<ObservedIlmRecoverySource> {
+    observe_recovery_source_with_options(api, canonical_path, source_schema, true).await
+}
+
+async fn observe_recovery_source_with_options(
+    api: Arc<ECStore>,
+    canonical_path: &str,
+    source_schema: &str,
+    no_lock: bool,
 ) -> EcstoreResult<ObservedIlmRecoverySource> {
    validate_canonical_source_path(canonical_path).map_err(recovery_control_store_error)?;
    if source_schema.trim().is_empty() {
@@ -837,7 +934,16 @@ pub async fn observe_recovery_source(
    let mut observations = Vec::new();
    for set in api.all_set_disks() {
        let authority = format!("pool-{}/set-{}", set.pool_index, set.set_index);
-        match config_boundary::read_config_with_metadata(set, canonical_path, &ObjectOptions::default()).await {
+        match config_boundary::read_config_with_metadata(
+            set,
+            canonical_path,
+            &ObjectOptions {
+                no_lock,
+                ..Default::default()
+            },
+        )
+        .await
+        {
            Ok((data, metadata)) => {
                let etag = metadata
                    .etag
@@ -1255,6 +1361,99 @@ mod tests {
        ));
    }

+    #[test]
+    fn operator_abandonment_is_an_exact_ownerless_retained_successor() {
+        let mut retained = IlmRecoveryControl::new(
+            control().identity,
+            generation(),
+            IlmRecoveryClassification::RetainedAmbiguous,
+            1_000_000_000,
+            IlmRecoveryErrorCode::OperatorDispositionRequired,
+        )
+        .expect("retained control should build");
+        let previous = retained.clone();
+        retained
+            .abandon_for_operator(&previous.observed_source_generation)
+            .expect("exact retained generation should be abandonable");
+        previous
+            .validate_successor(&retained)
+            .expect("operator abandonment should be a valid successor");
+        assert_eq!(retained.classification, IlmRecoveryClassification::Abandoned);
+        assert_eq!(retained.revision, previous.revision + 1);
+
+        let mut wrong_generation = previous.clone();
+        let mut generation = previous.observed_source_generation.clone();
+        generation.source_etag = "different".to_string();
+        assert!(wrong_generation.abandon_for_operator(&generation).is_err());
+
+        let mut mutated_history = retained.clone();
+        mutated_history.attempt_count += 1;
+        assert!(previous.validate_successor(&mutated_history).is_err());
+    }
+
+    #[test]
+    fn operator_retry_rearms_exact_retained_generation_without_resetting_history() {
+        for classification in [
+            IlmRecoveryClassification::RetainedAmbiguous,
+            IlmRecoveryClassification::OperatorRequired,
+        ] {
+            let mut retained = control();
+            retained
+                .claim("node-a", Uuid::new_v4(), 2_000_000_000, 1)
+                .expect("attempt should claim");
+            retained
+                .record_retryable_failure(2_000_000_001, IlmRecoveryErrorCode::BackendTimeout)
+                .expect("failure should persist");
+            retained.classification = classification;
+            retained.next_attempt_at_unix_nanos = None;
+            if classification == IlmRecoveryClassification::OperatorRequired {
+                retained.attempt_count = u64::from(MAX_RECOVERY_ATTEMPTS);
+                retained.consecutive_failure_count = MAX_RECOVERY_ATTEMPTS;
+            }
+            retained.validate().expect("retained control should remain valid");
+
+            let previous = retained.clone();
+            retained
+                .retry_for_operator(&previous.observed_source_generation)
+                .expect("exact retained generation should be retryable");
+            previous
+                .validate_successor(&retained)
+                .expect("operator retry should be a valid successor");
+            assert_eq!(retained.classification, IlmRecoveryClassification::Retrying);
+            assert_eq!(retained.revision, previous.revision + 1);
+            assert_eq!(retained.attempt_count, previous.attempt_count);
+            assert_eq!(retained.first_failure_at_unix_nanos, previous.first_failure_at_unix_nanos);
+            assert_eq!(retained.last_failure_at_unix_nanos, previous.last_failure_at_unix_nanos);
+            assert_eq!(retained.last_error_code, previous.last_error_code);
+            assert_eq!(retained.consecutive_failure_count, 0);
+            assert_eq!(retained.next_attempt_at_unix_nanos, None);
+            assert!(retained.should_attempt_at(2_000_000_002));
+
+            if classification == IlmRecoveryClassification::OperatorRequired {
+                retained
+                    .claim("node-b", Uuid::new_v4(), 2_000_000_002, 1)
+                    .expect("operator retry should authorize one new bounded attempt");
+                retained
+                    .record_retryable_failure(2_000_000_003, IlmRecoveryErrorCode::BackendTimeout)
+                    .expect("the bounded attempt failure should persist");
+                assert_eq!(retained.classification, IlmRecoveryClassification::OperatorRequired);
+                assert_eq!(retained.attempt_count, u64::from(MAX_RECOVERY_ATTEMPTS) + 1);
+                assert_eq!(retained.consecutive_failure_count, 1);
+            }
+
+            let mut wrong_generation = previous;
+            let mut changed_generation = wrong_generation.observed_source_generation.clone();
+            changed_generation.source_etag = "changed".to_string();
+            assert!(wrong_generation.retry_for_operator(&changed_generation).is_err());
+        }
+
+        let mut exhausted = control();
+        exhausted.classification = IlmRecoveryClassification::OperatorRequired;
+        exhausted.attempt_count = u64::MAX;
+        let generation = exhausted.observed_source_generation.clone();
+        assert!(exhausted.retry_for_operator(&generation).is_err());
+    }
+
    #[test]
    fn recovery_control_view_redacts_source_and_owner_details() {
        let mut control = control();
@@ -0,0 +1,840 @@
+// Copyright 2024 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::{collections::HashSet, sync::Arc};
+
+use rustfs_utils::crypto::{hex_sha256, is_sha256_checksum};
+use serde::{Deserialize, Serialize};
+
+use super::config_boundary;
+use super::recovery_control::{
+    IlmRecoveryClassification, IlmRecoveryControl, IlmRecoveryProtocol, IlmRecoverySourceCopy, IlmRecoverySourceGeneration,
+    MAX_ILM_RECOVERY_CONTROL_SIZE, ObservedIlmRecoveryControl, ObservedIlmRecoverySource, recovery_control_record_object_name,
+};
+use super::tier_delete_journal::{
+    TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA, TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA, validate_legacy_tier_delete_recovery_source,
+};
+use crate::disk::RUSTFS_META_BUCKET;
+use crate::error::{Error, Result};
+use crate::object_api::{ObjectOptions, WriteCompletion};
+use crate::services::notification_sys::{
+    acquire_ilm_recovery_export_fleet_proof, ilm_recovery_export_fleet_proof_matches, ilm_recovery_export_member_epochs_sha256,
+    ilm_recovery_export_topology_generation,
+};
+use crate::storage_api_contracts::{list::ListOperations as _, namespace::NamespaceLocking as _, object::HTTPPreconditions};
+use crate::store::ECStore;
+
+pub const ILM_RECOVERY_EXPORT_SCHEMA: &str = "rustfs-ilm-recovery-export-v1";
+pub const ILM_RECOVERY_EXPORT_PREFIX: &str = "ilm/recovery-exports";
+pub const MAX_ILM_RECOVERY_EXPORT_SIZE: usize = 128 * 1024;
+const MAX_ILM_RECOVERY_EXPORTS: usize = 10_000;
+const MAX_ILM_RECOVERY_EXPORT_BYTES: u64 = 1024 * 1024 * 1024;
+const MAX_ACTOR_EXPORTS_PER_MINUTE: usize = 10;
+const MAX_CLUSTER_EXPORTS_PER_MINUTE: usize = 100;
+const EXPORT_RETENTION_NANOS: i64 = 90 * 24 * 60 * 60 * 1_000_000_000;
+const EXPORT_ADMISSION_LOCK: &str = "ilm/recovery-admission/export.lock";
+const MAX_LEGACY_TIER_DELETE_SOURCE_SIZE: usize = 64 * 1024;
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct IlmRecoveryExportObservation {
+    pub control_id: String,
+    pub protocol: IlmRecoveryProtocol,
+    pub control_etag: String,
+    pub control_revision: u64,
+    pub classification: IlmRecoveryClassification,
+    pub canonical_source_path: String,
+    pub source_generation: IlmRecoverySourceGeneration,
+    pub topology_generation: String,
+    pub member_epochs_sha256: String,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct IlmRecoveryExport {
+    pub export_id: String,
+    pub control_id: String,
+    pub protocol: IlmRecoveryProtocol,
+    pub control_etag: String,
+    pub control_revision: u64,
+    pub classification: IlmRecoveryClassification,
+    pub canonical_source_path: String,
+    pub source_generation: IlmRecoverySourceGeneration,
+    pub topology_generation: String,
+    pub member_epochs_sha256: String,
+    pub creator_sha256: String,
+    pub created_at_unix_nanos: i64,
+    pub retain_until_unix_nanos: i64,
+    pub source_bytes_base64: String,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+struct PersistedIlmRecoveryExport {
+    schema: String,
+    content_sha256: String,
+    export: IlmRecoveryExport,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+pub struct IlmRecoveryExportCreated {
+    pub export_id: String,
+    pub content_sha256: String,
+    pub encoded: Vec<u8>,
+    pub replayed: bool,
+}
+
+impl IlmRecoveryExport {
+    fn validate(&self) -> Result<()> {
+        self.source_generation.validate().map_err(Error::other)?;
+        validate_sha256(&self.export_id, "ILM recovery export ID is invalid")?;
+        validate_sha256(&self.control_id, "ILM recovery export control ID is invalid")?;
+        validate_sha256(&self.topology_generation, "ILM recovery export topology generation is invalid")?;
+        validate_sha256(&self.member_epochs_sha256, "ILM recovery export member epoch digest is invalid")?;
+        validate_sha256(&self.creator_sha256, "ILM recovery export creator digest is invalid")?;
+        if self.protocol != IlmRecoveryProtocol::TierDeleteJournal
+            || self.classification != IlmRecoveryClassification::RetainedAmbiguous
+            || !is_legacy_export_schema(&self.source_generation.source_schema)
+        {
+            return Err(Error::other("ILM recovery export source is not an exportable legacy journal"));
+        }
+        if self.control_etag.trim().is_empty() || self.control_revision == 0 {
+            return Err(Error::other("ILM recovery export control generation is invalid"));
+        }
+        if self.canonical_source_path.is_empty()
+            || self.canonical_source_path.starts_with('/')
+            || self.canonical_source_path.ends_with('/')
+            || self.canonical_source_path.split('/').any(str::is_empty)
+        {
+            return Err(Error::other("ILM recovery export source path is invalid"));
+        }
+        if self.created_at_unix_nanos <= 0
+            || self.retain_until_unix_nanos < self.created_at_unix_nanos.saturating_add(EXPORT_RETENTION_NANOS)
+        {
+            return Err(Error::other("ILM recovery export retention is invalid"));
+        }
+        let source = base64_simd::STANDARD
+            .decode_to_vec(self.source_bytes_base64.as_bytes())
+            .map_err(|_| Error::other("ILM recovery export source encoding is invalid"))?;
+        validate_legacy_tier_delete_recovery_source(&self.canonical_source_path, &self.source_generation.source_schema, &source)?;
+        let encoded_len = u64::try_from(source.len()).map_err(|_| Error::other("ILM recovery export source length overflow"))?;
+        if source.is_empty()
+            || source.len() > MAX_LEGACY_TIER_DELETE_SOURCE_SIZE
+            || hex_sha256(&source, ToOwned::to_owned) != self.source_generation.content_sha256
+            || self.source_generation.copies.iter().any(|copy| {
+                copy.canonical_path != self.canonical_source_path
+                    || copy.etag != self.source_generation.source_etag
+                    || copy.content_sha256 != self.source_generation.content_sha256
+                    || copy.encoded_len != encoded_len
+            })
+        {
+            return Err(Error::other("ILM recovery export source bytes do not match the observed generation"));
+        }
+        if recovery_export_id(&self.control_id, &self.source_generation)? != self.export_id {
+            return Err(Error::other("ILM recovery export ID does not match its source generation"));
+        }
+        Ok(())
+    }
+
+    pub fn encode(&self) -> Result<Vec<u8>> {
+        self.validate()?;
+        let export_bytes = serde_json::to_vec(self).map_err(Error::other)?;
+        let persisted = PersistedIlmRecoveryExport {
+            schema: ILM_RECOVERY_EXPORT_SCHEMA.to_string(),
+            content_sha256: hex_sha256(&export_bytes, ToOwned::to_owned),
+            export: self.clone(),
+        };
+        let encoded = serde_json::to_vec(&persisted).map_err(Error::other)?;
+        if encoded.len() > MAX_ILM_RECOVERY_EXPORT_SIZE {
+            return Err(Error::other("encoded ILM recovery export exceeds maximum size"));
+        }
+        Ok(encoded)
+    }
+
+    pub fn decode(expected_export_id: &str, data: &[u8]) -> Result<Self> {
+        validate_sha256(expected_export_id, "ILM recovery export ID is invalid")?;
+        if data.len() > MAX_ILM_RECOVERY_EXPORT_SIZE {
+            return Err(Error::other("encoded ILM recovery export exceeds maximum size"));
+        }
+        let persisted: PersistedIlmRecoveryExport = serde_json::from_slice(data).map_err(Error::other)?;
+        if persisted.schema != ILM_RECOVERY_EXPORT_SCHEMA {
+            return Err(Error::other("ILM recovery export schema is unsupported"));
+        }
+        validate_sha256(&persisted.content_sha256, "ILM recovery export checksum is invalid")?;
+        let export_bytes = serde_json::to_vec(&persisted.export).map_err(Error::other)?;
+        if hex_sha256(&export_bytes, ToOwned::to_owned) != persisted.content_sha256 {
+            return Err(Error::other("ILM recovery export checksum mismatch"));
+        }
+        persisted.export.validate()?;
+        if persisted.export.export_id != expected_export_id {
+            return Err(Error::other("ILM recovery export ID does not match record key"));
+        }
+        Ok(persisted.export)
+    }
+}
+
+pub fn recovery_export_record_object_name(protocol: IlmRecoveryProtocol, export_id: &str) -> Result<String> {
+    validate_sha256(export_id, "ILM recovery export ID is invalid")?;
+    Ok(format!(
+        "{}/{}/{}/{}/{}.json",
+        ILM_RECOVERY_EXPORT_PREFIX,
+        protocol.as_str(),
+        &export_id[..2],
+        &export_id[2..4],
+        export_id
+    ))
+}
+
+pub fn recovery_export_id_from_record_object_name(object: &str) -> Result<(IlmRecoveryProtocol, String)> {
+    let suffix = object
+        .strip_prefix(ILM_RECOVERY_EXPORT_PREFIX)
+        .and_then(|suffix| suffix.strip_prefix('/'))
+        .ok_or_else(|| Error::other("ILM recovery export path has wrong prefix"))?;
+    let mut parts = suffix.split('/');
+    let protocol = match parts.next() {
+        Some("tier_delete_journal") => IlmRecoveryProtocol::TierDeleteJournal,
+        _ => return Err(Error::other("ILM recovery export protocol is invalid")),
+    };
+    let shard_a = parts
+        .next()
+        .ok_or_else(|| Error::other("ILM recovery export path is incomplete"))?;
+    let shard_b = parts
+        .next()
+        .ok_or_else(|| Error::other("ILM recovery export path is incomplete"))?;
+    let export_id = parts
+        .next()
+        .and_then(|name| name.strip_suffix(".json"))
+        .ok_or_else(|| Error::other("ILM recovery export suffix is invalid"))?;
+    if parts.next().is_some() {
+        return Err(Error::other("ILM recovery export path is not canonical"));
+    }
+    validate_sha256(export_id, "ILM recovery export ID is invalid")?;
+    if shard_a != &export_id[..2] || shard_b != &export_id[2..4] {
+        return Err(Error::other("ILM recovery export shard does not match export ID"));
+    }
+    Ok((protocol, export_id.to_string()))
+}
+
+pub async fn inspect_recovery_export_observation(api: Arc<ECStore>, control_id: &str) -> Result<IlmRecoveryExportObservation> {
+    let proof = acquire_ilm_recovery_export_fleet_proof()
+        .await
+        .ok_or_else(|| Error::other("ILM recovery export fleet proof is unavailable"))?;
+    let observed_control = load_exportable_control(api.clone(), control_id).await?;
+    let observed_source = observe_export_source(
+        api,
+        &observed_control.control.identity.canonical_source_path,
+        &observed_control.control.observed_source_generation.source_schema,
+    )
+    .await?;
+    if !observed_source.is_consistent()
+        || observed_source.generation != observed_control.control.observed_source_generation
+        || !ilm_recovery_export_fleet_proof_matches(&proof).await
+    {
+        return Err(Error::other("ILM recovery export observation changed or is incomplete"));
+    }
+    Ok(IlmRecoveryExportObservation {
+        control_id: control_id.to_string(),
+        protocol: observed_control.control.identity.protocol,
+        control_etag: observed_control.etag,
+        control_revision: observed_control.control.revision,
+        classification: observed_control.control.classification,
+        canonical_source_path: observed_control.control.identity.canonical_source_path,
+        source_generation: observed_source.generation,
+        topology_generation: ilm_recovery_export_topology_generation(&proof),
+        member_epochs_sha256: ilm_recovery_export_member_epochs_sha256(&proof),
+    })
+}
+
+pub async fn create_recovery_export(
+    api: Arc<ECStore>,
+    observation: &IlmRecoveryExportObservation,
+    creator_sha256: &str,
+) -> Result<IlmRecoveryExportCreated> {
+    validate_sha256(creator_sha256, "ILM recovery export creator digest is invalid")?;
+    let lock = api.new_ns_lock(RUSTFS_META_BUCKET, EXPORT_ADMISSION_LOCK).await?;
+    let admission_guard = lock.get_write_lock(crate::set_disk::get_lock_acquire_timeout()).await?;
+
+    let proof = acquire_ilm_recovery_export_fleet_proof()
+        .await
+        .ok_or_else(|| Error::other("ILM recovery export fleet proof is unavailable"))?;
+    if ilm_recovery_export_topology_generation(&proof) != observation.topology_generation
+        || ilm_recovery_export_member_epochs_sha256(&proof) != observation.member_epochs_sha256
+    {
+        return Err(Error::PreconditionFailed);
+    }
+    let control_object = recovery_control_record_object_name(IlmRecoveryProtocol::TierDeleteJournal, &observation.control_id)
+        .map_err(Error::other)?;
+    let control_lock = api.new_ns_lock(RUSTFS_META_BUCKET, &control_object).await?;
+    let control_guard = control_lock
+        .get_read_lock(crate::set_disk::get_lock_acquire_timeout())
+        .await?;
+    let source_lock = api
+        .new_ns_lock(RUSTFS_META_BUCKET, &observation.canonical_source_path)
+        .await?;
+    let source_guard = source_lock.get_read_lock(crate::set_disk::get_lock_acquire_timeout()).await?;
+    let locks_current = || !admission_guard.is_lock_lost() && !control_guard.is_lock_lost() && !source_guard.is_lock_lost();
+    let (current, current_source_bytes) = current_observation_under_proof_no_lock(api.clone(), observation, &proof).await?;
+    if &current != observation || !locks_current() {
+        return Err(Error::PreconditionFailed);
+    }
+    let current_source_base64 = base64_simd::STANDARD.encode_to_string(current_source_bytes);
+    let candidate_export_id = recovery_export_id(&current.control_id, &current.source_generation)?;
+    let object = recovery_export_record_object_name(current.protocol, &candidate_export_id)?;
+    match load_recovery_export_decoded(api.clone(), &candidate_export_id).await {
+        Ok((existing, export)) if export_matches_observation(&export, observation) => {
+            if !locks_current() || !ilm_recovery_export_fleet_proof_matches(&proof).await {
+                return Err(Error::PreconditionFailed);
+            }
+            api.record_durable_ilm_decommission_progress(&object, &existing.encoded)
+                .await?;
+            if !locks_current() {
+                return Err(Error::PreconditionFailed);
+            }
+            return Ok(existing.with_replayed());
+        }
+        Ok(_) => return Err(Error::PreconditionFailed),
+        Err(Error::ConfigNotFound) => {}
+        Err(err) => return Err(err),
+    }
+    let inventory = collect_export_inventory(api.clone()).await?;
+    if !locks_current() || !ilm_recovery_export_fleet_proof_matches(&proof).await {
+        return Err(Error::PreconditionFailed);
+    }
+    let created_at_unix_nanos = now_unix_nanos()?;
+    let export = build_export_from_source(&current, creator_sha256, created_at_unix_nanos, &current_source_base64)?;
+    let encoded = export.encode()?;
+    inventory.check(creator_sha256, encoded.len(), created_at_unix_nanos)?;
+
+    let mut write_options = ObjectOptions {
+        max_parity: true,
+        write_completion: WriteCompletion::TailDrained,
+        http_preconditions: Some(HTTPPreconditions {
+            if_none_match: Some("*".to_string()),
+            ..Default::default()
+        }),
+        ..Default::default()
+    };
+    write_options.add_namespace_lock_guard(&admission_guard);
+    write_options.add_namespace_lock_guard(&control_guard);
+    write_options.add_namespace_lock_guard(&source_guard);
+    if !locks_current() {
+        return Err(Error::PreconditionFailed);
+    }
+    let write_result = config_boundary::save_config_with_opts(api.clone(), &object, encoded.clone(), &write_options).await;
+    let stored = match load_recovery_export(api.clone(), &export.export_id).await {
+        Ok(stored) if stored.encoded == encoded => stored,
+        Ok(_) => return Err(Error::PreconditionFailed),
+        Err(read_err) => return Err(write_result.err().unwrap_or(read_err)),
+    };
+    if !locks_current() || !ilm_recovery_export_fleet_proof_matches(&proof).await {
+        return Err(Error::PreconditionFailed);
+    }
+    api.record_durable_ilm_decommission_progress(&object, &encoded).await?;
+    if !locks_current() {
+        return Err(Error::PreconditionFailed);
+    }
+    Ok(stored)
+}
+
+pub async fn load_recovery_export(api: Arc<ECStore>, export_id: &str) -> Result<IlmRecoveryExportCreated> {
+    let (created, _) = load_recovery_export_decoded(api, export_id).await?;
+    Ok(created)
+}
+
+async fn load_recovery_export_decoded(
+    api: Arc<ECStore>,
+    export_id: &str,
+) -> Result<(IlmRecoveryExportCreated, IlmRecoveryExport)> {
+    let object = recovery_export_record_object_name(IlmRecoveryProtocol::TierDeleteJournal, export_id)?;
+    let encoded = config_boundary::read_config_limited_preserve_empty(api, &object, MAX_ILM_RECOVERY_EXPORT_SIZE).await?;
+    let export = IlmRecoveryExport::decode(export_id, &encoded)?;
+    let content_sha256 = hex_sha256(&encoded, ToOwned::to_owned);
+    Ok((
+        IlmRecoveryExportCreated {
+            export_id: export.export_id.clone(),
+            content_sha256,
+            encoded,
+            replayed: false,
+        },
+        export,
+    ))
+}
+
+impl IlmRecoveryExportCreated {
+    fn with_replayed(mut self) -> Self {
+        self.replayed = true;
+        self
+    }
+}
+
+async fn load_exportable_control(api: Arc<ECStore>, control_id: &str) -> Result<ObservedIlmRecoveryControl> {
+    load_exportable_control_with_options(api, control_id, &ObjectOptions::default()).await
+}
+
+async fn load_exportable_control_no_lock(api: Arc<ECStore>, control_id: &str) -> Result<ObservedIlmRecoveryControl> {
+    load_exportable_control_with_options(
+        api,
+        control_id,
+        &ObjectOptions {
+            no_lock: true,
+            ..Default::default()
+        },
+    )
+    .await
+}
+
+async fn load_exportable_control_with_options(
+    api: Arc<ECStore>,
+    control_id: &str,
+    options: &ObjectOptions,
+) -> Result<ObservedIlmRecoveryControl> {
+    let object = recovery_control_record_object_name(IlmRecoveryProtocol::TierDeleteJournal, control_id).map_err(Error::other)?;
+    let (data, metadata) =
+        config_boundary::read_config_limited_preserve_empty_with_metadata(api, &object, options, MAX_ILM_RECOVERY_CONTROL_SIZE)
+            .await?;
+    let etag = metadata
+        .etag
+        .filter(|etag| !etag.trim().is_empty())
+        .ok_or_else(|| Error::other("ILM recovery control is missing an ETag"))?;
+    let control = IlmRecoveryControl::decode(control_id, &data).map_err(Error::other)?;
+    if control.identity.protocol != IlmRecoveryProtocol::TierDeleteJournal
+        || control.classification != IlmRecoveryClassification::RetainedAmbiguous
+        || !is_legacy_export_schema(&control.observed_source_generation.source_schema)
+    {
+        return Err(Error::other("ILM recovery control is not exportable"));
+    }
+    Ok(ObservedIlmRecoveryControl { control, etag })
+}
+
+async fn current_observation_under_proof_no_lock(
+    api: Arc<ECStore>,
+    expected: &IlmRecoveryExportObservation,
+    proof: &crate::services::notification_sys::IlmRecoveryExportFleetProofToken,
+) -> Result<(IlmRecoveryExportObservation, Vec<u8>)> {
+    let observed_control = load_exportable_control_no_lock(api.clone(), &expected.control_id).await?;
+    let observed_source = observe_export_source_no_lock(
+        api,
+        &observed_control.control.identity.canonical_source_path,
+        &observed_control.control.observed_source_generation.source_schema,
+    )
+    .await?;
+    let source_bytes = observed_source
+        .canonical_data
+        .clone()
+        .ok_or_else(|| Error::other("ILM recovery export source copies diverge"))?;
+    if !observed_source.is_consistent()
+        || observed_source.generation != observed_control.control.observed_source_generation
+        || !ilm_recovery_export_fleet_proof_matches(proof).await
+    {
+        return Err(Error::PreconditionFailed);
+    }
+    Ok((
+        IlmRecoveryExportObservation {
+            control_id: expected.control_id.clone(),
+            protocol: observed_control.control.identity.protocol,
+            control_etag: observed_control.etag,
+            control_revision: observed_control.control.revision,
+            classification: observed_control.control.classification,
+            canonical_source_path: observed_control.control.identity.canonical_source_path,
+            source_generation: observed_source.generation,
+            topology_generation: ilm_recovery_export_topology_generation(proof),
+            member_epochs_sha256: ilm_recovery_export_member_epochs_sha256(proof),
+        },
+        source_bytes,
+    ))
+}
+
+async fn observe_export_source(
+    api: Arc<ECStore>,
+    canonical_path: &str,
+    source_schema: &str,
+) -> Result<ObservedIlmRecoverySource> {
+    if canonical_path.is_empty()
+        || canonical_path.starts_with('/')
+        || canonical_path.ends_with('/')
+        || canonical_path.split('/').any(str::is_empty)
+        || !is_legacy_export_schema(source_schema)
+    {
+        return Err(Error::other("ILM recovery export source identity is invalid"));
+    }
+    let lock = api.new_ns_lock(RUSTFS_META_BUCKET, canonical_path).await?;
+    let _guard = lock.get_read_lock(crate::set_disk::get_lock_acquire_timeout()).await?;
+    observe_export_source_no_lock(api, canonical_path, source_schema).await
+}
+
+async fn observe_export_source_no_lock(
+    api: Arc<ECStore>,
+    canonical_path: &str,
+    source_schema: &str,
+) -> Result<ObservedIlmRecoverySource> {
+    let mut copies = Vec::new();
+    let mut canonical: Option<(String, String, Vec<u8>)> = None;
+    let mut consistent = true;
+    for set in api.all_set_disks() {
+        let authority = format!("pool-{}/set-{}", set.pool_index, set.set_index);
+        let result = config_boundary::read_config_limited_preserve_empty_with_metadata(
+            set,
+            canonical_path,
+            &ObjectOptions {
+                no_lock: true,
+                ..Default::default()
+            },
+            MAX_LEGACY_TIER_DELETE_SOURCE_SIZE,
+        )
+        .await;
+        match result {
+            Ok((data, metadata)) => {
+                if data.is_empty() || data.len() > MAX_LEGACY_TIER_DELETE_SOURCE_SIZE {
+                    return Err(Error::other("ILM recovery export source exceeds its protocol size limit"));
+                }
+                validate_legacy_tier_delete_recovery_source(canonical_path, source_schema, &data)?;
+                let etag = metadata
+                    .etag
+                    .filter(|etag| !etag.trim().is_empty())
+                    .ok_or_else(|| Error::other("ILM recovery export source copy is missing an ETag"))?;
+                let content_sha256 = hex_sha256(&data, ToOwned::to_owned);
+                let encoded_len =
+                    u64::try_from(data.len()).map_err(|_| Error::other("ILM recovery export source length does not fit u64"))?;
+                copies.push(IlmRecoverySourceCopy {
+                    authority,
+                    canonical_path: canonical_path.to_string(),
+                    etag: etag.clone(),
+                    encoded_len,
+                    content_sha256: content_sha256.clone(),
+                });
+                match canonical.as_ref() {
+                    Some((first_etag, first_digest, first_data)) => {
+                        consistent &= first_etag == &etag && first_digest == &content_sha256 && first_data == &data;
+                    }
+                    None => canonical = Some((etag, content_sha256, data)),
+                }
+            }
+            Err(err) if export_source_is_missing(&err) => {}
+            Err(err) => return Err(err),
+        }
+    }
+    let Some((source_etag, content_sha256, source_bytes)) = canonical else {
+        return Err(Error::ConfigNotFound);
+    };
+    let generation =
+        IlmRecoverySourceGeneration::new(source_schema, source_etag, content_sha256, copies).map_err(Error::other)?;
+    Ok(ObservedIlmRecoverySource {
+        generation,
+        canonical_data: consistent.then_some(source_bytes),
+    })
+}
+
+fn export_source_is_missing(err: &Error) -> bool {
+    matches!(
+        err,
+        Error::ConfigNotFound | Error::FileNotFound | Error::ObjectNotFound(_, _) | Error::VersionNotFound(_, _, _)
+    )
+}
+
+fn build_export_from_source(
+    observation: &IlmRecoveryExportObservation,
+    creator_sha256: &str,
+    created_at_unix_nanos: i64,
+    source_bytes_base64: &str,
+) -> Result<IlmRecoveryExport> {
+    let retain_until_unix_nanos = created_at_unix_nanos
+        .checked_add(EXPORT_RETENTION_NANOS)
+        .ok_or_else(|| Error::other("ILM recovery export retention timestamp overflow"))?;
+    let export = IlmRecoveryExport {
+        export_id: recovery_export_id(&observation.control_id, &observation.source_generation)?,
+        control_id: observation.control_id.clone(),
+        protocol: observation.protocol,
+        control_etag: observation.control_etag.clone(),
+        control_revision: observation.control_revision,
+        classification: observation.classification,
+        canonical_source_path: observation.canonical_source_path.clone(),
+        source_generation: observation.source_generation.clone(),
+        topology_generation: observation.topology_generation.clone(),
+        member_epochs_sha256: observation.member_epochs_sha256.clone(),
+        creator_sha256: creator_sha256.to_string(),
+        created_at_unix_nanos,
+        retain_until_unix_nanos,
+        source_bytes_base64: source_bytes_base64.to_string(),
+    };
+    export.validate()?;
+    Ok(export)
+}
+
+pub(crate) fn recovery_export_id(control_id: &str, generation: &IlmRecoverySourceGeneration) -> Result<String> {
+    validate_sha256(control_id, "ILM recovery export control ID is invalid")?;
+    validate_sha256(&generation.content_sha256, "ILM recovery export source checksum is invalid")?;
+    validate_sha256(&generation.copy_set_sha256, "ILM recovery export copy-set checksum is invalid")?;
+    let mut data = Vec::new();
+    for part in [control_id, &generation.content_sha256, &generation.copy_set_sha256] {
+        data.extend_from_slice(&(part.len() as u64).to_be_bytes());
+        data.extend_from_slice(part.as_bytes());
+    }
+    Ok(hex_sha256(&data, ToOwned::to_owned))
+}
+
+fn export_matches_observation(export: &IlmRecoveryExport, observation: &IlmRecoveryExportObservation) -> bool {
+    export.control_id == observation.control_id
+        && export.protocol == observation.protocol
+        && export.classification == observation.classification
+        && export.canonical_source_path == observation.canonical_source_path
+        && export.source_generation == observation.source_generation
+}
+
+#[derive(Debug, Default)]
+struct IlmRecoveryExportInventory {
+    count: usize,
+    bytes: u64,
+    creations: Vec<(i64, String)>,
+}
+
+impl IlmRecoveryExportInventory {
+    fn check(&self, creator_sha256: &str, candidate_len: usize, now: i64) -> Result<()> {
+        let recent_after = now.saturating_sub(60 * 1_000_000_000);
+        let cluster_recent = self
+            .creations
+            .iter()
+            .filter(|(created_at, _)| *created_at > recent_after)
+            .count();
+        let actor_recent = self
+            .creations
+            .iter()
+            .filter(|(created_at, creator)| *created_at > recent_after && creator == creator_sha256)
+            .count();
+        check_export_admission(self.count, self.bytes, actor_recent, cluster_recent, candidate_len)
+    }
+}
+
+async fn collect_export_inventory(api: Arc<ECStore>) -> Result<IlmRecoveryExportInventory> {
+    let mut marker = None;
+    let mut seen_markers = HashSet::new();
+    let mut inventory = IlmRecoveryExportInventory::default();
+    loop {
+        let page = api
+            .clone()
+            .list_objects_v2(
+                RUSTFS_META_BUCKET,
+                &format!("{ILM_RECOVERY_EXPORT_PREFIX}/"),
+                marker.clone(),
+                None,
+                1_000,
+                false,
+                None,
+                false,
+            )
+            .await?;
+        for object in page.objects {
+            let (_, export_id) = recovery_export_id_from_record_object_name(&object.name)?;
+            let (stored, export) = load_recovery_export_decoded(api.clone(), &export_id).await?;
+            inventory.count = inventory
+                .count
+                .checked_add(1)
+                .ok_or_else(|| Error::other("ILM recovery export count overflow"))?;
+            inventory.bytes = inventory
+                .bytes
+                .checked_add(u64::try_from(stored.encoded.len()).map_err(|_| Error::other("ILM recovery export size overflow"))?)
+                .ok_or_else(|| Error::other("ILM recovery export byte total overflow"))?;
+            inventory
+                .creations
+                .push((export.created_at_unix_nanos, export.creator_sha256));
+        }
+        if !page.is_truncated {
+            break;
+        }
+        let next = page
+            .next_continuation_token
+            .ok_or_else(|| Error::other("ILM recovery export inventory omitted its continuation marker"))?;
+        marker = Some(record_export_inventory_marker(&mut seen_markers, next)?);
+    }
+    Ok(inventory)
+}
+
+fn record_export_inventory_marker(seen_markers: &mut HashSet<String>, next: String) -> Result<String> {
+    if !seen_markers.insert(next.clone()) {
+        return Err(Error::other("ILM recovery export inventory repeated its continuation marker"));
+    }
+    Ok(next)
+}
+
+fn check_export_admission(
+    count: usize,
+    bytes: u64,
+    actor_recent: usize,
+    cluster_recent: usize,
+    candidate_len: usize,
+) -> Result<()> {
+    let candidate_len = u64::try_from(candidate_len).map_err(|_| Error::other("ILM recovery export size does not fit u64"))?;
+    if count >= MAX_ILM_RECOVERY_EXPORTS
+        || bytes
+            .checked_add(candidate_len)
+            .is_none_or(|total| total > MAX_ILM_RECOVERY_EXPORT_BYTES)
+        || actor_recent >= MAX_ACTOR_EXPORTS_PER_MINUTE
+        || cluster_recent >= MAX_CLUSTER_EXPORTS_PER_MINUTE
+    {
+        return Err(Error::SlowDown);
+    }
+    Ok(())
+}
+
+fn is_legacy_export_schema(schema: &str) -> bool {
+    matches!(schema, TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA | TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA)
+}
+
+fn validate_sha256(value: &str, message: &'static str) -> Result<()> {
+    if !is_sha256_checksum(value)
+        || value
+            .bytes()
+            .any(|byte| byte.is_ascii_hexdigit() && byte.is_ascii_uppercase())
+    {
+        return Err(Error::other(message));
+    }
+    Ok(())
+}
+
+fn now_unix_nanos() -> Result<i64> {
+    i64::try_from(time::OffsetDateTime::now_utc().unix_timestamp_nanos())
+        .map_err(|_| Error::other("ILM recovery export timestamp does not fit i64"))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::bucket::lifecycle::recovery_control::IlmRecoverySourceCopy;
+
+    const PINNED_V1_EXPORT: &[u8] = br#"{"schema":"rustfs-ilm-recovery-export-v1","content_sha256":"3dfb3ec3892256e909de1211c1a963ca7008963ff32b3a869f7161a7b9b44028","export":{"export_id":"2b78e7a825bfc2edbf7f773d0b6ed3bf93e360ff1702d73a449109c11bfaa105","control_id":"0fcd568a5cb9bdb4677b69354b11ee415af8f784519cff3da49a26f84eaee7f2","protocol":"tier_delete_journal","control_etag":"control-etag","control_revision":1,"classification":"retained_ambiguous","canonical_source_path":"ilm/tier-delete-journal/872072554f66ab326f10ce7adbae11422b7a4b0663aa7112d6061a8f6ed41b94.json","source_generation":{"source_schema":"rustfs-tier-delete-journal-v1","source_etag":"etag-a","content_sha256":"0e0b010ebdeeb7b41473fe8575e989d6bb1303c0ca551dd984e9400f0ae306bd","copy_set_sha256":"5a7406115b6c3923ffe79dcd1f43ccae7beed786e557163f019dd10ec409a653","copies":[{"authority":"pool-0/set-0","canonical_path":"ilm/tier-delete-journal/872072554f66ab326f10ce7adbae11422b7a4b0663aa7112d6061a8f6ed41b94.json","etag":"etag-a","encoded_len":81,"content_sha256":"0e0b010ebdeeb7b41473fe8575e989d6bb1303c0ca551dd984e9400f0ae306bd"}]},"topology_generation":"e6e2b826e31fca5c36125c48f130dcb6f961e698ff8a8776a1f290cf0892e8e6","member_epochs_sha256":"612dd8a861161819a4ad8f6f3e2a0567602877c043a2353ca933a13c78dc0ed4","creator_sha256":"50c9c4aeb40b5b206b6d98f516f8b8c0efd29ce2e56a76b345fb9240c225a1b7","created_at_unix_nanos":1000000000,"retain_until_unix_nanos":7776001000000000,"source_bytes_base64":"eyJ2ZXJzaW9uIjoxLCJvYmpfbmFtZSI6ImxlZ2FjeS9yZW1vdGUiLCJ2ZXJzaW9uX2lkIjoib3BhcXVlIiwidGllcl9uYW1lIjoiV0FSTSJ9"}}"#;
+
+    fn legacy_source() -> Vec<u8> {
+        br#"{"version":1,"obj_name":"legacy/remote","version_id":"opaque","tier_name":"WARM"}"#.to_vec()
+    }
+
+    fn observation() -> IlmRecoveryExportObservation {
+        let source = legacy_source();
+        let source_path = super::super::tier_delete_journal::tier_delete_journal_object_name(
+            &super::super::tier_delete_journal::decode_tier_delete_journal_entry(&source).expect("legacy fixture should decode"),
+        );
+        let source_sha256 = hex_sha256(&source, ToOwned::to_owned);
+        let generation = IlmRecoverySourceGeneration::new(
+            TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA,
+            "etag-a",
+            source_sha256.clone(),
+            vec![IlmRecoverySourceCopy {
+                authority: "pool-0/set-0".to_string(),
+                canonical_path: source_path.clone(),
+                etag: "etag-a".to_string(),
+                encoded_len: source.len() as u64,
+                content_sha256: source_sha256,
+            }],
+        )
+        .expect("generation should be valid");
+        IlmRecoveryExportObservation {
+            control_id: hex_sha256(b"control", ToOwned::to_owned),
+            protocol: IlmRecoveryProtocol::TierDeleteJournal,
+            control_etag: "control-etag".to_string(),
+            control_revision: 1,
+            classification: IlmRecoveryClassification::RetainedAmbiguous,
+            canonical_source_path: source_path,
+            source_generation: generation,
+            topology_generation: hex_sha256(b"topology", ToOwned::to_owned),
+            member_epochs_sha256: hex_sha256(b"epochs", ToOwned::to_owned),
+        }
+    }
+
+    #[test]
+    fn recovery_export_round_trip_is_strict_and_deterministic() {
+        let observed = observation();
+        let creator = hex_sha256(b"actor", ToOwned::to_owned);
+        let export = build_export_from_source(
+            &observed,
+            &creator,
+            1_000_000_000,
+            &base64_simd::STANDARD.encode_to_string(legacy_source()),
+        )
+        .expect("export should be valid");
+        assert_eq!(
+            export.export_id,
+            recovery_export_id(&observed.control_id, &observed.source_generation).unwrap()
+        );
+        let encoded = export.encode().expect("export should encode");
+        assert_eq!(encoded, PINNED_V1_EXPORT, "v1 export wire format must remain pinned");
+        assert_eq!(IlmRecoveryExport::decode(&export.export_id, &encoded).unwrap(), export);
+        assert_eq!(
+            IlmRecoveryExport::decode("2b78e7a825bfc2edbf7f773d0b6ed3bf93e360ff1702d73a449109c11bfaa105", PINNED_V1_EXPORT)
+                .unwrap(),
+            export,
+        );
+
+        let path = recovery_export_record_object_name(export.protocol, &export.export_id).unwrap();
+        let durable = super::super::durable_namespace::validate_durable_ilm_record(&path, &encoded)
+            .expect("export should be registered as a durable ILM record");
+        assert_eq!(durable.namespace, "recovery-export");
+        assert_eq!(durable.id_kind, "export_id");
+        assert_eq!(durable.id, export.export_id);
+
+        let mut wrong_source = export.clone();
+        wrong_source.source_bytes_base64 = base64_simd::STANDARD.encode_to_string(b"changed");
+        assert!(wrong_source.encode().is_err());
+
+        let mut persisted: serde_json::Value = serde_json::from_slice(&encoded).unwrap();
+        persisted["unknown"] = serde_json::json!(true);
+        assert!(IlmRecoveryExport::decode(&export.export_id, &serde_json::to_vec(&persisted).unwrap()).is_err());
+    }
+
+    #[test]
+    fn export_inventory_rejects_non_adjacent_continuation_cycles() {
+        let mut seen = HashSet::new();
+        assert_eq!(record_export_inventory_marker(&mut seen, "a".to_string()).unwrap(), "a");
+        assert_eq!(record_export_inventory_marker(&mut seen, "b".to_string()).unwrap(), "b");
+        record_export_inventory_marker(&mut seen, "a".to_string())
+            .expect_err("a non-adjacent continuation marker cycle must fail closed");
+    }
+
+    #[test]
+    fn recovery_export_path_rejects_noncanonical_shards() {
+        let id = hex_sha256(b"export", ToOwned::to_owned);
+        let path = recovery_export_record_object_name(IlmRecoveryProtocol::TierDeleteJournal, &id).unwrap();
+        assert_eq!(recovery_export_id_from_record_object_name(&path).unwrap().1, id);
+        let wrong_shard = path.replacen(&format!("/{}/", &id[..2]), "/zz/", 1);
+        assert!(recovery_export_id_from_record_object_name(&wrong_shard).is_err());
+    }
+
+    #[test]
+    fn canonical_replay_survives_fleet_rotation_but_not_source_change() {
+        let observed = observation();
+        let creator = hex_sha256(b"actor", ToOwned::to_owned);
+        let export = build_export_from_source(
+            &observed,
+            &creator,
+            1_000_000_000,
+            &base64_simd::STANDARD.encode_to_string(legacy_source()),
+        )
+        .unwrap();
+        let mut rotated = observed;
+        rotated.control_etag = "new-control-etag".to_string();
+        rotated.control_revision += 1;
+        rotated.topology_generation = hex_sha256(b"new-topology", ToOwned::to_owned);
+        rotated.member_epochs_sha256 = hex_sha256(b"new-members", ToOwned::to_owned);
+        assert!(export_matches_observation(&export, &rotated));
+
+        rotated.source_generation.content_sha256 = hex_sha256(b"changed", ToOwned::to_owned);
+        assert!(!export_matches_observation(&export, &rotated));
+    }
+
+    #[test]
+    fn export_admission_enforces_exact_count_byte_and_rate_boundaries() {
+        assert!(check_export_admission(9_999, MAX_ILM_RECOVERY_EXPORT_BYTES - 1, 9, 99, 1).is_ok());
+        assert!(check_export_admission(10_000, 0, 0, 0, 1).is_err());
+        assert!(check_export_admission(0, MAX_ILM_RECOVERY_EXPORT_BYTES, 0, 0, 1).is_err());
+        assert!(check_export_admission(0, 0, 10, 0, 1).is_err());
+        assert!(check_export_admission(0, 0, 0, 100, 1).is_err());
+    }
+}
@@ -82,8 +82,8 @@ const TIER_DELETE_DISPATCH_MEMBER_DELETE_CONCURRENCY: usize = 32;
 const TIER_DELETE_DISPATCH_PREPARE_CONCURRENCY: usize = 16;
 const TIER_DELETE_DISPATCH_CAS_CONCURRENCY: usize = 32;
 const TIER_DELETE_JOURNAL_VERSION: u8 = 2;
-const TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA: &str = "rustfs-tier-delete-journal-v1";
-const TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA: &str = "rustfs-tier-delete-journal-v2";
+pub(crate) const TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA: &str = "rustfs-tier-delete-journal-v1";
+pub(crate) const TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA: &str = "rustfs-tier-delete-journal-v2";
 const TIER_DELETE_JOURNAL_UNKNOWN_RECOVERY_SCHEMA: &str = "rustfs-tier-delete-journal-unknown";
 const TIER_DELETE_JOURNAL_V1_RECOVERY_CLASS: &str = "tier_delete_journal_v1";
 const TIER_DELETE_JOURNAL_V2_RECOVERY_CLASS: &str = "tier_delete_journal_v2";
@@ -884,6 +884,23 @@ struct PersistedTierDeleteJournalEntry {
 }

 impl PersistedTierDeleteJournalEntry {
+    fn validate_legacy_recovery_shape(&self) -> Result<()> {
+        let has_later_version_fields = self.version_id_exact.is_some()
+            || self.version_state.is_some()
+            || self.state.is_some()
+            || self.source.is_some()
+            || self.dispatch.is_some();
+        match self.version {
+            1 if self.backend_identity.is_none() && !has_later_version_fields => Ok(()),
+            TIER_DELETE_JOURNAL_VERSION if self.backend_identity.is_some() && !has_later_version_fields => Ok(()),
+            1 => Err(Error::other("tier delete journal v1 entry contains fields from a later version")),
+            TIER_DELETE_JOURNAL_VERSION => Err(Error::other(
+                "tier delete journal v2 entry is missing its identity or contains fields from a later version",
+            )),
+            _ => Err(Error::other("tier delete journal is not an exportable legacy version")),
+        }
+    }
+
    fn from_jentry(je: &Jentry) -> Result<Self> {
        validate_version_state(je.version_state, &je.version_id, je.version_id_exact)?;
        let legacy_unknown = je.version_state == rustfs_filemeta::TransitionVersionState::Unknown;
@@ -5531,6 +5548,12 @@ fn canonical_legacy_tier_delete_journal_identity(object_name: &str) -> Option<&s
    .then_some(identity)
 }

+pub(crate) fn validate_legacy_tier_delete_recovery_path(object_name: &str) -> Result<()> {
+    canonical_legacy_tier_delete_journal_identity(object_name)
+        .map(|_| ())
+        .ok_or_else(|| Error::other("legacy tier delete journal path is not canonical"))
+}
+
 fn legacy_tier_delete_recovery_descriptor(entry: &Jentry) -> Option<(&'static str, &'static str)> {
    match entry.persisted_version {
        1 => Some((TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA, TIER_DELETE_JOURNAL_V1_RECOVERY_CLASS)),
@@ -5539,6 +5562,21 @@ fn legacy_tier_delete_recovery_descriptor(entry: &Jentry) -> Option<(&'static st
    }
 }

+pub(crate) fn validate_legacy_tier_delete_recovery_source(object_name: &str, source_schema: &str, data: &[u8]) -> Result<()> {
+    validate_legacy_tier_delete_recovery_path(object_name)?;
+    let persisted: PersistedTierDeleteJournalEntry =
+        serde_json::from_slice(data).map_err(|err| Error::other_with_context("decode tier delete journal failed", err))?;
+    persisted.validate_legacy_recovery_shape()?;
+    let entry = persisted.into_jentry()?;
+    let Some((decoded_schema, _)) = legacy_tier_delete_recovery_descriptor(&entry) else {
+        return Err(Error::other("tier delete journal is not an exportable legacy version"));
+    };
+    if decoded_schema != source_schema || tier_delete_journal_object_name(&entry) != object_name {
+        return Err(Error::other("legacy tier delete journal identity does not match its recovery source"));
+    }
+    Ok(())
+}
+
 fn legacy_tier_delete_control_matches(
    control: &IlmRecoveryControl,
    identity: &IlmRecoveryControlIdentity,
@@ -6140,17 +6178,18 @@ where
 #[cfg(test)]
 mod tests {
    use super::{
-        TIER_DELETE_DISPATCH_MANIFEST_VERSION, TIER_DELETE_DISPATCH_PARENT_RECORD_TYPE, TIER_DELETE_DISPATCH_PARENT_VERSION,
-        TIER_DELETE_JOURNAL_EXACT_VERSION, TIER_DELETE_JOURNAL_LEGACY_PREFIX, TIER_DELETE_JOURNAL_SOLE_OWNER_VERSION,
-        TIER_DELETE_JOURNAL_STATE_VERSION, TIER_DELETE_JOURNAL_TRANSACTION_VERSION, TIER_DELETE_JOURNAL_V6_PREFIX,
-        TierDeleteDispatchChunkBinding, TierDeleteDispatchManifest, TierDeleteDispatchManifestState, TierDeleteDispatchParent,
-        TierDeleteDispatchParentState, TierDeleteDispatchRecord, await_tier_delete_journal_recovery,
+        PersistedTierDeleteJournalEntry, TIER_DELETE_DISPATCH_MANIFEST_VERSION, TIER_DELETE_DISPATCH_PARENT_RECORD_TYPE,
+        TIER_DELETE_DISPATCH_PARENT_VERSION, TIER_DELETE_JOURNAL_EXACT_VERSION, TIER_DELETE_JOURNAL_LEGACY_PREFIX,
+        TIER_DELETE_JOURNAL_SOLE_OWNER_VERSION, TIER_DELETE_JOURNAL_STATE_VERSION, TIER_DELETE_JOURNAL_TRANSACTION_VERSION,
+        TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA, TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA, TIER_DELETE_JOURNAL_V6_PREFIX,
+        TIER_DELETE_JOURNAL_VERSION, TierDeleteDispatchChunkBinding, TierDeleteDispatchManifest, TierDeleteDispatchManifestState,
+        TierDeleteDispatchParent, TierDeleteDispatchParentState, TierDeleteDispatchRecord, await_tier_delete_journal_recovery,
        decode_tier_delete_dispatch_record, decode_tier_delete_journal_entry, encode_tier_delete_dispatch_manifest,
        encode_tier_delete_dispatch_parent, encode_tier_delete_journal_entry, object_info_references_tier_delete,
        record_tier_delete_journal_backend_identity, same_tier_delete_authorization_identity, same_tier_delete_journal_identity,
        tier_delete_dispatch_child_matches_parent, tier_delete_dispatch_chunk_manifest_object_name,
        tier_delete_dispatch_journal_set_digest, tier_delete_dispatch_manifest_object_name, tier_delete_journal_object_name,
-        tier_delete_source_matches_dispatch_scope,
+        tier_delete_source_matches_dispatch_scope, validate_legacy_tier_delete_recovery_source,
    };
    use crate::bucket::lifecycle::tier_sweeper::{
        Jentry, TierDeleteDispatchBinding, TierDeleteJournalState, TierDeleteSourceIdentity,
@@ -6609,6 +6648,72 @@ mod tests {
        }
    }

+    #[test]
+    fn legacy_recovery_export_rejects_fields_from_later_journal_versions() {
+        let later = bound_v6_journal_entry(TierDeleteJournalState::Prepared);
+        let v1 = PersistedTierDeleteJournalEntry {
+            version: 1,
+            obj_name: "remote/object".to_string(),
+            version_id: "opaque".to_string(),
+            tier_name: "WARM".to_string(),
+            backend_identity: None,
+            version_id_exact: None,
+            version_state: None,
+            state: None,
+            source: None,
+            dispatch: None,
+        };
+        let mut v2 = v1.clone();
+        v2.version = TIER_DELETE_JOURNAL_VERSION;
+        v2.backend_identity = Some([7; 32]);
+
+        let assert_rejected = |persisted: PersistedTierDeleteJournalEntry, schema: &str| {
+            let normalized = persisted
+                .clone()
+                .into_jentry()
+                .expect("the generic compatibility decoder should demonstrate the discarded field");
+            let object_name = tier_delete_journal_object_name(&normalized);
+            let encoded = serde_json::to_vec(&persisted).expect("mixed-version journal fixture should encode");
+            let err = validate_legacy_tier_delete_recovery_source(&object_name, schema, &encoded)
+                .expect_err("legacy recovery export must reject fields from later versions");
+            assert!(err.to_string().contains("later version"));
+        };
+
+        let mut invalid_v1 = Vec::new();
+        let mut with_backend = v1.clone();
+        with_backend.backend_identity = Some([7; 32]);
+        invalid_v1.push(with_backend);
+        for persisted in [&v1, &v2] {
+            let schema = if persisted.version == 1 {
+                TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA
+            } else {
+                TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA
+            };
+            let mut invalid = Vec::new();
+            let mut with_exact = persisted.clone();
+            with_exact.version_id_exact = Some(false);
+            invalid.push(with_exact);
+            let mut with_version_state = persisted.clone();
+            with_version_state.version_state = Some(rustfs_filemeta::TransitionVersionState::Unknown);
+            invalid.push(with_version_state);
+            let mut with_state = persisted.clone();
+            with_state.state = Some(TierDeleteJournalState::Committed);
+            invalid.push(with_state);
+            let mut with_source = persisted.clone();
+            with_source.source = later.source.clone();
+            invalid.push(with_source);
+            let mut with_dispatch = persisted.clone();
+            with_dispatch.dispatch = later.dispatch.clone();
+            invalid.push(with_dispatch);
+            for record in invalid {
+                assert_rejected(record, schema);
+            }
+        }
+        for record in invalid_v1 {
+            assert_rejected(record, TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA);
+        }
+    }
+
    #[test]
    fn tier_delete_journal_path_is_stable_and_sanitized() {
        let je = journal_entry();
@@ -34,7 +34,7 @@ use crate::bucket::lifecycle::tier_sweeper::{
 };
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::{Error, Result as EcstoreResult};
-use crate::object_api::ObjectOptions;
+use crate::object_api::{ObjectInfo, ObjectOptions};
 use crate::services::tier::{tier::TierConfigMgr, warm_backend::TransitionCandidateProbe};
 use crate::storage_api_contracts::{
    list::ListOperations as _,
@@ -273,6 +273,14 @@ pub struct TransitionTransactionInit {

 impl TransitionTransaction {
    pub fn new(init: TransitionTransactionInit) -> Result<Self> {
+        Self::new_with_initial_state(init, TransitionTransactionState::UploadStarted)
+    }
+
+    pub(crate) fn new_compact(init: TransitionTransactionInit) -> Result<Self> {
+        Self::new_with_initial_state(init, TransitionTransactionState::UploadOutcomeUnknown)
+    }
+
+    fn new_with_initial_state(init: TransitionTransactionInit, state: TransitionTransactionState) -> Result<Self> {
        let remote_object =
            canonical_transition_remote_object(init.deployment_id, &init.source.bucket, init.transaction_id, init.write_id)?;
        let transaction = Self {
@@ -286,7 +294,7 @@ impl TransitionTransaction {
            backend_fingerprint: init.backend_fingerprint,
            remote_object,
            remote_version: TransitionRemoteVersion::unknown(),
-            state: TransitionTransactionState::UploadStarted,
+            state,
            not_after_unix_nanos: init.not_after_unix_nanos,
        };
        transaction.validate()?;
@@ -356,7 +364,7 @@ impl TransitionTransaction {
        remote_version: Option<TransitionRemoteVersion>,
    ) -> Result<TransitionTransactionFence> {
        self.check_fence(fence)?;
-        if !state_change_allowed(self.state, next) {
+        if !state_change_allowed_at(self.state, next, self.revision) {
            return Err(TransitionTransactionError::InvalidStateChange {
                from: self.state,
                to: next,
@@ -388,6 +396,14 @@ impl TransitionTransaction {
                }
                self.remote_version = TransitionRemoteVersion::unknown();
            }
+            TransitionTransactionState::LocalCommitStarted if self.state == TransitionTransactionState::UploadOutcomeUnknown => {
+                let remote_version =
+                    remote_version.ok_or(TransitionTransactionError::Corrupt("compact local commit requires remote version"))?;
+                if remote_version.is_unknown() {
+                    return Err(TransitionTransactionError::Corrupt("compact local commit requires known remote version"));
+                }
+                self.remote_version = remote_version;
+            }
            TransitionTransactionState::LocalCommitStarted | TransitionTransactionState::Committed => {
                if let Some(remote_version) = remote_version
                    && remote_version != self.remote_version
@@ -640,7 +656,7 @@ pub(crate) async fn save_transition_transaction_record_if_current(
 ) -> EcstoreResult<()> {
    let object = transition_transaction_record_object_name(next.transaction_id).map_err(transition_transaction_store_error)?;
    let revision_is_next = expected.revision.checked_add(1) == Some(next.revision);
-    let state_is_next = state_change_allowed(expected.state, next.state)
+    let state_is_next = state_change_allowed_at(expected.state, next.state, expected.revision)
        || matches!(
            (expected.state, next.state),
            (
@@ -954,6 +970,10 @@ pub enum TransitionOperatorError {
        expected: String,
        actual: TransitionOperatorProbe,
    },
+    #[error("transition recovery control is stale")]
+    StaleRecoveryControl,
+    #[error("transition recovery control is not eligible for operator retry")]
+    RetryNotAllowed,
    #[error("transition transaction store failed: {0}")]
    Store(#[source] Error),
    #[error("remote tier reconciliation failed: {0}")]
@@ -962,6 +982,179 @@ pub enum TransitionOperatorError {

 type TransitionOperatorResult<T> = std::result::Result<T, TransitionOperatorError>;

+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+pub struct TransitionRecoveryRetryStatus {
+    pub control_id: String,
+    pub transaction_id: Uuid,
+    pub state: TransitionTransactionState,
+    pub classification: IlmRecoveryClassification,
+    pub control_revision: u64,
+    pub attempt_count: u64,
+    pub consecutive_failure_count: u32,
+    pub last_error_code: IlmRecoveryErrorCode,
+    pub source_generation_sha256: String,
+    pub copy_set_sha256: String,
+    pub retry_ready: bool,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub retry_not_ready_reason: Option<&'static str>,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+pub struct TransitionRecoveryRetryResult {
+    pub control_id: String,
+    pub transaction_id: Uuid,
+    pub previous_revision: u64,
+    pub revision: u64,
+    pub classification: IlmRecoveryClassification,
+    pub attempt_count: u64,
+    pub source_generation_sha256: String,
+}
+
+struct TransitionRecoveryRetryContext {
+    observed: ObservedIlmRecoveryControl,
+    transaction: TransitionTransaction,
+    source_generation_sha256: String,
+}
+
+fn transition_recovery_retry_readiness(control: &IlmRecoveryControl) -> (bool, Option<&'static str>) {
+    if control.owner.is_some() {
+        return (false, Some("attempt_owned"));
+    }
+    match control.classification {
+        IlmRecoveryClassification::RetainedAmbiguous | IlmRecoveryClassification::OperatorRequired => (true, None),
+        IlmRecoveryClassification::Retrying => (false, Some("already_retrying")),
+        IlmRecoveryClassification::Corrupt => (false, Some("source_corrupt")),
+        IlmRecoveryClassification::Abandoned => (false, Some("source_abandoned")),
+        IlmRecoveryClassification::Terminal => (false, Some("source_terminal")),
+    }
+}
+
+async fn load_transition_recovery_retry_context(
+    api: Arc<ECStore>,
+    control_id: &str,
+) -> TransitionOperatorResult<TransitionRecoveryRetryContext> {
+    let observed = match load_recovery_control(api.clone(), IlmRecoveryProtocol::TransitionTransaction, control_id).await {
+        Ok(observed) => observed,
+        Err(Error::ConfigNotFound) => return Err(TransitionOperatorError::NotFound),
+        Err(err) => return Err(TransitionOperatorError::Store(err)),
+    };
+    let transaction_id = Uuid::parse_str(&observed.control.identity.stable_operation_identity)
+        .ok()
+        .filter(|transaction_id| !transaction_id.is_nil())
+        .ok_or(TransitionOperatorError::StaleRecoveryControl)?;
+    let canonical_path = transition_transaction_record_object_name(transaction_id)
+        .map_err(|err| TransitionOperatorError::Store(Error::other(err)))?;
+    if observed.control.identity.canonical_source_path != canonical_path
+        || observed.control.identity.record_class != "transition_transaction_v1"
+    {
+        return Err(TransitionOperatorError::StaleRecoveryControl);
+    }
+    let transaction = match load_transition_transaction_record(api.clone(), transaction_id).await {
+        Ok(transaction) => transaction,
+        Err(Error::ConfigNotFound) => return Err(TransitionOperatorError::NotFound),
+        Err(err) => return Err(TransitionOperatorError::Store(err)),
+    };
+    let source = observe_recovery_source(api, &canonical_path, TRANSITION_TRANSACTION_SCHEMA)
+        .await
+        .map_err(TransitionOperatorError::Store)?;
+    let exact_source = source.is_consistent()
+        && source.generation == observed.control.observed_source_generation
+        && source
+            .canonical_data
+            .as_deref()
+            .is_some_and(|data| TransitionTransaction::decode(transaction_id, data).is_ok_and(|decoded| decoded == transaction));
+    if !exact_source {
+        return Err(TransitionOperatorError::StaleRecoveryControl);
+    }
+    let generation = serde_json::to_vec(&observed.control.observed_source_generation)
+        .map_err(|err| TransitionOperatorError::Store(Error::other(err)))?;
+    Ok(TransitionRecoveryRetryContext {
+        observed,
+        transaction,
+        source_generation_sha256: hex_sha256(&generation, ToOwned::to_owned),
+    })
+}
+
+pub async fn inspect_transition_recovery_retry_for_operator(
+    api: Arc<ECStore>,
+    control_id: &str,
+) -> TransitionOperatorResult<TransitionRecoveryRetryStatus> {
+    let context = load_transition_recovery_retry_context(api, control_id).await?;
+    let (retry_ready, retry_not_ready_reason) = transition_recovery_retry_readiness(&context.observed.control);
+    Ok(TransitionRecoveryRetryStatus {
+        control_id: control_id.to_string(),
+        transaction_id: context.transaction.transaction_id,
+        state: context.transaction.state,
+        classification: context.observed.control.classification,
+        control_revision: context.observed.control.revision,
+        attempt_count: context.observed.control.attempt_count,
+        consecutive_failure_count: context.observed.control.consecutive_failure_count,
+        last_error_code: context.observed.control.last_error_code,
+        source_generation_sha256: context.source_generation_sha256,
+        copy_set_sha256: context.observed.control.observed_source_generation.copy_set_sha256.clone(),
+        retry_ready,
+        retry_not_ready_reason,
+    })
+}
+
+pub async fn retry_transition_recovery_for_operator(
+    api: Arc<ECStore>,
+    control_id: &str,
+    expected_control_revision: u64,
+    expected_source_generation_sha256: &str,
+) -> TransitionOperatorResult<TransitionRecoveryRetryResult> {
+    let control_object = recovery_control_record_object_name(IlmRecoveryProtocol::TransitionTransaction, control_id)
+        .map_err(|err| TransitionOperatorError::Store(Error::other(err)))?;
+    let retry_lock = api
+        .new_ns_lock(RUSTFS_META_BUCKET, &format!("{control_object}.recovery-lock"))
+        .await
+        .map_err(TransitionOperatorError::Store)?;
+    let retry_guard = retry_lock
+        .get_write_lock(crate::set_disk::get_lock_acquire_timeout())
+        .await
+        .map_err(|err| TransitionOperatorError::Store(Error::other(err)))?;
+    let context = load_transition_recovery_retry_context(api.clone(), control_id).await?;
+    let (retry_ready, _) = transition_recovery_retry_readiness(&context.observed.control);
+    if !retry_ready {
+        return Err(TransitionOperatorError::RetryNotAllowed);
+    }
+    if retry_guard.is_lock_lost()
+        || expected_control_revision == 0
+        || context.observed.control.revision != expected_control_revision
+        || context.source_generation_sha256 != expected_source_generation_sha256
+    {
+        return Err(TransitionOperatorError::StaleRecoveryControl);
+    }
+    let previous_revision = context.observed.control.revision;
+    let mut next = context.observed.control.clone();
+    next.retry_for_operator(&context.observed.control.observed_source_generation)
+        .map_err(|_| TransitionOperatorError::RetryNotAllowed)?;
+    if retry_guard.is_lock_lost() {
+        return Err(TransitionOperatorError::StaleRecoveryControl);
+    }
+    save_recovery_control_if_current(api.clone(), &context.observed, &next)
+        .await
+        .map_err(|err| match err {
+            Error::PreconditionFailed => TransitionOperatorError::StaleRecoveryControl,
+            err => TransitionOperatorError::Store(err),
+        })?;
+    let persisted = load_recovery_control(api, IlmRecoveryProtocol::TransitionTransaction, control_id)
+        .await
+        .map_err(TransitionOperatorError::Store)?;
+    if retry_guard.is_lock_lost() || persisted.control != next {
+        return Err(TransitionOperatorError::StaleRecoveryControl);
+    }
+    Ok(TransitionRecoveryRetryResult {
+        control_id: control_id.to_string(),
+        transaction_id: context.transaction.transaction_id,
+        previous_revision,
+        revision: persisted.control.revision,
+        classification: persisted.control.classification,
+        attempt_count: persisted.control.attempt_count,
+        source_generation_sha256: context.source_generation_sha256,
+    })
+}
+
 fn validate_operator_reconcile_transaction(
    transaction: &TransitionTransaction,
    now_unix_nanos: i128,
@@ -1706,12 +1899,27 @@ async fn local_commit_matches_transaction(api: Arc<ECStore>, transaction: &Trans
        .get_object_info(&transaction.source.bucket, &transaction.source.object, &opts)
        .await?;
    let transitioned = &object.transitioned_object;
-    Ok(transitioned.status == TRANSITION_COMPLETE
+    Ok(local_object_matches_transition_source(&object, &transaction.source)
+        && transitioned.status == TRANSITION_COMPLETE
        && transitioned.name == transaction.remote_object
        && transitioned.tier == transaction.tier_name
        && transitioned.version_id == transaction.remote_version.tier_delete_version_id().unwrap_or_default())
 }

+fn local_object_matches_transition_source(object: &ObjectInfo, source: &TransitionSourceIdentity) -> bool {
+    let observed_version_id = object.version_id.filter(|version_id| !version_id.is_nil());
+    let observed_mod_time = object
+        .mod_time
+        .and_then(|mod_time| i64::try_from(mod_time.unix_timestamp_nanos()).ok());
+    object.bucket == source.bucket
+        && object.name == source.object
+        && observed_version_id == source.version_id
+        && object.data_dir == Some(source.data_dir)
+        && observed_mod_time == Some(source.mod_time_unix_nanos)
+        && object.size == source.size
+        && object.etag.as_deref() == Some(source.etag.as_str())
+}
+
 fn transition_source_lookup_options(transaction: &TransitionTransaction) -> ObjectOptions {
    ObjectOptions {
        version_id: match transaction.source.version_mode {
@@ -1954,7 +2162,7 @@ where
    }
 }

-fn state_change_allowed(from: TransitionTransactionState, to: TransitionTransactionState) -> bool {
+fn state_change_allowed_at(from: TransitionTransactionState, to: TransitionTransactionState, revision: u64) -> bool {
    matches!(
        (from, to),
        (TransitionTransactionState::UploadStarted, TransitionTransactionState::Uploaded)
@@ -1966,7 +2174,9 @@ fn state_change_allowed(from: TransitionTransactionState, to: TransitionTransact
            | (TransitionTransactionState::UploadOutcomeUnknown, TransitionTransactionState::Uploaded)
            | (TransitionTransactionState::Uploaded, TransitionTransactionState::LocalCommitStarted)
            | (TransitionTransactionState::LocalCommitStarted, TransitionTransactionState::Committed)
-    )
+    ) || (revision == 1
+        && from == TransitionTransactionState::UploadOutcomeUnknown
+        && to == TransitionTransactionState::LocalCommitStarted)
 }

 fn state_requires_known_remote_version(state: TransitionTransactionState) -> bool {
@@ -2183,6 +2393,41 @@ mod tests {
        }
    }

+    #[test]
+    fn local_commit_proof_requires_the_complete_source_identity() {
+        let source = source_identity(TransitionSourceVersionMode::Versioned);
+        let exact = ObjectInfo {
+            bucket: source.bucket.clone(),
+            name: source.object.clone(),
+            version_id: source.version_id,
+            data_dir: Some(source.data_dir),
+            mod_time: Some(
+                time::OffsetDateTime::from_unix_timestamp_nanos(i128::from(source.mod_time_unix_nanos))
+                    .expect("source timestamp should be valid"),
+            ),
+            size: source.size,
+            etag: Some(source.etag.clone()),
+            ..Default::default()
+        };
+        assert!(local_object_matches_transition_source(&exact, &source));
+
+        let mut changed = exact.clone();
+        changed.version_id = Some(Uuid::new_v4());
+        assert!(!local_object_matches_transition_source(&changed, &source));
+        changed = exact.clone();
+        changed.data_dir = Some(Uuid::new_v4());
+        assert!(!local_object_matches_transition_source(&changed, &source));
+        changed = exact.clone();
+        changed.mod_time = changed.mod_time.map(|value| value + Duration::from_nanos(1));
+        assert!(!local_object_matches_transition_source(&changed, &source));
+        changed = exact.clone();
+        changed.size += 1;
+        assert!(!local_object_matches_transition_source(&changed, &source));
+        changed = exact;
+        changed.etag = Some("different-etag".to_string());
+        assert!(!local_object_matches_transition_source(&changed, &source));
+    }
+
    fn cleanup_proof(transaction: &TransitionTransaction, decision: TransitionCleanupDecision) -> TransitionCleanupProof {
        TransitionCleanupProof {
            transaction_id: transaction.transaction_id,
@@ -2384,6 +2629,57 @@ mod tests {
        assert_eq!(transaction.state, TransitionTransactionState::Committed);
    }

+    #[test]
+    fn compact_state_sequence_is_distinguishable_and_keeps_legacy_edges_strict() {
+        let init = TransitionTransactionInit {
+            deployment_id: Uuid::new_v4(),
+            transaction_id: Uuid::new_v4(),
+            owner_epoch: Uuid::new_v4(),
+            write_id: Uuid::new_v4(),
+            source: source_identity(TransitionSourceVersionMode::Versioned),
+            tier_name: "warm-tier".to_string(),
+            backend_fingerprint: BACKEND_FINGERPRINT,
+            not_after_unix_nanos: 1_780_000_000_000_000_000,
+        };
+        let mut compact = TransitionTransaction::new_compact(init).expect("compact transaction should be created");
+        assert_eq!(compact.state, TransitionTransactionState::UploadOutcomeUnknown);
+        assert_eq!(compact.revision, 1);
+        let remote_version = TransitionRemoteVersion::versioned(Uuid::new_v4().to_string());
+        let fence = compact
+            .advance(
+                compact.fence(),
+                TransitionTransactionState::LocalCommitStarted,
+                Some(remote_version.clone()),
+            )
+            .expect("compact upload should persist its exact candidate at the local commit fence");
+        assert_eq!(fence.revision, 2);
+        assert_eq!(compact.remote_version, remote_version);
+        assert_eq!(compact.state, TransitionTransactionState::LocalCommitStarted);
+        let encoded = compact
+            .encode()
+            .expect("compact transaction should encode as v1-compatible bytes");
+        assert_eq!(
+            TransitionTransaction::decode(compact.transaction_id, &encoded).expect("compact transaction should decode"),
+            compact
+        );
+
+        let mut legacy_unknown = new_transaction();
+        legacy_unknown
+            .advance(legacy_unknown.fence(), TransitionTransactionState::UploadOutcomeUnknown, None)
+            .expect("legacy transaction should persist its pre-upload fence");
+        assert!(matches!(
+            legacy_unknown.advance(
+                legacy_unknown.fence(),
+                TransitionTransactionState::LocalCommitStarted,
+                Some(TransitionRemoteVersion::unversioned()),
+            ),
+            Err(TransitionTransactionError::InvalidStateChange {
+                from: TransitionTransactionState::UploadOutcomeUnknown,
+                to: TransitionTransactionState::LocalCommitStarted,
+            })
+        ));
+    }
+
    #[test]
    fn cleanup_pending_requires_exact_proof_and_state_specific_decision() {
        let mut transaction = new_transaction();
@@ -66,6 +66,7 @@ pub(crate) use replication_lifecycle_bridge::ReplicationLifecycleBridge;
 pub(crate) use replication_migration_bridge::ReplicationMigrationBridge;
 pub use replication_object_bridge::ReplicationObjectBridge;
 pub use replication_object_config::{DeleteReplicationConfigSnapshot, ReplicationConfig};
+pub(crate) use replication_object_decision_boundary::replication_etags_match;
 pub use replication_object_decision_boundary::{
    MustReplicateOptions, ReplicationDeleteScheduleInput, ReplicationDeleteStateSource, delete_replication_state_from_config,
    delete_replication_version_id, should_schedule_delete_replication, should_use_existing_delete_replication_info,
@@ -88,5 +89,6 @@ pub use replication_state::{ReplicationStats, RuntimeReplicationTargetBacklog};
 pub use replication_stats_boundary::{BucketReplicationStat, BucketReplicationStats, BucketStats, InQueueMetric, XferStats};
 pub use replication_storage_boundary::{ReplicationObjectIO, ReplicationStorage};
 pub use replication_target_boundary::SsecPassthroughCapability;
+pub use replication_target_boundary::VersionIdentityCapability;
 pub use replication_target_boundary::{ObjectLockIntegrity, object_lock_put_integrity};
 pub(crate) use replication_target_config_bridge::ReplicationTargetConfigBridge;
@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

+#[cfg(test)]
+pub(crate) use rustfs_filemeta::ObjectPartInfo;
 pub use rustfs_replication::{MrfOpKind, MrfReplicateEntry};
 pub(crate) use rustfs_replication::{
    REPLICATE_EXISTING, REPLICATE_HEAL_DELETE, ReplicateTargetDecision, ReplicatedInfos, ReplicatedTargetInfo, ReplicationAction,
@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

+#[cfg(test)]
+pub(crate) use rustfs_replication::ReplicationMultipartPlanError;
 pub use rustfs_replication::{
    MustReplicateOptions, ReplicationDeleteScheduleInput, ReplicationDeleteStateSource, delete_replication_state_from_config,
    delete_replication_version_id, should_schedule_delete_replication, should_use_existing_delete_replication_info,
@@ -56,6 +56,8 @@ use super::replication_storage_boundary::{
 };
 #[cfg(test)]
 use super::replication_storage_boundary::{NamespaceLockFence, NamespaceLockSignalTestFence, ReplicationDeletedObject};
+#[cfg(test)]
+use super::replication_target_boundary::VersionIdentityCapability;
 use super::replication_target_boundary::{
    ERR_REPLICATION_SSEC_PASSTHROUGH_UNSUPPORTED, HeadObjectSdkError, PutObjectOptions, PutObjectPartOptions,
    RemotePutObjectResponse, ReplicationTargetStore, S3ClientError, SsecPassthroughCapability, SsecPassthroughGate, TargetClient,
@@ -63,7 +65,7 @@ use super::replication_target_boundary::{
    replication_delete_marker_purge_remove_options, replication_delete_remove_options, replication_force_delete_remove_options,
    replication_object_is_ssec_encrypted, replication_put_object_header_size, replication_put_object_options,
    replication_target_head_is_newer_null_version, resolve_read_api_version_id, ssec_passthrough_evidence_present,
-    ssec_passthrough_gate, version_identity_drifted,
+    ssec_passthrough_gate, version_identity_capability_from_put, version_identity_drifted,
 };
 use super::replication_versioning_boundary::ReplicationVersioningStore;
 use super::runtime_boundary as runtime_sources;
@@ -123,6 +125,7 @@ const EVENT_DELETE_MARKER_PURGE_FAILED: &str = "replication_delete_marker_purge_
 const EVENT_DELETE_MARKER_PURGE_MRF: &str = "replication_delete_marker_purge_mrf";
 const METRIC_DELETE_MARKER_PURGE_TOTAL: &str = "rustfs_replication_delete_marker_purge_total";
 const EVENT_REPLICATION_VERSION_IDENTITY_DRIFT: &str = "replication_version_identity_drift";
+const EVENT_REPLICATION_DRIFTED_REPLICA_LOCATED: &str = "replication_drifted_replica_located";
 const EVENT_REPLICATION_OBJECT_FAILED: &str = "replication_object_failed";
 const EVENT_REPLICATION_PURGE_OBJECT_LOCK_DENIED: &str = "replication_purge_object_lock_denied";

@@ -332,6 +335,12 @@ fn verify_single_part_replica(
 const REPLICA_ETAG_MISMATCH_ERROR: &str = "replica etag mismatch: the target persisted different bytes than were sent";

 fn audit_target_version_identity(tgt_client: &TargetClient, source_version_id: &str, assigned_version_id: Option<&str>) {
+    // Every write refreshes the cached verdict, so the convergence fallback
+    // below (`replica_head_fallback`) knows whether a 404 on a
+    // version-addressed HEAD can mean "replica missing" on this target.
+    if let Some(capability) = version_identity_capability_from_put(source_version_id, assigned_version_id) {
+        ReplicationTargetStore::record_version_identity_capability(&tgt_client.arn, capability);
+    }
    if !version_identity_drifted(source_version_id, assigned_version_id) {
        return;
    }
@@ -404,11 +413,70 @@ async fn head_object_fallback(
 ) -> std::result::Result<Option<HeadObjectOutput>, HeadObjectSdkError> {
    match head_object_for_worker(tgt_client, &tgt_client.bucket, object, None).await {
        Ok(oi) => Ok(Some(oi)),
-        Err(e) if e.as_service_error().is_some_and(|se| se.is_not_found()) || has_raw_status(&e, 404) => Ok(None),
+        Err(e) if head_object_not_found(&e) => Ok(None),
        Err(e) => Err(e),
    }
 }

+fn head_object_not_found(err: &HeadObjectSdkError) -> bool {
+    err.as_service_error().is_some_and(|se| se.is_not_found()) || has_raw_status(err, 404)
+}
+
+/// Second look at a replica whose version-addressed HEAD failed, for the two
+/// target shapes where that failure is not a verdict on the replica:
+///
+/// - AWS-style 400/403 (the RustFS uuid is rejected as malformed): HEAD the
+///   current version without a version id; callers compare ETags.
+/// - 404 on a target known to mint its own version ids (the Wasabi shape,
+///   rustfs/backlog#2340): the source id never existed there, so locate the
+///   replica by exact key and ETag through ListObjectVersions and HEAD the id
+///   the target assigned. Without this, every heal, MRF retry and
+///   existing-object resync re-drive PUTs the object again and mints one
+///   more target version.
+///
+/// `None` when the error stands as-is: a real miss on an adopting target, or
+/// a target whose identity contract is still unknown. A failed lookup is
+/// returned as a HEAD-shaped error so callers keep their "target operation
+/// failed" handling (retry later) instead of re-driving the PUT.
+async fn replica_head_fallback(
+    tgt_client: &TargetClient,
+    object: &str,
+    source_etag: Option<&str>,
+    err: &HeadObjectSdkError,
+) -> Option<std::result::Result<Option<HeadObjectOutput>, HeadObjectSdkError>> {
+    if is_version_id_format_mismatch(err) {
+        return Some(head_object_fallback(tgt_client, object).await);
+    }
+    if !head_object_not_found(err)
+        || !ReplicationTargetStore::version_identity_capability(&tgt_client.arn).version_addressing_unreliable()
+    {
+        return None;
+    }
+    let etag = source_etag.filter(|etag| !etag.trim().is_empty())?;
+    Some(match tgt_client.find_version_by_etag(&tgt_client.bucket, object, etag).await {
+        Ok(Some(assigned_version_id)) => {
+            debug!(
+                event = EVENT_REPLICATION_DRIFTED_REPLICA_LOCATED,
+                component = LOG_COMPONENT_ECSTORE,
+                subsystem = LOG_SUBSYSTEM_REPLICATION_RESYNC,
+                bucket = %tgt_client.bucket,
+                object = %object,
+                arn = %tgt_client.arn,
+                assigned_version_id = %assigned_version_id,
+                "Located replica by content identity on a target that mints its own version ids"
+            );
+            match head_object_for_worker(tgt_client, &tgt_client.bucket, object, Some(assigned_version_id)).await {
+                Ok(oi) => Ok(Some(oi)),
+                // The located version disappeared between LIST and HEAD.
+                Err(e) if head_object_not_found(&e) => Ok(None),
+                Err(e) => Err(e),
+            }
+        }
+        Ok(None) => Ok(None),
+        Err(list_err) => Err(Box::new(SdkError::construction_failure(*list_err))),
+    })
+}
+
 /// Resolve the N2 fail-closed gate for an SSE-C passthrough attempt against
 /// this target. Returns `Some(audit_required)` when replication may proceed;
 /// on a freshly-flagged header-dropping target it settles `rinfo` as FAILED
@@ -1400,31 +1468,26 @@ async fn verify_resync_head_result(
                (0, None)
            }
        }
-        Err(err) if is_version_id_format_mismatch(&err) => {
-            // AWS-style target rejects the RustFS UUID versionId
-            // (400). Re-verify without the versionId before
-            // concluding the object failed to replicate, instead
-            // of counting a well-replicated object as failed.
-            match head_object_fallback(target_client.as_ref(), &roi.name).await {
-                Ok(Some(_)) => {
+        Err(err) => {
+            // A version-addressed HEAD is not the last word on every target:
+            // re-verify through the fallback before counting a well-replicated
+            // object as failed (see `replica_head_fallback`).
+            match replica_head_fallback(target_client.as_ref(), &roi.name, roi.etag.as_deref(), &err).await {
+                Some(Ok(Some(_))) => {
                    st.replicated_count += 1;
                    st.replicated_size += roi.size;
                    (roi.size, None)
                }
-                Ok(None) => {
+                Some(Ok(None)) | None => {
                    st.failed_count += 1;
                    (0, Some(err))
                }
-                Err(e2) => {
+                Some(Err(e2)) => {
                    st.failed_count += 1;
                    (0, Some(e2))
                }
            }
        }
-        Err(err) => {
-            st.failed_count += 1;
-            (0, Some(err))
-        }
    }
 }

@@ -3597,11 +3660,8 @@ impl ReplicateObjectInfoExt for ReplicateObjectInfo {
                }
            }
            Err(e) => {
-                if e.as_service_error().is_some_and(|se| se.is_not_found()) || has_raw_status(&e, 404) {
-                    // Object not on target yet → fall through to PUT.
-                } else if is_version_id_format_mismatch(&e) {
-                    // Version-ID format mismatch: retry without versionId and compare ETags.
-                    match head_object_fallback(&tgt_client, &object).await {
+                if let Some(fallback) = replica_head_fallback(&tgt_client, &object, object_info.etag.as_deref(), &e).await {
+                    match fallback {
                        Ok(Some(oi)) if replication_etags_match(object_info.etag.as_deref(), oi.e_tag.as_deref()) => {
                            if ssec_audit_required
                                && !settle_ssec_passthrough_evidence(&oi, &tgt_client, &bucket, &object, &mut rinfo).await
@@ -3631,6 +3691,8 @@ impl ReplicateObjectInfoExt for ReplicateObjectInfo {
                            return rinfo;
                        }
                    }
+                } else if head_object_not_found(&e) {
+                    // Object not on target yet → fall through to PUT.
                } else {
                    rinfo.error = Some(e.to_string());
                    warn!(
@@ -4230,9 +4292,8 @@ async fn resolve_replicate_all_action(
            }
        }
        Err(e) => {
-            if is_version_id_format_mismatch(&e) {
-                // Version-ID format mismatch: retry without versionId and compare ETags.
-                match head_object_fallback(tgt_client, object).await {
+            if let Some(fallback) = replica_head_fallback(tgt_client, object, object_info.etag.as_deref(), &e).await {
+                match fallback {
                    Ok(Some(oi)) => {
                        let etags_match = replication_etags_match(object_info.etag.as_deref(), oi.e_tag.as_deref());
                        if require_existing_target && !etags_match {
@@ -4284,7 +4345,7 @@ async fn resolve_replicate_all_action(
                        return None;
                    }
                }
-            } else if e.as_service_error().is_some_and(|se| se.is_not_found()) || has_raw_status(&e, 404) {
+            } else if head_object_not_found(&e) {
                if require_existing_target {
                    rinfo.error = Some("replica metadata target does not contain this object version".to_string());
                    rinfo.duration = (OffsetDateTime::now_utc() - start_time).unsigned_abs();
@@ -4656,6 +4717,58 @@ where
    result
 }

+#[derive(Debug)]
+struct MultipartReplicationReadPlan {
+    part_number: i32,
+    part_size: i64,
+    range: Option<HTTPRangeSpec>,
+    next_offset: i64,
+}
+
+fn multipart_replication_read_plan(
+    object_info: &ObjectInfo,
+    obj_opts: &ObjectOptions,
+    mut input: ReplicationMultipartPartInput,
+    stored_size: usize,
+    is_last: bool,
+) -> std::io::Result<MultipartReplicationReadPlan> {
+    let empty_last_part = is_last && input.part_size == 0 && stored_size == 0;
+    // Raw reads address stored bytes. Only untransformed legacy parts may
+    // substitute their stored size for a missing logical size.
+    if obj_opts.raw_data_movement_read || (input.part_size == 0 && !object_info.is_compressed() && !object_info.is_encrypted()) {
+        input.part_size = i64::try_from(stored_size).map_err(|_| {
+            std::io::Error::new(std::io::ErrorKind::InvalidData, "multipart replication stored part size exceeds i64")
+        })?;
+    }
+    if empty_last_part {
+        if input.offset < 0 {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidData,
+                "empty multipart replication part has a negative offset",
+            ));
+        }
+        let part_number = i32::try_from(input.part_number)
+            .map_err(|_| std::io::Error::new(std::io::ErrorKind::InvalidData, "multipart replication part number exceeds i32"))?;
+        return Ok(MultipartReplicationReadPlan {
+            part_number,
+            part_size: 0,
+            range: None,
+            next_offset: input.offset,
+        });
+    }
+    let plan = replication_multipart_part_plan(input).map_err(std::io::Error::other)?;
+    Ok(MultipartReplicationReadPlan {
+        part_number: plan.part_number,
+        part_size: plan.part_size,
+        range: Some(HTTPRangeSpec {
+            is_suffix_length: false,
+            start: plan.range.start,
+            end: plan.range.end,
+        }),
+        next_offset: plan.next_offset,
+    })
+}
+
 async fn replicate_multipart_parts_and_complete<S: ReplicationObjectIO>(
    ctx: MultipartReplicationContext<'_, S>,
    upload_id: &str,
@@ -4676,35 +4789,31 @@ async fn replicate_multipart_parts_and_complete<S: ReplicationObjectIO>(

    let mut header_size = replication_put_object_header_size(&put_opts);
    let mut offset: i64 = 0;
-    for part_info in object_info.parts.iter() {
-        // Ciphertext passthrough (raw read) ranges over the stored part
-        // bytes; decrypted reads range over the logical plaintext parts.
-        let part_size = if obj_opts.raw_data_movement_read {
-            part_info.size as i64
-        } else {
-            part_info.actual_size
-        };
-        let part_plan = replication_multipart_part_plan(ReplicationMultipartPartInput {
-            offset,
-            part_number: part_info.number,
-            part_size,
-        })
-        .map_err(|err| std::io::Error::other(err.to_string()))?;
-        let range_spec = HTTPRangeSpec {
-            is_suffix_length: false,
-            start: part_plan.range.start,
-            end: part_plan.range.end,
-        };
+    for (index, part_info) in object_info.parts.iter().enumerate() {
+        let part_plan = multipart_replication_read_plan(
+            object_info,
+            obj_opts,
+            ReplicationMultipartPartInput {
+                offset,
+                part_number: part_info.number,
+                part_size: part_info.actual_size,
+            },
+            part_info.size,
+            index + 1 == object_info.parts.len(),
+        )?;
        offset = part_plan.next_offset;

-        let part_reader = storage
-            .get_object_reader(src_bucket, object, Some(range_spec), HeaderMap::new(), obj_opts)
-            .await
-            .map_err(|e| std::io::Error::other(e.to_string()))?;
-
-        let part_stream = wrap_with_bandwidth_monitor_with_header(part_reader.stream, src_bucket, arn, header_size);
+        let byte_stream = if let Some(range_spec) = part_plan.range {
+            let part_reader = storage
+                .get_object_reader(src_bucket, object, Some(range_spec), HeaderMap::new(), obj_opts)
+                .await
+                .map_err(|e| std::io::Error::other(e.to_string()))?;
+            let part_stream = wrap_with_bandwidth_monitor_with_header(part_reader.stream, src_bucket, arn, header_size);
+            async_read_to_bytestream(part_stream)
+        } else {
+            ByteStream::from_static(b"")
+        };
        header_size = 0;
-        let byte_stream = async_read_to_bytestream(part_stream);

        let object_part = cli
            .put_object_part(
@@ -4760,6 +4869,173 @@ async fn replicate_multipart_parts_and_complete<S: ReplicationObjectIO>(
 #[cfg(test)]
 mod tests {
    use super::super::replication_filemeta_boundary::ReplicateTargetDecision;
+    use super::super::replication_object_decision_boundary::ReplicationMultipartPlanError;
+
+    #[test]
+    fn multipart_read_plan_preserves_legacy_plain_part_ranges() {
+        const MIB: usize = 1024 * 1024;
+        let object_info = ObjectInfo {
+            etag: Some("0123456789abcdef0123456789abcdef".to_string()),
+            size: 6 * 1024 * 1024,
+            ..Default::default()
+        };
+        let mut offset = 0;
+        for (part_number, stored_size, start, end) in [
+            (1, 5 * MIB, 0, 5 * 1024 * 1024 - 1),
+            (2, MIB, 5 * 1024 * 1024, 6 * 1024 * 1024 - 1),
+        ] {
+            let plan = multipart_replication_read_plan(
+                &object_info,
+                &ObjectOptions::default(),
+                ReplicationMultipartPartInput {
+                    offset,
+                    part_number,
+                    part_size: 0,
+                },
+                stored_size,
+                part_number == 2,
+            )
+            .expect("legacy plain parts must use their stored sizes");
+            assert_eq!(plan.part_number, i32::try_from(part_number).expect("part number fits"));
+            assert_eq!(plan.part_size, i64::try_from(stored_size).expect("stored size fits"));
+            let range = plan.range.expect("a nonempty part must read a range");
+            assert!(!range.is_suffix_length);
+            assert_eq!((range.start, range.end), (start, end));
+            assert_eq!(plan.next_offset, end + 1);
+            offset = plan.next_offset;
+        }
+        assert_eq!(offset, object_info.size);
+    }
+
+    #[test]
+    fn multipart_read_plan_distinguishes_transformed_and_raw_sizes() {
+        for metadata in [
+            HashMap::from([("x-rustfs-internal-compression".to_string(), "klauspost/compress/s2".to_string())]),
+            HashMap::from([("x-amz-server-side-encryption".to_string(), "AES256".to_string())]),
+        ] {
+            let object_info = ObjectInfo {
+                user_defined: Arc::new(metadata),
+                ..Default::default()
+            };
+            assert!(object_info.is_compressed() || object_info.is_encrypted());
+            for raw in [false, true] {
+                for actual_size in [-1, 0, 5] {
+                    let result = multipart_replication_read_plan(
+                        &object_info,
+                        &ObjectOptions {
+                            raw_data_movement_read: raw,
+                            ..Default::default()
+                        },
+                        ReplicationMultipartPartInput {
+                            offset: 7,
+                            part_number: 2,
+                            part_size: actual_size,
+                        },
+                        9,
+                        true,
+                    );
+                    if !raw && actual_size <= 0 {
+                        let err = result.expect_err("transformed reads cannot substitute physical bytes for unknown plaintext");
+                        assert!(matches!(
+                            err.get_ref().and_then(|err| err.downcast_ref::<ReplicationMultipartPlanError>()),
+                            Some(ReplicationMultipartPlanError::InvalidPartSize { part_size })
+                                if *part_size == actual_size
+                        ));
+                    } else {
+                        let plan = result.expect("the selected representation has a known positive size");
+                        let expected_size = if raw { 9 } else { 5 };
+                        assert_eq!(plan.part_number, 2);
+                        assert_eq!(plan.part_size, expected_size);
+                        let range = plan.range.expect("a nonempty part must read a range");
+                        assert_eq!((range.start, range.end), (7, 7 + expected_size - 1));
+                        assert_eq!(plan.next_offset, 7 + expected_size);
+                    }
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn multipart_read_plan_retains_an_empty_last_part_without_advancing() {
+        for offset in [5 * 1024 * 1024, i64::MAX] {
+            for raw in [false, true] {
+                let plan = multipart_replication_read_plan(
+                    &ObjectInfo::default(),
+                    &ObjectOptions {
+                        raw_data_movement_read: raw,
+                        ..Default::default()
+                    },
+                    ReplicationMultipartPartInput {
+                        offset,
+                        part_number: 2,
+                        part_size: 0,
+                    },
+                    0,
+                    true,
+                )
+                .expect("an empty final part needs no range read");
+                assert_eq!(plan.part_number, 2);
+                assert_eq!(plan.part_size, 0);
+                assert!(plan.range.is_none());
+                assert_eq!(plan.next_offset, offset);
+            }
+        }
+    }
+
+    #[test]
+    fn multipart_read_plan_rejects_invalid_empty_parts_and_ranges() {
+        for (offset, part_number, actual_size, stored_size, is_last) in [
+            (0, 1, 0, 0, false),
+            (0, 2, -1, 0, true),
+            (0, 2, -1, 9, true),
+            (-1, 2, 0, 0, true),
+            (0, usize::try_from(i32::MAX).expect("i32 fits usize") + 1, 0, 0, true),
+            (i64::MAX, 2, 1, 1, true),
+            (i64::MAX, 2, 2, 2, true),
+        ] {
+            let err = multipart_replication_read_plan(
+                &ObjectInfo::default(),
+                &ObjectOptions::default(),
+                ReplicationMultipartPartInput {
+                    offset,
+                    part_number,
+                    part_size: actual_size,
+                },
+                stored_size,
+                is_last,
+            )
+            .expect_err("invalid part metadata must not become a successful transport plan");
+            assert!(
+                err.kind() == std::io::ErrorKind::InvalidData
+                    || err.get_ref().is_some_and(|err| { err.is::<ReplicationMultipartPlanError>() }),
+                "the failure must preserve a typed metadata or planner error: {err}"
+            );
+        }
+    }
+
+    #[cfg(target_pointer_width = "64")]
+    #[test]
+    fn multipart_read_plan_rejects_physical_size_overflow() {
+        for raw in [false, true] {
+            let err = multipart_replication_read_plan(
+                &ObjectInfo::default(),
+                &ObjectOptions {
+                    raw_data_movement_read: raw,
+                    ..Default::default()
+                },
+                ReplicationMultipartPartInput {
+                    offset: 0,
+                    part_number: 1,
+                    part_size: 0,
+                },
+                usize::MAX,
+                true,
+            )
+            .expect_err("a physical size outside the range API must be rejected before casting");
+            assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+            assert_eq!(err.to_string(), "multipart replication stored part size exceeds i64");
+        }
+    }

    #[test]
    fn same_state_terminal_retry_uses_validate_only() {
@@ -5168,6 +5444,178 @@ mod tests {
        ReplicationTargetStore::register_test_target(target).await;
    }

