chore(release): prepare 1.0.0-beta.12 (#5461 )

* chore(release): prepare 1.0.0-beta.12 * chore(release): align release assets for 1.0.0-beta.12 * chore(deps): refresh release dependencies Co-Authored-By: heihutu <heihutu@gmail.com> --------- Co-authored-by: houseme <housemecn@gmail.com> Co-authored-by: heihutu <heihutu@gmail.com>
fix(iam): add correctly named policy APIs (#5457 )
2026-07-30 01:58:59 +00:00 · 2026-07-30 01:31:28 +00:00 · 2026-07-30 00:55:23 +00:00 · 2026-07-30 00:53:16 +00:00 · 2026-07-30 00:52:37 +00:00 · 2026-07-30 00:10:38 +00:00
225 changed files with 24389 additions and 3979 deletions
@@ -10,10 +10,16 @@ never weaken a check to get green.

 ## `check_layer_dependencies.sh` — layer DAG in `rustfs/src`

-Enforces `interface (admin, storage/ecfs, storage/s3_api) → app → infra`; no
-upward imports. Known legacy violations live in
+Enforces `composition (server, startup/init) → interface (admin,
+storage/ecfs, storage/s3_api) → app → infra`; no upward imports. Server source
+files are composition roots, while imports of their exported HTTP contracts
+are classified as interface dependencies. Known legacy violations live in
 `scripts/layer-dependency-baseline.txt`.

+Dedicated `*_test.rs` and `tests/` modules are outside this production guard.
+Inline `#[cfg(test)]` imports remain checked under their source file's layer;
+move architecture-crossing test scaffolding into a dedicated test module.
+
 - **New violation**: restructure your change so the dependency points
  downward (move the shared type/function to the lower layer).
 - **You legitimately removed a baseline entry**: run
@@ -1,17 +1,14 @@
 # nextest configuration for RustFS.
 #
-# Serialize two known load-sensitive / global-state-sharing ecstore test groups
-# so the full parallel nextest suite stops producing spurious failures
-# (backlog #937). These tests pass in isolation but flake under the loaded
-# parallel run for two distinct reasons:
+# Serialize the ecstore tests that share the process-wide disk registry or
+# exercise a multi-disk commit handoff across nextest process boundaries.
 #
 #   * store::bucket::tests::bucket_delete_* share process/global state (disk
 #     registry, lock client) and race make_bucket into InsufficientWriteQuorum
 #     when run concurrently with other ecstore tests.
 #   * bucket_lifecycle_ops::tests::concurrent_resend_same_part_commits_one_generation
-#     asserts a lock-acquire correctness property whose serialized cross-disk
-#     commits exceed the (already max'd, 60s) acquire deadline only when the
-#     suite saturates disk I/O.
+#     uses the shared multipart fixture and a deterministic uploadId-lock
+#     handoff, so it must not overlap another process mutating that fixture.
 #
 # serial_test's #[serial] attribute does NOT serialize these across runs:
 # nextest executes each test in its own process, where the in-process
@@ -55,12 +52,6 @@ test-group = 'ecstore-serial-flaky'
 filter = 'package(rustfs-ecstore) & test(/^set_disk::ops::multipart::tests::crash_consistency::/)'
 test-group = 'ecstore-serial-flaky'

-# This regression test builds three multipart objects on a four-disk set and
-# coordinates live GET/DELETE races under both lock modes.
-[[profile.default.overrides]]
-filter = 'package(rustfs-ecstore) & test(multipart_get_keeps_delete_blocked_at_part_boundary_in_both_lock_modes)'
-test-group = 'ecstore-serial-flaky'
-
 # Serialize the durable manual-transition checkpoint test across nextest's
 # process boundary; it mutates bucket lifecycle metadata and is not quarantined.
 [[profile.default.overrides]]
@@ -106,13 +97,6 @@ path = "junit.xml"
 # profile's own overrides list, not the default profile's).
 # ===========================================================================

-# QUARANTINE: OPEN backlog#937 — concurrent_resend lock-acquire deadline flakes
-# under saturated disk I/O in the full parallel suite.
-[[profile.ci.overrides]]
-filter = 'package(rustfs-ecstore) & test(concurrent_resend_same_part_commits_one_generation)'
-test-group = 'ecstore-serial-flaky'
-retries = 2
-
 # QUARANTINE: OPEN backlog#937 — store::bucket::tests::bucket_delete_* race
 # make_bucket into InsufficientWriteQuorum via shared global state under load.
 [[profile.ci.overrides]]
@@ -120,6 +104,11 @@ filter = 'package(rustfs-ecstore) & test(/^store::bucket::tests::bucket_delete_(
 test-group = 'ecstore-serial-flaky'
 retries = 2

+# Keep the deterministic multipart handoff isolated across nextest processes.
+[[profile.ci.overrides]]
+filter = 'package(rustfs-ecstore) & test(concurrent_resend_same_part_commits_one_generation)'
+test-group = 'ecstore-serial-flaky'
+
 # QUARANTINE: OPEN rustfs#4690 — walk_dir stall-budget accounting test depends
 # on producer/consumer timing windows that stretch past the budget on loaded
 # CI runners (regression test for rustfs#4644; failed on a zero-Rust-diff PR).
@@ -142,10 +131,6 @@ test-group = 'e2e-reliability'
 filter = 'package(rustfs-ecstore) & test(/^set_disk::ops::multipart::tests::crash_consistency::/)'
 test-group = 'ecstore-serial-flaky'

-[[profile.ci.overrides]]
-filter = 'package(rustfs-ecstore) & test(multipart_get_keeps_delete_blocked_at_part_boundary_in_both_lock_modes)'
-test-group = 'ecstore-serial-flaky'
-
 # Serialize the durable manual-transition checkpoint test under the ci profile
 # too. No retries: failures stay visible.
 [[profile.ci.overrides]]
@@ -141,7 +141,7 @@ jobs:
    name: Test and Lint
    if: github.event_name != 'pull_request' || github.event.action != 'closed'
    runs-on: sm-standard-4
-    timeout-minutes: 60
+    timeout-minutes: 90
    env:
      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
    steps:
@@ -156,16 +156,103 @@ jobs:
          github-token: ${{ secrets.GITHUB_TOKEN }}
          cache-save-if: ${{ github.ref == 'refs/heads/main' }}

+      - name: Prepare test evidence
+        run: |
+          mkdir -p artifacts/test-and-lint
+          {
+            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
+
      # Clippy runs before the test pass: lint failures are the most common
      # CI-only breakage and should surface in minutes, not after 20+ minutes
      # of tests.
      - name: Run clippy lints
        run: cargo clippy --all-targets -- -D warnings

-      - name: Run tests
+      - name: Run nextest tests
+        env:
+          # Three concurrent workspace test links saturate the self-hosted
+          # runner's overlay I/O and can wedge Cargo until the 75m timeout.
+          CARGO_BUILD_JOBS: "2"
        run: |
-          cargo nextest run --profile ci --all --exclude e2e_test
-          cargo test --all --doc
+          mkdir -p artifacts/test-and-lint
+          # Evidence sampler for issue #5394: the post-mortem pgrep below runs
+          # only after `timeout` has already TERM'd the whole cargo process
+          # group, so it cannot name a wedged process. Sample system and
+          # process state every 60s instead; the last samples before the
+          # timeout show what was stuck (rustc, linker, build script, memory
+          # pressure, ...). The log rides along in the existing artifact.
+          (
+            while true; do
+              {
+                echo "=== $(date --utc --iso-8601=seconds)"
+                echo "--- load";  cat /proc/loadavg
+                echo "--- psi";   grep -H . /proc/pressure/* 2>/dev/null || true
+                echo "--- mem";   free -m
+                echo "--- disk";  df -h / /home/runner 2>/dev/null || df -h /
+                echo "--- top-rss"
+                ps -eo pid,ppid,stat,etime,rss,pcpu,args --sort=-rss | head -15
+                echo "--- build/test processes"
+                ps -eo pid,ppid,stat,etime,rss,pcpu,args | grep -E '[c]argo|[r]ustc|[n]extest|[c]ollect2|rust-ll[d]|[b]uild-script|deps[/]' || true
+                echo "--- d-state (uninterruptible IO)"
+                ps -eo pid,stat,etime,args | awk 'NR > 1 && $2 ~ /D/' || true
+                echo
+              } >> artifacts/test-and-lint/sampler.log 2>&1 || true
+              sleep 60
+            done
+          ) &
+          sampler_pid=$!
+          trap 'kill "${sampler_pid}" 2>/dev/null || true' EXIT
+          set +e
+          NEXTEST_HIDE_PROGRESS_BAR=1 timeout --verbose --signal=TERM --kill-after=30s 75m \
+            cargo nextest run --profile ci --all --exclude e2e_test \
+            --status-level all --final-status-level all \
+            2>&1 | tee artifacts/test-and-lint/nextest.log
+          status=${PIPESTATUS[0]}
+          {
+            echo "command=cargo nextest run --profile ci --all --exclude e2e_test"
+            echo "exit_status=${status}"
+            echo "finished_at=$(date --utc --iso-8601=seconds)"
+            echo
+            echo "Remaining test-related processes:"
+            pgrep -af 'cargo|nextest|target/.*/deps/' || true
+            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
+          exit "${status}"
+
+      - name: Run documentation tests
+        run: |
+          mkdir -p artifacts/test-and-lint
+          set +e
+          timeout --verbose --signal=TERM --kill-after=30s 15m \
+            cargo test --all --doc \
+            2>&1 | tee artifacts/test-and-lint/doctest.log
+          status=${PIPESTATUS[0]}
+          {
+            echo "command=cargo test --all --doc"
+            echo "exit_status=${status}"
+            echo "finished_at=$(date --utc --iso-8601=seconds)"
+            echo
+            echo "Remaining test-related processes:"
+            pgrep -af 'cargo|rustdoc|target/.*/deps/' || true
+          } > artifacts/test-and-lint/doctest-diagnostics.txt
+          exit "${status}"
+
+      - name: Upload test reports and diagnostics
+        if: always()
+        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
+        with:
+          name: junit-test-and-lint-${{ github.run_number }}
+          path: |
+            target/nextest/ci/junit.xml
+            artifacts/test-and-lint
+          retention-days: 3
+          if-no-files-found: error

      # rustfs/backlog#1289: fail if a seed rule's log anchor no longer exists
      # verbatim in the source tree (log message drifted without updating the
@@ -174,15 +261,6 @@ jobs:
      - name: Check log-analyzer rule anchors
        run: ./scripts/check_log_analyzer_rules.sh

-      - name: Upload test junit report
-        if: always()
-        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
-        with:
-          name: junit-test-and-lint-${{ github.run_number }}
-          path: target/nextest/ci/junit.xml
-          retention-days: 3
-          if-no-files-found: ignore
-
      # Explicit gate for migration-critical suites. These tests already ran in
      # the full nextest pass above; a single filtered nextest invocation keeps
      # the named gate without rebuilding or re-running them one package at a time.
@@ -328,7 +406,7 @@ jobs:
          github-token: ${{ secrets.GITHUB_TOKEN }}

      - name: Build debug binary
-        run: cargo build -p rustfs --bins
+        run: cargo build -p rustfs --bins --features e2e-test-hooks

      - name: Upload debug binary
        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
@@ -358,7 +436,7 @@ jobs:
          github-token: ${{ secrets.GITHUB_TOKEN }}

      - name: Build debug binary with rio-v2
-        run: cargo build -p rustfs --bins --features rio-v2
+        run: cargo build -p rustfs --bins --features rio-v2,e2e-test-hooks

      - name: Upload debug binary
        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
@@ -66,7 +66,7 @@ env:
  CARGO_TERM_COLOR: always
  REGISTRY_DOCKERHUB: rustfs/rustfs
  REGISTRY_GHCR: ghcr.io/${{ github.repository }}
-  REGISTRY_QUAY: quay.io/${{ secrets.QUAY_USERNAME }}/rustfs
+  REGISTRY_QUAY: quay.io/rustfs/rustfs
  DOCKER_PLATFORMS: linux/amd64,linux/arm64

 jobs:
@@ -172,7 +172,7 @@
            ],
        },
        {
-            "name": "Debug executable target/debug/rustfs with sse",
+            "name": "Debug executable target/debug/rustfs with sse kms",
            "type":  "lldb",
            "request": "launch",
            "program": "${workspaceFolder}/target/debug/rustfs",
@@ -200,7 +200,7 @@
                // 2. kms local backend test key
                // "RUSTFS_KMS_ENABLE": "true",
                // "RUSTFS_KMS_BACKEND": "local",
-                // "RUSTFS_KMS_KEY_DIR": "./target/kms-key-dir",
+                // "RUSTFS_KMS_KEY_DIR": "/tmp/kms-key-dir",
                // "RUSTFS_KMS_LOCAL_MASTER_KEY": "my-secret-key", // Some Password
                // "RUSTFS_KMS_DEFAULT_KEY_ID": "rustfs-master-key",

@@ -228,6 +228,29 @@
                "rust"
            ],
        },
+        {
+            "name": "Debug executable target/debug/rustfs with local sse",
+            "type": "lldb",
+            "request": "launch",
+            "program": "${workspaceFolder}/target/debug/rustfs",
+            "args": [],
+            "cwd": "${workspaceFolder}",
+            "env": {
+                "RUSTFS_ACCESS_KEY": "rustfsadmin",
+                "RUSTFS_SECRET_KEY": "rustfsadmin",
+                "RUSTFS_VOLUMES": "./target/volumes/test{1...4}",
+                "RUSTFS_ADDRESS": ":9000",
+                "RUSTFS_CONSOLE_ENABLE": "true",
+                "RUSTFS_CONSOLE_ADDRESS": "127.0.0.1:9001",
+                "RUSTFS_OBS_LOG_DIRECTORY": "./target/logs",
+                "RUSTFS_UNSAFE_BYPASS_DISK_CHECK": "true",
+                "RUSTFS_SSE_S3_MASTER_KEY": "xGb3aYSp825j2tPpg8JrUzghiXsIkfdOtmrsJ/iafiM=",
+                "RUST_LOG": "rustfs=debug,ecstore=debug,s3s=debug,iam=debug",
+            },
+            "sourceLanguages": [
+                "rust"
+            ],
+        },
        {
            "type": "lldb",
            "request": "launch",
@@ -69,7 +69,7 @@ edition = "2024"
 license = "Apache-2.0"
 repository = "https://github.com/rustfs/rustfs"
 rust-version = "1.97.1"
-version = "1.0.0-beta.11"
+version = "1.0.0-beta.12"
 homepage = "https://rustfs.com"
 description = "RustFS is a high-performance distributed object storage software built using Rust, one of the most popular languages worldwide. "
 keywords = ["RustFS", "Minio", "object-storage", "filesystem", "s3"]
@@ -86,58 +86,58 @@ redundant_clone = "warn"

 [workspace.dependencies]
 # RustFS Internal Crates
-rustfs = { path = "./rustfs", version = "1.0.0-beta.11" }
-rustfs-heal = { path = "crates/heal", version = "1.0.0-beta.11" }
-rustfs-audit = { path = "crates/audit", version = "1.0.0-beta.11" }
-rustfs-checksums = { path = "crates/checksums", version = "1.0.0-beta.11" }
-rustfs-common = { path = "crates/common", version = "1.0.0-beta.11" }
-rustfs-data-usage = { path = "crates/data-usage", version = "1.0.0-beta.11" }
-rustfs-config = { path = "./crates/config", version = "1.0.0-beta.11" }
-rustfs-concurrency = { path = "./crates/concurrency", version = "1.0.0-beta.11" }
-rustfs-credentials = { path = "crates/credentials", version = "1.0.0-beta.11" }
-rustfs-crypto = { path = "crates/crypto", version = "1.0.0-beta.11" }
-rustfs-ecstore = { path = "crates/ecstore", version = "1.0.0-beta.11" }
-rustfs-filemeta = { path = "crates/filemeta", version = "1.0.0-beta.11" }
-rustfs-iam = { path = "crates/iam", version = "1.0.0-beta.11" }
-rustfs-keystone = { path = "crates/keystone", version = "1.0.0-beta.11" }
-rustfs-lifecycle = { path = "crates/lifecycle", version = "1.0.0-beta.11" }
-rustfs-kms = { path = "crates/kms", version = "1.0.0-beta.11" }
-rustfs-lock = { path = "crates/lock", version = "1.0.0-beta.11" }
-rustfs-madmin = { path = "crates/madmin", version = "1.0.0-beta.11" }
-rustfs-notify = { path = "crates/notify", version = "1.0.0-beta.11" }
-rustfs-io-metrics = { path = "crates/io-metrics", version = "1.0.0-beta.11" }
-rustfs-io-core = { path = "crates/io-core", version = "1.0.0-beta.11" }
-rustfs-object-capacity = { path = "crates/object-capacity", version = "1.0.0-beta.11" }
-rustfs-object-data-cache = { path = "crates/object-data-cache", version = "1.0.0-beta.11" }
-rustfs-log-analyzer = { path = "crates/log-analyzer", version = "1.0.0-beta.11" }
-rustfs-obs = { path = "crates/obs", version = "1.0.0-beta.11" }
-rustfs-policy = { path = "crates/policy", version = "1.0.0-beta.11" }
-rustfs-protos = { path = "crates/protos", version = "1.0.0-beta.11" }
-rustfs-protocols = { path = "crates/protocols", version = "1.0.0-beta.11" }
-rustfs-replication = { path = "crates/replication", version = "1.0.0-beta.11" }
-rustfs-rio = { path = "crates/rio", version = "1.0.0-beta.11" }
-rustfs-rio-v2 = { path = "crates/rio-v2", version = "1.0.0-beta.11" }
-rustfs-s3-types = { path = "crates/s3-types", version = "1.0.0-beta.11" }
-rustfs-s3-ops = { path = "crates/s3-ops", version = "1.0.0-beta.11" }
-rustfs-s3select-api = { path = "crates/s3select-api", version = "1.0.0-beta.11" }
-rustfs-s3select-query = { path = "crates/s3select-query", version = "1.0.0-beta.11" }
-rustfs-scanner = { path = "crates/scanner", version = "1.0.0-beta.11" }
-rustfs-security-governance = { path = "crates/security-governance", version = "1.0.0-beta.11" }
-rustfs-extension-schema = { path = "crates/extension-schema", version = "1.0.0-beta.11" }
-rustfs-signer = { path = "crates/signer", version = "1.0.0-beta.11" }
-rustfs-storage-api = { path = "crates/storage-api", version = "1.0.0-beta.11" }
-rustfs-trusted-proxies = { path = "crates/trusted-proxies", version = "1.0.0-beta.11" }
-rustfs-targets = { path = "crates/targets", version = "1.0.0-beta.11" }
-rustfs-test-utils = { path = "crates/test-utils", version = "1.0.0-beta.11" }
-rustfs-tls-runtime = { path = "crates/tls-runtime", version = "1.0.0-beta.11" }
-rustfs-utils = { path = "crates/utils", version = "1.0.0-beta.11" }
-rustfs-zip = { path = "./crates/zip", version = "1.0.0-beta.11" }
+rustfs = { path = "./rustfs", version = "1.0.0-beta.12" }
+rustfs-heal = { path = "crates/heal", version = "1.0.0-beta.12" }
+rustfs-audit = { path = "crates/audit", version = "1.0.0-beta.12" }
+rustfs-checksums = { path = "crates/checksums", version = "1.0.0-beta.12" }
+rustfs-common = { path = "crates/common", version = "1.0.0-beta.12" }
+rustfs-data-usage = { path = "crates/data-usage", version = "1.0.0-beta.12" }
+rustfs-config = { path = "./crates/config", version = "1.0.0-beta.12" }
+rustfs-concurrency = { path = "./crates/concurrency", version = "1.0.0-beta.12" }
+rustfs-credentials = { path = "crates/credentials", version = "1.0.0-beta.12" }
+rustfs-crypto = { path = "crates/crypto", version = "1.0.0-beta.12" }
+rustfs-ecstore = { path = "crates/ecstore", version = "1.0.0-beta.12" }
+rustfs-filemeta = { path = "crates/filemeta", version = "1.0.0-beta.12" }
+rustfs-iam = { path = "crates/iam", version = "1.0.0-beta.12" }
+rustfs-keystone = { path = "crates/keystone", version = "1.0.0-beta.12" }
+rustfs-lifecycle = { path = "crates/lifecycle", version = "1.0.0-beta.12" }
+rustfs-kms = { path = "crates/kms", version = "1.0.0-beta.12" }
+rustfs-lock = { path = "crates/lock", version = "1.0.0-beta.12" }
+rustfs-madmin = { path = "crates/madmin", version = "1.0.0-beta.12" }
+rustfs-notify = { path = "crates/notify", version = "1.0.0-beta.12" }
+rustfs-io-metrics = { path = "crates/io-metrics", version = "1.0.0-beta.12" }
+rustfs-io-core = { path = "crates/io-core", version = "1.0.0-beta.12" }
+rustfs-object-capacity = { path = "crates/object-capacity", version = "1.0.0-beta.12" }
+rustfs-object-data-cache = { path = "crates/object-data-cache", version = "1.0.0-beta.12" }
+rustfs-log-analyzer = { path = "crates/log-analyzer", version = "1.0.0-beta.12" }
+rustfs-obs = { path = "crates/obs", version = "1.0.0-beta.12" }
+rustfs-policy = { path = "crates/policy", version = "1.0.0-beta.12" }
+rustfs-protos = { path = "crates/protos", version = "1.0.0-beta.12" }
+rustfs-protocols = { path = "crates/protocols", version = "1.0.0-beta.12" }
+rustfs-replication = { path = "crates/replication", version = "1.0.0-beta.12" }
+rustfs-rio = { path = "crates/rio", version = "1.0.0-beta.12" }
+rustfs-rio-v2 = { path = "crates/rio-v2", version = "1.0.0-beta.12" }
+rustfs-s3-types = { path = "crates/s3-types", version = "1.0.0-beta.12" }
+rustfs-s3-ops = { path = "crates/s3-ops", version = "1.0.0-beta.12" }
+rustfs-s3select-api = { path = "crates/s3select-api", version = "1.0.0-beta.12" }
+rustfs-s3select-query = { path = "crates/s3select-query", version = "1.0.0-beta.12" }
+rustfs-scanner = { path = "crates/scanner", version = "1.0.0-beta.12" }
+rustfs-security-governance = { path = "crates/security-governance", version = "1.0.0-beta.12" }
+rustfs-extension-schema = { path = "crates/extension-schema", version = "1.0.0-beta.12" }
+rustfs-signer = { path = "crates/signer", version = "1.0.0-beta.12" }
+rustfs-storage-api = { path = "crates/storage-api", version = "1.0.0-beta.12" }
+rustfs-trusted-proxies = { path = "crates/trusted-proxies", version = "1.0.0-beta.12" }
+rustfs-targets = { path = "crates/targets", version = "1.0.0-beta.12" }
+rustfs-test-utils = { path = "crates/test-utils", version = "1.0.0-beta.12" }
+rustfs-tls-runtime = { path = "crates/tls-runtime", version = "1.0.0-beta.12" }
+rustfs-utils = { path = "crates/utils", version = "1.0.0-beta.12" }
+rustfs-zip = { path = "./crates/zip", version = "1.0.0-beta.12" }

 # Async Runtime and Networking
 async-channel = "2.5.0"
 async_zip = { default-features = false, version = "0.0.18" }
 mysql_async = { default-features = false, version = "0.37" }
-async-compression = { version = "0.4.42" }
+async-compression = { version = "0.4.43" }
 async-recursion = "1.1.1"
 async-trait = "0.1.91"
 async-nats = { version = "0.50.0", default-features = false }
@@ -152,7 +152,7 @@ lapin = { default-features = false, version = "4.10.0" }
 hyper = { version = "1.11.0" }
 hyper-rustls = { default-features = false, version = "0.27.9" }
 hyper-util = { version = "0.1.20" }
-http = "1.4.2"
+http = "1.5.0"
 http-body = "1.1.0"
 http-body-util = "0.1.4"
 minlz = "1.2.3"
@@ -200,7 +200,7 @@ jsonwebtoken = { version = "11.0.0" }
 openidconnect = { default-features = false, version = "4.0" }
 pbkdf2 = "0.13.0"
 rsa = { version = "=0.10.0-rc.18" }
-rustls = { default-features = false, version = "0.23.42" }
+rustls = { default-features = false, version = "0.23.43" }
 rustls-native-certs = "0.8"
 rustls-pki-types = "1.15.1"
 sha1 = "0.11.0"
@@ -265,7 +265,6 @@ memmap2 = "0.9.11"
 lz4 = "1.28.1"
 matchit = "0.9.2"
 md-5 = "0.11.0"
-md5 = "0.8.1"
 mime_guess = "2.0.5"
 moka = { version = "0.12.15" }
 netif = "0.1.6"
@@ -284,11 +283,11 @@ reed-solomon-erasure = { package = "rustfs-erasure-codec", version = "8.0.2" }
 reed-solomon-simd = "3.1.0"
 regex = { version = "1.13.1" }
 rumqttc = { package = "rumqttc-next", version = "0.33.3" }
-redis = { version = "1.4.1" }
+redis = { version = "1.5.0" }
 rustix = { version = "1.1.4" }
 rust-embed = { version = "8.12.0" }
 rustc-hash = { version = "2.1.3" }
-s3s = { git = "https://github.com/s3s-project/s3s.git", rev = "8136db4ac8253c0ccfd86f8216e00e0235747757" }
+s3s = { git = "https://github.com/cxymds/s3s.git", rev = "fe3941d91fa1c69956f209a9145995c9f0235bff" }
 serial_test = "4.0.1"
 shadow-rs = { default-features = false, version = "2.0.0" }
 siphasher = "1.0.3"
@@ -304,6 +303,7 @@ test-case = "3.3.1"
 thiserror = "2.0.19"
 tracing = { version = "0.1.44" }
 tracing-appender = "0.2.5"
+tracing-core = "0.1.36"
 tracing-error = "0.2.1"
 tracing-opentelemetry = { version = "0.33" }
 tracing-subscriber = { version = "0.3.23" }
@@ -314,7 +314,9 @@ uuid = { version = "1.24.0" }
 vaultrs = { version = "0.8.0" }
 tar = "0.4.46"
 walkdir = "2.5.0"
+winapi-util = "0.1.11"
 windows = { version = "0.62.2" }
+windows-sys = "0.61.2"
 xxhash-rust = { version = "0.8.18" }
 zip = "8.6.0"
 zstd = "0.13.3"
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-FROM rust:1.97-trixie
+FROM rust:1.97.1-trixie

