Merge branch 'main' into ci/validation-sweep

chore(deps): update flake.lock (#9108 )
Flake lock file updates: • Updated input 'nixpkgs': 'github:NixOS/nixpkgs/674a202' (2026-10-24) → 'github:NixOS/nixpkgs/a078812' (2026-10-31) • Updated input 'rust-overlay': 'github:oxalica/rust-overlay/5b9f119' (2026-10-24) → 'github:oxalica/rust-overlay/6ecf534' (2026-10-31)
2026-11-01 08:40:23 +00:00 · 2026-11-01 16:22:00 +08:00 · 2026-11-01 06:26:39 +00:00 · 2026-11-01 04:49:50 +00:00 · 2026-11-01 13:07:42 +08:00 · 2026-11-01 12:47:38 +08:00
209 changed files with 17103 additions and 1250 deletions
@@ -1,2 +1,2 @@
-sha256-darwin=67251cb7d9978d38e899268cac4d7898297ff7e5b5c48c9c1546dc760b93568d
-sha256-linux=673da76c2e5710c4622fe2fa32d2b297730895769487fcf1152ee1c2366f32f4
+sha256-darwin=7c705942d3a9e583ebaa85845daa21c4f81fdbf36def24db931aaef38636c53e
+sha256-linux=5c33dc395233671cb77452f8fb12900ecdbeb368632222b96835d310bce1aab8
@@ -1 +1 @@
-sha256=e9dc896774122a304e2b8b90f28c82d4d4e0af8ded36a82b147b30945ec994e3
+sha256=00f8a56b75990b487a594b97c36610b94476a6944ba00473e37187ea654d42bf
@@ -0,0 +1,77 @@
+{
+  "schema_version": 1,
+  "plan_id": "rustfs-release-validation",
+  "description": "Versioned certification plan for every public RustFS release exit (rustfs/backlog#2662, R01). preview, prerelease and stable all require this explicit plan. Scripts/release/release_gate.py fail-closes publication on any missing, duplicate, stale, substituted or non-PASS required item.",
+  "release_types": ["preview", "prerelease", "stable"],
+  "targets": [
+    {"os": "linux", "arch": "x86_64", "abi": "musl"},
+    {"os": "linux", "arch": "aarch64", "abi": "musl"},
+    {"os": "linux", "arch": "x86_64", "abi": "gnu"},
+    {"os": "linux", "arch": "aarch64", "abi": "gnu"},
+    {"os": "macos", "arch": "aarch64", "abi": "none"},
+    {"os": "windows", "arch": "x86_64", "abi": "msvc"}
+  ],
+  "targets_basis": "Platform matrix of .github/workflows/build.yml (prepare-platform-matrix). gnu and macOS targets build with the pyroscope feature; musl and Windows targets build with no extra features.",
+  "features": ["pyroscope"],
+  "trusted_workflows": [
+    {"repository": "rustfs/s3chaos", "workflow_id": "release-gate.yml"}
+  ],
+  "trusted_workflows_basis": "The shared evidence envelope producer lives in rustfs/s3chaos (G01, fault/release_gate.rs) and its plan/live split workflow is .github/workflows/release-gate.yml there. Only results from these identities may enter a release certification; PR-uploaded JSON never gains release authority. R02 must extend this list when additional verifier workflows land, never widen it to untrusted sources.",
+  "lanes": [
+    {
+      "lane": "identity",
+      "harness_repository": "rustfs/s3chaos",
+      "harness_sha": "9a48f55c1d6d58ce63a8fb1ea7317e024f32f11d",
+      "harness_basis": "Fixed G01 evidence-envelope producer revision (rustfs/s3chaos PR #102 head recorded in rustfs/backlog#2658). R02 re-pins this when the live verifier ships.",
+      "timeout_seconds": 1800,
+      "environment_capabilities": ["process-exec"],
+      "cases": ["artifact-identity", "startup"],
+      "applies_to": {
+        "release_types": ["preview", "prerelease", "stable"],
+        "platforms": "all"
+      },
+      "applies_basis": "Artifact identity and startup checks are mandatory for every release type, preview included, on every shipped target (rustfs/backlog#2662 frozen design item 5)."
+    },
+    {
+      "lane": "upgrade",
+      "harness_repository": "rustfs/s3chaos",
+      "harness_sha": "9a48f55c1d6d58ce63a8fb1ea7317e024f32f11d",
+      "harness_basis": "Same fixed envelope producer as the identity lane; the workflow SHA and the product build SHA are bound separately and need not be equal.",
+      "timeout_seconds": 5400,
+      "environment_capabilities": ["process-exec", "previous-release-artifact"],
+      "cases": [
+        "direct-upgrade",
+        "mixed-version-upgrade",
+        "bucket-config-upgrade",
+        "bucket-config-rollback",
+        "multipart-layout-upgrade",
+        "multipart-layout-baseline"
+      ],
+      "cases_basis": "Case IDs of the pinned previous-release matrix in .github/workflows/e2e-upgrade.yml (source.version/asset/sha256 and the case list mapped there).",
+      "applies_to": {
+        "release_types": ["prerelease", "stable"],
+        "platforms": [{"os": "linux", "arch": "x86_64", "abi": "gnu"}]
+      },
+      "applies_basis": "preview releases use the pre-declared smaller set (identity lane only). The upgrade lane is only exercised on linux-x86_64-gnu, the platform of the pinned source assets."
+    }
+  ],
+  "upgrade_sources": {
+    "basis": "Exactly the two previous-release sources pinned in .github/workflows/e2e-upgrade.yml are exercised today (source.version / source.asset / source.sha256). No other N-1/N-2 compatibility is claimed, and these digests are not inferred from tags.",
+    "sources": [
+      {
+        "version": "1.0.1",
+        "role": "N-1",
+        "artifact_name": "rustfs-linux-x86_64-gnu-v1.0.1.zip",
+        "artifact_sha256": "9742e183cb0dfb0d48fa909797af572fe47b7f827569036baa4156e0d3958ee5"
+      },
+      {
+        "version": "1.0.0-rc.5",
+        "role": "N-2",
+        "artifact_name": "rustfs-linux-x86_64-gnu-v1.0.0-rc.5.zip",
+        "artifact_sha256": "3ee8df71e8edcfada533be452c4135868f697bc515460ae97b027313eade7a3d"
+      }
+    ],
+    "rollback_supported": false,
+    "rollback_basis": "Only the bucket-config-rollback case (rollback_to_previous_release_reads_current_bucket_metadata) is exercised. Full data-plane rollback is not certified and is never a default claim."
+  }
+}
@@ -296,6 +296,32 @@
      "origin": "baseline",
      "alert_contract": "scheduled-validation"
    },
+    {
+      "workflow": ".github/workflows/rustfs-validation-sweep.yml",
+      "name": "RustFS Validation Sweep",
+      "roles": [
+        "periodic",
+        "manual"
+      ],
+      "events": [
+        "schedule",
+        "workflow_dispatch"
+      ],
+      "cadence": "nightly",
+      "timezone": "UTC",
+      "schedule": [
+        "30 19 * * *"
+      ],
+      "owner": "@overtrue",
+      "policy": "active",
+      "monitor": "workflow",
+      "max_age_hours": 60,
+      "result_owner": "rustfs-validation-sweep.yml",
+      "disposition": "canonical validation cases executed in rotating batches",
+      "failure_owner": "workflow",
+      "origin": "baseline",
+      "alert_contract": "scheduled-validation"
+    },
    {
      "workflow": ".github/workflows/btf-prototype-reference.yml",
      "name": "BTF prototype reference",
@@ -814,13 +814,18 @@ jobs:
          export AWS_SECRET_ACCESS_KEY="$R2_SECRET_ACCESS_KEY"
          export AWS_DEFAULT_REGION="auto"

-          # Determine upload path based on build type
+          # Determine upload path based on build type. Release and prerelease
+          # builds only stage their bytes here: the official
+          # artifacts/rustfs/release/ path is written solely by the
+          # certification-gated promote-release-artifacts job, from these same
+          # staged objects. Development builds keep the uncertified dev channel
+          # and never touch official or latest channels.
          if [[ "$BUILD_TYPE" == "development" ]]; then
            R2_PATH="s3://${R2_BUCKET}/artifacts/rustfs/dev/"
            echo "📤 Uploading development build to R2 dev directory"
          else
-            R2_PATH="s3://${R2_BUCKET}/artifacts/rustfs/release/"
-            echo "📤 Uploading release build to R2 release directory"
+            R2_PATH="s3://${R2_BUCKET}/artifacts/rustfs/staging/${GITHUB_RUN_ID}/${GITHUB_RUN_ATTEMPT}/"
+            echo "📤 Staging release build for certification-gated promotion"
          fi

          # Upload all rustfs zip files to R2 using S3-compatible API
@@ -869,6 +874,221 @@ jobs:
            echo "This run does not trigger Docker publishing."
          fi

+  # Aggregate the parent candidate batch. This job always collects results and
+  # retains its evidence, including on failure, and every publication job
+  # below requires its success. Missing, stale, substituted or non-PASS
+  # evidence fails closed here: only the trusted verification sources
+  # (rustfs/s3chaos release-gate envelopes, rustfs/backlog#2658) can satisfy
+  # it, so a PR-uploaded JSON set never gains release authority.
+  release-certification:
+    name: Release Certification
+    needs: [ build-check, build-rustfs ]
+    if: ${{ !cancelled() && needs.build-check.outputs.should_publish == 'true' && startsWith(github.ref, 'refs/tags/') }}
+    runs-on: ubuntu-latest # gh CLI required; the sm-standard-2 fleet does not ship it
+    timeout-minutes: 30
+    permissions:
+      contents: read
+      actions: read
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          persist-credentials: false
+
+      - name: Download all build artifacts
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
+        with:
+          path: ./artifacts
+          pattern: rustfs-*
+          merge-multiple: true
+
+      - name: Write candidate and collect trusted validation results
+        env:
+          # Cross-repo read of the trusted verification run. Falls back to the
+          # workflow token, which cannot read rustfs/s3chaos; a missing token
+          # therefore fail-closes instead of certifying.
+          GH_TOKEN: ${{ secrets.RELEASE_VALIDATION_TOKEN || github.token }}
+          PRODUCT_COMMIT: ${{ github.sha }}
+          VERSION: ${{ needs.build-check.outputs.version }}
+          BUILD_TYPE: ${{ needs.build-check.outputs.build_type }}
+          FEATURES: ${{ needs.build-check.outputs.build_type == 'development' && '' || 'pyroscope' }}
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          mkdir -p results release-gate-output
+          PLAN_SHA256=$(python3 -c "import hashlib,json,sys; sys.path.insert(0,'scripts/release'); import release_gate; print(release_gate.canonical_sha256(json.load(open('.config/release-validation-plan.json'))))")
+
+          python3 - "$PLAN_SHA256" > candidate.json <<'PY'
+          import hashlib, json, os, sys, zipfile
+
+          plan_sha256 = sys.argv[1]
+          profile = {"release": "stable", "prerelease": "prerelease", "preview": "preview"}.get(os.environ["BUILD_TYPE"])
+          if profile is None:
+              raise SystemExit("release certification only covers preview, prerelease and stable builds")
+
+          def sha256(path):
+              digest = hashlib.sha256()
+              with open(path, "rb") as handle:
+                  for chunk in iter(lambda: handle.read(1024 * 1024), b""):
+                      digest.update(chunk)
+              return digest.hexdigest()
+
+          deliverables = []
+          for name in sorted(os.listdir("artifacts")):
+              if not name.startswith("rustfs-") or not name.endswith(".zip"):
+                  continue
+              path = os.path.join("artifacts", name)
+              members = zipfile.ZipFile(path).namelist()
+              binary_name = "rustfs.exe" if "rustfs.exe" in members else "rustfs"
+              with zipfile.ZipFile(path) as archive:
+                  binary_bytes = archive.read(binary_name)
+              import tempfile
+              with tempfile.NamedTemporaryFile() as handle:
+                  handle.write(binary_bytes)
+                  handle.flush()
+                  binary_sha = sha256(handle.name)
+              deliverables.append({
+                  "kind": "zip",
+                  "name": name,
+                  "location": "https://github.com/rustfs/rustfs/releases/download/{}/{}".format(os.environ["VERSION"], name),
+                  "sha256": sha256(path),
+                  "size": os.path.getsize(path),
+                  "binary_sha256": binary_sha,
+              })
+          if not deliverables:
+              raise SystemExit("no built deliverables found to certify")
+          candidate = {
+              "schema_version": 1,
+              "product_commit": os.environ["PRODUCT_COMMIT"],
+              "version": os.environ["VERSION"],
+              "release_profile": profile,
+              "features": [f for f in os.environ.get("FEATURES", "").split(",") if f],
+              "plan_sha256": plan_sha256,
+              "validation_batch": {"run_id": os.environ["GITHUB_RUN_ID"], "run_attempt": int(os.environ["GITHUB_RUN_ATTEMPT"])},
+              "deliverables": deliverables,
+          }
+          json.dump(candidate, sys.stdout, indent=2, sort_keys=True)
+          print()
+          PY
+
+          # Trusted verification evidence only: the release-gate envelopes of
+          # rustfs/s3chaos for this exact product commit. Absence blocks below.
+          TRUSTED_REPO="rustfs/s3chaos"
+          TRUSTED_WORKFLOW="release-gate.yml"
+          TRUSTED_RUN=$(gh api "repos/${TRUSTED_REPO}/actions/workflows/${TRUSTED_WORKFLOW}/runs?head_sha=${PRODUCT_COMMIT}&status=success&per_page=1" --jq '.workflow_runs[0].id // empty' || true)
+          if [[ -z "$TRUSTED_RUN" ]]; then
+            echo "❌ No trusted verification run of ${TRUSTED_REPO}/${TRUSTED_WORKFLOW} for ${PRODUCT_COMMIT}; refusing to certify" >&2
+          else
+            gh run download "$TRUSTED_RUN" -R "$TRUSTED_REPO" -n release-validation-evidence -D results || \
+              echo "❌ Trusted verification run ${TRUSTED_RUN} published no release-validation-evidence artifact" >&2
+          fi
+
+          python3 scripts/release/release_gate.py validate \
+            --plan .config/release-validation-plan.json \
+            --candidate candidate.json \
+            --results results \
+            --output release-gate-output
+
+      - name: Upload certification evidence
+        if: always()
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: release-certification
+          path: |
+            candidate.json
+            results/
+            release-gate-output/
+          if-no-files-found: warn
+          retention-days: 30
+
+  # The only writer of the official R2 release path. Promotion re-checks the
+  # exact bytes about to be published against the passed certification list
+  # and copies the staged objects byte-for-byte: no rebuild-after-verify.
+  promote-release-artifacts:
+    name: Promote Release Artifacts
+    needs: [ build-check, build-rustfs, release-certification ]
+    # Target-channel lock for the official R2 release path.
+    concurrency:
+      group: release-channel-r2-release
+      cancel-in-progress: false
+    if: needs.build-check.outputs.should_publish == 'true' && startsWith(github.ref, 'refs/tags/') && (needs.build-check.outputs.build_type == 'release' || needs.build-check.outputs.build_type == 'prerelease')
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    permissions:
+      contents: read
+      actions: read
+    env:
+      R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+      R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+      R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
+      R2_BUCKET: ${{ secrets.R2_BUCKET }}
+      AWS_EC2_METADATA_DISABLED: true
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          persist-credentials: false
+
+      - name: Download build and certification artifacts
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
+        with:
+          path: ./artifacts
+          pattern: rustfs-*
+          merge-multiple: true
+
+      - name: Download certification evidence
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
+        with:
+          name: release-certification
+          path: ./certification
+
+      - name: Verify and promote staged release artifacts
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          if [[ -z "$R2_ACCESS_KEY_ID" || -z "$R2_SECRET_ACCESS_KEY" || -z "$R2_ENDPOINT" || -z "$R2_BUCKET" ]]; then
+            echo "❌ R2 credentials missing; refusing to promote to the official channel" >&2
+            exit 1
+          fi
+
+          for zip_file in artifacts/rustfs-*.zip; do
+            [[ -f "$zip_file" ]] || continue
+            name="$(basename "$zip_file")"
+            python3 scripts/release/release_gate.py verify-promotion \
+              --plan .config/release-validation-plan.json \
+              --candidate certification/candidate.json \
+              --summary certification/release-gate-output \
+              --artifact "$zip_file" \
+              --artifact-name "$name"
+          done
+
+          if ! command -v aws >/dev/null 2>&1; then
+            sudo apt-get update && sudo apt-get install -y awscli
+          fi
+          export AWS_ACCESS_KEY_ID="$R2_ACCESS_KEY_ID"
+          export AWS_SECRET_ACCESS_KEY="$R2_SECRET_ACCESS_KEY"
+          export AWS_DEFAULT_REGION="auto"
+
+          STAGING_PATH="s3://${R2_BUCKET}/artifacts/rustfs/staging/${GITHUB_RUN_ID}/${GITHUB_RUN_ATTEMPT}/"
+          RELEASE_PATH="s3://${R2_BUCKET}/artifacts/rustfs/release/"
+          for zip_file in artifacts/rustfs-*.zip; do
+            [[ -f "$zip_file" ]] || continue
+            name="$(basename "$zip_file")"
+            staged="$(mktemp)"
+            aws s3 cp "${STAGING_PATH}${name}" "$staged" --endpoint-url "$R2_ENDPOINT" --only-show-errors
+            staged_sha="$(sha256sum "$staged" | awk '{print $1}')"
+            built_sha="$(sha256sum "$zip_file" | awk '{print $1}')"
+            if [[ "$staged_sha" != "$built_sha" ]]; then
+              echo "❌ Staged object ${name} does not match the verified bytes" >&2
+              exit 1
+            fi
+            aws s3 cp "$staged" "${RELEASE_PATH}${name}" --endpoint-url "$R2_ENDPOINT" --only-show-errors
+            rm -f "$staged"
+            echo "✅ Promoted ${name} (${built_sha})"
+          done
+
  # Create GitHub Release for every valid release tag, including previews
  create-release:
    name: Create GitHub Release
@@ -1028,6 +1248,10 @@ jobs:
  update-latest-version:
    name: Update Latest Version
    needs: [ build-check, publish-release ]
+    # Target-channel lock for the latest.json version pointer.
+    concurrency:
+      group: release-channel-oss-latest
+      cancel-in-progress: false
    if: needs.build-check.outputs.should_publish == 'true' && startsWith(github.ref, 'refs/tags/') && (needs.build-check.outputs.build_type == 'release' || needs.build-check.outputs.build_type == 'prerelease')
    runs-on: ubuntu-latest
    timeout-minutes: 30
@@ -1068,6 +1292,45 @@ jobs:
          rm -rf "$OSSUTIL_DIR" "$OSSUTIL_ZIP"
          chmod +x "$OSSUTIL_BIN"

+          # Candidate/pointer ordering: an older candidate must never overwrite
+          # a newer pointer, even if its promotion job finishes later.
+          CURRENT_LATEST="${RUNNER_TEMP}/current-latest.json"
+          if "$OSSUTIL_BIN" cp oss://rustfs-version/latest.json "$CURRENT_LATEST" 2>/dev/null; then
+            python3 - "$CURRENT_LATEST" "$VERSION" <<'PY'
+          import json
+          import re
+          import sys
+
+          def key(version):
+              match = re.match(r"^(\d+)\.(\d+)\.(\d+)(?:-([0-9A-Za-z.-]+))?$", version)
+              if not match:
+                  raise SystemExit(3)
+              core = tuple(int(part) for part in match.groups()[:3])
+              prerelease = match.group(4)
+              if prerelease is None:
+                  return core + (1,)
+              return core + (0, tuple(int(p) if p.isdigit() else p for p in re.split(r"[.-]", prerelease)))
+
+          try:
+              current = json.load(open(sys.argv[1], encoding="utf-8")).get("version", "")
+              proposed = sys.argv[2]
+              if current and key(current) > key(proposed):
+                  raise SystemExit(1)
+          except SystemExit:
+              raise
+          except (OSError, ValueError, TypeError):
+              raise SystemExit(3)
+          PY
+            guard_status=$?
+            if [[ "$guard_status" -eq 1 ]]; then
+              echo "Current latest.json points at a newer version; refusing to downgrade" >&2
+              exit 0
+            elif [[ "$guard_status" -ne 0 ]]; then
+              echo "Existing latest.json is unparsable; refusing to overwrite it blindly" >&2
+              exit 1
+            fi
+          fi
+
          # Create latest.json
          cat > latest.json << EOF
          {
@@ -1084,16 +1347,59 @@ jobs:

          echo "✅ Updated latest.json for ${RELEASE_TYPE} release $VERSION"

-  # Publish release (remove draft status)
+  # Publish release (remove draft status). Requires the release-certification
+  # aggregate: without a complete legal success there is no public write.
  publish-release:
    name: Publish Release
-    needs: [ build-check, create-release, upload-release-assets ]
+    needs: [ build-check, create-release, upload-release-assets, release-certification ]
+    # Target-channel lock: concurrent releases serialize on this channel so an
+    # older candidate cannot interleave its publish with a newer one.
+    concurrency:
+      group: release-channel-github-release
+      cancel-in-progress: false
    if: needs.build-check.outputs.should_publish == 'true' && startsWith(github.ref, 'refs/tags/') && (needs.build-check.outputs.build_type == 'preview' || needs.build-check.outputs.build_type == 'release' || needs.build-check.outputs.build_type == 'prerelease')
    runs-on: ubuntu-latest # gh CLI required; the sm-standard-2 fleet does not ship it
    timeout-minutes: 30
    permissions:
      contents: write
+      actions: read
    steps:
+      - name: Checkout repository
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          persist-credentials: false
+
+      - name: Download certification evidence
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
+        with:
+          name: release-certification
+          path: ./certification
+
+      - name: Verify release assets against certification
+        env:
+          GH_TOKEN: ${{ github.token }}
+        shell: bash
+        run: |
+          set -euo pipefail
+          TAG="${{ needs.build-check.outputs.version }}"
+          mkdir -p "$RUNNER_TEMP/promotion-assets"
+          gh release download "$TAG" --dir "$RUNNER_TEMP/promotion-assets" --pattern '*.zip'
+          found=0
+          for zip_file in "$RUNNER_TEMP"/promotion-assets/*.zip; do
+            [[ -f "$zip_file" ]] || continue
+            found=1
+            python3 scripts/release/release_gate.py verify-promotion \
+              --plan .config/release-validation-plan.json \
+              --candidate certification/candidate.json \
+              --summary certification/release-gate-output \
+              --artifact "$zip_file" \
+              --artifact-name "$(basename "$zip_file")"
+          done
+          if [[ "$found" -ne 1 ]]; then
+            echo "No release zip assets to verify against the certification" >&2
+            exit 1
+          fi
+
      - name: Publish release
        env:
          GH_TOKEN: ${{ github.token }}
@@ -96,7 +96,7 @@ env:
  RUST_BACKTRACE: 1

 jobs:
-  # Readers: test-and-lint (workspace and ILM), build-rustfs-debug-binary,
+  # Readers: workspace-lint-and-build, workspace-tests (workspace and ILM), build-rustfs-debug-binary,
  # e2e-tests, e2e-full.
  warm-ci-dev:
    name: Warm ci-dev
@@ -146,7 +146,7 @@ jobs:
          if-no-files-found: error

      # --all-targets covers the test binaries nextest builds, including
-      # e2e_test, which test-and-lint's own run excludes. The second build adds
+      # e2e_test, which the workspace run excludes. The second build adds
      # the e2e-test-hooks feature resolution that build-rustfs-debug-binary uses
      # and that no lint lane enables.
      - name: Build ci-dev superset
@@ -113,7 +113,7 @@ jobs:
            printf '%s\n' 'mode=full' >> "$GITHUB_OUTPUT"
            echo "Base CI policy unavailable; running the full matrix."
          fi
-          if ! grep -q '^plan=' "$GITHUB_OUTPUT" || ! grep -q '^rehearse_shards=' "$GITHUB_OUTPUT"; then
+          if ! grep -q '^plan=' "$GITHUB_OUTPUT" || ! grep -q '^plan_contract=2$' "$GITHUB_OUTPUT"; then
            printf '%s\n' 'mode=full' 'connect=true' 'full_validation=true' 'rehearse_shards=false' 'packages=' 'plan=' >> "$GITHUB_OUTPUT"
            echo "Base policy does not cover this workflow contract; retaining the full matrix."
          fi
@@ -160,7 +160,7 @@ jobs:
      - name: Run shared quick checks
        uses: ./.github/actions/quick-checks

-  test-and-lint:
+  workspace-lint-and-build:
    name: Workspace Lint and Build
    if: contains(fromJSON('["full", "e2e"]'), needs.classify-changes.outputs.mode) && (github.event_name != 'pull_request' || github.event.action != 'closed')
    needs: [ quick-checks, classify-changes ]
@@ -193,13 +193,13 @@ jobs:

      - name: Prepare test evidence
        run: |
-          mkdir -p artifacts/test-and-lint
+          mkdir -p artifacts/workspace
          {
            echo "run_id=${GITHUB_RUN_ID}"
            echo "job=${GITHUB_JOB}"
            echo "runner=${RUNNER_NAME}"
            echo "started_at=$(date --utc --iso-8601=seconds)"
-          } > artifacts/test-and-lint/run-metadata.txt
+          } > artifacts/workspace/run-metadata.txt

      # Clippy gates every event, pull requests included. When it ran only after
      # merge, a lint regression merged freely and then failed every main run
@@ -233,11 +233,11 @@ jobs:

      - name: Run documentation tests
        run: |
-          mkdir -p artifacts/test-and-lint
+          mkdir -p artifacts/workspace
          set +e
          timeout --verbose --signal=TERM --kill-after=30s 15m \
            python3 scripts/ci_test_artifact.py cargo doc --configuration workspace \
-            2>&1 | tee artifacts/test-and-lint/doctest.log
+            2>&1 | tee artifacts/workspace/doctest.log
          status=${PIPESTATUS[0]}
          {
            echo "command=cargo test --all --exclude e2e_test --doc"
@@ -246,7 +246,7 @@ jobs:
            echo
            echo "Remaining test-related processes:"
            pgrep -af 'cargo|rustdoc|target/.*/deps/' || true
-          } > artifacts/test-and-lint/doctest-diagnostics.txt
+          } > artifacts/workspace/doctest-diagnostics.txt
          exit "${status}"

      - name: Upload test reports and diagnostics
@@ -255,7 +255,7 @@ jobs:
        with:
          name: junit-test-and-lint-${{ github.run_number }}-${{ github.run_attempt }}
          path: |
-            artifacts/test-and-lint
+            artifacts/workspace
          retention-days: 3
          if-no-files-found: error

@@ -268,7 +268,7 @@ jobs:

  workspace-tests:
    name: Workspace tests
-    needs: [ test-and-lint, classify-changes ]
+    needs: [ workspace-lint-and-build, classify-changes ]
    # The archive retains compile-time fixture paths and the large test set
    # requires the same runner image until the hosted-runtime pilot passes.
    runs-on: sm-standard-4
@@ -293,14 +293,14 @@ jobs:

      - name: Extract and list workspace tests
        run: |
-          mkdir -p artifacts/test-and-lint
+          mkdir -p artifacts/workspace
          cargo nextest list --profile ci --archive-file "$RUNNER_TEMP/workspace-tests.tar.zst" --workspace-remap "$GITHUB_WORKSPACE" --extract-to "$GITHUB_WORKSPACE" --extract-overwrite --message-format json \
-            > artifacts/test-and-lint/core-test-listing.json
-          python3 scripts/check_test_wiring.py --check-core artifacts/test-and-lint/core-test-listing.json
+            > artifacts/workspace/core-test-listing.json
+          python3 scripts/check_test_wiring.py --check-core artifacts/workspace/core-test-listing.json

      - name: Run nextest tests
        run: |
-          mkdir -p artifacts/test-and-lint
+          mkdir -p artifacts/workspace
          rm -f target/nextest/ci/junit.xml
          ./scripts/ci/resource_sampler.sh start nextest
          trap './scripts/ci/resource_sampler.sh stop' EXIT
@@ -309,7 +309,7 @@ jobs:
            cargo nextest run --profile ci --cargo-metadata target/nextest/cargo-metadata.json --binaries-metadata target/nextest/binaries-metadata.json \
            --workspace-remap "$GITHUB_WORKSPACE" --target-dir-remap "$GITHUB_WORKSPACE/target" \
            --status-level all --final-status-level all \
-            2>&1 | tee artifacts/test-and-lint/nextest.log
+            2>&1 | tee artifacts/workspace/nextest.log
          status=${PIPESTATUS[0]}
          {
            echo "command=workspace archive execution"
@@ -321,13 +321,13 @@ jobs:
            echo
            echo "Kernel OOM / kill events:"
            dmesg -T 2>/dev/null | grep -iE 'oom|out of memory|killed process' | tail -20 || true
-          } > artifacts/test-and-lint/nextest-diagnostics.txt
+          } > artifacts/workspace/nextest-diagnostics.txt
          exit "${status}"

      - name: Verify workspace execution completeness
        run: |
-          python3 scripts/ci_test_artifact.py evidence artifacts/test-and-lint/core-test-listing.json target/nextest/ci/junit.xml
-          ./scripts/check_migration_gate_count.sh evidence artifacts/test-and-lint/core-test-listing.json target/nextest/ci/junit.xml
+          python3 scripts/ci_test_artifact.py evidence artifacts/workspace/core-test-listing.json target/nextest/ci/junit.xml
+          ./scripts/check_migration_gate_count.sh evidence artifacts/workspace/core-test-listing.json target/nextest/ci/junit.xml

      - name: Run ignored ILM integration tests serially
        if: needs.classify-changes.outputs.mode == 'full'
@@ -358,7 +358,7 @@ jobs:
          name: workspace-test-evidence-${{ github.run_attempt }}
          path: |
            target/nextest/ci/junit.xml
-            artifacts/test-and-lint
+            artifacts/workspace
            artifacts/ilm-integration
          retention-days: 3
          if-no-files-found: error
@@ -368,7 +368,7 @@ jobs:
        if: needs.classify-changes.outputs.rehearse_shards == 'true'
        run: |
          python3 scripts/ci_test_shards.py prepare \
-            --listing artifacts/test-and-lint/core-test-listing.json \
+            --listing artifacts/workspace/core-test-listing.json \
            --junit target/nextest/ci/junit.xml --output artifacts/workspace-shard-plan
      - name: Upload same-run shard plan
        if: needs.classify-changes.outputs.rehearse_shards == 'true'
@@ -449,7 +449,7 @@ jobs:

  connect-feature-boundaries:
    name: Connect Feature Boundaries
-    if: needs.classify-changes.outputs.mode == 'full' && needs.classify-changes.outputs.connect != 'false' && (github.event_name != 'pull_request' || github.event.action != 'closed')
+    if: needs.classify-changes.outputs.mode == 'full' && needs.classify-changes.outputs.connect != 'false' && contains(fromJSON('["schedule", "workflow_dispatch"]'), github.event_name)
    needs: [ quick-checks, classify-changes ]
    runs-on: sm-standard-4
    timeout-minutes: 90
@@ -499,11 +499,13 @@ jobs:

  test-and-lint-rio-v2:
    name: Test and Lint (rio-v2)
-    # Alternate feature configurations run on the nightly schedule, merge
-    # groups and manual dispatch. The default workspace remains the PR gate and
-    # the per-push main baseline: repeating four feature builds on every merge
-    # bought nothing the nightly run does not report within a day.
-    if: needs.classify-changes.outputs.mode == 'full' && contains(fromJSON('["merge_group", "schedule", "workflow_dispatch"]'), github.event_name)
+    # Specialized boundary, hardware and alternate-feature lanes run only on
+    # the nightly schedule and manual dispatch. They do not gate a merge: the
+    # nightly run reports a regression within a day, and repeating these builds
+    # on every PR revision and merge group bought no earlier signal worth the
+    # runner occupancy. Workspace lint and tests, the core contract, the debug
+    # build, E2E smoke and the S3 lanes remain the pull-request gate.
+    if: needs.classify-changes.outputs.mode == 'full' && contains(fromJSON('["schedule", "workflow_dispatch"]'), github.event_name)
    needs: [ quick-checks, classify-changes ]
    runs-on: sm-standard-4
    timeout-minutes: 90
@@ -540,12 +542,12 @@ jobs:
          # --profile ci so the quarantine list (and its junit flaky markers)
          # covers this leg too; the default profile is the local no-retry
          # profile and silently ignored quarantined flakes here (rustfs#6703).
-          mkdir -p artifacts/test-and-lint
+          mkdir -p artifacts/workspace
          rm -f target/nextest/ci/junit.xml
          ./scripts/ci/resource_sampler.sh start nextest-rio-v2
          trap './scripts/ci/resource_sampler.sh stop' EXIT
          cargo nextest run --profile ci -p rustfs -p rustfs-ecstore --features rio-v2 \
-            2>&1 | tee artifacts/test-and-lint/nextest.log
+            2>&1 | tee artifacts/workspace/nextest.log
          cargo test -p rustfs --doc --features rio-v2

      - name: Upload rio-v2 test reports and diagnostics
@@ -555,14 +557,14 @@ jobs:
          name: junit-test-and-lint-rio-v2-${{ github.run_number }}-${{ github.run_attempt }}
          path: |
            target/nextest/ci/junit.xml
-            artifacts/test-and-lint
+            artifacts/workspace
          retention-days: 3
          if-no-files-found: error

  test-and-lint-protocols:
    name: "Test and Lint (${{ matrix.features.name }})"
    # Same cadence as test-and-lint-rio-v2 above.
-    if: needs.classify-changes.outputs.mode == 'full' && contains(fromJSON('["merge_group", "schedule", "workflow_dispatch"]'), github.event_name)
+    if: needs.classify-changes.outputs.mode == 'full' && contains(fromJSON('["schedule", "workflow_dispatch"]'), github.event_name)
    needs: [ quick-checks, classify-changes ]
    runs-on: sm-standard-4
    timeout-minutes: 90
@@ -631,7 +633,7 @@ jobs:
          path: |
            target/nextest/ci/junit.xml
            artifacts/protocol-tests
-            artifacts/test-and-lint
+            artifacts/workspace
          retention-days: 3
          if-no-files-found: warn

@@ -783,7 +785,7 @@ jobs:
    # job had neither, so each closed/merged PR really ran the whole io_uring
    # suite (measured 4m17s / 7m19s / 7m31s on runs 30678272341 / 30678117601 /
    # 30662728539) and kept the cancellation run in progress for minutes.
-    if: needs.classify-changes.outputs.mode == 'full' && (github.event_name != 'pull_request' || github.event.action != 'closed')
+    if: needs.classify-changes.outputs.mode == 'full' && contains(fromJSON('["schedule", "workflow_dispatch"]'), github.event_name)
    needs: [ quick-checks, classify-changes ]
    # GitHub-hosted ubuntu-latest runs a recent kernel with io_uring and, unlike
    # a container, applies no seccomp filter that would block io_uring_setup — so
@@ -978,7 +980,7 @@ jobs:
          path: |
            ${{ runner.temp }}/rustfs-e2e-smoke-logs/
            ${{ runner.temp }}/rustfs-e2e-smoke-list.json
-            artifacts/test-and-lint
+            artifacts/workspace
          if-no-files-found: warn

      - name: Upload e2e smoke JUnit report
@@ -1389,7 +1391,7 @@ jobs:
      - classify-changes
      - typos
      - quick-checks
-      - test-and-lint
+      - workspace-lint-and-build
      - workspace-tests
      - workspace-shards
      - workspace-shard-evidence
@@ -1425,7 +1427,7 @@ jobs:
      - required-checks
      - typos
      - quick-checks
-      - test-and-lint
+      - workspace-lint-and-build
      - workspace-tests
      - core-contract
      - test-and-lint-rio-v2
@@ -329,13 +329,75 @@ jobs:
          echo "  - Create latest: $create_latest"
          echo "  - Source ref: $source_ref"

+  # Consume the trusted release certification for the exact version this
+  # workflow wraps and verify the certified input digests. A manual dispatch
+  # without that certification fails here and pushes no official tag; only
+  # results from the trusted verification sources carry release authority.
+  release-certification:
+    name: Release Certification
+    needs: build-check
+    if: needs.build-check.outputs.should_build == 'true'
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    permissions:
+      contents: read
+      actions: read
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          persist-credentials: false
+          ref: ${{ needs.build-check.outputs.source_ref }}
+
+      - name: Download and verify release certification
+        env:
+          GH_TOKEN: ${{ secrets.RELEASE_VALIDATION_TOKEN || github.token }}
+          TRIGGERING_RUN_ID: ${{ github.event.workflow_run.id }}
+          VERSION: ${{ needs.build-check.outputs.version }}
+        shell: bash
+        run: |
+          set -euo pipefail
+          mkdir -p certification
+          run_id="$TRIGGERING_RUN_ID"
+          if [[ -z "$run_id" ]]; then
+            run_id="$(gh api --method GET \
+              "repos/${GITHUB_REPOSITORY}/actions/workflows/build.yml/runs" \
+              -f branch="$VERSION" -f status=success -F per_page=1 \
+              --jq '.workflow_runs[0].id // empty' || true)"
+          fi
+          if [[ -z "$run_id" ]]; then
+            echo "No certified Build and Release run found for ${VERSION}; failing closed" >&2
+            exit 1
+          fi
+          if ! gh run download "$run_id" -n release-certification -D certification; then
+            echo "Build run ${run_id} published no release-certification artifact; failing closed" >&2
+            exit 1
+          fi
+          python3 scripts/release/release_gate.py verify-promotion \
+            --plan .config/release-validation-plan.json \
+            --candidate certification/candidate.json \
+            --summary certification/release-gate-output \
+            --artifact-sha256 "$(jq -r '.deliverables[0].sha256' certification/candidate.json)" \
+            --artifact-name "$(jq -r '.deliverables[0].name' certification/candidate.json)"
+
+          # Input digest binding: every zip this image wraps must be the
+          # certified byte-for-byte content, never a mutable latest pointer.
+          while IFS=$'\t' read -r name location expected; do
+            echo "Verifying certified input ${name}"
+            actual="$(curl -fsSL "$location" | sha256sum | awk '{print $1}')"
+            if [[ "$actual" != "$expected" ]]; then
+              echo "Certified input ${name} does not match its recorded digest" >&2
+              exit 1
+            fi
+          done < <(jq -r '.deliverables[] | select(.name | contains("linux")) | [.name, .location, .sha256] | @tsv' certification/candidate.json)
+
  # Build multi-arch Docker images
  # Strategy: Build images using pre-built binaries from dl.rustfs.com
  # Supports both release and dev channel binaries based on build context
  # Only runs when should_build is true (which includes workflow success check)
  build-docker:
    name: Build Docker Images
-    needs: build-check
+    needs: [ build-check, release-certification ]
    if: needs.build-check.outputs.should_build == 'true'
    runs-on: ubuntu-latest
    timeout-minutes: 60
@@ -443,6 +505,14 @@ jobs:
          # Output tags
          echo "tags=$TAGS" >> "$GITHUB_OUTPUT"

+          # Staging identity for this run's promotion: official tags are only
+          # created after the staged image verifies against the certification,
+          # and they point at this same digest (no rebuild-after-verify).
+          STAGING_TAG="staging-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}${VARIANT_SUFFIX}"
+          STAGING_TAGS="${{ env.REGISTRY_DOCKERHUB }}:${STAGING_TAG},${{ env.REGISTRY_GHCR }}:${STAGING_TAG},${{ env.REGISTRY_QUAY }}:${STAGING_TAG}"
+          echo "staging_tags=$STAGING_TAGS" >> "$GITHUB_OUTPUT"
+          echo "staging_tag=$STAGING_TAG" >> "$GITHUB_OUTPUT"
+
          # Generate labels
          LABELS="org.opencontainers.image.title=RustFS"
          LABELS="$LABELS,org.opencontainers.image.description=RustFS distributed object storage system"
@@ -467,7 +537,9 @@ jobs:
          file: ${{ matrix.file }}
          platforms: ${{ env.DOCKER_PLATFORMS }}
          push: ${{ needs.build-check.outputs.should_push == 'true' }}
-          tags: ${{ steps.meta.outputs.tags }}
+          # Push to staging tags only: official tags are created by the
+          # certification-gated promotion step below from this same digest.
+          tags: ${{ steps.meta.outputs.staging_tags }}
          labels: ${{ steps.meta.outputs.labels }}
          # No layer cache. This build compiles nothing — it downloads a
          # release zip and runs apk/apt — so the cache could only save the
@@ -493,6 +565,114 @@ jobs:
          no-cache: false
          pull: true

+      # A later-built wrapper: the staged image must complete this platform's
+      # own binary identity verification before any official tag is written.
+      # It may not inherit an unverified package's conclusion.
+      - name: Verify staged image against certification
+        if: needs.build-check.outputs.should_push == 'true'
+        env:
+          GH_TOKEN: ${{ secrets.RELEASE_VALIDATION_TOKEN || github.token }}
+          TRIGGERING_RUN_ID: ${{ github.event.workflow_run.id }}
+          VERSION: ${{ needs.build-check.outputs.version }}
+        shell: bash
+        run: |
+          set -euo pipefail
+          mkdir -p certification
+          run_id="$TRIGGERING_RUN_ID"
+          if [[ -z "$run_id" ]]; then
+            run_id="$(gh api --method GET \
+              "repos/${GITHUB_REPOSITORY}/actions/workflows/build.yml/runs" \
+              -f branch="$VERSION" -f status=success -F per_page=1 \
+              --jq '.workflow_runs[0].id // empty' || true)"
+          fi
+          gh run download "$run_id" -n release-certification -D certification
+
+          STAGING_REF="${{ env.REGISTRY_DOCKERHUB }}:staging-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}${{ matrix.suffix }}"
+          case "${{ matrix.variant }}" in
+            musl) libc="musl" ;;
+            glibc) libc="gnu" ;;
+          esac
+          for platform in linux/amd64 linux/arm64; do
+            case "$platform" in
+              linux/amd64) arch_sub="x86_64" ;;
+              linux/arm64) arch_sub="aarch64" ;;
+            esac
+            docker pull --platform "$platform" "$STAGING_REF"
+            container="$(docker create --platform "$platform" "$STAGING_REF")"
+            docker cp "$container:/usr/bin/rustfs" "$RUNNER_TEMP/staged-rustfs"
+            docker rm "$container" >/dev/null
+            actual="$(sha256sum "$RUNNER_TEMP/staged-rustfs" | awk '{print $1}')"
+            expected="$(jq -r --arg needle "linux-${arch_sub}-${libc}" \
+              '[.deliverables[] | select(.name | contains($needle)) | .binary_sha256][0] // empty' \
+              certification/candidate.json)"
+            if [[ -z "$expected" || "$actual" != "$expected" ]]; then
+              echo "Staged image binary for ${platform} does not match the certified binary digest" >&2
+              exit 1
+            fi
+            python3 scripts/release/release_gate.py verify-promotion \
+              --plan .config/release-validation-plan.json \
+              --candidate certification/candidate.json \
+              --summary certification/release-gate-output \
+              --artifact "$RUNNER_TEMP/staged-rustfs" \
+              --artifact-role binary
+          done
+
+      - name: Promote staged image to official tags
+        if: needs.build-check.outputs.should_push == 'true'
+        shell: bash
+        run: |
+          set -euo pipefail
+          STAGING_REF="${{ env.REGISTRY_DOCKERHUB }}:staging-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}${{ matrix.suffix }}"
+          IFS=',' read -r -a official_tags <<< "${{ steps.meta.outputs.tags }}"
+          for tag in "${official_tags[@]}"; do
+            docker buildx imagetools create -t "$tag" "$STAGING_REF"
+            echo "Promoted $tag -> $STAGING_REF"
+          done
+
+      # Downstream consumers (Helm) verify this record instead of inheriting
+      # an unverified wrapper's conclusion.
+      - name: Record image promotion
+        if: needs.build-check.outputs.should_push == 'true'
+        shell: bash
+        env:
+          GH_TOKEN: ${{ secrets.RELEASE_VALIDATION_TOKEN || github.token }}
+          TRIGGERING_RUN_ID: ${{ github.event.workflow_run.id }}
+          VERSION: ${{ needs.build-check.outputs.version }}
+        run: |
+          set -euo pipefail
+          mkdir -p certification
+          run_id="$TRIGGERING_RUN_ID"
+          if [[ -z "$run_id" ]]; then
+            run_id="$(gh api --method GET \
+              "repos/${GITHUB_REPOSITORY}/actions/workflows/build.yml/runs" \
+              -f branch="$VERSION" -f status=success -F per_page=1 \
+              --jq '.workflow_runs[0].id // empty' || true)"
+          fi
+          gh run download "$run_id" -n release-certification -D certification
+          STAGING_REF="${{ env.REGISTRY_DOCKERHUB }}:staging-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}${{ matrix.suffix }}"
+          digest="$(docker buildx imagetools inspect "$STAGING_REF" --format '{{json .Manifest}}' | jq -r '.digest')"
+          candidate_sha="$(python3 -c "import json,hashlib,sys; sys.path.insert(0,'scripts/release'); import release_gate; print(release_gate.canonical_sha256(json.load(open('certification/candidate.json'))))")"
+          jq -n \
+            --arg version "$VERSION" \
+            --arg digest "$digest" \
+            --arg staging_ref "$STAGING_REF" \
+            --arg candidate_sha256 "$candidate_sha" \
+            --arg tags "${{ steps.meta.outputs.tags }}" \
+            --slurpfile candidate certification/candidate.json \
+            '{schema_version: 1, version: $version, image_digest: $digest, staging_ref: $staging_ref,
+              promoted_tags: ($tags | split(",")), candidate_sha256: $candidate_sha256,
+              binary_sha256s: ($candidate[0].deliverables | map(.binary_sha256) | unique)}' \
+            > image-promotion-record.json
+
+      - name: Upload image promotion record
+        if: needs.build-check.outputs.should_push == 'true'
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: image-promotion-record${{ matrix.suffix }}
+          path: image-promotion-record.json
+          if-no-files-found: error
+          retention-days: 30
+
  # Note: Manifest creation is no longer needed as we only build one variant
  # Multi-arch manifests are automatically created by docker/build-push-action

@@ -30,9 +30,102 @@ permissions:
  contents: read

 jobs:
+  # Consume the same trusted release certification as every other publication
+  # exit and verify the image promotion record of the Docker workflow. The
+  # chart wraps no binary of its own, so it may only reference an image whose
+  # platform verification already passed; it must not inherit an unverified
+  # package's conclusion. A run without that evidence fails closed here and
+  # never reaches the chart repository.
+  release-certification:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    permissions:
+      contents: read
+      actions: read
+    outputs:
+      version: ${{ steps.version.outputs.raw_tag }}
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          persist-credentials: false
+
+      - name: Normalize release version
+        id: version
+        env:
+          RAW_INPUT: ${{ github.event.inputs.version }}
+          RAW_BRANCH: ${{ github.event.workflow_run.head_branch }}
+        run: |
+          set -euo pipefail
+          if [ "${{ github.event_name }}" = "workflow_dispatch" ]; then
+            RAW="$RAW_INPUT"
+          else
+            RAW="$RAW_BRANCH"
+          fi
+          case "$RAW" in
+            refs/tags/*) RAW_TAG="${RAW#refs/tags/}" ;;
+            *) RAW_TAG="$RAW" ;;
+          esac
+          echo "raw_tag=$RAW_TAG" >> "$GITHUB_OUTPUT"
+
+      - name: Download and verify release certification and image record
+        env:
+          GH_TOKEN: ${{ secrets.RELEASE_VALIDATION_TOKEN || github.token }}
+          TRIGGERING_RUN_ID: ${{ github.event.workflow_run.id }}
+          RAW_TAG: ${{ steps.version.outputs.raw_tag }}
+        shell: bash
+        run: |
+          set -euo pipefail
+          mkdir -p certification image-record
+
+          build_run="$TRIGGERING_RUN_ID"
+          if [[ -z "$build_run" ]]; then
+            build_run="$(gh api --method GET \
+              "repos/${GITHUB_REPOSITORY}/actions/workflows/build.yml/runs" \
+              -f branch="$RAW_TAG" -f status=success -F per_page=1 \
+              --jq '.workflow_runs[0].id // empty' || true)"
+          fi
+          if [[ -z "$build_run" ]]; then
+            echo "No certified Build and Release run found for ${RAW_TAG}; failing closed" >&2
+            exit 1
+          fi
+          if ! gh run download "$build_run" -n release-certification -D certification; then
+            echo "Build run ${build_run} published no release-certification artifact; failing closed" >&2
+            exit 1
+          fi
+          python3 scripts/release/release_gate.py verify-promotion \
+            --plan .config/release-validation-plan.json \
+            --candidate certification/candidate.json \
+            --summary certification/release-gate-output \
+            --artifact-sha256 "$(jq -r '.deliverables[0].sha256' certification/candidate.json)" \
+            --artifact-name "$(jq -r '.deliverables[0].name' certification/candidate.json)"
+
+          docker_run="$(gh api --method GET \
+            "repos/${GITHUB_REPOSITORY}/actions/workflows/docker.yml/runs" \
+            -f branch="$RAW_TAG" -f status=success -F per_page=1 \
+            --jq '.workflow_runs[0].id // empty' || true)"
+          if [[ -z "$docker_run" ]] || ! gh run download "$docker_run" -n image-promotion-record -D image-record; then
+            echo "No image promotion record for ${RAW_TAG}; a chart may not wrap an unverified image" >&2
+            exit 1
+          fi
+          candidate_sha="$(python3 -c "import json,hashlib,sys; sys.path.insert(0,'scripts/release'); import release_gate; print(release_gate.canonical_sha256(json.load(open('certification/candidate.json'))))")"
+          record_sha="$(jq -r '.candidate_sha256 // empty' image-record/image-promotion-record.json)"
+          if [[ "$record_sha" != "$candidate_sha" ]]; then
+            echo "Image promotion record does not belong to this certification" >&2
+            exit 1
+          fi
+          overlap="$(jq -r --slurpfile candidate certification/candidate.json \
+            '[.binary_sha256s[]] - ($candidate[0].deliverables | map(.binary_sha256)) | length' \
+            image-record/image-promotion-record.json)"
+          if [[ "$overlap" -ne 0 ]]; then
+            echo "Image promotion record carries binary digests outside the certification" >&2
+            exit 1
+          fi
+
  build-helm-package:
    runs-on: ubuntu-latest
    timeout-minutes: 30
+    needs: [ release-certification ]
    if: |
      (github.event_name == 'workflow_dispatch' && !contains(github.event.inputs.version, '-preview')) ||
      (
@@ -225,10 +225,57 @@ jobs:
          echo "  Build run ID: $BUILD_RUN_ID"
          echo "  Build run number: $RUN_NUMBER"

+  # Consume the trusted release certification of the resolved build run and
+  # verify the certified input digests before any package is wrapped. A run
+  # without that certification fails closed here; packages built later must
+  # complete their own verification and cannot inherit an unverified
+  # conclusion.
+  release-certification:
+    name: Release Certification
+    needs: resolve
+    if: needs.resolve.result == 'success'
+    runs-on: ubuntu-latest # gh CLI required; the sm-standard-2 fleet does not ship it
+    timeout-minutes: 30
+    permissions:
+      contents: read
+      actions: read
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          persist-credentials: false
+
+      - name: Download and verify release certification
+        env:
+          GH_TOKEN: ${{ secrets.RELEASE_VALIDATION_TOKEN || github.token }}
+          BUILD_RUN_ID: ${{ needs.resolve.outputs.build_run_id }}
+          BUILD_TYPE: ${{ needs.resolve.outputs.build_type }}
+        shell: bash
+        run: |
+          set -euo pipefail
+          mkdir -p certification
+          if [[ "$BUILD_TYPE" == "development" ]]; then
+            # Development artifacts are explicitly uncertified and never touch
+            # official or latest channels; only the staging/dev paths remain.
+            jq -n '{schema_version: 1, certified: false, reason: "development artifacts are explicitly uncertified"}' \
+              > certification/development-uncertified.json
+            exit 0
+          fi
+          if ! gh run download "$BUILD_RUN_ID" -n release-certification -D certification; then
+            echo "Build run ${BUILD_RUN_ID} published no release-certification artifact; failing closed" >&2
+            exit 1
+          fi
+          python3 scripts/release/release_gate.py verify-promotion \
+            --plan .config/release-validation-plan.json \
+            --candidate certification/candidate.json \
+            --summary certification/release-gate-output \
+            --artifact-sha256 "$(jq -r '.deliverables[0].sha256' certification/candidate.json)" \
+            --artifact-name "$(jq -r '.deliverables[0].name' certification/candidate.json)"
+
  # Build DEB and RPM packages for each architecture and libc variant
  package:
    name: Package (${{ matrix.arch }}-${{ matrix.libc }})
-    needs: resolve
+    needs: [ resolve, release-certification ]
    runs-on: ubuntu-latest # gh CLI required; the sm-standard-2 fleet does not ship it
    timeout-minutes: 30
    strategy:
@@ -494,6 +541,55 @@ jobs:
          echo "rpm_file=$RPM_FILE" >> "$GITHUB_OUTPUT"
          echo "✅ RPM built: $RPM_FILE"

+      - name: Download certification evidence
+        if: needs.resolve.outputs.build_type != 'development'
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
+        with:
+          name: release-certification
+          path: ./certification
+
+      # Later-built wrapper verification: the binary inside each package must
+      # match the certified binary digest for this platform before anything is
+      # uploaded, and the package itself stays on staging until promotion.
+      - name: Verify wrapped binary against certification
+        if: needs.resolve.outputs.build_type != 'development'
+        shell: bash
+        env:
+          DEB_FILE: ${{ steps.deb.outputs.deb_file }}
+          RPM_FILE: ${{ steps.rpm.outputs.rpm_file }}
+          ARTIFACT_NAME: ${{ matrix.artifact_name }}
+        run: |
+          set -euo pipefail
+          expected="$(jq -r --arg needle "$ARTIFACT_NAME" \
+            '[.deliverables[] | select(.name | contains($needle)) | .binary_sha256][0] // empty' \
+            certification/candidate.json)"
+          if [[ -z "$expected" ]]; then
+            echo "No certified binary digest recorded for ${ARTIFACT_NAME}" >&2
+            exit 1
+          fi
+          if ! command -v rpm2cpio >/dev/null 2>&1; then
+            sudo apt-get update && sudo apt-get install -y rpm2cpio
+          fi
+          if ! command -v cpio >/dev/null 2>&1; then
+            sudo apt-get update && sudo apt-get install -y cpio
+          fi
+          for package_file in "$DEB_FILE" "$RPM_FILE"; do
+            case "$package_file" in
+              *.deb)
+                dpkg-deb --fsys-tarfile "$package_file" | tar -xOf - ./usr/bin/rustfs > "$RUNNER_TEMP/wrapped-rustfs"
+                ;;
+              *.rpm)
+                rpm2cpio "$package_file" | cpio -i --to-stdout ./usr/bin/rustfs > "$RUNNER_TEMP/wrapped-rustfs"
+                ;;
+            esac
+            python3 scripts/release/release_gate.py verify-promotion \
+              --plan .config/release-validation-plan.json \
+              --candidate certification/candidate.json \
+              --summary certification/release-gate-output \
+              --artifact "$RUNNER_TEMP/wrapped-rustfs" \
+              --artifact-role binary
+          done
+
      - name: Upload packages to artifacts
        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
        with:
@@ -531,10 +627,14 @@ jobs:
          export AWS_SECRET_ACCESS_KEY="$R2_SECRET_ACCESS_KEY"
          export AWS_DEFAULT_REGION="auto"

+          # Build-phase writes stay on staging for release types: the official
+          # packages/release and packages/latest paths are written only by the
+          # certification-gated promote-packages job, from these same staged
+          # objects. Development packages keep the uncertified dev channel.
          if [[ "$BUILD_TYPE" == "development" ]]; then
            R2_PREFIX="artifacts/rustfs/packages/dev"
          else
-            R2_PREFIX="artifacts/rustfs/packages/release"
+            R2_PREFIX="artifacts/rustfs/packages/staging/${GITHUB_RUN_ID}/${GITHUB_RUN_ATTEMPT}"
          fi
          R2_PATH="s3://${R2_BUCKET}/${R2_PREFIX}/"

@@ -549,18 +649,6 @@ jobs:

          echo "✅ Upload complete"

-          # Also upload as latest for release/prerelease
-          if [[ "$BUILD_TYPE" == "release" || "$BUILD_TYPE" == "prerelease" ]]; then
-            LATEST_PATH="s3://${R2_BUCKET}/artifacts/rustfs/packages/latest/"
-            for f in "$DEB_FILE" "$RPM_FILE"; do
-              if [[ -n "$f" && -f "$f" ]]; then
-                echo "Uploading latest: $(basename "$f")"
-                aws s3 cp "$f" "$LATEST_PATH" --endpoint-url "$R2_ENDPOINT" --only-show-errors
-              fi
-            done
-            echo "✅ Latest packages updated"
-          fi
-
      - name: Upload packages to GitHub Release
        if: needs.resolve.outputs.tag != ''
        env:
@@ -614,6 +702,114 @@ jobs:

          echo "✅ GitHub release assets updated"

+  # The only writer of the official packages/release and packages/latest R2
+  # paths. Promotion re-checks each package's wrapped binary against the
+  # passed certification list and copies the staged objects byte-for-byte:
+  # no rebuild-after-verify, no mutable latest resolution downstream.
+  promote-packages:
+    name: Promote Packages
+    needs: [ resolve, package, release-certification ]
+    # Target-channel lock for the official package paths.
+    concurrency:
+      group: release-channel-r2-packages
+      cancel-in-progress: false
+    if: needs.resolve.outputs.tag != '' && (needs.resolve.outputs.build_type == 'release' || needs.resolve.outputs.build_type == 'prerelease')
+    runs-on: ubuntu-latest # gh CLI required; the sm-standard-2 fleet does not ship it
+    timeout-minutes: 30
+    permissions:
+      contents: read
+      actions: read
+    env:
+      R2_ACCESS_KEY_ID: ${{ secrets.R2_ACCESS_KEY_ID }}
+      R2_SECRET_ACCESS_KEY: ${{ secrets.R2_SECRET_ACCESS_KEY }}
+      R2_ENDPOINT: ${{ secrets.R2_ENDPOINT }}
+      R2_BUCKET: ${{ secrets.R2_BUCKET }}
+      AWS_EC2_METADATA_DISABLED: true
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          persist-credentials: false
+
+      - name: Download packages and certification
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
+        with:
+          path: ./packages
+          pattern: packages-*
+          merge-multiple: true
+
+      - name: Download certification evidence
+        uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8.0.1
+        with:
+          name: release-certification
+          path: ./certification
+
+      - name: Verify and promote staged packages
+        shell: bash
+        run: |
+          set -euo pipefail
+
+          if [[ -z "$R2_ACCESS_KEY_ID" || -z "$R2_SECRET_ACCESS_KEY" || -z "$R2_ENDPOINT" || -z "$R2_BUCKET" ]]; then
+            echo "❌ R2 credentials missing; refusing to promote to the official channel" >&2
+            exit 1
+          fi
+
+          STAGING_PREFIX="s3://${R2_BUCKET}/artifacts/rustfs/packages/staging/${GITHUB_RUN_ID}/${GITHUB_RUN_ATTEMPT}/"
+          RELEASE_PATH="s3://${R2_BUCKET}/artifacts/rustfs/packages/release/"
+          LATEST_PATH="s3://${R2_BUCKET}/artifacts/rustfs/packages/latest/"
+
+          if ! command -v aws >/dev/null 2>&1; then
+            sudo apt-get update && sudo apt-get install -y awscli
+          fi
+          export AWS_ACCESS_KEY_ID="$R2_ACCESS_KEY_ID"
+          export AWS_SECRET_ACCESS_KEY="$R2_SECRET_ACCESS_KEY"
+          export AWS_DEFAULT_REGION="auto"
+
+          if ! command -v rpm2cpio >/dev/null 2>&1; then
+            sudo apt-get install -y rpm2cpio
+          fi
+          if ! command -v cpio >/dev/null 2>&1; then
+            sudo apt-get install -y cpio
+          fi
+
+          promoted=0
+          for package_file in packages/*.deb packages/*.rpm; do
+            [[ -f "$package_file" ]] || continue
+            name="$(basename "$package_file")"
+            case "$package_file" in
+              *.deb)
+                dpkg-deb --fsys-tarfile "$package_file" | tar -xOf - ./usr/bin/rustfs > "$RUNNER_TEMP/wrapped-rustfs"
+                ;;
+              *.rpm)
+                rpm2cpio "$package_file" | cpio -i --to-stdout ./usr/bin/rustfs > "$RUNNER_TEMP/wrapped-rustfs"
+                ;;
+            esac
+            python3 scripts/release/release_gate.py verify-promotion \
+              --plan .config/release-validation-plan.json \
+              --candidate certification/candidate.json \
+              --summary certification/release-gate-output \
+              --artifact "$RUNNER_TEMP/wrapped-rustfs" \
+              --artifact-role binary
+
+            staged="$(mktemp)"
+            aws s3 cp "${STAGING_PREFIX}${name}" "$staged" --endpoint-url "$R2_ENDPOINT" --only-show-errors
+            staged_sha="$(sha256sum "$staged" | awk '{print $1}')"
+            built_sha="$(sha256sum "$package_file" | awk '{print $1}')"
+            if [[ "$staged_sha" != "$built_sha" ]]; then
+              echo "❌ Staged package ${name} does not match the verified bytes" >&2
+              exit 1
+            fi
+            aws s3 cp "$staged" "${RELEASE_PATH}${name}" --endpoint-url "$R2_ENDPOINT" --only-show-errors
+            aws s3 cp "$staged" "${LATEST_PATH}${name}" --endpoint-url "$R2_ENDPOINT" --only-show-errors
+            rm -f "$staged"
+            promoted=$((promoted + 1))
+            echo "✅ Promoted ${name} (${built_sha})"
+          done
+          if [[ "$promoted" -eq 0 ]]; then
+            echo "No staged packages found to promote" >&2
+            exit 1
+          fi
+
  # Summary
  summary:
    name: Summary
@@ -0,0 +1,410 @@
+# 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.
+
+name: RustFS Validation Sweep
+
+# Executes the migrated canonical validation cases (rustfs/auto-testing#113)
+# on the shared functional fleet. The catalog's implemented cases (66 today,
+# all non-heal) are swept in rotating nightly batches so the whole catalog is
+# covered roughly weekly; a single case or one module can be dispatched
+# manually. heal.* cases stay on the interactive validation flow for now:
+# many of them require reviewed native/tool collectors built by separate
+# validation-tools runs before fleet execution.
+
+on:
+  schedule:
+    # After the nightly functional chain settles (chain starts ~16:48 UTC and
+    # runs for about two hours).
+    - cron: "30 19 * * *"
+  workflow_dispatch:
+    inputs:
+      case:
+        description: 'Single canonical validation case id (overrides module and rotation), e.g. pool.decommission'
+        type: string
+        required: false
+        default: ''
+      module:
+        description: 'Restrict the sweep to one catalog module, e.g. bucket-replication'
+        type: string
+        required: false
+        default: ''
+      limit:
+        description: 'Maximum cases per sweep (0 = run everything selected)'
+        type: number
+        required: false
+        default: 12
+      build_run_id:
+        description: 'Successful nightly package run; empty resolves the latest'
+        type: string
+        required: false
+        default: ''
+      build_run_attempt:
+        description: 'Attempt of that package run'
+        type: string
+        required: false
+        default: '1'
+
+permissions:
+  contents: read
+
+concurrency:
+  group: rustfs-validation-sweep
+  cancel-in-progress: false
+
+env:
+  RUSTFS_NODES: ${{ secrets.RUSTFS_NODES || vars.RUSTFS_NODES || 'vm000 vm001 vm002 vm003' }}
+  RUSTFS_SSH_USER: ${{ secrets.RUSTFS_SSH_USER || vars.RUSTFS_SSH_USER || 'azureuser' }}
+  RUSTFS_ACCESS_KEY: ${{ secrets.RUSTFS_ACCESS_KEY }}
+  RUSTFS_SECRET_KEY: ${{ secrets.RUSTFS_SECRET_KEY }}
+  RUSTFS_NIGHTLY_PACKAGE_URL: ${{ vars.RUSTFS_NIGHTLY_PACKAGE_URL || 'https://dl.rustfs.com/artifacts/rustfs/packages/nightly/rustfs-nightly-latest.deb' }}
+  PF_TESTING_GH_TOKEN: ${{ secrets.PF_TESTING_GH_TOKEN }}
+
+jobs:
+  sweep:
+    if: github.repository == 'rustfs/rustfs'
+    runs-on: smoke-testing
+    timeout-minutes: 150
+    steps:
+      - name: Checkout chain tooling
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          persist-credentials: false
+
+      - name: Resolve nightly candidate
+        id: candidate
+        env:
+          GH_TOKEN: ${{ secrets.PF_TESTING_GH_TOKEN }}
+          BUILD_RUN_ID: ${{ inputs.build_run_id }}
+          BUILD_RUN_ATTEMPT: ${{ inputs.build_run_attempt }}
+          CHAIN_OUTPUT: ${{ runner.temp }}/validation-sweep-candidate.json
+        run: |
+          set -euo pipefail
+          python3 scripts/resolve_functional_candidate.py
+          python3 - <<'PY'
+          import json
+          import os
+          from pathlib import Path
+          root = Path(os.environ["RUNNER_TEMP"]) / "validation-sweep"
+          root.mkdir(parents=True, exist_ok=True)
+          candidate = json.loads((Path(os.environ["RUNNER_TEMP"]) / "validation-sweep-candidate.json").read_text())
+          manifest = candidate["manifest"]
+          (root / "candidate.json").write_text(json.dumps({
+              "package_url": manifest["package_url"],
+              "package_sha256": manifest["package_sha256"],
+              "source_sha": manifest["source_sha"],
+          }, sort_keys=True) + "\n")
+          with open(os.environ["GITHUB_OUTPUT"], "a") as output:
+              output.write("testing_sha=" + candidate["testing_sha"] + "\n")
+              output.write("package_url=" + manifest["package_url"] + "\n")
+              output.write("source_sha=" + manifest["source_sha"] + "\n")
+          PY
+
+      - name: Checkout auto-testing scripts
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          repository: rustfs/auto-testing
+          ref: ${{ steps.candidate.outputs.testing_sha }}
+          token: ${{ secrets.PF_TESTING_GH_TOKEN }}
+          path: auto-testing
+          persist-credentials: false
+
+      - name: Select sweep cases
+        id: select
+        env:
+          SELECTED_CASE: ${{ inputs.case }}
+          SELECTED_MODULE: ${{ inputs.module }}
+          CASE_LIMIT: ${{ inputs.limit }}
+          SWEEP_EVENT: ${{ github.event_name }}
+          SWEEP_ROOT: ${{ runner.temp }}/validation-sweep
+        run: |
+          set -euo pipefail
+          python3 - <<'PY'
+          import datetime
+          import json
+          import os
+          from pathlib import Path
+          sweep_root = Path(os.environ["SWEEP_ROOT"])
+
+          catalog = json.loads(Path("auto-testing/validation-cases.json").read_text())
+          selected = [(case["id"], module["id"]) for module in catalog["modules"]
+                      for case in module["cases"]
+                      if case.get("source_status") == "implemented"]
+          if not selected:
+              raise SystemExit("catalog has no implemented cases")
+
+          one_case = os.environ.get("SELECTED_CASE", "").strip()
+          module_filter = os.environ.get("SELECTED_MODULE", "").strip()
+          known = {case_id for case_id, _ in selected}
+          modules = {module_id for _, module_id in selected}
+          if one_case:
+              if one_case not in known:
+                  raise SystemExit(f"unknown or not-implemented case: {one_case}")
+              selected = [(one_case, dict(selected)[one_case])]
+          elif module_filter:
+              if module_filter not in modules:
+                  raise SystemExit(f"unknown module or no implemented cases: {module_filter}")
+              selected = [pair for pair in selected if pair[1] == module_filter]
+          elif os.environ.get("SWEEP_EVENT") == "schedule":
+              # Stateless rotation: every night take one window of the catalog
+              # so a full sweep needs len(selected)/limit nights and no state
+              # has to be carried between runs.
+              limit = max(1, int(os.environ.get("CASE_LIMIT") or 12))
+              day = datetime.datetime.now(datetime.timezone.utc).timetuple().tm_yday
+              start = (day * limit) % len(selected)
+              selected = (selected[start:] + selected[:start])[:limit]
+          limit = int(os.environ.get("CASE_LIMIT") or 0)
+          if one_case or module_filter or (not os.environ.get("SWEEP_EVENT") == "schedule"):
+              if limit > 0:
+                  selected = selected[:limit]
+
+          (sweep_root / "sweep-cases.txt").write_text("".join(f"{case_id}\n" for case_id, _ in selected))
+          with open(os.environ["GITHUB_OUTPUT"], "a") as output:
+              output.write(f"count={len(selected)}\n")
+          print(f"selected {len(selected)} case(s):", ", ".join(c for c, _ in selected))
+          PY
+
+      - name: Admit clean shared environment
+        id: environment
+        timeout-minutes: 5
+        env:
+          FUNCTIONAL_ENVIRONMENT_RECEIPT: ${{ runner.temp }}/environment-validation-sweep.json
+        run: python3 scripts/functional_environment.py enter --suite heal
+
+      - name: Run validation cases
+        id: run_cases
+        timeout-minutes: 120
+        env:
+          AUTO_TESTING_SHA: ${{ steps.candidate.outputs.testing_sha }}
+          SWEEP_ROOT: ${{ runner.temp }}/validation-sweep
+        run: |
+          set -uo pipefail
+          mkdir -p "${SWEEP_ROOT}/results"
+          : > "${SWEEP_ROOT}/results.txt"
+          while read -r CASE_ID; do
+            [ -n "${CASE_ID}" ] || continue
+            echo "::group::case ${CASE_ID}"
+            rm -rf "${SWEEP_ROOT}/run"
+            if ./auto-testing/scripts/run_validation_case.sh \
+                 "${CASE_ID}" "${SWEEP_ROOT}/candidate.json" \
+                 "${AUTO_TESTING_SHA}" "${SWEEP_ROOT}/run" \
+                 > "${SWEEP_ROOT}/results/${CASE_ID//\//_}.log" 2>&1; then
+              printf '%s\tPASS\n' "${CASE_ID}" >> "${SWEEP_ROOT}/results.txt"
+              echo "case ${CASE_ID}: PASS"
+            else
+              printf '%s\tFAIL\n' "${CASE_ID}" >> "${SWEEP_ROOT}/results.txt"
+              echo "case ${CASE_ID}: FAIL (log retained in the sweep artifact)"
+            fi
+            if [ -f "${SWEEP_ROOT}/run/report.md" ]; then
+              cp "${SWEEP_ROOT}/run/report.md" "${SWEEP_ROOT}/results/${CASE_ID//\//_}.report.md" 2>/dev/null || true
+            fi
+            echo "::endgroup::"
+            # Fleet hygiene between cases: the case runners set up their own
+            # topology, this purge guarantees the next case starts clean even
+            # when the previous one died mid-run.
+            for node in ${RUSTFS_NODES}; do
+              ssh -o BatchMode=yes -o ConnectTimeout=10 "${RUSTFS_SSH_USER}@${node}" '
+                SUDO=""; [ "$(id -u)" -ne 0 ] && SUDO="sudo -n"
+                ${SUDO} systemctl stop rustfs 2>/dev/null || true
+                if ${SUDO} dpkg -l rustfs 2>/dev/null | grep -q "^ii"; then
+                  ${SUDO} dpkg -P rustfs
+                fi
+                for i in 1 2 3 4; do ${SUDO} rm -rf /data/rustfs${i}/mnmd; done
+                ${SUDO} rm -f /etc/default/rustfs
+              ' || echo "inter-case cleanup failed on ${node} (continuing)"
+            done
+          done < "${SWEEP_ROOT}/sweep-cases.txt"
+          echo "sweep finished: $(wc -l < "${SWEEP_ROOT}/results.txt") case(s) executed"
+
+      - name: Generate sweep report
+        if: ${{ always() && steps.environment.outcome == 'success' }}
+        env:
+          SWEEP_ROOT: ${{ runner.temp }}/validation-sweep
+          PACKAGE_URL: ${{ steps.candidate.outputs.package_url }}
+          SOURCE_SHA: ${{ steps.candidate.outputs.source_sha }}
+        run: |
+          set -euo pipefail
+          python3 - <<'PY'
+          import json
+          import os
+          from pathlib import Path
+
+          root = Path(os.environ["SWEEP_ROOT"])
+          catalog = json.loads(Path("auto-testing/validation-cases.json").read_text())
+          module_of = {case["id"]: module["id"] for module in catalog["modules"] for case in module["cases"]}
+          rows = []
+          for line in (root / "results.txt").read_text().splitlines():
+            if not line:
+                continue
+            case_id, status = line.split("\t")
+            rows.append((case_id, module_of.get(case_id, "?"), status))
+          passed = sum(1 for _, _, s in rows if s == "PASS")
+          failed = len(rows) - passed
+          run_url = (f"{os.environ.get('GITHUB_SERVER_URL', 'https://github.com')}/"
+                     f"{os.environ.get('GITHUB_REPOSITORY', 'rustfs/rustfs')}/actions/runs/"
+                     f"{os.environ.get('GITHUB_RUN_ID', '')}")
+
+          report = [
+              "# RustFS validation sweep report", "",
+              f"- Run: {run_url}",
+              f"- Attempt: {os.environ.get('GITHUB_RUN_ATTEMPT', '')}",
+              f"- Trigger: {os.environ.get('GITHUB_EVENT_NAME', '')}",
+              f"- Package: {os.environ.get('PACKAGE_URL', 'nightly (R2 latest)')}",
+              f"- Source SHA: {os.environ.get('SOURCE_SHA', '')}",
+              f"- auto-testing: {os.environ.get('AUTO_TESTING_SHA', '')}", "",
+              "Canonical validation cases (rustfs/auto-testing#113) executed by the sweep, grouped by catalog module.", "",
+          ]
+          # Group the executed cases by catalog module (the major category:
+          # pool lifecycle, heal, ILM expiration, bucket/site replication, kms,
+          # data services, identity, s3 compatibility). Per-module subtotals
+          # deliberately use "- Subtotal:" so the dashboard overview regexes
+          # (which anchor on "- Total:") keep reading the grand totals below.
+          by_module = {}
+          module_order = []
+          for case_id, module, status in rows:
+              if module not in by_module:
+                  by_module[module] = []
+                  module_order.append(module)
+              by_module[module].append((case_id, status))
+          report += ["## Case Summary", ""]
+          for module in module_order:
+              cases = by_module[module]
+              mod_pass = sum(1 for _, st in cases if st == "PASS")
+              mod_fail = len(cases) - mod_pass
+              report += [f"### {module} ({len(cases)})", "",
+                         "| Case | Status |", "| --- | --- |"]
+              report += [f"| {case_id} | {status} |" for case_id, status in cases]
+              report += ["", f"- Subtotal: {mod_pass} passed, {mod_fail} failed", ""]
+          report += [
+              "## Totals", "",
+              f"- Total: {len(rows)}",
+              f"- PASS: {passed}",
+              f"- FAIL: {failed}",
+              "- UNSUPPORTED: 0", "",
+              "Failing cases keep this workflow red only as a report; the sweep artifact",
+              "retains every case log and per-case report for diagnosis.", "",
+          ]
+          (root / "report.md").write_text("\n".join(report))
+          (root / "cases.md").write_text("\n".join(f"| {c} | {m} | {s} |" for c, m, s in rows) + "\n")
+          PY
+          cat "${SWEEP_ROOT}/report.md" >> "${GITHUB_STEP_SUMMARY}"
+
+      - name: Upload functional report to dashboard
+        if: ${{ always() && steps.environment.outcome == 'success' }}
+        timeout-minutes: 2
+        continue-on-error: true
+        env:
+          GH_TOKEN: ${{ env.PF_TESTING_GH_TOKEN }}
+          REPORT_FILE: ${{ runner.temp }}/validation-sweep/report.md
+          SUITE: validation
+        run: |
+          set -euo pipefail
+          if [ -z "${GH_TOKEN:-}" ]; then
+            echo "PF_TESTING_GH_TOKEN is not configured; skipping dashboard upload"
+            exit 0
+          fi
+          DATE="$(date -u +%Y-%m-%d)"
+          REPORT_PATH="functional-reports/${SUITE}/${DATE}.md"
+          B64_FILE="$(mktemp)"
+          python3 -c 'import base64,sys;print(base64.b64encode(open(sys.argv[1],"rb").read()).decode())' "${REPORT_FILE}" > "${B64_FILE}"
+          SHA="$(gh api "repos/rustfs/dashboard/contents/${REPORT_PATH}" -q '.sha' 2>/dev/null || true)"
+          if [ -n "${SHA}" ]; then
+            jq -n --arg msg "report(${SUITE}): ${DATE}" --rawfile content "${B64_FILE}" --arg sha "${SHA}" \
+              '{message:$msg, content:($content|rtrimstr("\n")), sha:$sha}' \
+              | gh api --method PUT "repos/rustfs/dashboard/contents/${REPORT_PATH}" --input - >/dev/null
+          else
+            jq -n --arg msg "report(${SUITE}): ${DATE}" --rawfile content "${B64_FILE}" \
+              '{message:$msg, content:($content|rtrimstr("\n"))}' \
+              | gh api --method PUT "repos/rustfs/dashboard/contents/${REPORT_PATH}" --input - >/dev/null
+          fi
+          echo "dashboard updated: ${REPORT_PATH}"
+
+      - name: Retain sweep evidence
+        if: ${{ always() }}
+        uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        with:
+          name: rustfs-validation-sweep-${{ github.run_id }}-${{ github.run_attempt }}
+          path: |
+            ${{ runner.temp }}/validation-sweep/report.md
+            ${{ runner.temp }}/validation-sweep/results.txt
+            ${{ runner.temp }}/validation-sweep/sweep-cases.txt
+            ${{ runner.temp }}/validation-sweep/candidate.json
+            ${{ runner.temp }}/validation-sweep/results/
+          if-no-files-found: error
+          retention-days: 30
+
+      - name: Cleanup environment (after)
+        id: environment_cleanup
+        timeout-minutes: 5
+        if: ${{ always() && steps.environment.outcome == 'success' }}
+        run: |
+          set -uo pipefail
+          cleanup_status=0
+          for node in ${RUSTFS_NODES}; do
+            ssh -o BatchMode=yes -o ConnectTimeout=10 "${RUSTFS_SSH_USER}@${node}" '
+              set -euo pipefail
+              SUDO=""; [ "$(id -u)" -ne 0 ] && SUDO="sudo -n"
+              ${SUDO} systemctl stop rustfs 2>/dev/null || true
+              if ${SUDO} dpkg -l rustfs 2>/dev/null | grep -q "^ii"; then
+                ${SUDO} dpkg -P rustfs
+              fi
+              for i in 1 2 3 4; do ${SUDO} rm -rf /data/rustfs${i}/mnmd; done
+              ${SUDO} rm -f /etc/default/rustfs
+            ' || cleanup_status=$?
+          done
+          exit "${cleanup_status}"
+
+      - name: Release environment only after successful cleanup
+        if: ${{ always() && steps.environment.outcome == 'success' }}
+        timeout-minutes: 5
+        env:
+          FUNCTIONAL_ENVIRONMENT_RECEIPT: ${{ runner.temp }}/environment-validation-sweep.json
+          CLEANUP_OUTCOME: ${{ steps.environment_cleanup.outcome }}
+        run: python3 scripts/functional_environment.py finish --suite heal --cleanup-outcome "$CLEANUP_OUTCOME"
+
+      - name: Notify on failure
+        if: failure()
+        run: echo "RustFS validation sweep failed (harness/environment breakdown); see the sweep artifact for case logs."
+
+      - name: Enforce sweep result
+        if: ${{ always() && steps.environment.outcome == 'success' }}
+        env:
+          SWEEP_ROOT: ${{ runner.temp }}/validation-sweep
+        run: |
+          set -euo pipefail
+          if grep -q $'\tFAIL$' "${SWEEP_ROOT}/results.txt"; then
+            echo "validation sweep has failing cases:"
+            grep $'\tFAIL$' "${SWEEP_ROOT}/results.txt"
+            exit 1
+          fi
+
+  alert-on-failure:
+    name: Alert on scheduled failure
+    # A scheduled sweep failure usually means a canonical validation case
+    # regressed or the fleet environment broke; surface it as a tracking
+    # issue instead of leaving it in the Actions tab.
+    needs: [sweep]
+    if: github.repository == 'rustfs/rustfs' && (always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure'))
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+      issues: write
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v5.0.0
+        with:
+          persist-credentials: false
+      - name: Open or update failure-tracking issue
+        uses: ./.github/actions/schedule-failure-issue
+        with:
+          github-token: ${{ secrets.BACKLOG_ISSUE_TOKEN }}
@@ -1321,9 +1321,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-schema"
-version = "1.1.0"
+version = "1.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bec8dae1015a6d959a5faf86fbdf77486cfe19e4ccb6ee009d8a46efc45028a4"
+checksum = "9543fc722ed4865ab29fb66bd7ed0bef7b029d265ba8c587144511e30006d431"
 dependencies = [
 "aws-smithy-runtime-api",
 "aws-smithy-types",
@@ -4198,9 +4198,9 @@ checksum = "34aa73646ffb006b8f5147f3dc182bd4bcb190227ce861fc4a4844bf8e3cb2c0"

 [[package]]
 name = "encoding_rs"
-version = "0.8.42"
+version = "0.8.44"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8e985e0451871ad22fb8d2b6b076e2028a502a0d3950998c2c5c0a4f9b5d9679"
+checksum = "89253175e704ab869b04f9db224d38630e0211a53841a79181deae0889e169f5"
 dependencies = [
 "cfg-if",
 "core_detect",
@@ -5277,9 +5277,9 @@ dependencies = [

 [[package]]
 name = "hotpath"
-version = "0.30.1"
+version = "0.30.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "502dc37a7c218b9b1d3a0ccfb904da81d10d6dda0e899c722c503092d55576ea"
+checksum = "b39e941010d86063ff233bf8325269bf68097fb24ba67000fbddc0a410e8056e"
 dependencies = [
 "arc-swap",
 "async-channel",
@@ -5312,15 +5312,15 @@ dependencies = [

 [[package]]
 name = "hotpath-drain"
-version = "0.30.1"
+version = "0.30.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0ea17d04c231b90b1f68a91f6f80fb0762c5dea07c611dce9ae157b56a50ff4c"
+checksum = "d0faae1a9e1fd8ce5debaaaf452dcbe833fc8a8a4f3779d2ec8828b116c3c16a"

 [[package]]
 name = "hotpath-macros"
-version = "0.30.1"
+version = "0.30.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a76e0af911930fc97abc766b393b48cdca77fea9f5f6b388cf2123a6d9c30deb"
+checksum = "a24f5179ff680132003306354efd0ac8cde12e87b1a788719c836e91af35cb58"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -5899,9 +5899,9 @@ dependencies = [