+    const DRIFTED_ASSIGNED_VERSION_ID: &str = "001788697733811332140-fR6j6uXKV-";
+    const DRIFTED_ETAG: &str = "9a0364b9e99bb480dd25e1f0284c8555";
+
+    /// The Wasabi shape (rustfs/backlog#2340): a version-addressed HEAD with
+    /// the source uuid answers 404 (not the AWS 400), ListObjectVersions shows
+    /// the id the target minted, and a HEAD by that id succeeds. Serves exactly
+    /// `requests` connections and returns the request lines it saw.
+    fn spawn_drifted_target_server(requests: usize) -> (String, std::thread::JoinHandle<Vec<String>>) {
+        use std::io::{Read, Write};
+
+        let listener = std::net::TcpListener::bind(("127.0.0.1", 0)).expect("test HTTP listener should bind");
+        let endpoint = format!("http://{}", listener.local_addr().expect("test HTTP listener should have an address"));
+        let handle = std::thread::spawn(move || {
+            let mut seen = Vec::new();
+            for _ in 0..requests {
+                let (mut stream, _) = listener.accept().expect("test HTTP client should connect");
+                let mut request = [0_u8; 8192];
+                let bytes_read = stream.read(&mut request).expect("test HTTP request should be read");
+                let text = String::from_utf8_lossy(&request[..bytes_read]).to_string();
+                let request_line = text.lines().next().unwrap_or_default().to_string();
+                let response = if request_line.starts_with("HEAD ") {
+                    if request_line.contains(&format!("versionId={DRIFTED_ASSIGNED_VERSION_ID}")) {
+                        format!(
+                            "HTTP/1.1 200 OK\r\nETag: \"{DRIFTED_ETAG}\"\r\nContent-Length: 4\r\nLast-Modified: Sun, 06 Sep 2026 10:00:00 GMT\r\nConnection: close\r\n\r\n"
+                        )
+                    } else {
+                        "HTTP/1.1 404 Not Found\r\nContent-Length: 0\r\nConnection: close\r\n\r\n".to_string()
+                    }
+                } else if request_line.starts_with("GET ") && request_line.contains("versions") {
+                    let body = format!(
+                        "<?xml version=\"1.0\" encoding=\"UTF-8\"?><ListVersionsResult xmlns=\"http://s3.amazonaws.com/doc/2006-03-01/\"><Name>target-bucket</Name><Prefix>object</Prefix><MaxKeys>1000</MaxKeys><IsTruncated>false</IsTruncated><Version><Key>object</Key><VersionId>{DRIFTED_ASSIGNED_VERSION_ID}</VersionId><IsLatest>true</IsLatest><LastModified>2026-09-06T10:00:00.000Z</LastModified><ETag>&quot;{DRIFTED_ETAG}&quot;</ETag><Size>4</Size><StorageClass>STANDARD</StorageClass></Version></ListVersionsResult>"
+                    );
+                    format!(
+                        "HTTP/1.1 200 OK\r\nContent-Type: application/xml\r\nContent-Length: {}\r\nConnection: close\r\n\r\n{body}",
+                        body.len()
+                    )
+                } else {
+                    "HTTP/1.1 500 Unexpected\r\nContent-Length: 0\r\nConnection: close\r\n\r\n".to_string()
+                };
+                stream
+                    .write_all(response.as_bytes())
+                    .expect("test HTTP response should be written");
+                seen.push(request_line);
+            }
+            seen
+        });
+        (endpoint, handle)
+    }
+
+    fn drifted_roi_and_object() -> (ReplicateObjectInfo, ObjectInfo) {
+        let roi = ReplicateObjectInfo {
+            bucket: "source".to_string(),
+            name: "object".to_string(),
+            version_id: Some(Uuid::new_v4()),
+            op_type: ReplicationType::Heal,
+            replication_status: ReplicationStatusType::Pending,
+            etag: Some(DRIFTED_ETAG.to_string()),
+            size: 4,
+            ..Default::default()
+        };
+        let object_info = ObjectInfo {
+            bucket: roi.bucket.clone(),
+            name: roi.name.clone(),
+            version_id: roi.version_id,
+            etag: Some(DRIFTED_ETAG.to_string()),
+            size: 4,
+            ..Default::default()
+        };
+        (roi, object_info)
+    }
+
+    #[tokio::test]
+    async fn heal_redrive_locates_replica_by_etag_on_target_that_mints_own_version_ids() {
+        let (endpoint, server) = spawn_drifted_target_server(3);
+        let target = test_target_client(endpoint);
+        ReplicationTargetStore::record_version_identity_capability(&target.arn, VersionIdentityCapability::MintsOwn);
+        let (roi, object_info) = drifted_roi_and_object();
+        let mut rinfo = replicate_all_target_info(&roi, &target);
+
+        let action = resolve_replicate_all_action(
+            ReplicateAllActionContext {
+                roi: &roi,
+                tgt_client: &target,
+                bucket: &roi.bucket,
+                object: &roi.name,
+                start_time: OffsetDateTime::now_utc(),
+                ssec_audit_required: false,
+            },
+            object_info,
+            &mut rinfo,
+        )
+        .await;
+
+        assert!(
+            matches!(action, Some((ReplicationAction::None, _))),
+            "a replica located by content identity must not be re-driven: {action:?}"
+        );
+        assert!(rinfo.error.is_none(), "{:?}", rinfo.error);
+        let seen = server.join().expect("test HTTP server should finish");
+        assert_eq!(seen.len(), 3, "HEAD by source id, ListObjectVersions, HEAD by assigned id: {seen:?}");
+        assert!(seen[0].starts_with("HEAD ") && seen[0].contains(&roi.version_id.unwrap().to_string()));
+        assert!(seen[1].starts_with("GET ") && seen[1].contains("prefix=object"), "{}", seen[1]);
+        assert!(seen[2].starts_with("HEAD ") && seen[2].contains(DRIFTED_ASSIGNED_VERSION_ID));
+    }
+
+    #[tokio::test]
+    async fn head_not_found_still_replicates_when_identity_contract_is_unknown() {
+        // Same 404, but the target never revealed whether it adopts version
+        // ids: a 404 keeps meaning "replica missing" (adopting targets, e.g.
+        // RustFS/MinIO peers, must not skip a genuinely missing version).
+        let (endpoint, server) = spawn_head_status_server(404);
+        let target = test_target_client(endpoint);
+        let (roi, object_info) = drifted_roi_and_object();
+        let mut rinfo = replicate_all_target_info(&roi, &target);
+
+        let action = resolve_replicate_all_action(
+            ReplicateAllActionContext {
+                roi: &roi,
+                tgt_client: &target,
+                bucket: &roi.bucket,
+                object: &roi.name,
+                start_time: OffsetDateTime::now_utc(),
+                ssec_audit_required: false,
+            },
+            object_info,
+            &mut rinfo,
+        )
+        .await;
+
+        assert!(matches!(action, Some((ReplicationAction::All, _))));
+        server.join().expect("test HTTP server should finish");
+    }
+
+    #[tokio::test]
+    async fn resync_verification_counts_drifted_replica_as_replicated() {
+        let (endpoint, server) = spawn_drifted_target_server(3);
+        let target = test_target_client(endpoint);
+        ReplicationTargetStore::record_version_identity_capability(&target.arn, VersionIdentityCapability::MintsOwn);
+        let (roi, _) = drifted_roi_and_object();
+        let mut st = TargetReplicationResyncStatus::default();
+
+        let head_result =
+            head_object_for_worker(target.as_ref(), &target.bucket, &roi.name, roi.version_id.map(|v| v.to_string())).await;
+        let (size, err) = verify_resync_head_result(head_result, &roi, &mut st, &target).await;
+
+        assert!(err.is_none(), "{err:?}");
+        assert_eq!((size, st.replicated_count, st.failed_count), (4, 1, 0));
+        server.join().expect("test HTTP server should finish");
+    }
+
+    #[test]
+    fn put_response_audit_records_identity_verdict() {
+        let target = test_target_client("http://127.0.0.1:1".to_string());
+        let source = Uuid::new_v4().to_string();
+        audit_target_version_identity(&target, &source, Some(DRIFTED_ASSIGNED_VERSION_ID));
+        assert_eq!(
+            ReplicationTargetStore::version_identity_capability(&target.arn),
+            VersionIdentityCapability::MintsOwn
+        );
+        audit_target_version_identity(&target, &source, Some(&source));
+        assert_eq!(
+            ReplicationTargetStore::version_identity_capability(&target.arn),
+            VersionIdentityCapability::Adopts
+        );
+        // An unversioned write carries no contract and must not overwrite it.
+        audit_target_version_identity(&target, "null", None);
+        assert_eq!(
+            ReplicationTargetStore::version_identity_capability(&target.arn),
+            VersionIdentityCapability::Adopts
+        );
+    }
+
    #[test]
    fn resync_admission_configuration_is_bounded() {
        assert_eq!(ENV_REPL_RESYNC_MAX_JOBS, "RUSTFS_REPL_RESYNC_MAX_JOBS");
@@ -6339,4 +6787,326 @@ mod tests {
            "one target's report must not silence another's"
        );
    }
+    mod multipart_transport_tests {
+        use super::super::super::replication_filemeta_boundary::ObjectPartInfo;
+        use super::super::super::replication_storage_boundary::ObjectIO as _;
+        use super::*;
+        use bytes::Bytes;
+        use http_body_util::{BodyExt, Full};
+        use std::convert::Infallible;
+
+        #[derive(Debug)]
+        struct Source {
+            body: Bytes,
+            info: ObjectInfo,
+            ranges: StdMutex<Vec<(i64, i64)>>,
+            full_reads: std::sync::atomic::AtomicUsize,
+        }
+
+        #[async_trait::async_trait]
+        impl super::super::super::replication_storage_boundary::ObjectIO for Source {
+            type Error = Error;
+            type RangeSpec = HTTPRangeSpec;
+            type HeaderMap = HeaderMap;
+            type ObjectOptions = ObjectOptions;
+            type ObjectInfo = ObjectInfo;
+            type GetObjectReader = GetObjectReader;
+            type PutObjectReader = super::super::super::replication_storage_boundary::PutObjReader;
+
+            async fn get_object_reader(
+                &self,
+                _bucket: &str,
+                _object: &str,
+                range: Option<HTTPRangeSpec>,
+                _headers: HeaderMap,
+                opts: &ObjectOptions,
+            ) -> Result<GetObjectReader> {
+                assert_eq!(
+                    opts.version_id,
+                    self.info.version_id.map(|id| id.to_string()),
+                    "every read retains the selected source version"
+                );
+                if range.is_none() {
+                    self.full_reads.fetch_add(1, Ordering::Relaxed);
+                    return Ok(GetObjectReader {
+                        stream: Box::new(std::io::Cursor::new(self.body.clone())),
+                        object_info: self.info.clone(),
+                        buffered_body: None,
+                        body_source: Default::default(),
+                    });
+                }
+                let range = range.expect("multipart transport must request an explicit nonempty range");
+                assert!(!range.is_suffix_length);
+                assert!(range.start <= range.end, "empty parts must not issue an inverted range");
+                self.ranges.lock().expect("range journal lock").push((range.start, range.end));
+                let start = usize::try_from(range.start).expect("nonnegative start");
+                let end = usize::try_from(range.end).expect("nonnegative end");
+                let body = self.body.slice(start..=end);
+                Ok(GetObjectReader {
+                    stream: Box::new(std::io::Cursor::new(body)),
+                    object_info: self.info.clone(),
+                    buffered_body: None,
+                    body_source: Default::default(),
+                })
+            }
+
+            async fn put_object(
+                &self,
+                _bucket: &str,
+                _object: &str,
+                _data: &mut Self::PutObjectReader,
+                _opts: &ObjectOptions,
+            ) -> Result<ObjectInfo> {
+                panic!("replication must not overwrite its source")
+            }
+        }
+
+        #[derive(Debug)]
+        struct RequestRecord {
+            method: http::Method,
+            query: HashMap<String, String>,
+            headers: HeaderMap,
+            body: Bytes,
+        }
+
+        #[tokio::test]
+        async fn multipart_transport_preserves_legacy_zero_actual_sizes() {
+            run_transport(4096, None).await;
+        }
+
+        #[tokio::test]
+        async fn multipart_transport_uploads_an_empty_last_part_without_reading_a_range() {
+            run_transport(0, None).await;
+        }
+
+        #[tokio::test]
+        async fn multipart_transport_preserves_transformed_unknown_nonempty_parts() {
+            for unknown_part in [(0, 0), (1, 0), (0, -1), (1, -1)] {
+                run_transport(4096, Some(unknown_part)).await;
+            }
+        }
+
+        #[tokio::test]
+        async fn multipart_transport_preserves_transformed_empty_tail() {
+            run_transport(0, Some((1, 0))).await;
+        }
+
+        async fn run_transport(tail_size: usize, unknown_part: Option<(usize, i64)>) {
+            const FIRST_SIZE: usize = 5 * 1024 * 1024;
+            let body = Bytes::from([vec![0x35; FIRST_SIZE], vec![0xa7; tail_size]].concat());
+            let etag = faster_hex::hex_string(rustfs_utils::hash::HashAlgorithm::Md5.hash_encode(&body).as_ref());
+            let source = Arc::new(Source {
+                info: ObjectInfo {
+                    size: i64::try_from(body.len() + if unknown_part.is_some() { 16 } else { 0 }).expect("stored size"),
+                    actual_size: i64::try_from(body.len()).expect("body size"),
+                    etag: Some(etag.clone()),
+                    version_id: Some(Uuid::new_v4()),
+                    user_defined: Arc::new(if unknown_part.is_some() {
+                        HashMap::from([("x-amz-server-side-encryption".to_string(), "AES256".to_string())])
+                    } else {
+                        HashMap::new()
+                    }),
+                    parts: Arc::new(vec![
+                        ObjectPartInfo {
+                            number: 1,
+                            size: FIRST_SIZE + if unknown_part.is_some() { 8 } else { 0 },
+                            actual_size: if let Some((0, size)) = unknown_part {
+                                size
+                            } else if unknown_part.is_some() || tail_size == 0 {
+                                i64::try_from(FIRST_SIZE).expect("first part size")
+                            } else {
+                                0
+                            },
+                            ..Default::default()
+                        },
+                        ObjectPartInfo {
+                            number: 2,
+                            size: tail_size + if unknown_part.is_some() { 8 } else { 0 },
+                            actual_size: if let Some((1, size)) = unknown_part {
+                                size
+                            } else if unknown_part.is_some() {
+                                i64::try_from(tail_size).expect("tail logical size")
+                            } else {
+                                0
+                            },
+                            ..Default::default()
+                        },
+                    ]),
+                    ..Default::default()
+                },
+                body: body.clone(),
+                ranges: StdMutex::new(Vec::new()),
+                full_reads: std::sync::atomic::AtomicUsize::new(0),
+            });
+            let journal = Arc::new(StdMutex::new(Vec::<RequestRecord>::new()));
+            let listener = tokio::net::TcpListener::bind(("127.0.0.1", 0))
+                .await
+                .expect("bind multipart target");
+            let endpoint = format!("http://{}", listener.local_addr().expect("multipart target address"));
+            let server_journal = journal.clone();
+            let server = tokio::spawn(async move {
+                let mut connections = JoinSet::new();
+                loop {
+                    let (stream, _) = listener.accept().await.expect("accept multipart request");
+                    let journal = server_journal.clone();
+                    connections.spawn(async move {
+                            let service = hyper::service::service_fn(move |request: hyper::Request<hyper::body::Incoming>| {
+                                let journal = journal.clone();
+                                async move {
+                                    let (request, body) = request.into_parts();
+                                    let query: HashMap<String, String> = url::form_urlencoded::parse(
+                                        request.uri.query().unwrap_or_default().as_bytes(),
+                                    ).into_owned().collect();
+                                    let body = body.collect().await.expect("read complete multipart request body").to_bytes();
+                                    let response = if request.method == http::Method::POST && query.contains_key("uploads") {
+                                        "<InitiateMultipartUploadResult><Bucket>target-bucket</Bucket><Key>object</Key><UploadId>upload-1</UploadId></InitiateMultipartUploadResult>"
+                                    } else if request.method == http::Method::PUT {
+                                        ""
+                                    } else if request.method == http::Method::POST && query.contains_key("uploadId") {
+                                        "<CompleteMultipartUploadResult><Location>http://localhost/object</Location><Bucket>target-bucket</Bucket><Key>object</Key><ETag>&quot;target-2&quot;</ETag></CompleteMultipartUploadResult>"
+                                    } else if request.method == http::Method::DELETE && query.contains_key("uploadId") {
+ ""
+ } else {
+                                        panic!("unexpected multipart request: {} {}", request.method, request.uri)
+                                    };
+                                    let response_etag = if request.method == http::Method::PUT && !query.contains_key("partNumber") {
+                                        format!("\"{}\"", faster_hex::hex_string(rustfs_utils::hash::HashAlgorithm::Md5.hash_encode(&body).as_ref()))
+                                    } else {
+                                        "\"uploaded-part\"".to_string()
+                                    };
+                                    journal.lock().expect("request journal lock").push(RequestRecord {
+                                        method: request.method, query, headers: request.headers, body,
+                                    });
+                                    Ok::<_, Infallible>(hyper::Response::builder()
+                                        .header("content-type", "application/xml")
+                                        .header("etag", response_etag)
+                                        .body(Full::new(Bytes::from_static(response.as_bytes())))
+                                        .expect("multipart response"))
+                                }
+                            });
+                            hyper::server::conn::http1::Builder::new()
+                                .serve_connection(hyper_util::rt::TokioIo::new(stream), service)
+                                .await.expect("serve multipart connection");
+                        });
+                }
+            });
+            let mut target = test_target_client(endpoint);
+            let config = target
+                .client
+                .config()
+                .to_builder()
+                .request_checksum_calculation(aws_sdk_s3::config::RequestChecksumCalculation::WhenRequired)
+                .force_path_style(true)
+                .build();
+            Arc::get_mut(&mut target).expect("unshared test target").client = Arc::new(aws_sdk_s3::Client::from_conf(config));
+            let (put_opts, is_multipart) = replication_put_object_options("STANDARD", &source.info).expect("replication options");
+            let opts = ObjectOptions {
+                version_id: source.info.version_id.map(|id| id.to_string()),
+                ..Default::default()
+            };
+            let reader = source
+                .get_object_reader("source", "object", None, HeaderMap::new(), &opts)
+                .await
+                .expect("open the existing full-object stream");
+            let result = tokio::time::timeout(
+                std::time::Duration::from_secs(30),
+                replicate_all_payload_to_target(
+                    ReplicateAllPayloadContext {
+                        storage: &source,
+                        tgt_client: &target,
+                        bucket: "source",
+                        object: "object",
+                        object_info: &source.info,
+                        obj_opts: &opts,
+                        arn: &target.arn,
+                        transfer_size: i64::try_from(body.len()).expect("plaintext size"),
+                        is_multipart,
+                        put_opts,
+                    },
+                    reader,
+                ),
+            )
+            .await;
+            server.abort();
+            assert!(server.await.expect_err("fixture server is stopped").is_cancelled());
+            if let Some(error) = result.expect("replication must finish") {
+                panic!("legacy parts must replicate successfully: {error}");
+            }
+            assert_eq!(
+                source.full_reads.load(Ordering::Relaxed),
+                1,
+                "reuse the initial full stream without an extra read"
+            );
+            if unknown_part.is_some() {
+                let requests = journal.lock().expect("request journal lock");
+                assert_eq!(requests.len(), 1, "unknown transformed boundaries retain one streaming PUT");
+                let request = &requests[0];
+                assert_eq!(request.method, http::Method::PUT);
+                let source_version = source.info.version_id.map(|id| id.to_string()).expect("versioned fixture");
+                assert_eq!(
+                    request.query,
+                    HashMap::from([
+                        ("x-id".to_string(), "PutObject".to_string()),
+                        ("versionId".to_string(), source_version.clone()),
+                    ]),
+                    "single PUT carries only the SDK operation query and the source versionId the target must reuse"
+                );
+                assert_eq!(request.body, body, "single PUT includes every byte of both source parts");
+                assert_eq!(
+                    request.headers.get("content-length").expect("body length"),
+                    body.len().to_string().as_str()
+                );
+                assert_eq!(
+                    rustfs_utils::http::get_header(&request.headers, rustfs_utils::http::SUFFIX_SOURCE_ETAG).as_deref(),
+                    Some(etag.as_str())
+                );
+                assert_eq!(
+                    rustfs_utils::http::get_header(&request.headers, rustfs_utils::http::SUFFIX_SOURCE_VERSION_ID)
+                        .map(|value| value.into_owned()),
+                    Some(source_version),
+                    "single PUT preserves the selected source version"
+                );
+                assert!(
+                    source.ranges.lock().expect("range journal lock").is_empty(),
+                    "unknown logical boundaries must not issue guessed ranges"
+                );
+                return;
+            }
+
+            let requests = journal.lock().expect("request journal lock");
+            assert_eq!(requests.len(), 4, "initiate, two upload parts, and complete without retries");
+            assert!(requests[0].query.contains_key("uploads"));
+            for (index, expected) in [(1, body.slice(..FIRST_SIZE)), (2, body.slice(FIRST_SIZE..))] {
+                assert_eq!(requests[index].method, http::Method::PUT);
+                assert_eq!(requests[index].query.get("partNumber"), Some(&index.to_string()));
+                assert_eq!(requests[index].body, expected, "upload part contains the exact source range");
+                assert_eq!(
+                    requests[index].headers.get("content-length").expect("part content length"),
+                    expected.len().to_string().as_str()
+                );
+            }
+            let complete = &requests[3];
+            assert_eq!(complete.method, http::Method::POST);
+            assert_eq!(
+                rustfs_utils::http::get_header(&complete.headers, rustfs_utils::http::SUFFIX_SOURCE_ETAG).as_deref(),
+                Some(etag.as_str())
+            );
+            let complete_xml = std::str::from_utf8(&complete.body).expect("complete XML");
+            assert_eq!(
+                complete_xml.matches("<Part>").count(),
+                2,
+                "the empty final part must remain in the completion list"
+            );
+            assert!(complete_xml.contains("<PartNumber>1</PartNumber>"));
+            assert!(complete_xml.contains("<PartNumber>2</PartNumber>"));
+            let mut expected_ranges = vec![(0, i64::try_from(FIRST_SIZE - 1).expect("first end"))];
+            if tail_size > 0 {
+                expected_ranges.push((
+                    i64::try_from(FIRST_SIZE).expect("tail start"),
+                    i64::try_from(body.len() - 1).expect("tail end"),
+                ));
+            }
+            assert_eq!(*source.ranges.lock().expect("range journal lock"), expected_ranges);
+        }
+    }
 }
@@ -48,6 +48,7 @@ pub use rustfs_replication::{ObjectLockIntegrity, object_lock_put_integrity};
 pub(crate) use rustfs_replication::{
    SsecPassthroughGate, is_replication_target_offline_error, ssec_passthrough_gate, version_identity_drifted,
 };
+pub use rustfs_replication::{VersionIdentityCapability, version_identity_capability_from_put};

 use super::replication_config_store::ReplicationConfigStore;
 use super::replication_error_boundary::{Error, Result};
@@ -192,6 +193,14 @@ impl ReplicationTargetStore {
            .await
    }

+    pub(crate) fn version_identity_capability(arn: &str) -> VersionIdentityCapability {
+        BucketTargetSys::get().version_identity_capability(arn)
+    }
+
+    pub(crate) fn record_version_identity_capability(arn: &str, capability: VersionIdentityCapability) {
+        BucketTargetSys::get().record_version_identity_capability(arn, capability)
+    }
+
    #[cfg(test)]
    pub(crate) async fn register_test_target(target_client: &Arc<TargetClient>) {
        BucketTargetSys::get().arn_remotes_map.write().await.insert(
@@ -248,7 +257,16 @@ pub(crate) fn replication_put_object_options(sc: &str, object_info: &ObjectInfo)
        meta.insert(AMZ_SERVER_SIDE_ENCRYPTION.to_string(), "aws:kms".to_string());
    }

-    let mut is_multipart = object_info.is_multipart();
+    // Older transformed objects can have physical parts without logical part
+    // lengths. Keep their existing whole-object transport: physical sizes are
+    // not plaintext boundaries for a multipart replication read.
+    let legacy_single_put = object_info.etag.as_deref().is_none_or(|etag| etag.len() == 32);
+    let base_is_multipart = object_info.is_multipart()
+        && !(legacy_single_put
+            && object_info.parts.len() > 1
+            && (object_info.is_compressed() || object_info.is_encrypted())
+            && object_info.parts.iter().any(|part| part.actual_size <= 0));
+    let mut is_multipart = base_is_multipart;

    if let Some(checksum_data) = &object_info.checksum
        && !checksum_data.is_empty()
@@ -259,8 +277,8 @@ pub(crate) fn replication_put_object_options(sc: &str, object_info: &ObjectInfo)
        } else if object_info.is_encrypted() {
            // Encrypted checksums cannot be exposed as plaintext headers, and
            // decrypt_checksums reports is_multipart=false for them (a value
-            // the response path relies on). Keep the object's own multipart
-            // flag so encrypted objects stay on the multipart route.
+            // the response path relies on). Keep the transport selected from
+            // the object's layout and readable part boundaries.
        } else {
            let (checksum_meta, checksum_record_is_multipart) = object_info.decrypt_checksums(0, &HeaderMap::new())?;
            // The checksum record describes how the *checksum* is composed,
@@ -268,23 +286,37 @@ pub(crate) fn replication_put_object_options(sc: &str, object_info: &ObjectInfo)
            // MULTIPART flag even on a multipart upload, so trusting it here
            // routed a 768-part object through a single PutObject and the
            // target rejected the 6 GiB body with EntityTooLarge
-            // (rustfs#6825). The object's own shape is the authority: the
+            // (rustfs#6825). The usable part layout is the authority: the
            // record may only add multipart-ness, never take it away.
-            is_multipart = object_info.is_multipart() || checksum_record_is_multipart;
+            is_multipart = base_is_multipart || checksum_record_is_multipart;

-            for (key, value) in checksum_meta.iter() {
-                if key != AMZ_CHECKSUM_TYPE {
-                    meta.insert(key.clone(), value.clone());
-                }
-            }
-
-            if !object_info.is_multipart()
+            if !base_is_multipart
                && checksum_meta
                    .get(AMZ_CHECKSUM_TYPE)
                    .is_some_and(|value| value == AMZ_CHECKSUM_TYPE_FULL_OBJECT)
            {
                is_multipart = false;
            }
+
+            // The record keys each checksum by algorithm name ("CRC32"); the
+            // target only reads `x-amz-checksum-<algorithm>`. Inserting the bare
+            // name here made `PutObjectOptions::header()` send it as user
+            // metadata (`x-amz-meta-crc32`), so no replica ever carried the
+            // source checksum (rustfs/backlog#2340). The object-level record
+            // describes one PUT body: a multipart replica is rebuilt part by
+            // part, and its CreateMultipartUpload must not announce a checksum
+            // the parts do not carry, so the record is forwarded on the
+            // single-PUT route only (MinIO `getCRCMeta` parity).
+            if !is_multipart {
+                for (key, value) in checksum_meta.iter() {
+                    if key == AMZ_CHECKSUM_TYPE {
+                        continue;
+                    }
+                    if let Some(header) = rustfs_rio::ChecksumType::from_string(key).key() {
+                        meta.insert(header.to_string(), value.clone());
+                    }
+                }
+            }
        }
    }

@@ -516,6 +548,7 @@ fn is_standard_header(key: &str) -> bool {

 #[cfg(test)]
 mod tests {
+    use super::super::replication_filemeta_boundary::ObjectPartInfo;
    use super::*;
    use aws_smithy_types::DateTime;
    use rustfs_replication::content_matches_by_etag;
@@ -550,6 +583,109 @@ mod tests {
        checksum.to_bytes(&combined)
    }

+    fn replication_route_metadata() -> [(&'static str, Arc<HashMap<String, String>>); 4] {
+        let mut compressed = HashMap::new();
+        rustfs_utils::http::insert_str(&mut compressed, rustfs_utils::http::SUFFIX_COMPRESSION, "zstd".to_string());
+        [
+            ("plain", Arc::new(HashMap::new())),
+            ("compressed", Arc::new(compressed)),
+            (
+                "encrypted",
+                Arc::new(HashMap::from([(AMZ_SERVER_SIDE_ENCRYPTION.to_string(), "AES256".to_string())])),
+            ),
+            (
+                "ssec",
+                Arc::new(HashMap::from([(SSEC_ALGORITHM_HEADER.to_string(), "AES256".to_string())])),
+            ),
+        ]
+    }
+
+    fn replication_route_object(
+        etag: Option<&str>,
+        actual_sizes: [i64; 3],
+        metadata: Arc<HashMap<String, String>>,
+    ) -> ObjectInfo {
+        ObjectInfo {
+            etag: etag.map(str::to_string),
+            size: 48,
+            actual_size: 12,
+            user_defined: metadata,
+            parts: Arc::new(
+                actual_sizes
+                    .into_iter()
+                    .enumerate()
+                    .map(|(index, actual_size)| ObjectPartInfo {
+                        number: index + 1,
+                        size: 16,
+                        actual_size,
+                        ..Default::default()
+                    })
+                    .collect(),
+            ),
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn legacy_transformed_single_put_parts_keep_the_previous_replication_route() {
+        let [_, (_, compressed), (_, encrypted), (_, ssec)] = replication_route_metadata();
+        let cases = [
+            (
+                "compressed middle zero",
+                compressed.clone(),
+                Some("0123456789abcdef0123456789abcdef"),
+                [4, 0, 4],
+            ),
+            ("compressed tail unknown", compressed, None, [4, 4, -1]),
+            (
+                "encrypted middle unknown",
+                encrypted,
+                Some("gggggggggggggggggggggggggggggggg"),
+                [4, -1, 4],
+            ),
+            ("ssec tail zero", ssec.clone(), None, [4, 4, 0]),
+            ("ssec middle unknown", ssec, Some("gggggggggggggggggggggggggggggggg"), [4, -1, 4]),
+        ];
+        for (name, metadata, etag, actual_sizes) in cases {
+            for checksum in [None, Some(full_object_multipart_checksum_record())] {
+                let mut object_info = replication_route_object(etag, actual_sizes, metadata.clone());
+                object_info.checksum = checksum;
+                assert!(object_info.is_multipart(), "{name}: physical parts remain visible to metadata APIs");
+                assert!(object_info.is_compressed() || object_info.is_encrypted());
+
+                let (options, is_multipart) =
+                    replication_put_object_options("STANDARD", &object_info).expect("legacy transformed put options");
+                assert!(
+                    !is_multipart,
+                    "{name}: unknown logical part sizes must preserve the old whole-object route"
+                );
+                assert_eq!(options.internal.source_etag, etag.unwrap_or_default());
+                if metadata.contains_key(SSEC_ALGORITHM_HEADER) {
+                    assert_eq!(
+                        get_header_map(&options.user_metadata, SUFFIX_REPLICATION_SSEC_CRC).is_some(),
+                        object_info.checksum.is_some(),
+                        "SSE-C checksums retain their raw passthrough transport"
+                    );
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn positive_part_sizes_and_legacy_multipart_etags_keep_the_replication_route() {
+        for (name, metadata) in replication_route_metadata() {
+            for (etag, actual_sizes) in [
+                ("0123456789abcdef0123456789abcdef", [4, 4, 4]),
+                ("0123456789abcdef0123456789abcdef-3", [4, 0, -1]),
+            ] {
+                let mut object_info = replication_route_object(Some(etag), actual_sizes, metadata.clone());
+                object_info.checksum = Some(full_object_multipart_checksum_record());
+                let (_, is_multipart) = replication_put_object_options("STANDARD", &object_info).expect("multipart put options");
+                assert!(is_multipart, "{name}/{etag}: usable sizes and old multipart ETags must retain MPU");
+            }
+        }
+    }
+
    #[test]
    fn multipart_object_with_full_object_checksum_keeps_the_multipart_route() {
        // rustfs#6825: a 768-part upload was replicated with a single
@@ -582,6 +718,36 @@ mod tests {
        );
    }

+    #[test]
+    fn stored_multipart_parts_keep_the_replication_route_without_a_multipart_etag() {
+        for etag in [Some("0123456789abcdef0123456789abcdef"), None] {
+            for checksum in [None, Some(full_object_multipart_checksum_record())] {
+                let object_info = ObjectInfo {
+                    etag: etag.map(str::to_string),
+                    checksum,
+                    parts: Arc::new(
+                        (1..=2)
+                            .map(|number| ObjectPartInfo {
+                                number,
+                                ..Default::default()
+                            })
+                            .collect(),
+                    ),
+                    ..Default::default()
+                };
+                let (options, is_multipart) =
+                    replication_put_object_options("STANDARD", &object_info).expect("build put options");
+
+                assert!(
+                    is_multipart,
+                    "stored parts must retain multipart routing: etag={etag:?}, checksum={:?}",
+                    object_info.checksum
+                );
+                assert_eq!(options.internal.source_etag, etag.unwrap_or_default());
+            }
+        }
+    }
+
    #[test]
    fn checksum_record_never_changes_the_transport_a_single_part_object_needs() {
        // The mirror of the rustfs#6825 guard: an object stored as one PUT
@@ -592,6 +758,10 @@ mod tests {
        let object_info = ObjectInfo {
            etag: Some("0123456789abcdef0123456789abcdef".to_string()),
            checksum: Some(checksum.to_bytes(&[])),
+            parts: Arc::new(vec![ObjectPartInfo {
+                number: 1,
+                ..Default::default()
+            }]),
            ..Default::default()
        };

@@ -628,6 +798,19 @@ mod tests {
        let (_, is_multipart) = replication_put_object_options("STANDARD", &object_info).expect("build put options");

        assert!(is_multipart, "a composite-checksum multipart object must stay on the multipart transport");
+
+        for (name, metadata) in replication_route_metadata() {
+            let mut legacy = replication_route_object(Some("0123456789abcdef0123456789abcdef"), [4, 0, 4], metadata);
+            legacy.checksum = Some(checksum.to_bytes(&combined));
+            let (_, record_is_multipart) = legacy.decrypt_checksums(0, &HeaderMap::new()).expect("decode checksum");
+            let (_, is_multipart) = replication_put_object_options("STANDARD", &legacy).expect("legacy checksum put options");
+            if legacy.is_encrypted() {
+                assert!(!is_multipart, "{name}: encrypted checksum records must not change the old transport");
+            } else {
+                assert!(record_is_multipart, "the composite checksum must carry its own multipart signal");
+                assert!(is_multipart, "{name}: a composite record can still promote the legacy route to MPU");
+            }
+        }
    }

    #[test]
@@ -1308,12 +1491,63 @@ mod tests {
                ..Default::default()
            };

-            let (opts, _is_multipart) = replication_put_object_options("", &object_info).expect("build replication put options");
+            let (opts, is_multipart) = replication_put_object_options("", &object_info).expect("build replication put options");

+            assert!(!is_multipart, "{name}: a single-part checksum record must keep the single-PUT route");
+            let header = ty.key().expect("every forwarded algorithm has an x-amz-checksum header");
            assert_eq!(
-                opts.user_metadata.get(name),
+                opts.user_metadata.get(header),
                Some(&checksum.encoded),
-                "replication must forward the {name} checksum into user_metadata identically to the classic algorithms"
+                "replication must forward the {name} checksum as the {header} header"
+            );
+            assert!(
+                !opts.user_metadata.contains_key(name),
+                "{name}: the bare algorithm name would leave as x-amz-meta user metadata"
+            );
+        }
+    }
+
+    /// The object-level record of a multipart upload (composite or full-object)
+    /// must not become a PutObject checksum header: the replica is rebuilt
+    /// through CreateMultipartUpload/UploadPart, and a checksum announced there
+    /// that the parts do not carry would be rejected by the target.
+    #[test]
+    fn replication_put_object_options_keeps_multipart_checksum_records_off_the_wire() {
+        let mut composite_type = rustfs_rio::ChecksumType::from_string("crc32");
+        composite_type
+            .merge(rustfs_rio::ChecksumType::MULTIPART)
+            .merge(rustfs_rio::ChecksumType::INCLUDES_MULTIPART);
+        let mut combined = Vec::new();
+        for part in [b"part-one".as_slice(), b"part-two".as_slice()] {
+            let part_checksum =
+                rustfs_rio::Checksum::new_from_data(rustfs_rio::ChecksumType::from_string("crc32"), part).expect("part checksum");
+            combined.extend_from_slice(part_checksum.raw.as_slice());
+        }
+        let composite = rustfs_rio::Checksum::new_from_data(composite_type, &combined)
+            .expect("composite checksum")
+            .to_bytes(&combined);
+
+        for (label, checksum, etag) in [
+            ("composite", composite, "0123456789abcdef0123456789abcdef-2"),
+            (
+                "full-object",
+                full_object_multipart_checksum_record(),
+                "0123456789abcdef0123456789abcdef-3",
+            ),
+        ] {
+            let object_info = ObjectInfo {
+                etag: Some(etag.to_string()),
+                checksum: Some(checksum),
+                ..Default::default()
+            };
+            let (opts, is_multipart) = replication_put_object_options("", &object_info).expect("build replication put options");
+            assert!(is_multipart, "{label}: a multipart object must keep the multipart route");
+            assert!(
+                opts.user_metadata
+                    .keys()
+                    .all(|key| !key.starts_with("x-amz-checksum-") && key != "CRC32"),
+                "{label}: no object-level checksum may reach the target's CreateMultipartUpload: {:?}",
+                opts.user_metadata
            );
        }
    }
@@ -48,7 +48,8 @@ pub use internode_data_transport::build_internode_data_transport_from_env;
 pub(crate) use peer_rest_client::TierConfigReloadOutcome;
 pub use peer_rest_client::{
    KMS_SIGNAL_SUBSYSTEM, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS, PeerRestClient, SERVICE_SIGNAL_REFRESH_CONFIG,
-    SERVICE_SIGNAL_RELOAD_DYNAMIC, ScannerPeerActivity, ScannerPeerDirtyUsageSnapshot, ScannerPublicationLease,
+    SERVICE_SIGNAL_RELOAD_DYNAMIC, ScannerDirtyUsageAcknowledgement, ScannerPeerActivity, ScannerPeerDirtyUsageBucket,
+    ScannerPeerDirtyUsageSnapshot, ScannerPublicationLease, ScannerScopedDirtyUsageAckEntry,
 };
 pub(crate) use peer_s3_client::heal_bucket_local_on_disks;
 pub use peer_s3_client::{
@@ -49,10 +49,11 @@ use rustfs_protos::proto_gen::node_service::{
    LoadRebalanceMetaRequest, LoadServiceAccountRequest, LoadTransitionTierConfigRequest, LoadUserRequest,
    LocalStorageInfoRequest, Mss, ReloadPoolMetaRequest, ReloadSiteReplicationConfigRequest, ReplacementRecoveryStatusRequest,
    ScannerActivityRequest, ScannerActivityResponse, ScannerDirtyUsageSnapshotRequest, ScannerDirtyUsageSnapshotResponse,
-    ScannerPublicationLeaseReleaseRequest, ScannerPublicationLeaseRequest, ScannerPublicationLeaseResponse, ServerInfoRequest,
-    SignalServiceRequest, SignalServiceResponse, StartDecommissionRequest, StartProfilingRequest, StopRebalanceRequest,
-    TierDailyStatsRequest, TierMutationAbortRequest, TierMutationCommitRequest, TierMutationControlResponse,
-    TierMutationFailureClass, TierMutationPeerState, TierMutationPrepareRequest, node_service_client::NodeServiceClient,
+    ScannerPublicationLeaseReleaseRequest, ScannerPublicationLeaseRequest, ScannerPublicationLeaseResponse,
+    ScannerScopedDirtyUsageAckRequest, ScannerScopedDirtyUsageEntry, ServerInfoRequest, SignalServiceRequest,
+    SignalServiceResponse, StartDecommissionRequest, StartProfilingRequest, StopRebalanceRequest, TierDailyStatsRequest,
+    TierMutationAbortRequest, TierMutationCommitRequest, TierMutationControlResponse, TierMutationFailureClass,
+    TierMutationPeerState, TierMutationPrepareRequest, node_service_client::NodeServiceClient,
    tier_mutation_control_service_client::TierMutationControlServiceClient,
 };
 pub use rustfs_protos::{PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS};
@@ -92,6 +93,7 @@ const HEAL_CONTROL_PAYLOAD_MAX_SIZE: usize = 64 * 1024;
 const PEER_REST_RECOVERY_MAX_ATTEMPTS: u32 = 60;
 const PEER_REST_RECOVERY_MAX_BACKOFF: Duration = Duration::from_secs(30);
 const SCANNER_ACTIVITY_MAX_MESSAGE_SIZE: usize = 1024;
+const SCANNER_SCOPED_DIRTY_USAGE_STAGE_TIMEOUT: Duration = Duration::from_secs(5);
 /// Reserve time for the acquire response's network/clock uncertainty.  The
 /// server owns the real expiry; this local deadline is intentionally earlier
 /// so a coordinator never starts a bounded persistence operation at the edge
@@ -192,12 +194,102 @@ pub struct ScannerPeerActivity {

 #[derive(Clone, Debug, PartialEq, Eq)]
 pub struct ScannerPeerDirtyUsageSnapshot {
+    pub owner_id: String,
    pub instance_id: String,
    pub generation: u64,
    pub pending_bucket_count: u64,
    pub protocol_version: u32,
    pub complete: bool,
-    pub buckets: BTreeMap<String, u64>,
+    pub buckets: BTreeMap<String, ScannerPeerDirtyUsageBucket>,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct ScannerPeerDirtyUsageBucket {
+    pub bucket_incarnation: Uuid,
+    pub generation: u64,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct ScannerScopedDirtyUsageAckEntry {
+    pub bucket: String,
+    pub bucket_incarnation: Uuid,
+    pub generation: u64,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub enum ScannerDirtyUsageAcknowledgement {
+    Generation {
+        host: String,
+        instance_id: String,
+        generation: u64,
+    },
+    Scoped {
+        host: String,
+        owner_id: String,
+        instance_id: String,
+        entries: Vec<ScannerScopedDirtyUsageAckEntry>,
+    },
+}
+
+fn scanner_scoped_dirty_usage_ack_payloads(
+    owner_id: String,
+    instance_id: String,
+    probe_only: bool,
+    entries: Vec<ScannerScopedDirtyUsageAckEntry>,
+) -> Result<Vec<ScannerScopedDirtyUsageAckRequest>> {
+    use rustfs_protos::scoped_dirty_usage::*;
+
+    if entries.is_empty() {
+        return Err(Error::other("scoped dirty usage acknowledgement entries must be nonempty"));
+    }
+
+    let mut payloads = Vec::with_capacity(entries.len().div_ceil(SCOPED_DIRTY_USAGE_MAX_ENTRIES as usize));
+    let mut batch = Vec::with_capacity(SCOPED_DIRTY_USAGE_MAX_ENTRIES as usize);
+    for entry in entries {
+        batch.push(ScannerScopedDirtyUsageEntry {
+            bucket: entry.bucket,
+            bucket_incarnation: entry.bucket_incarnation.as_bytes().to_vec().into(),
+            generation: entry.generation,
+        });
+        if batch.len() == SCOPED_DIRTY_USAGE_MAX_ENTRIES as usize {
+            payloads.push(scanner_scoped_dirty_usage_ack_payload(
+                &owner_id,
+                &instance_id,
+                probe_only,
+                std::mem::take(&mut batch),
+            )?);
+        }
+    }
+    if !batch.is_empty() {
+        payloads.push(scanner_scoped_dirty_usage_ack_payload(&owner_id, &instance_id, probe_only, batch)?);
+    }
+
+    Ok(payloads)
+}
+
+fn scanner_scoped_dirty_usage_ack_payload(
+    owner_id: &str,
+    instance_id: &str,
+    probe_only: bool,
+    entries: Vec<ScannerScopedDirtyUsageEntry>,
+) -> Result<ScannerScopedDirtyUsageAckRequest> {
+    use rustfs_protos::scoped_dirty_usage::*;
+
+    let payload = ScannerScopedDirtyUsageAckRequest {
+        challenge: Uuid::new_v4().as_bytes().to_vec().into(),
+        protocol_version: SCOPED_DIRTY_USAGE_PROTOCOL_VERSION,
+        owner_id: owner_id.to_string(),
+        instance_id: instance_id.to_string(),
+        scope: SCOPED_DIRTY_USAGE_BUCKET_SCOPE,
+        probe_only,
+        entries,
+    };
+    canonical_scoped_dirty_usage_request(&payload).map_err(|err| Error::other(err.to_string()))?;
+    Ok(payload)
+}
+
+fn scanner_scoped_dirty_usage_ack_reconciled(activity: &ScannerPeerActivity, expected_instance_id: &str) -> bool {
+    activity.instance_id == expected_instance_id && activity.dirty_usage_pending == Some(false)
 }

 fn scanner_instance_id_is_valid(instance_id: &str) -> bool {
@@ -351,6 +443,11 @@ fn decode_scanner_dirty_usage_snapshot_with_verifier(
    if !scanner_instance_id_is_valid(&response.instance_id) {
        return Err(Error::other("peer returned an invalid scanner dirty usage snapshot instance ID"));
    }
+    let owner_id = Uuid::parse_str(&response.owner_id)
+        .ok()
+        .filter(|owner_id| !owner_id.is_nil())
+        .map(|owner_id| owner_id.to_string())
+        .ok_or_else(|| Error::other("peer returned an invalid scanner dirty usage snapshot owner"))?;
    if response.generation == u64::MAX {
        return Err(Error::other("peer scanner dirty usage snapshot exhausted its generation"));
    }
@@ -386,9 +483,14 @@ fn decode_scanner_dirty_usage_snapshot_with_verifier(
        if bucket.generation == 0 || bucket.generation > response.generation {
            return Err(Error::other("peer scanner dirty usage snapshot contains an invalid bucket generation"));
        }
+        Uuid::from_slice(bucket.bucket_incarnation.as_ref())
+            .ok()
+            .filter(|bucket_incarnation| !bucket_incarnation.is_nil())
+            .ok_or_else(|| Error::other("peer scanner dirty usage snapshot contains an invalid bucket incarnation"))?;
    }

    Ok(ScannerPeerDirtyUsageSnapshot {
+        owner_id,
        instance_id: response.instance_id,
        generation: response.generation,
        pending_bucket_count: response.pending_bucket_count,
@@ -397,7 +499,16 @@ fn decode_scanner_dirty_usage_snapshot_with_verifier(
        buckets: response
            .buckets
            .into_iter()
-            .map(|bucket| (bucket.bucket, bucket.generation))
+            .map(|bucket| {
+                (
+                    bucket.bucket,
+                    ScannerPeerDirtyUsageBucket {
+                        bucket_incarnation: Uuid::from_slice(bucket.bucket_incarnation.as_ref())
+                            .expect("bucket incarnation was validated"),
+                        generation: bucket.generation,
+                    },
+                )
+            })
            .collect(),
    })
 }
@@ -1688,6 +1799,24 @@ impl PeerRestClient {
        Ok((self.topology_member.clone(), supported_version, epoch))
    }

+    pub async fn probe_ilm_recovery_export(&self, topology_fingerprint: String) -> Result<(String, Uuid)> {
+        let probe = rustfs_protos::ilm_recovery_export_capability_probe(Uuid::new_v4().as_bytes());
+        let result = self
+            .heal_control(rustfs_protos::HEAL_CONTROL_PROTOCOL_VERSION, topology_fingerprint, probe)
+            .await?;
+        let epoch = decode_remote_version_state_capability(&self.topology_member, &result)?;
+        Ok((self.topology_member.clone(), epoch))
+    }
+
+    pub async fn probe_transition_transaction_compaction(&self, topology_fingerprint: String) -> Result<(String, Uuid)> {
+        let probe = rustfs_protos::transition_transaction_compaction_capability_probe(Uuid::new_v4().as_bytes());
+        let result = self
+            .heal_control(rustfs_protos::HEAL_CONTROL_PROTOCOL_VERSION, topology_fingerprint, probe)
+            .await?;
+        let epoch = decode_remote_version_state_capability(&self.topology_member, &result)?;
+        Ok((self.topology_member.clone(), epoch))
+    }
+
    pub async fn load_bucket_metadata(&self, bucket: &str, scanner_maintenance_change: bool) -> Result<()> {
        let result = tokio::time::timeout(BUCKET_METADATA_RELOAD_TIMEOUT, async {
            let result = self.load_bucket_metadata_once(bucket, scanner_maintenance_change).await;
@@ -2068,19 +2197,10 @@ impl PeerRestClient {
        &self,
        owner_id: String,
        instance_id: String,
-        entries: Vec<rustfs_protos::proto_gen::node_service::ScannerScopedDirtyUsageEntry>,
+        entries: Vec<ScannerScopedDirtyUsageAckEntry>,
    ) -> Result<bool> {
        use rustfs_protos::scoped_dirty_usage::*;
-        let payload = rustfs_protos::proto_gen::node_service::ScannerScopedDirtyUsageAckRequest {
-            challenge: Uuid::new_v4().as_bytes().to_vec().into(),
-            protocol_version: SCOPED_DIRTY_USAGE_PROTOCOL_VERSION,
-            owner_id,
-            instance_id,
-            scope: SCOPED_DIRTY_USAGE_BUCKET_SCOPE,
-            probe_only: true,
-            entries,
-        };
-        let canonical = canonical_scoped_dirty_usage_request(&payload).map_err(|err| Error::other(err.to_string()))?;
+        let payloads = scanner_scoped_dirty_usage_ack_payloads(owner_id, instance_id, true, entries)?;
        self.finalize_result(
            async {
                let mut client = super::client::scanner_control_time_out_client(
@@ -2088,26 +2208,106 @@ impl PeerRestClient {
                    TonicInterceptor::Signature(gen_tonic_signature_interceptor()),
                )
                .await?;
+                for payload in payloads {
+                    let canonical =
+                        canonical_scoped_dirty_usage_request(&payload).map_err(|err| Error::other(err.to_string()))?;
+                    let mut request = Request::new(payload.clone());
+                    set_tonic_canonical_body_digest(&mut request, &canonical)?;
+                    let response = client.scanner_scoped_dirty_usage_ack(request).await?.into_inner();
+                    let body = canonical_scoped_dirty_usage_response(&canonical, &response)
+                        .map_err(|_| Error::other("scoped dirty usage capability response is too large"))?;
+                    verify_tonic_rpc_response_proof(&body, response.response_proof.as_ref())?;
+                    if response.protocol_version != SCOPED_DIRTY_USAGE_PROTOCOL_VERSION
+                        || response.owner_id != payload.owner_id
+                        || response.instance_id != payload.instance_id
+                        || response.max_entries != SCOPED_DIRTY_USAGE_MAX_ENTRIES
+                        || response.max_request_bytes != SCOPED_DIRTY_USAGE_MAX_REQUEST_BYTES
+                        || response.cleared != 0
+                    {
+                        return Err(Error::other("scoped dirty usage capability response does not match request"));
+                    }
+                    if !response.supported {
+                        return Ok(false);
+                    }
+                }
+                Ok(true)
+            }
+            .await,
+        )
+        .await
+    }
+
+    pub async fn acknowledge_scanner_scoped_dirty_usage(
+        &self,
+        owner_id: String,
+        instance_id: String,
+        entries: Vec<ScannerScopedDirtyUsageAckEntry>,
+    ) -> Result<ScannerPeerActivity> {
+        use rustfs_protos::scoped_dirty_usage::*;
+        let payloads = scanner_scoped_dirty_usage_ack_payloads(owner_id, instance_id.clone(), false, entries)?;
+        let ack_attempt = async {
+            let mut client = super::client::scanner_control_time_out_client(
+                &self.grid_host,
+                TonicInterceptor::Signature(gen_tonic_signature_interceptor()),
+            )
+            .await?;
+            for payload in payloads {
+                let canonical = canonical_scoped_dirty_usage_request(&payload).map_err(|err| Error::other(err.to_string()))?;
                let mut request = Request::new(payload.clone());
                set_tonic_canonical_body_digest(&mut request, &canonical)?;
                let response = client.scanner_scoped_dirty_usage_ack(request).await?.into_inner();
                let body = canonical_scoped_dirty_usage_response(&canonical, &response)
-                    .map_err(|_| Error::other("scoped dirty usage capability response is too large"))?;
+                    .map_err(|_| Error::other("scoped dirty usage acknowledgement response is too large"))?;
                verify_tonic_rpc_response_proof(&body, response.response_proof.as_ref())?;
                if response.protocol_version != SCOPED_DIRTY_USAGE_PROTOCOL_VERSION
                    || response.owner_id != payload.owner_id
                    || response.instance_id != payload.instance_id
                    || response.max_entries != SCOPED_DIRTY_USAGE_MAX_ENTRIES
                    || response.max_request_bytes != SCOPED_DIRTY_USAGE_MAX_REQUEST_BYTES
-                    || response.cleared != 0
+                    || !response.supported
                {
-                    return Err(Error::other("scoped dirty usage capability response does not match request"));
+                    return Err(Error::other("scoped dirty usage acknowledgement response does not match request"));
                }
-                Ok(response.supported)
            }
-            .await,
-        )
-        .await
+            Ok(())
+        };
+        let result = match timeout(SCANNER_SCOPED_DIRTY_USAGE_STAGE_TIMEOUT, ack_attempt).await {
+            Ok(result) => self.finalize_result(result).await,
+            Err(_) => {
+                self.prepare_retry_with_timeout(SCANNER_SCOPED_DIRTY_USAGE_STAGE_TIMEOUT)
+                    .await;
+                Err(Error::other("scoped dirty usage acknowledgement deadline elapsed"))
+            }
+        };
+
+        match result {
+            Ok(()) => {
+                let activity = self.scanner_scoped_dirty_usage_activity_confirmation().await?;
+                if activity.instance_id == instance_id {
+                    Ok(activity)
+                } else {
+                    Err(Error::other(
+                        "scoped dirty usage acknowledgement peer restarted before activity confirmation",
+                    ))
+                }
+            }
+            Err(err) => {
+                if Self::is_network_like_error(&err) {
+                    self.prepare_retry_with_timeout(SCANNER_SCOPED_DIRTY_USAGE_STAGE_TIMEOUT)
+                        .await;
+                }
+                match self.scanner_scoped_dirty_usage_activity_confirmation().await {
+                    Ok(activity) if scanner_scoped_dirty_usage_ack_reconciled(&activity, &instance_id) => Ok(activity),
+                    _ => Err(err),
+                }
+            }
+        }
+    }
+
+    async fn scanner_scoped_dirty_usage_activity_confirmation(&self) -> Result<ScannerPeerActivity> {
+        timeout(SCANNER_SCOPED_DIRTY_USAGE_STAGE_TIMEOUT, self.scanner_activity())
+            .await
+            .map_err(|_| Error::other("scoped dirty usage activity confirmation timed out"))?
    }

    pub async fn acknowledge_scanner_dirty_usage(&self, instance_id: String, generation: u64) -> Result<ScannerPeerActivity> {
@@ -2836,16 +3036,79 @@ mod tests {
                rustfs_protos::proto_gen::node_service::ScannerDirtyUsageBucket {
                    bucket: "archive".to_string(),
                    generation: 3,
+                    bucket_incarnation: Uuid::from_u128(0x11111111111111111111111111111111).as_bytes().to_vec().into(),
                },
                rustfs_protos::proto_gen::node_service::ScannerDirtyUsageBucket {
                    bucket: "photos".to_string(),
                    generation: 7,
+                    bucket_incarnation: Uuid::from_u128(0x22222222222222222222222222222222).as_bytes().to_vec().into(),
                },
            ],
            response_proof: b"proof".to_vec().into(),
+            owner_id: "33333333-3333-3333-3333-333333333333".to_string(),
        }
    }

+    #[test]
+    fn scanner_scoped_dirty_usage_ack_payloads_split_at_protocol_limit() {
+        use rustfs_protos::scoped_dirty_usage::{SCOPED_DIRTY_USAGE_MAX_ENTRIES, canonical_scoped_dirty_usage_request};
+
+        let entries = (0..=SCOPED_DIRTY_USAGE_MAX_ENTRIES)
+            .map(|index| ScannerScopedDirtyUsageAckEntry {
+                bucket: format!("bucket-{index:02}"),
+                bucket_incarnation: Uuid::from_u128(0x11111111111111111111111111111111),
+                generation: 9,
+            })
+            .collect::<Vec<_>>();
+
+        let payloads = scanner_scoped_dirty_usage_ack_payloads(
+            "33333333-3333-3333-3333-333333333333".to_string(),
+            "0123456789abcdef0123456789abcdef".to_string(),
+            false,
+            entries,
+        )
+        .expect("33 entries should split into valid scoped dirty usage requests");
+
+        assert_eq!(payloads.len(), 2);
+        assert_eq!(payloads[0].entries.len(), SCOPED_DIRTY_USAGE_MAX_ENTRIES as usize);
+        assert_eq!(payloads[1].entries.len(), 1);
+        assert_eq!(payloads[0].entries.first().map(|entry| entry.bucket.as_str()), Some("bucket-00"));
+        assert_eq!(payloads[0].entries.last().map(|entry| entry.bucket.as_str()), Some("bucket-31"));
+        assert_eq!(payloads[1].entries.first().map(|entry| entry.bucket.as_str()), Some("bucket-32"));
+        for payload in payloads {
+            canonical_scoped_dirty_usage_request(&payload).expect("each split scoped ACK payload should be canonical");
+        }
+    }
+
+    #[test]
+    fn scanner_scoped_dirty_usage_ack_reconciliation_requires_same_clean_instance() {
+        let activity = |instance_id: &str, pending| ScannerPeerActivity {
+            instance_id: instance_id.to_string(),
+            namespace_generation: 1,
+            maintenance_generation: 1,
+            protocol_version: SCANNER_ACTIVITY_PROTOCOL_VERSION,
+            topology_digest: Some([1; 32]),
+            data_movement_active: Some(false),
+            dirty_usage_generation: Some(9),
+            dirty_usage_pending: pending,
+            movement_generation: Some(1),
+            publication_blocked: Some(false),
+        };
+
+        assert!(scanner_scoped_dirty_usage_ack_reconciled(
+            &activity("0123456789abcdef0123456789abcdef", Some(false)),
+            "0123456789abcdef0123456789abcdef"
+        ));
+        assert!(!scanner_scoped_dirty_usage_ack_reconciled(
+            &activity("0123456789abcdef0123456789abcdef", Some(true)),
+            "0123456789abcdef0123456789abcdef"
+        ));
+        assert!(!scanner_scoped_dirty_usage_ack_reconciled(
+            &activity("fedcba9876543210fedcba9876543210", Some(false)),
+            "0123456789abcdef0123456789abcdef"
+        ));
+    }
+
    #[test]
    fn scanner_dirty_usage_snapshot_requires_a_complete_authenticated_ordered_view() {
        let decoded = decode_test_scanner_dirty_usage_snapshot(test_scanner_dirty_usage_snapshot_response())
@@ -2854,9 +3117,18 @@ mod tests {
        assert_eq!(decoded.generation, 7);
        assert_eq!(decoded.pending_bucket_count, 2);
        assert_eq!(decoded.protocol_version, SCANNER_DIRTY_USAGE_SNAPSHOT_PROTOCOL_VERSION);
+        assert_eq!(decoded.owner_id, "33333333-3333-3333-3333-333333333333");
        assert!(decoded.complete);
-        assert_eq!(decoded.buckets.get("archive"), Some(&3));
-        assert_eq!(decoded.buckets.get("photos"), Some(&7));
+        assert_eq!(
+            decoded.buckets.get("archive").map(|bucket| bucket.bucket_incarnation),
+            Some(Uuid::from_u128(0x11111111111111111111111111111111))
+        );
+        assert_eq!(decoded.buckets.get("archive").map(|bucket| bucket.generation), Some(3));
+        assert_eq!(
+            decoded.buckets.get("photos").map(|bucket| bucket.bucket_incarnation),
+            Some(Uuid::from_u128(0x22222222222222222222222222222222))
+        );
+        assert_eq!(decoded.buckets.get("photos").map(|bucket| bucket.generation), Some(7));

        let overflow_count =
            u64::try_from(SCANNER_DIRTY_USAGE_SNAPSHOT_MAX_ENTRIES + 1).expect("the test snapshot entry limit should fit in u64");
@@ -2907,6 +3179,14 @@ mod tests {
        empty_bucket.buckets[0].bucket.clear();
        cases.push((empty_bucket, "empty bucket name"));

+        let mut invalid_owner = test_scanner_dirty_usage_snapshot_response();
+        invalid_owner.owner_id.clear();
+        cases.push((invalid_owner, "snapshot owner"));
+
+        let mut invalid_incarnation = test_scanner_dirty_usage_snapshot_response();
+        invalid_incarnation.buckets[0].bucket_incarnation = Uuid::nil().as_bytes().to_vec().into();
+        cases.push((invalid_incarnation, "bucket incarnation"));
+
        let mut partial = test_scanner_dirty_usage_snapshot_response();
        partial.complete = false;
        cases.push((partial, "entry-limit overflow"));
@@ -2919,6 +3199,7 @@ mod tests {
                .map(|index| rustfs_protos::proto_gen::node_service::ScannerDirtyUsageBucket {
                    bucket: format!("bucket-{index:04}"),
                    generation: 1,
+                    bucket_incarnation: Uuid::from_u128(0x11111111111111111111111111111111).as_bytes().to_vec().into(),
                })
                .collect(),
            ..test_scanner_dirty_usage_snapshot_response()
@@ -449,14 +449,6 @@ where
    Ok(data)
 }

-pub(crate) async fn read_config_limited<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<Vec<u8>>
-where
-    S: EcstoreObjectIO,
-{
-    let (data, _obj) = read_config_with_metadata_inner(api, file, &ObjectOptions::default(), false, Some(max_bytes)).await?;
-    Ok(data)
-}
-
 pub(crate) async fn read_config_limited_preserve_empty<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<Vec<u8>>
 where
    S: EcstoreObjectIO,
@@ -465,6 +457,14 @@ where
    Ok(data)
 }

+pub(crate) async fn read_config_limited<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<Vec<u8>>
+where
+    S: EcstoreObjectIO,
+{
+    let (data, _obj) = read_config_with_metadata_inner(api, file, &ObjectOptions::default(), false, Some(max_bytes)).await?;
+    Ok(data)
+}
+
 pub(crate) async fn read_config_limited_preserve_empty_with_metadata<S>(
    api: Arc<S>,
    file: &str,
@@ -476,6 +476,18 @@ where
    read_config_with_metadata_inner(api, file, &ObjectOptions::default(), true, Some(max_bytes)).await
 }

+pub(crate) async fn read_config_limited_preserve_empty_with_metadata_opts<S>(
+    api: Arc<S>,
+    file: &str,
+    opts: &ObjectOptions,
+    max_bytes: usize,
+) -> Result<(Vec<u8>, ObjectInfo)>
+where
+    S: EcstoreObjectIO,
+{
+    read_config_with_metadata_inner(api, file, opts, true, Some(max_bytes)).await
+}
+
 /// Read an existing config object without treating an empty payload as absent.
 /// Callers that validate their own payload format need to distinguish corruption
 /// from `ConfigNotFound`.
@@ -3385,7 +3385,7 @@ pub(crate) async fn acquire_pool_activation_fleet_proof(
        .ok_or_else(|| Error::other(POOL_ACTIVATION_FLEET_PROOF_REQUIRED))
 }

-pub(crate) fn is_pool_activation_fleet_proof_error(err: &Error) -> bool {
+pub fn is_pool_activation_fleet_proof_error(err: &Error) -> bool {
    // Save-stage helpers add context by formatting the original error, so the
    // marker may be nested in the display string. Restrict matching to the
    // `Error::other` I/O shape used by this activation path.
@@ -5108,49 +5108,7 @@ async fn read_pool_meta_replicas<S>(pools: Vec<Arc<S>>, no_lock: bool) -> Vec<Po
 where
    S: EcstoreObjectIO,
 {
-    let reads = join_all(pools.into_iter().map(|pool| read_pool_meta_replica(pool, no_lock))).await;
-    #[cfg(feature = "e2e-test-hooks")]
-    if STARTUP_CAS_OBSERVATION.try_with(|_| ()).is_ok() {
-        let batch = uuid::Uuid::new_v4();
-        for (pool, read) in reads.iter().enumerate() {
-            let mut observation = serde_json::json!({
-                "kind": "replica-read", "object": POOL_META_NAME, "batch": batch, "pool": pool,
-                "cas": match &read.cas {
-                    PoolMetaCasToken::Missing => "missing",
-                    PoolMetaCasToken::Existing(_) => "existing",
-                    PoolMetaCasToken::Unsafe => "unsafe",
-                },
-                "etag": match &read.cas { PoolMetaCasToken::Existing(etag) => Some(etag), _ => None },
-            });
-            match &read.replica {
-                PoolMetaReplica::Valid {
-                    raw,
-                    canonical,
-                    meta,
-                    revision,
-                    committed,
-                    ..
-                } => {
-                    observation["state"] = serde_json::json!("valid");
-                    observation["committed"] = serde_json::json!(committed);
-                    observation["version"] = serde_json::json!(revision.version);
-                    observation["cluster_id"] = serde_json::json!(revision.cluster_id);
-                    observation["epoch"] = serde_json::json!(revision.epoch);
-                    observation["generation"] = serde_json::json!(revision.generation);
-                    observation["transaction_id"] = serde_json::json!(revision.transaction_id);
-                    observation["pool_count"] = serde_json::json!(meta.pools.len());
-                    observation["payload_sha256"] = serde_json::json!(rustfs_utils::crypto::hex(Sha256::digest(canonical)));
-                    observation["raw_sha256"] = serde_json::json!(rustfs_utils::crypto::hex(Sha256::digest(raw)));
-                }
-                PoolMetaReplica::Missing => observation["state"] = serde_json::json!("missing"),
-                PoolMetaReplica::Corrupt(_) => observation["state"] = serde_json::json!("corrupt"),
-                PoolMetaReplica::Incompatible(_) => observation["state"] = serde_json::json!("incompatible"),
-                PoolMetaReplica::Unreadable(_) => observation["state"] = serde_json::json!("unreadable"),
-            }
-            startup_cas_test_observe(observation);
-        }
-    }
-    reads
+    join_all(pools.into_iter().map(|pool| read_pool_meta_replica(pool, no_lock))).await
 }

 fn select_pool_meta_replicas_observing<R>(write_state: &mut PoolMetaWriteState, replicas: Vec<R>) -> Result<PoolMetaSelection>
@@ -5522,60 +5480,6 @@ fn pool_meta_cas_preconditions(token: &PoolMetaCasToken, object: &str) -> Result
    }
 }

-#[cfg(feature = "e2e-test-hooks")]
-struct StartupCasObservation {
-    attempt: uuid::Uuid,
-    phase: &'static str,
-    pools: Vec<usize>,
-}
-
-#[cfg(feature = "e2e-test-hooks")]
-tokio::task_local! {
-    static STARTUP_CAS_OBSERVATION: StartupCasObservation;
-}
-
-// This scope follows only the directly polled startup future. Spawned work
-// does not inherit it; receiver evidence retains its existing RPC tuple.
-#[cfg(feature = "e2e-test-hooks")]
-pub(crate) async fn startup_cas_test_scope<S, F: std::future::Future>(
-    attempt: uuid::Uuid,
-    phase: &'static str,
-    pools: &[Arc<S>],
-    future: F,
-) -> F::Output {
-    STARTUP_CAS_OBSERVATION
-        .scope(
-            StartupCasObservation {
-                attempt,
-                phase,
-                // These identities are never dereferenced or logged. The
-                // caller and operation keep the same pool Arcs alive.
-                pools: pools.iter().map(|pool| Arc::as_ptr(pool) as usize).collect(),
-            },
-            future,
-        )
-        .await
-}
-
-// Direct JSON diagnostics are independent of the startup tracing subscriber.
-#[cfg(feature = "e2e-test-hooks")]
-pub(crate) fn startup_cas_test_observe(mut observation: serde_json::Value) {
-    let Some(nonce) = std::env::var("RUSTFS_E2E_STARTUP_CAS_NONCE")
-        .ok()
-        .and_then(|value| uuid::Uuid::parse_str(&value).ok())
-    else {
-        return;
-    };
-    observation["nonce"] = serde_json::json!(nonce);
-    observation["pid"] = serde_json::json!(std::process::id());
-    let _ = STARTUP_CAS_OBSERVATION.try_with(|scope| {
-        observation["attempt"] = serde_json::json!(scope.attempt);
-        observation["startup_phase"] = serde_json::json!(scope.phase);
-    });
-    let line = format!("RUSTFS_E2E_STARTUP_CAS {observation}\n");
-    let _ = std::io::Write::write_all(&mut std::io::stderr().lock(), line.as_bytes());
-}
-
 async fn save_pool_meta_object_cas<S>(
    pool: Arc<S>,
    object: &str,
@@ -5596,43 +5500,13 @@ where
        ..Default::default()
    };
    fence.add_to_options(&mut opts);
-    #[cfg(feature = "e2e-test-hooks")]
-    let observation = std::env::var_os("RUSTFS_E2E_STARTUP_CAS_NONCE").map(|_| {
-        serde_json::json!({
-            "kind": "cas", "object": object, "phase": phase,
-            "pool": STARTUP_CAS_OBSERVATION.try_with(|scope| {
-                scope.pools.iter().position(|identity| *identity == Arc::as_ptr(&pool) as usize)
-            }).ok().flatten(),
-            "payload_sha256": rustfs_utils::crypto::hex(Sha256::digest(&data)),
-            "if_match": opts.http_preconditions.as_ref().and_then(|p| p.if_match.as_deref()),
-            "if_none_match": opts.http_preconditions.as_ref().and_then(|p| p.if_none_match.as_deref()),
-            "tail_drained": opts.write_completion == crate::object_api::WriteCompletion::TailDrained,
-            "no_lock": opts.no_lock,
-        })
-    });
    let result = save_config_with_opts_and_metadata(pool, object, data, &opts).await;
    if matches!(&result, Err(Error::PreconditionFailed)) {
        record_pool_meta_stale_write_rejection(phase);
    }
-    let result = result.and_then(|object_info| {
-        fence.ensure_held()?;
-        Ok(object_info)
-    });
-    #[cfg(feature = "e2e-test-hooks")]
-    if let Some(mut observation) = observation {
-        observation["ok"] = serde_json::json!(result.is_ok());
-        observation["etag"] = serde_json::json!(result.as_ref().ok().and_then(|info| info.etag.as_deref()));
-        observation["mod_time"] = serde_json::json!(
-            result
-                .as_ref()
-                .ok()
-                .and_then(|info| info.mod_time)
-                .map(|time| time.unix_timestamp_nanos().to_string())
-        );
-        observation["error"] = serde_json::json!(result.as_ref().err().map(ToString::to_string));
-        startup_cas_test_observe(observation);
-    }
-    result
+    let object_info = result?;
+    fence.ensure_held()?;
+    Ok(object_info)
 }

 async fn persist_pool_meta_identity<S>(
@@ -6931,13 +6805,6 @@ impl PoolMeta {
        };
        if confirmed.revision == revision && confirmed.canonical.as_ref() == Some(&durable) {
            persist_pool_meta_identity(pools, write_state, true, fence).await?;
-            #[cfg(feature = "e2e-test-hooks")]
-            startup_cas_test_observe(serde_json::json!({
-                "kind": "confirmed", "object": POOL_META_NAME,
-                "payload_sha256": rustfs_utils::crypto::hex(Sha256::digest(&durable)),
-                "generation": confirmed.revision.generation,
-                "transaction_id": confirmed.revision.transaction_id,
-            }));
            return Ok(confirmed.meta);
        }
        if !commit_succeeded {
@@ -2673,7 +2673,7 @@ mod tests {
            ]),
            ..Default::default()
        };
-        assert!(!object_info.is_multipart());
+        assert!(object_info.is_multipart());
        assert!(should_use_multipart_data_movement(&object_info, false));

        let single_nonstandard_part = ObjectInfo {
@@ -3050,7 +3050,7 @@ mod tests {
            ..Default::default()
        };

-        assert!(!object_info.is_multipart());
+        assert!(object_info.is_multipart());
        assert!(object_info.parts.iter().any(|part| part.checksums.is_some()));
        let opts = data_movement_put_object_opts(&object_info, 0);
        assert!(!rustfs_utils::http::contains_key_str(&opts.user_defined, SUFFIX_PART_CHECKSUMS));
@@ -324,6 +324,46 @@ impl DiskStoreRenameDataExt for LocalDiskWrapper {
 }

 impl LocalDiskWrapper {
+    pub(in crate::disk) async fn delete_version_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        fi: FileInfo,
+        force_del_marker: bool,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        self.track_disk_health_mutation(
+            "delete_version",
+            DiskMetricMutation::Delete,
+            || async {
+                Box::pin(
+                    self.disk
+                        .delete_version_with_namespace_owner(volume, path, fi, force_del_marker, opts, namespace_owner),
+                )
+                .await
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    pub(in crate::disk) async fn delete_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        self.track_disk_health_mutation(
+            "delete",
+            DiskMetricMutation::Delete,
+            || async { Box::pin(self.disk.delete_with_namespace_owner(volume, path, opts, namespace_owner)).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
    pub(in crate::disk) async fn undo_write_with_namespace_owner(
        &self,
        volume: &str,
@@ -191,11 +191,33 @@ fn restore_part_transaction_file(current: &Path, backup: &Path, absent: &Path, r
 }

 async fn write_metadata_rollback_backup(object_dir: &Path, rollback_dir: Uuid, data: &[u8]) -> Result<()> {
+    write_delete_rollback_file(object_dir, rollback_dir, STORAGE_FORMAT_FILE_BACKUP, data, None).await
+}
+
+async fn write_delete_rollback_file(
+    object_dir: &Path,
+    rollback_dir: Uuid,
+    name: &str,
+    data: &[u8],
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> Result<()> {
    let backup_dir = object_dir.join(rollback_dir.to_string());
-    fs::create_dir_all(&backup_dir).await.map_err(to_file_error)?;
-    fs::write(backup_dir.join(STORAGE_FORMAT_FILE_BACKUP), data)
-        .await
-        .map_err(to_file_error)?;
+    let path = backup_dir.join(name);
+    if namespace_owner.is_none() {
+        fs::create_dir_all(&backup_dir).await.map_err(to_file_error)?;
+        fs::write(path, data).await.map_err(to_file_error)?;
+        return Ok(());
+    }
+    let lease = os::acquire_namespace_mutation_lease_with_owner(&path, namespace_owner).await;
+    let data = data.to_vec();
+    os::run_blocking_namespace_operation(lease, move || {
+        std::fs::create_dir_all(&backup_dir)?;
+        #[cfg(test)]
+        run_owned_file_write_before_open(&path);
+        std::fs::write(path, data)
+    })
+    .await
+    .map_err(to_file_error)?;
    Ok(())
 }

@@ -342,6 +364,7 @@ struct DeleteVersionMutation {
 struct DeleteRollbackFailure {
    stage: &'static str,
    error: DiskError,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
 }

 async fn restore_delete_rollback_after_error(
@@ -353,12 +376,18 @@ async fn restore_delete_rollback_after_error(
    failure: DeleteRollbackFailure,
    publication_root: &os::PublicationRoot,
 ) -> DiskError {
-    let DeleteRollbackFailure { stage, error } = failure;
+    let DeleteRollbackFailure {
+        stage,
+        error,
+        namespace_owner,
+    } = failure;
    let Some(rollback_dir) = rollback_dir else {
        return error;
    };

-    if let Err(restore_err) = restore_delete_rollback(object_dir, xl_path, rollback_dir, publication_root).await {
+    if let Err(restore_err) =
+        restore_delete_rollback_with_namespace_owner(object_dir, xl_path, rollback_dir, publication_root, namespace_owner).await
+    {
        warn!(
            volume,
            path,
@@ -5654,6 +5683,123 @@ impl LocalDisk {
    //     })
    // }

+    #[tracing::instrument(name = "delete_version", level = "trace", skip_all)]
+    pub(in crate::disk) async fn delete_version_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        fi: FileInfo,
+        force_del_marker: bool,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        self.delete_version_inner(
+            volume,
+            path,
+            fi,
+            DeleteVersionMutation {
+                force_del_marker,
+                opts,
+                namespace_owner,
+            },
+        )
+        .await
+    }
+
+    #[tracing::instrument(name = "write_metadata", level = "trace", skip_all)]
+    async fn write_metadata_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        fi: FileInfo,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        crate::hp_guard!("LocalDisk::write_metadata");
+        fi.validate_for_metadata_read()?;
+        let p = self.io_get_object_path(volume, format!("{path}/{STORAGE_FORMAT_FILE}").as_str())?;
+
+        let mut meta = FileMeta::new();
+        if !fi.fresh {
+            let (buf, _) = read_file_exists(&p).await?;
+            if !buf.is_empty() {
+                let _ = meta.unmarshal_msg(&buf).map_err(|_| {
+                    meta = FileMeta::new();
+                });
+            }
+        }
+
+        meta.add_version(fi)?;
+
+        let fm_data = meta.marshal_msg()?;
+
+        // Atomic temp+rename: this path also rewrites live xl.meta (delete markers,
+        // decommission), where an in-place truncate would expose torn metadata.
+        self.write_all_meta_with_namespace_owner(
+            volume,
+            format!("{path}/{STORAGE_FORMAT_FILE}").as_str(),
+            &fm_data,
+            true,
+            namespace_owner,
+        )
+        .await?;
+
+        Ok(())
+    }
+
+    async fn delete_data_dir_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<DataDirDeleteStatus> {
+        let key = SnapshotLeaseKey {
+            volume: volume.to_string(),
+            path: path.to_string(),
+        };
+        {
+            let mut registry = self.snapshot_leases.lock().await;
+            if let Some(entry) = registry.entries.get_mut(&key) {
+                if !entry.tokens.is_empty() {
+                    entry.pending_delete.get_or_insert_with(|| opts.clone());
+                    return Ok(DataDirDeleteStatus::Deferred);
+                }
+                if entry.deleting {
+                    entry.pending_delete.get_or_insert_with(|| opts.clone());
+                    return Ok(DataDirDeleteStatus::Deferred);
+                }
+                entry.deleting = true;
+                entry.pending_delete.get_or_insert_with(|| opts.clone());
+            } else {
+                registry.entries.insert(
+                    key.clone(),
+                    SnapshotLeaseEntry {
+                        pending_delete: Some(opts.clone()),
+                        deleting: true,
+                        ..Default::default()
+                    },
+                );
+            }
+        }
+
+        let result = self
+            .delete_unleased_with_namespace_owner(volume, path, &opts, namespace_owner)
+            .await;
+        let mut registry = self.snapshot_leases.lock().await;
+        match result {
+            Ok(()) => {
+                registry.entries.remove(&key);
+                Ok(DataDirDeleteStatus::Deleted)
+            }
+            Err(err) => {
+                if let Some(entry) = registry.entries.get_mut(&key) {
+                    entry.deleting = false;
+                }
+                Err(err)
+            }
+        }
+    }
+
    async fn delete_version_inner(&self, volume: &str, path: &str, fi: FileInfo, mutation: DeleteVersionMutation) -> Result<()> {
        let DeleteVersionMutation {
            force_del_marker,
@@ -5696,7 +5842,7 @@ impl LocalDisk {

                if fi.deleted && force_del_marker {
                    return self
-                        .write_missing_delete_marker(volume, path, fi, file_path.as_path(), &xl_path, rollback_dir)
+                        .write_missing_delete_marker(volume, path, fi, file_path.as_path(), rollback_dir, namespace_owner.clone())
                        .await;
                }

@@ -5712,7 +5858,14 @@ impl LocalDisk {
        let old_dir = meta.delete_version(&fi)?;
        let mut reserved_version_delete = false;
        if let Some(rollback_dir) = rollback_dir {
-            write_metadata_rollback_backup(file_path.as_path(), rollback_dir, &buf).await?;
+            write_delete_rollback_file(
+                file_path.as_path(),
+                rollback_dir,
+                STORAGE_FORMAT_FILE_BACKUP,
+                &buf,
+                namespace_owner.clone(),
+            )
+            .await?;
        }

        if let Some(uuid) = old_dir {
@@ -5728,6 +5881,7 @@ impl LocalDisk {
                    DeleteRollbackFailure {
                        stage: "delete_version_metadata_update",
                        error: err,
+                        namespace_owner: namespace_owner.clone(),
                    },
                    &self.publication_root,
                )
@@ -5745,6 +5899,7 @@ impl LocalDisk {
                    DeleteRollbackFailure {
                        stage: "delete_version_data_path",
                        error: err,
+                        namespace_owner: namespace_owner.clone(),
                    },
                    &self.publication_root,
                )
@@ -5753,7 +5908,7 @@ impl LocalDisk {

            if let Some(rollback_dir) = rollback_dir {
                let rollback_path = file_path.join(rollback_dir.to_string());
-                if let Err(err) = fs::create_dir_all(&rollback_path).await {
+                if let Err(err) = os::create_dir_all_with_namespace_owner(&rollback_path, namespace_owner.clone()).await {
                    let err: DiskError = to_file_error(err).into();
                    return Err(restore_delete_rollback_after_error(
                        file_path.as_path(),
@@ -5764,12 +5919,16 @@ impl LocalDisk {
                        DeleteRollbackFailure {
                            stage: "delete_version_rollback_dir",
                            error: err,
+                            namespace_owner: namespace_owner.clone(),
                        },
                        &self.publication_root,
                    )
                    .await);
                }
-                reserved_version_delete = match self.reserve_version_delete(volume, path, uuid, rollback_dir).await {
+                reserved_version_delete = match self
+                    .reserve_version_delete_with_namespace_owner(volume, path, uuid, rollback_dir, namespace_owner.clone())
+                    .await
+                {
                    Ok(reserved) => reserved,
                    Err(err) => {
                        return Err(restore_delete_rollback_after_error(
@@ -5781,6 +5940,7 @@ impl LocalDisk {
                            DeleteRollbackFailure {
                                stage: "delete_version_reserve_data",
                                error: err,
+                                namespace_owner: namespace_owner.clone(),
                            },
                            &self.publication_root,
                        )
@@ -5789,9 +5949,14 @@ impl LocalDisk {
                };
                let rollback_data_path = rollback_path.join(uuid.to_string());
                if !reserved_version_delete
-                    && let Err(err) =
-                        rename_all_ignore_missing_source(&old_path, &rollback_data_path, &rollback_path, &self.publication_root)
-                            .await
+                    && let Err(err) = os::rename_all_ignore_missing_source_with_owner(
+                        &old_path,
+                        &rollback_data_path,
+                        &rollback_path,
+                        &self.publication_root,
+                        namespace_owner.clone(),
+                    )
+                    .await
                {
                    return Err(restore_delete_rollback_after_error(
                        file_path.as_path(),
@@ -5802,6 +5967,7 @@ impl LocalDisk {
                        DeleteRollbackFailure {
                            stage: "delete_version_stage_data",
                            error: err,
+                            namespace_owner: namespace_owner.clone(),
                        },
                        &self.publication_root,
                    )
@@ -5810,13 +5976,16 @@ impl LocalDisk {
                if should_fail_after_delete_data_staged(path) {
                    if reserved_version_delete {
                        return Err(self
-                            .abort_reserved_version_delete(
+                            .abort_reserved_version_delete_with_failure(
                                file_path.as_path(),
                                rollback_dir,
                                volume,
                                path,
-                                "delete_version_test_after_stage",
-                                DiskError::Unexpected,
+                                DeleteRollbackFailure {
+                                    stage: "delete_version_test_after_stage",
+                                    error: DiskError::Unexpected,
+                                    namespace_owner: namespace_owner.clone(),
+                                },
                            )
                            .await);
                    }
@@ -5829,6 +5998,7 @@ impl LocalDisk {
                        DeleteRollbackFailure {
                            stage: "delete_version_test_after_stage",
                            error: DiskError::Unexpected,
+                            namespace_owner: namespace_owner.clone(),
                        },
                        &self.publication_root,
                    )
@@ -5858,13 +6028,16 @@ impl LocalDisk {
                    let err: DiskError = err.into();
                    if reserved_version_delete && let Some(rollback_dir) = rollback_dir {
                        return Err(self
-                            .abort_reserved_version_delete(
+                            .abort_reserved_version_delete_with_failure(
                                file_path.as_path(),
                                rollback_dir,
                                volume,
                                path,
-                                "delete_version_metadata_encode",
-                                err,
+                                DeleteRollbackFailure {
+                                    stage: "delete_version_metadata_encode",
+                                    error: err,
+                                    namespace_owner: namespace_owner.clone(),
+                                },
                            )
                            .await);
                    }
@@ -5877,6 +6050,7 @@ impl LocalDisk {
                        DeleteRollbackFailure {
                            stage: "delete_version_metadata_encode",
                            error: err,
+                            namespace_owner: namespace_owner.clone(),
                        },
                        &self.publication_root,
                    )
@@ -5899,7 +6073,17 @@ impl LocalDisk {
        if let Err(err) = commit_result {
            if reserved_version_delete && let Some(rollback_dir) = rollback_dir {
                return Err(self
-                    .abort_reserved_version_delete(file_path.as_path(), rollback_dir, volume, path, "delete_version_commit", err)
+                    .abort_reserved_version_delete_with_failure(
+                        file_path.as_path(),
+                        rollback_dir,
+                        volume,
+                        path,
+                        DeleteRollbackFailure {
+                            stage: "delete_version_commit",
+                            error: err,
+                            namespace_owner: namespace_owner.clone(),
+                        },
+                    )
                    .await);
            }
            return Err(restore_delete_rollback_after_error(
@@ -5911,6 +6095,7 @@ impl LocalDisk {
                DeleteRollbackFailure {
                    stage: "delete_version_commit",
                    error: err,
+                    namespace_owner: namespace_owner.clone(),
                },
                &self.publication_root,
            )
@@ -5919,16 +6104,21 @@ impl LocalDisk {

        if reserved_version_delete
            && let Some(rollback_dir) = rollback_dir
-            && let Err(err) = self.commit_reserved_version_delete(volume, path, rollback_dir).await
+            && let Err(err) = self
+                .commit_reserved_version_delete_with_namespace_owner(volume, path, rollback_dir, namespace_owner.clone())
+                .await
        {
            return Err(self
-                .abort_reserved_version_delete(
+                .abort_reserved_version_delete_with_failure(
                    file_path.as_path(),
                    rollback_dir,
                    volume,
                    path,
-                    "delete_version_commit_intent",
-                    err,
+                    DeleteRollbackFailure {
+                        stage: "delete_version_commit_intent",
+                        error: err,
+                        namespace_owner: namespace_owner.clone(),
+                    },
                )
                .await);
        }
@@ -6098,7 +6288,7 @@ impl LocalDisk {
    }

    #[tracing::instrument(name = "delete", level = "trace", skip_all)]
-    async fn delete_with_namespace_owner(
+    pub(in crate::disk) async fn delete_with_namespace_owner(
        &self,
        volume: &str,
        path: &str,
@@ -6111,7 +6301,8 @@ impl LocalDisk {
            && let Some((object, transaction_id)) = path.rsplit_once('/')
            && let Ok(transaction_id) = Uuid::parse_str(transaction_id)
        {
-            self.finish_version_delete(volume, object, transaction_id).await?
+            self.finish_version_delete(volume, object, transaction_id, namespace_owner.clone())
+                .await?
        } else {
            false
        };
@@ -6453,19 +6644,27 @@ impl LocalDisk {
        path: &str,
        fi: FileInfo,
        object_dir: &Path,
-        xl_path: &Path,
        rollback_dir: Option<Uuid>,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
    ) -> Result<()> {
+        let xl_path = object_dir.join(STORAGE_FORMAT_FILE);
        if let Some(rollback_dir) = rollback_dir {
-            let rollback_path = object_dir.join(rollback_dir.to_string());
-            fs::create_dir_all(&rollback_path).await.map_err(to_file_error)?;
-            fs::write(rollback_path.join(DELETE_MARKER_ROLLBACK_FILE), [])
-                .await
-                .map_err(to_file_error)?;
+            write_delete_rollback_file(object_dir, rollback_dir, DELETE_MARKER_ROLLBACK_FILE, &[], namespace_owner.clone())
+                .await?;
        }
-        if let Err(err) = self.write_metadata("", volume, path, fi).await {
+        if let Err(err) = self
+            .write_metadata_with_namespace_owner(volume, path, fi, namespace_owner.clone())
+            .await
+        {
            if let Some(rollback_dir) = rollback_dir
-                && let Err(restore_err) = restore_delete_rollback(object_dir, xl_path, rollback_dir, &self.publication_root).await
+                && let Err(restore_err) = restore_delete_rollback_with_namespace_owner(
+                    object_dir,
+                    &xl_path,
+                    rollback_dir,
+                    &self.publication_root,
+                    namespace_owner,
+                )
+                .await
            {
                warn!(
                    event = EVENT_DISK_LOCAL_DELETE_ROLLBACK_FAILED,
@@ -6510,7 +6709,7 @@ impl LocalDisk {
                    return Err(DiskError::FileNotFound);
                };
                return self
-                    .write_missing_delete_marker(volume, path, delete_marker, object_dir, &xlpath, opts.old_data_dir)
+                    .write_missing_delete_marker(volume, path, delete_marker, object_dir, opts.old_data_dir, None)
                    .await;
            }
            Err(err) => return Err(err),
@@ -6559,6 +6758,7 @@ impl LocalDisk {
                        DeleteRollbackFailure {
                            stage: "delete_versions_metadata_update",
                            error: err,
+                            namespace_owner: None,
                        },
                        &self.publication_root,
                    )
@@ -6594,6 +6794,7 @@ impl LocalDisk {
                            DeleteRollbackFailure {
                                stage: "delete_versions_data_path",
                                error: err,
+                                namespace_owner: None,
                            },
                            &self.publication_root,
                        )
@@ -6625,6 +6826,7 @@ impl LocalDisk {
                            DeleteRollbackFailure {
                                stage: "delete_versions_rollback_dir",
                                error: err,
+                                namespace_owner: None,
                            },
                            &self.publication_root,
                        )
@@ -6665,6 +6867,7 @@ impl LocalDisk {
                            DeleteRollbackFailure {
                                stage: "delete_versions_stage_data",
                                error: err,
+                                namespace_owner: None,
                            },
                            &self.publication_root,
                        )
@@ -6692,6 +6895,7 @@ impl LocalDisk {
                            DeleteRollbackFailure {
                                stage: "delete_versions_test_after_stage",
                                error: DiskError::Unexpected,
+                                namespace_owner: None,
                            },
                            &self.publication_root,
                        )
@@ -6734,6 +6938,7 @@ impl LocalDisk {
                    DeleteRollbackFailure {
                        stage: "delete_versions_commit_delete",
                        error: err,
+                        namespace_owner: None,
                    },
                    &self.publication_root,
                )
@@ -6780,6 +6985,7 @@ impl LocalDisk {
                    DeleteRollbackFailure {
                        stage: "delete_versions_metadata_encode",
                        error: err,
+                        namespace_owner: None,
                    },
                    &self.publication_root,
                )
@@ -6805,6 +7011,7 @@ impl LocalDisk {
                DeleteRollbackFailure {
                    stage: "delete_versions_commit_write",
                    error: err,
+                    namespace_owner: None,
                },
                &self.publication_root,
            )
@@ -8015,6 +8222,18 @@ impl LocalDisk {
    }

    async fn reserve_version_delete(&self, volume: &str, object: &str, data_dir: Uuid, rollback_dir: Uuid) -> Result<bool> {
+        self.reserve_version_delete_with_namespace_owner(volume, object, data_dir, rollback_dir, None)
+            .await
+    }
+
+    async fn reserve_version_delete_with_namespace_owner(
+        &self,
+        volume: &str,
+        object: &str,
+        data_dir: Uuid,
+        rollback_dir: Uuid,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<bool> {
        let path = format!("{object}/{data_dir}");
        let data_path = self.io_get_object_path(volume, &path)?;
        match fs::metadata(&data_path).await {
@@ -8024,6 +8243,28 @@ impl LocalDisk {
            Err(err) => return Err(to_file_error(err).into()),
        }
        let marker_path = data_path.join(format!("{RESERVED_DELETE_DATA_DIR_MARKER_PREFIX}{rollback_dir}"));
+        if namespace_owner.is_some() {
+            let lease = os::acquire_namespace_mutation_lease_with_owner(&marker_path, namespace_owner.clone()).await;
+            let volume = volume.to_string();
+            let sync = os::run_blocking_namespace_operation(lease, move || {
+                #[cfg(test)]
+                run_owned_file_write_before_open(&marker_path);
+                let marker = std::fs::File::create(marker_path)?;
+                let sync = effective_durability(&volume).syncs_commit_metadata();
+                if sync {
+                    marker.sync_all()?;
+                }
+                Ok(sync)
+            })
+            .await
+            .map_err(to_file_error)?;
+            if sync {
+                os::fsync_dir_with_owner(&data_path, namespace_owner)
+                    .await
+                    .map_err(to_file_error)?;
+            }
+            return Ok(true);
+        }
        let marker = File::create(marker_path).await.map_err(to_file_error)?;
        if effective_durability(volume).syncs_commit_metadata() {
            marker.sync_all().await.map_err(to_file_error)?;
@@ -8033,6 +8274,17 @@ impl LocalDisk {
    }

    async fn commit_reserved_version_delete(&self, volume: &str, object: &str, rollback_dir: Uuid) -> Result<()> {
+        self.commit_reserved_version_delete_with_namespace_owner(volume, object, rollback_dir, None)
+            .await
+    }
+
+    async fn commit_reserved_version_delete_with_namespace_owner(
+        &self,
+        volume: &str,
+        object: &str,
+        rollback_dir: Uuid,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
        let object_path = self.io_get_object_path(volume, object)?;
        let mut entries = match fs::read_dir(object_path).await {
            Ok(entries) => entries,
@@ -8048,10 +8300,14 @@ impl LocalDisk {
                continue;
            }
            let reserved_path = entry.path().join(&reserved_name);
-            match fs::rename(&reserved_path, entry.path().join(&committed_name)).await {
+            match os::rename_with_namespace_owner(&reserved_path, &entry.path().join(&committed_name), namespace_owner.clone())
+                .await
+            {
                Ok(()) => {
                    if effective_durability(volume).syncs_commit_metadata() {
-                        os::fsync_dir(&entry.path()).await.map_err(to_file_error)?;
+                        os::fsync_dir_with_owner(&entry.path(), namespace_owner.clone())
+                            .await
+                            .map_err(to_file_error)?;
                    }
                }
                Err(err) if err.kind() == ErrorKind::NotFound => {}
@@ -8061,7 +8317,13 @@ impl LocalDisk {
        Ok(())
    }

-    async fn finish_version_delete(&self, volume: &str, object: &str, rollback_dir: Uuid) -> Result<bool> {
+    async fn finish_version_delete(
+        &self,
+        volume: &str,
+        object: &str,
+        rollback_dir: Uuid,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<bool> {
        let object_path = self.io_get_object_path(volume, object)?;
        let mut entries = match fs::read_dir(object_path).await {
            Ok(entries) => entries,
@@ -8082,13 +8344,14 @@ impl LocalDisk {
                Err(err) => return Err(to_file_error(err).into()),
            }
            if let Err(err) = self
-                .delete_data_dir(
+                .delete_data_dir_with_namespace_owner(
                    volume,
                    &format!("{object}/{data_dir}"),
                    DeleteOptions {
                        recursive: true,
                        ..Default::default()
                    },
+                    namespace_owner.clone(),
                )
                .await
                && first_err.is_none()
@@ -8109,6 +8372,28 @@ impl LocalDisk {
        object: &str,
        stage: &'static str,
        err: DiskError,
+    ) -> DiskError {
+        self.abort_reserved_version_delete_with_failure(
+            object_dir,
+            rollback_dir,
+            volume,
+            object,
+            DeleteRollbackFailure {
+                stage,
+                error: err,
+                namespace_owner: None,
+            },
+        )
+        .await
+    }
+
+    async fn abort_reserved_version_delete_with_failure(
+        &self,
+        object_dir: &Path,
+        rollback_dir: Uuid,
+        volume: &str,
+        object: &str,
+        failure: DeleteRollbackFailure,
    ) -> DiskError {
        let xl_path = object_dir.join(STORAGE_FORMAT_FILE);
        restore_delete_rollback_after_error(
@@ -8117,7 +8402,7 @@ impl LocalDisk {
            Some(rollback_dir),
            volume,
            object,
-            DeleteRollbackFailure { stage, error: err },
+            failure,
            &self.publication_root,
        )
        .await
@@ -9608,49 +9893,7 @@ impl DiskAPI for LocalDisk {
    }

    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
-        let key = SnapshotLeaseKey {
-            volume: volume.to_string(),
-            path: path.to_string(),
-        };
-        {
-            let mut registry = self.snapshot_leases.lock().await;
-            if let Some(entry) = registry.entries.get_mut(&key) {
-                if !entry.tokens.is_empty() {
-                    entry.pending_delete.get_or_insert_with(|| opts.clone());
-                    return Ok(DataDirDeleteStatus::Deferred);
-                }
-                if entry.deleting {
-                    entry.pending_delete.get_or_insert_with(|| opts.clone());
-                    return Ok(DataDirDeleteStatus::Deferred);
-                }
-                entry.deleting = true;
-                entry.pending_delete.get_or_insert_with(|| opts.clone());
-            } else {
-                registry.entries.insert(
-                    key.clone(),
-                    SnapshotLeaseEntry {
-                        pending_delete: Some(opts.clone()),
-                        deleting: true,
-                        ..Default::default()
-                    },
-                );
-            }
-        }
-
-        let result = self.delete_unleased(volume, path, &opts).await;
-        let mut registry = self.snapshot_leases.lock().await;
-        match result {
-            Ok(()) => {
-                registry.entries.remove(&key);
-                Ok(DataDirDeleteStatus::Deleted)
-            }
-            Err(err) => {
-                if let Some(entry) = registry.entries.get_mut(&key) {
-                    entry.deleting = false;
-                }
-                Err(err)
-            }
-        }
+        self.delete_data_dir_with_namespace_owner(volume, path, opts, None).await
    }

    #[tracing::instrument(level = "trace", skip_all)]
@@ -9689,32 +9932,8 @@ impl DiskAPI for LocalDisk {
        Err(Error::other("Invalid Argument"))
    }

-    #[tracing::instrument(level = "trace", skip_all)]
    async fn write_metadata(&self, _org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()> {
-        crate::hp_guard!("LocalDisk::write_metadata");
-        fi.validate_for_metadata_read()?;
-        let p = self.io_get_object_path(volume, format!("{path}/{STORAGE_FORMAT_FILE}").as_str())?;
-
-        let mut meta = FileMeta::new();
-        if !fi.fresh {
-            let (buf, _) = read_file_exists(&p).await?;
-            if !buf.is_empty() {
-                let _ = meta.unmarshal_msg(&buf).map_err(|_| {
-                    meta = FileMeta::new();
-                });
-            }
-        }
-
-        meta.add_version(fi)?;
-
-        let fm_data = meta.marshal_msg()?;
-
-        // Atomic temp+rename: this path also rewrites live xl.meta (delete markers,
-        // decommission), where an in-place truncate would expose torn metadata.
-        self.write_all_meta(volume, format!("{path}/{STORAGE_FORMAT_FILE}").as_str(), &fm_data, true)
-            .await?;
-
-        Ok(())
+        self.write_metadata_with_namespace_owner(volume, path, fi, None).await
    }

    #[tracing::instrument(level = "trace", skip_all)]
@@ -22177,4 +22396,135 @@ mod test {
        assert_eq!(mount_id_from_mountinfo_contents(mountinfo, Path::new("/mnt/replacement disk")), Some(202));
        assert_eq!(mount_id_from_mountinfo_contents(mountinfo, Path::new("/mnt/replacement")), None);
    }
+
+    #[cfg(not(windows))]
+    #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
+    async fn single_delete_internal_restore_keeps_owner_after_cancellation() {
+        use crate::disk::os::prepared_publication_test_hooks as hooks;
+        use futures::FutureExt;
+
+        let dir = tempfile::tempdir().expect("fixture directory");
+        let endpoint = Endpoint::try_from(dir.path().to_str().expect("UTF-8 fixture path")).expect("endpoint");
+        let disk = Arc::new(LocalDisk::new(&endpoint, false).await.expect("local disk"));
+        let bucket = "single-delete-internal-restore";
+        let object = format!("object-{}", Uuid::new_v4());
+        ensure_test_volume(&disk, bucket).await;
+        let version = Uuid::new_v4();
+        let data_dir = Uuid::new_v4();
+        let rollback_dir = Uuid::new_v4();
+        let fi = test_file_info(&object, version, Some(data_dir), None);
+        let original = test_meta(fi.clone());
+        let object_dir = disk.io_get_object_path(bucket, &object).expect("object IO path");
+        let part = object_dir.join(data_dir.to_string()).join("part.1");
+        let metadata = object_dir.join(STORAGE_FORMAT_FILE);
+        let backup = object_dir.join(rollback_dir.to_string()).join(STORAGE_FORMAT_FILE_BACKUP);
+        fs::create_dir_all(part.parent().expect("data parent"))
+            .await
+            .expect("data directory");
+        fs::write(&part, b"x").await.expect("real shard");
+        fs::write(&metadata, &original).await.expect("real version metadata");
+        set_delete_version_fail_after_data_staged(&object);
+        let (entered_tx, entered_rx) = tokio::sync::oneshot::channel();
+        let (release, release_rx) = std::sync::mpsc::channel::<()>();
+        let hook = hooks::install_at(hooks::Stage::Rename, &metadata, move || {
+            let _ = entered_tx.send(());
+            let _ = release_rx.recv();
+        });
+        let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let before = ctx.namespace_commit_generation();
+        let owner = ctx.begin_namespace_commit();
+        let deleting_disk = Arc::clone(&disk);
+        let deleting_object = object.clone();
+        let mut delete = tokio::spawn(async move {
+            deleting_disk
+                .delete_version_inner(
+                    bucket,
+                    &deleting_object,
+                    fi,
+                    DeleteVersionMutation {
+                        force_del_marker: false,
+                        opts: DeleteOptions {
+                            old_data_dir: Some(rollback_dir),
+                            ..Default::default()
+                        },
+                        namespace_owner: Some(owner),
+                    },
+                )
+                .await
+        });
+        let mut joined = false;
+        let mut entered = false;
+        let mut counts = None;
+        let observations = std::panic::AssertUnwindSafe(async {
+            tokio::time::timeout(Duration::from_secs(10), async {
+                tokio::select! {
+                    result = entered_rx => {
+                        result.expect("actual internal restore entry");
+                        entered = true;
+                    }
+                    result = &mut delete => {
+                        joined = true;
+                        panic!("delete returned before internal physical restore: {result:?}");
+                    }
+                }
+            })
+            .await
+            .expect("internal restore must reach the physical rename");
+            assert_eq!(std::fs::read(&backup).expect("real undo backup"), original);
+            assert_eq!(std::fs::read(&part).expect("reserved shard"), b"x");
+            delete.abort();
+            let result = tokio::time::timeout(Duration::from_secs(5), &mut delete).await;
+            joined = result.is_ok();
+            assert!(
+                result
+                    .expect("cancelled caller joins")
+                    .expect_err("cancelled caller")
+                    .is_cancelled()
+            );
+            counts = Some((ctx.namespace_commits_pending(), ctx.namespace_commit_generation()));
+            assert!(hooks::drain_namespace_key(&metadata).now_or_never().is_none());
+        })
+        .catch_unwind()
+        .await;
+
+        drop(release);
+        drop(hook);
+        let coordinator_drained = joined || tokio::time::timeout(Duration::from_secs(10), &mut delete).await.is_ok();
+        if !coordinator_drained {
+            delete.abort();
+            let _ = tokio::time::timeout(Duration::from_secs(5), &mut delete).await;
+        }
+        let physical_drained = tokio::time::timeout(Duration::from_secs(5), hooks::drain_namespace_key(&metadata))
+            .await
+            .is_ok();
+        let owner_drained = tokio::time::timeout(Duration::from_secs(5), async {
+            while ctx.namespace_commits_pending() {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .is_ok();
+        if !entered || !coordinator_drained || !physical_drained || !owner_drained {
+            eprintln!("internal restore cleanup incomplete; retained root: {:?}", dir.keep());
+            if let Err(panic) = observations {
+                std::panic::resume_unwind(panic);
+            }
+            panic!("internal restore cleanup must drain before removing its root");
+        }
+        if let Err(panic) = observations {
+            std::panic::resume_unwind(panic);
+        }
+        assert_eq!(std::fs::read(&metadata).expect("late restored metadata"), original);
+        assert!(!backup.exists(), "the actual backup rename must have completed");
+        assert_eq!(std::fs::read(&part).expect("old shard survives"), b"x");
+        disk.read_version("", bucket, &object, &version.to_string(), &ReadOptions::default())
+            .await
+            .expect("restored version");
+        let (pending, generation) = counts.expect("observations completed");
+        assert!(pending, "internal error recovery lost the physical namespace owner");
+        assert_eq!(generation, before + 1);
+        assert_eq!(ctx.namespace_commit_generation(), before + 2);
+        assert!(!ctx.namespace_commits_pending());
+    }
 }
@@ -732,6 +732,46 @@ impl Disk {
        }
    }

+    pub(crate) async fn delete_version_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        fi: FileInfo,
+        force_del_marker: bool,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        match self {
+            Self::Local(disk) => {
+                disk.delete_version_with_namespace_owner(volume, path, fi, force_del_marker, opts, namespace_owner)
+                    .await
+            }
+            Self::Remote(disk) => {
+                let result = disk.delete_version(volume, path, fi, force_del_marker, opts).await;
+                // This is sender lifetime only, not proof of a remote physical drain.
+                drop(namespace_owner);
+                result
+            }
+        }
+    }
+
+    pub(crate) async fn delete_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        match self {
+            Self::Local(disk) => disk.delete_with_namespace_owner(volume, path, opts, namespace_owner).await,
+            Self::Remote(disk) => {
+                let result = disk.delete(volume, path, opts).await;
+                drop(namespace_owner);
+                result
+            }
+        }
+    }
+
    /// Keep local undo publication owned independently of the wrapper deadline.
    /// Remote undo retains its existing RPC contract; this is not a remote drain proof.
    pub(crate) async fn undo_write_with_namespace_owner(
@@ -247,7 +247,7 @@ pub(crate) mod fsync_dir_recorder {
 }

 /// Pause a real namespace mutation inside its physical executor.
-#[cfg(all(any(test, feature = "test-util"), not(windows)))]
+#[cfg(all(test, not(windows)))]
 pub(crate) mod prepared_publication_test_hooks {
    use super::*;

@@ -256,9 +256,7 @@ pub(crate) mod prepared_publication_test_hooks {
        PreparedRename,
        Rename,
        Remove,
-        #[cfg(test)]
        Rollback,
-        #[cfg(test)]
        DirFsync,
    }

@@ -274,7 +272,6 @@ pub(crate) mod prepared_publication_test_hooks {
        }
    }

-    #[cfg(test)]
    pub(crate) fn install(path: &Path, hook: impl FnOnce() + Send + 'static) -> Guard {
        install_at(Stage::PreparedRename, path, hook)
    }
@@ -291,50 +288,45 @@ pub(crate) mod prepared_publication_test_hooks {
            hook();
        }
    }
-}

-/// Controlled application-test pause at an existing physical executor boundary.
-#[cfg(all(feature = "test-util", not(windows)))]
-pub struct LocalPublicationPause {
-    _hook: prepared_publication_test_hooks::Guard,
-    entered: oneshot::Receiver<()>,
-    _release: std::sync::mpsc::Sender<()>,
-}
+    #[cfg(test)]
+    type RenameDestinationHook = Box<dyn FnOnce(&Path) + Send>;
+    #[cfg(test)]
+    static RENAME_DESTINATIONS: LazyLock<Mutex<HashMap<PathBuf, RenameDestinationHook>>> =
+        LazyLock::new(|| Mutex::new(HashMap::new()));