 RUN set -eux; \
    export DEBIAN_FRONTEND=noninteractive; \
@@ -32,7 +32,7 @@ ARG RUSTFS_BUILD_FEATURES=""
 # -----------------------------
 # Build stage
 # -----------------------------
-FROM rust:1.97-trixie AS builder
+FROM rust:1.97.1-trixie AS builder

 # Re-declare args after FROM
 ARG TARGETPLATFORM
@@ -116,7 +116,7 @@ chown -R 10001:10001 data logs
 docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:latest

 # Using specific version
-docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:1.0.0-beta.11
+docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:1.0.0-beta.12
 ```

 If you use [podman](https://github.com/containers/podman) instead of docker, you can install the RustFS with the below command
@@ -113,7 +113,7 @@ chown -R 10001:10001 data logs
 docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:latest

 # 使用指定版本运行
-docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:1.0.0-beta.11
+docker run -d -p 9000:9000 -p 9001:9001 -v $(pwd)/data:/data -v $(pwd)/logs:/logs rustfs/rustfs:1.0.0-beta.12
 ```

 如果您通过绑定挂载启用 TLS 证书目录，也请用同样方式准备该目录：
@@ -25,6 +25,9 @@ keywords = ["checksum-calculation", "verification", "integrity", "authenticity",
 categories = ["web-programming", "development-tools", "network-programming"]
 documentation = "https://docs.rs/rustfs-checksums/latest/rustfs_checksum/"

+[lints]
+workspace = true
+
 [dependencies]
 bytes = { workspace = true, features = ["serde"] }
 crc-fast = { workspace = true }
@@ -1114,6 +1114,8 @@ pub struct ScannerLastMinute {
 pub struct ScannerMetricsReport {
    pub collected_at: DateTime<Utc>,
    pub current_cycle: u64,
+    #[serde(default)]
+    pub current_cycle_active: bool,
    pub current_started: DateTime<Utc>,
    pub cycles_completed_at: Vec<DateTime<Utc>>,
    pub ongoing_buckets: usize,
@@ -2051,7 +2053,7 @@ impl Metrics {
    pub fn record_scanner_transition_failed(&self, count: u64) {
        self.scanner_transition_failed.fetch_add(count, Ordering::Relaxed);
        self.record_scanner_source_failed(ScannerWorkSource::Lifecycle, count);
-        if !self.current_scan_cycle_work_active.load(Ordering::Relaxed) {
+        if !self.current_scan_cycle_work_active.load(Ordering::Acquire) {
            self.record_last_cycle_scanner_source_work(ScannerWorkSource::Lifecycle, ScannerSourceWorkUpdate::failed(count));
        }
    }
@@ -2336,6 +2338,21 @@ impl Metrics {
        *self.cycle_info.write().await = cycle;
    }

+    /// Publish a scanner cycle and its work-accounting baseline as one state transition.
+    pub async fn start_scan_cycle_work_with_cycle(&self, cycle: CurrentCycle) -> ScanCycleWorkSnapshot {
+        let mut current_cycle = self.cycle_info.write().await;
+        let snapshot = self.start_scan_cycle_work();
+        *current_cycle = Some(cycle);
+        snapshot
+    }
+
+    /// Publish the completed work snapshot and idle cycle state as one state transition.
+    pub async fn finish_scan_cycle_work_with_cycle(&self, start: ScanCycleWorkSnapshot, cycle: CurrentCycle) {
+        let mut current_cycle = self.cycle_info.write().await;
+        self.finish_scan_cycle_work(start);
+        *current_cycle = Some(cycle);
+    }
+
    /// Read the current cycle record.
    pub async fn get_cycle(&self) -> Option<CurrentCycle> {
        self.cycle_info.read().await.clone()
@@ -2464,7 +2481,7 @@ impl Metrics {
            &self.current_scan_cycle_replication_repair_work_start,
            &replication_repair_snapshot,
        );
-        self.current_scan_cycle_work_active.store(true, Ordering::Relaxed);
+        self.current_scan_cycle_work_active.store(true, Ordering::Release);
        snapshot
    }

@@ -2476,11 +2493,11 @@ impl Metrics {
        self.record_scan_cycle_work(work);
        self.record_scan_cycle_source_work(&source_work);
        self.record_scan_cycle_replication_repair_work(&replication_repair_work);
-        self.current_scan_cycle_work_active.store(false, Ordering::Relaxed);
+        self.current_scan_cycle_work_active.store(false, Ordering::Release);
    }

    pub fn current_scan_cycle_has_unresolved_heal_work(&self) -> bool {
-        if !self.current_scan_cycle_work_active.load(Ordering::Relaxed) {
+        if !self.current_scan_cycle_work_active.load(Ordering::Acquire) {
            return false;
        }

@@ -2746,13 +2763,41 @@ impl Metrics {
    pub async fn report(&self) -> ScannerMetricsReport {
        let mut m = ScannerMetricsReport::default();

-        let has_cycle = if let Some(cycle) = self.get_cycle().await {
-            m.current_cycle = cycle.current;
-            m.cycles_completed_at = cycle.cycle_completed;
-            m.current_started = cycle.started;
-            true
-        } else {
-            false
+        let has_cycle = {
+            let cycle = self.cycle_info.read().await;
+            let has_cycle = if let Some(cycle) = cycle.as_ref() {
+                m.current_cycle = cycle.current;
+                m.cycles_completed_at = cycle.cycle_completed.clone();
+                m.current_started = cycle.started;
+                true
+            } else {
+                false
+            };
+            m.current_cycle_active = self.current_scan_cycle_work_active.load(Ordering::Acquire);
+            if m.current_cycle_active {
+                let current_work = self.scan_cycle_work_since(self.current_scan_cycle_work_start());
+                let current_source_work = self.scanner_source_work_since(&self.current_scan_cycle_source_work_start_values());
+                let current_replication_repair_work =
+                    self.scanner_replication_repair_work_since(&self.current_scan_cycle_replication_repair_work_start_values());
+                m.current_cycle_objects_scanned = current_work.objects_scanned;
+                m.current_cycle_directories_scanned = current_work.directories_scanned;
+                m.current_cycle_bucket_drive_scans = current_work.bucket_drive_scans;
+                m.current_cycle_bucket_drive_failures = current_work.bucket_drive_failures;
+                m.current_cycle_yield_events = current_work.yield_events;
+                m.current_cycle_yield_duration_seconds = current_work.yield_duration_millis as f64 / 1000.0;
+                m.current_cycle_throttle_sleep_events = current_work.throttle_sleep_events;
+                m.current_cycle_throttle_sleep_duration_seconds = current_work.throttle_sleep_duration_millis as f64 / 1000.0;
+                m.current_cycle_ilm_actions = current_work.ilm_actions;
+                m.current_cycle_lifecycle_expiry_actions = current_work.lifecycle_expiry_actions;
+                m.current_cycle_lifecycle_transition_actions = current_work.lifecycle_transition_actions;
+                m.current_cycle_heal_objects = current_work.heal_objects;
+                m.current_cycle_replication_checks = current_work.replication_checks;
+                m.current_cycle_usage_saves = current_work.usage_saves;
+                m.current_cycle_source_work = self.scanner_source_work_snapshots(&current_source_work);
+                m.current_cycle_replication_repair =
+                    self.scanner_replication_repair_work_snapshots(&current_replication_repair_work);
+            }
+            has_cycle
        };

        if !has_cycle && let Some(init_time) = crate::get_global_init_time().await {
@@ -2793,28 +2838,6 @@ impl Metrics {
        m.current_disk_scan_concurrency_limit = disk_scan_concurrency_limit;
        m.current_disk_bucket_scans_queued = disk_bucket_scans_queued;
        m.current_disk_bucket_scans_active = disk_bucket_scans_active;
-        if self.current_scan_cycle_work_active.load(Ordering::Relaxed) {
-            let current_work = self.scan_cycle_work_since(self.current_scan_cycle_work_start());
-            let current_source_work = self.scanner_source_work_since(&self.current_scan_cycle_source_work_start_values());
-            let current_replication_repair_work =
-                self.scanner_replication_repair_work_since(&self.current_scan_cycle_replication_repair_work_start_values());
-            m.current_cycle_objects_scanned = current_work.objects_scanned;
-            m.current_cycle_directories_scanned = current_work.directories_scanned;
-            m.current_cycle_bucket_drive_scans = current_work.bucket_drive_scans;
-            m.current_cycle_bucket_drive_failures = current_work.bucket_drive_failures;
-            m.current_cycle_yield_events = current_work.yield_events;
-            m.current_cycle_yield_duration_seconds = current_work.yield_duration_millis as f64 / 1000.0;
-            m.current_cycle_throttle_sleep_events = current_work.throttle_sleep_events;
-            m.current_cycle_throttle_sleep_duration_seconds = current_work.throttle_sleep_duration_millis as f64 / 1000.0;
-            m.current_cycle_ilm_actions = current_work.ilm_actions;
-            m.current_cycle_lifecycle_expiry_actions = current_work.lifecycle_expiry_actions;
-            m.current_cycle_lifecycle_transition_actions = current_work.lifecycle_transition_actions;
-            m.current_cycle_heal_objects = current_work.heal_objects;
-            m.current_cycle_replication_checks = current_work.replication_checks;
-            m.current_cycle_usage_saves = current_work.usage_saves;
-            m.current_cycle_source_work = self.scanner_source_work_snapshots(&current_source_work);
-            m.current_cycle_replication_repair = self.scanner_replication_repair_work_snapshots(&current_replication_repair_work);
-        }
        let last_cycle_result = self.last_scan_cycle_result.load(Ordering::Relaxed);
        m.last_cycle_result = scan_cycle_result_label(last_cycle_result).to_string();
        m.last_cycle_result_code = last_cycle_result as u64;
@@ -4142,6 +4165,8 @@ mod tests {

        let report = metrics.report().await;

+        assert!(report.current_cycle_active);
+        assert_eq!(report.current_cycle, 0);
        assert_eq!(report.current_cycle_objects_scanned, 7);
        assert_eq!(report.current_cycle_directories_scanned, 3);
        assert_eq!(report.current_cycle_bucket_drive_scans, 2);
@@ -4158,6 +4183,8 @@ mod tests {
        metrics.finish_scan_cycle_work(start);
        let report = metrics.report().await;

+        assert!(!report.current_cycle_active);
+        assert_eq!(report.current_cycle, 0);
        assert_eq!(report.current_cycle_objects_scanned, 0);
        assert_eq!(report.current_cycle_directories_scanned, 0);
        assert_eq!(report.current_cycle_bucket_drive_scans, 0);
@@ -4184,6 +4211,91 @@ mod tests {
        assert_eq!(report.last_cycle_usage_saves, 2);
    }

+    #[tokio::test]
+    async fn scan_cycle_activity_and_cycle_state_publish_together() {
+        let metrics = Metrics::new();
+        let cycle_started = Utc::now() - chrono::Duration::seconds(5);
+        let active_cycle = CurrentCycle {
+            current: 12,
+            next: 13,
+            started: cycle_started,
+            ..Default::default()
+        };
+
+        let cycle_state = metrics.cycle_info.read().await;
+        let mut start_transition = Box::pin(metrics.start_scan_cycle_work_with_cycle(active_cycle));
+        let waker = std::task::Waker::noop();
+        let mut context = std::task::Context::from_waker(waker);
+        assert!(start_transition.as_mut().poll(&mut context).is_pending());
+        assert!(!metrics.current_scan_cycle_work_active.load(Ordering::Acquire));
+        drop(cycle_state);
+
+        let start = start_transition.await;
+        let active = metrics.report().await;
+        assert!(active.current_cycle_active);
+        assert_eq!(active.current_cycle, 12);
+        assert_eq!(active.current_started, cycle_started);
+
+        let idle_cycle = CurrentCycle {
+            current: 0,
+            next: 13,
+            started: cycle_started,
+            ..Default::default()
+        };
+        let cycle_state = metrics.cycle_info.read().await;
+        let mut finish_transition = Box::pin(metrics.finish_scan_cycle_work_with_cycle(start, idle_cycle));
+        assert!(finish_transition.as_mut().poll(&mut context).is_pending());
+        assert!(metrics.current_scan_cycle_work_active.load(Ordering::Acquire));
+        drop(cycle_state);
+
+        finish_transition.await;
+        let idle = metrics.report().await;
+        assert!(!idle.current_cycle_active);
+        assert_eq!(idle.current_cycle, 0);
+    }
+
+    #[tokio::test]
+    async fn report_keeps_cycle_identity_and_work_in_one_snapshot() {
+        let metrics = Metrics::new();
+        let cycle_ten = CurrentCycle {
+            current: 10,
+            next: 11,
+            started: Utc::now() - chrono::Duration::seconds(10),
+            ..Default::default()
+        };
+        let cycle_ten_start = metrics.start_scan_cycle_work_with_cycle(cycle_ten.clone()).await;
+        metrics.operations[Metric::ScanObject as usize].store(1, Ordering::Relaxed);
+
+        let paths = metrics.current_paths.write().await;
+        let mut report = Box::pin(metrics.report());
+        let waker = std::task::Waker::noop();
+        let mut context = std::task::Context::from_waker(waker);
+        assert!(report.as_mut().poll(&mut context).is_pending());
+
+        metrics
+            .finish_scan_cycle_work_with_cycle(cycle_ten_start, CurrentCycle { current: 0, ..cycle_ten })
+            .await;
+        let cycle_eleven_start = metrics
+            .start_scan_cycle_work_with_cycle(CurrentCycle {
+                current: 11,
+                next: 12,
+                started: Utc::now(),
+                ..Default::default()
+            })
+            .await;
+        metrics.operations[Metric::ScanObject as usize].store(101, Ordering::Relaxed);
+
+        drop(paths);
+        let snapshot = report.await;
+
+        assert_eq!(snapshot.current_cycle, 10);
+        assert_eq!(snapshot.current_cycle_objects_scanned, 1);
+
+        metrics
+            .finish_scan_cycle_work_with_cycle(cycle_eleven_start, CurrentCycle::default())
+            .await;
+    }
+
    #[tokio::test]
    async fn scanner_cycle_ilm_actions_ignore_global_ilm_work() {
        let metrics = Metrics::new();
@@ -10,6 +10,9 @@ description = "Shared concurrency contract types for RustFS - workload admission
 keywords = ["rustfs", "concurrency", "admission", "backpressure", "workers"]
 categories = ["concurrency", "filesystem"]

+[lints]
+workspace = true
+
 [dependencies]
 # Internal crates
 rustfs-io-core = { workspace = true }
@@ -158,17 +158,33 @@ pub const DEFAULT_INTERNODE_RPC_BODY_DIGEST_STRICT: bool = false;
 // rolling upgrades until the fleet-wide body-digest fallback counter reads zero.
 const _: () = assert!(!DEFAULT_INTERNODE_RPC_BODY_DIGEST_STRICT);

-/// Capacity (distinct nonces) of the process-local internode RPC replay cache that enforces
-/// one-time consumption of body-bound v2 signatures.
+/// Require the replay-scoped internode RPC signature after the fleet has converged on it.
 ///
-/// The cache retains each nonce for the ~10-minute signature freshness envelope, so the steady
-/// state holds roughly `mutating RPS x 601s` entries; the default sustains ~1,700 body-bound
-/// mutating RPCs per second (about 120 MiB worst case, allocated only under sustained load).
-/// Overflow fails closed — legitimate signed traffic is the only thing that can fill the cache
-/// (replays are rejected before insertion, and an attacker cannot mint valid nonces without the
-/// shared secret) — and increments
+/// The default keeps v1/v2 peers available during a rolling upgrade. Operators may set this only
+/// after `rustfs_system_network_internode_replay_scope_fallback_total` remains zero for a full
+/// release window. The node still accepts a v2-authenticated `Ping` carrying an epoch challenge:
+/// that narrowly scoped bootstrap lets an upgraded client learn the receiving process epoch and
+/// immediately retry with the replay-scoped signature after a peer restart.
+pub const ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT: &str = "RUSTFS_INTERNODE_RPC_REPLAY_SCOPE_STRICT";
+pub const DEFAULT_INTERNODE_RPC_REPLAY_SCOPE_STRICT: bool = false;
+
+// Compile-time invariant: mixed-version clusters must remain available until operators make the
+// observed fallback counter an explicit strictness decision.
+const _: () = assert!(!DEFAULT_INTERNODE_RPC_REPLAY_SCOPE_STRICT);
+
+/// Capacity (distinct nonces) of the process-local internode RPC replay cache that enforces
+/// one-time consumption of authenticated RPC signatures.
+///
+/// The cache retains each nonce for the ~10-minute signature freshness envelope. Once peers use
+/// replay-scoped v3 authentication, every authenticated RPC consumes one entry, so the steady
+/// state holds roughly `authenticated RPC RPS x 601s` entries. The default sustains about 1,700
+/// authenticated RPCs per second (about 120 MiB worst case, allocated only under sustained load);
+/// operators must size it for the node's aggregate peak RPC rate before enabling strict replay
+/// scope. Overflow fails closed — legitimate signed traffic is the only thing that can fill the
+/// cache (replays are rejected before insertion, and an attacker cannot mint valid nonces without
+/// the shared secret) — and increments
 /// `rustfs_system_network_internode_replay_cache_overflow_total`, so a sustained non-zero overflow
-/// counter means this capacity is undersized for the node's peak mutation rate.
+/// counter means this capacity is undersized for the node's peak authenticated RPC rate.
 pub const ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY: &str = "RUSTFS_INTERNODE_RPC_REPLAY_CACHE_CAPACITY";
 pub const DEFAULT_INTERNODE_RPC_REPLAY_CACHE_CAPACITY: usize = 1_048_576;

@@ -354,6 +370,12 @@ mod tests {
        assert_eq!(ENV_INTERNODE_RPC_BODY_DIGEST_STRICT, "RUSTFS_INTERNODE_RPC_BODY_DIGEST_STRICT");
    }

+    #[test]
+    fn internode_replay_scope_strict_env_name_is_stable() {
+        // The fail-open default invariant is asserted at compile time next to the definition.
+        assert_eq!(ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT, "RUSTFS_INTERNODE_RPC_REPLAY_SCOPE_STRICT");
+    }
+
    #[test]
    fn internode_replay_cache_capacity_defaults_and_env_name() {
        assert_eq!(ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY, "RUSTFS_INTERNODE_RPC_REPLAY_CACHE_CAPACITY");
@@ -116,6 +116,27 @@ pub const ENV_OBJECT_GET_SKIP_BITROT_VERIFY: &str = "RUSTFS_OBJECT_GET_SKIP_BITR
 /// Default: bitrot verification is enabled on GetObject reads (do not skip).
 pub const DEFAULT_OBJECT_GET_SKIP_BITROT_VERIFY: bool = false;

+/// Request writing the complete remote-tier version state into object metadata.
+///
+/// This remains ineffective until
+/// [`ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED`] is also enabled.
+pub const ENV_TIER_REMOTE_VERSION_STATE_WRITE: &str = "RUSTFS_TIER_REMOTE_VERSION_STATE_WRITE";
+pub const DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE: bool = false;
+
+/// Operator-attested fleet-wide confirmation for
+/// [`ENV_TIER_REMOTE_VERSION_STATE_WRITE`].
+///
+/// This flag is an operational contract, not automatic capability discovery.
+/// Operators may enable it only after every node that can write or read
+/// transitioned object metadata supports the remote version-state schema and
+/// semantics. Keeping the confirmation separate makes a single-node request or
+/// a writer whose local opt-in is removed fail closed.
+pub const ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED: &str = "RUSTFS_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED";
+pub const DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED: bool = false;
+
+const _: () = assert!(!DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE);
+const _: () = assert!(!DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED);
+
 // =============================================================================
 // Concurrent Request Fix - Timeout and Backpressure Configuration
 // =============================================================================
@@ -617,3 +638,15 @@ pub const ENV_OBJECT_IO_RANDOM_READAHEAD_DISABLE_CONCURRENCY: &str = "RUSTFS_OBJ

 /// Default read-ahead disable concurrency threshold: 4.
 pub const DEFAULT_OBJECT_IO_RANDOM_READAHEAD_DISABLE_CONCURRENCY: usize = 4;
+
+#[cfg(test)]
+mod remote_version_state_tests {
+    #[test]
+    fn remote_version_state_gate_uses_stable_environment_names() {
+        assert_eq!(super::ENV_TIER_REMOTE_VERSION_STATE_WRITE, "RUSTFS_TIER_REMOTE_VERSION_STATE_WRITE");
+        assert_eq!(
+            super::ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED,
+            "RUSTFS_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED"
+        );
+    }
+}
@@ -29,7 +29,6 @@ workspace = true

 [dependencies]
 serde = { workspace = true, features = ["derive"] }
-path-clean = { workspace = true }
 rmp-serde = { workspace = true }
 async-trait = { workspace = true }
 rustfs-filemeta = { workspace = true }
@@ -12,12 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

-use path_clean::PathClean;
 use serde::{Deserialize, Serialize};
 use std::{
    collections::{HashMap, HashSet},
    hash::{DefaultHasher, Hash, Hasher},
-    path::Path,
    time::{Duration, SystemTime},
 };

@@ -334,7 +332,7 @@ impl<'de> Deserialize<'de> for SizeHistogram {
 impl SizeHistogram {
    pub fn add(&mut self, size: u64) {
        let intervals = [
-            (0, 1024),                                  // LESS_THAN_1024_B
+            (0, 1024 - 1),                              // LESS_THAN_1024_B
            (1024, 64 * 1024 - 1),                      // BETWEEN_1024_B_AND_64_KB
            (64 * 1024, 256 * 1024 - 1),                // BETWEEN_64_KB_AND_256_KB
            (256 * 1024, 512 * 1024 - 1),               // BETWEEN_256_KB_AND_512_KB
@@ -362,7 +360,7 @@ impl SizeHistogram {
        // the sub-ranges in [1 KiB, 512 KiB).
        const ONE_MIB: u64 = 1024 * 1024;
        let intervals = [
-            (0, 1024),                          // LESS_THAN_1024_B
+            (0, 1024 - 1),                      // LESS_THAN_1024_B
            (1024, 64 * 1024 - 1),              // BETWEEN_1024_B_AND_64_KB
            (64 * 1024, 256 * 1024 - 1),        // BETWEEN_64_KB_AND_256_KB
            (256 * 1024, 512 * 1024 - 1),       // BETWEEN_256_KB_AND_512_KB
@@ -506,8 +504,35 @@ pub struct ReplicationStats {
 }

 impl ReplicationStats {
+    pub fn is_empty(&self) -> bool {
+        let Self {
+            pending_size,
+            replicated_size,
+            failed_size,
+            failed_count,
+            pending_count,
+            missed_threshold_size,
+            after_threshold_size,
+            missed_threshold_count,
+            after_threshold_count,
+            replicated_count,
+        } = self;
+
+        *pending_size == 0
+            && *replicated_size == 0
+            && *failed_size == 0
+            && *failed_count == 0
+            && *pending_count == 0
+            && *missed_threshold_size == 0
+            && *after_threshold_size == 0
+            && *missed_threshold_count == 0
+            && *after_threshold_count == 0
+            && *replicated_count == 0
+    }
+
+    #[deprecated(note = "use is_empty instead")]
    pub fn empty(&self) -> bool {
-        self.replicated_size == 0 && self.failed_size == 0 && self.failed_count == 0
+        self.is_empty()
    }
 }