 [[package]]
 name = "js-sys"
-version = "0.3.106"
+version = "0.3.107"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7883d941dae510fb2d978fc3fe018c71c9e2892fd38854de3e8b92c2e5ad9cc5"
+checksum = "8dc09d2154098bbf4ef54eab6b117a3c24cbbffe94046b16ee6d6d4510584d3a"
 dependencies = [
 "cfg-if",
 "futures-util",
@@ -6094,9 +6094,9 @@ dependencies = [

 [[package]]
 name = "libbz2-rs-sys"
-version = "0.2.5"
+version = "0.2.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "34b357333733e8260735ba5894eb928c02ecc69c78715f01a8019e7fa7f2db4c"
+checksum = "32f454fd6831a29c0e827ed075bca2a15be201bd57f4168f0fd1041124eab17e"

 [[package]]
 name = "libc"
@@ -8305,9 +8305,9 @@ dependencies = [

 [[package]]
 name = "pulldown-cmark"
-version = "0.13.5"
+version = "0.13.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "278e9268c93b9e498a2da46c34d90ba93a3255eb99009646398f34bc73ab36eb"
+checksum = "3db2be46835f013dd3f4fa7f5f7402c537787b54ebb6fe2b94bbdd7e2669bb7b"
 dependencies = [
 "bitflags 2.13.2",
 "memchr",
@@ -8823,9 +8823,9 @@ dependencies = [

 [[package]]
 name = "redis"
-version = "1.8.0"
+version = "1.8.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7acd809e80afa754c63f586955699d8990efe1c4b6613183ee5f027a11fbc8b5"
+checksum = "db7f3d75b5a3301746bff39a767bf4c5c42506b08513527308b3875bfb64a031"
 dependencies = [
 "arc-swap",
 "arcstr",
@@ -12217,9 +12217,9 @@ dependencies = [

 [[package]]
 name = "sysinfo"
-version = "0.40.0"
+version = "0.40.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "110e37d98880fd804cd47509b90a8444f7aaaab655ae625e83cd442a8bc2ebf2"
+checksum = "0e8a3edbe0394ce115f3917a00349dd242be09eb5a091ed509eaf67b55cfb8a8"
 dependencies = [
 "libc",
 "memchr",
@@ -12710,9 +12710,9 @@ checksum = "06bdbd8cfc056b8d2e2e85f29b56a3bdbecb527cef81eb39e3e7b98af4652770"

 [[package]]
 name = "tonic"
-version = "0.14.8"
+version = "0.14.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cd1fa9b1a51188f3266a823208078315811269d06d272c3a61be0b9a194194ff"
+checksum = "e2cdfc673aabd30fe8fee92e65837b644c37c5d47f0f97ae6d5e0f1237d15e1b"
 dependencies = [
 "async-trait",
 "axum",
@@ -12743,9 +12743,9 @@ dependencies = [

 [[package]]
 name = "tonic-build"
-version = "0.14.8"
+version = "0.14.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2c07bd4e9f99e2a3ed8bebad596ae02a94e53745b68acc6900e22c3475c8bc44"
+checksum = "ac6d916d206ed7bdda1778ddff9bda43303789c258877e7bc22fb44ee5c984b4"
 dependencies = [
 "prettyplease 0.2.37",
 "proc-macro2",
@@ -12755,9 +12755,9 @@ dependencies = [

 [[package]]
 name = "tonic-prost"
-version = "0.14.8"
+version = "0.14.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f63c616ffaa648f7c5dd5657e97d9a285fcb44ff3b70da761d436f5a13ff7be6"
+checksum = "eba464054e2226f182a34489b8b079bf4c6e3e7ab5fd7fc30e024f24aa92a484"
 dependencies = [
 "bytes",
 "prost 0.14.4",
@@ -12766,9 +12766,9 @@ dependencies = [

 [[package]]
 name = "tonic-prost-build"
-version = "0.14.8"
+version = "0.14.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4a05d8957fac6b3542999796cb4449d45da790751962f974a457e5e00c43d78c"
+checksum = "2ea216ddc448e08e7839ab09e20e60b06b6a1c32745ccb212c0ebd161c199f1b"
 dependencies = [
 "prettyplease 0.2.37",
 "proc-macro2",
@@ -13161,9 +13161,12 @@ checksum = "daf8dba3b7eb870caf1ddeed7bc9d2a049f3cfdfae7cb521b087cc33ae4c49da"

 [[package]]
 name = "utf8_iter"
-version = "1.0.4"
+version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b6c140620e7ffbb22c2dee59cafe6084a59b5ffc27a8859a5f0d494b5d52b6be"
+checksum = "af3bd9d3e571c26fcce9d033b87ea5ab061b8d2ffe9b7a5675d810d68f03edf0"
+dependencies = [
+ "utf_types",
+]

 [[package]]
 name = "utf8parse"
@@ -13171,6 +13174,12 @@ version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "06abde3611657adf66d383f00b093d7faecc7fa57071cce2578660c9f1010821"

+[[package]]
+name = "utf_types"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3a3b425286980a367278453f3b6d28cdcb134b574e5a35824a6ff5eebad50bb6"
+
 [[package]]
 name = "uuid"
 version = "1.29.0"
@@ -13285,12 +13294,11 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen"
-version = "0.2.129"
+version = "0.2.130"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9bb54f33acc68fd454578d9820b0bde1a1a3d17aa17bb7b6595806d02886d409"
+checksum = "a0181c14ddd6c7b8b0b0132a5f1ea2f32be0d893cd3ea97b6e3b28886b8a3d1c"
 dependencies = [
 "cfg-if",
- "once_cell",
 "rustversion",
 "wasm-bindgen-macro",
 "wasm-bindgen-shared",
@@ -13298,9 +13306,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-futures"
-version = "0.4.79"
+version = "0.4.80"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3cbab34de2d982e9b48e18d216d04c4a6f641066ff19ffb699980f591ee3610e"
+checksum = "5e14e0327cb5368ad60ccb22d04a9827a693c95a926d8d125e9fa39e0baec501"
 dependencies = [
 "js-sys",
 "tokio",
@@ -13309,9 +13317,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-macro"
-version = "0.2.129"
+version = "0.2.130"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2e29d0c35b16e224a7eeb5cd2d25e3e1968fbd65604117b44d3b789d00ee8535"
+checksum = "6e931a02a39abdbf66966fc35902b5ed58344750a70885bed4ebb3d8eaf28ed9"
 dependencies = [
 "quote",
 "wasm-bindgen-macro-support",
@@ -13319,9 +13327,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-macro-support"
-version = "0.2.129"
+version = "0.2.130"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6f501a8bc3719dba86ef8ae4728879c08001bea749eb1333ac5b91e040e2a6b7"
+checksum = "2ca4b54280e455bd831e33a6b8e734baf432fe63fbc35c42ac14ce67a043a53a"
 dependencies = [
 "bumpalo",
 "proc-macro2",
@@ -13332,9 +13340,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-shared"
-version = "0.2.129"
+version = "0.2.130"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "23f0c9c52aa7cd7d77769a4cfe2a9adb1b331f489a41d912ce14513d5ab995c6"
+checksum = "3b1f9c3798725016c6ee5f2f57869ec0831e8bbc0ab0310a86e787e97f41389e"
 dependencies = [
 "unicode-ident",
 ]
@@ -13354,9 +13362,9 @@ dependencies = [

 [[package]]
 name = "web-sys"
-version = "0.3.106"
+version = "0.3.107"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "88261b9deccee56594c11a3460c462c41f58d148598fe70ad77070126a68aba4"
+checksum = "d7d179596ff5ec0f6de4db64d54011bde6289cc49603d3a6e4ce3f329207ab88"
 dependencies = [
 "js-sys",
 "wasm-bindgen",
@@ -175,9 +175,9 @@ tokio-rustls = { default-features = false, version = "0.26.6" }
 tokio-stream = { version = "0.1.19" }
 tokio-test = "0.4.6"
 tokio-util = { version = "0.7.20" }
-tonic = { version = "0.14.8" }
-tonic-prost = { version = "0.14.8" }
-tonic-prost-build = { version = "0.14.8" }
+tonic = { version = "0.14.9" }
+tonic-prost = { version = "0.14.9" }
+tonic-prost-build = { version = "0.14.9" }
 tower = { version = "0.5.3" }
 tower-http = { version = "0.7.1" }
 h3 = "0.0.8"
@@ -319,7 +319,7 @@ rustfs-erasure-codec = { version = "9.0.0" }
 reed-solomon-simd = "3.1.0"
 regex = { version = "1.13.1" }
 rumqttc = { package = "rumqttc-next", version = "0.34.0" }
-redis = { version = "1.8.0" }
+redis = { version = "1.8.1" }
 rustify = { version = "0.7", default-features = false }
 rustix = { version = "1.1.6" }
 rust-embed = { version = "8.13.0" }
@@ -334,7 +334,7 @@ compact_str = "0.10.0"
 snap = "1.1.2"
 starshard = { version = "2.4.0" }
 strum = { version = "0.28.0" }
-sysinfo = "0.40.0"
+sysinfo = "0.40.1"
 temp-env = "0.3.6"
 tempfile = "3.27.0"
 test-case = "3.4.0"
@@ -386,7 +386,7 @@ dav-server = "0.11.0"
 # Performance Analysis and Memory Profiling
 rustfs-mimalloc = { version = "0.7.2" }
 # Preserve Unicode focus filters until rustfs/backlog#2302 is resolved.
-hotpath = { version = "0.30.1", default-features = false }
+hotpath = { version = "0.30.2", default-features = false }
 # Snapshot testing for output format regression detection
 insta = { version = "1.49" }

@@ -16,8 +16,193 @@
 mod tests {
    use crate::common::{RustFSTestEnvironment, init_logging};
    use aws_sdk_s3::Client;
+    use aws_sdk_s3::config::http::HttpResponse;
+    use aws_sdk_s3::error::{ProvideErrorMetadata, SdkError};
    use aws_sdk_s3::primitives::ByteStream;
-    use aws_sdk_s3::types::{BucketVersioningStatus, VersioningConfiguration};
+    use aws_sdk_s3::types::{BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, VersioningConfiguration};
+
+    fn assert_not_found<E: ProvideErrorMetadata + std::fmt::Debug>(error: &SdkError<E, HttpResponse>) {
+        assert_eq!(error.raw_response().map(|r| r.status().as_u16()), Some(404), "{error:?}");
+        assert_eq!(
+            error.as_service_error().and_then(ProvideErrorMetadata::code),
+            Some("NoSuchKey"),
+            "{error:?}"
+        );
+    }
+
+    #[tokio::test]
+    async fn conditional_writes_cannot_replace_delete_marker_with_if_match() {
+        init_logging();
+        let mut env = RustFSTestEnvironment::new().await.expect("create test environment");
+        env.start_rustfs_server(vec![]).await.expect("start RustFS");
+        let client = env.create_s3_client();
+        let bucket = "conditional-delete-marker-proof";
+        create_versioned_bucket(&client, bucket).await;
+        client
+            .put_object()
+            .bucket(bucket)
+            .key("source")
+            .body(ByteStream::from_static(b"source"))
+            .send()
+            .await
+            .expect("write copy source");
+
+        for operation in ["put", "copy", "multipart"] {
+            let old = client
+                .put_object()
+                .bucket(bucket)
+                .key(operation)
+                .body(ByteStream::from_static(b"old"))
+                .send()
+                .await
+                .expect("write historical version");
+            let old_etag = old.e_tag().expect("historical ETag");
+            let marker = client
+                .delete_object()
+                .bucket(bucket)
+                .key(operation)
+                .send()
+                .await
+                .expect("create delete marker");
+            let marker_id = marker.version_id().expect("delete marker version");
+
+            let (upload_id, completed) = if operation == "multipart" {
+                let upload = client
+                    .create_multipart_upload()
+                    .bucket(bucket)
+                    .key(operation)
+                    .send()
+                    .await
+                    .expect("initiate multipart upload");
+                let upload_id = upload.upload_id().expect("upload ID").to_owned();
+                let part = client
+                    .upload_part()
+                    .bucket(bucket)
+                    .key(operation)
+                    .upload_id(&upload_id)
+                    .part_number(1)
+                    .body(ByteStream::from_static(b"new"))
+                    .send()
+                    .await
+                    .expect("upload part");
+                let completed = CompletedMultipartUpload::builder()
+                    .parts(
+                        CompletedPart::builder()
+                            .part_number(1)
+                            .set_e_tag(part.e_tag().map(str::to_owned))
+                            .build(),
+                    )
+                    .build();
+                (Some(upload_id), Some(completed))
+            } else {
+                (None, None)
+            };
+
+            for etag in [old_etag, "\"00000000000000000000000000000000\""] {
+                match operation {
+                    "put" => {
+                        let err = client
+                            .put_object()
+                            .bucket(bucket)
+                            .key(operation)
+                            .if_match(etag)
+                            .body(ByteStream::from_static(b"new"))
+                            .send()
+                            .await
+                            .expect_err("If-Match must reject marker");
+                        assert_not_found(&err);
+                    }
+                    "copy" => {
+                        let err = client
+                            .copy_object()
+                            .bucket(bucket)
+                            .key(operation)
+                            .copy_source(format!("{bucket}/source"))
+                            .if_match(etag)
+                            .send()
+                            .await
+                            .expect_err("If-Match copy must reject marker");
+                        assert_not_found(&err);
+                    }
+                    _ => {
+                        let err = client
+                            .complete_multipart_upload()
+                            .bucket(bucket)
+                            .key(operation)
+                            .upload_id(upload_id.as_deref().expect("upload ID"))
+                            .multipart_upload(completed.clone().expect("completed parts"))
+                            .if_match(etag)
+                            .send()
+                            .await
+                            .expect_err("If-Match completion must reject marker");
+                        assert_not_found(&err);
+                    }
+                }
+                let versions = client
+                    .list_object_versions()
+                    .bucket(bucket)
+                    .prefix(operation)
+                    .send()
+                    .await
+                    .expect("list versions after rejected write");
+                assert_eq!(versions.versions().len(), 1, "{operation}: failed write created a version");
+                assert_eq!(versions.delete_markers().len(), 1);
+                assert_eq!(versions.delete_markers()[0].version_id(), Some(marker_id));
+                assert_eq!(versions.delete_markers()[0].is_latest(), Some(true));
+            }
+
+            match operation {
+                "put" => {
+                    client
+                        .put_object()
+                        .bucket(bucket)
+                        .key(operation)
+                        .if_none_match("*")
+                        .body(ByteStream::from_static(b"new"))
+                        .send()
+                        .await
+                        .expect("If-None-Match PUT over marker");
+                }
+                "copy" => {
+                    client
+                        .copy_object()
+                        .bucket(bucket)
+                        .key(operation)
+                        .copy_source(format!("{bucket}/source"))
+                        .if_none_match("*")
+                        .send()
+                        .await
+                        .expect("If-None-Match copy over marker");
+                }
+                _ => {
+                    client
+                        .complete_multipart_upload()
+                        .bucket(bucket)
+                        .key(operation)
+                        .upload_id(upload_id.as_deref().expect("upload ID"))
+                        .multipart_upload(completed.expect("completed parts"))
+                        .if_none_match("*")
+                        .send()
+                        .await
+                        .expect("If-None-Match completion over marker");
+                }
+            }
+            let current = client
+                .get_object()
+                .bucket(bucket)
+                .key(operation)
+                .send()
+                .await
+                .expect("read successful conditional write")
+                .body
+                .collect()
+                .await
+                .expect("read object body")
+                .into_bytes();
+            let expected: &[u8] = if operation == "copy" { b"source" } else { b"new" };
+            assert_eq!(current.as_ref(), expected, "{operation}: unexpected current object");
+        }
+    }

    async fn create_versioned_bucket(client: &Client, bucket: &str) {
        client
@@ -1636,20 +1636,53 @@ async fn hot_window_with_old_followers(
    let mixed_initial_physical = if mixed { json!(converged_shards(&run)?) } else { Value::Null };
    let scheduler_config = run.assert_default_cadence().await?;
    let baseline = run.quiet_baseline().await?;
-    let leader = run.leader().await?;
-    let leader_pid = run.pid(leader)?;
+    // A leader-lock refresh gap during old-binary replacement can make
+    // `leader()` briefly observe a different node. Poll until the coordinator
+    // is stable before comparing against the measured fleet. A persistent
+    // change to an old-binary node is a genuine election outcome and fails.
+    let (leader, leader_pid) = if mixed {
+        let expected_leader = usize::try_from(
+            mixed_fleet["leader"]
+                .as_u64()
+                .ok_or("mixed fleet omitted its measured leader")?,
+        )?;
+        let expected_pid = mixed_fleet["leader_pid"]
+            .as_u64()
+            .ok_or("mixed fleet omitted its measured leader pid")?;
+        let mut resolved = None;
+        for _ in 0..20 {
+            let candidate = run.leader().await?;
+            let candidate_pid = u64::from(run.pid(candidate)?);
+            if Some(u64::try_from(candidate)?) == mixed_fleet["leader"].as_u64() && Some(candidate_pid) == Some(expected_pid) {
+                resolved = Some((candidate, run.pid(candidate)?));
+                break;
+            }
+            tokio::time::sleep(std::time::Duration::from_millis(500)).await;
+        }
+        match resolved {
+            Some(pair) => pair,
+            None => {
+                let current = match run.leader().await {
+                    Ok(node) => Some((node, run.pid(node).ok())),
+                    Err(_) => None,
+                };
+                return Err(format!(
+                    "the hot window did not retain its measured new coordinator after stabilization \
+                     (expected node {expected_leader} pid {expected_pid}, observed {current:?})"
+                )
+                .into());
+            }
+        }
+    } else {
+        let leader = run.leader().await?;
+        (leader, run.pid(leader)?)
+    };
    let producer = (leader + 1) % 4;
    let producers = if fault.is_some() {
        vec![producer, (leader + 2) % 4]
    } else {
        vec![producer]
    };
-    if mixed
-        && (mixed_fleet["leader"].as_u64() != Some(u64::try_from(leader)?)
-            || mixed_fleet["leader_pid"].as_u64() != Some(u64::from(leader_pid)))
-    {
-        return Err("the hot window did not retain its measured new coordinator".into());
-    }
    let producer_pid = run.pid(producer)?;
    let producer_address = run.peer_address(producer);
    let [cold_a, cold_b, hot] = run.buckets.clone();
@@ -2146,3 +2146,4 @@ mod tests {
        );
    }
 }
+// event-drill: isolated harness-only change
@@ -1112,8 +1112,9 @@ async fn wait_for_upgrade_writes(
            match response {
                Ok(Ok(_)) => break,
                Ok(Err(error)) => {
+                    let code = error.as_service_error().and_then(ProvideErrorMetadata::code);
                    if error.raw_response().map(|response| response.status().as_u16()) != Some(503)
-                        || error.as_service_error().and_then(ProvideErrorMetadata::code) != Some("ServiceUnavailable")
+                        || !matches!(code, Some("ServiceUnavailable" | "SlowDown" | "SlowDownRead"))
                    {
                        return Err(format!("{phase}: node {node} {probe} failed: {error:?}").into());
                    }
@@ -683,6 +683,7 @@ pub(crate) async fn update_validated_in<F>(
    bucket: &str,
    config_file: &str,
    data: Vec<u8>,
+    expected_incarnation_id: Option<Uuid>,
    validate: F,
 ) -> Result<OffsetDateTime>
 where
@@ -690,7 +691,7 @@ where
 {
    Box::pin(async {
        let sys = require_bucket_metadata_sys_in(ctx)?;
-        let guard = acquire_config_write_guard(sys.clone(), bucket).await?;
+        let guard = acquire_config_write_guard_for_incarnation(sys.clone(), bucket, expected_incarnation_id).await?;
        validate()?;
        update_under_config_write_guard(sys, &guard, config_file, data, None).await
    })
@@ -1901,6 +1902,15 @@ impl BucketMetadataSys {
        self.api.upgrade()
    }

+    /// Whether the backing `ECStore` has been dropped.
+    ///
+    /// Used by `InstanceContext::clear_bucket_metadata_sys_if_stale` during
+    /// `ECStore::drop` so a shared bootstrap context does not retain a
+    /// `BucketMetadataSys` whose weak handle is dangling.
+    pub(crate) fn is_store_stale(&self) -> bool {
+        self.api.upgrade().is_none()
+    }
+
    fn metadata_publish_lock(&self, bucket: &str) -> Arc<Mutex<MetadataPublishLockState>> {
        Self::bucket_lock_in(&self.metadata_publish_locks, bucket)
    }
@@ -5200,6 +5210,61 @@ mod tests {
        );
    }

+    #[tokio::test]
+    #[serial]
+    async fn validated_config_update_rejects_a_different_bucket_incarnation() {
+        let (_dirs, store) = isolated_store_over_temp_disks().await;
+        init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let bucket = "catalog-cleanup-incarnation";
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create bucket");
+        let incarnation = store
+            .bucket_incarnation_id_from_disk(bucket)
+            .await
+            .expect("read bucket incarnation");
+        let called = std::sync::atomic::AtomicBool::new(false);
+        let result = store
+            .update_bucket_metadata_config_for_incarnation_validated(
+                bucket,
+                "quota.json",
+                br#"{"quota":1}"#.to_vec(),
+                Uuid::new_v4(),
+                || {
+                    called.store(true, std::sync::atomic::Ordering::SeqCst);
+                    Ok(())
+                },
+            )
+            .await;
+        assert!(matches!(result, Err(Error::BucketNotFound(name)) if name == bucket));
+        assert!(
+            !called.load(std::sync::atomic::Ordering::SeqCst),
+            "incarnation must be checked before the external fence"
+        );
+        let (metadata, _) = get_config_from_disk_with_presence_in(&store.ctx, bucket)
+            .await
+            .expect("read unchanged bucket metadata");
+        assert!(
+            metadata.quota_config_json.is_empty(),
+            "a stale cleanup must not modify the replacement bucket"
+        );
+        store
+            .update_bucket_metadata_config_for_incarnation_validated(
+                bucket,
+                "quota.json",
+                br#"{"quota":1}"#.to_vec(),
+                incarnation,
+                || Ok(()),
+            )
+            .await
+            .expect("the current incarnation permits the validated update");
+        let (metadata, _) = get_config_from_disk_with_presence_in(&store.ctx, bucket)
+            .await
+            .expect("read updated bucket metadata");
+        assert_eq!(metadata.quota_config_json, br#"{"quota":1}"#);
+    }
+
    #[tokio::test]
    #[serial]
    async fn quota_admission_rejects_stale_cached_no_quota() {
@@ -747,6 +747,16 @@ async fn list_path_raw_inner(
            // Start every missing head read in the same round so one stalled
            // disk cannot multiply the wait budget by the erasure-set width.
            // Outcomes are still consumed below in stable disk-index order.
+            if !revjob_rx.is_cancelled() {
+                for (i, reader) in readers.iter_mut().enumerate() {
+                    if errs[i].is_none()
+                        && pending_entries[i].is_none()
+                        && let Some(Ok(Some(entry))) = reader.try_peek_buffered()
+                    {
+                        pending_entries[i] = Some(entry);
+                    }
+                }
+            }
            let concurrent_peeks = readers.iter_mut().enumerate().filter_map(|(i, reader)| {
                if errs[i].is_some() || pending_entries[i].is_some() {
                    return None;
@@ -878,7 +888,7 @@ async fn list_path_raw_inner(
                }
                if entry.name == current.name {
                    // If exact match, we agree.
-                    if current.matches_strict(&entry) {
+                    if current.matches_strict_borrowed(&entry) {
                        top_entries[i] = Some(entry);
                        agree += 1;
                    } else {
@@ -986,7 +996,9 @@ async fn list_path_raw_inner(

            if agree == readers.len() && !opts.preserve_replica_metadata {
                for r in readers.iter_mut() {
-                    let _ = r.skip(1).await;
+                    if !r.skip_current() {
+                        let _ = r.skip(1).await;
+                    }
                }

                if let Some(agreed_fn) = opts.agreed.as_ref() {
@@ -1009,7 +1021,7 @@ async fn list_path_raw_inner(
            }

            for (i, r) in readers.iter_mut().enumerate() {
-                if top_entries[i].is_some() {
+                if top_entries[i].is_some() && !r.skip_current() {
                    let _ = r.skip(1).await;
                }
            }
@@ -3506,6 +3506,7 @@ mod tests {
                replication_decision: String::new(),
                checksum: None,
                multipart_completion_replayed: false,
+                backfilled_old_current_size: None,
            }
        }
    }
@@ -2398,6 +2398,10 @@ impl LocalDiskWrapper {
        .await
    }

+    pub(crate) fn has_clean_active_multipart_source(&self, upload_or_part_path: &str, part_path: bool) -> bool {
+        self.disk.has_clean_active_multipart_source(upload_or_part_path, part_path)
+    }
+
    pub(crate) async fn try_publish_prepared_delete_local(
        &self,
        volume: &str,
@@ -175,17 +175,30 @@ pub(super) fn create_local_inline_rollback_backup(
    dst_file_path: &Path,
    staging_file_path: &Path,
    old_metadata: &[u8],
+    sync: bool,
 ) -> std::io::Result<PathBuf> {
    let Some(staging_parent) = staging_file_path.parent() else {
        return Err(std::io::Error::new(ErrorKind::InvalidInput, "missing staging metadata parent"));
    };
    let backup_path = staging_parent.join(STORAGE_FORMAT_FILE_BACKUP);
    remove_file_if_exists(&backup_path)?;
-    if (should_fail_local_inline_rollback_hardlink(dst_file_path) || std::fs::hard_link(dst_file_path, &backup_path).is_err())
-        && let Err(err) = std::fs::write(&backup_path, old_metadata)
-    {
-        let _ = remove_file_if_exists(&backup_path);
-        return Err(err);
+    if should_fail_local_inline_rollback_hardlink(dst_file_path) || std::fs::hard_link(dst_file_path, &backup_path).is_err() {
+        let write_res = (|| -> std::io::Result<()> {
+            #[cfg(test)]
+            run_owned_file_before_open(&backup_path);
+            let mut file = std::fs::File::create(&backup_path)?;
+            std::io::Write::write_all(&mut file, old_metadata)?;
+            if sync && !old_metadata.is_empty() {
+                file.sync_data()?;
+            }
+            Ok(())
+        })();
+        if let Err(err) = write_res {
+            let _ = remove_file_if_exists(&backup_path);
+            return Err(err);
+        }
+    } else if sync && !old_metadata.is_empty() {
+        std::fs::File::open(&backup_path)?.sync_data()?;
    }
    Ok(backup_path)
 }
@@ -356,9 +369,12 @@ fn prepare_rename_commit_and_destination_metadata_std(

 #[cfg(target_os = "linux")]
 enum FusedCommitAttempt {
-    Committed(RenameDataResp),
+    Committed(Box<RenameDataResp>),
    CommitErr(DiskError),
-    Fallback(os::RenameCommitGuard, Option<Bytes>, FileInfo),
+    RestoreDataSource(os::RenameCommitGuard, DiskError),
+    Fallback(os::RenameCommitGuard, Option<Bytes>, Box<FileInfo>),
+    #[cfg(test)]
+    ResumePanic(Box<dyn std::any::Any + Send + 'static>),
 }

 async fn prepare_rename_commit_and_destination_metadata(
@@ -540,7 +556,9 @@ impl LocalDisk {
            None
        };
        let multipart_publication = if let Some(staging) = multipart_staging.as_ref() {
-            Some(pin_multipart_future(|| self.prepare_native_multipart_publication(src_path, &mut fi, staging)).await?)
+            Some(Arc::new(
+                pin_multipart_future(|| self.prepare_native_multipart_publication_deferred(src_path, &mut fi, staging)).await?,
+            ))
        } else {
            None
        };
@@ -741,11 +759,6 @@ impl LocalDisk {
            && !fi_healing
            && !os::dst_dir_fsync_group_commit_enabled()
            && !os::file_fdatasync_group_commit_enabled()
-            && !super::has_rename_data_test_failpoints(dst_path)
-            && !crash_inject::is_armed(CrashPoint::RenameAfterDataRename, dst_path)
-            && !crash_inject::is_armed(CrashPoint::RenameAfterBackupBeforeMetaCommit, dst_path)
-            && !crash_inject::is_armed(CrashPoint::RenameAfterMetaCommit, dst_path)
-            && !super::has_delete_version_test_failpoints()
        {
            #[cfg(test)]
            if !no_inline {
@@ -785,10 +798,11 @@ impl LocalDisk {
                    &dst_volume_dir_owned,
                    &publication_root_owned,
                )?;
-                // An existing xl.meta can share versions or require rollback
-                // restoration when a post-rename directory sync fails.
-                if is_cancelled() || (sync_commit_metadata && has_dst_buf.is_some()) {
-                    return Ok(FusedCommitAttempt::Fallback(guard, has_dst_buf, fi_moved));
+                // An existing xl.meta that replaces an existing version slot
+                // requires a rollback directory; appending a new version keeps
+                // an inline rollback backup for post-rename directory sync errors.
+                if is_cancelled() {
+                    return Ok(FusedCommitAttempt::Fallback(guard, has_dst_buf, Box::new(fi_moved)));
                }
                let mut xlmeta = FileMeta::new();
                let mut dst_meta_unparsable = false;
@@ -811,7 +825,19 @@ impl LocalDisk {
                    }
                });
                if rollback_data_dir.is_some() {
-                    return Ok(FusedCommitAttempt::Fallback(guard, has_dst_buf, fi_moved));
+                    return Ok(FusedCommitAttempt::Fallback(guard, has_dst_buf, Box::new(fi_moved)));
+                }
+                #[cfg(test)]
+                let probe_active = os::file_sync_probe::is_active();
+                #[cfg(not(test))]
+                let probe_active = false;
+                if no_inline
+                    && sync_data_shards
+                    && let Some((src_data_path, _)) = has_data_dir_path_owned.as_ref()
+                    && (probe_active || !super::is_presynced_tmp_data_dir(src_data_path))
+                    && os::regular_files(src_data_path).is_ok_and(|files| files.len() >= os::PARALLEL_FILE_SYNC_THRESHOLD)
+                {
+                    return Ok(FusedCommitAttempt::Fallback(guard, has_dst_buf, Box::new(fi_moved)));
                }
                let (old_current_size, old_current_source_checked, old_current_source) = observe_old_current(
                    has_dst_buf.is_some(),
@@ -830,13 +856,31 @@ impl LocalDisk {
                    skip_parent = dst_parent_owned.clone();
                }

+                #[cfg(test)]
+                run_owned_file_before_open(&src_file_path_owned);
                let mut prepared_metadata_source =
                    os::create_prepared_rename_source_with_commit_guard(&src_file_path_owned, &dst_file_path_owned, &guard)?;
-                prepared_metadata_source.write_all(&new_dst_buf, sync_commit_metadata)?;
+                prepared_metadata_source.write_all(&new_dst_buf, false)?;
+                if !no_inline {
+                    run_inline_preparation_before_backup(&dst_path_str);
+                }
+
+                let mut staged_rollback_path = None;
+                if let Some(ref old_metadata) = has_dst_buf
+                    && (sync_commit_metadata || cfg!(test))
+                {
+                    let backup_path = create_local_inline_rollback_backup(
+                        &dst_file_path_owned,
+                        &src_file_path_owned,
+                        old_metadata,
+                        sync_commit_metadata,
+                    )?;
+                    staged_rollback_path = Some(StagedRollbackBackup(backup_path));
+                }

                if no_inline {
                    if sync_data_shards && let Some((src_data_path, _)) = has_data_dir_path_owned.as_ref() {
-                        let sync_result = if super::take_presynced_tmp_data_dir(src_data_path) {
+                        let sync_result = if !probe_active && super::take_presynced_tmp_data_dir(src_data_path) {
                            // The payload was synced, but its part.1 directory entry
                            // must also be persisted before moving the data directory.
                            os::fsync_dir_std(src_data_path)
@@ -846,49 +890,157 @@ impl LocalDisk {
                        if let Err(err) = sync_result
                            && err.kind() != ErrorKind::NotFound
                        {
+                            if let Some(backup_path) = staged_rollback_path.as_ref().map(StagedRollbackBackup::path) {
+                                let _ = remove_file_if_exists(backup_path);
+                            }
                            return Err(err);
                        }
                    }
+                    if sync_commit_metadata && let Err(err) = prepared_metadata_source.sync_data() {
+                        if let Some(backup_path) = staged_rollback_path.as_ref().map(StagedRollbackBackup::path) {
+                            let _ = remove_file_if_exists(backup_path);
+                        }
+                        return Err(err);
+                    }
+                } else if sync_commit_metadata && let Err(err) = prepared_metadata_source.sync_data() {
+                    if let Some(backup_path) = staged_rollback_path.as_ref().map(StagedRollbackBackup::path) {
+                        let _ = remove_file_if_exists(backup_path);
+                    }
+                    return Err(err);
+                }
+
+                if is_cancelled() {
+                    drop(prepared_metadata_source);
+                    if let Some(backup_path) = staged_rollback_path.as_ref().map(StagedRollbackBackup::path) {
+                        let _ = remove_file_if_exists(backup_path);
+                    }
+                    return Err(std::io::Error::new(ErrorKind::Interrupted, "commit cancelled"));
+                }
+
+                let guard = match super::remove_dst_base_before_commit_std(
+                    &dst_path_str,
+                    guard,
+                    &src_parent_owned,
+                    &dst_parent_owned,
+                    &dst_volume_dir_owned,
+                    &publication_root_owned,
+                ) {
+                    Ok(guard) => guard,
+                    Err(err) => {
+                        if let Some(backup_path) = staged_rollback_path.as_ref().map(StagedRollbackBackup::path) {
+                            let _ = remove_file_if_exists(backup_path);
+                        }
+                        return Ok(FusedCommitAttempt::CommitErr(err));
+                    }
+                };
+                if should_remove_staged_meta_before_commit(&dst_path_str) {
+                    drop(prepared_metadata_source);
+                    if let Some(backup_path) = staged_rollback_path.as_ref().map(StagedRollbackBackup::path) {
+                        let _ = remove_file_if_exists(backup_path);
+                    }
+                    std::fs::remove_file(&src_file_path_owned)?;
+                    return Ok(FusedCommitAttempt::CommitErr(DiskError::FileNotFound));
+                }
+
+                if no_inline {
                    if let Some((src_data_path, dst_data_path)) = has_data_dir_path_owned.as_ref()
                        && let Err(err) = os::rename_all_with_commit_guard_on_thread(src_data_path, dst_data_path, &guard)
                    {
-                        return Ok(FusedCommitAttempt::CommitErr(err));
+                        if let Some(backup_path) = staged_rollback_path.as_ref().map(StagedRollbackBackup::path) {
+                            let _ = remove_file_if_exists(backup_path);
+                        }
+                        return Ok(FusedCommitAttempt::RestoreDataSource(guard, err));
                    }
                    #[cfg(test)]
-                    if has_data_dir_path_owned.is_some() {
-                        run_rename_data_after_first_publication(&root_owned, &dst_volume_str, &dst_path_str);
+                    if has_data_dir_path_owned.is_some()
+                        && let Err(panic_payload) = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
+                            run_rename_data_after_first_publication(&root_owned, &dst_volume_str, &dst_path_str);
+                        }))
+                    {
+                        return Ok(FusedCommitAttempt::ResumePanic(panic_payload));
+                    }
+                    if crash_inject::should_crash_at(CrashPoint::RenameAfterDataRename, &dst_path_str) {
+                        return Ok(FusedCommitAttempt::CommitErr(DiskError::Unexpected));
+                    }
+                    if should_fail_before_old_metadata_backup(&dst_path_str) {
+                        if let Some(backup_path) = staged_rollback_path.as_ref().map(StagedRollbackBackup::path) {
+                            let _ = remove_file_if_exists(backup_path);
+                        }
+                        return Ok(FusedCommitAttempt::RestoreDataSource(guard, DiskError::Unexpected));
+                    }
+                    if crash_inject::should_crash_at(CrashPoint::RenameAfterBackupBeforeMetaCommit, &dst_path_str) {
+                        return Ok(FusedCommitAttempt::CommitErr(DiskError::Unexpected));
                    }
                }

-                if let Err(err) = os::rename_prepared_source_on_thread(
-                    prepared_metadata_source,
-                    &src_file_path_owned,
-                    &dst_file_path_owned,
-                    &skip_parent,
-                    &publication_root_owned,
-                    &guard,
-                ) {
-                    if no_inline && let Some((src_data_path, dst_data_path)) = has_data_dir_path_owned.as_ref() {
-                        let _ = std::fs::rename(dst_data_path, src_data_path);
+                let rename_res = if should_fail_commit_rename(&dst_path_str) {
+                    drop(prepared_metadata_source);
+                    Err(DiskError::other("test fail during metadata commit rename"))
+                } else {
+                    os::rename_prepared_source_on_thread(
+                        prepared_metadata_source,
+                        &src_file_path_owned,
+                        &dst_file_path_owned,
+                        &skip_parent,
+                        &publication_root_owned,
+                        &guard,
+                    )
+                };
+                if let Err(err) = rename_res {
+                    if let Some(backup_path) = staged_rollback_path.as_ref().map(StagedRollbackBackup::path) {
+                        let _ = remove_file_if_exists(backup_path);
+                    }
+                    if no_inline {
+                        return Ok(FusedCommitAttempt::RestoreDataSource(guard, err));
                    }
                    return Ok(FusedCommitAttempt::CommitErr(err));
                }

-                if sync_commit_metadata
-                    && let Err(err) =
-                        fsync_inline_commit_std(&dst_file_path_owned, &dst_volume_dir_owned, destination_was_absent, None, &guard)
-                {
-                    if no_inline && let Some((src_data_path, dst_data_path)) = has_data_dir_path_owned.as_ref() {
-                        let _ = std::fs::rename(dst_data_path, src_data_path);
+                if should_fail_after_metadata_commit(&dst_path_str) {
+                    if no_inline {
+                        rollback_inline_metadata_std(&dst_file_path_owned, None, staged_rollback_path.as_ref())?;
+                        if let Some((_, new_data_path)) = has_data_dir_path_owned.as_ref() {
+                            remove_dir_all_if_exists(new_data_path)?;
+                        }
+                        return Ok(FusedCommitAttempt::CommitErr(DiskError::Unexpected));
                    }
-                    return Err(err);
+                    let err = rollback_inline_metadata_std(&dst_file_path_owned, None, staged_rollback_path.as_ref())
+                        .err()
+                        .unwrap_or_else(|| std::io::Error::other("test fail after metadata commit"));
+                    return Ok(FusedCommitAttempt::CommitErr(DiskError::from(err)));
+                }
+
+                if crash_inject::should_crash_at(CrashPoint::RenameAfterMetaCommit, &dst_path_str) {
+                    return Ok(FusedCommitAttempt::CommitErr(DiskError::Unexpected));
+                }
+
+                if sync_commit_metadata
+                    && let Err(err) = fsync_inline_commit_std(
+                        &dst_file_path_owned,
+                        &dst_volume_dir_owned,
+                        destination_was_absent,
+                        staged_rollback_path.as_ref(),
+                        &guard,
+                    )
+                {
+                    if no_inline {
+                        if let Some((_, new_data_path)) = has_data_dir_path_owned.as_ref() {
+                            remove_dir_all_if_exists(new_data_path)?;
+                        }
+                        return Ok(FusedCommitAttempt::CommitErr(to_file_error(err).into()));
+                    }
+                    return Ok(FusedCommitAttempt::CommitErr(DiskError::from(err)));
+                }
+
+                if let Some(backup_path) = staged_rollback_path.as_ref().map(StagedRollbackBackup::path) {
+                    let _ = remove_file_if_exists(backup_path);
                }

                if let Some(parent) = src_file_path_owned.parent() {
                    let _ = std::fs::remove_dir(parent);
                }

-                Ok(FusedCommitAttempt::Committed(RenameDataResp {
+                Ok(FusedCommitAttempt::Committed(Box::new(RenameDataResp {
                    multipart_source: None,
                    old_data_dir: None,
                    rollback_data_dir: None,
@@ -897,7 +1049,7 @@ impl LocalDisk {
                    old_current_size,
                    old_current_source_checked,
                    old_current_source,
-                }))
+                })))
            };

            let attempt_res = if let Some(admission) = file_sync_admission.as_ref() {
@@ -917,7 +1069,23 @@ impl LocalDisk {
                    if let Some(receipt_mutation) = receipt_mutation.as_mut() {
                        receipt_mutation.complete();
                    }
-                    return Ok(resp);
+                    return Ok(*resp);
+                }
+                Ok(FusedCommitAttempt::RestoreDataSource(directory_guard, err)) => {
+                    let commit = RenameCommitContext {
+                        directory_guard,
+                        publication_root: &self.publication_root,
+                        mutation_lease,
+                        source_metadata: &src_file_path,
+                        destination_metadata: &dst_file_path,
+                        source_volume_dir: &src_volume_dir,
+                        data_paths: has_data_dir_path.as_ref(),
+                    };
+                    commit.restore_published_data_source().await?;
+                    for part_path in &invalidate_part_paths {
+                        self.io_backend.invalidate_cached_fd(dst_volume, part_path).await;
+                    }
+                    return Err(err);
                }
                Ok(FusedCommitAttempt::CommitErr(err)) | Err(err) => {
                    for part_path in &invalidate_part_paths {
@@ -925,11 +1093,269 @@ impl LocalDisk {
                    }
                    return Err(err);
                }
+                #[cfg(test)]
+                Ok(FusedCommitAttempt::ResumePanic(panic_payload)) => {
+                    std::panic::resume_unwind(panic_payload);
+                }
                Ok(FusedCommitAttempt::Fallback(guard, dst_buf, restored_fi)) => {
-                    fi = restored_fi;
+                    fi = *restored_fi;
                    (guard, dst_buf)
                }
            }
+        } else if let Some(publication) = multipart_publication.as_ref()
+            && no_inline
+            && !preserve_upload_metadata
+            && !publication.legacy
+            && !fi_healing
+            && !os::dst_dir_fsync_group_commit_enabled()
+        {
+            enum MultipartFusedCommitOutcome {
+                Committed(Box<RenameDataResp>),
+                AbortBeforeDataRename(DiskError),
+                RestoreDataSource(os::RenameCommitGuard, DiskError),
+                RestoreTargetAfterError(DiskError, bool),
+            }
+            let publication_owned = Arc::clone(publication);
+            let src_parent_owned = src_file_parent.to_path_buf();
+            let dst_parent_owned = dst_file_parent.to_path_buf();
+            let dst_file_path_owned = dst_file_path.clone();
+            let src_file_path_owned = src_file_path.clone();
+            let dst_volume_dir_owned = dst_volume_dir.clone();
+            let publication_root_owned = self.publication_root.clone();
+            let has_data_dir_path_owned = has_data_dir_path.clone();
+            let sync_data_shards = durability.syncs_data_shards();
+            let sync_commit_metadata = durability.syncs_commit_metadata();
+            let staging_cleanup = multipart_staging.as_ref().map(|s| s.cleanup_handle());
+            let dst_volume_str = dst_volume.to_string();
+            let dst_path_str = dst_path.to_string();
+            let fi_moved = fi;
+            #[cfg(test)]
+            let root_owned = self.root.clone();
+
+            let outcome = os::run_blocking_namespace_operation_with_sync(
+                mutation_lease.clone(),
+                true,
+                move || -> std::io::Result<MultipartFusedCommitOutcome> {
+                    let (directory_guard_owned, has_dst_buf_owned) = match prepare_rename_commit_and_destination_metadata_std(
+                        false,
+                        &src_parent_owned,
+                        &dst_parent_owned,
+                        &dst_file_path_owned,
+                        &dst_volume_dir_owned,
+                        &publication_root_owned,
+                    ) {
+                        Ok(res) => res,
+                        Err(err) => return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(to_file_error(err).into())),
+                    };
+                    let mut xlmeta = FileMeta::new();
+                    let mut dst_meta_unparsable = false;
+                    if let Some(dst_buf) = has_dst_buf_owned.as_ref() {
+                        if FileMeta::is_xl2_v1_format(dst_buf)
+                            && let Ok(nmeta) = FileMeta::load(dst_buf)
+                        {
+                            xlmeta = nmeta;
+                        } else {
+                            dst_meta_unparsable = true;
+                        }
+                    }
+                    let skip_parent_owned = if has_dst_buf_owned.is_some() {
+                        dst_parent_owned.clone()
+                    } else {
+                        dst_volume_dir_owned.clone()
+                    };
+                    let version_id = fi_moved.version_id.unwrap_or_default();
+                    let (old_current_size, old_current_source_checked, old_current_source) = observe_old_current(
+                        has_dst_buf_owned.is_some(),
+                        dst_meta_unparsable,
+                        &xlmeta,
+                        &dst_volume_str,
+                        &dst_path_str,
+                        version_id,
+                    );
+                    let has_old_data_dir = xlmeta.find_unshared_data_dir_for_version(Some(version_id));
+                    if let Some(old_data_dir) = has_old_data_dir.as_ref() {
+                        let _ = xlmeta.data.remove_two(version_id, *old_data_dir);
+                    }
+                    if let Err(err) = xlmeta.add_version(fi_moved) {
+                        return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(err.into()));
+                    }
+                    let version_signature = rename_data_versions_signature(&xlmeta);
+                    let new_dst_buf_owned = match xlmeta.marshal_msg() {
+                        Ok(buf) => buf,
+                        Err(err) => return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(err.into())),
+                    };
+
+                    let prepared_metadata_source = match publication_owned.prepare_and_install_head_on_thread(
+                        &new_dst_buf_owned,
+                        has_dst_buf_owned.as_deref(),
+                        &dst_file_path_owned,
+                        &directory_guard_owned,
+                    ) {
+                        Ok(source) => source,
+                        Err(err) => return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(err)),
+                    };
+                    #[cfg(test)]
+                    let probe_active = os::file_sync_probe::is_active();
+                    #[cfg(not(test))]
+                    let probe_active = false;
+                    // When presynced, shard payloads were synced during UploadPart and
+                    // prepare_and_install_head_on_thread just fsynced src_data_path after
+                    // writing RETIREMENT_ANCHOR.
+                    if sync_data_shards
+                        && let Some((src_data_path, _)) = has_data_dir_path_owned.as_ref()
+                        && (probe_active || !super::take_presynced_tmp_data_dir(src_data_path))
+                        && let Err(err) = os::sync_dir_files_std(src_data_path)
+                        && err.kind() != ErrorKind::NotFound
+                    {
+                        return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(to_file_error(err).into()));
+                    }
+                    let directory_guard_owned = match super::remove_dst_base_before_commit_std(
+                        &dst_path_str,
+                        directory_guard_owned,
+                        &src_parent_owned,
+                        &dst_parent_owned,
+                        &dst_volume_dir_owned,
+                        &publication_root_owned,
+                    ) {
+                        Ok(guard) => guard,
+                        Err(err) => return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(err)),
+                    };
+                    if should_remove_staged_meta_before_commit(&dst_path_str) {
+                        drop(prepared_metadata_source);
+                        if let Err(err) = std::fs::remove_file(&src_file_path_owned) {
+                            return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(to_file_error(err).into()));
+                        }
+                        return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(DiskError::FileNotFound));
+                    }
+                    if let Some((src_data_path, dst_data_path)) = has_data_dir_path_owned.as_ref() {
+                        let Some(src_parent) = src_data_path.parent() else {
+                            return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(DiskError::InvalidPath));
+                        };
+                        let payload_guard = match os::multipart_payload_commit_guard_std(&directory_guard_owned, src_parent) {
+                            Ok(guard) => guard,
+                            Err(err) => return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(to_file_error(err).into())),
+                        };
+                        if let Err(err) = os::rename_all_with_commit_guard_on_thread(src_data_path, dst_data_path, &payload_guard)
+                        {
+                            return Ok(MultipartFusedCommitOutcome::RestoreDataSource(directory_guard_owned, err));
+                        }
+                        #[cfg(test)]
+                        run_rename_data_after_first_publication(&root_owned, &dst_volume_str, &dst_path_str);
+                        if crash_inject::should_crash_at(CrashPoint::RenameAfterDataRename, &dst_path_str) {
+                            return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(DiskError::Unexpected));
+                        }
+                        if should_fail_before_old_metadata_backup(&dst_path_str) {
+                            return Ok(MultipartFusedCommitOutcome::RestoreDataSource(
+                                directory_guard_owned,
+                                DiskError::Unexpected,
+                            ));
+                        }
+                        if crash_inject::should_crash_at(CrashPoint::RenameAfterBackupBeforeMetaCommit, &dst_path_str) {
+                            return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(DiskError::Unexpected));
+                        }
+                    }
+                    let rename_res = if should_fail_commit_rename(&dst_path_str) {
+                        drop(prepared_metadata_source);
+                        Err(DiskError::other("test fail during metadata commit rename"))
+                    } else {
+                        os::rename_prepared_source_on_thread(
+                            prepared_metadata_source,
+                            &src_file_path_owned,
+                            &dst_file_path_owned,
+                            &skip_parent_owned,
+                            &publication_root_owned,
+                            &directory_guard_owned,
+                        )
+                    };
+                    if let Err(err) = rename_res {
+                        return Ok(MultipartFusedCommitOutcome::RestoreDataSource(directory_guard_owned, err));
+                    }
+                    if should_fail_after_metadata_commit(&dst_path_str) {
+                        return Ok(MultipartFusedCommitOutcome::RestoreTargetAfterError(DiskError::Unexpected, false));
+                    }
+                    if crash_inject::should_crash_at(CrashPoint::RenameAfterMetaCommit, &dst_path_str) {
+                        return Ok(MultipartFusedCommitOutcome::AbortBeforeDataRename(DiskError::Unexpected));
+                    }
+                    if sync_commit_metadata {
+                        if let Some(parent) = dst_file_path_owned.parent()
+                            && let Err(err) = os::fsync_commit_directory_limited_std(parent, &directory_guard_owned)
+                        {
+                            return Ok(MultipartFusedCommitOutcome::RestoreTargetAfterError(to_file_error(err).into(), true));
+                        }
+                        if has_dst_buf_owned.is_none() {
+                            let mut ancestor = dst_file_path_owned.parent().and_then(|object_dir| object_dir.parent());
+                            while let Some(dir) = ancestor {
+                                if !dir.starts_with(&dst_volume_dir_owned) {
+                                    break;
+                                }
+                                if let Err(err) = os::fsync_dir_limited_std(dir) {
+                                    return Ok(MultipartFusedCommitOutcome::RestoreTargetAfterError(
+                                        to_file_error(err).into(),
+                                        true,
+                                    ));
+                                }
+                                if dir == dst_volume_dir_owned.as_path() {
+                                    break;
+                                }
+                                ancestor = dir.parent();
+                            }
+                        }
+                    }
+                    match publication_owned.finish_on_thread(sync_commit_metadata, staging_cleanup) {
+                        Ok(result) => Ok(MultipartFusedCommitOutcome::Committed(Box::new(RenameDataResp {
+                            multipart_source: Some(Box::new(result)),
+                            old_data_dir: has_old_data_dir,
+                            rollback_data_dir: None,
+                            cleanup_data_dir: has_old_data_dir,
+                            sign: version_signature,
+                            old_current_size,
+                            old_current_source_checked,
+                            old_current_source,
+                        }))),
+                        Err(err) => Ok(MultipartFusedCommitOutcome::RestoreTargetAfterError(err, false)),
+                    }
+                },
+            )
+            .await
+            .map_err(to_file_error)
+            .map_err(DiskError::from)?;
+
+            match outcome {
+                MultipartFusedCommitOutcome::Committed(resp) => {
+                    for part_path in &invalidate_part_paths {
+                        self.io_backend.invalidate_cached_fd(dst_volume, part_path).await;
+                    }
+                    return Ok(*resp);
+                }
+                MultipartFusedCommitOutcome::AbortBeforeDataRename(err) => return Err(err),
+                MultipartFusedCommitOutcome::RestoreDataSource(directory_guard, err) => {
+                    let commit = RenameCommitContext {
+                        directory_guard,
+                        publication_root: &self.publication_root,
+                        mutation_lease,
+                        source_metadata: &src_file_path,
+                        destination_metadata: &dst_file_path,
+                        source_volume_dir: &src_volume_dir,
+                        data_paths: has_data_dir_path.as_ref(),
+                    };
+                    commit.restore_published_data_source().await?;
+                    return Err(err);
+                }
+                MultipartFusedCommitOutcome::RestoreTargetAfterError(err, invalidate_before_return) => {
+                    if !invalidate_before_return {
+                        for part_path in &invalidate_part_paths {
+                            self.io_backend.invalidate_cached_fd(dst_volume, part_path).await;
+                        }
+                    }
+                    pin_multipart_future(|| publication.restore_target_after_error(mutation_lease.clone())).await?;
+                    if invalidate_before_return {
+                        for part_path in &invalidate_part_paths {
+                            self.io_backend.invalidate_cached_fd(dst_volume, part_path).await;
+                        }
+                    }
+                    return Err(err);
+                }
+            }
        } else {
            prepare_rename_commit_and_destination_metadata(
                !no_inline && !preserve_upload_metadata,
@@ -1053,71 +1479,81 @@ impl LocalDisk {
            // after the commit rename, and a crash before the rename means the
            // PUT was never acknowledged. A metadata commit: relaxed tiers
            // leave it to the page cache.
-            if let Some(publication) = multipart_publication.as_ref() {
-                pin_multipart_future(|| publication.prepare_metadata(&new_dst_buf, has_dst_buf.as_deref())).await?;
-            }
-            let tmp_meta_sync = if durability.syncs_commit_metadata() || multipart_publication.is_some() {
-                SyncMode::FileOnly
-            } else {
-                SyncMode::None
-            };
-            // The tmp xl.meta write and the shard-file fdatasync are independent
-            // (disjoint paths) and both only need to be durable before the commit
-            // renames below, so run them concurrently to drop a blocking
-            // round-trip from the PUT commit critical path (rustfs/backlog#922
-            // step 2). The "contents durable -> rename -> dst dir fsync" ordering
-            // is unchanged — both futures complete before any rename — which the
-            // rename_data crash-consistency harness (backlog#935) exercises.
-            //
-            // Shard durability: once rename_data succeeds the write is
-            // acknowledged, so data must not live only in the page cache.
-            // Multipart parts were already synced during rename_part, so their
-            // fdatasync here is a cheap no-op. A missing source dir is left for the
-            // rename below to report through the existing rollback path. Payload
-            // durability is kept by both strict and relaxed.
-            let tmp_meta_write = {
-                let src_file_path = multipart_publication
-                    .as_ref()
-                    .map_or_else(|| commit.source_metadata.to_path_buf(), |publication| publication.prepared_path.clone());
-                let dst_file_path = dst_file_path.clone();
-                let owned_directory_guard = commit.directory_guard.clone();
-                let owned_mutation_lease = commit.mutation_lease.clone();
-                async move {
-                    os::run_blocking_namespace_operation(owned_mutation_lease, move || {
-                        #[cfg(test)]
-                        run_owned_file_before_open(&src_file_path);
-                        let mut prepared_metadata_source = os::create_prepared_rename_source_with_commit_guard(
-                            &src_file_path,
-                            &dst_file_path,
-                            &owned_directory_guard,
-                        )?;
-                        prepared_metadata_source.write_all(&new_dst_buf, tmp_meta_sync != SyncMode::None)?;
-                        Ok(prepared_metadata_source)
-                    })
-                    .await
-                    .map_err(to_file_error)
-                    .map_err(DiskError::from)
-                }
-            };
+            #[cfg(all(test, target_os = "linux"))]
+            let probe_active = os::file_sync_probe::is_active();
+            #[cfg(all(not(test), target_os = "linux"))]
+            let probe_active = false;
            let shard_sync = async {
                if durability.syncs_data_shards()
                    && let Some((src_data_path, _)) = has_data_dir_path.as_ref()
-                    && let Err(err) = os::sync_dir_files_with_limiter(src_data_path, self.file_sync_permits.clone()).await
-                    && err.kind() != ErrorKind::NotFound
                {
-                    return Err::<(), DiskError>(to_file_error(err).into());
+                    #[cfg(target_os = "linux")]
+                    if !probe_active && super::take_presynced_tmp_data_dir(src_data_path) {
+                        if multipart_publication.is_none()
+                            && let Err(err) = os::fsync_dir_with_owner(src_data_path, Some(commit.mutation_lease.clone())).await
+                            && err.kind() != ErrorKind::NotFound
+                        {
+                            return Err::<(), DiskError>(to_file_error(err).into());
+                        }
+                        return Ok(());
+                    }
+                    if let Err(err) = os::sync_dir_files_with_limiter(src_data_path, self.file_sync_permits.clone()).await
+                        && err.kind() != ErrorKind::NotFound
+                    {
+                        return Err::<(), DiskError>(to_file_error(err).into());
+                    }
                }
                Ok(())
            };
-            let (tmp_meta_res, shard_sync_res) = tokio::join!(tmp_meta_write, shard_sync);
-            // Surface a tmp-meta failure first (its prior serial position), then a
-            // shard-sync failure; either aborts before any rename, exactly as the
-            // sequential version did.
-            let prepared_metadata_source = tmp_meta_res?;
-            shard_sync_res?;
-            let prepared_metadata_source = match multipart_publication.as_ref() {
-                Some(publication) => pin_multipart_future(|| publication.install_head(prepared_metadata_source)).await?,
-                None => prepared_metadata_source,
+            let prepared_metadata_source = if let Some(publication) = multipart_publication.as_ref() {
+                let dst_file_path_ref = &dst_file_path;
+                let dir_guard_ref = &commit.directory_guard;
+                let lease_clone = commit.mutation_lease.clone();
+                let prepare_fut = pin_multipart_future(|| {
+                    publication.prepare_and_install_head(
+                        &new_dst_buf,
+                        has_dst_buf.as_deref(),
+                        dst_file_path_ref,
+                        dir_guard_ref,
+                        lease_clone,
+                    )
+                });
+                let (tmp_meta_res, shard_sync_res) = tokio::join!(prepare_fut, shard_sync);
+                let prepared_metadata_source = tmp_meta_res?;
+                shard_sync_res?;
+                prepared_metadata_source
+            } else {
+                let tmp_meta_sync = if durability.syncs_commit_metadata() {
+                    SyncMode::FileOnly
+                } else {
+                    SyncMode::None
+                };
+                let tmp_meta_write = {
+                    let src_file_path = commit.source_metadata.to_path_buf();
+                    let dst_file_path = dst_file_path.clone();
+                    let owned_directory_guard = commit.directory_guard.clone();
+                    let owned_mutation_lease = commit.mutation_lease.clone();
+                    async move {
+                        os::run_blocking_namespace_operation(owned_mutation_lease, move || {
+                            #[cfg(test)]
+                            run_owned_file_before_open(&src_file_path);
+                            let mut prepared_metadata_source = os::create_prepared_rename_source_with_commit_guard(
+                                &src_file_path,
+                                &dst_file_path,
+                                &owned_directory_guard,
+                            )?;
+                            prepared_metadata_source.write_all(&new_dst_buf, tmp_meta_sync != SyncMode::None)?;
+                            Ok(prepared_metadata_source)
+                        })
+                        .await
+                        .map_err(to_file_error)
+                        .map_err(DiskError::from)
+                    }
+                };
+                let (tmp_meta_res, shard_sync_res) = tokio::join!(tmp_meta_write, shard_sync);
+                let prepared_metadata_source = tmp_meta_res?;
+                shard_sync_res?;
+                prepared_metadata_source
            };
            commit.directory_guard = remove_dst_base_before_commit(
                dst_path,
@@ -1457,13 +1893,22 @@ impl LocalDisk {
            }

            let multipart_source = match multipart_publication.as_ref() {
-                Some(publication) => match pin_multipart_future(|| publication.finish(commit.mutation_lease.clone())).await {
-                    Ok(result) => Some(Box::new(result)),
-                    Err(error) => {
-                        pin_multipart_future(|| publication.restore_target_after_error(commit.mutation_lease.clone())).await?;
-                        return Err(error);
+                Some(publication) => {
+                    let target_already_synced = durability.syncs_commit_metadata();
+                    let staging_cleanup = multipart_staging.as_ref().map(|s| s.cleanup_handle());
+                    match pin_multipart_future(|| {
+                        publication.finish(commit.mutation_lease.clone(), target_already_synced, staging_cleanup)
+                    })
+                    .await
+                    {
+                        Ok(result) => Some(Box::new(result)),
+                        Err(error) => {
+                            pin_multipart_future(|| publication.restore_target_after_error(commit.mutation_lease.clone()))
+                                .await?;
+                            return Err(error);
+                        }
                    }
-                },
+                }
                None => None,
            };
            Ok(RenameDataResp {
@@ -1663,13 +2108,8 @@ impl LocalDisk {
                    #[cfg(not(windows))]
                    let backup_path = {
                        let _backup = os::StageTimer::new(native_timing, "pool_meta_native_backup_create");
-                        create_local_inline_rollback_backup(&dst, &src, old_metadata)?
+                        create_local_inline_rollback_backup(&dst, &src, old_metadata, sync)?
                    };
-                    #[cfg(not(windows))]
-                    if sync {
-                        let _backup_sync = os::StageTimer::new(native_timing, "pool_meta_native_backup_file_sync");
-                        std::fs::File::open(&backup_path)?.sync_data()?;
-                    }
                    staged_rollback_path = Some(StagedRollbackBackup(backup_path));
                }

@@ -1000,6 +1000,25 @@ pub(super) fn inspection_with_markers(
    Ok((inspection, has_physical_data_dirs_or_markers))
 }

+type PreparedDeleteReceiptEntry = (DeleteRecoveryJournal, bool);
+
+static PREPARED_DELETE_RECEIPTS: std::sync::LazyLock<parking_lot::Mutex<HashMap<PathBuf, PreparedDeleteReceiptEntry>>> =
+    std::sync::LazyLock::new(|| parking_lot::Mutex::new(HashMap::new()));
+
+fn put_prepared_delete_receipt(rollback: PathBuf, entry: PreparedDeleteReceiptEntry) {
+    let mut map = PREPARED_DELETE_RECEIPTS.lock();
+    if map.len() >= 4096
+        && let Some(evict_key) = map.keys().next().cloned()
+    {
+        map.remove(&evict_key);
+    }
+    map.insert(rollback, entry);
+}
+
+fn take_prepared_delete_receipt(rollback: &Path) -> Option<PreparedDeleteReceiptEntry> {
+    PREPARED_DELETE_RECEIPTS.lock().remove(rollback)
+}
+
 impl LocalDisk {
    // Native MPU source closure has its own conditional protocol. Generic
    // object recovery must never restore its head or acquire SourceRW under
@@ -1523,6 +1542,9 @@ impl LocalDisk {
        owner: Option<Arc<dyn Send + Sync>>,
    ) -> Result<DeleteRecoveryInspection> {
        self.reject_multipart_delete_recovery_mutation(volume, object)?;
+        if let Ok(root) = self.io_get_object_path(volume, object) {
+            let _ = take_prepared_delete_receipt(&root.join(transaction.to_string()));
+        }

        let (mut observed, mut has_data_dir_work) = self
            .delete_recovery_inspect_owned_with_markers(volume, object, transaction, owner.clone())
@@ -1911,8 +1933,14 @@ impl LocalDisk {
                    return Err(DiskError::DiskOngoingReq);
                }
                let bucket_path = self.io_get_bucket_path(volume)?;
-                let mut file_leases = Vec::with_capacity(3);
-                for name in [TARGET, PREPARED, FINISHED] {
+                let mut file_leases = Vec::with_capacity(5);
+                for name in [
+                    STORAGE_FORMAT_FILE_BACKUP,
+                    DELETE_MARKER_ROLLBACK_FILE,
+                    TARGET,
+                    PREPARED,
+                    FINISHED,
+                ] {
                    let path = rollback.join(name);
                    let lease = os::acquire_namespace_mutation_lease_with_owner(&path, owner.clone()).await;
                    file_leases.push((path, lease));
@@ -2132,119 +2160,151 @@ impl LocalDisk {
                    Err(err) if err.kind() == ErrorKind::NotFound => return Ok(Ok(None)),
                    Err(err) => return Err(err),
                };
-                let (old_dir, journal_raw) = match (|| -> Result<Option<(Option<Uuid>, Vec<u8>)>> {
-                    let mut meta = FileMeta::load(&buf)?;
-                    let deleted_versions: Vec<_> = meta
-                        .versions
-                        .iter()
-                        .map(|v| v.header.version_id.unwrap_or_default())
-                        .collect();
-                    let original_versions: Vec<_> = meta
-                        .versions
-                        .iter()
-                        .map(|v| generation(v.into_fileinfo(&volume_str, &object_str, true)?))
-                        .collect::<Result<_>>()?;
-                    let old_dir = meta.delete_version(&fi_clone)?;
-                    if old_dir.is_some_and(|dir| dir == transaction || dir.is_nil()) {
-                        return Err(DiskError::FileCorrupt);
-                    }
-                    if let Some(uuid) = old_dir {
-                        meta.data.remove(vec![fi_clone.version_id.unwrap_or_default(), uuid])?;
-                        let old_path = path_join(&[file_path.as_path(), Path::new(uuid.to_string().as_str())]);
-                        check_path_length(old_path.to_string_lossy().as_ref())?;
-                        if old_path.join(STORAGE_FORMAT_FILE).exists() {
-                            return Ok(None);
+                type PreparedDeleteTuple = (Option<Uuid>, Option<Vec<u8>>, Vec<u8>, DeleteRecoveryJournal, bool);
+                let (old_dir, target, journal_raw, journal, sole_transaction) =
+                    match (|| -> Result<Option<PreparedDeleteTuple>> {
+                        let mut meta = FileMeta::load(&buf)?;
+                        let mut deleted_headers: Vec<_> = meta
+                            .versions
+                            .iter()
+                            .map(|v| (v.header.version_id, std::mem::discriminant(&v.header.version_type)))
+                            .collect();
+                        let original_versions: Vec<_> = meta
+                            .versions
+                            .iter()
+                            .map(|v| generation(v.into_fileinfo(&volume_str, &object_str, true)?))
+                            .collect::<Result<_>>()?;
+                        let old_dir = meta.delete_version(&fi_clone)?;
+                        if old_dir.is_some_and(|dir| dir == transaction || dir.is_nil()) {
+                            return Err(DiskError::FileCorrupt);
                        }
-                    }
-                    if !meta.versions.is_empty() {
-                        return Ok(None);
-                    }
-                    let mut predecessors = BTreeSet::new();
-                    if let Ok(entries) = std::fs::read_dir(&file_path) {
-                        for (visited, entry) in entries.enumerate() {
-                            if visited >= 65536 {
-                                return Err(DiskError::DiskOngoingReq);
+                        if let Some(uuid) = old_dir {
+                            meta.data.remove(vec![fi_clone.version_id.unwrap_or_default(), uuid])?;
+                            let old_path = path_join(&[file_path.as_path(), Path::new(uuid.to_string().as_str())]);
+                            check_path_length(old_path.to_string_lossy().as_ref())?;
+                            if old_path.join(STORAGE_FORMAT_FILE).exists() {
+                                return Ok(None);
                            }
-                            let entry = entry.map_err(to_file_error)?;
-                            let Some(id) = entry
-                                .file_name()
-                                .to_str()
-                                .and_then(|name| Uuid::parse_str(name).ok())
-                                .filter(|id| !id.is_nil() && *id != transaction)
-                            else {
-                                continue;
-                            };
-                            if !entry.file_type().map_err(to_file_error)?.is_dir() {
-                                return Err(DiskError::InvalidPath);
-                            }
-                            let prepared = regular_bytes(&entry.path().join(PREPARED))?;
-                            let receipt = regular_bytes(&entry.path().join(DELETE_QUORUM_COMMITTED_FILE))?;
-                            let mut records = Vec::new();
-                            if let Some(bytes) = prepared {
-                                records.push(
-                                    serde_json::from_slice::<DeleteRecoveryJournal>(&bytes)
-                                        .map_err(|_| DiskError::FileCorrupt)?,
-                                );
-                            }
-                            if let Some(bytes) = receipt.filter(|bytes| !bytes.is_empty()) {
-                                records.push(
-                                    serde_json::from_slice::<CommitReceipt>(&bytes)
-                                        .map_err(|_| DiskError::FileCorrupt)?
-                                        .journal,
-                                );
-                            }
-                            for record in records {
-                                if record.protocol_version != DELETE_RECOVERY_PROTOCOL_VERSION
-                                    || record.volume != volume_str
-                                    || record.object != object_str
-                                    || record.transaction != id
-                                    || incarnation.zip(record.bucket_incarnation).is_some_and(|(a, b)| a != b)
-                                    || record.predecessors.iter().any(|before| before.is_nil() || *before == id)
-                                {
-                                    return Err(DiskError::FileCorrupt);
+                        }
+                        let surviving: HashSet<_> = meta
+                            .versions
+                            .iter()
+                            .map(|v| (v.header.version_id, std::mem::discriminant(&v.header.version_type)))
+                            .collect();
+                        deleted_headers.retain(|header| !surviving.contains(header));
+                        let deleted_versions = deleted_headers.into_iter().map(|(id, _)| id.unwrap_or_default()).collect();
+                        let target = (!meta.versions.is_empty()).then(|| meta.marshal_msg()).transpose()?;
+                        let mut predecessors = BTreeSet::new();
+                        let mut sole_transaction = true;
+                        if let Ok(entries) = std::fs::read_dir(&file_path) {
+                            for (visited, entry) in entries.enumerate() {
+                                if visited >= 65536 {
+                                    return Err(DiskError::DiskOngoingReq);
+                                }
+                                let entry = entry.map_err(to_file_error)?;
+                                let file_name = entry.file_name();
+                                let Some(name_str) = file_name.to_str() else {
+                                    sole_transaction = false;
+                                    continue;
+                                };
+                                let Some(id) = Uuid::parse_str(name_str).ok().filter(|id| !id.is_nil() && *id != transaction)
+                                else {
+                                    continue;
+                                };
+                                if !entry.file_type().map_err(to_file_error)?.is_dir() {
+                                    return Err(DiskError::InvalidPath);
+                                }
+                                let dir_path = entry.path();
+                                let prepared = regular_bytes(&dir_path.join(PREPARED))?;
+                                let receipt = regular_bytes(&dir_path.join(DELETE_QUORUM_COMMITTED_FILE))?;
+                                if sole_transaction && name_str == id.hyphenated().to_string() && old_dir != Some(id) {
+                                    let is_surviving_data_dir = original_versions.iter().any(|v| v.data_dir == Some(id))
+                                        && prepared.is_none()
+                                        && receipt.is_none()
+                                        && !dir_path.join(STORAGE_FORMAT_FILE_BACKUP).exists();
+                                    if !is_surviving_data_dir {
+                                        sole_transaction = false;
+                                    }
+                                }
+                                let mut records = Vec::new();
+                                if let Some(bytes) = prepared {
+                                    records.push(
+                                        serde_json::from_slice::<DeleteRecoveryJournal>(&bytes)
+                                            .map_err(|_| DiskError::FileCorrupt)?,
+                                    );
+                                }
+                                if let Some(bytes) = receipt.filter(|bytes| !bytes.is_empty()) {
+                                    records.push(
+                                        serde_json::from_slice::<CommitReceipt>(&bytes)
+                                            .map_err(|_| DiskError::FileCorrupt)?
+                                            .journal,
+                                    );
+                                }
+                                for record in records {
+                                    if record.protocol_version != DELETE_RECOVERY_PROTOCOL_VERSION
+                                        || record.volume != volume_str
+                                        || record.object != object_str
+                                        || record.transaction != id
+                                        || incarnation.zip(record.bucket_incarnation).is_some_and(|(a, b)| a != b)
+                                        || record.predecessors.iter().any(|before| before.is_nil() || *before == id)
+                                    {
+                                        return Err(DiskError::FileCorrupt);
+                                    }
+                                    predecessors.insert(id);
                                }
-                                predecessors.insert(id);
                            }
                        }
-                    }
-                    let mut sparse = fi_clone;
-                    sparse.data = None;
-                    let journal = DeleteRecoveryJournal {
-                        protocol_version: DELETE_RECOVERY_PROTOCOL_VERSION,
-                        volume: volume_str,
-                        object: object_str,
-                        transaction,
-                        bucket_incarnation: incarnation,
-                        original_digest: Some(digest(&buf)),
-                        target_digest: None,
-                        data_dirs: old_dir.into_iter().collect(),
-                        deleted_versions,
-                        requests: vec![sparse],
-                        original_versions,
-                        predecessors: predecessors.into_iter().collect(),
+                        sole_transaction &= predecessors.is_empty();
+                        let mut sparse = fi_clone;
+                        sparse.data = None;
+                        let journal = DeleteRecoveryJournal {
+                            protocol_version: DELETE_RECOVERY_PROTOCOL_VERSION,
+                            volume: volume_str,
+                            object: object_str,
+                            transaction,
+                            bucket_incarnation: incarnation,
+                            original_digest: Some(digest(&buf)),
+                            target_digest: target.as_deref().map(digest),
+                            data_dirs: old_dir.into_iter().collect(),
+                            deleted_versions,
+                            requests: vec![sparse],
+                            original_versions,
+                            predecessors: predecessors.into_iter().collect(),
+                        };
+                        let raw = journal_bytes(&journal)?;
+                        Ok(Some((old_dir, target, raw, journal, sole_transaction)))
+                    })() {
+                        Ok(Some(tuple)) => tuple,
+                        Ok(None) => return Ok(Ok(None)),
+                        Err(err) => return Ok(Err(err)),
                    };
-                    Ok(Some((old_dir, journal_bytes(&journal)?)))
-                })() {
-                    Ok(Some(pair)) => pair,
-                    Ok(None) => return Ok(Ok(None)),
-                    Err(err) => return Ok(Err(err)),
-                };
                std::fs::create_dir_all(&staging_root)?;
-                let staged = tempfile::Builder::new()
-                    .prefix("delete-preparation-")
-                    .tempdir_in(&staging_root)?;
-                for (name, bytes) in [
-                    (STORAGE_FORMAT_FILE_BACKUP, buf.as_slice()),
-                    (PREPARED, journal_raw.as_slice()),
-                ] {
-                    let mut file = std::fs::File::create(staged.path().join(name))?;
-                    std::io::Write::write_all(&mut file, bytes)?;
-                    if sync && !bytes.is_empty() {
+                let staged = tempfile::Builder::new().prefix("delete-prep-").tempdir_in(&staging_root)?;
+                let staged_dir = staged.path().to_path_buf();
+                let backup_path = staged_dir.join(STORAGE_FORMAT_FILE_BACKUP);
+                if std::fs::hard_link(&xl_path, &backup_path).is_err() {
+                    let mut file = std::fs::File::create(&backup_path)?;
+                    std::io::Write::write_all(&mut file, &buf)?;
+                    if sync && !buf.is_empty() {
                        file.sync_data()?;
                    }
                }
+                let mut written_files = Vec::with_capacity(2);
+                for (name, bytes) in [(TARGET, target.as_deref()), (PREPARED, Some(journal_raw.as_slice()))] {
+                    let Some(bytes) = bytes else {
+                        continue;
+                    };
+                    let mut file = std::fs::File::create(staged_dir.join(name))?;
+                    std::io::Write::write_all(&mut file, bytes)?;
+                    if sync && !bytes.is_empty() {
+                        written_files.push(file);
+                    }
+                }
+                for file in written_files {
+                    file.sync_data()?;
+                }
                if sync {
-                    os::fsync_dir_std(staged.path())?;
+                    os::fsync_dir_std(&staged_dir)?;
                }
                match std::fs::symlink_metadata(&destination) {
                    Ok(metadata) if metadata.is_dir() => match std::fs::remove_dir(&destination) {
@@ -2257,27 +2317,62 @@ impl LocalDisk {
                    Err(err) if err.kind() == ErrorKind::NotFound => {}
                    Err(err) => return Err(err),
                }
-                if let Some(parent) = destination.parent() {
-                    std::fs::create_dir_all(parent)?;
+                let staged_dir = staged.keep();
+                if let Err(err) = std::fs::rename(&staged_dir, &destination) {
+                    let _ = std::fs::remove_dir_all(&staged_dir);
+                    return Err(err);
                }
-                std::fs::rename(staged.path(), &destination)?;
-                if sync {
-                    os::fsync_dir_std(&file_path)?;
+                let old_data_path = old_dir.map(|uuid| file_path.join(uuid.to_string())).filter(|p| p.is_dir());
+                if sync && let Err(err) = os::fsync_dir_std(&file_path) {
+                    let _ = std::fs::remove_dir_all(&destination);
+                    return Err(err);
                }
                let mut reserved_data_path = None;
-                if let Some(uuid) = old_dir {
-                    let old_path = file_path.join(uuid.to_string());
-                    if old_path.is_dir() {
-                        let reserved_marker = old_path.join(format!("{RESERVED_DELETE_DATA_DIR_MARKER_PREFIX}{transaction}"));
+                if let Some(old_path) = old_data_path {
+                    let reserved_marker = old_path.join(format!("{RESERVED_DELETE_DATA_DIR_MARKER_PREFIX}{transaction}"));
+                    let reserve_res = (|| -> std::io::Result<()> {
                        let marker = std::fs::File::create(&reserved_marker)?;
                        if sync {
                            marker.sync_all()?;
                            os::fsync_dir_std(&old_path)?;
                        }
-                        reserved_data_path = Some((old_path, reserved_marker));
+                        Ok(())
+                    })();
+                    if let Err(err) = reserve_res {
+                        let _ = std::fs::remove_file(&reserved_marker);
+                        let _ = std::fs::remove_dir_all(&destination);
+                        return Err(err);
                    }
+                    reserved_data_path = Some((old_path, reserved_marker));
                }
-                if let Err(err) = std::fs::remove_file(&xl_path)
+                if let Some(ref target_bytes) = target {
+                    let target_in_rollback = destination.join(TARGET);
+                    let linked_xl = staging_root.join(format!("delete-xl-{transaction}"));
+                    let _ = std::fs::remove_file(&linked_xl);
+                    let write_res = if std::fs::hard_link(&target_in_rollback, &linked_xl).is_ok() {
+                        let rename_res = std::fs::rename(&linked_xl, &xl_path);
+                        if rename_res.is_err() {
+                            let _ = std::fs::remove_file(&linked_xl);
+                        }
+                        rename_res
+                    } else {
+                        (|| -> std::io::Result<()> {
+                            let mut staged_xl = tempfile::NamedTempFile::new_in(&staging_root)?;
+                            std::io::Write::write_all(staged_xl.as_file_mut(), target_bytes)?;
+                            if sync && !target_bytes.is_empty() {
+                                staged_xl.as_file().sync_data()?;
+                            }
+                            std::fs::rename(staged_xl.path(), &xl_path)
+                        })()
+                    };
+                    if let Err(err) = write_res {
+                        if let Some((_, reserved_marker)) = &reserved_data_path {
+                            let _ = std::fs::remove_file(reserved_marker);
+                        }
+                        let _ = std::fs::remove_dir_all(&destination);
+                        return Err(err);
+                    }
+                } else if let Err(err) = std::fs::remove_file(&xl_path)
                    && err.kind() != ErrorKind::NotFound
                {
                    if let Some((_, reserved_marker)) = &reserved_data_path {
@@ -2298,6 +2393,7 @@ impl LocalDisk {
                        Err(err) => return Err(err),
                    }
                }
+                put_prepared_delete_receipt(destination, (journal, sole_transaction));
                Ok(Ok(Some(old_dir)))
            },
        )
@@ -2343,6 +2439,26 @@ impl LocalDisk {
            lease,
            sync,
            move || -> std::io::Result<Result<Option<(DeleteRecoveryJournal, bool, bool)>>> {
+                if let Some((journal, sole_transaction)) = take_prepared_delete_receipt(&rollback)
+                    && rollback.join(PREPARED).is_file()
+                {
+                    let receipt = CommitReceipt {
+                        journal: journal.clone(),
+                        finished: false,
+                    };
+                    let receipt_bytes = serde_json::to_vec(&receipt).map_err(std::io::Error::other)?;
+                    std::fs::create_dir_all(&staging_root)?;
+                    let mut staged = tempfile::NamedTempFile::new_in(&staging_root)?;
+                    std::io::Write::write_all(staged.as_file_mut(), &receipt_bytes)?;
+                    if sync && !receipt_bytes.is_empty() {
+                        staged.as_file().sync_data()?;
+                    }
+                    std::fs::rename(staged.path(), &receipt_path)?;
+                    if sync {
+                        os::fsync_dir_std(&rollback)?;
+                    }
+                    return Ok(Ok(Some((journal, sole_transaction, true))));
+                }
                let prepared = match regular_bytes(&rollback.join(PREPARED)) {
                    Ok(Some(bytes)) => bytes,
                    Ok(None) => return Ok(Ok(None)),
@@ -2360,12 +2476,18 @@ impl LocalDisk {
                {
                    return Ok(Err(DiskError::FileCorrupt));
                }
-                let current_exists = match regular_file_exists(&root.join(STORAGE_FORMAT_FILE)) {
-                    Ok(exists) => exists,
-                    Err(e) => return Ok(Err(e)),
+                let target_satisfied = match journal.target_digest {
+                    None => match regular_file_exists(&root.join(STORAGE_FORMAT_FILE)) {
+                        Ok(exists) => !exists,
+                        Err(e) => return Ok(Err(e)),
+                    },
+                    Some(expected_target) => match regular_bytes(&root.join(STORAGE_FORMAT_FILE)) {
+                        Ok(Some(bytes)) => digest(&bytes) == expected_target,
+                        Ok(None) => false,
+                        Err(e) => return Ok(Err(e)),
+                    },
                };
-                let unversioned_target_absent = journal.target_digest.is_none() && !current_exists;
-                if !unversioned_target_absent {
+                if !target_satisfied {
                    return Ok(Ok(None));
                }
                let mut sole_transaction = journal.predecessors.is_empty();
@@ -2382,8 +2504,15 @@ impl LocalDisk {
                                    && id != transaction
                                    && !journal.data_dirs.contains(&id)
                                {
-                                    sole_transaction = false;
-                                    break;
+                                    let dir_path = entry.path();
+                                    let is_surviving_data_dir = journal.original_versions.iter().any(|v| v.data_dir == Some(id))
+                                        && !dir_path.join(PREPARED).exists()
+                                        && !dir_path.join(DELETE_QUORUM_COMMITTED_FILE).exists()
+                                        && !dir_path.join(STORAGE_FORMAT_FILE_BACKUP).exists();
+                                    if !is_surviving_data_dir {
+                                        sole_transaction = false;
+                                        break;
+                                    }
                                }
                            }
                        }
@@ -2406,7 +2535,7 @@ impl LocalDisk {
                if sync {
                    os::fsync_dir_std(&rollback)?;
                }
-                Ok(Ok(Some((journal, sole_transaction, unversioned_target_absent))))
+                Ok(Ok(Some((journal, sole_transaction, target_satisfied))))
            },
        )
        .await
@@ -2458,12 +2587,15 @@ impl LocalDisk {
        }