-#[cfg(all(feature = "test-util", not(windows)))]
-#[derive(Clone, Copy)]
-pub enum LocalPublicationStage {
-    PreparedRename,
-    Rename,
-    Remove,
-}
+    #[cfg(test)]
+    pub(crate) struct RenameDestinationGuard(PathBuf);

-#[cfg(all(feature = "test-util", not(windows)))]
-impl LocalPublicationPause {
-    pub fn install(disk: &crate::disk::Disk, volume: &str, path: &str, stage: LocalPublicationStage) -> Result<Self> {
-        let path = disk
-            .get_object_path_for_io_if_local(volume, path)
-            .ok_or(DiskError::DiskNotFound)??;
-        let stage = match stage {
-            LocalPublicationStage::PreparedRename => prepared_publication_test_hooks::Stage::PreparedRename,
-            LocalPublicationStage::Rename => prepared_publication_test_hooks::Stage::Rename,
-            LocalPublicationStage::Remove => prepared_publication_test_hooks::Stage::Remove,
-        };
-        let (entered_tx, entered) = oneshot::channel();
-        let (release, release_rx) = std::sync::mpsc::channel::<()>();
-        let hook = prepared_publication_test_hooks::install_at(stage, &path, move || {
-            let _ = entered_tx.send(());
-            let _ = release_rx.recv();
-        });
-        Ok(Self {
-            _hook: hook,
-            entered,
-            _release: release,
-        })
+    #[cfg(test)]
+    impl Drop for RenameDestinationGuard {
+        fn drop(&mut self) {
+            RENAME_DESTINATIONS.lock().remove(&self.0);
+        }
    }

-    pub async fn entered(&mut self) -> std::result::Result<(), oneshot::error::RecvError> {
-        (&mut self.entered).await
+    #[cfg(test)]
+    pub(crate) fn observe_rename_destination(source: &Path, hook: impl FnOnce(&Path) + Send + 'static) -> RenameDestinationGuard {
+        assert!(
+            RENAME_DESTINATIONS
+                .lock()
+                .insert(source.to_path_buf(), Box::new(hook))
+                .is_none()
+        );
+        RenameDestinationGuard(source.to_path_buf())
+    }
+
+    #[cfg(test)]
+    pub(crate) async fn drain_namespace_key(path: &Path) {
+        drop(super::acquire_namespace_mutation_lease(path).await);
+    }
+
+    #[cfg(test)]
+    pub(super) fn run_rename_destination(source: &Path, destination: &Path) {
+        let hook = RENAME_DESTINATIONS.lock().remove(source);
+        if let Some(hook) = hook {
+            hook(destination);
+        }
    }
 }

@@ -1409,7 +1401,7 @@ async fn acquire_namespace_mutation_lease(path: &Path) -> Arc<NamespaceMutationL
    acquire_namespace_mutation_lease_with_owner(path, None).await
 }

-async fn acquire_namespace_mutation_lease_with_owner(
+pub(in crate::disk) async fn acquire_namespace_mutation_lease_with_owner(
    path: &Path,
    namespace_owner: Option<Arc<dyn Send + Sync>>,
 ) -> Arc<NamespaceMutationLease> {
@@ -1964,7 +1956,7 @@ pub(crate) async fn remove_file_with_owner(
    let path = path.as_ref().to_path_buf();
    let lease = acquire_namespace_mutation_lease_with_owner(&path, namespace_owner).await;
    run_blocking_namespace_operation(lease, move || {
-        #[cfg(all(any(test, feature = "test-util"), not(windows)))]
+        #[cfg(all(test, not(windows)))]
        prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Remove, &path);
        std::fs::remove_file(path)
    })
@@ -1984,6 +1976,42 @@ pub(crate) async fn remove_dir_with_owner(
    run_blocking_namespace_operation(lease, move || std::fs::remove_dir(path)).await
 }

+/// Preserve raw rename semantics while retaining a counted owner in the syscall.
+/// Unlike reliable rename, this never creates parents or retries a missing source.
+pub(in crate::disk) async fn rename_with_namespace_owner(
+    src: &Path,
+    dst: &Path,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> io::Result<()> {
+    if namespace_owner.is_none() {
+        return tokio::fs::rename(src, dst).await;
+    }
+    let src = src.to_path_buf();
+    let dst = dst.to_path_buf();
+    let lease = acquire_namespace_mutation_lease_with_owner(&dst, namespace_owner).await;
+    run_blocking_namespace_operation(lease, move || {
+        #[cfg(all(test, not(windows)))]
+        {
+            prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Rename, &src);
+            prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Rename, &dst);
+        }
+        std::fs::rename(src, dst)
+    })
+    .await
+}
+
+pub(in crate::disk) async fn create_dir_all_with_namespace_owner(
+    path: &Path,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> io::Result<()> {
+    if namespace_owner.is_none() {
+        return tokio::fs::create_dir_all(path).await;
+    }
+    let path = path.to_path_buf();
+    let lease = acquire_namespace_mutation_lease_with_owner(&path, namespace_owner).await;
+    run_blocking_namespace_operation(lease, move || std::fs::create_dir_all(path)).await
+}
+
 #[tracing::instrument(name = "rename_all", level = "debug", skip_all)]
 pub(crate) async fn rename_all_with_owner(
    src_file_path: impl AsRef<Path>,
@@ -2188,7 +2216,7 @@ pub(crate) async fn rename_all_with_prepared_source(
        move || {
            validate_prepared_rename_source(&prepared_source, &src_file_path)?;
            let preparation = prepare_rename_with_retry(&src_file_path, &dst_file_path, &base_dir, &publication_root)?;
-            #[cfg(any(test, feature = "test-util"))]
+            #[cfg(test)]
            prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::PreparedRename, &dst_file_path);
            rename_prepared(&src_file_path, &dst_file_path, &preparation)
        }
@@ -2319,7 +2347,9 @@ async fn reliable_rename_inner_with_lease(
        let base_dir = base_dir.clone();
        move || {
            let preparation = prepare_rename_with_retry(&src_file_path, &dst_file_path, &base_dir, &publication_root)?;
-            #[cfg(all(any(test, feature = "test-util"), not(windows)))]
+            #[cfg(all(test, not(windows)))]
+            prepared_publication_test_hooks::run_rename_destination(&src_file_path, &dst_file_path);
+            #[cfg(all(test, not(windows)))]
            {
                prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Rename, &src_file_path);
                prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Rename, &dst_file_path);
@@ -2278,6 +2278,121 @@ mod tests {
        assert_eq!(read, fixture.plaintext, "SSE-C + compression full GET must reassemble all parts");
    }

+    #[tokio::test]
+    async fn multipart_empty_tail_full_reads_preserve_plaintext() {
+        let key = [0x6Eu8; 32];
+        let part_sizes = [5 * 1024 * 1024, 0];
+        let encrypted = build_legacy_ssec_multipart_fixture(key, &part_sizes).await;
+        for (kind, mut fixture, headers) in [
+            (
+                "encrypted",
+                CompressedMultipartFixture {
+                    object_info: encrypted.object_info,
+                    stored: encrypted.ciphertext,
+                    plaintext: encrypted.plaintext,
+                },
+                ssec_headers_from_key(key),
+            ),
+            ("compressed", compressed_multipart_fixture(&part_sizes).await, HeaderMap::new()),
+            (
+                "compressed and encrypted",
+                compressed_encrypted_multipart_fixture(key, &part_sizes).await,
+                ssec_headers_from_key(key),
+            ),
+        ] {
+            fixture.object_info.etag = Some(faster_hex::hex_string(Md5::digest(&fixture.plaintext).as_ref()));
+            assert_eq!(fixture.object_info.etag.as_ref().expect("source ETag").len(), 32);
+            assert_eq!(fixture.object_info.parts.len(), 2);
+            let tail = &fixture.object_info.parts[1];
+            assert_eq!(tail.actual_size, 0, "{kind}: final part has no plaintext");
+            if kind == "compressed" {
+                assert_eq!(tail.size, 0, "unpadded compression emits no bytes for an empty part");
+            } else {
+                assert!(tail.size > 0, "{kind}: the empty part still has a stored frame");
+            }
+            let stored_size = i64::try_from(fixture.stored.len()).expect("fixture size fits i64");
+            let (mut reader, offset, length) = GetObjectReader::new(
+                Box::new(Cursor::new(fixture.stored)),
+                None,
+                &fixture.object_info,
+                &ObjectOptions::default(),
+                &headers,
+            )
+            .await
+            .expect("full transformed read must include the empty tail");
+            assert_eq!((offset, length), (0, stored_size), "{kind}: full read includes all stored parts");
+            let mut body = Vec::new();
+            reader
+                .stream
+                .read_to_end(&mut body)
+                .await
+                .expect("read through the complete decoder EOF");
+            assert_eq!(body, fixture.plaintext, "{kind}: no plaintext is added or lost by the empty tail");
+        }
+    }
+
+    #[tokio::test]
+    async fn multipart_empty_tail_full_read_authenticates_v2_final_frame() {
+        let key = [0x6Eu8; 32];
+        let plaintext = legacy_fixture_part_plaintext(1, 5 * 1024 * 1024);
+        let mut ciphertext = Vec::new();
+        let mut parts = Vec::new();
+        for (number, body) in [(1, plaintext.as_slice()), (2, b"".as_slice())] {
+            let start = ciphertext.len();
+            rustfs_rio::EncryptReader::new_multipart_v2(Cursor::new(body), key, LEGACY_FIXTURE_BASE_NONCE, number)
+                .read_to_end(&mut ciphertext)
+                .await
+                .expect("encrypt a v2 fixture part with an authenticated final frame");
+            parts.push(ObjectPartInfo {
+                number,
+                size: ciphertext.len() - start,
+                actual_size: i64::try_from(body.len()).expect("fixture plaintext size fits"),
+                ..Default::default()
+            });
+        }
+        let tail_start = parts[0].size;
+        assert_eq!(parts[1].actual_size, 0);
+        assert!(parts[1].size > 8, "the empty final frame carries more than an END marker");
+        let object_info = ObjectInfo {
+            bucket: "bucket".to_string(),
+            name: "v2-empty-tail".to_string(),
+            size: i64::try_from(ciphertext.len()).expect("fixture ciphertext size fits"),
+            etag: Some(faster_hex::hex_string(Md5::digest(&plaintext).as_ref())),
+            parts: Arc::new(parts),
+            user_defined: Arc::new(legacy_ssec_multipart_metadata(key, plaintext.len())),
+            ..Default::default()
+        };
+        for corrupt_tail in [false, true] {
+            let mut stored = ciphertext.clone();
+            if corrupt_tail {
+                // The v2 header is authenticated associated data, including
+                // the header of a final frame containing zero plaintext.
+                stored[tail_start + 5] ^= 1;
+            }
+            let (mut reader, offset, length) = GetObjectReader::new(
+                Box::new(Cursor::new(stored)),
+                None,
+                &object_info,
+                &ObjectOptions::default(),
+                &ssec_headers_from_key(key),
+            )
+            .await
+            .expect("construct the full reader before consuming the final frame");
+            assert_eq!((offset, length), (0, object_info.size));
+            let result = tokio::io::copy(&mut reader.stream, &mut tokio::io::sink()).await;
+            if corrupt_tail {
+                let err = result.expect_err("EOF must authenticate the empty final frame after all plaintext is returned");
+                assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+                assert_eq!(err.to_string(), "v2 encrypted frame failed authentication");
+            } else {
+                assert_eq!(
+                    result.expect("valid empty final frame must reach EOF"),
+                    u64::try_from(plaintext.len()).expect("plaintext length fits")
+                );
+            }
+        }
+    }
+
    #[tokio::test]
    async fn compressed_encrypted_multipart_range_crosses_part_boundary() {
        let key_bytes = [0x6Eu8; 32];
@@ -3656,6 +3771,61 @@ mod tests {
        .await;
    }

+    #[tokio::test]
+    async fn multipart_full_read_preserves_legacy_zero_and_negative_part_sizes() {
+        let key = [0x77; 32];
+        let part_sizes = [5 * 1024 * 1024, 1024 * 1024];
+        let encrypted = build_legacy_ssec_multipart_fixture(key, &part_sizes).await;
+        // The encrypted case supplies the fixture key explicitly. This covers
+        // full decrypted reads, not managed-key acquisition.
+        for (kind, fixture, headers) in [
+            ("compressed", compressed_multipart_fixture(&part_sizes).await, HeaderMap::new()),
+            (
+                "encrypted with supplied key",
+                CompressedMultipartFixture {
+                    object_info: encrypted.object_info,
+                    stored: encrypted.ciphertext,
+                    plaintext: encrypted.plaintext,
+                },
+                ssec_headers_from_key(key),
+            ),
+        ] {
+            let source_etag = faster_hex::hex_string(Md5::digest(&fixture.plaintext).as_ref());
+            assert_eq!(source_etag.len(), 32);
+            assert_eq!(fixture.plaintext.len(), 6 * 1024 * 1024);
+            for part_index in 0..part_sizes.len() {
+                assert!(fixture.object_info.parts[part_index].actual_size > 0, "the selected part is nonempty");
+                for actual_size in [0, -1] {
+                    let mut object_info = fixture.object_info.clone();
+                    object_info.etag = Some(source_etag.clone());
+                    Arc::make_mut(&mut object_info.parts)[part_index].actual_size = actual_size;
+                    let (mut reader, offset, length) = GetObjectReader::new(
+                        Box::new(Cursor::new(fixture.stored.clone())),
+                        None,
+                        &object_info,
+                        &ObjectOptions::default(),
+                        &headers,
+                    )
+                    .await
+                    .expect("the authoritative total size must keep full legacy reads available");
+                    assert_eq!(offset, 0);
+                    assert_eq!(length, i64::try_from(fixture.stored.len()).expect("stored size fits"));
+                    let mut body = Vec::new();
+                    reader
+                        .stream
+                        .read_to_end(&mut body)
+                        .await
+                        .expect("full read must reach EOF despite an unspecified per-part logical size");
+                    assert_eq!(
+                        body, fixture.plaintext,
+                        "{kind}: part {part_index} with actual_size={actual_size} must not lose readable data"
+                    );
+                    assert_eq!(reader.object_info.etag.as_deref(), Some(source_etag.as_str()));
+                }
+            }
+        }
+    }
+
    /// The physical part sizes must add up to `oi.size` for a seek to be safe;
    /// inconsistent metadata must fall back to the previous full-object read
    /// instead of scheduling an erasure read past the object end.
@@ -1597,7 +1597,7 @@ impl ObjectInfo {
    }

    pub fn is_multipart(&self) -> bool {
-        self.etag.as_ref().is_some_and(|v| v.len() != 32)
+        self.parts.len() > 1 || self.etag.as_ref().is_some_and(|v| v.len() != 32)
    }

    pub fn is_encrypted(&self) -> bool {
@@ -2235,6 +2235,35 @@ mod tests {
    }
    use rustfs_filemeta::{FileInfo, FileMeta, MetaCacheEntry, TRANSITION_COMPLETE};

+    #[test]
+    fn multipart_identity_uses_stored_parts_and_preserves_the_etag_fallback() {
+        let plain_etag = "0123456789abcdef0123456789abcdef";
+        let multipart_etag = "0123456789abcdef0123456789abcdef-1";
+        for (case, part_count, etag, expected) in [
+            ("preserved source ETag", 2, Some(plain_etag), true),
+            ("missing ETag", 2, None, true),
+            ("ordinary PUT", 1, Some(plain_etag), false),
+            ("ordinary PUT without ETag", 1, None, false),
+            ("single-part MPU", 1, Some(multipart_etag), true),
+            ("legacy MPU without parts", 0, Some(multipart_etag), true),
+        ] {
+            let object = ObjectInfo {
+                etag: etag.map(str::to_string),
+                parts: Arc::new(
+                    (1..=part_count)
+                        .map(|number| ObjectPartInfo {
+                            number,
+                            ..Default::default()
+                        })
+                        .collect(),
+                ),
+                ..Default::default()
+            };
+
+            assert_eq!(object.is_multipart(), expected, "{case}");
+        }
+    }
+
    fn inline_fast_path_object(size: i64, versioned: bool) -> ObjectInfo {
        ObjectInfo {
            size,
@@ -14,7 +14,8 @@

 use crate::bucket::lifecycle::tier_last_day_stats::DailyAllTierStats;
 use crate::cluster::rpc::{
-    PeerRestClient, ScannerPeerActivity, ScannerPeerDirtyUsageSnapshot, ScannerPublicationLease, TierConfigReloadOutcome,
+    PeerRestClient, ScannerDirtyUsageAcknowledgement, ScannerPeerActivity, ScannerPeerDirtyUsageSnapshot,
+    ScannerPublicationLease, TierConfigReloadOutcome,
 };
 use crate::diagnostics::admin_server_info::get_commit_id;
 use crate::disk::DiskAPI;
@@ -33,11 +34,11 @@ use rustfs_madmin::net::NetInfo;
 use rustfs_madmin::{ItemState, ServerProperties, StorageInfo};
 use rustfs_utils::XHost;
 use sha2::{Digest, Sha256};
-use std::collections::{BTreeMap, HashMap, hash_map::DefaultHasher};
+use std::collections::{BTreeMap, BTreeSet, HashMap, hash_map::DefaultHasher};
 use std::future::Future;
 use std::hash::{Hash, Hasher};
 use std::sync::{
-    Arc, Mutex, OnceLock,
+    Arc, LazyLock, Mutex, OnceLock,
    atomic::{AtomicBool, AtomicUsize, Ordering},
 };
 use std::time::{Duration, Instant, SystemTime};
@@ -311,12 +312,29 @@ pub struct LegacyTransitionStateReconcileFleetProofToken {
    _permit: FleetCapabilityProofPermit,
 }

+/// Effect-window authority for one immutable ILM recovery export.
+pub struct IlmRecoveryExportFleetProofToken {
+    token: FleetCapabilityProofToken,
+    _permit: FleetCapabilityProofPermit,
+}
+
+/// Effect-window authority for emitting the compact transition-transaction
+/// state sequence. The generation permit prevents a successor proof from
+/// being published until the admitted writer has finished.
+pub(crate) struct TransitionTransactionCompactionFleetProofToken {
+    token: FleetCapabilityProofToken,
+    _permit: FleetCapabilityProofPermit,
+}
+
 static REMOTE_VERSION_STATE_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static CROSS_POOL_FENCE_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static TIER_DELETE_JOURNAL_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static DECOMMISSION_TARGET_FENCE_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static LEGACY_TRANSITION_STATE_RECONCILE_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
+static ILM_RECOVERY_EXPORT_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
+static TRANSITION_TRANSACTION_COMPACTION_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static REMOTE_VERSION_STATE_PROBE_TOPOLOGY: OnceLock<String> = OnceLock::new();
+static ILM_RECOVERY_EXPORT_LOCAL_PROCESS_EPOCH: LazyLock<Uuid> = LazyLock::new(Uuid::new_v4);

 fn cross_pool_fence_fleet_proof_slot() -> &'static std::sync::RwLock<FleetCapabilityProofState> {
    CROSS_POOL_FENCE_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
@@ -338,6 +356,14 @@ fn legacy_transition_state_reconcile_fleet_proof_slot() -> &'static std::sync::R
    LEGACY_TRANSITION_STATE_RECONCILE_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
 }

+fn ilm_recovery_export_fleet_proof_slot() -> &'static std::sync::RwLock<FleetCapabilityProofState> {
+    ILM_RECOVERY_EXPORT_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
+}
+
+fn transition_transaction_compaction_fleet_proof_slot() -> &'static std::sync::RwLock<FleetCapabilityProofState> {
+    TRANSITION_TRANSACTION_COMPACTION_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
+}
+
 fn revoke_fleet_capability_proof_state(state: &mut FleetCapabilityProofState) {
    if let Some(proof) = state.proof.take() {
        proof.generation.revoke();
@@ -438,6 +464,33 @@ pub(crate) fn remote_version_state_fleet_proof_matches(proof: &RemoteVersionStat
    fleet_capability_proof_matches(remote_version_state_fleet_proof_slot(), &proof.0)
 }

+pub(crate) fn acquire_transition_transaction_compaction_fleet_proof() -> Option<TransitionTransactionCompactionFleetProofToken> {
+    let expected_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get()?;
+    let state = transition_transaction_compaction_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    let token = acquire_fleet_capability_proof_from(&state, expected_topology, Instant::now())?;
+    let permit = state.proof.as_ref()?.generation.try_acquire()?;
+    Some(TransitionTransactionCompactionFleetProofToken { token, _permit: permit })
+}
+
+pub(crate) fn transition_transaction_compaction_fleet_proof_matches(
+    proof: &TransitionTransactionCompactionFleetProofToken,
+) -> bool {
+    let Some(expected_topology) = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() else {
+        return false;
+    };
+    let state = transition_transaction_compaction_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    proof._permit.generation.is_accepting()
+        && fleet_capability_proof_matches_at(&state, &proof.token, expected_topology, Instant::now())
+        && state
+            .proof
+            .as_ref()
+            .is_some_and(|current| Arc::ptr_eq(&current.generation, &proof._permit.generation))
+}
+
 pub fn acquire_cross_pool_fence_fleet_proof() -> Option<CrossPoolFenceFleetProofToken> {
    let expected_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get()?;
    let state = cross_pool_fence_fleet_proof_slot()
@@ -573,6 +626,117 @@ pub async fn legacy_transition_state_reconcile_fleet_proof_matches(
    .await
 }

+pub async fn acquire_ilm_recovery_export_fleet_proof() -> Option<IlmRecoveryExportFleetProofToken> {
+    let expected_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get()?;
+    let proof = {
+        let state = ilm_recovery_export_fleet_proof_slot()
+            .read()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        acquire_ilm_recovery_export_fleet_proof_from(&state, expected_topology, Instant::now())?
+    };
+    let observed = observe_ilm_recovery_export_fleet(expected_topology).await?;
+    let state = ilm_recovery_export_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    ilm_recovery_export_fleet_proof_matches_observation_at(&state, &proof, expected_topology, &observed, Instant::now())
+        .then_some(proof)
+}
+
+fn acquire_ilm_recovery_export_fleet_proof_from(
+    state: &FleetCapabilityProofState,
+    expected_topology: &str,
+    now: Instant,
+) -> Option<IlmRecoveryExportFleetProofToken> {
+    let token = acquire_fleet_capability_proof_from(state, expected_topology, now)?;
+    let permit = state.proof.as_ref()?.generation.try_acquire()?;
+    Some(IlmRecoveryExportFleetProofToken { token, _permit: permit })
+}
+
+pub async fn ilm_recovery_export_fleet_proof_matches(proof: &IlmRecoveryExportFleetProofToken) -> bool {
+    let Some(expected_topology) = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() else {
+        return false;
+    };
+    {
+        let state = ilm_recovery_export_fleet_proof_slot()
+            .read()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        if !ilm_recovery_export_fleet_proof_matches_at(&state, proof, expected_topology, Instant::now()) {
+            return false;
+        }
+    }
+    let Some(observed) = observe_ilm_recovery_export_fleet(expected_topology).await else {
+        return false;
+    };
+    let state = ilm_recovery_export_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    ilm_recovery_export_fleet_proof_matches_observation_at(&state, proof, expected_topology, &observed, Instant::now())
+}
+
+pub fn ilm_recovery_export_topology_generation(proof: &IlmRecoveryExportFleetProofToken) -> String {
+    let mut hasher = Sha256::new();
+    hasher.update(b"rustfs-ilm-recovery-export-topology-v1\0");
+    hasher.update(proof.token.topology_fingerprint.as_bytes());
+    rustfs_utils::crypto::hex(hasher.finalize().as_slice())
+}
+
+pub fn ilm_recovery_export_member_epochs_sha256(proof: &IlmRecoveryExportFleetProofToken) -> String {
+    let encoded = serde_json::to_vec(proof.token.peer_epochs.as_ref()).expect("member epoch map is JSON encodable");
+    let mut hasher = Sha256::new();
+    hasher.update(b"rustfs-ilm-recovery-export-members-v1\0");
+    hasher.update(encoded);
+    rustfs_utils::crypto::hex(hasher.finalize().as_slice())
+}
+
+pub fn ilm_recovery_export_local_process_epoch() -> Uuid {
+    *ILM_RECOVERY_EXPORT_LOCAL_PROCESS_EPOCH
+}
+
+fn ilm_recovery_export_fleet_proof_matches_at(
+    state: &FleetCapabilityProofState,
+    proof: &IlmRecoveryExportFleetProofToken,
+    expected_topology: &str,
+    now: Instant,
+) -> bool {
+    proof._permit.generation.is_accepting()
+        && fleet_capability_proof_matches_at(state, &proof.token, expected_topology, now)
+        && state
+            .proof
+            .as_ref()
+            .is_some_and(|current| Arc::ptr_eq(&current.generation, &proof._permit.generation))
+}
+
+fn ilm_recovery_export_fleet_proof_matches_observation_at(
+    state: &FleetCapabilityProofState,
+    proof: &IlmRecoveryExportFleetProofToken,
+    expected_topology: &str,
+    observed: &BTreeMap<String, Uuid>,
+    now: Instant,
+) -> bool {
+    ilm_recovery_export_fleet_proof_matches_at(state, proof, expected_topology, now)
+        && proof.token.peer_epochs.as_ref() == observed
+}
+
+async fn observe_ilm_recovery_export_fleet(expected_topology: &str) -> Option<BTreeMap<String, Uuid>> {
+    #[cfg(test)]
+    {
+        let state = ilm_recovery_export_fleet_proof_slot()
+            .read()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        if fleet_capability_proof_valid_at(state.proof.as_ref(), expected_topology, Instant::now()) {
+            return state.proof.as_ref().map(|proof| proof.peer_epochs.as_ref().clone());
+        }
+    }
+    let notification_sys = get_global_notification_sys()?;
+    timeout(
+        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+        notification_sys.probe_ilm_recovery_export_fleet(expected_topology),
+    )
+    .await
+    .ok()?
+    .ok()
+}
+
 async fn legacy_transition_state_reconcile_fleet_proof_matches_with_observer<F, Fut>(
    slot: &std::sync::RwLock<FleetCapabilityProofState>,
    proof: &LegacyTransitionStateReconcileFleetProofToken,
@@ -661,7 +825,7 @@ pub(crate) fn install_cross_pool_fence_fleet_proof_for_test() {
        state.proof.clone()
    } else {
        Some(FleetCapabilityProof::new(
-            topology,
+            topology.clone(),
            Arc::new(BTreeMap::new()),
            now + Duration::from_secs(60 * 60),
        ))
@@ -694,6 +858,21 @@ pub(crate) fn install_cross_pool_fence_fleet_proof_for_test() {
    decommission_state.topology_conflict = false;
    decommission_state.draining_generation = None;
    decommission_state.proof = proof.as_ref().map(FleetCapabilityProof::with_fresh_generation);
+    drop(decommission_state);
+    let mut export_state = ilm_recovery_export_fleet_proof_slot()
+        .write()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    if !fleet_capability_proof_valid_at(export_state.proof.as_ref(), &topology, now) {
+        debug_assert!(
+            export_state
+                .proof
+                .as_ref()
+                .is_none_or(|current| current.generation.is_drained())
+        );
+        export_state.topology_conflict = false;
+        export_state.draining_generation = None;
+        export_state.proof = proof.as_ref().map(FleetCapabilityProof::with_fresh_generation);
+    }
 }

 #[cfg(test)]
@@ -934,6 +1113,35 @@ pub(crate) fn install_remote_version_state_fleet_proof_for_test(topology_fingerp
    RemoteVersionStateFleetProofGuard
 }

+#[cfg(all(test, feature = "test-util"))]
+pub(crate) struct TransitionTransactionCompactionFleetProofGuard;
+
+#[cfg(all(test, feature = "test-util"))]
+impl Drop for TransitionTransactionCompactionFleetProofGuard {
+    fn drop(&mut self) {
+        revoke_fleet_capability_proof(transition_transaction_compaction_fleet_proof_slot());
+    }
+}
+
+#[cfg(all(test, feature = "test-util"))]
+pub(crate) fn install_transition_transaction_compaction_fleet_proof_for_test(
+    topology_fingerprint: &str,
+) -> TransitionTransactionCompactionFleetProofGuard {
+    let _ = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.set(topology_fingerprint.to_string());
+    let effective_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY
+        .get()
+        .expect("transition transaction compaction test topology should be initialized");
+    if let Some(err) = publish_fleet_capability_probe_result(
+        transition_transaction_compaction_fleet_proof_slot(),
+        effective_topology,
+        Ok(BTreeMap::new()),
+        Instant::now(),
+    ) {
+        panic!("test proof installation must not fail: {err}");
+    }
+    TransitionTransactionCompactionFleetProofGuard
+}
+
 fn insert_remote_version_state_peer(peer_epochs: &mut BTreeMap<String, Uuid>, peer: String, epoch: Uuid) -> Result<()> {
    if epoch.is_nil() || peer_epochs.values().any(|existing| *existing == epoch) || peer_epochs.insert(peer, epoch).is_some() {
        return Err(Error::other("remote version state capability peer identity is invalid"));
@@ -950,6 +1158,8 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
                tier_delete_journal_fleet_proof_slot(),
                decommission_target_fence_fleet_proof_slot(),
                legacy_transition_state_reconcile_fleet_proof_slot(),
+                ilm_recovery_export_fleet_proof_slot(),
+                transition_transaction_compaction_fleet_proof_slot(),
            ] {
                mark_fleet_capability_topology_conflict(slot);
            }
@@ -959,29 +1169,59 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {

    tokio::spawn(async move {
        loop {
-            let result = match get_global_notification_sys() {
-                Some(notification_sys) => {
-                    match timeout(
+            let notification_sys = get_global_notification_sys();
+            let remote_version_state_probe = async {
+                match notification_sys.as_ref() {
+                    Some(notification_sys) => timeout(
                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
                        notification_sys.probe_remote_version_state_fleet(&topology_fingerprint),
                    )
                    .await
-                    {
-                        Ok(result) => result,
-                        Err(_) => Err(Error::other("remote version state fleet capability probe timed out")),
-                    }
+                    .unwrap_or_else(|_| Err(Error::other("remote version state fleet capability probe timed out"))),
+                    None => Err(Error::other("remote version state fleet capability notification system is unavailable")),
                }
-                None => Err(Error::other("remote version state fleet capability notification system is unavailable")),
            };
-            let fence_probe = match get_global_notification_sys() {
-                Some(notification_sys) => timeout(
-                    REMOTE_VERSION_STATE_PROBE_TIMEOUT,
-                    notification_sys.probe_cross_pool_fence_fleet(&topology_fingerprint),
-                )
-                .await
-                .unwrap_or_else(|_| Err(Error::other("cross-pool fence fleet capability probe timed out"))),
-                None => Err(Error::other("cross-pool fence fleet capability notification system is unavailable")),
+            let cross_pool_fence_probe = async {
+                match notification_sys.as_ref() {
+                    Some(notification_sys) => timeout(
+                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                        notification_sys.probe_cross_pool_fence_fleet(&topology_fingerprint),
+                    )
+                    .await
+                    .unwrap_or_else(|_| Err(Error::other("cross-pool fence fleet capability probe timed out"))),
+                    None => Err(Error::other("cross-pool fence fleet capability notification system is unavailable")),
+                }
            };
+            let recovery_export_probe = async {
+                match notification_sys.as_ref() {
+                    Some(notification_sys) => timeout(
+                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                        notification_sys.probe_ilm_recovery_export_fleet(&topology_fingerprint),
+                    )
+                    .await
+                    .unwrap_or_else(|_| Err(Error::other("ILM recovery export fleet capability probe timed out"))),
+                    None => Err(Error::other("ILM recovery export fleet capability notification system is unavailable")),
+                }
+            };
+            let transition_transaction_compaction_probe = async {
+                match notification_sys.as_ref() {
+                    Some(notification_sys) => timeout(
+                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                        notification_sys.probe_transition_transaction_compaction_fleet(&topology_fingerprint),
+                    )
+                    .await
+                    .unwrap_or_else(|_| Err(Error::other("transition transaction compaction fleet capability probe timed out"))),
+                    None => Err(Error::other(
+                        "transition transaction compaction fleet capability notification system is unavailable",
+                    )),
+                }
+            };
+            let (result, fence_probe, recovery_export_result, transition_transaction_compaction_result) = tokio::join!(
+                remote_version_state_probe,
+                cross_pool_fence_probe,
+                recovery_export_probe,
+                transition_transaction_compaction_probe
+            );
            let (fence_result, journal_result, decommission_target_fence_result, reconcile_result) = match fence_probe {
                Ok((peer_epochs, minimum_version)) => cross_pool_fence_policy_results(peer_epochs, minimum_version),
                Err(err) => {
@@ -1004,6 +1244,8 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
                revoke_fleet_capability_proof(tier_delete_journal_fleet_proof_slot());
                revoke_fleet_capability_proof(decommission_target_fence_fleet_proof_slot());
                revoke_fleet_capability_proof(legacy_transition_state_reconcile_fleet_proof_slot());
+                revoke_fleet_capability_proof(ilm_recovery_export_fleet_proof_slot());
+                revoke_fleet_capability_proof(transition_transaction_compaction_fleet_proof_slot());
            } else if let Some(err) = publish_fleet_capability_probe_result(
                remote_version_state_fleet_proof_slot(),
                &topology_fingerprint,
@@ -1030,6 +1272,42 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
                    "notification capability probe"
                );
            }
+            if !topology_conflict
+                && let Some(err) = publish_fleet_capability_probe_result(
+                    ilm_recovery_export_fleet_proof_slot(),
+                    &topology_fingerprint,
+                    recovery_export_result,
+                    Instant::now(),
+                )
+            {
+                debug!(
+                    event = EVENT_NOTIFICATION_CAPABILITY_PROBE,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_NOTIFICATION,
+                    capability = "ilm_recovery_export_v1",
+                    state = "failed_closed",
+                    error = %err,
+                    "notification capability probe"
+                );
+            }
+            if !topology_conflict
+                && let Some(err) = publish_fleet_capability_probe_result(
+                    transition_transaction_compaction_fleet_proof_slot(),
+                    &topology_fingerprint,
+                    transition_transaction_compaction_result,
+                    Instant::now(),
+                )
+            {
+                debug!(
+                    event = EVENT_NOTIFICATION_CAPABILITY_PROBE,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_NOTIFICATION,
+                    capability = "transition_transaction_compaction_v1",
+                    state = "failed_closed",
+                    error = %err,
+                    "notification capability probe"
+                );
+            }
            if !topology_conflict
                && let Some(err) = publish_fleet_capability_probe_result(
                    tier_delete_journal_fleet_proof_slot(),
@@ -1152,6 +1430,28 @@ impl NotificationSys {
        Ok(peer_epochs)
    }

+    async fn probe_transition_transaction_compaction_fleet(&self, topology_fingerprint: &str) -> Result<BTreeMap<String, Uuid>> {
+        if self.peer_clients.len() != self.peer_topology_hosts.len() {
+            return Err(Error::other(
+                "transition transaction compaction capability fleet membership is incomplete",
+            ));
+        }
+        let probes = self.peer_clients.iter().map(|client| async {
+            let client = client
+                .as_ref()
+                .ok_or_else(|| Error::other("transition transaction compaction capability peer is unreachable"))?;
+            client
+                .probe_transition_transaction_compaction(topology_fingerprint.to_string())
+                .await
+        });
+        let mut peer_epochs = BTreeMap::new();
+        for result in join_all(probes).await {
+            let (peer, epoch) = result?;
+            insert_remote_version_state_peer(&mut peer_epochs, peer, epoch)?;
+        }
+        Ok(peer_epochs)
+    }
+
    async fn probe_cross_pool_fence_fleet(&self, topology_fingerprint: &str) -> Result<(BTreeMap<String, Uuid>, u32)> {
        if self.peer_clients.len() != self.peer_topology_hosts.len() {
            return Err(Error::other("cross-pool fence capability fleet membership is incomplete"));
@@ -1174,6 +1474,46 @@ impl NotificationSys {
        }
        Ok((peer_epochs, minimum_version))
    }
+
+    async fn probe_ilm_recovery_export_fleet(&self, topology_fingerprint: &str) -> Result<BTreeMap<String, Uuid>> {
+        if self.peer_clients.len() != self.peer_topology_hosts.len() {
+            return Err(Error::other("ILM recovery export capability fleet membership is incomplete"));
+        }
+        let local_member = runtime_sources::local_node_name().await;
+        if local_member.trim().is_empty() {
+            return Err(Error::other("ILM recovery export local member identity is unavailable"));
+        }
+        let mut peer_epochs = BTreeMap::new();
+        insert_remote_version_state_peer(&mut peer_epochs, local_member.clone(), ilm_recovery_export_local_process_epoch())?;
+        let probes = self.peer_clients.iter().map(|client| async {
+            let client = client
+                .as_ref()
+                .ok_or_else(|| Error::other("ILM recovery export capability peer is unreachable"))?;
+            client.probe_ilm_recovery_export(topology_fingerprint.to_string()).await
+        });
+        for result in join_all(probes).await {
+            let (peer, epoch) = result?;
+            insert_remote_version_state_peer(&mut peer_epochs, peer, epoch)?;
+        }
+        validate_ilm_recovery_export_members(&self.peer_topology_hosts, &local_member, &peer_epochs)?;
+        Ok(peer_epochs)
+    }
+}
+
+fn validate_ilm_recovery_export_members(
+    expected_remote_members: &[String],
+    local_member: &str,
+    observed: &BTreeMap<String, Uuid>,
+) -> Result<()> {
+    let expected = expected_remote_members
+        .iter()
+        .cloned()
+        .chain(std::iter::once(local_member.to_string()))
+        .collect::<BTreeSet<_>>();
+    if expected.len() != expected_remote_members.len().saturating_add(1) || observed.keys().ne(expected.iter()) {
+        return Err(Error::other("ILM recovery export capability fleet membership does not match topology"));
+    }
+    Ok(())
 }

 /// Rolling tier activity summed over every cluster member that answered, with
@@ -2341,11 +2681,70 @@ impl NotificationSys {
        Ok(snapshots)
    }

-    pub async fn acknowledge_scanner_dirty_usage(&self, acknowledgements: Vec<(String, String, u64)>) -> Result<bool> {
+    pub async fn scanner_scoped_dirty_usage_capabilities(
+        &self,
+        acknowledgements: Vec<ScannerDirtyUsageAcknowledgement>,
+    ) -> Result<bool> {
        let mut by_host = HashMap::with_capacity(acknowledgements.len());
-        for (host, instance_id, generation) in acknowledgements {
-            if by_host.insert(host.clone(), (instance_id, generation)).is_some() {
-                return Err(Error::other(format!("duplicate scanner dirty usage acknowledgement target: {host}")));
+        for acknowledgement in acknowledgements {
+            let host = match &acknowledgement {
+                ScannerDirtyUsageAcknowledgement::Scoped { host, .. } => host.clone(),
+                ScannerDirtyUsageAcknowledgement::Generation { .. } => {
+                    return Err(Error::other("scanner scoped dirty usage capability requires scoped acknowledgements"));
+                }
+            };
+            if by_host.insert(host.clone(), acknowledgement).is_some() {
+                return Err(Error::other("duplicate scanner dirty usage acknowledgement target"));
+            }
+        }
+
+        let clients = self
+            .peer_clients
+            .iter()
+            .flatten()
+            .map(|client| (client.grid_host.clone(), client.clone()))
+            .collect::<HashMap<_, _>>();
+        let mut futures = Vec::with_capacity(by_host.len());
+        for (host, acknowledgement) in by_host {
+            let Some(client) = clients.get(&host).cloned() else {
+                return Err(Error::other("scanner scoped dirty usage capability failed: peer is not reachable"));
+            };
+            futures.push(async move {
+                let ScannerDirtyUsageAcknowledgement::Scoped {
+                    owner_id,
+                    instance_id,
+                    entries,
+                    ..
+                } = acknowledgement
+                else {
+                    unreachable!("scoped acknowledgement was validated before probing");
+                };
+                timeout(
+                    SCANNER_ACTIVITY_PROBE_TIMEOUT,
+                    client.scanner_scoped_dirty_usage_capability(owner_id, instance_id, entries),
+                )
+                .await
+                .map_err(|_| Error::other("scanner scoped dirty usage capability timed out"))?
+            });
+        }
+
+        for result in join_all(futures).await {
+            if !result? {
+                return Ok(false);
+            }
+        }
+        Ok(true)
+    }
+
+    pub async fn acknowledge_scanner_dirty_usage(&self, acknowledgements: Vec<ScannerDirtyUsageAcknowledgement>) -> Result<bool> {
+        let mut by_host = HashMap::with_capacity(acknowledgements.len());
+        for acknowledgement in acknowledgements {
+            let host = match &acknowledgement {
+                ScannerDirtyUsageAcknowledgement::Generation { host, .. }
+                | ScannerDirtyUsageAcknowledgement::Scoped { host, .. } => host.clone(),
+            };
+            if by_host.insert(host.clone(), acknowledgement).is_some() {
+                return Err(Error::other("duplicate scanner dirty usage acknowledgement target"));
            }
        }

@@ -2357,18 +2756,34 @@ impl NotificationSys {
            .collect::<HashMap<_, _>>();
        let mut failures = Vec::new();
        let mut futures = Vec::with_capacity(by_host.len());
-        for (host, (instance_id, generation)) in by_host {
+        for (host, acknowledgement) in by_host {
            let Some(client) = clients.get(&host).cloned() else {
                failures.push(format!("peer {host} scanner dirty usage acknowledgement failed: peer is not reachable"));
                continue;
            };
            futures.push(async move {
-                let result = scanner_activity_with_timeout(
-                    SCANNER_ACTIVITY_PROBE_TIMEOUT,
-                    &host,
-                    client.acknowledge_scanner_dirty_usage(instance_id, generation),
-                )
-                .await;
+                let result = match acknowledgement {
+                    ScannerDirtyUsageAcknowledgement::Generation {
+                        instance_id, generation, ..
+                    } => {
+                        scanner_activity_with_timeout(
+                            SCANNER_ACTIVITY_PROBE_TIMEOUT,
+                            &host,
+                            client.acknowledge_scanner_dirty_usage(instance_id, generation),
+                        )
+                        .await
+                    }
+                    ScannerDirtyUsageAcknowledgement::Scoped {
+                        owner_id,
+                        instance_id,
+                        entries,
+                        ..
+                    } => {
+                        client
+                            .acknowledge_scanner_scoped_dirty_usage(owner_id, instance_id, entries)
+                            .await
+                    }
+                };
                (host, result)
            });
        }
@@ -3507,6 +3922,81 @@ mod tests {
        assert!(captured != restarted.token());
    }

+    #[test]
+    fn ilm_recovery_export_member_digest_is_order_independent_and_epoch_bound() {
+        let now = Instant::now();
+        let local_epoch = ilm_recovery_export_local_process_epoch();
+        assert!(!local_epoch.is_nil());
+        assert_eq!(local_epoch, ilm_recovery_export_local_process_epoch());
+        let remote_epoch = Uuid::new_v4();
+        let slot = std::sync::RwLock::new(FleetCapabilityProofState::default());
+        let peers = BTreeMap::from([("node-b".to_string(), remote_epoch), ("node-a".to_string(), local_epoch)]);
+        assert!(publish_fleet_capability_probe_result(&slot, "topology-a", Ok(peers), now).is_none());
+        let proof = {
+            let state = slot.read().expect("export proof slot should not poison");
+            acquire_ilm_recovery_export_fleet_proof_from(&state, "topology-a", now).expect("complete fleet should admit export")
+        };
+        let digest = ilm_recovery_export_member_epochs_sha256(&proof);
+
+        let changed_slot = std::sync::RwLock::new(FleetCapabilityProofState::default());
+        let changed = BTreeMap::from([("node-a".to_string(), local_epoch), ("node-b".to_string(), Uuid::new_v4())]);
+        assert!(publish_fleet_capability_probe_result(&changed_slot, "topology-a", Ok(changed), now).is_none());
+        let changed_proof = {
+            let state = changed_slot.read().expect("export proof slot should not poison");
+            acquire_ilm_recovery_export_fleet_proof_from(&state, "topology-a", now).expect("complete fleet should admit export")
+        };
+        assert_ne!(digest, ilm_recovery_export_member_epochs_sha256(&changed_proof));
+    }
+
+    #[test]
+    fn ilm_recovery_export_members_must_match_the_exact_topology() {
+        let expected_remote = vec!["node-b".to_string()];
+        let local = "node-a";
+        let complete = BTreeMap::from([
+            (local.to_string(), Uuid::new_v4()),
+            (expected_remote[0].clone(), Uuid::new_v4()),
+        ]);
+        assert!(validate_ilm_recovery_export_members(&expected_remote, local, &complete).is_ok());
+
+        let unexpected = BTreeMap::from([(local.to_string(), Uuid::new_v4()), ("node-c".to_string(), Uuid::new_v4())]);
+        assert!(validate_ilm_recovery_export_members(&expected_remote, local, &unexpected).is_err());
+        assert!(
+            validate_ilm_recovery_export_members(&[local.to_string()], local, &complete).is_err(),
+            "the configured remote set cannot repeat the local member"
+        );
+    }
+
+    #[test]
+    fn ilm_recovery_export_restart_revokes_authority_until_permit_drains() {
+        let slot = std::sync::RwLock::new(FleetCapabilityProofState::default());
+        let now = Instant::now();
+        let original = BTreeMap::from([("node-a".to_string(), Uuid::new_v4())]);
+        assert!(publish_fleet_capability_probe_result(&slot, "topology-a", Ok(original), now).is_none());
+        let admitted = {
+            let state = slot.read().expect("export proof slot should not poison");
+            acquire_ilm_recovery_export_fleet_proof_from(&state, "topology-a", now).expect("fresh fleet should admit export")
+        };
+
+        let restarted = BTreeMap::from([("node-a".to_string(), Uuid::new_v4())]);
+        let draining = publish_fleet_capability_probe_result(&slot, "topology-a", Ok(restarted.clone()), now)
+            .expect("restart must wait for the admitted export effect window");
+        assert!(draining.to_string().contains("previous generation to drain"));
+        {
+            let state = slot.read().expect("export proof slot should not poison");
+            assert!(!ilm_recovery_export_fleet_proof_matches_at(&state, &admitted, "topology-a", now));
+            assert!(
+                acquire_ilm_recovery_export_fleet_proof_from(&state, "topology-a", now).is_none(),
+                "successor authority must wait for the old effect window to drain"
+            );
+        }
+        drop(admitted);
+        assert!(
+            publish_fleet_capability_probe_result(&slot, "topology-a", Ok(restarted), now + Duration::from_millis(1)).is_none()
+        );
+        let state = slot.read().expect("export proof slot should not poison");
+        assert!(acquire_ilm_recovery_export_fleet_proof_from(&state, "topology-a", now).is_some());
+    }
+
    #[test]
    fn tier_delete_journal_generation_is_stable_across_members_and_process_restarts() {
        let topology = "topology-a";
@@ -4458,15 +4948,28 @@ mod tests {
            peer_topology_hosts: Vec::new(),
        };
        let missing = sys
-            .acknowledge_scanner_dirty_usage(vec![("peer-1".to_string(), "0123456789abcdef0123456789abcdef".to_string(), 7)])
+            .acknowledge_scanner_dirty_usage(vec![ScannerDirtyUsageAcknowledgement::Generation {
+                host: "peer-1".to_string(),
+                instance_id: "0123456789abcdef0123456789abcdef".to_string(),
+                generation: 7,
+            }])
            .await
            .expect_err("a missing acknowledgement target must remain pending");
        assert!(missing.to_string().contains("peer is not reachable"));

        let duplicate = sys
            .acknowledge_scanner_dirty_usage(vec![
-                ("peer-1".to_string(), "0123456789abcdef0123456789abcdef".to_string(), 7),
-                ("peer-1".to_string(), "0123456789abcdef0123456789abcdef".to_string(), 7),
+                ScannerDirtyUsageAcknowledgement::Generation {
+                    host: "peer-1".to_string(),
+                    instance_id: "0123456789abcdef0123456789abcdef".to_string(),
+                    generation: 7,
+                },
+                ScannerDirtyUsageAcknowledgement::Scoped {
+                    host: "peer-1".to_string(),
+                    owner_id: "11111111-1111-1111-1111-111111111111".to_string(),
+                    instance_id: "0123456789abcdef0123456789abcdef".to_string(),
+                    entries: Vec::new(),
+                },
            ])
            .await
            .expect_err("duplicate acknowledgement targets must be rejected");
@@ -14,6 +14,7 @@

 #[cfg(any(test, feature = "test-util"))]
 pub mod test_util;
+#[allow(clippy::module_inception, reason = "preserve the public services::tier::tier path")]
 pub mod tier;
 pub mod tier_admin;
 pub mod tier_config;
@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]

 use byteorder::{ByteOrder, LittleEndian};
 use bytes::Bytes;
@@ -802,7 +800,7 @@ pub enum TierConfigUpdateError {
 }

 enum TierCandidateMutation {
-    Add(TierConfig, bool),
+    Add(Box<TierConfig>, bool),
    Edit(String, TierCreds),
    Remove(String, bool),
    Clear(bool),
@@ -823,7 +821,7 @@ struct PrevalidatedTierCandidateMutation {
 impl TierCandidateMutation {
    fn add(mut config: TierConfig, force: bool) -> std::result::Result<Self, AdminError> {
        normalize_s3_gcs_add_tier_name(&mut config)?;
-        Ok(Self::Add(config, force))
+        Ok(Self::Add(Box::new(config), force))
    }

    fn normalize_add_tier_name(&mut self) -> std::result::Result<(), AdminError> {
@@ -910,7 +908,7 @@ impl TierCandidateMutation {
        match self {
            Self::Add(config, force) => {
                let tier_name = config.name.clone();
-                candidate.add_with_deadline(config, force, deadline).await?;
+                candidate.add_with_deadline(*config, force, deadline).await?;
                Ok(Some(tier_name))
            }
            Self::Edit(tier_name, credentials) => {
@@ -2990,7 +2988,7 @@ fn from_external_tier_config(name: String, ext: ExternalTierConfig) -> io::Resul
    let tier_type = if wasabi_version {
        TierType::Wasabi
    } else {
-        tier_type_from_hint(ext.tier_type_hint.as_deref()).unwrap_or_else(|| match ext.tier_type {
+        tier_type_from_hint(ext.tier_type_hint.as_deref()).unwrap_or(match ext.tier_type {
            EXTERNAL_TIER_TYPE_S3 => TierType::S3,
            EXTERNAL_TIER_TYPE_AZURE => TierType::Azure,
            EXTERNAL_TIER_TYPE_GCS => TierType::GCS,
@@ -3372,28 +3370,23 @@ impl TierConfigMgr {

    pub async fn remove(&mut self, tier_name: &str, force: bool) -> std::result::Result<(), AdminError> {
        self.ensure_generation_is_idle(tier_name)?;
-        let d = self.get_driver(tier_name).await;
-        if let Err(err) = d {
-            if err.code == ERR_TIER_NOT_FOUND.code {
-                return Ok(());
-            } else {
-                return Err(err);
-            }
-        }
+        let driver = match self.get_driver(tier_name).await {
+            Ok(driver) => driver,
+            Err(err) if err.code == ERR_TIER_NOT_FOUND.code => return Ok(()),
+            Err(err) => return Err(err),
+        };
        if !force {
-            if let Ok(driver) = d {
-                match driver.in_use().await {
-                    Err(err) => {
-                        let mut e = ERR_TIER_PERM_ERR.clone();
-                        e.message.push('.');
-                        e.message.push_str(&err.to_string());
-                        return Err(e);
-                    }
-                    Ok(in_use) if in_use => {
-                        return Err(ERR_TIER_BACKEND_NOT_EMPTY.clone());
-                    }
-                    _ => {}
+            match driver.in_use().await {
+                Err(err) => {
+                    let mut e = ERR_TIER_PERM_ERR.clone();
+                    e.message.push('.');
+                    e.message.push_str(&err.to_string());
+                    return Err(e);
                }
+                Ok(in_use) if in_use => {
+                    return Err(ERR_TIER_BACKEND_NOT_EMPTY.clone());
+                }
+                _ => {}
            }
        }
        self.tiers.remove(tier_name);
@@ -3402,21 +3395,12 @@ impl TierConfigMgr {
    }

    pub async fn verify(&mut self, tier_name: &str) -> std::result::Result<(), std::io::Error> {
-        let d = match self.get_driver(tier_name).await {
-            Ok(d) => d,
-            Err(err) => {
-                return Err(std::io::Error::other(err));
-            }
-        };
-        if let Err(err) = check_warm_backend(Some(d)).await {
-            return Err(std::io::Error::other(err));
-        } else {
-            return Ok(());
-        }
+        let driver = self.get_driver(tier_name).await.map_err(std::io::Error::other)?;
+        check_warm_backend(Some(driver)).await.map_err(std::io::Error::other)
    }

    pub fn empty(&self) -> bool {
-        self.list_tiers().len() == 0
+        self.tiers.is_empty()
    }

    pub fn tier_type(&self, tier_name: &str) -> String {
@@ -3429,7 +3413,7 @@ impl TierConfigMgr {

    pub fn list_tiers(&self) -> Vec<TierConfig> {
        let mut tier_cfgs = Vec::<TierConfig>::new();
-        for (_, tier) in self.tiers.iter() {
+        for tier in self.tiers.values() {
            let tier = tier.redacted();
            tier_cfgs.push(tier);
        }
@@ -5458,7 +5442,7 @@ impl TierConfigMgr {
        let manager = handle.read().await;
        let published_digest = if intents
            .iter()
-            .any(|recovered| recovered.is_peer_only_terminal() && recovered.intent.state == TierMutationIntentState::Committed)
+            .any(|recovered| recovered.intent.state == TierMutationIntentState::Committed)
        {
            Some(tier_config_candidate_digest(&manager).map_err(|err| {
                let mut admin_err = ERR_TIER_INVALID_CONFIG.clone();
@@ -5468,18 +5452,29 @@ impl TierConfigMgr {
        } else {
            None
        };
+        let locally_published_committed_mutations = intents
+            .iter()
+            .filter(|recovered| {
+                recovered.intent.state == TierMutationIntentState::Committed
+                    && published_digest == Some(recovered.intent.candidate_digest)
+            })
+            .map(|recovered| recovered.intent.mutation_id)
+            .collect::<HashSet<_>>();
        let mut prepared_mutation_blocks = HashMap::new();
        let mut committed_mutation_blocks: HashMap<String, HashSet<uuid::Uuid>> = HashMap::new();
        for recovered in intents {
-            let settled_tombstone = recovered.is_peer_only_terminal()
-                && match recovered.intent.state {
-                    TierMutationIntentState::Aborted => true,
-                    TierMutationIntentState::Committed => {
-                        !retain_missing_mutation_blocks || published_digest == Some(recovered.intent.candidate_digest)
-                    }
-                    TierMutationIntentState::Prepared => false,
-                };
-            if settled_tombstone {
+            // A matching in-memory manager has already crossed the local
+            // publication boundary. Keep replaying and durably cleaning the
+            // record, but do not re-fence object operations while that
+            // terminal work finishes.
+            let skip_runtime_fence = locally_published_committed_mutations.contains(&recovered.intent.mutation_id)
+                || (recovered.is_peer_only_terminal()
+                    && match recovered.intent.state {
+                        TierMutationIntentState::Aborted => true,
+                        TierMutationIntentState::Committed => !retain_missing_mutation_blocks,
+                        TierMutationIntentState::Prepared => false,
+                    });
+            if skip_runtime_fence {
                continue;
            }
            Self::collect_prepared_mutation_intent_block(&mut prepared_mutation_blocks, &recovered.intent)?;
@@ -5492,6 +5487,9 @@ impl TierConfigMgr {
        }
        if retain_missing_mutation_blocks {
            for (tier_name, mutation_id) in &runtime.prepared_mutation_blocks {
+                if locally_published_committed_mutations.contains(mutation_id) {
+                    continue;
+                }
                match prepared_mutation_blocks.entry(tier_name.clone()) {
                    Entry::Vacant(entry) => {
                        entry.insert(*mutation_id);
@@ -5505,10 +5503,15 @@ impl TierConfigMgr {
                }
            }
            for (tier_name, mutation_ids) in &runtime.committed_mutation_blocks {
-                committed_mutation_blocks
-                    .entry(tier_name.clone())
-                    .or_default()
-                    .extend(mutation_ids);
+                for mutation_id in mutation_ids {
+                    if locally_published_committed_mutations.contains(mutation_id) {
+                        continue;
+                    }
+                    committed_mutation_blocks
+                        .entry(tier_name.clone())
+                        .or_default()
+                        .insert(*mutation_id);
+                }
            }
        }
        let changed = runtime.prepared_mutation_blocks != prepared_mutation_blocks
@@ -7116,7 +7119,8 @@ mod tests {
        let err = expect_decode_err(&encode_fixture(&wrong_hint));
        assert!(err.to_string().contains("inconsistent Wasabi type discriminators"), "{err}");

-        let poison_fields: [(&str, fn(&mut ExternalTierS3)); 6] = [
+        type WasabiPoisonField = (&'static str, fn(&mut ExternalTierS3));
+        let poison_fields: [WasabiPoisonField; 6] = [
            ("storage_class", |s3| s3.storage_class = "GLACIER".to_string()),
            ("aws_role", |s3| s3.aws_role = true),
            ("web_identity_token", |s3| s3.aws_role_web_identity_token_file = "/tmp/token".to_string()),
@@ -8237,7 +8241,11 @@ mod tests {
                            peer_calls.clone(),
                            Ok(PeerTierMutationState::Committed),
                        )],
-                        TierConfigMgr::update_candidate_with_config_lock(&manager, store, TierCandidateMutation::Add(tier, true)),
+                        TierConfigMgr::update_candidate_with_config_lock(
+                            &manager,
+                            store,
+                            TierCandidateMutation::Add(Box::new(tier), true),
+                        ),
                    ),
                )
                .await
@@ -8276,7 +8284,7 @@ mod tests {
            let add = TIER_DRIVER_TEST_FACTORY.scope(
                factory,
                apply_tier_candidate_mutation(
-                    TierCandidateMutation::Add(build_rustfs_tier("COLD-DEADLINE"), false),
+                    TierCandidateMutation::Add(Box::new(build_rustfs_tier("COLD-DEADLINE")), false),
                    &mut candidate,
                    deadline,
                ),
@@ -9095,7 +9103,9 @@ mod tests {
    fn decode_hex_fixture(hex: &str) -> Vec<u8> {
        assert_eq!(hex.len() % 2, 0, "hex fixture must contain complete bytes");
        hex.as_bytes()
-            .chunks_exact(2)
+            .as_chunks::<2>()
+            .0
+            .iter()
            .map(|pair| {
                let pair = std::str::from_utf8(pair).expect("hex fixture should be ASCII");
                u8::from_str_radix(pair, 16).expect("hex fixture should contain only hexadecimal digits")
@@ -11109,7 +11119,7 @@ mod tests {
                            store.clone(),
                            candidate,
                            version,
-                            TierCandidateMutation::Add(build_rustfs_tier("COLD-A"), true),
+                            TierCandidateMutation::Add(Box::new(build_rustfs_tier("COLD-A")), true),
                            update,
                            None,
                        )
@@ -11298,6 +11308,115 @@ mod tests {
        );
    }

+    #[tokio::test]
+    async fn published_dual_terminal_intent_does_not_restore_local_runtime_fence_during_replay() {
+        use crate::services::tier::tier_mutation_intent::save_tier_mutation_intent_record;
+
+        let store = Arc::new(CasConfigStore::default());
+        let mut persisted = empty_mgr();
+        persisted.tiers.insert("COLD-A".to_string(), build_rustfs_tier("COLD-A"));
+        persisted
+            .save_tiering_config_if_current(store.clone(), None)
+            .await
+            .expect("published tier config fixture should persist");
+        let (_, current_etag) = load_tier_config_for_update(store.clone())
+            .await
+            .expect("published tier config fixture should load with metadata");
+        let current_etag = current_etag.expect("published tier config fixture should have an ETag");
+        let affected_targets = build_tier_mutation_affected_targets(
+            TierMutationIntentKind::Add,
+            HashSet::from(["COLD-A".to_string()]),
+            &empty_mgr(),
+            &persisted,
+        )
+        .expect("published AddTier targets should build");
+        let mut intent = build_coordinator_tier_mutation_intent(TierMutationIntentKind::Add, None, &persisted, affected_targets)
+            .expect("published AddTier intent should build")
+            .expect("published AddTier should require a durable intent");
+        intent
+            .advance(TierMutationIntentState::Committed, Some(current_etag))
+            .expect("published AddTier intent should commit");
+        save_tier_coordinator_mutation_intent_record_if_absent(store.clone(), &intent)
+            .await
+            .expect("published coordinator intent should persist");
+        save_tier_mutation_intent_record(store.clone(), &intent)
+            .await
+            .expect("published peer intent should persist");
+
+        let manager = TierConfigMgr::new();
+        {
+            let mut guard = manager.write().await;
+            install_lease_backend(&mut guard, "COLD-A", LeaseTestBackend::ready("published"));
+        }
+        {
+            let guard = manager.read().await;
+            assert_eq!(
+                tier_config_candidate_digest(&guard).expect("published manager digest should build"),
+                intent.candidate_digest
+            );
+        }
+        TierConfigMgr::apply_committed_mutation_intent_block(&manager, &intent)
+            .await
+            .expect("pre-existing committed runtime fence should install");
+        assert!(
+            TierConfigMgr::acquire_operation_lease(&manager, "COLD-A").await.is_err(),
+            "fixture must begin with the committed runtime fence installed"
+        );
+
+        let started = Arc::new(Notify::new());
+        let release = Arc::new(tokio::sync::Semaphore::new(0));
+        TIER_MUTATION_TEST_PEERS
+            .scope(
+                vec![Arc::new(BlockingCommitTierMutationPeer {
+                    started: started.clone(),
+                    release: release.clone(),
+                })],
+                async {
+                    let reload = TierConfigMgr::reload_handle_with(&manager, store.clone());
+                    tokio::pin!(reload);
+                    tokio::time::timeout(Duration::from_secs(5), async {
+                        tokio::select! {
+                            result = &mut reload => panic!("reload finished before terminal replay was released: {result:?}"),
+                            _ = started.notified() => {}
+                        }
+                    })
+                    .await
+                    .expect("terminal replay should reach the blocking peer");
+
+                    let lease = TierConfigMgr::acquire_operation_lease(&manager, "COLD-A")
+                        .await
+                        .expect("terminal cleanup must not re-fence an already-published tier");
+                    drop(lease);
+                    release.add_permits(1);
+                    tokio::time::timeout(Duration::from_secs(5), &mut reload)
+                        .await
+                        .expect("terminal replay should finish after the peer responds")
+                        .expect("terminal replay should succeed after the peer responds");
+                },
+            )
+            .await;
+
+        assert!(manager.read().await.tiers.contains_key("COLD-A"));
+        assert!(
+            TierConfigMgr::load_coordinator_mutation_intents(store.clone())
+                .await
+                .expect("coordinator cleanup should be readable")
+                .is_empty()
+        );
+        assert_eq!(
+            TierConfigMgr::load_tier_mutation_intents(store)
+                .await
+                .expect("retained peer tombstone should be readable"),
+            vec![intent]
+        );
+        let guard = manager.read().await;
+        let runtime = registered_tier_driver_runtime(&guard).expect("runtime should remain registered");
+        assert!(
+            lock_unpoisoned(&runtime).committed_mutation_blocks.is_empty(),
+            "retained terminal evidence must not restore the published runtime fence"
+        );
+    }
+
    #[tokio::test]
    async fn peer_terminal_tombstone_gc_uses_etag_and_retains_racing_replacement() {
        use crate::services::tier::tier_mutation_intent::{
@@ -11597,8 +11716,9 @@ mod tests {
            assert!(merged[0].has_peer_record && merged[0].has_coordinator_record);
        }

-        let err = TierConfigMgr::merge_mutation_recovery_intents(&[committed.clone()], &[prepared.clone()])
-            .expect_err("a peer committed record cannot outrun the coordinator commit order");
+        let err =
+            TierConfigMgr::merge_mutation_recovery_intents(std::slice::from_ref(&committed), std::slice::from_ref(&prepared))
+                .expect_err("a peer committed record cannot outrun the coordinator commit order");
        assert!(err.to_string().contains("conflicting states"), "{err}");

        let mut conflicting_identity = prepared.clone();
@@ -13443,9 +13563,11 @@ mod tests {
        let build = tokio::spawn(async move { TierConfigMgr::acquire_operation_lease(&build_manager, cold_tier).await });
        barrier.arrived.notified().await;

-        tokio::time::timeout(Duration::from_millis(100), manager.read())
-            .await
-            .expect("cold driver construction must not block manager readers");
+        drop(
+            tokio::time::timeout(Duration::from_millis(100), manager.read())
+                .await
+                .expect("cold driver construction must not block manager readers"),
+        );
        let tier_b = tokio::time::timeout(Duration::from_millis(100), TierConfigMgr::acquire_operation_lease(&manager, "COLD-B"))
            .await
            .expect("cold tier A construction must not block tier B")
@@ -13648,9 +13770,11 @@ mod tests {
        let verify_manager = manager.clone();
        let verify = tokio::spawn(async move { TierConfigMgr::verify_without_manager_lock(&verify_manager, "COLD-A").await });
        started.notified().await;
-        tokio::time::timeout(Duration::from_millis(100), manager.read())
-            .await
-            .expect("slow verify must not hold the manager lock");
+        drop(
+            tokio::time::timeout(Duration::from_millis(100), manager.read())
+                .await
+                .expect("slow verify must not hold the manager lock"),
+        );
        release.add_permits(1);
        verify.await.expect("verify task should join").expect("verify should finish");
    }
@@ -14058,9 +14182,11 @@ mod tests {
                vec!["COLD-A".to_string()]
            );
        }
-        tokio::time::timeout(Duration::from_secs(1), manager.read())
-            .await
-            .expect("manager reads must not wait for tier A leases");
+        drop(
+            tokio::time::timeout(Duration::from_secs(1), manager.read())
+                .await
+                .expect("manager reads must not wait for tier A leases"),
+        );
        let next_b = tokio::time::timeout(Duration::from_secs(1), TierConfigMgr::acquire_operation_lease(&manager, "COLD-B"))
            .await
            .expect("tier B lease acquisition must not wait for tier A")
@@ -14493,7 +14619,7 @@ mod tests {
            "https://example-compat.invalid"
        );
        let runtime = registered_tier_driver_runtime(&manager_guard).expect("runtime sidecar should remain registered");
-        assert!(lock_unpoisoned(&runtime).generations.get("COLD-A").is_none());
+        assert!(!lock_unpoisoned(&runtime).generations.contains_key("COLD-A"));
    }

    #[derive(Debug)]
@@ -15133,15 +15259,12 @@ mod tests {
                .filter(|object| object.bucket == bucket && object.name.starts_with(prefix))
                .cloned()
                .collect();
-            objects.sort_by(|left, right| tier_test_object_marker(left).cmp(&tier_test_object_marker(right)));
+            objects.sort_by_key(tier_test_object_marker);
            if marker.is_some() || version_marker.is_some() {
                let marker = (marker.unwrap_or_default(), version_marker.unwrap_or_default());
                objects.retain(|object| tier_test_object_marker(object) > marker);
            }
-            let limit = match usize::try_from(max_keys) {
-                Ok(limit) => limit,
-                Err(_) => 0,
-            };
+            let limit: usize = usize::try_from(max_keys).unwrap_or_default();
            let is_truncated = objects.len() > limit;
            if is_truncated {
                objects.truncate(limit);
@@ -15171,17 +15294,16 @@ mod tests {
            result: Self::WalkResultSender,
            opts: Self::WalkOptions,
        ) -> Result<()> {
-            if self.fail_reference_walk.load(Ordering::SeqCst) {
-                if result
+            if self.fail_reference_walk.load(Ordering::SeqCst)
+                && result
                    .send(StorageObjectInfoOrErr {
                        item: None,
                        err: Some(Error::other("injected tier reference walk failure")),
                    })
                    .await
                    .is_err()
-                {
-                    return Ok(());
-                }
+            {
+                return Ok(());
            }
            let mut objects = self
                .listed_versions
@@ -15192,7 +15314,7 @@ mod tests {
                .filter(|object| opts.include_free_versions || !object.transitioned_object.free_version)
                .cloned()
                .collect::<Vec<_>>();
-            objects.sort_by(|left, right| tier_test_object_marker(left).cmp(&tier_test_object_marker(right)));
+            objects.sort_by_key(tier_test_object_marker);
            if let Some(marker) = opts.marker.as_deref() {
                objects.retain(|object| object.name.as_str() > marker);
            }
@@ -15370,17 +15492,18 @@ mod tests {
            api_view.rustfs.expect("admin RustFS payload should exist").secret_key,
            TIER_CREDENTIAL_REDACTED
        );
-        let observed = lock_unpoisoned(&observed);
-        assert_eq!(observed.len(), 1);
-        assert_eq!(
-            observed[0]
-                .rustfs
-                .as_ref()
-                .expect("backend factory should observe the RustFS payload")
-                .secret_key,
-            SECRET_KEY
-        );
-        drop(observed);
+        {
+            let observed = lock_unpoisoned(&observed);
+            assert_eq!(observed.len(), 1);
+            assert_eq!(
+                observed[0]
+                    .rustfs
+                    .as_ref()
+                    .expect("backend factory should observe the RustFS payload")
+                    .secret_key,
+                SECRET_KEY
+            );
+        }

        let operations = backend.op_log().await;
        assert_eq!(operations.len(), 5);
@@ -16581,7 +16704,7 @@ mod tests {
        candidate.tiers.insert("COLD-A".to_string(), build_rustfs_tier("COLD-A"));
        candidate.tiers.insert("COLD-B".to_string(), build_rustfs_tier("COLD-B"));

-        let targets = TierCandidateMutation::Add(build_rustfs_tier("COLD-B"), true)
+        let targets = TierCandidateMutation::Add(Box::new(build_rustfs_tier("COLD-B")), true)
            .affected_targets(&current, &candidate)
            .expect("add proof should ignore unchanged durable tiers");
        assert_eq!(targets.len(), 1);
@@ -16607,7 +16730,7 @@ mod tests {
                TierConfigMgr::update_candidate_with_config_lock(
                    &manager,
                    store.clone(),
-                    TierCandidateMutation::Add(build_rustfs_tier("COLD-B"), true),
+                    TierCandidateMutation::Add(Box::new(build_rustfs_tier("COLD-B")), true),
                ),
            )
            .await
@@ -16666,14 +16789,15 @@ mod tests {
            .await
            .expect("legacy nested-name Add must run the full coordinator fanout");

-        let prepared_intents = lock_unpoisoned(&prepared_intents);
-        assert_eq!(prepared_intents.len(), 1);
-        assert_eq!(prepared_intents[0].kind, TierMutationIntentKind::Add);
-        assert_eq!(prepared_intents[0].affected_targets.len(), 1);
-        assert_eq!(prepared_intents[0].affected_targets[0].tier_name, "COLD-LEGACY");
-        assert!(prepared_intents[0].affected_targets[0].old_backend_identity.is_none());
-        assert!(prepared_intents[0].affected_targets[0].new_backend_identity.is_some());
-        drop(prepared_intents);
+        {
+            let prepared_intents = lock_unpoisoned(&prepared_intents);
+            assert_eq!(prepared_intents.len(), 1);
+            assert_eq!(prepared_intents[0].kind, TierMutationIntentKind::Add);
+            assert_eq!(prepared_intents[0].affected_targets.len(), 1);
+            assert_eq!(prepared_intents[0].affected_targets[0].tier_name, "COLD-LEGACY");
+            assert!(prepared_intents[0].affected_targets[0].old_backend_identity.is_none());
+            assert!(prepared_intents[0].affected_targets[0].new_backend_identity.is_some());
+        }

        let peer_calls = lock_unpoisoned(&peer_calls).clone();
        let prepare_index = peer_calls
@@ -16755,7 +16879,7 @@ mod tests {
        let err = TierConfigMgr::update_candidate_with_config_lock(
            &manager,
            store.clone(),
-            TierCandidateMutation::Add(build_rustfs_tier("COLD-B"), true),
+            TierCandidateMutation::Add(Box::new(build_rustfs_tier("COLD-B")), true),
        )
        .await
        .expect_err("a new tier config update must wait for pending mutation recovery");
@@ -17031,7 +17155,7 @@ mod tests {
                    TierConfigMgr::update_candidate_with_config_lock(
                        &update_manager,
                        update_store,
-                        TierCandidateMutation::Add(build_rustfs_tier("COLD-A"), true),
+                        TierCandidateMutation::Add(Box::new(build_rustfs_tier("COLD-A")), true),
                    ),
                )
                .await
@@ -17082,7 +17206,7 @@ mod tests {
                        TierConfigMgr::update_candidate_with_config_lock(
                            &update_manager,
                            update_store,
-                            TierCandidateMutation::Add(build_rustfs_tier("COLD-A"), true),
+                            TierCandidateMutation::Add(Box::new(build_rustfs_tier("COLD-A")), true),
                        ),
                    ),
                )
@@ -17145,7 +17269,7 @@ mod tests {
                TierConfigMgr::prevalidate_candidate_owned(
                    empty_mgr(),
                    None,
-                    TierCandidateMutation::Add(build_rustfs_tier("COLD-A"), true),
+                    TierCandidateMutation::Add(Box::new(build_rustfs_tier("COLD-A")), true),
                ),
            )
            .await;
@@ -17809,7 +17933,7 @@ mod tests {

    #[tokio::test]
    async fn tier_add_succeeds_with_refresh_during_coordinator_commit() {
-        assert_coordinator_commit_refresh_succeeds(TierCandidateMutation::Add(build_rustfs_tier("COLD-A"), true)).await;
+        assert_coordinator_commit_refresh_succeeds(TierCandidateMutation::Add(Box::new(build_rustfs_tier("COLD-A")), true)).await;
    }

    #[tokio::test]
@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]

 use crate::error::is_err_bucket_not_found;
 #[cfg(feature = "gcs")]
@@ -719,17 +717,7 @@ async fn check_warm_backend_with_deadlines(
    if !matches!(cleanup_result, Ok(Ok(()))) {
        return Err(probe_cleanup_incomplete_error());
    }
-    if let Err(err) = read_result {
-        //if is_err_bucket_not_found(&err) {
-        //    return Err(ERR_TIER_BUCKET_NOT_FOUND);
-        //}
-        /*else if is_err_signature_does_not_match(err) {
-            return Err(ERR_TIER_MISSING_CREDENTIALS);
-        }*/
-        //else {
-        return Err(err);
-        //}
-    }
+    read_result?;
    Ok(())
 }

@@ -759,7 +747,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                    warn!("{}", err);
                    return Err(AdminError {
                        code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                        status_code: StatusCode::BAD_REQUEST,
                    });
                }
@@ -800,7 +788,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                    warn!("{}", err);
                    return Err(AdminError {
                        code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                        status_code: StatusCode::BAD_REQUEST,
                    });
                }
@@ -820,7 +808,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                    warn!("{}", err);
                    return Err(AdminError {
                        code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                        status_code: StatusCode::BAD_REQUEST,
                    });
                }
@@ -840,7 +828,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                    warn!("{}", err);
                    return Err(AdminError {
                        code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                        status_code: StatusCode::BAD_REQUEST,
                    });
                }
@@ -860,7 +848,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                    warn!("{}", err);
                    return Err(AdminError {
                        code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                        status_code: StatusCode::BAD_REQUEST,
                    });
                }
@@ -880,7 +868,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                    warn!("{}", err);
                    return Err(AdminError {
                        code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                        status_code: StatusCode::BAD_REQUEST,
                    });
                }
@@ -900,7 +888,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                    warn!("{}", err);
                    return Err(AdminError {
                        code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                        status_code: StatusCode::BAD_REQUEST,
                    });
                }
@@ -929,7 +917,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                    warn!("{}", err);
                    return Err(AdminError {
                        code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                        status_code: StatusCode::BAD_REQUEST,
                    });
                }
@@ -949,7 +937,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                    warn!("{}", err);
                    return Err(AdminError {
                        code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                        status_code: StatusCode::BAD_REQUEST,
                    });
                }
@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]

 use std::collections::HashMap;
 use std::sync::Arc;
@@ -106,39 +104,35 @@ impl WarmBackendS3 {
        };
        validate_outbound_url(&u).map_err(|err| std::io::Error::other(format!("tier endpoint is not allowed: {err}")))?;

-        if conf.aws_role_web_identity_token_file == "" && conf.aws_role_arn != ""
-            || conf.aws_role_web_identity_token_file != "" && conf.aws_role_arn == ""
-        {
+        let has_web_identity_token_file = !conf.aws_role_web_identity_token_file.is_empty();
+        let has_role_arn = !conf.aws_role_arn.is_empty();
+        let has_access_key = !conf.access_key.is_empty();
+        let has_secret_key = !conf.secret_key.is_empty();
+
+        if has_web_identity_token_file != has_role_arn {
            return Err(std::io::Error::other("both the token file and the role ARN are required"));
-        } else if conf.access_key == "" && conf.secret_key != "" || conf.access_key != "" && conf.secret_key == "" {
+        } else if has_access_key != has_secret_key {
            return Err(std::io::Error::other("both the access and secret keys are required"));
-        } else if conf.aws_role
-            && (conf.aws_role_web_identity_token_file != ""
-                || conf.aws_role_arn != ""
-                || conf.access_key != ""
-                || conf.secret_key != "")
-        {
+        } else if conf.aws_role && (has_web_identity_token_file || has_role_arn || has_access_key || has_secret_key) {
            return Err(std::io::Error::other(
                "AWS Role cannot be activated with static credentials or the web identity token file",
            ));
-        } else if conf.bucket == "" {
+        } else if conf.bucket.is_empty() {
            return Err(std::io::Error::other("no bucket name was provided"));
        }

-        let creds: Credentials<Static>;
-
-        if conf.access_key != "" && conf.secret_key != "" {
+        let creds = if has_access_key && has_secret_key {
            //creds = Credentials::new_static_v4(conf.access_key, conf.secret_key, "");
-            creds = Credentials::new(Static(Value {
+            Credentials::new(Static(Value {
                access_key_id: conf.access_key.clone(),
                secret_access_key: conf.secret_key.clone(),
                session_token: "".to_string(),
                signer_type: SignatureType::SignatureV4,
                ..Default::default()
-            }));
+            }))
        } else {
            return Err(std::io::Error::other("insufficient parameters for S3 backend authentication"));
-        }
+        };
        let timeouts = transition_client_timeouts_from_env();
        let opts = Options {
            creds,
@@ -162,11 +156,11 @@ impl WarmBackendS3 {
    }

    pub fn get_dest(&self, object: &str) -> String {
-        let mut dest_obj = object.to_string();
-        if self.prefix != "" {
-            dest_obj = format!("{}/{}", &self.prefix, object);
+        if self.prefix.is_empty() {
+            object.to_string()
+        } else {
+            format!("{}/{}", self.prefix, object)
        }
-        return dest_obj;
    }

    pub(crate) async fn remove_with_result(&self, object: &str, rv: &str) -> Result<RemoveObjectResult, std::io::Error> {
@@ -413,6 +407,10 @@ impl TransitionCandidateVersions {
 }

 #[cfg(test)]
+#[allow(
+    clippy::items_after_test_module,
+    reason = "keep parsing tests adjacent to the helpers they cover"
+)]
 mod tests {
    use super::*;
    use rustfs_s3_client::api_s3_datatypes::{ListVersionsResult, Version};
@@ -917,7 +915,7 @@ impl WarmBackend for WarmBackendS3 {
            .list_objects_v2(&self.bucket, &self.prefix, "", "", SLASH_SEPARATOR, 1)
            .await?;

-        Ok(result.common_prefixes.len() > 0 || result.contents.len() > 0)
+        Ok(!result.common_prefixes.is_empty() || !result.contents.is_empty())
    }
 }

@@ -883,6 +883,17 @@ pub(crate) use ops::object::body_cache_plaintext_len;
 pub(crate) use ops::object::cleanup_rejected_transition_upload_durably;
 #[cfg(any(test, feature = "test-util"))]
 pub use ops::object::{PutObjectCommitBarrier, PutObjectCommitPause};
+#[cfg(all(test, feature = "test-util"))]
+pub(crate) use ops::object::{
+    TransitionTransactionKillPoint as SetDiskTransitionTransactionKillPoint,
+    TransitionTransactionKillPointBarrier as SetDiskTransitionTransactionKillPointBarrier,
+};
+#[cfg(all(test, feature = "test-util"))]
+pub(crate) use ops::object::{
+    TransitionTransactionMutationKind as SetDiskTransitionTransactionMutationKind,
+    TransitionTransactionMutationObservation as SetDiskTransitionTransactionMutationObservation,
+    TransitionTransactionMutationProbe as SetDiskTransitionTransactionMutationProbe,
+};
 mod read;
 mod replication;
 pub(crate) mod shard_source;
@@ -3301,13 +3301,18 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
            // (rustfs/backlog#1009): CompleteMultipartUpload keeps its pre-commit
            // `get_object_info` lookup, so the backfill has no consumer here yet.
            Self::assign_rename_data_indexes(&mut parts_metadatas);
-            let mut rename_result = SetDisks::rename_data_owned(
+            // Disk deadlines can expire before physical publication or failure undo drains.
+            let mut rename_result = SetDisks::rename_data_owned_with_fence(
                &commit_disks,
                (RUSTFS_META_MULTIPART_BUCKET, &commit_upload_id_path),
                parts_metadatas,
                (&commit_bucket, &commit_object),
-                write_quorum,
                commit_allows_early_ack,
+                crate::set_disk::core::io_primitives::RenameDataFenceOptions::new(write_quorum, None)
+                    .with_namespace_commit_guard(
+                        (!crate::bucket::utils::is_meta_bucketname(&commit_bucket))
+                            .then(|| commit_set.ctx.begin_namespace_commit()),
+                    ),
            )
            .await;
            if let Ok(rename_commit) = rename_result.as_mut() {
@@ -6758,6 +6763,301 @@ mod tests {
        .await;
    }

+    #[tokio::test]
+    #[serial(capacity_dirty_scope)]
+    async fn complete_multipart_advances_namespace_generation_after_commit() {
+        let (dirs, disks, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "multipart-namespace-commit";
+        let object = "completed-object";
+        let body = vec![0x65; 4096];
+        make_bucket_on_all(&disks, bucket).await;
+        let before = set_disks.ctx.namespace_commit_generation();
+        let (upload_id, parts) =
+            stage_upload_with_create_opts(&set_disks, bucket, object, &body, &ObjectOptions::default()).await;
+        assert_eq!(set_disks.ctx.namespace_commit_generation(), before, "staging is not publication");
+        assert!(!set_disks.ctx.namespace_commits_pending());
+        tokio::time::timeout(
+            Duration::from_secs(10),
+            set_disks
+                .clone()
+                .complete_multipart_upload(bucket, object, &upload_id, parts, &ObjectOptions::default()),
+        )
+        .await
+        .expect("completion must finish")
+        .expect("the four real shards must commit");
+        let mut reader = tokio::time::timeout(
+            Duration::from_secs(5),
+            set_disks.get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default()),
+        )
+        .await
+        .expect("GET after completion must finish")
+        .expect("a successful completion must be immediately readable");
+        let mut actual = Vec::new();
+        tokio::time::timeout(Duration::from_secs(5), reader.stream.read_to_end(&mut actual))
+            .await
+            .expect("the completed body stream must finish")
+            .expect("read completed body");
+        assert_eq!(actual, body);
+        let upload_path = SetDisks::get_upload_id_dir(bucket, object, &upload_id);
+        for dir in &dirs {
+            assert!(!dir.path().join(RUSTFS_META_MULTIPART_BUCKET).join(&upload_path).exists());
+        }
+        assert!(matches!(
+            set_disks.check_upload_id_exists(bucket, object, &upload_id, false).await,
+            Err(StorageError::InvalidUploadID(..))
+        ));
+        assert!(!set_disks.ctx.namespace_commits_pending());
+        assert_eq!(
+            set_disks.ctx.namespace_commit_generation(),
+            before + 2,
+            "one completed MPU must invalidate snapshots at namespace admission and physical retirement"
+        );
+    }
+
+    #[cfg(not(windows))]
+    async fn assert_complete_multipart_physical_namespace_owner(undo: bool) {
+        use crate::disk::os::{self, prepared_publication_test_hooks as hooks};
+        use crate::set_disk::core::io_primitives::rename_fault_injection;
+        use futures::FutureExt;
+
+        temp_env::async_with_vars([(rustfs_config::ENV_DRIVE_MAX_TIMEOUT_DURATION, Some("60"))], async {
+            let (dirs, disks, set_disks) = hermetic_set_disks(4).await;
+            let bucket = "multipart-physical-namespace";
+            let object = if undo { "undo-tail" } else { "publication-tail" };
+            let old_body = vec![0x41; 1024];
+            let new_body = vec![0x62; 4096];
+            make_bucket_on_all(&disks, bucket).await;
+            let old = set_disks
+                .put_object(
+                    bucket,
+                    object,
+                    &mut PutObjReader::from_vec(old_body.clone()),
+                    &ObjectOptions {
+                        write_completion: crate::object_api::WriteCompletion::TailDrained,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("seed a real readable old version");
+            let old_etag = old.etag.expect("the committed old object must have an ETag");
+            let before = set_disks.ctx.namespace_commit_generation();
+            assert!(!set_disks.ctx.namespace_commits_pending());
+            let (upload_id, parts) =
+                stage_upload_with_create_opts(&set_disks, bucket, object, &new_body, &ObjectOptions::default()).await;
+            let new_etag = get_complete_multipart_md5(&parts);
+            let upload_path = SetDisks::get_upload_id_dir(bucket, object, &upload_id);
+            for dir in &dirs {
+                assert!(dir.path().join(RUSTFS_META_MULTIPART_BUCKET).join(&upload_path).exists());
+            }
+            assert_eq!(set_disks.ctx.namespace_commit_generation(), before);
+            let _fault = undo.then(|| rename_fault_injection::fail_rename_on(object, &[2, 3]));
+            let stage = if undo {
+                hooks::Stage::Rename
+            } else {
+                hooks::Stage::PreparedRename
+            };
+            let (entered_tx, mut entered_rx) = tokio::sync::mpsc::unbounded_channel();
+            let mut hooks = Vec::new();
+            let mut releases = Vec::new();
+            let mut mutation_paths = Vec::new();
+            for (index, disk) in disks.iter().enumerate() {
+                let disk::Disk::Local(local) = disk.as_ref() else {
+                    panic!("physical MPU fixture requires local disks");
+                };
+                let destination = local
+                    .get_disk()
+                    .get_object_path_for_io(bucket, object)
+                    .expect("leased IO path");
+                let entered_tx = entered_tx.clone();
+                let (release_tx, release_rx) = std::sync::mpsc::channel::<()>();
+                hooks.push(hooks::install_at(stage, &destination.join(STORAGE_FORMAT_FILE), move || {
+                    let _ = entered_tx.send(index);
+                    // Dropping senders releases every syscall on assertion failure, too.
+                    let _ = release_rx.recv();
+                }));
+                releases.push(release_tx);
+                // Canonical rename serializes the object directory; backup restore
+                // serializes its xl.meta destination. Drain the actual executor key.
+                mutation_paths.push(if undo {
+                    destination.join(STORAGE_FORMAT_FILE)
+                } else {
+                    destination
+                });
+            }
+            drop(entered_tx);
+            let complete_set = set_disks.clone();
+            let complete_upload = upload_id.clone();
+            let mut complete = tokio::spawn(async move {
+                complete_set
+                    .complete_multipart_upload(bucket, object, &complete_upload, parts, &ObjectOptions::default())
+                    .await
+            });
+            let mut complete_joined = false;
+            let mut observed_counts = None;
+            let observations = std::panic::AssertUnwindSafe(async {
+                let expected_publishers = if undo { 2 } else { 4 };
+                let entered = tokio::time::timeout(Duration::from_secs(10), async {
+                    let mut entered = HashSet::new();
+                    while entered.len() < expected_publishers {
+                        tokio::select! {
+                            index = entered_rx.recv() => {
+                                assert!(entered.insert(index.expect("physical publisher must signal entry")));
+                            }
+                            result = &mut complete => {
+                                complete_joined = true;
+                                panic!("completion returned before physical entry: {result:?}");
+                            }
+                        }
+                    }
+                    entered
+                })
+                .await
+                .expect("all expected physical metadata operations must enter");
+                let pending_at_entry = set_disks.ctx.namespace_commits_pending();
+                let generation_at_entry = set_disks.ctx.namespace_commit_generation();
+                for &index in &entered {
+                    let metadata = tokio::time::timeout(
+                        Duration::from_secs(5),
+                        disks[index].read_version("", bucket, object, "", &ReadOptions::default()),
+                    )
+                    .await
+                    .expect("metadata observation must finish while publication is paused")
+                    .expect("metadata before the paused physical action must be readable");
+                    assert_eq!(
+                        metadata.metadata.get("etag"),
+                        Some(if undo { &new_etag } else { &old_etag }),
+                        "undo must follow actual publication; prepared rename must precede publication"
+                    );
+                }
+
+                // The entry signals run inside the real blocking closures, after each
+                // wrapper installed its normal deadline. No quota/external guard disables it.
+                tokio::time::pause();
+                tokio::time::advance(Duration::from_secs(61)).await;
+                tokio::time::resume();
+                let joined = tokio::time::timeout(Duration::from_secs(5), &mut complete).await;
+                complete_joined = joined.is_ok();
+                let result = joined
+                    .expect("ordinary MPU disk/undo deadlines must still return before physical drain")
+                    .expect("completion task must not panic");
+                assert!(result.is_err(), "a timed-out or two-shard commit cannot acknowledge success");
+                let pending_after_timeout = set_disks.ctx.namespace_commits_pending();
+                let generation_after_timeout = set_disks.ctx.namespace_commit_generation();
+                observed_counts = Some((pending_at_entry, generation_at_entry, pending_after_timeout, generation_after_timeout));
+                for &index in &entered {
+                    assert!(
+                        os::acquire_rename_data_mutation_lease(&disks[index].path(), bucket, &mutation_paths[index])
+                            .now_or_never()
+                            .is_none(),
+                        "timed-out physical work must still own its object serialization"
+                    );
+                }
+                for dir in &dirs {
+                    assert!(
+                        dir.path().join(RUSTFS_META_MULTIPART_BUCKET).join(&upload_path).exists(),
+                        "failed completion must not clean the upload staging"
+                    );
+                }
+            })
+            .catch_unwind()
+            .await;
+
+            // Release even after a failed observation, then finish dispatch before
+            // draining every physical key. No per-disk assertion may skip a later drain.
+            drop(releases);
+            drop(hooks);
+            let coordinator_drained = complete_joined
+                || tokio::time::timeout(Duration::from_secs(10), &mut complete).await.is_ok();
+            if !coordinator_drained {
+                complete.abort();
+                let _ = tokio::time::timeout(Duration::from_secs(5), &mut complete).await;
+            }
+            let drains = futures::future::join_all(disks.iter().zip(&mutation_paths).map(|(disk, destination)| async move {
+                tokio::time::timeout(
+                    Duration::from_secs(5),
+                    os::acquire_rename_data_mutation_lease(&disk.path(), bucket, destination),
+                )
+                .await
+                .map(drop)
+            }))
+            .await;
+            let owner_drained = tokio::time::timeout(Duration::from_secs(5), async {
+                while set_disks.ctx.namespace_commits_pending() {
+                    tokio::task::yield_now().await;
+                }
+            })
+            .await;
+            let physical_drained = drains.iter().all(|drain| drain.is_ok());
+            if !coordinator_drained || !physical_drained || owner_drained.is_err() {
+                // A bounded cleanup failure cannot justify deleting roots that a
+                // detached executor might still use. Keep them for diagnosis.
+                let retained = dirs.into_iter().map(TempDir::keep).collect::<Vec<_>>();
+                eprintln!("MPU cleanup incomplete: coordinator={coordinator_drained}, physical={physical_drained}, retained={retained:?}");
+                if let Err(panic) = observations {
+                    std::panic::resume_unwind(panic);
+                }
+                panic!("MPU cleanup did not drain: coordinator={coordinator_drained}, physical={physical_drained}, retained={retained:?}");
+            }
+            if let Err(panic) = observations {
+                std::panic::resume_unwind(panic);
+            }
+            // Preserve the original drain checks after collecting every result.
+            for drained in drains {
+                drained.expect("released physical MPU work must drain");
+            }
+            owner_drained.expect("physical retirement must finish its namespace counter decrement");
+            for disk in &disks {
+                let metadata = disk
+                    .read_version("", bucket, object, "", &ReadOptions::default())
+                    .await
+                    .expect("all disks must expose the expected final metadata");
+                assert_eq!(metadata.metadata.get("etag"), Some(if undo { &old_etag } else { &new_etag }));
+            }
+            let (pending_at_entry, generation_at_entry, pending_after_timeout, generation_after_timeout) =
+                observed_counts.expect("successful observations must record namespace counters");
+            let mut reader = tokio::time::timeout(
+                Duration::from_secs(5),
+                set_disks.get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default()),
+            )
+            .await
+            .expect("GET after physical drain must finish")
+            .expect("the real final object must be readable");
+            let mut actual = Vec::new();
+            tokio::time::timeout(Duration::from_secs(5), reader.stream.read_to_end(&mut actual))
+                .await
+                .expect("the final object stream must finish")
+                .expect("read final object bytes");
+            assert_eq!(actual, if undo { old_body } else { new_body });
+            assert!(
+                pending_at_entry && pending_after_timeout,
+                "physical MPU work outlived namespace accounting: undo={undo}"
+            );
+            assert_eq!(generation_at_entry, before + 1);
+            assert_eq!(
+                generation_after_timeout, generation_at_entry,
+                "the blocked physical owner cannot retire early"
+            );
+            assert_eq!(set_disks.ctx.namespace_commit_generation(), before + 2);
+            assert!(!set_disks.ctx.namespace_commits_pending());
+
+        })
+        .await;
+    }
+
+    #[cfg(not(windows))]
+    #[tokio::test]
+    #[serial(capacity_dirty_scope)]
+    async fn complete_multipart_timeout_keeps_namespace_owner_until_physical_publication() {
+        assert_complete_multipart_physical_namespace_owner(false).await;
+    }
+
+    #[cfg(not(windows))]
+    #[tokio::test]
+    #[serial(capacity_dirty_scope)]
+    async fn complete_multipart_failed_quorum_keeps_namespace_owner_until_physical_undo() {
+        assert_complete_multipart_physical_namespace_owner(true).await;
+    }
+
    #[tokio::test(flavor = "multi_thread")]
    #[serial]
    async fn complete_multipart_releases_disk_snapshot_before_cleanup() {
@@ -630,27 +630,14 @@ impl ECStore {
            .pools
            .first()
            .is_some_and(|pool| pool_first_endpoint_is_local(&pool.endpoints));
-        #[cfg(feature = "e2e-test-hooks")]
-        let startup_attempt = uuid::Uuid::new_v4();
        let (meta, pool_meta_replica_state) = {
            let mut write_state = self.pool_meta_save_gate.lock().await;
            establish_pool_meta_bootstrap_identity_if_proven(self.pools.clone(), &mut write_state, should_persist_pool_meta)
                .await
                .map_err(|err| Error::other(format!("store init failed during establish_pool_meta_bootstrap_identity: {err}")))?;
-            let load = load_pool_meta_for_startup(self.pools.clone(), &mut write_state);
-            #[cfg(feature = "e2e-test-hooks")]
-            let load = crate::core::pools::startup_cas_test_scope(startup_attempt, "load", &self.pools, load);
-            load.await?
+            load_pool_meta_for_startup(self.pools.clone(), &mut write_state).await?
        };
        let update = meta.validate(self.pools.clone())?;
-        #[cfg(feature = "e2e-test-hooks")]
-        crate::core::pools::startup_cas_test_observe(serde_json::json!({
-            "kind": "startup-classifier", "attempt": startup_attempt,
-            "elected_writer": should_persist_pool_meta,
-            "needs_repair": pool_meta_replica_state.needs_repair,
-            "repair_write_safe": pool_meta_replica_state.repair_write_safe,
-            "topology_update": update,
-        }));
        let endpoints = runtime_sources::endpoint_pools_or_default();

        let mut installed_pool_meta = if update {
@@ -662,17 +649,15 @@ impl ECStore {
        // distributed startup can race on the same lock and replay the prior init bug.
        {
            let mut write_state = self.pool_meta_save_gate.lock().await;
-            let persist = persist_pool_meta_for_startup_if_safe(
+            installed_pool_meta = persist_pool_meta_for_startup_if_safe(
                &installed_pool_meta,
                self.pools.clone(),
                pool_meta_replica_state,
                &mut write_state,
                update,
                should_persist_pool_meta,
-            );
-            #[cfg(feature = "e2e-test-hooks")]
-            let persist = crate::core::pools::startup_cas_test_scope(startup_attempt, "persist", &self.pools, persist);
-            installed_pool_meta = persist.await?;
+            )
+            .await?;
        }

        {
@@ -843,7 +828,17 @@ mod tests {
            recovery_control::{
                IlmRecoveryClassification, IlmRecoveryControl, IlmRecoveryControlIdentity, IlmRecoveryErrorCode,
                IlmRecoveryProtocol, MAX_RECOVERY_ATTEMPTS, list_recovery_controls, load_recovery_control,
-                observe_recovery_source, save_recovery_control_if_absent,
+                observe_recovery_source, recovery_control_record_object_name, save_recovery_control_if_absent,
+            },
+            recovery_disposition::{
+                IlmRecoveryDispositionExecutionOutcome, IlmRecoveryDispositionState, RecoveryDispositionCrashStage,
+                dry_run_recovery_disposition, execute_recovery_disposition, inject_recovery_disposition_crash_once,
+                load_recovery_disposition,
+            },
+            recovery_disposition_runtime::garbage_collect_completed_recovery_disposition,
+            recovery_export::{
+                create_recovery_export, inspect_recovery_export_observation, load_recovery_export,
+                recovery_export_record_object_name,
            },
            tier_delete_journal::{
                DecommissionCheckpointTargetFailureHook, TIER_DELETE_DISPATCH_MANIFEST_PREFIX, TIER_DELETE_JOURNAL_PREFIX,
@@ -867,9 +862,10 @@ mod tests {
                TransitionOperatorProbe, TransitionRecoveryClaimBarrier, TransitionRecoveryTerminalBarrier,
                TransitionRemoteVersion, TransitionSourceIdentity, TransitionSourceVersionMode, TransitionTransaction,
                TransitionTransactionInit, TransitionTransactionState, delete_transition_candidate_for_operator,
-                finalize_missing_transition_transaction_for_operator, inspect_transition_transaction_for_operator,
-                load_transition_transaction_record, recover_transition_transaction_records,
-                recover_transition_transaction_records_at, save_transition_transaction_record,
+                finalize_missing_transition_transaction_for_operator, inspect_transition_recovery_retry_for_operator,
+                inspect_transition_transaction_for_operator, load_transition_transaction_record,
+                recover_transition_transaction_records, recover_transition_transaction_records_at,
+                retry_transition_recovery_for_operator, save_transition_transaction_record,
                save_transition_transaction_record_if_current, transition_recovery_control_id,
                transition_transaction_record_object_name,
            },
@@ -881,7 +877,11 @@ mod tests {
        data_movement::SourceCleanupDeleteBarrier,
        disk::{BUCKET_META_PREFIX, RUSTFS_META_BUCKET, STORAGE_FORMAT_FILE},
        runtime::{global::set_object_store_resolver, sources as runtime_sources},
-        services::notification_sys::acquire_tier_delete_journal_fleet_proof,
+        services::notification_sys::{
+            acquire_tier_delete_journal_fleet_proof, acquire_transition_transaction_compaction_fleet_proof,
+            install_transition_transaction_compaction_fleet_proof_for_test,
+            transition_transaction_compaction_fleet_proof_matches,
+        },
        services::tier::{
            test_util::{MockWarmBackend, MockWarmOp, TransitionCleanupStoreBarrier, register_mock_tier},
            tier::{
@@ -903,7 +903,14 @@ mod tests {
            },
            warm_backend::{TransitionCandidateProbe, WarmBackend},
        },
-        set_disk::SetDiskTransitionUploadedCommitBarrier as TransitionUploadedCommitBarrier,
+        set_disk::{
+            SetDiskTransitionTransactionKillPoint as TransitionTransactionKillPoint,
+            SetDiskTransitionTransactionKillPointBarrier as TransitionTransactionKillPointBarrier,
+            SetDiskTransitionTransactionMutationKind as TransitionTransactionMutationKind,
+            SetDiskTransitionTransactionMutationObservation as TransitionTransactionMutationObservation,
+            SetDiskTransitionTransactionMutationProbe as TransitionTransactionMutationProbe,
+            SetDiskTransitionUploadedCommitBarrier as TransitionUploadedCommitBarrier,
+        },
        storage_api_contracts::list::ListOperations as _,
    };
    #[cfg(feature = "test-util")]
@@ -4384,7 +4391,7 @@ mod tests {
                    }
                    retry_source_info.parts = Arc::new(retry_source_parts);
                    assert_eq!(retry_source_info.etag.as_deref(), Some(retry_object_etag.as_str()));
-                    assert!(!retry_source_info.is_multipart());
+                    assert!(retry_source_info.is_multipart());
                    assert!(retry_source_info.parts.iter().all(|part| part.checksums.is_some()));
                    assert_eq!(retry_source_info.checksum.as_deref(), Some(retry_object_checksum_bytes.as_ref()));
                    assert!(
@@ -11590,6 +11597,36 @@ mod tests {
            .len()
    }