@@ -520,16 +545,19 @@ pub struct ReplicationAllStats {
 }

 impl ReplicationAllStats {
+    pub fn is_empty(&self) -> bool {
+        let Self {
+            replica_size,
+            replica_count,
+            targets,
+        } = self;
+
+        *replica_size == 0 && *replica_count == 0 && targets.values().all(ReplicationStats::is_empty)
+    }
+
+    #[deprecated(note = "use is_empty instead")]
    pub fn empty(&self) -> bool {
-        if self.replica_size != 0 && self.replica_count != 0 {
-            return false;
-        }
-        for v in self.targets.values() {
-            if !v.empty() {
-                return false;
-            }
-        }
-        true
+        self.is_empty()
    }
 }

@@ -783,7 +811,7 @@ impl DataUsageCache {
                    return Some(root);
                }
                let mut flat = self.flatten(&root);
-                if flat.replication_stats.as_ref().is_some_and(|stats| stats.empty()) {
+                if flat.replication_stats.as_ref().is_some_and(ReplicationAllStats::is_empty) {
                    flat.replication_stats = None;
                }
                Some(flat)
@@ -1080,9 +1108,39 @@ fn mark(duc: &DataUsageCache, entry: &DataUsageEntry, found: &mut HashSet<String
    }
 }

-/// Hash a path for data usage caching
+fn clean_data_usage_path(data: &str) -> String {
+    let rooted = data.starts_with('/');
+    let mut parts = Vec::new();
+
+    for part in data.split('/') {
+        match part {
+            "" | "." => {}
+            ".." => {
+                if parts.last().is_some_and(|last| *last != "..") {
+                    parts.pop();
+                } else if !rooted {
+                    parts.push(part);
+                }
+            }
+            _ => parts.push(part),
+        }
+    }
+
+    let clean = parts.join("/");
+    match (rooted, clean.is_empty()) {
+        (true, true) => "/".to_string(),
+        (true, false) => format!("/{clean}"),
+        (false, true) => ".".to_string(),
+        (false, false) => clean,
+    }
+}
+
+/// Hash a slash-separated path for data usage caching.
+///
+/// Cache identifiers are persisted and exchanged across nodes, so their
+/// normalization must not depend on the host operating system.
 pub fn hash_path(data: &str) -> DataUsageHash {
-    DataUsageHash(Path::new(&data).clean().to_string_lossy().to_string())
+    DataUsageHash(clean_data_usage_path(data))
 }

 impl DataUsageInfo {
@@ -1467,6 +1525,23 @@ mod tests {
        buckets_count: u64,
    }

+    #[test]
+    fn hash_path_uses_portable_slash_semantics() {
+        for (input, expected) in [
+            ("", "."),
+            (".", "."),
+            ("/", "/"),
+            ("//bucket///prefix/", "/bucket/prefix"),
+            ("bucket/./prefix//object", "bucket/prefix/object"),
+            ("bucket/a/../b", "bucket/b"),
+            ("../bucket/..", ".."),
+            ("/../../bucket", "/bucket"),
+            ("bucket\\prefix/object", "bucket\\prefix/object"),
+        ] {
+            assert_eq!(hash_path(input).key(), expected, "unexpected portable cache key for {input:?}");
+        }
+    }
+
    #[test]
    fn completeness_marker_is_additive_for_legacy_named_readers() {
        let current = DataUsageInfo {
@@ -1571,6 +1646,49 @@ mod tests {
        assert_eq!(map["BETWEEN_512_KB_AND_1_MB"], 1);
    }

+    #[test]
+    fn test_size_histogram_classifies_adjacent_boundaries_once() {
+        let cases = [
+            (1023, 0),
+            (1024, 1),
+            (64 * 1024 - 1, 1),
+            (64 * 1024, 2),
+            (256 * 1024 - 1, 2),
+            (256 * 1024, 3),
+            (512 * 1024 - 1, 3),
+            (512 * 1024, 4),
+            (1024 * 1024 - 1, 4),
+            (1024 * 1024, 6),
+            (10 * 1024 * 1024 - 1, 6),
+            (10 * 1024 * 1024, 7),
+            (64 * 1024 * 1024 - 1, 7),
+            (64 * 1024 * 1024, 8),
+            (128 * 1024 * 1024 - 1, 8),
+            (128 * 1024 * 1024, 9),
+            (512 * 1024 * 1024 - 1, 9),
+            (512 * 1024 * 1024, 10),
+        ];
+
+        for (size, expected_bucket) in cases {
+            let mut hist = SizeHistogram::default();
+            hist.add(size);
+
+            assert_eq!(hist.0.iter().sum::<u64>(), 1, "size {size} must have exactly one physical bucket");
+            assert_eq!(hist.0[expected_bucket], 1, "size {size} must select the expected bucket");
+        }
+    }
+
+    #[test]
+    fn test_size_histogram_1024_bytes_contributes_to_compat_rollup() {
+        let mut hist = SizeHistogram::default();
+        hist.add(1024);
+
+        let map = hist.to_map();
+        assert_eq!(map["LESS_THAN_1024_B"], 0);
+        assert_eq!(map["BETWEEN_1024_B_AND_64_KB"], 1);
+        assert_eq!(map["BETWEEN_1024B_AND_1_MB"], 1);
+    }
+
    #[test]
    fn test_size_histogram_compat_rollup_saturates_on_corrupt_counts() {
        let mut hist = SizeHistogram::default();
@@ -1582,6 +1700,126 @@ mod tests {
        assert_eq!(map["BETWEEN_1024B_AND_1_MB"], u64::MAX);
    }

+    #[test]
+    fn replication_stats_empty_checks_every_field() {
+        type SetField = fn(&mut ReplicationStats);
+
+        let cases: [(&str, SetField); 10] = [
+            ("pending_size", |stats| stats.pending_size = 1),
+            ("replicated_size", |stats| stats.replicated_size = 1),
+            ("failed_size", |stats| stats.failed_size = 1),
+            ("failed_count", |stats| stats.failed_count = 1),
+            ("pending_count", |stats| stats.pending_count = 1),
+            ("missed_threshold_size", |stats| stats.missed_threshold_size = 1),
+            ("after_threshold_size", |stats| stats.after_threshold_size = 1),
+            ("missed_threshold_count", |stats| stats.missed_threshold_count = 1),
+            ("after_threshold_count", |stats| stats.after_threshold_count = 1),
+            ("replicated_count", |stats| stats.replicated_count = 1),
+        ];
+
+        assert!(ReplicationStats::default().is_empty());
+        for (field, set_nonzero) in cases {
+            let mut stats = ReplicationStats::default();
+            set_nonzero(&mut stats);
+            assert!(!stats.is_empty(), "{field} must make replication stats non-empty");
+        }
+    }
+
+    #[test]
+    fn replication_all_stats_empty_checks_aggregate_fields_independently() {
+        let cases = [
+            (
+                "replica_size",
+                ReplicationAllStats {
+                    replica_size: 1,
+                    ..Default::default()
+                },
+            ),
+            (
+                "replica_count",
+                ReplicationAllStats {
+                    replica_count: 1,
+                    ..Default::default()
+                },
+            ),
+        ];
+
+        assert!(ReplicationAllStats::default().is_empty());
+        for (field, stats) in cases {
+            assert!(!stats.is_empty(), "{field} must make aggregate replication stats non-empty");
+        }
+
+        let empty_targets = ReplicationAllStats {
+            targets: HashMap::from([("arn:test:empty".to_string(), ReplicationStats::default())]),
+            ..Default::default()
+        };
+        assert!(empty_targets.is_empty(), "all-empty targets must keep aggregate stats empty");
+
+        let stats = ReplicationAllStats {
+            targets: HashMap::from([
+                ("arn:test:empty".to_string(), ReplicationStats::default()),
+                (
+                    "arn:test:non-empty".to_string(),
+                    ReplicationStats {
+                        pending_count: 1,
+                        ..Default::default()
+                    },
+                ),
+            ]),
+            ..Default::default()
+        };
+        assert!(!stats.is_empty(), "a non-empty target must make aggregate replication stats non-empty");
+    }
+
+    #[test]
+    fn size_recursive_prunes_empty_and_preserves_pending_replication_stats() {
+        let root = hash_path("bucket");
+        let child = hash_path("bucket/child");
+        let mut cache = DataUsageCache::default();
+        cache.replace_hashed(&root, &None, &DataUsageEntry::default());
+        cache.replace_hashed(
+            &child,
+            &Some(root.clone()),
+            &DataUsageEntry {
+                replication_stats: Some(ReplicationAllStats::default()),
+                ..Default::default()
+            },
+        );
+
+        assert!(
+            cache
+                .size_recursive("bucket")
+                .expect("bucket usage should flatten")
+                .replication_stats
+                .is_none()
+        );
+
+        cache.replace_hashed(
+            &child,
+            &Some(root.clone()),
+            &DataUsageEntry {
+                replication_stats: Some(ReplicationAllStats {
+                    targets: HashMap::from([(
+                        "arn:test:pending".to_string(),
+                        ReplicationStats {
+                            pending_count: 1,
+                            ..Default::default()
+                        },
+                    )]),
+                    ..Default::default()
+                }),
+                ..Default::default()
+            },
+        );
+
+        let flattened = cache.size_recursive("bucket").expect("bucket usage should flatten");
+        let replication = flattened
+            .replication_stats
+            .expect("pending-only replication stats must survive pruning");
+
+        assert_eq!(replication.targets["arn:test:pending"].pending_count, 1);
+    }
+
    #[test]
    fn test_data_usage_cache_merge_adds_missing_child() {
        let mut base = DataUsageCache::default();
@@ -70,7 +70,8 @@ walkdir.workspace = true
 base64 = { workspace = true }
 rand = { workspace = true, features = ["serde"] }
 chrono = { workspace = true, features = ["serde"] }
-md5 = { workspace = true }
+hex = { workspace = true }
+md-5 = { workspace = true }
 opentelemetry-proto = { workspace = true }
 prost.workspace = true
 sha2 = { workspace = true }
@@ -24,9 +24,10 @@ mod tests {
    use aws_sdk_s3::types::{ChecksumAlgorithm, ChecksumMode, CompletedMultipartUpload, CompletedPart};
    use aws_smithy_http_client::Builder as SmithyHttpClientBuilder;
    use base64::Engine;
+    use md5::{Digest as Md5Digest, Md5};
    use rustfs_rio::{Checksum, ChecksumType as RioChecksumType};
    use serial_test::serial;
-    use sha2::{Digest, Sha256};
+    use sha2::Sha256;
    use tracing::info;

    fn create_s3_client(env: &RustFSTestEnvironment) -> Client {
@@ -70,7 +71,9 @@ mod tests {
    }

    fn content_md5_base64(body: &[u8]) -> String {
-        let digest = md5::compute(body);
+        let mut hasher = Md5::new();
+        hasher.update(body);
+        let digest = hasher.finalize();
        base64::engine::general_purpose::STANDARD.encode(digest.as_slice())
    }

@@ -25,6 +25,7 @@ use hyper::body::Incoming;
 use hyper::server::conn::http1;
 use hyper::service::service_fn;
 use hyper_util::rt::{TokioIo, TokioTimer};
+use md5::{Digest as Md5Digest, Md5};
 use s3s::access::{S3Access, S3AccessContext};
 use s3s::auth::SimpleAuth;
 use s3s::dto::{
@@ -827,13 +828,25 @@ fn ensure_body_growth(current: usize, added: usize) -> S3Result {

 async fn md5_digest(body: Bytes, permit: OwnedSemaphorePermit) -> S3Result<([u8; 16], OwnedSemaphorePermit)> {
    if body.len() < 1024 * 1024 {
-        return Ok((md5::compute(body).0, permit));
+        return Ok((md5_bytes(body), permit));
    }
-    tokio::task::spawn_blocking(move || (md5::compute(body).0, permit))
+    tokio::task::spawn_blocking(move || (md5_bytes(body), permit))
        .await
        .map_err(|error| s3s::s3_error!(InternalError, "MD5 worker failed: {error}"))
 }

+fn md5_bytes(input: impl AsRef<[u8]>) -> [u8; 16] {
+    let mut hasher = Md5::new();
+    hasher.update(input.as_ref());
+    hasher.finalize().into()
+}
+
+fn md5_hex(input: impl AsRef<[u8]>) -> String {
+    let mut hasher = Md5::new();
+    hasher.update(input.as_ref());
+    hex::encode(hasher.finalize())
+}
+
 fn ensure_store_budget(state: &StoreState, removed_bytes: usize, added_bytes: usize, adds_version: bool) -> S3Result {
    let total_bytes = state
        .total_bytes
@@ -1005,7 +1018,7 @@ impl S3 for FakeBackend {
            Some(value) => value,
            None => {
                let (digest, _body_permit) = md5_digest(body.clone(), _body_permit).await?;
-                format!("{:x}", md5::Digest(digest))
+                hex::encode(digest)
            }
        };
        let version = ObjectVersion {
@@ -1208,7 +1221,7 @@ impl S3 for FakeBackend {
        }
        let body = collect_stream(input.body, input.content_length, fault.as_ref(), &self.control).await?;
        let (digest, _body_permit) = md5_digest(body.clone(), _body_permit).await?;
-        let e_tag = format!("{:x}", md5::Digest(digest));
+        let e_tag = hex::encode(digest);
        let mut state = lock(&self.store);
        let existing_bytes = state
            .uploads
@@ -1336,7 +1349,7 @@ impl S3 for FakeBackend {
            .collect();
        let (body, digests, _body_permits) = assemble_multipart(assembly_parts, total_len, _body_permits).await?;
        let part_count = requested.len();
-        let e_tag = source_etag(&headers)?.unwrap_or_else(|| format!("{:x}-{part_count}", md5::compute(digests)));
+        let e_tag = source_etag(&headers)?.unwrap_or_else(|| format!("{}-{part_count}", md5_hex(digests)));
        let version = ObjectVersion {
            version_id: upload.version_id.clone(),
            body,
@@ -45,10 +45,10 @@
 //!   * Parity reconstruction: one data disk is taken offline
 //!     (`take_disk_offline`) and the SAME object matrix is GET both ways while
 //!     the EC 2+2 set rebuilds each large object from the surviving shards. The
-//!     codec-streaming reader gate never inspects drive health, so the codec
-//!     fast path is exercised end-to-end through reconstruction; the test
-//!     asserts byte- and header-equality vs the legacy path AND that the codec
-//!     phase never fell back to a duplex pipe while reconstructing.
+//!     eager first/single-part setup may keep its conservative whole-request
+//!     fallback when shard placement makes codec streaming unsafe, so this phase
+//!     asserts byte- and header-equality vs the legacy path rather than requiring
+//!     zero duplex fallbacks under degraded drive health.
 //!   * Missing object: a GET for an absent key is compared across both phases
 //!     to prove the error semantics (HTTP status + S3 error code) are identical
 //!     — the codec env must not perturb the NoSuchKey negative path.
@@ -475,15 +475,14 @@ mod tests {
            "ranged GET length diverged with codec streaming enabled"
        );

-        // ---- Phase B degraded: the same reconstruction, now on the codec path ----
-        // Re-run the reconstruction A/B with the codec-streaming gates still
-        // open. The reader gate decision is independent of drive health (it
-        // never inspects disk state), so the codec fast path is exercised
-        // end-to-end while the EC set rebuilds each large object from the
-        // surviving shards — this is a real codec-vs-legacy reconstruction test,
-        // not legacy-vs-legacy. Snapshot the duplex count first (the range GET
-        // above already used the duplex path) so we can measure only the markers
-        // these degraded codec GETs add.
+        // ---- Phase B degraded: the same reconstruction, with codec gates open ----
+        // Re-run the reconstruction A/B with codec-streaming enabled. If eager
+        // first/single-part setup cannot prove the codec path is safe for the
+        // surviving shards, the implementation intentionally preserves the
+        // whole-request legacy fallback; later multipart parts can degrade in
+        // place. This phase verifies parity-reconstructed bytes and headers,
+        // while the healthy phase above remains the strict zero-duplex path
+        // confirmation.
        let dup_codec_before_degraded = count_marker(&codec_log, DUPLEX_MARKER);
        harness.take_disk_offline(0)?;
        let mut codec_degraded: BTreeMap<String, GetView> = BTreeMap::new();
@@ -511,16 +510,11 @@ mod tests {
            );
        }

-        // Path confirmation under reconstruction: the codec fast path must have
-        // served the reconstructed large objects without ever falling back to
-        // the legacy duplex pipe. Without this, the equivalence above could be
-        // legacy-vs-legacy and prove nothing about codec reconstruction.
+        // Keep degraded duplex markers as diagnostic evidence only: eager setup
+        // may fall back before streaming when shard safety cannot be proven.
        sleep(Duration::from_millis(300)).await;
        let dup_codec_degraded = count_marker(&codec_log, DUPLEX_MARKER).saturating_sub(dup_codec_before_degraded);
-        assert_eq!(
-            dup_codec_degraded, 0,
-            "codec phase created {dup_codec_degraded} duplex pipe(s) while reconstructing large objects with disk0 offline; the codec fast path was not exercised under degraded reads (see {codec_log})"
-        );
+        info!(dup_codec_degraded, "codec phase degraded-read legacy duplex marker count");

        info!(
            objects = baseline.len(),
@@ -520,28 +520,33 @@ async fn assert_msgpack_fallback_unchanged(
    Ok(())
 }

-async fn assert_msgpack_decode_observed(
-    collector: &OtlpMetricCollector,
-    before: &BTreeMap<String, u64>,
-    series: &[(&str, &str)],
-) -> TestResult {
+async fn assert_msgpack_decode_observed(collector: &OtlpMetricCollector, before: &BTreeMap<String, u64>) -> TestResult {
    let deadline = Instant::now() + Duration::from_secs(20);
    loop {
        let after = collector.msgpack_json_decode_totals().await;
-        let missing = series
+        let missing = [FALLBACK_REQUEST_DIRECTION, FALLBACK_RESPONSE_DIRECTION]
            .iter()
-            .filter_map(|(direction, message)| {
-                let key = msgpack_decode_metric_key(direction, message, MSGPACK_CODEC_MSGPACK);
-                let before_value = before.get(&key).copied().unwrap_or_default();
-                let after_value = after.get(&key).copied().unwrap_or_default();
-                (after_value <= before_value).then_some(format!("{direction}/{message}/{}", MSGPACK_CODEC_MSGPACK))
+            .filter_map(|direction| {
+                let prefix = format!("{direction}\u{1f}");
+                let suffix = format!("\u{1f}{MSGPACK_CODEC_MSGPACK}");
+                let before_total = before
+                    .iter()
+                    .filter(|(key, _)| key.starts_with(&prefix) && key.ends_with(&suffix))
+                    .map(|(_, value)| *value)
+                    .sum::<u64>();
+                let after_total = after
+                    .iter()
+                    .filter(|(key, _)| key.starts_with(&prefix) && key.ends_with(&suffix))
+                    .map(|(_, value)| *value)
+                    .sum::<u64>();
+                (after_total <= before_total).then_some(format!("{direction}/{}", MSGPACK_CODEC_MSGPACK))
            })
            .collect::<Vec<_>>();
        if missing.is_empty() {
            return Ok(());
        }
        if Instant::now() >= deadline {
-            return Err(format!("timed out waiting for msgpack decode traffic on {missing:?}; totals={after:?}").into());
+            return Err(format!("timed out waiting for msgpack decode traffic for {missing:?}; totals={after:?}").into());
        }
        sleep(Duration::from_millis(100)).await;
    }
@@ -1857,7 +1862,7 @@ async fn four_node_mixed_msgpack_compat_mode_preserves_fallback_controls() -> Te
        PartNumberReaderPathExpectation::new(bucket, multipart_key, &second_part, multipart_body.len(), MULTIPART, LEGACY_DUPLEX),
    )
    .await?;
-    assert_msgpack_decode_observed(&collector, &decode_before, &MSGPACK_FALLBACK_CONTROL_SERIES).await?;
+    assert_msgpack_decode_observed(&collector, &decode_before).await?;
    assert_msgpack_fallback_unchanged(&collector, &fallback_before, &MSGPACK_FALLBACK_CONTROL_SERIES).await?;
    assert_msgpack_decode_errors_unchanged(&collector, &decode_errors_before, &MSGPACK_FALLBACK_CONTROL_SERIES).await?;

@@ -1948,8 +1953,8 @@ async fn four_node_add_tier_converges_after_offline_node_restart_without_second_
    hot.start().await?;

    let tier_name = unique_tier_name();
-    hot.stop_node(3)?;
    let add_tier_response = submit_rustfs_tier(&hot, &cold, &tier_name).await?;
+    hot.stop_node(3)?;
    hot.start_node(3).await?;

    wait_for_tier_converged(&hot, &tier_name, &add_tier_response).await
@@ -2003,7 +2008,7 @@ async fn four_node_manual_transition_job_status_survives_node_restart() -> TestR
    assert_eq!(cancel_value["job_id"].as_str(), Some(job_id.as_str()));
    assert_eq!(cancel_value["bucket"].as_str(), Some(bucket.as_str()));
    assert_eq!(cancel_value["dry_run"].as_bool(), Some(true));
-    assert_eq!(cancel_value["cancel_requested"].as_bool(), Some(true));
+    assert_eq!(cancel_value["cancel_requested"].as_bool(), Some(false));

    let missing_job_id = Uuid::new_v4();
    let (missing_status, missing_body) = signed_admin_request(
@@ -2325,7 +2330,7 @@ async fn four_node_mixed_msgpack_compat_mode_preserves_fallback_controls_during_
        PartNumberReaderPathExpectation::new(bucket, key, &second_part, body.len(), REMOTE, REMOTE_TRANSITION),
    )
    .await?;
-    assert_msgpack_decode_observed(&collector, &decode_before, &MSGPACK_FALLBACK_CONTROL_SERIES).await?;
+    assert_msgpack_decode_observed(&collector, &decode_before).await?;

    let encrypted_key = "transition/encrypted-sse.bin";
    let encrypted_body = payload(16 * KIB, 0xAB);
@@ -13,8 +13,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

-//! Cross-process replay / tamper acceptance for the internode NodeService v2 RPC
-//! signature (<https://github.com/rustfs/backlog/issues/1327>).
+//! Cross-process replay / tamper acceptance for internode NodeService RPC
+//! signatures (<https://github.com/rustfs/backlog/issues/1327>,
+//! <https://github.com/rustfs/backlog/issues/1542>).
 //!
 //! # Why this exists on top of the in-process tests
 //!
@@ -78,6 +79,8 @@
 //! | mixed version: legacy-only still served, not blocked | [`legacy_only_signature_is_accepted_in_default_posture`] |
 //! | strict flip closes the signature downgrade | [`signature_strict_rejects_legacy_only_downgrade`] |
 //! | strict flip closes the body-digest downgrade, incl. v1 | [`body_digest_strict_rejects_digestless_mutation`] |
+//! | replay scope binds every RPC and rejects restart replay | [`replay_scope_rejects_replay_path_transplant_and_stale_epoch_e2e`] |
+//! | strict replay scope allows only Ping bootstrap before v3 | [`replay_scope_strict_requires_v3_after_ping_bootstrap_e2e`] |
 //!
 //! Two acceptance items are deliberately left to the in-process tests. A stale
 //! timestamp cannot be forged from outside — it is inside the HMAC — so
@@ -88,18 +91,20 @@

 use crate::common::{RustFSTestEnvironment, init_logging};
 use crate::storage_api::internode_rpc_signature::{
-    TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_signature_headers, node_service_time_out_client_no_auth,
+    TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
+    node_service_time_out_client_no_auth, verify_tonic_boot_epoch_response,
 };
 use http::{HeaderMap, Method};
 use rustfs_config::{
-    ENV_INTERNODE_RPC_BODY_DIGEST_STRICT, ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY, ENV_INTERNODE_RPC_SIGNATURE_STRICT,
+    ENV_INTERNODE_RPC_BODY_DIGEST_STRICT, ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY, ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT,
+    ENV_INTERNODE_RPC_SIGNATURE_STRICT,
 };
 use rustfs_protos::canonical_make_volume_request_body;
-use rustfs_protos::proto_gen::node_service::{MakeVolumeRequest, MakeVolumeResponse};
+use rustfs_protos::proto_gen::node_service::{MakeVolumeRequest, MakeVolumeResponse, PingRequest, PingResponse};
 use serial_test::serial;
 use sha2::{Digest, Sha256};
 use std::error::Error;
-use tonic::{Code, Request, Status};
+use tonic::{Code, Request, Response, Status};
 use uuid::Uuid;

 type TestResult = Result<(), Box<dyn Error + Send + Sync>>;
@@ -115,12 +120,14 @@ const TEST_RPC_SECRET: &str = "rustfs-internode-signature-e2e-secret";
 /// clears authentication stops harmlessly at `find_disk`.
 const ABSENT_DISK: &str = "/nonexistent/rustfs-signature-e2e-disk";