        let bucket_path = self.io_get_bucket_path(volume)?;
+        let staging_root = self.io_get_bucket_path(RUSTFS_META_TMP_BUCKET)?;
        let root = self.io_get_object_path(volume, object)?;
        let rollback = root.join(transaction.to_string());
        let receipt_path = rollback.join(DELETE_QUORUM_COMMITTED_FILE);
        let sync = effective_durability(volume).syncs_commit_metadata();
-        let lease = os::acquire_namespace_mutation_lease_with_owner(&receipt_path, owner).await;
        let data_dirs = journal.data_dirs.clone();
+        let has_target = journal.target_digest.is_some();
+        let journal_for_receipt = journal.clone();
+        let lease = os::acquire_namespace_mutation_lease_with_owner(&receipt_path, owner).await;
        let prune_paths: Vec<_> = rollback
            .ancestors()
            .take_while(|dir| *dir != bucket_path.as_path() && dir.starts_with(&bucket_path))
@@ -2474,11 +2606,13 @@ impl LocalDisk {
            let committed = format!("{DELETE_DATA_DIR_MARKER_PREFIX}{transaction}");
            let recovery_committed = format!("{RECOVERY_DELETE_DATA_DIR_MARKER_PREFIX}{transaction}");
            let reserved = format!("{RESERVED_DELETE_DATA_DIR_MARKER_PREFIX}{transaction}");
+            let mut had_physical_data_dirs = false;
            for data_dir in &data_dirs {
                for base in [&root, &rollback] {
                    let dir_path = base.join(data_dir.to_string());
                    match std::fs::read_dir(&dir_path) {
                        Ok(entries) => {
+                            had_physical_data_dirs = true;
                            let mut removable = Vec::new();
                            let mut preserve_dir = false;
                            for entry in entries {
@@ -2519,23 +2653,45 @@ impl LocalDisk {
                    }
                }
            }
-            if !data_dirs.is_empty() && sync && root.is_dir() {
+            if had_physical_data_dirs && sync && root.is_dir() {
                os::fsync_dir_std(&root)?;
            }
-            for name in [
-                STORAGE_FORMAT_FILE_BACKUP,
-                DELETE_MARKER_ROLLBACK_FILE,
-                TARGET,
-                PREPARED,
-                FINISHED,
-            ] {
+            for name in [STORAGE_FORMAT_FILE_BACKUP, DELETE_MARKER_ROLLBACK_FILE] {
                ignore_not_found(std::fs::remove_file(rollback.join(name)))?;
            }
+            if has_target || had_physical_data_dirs {
+                let finished_receipt = CommitReceipt {
+                    journal: journal_for_receipt,
+                    finished: true,
+                };
+                let receipt_bytes = serde_json::to_vec(&finished_receipt).map_err(std::io::Error::other)?;
+                std::fs::create_dir_all(&staging_root)?;
+                let mut staged_receipt = tempfile::NamedTempFile::new_in(&staging_root)?;
+                std::io::Write::write_all(staged_receipt.as_file_mut(), &receipt_bytes)?;
+                if sync && !receipt_bytes.is_empty() {
+                    staged_receipt.as_file().sync_data()?;
+                }
+                std::fs::rename(staged_receipt.path(), &receipt_path)?;
+                let _ = std::fs::File::create(rollback.join(FINISHED))?;
+                if sync {
+                    os::fsync_dir_std(&rollback)?;
+                }
+            }
+            for name in [TARGET, PREPARED, FINISHED] {
+                let path = rollback.join(name);
+                #[cfg(all(any(test, feature = "test-util"), not(windows)))]
+                os::prepared_publication_test_hooks::run(os::prepared_publication_test_hooks::Stage::Remove, &path);
+                ignore_not_found(std::fs::remove_file(path))?;
+            }
            if sync {
                os::fsync_dir_std(&rollback)?;
            }
+            #[cfg(all(any(test, feature = "test-util"), not(windows)))]
+            os::prepared_publication_test_hooks::run(os::prepared_publication_test_hooks::Stage::Remove, &receipt_path);
            ignore_not_found(std::fs::remove_file(&receipt_path))?;
            for dir_path in &prune_paths {
+                #[cfg(all(any(test, feature = "test-util"), not(windows)))]
+                os::prepared_publication_test_hooks::run(os::prepared_publication_test_hooks::Stage::Remove, dir_path);
                if let Err(err) = std::fs::remove_dir(dir_path) {
                    if is_dir_not_empty_error(&err) {
                        break;
@@ -6282,6 +6438,214 @@ mod tests {
            b"outside bytes"
        );
    }
+
+    #[tokio::test]
+    async fn delete_recovery_fused_versioned_delete_marker_and_publish_fast_path() {
+        let (_temp, disk, info, _) = fixture().await;
+        let root = disk.io_get_object_path("recovery", "object").expect("object root");
+        let transaction = Uuid::new_v4();
+        let mut marker = FileInfo {
+            name: "object".to_owned(),
+            version_id: Some(Uuid::new_v4()),
+            deleted: true,
+            mark_deleted: true,
+            mod_time: info.mod_time.map(|t| t + time::Duration::seconds(1)),
+            ..Default::default()
+        };
+        let opts = DeleteOptions {
+            old_data_dir: Some(transaction),
+            recoverable_delete: true,
+            ..Default::default()
+        };
+        let fused = disk
+            .try_fused_recoverable_delete_version("recovery", "object", &marker, &opts, transaction, None)
+            .await
+            .expect("fused delete marker preparation must succeed");
+        assert_eq!(fused, Some(()), "versioned delete marker must use fused path");
+        assert!(root.join(transaction.to_string()).join(TARGET).is_file());
+        assert!(root.join(STORAGE_FORMAT_FILE).is_file());
+
+        let published = disk
+            .try_publish_prepared_delete_local("recovery", "object", transaction, None)
+            .await
+            .expect("local publish must succeed")
+            .expect("versioned target must satisfy local publish");
+        assert!(published.1, "sole transaction must be true");
+        assert!(published.2, "target_satisfied must be true");
+
+        let finalized = disk
+            .try_finish_and_finalize_prepared_delete_local("recovery", "object", transaction, &published.0, None)
+            .await
+            .expect("local finalize must succeed");
+        assert!(finalized);
+        assert!(!root.join(transaction.to_string()).exists());
+        let final_meta = FileMeta::load(&fs::read(root.join(STORAGE_FORMAT_FILE)).await.expect("read final xl.meta"))
+            .expect("decode final xl.meta");
+        assert_eq!(final_meta.versions.len(), 2);
+        assert_eq!(final_meta.versions[0].header.version_id, marker.version_id);
+        marker.deleted = false;
+    }
+
+    #[tokio::test]
+    async fn delete_recovery_fused_versioned_delete_with_surviving_version_and_interrupted_target_cleanup() {
+        let (_temp, disk, mut v1, _) = fixture().await;
+        let root = disk.io_get_object_path("recovery", "object").expect("object path");
+
+        // Configure v1 and v2 with distinct physical data directories so deleting v2
+        // exercises both `old_dir` reservation and `is_surviving_data_dir`.
+        v1.data = None;
+        rustfs_utils::http::metadata_compat::remove_str(
+            &mut v1.metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_INLINE_DATA,
+        );
+        let v1_dir = Uuid::new_v4();
+        v1.data_dir = Some(v1_dir);
+        fs::create_dir_all(root.join(v1_dir.to_string()))
+            .await
+            .expect("create v1 dir");
+        fs::write(root.join(v1_dir.to_string()).join("part.1"), b"v1-data")
+            .await
+            .expect("write v1 part");
+
+        let mut v2 = v1.clone();
+        v2.version_id = Some(Uuid::new_v4());
+        let v2_dir = Uuid::new_v4();
+        v2.data_dir = Some(v2_dir);
+        fs::create_dir_all(root.join(v2_dir.to_string()))
+            .await
+            .expect("create v2 dir");
+        fs::write(root.join(v2_dir.to_string()).join("part.1"), b"v2-data")
+            .await
+            .expect("write v2 part");
+
+        let mut meta = FileMeta::new();
+        meta.add_version(v1.clone()).expect("add v1");
+        meta.add_version(v2.clone()).expect("add v2");
+        fs::write(root.join(STORAGE_FORMAT_FILE), meta.marshal_msg().expect("marshal meta"))
+            .await
+            .expect("write xl.meta");
+
+        let transaction = Uuid::new_v4();
+        let opts = DeleteOptions {
+            recoverable_delete: true,
+            ..Default::default()
+        };
+        let fused = disk
+            .try_fused_recoverable_delete_version("recovery", "object", &v2, &opts, transaction, None)
+            .await
+            .expect("fused versioned delete preparation must succeed");
+        assert_eq!(fused, Some(()));
+        assert!(
+            root.join(v2_dir.to_string())
+                .join(format!("{RECOVERY_DELETE_DATA_DIR_MARKER_PREFIX}{transaction}"))
+                .is_file(),
+            "deleted version data_dir must have recovery marker"
+        );
+
+        let (journal, sole_transaction, target_satisfied) = disk
+            .try_publish_prepared_delete_local("recovery", "object", transaction, None)
+            .await
+            .expect("local publish must succeed")
+            .expect("versioned target must satisfy local publish");
+        assert!(sole_transaction, "surviving v1 data_dir must not clear sole_transaction");
+        assert!(target_satisfied);
+        assert!(journal.target_digest.is_some());
+
+        // Simulate interruption during finish-before-finalize right after TARGET is removed
+        // (after finished receipt + FINISHED marker were persisted, before receipt_path is removed).
+        let rollback = root.join(transaction.to_string());
+        let finished_receipt = CommitReceipt {
+            journal: journal.clone(),
+            finished: true,
+        };
+        fs::write(
+            rollback.join(DELETE_QUORUM_COMMITTED_FILE),
+            serde_json::to_vec(&finished_receipt).expect("serialize finished receipt"),
+        )
+        .await
+        .expect("write finished receipt");
+        fs::write(rollback.join(FINISHED), []).await.expect("write FINISHED");
+        fs::remove_file(rollback.join(STORAGE_FORMAT_FILE_BACKUP))
+            .await
+            .expect("remove backup");
+        fs::remove_file(rollback.join(TARGET)).await.expect("remove TARGET");
+
+        let inspected = disk
+            .inspect_delete_transaction("recovery", "object", transaction)
+            .await
+            .expect("interrupted post-TARGET state must remain inspectable without FileCorrupt");
+        assert_eq!(inspected.state, DeleteRecoveryState::CommitFinished);
+
+        // Complete cleanup via try_finish_and_finalize_prepared_delete_local.
+        let finalized = disk
+            .try_finish_and_finalize_prepared_delete_local("recovery", "object", transaction, &journal, None)
+            .await
+            .expect("local finalize must succeed after interruption");
+        assert!(finalized);
+        assert!(!rollback.exists(), "rollback directory must be pruned");
+        assert!(!root.join(v2_dir.to_string()).exists(), "deleted v2 data_dir must be removed");
+        assert!(
+            root.join(v1_dir.to_string()).join("part.1").is_file(),
+            "surviving v1 data_dir must remain intact"
+        );
+    }
+
+    #[tokio::test]
+    async fn delete_recovery_fused_unversioned_physical_delete_retains_receipt_until_finalize() {
+        let (_temp, disk, mut info, _) = fixture().await;
+        let root = disk.io_get_object_path("recovery", "object").expect("object path");
+
+        info.data = None;
+        rustfs_utils::http::metadata_compat::remove_str(
+            &mut info.metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_INLINE_DATA,
+        );
+        let data_dir = Uuid::new_v4();
+        info.data_dir = Some(data_dir);
+        fs::create_dir_all(root.join(data_dir.to_string()))
+            .await
+            .expect("create data_dir");
+        fs::write(root.join(data_dir.to_string()).join("part.1"), b"physical-payload")
+            .await
+            .expect("write part.1");
+        let mut meta = FileMeta::new();
+        meta.add_version(info.clone()).expect("add version");
+        fs::write(root.join(STORAGE_FORMAT_FILE), meta.marshal_msg().expect("marshal meta"))
+            .await
+            .expect("write xl.meta");
+
+        let transaction = Uuid::new_v4();
+        let opts = DeleteOptions {
+            recoverable_delete: true,
+            ..Default::default()
+        };
+        let fused = disk
+            .try_fused_recoverable_delete_version("recovery", "object", &info, &opts, transaction, None)
+            .await
+            .expect("fused unversioned physical delete must succeed");
+        assert_eq!(fused, Some(()));
+
+        let (journal, sole_transaction, target_satisfied) = disk
+            .try_publish_prepared_delete_local("recovery", "object", transaction, None)
+            .await
+            .expect("local publish must succeed")
+            .expect("absent target must satisfy local publish");
+        assert!(sole_transaction);
+        assert!(target_satisfied);
+        assert!(journal.target_digest.is_none());
+
+        let rollback = root.join(transaction.to_string());
+        assert!(rollback.join(DELETE_QUORUM_COMMITTED_FILE).is_file());
+
+        let finalized = disk
+            .try_finish_and_finalize_prepared_delete_local("recovery", "object", transaction, &journal, None)
+            .await
+            .expect("local finalize must succeed");
+        assert!(finalized);
+        assert!(!rollback.exists(), "rollback directory must be pruned");
+        assert!(!root.join(data_dir.to_string()).exists(), "physical data_dir must be removed");
+        assert!(!root.exists(), "empty object directory must be pruned");
+    }
 }

 #[cfg(test)]
@@ -421,6 +421,7 @@ impl LocalDisk {
                    return Err(DiskError::ConditionalDeleteConflict);
                }
                let permit = AbortDeletePermit::new(&store, &expected, certificate.clone())?;
+                super::multipart_source_state::invalidate_active_multipart_source_head(&store.source);
                let destination = self.io_get_object_path(RUSTFS_META_TMP_DELETED_BUCKET, &Uuid::new_v4().to_string())?;
                let source = store.source.clone();
                let base = store.base.clone();
@@ -618,6 +619,9 @@ pub(super) struct OrdinaryMultipartDeleteCheck {
    root: os::PublicationRoot,
 }
 impl OrdinaryMultipartDeleteCheck {
+    pub(super) fn source(&self) -> &Path {
+        &self.source
+    }
    pub(super) fn validate(&self) -> Result<()> {
        let pinned = match os::pin_multipart_source_directory(&self.source, &self.base, &self.root) {
            Ok(pinned) => pinned,
@@ -238,9 +238,192 @@ pub(super) fn read_optional_in_pinned(
        Err(error) => Err(error.into()),
    }
 }
+#[derive(Clone, Copy, PartialEq, Eq)]
+struct FileStamp {
+    len: u64,
+    modified: Option<std::time::SystemTime>,
+    #[cfg(unix)]
+    ino: u64,
+}
+
+impl FileStamp {
+    fn from_path(path: &Path) -> Option<Self> {
+        let meta = std::fs::symlink_metadata(path).ok()?;
+        Some(Self {
+            len: meta.len(),
+            modified: meta.modified().ok(),
+            #[cfg(unix)]
+            ino: {
+                use std::os::unix::fs::MetadataExt;
+                meta.ino()
+            },
+        })
+    }
+}
+
+#[derive(Clone)]
+struct ActiveMultipartSourceEntry {
+    raw: Bytes,
+    fi: FileInfo,
+    head_stamp: FileStamp,
+    unsettled_parts: std::collections::HashSet<usize>,
+    parts: std::collections::HashMap<usize, (FileStamp, rustfs_filemeta::ObjectPartInfo)>,
+}
+
+static ACTIVE_MULTIPART_SOURCE_HEADS: std::sync::LazyLock<
+    parking_lot::Mutex<std::collections::HashMap<PathBuf, ActiveMultipartSourceEntry>>,
+> = std::sync::LazyLock::new(|| parking_lot::Mutex::new(std::collections::HashMap::new()));
+
+pub(super) fn put_active_multipart_source_head(source: PathBuf, raw: Bytes, fi: FileInfo) {
+    let Some(head_stamp) = FileStamp::from_path(&source.join(STORAGE_FORMAT_FILE)) else {
+        return;
+    };
+    let mut map = ACTIVE_MULTIPART_SOURCE_HEADS.lock();
+    if map.len() >= 4096
+        && let Some(evict_key) = map.keys().next().cloned()
+    {
+        map.remove(&evict_key);
+    }
+    map.insert(
+        source,
+        ActiveMultipartSourceEntry {
+            raw,
+            fi,
+            head_stamp,
+            unsettled_parts: std::collections::HashSet::new(),
+            parts: std::collections::HashMap::new(),
+        },
+    );
+}
+
+fn stamp_matches_active_disk(source: &Path, expected_head_stamp: FileStamp) -> bool {
+    FileStamp::from_path(&source.join(STORAGE_FORMAT_FILE)) == Some(expected_head_stamp)
+        && !source.join(STATE).exists()
+        && !source.join(crate::disk::MULTIPART_RECOVERY_DIR).exists()
+        && !source.join(".mpu-completions").exists()
+}
+
+pub(super) fn get_active_multipart_source_head(source: &Path) -> Option<(Bytes, FileInfo)> {
+    let (raw, fi, head_stamp) = {
+        let map = ACTIVE_MULTIPART_SOURCE_HEADS.lock();
+        let entry = map.get(source)?;
+        (entry.raw.clone(), entry.fi.clone(), entry.head_stamp)
+    };
+    if !stamp_matches_active_disk(source, head_stamp) {
+        ACTIVE_MULTIPART_SOURCE_HEADS.lock().remove(source);
+        return None;
+    }
+    Some((raw, fi))
+}
+
+pub(super) fn has_active_multipart_source_head(source: &Path) -> bool {
+    let Some(head_stamp) = ACTIVE_MULTIPART_SOURCE_HEADS.lock().get(source).map(|entry| entry.head_stamp) else {
+        return false;
+    };
+    if !stamp_matches_active_disk(source, head_stamp) {
+        ACTIVE_MULTIPART_SOURCE_HEADS.lock().remove(source);
+        return false;
+    }
+    true
+}
+
+pub(super) fn has_clean_active_multipart_source(source: &Path) -> bool {
+    let Some((head_stamp, clean)) = ACTIVE_MULTIPART_SOURCE_HEADS
+        .lock()
+        .get(source)
+        .map(|entry| (entry.head_stamp, entry.unsettled_parts.is_empty()))
+    else {
+        return false;
+    };
+    if !stamp_matches_active_disk(source, head_stamp) {
+        ACTIVE_MULTIPART_SOURCE_HEADS.lock().remove(source);
+        return false;
+    }
+    clean
+}
+
+pub(super) fn begin_active_multipart_part_transaction(source: &Path, part_number: usize) {
+    if let Some(entry) = ACTIVE_MULTIPART_SOURCE_HEADS.lock().get_mut(source) {
+        entry.unsettled_parts.insert(part_number);
+        entry.parts.remove(&part_number);
+    }
+}
+
+pub(super) fn record_active_multipart_part_committed(source: &Path, part: rustfs_filemeta::ObjectPartInfo) {
+    let Some(data_dir) = ACTIVE_MULTIPART_SOURCE_HEADS
+        .lock()
+        .get(source)
+        .and_then(|entry| entry.fi.data_dir)
+        .filter(|dir| !dir.is_nil())
+    else {
+        return;
+    };
+    let part_meta_path = source.join(data_dir.to_string()).join(format!("part.{}.meta", part.number));
+    let Some(part_stamp) = FileStamp::from_path(&part_meta_path) else {
+        return;
+    };
+    if let Some(entry) = ACTIVE_MULTIPART_SOURCE_HEADS.lock().get_mut(source) {
+        entry.parts.insert(part.number, (part_stamp, part));
+    }
+}
+
+pub(super) fn finish_active_multipart_part_transaction(
+    source: &Path,
+    part_number: Option<usize>,
+    committed: bool,
+    retired_existing_txn: bool,
+) {
+    let mut map = ACTIVE_MULTIPART_SOURCE_HEADS.lock();
+    let Some(part_number) = part_number.filter(|_| committed) else {
+        map.remove(source);
+        return;
+    };
+    if !retired_existing_txn {
+        return;
+    }
+    if let Some(entry) = map.get_mut(source)
+        && !entry.unsettled_parts.remove(&part_number)
+    {
+        map.remove(source);
+    }
+}
+
+pub(super) fn get_active_multipart_source_part(source: &Path, part_number: usize) -> Option<rustfs_filemeta::ObjectPartInfo> {
+    let (head_stamp, data_dir, expected_stamp, part) = {
+        let map = ACTIVE_MULTIPART_SOURCE_HEADS.lock();
+        let entry = map.get(source)?;
+        if !entry.unsettled_parts.is_empty() {
+            return None;
+        }
+        let data_dir = entry.fi.data_dir.filter(|dir| !dir.is_nil())?;
+        let (expected_stamp, part) = entry.parts.get(&part_number)?;
+        (entry.head_stamp, data_dir, *expected_stamp, part.clone())
+    };
+    if !stamp_matches_active_disk(source, head_stamp) {
+        ACTIVE_MULTIPART_SOURCE_HEADS.lock().remove(source);
+        return None;
+    }
+    let part_meta_path = source.join(data_dir.to_string()).join(format!("part.{part_number}.meta"));
+    if FileStamp::from_path(&part_meta_path) != Some(expected_stamp) {
+        return None;
+    }
+    Some(part)
+}
+
+pub(super) fn invalidate_active_multipart_source_head(source: &Path) {
+    ACTIVE_MULTIPART_SOURCE_HEADS.lock().remove(source);
+}
+
+pub(super) fn invalidate_active_multipart_sources_under(prefix: &Path) {
+    ACTIVE_MULTIPART_SOURCE_HEADS.lock().retain(|key, _| !key.starts_with(prefix));
+}
+
 /// Canonical source mutations are admitted under SourceRW at the syscall,
 /// not on the earlier async caller that may have timed out.
 pub(super) fn validate_source_mutation(source: &Path, base: &Path, root: &os::PublicationRoot, initial: bool) -> Result<()> {
+    if !initial && has_active_multipart_source_head(source) {
+        return Ok(());
+    }
    let pinned = match os::pin_multipart_source_directory(source, base, root) {
        Ok(pinned) => pinned,
        Err(error) if error.kind() == std::io::ErrorKind::NotFound => {
@@ -440,6 +623,7 @@ impl SourceStore {
        terminal: Option<&Path>,
        sync_source_before: bool,
    ) -> std::io::Result<()> {
+        invalidate_active_multipart_source_head(source);
        #[cfg(unix)]
        let mut pinned = os::pin_multipart_source_directory(source, base, root)?;
        #[cfg(not(unix))]
@@ -470,7 +654,7 @@ impl SourceStore {
        let terminal = terminal_record_path(&source, state);
        let bytes = bytes.to_vec();
        os::run_blocking_namespace_operation_with_sync(self.lease.clone(), true, move || {
-            Self::write_owned_backup_std(&source, &base, &root, owner, target, &bytes)?;
+            Self::write_owned_backup_std(&source, &base, &root, owner, target, &bytes, false)?;
            Self::persist_state_std(&source, &base, &root, &raw, terminal.as_deref(), false)
        })
        .await
@@ -539,7 +723,7 @@ impl SourceStore {
        let root = self.publication_root.clone();
        let bytes = bytes.to_vec();
        os::run_blocking_namespace_operation(self.lease.clone(), move || {
-            Self::write_owned_backup_std(&source, &base, &root, owner, target, &bytes)
+            Self::write_owned_backup_std(&source, &base, &root, owner, target, &bytes, true)
        })
        .await
        .map_err(DiskError::from)
@@ -551,6 +735,7 @@ impl SourceStore {
        owner: Uuid,
        target: bool,
        bytes: &[u8],
+        sync_source_root: bool,
    ) -> std::io::Result<()> {
        let path = if target {
            target_backup_path(source, owner)
@@ -582,9 +767,10 @@ impl SourceStore {
        if !existed {
            os::replace_multipart_source_metadata(&path, bytes, base, root)?;
        }
-        os::sync_multipart_owned_record_parents(source, parent, existed, base, root)
+        os::sync_multipart_owned_record_parents_options(source, parent, existed, sync_source_root, base, root)
    }
    pub async fn replace_head(&self, bytes: &[u8]) -> Result<()> {
+        invalidate_active_multipart_source_head(&self.source);
        let path = self.source.join(STORAGE_FORMAT_FILE);
        let base = self.base.clone();
        let root = self.publication_root.clone();
@@ -1097,6 +1283,82 @@ pub(super) fn load_source(
    })
 }

+fn load_uninitialized_source(
+    raw: Bytes,
+    path: &str,
+    disk_id: Uuid,
+    physical_slot: usize,
+    disk_ids: &[Uuid],
+) -> Result<LoadedSource> {
+    let sha = raw_sha(&raw);
+    let length = raw.len() as u64;
+    let current = get_file_info(
+        &raw,
+        RUSTFS_META_MULTIPART_BUCKET,
+        path,
+        "",
+        FileInfoOpts {
+            data: false,
+            include_free_versions: false,
+            include_part_checksums: true,
+        },
+    )
+    .map_err(Error::other)?;
+    native_source_protocol(&current)?;
+    if source_status(&current)?.is_some()
+        || publication_identity(&current)?.is_some()
+        || SourceConditionV1::take(&mut current.clone())?.is_some()
+    {
+        return Err(DiskError::FileCorrupt);
+    }
+    let bucket = rustfs_utils::http::metadata_compat::get_consistent_str(&current.metadata, "multipart-bucket")
+        .ok_or(DiskError::FileCorrupt)?;
+    let object = rustfs_utils::http::metadata_compat::get_consistent_str(&current.metadata, "multipart-object")
+        .ok_or(DiskError::FileCorrupt)?;
+    let condition = capture_source_condition(bucket, object, path, &current, current.data_dir.ok_or(DiskError::FileCorrupt)?)?;
+    if condition.generation.epoch != 0 {
+        return Err(DiskError::FileCorrupt);
+    }
+    let slot = condition
+        .lineage
+        .geometry
+        .distribution
+        .iter()
+        .position(|index| *index == current.erasure.index)
+        .ok_or(DiskError::FileCorrupt)?;
+    if slot != physical_slot || condition.lineage.geometry.disks() != disk_ids.len() {
+        return Err(DiskError::ConditionalDeleteConflict);
+    }
+    let inventory = raw_sha(&encode(&(None::<[u8; 32]>, Some(sha), sha, None::<[u8; 32]>, None::<[u8; 32]>))?);
+    let snapshot = Some(MultipartSourceSnapshot {
+        disk_id,
+        physical_slot: slot,
+        lineage: condition.lineage,
+        generation: condition.generation,
+        revision: 0,
+        phase: MultipartSourcePhase::Active,
+        state_sha256: [0; 32],
+        head_sha256: Some(sha),
+        inventory_sha256: inventory,
+        capability_sha256: None,
+        publication: None,
+    });
+    Ok(LoadedSource {
+        root_missing: false,
+        state: None,
+        head: Some(SourceHead {
+            sha256: sha,
+            length,
+            raw: Some(raw.clone()),
+        }),
+        original: Some(raw),
+        info: Some(current),
+        snapshot,
+        head_matches_state: true,
+        head_matches_activation: false,
+    })
+}
+
 pub(super) const RETIREMENT_ANCHOR: &str = ".mpu-source-retirement";

 #[derive(Serialize, Deserialize)]
@@ -1323,41 +1585,6 @@ impl LocalDisk {
 }

 impl SourceStore {
-    /// Only native source validation may create provenance. Ordinary TMP/heal
-    /// publication can retain a valid old local anchor, never issue/rebind one.
-    pub async fn write_retirement_anchor(
-        &self,
-        state: &SourceState,
-        publication: &PublicationIdentity,
-        directory: &Path,
-        lease: Arc<os::NamespaceMutationLease>,
-    ) -> Result<()> {
-        if state.lineage != publication.lineage
-            || state.generation != publication.source_generation
-            || state.disk_id != self.disk_id
-            || state.physical_slot != self.physical_slot
-        {
-            return Err(DiskError::ConditionalDeleteConflict);
-        }
-        let record = NativeRetirementAnchor {
-            protocol_version: MULTIPART_SOURCE_PROTOCOL_VERSION,
-            disk_id: self.disk_id,
-            physical_slot: self.physical_slot,
-            publication: publication.clone(),
-            source_original_sha256: state.original_sha256,
-            source_validation_receipt_id: state.backup_owner,
-        };
-        let bytes = encode(&record)?;
-        let directory = directory.to_path_buf();
-        let base = self.base.clone();
-        let root = self.publication_root.clone();
-        os::run_blocking_namespace_operation(lease, move || {
-            let _parent = os::mkdir_all_below_existing_base_std(&directory, &base, &root)?;
-            os::replace_multipart_source_metadata(&directory.join(RETIREMENT_ANCHOR), &bytes, &base, &root)
-        })
-        .await
-        .map_err(DiskError::from)
-    }
    /// This caller already owns SourceRW and destination metadata ownership.
    /// An Aborting survivor can confirm a historical target only with its
    /// independent, same-disk publication anchor. Closure never creates one.
@@ -1456,6 +1683,8 @@ struct ValidatedPreparedPublication {
    length: u64,
 }

+type DeferredInitialPublication = (Vec<u8>, PathBuf, Vec<u8>);
+
 pub(super) struct NativeMultipartPublication {
    pub store: Arc<SourceStore>,
    pub publication: PublicationIdentity,
@@ -1463,22 +1692,49 @@ pub(super) struct NativeMultipartPublication {
    state: parking_lot::Mutex<SourceState>,
    capability: parking_lot::Mutex<Option<MultipartUndoCapability>>,
    validated_prepared: parking_lot::Mutex<Option<ValidatedPreparedPublication>>,
+    deferred_initial: parking_lot::Mutex<Option<DeferredInitialPublication>>,
    pub legacy: bool,
 }

 impl LocalDisk {
+    #[cfg(test)]
    pub(super) async fn prepare_native_multipart_publication(
        &self,
        path: &str,
        incoming: &mut FileInfo,
        staging: &MultipartCommitStaging,
+    ) -> Result<NativeMultipartPublication> {
+        self.prepare_native_multipart_publication_inner(path, incoming, staging, false)
+            .await
+    }
+
+    pub(super) async fn prepare_native_multipart_publication_deferred(
+        &self,
+        path: &str,
+        incoming: &mut FileInfo,
+        staging: &MultipartCommitStaging,
+    ) -> Result<NativeMultipartPublication> {
+        self.prepare_native_multipart_publication_inner(path, incoming, staging, true)
+            .await
+    }
+
+    async fn prepare_native_multipart_publication_inner(
+        &self,
+        path: &str,
+        incoming: &mut FileInfo,
+        staging: &MultipartCommitStaging,
+        defer_persist: bool,
    ) -> Result<NativeMultipartPublication> {
        // Real Complete is parts-backed, including a zero-byte last part.
        if incoming.data.is_some() || incoming.inline_data() {
            return Err(DiskError::UnsupportedDisk);
        }
        let store = Arc::new(self.multipart_store_from_staging(path, staging).await?);
-        let loaded = store.load().await?;
+        let loaded = if defer_persist && let Some(raw) = staging.fresh_source_head() {
+            load_uninitialized_source(raw, &store.path, store.disk_id, store.physical_slot, &store.disk_ids)?
+        } else {
+            store.load().await?
+        };
        let snapshot = loaded.snapshot.as_ref().ok_or(DiskError::ConditionalDeleteConflict)?;
        let original = loaded
            .head
@@ -1569,7 +1825,44 @@ impl LocalDisk {
        };
        // Copy the small source FI to an independent immutable inode. Its
        // canonical inode is never truncated to build the destination proxy.
-        store.write_owned_backup_and_persist(owner, false, original, &state).await?;
+        let directory = staging.private_directory(publication.final_data_dir);
+        let anchor_record = NativeRetirementAnchor {
+            protocol_version: MULTIPART_SOURCE_PROTOCOL_VERSION,
+            disk_id: store.disk_id,
+            physical_slot: store.physical_slot,
+            publication: publication.clone(),
+            source_original_sha256: state.original_sha256,
+            source_validation_receipt_id: state.backup_owner,
+        };
+        let anchor_bytes = encode(&anchor_record)?;
+        state.validate(store.disk_id, &store.path)?;
+        let original_bytes = original.to_vec();
+        let deferred_initial = if defer_persist {
+            Some((original_bytes, directory, anchor_bytes))
+        } else {
+            let state_raw = encode(&state)?;
+            let source_dir = store.source.clone();
+            let base_dir = store.base.clone();
+            let pub_root = store.publication_root.clone();
+            let terminal = terminal_record_path(&source_dir, &state);
+            #[cfg(target_os = "linux")]
+            let can_mark_presynced = has_active_multipart_source_head(&source_dir)
+                && super::effective_durability(RUSTFS_META_MULTIPART_BUCKET).syncs_data_shards();
+            os::run_blocking_namespace_operation_with_sync(staging.lease.clone(), true, move || {
+                SourceStore::write_owned_backup_std(&source_dir, &base_dir, &pub_root, owner, false, &original_bytes, false)?;
+                SourceStore::persist_state_std(&source_dir, &base_dir, &pub_root, &state_raw, terminal.as_deref(), false)?;
+                let _parent = os::mkdir_all_below_existing_base_std(&directory, &base_dir, &pub_root)?;
+                os::replace_multipart_source_metadata(&directory.join(RETIREMENT_ANCHOR), &anchor_bytes, &base_dir, &pub_root)?;
+                #[cfg(target_os = "linux")]
+                if can_mark_presynced {
+                    super::mark_presynced_tmp_data_dir(directory);
+                }
+                Ok(())
+            })
+            .await
+            .map_err(DiskError::from)?;
+            None
+        };
        let prepared_path = store.source.join(format!(".mpu-prepared-{owner}"));
        let context = NativeMultipartPublication {
            store,
@@ -1578,14 +1871,9 @@ impl LocalDisk {
            state: parking_lot::Mutex::new(state),
            capability: parking_lot::Mutex::new(Some(capability)),
            validated_prepared: parking_lot::Mutex::new(None),
+            deferred_initial: parking_lot::Mutex::new(deferred_initial),
            legacy,
        };
-        let directory = staging.private_directory(context.publication.final_data_dir);
-        let reserved = context.state.lock().clone();
-        context
-            .store
-            .write_retirement_anchor(&reserved, &context.publication, &directory, staging.lease.clone())
-            .await?;
        Ok(context)
    }
 }
@@ -1594,18 +1882,15 @@ impl NativeMultipartPublication {
    pub fn canonical_head(&self) -> PathBuf {
        self.store.source.join(STORAGE_FORMAT_FILE)
    }
+    #[cfg(test)]
    pub async fn prepare_metadata(&self, new_metadata: &[u8], old_metadata: Option<&[u8]>) -> Result<()> {
        let mut state = self.state.lock().clone();
-        let current = self.store.load().await?;
-        let current_state = current.state.as_ref().ok_or(DiskError::ConditionalDeleteConflict)?;
-        if encode(current_state)? != encode(&state)? || !current.head_matches_state {
-            return Err(DiskError::ConditionalDeleteConflict);
-        }
        // Validate the actual persisted representation once, after FileMeta's
        // canonicalization. No unvalidated input can install a source proxy.
        if !self.store.target_matches(&self.publication, new_metadata)? {
            return Err(DiskError::ConditionalDeleteConflict);
        }
+        let expected_state_raw = encode(&state)?;
        let validated = ValidatedPreparedPublication {
            publication: self.publication.clone(),
            sha256: raw_sha(new_metadata),
@@ -1618,17 +1903,297 @@ impl NativeMultipartPublication {
        state.prepared_sha256 = Some(validated.sha256);
        state.prepared_length = Some(validated.length);
        state.advance(MultipartSourcePhase::Prepared)?;
-        if let Some(previous) = old_metadata {
-            self.store
-                .write_owned_backup_and_persist(state.backup_owner, true, previous, &state)
-                .await?;
-        } else {
-            self.store.persist(&state).await?;
-        }
+        state.validate(self.store.disk_id, &self.store.path)?;
+        let next_state_raw = encode(&state)?;
+        let source = self.store.source.clone();
+        let base = self.store.base.clone();
+        let root = self.store.publication_root.clone();
+        let path = self.store.path.clone();
+        let disk_id = self.store.disk_id;
+        let physical_slot = self.store.physical_slot;
+        let disk_ids = self.store.disk_ids.clone();
+        let backup_owner = state.backup_owner;
+        let old_bytes = old_metadata.map(|b| b.to_vec());
+        let terminal = terminal_record_path(&source, &state);
+        os::run_blocking_namespace_operation_with_sync(self.store.lease.clone(), true, move || {
+            let current = load_source(&source, &base, &root, &path, disk_id, physical_slot, &disk_ids).map_err(IoError::other)?;
+            let current_state = current
+                .state
+                .as_ref()
+                .ok_or_else(|| IoError::other(DiskError::ConditionalDeleteConflict))?;
+            let current_raw = encode(current_state).map_err(IoError::other)?;
+            if current_raw != expected_state_raw || !current.head_matches_state {
+                return Err(IoError::other(DiskError::ConditionalDeleteConflict));
+            }
+            if let Some(previous) = old_bytes.as_deref() {
+                SourceStore::write_owned_backup_std(&source, &base, &root, backup_owner, true, previous, false)?;
+            }
+            SourceStore::persist_state_std(&source, &base, &root, &next_state_raw, terminal.as_deref(), false)
+        })
+        .await
+        .map_err(DiskError::from)?;
        *self.state.lock() = state;
        *self.validated_prepared.lock() = Some(validated);
        Ok(())
    }
+    pub(super) fn prepare_and_install_head_on_thread(
+        &self,
+        new_metadata: &[u8],
+        old_metadata: Option<&[u8]>,
+        dst_file_path: &Path,
+        directory_guard: &os::RenameCommitGuard,
+    ) -> Result<os::PreparedRenameSource> {
+        let mut state = self.state.lock().clone();
+        if !self.store.target_matches(&self.publication, new_metadata)? {
+            return Err(DiskError::ConditionalDeleteConflict);
+        }
+        let deferred_initial = self.deferred_initial.lock().take();
+        let expected_state_raw = encode(&state)?;
+        let expected_lineage = state.lineage.clone();
+        let expected_generation = state.generation;
+        let expected_revision = state.revision;
+        let validated = ValidatedPreparedPublication {
+            publication: self.publication.clone(),
+            sha256: raw_sha(new_metadata),
+            length: new_metadata.len() as u64,
+        };
+        if let Some(previous) = old_metadata {
+            state.previous_target_sha256 = Some(raw_sha(previous));
+            state.previous_target_length = Some(previous.len() as u64);
+        }
+        state.prepared_sha256 = Some(validated.sha256);
+        state.prepared_length = Some(validated.length);
+        state.advance(MultipartSourcePhase::Prepared)?;
+        state.validate(self.store.disk_id, &self.store.path)?;
+        let next_state_raw = encode(&state)?;
+        let source = &self.store.source;
+        let base = &self.store.base;
+        let root = &self.store.publication_root;
+        let path = &self.store.path;
+        let disk_id = self.store.disk_id;
+        let physical_slot = self.store.physical_slot;
+        let disk_ids = &self.store.disk_ids;
+        let backup_owner = state.backup_owner;
+        #[cfg(target_os = "linux")]
+        let can_mark_presynced = has_active_multipart_source_head(source)
+            && super::effective_durability(RUSTFS_META_MULTIPART_BUCKET).syncs_data_shards();
+        let terminal = terminal_record_path(source, &state);
+        let prepared_path = &self.prepared_path;
+        let canonical = self.canonical_head();
+        let prepared = (|| -> std::io::Result<os::PreparedRenameSource> {
+            if let Some((ref original_bytes, _, _)) = deferred_initial {
+                let current = load_source(source, base, root, path, disk_id, physical_slot, disk_ids).map_err(IoError::other)?;
+                let snapshot = current
+                    .snapshot
+                    .as_ref()
+                    .ok_or_else(|| IoError::other(DiskError::ConditionalDeleteConflict))?;
+                let current_head = current
+                    .head
+                    .as_ref()
+                    .and_then(SourceHead::raw)
+                    .ok_or_else(|| IoError::other(DiskError::ConditionalDeleteConflict))?;
+                if !current.head_matches_state
+                    || snapshot.phase != MultipartSourcePhase::Active
+                    || snapshot.capability_sha256.is_some()
+                    || snapshot.lineage != expected_lineage
+                    || snapshot.generation != expected_generation
+                    || snapshot.revision.checked_add(1) != Some(expected_revision)
+                    || current_head != original_bytes.as_slice()
+                {
+                    return Err(IoError::other(DiskError::ConditionalDeleteConflict));
+                }
+            } else {
+                let current = load_source(source, base, root, path, disk_id, physical_slot, disk_ids).map_err(IoError::other)?;
+                let current_state = current
+                    .state
+                    .as_ref()
+                    .ok_or_else(|| IoError::other(DiskError::ConditionalDeleteConflict))?;
+                let current_raw = encode(current_state).map_err(IoError::other)?;
+                if current_raw != expected_state_raw || !current.head_matches_state {
+                    return Err(IoError::other(DiskError::ConditionalDeleteConflict));
+                }
+            }
+            #[cfg(test)]
+            super::run_owned_file_before_open(prepared_path);
+            let mut prepared_metadata_source =
+                os::create_prepared_rename_source_with_commit_guard(prepared_path, dst_file_path, directory_guard)?;
+            if let Err(err) = prepared_metadata_source.write_all(new_metadata, false) {
+                drop(prepared_metadata_source);
+                let _ = std::fs::remove_file(prepared_path);
+                return Err(err);
+            }
+            let setup_res: std::io::Result<()> = (|| {
+                if let Some((original_bytes, directory, anchor_bytes)) = deferred_initial {
+                    SourceStore::write_owned_backup_std(source, base, root, backup_owner, false, &original_bytes, false)?;
+                    if let Some(previous) = old_metadata {
+                        SourceStore::write_owned_backup_std(source, base, root, backup_owner, true, previous, false)?;
+                    }
+                    SourceStore::persist_state_std(source, base, root, &next_state_raw, terminal.as_deref(), false)?;
+                    let _parent = os::mkdir_all_below_existing_base_std(&directory, base, root)?;
+                    os::replace_multipart_source_metadata(&directory.join(RETIREMENT_ANCHOR), &anchor_bytes, base, root)?;
+                    #[cfg(target_os = "linux")]
+                    if can_mark_presynced {
+                        super::mark_presynced_tmp_data_dir(directory);
+                    }
+                } else {
+                    if let Some(previous) = old_metadata {
+                        SourceStore::write_owned_backup_std(source, base, root, backup_owner, true, previous, false)?;
+                    }
+                    SourceStore::persist_state_std(source, base, root, &next_state_raw, terminal.as_deref(), false)?;
+                }
+                Ok(())
+            })();
+            if let Err(err) = setup_res {
+                drop(prepared_metadata_source);
+                let _ = std::fs::remove_file(prepared_path);
+                return Err(err);
+            }
+            if let Err(err) = prepared_metadata_source.sync_data() {
+                drop(prepared_metadata_source);
+                let _ = std::fs::remove_file(prepared_path);
+                return Err(err);
+            }
+            os::install_multipart_prepared_head(prepared_metadata_source, &canonical, base, root)
+        })()
+        .map_err(DiskError::from)?;
+        *self.state.lock() = state;
+        *self.validated_prepared.lock() = Some(validated);
+        Ok(prepared)
+    }
+
+    #[cfg(not(target_os = "linux"))]
+    pub async fn prepare_and_install_head(
+        &self,
+        new_metadata: &[u8],
+        old_metadata: Option<&[u8]>,
+        dst_file_path: &Path,
+        directory_guard: &os::RenameCommitGuard,
+        mutation_lease: Arc<os::NamespaceMutationLease>,
+    ) -> Result<os::PreparedRenameSource> {
+        let mut state = self.state.lock().clone();
+        if !self.store.target_matches(&self.publication, new_metadata)? {
+            return Err(DiskError::ConditionalDeleteConflict);
+        }
+        let deferred_initial = self.deferred_initial.lock().take();
+        let expected_state_raw = encode(&state)?;
+        let expected_lineage = state.lineage.clone();
+        let expected_generation = state.generation;
+        let expected_revision = state.revision;
+        let validated = ValidatedPreparedPublication {
+            publication: self.publication.clone(),
+            sha256: raw_sha(new_metadata),
+            length: new_metadata.len() as u64,
+        };
+        if let Some(previous) = old_metadata {
+            state.previous_target_sha256 = Some(raw_sha(previous));
+            state.previous_target_length = Some(previous.len() as u64);
+        }
+        state.prepared_sha256 = Some(validated.sha256);
+        state.prepared_length = Some(validated.length);
+        state.advance(MultipartSourcePhase::Prepared)?;
+        state.validate(self.store.disk_id, &self.store.path)?;
+        let next_state_raw = encode(&state)?;
+        let source = self.store.source.clone();
+        let base = self.store.base.clone();
+        let root = self.store.publication_root.clone();
+        let path = self.store.path.clone();
+        let disk_id = self.store.disk_id;
+        let physical_slot = self.store.physical_slot;
+        let disk_ids = self.store.disk_ids.clone();
+        let backup_owner = state.backup_owner;
+        #[cfg(target_os = "linux")]
+        let can_mark_presynced = has_active_multipart_source_head(&source)
+            && super::effective_durability(RUSTFS_META_MULTIPART_BUCKET).syncs_data_shards();
+        let old_bytes = old_metadata.map(|b| b.to_vec());
+        let new_bytes = new_metadata.to_vec();
+        let terminal = terminal_record_path(&source, &state);
+        let prepared_path = self.prepared_path.clone();
+        let canonical = self.canonical_head();
+        let dst_file_path = dst_file_path.to_path_buf();
+        let owned_directory_guard = directory_guard.clone();
+        let prepared = os::run_blocking_namespace_operation_with_sync(mutation_lease, true, move || {
+            let current = load_source(&source, &base, &root, &path, disk_id, physical_slot, &disk_ids).map_err(IoError::other)?;
+            if let Some((original_bytes, directory, anchor_bytes)) = deferred_initial {
+                let snapshot = current
+                    .snapshot
+                    .as_ref()
+                    .ok_or_else(|| IoError::other(DiskError::ConditionalDeleteConflict))?;
+                let current_head = current
+                    .head
+                    .as_ref()
+                    .and_then(SourceHead::raw)
+                    .ok_or_else(|| IoError::other(DiskError::ConditionalDeleteConflict))?;
+                if !current.head_matches_state
+                    || snapshot.phase != MultipartSourcePhase::Active
+                    || snapshot.capability_sha256.is_some()
+                    || snapshot.lineage != expected_lineage
+                    || snapshot.generation != expected_generation
+                    || snapshot.revision.checked_add(1) != Some(expected_revision)
+                    || current_head != original_bytes.as_slice()
+                {
+                    return Err(IoError::other(DiskError::ConditionalDeleteConflict));
+                }
+                SourceStore::write_owned_backup_std(&source, &base, &root, backup_owner, false, &original_bytes, false)?;
+                if let Some(previous) = old_bytes.as_deref() {
+                    SourceStore::write_owned_backup_std(&source, &base, &root, backup_owner, true, previous, false)?;
+                }
+                SourceStore::persist_state_std(&source, &base, &root, &next_state_raw, terminal.as_deref(), false)?;
+                let _parent = os::mkdir_all_below_existing_base_std(&directory, &base, &root)?;
+                os::replace_multipart_source_metadata(&directory.join(RETIREMENT_ANCHOR), &anchor_bytes, &base, &root)?;
+                #[cfg(target_os = "linux")]
+                if can_mark_presynced {
+                    super::mark_presynced_tmp_data_dir(directory);
+                }
+            } else {
+                let current_state = current
+                    .state
+                    .as_ref()
+                    .ok_or_else(|| IoError::other(DiskError::ConditionalDeleteConflict))?;
+                let current_raw = encode(current_state).map_err(IoError::other)?;
+                if current_raw != expected_state_raw || !current.head_matches_state {
+                    return Err(IoError::other(DiskError::ConditionalDeleteConflict));
+                }
+                if let Some(previous) = old_bytes.as_deref() {
+                    SourceStore::write_owned_backup_std(&source, &base, &root, backup_owner, true, previous, false)?;
+                }
+                SourceStore::persist_state_std(&source, &base, &root, &next_state_raw, terminal.as_deref(), false)?;
+            }
+            #[cfg(test)]
+            super::run_owned_file_before_open(&prepared_path);
+            let mut prepared_metadata_source =
+                os::create_prepared_rename_source_with_commit_guard(&prepared_path, &dst_file_path, &owned_directory_guard)?;
+            prepared_metadata_source.write_all(&new_bytes, true)?;
+            os::install_multipart_prepared_head(prepared_metadata_source, &canonical, &base, &root)
+        })
+        .await
+        .map_err(DiskError::from)?;
+        *self.state.lock() = state;
+        *self.validated_prepared.lock() = Some(validated);
+        Ok(prepared)
+    }
+
+    #[cfg(target_os = "linux")]
+    pub async fn prepare_and_install_head(
+        self: &Arc<Self>,
+        new_metadata: &[u8],
+        old_metadata: Option<&[u8]>,
+        dst_file_path: &Path,
+        directory_guard: &os::RenameCommitGuard,
+        mutation_lease: Arc<os::NamespaceMutationLease>,
+    ) -> Result<os::PreparedRenameSource> {
+        let this = Arc::clone(self);
+        let new_bytes = new_metadata.to_vec();
+        let old_bytes = old_metadata.map(|b| b.to_vec());
+        let dst_file_path = dst_file_path.to_path_buf();
+        let owned_directory_guard = directory_guard.clone();
+        os::run_blocking_namespace_operation_with_sync(mutation_lease, true, move || {
+            this.prepare_and_install_head_on_thread(&new_bytes, old_bytes.as_deref(), &dst_file_path, &owned_directory_guard)
+                .map_err(IoError::other)
+        })
+        .await
+        .map_err(DiskError::from)
+    }
+    #[cfg(test)]
    pub async fn install_head(&self, prepared: os::PreparedRenameSource) -> Result<os::PreparedRenameSource> {
        {
            let guard = self.validated_prepared.lock();
@@ -1699,13 +2264,13 @@ impl NativeMultipartPublication {
        // return an error after consuming the last claimable source heads.
        self.store.restore_exact_target_with_lease(&state, lease).await
    }
-    pub async fn finish(&self, target_lease: Arc<os::NamespaceMutationLease>) -> Result<MultipartPublicationResult> {
+    pub(super) fn finish_on_thread(
+        &self,
+        target_already_synced: bool,
+        staging_cleanup: Option<(PathBuf, Arc<std::sync::atomic::AtomicBool>)>,
+    ) -> Result<MultipartPublicationResult> {
        let mut state = self.state.lock().clone();
-        let current = self.store.load().await?;
-        let current_state = current.state.as_ref().ok_or(DiskError::ConditionalDeleteConflict)?;
-        if current.head.is_some() || encode(current_state)? != encode(&state)? {
-            return Err(DiskError::ConditionalDeleteConflict);
-        }
+        let expected_state_raw = encode(&state)?;
        let (expected_sha256, expected_length) = {
            let guard = self.validated_prepared.lock();
            let validated = guard.as_ref().ok_or(DiskError::ConditionalDeleteConflict)?;
@@ -1717,16 +2282,88 @@ impl NativeMultipartPublication {
            }
            (validated.sha256, validated.length)
        };
-        self.store
-            .read_and_sync_target_with_lease(&self.publication, target_lease, move |target| {
-                if target.len() as u64 != expected_length || raw_sha(target) != expected_sha256 {
-                    return Err(DiskError::ConditionalDeleteConflict);
-                }
-                Ok(())
-            })
-            .await?;
+        let target_dir = self.store.target_path(&self.publication)?;
        state.advance(MultipartSourcePhase::ConsumedUnsealed)?;
-        let snapshot = self.store.persist_with_source_sync(&state, true).await?;
+        state.validate(self.store.disk_id, &self.store.path)?;
+        let next_state_raw = encode(&state)?;
+        let snapshot = snapshot_for_state(self.store.disk_id, &state, &next_state_raw, None)?;
+        let source = &self.store.source;
+        let base = &self.store.base;
+        let root = &self.store.publication_root;
+        let path = &self.store.path;
+        let disk_id = self.store.disk_id;
+        let physical_slot = self.store.physical_slot;
+        let disk_ids = &self.store.disk_ids;
+        let terminal = terminal_record_path(source, &state);
+        (|| -> std::io::Result<()> {
+            let current = load_source(source, base, root, path, disk_id, physical_slot, disk_ids).map_err(IoError::other)?;
+            let current_state = current
+                .state
+                .as_ref()
+                .ok_or_else(|| IoError::other(DiskError::ConditionalDeleteConflict))?;
+            let current_raw = encode(current_state).map_err(IoError::other)?;
+            if current.head.is_some() || current_raw != expected_state_raw {
+                return Err(IoError::other(DiskError::ConditionalDeleteConflict));
+            }
+            if target_already_synced {
+                let pinned = os::pin_multipart_source_directory(&target_dir, base, root).map_err(|error| {
+                    if error.kind() == std::io::ErrorKind::NotFound {
+                        IoError::other(DiskError::ConditionalDeleteConflict)
+                    } else {
+                        error
+                    }
+                })?;
+                let target = os::read_multipart_source_metadata_in_pinned(
+                    &pinned,
+                    &target_dir,
+                    std::ffi::OsStr::new(STORAGE_FORMAT_FILE),
+                    base,
+                    root,
+                )
+                .map_err(|error| {
+                    if error.kind() == std::io::ErrorKind::NotFound {
+                        IoError::other(DiskError::ConditionalDeleteConflict)
+                    } else {
+                        error
+                    }
+                })?;
+                if target.len() as u64 != expected_length || raw_sha(&target) != expected_sha256 {
+                    return Err(IoError::other(DiskError::ConditionalDeleteConflict));
+                }
+            } else {
+                os::read_and_sync_multipart_target_metadata(
+                    &target_dir,
+                    std::ffi::OsStr::new(STORAGE_FORMAT_FILE),
+                    base,
+                    root,
+                    |target| {
+                        if target.len() as u64 != expected_length || raw_sha(target) != expected_sha256 {
+                            return Err(IoError::other(DiskError::ConditionalDeleteConflict));
+                        }
+                        Ok(())
+                    },
+                )
+                .map_err(|error| {
+                    if error.kind() == std::io::ErrorKind::NotFound {
+                        IoError::other(DiskError::ConditionalDeleteConflict)
+                    } else {
+                        error
+                    }
+                })?;
+            }
+            SourceStore::persist_state_std(source, base, root, &next_state_raw, terminal.as_deref(), true)?;
+            if let Some((io_path, cleaned)) = staging_cleanup {
+                match std::fs::remove_dir(&io_path) {
+                    Ok(()) => cleaned.store(true, std::sync::atomic::Ordering::Release),
+                    Err(err) if err.kind() == std::io::ErrorKind::NotFound => {
+                        cleaned.store(true, std::sync::atomic::Ordering::Release);
+                    }
+                    Err(_) => {}
+                }
+            }
+            Ok(())
+        })()
+        .map_err(DiskError::from)?;
        *self.state.lock() = state;
        let proof = self
            .store
@@ -1738,6 +2375,21 @@ impl NativeMultipartPublication {
            undo_capability: capability,
        })
    }
+
+    pub async fn finish(
+        self: &Arc<Self>,
+        target_lease: Arc<os::NamespaceMutationLease>,
+        target_already_synced: bool,
+        staging_cleanup: Option<(PathBuf, Arc<std::sync::atomic::AtomicBool>)>,
+    ) -> Result<MultipartPublicationResult> {
+        let this = Arc::clone(self);
+        os::run_blocking_namespace_operation_with_sync(target_lease, true, move || {
+            this.finish_on_thread(target_already_synced, staging_cleanup)
+                .map_err(IoError::other)
+        })
+        .await
+        .map_err(DiskError::from)
+    }
 }

 impl SourceStore {
@@ -207,6 +207,12 @@ pub(super) struct MultipartSourceAdmission {
    initial: bool,
 }
 impl MultipartSourceAdmission {
+    pub(super) fn source(&self) -> &Path {
+        &self.source
+    }
+    pub(super) fn is_initial(&self) -> bool {
+        self.initial
+    }
    pub(super) fn precondition(&self) -> Arc<dyn Fn() -> std::io::Result<()> + Send + Sync> {
        let admission = self.clone();
        Arc::new(move || admission.validate())
@@ -250,6 +256,8 @@ pub(super) struct MultipartCommitStaging {
    logical_path: PathBuf,
    io_path: PathBuf,
    owner: Option<Arc<dyn Send + Sync>>,
+    cleaned: Arc<std::sync::atomic::AtomicBool>,
+    fresh_source_head: Option<bytes::Bytes>,
 }

 impl MultipartCommitStaging {
@@ -259,10 +267,19 @@ impl MultipartCommitStaging {
    pub(super) fn private_directory(&self, directory: Uuid) -> PathBuf {
        self.io_path.join(directory.to_string())
    }
+    pub(super) fn cleanup_handle(&self) -> (PathBuf, Arc<std::sync::atomic::AtomicBool>) {
+        (self.io_path.clone(), Arc::clone(&self.cleaned))
+    }
+    pub(super) fn fresh_source_head(&self) -> Option<bytes::Bytes> {
+        self.fresh_source_head.clone()
+    }
 }

 impl Drop for MultipartCommitStaging {
    fn drop(&mut self) {
+        if self.cleaned.load(std::sync::atomic::Ordering::Acquire) {
+            return;
+        }
        let logical_path = self.logical_path.clone();
        let io_path = self.io_path.clone();
        let owner = self.owner.clone();
@@ -285,6 +302,14 @@ impl Drop for MultipartCommitStaging {
 }

 impl LocalDisk {
+    pub(crate) fn has_clean_active_multipart_source(&self, upload_or_part_path: &str, part_path: bool) -> bool {
+        let root = upload_scope_root(upload_or_part_path, part_path);
+        let Ok(source) = self.io_get_object_path(RUSTFS_META_MULTIPART_BUCKET, root) else {
+            return false;
+        };
+        super::multipart_source_state::has_clean_active_multipart_source(&source)
+    }
+
    pub(super) fn multipart_source_admission(
        &self,
        volume: &str,
@@ -356,12 +381,14 @@ impl LocalDisk {
        let io_path = self.io_get_object_path(RUSTFS_META_TMP_BUCKET, &attempt)?;
        // The attempt root is exactly the namespace used by the tmp sweeper.
        let lease = os::acquire_namespace_mutation_lease_with_owner(&logical_path, owner.clone()).await;
-        let staging = MultipartCommitStaging {
+        let mut staging = MultipartCommitStaging {
            path: attempt,
            lease,
            logical_path,
            io_path,
            owner,
+            cleaned: Arc::new(std::sync::atomic::AtomicBool::new(false)),
+            fresh_source_head: None,
        };
        let canonical_metadata = self.io_get_object_path(RUSTFS_META_MULTIPART_BUCKET, &format!("{source_path}/xl.meta"))?;
        let source_object = self.io_get_object_path(RUSTFS_META_MULTIPART_BUCKET, source_path)?;
@@ -371,31 +398,37 @@ impl LocalDisk {
        let source_path = source_path.to_owned();
        let incoming = incoming.clone();
        let sync_directory = effective_durability(destination_volume).syncs_data_shards();
-        os::run_cancellable_blocking_namespace_operation(Arc::clone(&staging.lease), move |cancelled| {
+        let fresh_head = os::run_cancellable_blocking_namespace_operation(Arc::clone(&staging.lease), move |cancelled| {
            if cancelled() {
                return Err(std::io::Error::new(std::io::ErrorKind::Interrupted, "multipart staging cancelled"));
            }
-            let raw = os::read_multipart_source_metadata(&canonical_metadata, &base, &publication_root).map_err(|error| {
-                // A retired source has no authority for this delayed completion.
-                // Preserve confinement and other I/O failures as their real errors.
-                if error.kind() == std::io::ErrorKind::NotFound {
-                    DiskError::FileNotFound.into()
-                } else {
-                    error
-                }
-            })?;
-            let source = get_file_info(
-                &raw,
-                RUSTFS_META_MULTIPART_BUCKET,
-                &source_path,
-                "",
-                FileInfoOpts {
-                    data: false,
-                    include_free_versions: false,
-                    include_part_checksums: true,
-                },
-            )
-            .map_err(std::io::Error::other)?;
+            let cached_head = super::multipart_source_state::get_active_multipart_source_head(&source_object);
+            let (raw, source, from_active_cache) = if let Some((raw, source)) = cached_head {
+                (raw, source, true)
+            } else {
+                let raw = os::read_multipart_source_metadata(&canonical_metadata, &base, &publication_root).map_err(|error| {
+                    // A retired source has no authority for this delayed completion.
+                    // Preserve confinement and other I/O failures as their real errors.
+                    if error.kind() == std::io::ErrorKind::NotFound {
+                        DiskError::FileNotFound.into()
+                    } else {
+                        error
+                    }
+                })?;
+                let source = get_file_info(
+                    &raw,
+                    RUSTFS_META_MULTIPART_BUCKET,
+                    &source_path,
+                    "",
+                    FileInfoOpts {
+                        data: false,
+                        include_free_versions: false,
+                        include_part_checksums: true,
+                    },
+                )
+                .map_err(std::io::Error::other)?;
+                (bytes::Bytes::from(raw), source, false)
+            };
            let source_dir = source
                .data_dir
                .filter(|dir| !dir.is_nil())
@@ -412,23 +445,31 @@ impl LocalDisk {
            let _attempt = os::mkdir_all_below_existing_base_std(&private_object, &base, &publication_root)?;
            if incoming.data.is_none() && !incoming.is_remote() {
                let source_data_dir = source_object.join(source_dir.to_string());
-                let pinned_source_data = (!incoming.parts.is_empty())
-                    .then(|| os::pin_multipart_source_directory(&source_data_dir, &base, &publication_root))
-                    .transpose()?;
+                let mut pinned_source_data = None;
                let mut names = Vec::with_capacity(incoming.parts.len().saturating_mul(3));
                for part in &incoming.parts {
                    if cancelled() {
                        return Err(std::io::Error::new(std::io::ErrorKind::Interrupted, "multipart staging cancelled"));
                    }
                    let meta_name = format!("part.{}.meta", part.number);
-                    let bytes = os::read_multipart_source_metadata_in_pinned(
-                        pinned_source_data.as_ref().expect("pinned source data directory"),
-                        &source_data_dir,
-                        std::ffi::OsStr::new(&meta_name),
-                        &base,
-                        &publication_root,
-                    )?;
-                    let actual = rustfs_filemeta::ObjectPartInfo::unmarshal(&bytes).map_err(std::io::Error::other)?;
+                    let actual = if let Some(cached_part) =
+                        super::multipart_source_state::get_active_multipart_source_part(&source_object, part.number)
+                    {
+                        cached_part
+                    } else {
+                        if pinned_source_data.is_none() {
+                            pinned_source_data =
+                                Some(os::pin_multipart_source_directory(&source_data_dir, &base, &publication_root)?);
+                        }
+                        let bytes = os::read_multipart_source_metadata_in_pinned(
+                            pinned_source_data.as_ref().expect("pinned source data directory"),
+                            &source_data_dir,
+                            std::ffi::OsStr::new(&meta_name),
+                            &base,
+                            &publication_root,
+                        )?;
+                        rustfs_filemeta::ObjectPartInfo::unmarshal(&bytes).map_err(std::io::Error::other)?
+                    };
                    let normalized = i64::try_from(actual.size)
                        .map_err(std::io::Error::other)?
                        .max(actual.actual_size);
@@ -464,10 +505,15 @@ impl LocalDisk {
                    cancelled,
                )?;
            }
-            Ok(())
+            let has_state = !from_active_cache
+                && (source_object.join(".mpu-source-state").exists()
+                    || source_object.join(".mpu-recovery").exists()
+                    || source_object.join(".mpu-completions").exists());
+            Ok((!has_state).then_some(raw))
        })
        .await
        .map_err(DiskError::from)?;
+        staging.fresh_source_head = fresh_head;
        Ok(staging)
    }
 }
@@ -1219,4 +1265,211 @@ mod tests {
            .expect("non-part empty hash cleanup must remain supported");
        assert!(!directory.exists(), "a per-upload safety guard must not strand empty hash directories");
    }
+
+    #[tokio::test]
+    async fn multipart_source_active_head_cache_populates_and_invalidates_across_part_and_delete_paths() {
+        use crate::disk::{DiskAPI as _, PartTransactionAction, ReadOptions};
+        let (_root, disk, completed, _) = fixture().await;
+        std::fs::create_dir_all(disk.io_get_bucket_path(RUSTFS_META_TMP_BUCKET).expect("TMP path")).expect("TMP bucket");
+        let upload = format!("{}/{}", "c".repeat(64), Uuid::new_v4());
+        let data_dir = Uuid::new_v4();
+        let mut init_fi = FileInfo::new(&upload, 2, 2);
+        init_fi.erasure.index = 1;
+        init_fi.data_dir = Some(data_dir);
+        init_fi.fresh = true;
+        init_fi.mod_time = Some(time::OffsetDateTime::now_utc());
+
+        disk.write_metadata("", RUSTFS_META_MULTIPART_BUCKET, &upload, init_fi.clone())
+            .await
+            .expect("fresh multipart upload metadata");
+        assert!(disk.has_clean_active_multipart_source(&upload, false));
+
+        let read_fi = disk
+            .read_version("", RUSTFS_META_MULTIPART_BUCKET, &upload, "", &ReadOptions::default())
+            .await
+            .expect("cached read_version");
+        assert_eq!(read_fi.data_dir, Some(data_dir));
+
+        let destination = format!("{upload}/{data_dir}/part.1");
+        let part_meta = bytes::Bytes::from(completed.parts[0].marshal_msg().expect("part meta"));
+        disk.write_all(
+            RUSTFS_META_TMP_BUCKET,
+            "cache-test-part/part.1",
+            bytes::Bytes::from_static(b"immutable shard"),
+        )
+        .await
+        .expect("stage tmp part");
+        disk.prepare_part_transaction(
+            RUSTFS_META_TMP_BUCKET,
+            "cache-test-part/part.1",
+            RUSTFS_META_MULTIPART_BUCKET,
+            &destination,
+            part_meta.clone(),
+        )
+        .await
+        .expect("prepare part transaction");
+        assert!(
+            !disk.has_clean_active_multipart_source(&upload, false),
+            "in-flight part transaction must mark active source as having unsettled part txn"
+        );
+        disk.rename_part(
+            RUSTFS_META_TMP_BUCKET,
+            "cache-test-part/part.1",
+            RUSTFS_META_MULTIPART_BUCKET,
+            &destination,
+            part_meta,
+        )
+        .await
+        .expect("rename part");
+        disk.settle_part_transaction(RUSTFS_META_MULTIPART_BUCKET, &destination, PartTransactionAction::Commit)
+            .await
+            .expect("settle part transaction");
+        assert!(
+            disk.has_clean_active_multipart_source(&upload, false),
+            "settled commit must restore clean active multipart source status"
+        );
+        let read_parts = disk
+            .read_parts(RUSTFS_META_MULTIPART_BUCKET, &[format!("{destination}.meta")])
+            .await
+            .expect("cached read_parts");
+        assert_eq!(read_parts.len(), 1);
+        assert_eq!(read_parts[0].number, 1);
+
+        // Prepare part 2, then replay Commit for already-settled part 1: part 2's
+        // pending transaction must keep the upload marked as having unsettled part txns.
+        let destination2 = format!("{upload}/{data_dir}/part.2");
+        let mut part2_info = completed.parts[0].clone();
+        part2_info.number = 2;
+        let part2_meta = bytes::Bytes::from(part2_info.marshal_msg().expect("part 2 meta"));
+        disk.write_all(
+            RUSTFS_META_TMP_BUCKET,
+            "cache-test-part/part.2",
+            bytes::Bytes::from_static(b"immutable shard"),
+        )
+        .await
+        .expect("stage tmp part 2");
+        disk.prepare_part_transaction(
+            RUSTFS_META_TMP_BUCKET,
+            "cache-test-part/part.2",
+            RUSTFS_META_MULTIPART_BUCKET,
+            &destination2,
+            part2_meta.clone(),
+        )
+        .await
+        .expect("prepare part 2 transaction");
+        assert!(!disk.has_clean_active_multipart_source(&upload, false));
+        disk.settle_part_transaction(RUSTFS_META_MULTIPART_BUCKET, &destination, PartTransactionAction::Commit)
+            .await
+            .expect("replayed part 1 commit settlement");
+        assert!(
+            !disk.has_clean_active_multipart_source(&upload, false),
+            "replayed settlement for part 1 must not clear part 2's pending transaction"
+        );
+        disk.rename_part(
+            RUSTFS_META_TMP_BUCKET,
+            "cache-test-part/part.2",
+            RUSTFS_META_MULTIPART_BUCKET,
+            &destination2,
+            part2_meta,
+        )
+        .await
+        .expect("rename part 2");
+        disk.settle_part_transaction(RUSTFS_META_MULTIPART_BUCKET, &destination2, PartTransactionAction::Commit)
+            .await
+            .expect("settle part 2 transaction");
+        assert!(disk.has_clean_active_multipart_source(&upload, false));
+
+        // Deleting {upload}/xl.meta through delete_paths must invalidate the canonical {upload} cache key.
+        disk.delete_paths(RUSTFS_META_MULTIPART_BUCKET, &[format!("{upload}/xl.meta")])
+            .await
+            .expect("delete xl.meta via delete_paths");
+        assert!(
+            !disk.has_clean_active_multipart_source(&upload, false),
+            "per-file head deletion must invalidate the canonical upload cache key"
+        );
+        assert!(matches!(
+            disk.read_version("", RUSTFS_META_MULTIPART_BUCKET, &upload, "", &ReadOptions::default())
+                .await,
+            Err(DiskError::FileNotFound)
+        ));
+    }
+
+    #[cfg(not(windows))]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    async fn multipart_source_cache_publication_holds_source_lock_against_concurrent_retirement() {
+        use crate::disk::{DiskAPI as _, disk_store::LocalDiskWrapper, os::prepared_publication_test_hooks as hooks};
+
+        let (_root, disk, _, _) = fixture().await;
+        std::fs::create_dir_all(disk.io_get_bucket_path(RUSTFS_META_TMP_BUCKET).expect("TMP path")).expect("TMP bucket");
+        let disk = Arc::new(disk);
+        let wrapper = LocalDiskWrapper::new(Arc::clone(&disk), false);
+        let upload = format!("{}/{}", "d".repeat(64), Uuid::new_v4());
+        let data_dir = Uuid::new_v4();
+        let mut init_fi = FileInfo::new(&upload, 2, 2);
+        init_fi.erasure.index = 1;
+        init_fi.data_dir = Some(data_dir);
+        init_fi.fresh = true;
+        init_fi.mod_time = Some(time::OffsetDateTime::now_utc());
+
+        let source_dir = disk
+            .io_get_object_path(RUSTFS_META_MULTIPART_BUCKET, &upload)
+            .expect("source dir");
+        let base_dir = disk
+            .io_get_bucket_path(RUSTFS_META_MULTIPART_BUCKET)
+            .expect("multipart bucket path");
+        let (entered_tx, entered_rx) = tokio::sync::oneshot::channel();
+        let (release_tx, release_rx) = std::sync::mpsc::channel();
+        let _hook = hooks::install_at(hooks::Stage::DirFsync, &source_dir, move || {
+            let _ = entered_tx.send(());
+            let _ = release_rx.recv();
+        });
+
+        let init_wrapper = wrapper.clone();
+        let init_upload = upload.clone();
+        let init_task = tokio::spawn(async move {
+            init_wrapper
+                .write_metadata("", RUSTFS_META_MULTIPART_BUCKET, &init_upload, init_fi)
+                .await
+        });
+
+        tokio::time::timeout(std::time::Duration::from_secs(5), entered_rx)
+            .await
+            .expect("initial write_metadata reaches parent dir fsync")
+            .expect("entered channel");
+
+        let retire_wrapper = wrapper.clone();
+        let retire_path = format!("{upload}/xl.meta");
+        let mut retire_fut = Box::pin(async move {
+            retire_wrapper
+                .delete_paths(RUSTFS_META_MULTIPART_BUCKET, &[retire_path])
+                .await
+        });
+        assert!(
+            futures::poll!(tokio::task::unconstrained(retire_fut.as_mut())).is_pending(),
+            "concurrent retirement must block on SourceRW while initial head publication is in flight"
+        );
+
+        let _ = release_tx.send(());
+        init_task.await.expect("init join").expect("initial write_metadata succeeds");
+        retire_fut
+            .await
+            .expect("concurrent retirement completes after SourceRW release");
+
+        assert!(
+            !super::super::multipart_source_state::has_active_multipart_source_head(&source_dir),
+            "retired source must not remain active in cache after interleaved initial publication"
+        );
+        assert!(
+            matches!(
+                super::super::multipart_source_state::validate_source_mutation(
+                    &source_dir,
+                    &base_dir,
+                    &disk.publication_root,
+                    false
+                ),
+                Err(DiskError::FileNotFound)
+            ),
+            "validate_source_mutation must reject non-initial mutation on retired source"
+        );
+    }
 }
@@ -401,6 +401,13 @@ impl Disk {
        }
    }

+    pub(crate) fn has_clean_active_multipart_source(&self, upload_or_part_path: &str, part_path: bool) -> bool {
+        match self {
+            Self::Local(disk) => disk.has_clean_active_multipart_source(upload_or_part_path, part_path),
+            Self::Remote(_) => false,
+        }
+    }
+
    pub(crate) async fn try_publish_prepared_delete_local(
        &self,
        volume: &str,
@@ -1668,7 +1668,7 @@ fn clear_file_fdatasync_group_commit_for_test() {

 // Small object directories are cheaper to flush in one blocking task. Multipart
 // directories fan out only once enough files can amortize per-task scheduling.
-const PARALLEL_FILE_SYNC_THRESHOLD: usize = 16;
+pub(crate) const PARALLEL_FILE_SYNC_THRESHOLD: usize = 16;
 pub(crate) const MAX_PARALLEL_FILE_SYNCS: usize = 16;
 // Scale aggregate fan-out for wider nodes while reserving at least half of the
 // configured Tokio blocking pool for unrelated filesystem work.
@@ -1782,10 +1782,13 @@ pub(crate) fn disk_file_sync_limiter(root: &Path) -> Arc<Semaphore> {
    limiter
 }

-// Keep one warm draining worker per disk/lane and spin briefly in user space
+// Keep warm draining workers per disk/lane and spin briefly in user space
 // before exiting `spawn_blocking` so back-to-back stages of `PUT` and `DELETE`
-// on the same drive do not thrash Tokio's `BlockingPool` (`futex_wake`/`futex_wait`).
+// on the same drive do not thrash Tokio's `BlockingPool` (`futex_wake`/`futex_wait`),
+// while allowing concurrent fsync callers to enter the kernel simultaneously and
+// coalesce in the filesystem journal (`jbd2`/`xfs` group commit).
 const MAX_NAMESPACE_WORKERS_PER_DISK: usize = 1;
+pub(crate) const MAX_FSYNC_WORKERS_PER_DISK: usize = 8;
 // Return busy workers to the shared pool so other disks can acquire a slot.
 const MAX_BLOCKING_JOBS_PER_WORKER: usize = 64;
 const WORKER_IDLE_SPIN_ITERS: usize = 256;
@@ -1813,6 +1816,7 @@ impl DiskBlockingLane {
 struct DiskBlockingQueueState {
    jobs: VecDeque<Box<dyn FnOnce() + Send>>,
    active_workers: usize,
+    idle_workers: usize,
    #[cfg(any(test, feature = "test-util"))]
    paused_workers: usize,
 }
@@ -1856,6 +1860,7 @@ impl Drop for DiskBlockingWorkerGuard {
            let cancelled = {
                let mut state = self.queue.lane(self.lane).state.lock();
                state.active_workers = state.active_workers.saturating_sub(1);
+                state.idle_workers = state.idle_workers.saturating_sub(1);
                if state.active_workers == 0 {
                    self.queue
                        .lane(self.lane)
@@ -1874,9 +1879,10 @@ impl Drop for DiskBlockingWorkerGuard {
        let should_respawn = {
            let mut state = self.queue.lane(self.lane).state.lock();
            state.active_workers = state.active_workers.saturating_sub(1);
-            let max_workers = self.queue.effective_max_workers(&state);
-            if !state.jobs.is_empty() && state.active_workers < max_workers {
+            let max_workers = self.queue.effective_max_workers(self.lane, &state);
+            if state.jobs.len() > state.idle_workers && state.active_workers < max_workers {
                state.active_workers += 1;
+                state.idle_workers += 1;
                true
            } else {
                false
@@ -1901,15 +1907,23 @@ impl DiskBlockingQueue {
    }

    #[inline]
-    fn effective_max_workers(&self, state: &DiskBlockingQueueState) -> usize {
+    fn lane_max_workers(&self, lane: DiskBlockingLane) -> usize {
+        match lane {
+            DiskBlockingLane::Foreground | DiskBlockingLane::Isolated => MAX_NAMESPACE_WORKERS_PER_DISK,
+            DiskBlockingLane::Fsync => MAX_FSYNC_WORKERS_PER_DISK,
+        }
+    }
+
+    #[inline]
+    fn effective_max_workers(&self, lane: DiskBlockingLane, state: &DiskBlockingQueueState) -> usize {
        #[cfg(any(test, feature = "test-util"))]
        {
-            MAX_NAMESPACE_WORKERS_PER_DISK.saturating_add(state.paused_workers)
+            self.lane_max_workers(lane).saturating_add(state.paused_workers)
        }
        #[cfg(not(any(test, feature = "test-util")))]
        {
            let _ = state;
-            MAX_NAMESPACE_WORKERS_PER_DISK
+            self.lane_max_workers(lane)
        }
    }

@@ -1918,8 +1932,9 @@ impl DiskBlockingQueue {
            let mut state = self.lane(lane).state.lock();
            state.jobs.push_back(job);
            self.lane(lane).ready.store(true, std::sync::atomic::Ordering::Relaxed);
-            if state.active_workers < self.effective_max_workers(&state) {
+            if state.jobs.len() > state.idle_workers && state.active_workers < self.effective_max_workers(lane, &state) {
                state.active_workers += 1;
+                state.idle_workers += 1;
                true
            } else {
                false
@@ -1962,13 +1977,19 @@ impl DiskBlockingQueue {
            CURRENT_DISK_BLOCKING_QUEUE.with(|slot| {
                *slot.borrow_mut() = Some((Arc::clone(&queue), lane));
            });
-            guard.state = DiskBlockingWorkerState::Running;
            let mut remaining_jobs = MAX_BLOCKING_JOBS_PER_WORKER;
            let mut idle_polled = false;
+            let mut counted_idle = true;
            loop {
                let job = {
                    let mut state = queue.lane(lane).state.lock();
-                    if state.active_workers > queue.effective_max_workers(&state) || (idle_polled && state.jobs.is_empty()) {
+                    if counted_idle {
+                        state.idle_workers = state.idle_workers.saturating_sub(1);
+                        counted_idle = false;
+                    }
+                    guard.state = DiskBlockingWorkerState::Running;
+                    if state.active_workers > queue.effective_max_workers(lane, &state) || (idle_polled && state.jobs.is_empty())
+                    {
                        state.active_workers = state.active_workers.saturating_sub(1);
                        guard.state = DiskBlockingWorkerState::Retired;
                        return;
@@ -1978,6 +1999,15 @@ impl DiskBlockingQueue {
                        .lane(lane)
                        .ready
                        .store(!state.jobs.is_empty(), std::sync::atomic::Ordering::Relaxed);
+                    if job.is_none() {
+                        if state.idle_workers > 0 {
+                            state.active_workers = state.active_workers.saturating_sub(1);
+                            guard.state = DiskBlockingWorkerState::Retired;
+                            return;
+                        }
+                        state.idle_workers += 1;
+                        counted_idle = true;
+                    }
                    job
                };
                if let Some(job) = job {
@@ -1988,15 +2018,11 @@ impl DiskBlockingQueue {
                    }
                    idle_polled = false;
                } else {
-                    for i in 0..WORKER_IDLE_SPIN_ITERS {
+                    for _ in 0..WORKER_IDLE_SPIN_ITERS {
                        if queue.lane(lane).ready.load(std::sync::atomic::Ordering::Relaxed) {
                            break;
                        }
-                        if i & 15 == 15 {
-                            std::thread::yield_now();
-                        } else {
-                            std::hint::spin_loop();
-                        }
+                        std::hint::spin_loop();
                    }
                    // Recheck and retire under the same mutex used by enqueue.
                    idle_polled = true;
@@ -2106,8 +2132,9 @@ pub(crate) fn enter_blocking_test_hook() -> BlockingTestHookGuard {
        let should_spawn = {
            let mut state = queue.lane(lane).state.lock();
            state.paused_workers = state.paused_workers.saturating_add(1);
-            if !state.jobs.is_empty() && state.active_workers < queue.effective_max_workers(&state) {
+            if state.jobs.len() > state.idle_workers && state.active_workers < queue.effective_max_workers(lane, &state) {
                state.active_workers += 1;
+                state.idle_workers += 1;
                true
            } else {
                false
@@ -2569,7 +2596,7 @@ fn sync_files(paths: &[PathBuf]) -> io::Result<()> {
    Ok(())
 }

-fn regular_files(dir: &Path) -> io::Result<Vec<PathBuf>> {
+pub(crate) fn regular_files(dir: &Path) -> io::Result<Vec<PathBuf>> {
    let mut files = Vec::with_capacity(PARALLEL_FILE_SYNC_THRESHOLD);
    for entry in std::fs::read_dir(dir)? {
        let entry = entry?;
@@ -2752,22 +2779,114 @@ pub(crate) fn read_dir_probe(path: impl AsRef<Path>, entry_limit: usize) -> io::
    })
 }

+#[cfg(test)]
 pub(crate) fn has_non_data_dir_entry(path: impl AsRef<Path>, data_dirs: &[uuid::Uuid]) -> io::Result<bool> {
-    match has_non_data_dir_entry_inner(path.as_ref(), data_dirs) {
-        Ok(res) => Ok(res),
-        Err(err) if err.kind() == io::ErrorKind::NotFound => Ok(false),
-        Err(err) => Err(err),
-    }
+    has_non_data_dir_entry_lazy(path, || data_dirs.to_vec()).map(|(_, res)| res)
 }