+    #[cfg(feature = "test-util")]
+    async fn only_transition_transaction(store: Arc<crate::store::ECStore>) -> TransitionTransaction {
+        let records = store
+            .clone()
+            .list_objects_v2(
+                RUSTFS_META_BUCKET,
+                TRANSITION_TRANSACTION_RECORD_PREFIX,
+                None,
+                None,
+                100,
+                false,
+                None,
+                false,
+            )
+            .await
+            .expect("transition transaction records should be listable")
+            .objects;
+        assert_eq!(records.len(), 1, "test fixture should have exactly one transition transaction");
+        let transaction_id = records[0]
+            .name
+            .rsplit('/')
+            .next()
+            .and_then(|name| name.strip_suffix(".json"))
+            .and_then(|name| uuid::Uuid::parse_str(name).ok())
+            .expect("transition transaction path should end in its UUID");
+        load_transition_transaction_record(store, transaction_id)
+            .await
+            .expect("transition transaction should load")
+    }
+
    #[cfg(feature = "test-util")]
    async fn register_transition_reconcile_test_tier(
        handle: &Arc<tokio::sync::RwLock<TierConfigMgr>>,
@@ -16885,6 +16922,43 @@ mod tests {
            }
        }

+        let creator_sha256 = rustfs_utils::crypto::hex_sha256(b"legacy-export-actor", ToOwned::to_owned);
+        let mut created_exports = Vec::new();
+        for exportable in first_controls
+            .iter()
+            .filter(|control| control.classification == IlmRecoveryClassification::RetainedAmbiguous)
+        {
+            let observation = inspect_recovery_export_observation(store.clone(), &exportable.control_id)
+                .await
+                .expect("fresh legacy recovery observation should be exportable");
+            let created = create_recovery_export(store.clone(), &observation, &creator_sha256)
+                .await
+                .expect("legacy recovery export should be created exactly once");
+            assert!(!created.replayed);
+            let loaded = load_recovery_export(store.clone(), &created.export_id)
+                .await
+                .expect("created legacy recovery export should load");
+            assert_eq!(loaded.encoded, created.encoded, "export readback must preserve the exact committed bytes");
+            let replayed = create_recovery_export(store.clone(), &observation, &creator_sha256)
+                .await
+                .expect("the same observed generation should replay its immutable export");
+            assert!(replayed.replayed);
+            assert_eq!(replayed.encoded, created.encoded);
+            created_exports.push(created);
+        }
+        assert_eq!(created_exports.len(), 2, "both v1 and v2 legacy journals must have an export path");
+
+        let corrupt_export_id = &created_exports[0].export_id;
+        let export_path = recovery_export_record_object_name(IlmRecoveryProtocol::TierDeleteJournal, corrupt_export_id)
+            .expect("export path should build");
+        com::save_config(store.clone(), &export_path, Vec::new())
+            .await
+            .expect("zero-byte corruption fixture should persist");
+        let corrupt_export = load_recovery_export(store.clone(), corrupt_export_id)
+            .await
+            .expect_err("an existing zero-byte export must fail closed");
+        assert!(!matches!(corrupt_export, Error::ConfigNotFound));
+
        com::save_config(
            store.clone(),
            &journal_paths[0],
@@ -16910,6 +16984,301 @@ mod tests {
        );
    }

+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn legacy_recovery_disposition_removes_only_local_journals_and_replays() {
+        Box::pin(legacy_recovery_disposition_removes_only_local_journals_and_replays_case()).await;
+    }
+
+    #[cfg(feature = "test-util")]
+    async fn legacy_recovery_disposition_removes_only_local_journals_and_replays_case() {
+        let temp_dir = tempfile::tempdir().expect("create legacy disposition store dir");
+        let (ctx, store, _shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "legacy-recovery-disposition", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+        let tier_name = "LEGACY-DISPOSITION";
+        let backend = register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let backend_identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
+            .await
+            .expect("legacy disposition tier lease should resolve")
+            .backend_identity();
+        let fixtures = [
+            serde_json::json!({
+                "version": 1,
+                "obj_name": "legacy/disposition-v1",
+                "version_id": "opaque-disposition-v1",
+                "tier_name": tier_name,
+            }),
+            serde_json::json!({
+                "version": 2,
+                "obj_name": "legacy/disposition-v2",
+                "version_id": "opaque-disposition-v2",
+                "tier_name": tier_name,
+                "backend_identity": backend_identity,
+            }),
+        ];
+        let mut journal_paths = Vec::new();
+        for fixture in &fixtures {
+            let data = serde_json::to_vec(fixture).expect("legacy disposition fixture should encode");
+            let entry = crate::bucket::lifecycle::tier_delete_journal::decode_tier_delete_journal_entry(&data)
+                .expect("legacy disposition fixture should decode");
+            let path = tier_delete_journal_object_name(&entry);
+            com::save_config(store.clone(), &path, data)
+                .await
+                .expect("legacy disposition fixture should persist");
+            journal_paths.push(path);
+        }
+
+        let recovered = recover_tier_delete_journal_entries(store.clone(), 100, None)
+            .await
+            .expect("legacy disposition recovery scan should finish");
+        assert_eq!((recovered.scanned, recovered.deleted, recovered.failed), (2, 0, 0));
+        assert_eq!(tier_delete_journal_count(store.clone()).await, 2);
+
+        let mut controls = list_recovery_controls(
+            store.clone(),
+            IlmRecoveryProtocol::TierDeleteJournal,
+            Some(IlmRecoveryClassification::RetainedAmbiguous),
+            100,
+            None,
+        )
+        .await
+        .expect("legacy disposition controls should be listable")
+        .records;
+        controls.sort_by(|left, right| left.control_id.cmp(&right.control_id));
+        assert_eq!(controls.len(), 2, "both legacy schemas must support disposition");
+
+        let actor_sha256 = rustfs_utils::crypto::hex_sha256(b"legacy-disposition-actor", ToOwned::to_owned);
+        let wrong_actor_sha256 = rustfs_utils::crypto::hex_sha256(b"different-disposition-actor", ToOwned::to_owned);
+        let wrong_export_sha256 = "ff".repeat(32);
+        for (index, control) in controls.iter().enumerate() {
+            let observation = inspect_recovery_export_observation(store.clone(), &control.control_id)
+                .await
+                .expect("legacy disposition source should be observable");
+            let export = create_recovery_export(store.clone(), &observation, &actor_sha256)
+                .await
+                .expect("legacy disposition export should persist");
+            let confirmed_at_unix_nanos = i64::try_from(OffsetDateTime::now_utc().unix_timestamp_nanos())
+                .expect("legacy disposition timestamp should fit i64");
+
+            if index == 0 {
+                let wrong_hash = Box::pin(execute_recovery_disposition(
+                    store.clone(),
+                    &observation,
+                    &export.export_id,
+                    &wrong_export_sha256,
+                    &actor_sha256,
+                    confirmed_at_unix_nanos,
+                ))
+                .await
+                .expect_err("a mismatched export checksum must fail before local deletion");
+                assert_eq!(wrong_hash, Error::PreconditionFailed);
+                assert_eq!(tier_delete_journal_count(store.clone()).await, 2);
+            }
+
+            let dry_run = dry_run_recovery_disposition(
+                store.clone(),
+                &observation,
+                &export.export_id,
+                &export.content_sha256,
+                &actor_sha256,
+                confirmed_at_unix_nanos,
+            )
+            .await
+            .expect("legacy disposition dry-run should validate exact local state");
+            assert_eq!(dry_run.source_copy_count, observation.source_generation.copies.len());
+            assert_eq!(
+                tier_delete_journal_count(store.clone()).await,
+                fixtures.len() - index,
+                "dry-run must not delete a legacy journal"
+            );
+            assert!(
+                matches!(
+                    load_recovery_disposition(store.clone(), IlmRecoveryProtocol::TierDeleteJournal, &dry_run.disposition_id,)
+                        .await,
+                    Err(Error::ConfigNotFound)
+                ),
+                "dry-run must not persist a disposition record"
+            );
+
+            if index == 0 {
+                inject_recovery_disposition_crash_once(RecoveryDispositionCrashStage::AfterLocalDelete);
+                Box::pin(execute_recovery_disposition(
+                    store.clone(),
+                    &observation,
+                    &export.export_id,
+                    &export.content_sha256,
+                    &actor_sha256,
+                    confirmed_at_unix_nanos,
+                ))
+                .await
+                .expect_err("the injected crash must stop after local delete commits");
+                assert_eq!(tier_delete_journal_count(store.clone()).await, 1);
+                let interrupted =
+                    load_recovery_disposition(store.clone(), IlmRecoveryProtocol::TierDeleteJournal, &dry_run.disposition_id)
+                        .await
+                        .expect("the applying disposition must survive the post-delete crash");
+                assert_eq!(interrupted.disposition.state, IlmRecoveryDispositionState::Applying);
+                assert!(
+                    interrupted.disposition.confirmed_absent.is_empty(),
+                    "the crash must occur before absence progress is persisted"
+                );
+            } else {
+                inject_recovery_disposition_crash_once(RecoveryDispositionCrashStage::AfterControlAbandon);
+                Box::pin(execute_recovery_disposition(
+                    store.clone(),
+                    &observation,
+                    &export.export_id,
+                    &export.content_sha256,
+                    &actor_sha256,
+                    confirmed_at_unix_nanos,
+                ))
+                .await
+                .expect_err("the injected crash must stop after control abandonment commits");
+                let interrupted =
+                    load_recovery_disposition(store.clone(), IlmRecoveryProtocol::TierDeleteJournal, &dry_run.disposition_id)
+                        .await
+                        .expect("the applying disposition must survive the post-control crash");
+                assert_eq!(interrupted.disposition.state, IlmRecoveryDispositionState::Applying);
+                assert_eq!(
+                    interrupted.disposition.confirmed_absent.len(),
+                    interrupted.disposition.identity.source_generation.copies.len()
+                );
+
+                let abandoned =
+                    load_recovery_control(store.clone(), IlmRecoveryProtocol::TierDeleteJournal, &observation.control_id)
+                        .await
+                        .expect("the abandoned control must survive the injected crash");
+                let exact_abandoned = abandoned.control.encode().expect("the exact abandoned control should encode");
+                let mut wrong_history = abandoned.control;
+                wrong_history.last_error_code = IlmRecoveryErrorCode::CleanupFailed;
+                let control_path = recovery_control_record_object_name(observation.protocol, &observation.control_id)
+                    .expect("control path should remain canonical");
+                com::save_config(
+                    store.clone(),
+                    &control_path,
+                    wrong_history
+                        .encode()
+                        .expect("the alternate valid control history should encode"),
+                )
+                .await
+                .expect("the alternate control history fixture should persist");
+                let wrong_history_err = Box::pin(execute_recovery_disposition(
+                    store.clone(),
+                    &observation,
+                    &export.export_id,
+                    &export.content_sha256,
+                    &actor_sha256,
+                    confirmed_at_unix_nanos + 1,
+                ))
+                .await
+                .expect_err("a different abandoned control history must not bridge to completion");
+                assert_eq!(wrong_history_err, Error::PreconditionFailed);
+                com::save_config(store.clone(), &control_path, exact_abandoned)
+                    .await
+                    .expect("the exact abandoned control fixture should be restored");
+            }
+
+            let replay_confirmed_at_unix_nanos = confirmed_at_unix_nanos + 2;
+            let executed = Box::pin(execute_recovery_disposition(
+                store.clone(),
+                &observation,
+                &export.export_id,
+                &export.content_sha256,
+                &actor_sha256,
+                replay_confirmed_at_unix_nanos,
+            ))
+            .await
+            .expect("a later request must resume and complete the interrupted disposition");
+            assert_eq!(executed.state, IlmRecoveryDispositionState::Completed);
+            assert_eq!(executed.outcome, IlmRecoveryDispositionExecutionOutcome::Completed);
+            assert_eq!(executed.confirmed_absent_copy_count, executed.source_copy_count);
+            assert_eq!(tier_delete_journal_count(store.clone()).await, fixtures.len() - index - 1);
+            assert!(matches!(
+                com::read_config(store.clone(), &observation.canonical_source_path).await,
+                Err(Error::ConfigNotFound)
+            ));
+            if index == 0 {
+                let untouched = journal_paths
+                    .iter()
+                    .find(|path| *path != &observation.canonical_source_path)
+                    .expect("the other legacy journal should remain");
+                com::read_config(store.clone(), untouched)
+                    .await
+                    .expect("disposition must not remove a different legacy journal");
+            }
+
+            let abandoned = load_recovery_control(store.clone(), IlmRecoveryProtocol::TierDeleteJournal, &observation.control_id)
+                .await
+                .expect("abandoned recovery control should remain inspectable");
+            assert_eq!(abandoned.control.classification, IlmRecoveryClassification::Abandoned);
+            assert_eq!(abandoned.control.revision, observation.control_revision + 1);
+            assert_eq!(abandoned.control.observed_source_generation, observation.source_generation);
+
+            let persisted =
+                load_recovery_disposition(store.clone(), IlmRecoveryProtocol::TierDeleteJournal, &executed.disposition_id)
+                    .await
+                    .expect("completed disposition should remain durable");
+            assert_eq!(persisted.disposition.state, IlmRecoveryDispositionState::Completed);
+
+            let replayed = Box::pin(execute_recovery_disposition(
+                store.clone(),
+                &observation,
+                &export.export_id,
+                &export.content_sha256,
+                &actor_sha256,
+                replay_confirmed_at_unix_nanos + 1,
+            ))
+            .await
+            .expect("same actor should replay the completed disposition");
+            assert_eq!(replayed.state, IlmRecoveryDispositionState::Completed);
+            assert_eq!(replayed.outcome, IlmRecoveryDispositionExecutionOutcome::Replayed);
+
+            let wrong_actor = Box::pin(execute_recovery_disposition(
+                store.clone(),
+                &observation,
+                &export.export_id,
+                &export.content_sha256,
+                &wrong_actor_sha256,
+                replay_confirmed_at_unix_nanos + 2,
+            ))
+            .await
+            .expect_err("a different actor must not replay a completed disposition");
+            assert_eq!(wrong_actor, Error::PreconditionFailed);
+
+            assert!(
+                !Box::pin(garbage_collect_completed_recovery_disposition(
+                    store.clone(),
+                    &persisted,
+                    persisted.disposition.retain_until_unix_nanos - 1,
+                ))
+                .await
+                .expect("completed disposition should remain before retention expires")
+            );
+            assert!(
+                Box::pin(garbage_collect_completed_recovery_disposition(
+                    store.clone(),
+                    &persisted,
+                    persisted.disposition.retain_until_unix_nanos,
+                ))
+                .await
+                .expect("expired completed disposition should be garbage collected")
+            );
+            assert!(matches!(
+                load_recovery_disposition(store.clone(), IlmRecoveryProtocol::TierDeleteJournal, &executed.disposition_id).await,
+                Err(Error::ConfigNotFound)
+            ));
+            assert_eq!(backend.remove_count().await, 0, "legacy disposition must not call the remote tier");
+            assert_eq!(backend.exact_remove_count(), 0, "legacy disposition must not issue exact remote DELETE");
+            assert!(
+                backend.op_log().await.is_empty(),
+                "legacy disposition must not invoke any backend operation"
+            );
+        }
+    }
+
    #[cfg(feature = "test-util")]
    #[tokio::test]
    #[serial_test::serial(storage_class_env)]
@@ -19257,6 +19626,330 @@ mod tests {
        assert_eq!(loaded_ids, intent_ids);
    }

+    #[cfg(feature = "test-util")]
+    fn transition_mutation_measurement(observations: &[TransitionTransactionMutationObservation]) -> (usize, usize, u128) {
+        (
+            observations.len(),
+            observations.iter().map(|observation| observation.encoded_bytes).sum(),
+            observations.iter().map(|observation| observation.elapsed.as_micros()).sum(),
+        )
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn compact_transition_transactions_halve_success_path_quorum_mutations() {
+        let temp_dir = tempfile::tempdir().expect("create transition mutation measurement dir");
+        let (ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "transition-mutation-measurement", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let tier_name = "MUTATION-MEASURE";
+        register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let bucket = "transition-mutation-measurement";
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("measurement bucket should be created");
+
+        let run_case = |profile: &'static str, size: usize| {
+            let store = store.clone();
+            async move {
+                let object = format!("{profile}-{size}.bin");
+                let mut reader = PutObjReader::from_vec(vec![b'm'; size]);
+                let original = store
+                    .put_object(bucket, &object, &mut reader, &ObjectOptions::default())
+                    .await
+                    .expect("measurement source should be written");
+                let probe = TransitionTransactionMutationProbe::install(bucket, &object);
+                store
+                    .transition_object(
+                        bucket,
+                        &object,
+                        &ObjectOptions {
+                            transition: TransitionOptions {
+                                status: TRANSITION_PENDING.to_string(),
+                                tier: tier_name.to_string(),
+                                etag: original.etag.clone().expect("measurement source should have an ETag"),
+                                ..Default::default()
+                            },
+                            version_id: original.version_id.map(|version| version.to_string()),
+                            mod_time: original.mod_time,
+                            ..Default::default()
+                        },
+                    )
+                    .await
+                    .expect("measurement transition should commit");
+                probe.observations()
+            }
+        };
+
+        let sizes = [4 * 1024, 1024 * 1024];
+        let mut legacy = Vec::with_capacity(sizes.len());
+        for size in sizes {
+            legacy.push((size, run_case("legacy", size).await));
+        }
+
+        let compaction_proof = install_transition_transaction_compaction_fleet_proof_for_test("object-transaction-fencing-test");
+        for ((size, legacy_observations), compact_size) in legacy.into_iter().zip(sizes) {
+            assert_eq!(size, compact_size);
+            let compact_observations = run_case("compact", size).await;
+            let legacy_measurement = transition_mutation_measurement(&legacy_observations);
+            let compact_measurement = transition_mutation_measurement(&compact_observations);
+            assert_eq!(legacy_measurement.0, 6, "legacy success should use five saves and one delete");
+            assert_eq!(compact_measurement.0, 3, "compact success should use two saves and one delete");
+            assert!(
+                compact_measurement.1 < legacy_measurement.1,
+                "compact transaction bodies should write fewer aggregate bytes"
+            );
+            assert_eq!(
+                legacy_observations
+                    .iter()
+                    .map(|observation| (observation.kind, observation.previous_state, observation.state))
+                    .collect::<Vec<_>>(),
+                vec![
+                    (TransitionTransactionMutationKind::Create, None, TransitionTransactionState::UploadStarted),
+                    (
+                        TransitionTransactionMutationKind::CompareAndSave,
+                        Some(TransitionTransactionState::UploadStarted),
+                        TransitionTransactionState::UploadOutcomeUnknown,
+                    ),
+                    (
+                        TransitionTransactionMutationKind::CompareAndSave,
+                        Some(TransitionTransactionState::UploadOutcomeUnknown),
+                        TransitionTransactionState::Uploaded,
+                    ),
+                    (
+                        TransitionTransactionMutationKind::CompareAndSave,
+                        Some(TransitionTransactionState::Uploaded),
+                        TransitionTransactionState::LocalCommitStarted,
+                    ),
+                    (
+                        TransitionTransactionMutationKind::CompareAndSave,
+                        Some(TransitionTransactionState::LocalCommitStarted),
+                        TransitionTransactionState::Committed,
+                    ),
+                    (
+                        TransitionTransactionMutationKind::Delete,
+                        Some(TransitionTransactionState::Committed),
+                        TransitionTransactionState::Committed,
+                    ),
+                ]
+            );
+            assert_eq!(
+                compact_observations
+                    .iter()
+                    .map(|observation| (observation.kind, observation.previous_state, observation.state))
+                    .collect::<Vec<_>>(),
+                vec![
+                    (
+                        TransitionTransactionMutationKind::Create,
+                        None,
+                        TransitionTransactionState::UploadOutcomeUnknown,
+                    ),
+                    (
+                        TransitionTransactionMutationKind::CompareAndSave,
+                        Some(TransitionTransactionState::UploadOutcomeUnknown),
+                        TransitionTransactionState::LocalCommitStarted,
+                    ),
+                    (
+                        TransitionTransactionMutationKind::Delete,
+                        Some(TransitionTransactionState::LocalCommitStarted),
+                        TransitionTransactionState::LocalCommitStarted,
+                    ),
+                ]
+            );
+            assert!(legacy_observations.iter().all(|observation| observation.succeeded));
+            assert!(compact_observations.iter().all(|observation| observation.succeeded));
+            println!(
+                "transition_mutation_measurement,profile=legacy,size={size},mutations={},encoded_bytes={},latency_us={},quorum_operations={}",
+                legacy_measurement.0, legacy_measurement.1, legacy_measurement.2, legacy_measurement.0
+            );
+            println!(
+                "transition_mutation_measurement,profile=compact,size={size},mutations={},encoded_bytes={},latency_us={},quorum_operations={}",
+                compact_measurement.0, compact_measurement.1, compact_measurement.2, compact_measurement.0
+            );
+        }
+        assert_eq!(transition_transaction_record_count(store.clone()).await, 0);
+        let admitted = acquire_transition_transaction_compaction_fleet_proof()
+            .expect("published homogeneous proof should admit one compact writer");
+        assert!(transition_transaction_compaction_fleet_proof_matches(&admitted));
+        drop(compaction_proof);
+        assert!(
+            !transition_transaction_compaction_fleet_proof_matches(&admitted),
+            "revocation must fence a writer admitted by the previous process-epoch snapshot"
+        );
+        drop(admitted);
+        assert!(
+            acquire_transition_transaction_compaction_fleet_proof().is_none(),
+            "revoking the homogeneous proof must restore the legacy writer profile"
+        );
+        shutdown.cancel();
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn compact_transition_kill_points_preserve_the_only_remote_owner() {
+        let temp_dir = tempfile::tempdir().expect("create compact transition kill-point dir");
+        let (ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "compact-transition-kill-points", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let tier_name = "COMPACT-KILL";
+        let backend = register_transition_reconcile_test_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let _tier_lease = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
+            .await
+            .expect("mock tier lease should remain available during recovery");
+        let bucket = "compact-transition-kill-points";
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("kill-point bucket should be created");
+        let _compaction_proof = install_transition_transaction_compaction_fleet_proof_for_test("object-transaction-fencing-test");
+
+        for (index, point, expected_state, expected_revision, expected_committed, expected_recovered) in [
+            (
+                0,
+                TransitionTransactionKillPoint::PrePutFence,
+                TransitionTransactionState::UploadOutcomeUnknown,
+                1,
+                false,
+                true,
+            ),
+            (
+                1,
+                TransitionTransactionKillPoint::UploadBeforeCommitFence,
+                TransitionTransactionState::UploadOutcomeUnknown,
+                1,
+                false,
+                true,
+            ),
+            (
+                2,
+                TransitionTransactionKillPoint::LocalCommitBeforeDelete,
+                TransitionTransactionState::LocalCommitStarted,
+                2,
+                true,
+                true,
+            ),
+            (
+                3,
+                TransitionTransactionKillPoint::CommitFenceBeforeLocalCommit,
+                TransitionTransactionState::LocalCommitStarted,
+                2,
+                false,
+                false,
+            ),
+        ] {
+            let object = format!("kill-point-{index}.bin");
+            let payload = vec![b'k' + u8::try_from(index).expect("small case index should fit u8"); 64 * 1024];
+            let mut reader = PutObjReader::from_vec(payload.clone());
+            let source = store
+                .put_object(bucket, &object, &mut reader, &ObjectOptions::default())
+                .await
+                .expect("kill-point source should be written");
+            let remote_before = backend.object_count().await;
+            let barrier = TransitionTransactionKillPointBarrier::install(bucket, &object, point);
+            let transition_store = store.clone();
+            let transition_object = object.clone();
+            let transition = tokio::spawn(async move {
+                transition_store
+                    .transition_object(
+                        bucket,
+                        &transition_object,
+                        &ObjectOptions {
+                            transition: TransitionOptions {
+                                status: TRANSITION_PENDING.to_string(),
+                                tier: tier_name.to_string(),
+                                etag: source.etag.clone().expect("kill-point source should have an ETag"),
+                                ..Default::default()
+                            },
+                            version_id: source.version_id.map(|version| version.to_string()),
+                            mod_time: source.mod_time,
+                            ..Default::default()
+                        },
+                    )
+                    .await
+            });
+            barrier.wait_until_paused().await;
+            transition.abort();
+            assert!(
+                transition
+                    .await
+                    .expect_err("kill-point transition should be cancelled")
+                    .is_cancelled(),
+                "kill-point transition should stop without unwinding"
+            );
+            drop(barrier);
+
+            let transaction = only_transition_transaction(store.clone()).await;
+            assert_eq!((transaction.state, transaction.revision), (expected_state, expected_revision));
+            let paused_source = store
+                .get_object_info(
+                    bucket,
+                    &object,
+                    &ObjectOptions {
+                        metadata_cache_safe: false,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("kill-point source metadata should remain readable");
+            assert_eq!(
+                paused_source.transitioned_object.status == rustfs_filemeta::TRANSITION_COMPLETE,
+                expected_committed
+            );
+            let expected_remote_at_pause = remote_before + usize::from(point != TransitionTransactionKillPoint::PrePutFence);
+            assert_eq!(backend.object_count().await, expected_remote_at_pause);
+
+            let stats = recover_transition_transaction_records_at(
+                store.clone(),
+                100,
+                None,
+                i128::from(transaction.not_after_unix_nanos) + 1,
+            )
+            .await
+            .expect("kill-point transaction recovery should complete");
+            if expected_recovered {
+                assert_eq!((stats.scanned, stats.recovered, stats.retained, stats.failed), (1, 1, 0, 0));
+                assert_eq!(transition_transaction_record_count(store.clone()).await, 0);
+            } else {
+                assert_eq!((stats.scanned, stats.recovered, stats.retained, stats.failed), (1, 0, 1, 0));
+                assert_eq!(
+                    transition_transaction_record_count(store.clone()).await,
+                    1,
+                    "an uncommitted local-commit fence must retain its exact remote owner"
+                );
+            }
+            let expected_remote_after_recovery = if matches!(
+                point,
+                TransitionTransactionKillPoint::LocalCommitBeforeDelete
+                    | TransitionTransactionKillPoint::CommitFenceBeforeLocalCommit
+            ) {
+                remote_before + 1
+            } else {
+                remote_before
+            };
+            assert_eq!(backend.object_count().await, expected_remote_after_recovery);
+            assert_eq!(backend.remove_count().await, usize::from(index >= 1));
+
+            let mut restored = Vec::new();
+            store
+                .get_object_reader(bucket, &object, None, HeaderMap::new(), &ObjectOptions::default())
+                .await
+                .expect("kill-point source should remain readable through its authoritative location")
+                .stream
+                .read_to_end(&mut restored)
+                .await
+                .expect("kill-point source body should drain");
+            assert_eq!(restored, payload);
+
+            if !expected_recovered {
+                break;
+            }
+        }
+        shutdown.cancel();
+    }
+
    #[cfg(feature = "test-util")]
    #[tokio::test]
    #[serial_test::serial(storage_class_env)]
@@ -20302,7 +20995,7 @@ mod tests {
            bucket: bucket.to_string(),
            object: object.to_string(),
            version_id: None,
-            data_dir: uuid::Uuid::new_v4(),
+            data_dir: original.data_dir.expect("source object should have data_dir"),
            mod_time_unix_nanos: original
                .mod_time
                .expect("source object should have mod_time")
@@ -20436,7 +21129,7 @@ mod tests {
                bucket: bucket.to_string(),
                object: object.to_string(),
                version_id: None,
-                data_dir: uuid::Uuid::new_v4(),
+                data_dir: original.data_dir.expect("source object should have data_dir"),
                mod_time_unix_nanos: original
                    .mod_time
                    .expect("source object should have mod_time")
@@ -20577,10 +21270,77 @@ mod tests {
            IlmRecoveryClassification::RetainedAmbiguous
        );
        let local_commit_control =
-            load_recovery_control(store, IlmRecoveryProtocol::TransitionTransaction, &local_commit_control_id)
+            load_recovery_control(store.clone(), IlmRecoveryProtocol::TransitionTransaction, &local_commit_control_id)
                .await
                .expect("local-commit control should persist");
        assert_eq!(local_commit_control.control.classification, IlmRecoveryClassification::OperatorRequired);
+
+        let upload_status = inspect_transition_recovery_retry_for_operator(store.clone(), &upload_started_control_id)
+            .await
+            .expect("retained upload should be inspectable for a bounded retry");
+        let local_status = inspect_transition_recovery_retry_for_operator(store.clone(), &local_commit_control_id)
+            .await
+            .expect("operator-required local commit should be inspectable for a bounded retry");
+        assert!(upload_status.retry_ready);
+        assert!(local_status.retry_ready);
+        assert!(matches!(
+            retry_transition_recovery_for_operator(
+                store.clone(),
+                &upload_started_control_id,
+                upload_status.control_revision + 1,
+                &upload_status.source_generation_sha256,
+            )
+            .await,
+            Err(TransitionOperatorError::StaleRecoveryControl)
+        ));
+
+        let put_count_before_retry = backend.put_count().await;
+        let get_count_before_retry = backend.get_count().await;
+        let remove_count_before_retry = backend.remove_count().await;
+        let upload_retry = retry_transition_recovery_for_operator(
+            store.clone(),
+            &upload_started_control_id,
+            upload_status.control_revision,
+            &upload_status.source_generation_sha256,
+        )
+        .await
+        .expect("exact retained upload generation should be rearmed");
+        let local_retry = retry_transition_recovery_for_operator(
+            store.clone(),
+            &local_commit_control_id,
+            local_status.control_revision,
+            &local_status.source_generation_sha256,
+        )
+        .await
+        .expect("exact operator-required local commit generation should be rearmed");
+        assert_eq!(upload_retry.classification, IlmRecoveryClassification::Retrying);
+        assert_eq!(local_retry.classification, IlmRecoveryClassification::Retrying);
+        assert_eq!(upload_retry.attempt_count, upload_status.attempt_count);
+        assert_eq!(local_retry.attempt_count, local_status.attempt_count);
+        assert_eq!(backend.put_count().await, put_count_before_retry);
+        assert_eq!(backend.get_count().await, get_count_before_retry);
+        assert_eq!(backend.remove_count().await, remove_count_before_retry);
+        assert_eq!(backend.exact_remove_count(), 0, "operator retry must not directly issue remote DELETE");
+
+        let retried = recover_transition_transaction_records(store.clone(), 100, None)
+            .await
+            .expect("rearmed records should be re-evaluated through normal recovery");
+        assert_eq!((retried.scanned, retried.recovered, retried.retained, retried.failed), (2, 0, 2, 0));
+        let upload_retained =
+            load_recovery_control(store.clone(), IlmRecoveryProtocol::TransitionTransaction, &upload_started_control_id)
+                .await
+                .expect("upload retry result should persist");
+        let local_retained = load_recovery_control(store, IlmRecoveryProtocol::TransitionTransaction, &local_commit_control_id)
+            .await
+            .expect("local commit retry result should persist");
+        assert_eq!(upload_retained.control.classification, IlmRecoveryClassification::RetainedAmbiguous);
+        assert_eq!(local_retained.control.classification, IlmRecoveryClassification::OperatorRequired);
+        assert_eq!(upload_retained.control.attempt_count, upload_status.attempt_count + 1);
+        assert_eq!(local_retained.control.attempt_count, local_status.attempt_count + 1);
+        assert_eq!(backend.put_count().await, put_count_before_retry);
+        assert_eq!(backend.get_count().await, get_count_before_retry);
+        assert_eq!(backend.remove_count().await, remove_count_before_retry);
+        assert_eq!(backend.exact_remove_count(), 0);
    }

    #[cfg(feature = "test-util")]
@@ -20841,6 +21601,7 @@ mod tests {

        let tier_name = "TXRESPONSELOSS";
        let backend = register_transition_reconcile_test_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let _compaction_proof = install_transition_transaction_compaction_fleet_proof_for_test("object-transaction-fencing-test");
        let bucket = "transition-response-loss-bucket";
        let object = "source.bin";
        store
@@ -20909,6 +21670,7 @@ mod tests {
            TransitionTransactionState::UploadOutcomeUnknown,
            "a response-lost PUT must not remain in UploadStarted"
        );
+        assert_eq!(transaction.revision, 1, "compact response loss must retain the pre-PUT fence generation");
        assert!(
            backend.contains(&transaction.remote_object).await,
            "the test backend must retain the remote candidate"
@@ -442,7 +442,7 @@ pub(crate) mod utils;

 use peer::init_local_peer;
 pub use peer::{
-    BootstrapLocalTarget, all_local_disk, all_local_disk_path, find_local_disk_by_ref, get_disk_infos, init_local_disks,
+    all_local_disk, all_local_disk_path, find_local_disk_by_ref, get_disk_infos, init_local_disks,
    init_local_disks_with_instance_ctx, init_lock_clients, prewarm_local_disk_id_map,
    prewarm_local_disk_id_map_with_instance_ctx,
 };
@@ -1787,7 +1787,7 @@ mod tests {

    // Build a minimal ECStore carrying an explicit instance context. Empty
    // pools/disks are sufficient: the Phase 5 accessors read only `self.ctx`.
-    pub(super) fn build_store_with_ctx(ctx: Arc<InstanceContext>) -> Arc<ECStore> {
+    fn build_store_with_ctx(ctx: Arc<InstanceContext>) -> Arc<ECStore> {
        let endpoint_pools = EndpointServerPools::default();
        Arc::new(ECStore {
            id: uuid::Uuid::new_v4(),
@@ -13,10 +13,7 @@
 // limitations under the License.

 use super::*;
-use crate::bucket::utils::has_bad_path_component;
-use crate::disk::error::{DiskError, Result as DiskResult};
-use crate::disk::{DeleteOptions, Disk, RenameDataGuards, RenameDataResp};
-use crate::runtime::instance::{InstanceContext, NamespaceCommitGuard};
+use crate::runtime::instance::InstanceContext;
 use crate::runtime::sources as runtime_sources;
 use tracing::{debug, error};

@@ -25,203 +22,6 @@ const LOG_SUBSYSTEM_DISK_STARTUP: &str = "disk_startup";
 const EVENT_LOCAL_DISK_ID_PREWARM_SKIPPED: &str = "local_disk_id_prewarm_skipped";
 const EVENT_LOCK_CLIENT_INITIALIZATION_FAILED: &str = "lock_client_initialization_failed";

-/// An instance-bound capability for internal writes before ECStore/IAM startup.
-/// Its private context and volume checks cannot be replaced by a caller guard.
-#[derive(Clone)]
-pub struct BootstrapLocalTarget {
-    ctx: Arc<InstanceContext>,
-}
-
-impl BootstrapLocalTarget {
-    pub fn new(ctx: Arc<InstanceContext>) -> Self {
-        Self { ctx }
-    }
-
-    pub fn is_for_store(&self, store: &ECStore) -> bool {
-        Arc::ptr_eq(&self.ctx, &store.ctx)
-    }
-
-    pub async fn rename_local_data(
-        &self,
-        disk_ref: &str,
-        source: (&str, &str),
-        fi: &FileInfo,
-        destination: (&str, &str),
-        scanner_token: Option<Uuid>,
-    ) -> DiskResult<RenameDataResp> {
-        if scanner_token.is_some() {
-            return Err(DiskError::other("bootstrap rename cannot use a scanner publication lease"));
-        }
-        validate_bootstrap_volume(source.0)?;
-        validate_bootstrap_volume(destination.0)?;
-        rename_local_data_with_ctx(&self.ctx, disk_ref, source, fi, destination, RenameDataGuards::default()).await
-    }
-
-    pub async fn undo_local_write(
-        &self,
-        disk_ref: &str,
-        volume: &str,
-        path: &str,
-        fi: FileInfo,
-        opts: DeleteOptions,
-    ) -> DiskResult<()> {
-        validate_bootstrap_volume(volume)?;
-        undo_local_write_with_ctx(&self.ctx, disk_ref, volume, path, fi, opts).await
-    }
-}
-
-fn validate_bootstrap_volume(volume: &str) -> DiskResult<()> {
-    // Prefix membership alone permits aliases such as .rustfs.sys/../bucket.
-    // Validate both raw rename volumes before any disk lookup or admission.
-    if has_bad_path_component(volume) || !is_meta_bucketname(volume) {
-        return Err(DiskError::FileAccessDenied);
-    }
-    Ok(())
-}
-
-impl ECStore {
-    /// Execute on this instance's active local disk through the physical owner.
-    pub async fn rename_local_data(
-        &self,
-        disk_ref: &str,
-        source: (&str, &str),
-        fi: &FileInfo,
-        destination: (&str, &str),
-        scanner_token: Option<Uuid>,
-    ) -> DiskResult<RenameDataResp> {
-        let external_guard: Option<Arc<dyn Send + Sync>> = if let Some(token) = scanner_token {
-            Some(Arc::new(
-                self.acquire_scanner_publication_lease_guard(token)
-                    .await
-                    .map_err(|err| DiskError::other(err.to_string()))?,
-            ))
-        } else {
-            None
-        };
-        rename_local_data_with_ctx(
-            &self.ctx,
-            disk_ref,
-            source,
-            fi,
-            destination,
-            RenameDataGuards {
-                scanner_publication_lease_token: scanner_token,
-                external_guard,
-                namespace_owner: None,
-            },
-        )
-        .await
-    }
-
-    pub async fn undo_local_write(
-        &self,
-        disk_ref: &str,
-        volume: &str,
-        path: &str,
-        fi: FileInfo,
-        opts: DeleteOptions,
-    ) -> DiskResult<()> {
-        undo_local_write_with_ctx(&self.ctx, disk_ref, volume, path, fi, opts).await
-    }
-}
-
-// The optional ID is a cold lookup to cache only after final admission.
-async fn local_disk_candidate(ctx: &Arc<InstanceContext>, disk_ref: &str) -> DiskResult<(DiskStore, Option<Uuid>)> {
-    let map = ctx.local_disk_map();
-    if let Some(disk) = map.read().await.get(disk_ref).and_then(Option::as_ref).cloned() {
-        return Ok((disk, None));
-    }
-    let disk_id = Uuid::parse_str(disk_ref).map_err(|_| DiskError::DiskNotFound)?;
-    let cached_path = ctx.local_disk_id_map().read().await.get(&disk_id).cloned();
-    if let Some(path) = cached_path {
-        let cached_disk = map.read().await.get(&path).and_then(Option::as_ref).cloned();
-        if let Some(disk) = cached_disk
-            && matches!(disk.as_ref(), Disk::Local(_))
-            && disk.get_disk_id().await? == Some(disk_id)
-        {
-            return Ok((disk, None));
-        }
-    }
-    let disks: Vec<_> = map.read().await.values().filter_map(Clone::clone).collect();
-    // Disk identity may perform format I/O. No registry guard spans this await.
-    for disk in disks {
-        if matches!(disk.as_ref(), Disk::Local(_)) && disk.get_disk_id().await.ok().flatten() == Some(disk_id) {
-            return Ok((disk, Some(disk_id)));
-        }
-    }
-    Err(DiskError::DiskNotFound)
-}
-
-async fn admit_local_disk(
-    ctx: &Arc<InstanceContext>,
-    disk: &DiskStore,
-    disk_id: Option<Uuid>,
-    volume: &str,
-) -> DiskResult<Option<Arc<NamespaceCommitGuard>>> {
-    if !matches!(disk.as_ref(), Disk::Local(_)) {
-        return Err(DiskError::DiskNotFound);
-    }
-    let map = ctx.local_disk_map();
-    let active = map.read().await;
-    if !active
-        .get(&disk.endpoint().to_string())
-        .and_then(Option::as_ref)
-        .is_some_and(|current| Arc::ptr_eq(current, disk))
-    {
-        return Err(DiskError::DiskNotFound);
-    }
-    // Preserve registry -> ID-cache lock order; no filesystem I/O under either.
-    if let Some(disk_id) = disk_id {
-        ctx.local_disk_id_map()
-            .write()
-            .await
-            .insert(disk_id, disk.endpoint().to_string());
-    }
-    // Admission linearizes under the registry read: replacement/quarantine
-    // before this point rejects; later changes do not revoke physical I/O.
-    Ok((!is_meta_bucketname(volume)).then(|| ctx.begin_namespace_commit()))
-}
-
-async fn rename_local_data_with_ctx(
-    ctx: &Arc<InstanceContext>,
-    disk_ref: &str,
-    source: (&str, &str),
-    fi: &FileInfo,
-    destination: (&str, &str),
-    mut guards: RenameDataGuards,
-) -> DiskResult<RenameDataResp> {
-    let (disk, disk_id) = local_disk_candidate(ctx, disk_ref).await?;
-    let owner = admit_local_disk(ctx, &disk, disk_id, destination.0).await?;
-    guards.namespace_owner = owner.as_ref().map(|owner| owner.clone() as Arc<dyn Send + Sync>);
-    let result = disk
-        .rename_data_borrowed_with_fence_observed(source.0, source.1, fi, destination.0, destination.1, guards)
-        .await
-        .result;
-    drop(owner);
-    result
-}
-
-async fn undo_local_write_with_ctx(
-    ctx: &Arc<InstanceContext>,
-    disk_ref: &str,
-    volume: &str,
-    path: &str,
-    fi: FileInfo,
-    opts: DeleteOptions,
-) -> DiskResult<()> {
-    if !opts.undo_write {
-        return Err(DiskError::other("target undo requires undo_write"));
-    }
-    let (disk, disk_id) = local_disk_candidate(ctx, disk_ref).await?;
-    let owner = admit_local_disk(ctx, &disk, disk_id, volume).await?;
-    let physical_owner = owner.as_ref().map(|owner| owner.clone() as Arc<dyn Send + Sync>);
-    let result = disk
-        .undo_write_with_namespace_owner(volume, path, fi, opts, physical_owner)
-        .await;
-    drop(owner);
-    result
-}
-
 async fn remember_local_disk_id(disk: &DiskStore) -> Option<Uuid> {
    remember_local_disk_id_with_instance_ctx(&crate::runtime::global::current_ctx(), disk).await
 }
@@ -465,522 +265,6 @@ mod tests {
        }])
    }

-    async fn target_disk(ctx: &Arc<InstanceContext>, root: &std::path::Path, id: Uuid) -> DiskStore {
-        let mut format = crate::layout::format::FormatV3::new(1, 1);
-        format.erasure.this = id;
-        format.erasure.sets[0][0] = id;
-        let meta = root.join(crate::disk::RUSTFS_META_BUCKET);
-        tokio::fs::create_dir_all(&meta).await.expect("create format volume");
-        tokio::fs::write(
-            meta.join(crate::disk::FORMAT_CONFIG_FILE),
-            serde_json::to_vec(&format).expect("encode format"),
-        )
-        .await
-        .expect("write real disk identity");
-        let mut endpoint = Endpoint::try_from(root.to_str().expect("UTF-8 root")).expect("endpoint");
-        endpoint.set_pool_index(0);
-        endpoint.set_set_index(0);
-        endpoint.set_disk_index(0);
-        let disk = new_disk(
-            &endpoint,
-            &DiskOption {
-                cleanup: false,
-                health_check: false,
-            },
-        )
-        .await
-        .expect("open real local disk");
-        assert_eq!(disk.get_disk_id().await.expect("read disk format identity"), Some(id));
-        ctx.local_disk_map()
-            .write()
-            .await
-            .insert(disk.endpoint().to_string(), Some(disk.clone()));
-        disk
-    }
-
-    fn target_file_info(object: &str, version: Uuid, body: &'static [u8]) -> FileInfo {
-        let mut fi = FileInfo::new(object, 1, 0);
-        fi.erasure.index = 1;
-        fi.version_id = Some(version);
-        fi.mod_time = Some(OffsetDateTime::now_utc());
-        fi.size = i64::try_from(body.len()).expect("fixture length");
-        fi.parts = vec![rustfs_filemeta::ObjectPartInfo {
-            number: 1,
-            size: body.len(),
-            actual_size: fi.size,
-            ..Default::default()
-        }];
-        fi.data = Some(bytes::Bytes::from_static(body));
-        fi.set_inline_data();
-        fi
-    }
-
-    async fn seed_target(disk: &DiskStore, volume: &str, object: &str, fi: FileInfo) -> Vec<u8> {
-        let dir = disk.path().join(volume);
-        tokio::fs::create_dir_all(&dir).await.expect("real fixture volume");
-        disk.write_metadata(volume, volume, object, fi.clone())
-            .await
-            .expect("seed real metadata");
-        let read = disk
-            .read_version(
-                volume,
-                volume,
-                object,
-                &fi.version_id.expect("fixture version").to_string(),
-                &crate::disk::ReadOptions {
-                    read_data: true,
-                    ..Default::default()
-                },
-            )
-            .await
-            .expect("read fixture before mutation");
-        assert_eq!(read.data, fi.data, "fixture must contain readable inline bytes");
-        tokio::fs::read(dir.join(object).join(crate::disk::STORAGE_FORMAT_FILE))
-            .await
-            .expect("seeded metadata bytes")
-    }
-
-    #[tokio::test]
-    async fn target_uuid_lookup_binds_real_disk_and_owner_to_one_instance() {
-        for warm in [false, true] {
-            let ctx_a = Arc::new(InstanceContext::new());
-            let ctx_b = Arc::new(InstanceContext::new());
-            let a = tempfile::tempdir().expect("A root");
-            let b = tempfile::tempdir().expect("B root");
-            let id = Uuid::new_v4();
-            let disk_a = target_disk(&ctx_a, a.path(), id).await;
-            let disk_b = target_disk(&ctx_b, b.path(), id).await;
-            if warm {
-                assert!(record_local_disk_id_if_active(&ctx_a, &disk_a, id).await);
-                assert!(record_local_disk_id_if_active(&ctx_b, &disk_b, id).await);
-            }
-            let version = Uuid::new_v4();
-            let fi = target_file_info("destination", version, b"new-A");
-            for disk in [&disk_a, &disk_b] {
-                seed_target(disk, "target-bucket", "staged", fi.clone()).await;
-            }
-            let b_before = seed_target(
-                &disk_b,
-                "target-bucket",
-                "destination",
-                target_file_info("destination", version, b"old-B"),
-            )
-            .await;
-            let store = super::super::tests::build_store_with_ctx(ctx_a.clone());
-            store
-                .rename_local_data(&id.to_string(), ("target-bucket", "staged"), &fi, ("target-bucket", "destination"), None)
-                .await
-                .expect("rename on A");
-            let read = disk_a
-                .read_version(
-                    "target-bucket",
-                    "target-bucket",
-                    "destination",
-                    &version.to_string(),
-                    &crate::disk::ReadOptions {
-                        read_data: true,
-                        ..Default::default()
-                    },
-                )
-                .await
-                .expect("read committed A");
-            assert_eq!(read.data, fi.data, "warm={warm}");
-            assert_eq!(
-                tokio::fs::read(b.path().join("target-bucket/destination/xl.meta"))
-                    .await
-                    .expect("B metadata"),
-                b_before
-            );
-            assert!(b.path().join("target-bucket/staged/xl.meta").exists());
-            assert!(ctx_a.namespace_commit_generation() > 0);
-            assert_eq!(ctx_b.namespace_commit_generation(), 0);
-            assert!(!ctx_a.namespace_commits_pending());
-            assert!(!ctx_b.namespace_commits_pending());
-            assert_eq!(ctx_a.local_disk_id_map().read().await.get(&id), Some(&disk_a.endpoint().to_string()));
-        }
-    }
-
-    #[tokio::test]
-    async fn target_admission_rejects_removed_quarantined_and_replaced_arcs() {
-        let ctx = Arc::new(InstanceContext::new());
-        let root = tempfile::tempdir().expect("root");
-        let disk = target_disk(&ctx, root.path(), Uuid::new_v4()).await;
-        let endpoint = disk.endpoint().to_string();
-        for state in ["removed", "quarantined", "replaced"] {
-            let replacement = new_disk(
-                &disk.endpoint(),
-                &DiskOption {
-                    cleanup: false,
-                    health_check: false,
-                },
-            )
-            .await
-            .expect("separate active Arc");
-            let map = ctx.local_disk_map();
-            let mut entries = map.write().await;
-            match state {
-                "removed" => {
-                    entries.remove(&endpoint);
-                }
-                "quarantined" => {
-                    entries.insert(endpoint.clone(), None);
-                }
-                _ => {
-                    entries.insert(endpoint.clone(), Some(replacement));
-                }
-            }
-            drop(entries);
-            assert!(
-                matches!(admit_local_disk(&ctx, &disk, None, "target-bucket").await, Err(DiskError::DiskNotFound)),
-                "{state}"
-            );
-            assert!(!ctx.namespace_commits_pending());
-            assert_eq!(ctx.namespace_commit_generation(), 0);
-        }
-    }
-
-    #[tokio::test]
-    async fn target_uuid_cache_cannot_admit_a_different_format_at_the_same_path() {
-        let ctx = Arc::new(InstanceContext::new());
-        let root = tempfile::tempdir().expect("root");
-        let old_id = Uuid::new_v4();
-        let old = target_disk(&ctx, root.path(), old_id).await;
-        assert!(record_local_disk_id_if_active(&ctx, &old, old_id).await);
-        let replacement_id = Uuid::new_v4();
-        let replacement = target_disk(&ctx, root.path(), replacement_id).await;
-        assert!(!Arc::ptr_eq(&old, &replacement));
-        assert!(matches!(
-            local_disk_candidate(&ctx, &old_id.to_string()).await,
-            Err(DiskError::DiskNotFound)
-        ));
-        let (candidate, verified) = local_disk_candidate(&ctx, &replacement_id.to_string())
-            .await
-            .expect("replacement UUID");
-        assert!(Arc::ptr_eq(&candidate, &replacement));
-        assert_eq!(verified, Some(replacement_id));
-        assert!(!ctx.namespace_commits_pending());
-    }
-
-    #[tokio::test]
-    async fn bootstrap_rejects_user_volumes_aliases_and_scanner_tokens_without_mutation() {
-        let ctx = Arc::new(InstanceContext::new());
-        let root = tempfile::tempdir().expect("root");
-        let disk = target_disk(&ctx, root.path(), Uuid::new_v4()).await;
-        let target = BootstrapLocalTarget::new(ctx.clone());
-        let fi = target_file_info("destination", Uuid::new_v4(), b"body");
-        let user_before = seed_target(&disk, "victim", "staged", fi.clone()).await;
-        let meta_before = seed_target(&disk, ".rustfs.sys/tmp", "staged", fi.clone()).await;
-        for invalid in [
-            "victim",
-            ".rustfs.sys/../victim",
-            ".rustfs.sys/./tmp",
-            ".rustfs.sys/ .. /victim",
-            ".rustfs.sys\\..\\victim",
-            ".minio.sys/../victim",
-        ] {
-            for (src, dst) in [(invalid, ".rustfs.sys/tmp"), (".rustfs.sys/tmp", invalid)] {
-                assert!(
-                    target
-                        .rename_local_data(&disk.endpoint().to_string(), (src, "staged"), &fi, (dst, "destination"), None)
-                        .await
-                        .is_err(),
-                    "src={src}, dst={dst}"
-                );
-            }
-            assert!(
-                target
-                    .undo_local_write(
-                        &disk.endpoint().to_string(),
-                        invalid,
-                        "staged",
-                        fi.clone(),
-                        DeleteOptions {
-                            undo_write: true,
-                            ..Default::default()
-                        }
-                    )
-                    .await
-                    .is_err(),
-                "{invalid}"
-            );
-        }
-        assert!(
-            target
-                .rename_local_data(
-                    &disk.endpoint().to_string(),
-                    (".rustfs.sys/tmp", "staged"),
-                    &fi,
-                    (".rustfs.sys/tmp", "destination"),
-                    Some(Uuid::new_v4())
-                )
-                .await
-                .is_err()
-        );
-        assert_eq!(
-            tokio::fs::read(root.path().join("victim/staged/xl.meta"))
-                .await
-                .expect("user source"),
-            user_before
-        );
-        assert_eq!(
-            tokio::fs::read(root.path().join(".rustfs.sys/tmp/staged/xl.meta"))
-                .await
-                .expect("metadata source"),
-            meta_before
-        );
-        assert!(!root.path().join("victim/destination").exists());
-        assert!(!root.path().join(".rustfs.sys/tmp/destination").exists());
-        assert_eq!(ctx.namespace_commit_generation(), 0);
-        assert!(!ctx.namespace_commits_pending());
-    }
-
-    #[tokio::test]
-    async fn bootstrap_allows_internal_multisegment_rename_without_namespace_owner() {
-        for volume in [".rustfs.sys/tmp", ".rustfs.sys/multipart", ".minio.sys/config"] {
-            let ctx = Arc::new(InstanceContext::new());
-            let root = tempfile::tempdir().expect("root");
-            let disk = target_disk(&ctx, root.path(), Uuid::new_v4()).await;
-            let fi = target_file_info("destination", Uuid::new_v4(), b"internal-CAS-body");
-            seed_target(&disk, volume, "staged", fi.clone()).await;
-            BootstrapLocalTarget::new(ctx.clone())
-                .rename_local_data(&disk.endpoint().to_string(), (volume, "staged"), &fi, (volume, "destination"), None)
-                .await
-                .expect("legitimate bootstrap metadata write");
-            let read = disk
-                .read_version(
-                    volume,
-                    volume,
-                    "destination",
-                    &fi.version_id.expect("version").to_string(),
-                    &crate::disk::ReadOptions {
-                        read_data: true,
-                        ..Default::default()
-                    },
-                )
-                .await
-                .expect("read bootstrap result");
-            assert_eq!(read.data, fi.data);
-            assert_eq!(ctx.namespace_commit_generation(), 0);
-            assert!(!ctx.namespace_commits_pending());
-        }
-    }
-
-    #[cfg(not(windows))]
-    #[tokio::test]
-    async fn target_rename_cancellation_retains_real_namespace_and_scanner_owners() {
-        use crate::disk::os::prepared_publication_test_hooks as hooks;
-        let ctx = Arc::new(InstanceContext::new());
-        let sibling = Arc::new(InstanceContext::new());
-        let root = tempfile::tempdir().expect("root");
-        let disk = target_disk(&ctx, root.path(), Uuid::new_v4()).await;
-        let store = super::super::tests::build_store_with_ctx(ctx.clone());
-        let fi = target_file_info("destination", Uuid::new_v4(), b"physically-owned");
-        seed_target(&disk, "target-bucket", "staged", fi.clone()).await;
-        let (token, _) = store
-            .acquire_scanner_publication_lease(0, crate::runtime::instance::SCANNER_PUBLICATION_LEASE_TTL)
-            .await
-            .expect("real scanner token in A");
-        let destination = disk
-            .get_object_path_for_io_if_local("target-bucket", "destination/xl.meta")
-            .expect("local disk")
-            .expect("destination IO path");
-        let (entered_tx, entered_rx) = tokio::sync::oneshot::channel();
-        let (release_tx, release_rx) = std::sync::mpsc::channel::<()>();
-        let _hook = hooks::install(&destination, move || {
-            let _ = entered_tx.send(());
-            let _ = release_rx.recv();
-        });
-        let disk_ref = disk.endpoint().to_string();
-        let mut rename = Box::pin(store.rename_local_data(
-            &disk_ref,
-            ("target-bucket", "staged"),
-            &fi,
-            ("target-bucket", "destination"),
-            Some(token),
-        ));
-        tokio::time::timeout(std::time::Duration::from_secs(10), async {
-            tokio::select! {
-                result = &mut rename => panic!("rename completed before physical pause: {result:?}"),
-                entered = entered_rx => entered.expect("physical rename entered"),
-            }
-        })
-        .await
-        .expect("bounded physical entry");
-        drop(rename);
-        assert!(store.scanner_data_usage_publication_blocked().await);
-        assert!(ctx.namespace_commits_pending());
-        assert!(!sibling.namespace_commits_pending());
-        assert!(
-            store
-                .rename_local_data(&disk_ref, ("target-bucket", "staged"), &fi, ("target-bucket", "another"), Some(token))
-                .await
-                .is_err(),
-            "real pending rename blocks another scanner publication"
-        );
-        assert!(store.release_scanner_publication_lease(token).await, "remove registered token");
-        let gate = ctx.data_movement_operation_gate();
-        assert!(
-            gate.clone().try_write_owned().is_err(),
-            "physical operation still owns the scanner read guard"
-        );
-        drop(release_tx);
-        let _drained = tokio::time::timeout(std::time::Duration::from_secs(10), gate.write_owned())
-            .await
-            .expect("physical tail must release scanner guard");
-        tokio::time::timeout(std::time::Duration::from_secs(10), async {
-            while ctx.namespace_commits_pending() {
-                tokio::task::yield_now().await;
-            }
-        })
-        .await
-        .expect("namespace owner drains");
-        let read = disk
-            .read_version(
-                "target-bucket",
-                "target-bucket",
-                "destination",
-                &fi.version_id.expect("version").to_string(),
-                &crate::disk::ReadOptions {
-                    read_data: true,
-                    ..Default::default()
-                },
-            )
-            .await
-            .expect("read actual late commit");
-        assert_eq!(read.data, fi.data);
-        assert!(ctx.namespace_commit_generation() >= 2);
-        assert_eq!(sibling.namespace_commit_generation(), 0);
-    }
-
-    #[tokio::test]
-    async fn target_ready_rejects_unknown_foreign_released_and_expired_scanner_tokens() {
-        let ctx = Arc::new(InstanceContext::new());
-        let other = Arc::new(InstanceContext::new());
-        let store = super::super::tests::build_store_with_ctx(ctx.clone());
-        let other_store = super::super::tests::build_store_with_ctx(other);
-        let root = tempfile::tempdir().expect("root");
-        let disk = target_disk(&ctx, root.path(), Uuid::new_v4()).await;
-        let fi = target_file_info("destination", Uuid::new_v4(), b"unchanged");
-        let before = seed_target(&disk, "target-bucket", "staged", fi.clone()).await;
-        let ttl = crate::runtime::instance::SCANNER_PUBLICATION_LEASE_TTL;
-        let (foreign, _) = other_store.acquire_scanner_publication_lease(0, ttl).await.expect("B token");
-        let (released, _) = store.acquire_scanner_publication_lease(0, ttl).await.expect("A token");
-        assert!(store.release_scanner_publication_lease(released).await);
-        let (valid, _) = store.acquire_scanner_publication_lease(0, ttl).await.expect("new A token");
-        for token in [Uuid::new_v4(), foreign, released] {
-            assert!(
-                store
-                    .rename_local_data(
-                        &disk.endpoint().to_string(),
-                        ("target-bucket", "staged"),
-                        &fi,
-                        ("target-bucket", "destination"),
-                        Some(token)
-                    )
-                    .await
-                    .is_err()
-            );
-        }
-        tokio::time::pause();
-        tokio::time::advance(ttl + std::time::Duration::from_secs(1)).await;
-        tokio::time::resume();
-        assert!(
-            store
-                .rename_local_data(
-                    &disk.endpoint().to_string(),
-                    ("target-bucket", "staged"),
-                    &fi,
-                    ("target-bucket", "destination"),
-                    Some(valid)
-                )
-                .await
-                .is_err(),
-            "expired real token"
-        );
-        let _ = other_store.release_scanner_publication_lease(foreign).await;
-        assert_eq!(
-            tokio::fs::read(root.path().join("target-bucket/staged/xl.meta"))
-                .await
-                .expect("source bytes"),
-            before
-        );
-        assert!(!root.path().join("target-bucket/destination").exists());
-        assert!(!ctx.namespace_commits_pending());
-    }
-
-    #[cfg(not(windows))]
-    #[tokio::test]
-    #[serial_test::serial]
-    async fn target_ordinary_timeout_keeps_its_physical_namespace_owner() {
-        use crate::disk::os::prepared_publication_test_hooks as hooks;
-        temp_env::async_with_vars([(rustfs_config::ENV_DRIVE_MAX_TIMEOUT_DURATION, Some("1"))], async {
-            let ctx = Arc::new(InstanceContext::new());
-            let store = super::super::tests::build_store_with_ctx(ctx.clone());
-            let root = tempfile::tempdir().expect("root");
-            let disk = target_disk(&ctx, root.path(), Uuid::new_v4()).await;
-            let fi = target_file_info("destination", Uuid::new_v4(), b"timed-out-physical-commit");
-            seed_target(&disk, "target-bucket", "staged", fi.clone()).await;
-            let path = disk
-                .get_object_path_for_io_if_local("target-bucket", "destination/xl.meta")
-                .expect("local")
-                .expect("destination IO path");
-            let (entered_tx, entered_rx) = tokio::sync::oneshot::channel();
-            let (release_tx, release_rx) = std::sync::mpsc::channel::<()>();
-            let _hook = hooks::install(&path, move || {
-                let _ = entered_tx.send(());
-                let _ = release_rx.recv();
-            });
-            let disk_ref = disk.endpoint().to_string();
-            let mut rename = Box::pin(store.rename_local_data(
-                &disk_ref,
-                ("target-bucket", "staged"),
-                &fi,
-                ("target-bucket", "destination"),
-                None,
-            ));
-            tokio::time::timeout(std::time::Duration::from_secs(10), async {
-                tokio::select! {
-                    result = &mut rename => panic!("completed before physical pause: {result:?}"),
-                    entered = entered_rx => entered.expect("physical entry"),
-                }
-            })
-            .await
-            .expect("bounded entry");
-            tokio::time::pause();
-            tokio::time::advance(std::time::Duration::from_secs(2)).await;
-            tokio::time::resume();
-            let result = tokio::time::timeout(std::time::Duration::from_secs(5), &mut rename)
-                .await
-                .expect("ordinary deadline remains enabled");
-            assert!(matches!(result, Err(DiskError::Timeout)), "{result:?}");
-            drop(rename);
-            assert!(ctx.namespace_commits_pending(), "timeout is not a physical drain");
-            drop(release_tx);
-            tokio::time::timeout(std::time::Duration::from_secs(10), async {
-                while ctx.namespace_commits_pending() {
-                    tokio::task::yield_now().await;
-                }
-            })
-            .await
-            .expect("late physical owner drains");
-            let read = disk
-                .read_version(
-                    "target-bucket",
-                    "target-bucket",
-                    "destination",
-                    &fi.version_id.expect("version").to_string(),
-                    &crate::disk::ReadOptions {
-                        read_data: true,
-                        ..Default::default()
-                    },
-                )
-                .await
-                .expect("read actual timeout tail");
-            assert_eq!(read.data, fi.data);
-        })
-        .await;
-    }
-
    #[test]
    fn endpoint_rpc_authority_preserves_port_and_ipv6_brackets() {
        let endpoint = Endpoint::try_from("https://127.0.0.1:9001/d1").expect("URL endpoint");
@@ -692,10 +692,15 @@ impl FileMeta {
                            }
                        }

-                        let old_dir = v.object.as_ref().map(|v| v.data_dir).unwrap_or_default();
+                        // The version stays on disk while the purge replicates
+                        // (status PENDING/FAILED); its data dir must stay with
+                        // it. Returning the dir here made the disk layer delete
+                        // it, which turned every non-inline retained version
+                        // into an unreadable zombie: the purge state could never
+                        // be applied and the bucket could never be deleted.
                        self.set_idx(i, v)?;

-                        return Ok(old_dir);
+                        return Ok(None);
                    }
                    found_index = Some(i);
                }
@@ -2702,6 +2707,58 @@ mod test {
        );
    }

+    /// Regression for rustfs/backlog#2340: a version purge that still awaits
+    /// the replication target keeps the object version on disk with a pending
+    /// purge status. Its data dir must be retained with it; handing the dir
+    /// back here made the disk layer delete it, leaving every non-inline
+    /// retained version unreadable. The dir is released only once the purge
+    /// completes and the version itself goes away.
+    #[test]
+    fn delete_version_pending_version_purge_retains_object_data_dir() {
+        let version_id = Uuid::new_v4();
+        let data_dir = Uuid::new_v4();
+        let mut fm = FileMeta::new();
+        let mut fi = FileInfo::new("object", 2, 2);
+        fi.version_id = Some(version_id);
+        fi.data_dir = Some(data_dir);
+        fi.mod_time = Some(OffsetDateTime::now_utc());
+        fm.add_version(fi).unwrap();
+
+        let pending_purge = FileInfo {
+            name: "object".to_string(),
+            version_id: Some(version_id),
+            mark_deleted: true,
+            replication_state_internal: Some(ReplicationState {
+                version_purge_status_internal: Some("target=PENDING;".to_string()),
+                purge_targets: version_purge_statuses_map("target=PENDING;"),
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let freed = fm.delete_version(&pending_purge).unwrap();
+        assert_eq!(freed, None, "a pending purge must not release the retained version's data dir");
+        assert_eq!(fm.versions.len(), 1, "the version must stay until the purge replicates");
+        let retained = fm
+            .into_fileinfo("vol", "object", &version_id.to_string(), false, false, true)
+            .unwrap();
+        assert_eq!(retained.data_dir, Some(data_dir));
+        assert_eq!(retained.version_purge_status(), VersionPurgeStatusType::Pending);
+
+        let completed_purge = FileInfo {
+            name: "object".to_string(),
+            version_id: Some(version_id),
+            replication_state_internal: Some(ReplicationState {
+                version_purge_status_internal: Some("target=COMPLETE;".to_string()),
+                purge_targets: version_purge_statuses_map("target=COMPLETE;"),
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let freed = fm.delete_version(&completed_purge).unwrap();
+        assert_eq!(freed, Some(data_dir), "a completed purge removes the version and releases its data dir");
+        assert!(fm.versions.is_empty());
+    }
+
    #[test]
    fn delete_version_accepts_delete_only_marker_and_free_version_paths() {
        let marker_version_id = Uuid::new_v4();
@@ -76,6 +76,8 @@ struct HealTaskStatusPayload<'a> {
    min_seq: u64,
    #[serde(skip_serializing_if = "Option::is_none")]
    progress: Option<&'a HealProgress>,
+    #[serde(skip_serializing_if = "Option::is_none")]
+    outcome: Option<&'a super::outcome::HealTaskOutcome>,
 }

 fn u64_is_zero(value: &u64) -> bool {
@@ -87,17 +89,18 @@ fn encode_heal_task_status_payload(
    mut items: Vec<HealResultItem>,
    progress: Option<&HealProgress>,
    mut truncated: bool,
-    next_seq: u64,
-    min_seq: u64,
+    sequence: (u64, u64),
+    outcome: Option<&super::outcome::HealTaskOutcome>,
 ) -> Result<(Vec<u8>, bool)> {
    loop {
        let data = serde_json::to_vec(&HealTaskStatusPayload {
            summary,
            items: &items,
            truncated,
-            next_seq,
-            min_seq,
+            next_seq: sequence.0,
+            min_seq: sequence.1,
            progress,
+            outcome,
        })
        .map_err(|e| Error::Serialization(format!("failed to serialize heal task status: {e}")))?;
        if data.len() <= MAX_HEAL_STATUS_PAYLOAD_SIZE {
@@ -111,25 +114,21 @@ fn encode_heal_task_status_payload(
    }
 }

-fn heal_status_detail(detail: Option<String>, truncated: bool) -> Option<String> {
-    if !truncated {
-        return detail;
-    }
-    let truncation = "heal result items were truncated";
-    Some(detail.map_or_else(|| truncation.to_string(), |detail| format!("{detail}; {truncation}")))
-}
-
 fn encode_heal_status_response(
    summary: &str,
    items: Vec<HealResultItem>,
    progress: Option<&HealProgress>,
    detail: Option<String>,
    truncated: bool,
-    next_seq: u64,
-    min_seq: u64,
+    sequence: (u64, u64),
+    outcome: Option<&super::outcome::HealTaskOutcome>,
 ) -> Result<(Vec<u8>, Option<String>)> {
-    let (data, truncated) = encode_heal_task_status_payload(summary, items, progress, truncated, next_seq, min_seq)?;
-    Ok((data, heal_status_detail(detail, truncated)))
+    let (summary, detail) = match outcome {
+        Some(outcome) => outcome.legacy_status(summary, detail),
+        None => (summary, detail),
+    };
+    let (data, truncated) = encode_heal_task_status_payload(summary, items, progress, truncated, sequence, outcome)?;
+    Ok((data, super::outcome::heal_status_detail(detail, truncated)))
 }

 impl HealChannelProcessor {
@@ -439,6 +438,7 @@ impl HealChannelProcessor {
                .await
        };

+        let outcome = report.as_ref().ok().and_then(|report| report.outcome.clone());
        let (summary, detail, items, truncated, progress, next_seq, min_seq) = match report {
            Ok(HealTaskReport {
                status: HealTaskStatus::Pending | HealTaskStatus::Running,
@@ -576,8 +576,15 @@ impl HealChannelProcessor {
            }
        };

-        let (data, detail) =
-            encode_heal_status_response(&summary, items, progress.as_ref(), detail, truncated, next_seq, min_seq)?;
+        let (data, detail) = encode_heal_status_response(
+            &summary,
+            items,
+            progress.as_ref(),
+            detail,
+            truncated,
+            (next_seq, min_seq),
+            outcome.as_deref(),
+        )?;

        let response = HealChannelResponse {
            request_id: client_token,
@@ -866,7 +873,7 @@ mod tests {
            ..Default::default()
        }];

-        let (data, detail) = encode_heal_status_response("running", items, None, None, false, 0, 0).unwrap();
+        let (data, detail) = encode_heal_status_response("running", items, None, None, false, (0, 0), None).unwrap();

        assert!(data.len() <= MAX_HEAL_STATUS_PAYLOAD_SIZE);
        let payload: serde_json::Value = serde_json::from_slice(&data).unwrap();
@@ -875,6 +882,135 @@ mod tests {
        assert_eq!(detail.as_deref(), Some("heal result items were truncated"));
    }

+    #[test]
+    fn outcome_v3_fixture_matches_canonical_owner_and_preserves_legacy_terminals() {
+        use crate::heal::outcome::*;
+        let cases: serde_json::Value = serde_json::from_str(include_str!("../../../madmin/tests/fixtures/heal-outcome-v3.json"))
+            .expect("shared client fixtures");
+        for case in cases.as_array().expect("fixture cases") {
+            if case.get("remoteResponse").is_some() {
+                continue;
+            }
+            let mut outcome = HealTaskOutcome::default();
+            let name = case["name"].as_str().expect("case name");
+            if matches!(name, "unknown" | "completed_with_errors") {
+                let disposition = if name == "unknown" {
+                    HealObjectDisposition::Unknown
+                } else {
+                    outcome.attempt_failed();
+                    HealObjectDisposition::Failed(HealFailureClass::RetryExhausted)
+                };
+                outcome.record(HealObjectOutcome {
+                    identity: HealObjectIdentity {
+                        kind: HealObjectKind::Object,
+                        bucket: "bucket".into(),
+                        object: "object".into(),
+                        version_id: None,
+                        bucket_incarnation_id: None,
+                        pool_index: None,
+                        set_index: None,
+                    },
+                    disposition,
+                    detail: None,
+                });
+            }
+            let abort = match name {
+                "cancelled" => Some(HealAbortReason::Cancelled),
+                "deadline" => Some(HealAbortReason::Deadline),
+                "untraversable" => Some(HealAbortReason::Untraversable),
+                _ => None,
+            };
+            outcome.finish(abort);
+            let expected = &case["response"];
+            let initial_detail = abort.map(|reason| {
+                match reason {
+                    HealAbortReason::Cancelled => "heal task cancelled",
+                    HealAbortReason::Deadline => "heal task timed out",
+                    HealAbortReason::Untraversable => "heal listing is untraversable",
+                }
+                .to_string()
+            });
+            let (bytes, detail) = encode_heal_status_response(
+                if abort.is_some() { "stopped" } else { "finished" },
+                Vec::new(),
+                None,
+                initial_detail,
+                true,
+                (9, 4),
+                Some(&outcome),
+            )
+            .expect("canonical owner encoding");
+            let decoded: serde_json::Value = serde_json::from_slice(&bytes).expect("wire payload");
+            assert_eq!(decoded["summary"], expected["summary"], "{name}");
+            assert_eq!(detail.unwrap_or_default(), expected["detail"].as_str().expect("detail"), "{name}");
+            assert_eq!(decoded["outcome"], expected["outcome"], "{name}");
+            assert_eq!((decoded["next_seq"].as_u64(), decoded["min_seq"].as_u64()), (Some(9), Some(4)));
+            assert!(decoded["outcome"].get("retainedObjectBytes").is_none());
+            assert!(decoded["outcome"].get("untraversable").is_none());
+        }
+    }
+
+    #[test]
+    fn outcome_v3_abort_cannot_be_hidden_by_a_finished_status() {
+        use crate::heal::outcome::{HealAbortReason, HealTaskOutcome};
+        for reason in [
+            HealAbortReason::Cancelled,
+            HealAbortReason::Deadline,
+            HealAbortReason::Untraversable,
+        ] {
+            let mut outcome = HealTaskOutcome::default();
+            outcome.finish(Some(reason));
+            let (data, detail) = encode_heal_status_response("finished", Vec::new(), None, None, false, (0, 0), Some(&outcome))
+                .expect("canonical abort adapter");
+            let json: serde_json::Value = serde_json::from_slice(&data).expect("public state");
+            assert_eq!(json["summary"], "stopped");
+            assert_eq!(json["outcome"]["execution"]["state"], "aborted");
+            assert!(detail.is_some());
+        }
+    }
+
+    #[test]
+    fn outcome_v3_payload_bound_keeps_cumulative_outcome_and_cursors() {
+        use crate::heal::outcome::*;
+        let mut outcome = HealTaskOutcome::default();
+        outcome.start();
+        for index in 0..256 {
+            outcome.record(HealObjectOutcome {
+                identity: HealObjectIdentity {
+                    kind: HealObjectKind::Object,
+                    bucket: "bucket".into(),
+                    object: format!("object-{index}"),
+                    version_id: None,
+                    bucket_incarnation_id: None,
+                    pool_index: None,
+                    set_index: None,
+                },
+                disposition: HealObjectDisposition::Unknown,
+                detail: Some("\"".repeat(1024)),
+            });
+        }
+        let retained = outcome.objects.len();
+        let items = vec![
+            HealResultItem::default(),
+            HealResultItem {
+                detail: "x".repeat(MAX_HEAL_STATUS_PAYLOAD_SIZE + 1),
+                ..Default::default()
+            },
+        ];
+        let (bytes, detail) = encode_heal_status_response("running", items, None, None, false, (9, 4), Some(&outcome))
+            .expect("bounded status with cumulative outcome");
+        assert!(bytes.len() <= MAX_HEAL_STATUS_PAYLOAD_SIZE);
+        let wire: serde_json::Value = serde_json::from_slice(&bytes).expect("bounded payload");
+        assert_eq!(wire["items"].as_array().expect("items").len(), 1);
+        assert_eq!(wire["truncated"], true);
+        assert_eq!((wire["next_seq"].as_u64(), wire["min_seq"].as_u64()), (Some(9), Some(4)));
+        assert_eq!(wire["outcome"]["counters"]["processed"], 256);
+        assert_eq!(wire["outcome"]["counters"]["healed"], 0);
+        assert_eq!(wire["outcome"]["objects"].as_array().expect("outcome window").len(), retained);
+        assert!(retained < 256 && outcome.objects_truncated);
+        assert_eq!(detail.as_deref(), Some("heal result items were truncated"));
+    }
+
    #[test]
    fn admission_response_preserves_all_admission_outcomes() {
        let cases = [
@@ -150,6 +150,7 @@ impl HealManager {
                                Err(DiskError::UnformattedDisk) => {
                                    if !super::super::replacement_readiness::auto_replacement_target_ready(disk, &local_disks)
                                        .await
+                                        && !super::super::replacement_readiness::directory_backed_replacement_fallback_enabled()
                                    {
                                        deferred_replacement_endpoints.insert(endpoint.to_string());
                                        skipped_invalid_count += 1;
@@ -313,7 +313,6 @@ impl HealManager {
                        completed_status_entry.outcome = Some(Arc::new(task.get_outcome().await));
                    }
                    let terminal_completion = !matches!(completed_status, HealTaskStatus::Retrying { .. });
-                    let successful_completion = matches!(completed_status, HealTaskStatus::Completed);
                    // Keep retry ownership continuous: status snapshots acquire
                    // these locks in the same active -> retrying order.
                    let mut retrying_heals_guard = if let (Some((request, _, error)), Some(cancel_token)) =
@@ -375,11 +374,11 @@ impl HealManager {
                        drop(stats);
                        if terminal_completion {
                            let notice_targets = take_mrf_repair_notice_targets(&mrf_repair_notice_targets_clone, &task_id);
-                            if successful_completion {
-                                emit_mrf_repaired_events(notice_targets);
-                            } else {
-                                release_mrf_repair_notice_targets(notice_targets);
-                            }
+                            // Neither task status nor the diagnostic outcome
+                            // window supplies a storage-owned repair receipt.
+                            // Release only the ingress lease for rediscovery;
+                            // preserve the producer's existing retry hints.
+                            release_mrf_repair_notice_targets(notice_targets);
                        }
                    }

@@ -706,20 +705,6 @@ fn move_mrf_repair_notice_targets(
    }
 }

-fn emit_mrf_repaired_events(targets: Vec<MrfRepairNoticeTarget>) {
-    for target in targets {
-        rustfs_common::mrf_channel::note_mrf_repaired(&target.bucket, &target.object, target.version_id);
-        rustfs_common::mrf_channel::release_mrf_identity(
-            target.kind,
-            &target.bucket,
-            &target.object,
-            target.version_id,
-            target.scope,
-            target.lease,
-        );
-    }
-}
-
 fn release_mrf_repair_notice_targets(targets: Vec<MrfRepairNoticeTarget>) {
    for target in targets {
        rustfs_common::mrf_channel::release_mrf_identity(
@@ -746,11 +731,8 @@ pub(super) fn prune_completed_heal_statuses(completed_heals: &mut HashMap<String
 }

 pub(super) fn prune_completed_heal_statuses_at(completed_heals: &mut HashMap<String, Arc<CompletedHealStatus>>, now: SystemTime) {
-    completed_heals.retain(|_, completed| {
-        now.duration_since(completed.completed_at)
-            .map(|age| age <= KEEP_HEAL_TASK_STATUS_DURATION)
-            .unwrap_or(false)
-    });
+    completed_heals
+        .retain(|_, completed| now.duration_since(completed.completed_at).unwrap_or_default() <= KEEP_HEAL_TASK_STATUS_DURATION);
    let entry_bytes = |key: &String, value: &Arc<CompletedHealStatus>| {
        key.capacity()
            .saturating_add(size_of::<(String, Arc<CompletedHealStatus>)>())
@@ -189,37 +189,104 @@ fn completed_retention_count_ttl_and_alias_eviction_are_bounded() {
    );
    entries.insert("future".to_string(), Arc::new(completed_retention_fixture(now + Duration::from_nanos(1))));
    prune_completed_heal_statuses_at(&mut entries, now);
-    assert_eq!(entries.len(), 1);
+    assert_eq!(entries.len(), 2);
    assert!(entries.contains_key("ttl-boundary"));
+    assert!(entries.contains_key("future"), "clock rollback must not expire a new completion");
    prune_completed_heal_statuses_at(&mut entries, now + Duration::from_nanos(1));
+    assert_eq!(entries.len(), 1);
+    assert!(entries.contains_key("future"));
+    prune_completed_heal_statuses_at(&mut entries, now + Duration::from_nanos(1) + KEEP_HEAL_TASK_STATUS_DURATION);
+    assert!(entries.contains_key("future"), "the exact TTL boundary remains retained");
+    prune_completed_heal_statuses_at(&mut entries, now + Duration::from_nanos(2) + KEEP_HEAL_TASK_STATUS_DURATION);
    assert!(entries.is_empty());
 }

+#[tokio::test]
+async fn completed_retention_clock_rollback_preserves_terminal_alias_queries() {
+    let completed_at = SystemTime::now() + Duration::from_secs(3600);
+    for status in [
+        HealTaskStatus::Completed,
+        HealTaskStatus::Failed {
+            error: "fixture failure".to_string(),
+        },
+        HealTaskStatus::Cancelled,
+    ] {
+        let manager = HealManager::new(Arc::new(MockStorage), None);
+        let mut snapshot = completed_retention_fixture(completed_at);
+        snapshot.status = status.clone();
+        let expected_progress = snapshot.progress.clone();
+        let snapshot = Arc::new(snapshot);
+        {
+            let mut completed = manager.completed_heals.lock().await;
+            completed.insert("canonical".to_string(), Arc::clone(&snapshot));
+            completed.insert("alias".to_string(), Arc::clone(&snapshot));
+        }
+        for token in ["canonical", "alias"] {
+            let report = manager
+                .get_task_report_since(token, Some(3))
+                .await
+                .expect("a clock rollback must retain terminal queries");
+            assert_eq!(report.status, status);
+            assert_eq!(report.progress, expected_progress);
+            assert_eq!(report.result_items.len(), 1);
+            assert_eq!((report.min_seq, report.next_seq), (3, 5));
+            assert!(!report.result_items_truncated);
+        }
+        let mut completed = manager.completed_heals.lock().await;
+        prune_completed_heal_statuses_at(&mut completed, completed_at + KEEP_HEAL_TASK_STATUS_DURATION);
+        assert_eq!(completed.len(), 2, "both tokens remain at the exact TTL boundary");
+        prune_completed_heal_statuses_at(&mut completed, completed_at + KEEP_HEAL_TASK_STATUS_DURATION + Duration::from_nanos(1));
+        assert!(completed.is_empty(), "both tokens expire after the TTL");
+    }
+}
+
+#[test]
+fn completed_retention_clock_rollback_keeps_count_and_alias_eviction_bounded() {
+    let now = SystemTime::UNIX_EPOCH + Duration::from_secs(3600);
+    let oldest = Arc::new(completed_retention_fixture(now + Duration::from_secs(1)));
+    let mut entries = HashMap::from([
+        ("oldest".to_string(), Arc::clone(&oldest)),
+        ("oldest-alias".to_string(), oldest),
+    ]);
+    for index in 2..=MAX_COMPLETED_HEAL_TOKENS {
+        entries.insert(
+            format!("task-{index}"),
+            Arc::new(completed_retention_fixture(now + Duration::from_secs(2))),
+        );
+    }
+    prune_completed_heal_statuses_at(&mut entries, now);
+    assert_eq!(entries.len(), MAX_COMPLETED_HEAL_TOKENS - 1);
+    assert!(!entries.contains_key("oldest"));
+    assert!(!entries.contains_key("oldest-alias"));
+}
+
 #[test]
 fn completed_retention_total_byte_cap_and_cap_plus_one() {
    let now = SystemTime::now();
-    let key = "large".to_string();
-    let mut entry = completed_retention_fixture(now);
-    let base_bytes = entry.retained_bytes() + key.capacity() + size_of::<(String, Arc<CompletedHealStatus>)>();
-    entry.retained_bytes.take();
-    entry.status = HealTaskStatus::Failed {
-        error: "x".repeat(MAX_COMPLETED_HEAL_BYTES - base_bytes),
-    };
-    assert_eq!(
-        entry.retained_bytes() + key.capacity() + size_of::<(String, Arc<CompletedHealStatus>)>(),
-        MAX_COMPLETED_HEAL_BYTES
-    );
-    let mut entries = HashMap::from([(key, Arc::new(entry))]);
-    prune_completed_heal_statuses_at(&mut entries, now);
-    assert_eq!(entries.len(), 1, "exact byte cap remains retained");
-    let mut over = Arc::try_unwrap(entries.remove("large").expect("entry retained")).expect("entry not shared");
-    over.retained_bytes.take();
-    if let HealTaskStatus::Failed { error } = &mut over.status {
-        *error = "x".repeat(error.len() + 1);
+    for completed_at in [now, now + Duration::from_secs(1)] {
+        let key = "large".to_string();
+        let mut entry = completed_retention_fixture(completed_at);
+        let base_bytes = entry.retained_bytes() + key.capacity() + size_of::<(String, Arc<CompletedHealStatus>)>();
+        entry.retained_bytes.take();
+        entry.status = HealTaskStatus::Failed {
+            error: "x".repeat(MAX_COMPLETED_HEAL_BYTES - base_bytes),
+        };
+        assert_eq!(
+            entry.retained_bytes() + key.capacity() + size_of::<(String, Arc<CompletedHealStatus>)>(),
+            MAX_COMPLETED_HEAL_BYTES
+        );
+        let mut entries = HashMap::from([(key, Arc::new(entry))]);
+        prune_completed_heal_statuses_at(&mut entries, now);
+        assert_eq!(entries.len(), 1, "exact byte cap remains retained");
+        let mut over = Arc::try_unwrap(entries.remove("large").expect("entry retained")).expect("entry not shared");
+        over.retained_bytes.take();
+        if let HealTaskStatus::Failed { error } = &mut over.status {
+            *error = "x".repeat(error.len() + 1);
+        }
+        entries.insert("large".to_string(), Arc::new(over));
+        prune_completed_heal_statuses_at(&mut entries, now);
+        assert!(entries.is_empty(), "oversized metadata cannot escape total byte bound");
    }
-    entries.insert("large".to_string(), Arc::new(over));
-    prune_completed_heal_statuses_at(&mut entries, now);
-    assert!(entries.is_empty(), "oversized metadata cannot escape total byte bound");
 }

 #[tokio::test]
@@ -877,6 +944,84 @@ fn bucket_request(bucket: &str, priority: HealPriority, source: HealRequestSourc
    request
 }

+fn scoped_object_request(bucket: &str, object: &str, pool_index: usize, set_index: usize) -> HealRequest {
+    HealRequest::new(
+        HealType::Object {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            version_id: None,
+        },
+        HealOptions {
+            pool_index: Some(pool_index),
+            set_index: Some(set_index),
+            ..Default::default()
+        },
+        HealPriority::Normal,
+    )
+}
+
+#[tokio::test]
+async fn scheduler_bulkhead_starts_other_sets_and_retains_same_set_tail() {
+    let manager = HealManager::new(Arc::new(MockStorage), None);
+    {
+        let mut config = manager.config.write().await;
+        config.max_concurrent_heals = 2;
+        config.max_concurrent_per_set = 1;
+        config.set_bulkhead_enable = true;
+        config.event_driven_scheduler_enable = false;
+        config.mainline_throttle_enable = false;
+    }
+
+    let first_set = scoped_object_request("scheduler-bulkhead-set-a-first", "object-a", 0, 1);
+    let first_set_id = first_set.id.clone();
+    let same_set_tail = scoped_object_request("scheduler-bulkhead-set-a-tail", "object-b", 0, 1);
+    let same_set_tail_id = same_set_tail.id.clone();
+    let other_set = scoped_object_request("scheduler-bulkhead-set-b", "object-c", 0, 2);
+    let other_set_id = other_set.id.clone();
+
+    let first_hook = Arc::new(CompletedRetentionHook::default());
+    let other_hook = Arc::new(CompletedRetentionHook::default());
+    {
+        let mut hooks = COMPLETED_RETENTION_HOOKS.lock().await;
+        hooks.insert("scheduler-bulkhead-set-a-first".to_string(), first_hook.clone());
+        hooks.insert("scheduler-bulkhead-set-b".to_string(), other_hook.clone());
+    }
+    {
+        let mut queue = manager.heal_queue.lock().await;
+        assert_eq!(queue.push(first_set), QueuePushOutcome::Accepted);
+        assert_eq!(queue.push(same_set_tail), QueuePushOutcome::Accepted);
+        assert_eq!(queue.push(other_set), QueuePushOutcome::Accepted);
+    }
+
+    process_manager_queue_once(&manager).await;
+    tokio::time::timeout(Duration::from_secs(5), first_hook.started.notified())
+        .await
+        .expect("first set task should start");
+    tokio::time::timeout(Duration::from_secs(5), other_hook.started.notified())
+        .await
+        .expect("other set task should start despite same-set tail");
+
+    assert_eq!(manager.get_active_task_count().await, 2);
+    assert_eq!(manager.get_queue_length().await, 1);
+    assert!(matches!(manager.get_task_status(&same_set_tail_id).await, Ok(HealTaskStatus::Pending)));
+    {
+        let active = manager.active_heals.lock().await;
+        assert!(active.contains_key(&first_set_id));
+        assert!(active.contains_key(&other_set_id));
+        assert!(!active.contains_key(&same_set_tail_id));
+        let counts = running_heal_set_counts(&active);
+        assert_eq!(counts.get("pool_0_set_1"), Some(&1));
+        assert_eq!(counts.get("pool_0_set_2"), Some(&1));
+    }
+
+    manager.cancel_task(&first_set_id).await.expect("cancel first active task");
+    manager.cancel_task(&other_set_id).await.expect("cancel other active task");
+    COMPLETED_RETENTION_HOOKS
+        .lock()
+        .await
+        .retain(|bucket, _| !bucket.starts_with("scheduler-bulkhead-"));
+}
+
 #[test]
 fn test_push_displacing_lower_priority_actually_enqueues_new_request() {
    // Regression for the release-build defect where the enqueue side effect lived inside
@@ -3096,7 +3241,7 @@ async fn test_cancel_task_removes_queued_request() {
 }

 #[tokio::test]
-async fn test_mrf_repaired_notice_waits_for_successful_completion() {
+async fn mrf_ownership_unverified_completion_does_not_emit_repaired() {
    let bucket = "mrf-completion-success";
    let object = "object";
    let version_id = Some([9u8; 16]);
@@ -3126,21 +3271,81 @@ async fn test_mrf_repaired_notice_waits_for_successful_completion() {
    );

    process_manager_queue_once(&manager).await;
-    for _ in 0..100 {
-        let events = rustfs_common::mrf_channel::take_mrf_repaired_events_for(bucket);
-        if !events.is_empty() {
-            assert_eq!(events.len(), 1);
-            assert_eq!(events[0].object.as_ref(), object);
-            assert_eq!(events[0].version_id, version_id);
-            return;
+    tokio::time::timeout(Duration::from_secs(5), async {
+        loop {
+            let stats = manager.get_statistics().await;
+            if stats.successful_tasks + stats.failed_tasks > 0 {
+                break;
+            }
+            tokio::task::yield_now().await;
        }
-        tokio::time::sleep(Duration::from_millis(10)).await;
-    }
-    panic!("successful MRF-owned heal should emit one repaired event");
+    })
+    .await
+    .expect("scheduler completes the task");
+    assert!(rustfs_common::mrf_channel::take_mrf_repaired_events_for(bucket).is_empty());
+    assert!(!lock_mrf_repair_notice_targets(&manager.mrf_repair_notice_targets).contains_key(&receipt.task_id));
 }

 #[tokio::test]
-async fn test_mrf_repaired_notice_removed_on_queued_cancel_without_event() {
+async fn mrf_ownership_dry_run_and_empty_window_do_not_emit_repaired() {
+    for empty_window in [false, true] {
+        let bucket = if empty_window {
+            "mrf-empty-outcome"
+        } else {
+            "mrf-dry-run-outcome"
+        };
+        let manager = HealManager::new(Arc::new(MockStorage), None);
+        let request = HealRequest::new(
+            if empty_window {
+                HealType::Cluster
+            } else {
+                HealType::Object {
+                    bucket: bucket.to_string(),
+                    object: "object".to_string(),
+                    version_id: None,
+                }
+            },
+            HealOptions {
+                recursive: true,
+                dry_run: !empty_window,
+                recreate_missing: true,
+                ..Default::default()
+            },
+            HealPriority::Normal,
+        );
+        let receipt = manager
+            .submit_mrf_heal_request_with_receipt(request, Arc::from(bucket), Arc::from("object"), None)
+            .await
+            .expect("notice target registered");
+        process_manager_queue_once(&manager).await;
+        tokio::time::timeout(Duration::from_secs(5), async {
+            loop {
+                let stats = manager.get_statistics().await;
+                if stats.successful_tasks + stats.failed_tasks > 0 {
+                    break;
+                }
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("scheduler completed");
+        let report = manager.get_task_report(&receipt.task_id).await.expect("completed report");
+        assert_eq!(report.status, HealTaskStatus::Completed);
+        let outcome = report.outcome.expect("canonical outcome");
+        if empty_window {
+            assert!(outcome.objects.is_empty());
+        } else {
+            assert_eq!(
+                outcome.objects[0].disposition,
+                crate::heal::outcome::HealObjectDisposition::DryRunObserved
+            );
+        }
+        assert!(rustfs_common::mrf_channel::take_mrf_repaired_events_for(bucket).is_empty());
+    }
+}
+
+#[tokio::test]
+async fn mrf_ownership_queued_cancel_does_not_emit_repaired() {
    let bucket = "mrf-completion-cancel";
    let object = "object";
    let _ = rustfs_common::mrf_channel::take_mrf_repaired_events_for(bucket);
@@ -25,12 +25,13 @@
 //! set, rewritten on a group-commit cadence (every flush interval or flush
 //! threshold new intents). A rewrite is atomic at the record level only — a
 //! torn tail simply truncates during replay because every record carries its
-//! own CRC32. Losing the last flush window (≤500 ms) is acceptable because
-//! every producer keeps its own safety net: read-repair re-detects on the
-//! next failing read, and the scanner's corrupt-metadata branch leaves a
-//! pending-ledger entry behind even when its MRF intent is accepted
-//! (backlog#1894 axis A), so a lost intent is retried by the ledger rather
-//! than waiting for the failed-object TTL to re-scan the path.
+//! own CRC32. Neither ingress nor manager admission is a durable ownership
+//! receipt. The last flush window can be lost. Read-repair can rediscover a
+//! failed read; the scanner retains bounded, expiring retry hints. Partial
+//! writes also use a best-effort in-memory fast path, not a durable successor.
+//! These mechanisms must not be reported as verified repair completion.
+//! The partial-write caller's restart-survival requirement remains unmet by
+//! admission alone; a verified durable handoff is still required.

 use super::{DiskStore, HealDiskExt as _, local_disk_map_read};
 use crate::heal::manager::{HealManager, MrfRepairNoticeTarget};
@@ -585,8 +586,8 @@ impl MrfRuntime {
            self.dirty = true;
            match submit_mrf_heal_request(manager, &intent).await {
                // Accepted intents leave the pending set; the next flush persists the
-                // smaller snapshot. The scanner ledger is cleared later, when the
-                // canonical heal task reaches a successful terminal completion.
+                // smaller snapshot. This is not a durable successor receipt and
+                // does not discharge the producer's existing retry hints.
                Ok(HealAdmissionResult::Accepted) | Ok(HealAdmissionResult::Merged) => {}
                Ok(HealAdmissionResult::Full) | Ok(HealAdmissionResult::Dropped(HealAdmissionDropReason::QueueFull)) => {
                    intent.attempts = intent.attempts.saturating_add(1);
@@ -123,6 +123,13 @@ pub struct CommittedSnapshot {
    payload: Vec<u8>,
 }

+#[derive(Default)]
+struct SnapshotReadStats {
+    file_reads: usize,
+    bytes_read: usize,
+    peak_file_bytes: usize,
+}
+
 impl CommittedSnapshot {
    /// Persistent single-writer sequence, not a process UUID ordering.
    pub fn sequence(&self) -> u64 {
@@ -162,6 +169,15 @@ pub enum RecoverySnapshot {
 }

 async fn read_bounded(disk: &EcstoreDiskStore, path: &str, limit: usize) -> Result<Option<Vec<u8>>, SnapshotError> {
+    read_bounded_with_stats(disk, path, limit, None).await
+}
+
+async fn read_bounded_with_stats(
+    disk: &EcstoreDiskStore,
+    path: &str,
+    limit: usize,
+    mut stats: Option<&mut SnapshotReadStats>,
+) -> Result<Option<Vec<u8>>, SnapshotError> {
    let reader = match EcstoreDiskAPI::read_file(disk.as_ref(), RUSTFS_META_BUCKET, path).await {
        Ok(reader) => reader,
        Err(EcstoreDiskError::FileNotFound | EcstoreDiskError::VolumeNotFound) => return Ok(None),
@@ -178,6 +194,11 @@ async fn read_bounded(disk: &EcstoreDiskStore, path: &str, limit: usize) -> Resu
    if bytes.len() > limit {
        return Err(SnapshotError::TooLarge);
    }
+    if let Some(stats) = stats.as_mut() {
+        stats.file_reads += 1;
+        stats.bytes_read = stats.bytes_read.checked_add(bytes.len()).ok_or(SnapshotError::TooLarge)?;
+        stats.peak_file_bytes = stats.peak_file_bytes.max(bytes.len());
+    }
    Ok(Some(bytes))
 }

@@ -197,17 +218,26 @@ fn select_snapshot(selected: &mut Option<CommittedSnapshot>, candidate: Committe
 }

 async fn read_committed(disks: &[EcstoreDiskStore], limit: usize) -> Result<Option<CommittedSnapshot>, SnapshotError> {
+    read_committed_with_stats(disks, limit, None).await
+}
+
+async fn read_committed_with_stats(
+    disks: &[EcstoreDiskStore],
+    limit: usize,
+    mut stats: Option<&mut SnapshotReadStats>,
+) -> Result<Option<CommittedSnapshot>, SnapshotError> {
    let mut selected = None;
    let mut damaged = None;
    let mut identities = HashMap::new();
    for disk in disks {
        for (manifest_path, payload_path) in MANIFEST_PATHS.into_iter().zip(PAYLOAD_PATHS) {
            let candidate = async {
-                let Some(manifest) = read_bounded(disk, manifest_path, MANIFEST_LEN).await? else {
+                let Some(manifest) = read_bounded_with_stats(disk, manifest_path, MANIFEST_LEN, stats.as_deref_mut()).await?
+                else {
                    return Ok(None);
                };
                let header = Manifest::decode(&manifest, limit)?;
-                let payload = read_bounded(disk, payload_path, header.payload_len)
+                let payload = read_bounded_with_stats(disk, payload_path, header.payload_len, stats.as_deref_mut())
                    .await?
                    .ok_or(SnapshotError::Corrupt)?;
                CommittedSnapshot::decode(&manifest, payload, limit).map(Some)
@@ -492,6 +522,69 @@ mod tests {
        assert_eq!(recovered.manifest.sequence, 1);
    }

+    #[tokio::test]
+    async fn committed_reader_reopens_previous_anchor_across_publication_boundaries() {
+        let owner = Uuid::new_v4();
+        let old = payload("old");
+        let next = payload("next");
+        let boundaries = [
+            ("payload-only", next.clone(), None),
+            ("torn-manifest", next.clone(), Some(manifest(owner, 2, &next)[..20].to_vec())),
+            ("stale-payload", old.clone(), Some(manifest(owner, 2, &next))),
+        ];
+        for (case, successor_payload, successor_manifest) in boundaries {
+            let root = TempDir::new().expect("test directory");
+            let store = disk(&root, "disk").await;
+            commit(&store, 0, owner, 1, &old).await;
+            install(&store, PAYLOAD_PATHS[1], &successor_payload).await;
+            if let Some(manifest) = &successor_manifest {
+                install(&store, MANIFEST_PATHS[1], manifest).await;
+            }
+
+            let reopened = disk(&root, "disk").await;
+            let recovered = read_committed(std::slice::from_ref(&reopened), 4096)
+                .await
+                .unwrap_or_else(|error| panic!("{case}: old anchor must remain readable after reopen: {error:?}"))
+                .unwrap_or_else(|| panic!("{case}: previous committed anchor missing after reopen"));
+            assert_eq!(recovered.manifest.sequence, 1, "{case}: successor must not become authoritative");
+            assert_eq!(recovered.payload, old, "{case}: previous payload must survive");
+            assert_eq!(
+                EcstoreDiskAPI::read_all(reopened.as_ref(), RUSTFS_META_BUCKET, PAYLOAD_PATHS[0])
+                    .await
+                    .expect("old payload retained")
+                    .as_ref(),
+                old.as_slice(),
+                "{case}: previous payload bytes changed"
+            );
+            assert_eq!(
+                EcstoreDiskAPI::read_all(reopened.as_ref(), RUSTFS_META_BUCKET, MANIFEST_PATHS[0])
+                    .await
+                    .expect("old manifest retained")
+                    .as_ref(),
+                manifest(owner, 1, &old).as_slice(),
+                "{case}: previous manifest bytes changed"
+            );
+            assert_eq!(
+                EcstoreDiskAPI::read_all(reopened.as_ref(), RUSTFS_META_BUCKET, PAYLOAD_PATHS[1])
+                    .await
+                    .expect("successor payload retained")
+                    .as_ref(),
+                successor_payload.as_slice(),
+                "{case}: successor evidence changed"
+            );
+            if let Some(manifest) = &successor_manifest {
+                assert_eq!(
+                    EcstoreDiskAPI::read_all(reopened.as_ref(), RUSTFS_META_BUCKET, MANIFEST_PATHS[1])
+                        .await
+                        .expect("successor manifest retained")
+                        .as_ref(),
+                    manifest.as_slice(),
+                    "{case}: successor manifest evidence changed"
+                );
+            }
+        }
+    }
+
    #[tokio::test]
    async fn stale_manifest_cas_cannot_replace_committed_anchor() {
        let root = TempDir::new().expect("test directory");
@@ -571,6 +664,35 @@ mod tests {
        );
    }

+    #[tokio::test]
+    async fn committed_reader_resource_bounds_are_measured() {
+        let root = TempDir::new().expect("test directory");
+        let first = disk(&root, "first").await;
+        let second = disk(&root, "second").await;
+        let owner = Uuid::new_v4();
+        let old = [payload("old-0"), payload("old-1")].concat();
+        let new = [payload("new-0"), payload("new-1"), payload("new-2")].concat();
+        commit(&first, 0, owner, 1, &old).await;
+        commit(&second, 1, owner, 2, &new).await;
+
+        let mut stats = SnapshotReadStats::default();
+        let recovered = read_committed_with_stats(&[first, second], 4096, Some(&mut stats))
+            .await
+            .expect("read committed replicas")
+            .expect("committed snapshot");
+
+        assert_eq!(recovered.sequence(), 2);
+        assert_eq!(recovered.payload(), new.as_slice());
+        assert_eq!(recovered.manifest.payload_len, new.len());
+        assert_eq!(stats.file_reads, 4, "only committed manifests and their payloads are materialized");
+        assert_eq!(stats.bytes_read, (MANIFEST_LEN * 2) + old.len() + new.len());
+        assert_eq!(
+            stats.peak_file_bytes,
+            new.len().max(MANIFEST_LEN),
+            "reader peak allocation remains bounded by one manifest or payload file"
+        );
+    }
+
    #[tokio::test]
    async fn legacy_inspection_rejects_complete_subsets_and_scope_ambiguity() {
        let scoped = |set_index| {
@@ -15,6 +15,7 @@
 //! Execution results are separate from repair responsibility. A legacy
 //! successful storage call supplies no authoritative repair receipt.