-/// Wire names of the two v2 headers these tests edit. They are `pub(crate)` in
-/// ecstore, so they are repeated here rather than imported — [`overwrite_header`]
-/// asserts the header it replaces was actually present, which turns a rename
-/// into a loud failure instead of silently reducing an attack to a no-op.
+/// Wire names of the v2 and replay-scope headers these black-box tests edit. They are
+/// `pub(crate)` in ecstore, so they are repeated here rather than imported.
+/// [`overwrite_header`] asserts the header it replaces was actually present, which turns a
+/// rename into a loud failure instead of silently reducing an attack to a no-op.
 const CONTENT_SHA256_HEADER: &str = "x-rustfs-content-sha256";
 const NONCE_HEADER: &str = "x-rustfs-rpc-nonce";
+const TIMESTAMP_HEADER: &str = "x-rustfs-timestamp";
+const BOOT_EPOCH_CHALLENGE_HEADER: &str = "x-rustfs-rpc-boot-epoch-challenge";

 /// gRPC service name carried in the signed scope, i.e. `TONIC_RPC_PREFIX`
 /// without its leading `/`.
@@ -155,19 +162,28 @@ fn align_rpc_secret_with_server() {
 ///
 /// Uses the no-cleanup spawn so a `pkill` pattern cannot reap servers belonging
 /// to other tests running in the same binary.
-async fn start_server(extra_env: &[(&str, &str)]) -> Result<RustFSTestEnvironment, Box<dyn Error + Send + Sync>> {
-    let mut env = RustFSTestEnvironment::new().await?;
+fn server_env(extra_env: &[(&'static str, &'static str)]) -> Vec<(&'static str, &'static str)> {
    let mut child_env = vec![
        ("RUSTFS_RPC_SECRET", TEST_RPC_SECRET),
        (ENV_INTERNODE_RPC_SIGNATURE_STRICT, "false"),
        (ENV_INTERNODE_RPC_BODY_DIGEST_STRICT, "false"),
+        (ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT, "false"),
        (ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY, "1048576"),
    ];
    child_env.extend_from_slice(extra_env);
-    env.start_rustfs_server_without_cleanup_with_env(&child_env).await?;
+    child_env
+}
+
+async fn start_server_with_env(child_env: &[(&str, &str)]) -> Result<RustFSTestEnvironment, Box<dyn Error + Send + Sync>> {
+    let mut env = RustFSTestEnvironment::new().await?;
+    env.start_rustfs_server_without_cleanup_with_env(child_env).await?;
    Ok(env)
 }

+async fn start_server(extra_env: &[(&'static str, &'static str)]) -> Result<RustFSTestEnvironment, Box<dyn Error + Send + Sync>> {
+    start_server_with_env(&server_env(extra_env)).await
+}
+
 /// Stop the child and drop the cached gRPC channel for its address.
 ///
 /// `node_service_time_out_client_no_auth` memoises channels in a process-global
@@ -238,12 +254,83 @@ fn overwrite_header(headers: &mut HeaderMap, name: &'static str, value: &str) {
 /// and nothing else — no interceptor adds or rewrites auth metadata, so the
 /// bytes on the wire are the ones the test chose.
 async fn call_make_volume(url: &str, request: MakeVolumeRequest, headers: HeaderMap) -> Result<MakeVolumeResponse, Status> {
+    call_make_volume_response(url, request, headers)
+        .await
+        .map(Response::into_inner)
+}
+
+async fn call_make_volume_response(
+    url: &str,
+    request: MakeVolumeRequest,
+    headers: HeaderMap,
+) -> Result<Response<MakeVolumeResponse>, Status> {
    let mut client = node_service_time_out_client_no_auth(&url.to_string())
        .await
        .map_err(|err| Status::unavailable(format!("cannot reach the node service: {err}")))?;
    let mut rpc_request = Request::new(request);
    rpc_request.metadata_mut().as_mut().extend(headers);
-    client.make_volume(rpc_request).await.map(|response| response.into_inner())
+    client.make_volume(rpc_request).await
+}
+
+async fn call_ping_response(url: &str, headers: HeaderMap) -> Result<Response<PingResponse>, Status> {
+    let mut client = node_service_time_out_client_no_auth(&url.to_string())
+        .await
+        .map_err(|err| Status::unavailable(format!("cannot reach the node service: {err}")))?;
+    let mut rpc_request = Request::new(PingRequest {
+        version: 1,
+        body: bytes::Bytes::new(),
+    });
+    rpc_request.metadata_mut().as_mut().extend(headers);
+    client.ping(rpc_request).await
+}
+
+fn attach_boot_epoch_challenge(headers: &mut HeaderMap) -> Uuid {
+    let challenge = Uuid::new_v4();
+    headers.insert(
+        BOOT_EPOCH_CHALLENGE_HEADER,
+        challenge.to_string().parse().expect("UUID must be a valid header value"),
+    );
+    challenge
+}
+
+fn mint_replay_scope_headers(audience: &str, path: &str, content_sha256: &str, boot_epoch: Uuid) -> HeaderMap {
+    let mut headers = mint_v2_headers(audience, "MakeVolume", Some(content_sha256));
+    let timestamp = headers
+        .get(TIMESTAMP_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .expect("v2 headers must carry a timestamp")
+        .to_string();
+    headers.extend(
+        gen_tonic_replay_scope_headers(audience, path, &timestamp, content_sha256, boot_epoch)
+            .expect("replay-scope headers must mint with the aligned RPC secret"),
+    );
+    headers
+}
+
+async fn learn_boot_epoch_from_make_volume(url: &str, audience: &str) -> Uuid {
+    let request = make_volume_request("signature-e2e-epoch-bootstrap");
+    let mut headers = mint_v2_headers(audience, "MakeVolume", Some(&canonical_digest(&request)));
+    let challenge = attach_boot_epoch_challenge(&mut headers);
+    let response = call_make_volume_response(url, request, headers)
+        .await
+        .expect("v2 request with epoch challenge must clear default authentication");
+    let boot_epoch = verify_tonic_boot_epoch_response(audience, challenge, response.metadata().as_ref())
+        .expect("server must HMAC-authenticate the advertised boot epoch");
+    assert_authenticated(
+        Ok(response.into_inner()),
+        "a v2 epoch-challenge request in the default replay-scope posture",
+    );
+    boot_epoch
+}
+
+async fn learn_boot_epoch_from_ping(url: &str, audience: &str) -> Uuid {
+    let mut headers = mint_v2_headers(audience, "Ping", None);
+    let challenge = attach_boot_epoch_challenge(&mut headers);
+    let response = call_ping_response(url, headers)
+        .await
+        .expect("v2 Ping with an epoch challenge must bootstrap strict replay scope");
+    verify_tonic_boot_epoch_response(audience, challenge, response.metadata().as_ref())
+        .expect("strict replay-scope Ping must return a valid boot epoch proof")
 }

 /// Assert a call cleared authentication.
@@ -332,6 +419,122 @@ async fn internode_rpc_signature_default_posture_e2e() -> TestResult {
    Ok(())
 }

+/// A replay-scoped signature is usable exactly once against the exact gRPC path and the server
+/// process epoch that minted it. This crosses the child-process boundary twice: the HMAC-protected
+/// epoch is learned from a real response, then the same server is restarted in place to prove its
+/// replacement epoch rejects the captured request even though the nonce cache is necessarily new.
+#[tokio::test]
+#[serial]
+async fn replay_scope_rejects_replay_path_transplant_and_stale_epoch_e2e() -> TestResult {
+    init_logging();
+    align_rpc_secret_with_server();
+    let child_env = server_env(&[]);
+    let mut env = start_server_with_env(&child_env).await?;
+    let url = env.url.clone();
+    let audience = audience_of(&env);
+    let boot_epoch = learn_boot_epoch_from_make_volume(&url, &audience).await;
+
+    let request = make_volume_request("replay-scope-e2e-once");
+    let captured = mint_replay_scope_headers(
+        &audience,
+        &format!("{TONIC_RPC_PREFIX}/MakeVolume"),
+        &canonical_digest(&request),
+        boot_epoch,
+    );
+    assert_authenticated(
+        call_make_volume(&url, request.clone(), captured.clone()).await,
+        "the first replay-scoped mutation delivery",
+    );
+    assert_rejected(
+        call_make_volume(&url, request.clone(), captured).await,
+        Code::Unauthenticated,
+        None,
+        "the same replay-scoped mutation delivered twice",
+    );
+
+    let transplanted =
+        mint_replay_scope_headers(&audience, &format!("{TONIC_RPC_PREFIX}/Ping"), &canonical_digest(&request), boot_epoch);
+    assert_rejected(
+        call_make_volume(&url, request.clone(), transplanted).await,
+        Code::Unauthenticated,
+        None,
+        "a replay-scoped Ping signature transplanted onto MakeVolume",
+    );
+
+    let stale_epoch = mint_replay_scope_headers(
+        &audience,
+        &format!("{TONIC_RPC_PREFIX}/MakeVolume"),
+        &canonical_digest(&request),
+        boot_epoch,
+    );
+    env.restart_server_preserving_data(Vec::new(), &child_env).await?;
+    rustfs_protos::evict_failed_connection(&url).await;
+    assert_rejected(
+        call_make_volume(&url, request.clone(), stale_epoch).await,
+        Code::Unauthenticated,
+        None,
+        "a replay-scoped signature captured before the receiving process restart",
+    );
+
+    let restarted_epoch = learn_boot_epoch_from_make_volume(&url, &audience).await;
+    assert_ne!(boot_epoch, restarted_epoch, "a restarted child process must advertise a new boot epoch");
+    let fresh_epoch = mint_replay_scope_headers(
+        &audience,
+        &format!("{TONIC_RPC_PREFIX}/MakeVolume"),
+        &canonical_digest(&request),
+        restarted_epoch,
+    );
+    assert_authenticated(
+        call_make_volume(&url, request, fresh_epoch).await,
+        "a replay-scoped mutation signed with the replacement process epoch",
+    );
+
+    stop_server(env, &url).await;
+    Ok(())
+}
+
+/// Strict replay scope leaves one authenticated v2 bootstrap: `Ping` carrying a fresh challenge.
+/// A mutating v2 request cannot use that lane; once the epoch proof is returned, the first v3
+/// mutation succeeds. This protects a server restart without reopening a general downgrade path.
+#[tokio::test]
+#[serial]
+async fn replay_scope_strict_requires_v3_after_ping_bootstrap_e2e() -> TestResult {
+    init_logging();
+    align_rpc_secret_with_server();
+    let env = start_server(&[(ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT, "true")]).await?;
+    let url = env.url.clone();
+    let audience = audience_of(&env);
+
+    let v2_request = make_volume_request("replay-scope-e2e-strict-v2");
+    assert_rejected(
+        call_make_volume(
+            &url,
+            v2_request.clone(),
+            mint_v2_headers(&audience, "MakeVolume", Some(&canonical_digest(&v2_request))),
+        )
+        .await,
+        Code::Unauthenticated,
+        None,
+        "a v2 mutation after replay-scope strictness is enabled",
+    );
+
+    let boot_epoch = learn_boot_epoch_from_ping(&url, &audience).await;
+    let request = make_volume_request("replay-scope-e2e-strict-v3");
+    let replay_scoped = mint_replay_scope_headers(
+        &audience,
+        &format!("{TONIC_RPC_PREFIX}/MakeVolume"),
+        &canonical_digest(&request),
+        boot_epoch,
+    );
+    assert_authenticated(
+        call_make_volume(&url, request, replay_scoped).await,
+        "a replay-scoped mutation after Ping bootstrap under strict replay scope",
+    );
+
+    stop_server(env, &url).await;
+    Ok(())
+}
+
 /// Baseline: correctly signed mutations are accepted, both with and without a
 /// body digest.
 ///
@@ -30,6 +30,7 @@ use aws_sdk_s3::primitives::ByteStream;
 use aws_sdk_s3::types::ServerSideEncryption;
 use base64::{Engine, engine::general_purpose::STANDARD as BASE64};
 use http::header::{CONTENT_TYPE, HOST};
+use md5::{Digest as Md5Digest, Md5};
 use rustfs_signer::constants::UNSIGNED_PAYLOAD;
 use rustfs_signer::sign_v4;
 use s3s::Body;
@@ -68,7 +69,9 @@ pub fn skip_if_kms_admin_tool_unavailable(test_name: &str) -> bool {
 }

 pub fn sse_customer_key_md5_base64(key: &str) -> String {
-    BASE64.encode(md5::compute(key).0)
+    let mut hasher = Md5::new();
+    hasher.update(key.as_bytes());
+    BASE64.encode(hasher.finalize())
 }

 pub async fn kms_admin_request(
@@ -25,12 +25,18 @@ use super::common::{LocalKMSTestEnvironment, sse_customer_key_md5_base64};
 use crate::common::{TEST_BUCKET, init_logging};
 use aws_sdk_s3::types::ServerSideEncryption;
 use base64::Engine;
-use md5::compute;
+use md5::{Digest as Md5Digest, Md5};
 use serial_test::serial;
 use std::sync::Arc;
 use tokio::sync::Semaphore;
 use tracing::{info, warn};

+fn md5_hex(input: impl AsRef<[u8]>) -> String {
+    let mut hasher = Md5::new();
+    hasher.update(input.as_ref());
+    hex::encode(hasher.finalize())
+}
+
 /// Test encryption of zero-byte files (empty files)
 #[tokio::test]
 #[serial]
@@ -294,7 +300,7 @@ async fn test_kms_invalid_key_scenarios() -> Result<(), Box<dyn std::error::Erro
    info!("🔍 Testing invalid SSE-C key length");
    let invalid_short_key = "short"; // Too short
    let invalid_key_b64 = base64::engine::general_purpose::STANDARD.encode(invalid_short_key);
-    let invalid_key_md5 = format!("{:x}", compute(invalid_short_key));
+    let invalid_key_md5 = md5_hex(invalid_short_key);

    let invalid_key_result = s3_client
        .put_object()
@@ -26,6 +26,7 @@ use chrono::{Duration as ChronoDuration, Utc};
 use flate2::{Compression, write::GzEncoder};
 use http::HeaderValue;
 use http::header::{CONTENT_TYPE, HOST};
+use md5::{Digest as Md5Digest, Md5};
 use rustfs_signer::constants::UNSIGNED_PAYLOAD;
 use rustfs_signer::sign_v4;
 use s3s::Body;
@@ -50,7 +51,15 @@ fn encode_post_policy(conditions: Vec<serde_json::Value>) -> String {
 }

 fn sse_customer_key_md5_base64(key: &str) -> String {
-    base64::engine::general_purpose::STANDARD.encode(md5::compute(key).0)
+    let mut hasher = Md5::new();
+    hasher.update(key.as_bytes());
+    base64::engine::general_purpose::STANDARD.encode(hasher.finalize())
+}
+
+fn md5_hex(input: impl AsRef<[u8]>) -> String {
+    let mut hasher = Md5::new();
+    hasher.update(input.as_ref());
+    hex::encode(hasher.finalize())
 }

 /// Env var consumed by the local SSE-S3 DEK provider when KMS is not configured.
@@ -5664,7 +5673,7 @@ async fn test_signed_put_object_extract_returns_archive_etag() -> Result<(), Box
    client.create_bucket().bucket(bucket).send().await?;

    let archive = make_tar(&[("alpha.txt", b"alpha-body")], &[]).await;
-    let expected_etag = format!("\"{:x}\"", md5::compute(&archive));
+    let expected_etag = format!("\"{}\"", md5_hex(&archive));

    let response = client
        .put_object()
@@ -21,18 +21,22 @@
 //! function, never as an S3 event sink.
 //!
 //! Coverage:
-//!   * PUT / multipart-complete / DELETE each deliver one event with the correct
+//!   * PUT / multipart-complete / DeleteObject / DeleteObjects each deliver one event with the correct
 //!     eventName, bucket, key, versionId and eTag.
 //!   * prefix/suffix filters drop non-matching keys (rule-engine gate).
 //!   * an event queued while the target endpoint is unreachable is redelivered
 //!     from the on-disk store once the endpoint recovers (store-and-forward).
+//!   * responseElements and the S3 response use the canonical request ID while
+//!     requestParameters preserve a conflicting client-supplied value.

 use crate::common::{RustFSTestEnvironment, init_logging};
 use aws_sdk_s3::Client;
+use aws_sdk_s3::operation::RequestId;
 use aws_sdk_s3::primitives::ByteStream;
 use aws_sdk_s3::types::{
-    BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, Event, FilterRule, FilterRuleName,
-    NotificationConfiguration, NotificationConfigurationFilter, QueueConfiguration, S3KeyFilter, VersioningConfiguration,
+    BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, Delete, Event, FilterRule, FilterRuleName,
+    NotificationConfiguration, NotificationConfigurationFilter, ObjectIdentifier, QueueConfiguration, S3KeyFilter,
+    VersioningConfiguration,
 };
 use http::header::{CONTENT_TYPE, HOST};
 use local_ip_address::local_ip;
@@ -40,10 +44,12 @@ use reqwest::StatusCode;
 use rustfs_signer::constants::UNSIGNED_PAYLOAD;
 use rustfs_signer::sign_v4;
 use rustfs_utils::egress::ENV_OUTBOUND_ALLOW_ORIGINS;
+use rustfs_utils::http::headers::{AMZ_REQUEST_ID, REQUEST_ID_HEADER};
 use s3s::Body;
 use serde_json::Value;
 use serial_test::serial;
 use std::error::Error;
+use std::io::Cursor;
 use std::path::Path;
 use std::sync::{
    Arc, Once,
@@ -63,6 +69,8 @@ type BoxError = Box<dyn Error + Send + Sync>;
 /// for target lookup (see `process_queue_configurations`), so the region is
 /// nominal, but it must be present for `ARN::parse` to succeed.
 const NOTIFY_REGION: &str = "us-east-1";
+const CLIENT_REQUEST_ID: &str = "client-supplied-request-id";
+const CLIENT_AMZ_REQUEST_ID: &str = "client-supplied-amz-request-id";

 /// Webhook targets are registered as `TargetID { id: <name>, name: "webhook" }`,
 /// so the ARN a notification rule references is
@@ -579,6 +587,36 @@ fn trimmed_etag(value: Option<&str>) -> Option<String> {
    value.map(|e| e.trim_matches('"').to_string())
 }

+fn assert_conflicting_request_id_correlation(record: &Value, server_request_id: &str) {
+    assert_eq!(
+        record["requestParameters"][REQUEST_ID_HEADER].as_str(),
+        Some(CLIENT_REQUEST_ID),
+        "notification request parameters should retain the actual client header: {record}"
+    );
+    assert_eq!(
+        record["requestParameters"][AMZ_REQUEST_ID].as_str(),
+        Some(CLIENT_AMZ_REQUEST_ID),
+        "notification request parameters should retain the actual client header: {record}"
+    );
+    assert_eq!(
+        record["responseElements"][AMZ_REQUEST_ID].as_str(),
+        Some(server_request_id),
+        "notification response elements should use the canonical request ID: {record}"
+    );
+}
+
+fn assert_generated_request_id_correlation(record: &Value, request_id: &str) {
+    assert!(
+        record["requestParameters"][AMZ_REQUEST_ID].is_null(),
+        "notification request parameters must not invent a client request header: {record}"
+    );
+    assert_eq!(
+        record["responseElements"][AMZ_REQUEST_ID].as_str(),
+        Some(request_id),
+        "notification response elements should match the S3 response request ID: {record}"
+    );
+}
+
 // ---------------------------------------------------------------------------
 // Tests
 // ---------------------------------------------------------------------------
@@ -671,8 +709,17 @@ async fn test_webhook_event_delivery_and_filtering() -> TestResult {
        .bucket(bucket)
        .key(put_key)
        .body(ByteStream::from_static(b"peri-1 put body"))
+        .customize()
+        .mutate_request(|request| {
+            request.headers_mut().insert(REQUEST_ID_HEADER, CLIENT_REQUEST_ID);
+            request.headers_mut().insert(AMZ_REQUEST_ID, CLIENT_AMZ_REQUEST_ID);
+        })
        .send()
        .await?;
+    let put_request_id = put.request_id().ok_or("PUT response missing request ID")?.to_owned();
+    assert!(uuid::Uuid::parse_str(&put_request_id).is_ok());
+    assert_ne!(put_request_id, CLIENT_REQUEST_ID);
+    assert_ne!(put_request_id, CLIENT_AMZ_REQUEST_ID);
    let put_version = put
        .version_id()
        .ok_or("PUT response missing versionId (versioning not enabled?)")?;
@@ -692,6 +739,7 @@ async fn test_webhook_event_delivery_and_filtering() -> TestResult {
        "record eventName: {record}"
    );
    assert_eq!(record["s3"]["bucket"]["name"].as_str(), Some(bucket), "bucket in event: {record}");
+    assert_conflicting_request_id_correlation(record, &put_request_id);
    assert_eq!(object["versionId"].as_str(), Some(put_version), "versionId in event: {object}");
    assert_eq!(
        trimmed_etag(object["eTag"].as_str()),
@@ -744,6 +792,35 @@ async fn test_webhook_event_delivery_and_filtering() -> TestResult {
        "multipart eTag in event: {mp_record}"
    );

+    // --- Snowball extract: direct notification path keeps response correlation
+    let snowball_key = "uploads/snowball.dat";
+    let snowball_body = b"snowball notification body";
+    let mut archive_builder = tokio_tar::Builder::new(Cursor::new(Vec::new()));
+    let mut archive_header = tokio_tar::Header::new_gnu();
+    archive_header.set_size(u64::try_from(snowball_body.len()).expect("snowball fixture length should fit in u64"));
+    archive_header.set_mode(0o644);
+    archive_header.set_cksum();
+    archive_builder
+        .append_data(&mut archive_header, snowball_key, Cursor::new(snowball_body))
+        .await?;
+    let archive = archive_builder.into_inner().await?.into_inner();
+    let snowball = client
+        .put_object()
+        .bucket(bucket)
+        .key("snowball-fixture.tar")
+        .body(ByteStream::from(archive))
+        .customize()
+        .mutate_request(|request| {
+            request.headers_mut().insert("x-amz-meta-snowball-auto-extract", "true");
+        })
+        .send()
+        .await?;
+    let snowball_request_id = snowball.request_id().ok_or("Snowball response missing request ID")?;
+
+    let snowball_event = wait_for_event(&mut rx, snowball_key, "s3:ObjectCreated:", Duration::from_secs(20)).await?;
+    let snowball_record = &snowball_event["Records"][0];
+    assert_generated_request_id_correlation(snowball_record, snowball_request_id);
+
    // --- Filter: non-matching prefix and suffix must never be delivered ------
    let wrong_prefix = "logs/report.dat"; // right suffix, wrong prefix
    let wrong_suffix = "uploads/report.txt"; // right prefix, wrong suffix
@@ -772,7 +849,32 @@ async fn test_webhook_event_delivery_and_filtering() -> TestResult {
        "wrong-suffix key {wrong_suffix} bypassed the filter; saw {seen:?}"
    );

-    // --- DELETE on a versioned bucket: ObjectRemoved:* with delete-marker version
+    // --- DeleteObjects: direct notification path keeps response correlation --
+    let delete_many_key = "uploads/delete-many.dat";
+    client
+        .put_object()
+        .bucket(bucket)
+        .key(delete_many_key)
+        .body(ByteStream::from_static(b"delete objects notification body"))
+        .send()
+        .await?;
+    wait_for_event(&mut rx, delete_many_key, "s3:ObjectCreated:", Duration::from_secs(20)).await?;
+
+    let delete_many = client
+        .delete_objects()
+        .bucket(bucket)
+        .delete(
+            Delete::builder()
+                .objects(ObjectIdentifier::builder().key(delete_many_key).build()?)
+                .build()?,
+        )
+        .send()
+        .await?;
+    let delete_many_request_id = delete_many.request_id().ok_or("DeleteObjects response missing request ID")?;
+    let delete_many_event = wait_for_event(&mut rx, delete_many_key, "s3:ObjectRemoved:", Duration::from_secs(20)).await?;
+    assert_generated_request_id_correlation(&delete_many_event["Records"][0], delete_many_request_id);
+
+    // --- DeleteObject on a versioned bucket: delete-marker version ----------
    let delete = client.delete_object().bucket(bucket).key(put_key).send().await?;
    let removed = wait_for_event(&mut rx, put_key, "s3:ObjectRemoved:", Duration::from_secs(20)).await?;
    let removed_record = &removed["Records"][0];
@@ -24,7 +24,7 @@ use rustfs_protos::proto_gen::node_service::{BatchGenerallyLockRequest, Generall
 use tonic::Request;
 use tracing::{info, warn};

-use crate::storage_api::grpc_lock::{TonicInterceptor, node_service_time_out_client_no_auth};
+use crate::storage_api::grpc_lock::{AuthenticatedChannel, TonicInterceptor, node_service_time_out_client_no_auth};