-fn has_non_data_dir_entry_inner(path: &Path, data_dirs: &[uuid::Uuid]) -> io::Result<bool> {
-    let mut dir = std::fs::read_dir(path)?;
-    let entry_limit = data_dirs.len().saturating_add(2);
+#[cfg(target_os = "linux")]
+pub(crate) fn has_non_data_dir_entry_lazy_from_fd(
+    dir_fd: &std::fs::File,
+    load_data_dirs: impl FnOnce() -> Vec<uuid::Uuid>,
+) -> io::Result<(Vec<uuid::Uuid>, bool)> {
+    use rustix::fs::{AtFlags, FileType, RawDir, statat};
+
+    let mut raw_buf = [std::mem::MaybeUninit::<u8>::uninit(); 1024];
+    let mut dir = RawDir::new(dir_fd, &mut raw_buf);
+    let mut data_dirs: Option<Vec<uuid::Uuid>> = None;
+    let mut load = Some(load_data_dirs);
    let mut valid_entries = 0usize;

    for _ in 0..READ_DIR_PROBE_RAW_LIMIT {
        let Some(entry) = dir.next() else {
-            return Ok(false);
+            return Ok((data_dirs.unwrap_or_default(), false));
+        };
+        let entry = match entry {
+            Ok(entry) => entry,
+            Err(rustix::io::Errno::NOENT) => return Ok((data_dirs.unwrap_or_default(), false)),
+            Err(err) => return Err(io::Error::from(err)),
+        };
+        let c_name = entry.file_name();
+        let name_bytes = c_name.to_bytes();
+        if name_bytes.is_empty() || name_bytes == b"." || name_bytes == b".." {
+            continue;
+        }
+
+        let file_type = match entry.file_type() {
+            FileType::Unknown => match statat(dir_fd, c_name, AtFlags::SYMLINK_NOFOLLOW) {
+                Ok(st) => FileType::from_raw_mode(st.st_mode),
+                Err(rustix::io::Errno::NOENT) => continue,
+                Err(err) => return Err(io::Error::from(err)),
+            },
+            ft => ft,
+        };
+        if file_type == FileType::Directory {
+            let parsed_uuid = if name_bytes.len() == uuid::fmt::Hyphenated::LENGTH {
+                std::str::from_utf8(name_bytes).ok().and_then(|s| {
+                    let parsed = uuid::Uuid::try_parse(s).ok()?;
+                    let mut buf = [0u8; uuid::fmt::Hyphenated::LENGTH];
+                    (parsed.hyphenated().encode_lower(&mut buf) == s).then_some(parsed)
+                })
+            } else {
+                None
+            };
+            let Some(parsed) = parsed_uuid else {
+                let dirs = data_dirs
+                    .take()
+                    .unwrap_or_else(|| (load.take().expect("load_data_dirs called once"))());
+                return Ok((dirs, true));
+            };
+            let dirs = data_dirs.get_or_insert_with(|| (load.take().expect("load_data_dirs called once"))());
+            if dirs.binary_search(&parsed).is_err() {
+                return Ok((std::mem::take(dirs), true));
+            }
+            valid_entries += 1;
+        } else if file_type == FileType::RegularFile {
+            valid_entries += 1;
+        } else {
+            continue;
+        }
+
+        let entry_limit = data_dirs.as_ref().map_or(2, |dirs| dirs.len().saturating_add(2));
+        if valid_entries == entry_limit {
+            let dirs = data_dirs.get_or_insert_with(|| (load.take().expect("load_data_dirs called once"))());
+            if valid_entries == dirs.len().saturating_add(2) {
+                return Ok((std::mem::take(dirs), true));
+            }
+        }
+    }
+
+    let dirs = data_dirs
+        .take()
+        .unwrap_or_else(|| (load.take().expect("load_data_dirs called once"))());
+    Ok((dirs, true))
+}
+
+#[cfg(any(not(target_os = "linux"), test))]
+pub(crate) fn has_non_data_dir_entry_lazy(
+    path: impl AsRef<Path>,
+    load_data_dirs: impl FnOnce() -> Vec<uuid::Uuid>,
+) -> io::Result<(Vec<uuid::Uuid>, bool)> {
+    match has_non_data_dir_entry_lazy_inner(path.as_ref(), load_data_dirs) {
+        Ok(res) => Ok(res),
+        Err(err) if err.kind() == io::ErrorKind::NotFound => Ok((Vec::new(), false)),
+        Err(err) => Err(err),
+    }
+}
+
+#[cfg(any(not(target_os = "linux"), test))]
+fn has_non_data_dir_entry_lazy_inner(
+    path: &Path,
+    load_data_dirs: impl FnOnce() -> Vec<uuid::Uuid>,
+) -> io::Result<(Vec<uuid::Uuid>, bool)> {
+    let mut dir = std::fs::read_dir(path)?;
+    let mut data_dirs: Option<Vec<uuid::Uuid>> = None;
+    let mut load = Some(load_data_dirs);
+    let mut valid_entries = 0usize;
+
+    for _ in 0..READ_DIR_PROBE_RAW_LIMIT {
+        let Some(entry) = dir.next() else {
+            return Ok((data_dirs.unwrap_or_default(), false));
        };
        let entry = entry?;
        let file_name = entry.file_name();
@@ -2777,18 +2896,24 @@ fn has_non_data_dir_entry_inner(path: &Path, data_dirs: &[uuid::Uuid]) -> io::Re

        let file_type = entry.file_type()?;
        if file_type.is_dir() {
-            let is_data_dir = file_name.len() == uuid::fmt::Hyphenated::LENGTH
-                && file_name
-                    .to_str()
-                    .and_then(|s| {
-                        let parsed = uuid::Uuid::try_parse(s).ok()?;
-                        let mut buf = [0u8; uuid::fmt::Hyphenated::LENGTH];
-                        (parsed.hyphenated().encode_lower(&mut buf) == s && data_dirs.binary_search(&parsed).is_ok())
-                            .then_some(())
-                    })
-                    .is_some();
-            if !is_data_dir {
-                return Ok(true);
+            let parsed_uuid = if file_name.len() == uuid::fmt::Hyphenated::LENGTH {
+                file_name.to_str().and_then(|s| {
+                    let parsed = uuid::Uuid::try_parse(s).ok()?;
+                    let mut buf = [0u8; uuid::fmt::Hyphenated::LENGTH];
+                    (parsed.hyphenated().encode_lower(&mut buf) == s).then_some(parsed)
+                })
+            } else {
+                None
+            };
+            let Some(parsed) = parsed_uuid else {
+                let dirs = data_dirs
+                    .take()
+                    .unwrap_or_else(|| (load.take().expect("load_data_dirs called once"))());
+                return Ok((dirs, true));
+            };
+            let dirs = data_dirs.get_or_insert_with(|| (load.take().expect("load_data_dirs called once"))());
+            if dirs.binary_search(&parsed).is_err() {
+                return Ok((std::mem::take(dirs), true));
            }
            valid_entries += 1;
        } else if file_type.is_file() {
@@ -2797,12 +2922,19 @@ fn has_non_data_dir_entry_inner(path: &Path, data_dirs: &[uuid::Uuid]) -> io::Re
            continue;
        }

+        let entry_limit = data_dirs.as_ref().map_or(2, |dirs| dirs.len().saturating_add(2));
        if valid_entries == entry_limit {
-            return Ok(true);
+            let dirs = data_dirs.get_or_insert_with(|| (load.take().expect("load_data_dirs called once"))());
+            if valid_entries == dirs.len().saturating_add(2) {
+                return Ok((std::mem::take(dirs), true));
+            }
        }
    }

-    Ok(true)
+    let dirs = data_dirs
+        .take()
+        .unwrap_or_else(|| (load.take().expect("load_data_dirs called once"))());
+    Ok((dirs, true))
 }

 // read_dir  count read limit. when count == 0 unlimit.
@@ -3637,16 +3769,21 @@ impl PreparedRenameSource {
        #[cfg(not(windows))]
        std::io::Write::write_all(&mut self.source, data)?;
        if sync {
-            #[cfg(test)]
-            multipart_source_syscall_failures::check(multipart_source_syscall_failures::Stage::FileSync, &self.test_destination)?;
-            let _sync = StageTimer::new(true, "disk_prepared_metadata_fdatasync");
-            #[cfg(windows)]
-            self.source.as_file().sync_data()?;
-            #[cfg(not(windows))]
-            self.source.sync_data()?;
+            self.sync_data()?;
        }
        Ok(())
    }
+
+    pub(crate) fn sync_data(&self) -> io::Result<()> {
+        #[cfg(test)]
+        multipart_source_syscall_failures::check(multipart_source_syscall_failures::Stage::FileSync, &self.test_destination)?;
+        let _sync = StageTimer::new(true, "disk_prepared_metadata_fdatasync");
+        #[cfg(windows)]
+        self.source.as_file().sync_data()?;
+        #[cfg(not(windows))]
+        self.source.sync_data()?;
+        Ok(())
+    }
 }

 pub(crate) fn create_prepared_rename_source_with_commit_guard(
@@ -4832,6 +4969,17 @@ impl PublicationRoot {
        )
    }

+    #[cfg(target_os = "linux")]
+    pub(crate) fn open_directory_and_readonly_child(
+        &self,
+        dir_path: &Path,
+        child_name: &str,
+    ) -> io::Result<(std::fs::File, std::fs::File)> {
+        let dir = self.open_directory(dir_path)?;
+        let child = open_linux_relative(&dir, Path::new(child_name), rustix::fs::OFlags::RDONLY, rustix::fs::Mode::empty())?;
+        Ok((dir, child))
+    }
+
    #[cfg(target_os = "linux")]
    fn relative_path<'a>(&self, path: &'a Path) -> io::Result<&'a Path> {
        path.strip_prefix(self.io_path.as_path())
@@ -5297,6 +5445,34 @@ impl RenameCommitGuard {
    }
 }

+#[cfg(target_os = "linux")]
+pub(crate) fn multipart_payload_commit_guard_std(
+    metadata_guard: &RenameCommitGuard,
+    source_parent: &Path,
+) -> io::Result<RenameCommitGuard> {
+    let original = &metadata_guard.directories;
+    original.validate()?;
+    let source = LinuxCommitDirectory::new(original.root.open_directory(source_parent)?)?;
+    let destination = original.prepare_destination()?;
+    // Clone the pinned file identity, never recapture the target pathname
+    // between payload and canonical-head publication.
+    let destination = LinuxCommitDirectory {
+        file: Arc::clone(&destination.file),
+        device: destination.device,
+        inode: destination.inode,
+    };
+    Ok(RenameCommitGuard {
+        directories: Arc::new(LinuxCommitDirectories {
+            root: original.root.clone(),
+            source_path: source_parent.to_path_buf(),
+            destination_path: original.destination_path.clone(),
+            source,
+            destination: std::sync::OnceLock::from(destination),
+            destination_base: original.destination_base.clone(),
+        }),
+    })
+}
+
 /// Native MPU metadata and payload have distinct source parents. Reuse the
 /// exact target identity already held by the metadata transaction, while
 /// independently pinning the private payload source under the same root.
@@ -5308,30 +5484,7 @@ pub(crate) async fn multipart_payload_commit_guard(
 ) -> io::Result<RenameCommitGuard> {
    let metadata_guard = metadata_guard.clone();
    let source_parent = source_parent.to_path_buf();
-    run_blocking_namespace_operation(lease, move || {
-        let original = &metadata_guard.directories;
-        original.validate()?;
-        let source = LinuxCommitDirectory::new(original.root.open_directory(&source_parent)?)?;
-        let destination = original.prepare_destination()?;
-        // Clone the pinned file identity, never recapture the target pathname
-        // between payload and canonical-head publication.
-        let destination = LinuxCommitDirectory {
-            file: Arc::clone(&destination.file),
-            device: destination.device,
-            inode: destination.inode,
-        };
-        Ok(RenameCommitGuard {
-            directories: Arc::new(LinuxCommitDirectories {
-                root: original.root.clone(),
-                source_path: source_parent,
-                destination_path: original.destination_path.clone(),
-                source,
-                destination: std::sync::OnceLock::from(destination),
-                destination_base: original.destination_base.clone(),
-            }),
-        })
-    })
-    .await
+    run_blocking_namespace_operation(lease, move || multipart_payload_commit_guard_std(&metadata_guard, &source_parent)).await
 }

 pub(crate) fn prepare_rename_commit_guard(
@@ -6513,6 +6666,17 @@ pub(crate) fn sync_multipart_owned_record_parents(
    include_leaf: bool,
    base: &Path,
    publication_root: &PublicationRoot,
+) -> io::Result<()> {
+    sync_multipart_owned_record_parents_options(source, record_parent, include_leaf, true, base, publication_root)
+}
+
+pub(crate) fn sync_multipart_owned_record_parents_options(
+    source: &Path,
+    record_parent: &Path,
+    include_leaf: bool,
+    sync_source_root: bool,
+    base: &Path,
+    publication_root: &PublicationRoot,
 ) -> io::Result<()> {
    let relative = record_parent
        .strip_prefix(source)
@@ -6536,6 +6700,9 @@ pub(crate) fn sync_multipart_owned_record_parents(
        if !directory.starts_with(source) {
            return Err(io::Error::new(io::ErrorKind::InvalidInput, "native record barrier escaped source"));
        }
+        if directory == source && !sync_source_root {
+            return Ok(());
+        }
        #[cfg_attr(not(unix), allow(unused_mut))]
        let mut pinned = pin_multipart_source_directory(directory, base, publication_root)?;
        #[cfg(unix)]
@@ -10985,6 +11152,54 @@ mod tests {
        assert_eq!(err.kind(), io::ErrorKind::NotFound);
    }

+    #[test]
+    fn has_non_data_dir_entry_lazy_skips_loader_when_only_xl_meta_exists() {
+        let temp_dir = tempdir().expect("create temp dir");
+        std::fs::write(temp_dir.path().join("xl.meta"), b"meta").expect("write xl.meta");
+        let mut called = 0usize;
+        let (dirs, has_child) = has_non_data_dir_entry_lazy(temp_dir.path(), || {
+            called += 1;
+            vec![uuid::Uuid::new_v4()]
+        })
+        .expect("lazy check must succeed");
+        assert_eq!(called, 0, "inline object dir without subdirs must not invoke loader");
+        assert!(dirs.is_empty());
+        assert!(!has_child);
+
+        #[cfg(target_os = "linux")]
+        {
+            let pub_root = PublicationRoot::new(temp_dir.path()).expect("open pub root");
+            let (dir_fd, _meta) = pub_root
+                .open_directory_and_readonly_child(temp_dir.path(), "xl.meta")
+                .expect("open dir and child");
+            let mut fd_called = 0usize;
+            let (fd_dirs, fd_has_child) = has_non_data_dir_entry_lazy_from_fd(&dir_fd, || {
+                fd_called += 1;
+                vec![uuid::Uuid::new_v4()]
+            })
+            .expect("fd lazy check must succeed");
+            assert_eq!(fd_called, 0);
+            assert!(fd_dirs.is_empty());
+            assert!(!fd_has_child);
+        }
+
+        let data_dir = uuid::Uuid::new_v4();
+        std::fs::create_dir(temp_dir.path().join(data_dir.to_string())).expect("create data dir");
+        let (dirs, has_child) = has_non_data_dir_entry_lazy(temp_dir.path(), || {
+            called += 1;
+            vec![data_dir]
+        })
+        .expect("lazy check with data dir must succeed");
+        assert_eq!(called, 1, "data dir presence must invoke loader once");
+        assert_eq!(dirs, vec![data_dir]);
+        assert!(!has_child);
+
+        std::fs::create_dir(temp_dir.path().join("nested-key")).expect("create nested child dir");
+        let (dirs, has_child) = has_non_data_dir_entry_lazy(temp_dir.path(), || vec![data_dir]).expect("lazy check child");
+        assert_eq!(dirs, vec![data_dir], "child presence must still return data_dirs for skip set");
+        assert!(has_child);
+    }
+
    #[tokio::test]
    #[serial_test::serial(file_sync_probe)]
    async fn resource_grouped_fdatasync_retired_limiter_keeps_its_budget_until_worker_cleanup() {
@@ -11228,19 +11443,24 @@ mod tests {
            let queue = register_disk_blocking_queue(root.path(), root.path());
            std::fs::write(root.path().join("intermediate"), b"backup").expect("backup");
            std::fs::write(root.path().join("receipt"), b"committed").expect("receipt");
-            let (started_tx, started_rx) = oneshot::channel();
-            let (release_tx, release_rx) = oneshot::channel();
-            let blocker_queue = Arc::clone(&queue);
-            let blocker = tokio::spawn(async move {
-                blocker_queue
-                    .run_fsync_cancellable(move |_| {
-                        let _ = started_tx.send(());
-                        let _ = release_rx.blocking_recv();
-                        Ok(())
-                    })
-                    .await
-            });
-            started_rx.await.expect("fsync lane blocked");
+            let mut release_txs = Vec::new();
+            let mut blockers = Vec::new();
+            for _ in 0..MAX_FSYNC_WORKERS_PER_DISK {
+                let (started_tx, started_rx) = oneshot::channel();
+                let (release_tx, release_rx) = oneshot::channel();
+                let blocker_queue = Arc::clone(&queue);
+                blockers.push(tokio::spawn(async move {
+                    blocker_queue
+                        .run_fsync_cancellable(move |_| {
+                            let _ = started_tx.send(());
+                            let _ = release_rx.blocking_recv();
+                            Ok(())
+                        })
+                        .await
+                }));
+                started_rx.await.expect("fsync worker blocked");
+                release_txs.push(release_tx);
+            }
            let mut cleanup = std::pin::pin!(remove_files_and_fsync_with_owner(
                root.path(),
                &["intermediate"],
@@ -11251,8 +11471,12 @@ mod tests {
            assert!(futures::FutureExt::now_or_never(&mut cleanup).is_none());
            let foreground = tokio::time::timeout(Duration::from_secs(5), queue.run(|| Ok(()))).await;
            let retained = root.path().join("receipt").exists();
-            release_tx.send(()).expect("release fsync lane");
-            blocker.await.expect("blocker joined").expect("blocker completed");
+            for release_tx in release_txs {
+                release_tx.send(()).expect("release fsync worker");
+            }
+            for blocker in blockers {
+                blocker.await.expect("blocker joined").expect("blocker completed");
+            }
            cleanup.await.expect("cleanup completed");
            foreground.expect("foreground must progress").expect("foreground operation");
            assert_eq!(retained, sync, "only durable cleanup must wait for the fsync lane");
@@ -11474,6 +11698,25 @@ mod tests {

        for lane in [DiskBlockingLane::Foreground, DiskBlockingLane::Fsync] {
            let queue = DiskBlockingQueue::new();
+            let max_workers = queue.lane_max_workers(lane);
+            let mut hold_txs = Vec::new();
+            let mut hold_tasks = Vec::new();
+            for _ in 1..max_workers {
+                let (started_tx, started_rx) = oneshot::channel();
+                let (release_tx, release_rx) = oneshot::channel();
+                let hold_queue = Arc::clone(&queue);
+                hold_tasks.push(tokio::spawn(async move {
+                    hold_queue
+                        .run_on_lane(lane, "hold operation failed", DiskBlockingQueue::test_stages_for_lane(lane), move |_| {
+                            let _ = started_tx.send(());
+                            let _ = release_rx.blocking_recv();
+                            Ok(())
+                        })
+                        .await
+                }));
+                started_rx.await.expect("hold worker started");
+                hold_txs.push(release_tx);
+            }
            let gate = AdaptiveAdmission::new(NonZeroU16::new(1).expect("one unit"), 1);
            let (first_started_tx, first_started_rx) = oneshot::channel();
            let (first_release_tx, first_release_rx) = oneshot::channel();
@@ -11507,10 +11750,91 @@ mod tests {
            let retained = gate.snapshot().active;
            second_release_tx.send(()).expect("release second producer");
            second.await.expect("second producer completed");
+            for tx in hold_txs {
+                let _ = tx.send(());
+            }
+            for task in hold_tasks {
+                task.await.expect("hold task joined").expect("hold task completed");
+            }
            assert_eq!(retained, 0, "shared worker must not retain a completed producer's unit");
        }
    }

+    #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+    async fn disk_blocking_queue_fsync_lane_scales_workers_and_retires_idle() {
+        use std::sync::atomic::{AtomicUsize, Ordering};
+
+        let queue = DiskBlockingQueue::new();
+        let active = Arc::new(AtomicUsize::new(0));
+        let max_seen = Arc::new(AtomicUsize::new(0));
+        let entered = Arc::new(AtomicUsize::new(0));
+        let (all_entered_tx, all_entered_rx) = oneshot::channel();
+        let all_entered_tx = Arc::new(Mutex::new(Some(all_entered_tx)));
+        let (release_tx, _) = tokio::sync::broadcast::channel::<()>(1);
+
+        let mut handles = Vec::new();
+        for i in 0..(MAX_FSYNC_WORKERS_PER_DISK + 4) {
+            let q = Arc::clone(&queue);
+            let active = Arc::clone(&active);
+            let max_seen = Arc::clone(&max_seen);
+            let entered = Arc::clone(&entered);
+            let all_entered_tx = Arc::clone(&all_entered_tx);
+            let mut release_rx = release_tx.subscribe();
+            handles.push(tokio::spawn(async move {
+                use_current_runtime_for_fsync();
+                q.run_on_lane(
+                    DiskBlockingLane::Fsync,
+                    "fsync scale test failed",
+                    DiskBlockingQueue::test_stages_for_lane(DiskBlockingLane::Fsync),
+                    move |_| {
+                        let cur = active.fetch_add(1, Ordering::SeqCst) + 1;
+                        max_seen.fetch_max(cur, Ordering::SeqCst);
+                        let ord = entered.fetch_add(1, Ordering::SeqCst) + 1;
+                        if ord == MAX_FSYNC_WORKERS_PER_DISK
+                            && let Some(tx) = all_entered_tx.lock().expect("lock completion sender").take()
+                        {
+                            let _ = tx.send(());
+                        }
+                        if ord <= MAX_FSYNC_WORKERS_PER_DISK {
+                            let _ = release_rx.blocking_recv();
+                        }
+                        active.fetch_sub(1, Ordering::SeqCst);
+                        Ok(i)
+                    },
+                )
+                .await
+            }));
+        }
+
+        tokio::time::timeout(std::time::Duration::from_secs(5), all_entered_rx)
+            .await
+            .expect("all fsync workers should enter within timeout")
+            .expect("all_entered channel");
+        let _ = release_tx.send(());
+        for (i, handle) in handles.into_iter().enumerate() {
+            assert_eq!(handle.await.expect("task join").expect("fsync job"), i);
+        }
+        assert_eq!(
+            max_seen.load(Ordering::SeqCst),
+            MAX_FSYNC_WORKERS_PER_DISK,
+            "fsync lane must run MAX_FSYNC_WORKERS_PER_DISK concurrent workers up to its configured bound"
+        );
+        tokio::time::timeout(std::time::Duration::from_secs(5), async {
+            loop {
+                {
+                    let state = queue.lane(DiskBlockingLane::Fsync).state.lock();
+                    if state.active_workers == 0 && state.idle_workers == 0 {
+                        assert!(state.jobs.is_empty());
+                        break;
+                    }
+                }
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("all fsync workers must retire with zero idle count");
+    }
+
    #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
    async fn disk_blocking_queue_bounds_concurrency_and_recovers_from_panic() {
        use std::sync::atomic::{AtomicUsize, Ordering};
@@ -11518,6 +11842,7 @@ mod tests {
        let queue = DiskBlockingQueue::new();
        let active = Arc::new(AtomicUsize::new(0));
        let max_seen = Arc::new(AtomicUsize::new(0));
+        let entered = Arc::new(AtomicUsize::new(0));
        let barrier = Arc::new(std::sync::Barrier::new(MAX_NAMESPACE_WORKERS_PER_DISK + 1));

        let mut handles = Vec::new();
@@ -11525,12 +11850,13 @@ mod tests {
            let q = Arc::clone(&queue);
            let active = Arc::clone(&active);
            let max_seen = Arc::clone(&max_seen);
+            let entered = Arc::clone(&entered);
            let barrier = Arc::clone(&barrier);
            handles.push(tokio::spawn(async move {
                q.run(move || {
                    let cur = active.fetch_add(1, Ordering::SeqCst) + 1;
                    max_seen.fetch_max(cur, Ordering::SeqCst);
-                    if i < MAX_NAMESPACE_WORKERS_PER_DISK {
+                    if entered.fetch_add(1, Ordering::SeqCst) < MAX_NAMESPACE_WORKERS_PER_DISK {
                        barrier.wait();
                    }
                    std::thread::sleep(std::time::Duration::from_millis(5));
@@ -828,13 +828,27 @@ pub async fn create_bitrot_writer(
    length: i64,
    shard_size: usize,
    checksum_algo: HashAlgorithm,
+) -> disk::error::Result<BitrotWriterWrapper> {
+    create_bitrot_writer_with_orig_volume(is_inline_buffer, disk, "", volume, path, length, shard_size, checksum_algo).await
+}
+
+#[allow(clippy::too_many_arguments)]
+pub async fn create_bitrot_writer_with_orig_volume(
+    is_inline_buffer: bool,
+    disk: Option<&DiskStore>,
+    origvolume: &str,
+    volume: &str,
+    path: &str,
+    length: i64,
+    shard_size: usize,
+    checksum_algo: HashAlgorithm,
 ) -> disk::error::Result<BitrotWriterWrapper> {
    let writer = if is_inline_buffer {
        CustomWriter::new_inline_buffer()
    } else if let Some(disk) = disk {
        let length = bitrot_create_file_size(length, shard_size, &checksum_algo);

-        let file = disk.create_file("", volume, path, length).await?;
+        let file = disk.create_file(origvolume, volume, path, length).await?;
        #[cfg(feature = "hotpath")]
        let file = instrument_raw_shard_writer(file, disk.is_local());
        CustomWriter::new_tokio_writer(file)
@@ -333,6 +333,34 @@ struct ProofCache {
    preferred: usize,
 }

+type VerifiedStripeKey = (uuid::Uuid, [u8; 32], u32);
+
+static VERIFIED_STRIPE_RECORDS: std::sync::LazyLock<parking_lot::Mutex<std::collections::HashMap<VerifiedStripeKey, Bytes>>> =
+    std::sync::LazyLock::new(|| parking_lot::Mutex::new(std::collections::HashMap::new()));
+
+fn get_verified_stripe_record(part: &PartIntegrity, stripe: u32) -> Option<Bytes> {
+    VERIFIED_STRIPE_RECORDS
+        .lock()
+        .get(&(part.generation, part.root, stripe))
+        .cloned()
+}
+
+fn put_verified_stripe_record(part: &PartIntegrity, stripe: u32, record: Bytes) {
+    let mut map = VERIFIED_STRIPE_RECORDS.lock();
+    if map.len() >= 4096
+        && let Some(evict_key) = map.keys().next().copied()
+    {
+        map.remove(&evict_key);
+    }
+    map.insert((part.generation, part.root, stripe), record);
+}
+
+pub(crate) fn invalidate_verified_stripe_records(generation: uuid::Uuid) {
+    VERIFIED_STRIPE_RECORDS
+        .lock()
+        .retain(|(stored_gen, _, _), _| *stored_gen != generation);
+}
+
 impl PartProofReader {
    pub(crate) async fn reconstruction_proof(
        &self,
@@ -410,6 +438,9 @@ impl PartProofReader {

    pub(crate) async fn record(&self, stripe: u32) -> io::Result<Bytes> {
        let offset = self.part.record_offset(stripe).map_err(|_| corrupt())?;
+        if let Some(bytes) = get_verified_stripe_record(&self.part, stripe) {
+            return Ok(bytes);
+        }
        let mut cached = self.cached.lock().await;
        if let Some((_, bytes)) = cached.records.iter().find(|(number, _)| *number == stripe) {
            return Ok(bytes.clone());
@@ -465,6 +496,9 @@ impl PartProofReader {
            bytes
        };
        self.part.verify_record(stripe, &record).map_err(|_| corrupt())?;
+        if self.part.stripes <= 4 {
+            put_verified_stripe_record(&self.part, stripe, record.clone());
+        }
        if cached.records.len() == 2 {
            cached.records.pop_front();
        }
@@ -863,4 +897,41 @@ mod tests {
            prepared.part.verify_shard(stripe, 15, b"x", &record).expect("spilled proof");
        }
    }
+
+    #[tokio::test]
+    async fn verified_stripe_record_cache_serves_verified_record_and_still_verifies_payload() {
+        let layout = IntegrityLayout::new(2, 2, 8, false).expect("layout");
+        let mut builder = IntegrityBuilder::new(layout, 1).expect("builder");
+        let shards = [b"ABCD".as_slice(), b"EFGH", b"IJKL", b"MNOP"];
+        builder.push(shards.into_iter()).await.expect("stripe");
+        let prepared = builder.finish(8).await.expect("index");
+        let inline_bytes = prepared.inline_bytes().expect("inline bytes");
+        let part = prepared.part;
+
+        let warm_reader = PartProofReader {
+            part: part.clone(),
+            sources: Vec::new(),
+            volume: "bucket".to_owned(),
+            inline: Some(inline_bytes),
+            cached: Mutex::new(ProofCache::default()),
+        };
+        warm_reader.verify(0, 0, shards[0]).await.expect("warm verify");
+        assert!(
+            get_verified_stripe_record(&part, 0).is_some(),
+            "small verified stripe record should populate the verified stripe cache"
+        );
+
+        let cached_reader = Arc::new(PartProofReader {
+            part,
+            sources: Vec::new(),
+            volume: "bucket".to_owned(),
+            inline: None,
+            cached: Mutex::new(ProofCache::default()),
+        });
+        cached_reader.verify(0, 0, shards[0]).await.expect("cached verify");
+        assert!(
+            cached_reader.verify(0, 0, shards[1]).await.is_err(),
+            "cached record must still reject corrupt or mismatched shard payloads"
+        );
+    }
 }
@@ -1543,6 +1543,8 @@ pub struct ObjectInfo {
    /// published by the same upload. Callers must not repeat quota, replication,
    /// lifecycle, or object-created side effects for that response.
    pub multipart_completion_replayed: bool,
+    /// Previous current version's size observed during `rename_data` commit.
+    pub backfilled_old_current_size: Option<crate::disk::OldCurrentSize>,
 }

 impl Default for ObjectInfo {
@@ -1585,6 +1587,7 @@ impl Default for ObjectInfo {
            replication_decision: String::new(),
            checksum: None,
            multipart_completion_replayed: false,
+            backfilled_old_current_size: None,
        }
    }
 }
@@ -1628,6 +1631,7 @@ impl Clone for ObjectInfo {
            replication_decision: self.replication_decision.clone(),
            checksum: self.checksum.clone(),
            multipart_completion_replayed: self.multipart_completion_replayed,
+            backfilled_old_current_size: self.backfilled_old_current_size,
            expires: self.expires,
        }
    }
@@ -941,6 +941,32 @@ impl InstanceContext {
            .clone()
    }

+    /// Remove the `BucketMetadataSys` when its backing `ECStore` is gone.
+    ///
+    /// Called from `ECStore::drop`. When test fixtures share the bootstrap
+    /// context, the `BucketMetadataSys` stores a `Weak<ECStore>` to prevent a
+    /// reference cycle; once the store drops the weak is dangling and any
+    /// `object_store()` call panics. Clearing the slot here lets the next
+    /// fixture install a fresh `BucketMetadataSys` without hitting the stale one.
+    pub fn clear_bucket_metadata_sys_if_stale(&self) {
+        let mut guard = self.bucket_metadata_sys.lock().expect("bucket_metadata_sys lock poisoned");
+        if let Some(sys) = guard.as_ref() {
+            // `read()` may block briefly; the async RwLock is acceptable here
+            // because Drop is sync and the lock is held only long enough to
+            // check the weak reference.
+            let is_stale = {
+                let s = sys.try_read();
+                match s {
+                    Ok(s) => s.is_store_stale(),
+                    Err(_) => true, // contended: treat as stale-safe to clear
+                }
+            };
+            if is_stale {
+                *guard = None;
+            }
+        }
+    }
+
    /// Set this instance's background-services cancellation token (once).
    pub fn init_background_cancel_token(&self, token: CancellationToken) -> Result<(), CancellationToken> {
        self.background_cancel_token.set(token)
@@ -5340,6 +5340,14 @@ impl SetDisks {
        }

        let disks = self.get_disks_internal().await;
+        if disks.len() == self.set_drive_count
+            && disks.iter().all(|disk| {
+                disk.as_ref()
+                    .is_some_and(|disk| disk.has_clean_active_multipart_source(dst_object, true))
+            })
+        {
+            return Ok(false);
+        }
        let transaction_path = part_transaction_path(dst_object);
        let transaction_meta_path = format!("{transaction_path}/{PART_TRANSACTION_NEW_META}");

@@ -5539,6 +5547,14 @@ impl SetDisks {
        write_quorum: usize,
    ) -> disk::error::Result<()> {
        let disks = self.get_disks_internal().await;
+        if disks.len() == self.set_drive_count
+            && disks.iter().all(|disk| {
+                disk.as_ref()
+                    .is_some_and(|disk| disk.has_clean_active_multipart_source(part_path, false))
+            })
+        {
+            return Ok(());
+        }
        let listings = join_all(disks.iter().map(|disk| {
            let disk = disk.clone();
            async move {
@@ -5598,10 +5614,11 @@ impl SetDisks {
        write_quorum: usize,
        quorum_context: Option<MultipartWriteQuorumContext<'_>>,
    ) -> disk::error::Result<Vec<Option<DiskStore>>> {
-        self.recover_part_transaction(dst_object, write_quorum).await?;
        let part = ObjectPartInfo::unmarshal(&meta)?;
        let integrity = part.integrity.map(Arc::new);

+        self.recover_part_transaction(dst_object, write_quorum).await?;
+
        let src_bucket = Arc::new(src_bucket.to_string());
        let src_object = Arc::new(src_object.to_string());
        let dst_bucket = Arc::new(dst_bucket.to_string());
@@ -5623,7 +5640,9 @@ impl SetDisks {
                if let Err(error) = prepared {
                    return Some(Err(error));
                }
-                if let Some(integrity) = integrity {
+                if let Some(integrity) = integrity
+                    && !disk.is_local()
+                {
                    let Some((directory, _)) = dst_object.rsplit_once('/') else {
                        return Some(Err(DiskError::FileCorrupt));
                    };
@@ -5740,7 +5759,36 @@ impl SetDisks {
            return Err(err);
        }

-        let committed = self.recover_part_transaction(&dst_object, write_quorum).await?;
+        let all_prepared_succeeded = errs
+            .iter()
+            .zip(prepared_disks.iter())
+            .all(|(err, disk)| err.is_none() && disk.as_ref().is_some_and(|d| d.is_local()));
+        let committed = if all_prepared_succeeded {
+            let settles = prepared_disks.iter().map(|disk| {
+                let disk = disk.clone();
+                let dst_object = dst_object.clone();
+                async move {
+                    let Some(disk) = disk else {
+                        return Err(DiskError::DiskNotFound);
+                    };
+                    disk.settle_part_transaction(RUSTFS_META_MULTIPART_BUCKET, &dst_object, PartTransactionAction::Commit)
+                        .await
+                }
+            });
+            let settle_results = join_all(settles).await;
+            if settle_results.iter().filter(|res| res.is_ok()).count() >= write_quorum
+                && settle_results
+                    .iter()
+                    .zip(prepared_disks.iter())
+                    .all(|(res, disk)| disk.is_none() || res.is_ok())
+            {
+                true
+            } else {
+                self.recover_part_transaction(&dst_object, write_quorum).await?
+            }
+        } else {
+            self.recover_part_transaction(&dst_object, write_quorum).await?
+        };
        if !committed {
            let err = DiskError::ErasureWriteQuorum;
            if let Some(context) = quorum_context {
@@ -6737,9 +6785,7 @@ impl SetDisks {
                        object,
                    );
                }
-                let if_none_match = http_preconditions.if_none_match_value().map(str::to_owned);
-                let if_match = http_preconditions.if_match_value().map(str::to_owned);
-                if should_prevent_write(&oi, if_none_match, if_match)
+                if should_prevent_write(&oi, http_preconditions.if_none_match_value(), http_preconditions.if_match_value())
                    && !crate::data_movement::can_replace_stale_data_movement_target(&oi, opts)
                {
                    return Some(StorageError::PreconditionFailed);
@@ -161,7 +161,7 @@ impl SetDisks {
            )
        });
        let mut metadata = FileMeta::new();
-        for parts in versions.into_values() {
+        for (version_id, parts) in versions {
            let errors = parts
                .iter()
                .zip(&missing)
@@ -173,15 +173,25 @@ impl SetDisks {
                    }
                })
                .collect::<Vec<_>>();
-            let read_quorum = match Self::object_quorum_from_meta(&parts, &errors, default_parity) {
-                Ok((quorum, _)) => usize::try_from(quorum).map_err(|_| DiskError::FileCorrupt)?,
+            let (read_quorum, write_quorum) = match Self::object_quorum_from_meta(&parts, &errors, default_parity) {
+                Ok((read, write)) => (
+                    usize::try_from(read).map_err(|_| DiskError::FileCorrupt)?,
+                    usize::try_from(write).map_err(|_| DiskError::FileCorrupt)?,
+                ),
                Err(DiskError::FileNotFound | DiskError::FileVersionNotFound) if complete => continue,
                Err(DiskError::FileNotFound | DiskError::FileVersionNotFound) => return Err(DiskError::ErasureReadQuorum),
                Err(DiskError::ErasureReadQuorum) if complete => continue,
                Err(error) => return Err(error),
            };
-            let mod_time = Self::common_time(&Self::list_object_modtimes(&parts, &errors), read_quorum);
-            let selected = Self::pick_valid_fileinfo(&parts, mod_time, None, read_quorum)?;
+            let selected = if crate::bucket::utils::is_meta_bucketname(bucket) && version_id.is_none() {
+                // A null-version system record must carry the same commit proof
+                // as GET. Listing only needs the selected metadata, so no disk
+                // handles are retained from the shared read selector.
+                Self::select_committed_fileinfo(&vec![None; disk_count], &parts, &errors, "", read_quorum, write_quorum)?.1
+            } else {
+                let mod_time = Self::common_time(&Self::list_object_modtimes(&parts, &errors), read_quorum);
+                Self::pick_valid_fileinfo(&parts, mod_time, None, read_quorum)?
+            };
            metadata.add_version(selected)?;
        }
        if metadata.versions.is_empty() {
@@ -80,7 +80,7 @@ use crate::erasure::codec::bridge::{
 use crate::erasure::coding;
 use crate::error::{Error, Result, is_err_version_not_found};
 use crate::error::{GenericError, ObjectApiError, is_err_object_not_found};
-use crate::io_support::bitrot::{create_bitrot_reader_from_bytes, create_bitrot_writer};
+use crate::io_support::bitrot::{create_bitrot_reader_from_bytes, create_bitrot_writer, create_bitrot_writer_with_orig_volume};
 use crate::object_api::ObjectOptions;
 use crate::object_api::get_object_body_cache_hook;
 use crate::object_api::object_api_utils::get_raw_etag;
@@ -120,7 +120,6 @@ use bytesize::ByteSize;
 use futures::future::join_all;
 use http::HeaderMap;
 use md5::{Digest as Md5Digest, Md5};
-use regex::Regex;
 use rustfs_config::MI_B;
 use rustfs_filemeta::metadata_keys;
 use rustfs_filemeta::{
@@ -8036,16 +8035,29 @@ fn parts_after_marker(part_numbers: &[usize], part_number_marker: usize) -> Opti
    Some(&part_numbers[index..])
 }

-pub fn canonicalize_etag(etag: &str) -> String {
-    let re = Regex::new("\"*?([^\"]*?)\"*?$").unwrap();
-    re.replace_all(etag, "$1").to_string()
+fn canonicalized_etag(etag: &str) -> Cow<'_, str> {
+    // Preserve the legacy suffix-matching behavior for embedded quotes.
+    let trimmed = etag.trim_end_matches('"');
+    let Some(last_quote) = trimmed.rfind('"') else {
+        return Cow::Borrowed(trimmed);
+    };
+    let prefix = trimmed[..last_quote].trim_end_matches('"');
+    let suffix = &trimmed[last_quote + 1..];
+    if prefix.is_empty() {
+        Cow::Borrowed(suffix)
+    } else {
+        let mut result = String::with_capacity(prefix.len() + suffix.len());
+        result.push_str(prefix);
+        result.push_str(suffix);
+        Cow::Owned(result)
+    }
 }

 pub fn e_tag_matches(etag: &str, condition: &str) -> bool {
    if condition.trim() == "*" {
        return true;
    }
-    canonicalize_etag(etag) == canonicalize_etag(condition)
+    canonicalized_etag(etag) == canonicalized_etag(condition)
 }

 fn weak_etag_matches(etag: &str, condition: &str) -> bool {
@@ -8060,15 +8072,12 @@ fn weak_etag_matches(etag: &str, condition: &str) -> bool {
            .or_else(|| trimmed.strip_prefix("w/"))
            .unwrap_or(trimmed)
    }
-    canonicalize_etag(strip_weak(etag)) == canonicalize_etag(strip_weak(condition))
+    canonicalized_etag(strip_weak(etag)) == canonicalized_etag(strip_weak(condition))
 }

-pub fn should_prevent_write(oi: &ObjectInfo, if_none_match: Option<String>, if_match: Option<String>) -> bool {
-    let if_none_match = if_none_match
-        .as_deref()
-        .map(str::trim)
-        .filter(|condition| !condition.is_empty());
-    let if_match = if_match.as_deref().map(str::trim).filter(|condition| !condition.is_empty());
+pub fn should_prevent_write(oi: &ObjectInfo, if_none_match: Option<&str>, if_match: Option<&str>) -> bool {
+    let if_none_match = if_none_match.map(str::trim).filter(|condition| !condition.is_empty());
+    let if_match = if_match.map(str::trim).filter(|condition| !condition.is_empty());

    match &oi.etag {
        Some(etag) => {
@@ -13772,6 +13781,22 @@ mod tests {
        assert!(result2.iter().all(|d| d.is_none()));
    }

+    #[test]
+    fn canonicalized_etag_preserves_legacy_regex_semantics() {
+        let regex = regex::Regex::new("\"*?([^\"]*?)\"*?$").expect("valid legacy ETag pattern");
+        let alphabet = ['a', '"', '\n', 'é', '\0'];
+        for length in 0..=6_u32 {
+            for mut number in 0..alphabet.len().pow(length) {
+                let mut input = String::new();
+                for _ in 0..length {
+                    input.push(alphabet[number % alphabet.len()]);
+                    number /= alphabet.len();
+                }
+                assert_eq!(canonicalized_etag(&input), regex.replace_all(&input, "$1"), "input: {input:?}");
+            }
+        }
+    }
+
    #[test]
    fn test_etag_matches() {
        assert!(e_tag_matches("abc", "abc"));
@@ -14347,61 +14372,61 @@ mod tests {
            etag: Some("abc".to_string()),
            ..Default::default()
        };
-        let if_none_match = Some("abc".to_string());
+        let if_none_match = Some("abc");
        let if_match = None;
        assert!(should_prevent_write(&oi, if_none_match, if_match));

-        let if_none_match = Some("*".to_string());
+        let if_none_match = Some("*");
        let if_match = None;
        assert!(should_prevent_write(&oi, if_none_match, if_match));

        let if_none_match = None;
-        let if_match = Some("def".to_string());
+        let if_match = Some("def");
        assert!(should_prevent_write(&oi, if_none_match, if_match));

        let if_none_match = None;
-        let if_match = Some("*".to_string());
+        let if_match = Some("*");
        assert!(!should_prevent_write(&oi, if_none_match, if_match));

-        let if_none_match = Some("def".to_string());
+        let if_none_match = Some("def");
        let if_match = None;
        assert!(!should_prevent_write(&oi, if_none_match, if_match));

-        let if_none_match = Some("def".to_string());
-        let if_match = Some("*".to_string());
+        let if_none_match = Some("def");
+        let if_match = Some("*");
        assert!(!should_prevent_write(&oi, if_none_match, if_match));

-        let if_none_match = Some("def".to_string());
-        let if_match = Some("\"abc\"".to_string());
+        let if_none_match = Some("def");
+        let if_match = Some("\"abc\"");
        assert!(!should_prevent_write(&oi, if_none_match, if_match));

-        let if_none_match = Some("*".to_string());
-        let if_match = Some("\"abc\"".to_string());
+        let if_none_match = Some("*");
+        let if_match = Some("\"abc\"");
        assert!(should_prevent_write(&oi, if_none_match, if_match));

        let oi = ObjectInfo {
            etag: None,
            ..Default::default()
        };
-        let if_none_match = Some("*".to_string());
-        let if_match = Some("\"abc\"".to_string());
+        let if_none_match = Some("*");
+        let if_match = Some("\"abc\"");
        assert!(should_prevent_write(&oi, if_none_match, if_match));

        let if_none_match = None;
        let if_match = None;
        assert!(!should_prevent_write(&oi, if_none_match, if_match));

-        let if_none_match = Some(String::new());
-        let if_match = Some(" ".to_string());
+        let if_none_match = Some("");
+        let if_match = Some(" ");
        assert!(!should_prevent_write(&oi, if_none_match, if_match));

        let oi = ObjectInfo {
            etag: Some("abc".to_string()),
            ..Default::default()
        };
-        assert!(should_prevent_write(&oi, Some("W/\"abc\"".to_string()), None));
-        assert!(!should_prevent_write(&oi, Some("W/\"def\"".to_string()), None));
-        assert!(should_prevent_write(&oi, None, Some("W/\"abc\"".to_string())));
+        assert!(should_prevent_write(&oi, Some("W/\"abc\""), None));
+        assert!(!should_prevent_write(&oi, Some("W/\"def\""), None));
+        assert!(should_prevent_write(&oi, None, Some("W/\"abc\"")));
    }

    #[test]
@@ -41,8 +41,8 @@ use super::super::{
    SUFFIX_REPLICATION_SSEC_CRC, SUFFIX_RESTORE_OPERATION_ID, SUFFIX_RESTORE_WORKER_LOCK, SetDisks, SmallWritePath, StorageError,
    Uuid, WriteLayout, check_object_lock_for_deletion_with_state, classify_multipart_part_write_path, coding,
    complete_multipart_part_error, complete_multipart_part_error_result, complete_part_checksum, contains_key_str,
-    create_bitrot_writer, debug, disk, error, file_info_is_valid_for_metadata, get_complete_multipart_md5, get_header_map,
-    get_str, insert_str, is_err_object_not_found, is_err_version_not_found, is_min_allowed_part_size,
+    create_bitrot_writer_with_orig_volume, debug, disk, error, file_info_is_valid_for_metadata, get_complete_multipart_md5,
+    get_header_map, get_str, insert_str, is_err_object_not_found, is_err_version_not_found, is_min_allowed_part_size,
    log_multipart_write_quorum_failure, multipart_part_logical_size, parts_after_marker, path_join_buf,
    prepare_complete_multipart_parts, record_compression_total_memory, record_multipart_write_quorum_failure,
    reduce_read_quorum_errs, reduce_write_quorum_errs, remove_header_map, resolve_write_layout,
@@ -124,9 +124,10 @@ async fn create_part_writers(
        let Some(disk) = disk else {
            return (None, Some(DiskError::DiskNotFound));
        };
-        match create_bitrot_writer(
+        match create_bitrot_writer_with_orig_volume(
            false,
            Some(disk),
+            RUSTFS_META_MULTIPART_BUCKET,
            RUSTFS_META_TMP_BUCKET,
            path,
            length,
@@ -977,7 +978,7 @@ async fn multipart_upload_paths_on_disk(disk: DiskStore, bucket: &str, root_pref
        return Err(DiskError::DiskNotFound);
    }

-    if !root_prefix.is_empty() {
+    if !root_prefix.is_empty() && root_prefix != DATA_MOVEMENT_MULTIPART_PREFIX {
        let upload_dirs = match disk.list_dir(bucket, RUSTFS_META_MULTIPART_BUCKET, root_prefix, -1).await {
            Ok(entries) => entries,
            Err(DiskError::FileNotFound | DiskError::VolumeNotFound) => return Ok(Vec::new()),
@@ -1002,6 +1003,13 @@ async fn multipart_upload_paths_on_disk(disk: DiskStore, bucket: &str, root_pref
        .into_iter()
        .map(|sha_dir| sha_dir.trim_end_matches('/').to_owned())
        .filter(|sha_dir| sha_dir.len() == 64 && sha_dir.bytes().all(|byte| byte.is_ascii_hexdigit()))
+        .map(|sha_dir| {
+            if root_prefix.is_empty() {
+                sha_dir
+            } else {
+                format!("{root_prefix}/{sha_dir}")
+            }
+        })
        .collect::<Vec<_>>();
    let listed_upload_dirs = stream::iter(sha_dirs)
        .map(|sha_dir| {
@@ -1124,6 +1132,7 @@ impl SetDisks {
        orig_bucket: &str,
        error_path: &str,
        root_prefix: &str,
+        require_all_disks: bool,
    ) -> Result<(Vec<Option<DiskStore>>, Vec<String>, usize, bool)> {
        let disks = self.disks.read().await.clone();
        if disks.is_empty() {
@@ -1165,7 +1174,8 @@ impl SetDisks {
            }
        }

-        if let Some(err) = reduce_read_quorum_errs(&discovery_errors, OBJECT_OP_IGNORED_ERRS, discovery_quorum) {
+        let required_discoveries = if require_all_disks { disks.len() } else { discovery_quorum };
+        if let Some(err) = reduce_read_quorum_errs(&discovery_errors, OBJECT_OP_IGNORED_ERRS, required_discoveries) {
            return Err(to_object_err(err.into(), vec![orig_bucket, error_path]));
        }

@@ -1187,7 +1197,7 @@ impl SetDisks {

    pub(crate) async fn first_multipart_upload_path_for_decommission(&self, bucket: &str) -> Result<Option<String>> {
        let (_, paths, _, _) = self
-            .discover_multipart_upload_paths(bucket, RUSTFS_META_MULTIPART_BUCKET, "")
+            .discover_multipart_upload_paths(bucket, RUSTFS_META_MULTIPART_BUCKET, "", false)
            .await?;
        Ok(paths.into_iter().next())
    }
@@ -1200,8 +1210,9 @@ impl SetDisks {
        upload_identity: &str,
    ) -> Result<Vec<String>> {
        let expected_parent = format!("{DATA_MOVEMENT_MULTIPART_PREFIX}/{}", Self::get_multipart_sha_dir(bucket, object));
-        let (_, candidate_paths, _, has_minority_candidate) =
-            self.discover_multipart_upload_paths(bucket, object, &expected_parent).await?;
+        let (_, candidate_paths, _, has_minority_candidate) = self
+            .discover_multipart_upload_paths(bucket, object, &expected_parent, false)
+            .await?;
        if has_minority_candidate {
            return Err(Error::DecommissionCapacityBlocked {
                message: "data movement multipart cleanup found an upload path on fewer than the discovery quorum of disks"
@@ -1750,6 +1761,54 @@ impl SetDisks {
        })
    }

+    pub(crate) async fn ensure_no_multipart_uploads(&self, bucket: &str) -> Result<()> {
+        // Every disk must confirm absence in both namespaces before cleanup can
+        // remove catalog protection from the bucket.
+        for root_prefix in ["", DATA_MOVEMENT_MULTIPART_PREFIX] {
+            let (disks, paths, _, has_minority_candidate) =
+                self.discover_multipart_upload_paths(bucket, "", root_prefix, true).await?;
+            if has_minority_candidate {
+                return Err(Error::ErasureReadQuorum);
+            }
+            for upload_path in paths {
+                let (sha_dir, _) = upload_path
+                    .rsplit_once('/')
+                    .filter(|(directory, upload_id)| !directory.is_empty() && !upload_id.is_empty())
+                    .ok_or(DiskError::CorruptedFormat)?;
+                let (metadata, errors) =
+                    Self::read_all_fileinfo(&disks, bucket, RUSTFS_META_MULTIPART_BUCKET, &upload_path, "", false, false, false)
+                        .await?;
+                let (read_quorum, _) = Self::object_quorum_from_meta(&metadata, &errors, self.default_parity_count)?;
+                let read_quorum = usize::try_from(read_quorum).map_err(|_| DiskError::ErasureReadQuorum)?;
+                if let Some(error) = reduce_read_quorum_errs(&errors, OBJECT_OP_IGNORED_ERRS, read_quorum) {
+                    return Err(to_object_err(error.into(), vec![bucket, &upload_path]));
+                }
+                let (_, modified, etag) = Self::list_online_disks(&disks, &metadata, &errors, read_quorum);
+                let file = Self::pick_valid_fileinfo(&metadata, modified, etag, read_quorum)?;
+                let (Some(owner_bucket), Some(owner_object)) = (
+                    file.metadata.get(RUSTFS_MULTIPART_BUCKET_KEY),
+                    file.metadata.get(RUSTFS_MULTIPART_OBJECT_KEY),
+                ) else {
+                    return Err(to_object_err(DiskError::CorruptedFormat.into(), vec![bucket, &upload_path]));
+                };
+                if owner_bucket.is_empty()
+                    || owner_object.is_empty()
+                    || Self::get_multipart_sha_dir(owner_bucket, owner_object)
+                        != sha_dir
+                            .strip_prefix(DATA_MOVEMENT_MULTIPART_PREFIX)
+                            .and_then(|suffix| suffix.strip_prefix('/'))
+                            .unwrap_or(sha_dir)
+                {
+                    return Err(to_object_err(DiskError::CorruptedFormat.into(), vec![bucket, &upload_path]));
+                }
+                if owner_bucket == bucket {
+                    return Err(Error::BucketNotEmpty(bucket.to_string()));
+                }
+            }
+        }
+        Ok(())
+    }
+
    #[allow(clippy::too_many_arguments)]
    pub(crate) async fn list_multipart_uploads_for_incarnation(
        &self,
@@ -1761,7 +1820,8 @@ impl SetDisks {
        max_uploads: usize,
        expected_incarnation_id: Option<Uuid>,
    ) -> Result<ListMultipartsInfo> {
-        let (disks, candidate_paths, discovery_quorum, _) = self.discover_multipart_upload_paths(bucket, prefix, "").await?;
+        let (disks, candidate_paths, discovery_quorum, _) =
+            self.discover_multipart_upload_paths(bucket, prefix, "", false).await?;
        let listed_uploads = stream::iter(candidate_paths)
            .map(|upload_path| {
                let disks = &disks;
@@ -2436,7 +2496,16 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
            }

            let part_integrity = if let Some(integrity) = integrity {
-                Some(integrity.write(&mut shuffle_disks, bucket, RUSTFS_META_TMP_BUCKET, &tmp_part).await?)
+                Some(
+                    integrity
+                        .write(
+                            &mut shuffle_disks,
+                            RUSTFS_META_MULTIPART_BUCKET,
+                            RUSTFS_META_TMP_BUCKET,
+                            &tmp_part,
+                        )
+                        .await?,
+                )
            } else { None };
            if shuffle_disks.iter().filter(|disk| disk.is_some()).count() < write_quorum {
                return Err(Error::ErasureWriteQuorum);
@@ -3471,10 +3540,6 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
        bind_multipart_proof_snapshot(&disks, &self.multipart_proof_verification_context())?;

        let part_path = format!("{}/{}/", upload_id_path, fi.data_dir.unwrap_or(Uuid::nil()));
-        self.recover_part_transactions(&part_path, read_quorum, write_quorum)
-            .await
-            .map_err(|err| to_object_err(err.into(), vec![bucket, object]))?;
-
        let part_meta_paths = uploaded_parts
            .iter()
            .map(|v| format!("{part_path}part.{0}.meta", v.part_num))
@@ -3482,6 +3547,10 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {

        let part_numbers = uploaded_parts.iter().map(|v| v.part_num).collect::<Vec<usize>>();

+        self.recover_part_transactions(&part_path, read_quorum, write_quorum)
+            .await
+            .map_err(|err| to_object_err(err.into(), vec![bucket, object]))?;
+
        let object_parts = Self::read_parts(&disks, RUSTFS_META_MULTIPART_BUCKET, &part_meta_paths, &part_numbers, read_quorum)
            .await
            .map_err(|err| to_object_err(err.into(), vec![bucket, object]))?;
@@ -4389,6 +4458,7 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
            let needs_immediate_heal = rename_commit.needs_immediate_heal();
            let op_old_dir = rename_commit.data_dir;
            let cleanup_disks = rename_commit.cleanup_disks;
+            let old_current_size = rename_commit.old_current_size;
            let committed_file_info = rename_commit.committed_file_info;

            if needs_immediate_heal {
@@ -4457,7 +4527,13 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
            // parts; deleting them before the commit would strand the upload
            // permanently. This mirrors the "clean up only after commit" pattern
            // already used for the old data-dir GC and the upload-dir delete_all below.
-            commit_set.cleanup_multipart_path(&parts).await;
+            let can_skip_part_cleanup = !needs_upgrade
+                && commit_disks
+                    .iter()
+                    .all(|disk| disk.as_ref().is_some_and(|disk| disk.is_local()));
+            if !can_skip_part_cleanup {
+                commit_set.cleanup_multipart_path(&parts).await;
+            }

            if let Some(old_dir) = op_old_dir {
                // backlog#898: best-effort reclaim of the dereferenced old data dir.
@@ -4503,7 +4579,9 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {

            drop(_upload_guard.take());

-            Ok(ObjectInfo::from_file_info(&fi, &commit_bucket, &commit_object, commit_is_versioned))
+            let mut obj_info = ObjectInfo::from_file_info(&fi, &commit_bucket, &commit_object, commit_is_versioned);
+            obj_info.backfilled_old_current_size = old_current_size;
+            Ok(obj_info)
        };

        let result = if detach_commit_owner {
@@ -11194,6 +11272,163 @@ mod tests {
        assert_eq!(upload_uuid_suffix(&recovered[0]), upload_uuid_suffix(&upload.upload_id));
    }

+    #[tokio::test]
+    async fn multipart_absence_inspection_distinguishes_bucket_ownership() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "multipart-cleanup-target";
+        let other = "multipart-cleanup-other";
+        for disk in &disk_stores {
+            disk.make_volume(bucket).await.expect("create target bucket");
+            disk.make_volume(other).await.expect("create other bucket");
+        }
+        set_disks
+            .ensure_no_multipart_uploads(bucket)
+            .await
+            .expect("an empty upload namespace permits cleanup");
+        set_disks
+            .new_multipart_upload(other, "object.bin", &ObjectOptions::default())
+            .await
+            .expect("create upload in other bucket");
+        set_disks
+            .ensure_no_multipart_uploads(bucket)
+            .await
+            .expect("an upload owned by another bucket permits cleanup");
+        set_disks
+            .new_multipart_upload(bucket, "object.bin", &ObjectOptions::default())
+            .await
+            .expect("create upload in target bucket");
+        assert!(
+            matches!(set_disks.ensure_no_multipart_uploads(bucket).await, Err(Error::BucketNotEmpty(name)) if name == bucket)
+        );
+    }
+
+    #[tokio::test]
+    async fn multipart_absence_inspection_requires_every_disk_online() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "multipart-cleanup-offline";
+        for disk in &disk_stores {
+            disk.make_volume(bucket).await.expect("create bucket");
+        }
+        set_disks.disks.write().await[3] = None;
+        let listed = set_disks
+            .list_multipart_uploads_for_incarnation(bucket, "", None, None, None, 1000, None)
+            .await
+            .expect("ordinary listing preserves its discovery quorum");
+        assert!(listed.uploads.is_empty());
+        let error = set_disks
+            .ensure_no_multipart_uploads(bucket)
+            .await
+            .expect_err("an offline disk must prevent confirmation of an empty multipart namespace");
+        assert!(
+            matches!(&error, Error::InsufficientReadQuorum(error_bucket, error_path) if error_bucket == bucket && error_path.is_empty()),
+            "the offline disk must produce a bucket-scoped read quorum error: {error:?}"
+        );
+
+        set_disks.disks.write().await[3] = Some(disk_stores[3].clone());
+        set_disks
+            .ensure_no_multipart_uploads(bucket)
+            .await
+            .expect("all disks can confirm the empty namespace after recovery");
+    }
+
+    #[tokio::test]
+    async fn multipart_absence_inspection_rejects_a_failed_disk_enumeration() {
+        let (temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "multipart-cleanup-list-error";
+        for disk in &disk_stores {
+            disk.make_volume(bucket).await.expect("create bucket");
+        }
+        let sha_dir = SetDisks::get_multipart_sha_dir(bucket, "object.bin");
+        let multipart_dir = temp_dirs[0].path().join(RUSTFS_META_MULTIPART_BUCKET);
+        tokio::fs::create_dir_all(&multipart_dir)
+            .await
+            .expect("create multipart staging directory");
+        tokio::fs::write(multipart_dir.join(&sha_dir), b"not a directory")
+            .await
+            .expect("replace one upload parent directory with a file");
+        multipart_upload_paths_on_disk(disk_stores[0].clone(), bucket, "")
+            .await
+            .expect_err("the affected disk must fail directory enumeration");
+        let listed = set_disks
+            .list_multipart_uploads_for_incarnation(bucket, "", None, None, None, 1000, None)
+            .await
+            .expect("ordinary listing still accepts the remaining discovery quorum");
+        assert!(listed.uploads.is_empty());
+        let error = set_disks
+            .ensure_no_multipart_uploads(bucket)
+            .await
+            .expect_err("a failed disk enumeration must prevent confirmation of an empty multipart namespace");
+        assert!(
+            matches!(&error, Error::InsufficientReadQuorum(error_bucket, error_path) if error_bucket == bucket && error_path.is_empty()),
+            "the failed enumeration must produce a bucket-scoped read quorum error: {error:?}"
+        );
+
+        tokio::fs::remove_file(temp_dirs[0].path().join(RUSTFS_META_MULTIPART_BUCKET).join(sha_dir))
+            .await
+            .expect("remove the invalid directory entry");
+        set_disks
+            .ensure_no_multipart_uploads(bucket)
+            .await
+            .expect("all disks can confirm the empty namespace after repair");
+    }
+
+    #[tokio::test]
+    async fn multipart_absence_inspection_rejects_data_movement_uploads() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "multipart-cleanup-data-movement";
+        for disk in &disk_stores {
+            disk.make_volume(bucket).await.expect("create bucket");
+        }
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_DATA_MOVEMENT_UPLOAD,
+            format!("v1:{}:{}", Uuid::new_v4(), OffsetDateTime::now_utc().unix_timestamp_nanos()),
+        );
+        set_disks
+            .new_multipart_upload(
+                bucket,
+                "object.bin",
+                &ObjectOptions {
+                    data_movement: true,
+                    user_defined: metadata,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("create data movement upload");
+        assert!(
+            matches!(set_disks.ensure_no_multipart_uploads(bucket).await, Err(Error::BucketNotEmpty(name)) if name == bucket)
+        );
+    }
+
+    #[tokio::test]
+    async fn multipart_absence_inspection_rejects_unreadable_ownership() {
+        let (temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "multipart-cleanup-corrupt";
+        for disk in &disk_stores {
+            disk.make_volume(bucket).await.expect("create bucket");
+        }
+        set_disks
+            .new_multipart_upload(bucket, "object.bin", &ObjectOptions::default())
+            .await
+            .expect("create multipart upload");
+        let mut corrupted = 0;
+        for temp_dir in &temp_dirs {
+            for path in multipart_meta_files_on_disk(temp_dir, "xl.meta").await {
+                tokio::fs::write(path, b"invalid upload metadata")
+                    .await
+                    .expect("corrupt upload metadata");
+                corrupted += 1;
+            }
+        }
+        assert!(corrupted > 0, "upload metadata must exist before corruption");
+        set_disks
+            .ensure_no_multipart_uploads(bucket)
+            .await
+            .expect_err("unreadable upload ownership must block cleanup");
+    }
+
    /// Regression (issue #5716): a single upload directory whose `xl.meta` was
    /// destroyed (crash mid-write, torn disk state) must degrade to that upload
    /// alone. Failing the whole ListMultipartUploads turns one piece of stale
@@ -46,9 +46,9 @@ use super::super::{
    build_get_object_info, build_inline_bitrot_readers, build_inline_bitrot_readers_from_refs, can_try_inline_data_shards_direct,
    check_object_lock_delete, check_object_lock_for_deletion_with_state, check_object_lock_retention_update,
    classify_get_codec_streaming_object_class, classify_put_write_path, classify_storage_error,
-    collect_inline_data_shard_fileinfos_by_index, contains_key_str, create_bitrot_writer, debug, delete_file_info_version_id,
-    disk, ensure_delete_commit_locks_held, error, explicit_delete_removed_marker, finish_set_disk_read_lock,
-    get_codec_streaming_reader_gate_with_plan, get_object_body_cache_hook, get_raw_etag,
+    collect_inline_data_shard_fileinfos_by_index, contains_key_str, create_bitrot_writer_with_orig_volume, debug,
+    delete_file_info_version_id, disk, ensure_delete_commit_locks_held, error, explicit_delete_removed_marker,
+    finish_set_disk_read_lock, get_codec_streaming_reader_gate_with_plan, get_object_body_cache_hook, get_raw_etag,
    get_small_object_direct_memory_decision_with_threshold_and_plan, get_small_object_direct_memory_threshold,
    get_stage_timer_if_enabled, get_str, get_transitioned_object_reader_with_tier_manager, inline_admission_shard_size,
    inline_erasure_shard_file_offset, inline_erasure_shard_size, insert_str, is_deadlock_detection_enabled,
@@ -4459,9 +4459,10 @@ impl SetDisks {
                            if let Some(disk) = disk_op
                                && disk.is_online().await
                            {
-                                match create_bitrot_writer(
+                                match create_bitrot_writer_with_orig_volume(
                                    is_inline_buffer,
                                    Some(disk),
+                                    bucket,
                                    RUSTFS_META_TMP_BUCKET,
                                    &tmp_obj,
                                    shard_file_size,
@@ -6128,7 +6129,11 @@ impl SetDisks {
                    );
                }

-                if !tail_owns_tmp_cleanup {
+                let all_local_disks_cleaned_tmp = all_commit_disks_local
+                    && online_disks.len() == commit_set.set_drive_count
+                    && online_disks.iter().all(Option::is_some)
+                    && !issue3031_diag_enabled();
+                if !tail_owns_tmp_cleanup && !all_local_disks_cleaned_tmp {
                    let cleanup_set = commit_set.clone();
                    let cleanup_tmp_dir = commit_tmp_dir.clone();
                    cleanup_runtime::spawn(commit_cleanup_reservation.clone(), async move {
@@ -11064,7 +11069,9 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
        if let Some(err) = &gerr
            && goi.name.is_empty()
        {
-            if should_force_delete_marker_for_missing_version(&opts) {
+            if (is_err_object_not_found(err) || is_err_version_not_found(err))
+                && should_force_delete_marker_for_missing_version(&opts)
+            {
                version_found = false;
            } else {
                if is_err_object_not_found(err) {
@@ -44,15 +44,16 @@ pub(crate) mod lifecycle {

 pub(crate) mod list {
    pub(crate) use rustfs_storage_api::{
-        ListObjectVersionsInfo as StorageListObjectVersionsInfo, ListObjectsInfo, ListObjectsV2Info as StorageListObjectsV2Info,
-        ListOperations, ObjectInfoOrErr as StorageObjectInfoOrErr, VersionMarker, WalkOptions as StorageWalkOptions,
-        WalkVersionsSortOrder,
+        AuthoritativeListOperations, ListObjectVersionsInfo as StorageListObjectVersionsInfo, ListObjectsInfo,
+        ListObjectsV2Info as StorageListObjectsV2Info, ListOperations, ObjectInfoOrErr as StorageObjectInfoOrErr, VersionMarker,
+        WalkOptions as StorageWalkOptions, WalkVersionsSortOrder,
    };
 }

 pub(crate) mod multipart {
    pub(crate) use rustfs_storage_api::{
-        CompletePart, ListMultipartsInfo, ListPartsInfo, MultipartInfo, MultipartOperations, MultipartUploadResult, PartInfo,
+        CompletePart, ListMultipartsInfo, ListPartsInfo, MultipartInfo, MultipartOperations, MultipartUploadInspection,
+        MultipartUploadResult, PartInfo,
    };
 }

@@ -523,7 +523,22 @@ impl ECStore {
    where
        F: FnOnce() -> Result<()> + Send,
    {
-        metadata_sys::update_validated_in(&self.ctx, bucket, config_file, data, validate).await
+        metadata_sys::update_validated_in(&self.ctx, bucket, config_file, data, None, validate).await
+    }
+
+    /// Validate an external fence and the bucket incarnation in one metadata transaction.
+    pub async fn update_bucket_metadata_config_for_incarnation_validated<F>(
+        &self,
+        bucket: &str,
+        config_file: &str,
+        data: Vec<u8>,
+        expected_incarnation_id: Uuid,
+        validate: F,
+    ) -> Result<OffsetDateTime>
+    where
+        F: FnOnce() -> Result<()> + Send,
+    {
+        metadata_sys::update_validated_in(&self.ctx, bucket, config_file, data, Some(expected_incarnation_id), validate).await
    }

    pub async fn bucket_incarnation_id(&self, bucket: &str) -> Result<Uuid> {
@@ -5340,6 +5340,203 @@ mod tests {
        shutdown.cancel();
    }

+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn conditional_multipart_must_not_commit_against_an_older_readable_pool() {
+        use crate::storage_api_contracts::multipart::CompletePart;
+        use crate::storage_api_contracts::object::HTTPPreconditions;
+
+        let temp_dir = tempfile::tempdir().expect("create conditional multipart test directory");
+        let (_ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "conditional-mpu-unreadable", &[4, 4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let bucket = format!("conditional-mpu-{}", uuid::Uuid::new_v4().simple());
+        let object = "object";
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create conditional multipart test bucket");
+        let mut old_body = PutObjReader::from_vec(b"older-pool".to_vec());
+        let old = store.pools[0]
+            .put_object(
+                &bucket,
+                object,
+                &mut old_body,
+                &ObjectOptions {
+                    mod_time: Some(OffsetDateTime::UNIX_EPOCH + time::Duration::seconds(1)),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("write old object to staging pool");
+        let upload = store.pools[0]
+            .new_multipart_upload(&bucket, object, &ObjectOptions::default())
+            .await
+            .expect("create multipart upload in pool 0");
+        let mut part_body = PutObjReader::from_vec(b"stale-writer".to_vec());
+        let part = store.pools[0]
+            .put_object_part(&bucket, object, &upload.upload_id, 1, &mut part_body, &ObjectOptions::default())
+            .await
+            .expect("stage multipart part in pool 0");
+        let mut new_body = PutObjReader::from_vec(b"newer-pool".to_vec());
+        let new = store.pools[1]
+            .put_object(
+                &bucket,
+                object,
+                &mut new_body,
+                &ObjectOptions {
+                    mod_time: Some(OffsetDateTime::UNIX_EPOCH + time::Duration::seconds(2)),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("write newer object to pool 1");
+        let old_etag = old.etag.clone().expect("old object must have an ETag");
+        let before = store.get_object_info(&bucket, object, &ObjectOptions::default()).await;
+
+        let mut metadata = Vec::new();
+        for disk_index in 0..4 {
+            let path = temp_dir
+                .path()
+                .join(format!("pool1/set0/disk{disk_index}/{bucket}/{object}/xl.meta"));
+            let original = tokio::fs::read(&path)
+                .await
+                .expect("save original object metadata before fault injection");
+            metadata.push((path, original));
+        }
+        let mut fault_error = None;
+        for (path, _) in &metadata {
+            if let Err(err) = tokio::fs::write(path, b"invalid object metadata").await {
+                fault_error = Some(err);
+                break;
+            }
+        }
+        let mut fault_contents = Vec::new();
+        if fault_error.is_none() {
+            for (path, _) in &metadata {
+                fault_contents.push(tokio::fs::read(path).await);
+            }
+            store.pools[1]
+                .get_disks_by_key(object)
+                .invalidate_get_object_metadata_cache(&bucket, object)
+                .await;
+        }
+        let unreadable = if fault_error.is_none() {
+            Some(
+                store.pools[1]
+                    .get_object_info(
+                        &bucket,
+                        object,
+                        &ObjectOptions {
+                            no_lock: true,
+                            metadata_cache_safe: false,
+                            suppress_read_repair: true,
+                            ..Default::default()
+                        },
+                    )
+                    .await,
+            )
+        } else {
+            None
+        };
+        let completed = if fault_error.is_none() {
+            let completion_store = Arc::clone(&store);
+            let completion_bucket = bucket.clone();
+            let upload_id = upload.upload_id.clone();
+            let part_etag = part.etag.clone();
+            Some(
+                tokio::spawn(async move {
+                    completion_store
+                        .complete_multipart_upload(
+                            &completion_bucket,
+                            object,
+                            &upload_id,
+                            vec![CompletePart {
+                                part_num: 1,
+                                etag: part_etag,
+                                ..Default::default()
+                            }],
+                            &ObjectOptions {
+                                http_preconditions: Some(HTTPPreconditions {
+                                    if_match: Some(old_etag),
+                                    ..Default::default()
+                                }),
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                })
+                .await,
+            )
+        } else {
+            None
+        };
+        let mut restore_error = None;
+        for (path, original) in &metadata {
+            if let Err(err) = tokio::fs::write(path, original).await {
+                restore_error = Some(err);
+            }
+        }
+        store.pools[1]
+            .get_disks_by_key(object)
+            .invalidate_get_object_metadata_cache(&bucket, object)
+            .await;
+        let after = store.get_object_info(&bucket, object, &ObjectOptions::default()).await;
+        let recovered_body = match store
+            .get_object_reader(&bucket, object, None, HeaderMap::new(), &ObjectOptions::default())
+            .await
+        {
+            Ok(mut reader) => {
+                let mut body = Vec::new();
+                let result = reader
+                    .read_to_end(&mut body)
+                    .await
+                    .map(|_| body)
+                    .map_err(|err| err.to_string());
+                drop(reader);
+                result
+            }
+            Err(err) => Err(err.to_string()),
+        };
+        let parts = store
+            .list_object_parts(&bucket, object, &upload.upload_id, None, 100, &ObjectOptions::default())
+            .await;
+        shutdown.cancel();
+
+        assert!(fault_error.is_none(), "failed to inject metadata fault: {fault_error:?}");
+        assert!(restore_error.is_none(), "failed to restore metadata: {restore_error:?}");
+        assert_eq!(fault_contents.len(), 4, "all four pool 1 metadata files must be inspected");
+        for observed in fault_contents {
+            assert_eq!(
+                observed.expect("read injected metadata"),
+                b"invalid object metadata",
+                "each pool 1 metadata file must contain the injected corruption"
+            );
+        }
+        assert_eq!(before.expect("latest object before fault").etag, new.etag);
+        assert_ne!(old.etag, new.etag, "fixture bodies must produce different ETags");
+        let read_error = unreadable
+            .expect("metadata fault should have been injected")
+            .expect_err("the newer pool must fail its object metadata read during the fault");
+        assert!(
+            !crate::error::is_err_object_not_found(&read_error) && !crate::error::is_err_version_not_found(&read_error),
+            "corrupt metadata must produce a non-not-found read error: {read_error:?}"
+        );
+        let completed = completed
+            .expect("completion should have been attempted")
+            .expect("completion task should not panic");
+        let Err(completion_error) = completed else {
+            panic!("stale If-Match must not commit while the newer object's metadata is unreadable");
+        };
+        assert_eq!(completion_error, read_error, "completion must propagate the unreadable pool's error");
+        assert_eq!(after.expect("newer object after recovery").etag, new.etag);
+        assert_eq!(recovered_body.expect("newer body after recovery"), b"newer-pool");
+        let parts = parts.expect("failed completion must leave the upload retryable");
+        assert_eq!(parts.parts.len(), 1);
+        assert_eq!(parts.parts[0].etag, part.etag);
+    }
+
    async fn migrate_versioned_decommission_test_object(
        store: &Arc<crate::store::ECStore>,
        bucket: &str,
@@ -37,8 +37,8 @@ use crate::error::{
 use crate::object_api::{ObjectInfo, ObjectOptions};
 use crate::set_disk::SetDisks;
 use crate::storage_api_contracts::list::{
-    ListObjectsInfo as StorageListObjectsInfo, StorageListObjectVersionsInfo, StorageListObjectsV2Info, StorageObjectInfoOrErr,
-    StorageWalkOptions, VersionMarker, WalkVersionsSortOrder,
+    AuthoritativeListOperations, ListObjectsInfo as StorageListObjectsInfo, StorageListObjectVersionsInfo,
+    StorageListObjectsV2Info, StorageObjectInfoOrErr, StorageWalkOptions, VersionMarker, WalkVersionsSortOrder,
 };
 use crate::store::ECStore;
 use crate::store::utils::is_reserved_or_invalid_bucket;
@@ -80,6 +80,21 @@ type ListObjectVersionsInfo = StorageListObjectVersionsInfo<ObjectInfo>;
 type ObjectInfoOrErr = StorageObjectInfoOrErr<ObjectInfo, Error>;
 type WalkOptions = StorageWalkOptions<fn(&rustfs_filemeta::FileInfo) -> bool>;

+#[async_trait::async_trait]
+impl AuthoritativeListOperations for ECStore {
+    async fn walk_authoritative(
+        self: Arc<Self>,
+        rx: CancellationToken,
+        bucket: &str,
+        prefix: &str,
+        result: Sender<ObjectInfoOrErr>,
+        opts: WalkOptions,
+    ) -> Result<()> {
+        check_list_objs_args(bucket, prefix, &None)?;
+        self.walk_authoritatively(rx, bucket, prefix, result, opts).await
+    }
+}
+
 /// Build the versioned listing options used by recursive object walks.
 ///
 /// `list_path_result` normally derives these scan bounds before dispatching to
@@ -1292,6 +1307,37 @@ impl ECStore {
        );

        let rx = rx.child_token();
+        if self.pools.len() == 1 && self.pools[0].disk_set.len() == 1 {
+            rustfs_io_metrics::record_list_objects_merge(LIST_OBJECTS_SOURCE_WALKER, 1, 1);
+            let res = self.pools[0].disk_set[0].list_path(rx, opts, sender).await;
+            let mut outcome = ListPathRawOutcome::default();
+            let err = match res {
+                Ok(set_outcome) => {
+                    outcome.limit_reached = set_outcome.limit_reached;
+                    None
+                }
+                Err(err) => Some(err),
+            };
+            let errs = [err];
+            if is_all_not_found(&errs) {
+                return Ok(outcome);
+            }
+            if let Some(err) = errs.into_iter().flatten().next() {
+                return Err(err);
+            }
+            rustfs_io_metrics::record_stage_duration(
+                "store_list_objects_list_merged",
+                merge_started.elapsed().as_secs_f64() * 1000.0,
+            );
+            debug!(
+                set_pair_count = 1,
+                limit_reached = outcome.limit_reached,
+                error_count = 0,
+                "store list_merged finished"
+            );
+            return Ok(outcome);
+        }
+
        let mut futures = Vec::new();

        let mut inputs = Vec::new();
@@ -1519,6 +1565,57 @@ impl ECStore {

        result
    }
+
+    async fn walk_authoritatively(
+        self: Arc<Self>,
+        cancellation: CancellationToken,
+        bucket: &str,
+        prefix: &str,
+        result: Sender<ObjectInfoOrErr>,
+        opts: WalkOptions,
+    ) -> Result<()> {
+        let cancel = cancellation.child_token();
+        let _cancel_on_drop = cancel.clone().drop_guard();
+        let mut scans = Vec::new();
+        let mut inputs = Vec::new();
+        for pool in &self.pools {
+            for set in &pool.disk_set {
+                let (sender, receiver) = mpsc::channel(1);
+                inputs.push(receiver);
+                let mut scan = recursive_walk_list_path_options(bucket.to_owned(), prefix, &opts);
+                // Internal absence checks must include reserved object prefixes.
+                scan.skip_hidden_prefix_check = true;
+                scans.push(set.list_versions_authoritatively(cancel.clone(), scan, sender));
+            }
+        }
+        if scans.is_empty() {
+            return Err(StorageError::ErasureReadQuorum);
+        }
+
+        let (sender, receiver) = mpsc::channel(100);
+        let forwarder = spawn_walk_entry_forwarder(cancel.clone(), receiver, result, opts, bucket.to_owned());
+        let merge = spawn_listing_merge(cancel.clone(), inputs, sender);
+        let scan_results = join_all(scans).await;
+        let merge_result = merge.await.map_err(Error::from)?;
+        let forward_result = forwarder.await.map_err(Error::from);
+
+        // An unavailable set cannot establish absence, even when another set
+        // completed. Only an explicitly missing directory or volume is empty.
+        for scan in scan_results {
+            match scan {
+                Ok(outcome) if !outcome.limit_reached => {}
+                Ok(_) => return Err(StorageError::ErasureReadQuorum),
+                Err(StorageError::FileNotFound | StorageError::VolumeNotFound) => {}
+                Err(err) => return Err(err),
+            }
+        }
+        merge_result?;
+        forward_result?;
+        if cancellation.is_cancelled() {
+            return Err(Error::other("authoritative object walk was cancelled"));
+        }
+        Ok(())
+    }
 }

 /// Forwards merged raw entries to a walk consumer as `ObjectInfoOrErr`,
@@ -1811,7 +1908,7 @@ async fn gather_results(
        if opts.versioned
            && is_object
            && entry
-                .xl_meta()
+                .xl_meta_ref()
                .is_ok_and(|meta| !meta.versions.is_empty() && meta.versions.iter().all(|version| version.header.free_version()))
        {
            continue;
@@ -2489,7 +2586,8 @@ impl SetDisks {
    /// Version scan for set-scoped callers that act on absence: set walks
    /// (tier reference proofs, tier-delete journals, pool maintenance) and the
    /// recursive prefix-delete scan. Every online drive is read and each
-    /// version is resolved at the read quorum of its own erasure layout; a
+    /// version uses the read quorum of its own erasure layout. Null versions
+    /// in the system bucket also retain GET's commit-proof selection. A
    /// version that cannot be ruled out fails the scan instead of being
    /// omitted. Merged `ListObjectVersions` pages go through `list_path`.
    async fn list_versions_authoritatively(
@@ -2828,6 +2926,7 @@ mod test {
    use crate::error::{Result, StorageError};
    use crate::object_api::{ObjectInfo, ObjectOptions, PutObjReader};
    use crate::set_disk::{SetDisks, hermetic_set_disks_isolated};
+    use crate::storage_api_contracts::list::AuthoritativeListOperations as _;
    use crate::storage_api_contracts::namespace::NamespaceLocking as _;
    use crate::storage_api_contracts::object::ObjectIO as _;
    use bytes::Bytes;
@@ -5688,6 +5787,126 @@ mod test {
        ));
    }