+use serde::{Deserialize, Serialize};
 use std::{collections::VecDeque, time::SystemTime};
 use uuid::Uuid;

@@ -22,14 +23,16 @@ const MAX_OUTCOME_ITEMS: usize = 128;
 const MAX_OUTCOME_BYTES: usize = 64 * 1024;
 const MAX_OUTCOME_DETAIL_BYTES: usize = 1024;

-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "snake_case")]
 pub enum HealObjectKind {
    Object,
    Metadata,
    Decode,
 }

-#[derive(Debug, Clone, PartialEq, Eq)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "camelCase")]
 pub struct HealObjectIdentity {
    pub kind: HealObjectKind,
    pub bucket: String,
@@ -41,7 +44,8 @@ pub struct HealObjectIdentity {
    pub set_index: Option<usize>,
 }

-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "snake_case")]
 pub enum HealDeferredReason {
    DanglingDeleteGrace,
    TransientUsageCache,
@@ -49,14 +53,21 @@ pub enum HealDeferredReason {
    Deadline,
 }

-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "snake_case")]
 pub enum HealFailureClass {
    Recoverable,
    RetryExhausted,
    Permanent,
 }

-#[derive(Debug, Clone, PartialEq, Eq)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+#[serde(
+    tag = "state",
+    content = "details",
+    rename_all = "snake_case",
+    rename_all_fields = "camelCase"
+)]
 pub enum HealObjectDisposition {
    /// The legacy storage response does not prove the requested check or commit.
    Unknown,
@@ -72,7 +83,8 @@ pub enum HealObjectDisposition {
    DryRunObserved,
 }

-#[derive(Debug, Clone, PartialEq, Eq)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "camelCase")]
 pub struct HealObjectOutcome {
    pub identity: HealObjectIdentity,
    pub disposition: HealObjectDisposition,
@@ -89,7 +101,8 @@ impl HealObjectOutcome {
    }
 }

-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "snake_case")]
 pub enum HealTraversalCoverage {
    #[default]
    Unknown,
@@ -97,14 +110,16 @@ pub enum HealTraversalCoverage {
    Complete,
 }

-#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
 pub enum HealAbortReason {
    Cancelled,
    Deadline,
    Untraversable,
 }

-#[derive(Debug, Clone, Copy, Default, PartialEq, Eq)]
+#[derive(Debug, Clone, Copy, Default, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(tag = "state", content = "reason", rename_all = "snake_case")]
 pub enum HealExecutionOutcome {
    #[default]
    Pending,
@@ -114,7 +129,8 @@ pub enum HealExecutionOutcome {
    Aborted(HealAbortReason),
 }

-#[derive(Debug, Clone, Default, PartialEq, Eq)]
+#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
 pub struct HealOutcomeCounters {
    pub processed: u64,
    pub healed: u64,
@@ -127,7 +143,8 @@ pub struct HealOutcomeCounters {
    pub overflowed: bool,
 }

-#[derive(Debug, Clone, Default, PartialEq, Eq)]
+#[derive(Debug, Clone, Default, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "camelCase")]
 pub struct HealTaskOutcome {
    pub execution: HealExecutionOutcome,
    pub coverage: HealTraversalCoverage,
@@ -135,11 +152,99 @@ pub struct HealTaskOutcome {
    /// A bounded diagnostic window, not a complete responsibility ledger.
    pub objects: VecDeque<HealObjectOutcome>,
    pub objects_truncated: bool,
+    #[serde(skip)]
    retained_object_bytes: usize,
+    #[serde(skip)]
    untraversable: bool,
 }

+#[derive(Debug, thiserror::Error)]
+pub enum HealOutcomeWireError {
+    #[error("heal outcome is missing execution or counters")]
+    MissingFields,
+    #[error("heal outcome has invalid or unsupported execution fields")]
+    InvalidFields(#[from] serde_json::Error),
+    #[error("finished heal summary contradicts its canonical outcome")]
+    ContradictoryCompletion,
+}
+
+/// Reconcile a peer's successful legacy summary using the canonical owner types.
+/// A running retry may legitimately retain the preceding attempt's outcome.
+pub fn legacy_wire_status<'a>(
+    summary: &'a str,
+    wire: &serde_json::Value,
+    truncated: bool,
+) -> Result<(&'a str, Option<String>), HealOutcomeWireError> {
+    if summary != "finished" {
+        return Ok((summary, None));
+    }
+    let execution = HealExecutionOutcome::deserialize(wire.get("execution").ok_or(HealOutcomeWireError::MissingFields)?)?;
+    let counters = HealOutcomeCounters::deserialize(wire.get("counters").ok_or(HealOutcomeWireError::MissingFields)?)?;
+    if matches!(execution, HealExecutionOutcome::Pending | HealExecutionOutcome::Running)
+        || (execution == HealExecutionOutcome::Completed && counters.failed > 0)
+    {
+        return Err(HealOutcomeWireError::ContradictoryCompletion);
+    }
+    let (adapted, detail) = legacy_execution_status(summary, None, execution, &counters);
+    Ok((
+        adapted,
+        if adapted != summary {
+            heal_status_detail(detail, truncated)
+        } else {
+            None
+        },
+    ))
+}
+
+pub(crate) fn heal_status_detail(detail: Option<String>, truncated: bool) -> Option<String> {
+    if !truncated {
+        return detail;
+    }
+    let truncation = "heal result items were truncated";
+    Some(detail.map_or_else(|| truncation.to_string(), |detail| format!("{detail}; {truncation}")))
+}
+
+fn legacy_execution_status<'a>(
+    summary: &'a str,
+    detail: Option<String>,
+    execution: HealExecutionOutcome,
+    counters: &HealOutcomeCounters,
+) -> (&'a str, Option<String>) {
+    if summary != "finished" {
+        return (summary, detail);
+    }
+    match execution {
+        HealExecutionOutcome::CompletedWithErrors => (
+            "stopped",
+            Some(format!("heal traversal completed with errors: {} failed objects", counters.failed)),
+        ),
+        HealExecutionOutcome::Aborted(reason) => {
+            let reason = match reason {
+                HealAbortReason::Cancelled => "cancelled",
+                HealAbortReason::Deadline => "timed out",
+                HealAbortReason::Untraversable => "untraversable",
+            };
+            ("stopped", Some(format!("heal task {reason}")))
+        }
+        HealExecutionOutcome::Completed if counters.unknown > 0 => (
+            summary,
+            Some(format!(
+                "heal traversal completed; authoritative storage proof is unavailable for {} objects",
+                counters.unknown
+            )),
+        ),
+        HealExecutionOutcome::Pending | HealExecutionOutcome::Running => {
+            ("running", Some("heal execution has not reached a terminal outcome".to_string()))
+        }
+        HealExecutionOutcome::Completed => (summary, detail),
+    }
+}
+
 impl HealTaskOutcome {
+    pub(crate) fn legacy_status<'a>(&self, summary: &'a str, detail: Option<String>) -> (&'a str, Option<String>) {
+        legacy_execution_status(summary, detail, self.execution, &self.counters)
+    }
+
    pub(crate) fn start(&mut self) {
        if self.execution != HealExecutionOutcome::Aborted(HealAbortReason::Cancelled) {
            self.execution = HealExecutionOutcome::Running;
@@ -292,6 +397,69 @@ mod canonical_outcome_tests {
        assert!(outcome.objects.len() < MAX_OUTCOME_ITEMS);
    }

+    #[test]
+    fn outcome_v3_serialization_keeps_unverified_dispositions_and_window_bounds() {
+        let mut outcome = HealTaskOutcome::default();
+        for disposition in [
+            HealObjectDisposition::Unknown,
+            HealObjectDisposition::Deferred {
+                reason: HealDeferredReason::DanglingDeleteGrace,
+                retry_not_before: None,
+            },
+            HealObjectDisposition::DryRunObserved,
+        ] {
+            outcome.record(item(disposition));
+        }
+        outcome.finish(None);
+        let wire = serde_json::to_value(&outcome).expect("canonical wire view");
+        assert_eq!(wire["execution"]["state"], "completed");
+        assert_eq!(wire["counters"]["healed"], 0);
+        assert_eq!(wire["counters"]["skipped"], 3);
+        assert_eq!(wire["objects"][1]["disposition"]["details"]["reason"], "dangling_delete_grace");
+        assert!(wire["objects"][1]["identity"]["bucketIncarnationId"].is_null());
+        for _ in 0..MAX_OUTCOME_ITEMS + 1 {
+            let mut result = item(HealObjectDisposition::Unknown);
+            result.detail = Some("\"".repeat(MAX_OUTCOME_DETAIL_BYTES));
+            outcome.record(result);
+        }
+        let bytes = serde_json::to_vec(&outcome).expect("bounded canonical samples");
+        assert!(
+            bytes.len() < 8 * MAX_OUTCOME_BYTES,
+            "JSON escaping remains bounded independently of object count"
+        );
+        assert!(outcome.objects_truncated);
+    }
+
+    #[test]
+    fn outcome_v3_wire_consistency_rejects_unknown_success_without_rejecting_extensions() {
+        let mut outcome = HealTaskOutcome::default();
+        outcome.finish(None);
+        let mut wire = serde_json::to_value(&outcome).expect("canonical snapshot");
+        wire["execution"]["futureField"] = serde_json::json!({"new": true});
+        wire["counters"]["futureCounter"] = serde_json::json!(42);
+        assert_eq!(
+            legacy_wire_status("finished", &wire, false).expect("unknown extension fields"),
+            ("finished", None)
+        );
+        wire["execution"]["state"] = serde_json::json!("future_execution");
+        assert!(legacy_wire_status("finished", &wire, false).is_err());
+        assert_eq!(
+            legacy_wire_status("running", &wire, false).expect("unknown nonterminal outcome"),
+            ("running", None)
+        );
+        wire["execution"] = serde_json::json!({"state":"completed"});
+        wire["counters"]["failed"] = serde_json::json!(1);
+        assert!(matches!(
+            legacy_wire_status("finished", &wire, false),
+            Err(HealOutcomeWireError::ContradictoryCompletion)
+        ));
+        wire.as_object_mut().expect("object").remove("execution");
+        assert!(matches!(
+            legacy_wire_status("finished", &wire, false),
+            Err(HealOutcomeWireError::MissingFields)
+        ));
+    }
+
    #[test]
    fn canonical_outcome_counter_overflow_cannot_claim_complete_coverage() {
        let mut outcome = HealTaskOutcome::default();
@@ -18,6 +18,25 @@ use super::{
    DiskOption, DiskStore, Endpoint, HealDiskExt as _, local_disk_map_read, new_disk, resume::ReplacementTargetIdentity,
 };

+/// Whether automatic replacement may fall back to the set-wide format heal when
+/// a target cannot pass the independent-mount admission.
+///
+/// Directory-backed deployments already declare, through
+/// `RUSTFS_UNSAFE_BYPASS_DISK_CHECK`, that their endpoints are plain
+/// directories sharing a device with the host root. Those endpoints can never
+/// satisfy [`auto_replacement_target_identity`], so without this fallback a
+/// runtime-wiped or replaced directory disk would stay deferred forever. The
+/// admission check itself is never bypassed; the fallback only routes the heal
+/// through the ordinary format path that formats every unformatted disk in the
+/// set, which is exactly what the pre-admission `heal_disk` path did.
+pub(crate) fn directory_backed_replacement_fallback_enabled() -> bool {
+    rustfs_utils::get_env_bool_with_aliases(
+        rustfs_config::ENV_UNSAFE_BYPASS_DISK_CHECK,
+        &[rustfs_config::ENV_MINIO_CI],
+        rustfs_config::DEFAULT_UNSAFE_BYPASS_DISK_CHECK,
+    )
+}
+
 pub(crate) async fn auto_replacement_target_ready(disk: &DiskStore, local_disks: &[DiskStore]) -> bool {
    auto_replacement_target_identity(disk, local_disks).await.is_some()
 }
@@ -184,12 +203,47 @@ mod tests {
        assert!(endpoint.is_local);
    }

+    #[test]
+    fn directory_backed_fallback_is_off_by_default() {
+        temp_env::with_vars(
+            [
+                (rustfs_config::ENV_UNSAFE_BYPASS_DISK_CHECK, None::<&str>),
+                (rustfs_config::ENV_MINIO_CI, None::<&str>),
+            ],
+            || assert!(!directory_backed_replacement_fallback_enabled()),
+        );
+    }
+
+    #[test]
+    fn directory_backed_fallback_follows_the_disk_check_bypass() {
+        temp_env::with_vars(
+            [
+                (rustfs_config::ENV_UNSAFE_BYPASS_DISK_CHECK, Some("true")),
+                (rustfs_config::ENV_MINIO_CI, None::<&str>),
+            ],
+            || assert!(directory_backed_replacement_fallback_enabled()),
+        );
+        temp_env::with_vars(
+            [
+                (rustfs_config::ENV_UNSAFE_BYPASS_DISK_CHECK, Some("false")),
+                (rustfs_config::ENV_MINIO_CI, Some("true")),
+            ],
+            || {
+                assert!(
+                    !directory_backed_replacement_fallback_enabled(),
+                    "the canonical key must win over the alias"
+                )
+            },
+        );
+    }
+
    #[tokio::test]
    async fn runtime_environment_cannot_bypass_mount_admission() {
        temp_env::async_with_vars(
            [
                ("RUSTFS_TEST_AUTO_REPLACEMENT_READINESS_BYPASS", Some("1")),
                ("RUSTFS_E2E_AUTO_REPLACEMENT_READINESS_BYPASS", Some("1")),
+                (rustfs_config::ENV_UNSAFE_BYPASS_DISK_CHECK, Some("true")),
            ],
            async {
                let temp = TempDir::new().expect("temporary replacement root should be created");
@@ -14,8 +14,107 @@
 /// bucket/cluster/prefix heal: the recursive bucket-objects sweep and the erasure-set usage baseline
 use super::*;
 use crate::heal::progress::{add_bytes, increment_counter, stable_generation};
+use crate::heal::storage::HealListItem;
 use crate::heal::utils::format_set_disk_id;

+const MAX_DEFERRED_OBJECTS: usize = 256;
+const MAX_DEFERRED_BYTES: usize = 256 * 1024;
+const MAX_DEFERRED_FORWARD_PAGES: u64 = 2;
+const MAX_DEFERRED_AGE: Duration = Duration::from_secs(30);
+
+struct DeferredObject {
+    name: String,
+    version_id: Option<String>,
+    attempt: u32,
+    page: u64,
+    first_failure: Option<tokio::time::Instant>,
+    due: tokio::time::Instant,
+}
+
+impl DeferredObject {
+    fn new(item: HealListItem, page: u64) -> Self {
+        Self {
+            name: item.name,
+            version_id: item.version_id,
+            attempt: 0,
+            page,
+            first_failure: None,
+            due: tokio::time::Instant::now(),
+        }
+    }
+
+    fn payload_bytes(&self) -> usize {
+        self.name
+            .capacity()
+            .saturating_add(self.version_id.as_ref().map_or(0, String::capacity))
+    }
+
+    fn expired(&self) -> bool {
+        self.first_failure.is_some_and(|first| first.elapsed() >= MAX_DEFERRED_AGE)
+    }
+
+    fn defer(&mut self, delay: Duration) {
+        let now = tokio::time::Instant::now();
+        let first = *self.first_failure.get_or_insert(now);
+        self.attempt += 1;
+        self.due = (now + delay).min(first + MAX_DEFERRED_AGE);
+    }
+}
+
+// Only failed identities are retained. The current listing page remains owned
+// by the caller; capacity pressure stops fetching, never discards that page.
+struct DeferredWindow {
+    objects: VecDeque<DeferredObject>,
+    bytes: usize,
+}
+
+impl Default for DeferredWindow {
+    fn default() -> Self {
+        Self {
+            objects: VecDeque::new(),
+            // Charge every possible slot up front, including spare capacity.
+            bytes: MAX_DEFERRED_OBJECTS * size_of::<DeferredObject>(),
+        }
+    }
+}
+
+impl DeferredWindow {
+    fn push(&mut self, item: DeferredObject) -> std::result::Result<(), DeferredObject> {
+        let bytes = item.payload_bytes();
+        if self.objects.len() >= MAX_DEFERRED_OBJECTS || bytes > MAX_DEFERRED_BYTES.saturating_sub(self.bytes) {
+            return Err(item);
+        }
+        self.bytes += bytes;
+        self.objects.push_back(item);
+        Ok(())
+    }
+
+    fn pop_due(&mut self) -> Option<DeferredObject> {
+        let now = tokio::time::Instant::now();
+        let index = self.objects.iter().position(|item| item.due <= now)?;
+        let item = self.objects.remove(index)?;
+        self.bytes -= item.payload_bytes();
+        Some(item)
+    }
+
+    fn next_due(&self) -> Option<tokio::time::Instant> {
+        self.objects.iter().map(|item| item.due).min()
+    }
+
+    fn can_advance(&self, page: u64) -> bool {
+        self.objects.len() < MAX_DEFERRED_OBJECTS
+            && self.bytes < MAX_DEFERRED_BYTES
+            && self
+                .objects
+                .iter()
+                .all(|item| page.saturating_sub(item.page) < MAX_DEFERRED_FORWARD_PAGES)
+    }
+}
+
+#[cfg(test)]
+#[path = "tests/deferred_retry_window.rs"]
+mod deferred_retry_window;
+
 fn unavailable_recreate_error(result: &HealResultItem, opts: &HealOpts) -> Option<Error> {
    if opts.dry_run || !opts.recreate {
        return None;
@@ -305,83 +404,122 @@ impl HealTask {

        for (set_disk_id, heal_opts) in listing_scopes {
            let mut continuation_token: Option<String> = None;
-            loop {
-                self.check_control_flags().await?;
-                let mut listing_attempt = 0;
-                let (objects, next_token, is_truncated) = loop {
+            let mut deferred = DeferredWindow::default();
+            let mut inline_retry: Option<DeferredObject> = None;
+            let mut page_number = 0_u64;
+            let mut aborted_progress_unknown = false;
+            let mut pending = Vec::<HealListItem>::new().into_iter();
+            let mut listing_finished = false;
+            let mut listing_attempt = 0;
+            let mut listing_due = tokio::time::Instant::now();
+            let scope_result: Result<()> = async {
+                loop {
+                    self.check_control_flags().await?;
+                    if listing_finished && pending.as_slice().is_empty() && deferred.objects.is_empty() && inline_retry.is_none()
+                    {
+                        break;
+                    }
                    self.pace_mainline().await?;
-                    let page = if let Some(set_disk_id) = set_disk_id.as_deref() {
-                        self.await_with_control(self.storage.list_versions_for_heal_page_disk_walk(
-                            set_disk_id,
-                            bucket,
-                            prefix,
-                            continuation_token.as_deref(),
-                            false,
-                        ))
-                        .await
-                    } else {
-                        self.await_with_control(self.storage.list_objects_for_heal_page(
-                            bucket,
-                            prefix,
-                            continuation_token.as_deref(),
-                            false,
-                        ))
-                        .await
-                    };
-                    match page {
-                        Ok(page) => break page,
-                        Err(error @ (Error::TaskCancelled | Error::TaskTimeout)) => return Err(error),
-                        Err(error) => {
-                            self.outcome.write().await.attempt_failed();
-                            if error.is_recoverable_heal() && listing_attempt < MAX_BUCKET_OBJECT_HEAL_RETRIES {
-                                listing_attempt += 1;
+                    // Listing and object retries share this safe boundary. A
+                    // failed listing never hides an already-due object retry.
+                    let item = deferred.pop_due().or_else(|| {
+                        if inline_retry
+                            .as_ref()
+                            .is_some_and(|item| item.due <= tokio::time::Instant::now())
+                        {
+                            inline_retry.take()
+                        } else if inline_retry.is_none() {
+                            pending.next().map(|item| DeferredObject::new(item, page_number))
+                        } else {
+                            None
+                        }
+                    });
+                    let Some(mut item) = item else {
+                        let can_list = !listing_finished && inline_retry.is_none() && deferred.can_advance(page_number);
+                        if can_list && listing_due <= tokio::time::Instant::now() {
+                            let page = if let Some(set_disk_id) = set_disk_id.as_deref() {
+                                self.await_with_control(self.storage.list_versions_for_heal_page_disk_walk(
+                                    set_disk_id,
+                                    bucket,
+                                    prefix,
+                                    continuation_token.as_deref(),
+                                    false,
+                                ))
+                                .await
+                            } else {
+                                self.await_with_control(self.storage.list_objects_for_heal_page(
+                                    bucket,
+                                    prefix,
+                                    continuation_token.as_deref(),
+                                    false,
+                                ))
+                                .await
+                            };
+                            match page {
+                                Ok((objects, next_token, is_truncated)) => {
+                                    page_number = page_number.saturating_add(1);
+                                    continuation_token = next_heal_listing_token(bucket, prefix, next_token, is_truncated)?;
+                                    listing_finished = continuation_token.is_none();
+                                    listing_attempt = 0;
+                                    listing_due = tokio::time::Instant::now();
+                                    pending = objects.into_iter();
+                                }
+                                Err(error @ (Error::TaskCancelled | Error::TaskTimeout)) => return Err(error),
+                                Err(error) => {
+                                    self.outcome.write().await.attempt_failed();
+                                    if error.is_recoverable_heal() && listing_attempt < MAX_BUCKET_OBJECT_HEAL_RETRIES {
+                                        listing_attempt += 1;
+                                        listing_due =
+                                            tokio::time::Instant::now() + self.bucket_object_retry_delay(listing_attempt);
+                                        continue;
+                                    }
+                                    self.outcome.write().await.mark_untraversable();
+                                    return Err(Error::HealListingFailed {
+                                        bucket: bucket.to_string(),
+                                        source: Box::new(error),
+                                    });
+                                }
+                            }
+                            continue;
+                        } else {
+                            let due = deferred
+                                .next_due()
+                                .into_iter()
+                                .chain(inline_retry.as_ref().map(|item| item.due))
+                                .chain(can_list.then_some(listing_due))
+                                .min();
+                            if let Some(due) = due {
                                self.await_with_control(async {
-                                    tokio::time::sleep(self.bucket_object_retry_delay(listing_attempt)).await;
+                                    tokio::time::sleep_until(due).await;
                                    Ok(())
                                })
                                .await?;
-                                continue;
                            }
-                            self.outcome.write().await.mark_untraversable();
-                            return Err(Error::HealListingFailed {
-                                bucket: bucket.to_string(),
-                                source: Box::new(error),
-                            });
                        }
+                        continue;
+                    };
+                    let retry_attempt = item.attempt;
+                    let mut telemetry_unknown = false;
+                    let object = item.name.as_str();
+                    let identity =
+                        self.outcome_identity(bucket, object, item.version_id.as_deref(), heal_opts.pool, heal_opts.set);
+                    let mut disposition = if heal_opts.dry_run {
+                        HealObjectDisposition::DryRunObserved
+                    } else {
+                        HealObjectDisposition::Unknown
+                    };
+                    let mut detail = None;
+                    {
+                        let mut progress = self.progress.write().await;
+                        progress.set_current_object(Some(format!("{bucket}/{object}")));
                    }
-                };

-                let mut pending = objects;
-                let mut retry_attempt = 0_u32;
-                while !pending.is_empty() {
-                    if retry_attempt > 0 {
-                        self.await_with_control(async {
-                            tokio::time::sleep(self.bucket_object_retry_delay(retry_attempt)).await;
-                            Ok(())
-                        })
-                        .await?;
-                    }
-                    let mut retry = Vec::with_capacity(pending.len());
-                    for item in pending {
-                        self.check_control_flags().await?;
-                        self.pace_mainline().await?;
-                        let mut telemetry_unknown = false;
-                        let object = item.name.as_str();
-                        let identity =
-                            self.outcome_identity(bucket, object, item.version_id.as_deref(), heal_opts.pool, heal_opts.set);
-                        let mut disposition = if heal_opts.dry_run {
-                            HealObjectDisposition::DryRunObserved
-                        } else {
-                            HealObjectDisposition::Unknown
-                        };
-                        let mut detail = None;
-                        {
-                            let mut progress = self.progress.write().await;
-                            progress.set_current_object(Some(format!("{bucket}/{object}")));
-                        }
-
-                        let mut terminal_outcome = true;
-                        let error = match self
+                    let mut terminal_outcome = true;
+                    let age_exhausted = item.expired();
+                    let error = if age_exhausted {
+                        Some(Error::other("heal object retry age exhausted"))
+                    } else {
+                        match self
                            .await_with_control(
                                self.storage
                                    .heal_object(bucket, object, item.version_id.as_deref(), &heal_opts),
@@ -414,152 +552,188 @@ impl HealTask {
                                None
                            }
                            Ok((_, Some(err))) | Err(err) => Some(err),
-                        };
+                        }
+                    };

-                        if let Some(err) = error {
-                            match err {
-                                Error::TaskCancelled | Error::TaskTimeout => {
-                                    let disposition = if matches!(err, Error::TaskCancelled) {
-                                        HealObjectDisposition::Cancelled
-                                    } else {
-                                        HealObjectDisposition::Deferred {
-                                            reason: HealDeferredReason::Deadline,
-                                            retry_not_before: None,
-                                        }
-                                    };
-                                    self.outcome.write().await.record(HealObjectOutcome {
-                                        identity,
-                                        disposition,
-                                        detail: None,
-                                    });
-                                    return Err(err);
-                                }
-                                _ => self.outcome.write().await.attempt_failed(),
-                            }
-                            detail = Some(err.to_string());
-                            if Self::is_dangling_delete_grace_error(&err) {
-                                disposition = HealObjectDisposition::Deferred {
-                                    reason: HealDeferredReason::DanglingDeleteGrace,
-                                    retry_not_before: None,
-                                };
-                                telemetry_unknown |= !increment_counter(&mut skipped);
-                                warn!(
-                                    target: "rustfs::heal::task",
-                                    event = EVENT_HEAL_BUCKET_RESULT,
-                                    component = LOG_COMPONENT_HEAL,
-                                    subsystem = LOG_SUBSYSTEM_TASK,
-                                    task_id = %self.id,
-                                    bucket,
-                                    object,
-                                    result = "dangling_delete_grace_skip",
-                                    error = %err,
-                                    "Heal bucket object dangling cleanup deferred by grace window"
-                                );
-                            } else if Self::should_skip_data_usage_cache_heal_error(bucket, object, &err) {
-                                disposition = HealObjectDisposition::Deferred {
-                                    reason: HealDeferredReason::TransientUsageCache,
-                                    retry_not_before: None,
-                                };
-                                telemetry_unknown |= !increment_counter(&mut skipped);
-                                warn!(
-                                    target: "rustfs::heal::task",
-                                    event = EVENT_HEAL_BUCKET_RESULT,
-                                    component = LOG_COMPONENT_HEAL,
-                                    subsystem = LOG_SUBSYSTEM_TASK,
-                                    task_id = %self.id,
-                                    bucket,
-                                    object,
-                                    result = "transient_skip",
-                                    error = %err,
-                                    "Heal bucket object repair skipped due to transient metadata error"
-                                );
-                            } else if err.is_recoverable_heal() && retry_attempt < MAX_BUCKET_OBJECT_HEAL_RETRIES {
-                                terminal_outcome = false;
-                                debug!(
-                                    target: "rustfs::heal::task",
-                                    event = EVENT_HEAL_BUCKET_RESULT,
-                                    component = LOG_COMPONENT_HEAL,
-                                    subsystem = LOG_SUBSYSTEM_TASK,
-                                    task_id = %self.id,
-                                    bucket,
-                                    object,
-                                    retry_attempt = retry_attempt.saturating_add(1),
-                                    error = %err,
-                                    result = "object_retry_scheduled",
-                                    "Heal bucket object retry scheduled"
-                                );
-                                retry.push(item);
-                            } else {
-                                disposition = HealObjectDisposition::Failed(if err.is_recoverable_heal() {
-                                    HealFailureClass::RetryExhausted
+                    if let Some(err) = error {
+                        match err {
+                            Error::TaskCancelled | Error::TaskTimeout => {
+                                let disposition = if matches!(err, Error::TaskCancelled) {
+                                    HealObjectDisposition::Cancelled
                                } else {
-                                    HealFailureClass::Permanent
+                                    HealObjectDisposition::Deferred {
+                                        reason: HealDeferredReason::Deadline,
+                                        retry_not_before: None,
+                                    }
+                                };
+                                self.outcome.write().await.record(HealObjectOutcome {
+                                    identity,
+                                    disposition,
+                                    detail: None,
                                });
-                                telemetry_unknown |= !increment_counter(&mut failed);
-                                if err.is_recoverable_heal() {
-                                    retryable_failed = retryable_failed.saturating_add(1);
-                                } else {
-                                    permanent_failed = permanent_failed.saturating_add(1);
-                                }
-                                first_failed_object.get_or_insert_with(|| object.to_string());
-                                first_error.get_or_insert_with(|| err.to_string());
-                                if take_failure_log_sample(&mut failure_samples_logged) {
-                                    warn!(
-                                        target: "rustfs::heal::task",
-                                        event = EVENT_HEAL_BUCKET_RESULT,
-                                        component = LOG_COMPONENT_HEAL,
-                                        subsystem = LOG_SUBSYSTEM_TASK,
-                                        task_id = %self.id,
-                                        bucket,
-                                        object,
-                                        retry_attempt,
-                                        error = %err,
-                                        result = "object_failed",
-                                        "Heal bucket object repair failed"
-                                    );
-                                }
+                                aborted_progress_unknown |= !increment_counter(&mut scanned);
+                                aborted_progress_unknown |= !increment_counter(&mut skipped);
+                                return Err(err);
                            }
+                            _ if !age_exhausted => self.outcome.write().await.attempt_failed(),
+                            _ => {}
                        }
-
-                        if terminal_outcome {
-                            telemetry_unknown |= !increment_counter(&mut scanned);
-                        }
-
-                        if !terminal_outcome {
-                            continue;
-                        }
-
-                        self.outcome.write().await.record(HealObjectOutcome {
-                            identity,
-                            disposition,
-                            detail,
-                        });
-
-                        let mut progress = self.progress.write().await;
-                        progress.update_object_progress(
-                            previous_progress.objects_scanned.saturating_add(scanned),
-                            previous_progress.objects_healed.saturating_add(healed),
-                            previous_progress.objects_failed.saturating_add(failed),
-                            previous_progress.skipped_objects.saturating_add(skipped),
-                            previous_progress.bytes_processed.saturating_add(bytes),
-                        );
-                        if telemetry_unknown {
-                            progress.mark_unknown();
+                        detail = Some(err.to_string());
+                        if Self::is_dangling_delete_grace_error(&err) {
+                            disposition = HealObjectDisposition::Deferred {
+                                reason: HealDeferredReason::DanglingDeleteGrace,
+                                retry_not_before: None,
+                            };
+                            telemetry_unknown |= !increment_counter(&mut skipped);
+                            warn!(
+                                target: "rustfs::heal::task",
+                                event = EVENT_HEAL_BUCKET_RESULT,
+                                component = LOG_COMPONENT_HEAL,
+                                subsystem = LOG_SUBSYSTEM_TASK,
+                                task_id = %self.id,
+                                bucket,
+                                object,
+                                result = "dangling_delete_grace_skip",
+                                error = %err,
+                                "Heal bucket object dangling cleanup deferred by grace window"
+                            );
+                        } else if Self::should_skip_data_usage_cache_heal_error(bucket, object, &err) {
+                            disposition = HealObjectDisposition::Deferred {
+                                reason: HealDeferredReason::TransientUsageCache,
+                                retry_not_before: None,
+                            };
+                            telemetry_unknown |= !increment_counter(&mut skipped);
+                            warn!(
+                                target: "rustfs::heal::task",
+                                event = EVENT_HEAL_BUCKET_RESULT,
+                                component = LOG_COMPONENT_HEAL,
+                                subsystem = LOG_SUBSYSTEM_TASK,
+                                task_id = %self.id,
+                                bucket,
+                                object,
+                                result = "transient_skip",
+                                error = %err,
+                                "Heal bucket object repair skipped due to transient metadata error"
+                            );
+                        } else if !age_exhausted && err.is_recoverable_heal() && retry_attempt < MAX_BUCKET_OBJECT_HEAL_RETRIES {
+                            terminal_outcome = false;
+                            debug!(
+                                target: "rustfs::heal::task",
+                                event = EVENT_HEAL_BUCKET_RESULT,
+                                component = LOG_COMPONENT_HEAL,
+                                subsystem = LOG_SUBSYSTEM_TASK,
+                                task_id = %self.id,
+                                bucket,
+                                object,
+                                retry_attempt = retry_attempt.saturating_add(1),
+                                error = %err,
+                                result = "object_retry_scheduled",
+                                "Heal bucket object retry scheduled"
+                            );
+                            item.defer(self.bucket_object_retry_delay(retry_attempt + 1));
+                            if let Err(item) = deferred.push(item) {
+                                inline_retry = Some(item);
+                            }
+                        } else {
+                            disposition = HealObjectDisposition::Failed(if age_exhausted || err.is_recoverable_heal() {
+                                HealFailureClass::RetryExhausted
+                            } else {
+                                HealFailureClass::Permanent
+                            });
+                            telemetry_unknown |= !increment_counter(&mut failed);
+                            if age_exhausted || err.is_recoverable_heal() {
+                                retryable_failed = retryable_failed.saturating_add(1);
+                            } else {
+                                permanent_failed = permanent_failed.saturating_add(1);
+                            }
+                            first_failed_object.get_or_insert_with(|| object.to_string());
+                            first_error.get_or_insert_with(|| err.to_string());
+                            if take_failure_log_sample(&mut failure_samples_logged) {
+                                warn!(
+                                    target: "rustfs::heal::task",
+                                    event = EVENT_HEAL_BUCKET_RESULT,
+                                    component = LOG_COMPONENT_HEAL,
+                                    subsystem = LOG_SUBSYSTEM_TASK,
+                                    task_id = %self.id,
+                                    bucket,
+                                    object,
+                                    retry_attempt,
+                                    error = %err,
+                                    result = "object_failed",
+                                    "Heal bucket object repair failed"
+                                );
+                            }
                        }
                    }
-                    pending = retry;
-                    retry_attempt = retry_attempt.saturating_add(1);
-                }

-                if !is_truncated {
-                    break;
-                }
+                    if terminal_outcome {
+                        telemetry_unknown |= !increment_counter(&mut scanned);
+                    }

-                continuation_token = next_heal_listing_token(bucket, prefix, next_token, is_truncated)?;
-                if continuation_token.is_none() {
-                    // Truncated without a continuation token is a compatibility EOF.
-                    break;
+                    if !terminal_outcome {
+                        continue;
+                    }
+
+                    self.outcome.write().await.record(HealObjectOutcome {
+                        identity,
+                        disposition,
+                        detail,
+                    });
+
+                    let mut progress = self.progress.write().await;
+                    progress.update_object_progress(
+                        previous_progress.objects_scanned.saturating_add(scanned),
+                        previous_progress.objects_healed.saturating_add(healed),
+                        previous_progress.objects_failed.saturating_add(failed),
+                        previous_progress.skipped_objects.saturating_add(skipped),
+                        previous_progress.bytes_processed.saturating_add(bytes),
+                    );
+                    if telemetry_unknown {
+                        progress.mark_unknown();
+                    }
                }
+                Ok(())
+            }
+            .await;
+            if let Err(error) = scope_result {
+                let disposition = match error {
+                    Error::TaskCancelled => HealObjectDisposition::Cancelled,
+                    Error::TaskTimeout => HealObjectDisposition::Deferred {
+                        reason: HealDeferredReason::Deadline,
+                        retry_not_before: None,
+                    },
+                    _ => HealObjectDisposition::Unknown,
+                };
+                // Only attempted identities have terminal outcomes. Unstarted
+                // page tails remain unprocessed under the task's partial coverage.
+                // No detached sleepers survive abort.
+                for item in deferred.objects.into_iter().chain(inline_retry) {
+                    self.outcome.write().await.record(HealObjectOutcome {
+                        identity: self.outcome_identity(
+                            bucket,
+                            &item.name,
+                            item.version_id.as_deref(),
+                            heal_opts.pool,
+                            heal_opts.set,
+                        ),
+                        disposition: disposition.clone(),
+                        detail: None,
+                    });
+                    aborted_progress_unknown |= !increment_counter(&mut scanned);
+                    aborted_progress_unknown |= !increment_counter(&mut skipped);
+                }
+                let mut progress = self.progress.write().await;
+                progress.update_object_progress(
+                    previous_progress.objects_scanned.saturating_add(scanned),
+                    previous_progress.objects_healed.saturating_add(healed),
+                    previous_progress.objects_failed.saturating_add(failed),
+                    previous_progress.skipped_objects.saturating_add(skipped),
+                    previous_progress.bytes_processed.saturating_add(bytes),
+                );
+                if aborted_progress_unknown {
+                    progress.mark_unknown();
+                }
+                return Err(error);
            }
        }

@@ -42,7 +42,36 @@ impl HealTask {
            progress.update_stage(0, 4);
        }

-        let is_auto_replacement = matches!(self.source, HealRequestSource::AutoHeal) && !self.heal_endpoints.is_empty();
+        let mut is_auto_replacement = matches!(self.source, HealRequestSource::AutoHeal) && !self.heal_endpoints.is_empty();
+        if is_auto_replacement
+            && crate::heal::replacement_readiness::directory_backed_replacement_fallback_enabled()
+            && self
+                .await_with_control(self.storage.replacement_target_identities(&self.heal_endpoints))
+                .await
+                .is_err()
+        {
+            // Directory-backed endpoints cannot pass replacement admission; the
+            // operator opted out of disk checks, so heal the set the way the
+            // pre-admission `heal_disk` path did instead of deferring forever.
+            warn!(
+                target: "rustfs::heal::task",
+                event = EVENT_HEAL_ERASURE_SET_STAGE,
+                component = LOG_COMPONENT_HEAL,
+                subsystem = LOG_SUBSYSTEM_TASK,
+                task_id = %self.id,
+                set_disk_id,
+                stage = "replacement_admission",
+                result = "directory_backed_fallback",
+                target_count = self.heal_endpoints.len(),
+                "Heal erasure set falls back to set-wide format heal for a replacement target that is not an independently mounted disk"
+            );
+            is_auto_replacement = false;
+        }
+        let replacement_targets = if is_auto_replacement {
+            self.heal_endpoints.clone()
+        } else {
+            Vec::new()
+        };
        let replacement_resume_disk = if is_auto_replacement {
            let mut requested_targets = self.heal_endpoints.clone();
            requested_targets.sort_unstable();
@@ -421,7 +450,7 @@ impl HealTask {
            heal_opts,
            self.source,
        )
-        .with_replacement_targets(self.heal_endpoints.clone(), is_auto_replacement.then(|| self.id.clone()))
+        .with_replacement_targets(replacement_targets, is_auto_replacement.then(|| self.id.clone()))
        .with_replacement_identity_fence(replacement_target_identities.clone())
        .with_mainline_pacer(self.mainline_pacer.clone());

@@ -15,6 +15,8 @@
 use super::super::{DiskOption, DiskStore, Endpoint, new_disk};
 use super::*;

+mod deferred_retry;
+
 mod canonical_outcome {
    use super::*;
    use crate::heal::outcome::{HealExecutionOutcome, HealTraversalCoverage};
@@ -478,6 +480,95 @@ async fn automatic_replacement_uses_target_scoped_format() {
    );
 }

+fn directory_backed_replacement_request() -> HealRequest {
+    let mut request = HealRequest::new(
+        HealType::ErasureSet {
+            buckets: Vec::new(),
+            set_disk_id: "pool_0_set_0".to_string(),
+        },
+        HealOptions {
+            pool_index: Some(0),
+            set_index: Some(0),
+            ..Default::default()
+        },
+        HealPriority::Low,
+    );
+    request.source = HealRequestSource::AutoHeal;
+    request.heal_endpoints = vec!["/data/disk0".to_string()];
+    request
+}
+
+#[tokio::test]
+async fn directory_backed_replacement_falls_back_to_set_format_when_disk_checks_are_bypassed() {
+    temp_env::async_with_vars(
+        [
+            (rustfs_config::ENV_UNSAFE_BYPASS_DISK_CHECK, Some("true")),
+            (rustfs_config::ENV_MINIO_CI, None::<&str>),
+        ],
+        async {
+            let storage = Arc::new(MockStorage {
+                replacement_target_identities_ready: Mutex::new(false),
+                global_format_ok_endpoints: Mutex::new(vec!["/data/disk0".to_string()]),
+                ..Default::default()
+            });
+            let task = HealTask::from_request(directory_backed_replacement_request(), storage.clone());
+
+            // The mock has no local disk behind "/data/disk0", so the run stops at
+            // the healing-marker step that follows the format stage, exactly like
+            // `automatic_replacement_uses_target_scoped_format`. The assertions
+            // below pin which format path ran before that point.
+            let err = task.execute().await.expect_err("the mock has no local healing marker target");
+            assert!(
+                err.to_string().contains("healing marker target is unavailable"),
+                "the fallback must reach the post-format marker step, got: {err}"
+            );
+
+            assert_eq!(
+                *storage.global_format_calls.lock().unwrap(),
+                1,
+                "the fallback must run exactly one set-wide format heal"
+            );
+            assert!(
+                storage.replacement_format_calls.lock().unwrap().is_empty(),
+                "the fallback must not run the target-scoped replacement format"
+            );
+        },
+    )
+    .await;
+}
+
+#[tokio::test]
+async fn directory_backed_replacement_stays_fail_closed_without_disk_check_bypass() {
+    temp_env::async_with_vars(
+        [
+            (rustfs_config::ENV_UNSAFE_BYPASS_DISK_CHECK, None::<&str>),
+            (rustfs_config::ENV_MINIO_CI, None::<&str>),
+        ],
+        async {
+            let storage = Arc::new(MockStorage {
+                replacement_target_identities_ready: Mutex::new(false),
+                ..Default::default()
+            });
+            let task = HealTask::from_request(directory_backed_replacement_request(), storage.clone());
+
+            task.execute()
+                .await
+                .expect_err("an inadmissible replacement target must keep failing closed");
+
+            assert_eq!(
+                *storage.global_format_calls.lock().unwrap(),
+                0,
+                "fail-closed admission must not format the set"
+            );
+            assert!(
+                storage.replacement_format_calls.lock().unwrap().is_empty(),
+                "fail-closed admission must not format the target"
+            );
+        },
+    )
+    .await;
+}
+
 #[tokio::test]
 async fn automatic_replacement_persists_intent_before_format() {
    let storage = Arc::new(MockStorage {
@@ -939,6 +1030,10 @@ async fn verified_recovery_keeps_state_when_marker_clear_fails() {

 #[derive(Default)]
 struct MockStorage {
+    retry_test_pages: Option<Vec<Vec<HealListItem>>>,
+    retry_test_delays: HashMap<String, Duration>,
+    retry_test_listing_delays: Mutex<VecDeque<Duration>>,
+    retry_test_events: Mutex<Vec<String>>,
    listed: Mutex<bool>,
    list_each_bucket: bool,
    fail_second_listing_page: bool,
@@ -956,6 +1051,8 @@ struct MockStorage {
    format_no_heal_required: Mutex<bool>,
    format_error: Mutex<Option<Error>>,
    global_format_calls: Mutex<u32>,
+    /// Endpoints the set-wide format mock reports as freshly formatted (`state == "ok"`).
+    global_format_ok_endpoints: Mutex<Vec<String>>,
    replacement_format_calls: Mutex<Vec<(usize, usize, Vec<String>)>>,
    replacement_target_identities_ready: Mutex<bool>,
    replacement_target_identity_sequences: Mutex<VecDeque<Vec<crate::heal::resume::ReplacementTargetIdentity>>>,
@@ -1109,6 +1206,7 @@ fn replacement_identity(
 }

 enum MockHealObjectOutcome {
+    RetryableLock,
    OkWithOtherError(&'static str),
    ErrOther(&'static str),
    DanglingGraceDeferred,
@@ -1213,6 +1311,10 @@ impl HealStorageAPI for MockStorage {
        opts: &HealOpts,
    ) -> Result<(HealResultItem, Option<Error>)> {
        self.heal_object_calls.lock().unwrap().push(object.to_string());
+        self.retry_test_events.lock().expect("events").push(format!("heal:{object}"));
+        if let Some(delay) = self.retry_test_delays.get(object) {
+            tokio::time::sleep(*delay).await;
+        }
        self.heal_object_version_ids
            .lock()
            .unwrap()
@@ -1241,6 +1343,13 @@ impl HealStorageAPI for MockStorage {
                    bucket.to_string(),
                    object.to_string(),
                ))),
+                MockHealObjectOutcome::RetryableLock => Ok((
+                    HealResultItem::default(),
+                    Some(Error::Storage(EcstoreError::Lock(rustfs_lock::LockError::AlreadyLocked {
+                        resource: object.to_string(),
+                        owner: "competing-writer".to_string(),
+                    }))),
+                )),
                MockHealObjectOutcome::RetryableSlowDown => {
                    Ok((HealResultItem::default(), Some(Error::Storage(EcstoreError::SlowDown))))
                }
@@ -1266,6 +1375,13 @@ impl HealStorageAPI for MockStorage {
                    bucket.to_string(),
                    object.to_string(),
                ))),
+                MockHealObjectOutcome::RetryableLock => Ok((
+                    HealResultItem::default(),
+                    Some(Error::Storage(EcstoreError::Lock(rustfs_lock::LockError::AlreadyLocked {
+                        resource: object.to_string(),
+                        owner: "competing-writer".to_string(),
+                    }))),
+                )),
                MockHealObjectOutcome::RetryableSlowDown => {
                    Ok((HealResultItem::default(), Some(Error::Storage(EcstoreError::SlowDown))))
                }
@@ -1313,10 +1429,26 @@ impl HealStorageAPI for MockStorage {
            return Err(error);
        }
        let no_heal_required = *self.format_no_heal_required.lock().unwrap();
+        let result = HealResultItem {
+            after: Infos {
+                drives: self
+                    .global_format_ok_endpoints
+                    .lock()
+                    .unwrap()
+                    .iter()
+                    .map(|endpoint| HealDriveInfo {
+                        endpoint: endpoint.clone(),
+                        state: "ok".to_string(),
+                        ..Default::default()
+                    })
+                    .collect(),
+            },
+            ..Default::default()
+        };
        if no_heal_required {
-            Ok((HealResultItem::default(), Some(Error::Storage(EcstoreError::NoHealRequired))))
+            Ok((result, Some(Error::Storage(EcstoreError::NoHealRequired))))
        } else {
-            Ok((HealResultItem::default(), None))
+            Ok((result, None))
        }
    }

@@ -1361,6 +1493,19 @@ impl HealStorageAPI for MockStorage {
            .lock()
            .expect("listing tokens")
            .push(continuation_token.map(ToOwned::to_owned));
+        self.retry_test_events
+            .lock()
+            .expect("events")
+            .push(format!("list:{}", continuation_token.unwrap_or("first")));
+        let delay = self.retry_test_listing_delays.lock().expect("listing delays").pop_front();
+        if let Some(delay) = delay {
+            tokio::time::sleep(delay).await;
+        }
+        if let Some(pages) = &self.retry_test_pages {
+            let page = continuation_token.map_or(0, |token| token.parse::<usize>().expect("test page token"));
+            let next = (page + 1 < pages.len()).then(|| (page + 1).to_string());
+            return Ok((pages[page].clone(), next.clone(), next.is_some()));
+        }
        if let Some(remaining) = self
            .recoverable_second_page_failures
            .lock()
@@ -0,0 +1,487 @@
+// Copyright 2024 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use super::*;
+
+fn bucket_task(storage: Arc<MockStorage>) -> HealTask {
+    HealTask::from_request(
+        HealRequest::new(
+            HealType::Bucket {
+                bucket: "bucket-a".to_string(),
+            },
+            HealOptions {
+                recursive: true,
+                timeout: None,
+                ..Default::default()
+            },
+            HealPriority::Normal,
+        ),
+        storage,
+    )
+}
+
+fn pages_storage(pages: &[&[&str]]) -> MockStorage {
+    MockStorage {
+        retry_test_pages: Some(
+            pages
+                .iter()
+                .map(|page| page.iter().map(|name| heal_item(name)).collect())
+                .collect(),
+        ),
+        ..Default::default()
+    }
+}
+
+fn fail_once(storage: &MockStorage, name: &str) {
+    storage
+        .heal_object_outcomes
+        .lock()
+        .expect("outcomes")
+        .insert(name.to_string(), VecDeque::from([MockHealObjectOutcome::RetryableLock]));
+}
+
+#[tokio::test(start_paused = true)]
+async fn slow_listing_retry_services_due_object_then_age_before_next_listing() {
+    let storage = Arc::new(MockStorage {
+        recoverable_second_page_failures: Mutex::new(Some(1)),
+        retry_test_listing_delays: Mutex::new(VecDeque::from([Duration::ZERO, Duration::from_secs(29)])),
+        ..Default::default()
+    });
+    storage.heal_object_outcomes.lock().expect("outcomes").insert(
+        "object-a".to_string(),
+        VecDeque::from([
+            MockHealObjectOutcome::RetryableSlowDown,
+            MockHealObjectOutcome::RetryableSlowDown,
+        ]),
+    );
+    let task = bucket_task(storage.clone());
+    let execution = task.execute();
+    tokio::pin!(execution);
+    assert!(
+        tokio::time::timeout(Duration::from_millis(30_500), &mut execution)
+            .await
+            .is_err()
+    );
+    assert_eq!(
+        storage.retry_test_events.lock().expect("events").as_slice(),
+        ["list:first", "heal:object-a", "list:second", "heal:object-a"]
+    );
+    let outcome = task.get_outcome().await;
+    assert_eq!(outcome.counters.processed, 1);
+    assert_eq!(
+        outcome.objects[0].disposition,
+        HealObjectDisposition::Failed(HealFailureClass::RetryExhausted)
+    );
+    execution.await.expect_err("age exhausted object must remain a batch failure");
+    assert_eq!(
+        storage.retry_test_events.lock().expect("events").as_slice(),
+        [
+            "list:first",
+            "heal:object-a",
+            "list:second",
+            "heal:object-a",
+            "list:second",
+            "heal:object-b"
+        ]
+    );
+    let outcome = task.get_outcome().await;
+    assert_eq!((outcome.counters.processed, outcome.counters.attempt_failures), (2, 3));
+}
+
+#[tokio::test(start_paused = true)]
+async fn listing_return_after_age_expires_does_not_start_another_heal_attempt() {
+    let storage = Arc::new(MockStorage {
+        recoverable_second_page_failures: Mutex::new(Some(1)),
+        retry_test_listing_delays: Mutex::new(VecDeque::from([Duration::ZERO, Duration::from_secs(31)])),
+        ..Default::default()
+    });
+    fail_once(&storage, "object-a");
+    let task = bucket_task(storage.clone());
+    let execution = task.execute();
+    tokio::pin!(execution);
+    assert!(
+        tokio::time::timeout(Duration::from_millis(31_500), &mut execution)
+            .await
+            .is_err()
+    );
+    assert_eq!(
+        storage.retry_test_events.lock().expect("events").as_slice(),
+        ["list:first", "heal:object-a", "list:second"]
+    );
+    assert_eq!(task.get_outcome().await.counters.failed, 1);
+    execution.await.expect_err("age exhaustion remains a failure");
+    assert_eq!(
+        storage.retry_test_events.lock().expect("events").as_slice(),
+        ["list:first", "heal:object-a", "list:second", "list:second", "heal:object-b"]
+    );
+}
+
+#[tokio::test(start_paused = true)]
+async fn full_window_abort_accounts_inline_once_and_leaves_unstarted_tail_unprocessed() {
+    for cancel in [true, false] {
+        let names: Vec<String> = (0..258).map(|index| format!("blocked-{index}")).collect();
+        let mut page: Vec<HealListItem> = names.iter().map(|name| heal_item(name)).collect();
+        page[256].version_id = Some("inline-version".to_string());
+        let storage = Arc::new(MockStorage {
+            retry_test_pages: Some(vec![page, vec![heal_item("healthy")]]),
+            ..Default::default()
+        });
+        for name in &names {
+            fail_once(&storage, name);
+        }
+        let mut task = bucket_task(storage.clone());
+        if !cancel {
+            task.options.timeout = Some(Duration::from_secs(1));
+        }
+        let execution = task.execute();
+        tokio::pin!(execution);
+        assert!(
+            tokio::time::timeout(Duration::from_millis(500), &mut execution)
+                .await
+                .is_err()
+        );
+        assert_eq!(storage.heal_object_calls.lock().expect("calls").len(), 257);
+        assert_eq!(storage.listing_tokens.lock().expect("tokens").len(), 1);
+        if cancel {
+            task.cancel().await.expect("cancel");
+        }
+        let result = execution.await;
+        assert!(matches!(
+            (&result, cancel),
+            (Err(Error::TaskCancelled), true) | (Err(Error::TaskTimeout), false)
+        ));
+        let outcome = task.get_outcome().await;
+        assert_eq!(
+            (
+                outcome.counters.processed,
+                outcome.counters.skipped,
+                outcome.counters.failed,
+                outcome.counters.healed
+            ),
+            (257, 257, 0, 0)
+        );
+        assert_eq!(outcome.coverage, crate::heal::outcome::HealTraversalCoverage::Partial);
+        assert_eq!(
+            outcome.execution,
+            crate::heal::outcome::HealExecutionOutcome::Aborted(if cancel {
+                HealAbortReason::Cancelled
+            } else {
+                HealAbortReason::Deadline
+            })
+        );
+        let inline: Vec<_> = outcome
+            .objects
+            .iter()
+            .filter(|item| item.identity.object == "blocked-256")
+            .collect();
+        assert_eq!(inline.len(), 1);
+        assert_eq!(inline[0].identity.version_id.as_deref(), Some("inline-version"));
+        assert_eq!(
+            inline[0].disposition,
+            if cancel {
+                HealObjectDisposition::Cancelled
+            } else {
+                HealObjectDisposition::Deferred {
+                    reason: HealDeferredReason::Deadline,
+                    retry_not_before: None,
+                }
+            }
+        );
+        let progress = task.get_progress().await;
+        assert_eq!(
+            (
+                progress.objects_scanned,
+                progress.skipped_objects,
+                progress.objects_failed,
+                progress.objects_healed
+            ),
+            (257, 257, 0, 0)
+        );
+        assert!(
+            !outcome
+                .objects
+                .iter()
+                .any(|item| item.identity.object == "blocked-257" || item.identity.object == "healthy")
+        );
+        tokio::time::advance(Duration::from_secs(60)).await;
+        assert_eq!(storage.heal_object_calls.lock().expect("calls").len(), 257);
+        assert_eq!(storage.listing_tokens.lock().expect("tokens").len(), 1);
+    }
+}
+
+#[tokio::test(start_paused = true)]
+async fn repeated_slowdown_keeps_attempts_and_forward_pages_bounded() {
+    let storage = Arc::new(pages_storage(&[&["a"], &["b"], &["c"], &["d"]]));
+    for name in ["a", "b", "c", "d"] {
+        storage
+            .heal_object_outcomes
+            .lock()
+            .expect("outcomes")
+            .insert(name.to_string(), (0..4).map(|_| MockHealObjectOutcome::RetryableSlowDown).collect());
+    }
+    let task = bucket_task(storage.clone());
+    let execution = task.execute();
+    tokio::pin!(execution);
+    assert!(tokio::time::timeout(Duration::from_secs(1), &mut execution).await.is_err());
+    assert_eq!(storage.listing_tokens.lock().expect("tokens").len(), 3);
+    execution.await.expect_err("all four objects exhaust retries");
+    let outcome = task.get_outcome().await;
+    assert_eq!(
+        (outcome.counters.processed, outcome.counters.failed, outcome.counters.attempt_failures),
+        (4, 4, 16)
+    );
+    assert_eq!(storage.heal_object_calls.lock().expect("calls").len(), 16);
+    for name in ["a", "b", "c", "d"] {
+        assert_eq!(outcome.objects.iter().filter(|item| item.identity.object == name).count(), 1);
+    }
+}
+
+#[tokio::test(start_paused = true)]
+async fn listing_retry_keeps_cursor_and_does_not_replay_successful_objects() {
+    let storage = Arc::new(MockStorage {
+        recoverable_second_page_failures: Mutex::new(Some(1)),
+        ..Default::default()
+    });
+    fail_once(&storage, "object-a");
+    let task = bucket_task(storage.clone());
+    task.execute().await.expect("both retries complete");
+    assert_eq!(
+        storage.listing_tokens.lock().expect("tokens").as_slice(),
+        [None, Some("second".to_string()), Some("second".to_string())]
+    );
+    assert_eq!(
+        storage.heal_object_calls.lock().expect("calls").as_slice(),
+        ["object-a", "object-a", "object-b"]
+    );
+    let outcome = task.get_outcome().await;
+    assert_eq!((outcome.counters.processed, outcome.counters.attempt_failures), (2, 2));
+}
+
+#[tokio::test(start_paused = true)]
+async fn typed_lock_contention_allows_only_two_forward_pages() {
+    let storage = Arc::new(pages_storage(&[&["a"], &["b"], &["c"], &["d"]]));
+    fail_once(&storage, "a");
+    let task = bucket_task(storage.clone());
+    let execution = task.execute();
+    tokio::pin!(execution);
+    assert!(tokio::time::timeout(Duration::from_secs(1), &mut execution).await.is_err());
+    assert_eq!(storage.heal_object_calls.lock().expect("calls").as_slice(), ["a", "b", "c"]);
+    assert_eq!(storage.listing_tokens.lock().expect("tokens").len(), 3);
+    execution.await.expect("all objects complete");
+    assert_eq!(storage.heal_object_calls.lock().expect("calls").as_slice(), ["a", "b", "c", "a", "d"]);
+    assert_eq!(task.get_outcome().await.counters.processed, 4);
+}
+
+#[tokio::test(start_paused = true)]
+async fn due_retry_runs_before_next_object_in_a_slow_healthy_page() {
+    let mut storage = pages_storage(&[&["a"], &["b", "c"]]);
+    storage.retry_test_delays.insert("b".to_string(), Duration::from_secs(3));
+    fail_once(&storage, "a");
+    let storage = Arc::new(storage);
+    bucket_task(storage.clone()).execute().await.expect("all objects complete");
+    assert_eq!(storage.heal_object_calls.lock().expect("calls").as_slice(), ["a", "b", "a", "c"]);
+}
+
+#[tokio::test(start_paused = true)]
+async fn expired_retry_is_terminal_without_an_extra_storage_attempt() {
+    let mut storage = pages_storage(&[&["a"], &["b"]]);
+    storage.retry_test_delays.insert("b".to_string(), Duration::from_secs(31));
+    fail_once(&storage, "a");
+    let storage = Arc::new(storage);
+    let task = bucket_task(storage.clone());
+    task.execute().await.expect_err("aged pending responsibility is not success");
+    assert_eq!(storage.heal_object_calls.lock().expect("calls").as_slice(), ["a", "b"]);
+    let outcome = task.get_outcome().await;
+    assert_eq!(outcome.counters.processed, 2);
+    assert_eq!(outcome.counters.attempt_failures, 1);
+    assert_eq!(outcome.counters.failed, 1);
+    assert_eq!(
+        outcome
+            .objects
+            .iter()
+            .find(|item| item.identity.object == "a")
+            .expect("a outcome")
+            .disposition,
+        HealObjectDisposition::Failed(HealFailureClass::RetryExhausted)
+    );
+}
+
+#[tokio::test(start_paused = true)]
+async fn cancellation_drains_owned_retries_once() {
+    let storage = Arc::new(pages_storage(&[&["a"], &["b"]]));
+    fail_once(&storage, "a");
+    let task = bucket_task(storage.clone());
+    let execution = task.execute();
+    tokio::pin!(execution);
+    assert!(tokio::time::timeout(Duration::from_secs(1), &mut execution).await.is_err());
+    task.cancel().await.expect("cancel");
+    assert!(matches!(execution.await, Err(Error::TaskCancelled)));
+    tokio::time::advance(Duration::from_secs(60)).await;
+    assert_eq!(storage.heal_object_calls.lock().expect("calls").as_slice(), ["a", "b"]);
+    let outcome = task.get_outcome().await;
+    assert_eq!(outcome.counters.processed, 2);
+    assert_eq!(outcome.objects.iter().filter(|item| item.identity.object == "a").count(), 1);
+    assert_eq!(
+        outcome
+            .objects
+            .iter()
+            .find(|item| item.identity.object == "a")
+            .expect("a")
+            .disposition,
+        HealObjectDisposition::Cancelled
+    );
+}
+
+#[tokio::test(start_paused = true)]
+async fn deadline_drains_owned_retries_without_false_completion() {
+    let storage = Arc::new(pages_storage(&[&["a"], &["b"]]));
+    fail_once(&storage, "a");
+    let mut task = bucket_task(storage.clone());
+    task.options.timeout = Some(Duration::from_secs(1));
+    assert!(matches!(task.execute().await, Err(Error::TaskTimeout)));
+    let outcome = task.get_outcome().await;
+    assert_eq!(outcome.counters.processed, 2);
+    assert_eq!(
+        outcome
+            .objects
+            .iter()
+            .find(|item| item.identity.object == "a")
+            .expect("a")
+            .disposition,
+        HealObjectDisposition::Deferred {
+            reason: HealDeferredReason::Deadline,
+            retry_not_before: None
+        }
+    );
+    tokio::time::advance(Duration::from_secs(60)).await;
+    assert_eq!(storage.heal_object_calls.lock().expect("calls").as_slice(), ["a", "b"]);
+}
+
+#[tokio::test(start_paused = true)]
+async fn full_window_backpressures_without_losing_the_current_page_tail() {
+    let names: Vec<String> = (0..258).map(|index| format!("blocked-{index}")).collect();
+    let mut storage = MockStorage {
+        retry_test_pages: Some(vec![names.iter().map(|name| heal_item(name)).collect(), vec![heal_item("healthy")]]),
+        ..Default::default()
+    };
+    for name in &names {
+        fail_once(&storage, name);
+    }
+    // The last item has a version, proving the current-page tail is not rebuilt
+    // from names alone when the window fills.
+    storage.retry_test_pages.as_mut().expect("pages")[0][257].version_id = Some("version-tail".to_string());
+    let storage = Arc::new(storage);
+    let task = bucket_task(storage.clone());
+    let execution = task.execute();
+    tokio::pin!(execution);
+    assert!(tokio::time::timeout(Duration::from_secs(1), &mut execution).await.is_err());
+    assert_eq!(storage.heal_object_calls.lock().expect("calls").len(), 257);
+    assert_eq!(storage.listing_tokens.lock().expect("tokens").len(), 1);
+    execution.await.expect("every owned item eventually completes");
+    assert_eq!(task.get_outcome().await.counters.processed, 259);
+    let calls = storage.heal_object_calls.lock().expect("calls");
+    for name in &names {
+        assert_eq!(calls.iter().filter(|called| *called == name).count(), 2);
+    }
+    let versions = storage.heal_object_version_ids.lock().expect("versions");
+    for (name, version) in calls.iter().zip(versions.iter()) {
+        if name == "blocked-257" {
+            assert_eq!(version.as_deref(), Some("version-tail"));
+        }
+    }
+}
+
+#[tokio::test(start_paused = true)]
+async fn oversized_identity_stays_inline_without_losing_version() {
+    let name = "k".repeat(256 * 1024);
+    let mut item = heal_item(&name);
+    item.version_id = Some("v".repeat(1024));
+    let storage = Arc::new(MockStorage {
+        retry_test_pages: Some(vec![vec![item], vec![heal_item("healthy")]]),
+        ..Default::default()
+    });
+    fail_once(&storage, &name);
+    let task = bucket_task(storage.clone());
+    let execution = task.execute();
+    tokio::pin!(execution);
+    assert!(tokio::time::timeout(Duration::from_secs(1), &mut execution).await.is_err());
+    assert_eq!(storage.listing_tokens.lock().expect("tokens").len(), 1);
+    execution.await.expect("oversized identity retries inline");
+    assert_eq!(task.get_outcome().await.counters.processed, 2);
+    let versions = storage.heal_object_version_ids.lock().expect("versions");
+    assert_eq!(versions[0], versions[1]);
+    assert_eq!(versions[0].as_ref().expect("version").len(), 1024);
+}
+
+#[tokio::test(start_paused = true)]
+async fn terminal_listing_failure_keeps_deferred_identity_unknown() {
+    let storage = Arc::new(MockStorage {
+        fail_second_listing_page: true,
+        ..Default::default()
+    });
+    fail_once(&storage, "object-a");
+    let task = bucket_task(storage.clone());
+    task.execute().await.expect_err("listing cannot continue");
+    let outcome = task.get_outcome().await;
+    assert_eq!(outcome.counters.processed, 1);
+    assert_eq!(outcome.objects[0].disposition, HealObjectDisposition::Unknown);
+    assert_eq!(outcome.coverage, crate::heal::outcome::HealTraversalCoverage::Partial);
+    assert_eq!(storage.heal_object_calls.lock().expect("calls").as_slice(), ["object-a"]);
+}
+
+#[tokio::test(start_paused = true)]
+async fn healthy_second_page_advances_before_first_retry_is_due() {
+    let storage = Arc::new(MockStorage {
+        recoverable_second_page_failures: Mutex::new(Some(0)),
+        ..Default::default()
+    });
+    storage
+        .heal_object_outcomes
+        .lock()
+        .expect("outcomes")
+        .insert("object-a".to_string(), VecDeque::from([MockHealObjectOutcome::RetryableSlowDown]));
+    let task = HealTask::from_request(
+        HealRequest::new(
+            HealType::Bucket {
+                bucket: "bucket-a".to_string(),
+            },
+            HealOptions {
+                recursive: true,
+                timeout: None,
+                ..Default::default()
+            },
+            HealPriority::Normal,
+        ),
+        storage.clone(),
+    );
+    let execution = task.execute();
+    tokio::pin!(execution);
+    assert!(
+        tokio::time::timeout(Duration::from_secs(1), &mut execution).await.is_err(),
+        "the deferred first object must remain pending before its retry is due"
+    );
+    assert_eq!(
+        storage.heal_object_calls.lock().expect("calls").as_slice(),
+        ["object-a", "object-b"],
+        "a retryable page head must not hold the healthy second page behind its backoff"
+    );
+    execution.await.expect("retry eventually succeeds");
+    let outcome = task.get_outcome().await;
+    assert_eq!(outcome.counters.processed, 2);
+    assert_eq!(outcome.counters.attempt_failures, 1);
+    assert_eq!(
+        storage.heal_object_calls.lock().expect("calls").as_slice(),
+        ["object-a", "object-b", "object-a"]
+    );
+}
@@ -0,0 +1,76 @@
+// Copyright 2024 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use super::*;
+
+fn item(name: String, version_id: Option<String>) -> DeferredObject {
+    DeferredObject::new(
+        HealListItem {
+            name,
+            version_id,
+            mod_time_unix_nanos: None,
+            lifecycle_object_info: None,
+            is_delete_marker: false,
+        },
+        1,
+    )
+}
+
+#[tokio::test(start_paused = true)]
+async fn count_cap_and_next_item_preserve_ownership() {
+    let mut window = DeferredWindow::default();
+    for _ in 0..MAX_DEFERRED_OBJECTS {
+        assert!(window.push(item("key".to_string(), None)).is_ok());
+    }
+    let rejected = window
+        .push(item("next".to_string(), Some("version".to_string())))
+        .expect_err("count cap");
+    assert_eq!(rejected.name, "next");
+    assert_eq!(rejected.version_id.as_deref(), Some("version"));
+    assert_eq!(window.objects.len(), MAX_DEFERRED_OBJECTS);
+    assert!(window.bytes <= MAX_DEFERRED_BYTES);
+    assert!(!window.can_advance(1));
+    assert!(window.pop_due().is_some());
+    assert!(window.push(rejected).is_ok());
+}
+
+#[tokio::test(start_paused = true)]
+async fn byte_cap_counts_key_version_and_reserved_slots() {
+    let mut window = DeferredWindow::default();
+    let available = MAX_DEFERRED_BYTES - window.bytes;
+    let key = "k".repeat(available / 2);
+    let version = "v".repeat(available - key.capacity());
+    assert_eq!(key.capacity() + version.capacity(), available);
+    assert!(window.push(item(key, Some(version))).is_ok());
+    assert_eq!(window.bytes, MAX_DEFERRED_BYTES);
+    assert!(!window.can_advance(1));
+    assert!(window.push(item("x".to_string(), None)).is_err());
+    assert!(window.pop_due().is_some());
+    assert_eq!(window.bytes, MAX_DEFERRED_OBJECTS * size_of::<DeferredObject>());
+    assert!(window.push(item("x".to_string(), None)).is_ok());
+}
+
+#[tokio::test(start_paused = true)]
+async fn retry_age_caps_due_time_and_is_not_reset_by_rescheduling() {
+    let mut entry = item("a".to_string(), None);
+    entry.defer(Duration::from_secs(2));
+    let first = entry.first_failure.expect("first failure");
+    tokio::time::advance(Duration::from_secs(29)).await;
+    entry.defer(Duration::from_secs(8));
+    assert_eq!(entry.first_failure, Some(first));
+    assert_eq!(entry.due, first + MAX_DEFERRED_AGE);
+    assert!(!entry.expired());
+    tokio::time::advance(Duration::from_secs(1)).await;
+    assert!(entry.expired());
+}
@@ -66,9 +66,8 @@ const ERR_LIFECYCLE_EXPIRED_OBJECT_DELETE_MARKER_WITH_TAGS: &str =
 const ERR_LIFECYCLE_RULE_MUST_HAVE_ACTION: &str = "Rule must have at least one of Expiration, Transition, NoncurrentVersionExpiration, NoncurrentVersionTransition, or DelMarkerExpiration";
 const ERR_LIFECYCLE_PREFIX_FILTER_CONFLICT: &str = "Legacy Prefix and Filter cannot both be present in a lifecycle rule. Use Filter.Prefix instead of the top-level Prefix element.";
 const ERR_LIFECYCLE_INVALID_NEWER_NONCURRENT_VERSIONS: &str = "'NewerNoncurrentVersions' must be a non-negative integer";
-const ERR_LIFECYCLE_FILTER_TOO_MANY_PREDICATES: &str =
-    "Filter must have at most one of Prefix, Tag, ObjectSizeGreaterThan, ObjectSizeLessThan or And; combine predicates with And";
 const ERR_LIFECYCLE_FILTER_AND_TOO_FEW_PREDICATES: &str = "Filter And must contain at least two predicates";
+const ERR_LIFECYCLE_FILTER_TOO_MANY_PREDICATES: &str = "Filter has too many predicates";
 const ERR_LIFECYCLE_FILTER_DUPLICATE_TAG_KEY: &str = "Filter must not repeat a tag key";
 const ERR_LIFECYCLE_FILTER_INVALID_TAG: &str = "Tag key must be 1-128 characters and tag value must be at most 256 characters";
 const ERR_LIFECYCLE_FILTER_NEGATIVE_SIZE: &str = "ObjectSizeGreaterThan and ObjectSizeLessThan must not be negative";
@@ -167,6 +167,13 @@ pub struct HealTaskStatus {
    /// Live progress snapshot; the exact shape is owned by the heal runtime.
    #[serde(default)]
    pub progress: Option<serde_json::Value>,
+    /// Canonical heal-owner result. Missing or future states are not repair proof.
+    #[serde(default)]
+    pub outcome: Option<serde_json::Value>,
+    #[serde(default, alias = "next_seq")]
+    pub next_seq: Option<u64>,
+    #[serde(default, alias = "min_seq")]
+    pub min_seq: Option<u64>,
 }

 /// `POST /v3/background-heal/status` response. Known top-level fields are
@@ -223,6 +230,40 @@ pub struct ScannerStatus {
    pub extra: serde_json::Map<String, serde_json::Value>,
 }

+/// `POST /v3/scanner/cycle-state/reset` response for the legacy synchronous
+/// full-rescan reset path.
+#[derive(Debug, Clone, Deserialize)]
+pub struct ScannerCycleResetResponse {
+    pub status: String,
+    pub mode: String,
+    #[serde(flatten)]
+    pub extra: serde_json::Map<String, serde_json::Value>,
+}
+
+/// `POST /v3/scanner/usage-state/reset` response for the legacy synchronous
+/// full-rebuild reset path.
+#[derive(Debug, Clone, Deserialize)]
+pub struct ScannerUsageStateResetResponse {
+    pub status: String,
+    pub mode: String,
+    pub usage_state: String,
+    pub leader_epoch: u64,
+    pub next_cycle: u64,
+    pub reset_paths: Vec<String>,
+    #[serde(flatten)]
+    pub extra: serde_json::Map<String, serde_json::Value>,
+}
+
+#[derive(Debug, Serialize)]
+struct ScannerCycleResetRequest<'a> {
+    mode: &'a str,
+}
+
+#[derive(Debug, Serialize)]
+struct ScannerUsageStateResetRequest<'a> {
+    mode: &'a str,
+}
+
 /// Freshness block of the scanner status response.
 #[derive(Debug, Clone, Deserialize)]
 #[serde(rename_all = "camelCase")]
@@ -358,8 +399,22 @@ impl AdminClient {
        prefix: Option<&str>,
        client_token: &str,
    ) -> Result<HealTaskStatus, AdminClientError> {
-        self.post_json(&heal_path(bucket, prefix), &[("clientToken", client_token.to_string())], Vec::new())
-            .await
+        self.heal_status_since(bucket, prefix, client_token, None).await
+    }
+
+    /// Query a retained result window. Missing cursors and outcome remain unknown.
+    pub async fn heal_status_since(
+        &self,
+        bucket: Option<&str>,
+        prefix: Option<&str>,
+        client_token: &str,
+        since_seq: Option<u64>,
+    ) -> Result<HealTaskStatus, AdminClientError> {
+        let mut query = vec![("clientToken", client_token.to_string())];
+        if let Some(since_seq) = since_seq {
+            query.push(("sinceSeq", since_seq.to_string()));
+        }
+        self.post_json(&heal_path(bucket, prefix), &query, Vec::new()).await
    }

    /// Stop a heal: with a `client_token` only that task is cancelled and its
@@ -378,7 +433,7 @@ impl AdminClient {
        match client_token {
            Some(_) => {
                let status: HealTaskStatus = self.post_json(&heal_path(bucket, prefix), &query, Vec::new()).await?;
-                Ok(HealStopOutcome::Stopped(status))
+                Ok(HealStopOutcome::Stopped(Box::new(status)))
            }
            None => {
                let success: HealStartSuccess = self.post_json(&heal_path(bucket, prefix), &query, Vec::new()).await?;
@@ -398,6 +453,25 @@ impl AdminClient {
        self.get_json("/v3/scanner/status").await
    }

+    /// Request a legacy synchronous scanner cycle reset (`full-rescan`).
+    pub async fn scanner_cycle_state_reset_full_rescan(&self) -> Result<ScannerCycleResetResponse, AdminClientError> {
+        let body =
+            serde_json::to_vec(&ScannerCycleResetRequest { mode: "full-rescan" }).map_err(|err| AdminClientError::Decode {
+                message: err.to_string(),
+            })?;
+        self.post_json("/v3/scanner/cycle-state/reset", &[], body).await
+    }
+
+    /// Request a legacy synchronous scanner usage reset (`full-rebuild`).
+    pub async fn scanner_usage_state_reset_full_rebuild(&self) -> Result<ScannerUsageStateResetResponse, AdminClientError> {
+        let body = serde_json::to_vec(&ScannerUsageStateResetRequest { mode: "full-rebuild" }).map_err(|err| {
+            AdminClientError::Decode {
+                message: err.to_string(),
+            }
+        })?;
+        self.post_json("/v3/scanner/usage-state/reset", &[], body).await
+    }
+
    /// ILM expiry worker status. The payload is owned by the expiry
    /// subsystem and still evolving; returned verbatim.
    pub async fn ilm_expiry_status(&self) -> Result<serde_json::Value, AdminClientError> {
@@ -533,7 +607,7 @@ impl AdminClient {
 /// start-success-shaped receipt.
 #[derive(Debug, Clone)]
 pub enum HealStopOutcome {
-    Stopped(HealTaskStatus),
+    Stopped(Box<HealTaskStatus>),
    PathStopped(HealStartSuccess),
 }

@@ -566,7 +640,7 @@ pub(crate) fn percent_encode_path_segment(segment: &str) -> String {
 mod tests {
    use super::{
        AdminClient, AdminClientError, BackgroundHealStatus, HealOpts, HealScanMode, HealStartSuccess, HealTaskStatus,
-        ScannerStatus, heal_path, percent_encode_path_segment,
+        ScannerCycleResetResponse, ScannerStatus, ScannerUsageStateResetResponse, heal_path, percent_encode_path_segment,
    };
    use crate::test_support::TestServer;
    use serde_json::json;
@@ -635,6 +709,24 @@ mod tests {
        assert!(status.progress.is_none());
    }

+    #[test]
+    fn outcome_v3_decoder_preserves_canonical_unknown_and_future_fields() {
+        let cases: serde_json::Value =
+            serde_json::from_str(include_str!("../tests/fixtures/heal-outcome-v3.json")).expect("shared fixtures");
+        for case in cases.as_array().expect("cases") {
+            let status: HealTaskStatus = serde_json::from_value(case["response"].clone()).expect("optional outcome response");
+            assert_eq!(status.outcome.as_ref(), Some(&case["response"]["outcome"]));
+            assert_eq!((status.next_seq, status.min_seq), (Some(9), Some(4)));
+            assert!(status.truncated);
+        }
+        let old: HealTaskStatus = serde_json::from_value(json!({"summary":"finished"})).expect("legacy response");
+        assert!(old.outcome.is_none() && old.next_seq.is_none() && old.min_seq.is_none());
+        let future = json!({"execution":{"state":"future_state"},"newField":7});
+        let status: HealTaskStatus =
+            serde_json::from_value(json!({"summary":"running","outcome":future})).expect("future outcome remains opaque");
+        assert_eq!(status.outcome, Some(future));
+    }
+
    #[test]
    fn background_heal_status_types_known_fields_and_passes_the_rest_through() {
        let raw = json!({
@@ -691,6 +783,33 @@ mod tests {
        assert_eq!(bare.freshness(), "unknown");
    }

+    #[test]
+    fn scanner_reset_responses_preserve_future_fields() {
+        let cycle: ScannerCycleResetResponse =
+            serde_json::from_value(json!({"status": "reset", "mode": "full-rescan", "future": true})).unwrap();
+        assert_eq!(cycle.status, "reset");
+        assert_eq!(cycle.mode, "full-rescan");
+        assert_eq!(cycle.extra["future"], true);
+
+        let usage: ScannerUsageStateResetResponse = serde_json::from_value(json!({
+            "status": "reset",
+            "mode": "full-rebuild",
+            "usage_state": "bootstrap-pending",
+            "leader_epoch": 11,
+            "next_cycle": 42,
+            "reset_paths": [".usage.json"],
+            "future": {"accepted": false}
+        }))
+        .unwrap();
+        assert_eq!(usage.status, "reset");
+        assert_eq!(usage.mode, "full-rebuild");
+        assert_eq!(usage.usage_state, "bootstrap-pending");
+        assert_eq!(usage.leader_epoch, 11);
+        assert_eq!(usage.next_cycle, 42);
+        assert_eq!(usage.reset_paths, [".usage.json"]);
+        assert_eq!(usage.extra["future"]["accepted"], false);
+    }
+
    #[test]
    fn invalid_endpoint_is_rejected_without_io() {
        let err = AdminClient::new("not a url", "ak", "sk").unwrap_err();
@@ -733,6 +852,52 @@ mod tests {
        assert!(request.body.contains("\"recursive\":true"));
    }

+    #[tokio::test]
+    async fn scanner_cycle_reset_posts_legacy_full_rescan_request() {
+        let server = TestServer::spawn(r#"{"status":"reset","mode":"full-rescan"}"#, 200).await;
+        let client = AdminClient::new(&format!("http://{}", server.addr), "ak", "sk").unwrap();
+
+        let reset = client
+            .scanner_cycle_state_reset_full_rescan()
+            .await
+            .expect("cycle reset response decodes");
+
+        assert_eq!(reset.status, "reset");
+        assert_eq!(reset.mode, "full-rescan");
+        let request = server.recorded();
+        assert_eq!(request.method, "POST");
+        assert_eq!(request.path, "/rustfs/admin/v3/scanner/cycle-state/reset");
+        assert_eq!(request.query, "");
+        assert!(request.body.contains("\"mode\":\"full-rescan\""));
+    }
+
+    #[tokio::test]
+    async fn scanner_usage_reset_posts_legacy_full_rebuild_request() {
+        let server = TestServer::spawn(
+            r#"{"status":"reset","mode":"full-rebuild","usage_state":"bootstrap-pending","leader_epoch":11,"next_cycle":42,"reset_paths":[".usage.json"]}"#,
+            200,
+        )
+        .await;
+        let client = AdminClient::new(&format!("http://{}", server.addr), "ak", "sk").unwrap();
+
+        let reset = client
+            .scanner_usage_state_reset_full_rebuild()
+            .await
+            .expect("usage reset response decodes");
+
+        assert_eq!(reset.status, "reset");
+        assert_eq!(reset.mode, "full-rebuild");
+        assert_eq!(reset.usage_state, "bootstrap-pending");
+        assert_eq!(reset.leader_epoch, 11);
+        assert_eq!(reset.next_cycle, 42);
+        assert_eq!(reset.reset_paths, [".usage.json"]);
+        let request = server.recorded();
+        assert_eq!(request.method, "POST");
+        assert_eq!(request.path, "/rustfs/admin/v3/scanner/usage-state/reset");
+        assert_eq!(request.query, "");
+        assert!(request.body.contains("\"mode\":\"full-rebuild\""));
+    }
+
    #[tokio::test]
    async fn query_sends_client_token_on_the_same_path() {
        let body = r#"{"summary":"running","detail":"","settings":{"recursive":false},"items":[],"truncated":false}"#;
@@ -750,6 +915,26 @@ mod tests {
        assert!(!request.query.contains("forceStop"));
    }

+    #[tokio::test]
+    async fn outcome_v3_since_query_preserves_cursor_and_never_sends_force_start() {
+        let server = TestServer::spawn(
+            r#"{"summary":"running","nextSeq":9,"minSeq":4,"truncated":true,"outcome":{"execution":{"state":"future_state"}}}"#,
+            200,
+        )
+        .await;
+        let client = AdminClient::new(&format!("http://{}", server.addr), "ak", "sk").expect("test client");
+        let status = client
+            .heal_status_since(Some("bucket"), None, "token-1", Some(3))
+            .await
+            .expect("window response");
+        assert_eq!((status.next_seq, status.min_seq), (Some(9), Some(4)));
+        assert!(status.truncated);
+        assert_eq!(status.outcome.expect("future state is preserved")["execution"]["state"], "future_state");
+        let request = server.recorded();
+        assert!(request.query.contains("sinceSeq=3") && request.query.contains("clientToken=token-1"));
+        assert!(!request.query.contains("forceStart") && !request.query.contains("forceStop"));
+    }
+
    #[tokio::test]
    async fn stop_without_token_takes_the_path_cancel_branch() {
        let server = TestServer::spawn(r#"{"clientToken":"path","clientAddress":"c","startTime":"t"}"#, 200).await;
@@ -762,6 +947,25 @@ mod tests {
        assert!(!request.query.contains("clientToken"));
    }

+    #[tokio::test]
+    async fn stop_with_token_decodes_boxed_task_status() {
+        let body = r#"{"summary":"stopped","detail":"","settings":{"recursive":false},"items":[],"truncated":false}"#;
+        let server = TestServer::spawn(body, 200).await;
+        let client = AdminClient::new(&format!("http://{}", server.addr), "ak", "sk").unwrap();
+
+        let outcome = client
+            .heal_stop(Some("bucket"), None, Some("token-1"))
+            .await
+            .expect("token stop decodes");
+        let super::HealStopOutcome::Stopped(status) = outcome else {
+            panic!("token stop should return task status");
+        };
+        assert_eq!(status.summary, "stopped");
+        let request = server.recorded();
+        assert!(request.query.contains("forceStop=true"));
+        assert!(request.query.contains("clientToken=token-1"));
+    }
+
    #[tokio::test]
    async fn background_heal_status_posts_to_the_registered_route() {
        let body = r#"{"state":"idle","healQueueLength":0,"healActiveTasks":0,"clusterStatusComplete":true}"#;
@@ -0,0 +1,472 @@
+[
+  {
+    "name": "completed",
+    "cliExit": 0,
+    "response": {
+      "summary": "finished",
+      "detail": "heal result items were truncated",
+      "startTime": "2026-01-01T00:00:00Z",
+      "settings": {
+        "recursive": true,
+        "scanMode": 1
+      },
+      "items": [],
+      "truncated": true,
+      "nextSeq": 9,
+      "minSeq": 4,
+      "progress": {
+        "objectsScanned": 11,
+        "objectsHealed": 7
+      },
+      "outcome": {
+        "execution": {
+          "state": "completed"
+        },
+        "coverage": "complete",
+        "counters": {
+          "processed": 0,
+          "healed": 0,
+          "unchanged": 0,
+          "skipped": 0,
+          "failed": 0,
+          "unknown": 0,
+          "attemptFailures": 0,
+          "overflowed": false
+        },
+        "objects": [],
+        "objectsTruncated": false
+      }
+    }
+  },
+  {
+    "name": "unknown",
+    "cliExit": 0,
+    "response": {
+      "summary": "finished",
+      "detail": "heal traversal completed; authoritative storage proof is unavailable for 1 objects; heal result items were truncated",
+      "startTime": "2026-01-01T00:00:00Z",
+      "settings": {
+        "recursive": true,
+        "scanMode": 1
+      },
+      "items": [],
+      "truncated": true,
+      "nextSeq": 9,
+      "minSeq": 4,
+      "progress": {
+        "objectsScanned": 11,
+        "objectsHealed": 7
+      },
+      "outcome": {
+        "execution": {
+          "state": "completed"
+        },
+        "coverage": "complete",
+        "counters": {
+          "processed": 1,
+          "healed": 0,
+          "unchanged": 0,
+          "skipped": 1,
+          "failed": 0,
+          "unknown": 1,
+          "attemptFailures": 0,
+          "overflowed": false
+        },
+        "objects": [
+          {
+            "identity": {
+              "kind": "object",
+              "bucket": "bucket",
+              "object": "object",
+              "versionId": null,
+              "bucketIncarnationId": null,
+              "poolIndex": null,
+              "setIndex": null
+            },
+            "disposition": {
+              "state": "unknown"
+            },
+            "detail": null
+          }
+        ],
+        "objectsTruncated": false
+      }
+    }
+  },
+  {
+    "name": "completed_with_errors",
+    "cliExit": 1,
+    "response": {
+      "summary": "stopped",
+      "detail": "heal traversal completed with errors: 1 failed objects; heal result items were truncated",
+      "startTime": "2026-01-01T00:00:00Z",
+      "settings": {
+        "recursive": true,
+        "scanMode": 1
+      },
+      "items": [],
+      "truncated": true,
+      "nextSeq": 9,
+      "minSeq": 4,
+      "progress": {
+        "objectsScanned": 11,
+        "objectsHealed": 7
+      },
+      "outcome": {
+        "execution": {
+          "state": "completed_with_errors"
+        },
+        "coverage": "complete",
+        "counters": {
+          "processed": 1,
+          "healed": 0,
+          "unchanged": 0,
+          "skipped": 0,
+          "failed": 1,
+          "unknown": 0,
+          "attemptFailures": 1,
+          "overflowed": false
+        },
+        "objects": [
+          {
+            "identity": {
+              "kind": "object",
+              "bucket": "bucket",
+              "object": "object",
+              "versionId": null,
+              "bucketIncarnationId": null,
+              "poolIndex": null,
+              "setIndex": null
+            },
+            "disposition": {
+              "state": "failed",
+              "details": "retry_exhausted"
+            },
+            "detail": null
+          }
+        ],
+        "objectsTruncated": false
+      }
+    }
+  },
+  {
+    "name": "cancelled",
+    "cliExit": 1,
+    "response": {
+      "summary": "stopped",
+      "detail": "heal task cancelled; heal result items were truncated",
+      "startTime": "2026-01-01T00:00:00Z",
+      "settings": {
+        "recursive": true,
+        "scanMode": 1
+      },
+      "items": [],
+      "truncated": true,
+      "nextSeq": 9,
+      "minSeq": 4,
+      "progress": {
+        "objectsScanned": 11,
+        "objectsHealed": 7
+      },
+      "outcome": {
+        "execution": {
+          "state": "aborted",
+          "reason": "cancelled"
+        },
+        "coverage": "partial",
+        "counters": {
+          "processed": 0,
+          "healed": 0,
+          "unchanged": 0,
+          "skipped": 0,
+          "failed": 0,
+          "unknown": 0,
+          "attemptFailures": 0,
+          "overflowed": false
+        },
+        "objects": [],
+        "objectsTruncated": false
+      }
+    }
+  },
+  {
+    "name": "deadline",
+    "cliExit": 1,
+    "response": {
+      "summary": "stopped",
+      "detail": "heal task timed out; heal result items were truncated",
+      "startTime": "2026-01-01T00:00:00Z",
+      "settings": {
+        "recursive": true,
+        "scanMode": 1
+      },
+      "items": [],
+      "truncated": true,
+      "nextSeq": 9,
+      "minSeq": 4,
+      "progress": {
+        "objectsScanned": 11,
+        "objectsHealed": 7
+      },
+      "outcome": {
+        "execution": {
+          "state": "aborted",
+          "reason": "deadline"
+        },
+        "coverage": "partial",
+        "counters": {
+          "processed": 0,
+          "healed": 0,
+          "unchanged": 0,
+          "skipped": 0,
+          "failed": 0,
+          "unknown": 0,
+          "attemptFailures": 0,
+          "overflowed": false
+        },
+        "objects": [],
+        "objectsTruncated": false
+      }
+    }
+  },
+  {
+    "name": "untraversable",
+    "cliExit": 1,
+    "response": {
+      "summary": "stopped",
+      "detail": "heal listing is untraversable; heal result items were truncated",
+      "startTime": "2026-01-01T00:00:00Z",
+      "settings": {
+        "recursive": true,
+        "scanMode": 1
+      },
+      "items": [],
+      "truncated": true,
+      "nextSeq": 9,
+      "minSeq": 4,
+      "progress": {
+        "objectsScanned": 11,
+        "objectsHealed": 7
+      },
+      "outcome": {
+        "execution": {
+          "state": "aborted",
+          "reason": "untraversable"
+        },
+        "coverage": "partial",
+        "counters": {
+          "processed": 0,
+          "healed": 0,
+          "unchanged": 0,
+          "skipped": 0,
+          "failed": 0,
+          "unknown": 0,
+          "attemptFailures": 0,
+          "overflowed": false
+        },
+        "objects": [],
+        "objectsTruncated": false
+      }
+    }
+  },
+  {
+    "name": "remote_completed_with_errors",
+    "cliExit": 1,
+    "response": {
+      "summary": "stopped",
+      "detail": "heal traversal completed with errors: 1 failed objects; heal result items were truncated",
+      "startTime": "2026-01-01T00:00:00Z",
+      "settings": {
+        "recursive": true,
+        "scanMode": 1
+      },
+      "items": [],
+      "truncated": true,
+      "nextSeq": 9,
+      "minSeq": 4,
+      "progress": {
+        "objectsScanned": 11,
+        "objectsHealed": 7
+      },
+      "outcome": {
+        "execution": {
+          "state": "completed_with_errors",
+          "futureExtension": {
+            "value": 7
+          }
+        },
+        "coverage": "complete",
+        "counters": {
+          "processed": 1,
+          "healed": 0,
+          "unchanged": 0,
+          "skipped": 0,
+          "failed": 1,
+          "unknown": 0,
+          "attemptFailures": 1,
+          "overflowed": false,
+          "futureCounter": 11
+        },
+        "objects": [
+          {
+            "identity": {
+              "kind": "object",
+              "bucket": "bucket",
+              "object": "object",
+              "versionId": null,
+              "bucketIncarnationId": null,
+              "poolIndex": null,
+              "setIndex": null
+            },
+            "disposition": {
+              "state": "failed",
+              "details": "retry_exhausted"
+            },
+            "detail": null
+          }
+        ],
+        "objectsTruncated": false
+      }
+    },
+    "remoteResponse": {
+      "summary": "finished",
+      "detail": "",
+      "startTime": "2026-01-01T00:00:00Z",
+      "settings": {
+        "recursive": true,
+        "scanMode": 1
+      },
+      "items": [],
+      "truncated": true,
+      "nextSeq": 9,
+      "minSeq": 4,
+      "progress": {
+        "objectsScanned": 11,
+        "objectsHealed": 7
+      },
+      "outcome": {
+        "execution": {
+          "state": "completed_with_errors",
+          "futureExtension": {
+            "value": 7
+          }
+        },
+        "coverage": "complete",
+        "counters": {
+          "processed": 1,
+          "healed": 0,
+          "unchanged": 0,
+          "skipped": 0,
+          "failed": 1,
+          "unknown": 0,
+          "attemptFailures": 1,
+          "overflowed": false,
+          "futureCounter": 11
+        },
+        "objects": [
+          {
+            "identity": {
+              "kind": "object",
+              "bucket": "bucket",
+              "object": "object",
+              "versionId": null,
+              "bucketIncarnationId": null,
+              "poolIndex": null,
+              "setIndex": null
+            },
+            "disposition": {
+              "state": "failed",
+              "details": "retry_exhausted"
+            },
+            "detail": null
+          }
+        ],
+        "objectsTruncated": false
+      }
+    }
+  },
+  {
+    "name": "remote_cancelled",
+    "cliExit": 1,
+    "response": {
+      "summary": "stopped",
+      "detail": "heal task cancelled; heal result items were truncated",
+      "startTime": "2026-01-01T00:00:00Z",
+      "settings": {
+        "recursive": true,
+        "scanMode": 1
+      },
+      "items": [],
+      "truncated": true,
+      "nextSeq": 9,
+      "minSeq": 4,
+      "progress": {
+        "objectsScanned": 11,
+        "objectsHealed": 7
+      },
+      "outcome": {
+        "execution": {
+          "state": "aborted",
+          "reason": "cancelled",
+          "futureExtension": {
+            "value": 7
+          }
+        },
+        "coverage": "partial",
+        "counters": {
+          "processed": 0,
+          "healed": 0,
+          "unchanged": 0,
+          "skipped": 0,
+          "failed": 0,
+          "unknown": 0,
+          "attemptFailures": 0,
+          "overflowed": false,
+          "futureCounter": 11
+        },
+        "objects": [],
+        "objectsTruncated": false
+      }
+    },
+    "remoteResponse": {
+      "summary": "finished",
+      "detail": "",
+      "startTime": "2026-01-01T00:00:00Z",
+      "settings": {
+        "recursive": true,
+        "scanMode": 1
+      },
+      "items": [],
+      "truncated": true,
+      "nextSeq": 9,
+      "minSeq": 4,
+      "progress": {
+        "objectsScanned": 11,
+        "objectsHealed": 7
+      },
+      "outcome": {
+        "execution": {
+          "state": "aborted",
+          "reason": "cancelled",
+          "futureExtension": {
+            "value": 7
+          }
+        },
+        "coverage": "partial",
+        "counters": {
+          "processed": 0,
+          "healed": 0,
+          "unchanged": 0,
+          "skipped": 0,
+          "failed": 0,
+          "unknown": 0,
+          "attemptFailures": 0,
+          "overflowed": false,
+          "futureCounter": 11
+        },
+        "objects": [],
+        "objectsTruncated": false
+      }
+    }
+  }
+]
@@ -1257,6 +1257,8 @@ pub struct ScannerDirtyUsageBucket {
    pub bucket: ::prost::alloc::string::String,
    #[prost(uint64, tag = "2")]
    pub generation: u64,
+    #[prost(bytes = "bytes", tag = "3")]
+    pub bucket_incarnation: ::prost::bytes::Bytes,
 }
 #[derive(Clone, PartialEq, Eq, Hash, ::prost::Message)]
 pub struct ScannerDirtyUsageSnapshotRequest {
@@ -1282,6 +1284,8 @@ pub struct ScannerDirtyUsageSnapshotResponse {
    pub buckets: ::prost::alloc::vec::Vec<ScannerDirtyUsageBucket>,
    #[prost(bytes = "bytes", tag = "7")]
    pub response_proof: ::prost::bytes::Bytes,
+    #[prost(string, tag = "8")]
+    pub owner_id: ::prost::alloc::string::String,
 }
 /// Receiver-only protocol. Producers must retain whole-cycle ACK until they
 /// have a durable per-bucket publication proof.
@@ -175,6 +175,9 @@ pub const BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION: u32 = 2;
 pub const HEAL_CONTROL_CAPABILITY_PROBE_PREFIX: &[u8] = b"rustfs-heal-control-capability-v3\0";
 pub const REMOTE_VERSION_STATE_CAPABILITY_PROBE_PREFIX: &[u8] = b"rustfs-tier-remote-version-state-capability-v1\0";
 pub const CROSS_POOL_FENCE_CAPABILITY_PROBE_PREFIX: &[u8] = b"rustfs-cross-pool-fence-capability-v1\0";
+pub const ILM_RECOVERY_EXPORT_CAPABILITY_PROBE_PREFIX: &[u8] = b"rustfs-ilm-recovery-export-capability-v1\0";
+pub const TRANSITION_TRANSACTION_COMPACTION_CAPABILITY_PROBE_PREFIX: &[u8] =
+    b"rustfs-transition-transaction-compaction-capability-v1\0";
 pub const TIER_MUTATION_RPC_MAX_PREPARE_PAYLOAD_SIZE: usize = 64 * 1024;
 pub const TIER_MUTATION_RPC_MAX_COMMIT_PAYLOAD_SIZE: usize = 1024;
 pub const TIER_MUTATION_RPC_MAX_ABORT_PAYLOAD_SIZE: usize = TIER_MUTATION_RPC_MAX_PREPARE_PAYLOAD_SIZE;
@@ -219,6 +222,30 @@ pub fn is_cross_pool_fence_capability_probe(command: &[u8]) -> bool {
        && command.starts_with(CROSS_POOL_FENCE_CAPABILITY_PROBE_PREFIX)
 }

+pub fn ilm_recovery_export_capability_probe(nonce: &[u8; 16]) -> Vec<u8> {
+    let mut probe = Vec::with_capacity(ILM_RECOVERY_EXPORT_CAPABILITY_PROBE_PREFIX.len() + nonce.len());
+    probe.extend_from_slice(ILM_RECOVERY_EXPORT_CAPABILITY_PROBE_PREFIX);
+    probe.extend_from_slice(nonce);
+    probe
+}
+
+pub fn is_ilm_recovery_export_capability_probe(command: &[u8]) -> bool {
+    command.len() == ILM_RECOVERY_EXPORT_CAPABILITY_PROBE_PREFIX.len() + 16
+        && command.starts_with(ILM_RECOVERY_EXPORT_CAPABILITY_PROBE_PREFIX)
+}
+
+pub fn transition_transaction_compaction_capability_probe(nonce: &[u8; 16]) -> Vec<u8> {
+    let mut probe = Vec::with_capacity(TRANSITION_TRANSACTION_COMPACTION_CAPABILITY_PROBE_PREFIX.len() + nonce.len());
+    probe.extend_from_slice(TRANSITION_TRANSACTION_COMPACTION_CAPABILITY_PROBE_PREFIX);
+    probe.extend_from_slice(nonce);
+    probe
+}
+
+pub fn is_transition_transaction_compaction_capability_probe(command: &[u8]) -> bool {
+    command.len() == TRANSITION_TRANSACTION_COMPACTION_CAPABILITY_PROBE_PREFIX.len() + 16
+        && command.starts_with(TRANSITION_TRANSACTION_COMPACTION_CAPABILITY_PROBE_PREFIX)
+}
+
 pub fn encode_remote_version_state_capability(
    topology_member: &str,
    process_epoch: &[u8; 16],
@@ -562,11 +589,13 @@ pub fn canonical_scanner_dirty_usage_snapshot_response_body(
    body.push_u64(response.generation);
    body.push_u64(response.pending_bucket_count);
    body.push_u32(response.protocol_version);
+    body.push_str(&response.owner_id)?;
    body.push_bool(response.complete);
    body.push_count(response.buckets.len())?;
    for bucket in &response.buckets {
        body.push_str(&bucket.bucket)?;
        body.push_u64(bucket.generation);
+        body.push_bytes(bucket.bucket_incarnation.as_ref())?;
    }
    Ok(body.finish())
 }
@@ -1829,13 +1858,16 @@ mod scanner_activity_tests {
                ScannerDirtyUsageBucket {
                    bucket: "archive".to_string(),
                    generation: 3,
+                    bucket_incarnation: vec![1; 16].into(),
                },
                ScannerDirtyUsageBucket {
                    bucket: "photos".to_string(),
                    generation: 7,
+                    bucket_incarnation: vec![2; 16].into(),
                },
            ],
            response_proof: vec![9; 32].into(),
+            owner_id: "11111111-1111-1111-1111-111111111111".to_string(),
        };
        let baseline = canonical_scanner_dirty_usage_snapshot_response_body(&[1; 16], &response)
            .expect("scanner dirty usage snapshot response should encode");
@@ -1852,6 +1884,9 @@ mod scanner_activity_tests {
        let mut protocol = response.clone();
        protocol.protocol_version = 2;
        variants.push(protocol);
+        let mut owner = response.clone();
+        owner.owner_id = "22222222-2222-2222-2222-222222222222".to_string();
+        variants.push(owner);
        let mut complete = response.clone();
        complete.complete = false;
        variants.push(complete);
@@ -1861,6 +1896,9 @@ mod scanner_activity_tests {
        let mut bucket_generation = response.clone();
        bucket_generation.buckets[0].generation = 4;
        variants.push(bucket_generation);
+        let mut bucket_incarnation = response.clone();
+        bucket_incarnation.buckets[0].bucket_incarnation = vec![3; 16].into();
+        variants.push(bucket_incarnation);
        let mut bucket_order = response.clone();
        bucket_order.buckets.reverse();
        variants.push(bucket_order);
@@ -2127,12 +2165,15 @@ mod scanner_activity_tests {
 mod heal_control_tests {
    use super::{
        CROSS_POOL_FENCE_CAPABILITY_PROBE_PREFIX, HEAL_CONTROL_CAPABILITY_PROBE_PREFIX, HEAL_CONTROL_PROTOCOL_VERSION,
-        REMOTE_VERSION_STATE_CAPABILITY_PROBE_PREFIX, canonical_heal_control_capability_ack, canonical_heal_control_request_body,
-        canonical_heal_control_response_body, decode_remote_version_state_capability, encode_cross_pool_fence_capability,
-        encode_remote_version_state_capability, heal_control_capability_probe, heal_control_coordinator_epoch,
-        heal_control_execution_timeout, heal_control_execution_timeout_for, internode_rpc_timeout,
-        is_cross_pool_fence_capability_probe, is_heal_control_capability_probe, is_remote_version_state_capability_probe,
-        normalize_internode_rpc_timeout, remote_version_state_capability_probe,
+        ILM_RECOVERY_EXPORT_CAPABILITY_PROBE_PREFIX, REMOTE_VERSION_STATE_CAPABILITY_PROBE_PREFIX,
+        TRANSITION_TRANSACTION_COMPACTION_CAPABILITY_PROBE_PREFIX, canonical_heal_control_capability_ack,
+        canonical_heal_control_request_body, canonical_heal_control_response_body, decode_remote_version_state_capability,
+        encode_cross_pool_fence_capability, encode_remote_version_state_capability, heal_control_capability_probe,
+        heal_control_coordinator_epoch, heal_control_execution_timeout, heal_control_execution_timeout_for,
+        ilm_recovery_export_capability_probe, internode_rpc_timeout, is_cross_pool_fence_capability_probe,
+        is_heal_control_capability_probe, is_ilm_recovery_export_capability_probe, is_remote_version_state_capability_probe,
+        is_transition_transaction_compaction_capability_probe, normalize_internode_rpc_timeout,
+        remote_version_state_capability_probe, transition_transaction_compaction_capability_probe,
    };
    use crate::heal_control;
    use std::time::Duration;
@@ -2196,6 +2237,34 @@ mod heal_control_tests {
        assert!(!is_remote_version_state_capability_probe(REMOTE_VERSION_STATE_CAPABILITY_PROBE_PREFIX));
    }