 /// gRPC lock client without authentication for testing
 /// Similar to RemoteClient but uses no_auth client
@@ -42,7 +42,7 @@ impl GrpcLockClient {
        &self,
    ) -> Result<
        rustfs_protos::proto_gen::node_service::node_service_client::NodeServiceClient<
-            tonic::service::interceptor::InterceptedService<tonic::transport::Channel, TonicInterceptor>,
+            tonic::service::interceptor::InterceptedService<AuthenticatedChannel, TonicInterceptor>,
        >,
    > {
        node_service_time_out_client_no_auth(&self.addr)
@@ -23,7 +23,8 @@ use rustfs_protos::{
    proto_gen::node_service::{
        BatchGenerallyLockRequest, BatchGenerallyLockResponse, BatchReadVersionRequest, BatchReadVersionResponse,
        GenerallyLockRequest, GenerallyLockResponse, GenerallyLockResult, PingRequest, PingResponse,
-        node_service_server::NodeService,
+        SnapshotLeaseMutationResponse, SnapshotLeaseReleaseRequest, SnapshotLeaseRenewRequest, SnapshotLeaseRequest,
+        SnapshotLeaseResponse, node_service_server::NodeService,
    },
 };
 use std::pin::Pin;
@@ -104,6 +105,27 @@ impl NodeService for MinimalLockNodeService {
        Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
    }

+    async fn acquire_snapshot_lease(
+        &self,
+        _request: Request<SnapshotLeaseRequest>,
+    ) -> Result<Response<SnapshotLeaseResponse>, Status> {
+        Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
+    }
+
+    async fn renew_snapshot_lease(
+        &self,
+        _request: Request<SnapshotLeaseRenewRequest>,
+    ) -> Result<Response<SnapshotLeaseResponse>, Status> {
+        Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
+    }
+
+    async fn release_snapshot_lease(
+        &self,
+        _request: Request<SnapshotLeaseReleaseRequest>,
+    ) -> Result<Response<SnapshotLeaseMutationResponse>, Status> {
+        Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
+    }
+
    async fn lock(&self, request: Request<GenerallyLockRequest>) -> Result<Response<GenerallyLockResponse>, Status> {
        let request = request.into_inner();
        let args: LockRequest = match serde_json::from_str(&request.args) {
@@ -400,6 +422,20 @@ impl NodeService for MinimalLockNodeService {
        Err(Status::unimplemented("lock-only test server"))
    }

+    async fn prepare_part_transaction(
+        &self,
+        _request: Request<rustfs_protos::proto_gen::node_service::PreparePartTransactionRequest>,
+    ) -> Result<Response<rustfs_protos::proto_gen::node_service::PreparePartTransactionResponse>, Status> {
+        Err(Status::unimplemented("lock-only test server"))
+    }
+
+    async fn settle_part_transaction(
+        &self,
+        _request: Request<rustfs_protos::proto_gen::node_service::SettlePartTransactionRequest>,
+    ) -> Result<Response<rustfs_protos::proto_gen::node_service::SettlePartTransactionResponse>, Status> {
+        Err(Status::unimplemented("lock-only test server"))
+    }
+
    async fn rename_file(
        &self,
        _request: Request<rustfs_protos::proto_gen::node_service::RenameFileRequest>,
@@ -95,6 +95,7 @@ const MANUAL_ASYNC_PARALLEL_OBJECTS: usize = 64;
 const MANUAL_ACTIVE_CANCEL_OBJECTS: usize = 512;
 const MANUAL_RESTART_CANCEL_OBJECTS: usize = 512;
 const MANUAL_ACTIVE_CANCEL_RUNNING_TIMEOUT: StdDuration = StdDuration::from_secs(15);
+const MANUAL_TRANSITION_CANCEL_BARRIER_ENV: &str = "RUSTFS_E2E_MANUAL_TRANSITION_CANCEL_BARRIER";
 const MANUAL_ASYNC_CONFLICT_TERMINAL_TIMEOUT: StdDuration = StdDuration::from_secs(90);
 const MANUAL_RESTART_RECOVERY_TIMEOUT: StdDuration = StdDuration::from_secs(80);
 const OBJECT_KEY: &str = "tier/鲁A12345/report.bin";
@@ -1684,8 +1685,15 @@ async fn test_manual_transition_async_active_cancel_reports_terminal_cancelled()
    cold_client.create_bucket().bucket(TIER_BUCKET).send().await?;

    let mut hot = RustFSTestEnvironment::new().await?;
-    hot.start_rustfs_server_with_env(vec![], &[("RUSTFS_SCANNER_ENABLED", "false"), ("RUSTFS_SCANNER_CYCLE", "3600")])
-        .await?;
+    hot.start_rustfs_server_with_env(
+        vec![],
+        &[
+            ("RUSTFS_SCANNER_ENABLED", "false"),
+            ("RUSTFS_SCANNER_CYCLE", "3600"),
+            (MANUAL_TRANSITION_CANCEL_BARRIER_ENV, "1"),
+        ],
+    )
+    .await?;
    let hot_client = hot.create_s3_client();
    add_rustfs_tier(&hot, &cold).await?;

@@ -22,6 +22,7 @@ use aws_sdk_s3::primitives::ByteStream;
 use aws_sdk_s3::types::{BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, VersioningConfiguration};
 use aws_smithy_http_client::Builder as SmithyHttpClientBuilder;
 use base64::Engine;
+use md5::{Digest as Md5Digest, Md5};
 use std::collections::HashMap;
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex};
@@ -102,10 +103,12 @@ impl Intercept for ResponseHeaderCapture {

 fn customer_key(byte: u8) -> CustomerKey {
    let raw = [byte; 32];
+    let mut hasher = Md5::new();
+    hasher.update(raw);
    CustomerKey {
        raw: String::from_utf8_lossy(&raw).into_owned(),
        encoded: base64::engine::general_purpose::STANDARD.encode(raw),
-        md5: base64::engine::general_purpose::STANDARD.encode(md5::compute(raw).0),
+        md5: base64::engine::general_purpose::STANDARD.encode(hasher.finalize()),
    }
 }

@@ -16,9 +16,12 @@
 pub(crate) use rustfs_ecstore::api::bucket::bucket_target_sys::BucketTargetSys;
 #[cfg(test)]
 pub(crate) use rustfs_ecstore::api::disk::{VolumeInfo, WalkDirOptions};
+pub(crate) use rustfs_ecstore::api::rpc::{AuthenticatedChannel, TonicInterceptor, node_service_time_out_client_no_auth};
 #[cfg(test)]
-pub(crate) use rustfs_ecstore::api::rpc::{TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_signature_headers};
-pub(crate) use rustfs_ecstore::api::rpc::{TonicInterceptor, node_service_time_out_client_no_auth};
+pub(crate) use rustfs_ecstore::api::rpc::{
+    TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
+    verify_tonic_boot_epoch_response,
+};
 #[cfg(test)]
 pub(crate) use rustfs_ecstore::api::rpc::{gen_tonic_signature_interceptor, node_service_time_out_client};

@@ -30,7 +33,7 @@ pub(crate) mod node_interact {
 }

 pub(crate) mod grpc_lock {
-    pub(crate) use super::{TonicInterceptor, node_service_time_out_client_no_auth};
+    pub(crate) use super::{AuthenticatedChannel, TonicInterceptor, node_service_time_out_client_no_auth};
 }

 /// Signing/transport surface used by the cross-process internode RPC signature
@@ -40,7 +43,8 @@ pub(crate) mod grpc_lock {
 #[cfg(test)]
 pub(crate) mod internode_rpc_signature {
    pub(crate) use super::{
-        TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_signature_headers, node_service_time_out_client_no_auth,
+        TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
+        node_service_time_out_client_no_auth, verify_tonic_boot_epoch_response,
    };
 }

@@ -153,11 +153,18 @@ metrics = { workspace = true }
 [target.'cfg(target_os = "linux")'.dependencies]
 rustfs-uring = "0.2.1"

+[target.'cfg(windows)'.dependencies]
+winapi-util.workspace = true
+windows-sys = { workspace = true, features = ["Win32_Foundation", "Win32_Storage_FileSystem"] }
+
 [dev-dependencies]
 tokio = { workspace = true, features = ["rt-multi-thread", "macros", "test-util", "fs"] }
 criterion = { workspace = true, features = ["html_reports"] }
 temp-env = { workspace = true, features = ["async_closure"] }
 tracing-subscriber = { workspace = true, features = ["json", "env-filter", "time"] }
+# Only for `pin_callsite_interest_for_test`, which registers a `NoSubscriber`
+# dispatcher to keep tracing's process-global callsite-interest cache honest.
+tracing-core = { workspace = true }
 serial_test = { workspace = true }
 opentelemetry_sdk = { workspace = true, features = ["rt-tokio"] }
 proptest = "1"
@@ -67,6 +67,14 @@ pub mod bucket {
            };
        }

+        pub mod transition_transaction {
+            pub use crate::bucket::lifecycle::transition_transaction::{
+                TransitionOperatorDeleteResult, TransitionOperatorError, TransitionOperatorProbe, TransitionOperatorStatus,
+                delete_transition_candidate_for_operator, finalize_missing_transition_transaction_for_operator,
+                inspect_transition_transaction_for_operator,
+            };
+        }
+
        pub mod evaluator {
            pub use crate::bucket::lifecycle::evaluator::Evaluator;
        }
@@ -122,13 +130,13 @@ pub mod bucket {

    pub mod metadata_sys {
        pub use crate::bucket::metadata_sys::{
-            BucketMetadataSys, acquire_bucket_targets_transaction_lock, delete, get, get_accelerate_config, get_bucket_policy,
+            BucketMetadataSys, acquire_bucket_metadata_transaction_lock, delete, get, get_accelerate_config, get_bucket_policy,
            get_bucket_policy_raw, get_bucket_targets_config, get_config_from_disk, get_cors_config, get_durability_config,
            get_global_bucket_metadata_sys, get_lifecycle_config, get_logging_config, get_notification_config,
            get_object_lock_config, get_public_access_block_config, get_quota_config, get_replication_config,
            get_request_payment_config, get_sse_config, get_tagging_config, get_versioning_config, get_website_config,
-            init_bucket_metadata_sys, list_bucket_targets, remove_bucket_metadata, set_bucket_metadata, update,
-            update_bucket_targets_under_transaction_lock,
+            init_bucket_metadata_sys, list_bucket_targets, reload_bucket_metadata, remove_bucket_metadata, set_bucket_metadata,
+            update, update_bucket_targets_under_transaction_lock, update_config_with, update_under_transaction_lock,
        };
    }

@@ -303,9 +311,10 @@ pub mod disk {
    pub use crate::disk::{
        BATCH_READ_VERSION_MAX_ITEMS, BUCKET_META_PREFIX, BatchReadVersionItem, BatchReadVersionReq, BatchReadVersionResp,
        CheckPartsResp, DeleteOptions, Disk, DiskAPI, DiskInfo, DiskInfoOptions, DiskLocation, DiskOption, DiskStore,
-        FileInfoVersions, FileReader, FileWriter, HEALING_MARKER_PATH, NsScannerOpenRequest, OldCurrentSize, RUSTFS_META_BUCKET,
-        ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp, STORAGE_FORMAT_FILE, UpdateMetadataOpts, VolumeInfo,
-        WalkDirOptions, new_disk, validate_batch_read_version_item_count,
+        FileInfoVersions, FileReader, FileWriter, HEALING_MARKER_PATH, NsScannerOpenRequest, OldCurrentSize,
+        PartTransactionAction, RUSTFS_META_BUCKET, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp,
+        STORAGE_FORMAT_FILE, SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, new_disk,
+        validate_batch_read_version_item_count,
    };
    pub use bytes::Bytes;
    pub use endpoint::Endpoint;
@@ -376,6 +385,7 @@ pub mod metrics {
 pub mod notification {
    pub use crate::services::notification_sys::{
        NotificationPeerErr, NotificationSys, get_global_notification_sys, new_global_notification_sys,
+        start_remote_version_state_fleet_probe,
    };
 }

@@ -406,13 +416,15 @@ pub mod rio {

 pub mod rpc {
    pub use crate::cluster::rpc::{
-        LocalPeerS3Client, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS, PeerRestClient, PeerS3Client, S3PeerSys,
-        SERVICE_SIGNAL_REFRESH_CONFIG, SERVICE_SIGNAL_RELOAD_DYNAMIC, ScannerBucketListing, ScannerPeerActivity,
-        TONIC_RPC_PREFIX, TonicInterceptor, gen_signature_headers, gen_tonic_signature_headers, gen_tonic_signature_interceptor,
-        node_service_time_out_client, node_service_time_out_client_no_auth, normalize_tonic_rpc_audience,
-        set_tonic_canonical_body_digest, sign_ns_scanner_capability, sign_tonic_rpc_response_proof, verify_rpc_signature,
-        verify_tonic_canonical_body_digest, verify_tonic_mutation_body_digest, verify_tonic_rpc_response_proof,
-        verify_tonic_rpc_signature,
+        AuthenticatedChannel, LocalPeerS3Client, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS, PeerRestClient,
+        PeerS3Client, S3PeerSys, SERVICE_SIGNAL_REFRESH_CONFIG, SERVICE_SIGNAL_RELOAD_DYNAMIC, ScannerBucketListing,
+        ScannerPeerActivity, TONIC_RPC_PREFIX, TonicInterceptor, gen_signature_headers, gen_tonic_replay_scope_headers,
+        gen_tonic_signature_headers, gen_tonic_signature_interceptor, node_service_time_out_client,
+        node_service_time_out_client_no_auth, normalize_tonic_rpc_audience, set_tonic_canonical_body_digest,
+        sign_ns_scanner_capability, sign_tonic_rpc_response_proof, tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers,
+        verify_rpc_signature, verify_tonic_boot_epoch_response, verify_tonic_canonical_body_digest,
+        verify_tonic_mutation_body_digest, verify_tonic_rpc_response_proof, verify_tonic_rpc_signature,
+        verify_tonic_rpc_signature_with_bootstrap,
    };
 }

@@ -554,6 +554,7 @@ async fn delete_free_version_remote_object(
    oi: &ObjectInfo,
    tier_config_mgr: &Arc<RwLock<TierConfigMgr>>,
 ) -> Result<(), std::io::Error> {
+    let version_id_exact = validate_transition_remote_version(oi)?;
    let identity = tier_destination_id_from_metadata(&oi.user_defined)?
        .ok_or_else(|| std::io::Error::other("tier free-version has no durable backend identity"))?;
    delete_object_from_remote_tier_idempotent_with_manager_and_identity(
@@ -562,7 +563,7 @@ async fn delete_free_version_remote_object(
        &oi.transitioned_object.tier,
        identity,
        tier_config_mgr,
-        false,
+        version_id_exact,
    )
    .await?;
    Ok(())
@@ -1181,8 +1182,12 @@ impl TransitionState {
    }

    fn new_with_capacity(capacity: usize) -> Arc<Self> {
-        let capacity = capacity.max(1);
        let queue_send_timeout = resolve_transition_queue_send_timeout();
+        Self::new_with_capacity_and_timeout(capacity, queue_send_timeout)
+    }
+
+    fn new_with_capacity_and_timeout(capacity: usize, queue_send_timeout: StdDuration) -> Arc<Self> {
+        let capacity = capacity.max(1);
        let (tx1, rx1) = bounded(capacity);
        Arc::new(Self {
            transition_tx: tx1,
@@ -1249,13 +1254,12 @@ impl TransitionState {
            return false;
        }
        let bucket = bucket.to_string();
-        let scheduled = Arc::clone(&self.compensation_buckets);
        let state = Arc::clone(self);
        tokio::spawn(async move {
            Self::inc_counter(&state.compensation_running_tasks);
            state.record_scanner_transition_state();
            let Some(api) = runtime_sources::object_store_handle() else {
-                scheduled.lock().unwrap().remove(&bucket);
+                state.finish_bucket_compensation(&bucket);
                Self::add_counter(&state.compensation_running_tasks, -1);
                state.record_scanner_transition_state();
                debug!(
@@ -1291,13 +1295,25 @@ impl TransitionState {
                );
            }

-            scheduled.lock().unwrap().remove(&bucket);
+            state.finish_bucket_compensation(&bucket);
            Self::add_counter(&state.compensation_running_tasks, -1);
            state.record_scanner_transition_state();
        });
        true
    }

+    fn finish_bucket_compensation(&self, bucket: &str) {
+        match self.compensation_buckets.lock() {
+            Ok(mut scheduled) => {
+                scheduled.remove(bucket);
+            }
+            Err(poisoned) => {
+                poisoned.into_inner().remove(bucket);
+                self.compensation_buckets.clear_poison();
+            }
+        }
+    }
+
    #[inline]
    fn inc_counter(counter: &AtomicI64) {
        Self::add_counter(counter, 1);
@@ -1534,17 +1550,35 @@ impl TransitionState {

        let outcome = match self.transition_tx.try_send(Some(task)) {
            Ok(()) => TransitionEnqueueOutcome::Queued,
-            Err(async_channel::TrySendError::Full(_)) => {
+            Err(async_channel::TrySendError::Full(task)) => {
                Self::inc_counter(&self.queue_full_tasks);
-                debug!(
-                    bucket = %oi.bucket,
-                    object = %oi.name,
-                    source = ?src,
-                    "transition queue is full; deferring to scanner/backfill"
-                );
-                TransitionEnqueueOutcome::QueueFull
+                let send_timeout = self.transition_queue_send_timeout;
+                match tokio::time::timeout(send_timeout, self.transition_tx.send(task)).await {
+                    Ok(Ok(())) => TransitionEnqueueOutcome::Queued,
+                    Ok(Err(_)) => {
+                        self.schedule_bucket_compensation(&oi.bucket);
+                        TransitionEnqueueOutcome::QueueClosed
+                    }
+                    Err(_) => {
+                        Self::inc_counter(&self.queue_send_timeout_tasks);
+                        self.schedule_bucket_compensation(&oi.bucket);
+                        debug!(
+                            bucket = %oi.bucket,
+                            object = %oi.name,
+                            source = ?src,
+                            timeout_ms = send_timeout.as_millis() as u64,
+                            event = EVENT_LIFECYCLE_TRANSITION_COMPENSATION,
+                            component = LOG_COMPONENT_ECSTORE,
+                            subsystem = LOG_SUBSYSTEM_LIFECYCLE,
+                            state = "queue_send_timed_out",
+                            "Scanner transition enqueue timed out; scheduled bucket compensation"
+                        );
+                        TransitionEnqueueOutcome::QueueFull
+                    }
+                }
            }
            Err(async_channel::TrySendError::Closed(_)) => {
+                self.schedule_bucket_compensation(&oi.bucket);
                debug!(
                    bucket = %oi.bucket,
                    object = %oi.name,
@@ -1729,7 +1763,7 @@ impl TransitionState {
    }

    pub fn add_lastday_stats(&self, tier: &str, ts: TierStats) {
-        let mut tier_stats = self.last_day_stats.lock().unwrap();
+        let mut tier_stats = self.lock_last_day_stats();
        tier_stats
            .entry(tier.to_string())
            .and_modify(|e| e.add_stats(ts))
@@ -1737,7 +1771,7 @@ impl TransitionState {
    }

    pub fn get_daily_all_tier_stats(&self) -> DailyAllTierStats {
-        let tier_stats = self.last_day_stats.lock().unwrap();
+        let tier_stats = self.lock_last_day_stats();
        let mut res = DailyAllTierStats::with_capacity(tier_stats.len());
        for (tier, st) in tier_stats.iter() {
            res.insert(tier.clone(), st.clone());
@@ -1745,6 +1779,18 @@ impl TransitionState {
        res
    }

+    fn lock_last_day_stats(&self) -> std::sync::MutexGuard<'_, HashMap<String, LastDayTierStats>> {
+        match self.last_day_stats.lock() {
+            Ok(stats) => stats,
+            Err(poisoned) => {
+                let mut stats = poisoned.into_inner();
+                stats.clear();
+                self.last_day_stats.clear_poison();
+                stats
+            }
+        }
+    }
+
    pub async fn update_workers(api: Arc<ECStore>, n: i64) {
        Self::update_workers_inner(api, n).await;
    }
@@ -1766,7 +1812,27 @@ impl TransitionState {
    fn resize_workers_to(api: Arc<ECStore>, n: i64, requested: i64, absolute_max: i64) {
        let target = n as usize;
        let transition_state = runtime_sources::transition_state_handle();
-        let mut workers = transition_state.workers.lock().unwrap();
+        let runtime = match tokio::runtime::Handle::try_current() {
+            Ok(runtime) => runtime,
+            Err(err) => {
+                warn!(
+                    event = EVENT_LIFECYCLE_WORKER_STATE,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_LIFECYCLE,
+                    error = %err,
+                    state = "resize_failed",
+                    "Lifecycle worker pool requires a Tokio runtime"
+                );
+                return;
+            }
+        };
+        // Runtime lookup happens before locking, and the guard is dropped before
+        // metrics/logging callbacks. Poison therefore means a Vec mutation may
+        // have unwound and worker tracking cannot be reconstructed safely.
+        let mut workers = transition_state
+            .workers
+            .lock()
+            .expect("transition worker tracking mutex poisoned");
        let tracked_workers = workers.len();
        workers.retain(|worker| !worker.handle.is_finished());
        let pruned_finished_workers = tracked_workers.saturating_sub(workers.len());
@@ -1776,7 +1842,7 @@ impl TransitionState {
            let clone_api = api.clone();
            let cancel = CancellationToken::new();
            let worker_cancel = cancel.clone();
-            let handle = tokio::spawn(async move {
+            let handle = runtime.spawn(async move {
                TransitionState::worker_with_cancel(clone_api, worker_cancel).await;
            });
            workers.push(TransitionWorker { cancel, handle });
@@ -1790,6 +1856,7 @@ impl TransitionState {

        let current_workers = workers.len() as i64;
        transition_state.num_workers.store(current_workers, Ordering::SeqCst);
+        drop(workers);
        transition_state.record_scanner_transition_state();

        debug!(
@@ -4135,6 +4202,23 @@ pub async fn get_transitioned_object_reader(
    get_transitioned_object_reader_with_tier_manager(bucket, object, rs, h, oi, opts, &tier_config_mgr).await
 }

+fn validate_transition_remote_version(oi: &ObjectInfo) -> Result<bool, std::io::Error> {
+    let version = oi.transitioned_object.version_id.as_str();
+    match oi.transition_version_state {
+        rustfs_filemeta::TransitionVersionState::Unknown => Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier object version state is unknown",
+        )),
+        rustfs_filemeta::TransitionVersionState::KnownDisabled if version.is_empty() => Ok(false),
+        rustfs_filemeta::TransitionVersionState::SuspendedNull if version == "null" => Ok(true),
+        rustfs_filemeta::TransitionVersionState::Exact if !version.is_empty() && version != "null" => Ok(true),
+        _ => Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier object version state conflicts with its version ID",
+        )),
+    }
+}
+
 pub(crate) async fn get_transitioned_object_reader_with_tier_manager(
    bucket: &str,
    object: &str,
@@ -4144,6 +4228,7 @@ pub(crate) async fn get_transitioned_object_reader_with_tier_manager(
    opts: &ObjectOptions,
    tier_config_mgr: &Arc<RwLock<TierConfigMgr>>,
 ) -> Result<GetObjectReader, std::io::Error> {
+    validate_transition_remote_version(oi)?;
    let expected_identity = tier_destination_id_from_metadata(&oi.user_defined)?;
    let lease = match expected_identity {
        Some(identity) => {
@@ -4884,6 +4969,7 @@ mod tests {
        FreeVersionRecoveryStats, RecoveryWalkTestAction, list_tier_free_versions, recover_tier_free_versions_with_cancel,
        set_recovery_bucket_list_wait_hook, set_recovery_walk_test_hook,
    };
+    use crate::bucket::lifecycle::tier_last_day_stats::LastDayTierStats;
    use crate::bucket::lifecycle::tier_sweeper::Jentry;
    use crate::bucket::metadata::BUCKET_LIFECYCLE_CONFIG;
    use crate::bucket::metadata_sys;
@@ -4917,6 +5003,7 @@ mod tests {
    use http::HeaderMap;
    use rustfs_common::metrics::{IlmAction, global_metrics};
    use rustfs_config::ENV_TRANSITION_WORKERS_ABSOLUTE_MAX;
+    use rustfs_data_usage::TierStats;
    use rustfs_filemeta::{FileInfo, FileMeta};
    use s3s::dto::{
        BucketLifecycleConfiguration, ExpirationStatus, LifecycleExpiration, LifecycleRule, MetadataEntry, OutputLocation,
@@ -5438,6 +5525,7 @@ mod tests {
                tier: tier.clone(),
                ..Default::default()
            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Exact,
            ..Default::default()
        };

@@ -5501,6 +5589,7 @@ mod tests {
                tier,
                ..Default::default()
            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Exact,
            ..Default::default()
        };

@@ -5523,6 +5612,70 @@ mod tests {
        assert_eq!(backend.get_count().await, 0);
    }