+    #[test]
+    fn authoritative_system_bucket_scan_rejects_uncommitted_null_version_split() {
+        let time = time::OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid fixture time");
+        let old = test_unversioned_object_meta_entry("entry.json", time, "old", Uuid::from_u128(1));
+        for new_etag in ["new", "old"] {
+            let new =
+                test_unversioned_object_meta_entry("entry.json", time + time::Duration::seconds(1), new_etag, Uuid::from_u128(2));
+            let entries = MetaCacheEntries(vec![Some(old.clone()), Some(old.clone()), Some(new.clone()), Some(new)]);
+
+            assert!(matches!(
+                SetDisks::resolve_listed_versions(RUSTFS_META_BUCKET, entries, &[None, None, None, None], 4, 2),
+                Err(DiskError::ErasureReadQuorum)
+            ));
+        }
+    }
+
+    #[test]
+    fn authoritative_system_bucket_scan_selects_committed_null_version() {
+        let time = time::OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid fixture time");
+        let old = test_unversioned_object_meta_entry("entry.json", time, "old", Uuid::from_u128(1));
+        let new = test_unversioned_object_meta_entry("entry.json", time + time::Duration::seconds(1), "new", Uuid::from_u128(2));
+        for (new_copies, expected) in [(1, "old"), (3, "new"), (4, "new")] {
+            let entries = MetaCacheEntries(
+                (0..4)
+                    .map(|slot| Some(if slot < new_copies { new.clone() } else { old.clone() }))
+                    .collect(),
+            );
+            let resolved = SetDisks::resolve_listed_versions(RUSTFS_META_BUCKET, entries, &[None, None, None, None], 4, 2)
+                .expect("committed generation should resolve")
+                .expect("committed record should be present");
+            let info = resolved
+                .to_fileinfo(RUSTFS_META_BUCKET)
+                .expect("resolved metadata should decode");
+
+            assert_eq!(info.metadata.get("etag").map(String::as_str), Some(expected));
+        }
+    }
+
+    #[test]
+    fn authoritative_system_bucket_scan_matches_degraded_null_version_reads() {
+        let time = time::OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid fixture time");
+        let old = test_unversioned_object_meta_entry("entry.json", time, "old", Uuid::from_u128(1));
+        let new = test_unversioned_object_meta_entry("entry.json", time + time::Duration::seconds(1), "new", Uuid::from_u128(2));
+        for (old_copies, new_copies, expected) in [
+            (2, 1, Some("old")),
+            (1, 2, None),
+            (0, 3, Some("new")),
+            (0, 2, Some("new")),
+            (0, 1, None),
+        ] {
+            let entries = MetaCacheEntries(
+                (0..4)
+                    .map(|slot| {
+                        if slot < old_copies {
+                            Some(old.clone())
+                        } else if slot < old_copies + new_copies {
+                            Some(new.clone())
+                        } else {
+                            None
+                        }
+                    })
+                    .collect(),
+            );
+            let errors = (0..4)
+                .map(|slot| (slot >= old_copies + new_copies).then_some(DiskError::DiskNotFound))
+                .collect::<Vec<_>>();
+            let resolved = SetDisks::resolve_listed_versions(RUSTFS_META_BUCKET, entries, &errors, 4, 2);
+
+            if let Some(expected) = expected {
+                let info = resolved
+                    .expect("degraded committed record should resolve")
+                    .expect("degraded committed record should be present")
+                    .to_fileinfo(RUSTFS_META_BUCKET)
+                    .expect("resolved metadata should decode");
+                assert_eq!(info.metadata.get("etag").map(String::as_str), Some(expected));
+            } else {
+                assert!(matches!(resolved, Err(DiskError::ErasureReadQuorum)));
+            }
+        }
+    }
+
+    #[test]
+    fn authoritative_user_bucket_scan_keeps_null_version_read_quorum_selection() {
+        let time = time::OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid fixture time");
+        let old = test_unversioned_object_meta_entry("entry.json", time, "old", Uuid::from_u128(1));
+        let new = test_unversioned_object_meta_entry("entry.json", time + time::Duration::seconds(1), "new", Uuid::from_u128(2));
+        for old_copies in [1, 2] {
+            let entries = MetaCacheEntries(vec![
+                Some(new.clone()),
+                Some(new.clone()),
+                Some(old.clone()),
+                (old_copies == 2).then(|| old.clone()),
+            ]);
+            let errors = [None, None, None, (old_copies == 1).then_some(DiskError::DiskNotFound)];
+            let resolved = SetDisks::resolve_listed_versions("bucket", entries, &errors, 4, 2)
+                .expect("user bucket read quorum should resolve")
+                .expect("user bucket record should be present");
+            let info = resolved.to_fileinfo("bucket").expect("resolved metadata should decode");
+
+            assert_eq!(info.metadata.get("etag").map(String::as_str), Some("new"));
+        }
+    }
+
+    #[test]
+    fn authoritative_system_bucket_scan_keeps_explicit_version_read_quorum() {
+        let time = time::OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid fixture time");
+        let entry = test_object_meta_entry_with_erasure_versions("entry.json", &[(time, "version", 2, 2)]);
+        let entries = MetaCacheEntries(vec![Some(entry.clone()), Some(entry), None, None]);
+        let errors = [None, None, Some(DiskError::DiskNotFound), Some(DiskError::DiskNotFound)];
+        let resolved = SetDisks::resolve_listed_versions(RUSTFS_META_BUCKET, entries, &errors, 4, 2)
+            .expect("explicit version read quorum should resolve")
+            .expect("explicit version should be present");
+        let info = resolved
+            .to_fileinfo(RUSTFS_META_BUCKET)
+            .expect("resolved metadata should decode");
+
+        assert_eq!(info.version_id, Some(Uuid::from_u128(1)));
+        assert_eq!(info.metadata.get("etag").map(String::as_str), Some("version"));
+    }
+
    #[test]
    fn authoritative_version_scan_tolerates_corrupt_replicas_only_with_version_quorum() {
        let entry = test_object_meta_entry_with_erasure_versions(
@@ -6029,6 +6248,49 @@ mod test {
        Bytes::from(meta.marshal_msg().expect("encode fixture xl.meta"))
    }

+    #[tokio::test]
+    async fn authoritative_store_walk_includes_reserved_prefix_versions() {
+        let bucket = "catalog-cleanup-versions";
+        let name = ".rustfs-table/namespaces/ns/tables/table/metadata/metadata.json";
+        let (_dirs, store) = single_set_store(16).await;
+        let set = &store.pools[0].disk_set[0];
+        for (slot, disk) in set.disks.read().await.iter().enumerate() {
+            let disk = disk.as_ref().expect("fixture drive is online");
+            disk.make_volume(bucket).await.expect("create bucket volume");
+            disk.write_all(
+                bucket,
+                &format!("{name}/{STORAGE_FORMAT_FILE}"),
+                versioned_object_metadata(bucket, name, &[(2, 1_700_000_002), (1, 1_700_000_001)], slot),
+            )
+            .await
+            .expect("persist reserved metadata versions");
+        }
+        let (sender, mut receiver) = mpsc::channel(1);
+        let walk = store
+            .clone()
+            .walk_authoritative(CancellationToken::new(), bucket, "", sender, super::WalkOptions::default());
+        let collect = async {
+            let mut versions = Vec::new();
+            while let Some(item) = receiver.recv().await {
+                assert!(item.err.is_none(), "reserved metadata must decode: {:?}", item.err);
+                let item = item.item.expect("walk must return an object");
+                versions.push((item.name, item.version_id));
+            }
+            versions
+        };
+        let (walk, versions) = timeout(Duration::from_secs(10), async { tokio::join!(walk, collect) })
+            .await
+            .expect("authoritative reserved-prefix walk must terminate");
+        walk.expect("authoritative walk must succeed");
+        assert_eq!(
+            versions,
+            [
+                (name.to_string(), Some(Uuid::from_u128(1))),
+                (name.to_string(), Some(Uuid::from_u128(2)))
+            ]
+        );
+    }
+
    // A versioned DeleteObject(versionId) can reach write quorum while three
    // drives are offline; they keep the deleted version after they return. A
    // different drive going offline later must not turn that sub-quorum
@@ -6099,6 +6361,16 @@ mod test {
                .await
                .expect("set walk must terminate");
            assert!(matches!(walk, Err(StorageError::ErasureReadQuorum)), "{walk:?}");
+            let (sender, mut receiver) = mpsc::channel(16);
+            let walk =
+                store
+                    .clone()
+                    .walk_authoritative(CancellationToken::new(), bucket, "", sender, super::WalkOptions::default());
+            let drain = async { while receiver.recv().await.is_some() {} };
+            let (walk, ()) = timeout(Duration::from_secs(10), async { tokio::join!(walk, drain) })
+                .await
+                .expect("authoritative store walk must terminate");
+            assert!(matches!(walk, Err(StorageError::ErasureReadQuorum)), "{walk:?}");
            let recursive_delete_scan = set
                .clone()
                .inner_list_object_versions_for_recursive_delete(bucket, "", None, None, 1000)
@@ -498,6 +498,17 @@ pub struct ECStore {
    pub(crate) bucket_fence_registry: Arc<bucket_fence::BucketFenceRegistry>,
 }

+impl Drop for ECStore {
+    /// Clear the instance context's `BucketMetadataSys` when its backing store
+    /// is dropped. Without this, the shared bootstrap context retains a
+    /// `BucketMetadataSys` whose `Weak<ECStore>` is dangling after the store
+    /// goes away (e.g. between sequential test fixtures that share the
+    /// bootstrap), and the next `object_store()` call panics.
+    fn drop(&mut self) {
+        self.ctx.clear_bucket_metadata_sys_if_stale();
+    }
+}
+
 const METRIC_SCANNER_DATA_MOVEMENT_PAUSED: &str = "rustfs_scanner_data_movement_paused";
 const METRIC_SCANNER_DATA_MOVEMENT_PAUSE_DURATION_SECONDS: &str = "rustfs_scanner_data_movement_pause_duration_seconds";
 const METRIC_SCANNER_DATA_MOVEMENT_BACKLOG_WORK_ITEMS: &str = "rustfs_scanner_data_movement_backlog_work_items";
@@ -23,6 +23,23 @@ use std::collections::HashSet;

 const MULTIPART_LIST_SET_CONCURRENCY: usize = 4;

+#[async_trait::async_trait]
+impl crate::storage_api_contracts::multipart::MultipartUploadInspection for ECStore {
+    type Error = Error;
+
+    async fn ensure_no_multipart_uploads(&self, bucket: &str) -> Result<()> {
+        let (_, lifecycle_guard) = self
+            .guard_multipart_bucket_incarnation(bucket, &ObjectOptions::default())
+            .await?;
+        for pool in &self.pools {
+            for set in &pool.disk_set {
+                set.ensure_no_multipart_uploads(bucket).await?;
+            }
+        }
+        ensure_multipart_bucket_lifecycle_guard_held(lifecycle_guard.as_ref(), bucket, "")
+    }
+}
+
 #[cfg(test)]
 static DATA_MOVEMENT_MULTIPART_DISCOVERY_COUNTS: std::sync::OnceLock<std::sync::Mutex<std::collections::HashMap<Uuid, usize>>> =
    std::sync::OnceLock::new();
@@ -612,7 +629,7 @@ impl ECStore {
        }
        if opts.http_preconditions.is_some()
            && !opts.data_movement
-            && let Some(err) = self.check_write_precondition(bucket, object, &opts).await
+            && let Some(err) = self.check_write_precondition(bucket, object, &opts, false).await
        {
            return Err(err);
        }
@@ -1154,12 +1171,33 @@ impl ECStore {
        Ok(())
    }

-    async fn check_write_precondition(&self, bucket: &str, object: &str, opts: &ObjectOptions) -> Option<StorageError> {
+    async fn check_write_precondition(
+        &self,
+        bucket: &str,
+        object: &str,
+        opts: &ObjectOptions,
+        require_all_pools: bool,
+    ) -> Option<StorageError> {
        let mut lookup_opts = opts.clone();
        let http_preconditions = lookup_opts.http_preconditions?;
        lookup_opts.http_preconditions = None;
        lookup_opts.no_lock = true;
-        let oi = self.get_object_info_for_delete(bucket, object, &lookup_opts).await;
+        let oi = if require_all_pools {
+            lookup_opts.metadata_cache_safe = false;
+            lookup_opts.suppress_read_repair = true;
+            if let Err(err) = check_object_args(bucket, object) {
+                return Some(err);
+            }
+            let encoded_object = encode_dir_object(object);
+            self.get_latest_object_info_for_conditional_write(bucket, &encoded_object, &lookup_opts)
+                .await
+                .and_then(|(info, _)| {
+                    lookup_opts.precondition_check(&info)?;
+                    Ok(info)
+                })
+        } else {
+            self.get_object_info_for_delete(bucket, object, &lookup_opts).await
+        };
        crate::set_disk::SetDisks::evaluate_write_precondition(oi, &http_preconditions, bucket, object, opts)
    }

@@ -1186,11 +1224,13 @@ impl ECStore {
                .await;
            return enqueue_transition_after_write(self.as_ref(), result, LcEventSrc::S3CompleteMultipartUpload, opts).await;
        }
-        if opts.http_preconditions.is_some()
-            && !opts.data_movement
-            && let Some(err) = self.check_write_precondition(bucket, object, opts).await
-        {
-            return Err(err);
+        if opts.http_preconditions.is_some() && !opts.data_movement {
+            // Preserve explicitly scoped internal reads and borrowed-lock callers.
+            let require_all_pools =
+                !opts.replication_request && !opts.no_lock && !opts.skip_decommissioned && !opts.skip_rebalancing;
+            if let Some(err) = self.check_write_precondition(bucket, object, opts, require_all_pools).await {
+                return Err(err);
+            }
        }

        let result = match self.multipart_upload_pool_idx(bucket, object, upload_id, opts).await {
@@ -5637,13 +5637,11 @@ impl ECStore {
        }

        if self.single_pool()
-            && !creates_latest_marker
            && !opts.data_movement
            && opts.http_preconditions.is_none()
            && !opts.replication_request
-            && !opts.versioned
            && !opts.version_suspended
-            && opts.version_id.is_none()
+            && !opts.delete_marker
            && !self.is_suspended(0).await
            && !self.is_pool_rebalancing(0).await
        {
@@ -5655,6 +5653,10 @@ impl ECStore {
                .await
            {
                Ok(mut res) => {
+                    #[cfg(test)]
+                    if creates_latest_marker {
+                        pause_versioned_delete_marker_after_commit(bucket, object).await;
+                    }
                    res.name = decode_dir_object(res.name.as_str());
                    return Ok(res);
                }
@@ -837,7 +837,29 @@ impl ECStore {
        object: &str,
        opts: &ObjectOptions,
    ) -> Result<(ObjectInfo, usize)> {
-        let suspended_pools = if opts.skip_decommissioned {
+        self.get_latest_object_info_with_pool_policy(bucket, object, opts, false)
+            .await
+    }
+
+    /// Observe every pool for a conditional-write preflight. This does not
+    /// hold the namespace locks needed to prove a condition at publication.
+    pub(super) async fn get_latest_object_info_for_conditional_write(
+        &self,
+        bucket: &str,
+        object: &str,
+        opts: &ObjectOptions,
+    ) -> Result<(ObjectInfo, usize)> {
+        self.get_latest_object_info_with_pool_policy(bucket, object, opts, true).await
+    }
+
+    async fn get_latest_object_info_with_pool_policy(
+        &self,
+        bucket: &str,
+        object: &str,
+        opts: &ObjectOptions,
+        require_all_pools: bool,
+    ) -> Result<(ObjectInfo, usize)> {
+        let suspended_pools = if opts.skip_decommissioned && !require_all_pools {
            let pool_meta = self.pool_meta.read().await;
            Some(
                (0..self.pools.len())
@@ -853,7 +875,7 @@ impl ECStore {
                continue;
            }

-            if opts.skip_rebalancing && self.is_pool_rebalancing(idx).await {
+            if opts.skip_rebalancing && !require_all_pools && self.is_pool_rebalancing(idx).await {
                continue;
            }

@@ -873,6 +895,9 @@ impl ECStore {
                    });
                }
                Err(e) => {
+                    if require_all_pools && !is_err_object_not_found(&e) && !is_err_version_not_found(&e) {
+                        return Err(e);
+                    }
                    candidates.push(LatestObjectInfoCandidate {
                        info: None,
                        idx,
@@ -2070,6 +2070,53 @@ mod test {
        assert!(meta.versions.iter().all(|version| !version.header.free_version()));
    }

+    #[test]
+    fn bytes_count_toward_listing_limit_matches_fileinfo_visibility() {
+        let mut meta = FileMeta::new();
+        let mut obj = FileInfo::new("object", 2, 2);
+        obj.version_id = Some(Uuid::new_v4());
+        obj.mod_time = Some(OffsetDateTime::now_utc());
+        meta.add_version(obj.clone()).expect("add object version");
+        let obj_bytes = meta.marshal_msg().expect("marshal object");
+        assert!(FileMeta::bytes_count_toward_listing_limit(&obj_bytes));
+
+        let dm = FileInfo {
+            name: "object".to_string(),
+            version_id: Some(Uuid::new_v4()),
+            deleted: true,
+            mark_deleted: true,
+            mod_time: obj.mod_time.map(|t| t + time::Duration::seconds(1)),
+            ..Default::default()
+        };
+        meta.delete_version(&dm).expect("add delete marker");
+        let dm_bytes = meta.marshal_msg().expect("marshal delete marker");
+        assert!(!FileMeta::bytes_count_toward_listing_limit(&dm_bytes));
+
+        let (_, mut source) = tier_overwrite_fixture(crate::TransitionVersionState::Exact);
+        source.version_id = Some(Uuid::new_v4());
+        let mut transitioned = FileMeta::new();
+        transitioned.add_version(source.clone()).expect("seed transitioned version");
+        let mut del_req = FileInfo {
+            name: "object".to_string(),
+            version_id: source.version_id,
+            mod_time: source.mod_time,
+            ..Default::default()
+        };
+        del_req.set_tier_free_version_id(&Uuid::new_v4().to_string());
+        transitioned
+            .delete_version(&del_req)
+            .expect("delete transitioned version leaving free version");
+        assert!(!FileMeta::bytes_count_toward_listing_limit(
+            &transitioned.marshal_msg().expect("marshal free-only")
+        ));
+
+        transitioned.add_version(obj).expect("add live object version");
+        assert!(FileMeta::bytes_count_toward_listing_limit(
+            &transitioned.marshal_msg().expect("marshal free + live object")
+        ));
+        assert!(FileMeta::bytes_count_toward_listing_limit(b"corrupt-xl-meta"));
+    }
+
    #[test]
    fn add_version_filemata_uses_canonical_equal_time_order() {
        let mod_time = OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid test timestamp");
@@ -324,6 +324,63 @@ impl FileMeta {
        true
    }

+    pub(crate) fn buffers_match_strict(self_buf: &[u8], other_buf: &[u8]) -> bool {
+        let Ok((s_meta, _)) = Self::parse_validated_xl_envelope(self_buf) else {
+            return false;
+        };
+        let Ok((o_meta, _)) = Self::parse_validated_xl_envelope(other_buf) else {
+            return false;
+        };
+
+        if s_meta.is_empty() || o_meta.is_empty() {
+            return s_meta.is_empty() && o_meta.is_empty();
+        }
+
+        let Ok((s_len, s_header_ver, _s_meta_ver, s_meta)) = Self::decode_validated_xl_headers(s_meta) else {
+            return false;
+        };
+        let Ok((o_len, o_header_ver, _o_meta_ver, o_meta)) = Self::decode_validated_xl_headers(o_meta) else {
+            return false;
+        };
+
+        if s_len != o_len {
+            return false;
+        }
+
+        let mut s_cur: Cursor<&[u8]> = Cursor::new(s_meta);
+        let mut o_cur: Cursor<&[u8]> = Cursor::new(o_meta);
+        for _ in 0..s_len {
+            let Ok((s_header_raw, _)) = Self::read_next_raw_slices(s_meta, &mut s_cur) else {
+                return false;
+            };
+            let Ok((o_header_raw, _)) = Self::read_next_raw_slices(o_meta, &mut o_cur) else {
+                return false;
+            };
+            if s_header_ver == XL_HEADER_VERSION
+                && o_header_ver == XL_HEADER_VERSION
+                && s_header_raw.len() == 38
+                && o_header_raw.len() == 38
+                && s_header_raw == o_header_raw
+                && parse_canonical_v3_header_fields(s_header_raw).is_some()
+            {
+                continue;
+            }
+            let mut s_header = FileMetaVersionHeader::default();
+            if s_header.unmarshal_v(s_header_ver, s_header_raw).is_err() {
+                return false;
+            }
+            let mut o_header = FileMetaVersionHeader::default();
+            if o_header.unmarshal_v(o_header_ver, o_header_raw).is_err() {
+                return false;
+            }
+            if !Self::headers_match_strict(&s_header, &o_header) {
+                return false;
+            }
+        }
+
+        true
+    }
+
    #[inline]
    fn header_matches_canonical_v3_raw_strict(s_header: &FileMetaVersionHeader, buf: &[u8]) -> Option<bool> {
        if buf.len() != 38 {
@@ -625,6 +682,49 @@ impl FileMeta {
        }
    }

+    pub fn bytes_count_toward_listing_limit(buf: &[u8]) -> bool {
+        let Ok((meta, _inline_buf)) = Self::parse_validated_xl_envelope(buf) else {
+            return true;
+        };
+        if meta.is_empty() {
+            return false;
+        }
+        let Ok((versions_len, header_ver, _meta_ver, meta)) = Self::decode_validated_xl_headers(meta) else {
+            return true;
+        };
+        if versions_len == 0 {
+            return false;
+        }
+        let mut cur = Cursor::new(meta);
+        let mut selected: Option<(FileMetaVersionHeader, &[u8])> = None;
+        let mut any_non_hidden = false;
+        for _ in 0..versions_len {
+            let Ok((header, ver_meta)) = Self::read_next_shallow_slice(meta, &mut cur, header_ver) else {
+                return true;
+            };
+            if header.version_type != VersionType::Delete && !header.free_version() {
+                any_non_hidden = true;
+            }
+            if selected.is_some() {
+                continue;
+            }
+            if header.free_version() && header.version_id.is_some_and(|id| !id.is_nil()) {
+                continue;
+            }
+            selected = Some((header, ver_meta));
+        }
+        let Some((header, ver_meta)) = selected else {
+            return false;
+        };
+        if !header.free_version() && header.version_type != VersionType::Delete {
+            return true;
+        }
+        if !any_non_hidden {
+            return false;
+        }
+        FileMetaVersion::decode_fileinfo_from_slice(ver_meta, "", "", true, false).is_err()
+    }
+
    #[hotpath::measure(impl_type = "FileMeta")]
    pub fn marshal_msg(&self) -> Result<Vec<u8>> {
        let mut wr =
@@ -352,6 +352,22 @@ impl MetaCacheEntry {
        (prefer, true)
    }

+    pub fn matches_strict_borrowed(&mut self, other: &MetaCacheEntry) -> bool {
+        if self.name != other.name {
+            return false;
+        }
+
+        if other.is_dir() || self.is_dir() {
+            return other.is_dir() == self.is_dir();
+        }
+
+        if self.cached.is_none() && other.cached.is_none() {
+            return FileMeta::buffers_match_strict(&self.metadata, &other.metadata);
+        }
+
+        self.matches_strict(other)
+    }
+
    pub fn matches_strict(&mut self, other: &MetaCacheEntry) -> bool {
        if self.name != other.name {
            return false;
@@ -989,6 +1005,17 @@ impl<W: AsyncWrite + Unpin> MetacacheWriter<W> {
        Ok(())
    }

+    pub async fn write_obj_buffered(&mut self, obj: &MetaCacheEntry) -> Result<()> {
+        self.init().await?;
+
+        Self::encode_obj(&mut self.buf, obj)?;
+        if self.buf.len() >= METACACHE_READ_CHUNK_SIZE {
+            self.flush().await?;
+        }
+
+        Ok(())
+    }
+
    pub async fn close(&mut self) -> Result<()> {
        self.close_with_limit_reached(false).await
    }
@@ -1282,6 +1309,85 @@ impl<R: AsyncRead + Unpin> MetacacheReader<R> {
        Ok(u32::from_be_bytes(buf.try_into().expect("Slice with incorrect length")))
    }

+    pub fn skip_current(&mut self) -> bool {
+        if self.has_current && self.init && self.err.is_none() {
+            self.has_current = false;
+            true
+        } else {
+            false
+        }
+    }
+
+    pub fn try_peek_buffered(&mut self) -> Option<Result<Option<MetaCacheEntry>>> {
+        if !self.init || self.err.is_some() || self.terminated {
+            return None;
+        }
+
+        let avail = &self.buf[self.pos..];
+        let &first = avail.first()?;
+        if first != 0xc3 {
+            return None;
+        }
+
+        let str_marker = *avail.get(1)?;
+        let (str_hdr, str_len) = match str_marker {
+            0xa0..=0xbf => (1usize, usize::from(str_marker & 0x1f)),
+            0xd9 => (2usize, usize::from(*avail.get(2)?)),
+            0xda => (3usize, usize::from(u16::from_be_bytes([*avail.get(2)?, *avail.get(3)?]))),
+            0xdb => (
+                5usize,
+                usize::try_from(u32::from_be_bytes([*avail.get(2)?, *avail.get(3)?, *avail.get(4)?, *avail.get(5)?])).ok()?,
+            ),
+            _ => return None,
+        };
+
+        let bin_offset = 1usize.checked_add(str_hdr)?.checked_add(str_len)?;
+        let bin_marker = *avail.get(bin_offset)?;
+        let (bin_hdr, bin_len) = match bin_marker {
+            0xc4 => (2usize, usize::from(*avail.get(bin_offset + 1)?)),
+            0xc5 => (
+                3usize,
+                usize::from(u16::from_be_bytes([*avail.get(bin_offset + 1)?, *avail.get(bin_offset + 2)?])),
+            ),
+            0xc6 => (
+                5usize,
+                usize::try_from(u32::from_be_bytes([
+                    *avail.get(bin_offset + 1)?,
+                    *avail.get(bin_offset + 2)?,
+                    *avail.get(bin_offset + 3)?,
+                    *avail.get(bin_offset + 4)?,
+                ]))
+                .ok()?,
+            ),
+            _ => return None,
+        };
+
+        let total_len = bin_offset.checked_add(bin_hdr)?.checked_add(bin_len)?;
+        if avail.len() < total_len {
+            return None;
+        }
+
+        let Ok(decoded) = from_utf8(&avail[1 + str_hdr..bin_offset]) else {
+            return None;
+        };
+        let name = decoded.to_owned();
+        let metadata = if bin_len == 0 {
+            Vec::new()
+        } else {
+            avail[bin_offset + bin_hdr..total_len].to_vec()
+        };
+
+        self.pos += total_len;
+        self.reset();
+        self.has_current = true;
+        Some(Ok(Some(MetaCacheEntry {
+            name,
+            metadata,
+            cached: None,
+            reusable: false,
+        })))
+    }
+
    pub async fn skip(&mut self, size: usize) -> Result<()> {
        self.check_init().await?;

@@ -4257,4 +4363,95 @@ mod tests {
            .expect("resolve mixed empty");
        assert_eq!(res_mixed_empty.cached.expect("cached").versions.len(), 1);
    }
+
+    #[test]
+    fn matches_strict_borrowed_matches_shallow_headers_and_rejects_corrupt_or_divergent_buffers() {
+        let base_bytes = create_real_xlmeta().expect("create real xlmeta");
+        let mut peer_fm = FileMeta::load(&base_bytes).expect("load base xlmeta");
+        // Mutate inner version payload (ver.meta) without changing shallow header.
+        let mut ver = FileMetaVersion::try_from(peer_fm.versions[1].meta.as_slice()).expect("decode version");
+        ver.object.as_mut().expect("object version").erasure_index += 1;
+        peer_fm.versions[1].meta = ver.marshal_msg().expect("marshal peer version");
+        let peer_bytes = peer_fm.marshal_msg().expect("marshal peer xlmeta");
+        assert_ne!(base_bytes, peer_bytes, "peer shard has distinct erasure index payload");
+
+        let mut e1 = MetaCacheEntry {
+            name: "obj-a".to_owned(),
+            metadata: base_bytes.clone(),
+            cached: None,
+            reusable: false,
+        };
+        let e1_clone = e1.clone();
+        let e2 = MetaCacheEntry {
+            name: "obj-a".to_owned(),
+            metadata: peer_bytes,
+            cached: None,
+            reusable: false,
+        };
+        assert!(e1.matches_strict_borrowed(&e2));
+        assert!(e1.matches_strict_borrowed(&e1_clone));
+
+        // Different shallow header must not match.
+        let mut newer_fm = FileMeta::load(&base_bytes).expect("load base xlmeta");
+        newer_fm.versions[0].header.version_id = Some(Uuid::new_v4());
+        let newer_entry = MetaCacheEntry {
+            name: "obj-a".to_owned(),
+            metadata: newer_fm.marshal_msg().expect("marshal newer xlmeta"),
+            cached: None,
+            reusable: false,
+        };
+        assert!(!e1.matches_strict_borrowed(&newer_entry));
+
+        // Identical corrupt buffers must return false, never true via byte-equality shortcut.
+        let mut corrupt = MetaCacheEntry {
+            name: "obj-a".to_owned(),
+            metadata: b"XL2 corrupt-payload".to_vec(),
+            cached: None,
+            reusable: false,
+        };
+        let corrupt_peer = corrupt.clone();
+        assert!(!corrupt.matches_strict_borrowed(&corrupt_peer));
+    }
+
+    #[tokio::test]
+    async fn try_peek_buffered_and_skip_current_match_async_peek_and_skip() {
+        let entries: Vec<_> = (0..5)
+            .map(|i| MetaCacheEntry {
+                name: format!("buffered-key-{i}"),
+                metadata: vec![i as u8; 16],
+                cached: None,
+                reusable: false,
+            })
+            .collect();
+        let mut buf = Cursor::new(Vec::new());
+        let mut writer = MetacacheWriter::new(&mut buf);
+        writer.write(&entries).await.expect("write entries");
+        writer.close_with_limit_reached(false).await.expect("close stream");
+
+        let mut reader = MetacacheReader::new(Cursor::new(buf.into_inner()));
+        assert!(reader.try_peek_buffered().is_none(), "uninitialized reader must defer to async peek");
+        assert!(!reader.skip_current(), "no current entry before first peek");
+
+        let first = reader.peek().await.expect("first peek").expect("entry 0");
+        assert_eq!(first, entries[0]);
+        assert!(reader.skip_current(), "skip_current must consume the peeked entry");
+        assert!(!reader.skip_current(), "second skip_current without peek must return false");
+
+        for expected in &entries[1..] {
+            let next = reader
+                .try_peek_buffered()
+                .expect("remaining entries are already buffered")
+                .expect("valid entry")
+                .expect("present entry");
+            assert_eq!(&next, expected);
+            assert!(reader.skip_current());
+        }
+
+        assert!(
+            reader.try_peek_buffered().is_none(),
+            "stream terminator must defer to async peek for trailer handling"
+        );
+        assert!(reader.peek().await.expect("eof peek").is_none());
+        assert_eq!(reader.limit_reached(), Some(false));
+    }
 }
@@ -420,6 +420,7 @@ async fn real_storage_grace_round_trip() {
            &HealOpts {
                scan_mode: rustfs_heal_contracts::heal_channel::HealScanMode::Deep,
                recreate: true,
+                remove: true,
                pool: Some(0),
                set: Some(0),
                ..Default::default()
@@ -157,7 +157,7 @@ async fn wait_for_object_copies(disks: &[PathBuf], bucket: &str, object: &str) {
 fn deep_heal_opts() -> HealOpts {
    HealOpts {
        recreate: true,
-        remove: false,
+        remove: true,
        scan_mode: HealScanMode::Deep,
        ..Default::default()
    }
@@ -633,6 +633,7 @@ mod absence_receipt_regressions {
                version_id: Some(old.to_owned()),
            },
            HealOptions {
+                remove_corrupted: true,
                scan_mode: HealScanMode::Deep,
                pool_index: Some(0),
                set_index: Some(0),
@@ -823,6 +824,7 @@ mod absence_receipt_regressions {
            },
            HealOptions {
                recursive: true,
+                remove_corrupted: true,
                scan_mode: HealScanMode::Deep,
                ..Default::default()
            },
@@ -810,7 +810,13 @@ fn admin_mixed_batch_retains_repairs_after_object_retry_exhaustion() {
                    restarted.start().await.expect("restart with retained failed batch");
                    let restored = restarted.get_task_report(&token).await.expect("original token after restart");
                    assert_eq!(restored.status, report.status);
-                    assert_eq!(restored.outcome, report.outcome);
+                    // Quorum and Deep observations are local and never persisted.
+                    let mut expected_outcome = report.outcome.as_deref().cloned().expect("mixed batch outcome");
+                    for object in &mut expected_outcome.objects {
+                        object.quorum_selected_data_dir = None;
+                        object.deep_check_observed = false;
+                    }
+                    assert_eq!(restored.outcome.as_deref(), Some(&expected_outcome));
                    assert_eq!(restarted.get_queue_length().await, 0, "terminal batch must not replay");
                    restarted.stop().await.expect("stop restarted mixed batch manager");
                });
@@ -596,9 +596,205 @@ pub struct PendingScannerHeal {
    pub last_admission_reason: String,
 }

+/// One bucket's standing in its set's service order. The wait position, the
+/// quanta granted and the last certification are separate quantities: a bucket
+/// that keeps failing moves to the tail like any other and only adds debt.
+#[derive(Clone, Copy, Debug, Default, Deserialize, PartialEq, Eq)]
+pub struct DataUsageBucketService {
+    #[serde(default)]
+    pub waiting_since: u64,
+    #[serde(default)]
+    pub service_quanta: u64,
+    #[serde(default)]
+    pub last_certified: Option<u64>,
+    #[serde(default)]
+    pub failure_debt: u64,
+    #[serde(default)]
+    pub incarnation: Option<uuid::Uuid>,
+}
+
+// Nested derives encode positionally; a map lets later fields stay optional.
+impl Serialize for DataUsageBucketService {
+    fn serialize<S: serde::Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
+        let field_count = 2
+            + usize::from(self.last_certified.is_some())
+            + usize::from(self.failure_debt != 0)
+            + usize::from(self.incarnation.is_some());
+        let mut map = serializer.serialize_map(Some(field_count))?;
+        map.serialize_entry("waiting_since", &self.waiting_since)?;
+        map.serialize_entry("service_quanta", &self.service_quanta)?;
+        if let Some(last_certified) = self.last_certified {
+            map.serialize_entry("last_certified", &last_certified)?;
+        }
+        if self.failure_debt != 0 {
+            map.serialize_entry("failure_debt", &self.failure_debt)?;
+        }
+        if let Some(incarnation) = self.incarnation {
+            map.serialize_entry("incarnation", &incarnation)?;
+        }
+        map.end()
+    }
+}
+
+/// Durable first-in-first-out service order of one set, in generations: every
+/// bucket at or below `generation_floor` still waits for its turn and everything
+/// above it was served, so a new bucket queues behind waiters but ahead of
+/// buckets that already had a turn. It only orders work:
+/// no entry proves coverage, authorizes a skip or releases a responsibility,
+/// so losing it restarts fairness and nothing else. Like the set cache's bucket
+/// entries and incarnations, it holds one record per bucket of the inventory.
+#[derive(Clone, Debug, Default, Deserialize, PartialEq, Eq)]
+pub struct DataUsageBucketServiceLedger {
+    /// Movement epoch the order was built under; another epoch invalidates it.
+    #[serde(default)]
+    pub publication_epoch: u64,
+    /// Leader that saved the order; an incomplete save stamps nothing else.
+    #[serde(default)]
+    pub leader_epoch: u64,
+    #[serde(default)]
+    pub next_sequence: u64,
+    #[serde(default)]
+    pub generation_floor: u64,
+    #[serde(default)]
+    pub buckets: HashMap<String, DataUsageBucketService>,
+}
+
+impl Serialize for DataUsageBucketServiceLedger {
+    fn serialize<S: serde::Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
+        let mut map = serializer.serialize_map(Some(5))?;
+        map.serialize_entry("publication_epoch", &self.publication_epoch)?;
+        map.serialize_entry("leader_epoch", &self.leader_epoch)?;
+        map.serialize_entry("next_sequence", &self.next_sequence)?;
+        map.serialize_entry("generation_floor", &self.generation_floor)?;
+        map.serialize_entry("buckets", &self.buckets)?;
+        map.end()
+    }
+}
+
+impl DataUsageBucketServiceLedger {
+    /// Rebuilds the order for this cycle's inventory. Decoded state that fails
+    /// validation or belongs to another movement epoch starts a new order;
+    /// a bucket whose incarnation changed is a new arrival behind all waiters.
+    pub(crate) fn restore<'a>(
+        persisted: Option<Self>,
+        publication_epoch: u64,
+        inventory: impl Iterator<Item = &'a str>,
+        incarnations: Option<&HashMap<String, uuid::Uuid>>,
+    ) -> Self {
+        let mut ledger = persisted
+            .filter(|ledger| ledger.publication_epoch == publication_epoch && ledger.is_valid())
+            .unwrap_or_else(|| Self {
+                publication_epoch,
+                ..Default::default()
+            });
+        let previous = std::mem::take(&mut ledger.buckets);
+        let mut arrivals = Vec::new();
+        for bucket in inventory {
+            let incarnation = incarnations.and_then(|current| current.get(bucket)).copied();
+            match previous.get(bucket) {
+                Some(service) if service.incarnation.is_none() || incarnation.is_none() || service.incarnation == incarnation => {
+                    ledger.buckets.insert(
+                        bucket.to_owned(),
+                        DataUsageBucketService {
+                            incarnation: incarnation.or(service.incarnation),
+                            ..*service
+                        },
+                    );
+                }
+                _ => arrivals.push((bucket, incarnation)),
+            }
+        }
+        if arrivals.is_empty() {
+            let floor = ledger.generation_floor;
+            if ledger.buckets.values().all(|service| service.waiting_since > floor) {
+                ledger.generation_floor = ledger.next_sequence;
+            }
+            return ledger;
+        }
+        // Open one position between the waiters and the served buckets.
+        let floor = ledger.generation_floor;
+        ledger.generation_floor = floor.saturating_add(1);
+        ledger.next_sequence = ledger.next_sequence.saturating_add(1);
+        for service in ledger.buckets.values_mut().filter(|service| service.waiting_since > floor) {
+            service.waiting_since = service.waiting_since.saturating_add(1);
+        }
+        for (bucket, incarnation) in arrivals {
+            ledger.buckets.insert(
+                bucket.to_owned(),
+                DataUsageBucketService {
+                    waiting_since: ledger.generation_floor,
+                    incarnation,
+                    ..Default::default()
+                },
+            );
+        }
+        ledger
+    }
+
+    fn is_valid(&self) -> bool {
+        self.generation_floor <= self.next_sequence
+            && self
+                .buckets
+                .iter()
+                .all(|(bucket, service)| !bucket.is_empty() && service.waiting_since <= self.next_sequence)
+    }
+
+    /// Smaller positions have waited longer; an untracked bucket queues last.
+    pub(crate) fn wait_position(&self, bucket: &str) -> u64 {
+        self.buckets.get(bucket).map_or(u64::MAX, |service| service.waiting_since)
+    }
+
+    /// Moves a bucket to the tail and returns its previous wait position.
+    pub(crate) fn record_service(&mut self, bucket: &str) -> Option<u64> {
+        let service = self.buckets.get_mut(bucket)?;
+        self.next_sequence = self.next_sequence.saturating_add(1);
+        let previous = std::mem::replace(&mut service.waiting_since, self.next_sequence);
+        service.service_quanta = service.service_quanta.saturating_add(1);
+        Some(previous)
+    }
+
+    /// Undoes `record_service` for an admission a resource veto turned back.
+    pub(crate) fn return_service(&mut self, bucket: &str, waiting_since: u64) {
+        if let Some(service) = self.buckets.get_mut(bucket) {
+            service.waiting_since = waiting_since;
+            service.service_quanta = service.service_quanta.saturating_sub(1);
+        }
+    }
+
+    pub(crate) fn record_certified(&mut self, bucket: &str) {
+        if let Some(service) = self.buckets.get_mut(bucket) {
+            service.last_certified = Some(service.waiting_since);
+            service.failure_debt = 0;
+        }
+    }
+
+    pub(crate) fn record_failure(&mut self, bucket: &str) {
+        if let Some(service) = self.buckets.get_mut(bucket) {
+            service.failure_debt = service.failure_debt.saturating_add(1);
+        }
+    }
+
+    /// Buckets holding unpaid failures and the first of them by name.
+    pub(crate) fn failure_debt(&self) -> (usize, Option<&str>) {
+        let mut debtors = self
+            .buckets
+            .iter()
+            .filter(|(_, service)| service.failure_debt > 0)
+            .map(|(bucket, _)| bucket.as_str());
+        let first = debtors.next();
+        let (count, first) = debtors.fold((usize::from(first.is_some()), first), |(count, first), bucket| {
+            (count + 1, first.min(Some(bucket)))
+        });
+        (count, first)
+    }
+}
+
 /// Data usage cache info
 #[derive(Clone, Debug, Default, Deserialize)]
 pub struct DataUsageCacheInfo {
+    /// In-memory only: this partial stopped on admission pressure, not a failure.
+    #[serde(skip)]
+    pub scan_resource_deferred: bool,
    pub name: String,
    pub next_cycle: u64,
    pub last_update: Option<SystemTime>,
@@ -615,6 +811,11 @@ pub struct DataUsageCacheInfo {
    pub scan_raw_enumeration_cursor: Option<DataUsageRawEnumerationCursor>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub scan_raw_enumeration_page_index: Option<RawEnumerationPageIndex>,
+    /// Indexes of the other directories on the active walk path. A restart
+    /// must replay every wide ancestor before it reaches the selected one, and
+    /// recharging those replays would stop a fixed budget from ever advancing.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub scan_raw_enumeration_path_page_indexes: Vec<RawEnumerationPageIndex>,
    #[serde(default)]
    pub scan_identity: Option<DataUsageScanIdentity>,
    #[serde(default)]
@@ -674,6 +875,9 @@ pub struct DataUsageCacheInfo {
    /// skipping an unselected bucket in a later scoped set scan.
    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
    pub scan_bucket_incarnations: HashMap<String, uuid::Uuid>,
+    /// Bucket service order of a set cache; absent from bucket caches.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub scan_bucket_service: Option<DataUsageBucketServiceLedger>,
 }

 impl Serialize for DataUsageCacheInfo {
@@ -686,6 +890,7 @@ impl Serialize for DataUsageCacheInfo {
        let field_count = 16
            + usize::from(self.scan_raw_enumeration_cursor.is_some())
            + usize::from(self.scan_raw_enumeration_page_index.is_some())
+            + usize::from(!self.scan_raw_enumeration_path_page_indexes.is_empty())
            + usize::from(self.scan_identity.is_some())
            + usize::from(self.scan_progress.is_some())
            + usize::from(self.scan_coverage_receipt.is_some())
@@ -699,7 +904,8 @@ impl Serialize for DataUsageCacheInfo {
            + usize::from(self.lkg_scan_plan_digest.is_some())
            + usize::from(self.scan_execution_digest.is_some())
            + usize::from(self.segment_invalidation_proof.is_some())
-            + usize::from(!self.scan_bucket_incarnations.is_empty());
+            + usize::from(!self.scan_bucket_incarnations.is_empty())
+            + usize::from(self.scan_bucket_service.is_some());
        let mut state = serializer.serialize_map(Some(field_count))?;
        state.serialize_entry("name", &self.name)?;
        state.serialize_entry("next_cycle", &self.next_cycle)?;
@@ -717,6 +923,9 @@ impl Serialize for DataUsageCacheInfo {
        if let Some(index) = &self.scan_raw_enumeration_page_index {
            state.serialize_entry("scan_raw_enumeration_page_index", index)?;
        }
+        if !self.scan_raw_enumeration_path_page_indexes.is_empty() {
+            state.serialize_entry("scan_raw_enumeration_path_page_indexes", &self.scan_raw_enumeration_path_page_indexes)?;
+        }
        if let Some(identity) = self.scan_identity {
            state.serialize_entry("scan_identity", &identity)?;
        }
@@ -765,6 +974,9 @@ impl Serialize for DataUsageCacheInfo {
        if !self.scan_bucket_incarnations.is_empty() {
            state.serialize_entry("scan_bucket_incarnations", &self.scan_bucket_incarnations)?;
        }
+        if let Some(ledger) = &self.scan_bucket_service {
+            state.serialize_entry("scan_bucket_service", ledger)?;
+        }
        state.end()
    }
 }
@@ -902,6 +1114,30 @@ impl DataUsageCache {
        }
    }

+    /// Identifies how far durable resume state has advanced. Discovery alone
+    /// does not change it; only committed raw pages or coverage frontier do.
+    pub(crate) fn durable_scan_position(&self) -> (Option<String>, Option<String>, u64) {
+        let raw_entries = self
+            .info
+            .scan_raw_enumeration_page_index
+            .iter()
+            .chain(&self.info.scan_raw_enumeration_path_page_indexes)
+            .map(|index| match index.status() {
+                RawEnumerationPageOwnerStatus::Unsupported => 0,
+                RawEnumerationPageOwnerStatus::Building { indexed_entries, .. }
+                | RawEnumerationPageOwnerStatus::Ready { indexed_entries, .. } => indexed_entries,
+            })
+            .fold(0, u64::saturating_add);
+        (
+            self.info.scan_resume_after.clone(),
+            self.info
+                .scan_coverage_receipt
+                .as_ref()
+                .map(|receipt| receipt.through.clone()),
+            raw_entries,
+        )
+    }
+
    pub(crate) fn has_scan_position(&self) -> bool {
        self.info.scan_resume_after.is_some()
            || self.info.scan_checkpoint.is_some()
@@ -1044,7 +1280,14 @@ impl DataUsageCache {
        }
        if self.validated_raw_enumeration_page_index().is_none() {
            self.info.scan_raw_enumeration_page_index = None;
+            self.info.scan_raw_enumeration_path_page_indexes.clear();
        }
+        let path_indexes = std::mem::take(&mut self.info.scan_raw_enumeration_path_page_indexes);
+        self.info.scan_raw_enumeration_path_page_indexes = path_indexes
+            .into_iter()
+            .take(MAX_DATA_USAGE_CACHE_DEPTH)
+            .filter(|index| self.raw_enumeration_page_index_is_valid(index))
+            .collect();
        let cursor_is_valid = !self.has_scan_position()
            || self.validated_scan_frontier().is_some()
            || self.info.scan_raw_enumeration_cursor.is_some()
@@ -1087,6 +1330,7 @@ impl DataUsageCache {
            self.info.scan_checkpoint = None;
            self.info.scan_raw_enumeration_cursor = None;
            self.info.scan_raw_enumeration_page_index = None;
+            self.info.scan_raw_enumeration_path_page_indexes.clear();
            self.info.scan_coverage_receipt = None;
        }
        // Old readers do not understand coverage sweeps. An absent plan makes
@@ -1230,22 +1474,37 @@ impl DataUsageCache {
    }

    pub(crate) fn validated_raw_enumeration_page_index(&self) -> Option<&RawEnumerationPageIndex> {
-        let index = self.info.scan_raw_enumeration_page_index.as_ref()?;
+        self.info
+            .scan_raw_enumeration_page_index
+            .as_ref()
+            .filter(|index| self.raw_enumeration_page_index_is_valid(index))
+    }
+
+    /// The durable index of one directory on the saved walk path.
+    pub(crate) fn validated_raw_enumeration_page_index_for_parent(&self, parent: &str) -> Option<&RawEnumerationPageIndex> {
+        self.info
+            .scan_raw_enumeration_page_index
+            .iter()
+            .chain(&self.info.scan_raw_enumeration_path_page_indexes)
+            .find(|index| index.parent() == Some(parent))
+            .filter(|index| self.raw_enumeration_page_index_is_valid(index))
+    }
+
+    fn raw_enumeration_page_index_is_valid(&self, index: &RawEnumerationPageIndex) -> bool {
        if self.info.scan_progress.is_none()
            || !self.info.scan_identity.is_some_and(|identity| identity.is_valid())
            || self.info.source.is_none()
            || index.committed_entries().is_err()
            || index.indexed_entries().is_err()
        {
-            return None;
+            return false;
        }
-        let parent = match index.status() {
-            RawEnumerationPageOwnerStatus::Unsupported => return None,
+        match index.status() {
+            RawEnumerationPageOwnerStatus::Unsupported => false,
            RawEnumerationPageOwnerStatus::Building { parent, .. } | RawEnumerationPageOwnerStatus::Ready { parent, .. } => {
-                parent
+                path_is_in_bucket_scope(&self.info.name, &parent)
            }
-        };
-        path_is_in_bucket_scope(&self.info.name, &parent).then_some(index)
+        }
    }

    /// Seal only the frontier supplied by completed traversal, never a restored cursor.
@@ -57,6 +57,66 @@ fn scoped_scan_coverage_metadata_preserves_map_compatibility() {
    assert_eq!(decoded.scan_coverage_digest, Some(TEST_PLAN_DIGEST));
 }

+#[test]
+fn raw_enumeration_path_indexes_preserve_map_compatibility_and_validation() {
+    #[derive(serde::Deserialize)]
+    struct LegacyInfo {
+        name: String,
+        scan_raw_enumeration_page_index: Option<RawEnumerationPageIndex>,
+    }
+    let index_for = |parent: &str| RawEnumerationPageIndex::new(parent, 8).expect("fixture index");
+    let mut cache = DataUsageCache {
+        info: DataUsageCacheInfo {
+            name: "bucket".to_string(),
+            scan_progress: Some(DataUsageScanProgress {
+                started_plan: TEST_PLAN_DIGEST,
+                requested_plan: TEST_PLAN_DIGEST,
+            }),
+            scan_identity: Some(DataUsageScanIdentity {
+                version: 1,
+                bucket_incarnation: uuid::Uuid::from_u128(1),
+                set_layout: DataUsageScanPlanDigest([2; 32]),
+                publication_epoch: 1,
+                tier_registry_generation: 0,
+                scan_mode: HealScanMode::Normal,
+            }),
+            source: Some(DataUsageCacheSource::new(0, 0)),
+            scan_raw_enumeration_page_index: Some(index_for("bucket/wide")),
+            scan_raw_enumeration_path_page_indexes: vec![index_for("bucket"), index_for("other-bucket/wide")],
+            ..Default::default()
+        },
+        cache: HashMap::new(),
+    };
+    assert!(cache.validated_raw_enumeration_page_index_for_parent("bucket").is_some());
+    assert!(
+        cache
+            .validated_raw_enumeration_page_index_for_parent("other-bucket/wide")
+            .is_none(),
+        "a saved path index must stay inside its bucket"
+    );
+
+    let encoded = rmp_serde::to_vec(&cache.info).expect("path indexes should remain map encoded");
+    let legacy: LegacyInfo = rmp_serde::from_slice(&encoded).expect("old map readers should ignore the path indexes");
+    assert_eq!(legacy.name, "bucket");
+    assert_eq!(
+        legacy
+            .scan_raw_enumeration_page_index
+            .and_then(|index| index.parent().map(str::to_owned)),
+        Some("bucket/wide".to_string())
+    );
+    let decoded: DataUsageCacheInfo = rmp_serde::from_slice(&encoded).expect("new reader should restore the path indexes");
+    assert_eq!(decoded.scan_raw_enumeration_path_page_indexes.len(), 2);
+
+    cache.info.scan_raw_enumeration_path_page_indexes.clear();
+    let encoded = rmp_serde::to_vec(&cache.info).expect("empty path indexes should encode");
+    assert!(
+        !encoded
+            .windows(b"scan_raw_enumeration_path_page_indexes".len())
+            .any(|window| window == b"scan_raw_enumeration_path_page_indexes"),
+        "the default must stay absent from the map"
+    );
+}
+
 #[derive(Debug, PartialEq, Eq)]
 struct CachePutRecord {
    object: String,
@@ -80,6 +80,7 @@ impl RemoteScannerBudget {
            max_duration: self.max_duration_ms.map(Duration::from_millis),
            max_objects: self.max_objects,
            max_directories: self.max_directories,
+            ..Default::default()
        }
    }
 }
@@ -743,6 +743,7 @@ pub(crate) fn lookup_scanner_runtime_config(
        max_duration: cycle_max_duration.filter(|duration| !duration.is_zero()),
        max_objects: cycle_max_objects,
        max_directories: cycle_max_directories,
+        ..Default::default()
    };

    let (bitrot_cycle, bitrot_cycle_source) = if let Some(value) = rustfs_utils::get_env_opt_str(ENV_SCANNER_BITROT_CYCLE_SECS) {
@@ -1171,6 +1172,7 @@ mod tests {
                max_duration: Some(Duration::from_secs(1800)),
                max_objects: Some(1000),
                max_directories: Some(8000),
+                ..Default::default()
            },
            ..Default::default()
        };
@@ -474,14 +474,17 @@ fn scan_cycle_partial_reason(reason: Option<ScannerCycleBudgetReason>) -> ScanCy
    match reason {
        Some(ScannerCycleBudgetReason::Runtime) => ScanCyclePartialReason::Runtime,
        Some(ScannerCycleBudgetReason::Objects) => ScanCyclePartialReason::Objects,
-        Some(ScannerCycleBudgetReason::Directories) => ScanCyclePartialReason::Directories,
+        // Raw enumeration is directory work; the reason set is a stable metric label.
+        Some(ScannerCycleBudgetReason::Directories | ScannerCycleBudgetReason::Entries) => ScanCyclePartialReason::Directories,
        None => ScanCyclePartialReason::Unknown,
    }
 }

 fn scan_cycle_partial_source(reason: Option<ScannerCycleBudgetReason>) -> Option<ScannerWorkSource> {
    match reason {
-        Some(ScannerCycleBudgetReason::Objects | ScannerCycleBudgetReason::Directories) => Some(ScannerWorkSource::Usage),
+        Some(ScannerCycleBudgetReason::Objects | ScannerCycleBudgetReason::Directories | ScannerCycleBudgetReason::Entries) => {
+            Some(ScannerWorkSource::Usage)
+        }
        Some(ScannerCycleBudgetReason::Runtime) | None => None,
    }
 }
@@ -1059,6 +1062,31 @@ fn prepare_cycle_for_usage_floor_bootstrap(
    }
 }

+/// Overwrite an unreadable usage floor with a bootstrap marker using the
+/// file's current revision for CAS. A successful publish leaves a
+/// `BootstrapPending` floor that the normal load path picks up on retry.
+/// Returns `Err` when the floor is not even readable (transient I/O) so the
+/// caller keeps the supervisor-retry behavior instead of rebuilding.
+async fn rebuild_unreadable_usage_floor(
+    storeapi: Arc<impl ScannerObjectIO + ScannerConfigObjectDelete>,
+) -> Result<(), ScannerError> {
+    let Some(expected_epoch) = scanner_publication_epoch(storeapi.clone()).await else {
+        return Err(ScannerError::Other("scanner usage floor rebuild is blocked by data movement".to_string()));
+    };
+    let (_data, revision) = read_config_with_revision(storeapi.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
+        .await
+        .map_err(|err| ScannerError::Other(format!("scanner usage floor revision unreadable for rebuild: {err}")))?;
+    publish_scanner_usage_bootstrap_primary(
+        storeapi,
+        &revision,
+        expected_epoch,
+        None,
+        ScannerUsageBootstrapPublishContext::Recovery,
+        || true,
+    )
+    .await
+}
+
 pub async fn init_data_scanner(ctx: CancellationToken, storeapi: Arc<ECStore>) {
    init_data_scanner_with_storage(ctx, storeapi, ScannerRuntimeGuard::new()).await;
 }
@@ -3087,6 +3115,27 @@ where
            Err(err) => usage_floor_result = Err(err),
        }
    }
+    // backlog#2918: after dual-corruption recovery has been attempted an
+    // unreadable floor is no longer a quota proof. Report the failure, then
+    // rebuild through a CAS overwrite before giving up. Transient I/O errors
+    // fail the rebuild and fall through to the supervisor-retry stop below.
+    if usage_floor_result.is_err() {
+        let error = usage_floor_result.as_ref().unwrap_err().to_string();
+        error!(
+            target: "rustfs::scanner",
+            event = EVENT_SCANNER_PERSIST_STATE,
+            component = LOG_COMPONENT_SCANNER,
+            subsystem = LOG_SUBSYSTEM_RUNTIME,
+            path = %DATA_USAGE_OBJ_NAME_PATH.as_str(),
+            state = "usage_floor_load_failed",
+            error = %error,
+            "Scanner usage floor is unreadable after recovery; attempting rebuild"
+        );
+        record_scanner_usage_floor_failure(error);
+        if rebuild_unreadable_usage_floor(storeapi.clone()).await.is_ok() {
+            usage_floor_result = persisted_usage_floor_for_startup(storeapi.clone(), true).await;
+        }
+    }
    let (usage_floor, usage_floor_startup) = match usage_floor_result {
        Ok(result) => result,
        Err(err) => {
@@ -3099,7 +3148,7 @@ where
                path = %DATA_USAGE_OBJ_NAME_PATH.as_str(),
                state = "usage_floor_load_failed",
                error = %err,
-                "Scanner stopped because the persisted usage floor could not be loaded"
+                "Scanner usage floor could not be loaded or rebuilt; stopping leader iteration for supervisor retry"
            );
            record_scanner_usage_floor_failure(error.clone());
            global_metrics().set_cycle(None).await;
@@ -41,20 +41,6 @@ fn reset_recovery_status() {
    .expect("test must be able to reset the local recovery observation");
 }

-async fn wait_for_recovery_attempt(attempt: u32) -> ScannerCycleRecoveryStatus {
-    tokio::time::timeout(WAIT_BUDGET, async {
-        loop {
-            let status = scanner_cycle_recovery_status();
-            if status.retry_count >= u64::from(attempt) {
-                return status;
-            }
-            tokio::time::sleep(Duration::from_millis(10)).await;
-        }
-    })
-    .await
-    .unwrap_or_else(|_| panic!("recovery attempt {attempt} was never recorded: {:?}", scanner_cycle_recovery_status()))
-}
-
 async fn wait_for_recovery_state(state: &str) -> ScannerCycleRecoveryStatus {
    tokio::time::timeout(WAIT_BUDGET, async {
        loop {
@@ -130,20 +116,19 @@ fn started_rebalance() -> EcstoreRebalanceMeta {
    }
 }

-/// rustfs/rustfs#8670: a usage floor that is present but untrusted must not
-/// drive the supervisor into a terminal `paused, retryable=false` observation.
-/// Every attempt stays retryable with a bounded backoff, and repairing the
-/// floor recovers the scanner without a process restart.
+/// rustfs/rustfs#8670 + backlog#2918: a usage floor that is present but
+/// untrusted must not stop the scanner. The unreadable floor is reported and
+/// rebuilt through a CAS overwrite so the scanner recovers without a process
+/// restart or a terminal `paused, retryable=false` observation.
 #[tokio::test]
 #[serial]
-async fn usage_floor_load_failure_backs_off_and_recovers_without_restart() {
+async fn usage_floor_load_failure_rebuilds_and_recovers_without_restart() {
    temp_env::async_with_vars(SUPERVISOR_ENV, async {
        crate::runtime_config::refresh_scanner_runtime_config_for_tests();
        crate::scanner_io::clear_dirty_usage_buckets_for_tests();
        reset_recovery_status();
        let (_dir, store) = setup_scanner_cycle_store().await;
-        // Present but without a baseline identity: the shape #8670 reported,
-        // whose error points the operator at the usage-state reset.
+        // Present but without a baseline identity: the shape #8670 reported.
        save_config(
            store.clone(),
            DATA_USAGE_OBJ_NAME_PATH.as_str(),
@@ -156,41 +141,11 @@ async fn usage_floor_load_failure_backs_off_and_recovers_without_restart() {
        let _cancel_on_drop = ctx.clone().drop_guard();
        init_data_scanner_with_storage(ctx.clone(), store.clone(), ScannerRuntimeGuard::new()).await;

-        let attempts = MAX_SCANNER_CYCLE_RECOVERY_RETRIES + 2;
-        let mut first_detected = None;
-        for attempt in 1..=attempts {
-            let status = wait_for_recovery_attempt(attempt).await;
-            assert_eq!(status.classification.as_deref(), Some("usage_floor_load_failed"), "attempt {attempt}");
-            assert!(status.retryable, "attempt {attempt} must stay retryable: {status:?}");
-            assert_ne!(status.state, "paused", "attempt {attempt} must not report a terminal pause");
-            if attempt >= MAX_SCANNER_CYCLE_RECOVERY_RETRIES {
-                assert_eq!(status.state, "backoff", "attempt {attempt}: {status:?}");
-            } else {
-                assert_eq!(status.state, "usage_floor_load_failed", "attempt {attempt}: {status:?}");
-            }
-            assert!(
-                status
-                    .reason
-                    .as_deref()
-                    .is_some_and(|reason| reason.contains("usage-state/reset") && reason.contains("next automatic attempt")),
-                "the operator hint and the next attempt must both be visible: {status:?}"
-            );
-            assert!(status.next_attempt_at_unix_secs.is_some(), "attempt {attempt}: {status:?}");
-            match first_detected {
-                None => first_detected = status.first_detected_at_unix_secs,
-                Some(first) => assert_eq!(status.first_detected_at_unix_secs, Some(first), "attempt {attempt}"),
-            }
-            if attempt == attempts {
-                save_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str(), valid_usage_floor(0, 0))
-                    .await
-                    .expect("repaired usage floor should persist");
-            }
-            notify_scanner_cycle_recovery_wake();
-        }
-
+        // backlog#2918: the unreadable floor is rebuilt automatically; the
+        // scanner must reach healthy without a manual repair or terminal pause.
        let healthy = wait_for_recovery_state("healthy").await;
-        assert_eq!(healthy.retry_count, 0, "a successful attempt clears the retry budget: {healthy:?}");
-        assert!(!healthy.retryable);
+        assert!(!healthy.retryable, "rebuild must clear the retry budget: {healthy:?}");
+        assert_ne!(healthy.state, "paused", "rebuild must not report a terminal pause");
        wait_for_persisted_cycle(&store, |cycle, _| cycle.next >= 1).await;
        assert_eq!(scanner_cycle_schedule_status().execution_role, "leader");
        stop_supervisor(ctx).await;
@@ -24,13 +24,51 @@ const BUDGET_REASON_NONE: u8 = 0;
 const BUDGET_REASON_RUNTIME: u8 = 1;
 const BUDGET_REASON_OBJECTS: u8 = 2;
 const BUDGET_REASON_DIRECTORIES: u8 = 3;
+const BUDGET_REASON_ENTRIES: u8 = 4;
 const PROGRESS_CLOCK_SAMPLE_INTERVAL: u64 = 128;
+/// One raw-enumeration page, so a directory stop always commits durable work.
+const DIRECTORY_ENTRY_BUDGET_FLOOR: u64 = 128;
+/// Directory entries one object or directory budget unit pays for. A forward
+/// sweep classifies in sorted order and so needs a whole listing first; a
+/// directory a few times wider than the budget must still reach objects in
+/// its first round, while a far wider one builds its index across rounds.
+const DIRECTORY_ENTRIES_PER_COUNT_UNIT: u64 = 16;

 #[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
 pub(crate) struct ScannerCycleBudgetConfig {
    pub max_duration: Option<Duration>,
    pub max_objects: Option<u64>,
    pub max_directories: Option<u64>,
+    /// Raw directory entries first observed under this budget. Entries replayed
+    /// from a validated page index are not charged.
+    pub max_entries: Option<u64>,
+    /// Cumulative name bytes of the charged raw entries.
+    pub max_entry_name_bytes: Option<u64>,
+    /// One bucket's share. `None` lets a bucket consume the whole budget.
+    pub bucket_quantum: Option<ScannerBucketQuantum>,
+}
+
+/// Limits of one bucket admission, charged to the owning cycle budget as well.
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub(crate) struct ScannerBucketQuantum {
+    pub max_duration: Option<Duration>,
+    pub max_objects: Option<u64>,
+    pub max_directories: Option<u64>,
+    pub max_entries: Option<u64>,
+    pub max_entry_name_bytes: Option<u64>,
+}
+
+impl From<ScannerBucketQuantum> for ScannerCycleBudgetConfig {
+    fn from(quantum: ScannerBucketQuantum) -> Self {
+        Self {
+            max_duration: quantum.max_duration,
+            max_objects: quantum.max_objects,
+            max_directories: quantum.max_directories,
+            max_entries: quantum.max_entries,
+            max_entry_name_bytes: quantum.max_entry_name_bytes,
+            bucket_quantum: None,
+        }
+    }
 }

 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
@@ -38,6 +76,7 @@ pub(crate) enum ScannerCycleBudgetReason {
    Runtime,
    Objects,
    Directories,
+    Entries,
 }

 impl ScannerCycleBudgetReason {
@@ -46,6 +85,7 @@ impl ScannerCycleBudgetReason {
            Self::Runtime => BUDGET_REASON_RUNTIME,
            Self::Objects => BUDGET_REASON_OBJECTS,
            Self::Directories => BUDGET_REASON_DIRECTORIES,
+            Self::Entries => BUDGET_REASON_ENTRIES,
        }
    }

@@ -54,12 +94,15 @@ impl ScannerCycleBudgetReason {
            BUDGET_REASON_RUNTIME => Some(Self::Runtime),
            BUDGET_REASON_OBJECTS => Some(Self::Objects),
            BUDGET_REASON_DIRECTORIES => Some(Self::Directories),
+            BUDGET_REASON_ENTRIES => Some(Self::Entries),
            _ => None,
        }
    }
 }

 pub struct ScannerCycleBudget {
+    /// The cycle budget a bucket quantum also charges.
+    parent: Option<Arc<ScannerCycleBudget>>,
    token: CancellationToken,
    reason: Arc<AtomicU8>,
    started_at: Instant,
@@ -67,11 +110,17 @@ pub struct ScannerCycleBudget {
    max_duration: Option<Duration>,
    max_objects: Option<u64>,
    max_directories: Option<u64>,
+    max_entries: Option<u64>,
+    max_entry_name_bytes: Option<u64>,
+    directory_entry_limit: Option<u64>,
+    bucket_quantum: Option<ScannerBucketQuantum>,
    track_progress: bool,
    track_unbounded_counts: bool,
    objects_scanned: AtomicU64,
    directories_started: AtomicU64,
    entries_visited: AtomicU64,
+    entries_charged: AtomicU64,
+    entry_name_bytes_charged: AtomicU64,
    last_progress_millis: AtomicU64,
    cycle_state_persisted: AtomicBool,
    remote_scan_progress: Arc<RemoteScanProgress>,
@@ -80,16 +129,16 @@ pub struct ScannerCycleBudget {
 impl ScannerCycleBudget {
    #[cfg(test)]
    pub(crate) fn new(parent: &CancellationToken, config: ScannerCycleBudgetConfig) -> Arc<Self> {
-        Self::new_inner(parent, config, false, false)
+        Self::new_inner(parent, config, false, false, None)
    }

    pub(crate) fn new_with_progress_tracking(parent: &CancellationToken, config: ScannerCycleBudgetConfig) -> Arc<Self> {
-        Self::new_inner(parent, config, true, true)
+        Self::new_inner(parent, config, true, true, None)
    }

    pub(crate) fn new_with_runtime_progress_tracking(parent: &CancellationToken, config: ScannerCycleBudgetConfig) -> Arc<Self> {
        let track_progress = config.max_duration.is_some();
-        Self::new_inner(parent, config, track_progress, false)
+        Self::new_inner(parent, config, track_progress, false, None)
    }

    fn new_inner(
@@ -97,6 +146,7 @@ impl ScannerCycleBudget {
        config: ScannerCycleBudgetConfig,
        track_progress: bool,
        track_unbounded_counts: bool,
+        cycle_budget: Option<Arc<ScannerCycleBudget>>,
    ) -> Arc<Self> {
        let token = parent.child_token();
        let reason = Arc::new(AtomicU8::new(BUDGET_REASON_NONE));
@@ -124,7 +174,11 @@ impl ScannerCycleBudget {
            });
        }

+        let remote_scan_progress = cycle_budget
+            .as_ref()
+            .map_or_else(|| Arc::new(RemoteScanProgress::default()), |budget| budget.remote_scan_progress());
        Arc::new(Self {
+            parent: cycle_budget,
            token,
            reason,
            started_at,
@@ -132,14 +186,20 @@ impl ScannerCycleBudget {
            max_duration: config.max_duration,
            max_objects: config.max_objects,
            max_directories: config.max_directories,
+            max_entries: config.max_entries,
+            max_entry_name_bytes: config.max_entry_name_bytes,
+            directory_entry_limit: directory_entry_limit(&config),
+            bucket_quantum: config.bucket_quantum,
            track_progress,
            track_unbounded_counts,
            objects_scanned: AtomicU64::new(0),
            directories_started: AtomicU64::new(0),
            entries_visited: AtomicU64::new(0),
+            entries_charged: AtomicU64::new(0),
+            entry_name_bytes_charged: AtomicU64::new(0),
            last_progress_millis: AtomicU64::new(0),
            cycle_state_persisted: AtomicBool::new(false),
-            remote_scan_progress: Arc::new(RemoteScanProgress::default()),
+            remote_scan_progress,
        })
    }

@@ -147,12 +207,31 @@ impl ScannerCycleBudget {
        self.token.clone()
    }

+    /// Budget and cancellation scope of one bucket scan. Without a quantum the
+    /// bucket shares this budget; a quantum stops only its own scan token.
+    pub(crate) fn for_bucket(self: &Arc<Self>, ctx: &CancellationToken) -> (Arc<Self>, CancellationToken) {
+        let Some(quantum) = self.bucket_quantum else {
+            return (self.clone(), ctx.child_token());
+        };
+        let budget = Self::new_inner(ctx, quantum.into(), self.track_progress, self.track_unbounded_counts, Some(self.clone()));
+        let token = budget.token();
+        (budget, token)
+    }
+
+    /// True when only this bucket's share ended and the cycle can continue.
+    pub(crate) fn quantum_exhausted(&self) -> bool {
+        self.parent.as_ref().is_some_and(|parent| !parent.budget_elapsed())
+            && self.reason.load(Ordering::Relaxed) != BUDGET_REASON_NONE
+    }
+
    pub(crate) fn budget_elapsed(&self) -> bool {
        self.reason.load(Ordering::Relaxed) != BUDGET_REASON_NONE
+            || self.parent.as_ref().is_some_and(|parent| parent.budget_elapsed())
    }

    pub(crate) fn reason(&self) -> Option<ScannerCycleBudgetReason> {
        ScannerCycleBudgetReason::from_code(self.reason.load(Ordering::Relaxed))
+            .or_else(|| self.parent.as_ref().and_then(|parent| parent.reason()))
    }

    pub(crate) fn max_duration(&self) -> Option<Duration> {
@@ -176,7 +255,10 @@ impl ScannerCycleBudget {
    }

    pub(crate) fn requires_serial_progress_accounting(&self) -> bool {
-        self.max_objects.is_some() || self.max_directories.is_some()
+        self.max_objects.is_some()
+            || self.max_directories.is_some()
+            || self.max_entries.is_some()
+            || self.max_entry_name_bytes.is_some()
    }

    pub(crate) fn remaining_config(&self) -> ScannerCycleBudgetConfig {
@@ -195,6 +277,13 @@ impl ScannerCycleBudget {
            max_directories: self
                .max_directories
                .map(|max| max.saturating_sub(self.directories_started.load(Ordering::Relaxed))),
+            max_entries: self
+                .max_entries
+                .map(|max| max.saturating_sub(self.entries_charged.load(Ordering::Relaxed))),
+            max_entry_name_bytes: self
+                .max_entry_name_bytes
+                .map(|max| max.saturating_sub(self.entry_name_bytes_charged.load(Ordering::Relaxed))),
+            bucket_quantum: None,
        }
    }

@@ -234,6 +323,9 @@ impl ScannerCycleBudget {
    }

    pub(crate) fn record_entries_visited(&self, entries_visited: u64) {
+        if let Some(parent) = &self.parent {
+            parent.record_entries_visited(entries_visited);
+        }
        if self.track_progress {
            let entries = saturating_fetch_add(&self.entries_visited, entries_visited);
            self.record_progress_sample(entries);
@@ -245,7 +337,11 @@ impl ScannerCycleBudget {
    }

    pub(crate) fn record_remote_progress(&self, objects_scanned: u64, directories_started: u64) {
-        self.remote_scan_progress.record(objects_scanned, directories_started);
+        // A quantum shares the cycle's progress record; its parent writes it.
+        match &self.parent {
+            Some(parent) => parent.record_remote_progress(objects_scanned, directories_started),
+            None => self.remote_scan_progress.record(objects_scanned, directories_started),
+        }
        if self.track_progress || self.max_objects.is_some() {
            let objects = saturating_fetch_add(&self.objects_scanned, objects_scanned);
            if self.track_progress {
@@ -271,6 +367,9 @@ impl ScannerCycleBudget {
    }

    pub(crate) fn cancel_after_unreported_remote_progress(&self) {
+        if let Some(parent) = &self.parent {
+            parent.cancel_after_unreported_remote_progress();
+        }
        if self.max_objects.is_some() {
            self.cancel_for(ScannerCycleBudgetReason::Objects);
        } else if self.max_directories.is_some() {
@@ -279,6 +378,11 @@ impl ScannerCycleBudget {
    }

    pub(crate) fn try_start_directory(&self) -> bool {
+        let parent_allows = self.parent.as_ref().is_none_or(|parent| parent.try_start_directory());
+        self.try_start_own_directory() && parent_allows
+    }
+
+    fn try_start_own_directory(&self) -> bool {
        if self.max_directories.is_none() && !self.track_unbounded_counts {
            return true;
        }
@@ -298,7 +402,35 @@ impl ScannerCycleBudget {
        true
    }

+    /// Charges one raw entry first observed in this scan. `directory_entries`
+    /// counts the charged entries of the current directory: a count budget is
+    /// a total work limit, so one listing cannot read past it.
+    ///
+    /// Returns `false` after charging the entry that exhausts a limit; the
+    /// caller stops with the entry already recorded in its durable index.
+    pub(crate) fn try_charge_raw_entry(&self, name_bytes: u64, directory_entries: u64) -> bool {
+        let parent_allows = self
+            .parent
+            .as_ref()
+            .is_none_or(|parent| parent.try_charge_raw_entry(name_bytes, directory_entries));
+        let entries = self
+            .max_entries
+            .is_some_and(|max| saturating_fetch_add(&self.entries_charged, 1) >= max);
+        let name_bytes = self
+            .max_entry_name_bytes
+            .is_some_and(|max| saturating_fetch_add(&self.entry_name_bytes_charged, name_bytes) >= max);
+        let directory = self.directory_entry_limit.is_some_and(|limit| directory_entries >= limit);
+        if entries || name_bytes || directory {
+            self.cancel_for(ScannerCycleBudgetReason::Entries);
+            return false;
+        }
+        parent_allows
+    }
+
    pub(crate) fn record_object_scanned(&self) {
+        if let Some(parent) = &self.parent {
+            parent.record_object_scanned();
+        }
        if self.max_objects.is_none() && !self.track_unbounded_counts {
            return;
        }
@@ -337,6 +469,21 @@ fn saturating_fetch_add(value: &AtomicU64, delta: u64) -> u64 {
    }
 }

+fn directory_entry_limit(config: &ScannerCycleBudgetConfig) -> Option<u64> {
+    if config.max_objects.is_none() && config.max_directories.is_none() {
+        return None;
+    }
+    let total = config
+        .max_objects
+        .unwrap_or(0)
+        .saturating_add(config.max_directories.unwrap_or(0));
+    Some(
+        total
+            .saturating_mul(DIRECTORY_ENTRIES_PER_COUNT_UNIT)
+            .max(DIRECTORY_ENTRY_BUDGET_FLOOR),
+    )
+}
+
 fn directory_budget_exhausted(directories: u64, max_directories: u64) -> bool {
    // Saturation hides a remote max+1 update when the configured limit is the
    // largest representable counter. Treat that boundary as exhausted rather
@@ -354,6 +501,127 @@ impl Drop for ScannerCycleBudget {
 mod tests {
    use super::*;

+    #[tokio::test]
+    async fn fair_prefix_quantum_budget_limits_entries_and_name_bytes_with_one_reason() {
+        let parent = CancellationToken::new();
+        let by_entries = ScannerCycleBudget::new(
+            &parent,
+            ScannerCycleBudgetConfig {
+                max_entries: Some(2),
+                ..Default::default()
+            },
+        );
+        assert!(by_entries.try_charge_raw_entry(1, 1));
+        assert!(!by_entries.try_charge_raw_entry(1, 2));
+        assert_eq!(by_entries.reason(), Some(ScannerCycleBudgetReason::Entries));
+        assert!(by_entries.token().is_cancelled());
+
+        let by_bytes = ScannerCycleBudget::new(
+            &parent,
+            ScannerCycleBudgetConfig {
+                max_entry_name_bytes: Some(10),
+                ..Default::default()
+            },
+        );
+        assert!(by_bytes.try_charge_raw_entry(9, 1));
+        assert!(!by_bytes.try_charge_raw_entry(1, 2));
+        assert_eq!(by_bytes.reason(), Some(ScannerCycleBudgetReason::Entries));
+        assert!(!parent.is_cancelled(), "a budget stop must not cancel its parent token");
+    }
+
+    #[tokio::test]
+    async fn fair_prefix_quantum_budget_derives_a_directory_limit_only_from_count_budgets() {
+        let parent = CancellationToken::new();
+        let runtime_only = ScannerCycleBudget::new(&parent, ScannerCycleBudgetConfig::default());
+        assert!(runtime_only.try_charge_raw_entry(1, u64::MAX));
+
+        let small = ScannerCycleBudget::new(
+            &parent,
+            ScannerCycleBudgetConfig {
+                max_objects: Some(1),
+                ..Default::default()
+            },
+        );
+        assert!(small.try_charge_raw_entry(1, DIRECTORY_ENTRY_BUDGET_FLOOR - 1));
+        assert!(!small.try_charge_raw_entry(1, DIRECTORY_ENTRY_BUDGET_FLOOR));
+
+        let large = ScannerCycleBudget::new(
+            &parent,
+            ScannerCycleBudgetConfig {
+                max_objects: Some(1000),
+                max_directories: Some(24),
+                ..Default::default()
+            },
+        );
+        assert!(large.try_charge_raw_entry(1, 1024 * DIRECTORY_ENTRIES_PER_COUNT_UNIT - 1));
+        assert!(!large.try_charge_raw_entry(1, 1024 * DIRECTORY_ENTRIES_PER_COUNT_UNIT));
+    }
+
+    #[tokio::test]
+    async fn fair_prefix_quantum_bucket_share_stops_only_its_own_scan() {
+        let ctx = CancellationToken::new();
+        let cycle = ScannerCycleBudget::new_with_progress_tracking(
+            &ctx,
+            ScannerCycleBudgetConfig {
+                max_objects: Some(3),
+                bucket_quantum: Some(ScannerBucketQuantum {
+                    max_objects: Some(2),
+                    ..Default::default()
+                }),
+                ..Default::default()
+            },
+        );
+        let (first, first_token) = cycle.for_bucket(&cycle.token());
+        first.record_object_scanned();
+        first.record_object_scanned();
+        assert!(first.quantum_exhausted());
+        assert!(first_token.is_cancelled());
+        assert_eq!(first.reason(), Some(ScannerCycleBudgetReason::Objects));
+        assert!(!cycle.budget_elapsed(), "a bucket share leaves the cycle running");
+        assert!(!cycle.token().is_cancelled());
+        assert_eq!(cycle.progress().0, 2, "bucket work is charged to the cycle");
+
+        let (second, second_token) = cycle.for_bucket(&cycle.token());
+        assert!(!second.budget_elapsed());
+        second.record_object_scanned();
+        assert!(cycle.budget_elapsed(), "the total budget still bounds every share");
+        assert!(second_token.is_cancelled());
+        assert!(second.budget_elapsed());
+        assert!(!second.quantum_exhausted(), "a cycle stop is not a quantum handoff");
+    }
+
+    #[tokio::test]
+    async fn fair_prefix_quantum_bucket_share_runtime_expires_independently() {
+        let ctx = CancellationToken::new();
+        let cycle = ScannerCycleBudget::new(
+            &ctx,
+            ScannerCycleBudgetConfig {
+                bucket_quantum: Some(ScannerBucketQuantum {
+                    max_duration: Some(Duration::ZERO),
+                    ..Default::default()
+                }),
+                ..Default::default()
+            },
+        );
+        let (bucket, token) = cycle.for_bucket(&cycle.token());
+        tokio::time::timeout(Duration::from_secs(5), token.cancelled())
+            .await
+            .expect("quantum deadline should cancel the bucket scan");
+        assert_eq!(bucket.reason(), Some(ScannerCycleBudgetReason::Runtime));
+        assert!(!cycle.budget_elapsed());
+    }
+
+    #[tokio::test]
+    async fn bucket_without_a_quantum_shares_the_cycle_budget() {
+        let ctx = CancellationToken::new();
+        let cycle = ScannerCycleBudget::new(&ctx, ScannerCycleBudgetConfig::default());
+        let (bucket, token) = cycle.for_bucket(&cycle.token());
+        assert!(Arc::ptr_eq(&bucket, &cycle));
+        cycle.cancel_for_runtime();
+        assert!(token.is_cancelled());
+        assert!(!bucket.quantum_exhausted());
+    }
+
    #[tokio::test]
    async fn runtime_budget_cancels_child_token() {
        let parent = CancellationToken::new();
@@ -441,6 +709,7 @@ mod tests {
                max_duration: Some(Duration::from_secs(60)),
                max_objects: Some(10),
                max_directories: Some(5),
+                ..Default::default()
            },
        );

@@ -498,7 +498,10 @@ fn checkpoint_reason_from_budget(reason: Option<ScannerCycleBudgetReason>) -> Da
    match reason {
        Some(ScannerCycleBudgetReason::Runtime) => DataUsageScanCheckpointReason::Runtime,
        Some(ScannerCycleBudgetReason::Objects) => DataUsageScanCheckpointReason::Objects,
-        Some(ScannerCycleBudgetReason::Directories) => DataUsageScanCheckpointReason::Directories,
+        // This enum is persisted and older readers reject unknown variants.
+        Some(ScannerCycleBudgetReason::Directories | ScannerCycleBudgetReason::Entries) => {
+            DataUsageScanCheckpointReason::Directories
+        }
        None => DataUsageScanCheckpointReason::Unknown,
    }
 }
@@ -1302,9 +1305,11 @@ impl FolderScanner {
        }
    }

-    fn record_raw_enumeration_entry(&mut self, parent: &str, entry: &str) {
+    /// Records one raw entry. Returns whether a durable index holds it, which
+    /// is what makes stopping at this entry recoverable.
+    fn record_raw_enumeration_entry(&mut self, parent: &str, entry: &str) -> bool {
        if self.old_cache.info.scan_progress.is_none() {
-            return;
+            return false;
        }
        if let Some(position) = self
            .raw_enumeration_progress
@@ -1317,18 +1322,14 @@ impl FolderScanner {
            self.raw_enumeration_progress
                .push(RawEnumerationProgress::new(parent, page_index));
        }
-        if let Some(progress) = self.raw_enumeration_progress.last_mut() {
+        self.raw_enumeration_progress.last_mut().is_some_and(|progress| {
            progress.record_entry(entry);
-        }
+            progress.page_index.is_some()
+        })
    }

    fn raw_enumeration_page_index_for_parent(&self, parent: &str) -> Option<&RawEnumerationPageIndex> {
-        self.old_cache
-            .info
-            .scan_raw_enumeration_page_index
-            .as_ref()
-            .filter(|index| index.parent() == Some(parent))?;
-        self.old_cache.validated_raw_enumeration_page_index()
+        self.old_cache.validated_raw_enumeration_page_index_for_parent(parent)
    }

    fn raw_enumeration_committed_entry_oracle(&self, parent: &str) -> HashSet<String> {
@@ -1372,23 +1373,36 @@ impl FolderScanner {
        if self.raw_enumeration_progress.is_empty() {
            return (None, None);
        }
-        let progress_index = self
-            .raw_enumeration_progress
-            .iter()
-            .enumerate()
-            .max_by_key(|(index, progress)| (progress.checkpointable_entry_count(), std::cmp::Reverse(*index)))
-            .map(|(index, _)| index)
-            .unwrap_or(0);
+        let progress_index = self.selected_raw_enumeration_progress().unwrap_or(0);
        let progress = self.raw_enumeration_progress.swap_remove(progress_index);
        self.raw_enumeration_progress.clear();
        (progress.cursor(), progress.page_index())
    }