+    #[test]
+    fn ilm_recovery_export_capability_probe_requires_exact_prefix_and_nonce() {
+        let probe = ilm_recovery_export_capability_probe(&[7; 16]);
+        assert!(is_ilm_recovery_export_capability_probe(&probe));
+        assert!(!is_ilm_recovery_export_capability_probe(ILM_RECOVERY_EXPORT_CAPABILITY_PROBE_PREFIX));
+        let mut wrong_prefix = probe.clone();
+        wrong_prefix[0] ^= 1;
+        assert!(!is_ilm_recovery_export_capability_probe(&wrong_prefix));
+        let mut extra = probe;
+        extra.push(0);
+        assert!(!is_ilm_recovery_export_capability_probe(&extra));
+    }
+
+    #[test]
+    fn transition_transaction_compaction_probe_requires_exact_prefix_and_nonce() {
+        let probe = transition_transaction_compaction_capability_probe(&[7; 16]);
+        assert!(is_transition_transaction_compaction_capability_probe(&probe));
+        assert!(!is_transition_transaction_compaction_capability_probe(
+            TRANSITION_TRANSACTION_COMPACTION_CAPABILITY_PROBE_PREFIX,
+        ));
+        let mut wrong_prefix = probe.clone();
+        wrong_prefix[0] ^= 1;
+        assert!(!is_transition_transaction_compaction_capability_probe(&wrong_prefix));
+        let mut extra = probe;
+        extra.push(0);
+        assert!(!is_transition_transaction_compaction_capability_probe(&extra));
+    }
+
    #[test]
    fn remote_version_state_capability_binds_member_and_process_epoch() {
        let encoded =
@@ -885,6 +885,7 @@ message ScannerActivityResponse {
 message ScannerDirtyUsageBucket {
  string bucket = 1;
  uint64 generation = 2;
+  bytes bucket_incarnation = 3;
 }

 message ScannerDirtyUsageSnapshotRequest {
@@ -901,6 +902,7 @@ message ScannerDirtyUsageSnapshotResponse {
  bool complete = 5;
  repeated ScannerDirtyUsageBucket buckets = 6;
  bytes response_proof = 7;
+  string owner_id = 8;
 }

 // Receiver-only protocol. Producers must retain whole-cycle ACK until they
@@ -65,9 +65,10 @@ pub use multipart::{
 };
 pub use object::{
    ObjectLockIntegrity, ReplicationSourceObject, ReplicationTargetObject, SsecPassthroughCapability, SsecPassthroughGate,
-    content_matches_by_etag, is_replication_target_offline_error, object_lock_put_integrity, replication_action_for_target,
-    replication_etags_match, single_part_replica_etag_mismatch, ssec_passthrough_evidence_present, ssec_passthrough_gate,
-    target_is_newer_than_source_null_version, version_identity_drifted,
+    VersionIdentityCapability, content_matches_by_etag, is_replication_target_offline_error, object_lock_put_integrity,
+    replication_action_for_target, replication_etags_match, single_part_replica_etag_mismatch, ssec_passthrough_evidence_present,
+    ssec_passthrough_gate, target_is_newer_than_source_null_version, version_identity_capability_from_put,
+    version_identity_drifted,
 };
 pub use operation::{
    MustReplicateOptions, ReplicationDeleteScheduleInput, ReplicationDeleteSource, ReplicationDeleteStateSource,
@@ -234,18 +234,62 @@ fn comparable_metadata(metadata: Option<&HashMap<String, String>>) -> HashMap<St
 /// real (non-nil) version uuid, and drift means the target answered with
 /// anything else — including nothing at all.
 pub fn version_identity_drifted(source_version_id: &str, assigned_version_id: Option<&str>) -> bool {
-    if source_version_id.is_empty() {
-        return false;
+    version_identity_capability_from_put(source_version_id, assigned_version_id) == Some(VersionIdentityCapability::MintsOwn)
+}
+
+/// Whether a replication target adopts the source version id it is handed on
+/// PutObject / CompleteMultipartUpload, or mints its own.
+///
+/// A target that mints its own ids (AWS S3, Wasabi, Impossible Cloud) still
+/// stores the bytes, but every later version-addressed request from the
+/// source names an id the target never had. Its HEAD then answers 404 —
+/// indistinguishable from a replica that is really missing — so a heal, MRF
+/// retry or existing-object resync re-drive would PUT the object again and
+/// mint yet another target version (rustfs/backlog#2340). The replication
+/// worker learns the verdict from each PUT response (and replication-check's
+/// VersionFidelity phase) and, once `MintsOwn` is known, locates a replica by
+/// exact key and ETag before concluding that it is missing. The verdict cache
+/// is owned by the runtime's bucket target system; this crate owns only the
+/// vocabulary and the judgment.
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
+pub enum VersionIdentityCapability {
+    #[default]
+    Unknown,
+    Adopts,
+    MintsOwn,
+}
+
+impl VersionIdentityCapability {
+    /// True when a 404 from a version-addressed HEAD on this target cannot be
+    /// read as "replica missing": the source-side id was never the target's.
+    pub fn version_addressing_unreliable(self) -> bool {
+        self == VersionIdentityCapability::MintsOwn
+    }
+}
+
+/// Judge the identity contract from one replication write: `None` when no
+/// contract applies (the source addressed no real version — an empty or nil
+/// uuid travels as the literal "null", unversioned-source semantics),
+/// otherwise whether the target echoed the source id or answered with
+/// anything else — including nothing at all.
+pub fn version_identity_capability_from_put(
+    source_version_id: &str,
+    assigned_version_id: Option<&str>,
+) -> Option<VersionIdentityCapability> {
+    if source_version_id.is_empty() {
+        return None;
    }
-    // A nil source uuid travels as the literal "null" (unversioned-source
-    // semantics); no identity contract applies to it.
    if uuid::Uuid::parse_str(source_version_id)
        .map(|uuid| uuid.is_nil())
        .unwrap_or(true)
    {
-        return false;
+        return None;
    }
-    assigned_version_id != Some(source_version_id)
+    Some(if assigned_version_id == Some(source_version_id) {
+        VersionIdentityCapability::Adopts
+    } else {
+        VersionIdentityCapability::MintsOwn
+    })
 }

 const REPLICATION_TARGET_OFFLINE_ERROR_MARKERS: &[&str] = &[
@@ -377,9 +421,9 @@ mod tests {

    use super::{
        ReplicationSourceObject, ReplicationTargetObject, SsecPassthroughCapability, SsecPassthroughGate,
-        content_matches_by_etag, is_replication_target_offline_error, replication_action_for_target, replication_etags_match,
-        single_part_replica_etag_mismatch, ssec_passthrough_evidence_present, ssec_passthrough_gate,
-        target_is_newer_than_source_null_version, version_identity_drifted,
+        VersionIdentityCapability, content_matches_by_etag, is_replication_target_offline_error, replication_action_for_target,
+        replication_etags_match, single_part_replica_etag_mismatch, ssec_passthrough_evidence_present, ssec_passthrough_gate,
+        target_is_newer_than_source_null_version, version_identity_capability_from_put, version_identity_drifted,
    };
    use crate::filemeta::{ReplicationAction, ReplicationType};
    use crate::http::AMZ_OBJECT_LOCK_MODE;
@@ -508,6 +552,32 @@ mod tests {
        }
    }

+    #[test]
+    fn version_identity_capability_is_judged_only_for_real_source_versions() {
+        let source = "8e4d2f4c-2d5c-4f1b-9d0a-9c8b7a6f5e4d";
+        assert_eq!(
+            version_identity_capability_from_put(source, Some(source)),
+            Some(VersionIdentityCapability::Adopts)
+        );
+        // Wasabi / AWS shape: a minted id, or no id at all, both mean the
+        // source-side id is not addressable on the target.
+        assert_eq!(
+            version_identity_capability_from_put(source, Some("001788697733811332140-fR6j6uXKV-")),
+            Some(VersionIdentityCapability::MintsOwn)
+        );
+        assert_eq!(
+            version_identity_capability_from_put(source, None),
+            Some(VersionIdentityCapability::MintsOwn)
+        );
+        // No contract for an unversioned source write.
+        assert_eq!(version_identity_capability_from_put("", Some("anything")), None);
+        assert_eq!(version_identity_capability_from_put("00000000-0000-0000-0000-000000000000", None), None);
+        assert_eq!(version_identity_capability_from_put("null", Some("null")), None);
+        assert!(VersionIdentityCapability::MintsOwn.version_addressing_unreliable());
+        assert!(!VersionIdentityCapability::Adopts.version_addressing_unreliable());
+        assert!(!VersionIdentityCapability::Unknown.version_addressing_unreliable());
+    }
+
    #[test]
    fn replication_target_offline_error_classifier_is_network_scoped() {
        assert!(is_replication_target_offline_error("put_object dispatch failure: connector error"));
@@ -106,6 +106,7 @@ bytes.workspace = true
 hex-simd.workspace = true

 [dev-dependencies]
+rustfs-heal.workspace = true
 tracing-subscriber = { workspace = true, features = ["json", "env-filter", "time"] }
 serial_test = { workspace = true }
 temp-env = { workspace = true, features = ["async_closure"] }
@@ -85,14 +85,14 @@ pub use rustfs_scanner_metrics::last_minute;
 pub use scanner::{
    ScannerCycleRecoveryMarker, ScannerCycleRecoveryStatus, ScannerCycleScheduleStatus, ScannerPauseBacklogAlertReason,
    ScannerPauseBacklogPhase, ScannerPauseBacklogStatus, ScannerPauseBacklogThresholds, ScannerUsageStateResetResult,
-    init_data_scanner, reset_scanner_cycle_recovery, reset_scanner_usage_state_for_full_rebuild, scanner_cycle_recovery_status,
-    scanner_cycle_schedule_status, scanner_pause_backlog_status, scanner_topology_digest,
+    init_data_scanner, init_scanner_with_recovery, reset_scanner_cycle_recovery, reset_scanner_usage_state_for_full_rebuild,
+    scanner_cycle_recovery_status, scanner_cycle_schedule_status, scanner_pause_backlog_status, scanner_topology_digest,
 };
 pub use scanner_io::{
    ScannerDirtyUsageAckError, ScannerDirtyUsageBucket, ScannerDirtyUsageSnapshot, ScannerDirtyUsageState,
    acknowledge_dirty_usage_generation, acknowledge_scoped_dirty_usage, clear_dirty_usage_bucket, record_dirty_usage_bucket,
-    record_scanner_maintenance_change, scanner_activity_epoch, scanner_dirty_usage_snapshot, scanner_dirty_usage_state,
-    scanner_maintenance_generation,
+    record_dirty_usage_object, record_scanner_maintenance_change, scanner_activity_epoch, scanner_dirty_usage_snapshot,
+    scanner_dirty_usage_state, scanner_maintenance_generation,
 };
 pub use sleeper::{DynamicSleeper, SCANNER_IDLE_MODE, SCANNER_SLEEPER};
 use std::sync::atomic::{AtomicU64, Ordering};
@@ -100,6 +100,7 @@ pub use storage_api::ScannerReplicationConfig as ReplicationConfig;
 pub use storage_api::scan::{
    SCANNER_ACTIVITY_PROTOCOL_VERSION, SCANNER_ACTIVITY_V6_PROTOCOL_VERSION, SCANNER_DIRTY_USAGE_SNAPSHOT_MAX_ENTRIES,
    SCANNER_DIRTY_USAGE_SNAPSHOT_PROTOCOL_VERSION, SCANNER_DIRTY_USAGE_SNAPSHOT_RPC_MAX_MESSAGE_SIZE,
+    SCANNER_SCOPED_DIRTY_USAGE_ACK_MAX_ENTRIES,
 };
 pub use workload_admission::set_scanner_workload_admission_snapshot_provider;

@@ -16,8 +16,9 @@
 use crate::RUSTFS_META_BUCKET;
 use crate::scanner_budget::{ScannerCycleBudget, ScannerCycleBudgetConfig};
 use crate::scanner_io::{
-    DataUsageCacheReuseOptions, DataUsageCacheScanState, ScannerDiskScanOutcome, ScannerIODisk, acquire_scanner_cache_locks,
-    cache_root_entry_info, current_cache_root_or_prepare_with_generation, scanner_set_disk_inventory,
+    DataUsageCacheReuseOptions, DataUsageCacheScanState, ScannerDiskScanOptions, ScannerDiskScanOutcome, ScannerIODisk,
+    acquire_scanner_cache_locks, cache_root_entry_info, current_cache_root_or_prepare_with_generation,
+    scanner_set_disk_inventory,
 };
 use crate::storage_api::owner::NS_SCANNER_PROTOCOL_VERSION;
 use crate::{
@@ -779,7 +780,18 @@ async fn scan_and_persist_local_bucket(

    let set_disks = scanner_set_disk_inventory(set.as_ref()).await;
    let scan_ctx = ctx.child_token();
-    let scan = ScannerIODisk::nsscanner_disk(disk.clone(), scan_ctx.clone(), budget, set_disks, cache, None, scan_mode);
+    let scan = ScannerIODisk::nsscanner_disk(
+        disk.clone(),
+        scan_ctx.clone(),
+        budget,
+        set_disks,
+        cache,
+        None,
+        ScannerDiskScanOptions {
+            scan_mode,
+            prefix_scan_scope: None,
+        },
+    );
    tokio::pin!(scan);
    let fence_watch = watch_remote_scanner_request_fence(next_cycle, leader_epoch, store.clone(), NS_SCANNER_FENCE_POLL_INTERVAL);
    tokio::pin!(fence_watch);
@@ -14,7 +14,6 @@

 use std::collections::BTreeMap;
 use std::future::Future;
-#[cfg(test)]
 use std::sync::Mutex as StdMutex;
 use std::sync::atomic::{AtomicBool, Ordering};
 use std::sync::{Arc, LazyLock, RwLock};
@@ -948,6 +947,33 @@ pub async fn init_data_scanner(ctx: CancellationToken, storeapi: Arc<ECStore>) {
    init_data_scanner_with_storage(ctx, storeapi).await;
 }

+/// Start normal scanning when enabled, or one resume-only cleanup attempt.
+/// The disabled branch returns a finite task for the startup owner to join;
+/// it never enables ordinary namespace scanning or accepts a new reset intent.
+pub async fn init_scanner_with_recovery(
+    ctx: CancellationToken,
+    storeapi: Arc<ECStore>,
+    enabled: bool,
+) -> Option<tokio::task::JoinHandle<()>> {
+    if enabled {
+        init_data_scanner(ctx, storeapi).await;
+        return None;
+    }
+    Some(tokio::spawn(async move {
+        if let Err(error) = resume_scanner_cycle_cleanup(ctx, storeapi).await {
+            warn!(
+                target: "rustfs::scanner",
+                event = EVENT_SCANNER_PERSIST_STATE,
+                component = LOG_COMPONENT_SCANNER,
+                subsystem = LOG_SUBSYSTEM_RUNTIME,
+                state = "disabled_cleanup_deferred",
+                error = %error,
+                "Disabled scanner cleanup remains pending for an operator retry"
+            );
+        }
+    }))
+}
+
 async fn init_data_scanner_with_storage<S>(ctx: CancellationToken, storeapi: Arc<S>)
 where
    S: ScannerStorage,
@@ -1558,6 +1584,11 @@ async fn mark_scan_cycle_idle(cycle_info: &mut CurrentCycle, cycle_metrics_guard
    cycle_metrics_guard.finish(cycle_info.clone()).await;
 }

+struct ScannerCycleScheduling {
+    requires_full_scan: bool,
+    service_cohort: Option<Arc<StdMutex<crate::scanner_io::ScannerServiceCohort>>>,
+}
+
 #[cfg(test)]
 async fn run_data_scanner_cycle<S>(
    ctx: &CancellationToken,
@@ -1570,7 +1601,19 @@ where
    S: ScannerStorage,
 {
    let cycle_budget = ScannerCycleBudget::new(ctx, scanner_cycle_budget_config());
-    run_data_scanner_cycle_with_budget(ctx, storeapi, cycle_info, cycle_revision, leader_epoch, cycle_budget, true).await
+    run_data_scanner_cycle_with_budget(
+        ctx,
+        storeapi,
+        cycle_info,
+        cycle_revision,
+        leader_epoch,
+        cycle_budget,
+        ScannerCycleScheduling {
+            requires_full_scan: true,
+            service_cohort: None,
+        },
+    )
+    .await
 }

 #[instrument(skip_all)]
@@ -1582,7 +1625,7 @@ async fn run_data_scanner_cycle_with_budget<S>(
    cycle_revision: &mut DataUsageCacheRevision,
    leader_epoch: u64,
    cycle_budget: Arc<ScannerCycleBudget>,
-    requires_full_scan: bool,
+    scheduling: ScannerCycleScheduling,
 ) -> ScannerCycleOutcome
 where
    S: ScannerStorage,
@@ -1687,6 +1730,7 @@ where
    };
    let baseline_publication_epoch = baseline_publication_guard.epoch();
    let usage_persist_baseline_result = read_data_usage_persist_baseline(storeapi.clone()).await;
+    let observed_usage_candidate_result = read_config(storeapi.clone(), DATA_USAGE_OBSERVED_OBJ_NAME_PATH.as_str()).await;
    drop(baseline_publication_guard);
    let usage_persist_baseline = match usage_persist_baseline_result {
        Ok(baseline) => baseline,
@@ -1707,6 +1751,24 @@ where
            return ScannerCycleOutcome::Failed;
        }
    };
+    let observed_usage_candidate = match observed_usage_candidate_result {
+        Ok(candidate) => Some(Bytes::from(candidate)),
+        Err(EcstoreError::ConfigNotFound) => None,
+        Err(err) => {
+            debug!(
+                target: "rustfs::scanner",
+                event = EVENT_SCANNER_PERSIST_STATE,
+                component = LOG_COMPONENT_SCANNER,
+                subsystem = LOG_SUBSYSTEM_RUNTIME,
+                cycle = cycle_info.current,
+                path = %DATA_USAGE_OBSERVED_OBJ_NAME_PATH.as_str(),
+                state = "observed_candidate_load_failed",
+                error = %err,
+                "Scanner skipped an unavailable observed usage candidate for scoped refresh"
+            );
+            None
+        }
+    };
    let (sender, receiver) = mpsc::channel::<DataUsageInfo>(1);

    let done_cycle = Metrics::time(Metric::ScanCycle);
@@ -1721,7 +1783,9 @@ where
            scan_mode,
            scan_scope: crate::scanner_io::ScannerBucketScanScope::default(),
            persisted_usage_baseline: usage_persist_baseline.data.clone(),
-            requires_full_scan,
+            observed_usage_candidate,
+            requires_full_scan: scheduling.requires_full_scan,
+            service_cohort: scheduling.service_cohort,
            #[cfg(test)]
            resolved_scope_observer: None,
        },
@@ -1846,10 +1910,10 @@ where
        .as_ref()
        .map(|(notification_system, grants)| (Arc::clone(notification_system), grants.clone()));
    let remote_lease_release_safe = Arc::new(AtomicBool::new(true));
-    let mut usage_persist_outcome = match publication_defer_reason {
+    let mut usage_publication_result = match publication_defer_reason {
        Some(reason) => {
            drop(receiver);
-            DataUsagePersistOutcome::Deferred(reason)
+            DataUsagePublicationResult::from(DataUsagePersistOutcome::Deferred(reason))
        }
        None => {
            // ScannerIO emits its complete or observational update only after
@@ -1864,6 +1928,11 @@ where
                .as_ref()
                .map(|(_, grants)| grants.iter().map(|grant| grant.lease.token).collect())
                .unwrap_or_default();
+            let ack_expectation = scan_result
+                .as_ref()
+                .ok()
+                .filter(|result| result.has_dirty_usage_to_acknowledge())
+                .and_then(ScannerCycleResult::publication_expectation);
            let mut usage_persist_task = AbortOnDropHandle::new(tokio::spawn(async move {
                store_data_usage_in_backend_with_outcome_for_epoch_and_baseline_and_route_probe_for_publication_epoch_and_lease_fence(
                    ctx_clone,
@@ -1876,6 +1945,7 @@ where
                        remote_lease_deadline,
                        remote_lease_fence,
                    )
+                    .with_ack_expectation(ack_expectation)
                    .with_remote_lease_tokens(remote_lease_tokens)
                    .with_lease_release_flag(remote_lease_release_safe_for_task),
                    move || {
@@ -1909,7 +1979,7 @@ where
                        error = %err,
                        "Scanner data usage persistence task failed"
                    );
-                    DataUsagePersistOutcome::Failed
+                    DataUsagePublicationResult::from(DataUsagePersistOutcome::Failed)
                }
                DataUsagePersistTaskResult::Cancelled => {
                    debug!(
@@ -1921,7 +1991,7 @@ where
                        state = "usage_persist_task_cancelled",
                        "Scanner data usage persistence task cancelled"
                    );
-                    DataUsagePersistOutcome::Failed
+                    DataUsagePublicationResult::from(DataUsagePersistOutcome::Failed)
                }
                DataUsagePersistTaskResult::TimedOut => {
                    error!(
@@ -1934,11 +2004,12 @@ where
                        state = "usage_persist_task_timed_out",
                        "Scanner data usage persistence task timed out"
                    );
-                    DataUsagePersistOutcome::Failed
+                    DataUsagePublicationResult::from(DataUsagePersistOutcome::Failed)
                }
            }
        }
    };
+    let mut usage_persist_outcome = usage_publication_result.outcome();
    let lease_expired = remote_publication_leases
        .as_ref()
        .is_some_and(|(_, grants)| grants.iter().any(|grant| !grant.lease.is_valid()));
@@ -2138,16 +2209,14 @@ where
        };
    }

+    usage_publication_result.restrict_outcome(usage_persist_outcome);
    let (completion_outcome, scanner_pending_maintenance_work, remote_dirty_usage_acknowledgements) =
-        finalize_scanner_cycle_result(scan_cycle_result, usage_persist_outcome);
+        finalize_scanner_cycle_result(scan_cycle_result, usage_publication_result);
    let remote_dirty_usage_pending = if remote_dirty_usage_acknowledgements.is_empty() {
        false
    } else if let Some(notification_system) = storeapi.scanner_notification_system() {
        let acknowledgement_count = remote_dirty_usage_acknowledgements.len();
-        let acknowledgements = remote_dirty_usage_acknowledgements
-            .into_iter()
-            .map(|acknowledgement| (acknowledgement.host, acknowledgement.instance_id, acknowledgement.generation))
-            .collect();
+        let acknowledgements = remote_dirty_usage_acknowledgements.into_iter().map(Into::into).collect();
        remote_dirty_usage_acknowledgement_pending(
            cycle_info.current,
            acknowledgement_count,
@@ -2583,6 +2652,7 @@ where
    let mut clean_idle_backoff = ScannerCleanIdleBackoff::default();
    let mut superseded_backoff = ScannerRetryBackoff::default();
    let mut deferred_backoff = ScannerRetryBackoff::default();
+    let service_cohort = Arc::new(StdMutex::new(crate::scanner_io::ScannerServiceCohort::default()));
    let initial_runtime_config = resolve_scanner_runtime_config();
    if clean_idle_topology_supported && maintenance_generation_seen.is_none() {
        let Some((features, generation)) = detect_stable_scanner_maintenance_features(&ctx, &storeapi).await else {
@@ -2792,7 +2862,10 @@ where
                &mut cycle_revision,
                leader_epoch,
                cycle_budget.clone(),
-                true,
+                ScannerCycleScheduling {
+                    requires_full_scan: true,
+                    service_cohort: Some(service_cohort.clone()),
+                },
            ),
            guard.lock_lost_notified(),
        )
@@ -3083,11 +3156,14 @@ where
                &mut cycle_revision,
                leader_epoch,
                cycle_budget.clone(),
-                maintenance_features.requires_full_scan(
-                    maintenance_generation_seen,
-                    scanner_maintenance_generation(),
-                    wake_reason,
-                ),
+                ScannerCycleScheduling {
+                    requires_full_scan: maintenance_features.requires_full_scan(
+                        maintenance_generation_seen,
+                        scanner_maintenance_generation(),
+                        wake_reason,
+                    ),
+                    service_cohort: Some(service_cohort.clone()),
+                },
            ),
            guard.lock_lost_notified(),
        )
@@ -3366,21 +3442,35 @@ fn scanner_cycle_completion_outcome(

 fn finalize_scanner_cycle_result(
    scan_cycle_result: crate::scanner_io::ScannerCycleResult,
-    usage_persist_outcome: DataUsagePersistOutcome,
+    publication: DataUsagePublicationResult,
 ) -> (ScannerCycleOutcome, bool, Vec<ScannerDirtyUsageAcknowledgement>) {
+    let (usage_persist_outcome, proof) = publication.into_parts();
    let completion_outcome = scanner_cycle_completion_outcome_for_result(&scan_cycle_result, usage_persist_outcome);
-    let pending_maintenance_work = scan_cycle_result.has_pending_maintenance_work();
    let durable_complete_snapshot = scan_cycle_result.status == ScannerCycleStatus::Complete
        && matches!(
            usage_persist_outcome,
            DataUsagePersistOutcome::Saved | DataUsagePersistOutcome::AlreadyDurable
-        );
+        )
+        && scan_cycle_result.publication_expectation().as_ref().is_some_and(|expected| {
+            proof
+                .as_ref()
+                .is_some_and(|proof| proof.verified_version_for(expected).is_some())
+        });
+    let pending_maintenance_work = scan_cycle_result.has_pending_maintenance_work()
+        || (scan_cycle_result.has_dirty_usage_to_acknowledge() && !durable_complete_snapshot);
    let remote_dirty_usage_acknowledgements = if durable_complete_snapshot {
-        scan_cycle_result.acknowledge_durable_usage()
+        match proof {
+            Some(proof) => scan_cycle_result.acknowledge_durable_usage(&proof),
+            None => Vec::new(),
+        }
    } else {
        Vec::new()
    };
-    (completion_outcome, pending_maintenance_work, remote_dirty_usage_acknowledgements)
+    (
+        completion_outcome,
+        pending_maintenance_work || crate::scanner_io::dirty_usage_buckets_pending(),
+        remote_dirty_usage_acknowledgements,
+    )
 }

 fn scanner_cycle_completion_outcome_for_result(
@@ -3480,13 +3570,17 @@ use activity::*;
 use backlog::*;
 use cycle_state::*;
 use leadership::*;
+pub(crate) use usage_store::RootPublicationProof;
 use usage_store::*;

+#[cfg(test)]
+pub(crate) use activity::scanner_node_activity_for_tests;
 pub use activity::scanner_topology_digest;
 pub(crate) use activity::{
-    ScannerActivitySnapshot, ScannerDirtyUsageAcknowledgement, probe_scanner_activity, scanner_activity_allows_usage_publication,
-    scanner_activity_dirty_usage_state_for_host, scanner_activity_publication_lease_targets, scanner_activity_snapshot_digest,
-    scanner_activity_structural_digest, scanner_dirty_usage_acknowledgements,
+    ScannerActivitySnapshot, ScannerDirtyUsageAcknowledgement, ScannerDirtyUsageAcknowledgementKind, probe_scanner_activity,
+    scanner_activity_allows_usage_publication, scanner_activity_dirty_usage_state_for_host,
+    scanner_activity_publication_lease_targets, scanner_activity_snapshot_digest, scanner_activity_structural_digest,
+    scanner_dirty_usage_acknowledgements,
 };
 pub(crate) use activity::{ScannerCycleOutcome, scanner_cycle_outcome_with_pending_maintenance};
 pub use backlog::{
@@ -441,11 +441,59 @@ pub(crate) struct ScannerNodeActivity {

 pub(crate) type ScannerActivitySnapshot = BTreeMap<String, ScannerNodeActivity>;

+#[cfg(test)]
+pub(crate) fn scanner_node_activity_for_tests(
+    instance_id: &str,
+    namespace_generation: u64,
+    dirty_usage_generation: u64,
+    dirty_usage_pending: bool,
+) -> ScannerNodeActivity {
+    ScannerNodeActivity {
+        instance_id: instance_id.to_string(),
+        namespace_generation,
+        maintenance_generation: 0,
+        protocol_version: SCANNER_ACTIVITY_PROTOCOL_VERSION,
+        topology_digest: [0; 32],
+        data_movement_active: false,
+        dirty_usage_generation,
+        dirty_usage_pending,
+        movement_generation: 0,
+        publication_blocked: false,
+    }
+}
+
 #[derive(Clone, Debug, PartialEq, Eq)]
 pub(crate) struct ScannerDirtyUsageAcknowledgement {
    pub(crate) host: String,
    pub(crate) instance_id: String,
-    pub(crate) generation: u64,
+    pub(crate) kind: ScannerDirtyUsageAcknowledgementKind,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(crate) enum ScannerDirtyUsageAcknowledgementKind {
+    Generation(u64),
+    Scoped {
+        owner_id: String,
+        entries: Vec<crate::storage_api::EcstoreScannerScopedDirtyUsageAckEntry>,
+    },
+}
+
+impl From<ScannerDirtyUsageAcknowledgement> for crate::storage_api::EcstoreScannerDirtyUsageAcknowledgement {
+    fn from(acknowledgement: ScannerDirtyUsageAcknowledgement) -> Self {
+        match acknowledgement.kind {
+            ScannerDirtyUsageAcknowledgementKind::Generation(generation) => Self::Generation {
+                host: acknowledgement.host,
+                instance_id: acknowledgement.instance_id,
+                generation,
+            },
+            ScannerDirtyUsageAcknowledgementKind::Scoped { owner_id, entries } => Self::Scoped {
+                host: acknowledgement.host,
+                owner_id,
+                instance_id: acknowledgement.instance_id,
+                entries,
+            },
+        }
+    }
 }

 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
@@ -974,7 +1022,7 @@ pub(crate) fn scanner_dirty_usage_acknowledgements(snapshot: &ScannerActivitySna
        .map(|(host, activity)| ScannerDirtyUsageAcknowledgement {
            host: host.clone(),
            instance_id: activity.instance_id.clone(),
-            generation: activity.dirty_usage_generation,
+            kind: ScannerDirtyUsageAcknowledgementKind::Generation(activity.dirty_usage_generation),
        })
        .collect()
 }
@@ -18,6 +18,7 @@ use crate::data_usage_define::{
    DATA_USAGE_BLOOM_RECOVERY_PATH, DATA_USAGE_RECOVERY_PATH, usage_floor_primary_read_error_allows_backup,
 };
 use crate::storage_api::owner::ObjectIO as _;
+use std::sync::atomic::AtomicU64;
 use tokio::io::AsyncReadExt as _;

 const SCANNER_CYCLE_RECOVERY_SCHEMA_VERSION: u16 = 1;
@@ -34,6 +35,86 @@ const CACHE_CYCLE_AHEAD: &str = "cache_cycle_ahead";

 const SCANNER_USAGE_STATE_RESET_MODE_FULL_REBUILD: &str = "full-rebuild";

+#[cfg(test)]
+pub(super) mod cleanup_io_fault {
+    use super::*;
+
+    #[derive(Clone, Copy, PartialEq, Eq)]
+    pub(in crate::scanner) enum Stage {
+        PrimaryRead,
+        PrimaryWrite,
+        UsageFence,
+    }
+
+    struct Injection {
+        store: std::sync::Weak<ECStore>,
+        stage: Stage,
+        fired: AtomicBool,
+        owned: AtomicBool,
+        newer_completion: bool,
+    }
+
+    static INJECTION: StdMutex<Option<Arc<Injection>>> = StdMutex::new(None);
+    pub(in crate::scanner) struct Guard(Arc<Injection>);
+
+    impl Guard {
+        pub(in crate::scanner) fn fired_while_owned(&self) -> bool {
+            self.0.fired.load(Ordering::Relaxed) && self.0.owned.load(Ordering::Relaxed)
+        }
+    }
+
+    impl Drop for Guard {
+        fn drop(&mut self) {
+            let mut slot = INJECTION.lock().unwrap_or_else(|poisoned| poisoned.into_inner());
+            if slot.as_ref().is_some_and(|current| Arc::ptr_eq(current, &self.0)) {
+                *slot = None;
+            }
+        }
+    }
+
+    pub(in crate::scanner) fn install(store: &Arc<ECStore>, stage: Stage, newer_completion: bool) -> Guard {
+        let injection = Arc::new(Injection {
+            store: Arc::downgrade(store),
+            stage,
+            fired: AtomicBool::new(false),
+            owned: AtomicBool::new(false),
+            newer_completion,
+        });
+        let mut slot = INJECTION.lock().expect("cleanup injection slot");
+        assert!(slot.is_none(), "only one cleanup I/O injection may be installed");
+        *slot = Some(injection.clone());
+        Guard(injection)
+    }
+
+    pub(super) fn check(store: &Arc<ECStore>, stage: Stage, owned: bool) -> Result<(), ScannerError> {
+        let injection = {
+            let mut slot = INJECTION.lock().expect("cleanup injection slot");
+            if slot
+                .as_ref()
+                .is_some_and(|injection| injection.stage == stage && injection.store.ptr_eq(&Arc::downgrade(store)))
+            {
+                slot.take()
+            } else {
+                None
+            }
+        };
+        let Some(injection) = injection else {
+            return Ok(());
+        };
+        injection.fired.store(true, Ordering::Relaxed);
+        injection.owned.store(owned, Ordering::Relaxed);
+        if injection.newer_completion {
+            set_scanner_cycle_recovery_status(recovery_status("healthy", None, false));
+        }
+        let reason = match stage {
+            Stage::PrimaryRead => "injected primary read failure",
+            Stage::PrimaryWrite => "injected primary write failure",
+            Stage::UsageFence => "injected usage fence failure",
+        };
+        Err(ScannerError::Io(std::io::Error::other(reason)))
+    }
+}
+
 #[derive(Clone, Debug, Default, Serialize)]
 pub struct ScannerCycleRecoveryStatus {
    /// The immutable primary object whose revision is being guarded.
@@ -81,6 +162,8 @@ static SCANNER_CYCLE_RECOVERY_STATUS: LazyLock<RwLock<ScannerCycleRecoveryStatus
        ..Default::default()
    })
 });
+// An old startup observation must not overwrite a newer explicit reset status.
+static SCANNER_CYCLE_RECOVERY_STATUS_VERSION: AtomicU64 = AtomicU64::new(0);

 pub fn scanner_cycle_recovery_status() -> ScannerCycleRecoveryStatus {
    SCANNER_CYCLE_RECOVERY_STATUS
@@ -90,6 +173,24 @@ pub fn scanner_cycle_recovery_status() -> ScannerCycleRecoveryStatus {
 }

 fn set_scanner_cycle_recovery_status(status: ScannerCycleRecoveryStatus) {
+    let _ = publish_scanner_cleanup_status(status, None);
+}
+
+pub(super) fn scanner_cleanup_status_version() -> u64 {
+    let _status = SCANNER_CYCLE_RECOVERY_STATUS
+        .read()
+        .unwrap_or_else(|poisoned| poisoned.into_inner());
+    SCANNER_CYCLE_RECOVERY_STATUS_VERSION.load(Ordering::Relaxed)
+}
+
+pub(super) fn publish_scanner_cleanup_status(status: ScannerCycleRecoveryStatus, expected: Option<u64>) -> Option<u64> {
+    let mut current = SCANNER_CYCLE_RECOVERY_STATUS
+        .write()
+        .unwrap_or_else(|poisoned| poisoned.into_inner());
+    let version = SCANNER_CYCLE_RECOVERY_STATUS_VERSION.load(Ordering::Relaxed);
+    if expected.is_some_and(|expected| expected == u64::MAX || expected != version) {
+        return None;
+    }
    let recovery_required = if matches!(
        status.state.as_str(),
        "blocked"
@@ -106,9 +207,27 @@ fn set_scanner_cycle_recovery_status(status: ScannerCycleRecoveryStatus) {
    };
    metrics::gauge!(METRIC_SCANNER_CYCLE_RECOVERY_REQUIRED).set(recovery_required);
    metrics::gauge!(METRIC_SCANNER_CYCLE_RECOVERY_RETRY_COUNT).set(status.retry_count as f64);
-    *SCANNER_CYCLE_RECOVERY_STATUS
-        .write()
-        .unwrap_or_else(|poisoned| poisoned.into_inner()) = status;
+    *current = status;
+    let next = version.saturating_add(1);
+    SCANNER_CYCLE_RECOVERY_STATUS_VERSION.store(next, Ordering::Relaxed);
+    Some(next)
+}
+
+fn publish_scanner_cleanup_failure(reason: String, expected: u64) {
+    let mut status = {
+        let current = SCANNER_CYCLE_RECOVERY_STATUS
+            .read()
+            .unwrap_or_else(|poisoned| poisoned.into_inner());
+        if SCANNER_CYCLE_RECOVERY_STATUS_VERSION.load(Ordering::Relaxed) != expected {
+            return;
+        }
+        current.clone()
+    };
+    // Preserve the latest core progress, including a newly written primary's
+    // revision and epoch. The original marker may predate that durable write.
+    status.reason = Some(reason);
+    status.last_attempt_at_unix_secs = Some(unix_now_secs());
+    let _ = publish_scanner_cleanup_status(status, Some(expected));
 }

 pub(super) fn record_scanner_usage_floor_failure(reason: String) {
@@ -932,10 +1051,81 @@ pub(crate) async fn load_scanner_cycle_state_for_startup(
    }
 }

-/// Reset a blocked cycle state after an operator has explicitly requested a full
-/// usage rebuild. The primary object is changed first with its observed ETag;
-/// the recovery marker is removed only when its own ETag still matches.
+/// Reset a blocked cycle state after an explicit full-rescan request. Durable
+/// cleanup state fences primary rewrites; marker removal retains its ETag and
+/// usage-epoch checks.
 pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<ECStore>) -> Result<(), ScannerError> {
+    reset_scanner_cycle_recovery_for_intent(ctx, storeapi, None, None).await
+}
+
+/// Resume only an operator reset whose cleanup phase is already durable.
+/// Missing or merely blocked markers never authorize an automatic reset.
+pub(super) async fn resume_scanner_cycle_cleanup(ctx: CancellationToken, storeapi: Arc<ECStore>) -> Result<(), ScannerError> {
+    let status_version = scanner_cleanup_status_version();
+    let (marker, revision) = match read_scanner_cleanup_marker(storeapi.clone(), &ctx).await {
+        Ok(Some(marker)) => marker,
+        Ok(None) => return Ok(()),
+        Err(error) => {
+            let _ =
+                publish_scanner_cleanup_status(recovery_status("blocked", Some(&error.to_string()), false), Some(status_version));
+            return Err(error);
+        }
+    };
+    let mut observation =
+        publish_scanner_cleanup_status(recovery_status_from_marker(&marker, &marker.state), Some(status_version));
+    if marker.state != "cleanup-pending" {
+        return Ok(());
+    }
+    let result = reset_scanner_cycle_recovery_for_intent(ctx, storeapi, Some(revision), Some(&mut observation)).await;
+    if let Err(error) = &result
+        && let Some(observation) = observation
+    {
+        publish_scanner_cleanup_failure(error.to_string(), observation);
+    }
+    result
+}
+
+pub(super) async fn read_scanner_cleanup_marker(
+    storeapi: Arc<impl ScannerObjectIO>,
+    ctx: &CancellationToken,
+) -> Result<Option<(ScannerCycleRecoveryMarker, DataUsageCacheRevision)>, ScannerError> {
+    let (data, revision) = tokio::select! {
+        biased;
+        _ = ctx.cancelled() => return Err(ScannerError::Other("scanner cleanup recovery was cancelled".to_string())),
+        result = tokio::time::timeout(data_usage_persist_timeout(), read_cycle_recovery_marker_bytes(storeapi)) => {
+            result.map_err(|_| ScannerError::Other("scanner cleanup marker inspection timed out".to_string()))?
+                .map_err(|err| ScannerError::Other(format!("failed to inspect pending scanner cleanup: {err}")))?
+        }
+    };
+    let Some(data) = data else {
+        return Ok(None);
+    };
+    let marker: ScannerCycleRecoveryMarker = serde_json::from_slice(&data)
+        .map_err(|err| ScannerError::Other(format!("pending scanner cleanup marker is invalid: {err}")))?;
+    validate_recovery_marker(&marker)
+        .map_err(|err| ScannerError::Other(format!("pending scanner cleanup marker is invalid: {err}")))?;
+    Ok(Some((marker, revision)))
+}
+
+pub(super) async fn reset_scanner_cycle_recovery_for_intent(
+    ctx: CancellationToken,
+    storeapi: Arc<ECStore>,
+    expected_cleanup_revision: Option<DataUsageCacheRevision>,
+    mut observation: Option<&mut Option<u64>>,
+) -> Result<(), ScannerError> {
+    #[cfg(test)]
+    let resume_only = expected_cleanup_revision.is_some();
+    // The outer Some distinguishes a tracked resume whose version may be
+    // invalidated from an explicit v3 reset with no observation owner.
+    let mut publish_status = |status| {
+        if let Some(version) = observation.as_deref_mut() {
+            if let Some(expected) = *version {
+                *version = publish_scanner_cleanup_status(status, Some(expected));
+            }
+        } else {
+            set_scanner_cycle_recovery_status(status);
+        }
+    };
    let lock = storeapi
        .new_ns_lock(RUSTFS_META_BUCKET, "leader.lock")
        .await
@@ -964,6 +1154,21 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
        }
        Err(err) => return Err(ScannerError::Other(format!("failed to read cycle recovery marker: {err}"))),
    };
+    if let Some(expected) = expected_cleanup_revision {
+        if marker_revision != expected || !owns_reset() {
+            return Err(ScannerError::Other(
+                "pending scanner cleanup changed before recovery acquired ownership".to_string(),
+            ));
+        }
+        let marker: ScannerCycleRecoveryMarker = marker_data
+            .as_deref()
+            .and_then(|data| serde_json::from_slice(data).ok())
+            .filter(|marker| validate_recovery_marker(marker).is_ok() && marker.state == "cleanup-pending")
+            .ok_or_else(|| {
+                ScannerError::Other("scanner cleanup recovery requires an unchanged cleanup-pending marker".to_string())
+            })?;
+        publish_status(recovery_status_from_marker(&marker, "cleanup-pending"));
+    }
    let Some(marker_data) = marker_data else {
        // A delete may commit before its reply is lost. Confirm both durable
        // fences before treating a retry without its marker as completed.
@@ -981,7 +1186,7 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
                "scanner cycle recovery marker is absent without a completed reset fence".to_string(),
            ));
        }
-        set_scanner_cycle_recovery_status(recovery_status("healthy", None, false));
+        publish_status(recovery_status("healthy", None, false));
        super::notify_scanner_cycle_recovery_wake();
        return Ok(());
    };
@@ -997,6 +1202,10 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
        ));
    }

+    #[cfg(test)]
+    if resume_only {
+        cleanup_io_fault::check(&storeapi, cleanup_io_fault::Stage::PrimaryRead, owns_reset())?;
+    }
    let (mut primary_reader, primary_revision) = match storeapi
        .get_object_reader(
            RUSTFS_META_BUCKET,
@@ -1062,7 +1271,7 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
            let (cleanup_marker, cleanup_marker_revision) =
                mark_cycle_recovery_cleanup_pending(storeapi.clone(), marker.clone(), &marker_revision, reset_epoch, &owns_reset)
                    .await?;
-            set_scanner_cycle_recovery_status(recovery_status_from_marker(&cleanup_marker, "cleanup-pending"));
+            publish_status(recovery_status_from_marker(&cleanup_marker, "cleanup-pending"));
            let usage_floor = persisted_usage_floor(storeapi.clone()).await?;
            let fence_epoch = primary_epoch
                .max(usage_floor.leader_epoch)
@@ -1082,6 +1291,10 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
                ));
            }
            verify_cycle_reset_intent(storeapi.clone(), &cleanup_marker_revision, &owns_reset).await?;
+            #[cfg(test)]
+            if resume_only {
+                cleanup_io_fault::check(&storeapi, cleanup_io_fault::Stage::PrimaryWrite, owns_reset())?;
+            }
            let preserved_info = save_reset_config(
                storeapi.clone(),
                DATA_USAGE_BLOOM_NAME_PATH.as_str(),
@@ -1155,7 +1368,7 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
                    ScannerError::Other(format!("failed to clear stale cycle recovery marker: {err}"))
                }
            })?;
-            set_scanner_cycle_recovery_status(recovery_status("healthy", None, false));
+            publish_status(recovery_status("healthy", None, false));
            super::notify_scanner_cycle_recovery_wake();
            return Ok(());
        } else if !force_full_rescan && !marker_cleanup_pending {
@@ -1228,11 +1441,23 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
        ));
    }
    verify_cycle_reset_intent(storeapi.clone(), &marker_revision, &owns_reset).await?;
-    if let Err(err) =
-        fence_scanner_usage_epoch_with_expected_epoch(&ctx, storeapi.clone(), leader_epoch, Some(reset_epoch), false, &owns_reset)
-            .await
-    {
-        set_scanner_cycle_recovery_status(ScannerCycleRecoveryStatus {
+    let usage_fence = fence_scanner_usage_epoch_with_expected_epoch(
+        &ctx,
+        storeapi.clone(),
+        leader_epoch,
+        Some(reset_epoch),
+        false,
+        &owns_reset,
+    );
+    #[cfg(test)]
+    let usage_fence = async {
+        if resume_only {
+            cleanup_io_fault::check(&storeapi, cleanup_io_fault::Stage::UsageFence, owns_reset())?;
+        }
+        usage_fence.await
+    };
+    if let Err(err) = usage_fence.await {
+        publish_status(ScannerCycleRecoveryStatus {
            path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
            quarantine_path: Some(DATA_USAGE_BLOOM_RECOVERY_PATH.clone()),
            state: "cleanup-pending".to_string(),
@@ -1260,7 +1485,7 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
        }
    };
    if current_revision != rebuilt_revision {
-        set_scanner_cycle_recovery_status(ScannerCycleRecoveryStatus {
+        publish_status(ScannerCycleRecoveryStatus {
            path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
            quarantine_path: Some(DATA_USAGE_BLOOM_RECOVERY_PATH.clone()),
            state: "cleanup-pending".to_string(),
@@ -1281,7 +1506,7 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
    }

    if guard.is_lock_lost() {
-        set_scanner_cycle_recovery_status(ScannerCycleRecoveryStatus {
+        publish_status(ScannerCycleRecoveryStatus {
            path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
            quarantine_path: Some(DATA_USAGE_BLOOM_RECOVERY_PATH.clone()),
            state: "cleanup-pending".to_string(),
@@ -1318,7 +1543,7 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
    .await
    {
        if scanner_publication_epoch_changed(&err) {
-            set_scanner_cycle_recovery_status(ScannerCycleRecoveryStatus {
+            publish_status(ScannerCycleRecoveryStatus {
                path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
                quarantine_path: Some(DATA_USAGE_BLOOM_RECOVERY_PATH.clone()),
                state: "cleanup-pending".to_string(),
@@ -1336,7 +1561,7 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
                "scanner recovery reset deferred by a movement epoch change".to_string(),
            ));
        }
-        set_scanner_cycle_recovery_status(ScannerCycleRecoveryStatus {
+        publish_status(ScannerCycleRecoveryStatus {
            path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
            quarantine_path: Some(DATA_USAGE_BLOOM_RECOVERY_PATH.clone()),
            state: "cleanup-pending".to_string(),
@@ -1352,7 +1577,7 @@ pub async fn reset_scanner_cycle_recovery(ctx: CancellationToken, storeapi: Arc<
        });
        return Err(ScannerError::Other(format!("failed to clear cycle recovery marker: {err}")));
    }
-    set_scanner_cycle_recovery_status(ScannerCycleRecoveryStatus {
+    publish_status(ScannerCycleRecoveryStatus {
        path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
        quarantine_path: Some(DATA_USAGE_BLOOM_RECOVERY_PATH.clone()),
        state: "healthy".to_string(),
@@ -27,6 +27,7 @@ use crate::{
 use serial_test::serial;
 use std::collections::{HashMap, HashSet};
 use std::io::Cursor;
+use std::path::Path;
 use std::sync::atomic::{AtomicBool, AtomicU64, AtomicUsize, Ordering};
 use std::task::Poll;
 use temp_env::{with_var, with_var_unset};
@@ -36,6 +37,11 @@ use tokio::time::{Duration, advance};

 const TEST_DEFAULT_SCANNER_CYCLE_SECS: u64 = 24 * 60 * 60;

+mod quota_reset_preservation;
+
+mod recovery_control;
+mod scoped_ack_publication;
+
 async fn setup_scanner_cycle_store() -> (tempfile::TempDir, Arc<ECStore>) {
    setup_scanner_cycle_store_with_usage_baseline(true).await
 }
@@ -50,11 +56,17 @@ async fn setup_scanner_cycle_store_with_pool_count(
 ) -> (tempfile::TempDir, Arc<ECStore>) {
    init_ecstore_config_for_scanner_tests();
    let temp_dir = tempfile::tempdir().expect("scanner cycle test directory should be created");
+    let store = setup_scanner_cycle_store_at_path(temp_dir.path(), seed_usage_baseline, pool_count).await;
+    (temp_dir, store)
+}
+
+async fn setup_scanner_cycle_store_at_path(root: &Path, seed_usage_baseline: bool, pool_count: usize) -> Arc<ECStore> {
+    init_ecstore_config_for_scanner_tests();
    let mut pools = Vec::with_capacity(pool_count);
    for pool_index in 0..pool_count {
        let mut endpoints = Vec::new();
        for disk_index in 0..4 {
-            let disk_path = temp_dir.path().join(format!("pool{pool_index}/disk{disk_index}"));
+            let disk_path = root.join(format!("pool{pool_index}/disk{disk_index}"));
            tokio::fs::create_dir_all(&disk_path)
                .await
                .expect("scanner cycle test disk should be created");
@@ -104,7 +116,7 @@ async fn setup_scanner_cycle_store_with_pool_count(
        .expect("scanner cycle usage baseline should persist");
    }

-    (temp_dir, store)
+    store
 }

 async fn restart_scanner_cycle_store_from(store: &Arc<ECStore>) -> Arc<ECStore> {
@@ -1219,7 +1231,18 @@ async fn coordinator_walks_during_pending_put_without_persisting_or_acknowledgin
    let mut revision = DataUsageCacheRevision::Missing;
    let outcome = tokio::time::timeout(
        Duration::from_secs(30),
-        run_data_scanner_cycle_with_budget(&ctx, &store, &mut cycle_info, &mut revision, 1, Arc::clone(&budget), true),
+        run_data_scanner_cycle_with_budget(
+            &ctx,
+            &store,
+            &mut cycle_info,
+            &mut revision,
+            1,
+            Arc::clone(&budget),
+            ScannerCycleScheduling {
+                requires_full_scan: true,
+                service_cohort: None,
+            },
+        ),
    )
    .await
    .expect("the coordinator must finish its namespace walk while a PUT is pending");
@@ -1263,7 +1286,18 @@ async fn coordinator_walks_during_pending_put_without_persisting_or_acknowledgin
    let retry_budget = ScannerCycleBudget::new_with_progress_tracking(&ctx, ScannerCycleBudgetConfig::default());
    let outcome = tokio::time::timeout(
        Duration::from_secs(30),
-        run_data_scanner_cycle_with_budget(&ctx, &store, &mut cycle_info, &mut revision, 1, Arc::clone(&retry_budget), true),
+        run_data_scanner_cycle_with_budget(
+            &ctx,
+            &store,
+            &mut cycle_info,
+            &mut revision,
+            1,
+            Arc::clone(&retry_budget),
+            ScannerCycleScheduling {
+                requires_full_scan: true,
+                service_cohort: None,
+            },
+        ),
    )
    .await
    .expect("the same cycle must converge after the pending PUT drains");
@@ -5264,6 +5298,103 @@ async fn scanner_usage_state_reset_resumes_every_cleanup_boundary_without_rewrit
    }
 }

+#[tokio::test]
+#[serial]
+async fn scanner_usage_state_reset_resumes_real_store_cleanup_boundaries_after_reopen() {
+    let primary_path = DATA_USAGE_OBJ_NAME_PATH.as_str();
+    let cleanup_paths = [
+        format!("{primary_path}.bkp"),
+        LEGACY_DATA_USAGE_OBJ_NAME_PATH.as_str().to_string(),
+        format!("{}.bkp", LEGACY_DATA_USAGE_OBJ_NAME_PATH.as_str()),
+        DATA_USAGE_OBSERVED_OBJ_NAME_PATH.as_str().to_string(),
+    ];
+
+    for completed in 0..=cleanup_paths.len() {
+        let (_temp_dir, store) = setup_scanner_cycle_store().await;
+        let cycle = CurrentCycle {
+            current: 12,
+            next: 42,
+            cycle_completed: vec![Utc::now()],
+            started: Utc::now(),
+        };
+        save_config(
+            store.clone(),
+            DATA_USAGE_BLOOM_NAME_PATH.as_str(),
+            encode_scanner_cycle_state(&cycle, 3).expect("cycle state should encode"),
+        )
+        .await
+        .expect("cycle state should persist");
+        let marker = scanner_usage_bootstrap_marker(std::time::SystemTime::UNIX_EPOCH, Some(3));
+        save_config(
+            store.clone(),
+            primary_path,
+            serde_json::to_vec(&marker).expect("usage reset marker should encode"),
+        )
+        .await
+        .expect("usage reset marker should persist");
+
+        for path in cleanup_paths.iter().skip(completed) {
+            let mut usage = complete_usage_with_bucket_count(Some(std::time::SystemTime::UNIX_EPOCH), 0);
+            usage.scanner_epoch = Some(1);
+            usage.scanner_cycle = Some(12);
+            save_config(store.clone(), path, serde_json::to_vec(&usage).expect("cleanup slot should encode"))
+                .await
+                .expect("cleanup slot should persist");
+        }
+        for path in ["buckets/quota-reservations/ledger", "buckets/example/incarnation"] {
+            save_config(store.clone(), path, b"retain".to_vec())
+                .await
+                .expect("unrelated state should persist before reopen");
+        }
+
+        let restarted = restart_scanner_cycle_store_from(&store).await;
+        let intent_before = read_config_with_revision(restarted.clone(), primary_path)
+            .await
+            .expect("reopened reset intent should be readable");
+
+        let result = reset_scanner_usage_state_for_full_rebuild(CancellationToken::new(), restarted.clone())
+            .await
+            .expect("reopened usage reset should complete");
+
+        assert_eq!(result.leader_epoch, 3, "boundary {completed}");
+        assert_eq!(result.next_cycle, 42, "boundary {completed}");
+        assert_eq!(result.reset_paths.len(), cleanup_paths.len() + 1 - completed, "boundary {completed}");
+        assert_eq!(
+            read_config_with_revision(restarted.clone(), primary_path)
+                .await
+                .expect("completed reset intent should remain readable"),
+            intent_before,
+            "boundary {completed}: resumed cleanup must not rewrite the reset intent"
+        );
+
+        let (floor, state) = persisted_usage_floor_for_startup(restarted.clone(), false)
+            .await
+            .expect("completed reset marker should remain resumable");
+        assert_eq!(floor.leader_epoch, 3, "boundary {completed}");
+        assert_eq!(state, PersistedUsageFloorStartup::BootstrapPending, "boundary {completed}");
+        assert!(
+            persisted_usage_floor(restarted.clone()).await.is_err(),
+            "boundary {completed}: bootstrap marker must not become an authoritative floor"
+        );
+
+        for path in &cleanup_paths {
+            assert!(
+                matches!(read_config(restarted.clone(), path).await, Err(EcstoreError::ConfigNotFound)),
+                "boundary {completed}: reset should remove stale usage slot {path}"
+            );
+        }
+        for path in ["buckets/quota-reservations/ledger", "buckets/example/incarnation"] {
+            assert_eq!(
+                read_config(restarted.clone(), path)
+                    .await
+                    .expect("unrelated state should survive reopened reset"),
+                b"retain",
+                "boundary {completed}: reset must preserve non-scanner-state config"
+            );
+        }
+    }
+}
+
 #[tokio::test]
 async fn scanner_usage_state_reset_stops_usage_fence_after_owner_loss() {
    let store = Arc::new(MemoryConfigStore::default());
@@ -6160,7 +6291,7 @@ async fn coordinator_classifies_an_expired_publication_lease() {
        .await;

    assert_eq!(
-        outcome,
+        outcome.outcome(),
        DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::PublicationLeaseDeadlineExceeded)
    );
    assert!(store.put_counts.lock().await.is_empty(), "expired lease must prevent a PUT");
@@ -7301,7 +7432,7 @@ fn scanner_cycle_cache_floor_stays_pending_during_deferred_usage_publication() {

 #[test]
 #[serial]
-fn finalizing_a_saved_cycle_acknowledges_its_exact_dirty_snapshot() {
+fn finalizing_a_saved_enum_without_proof_keeps_dirty_pending() {
    crate::scanner_io::clear_dirty_usage_bucket("photos");
    crate::scanner_io::record_dirty_usage_bucket("photos");
    let dirty_snapshot = crate::scanner_io::dirty_usage_buckets_for_tests();
@@ -7309,21 +7440,23 @@ fn finalizing_a_saved_cycle_acknowledges_its_exact_dirty_snapshot() {
    let remote_acknowledgement = ScannerDirtyUsageAcknowledgement {
        host: "node-2".to_string(),
        instance_id: "0123456789abcdef0123456789abcdef".to_string(),
-        generation: 11,
+        kind: ScannerDirtyUsageAcknowledgementKind::Generation(11),
    };
    let unsaved = crate::scanner_io::ScannerCycleResult::new(ScannerCycleStatus::Complete, Some(dirty_snapshot.clone()))
        .with_remote_dirty_usage_acknowledgements(vec![remote_acknowledgement.clone()]);
-    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(unsaved, DataUsagePersistOutcome::NoUpdate);
+    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(unsaved, DataUsagePersistOutcome::NoUpdate.into());
    assert_eq!(outcome, ScannerCycleOutcome::Failed);
    assert!(acknowledgements.is_empty());
    assert!(crate::scanner_io::dirty_usage_buckets_pending());

    let saved = crate::scanner_io::ScannerCycleResult::new(ScannerCycleStatus::Complete, Some(dirty_snapshot))
-        .with_remote_dirty_usage_acknowledgements(vec![remote_acknowledgement.clone()]);
-    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(saved, DataUsagePersistOutcome::Saved);
+        .with_remote_dirty_usage_acknowledgements(vec![remote_acknowledgement]);
+    let (outcome, pending, acknowledgements) = finalize_scanner_cycle_result(saved, DataUsagePersistOutcome::Saved.into());
    assert_eq!(outcome, ScannerCycleOutcome::Completed);
-    assert_eq!(acknowledgements, vec![remote_acknowledgement]);
-    assert!(!crate::scanner_io::dirty_usage_buckets_pending());
+    assert!(acknowledgements.is_empty());
+    assert!(pending);
+    assert!(crate::scanner_io::dirty_usage_buckets_pending());
+    crate::scanner_io::clear_dirty_usage_bucket("photos");
 }

 #[test]
@@ -7335,7 +7468,7 @@ fn finalizing_a_deferred_usage_save_keeps_dirty_work_pending() {
    let deferred = crate::scanner_io::ScannerCycleResult::new(ScannerCycleStatus::Complete, Some(dirty_snapshot));

    let (outcome, _, acknowledgements) =
-        finalize_scanner_cycle_result(deferred, DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement));
+        finalize_scanner_cycle_result(deferred, DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement).into());

    assert_eq!(outcome, ScannerCycleOutcome::Deferred(ScannerCycleDeferReason::DataMovement));
    assert!(acknowledgements.is_empty());
@@ -7355,7 +7488,7 @@ fn finalizing_post_scan_observation_advances_partially_without_dirty_ack() {
    )
    .with_observational_snapshot_published(true);

-    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(observed, DataUsagePersistOutcome::Saved);
+    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(observed, DataUsagePersistOutcome::Saved.into());

    assert_eq!(outcome, ScannerCycleOutcome::Partial);
    assert!(acknowledgements.is_empty());
@@ -7391,17 +7524,20 @@ async fn scanner_cycle_keeps_remote_pending_acknowledgement() {

 #[test]
 #[serial]
-fn finalizing_an_already_durable_cycle_acknowledges_its_exact_dirty_snapshot() {
+fn finalizing_an_already_durable_enum_without_proof_keeps_dirty_pending() {
    crate::scanner_io::clear_dirty_usage_bucket("photos");
    crate::scanner_io::record_dirty_usage_bucket("photos");
    let dirty_snapshot = crate::scanner_io::dirty_usage_buckets_for_tests();

    let durable = crate::scanner_io::ScannerCycleResult::new(ScannerCycleStatus::Complete, Some(dirty_snapshot));
-    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(durable, DataUsagePersistOutcome::AlreadyDurable);
+    let (outcome, pending, acknowledgements) =
+        finalize_scanner_cycle_result(durable, DataUsagePersistOutcome::AlreadyDurable.into());

    assert_eq!(outcome, ScannerCycleOutcome::Completed);
    assert!(acknowledgements.is_empty());
-    assert!(!crate::scanner_io::dirty_usage_buckets_pending());
+    assert!(pending);
+    assert!(crate::scanner_io::dirty_usage_buckets_pending());
+    crate::scanner_io::clear_dirty_usage_bucket("photos");
 }

 #[test]
@@ -7412,7 +7548,8 @@ fn finalizing_a_prior_same_cycle_snapshot_keeps_new_dirty_work_pending() {
    let dirty_snapshot = crate::scanner_io::dirty_usage_buckets_for_tests();

    let durable = crate::scanner_io::ScannerCycleResult::new(ScannerCycleStatus::Complete, Some(dirty_snapshot));
-    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(durable, DataUsagePersistOutcome::PriorCycleDurable);
+    let (outcome, _, acknowledgements) =
+        finalize_scanner_cycle_result(durable, DataUsagePersistOutcome::PriorCycleDurable.into());

    assert_eq!(outcome, ScannerCycleOutcome::Completed);
    assert!(acknowledgements.is_empty());
@@ -7428,7 +7565,7 @@ fn finalizing_a_durable_superseded_snapshot_keeps_dirty_work_pending() {
    let dirty_snapshot = crate::scanner_io::dirty_usage_buckets_for_tests();

    let superseded = crate::scanner_io::ScannerCycleResult::new(ScannerCycleStatus::Superseded, Some(dirty_snapshot));
-    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(superseded, DataUsagePersistOutcome::Saved);
+    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(superseded, DataUsagePersistOutcome::Saved.into());

    assert_eq!(outcome, ScannerCycleOutcome::Superseded);
    assert!(acknowledgements.is_empty());
@@ -8514,6 +8651,56 @@ async fn test_wait_for_next_scanner_cycle_wakes_for_dirty_usage() {
    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
 }

+#[tokio::test(start_paused = true)]
+#[serial]
+async fn service_cohort_aging_preserves_explicit_cycle_wait() {
+    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
+    let config = ScannerRuntimeConfig {
+        cycle_interval: Duration::from_secs(3600),
+        cycle_interval_source: ScannerRuntimeConfigSource::Env,
+        ..Default::default()
+    };
+    let observed = ScannerCycleObservedGenerations::for_wait(
+        &config,
+        None,
+        crate::scanner_io::dirty_usage_generation(),
+        crate::runtime_config::scanner_runtime_config_generation(),
+        crate::scanner_io::scanner_maintenance_generation(),
+    );
+    assert_eq!(observed.dirty_usage, None);
+    let inventory = HashMap::from([(
+        crate::data_usage_define::DataUsageCacheSource::new(0, 0),
+        vec![crate::storage_api::scanner_io::BucketInfo {
+            name: "waiting-bootstrap".to_string(),
+            ..Default::default()
+        }],
+    )]);
+    let mut cohort = crate::scanner_io::ScannerServiceCohort::default();
+    cohort.refresh(&inventory);
+    let ctx = CancellationToken::new();
+    let mut wait = Box::pin(wait_for_next_scanner_cycle(
+        &ctx,
+        config.cycle_interval,
+        observed.dirty_usage,
+        observed.runtime_config,
+        observed.maintenance,
+        || false,
+    ));
+    assert!(matches!(futures::poll!(&mut wait), Poll::Pending));
+    for _ in 0..59 {
+        tokio::time::advance(Duration::from_secs(60)).await;
+        cohort.refresh(&inventory);
+        crate::scanner_io::record_dirty_usage_bucket("hot");
+        assert!(
+            matches!(futures::poll!(&mut wait), Poll::Pending),
+            "aging/dirty must not shorten the explicit hour"
+        );
+    }
+    tokio::time::advance(Duration::from_secs(60)).await;
+    assert_eq!(wait.await, ScannerCycleWakeReason::Timer);
+    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
+}
+
 #[tokio::test]
 #[serial]
 async fn test_wait_for_next_scanner_cycle_sees_unattempted_dirty_usage() {
@@ -8802,12 +8989,12 @@ fn post_lease_activity_proof_rejects_a_put_tail_that_finished_before_lease_acqui
        ScannerDirtyUsageAcknowledgement {
            host: "node-2".to_string(),
            instance_id: "epoch-a".to_string(),
-            generation: 5,
+            kind: ScannerDirtyUsageAcknowledgementKind::Generation(5),
        },
    ]);
    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(
        result,
-        DataUsagePersistOutcome::Deferred(reason.expect("changed namespace should defer publication")),
+        DataUsagePersistOutcome::Deferred(reason.expect("changed namespace should defer publication")).into(),
    );
    assert_eq!(
        outcome,
@@ -8962,7 +9149,7 @@ fn scanner_dirty_usage_acknowledgements_exclude_local_and_clean_nodes() {
        vec![ScannerDirtyUsageAcknowledgement {
            host: "node-3".to_string(),
            instance_id: "epoch-dirty".to_string(),
-            generation: 11,
+            kind: ScannerDirtyUsageAcknowledgementKind::Generation(11),
        }]
    );
 }
@@ -0,0 +1,148 @@
+// Copyright 2026 RustFS Team
+// Licensed under the Apache License, Version 2.0.
+
+use super::*;
+use crate::storage_api::owner::ObjectOperations as _;
+
+const BUCKET: &str = "quota-reset-preservation";
+const OPERATION: &str = "00000000-0000-0000-0000-000000000002";
+
+async fn reservation_fixture() -> (tempfile::TempDir, Arc<ECStore>, Uuid, String, Vec<u8>) {
+    let (directory, store) = setup_scanner_cycle_store().await;
+    store
+        .make_bucket(BUCKET, &crate::storage_api::scan::MakeBucketOptions::default())
+        .await
+        .expect("create the reservation fixture bucket through its owner");
+    let incarnation = store
+        .bucket_incarnation_id_from_disk(BUCKET)
+        .await
+        .expect("durable bucket incarnation");
+    assert!(!incarnation.is_nil());
+    let path = format!("config/quota-ledger/{BUCKET}.json");
+    let bytes = serde_json::to_vec(&serde_json::json!({
+        "version": 1,
+        "bucket_incarnation": incarnation,
+        "quota_revision_unix_nanos": 1,
+        "accounted_usage": 100,
+        "reservations": {
+            OPERATION: {
+                "object": "pending-object",
+                "old_size": 0,
+                "new_size": 64,
+                "created_at": 1,
+                "pool_index": 0,
+                "set_index": 0,
+                "commit_started": true
+            }
+        }
+    }))
+    .expect("encode the committed reservation fixture");
+    save_config(store.clone(), &path, bytes.clone())
+        .await
+        .expect("persist reservation bytes through the real storage owner");
+    (directory, store, incarnation, path, bytes)
+}
+
+async fn assert_reservation_retained(store: &Arc<ECStore>, path: &str, expected: &[u8], incarnation: Uuid) {
+    let bytes = read_config(store.clone(), path)
+        .await
+        .expect("read the actual reservation ledger");
+    assert_eq!(bytes, expected, "scanner reset must not rewrite the reservation ledger");
+    let ledger: serde_json::Value = serde_json::from_slice(&bytes).expect("persisted ledger JSON");
+    assert_eq!(ledger["version"], 1);
+    assert_eq!(ledger["bucket_incarnation"], incarnation.to_string());
+    assert_eq!(ledger["accounted_usage"], 100);
+    let reservations = ledger["reservations"].as_object().expect("reservation map");
+    assert_eq!(reservations.len(), 1);
+    let pending = &reservations[OPERATION];
+    assert_eq!(pending["old_size"], 0);
+    assert_eq!(pending["new_size"], 64);
+    assert_eq!(pending["commit_started"], true);
+    assert_eq!(
+        store
+            .bucket_incarnation_id_from_disk(BUCKET)
+            .await
+            .expect("owner incarnation after restart"),
+        incarnation
+    );
+}
+
+#[tokio::test]
+#[serial]
+async fn quota_reset_preservation_survives_storage_owner_reconstruction() {
+    let (_directory, store, incarnation, path, bytes) = reservation_fixture().await;
+    let reset = reset_scanner_usage_state_for_full_rebuild(CancellationToken::new(), store.clone())
+        .await
+        .expect("reset scanner usage through the fenced production entry");
+    assert_eq!(reset.usage_state, "bootstrap-pending");
+    let restarted = restart_scanner_cycle_store_from(&store).await;
+    assert!(
+        !Arc::ptr_eq(&store, &restarted),
+        "the assertion must read through a newly constructed ECStore"
+    );
+    assert_reservation_retained(&restarted, &path, &bytes, incarnation).await;
+    let usage = read_config(restarted.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+        .await
+        .expect("read reset usage through the reconstructed owner");
+    let usage: DataUsageInfo = serde_json::from_slice(&usage).expect("bootstrap usage JSON");
+    assert!(data_usage_info_is_bootstrap_pending(&usage));
+    assert!(!data_usage_info_has_persisted_baseline_identity(&usage));
+}
+
+#[tokio::test]
+#[serial]
+async fn quota_reset_preservation_unknown_protocol_rejects_put_after_restart() {
+    for quota_shape in ["zero", "null", "missing"] {
+        let (_directory, store, incarnation, path, bytes) = reservation_fixture().await;
+        let mut quota = serde_json::json!({
+            "quota_type": "Hard",
+            "reservation_protocol": 2,
+            "reservation_quota": 1024
+        });
+        match quota_shape {
+            "zero" => quota["quota"] = serde_json::json!(0),
+            "null" => quota["quota"] = serde_json::Value::Null,
+            "missing" => {}
+            _ => unreachable!("fixed quota shapes"),
+        }
+        let unknown_quota = serde_json::to_vec(&quota).expect("unknown but syntactically valid quota protocol");
+        store
+            .update_bucket_metadata_config(BUCKET, rustfs_config::QUOTA_CONFIG_FILE, unknown_quota)
+            .await
+            .expect("persist a future protocol using the real metadata owner");
+        assert_eq!(
+            store
+                .bucket_incarnation_id_from_disk(BUCKET)
+                .await
+                .expect("same metadata owner incarnation"),
+            incarnation
+        );
+        reset_scanner_usage_state_for_full_rebuild(CancellationToken::new(), store.clone())
+            .await
+            .expect("scanner reset must not change quota metadata");
+        let restarted = restart_scanner_cycle_store_from(&store).await;
+        assert!(!Arc::ptr_eq(&store, &restarted));
+        assert_reservation_retained(&restarted, &path, &bytes, incarnation).await;
+        let mut reader = PutObjReader::from_vec(b"must-not-commit".to_vec());
+        let result = restarted.pools[0].disk_set[0]
+            .put_object(BUCKET, "rejected-object", &mut reader, &ObjectOptions::default())
+            .await;
+        let error = match result {
+            Err(error) => error,
+            Ok(_) => panic!("unknown reservation protocol with quota={quota_shape} must not admit a PUT"),
+        };
+        assert!(
+            matches!(error, EcstoreError::PartMissingOrCorrupt),
+            "unexpected protocol rejection: {error}"
+        );
+        let missing = restarted.pools[0].disk_set[0]
+            .get_object_info(BUCKET, "rejected-object", &ObjectOptions::default())
+            .await
+            .expect_err("the rejected PUT must not create an object");
+        assert!(
+            matches!(missing, EcstoreError::FileNotFound | EcstoreError::ObjectNotFound(_, _)),
+            "object absence must not be confused with another storage failure: {missing}"
+        );
+        assert_reservation_retained(&restarted, &path, &bytes, incarnation).await;
+    }
+}
@@ -0,0 +1,607 @@
+// Copyright 2026 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use super::super::cycle_state::cleanup_io_fault;
+use super::*;
+use crate::storage_api::owner::{EcstoreRebalStatus, EcstoreRebalanceInfo, EcstoreRebalanceMeta, EcstoreRebalanceStats};
+
+async fn seed_cleanup(store: &Arc<ECStore>, state: &str) -> ScannerCycleRecoveryMarker {
+    let cycle = CurrentCycle {
+        current: 3,
+        next: 42,
+        ..Default::default()
+    };
+    save_config(
+        store.clone(),
+        DATA_USAGE_BLOOM_NAME_PATH.as_str(),
+        encode_scanner_cycle_state(&cycle, 7).expect("cycle encoding"),
+    )
+    .await
+    .expect("persist cycle");
+    let usage = DataUsageInfo {
+        scanner_epoch: Some(7),
+        scanner_cycle: Some(41),
+        ..complete_usage_with_bucket_count(Some(std::time::SystemTime::UNIX_EPOCH), 0)
+    };
+    save_config(
+        store.clone(),
+        DATA_USAGE_OBJ_NAME_PATH.as_str(),
+        serde_json::to_vec(&usage).expect("usage encoding"),
+    )
+    .await
+    .expect("persist usage floor");
+    let marker = ScannerCycleRecoveryMarker {
+        schema_version: 1,
+        primary_revision: "previous-primary".to_string(),
+        generation: 41,
+        leader_epoch: 7,
+        classification: "corrupt".to_string(),
+        first_detected_at_unix_secs: 1,
+        last_attempt_at_unix_secs: 2,
+        retry_count: 1,
+        reason: "operator reset in progress".to_string(),
+        path: DATA_USAGE_BLOOM_NAME_PATH.clone(),
+        quarantine_path: DATA_USAGE_BLOOM_RECOVERY_PATH.clone(),
+        state: state.to_string(),
+    };
+    save_config(
+        store.clone(),
+        DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(),
+        serde_json::to_vec(&marker).expect("marker encoding"),
+    )
+    .await
+    .expect("persist operator marker");
+    marker
+}
+
+async fn persisted_state(store: &Arc<ECStore>) -> Vec<(Option<Vec<u8>>, DataUsageCacheRevision)> {
+    let mut state = Vec::new();
+    for path in [
+        DATA_USAGE_BLOOM_NAME_PATH.as_str(),
+        DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(),
+        DATA_USAGE_OBJ_NAME_PATH.as_str(),
+    ] {
+        state.push(
+            read_config_with_revision(store.clone(), path)
+                .await
+                .expect("read exact metadata revision"),
+        );
+    }
+    state
+}
+
+async fn run_disabled_startup(ctx: CancellationToken, store: Arc<ECStore>) {
+    let initialized_before = crate::scanner_runtime_initialized();
+    let cleanup = init_scanner_with_recovery(ctx, store, false).await;
+    if let Some(cleanup) = cleanup {
+        tokio::time::timeout(Duration::from_secs(15), cleanup)
+            .await
+            .expect("finite disabled cleanup attempt")
+            .expect("cleanup task should not panic");
+    }
+    assert_eq!(
+        crate::scanner_runtime_initialized(),
+        initialized_before,
+        "disabled recovery must not start the normal scanner runtime"
+    );
+}
+
+async fn assert_reset_fences(store: &Arc<ECStore>) {
+    let data = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("cycle remains durable");
+    let (cycle, epoch) = decode_scanner_cycle_state(&data).expect("valid preserved cycle");
+    assert_eq!((cycle.current, cycle.next, epoch), (3, 42, 8));
+    let usage: DataUsageInfo = serde_json::from_slice(
+        &read_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+            .await
+            .expect("durable usage fence"),
+    )
+    .expect("valid usage");
+    assert_eq!(usage.scanner_epoch, Some(8));
+    assert_eq!(usage.scanner_cycle, Some(41));
+    assert!(matches!(
+        read_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+async fn assert_rebuilt_reset_fences(store: &Arc<ECStore>, expected_epoch: u64) {
+    let data = read_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str())
+        .await
+        .expect("rebuilt cycle remains durable");
+    let (cycle, epoch) = decode_scanner_cycle_state(&data).expect("valid rebuilt cycle");
+    assert_eq!((cycle.current, cycle.next, epoch), (0, 42, expected_epoch));
+    let usage: DataUsageInfo = serde_json::from_slice(
+        &read_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+            .await
+            .expect("durable rebuilt usage fence"),
+    )
+    .expect("valid usage");
+    assert_eq!(usage.scanner_epoch, Some(expected_epoch));
+    assert_eq!(usage.scanner_cycle, Some(41));
+    assert!(matches!(
+        read_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()).await,
+        Err(EcstoreError::ConfigNotFound)
+    ));
+}
+
+#[test]
+fn scanner_reset_crash_child_process_fixture() {
+    let Ok(root) = std::env::var("RUSTFS_SCANNER_RESET_CRASH_ROOT") else {
+        return;
+    };
+    let stage = match std::env::var("RUSTFS_SCANNER_RESET_CRASH_STAGE").as_deref() {
+        Ok("primary-read") => cleanup_io_fault::Stage::PrimaryRead,
+        Ok("primary-write") => cleanup_io_fault::Stage::PrimaryWrite,
+        Ok("usage-fence") => cleanup_io_fault::Stage::UsageFence,
+        other => panic!("unexpected reset crash stage: {other:?}"),
+    };
+    let runtime = tokio::runtime::Builder::new_current_thread()
+        .enable_all()
+        .build()
+        .expect("child runtime should build");
+    runtime.block_on(async {
+        let store = setup_scanner_cycle_store_at_path(std::path::Path::new(&root), false, 1).await;
+        if stage == cleanup_io_fault::Stage::UsageFence {
+            save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), b"corrupt-cycle".to_vec())
+                .await
+                .expect("force child through reconstruction branch");
+        }
+        let injection = cleanup_io_fault::install(&store, stage, false);
+        let error = resume_scanner_cycle_cleanup(CancellationToken::new(), store)
+            .await
+            .expect_err("child should stop at the injected owned I/O boundary");
+        assert!(injection.fired_while_owned());
+        let expected = match stage {
+            cleanup_io_fault::Stage::PrimaryRead => "injected primary read failure",
+            cleanup_io_fault::Stage::PrimaryWrite => "injected primary write failure",
+            cleanup_io_fault::Stage::UsageFence => "injected usage fence failure",
+        };
+        assert!(error.to_string().contains(expected), "{error}");
+    });
+    std::process::exit(77);
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_recovers_after_child_process_crash_boundaries() {
+    for (name, expected_rebuilt_epoch) in [("primary-read", None), ("primary-write", None), ("usage-fence", Some(9))] {
+        let temp_dir = tempfile::tempdir().expect("crash fixture directory");
+        let store = setup_scanner_cycle_store_at_path(temp_dir.path(), false, 1).await;
+        seed_cleanup(&store, "cleanup-pending").await;
+        drop(store);
+
+        let status = std::process::Command::new(std::env::current_exe().expect("test binary path"))
+            .arg("scanner::tests::recovery_control::scanner_reset_crash_child_process_fixture")
+            .arg("--exact")
+            .arg("--nocapture")
+            .env("RUSTFS_SCANNER_RESET_CRASH_ROOT", temp_dir.path())
+            .env("RUSTFS_SCANNER_RESET_CRASH_STAGE", name)
+            .status()
+            .expect("child crash fixture should start");
+        assert_eq!(status.code(), Some(77), "{name} child did not reach the owned crash boundary");
+
+        let restarted = setup_scanner_cycle_store_at_path(temp_dir.path(), false, 1).await;
+        run_disabled_startup(CancellationToken::new(), restarted.clone()).await;
+        if let Some(epoch) = expected_rebuilt_epoch {
+            assert_rebuilt_reset_fences(&restarted, epoch).await;
+        } else {
+            assert_reset_fences(&restarted).await;
+        }
+    }
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_reopens_persisted_intent_without_starting_scanner() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    seed_cleanup(&store, "cleanup-pending").await;
+    let restarted = restart_scanner_cycle_store_from(&store).await;
+    run_disabled_startup(CancellationToken::new(), restarted.clone()).await;
+    assert_reset_fences(&restarted).await;
+    let completed = persisted_state(&restarted).await;
+    run_disabled_startup(CancellationToken::new(), restarted.clone()).await;
+    assert_eq!(
+        persisted_state(&restarted).await,
+        completed,
+        "a later startup without an intent must not reset again"
+    );
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_does_not_authorize_blocked_unknown_or_corrupt_markers() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    for kind in ["blocked", "unknown-phase", "future-version", "unknown-field", "corrupt"] {
+        let marker = seed_cleanup(&store, "blocked").await;
+        let mut value = serde_json::to_value(marker).expect("marker value");
+        match kind {
+            "unknown-phase" => value["state"] = "future-phase".into(),
+            "future-version" => value["schema_version"] = 99.into(),
+            "unknown-field" => value["future_hint"] = true.into(),
+            _ => {}
+        }
+        let bytes = if kind == "corrupt" {
+            b"{broken".to_vec()
+        } else {
+            serde_json::to_vec(&value).expect("marker JSON")
+        };
+        save_config(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(), bytes)
+            .await
+            .expect("persist rejected marker");
+        let before = persisted_state(&store).await;
+        run_disabled_startup(CancellationToken::new(), store.clone()).await;
+        assert_eq!(persisted_state(&store).await, before, "{kind} must not become an automatic full rescan");
+    }
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_rechecks_revision_after_waiting_for_leader_lock() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    let mut marker = seed_cleanup(&store, "cleanup-pending").await;
+    let expected = read_config_with_revision(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str())
+        .await
+        .expect("intent revision")
+        .1;
+    let lock = store
+        .new_ns_lock(RUSTFS_META_BUCKET, "leader.lock")
+        .await
+        .expect("leader lock");
+    let guard = lock
+        .get_write_lock_quiet(Duration::from_secs(1))
+        .await
+        .expect("hold leader ownership");
+    let mut recovery = Box::pin(reset_scanner_cycle_recovery_for_intent(
+        CancellationToken::new(),
+        store.clone(),
+        Some(expected),
+        None,
+    ));
+    assert!(matches!(futures::poll!(&mut recovery), Poll::Pending));
+    marker.state = "blocked".to_string();
+    save_config(
+        store.clone(),
+        DATA_USAGE_BLOOM_RECOVERY_PATH.as_str(),
+        serde_json::to_vec(&marker).expect("replacement marker"),
+    )
+    .await
+    .expect("replace intent while the fixture owns leader lock");
+    let replaced = persisted_state(&store).await;
+    drop(guard);
+    let error = recovery.await.expect_err("old preflight cannot authorize replacement marker");
+    assert!(error.to_string().contains("changed before recovery acquired ownership"));
+    assert_eq!(persisted_state(&store).await, replaced);
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_requires_phase_even_when_revision_matches() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    seed_cleanup(&store, "blocked").await;
+    let before = persisted_state(&store).await;
+    let error = reset_scanner_cycle_recovery_for_intent(CancellationToken::new(), store.clone(), Some(before[1].1.clone()), None)
+        .await
+        .expect_err("a matching ETag alone is not operator cleanup authorization");
+    assert!(error.to_string().contains("unchanged cleanup-pending"));
+    assert_eq!(persisted_state(&store).await, before);
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("explicit v3 core retains full reset authorization");
+    assert_reset_fences(&store).await;
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_lock_busy_preserves_intent_without_force_unlock() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    seed_cleanup(&store, "cleanup-pending").await;
+    let before = persisted_state(&store).await;
+    let lock = store
+        .new_ns_lock(RUSTFS_META_BUCKET, "leader.lock")
+        .await
+        .expect("leader lock");
+    let guard = lock
+        .get_write_lock_quiet(Duration::from_secs(1))
+        .await
+        .expect("hold live leader");
+    let error = resume_scanner_cycle_cleanup(CancellationToken::new(), store.clone())
+        .await
+        .expect_err("busy leader must block recovery");
+    assert!(error.to_string().contains("leader lock is busy"));
+    let status = scanner_cycle_recovery_status();
+    assert_eq!(status.state, "cleanup-pending");
+    assert!(
+        status
+            .reason
+            .as_deref()
+            .is_some_and(|reason| reason.contains("leader lock is busy"))
+    );
+    assert!(!status.retryable, "disabled startup makes one attempt, not an automatic retry loop");
+    assert!(!guard.is_lock_lost(), "recovery must not revoke the live owner");
+    assert_eq!(persisted_state(&store).await, before);
+    drop(guard);
+    run_disabled_startup(CancellationToken::new(), store.clone()).await;
+    assert_reset_fences(&store).await;
+}
+
+#[tokio::test]
+#[serial]
+async fn concurrent_full_rescan_requests_converge_to_one_recovery_fence() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    seed_cleanup(&store, "blocked").await;
+    let before = persisted_state(&store).await;
+    let lock = store
+        .new_ns_lock(RUSTFS_META_BUCKET, "leader.lock")
+        .await
+        .expect("leader lock");
+    let guard = lock
+        .get_write_lock_quiet(Duration::from_secs(1))
+        .await
+        .expect("hold live leader before concurrent admin resets");
+
+    let first = tokio::spawn(reset_scanner_cycle_recovery(CancellationToken::new(), store.clone()));
+    let second = tokio::spawn(reset_scanner_cycle_recovery(CancellationToken::new(), store.clone()));
+    tokio::task::yield_now().await;
+    assert_eq!(
+        persisted_state(&store).await,
+        before,
+        "waiting admin reset requests must not mutate state before leader ownership"
+    );
+
+    drop(guard);
+    let (first, second) = tokio::time::timeout(Duration::from_secs(15), async { tokio::join!(first, second) })
+        .await
+        .expect("both admin reset requests should finish after the live owner releases the lock");
+    let first = first.expect("first admin reset task should not panic");
+    let second = second.expect("second admin reset task should not panic");
+    assert!(first.is_ok(), "first admin reset failed: {first:?}");
+    assert!(second.is_ok(), "second admin reset failed: {second:?}");
+
+    assert_reset_fences(&store).await;
+    let completed = persisted_state(&store).await;
+    reset_scanner_cycle_recovery(CancellationToken::new(), store.clone())
+        .await
+        .expect("lost-reply retry after a completed reset should be idempotent");
+    assert_eq!(
+        persisted_state(&store).await,
+        completed,
+        "a later retry must not create a second reset identity or rewrite durable fences"
+    );
+    assert_eq!(scanner_cycle_recovery_status().state, "healthy");
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_movement_pause_preserves_intent_for_later_startup() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    seed_cleanup(&store, "cleanup-pending").await;
+    let before = persisted_state(&store).await;
+    *store.rebalance_meta.write().await = Some(EcstoreRebalanceMeta {
+        id: "cleanup-movement".to_string(),
+        pool_stats: vec![EcstoreRebalanceStats {
+            participating: true,
+            info: EcstoreRebalanceInfo {
+                start_time: Some(time::OffsetDateTime::now_utc()),
+                status: EcstoreRebalStatus::Started,
+                ..Default::default()
+            },
+            ..Default::default()
+        }],
+        ..Default::default()
+    });
+    let error = resume_scanner_cycle_cleanup(CancellationToken::new(), store.clone())
+        .await
+        .expect_err("movement must block reset publication");
+    assert!(error.to_string().contains("blocked by data movement"));
+    let status = scanner_cycle_recovery_status();
+    assert_eq!(status.state, "cleanup-pending");
+    assert!(
+        status
+            .reason
+            .as_deref()
+            .is_some_and(|reason| reason.contains("blocked by data movement"))
+    );
+    assert_eq!(persisted_state(&store).await, before);
+    *store.rebalance_meta.write().await = None;
+    run_disabled_startup(CancellationToken::new(), store.clone()).await;
+    assert_reset_fences(&store).await;
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_cancelled_startup_preserves_persisted_work() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    seed_cleanup(&store, "cleanup-pending").await;
+    let before = persisted_state(&store).await;
+    let ctx = CancellationToken::new();
+    ctx.cancel();
+    run_disabled_startup(ctx, store.clone()).await;
+    assert_eq!(persisted_state(&store).await, before);
+}
+
+#[tokio::test(start_paused = true)]
+#[serial]
+async fn disabled_cleanup_probe_obeys_cancellation_and_existing_io_deadline() {
+    for cancel in [false, true] {
+        let store = Arc::new(MemoryConfigStore::default());
+        store.delayed_gets.lock().await.insert(
+            memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_RECOVERY_PATH.as_str()),
+            data_usage_persist_timeout().saturating_add(Duration::from_secs(60)),
+        );
+        let ctx = CancellationToken::new();
+        let mut probe = Box::pin(read_scanner_cleanup_marker(store.clone(), &ctx));
+        assert!(matches!(futures::poll!(&mut probe), Poll::Pending));
+        if cancel {
+            ctx.cancel();
+        } else {
+            advance(data_usage_persist_timeout()).await;
+        }
+        let error = probe.await.expect_err("pending read must be bounded");
+        assert!(error.to_string().contains(if cancel { "cancelled" } else { "timed out" }));
+        assert!(store.put_counts.lock().await.is_empty(), "probe must remain read-only");
+    }
+}
+
+#[test]
+#[serial]
+fn disabled_cleanup_old_observation_cannot_overwrite_a_new_completion() {
+    let original = scanner_cycle_recovery_status();
+    let healthy = ScannerCycleRecoveryStatus {
+        state: "healthy".to_string(),
+        ..Default::default()
+    };
+    publish_scanner_cleanup_status(healthy.clone(), None).expect("first status version");
+    let old = scanner_cleanup_status_version();
+    publish_scanner_cleanup_status(healthy, None).expect("a newer completion may have identical fields");
+    assert!(
+        publish_scanner_cleanup_status(
+            ScannerCycleRecoveryStatus {
+                state: "cleanup-pending".to_string(),
+                reason: Some("old lock wait failed".to_string()),
+                ..Default::default()
+            },
+            Some(old)
+        )
+        .is_none()
+    );
+    assert_eq!(scanner_cycle_recovery_status().state, "healthy");
+    publish_scanner_cleanup_status(original, None).expect("restore prior observation");
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_owned_read_and_write_failures_keep_specific_status() {
+    for (stage, newer_completion) in [
+        (cleanup_io_fault::Stage::PrimaryRead, false),
+        (cleanup_io_fault::Stage::PrimaryWrite, false),
+        (cleanup_io_fault::Stage::PrimaryWrite, true),
+    ] {
+        let (_dir, store) = setup_scanner_cycle_store().await;
+        seed_cleanup(&store, "cleanup-pending").await;
+        let before = persisted_state(&store).await;
+        let injection = cleanup_io_fault::install(&store, stage, newer_completion);
+        let error = resume_scanner_cycle_cleanup(CancellationToken::new(), store.clone())
+            .await
+            .expect_err("injected owned I/O boundary");
+        assert!(
+            injection.fired_while_owned(),
+            "fault must occur after real leader ownership and marker validation"
+        );
+        let expected_error = match stage {
+            cleanup_io_fault::Stage::PrimaryRead => "injected primary read failure",
+            cleanup_io_fault::Stage::PrimaryWrite => "injected primary write failure",
+            cleanup_io_fault::Stage::UsageFence => "injected usage fence failure",
+        };
+        assert!(error.to_string().contains(expected_error));
+        let status = scanner_cycle_recovery_status();
+        if newer_completion {
+            assert_eq!(status.state, "healthy", "old error cannot overwrite a newer completion observation");
+            assert!(status.reason.is_none());
+        } else {
+            assert_eq!(status.state, "cleanup-pending");
+            assert!(
+                status.reason.as_deref().is_some_and(|reason| reason.contains(expected_error)),
+                "{status:?}"
+            );
+        }
+        let after = persisted_state(&store).await;
+        assert_eq!(after[0], before[0], "failed primary I/O must preserve its prior revision");
+        assert_eq!(after[2], before[2], "failed primary I/O must not advance the usage fence");
+        let marker: ScannerCycleRecoveryMarker =
+            serde_json::from_slice(after[1].0.as_deref().expect("durable marker retained")).expect("valid cleanup marker");
+        assert_eq!(marker.state, "cleanup-pending");
+        drop(injection);
+        run_disabled_startup(CancellationToken::new(), store.clone()).await;
+        assert_reset_fences(&store).await;
+    }
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_invalidated_observation_never_becomes_unconditional() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    seed_cleanup(&store, "cleanup-pending").await;
+    let revision = read_config_with_revision(store.clone(), DATA_USAGE_BLOOM_RECOVERY_PATH.as_str())
+        .await
+        .expect("marker revision")
+        .1;
+    let newer = ScannerCycleRecoveryStatus {
+        state: "healthy".to_string(),
+        reason: Some("newer completion owner".to_string()),
+        ..Default::default()
+    };
+    publish_scanner_cleanup_status(newer.clone(), None).expect("newer observation");
+    let mut invalidated = None;
+    reset_scanner_cycle_recovery_for_intent(CancellationToken::new(), store.clone(), Some(revision), Some(&mut invalidated))
+        .await
+        .expect("metadata cleanup may complete without owning the newest status observation");
+    assert!(invalidated.is_none());
+    assert_eq!(
+        serde_json::to_value(scanner_cycle_recovery_status()).expect("observed status"),
+        serde_json::to_value(newer).expect("newer status")
+    );
+    assert_reset_fences(&store).await;
+}
+
+#[tokio::test]
+#[serial]
+async fn disabled_cleanup_later_failure_preserves_rebuilt_primary_status_identity() {
+    for newer_completion in [false, true] {
+        let (_dir, store) = setup_scanner_cycle_store().await;
+        seed_cleanup(&store, "cleanup-pending").await;
+        save_config(store.clone(), DATA_USAGE_BLOOM_NAME_PATH.as_str(), b"corrupt-cycle".to_vec())
+            .await
+            .expect("force the full reconstruction branch");
+        let before = persisted_state(&store).await;
+        let injection = cleanup_io_fault::install(&store, cleanup_io_fault::Stage::UsageFence, newer_completion);
+        let error = resume_scanner_cycle_cleanup(CancellationToken::new(), store.clone())
+            .await
+            .expect_err("fail after primary publication");
+        assert!(injection.fired_while_owned());
+        assert!(error.to_string().contains("injected usage fence failure"));
+        let after = persisted_state(&store).await;
+        assert_ne!(after[0].1, before[0].1, "the primary write must actually commit before this failure");
+        let (cycle, epoch) =
+            decode_scanner_cycle_state(after[0].0.as_deref().expect("rebuilt primary")).expect("valid durable reconstruction");
+        assert_eq!((cycle.current, cycle.next, epoch), (0, 42, 8));
+        assert_eq!(after[2], before[2], "usage fence publication was rejected");
+        let marker: ScannerCycleRecoveryMarker =
+            serde_json::from_slice(after[1].0.as_deref().expect("cleanup marker retained")).expect("valid cleanup marker");
+        assert_eq!(marker.state, "cleanup-pending");
+        let status = scanner_cycle_recovery_status();
+        if newer_completion {
+            assert_eq!(status.state, "healthy");
+            assert!(
+                status.reason.is_none(),
+                "old core and outer failure must both retain invalidated ownership"
+            );
+        } else {
+            let DataUsageCacheRevision::Etag(etag) = &after[0].1 else {
+                panic!("rebuilt primary must have a revision");
+            };
+            assert_eq!(status.state, "cleanup-pending");
+            assert_eq!(status.primary_revision.as_deref(), Some(etag.as_str()));
+            assert_eq!(status.generation, Some(cycle.next));
+            assert_eq!(status.leader_epoch, Some(epoch));
+            assert!(
+                status
+                    .reason
+                    .as_deref()
+                    .is_some_and(|reason| reason.contains("injected usage fence failure"))
+            );
+        }
+    }
+}
@@ -0,0 +1,490 @@
+// Copyright 2026 RustFS Team
+// Licensed under the Apache License, Version 2.0.
+
+use super::super::usage_store::DataUsagePublicationResult;
+use super::*;
+use crate::scanner_io::ScannerBucketScanScope;
+use rustfs_utils::path::path_join_buf;
+use sha2::Digest;
+use std::time::SystemTime;
+
+const PROOF_BUCKET: &str = "publication-proof-bucket";
+const PROOF_EPOCH: u64 = 7;
+const PROOF_CYCLE: u64 = 11;
+
+async fn settle_namespace_commits(store: &ECStore) {
+    tokio::time::timeout(Duration::from_secs(30), async {
+        while store.scanner_data_usage_publication_blocked().await {
+            tokio::time::sleep(Duration::from_millis(1)).await;
+        }
+    })
+    .await
+    .expect("fixture namespace commits must settle before collecting complete coverage");
+}
+
+async fn complete_candidate(store: &Arc<ECStore>, cycle: u64) -> (crate::scanner_io::ScannerCycleResult, DataUsageInfo) {
+    settle_namespace_commits(store).await;
+    let ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new_with_progress_tracking(
+        &ctx,
+        ScannerCycleBudgetConfig {
+            max_objects: Some(8),
+            ..Default::default()
+        },
+    );
+    let (updates, mut receiver) = mpsc::channel(1);
+    let result = crate::scanner_io::nsscanner_with_storage_status_scoped(
+        store.as_ref(),
+        crate::scanner_io::ScannerCycleRequest {
+            ctx,
+            budget,
+            updates,
+            want_cycle: cycle,
+            leader_epoch: PROOF_EPOCH,
+            scan_mode: HealScanMode::Normal,
+            scan_scope: ScannerBucketScanScope::default(),
+            persisted_usage_baseline: None,
+            observed_usage_candidate: None,
+            requires_full_scan: true,
+            service_cohort: None,
+            resolved_scope_observer: None,
+        },
+    )
+    .await
+    .expect("real scanner must produce the fixture candidate");
+    assert_eq!(result.status, ScannerCycleStatus::Complete);
+    let candidate = receiver.recv().await.expect("complete scanner snapshot");
+    assert!(candidate.usage_snapshot_complete);
+    assert_eq!(candidate.usage_snapshot_converged, Some(true));
+    assert_eq!(candidate.scanner_cycle, Some(cycle));
+    assert_eq!(candidate.scanner_epoch, Some(PROOF_EPOCH));
+    (result, candidate)
+}
+
+async fn candidate_store() -> (tempfile::TempDir, Arc<ECStore>) {
+    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
+    let (directory, store) = setup_scanner_cycle_store_with_usage_baseline(false).await;
+    store
+        .make_bucket(PROOF_BUCKET, &crate::storage_api::scan::MakeBucketOptions::default())
+        .await
+        .expect("create proof fixture bucket through the owner");
+    let mut reader = PutObjReader::from_vec(b"proof".to_vec());
+    store.pools[0].disk_set[0]
+        .put_object(PROOF_BUCKET, "initial", &mut reader, &ObjectOptions::default())
+        .await
+        .expect("persist fixture object through the owner");
+    crate::scanner_io::record_dirty_usage_bucket(PROOF_BUCKET);
+    settle_namespace_commits(&store).await;
+    (directory, store)
+}
+
+async fn read_root(store: &Arc<ECStore>) -> (Option<Vec<u8>>, DataUsageCacheRevision) {
+    read_config_with_revision(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+        .await
+        .expect("read actual v2 root bytes and revision")
+}
+
+async fn publish_candidate(
+    store: &Arc<ECStore>,
+    scan: &crate::scanner_io::ScannerCycleResult,
+    candidate: DataUsageInfo,
+    baseline: Option<DataUsagePersistBaseline>,
+) -> DataUsagePublicationResult {
+    let expectation = scan.publication_expectation();
+    assert!(expectation.is_some(), "only a real complete scan may supply the expectation");
+    let (sender, receiver) = mpsc::channel(1);
+    sender.send(candidate).await.expect("enqueue the real scan candidate");
+    drop(sender);
+    store_data_usage_in_backend_with_outcome_for_epoch_and_baseline_and_route_probe_for_publication_epoch_and_lease_fence(
+        CancellationToken::new(),
+        store.clone(),
+        receiver,
+        Some(PROOF_EPOCH),
+        baseline,
+        ScannerPublicationFence::new(scan.publication_epoch(), None, None).with_ack_expectation(expectation),
+        || async { None },
+    )
+    .await
+}
+
+#[tokio::test]
+#[serial]
+async fn scoped_ack_publication_companion_only_does_not_authorize_root_ack() {
+    for companion in [
+        format!("{}.bkp", DATA_USAGE_OBJ_NAME_PATH.as_str()),
+        LEGACY_DATA_USAGE_OBJ_NAME_PATH.to_string(),
+        format!("{}.bkp", LEGACY_DATA_USAGE_OBJ_NAME_PATH.as_str()),
+    ] {
+        let (_directory, store) = candidate_store().await;
+        let (scan, candidate) = complete_candidate(&store, PROOF_CYCLE).await;
+        let bytes = serde_json::to_vec(&candidate).expect("actual candidate JSON");
+        save_config(store.clone(), &companion, bytes.clone())
+            .await
+            .expect("persist the companion on real disks");
+        let baseline = read_data_usage_persist_baseline(store.clone())
+            .await
+            .expect("companion fallback baseline");
+        assert_eq!(baseline.data.as_deref(), Some(bytes.as_slice()));
+        assert_eq!(baseline.revision, DataUsageCacheRevision::Missing);
+        assert_eq!(read_root(&store).await.0, None);
+        let dirty = crate::scanner_io::dirty_usage_buckets_for_tests();
+
+        let publication = publish_candidate(&store, &scan, candidate, Some(baseline)).await;
+        assert_eq!(publication.outcome(), DataUsagePersistOutcome::AlreadyDurable);
+        let (_, pending, acknowledgements) = finalize_scanner_cycle_result(scan, publication);
+        assert!(pending, "unacknowledged durable companion work must remain pending");
+        assert!(acknowledgements.is_empty());
+        assert_eq!(
+            crate::scanner_io::dirty_usage_buckets_for_tests(),
+            dirty,
+            "a companion is not the v2 root target"
+        );
+        assert_eq!(read_root(&store).await, (None, DataUsageCacheRevision::Missing));
+        assert_eq!(read_config(store.clone(), &companion).await.expect("companion retained"), bytes);
+    }
+    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
+}
+
+#[tokio::test]
+#[serial]
+async fn scoped_ack_publication_actual_root_readback_accepts_semantic_json_equivalence() {
+    let (_directory, store) = candidate_store().await;
+    let (scan, candidate) = complete_candidate(&store, PROOF_CYCLE).await;
+    let canonical = serde_json::to_vec(&candidate).expect("candidate encoding");
+    let mut value = serde_json::to_value(&candidate).expect("candidate value");
+    value
+        .as_object_mut()
+        .expect("usage object")
+        .insert("fixture_unknown_field".into(), serde_json::json!({"retained": true}));
+    let different_bytes = serde_json::to_vec_pretty(&value).expect("noncanonical primary JSON");
+    assert_ne!(different_bytes, canonical);
+    assert_eq!(
+        serde_json::from_slice::<DataUsageInfo>(&different_bytes).expect("semantic primary"),
+        candidate
+    );
+    save_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str(), different_bytes.clone())
+        .await
+        .expect("persist actual primary representation");
+    let before = read_root(&store).await;
+    assert!(matches!(&before.1, DataUsageCacheRevision::Etag(etag) if !etag.is_empty()));
+    let baseline = read_data_usage_persist_baseline(store.clone())
+        .await
+        .expect("real primary revision");
+    assert!(crate::scanner_io::dirty_usage_buckets_pending());
+
+    let publication = publish_candidate(&store, &scan, candidate.clone(), Some(baseline)).await;
+    assert_eq!(publication.outcome(), DataUsagePersistOutcome::AlreadyDurable);
+    let (_, proof) = publication.into_parts();
+    let proof = proof.expect("actual primary readback must produce its own root proof");
+    let expected = scan.publication_expectation().expect("real scan expectation");
+    let (etag, raw_digest) = proof.verified_version_for(&expected).expect("proof must bind this candidate");
+    let DataUsageCacheRevision::Etag(expected_etag) = &before.1 else { panic!("actual root ETag") };
+    assert_eq!(etag, expected_etag);
+    let expected_digest: [u8; 32] = sha2::Sha256::digest(&different_bytes).into();
+    assert_eq!(
+        *raw_digest, expected_digest,
+        "proof must record actual bytes, not reserialized candidate bytes"
+    );
+    // Obtain another proof through the same real readback path rather than
+    // fabricating a publication result from the inspected proof above.
+    let publication = publish_candidate(&store, &scan, candidate, None).await;
+    let (outcome, _, acknowledgements) = finalize_scanner_cycle_result(scan, publication);
+    assert_eq!(outcome, ScannerCycleOutcome::Completed);
+    assert!(acknowledgements.is_empty(), "the single-node fixture has no remote targets");
+    assert!(
+        !crate::scanner_io::dirty_usage_buckets_pending(),
+        "actual root bytes plus a real revision authorize this scan"
+    );
+    assert_eq!(read_root(&store).await, before, "readback must not rewrite unknown fields or whitespace");
+}
+
+#[tokio::test]
+#[serial]
+async fn scoped_ack_publication_successful_root_cas_authorizes_its_scan() {
+    let (_directory, store) = candidate_store().await;
+    let (scan, candidate) = complete_candidate(&store, PROOF_CYCLE).await;
+    let baseline = read_data_usage_persist_baseline(store.clone())
+        .await
+        .expect("initial root revision");
+    assert_eq!(baseline.revision, DataUsageCacheRevision::Missing);
+    assert!(crate::scanner_io::dirty_usage_buckets_pending());
+    let publication = publish_candidate(&store, &scan, candidate.clone(), Some(baseline)).await;
+    assert_eq!(publication.outcome(), DataUsagePersistOutcome::Saved);
+    let (bytes, revision) = read_root(&store).await;
+    assert!(matches!(revision, DataUsageCacheRevision::Etag(etag) if !etag.is_empty()));
+    assert_eq!(
+        serde_json::from_slice::<DataUsageInfo>(&bytes.expect("actual saved root")).expect("root JSON"),
+        candidate
+    );
+    let (outcome, pending, acknowledgements) = finalize_scanner_cycle_result(scan, publication);
+    assert_eq!(outcome, ScannerCycleOutcome::Completed);
+    assert!(!pending);
+    assert!(acknowledgements.is_empty(), "the single-node fixture has no remote targets");
+    assert!(
+        !crate::scanner_io::dirty_usage_buckets_pending(),
+        "the real root CAS must authorize its matching scan"
+    );
+}
+
+#[tokio::test]
+#[serial]
+async fn scoped_ack_publication_observed_candidate_reuse_requires_a_new_root_proof() {
+    let (_directory, store) = candidate_store().await;
+    let bootstrap = scanner_usage_bootstrap_marker(SystemTime::now(), Some(PROOF_EPOCH));
+    save_config(
+        store.clone(),
+        DATA_USAGE_OBJ_NAME_PATH.as_str(),
+        serde_json::to_vec(&bootstrap).expect("bootstrap root encoding"),
+    )
+    .await
+    .expect("persist authoritative bootstrap root");
+    let (prior_scan, mut observed_candidate) = complete_candidate(&store, PROOF_CYCLE).await;
+    // Seed a complete but unconverged observation from real scanner coverage;
+    // the production writer attaches its authoritative baseline identity.
+    observed_candidate.usage_snapshot_converged = Some(false);
+    let observation = publish_candidate(&store, &prior_scan, observed_candidate, None).await;
+    let (outcome, proof) = observation.into_parts();
+    assert_eq!(outcome, DataUsagePersistOutcome::Saved);
+    assert!(proof.is_none(), "an observational write cannot authorize a root ACK");
+    let (root_before, revision_before) = read_root(&store).await;
+    let observed = read_config(store.clone(), DATA_USAGE_OBSERVED_OBJ_NAME_PATH.as_str())
+        .await
+        .expect("read real persisted observation");
+    let ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(&ctx, ScannerCycleBudgetConfig::default());
+    let (updates, mut receiver) = mpsc::channel(1);
+    let (observer, selected) = tokio::sync::oneshot::channel();
+    let scan = crate::scanner_io::nsscanner_with_storage_status_scoped(
+        store.as_ref(),
+        crate::scanner_io::ScannerCycleRequest {
+            ctx,
+            budget,
+            updates,
+            want_cycle: PROOF_CYCLE + 1,
+            leader_epoch: PROOF_EPOCH,
+            scan_mode: HealScanMode::Normal,
+            scan_scope: ScannerBucketScanScope::default(),
+            persisted_usage_baseline: root_before.clone().map(Bytes::from),
+            observed_usage_candidate: Some(Bytes::from(observed)),
+            requires_full_scan: false,
+            service_cohort: None,
+            resolved_scope_observer: Some(observer),
+        },
+    )
+    .await
+    .expect("observation-backed scope must run through the real scanner");
+    let scope = selected.await.expect("production resolver decision");
+    assert_eq!(scope.selected_buckets_for_tests(), Some(&HashSet::from([PROOF_BUCKET.to_string()])));
+    assert_eq!(scan.status, ScannerCycleStatus::Complete);
+    let expectation = scan.publication_expectation().expect("reused coverage must be revalidated");
+    assert!(
+        !expectation.same_candidate(&prior_scan.publication_expectation().expect("prior real candidate")),
+        "the observation cannot transfer the previous scan's expectation"
+    );
+    assert_eq!(read_root(&store).await, (root_before, revision_before));
+    assert!(crate::scanner_io::dirty_usage_buckets_pending());
+    let candidate = receiver.recv().await.expect("new validated root candidate");
+    assert_eq!(candidate.scanner_cycle, Some(PROOF_CYCLE + 1));
+    assert_eq!(candidate.usage_snapshot_converged, Some(true));
+    let publication = publish_candidate(&store, &scan, candidate, None).await;
+    assert_eq!(publication.outcome(), DataUsagePersistOutcome::Saved);
+    let (outcome, pending, acknowledgements) = finalize_scanner_cycle_result(scan, publication);
+    assert_eq!(outcome, ScannerCycleOutcome::Completed);
+    assert!(!pending);
+    assert!(acknowledgements.is_empty());
+    assert!(!crate::scanner_io::dirty_usage_buckets_pending());
+}
+
+#[tokio::test]
+#[serial]
+async fn scoped_ack_publication_stale_root_cas_keeps_dirty_after_bucket_save() {
+    let (_directory, store) = candidate_store().await;
+    let (scan, candidate) = complete_candidate(&store, PROOF_CYCLE).await;
+    let mut bucket_cache = DataUsageCache::default();
+    bucket_cache
+        .load(store.pools[0].disk_set[0].clone(), &path_join_buf(&[PROOF_BUCKET, DATA_USAGE_CACHE_NAME]))
+        .await
+        .expect("real bucket checkpoint must be persisted before root publication");
+    assert!(bucket_cache.info.snapshot_complete);
+    assert_eq!(
+        bucket_cache
+            .checked_flatten(PROOF_BUCKET)
+            .expect("persisted bucket root")
+            .objects,
+        1
+    );
+    let stale_baseline = read_data_usage_persist_baseline(store.clone())
+        .await
+        .expect("missing root revision");
+    assert_eq!(stale_baseline.revision, DataUsageCacheRevision::Missing);
+    let mut competing = candidate.clone();
+    competing.scanner_epoch = Some(PROOF_EPOCH + 1);
+    competing.scanner_cycle = Some(PROOF_CYCLE + 1);
+    for state in &mut competing.usage_snapshot_set_states {
+        state.scanner_epoch = Some(PROOF_EPOCH + 1);
+        state.scanner_cycle = Some(PROOF_CYCLE + 1);
+    }
+    let competing_bytes = serde_json::to_vec(&competing).expect("competing root");
+    save_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str(), competing_bytes.clone())
+        .await
+        .expect("another publisher wins the actual root slot");
+    let before = read_root(&store).await;
+    let dirty = crate::scanner_io::dirty_usage_buckets_for_tests();
+
+    let publication = publish_candidate(&store, &scan, candidate, Some(stale_baseline)).await;
+    assert_eq!(
+        publication.outcome(),
+        DataUsagePersistOutcome::Current,
+        "the old missing revision loses CAS and reconciles the newer root"
+    );
+    let (_, _, acknowledgements) = finalize_scanner_cycle_result(scan, publication);
+    assert!(acknowledgements.is_empty());
+    assert_eq!(crate::scanner_io::dirty_usage_buckets_for_tests(), dirty);
+    assert_eq!(
+        read_root(&store).await,
+        before,
+        "bucket durability must not authorize replacing the winning root"
+    );
+    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
+}
+
+#[tokio::test]
+#[serial]
+async fn scoped_ack_publication_cannot_transfer_proof_between_real_scan_results() {
+    let (_directory, store) = candidate_store().await;
+    let (first_scan, first_candidate) = complete_candidate(&store, PROOF_CYCLE).await;
+    let (second_scan, second_candidate) = complete_candidate(&store, PROOF_CYCLE).await;
+    assert_eq!(first_candidate.scanner_epoch, second_candidate.scanner_epoch);
+    assert_eq!(first_candidate.scanner_cycle, second_candidate.scanner_cycle);
+    assert_eq!(first_candidate.objects_total_count, second_candidate.objects_total_count);
+    let baseline = read_data_usage_persist_baseline(store.clone())
+        .await
+        .expect("initial root revision");
+    let dirty = crate::scanner_io::dirty_usage_buckets_for_tests();
+    let publication = publish_candidate(&store, &first_scan, first_candidate, Some(baseline)).await;
+    assert_eq!(publication.outcome(), DataUsagePersistOutcome::Saved);
+    assert!(read_root(&store).await.0.is_some(), "the first scan really published its root");
+
+    let (_, pending, acknowledgements) = finalize_scanner_cycle_result(second_scan, publication);
+    assert!(pending, "another scan's publication must not finish this scan's dirty maintenance work");
+    assert!(acknowledgements.is_empty());
+    assert_eq!(
+        crate::scanner_io::dirty_usage_buckets_for_tests(),
+        dirty,
+        "same counters and cycle cannot transfer another scan's proof"
+    );
+
+    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
+}
+
+#[tokio::test]
+#[serial]
+async fn scoped_ack_publication_stale_baseline_cannot_prove_a_replaced_root() {
+    let (_directory, store) = candidate_store().await;
+    let (first_scan, first_candidate) = complete_candidate(&store, PROOF_CYCLE).await;
+    save_config(
+        store.clone(),
+        DATA_USAGE_OBJ_NAME_PATH.as_str(),
+        serde_json::to_vec(&first_candidate).expect("first candidate"),
+    )
+    .await
+    .expect("persist the first candidate on real disks");
+    let stale_baseline = read_data_usage_persist_baseline(store.clone())
+        .await
+        .expect("capture the genuine first root revision");
+    let dirty = crate::scanner_io::dirty_usage_buckets_for_tests();
+    let mut reader = PutObjReader::from_vec(b"second".to_vec());
+    store.pools[0].disk_set[0]
+        .put_object(PROOF_BUCKET, "second", &mut reader, &ObjectOptions::default())
+        .await
+        .expect("commit a real namespace change");
+    assert_eq!(
+        crate::scanner_io::dirty_usage_buckets_for_tests(),
+        dirty,
+        "direct storage writes leave this fixture's scanner hint generation unchanged"
+    );
+    let (_, replacement) = complete_candidate(&store, PROOF_CYCLE).await;
+    assert_eq!(first_candidate.scanner_epoch, replacement.scanner_epoch);
+    assert_eq!(first_candidate.scanner_cycle, replacement.scanner_cycle);
+    assert_eq!((first_candidate.objects_total_count, replacement.objects_total_count), (1, 2));
+    save_config(
+        store.clone(),
+        DATA_USAGE_OBJ_NAME_PATH.as_str(),
+        serde_json::to_vec(&replacement).expect("replacement candidate"),
+    )
+    .await
+    .expect("publish the replacement root");
+    let current = read_root(&store).await;
+    assert_ne!(current.1, stale_baseline.revision);
+
+    // The supplied baseline still equals candidate A, but the actual target
+    // now contains B. Compatibility's AlreadyDurable outcome is not proof.
+    let publication = publish_candidate(&store, &first_scan, first_candidate, Some(stale_baseline)).await;
+    assert_eq!(publication.outcome(), DataUsagePersistOutcome::AlreadyDurable);
+    let (_, pending, acknowledgements) = finalize_scanner_cycle_result(first_scan, publication);
+    assert!(pending);
+    assert!(acknowledgements.is_empty());
+    assert_eq!(crate::scanner_io::dirty_usage_buckets_for_tests(), dirty);
+    assert_eq!(read_root(&store).await, current);
+    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
+}
+
+#[tokio::test]
+#[serial]
+async fn scoped_ack_publication_rejects_builder_mutation_after_real_root_publish() {
+    for mutation in ["remote_ack_target", "publication_epoch", "remote_lease_targets"] {
+        let (_directory, store) = candidate_store().await;
+        let (scan, candidate) = complete_candidate(&store, PROOF_CYCLE).await;
+        let baseline = read_data_usage_persist_baseline(store.clone())
+            .await
+            .expect("initial root revision");
+        let dirty = crate::scanner_io::dirty_usage_buckets_for_tests();
+        let changed_generation = dirty
+            .get(PROOF_BUCKET)
+            .expect("the real scan has dirty work")
+            .checked_add(1)
+            .expect("bounded fixture generation");
+        let changed_epoch = scan
+            .publication_epoch()
+            .expect("real scan publication epoch")
+            .checked_add(1)
+            .expect("bounded fixture epoch");
+        let publication = publish_candidate(&store, &scan, candidate.clone(), Some(baseline)).await;
+        assert_eq!(publication.outcome(), DataUsagePersistOutcome::Saved, "{mutation}");
+        let root_before = read_root(&store).await;
+        assert_eq!(
+            serde_json::from_slice::<DataUsageInfo>(root_before.0.as_deref().expect("actual saved root"))
+                .expect("persisted root JSON"),
+            candidate,
+            "{mutation}: the original candidate really reached root storage"
+        );
+
+        let changed = match mutation {
+            "remote_ack_target" => scan.with_remote_dirty_usage_acknowledgements(vec![ScannerDirtyUsageAcknowledgement {
+                host: "proof-peer:9000".to_string(),
+                instance_id: crate::scanner_activity_epoch().to_string(),
+                kind: crate::scanner::ScannerDirtyUsageAcknowledgementKind::Generation(changed_generation),
+            }]),
+            "publication_epoch" => scan.with_publication_epoch(Some(changed_epoch)),
+            "remote_lease_targets" => scan.with_remote_publication_lease_targets(vec![(
+                "proof-peer:9000".to_string(),
+                crate::scanner_activity_epoch().to_string(),
+                changed_generation,
+            )]),
+            _ => unreachable!("fixed mutation cases"),
+        };
+        let (_, pending, acknowledgements) = finalize_scanner_cycle_result(changed, publication);
+        assert!(
+            acknowledgements.is_empty(),
+            "{mutation}: the old root proof must not authorize changed ACK work"
+        );
+        assert!(pending, "{mutation}: changed maintenance work must remain pending");
+        assert_eq!(
+            crate::scanner_io::dirty_usage_buckets_for_tests(),
+            dirty,
+            "{mutation}: the changed scan must not clear local dirty work"
+        );
+        assert_eq!(read_root(&store).await, root_before, "{mutation}: the durable original root is retained");
+    }
+    crate::scanner_io::clear_dirty_usage_buckets_for_tests();
+}
@@ -34,6 +34,139 @@ pub(super) enum DataUsagePersistOutcome {
    Failed,
 }