+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn transitioned_get_rejects_unknown_version_state_before_backend_io() {
+        let manager = TierConfigMgr::new();
+        let tier = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let backend = register_mock_tier(&manager, &tier).await;
+        let object_info = ObjectInfo {
+            bucket: "bucket".to_string(),
+            name: "object".to_string(),
+            size: 1,
+            transitioned_object: TransitionedObject {
+                name: "remote/object".to_string(),
+                version_id: String::new(),
+                status: crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE.to_string(),
+                tier,
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            ..Default::default()
+        };
+
+        let err = match get_transitioned_object_reader_with_tier_manager(
+            &object_info.bucket,
+            &object_info.name,
+            &None,
+            &HeaderMap::new(),
+            &object_info,
+            &ObjectOptions::default(),
+            &manager,
+        )
+        .await
+        {
+            Ok(_) => panic!("unknown remote version state must fail before backend IO"),
+            Err(err) => err,
+        };
+
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.get_count().await, 0);
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn free_version_delete_rejects_unknown_version_state_before_backend_io() {
+        let manager = TierConfigMgr::new();
+        let backend = register_mock_tier(&manager, "WARM").await;
+        let object_info = ObjectInfo {
+            transitioned_object: TransitionedObject {
+                name: "remote/object".to_string(),
+                version_id: "legacy-version".to_string(),
+                tier: "WARM".to_string(),
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            ..Default::default()
+        };
+
+        let err = super::delete_free_version_remote_object(&object_info, &manager)
+            .await
+            .expect_err("unknown remote version state must fail before backend IO");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.remove_count().await, 0);
+    }
+
    #[cfg(feature = "test-util")]
    #[tokio::test]
    async fn free_version_remote_delete_requires_persisted_destination_identity() {
@@ -5572,6 +5725,7 @@ mod tests {
        oi.transitioned_object.tier = "WARM".to_string();
        oi.transitioned_object.name = "remote/object".to_string();
        oi.transitioned_object.version_id = "remote-version".to_string();
+        oi.transition_version_state = rustfs_filemeta::TransitionVersionState::Exact;
        let local_delete_calls = Arc::new(std::sync::atomic::AtomicUsize::new(0));

        let legacy_err = delete_free_version_remote_object_then(&oi, &manager, {
@@ -5582,7 +5736,8 @@ mod tests {
        })
        .await
        .expect_err("legacy free-version without identity must be retained");
-        assert!(legacy_err.to_string().contains("no durable backend identity"));
+        assert_eq!(legacy_err.kind(), std::io::ErrorKind::Other);
+        assert_eq!(old_backend.remove_count().await, 0);
        assert_eq!(local_delete_calls.load(Ordering::Relaxed), 0);

        let mut invalid_metadata = HashMap::new();
@@ -5713,6 +5868,7 @@ mod tests {
        oi.transitioned_object.tier = "WARM".to_string();
        oi.transitioned_object.name = "remote/object".to_string();
        oi.transitioned_object.version_id = "remote-version".to_string();
+        oi.transition_version_state = rustfs_filemeta::TransitionVersionState::Exact;

        let err = match get_transitioned_object_reader_with_tier_manager(
            "bucket",
@@ -5728,7 +5884,12 @@ mod tests {
            Ok(_) => panic!("identity-bound GET must reject a same-name tier rebind"),
            Err(err) => err,
        };
-        assert!(err.to_string().contains("identity no longer matches"));
+        assert_eq!(err.kind(), std::io::ErrorKind::Other);
+        let admin_err = err
+            .get_ref()
+            .and_then(|source| source.downcast_ref::<crate::client::admin_handler_utils::AdminError>())
+            .expect("identity mismatch should retain the typed tier error");
+        assert_eq!(admin_err.code, crate::services::tier::tier::ERR_TIER_INVALID_CONFIG.code);
        assert_eq!(new_backend.get_count().await, 0);

        oi.user_defined = Arc::new(HashMap::new());
@@ -5834,7 +5995,8 @@ mod tests {
            version_id: "remote-version".to_string(),
            tier_name: "WARM".to_string(),
            backend_identity: Some([1; 32]),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
        };

        let err = state
@@ -5945,7 +6107,8 @@ mod tests {
            version_id: "remote-version".to_string(),
            tier_name: "WARM".to_string(),
            backend_identity: Some([1; 32]),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
        };

        state
@@ -6459,6 +6622,96 @@ mod tests {
        assert_eq!(state.transition_rx.len(), 1);
    }

+    #[tokio::test]
+    #[serial]
+    async fn scanner_transition_enqueue_waits_for_saturated_queue_to_recover() {
+        let state = TransitionState::new_with_capacity(1);
+        let first_object = ObjectInfo {
+            bucket: "bucket".to_string(),
+            name: "first".to_string(),
+            ..Default::default()
+        };
+        let deferred_object = ObjectInfo {
+            bucket: "bucket".to_string(),
+            name: "deferred".to_string(),
+            ..Default::default()
+        };
+        let event = crate::bucket::lifecycle::lifecycle::Event {
+            action: IlmAction::TransitionAction,
+            ..Default::default()
+        };
+
+        assert!(
+            state.queue_transition_task(&first_object, &event, &LcEventSrc::Scanner).await,
+            "first scanner transition should fill the queue"
+        );
+
+        let deferred = state.queue_transition_task(&deferred_object, &event, &LcEventSrc::Scanner);
+        tokio::pin!(deferred);
+        assert!(
+            (&mut deferred).now_or_never().is_none(),
+            "a saturated scanner queue should apply bounded backpressure instead of dropping the task"
+        );
+
+        let first_task = state
+            .transition_rx
+            .recv()
+            .await
+            .expect("queue should remain open")
+            .expect("first queued transition task should be present");
+        state.release_transition(&first_task.obj_info);
+
+        assert!(
+            deferred.await,
+            "the deferred scanner transition should enqueue as soon as capacity recovers"
+        );
+        let recovered_task = state
+            .transition_rx
+            .recv()
+            .await
+            .expect("queue should remain open")
+            .expect("deferred transition task should be present");
+        assert_eq!(recovered_task.obj_info.name, deferred_object.name);
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn scanner_transition_sustained_saturation_schedules_compensation() {
+        let state = TransitionState::new_with_capacity_and_timeout(1, StdDuration::ZERO);
+        let first_object = ObjectInfo {
+            bucket: "saturated-bucket".to_string(),
+            name: "first".to_string(),
+            ..Default::default()
+        };
+        let missed_object = ObjectInfo {
+            bucket: "saturated-bucket".to_string(),
+            name: "missed".to_string(),
+            ..Default::default()
+        };
+        let event = crate::bucket::lifecycle::lifecycle::Event {
+            action: IlmAction::TransitionAction,
+            ..Default::default()
+        };
+
+        assert!(
+            state.queue_transition_task(&first_object, &event, &LcEventSrc::Scanner).await,
+            "first scanner transition should fill the queue"
+        );
+        assert!(
+            !state
+                .queue_transition_task(&missed_object, &event, &LcEventSrc::Scanner)
+                .await,
+            "a continuously saturated queue should report that the object was not admitted"
+        );
+        assert_eq!(state.queue_full_tasks(), 1);
+        assert_eq!(state.queue_send_timeout_tasks(), 1);
+        assert_eq!(
+            state.compensation_scheduled_tasks(),
+            1,
+            "timed-out scanner work must schedule a bounded bucket backfill"
+        );
+    }
+
    #[tokio::test]
    #[serial]
    async fn scanner_transition_enqueue_updates_transition_status() {
@@ -6984,6 +7237,46 @@ mod tests {
        assert_eq!(state.compensation_pending_tasks(), 1);
    }

+    #[test]
+    fn poisoned_compensation_set_can_release_completed_bucket() {
+        let state = TransitionState::new_with_capacity(1);
+        state
+            .compensation_buckets
+            .lock()
+            .expect("fresh mutex should lock")
+            .insert("bucket-a".to_string());
+        let poison_target = Arc::clone(&state.compensation_buckets);
+        let _ = std::thread::spawn(move || {
+            let _guard = poison_target.lock().expect("fresh mutex should lock");
+            panic!("poison compensation set");
+        })
+        .join();
+
+        state.finish_bucket_compensation("bucket-a");
+        assert_eq!(state.compensation_pending_tasks(), 0);
+        assert!(
+            state.compensation_buckets.lock().is_ok(),
+            "validated compensation state must clear poison"
+        );
+    }
+
+    #[test]
+    fn poisoned_tier_stats_are_reset_before_reuse() {
+        let state = TransitionState::new_with_capacity(1);
+        let poison_target = Arc::clone(&state.last_day_stats);
+        let _ = std::thread::spawn(move || {
+            let mut stats = poison_target.lock().expect("fresh mutex should lock");
+            stats.insert("stale".to_string(), LastDayTierStats::default());
+            panic!("poison tier stats");
+        })
+        .join();
+
+        state.add_lastday_stats("fresh", TierStats::default());
+        let stats = state.get_daily_all_tier_stats();
+        assert!(!stats.contains_key("stale"), "possibly partial statistics must be discarded");
+        assert!(stats.contains_key("fresh"), "statistics must accept new samples after recovery");
+    }
+
    #[tokio::test(flavor = "current_thread")]
    async fn scanner_transition_state_reports_compensation_pending_buckets() {
        let state = TransitionState::new_with_capacity(1);
@@ -7274,6 +7567,23 @@ mod tests {
        TransitionState::resize_workers_to(ecstore, original_workers, original_workers, absolute_max);
    }

+    #[tokio::test]
+    #[serial]
+    async fn transition_worker_resize_without_runtime_does_not_poison_tracking() {
+        let (_paths, ecstore) = setup_test_env().await;
+        let transition_state = runtime_sources::transition_state_handle();
+        let resize = std::thread::spawn(move || {
+            TransitionState::resize_workers_to(ecstore, 1, 1, resolve_transition_workers_absolute_max());
+        })
+        .join();
+
+        assert!(resize.is_ok(), "missing Tokio runtime must not panic while worker tracking is locked");
+        assert!(
+            transition_state.workers.lock().is_ok(),
+            "failed resize must leave worker tracking unpoisoned"
+        );
+    }
+
    #[test]
    fn should_defer_date_expiry_for_recent_config_update_respects_grace_window() {
        let now = OffsetDateTime::now_utc();
@@ -8664,7 +8974,7 @@ mod tests {
        .await
        .expect("first worker result should persist");
        assert_eq!(first.state, ManualTransitionJobState::Running);
-        assert_eq!(first.report.transition_completed, 1);
+        assert_eq!(first.report.transition_completed, 0);
        assert_eq!(first.report.transition_failed, 0);

        let duplicate = record_manual_transition_worker_result(
@@ -8677,11 +8987,11 @@ mod tests {
        .await
        .expect("duplicate worker result should be idempotent");
        assert_eq!(duplicate.state, ManualTransitionJobState::Running);
-        assert_eq!(duplicate.report.transition_completed, 1);
+        assert_eq!(duplicate.report.transition_completed, 0);
        assert_eq!(duplicate.report.transition_failed, 0);

        let second_key = manual_transition_worker_result_task_key(&bucket, "logs/b", None);
-        let final_record = record_manual_transition_worker_result(
+        let pending_record = record_manual_transition_worker_result(
            ecstore.clone(),
            job_id,
            &second_key,
@@ -8690,7 +9000,14 @@ mod tests {
        )
        .await
        .expect("second distinct worker result should persist");
+        assert_eq!(pending_record.state, ManualTransitionJobState::Running);
+        assert_eq!(pending_record.report.transition_completed, 0);
+        assert_eq!(pending_record.report.transition_failed, 0);

+        let final_record =
+            reconcile_manual_transition_worker_results(ecstore.clone(), job_id, ManualTransitionQueueSnapshot::default())
+                .await
+                .expect("worker result journal should reconcile");
        assert_eq!(final_record.state, ManualTransitionJobState::Partial);
        assert_eq!(final_record.report.transition_completed, 1);
        assert_eq!(final_record.report.transition_failed, 1);
@@ -8725,7 +9042,7 @@ mod tests {
            .expect("worker result job record should save");

        let task_key = manual_transition_worker_result_task_key(&bucket, "logs/fail", None);
-        let final_record = record_manual_transition_worker_result_with_reason(
+        let pending_record = record_manual_transition_worker_result_with_reason(
            ecstore.clone(),
            job_id,
            &task_key,
@@ -8735,7 +9052,12 @@ mod tests {
        )
        .await
        .expect("worker result with failure reason should persist");
+        assert!(pending_record.report.tier_failure_by_reason.is_empty());
+        assert_eq!(pending_record.report.transition_failed, 0);

+        let final_record = reconcile_manual_transition_worker_results(ecstore, job_id, ManualTransitionQueueSnapshot::default())
+            .await
+            .expect("worker failure reason should reconcile");
        assert_eq!(
            final_record
                .report
@@ -9866,6 +10188,87 @@ mod tests {
        (backend, identity_hex)
    }

+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn journal_replay_rejects_unknown_version_state_before_backend_io() {
+        let (_disk_paths, ecstore) = setup_test_env().await;
+        let (backend, _) = register_recovery_mock_tier(&ecstore).await;
+        let identity = TierConfigMgr::acquire_operation_lease(&ecstore.tier_config_mgr(), "WARM")
+            .await
+            .expect("mock tier lease should be available")
+            .backend_identity();
+        let je = Jentry {
+            obj_name: "remote/object".to_string(),
+            version_id: "legacy-version".to_string(),
+            tier_name: "WARM".to_string(),
+            backend_identity: Some(identity),
+            version_id_exact: false,
+            version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+        };
+
+        let err = crate::bucket::lifecycle::tier_delete_journal::process_tier_delete_journal_entry(ecstore, &je)
+            .await
+            .expect_err("unknown journal state must fail before backend IO");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.remove_count().await, 0);
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn journal_replay_deletes_confirmed_exact_provider_token() {
+        let (_disk_paths, ecstore) = setup_test_env().await;
+        let (backend, _) = register_recovery_mock_tier(&ecstore).await;
+        let lease = TierConfigMgr::acquire_operation_lease(&ecstore.tier_config_mgr(), "WARM")
+            .await
+            .expect("mock tier lease should be available");
+        let identity = lease.backend_identity();
+        backend
+            .set_put_remote_version(Some("provider-version-token".to_string()))
+            .await;
+        lease
+            .put(
+                "remote/object",
+                crate::client::transition_api::ReaderImpl::Body(bytes::Bytes::from_static(b"candidate")),
+                9,
+            )
+            .await
+            .expect("confirmed remote candidate should be seeded");
+        backend.set_remove_failure(true);
+        backend.set_reject_non_empty_remote_versions(true);
+        let je = Jentry {
+            obj_name: "remote/object".to_string(),
+            version_id: "provider-version-token".to_string(),
+            tier_name: "WARM".to_string(),
+            backend_identity: Some(identity),
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
+        };
+
+        crate::set_disk::cleanup_rejected_transition_upload_durably(
+            &lease,
+            &je.obj_name,
+            &je.version_id,
+            true,
+            Some(ecstore.clone()),
+        )
+        .await
+        .expect("failed immediate cleanup should remain durable in the journal");
+        assert!(backend.contains(&je.obj_name).await);
+
+        backend.set_remove_failure(false);
+        crate::bucket::lifecycle::tier_delete_journal::process_tier_delete_journal_entry(ecstore, &je)
+            .await
+            .expect("identity-bound exact journal must retry confirmed candidate cleanup");
+
+        assert!(!backend.contains(&je.obj_name).await);
+        assert_eq!(backend.exact_remove_count(), 2);
+        assert_eq!(
+            backend.remove_versions().await,
+            vec![("remote/object".to_string(), "provider-version-token".to_string())]
+        );
+    }
+
    async fn seed_recoverable_free_version(
        disk_paths: &[PathBuf],
        bucket: &str,
@@ -9885,6 +10288,7 @@ mod tests {
                identity,
            );
        }
+        let transition_version_id = Uuid::new_v4();
        let mut metadata = FileMeta::new();
        metadata
            .add_version(FileInfo {
@@ -9893,7 +10297,9 @@ mod tests {
                version_id: Some(object_version_id),
                transition_status: crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE.to_string(),
                transitioned_objname: format!("remote/{bucket}/{object}"),
-                transition_version_id: Some(Uuid::new_v4()),
+                transition_version_id: Some(transition_version_id),
+                transition_version: Some(transition_version_id.to_string()),
+                transition_version_state: rustfs_filemeta::TransitionVersionState::Exact,
                transition_tier: "WARM".to_string(),
                mod_time: Some(OffsetDateTime::now_utc()),
                metadata: transitioned_metadata,
@@ -10897,6 +11303,7 @@ mod tests {
    #[tokio::test(flavor = "multi_thread")]
    #[serial]
    async fn concurrent_resend_same_part_commits_one_generation() {
+        use crate::set_disk::{MultipartCommitBarrier, MultipartCommitPause};
        use crate::storage_api_contracts::object::ObjectIO as _;

        let (_paths, ecstore) = setup_test_env().await;
@@ -10911,58 +11318,44 @@ mod tests {

        // Distinct payloads with distinct sizes: a mixed-generation reassembly
        // would produce bytes matching none of them (or fail the read outright).
-        let candidates: Vec<Vec<u8>> = (0..3)
+        let candidates: Vec<Vec<u8>> = (0..2)
            .map(|g| {
                let len = 4096 + g * 512;
                vec![b'a' + g as u8; len]
            })
            .collect();

-        // Two independent causes can produce a spurious lock-acquire timeout
-        // here, and both must stay covered:
-        // 1. A lost/stolen fast-lock wakeup could strand a waiter until the
-        //    deadline — fixed for real in fast_lock::shard by bounding each
-        //    notification wait (NOTIFY_WAIT_CAP re-polling).
-        // 2. Under the full nextest suite on loaded CI disks, the
-        //    *legitimately serialized* cross-disk commits can exceed the
-        //    acquire deadline all by themselves — observed on CI at the 5s
-        //    default and the 30s production default with six resends, and
-        //    again at 60s, which is a hard ceiling: fast_lock clamps every
-        //    requested timeout to MAX_ACQUIRE_TIMEOUT (60s), so raising the
-        //    env override higher is a no-op (the Timeout error still reports
-        //    the requested value). Keep the guard about the correctness
-        //    property, not disk latency: request the full 60s ceiling and cap
-        //    the queue depth at three resends, so the last waiter sits behind
-        //    at most two serialized commits (~12s each on the slowest observed
-        //    CI runner, comfortably inside the deadline). Three concurrent
-        //    resends still race the streaming phase and contend on the commit
-        //    lock, which is all the generation-mixing regression needs.
-        //    `#[serial]` keeps the process-wide env override isolated.
-        let results = temp_env::async_with_vars([(rustfs_config::ENV_OBJECT_LOCK_ACQUIRE_TIMEOUT, Some("60"))], async {
-            let mut tasks = tokio::task::JoinSet::new();
-            for payload in candidates.iter().cloned() {
-                let store = ecstore.clone();
-                let bucket = bucket.clone();
-                let upload_id = upload.upload_id.clone();
-                tasks.spawn(async move {
-                    let mut data = PutObjReader::from_vec(payload.clone());
-                    store
-                        .put_object_part(&bucket, object, &upload_id, 1, &mut data, &ObjectOptions::default())
-                        .await
-                        .map(|info| (info, payload))
-                });
-            }
+        let commit_barrier = MultipartCommitBarrier::install_for_arrivals(
+            &bucket,
+            object,
+            MultipartCommitPause::PutPartBeforeLockAcquire,
+            candidates.len(),
+        );
+        let mut tasks = tokio::task::JoinSet::new();
+        for payload in candidates.iter().cloned() {
+            let store = ecstore.clone();
+            let bucket = bucket.clone();
+            let upload_id = upload.upload_id.clone();
+            tasks.spawn(async move {
+                let mut data = PutObjReader::from_vec(payload.clone());
+                store
+                    .put_object_part(&bucket, object, &upload_id, 1, &mut data, &ObjectOptions::default())
+                    .await
+                    .map(|info| (info, payload))
+            });
+        }

-            // Every concurrent resend must succeed; the commit lock must never
-            // starve a waiter into a timeout.
-            let mut results = Vec::new();
-            while let Some(joined) = tasks.join_next().await {
-                let outcome = joined.expect("put_object_part task should not panic");
-                results.push(outcome.expect("every concurrent same-part resend must succeed without lock timeout"));
-            }
-            results
-        })
-        .await;
+        // Both writers finish streaming before racing for the uploadId commit
+        // lock. Two generations are sufficient to exercise the mixed-shard
+        // hazard, while each waiter sits behind at most one cross-disk rename.
+        commit_barrier.wait_until_paused().await;
+        commit_barrier.release();
+
+        let mut results = Vec::new();
+        while let Some(joined) = tasks.join_next().await {
+            let outcome = joined.expect("put_object_part task should not panic");
+            results.push(outcome.expect("every concurrent same-part resend must succeed without lock timeout"));
+        }
        assert_eq!(results.len(), candidates.len());

        // Exactly one generation is visible after the serialized commits, and its
@@ -1646,40 +1646,14 @@ pub async fn record_manual_transition_worker_result_with_reason(
    job_id: Uuid,
    task_key: &str,
    result: ManualTransitionWorkerResult,
-    queue_snapshot: ManualTransitionQueueSnapshot,
+    _queue_snapshot: ManualTransitionQueueSnapshot,
    failure_reason: Option<ManualTransitionWorkerFailureReason>,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
    let result_record = ManualTransitionWorkerResultRecord::new_with_reason(job_id, task_key, result, failure_reason);
    if !save_manual_transition_worker_result_if_absent(api.clone(), &result_record).await? {
        return load_manual_transition_job_record(api, job_id).await;
    }
-
-    for _ in 0..4 {
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-        if record.is_terminal() {
-            return Ok(record);
-        }
-        record.record_worker_result_with_reason(result, queue_snapshot, failure_reason);
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
-            Ok(()) => {
-                if record.is_terminal() {
-                    delete_manual_transition_scope_admission_if_current(
-                        api.clone(),
-                        &record.scope_key,
-                        record.job_id,
-                        record.lease_id,
-                    )
-                    .await?;
-                } else {
-                    renew_manual_transition_scope_admission_from_job(api, &record).await?;
-                }
-                return Ok(record);
-            }
-            Err(Error::PreconditionFailed) => continue,
-            Err(err) => return Err(err),
-        }
-    }
-    Err(Error::PreconditionFailed)
+    load_manual_transition_job_record(api, job_id).await
 }

 pub async fn renew_manual_transition_job_lease(
@@ -20,7 +20,10 @@ use tokio_util::sync::CancellationToken;
 use tracing::{debug, warn};

 use crate::bucket::lifecycle::config_boundary;
-use crate::bucket::lifecycle::tier_sweeper::{Jentry, delete_object_from_remote_tier_idempotent_with_manager_and_identity};
+use crate::bucket::lifecycle::tier_sweeper::{
+    Jentry, delete_confirmed_transition_candidate_exact_with_manager_and_identity,
+    delete_object_from_remote_tier_idempotent_with_manager_and_identity,
+};
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::{Error, Result};
 use crate::object_api::{GetObjectReader, ObjectInfo, ObjectOptions, PutObjReader};
@@ -42,6 +45,7 @@ const TIER_DELETE_JOURNAL_RECOVERY_INTERVAL: Duration = Duration::from_secs(60);
 const TIER_DELETE_JOURNAL_RECOVERY_TIMEOUT: Duration = Duration::from_secs(300);
 const TIER_DELETE_JOURNAL_VERSION: u8 = 2;
 const TIER_DELETE_JOURNAL_EXACT_VERSION: u8 = 3;
+const TIER_DELETE_JOURNAL_STATE_VERSION: u8 = 4;
 pub(crate) const TIER_DELETE_JOURNAL_PREFIX: &str = "ilm/tier-delete-journal/";