-    fn peek_raw_enumeration_resume_state(&self) -> (Option<DataUsageRawEnumerationCursor>, Option<RawEnumerationPageIndex>) {
+    fn selected_raw_enumeration_progress(&self) -> Option<usize> {
        self.raw_enumeration_progress
            .iter()
-            .max_by_key(|progress| progress.checkpointable_entry_count())
-            .map_or((None, None), |progress| (progress.cursor(), progress.page_index()))
+            .enumerate()
+            .max_by_key(|(index, progress)| (progress.checkpointable_entry_count(), std::cmp::Reverse(*index)))
+            .map(|(index, _)| index)
+    }
+
+    /// The walk path except the directory the resume state selects.
+    fn raw_enumeration_path_page_indexes(&self) -> Vec<RawEnumerationPageIndex> {
+        let selected = self.selected_raw_enumeration_progress();
+        self.raw_enumeration_progress
+            .iter()
+            .enumerate()
+            .filter(|(index, _)| Some(*index) != selected)
+            .filter_map(|(_, progress)| progress.page_index())
+            .collect()
+    }
+
+    fn peek_raw_enumeration_resume_state(&self) -> (Option<DataUsageRawEnumerationCursor>, Option<RawEnumerationPageIndex>) {
+        self.selected_raw_enumeration_progress().map_or((None, None), |index| {
+            let progress = &self.raw_enumeration_progress[index];
+            (progress.cursor(), progress.page_index())
+        })
    }

    async fn scan_child_folder(
@@ -1480,6 +1494,7 @@ impl FolderScanner {
        snapshot.info.last_update = Some(SystemTime::now());
        snapshot.info.snapshot_complete = false;
        let (cursor, page_index) = self.peek_raw_enumeration_resume_state();
+        snapshot.info.scan_raw_enumeration_path_page_indexes = self.raw_enumeration_path_page_indexes();
        snapshot.info.scan_raw_enumeration_cursor = cursor;
        snapshot.info.scan_raw_enumeration_page_index = page_index;
        snapshot.seal_scan_frontier(self.coverage_frontier.as_deref()).ok()?;
@@ -1976,6 +1991,9 @@ impl FolderScanner {
            let mut last_entry_progress = Instant::now();
            let mut raw_enumeration_complete = false;
            let raw_enumeration_committed_entries = self.raw_enumeration_committed_entry_oracle(&folder.name);
+            let raw_replay_floor = u64::try_from(raw_enumeration_committed_entries.len()).unwrap_or(u64::MAX);
+            let mut raw_entries_observed = 0_u64;
+            let mut raw_entries_charged = 0_u64;

            let mut candidate_unclassified = mode != FolderScanMode::Namespace;
            let mut deferred_metadata = None;
@@ -2061,8 +2079,13 @@ impl FolderScanner {
                    continue;
                }
                if !replayed_metadata {
-                    let raw_entry_consumed_by_owner_index = raw_enumeration_committed_entries.contains(&file_name);
-                    if !raw_entry_consumed_by_owner_index {
+                    // A restored index appends one pending name per entry past
+                    // its old size, so charging only that tail keeps each
+                    // budget unit durable even if the listing order changed.
+                    raw_entries_observed = raw_entries_observed.saturating_add(1);
+                    let raw_entry_is_new =
+                        raw_entries_observed > raw_replay_floor && !raw_enumeration_committed_entries.contains(&file_name);
+                    if raw_entry_is_new {
                        #[cfg(test)]
                        tests::enumeration_restart::observe_raw_entry(&dir_path, &entry.file_name(), &self.budget);
                        pending_entry_progress = pending_entry_progress.saturating_add(1);
@@ -2074,7 +2097,18 @@ impl FolderScanner {
                            last_entry_progress = Instant::now();
                        }
                    }
-                    self.record_raw_enumeration_entry(&folder.name, &file_name);
+                    let raw_entry_is_durable = self.record_raw_enumeration_entry(&folder.name, &file_name);
+                    // Object metadata and the entries behind it are charged to
+                    // the object budget; this one covers listing work.
+                    let object_entry = file_name == STORAGE_FORMAT_FILE || deferred_metadata.is_some();
+                    if raw_entry_is_new && raw_entry_is_durable && !object_entry {
+                        raw_entries_charged = raw_entries_charged.saturating_add(1);
+                        let name_bytes = u64::try_from(file_name.len()).unwrap_or(u64::MAX);
+                        if !self.budget.try_charge_raw_entry(name_bytes, raw_entries_charged) {
+                            self.budget.record_entries_visited(pending_entry_progress);
+                            return Err(ScannerError::Other("Operation cancelled".to_string()));
+                        }
+                    }
                    self.maybe_send_checkpoint(&folder, &this_hash, into);
                }
                let is_storage_format_entry = file_name == STORAGE_FORMAT_FILE;
@@ -2474,7 +2508,9 @@ impl FolderScanner {
                }
            }
            self.budget.record_entries_visited(pending_entry_progress);
-            if raw_enumeration_complete {
+            // A directory with queued children keeps its index until they end:
+            // a restart replays this listing to reach the unfinished child.
+            if raw_enumeration_complete && queued_folders.is_empty() {
                self.finish_raw_enumeration_parent(&folder.name);
            }

@@ -2683,10 +2719,12 @@ impl FolderScanner {
                                "Scanner child folder scan failed"
                            );
                            self.coverage_gap |= forward_sweep;
+                            self.finish_raw_enumeration_parent(&folder_item.name);
                            continue;
                        }
                    }
                };
+                self.finish_raw_enumeration_parent(&folder_item.name);
                if outcome == FolderScanOutcome::Vanished {
                    self.new_cache.delete_recursive(&h);
                    if let Some(updates) = &mut self.usage_updates {
@@ -3405,6 +3443,7 @@ pub(crate) async fn scan_data_folder_scoped(
            new_cache.info.scan_checkpoint = None;
            new_cache.info.scan_raw_enumeration_cursor = None;
            new_cache.info.scan_raw_enumeration_page_index = None;
+            new_cache.info.scan_raw_enumeration_path_page_indexes.clear();
            new_cache.info.scan_coverage_receipt = None;
            if had_scan_checkpoint {
                global_metrics().record_scanner_checkpoint_cleared();
@@ -3443,6 +3482,7 @@ pub(crate) async fn scan_data_folder_scoped(
                let root_hash = hash_path(&bucket_name);
                let root_has_progress = data_usage_root_has_progress(&root);
                let pending_heals_changed = scanner.pending_heals_changed;
+                let raw_enumeration_path_page_indexes = scanner.raw_enumeration_path_page_indexes();
                let (raw_enumeration_cursor, raw_enumeration_page_index) = scanner.take_raw_enumeration_resume_state();
                let carry_forward_cache = ((raw_enumeration_cursor.is_some() || raw_enumeration_page_index.is_some())
                    && !root_has_progress)
@@ -3465,6 +3505,7 @@ pub(crate) async fn scan_data_folder_scoped(
                }
                if raw_enumeration_page_index.is_some() {
                    new_cache.info.scan_raw_enumeration_page_index = raw_enumeration_page_index;
+                    new_cache.info.scan_raw_enumeration_path_page_indexes = raw_enumeration_path_page_indexes;
                    new_cache.info.scan_checkpoint = None;
                    new_cache.info.scan_resume_after = None;
                    new_cache.info.scan_coverage_receipt = None;
@@ -3495,6 +3536,7 @@ pub(crate) async fn scan_data_folder_scoped(
                            global_metrics().record_scanner_checkpoint_cleared();
                        }
                    }
+                    new_cache.info.scan_resource_deferred = matches!(&e, ScannerError::ResourceDeferred);
                    close_disk_guard.close().await;
                    return Err(ScannerError::PartialCache(Box::new(std::mem::take(new_cache))));
                }
@@ -3524,6 +3566,7 @@ pub(crate) async fn scan_data_folder_scoped(
                        .insert(hash_path(&bucket_name).0, DataUsageEntry::default());
                }
                partial_cache.info.snapshot_complete = false;
+                partial_cache.info.scan_resource_deferred = true;
                close_disk_guard.close().await;
                return Err(ScannerError::PartialCache(Box::new(partial_cache)));
            }
@@ -30,6 +30,7 @@ use std::sync::Mutex;

 pub(super) mod checkpoint_fixture;
 pub(super) mod enumeration_restart;
+mod fair_prefix_quantum;
 mod incremental_enumeration;
 mod listing_completion;
 mod metadata_visits;
@@ -51,6 +51,10 @@ impl Drop for ObservationGuard {
    }
 }

+pub(in crate::scanner_folder) fn install_raw_entry_observer(root: PathBuf) -> ObservationGuard {
+    install_raw_entry_budget(root, u64::MAX)
+}
+
 pub(in crate::scanner_folder) fn install_raw_entry_budget(root: PathBuf, limit: u64) -> ObservationGuard {
    *OBSERVATION.lock().expect("install raw-entry observation") = Some(Observation {
        root,
@@ -83,6 +87,45 @@ pub(super) async fn read_bounded(path: &Path) -> Vec<u8> {
    bytes
 }

+/// Creates a flat synthetic bucket and the checkpointable cache a first sweep
+/// starts from.
+pub(super) async fn seed_flat_bucket(disk_root: &Path, objects: usize) -> DataUsageCache {
+    tokio::fs::create_dir(disk_root).await.expect("create fresh synthetic disk");
+    for index in 0..objects {
+        let object = format!("object-{index:04}");
+        let version = Uuid::from_u128(u128::try_from(index).expect("fixture index") + 1);
+        let bytes = metadata_for_object_version("bucket", &object, Some(version));
+        write_test_object_metadata_bytes(disk_root, "bucket", &object, &bytes).await;
+    }
+    let mut initial = DataUsageCache::default();
+    initial.info.name = "bucket".to_string();
+    let source = crate::data_usage_define::DataUsageCacheSource::new(0, 0);
+    let plan = crate::data_usage_define::DataUsageScanPlanDigest([31; 32]);
+    let identity = crate::data_usage_define::DataUsageScanIdentity {
+        version: 1,
+        bucket_incarnation: Uuid::from_u128(31),
+        set_layout: crate::data_usage_define::DataUsageScanPlanDigest([32; 32]),
+        publication_epoch: 1,
+        tier_registry_generation: 0,
+        scan_mode: HealScanMode::Normal,
+    };
+    assert_eq!(
+        initial.prepare_bucket_checkpoint("bucket", 1, 0, source, plan, identity),
+        crate::data_usage_define::DataUsageCachePrepareOutcome::Reset
+    );
+    initial.info.skip_healing = true;
+    initial.replace("bucket", "", DataUsageEntry::default());
+    initial
+}
+
+pub(super) fn observed_raw_entries() -> u64 {
+    OBSERVATION
+        .lock()
+        .expect("read raw-entry observation")
+        .as_ref()
+        .map_or(0, |observation| observation.entries)
+}
+
 async fn round(request: &Request) -> serde_json::Value {
    assert!((1..=1024).contains(&request.objects));
    assert!((1..=4096).contains(&request.raw_entry_budget));
@@ -90,31 +133,7 @@ async fn round(request: &Request) -> serde_json::Value {
    let disk_root = request.workspace.join("disk");
    let cache_path = request.workspace.join("cache.bin");
    if request.round == 0 {
-        tokio::fs::create_dir(&disk_root).await.expect("create fresh synthetic disk");
-        for index in 0..request.objects {
-            let object = format!("object-{index:04}");
-            let version = Uuid::from_u128(u128::try_from(index).expect("fixture index") + 1);
-            let bytes = metadata_for_object_version("bucket", &object, Some(version));
-            write_test_object_metadata_bytes(&disk_root, "bucket", &object, &bytes).await;
-        }
-        let mut initial = DataUsageCache::default();
-        initial.info.name = "bucket".to_string();
-        let source = crate::data_usage_define::DataUsageCacheSource::new(0, 0);
-        let plan = crate::data_usage_define::DataUsageScanPlanDigest([31; 32]);
-        let identity = crate::data_usage_define::DataUsageScanIdentity {
-            version: 1,
-            bucket_incarnation: Uuid::from_u128(31),
-            set_layout: crate::data_usage_define::DataUsageScanPlanDigest([32; 32]),
-            publication_epoch: 1,
-            tier_registry_generation: 0,
-            scan_mode: HealScanMode::Normal,
-        };
-        assert_eq!(
-            initial.prepare_bucket_checkpoint("bucket", 1, 0, source, plan, identity),
-            crate::data_usage_define::DataUsageCachePrepareOutcome::Reset
-        );
-        initial.info.skip_healing = true;
-        initial.replace("bucket", "", DataUsageEntry::default());
+        let initial = seed_flat_bucket(&disk_root, request.objects).await;
        tokio::fs::write(&cache_path, initial.marshal_msg().expect("initial cache codec"))
            .await
            .expect("persist initial cache");
@@ -137,10 +156,11 @@ async fn round(request: &Request) -> serde_json::Value {
        &parent,
        crate::scanner_budget::ScannerCycleBudgetConfig {
            max_objects: Some(request.raw_entry_budget),
+            max_entries: Some(request.raw_entry_budget),
            ..Default::default()
        },
    );
-    let _observation_guard = install_raw_entry_budget(disk.path(), request.raw_entry_budget);
+    let _observation_guard = install_raw_entry_observer(disk.path());
    let result = scan_data_folder(
        budget.token(),
        budget.clone(),
@@ -0,0 +1,233 @@
+// Copyright 2026 RustFS Team
+// Licensed under the Apache License, Version 2.0.
+
+use super::enumeration_restart::{install_raw_entry_observer, observed_raw_entries, seed_flat_bucket};
+use super::*;
+use crate::scanner_budget::ScannerCycleBudgetConfig;
+
+async fn open_fixture_disk(disk_root: &std::path::Path) -> Arc<Disk> {
+    let endpoint = Endpoint::try_from(disk_root.to_string_lossy().as_ref()).expect("fixture endpoint");
+    new_disk(
+        &endpoint,
+        &DiskOption {
+            cleanup: false,
+            health_check: false,
+        },
+    )
+    .await
+    .expect("open synthetic disk")
+}
+
+/// A count budget is the operator's total work limit. It must also stop the
+/// first listing of a directory far wider than that budget, instead of letting
+/// one read_dir pass finish before the first object is charged.
+#[tokio::test]
+#[serial]
+async fn fair_prefix_quantum_count_budget_bounds_the_first_directory_build() {
+    const OBJECTS: usize = 300;
+    const FIRST_BUILD_FLOOR: u64 = 128;
+    let temp = tempfile::tempdir().expect("fixture directory");
+    let disk_root = temp.path().join("disk");
+    let cache = seed_flat_bucket(&disk_root, OBJECTS).await;
+    let disk = open_fixture_disk(&disk_root).await;
+    let parent = CancellationToken::new();
+    let budget = ScannerCycleBudget::new_with_progress_tracking(
+        &parent,
+        ScannerCycleBudgetConfig {
+            max_objects: Some(8),
+            ..Default::default()
+        },
+    );
+    let _observation = install_raw_entry_observer(disk.path());
+
+    let result = scan_data_folder(
+        budget.token(),
+        budget.clone(),
+        vec![disk.clone()],
+        disk,
+        cache,
+        None,
+        HealScanMode::Normal,
+        SCANNER_SLEEPER.clone(),
+    )
+    .await;
+
+    assert!(
+        matches!(result, Err(ScannerError::PartialCache(_))),
+        "an exhausted first build must return recoverable progress"
+    );
+    assert!(
+        observed_raw_entries() <= FIRST_BUILD_FLOOR,
+        "first build read {} raw entries under an 8-object budget",
+        observed_raw_entries()
+    );
+}
+
+struct RestartRound {
+    outcome: DataUsageCache,
+    finished: bool,
+    observed_entries: u64,
+}
+
+/// One scanner lifetime: the input was decoded from persisted bytes, and the
+/// result is encoded and decoded again before the next round may use it.
+async fn scan_one_restart_round(
+    disk_root: &std::path::Path,
+    cache: DataUsageCache,
+    config: ScannerCycleBudgetConfig,
+) -> RestartRound {
+    let disk = open_fixture_disk(disk_root).await;
+    let parent = CancellationToken::new();
+    let budget = ScannerCycleBudget::new_with_progress_tracking(&parent, config);
+    let _observation = install_raw_entry_observer(disk.path());
+    let result = scan_data_folder(
+        budget.token(),
+        budget.clone(),
+        vec![disk.clone()],
+        disk,
+        cache.clone(),
+        None,
+        HealScanMode::Normal,
+        SCANNER_SLEEPER.clone(),
+    )
+    .await;
+    let observed_entries = observed_raw_entries();
+    let (returned, finished) = match result {
+        Ok(cache) => (cache, true),
+        Err(ScannerError::PartialCache(cache)) => (*cache, false),
+        Err(error) => panic!("a budget stop must keep recoverable progress: {error}"),
+    };
+    let encoded = returned.marshal_msg().expect("encode returned cache");
+    RestartRound {
+        outcome: DataUsageCache::unmarshal(&encoded).expect("decode returned cache"),
+        finished,
+        observed_entries,
+    }
+}
+
+async fn run_until_finished(
+    disk_root: &std::path::Path,
+    mut cache: DataUsageCache,
+    config: ScannerCycleBudgetConfig,
+    rounds: usize,
+) -> DataUsageCache {
+    for round in 0..rounds {
+        let result = scan_one_restart_round(disk_root, cache, config).await;
+        if let Some(limit) = config.max_entries {
+            assert!(result.observed_entries <= limit, "round {round} read {} entries", result.observed_entries);
+        }
+        cache = result.outcome;
+        if result.finished {
+            assert!(cache.info.snapshot_complete, "a finished sweep must certify its bucket");
+            return cache;
+        }
+    }
+    panic!("the sweep did not finish in {rounds} restarts");
+}
+
+/// Both a wide bucket root and a wide nested prefix must keep their indexes:
+/// losing either one recharges its replay and a fixed budget stops advancing.
+#[tokio::test]
+#[serial]
+async fn fair_prefix_quantum_nested_wide_directories_converge_across_restarts() {
+    const WIDTH: usize = 40;
+    let temp = tempfile::tempdir().expect("fixture directory");
+    let disk_root = temp.path().join("disk");
+    let cache = seed_flat_bucket(&disk_root, WIDTH).await;
+    for index in 0..WIDTH {
+        let object = format!("wide/nested-{index:04}");
+        let version = Uuid::from_u128(u128::try_from(index).expect("fixture index") + 1000);
+        let bytes = metadata_for_object_version("bucket", &object, Some(version));
+        write_test_object_metadata_bytes(&disk_root, "bucket", &object, &bytes).await;
+    }
+    let config = ScannerCycleBudgetConfig {
+        max_objects: Some(8),
+        max_entries: Some(8),
+        ..Default::default()
+    };
+    // 41 root and 40 prefix entries at 8 per round, then 80 objects at the 7
+    // that survive each object-budget stop. Rebuilding a dropped index costs
+    // a full prefix enumeration per classification round and overruns this.
+    const ROUND_BOUND: usize = 6 + 5 + 12 + 2;
+    let finished = run_until_finished(&disk_root, cache, config, ROUND_BOUND).await;
+    let totals = finished.checked_flatten("bucket").expect("bucket root");
+    let expected = u64::try_from(WIDTH * 2).expect("fixture size");
+    assert_eq!((totals.objects as u64, totals.versions as u64), (expected, expected));
+    assert!(finished.info.scan_raw_enumeration_page_index.is_none());
+    assert!(finished.info.scan_raw_enumeration_path_page_indexes.is_empty());
+}
+
+/// A name-byte share stops a listing the same way an entry share does, and the
+/// bytes kept for the next restart grow only by what each round charged.
+#[tokio::test]
+#[serial]
+async fn fair_prefix_quantum_name_bytes_bound_each_round_and_its_checkpoint_growth() {
+    const OBJECTS: usize = 64;
+    const NAME_BYTES_PER_ROUND: u64 = 88;
+    const INDEX_BYTES_PER_NAME_BYTE: usize = 16;
+    let temp = tempfile::tempdir().expect("fixture directory");
+    let disk_root = temp.path().join("disk");
+    let mut cache = seed_flat_bucket(&disk_root, OBJECTS).await;
+    let config = ScannerCycleBudgetConfig {
+        max_entry_name_bytes: Some(NAME_BYTES_PER_ROUND),
+        ..Default::default()
+    };
+    let mut indexed = 0;
+    let mut encoded_before = cache.marshal_msg().expect("encode initial cache").len();
+    for _ in 0..8 {
+        let result = scan_one_restart_round(&disk_root, cache, config).await;
+        assert!(!result.finished);
+        assert_eq!(result.observed_entries, 8, "each entry name in this fixture is 11 bytes");
+        cache = result.outcome;
+        let committed = cache
+            .validated_raw_enumeration_page_index()
+            .expect("durable index")
+            .committed_entries()
+            .expect("committed entries")
+            .len();
+        assert_eq!(committed, indexed + 8, "every charged entry must become durable");
+        indexed = committed;
+        let encoded = cache.marshal_msg().expect("encode cache").len();
+        let growth = encoded.saturating_sub(encoded_before);
+        assert!(
+            growth <= usize::try_from(NAME_BYTES_PER_ROUND).expect("fixture size") * INDEX_BYTES_PER_NAME_BYTE,
+            "checkpoint grew {growth} bytes in one round"
+        );
+        encoded_before = encoded;
+    }
+}
+
+/// Stopping for an entry share must leave the same exact totals a full sweep
+/// produces, and a Deep request must not inherit Normal raw coverage.
+#[tokio::test]
+#[serial]
+async fn fair_prefix_quantum_budget_stop_keeps_progress_without_certifying_it() {
+    const OBJECTS: usize = 24;
+    let temp = tempfile::tempdir().expect("fixture directory");
+    let disk_root = temp.path().join("disk");
+    let cache = seed_flat_bucket(&disk_root, OBJECTS).await;
+    let config = ScannerCycleBudgetConfig {
+        max_objects: Some(4),
+        max_entries: Some(4),
+        ..Default::default()
+    };
+    let round = scan_one_restart_round(&disk_root, cache, config).await;
+    assert!(!round.finished);
+    assert!(!round.outcome.info.snapshot_complete, "budget stops are progress-only");
+    assert!(round.outcome.info.scan_plan_digest.is_none(), "a partial sweep must not be reusable");
+    assert!(round.outcome.has_scan_position(), "the saved cache must carry the resume position");
+
+    let mut deep = round.outcome.clone();
+    let mut identity = deep.info.scan_identity.expect("checkpoint identity");
+    identity.scan_mode = HealScanMode::Deep;
+    let plan = crate::data_usage_define::DataUsageScanPlanDigest([31; 32]);
+    let source = crate::data_usage_define::DataUsageCacheSource::new(0, 0);
+    deep.prepare_bucket_checkpoint("bucket", 2, 0, source, plan, identity);
+    assert!(!deep.has_scan_position(), "Normal raw coverage must not carry into a Deep sweep");
+    assert!(deep.info.scan_raw_enumeration_path_page_indexes.is_empty());
+
+    let finished = run_until_finished(&disk_root, round.outcome, config, 16).await;
+    let totals = finished.checked_flatten("bucket").expect("bucket root");
+    let expected = u64::try_from(OBJECTS).expect("fixture size");
+    assert_eq!((totals.objects as u64, totals.versions as u64), (expected, expected));
+}
@@ -497,6 +497,7 @@ async fn scanner_node_admission_mid_walk_shrink_preserves_partial_and_resumes_al
        panic!("deferral below root cannot certify completion: {partial:?}");
    };
    assert!(!partial.info.snapshot_complete);
+    assert!(partial.info.scan_resource_deferred);
    assert_eq!(
        partial
            .size_recursive("bucket")
@@ -610,6 +611,7 @@ async fn check_early_admission_deferral(node_open: bool, physical_limit: u32) {
        panic!("resource pressure is a partial census, not a disk error: {result:?}");
    };
    assert!(!partial.info.snapshot_complete);
+    assert!(partial.info.scan_resource_deferred);
    let totals = partial.size_recursive("bucket").expect("old tree retained");
    assert_eq!((totals.objects, totals.size), (7, 19));
    assert_eq!(gate.completed(), Some(0));
@@ -18,9 +18,9 @@ use crate::scanner_budget::ScannerCycleBudget;
 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,
-    DataUsageCacheSource, DataUsageEntry, DataUsageEntryInfo, DataUsageInfo, DataUsageScanPlanDigest, DataUsageSnapshotSetState,
-    ScannerError, SizeSummary, TierStats,
+    DATA_USAGE_CACHE_NAME, DATA_USAGE_ROOT, DataUsageBucketServiceLedger, DataUsageCache, DataUsageCacheInfo,
+    DataUsageCachePrepareOutcome, DataUsageCacheSource, DataUsageEntry, DataUsageEntryInfo, DataUsageInfo,
+    DataUsageScanPlanDigest, DataUsageSnapshotSetState, ScannerError, SizeSummary, TierStats,
 };
 use bytes::Bytes;
 use futures::future::join_all;
@@ -1200,6 +1200,84 @@ fn log_missing_set_snapshot_metadata(store: &SetDisks, info: &DataUsageCacheInfo
    );
 }

+/// Saves a set's service order over its unchanged cache, which an incomplete
+/// cycle otherwise never writes. Under the set cache lock and its revisions the
+/// save cannot replace newer coverage; a failed save costs fairness only.
+pub(super) async fn persist_scanner_service_ledger(
+    store: Arc<SetDisks>,
+    source: DataUsageCacheSource,
+    mut ledger: DataUsageBucketServiceLedger,
+    cycle: u64,
+    leader_epoch: u64,
+    expected_publication_epoch: u64,
+) -> bool {
+    let log_skipped = |state: &str, error: &dyn std::fmt::Display| {
+        warn!(
+            target: "rustfs::scanner::io",
+            event = EVENT_SCANNER_CACHE_PERSIST_STATE,
+            component = LOG_COMPONENT_SCANNER,
+            subsystem = LOG_SUBSYSTEM_IO,
+            state,
+            pool = source.pool_index,
+            set = source.set_index,
+            cycle,
+            leader_epoch,
+            cache_name = DATA_USAGE_CACHE_NAME,
+            error = %error,
+            "Scanner kept the previous bucket service order"
+        );
+    };
+    let guard = match acquire_scanner_cache_locks(store.as_ref(), DATA_USAGE_CACHE_NAME, source).await {
+        Ok(guard) => guard,
+        Err(error) => {
+            log_skipped(error.state(), &error);
+            return false;
+        }
+    };
+    let mut persisted = DataUsageCache::default();
+    let revisions = match persisted.load_with_revisions(store.clone(), DATA_USAGE_CACHE_NAME).await {
+        Ok(revisions) => revisions,
+        Err(error) => {
+            log_skipped("service_order_load_failed", &error);
+            return false;
+        }
+    };
+    if persisted.info.next_cycle > cycle
+        || persisted.info.leader_epoch > leader_epoch
+        || persisted
+            .info
+            .scan_bucket_service
+            .as_ref()
+            .is_some_and(|saved| saved.leader_epoch > leader_epoch)
+    {
+        log_skipped("service_order_stale", &"a newer scanner owns the set cache");
+        return false;
+    }
+    if guard.is_lock_lost() {
+        log_skipped("service_order_lock_lost", &"set cache lock was lost");
+        return false;
+    }
+    if persisted.info.name.is_empty() {
+        // First incomplete cycle of a set: the order is its only content.
+        persisted.info.name = DATA_USAGE_ROOT.to_string();
+        persisted.info.source = Some(source);
+        persisted.info.leader_epoch = leader_epoch;
+        persisted.info.cache_key_format = DATA_USAGE_CACHE_KEY_FORMAT;
+    }
+    ledger.leader_epoch = leader_epoch;
+    persisted.info.scan_bucket_service = Some(ledger);
+    match persisted
+        .save_with_revisions_for_epoch(store, DATA_USAGE_CACHE_NAME, &revisions, expected_publication_epoch)
+        .await
+    {
+        Ok(()) => true,
+        Err(error) => {
+            log_skipped("service_order_save_failed", &error);
+            false
+        }
+    }
+}
+
 pub(super) async fn persist_and_publish_cache_snapshot(
    store: Arc<SetDisks>,
    updates: &mpsc::Sender<DataUsageCache>,
@@ -1638,6 +1638,7 @@ pub(super) async fn requeue_bucket_work(
    bucket: &BucketInfo,
    work_guard: &mut BucketWorkGuard,
 ) -> bool {
+    work_guard.settle_service_for_requeue();
    if bucket_tx.send(bucket.clone()).await.is_err() {
        return false;
    }
@@ -307,6 +307,33 @@ impl ScannerServiceCohort {
    }
 }

+/// Publishes one aggregate per set: each failure is already logged where it
+/// happened, and this keeps the unpaid total visible after a restart.
+pub(super) fn record_bucket_failure_debt(source: DataUsageCacheSource, cycle: u64, ledger: &DataUsageBucketServiceLedger) {
+    let (debtors, first_bucket) = ledger.failure_debt();
+    metrics::gauge!(
+        "rustfs_scanner_bucket_failure_debt",
+        "pool" => source.pool_index.to_string(),
+        "set" => source.set_index.to_string()
+    )
+    .set(debtors as f64);
+    if debtors > 0 {
+        warn!(
+            target: "rustfs::scanner::io",
+            event = EVENT_SCANNER_SET_STATE,
+            component = LOG_COMPONENT_SCANNER,
+            subsystem = LOG_SUBSYSTEM_IO,
+            state = "bucket_failure_debt",
+            pool = source.pool_index,
+            set = source.set_index,
+            cycle,
+            debtors,
+            first_bucket,
+            "Scanner buckets still owe a successful sweep"
+        );
+    }
+}
+
 pub(super) async fn wait_for_bucket_scan_permit(
    semaphore: &Arc<Semaphore>,
    ctx: &CancellationToken,
@@ -408,10 +435,102 @@ pub(super) struct DiskBucketScanActiveGuard {
    set: String,
 }

+#[derive(Default)]
+struct ScannerBucketServiceRoundState {
+    ledger: DataUsageBucketServiceLedger,
+    claimed: HashMap<String, u64>,
+    interrupted: HashSet<String>,
+}
+
+/// One cycle's view of a set's durable service order. Lock order: taken alone,
+/// and never held across an await.
+pub(super) struct ScannerBucketServiceRound(StdMutex<ScannerBucketServiceRoundState>);
+
+impl ScannerBucketServiceRound {
+    pub(super) fn new(ledger: DataUsageBucketServiceLedger) -> Self {
+        Self(StdMutex::new(ScannerBucketServiceRoundState {
+            ledger,
+            ..Default::default()
+        }))
+    }
+
+    fn state(&self) -> std::sync::MutexGuard<'_, ScannerBucketServiceRoundState> {
+        self.0.lock().unwrap_or_else(|poisoned| poisoned.into_inner())
+    }
+
+    pub(super) fn order_buckets(&self, buckets: &mut [BucketInfo]) {
+        // The stable sort keeps the caller's order for equal positions.
+        let state = self.state();
+        buckets.sort_by_key(|bucket| state.ledger.wait_position(&bucket.name));
+    }
+
+    fn claim(&self, bucket: &str) {
+        let mut state = self.state();
+        if let Some(previous) = state.ledger.record_service(bucket) {
+            // A second claim in one cycle keeps the position before the first.
+            state.claimed.entry(bucket.to_owned()).or_insert(previous);
+        }
+    }
+
+    fn release(&self, bucket: &str) {
+        let mut state = self.state();
+        if let Some(previous) = state.claimed.remove(bucket) {
+            state.ledger.return_service(bucket, previous);
+        }
+    }
+
+    fn interrupt(&self, bucket: &str) {
+        self.state().interrupted.insert(bucket.to_owned());
+    }
+
+    /// Closes the cycle. Only a real worker completion clears debt; a claimed
+    /// bucket that failed or stayed partial adds debt unless a budget stopped
+    /// it. `None` means no bucket was served, so nothing needs persisting.
+    pub(super) fn settle(
+        &self,
+        completed: &HashSet<String>,
+        failed: &HashSet<String>,
+        partial: &HashSet<String>,
+    ) -> Option<DataUsageBucketServiceLedger> {
+        let mut state = self.state();
+        let state = &mut *state;
+        if state.claimed.is_empty() {
+            return None;
+        }
+        for bucket in state.claimed.keys() {
+            if completed.contains(bucket) {
+                state.ledger.record_certified(bucket);
+            } else if (failed.contains(bucket) || partial.contains(bucket)) && !state.interrupted.contains(bucket) {
+                state.ledger.record_failure(bucket);
+            }
+        }
+        state.claimed.clear();
+        Some(std::mem::take(&mut state.ledger))
+    }
+}
+
+struct BucketServiceClaim {
+    round: Arc<ScannerBucketServiceRound>,
+    bucket: String,
+    ctx: CancellationToken,
+    budget: Arc<ScannerCycleBudget>,
+}
+
+impl BucketServiceClaim {
+    fn settle(self, requeued: bool) {
+        if self.budget.budget_elapsed() || self.ctx.is_cancelled() {
+            self.round.interrupt(&self.bucket);
+        } else if requeued {
+            self.round.release(&self.bucket);
+        }
+    }
+}
+
 pub(super) struct BucketWorkGuard {
    remaining: Arc<AtomicUsize>,
    complete: CancellationToken,
    requeued: bool,
+    service: Option<BucketServiceClaim>,
 }

 impl BucketWorkGuard {
@@ -420,16 +539,53 @@ impl BucketWorkGuard {
            remaining,
            complete,
            requeued: false,
+            service: None,
        }
    }

    pub(super) fn mark_requeued(&mut self) {
        self.requeued = true;
    }
+
+    /// Settles the claim before the bucket is visible to another worker, whose
+    /// own claim a later release would otherwise undo.
+    pub(super) fn settle_service_for_requeue(&mut self) {
+        if let Some(claim) = self.service.take() {
+            claim.settle(true);
+        }
+    }
+
+    /// Records a stop that is neither service nor failure, such as a resource veto.
+    pub(super) fn mark_service_interrupted(&self) {
+        if let Some(claim) = &self.service {
+            claim.round.interrupt(&claim.bucket);
+        }
+    }
+
+    /// Spends this bucket's service turn. The guard settles it at the end of
+    /// the admission: a requeue without a spent budget hands the turn back.
+    pub(super) fn claim_service(
+        &mut self,
+        round: Arc<ScannerBucketServiceRound>,
+        bucket: &str,
+        ctx: CancellationToken,
+        budget: Arc<ScannerCycleBudget>,
+    ) {
+        round.claim(bucket);
+        self.service = Some(BucketServiceClaim {
+            round,
+            bucket: bucket.to_owned(),
+            ctx,
+            budget,
+        });
+    }
 }

 impl Drop for BucketWorkGuard {
    fn drop(&mut self) {
+        if let Some(claim) = self.service.take() {
+            claim.settle(self.requeued);
+        }
        if !self.requeued && self.remaining.fetch_sub(1, Ordering::AcqRel) == 1 {
            self.complete.cancel();
        }
@@ -348,10 +348,21 @@ impl ScannerIOCache for SetDisks {
                None
            }
        };
+        let persisted_bucket_service = old_cache
+            .info
+            .scan_bucket_service
+            .take()
+            .filter(|_| old_cache.info.source == Some(source));
        // Captured before `prepare_for_scan` clears the mark: only a complete
        // snapshot may stand in for a bucket whose walk stops early.
        let old_cache_certified = old_cache.info.snapshot_complete && old_cache.info.last_update.is_some();
        let current_bucket_incarnations = scanner_current_bucket_incarnations(self.as_ref(), &all_buckets).await;
+        let service_round = Arc::new(ScannerBucketServiceRound::new(DataUsageBucketServiceLedger::restore(
+            clone_cache_value(&persisted_bucket_service),
+            expected_publication_epoch,
+            all_buckets.iter().map(|bucket| bucket.name.as_str()),
+            current_bucket_incarnations.as_ref(),
+        )));
        let scoped_scan = prepare_scoped_set_scan(
            &old_cache,
            &buckets,
@@ -412,6 +423,7 @@ impl ScannerIOCache for SetDisks {
            };
            cache.info.last_update = Some(now);
            cache.info.snapshot_complete = true;
+            cache.info.scan_bucket_service = persisted_bucket_service;
            let replication_execution = super::cache::scanner_replication_execution_digest(execution_digest, &cache);
            cache.info.scan_execution_digest = Some(replication_execution);
            cache.info.segment_invalidation_proof = completed_segment_invalidation_proof;
@@ -753,6 +765,9 @@ impl ScannerIOCache for SetDisks {
                .unwrap_or_else(|poisoned| poisoned.into_inner())
                .order_buckets(source, &mut scan_order);
        }
+        // Durable age outranks both orders above; they only break its ties,
+        // so a restart cannot hand the budget back to the same first buckets.
+        service_round.order_buckets(&mut scan_order);

        for bucket in scan_order.iter() {
            if let Some(c) = old_cache.find(&bucket.name) {
@@ -790,6 +805,7 @@ impl ScannerIOCache for SetDisks {
        } in walkers
        {
            let service_cohort_clone = service_cohort.clone();
+            let service_round_clone = service_round.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();
@@ -874,6 +890,13 @@ impl ScannerIOCache for SetDisks {
                            .unwrap_or_else(|poisoned| poisoned.into_inner())
                            .record_admitted(source, &bucket.name);
                    }
+                    let (bucket_budget, scan_ctx) = budget_clone.for_bucket(&ctx_clone);
+                    work_guard.claim_service(
+                        service_round_clone.clone(),
+                        &bucket.name,
+                        ctx_clone.clone(),
+                        bucket_budget.clone(),
+                    );

                    debug!(
                        target: "rustfs::scanner::io",
@@ -915,8 +938,8 @@ impl ScannerIOCache for SetDisks {
                        remote_session_sequence = next_sequence;
                        let remote_scan = crate::remote_scanner::scan_remote_bucket(
                            &disk,
-                            ctx_clone.clone(),
-                            budget_clone.clone(),
+                            scan_ctx.clone(),
+                            bucket_budget.clone(),
                            crate::remote_scanner::RemoteScannerScanSpec {
                                bucket: &bucket.name,
                                next_cycle: want_cycle,
@@ -959,6 +982,11 @@ impl ScannerIOCache for SetDisks {
                                }
                            }
                            Ok(crate::remote_scanner::RemoteScannerOutcome::Partial) => {
+                                // A remote count budget stops on entries it does
+                                // not report back, so the partial proves no failure.
+                                if bucket_budget.max_objects().is_some() || bucket_budget.max_directories().is_some() {
+                                    work_guard.mark_service_interrupted();
+                                }
                                record_partial_dirty_bucket(&partial_dirty_buckets_clone, &bucket.name).await;
                                debug!(
                                    target: "rustfs::scanner::io",
@@ -1096,6 +1124,11 @@ impl ScannerIOCache for SetDisks {
                    };
                    #[cfg(test)]
                    super::tests::observe_coordinator_bucket_scan_admitted(worker_index, &bucket.name).await;
+                    #[cfg(test)]
+                    if super::tests::coordinator_bucket_scan_fails(&bucket.name) {
+                        record_failed_dirty_bucket(&failed_dirty_buckets_clone, &bucket.name).await;
+                        continue;
+                    }

                    // Lock order: scanner leader fence -> set-scoped per-bucket cache lock ->
                    // cache object read/write.
@@ -1331,7 +1364,7 @@ impl ScannerIOCache for SetDisks {
                    let (checkpoint_tx, mut checkpoint_rx) = mpsc::channel::<DataUsageCache>(1);
                    let (node_tx, mut node_rx) = crate::workload_admission::scanner_node_request_channel();

-                    let scan_ctx = ctx_clone.child_token();
+                    let scan_position_before = cache.durable_scan_position();
                    #[cfg(feature = "e2e-test-hooks")]
                    {
                        let disk_uuid = if crate::e2e_barrier::worker_observation::enabled() {
@@ -1351,7 +1384,7 @@ impl ScannerIOCache for SetDisks {
                    }
                    let scan = crate::workload_admission::with_scanner_node_requests(node_tx, disk.clone().nsscanner_disk(
                        scan_ctx.clone(),
-                        budget_clone.clone(),
+                        bucket_budget.clone(),
                        set_disk_inventory_clone.as_ref().clone(),
                        cache,
                        None,
@@ -1501,6 +1534,20 @@ impl ScannerIOCache for SetDisks {
                    }
                    let scan_outcome = match scan_result {
                        Ok(scan_outcome) => scan_outcome,
+                        Err(e) if bucket_budget.quantum_exhausted() => {
+                            record_partial_dirty_bucket(&partial_dirty_buckets_clone, &bucket.name).await;
+                            debug!(
+                                target: "rustfs::scanner::io",
+                                event = EVENT_SCANNER_DISK_BUCKET_STATE,
+                                component = LOG_COMPONENT_SCANNER,
+                                subsystem = LOG_SUBSYSTEM_IO,
+                                bucket = %bucket.name,
+                                state = "quantum_exhausted_without_progress",
+                                error = %e,
+                                "Scanner bucket quantum ended before durable progress"
+                            );
+                            continue;
+                        }
                        Err(e) => {
                            record_failed_dirty_bucket(&failed_dirty_buckets_clone, &bucket.name).await;
                            if ctx_clone.is_cancelled() {
@@ -1542,7 +1589,12 @@ impl ScannerIOCache for SetDisks {
                            }
                            (completed_cache, false)
                        }
-                        ScannerDiskScanOutcome::Partial(partial_cache) => (partial_cache, true),
+                        ScannerDiskScanOutcome::Partial(partial_cache) => {
+                            if partial_cache.info.scan_resource_deferred {
+                                work_guard.mark_service_interrupted();
+                            }
+                            (partial_cache, true)
+                        }
                        ScannerDiskScanOutcome::NamespaceNotFound(_) => {
                            if requeue_bucket_work(&bucket_tx_clone, &bucket, &mut work_guard).await {
                                increment_disk_bucket_scans_queued(
@@ -1566,7 +1618,6 @@ impl ScannerIOCache for SetDisks {
                        }
                    };
                    if partial {
-                        record_partial_dirty_bucket(&partial_dirty_buckets_clone, &bucket.name).await;
                        let done_save = Metrics::time(Metric::SaveUsage);
                        let partial_saved = if cache_guard.is_lock_lost() {
                            false
@@ -1604,6 +1655,21 @@ impl ScannerIOCache for SetDisks {
                        if partial_saved {
                            log_saved_partial_bucket_cache(&cache, &cache_name);
                        }
+                        // Only saved forward motion earns another turn in this
+                        // cycle; otherwise the bucket waits for the next one.
+                        if partial_saved
+                            && bucket_budget.quantum_exhausted()
+                            && cache.durable_scan_position() != scan_position_before
+                            && requeue_bucket_work(&bucket_tx_clone, &bucket, &mut work_guard).await
+                        {
+                            increment_disk_bucket_scans_queued(
+                                &queued_disk_bucket_scans_clone,
+                                &pool_label_clone,
+                                &set_label_clone,
+                            );
+                            continue;
+                        }
+                        record_partial_dirty_bucket(&partial_dirty_buckets_clone, &bucket.name).await;

                        continue;
                    }
@@ -1824,6 +1890,14 @@ impl ScannerIOCache for SetDisks {
                .await
                .extend(carried_buckets.iter().cloned());
        }
+        let service_ledger = {
+            let failed = bucket_failures.hard.lock().await;
+            let partial = bucket_failures.partial.lock().await;
+            service_round.settle(&completed_buckets, &failed, &partial)
+        };
+        if let Some(ledger) = &service_ledger {
+            record_bucket_failure_debt(source, want_cycle, ledger);
+        }
        let completed_count = completed_buckets.len();
        let covered_count = completed_count.saturating_add(carried_buckets.len());
        let coverage_complete =
@@ -1851,6 +1925,7 @@ impl ScannerIOCache for SetDisks {
            if cold_zero_walk_reuse_candidate {
                cold_zero_walk_reuse_observed.store(true, Ordering::Release);
            }
+            cache.info.scan_bucket_service = service_ledger.or(persisted_bucket_service);
            let _ = persist_and_publish_cache_snapshot(
                self.clone(),
                &updates,
@@ -1908,6 +1983,20 @@ impl ScannerIOCache for SetDisks {
                cancelled = ctx.is_cancelled(),
                "Scanner set cache publish skipped because cycle did not complete cleanly"
            );
+            // Publish the scope first: a contended cache lock must not delay it.
+            // A stopped owner leaves the set cache untouched; only a budget stop
+            // or a finished incomplete cycle hands its order to the next one.
+            if let Some(ledger) = service_ledger.filter(|_| !ctx.is_cancelled() || budget.budget_elapsed()) {
+                persist_scanner_service_ledger(
+                    self.clone(),
+                    source,
+                    ledger,
+                    want_cycle,
+                    leader_epoch,
+                    expected_publication_epoch,
+                )
+                .await;
+            }
        }

        debug!(
@@ -45,6 +45,7 @@ use time::OffsetDateTime;
 use uuid::Uuid;

 mod checkpoint_coverage_identity;
+mod cold_bucket_service;
 mod collector_lifecycle;
 mod cycle_collector_diagnostics;
 mod expiry_lock;
@@ -5176,6 +5177,35 @@ pub(super) async fn observe_coordinator_bucket_scan_admitted(worker_index: usize
    probe.active.fetch_sub(1, Ordering::SeqCst);
 }

+static FAILING_COORDINATOR_BUCKETS: std::sync::LazyLock<StdMutex<HashSet<String>>> =
+    std::sync::LazyLock::new(|| StdMutex::new(HashSet::new()));
+
+pub(super) fn coordinator_bucket_scan_fails(bucket: &str) -> bool {
+    FAILING_COORDINATOR_BUCKETS
+        .lock()
+        .unwrap_or_else(std::sync::PoisonError::into_inner)
+        .contains(bucket)
+}
+
+struct FailingCoordinatorBucket(String);
+
+impl Drop for FailingCoordinatorBucket {
+    fn drop(&mut self) {
+        FAILING_COORDINATOR_BUCKETS
+            .lock()
+            .unwrap_or_else(std::sync::PoisonError::into_inner)
+            .remove(&self.0);
+    }
+}
+
+fn fail_coordinator_bucket_scan(bucket: &str) -> FailingCoordinatorBucket {
+    FAILING_COORDINATOR_BUCKETS
+        .lock()
+        .unwrap_or_else(std::sync::PoisonError::into_inner)
+        .insert(bucket.to_string());
+    FailingCoordinatorBucket(bucket.to_string())
+}
+
 pub(super) async fn observe_coordinator_bucket_worker_started(worker_index: usize) {
    collector_lifecycle::observe_worker_start(worker_index).await;
 }
@@ -94,15 +94,15 @@ async fn checkpoint_resumes_across_a_maintenance_publication() {
    assert!(result.clear_verified_usage().is_empty());
    assert!(dirty_usage_buckets_for_tests().is_empty());

-    // Cycle 2: the written bucket is the only dirty one, so it is walked
-    // first and the object budget stops its sweep with a checkpoint; the
-    // cycle itself is incomplete.
+    // Cycle 2: the object budget stops the written bucket's sweep with a
+    // checkpoint; the cycle itself is incomplete. Service order follows the
+    // wait, so the budget leaves it two objects even behind the idle bucket.
    for object in ["p6/g", "p7/h"] {
        put_fixture_object(&store, "hot-bucket", object).await;
    }
    record_dirty_usage_bucket("hot-bucket");
    wait_for_namespace_commit_tails(store.as_ref()).await;
-    let (result, _, budget) = run_full_cycle(&store, 2, Some(2)).await;
+    let (result, _, budget) = run_full_cycle(&store, 2, Some(3)).await;
    assert_eq!(result.status, ScannerCycleStatus::Incomplete);
    assert!(budget.budget_elapsed());
    let set = Arc::clone(&store.pools[0].disk_set[0]);
@@ -0,0 +1,380 @@
+// 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 std::path::Path;
+
+const BUCKETS: usize = 17;
+const QUANTA: usize = 4;
+const ROUND_BOUND: usize = BUCKETS.div_ceil(QUANTA);
+const HOT: &str = "bucket-00";
+const FAILING: &str = "bucket-01";
+const ROUND_REQUEST_ENV: &str = "RUSTFS_COLD_BUCKET_SERVICE_ROUND";
+const EPOCH: u64 = 9;
+
+fn inventory() -> Vec<BucketInfo> {
+    (0..BUCKETS)
+        .map(|index| BucketInfo {
+            name: format!("bucket-{index:02}"),
+            ..Default::default()
+        })
+        .collect()
+}
+
+fn restore(persisted: Option<DataUsageBucketServiceLedger>, buckets: &[BucketInfo]) -> DataUsageBucketServiceLedger {
+    DataUsageBucketServiceLedger::restore(persisted, EPOCH, buckets.iter().map(|bucket| bucket.name.as_str()), None)
+}
+
+/// One scheduler lifetime over the production order and queue accounting.
+/// `persisted` is decoded from durable bytes; the result is encoded again.
+fn serve_round(persisted: &[u8]) -> (Vec<u8>, Vec<String>) {
+    let buckets = inventory();
+    let decoded = DataUsageCache::unmarshal(persisted).expect("decode saved order");
+    let round = Arc::new(ScannerBucketServiceRound::new(restore(decoded.info.scan_bucket_service, &buckets)));
+    let dirty = DirtyUsageBuckets::from([(HOT.to_string(), 1)]);
+    let mut order = bucket_usage_scan_order(&buckets, &DataUsageCache::default(), &dirty);
+    round.order_buckets(&mut order);
+    let served = order
+        .iter()
+        .take(QUANTA)
+        .map(|bucket| bucket.name.clone())
+        .collect::<Vec<_>>();
+    let mut completed = HashSet::new();
+    let mut failed = HashSet::new();
+    let ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(&ctx, Default::default());
+    for bucket in &served {
+        let mut work = BucketWorkGuard::new(Arc::new(AtomicUsize::new(2)), CancellationToken::new());
+        work.claim_service(round.clone(), bucket, ctx.clone(), budget.clone());
+        drop(work);
+        if bucket == FAILING {
+            failed.insert(bucket.clone());
+        } else {
+            completed.insert(bucket.clone());
+        }
+    }
+    let ledger = round.settle(&completed, &failed, &HashSet::new()).expect("served buckets");
+    let mut cache = DataUsageCache::default();
+    cache.info.scan_bucket_service = Some(ledger);
+    (cache.marshal_msg().expect("encode saved order"), served)
+}
+
+fn assert_service_bound(rounds: &[Vec<String>]) {
+    for bucket in inventory() {
+        let mut waited = 0;
+        for (index, served) in rounds.iter().enumerate() {
+            waited = if served.contains(&bucket.name) { 0 } else { waited + 1 };
+            assert!(waited < ROUND_BOUND, "{} waited {waited} rounds at round {index}", bucket.name);
+        }
+    }
+}
+
+#[test]
+fn cold_bucket_service_bound_holds_through_the_saved_codec() {
+    let mut saved = DataUsageCache::default().marshal_msg().expect("empty cache");
+    let mut rounds = Vec::new();
+    for _ in 0..(ROUND_BOUND * 4) {
+        let (next, served) = serve_round(&saved);
+        saved = next;
+        rounds.push(served);
+    }
+    assert_service_bound(&rounds);
+    let ledger = DataUsageCache::unmarshal(&saved)
+        .expect("decode final order")
+        .info
+        .scan_bucket_service
+        .expect("saved order");
+    let failing = ledger.buckets[FAILING];
+    assert!(failing.failure_debt >= 4, "every failed turn must stay visible: {failing:?}");
+    assert_eq!(failing.last_certified, None, "failure is not certification");
+    assert!(failing.service_quanta >= 4, "a failing bucket still spends real turns");
+    assert_eq!(ledger.failure_debt(), (1, Some(FAILING)));
+    assert_eq!(ledger.buckets[HOT].failure_debt, 0);
+    let hot_quanta = ledger.buckets[HOT].service_quanta;
+    assert!(
+        ledger
+            .buckets
+            .values()
+            .all(|service| service.service_quanta.abs_diff(hot_quanta) <= 1),
+        "a hot bucket must not earn more turns than a cold one: {ledger:?}"
+    );
+}
+
+/// The driver and each round are separate OS processes, so the only carrier
+/// between rounds is the saved cache file.
+#[test]
+fn cold_bucket_service_bound_holds_across_fresh_os_processes() {
+    let temp = tempfile::tempdir().expect("fixture directory");
+    let order = temp.path().join("order.bin");
+    std::fs::write(&order, DataUsageCache::default().marshal_msg().expect("empty cache")).expect("seed order");
+    let test = format!(
+        "{}::cold_bucket_service_process_round_worker",
+        module_path!().split_once("::").expect("crate path").1
+    );
+    let binary = std::env::current_exe().expect("test binary");
+    let listed = std::process::Command::new(&binary)
+        .args([test.as_str(), "--exact", "--list"])
+        .output()
+        .expect("list worker");
+    assert!(
+        String::from_utf8_lossy(&listed.stdout).contains(&format!("{test}: test")),
+        "the restart driver must select a real worker"
+    );
+    let mut rounds = Vec::new();
+    let mut pids = HashSet::new();
+    for round in 0..(ROUND_BOUND * 2) {
+        let status = std::process::Command::new(&binary)
+            .args([test.as_str(), "--exact", "--test-threads=1"])
+            .env(ROUND_REQUEST_ENV, &order)
+            .stdout(std::process::Stdio::null())
+            .status()
+            .expect("start worker round");
+        assert!(status.success(), "worker round {round} failed");
+        let report = std::fs::read_to_string(order.with_extension("served")).expect("worker report");
+        let (pid, served) = report.split_once('\n').expect("worker report shape");
+        assert!(pids.insert(pid.to_string()), "round {round} reused a process");
+        rounds.push(served.lines().map(str::to_owned).collect::<Vec<_>>());
+    }
+    assert_service_bound(&rounds);
+}
+
+#[test]
+fn cold_bucket_service_process_round_worker() {
+    let Some(order) = std::env::var_os(ROUND_REQUEST_ENV) else {
+        let empty = DataUsageCache::default().marshal_msg().expect("empty cache");
+        assert_eq!(serve_round(&empty).1.len(), QUANTA);
+        return;
+    };
+    let order = Path::new(&order);
+    let (saved, served) = serve_round(&std::fs::read(order).expect("read saved order"));
+    std::fs::write(order, saved).expect("save order");
+    std::fs::write(order.with_extension("served"), format!("{}\n{}", std::process::id(), served.join("\n")))
+        .expect("write worker report");
+}
+
+#[test]
+fn cold_bucket_service_new_bucket_joins_behind_existing_waiters() {
+    let buckets = inventory();
+    let mut ledger = restore(None, &buckets);
+    ledger.record_service("bucket-00");
+    let with_new = buckets
+        .iter()
+        .cloned()
+        .chain([BucketInfo {
+            name: "a-new-bucket".to_string(),
+            ..Default::default()
+        }])
+        .collect::<Vec<_>>();
+    let ledger = restore(Some(ledger), &with_new);
+    let mut order = with_new;
+    order.sort_by(|left, right| left.name.cmp(&right.name));
+    ScannerBucketServiceRound::new(ledger.clone()).order_buckets(&mut order);
+    assert_eq!(order[0].name, "bucket-01", "a stream of new buckets must not displace waiters");
+    let arrival = ledger.buckets["a-new-bucket"].waiting_since;
+    assert!(arrival > ledger.buckets["bucket-01"].waiting_since);
+    assert!(
+        arrival < ledger.buckets["bucket-00"].waiting_since,
+        "an arrival must not wait behind a bucket that already had its turn"
+    );
+}
+
+#[test]
+fn cold_bucket_service_scope_change_invalidates_the_saved_order() {
+    let buckets = inventory();
+    let mut ledger = restore(None, &buckets);
+    ledger.record_service(FAILING);
+    ledger.record_failure(FAILING);
+
+    let moved = DataUsageBucketServiceLedger::restore(
+        Some(ledger.clone()),
+        EPOCH + 1,
+        buckets.iter().map(|bucket| bucket.name.as_str()),
+        None,
+    );
+    assert_eq!(moved.publication_epoch, EPOCH + 1);
+    assert_eq!(moved.buckets[FAILING].failure_debt, 0, "another movement epoch starts a new order");
+
+    let mut corrupt = ledger.clone();
+    corrupt.next_sequence = 0;
+    let rebuilt = restore(Some(corrupt), &buckets);
+    assert_eq!(rebuilt.buckets[FAILING].service_quanta, 0, "an impossible wait position is not trusted");
+
+    let remaining = restore(Some(ledger.clone()), &buckets[..1]);
+    assert_eq!(remaining.buckets.len(), 1, "a deleted bucket leaves the order");
+}
+
+#[test]
+fn cold_bucket_service_recreated_bucket_loses_its_old_standing() {
+    let buckets = inventory();
+    let first = Uuid::from_u128(1);
+    let incarnations = HashMap::from([(FAILING.to_string(), first)]);
+    let names = || buckets.iter().map(|bucket| bucket.name.as_str());
+    let mut ledger = DataUsageBucketServiceLedger::restore(None, EPOCH, names(), Some(&incarnations));
+    ledger.record_failure(FAILING);
+    let same = DataUsageBucketServiceLedger::restore(Some(ledger.clone()), EPOCH, names(), Some(&incarnations));
+    assert_eq!(same.buckets[FAILING].failure_debt, 1);
+
+    let unknown = DataUsageBucketServiceLedger::restore(Some(ledger.clone()), EPOCH, names(), None);
+    assert_eq!(
+        unknown.buckets[FAILING].incarnation,
+        Some(first),
+        "a transient identity gap must not erase debt"
+    );
+
+    let recreated = HashMap::from([(FAILING.to_string(), Uuid::from_u128(2))]);
+    let replaced = DataUsageBucketServiceLedger::restore(Some(ledger), EPOCH, names(), Some(&recreated));
+    assert_eq!(replaced.buckets[FAILING].failure_debt, 0, "a recreated bucket does not inherit debt");
+    assert_eq!(replaced.buckets[FAILING].incarnation, Some(Uuid::from_u128(2)));
+}
+
+#[test]
+fn cold_bucket_service_budget_stop_and_resource_veto_add_no_debt() {
+    let buckets = inventory();
+    let ledger = restore(None, &buckets);
+    let before = ledger.buckets["bucket-02"].waiting_since;
+    let round = Arc::new(ScannerBucketServiceRound::new(ledger));
+    let ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(&ctx, Default::default());
+    let work = || BucketWorkGuard::new(Arc::new(AtomicUsize::new(2)), CancellationToken::new());
+
+    let mut vetoed = work();
+    vetoed.claim_service(round.clone(), "bucket-02", ctx.clone(), budget.clone());
+    vetoed.mark_requeued();
+    drop(vetoed);
+
+    let mut stopped = work();
+    stopped.claim_service(round.clone(), "bucket-03", ctx, budget.clone());
+    budget.cancel_for_runtime();
+    drop(stopped);
+
+    let names = |names: &[&str]| names.iter().map(|name| name.to_string()).collect::<HashSet<_>>();
+    let ledger = round
+        .settle(&HashSet::new(), &names(&["bucket-02", "bucket-03"]), &HashSet::new())
+        .expect("one bucket was served");
+    assert_eq!(ledger.buckets["bucket-02"].waiting_since, before, "a veto hands the turn back");
+    assert_eq!(ledger.buckets["bucket-02"].service_quanta, 0);
+    assert_eq!(ledger.buckets["bucket-03"].service_quanta, 1, "the budget was really spent");
+    assert_eq!(ledger.failure_debt(), (0, None), "budget-cancelled work is progress-only");
+}
+
+#[tokio::test]
+async fn cold_bucket_service_requeue_cannot_undo_the_next_workers_claim() {
+    let buckets = inventory();
+    let ledger = restore(None, &buckets);
+    let before = ledger.buckets["bucket-02"].waiting_since;
+    let round = Arc::new(ScannerBucketServiceRound::new(ledger));
+    let ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(&ctx, Default::default());
+    let work = || BucketWorkGuard::new(Arc::new(AtomicUsize::new(2)), CancellationToken::new());
+    let (bucket_tx, mut bucket_rx) = tokio::sync::mpsc::channel(1);
+
+    let mut first = work();
+    first.claim_service(round.clone(), "bucket-02", ctx.clone(), budget.clone());
+    assert!(requeue_bucket_work(&bucket_tx, &buckets[2], &mut first).await);
+
+    // The next worker takes the bucket while the first guard is still alive.
+    bucket_rx.recv().await.expect("requeued bucket");
+    let mut second = work();
+    second.claim_service(round.clone(), "bucket-02", ctx, budget);
+    drop(first);
+    drop(second);
+
+    let completed = HashSet::from(["bucket-02".to_string()]);
+    let ledger = round
+        .settle(&completed, &HashSet::new(), &HashSet::new())
+        .expect("the second claim served the bucket");
+    let service = &ledger.buckets["bucket-02"];
+    assert!(service.waiting_since > before, "the completed bucket returns to the tail");
+    assert_eq!(service.service_quanta, 1, "only the completed turn counts");
+}
+
+#[test]
+fn cold_bucket_service_partial_resource_veto_adds_no_debt() {
+    let buckets = inventory();
+    let round = Arc::new(ScannerBucketServiceRound::new(restore(None, &buckets)));
+    let ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(&ctx, Default::default());
+    let work = || BucketWorkGuard::new(Arc::new(AtomicUsize::new(2)), CancellationToken::new());
+
+    let mut vetoed = work();
+    vetoed.claim_service(round.clone(), "bucket-02", ctx.clone(), budget.clone());
+    vetoed.mark_service_interrupted();
+    drop(vetoed);
+    let mut failed = work();
+    failed.claim_service(round.clone(), "bucket-03", ctx, budget);
+    drop(failed);
+
+    let partial = HashSet::from(["bucket-02".to_string(), "bucket-03".to_string()]);
+    let ledger = round
+        .settle(&HashSet::new(), &HashSet::new(), &partial)
+        .expect("two buckets were served");
+    assert_eq!(ledger.buckets["bucket-02"].failure_debt, 0, "a resource block is not a failure");
+    assert_eq!(ledger.buckets["bucket-03"].failure_debt, 1, "an uninterrupted partial stays visible");
+}
+
+#[test]
+fn cold_bucket_service_resource_deferral_marker_is_never_persisted() {
+    let mut cache = DataUsageCache::default();
+    cache.info.name = DATA_USAGE_ROOT.to_string();
+    cache.info.scan_resource_deferred = true;
+    let decoded = DataUsageCache::unmarshal(&cache.marshal_msg().expect("encode")).expect("decode");
+    assert!(!decoded.info.scan_resource_deferred);
+}
+
+#[test]
+fn cold_bucket_service_order_is_an_optional_map_field_for_old_readers() {
+    #[derive(serde::Deserialize)]
+    struct LegacyInfo {
+        name: String,
+        next_cycle: u64,
+    }
+    #[derive(serde::Deserialize)]
+    struct LegacyCache {
+        info: LegacyInfo,
+        #[serde(rename = "cache")]
+        _cache: serde::de::IgnoredAny,
+    }
+    let mut cache = DataUsageCache::default();
+    cache.info.name = DATA_USAGE_ROOT.to_string();
+    cache.info.next_cycle = 7;
+    let absent = cache.marshal_msg().expect("encode cache");
+    let field = b"scan_bucket_service";
+    assert!(
+        !absent.windows(field.len()).any(|window| window == field),
+        "an unused order must not change the encoded cache"
+    );
+
+    let buckets = inventory();
+    let mut ledger = restore(None, &buckets);
+    ledger.record_service(FAILING);
+    ledger.record_failure(FAILING);
+    cache.info.scan_bucket_service = Some(ledger.clone());
+    let encoded = cache.marshal_msg().expect("encode order");
+    let legacy: LegacyCache = rmp_serde::from_slice(&encoded).expect("old readers must ignore the order");
+    assert_eq!((legacy.info.name.as_str(), legacy.info.next_cycle), (DATA_USAGE_ROOT, 7));
+    assert_eq!(
+        DataUsageCache::unmarshal(&encoded)
+            .expect("decode order")
+            .info
+            .scan_bucket_service,
+        Some(ledger)
+    );
+    assert!(
+        DataUsageCache::unmarshal(&absent)
+            .expect("decode a cache from an older writer")
+            .info
+            .scan_bucket_service
+            .is_none()
+    );
+}
@@ -337,3 +337,221 @@ async fn service_cohort_cancelled_dispatch_does_not_consume_waiters_or_leak_perm
    assert_eq!(usage.expect("retry aggregate").objects_total_count, 2);
    clear_dirty_usage_buckets_for_tests();
 }
+
+/// N=17, q=4: every bucket reaches a worker within ceil(17/4)=5 rounds while
+/// one bucket stays hot and another fails forever, with leader-local state
+/// discarded after each round as a restart would.
+#[tokio::test]
+#[serial]
+async fn cold_bucket_service_bound_survives_restart_with_hot_and_failing_bucket() {
+    const BUCKETS: usize = 17;
+    const QUANTA: u64 = 4;
+    const ROUND_BOUND: usize = 5;
+    let (_dir, store) = setup_local_scanner_store(1, 4).await;
+    clear_dirty_usage_buckets_for_tests();
+    let names = (0..BUCKETS).map(|index| format!("cold-{index:02}")).collect::<Vec<_>>();
+    for name in &names {
+        create_cohort_bucket(&store, name).await;
+    }
+    let hot = names[0].clone();
+    let _failing = fail_coordinator_bucket_scan(&names[1]);
+    let mut waited = [0_usize; BUCKETS];
+    for round in 1..=(ROUND_BOUND * 3) {
+        record_dirty_usage_bucket(&hot);
+        let cohort = Arc::new(StdMutex::new(ScannerServiceCohort::default()));
+        let ctx = CancellationToken::new();
+        let budget = ScannerCycleBudget::new_with_progress_tracking(
+            &ctx,
+            ScannerCycleBudgetConfig {
+                max_objects: Some(QUANTA),
+                ..Default::default()
+            },
+        );
+        run_cohort_cycle(&store, cohort.clone(), u64::try_from(round).expect("round"), budget).await;
+        let admitted = cohort
+            .lock()
+            .expect("cohort lock")
+            .admitted_members()
+            .into_iter()
+            .map(|(_, bucket)| bucket)
+            .collect::<HashSet<_>>();
+        for (index, name) in names.iter().enumerate() {
+            waited[index] = if admitted.contains(name) { 0 } else { waited[index] + 1 };
+            assert!(
+                waited[index] < ROUND_BOUND,
+                "{name} waited {} rounds at round {round}; admitted {admitted:?}",
+                waited[index]
+            );
+        }
+    }
+    let saved = load_set_cache(&store).await;
+    let ledger = saved
+        .info
+        .scan_bucket_service
+        .expect("incomplete cycles must save the service order");
+    assert!(ledger.buckets[&names[1]].failure_debt > 0, "the failing bucket must keep visible debt");
+    assert_eq!(ledger.buckets[&names[1]].last_certified, None);
+    assert_eq!(ledger.failure_debt().0, 1, "only the failing bucket owes a sweep");
+    assert!(!saved.info.snapshot_complete, "the saved order must not certify the set");
+    clear_dirty_usage_buckets_for_tests();
+}
+
+async fn load_set_cache(store: &Arc<ECStore>) -> DataUsageCache {
+    let mut cache = DataUsageCache::default();
+    cache
+        .load(store.all_set_disks()[0].clone(), DATA_USAGE_CACHE_NAME)
+        .await
+        .expect("load set cache");
+    cache
+}
+
+fn quantum_budget(quantum: crate::scanner_budget::ScannerBucketQuantum) -> Arc<ScannerCycleBudget> {
+    ScannerCycleBudget::new_with_progress_tracking(
+        &CancellationToken::new(),
+        ScannerCycleBudgetConfig {
+            bucket_quantum: Some(quantum),
+            ..Default::default()
+        },
+    )
+}
+
+#[tokio::test]
+#[serial]
+async fn fair_prefix_quantum_spent_bucket_returns_behind_waiting_buckets() {
+    let (_dir, store) = setup_local_scanner_store(1, 4).await;
+    clear_dirty_usage_buckets_for_tests();
+    create_cohort_bucket_objects(&store, "a-flat", 6).await;
+    create_cohort_bucket(&store, "z-cold-one").await;
+    create_cohort_bucket(&store, "z-cold-two").await;
+    let cohort = Arc::new(StdMutex::new(ScannerServiceCohort::default()));
+    let budget = quantum_budget(crate::scanner_budget::ScannerBucketQuantum {
+        max_objects: Some(2),
+        ..Default::default()
+    });
+    let (result, usage) = run_cohort_cycle(&store, cohort, 1, budget.clone()).await;
+    assert_eq!(result.status, ScannerCycleStatus::Complete, "shares must add up to one whole sweep");
+    assert_eq!(usage.expect("complete aggregate").objects_total_count, 8);
+    assert!(!budget.budget_elapsed(), "only the bucket shares ended");
+
+    let ledger = load_set_cache(&store).await.info.scan_bucket_service.expect("saved order");
+    assert!(ledger.buckets["a-flat"].service_quanta >= 3, "{ledger:?}");
+    assert_eq!(ledger.buckets["z-cold-one"].service_quanta, 1, "{ledger:?}");
+    assert_eq!(ledger.buckets["z-cold-two"].service_quanta, 1, "{ledger:?}");
+    assert!(
+        ledger.buckets["a-flat"].waiting_since > ledger.buckets["z-cold-two"].waiting_since,
+        "the wide bucket must finish behind the buckets it yielded to: {ledger:?}"
+    );
+    assert_eq!(ledger.failure_debt(), (0, None), "a spent share is not a failure");
+    clear_dirty_usage_buckets_for_tests();
+}
+
+#[tokio::test]
+#[serial]
+async fn fair_prefix_quantum_without_durable_motion_ends_the_bucket_turn() {
+    let (_dir, store) = setup_local_scanner_store(1, 4).await;
+    clear_dirty_usage_buckets_for_tests();
+    create_cohort_bucket_objects(&store, "a-flat", 6).await;
+    let cohort = Arc::new(StdMutex::new(ScannerServiceCohort::default()));
+    let budget = quantum_budget(crate::scanner_budget::ScannerBucketQuantum {
+        max_duration: Some(Duration::ZERO),
+        ..Default::default()
+    });
+    let (result, usage) = run_cohort_cycle(&store, cohort, 1, budget).await;
+    assert_eq!(result.status, ScannerCycleStatus::Incomplete);
+    assert!(usage.is_none());
+    let ledger = load_set_cache(&store).await.info.scan_bucket_service.expect("saved order");
+    assert_eq!(
+        ledger.buckets["a-flat"].service_quanta, 1,
+        "a share that saved nothing must not requeue itself: {ledger:?}"
+    );
+    assert_eq!(ledger.failure_debt(), (0, None));
+    clear_dirty_usage_buckets_for_tests();
+}
+
+/// A one-object share stops inside its first child, so the frontier never
+/// moves. Requeuing that bucket would spin until the cycle deadline.
+#[tokio::test]
+#[serial]
+async fn fair_prefix_quantum_repeated_share_without_a_new_frontier_stops_requeuing() {
+    let (_dir, store) = setup_local_scanner_store(1, 4).await;
+    clear_dirty_usage_buckets_for_tests();
+    create_cohort_bucket_objects(&store, "a-flat", 3).await;
+    let cohort = Arc::new(StdMutex::new(ScannerServiceCohort::default()));
+    let budget = quantum_budget(crate::scanner_budget::ScannerBucketQuantum {
+        max_objects: Some(1),
+        ..Default::default()
+    });
+    let (result, usage) = run_cohort_cycle(&store, cohort, 1, budget).await;
+    assert_eq!(result.status, ScannerCycleStatus::Incomplete);
+    assert!(usage.is_none());
+    let ledger = load_set_cache(&store).await.info.scan_bucket_service.expect("saved order");
+    assert!(
+        ledger.buckets["a-flat"].service_quanta <= 3,
+        "turns without durable motion must end: {ledger:?}"
+    );
+    assert_eq!(ledger.failure_debt(), (0, None));
+    clear_dirty_usage_buckets_for_tests();
+}
+
+#[tokio::test]
+#[serial]
+async fn cold_bucket_service_saved_order_keeps_coverage_and_rejects_older_owners() {
+    let (_dir, store) = setup_local_scanner_store(1, 4).await;
+    clear_dirty_usage_buckets_for_tests();
+    create_cohort_bucket(&store, "cold-only").await;
+    let ctx = CancellationToken::new();
+    let cohort = Arc::new(StdMutex::new(ScannerServiceCohort::default()));
+    let (result, _) = run_cohort_cycle(&store, cohort, 3, ScannerCycleBudget::new(&ctx, Default::default())).await;
+    assert_eq!(result.status, ScannerCycleStatus::Complete);
+    let certified = load_set_cache(&store).await;
+    assert!(certified.info.snapshot_complete);
+    let set = store.all_set_disks()[0].clone();
+    let epoch = scanner_publication_epoch(set.clone()).await.expect("publication epoch");
+    let source = DataUsageCacheSource::new(set.pool_index, set.set_index);
+    let mut ledger = certified.info.scan_bucket_service.clone().expect("saved order");
+    ledger.next_sequence += 1;
+    ledger.leader_epoch = 11;
+
+    assert!(persist_scanner_service_ledger(set.clone(), source, ledger.clone(), 3, 11, epoch).await);
+    let saved = load_set_cache(&store).await;
+    assert_eq!(saved.info.scan_bucket_service.as_ref(), Some(&ledger));
+    assert_eq!(saved.info.snapshot_complete, certified.info.snapshot_complete);
+    assert_eq!(saved.info.scan_plan_digest, certified.info.scan_plan_digest);
+    assert_eq!(saved.info.scan_execution_digest, certified.info.scan_execution_digest);
+    assert_eq!(saved.info.next_cycle, certified.info.next_cycle);
+    assert_eq!(saved.info.leader_epoch, certified.info.leader_epoch);
+    assert_eq!(saved.cache.len(), certified.cache.len());
+
+    let mut stale = ledger.clone();
+    stale.next_sequence += 1;
+    assert!(
+        !persist_scanner_service_ledger(set.clone(), source, stale.clone(), 3, 10, epoch).await,
+        "a late write from an older leader must not replace the order"
+    );
+    assert!(
+        !persist_scanner_service_ledger(set.clone(), source, stale, 2, 11, epoch).await,
+        "a late write from an older cycle must not replace the order"
+    );
+    assert_eq!(load_set_cache(&store).await.info.scan_bucket_service.as_ref(), Some(&ledger));
+
+    // An incomplete save leaves the cache stamps at epoch 11, so only the order
+    // itself can reject the leader it replaced.
+    let mut newer = ledger.clone();
+    newer.next_sequence += 1;
+    assert!(persist_scanner_service_ledger(set.clone(), source, newer.clone(), 3, 12, epoch).await);
+    let saved = load_set_cache(&store).await;
+    assert_eq!(saved.info.leader_epoch, certified.info.leader_epoch);
+    assert!(
+        !persist_scanner_service_ledger(set.clone(), source, ledger.clone(), 3, 11, epoch).await,
+        "a late write must not replace an order saved by a newer leader"
+    );
+    assert_eq!(
+        load_set_cache(&store)
+            .await
+            .info
+            .scan_bucket_service
+            .map(|saved| saved.next_sequence),
+        Some(newer.next_sequence)
+    );
+    clear_dirty_usage_buckets_for_tests();
+}
@@ -90,7 +90,9 @@ pub use bucket::BUCKET_LIFECYCLE_LOCK_OBJECT;
 pub use bucket::{BucketInfo, BucketOperations, BucketOptions, DeleteBucketOptions, MakeBucketOptions, SRBucketDeleteOp};
 pub use capability::{CapabilitySnapshotError, CapabilityState, CapabilityStatus};
 pub use error::{StorageErrorCode, StorageResult};
+pub use multipart::MultipartUploadInspection;
 pub use multipart::{CompletePart, ListMultipartsInfo, ListPartsInfo, MultipartInfo, MultipartUploadResult, PartInfo};
+pub use object::AuthoritativeListOperations;
 pub use object::DeleteAccounting;
 pub use object::ObjectLockDeleteOptions;
 pub use object::ObjectToDeleteIdentity;
@@ -15,6 +15,15 @@
 use std::collections::HashMap;
 use time::OffsetDateTime;

+/// Checks upload absence before removing a bucket's data protections.
+#[async_trait::async_trait]
+pub trait MultipartUploadInspection: Send + Sync {
+    type Error: std::error::Error + Send + Sync + 'static;
+
+    /// Unknown or unreadable upload ownership must fail this check.
+    async fn ensure_no_multipart_uploads(&self, bucket: &str) -> Result<(), Self::Error>;
+}
+
 #[derive(Debug, Default, Clone)]
 pub struct MultipartUploadResult {
    pub upload_id: String,
@@ -566,6 +566,29 @@ pub trait ListOperations: Send + Sync + fmt::Debug {
    ) -> Result<(), Self::Error>;
 }

+/// Quorum-backed enumeration for decisions that require proving object absence.
+///
+/// Consumers must drain the result stream and await successful completion before
+/// trusting the scan. An error or cancellation can follow partial results and
+/// leaves absence unproven. Callers coordinate concurrent mutations separately;
+/// this capability does not provide a snapshot or retain a mutation lock.
+#[async_trait::async_trait]
+pub trait AuthoritativeListOperations: ListOperations {
+    /// Resolves every set at its version read quorum, including reserved prefixes.
+    ///
+    /// # Errors
+    /// Returns an error when a set cannot establish quorum, metadata cannot be
+    /// resolved, or enumeration is interrupted. Retry starts a new scan.
+    async fn walk_authoritative(
+        self: Arc<Self>,
+        rx: Self::WalkCancellation,
+        bucket: &str,
+        prefix: &str,
+        result: Self::WalkResultSender,
+        opts: Self::WalkOptions,
+    ) -> Result<(), Self::Error>;
+}
+
 #[async_trait::async_trait]
 #[allow(clippy::too_many_arguments)]
 pub trait MultipartOperations: Send + Sync + fmt::Debug {
@@ -91,6 +91,26 @@ Use `GET /iceberg/v1/buckets/{warehouse}` to check `enabled` and `disable-suppor

 Review the bucket lifecycle configuration before disabling: clearing table-bucket protection allows existing object expiration rules to run again. Objects are preserved by the disable operation itself. Catalog disablement requires no namespaces, retained table/view resources, or objects under `.rustfs-table/`; an active rename or backing migration also blocks it. Clean up through supported catalog operations; do not manually delete internal metadata to bypass these checks.

-The server persists an inactive catalog entry before clearing the bucket marker. If marker publication fails, retry the same DELETE; the marker retains object protections until the retry succeeds. A subsequent PUT explicitly enables the bucket again. Stale requests cannot initialize an inactive catalog entry. No table metadata or object data is deleted by either state change.
+The server persists an inactive catalog entry before clearing the bucket marker. If marker publication fails, retry the same DELETE; the marker retains object protections until the retry succeeds. A subsequent PUT explicitly enables the bucket again. Stale requests cannot initialize an inactive catalog entry. A plain DELETE and a subsequent PUT preserve table metadata and object data.
+
+### Clean Up Deleted Empty Tables
+
+For the object-backed catalog, `DELETE /iceberg/v1/buckets/{warehouse}?cleanup=true` removes the metadata and internal records left by deleted empty tables, then disables the table catalog. Ordinary table deletion continues to preserve files. Cleanup preserves existing table record formats and can process empty tables deleted before the server upgrade.
+
+Complete the server upgrade across all serving nodes and set the existing `RUSTFS_TABLE_CATALOG_PUBLICATION_FENCE_FLEET_CONFIRMED=true` setting before requesting cleanup. Delete the empty tables and their namespaces through the catalog APIs, then send the cleanup request with `admin:SetTableBucket` and `s3:GetObject`/`s3:DeleteObject` permissions for the target metadata objects. The `/_iceberg/v1` alias accepts the same option. Omitting `cleanup` or setting it to `false` keeps the ordinary disable behavior; repeated or invalid values are rejected.
+
+Cleanup accepts only recognized empty-table metadata in the default managed warehouse. Active catalog resources, snapshots or other data references, user objects, object versions, incomplete multipart uploads, corrupt or unknown metadata, and unfinished catalog operations block cleanup. The server validates the complete cleanup plan and object permissions before deleting any object. Durable-strong catalogs retain their existing disable procedure.
+
+If a storage operation fails during cleanup, retry the same request. The server checks the remaining objects again and keeps the bucket's table protection enabled until cleanup and catalog disablement complete. No separate cleanup task file is created. After a successful response with `enabled: false`, delete the empty bucket through the normal S3 `DeleteBucket` operation.
+
+## Object-Backed Registration Reads
+
+Table registration checks table ID uniqueness and warehouse directory overlap using the existing catalog and warehouse index records. Each registration obtains a complete inventory under the existing publication and migration guards. System-bucket inventory records must satisfy the same metadata commit checks as ordinary catalog reads. An incomplete or failed inventory rejects registration.
+
+Each server keeps a bounded memory cache of validated ownership records. A record can be reused only when its object identity and ETag match the current inventory. New or changed records are read and validated again. Cache eviction and process restarts cause additional reads; catalog state continues to come from the existing persisted records. The cache is automatic and requires no additional configuration or stored objects.
+
+Registration still enumerates the namespace and warehouse index prefixes, so enumeration work grows with catalog size. Cache hits reduce object-body reads and JSON parsing. A cache hit verifies the current metadata identity and reuses previously validated content; it does not re-read the current body to check its physical readability. Conditional writes, rollback, and lock-loss fencing continue to control publication.
+
+Existing catalog formats and Console and Iceberg client request formats are preserved. Concurrent writers must follow the existing publication, migration, and storage fencing protocol; the memory cache does not provide synchronization for writers that bypass those guards.