+#[derive(Debug)]
+pub(crate) struct RootPublicationProof {
+    candidate: crate::scanner_io::ScannerPublicationExpectation,
+    root_version: (String, [u8; 32]),
+}
+
+impl RootPublicationProof {
+    pub(crate) fn verified_version_for(
+        &self,
+        expected: &crate::scanner_io::ScannerPublicationExpectation,
+    ) -> Option<&(String, [u8; 32])> {
+        self.candidate.same_candidate(expected).then_some(&self.root_version)
+    }
+}
+
+#[derive(Debug)]
+pub(super) struct DataUsagePublicationResult {
+    outcome: DataUsagePersistOutcome,
+    proof: Option<RootPublicationProof>,
+}
+
+impl From<DataUsagePersistOutcome> for DataUsagePublicationResult {
+    fn from(outcome: DataUsagePersistOutcome) -> Self {
+        Self { outcome, proof: None }
+    }
+}
+
+impl DataUsagePublicationResult {
+    pub(super) fn outcome(&self) -> DataUsagePersistOutcome {
+        self.outcome
+    }
+    pub(super) fn restrict_outcome(&mut self, outcome: DataUsagePersistOutcome) {
+        if outcome != self.outcome {
+            self.proof = None;
+        }
+        self.outcome = outcome;
+    }
+    pub(super) fn into_parts(self) -> (DataUsagePersistOutcome, Option<RootPublicationProof>) {
+        (self.outcome, self.proof)
+    }
+}
+
+fn root_ack_write_is_confirmed<T, E>(
+    result: &std::result::Result<T, E>,
+    state: Option<ScannerPublicationCommitState>,
+    written_etag: Option<&str>,
+) -> bool {
+    result.is_ok() && state == Some(ScannerPublicationCommitState::Committed) && written_etag.is_some_and(|etag| !etag.is_empty())
+}
+
+#[cfg(test)]
+mod root_publication_confirmation_tests {
+    use super::*;
+
+    #[test]
+    fn root_publication_confirmation_requires_committed_state_and_write_revision() {
+        let saved = Ok::<(), ()>(());
+        for state in [
+            None,
+            Some(ScannerPublicationCommitState::Admitted),
+            Some(ScannerPublicationCommitState::InFlight),
+            Some(ScannerPublicationCommitState::AbortedBeforeCommit),
+            Some(ScannerPublicationCommitState::Indeterminate),
+        ] {
+            assert!(!root_ack_write_is_confirmed(&saved, state, Some("revision")));
+        }
+        for etag in [None, Some("")] {
+            assert!(!root_ack_write_is_confirmed(&saved, Some(ScannerPublicationCommitState::Committed), etag));
+        }
+        assert!(root_ack_write_is_confirmed(
+            &saved,
+            Some(ScannerPublicationCommitState::Committed),
+            Some("revision")
+        ));
+    }
+
+    #[test]
+    fn root_publication_confirmation_does_not_carry_state_across_cas_attempts() {
+        let attempts = [
+            (Err(()), Some(ScannerPublicationCommitState::Committed), Some("first")),
+            (Ok(()), Some(ScannerPublicationCommitState::AbortedBeforeCommit), Some("second")),
+            (Ok(()), None, Some("legacy")),
+            (Ok(()), Some(ScannerPublicationCommitState::Committed), Some("confirmed")),
+        ];
+        let confirmations = attempts
+            .iter()
+            .map(|(result, state, etag)| root_ack_write_is_confirmed(result, *state, *etag))
+            .collect::<Vec<_>>();
+        assert_eq!(confirmations, [false, false, false, true]);
+    }
+}
+
+async fn read_root_publication_proof<S: ScannerObjectIO + ScannerConfigObjectDelete>(
+    store: Arc<S>,
+    ctx: &CancellationToken,
+    deadline: tokio::time::Instant,
+    epoch: u64,
+    expected: &crate::scanner_io::ScannerPublicationExpectation,
+    candidate: &DataUsageInfo,
+    written_etag: Option<&str>,
+) -> Option<RootPublicationProof> {
+    let read = async {
+        let _admission = scanner_publication_admission_for_epoch(store.clone(), epoch).await?;
+        let (bytes, revision) = read_config_with_revision(store, DATA_USAGE_OBJ_NAME_PATH.as_str())
+            .await
+            .ok()?;
+        let bytes = bytes?;
+        let DataUsageCacheRevision::Etag(etag) = revision else {
+            return None;
+        };
+        if etag.is_empty() || written_etag.is_some_and(|written| written != etag) {
+            return None;
+        }
+        let persisted: DataUsageInfo = serde_json::from_slice(&bytes).ok()?;
+        if &persisted != candidate {
+            return None;
+        }
+        let root_digest = Sha256::digest(&bytes).into();
+        if ctx.is_cancelled() || tokio::time::Instant::now() >= deadline {
+            return None;
+        }
+        Some(RootPublicationProof {
+            candidate: expected.clone(),
+            root_version: (etag, root_digest),
+        })
+    };
+    tokio::select! {
+        biased;
+        _ = ctx.cancelled() => None,
+        result = tokio::time::timeout_at(deadline, read) => result.ok().flatten(),
+    }
+}
+
 fn remote_lease_expired(deadline: Option<std::time::Instant>) -> bool {
    deadline.is_some_and(|deadline| std::time::Instant::now() >= deadline)
 }
@@ -166,6 +299,7 @@ pub(super) struct ScannerPublicationFence {
    pub(super) scanner_publication_lease_fence: Option<String>,
    pub(super) remote_lease_tokens: Vec<Uuid>,
    pub(super) lease_release_safe: Arc<AtomicBool>,
+    pub(super) ack_expectation: Option<crate::scanner_io::ScannerPublicationExpectation>,
 }

 impl ScannerPublicationFence {
@@ -180,6 +314,7 @@ impl ScannerPublicationFence {
            scanner_publication_lease_fence,
            remote_lease_tokens: Vec::new(),
            lease_release_safe: Arc::new(AtomicBool::new(true)),
+            ack_expectation: None,
        }
    }

@@ -192,21 +327,26 @@ impl ScannerPublicationFence {
        self.lease_release_safe = lease_release_safe;
        self
    }
+
+    pub(super) fn with_ack_expectation(mut self, expected: Option<crate::scanner_io::ScannerPublicationExpectation>) -> Self {
+        self.ack_expectation = expected;
+        self
+    }
 }

 #[derive(Debug)]
-pub(super) enum DataUsagePersistTaskResult {
-    Completed(DataUsagePersistOutcome),
+pub(super) enum DataUsagePersistTaskResult<T = DataUsagePersistOutcome> {
+    Completed(T),
    Cancelled,
    TimedOut,
    JoinFailed(tokio::task::JoinError),
 }

-pub(super) async fn wait_for_data_usage_persist_task(
+pub(super) async fn wait_for_data_usage_persist_task<T>(
    ctx: &CancellationToken,
-    task: &mut AbortOnDropHandle<DataUsagePersistOutcome>,
+    task: &mut AbortOnDropHandle<T>,
    timeout: Duration,
-) -> DataUsagePersistTaskResult {
+) -> DataUsagePersistTaskResult<T> {
    tokio::select! {
        biased;
        result = &mut *task => match result {
@@ -320,6 +460,7 @@ where
        route_probe,
    )
    .await
+    .outcome()
 }

 pub(super) async fn store_data_usage_in_backend_with_outcome_for_epoch_and_baseline_and_route_probe_for_publication_epoch_and_lease_fence<
@@ -333,7 +474,7 @@ pub(super) async fn store_data_usage_in_backend_with_outcome_for_epoch_and_basel
    initial_baseline: Option<DataUsagePersistBaseline>,
    publication_fence: ScannerPublicationFence,
    route_probe: F,
-) -> DataUsagePersistOutcome
+) -> DataUsagePublicationResult
 where
    F: Fn() -> Fut + Send + Sync,
    Fut: Future<Output = Option<ScannerCycleDeferReason>> + Send,
@@ -344,11 +485,15 @@ where
        scanner_publication_lease_fence,
        remote_lease_tokens,
        lease_release_safe,
+        ack_expectation,
    } = publication_fence;
+    let ack_deadline = scanner_publication_scope_deadline(data_usage_persist_timeout(), remote_lease_deadline);
    let mut outcome = DataUsagePersistOutcome::NoUpdate;
+    let mut proof = None;
    let mut next_baseline = initial_baseline;

    'updates: while let Some(mut data_usage_info) = receiver.recv().await {
+        proof = None;
        let _activity_guard = ScannerActivityGuard::new();
        if ctx.is_cancelled() {
            break;
@@ -523,10 +668,14 @@ where
                continue;
            }
        };
-        let sha256hex = (!data.is_empty()).then(|| hex_simd::encode_to_string(Sha256::digest(&data), hex_simd::AsciiCase::Lower));
+        let data_digest: [u8; 32] = Sha256::digest(&data).into();
+        let sha256hex = (!data.is_empty()).then(|| hex_simd::encode_to_string(data_digest, hex_simd::AsciiCase::Lower));
        let data = Bytes::from(data);
        let backup_due = !observational && data_usage_backup_due(&data_usage_info);
        let mut cas_retry = 0usize;
+        let mut ack_epoch = None;
+        let mut write_confirmed = false;
+        let mut written_etag = None;
        let save_outcome = loop {
            if ctx.is_cancelled() {
                break 'updates;
@@ -557,6 +706,7 @@ where
            } else {
                None
            };
+            ack_epoch = Some(publication_epoch_for_save);
            let (existing_data, revision) = match baseline {
                Some(baseline) => (baseline.data, baseline.revision),
                None => match read_config_with_revision(storeapi.clone(), target_path).await {
@@ -645,7 +795,7 @@ where
            }

            let done_save = Metrics::time(Metric::SaveUsage);
-            let save_result = {
+            let (save_result, commit_state) = {
                let publication_scope = storeapi
                    .scanner_data_usage_publication_commit_scope_with_release_flag(
                        publication_epoch_for_save,
@@ -681,24 +831,33 @@ where
                .await;
                drop(legacy_publication_admission);
                if let Some(scope) = publication_scope {
-                    match scope.wait_for_completion().await {
-                        ScannerPublicationCommitState::Committed | ScannerPublicationCommitState::AbortedBeforeCommit => {
-                            save_result
-                        }
+                    let state = scope.wait_for_completion().await;
+                    let result = match state {
+                        ScannerPublicationCommitState::Committed => save_result,
+                        ScannerPublicationCommitState::AbortedBeforeCommit => save_result,
                        ScannerPublicationCommitState::Indeterminate
                        | ScannerPublicationCommitState::Admitted
                        | ScannerPublicationCommitState::InFlight => Err(EcstoreError::other(
                            "scanner publication commit scope did not reach a safe terminal state",
                        )),
-                    }
+                    };
+                    (result, Some(state))
                } else {
-                    save_result
+                    (save_result, None)
                }
            };
            done_save();

+            let attempt_confirmed = root_ack_write_is_confirmed(
+                &save_result,
+                commit_state,
+                save_result.as_ref().ok().and_then(|info| info.etag.as_deref()),
+            );
+
            match save_result {
                Ok(object_info) => {
+                    write_confirmed = attempt_confirmed;
+                    written_etag = object_info.etag.as_ref().filter(|etag| !etag.is_empty()).cloned();
                    if !observational {
                        next_baseline = object_info
                            .etag
@@ -909,9 +1068,27 @@ where
                break 'updates;
            }
        }
+        if !observational
+            && data_usage_info.usage_snapshot_converged == Some(true)
+            && matches!(outcome, DataUsagePersistOutcome::Saved | DataUsagePersistOutcome::AlreadyDurable)
+            && (outcome == DataUsagePersistOutcome::AlreadyDurable || write_confirmed)
+            && let (Some(expected), Some(epoch)) = (ack_expectation.as_ref(), ack_epoch)
+            && expected.matches_encoded_candidate(&data_digest)
+        {
+            proof = read_root_publication_proof(
+                storeapi.clone(),
+                &ctx,
+                ack_deadline,
+                epoch,
+                expected,
+                &data_usage_info,
+                written_etag.as_deref(),
+            )
+            .await;
+        }
    }

-    outcome
+    DataUsagePublicationResult { outcome, proof }
 }

 async fn cleanup_observed_data_usage_snapshot_for_epoch_and_lease(
@@ -675,6 +675,28 @@ fn partial_cache_is_useful(root: &DataUsageEntry, pending_heals_changed: bool) -
    data_usage_root_has_progress(root) || pending_heals_changed
 }

+/// Process-local hint that narrows a dirty bucket scan to known changed direct children.
+///
+/// The hint is used only while rebuilding a complete bucket cache. It never
+/// changes usage publication semantics and callers must discard it when the
+/// mutation source cannot be verified.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct ScannerBucketPrefixScanScope {
+    selected_top_level_entries: Arc<HashSet<String>>,
+}
+
+impl ScannerBucketPrefixScanScope {
+    pub(crate) fn from_dirty_top_level_entries(entries: HashSet<String>) -> Option<Self> {
+        (!entries.is_empty()).then(|| Self {
+            selected_top_level_entries: Arc::new(entries),
+        })
+    }
+
+    fn contains(&self, entry: &str) -> bool {
+        self.selected_top_level_entries.contains(entry)
+    }
+}
+
 /// Folder scanner for scanning directory structures
 pub struct FolderScanner {
    root: String,
@@ -686,6 +708,7 @@ pub struct FolderScanner {
    heal_object_select: u32,
    scan_mode: HealScanMode,
    is_erasure_mode: bool,
+    prefix_scan_scope: Option<ScannerBucketPrefixScanScope>,

    failed_object_ttl_secs: u64,
    failed_objects_max: usize,
@@ -710,6 +733,10 @@ pub struct FolderScanner {
    coverage_frontier: Option<String>,
    resume_frontier: Option<String>,
    coverage_gap: bool,
+    pending_heal_sync_deferred: bool,
+    pending_heal_batch_dirty: bool,
+    #[cfg(test)]
+    pending_heal_sync_count: usize,
    pending_size_reconciliation_keys: HashSet<String>,
    pending_size_reconciliation_scopes: HashSet<String>,
    pending_size_reconciliation_truncated: bool,
@@ -777,6 +804,47 @@ impl FolderScanner {
            .as_secs()
    }

+    fn should_reuse_clean_root_child(
+        &self,
+        folder: &CachedFolder,
+        into: &DataUsageEntry,
+        child: &CachedFolder,
+        child_hash: &DataUsageHash,
+        abandoned_children: &DataUsageHashMap,
+    ) -> bool {
+        let Some(prefix_scan_scope) = &self.prefix_scan_scope else {
+            return false;
+        };
+        // Erasure-mode usage scans also perform probabilistic object-health
+        // work. Reusing a clean subtree here would silently suppress that
+        // independent maintenance path, so prefix reuse is limited to the
+        // non-erasure data-usage scanner.
+        if self.is_erasure_mode
+            || folder.parent.is_some()
+            || folder.name != self.old_cache.info.name
+            || into.compacted
+            || !abandoned_children.contains(&child_hash.key())
+        {
+            return false;
+        }
+
+        let Some(entry) = child
+            .name
+            .strip_prefix(folder.name.as_str())
+            .and_then(|entry| entry.strip_prefix('/'))
+        else {
+            return false;
+        };
+        !entry.is_empty() && !entry.contains('/') && !prefix_scan_scope.contains(entry)
+    }
+
+    fn reuse_clean_root_child(&mut self, child: &CachedFolder, child_hash: &DataUsageHash, into: &mut DataUsageEntry) {
+        self.new_cache.copy_with_children(&self.old_cache, child_hash, &child.parent);
+        self.update_cache
+            .copy_with_children(&self.old_cache, child_hash, &child.parent);
+        into.add_child(child_hash);
+    }
+
    fn should_skip_failed(&self, path: &str) -> bool {
        let ttl = self.failed_object_ttl_secs;
        if ttl == 0 {
@@ -1080,7 +1148,7 @@ impl FolderScanner {
            scan_mode,
            result,
        );
-        if result.is_admitted() {
+        if result.is_admitted() || matches!(priority, HealChannelPriority::Low) {
            return Ok(result);
        }

@@ -1447,13 +1515,16 @@ impl FolderScanner {
                        continue;
                    }

-                    abandoned_children.remove(&h.key());
-
-                    if exists {
+                    if exists && self.should_reuse_clean_root_child(&folder, into, &this, &h, &abandoned_children) {
+                        abandoned_children.remove(&h.key());
+                        self.reuse_clean_root_child(&this, &h, into);
+                    } else if exists {
+                        abandoned_children.remove(&h.key());
                        existing_folders.push(this);
                        self.update_cache
                            .copy_with_children(&self.old_cache, &h, &Some(this_hash.clone()));
                    } else {
+                        abandoned_children.remove(&h.key());
                        new_folders.push(this);
                    }
                    continue;
@@ -1629,11 +1700,15 @@ impl FolderScanner {
            if !found_object_metadata && !found_erasure_data_directory {
                for (candidate, exists, _) in erasure_data_directory_candidates {
                    let h = hash_path(&candidate.name);
-                    abandoned_children.remove(&h.key());
-                    if exists {
+                    if exists && self.should_reuse_clean_root_child(&folder, into, &candidate, &h, &abandoned_children) {
+                        abandoned_children.remove(&h.key());
+                        self.reuse_clean_root_child(&candidate, &h, into);
+                    } else if exists {
+                        abandoned_children.remove(&h.key());
                        self.update_cache.copy_with_children(&self.old_cache, &h, &candidate.parent);
                        existing_folders.push(candidate);
                    } else {
+                        abandoned_children.remove(&h.key());
                        new_folders.push(candidate);
                    }
                }
@@ -1828,7 +1903,12 @@ impl FolderScanner {
                        }
                    }
                    FolderScanSource::Existing => {
-                        if !forward_sweep && !into.compacted && self.old_cache.is_compacted(&h) {
+                        // Usage sampling is not proof that a Deep check ran.
+                        if self.scan_mode != HealScanMode::Deep
+                            && !forward_sweep
+                            && !into.compacted
+                            && self.old_cache.is_compacted(&h)
+                        {
                            let next_cycle = self.old_cache.info.next_cycle as u32;
                            if !h.mod_(next_cycle, data_usage_update_dir_cycles()) {
                                // Transfer and add as child...
@@ -2368,6 +2448,21 @@ pub async fn scan_data_folder(
    updates: Option<mpsc::Sender<DataUsageEntry>>,
    scan_mode: HealScanMode,
    sleeper: DynamicSleeper,
+) -> Result<DataUsageCache, ScannerError> {
+    scan_data_folder_scoped(ctx, budget, disks, local_disk, cache, updates, scan_mode, sleeper, None).await
+}
+
+#[allow(clippy::too_many_arguments)]
+pub(crate) async fn scan_data_folder_scoped(
+    ctx: CancellationToken,
+    budget: Arc<ScannerCycleBudget>,
+    disks: Vec<Arc<Disk>>,
+    local_disk: Arc<Disk>,
+    cache: DataUsageCache,
+    updates: Option<mpsc::Sender<DataUsageEntry>>,
+    scan_mode: HealScanMode,
+    sleeper: DynamicSleeper,
+    prefix_scan_scope: Option<ScannerBucketPrefixScanScope>,
 ) -> Result<DataUsageCache, ScannerError> {
    use crate::data_usage_define::DATA_USAGE_ROOT;

@@ -2419,6 +2514,7 @@ pub async fn scan_data_folder(
        heal_object_select,
        scan_mode,
        is_erasure_mode,
+        prefix_scan_scope,
        failed_object_ttl_secs: failed_object_ttl,
        failed_objects_max,
        sleeper,
@@ -2435,6 +2531,10 @@ pub async fn scan_data_folder(
        coverage_frontier: resume_frontier.clone(),
        resume_frontier,
        coverage_gap: false,
+        pending_heal_sync_deferred: false,
+        pending_heal_batch_dirty: false,
+        #[cfg(test)]
+        pending_heal_sync_count: 0,
        pending_size_reconciliation_keys: HashSet::new(),
        pending_size_reconciliation_scopes: HashSet::new(),
        pending_size_reconciliation_truncated: false,
@@ -11,13 +11,58 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
-/// The pending-scanner-heal ledger: durable heal intents recorded during scans and retried after MRF consumption.
+/// Persisted best-effort retry hints. Existing TTL/count pruning applies only
+/// to this hint cache, never to a committed durable repair obligation.
 use super::*;

+const PENDING_HEAL_RETRY_BASE_SECS: u64 = 15 * 60;
+const PENDING_HEAL_RETRY_CAP_SECS: u64 = 6 * 60 * 60;
+
+pub(super) struct PendingHealSyncBatch<'a> {
+    pub(super) scanner: &'a mut FolderScanner,
+}
+
+impl<'a> PendingHealSyncBatch<'a> {
+    pub(super) fn new(scanner: &'a mut FolderScanner) -> Self {
+        scanner.pending_heal_sync_deferred = true;
+        scanner.pending_heal_batch_dirty = false;
+        Self { scanner }
+    }
+}
+
+impl Drop for PendingHealSyncBatch<'_> {
+    fn drop(&mut self) {
+        self.scanner.pending_heal_sync_deferred = false;
+        if self.scanner.pending_heal_batch_dirty {
+            self.scanner.pending_heal_batch_dirty = false;
+            self.scanner.sync_pending_heals();
+        }
+    }
+}
+
+pub(super) fn record_pending_heal_attempt(entry: &mut PendingScannerHeal, now: u64) {
+    entry.last_attempt = now;
+    entry.attempts = entry.attempts.saturating_add(1);
+}
+
+pub(super) fn observe_pending_heal_admission(entry: &mut PendingScannerHeal, result: HealAdmissionResult) {
+    // Rediscovery and coalesced admissions must not postpone an armed retry.
+    entry.last_admission_result = result.result_label().to_string();
+    entry.last_admission_reason = result.reason_label().to_string();
+}
+
 impl FolderScanner {
    pub(super) fn sync_pending_heals(&mut self) {
-        self.update_cache.info.pending_heals = self.new_cache.info.pending_heals.clone();
        self.pending_heals_changed = true;
+        if self.pending_heal_sync_deferred {
+            self.pending_heal_batch_dirty = true;
+            return;
+        }
+        self.update_cache.info.pending_heals = self.new_cache.info.pending_heals.clone();
+        #[cfg(test)]
+        {
+            self.pending_heal_sync_count += 1;
+        }
    }

    pub(super) fn clear_pending_scanner_heal(
@@ -37,32 +82,6 @@ impl FolderScanner {
        }
    }

-    /// Batched variant of [`Self::clear_pending_scanner_heal`] for repaired
-    /// notices (backlog#1894 axis B): one retain pass and one ledger sync
-    /// for the whole notice set, so a mass-recovery first sweep cannot turn
-    /// into thousands of full-table clones on the scan task. Only Object
-    /// entries match — bucket-level heals are never the MRF consumer's work.
-    pub(super) fn clear_pending_scanner_heals_for_repaired(&mut self, events: &[rustfs_common::mrf_channel::MrfRepairedEvent]) {
-        // Pre-resolve the notice version strings once; each ledger entry then
-        // compares against plain Option<&str>.
-        let targets: Vec<(&str, &str, Option<String>)> = events
-            .iter()
-            .map(|event| (event.bucket.as_ref(), event.object.as_ref(), mrf_repaired_version_id(event.version_id)))
-            .collect();
-        let before = self.new_cache.info.pending_heals.len();
-        self.new_cache.info.pending_heals.retain(|entry| {
-            entry.kind != PendingScannerHealKind::Object
-                || !targets.iter().any(|(bucket, object, version)| {
-                    entry.bucket.as_str() == *bucket
-                        && entry.object.as_deref() == Some(*object)
-                        && entry.version_id.as_deref() == version.as_deref()
-                })
-        });
-        if self.new_cache.info.pending_heals.len() != before {
-            self.sync_pending_heals();
-        }
-    }
-
    pub(super) fn record_pending_scanner_heal(
        &mut self,
        kind: PendingScannerHealKind,
@@ -80,10 +99,7 @@ impl FolderScanner {
            .iter_mut()
            .find(|entry| pending_scanner_heal_matches(entry, kind, bucket, object, version_id))
        {
-            entry.last_attempt = now;
-            entry.attempts = entry.attempts.saturating_add(1);
-            entry.last_admission_result = result.result_label().to_string();
-            entry.last_admission_reason = result.reason_label().to_string();
+            observe_pending_heal_admission(entry, result);
            self.sync_pending_heals();
            return;
        }
@@ -198,47 +214,54 @@ impl FolderScanner {
        result: HealAdmissionResult,
    ) {
        match result {
-            HealAdmissionResult::Accepted | HealAdmissionResult::Merged => {
-                self.clear_pending_scanner_heal(kind, bucket, object, version_id);
-            }
            HealAdmissionResult::Full | HealAdmissionResult::Dropped(HealAdmissionDropReason::QueueFull) => {
                self.record_pending_scanner_heal(kind, bucket, object, version_id, scan_mode, result);
            }
-            HealAdmissionResult::Dropped(HealAdmissionDropReason::PolicyDropped) => {
-                self.clear_pending_scanner_heal(kind, bucket, object, version_id);
-            }
-            // Admin-only overlap rejections (HS-06); the scanner never sees
-            // them, but if it ever does, treat them as terminal like any
-            // other policy drop rather than endlessly retrying.
-            HealAdmissionResult::Dropped(HealAdmissionDropReason::AlreadyRunning)
+            HealAdmissionResult::Accepted
+            | HealAdmissionResult::Merged
+            | HealAdmissionResult::Dropped(HealAdmissionDropReason::PolicyDropped)
+            | HealAdmissionResult::Dropped(HealAdmissionDropReason::AlreadyRunning)
            | HealAdmissionResult::Dropped(HealAdmissionDropReason::OverlappingPaths) => {
-                self.clear_pending_scanner_heal(kind, bucket, object, version_id);
+                // Admission is neither repair completion nor a durable
+                // successor receipt. Preserve existing responsibility without
+                // turning every newly admitted hint into a persisted intent.
+                if let Some(entry) = self
+                    .new_cache
+                    .info
+                    .pending_heals
+                    .iter_mut()
+                    .find(|entry| pending_scanner_heal_matches(entry, kind, bucket, object, version_id))
+                {
+                    observe_pending_heal_admission(entry, result);
+                    self.sync_pending_heals();
+                }
            }
        }
    }

    pub(super) async fn retry_pending_scanner_heals(&mut self) -> Result<(), ScannerError> {
-        if !self.should_heal().await {
+        let batch = PendingHealSyncBatch::new(self);
+        let scanner = &mut *batch.scanner;
+        if !scanner.should_heal().await {
            return Ok(());
        }

-        let bucket = self.new_cache.info.name.clone();
-        // Backlog#1894 axis B: repairs the MRF consumer landed hand the
-        // manager the heal task, so the matching pending-ledger entries are
-        // retried nowhere — drop them here. Best-effort: a lost notice just
-        // leaves the entry to expire through its own attempts/age limits.
+        let bucket = scanner.new_cache.info.name.clone();
+        // Legacy notices cannot bind a verified disposition to the current
+        // incarnation, kind, set scope and responsibility generation.
        let repaired = rustfs_common::mrf_channel::take_mrf_repaired_events_for(&bucket);
        if !repaired.is_empty() {
-            self.clear_pending_scanner_heals_for_repaired(&repaired);
+            counter!("rustfs_scanner_unverified_repair_notices_total")
+                .increment(u64::try_from(repaired.len()).unwrap_or(u64::MAX));
        }
-        self.prune_pending_scanner_heals();
-        for pending in pending_scanner_heal_retry_candidates(&self.new_cache.info.pending_heals, &bucket) {
-            if !self.should_heal().await {
+        scanner.prune_pending_scanner_heals();
+        for pending in pending_scanner_heal_retry_candidates(&scanner.new_cache.info.pending_heals, &bucket) {
+            if !scanner.should_heal().await {
                break;
            }

            let Some(request) = build_pending_scanner_heal_request(&pending) else {
-                self.clear_pending_scanner_heal(pending.kind, &pending.bucket, None, pending.version_id.as_deref());
+                scanner.clear_pending_scanner_heal(pending.kind, &pending.bucket, None, pending.version_id.as_deref());
                counter!(
                    METRIC_SCANNER_PENDING_HEAL_MALFORMED_TOTAL,
                    "bucket" => pending.bucket.clone(),
@@ -257,14 +280,27 @@ impl FolderScanner {
                continue;
            };

-            self.send_required_scanner_heal_request(
-                pending.kind,
-                pending.bucket.clone(),
-                pending.object.clone(),
-                pending.version_id.clone(),
-                request,
-            )
-            .await?;
+            if let Some(entry) = scanner.new_cache.info.pending_heals.iter_mut().find(|entry| {
+                pending_scanner_heal_matches(
+                    entry,
+                    pending.kind,
+                    &pending.bucket,
+                    pending.object.as_deref(),
+                    pending.version_id.as_deref(),
+                )
+            }) {
+                record_pending_heal_attempt(entry, Self::now_secs());
+                scanner.sync_pending_heals();
+            }
+            scanner
+                .send_required_scanner_heal_request(
+                    pending.kind,
+                    pending.bucket.clone(),
+                    pending.object.clone(),
+                    pending.version_id.clone(),
+                    request,
+                )
+                .await?;
        }

        Ok(())
@@ -295,16 +331,6 @@ pub(super) fn pending_scanner_heal_identity(entry: &PendingScannerHeal) -> (u8,
    (kind, entry.bucket.as_str(), entry.object.as_deref(), entry.version_id.as_deref())
 }

-/// Decode an MRF repaired-notice version id for ledger matching. A nil UUID
-/// means "no value" per the repo-wide defensive-UUID invariant, so it maps
-/// to `None` and matches unversioned ledger entries only.
-pub(super) fn mrf_repaired_version_id(version_id: Option<[u8; 16]>) -> Option<String> {
-    version_id
-        .map(uuid::Uuid::from_bytes)
-        .filter(|uuid| !uuid.is_nil())
-        .map(|uuid| uuid.to_string())
-}
-
 pub(super) fn sort_pending_scanner_heals_for_retry(entries: &mut [PendingScannerHeal]) {
    entries.sort_by(|a, b| {
        a.last_attempt
@@ -318,30 +344,56 @@ pub(super) fn pending_scanner_heal_retry_candidates(
    pending_heals: &[PendingScannerHeal],
    bucket: &str,
 ) -> Vec<PendingScannerHeal> {
-    let mut entries: Vec<PendingScannerHeal> = pending_heals.iter().filter(|entry| entry.bucket == bucket).cloned().collect();
-    sort_pending_scanner_heals_for_retry(&mut entries);
+    pending_scanner_heal_retry_candidates_at(pending_heals, bucket, FolderScanner::now_secs())
+}
+
+pub(super) fn pending_scanner_heal_retry_candidates_at(
+    pending_heals: &[PendingScannerHeal],
+    bucket: &str,
+    now: u64,
+) -> Vec<PendingScannerHeal> {
+    // Schedule across scanner cycles rather than allocating a timer per hint.
+    // A later Full response must not reset an already retried hint's backoff.
+    let mut entries: Vec<&PendingScannerHeal> = pending_heals
+        .iter()
+        .filter(|entry| {
+            let exponent = entry.attempts.saturating_sub(1).min(31);
+            let delay = PENDING_HEAL_RETRY_BASE_SECS
+                .saturating_mul(1_u64 << exponent)
+                .min(PENDING_HEAL_RETRY_CAP_SECS);
+            entry.bucket == bucket && now.checked_sub(entry.last_attempt).is_some_and(|age| age >= delay)
+        })
+        .collect();
+    entries.sort_by(|a, b| {
+        a.last_attempt
+            .cmp(&b.last_attempt)
+            .then_with(|| a.attempts.cmp(&b.attempts))
+            .then_with(|| pending_scanner_heal_identity(a).cmp(&pending_scanner_heal_identity(b)))
+    });
    entries.truncate(MAX_PENDING_SCANNER_HEAL_RETRIES_PER_BUCKET);
-    entries
+    entries.into_iter().cloned().collect()
 }

 pub(super) fn build_pending_scanner_heal_request(entry: &PendingScannerHeal) -> Option<HealChannelRequest> {
+    let priority = if entry.last_admission_result == "full"
+        || (entry.last_admission_result == "dropped" && entry.last_admission_reason == "queue_full")
+    {
+        HealChannelPriority::High
+    } else {
+        HealChannelPriority::Low
+    };
    match entry.kind {
-        PendingScannerHealKind::Bucket => Some(build_bucket_heal_request(entry.bucket.clone(), HealChannelPriority::High)),
+        PendingScannerHealKind::Bucket => Some(build_bucket_heal_request(entry.bucket.clone(), priority)),
        PendingScannerHealKind::Object => entry.object.as_ref().map(|object| {
            if entry.version_id.is_none() {
-                build_non_destructive_object_heal_request(
-                    entry.bucket.clone(),
-                    object.clone(),
-                    entry.scan_mode,
-                    HealChannelPriority::High,
-                )
+                build_non_destructive_object_heal_request(entry.bucket.clone(), object.clone(), entry.scan_mode, priority)
            } else {
                build_object_heal_request(
                    entry.bucket.clone(),
                    object.clone(),
                    entry.version_id.clone(),
                    entry.scan_mode,
-                    HealChannelPriority::High,
+                    priority,
                )
            }
        }),
@@ -15,6 +15,8 @@
 use crate::SCANNER_SLEEPER;

 use super::*;
+
+mod mrf_ownership;
 use crate::storage_api::VersionPurgeStatusType;
 use crate::{DiskOption, Endpoint, STORAGE_FORMAT_FILE, TierStats, new_disk, storageclass};
 use rustfs_filemeta::{FileInfo, FileMeta, MetadataResolutionParams};
@@ -330,6 +332,7 @@ async fn build_test_scanner() -> (FolderScanner, std::path::PathBuf) {
        heal_object_select: 0,
        scan_mode: HealScanMode::Normal,
        is_erasure_mode: false,
+        prefix_scan_scope: None,
        failed_object_ttl_secs: u64::MAX,
        failed_objects_max: usize::MAX,
        sleeper: SCANNER_SLEEPER.clone(),
@@ -350,6 +353,9 @@ async fn build_test_scanner() -> (FolderScanner, std::path::PathBuf) {
        coverage_frontier: None,
        resume_frontier: None,
        coverage_gap: false,
+        pending_heal_sync_deferred: false,
+        pending_heal_batch_dirty: false,
+        pending_heal_sync_count: 0,
        pending_size_reconciliation_keys: HashSet::new(),
        pending_size_reconciliation_scopes: HashSet::new(),
        pending_size_reconciliation_truncated: false,
@@ -1132,23 +1138,8 @@ fn pending_heal(
    }
 }

-/// The nil-UUID branch of the defensive-UUID invariant: a nil version in
-/// a repaired notice means "no value" and must match unversioned ledger
-/// entries only.
-#[test]
-fn test_mrf_repaired_version_id_maps_nil_to_none() {
-    assert_eq!(mrf_repaired_version_id(None), None);
-    assert_eq!(mrf_repaired_version_id(Some([0u8; 16])), None);
-    let uuid = Uuid::new_v4();
-    assert_eq!(mrf_repaired_version_id(Some(*uuid.as_bytes())), Some(uuid.to_string()));
-}
-
-/// Full wiring of backlog#1894 axis B: notes taken for the scanned bucket
-/// clear exactly the matching Object ledger entries — bucket-level
-/// entries, other buckets' entries, and version-mismatched entries
-/// survive; a real (non-nil) version matches only the same version.
 #[tokio::test]
-async fn test_mrf_repaired_notices_clear_matching_ledger_entries() {
+async fn mrf_ownership_legacy_notices_preserve_pending_entries() {
    use rustfs_common::mrf_channel::note_mrf_repaired;

    let (mut scanner, temp_dir) = build_test_scanner().await;
@@ -1176,8 +1167,7 @@ async fn test_mrf_repaired_notices_clear_matching_ledger_entries() {

    note_mrf_repaired("bucket", "object-a", None);
    note_mrf_repaired("bucket", "object-b", Some(*Uuid::parse_str(&version).unwrap().as_bytes()));
-    // A nil-UUID notice for object-c means "no value": it clears the
-    // unversioned entry but must not touch the versioned one.
+    // Neither nil nor a matching version proves incarnation, scope or owner.
    note_mrf_repaired("bucket", "object-c", Some([0u8; 16]));
    // A notice for a target the ledger does not track must be a no-op.
    note_mrf_repaired("bucket", "object-untracked", None);
@@ -1194,12 +1184,12 @@ async fn test_mrf_repaired_notices_clear_matching_ledger_entries() {
        .iter()
        .map(|entry| (entry.kind, entry.bucket.as_str(), entry.object.as_deref(), entry.version_id.as_deref()))
        .collect();
-    // Cleared: object-a (no version), object-b (exact version match), and
-    // object-c's unversioned entry (the nil branch matched no-version
-    // only — the versioned object-c entry survives).
    assert_eq!(
        survivors,
        vec![
+            (PendingScannerHealKind::Object, "bucket", Some("object-a"), None),
+            (PendingScannerHealKind::Object, "bucket", Some("object-b"), Some(version.as_str())),
+            (PendingScannerHealKind::Object, "bucket", Some("object-c"), None),
            (
                PendingScannerHealKind::Object,
                "bucket",
@@ -1338,7 +1328,7 @@ async fn test_pending_heal_update_keeps_stale_entry_until_retry_prune() {
    );

    assert_eq!(scanner.new_cache.info.pending_heals.len(), 1);
-    assert_eq!(scanner.new_cache.info.pending_heals[0].attempts, 2);
+    assert_eq!(scanner.new_cache.info.pending_heals[0].attempts, 1);
    assert_eq!(scanner.new_cache.info.pending_heals[0].object.as_deref(), Some("object"));
    assert_eq!(scanner.update_cache.info.pending_heals, scanner.new_cache.info.pending_heals);
 }
@@ -1371,7 +1361,7 @@ async fn test_pending_heal_queue_full_deduplicates_object_entry() {
    let pending = &scanner.new_cache.info.pending_heals[0];
    assert_eq!(pending.object.as_deref(), Some("object"));
    assert_eq!(pending.version_id.as_deref(), Some("version-a"));
-    assert_eq!(pending.attempts, 2);
+    assert_eq!(pending.attempts, 1);
    assert_eq!(pending.last_admission_result, "dropped");
    assert_eq!(pending.last_admission_reason, "queue_full");
    assert_eq!(scanner.update_cache.info.pending_heals, scanner.new_cache.info.pending_heals);
@@ -1379,7 +1369,7 @@ async fn test_pending_heal_queue_full_deduplicates_object_entry() {
 }

 #[tokio::test]
-async fn test_pending_heal_admitted_results_clear_matching_entry() {
+async fn mrf_ownership_admission_preserves_existing_pending() {
    let (mut scanner, temp_dir) = build_test_scanner().await;
    let _guard = TestGuard::new(u64::MAX, usize::MAX, &mut scanner, temp_dir);

@@ -1400,7 +1390,8 @@ async fn test_pending_heal_admitted_results_clear_matching_entry() {
        HealAdmissionResult::Accepted,
    );

-    assert!(scanner.new_cache.info.pending_heals.is_empty());
+    assert_eq!(scanner.new_cache.info.pending_heals.len(), 1);
+    assert_eq!(scanner.new_cache.info.pending_heals[0].last_admission_result, "accepted");

    scanner.update_pending_scanner_heal_after_admission(
        PendingScannerHealKind::Bucket,
@@ -1419,11 +1410,12 @@ async fn test_pending_heal_admitted_results_clear_matching_entry() {
        HealAdmissionResult::Merged,
    );

-    assert!(scanner.new_cache.info.pending_heals.is_empty());
+    assert_eq!(scanner.new_cache.info.pending_heals.len(), 2);
+    assert_eq!(scanner.new_cache.info.pending_heals[1].last_admission_result, "merged");
 }

 #[tokio::test]
-async fn test_pending_heal_policy_dropped_clears_without_creating_entry() {
+async fn mrf_ownership_policy_drop_does_not_discharge_existing_pending() {
    let (mut scanner, temp_dir) = build_test_scanner().await;
    let _guard = TestGuard::new(u64::MAX, usize::MAX, &mut scanner, temp_dir);

@@ -1454,7 +1446,7 @@ async fn test_pending_heal_policy_dropped_clears_without_creating_entry() {
        HealAdmissionResult::Dropped(HealAdmissionDropReason::PolicyDropped),
    );

-    assert!(scanner.new_cache.info.pending_heals.is_empty());
+    assert_eq!(scanner.new_cache.info.pending_heals.len(), 1);
 }

 #[test]
@@ -2393,6 +2385,115 @@ async fn test_scan_folder_non_erasure_metadata_keeps_namespace_descent() {
    assert_eq!(budget.reason(), Some(crate::scanner_budget::ScannerCycleBudgetReason::Directories));
 }

+#[tokio::test]
+#[serial]
+async fn scoped_root_scan_reuses_clean_top_level_entries_and_rescans_dirty_entries() {
+    let (mut scanner, temp_dir) = build_test_scanner().await;
+    let _guard = TestGuard::new(60, 100, &mut scanner, temp_dir.clone());
+    let bucket_dir = temp_dir.join("bucket");
+    tokio::fs::create_dir_all(bucket_dir.join("clean"))
+        .await
+        .expect("failed to create clean top-level directory");
+    tokio::fs::create_dir_all(bucket_dir.join("dirty"))
+        .await
+        .expect("failed to create dirty top-level directory");
+
+    scanner.old_cache.info.name = "bucket".to_string();
+    scanner.new_cache.info.name = "bucket".to_string();
+    scanner.update_cache.info.name = "bucket".to_string();
+    scanner.old_cache.replace("bucket", "", DataUsageEntry::default());
+    scanner.old_cache.replace(
+        "bucket/clean",
+        "bucket",
+        DataUsageEntry {
+            size: 17,
+            objects: 3,
+            ..Default::default()
+        },
+    );
+    scanner.old_cache.replace(
+        "bucket/dirty",
+        "bucket",
+        DataUsageEntry {
+            size: 23,
+            objects: 4,
+            ..Default::default()
+        },
+    );
+    scanner.prefix_scan_scope = ScannerBucketPrefixScanScope::from_dirty_top_level_entries(HashSet::from(["dirty".to_string()]));
+
+    let folder = CachedFolder {
+        name: "bucket".to_string(),
+        parent: None,
+        object_heal_prob_div: 1,
+    };
+    let mut root = DataUsageEntry::default();
+    scanner
+        .scan_folder(CancellationToken::new(), folder, &mut root)
+        .await
+        .expect("scoped root scan should finish successfully");
+
+    let clean = scanner
+        .new_cache
+        .size_recursive("bucket/clean")
+        .expect("clean entry should be copied from the complete cache");
+    assert_eq!((clean.size, clean.objects), (17, 3));
+    let dirty = scanner
+        .new_cache
+        .size_recursive("bucket/dirty")
+        .expect("dirty entry should be rescanned");
+    assert_eq!((dirty.size, dirty.objects), (0, 0));
+    let bucket = scanner
+        .new_cache
+        .size_recursive("bucket")
+        .expect("bucket root should include reused and rescanned entries");
+    assert_eq!((bucket.size, bucket.objects), (17, 3));
+}
+
+#[tokio::test]
+#[serial]
+async fn scoped_root_scan_preserves_erasure_health_walks() {
+    let (mut scanner, temp_dir) = build_test_scanner().await;
+    let _guard = TestGuard::new(60, 100, &mut scanner, temp_dir.clone());
+    let bucket_dir = temp_dir.join("bucket");
+    tokio::fs::create_dir_all(bucket_dir.join("clean"))
+        .await
+        .expect("failed to create clean top-level directory");
+
+    scanner.is_erasure_mode = true;
+    scanner.old_cache.info.name = "bucket".to_string();
+    scanner.new_cache.info.name = "bucket".to_string();
+    scanner.update_cache.info.name = "bucket".to_string();
+    scanner.old_cache.replace("bucket", "", DataUsageEntry::default());
+    scanner.old_cache.replace(
+        "bucket/clean",
+        "bucket",
+        DataUsageEntry {
+            size: 17,
+            objects: 3,
+            ..Default::default()
+        },
+    );
+    scanner.prefix_scan_scope = ScannerBucketPrefixScanScope::from_dirty_top_level_entries(HashSet::from(["dirty".to_string()]));
+
+    let folder = CachedFolder {
+        name: "bucket".to_string(),
+        parent: None,
+        object_heal_prob_div: 1,
+    };
+    let mut root = DataUsageEntry::default();
+    scanner
+        .scan_folder(CancellationToken::new(), folder, &mut root)
+        .await
+        .expect("erasure root scan should finish successfully");
+
+    let clean = scanner
+        .new_cache
+        .size_recursive("bucket/clean")
+        .expect("erasure scan should visit the clean entry");
+    assert_eq!((clean.size, clean.objects), (0, 0));
+}
+
 #[tokio::test]
 #[serial]
 async fn test_scan_folder_compacted_parent_sends_partial_update() {
@@ -14,12 +14,13 @@

 use super::*;
 use crate::scanner_budget::ScannerCycleBudgetConfig;
-use crate::scanner_io::{ScannerDiskScanOutcome, ScannerIODisk};
+use crate::scanner_io::{ScannerDiskScanOptions, ScannerDiskScanOutcome, ScannerIODisk};
 use crate::storage_api::scanner_io::ObjectIO;
 use crate::{DataUsageCacheSource, DataUsageScanPlanDigest};
 use std::io::Cursor;
 use tokio::io::AsyncReadExt;

+mod deep_compacted;
 mod segment_observation;

 const CACHE_NAME: &str = "bucket/checkpoint-fixture.bin";
@@ -28,6 +29,13 @@ const MAX_CACHE_BYTES: u64 = 1024 * 1024;
 const SOURCE: DataUsageCacheSource = DataUsageCacheSource::new(0, 0);
 const PLAN: DataUsageScanPlanDigest = DataUsageScanPlanDigest([17; 32]);

+fn scan_options(scan_mode: HealScanMode) -> ScannerDiskScanOptions {
+    ScannerDiskScanOptions {
+        scan_mode,
+        prefix_scan_scope: None,
+    }
+}
+
 /// Real cache persistence codec and CAS calls, backed by two bounded local files.
 #[derive(Debug)]
 struct FixtureStore {
@@ -537,7 +545,7 @@ async fn checkpoint_fixture_existing_uncovered_cursor_cannot_skip_to_complete()
                vec![scanner.local_disk.clone()],
                loaded,
                None,
-                HealScanMode::Normal,
+                scan_options(HealScanMode::Normal),
            )
            .await
            .expect("scan must revisit the prefix");
@@ -593,7 +601,7 @@ async fn checkpoint_fixture_failed_child_prevents_receipt_advancing_past_gap() {
            vec![scanner.local_disk.clone()],
            cache,
            None,
-            HealScanMode::Normal,
+            scan_options(HealScanMode::Normal),
        )
        .await
        .expect("scan with a known failed child");
@@ -644,7 +652,7 @@ async fn check_complete_sampling_resumption(resume_mode: HealScanMode) {
            vec![scanner.local_disk.clone()],
            cache,
            None,
-            HealScanMode::Normal,
+            scan_options(HealScanMode::Normal),
        )
        .await
        .expect("initial complete baseline");
@@ -689,7 +697,7 @@ async fn check_complete_sampling_resumption(resume_mode: HealScanMode) {
            vec![scanner.local_disk.clone()],
            cache,
            None,
-            HealScanMode::Normal,
+            scan_options(HealScanMode::Normal),
        )
        .await
        .expect("sampling interruption");
@@ -736,7 +744,14 @@ async fn check_complete_sampling_resumption(resume_mode: HealScanMode) {
        let result = scanner
            .local_disk
            .clone()
-            .nsscanner_disk(budget.token(), budget, vec![scanner.local_disk.clone()], loaded, None, resume_mode)
+            .nsscanner_disk(
+                budget.token(),
+                budget,
+                vec![scanner.local_disk.clone()],
+                loaded,
+                None,
+                scan_options(resume_mode),
+            )
            .await
            .expect("bounded recovery scan");
        let (cache, complete) = match result {
@@ -852,7 +867,10 @@ async fn run_checkpoint_fixture(change_digest: bool) {
                vec![scanner.local_disk.clone()],
                cache,
                None,
-                HealScanMode::Normal,
+                ScannerDiskScanOptions {
+                    scan_mode: HealScanMode::Normal,
+                    prefix_scan_scope: None,
+                },
            )
            .await
            .expect("budgeted local disk scan returns partial cache");
@@ -930,7 +948,10 @@ async fn run_checkpoint_fixture(change_digest: bool) {
            vec![scanner.local_disk.clone()],
            loaded.clone(),
            None,
-            HealScanMode::Normal,
+            ScannerDiskScanOptions {
+                scan_mode: HealScanMode::Normal,
+                prefix_scan_scope: None,
+            },
        )
        .await;
    assert!(result.is_err(), "pre-scan cancellation must not produce a complete root");
@@ -981,7 +1002,7 @@ async fn run_checkpoint_fixture(change_digest: bool) {
                vec![scanner.local_disk.clone()],
                cache,
                None,
-                HealScanMode::Normal,
+                scan_options(HealScanMode::Normal),
            )
            .await
            .expect("bounded sweep outcome");
@@ -0,0 +1,204 @@
+// Copyright 2026 RustFS Team
+// Licensed under the Apache License, Version 2.0.
+
+use super::*;
+
+const PREFIX: &str = "bucket/prefix";
+const OBJECTS: u64 = 4;
+
+struct CompactedFixture {
+    disk: Arc<Disk>,
+    root: std::path::PathBuf,
+    cache: DataUsageCache,
+    identity: crate::DataUsageScanIdentity,
+    store: Arc<FixtureStore>,
+    _cleanup: TestGuard,
+}
+
+async fn scan(
+    disk: &Arc<Disk>,
+    cache: DataUsageCache,
+    mode: HealScanMode,
+    max_objects: u64,
+) -> (ScannerDiskScanOutcome, Arc<ScannerCycleBudget>) {
+    let budget = ScannerCycleBudget::new_with_progress_tracking(
+        &CancellationToken::new(),
+        ScannerCycleBudgetConfig {
+            max_objects: Some(max_objects),
+            ..Default::default()
+        },
+    );
+    let outcome = disk
+        .clone()
+        .nsscanner_disk(budget.token(), budget.clone(), vec![disk.clone()], cache, None, super::scan_options(mode))
+        .await
+        .expect("bounded real disk scan");
+    (outcome, budget)
+}
+
+async fn save_reload(store: &Arc<FixtureStore>, cache: &DataUsageCache) -> DataUsageCache {
+    let revisions = DataUsageCache::default()
+        .load_with_revisions(store.clone(), CACHE_NAME)
+        .await
+        .expect("fixture save revisions");
+    cache
+        .save_with_revisions_for_epoch(store.clone(), CACHE_NAME, &revisions, 0)
+        .await
+        .expect("save compacted checkpoint through real codec and revision checks");
+    let loaded = store.strict_load().await;
+    assert_eq!(loaded.info.snapshot_complete, cache.info.snapshot_complete);
+    assert_eq!(loaded.info.scan_checkpoint, cache.info.scan_checkpoint);
+    assert_eq!(loaded.info.scan_identity, cache.info.scan_identity);
+    assert_eq!(loaded.info.scan_plan_digest, cache.info.scan_plan_digest);
+    assert_eq!(
+        loaded.checked_flatten("bucket").expect("reloaded root").size,
+        cache.checked_flatten("bucket").expect("returned root").size
+    );
+    loaded
+}
+
+impl CompactedFixture {
+    async fn new(mode: HealScanMode) -> Self {
+        let (scanner, root) = build_test_scanner().await;
+        let cleanup = TestGuard {
+            temp_dir: Some(root.clone()),
+        };
+        for index in 0..OBJECTS {
+            write_checkpoint_object(&root, &format!("prefix/{index:04}"), &[(None, 1)]).await;
+        }
+        let identity = crate::DataUsageScanIdentity {
+            scan_mode: mode,
+            tier_registry_generation: crate::runtime_tier_registry_for_cycle(11, 7).await.generation,
+            ..bound_checkpoint().1
+        };
+        let mut cache = DataUsageCache::default();
+        // The first real scan builds coverage; the second same-plan scan takes
+        // the normal compaction path. No synthetic compacted cache is injected.
+        for cycle in [11, 12] {
+            cache.prepare_bucket_checkpoint("bucket", cycle, 7, SOURCE, PLAN, identity);
+            cache.info.skip_healing = true;
+            let (outcome, budget) = scan(&scanner.local_disk, cache, mode, OBJECTS + 1).await;
+            let ScannerDiskScanOutcome::Complete(completed) = outcome else {
+                panic!("fixture baseline must complete")
+            };
+            assert_eq!(budget.progress().0, OBJECTS, "baseline must read every metadata object");
+            cache = completed;
+        }
+        let store = FixtureStore::new();
+        cache = save_reload(&store, &cache).await;
+        assert!(cache.find(PREFIX).expect("baseline prefix").compacted);
+        assert_eq!(cache.checked_flatten("bucket").expect("complete baseline").size, 4);
+        assert!(cache.info.scan_progress.is_none());
+        Self {
+            disk: scanner.local_disk,
+            root,
+            cache,
+            identity,
+            store,
+            _cleanup: cleanup,
+        }
+    }
+
+    fn next_cycle(&self, sampled: bool) -> u64 {
+        let first = self.cache.info.next_cycle + 1;
+        (first..first + 16)
+            .find(|cycle| hash_path(PREFIX).mod_(u32::try_from(*cycle).expect("bounded cycle"), 16) == sampled)
+            .expect("one selected cycle and non-selected cycles exist within the fixed rotation")
+    }
+
+    fn prepare(&mut self, cycle: u64) {
+        let state = crate::scanner_io::current_cache_root_or_prepare_with_generation(
+            &mut self.cache,
+            "bucket",
+            SOURCE,
+            cycle,
+            7,
+            PLAN,
+            crate::scanner_io::DataUsageCacheReuseOptions {
+                checkpoint_identity: Some(self.identity),
+                ..Default::default()
+            },
+        );
+        assert!(matches!(state, crate::scanner_io::DataUsageCacheScanState::Prepared { .. }));
+        assert_eq!(self.cache.info.scan_identity, Some(self.identity));
+        assert_eq!(self.cache.info.scan_plan_digest, Some(PLAN));
+        assert!(
+            self.cache.info.scan_progress.is_none(),
+            "same-strength complete baseline uses the existing tree"
+        );
+        assert!(self.cache.find(PREFIX).expect("prepared prefix").compacted);
+    }
+
+    async fn change_metadata_without_activity_event(&self) {
+        // This models a local metadata change not announced by a segment
+        // producer. Deep traversal must not depend on a usage-clean signal.
+        // Healing is disabled: the oracle proves metadata re-entry, not repair.
+        write_checkpoint_object(&self.root, "prefix/0000", &[(None, 7)]).await;
+    }
+}
+
+#[tokio::test]
+#[serial]
+async fn deep_compacted_same_plan_rechecks_unsampled_prefix() {
+    temp_env::async_with_vars([(ENV_DATA_USAGE_UPDATE_DIR_CYCLES, Some("16"))], async {
+        let mut fixture = CompactedFixture::new(HealScanMode::Deep).await;
+        fixture.change_metadata_without_activity_event().await;
+        fixture.prepare(fixture.next_cycle(false));
+        let (outcome, budget) = scan(&fixture.disk, fixture.cache.clone(), HealScanMode::Deep, OBJECTS + 1).await;
+        assert_eq!(
+            budget.progress().0,
+            OBJECTS,
+            "Deep must inspect compacted children even outside the usage sample cycle"
+        );
+        let ScannerDiskScanOutcome::Complete(cache) = outcome else {
+            panic!("bounded Deep scan must complete")
+        };
+        let loaded = save_reload(&fixture.store, &cache).await;
+        let root = loaded.checked_flatten("bucket").expect("Deep scan root");
+        assert_eq!((root.objects, root.size), (4, 10), "Deep must observe the changed metadata");
+    })
+    .await;
+}
+
+#[tokio::test]
+#[serial]
+async fn deep_compacted_normal_scan_preserves_periodic_sampling() {
+    temp_env::async_with_vars([(ENV_DATA_USAGE_UPDATE_DIR_CYCLES, Some("16"))], async {
+        let mut fixture = CompactedFixture::new(HealScanMode::Normal).await;
+        fixture.change_metadata_without_activity_event().await;
+        fixture.prepare(fixture.next_cycle(false));
+        let (outcome, budget) = scan(&fixture.disk, fixture.cache.clone(), HealScanMode::Normal, OBJECTS + 1).await;
+        assert_eq!(budget.progress().0, 0, "Normal retains its existing unsampled-subtree policy");
+        let ScannerDiskScanOutcome::Complete(cache) = outcome else { panic!("normal sampling completes") };
+        assert_eq!(cache.checked_flatten("bucket").expect("sampled root").size, 4);
+        fixture.cache = save_reload(&fixture.store, &cache).await;
+        fixture.prepare(fixture.next_cycle(true));
+        let (outcome, budget) = scan(&fixture.disk, fixture.cache.clone(), HealScanMode::Normal, OBJECTS + 1).await;
+        assert_eq!(budget.progress().0, OBJECTS);
+        let ScannerDiskScanOutcome::Complete(cache) = outcome else {
+            panic!("selected Normal rotation completes")
+        };
+        assert_eq!(cache.checked_flatten("bucket").expect("refreshed root").size, 10);
+    })
+    .await;
+}
+
+#[tokio::test]
+#[serial]
+async fn deep_compacted_budget_preserves_partial_checkpoint() {
+    temp_env::async_with_vars([(ENV_DATA_USAGE_UPDATE_DIR_CYCLES, Some("16"))], async {
+        let mut fixture = CompactedFixture::new(HealScanMode::Deep).await;
+        fixture.prepare(fixture.next_cycle(false));
+        let (outcome, budget) = scan(&fixture.disk, fixture.cache.clone(), HealScanMode::Deep, 2).await;
+        assert_eq!(budget.progress().0, 2);
+        assert_eq!(budget.reason(), Some(ScannerCycleBudgetReason::Objects));
+        let ScannerDiskScanOutcome::Partial(cache) = outcome else {
+            panic!("Deep must retain budget interruption as partial")
+        };
+        assert!(!cache.info.snapshot_complete);
+        let loaded = save_reload(&fixture.store, &cache).await;
+        assert!(!loaded.info.snapshot_complete);
+        assert!(loaded.checked_flatten("bucket").expect("partial root").objects > 0);
+    })
+    .await;
+}
@@ -0,0 +1,410 @@
+// Copyright 2026 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use super::*;
+use crate::storage_api::EcstoreHealResultItem as HealItem;
+use crate::storage_api::scanner_io::BucketInfo;
+use rustfs_common::mrf_channel::{
+    MrfIngressResult, MrfKind, MrfScope, note_mrf_repaired, take_mrf_repaired_events_for, try_send_mrf_intent_typed,
+};
+use rustfs_heal::heal::{
+    manager::{HealConfig, HealManager},
+    mrf_queue::spawn_mrf_consumer,
+    storage::{HealListItem, HealObjectInfo, HealStorageAPI},
+};
+use rustfs_heal_contracts::heal_channel::HealOpts;
+
+#[tokio::test]
+async fn mrf_ownership_admission_observation_does_not_postpone_retry() {
+    let (mut scanner, temp_dir) = build_test_scanner().await;
+    let _guard = TestGuard::new(u64::MAX, usize::MAX, &mut scanner, temp_dir);
+    scanner.new_cache.info.pending_heals.push(pending_heal(
+        PendingScannerHealKind::Object,
+        "bucket",
+        Some("object"),
+        None,
+        100,
+        2,
+    ));
+    for result in [
+        HealAdmissionResult::Accepted,
+        HealAdmissionResult::Merged,
+        HealAdmissionResult::Full,
+        HealAdmissionResult::Dropped(HealAdmissionDropReason::QueueFull),
+        HealAdmissionResult::Dropped(HealAdmissionDropReason::PolicyDropped),
+    ] {
+        scanner.update_pending_scanner_heal_after_admission(
+            PendingScannerHealKind::Object,
+            "bucket",
+            Some("object"),
+            None,
+            HealScanMode::Deep,
+            result,
+        );
+        let entry = &scanner.new_cache.info.pending_heals[0];
+        assert_eq!((entry.last_attempt, entry.attempts), (100, 2));
+        assert!(pending_scanner_heal_retry_candidates_at(&scanner.new_cache.info.pending_heals, "bucket", 1899).is_empty());
+        assert_eq!(
+            pending_scanner_heal_retry_candidates_at(&scanner.new_cache.info.pending_heals, "bucket", 1900).len(),
+            1
+        );
+    }
+    scanner.update_pending_scanner_heal_after_admission(
+        PendingScannerHealKind::Object,
+        "bucket",
+        Some("new-object"),
+        None,
+        HealScanMode::Deep,
+        HealAdmissionResult::Accepted,
+    );
+    assert_eq!(
+        scanner.new_cache.info.pending_heals.len(),
+        1,
+        "successful admission does not create a new ledger"
+    );
+}
+
+#[test]
+fn mrf_ownership_retry_due_boundaries_and_priority_are_bounded() {
+    let mut entry = pending_heal(PendingScannerHealKind::Object, "bucket", Some("object"), None, 100, 1);
+    entry.last_admission_result = "accepted".to_string();
+    assert!(pending_scanner_heal_retry_candidates_at(std::slice::from_ref(&entry), "bucket", 999).is_empty());
+    assert_eq!(
+        pending_scanner_heal_retry_candidates_at(std::slice::from_ref(&entry), "bucket", 1000).len(),
+        1
+    );
+    assert_eq!(
+        build_pending_scanner_heal_request(&entry).expect("request").priority,
+        HealChannelPriority::Low
+    );
+    record_pending_heal_attempt(&mut entry, 1000);
+    observe_pending_heal_admission(&mut entry, HealAdmissionResult::Full);
+    assert!(pending_scanner_heal_retry_candidates_at(std::slice::from_ref(&entry), "bucket", 2799).is_empty());
+    assert_eq!(
+        pending_scanner_heal_retry_candidates_at(std::slice::from_ref(&entry), "bucket", 2800).len(),
+        1
+    );
+    assert_eq!(
+        build_pending_scanner_heal_request(&entry).expect("request").priority,
+        HealChannelPriority::High
+    );
+    entry.attempts = u32::MAX;
+    assert!(pending_scanner_heal_retry_candidates_at(std::slice::from_ref(&entry), "bucket", 22599).is_empty());
+    assert_eq!(
+        pending_scanner_heal_retry_candidates_at(std::slice::from_ref(&entry), "bucket", 22600).len(),
+        1
+    );
+    entry.last_attempt = u64::MAX;
+    assert!(pending_scanner_heal_retry_candidates_at(&[entry], "bucket", 22600).is_empty());
+}
+
+#[tokio::test]
+async fn mrf_ownership_full_hint_table_has_bounded_multicycle_work_and_sync() {
+    let (mut scanner, temp_dir) = build_test_scanner().await;
+    let _guard = TestGuard::new(u64::MAX, usize::MAX, &mut scanner, temp_dir);
+    let base = 1_700_000_000;
+    scanner.new_cache.info.pending_heals = (0..MAX_PENDING_SCANNER_HEALS_PER_BUCKET)
+        .map(|index| {
+            let mut entry =
+                pending_heal(PendingScannerHealKind::Object, "bucket", Some(&format!("object-{index}")), None, base, 1);
+            entry.first_seen = base;
+            entry.last_admission_result = "accepted".to_string();
+            entry
+        })
+        .collect();
+    let indices: HashMap<String, usize> = scanner
+        .new_cache
+        .info
+        .pending_heals
+        .iter()
+        .enumerate()
+        .map(|(index, entry)| (entry.object.clone().expect("object identity"), index))
+        .collect();
+    scanner.sync_pending_heals();
+    let initial_syncs = scanner.pending_heal_sync_count;
+    let mut requests = 0usize;
+    let mut nonempty_batches = 0;
+    for minute in 0..24 * 60 {
+        let now = base + minute * 60;
+        let candidates = pending_scanner_heal_retry_candidates_at(&scanner.new_cache.info.pending_heals, "bucket", now);
+        assert!(candidates.len() <= MAX_PENDING_SCANNER_HEAL_RETRIES_PER_BUCKET);
+        if candidates.is_empty() {
+            continue;
+        }
+        nonempty_batches += 1;
+        let before = scanner.pending_heal_sync_count;
+        {
+            let batch = PendingHealSyncBatch::new(&mut scanner);
+            for candidate in candidates {
+                assert_eq!(
+                    build_pending_scanner_heal_request(&candidate)
+                        .expect("retry request")
+                        .priority,
+                    HealChannelPriority::Low
+                );
+                let index = indices[candidate.object.as_ref().expect("object identity")];
+                record_pending_heal_attempt(&mut batch.scanner.new_cache.info.pending_heals[index], now);
+                observe_pending_heal_admission(
+                    &mut batch.scanner.new_cache.info.pending_heals[index],
+                    HealAdmissionResult::Accepted,
+                );
+                batch.scanner.sync_pending_heals();
+                requests += 1;
+            }
+        }
+        assert_eq!(scanner.pending_heal_sync_count, before + 1, "one table clone per changed retry batch");
+        assert_eq!(scanner.new_cache.info.pending_heals.len(), MAX_PENDING_SCANNER_HEALS_PER_BUCKET);
+    }
+    assert!(requests >= MAX_PENDING_SCANNER_HEALS_PER_BUCKET, "every retained hint receives a retry");
+    assert!(
+        requests <= 7 * MAX_PENDING_SCANNER_HEALS_PER_BUCKET,
+        "15min..6h backoff bounds repeated work within 24h"
+    );
+    assert!(scanner.new_cache.info.pending_heals.iter().all(|entry| entry.attempts >= 2));
+    assert_eq!(scanner.pending_heal_sync_count - initial_syncs, nonempty_batches);
+    assert_eq!(scanner.update_cache.info.pending_heals, scanner.new_cache.info.pending_heals);
+}
+
+#[tokio::test]
+async fn mrf_ownership_cancelled_batch_restores_sync_without_per_item_clones() {
+    let (mut scanner, temp_dir) = build_test_scanner().await;
+    let _guard = TestGuard::new(u64::MAX, usize::MAX, &mut scanner, temp_dir);
+    scanner
+        .new_cache
+        .info
+        .pending_heals
+        .push(pending_heal(PendingScannerHealKind::Object, "bucket", Some("object"), None, 1, 1));
+    let before = scanner.pending_heal_sync_count;
+    let mut work = Box::pin(async {
+        let batch = PendingHealSyncBatch::new(&mut scanner);
+        record_pending_heal_attempt(&mut batch.scanner.new_cache.info.pending_heals[0], 100);
+        batch.scanner.sync_pending_heals();
+        batch.scanner.sync_pending_heals();
+        std::future::pending::<()>().await;
+    });
+    assert!(futures::poll!(&mut work).is_pending());
+    drop(work);
+    assert!(!scanner.pending_heal_sync_deferred);
+    assert!(!scanner.pending_heal_batch_dirty);
+    assert_eq!(scanner.pending_heal_sync_count, before + 1);
+    assert_eq!(scanner.new_cache.info.pending_heals[0].attempts, 2);
+    assert!(pending_scanner_heal_retry_candidates_at(&scanner.new_cache.info.pending_heals, "bucket", 101).is_empty());
+    assert_eq!(scanner.update_cache.info.pending_heals, scanner.new_cache.info.pending_heals);
+    let result: std::result::Result<(), &'static str> = async {
+        let batch = PendingHealSyncBatch::new(&mut scanner);
+        record_pending_heal_attempt(&mut batch.scanner.new_cache.info.pending_heals[0], 200);
+        observe_pending_heal_admission(&mut batch.scanner.new_cache.info.pending_heals[0], HealAdmissionResult::Merged);
+        batch.scanner.sync_pending_heals();
+        Err("injected retry batch failure")
+    }
+    .await;
+    assert!(result.is_err());
+    assert_eq!(scanner.pending_heal_sync_count, before + 2);
+    assert!(!scanner.pending_heal_sync_deferred);
+    assert_eq!(scanner.update_cache.info.pending_heals, scanner.new_cache.info.pending_heals);
+}
+
+#[derive(Default)]
+struct NoticeStorage {
+    calls: std::sync::Mutex<HashMap<String, u32>>,
+    retry_started: tokio::sync::Notify,
+}
+
+#[async_trait::async_trait]
+impl HealStorageAPI for NoticeStorage {
+    async fn get_object_meta(&self, _: &str, _: &str) -> rustfs_heal::Result<Option<HealObjectInfo>> {
+        Ok(None)
+    }
+    async fn ec_decode_rebuild(&self, _: &str, _: &str) -> rustfs_heal::Result<Vec<u8>> {
+        Err(rustfs_heal::Error::other("unused decode fixture"))
+    }
+    async fn get_bucket_info(&self, bucket: &str) -> rustfs_heal::Result<Option<BucketInfo>> {
+        Ok(Some(BucketInfo {
+            name: bucket.to_string(),
+            ..Default::default()
+        }))
+    }
+    async fn list_buckets(&self) -> rustfs_heal::Result<Vec<BucketInfo>> {
+        Ok(Vec::new())
+    }
+    async fn object_exists(&self, _: &str, _: &str) -> rustfs_heal::Result<bool> {
+        Ok(true)
+    }
+    async fn heal_object(
+        &self,
+        _: &str,
+        object: &str,
+        _: Option<&str>,
+        _: &HealOpts,
+    ) -> rustfs_heal::Result<(HealItem, Option<rustfs_heal::Error>)> {
+        let retry = {
+            let mut calls = self.calls.lock().expect("fixture calls");
+            let count = calls.entry(object.to_string()).or_default();
+            *count += 1;
+            *count > 1
+        };
+        if retry {
+            self.retry_started.notify_one();
+            std::future::pending::<()>().await;
+        }
+        match object {
+            "grace" => Ok((
+                HealItem::default(),
+                Some(rustfs_heal::Error::Disk(crate::DiskError::other(
+                    "dangling object deletion deferred by heal grace window; retry_after_secs=3599; grace_secs=3600",
+                ))),
+            )),
+            "failed" => Err(rustfs_heal::Error::other("permanent fixture failure")),
+            "cancelled" => Err(rustfs_heal::Error::TaskCancelled),
+            _ => Ok((HealItem::default(), None)),
+        }
+    }
+    async fn heal_bucket(&self, _: &str, _: &HealOpts) -> rustfs_heal::Result<HealItem> {
+        Ok(HealItem::default())
+    }
+    async fn heal_format(&self, _: bool) -> rustfs_heal::Result<(HealItem, Option<rustfs_heal::Error>)> {
+        Ok((HealItem::default(), None))
+    }
+    async fn list_objects_for_heal_page(
+        &self,
+        _: &str,
+        _: &str,
+        _: Option<&str>,
+        _: bool,
+    ) -> rustfs_heal::Result<(Vec<HealListItem>, Option<String>, bool)> {
+        Ok((Vec::new(), None, false))
+    }
+    async fn get_disk_for_resume(&self, _: &str) -> rustfs_heal::Result<crate::DiskStore> {
+        Err(rustfs_heal::Error::other("unused resume fixture"))
+    }
+}
+
+#[tokio::test]
+#[serial]
+async fn mrf_ownership_manager_completion_preserves_scanner_pending() {
+    const CHILD: &str = "RUSTFS_MRF_OWNERSHIP_TEST_CHILD";
+    if std::env::var_os(CHILD).is_none() {
+        let output = std::process::Command::new(std::env::current_exe().expect("test executable"))
+            .args([
+                "--exact",
+                "scanner_folder::tests::mrf_ownership::mrf_ownership_manager_completion_preserves_scanner_pending",
+                "--nocapture",
+            ])
+            .env(CHILD, "1")
+            .env("RUSTFS_HEAL_MRF_ENABLE", "true")
+            .output()
+            .expect("isolated ingress test process");
+        let stdout = String::from_utf8_lossy(&output.stdout);
+        assert!(
+            output.status.success() && stdout.contains("1 passed;"),
+            "{stdout}\n{}",
+            String::from_utf8_lossy(&output.stderr)
+        );
+        return;
+    }
+    // The production ingress channel is a process singleton; isolation keeps
+    // its receiver and lease generations independent from other scanner tests.
+    let (mut scanner, temp_dir) = build_test_scanner().await;
+    let _guard = TestGuard::new(u64::MAX, usize::MAX, &mut scanner, temp_dir);
+    let bucket = format!("mrf-ownership-{}", Uuid::new_v4());
+    scanner.new_cache.info.name = bucket.clone();
+    scanner.update_cache.info.name = bucket.clone();
+    scanner.heal_object_select = 1;
+    let storage = Arc::new(NoticeStorage::default());
+    let manager = Arc::new(HealManager::new(
+        storage.clone(),
+        Some(HealConfig {
+            enable_auto_heal: false,
+            mainline_throttle_enable: false,
+            heal_interval: Duration::from_millis(10),
+            ..Default::default()
+        }),
+    ));
+    manager.start().await.expect("production manager starts");
+    spawn_mrf_consumer(manager.clone());
+    for (index, object) in ["grace", "unknown", "failed", "cancelled"].iter().enumerate() {
+        let version = Uuid::new_v4();
+        scanner.new_cache.info.pending_heals.push(pending_heal(
+            PendingScannerHealKind::Object,
+            &bucket,
+            Some(object),
+            Some(&version.to_string()),
+            1,
+            1,
+        ));
+        let scope = Some(MrfScope {
+            pool_index: 0,
+            set_index: 0,
+        });
+        assert_eq!(
+            try_send_mrf_intent_typed(MrfKind::PartialWrite, &bucket, object, Some(version), scope),
+            MrfIngressResult::Enqueued
+        );
+        // Re-admission establishes that the first terminal callback released
+        // its ingress lease. Statistics alone precede notice publication.
+        tokio::time::timeout(Duration::from_secs(5), async {
+            loop {
+                match try_send_mrf_intent_typed(MrfKind::PartialWrite, &bucket, object, Some(version), scope) {
+                    MrfIngressResult::Enqueued => break,
+                    MrfIngressResult::Coalesced => tokio::task::yield_now().await,
+                    other => panic!("unexpected retry ingress result: {other:?}"),
+                }
+            }
+        })
+        .await
+        .expect("production terminal releases its ingress lease");
+        tokio::time::timeout(Duration::from_secs(5), storage.retry_started.notified())
+            .await
+            .expect("the real consumer starts the second generation");
+        assert!(
+            take_mrf_repaired_events_for(&bucket).is_empty(),
+            "{object}: task completion must not emit an unproved repair"
+        );
+        if *object == "unknown" {
+            assert_eq!(
+                manager.get_statistics().await.total_objects_healed,
+                1,
+                "legacy healed count is not repair proof"
+            );
+        }
+        assert_eq!(
+            try_send_mrf_intent_typed(MrfKind::PartialWrite, &bucket, object, Some(version), scope),
+            MrfIngressResult::Coalesced,
+            "the in-flight retry retains its new ingress lease"
+        );
+        note_mrf_repaired(&bucket, object, Some(*version.as_bytes()));
+        let syncs_before_retry = scanner.pending_heal_sync_count;
+        scanner
+            .retry_pending_scanner_heals()
+            .await
+            .expect("real scanner ledger retry");
+        assert_eq!(scanner.pending_heal_sync_count, syncs_before_retry + 1, "the real retry batch syncs once");
+        assert_eq!(
+            scanner.new_cache.info.pending_heals.len(),
+            index + 1,
+            "{object}: pending responsibility survives"
+        );
+        let restored = DataUsageCache::unmarshal(&scanner.new_cache.marshal_msg().expect("serialize pending cache"))
+            .expect("decode pending cache");
+        assert_eq!(restored.info.pending_heals.len(), index + 1);
+        assert_eq!(
+            manager
+                .cancel_tasks_for_path(&format!("{bucket}/{object}"))
+                .await
+                .expect("cancel blocked retry"),
+            1
+        );
+    }
+    manager.stop().await.expect("production manager stops");
+}
@@ -14,7 +14,7 @@

 use crate::data_usage_define::{DATA_USAGE_CACHE_KEY_FORMAT, DataUsageCacheRevisions};
 use crate::scanner_budget::ScannerCycleBudget;
-use crate::scanner_folder::{ScannerItem, scan_data_folder};
+use crate::scanner_folder::{ScannerBucketPrefixScanScope, ScannerItem, scan_data_folder_scoped};
 use crate::sleeper::SCANNER_SLEEPER;
 use crate::{
    DATA_USAGE_CACHE_NAME, DATA_USAGE_ROOT, DataUsageCache, DataUsageCacheInfo, DataUsageCachePrepareOutcome,
@@ -27,7 +27,7 @@ use metrics::counter;
 use rand::seq::SliceRandom as _;
 #[cfg(test)]
 use rustfs_config::{ENV_SCANNER_MAX_CONCURRENT_DISK_SCANS, ENV_SCANNER_MAX_CONCURRENT_SET_SCANS};
-use rustfs_data_usage::{BucketTargetUsageInfo, BucketUsageInfo};
+use rustfs_data_usage::{BucketTargetUsageInfo, BucketUsageInfo, observed_data_usage_is_newer};
 use rustfs_filemeta::FileMeta;
 use rustfs_heal_contracts::heal_channel::HealScanMode;
 use rustfs_lock::{LockError, NamespaceLockGuard};
@@ -103,6 +103,7 @@ pub type DirtyUsageBuckets = HashMap<String, u64>;
 #[derive(Clone, Debug)]
 struct DirtyUsageSnapshot {
    buckets: Arc<DirtyUsageBuckets>,
+    scopes: Arc<DirtyUsageBucketScopes>,
    generation: u64,
    covers_all_pending: bool,
 }
@@ -110,25 +111,41 @@ struct DirtyUsageSnapshot {
 #[derive(Clone, Debug, Default)]
 pub(crate) struct ScannerBucketScanScope {
    selected_buckets: Option<Arc<HashSet<String>>>,
+    selected_bucket_prefixes: Option<Arc<HashMap<String, ScannerBucketPrefixScanScope>>>,
    baseline_scan_plan_digest: Option<DataUsageScanPlanDigest>,
 }

 impl ScannerBucketScanScope {
-    fn is_default(&self) -> bool {
-        self.selected_buckets.is_none() && self.baseline_scan_plan_digest.is_none()
+    #[cfg(test)]
+    pub(crate) fn selected_buckets_for_tests(&self) -> Option<&HashSet<String>> {
+        self.selected_buckets.as_deref()
    }

-    fn from_dirty_buckets(selected_buckets: HashSet<String>, baseline_scan_plan_digest: DataUsageScanPlanDigest) -> Self {
+    fn is_default(&self) -> bool {
+        self.selected_buckets.is_none() && self.selected_bucket_prefixes.is_none() && self.baseline_scan_plan_digest.is_none()
+    }
+
+    fn from_dirty_buckets(
+        selected_buckets: HashSet<String>,
+        selected_bucket_prefixes: HashMap<String, ScannerBucketPrefixScanScope>,
+        baseline_scan_plan_digest: DataUsageScanPlanDigest,
+    ) -> Self {
        Self {
            selected_buckets: Some(Arc::new(selected_buckets)),
+            selected_bucket_prefixes: (!selected_bucket_prefixes.is_empty()).then(|| Arc::new(selected_bucket_prefixes)),
            baseline_scan_plan_digest: Some(baseline_scan_plan_digest),
        }
    }
+
+    pub(crate) fn prefix_scope_for(&self, bucket: &str) -> Option<ScannerBucketPrefixScanScope> {
+        self.selected_bucket_prefixes.as_ref()?.get(bucket).cloned()
+    }
 }

 #[derive(Clone, Copy)]
 pub(super) struct ScannerCacheBaselineProof<'a> {
-    pub(super) data: Option<&'a Bytes>,
+    pub(super) authoritative_data: Option<&'a Bytes>,
+    pub(super) observed_candidate_data: Option<&'a Bytes>,
    pub(super) expected_sources: &'a HashSet<DataUsageCacheSource>,
    pub(super) leader_epoch: u64,
    pub(super) want_cycle: u64,
@@ -142,19 +159,31 @@ struct ScannerPeerDirtyUsageExpectation {
    pending: bool,
 }

-fn verified_remote_dirty_usage_buckets(
+#[derive(Debug, PartialEq, Eq)]
+struct VerifiedRemoteDirtyUsage {
+    dirty_buckets: HashSet<String>,
+    acknowledgements: Vec<crate::scanner::ScannerDirtyUsageAcknowledgement>,
+}
+
+struct ScannerBucketScopeResolutionResult {
+    scope: ScannerBucketScanScope,
+    remote_dirty_usage_acknowledgements: Vec<crate::scanner::ScannerDirtyUsageAcknowledgement>,
+}
+
+fn verified_remote_dirty_usage(
    expected_peers: &HashMap<String, ScannerPeerDirtyUsageExpectation>,
    peer_snapshots: Vec<(String, EcstoreScannerPeerDirtyUsageSnapshot)>,
-) -> Option<HashSet<String>> {
+) -> Option<VerifiedRemoteDirtyUsage> {
    if expected_peers.is_empty() || peer_snapshots.len() != expected_peers.len() {
        return None;
    }

    let mut received_peers = HashSet::with_capacity(peer_snapshots.len());
    let mut dirty_buckets = HashSet::new();
+    let mut acknowledgements = Vec::new();
    for (host, snapshot) in peer_snapshots {
        let expected = expected_peers.get(&host)?;
-        if !received_peers.insert(host)
+        if !received_peers.insert(host.clone())
            || snapshot.instance_id != expected.instance_id
            || snapshot.generation != expected.generation
            || snapshot.generation == u64::MAX
@@ -165,27 +194,118 @@ fn verified_remote_dirty_usage_buckets(
        {
            return None;
        }
-        dirty_buckets.extend(snapshot.buckets.into_keys());
+        let entries = snapshot
+            .buckets
+            .iter()
+            .map(|(bucket, state)| crate::storage_api::EcstoreScannerScopedDirtyUsageAckEntry {
+                bucket: bucket.clone(),
+                bucket_incarnation: state.bucket_incarnation,
+                generation: state.generation,
+            })
+            .collect::<Vec<_>>();
+        dirty_buckets.extend(snapshot.buckets.keys().cloned());
+        if !entries.is_empty() {
+            acknowledgements.push(crate::scanner::ScannerDirtyUsageAcknowledgement {
+                host,
+                instance_id: snapshot.instance_id,
+                kind: crate::scanner::ScannerDirtyUsageAcknowledgementKind::Scoped {
+                    owner_id: snapshot.owner_id,
+                    entries,
+                },
+            });
+        }
    }

-    (received_peers.len() == expected_peers.len()).then_some(dirty_buckets)
+    (received_peers.len() == expected_peers.len()).then_some(VerifiedRemoteDirtyUsage {
+        dirty_buckets,
+        acknowledgements,
+    })
 }