 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
@@ -55,24 +59,35 @@ struct PersistedTierDeleteJournalEntry {
    backend_identity: Option<[u8; 32]>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    version_id_exact: Option<bool>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    version_state: Option<rustfs_filemeta::TransitionVersionState>,
 }

 impl PersistedTierDeleteJournalEntry {
-    fn from_jentry(je: &Jentry) -> Self {
-        Self {
-            version: if je.version_id_exact {
-                TIER_DELETE_JOURNAL_EXACT_VERSION
-            } else if je.backend_identity.is_some() {
+    fn from_jentry(je: &Jentry) -> Result<Self> {
+        validate_version_state(je.version_state, &je.version_id, je.version_id_exact)?;
+        let legacy_unknown = je.version_state == rustfs_filemeta::TransitionVersionState::Unknown;
+        let version = if legacy_unknown {
+            if je.backend_identity.is_some() {
                TIER_DELETE_JOURNAL_VERSION
            } else {
                1
-            },
+            }
+        } else {
+            if je.backend_identity.is_none() {
+                return Err(Error::other("new tier delete journal entry is missing its backend identity"));
+            }
+            TIER_DELETE_JOURNAL_STATE_VERSION
+        };
+        Ok(Self {
+            version,
            obj_name: je.obj_name.clone(),
            version_id: je.version_id.clone(),
            tier_name: je.tier_name.clone(),
            backend_identity: je.backend_identity,
            version_id_exact: je.version_id_exact.then_some(true),
-        }
+            version_state: (!legacy_unknown).then_some(je.version_state),
+        })
    }

    fn into_jentry(self) -> Result<Jentry> {
@@ -84,19 +99,23 @@ impl PersistedTierDeleteJournalEntry {
        if self.obj_name.is_empty() || self.tier_name.is_empty() {
            return Err(Error::other("tier delete journal entry is incomplete"));
        }
-        if self.version != TIER_DELETE_JOURNAL_EXACT_VERSION && self.version_id_exact.unwrap_or(false) {
+        if self.version != TIER_DELETE_JOURNAL_EXACT_VERSION
+            && self.version != TIER_DELETE_JOURNAL_STATE_VERSION
+            && self.version_id_exact.unwrap_or(false)
+        {
            return Err(Error::other(
                "legacy tier delete journal entry has an unsupported exact version constraint",
            ));
        }
-        let (backend_identity, version_id_exact) = match self.version {
-            1 => (None, false),
+        let (backend_identity, version_id_exact, version_state) = match self.version {
+            1 => (None, false, rustfs_filemeta::TransitionVersionState::Unknown),
            TIER_DELETE_JOURNAL_VERSION => (
                Some(
                    self.backend_identity
                        .ok_or_else(|| Error::other("tier delete journal v2 entry is missing its backend identity"))?,
                ),
                false,
+                rustfs_filemeta::TransitionVersionState::Unknown,
            ),
            TIER_DELETE_JOURNAL_EXACT_VERSION => {
                if self.version_id.is_empty() || self.version_id_exact != Some(true) {
@@ -108,6 +127,22 @@ impl PersistedTierDeleteJournalEntry {
                            .ok_or_else(|| Error::other("tier delete journal v3 entry is missing its backend identity"))?,
                    ),
                    true,
+                    rustfs_filemeta::TransitionVersionState::Exact,
+                )
+            }
+            TIER_DELETE_JOURNAL_STATE_VERSION => {
+                let state = self
+                    .version_state
+                    .ok_or_else(|| Error::other("tier delete journal v4 entry is missing its version state"))?;
+                let exact = self.version_id_exact.unwrap_or(false);
+                validate_version_state(state, &self.version_id, exact)?;
+                (
+                    Some(
+                        self.backend_identity
+                            .ok_or_else(|| Error::other("tier delete journal v4 entry is missing its backend identity"))?,
+                    ),
+                    exact,
+                    state,
                )
            }
            version => return Err(Error::other(format!("unsupported tier delete journal version {version}"))),
@@ -118,10 +153,30 @@ impl PersistedTierDeleteJournalEntry {
            tier_name: self.tier_name,
            backend_identity,
            version_id_exact,
+            version_state,
        })
    }
 }

+fn validate_version_state(
+    state: rustfs_filemeta::TransitionVersionState,
+    version_id: &str,
+    version_id_exact: bool,
+) -> Result<()> {
+    use rustfs_filemeta::TransitionVersionState::{Exact, KnownDisabled, SuspendedNull, Unknown};
+
+    let valid = match state {
+        Unknown => !version_id_exact,
+        KnownDisabled => version_id.is_empty() && !version_id_exact,
+        SuspendedNull => version_id == "null" && version_id_exact,
+        Exact => !version_id.is_empty() && version_id != "null" && version_id_exact,
+    };
+    if !valid {
+        return Err(Error::other("tier delete journal version state conflicts with its version id"));
+    }
+    Ok(())
+}
+
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct TierDeleteJournalRecoveryStats {
    pub scanned: usize,
@@ -159,7 +214,7 @@ pub(crate) fn decode_tier_delete_journal_entry(data: &[u8]) -> Result<Jentry> {
 }

 pub(crate) fn encode_tier_delete_journal_entry(je: &Jentry) -> Result<Vec<u8>> {
-    serde_json::to_vec(&PersistedTierDeleteJournalEntry::from_jentry(je))
+    serde_json::to_vec(&PersistedTierDeleteJournalEntry::from_jentry(je)?)
        .map_err(|err| Error::other(format!("encode tier delete journal failed: {err}")))
 }

@@ -209,18 +264,35 @@ where
 }

 pub async fn process_tier_delete_journal_entry(api: Arc<ECStore>, je: &Jentry) -> std::io::Result<()> {
+    if je.version_state == rustfs_filemeta::TransitionVersionState::Unknown {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "tier delete journal remote version state is unknown",
+        ));
+    }
    let backend_identity = je
        .backend_identity
        .ok_or_else(|| std::io::Error::other("legacy tier delete journal has no durable backend identity"))?;
-    delete_object_from_remote_tier_idempotent_with_manager_and_identity(
-        &je.obj_name,
-        &je.version_id,
-        &je.tier_name,
-        backend_identity,
-        &api.tier_config_mgr(),
-        je.version_id_exact,
-    )
-    .await?;
+    if je.version_id_exact {
+        delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+            &je.obj_name,
+            &je.version_id,
+            &je.tier_name,
+            backend_identity,
+            &api.tier_config_mgr(),
+        )
+        .await?;
+    } else {
+        delete_object_from_remote_tier_idempotent_with_manager_and_identity(
+            &je.obj_name,
+            &je.version_id,
+            &je.tier_name,
+            backend_identity,
+            &api.tier_config_mgr(),
+            false,
+        )
+        .await?;
+    }
    remove_tier_delete_journal_entry(api, je).await
 }

@@ -406,8 +478,9 @@ where
 #[cfg(test)]
 mod tests {
    use super::{
-        TIER_DELETE_JOURNAL_EXACT_VERSION, await_tier_delete_journal_recovery, decode_tier_delete_journal_entry,
-        encode_tier_delete_journal_entry, record_tier_delete_journal_backend_identity, tier_delete_journal_object_name,
+        TIER_DELETE_JOURNAL_EXACT_VERSION, TIER_DELETE_JOURNAL_STATE_VERSION, await_tier_delete_journal_recovery,
+        decode_tier_delete_journal_entry, encode_tier_delete_journal_entry, record_tier_delete_journal_backend_identity,
+        tier_delete_journal_object_name,
    };
    use crate::bucket::lifecycle::tier_sweeper::Jentry;
    use crate::error::Result;
@@ -420,7 +493,8 @@ mod tests {
            version_id: "remote-version".to_string(),
            tier_name: "WARM".to_string(),
            backend_identity: Some([7; 32]),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
        }
    }

@@ -436,6 +510,7 @@ mod tests {
        assert_eq!(decoded.tier_name, je.tier_name);
        assert_eq!(decoded.backend_identity, je.backend_identity);
        assert_eq!(decoded.version_id_exact, je.version_id_exact);
+        assert_eq!(decoded.version_state, je.version_state);
    }

    #[test]
@@ -450,7 +525,7 @@ mod tests {
        let persisted: serde_json::Value = serde_json::from_slice(&encoded).expect("exact journal JSON should decode");
        let decoded = decode_tier_delete_journal_entry(&encoded).expect("exact journal entry should decode");

-        assert_eq!(persisted["version"], TIER_DELETE_JOURNAL_EXACT_VERSION);
+        assert_eq!(persisted["version"], TIER_DELETE_JOURNAL_STATE_VERSION);
        assert_eq!(persisted["version_id_exact"], true);
        assert!(decoded.version_id_exact);
        assert_ne!(tier_delete_journal_object_name(&exact), tier_delete_journal_object_name(&normalized));
@@ -513,6 +588,46 @@ mod tests {
        }
    }

+    #[test]
+    fn tier_delete_journal_rejects_conflicting_v4_version_states() {
+        let identity = vec![7_u8; 32];
+        let invalid = [
+            ("known-disabled", "unexpected", false),
+            ("suspended-null", "", true),
+            ("suspended-null", "null", false),
+            ("exact", "", true),
+            ("exact", "null", true),
+            ("exact", "version", false),
+            ("unknown", "version", true),
+        ];
+
+        for (state, version_id, exact) in invalid {
+            let persisted = serde_json::json!({
+                "version": TIER_DELETE_JOURNAL_STATE_VERSION,
+                "obj_name": "remote/object",
+                "version_id": version_id,
+                "tier_name": "WARM",
+                "backend_identity": identity,
+                "version_id_exact": exact.then_some(true),
+                "version_state": state,
+            });
+            let encoded = serde_json::to_vec(&persisted).expect("invalid journal fixture should encode");
+            decode_tier_delete_journal_entry(&encoded).expect_err("conflicting v4 version state must fail closed");
+        }
+    }
+
+    #[test]
+    fn legacy_journals_decode_with_unknown_version_state() {
+        let v1 = br#"{"version":1,"obj_name":"remote/object","version_id":"opaque","tier_name":"WARM"}"#;
+        let v2 = br#"{"version":2,"obj_name":"remote/object","version_id":"opaque","tier_name":"WARM","backend_identity":[7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]}"#;
+
+        for payload in [v1.as_slice(), v2.as_slice()] {
+            let decoded = decode_tier_delete_journal_entry(payload).expect("legacy journal should decode");
+            assert_eq!(decoded.version_state, rustfs_filemeta::TransitionVersionState::Unknown);
+            assert!(!decoded.version_id_exact);
+        }
+    }
+
    #[test]
    fn tier_delete_journal_path_is_stable_and_sanitized() {
        let je = journal_entry();
@@ -530,6 +645,8 @@ mod tests {
    fn tier_delete_journal_paths_separate_legacy_and_backend_identities() {
        let mut legacy = journal_entry();
        legacy.backend_identity = None;
+        legacy.version_id_exact = false;
+        legacy.version_state = rustfs_filemeta::TransitionVersionState::Unknown;
        let mut backend_a = journal_entry();
        backend_a.backend_identity = Some([1; 32]);
        let mut backend_b = journal_entry();
@@ -575,6 +692,8 @@ mod tests {
    fn tier_delete_journal_without_transition_identity_stays_legacy() {
        let mut je = journal_entry();
        je.backend_identity = None;
+        je.version_id_exact = false;
+        je.version_state = rustfs_filemeta::TransitionVersionState::Unknown;

        let encoded = encode_tier_delete_journal_entry(&je).expect("legacy journal should remain encodable");
        let persisted: serde_json::Value = serde_json::from_slice(&encoded).expect("journal JSON should decode");
@@ -185,6 +185,7 @@ struct ObjSweeper {
    transition_status: String,
    transition_tier: String,
    transition_version_id: String,
+    transition_version_state: rustfs_filemeta::TransitionVersionState,
    remote_object: String,
 }

@@ -231,7 +232,9 @@ impl ObjSweeper {
    }

    pub fn should_remove_remote_object(&self) -> Option<Jentry> {
-        if self.transition_status != lifecycle::TRANSITION_COMPLETE {
+        if self.transition_status != lifecycle::TRANSITION_COMPLETE
+            || self.transition_version_state == rustfs_filemeta::TransitionVersionState::Unknown
+        {
            return None;
        }

@@ -249,7 +252,11 @@ impl ObjSweeper {
                version_id: self.transition_version_id.clone(),
                tier_name: self.transition_tier.clone(),
                backend_identity: None,
-                version_id_exact: false,
+                version_id_exact: matches!(
+                    self.transition_version_state,
+                    rustfs_filemeta::TransitionVersionState::SuspendedNull | rustfs_filemeta::TransitionVersionState::Exact
+                ),
+                version_state: self.transition_version_state,
            });
        }
        None
@@ -286,6 +293,7 @@ pub struct Jentry {
    pub(crate) tier_name: String,
    pub(crate) backend_identity: Option<TierDestinationId>,
    pub(crate) version_id_exact: bool,
+    pub(crate) version_state: rustfs_filemeta::TransitionVersionState,
 }

 impl ExpiryOp for Jentry {
@@ -330,7 +338,7 @@ async fn delete_object_from_remote_tier_raw_with_manager(
    let lease = TierConfigMgr::acquire_operation_lease(&tier_config_mgr, tier_name)
        .await
        .map_err(std::io::Error::other)?;
-    delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, &lease, false).await
+    delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, &lease, false, true).await
 }

 async fn delete_object_from_remote_tier_raw_with_lease(
@@ -338,8 +346,11 @@ async fn delete_object_from_remote_tier_raw_with_lease(
    rv_id: &str,
    lease: &TierOperationLease,
    version_id_exact: bool,
+    validate_remote_version_id: bool,
 ) -> Result<(), std::io::Error> {
-    lease.validate_remote_version_id(rv_id)?;
+    if validate_remote_version_id {
+        lease.validate_remote_version_id(rv_id)?;
+    }

    if remote_delete_breaker_is_open(Instant::now()).await {
        metrics::counter!(METRIC_DELETE_REMOTE_BREAKER_TOTAL).increment(1);
@@ -435,7 +446,53 @@ pub(crate) async fn delete_object_from_remote_tier_with_lease_idempotent(
    lease: &TierOperationLease,
    version_id_exact: bool,
 ) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
-    match delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, lease, version_id_exact).await {
+    delete_object_from_remote_tier_with_lease_idempotent_inner(obj_name, rv_id, lease, version_id_exact, true).await
+}
+
+pub(crate) async fn delete_confirmed_transition_candidate_exact_with_lease_idempotent(
+    obj_name: &str,
+    rv_id: &str,
+    lease: &TierOperationLease,
+) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
+    if rv_id.is_empty() {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidInput,
+            "confirmed versioned transition candidate requires a non-empty remote version",
+        ));
+    }
+    #[cfg(test)]
+    if obj_name == "remote/empty-guard-probe" {
+        CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+    }
+    delete_object_from_remote_tier_with_lease_idempotent_inner(obj_name, rv_id, lease, true, false).await
+}
+
+#[cfg(test)]
+static CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES: std::sync::atomic::AtomicUsize = std::sync::atomic::AtomicUsize::new(0);
+
+pub(crate) async fn delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+    obj_name: &str,
+    rv_id: &str,
+    tier_name: &str,
+    backend_identity: TierDestinationId,
+    tier_config_mgr: &Arc<tokio::sync::RwLock<TierConfigMgr>>,
+) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
+    let lease = TierConfigMgr::acquire_operation_lease_for_backend_identity(tier_config_mgr, tier_name, backend_identity)
+        .await
+        .map_err(std::io::Error::other)?;
+    delete_confirmed_transition_candidate_exact_with_lease_idempotent(obj_name, rv_id, &lease).await
+}
+
+async fn delete_object_from_remote_tier_with_lease_idempotent_inner(
+    obj_name: &str,
+    rv_id: &str,
+    lease: &TierOperationLease,
+    version_id_exact: bool,
+    validate_remote_version_id: bool,
+) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
+    match delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, lease, version_id_exact, validate_remote_version_id)
+        .await
+    {
        Ok(()) => Ok(RemoteTierDeleteOutcome::Deleted),
        Err(err) if is_remote_tier_not_found_error(&err) => Ok(RemoteTierDeleteOutcome::AlreadyRemoved),
        Err(err) => {
@@ -460,6 +517,7 @@ pub fn transitioned_delete_journal_entry(
    versioned: bool,
    suspended: bool,
    transitioned: &TransitionedObject,
+    transition_version_state: rustfs_filemeta::TransitionVersionState,
 ) -> Option<Jentry> {
    let sweeper = ObjSweeper {
        version_id,
@@ -468,6 +526,7 @@ pub fn transitioned_delete_journal_entry(
        transition_status: transitioned.status.clone(),
        transition_tier: transitioned.tier.clone(),
        transition_version_id: transitioned.version_id.clone(),
+        transition_version_state,
        remote_object: transitioned.name.clone(),
        ..Default::default()
    };
@@ -475,8 +534,13 @@ pub fn transitioned_delete_journal_entry(
    sweeper.should_remove_remote_object()
 }

-pub fn transitioned_force_delete_journal_entry(transitioned: &TransitionedObject) -> Option<Jentry> {
-    if transitioned.status != lifecycle::TRANSITION_COMPLETE {
+pub fn transitioned_force_delete_journal_entry(
+    transitioned: &TransitionedObject,
+    transition_version_state: rustfs_filemeta::TransitionVersionState,
+) -> Option<Jentry> {
+    if transitioned.status != lifecycle::TRANSITION_COMPLETE
+        || transition_version_state == rustfs_filemeta::TransitionVersionState::Unknown
+    {
        return None;
    }

@@ -485,7 +549,11 @@ pub fn transitioned_force_delete_journal_entry(transitioned: &TransitionedObject
        version_id: transitioned.version_id.clone(),
        tier_name: transitioned.tier.clone(),
        backend_identity: None,
-        version_id_exact: false,
+        version_id_exact: matches!(
+            transition_version_state,
+            rustfs_filemeta::TransitionVersionState::SuspendedNull | rustfs_filemeta::TransitionVersionState::Exact
+        ),
+        version_state: transition_version_state,
    })
 }

@@ -494,11 +562,14 @@ mod test {
    use crate::client::signer_error::invalid_utf8_header_error;

    use super::{
-        ERR_REMOTE_DELETE_BREAKER_OPEN, ERR_REMOTE_DELETE_LIMITER_CLOSED, RemoteDeleteBreaker, RemoteTierDeleteOutcome,
+        CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES, ERR_REMOTE_DELETE_BREAKER_OPEN, ERR_REMOTE_DELETE_LIMITER_CLOSED,
+        RemoteDeleteBreaker, RemoteTierDeleteOutcome, delete_confirmed_transition_candidate_exact_with_manager_and_identity,
        delete_object_from_remote_tier_idempotent, delete_object_from_remote_tier_idempotent_with_manager_and_identity,
-        is_remote_tier_not_found_error, is_signer_header_error, set_remote_tier_delete_test_hook,
-        should_record_remote_delete_failure,
+        is_remote_tier_not_found_error, is_signer_header_error, lifecycle, set_remote_tier_delete_test_hook,
+        should_record_remote_delete_failure, transitioned_delete_journal_entry, transitioned_force_delete_journal_entry,
    };
+    use crate::storage_api_contracts::lifecycle::TransitionedObject;
+    use rustfs_filemeta::TransitionVersionState;
    use std::io::{Error, ErrorKind};
    use std::time::{Duration, Instant};

@@ -542,6 +613,43 @@ mod test {
        assert!(should_record_remote_delete_failure(&Error::other("NoSuchVersion")));
    }

+    #[test]
+    fn transitioned_delete_journal_preserves_remote_version_state() {
+        let cases = [
+            (TransitionVersionState::Unknown, "legacy-version", None),
+            (TransitionVersionState::KnownDisabled, "", Some(false)),
+            (TransitionVersionState::SuspendedNull, "null", Some(true)),
+            (TransitionVersionState::Exact, "opaque-version", Some(true)),
+        ];
+
+        for (state, version_id, expected_exact) in cases {
+            let transitioned = TransitionedObject {
+                name: "remote/object".to_string(),
+                version_id: version_id.to_string(),
+                tier: "WARM".to_string(),
+                status: lifecycle::TRANSITION_COMPLETE.to_string(),
+                ..Default::default()
+            };
+            let regular = transitioned_delete_journal_entry(None, false, false, &transitioned, state);
+            let forced = transitioned_force_delete_journal_entry(&transitioned, state);
+
+            match expected_exact {
+                Some(expected_exact) => {
+                    let regular = regular.expect("known version state should produce a regular delete journal entry");
+                    assert_eq!(regular.version_state, state);
+                    assert_eq!(regular.version_id_exact, expected_exact);
+                    let forced = forced.expect("known version state should produce a forced delete journal entry");
+                    assert_eq!(forced.version_state, state);
+                    assert_eq!(forced.version_id_exact, expected_exact);
+                }
+                None => {
+                    assert!(regular.is_none());
+                    assert!(forced.is_none());
+                }
+            }
+        }
+    }
+
    #[tokio::test]
    #[serial_test::serial]
    async fn idempotent_remote_delete_treats_hooked_nosuchversion_as_already_removed() {
@@ -664,6 +772,55 @@ mod test {
        assert_eq!(backend.remove_versions().await, vec![("remote/object".to_string(), String::new())]);
    }

+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn confirmed_transition_cleanup_deletes_exact_provider_token() {
+        CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES.store(0, std::sync::atomic::Ordering::Relaxed);
+        let manager = crate::services::tier::tier::TierConfigMgr::new();
+        let backend = crate::services::tier::test_util::register_mock_tier(&manager, "WARM").await;
+        let lease = crate::services::tier::tier::TierConfigMgr::acquire_operation_lease(&manager, "WARM")
+            .await
+            .expect("test tier lease should be available");
+        let identity = lease.backend_identity();
+        drop(lease);
+        backend.set_reject_non_empty_remote_versions(true);
+
+        let outcome = delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+            "remote/object",
+            "provider-version-token",
+            "WARM",
+            identity,
+            &manager,
+        )
+        .await
+        .expect("confirmed upload compensation should delete the exact provider token");
+
+        assert_eq!(outcome, RemoteTierDeleteOutcome::Deleted);
+        assert_eq!(backend.exact_remove_count(), 1);
+        assert_eq!(
+            backend.remove_versions().await,
+            vec![("remote/object".to_string(), "provider-version-token".to_string())]
+        );
+
+        let err = delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+            "remote/empty-guard-probe",
+            "",
+            "WARM",
+            identity,
+            &manager,
+        )
+        .await
+        .expect_err("confirmed versioned cleanup must reject an empty token");
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+        assert_eq!(backend.remove_count().await, 1);
+        assert_eq!(
+            CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES.load(std::sync::atomic::Ordering::Relaxed),
+            0,
+            "empty remote versions must be rejected before exact cleanup dispatch"
+        );
+    }
+
    #[test]
    fn breaker_opens_at_threshold_and_recovers_after_window() {
        let mut breaker = RemoteDeleteBreaker::new(3, Duration::from_secs(30));
@@ -22,7 +22,11 @@ use uuid::Uuid;

 use crate::bucket::lifecycle::config_boundary;
 use crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE;
-use crate::bucket::lifecycle::tier_sweeper::delete_object_from_remote_tier_idempotent_with_manager_and_identity;
+use crate::bucket::lifecycle::tier_sweeper::{
+    delete_confirmed_transition_candidate_exact_with_lease_idempotent,
+    delete_confirmed_transition_candidate_exact_with_manager_and_identity,
+    delete_object_from_remote_tier_idempotent_with_manager_and_identity,
+};
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::{Error, Result as EcstoreResult};
 use crate::object_api::ObjectOptions;
@@ -613,6 +617,199 @@ pub enum TransitionTransactionRecoveryOutcome {
    Retained,
 }