 Before rolling back to a server version without this endpoint, re-enable disabled catalogs with PUT while the supporting version is running. Older enable handlers do not reactivate inactive catalog entries. Complete the server upgrade across the cluster before using disablement.
@@ -0,0 +1,49 @@
+# Event-class drills and retained evidence
+
+**Use this when:** you need to confirm that a CI event class has been exercised end to end, locate its retained run evidence, or repeat the drill. **Source of truth:** the live `main` ruleset for required status; `.github/workflows/ci.yml` for event triggers; `scripts/ci_gate.py` for per-event selection and expected results; the run URLs below for observed outcomes.
+
+Every event class the planner serves must have at least one reviewable drill: an actual run whose plan, job set and gate result can be re-inspected. A drill proves the wiring for that event, not product quality and not a cost saving. Static fixtures in `scripts/test_ci_plan.py` and `scripts/test_ci_merge_queue.py` cover the negative cases (missing, cancelled, mismatched, combined-candidate conflicts) and are not substitutes for the live drills listed here.
+
+## Event classes and expected selection
+
+| Event | Selection | Expected gate behavior |
+|---|---|---|
+| `pull_request` (ordinary code) | reverse-dependency closure + fixed core inventory; shared/unknown inputs expand to full | `Test and Lint` requires exactly the planned job results |
+| `pull_request` (documentation only) | `docs` mode: no code jobs | Quick Checks and Typos run; expensive jobs are required **skipped** |
+| `pull_request` (isolated E2E) | `e2e` mode: workspace + server build + E2E smoke + S3 lanes | Connect boundary, serial ILM, protocol/rio-v2, real io_uring required **skipped** |
+| `pull_request` (closed) | `Cancel Closed PR Runs` only | in-flight validation for that PR is cancelled; no new validation is claimed |
+| `pull_request` (merged) | entry into `merge_group` | combined candidate is validated as its own SHA/tree with the full matrix |
+| `push` to `main` | full | complete validation; no reuse inferred from tree equality |
+| `push` of a release tag | `build.yml` release matrix | build and publication path (not the PR planner) |
+| `schedule` | registry-selected cadence plan | suites run under their resource locks; failures feed the health monitors |
+| `workflow_dispatch` | full unless an explicit narrow input is passed | manual runs without a plan stay visible as unverified |
+| `workflow_run` | monitor/filter through the responsibility registry | independent health layer reports expected SHA, attempt and evidence completeness |
+
+## Retained evidence index
+
+Run URLs are `https://github.com/rustfs/rustfs/actions/runs/<id>`. Re-open the run and its `Select CI scope` job log to read the planner's recorded mode, reasons and expected job results.
+
+| Class | Drill | Run | Observed |
+|---|---|---|---|
+| Ordinary PR | `test(heal): align cleanup tests with removal-disabled heals` | [41153190859](https://github.com/rustfs/rustfs/actions/runs/41153190859) | full scope (12 real jobs: workspace build/tests, E2E, S3, core contract, io_uring); gate accepted exact planned results |
+| PR closed | `test(heal): align cleanup tests…` closure handler | [41160735383](https://github.com/rustfs/rustfs/actions/runs/41160735383) | only `Cancel Closed PR Runs` executed; every validation job required **skipped** |
+| `push` to `main` | `chore(deps): refresh workspace dependencies (#9080)` | [41160768241](https://github.com/rustfs/rustfs/actions/runs/41160768241) | full matrix on the merge commit; no selection shortcut |
+| Release tag | `1.0.4` | [40198812318](https://github.com/rustfs/rustfs/actions/runs/40198812318) | `Build and Release` release matrix succeeded on the tag ref |
+| `schedule` | `Functional chain health` | [41159991473](https://github.com/rustfs/rustfs/actions/runs/41159991473) | monitor refused to advance freshness on an incomplete chain attempt (failure is the expected answer while the chain is unhealthy) |
+| `workflow_dispatch` | manual full validation | [41152382212](https://github.com/rustfs/rustfs/actions/runs/41152382212) | full suite set ran and the aggregate accepted the complete plan (85 min) |
+| `workflow_run` | `Scheduled Validation Watchdog` | [41161431618](https://github.com/rustfs/rustfs/actions/runs/41161431618) | independent filter/monitor path executed against the registry |
+| `merge_group` | first queued candidate after queue enablement | see below | live evidence attached after the first queue entry completes |
+| Documentation-only PR | PR #9088, this document's own PR | [41162871854](https://github.com/rustfs/rustfs/actions/runs/41162871854) | `docs` mode: `Select CI scope` recorded `mode: docs`; `Test and Lint` accepted in 8 s with every expensive job required **skipped** |
+| Isolated E2E PR | PR #9089, comment-only change under `crates/e2e_test/src/` | [41163203959](https://github.com/rustfs/rustfs/actions/runs/41163203959) | `e2e` mode: Connect boundary, serial ILM, rio-v2/protocol lanes and real io_uring required **skipped**; workspace build, E2E and both S3 lanes retained |
+
+## Repeating a drill
+
+Ordinary, docs and E2E classes: open a pull request containing only that class's paths and inspect `Select CI scope` → the planner's `mode` and `reasons`, then the aggregate's expected-vs-actual comparison. Close the PR without merging to leave no product change; the closure handler itself is the closed-event drill.
+
+Merge-group class: queue a pull request through the enabled `merge_queue` rule (see [verification plans and artifacts](ci-verification-pipeline.md) for the switch record and rollback payload). The `merge_group` run validates the combined candidate's own SHA/tree. Two individually passing changes whose combination fails stay blocked until their combined candidate passes; that negative case is fixed in `scripts/test_ci_merge_queue.py`.
+
+Schedule and `workflow_run` classes cannot be forced on demand without an event; use the retained runs above and the registry's own `python3 scripts/validation_registry.py check` output. Tag class requires a real tag push; use the retained release runs or the preview-tag rehearsal described in [CI gate matrix](ci-gates.md).
+
+## Boundaries
+
+A completed drill is evidence that the event wiring behaves as planned for that candidate. It is not release certification, not a latency or cost result (use [CI timing samples](ci-timing.md) and [resource accounting](ci-resource-accounting.md)), and not a substitute for the required checks' own product coverage. Do not edit this file to claim a drill that has no run URL; attach the run first.
@@ -18,7 +18,7 @@ Every PR enters `ci.yml`. The `classify-changes` job uses the base revision of `

 PRs changing only Rust sources under `crates/e2e_test/src/` and known E2E selection digests (optionally with those documentation paths) use the `e2e` scope. They keep workspace validation, the server build, E2E smoke, and both S3 lanes; independent workflows retain their path selection. Connect boundary, serial ILM, optional protocol/rio-v2, and real io_uring jobs are skipped because their sources and test inputs are unchanged. Manifests, fixtures, shared nextest configuration, scripts, production sources, and unknown paths still select `full`. A dependency on `e2e_test` from another workspace member or reachable local dependency disables this shortcut. All non-PR events retain their full selection. The structured planner enforces affected-package and Connect selection from the trusted base policy, retaining fixed core coverage and full validation for shared or unknown inputs. See [verification plans and artifacts](ci-verification-pipeline.md) for identity, rollout and queue-switch contracts.

-Alternate-feature unit tests and Clippy for rio-v2, Swift, SFTP, and HTTP/3 run on merge groups and on scheduled (nightly) or manual validation. PRs and main pushes run the default workspace and retain ILM, S3, real io_uring, and Connect security boundaries in their applicable scopes. The aggregate requires the alternate-feature jobs to be skipped on PRs and main pushes and successful on merge-group, scheduled, and manual events. This avoids repeating four feature builds on every PR revision and on every merge; feature-specific regressions are detected by the nightly run when the merge queue is not used. For changes to these features, dispatch `ci.yml` against the candidate branch before merging when that signal is needed, including PRs targeting `gateway/integration`, which has no automatic push trigger in this workflow. Scheduled results supplement the main baseline and do not replace it.
+Specialized boundary, hardware and alternate-feature lanes run only on the scheduled (nightly) and manual full-validation runs: `Connect Feature Boundaries`, `io_uring Integration (real)`, and the rio-v2, Swift, SFTP and HTTP/3 feature lanes. They do not gate a merge. The nightly run reports their regressions within a day, and repeating these builds on every PR revision and merge group bought no earlier signal worth the runner occupancy (Connect alone measured 53.5 median minutes per run and occasionally set the wall). PRs, merge groups and main pushes run the default workspace and retain ILM, S3, E2E smoke, the short core contract and the debug build. The aggregate requires the specialized lanes to be skipped on pull-request, merge-group and push events and successful on scheduled and manual events. For changes to these features, dispatch `ci.yml` against the candidate branch before merging when that signal is needed, including PRs targeting `gateway/integration`, which has no automatic push trigger in this workflow. Scheduled results supplement the main baseline and do not replace it.

 Workspace Clippy runs on every event with a selected workspace Rust job, pull requests included, in `Workspace Lint and Build`. That job also creates the exact test archive and runs doctests; `Workspace tests` consumes the archive without rebuilding it. `cache-warm.yml` warms Clippy's check artifacts under `ci-dev` with the same invocation to reuse compatible dependency artifacts. `make pre-commit` does not run Clippy. Select local Clippy targets, features, and documentation exemptions from the root `AGENTS.md` verification tier; the workspace command below reproduces the CI lane and is not a universal pre-push requirement.

@@ -50,13 +50,13 @@ The four site-replication state-writer concurrency proofs reserve the available
 | Event | Check name | Workflow / job | Merge status | Reproduce |
 |---|---|---|---|---|
 | PR, non-doc change | `Quick Checks` | `ci.yml` `quick-checks` | Required | `make pre-commit` |
-| PR, non-doc change, and every non-PR event | `Workspace Lint and Build`, `Workspace tests` | `ci.yml` `test-and-lint` / `workspace-tests` | Via aggregate | `cargo clippy --all-targets -- -D warnings`; `cargo nextest run --profile ci --all --exclude e2e_test`; `cargo test --all --exclude e2e_test --doc`; migration evidence check described below |
+| PR, non-doc change, and every non-PR event | `Workspace Lint and Build`, `Workspace tests` | `ci.yml` `workspace-lint-and-build` / `workspace-tests` | Via aggregate | `cargo clippy --all-targets -- -D warnings`; `cargo nextest run --profile ci --all --exclude e2e_test`; `cargo test --all --exclude e2e_test --doc`; migration evidence check described below |
 | PR, non-doc change | `Typos` | `ci.yml` `typos` | Via aggregate | `typos` |
 | Full selection | `Workspace tests` (serial ILM step) | `ci.yml` `workspace-tests` | Via aggregate | exact `Run ignored ILM integration tests serially` command in the job |
-| Merge group, nightly schedule, manual dispatch | `Test and Lint (rio-v2)`, `Test and Lint (swift)`, `Test and Lint (sftp)`, `Test and Lint (http3)` | `ci.yml` `test-and-lint-rio-v2`, `test-and-lint-protocols` | Via aggregate on these events; skipped on PRs and main pushes | `cargo nextest run` with the job's `--features` |
-| PR, full selection | `Connect Feature Boundaries` | `ci.yml` `connect-feature-boundaries` | Via aggregate | `cargo test -p rustfs --test connect_registration --features connect-e2e-short-credentials`; `cargo check -p rustfs --release --features connect-e2e-short-credentials` must fail; `scripts/check_offline_enrollment_e2e.sh` |
+| Nightly schedule, manual dispatch | `Test and Lint (rio-v2)`, `Test and Lint (swift)`, `Test and Lint (sftp)`, `Test and Lint (http3)` | `ci.yml` `test-and-lint-rio-v2`, `test-and-lint-protocols` | Via aggregate on these events; skipped on PRs, merge groups and main pushes | `cargo nextest run` with the job's `--features` |
+| Nightly schedule, manual dispatch, full selection | `Connect Feature Boundaries` | `ci.yml` `connect-feature-boundaries` | Via aggregate on these events; skipped on PRs, merge groups and main pushes | `cargo test -p rustfs --test connect_registration --features connect-e2e-short-credentials`; `cargo check -p rustfs --release --features connect-e2e-short-credentials` must fail; `scripts/check_offline_enrollment_e2e.sh` |
 | PR, non-doc change | `Build RustFS Debug Binary` | `ci.yml` `build-rustfs-debug-binary` | Via aggregate; prerequisite for black-box jobs | `python3 scripts/e2e_binary.py build --bins --features e2e-test-hooks` (binary plus its `rustfs.e2e.json` sidecar) |
-| PR, full selection | `io_uring Integration (real)` | `ci.yml` `uring-integration` | Via aggregate | `cargo test -p rustfs-ecstore --lib uring_ -- --test-threads=1 --nocapture` |
+| Nightly schedule, manual dispatch, full selection | `io_uring Integration (real)` | `ci.yml` `uring-integration` | Via aggregate on these events; skipped on PRs, merge groups and main pushes | `cargo test -p rustfs-ecstore --lib uring_ -- --test-threads=1 --nocapture` |
 | PR, non-doc change | `End-to-End Tests` | `ci.yml` `e2e-tests` | Via aggregate | `python3 scripts/e2e_binary.py run --features e2e-test-hooks -- cargo nextest run --profile e2e-smoke -p e2e_test`, then `./scripts/e2e-run.sh ./target/debug/rustfs <data-dir>` under the same wrapper; membership guards `scripts/check_test_wiring.py --check-profile e2e-smoke <listing.json>` and `scripts/check_security_smoke_count.sh check <listing.json>` |
 | PR, non-doc change | `Core contract` | `ci.yml` `core-contract` | Via aggregate | existing write/read, anonymous denial and restart-byte tests via `core-contract` profile and the E2E archive |
 | PR, non-doc change | `S3 Implemented Tests` | `ci.yml` `s3-implemented-tests` | Via aggregate | build `rustfs`, then `scripts/s3-tests/run.sh` with the job's `DEPLOY_MODE` / `TEST_MODE` / `MAXFAIL` env |
@@ -117,18 +117,40 @@ Functional-chain health runs on a hosted runner independently of the lab fleet.

 ## Release validation

-Post-merge and tag-driven; not a substitute for a PR gate.
+Post-merge and tag-driven; not a substitute for a PR gate. Every public publication exit consumes the release certification described here; this certification path is separate from the reusable CI artifact contract in [verification plans](ci-verification-pipeline.md), whose archives expire after one day and are never used as release certification.
+
+The versioned plan is `.config/release-validation-plan.json` (`schema_version=1`): release types, target OS/arch/ABI, supported features, a fixed harness SHA per lane, required case IDs, environment capabilities, timeouts, applicability conditions, and an explicit `upgrade_sources` list carrying the real artifact digests of the supported N-1/N-2 sources plus a stated rollback basis — compatibility and rollback are never invented or defaulted. `scripts/release/release_gate.py validate` expands and freezes `plan_sha256` and the required case set before reading results, then binds that plan, a `candidate.json` (product 40-character SHA, version, release profile, features, immutable deliverable locations with SHA256 and wrapped-binary SHA256, image digest) and the shared G01 evidence-envelope set (rustfs/backlog#2658) of one parent candidate batch, and writes machine JSON, JUnit and a summary. Within a batch the unique key is `(lane, platform{os,arch,abi}, case_id)`; the top-level run/attempt is the parent batch and the nested workflow run/attempt is the child workflow identity, and both are validated separately (workflow SHA and product build SHA are bound independently and need not be equal). Missing, duplicate or extra cases, stale attempts, wrong digests, wrong platforms, wrong harness SHAs, zero tests, cancellation, timeout, untrusted sources and any non-PASS required item block with classified errors; only a complete legal success returns 0. `verify-promotion` re-checks the exact bytes about to be published against the passed list and fails on a swapped package or an edited plan after verification. preview, prerelease and stable all require an explicit plan; preview may use the plan's smaller pre-declared set but never bypasses artifact identity and startup checks. development artifacts are explicitly uncertified and never touch official or latest channels.

 | Trigger | Workflow / job | Result |
 |---|---|---|
-| Nightly schedule, dispatch | `build.yml` `build-rustfs` (development build over the full platform matrix) | build artifacts and the dev download channel; no release publication |
-| Valid release or preview tag | `build.yml` `build-rustfs`, `create-release`, `upload-release-assets`, `publish-release` | draft release, checksummed assets, publish |
-| Successful non-preview release-tag build (`workflow_run`) | `docker.yml` `build-docker`, `scan-docker-image` | multi-architecture images and vulnerability report |
-| Successful release-tag build (`workflow_run`) | `package.yml` `package` | DEB/RPM packages and checksums uploaded to the release |
-| Successful non-preview release-tag build (`workflow_run`) | `helm-package.yml` `build-helm-package`, `publish-helm-package` | versioned chart and repository index |
+| Nightly schedule, dispatch | `build.yml` `build-rustfs` (development build over the full platform matrix) | build artifacts and the uncertified dev download channel; no release publication |
+| Valid release or preview tag | `build.yml` `build-rustfs`, `release-certification`, `create-release`, `upload-release-assets`, `publish-release`, `promote-release-artifacts` | staged binaries, draft release, checksummed assets, then certified publish and promotion of the same bytes |
+| Successful non-preview release-tag build (`workflow_run`) | `docker.yml` `release-certification`, `build-docker`, `scan-docker-image` | multi-architecture images staged, verified against the certification and promoted, plus vulnerability report |
+| Successful release-tag build (`workflow_run`) | `package.yml` `release-certification`, `package`, `promote-packages` | DEB/RPM packages with wrapped-binary verification, staged then promoted with checksums |
+| Successful non-preview release-tag build (`workflow_run`) | `helm-package.yml` `release-certification`, `build-helm-package`, `publish-helm-package` | versioned chart and repository index, only over a certified image promotion record |
 | Final tag's release published | `build.yml` `cleanup-preview-releases` | deletes every `<target>-preview.<N>` Release for that target; the tags are kept |

-Use an exact preview tag for an end-to-end release rehearsal. Manual dispatches are backfill/debug paths and do not prove the automatic `workflow_run` chain.
+Publication exits and their gate:
+
+| Exit | Writer | Certification consumed |
+|---|---|---|
+| GitHub Release draft → public | `build.yml` `publish-release` | `release-certification` success and per-asset `verify-promotion` |
+| Official R2 binary path `artifacts/rustfs/release/` | `build.yml` `promote-release-artifacts` | `release-certification` plus `verify-promotion`; the build job itself only writes `artifacts/rustfs/staging/<run>/<attempt>/` |
+| `latest.json` version pointer | `build.yml` `update-latest-version` | transitively through `publish-release` |
+| Image channels (Docker Hub, GHCR, Quay tags) | `docker.yml` `build-docker` promotion step | certified input digests, staged image binary identity and `verify-promotion`, then promotion of the same digest with `imagetools create` |
+| Official R2 package paths `packages/release` and `packages/latest` | `package.yml` `promote-packages` | `release-certification` and wrapped-binary `verify-promotion`; the package job only stages |
+| GitHub release package assets | `package.yml` `package` | `release-certification` and wrapped-binary `verify-promotion` before upload |
+| Helm chart repository (`rustfs/helm` push) | `helm-package.yml` `publish-helm-package` | `release-certification` plus the Docker workflow's image-promotion record bound to the same candidate |
+
+Downstream workflows fail closed when the trusted certification is absent: a manual dispatch without certification fails and pushes no official tag. Only results from the trusted verification sources listed in the plan (currently the `rustfs/s3chaos` `release-gate.yml` envelope runs) carry release authority; a PR-uploaded JSON set cannot certify a release. Retries reuse the same candidate, only the staged bytes are promoted (no rebuild-after-verify), and no downstream step re-resolves a mutable `latest`.
+
+Use an exact preview tag for an end-to-end release rehearsal. Manual dispatches are backfill/debug paths and do not prove the automatic `workflow_run` chain. The committed fixtures under `testdata/release-gate/` and the commands below are contract rehearsals of this orchestration; they are explicitly not product release evidence and not live verifier results.
+
+```bash
+python3 -m unittest discover -s scripts/release -p 'test_release_gate.py'
+python3 scripts/release/release_gate.py validate --plan testdata/release-gate/plan.json --candidate testdata/release-gate/candidate.json --results testdata/release-gate/pass --output target/release-gate-check
+bash scripts/security/check_preview_release_workflow.sh
+```

 ## Change checklist

@@ -136,7 +158,7 @@ Update this file in the same PR when a job or check name changes, a workflow gai

 ## ECStore invariant selection

-The existing `ci.yml` test-and-lint job runs the ordinary ECStore and filemeta tests. After that run, `scripts/check_test_wiring.py --check-core` checks the same nextest profile and package selection against `.config/ecstore-required-tests.json`. Every named test must exist, match the filter, and be non-ignored; the job also requires a nonempty JUnit report. This checks membership without running the tests twice. `core-test-listing.json`, JUnit, and the run log are retained in the existing test-and-lint artifact.
+The existing `ci.yml` `workspace-tests` job runs the ordinary ECStore and filemeta tests. After that run, `scripts/check_test_wiring.py --check-core` checks the same nextest profile and package selection against `.config/ecstore-required-tests.json`. Every named test must exist, match the filter, and be non-ignored; the job also requires a nonempty JUnit report. This checks membership without running the tests twice. `core-test-listing.json`, JUnit, and the run log are retained in the existing `artifacts/workspace` evidence directory.

 The manifest records a minimum set of invariants: write quorum, metadata rollback, stale-writer lock loss, plaintext Range content, multipart cancellation, hiding uncommitted LIST versions, real MinIO metadata, corrupt part arrays, and the shared on-demand-migration source-backend contract for each provider dialect (S3, Azure, native GCS). The three contract entries live in the `rustfs` suite and reach the lane through that package's default features, so dropping `gcs` from `rustfs`'s defaults fails this check instead of silently deselecting the GCS contract (rustfs/backlog#2323). Renaming or moving a required test must update the manifest in the same change after checking the compiled listing. Extend this list as new deterministic regressions land; it is not a claim that all storage invariants are covered.

@@ -0,0 +1,72 @@
+# CI timing sample 2026-10-31 (backlog #2963 CI01 measurement)
+
+Operational measurement of ordinary-code PR feedback and runner occupancy against the CI01 experimental targets (feedback median 15–20 min, P95 ≤ 30 min, paid occupancy median 50–60 runner minutes). Methodology: [ci-timing.md](ci-timing.md) and [ci-resource-accounting.md](ci-resource-accounting.md). Raw snapshots: `/tmp/ci-timing.json` (limit 30 reproduction), `/tmp/ci-timing-200.json` (analysis sample), `/tmp/ci-acc-w{1,3}.json`. Nothing in this report is an acceptance result; the targets are **not met** by the measured sample.
+
+## Sample window and n
+
+| Item | Value |
+|---|---|
+| Collection time | 2026-10-31 ~02:50–03:15 UTC |
+| Window | completed `pull_request` runs of `ci.yml` created in the last 7 days |
+| Runs sampled | 30 (reproduction) and 200 (analysis, script max) |
+| Successful-code runs (classifier) | 1 of 30; **39 of 200** |
+| Other outcomes in 200-run sample | 68 cancelled (34 %), 36 failed, remainder non-code greens/closure handlers |
+| Lane shapes in sample | 24 combined-lane runs (`Workspace Test and Lint`), 15 split-lane runs (`Workspace Lint and Build` + `Workspace tests` + `Core contract`) |
+
+The limit-30 reproduction reports median wall 76.85 min and median job-sum 133.33 runner-min on n=1 (per methodology, an absent/small median is not zero). All medians/P95 below are from the n=39 analysis sample. Two of the 39 are rerun continuations (run_attempt > 1, wall 7.0 and 22.6 min re-executing only failed jobs under the rerun-start convention); they do not move the median.
+
+## Measured feedback and occupancy (ordinary successful code runs, n=39)
+
+| Metric | Median | P90 | P95 | Max |
+|---|---|---|---|---|
+| Wall feedback (created/rerun-start → last job end) | **76.3 min** | 94.5 | **97.0** | 103.2 |
+| Runner-minute job sum per run | **179.5** | 206.5 | **213.6** | 220.4 |
+
+Independent accounting cross-check (`scripts/ci/actions_cost_report.py`, 2026-10-30T00–12Z window, 547 attempts, collection **incomplete** — see gaps): successful-code PR feedback median 88.34 min (n=12), runner_sum median 199.68 min (n=12). A complete half-day window (2026-10-31T00–12Z, 78 attempts) shows total known runner occupancy 1379.72 min with 42 jobs of unknown occupancy; queue wait summed 28.43 min, so queueing is not the dominant cost.
+
+## Per-job critical path breakdown
+
+Execution minutes across the 39 runs (known only):
+
+| Job | n | Median | P95 |
+|---|---|---|---|
+| Workspace Test and Lint (combined lane) | 24 | 63.6 | 73.9 |
+| Connect Feature Boundaries | 38 | 53.5 | 70.4 |
+| Workspace Lint and Build (split lane) | 15 | 43.3 | 56.5 |
+| Workspace tests (split lane) | 15 | 29.2 | 43.0 |
+| Build RustFS Debug Binary | 39 | 18.5 | 24.9 |
+| End-to-End Tests | 39 | 11.4 | 16.7 |
+| S3 Implemented Tests | 39 | 9.8 | 11.1 |
+| io_uring Integration (real) | 39 | 7.7 | 8.8 |
+| Quick Checks | 39 | 7.2 | 9.2 |
+
+Per-run chain timeline (start = attempt start; see `/tmp/per-run-jobs.json`):
+
+- Split lane (15 runs): `Quick Checks` (~7–10) → `Workspace Lint and Build` (ends ~46–67) → `Workspace tests` (ends ~71–103) → shard evidence → aggregator `Test and Lint`/`required-checks`. The lint→tests chain sets the wall in most split runs.
+- Combined lane (24 runs): `Quick Checks` → `Workspace Test and Lint` alone runs 63–86 min and sets the wall.
+- `Connect Feature Boundaries` (median 53.5, P95 70.4) runs parallel to the lint chain and occasionally becomes the tail (e.g. runs 41021787700, 40750122159: connect end within 0.2 min of wall).
+- `Build RustFS Debug Binary` → E2E/Core-contract/S3 chain finishes with 30–60 min of slack in every sampled run.
+- The required-check aggregator `Test and Lint` (`required-checks`, `ci_gate.py verify`) finishes last by construction (~0.2 min).
+
+## Against the CI01 targets
+
+| Target | Measured | Verdict |
+|---|---|---|
+| Feedback median 15–20 min | 76.3 min (n=39) | **not met** (~3.8–5× over) |
+| Feedback P95 ≤ 30 min | 97.0 min | **not met** (~3.2× over) |
+| Occupancy median 50–60 runner-min | 179.5 runner-min | **not met** (~3–3.6× over) |
+
+These are the experimental-target measurements. The full merge-lifecycle cost is separate and was not established here (the accounting PR-allocation buckets were empty in the partial window; see gaps).
+
+## Bottlenecks (honest list)
+
+1. **Workspace lint+test compile/test chain (critical path):** combined `Workspace Test and Lint` median 63.6 min, or split `Workspace Lint and Build` 43.3 + `Workspace tests` 29.2 ≈ 72.5 min after Quick Checks. This is the wall clock.
+2. **Connect Feature Boundaries (parallel near-critical):** median 53.5 min on its own lane with `CARGO_BUILD_JOBS: "2"` limiting build concurrency on the credential-test step; becomes the tail when the lint chain is fast.
+3. **Duplicate compilation across lanes:** `Workspace Lint and Build` builds workspace test binaries; `Build RustFS Debug Binary` separately builds the debug binary + E2E binaries on another runner; Connect builds again. `cache-save-if: 'false'` on the Connect lane means that lane never warms the shared `ci-dev` cache.
+4. Runner queueing is small (28.43 queue-min vs 1379.72 runner-min in the complete half-day window) — not the bottleneck.
+
+## Gaps and caveats
+
+- Accounting windows: 2026-10-30T00–12Z collected 547 attempts but is flagged incomplete; 2026-10-30T12–24Z collection did not finish (timeout) and has no report; 2026-10-31T12–24Z was empty at collection time. Occupancy figures are per-window partial sums and must not be summed as a bill.
+- No CI workflow or selection change was made in this pass; strict gate semantics unchanged. No savings are claimed. Candidate fixes (build-artifact reuse between lanes, enabling cache save on the Connect lane, finer affected-selection granularity in shadow mode per ci-timing.md) are **unmeasured proposals**, not results.
+- Timing data alone does not establish functional coverage or escaped-regression rate; the 15–20 min target remains an experiment that currently misses by a wide margin.
@@ -32,7 +32,7 @@ Periodic and manual suites use the responsibility registry and retain the existi

 ## Merge-queue preparation and rollback

-The merge-group event plans and validates its actual combined SHA/tree with the full Linux and feature configuration matrix. Main pushes keep complete validation; no reuse is inferred from tree equality because commit identity and build inputs can matter. CLA remains an independent queue-entry requirement.
+The merge-group event plans and validates its actual combined SHA/tree with the default workspace matrix: workspace lint and tests, the short core contract, the debug build, E2E smoke, both S3 lanes and the full merge-gate E2E. The specialized boundary, hardware and alternate-feature lanes (Connect Feature Boundaries, real io_uring, rio-v2, Swift, SFTP, HTTP/3) run only in the scheduled and manual full-validation runs and are required to be skipped here. Main pushes keep complete validation; no reuse is inferred from tree equality because commit identity and build inputs can matter. CLA remains an independent queue-entry requirement.

 `scripts/ci_merge_queue.py` prepares a bounded queue proposal from a live ruleset snapshot without calling a mutation API. It preserves all existing checks, reviews, bypass actors and branch conditions, adds an `ALLGREEN` squash queue with at most two concurrent builds/two merged entries, and produces the original rollback payload. Its timeout accommodates the existing full gate while measurements are collected. API fields follow [GitHub's ruleset contract](https://docs.github.com/en/rest/repos/rules).

@@ -142,7 +142,7 @@ filtered and malformed listing cases, and missing or altered required fixtures.
 Those cases must fail; an optional legacy fixture test that returns early is
 not evidence for a required fixture.

-The `test-and-lint` artifact contains the native log, listing, JUnit and
+The `artifacts/workspace` evidence directory contains the native log, listing, JUnit and
 execution receipt. A native test that reopens a disk in the same process does
 not establish process-crash or power-loss durability.