-fn complete_scanner_cache_baseline_plan_digest(proof: ScannerCacheBaselineProof<'_>) -> Option<DataUsageScanPlanDigest> {
-    let data = proof.data?;
-    let baseline = serde_json::from_slice::<DataUsageInfo>(data).ok()?;
-    if !baseline.is_complete_bucket_usage_snapshot()
-        || baseline.usage_snapshot_partial
-        || baseline.usage_snapshot_converged != Some(true)
-        || baseline.scanner_epoch != Some(proof.leader_epoch)
-        || baseline.usage_snapshot_set_states.len() != proof.expected_sources.len()
+fn scanner_scoped_dirty_usage_ack_exceeds_cost_threshold(
+    acknowledgements: &[crate::scanner::ScannerDirtyUsageAcknowledgement],
+) -> bool {
+    acknowledgements.iter().any(|acknowledgement| {
+        matches!(
+            &acknowledgement.kind,
+            crate::scanner::ScannerDirtyUsageAcknowledgementKind::Scoped { entries, .. }
+                if entries.len() > crate::SCANNER_SCOPED_DIRTY_USAGE_ACK_MAX_ENTRIES
+        )
+    })
+}
+
+fn resolve_remote_dirty_usage_scope(
+    requested_scope: ScannerBucketScanScope,
+    mut dirty_buckets: HashSet<String>,
+    remote_dirty_usage: VerifiedRemoteDirtyUsage,
+    all_buckets: &[BucketInfo],
+    baseline_proof: ScannerCacheBaselineProof<'_>,
+) -> ScannerBucketScopeResolutionResult {
+    let default_result = |scope: ScannerBucketScanScope| ScannerBucketScopeResolutionResult {
+        scope,
+        remote_dirty_usage_acknowledgements: Vec::new(),
+    };
+
+    dirty_buckets.extend(remote_dirty_usage.dirty_buckets);
+    // Peer snapshots contribute bucket names only; the local prefix scopes
+    // would narrow a bucket a peer dirtied elsewhere, so the merged scope
+    // stays at bucket granularity (same rule as the local fallthrough).
+    let scope = scoped_scan_scope_from_dirty_buckets(requested_scope, dirty_buckets, None, true, all_buckets, baseline_proof);
+    if scope.is_default() {
+        return default_result(scope);
+    }
+    let Some(selected_buckets) = scope.selected_buckets.as_ref() else {
+        return default_result(scope);
+    };
+    let mut scoped_acknowledgements = Vec::with_capacity(remote_dirty_usage.acknowledgements.len());
+    for acknowledgement in remote_dirty_usage.acknowledgements {
+        let crate::scanner::ScannerDirtyUsageAcknowledgement {
+            host,
+            instance_id,
+            kind: crate::scanner::ScannerDirtyUsageAcknowledgementKind::Scoped { owner_id, entries },
+        } = acknowledgement
+        else {
+            return default_result(scope);
+        };
+        let entries = entries
+            .into_iter()
+            .filter(|entry| selected_buckets.contains(&entry.bucket))
+            .collect::<Vec<_>>();
+        if !entries.is_empty() {
+            scoped_acknowledgements.push(crate::scanner::ScannerDirtyUsageAcknowledgement {
+                host,
+                instance_id,
+                kind: crate::scanner::ScannerDirtyUsageAcknowledgementKind::Scoped { owner_id, entries },
+            });
+        }
+    }
+    if scanner_scoped_dirty_usage_ack_exceeds_cost_threshold(&scoped_acknowledgements) {
+        return default_result(ScannerBucketScanScope::default());
+    }
+
+    ScannerBucketScopeResolutionResult {
+        scope,
+        remote_dirty_usage_acknowledgements: scoped_acknowledgements,
+    }
+}
+
+fn complete_scanner_cache_snapshot_plan_digest(
+    snapshot: &DataUsageInfo,
+    proof: ScannerCacheBaselineProof<'_>,
+    expected_converged: bool,
+) -> Option<DataUsageScanPlanDigest> {
+    if !snapshot.is_complete_bucket_usage_snapshot()
+        || snapshot.usage_snapshot_partial
+        || snapshot.usage_snapshot_converged != Some(expected_converged)
+        || snapshot.scanner_epoch != Some(proof.leader_epoch)
+        || snapshot.usage_snapshot_set_states.len() != proof.expected_sources.len()
    {
        return None;
    }

    // Completed maintenance also covers ordinary usage. Keep its exact stored
    // proof for cache reuse, and reject mixtures of different set work proofs.
-    let baseline_plan_digest = DataUsageScanPlanDigest(baseline.usage_snapshot_set_states.first()?.scan_plan_digest?);
+    let baseline_plan_digest = DataUsageScanPlanDigest(snapshot.usage_snapshot_set_states.first()?.scan_plan_digest?);
    if ![
        proof.scan_plan_digest,
        scanner_bucket_work_digest(proof.scan_plan_digest, HealScanMode::Normal, true),
@@ -195,8 +315,8 @@ fn complete_scanner_cache_baseline_plan_digest(proof: ScannerCacheBaselineProof<
    {
        return None;
    }
-    let mut states = HashSet::with_capacity(baseline.usage_snapshot_set_states.len());
-    for state in &baseline.usage_snapshot_set_states {
+    let mut states = HashSet::with_capacity(snapshot.usage_snapshot_set_states.len());
+    for state in &snapshot.usage_snapshot_set_states {
        let source = DataUsageCacheSource::new(usize::try_from(state.pool_index).ok()?, usize::try_from(state.set_index).ok()?);
        if !proof.expected_sources.contains(&source)
            || !states.insert(source)
@@ -213,9 +333,34 @@ fn complete_scanner_cache_baseline_plan_digest(proof: ScannerCacheBaselineProof<
    (states == *proof.expected_sources).then_some(baseline_plan_digest)
 }

+fn complete_scanner_cache_baseline_plan_digest(proof: ScannerCacheBaselineProof<'_>) -> Option<DataUsageScanPlanDigest> {
+    let authoritative = serde_json::from_slice::<DataUsageInfo>(proof.authoritative_data?).ok()?;
+    if let Some(validated_digest) = complete_scanner_cache_snapshot_plan_digest(&authoritative, proof, true) {
+        return Some(validated_digest);
+    }
+
+    // A complete but superseded observation may reuse its per-set cache only
+    // when it was explicitly tied to the durable authoritative baseline. It
+    // remains observational: this proof grants bucket-scope reuse only and
+    // never changes authoritative usage publication or dirty acknowledgement.
+    let authoritative_has_identity = (crate::scanner::data_usage_info_has_persisted_baseline_identity(&authoritative)
+        && authoritative.usage_snapshot_converged != Some(false))
+        || crate::scanner::data_usage_info_is_bootstrap_pending(&authoritative);
+    if !authoritative_has_identity {
+        return None;
+    }
+    let observed = serde_json::from_slice::<DataUsageInfo>(proof.observed_candidate_data?).ok()?;
+    if !observed_data_usage_is_newer(&observed, &authoritative) {
+        return None;
+    }
+
+    complete_scanner_cache_snapshot_plan_digest(&observed, proof, false)
+}
+
 fn scoped_scan_scope_from_dirty_buckets(
    requested_scope: ScannerBucketScanScope,
    dirty_buckets: HashSet<String>,
+    dirty_scopes: Option<&DirtyUsageBucketScopes>,
    dirty_snapshot_complete: bool,
    all_buckets: &[BucketInfo],
    baseline_proof: ScannerCacheBaselineProof<'_>,
@@ -237,7 +382,22 @@ fn scoped_scan_scope_from_dirty_buckets(
        return requested_scope;
    };

-    ScannerBucketScanScope::from_dirty_buckets(selected_buckets, baseline_scan_plan_digest)
+    let selected_bucket_prefixes = dirty_scopes
+        .into_iter()
+        .flat_map(|dirty_scopes| {
+            selected_buckets
+                .iter()
+                .filter_map(|bucket| dirty_scopes.get(bucket).map(|scope| (bucket.clone(), scope)))
+        })
+        .filter_map(|(bucket, scope)| match scope {
+            DirtyUsageBucketScope::WholeBucket => None,
+            DirtyUsageBucketScope::TopLevelEntries(entries) => {
+                ScannerBucketPrefixScanScope::from_dirty_top_level_entries(entries.clone()).map(|scope| (bucket, scope))
+            }
+        })
+        .collect();
+
+    ScannerBucketScanScope::from_dirty_buckets(selected_buckets, selected_bucket_prefixes, baseline_scan_plan_digest)
 }

 pub(crate) fn is_scanner_metadata_corrupt_error(err: &StorageError) -> bool {
@@ -286,6 +446,7 @@ pub struct ScannerBucketScanPlan {
    /// Includes mutation generations even when the set planner uses a structural digest.
    bucket_coverage_digest: DataUsageScanPlanDigest,
    requires_full_scan: bool,
+    service_cohort: Option<Arc<StdMutex<ScannerServiceCohort>>>,
    // Cache work must invalidate on namespace completion even when its scoped baseline remains reusable.
    execution_digest: DataUsageScanPlanDigest,
    leader_epoch: u64,
@@ -732,6 +893,12 @@ pub trait ScannerIOCache: Send + Sync + Debug + 'static {
    ) -> Result<()>;
 }

+#[derive(Debug)]
+pub struct ScannerDiskScanOptions {
+    pub scan_mode: HealScanMode,
+    pub prefix_scan_scope: Option<ScannerBucketPrefixScanScope>,
+}
+
 #[async_trait::async_trait]
 pub trait ScannerIODisk: Send + Sync + Debug + 'static {
    async fn nsscanner_disk(
@@ -741,7 +908,7 @@ pub trait ScannerIODisk: Send + Sync + Debug + 'static {
        set_disks: Vec<Arc<Disk>>,
        cache: DataUsageCache,
        updates: Option<mpsc::Sender<DataUsageEntry>>,
-        scan_mode: HealScanMode,
+        options: ScannerDiskScanOptions,
    ) -> Result<ScannerDiskScanOutcome>;

    async fn get_size(&self, item: ScannerItem) -> Result<SizeSummary>;
@@ -868,6 +1035,7 @@ pub(crate) struct ScannerCycleResult {
    failed_dirty_usage: bool,
    pending_maintenance_work: bool,
    required_cycle_floor: Option<u64>,
+    publication_expectation: Option<ScannerPublicationExpectation>,
 }

 impl ScannerCycleResult {
@@ -883,10 +1051,12 @@ impl ScannerCycleResult {
            failed_dirty_usage: false,
            pending_maintenance_work: false,
            required_cycle_floor: None,
+            publication_expectation: None,
        }
    }

    pub(crate) fn with_publication_epoch(mut self, publication_epoch: Option<u64>) -> Self {
+        self.publication_expectation = None;
        self.publication_epoch = publication_epoch;
        self
    }
@@ -896,6 +1066,7 @@ impl ScannerCycleResult {
    }

    fn with_activity_digest(mut self, activity_digest: [u8; 32]) -> Self {
+        self.publication_expectation = None;
        self.activity_digest = Some(activity_digest);
        self
    }
@@ -905,6 +1076,7 @@ impl ScannerCycleResult {
    }

    pub(crate) fn with_observational_snapshot_published(mut self, published: bool) -> Self {
+        self.publication_expectation = None;
        self.observational_snapshot_published = published;
        self
    }
@@ -914,16 +1086,19 @@ impl ScannerCycleResult {
    }

    fn with_failed_dirty_usage(mut self, failed_dirty_usage: bool) -> Self {
+        self.publication_expectation = None;
        self.failed_dirty_usage = failed_dirty_usage;
        self
    }

    fn with_pending_maintenance_work(mut self, pending_maintenance_work: bool) -> Self {
+        self.publication_expectation = None;
        self.pending_maintenance_work = pending_maintenance_work;
        self
    }

    fn with_required_cycle_floor(mut self, required_cycle_floor: Option<u64>) -> Self {
+        self.publication_expectation = None;
        self.required_cycle_floor = required_cycle_floor;
        self
    }
@@ -932,11 +1107,13 @@ impl ScannerCycleResult {
        mut self,
        acknowledgements: Vec<crate::scanner::ScannerDirtyUsageAcknowledgement>,
    ) -> Self {
+        self.publication_expectation = None;
        self.remote_dirty_usage_acknowledgements = acknowledgements;
        self
    }

    pub(crate) fn with_remote_publication_lease_targets(mut self, targets: Vec<(String, String, u64)>) -> Self {
+        self.publication_expectation = None;
        self.remote_publication_lease_targets = targets;
        self
    }
@@ -945,7 +1122,32 @@ impl ScannerCycleResult {
        &self.remote_publication_lease_targets
    }

-    pub(crate) fn acknowledge_durable_usage(self) -> Vec<crate::scanner::ScannerDirtyUsageAcknowledgement> {
+    pub(crate) fn publication_expectation(&self) -> Option<ScannerPublicationExpectation> {
+        self.publication_expectation.clone()
+    }
+
+    fn with_publication_expectation(mut self, expectation: Option<ScannerPublicationExpectation>) -> Self {
+        // Seal only after all coverage and acknowledgement inputs are final.
+        self.publication_expectation = expectation;
+        self
+    }
+
+    pub(crate) fn acknowledge_durable_usage(
+        self,
+        proof: &crate::scanner::RootPublicationProof,
+    ) -> Vec<crate::scanner::ScannerDirtyUsageAcknowledgement> {
+        if self.status != ScannerCycleStatus::Complete
+            || self
+                .publication_expectation
+                .as_ref()
+                .is_none_or(|expected| proof.verified_version_for(expected).is_none())
+        {
+            return Vec::new();
+        }
+        self.clear_verified_usage()
+    }
+
+    fn clear_verified_usage(self) -> Vec<crate::scanner::ScannerDirtyUsageAcknowledgement> {
        if let Some(snapshot) = self.dirty_usage_clear {
            clear_dirty_usage_buckets(&snapshot);
        }
@@ -973,6 +1175,7 @@ impl ScannerCycleResult {
 mod cache;
 mod dirty_usage;
 mod guards;
+pub(crate) use guards::ScannerServiceCohort;
 mod io_cache;
 mod io_cycle;
 #[cfg(test)]
@@ -984,6 +1187,7 @@ mod publish_gate_tests;
 #[cfg(test)]
 mod tests;

+pub(crate) use cache::ScannerPublicationExpectation;
 use cache::*;
 use dirty_usage::*;
 use guards::*;
@@ -995,8 +1199,8 @@ pub(crate) use cache::{
 pub use dirty_usage::{
    ScannerDirtyUsageAckError, ScannerDirtyUsageBucket, ScannerDirtyUsageSnapshot, ScannerDirtyUsageState,
    acknowledge_dirty_usage_generation, acknowledge_scoped_dirty_usage, clear_dirty_usage_bucket, record_dirty_usage_bucket,
-    record_scanner_maintenance_change, scanner_activity_epoch, scanner_dirty_usage_snapshot, scanner_dirty_usage_state,
-    scanner_maintenance_generation,
+    record_dirty_usage_object, record_scanner_maintenance_change, scanner_activity_epoch, scanner_dirty_usage_snapshot,
+    scanner_dirty_usage_state, scanner_maintenance_generation,
 };
 #[cfg(test)]
 pub(crate) use dirty_usage::{clear_dirty_usage_buckets_for_tests, dirty_usage_buckets_for_tests};
@@ -308,6 +308,86 @@ impl<'a> ValidatedScannerSnapshot<'a> {
    }
 }

+#[derive(Clone, Debug)]
+pub(crate) struct ScannerPublicationExpectation {
+    candidate: Arc<([u8; 32], DataUsageScanPlanDigest)>,
+}
+
+impl ScannerPublicationExpectation {
+    pub(crate) fn matches_encoded_candidate(&self, digest: &[u8; 32]) -> bool {
+        &self.candidate.0 == digest
+    }
+
+    pub(crate) fn same_candidate(&self, other: &Self) -> bool {
+        Arc::ptr_eq(&self.candidate, &other.candidate) && self.candidate.1 == other.candidate.1
+    }
+}
+
+pub(super) struct ValidatedUsageCandidate {
+    data: DataUsageInfo,
+    #[cfg(test)]
+    last_update: SystemTime,
+    coverage_digest: DataUsageScanPlanDigest,
+}
+
+pub(super) fn empty_namespace_usage_candidate(
+    all_buckets: &[BucketInfo],
+    sources: &HashSet<DataUsageCacheSource>,
+    buckets_by_source: &HashMap<DataUsageCacheSource, Vec<BucketInfo>>,
+    identity: ScannerSnapshotIdentity,
+) -> Option<ValidatedUsageCandidate> {
+    if !all_buckets.is_empty()
+        || sources.is_empty()
+        || sources.len() != buckets_by_source.len()
+        || sources
+            .iter()
+            .any(|source| buckets_by_source.get(source).is_none_or(|buckets| !buckets.is_empty()))
+    {
+        return None;
+    }
+    let last_update = SystemTime::now();
+    Some(ValidatedUsageCandidate {
+        data: DataUsageInfo {
+            last_update: Some(last_update),
+            scanner_cycle: Some(identity.cycle),
+            scanner_epoch: Some(identity.leader_epoch),
+            usage_snapshot_complete: true,
+            ..Default::default()
+        },
+        #[cfg(test)]
+        last_update,
+        coverage_digest: identity.coverage_digest,
+    })
+}
+
+impl ValidatedUsageCandidate {
+    pub(super) fn prepare(mut self, status: ScannerCycleStatus) -> (DataUsageInfo, Option<ScannerPublicationExpectation>) {
+        self.data.usage_snapshot_converged = Some(status == ScannerCycleStatus::Complete);
+        let expectation = if status == ScannerCycleStatus::Complete {
+            struct DigestWriter(Sha256);
+            impl std::io::Write for DigestWriter {
+                fn write(&mut self, bytes: &[u8]) -> std::io::Result<usize> {
+                    self.0.update(bytes);
+                    Ok(bytes.len())
+                }
+                fn flush(&mut self) -> std::io::Result<()> {
+                    Ok(())
+                }
+            }
+            let mut writer = DigestWriter(Sha256::new());
+            serde_json::to_writer(&mut writer, &self.data)
+                .ok()
+                .map(|()| ScannerPublicationExpectation {
+                    candidate: Arc::new((writer.0.finalize().into(), self.coverage_digest)),
+                })
+        } else {
+            None
+        };
+        (self.data, expectation)
+    }
+}
+
+#[cfg(test)]
 pub(super) fn completed_data_usage_info(
    results: &[DataUsageCache],
    scope: &ScannerSnapshotScope<'_>,
@@ -316,6 +396,18 @@ pub(super) fn completed_data_usage_info(
    budget_elapsed: bool,
    cancelled: bool,
 ) -> Option<(DataUsageInfo, SystemTime)> {
+    completed_usage_candidate(results, scope, tier_registry_names, bucket_plan_complete, budget_elapsed, cancelled)
+        .map(|candidate| (candidate.data, candidate.last_update))
+}
+
+pub(super) fn completed_usage_candidate(
+    results: &[DataUsageCache],
+    scope: &ScannerSnapshotScope<'_>,
+    tier_registry_names: &[String],
+    bucket_plan_complete: bool,
+    budget_elapsed: bool,
+    cancelled: bool,
+) -> Option<ValidatedUsageCandidate> {
    if !bucket_plan_complete {
        return None;
    }
@@ -393,7 +485,12 @@ pub(super) fn completed_data_usage_info(
        usage_snapshot_set_states,
        ..Default::default()
    };
-    Some((data_usage_info, merged_last_update))
+    Some(ValidatedUsageCandidate {
+        data: data_usage_info,
+        #[cfg(test)]
+        last_update: merged_last_update,
+        coverage_digest: scope.identity.coverage_digest,
+    })
 }

 fn tier_accounting_proof_is_publishable(
@@ -16,6 +16,12 @@ use super::*;

 pub(super) static DIRTY_USAGE_BUCKET_GENERATION: AtomicU64 = AtomicU64::new(0);
 pub(super) static DIRTY_USAGE_BUCKETS: LazyLock<StdMutex<DirtyUsageBuckets>> = LazyLock::new(|| StdMutex::new(HashMap::new()));
+// Lock order when both dirty maps are needed is `DIRTY_USAGE_BUCKETS` followed
+// by `DIRTY_USAGE_BUCKET_SCOPES`. Both are held only for synchronous map
+// updates, so no scanner task can observe a bucket generation without its
+// matching scope.
+pub(super) static DIRTY_USAGE_BUCKET_SCOPES: LazyLock<StdMutex<DirtyUsageBucketScopes>> =
+    LazyLock::new(|| StdMutex::new(HashMap::new()));
 pub(super) static DIRTY_USAGE_BUCKET_NOTIFY: LazyLock<Notify> = LazyLock::new(Notify::new);
 pub(super) static SCANNER_ACTIVITY_EPOCH: LazyLock<String> = LazyLock::new(|| format!("{:032x}", rand::random::<u128>()));
 pub(super) static SCANNER_MAINTENANCE_GENERATION: AtomicU64 = AtomicU64::new(0);
@@ -33,6 +39,21 @@ pub struct ScannerDirtyUsageBucket {
    pub generation: u64,
 }

+/// A non-durable optimization hint for a dirty bucket.
+///
+/// A whole-bucket marker always wins over narrow path hints. The scanner never
+/// publishes a prefix-only result as authoritative usage; this only controls
+/// whether a complete per-bucket cache can reuse known-clean direct children.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub(super) enum DirtyUsageBucketScope {
+    WholeBucket,
+    TopLevelEntries(HashSet<String>),
+}
+
+pub(super) type DirtyUsageBucketScopes = HashMap<String, DirtyUsageBucketScope>;
+
+const MAX_DIRTY_USAGE_TOP_LEVEL_ENTRIES_PER_BUCKET: usize = 128;
+
 /// A point-in-time view of the local dirty bucket generations.
 ///
 /// `complete == false` is an all-or-nothing overflow signal: `buckets` is
@@ -67,6 +88,7 @@ pub fn acknowledge_scoped_dirty_usage(
    // No await or storage operation occurs while the dirty map is locked.
    let (cleared, pending) = {
        let mut dirty = dirty_usage_buckets();
+        let mut dirty_scopes = dirty_usage_bucket_scopes();
        let checked = entries
            .iter()
            .map(|(guard, generation)| {
@@ -81,6 +103,7 @@ pub fn acknowledge_scoped_dirty_usage(
            scanner_activity_epoch(),
            DIRTY_USAGE_BUCKET_GENERATION.load(Ordering::Acquire),
            &mut dirty,
+            &mut dirty_scopes,
            &checked,
            probe_only,
        )?;
@@ -100,6 +123,7 @@ fn apply_scoped_dirty_usage_ack(
    current_instance: &str,
    current_generation: u64,
    dirty: &mut DirtyUsageBuckets,
+    dirty_scopes: &mut DirtyUsageBucketScopes,
    entries: &[(&str, u64)],
    probe_only: bool,
 ) -> std::result::Result<usize, ScannerDirtyUsageAckError> {
@@ -118,6 +142,7 @@ fn apply_scoped_dirty_usage_ack(
        for (bucket, generation) in entries {
            if dirty.get(*bucket) == Some(generation) {
                dirty.remove(*bucket);
+                dirty_scopes.remove(*bucket);
                cleared += 1;
            }
        }
@@ -132,30 +157,60 @@ mod scoped_dirty_usage_tests {
    #[test]
    fn scoped_dirty_usage_preserves_uncovered_newer_and_replayed_generations() {
        let mut dirty = HashMap::from([("hot".to_string(), 7), ("cold".to_string(), 8)]);
-        assert_eq!(apply_scoped_dirty_usage_ack("p", "p", 8, &mut dirty, &[("cold", 8)], true), Ok(0));
+        let mut scopes = HashMap::from([
+            ("hot".to_string(), DirtyUsageBucketScope::WholeBucket),
+            (
+                "cold".to_string(),
+                DirtyUsageBucketScope::TopLevelEntries(HashSet::from(["first".to_string()])),
+            ),
+        ]);
+        assert_eq!(
+            apply_scoped_dirty_usage_ack("p", "p", 8, &mut dirty, &mut scopes, &[("cold", 8)], true),
+            Ok(0)
+        );
        assert_eq!(dirty.len(), 2);
-        assert_eq!(apply_scoped_dirty_usage_ack("p", "p", 8, &mut dirty, &[("cold", 8)], false), Ok(1));
+        assert!(scopes.contains_key("cold"));
+        assert_eq!(
+            apply_scoped_dirty_usage_ack("p", "p", 8, &mut dirty, &mut scopes, &[("cold", 8)], false),
+            Ok(1)
+        );
        assert_eq!(dirty.get("hot"), Some(&7));
-        assert_eq!(apply_scoped_dirty_usage_ack("p", "p", 8, &mut dirty, &[("cold", 8)], false), Ok(0));
+        assert!(!scopes.contains_key("cold"));
+        assert_eq!(
+            apply_scoped_dirty_usage_ack("p", "p", 8, &mut dirty, &mut scopes, &[("cold", 8)], false),
+            Ok(0)
+        );
        dirty.insert("cold".to_string(), 9);
-        assert_eq!(apply_scoped_dirty_usage_ack("p", "p", 9, &mut dirty, &[("cold", 8)], false), Ok(0));
+        scopes.insert("cold".to_string(), DirtyUsageBucketScope::WholeBucket);
+        assert_eq!(
+            apply_scoped_dirty_usage_ack("p", "p", 9, &mut dirty, &mut scopes, &[("cold", 8)], false),
+            Ok(0)
+        );
        assert_eq!(dirty.get("cold"), Some(&9));
+        assert!(scopes.contains_key("cold"));
    }

    #[test]
    fn scoped_dirty_usage_rejects_restart_and_invalid_batch_before_clearing() {
        let original = HashMap::from([("hot".to_string(), 7), ("cold".to_string(), 8)]);
        let mut dirty = original.clone();
+        let original_scopes = HashMap::from([
+            ("hot".to_string(), DirtyUsageBucketScope::WholeBucket),
+            ("cold".to_string(), DirtyUsageBucketScope::WholeBucket),
+        ]);
+        let mut scopes = original_scopes.clone();
        assert_eq!(
-            apply_scoped_dirty_usage_ack("old", "new", 8, &mut dirty, &[("cold", 8)], false),
+            apply_scoped_dirty_usage_ack("old", "new", 8, &mut dirty, &mut scopes, &[("cold", 8)], false),
            Err(ScannerDirtyUsageAckError::ProcessChanged)
        );
+        assert_eq!(scopes, original_scopes);
        for generation in [0, 9, u64::MAX] {
            assert_eq!(
-                apply_scoped_dirty_usage_ack("p", "p", 8, &mut dirty, &[("cold", 8), ("hot", generation)], false),
+                apply_scoped_dirty_usage_ack("p", "p", 8, &mut dirty, &mut scopes, &[("cold", 8), ("hot", generation)], false,),
                Err(ScannerDirtyUsageAckError::InvalidGeneration)
            );
            assert_eq!(dirty, original);
+            assert_eq!(scopes, original_scopes);
        }
    }
 }
@@ -164,6 +219,12 @@ pub(super) fn dirty_usage_buckets() -> MutexGuard<'static, DirtyUsageBuckets> {
    DIRTY_USAGE_BUCKETS.lock().unwrap_or_else(|poisoned| poisoned.into_inner())
 }

+fn dirty_usage_bucket_scopes() -> MutexGuard<'static, DirtyUsageBucketScopes> {
+    DIRTY_USAGE_BUCKET_SCOPES
+        .lock()
+        .unwrap_or_else(|poisoned| poisoned.into_inner())
+}
+
 pub(super) fn usize_to_u64_saturated(value: usize) -> u64 {
    u64::try_from(value).unwrap_or(u64::MAX)
 }
@@ -181,8 +242,10 @@ pub fn record_dirty_usage_bucket(bucket: &str) {

    let pending_buckets = {
        let mut dirty_buckets = dirty_usage_buckets();
+        let mut dirty_scopes = dirty_usage_bucket_scopes();
        let generation = advance_generation(&DIRTY_USAGE_BUCKET_GENERATION);
        dirty_buckets.insert(bucket.to_string(), generation);
+        dirty_scopes.insert(bucket.to_string(), DirtyUsageBucketScope::WholeBucket);
        dirty_buckets.len()
    };
    global_metrics().record_scanner_dirty_usage_pending(usize_to_u64_saturated(pending_buckets));
@@ -193,6 +256,56 @@ pub fn record_dirty_usage_bucket(bucket: &str) {
    DIRTY_USAGE_BUCKET_NOTIFY.notify_one();
 }

+/// Record a mutation whose affected top-level namespace entry is known.
+///
+/// Object names that cannot be represented as one safe direct child retain the
+/// conservative whole-bucket marker. This journal is intentionally process
+/// local: after restart or any unverified distributed path the scanner falls
+/// back to its ordinary bucket scan.
+pub fn record_dirty_usage_object(bucket: &str, object: &str) {
+    let Some(top_level_entry) = dirty_usage_top_level_entry(object) else {
+        record_dirty_usage_bucket(bucket);
+        return;
+    };
+    if bucket.is_empty() {
+        return;
+    }
+
+    let pending_buckets = {
+        let mut dirty_buckets = dirty_usage_buckets();
+        let mut dirty_scopes = dirty_usage_bucket_scopes();
+        let generation = advance_generation(&DIRTY_USAGE_BUCKET_GENERATION);
+        dirty_buckets.insert(bucket.to_string(), generation);
+        let scope = dirty_scopes
+            .entry(bucket.to_string())
+            .or_insert_with(|| DirtyUsageBucketScope::TopLevelEntries(HashSet::new()));
+        let overflowed = match scope {
+            DirtyUsageBucketScope::WholeBucket => false,
+            DirtyUsageBucketScope::TopLevelEntries(entries) => {
+                entries.insert(top_level_entry);
+                entries.len() > MAX_DIRTY_USAGE_TOP_LEVEL_ENTRIES_PER_BUCKET
+            }
+        };
+        if overflowed {
+            *scope = DirtyUsageBucketScope::WholeBucket;
+        }
+        dirty_buckets.len()
+    };
+    global_metrics().record_scanner_dirty_usage_pending(usize_to_u64_saturated(pending_buckets));
+    crate::prefix_usage::invalidate_prefix_usage_cache(bucket);
+    DIRTY_USAGE_BUCKET_NOTIFY.notify_one();
+}
+
+fn dirty_usage_top_level_entry(object: &str) -> Option<String> {
+    let (top_level_entry, _) = object.split_once('/').unwrap_or((object, ""));
+    (!top_level_entry.is_empty()
+        && top_level_entry != "."
+        && top_level_entry != ".."
+        && !object.starts_with('/')
+        && !top_level_entry.contains(['\\', '\0']))
+    .then(|| top_level_entry.to_string())
+}
+
 pub fn record_scanner_maintenance_change(bucket: &str) {
    if bucket.is_empty() {
        return;
@@ -262,6 +375,7 @@ pub fn acknowledge_dirty_usage_generation(

    let (cleared_buckets, pending_buckets) = {
        let mut dirty_buckets = dirty_usage_buckets();
+        let mut dirty_scopes = dirty_usage_bucket_scopes();
        let current_generation = DIRTY_USAGE_BUCKET_GENERATION.load(Ordering::Acquire);
        if generation == 0 || generation == u64::MAX || current_generation == u64::MAX || generation > current_generation {
            return Err(ScannerDirtyUsageAckError::InvalidGeneration);
@@ -269,6 +383,7 @@ pub fn acknowledge_dirty_usage_generation(

        let before = dirty_buckets.len();
        dirty_buckets.retain(|_, dirty_generation| *dirty_generation > generation);
+        dirty_scopes.retain(|bucket, _| dirty_buckets.contains_key(bucket));
        let cleared_buckets = before.saturating_sub(dirty_buckets.len());
        if cleared_buckets > 0 {
            advance_generation(&DIRTY_USAGE_BUCKET_GENERATION);
@@ -291,7 +406,9 @@ pub fn clear_dirty_usage_bucket(bucket: &str) {

    let pending_buckets = {
        let mut dirty_buckets = dirty_usage_buckets();
+        let mut dirty_scopes = dirty_usage_bucket_scopes();
        dirty_buckets.remove(bucket);
+        dirty_scopes.remove(bucket);
        advance_generation(&DIRTY_USAGE_BUCKET_GENERATION);
        dirty_buckets.len()
    };
@@ -299,8 +416,9 @@ pub fn clear_dirty_usage_bucket(bucket: &str) {
 }

 pub(super) fn snapshot_dirty_usage_buckets(buckets: &[BucketInfo], absent_generation_cutoff: u64) -> DirtyUsageSnapshot {
-    let (snapshot, generation, covers_all_pending) = {
+    let (snapshot, scopes, generation, covers_all_pending) = {
        let dirty_buckets = dirty_usage_buckets();
+        let dirty_scopes = dirty_usage_bucket_scopes();
        let listed_buckets = dirty_buckets
            .values()
            .any(|generation| *generation > absent_generation_cutoff)
@@ -315,13 +433,26 @@ pub(super) fn snapshot_dirty_usage_buckets(buckets: &[BucketInfo], absent_genera
            })
            .map(|(bucket, generation)| (bucket.clone(), *generation))
            .collect::<DirtyUsageBuckets>();
+        let scopes = snapshot
+            .keys()
+            .map(|bucket| {
+                (
+                    bucket.clone(),
+                    dirty_scopes
+                        .get(bucket)
+                        .cloned()
+                        .unwrap_or(DirtyUsageBucketScope::WholeBucket),
+                )
+            })
+            .collect::<DirtyUsageBucketScopes>();
        let generation = DIRTY_USAGE_BUCKET_GENERATION.load(Ordering::Acquire);
        let covers_all_pending = generation == absent_generation_cutoff && snapshot.len() == dirty_buckets.len();
-        (snapshot, generation, covers_all_pending)
+        (snapshot, scopes, generation, covers_all_pending)
    };
    global_metrics().record_scanner_dirty_usage_cycle_snapshot(usize_to_u64_saturated(snapshot.len()));
    DirtyUsageSnapshot {
        buckets: Arc::new(snapshot),
+        scopes: Arc::new(scopes),
        generation,
        covers_all_pending,
    }
@@ -338,10 +469,12 @@ pub(crate) async fn dirty_usage_bucket_notified() {
 pub(super) fn clear_dirty_usage_buckets(snapshot: &DirtyUsageBuckets) {
    let (cleared_buckets, pending_buckets) = {
        let mut dirty_buckets = dirty_usage_buckets();
+        let mut dirty_scopes = dirty_usage_bucket_scopes();
        let mut cleared_buckets = 0usize;
        for (bucket, generation) in snapshot {
            if dirty_buckets.get(bucket).is_some_and(|current| current == generation) {
                dirty_buckets.remove(bucket);
+                dirty_scopes.remove(bucket);
                cleared_buckets += 1;
            }
        }
@@ -443,9 +576,15 @@ pub(super) fn dirty_usage_bucket_count() -> usize {
 #[cfg(test)]
 pub(crate) fn clear_dirty_usage_buckets_for_tests() {
    dirty_usage_buckets().clear();
+    dirty_usage_bucket_scopes().clear();
 }

 #[cfg(test)]
 pub(crate) fn dirty_usage_buckets_for_tests() -> DirtyUsageBuckets {
    dirty_usage_buckets().clone()
 }
+
+#[cfg(test)]
+pub(crate) fn dirty_usage_bucket_scopes_for_tests() -> DirtyUsageBucketScopes {
+    dirty_usage_bucket_scopes().clone()
+}
@@ -14,6 +14,312 @@
 /// scan concurrency accounting: gauge recorders, RAII guards, and worker limits.
 use super::*;

+const SCANNER_SERVICE_COHORT_MAX_MEMBERS: usize = 4096;
+const SCANNER_SERVICE_COHORT_MAX_NAME_BYTES: usize = 128 * 1024;
+
+static SERVICE_COHORT_METRICS_OWNER: StdMutex<std::sync::Weak<()>> = StdMutex::new(std::sync::Weak::new());
+
+struct ScannerCohortMetricsOwner(Arc<()>);
+
+impl Default for ScannerCohortMetricsOwner {
+    fn default() -> Self {
+        let owner = Arc::new(());
+        let mut current = SERVICE_COHORT_METRICS_OWNER
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner());
+        *current = Arc::downgrade(&owner);
+        write_service_cohort_metrics(0, 0.0, false);
+        Self(owner)
+    }
+}
+
+impl Drop for ScannerCohortMetricsOwner {
+    fn drop(&mut self) {
+        let mut current = SERVICE_COHORT_METRICS_OWNER
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner());
+        if current.ptr_eq(&Arc::downgrade(&self.0)) {
+            *current = std::sync::Weak::new();
+            write_service_cohort_metrics(0, 0.0, false);
+        }
+    }
+}
+
+fn write_service_cohort_metrics(waiting: usize, oldest: f64, overflowed: bool) {
+    metrics::gauge!("rustfs_scanner_service_cohort_waiting").set(waiting as f64);
+    metrics::gauge!("rustfs_scanner_service_cohort_oldest_wait_seconds").set(oldest);
+    metrics::gauge!("rustfs_scanner_service_cohort_capacity_fallback").set(if overflowed { 1.0 } else { 0.0 });
+}
+
+struct ScannerCohortWait {
+    order: u64,
+    queued_at: Instant,
+    admitted: bool,
+    present: bool,
+}
+
+/// Leader-local admission order, never evidence of completed scan coverage.
+/// Retains at most 4096 members and 128 KiB of name payload, including the
+/// cursor shared with its last member. Candidate selection borrows at most
+/// 4096 inventory entries; the existing full inventory is not bounded here.
+pub(crate) struct ScannerServiceCohort {
+    members: HashMap<DataUsageCacheSource, HashMap<Arc<str>, ScannerCohortWait>>,
+    cursor: Option<(DataUsageCacheSource, Arc<str>)>,
+    next_order: u64,
+    max_members: usize,
+    max_name_bytes: usize,
+    overflowed: bool,
+    metrics_owner: ScannerCohortMetricsOwner,
+    waiting: usize,
+    oldest_wait_at_refresh: f64,
+    #[cfg(test)]
+    metric_members_examined: usize,
+}
+
+impl Default for ScannerServiceCohort {
+    fn default() -> Self {
+        Self {
+            members: HashMap::new(),
+            cursor: None,
+            next_order: 0,
+            max_members: SCANNER_SERVICE_COHORT_MAX_MEMBERS,
+            max_name_bytes: SCANNER_SERVICE_COHORT_MAX_NAME_BYTES,
+            overflowed: false,
+            metrics_owner: ScannerCohortMetricsOwner::default(),
+            waiting: 0,
+            oldest_wait_at_refresh: 0.0,
+            #[cfg(test)]
+            metric_members_examined: 0,
+        }
+    }
+}
+
+impl ScannerServiceCohort {
+    pub(crate) fn refresh(&mut self, inventory: &HashMap<DataUsageCacheSource, Vec<BucketInfo>>) {
+        let count = inventory
+            .values()
+            .fold(0usize, |count, buckets| count.saturating_add(buckets.len()));
+        let name_bytes = inventory
+            .values()
+            .flatten()
+            .fold(0usize, |bytes, bucket| bytes.saturating_add(bucket.name.len()));
+        self.overflowed = count > self.max_members || name_bytes > self.max_name_bytes;
+        for wait in self.members.values_mut().flat_map(HashMap::values_mut) {
+            wait.present = false;
+        }
+        for (source, buckets) in inventory {
+            for bucket in buckets {
+                if let Some(wait) = self
+                    .members
+                    .get_mut(source)
+                    .and_then(|members| members.get_mut(bucket.name.as_str()))
+                {
+                    wait.present = true;
+                }
+            }
+        }
+        self.members.retain(|_, buckets| {
+            buckets.retain(|_, wait| wait.present);
+            !buckets.is_empty()
+        });
+        if self.members.values().flat_map(HashMap::values).all(|wait| wait.admitted) {
+            self.members.clear();
+            self.next_order = 0;
+        }
+        if count == 0 {
+            self.cursor = None;
+        }
+        let mut member_count = self.members.values().map(HashMap::len).sum::<usize>();
+        let mut retained_bytes = self
+            .members
+            .values()
+            .flat_map(HashMap::keys)
+            .map(|name| name.len())
+            .sum::<usize>();
+        let mut incoming = self.admission_candidates(inventory, true);
+        if incoming.is_empty() {
+            incoming = self.admission_candidates(inventory, false);
+        }
+        for (pool, set, bucket) in incoming {
+            if member_count >= self.max_members {
+                break;
+            }
+            if retained_bytes.saturating_add(bucket.len()) > self.max_name_bytes {
+                continue;
+            }
+            let Some(next_order) = self.next_order.checked_add(1) else {
+                self.overflowed = true;
+                break;
+            };
+            let source = DataUsageCacheSource::new(pool, set);
+            let members = self.members.entry(source).or_default();
+            if members.contains_key(bucket) {
+                continue;
+            }
+            let name: Arc<str> = bucket.into();
+            retained_bytes += name.len();
+            member_count += 1;
+            members.insert(
+                name.clone(),
+                ScannerCohortWait {
+                    order: self.next_order,
+                    queued_at: Instant::now(),
+                    admitted: false,
+                    present: true,
+                },
+            );
+            self.cursor = Some((source, name));
+            self.next_order = next_order;
+        }
+        self.refresh_metrics();
+    }
+
+    fn admission_candidates<'a>(
+        &self,
+        inventory: &'a HashMap<DataUsageCacheSource, Vec<BucketInfo>>,
+        after_cursor: bool,
+    ) -> Vec<(usize, usize, &'a str)> {
+        let mut candidates = std::collections::BinaryHeap::new();
+        for (source, buckets) in inventory {
+            for bucket in buckets {
+                let key = (source.pool_index, source.set_index, bucket.name.as_str());
+                let after = self
+                    .cursor
+                    .as_ref()
+                    .is_none_or(|(source, name)| key > (source.pool_index, source.set_index, name.as_ref()));
+                if after != after_cursor
+                    || bucket.name.len() > self.max_name_bytes
+                    || self
+                        .members
+                        .get(source)
+                        .is_some_and(|members| members.contains_key(bucket.name.as_str()))
+                {
+                    continue;
+                }
+                if candidates.len() < self.max_members {
+                    candidates.push(key);
+                } else if candidates.peek().is_some_and(|last| key < *last) {
+                    candidates.pop();
+                    candidates.push(key);
+                }
+            }
+        }
+        candidates.into_sorted_vec()
+    }
+
+    pub(crate) fn order_set_indices(&self, sets: &[Arc<SetDisks>]) -> Vec<usize> {
+        let ranks = self
+            .members
+            .iter()
+            .map(|(source, buckets)| {
+                (
+                    *source,
+                    buckets
+                        .values()
+                        .filter(|wait| !wait.admitted)
+                        .map(|wait| wait.order)
+                        .min()
+                        .unwrap_or(u64::MAX),
+                )
+            })
+            .collect::<HashMap<_, _>>();
+        let mut indices = (0..sets.len()).collect::<Vec<_>>();
+        indices.sort_by_key(|index| {
+            (
+                ranks
+                    .get(&DataUsageCacheSource::new(sets[*index].pool_index, sets[*index].set_index))
+                    .copied()
+                    .unwrap_or(u64::MAX),
+                *index,
+            )
+        });
+        indices
+    }
+
+    pub(crate) fn order_buckets(&self, source: DataUsageCacheSource, buckets: &mut [BucketInfo]) {
+        let rank = |bucket: &str| {
+            self.members
+                .get(&source)
+                .and_then(|members| members.get(bucket))
+                .filter(|wait| !wait.admitted)
+                .map_or(u64::MAX, |wait| wait.order)
+        };
+        // Stable sorting preserves the existing dispatch order in the tail.
+        buckets.sort_by_key(|bucket| rank(&bucket.name));
+    }
+
+    pub(crate) fn record_admitted(&mut self, source: DataUsageCacheSource, bucket: &str) {
+        let Some(wait) = self.members.get_mut(&source).and_then(|members| members.get_mut(bucket)) else {
+            return;
+        };
+        if wait.admitted {
+            return;
+        }
+        wait.admitted = true;
+        self.waiting -= 1;
+        if self.waiting == 0 {
+            self.oldest_wait_at_refresh = 0.0;
+        }
+        self.record_metrics();
+    }
+
+    #[cfg(test)]
+    pub(super) fn admitted_members(&self) -> Vec<(DataUsageCacheSource, String)> {
+        self.members
+            .iter()
+            .flat_map(|(source, buckets)| {
+                buckets
+                    .iter()
+                    .filter(|(_, wait)| wait.admitted)
+                    .map(|(bucket, _)| (*source, bucket.to_string()))
+            })
+            .collect()
+    }
+
+    fn refresh_metrics(&mut self) {
+        // Oldest age is an inventory-refresh snapshot, not a per-admission
+        // scan of the cohort. Clear it immediately when no waiters remain.
+        let (mut waiting, mut oldest) = (0usize, 0.0f64);
+        for wait in self.members.values().flat_map(HashMap::values) {
+            #[cfg(test)]
+            {
+                self.metric_members_examined += 1;
+            }
+            if !wait.admitted {
+                waiting += 1;
+                oldest = oldest.max(wait.queued_at.elapsed().as_secs_f64());
+            }
+        }
+        self.waiting = waiting;
+        self.oldest_wait_at_refresh = oldest;
+        self.record_metrics();
+    }
+
+    fn record_metrics(&self) {
+        // Serialize owner replacement, publication and retirement. A retired
+        // scanner must neither publish nor clear a replacement's gauges.
+        let current = SERVICE_COHORT_METRICS_OWNER
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner());
+        if current.ptr_eq(&Arc::downgrade(&self.metrics_owner.0)) {
+            write_service_cohort_metrics(self.waiting, self.oldest_wait_at_refresh, self.overflowed);
+        }
+    }
+}
+
+pub(super) async fn wait_for_bucket_scan_permit(
+    semaphore: &Arc<Semaphore>,
+    ctx: &CancellationToken,
+    complete: &CancellationToken,
+) -> Option<tokio::sync::OwnedSemaphorePermit> {
+    tokio::select! {
+        biased;
+        _ = complete.cancelled() => None,
+        _ = ctx.cancelled() => None,
+        permit = semaphore.clone().acquire_owned() => permit.ok(),
+    }
+}
+
 pub(super) fn bucket_usage_scan_order(
    buckets: &[BucketInfo],
    old_cache: &DataUsageCache,
@@ -288,6 +594,340 @@ mod tests {
    use rustfs_scanner_metrics::metrics::{ScannerWorkSource, global_metrics};
    use tokio::sync::oneshot;

+    #[derive(Default)]
+    struct RecordedGauge(AtomicU64);
+
+    impl metrics::GaugeFn for RecordedGauge {
+        fn increment(&self, value: f64) {
+            self.set(f64::from_bits(self.0.load(Ordering::Relaxed)) + value);
+        }
+        fn decrement(&self, value: f64) {
+            self.increment(-value);
+        }
+        fn set(&self, value: f64) {
+            self.0.store(value.to_bits(), Ordering::Relaxed);
+        }
+    }
+
+    #[derive(Default)]
+    struct CohortGaugeRecorder(StdMutex<HashMap<String, Arc<RecordedGauge>>>);
+
+    impl metrics::Recorder for CohortGaugeRecorder {
+        fn describe_counter(&self, _: metrics::KeyName, _: Option<metrics::Unit>, _: metrics::SharedString) {}
+        fn describe_gauge(&self, _: metrics::KeyName, _: Option<metrics::Unit>, _: metrics::SharedString) {}
+        fn describe_histogram(&self, _: metrics::KeyName, _: Option<metrics::Unit>, _: metrics::SharedString) {}
+        fn register_counter(&self, _: &metrics::Key, _: &metrics::Metadata<'_>) -> metrics::Counter {
+            metrics::Counter::noop()
+        }
+        fn register_histogram(&self, _: &metrics::Key, _: &metrics::Metadata<'_>) -> metrics::Histogram {
+            metrics::Histogram::noop()
+        }
+        fn register_gauge(&self, key: &metrics::Key, _: &metrics::Metadata<'_>) -> metrics::Gauge {
+            metrics::Gauge::from_arc(
+                self.0
+                    .lock()
+                    .expect("gauge recorder")
+                    .entry(key.name().to_string())
+                    .or_default()
+                    .clone(),
+            )
+        }
+    }
+
+    impl CohortGaugeRecorder {
+        fn value(&self, name: &str) -> f64 {
+            f64::from_bits(self.0.lock().expect("gauge recorder")[name].0.load(Ordering::Relaxed))
+        }
+    }
+
+    #[test]
+    #[serial_test::serial]
+    fn service_cohort_metrics_retire_only_the_current_owner() {
+        let recorder = CohortGaugeRecorder::default();
+        metrics::with_local_recorder(&recorder, || {
+            let mut old = ScannerServiceCohort::default();
+            old.refresh(&cohort_inventory(&["old"]));
+            let mut current = ScannerServiceCohort {
+                max_members: 1,
+                ..Default::default()
+            };
+            current.refresh(&cohort_inventory(&["a", "b"]));
+            for wait in current.members.values_mut().flat_map(HashMap::values_mut) {
+                wait.queued_at = Instant::now() - Duration::from_secs(60);
+            }
+            current.refresh_metrics();
+            old.refresh(&cohort_inventory(&["old", "more"]));
+            drop(old);
+            assert_eq!(recorder.value("rustfs_scanner_service_cohort_waiting"), 1.0);
+            assert_eq!(recorder.value("rustfs_scanner_service_cohort_capacity_fallback"), 1.0);
+            assert!(recorder.value("rustfs_scanner_service_cohort_oldest_wait_seconds") >= 60.0);
+            drop(current);
+            for metric in [
+                "rustfs_scanner_service_cohort_waiting",
+                "rustfs_scanner_service_cohort_oldest_wait_seconds",
+                "rustfs_scanner_service_cohort_capacity_fallback",
+            ] {
+                assert_eq!(recorder.value(metric), 0.0, "owner retirement must clear {metric}");
+            }
+        });
+    }
+
+    #[test]
+    #[serial_test::serial]
+    fn service_cohort_admission_metrics_do_not_rescan_a_full_window() {
+        let recorder = CohortGaugeRecorder::default();
+        metrics::with_local_recorder(&recorder, || {
+            let source = DataUsageCacheSource::new(0, 0);
+            let names = (0..SCANNER_SERVICE_COHORT_MAX_MEMBERS)
+                .map(|index| format!("bucket-{index:04}"))
+                .collect::<Vec<_>>();
+            let inventory = cohort_inventory(&names.iter().map(String::as_str).collect::<Vec<_>>());
+            let mut cohort = ScannerServiceCohort::default();
+            cohort.refresh(&inventory);
+            assert_eq!(cohort.metric_members_examined, SCANNER_SERVICE_COHORT_MAX_MEMBERS);
+            assert_eq!(cohort.waiting, SCANNER_SERVICE_COHORT_MAX_MEMBERS);
+            for (index, name) in names.iter().enumerate() {
+                cohort.record_admitted(source, name);
+                for _ in 0..10 {
+                    cohort.record_admitted(source, name);
+                    cohort.record_admitted(source, "untracked-overflow-name");
+                    cohort.record_admitted(DataUsageCacheSource::new(99, 0), name);
+                }
+                assert_eq!(cohort.waiting, SCANNER_SERVICE_COHORT_MAX_MEMBERS - index - 1);
+                assert_eq!(
+                    cohort.metric_members_examined, SCANNER_SERVICE_COHORT_MAX_MEMBERS,
+                    "tracked, repeated and overflow admissions must not scan cohort members"
+                );
+            }
+            assert_eq!(recorder.value("rustfs_scanner_service_cohort_waiting"), 0.0);
+            assert_eq!(recorder.value("rustfs_scanner_service_cohort_oldest_wait_seconds"), 0.0);
+            cohort.refresh(&inventory);
+            assert_eq!(
+                cohort.metric_members_examined,
+                2 * SCANNER_SERVICE_COHORT_MAX_MEMBERS,
+                "one inventory refresh performs one metrics traversal"
+            );
+        });
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn service_cohort_queued_permit_cancel_and_drop_return_the_same_capacity() {
+        let semaphore = Arc::new(Semaphore::new(1));
+        let active = Arc::new(AtomicUsize::new(0));
+        let mut cohort = ScannerServiceCohort::default();
+        cohort.refresh(&cohort_inventory(&["waiting"]));
+        let gauge_reset = DiskBucketScanGaugeReset::new("cohort-wait".to_string(), "0".to_string());
+        record_disk_bucket_scans_queued(1, "cohort-wait", "0");
+        record_disk_bucket_scans_active(0, "cohort-wait", "0");
+        for cancel in [true, false] {
+            let held = semaphore
+                .clone()
+                .acquire_owned()
+                .await
+                .expect("hold the sole permit as a barrier");
+            let ctx = CancellationToken::new();
+            let complete = CancellationToken::new();
+            let mut waiter = Box::pin(wait_for_bucket_scan_permit(&semaphore, &ctx, &complete));
+            assert!(
+                futures::poll!(&mut waiter).is_pending(),
+                "the production wait must actually enqueue behind the barrier"
+            );
+            assert_eq!(semaphore.available_permits(), 0);
+            if cancel {
+                ctx.cancel();
+                assert!(waiter.as_mut().await.is_none());
+            }
+            drop(waiter);
+            assert_eq!(semaphore.available_permits(), 0, "cancelling a waiter must not release the held permit");
+            assert!(cohort.admitted_members().is_empty());
+            drop(held);
+            assert_eq!(semaphore.available_permits(), 1, "no queued waiter may leak or steal released capacity");
+        }
+        let ctx = CancellationToken::new();
+        let complete = CancellationToken::new();
+        let permit = wait_for_bucket_scan_permit(&semaphore, &ctx, &complete)
+            .await
+            .expect("same semaphore remains usable");
+        let active_guard = DiskBucketScanActiveGuard::new(active.clone(), "cohort-wait".to_string(), "0".to_string());
+        assert_eq!(active.load(Ordering::Relaxed), 1);
+        drop(active_guard);
+        drop(permit);
+        drop(gauge_reset);
+        assert_eq!(active.load(Ordering::Relaxed), 0);
+        assert_eq!(semaphore.available_permits(), 1);
+        let state = global_metrics()
+            .scanner_runtime_details_report()
+            .disk_bucket_scan_states
+            .into_iter()
+            .find(|state| state.pool == "cohort-wait" && state.set == "0")
+            .expect("fixture gauges");
+        assert_eq!((state.queued, state.active), (0, 0));
+        assert!(cohort.admitted_members().is_empty(), "permit ownership alone does not admit a bucket");
+    }
+
+    fn cohort_inventory(names: &[&str]) -> HashMap<DataUsageCacheSource, Vec<BucketInfo>> {
+        HashMap::from([(
+            DataUsageCacheSource::new(0, 0),
+            names
+                .iter()
+                .map(|name| BucketInfo {
+                    name: (*name).to_string(),
+                    ..Default::default()
+                })
+                .collect(),
+        )])
+    }
+
+    #[test]
+    #[serial_test::serial]
+    fn service_cohort_visits_fixed_members_within_service_round_bound() {
+        let inventory = cohort_inventory(&["a", "b", "c", "d", "e"]);
+        let source = DataUsageCacheSource::new(0, 0);
+        let mut cohort = ScannerServiceCohort::default();
+        let mut admitted = HashSet::new();
+        for _ in 0..3 {
+            cohort.refresh(&inventory);
+            let mut buckets = inventory[&source].clone();
+            cohort.order_buckets(source, &mut buckets);
+            for bucket in buckets.iter().take(2) {
+                admitted.insert(bucket.name.clone());
+                cohort.record_admitted(source, &bucket.name);
+            }
+        }
+        assert_eq!(admitted.len(), 5, "ceil(5/2) service rounds must include every member");
+    }
+
+    #[test]
+    #[serial_test::serial]
+    fn service_cohort_keeps_waiting_bootstrap_ahead_of_new_work() {
+        let source = DataUsageCacheSource::new(0, 0);
+        let mut cohort = ScannerServiceCohort::default();
+        cohort.refresh(&cohort_inventory(&["a-hot", "z-bootstrap"]));
+        cohort.record_admitted(source, "a-hot");
+        let queued_at = cohort.members[&source]["z-bootstrap"].queued_at;
+        let inventory = cohort_inventory(&["a-hot", "aaa-new-bootstrap", "z-bootstrap"]);
+        for _ in 0..10 {
+            cohort.refresh(&inventory);
+            let mut buckets = inventory[&source].clone();
+            cohort.order_buckets(source, &mut buckets);
+            assert_eq!(buckets[0].name, "z-bootstrap");
+            assert_eq!(buckets[1].name, "aaa-new-bootstrap");
+            assert_eq!(cohort.members[&source]["z-bootstrap"].queued_at, queued_at);
+        }
+    }
+
+    #[test]
+    #[serial_test::serial]
+    fn service_cohort_overflow_preserves_waiters_and_rotates_finished_windows() {
+        let source = DataUsageCacheSource::new(0, 0);
+        let mut cohort = ScannerServiceCohort {
+            max_members: 2,
+            max_name_bytes: 4,
+            ..Default::default()
+        };
+        cohort.refresh(&cohort_inventory(&["aa", "bb"]));
+        cohort.record_admitted(source, "aa");
+        for names in [["aa", "bb", "c"], ["aa", "bb", "d"]] {
+            let inventory = cohort_inventory(&names);
+            cohort.refresh(&inventory);
+            assert!(cohort.overflowed);
+            assert_eq!(cohort.members[&source].len(), 2);
+            assert!(cohort.members[&source].contains_key("aa"));
+            let mut fallback = inventory[&source].clone();
+            fallback.reverse();
+            cohort.order_buckets(source, &mut fallback);
+            assert_eq!(fallback[0].name, "bb", "overflow must not discard a waiting member's priority");
+            assert_eq!(fallback.len(), 3, "unknown tail must remain dispatchable");
+        }
+        cohort.record_admitted(source, "bb");
+        let inventory = cohort_inventory(&["aa", "bb", "c", "d"]);
+        cohort.refresh(&inventory);
+        assert_eq!(
+            cohort.members[&source].keys().map(AsRef::as_ref).collect::<HashSet<&str>>(),
+            HashSet::from(["c", "d"])
+        );
+        cohort.record_admitted(source, "c");
+        cohort.record_admitted(source, "d");
+        cohort.refresh(&inventory);
+        assert!(
+            cohort.members[&source].contains_key("aa"),
+            "finite inventory must wrap after the last window"
+        );
+        cohort.refresh(&cohort_inventory(&["bb"]));
+        assert!(!cohort.overflowed);
+        assert_eq!(cohort.members[&source].len(), 1);
+        cohort.next_order = u64::MAX;
+        cohort.refresh(&cohort_inventory(&["bb", "c"]));
+        assert!(cohort.overflowed);
+        assert_eq!(cohort.members[&source].len(), 1);
+    }
+
+    #[test]
+    #[serial_test::serial]
+    fn service_cohort_overflow_eventually_services_every_stable_member() {
+        let source = DataUsageCacheSource::new(0, 0);
+        let inventory = cohort_inventory(&["aa", "bb", "cc", "dd", "ee", "ff"]);
+        let mut cohort = ScannerServiceCohort {
+            max_members: 2,
+            max_name_bytes: 4,
+            ..Default::default()
+        };
+        let mut admitted = HashSet::new();
+        for _ in 0..3 {
+            cohort.refresh(&inventory);
+            let mut buckets = inventory[&source].clone();
+            cohort.order_buckets(source, &mut buckets);
+            for bucket in buckets.iter().take(2) {
+                admitted.insert(bucket.name.clone());
+                cohort.record_admitted(source, &bucket.name);
+            }
+        }
+        assert_eq!(
+            admitted,
+            HashSet::from([
+                "aa".to_string(),
+                "bb".to_string(),
+                "cc".to_string(),
+                "dd".to_string(),
+                "ee".to_string(),
+                "ff".to_string(),
+            ]),
+            "stable overflow inventory must rotate every member through the tracked window"
+        );
+        assert!(cohort.overflowed);
+    }
+
+    #[test]
+    #[serial_test::serial]
+    fn service_cohort_bounds_names_and_does_not_reset_duplicate_dirty_age() {
+        let source = DataUsageCacheSource::new(0, 0);
+        let mut cohort = ScannerServiceCohort {
+            max_members: 2,
+            max_name_bytes: 4,
+            ..Default::default()
+        };
+        cohort.refresh(&cohort_inventory(&["aa", "bb", "long-name"]));
+        let queued_at = cohort.members[&source]["bb"].queued_at;
+        for _ in 0..10 {
+            cohort.refresh(&cohort_inventory(&["aa", "aa", "bb", "long-name"]));
+            assert_eq!(cohort.members.values().map(HashMap::len).sum::<usize>(), 2);
+            assert_eq!(
+                cohort
+                    .members
+                    .values()
+                    .flat_map(HashMap::keys)
+                    .map(|name| name.len())
+                    .sum::<usize>(),
+                4
+            );
+            assert_eq!(cohort.members[&source]["bb"].queued_at, queued_at);
+        }
+        cohort.refresh(&cohort_inventory(&[]));
+        assert!(cohort.members.is_empty());
+        assert!(cohort.cursor.is_none());
+    }
+
    fn active_bucket_drive_count(source: ScannerWorkSource, bucket: &str, drive: &str) -> u64 {
        global_metrics()
            .scanner_runtime_details_report()
@@ -117,6 +117,7 @@ impl ScannerIOCache for SetDisks {
            digest: scan_plan_digest,
            bucket_coverage_digest,
            requires_full_scan,
+            service_cohort,
            execution_digest,
            leader_epoch,
            tier_registry_generation,
@@ -500,7 +501,13 @@ impl ScannerIOCache for SetDisks {

        let mut permutes = buckets.clone();
        permutes.shuffle(&mut rand::rng());
-        let scan_order = bucket_usage_scan_order(&permutes, &old_cache, &dirty_usage_buckets);
+        let mut scan_order = bucket_usage_scan_order(&permutes, &old_cache, &dirty_usage_buckets);
+        if let Some(cohort) = &service_cohort {
+            cohort
+                .lock()
+                .unwrap_or_else(|poisoned| poisoned.into_inner())
+                .order_buckets(source, &mut scan_order);
+        }

        for bucket in scan_order.iter() {
            if let Some(c) = old_cache.find(&bucket.name) {
@@ -558,6 +565,7 @@ impl ScannerIOCache for SetDisks {
        let remaining_bucket_work = Arc::new(AtomicUsize::new(buckets.len()));
        let bucket_work_complete = CancellationToken::new();
        for (disk, worker_mode) in workers {
+            let service_cohort_clone = service_cohort.clone();
            let bucket_rx_mutex_clone = bucket_rx_mutex.clone();
            let bucket_tx_clone = bucket_tx.clone();
            let remaining_bucket_work_clone = remaining_bucket_work.clone();
@@ -577,6 +585,7 @@ impl ScannerIOCache for SetDisks {
            let partial_dirty_buckets_clone = bucket_failures.partial.clone();
            let pending_maintenance_work_clone = pending_maintenance_work.clone();
            let dirty_usage_buckets_clone = dirty_usage_buckets.clone();
+            let scope_clone = scope.clone();
            let cache_cycle_floor_clone = cache_cycle_floor.clone();
            let expected_publication_epoch_clone = expected_publication_epoch;
            let remote_server_epoch = match worker_mode {
@@ -587,6 +596,18 @@ impl ScannerIOCache for SetDisks {
                let remote_session_id = uuid::Uuid::new_v4();
                let mut remote_session_sequence = 0_u64;
                loop {
+                    // Do not prefetch a FIFO member into an independently
+                    // scheduled permit waiter: that can reorder admissions.
+                    let permit_wait_start = Instant::now();
+                    let Some(_permit) =
+                        wait_for_bucket_scan_permit(&disk_scan_semaphore_clone, &ctx_clone, &bucket_work_complete_clone).await
+                    else {
+                        break;
+                    };
+                    if ctx_clone.is_cancelled() || budget_clone.budget_elapsed() {
+                        break;
+                    }
+                    let permit_wait_elapsed = permit_wait_start.elapsed();
                    let bucket = tokio::select! {
                        _ = bucket_work_complete_clone.cancelled() => break,
                        _ = ctx_clone.cancelled() => break,
@@ -599,42 +620,34 @@ impl ScannerIOCache for SetDisks {
                    };
                    let mut work_guard =
                        BucketWorkGuard::new(remaining_bucket_work_clone.clone(), bucket_work_complete_clone.clone());
+                    // Prefix hints are process-local. Never hand one to a
+                    // remote or legacy-coordinator disk path.
+                    let prefix_scan_scope = disk_clone
+                        .is_local()
+                        .then(|| scope_clone.prefix_scope_for(&bucket.name))
+                        .flatten();

-                    let permit_wait = ctx_clone.clone();
-                    let permit_wait_start = Instant::now();
-                    let _permit = tokio::select! {
-                        permit = disk_scan_semaphore_clone.clone().acquire_owned() => match permit {
-                            Ok(permit) => permit,
-                            Err(_) => {
-                                decrement_disk_bucket_scans_queued(
-                                    &queued_disk_bucket_scans_clone,
-                                    &pool_label_clone,
-                                    &set_label_clone,
-                                );
-                                break;
-                            },
-                        },
-                        _ = permit_wait.cancelled() => {
-                            decrement_disk_bucket_scans_queued(
-                                &queued_disk_bucket_scans_clone,
-                                &pool_label_clone,
-                                &set_label_clone,
-                            );
-                            break;
-                        },
-                    };
                    metrics::histogram!(
                        METRIC_SCANNER_DISK_SCAN_WAIT_SECONDS,
                        "pool" => pool_label_clone.clone(),
                        "set" => set_label_clone.clone()
                    )
-                    .record(permit_wait_start.elapsed().as_secs_f64());
+                    .record(permit_wait_elapsed.as_secs_f64());
                    decrement_disk_bucket_scans_queued(&queued_disk_bucket_scans_clone, &pool_label_clone, &set_label_clone);
                    let _active_guard = DiskBucketScanActiveGuard::new(
                        active_disk_bucket_scans_clone.clone(),
                        pool_label_clone.clone(),
                        set_label_clone.clone(),
                    );
+                    if ctx_clone.is_cancelled() || budget_clone.budget_elapsed() {
+                        break;
+                    }
+                    if let Some(cohort) = &service_cohort_clone {
+                        cohort
+                            .lock()
+                            .unwrap_or_else(|poisoned| poisoned.into_inner())
+                            .record_admitted(source, &bucket.name);
+                    }

                    debug!(
                        target: "rustfs::scanner::io",
@@ -1048,7 +1061,10 @@ impl ScannerIOCache for SetDisks {
                        set_disk_inventory_clone.as_ref().clone(),
                        cache.clone(),
                        None,
-                        scan_mode,
+                        ScannerDiskScanOptions {
+                            scan_mode,
+                            prefix_scan_scope,
+                        },
                    );
                    tokio::pin!(scan);
                    let mut lock_watch = tokio::time::interval(SCANNER_CACHE_LOCK_POLL_INTERVAL);
@@ -72,7 +72,9 @@ where
        scan_mode,
        scan_scope: ScannerBucketScanScope::default(),
        persisted_usage_baseline: None,
+        observed_usage_candidate: None,
        requires_full_scan: true,
+        service_cohort: None,
        #[cfg(test)]
        resolved_scope_observer: None,
    };
@@ -88,52 +90,79 @@ pub(crate) struct ScannerCycleRequest {
    pub(crate) scan_mode: HealScanMode,
    pub(crate) scan_scope: ScannerBucketScanScope,
    pub(crate) persisted_usage_baseline: Option<Bytes>,
+    pub(crate) observed_usage_candidate: Option<Bytes>,
    /// Scheduled maintenance must visit clean buckets even with a valid dirty scope.
    pub(crate) requires_full_scan: bool,
+    pub(crate) service_cohort: Option<Arc<StdMutex<ScannerServiceCohort>>>,
    #[cfg(test)]
    pub(crate) resolved_scope_observer: Option<tokio::sync::oneshot::Sender<ScannerBucketScanScope>>,
 }

-struct ScannerBucketScopeResolution<'a> {
-    requested_scope: ScannerBucketScanScope,
-    baseline_proof: ScannerCacheBaselineProof<'a>,
-    activity_before: &'a crate::scanner::ScannerActivitySnapshot,
-    dirty_usage_snapshot: &'a DirtyUsageSnapshot,
-    all_buckets: &'a [BucketInfo],
-    requires_full_scan: bool,
+pub(super) struct ScannerBucketScopeResolution<'a> {
+    pub(super) requested_scope: ScannerBucketScanScope,
+    pub(super) baseline_proof: ScannerCacheBaselineProof<'a>,
+    pub(super) activity_before: &'a crate::scanner::ScannerActivitySnapshot,
+    pub(super) dirty_usage_snapshot: &'a DirtyUsageSnapshot,
+    pub(super) all_buckets: &'a [BucketInfo],
+    pub(super) requires_full_scan: bool,
+    #[cfg(test)]
+    pub(super) test_peer_snapshots: Option<Vec<(String, crate::storage_api::EcstoreScannerPeerDirtyUsageSnapshot)>>,
+    #[cfg(test)]
+    pub(super) test_scoped_dirty_usage_capability: Option<bool>,
 }

 async fn resolve_scanner_bucket_scan_scope<S>(
    store: &S,
    distributed: bool,
    resolution: ScannerBucketScopeResolution<'_>,
-) -> ScannerBucketScanScope
+) -> ScannerBucketScopeResolutionResult
 where
    S: ScannerStorage,
 {
+    let default_result = |scope: ScannerBucketScanScope| ScannerBucketScopeResolutionResult {
+        scope,
+        remote_dirty_usage_acknowledgements: Vec::new(),
+    };
    if resolution.requires_full_scan {
-        return ScannerBucketScanScope::default();
+        return default_result(ScannerBucketScanScope::default());
    }
    if !resolution.requested_scope.is_default()
        || !resolution.dirty_usage_snapshot.covers_all_pending
        || resolution.dirty_usage_snapshot.generation == u64::MAX
        || resolution.dirty_usage_snapshot.buckets.len() > crate::SCANNER_DIRTY_USAGE_SNAPSHOT_MAX_ENTRIES
    {
-        return resolution.requested_scope;
+        return default_result(resolution.requested_scope);
    }

-    let mut dirty_buckets = resolution
+    let dirty_buckets = resolution
        .dirty_usage_snapshot
        .buckets
        .keys()
        .cloned()
        .collect::<HashSet<_>>();
    if distributed {
-        let Some(notification_system) = store.scanner_notification_system() else {
-            return resolution.requested_scope;
+        let notification_system = store.scanner_notification_system();
+        #[cfg(test)]
+        let peer_snapshots = if let Some(peer_snapshots) = resolution.test_peer_snapshots.clone() {
+            peer_snapshots
+        } else {
+            let Some(notification_system) = notification_system.as_ref() else {
+                return default_result(resolution.requested_scope);
+            };
+            let Ok(peer_snapshots) = notification_system.scanner_dirty_usage_snapshots().await else {
+                return default_result(resolution.requested_scope);
+            };
+            peer_snapshots
        };
-        let Ok(peer_snapshots) = notification_system.scanner_dirty_usage_snapshots().await else {
-            return resolution.requested_scope;
+        #[cfg(not(test))]
+        let peer_snapshots = {
+            let Some(notification_system) = notification_system.as_ref() else {
+                return default_result(resolution.requested_scope);
+            };
+            let Ok(peer_snapshots) = notification_system.scanner_dirty_usage_snapshots().await else {
+                return default_result(resolution.requested_scope);
+            };
+            peer_snapshots
        };
        let mut expected_peers = HashMap::new();
        for (host, lease_instance_id, _) in crate::scanner::scanner_activity_publication_lease_targets(resolution.activity_before)
@@ -141,10 +170,10 @@ where
            let Some((activity_instance_id, generation, pending)) =
                crate::scanner::scanner_activity_dirty_usage_state_for_host(resolution.activity_before, &host)
            else {
-                return resolution.requested_scope;
+                return default_result(resolution.requested_scope);
            };
            if activity_instance_id != lease_instance_id || expected_peers.contains_key(&host) {
-                return resolution.requested_scope;
+                return default_result(resolution.requested_scope);
            }
            expected_peers.insert(
                host,
@@ -155,19 +184,82 @@ where
                },
            );
        }
-        let Some(remote_dirty_buckets) = verified_remote_dirty_usage_buckets(&expected_peers, peer_snapshots) else {
-            return resolution.requested_scope;
+        let Some(remote_dirty_usage) = verified_remote_dirty_usage(&expected_peers, peer_snapshots) else {
+            return default_result(resolution.requested_scope);
        };
-        dirty_buckets.extend(remote_dirty_buckets);
+        let remote_resolution = resolve_remote_dirty_usage_scope(
+            resolution.requested_scope,
+            dirty_buckets,
+            remote_dirty_usage,
+            resolution.all_buckets,
+            resolution.baseline_proof,
+        );
+        if !remote_resolution.remote_dirty_usage_acknowledgements.is_empty() {
+            #[cfg(test)]
+            let capability_supported = if let Some(capability_supported) = resolution.test_scoped_dirty_usage_capability {
+                capability_supported
+            } else {
+                let Some(notification_system) = notification_system.as_ref() else {
+                    return default_result(ScannerBucketScanScope::default());
+                };
+                let capability_acknowledgements = remote_resolution
+                    .remote_dirty_usage_acknowledgements
+                    .clone()
+                    .into_iter()
+                    .map(Into::into)
+                    .collect::<Vec<crate::storage_api::EcstoreScannerDirtyUsageAcknowledgement>>();
+                matches!(
+                    notification_system
+                        .scanner_scoped_dirty_usage_capabilities(capability_acknowledgements)
+                        .await,
+                    Ok(true)
+                )
+            };
+            #[cfg(not(test))]
+            let capability_supported = {
+                let Some(notification_system) = notification_system.as_ref() else {
+                    return default_result(ScannerBucketScanScope::default());
+                };
+                let capability_acknowledgements = remote_resolution
+                    .remote_dirty_usage_acknowledgements
+                    .clone()
+                    .into_iter()
+                    .map(Into::into)
+                    .collect::<Vec<crate::storage_api::EcstoreScannerDirtyUsageAcknowledgement>>();
+                matches!(
+                    notification_system
+                        .scanner_scoped_dirty_usage_capabilities(capability_acknowledgements)
+                        .await,
+                    Ok(true)
+                )
+            };
+            if !capability_supported {
+                return default_result(ScannerBucketScanScope::default());
+            }
+        }
+        return remote_resolution;
    }

-    scoped_scan_scope_from_dirty_buckets(
+    default_result(scoped_scan_scope_from_dirty_buckets(
        resolution.requested_scope,
        dirty_buckets,
+        (!distributed).then_some(resolution.dirty_usage_snapshot.scopes.as_ref()),
        true,
        resolution.all_buckets,
        resolution.baseline_proof,
-    )
+    ))
+}
+
+#[cfg(test)]
+pub(super) async fn resolve_scanner_bucket_scan_scope_for_tests<S>(
+    store: &S,
+    distributed: bool,
+    resolution: ScannerBucketScopeResolution<'_>,
+) -> ScannerBucketScopeResolutionResult
+where
+    S: ScannerStorage,
+{
+    resolve_scanner_bucket_scan_scope(store, distributed, resolution).await
 }

 pub(crate) async fn nsscanner_with_storage_status_scoped<S>(store: &S, request: ScannerCycleRequest) -> Result<ScannerCycleResult>
@@ -183,7 +275,9 @@ where
        scan_mode,
        scan_scope,
        persisted_usage_baseline,
+        observed_usage_candidate,
        requires_full_scan,
+        service_cohort,
        #[cfg(test)]
        resolved_scope_observer,
    } = request;
@@ -275,6 +369,12 @@ where
    }
    bucket_plan_complete &= buckets_by_source.keys().copied().collect::<HashSet<_>>() == *expected_sources;
    bucket_plan_complete &= scanner_bucket_inventory_is_complete(&all_buckets, &buckets_by_source);
+    if bucket_plan_complete && let Some(cohort) = &service_cohort {
+        cohort
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .refresh(&buckets_by_source);
+    }
    let structural_scan_plan_digest =
        scanner_bucket_plan_digest(&all_buckets, crate::scanner::scanner_activity_structural_digest(&activity_before));
    let scan_plan_digest = scanner_bucket_work_digest(structural_scan_plan_digest, scan_mode, requires_full_scan);
@@ -282,13 +382,14 @@ where
    let bucket_coverage_digest = scanner_bucket_plan_digest(&all_buckets, activity_digest);
    let execution_digest = scanner_bucket_work_digest(bucket_coverage_digest, scan_mode, requires_full_scan);
    let dirty_usage_snapshot = Arc::new(snapshot_dirty_usage_buckets(&all_buckets, dirty_generation_before_bucket_list));
-    let scan_scope = resolve_scanner_bucket_scan_scope(
+    let scope_resolution = resolve_scanner_bucket_scan_scope(
        store,
        distributed,
        ScannerBucketScopeResolution {
            requested_scope: scan_scope,
            baseline_proof: ScannerCacheBaselineProof {
-                data: persisted_usage_baseline.as_ref(),
+                authoritative_data: persisted_usage_baseline.as_ref(),
+                observed_candidate_data: observed_usage_candidate.as_ref(),
                expected_sources: &expected_sources,
                leader_epoch,
                want_cycle,
@@ -298,9 +399,15 @@ where
            dirty_usage_snapshot: &dirty_usage_snapshot,
            all_buckets: &all_buckets,
            requires_full_scan: requires_full_scan || scan_mode == HealScanMode::Deep,
+            #[cfg(test)]
+            test_peer_snapshots: None,
+            #[cfg(test)]
+            test_scoped_dirty_usage_capability: None,
        },
    )
    .await;
+    let remote_dirty_usage_acknowledgements = scope_resolution.remote_dirty_usage_acknowledgements;
+    let scan_scope = scope_resolution.scope;
    #[cfg(test)]
    if let Some(observer) = resolved_scope_observer {
        let _ = observer.send(scan_scope.clone());
@@ -325,12 +432,21 @@ where
            dirty_usage_status,
            activity_status,
        );
-        let empty_usage = DataUsageInfo {
-            last_update: Some(SystemTime::now()),
-            scanner_cycle: Some(want_cycle),
-            usage_snapshot_complete: true,
-            ..Default::default()
+        let Some(candidate) = empty_namespace_usage_candidate(
+            &all_buckets,
+            &expected_sources,
+            &buckets_by_source,
+            ScannerSnapshotIdentity {
+                cycle: want_cycle,
+                leader_epoch,
+                plan_digest: scan_plan_digest,
+                coverage_digest: bucket_coverage_digest,
+                tier_registry_generation: Some(tier_registry_generation),
+            },
+        ) else {
+            return Ok(ScannerCycleResult::new(ScannerCycleStatus::Incomplete, None).with_publication_epoch(publication_epoch));
        };
+        let (empty_usage, publication_expectation) = candidate.prepare(status);
        let observational_snapshot_published = if should_publish_observational_snapshot(status) {
            publish_observational_snapshot(&updates, empty_usage).await?
        } else {
@@ -353,7 +469,8 @@ where
            .with_activity_digest(activity_digest)
            .with_observational_snapshot_published(observational_snapshot_published)
            .with_remote_publication_lease_targets(remote_publication_lease_targets)
-            .with_remote_dirty_usage_acknowledgements(remote_dirty_usage_acknowledgements));
+            .with_remote_dirty_usage_acknowledgements(remote_dirty_usage_acknowledgements)
+            .with_publication_expectation(publication_expectation));
    }

    let total_results = expected_sources.len();
@@ -399,7 +516,32 @@ where
    let first_err_mutex: Arc<Mutex<Option<Error>>> = Arc::new(Mutex::new(None));
    let mut wait_futs = Vec::new();

-    for (results_index, set) in set_disks.iter().enumerate() {
+    let set_order = service_cohort.as_ref().map_or_else(
+        || (0..set_disks.len()).collect::<Vec<_>>(),
+        |cohort| {
+            cohort
+                .lock()
+                .unwrap_or_else(|poisoned| poisoned.into_inner())
+                .order_set_indices(&set_disks)
+        },
+    );
+    for results_index in set_order {
+        let set = &set_disks[results_index];
+        // Acquire in dispatch order, not in independently scheduled tasks.
+        // A whole set still shares the existing parent budget; this is not
+        // a per-bucket quantum or a cross-source completion guarantee.
+        let permit_wait_start = Instant::now();
+        let permit = tokio::select! {
+            biased;
+            _ = child_token.cancelled() => break,
+            permit = set_scan_semaphore.clone().acquire_owned() => match permit {
+                Ok(permit) => permit,
+                Err(_) => break,
+            },
+        };
+        if child_token.is_cancelled() || budget.budget_elapsed() {
+            break;
+        }
        let results_index_clone = results_index;
        // Clone the Arc to move it into the spawned task
        let set_clone: Arc<SetDisks> = Arc::clone(set);
@@ -414,7 +556,6 @@ where
        let scan_mode_clone = scan_mode;
        let results_mutex_clone = results_mutex.clone();
        let first_err_mutex_clone = first_err_mutex.clone();
-        let set_scan_semaphore_clone = set_scan_semaphore.clone();
        let queued_set_scans_clone = queued_set_scans.clone();
        let active_set_scans_clone = active_set_scans.clone();

@@ -437,6 +578,7 @@ where
            digest: structural_scan_plan_digest,
            bucket_coverage_digest,
            requires_full_scan,
+            service_cohort: service_cohort.clone(),
            execution_digest,
            leader_epoch,
            tier_registry_generation,
@@ -448,15 +590,10 @@ where
        };
        // Spawn task to run the scanner
        let scanner_fut = tokio::spawn(async move {
-            let permit_wait = child_token_clone.clone();
-            let permit_wait_start = Instant::now();
-            let _permit = tokio::select! {
-                permit = set_scan_semaphore_clone.acquire_owned() => match permit {
-                    Ok(permit) => permit,
-                    Err(_) => return,
-                },
-                _ = permit_wait.cancelled() => return,
-            };
+            let _permit = permit;
+            if child_token_clone.is_cancelled() || budget_clone.budget_elapsed() {
+                return;
+            }
            metrics::histogram!(
                METRIC_SCANNER_SET_SCAN_WAIT_SECONDS,
                "pool" => pool_label.clone(),
@@ -562,7 +699,7 @@ where
    let (activity_status, remote_publication_lease_targets) =
        scanner_cycle_activity_status(store, distributed, &activity_before).await;
    let all_bucket_names = all_buckets.iter().map(|bucket| bucket.name.clone()).collect::<Vec<_>>();
-    let completed_usage = completed_data_usage_info(
+    let completed_usage = completed_usage_candidate(
        &results,
        &ScannerSnapshotScope {
            sources: &expected_sources,
@@ -603,7 +740,10 @@ where
        dirty_usage_status,
        activity_status,
    );
-    let observational_snapshot_published = if let Some((data_usage_info, _)) = completed_usage {
+    let mut publication_expectation = None;
+    let observational_snapshot_published = if let Some(candidate) = completed_usage {
+        let (data_usage_info, expectation) = candidate.prepare(cycle_status);
+        publication_expectation = expectation;
        if should_publish_observational_snapshot(cycle_status) {
            publish_observational_snapshot(&updates, data_usage_info).await?
        } else {
@@ -628,11 +768,14 @@ where
    if cycle_status == ScannerCycleStatus::Complete {
        complete_tier_registry_cycle(want_cycle, leader_epoch);
    }
-    let remote_dirty_usage_acknowledgements = if cycle_status == ScannerCycleStatus::Complete {
-        crate::scanner::scanner_dirty_usage_acknowledgements(&activity_before)
-    } else {
-        Vec::new()
-    };
+    let remote_dirty_usage_acknowledgements =
+        if cycle_status == ScannerCycleStatus::Complete && !remote_dirty_usage_acknowledgements.is_empty() {
+            remote_dirty_usage_acknowledgements
+        } else if cycle_status == ScannerCycleStatus::Complete && scan_scope.is_default() {
+            crate::scanner::scanner_dirty_usage_acknowledgements(&activity_before)
+        } else {
+            Vec::new()
+        };
    Ok(ScannerCycleResult::new(cycle_status, dirty_usage_clear)
        .with_publication_epoch(publication_epoch)
        .with_activity_digest(activity_digest)
@@ -641,5 +784,6 @@ where
        .with_remote_dirty_usage_acknowledgements(remote_dirty_usage_acknowledgements)
        .with_failed_dirty_usage(!failed_buckets.is_empty())
        .with_pending_maintenance_work(pending_maintenance_work)
-        .with_required_cycle_floor(required_cycle_floor))
+        .with_required_cycle_floor(required_cycle_floor)
+        .with_publication_expectation(publication_expectation))
 }
@@ -146,7 +146,7 @@ impl ScannerIODisk for Disk {
        Ok(size_summary)
    }

-    #[tracing::instrument(skip(self, budget, updates, cache, set_disks))]
+    #[tracing::instrument(skip(self, budget, updates, cache, set_disks, options), fields(scan_mode = ?options.scan_mode))]
    async fn nsscanner_disk(
        self: Arc<Self>,
        ctx: CancellationToken,
@@ -154,8 +154,12 @@ impl ScannerIODisk for Disk {
        set_disks: Vec<Arc<Disk>>,
        cache: DataUsageCache,
        updates: Option<mpsc::Sender<DataUsageEntry>>,
-        scan_mode: HealScanMode,
+        options: ScannerDiskScanOptions,
    ) -> Result<ScannerDiskScanOutcome> {
+        let ScannerDiskScanOptions {
+            scan_mode,
+            prefix_scan_scope,
+        } = options;
        let done_drive = Metrics::time(Metric::ScanBucketDrive);
        let drive_start = std::time::Instant::now();
        let bucket = cache.info.name.clone();
@@ -198,7 +202,19 @@ impl ScannerIODisk for Disk {
            cache.info.object_lock = Some(Arc::new(object_lock_config));
        }

-        let result = scan_data_folder(
+        // Prefix reuse never crosses semantic maintenance boundaries. A
+        // lifecycle, replication, Object Lock, or erasure health walk can
+        // make a clean data subtree require scanner-side work even without a
+        // direct object mutation in the local journal. The folder scanner
+        // separately rejects scopes in erasure mode.
+        let prefix_scan_scope = (scan_mode == HealScanMode::Normal
+            && cache.info.lifecycle.is_none()
+            && cache.info.replication.is_none()
+            && cache.info.object_lock.is_none())
+        .then_some(prefix_scan_scope)
+        .flatten();
+
+        let result = scan_data_folder_scoped(
            ctx.clone(),
            budget,
            set_disks,
@@ -207,6 +223,7 @@ impl ScannerIODisk for Disk {
            updates,
            scan_mode,
            SCANNER_SLEEPER.clone(),
+            prefix_scan_scope,
        )
        .await;

--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
houseme	df06b7e168	test(heal): cover set bulkhead scheduler isolation Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com>	2026-09-07 12:12:16 +08:00
Zhengchao An	672087ec0d	ci(docker): move latest on every prerelease until first stable tag (#7341 ) Since rc.1 the latest tag has been frozen at 1.0.0-beta.12 because the docker workflow only allowed alpha/beta prereleases to update latest (#2732 dropped the rc case). Before 1.0.0 GA, latest is expected to track the newest test build. - Prereleases (alpha/beta/rc) now update latest as long as no stable vX.Y.Z tag exists on origin, so the rule retires itself at GA. - Channel tags (alpha/beta/rc) are now always added for prereleases; the previous if/elif skipped the channel tag whenever latest was created, which is why no :beta tag was ever published.	2026-09-07 11:25:40 +08:00
houseme	9dbeae1b45	chore(deps): refresh scanner batch baseline (#7336 ) Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-07 11:04:34 +08:00
cxymds	6018dd372f	perf(ilm): reduce transition transaction mutations (#7320 ) * perf(ilm): reduce transition transaction mutations * test(ilm): rename transition kill points --------- Co-authored-by: Zhengchao An <anzhengchao@gmail.com>	2026-09-07 02:40:57 +00:00
houseme	d633a635ec	feat(madmin): add scanner reset client helpers (#7333 ) Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-07 10:28:42 +08:00
houseme	fbd30a6f43	test(scanner): cover concurrent cycle reset convergence (#7335 ) Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-07 10:28:37 +08:00
houseme	a32360198c	test(scanner): cover scoped ack resolver fallback (#7332 ) * test(scanner): cover scoped ack resolver fallback Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * test(scanner): reduce scoped resolver helper args Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> --------- Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-07 10:28:19 +08:00
GatewayJ	05efc584b0	perf(s3-tables): avoid duplicate registration prefix scans (#7310 ) Co-authored-by: houseme <housemecn@gmail.com> Co-authored-by: Zhengchao An <anzhengchao@gmail.com>	2026-09-07 09:55:42 +08:00
cxymds	a70c96d520	feat(ilm): execute legacy recovery dispositions (#7304 ) * feat(ilm): execute legacy recovery dispositions * feat(ilm): retry retained transition recovery (#7308) * fix(admin): use gateway errors for recovery retries --------- Co-authored-by: houseme <housemecn@gmail.com> Co-authored-by: Zhengchao An <anzhengchao@gmail.com>	2026-09-07 01:46:43 +00:00
Zhengchao An	b0f68f5d0a	test(upgrade): read and replicate rc.5 multipart layouts (#7334 )	2026-09-07 09:37:21 +08:00
Zhengchao An	df554439b0	fix(heal): fall back to set-wide format for directory-backed targets (#7331 ) fix(heal): fall back to set-wide format for directory-backed replacement targets Since #7018 `renew_disk` routes an unformatted local endpoint through the automatic replacement heal, which requires the target to be an independently mounted disk. Directory-backed deployments (the operator set `RUSTFS_UNSAFE_BYPASS_DISK_CHECK`, which the startup disk-independence check already requires for endpoints sharing a device) can never pass that admission, so a runtime-wiped or replaced directory disk stayed unformatted forever: the heal task failed with "replacement target is not a stable mounted disk" and the auto-scan kept deferring the endpoint. This broke the Issue #1533 contract and the `heal_erasure_disk_rebuild_test` lane on main since 2026-09-02. When the disk-check bypass is set and the replacement target fails admission, the erasure-set heal now logs a warning and downgrades to the ordinary set-wide `heal_format` path that predated replacement admission, and the auto-scan no longer defers such endpoints. The mount admission itself is unchanged and still cannot be bypassed by any environment variable; deployments without the bypass keep failing closed. The endpoint-blackhole heal scenario now probes whether `iptables` can read the OUTPUT chain and logs an explicit skip when the host lacks `CAP_NET_ADMIN` (containerised runners report "Permission denied" from the nf_tables backend even under sudo); `RUSTFS_E2E_REQUIRE_NET_FAULT_INJECTION=1` turns that into a failure for lanes that provision the capability. The CI full-gate job surfaces the missing capability as a workflow warning, and the runtime-wipe fixture retries `remove_dir_all` on the listing race macOS surfaces as `DirectoryNotEmpty`. Refs rustfs/backlog#2357.	2026-09-07 09:25:44 +08:00
唐小鸭	3b404e56c0	fix(replication): forward single-part object checksums as headers (#7313 )	2026-09-07 01:17:24 +00:00
Zhengchao An	ff00872922	docs(ilm): restore transition gate doc comment after handler reorder (#7325 ) Commit `086ee8e48` moved authorize_recovery_admin_request above authorize_transition_admin_request and left the transition gate's doc comment attached to the recovery gate, which returns a hashed actor rather than the masked access key the comment describes. Move the comment back onto the transition gate and give the recovery gate its own accurate description. No behavior change.	2026-09-07 08:59:20 +08:00
唐小鸭	4d1ce9618a	fix(scanner): pass dirty scopes to distributed scope resolution (#7329 )	2026-09-07 07:42:17 +08:00
houseme	a2bad0953f	test(scanner): collect heal evidence oracles from registry (#7328 )	2026-09-06 21:44:01 +00:00
唐小鸭	14c99a994c	fix(filemeta): keep data dir of a version awaiting purge replication (#7307 )	2026-09-07 05:30:44 +08:00
houseme	f6c2a9bfe0	test(scanner): cover overflow service cohort rotation (#7324 )	2026-09-07 05:29:59 +08:00
houseme	640d7e0e3c	test(heal): cover MRF snapshot recovery bounds (#7327 )	2026-09-07 05:29:25 +08:00
houseme	409ac3de66	test(scanner): cover reset cleanup process crash boundaries (#7326 )	2026-09-07 05:29:14 +08:00
唐小鸭	2a63fcbea6	fix(replication): stop duplicate re-drives on own-version-id targets (#7323 )	2026-09-07 05:29:00 +08:00
houseme	f4049598e4	feat(scanner): send scoped dirty usage acknowledgements (#7322 )	2026-09-07 05:28:50 +08:00
GatewayJ	1651541d38	fix(s3-tables): preserve encoded paths in client signatures (#7309 )	2026-09-07 05:28:06 +08:00
cxymds	85849788af	chore(tier): remove remaining blanket lint allowances (#7306 )	2026-09-07 05:27:29 +08:00
Henry Guo	975983abdd	feat(scanner): reuse clean local bucket prefixes (#7208 )	2026-09-07 05:26:56 +08:00
唐小鸭	cf1c45eb91	test(replication): pin directory-marker null version and replication (#7315 )	2026-09-06 19:26:09 +00:00
唐小鸭	536c283716	test(ilm): realign source-marker assertions with handler order (#7321 )	2026-09-07 02:10:13 +08:00
houseme	a04ddc237e	chore: remove Go heal outcome compatibility fixture (#7319 ) * del go code * chore: record ILM lifecycle validation on go-del (#7318) Co-authored-by: zhi22915 <qiuzgang@gmail.com> --------- Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 23:49:56 +08:00
houseme	7de6ac82e1	test(scanner): cover reset cleanup across store reopen (#7317 ) Add a real ECStore reopen regression for scanner usage-state reset cleanup boundaries. The fixture seeds each partially completed cleanup state, recreates the store, then verifies the reset resumes without rewriting the bootstrap intent or deleting unrelated metadata. Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 23:40:31 +08:00
唐小鸭	7993e11058	docs(s3): record object-key path-segment validation as intentional (#7314 )	2026-09-06 23:27:10 +08:00
houseme	37ead2f69c	test(ilm): align feature matrix lifecycle assertions (#7311 ) Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 23:27:01 +08:00
houseme	1b1e590df7	test(scanner): verify quota state across reset and owner restart (#7303 ) * test(scanner): verify quota state across reset and owner restart Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * fix: restore ILM transition and lifecycle validation (#7312) Co-authored-by: zhi22915 <qiuzgang@gmail.com> --------- Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 23:26:01 +08:00
houseme	086ee8e48a	fix(scanner): require root publication proof before dirty ack (#7297 ) * fix(scanner): require root publication proof before dirty ack Bind ACK expectations to validated scan candidates and confirm the actual primary-root revision and readback. Retain saved outcomes and dirty responsibility when stronger evidence is unavailable. Isolate CAS attempt confirmation and invalidate proof after scope mutations. Revalidate observed candidate reuse before issuing a new publication proof, preserve the exact validated authoritative baseline work digest, and settle fixture commit tails before stable maintenance scans. Keep scoped ACK production disabled. Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * fix: restore ILM transition and lifecycle validation (#7316) Co-authored-by: zhi22915 <qiuzgang@gmail.com> --------- Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 23:25:40 +08:00
Zhengchao An	22bff27aee	fix(storage): derive multipart identity from stored parts (#7305 )	2026-09-06 22:04:15 +08:00
houseme	8a20498705	fix(heal): retain completed reports during clock rollback (#7299 ) Treat a negative wall-clock age as zero without bypassing count or byte eviction. Cover canonical and alias queries, terminal outcomes, exact TTL expiry, and capacity limits during rollback. Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 21:28:22 +08:00
houseme	c3e6d90c6f	docs(heal): clarify start retries and execution budgets (#7300 ) Distinguish control requests, cumulative task execution and object retries. Explain ambiguous responses and bounded envelope replay without promising HTTP idempotency or changing runtime policy. Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 21:27:37 +08:00
Zhengchao An	d2af8f073e	test(s3): exclude invalid lifecycle filter vector (#7295 ) Co-authored-by: houseme <housemecn@gmail.com>	2026-09-06 21:21:12 +08:00
Zhengchao An	2baba1bda7	fix(scanner): preserve verified maintenance digest (#7293 ) Co-authored-by: houseme <housemecn@gmail.com>	2026-09-06 21:20:58 +08:00
Zhengchao An	bf2ca9113f	test(scanner): settle fixture writes before activity baseline (#7290 )	2026-09-06 21:20:46 +08:00
Zhengchao An	f474ea30d4	fix(admin): preserve decommission readiness error (#7285 )	2026-09-06 21:20:32 +08:00
cxymds	002ac9544c	feat(ilm): add immutable legacy recovery exports (#7283 ) * feat(ilm): add immutable legacy recovery exports * feat(ilm): add legacy recovery disposition records (#7292) * fix(ilm): stabilize legacy recovery decode errors * fix(admin): route recovery auth errors through gateway * fix(ilm): resolve recovery disposition clippy errors * fix(ilm): remove redundant recovery test clones	2026-09-06 21:20:16 +08:00
Zhengchao An	d74d970e24	fix(ecstore): retain namespace ownership during multipart commit (#7282 )	2026-09-06 21:20:02 +08:00
Zhengchao An	cc15eae479	fix(lifecycle): allow multiple filter predicates without And wrapper (#7298 )	2026-09-06 21:09:11 +08:00
dependabot[bot]	c130d00d4b	chore(deps): bump golang.org/x/net from 0.54.0 to 0.55.0 in /scripts/compat/heal-outcome in the go_modules group across 1 directory (#7296 ) chore(deps): bump golang.org/x/net Bumps the go_modules group with 1 update in the /scripts/compat/heal-outcome directory: [golang.org/x/net](https://github.com/golang/net). Updates `golang.org/x/net` from 0.54.0 to 0.55.0 - [Commits](https://github.com/golang/net/compare/v0.54.0...v0.55.0) --- updated-dependencies: - dependency-name: golang.org/x/net dependency-version: 0.55.0 dependency-type: indirect dependency-group: go_modules ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>	2026-09-06 19:13:21 +08:00
dependabot[bot]	e49d9cdea2	chore(deps): bump the go_modules group across 1 directory with 2 updates (#7288 ) Bumps the go_modules group with 1 update in the /scripts/compat/heal-outcome directory: [golang.org/x/crypto](https://github.com/golang/crypto). Updates `golang.org/x/crypto` from 0.37.0 to 0.52.0 - [Commits](https://github.com/golang/crypto/compare/v0.37.0...v0.52.0) Updates `golang.org/x/net` from 0.39.0 to 0.54.0 - [Commits](https://github.com/golang/net/compare/v0.39.0...v0.54.0) --- updated-dependencies: - dependency-name: golang.org/x/crypto dependency-version: 0.52.0 dependency-type: indirect dependency-group: go_modules - dependency-name: golang.org/x/net dependency-version: 0.54.0 dependency-type: indirect dependency-group: go_modules ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com>	2026-09-06 18:56:54 +08:00
houseme	f5778d8b97	chore(deps): refresh concurrency and protocol dependencies (#7289 ) chore(deps): update Crossbeam, Redis, DER, and ipnet Refresh the shared concurrency, Redis client, DER decoding and IP network dependencies while retaining existing feature selections and the hotpath pin. Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 18:51:12 +08:00
Zhengchao An	9da42a899a	test(connect): refresh offline enrollment e2e chain (#7294 )	2026-09-06 18:30:04 +08:00
Henry Guo	3d46ed312a	feat(scanner): reuse complete observed scan candidates (#7206 )	2026-09-06 17:52:52 +08:00
Zhengchao An	3496277e7c	test(ecstore): drain source writes before corrupting shards (#7291 )	2026-09-06 17:51:07 +08:00
cxymds	0b72f39023	fix(tier): avoid re-fencing published mutations (#7274 )	2026-09-06 16:57:03 +08:00
Zhengchao An	30a0937a7d	fix(ecstore): retain single-delete physical namespace ownership (#7287 )	2026-09-06 16:56:24 +08:00
houseme	5b962b6c58	feat(heal): expose compatible canonical v3 outcomes (#7256 ) * feat(heal): expose compatible canonical v3 outcomes Serialize the existing canonical heal outcome, retain the v3 summary vocabulary, and reject or conservatively adapt contradictory peer success responses. Preserve progress and bounded result cursors without treating legacy storage responses as repair proof. Add optional SDK outcome and cursor support with shared Rust fixtures, pinned legacy Go decoder and mc polling checks, and explicit compatibility limits. Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * fix(madmin): box heal stop task status outcome Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> --------- Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 16:46:42 +08:00
houseme	0ee5408b94	feat(scanner): preserve bounded bootstrap admission fairness (#7260 ) feat(scanner): retain bounded bootstrap admission fairness Keep a leader-local bounded cohort across scanner retries and preserve waiting bucket priority during dirty arrivals and capacity overflow. Order source permits in the dispatcher without changing result identity, parent budgets, explicit cycle timing, or persistent coverage evidence. Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 16:45:57 +08:00
houseme	cb3100a252	fix(scanner): visit compacted subtrees during deep scans (#7270 ) Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 16:45:17 +08:00
houseme	3a4afe9b38	fix(heal): bound cross-page object retry delays (#7273 ) Retain failed identities in an execution-local count and byte bounded window so healthy later pages can advance. Preserve retry jitter, deadlines, terminal accounting and pressure pacing, with deterministic head-of-line and capacity regressions. Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 16:44:59 +08:00
houseme	71859ff83c	fix(scanner): resume committed cleanup when scanning is disabled (#7281 ) Run one supervised cleanup attempt for an existing operator reset, with strict phase and revision checks under the original leader lock. Keep v3 reset authorization and responses unchanged, report deferred status, and bound probe and shutdown waits without aborting in-flight reset ownership. Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 16:44:45 +08:00
houseme	cb72df269a	fix(heal): retain hints without verified repair receipts (#7275 ) Stop task completion and legacy notices from discharging scanner retry hints. Preserve existing hints and their retry due time across admission observations, bound retry scheduling, and synchronize changed batches once even on cancellation. Exercise the production MRF consumer, manager, event channel and scanner ledger. Document producer durability gaps without enabling successor activation or garbage collection. Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-09-06 16:44:30 +08:00