@@ -16,20 +16,32 @@ Use the `executable` from the scanner library test `compiler-artifact` JSON reco
 python3 scripts/diagnose_scanner_enumeration_restart.py \
  --test-binary /path/to/compiled/scanner-libtest \
  --output /tmp/scanner-enumeration-new-run \
-  --objects 128 --raw-entry-budget 8 --rounds 8
+  --objects 128 --raw-entry-budget 8 --rounds 40
 ```

+The round count must cover both stages: at `E` first-seen entries and `O` objects with budget `B`, enumeration takes `E/B` rounds and classification about `O/(B-1)` more, so 128 objects at budget 8 converge in 35 rounds. `--rounds 8` cannot converge at that budget on any correct scanner and reports exit 2 because no round reached classification.
+
 The output directory must not exist. Each round starts a new OS test-worker process, opens the same synthetic disk, decodes the preceding cache, invokes the real scanner, encodes the returned cache, and decodes it again. When cancellation returns no useful partial cache, it preserves the previous cache. Reports identify the actual child PID, round, raw entries and name bytes observed, raw page-index parent/entries, processed/classified objects, retained object/version/byte counts, and completeness. The driver rejects retained coverage that advances beyond classified object work, root raw page indexes that outrun the fixture namespace, committed page coverage that exceeds indexed coverage, same-parent committed coverage regressions before completion, and any process-restart regression in retained coverage. No observed-name set, `readdir` offset, or assumed stable ordering is used as durable progress. Namespace creation happens only during fixture setup, before scan accounting.

-The `cfg(test)` hook observes actual entries delivered by `read_dir` and cancels the existing cycle token at the fixed entry limit. This is a deterministic injected **raw-entry work budget**, not a wall-clock performance measurement or a claim that kernel prefetch, probes, allocations, name bytes, or cache I/O are independently budgeted. The watchdog timeout only bounds worker lifetime. The hook does not replace enumeration, classification, or recursion, and does not exist in production builds. In particular, `xl.meta` object-boundary classification is unchanged.
+The worker sets `max_entries` on the production `ScannerCycleBudget`, and the `cfg(test)` hook only observes actual entries delivered by `read_dir`. This is a deterministic injected **raw-entry work budget**, not a wall-clock performance measurement or a claim that kernel prefetch, probes, allocations, or cache I/O are independently budgeted. Entries replayed up to the size of a validated index are read again on every restart and are bounded only by the runtime budget. The watchdog timeout only bounds worker lifetime. The hook does not replace enumeration, classification, or recursion, and does not exist in production builds. In particular, `xl.meta` object-boundary classification is unchanged.

-Exit 0 requires exact complete object/version/byte coverage within the same fixed budget on every executed round and positive evidence for all three stages: raw enumeration/indexing, object classification/processing, and durable cache retention after a fresh worker process reloads the previous report. Exit 1 means the strict convergence oracle remains unmet, including the current flat-directory enumeration starvation case. Exit 2 means invalid input, worker failure, or invalid evidence; it is not a successful reproduction. There is no final unbudgeted sweep. Small fixtures can pass; that does not establish the general R-E gate from [the scanner review comment](https://github.com/rustfs/backlog/issues/2240#issuecomment-5549222480). Raw entries observed are not a retained enumeration watermark. This is scanner-worker process restart plus codec evidence, **not** whole-daemon restart, EC quorum persistence, crash/fsync durability, remote RPC, or a throughput benchmark. The caller owns the bounded evidence directory and may remove it after inspection.
+Exit 0 requires exact complete object/version/byte coverage within the same fixed budget on every executed round and positive evidence for all three stages: raw enumeration/indexing, object classification/processing, and durable cache retention after a fresh worker process reloads the previous report. Exit 1 means the strict convergence oracle remains unmet. Exit 2 means invalid input, worker failure, or invalid evidence; it is not a successful reproduction. There is no final unbudgeted sweep. Small fixtures can pass; that does not establish the general R-E gate from [the scanner review comment](https://github.com/rustfs/backlog/issues/2240#issuecomment-5549222480). Raw entries observed are not a retained enumeration watermark. This is scanner-worker process restart plus codec evidence, **not** whole-daemon restart, EC quorum persistence, crash/fsync durability, remote RPC, or a throughput benchmark. The caller owns the bounded evidence directory and may remove it after inspection.
+
+### Budgeted First Build
+
+`FolderScanner::scan_folder` streams `read_dir`, charges each entry past the restored index size to the budget, and records it in the raw page index. Object metadata stays on the object budget. `max_entries` and `max_entry_name_bytes` stop a scan directly; `max_objects` and `max_directories` also bound one directory listing to 16 entries per budget unit, with a 128-entry floor, because a forward sweep needs the whole listing before it classifies in sorted order. A directory keeps its index until its children end, and the checkpoint saves every index on the active walk path, so a restart does not recharge any wide ancestor. A stop checkpoint grows only by the charged names. The total index stays proportional to the directory width.
+
+### Bucket Service Order
+
+Each set cache carries an optional `scan_bucket_service` map: a first-in-first-out wait position per bucket with its granted quanta, last certification and failure debt. A new or recreated bucket joins behind waiters but ahead of buckets that already had a turn in the current generation. The order is saved after budget stops and finished incomplete cycles too, under the set cache lock and revisions, and never when the owner was cancelled. It orders work and authorizes nothing, so another movement epoch or an invalid value starts a new order. Service recorded after the last save is not durable: a crash or cancellation inside a cycle repeats that cycle's order. Set dispatch order across restarts is not persisted.
+
+`cold_bucket_service_` tests fix N=17 buckets and q=4 service quanta with one hot and one failing bucket, and require service within ceil(17/4)=5 rounds through the real worker queue, the saved codec, and fresh OS processes. `fair_prefix_quantum_` tests inject small bucket shares. Production keeps `bucket_quantum` unset, so a bucket may use the total budget and then moves to the tail.

 ### Missing Storage Capability

-The current `scanner_folder::FolderScanner::scan_folder` collects child folders before recursing. `LocalDisk::scan_dir` also reads the whole parent before sorting and applying `forward_to`. Neither can supply a restartable first-build cursor over per-disk raw entries, orphan directories, and metadata boundaries. Repeated listing from the beginning is real work, not free pagination.
+`read_dir` has no restartable cursor, so resuming a listing rereads it from the start. Repeated listing from the beginning is real work, not free pagination.

-A future storage-owner capability must expose an explicit unsupported/building/ready state and a durable snapshot/index identity bound to disk mount, bucket incarnation, and directory identity. It must budget the first build and every page, including entry count, name bytes, metadata probes, I/O and time; survive a process restart during first build; seal page data before advancing the manifest; and distinguish enumerated, classified, and fully processed frontiers. An uncommitted page may be replayed only within a bounded cost. `xl.meta` classification must finish before descendants become traversable namespace. Missing capability or invalid identities must not become fabricated progress or completeness. No such capability is implemented by this diagnostic, and ordinary local storage remains without this R-E guarantee.
+A future storage-owner capability must expose an explicit unsupported/building/ready state and a durable snapshot/index identity bound to disk mount, bucket incarnation, and directory identity. It must budget the first build and every page, including entry count, name bytes, metadata probes, I/O and time; survive a process restart during first build; seal page data before advancing the manifest; and distinguish enumerated, classified, and fully processed frontiers. An uncommitted page may be replayed only within a bounded cost. `xl.meta` classification must finish before descendants become traversable namespace. Missing capability or invalid identities must not become fabricated progress or completeness. No such capability is implemented by this diagnostic. The page index above budgets new entries only, and ordinary local storage remains without this R-E guarantee.

 ## Completed Subtree Checkpoint Fixture

@@ -2,11 +2,11 @@
  "nodes": {
    "nixpkgs": {
      "locked": {
-        "lastModified": 1792812153,
-        "narHash": "sha256-T2+XBfNbf9pecLUWS4+RtuVNZ3W8K5taG5QdV71mWlk=",
+        "lastModified": 1793466162,
+        "narHash": "sha256-pt+peAEoOK3lLLdtbnQ1AGIBTZbdB6jrkgmurfcQwig=",
        "owner": "NixOS",
        "repo": "nixpkgs",
-        "rev": "674a2028661e550a97400c1cc992b9cca9a668a1",
+        "rev": "a0788123ada57ec92a79ab2581cda6d3c6cca75f",
        "type": "github"
      },
      "original": {
@@ -29,11 +29,11 @@
        ]
      },
      "locked": {
-        "lastModified": 1792825530,
-        "narHash": "sha256-aphXA2IFNd3q48++TKmfuuvDDjkrgGdJDnDIhij9684=",
+        "lastModified": 1793434469,
+        "narHash": "sha256-+sgVNI2c67di+XIeMaz5KSei4T65kX68s5Nl0E5Gzbc=",
        "owner": "oxalica",
        "repo": "rust-overlay",
-        "rev": "5b9f1198b59bcb31c5c3250b332e7d77590e6306",
+        "rev": "6ecf534a7c16e7c969f26838fdd8c93597ae93f3",
        "type": "github"
      },
      "original": {
@@ -495,9 +495,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-schema"
-version = "1.1.0"
+version = "1.1.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bec8dae1015a6d959a5faf86fbdf77486cfe19e4ccb6ee009d8a46efc45028a4"
+checksum = "9543fc722ed4865ab29fb66bd7ed0bef7b029d265ba8c587144511e30006d431"
 dependencies = [
 "aws-smithy-runtime-api",
 "aws-smithy-types",
@@ -1418,9 +1418,9 @@ dependencies = [

 [[package]]
 name = "encoding_rs"
-version = "0.8.42"
+version = "0.8.44"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8e985e0451871ad22fb8d2b6b076e2028a502a0d3950998c2c5c0a4f9b5d9679"
+checksum = "89253175e704ab869b04f9db224d38630e0211a53841a79181deae0889e169f5"
 dependencies = [
 "cfg-if",
 "core_detect",
@@ -1837,18 +1837,18 @@ dependencies = [

 [[package]]
 name = "hotpath"
-version = "0.30.1"
+version = "0.30.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "502dc37a7c218b9b1d3a0ccfb904da81d10d6dda0e899c722c503092d55576ea"
+checksum = "b39e941010d86063ff233bf8325269bf68097fb24ba67000fbddc0a410e8056e"
 dependencies = [
 "hotpath-macros",
 ]

 [[package]]
 name = "hotpath-macros"
-version = "0.30.1"
+version = "0.30.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a76e0af911930fc97abc766b393b48cdca77fea9f5f6b388cf2123a6d9c30deb"
+checksum = "a24f5179ff680132003306354efd0ac8cde12e87b1a788719c836e91af35cb58"

 [[package]]
 name = "http"
@@ -2305,9 +2305,9 @@ dependencies = [

 [[package]]
 name = "js-sys"
-version = "0.3.106"
+version = "0.3.107"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7883d941dae510fb2d978fc3fe018c71c9e2892fd38854de3e8b92c2e5ad9cc5"
+checksum = "8dc09d2154098bbf4ef54eab6b117a3c24cbbffe94046b16ee6d6d4510584d3a"
 dependencies = [
 "cfg-if",
 "futures-util",
@@ -3162,9 +3162,9 @@ dependencies = [

 [[package]]
 name = "pulldown-cmark"
-version = "0.13.5"
+version = "0.13.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "278e9268c93b9e498a2da46c34d90ba93a3255eb99009646398f34bc73ab36eb"
+checksum = "3db2be46835f013dd3f4fa7f5f7402c537787b54ebb6fe2b94bbdd7e2669bb7b"
 dependencies = [
 "bitflags 2.13.2",
 "memchr",
@@ -4822,9 +4822,9 @@ dependencies = [

 [[package]]
 name = "sysinfo"
-version = "0.40.0"
+version = "0.40.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "110e37d98880fd804cd47509b90a8444f7aaaab655ae625e83cd442a8bc2ebf2"
+checksum = "0e8a3edbe0394ce115f3917a00349dd242be09eb5a091ed509eaf67b55cfb8a8"
 dependencies = [
 "libc",
 "memchr",
@@ -5017,9 +5017,9 @@ dependencies = [

 [[package]]
 name = "tonic"
-version = "0.14.8"
+version = "0.14.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cd1fa9b1a51188f3266a823208078315811269d06d272c3a61be0b9a194194ff"
+checksum = "e2cdfc673aabd30fe8fee92e65837b644c37c5d47f0f97ae6d5e0f1237d15e1b"
 dependencies = [
 "async-trait",
 "axum",
@@ -5050,9 +5050,9 @@ dependencies = [

 [[package]]
 name = "tonic-build"
-version = "0.14.8"
+version = "0.14.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2c07bd4e9f99e2a3ed8bebad596ae02a94e53745b68acc6900e22c3475c8bc44"
+checksum = "ac6d916d206ed7bdda1778ddff9bda43303789c258877e7bc22fb44ee5c984b4"
 dependencies = [
 "prettyplease",
 "proc-macro2",
@@ -5062,9 +5062,9 @@ dependencies = [

 [[package]]
 name = "tonic-prost"
-version = "0.14.8"
+version = "0.14.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f63c616ffaa648f7c5dd5657e97d9a285fcb44ff3b70da761d436f5a13ff7be6"
+checksum = "eba464054e2226f182a34489b8b079bf4c6e3e7ab5fd7fc30e024f24aa92a484"
 dependencies = [
 "bytes",
 "prost",
@@ -5073,9 +5073,9 @@ dependencies = [

 [[package]]
 name = "tonic-prost-build"
-version = "0.14.8"
+version = "0.14.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4a05d8957fac6b3542999796cb4449d45da790751962f974a457e5e00c43d78c"
+checksum = "2ea216ddc448e08e7839ab09e20e60b06b6a1c32745ccb212c0ebd161c199f1b"
 dependencies = [
 "prettyplease",
 "proc-macro2",
@@ -5293,9 +5293,18 @@ checksum = "daf8dba3b7eb870caf1ddeed7bc9d2a049f3cfdfae7cb521b087cc33ae4c49da"

 [[package]]
 name = "utf8_iter"
-version = "1.0.4"
+version = "1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b6c140620e7ffbb22c2dee59cafe6084a59b5ffc27a8859a5f0d494b5d52b6be"
+checksum = "af3bd9d3e571c26fcce9d033b87ea5ab061b8d2ffe9b7a5675d810d68f03edf0"
+dependencies = [
+ "utf_types",
+]
+
+[[package]]
+name = "utf_types"
+version = "1.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3a3b425286980a367278453f3b6d28cdcb134b574e5a35824a6ff5eebad50bb6"

 [[package]]
 name = "uuid"
@@ -5355,12 +5364,11 @@ checksum = "ccf3ec651a847eb01de73ccad15eb7d99f80485de043efb2f370cd654f4ea44b"

 [[package]]
 name = "wasm-bindgen"
-version = "0.2.129"
+version = "0.2.130"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9bb54f33acc68fd454578d9820b0bde1a1a3d17aa17bb7b6595806d02886d409"
+checksum = "a0181c14ddd6c7b8b0b0132a5f1ea2f32be0d893cd3ea97b6e3b28886b8a3d1c"
 dependencies = [
 "cfg-if",
- "once_cell",
 "rustversion",
 "wasm-bindgen-macro",
 "wasm-bindgen-shared",
@@ -5368,9 +5376,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-futures"
-version = "0.4.79"
+version = "0.4.80"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3cbab34de2d982e9b48e18d216d04c4a6f641066ff19ffb699980f591ee3610e"
+checksum = "5e14e0327cb5368ad60ccb22d04a9827a693c95a926d8d125e9fa39e0baec501"
 dependencies = [
 "js-sys",
 "tokio",
@@ -5379,9 +5387,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-macro"
-version = "0.2.129"
+version = "0.2.130"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2e29d0c35b16e224a7eeb5cd2d25e3e1968fbd65604117b44d3b789d00ee8535"
+checksum = "6e931a02a39abdbf66966fc35902b5ed58344750a70885bed4ebb3d8eaf28ed9"
 dependencies = [
 "quote",
 "wasm-bindgen-macro-support",
@@ -5389,9 +5397,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-macro-support"
-version = "0.2.129"
+version = "0.2.130"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6f501a8bc3719dba86ef8ae4728879c08001bea749eb1333ac5b91e040e2a6b7"
+checksum = "2ca4b54280e455bd831e33a6b8e734baf432fe63fbc35c42ac14ce67a043a53a"
 dependencies = [
 "bumpalo",
 "proc-macro2",
@@ -5402,9 +5410,9 @@ dependencies = [

 [[package]]
 name = "wasm-bindgen-shared"
-version = "0.2.129"
+version = "0.2.130"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "23f0c9c52aa7cd7d77769a4cfe2a9adb1b331f489a41d912ce14513d5ab995c6"
+checksum = "3b1f9c3798725016c6ee5f2f57869ec0831e8bbc0ab0310a86e787e97f41389e"
 dependencies = [
 "unicode-ident",
 ]
@@ -5424,9 +5432,9 @@ dependencies = [

 [[package]]
 name = "web-sys"
-version = "0.3.106"
+version = "0.3.107"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "88261b9deccee56594c11a3460c462c41f58d148598fe70ad77070126a68aba4"
+checksum = "d7d179596ff5ec0f6de4db64d54011bde6289cc49603d3a6e4ce3f329207ab88"
 dependencies = [
 "js-sys",
 "wasm-bindgen",
@@ -433,6 +433,47 @@ async fn config_handler(uri: Uri, headers: HeaderMap) -> impl IntoResponse {
        .unwrap()
 }

+/// Buffers brotli-encoded responses to add a `Content-Length` header.
+///
+/// `tower_http`'s `CompressionLayer` streams brotli via `Transfer-Encoding: chunked`
+/// without a `Content-Length`. Firefox's incremental brotli decoder aborts on this
+/// combination (`NS_BINDING_ERROR`). Chrome and Safari buffer the full chunked body
+/// before decompressing and are unaffected.
+///
+/// This middleware collects the complete body for any brotli response that lacks a
+/// `Content-Length`, then re-emits it with the correct length so Firefox can decode it.
+pub(crate) async fn buffer_brotli_responses(request: Request, next: middleware::Next) -> Response {
+    let response = next.run(request).await;
+
+    let is_brotli = response
+        .headers()
+        .get(http::header::CONTENT_ENCODING)
+        .and_then(|v| v.to_str().ok())
+        .is_some_and(|v| v == "br");
+
+    if !is_brotli || response.headers().contains_key(http::header::CONTENT_LENGTH) {
+        return response;
+    }
+
+    let (mut parts, body) = response.into_parts();
+    match axum::body::to_bytes(body, 32 * 1024 * 1024).await {
+        Ok(bytes) => {
+            if let Ok(value) = HeaderValue::from_str(&bytes.len().to_string()) {
+                parts.headers.insert(http::header::CONTENT_LENGTH, value);
+            }
+            Response::from_parts(parts, Body::from(bytes))
+        }
+        Err(err) => {
+            warn!(
+                error = %err,
+                "Failed to buffer brotli console response; sending empty body without Content-Encoding"
+            );
+            parts.headers.remove(http::header::CONTENT_ENCODING);
+            Response::from_parts(parts, Body::empty())
+        }
+    }
+}
+
 /// Console access logging middleware
 /// Logs each console access with method, URI, status code, and duration.
 ///
@@ -605,6 +646,8 @@ fn setup_console_middleware_stack(
        .layer(SetRequestIdLayer::x_request_id(MakeRequestUuid))
        // Compress responses
        .layer(CompressionLayer::new())
+        // Buffer brotli responses to add Content-Length
+        .layer(middleware::from_fn(buffer_brotli_responses))
        .layer(middleware::from_fn(console_logging_middleware))
        .layer(cors_layer)
        // Add timeout layer - convert auth_timeout from seconds to Duration
@@ -43,6 +43,163 @@ mod tests {
            .map(|value| value.to_str().expect("allow-origin header must be valid UTF-8").to_string())
    }

+    /// Sends a brotli-encoded response through `buffer_brotli_responses` and returns
+    /// the `Content-Length` header value and body bytes.
+    async fn run_brotli_middleware(body: &'static str, has_content_length: bool) -> (Option<String>, bytes::Bytes) {
+        use crate::admin::console::buffer_brotli_responses;
+        use axum::middleware;
+        use axum::{Router, body::Body, routing::get};
+        use tower::ServiceExt;
+
+        let app = Router::new()
+            .route(
+                "/",
+                get(move || async move {
+                    let mut builder = http::Response::builder().header(http::header::CONTENT_ENCODING, "br");
+                    if has_content_length {
+                        builder = builder.header(http::header::CONTENT_LENGTH, body.len().to_string());
+                    }
+                    builder.body(Body::from(body)).expect("response must build")
+                }),
+            )
+            .layer(middleware::from_fn(buffer_brotli_responses));
+
+        let response = app
+            .oneshot(
+                http::Request::builder()
+                    .uri("/")
+                    .body(Body::empty())
+                    .expect("request must build"),
+            )
+            .await
+            .expect("request must not fail");
+
+        let content_length = response
+            .headers()
+            .get(http::header::CONTENT_LENGTH)
+            .map(|value| value.to_str().expect("content-length header must be valid UTF-8").to_string());
+        let body_bytes = axum::body::to_bytes(response.into_body(), 1024 * 1024)
+            .await
+            .expect("body must be readable");
+        (content_length, body_bytes)
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn test_buffer_brotli_responses() {
+        // Brotli response without Content-Length must have Content-Length added with correct value.
+        let (content_length, body) = run_brotli_middleware("fake-brotli-data", false).await;
+        assert_eq!(
+            content_length.as_deref(),
+            Some("16"),
+            "brotli response without Content-Length must have Content-Length set to the body length"
+        );
+        assert_eq!(
+            body.as_ref(),
+            b"fake-brotli-data",
+            "body bytes must be preserved through the buffer round-trip"
+        );
+
+        // Brotli response that already carries Content-Length must pass through unchanged.
+        let (content_length, body) = run_brotli_middleware("data", true).await;
+        assert_eq!(
+            content_length.as_deref(),
+            Some("4"),
+            "brotli response that already has Content-Length must not be re-buffered"
+        );
+        assert_eq!(body.as_ref(), b"data", "body bytes must be preserved when passing through unchanged");
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn test_buffer_brotli_responses_non_brotli_passthrough() {
+        use crate::admin::console::buffer_brotli_responses;
+        use axum::middleware;
+        use axum::{Router, body::Body, routing::get};
+        use tower::ServiceExt;
+
+        // Non-brotli response must pass through with no modification.
+        let app = Router::new()
+            .route(
+                "/",
+                get(|| async {
+                    http::Response::builder()
+                        .header(http::header::CONTENT_ENCODING, "gzip")
+                        .body(Body::from("gzip-data"))
+                        .expect("response must build")
+                }),
+            )
+            .layer(middleware::from_fn(buffer_brotli_responses));
+
+        let response = app
+            .oneshot(
+                http::Request::builder()
+                    .uri("/")
+                    .body(Body::empty())
+                    .expect("request must build"),
+            )
+            .await
+            .expect("request must not fail");
+
+        assert_eq!(
+            response
+                .headers()
+                .get(http::header::CONTENT_ENCODING)
+                .and_then(|v| v.to_str().ok()),
+            Some("gzip"),
+            "non-brotli Content-Encoding must pass through unchanged"
+        );
+        let body_bytes = axum::body::to_bytes(response.into_body(), 1024 * 1024)
+            .await
+            .expect("body must be readable");
+        assert_eq!(body_bytes.as_ref(), b"gzip-data", "body bytes must pass through unchanged");
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn test_buffer_brotli_responses_body_stream_error() {
+        use crate::admin::console::buffer_brotli_responses;
+        use axum::middleware;
+        use axum::{Router, body::Body, routing::get};
+        use tower::ServiceExt;
+
+        // A brotli response whose body stream errors must have Content-Encoding stripped
+        // so the client does not attempt to decode an empty body as brotli.
+        let app = Router::new()
+            .route(
+                "/",
+                get(|| async {
+                    let broken = futures::stream::once(async {
+                        Err::<bytes::Bytes, std::io::Error>(std::io::Error::other("simulated stream error"))
+                    });
+                    http::Response::builder()
+                        .header(http::header::CONTENT_ENCODING, "br")
+                        .body(Body::from_stream(broken))
+                        .expect("response must build")
+                }),
+            )
+            .layer(middleware::from_fn(buffer_brotli_responses));
+
+        let response = app
+            .oneshot(
+                http::Request::builder()
+                    .uri("/")
+                    .body(Body::empty())
+                    .expect("request must build"),
+            )
+            .await
+            .expect("request must not fail");
+
+        assert!(
+            response.headers().get(http::header::CONTENT_ENCODING).is_none(),
+            "Content-Encoding must be stripped when body buffering fails so the client does not decode an empty body as brotli"
+        );
+        let body_bytes = axum::body::to_bytes(response.into_body(), 1024 * 1024)
+            .await
+            .expect("body must be readable");
+        assert!(body_bytes.is_empty(), "body must be empty after buffering failure");
+    }
+
    #[tokio::test]
    #[serial]
    async fn test_console_cors_configuration() {
@@ -50,17 +50,23 @@ pub struct DisableTableBucketHandler {}

 #[async_trait::async_trait]
 impl Operation for DisableTableBucketHandler {
-    async fn call(&self, req: S3Request<Body>, params: Params<'_, '_>) -> S3Result<S3Response<(StatusCode, Body)>> {
+    async fn call(&self, mut req: S3Request<Body>, params: Params<'_, '_>) -> S3Result<S3Response<(StatusCode, Body)>> {
        let warehouse = warehouse_from_params(&params)?;
        let resource = TableCatalogResource::warehouse(&warehouse);
-        authorize_table_catalog_resource_request(&req, &resource, AdminAction::SetTableBucketAction).await?;
+        let principal = authorize_table_catalog_resource_request(&req, &resource, AdminAction::SetTableBucketAction).await?;
+        let cleanup = table_bucket_cleanup_requested_from_query(&req.uri)?;
        let app_context = runtime_sources::app_context_from_extensions(&req.extensions);
        let backend = table_catalog_backend_from_extensions(&req.extensions)?;
        let object_store = runtime_sources::object_store_from_req(&req)
            .ok_or_else(|| table_catalog_internal_error("request object store is not initialized"))?;
        let store = table_catalog_store_from_backend(backend.clone())?;
-        let publication = TableCommitObjectBackend::preauthorized(backend);
-        let response = disable_table_bucket_response(&store, &publication, object_store.as_ref(), &warehouse).await?;
+        install_table_catalog_s3_request_info(&mut req, &principal)?;
+        let publication = TableCommitObjectBackend::for_request(backend, req);
+        let response = if cleanup {
+            cleanup_empty_table_bucket_response(&store, &publication, object_store.as_ref(), &warehouse).await?
+        } else {
+            disable_table_bucket_response(&store, &publication, object_store.as_ref(), &warehouse).await?
+        };
        crate::app::bucket_usecase::notify_table_bucket_metadata_reload(warehouse.clone(), "disable table bucket", app_context)
            .await;
        build_json_response(StatusCode::OK, &response)
@@ -2092,17 +2092,25 @@ fn warehouse_from_config_query(uri: &http::Uri) -> S3Result<Option<String>> {
 }

 fn rest_purge_requested_from_query(uri: &http::Uri) -> S3Result<bool> {
-    let mut purge_requested = None;
+    table_catalog_boolean_query(uri, "purgeRequested")
+}
+
+fn table_bucket_cleanup_requested_from_query(uri: &http::Uri) -> S3Result<bool> {
+    table_catalog_boolean_query(uri, "cleanup")
+}
+
+fn table_catalog_boolean_query(uri: &http::Uri, parameter: &str) -> S3Result<bool> {
+    let mut requested = None;
    if let Some(query) = uri.query() {
        for (key, value) in url::form_urlencoded::parse(query.as_bytes()) {
-            if key != "purgeRequested" {
+            if key != parameter {
                continue;
            }
-            if purge_requested.is_some() {
+            if requested.is_some() {
                return Err(iceberg_rest_error(
                    ICEBERG_ERROR_BAD_REQUEST,
                    StatusCode::BAD_REQUEST,
-                    "purgeRequested query parameter must not be repeated",
+                    format!("{parameter} query parameter must not be repeated"),
                ));
            }
            let value = if value.eq_ignore_ascii_case("true") {
@@ -2113,13 +2121,13 @@ fn rest_purge_requested_from_query(uri: &http::Uri) -> S3Result<bool> {
                return Err(iceberg_rest_error(
                    ICEBERG_ERROR_BAD_REQUEST,
                    StatusCode::BAD_REQUEST,
-                    "purgeRequested query parameter must be true or false",
+                    format!("{parameter} query parameter must be true or false"),
                ));
            };
-            purge_requested = Some(value);
+            requested = Some(value);
        }
    }
-    Ok(purge_requested.unwrap_or(false))
+    Ok(requested.unwrap_or(false))
 }

 fn rest_table_snapshot_selection_from_query(uri: &http::Uri) -> S3Result<RestTableSnapshotSelection> {
@@ -2417,9 +2425,10 @@ fn table_catalog_backend_from_extensions(
 ) -> S3Result<crate::table_catalog::EcStoreTableCatalogObjectBackend<ECStore>> {
    let context = runtime_sources::app_context_from_extensions(extensions)
        .ok_or_else(|| table_catalog_internal_error("request application context is not initialized"))?;
-    Ok(crate::table_catalog::EcStoreTableCatalogObjectBackend::new_with_strong_runtime(
+    Ok(crate::table_catalog::EcStoreTableCatalogObjectBackend::new_with_runtimes(
        context.object_store(),
        context.table_catalog_strong_runtime(),
+        context.table_catalog_registration_cache(),
    ))
 }

@@ -2572,6 +2581,61 @@ where
    table_bucket_response(store, bucket, true).await
 }

+async fn cleanup_empty_table_bucket_response<B>(
+    store: &crate::table_catalog::ConfiguredTableCatalogStore<B>,
+    publication: &TableCommitObjectBackend<B>,
+    object_store: &ECStore,
+    bucket: &str,
+) -> S3Result<TableBucketResponse>
+where
+    B: crate::table_catalog::TableCatalogObjectBackend,
+{
+    if !publication.publication_fence_fleet_confirmed {
+        return Err(catalog_store_error(crate::table_catalog::TableCatalogStoreError::Conflict(
+            "empty table bucket cleanup requires publication fences on every cluster node".to_string(),
+        )));
+    }
+    crate::table_catalog::TableCommitPublication::begin_table_bucket(publication, bucket)
+        .await
+        .map_err(catalog_store_error)?;
+    let _publication_completion = crate::table_catalog::TableCommitPublicationCompletion::new(publication);
+    let incarnation = object_store
+        .bucket_incarnation_id_from_disk(bucket)
+        .await
+        .map_err(|err| ApiError::other(format!("failed to read table bucket identity: {err}")))?;
+    let prepared = store
+        .prepare_empty_table_bucket_cleanup(table_bucket_entry_from_metadata_marker(bucket), publication)
+        .await
+        .map_err(catalog_store_error);
+    let mut cleanup = publication.finish(prepared).await?;
+    for object in cleanup.metadata_objects() {
+        if let Err(err) = publication.authorize(bucket, object, S3Action::DeleteObjectAction).await {
+            return publication.finish(Err(catalog_store_error(err))).await;
+        }
+    }
+    store
+        .execute_empty_table_bucket_cleanup(&mut cleanup, incarnation)
+        .await
+        .map_err(catalog_store_error)?;
+    let fence = cleanup.mutation_fence();
+    object_store
+        .update_bucket_metadata_config_for_incarnation_validated(
+            bucket,
+            crate::table_catalog::TABLE_BUCKET_MARKER_CONFIG,
+            Vec::new(),
+            incarnation,
+            || {
+                fence
+                    .ensure_held()
+                    .map_err(|err| crate::admin::storage_api::error::StorageError::other(err.to_string()))
+            },
+        )
+        .await
+        .map_err(|err| ApiError::other(format!("failed to disable table bucket: {err}")))?;
+    drop(cleanup);
+    table_bucket_response(store, bucket, false).await
+}
+
 async fn disable_table_bucket_response<S, B>(
    store: &S,
    publication: &TableCommitObjectBackend<B>,
@@ -87,6 +87,15 @@ async fn call_load_table_handler(request: S3Request<Body>) -> S3Result<S3Respons
    RestLoadTableHandler {}.call(request, matched.params).await
 }

+async fn call_disable_table_bucket_handler(request: S3Request<Body>) -> S3Result<S3Response<(StatusCode, Body)>> {
+    let mut router = matchit::Router::new();
+    router.insert("/iceberg/v1/buckets/{warehouse}", ()).unwrap();
+    router.insert("/_iceberg/v1/buckets/{warehouse}", ()).unwrap();
+    let path = request.uri.path().to_string();
+    let matched = router.at(&path).unwrap();
+    DisableTableBucketHandler {}.call(request, matched.params).await
+}
+
 async fn call_maintenance_worker_handler(request: S3Request<Body>) -> S3Result<S3Response<(StatusCode, Body)>> {
    let mut router = matchit::Router::new();
    router
@@ -1255,6 +1264,28 @@ fn drop_table_purge_query_is_explicit_and_strict() {
    }
 }

+#[test]
+fn table_bucket_cleanup_query_is_explicit_and_strict() {
+    for (uri, expected) in [
+        ("/iceberg/v1/buckets/analytics", false),
+        ("/iceberg/v1/buckets/analytics?cleanup=false", false),
+        ("/iceberg/v1/buckets/analytics?cleanup=true", true),
+        ("/_iceberg/v1/buckets/analytics?cleanup=true", true),
+    ] {
+        assert_eq!(table_bucket_cleanup_requested_from_query(&uri.parse().unwrap()).unwrap(), expected);
+    }
+    for query in [
+        "cleanup=",
+        "cleanup=1",
+        "cleanup=yes",
+        "cleanup=true&cleanup=false",
+        "cleanup=true&cleanup=true",
+    ] {
+        let uri = format!("/iceberg/v1/buckets/analytics?{query}").parse().unwrap();
+        assert!(table_bucket_cleanup_requested_from_query(&uri).is_err(), "invalid query: {query}");
+    }
+}
+
 #[test]
 fn catalog_conflicts_use_operation_specific_iceberg_errors() {
    let already_exists = catalog_store_already_exists_error(crate::table_catalog::TableCatalogStoreError::Conflict(
@@ -1689,9 +1720,7 @@ fn table_bucket_toggle_handlers_reload_peer_metadata_after_publication() {
        (enable, "enable_table_bucket_response", "enable table bucket"),
        (disable, "disable_table_bucket_response", "disable table bucket"),
    ] {
-        let response_offset = block
-            .find(&format!("let response = {response}"))
-            .expect("table bucket toggle should publish local state");
+        let response_offset = block.find(response).expect("table bucket toggle should publish local state");
        let reload = block
            .find(&format!("notify_table_bucket_metadata_reload(warehouse.clone(), \"{operation}\""))
            .expect("table bucket toggle should reload peer metadata");
@@ -16274,3 +16303,799 @@ async fn disable_table_bucket_preserves_objects_and_can_be_reenabled() {
            .table_bucket_enabled()
    );
 }
+
+async fn create_dropped_empty_tables_for_cleanup<S, B>(
+    catalog: &S,
+    backend: &B,
+    object_store: &ECStore,
+    bucket: &str,
+    table_count: usize,
+) -> Vec<String>
+where
+    S: TableCatalogStore + ?Sized,
+    B: TableCatalogObjectBackend,
+{
+    object_store.make_bucket(bucket, &MakeBucketOptions::default()).await.unwrap();
+    enable_table_bucket_response(catalog, &TableCommitObjectBackend::trusted(backend.clone()), object_store, bucket)
+        .await
+        .unwrap();
+    let namespace = crate::table_catalog::Namespace::parse("analytics").unwrap();
+    create_namespace_response(
+        catalog,
+        bucket,
+        CreateNamespaceRequest {
+            namespace: vec!["analytics".to_string()],
+            properties: BTreeMap::new(),
+        },
+        true,
+    )
+    .await
+    .unwrap();
+    let mut metadata_objects = Vec::new();
+    for number in 0..table_count {
+        let name = format!("events_{number:02}");
+        let request = serde_json::from_value(serde_json::json!({
+            "name": name,
+            "schema": {
+                "type": "struct",
+                "schema-id": 0,
+                "fields": [{"id": 1, "name": "id", "required": true, "type": "long"}]
+            }
+        }))
+        .unwrap();
+        create_table_response(
+            catalog,
+            &TableCommitObjectBackend::trusted(backend.clone()),
+            bucket,
+            &namespace,
+            request,
+            true,
+        )
+        .await
+        .unwrap();
+        metadata_objects.push(
+            catalog
+                .load_table(bucket, &namespace.public_name(), &name)
+                .await
+                .unwrap()
+                .unwrap()
+                .metadata_location,
+        );
+    }
+    for number in 0..table_count {
+        catalog
+            .drop_table(bucket, &namespace.public_name(), &format!("events_{number:02}"))
+            .await
+            .unwrap();
+    }
+    catalog.drop_namespace(bucket, &namespace.public_name()).await.unwrap();
+    metadata_objects
+}
+
+#[tokio::test]
+#[serial_test::serial]
+async fn object_catalog_registration_cache_checks_real_records_without_additional_objects() {
+    use crate::admin::storage_api::RUSTFS_META_BUCKET;
+    use crate::table_catalog::{
+        ConfiguredTableCatalogStore, EcStoreTableCatalogObjectBackend, StrongTableCatalogRuntime, TableCatalogBackingMode,
+        TableCatalogObjectPaths, TableCatalogStoreError,
+    };
+
+    let (_temp_dir, _disk_paths, object_store) = crate::app::gating_test_env::isolated_multi_pool_ecstore().await;
+    let bucket = format!("registration-cache-{}", Uuid::new_v4().simple());
+    let backend =
+        EcStoreTableCatalogObjectBackend::new_with_strong_runtime(object_store.clone(), StrongTableCatalogRuntime::default());
+    let catalog = ConfiguredTableCatalogStore::new_for_test(backend.clone(), TableCatalogBackingMode::ObjectBacked);
+    object_store
+        .make_bucket(&bucket, &MakeBucketOptions::default())
+        .await
+        .unwrap();
+    enable_table_bucket_response(&catalog, &TableCommitObjectBackend::trusted(backend.clone()), &object_store, &bucket)
+        .await
+        .unwrap();
+    let namespace = crate::table_catalog::Namespace::parse("analytics").unwrap();
+    create_namespace_response(
+        &catalog,
+        &bucket,
+        CreateNamespaceRequest {
+            namespace: vec!["analytics".to_string()],
+            properties: BTreeMap::new(),
+        },
+        true,
+    )
+    .await
+    .unwrap();
+    let paths = TableCatalogObjectPaths::default();
+    let bucket_entry = paths.table_bucket_entry_path(&bucket);
+    let catalog_prefix = format!("{}/", bucket_entry.rsplit_once('/').unwrap().0);
+    let mut create_micros = Vec::new();
+    for number in 0..100 {
+        let store = ConfiguredTableCatalogStore::new_for_test(backend.clone(), TableCatalogBackingMode::ObjectBacked);
+        let request = serde_json::from_value(serde_json::json!({
+            "name": format!("events_{number:03}"),
+            "schema": {
+                "type": "struct",
+                "schema-id": 0,
+                "fields": [{"id": 1, "name": "id", "required": true, "type": "long"}]
+            }
+        }))
+        .unwrap();
+        let started = std::time::Instant::now();
+        create_table_response(
+            &store,
+            &TableCommitObjectBackend::trusted(backend.clone()),
+            &bucket,
+            &namespace,
+            request,
+            true,
+        )
+        .await
+        .expect("fresh store instances share the runtime cache and create independent tables");
+        create_micros.push(started.elapsed().as_micros());
+        if matches!(number + 1, 20 | 100) {
+            let before = backend.list_objects(RUSTFS_META_BUCKET, &catalog_prefix).await.unwrap();
+            let cold_backend = EcStoreTableCatalogObjectBackend::new_with_strong_runtime(
+                object_store.clone(),
+                StrongTableCatalogRuntime::default(),
+            );
+            let probe = ConfiguredTableCatalogStore::new_for_test(cold_backend, TableCatalogBackingMode::ObjectBacked);
+            let mut duplicate = catalog
+                .load_table(&bucket, "analytics", &format!("events_{number:03}"))
+                .await
+                .unwrap()
+                .unwrap();
+            duplicate.table = "duplicate_probe".to_string();
+            for cache_state in ["cold", "warm"] {
+                let started = std::time::Instant::now();
+                assert!(matches!(
+                    probe.create_table(duplicate.clone()).await,
+                    Err(TableCatalogStoreError::Conflict(_))
+                ));
+                eprintln!(
+                    "registration_check tables={} cache={} elapsed_us={}",
+                    number + 1,
+                    cache_state,
+                    started.elapsed().as_micros()
+                );
+            }
+            assert_eq!(before, backend.list_objects(RUSTFS_META_BUCKET, &catalog_prefix).await.unwrap());
+            let mut recent = create_micros[create_micros.len() - 5..].to_vec();
+            recent.sort_unstable();
+            eprintln!("create_table_response tables={} last_five_median_us={}", number + 1, recent[2]);
+        }
+    }
+    assert_eq!(catalog.list_all_tables(&bucket).await.unwrap().len(), 100);
+    let namespace_entry = paths.namespace_entry_path(&bucket, &namespace);
+    let index_prefix = paths.warehouse_index_entries_prefix(&bucket);
+    let objects = backend.list_objects(RUSTFS_META_BUCKET, &catalog_prefix).await.unwrap();
+    assert!(objects.iter().all(|key| {
+        key == &bucket_entry || key == &namespace_entry || key.starts_with(&index_prefix) || key.ends_with("/table-entry.json")
+    }));
+}
+
+#[tokio::test]
+#[serial_test::serial]
+async fn cleanup_empty_table_bucket_reclaims_twenty_dropped_tables_and_allows_bucket_deletion() {
+    let (_temp_dir, _disk_paths, object_store) = crate::app::gating_test_env::isolated_multi_pool_ecstore().await;
+    let bucket = format!("cleanup-{}", Uuid::new_v4().simple());
+    let backend = crate::table_catalog::EcStoreTableCatalogObjectBackend::new_with_strong_runtime(
+        object_store.clone(),
+        crate::table_catalog::StrongTableCatalogRuntime::default(),
+    );
+    let catalog = crate::table_catalog::ConfiguredTableCatalogStore::new_for_test(
+        backend.clone(),
+        crate::table_catalog::TableCatalogBackingMode::ObjectBacked,
+    );
+    let metadata_objects = create_dropped_empty_tables_for_cleanup(&catalog, &backend, object_store.as_ref(), &bucket, 20).await;
+    assert!(
+        disable_table_bucket_response(
+            &catalog,
+            &TableCommitObjectBackend::trusted(backend.clone()),
+            object_store.as_ref(),
+            &bucket,
+        )
+        .await
+        .is_err()
+    );
+    for object in &metadata_objects {
+        assert!(backend.object_exists(&bucket, object).await.unwrap());
+    }
+
+    let authorizations = Arc::new(tokio::sync::Mutex::new(Vec::new()));
+    let publication = TableCommitObjectBackend::test(backend.clone(), authorizations.clone(), None);
+    let result = cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket).await;
+    assert!(!publication.finish(result).await.unwrap().enabled);
+    for object in &metadata_objects {
+        assert!(!backend.object_exists(&bucket, object).await.unwrap());
+        let authorized = authorizations.lock().await;
+        assert!(authorized.contains(&(object.clone(), S3Action::GetObjectAction)));
+        assert!(authorized.contains(&(object.clone(), S3Action::DeleteObjectAction)));
+    }
+    assert!(
+        !object_store
+            .get_bucket_metadata(&bucket)
+            .await
+            .unwrap()
+            .table_bucket_enabled()
+    );
+    let publication = TableCommitObjectBackend::trusted(backend);
+    assert!(
+        !cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket)
+            .await
+            .unwrap()
+            .enabled
+    );
+    object_store.delete_bucket(&bucket, &Default::default()).await.unwrap();
+}
+
+#[tokio::test]
+#[serial_test::serial]
+async fn cleanup_empty_table_bucket_validates_authorization_and_metadata_before_deleting() {
+    let (_temp_dir, _disk_paths, object_store) = crate::app::gating_test_env::isolated_multi_pool_ecstore().await;
+    let bucket = format!("cleanup-check-{}", Uuid::new_v4().simple());
+    let backend = crate::table_catalog::EcStoreTableCatalogObjectBackend::new_with_strong_runtime(
+        object_store.clone(),
+        crate::table_catalog::StrongTableCatalogRuntime::default(),
+    );
+    let catalog = crate::table_catalog::ConfiguredTableCatalogStore::new_for_test(
+        backend.clone(),
+        crate::table_catalog::TableCatalogBackingMode::ObjectBacked,
+    );
+    let metadata_objects = create_dropped_empty_tables_for_cleanup(&catalog, &backend, object_store.as_ref(), &bucket, 2).await;
+    let publication = TableCommitObjectBackend::rolling_upgrade(backend.clone());
+    assert!(
+        cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket)
+            .await
+            .is_err()
+    );
+    for object in &metadata_objects {
+        assert!(backend.object_exists(&bucket, object).await.unwrap());
+    }
+
+    let authorizations = Arc::new(tokio::sync::Mutex::new(Vec::new()));
+    let denied_object = metadata_objects.last().unwrap().clone();
+    let publication = TableCommitObjectBackend::test(backend.clone(), authorizations, Some(denied_object.clone()));
+    let result = cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket).await;
+    assert_eq!(publication.finish(result).await.unwrap_err().code(), &S3ErrorCode::AccessDenied);
+    for object in &metadata_objects {
+        assert!(backend.object_exists(&bucket, object).await.unwrap());
+    }
+
+    let original = backend.read_object(&bucket, &denied_object).await.unwrap().unwrap().data;
+    backend
+        .put_object(
+            &bucket,
+            &denied_object,
+            b"{".to_vec(),
+            crate::table_catalog::TableCatalogPutPrecondition::Any,
+        )
+        .await
+        .unwrap();
+    let publication = TableCommitObjectBackend::trusted(backend.clone());
+    assert!(
+        cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket)
+            .await
+            .is_err()
+    );
+    for object in &metadata_objects {
+        assert!(backend.object_exists(&bucket, object).await.unwrap());
+    }
+    assert!(
+        object_store
+            .get_bucket_metadata(&bucket)
+            .await
+            .unwrap()
+            .table_bucket_enabled()
+    );
+
+    backend
+        .put_object(&bucket, &denied_object, original, crate::table_catalog::TableCatalogPutPrecondition::Any)
+        .await
+        .unwrap();
+    let publication = TableCommitObjectBackend::trusted(backend);
+    assert!(
+        !cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket)
+            .await
+            .unwrap()
+            .enabled
+    );
+    object_store.delete_bucket(&bucket, &Default::default()).await.unwrap();
+}
+
+#[tokio::test]
+#[serial_test::serial]
+async fn cleanup_empty_table_bucket_handler_enforces_request_object_permissions() {
+    temp_env::async_with_vars(
+        [
+            (
+                crate::table_catalog::ENV_TABLE_CATALOG_BACKING,
+                Some(crate::table_catalog::TABLE_CATALOG_BACKING_OBJECT),
+            ),
+            (crate::table_catalog::ENV_TABLE_CATALOG_PUBLICATION_FENCE_FLEET_CONFIRMED, Some("true")),
+        ],
+        async {
+            let (_temp_dir, _disk_paths, object_store) = crate::app::gating_test_env::isolated_multi_pool_ecstore().await;
+            let bucket = format!("cleanup-request-{}", Uuid::new_v4().simple());
+            rustfs_iam::store::object::ObjectStore::new(object_store.clone())
+                .save_iam_config(
+                    serde_json::json!({"version": 1}),
+                    format!("{}/format.json", *rustfs_iam::store::object::IAM_CONFIG_PREFIX),
+                )
+                .await
+                .unwrap();
+            let iam = rustfs_iam::build_iam_sys(object_store.clone()).await.unwrap();
+            let access_key = "cleanup-request-user";
+            let secret_key = "cleanup-request-user-secret";
+            iam.create_user(
+                access_key,
+                &AddOrUpdateUserReq {
+                    secret_key: secret_key.to_string(),
+                    policy: None,
+                    status: AccountStatus::Enabled,
+                },
+            )
+            .await
+            .unwrap();
+            let context = Arc::new(AppContext::new(
+                object_store.clone(),
+                Arc::new(RequestIam { handle: iam.clone() }),
+                Arc::new(RequestKms),
+            ));
+            assert!(context.publish_action_credentials(rustfs_credentials::Credentials {
+                access_key: "cleanup-request-root".to_string(),
+                secret_key: "cleanup-request-root-secret".to_string(),
+                status: "on".to_string(),
+                ..Default::default()
+            }));
+            let backend = crate::table_catalog::EcStoreTableCatalogObjectBackend::new_with_strong_runtime(
+                object_store.clone(),
+                context.table_catalog_strong_runtime(),
+            );
+            let catalog = crate::table_catalog::ConfiguredTableCatalogStore::new_for_test(
+                backend.clone(),
+                crate::table_catalog::TableCatalogBackingMode::ObjectBacked,
+            );
+            let metadata_objects =
+                create_dropped_empty_tables_for_cleanup(&catalog, &backend, object_store.as_ref(), &bucket, 2).await;
+            let request = |prefix: &str| {
+                let mut request = table_catalog_handler_request(context.clone(), access_key, secret_key);
+                request.method = Method::DELETE;
+                request.uri = format!("{prefix}/v1/buckets/{bucket}?cleanup=true").parse().unwrap();
+                request
+            };
+            let object_resource = format!("arn:aws:s3:::{bucket}/*");
+            let first_metadata_resource = format!("arn:aws:s3:::{bucket}/{}", metadata_objects[0]);
+            for (policy_name, object_statements, allowed) in [
+                ("cleanup-request-no-read", Vec::new(), false),
+                (
+                    "cleanup-request-partial-delete",
+                    vec![
+                        serde_json::json!({"Effect":"Allow","Action":["s3:GetObject"],"Resource":[object_resource]}),
+                        serde_json::json!({"Effect":"Allow","Action":["s3:DeleteObject"],"Resource":[first_metadata_resource]}),
+                    ],
+                    false,
+                ),
+                (
+                    "cleanup-request-full",
+                    vec![serde_json::json!({
+                        "Effect":"Allow",
+                        "Action":["s3:GetObject","s3:DeleteObject"],
+                        "Resource":[object_resource],
+                    })],
+                    true,
+                ),
+            ] {
+                let mut statements = vec![serde_json::json!({"Effect":"Allow","Action":["admin:SetTableBucket"]})];
+                statements.extend(object_statements);
+                let policy = serde_json::to_vec(&serde_json::json!({"Version":"2012-10-17","Statement":statements})).unwrap();
+                iam.set_policy(policy_name, Policy::parse_config(&policy).unwrap())
+                    .await
+                    .unwrap();
+                iam.policy_db_set(access_key, UserType::Reg, false, policy_name)
+                    .await
+                    .unwrap();
+
+                let result = call_disable_table_bucket_handler(request(if allowed { "/_iceberg" } else { "/iceberg" })).await;
+                if allowed {
+                    let response = result.expect("request-scoped object permissions should authorize cleanup");
+                    assert_eq!(response.output.0, StatusCode::OK);
+                    let body: serde_json::Value = serde_json::from_slice(&response.output.1.bytes().unwrap()).unwrap();
+                    assert_eq!(body["enabled"], false);
+                    for object in &metadata_objects {
+                        assert!(!backend.object_exists(&bucket, object).await.unwrap());
+                    }
+                } else {
+                    assert_eq!(result.unwrap_err().code(), &S3ErrorCode::AccessDenied);
+                    for object in &metadata_objects {
+                        assert!(backend.object_exists(&bucket, object).await.unwrap());
+                    }
+                    assert!(
+                        object_store
+                            .get_bucket_metadata(&bucket)
+                            .await
+                            .unwrap()
+                            .table_bucket_enabled()
+                    );
+                    assert_eq!(
+                        catalog.get_table_bucket(&bucket).await.unwrap().unwrap().state,
+                        crate::table_catalog::TableCatalogEntryState::Active,
+                    );
+                }
+            }
+            object_store.delete_bucket(&bucket, &Default::default()).await.unwrap();
+        },
+    )
+    .await;
+}
+
+#[tokio::test]
+#[serial_test::serial]
+async fn cleanup_empty_table_bucket_retries_after_marker_write_rejection() {
+    use crate::table_catalog::TableCommitPublication;
+
+    let (_temp_dir, _disk_paths, object_store) = crate::app::gating_test_env::isolated_multi_pool_ecstore().await;
+    let bucket = format!("cleanup-marker-{}", Uuid::new_v4().simple());
+    let backend = TestTableCatalogObjectBackend::default();
+    let catalog = crate::table_catalog::ConfiguredTableCatalogStore::new_for_test(
+        backend.clone(),
+        crate::table_catalog::TableCatalogBackingMode::ObjectBacked,
+    );
+    let metadata_objects = create_dropped_empty_tables_for_cleanup(&catalog, &backend, object_store.as_ref(), &bucket, 2).await;
+    let publication = TableCommitObjectBackend::trusted(backend.clone());
+    TableCommitPublication::begin_table_bucket(&publication, &bucket)
+        .await
+        .unwrap();
+    let incarnation = object_store.bucket_incarnation_id_from_disk(&bucket).await.unwrap();
+    let mut cleanup = catalog
+        .prepare_empty_table_bucket_cleanup(table_bucket_entry_from_metadata_marker(&bucket), &publication)
+        .await
+        .unwrap();
+    catalog
+        .execute_empty_table_bucket_cleanup(&mut cleanup, incarnation)
+        .await
+        .unwrap();
+    let marker_result = object_store
+        .update_bucket_metadata_config_for_incarnation_validated(
+            &bucket,
+            crate::table_catalog::TABLE_BUCKET_MARKER_CONFIG,
+            Vec::new(),
+            incarnation,
+            || {
+                Err(crate::admin::storage_api::error::StorageError::other(
+                    "injected marker publication rejection",
+                ))
+            },
+        )
+        .await;
+    assert!(marker_result.is_err());
+    drop(cleanup);
+    TableCommitPublication::complete(&publication);
+    for object in &metadata_objects {
+        assert!(!backend.object_exists(&bucket, object).await.unwrap());
+    }
+    assert_eq!(
+        catalog.get_table_bucket(&bucket).await.unwrap().unwrap().state,
+        crate::table_catalog::TableCatalogEntryState::Deleted,
+    );
+    assert!(
+        object_store
+            .get_bucket_metadata(&bucket)
+            .await
+            .unwrap()
+            .table_bucket_enabled()
+    );
+
+    let publication = TableCommitObjectBackend::trusted(backend);
+    assert!(
+        !cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket)
+            .await
+            .unwrap()
+            .enabled
+    );
+    assert!(
+        !object_store
+            .get_bucket_metadata(&bucket)
+            .await
+            .unwrap()
+            .table_bucket_enabled()
+    );
+    object_store.delete_bucket(&bucket, &Default::default()).await.unwrap();
+}
+
+#[tokio::test]
+#[serial_test::serial]
+async fn cleanup_empty_table_bucket_reads_incarnation_after_publication_wait() {
+    let (_temp_dir, _disk_paths, object_store) = crate::app::gating_test_env::isolated_multi_pool_ecstore().await;
+    let bucket = format!("cleanup-recreated-{}", Uuid::new_v4().simple());
+    let backend = TestTableCatalogObjectBackend::default();
+    let catalog = crate::table_catalog::ConfiguredTableCatalogStore::new_for_test(
+        backend.clone(),
+        crate::table_catalog::TableCatalogBackingMode::ObjectBacked,
+    );
+    create_dropped_empty_tables_for_cleanup(&catalog, &backend, object_store.as_ref(), &bucket, 0).await;
+    let original_incarnation = object_store.bucket_incarnation_id_from_disk(&bucket).await.unwrap();
+    let publication_lock = crate::table_catalog::default_table_bucket_publication_lock_path();
+    let guard = backend.acquire_write_lock(&bucket, &publication_lock).await.unwrap();
+    let attempts = backend.write_lock_acquisition_count(&bucket, &publication_lock).await;
+    let cleanup_bucket = bucket.clone();
+    let cleanup_store = object_store.clone();
+    let publication = TableCommitObjectBackend::trusted(backend.clone());
+    let cleanup = tokio::spawn(async move {
+        cleanup_empty_table_bucket_response(&catalog, &publication, cleanup_store.as_ref(), &cleanup_bucket).await
+    });
+    tokio::time::timeout(StdDuration::from_secs(5), async {
+        while backend.write_lock_acquisition_count(&bucket, &publication_lock).await == attempts {
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .expect("cleanup should wait for its publication lock");
+
+    // Replace the real bucket while the in-memory publication backend holds cleanup at its lock boundary.
+    object_store.delete_bucket(&bucket, &Default::default()).await.unwrap();
+    object_store
+        .make_bucket(&bucket, &MakeBucketOptions::default())
+        .await
+        .unwrap();
+    enable_table_bucket_marker(object_store.as_ref(), &bucket).await.unwrap();
+    let replacement_incarnation = object_store.bucket_incarnation_id_from_disk(&bucket).await.unwrap();
+    assert_ne!(replacement_incarnation, original_incarnation);
+    drop(guard);
+    let response = tokio::time::timeout(StdDuration::from_secs(10), cleanup)
+        .await
+        .expect("cleanup should resume after publication becomes available")
+        .unwrap()
+        .expect("cleanup must use the incarnation observed while holding its publication lock");
+    assert!(!response.enabled);
+    assert_eq!(
+        object_store.bucket_incarnation_id_from_disk(&bucket).await.unwrap(),
+        replacement_incarnation
+    );
+    object_store.delete_bucket(&bucket, &Default::default()).await.unwrap();
+}
+
+#[tokio::test]
+#[serial_test::serial]
+async fn cleanup_empty_table_bucket_retries_metadata_delete_failures_without_removing_protection() {
+    use crate::admin::storage_api::RUSTFS_META_BUCKET;
+    use crate::table_catalog::{TableCatalogEntryState, TableCatalogObjectPaths};
+
+    let (_temp_dir, _disk_paths, object_store) = crate::app::gating_test_env::isolated_multi_pool_ecstore().await;
+    for fail_after_delete in [false, true] {
+        let bucket = format!("cleanup-retry-{}", Uuid::new_v4().simple());
+        let backend = TestTableCatalogObjectBackend::content_addressed();
+        let catalog = crate::table_catalog::ConfiguredTableCatalogStore::new_for_test(
+            backend.clone(),
+            crate::table_catalog::TableCatalogBackingMode::ObjectBacked,
+        );
+        let mut metadata_objects =
+            create_dropped_empty_tables_for_cleanup(&catalog, &backend, object_store.as_ref(), &bucket, 3).await;
+        metadata_objects.sort();
+        let paths = TableCatalogObjectPaths::default();
+        let index_prefix = paths.warehouse_index_entries_prefix(&bucket);
+        let indexes = backend.list_objects(RUSTFS_META_BUCKET, &index_prefix).await.unwrap();
+        assert!(indexes.len() >= metadata_objects.len());
+        if fail_after_delete {
+            backend.fail_after_next_delete(&bucket, &metadata_objects[1]).await;
+        } else {
+            backend.fail_delete_attempt(&bucket, &metadata_objects[1], 1).await;
+        }
+
+        let publication = TableCommitObjectBackend::trusted(backend.clone());
+        let error = cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket)
+            .await
+            .unwrap_err();
+        assert!(error.to_string().contains("injected"));
+        assert!(!backend.object_exists(&bucket, &metadata_objects[0]).await.unwrap());
+        assert_eq!(backend.object_exists(&bucket, &metadata_objects[1]).await.unwrap(), !fail_after_delete);
+        assert!(backend.object_exists(&bucket, &metadata_objects[2]).await.unwrap());
+        assert_eq!(backend.list_objects(RUSTFS_META_BUCKET, &index_prefix).await.unwrap(), indexes);
+        assert_eq!(
+            catalog.get_table_bucket(&bucket).await.unwrap().unwrap().state,
+            TableCatalogEntryState::Active
+        );
+        assert!(
+            object_store
+                .get_bucket_metadata(&bucket)
+                .await
+                .unwrap()
+                .table_bucket_enabled()
+        );
+
+        let restarted = crate::table_catalog::ConfiguredTableCatalogStore::new_for_test(
+            backend.clone(),
+            crate::table_catalog::TableCatalogBackingMode::ObjectBacked,
+        );
+        let publication = TableCommitObjectBackend::trusted(backend.clone());
+        assert!(
+            !cleanup_empty_table_bucket_response(&restarted, &publication, object_store.as_ref(), &bucket)
+                .await
+                .unwrap()
+                .enabled
+        );
+        assert!(backend.list_objects(&bucket, "").await.unwrap().is_empty());
+        assert!(
+            backend
+                .list_objects(RUSTFS_META_BUCKET, &index_prefix)
+                .await
+                .unwrap()
+                .is_empty()
+        );
+        assert_eq!(
+            restarted.get_table_bucket(&bucket).await.unwrap().unwrap().state,
+            TableCatalogEntryState::Deleted
+        );
+        assert!(
+            !object_store
+                .get_bucket_metadata(&bucket)
+                .await
+                .unwrap()
+                .table_bucket_enabled()
+        );
+        object_store.delete_bucket(&bucket, &Default::default()).await.unwrap();
+    }
+}
+
+#[tokio::test]
+#[serial_test::serial]
+async fn cleanup_empty_table_bucket_retries_partial_warehouse_index_deletion() {
+    use crate::admin::storage_api::RUSTFS_META_BUCKET;
+    use crate::table_catalog::{TableCatalogEntryState, TableCatalogObjectPaths};
+
+    let (_temp_dir, _disk_paths, object_store) = crate::app::gating_test_env::isolated_multi_pool_ecstore().await;
+    let bucket = format!("cleanup-index-retry-{}", Uuid::new_v4().simple());
+    let backend = TestTableCatalogObjectBackend::content_addressed();
+    let catalog = crate::table_catalog::ConfiguredTableCatalogStore::new_for_test(
+        backend.clone(),
+        crate::table_catalog::TableCatalogBackingMode::ObjectBacked,
+    );
+    create_dropped_empty_tables_for_cleanup(&catalog, &backend, object_store.as_ref(), &bucket, 2).await;
+    let index_prefix = TableCatalogObjectPaths::default().warehouse_index_entries_prefix(&bucket);
+    let indexes = backend.list_objects(RUSTFS_META_BUCKET, &index_prefix).await.unwrap();
+    assert!(indexes.len() >= 2);
+    let last_index = indexes.last().unwrap();
+    backend.fail_delete_attempt(RUSTFS_META_BUCKET, last_index, 1).await;
+
+    let publication = TableCommitObjectBackend::trusted(backend.clone());
+    let error = cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket)
+        .await
+        .unwrap_err();
+    assert!(error.to_string().contains("injected"));
+    assert!(backend.list_objects(&bucket, "").await.unwrap().is_empty());
+    assert_eq!(
+        backend.list_objects(RUSTFS_META_BUCKET, &index_prefix).await.unwrap(),
+        vec![last_index.clone()]
+    );
+    assert_eq!(
+        catalog.get_table_bucket(&bucket).await.unwrap().unwrap().state,
+        TableCatalogEntryState::Active
+    );
+    assert!(
+        object_store
+            .get_bucket_metadata(&bucket)
+            .await
+            .unwrap()
+            .table_bucket_enabled()
+    );
+
+    let publication = TableCommitObjectBackend::trusted(backend.clone());
+    assert!(
+        !cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket)
+            .await
+            .unwrap()
+            .enabled
+    );
+    assert!(
+        backend
+            .list_objects(RUSTFS_META_BUCKET, &index_prefix)
+            .await
+            .unwrap()
+            .is_empty()
+    );
+    assert!(
+        !object_store
+            .get_bucket_metadata(&bucket)
+            .await
+            .unwrap()
+            .table_bucket_enabled()
+    );
+    object_store.delete_bucket(&bucket, &Default::default()).await.unwrap();
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+#[serial_test::serial]
+async fn cleanup_empty_table_bucket_stops_before_delete_when_publication_lease_is_lost() {
+    use crate::admin::storage_api::RUSTFS_META_BUCKET;
+    use crate::table_catalog::test_support::RecordingLeaseClient;
+    use crate::table_catalog::{TableCatalogEntryState, TableCatalogLockGuard, TableCatalogObjectPaths, TableCommitPublication};
+    use rustfs_lock::LockClient;
+
+    let (_temp_dir, _disk_paths, object_store) = crate::app::gating_test_env::isolated_multi_pool_ecstore().await;
+    let bucket = format!("cleanup-lease-{}", Uuid::new_v4().simple());
+    let backend = TestTableCatalogObjectBackend::content_addressed();
+    let catalog = crate::table_catalog::ConfiguredTableCatalogStore::new_for_test(
+        backend.clone(),
+        crate::table_catalog::TableCatalogBackingMode::ObjectBacked,
+    );
+    let metadata_objects = create_dropped_empty_tables_for_cleanup(&catalog, &backend, object_store.as_ref(), &bucket, 1).await;
+    let index_prefix = TableCatalogObjectPaths::default().warehouse_index_entries_prefix(&bucket);
+    let indexes = backend.list_objects(RUSTFS_META_BUCKET, &index_prefix).await.unwrap();
+    let publication = TableCommitObjectBackend::trusted(backend.clone());
+    TableCommitPublication::begin_table_bucket(&publication, &bucket)
+        .await
+        .unwrap();
+    let client = Arc::new(RecordingLeaseClient {
+        inner: rustfs_lock::LocalClient::with_manager(Arc::new(rustfs_lock::GlobalLockManager::Enabled(Arc::new(
+            rustfs_lock::FastObjectLockManager::new(),
+        )))),
+        acquired: std::sync::Mutex::new(None),
+    });
+    let lock = rustfs_lock::NamespaceLockWrapper::new(
+        rustfs_lock::NamespaceLock::with_clients_and_quorum(
+            "cleanup-publication-fence".to_string(),
+            vec![client.clone() as Arc<dyn LockClient>],
+            1,
+        ),
+        rustfs_lock::ObjectKey::new(bucket.as_str(), "cleanup-publication-fence"),
+        "cleanup-publication-owner".to_string(),
+    );
+    let guard = lock.get_read_lock(StdDuration::from_secs(5)).await.unwrap();
+    let signal = guard.lock_lost_signal().unwrap();
+    let lock_id = client.acquired.lock().unwrap().clone().unwrap();
+    publication
+        .publication
+        .lock()
+        .guards
+        .push(TableCatalogLockGuard::test_namespace(guard));
+    let mut cleanup = catalog
+        .prepare_empty_table_bucket_cleanup(table_bucket_entry_from_metadata_marker(&bucket), &publication)
+        .await
+        .unwrap();
+    let incarnation = object_store.bucket_incarnation_id_from_disk(&bucket).await.unwrap();
+    let pause = backend.pause_next_delete(&bucket, &metadata_objects[0]).await;
+    let operation = catalog.execute_empty_table_bucket_cleanup(&mut cleanup, incarnation);
+    let interfere = async {
+        tokio::time::timeout(StdDuration::from_secs(5), pause.wait_started())
+            .await
+            .unwrap();
+        assert!(client.force_release(&lock_id).await.unwrap());
+        tokio::time::timeout(StdDuration::from_secs(15), signal.notified())
+            .await
+            .expect("cleanup publication must observe lease loss");
+        pause.release();
+    };
+    let (result, ()) = tokio::join!(operation, interfere);
+    assert!(matches!(result, Err(crate::table_catalog::TableCatalogStoreError::Unavailable(_))));
+    drop(cleanup);
+    TableCommitPublication::complete(&publication);
+    assert!(backend.object_exists(&bucket, &metadata_objects[0]).await.unwrap());
+    assert_eq!(backend.list_objects(RUSTFS_META_BUCKET, &index_prefix).await.unwrap(), indexes);
+    assert_eq!(
+        catalog.get_table_bucket(&bucket).await.unwrap().unwrap().state,
+        TableCatalogEntryState::Active
+    );
+    assert!(
+        object_store
+            .get_bucket_metadata(&bucket)
+            .await
+            .unwrap()
+            .table_bucket_enabled()
+    );
+
+    let publication = TableCommitObjectBackend::trusted(backend.clone());
+    assert!(
+        !cleanup_empty_table_bucket_response(&catalog, &publication, object_store.as_ref(), &bucket)
+            .await
+            .unwrap()
+            .enabled
+    );
+    assert!(!backend.object_exists(&bucket, &metadata_objects[0]).await.unwrap());
+    assert!(
+        !object_store
+            .get_bucket_metadata(&bucket)
+            .await
+            .unwrap()
+            .table_bucket_enabled()
+    );
+    object_store.delete_bucket(&bucket, &Default::default()).await.unwrap();
+}
@@ -76,6 +76,7 @@ pub struct AppContext {
    object_data_cache: Arc<ObjectDataCacheAdapter>,
    object_traffic_health: Arc<ObjectTrafficHealth>,
    table_catalog_strong_runtime: crate::table_catalog::StrongTableCatalogRuntime,
+    table_catalog_registration_cache: crate::table_catalog::TableCatalogRegistrationCache,
 }

 impl AppContext {
@@ -127,6 +128,7 @@ impl AppContext {
            object_data_cache,
            object_traffic_health: Arc::new(ObjectTrafficHealth::from_env()),
            table_catalog_strong_runtime: crate::table_catalog::StrongTableCatalogRuntime::default(),
+            table_catalog_registration_cache: crate::table_catalog::TableCatalogRegistrationCache::default(),
        }
    }

@@ -150,6 +152,10 @@ impl AppContext {
        self.table_catalog_strong_runtime.clone()
    }

+    pub(crate) fn table_catalog_registration_cache(&self) -> crate::table_catalog::TableCatalogRegistrationCache {
+        self.table_catalog_registration_cache.clone()
+    }
+
    pub fn iam(&self) -> Arc<dyn IamInterface> {
        self.iam.clone()
    }
@@ -360,6 +366,7 @@ impl AppContext {
            object_data_cache: ObjectDataCacheAdapter::disabled_arc(),
            object_traffic_health: Arc::new(ObjectTrafficHealth::from_env()),
            table_catalog_strong_runtime: crate::table_catalog::StrongTableCatalogRuntime::default(),
+            table_catalog_registration_cache: crate::table_catalog::TableCatalogRegistrationCache::default(),
        }
    }

@@ -39,6 +39,7 @@ use super::storage_api::multipart_usecase::contract::object::ObjectOperations as
 use super::storage_api::multipart_usecase::contract::range::HTTPRangeSpec;
 use super::storage_api::multipart_usecase::data_usage::{
    quota_object_size, record_bucket_object_version_write_memory, record_bucket_object_write_memory,
+    record_bucket_object_write_unknown_previous_memory,
 };
 use super::storage_api::multipart_usecase::error::{
    StorageError, is_err_invalid_upload_id, is_err_object_not_found, is_err_version_not_found,
@@ -690,27 +691,6 @@ impl DefaultMultipartUsecase {
            return Err(S3Error::with_message(S3ErrorCode::InternalError, "Not init".to_string()));
        };

-        let current_opts = internal_object_info_lookup_opts(
-            get_opts(&bucket, &key, None, None, &req.headers)
-                .await
-                .map_err(ApiError::from)?,
-        );
-        let object_lock_config_state = Box::pin(load_bucket_object_lock_config_state(&bucket)).await?;
-        let previous_current_sizes = match store.get_object_info(&bucket, &key, &current_opts).await {
-            Ok(existing_obj_info) => {
-                validate_existing_object_lock_for_write(&object_lock_config_state, &existing_obj_info, &current_opts)?;
-                let physical_size = existing_obj_info.size.max(0) as u64;
-                let logical_size = quota_object_size(&existing_obj_info);
-                Some((physical_size, logical_size))
-            }
-            Err(err) => {
-                if !is_err_object_not_found(&err) && !is_err_version_not_found(&err) {
-                    return Err(ApiError::from(err).into());
-                }
-                None
-            }
-        };
-
        // A lost Complete response retries the same upload id after staging is
        // gone. `get_multipart_info` is NoSuchUpload in that case; completion
        // below replays the committed object instead of failing the retry.
@@ -740,11 +720,6 @@ impl DefaultMultipartUsecase {
            }
            validate_complete_multipart_checksum_type(&req.headers, &info.user_defined)?;
        }
-        let cache_adapter = self.object_data_cache();
-        let _ = invalidate_object_data_cache_before_mutation(&cache_adapter, &bucket, &key).await;
-
-        let upload_user_defined = multipart_info.as_ref().map(|info| info.user_defined.clone());
-        let table_data_plane_publication_guards = retained_table_data_plane_publication_guards(&req);

        let quota_metadata_sys = self.bucket_metadata_sys();
        let quota_tracking = quota_metadata_sys.is_some();
@@ -757,13 +732,42 @@ impl DefaultMultipartUsecase {
            apply_quota_admission(&mut opts, &check_result)?;
        }

-        let previous_current_size = match previous_current_sizes {
-            Some((_, Ok(logical_size))) if quota_enabled => Some(logical_size),
-            Some((_, Err(err))) if quota_enabled => return Err(ApiError::from(err).into()),
-            Some((physical_size, _)) => Some(physical_size),
-            None => None,
+        let object_lock_config_state = Box::pin(load_bucket_object_lock_config_state(&bucket)).await?;
+        let prelookup_required = quota_enabled
+            || (!super::object::shared::put_like_write_creates_new_version(&opts)
+                && super::object::shared::object_lock_checks_required_for_state(&object_lock_config_state));
+        let prelookup_previous_current_size: Option<Option<u64>> = if prelookup_required {
+            let current_opts = internal_object_info_lookup_opts(
+                get_opts(&bucket, &key, None, None, &req.headers)
+                    .await
+                    .map_err(ApiError::from)?,
+            );
+            match store.get_object_info(&bucket, &key, &current_opts).await {
+                Ok(existing_obj_info) => {
+                    validate_existing_object_lock_for_write(&object_lock_config_state, &existing_obj_info, &current_opts)?;
+                    Some(Some(if quota_enabled {
+                        quota_object_size(&existing_obj_info).map_err(ApiError::from)?
+                    } else {
+                        existing_obj_info.size.max(0) as u64
+                    }))
+                }
+                Err(err) => {
+                    if !is_err_object_not_found(&err) && !is_err_version_not_found(&err) {
+                        return Err(ApiError::from(err).into());
+                    }
+                    Some(None)
+                }
+            }
+        } else {
+            None
        };

+        let cache_adapter = self.object_data_cache();
+        let _ = invalidate_object_data_cache_before_mutation(&cache_adapter, &bucket, &key).await;
+
+        let upload_user_defined = multipart_info.as_ref().map(|info| info.user_defined.clone());
+        let table_data_plane_publication_guards = retained_table_data_plane_publication_guards(&req);
+
        // Replication admission is decided at the object-creation commit, not
        // at multipart-session creation. Configuration and matching targets
        // can change while parts are uploaded, so persist the exact decision's
@@ -831,11 +835,20 @@ impl DefaultMultipartUsecase {
                if !obj_info.multipart_completion_replayed {
                    if quota_tracking {
                        let committed_size = quota_accounting_object_size(&obj_info, quota_enabled)?;
-
-                        if versioned {
-                            record_bucket_object_version_write_memory(&bucket, previous_current_size, committed_size).await;
-                        } else {
-                            record_bucket_object_write_memory(&bucket, previous_current_size, committed_size).await;
+                        match prelookup_previous_current_size
+                            .or_else(|| super::object::previous_current_size_from_backfill(obj_info.backfilled_old_current_size))
+                        {
+                            Some(previous_current_size) => {
+                                if versioned {
+                                    record_bucket_object_version_write_memory(&bucket, previous_current_size, committed_size)
+                                        .await;
+                                } else {
+                                    record_bucket_object_write_memory(&bucket, previous_current_size, committed_size).await;
+                                }
+                            }
+                            None => {
+                                record_bucket_object_write_unknown_previous_memory(&bucket, committed_size, versioned).await;
+                            }
                        }
                    }

@@ -1159,18 +1172,22 @@ impl DefaultMultipartUsecase {
            );
        }

-        let current_opts: ObjectOptions = internal_object_info_lookup_opts(
-            get_opts(&bucket, &key, opts.version_id.clone(), None, &req.headers)
-                .await
-                .map_err(ApiError::from)?,
-        );
-        match store.get_object_info(&bucket, &key, &current_opts).await {
-            Ok(existing_obj_info) => {
-                validate_existing_object_lock_for_write(&object_lock_config_state, &existing_obj_info, &opts)?
-            }
-            Err(err) => {
-                if !is_err_object_not_found(&err) && !is_err_version_not_found(&err) {
-                    return Err(ApiError::from(err).into());
+        if !super::object::shared::put_like_write_creates_new_version(&opts)
+            && super::object::shared::object_lock_checks_required_for_state(&object_lock_config_state)
+        {
+            let current_opts: ObjectOptions = internal_object_info_lookup_opts(
+                get_opts(&bucket, &key, opts.version_id.clone(), None, &req.headers)
+                    .await
+                    .map_err(ApiError::from)?,
+            );
+            match store.get_object_info(&bucket, &key, &current_opts).await {
+                Ok(existing_obj_info) => {
+                    validate_existing_object_lock_for_write(&object_lock_config_state, &existing_obj_info, &opts)?
+                }
+                Err(err) => {
+                    if !is_err_object_not_found(&err) && !is_err_version_not_found(&err) {
+                        return Err(ApiError::from(err).into());
+                    }
                }
            }
        }
@@ -230,7 +230,10 @@ pub(crate) use self::on_demand_migration_put::*;
 #[cfg(test)]
 pub(crate) use self::put::PollCountingBody;
 use self::put::*;
-pub(crate) use self::put::{guard_put_object_body_read_timeout, put_object_body_read_timeout, reject_oversize_single_upload};
+pub(crate) use self::put::{
+    guard_put_object_body_read_timeout, previous_current_size_from_backfill, put_object_body_read_timeout,
+    reject_oversize_single_upload,
+};
 #[cfg(test)]
 pub(crate) use self::restore::RestoreStatusCommitBarrier;
 pub(crate) use self::shared::*;
@@ -2324,7 +2324,7 @@ impl DefaultObjectUsecase {
 /// `previous_current_size` value the usage-accounting helpers expect. Outer
 /// `None` = unknown (no quorum agreement, or a peer predates the field) — the
 /// caller must fall back to the degraded accounting path.
-pub(super) fn previous_current_size_from_backfill(backfill: Option<OldCurrentSize>) -> Option<Option<u64>> {
+pub(crate) fn previous_current_size_from_backfill(backfill: Option<OldCurrentSize>) -> Option<Option<u64>> {
    backfill.map(|observation| match observation {
        OldCurrentSize::Present(size) => Some(size.max(0) as u64),
        OldCurrentSize::Absent => None,
@@ -874,7 +874,7 @@ pub(crate) fn build_put_like_object_lock_metadata(
    Ok(Some(eval_metadata))
 }

-pub(super) fn put_like_write_creates_new_version(opts: &ObjectOptions) -> bool {
+pub(in crate::app) fn put_like_write_creates_new_version(opts: &ObjectOptions) -> bool {
    opts.version_id.is_none() && opts.versioned && !opts.version_suspended
 }

@@ -997,7 +997,7 @@ pub(crate) async fn object_lock_checks_required(bucket: &str) -> bool {
        .map_or(true, |metadata| metadata.object_lock_checks_required())
 }

-pub(super) fn object_lock_checks_required_for_state(state: &metadata_sys::ObjectLockConfigState) -> bool {
+pub(in crate::app) fn object_lock_checks_required_for_state(state: &metadata_sys::ObjectLockConfigState) -> bool {
    match state {
        metadata_sys::ObjectLockConfigState::Configured { .. } | metadata_sys::ObjectLockConfigState::Fabricated => true,
        metadata_sys::ObjectLockConfigState::ConfirmedAbsent => false,
@@ -490,7 +490,8 @@ pub(crate) mod table {

        pub(crate) mod list {
            pub(crate) use super::super::super::storage_contracts::{
-                ListObjectVersionsInfo, ListObjectsV2Info, ListOperations, ObjectInfoOrErr, WalkOptions,
+                AuthoritativeListOperations, ListObjectVersionsInfo, ListObjectsV2Info, ListOperations, ObjectInfoOrErr,
+                WalkOptions,
            };
        }

@@ -498,6 +499,10 @@ pub(crate) mod table {
            pub(crate) use super::super::super::storage_contracts::NamespaceLocking;
        }

+        pub(crate) mod multipart {
+            pub(crate) use super::super::super::storage_contracts::MultipartUploadInspection;
+        }
+
        pub(crate) mod object {
            pub(crate) use super::super::super::storage_contracts::{ObjectIO, ObjectOperations};
        }
@@ -0,0 +1,342 @@
+// 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::*;
+
+/// Validates one metadata object; callers must also validate the complete object inventory.
+pub(crate) fn validate_empty_table_cleanup_metadata(
+    table_bucket: &str,
+    object_key: &str,
+    metadata: &[u8],
+) -> TableCatalogStoreResult<()> {
+    let (namespace, table) = table_identity_from_metadata_object_key(object_key).ok_or_else(|| {
+        TableCatalogStoreError::Invalid("cleanup metadata must use the default table metadata directory".to_string())
+    })?;
+    let metadata = decode_table_metadata_json(object_key, metadata)?;
+    validate_supported_table_metadata(&metadata)?;
+
+    // Future metadata fields may introduce additional object references.
+    for field in metadata.as_object().expect("validated metadata is a JSON object").keys() {
+        if !matches!(
+            field.as_str(),
+            "format-version"
+                | "table-uuid"
+                | "location"
+                | "last-sequence-number"
+                | "last-updated-ms"
+                | "last-column-id"
+                | "schema"
+                | "schemas"
+                | "current-schema-id"
+                | "partition-spec"
+                | "partition-specs"
+                | "default-spec-id"
+                | "last-partition-id"
+                | "properties"
+                | "current-snapshot-id"
+                | "snapshots"
+                | "snapshot-log"
+                | "metadata-log"
+                | "sort-orders"
+                | "default-sort-order-id"
+                | "refs"
+                | "statistics"
+                | "partition-statistics"
+        ) {
+            return Err(TableCatalogStoreError::Unsupported(format!(
+                "empty table cleanup does not support metadata field {field}"
+            )));
+        }
+    }
+
+    for field in ["snapshots", "snapshot-log", "statistics", "partition-statistics"] {
+        let Some(value) = metadata.get(field) else {
+            continue;
+        };
+        let entries = value
+            .as_array()
+            .ok_or_else(|| TableCatalogStoreError::Invalid(format!("cleanup metadata field {field} must be an array")))?;
+        if !entries.is_empty() {
+            return Err(TableCatalogStoreError::Conflict(format!(
+                "empty table cleanup requires metadata field {field} to be empty"
+            )));
+        }
+    }
+    if let Some(value) = metadata.get("refs") {
+        let refs = value
+            .as_object()
+            .ok_or_else(|| TableCatalogStoreError::Invalid("cleanup metadata refs must be an object".to_string()))?;
+        if !refs.is_empty() {
+            return Err(TableCatalogStoreError::Conflict(
+                "empty table cleanup requires metadata refs to be empty".to_string(),
+            ));
+        }
+    }
+    let table_uuid = table_metadata_uuid(&metadata)?;
+    if Uuid::parse_str(table_uuid).ok().filter(|uuid| !uuid.is_nil()).is_none() {
+        return Err(TableCatalogStoreError::Invalid(
+            "cleanup metadata table-uuid must be a non-nil UUID".to_string(),
+        ));
+    }
+    let warehouse_prefix = table_warehouse_object_prefix_from_location(table_bucket, table_metadata_location(&metadata)?)?;
+    if warehouse_prefix
+        .strip_prefix("tables/")
+        .and_then(|prefix| prefix.strip_suffix('/'))
+        .and_then(|table_id| Uuid::parse_str(table_id).ok())
+        .filter(|table_id| !table_id.is_nil())
+        .is_none()
+    {
+        return Err(TableCatalogStoreError::Conflict(
+            "empty table cleanup requires the default tables/<table-id> warehouse location".to_string(),
+        ));
+    }
+    if let Some(entries) = metadata.get("metadata-log").and_then(serde_json::Value::as_array) {
+        for entry in entries {
+            let entry = entry.as_object().expect("validated metadata log entries are JSON objects");
+            if entry
+                .keys()
+                .any(|field| !matches!(field.as_str(), "timestamp-ms" | "metadata-file"))
+            {
+                return Err(TableCatalogStoreError::Unsupported(
+                    "empty table cleanup does not support unknown metadata-log fields".to_string(),
+                ));
+            }
+            let location = entry["metadata-file"]
+                .as_str()
+                .expect("validated metadata log entries contain metadata-file");
+            table_catalog_object_key_from_location(table_bucket, location)
+                .filter(|location| is_valid_table_metadata_location(&namespace, &table, location))
+                .filter(|location| location != object_key)
+                .ok_or_else(|| {
+                    TableCatalogStoreError::Invalid(
+                        "cleanup metadata-log must reference another file in the same table metadata directory".to_string(),
+                    )
+                })?;
+        }
+    }
+
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::table_catalog::test_support::table_metadata_json;
+    use serde_json::{Value, json};
+    use std::io::Write;
+
+    const BUCKET: &str = "warehouse";
+    const TABLE_UUID: &str = "f6dd46c2-4971-4d41-b188-452d5f5c2c01";
+    const WAREHOUSE_LOCATION: &str = "s3://warehouse/tables/a9d8f3b1-abde-4a0b-b6fb-e36a30abff16";
+    const OBJECT: &str =
+        ".rustfs-table/warehouses/default/namespaces/analytics/events/tables/orders/metadata/00002.metadata.json";
+    const PREVIOUS: &str =
+        ".rustfs-table/warehouses/default/namespaces/analytics/events/tables/orders/metadata/00001.metadata.json";
+
+    fn empty_metadata() -> Value {
+        table_metadata_json(TABLE_UUID, WAREHOUSE_LOCATION)
+    }
+
+    fn validate(metadata: &Value) -> TableCatalogStoreResult<()> {
+        validate_empty_table_cleanup_metadata(BUCKET, OBJECT, &serde_json::to_vec(metadata).unwrap())
+    }
+
+    #[test]
+    fn empty_table_cleanup_accepts_v2_and_same_directory_history() {
+        let mut metadata = empty_metadata();
+        metadata["metadata-log"] = json!([
+            {"timestamp-ms": 1, "metadata-file": PREVIOUS},
+            {"timestamp-ms": 1, "metadata-file": format!("s3://{BUCKET}/{PREVIOUS}")}
+        ]);
+        assert!(validate(&metadata).is_ok());
+    }
+
+    #[test]
+    fn empty_table_cleanup_accepts_v1_without_snapshots() {
+        let mut metadata = json!({
+            "format-version": 1,
+            "table-uuid": TABLE_UUID,
+            "location": WAREHOUSE_LOCATION,
+            "last-updated-ms": 1,
+            "last-column-id": 0,
+            "schema": {"type": "struct", "fields": []},
+            "partition-spec": [],
+            "current-snapshot-id": -1
+        });
+        assert!(validate(&metadata).is_ok());
+        metadata["snapshots"] = json!([{
+            "snapshot-id": 1,
+            "timestamp-ms": 1,
+            "manifests": ["s3://warehouse/tables/table-id/metadata/manifest.avro"]
+        }]);
+        validate_supported_table_metadata(&metadata).unwrap();
+        assert!(matches!(validate(&metadata), Err(TableCatalogStoreError::Conflict(_))));
+    }
+
+    #[test]
+    fn empty_table_cleanup_rejects_unsupported_versions_and_unknown_references() {
+        for version in [0, 3, 4] {
+            let mut metadata = empty_metadata();
+            metadata["format-version"] = json!(version);
+            assert!(validate(&metadata).is_err(), "version {version}");
+        }
+        for field in [
+            "manifest-list",
+            "manifests",
+            "data-files",
+            "delete-files",
+            "next-row-id",
+            "future-field",
+        ] {
+            let mut metadata = empty_metadata();
+            metadata[field] = json!([]);
+            assert!(validate(&metadata).is_err(), "field {field}");
+        }
+    }
+
+    #[test]
+    fn empty_table_cleanup_rejects_snapshot_and_history_references() {
+        let mut metadata = empty_metadata();
+        metadata["snapshots"] = json!([{
+            "snapshot-id": 1,
+            "sequence-number": 0,
+            "timestamp-ms": 1,
+            "manifest-list": "s3://warehouse/tables/table-id/metadata/list.avro",
+            "summary": {"operation": "append"}
+        }]);
+        validate_supported_table_metadata(&metadata).unwrap();
+        assert!(matches!(validate(&metadata), Err(TableCatalogStoreError::Conflict(_))));
+
+        let mut metadata = empty_metadata();
+        metadata["snapshot-log"] = json!([{"timestamp-ms": 1, "snapshot-id": 1}]);
+        assert!(validate(&metadata).is_err());
+        metadata = empty_metadata();
+        metadata["refs"] = json!({"main": {"snapshot-id": 1, "type": "branch"}});
+        assert!(validate(&metadata).is_err());
+        metadata = empty_metadata();
+        metadata["current-snapshot-id"] = json!(1);
+        assert!(validate(&metadata).is_err());
+    }
+
+    #[test]
+    fn empty_table_cleanup_accepts_only_absent_null_or_minus_one_current_snapshot() {
+        for version in [1, 2] {
+            let mut metadata = empty_metadata();
+            metadata["format-version"] = json!(version);
+            if version == 1 {
+                metadata["schema"] = metadata["schemas"][0].clone();
+                metadata["partition-spec"] = json!([]);
+            }
+            assert!(validate(&metadata).is_ok());
+            for value in [Value::Null, json!(-1)] {
+                metadata["current-snapshot-id"] = value;
+                assert!(validate(&metadata).is_ok(), "version {version}");
+            }
+            for value in [json!(1), json!(0), json!(-2), json!("1"), json!(false), json!([])] {
+                metadata["current-snapshot-id"] = value;
+                assert!(validate(&metadata).is_err(), "version {version}");
+            }
+        }
+    }
+
+    #[test]
+    fn empty_table_cleanup_rejects_statistics_references_and_malformed_empty_fields() {
+        for field in ["snapshots", "snapshot-log", "statistics", "partition-statistics"] {
+            for value in [Value::Null, json!({}), json!("empty"), json!([{}])] {
+                let mut metadata = empty_metadata();
+                metadata[field] = value;
+                assert!(validate(&metadata).is_err(), "field {field}");
+            }
+        }
+        for value in [Value::Null, json!([]), json!("empty")] {
+            let mut metadata = empty_metadata();
+            metadata["refs"] = value;
+            assert!(validate(&metadata).is_err());
+        }
+    }
+
+    #[test]
+    fn empty_table_cleanup_rejects_unknown_or_cross_table_history() {
+        for location in [
+            format!("s3://other/{PREVIOUS}"),
+            PREVIOUS.replace("/orders/", "/other/"),
+            PREVIOUS.replace("/analytics/events/", "/another/"),
+            PREVIOUS.replace("00001.metadata.json", "../00001.metadata.json"),
+            PREVIOUS.replace("00001.metadata.json", "%2e%2e.json"),
+            format!("{PREVIOUS}?versionId=previous"),
+            format!("/{PREVIOUS}"),
+            OBJECT.to_string(),
+        ] {
+            let mut metadata = empty_metadata();
+            metadata["metadata-log"] = json!([{"timestamp-ms": 1, "metadata-file": location}]);
+            assert!(validate(&metadata).is_err(), "location {location}");
+        }
+        let mut metadata = empty_metadata();
+        metadata["metadata-log"] = json!([{"timestamp-ms": 1, "metadata-file": PREVIOUS, "future-file": "data.parquet"}]);
+        assert!(validate(&metadata).is_err());
+    }
+
+    #[test]
+    fn empty_table_cleanup_requires_valid_schema_uuid_and_warehouse() {
+        for uuid in ["table-uuid", "00000000-0000-0000-0000-000000000000", ""] {
+            let mut metadata = empty_metadata();
+            metadata["table-uuid"] = json!(uuid);
+            assert!(validate(&metadata).is_err());
+        }
+        for location in [
+            "s3://other/tables/table-id",
+            "s3://warehouse/",
+            "s3://warehouse/.rustfs-table/metadata",
+            "file:///table",
+            "s3://warehouse/custom/table-id",
+            "s3://warehouse/tables/table-id",
+            "s3://warehouse/tables/00000000-0000-0000-0000-000000000000",
+            "s3://warehouse/tables/a9d8f3b1-abde-4a0b-b6fb-e36a30abff16/nested",
+        ] {
+            let mut metadata = empty_metadata();
+            metadata["location"] = json!(location);
+            assert!(validate(&metadata).is_err());
+        }
+        let mut metadata = empty_metadata();
+        metadata["schemas"] = json!([]);
+        assert!(validate(&metadata).is_err());
+    }
+
+    #[test]
+    fn empty_table_cleanup_rejects_non_default_metadata_paths() {
+        let bytes = serde_json::to_vec(&empty_metadata()).unwrap();
+        for object in [
+            OBJECT.replace("/default/", "/custom/"),
+            OBJECT.replace("/tables/", "/views/"),
+            OBJECT.replace("/metadata/", "/data/"),
+            OBJECT.replace("00002.metadata.json", "part.parquet"),
+            OBJECT.replace("00002.metadata.json", "nested/00002.metadata.json"),
+            format!("s3://{BUCKET}/{OBJECT}"),
+        ] {
+            assert!(validate_empty_table_cleanup_metadata(BUCKET, &object, &bytes).is_err(), "object {object}");
+        }
+    }
+
+    #[test]
+    fn empty_table_cleanup_decodes_gzip_and_rejects_invalid_json() {
+        let mut encoder = flate2::write::GzEncoder::new(Vec::new(), flate2::Compression::default());
+        encoder.write_all(&serde_json::to_vec(&empty_metadata()).unwrap()).unwrap();
+        let bytes = encoder.finish().unwrap();
+        let object = OBJECT.replace(".metadata.json", ".gz.metadata.json");
+        assert!(validate_empty_table_cleanup_metadata(BUCKET, &object, &bytes).is_ok());
+        assert!(validate_empty_table_cleanup_metadata(BUCKET, OBJECT, b"{}").is_err());
+        assert!(validate_empty_table_cleanup_metadata(BUCKET, OBJECT, b"[").is_err());
+        assert!(validate_empty_table_cleanup_metadata(BUCKET, &object, b"invalid gzip").is_err());
+    }
+}
@@ -28,9 +28,11 @@ use std::{

 use crate::storage_api::table::contract::http::HTTPPreconditions;
 use crate::storage_api::table::contract::list::{
-    ListObjectVersionsInfo as StorageListObjectVersionsInfo, ListObjectsV2Info as StorageListObjectsV2Info,
-    ListOperations as StorageListOperations, ObjectInfoOrErr as StorageObjectInfoOrErr, WalkOptions as StorageWalkOptions,
+    AuthoritativeListOperations as StorageAuthoritativeListOperations, ListObjectVersionsInfo as StorageListObjectVersionsInfo,
+    ListObjectsV2Info as StorageListObjectsV2Info, ListOperations as StorageListOperations,
+    ObjectInfoOrErr as StorageObjectInfoOrErr, WalkOptions as StorageWalkOptions,
 };
+use crate::storage_api::table::contract::multipart::MultipartUploadInspection as StorageMultipartUploadInspection;
 use crate::storage_api::table::contract::namespace::NamespaceLocking as StorageNamespaceLocking;
 use crate::storage_api::table::contract::object::{ObjectIO as StorageObjectIO, ObjectOperations as StorageObjectOperations};
 use crate::storage_api::table::contract::range::HTTPRangeSpec;
@@ -58,6 +60,7 @@ use crate::storage_api::table::{
    StorageObjectOptions as ObjectOptions, StorageObjectToDelete as ObjectToDelete, StoragePutObjReader as PutObjReader,
 };

+mod cleanup_metadata;
 mod error;
 mod iceberg;
 mod identifier;
@@ -246,7 +249,9 @@ pub(crate) trait TableCatalogStorage:
        WalkOptions = CatalogWalkOptions,
        WalkCancellation = tokio_util::sync::CancellationToken,
        WalkResultSender = tokio::sync::mpsc::Sender<CatalogObjectInfoOrErr>,
-    > + StorageNamespaceLocking<Error = EcstoreError, NamespaceLock = rustfs_lock::NamespaceLockWrapper>
+    > + StorageAuthoritativeListOperations
+    + StorageNamespaceLocking<Error = EcstoreError, NamespaceLock = rustfs_lock::NamespaceLockWrapper>
+    + StorageMultipartUploadInspection<Error = EcstoreError>
 {
 }

@@ -274,7 +279,9 @@ impl<T> TableCatalogStorage for T where
            WalkOptions = CatalogWalkOptions,
            WalkCancellation = tokio_util::sync::CancellationToken,
            WalkResultSender = tokio::sync::mpsc::Sender<CatalogObjectInfoOrErr>,
-        > + StorageNamespaceLocking<Error = EcstoreError, NamespaceLock = rustfs_lock::NamespaceLockWrapper>
+        > + StorageAuthoritativeListOperations
+        + StorageNamespaceLocking<Error = EcstoreError, NamespaceLock = rustfs_lock::NamespaceLockWrapper>
+        + StorageMultipartUploadInspection<Error = EcstoreError>
 {
 }

@@ -0,0 +1,267 @@
+// 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::object::{
+    next_table_catalog_update_time, table_warehouse_index_state_ready, validate_table_warehouse_index_entry_object,
+};
+use super::*;
+
+pub(super) struct TableBucketDisablement {
+    pub(super) entry: TableBucketEntry,
+    guards: Vec<TableCatalogLockGuard>,
+}
+
+/// An in-memory, fenced cleanup plan. No mutation occurs until every object has
+/// been validated and the caller has authorized all metadata deletions.
+pub(crate) struct EmptyTableBucketCleanup {
+    disablement: TableBucketDisablement,
+    metadata: Vec<(String, String)>,
+    indexes: Vec<(String, String)>,
+    fence: TableCatalogObjectMutationFence,
+}
+
+impl EmptyTableBucketCleanup {
+    pub(crate) fn metadata_objects(&self) -> impl Iterator<Item = &str> {
+        self.metadata.iter().map(|(key, _)| key.as_str())
+    }
+
+    pub(crate) fn ensure_held(&self) -> TableCatalogStoreResult<()> {
+        self.fence.ensure_held()
+    }
+
+    pub(crate) fn mutation_fence(&self) -> TableCatalogObjectMutationFence {
+        self.fence.clone()
+    }
+}
+
+fn cleanup_conflict(message: impl Into<String>) -> TableCatalogStoreError {
+    TableCatalogStoreError::Conflict(message.into())
+}
+
+impl<B: TableCatalogObjectBackend> ObjectTableCatalogStore<B> {
+    pub(super) async fn begin_table_bucket_disablement(
+        &self,
+        mut entry: TableBucketEntry,
+        publication: Option<&dyn TableCommitPublication>,
+    ) -> TableCatalogStoreResult<TableBucketDisablement> {
+        validate_table_bucket_entry(&entry)?;
+        let mut guards = Vec::new();
+        if publication.is_some_and(|publication| publication.catalog_migration_read_permit_status() == Some(true)) {
+            if self
+                .backend
+                .read_object(self.catalog_bucket(), &self.paths.backing_migration_global_fence_path())
+                .await?
+                .is_some()
+            {
+                return Err(cleanup_conflict("table bucket registry migration is in progress"));
+            }
+            self.ensure_object_backed_catalog_write_permit_after_lock(&entry.table_bucket)
+                .await?;
+        } else {
+            guards.push(self.acquire_table_bucket_registry_write_permit().await?);
+            guards.push(self.acquire_object_backed_catalog_write_permit(&entry.table_bucket).await?);
+        }
+        let object = self.paths.table_bucket_entry_path(&entry.table_bucket);
+        guards.push(self.backend.acquire_write_lock(self.catalog_bucket(), &object).await?);
+        if let Some((current, _)) = self.read_table_bucket_with_etag_unlocked(&entry.table_bucket).await? {
+            if current.active_rename_id.is_some()
+                || !matches!(current.state, TableCatalogEntryState::Active | TableCatalogEntryState::Deleted)
+            {
+                return Err(cleanup_conflict("table bucket has an operation in progress"));
+            }
+            entry = current;
+        }
+        Ok(TableBucketDisablement { entry, guards })
+    }
+
+    pub(super) async fn finish_table_bucket_disablement(
+        &self,
+        disablement: &mut TableBucketDisablement,
+    ) -> TableCatalogStoreResult<()> {
+        let mut entry = disablement.entry.clone();
+        entry.state = TableCatalogEntryState::Deleted;
+        entry.updated_at = Some(next_table_catalog_update_time(entry.updated_at.as_deref()));
+        self.write_entry_unlocked(
+            self.catalog_bucket(),
+            &self.paths.table_bucket_entry_path(&entry.table_bucket),
+            &entry,
+            TableCatalogPutPrecondition::Any,
+            disablement
+                .guards
+                .iter()
+                .flat_map(TableCatalogLockGuard::write_commit_guards)
+                .collect(),
+        )
+        .await
+    }
+
+    async fn prepare_empty_table_bucket_cleanup(
+        &self,
+        entry: TableBucketEntry,
+        publication: &impl TableCatalogObjectBackend,
+    ) -> TableCatalogStoreResult<EmptyTableBucketCleanup> {
+        if !publication.holds_table_bucket(&entry.table_bucket) {
+            return Err(cleanup_conflict("empty table bucket cleanup requires a publication fence"));
+        }
+        let disablement = self.begin_table_bucket_disablement(entry, Some(publication)).await?;
+        let bucket = &disablement.entry.table_bucket;
+        let mut fence = publication
+            .object_mutation_fence()
+            .ok_or_else(|| cleanup_conflict("cleanup requires storage fencing"))?;
+        fence
+            .signals
+            .extend(disablement.guards.iter().filter_map(TableCatalogLockGuard::lock_lost_signal));
+        fence.ensure_held()?;
+
+        let catalog_objects = self
+            .backend
+            .list_objects_for_cleanup(self.catalog_bucket(), &self.paths.table_bucket_root_prefix(bucket))
+            .await?;
+        let bucket_path = self.paths.table_bucket_entry_path(bucket);
+        let index_state_path = self.paths.warehouse_index_state_path(bucket);
+        let index_prefix = self.paths.warehouse_index_entries_prefix(bucket);
+        let mut indexes = Vec::new();
+        for object in catalog_objects {
+            if object.delete_marker
+                || object
+                    .version_id
+                    .as_deref()
+                    .is_some_and(|version| !version.is_empty() && version != "null")
+            {
+                return Err(cleanup_conflict("catalog contains versioned or deleted records"));
+            }
+            if object.key == bucket_path {
+                continue;
+            }
+            let etag = if object.key == index_state_path {
+                let (state, etag) = self
+                    .read_entry::<TableWarehouseIndexStateEntry>(self.catalog_bucket(), &object.key)
+                    .await?
+                    .ok_or_else(|| cleanup_conflict("warehouse index state changed during cleanup"))?;
+                table_warehouse_index_state_ready(&state, bucket)?;
+                if !matches!(state.state, TableCatalogEntryState::Active | TableCatalogEntryState::Deleted) {
+                    return Err(cleanup_conflict("warehouse index has an operation in progress"));
+                }
+                etag
+            } else if object.key.starts_with(&index_prefix) {
+                let (index, etag) = self
+                    .read_entry::<TableWarehouseIndexEntry>(self.catalog_bucket(), &object.key)
+                    .await?
+                    .ok_or_else(|| cleanup_conflict("warehouse index changed during cleanup"))?;
+                validate_table_warehouse_index_entry_object(&self.paths, &object.key, &index)?;
+                if index.table_bucket != *bucket || index.state != TableCatalogEntryState::Deleted {
+                    return Err(cleanup_conflict("table bucket contains an active warehouse index"));
+                }
+                Namespace::parse(&index.namespace).map_err(|err| cleanup_conflict(err.to_string()))?;
+                IdentifierSegment::parse(index.table.clone()).map_err(|err| cleanup_conflict(err.to_string()))?;
+                if Uuid::parse_str(&index.table_id).ok().filter(|id| !id.is_nil()).is_none() {
+                    return Err(cleanup_conflict("warehouse index table identity is invalid"));
+                }
+                let expected = format!("tables/{}/", index.table_id);
+                if index.warehouse_object_prefix != expected {
+                    return Err(cleanup_conflict("cleanup requires the default table warehouse location"));
+                }
+                etag
+            } else {
+                return Err(cleanup_conflict("table bucket contains namespaces or retained catalog resources"));
+            };
+            let etag = etag.ok_or_else(|| cleanup_conflict("catalog cleanup object has no etag"))?;
+            indexes.push((object.key, etag));
+        }
+        self.backend.ensure_no_multipart_uploads(bucket).await?;
+        let mut metadata = Vec::new();
+        for object in self.backend.list_objects_for_cleanup(bucket, "").await? {
+            if object.delete_marker
+                || object
+                    .version_id
+                    .as_deref()
+                    .is_some_and(|version| !version.is_empty() && version != "null")
+            {
+                return Err(cleanup_conflict("table bucket contains object versions or delete markers"));
+            }
+            let value = publication
+                .read_object_limited(bucket, &object.key, TABLE_METADATA_JSON_MAX_SIZE)
+                .await?
+                .ok_or_else(|| cleanup_conflict("metadata object changed during cleanup"))?;
+            crate::table_catalog::cleanup_metadata::validate_empty_table_cleanup_metadata(bucket, &object.key, &value.data)?;
+            let etag = value
+                .etag
+                .ok_or_else(|| cleanup_conflict("metadata cleanup object has no etag"))?;
+            metadata.push((object.key, etag));
+        }
+        fence.ensure_held()?;
+        Ok(EmptyTableBucketCleanup {
+            disablement,
+            metadata,
+            indexes,
+            fence,
+        })
+    }
+
+    async fn execute_empty_table_bucket_cleanup(
+        &self,
+        cleanup: &mut EmptyTableBucketCleanup,
+        incarnation: Uuid,
+    ) -> TableCatalogStoreResult<()> {
+        let bucket = &cleanup.disablement.entry.table_bucket;
+        for (key, etag) in &cleanup.metadata {
+            cleanup.ensure_held()?;
+            self.backend
+                .delete_object_if_match_fenced_for_incarnation(bucket, key, etag, incarnation, &cleanup.fence)
+                .await?;
+        }
+        // A retry inventories the remaining metadata. Keep warehouse protection
+        // until all user objects and uploads have been checked again.
+        if !self.backend.list_objects_for_cleanup(bucket, "").await?.is_empty() {
+            return Err(cleanup_conflict("table bucket changed during cleanup"));
+        }
+        self.backend.ensure_no_multipart_uploads(bucket).await?;
+        for (key, etag) in &cleanup.indexes {
+            cleanup.ensure_held()?;
+            self.backend
+                .delete_object_if_match_fenced(self.catalog_bucket(), key, etag, &cleanup.fence)
+                .await?;
+        }
+        cleanup.ensure_held()?;
+        self.finish_table_bucket_disablement(&mut cleanup.disablement).await
+    }
+}
+
+impl<B: TableCatalogObjectBackend> ConfiguredTableCatalogStore<B> {
+    pub(crate) async fn prepare_empty_table_bucket_cleanup(
+        &self,
+        entry: TableBucketEntry,
+        publication: &impl TableCatalogObjectBackend,
+    ) -> TableCatalogStoreResult<EmptyTableBucketCleanup> {
+        match self {
+            Self::ObjectBacked(store) => store.prepare_empty_table_bucket_cleanup(entry, publication).await,
+            Self::DurableStrong(_) => Err(TableCatalogStoreError::Unsupported(
+                "empty table bucket cleanup requires object-backed catalog storage".to_string(),
+            )),
+        }
+    }
+
+    pub(crate) async fn execute_empty_table_bucket_cleanup(
+        &self,
+        cleanup: &mut EmptyTableBucketCleanup,
+        incarnation: Uuid,
+    ) -> TableCatalogStoreResult<()> {
+        match self {
+            Self::ObjectBacked(store) => store.execute_empty_table_bucket_cleanup(cleanup, incarnation).await,
+            Self::DurableStrong(_) => Err(TableCatalogStoreError::Unsupported(
+                "empty table bucket cleanup requires object-backed catalog storage".to_string(),
+            )),
+        }
+    }
+}
@@ -17,8 +17,13 @@ use super::*;
 use crate::storage_api::table::{WriteCommitGuard, WriteCompletion};

 mod backup;
+mod cleanup;
 mod migration;
 mod object;
+mod registration_cache;
+#[cfg(test)]
+mod registration_cache_tests;
+pub(crate) use registration_cache::TableCatalogRegistrationCache;
 mod strong;

 pub(crate) use backup::{TableCatalogBackupReport, TableCatalogRestoreReport};
@@ -775,6 +780,14 @@ pub(crate) struct TableCatalogObjectListPage {
    pub is_truncated: bool,
 }

+#[derive(Debug, Clone, PartialEq, Eq)]
+pub(crate) struct TableCatalogCleanupObject {
+    pub key: String,
+    pub version_id: Option<String>,
+    pub etag: Option<String>,
+    pub delete_marker: bool,
+}
+
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub(crate) enum TableCatalogPutPrecondition {
    Any,
@@ -803,6 +816,10 @@ pub(in crate::table_catalog) fn catalog_list_next_continuation(

 #[async_trait::async_trait]
 pub(crate) trait TableCatalogObjectBackend: Clone + Send + Sync + 'static {
+    fn registration_cache(&self) -> Option<TableCatalogRegistrationCache> {
+        None
+    }
+
    fn strong_catalog_runtime(&self) -> Option<StrongTableCatalogRuntime> {
        None
    }
@@ -918,6 +935,30 @@ pub(crate) trait TableCatalogObjectBackend: Clone + Send + Sync + 'static {
        fence.ensure_held()
    }

+    async fn put_object_unlocked_fenced(
+        &self,
+        bucket: &str,
+        object: &str,
+        data: Vec<u8>,
+        precondition: TableCatalogPutPrecondition,
+        _write_guards: Vec<WriteCommitGuard>,
+        fence: &TableCatalogObjectMutationFence,
+    ) -> TableCatalogStoreResult<()> {
+        self.put_object_fenced(bucket, object, data, precondition, fence).await
+    }
+
+    async fn delete_object_unlocked_fenced(
+        &self,
+        bucket: &str,
+        object: &str,
+        _write_guards: Vec<WriteCommitGuard>,
+        fence: &TableCatalogObjectMutationFence,
+    ) -> TableCatalogStoreResult<()> {
+        fence.ensure_held()?;
+        self.delete_object_unlocked(bucket, object).await?;
+        fence.ensure_held()
+    }
+
    async fn delete_object(&self, bucket: &str, object: &str) -> TableCatalogStoreResult<()>;

    async fn delete_object_if_match(&self, bucket: &str, object: &str, expected_etag: &str) -> TableCatalogStoreResult<()>;
@@ -934,12 +975,41 @@ pub(crate) trait TableCatalogObjectBackend: Clone + Send + Sync + 'static {
        fence.ensure_held()
    }

+    async fn delete_object_if_match_fenced_for_incarnation(
+        &self,
+        _bucket: &str,
+        _object: &str,
+        _expected_etag: &str,
+        _expected_incarnation_id: Uuid,
+        _fence: &TableCatalogObjectMutationFence,
+    ) -> TableCatalogStoreResult<()> {
+        Err(TableCatalogStoreError::Unsupported(
+            "incarnation-checked cleanup is not supported by this backend".to_string(),
+        ))
+    }
+
    async fn delete_object_unlocked(&self, bucket: &str, object: &str) -> TableCatalogStoreResult<()> {
        self.delete_object(bucket, object).await
    }

    async fn list_objects(&self, bucket: &str, prefix: &str) -> TableCatalogStoreResult<Vec<String>>;

+    async fn list_objects_for_cleanup(
+        &self,
+        _bucket: &str,
+        _prefix: &str,
+    ) -> TableCatalogStoreResult<Vec<TableCatalogCleanupObject>> {
+        Err(TableCatalogStoreError::Unsupported(
+            "authoritative cleanup listing is not supported by this backend".to_string(),
+        ))
+    }
+
+    async fn ensure_no_multipart_uploads(&self, _bucket: &str) -> TableCatalogStoreResult<()> {
+        Err(TableCatalogStoreError::Unsupported(
+            "multipart cleanup validation is not supported by this backend".to_string(),
+        ))
+    }
+
    async fn list_objects_page(
        &self,
        bucket: &str,
@@ -2204,6 +2274,7 @@ where
 pub(crate) struct EcStoreTableCatalogObjectBackend<S> {
    store: Arc<S>,
    strong_runtime: StrongTableCatalogRuntime,
+    registration_cache: TableCatalogRegistrationCache,
 }

 impl<S> Clone for EcStoreTableCatalogObjectBackend<S> {
@@ -2211,6 +2282,7 @@ impl<S> Clone for EcStoreTableCatalogObjectBackend<S> {
        Self {
            store: self.store.clone(),
            strong_runtime: self.strong_runtime.clone(),
+            registration_cache: self.registration_cache.clone(),
        }
    }
 }
@@ -2220,7 +2292,19 @@ where
    S: TableCatalogStorage,
 {
    pub fn new_with_strong_runtime(store: Arc<S>, strong_runtime: StrongTableCatalogRuntime) -> Self {
-        Self { store, strong_runtime }
+        Self::new_with_runtimes(store, strong_runtime, TableCatalogRegistrationCache::default())
+    }
+
+    pub fn new_with_runtimes(
+        store: Arc<S>,
+        strong_runtime: StrongTableCatalogRuntime,
+        registration_cache: TableCatalogRegistrationCache,
+    ) -> Self {
+        Self {
+            store,
+            strong_runtime,
+            registration_cache,
+        }
    }
 }

@@ -2231,6 +2315,10 @@ impl<S> TableCatalogObjectBackend for EcStoreTableCatalogObjectBackend<S>
 where
    S: TableCatalogStorage,
 {
+    fn registration_cache(&self) -> Option<TableCatalogRegistrationCache> {
+        Some(self.registration_cache.clone())
+    }
+
    fn strong_catalog_runtime(&self) -> Option<StrongTableCatalogRuntime> {
        Some(self.strong_runtime.clone())
    }
@@ -2390,6 +2478,46 @@ where
            .await
    }

+    async fn put_object_unlocked_fenced(
+        &self,
+        bucket: &str,
+        object: &str,
+        data: Vec<u8>,
+        precondition: TableCatalogPutPrecondition,
+        write_guards: Vec<WriteCommitGuard>,
+        fence: &TableCatalogObjectMutationFence,
+    ) -> TableCatalogStoreResult<()> {
+        fence.ensure_held()?;
+        self.put_object_with_options(bucket, object, data, precondition, Some(write_guards), Some(fence))
+            .await
+    }
+
+    async fn delete_object_unlocked_fenced(
+        &self,
+        bucket: &str,
+        object: &str,
+        write_guards: Vec<WriteCommitGuard>,
+        fence: &TableCatalogObjectMutationFence,
+    ) -> TableCatalogStoreResult<()> {
+        fence.ensure_held()?;
+        if write_guards.is_empty() {
+            return Err(TableCatalogStoreError::Internal("catalog delete has no namespace owner".to_string()));
+        }
+        let mut options = ObjectOptions {
+            no_lock: true,
+            ..Default::default()
+        };
+        fence.add_to_object_options(&mut options);
+        for guard in &write_guards {
+            options.add_write_commit_guard(guard);
+        }
+        match self.store.delete_object(bucket, object, options).await {
+            Ok(_) => fence.ensure_held(),
+            Err(err) if is_missing_storage_error(&err) => fence.ensure_held(),
+            Err(err) => Err(storage_error_to_catalog("delete fenced catalog object", err)),
+        }
+    }
+
    async fn put_object_fenced(
        &self,
        bucket: &str,
@@ -2412,7 +2540,7 @@ where
    }

    async fn delete_object_if_match(&self, bucket: &str, object: &str, expected_etag: &str) -> TableCatalogStoreResult<()> {
-        self.delete_object_if_match_with_fence(bucket, object, expected_etag, None)
+        self.delete_object_if_match_with_fence(bucket, object, expected_etag, None, None)
            .await
    }

@@ -2424,7 +2552,25 @@ where
        fence: &TableCatalogObjectMutationFence,
    ) -> TableCatalogStoreResult<()> {
        fence.ensure_held()?;
-        self.delete_object_if_match_with_fence(bucket, object, expected_etag, Some(fence))
+        self.delete_object_if_match_with_fence(bucket, object, expected_etag, Some(fence), None)
+            .await
+    }
+
+    async fn delete_object_if_match_fenced_for_incarnation(
+        &self,
+        bucket: &str,
+        object: &str,
+        expected_etag: &str,
+        expected_incarnation_id: Uuid,
+        fence: &TableCatalogObjectMutationFence,
+    ) -> TableCatalogStoreResult<()> {
+        if expected_incarnation_id.is_nil() {
+            return Err(TableCatalogStoreError::Invalid(
+                "cleanup requires a non-nil bucket incarnation".to_string(),
+            ));
+        }
+        fence.ensure_held()?;
+        self.delete_object_if_match_with_fence(bucket, object, expected_etag, Some(fence), Some(expected_incarnation_id))
            .await
    }

@@ -2498,6 +2644,57 @@ where
        })
    }

+    async fn list_objects_for_cleanup(
+        &self,
+        bucket: &str,
+        prefix: &str,
+    ) -> TableCatalogStoreResult<Vec<TableCatalogCleanupObject>> {
+        let cancel = tokio_util::sync::CancellationToken::new();
+        let _cancel_on_drop = cancel.clone().drop_guard();
+        let (sender, mut receiver) = tokio::sync::mpsc::channel::<CatalogObjectInfoOrErr>(100);
+        let walk = self.store.clone().walk_authoritative(
+            cancel,
+            bucket,
+            prefix,
+            sender,
+            CatalogWalkOptions {
+                include_free_versions: true,
+                ..Default::default()
+            },
+        );
+        let collect = async move {
+            let mut objects = Vec::new();
+            while let Some(result) = receiver.recv().await {
+                if let Some(err) = result.err {
+                    return Err(storage_error_to_catalog("inspect cleanup object versions", err));
+                }
+                let object = result.item.ok_or_else(|| {
+                    TableCatalogStoreError::Internal("cleanup object walk returned an empty result".to_string())
+                })?;
+                objects.push(TableCatalogCleanupObject {
+                    key: object.name,
+                    version_id: object.version_id.filter(|id| !id.is_nil()).map(|id| id.to_string()),
+                    etag: object.etag,
+                    delete_marker: object.delete_marker,
+                });
+            }
+            Ok(objects)
+        };
+        let (walk_result, objects) = tokio::join!(walk, collect);
+        let objects = objects?;
+        walk_result.map_err(|err| storage_error_to_catalog("list cleanup object versions", err))?;
+        Ok(objects)
+    }
+
+    async fn ensure_no_multipart_uploads(&self, bucket: &str) -> TableCatalogStoreResult<()> {
+        self.store.ensure_no_multipart_uploads(bucket).await.map_err(|err| match err {
+            StorageError::BucketNotEmpty(_) => {
+                TableCatalogStoreError::Conflict(format!("table bucket {bucket} contains unfinished multipart uploads"))
+            }
+            other => storage_error_to_catalog("inspect unfinished multipart uploads", other),
+        })
+    }
+
    async fn acquire_write_lock(&self, bucket: &str, object: &str) -> TableCatalogStoreResult<TableCatalogLockGuard> {
        let lock = self
            .store
@@ -2534,12 +2731,14 @@ where
        object: &str,
        expected_etag: &str,
        fence: Option<&TableCatalogObjectMutationFence>,
+        expected_incarnation_id: Option<Uuid>,
    ) -> TableCatalogStoreResult<()> {
        let mut options = ObjectOptions {
            http_preconditions: Some(HTTPPreconditions {
                if_match: Some(expected_etag.to_string()),
                ..Default::default()
            }),
+            expected_bucket_incarnation_id: expected_incarnation_id,
            ..Default::default()
        };
        if let Some(fence) = fence {
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Hauser Zhi	4e90946398	Merge branch 'main' into ci/validation-sweep	2026-11-01 16:22:00 +08:00
Hauser Zhi	647846956f	chore(deps): update flake.lock (#9108 ) Flake lock file updates: • Updated input 'nixpkgs': 'github:NixOS/nixpkgs/674a202' (2026-10-24) → 'github:NixOS/nixpkgs/a078812' (2026-10-31) • Updated input 'rust-overlay': 'github:oxalica/rust-overlay/5b9f119' (2026-10-24) → 'github:oxalica/rust-overlay/6ecf534' (2026-10-31)	2026-11-01 06:26:39 +00:00
GatewayJ	04690f7da5	fix(s3-tables): clean empty catalogs and cache registration reads (#9063 ) * fix(s3-tables): clean up empty tables and parallelize catalog reads * fix(s3-tables): bound catalog registration index reads * fix(s3-tables): cache validated registration ownership in memory	2026-11-01 04:49:50 +00:00
Ramakrishna Chilaka	3c0829049d	perf(ecstore): optimize multipart commit, delete receipts, and fsync (#9105 ) * perf(ecstore): fuse multipart commit, cache active upload heads, and scale fsync workers (#9077) - Cache active multipart upload heads, committed part records, and per-part transaction state in ACTIVE_MULTIPART_SOURCE_HEADS with invalidation on mutation, retirement, abort, and delete paths. - Fuse destination metadata preparation into MultipartFusedCommitOutcome, coalesce prepared metadata writes before sync_data(), and gate direct local part transaction settlement on disk.is_local(). - Retain quorum commit receipt publication before physical delete cleanup using PREPARED_DELETE_RECEIPTS, batch TARGET and PREPARED syncs, and skip redundant pre-reservation directory syncs for inline deletes. - Scale DiskBlockingLane::Fsync to 8 workers per disk with idle_workers accounting and bounded per-disk batch fairness so concurrent fsyncs coalesce in the filesystem journal. - Cache verified small stripe proof records in VERIFIED_STRIPE_RECORDS while recomputing payload digests on every read. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: d85aaf47-2a4b-4c2c-97a3-dbe8c7717e2a * fix(ecstore): align test/prod commit and cache paths and hold source lock across cache publication * fix(ecstore): preserve fused commit cancellation/panic semantics and retire excess idle fsync workers Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: d85aaf47-2a4b-4c2c-97a3-dbe8c7717e2a --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: d85aaf47-2a4b-4c2c-97a3-dbe8c7717e2a	2026-11-01 13:07:42 +08:00
hector	e5c4f8cc53	ci(validation): group the sweep report by catalog module The validation cases cover distinct areas (pool lifecycle, ILM expiration, bucket/site replication, kms, data services, identity, s3 compatibility), so a flat case table hides the category breakdown. Render one section per catalog module with its own case table and a Subtotal line, and keep the single grand - Total:/- PASS:/- FAIL: block for the dashboard overview parser. Subtotals deliberately use a different bullet prefix so the overview regexes cannot mistake a module subtotal for the grand total.	2026-11-01 12:47:38 +08:00
hector	46d0f1d76a	ci(validation): sweep the migrated canonical validation cases nightly The canonical release-validation cases migrated into auto-testing by rustfs/auto-testing#113 (66 implemented, all non-heal) currently run only through the manual validation flow; the last fleet executions stopped on 2026-10-27 and nothing schedules them. Add a scheduled validation sweep that closes the gap: - nightly at 19:30 UTC (after the functional chain settles), rotating windows of the implemented catalog cases so the whole catalog is covered roughly weekly, with a bounded batch per night - workflow_dispatch for a single case (--case), one module (--module), or an explicit batch (--limit), mirroring how the lanes are testable standalone - fleet admission through functional_environment enter/finish with the same quarantine discipline as the lanes, inter-case node cleanup, per-case logs + reports retained as a sweep artifact - a markdown sweep report uploaded to rustfs/dashboard under functional-reports/validation/ so the cases appear next to the lane reports, plus an alert-on-failure job for scheduled runs - heal.* cases stay on the interactive validation flow: many need the reviewed native/tool collectors built by separate validation-tools runs before fleet execution The registry gains the workflow responsibility (monitor: workflow, max-age 60h so the watchdog notices a dead sweep) and the wiring test baseline count moves with it. Verified locally: check_wiring, check_test_wiring, test_validation_registry (9/9), test_functional_chain + test_functional_chain_health + test_security_workflow (99/99).	2026-11-01 12:42:09 +08:00
Chris An	890a463218	test(e2e): stabilize mixed-version coordinator assertion against leader-lock refresh gaps (#9103 ) Replace the immediate post-replacement leader identity comparison with a polling stabilization loop (up to 20 x 500ms). A leader-lock refresh gap during old-binary replacement can make leader() briefly observe a different node; the loop waits for the measured new coordinator to be confirmed. A persistent change to an old-binary node still fails with the observed identity in the error message. Removes the redundant immediate assertion that the stabilization loop supersedes. Also consolidates the non-mixed leader/pid resolution to avoid a double leader query.	2026-11-01 07:03:54 +08:00
Joep	25f45f6309	test(console): strengthen brotli buffer middleware test coverage (#9098 ) Extend test_buffer_brotli_responses to verify body bytes are preserved and Content-Length matches the exact body length. Add two new cases: non-brotli responses pass through unmodified, and a body stream error strips Content-Encoding so clients do not attempt to decode an empty body as brotli (covers the fix from #9085).	2026-11-01 05:10:43 +08:00
Ramakrishna Chilaka	e3700614a1	perf(ecstore): optimize listing, versioned put, delete, and multipart (#9077 ) (#9094 ) * perf(ecstore): optimize listing, versioned put, delete, and multipart (#9077) - Buffer leaf object entries in LocalDisk::scan_dir and add zero-alloc FileMeta::buffers_match_strict, MetaCacheEntry::matches_strict_borrowed, and MetacacheReader::try_peek_buffered / skip_current for listing - Add Linux PublicationRoot::open_directory_and_readonly_child and os::has_non_data_dir_entry_lazy_from_fd with RawDir stack buffer - Fuse local versioned PUT overwrite commit, recoverable DELETE preparation/settlement, and multipart publication while preserving crash-consistency and fsync ordering invariants Fixes #9077 Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Copilot-Session: d85aaf47-2a4b-4c2c-97a3-dbe8c7717e2a * fix(app): use multipart storage domain facade Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * fix(ecstore): remove redundant multipart borrow Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com> Co-authored-by: Chris An <anzhengchao@gmail.com> Co-authored-by: Hauser Zhi <housemecn@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com> Copilot-Session: d85aaf47-2a4b-4c2c-97a3-dbe8c7717e2a	2026-10-31 16:26:52 +00:00
Chris An	d4ab7704b3	fix(scanner): rebuild unreadable usage floor instead of stopping leader iteration (#9101 ) backlog#2918 retires the usage floor as a quota proof. After dual-corruption recovery has been attempted, an unreadable floor is reported and rebuilt through a CAS overwrite with a bootstrap marker rather than stopping the leader iteration for a supervisor retry. Transient I/O errors that prevent reading the floor revision fall through to the existing stop-and-retry path. The identity-less floor regression from rustfs/rustfs#8670 now verifies automatic rebuild and recovery instead of bounded backoff.	2026-11-01 01:38:39 +08:00
Chris An	422e724051	fix(ecstore): clear stale BucketMetadataSys on ECStore drop (#9097 ) When sequential test fixtures share the bootstrap InstanceContext, the BucketMetadataSys stores a Weak<ECStore> to avoid a reference cycle. The ExpiryState holds an Arc<ECStore> that keeps the store alive past the test's own Arc drop; the next fixture's resize_workers releases it and the weak goes dangling before init_bucket_metadata_sys installs a fresh BucketMetadataSys. Any object_store() call in that window panicked. Add Drop for ECStore that clears the context's BucketMetadataSys slot when its weak handle is stale, so the next fixture always finds a clean slot. Fixes the heal absence_receipt_regressions and dangling_grace test failures under sequential execution.	2026-10-31 22:25:08 +08:00
overtrue	e0bf54197d	fix(ecstore): clear stale BucketMetadataSys on ECStore drop When sequential test fixtures share the bootstrap InstanceContext, the BucketMetadataSys stores a Weak<ECStore> to avoid a reference cycle. The ExpiryState holds an Arc<ECStore> that keeps the store alive past the test's own Arc drop; the next fixture's resize_workers releases it and the weak goes dangling before init_bucket_metadata_sys installs a fresh BucketMetadataSys. Any object_store() call in that window panicked. Add Drop for ECStore that clears the context's BucketMetadataSys slot when its weak handle is stale, so the next fixture always finds a clean slot. Fixes the heal absence_receipt_regressions and dangling_grace test failures under sequential execution.	2026-10-31 21:58:51 +08:00
Hauser Zhi	b4af43c884	fix(s3): guard conditional multipart pool errors (#9092 ) * test(s3): cover conditional writes over delete markers Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * perf(s3): avoid repeated ETag regex compilation Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * perf(s3): borrow conditional write headers Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * fix(s3): reject conditional completion on unknown pool Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * test(ci): update e2e smoke selection digest Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * test(ci): update full E2E selection for marker regression 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> Co-authored-by: Chris An <anzhengchao@gmail.com>	2026-10-31 20:48:34 +08:00
Chris An	1cb4c0ebe2	ci: stamp the plan contract and rename the workspace lint-and-build key (#9095 ) * ci: gate the plan on a contract marker A base planner predating the NIGHTLY_JOBS split still emits the old required set, and the strict verifier would reject it. Select now stamps plan_contract=2 and the workflow refuses any plan without the current marker, falling back to the full matrix, so a stale base planner can neither authorize the new narrower selection nor force a job the new workflow no longer schedules. * ci: rename the workspace lint-and-build job key and evidence path The job key test-and-lint named a job that runs no tests: it does Clippy, the test-binary archive, doctests and the log-rule anchors, while the tests live in workspace-tests. Anyone reading needs: [test-and-lint] would conclude tests had already gated that dependency. The key becomes workspace-lint-and-build to match its display name, and the shared evidence directory becomes artifacts/workspace so the path stops asserting the same thing. The aggregate's display name stays Test and Lint because that is the live ruleset check context. ci-gates.md also stops attributing the ECStore test run to the wrong job.	2026-10-31 17:41:02 +08:00
Chris An	e5563e6004	feat(release): gate publication on candidate certification (R01) (#9090 ) Add a fail-closed release certification path so incomplete or substituted evidence blocks every official publication exit. scripts/release/release_gate.py validates the versioned release-validation plan, the candidate identity and the G01 evidence-envelope set, then re-checks promoted bytes against the passed list. build/docker/package/helm publication jobs require that certification and stage artifacts before promoting identical bytes.	2026-10-31 16:53:18 +08:00
Chris An	0a72b29a34	chore(e2e): isolated harness event drill (close after evidence) (#9089 ) chore(e2e): isolated harness drill	2026-10-31 16:52:19 +08:00
Chris An	51275cb7d8	ci: move specialized lanes off the pull-request path (#9093 ) Connect Feature Boundaries, real io_uring and the rio-v2, Swift, SFTP and HTTP/3 feature lanes no longer gate a merge. They run only in the scheduled and manual full-validation runs, which report a regression within a day; on the pull-request path they bought no earlier signal worth the runner occupancy (Connect alone measured 53.5 median minutes and occasionally set the wall). Workspace lint and tests, the short core contract, the debug build, E2E smoke and the S3 lanes stay required on PRs and merge groups.	2026-10-31 16:52:03 +08:00
Chris An	3bf270b49f	ci: record CI01 measurements, drill evidence and fix the OIDC live gate (#9091 ) * ci: retain CI01 timing sample and fix the OIDC live gate tool Record the measured PR feedback and runner occupancy for the CI plan's CI01 criterion, including the honest gap between the experimental targets and the sampled medians, and replace the awscurl dependency in the live Keycloak gate with the AWS CLI so the security suite runs on PEP 668 managed fleets. * docs(testing): attach live run evidence to the event-drill index	2026-10-31 05:10:38 +00:00
Chris An	d8d064bd18	docs(testing): index event-class drills and retained CI evidence (#9088 )	2026-10-31 11:21:46 +08:00
Hauser Zhi	e5653ca055	chore(deps): refresh workspace dependencies (#9080 ) Run cargo update --verbose and cargo upgrade --verbose twice from the current main baseline. Upgrade the Tonic family to 0.14.9, Redis to 1.8.1, sysinfo to 0.40.1, and hotpath to 0.30.2 with compatible transitive updates. Synchronize the standalone fuzz lockfile and retain its PKCS#1 rc.4 resolution for compatibility with pinned RSA rc.18. Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com> Co-authored-by: Chris An <anzhengchao@gmail.com>	2026-10-31 10:36:21 +08:00
Chris An	6413a11889	test(heal): align cleanup tests with removal-disabled heals (#9087 )	2026-10-31 10:35:56 +08:00
Chris An	9e4ea9fdd1	fix(console): strip Content-Encoding on brotli buffer failure (#9085 ) When to_bytes fails in buffer_brotli_responses the error path replaced the body with empty while keeping Content-Encoding: br, producing a malformed response (empty body claiming brotli encoding). Strip the header and log a warning so the degradation is visible and the client does not attempt to decode a missing body.	2026-10-31 09:02:10 +08:00
Joep	1a841fb7c7	fix(console): buffer brotli responses to set Content-Length for Firefox (#9076 ) (#9079 ) tower_http's CompressionLayer streams brotli via Transfer-Encoding: chunked without a Content-Length header. Firefox's incremental brotli decoder aborts on this combination (NS_BINDING_ERROR), making the web console unusable in Firefox. Chrome and Safari buffer the full body before decoding so they are unaffected. Add a thin middleware (buffer_brotli_responses) that runs after CompressionLayer, collects the complete body for any brotli response that lacks Content-Length, and re-emits it with the correct length header.	2026-10-31 08:07:28 +08:00
Chris An	b8ee0301e9	feat(scanner): fair bucket service across restarts and budgeted first enumeration (#9083 ) * feat(scanner): bound first-build enumeration with the cycle budget Charge raw directory entries and their name bytes, and let one bucket scan hold a child budget that stops its own token while charging the cycle. A count budget also bounds one directory listing, so a wide first build can no longer finish a full read_dir before the first object is charged. The new stop reason maps onto the persisted Directories variant so older readers keep decoding checkpoints. A directory with queued children keeps its index until they end, and every index on the active walk path is saved as an optional map field: losing either one would recharge its replay and stop a fixed budget from advancing. * feat(scanner): keep bucket service order across restarts Save each set's first-in-first-out bucket order as an optional map field of its usage cache, including after an incomplete cycle, which otherwise writes nothing. Durable age now outranks the dirty-first and leader-local cohort orders, which only break its ties, so a restart cannot hand the budget back to the same first buckets. Failed turns add visible debt without moving a bucket forward; a budget stop or a resource veto adds none. A bucket that spends its quantum and saves forward motion returns to the queue tail within the cycle. The order authorizes nothing: another movement epoch or a corrupt value starts a new one, and a recreated bucket joins as a new arrival. * test(scanner): cover quantum handoff and saved service order Run the real worker queue with injected bucket shares and verify a spent bucket yields to waiters, a share without durable motion ends its turn, and the saved order keeps certified coverage while rejecting older owners. * docs(scanner): describe budgeted first build and service order * test(scanner): pin index retention and requeue motion Bound the nested restart fixture tightly enough that dropping a directory index fails it, and cover a repeated share whose frontier never moves. * fix(scanner): keep service generations and owner stops intact Queue a new bucket behind waiters but ahead of buckets that already had a turn, matching the leader-local cohort, and keep the set cache untouched when the owner was cancelled so an interrupted cycle cannot advance its revision. The order holds one record per inventory bucket, like the set cache's entries, so a truncating bound that could starve the long-waiting buckets is removed. Publish the incomplete scope before saving so a contended lock cannot delay it. * ci(e2e): refresh the full selection digests * fix(scanner): correct a comment spelling * fix(scanner): separate service debt from vetoes and stale owners Keep a resource veto and a remote count-budget stop out of failure debt, settle the claim before a requeue exposes the bucket to another worker, and stamp the service order with its leader so an older owner cannot replace it after an incomplete save. The checkpoint fixture's budget now covers both service orders. * fix(ecstore): create parents before resource-scoped writes #9064 moved parent creation into the lazy writer, so a write that holds a resource unit reached the native open without its parent directory and failed with FileNotFound on nested paths. * fix(e2e): retry SlowDown codes during upgrade write readiness The rolling-upgrade readiness probe only retried 503 ServiceUnavailable, so a transient 503 SlowDown from a still-recovering node failed the four_node_rolling_upgrade test immediately. Accept SlowDown and SlowDownRead alongside ServiceUnavailable, matching the existing chaos test transient-error set.	2026-10-31 07:57:18 +08:00