+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "snake_case")]
+pub enum TransitionOperatorProbe {
+    Missing,
+    UnversionedPresent,
+    VersionedPresent(String),
+    Ambiguous,
+    Unsupported,
+}
+
+impl From<TransitionCandidateProbe> for TransitionOperatorProbe {
+    fn from(value: TransitionCandidateProbe) -> Self {
+        match value {
+            TransitionCandidateProbe::Missing => Self::Missing,
+            TransitionCandidateProbe::UnversionedPresent => Self::UnversionedPresent,
+            TransitionCandidateProbe::VersionedPresent(version_id) => Self::VersionedPresent(version_id),
+            TransitionCandidateProbe::Ambiguous => Self::Ambiguous,
+            TransitionCandidateProbe::Unsupported => Self::Unsupported,
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+pub struct TransitionOperatorStatus {
+    pub transaction_id: Uuid,
+    pub state: TransitionTransactionState,
+    pub tier_name: String,
+    pub remote_object: String,
+    pub not_after_unix_nanos: i64,
+    pub probe: TransitionOperatorProbe,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+pub struct TransitionOperatorDeleteResult {
+    pub status: TransitionOperatorStatus,
+    pub journal_observed_after_delete: bool,
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum TransitionOperatorError {
+    #[error("transition transaction was not found")]
+    NotFound,
+    #[error("transition transaction is still inside its active ownership window")]
+    NotExpired,
+    #[error("transition transaction state is not eligible for operator reconciliation: {0:?}")]
+    InvalidState(TransitionTransactionState),
+    #[error("an exact non-empty remote version is required")]
+    RemoteVersionRequired,
+    #[error("remote candidate is not proven missing: {0:?}")]
+    CandidateNotMissing(TransitionOperatorProbe),
+    #[error("remote candidate version does not match requested exact version: expected {expected}, observed {actual:?}")]
+    CandidateVersionMismatch {
+        expected: String,
+        actual: TransitionOperatorProbe,
+    },
+    #[error("transition transaction store failed: {0}")]
+    Store(#[source] Error),
+    #[error("remote tier reconciliation failed: {0}")]
+    Remote(#[source] std::io::Error),
+}
+
+type TransitionOperatorResult<T> = std::result::Result<T, TransitionOperatorError>;
+
+fn validate_operator_reconcile_transaction(
+    transaction: &TransitionTransaction,
+    now_unix_nanos: i128,
+) -> TransitionOperatorResult<()> {
+    transaction
+        .validate()
+        .map_err(|err| TransitionOperatorError::Store(Error::other(err)))?;
+    if transaction.state != TransitionTransactionState::UploadOutcomeUnknown {
+        return Err(TransitionOperatorError::InvalidState(transaction.state));
+    }
+    if now_unix_nanos < i128::from(transaction.not_after_unix_nanos) {
+        return Err(TransitionOperatorError::NotExpired);
+    }
+    Ok(())
+}
+
+async fn load_operator_reconcile_transaction(
+    api: Arc<ECStore>,
+    transaction_id: Uuid,
+) -> TransitionOperatorResult<TransitionTransaction> {
+    match load_transition_transaction_record(api, transaction_id).await {
+        Ok(transaction) => Ok(transaction),
+        Err(Error::ConfigNotFound) => Err(TransitionOperatorError::NotFound),
+        Err(err) => Err(TransitionOperatorError::Store(err)),
+    }
+}
+
+async fn operator_probe_transition_candidate(
+    api: Arc<ECStore>,
+    transaction: &TransitionTransaction,
+) -> TransitionOperatorResult<TransitionOperatorProbe> {
+    let lease = TierConfigMgr::acquire_operation_lease_for_backend_identity(
+        &api.tier_config_mgr(),
+        &transaction.tier_name,
+        transaction.backend_fingerprint,
+    )
+    .await
+    .map_err(|err| TransitionOperatorError::Remote(std::io::Error::other(err)))?;
+    lease
+        .probe_transition_candidate_for(&transaction.remote_object, transaction.transaction_id)
+        .await
+        .map(TransitionOperatorProbe::from)
+        .map_err(TransitionOperatorError::Remote)
+}
+
+pub async fn inspect_transition_transaction_for_operator(
+    api: Arc<ECStore>,
+    transaction_id: Uuid,
+) -> TransitionOperatorResult<TransitionOperatorStatus> {
+    let transaction = load_operator_reconcile_transaction(api.clone(), transaction_id).await?;
+    validate_operator_reconcile_transaction(&transaction, time::OffsetDateTime::now_utc().unix_timestamp_nanos())?;
+    let probe = operator_probe_transition_candidate(api, &transaction).await?;
+    Ok(TransitionOperatorStatus {
+        transaction_id,
+        state: transaction.state,
+        tier_name: transaction.tier_name,
+        remote_object: transaction.remote_object,
+        not_after_unix_nanos: transaction.not_after_unix_nanos,
+        probe,
+    })
+}
+
+pub async fn delete_transition_candidate_for_operator(
+    api: Arc<ECStore>,
+    transaction_id: Uuid,
+    remote_version_id: &str,
+) -> TransitionOperatorResult<TransitionOperatorDeleteResult> {
+    if remote_version_id.is_empty() {
+        return Err(TransitionOperatorError::RemoteVersionRequired);
+    }
+    let transaction = load_operator_reconcile_transaction(api.clone(), transaction_id).await?;
+    validate_operator_reconcile_transaction(&transaction, time::OffsetDateTime::now_utc().unix_timestamp_nanos())?;
+    let lease = TierConfigMgr::acquire_operation_lease_for_backend_identity(
+        &api.tier_config_mgr(),
+        &transaction.tier_name,
+        transaction.backend_fingerprint,
+    )
+    .await
+    .map_err(|err| TransitionOperatorError::Remote(std::io::Error::other(err)))?;
+    lease
+        .validate_remote_version_id(remote_version_id)
+        .map_err(TransitionOperatorError::Remote)?;
+    let before_delete_probe = lease
+        .probe_transition_candidate_for(&transaction.remote_object, transaction.transaction_id)
+        .await
+        .map(TransitionOperatorProbe::from)
+        .map_err(TransitionOperatorError::Remote)?;
+    if !matches!(&before_delete_probe, TransitionOperatorProbe::VersionedPresent(version_id) if version_id == remote_version_id) {
+        return Err(TransitionOperatorError::CandidateVersionMismatch {
+            expected: remote_version_id.to_string(),
+            actual: before_delete_probe,
+        });
+    }
+    delete_confirmed_transition_candidate_exact_with_lease_idempotent(&transaction.remote_object, remote_version_id, &lease)
+        .await
+        .map_err(TransitionOperatorError::Remote)?;
+    let probe = operator_probe_transition_candidate(api.clone(), &transaction).await?;
+    let journal_observed_after_delete = match load_transition_transaction_record(api, transaction_id).await {
+        Ok(_) => true,
+        Err(Error::ConfigNotFound) => false,
+        Err(err) => return Err(TransitionOperatorError::Store(err)),
+    };
+    Ok(TransitionOperatorDeleteResult {
+        status: TransitionOperatorStatus {
+            transaction_id,
+            state: transaction.state,
+            tier_name: transaction.tier_name,
+            remote_object: transaction.remote_object,
+            not_after_unix_nanos: transaction.not_after_unix_nanos,
+            probe,
+        },
+        journal_observed_after_delete,
+    })
+}
+
+pub async fn finalize_missing_transition_transaction_for_operator(
+    api: Arc<ECStore>,
+    transaction_id: Uuid,
+) -> TransitionOperatorResult<()> {
+    let transaction = load_operator_reconcile_transaction(api.clone(), transaction_id).await?;
+    validate_operator_reconcile_transaction(&transaction, time::OffsetDateTime::now_utc().unix_timestamp_nanos())?;
+    let probe = operator_probe_transition_candidate(api.clone(), &transaction).await?;
+    if probe != TransitionOperatorProbe::Missing {
+        return Err(TransitionOperatorError::CandidateNotMissing(probe));
+    }
+    delete_transition_transaction_record(api, transaction_id)
+        .await
+        .map_err(TransitionOperatorError::Store)
+}
+
 pub(crate) fn decode_transition_transaction_record(object: &str, data: &[u8]) -> Result<TransitionTransaction> {
    let transaction_id = transition_transaction_id_from_record_object_name(object)?;
    TransitionTransaction::decode(transaction_id, data)
@@ -697,7 +894,7 @@ async fn recover_unknown_upload_outcome(
    .map_err(Error::other)?;

    match lease
-        .probe_transition_candidate(&transaction.remote_object)
+        .probe_transition_candidate_for(&transaction.remote_object, transaction.transaction_id)
        .await
        .map_err(Error::other)?
    {
@@ -708,6 +905,21 @@ async fn recover_unknown_upload_outcome(
        TransitionCandidateProbe::UnversionedPresent => {
            cleanup_recovered_unknown_upload_candidate(api, transaction, TransitionRemoteVersion::unversioned()).await
        }
+        TransitionCandidateProbe::VersionedPresent(version_id)
+            if Uuid::parse_str(&version_id).is_ok_and(|version_id| version_id.is_nil()) =>
+        {
+            delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+                &transaction.remote_object,
+                &version_id,
+                &transaction.tier_name,
+                transaction.backend_fingerprint,
+                &api.tier_config_mgr(),
+            )
+            .await
+            .map_err(Error::other)?;
+            delete_transition_transaction_record(api, transaction.transaction_id).await?;
+            Ok(TransitionTransactionRecoveryOutcome::RemoteCandidateDeleted)
+        }
        TransitionCandidateProbe::VersionedPresent(version_id) => {
            cleanup_recovered_unknown_upload_candidate(api, transaction, TransitionRemoteVersion::versioned(version_id)).await
        }
@@ -1079,6 +1291,27 @@ mod tests {
            .expect("upload state change should succeed")
    }

+    #[test]
+    fn operator_reconcile_requires_expired_unknown_upload_outcome() {
+        let mut transaction = new_transaction();
+        let active_deadline = transaction.not_after_unix_nanos;
+
+        assert!(matches!(
+            validate_operator_reconcile_transaction(&transaction, i128::from(active_deadline) + 1),
+            Err(TransitionOperatorError::InvalidState(TransitionTransactionState::UploadStarted))
+        ));
+
+        transaction
+            .advance(transaction.fence(), TransitionTransactionState::UploadOutcomeUnknown, None)
+            .expect("unknown upload outcome should be recorded");
+        assert!(matches!(
+            validate_operator_reconcile_transaction(&transaction, i128::from(active_deadline) - 1),
+            Err(TransitionOperatorError::NotExpired)
+        ));
+        validate_operator_reconcile_transaction(&transaction, i128::from(active_deadline))
+            .expect("expired unknown upload outcome should be eligible");
+    }
+
    fn cleanup_proof(transaction: &TransitionTransaction, decision: TransitionCleanupDecision) -> TransitionCleanupProof {
        TransitionCleanupProof {
            transaction_id: transaction.transaction_id,
@@ -737,6 +737,9 @@ impl BucketMetadata {
            }
            BUCKET_TAGGING_CONFIG => {
                self.tagging_config_xml = data;
+                // Drop the parsed form (like lifecycle above) so clearing the
+                // payload can't leave stale parsed tags to be cached.
+                self.tagging_config = None;
                self.tagging_config_updated_at = updated;
            }
            BUCKET_QUOTA_CONFIG_FILE => {
@@ -1318,6 +1321,30 @@ mod test {
        assert!(bm.lifecycle_config.is_none());
    }

+    /// Companion to the lifecycle case above. `parse_all_configs` skips empty
+    /// XML rather than clearing, so without the explicit reset a cleared
+    /// tagging config would keep serving the previously parsed tags.
+    #[test]
+    fn tagging_update_config_clears_parsed_config_on_delete() {
+        let mut bm = BucketMetadata::new("test-bucket");
+        let tagging_xml = br#"<Tagging><TagSet><Tag><Key>env</Key><Value>prod</Value></Tag></TagSet></Tagging>"#;
+
+        bm.update_config(BUCKET_TAGGING_CONFIG, tagging_xml.to_vec())
+            .expect("tagging config should update");
+        bm.parse_all_configs().expect("tagging config should parse");
+        assert!(bm.tagging_config.is_some());
+
+        bm.update_config(BUCKET_TAGGING_CONFIG, Vec::new())
+            .expect("tagging config delete should update metadata");
+
+        assert!(bm.tagging_config_xml.is_empty());
+        assert!(bm.tagging_config.is_none());
+
+        // A re-parse must not resurrect them either.
+        bm.parse_all_configs().expect("cleared tagging should parse");
+        assert!(bm.tagging_config.is_none());
+    }
+
    #[tokio::test]
    async fn marshal_msg_complete_example() {
        // Create a complete BucketMetadata with various configurations
@@ -15,23 +15,26 @@
 use super::metadata_sys::get_bucket_metadata_sys;
 use crate::error::{Result, StorageError};
 use rustfs_policy::policy::{BucketPolicy, BucketPolicyArgs};
-use tracing::info;

 pub struct PolicySys {}

 impl PolicySys {
    pub async fn is_allowed(args: &BucketPolicyArgs<'_>) -> bool {
-        match Self::get(args.bucket).await {
-            Ok(cfg) => return cfg.is_allowed(args).await,
-            Err(err) => {
-                if err != StorageError::ConfigNotFound {
-                    info!("config get err {:?}", err);
-                }
-            }
-        }
-
-        args.is_owner
+        matches!(Self::try_is_allowed(args).await, Ok(true))
    }
+
+    pub async fn try_is_allowed(args: &BucketPolicyArgs<'_>) -> Result<bool> {
+        Self::is_allowed_with_policy(args, Self::get(args.bucket).await).await
+    }
+
+    async fn is_allowed_with_policy(args: &BucketPolicyArgs<'_>, policy: Result<BucketPolicy>) -> Result<bool> {
+        match policy {
+            Ok(policy) => Ok(policy.is_allowed(args).await),
+            Err(StorageError::ConfigNotFound) => Ok(args.is_owner),
+            Err(err) => Err(err),
+        }
+    }
+
    pub async fn get(bucket: &str) -> Result<BucketPolicy> {
        let bucket_meta_sys_lock = get_bucket_metadata_sys()?;
        let bucket_meta_sys = bucket_meta_sys_lock.read().await;
@@ -41,3 +44,91 @@ impl PolicySys {
        Ok(cfg)
    }
 }
+
+#[cfg(test)]
+mod tests {
+    use super::{PolicySys, StorageError};
+    use rustfs_policy::policy::action::{Action, S3Action};
+    use rustfs_policy::policy::{BucketPolicy, BucketPolicyArgs};
+    use std::collections::HashMap;
+
+    fn args<'a>(
+        is_owner: bool,
+        groups: &'a Option<Vec<String>>,
+        conditions: &'a HashMap<String, Vec<String>>,
+    ) -> BucketPolicyArgs<'a> {
+        BucketPolicyArgs {
+            bucket: "bucket",
+            action: Action::S3Action(S3Action::GetObjectAction),
+            is_owner,
+            account: "account",
+            groups,
+            conditions,
+            object: "object",
+        }
+    }
+
+    #[tokio::test]
+    async fn missing_policy_preserves_owner_and_iam_fallback_semantics() {
+        let groups = None;
+        let conditions = HashMap::new();
+        assert!(
+            PolicySys::is_allowed_with_policy(&args(true, &groups, &conditions), Err(StorageError::ConfigNotFound),)
+                .await
+                .expect("missing policy should preserve owner access")
+        );
+        assert!(
+            !PolicySys::is_allowed_with_policy(&args(false, &groups, &conditions), Err(StorageError::ConfigNotFound),)
+                .await
+                .expect("missing policy should defer non-owner access to IAM")
+        );
+    }
+
+    #[tokio::test]
+    async fn policy_load_failures_propagate() {
+        let groups = None;
+        let conditions = HashMap::new();
+        for (failure, expected_message) in [
+            (StorageError::Io(std::io::Error::other("policy read failed")), "policy read failed"),
+            (
+                StorageError::other("bucket metadata sys not initialized for this instance"),
+                "bucket metadata sys not initialized for this instance",
+            ),
+        ] {
+            let result = PolicySys::is_allowed_with_policy(&args(true, &groups, &conditions), Err(failure)).await;
+
+            assert!(
+                matches!(result, Err(StorageError::Io(ref err)) if err.to_string().contains(expected_message)),
+                "policy I/O and uninitialized metadata failures must propagate instead of granting owner access"
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn explicit_bucket_deny_precedes_iam_allow() {
+        let groups = None;
+        let conditions = HashMap::new();
+        let policy: BucketPolicy = serde_json::from_str(
+            r#"{
+  "Version":"2012-10-17",
+  "Statement":[{
+    "Effect":"Deny",
+    "Principal":{"AWS":"*"},
+    "Action":["s3:GetObject"],
+    "Resource":["arn:aws:s3:::bucket/*"]
+  }]
+}"#,
+        )
+        .expect("deny policy should parse");
+
+        let bucket_allowed = PolicySys::is_allowed_with_policy(&args(true, &groups, &conditions), Ok(policy))
+            .await
+            .expect("loaded bucket policy should evaluate");
+        let iam_allowed = true;
+        let request_allowed = bucket_allowed && iam_allowed;
+
+        assert!(iam_allowed, "test precondition: IAM grants the action");
+        assert!(!bucket_allowed, "test precondition: bucket policy explicitly denies the action");
+        assert!(!request_allowed, "explicit bucket Deny must reject before IAM Allow fallback");
+    }
+}
@@ -417,7 +417,7 @@ impl TransitionClient {
            bucket: complete_multipart_upload_result.bucket,
            key: complete_multipart_upload_result.key,
            etag: trim_etag(&complete_multipart_upload_result.etag),
-            version_id: self.raw_version_id(&h)?.unwrap_or_default().to_string(),
+            version_id: self.legacy_remote_version_id(&h)?,
            location: complete_multipart_upload_result.location,
            expiration: exp_time,
            expiration_rule_id: rule_id,
@@ -22,7 +22,7 @@ use bytes::Bytes;
 use futures::future::join_all;
 use http::{HeaderMap, HeaderName, HeaderValue, StatusCode};
 use std::io::Error;
-use std::sync::RwLock;
+use std::sync::{Mutex, MutexGuard, RwLock};
 use std::{collections::HashMap, sync::Arc};
 use time::{OffsetDateTime, format_description};
 use tokio::io::AsyncReadExt;
@@ -46,6 +46,14 @@ use crate::client::utils::base64_encode;
 use rustfs_utils::path::trim_etag;
 use s3s::header::X_AMZ_EXPIRATION;

+fn lock_md5_hasher(
+    md5_hasher: &Mutex<Option<rustfs_utils::hash::HashAlgorithm>>,
+) -> Result<MutexGuard<'_, Option<rustfs_utils::hash::HashAlgorithm>>, std::io::Error> {
+    md5_hasher
+        .lock()
+        .map_err(|_| std::io::Error::other("MD5 hasher state is unavailable"))
+}
+
 /// Read exactly `want` bytes for a single multipart part, or fewer if the reader
 /// reaches EOF first. Advances the reader so the next call returns the following
 /// part. Replaces the previous per-part `read_all()`/`to_vec()`, which drained
@@ -177,7 +185,7 @@ impl TransitionClient {
            let length = buf.len();

            if opts.send_content_md5 {
-                let mut md5_hasher = self.md5_hasher.lock().unwrap();
+                let mut md5_hasher = lock_md5_hasher(&self.md5_hasher)?;
                let md5_hash = match md5_hasher.as_mut() {
                    Some(hasher) => hasher,
                    None => return Err(std::io::Error::other("MD5 hasher not initialized")),
@@ -370,7 +378,7 @@ impl TransitionClient {
                let mut md5_base64: String = "".to_string();

                if opts.send_content_md5 {
-                    let mut md5_hasher = clone_self.md5_hasher.lock().unwrap();
+                    let mut md5_hasher = lock_md5_hasher(&clone_self.md5_hasher)?;
                    let md5_hash = match md5_hasher.as_mut() {
                        Some(hasher) => hasher,
                        None => {
@@ -418,6 +426,9 @@ impl TransitionClient {
        }

        let results = join_all(futures).await;
+        for result in results {
+            result?;
+        }

        select! {
            err = err_rx.recv() => {
@@ -567,7 +578,7 @@ impl TransitionClient {
            key: object_name.to_string(),
            etag: trim_etag(h.get("ETag").and_then(|v| v.to_str().ok()).unwrap_or("")),

-            version_id: self.raw_version_id(h)?.unwrap_or_default().to_string(),
+            version_id: self.legacy_remote_version_id(h)?,
            size,
            expiration: exp_time,
            expiration_rule_id: rule_id,
@@ -620,10 +631,12 @@ fn collect_complete_parts(parts_info: &HashMap<i64, ObjectPart>, total_parts_cou

 #[cfg(test)]
 mod tests {
-    use super::{ObjectPart, ReaderImpl, collect_complete_parts, read_multipart_part};
+    use super::{ObjectPart, ReaderImpl, collect_complete_parts, lock_md5_hasher, read_multipart_part};
    use crate::object_api::GetObjectReader;
    use bytes::Bytes;
+    use rustfs_utils::hash::HashAlgorithm;
    use std::collections::HashMap;
+    use std::sync::{Arc, Mutex};

    // Drive a reader through the same per-part loop the multipart stream uses and
    // collect the size of every part. Regression for rustfs/rustfs#4811: the old
@@ -733,4 +746,18 @@ mod tests {
            "a gap in the parts map must be an error, not a panic"
        );
    }
+
+    #[test]
+    fn poisoned_md5_state_fails_closed() {
+        let hasher = Arc::new(Mutex::new(Some(HashAlgorithm::Md5)));
+        let poison_target = Arc::clone(&hasher);
+        let _ = std::thread::spawn(move || {
+            let _guard = poison_target.lock().expect("fresh mutex should lock");
+            panic!("poison MD5 state");
+        })
+        .join();
+
+        let error = lock_md5_hasher(&hasher).expect_err("poisoned hash state must not be reused");
+        assert_eq!(error.kind(), std::io::ErrorKind::Other);
+    }
 }
@@ -201,7 +201,7 @@ impl TransitionClient {
            object_name: object_name.to_string(),
            object_version_id: opts.version_id,
            delete_marker: resp.headers().get(X_AMZ_DELETE_MARKER).map_or(false, |v| v == "true"),
-            delete_marker_version_id: self.raw_version_id(resp.headers())?.unwrap_or_default().to_string(),
+            delete_marker_version_id: self.legacy_remote_version_id(resp.headers())?,
            ..Default::default()
        })
    }
@@ -46,11 +46,33 @@ impl RemoteVersion {
            Self::Unknown | Self::Disabled => None,
        }
    }
+
+    pub(crate) fn exact_request_id(&self) -> Result<Option<&str>, Error> {
+        match self {
+            Self::Unknown => Err(Error::new(
+                ErrorKind::InvalidData,
+                "remote object version is unknown; exact version routing is unsafe",
+            )),
+            Self::Disabled => Ok(None),
+            Self::SuspendedNull => Ok(Some("null")),
+            Self::Exact(version_id) => Ok(Some(version_id)),
+        }
+    }
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub(crate) enum ConditionalCreateCapability {
+    Unsupported,
+    IfNoneMatchStar,
+    GenerationMatchZero,
 }

 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 pub(crate) struct ProviderVersionCapabilities {
    raw_version_header: Option<&'static str>,
+    pub(crate) bucket_versioning_state: bool,
+    pub(crate) list_object_versions: bool,
+    pub(crate) conditional_create: ConditionalCreateCapability,
    pub(crate) exact_get_delete: bool,
 }

@@ -62,28 +84,59 @@ impl ProviderVersionCapabilities {
            || tier_type.eq_ignore_ascii_case("r2")
            || tier_type.eq_ignore_ascii_case("wasabi")
        {
+            let list_object_versions = tier_type.eq_ignore_ascii_case("s3")
+                || tier_type.eq_ignore_ascii_case("rustfs")
+                || tier_type.eq_ignore_ascii_case("minio")
+                || tier_type.eq_ignore_ascii_case("r2");
            Self {
                raw_version_header: Some(X_AMZ_VERSION_ID),
+                bucket_versioning_state: list_object_versions,
+                list_object_versions,
+                conditional_create: if tier_type.eq_ignore_ascii_case("s3") || tier_type.eq_ignore_ascii_case("r2") {
+                    ConditionalCreateCapability::IfNoneMatchStar
+                } else {
+                    ConditionalCreateCapability::Unsupported
+                },
                exact_get_delete: true,
            }
        } else if tier_type.eq_ignore_ascii_case("aliyun") {
            Self {
                raw_version_header: Some(X_OSS_VERSION_ID),
+                bucket_versioning_state: false,
+                list_object_versions: false,
+                conditional_create: ConditionalCreateCapability::Unsupported,
                exact_get_delete: true,
            }
        } else if tier_type.eq_ignore_ascii_case("tencent") {
            Self {
                raw_version_header: Some(X_COS_VERSION_ID),
+                bucket_versioning_state: false,
+                list_object_versions: false,
+                conditional_create: ConditionalCreateCapability::Unsupported,
                exact_get_delete: true,
            }
        } else if tier_type.eq_ignore_ascii_case("huaweicloud") {
            Self {
                raw_version_header: Some(X_OBS_VERSION_ID),
+                bucket_versioning_state: false,
+                list_object_versions: false,
+                conditional_create: ConditionalCreateCapability::Unsupported,
                exact_get_delete: true,
            }
+        } else if tier_type.eq_ignore_ascii_case("gcs") {
+            Self {
+                raw_version_header: None,
+                bucket_versioning_state: false,
+                list_object_versions: false,
+                conditional_create: ConditionalCreateCapability::GenerationMatchZero,
+                exact_get_delete: false,
+            }
        } else {
            Self {
                raw_version_header: None,
+                bucket_versioning_state: false,
+                list_object_versions: false,
+                conditional_create: ConditionalCreateCapability::Unsupported,
                exact_get_delete: false,
            }
        }
@@ -119,7 +172,7 @@ impl ProviderVersionCapabilities {
    }
 }

-fn validate_remote_version_id(version_id: &str) -> Result<(), Error> {
+pub(crate) fn validate_remote_version_id(version_id: &str) -> Result<(), Error> {
    if version_id.is_empty() {
        return Err(Error::new(
            ErrorKind::InvalidData,
@@ -143,7 +196,7 @@ fn validate_remote_version_id(version_id: &str) -> Result<(), Error> {

 #[cfg(test)]
 mod tests {
-    use super::{BucketVersioningState, ProviderVersionCapabilities, RemoteVersion};
+    use super::{BucketVersioningState, ConditionalCreateCapability, ProviderVersionCapabilities, RemoteVersion};
    use http::{HeaderMap, HeaderValue};

    #[test]
@@ -219,6 +272,71 @@ mod tests {
        );
    }

+    #[test]
+    fn provider_capability_matrix_is_conservative_and_provider_specific() {
+        for (tier_type, state, list, conditional_create, exact_get_delete) in [
+            ("s3", true, true, ConditionalCreateCapability::IfNoneMatchStar, true),
+            ("rustfs", true, true, ConditionalCreateCapability::Unsupported, true),
+            ("minio", true, true, ConditionalCreateCapability::Unsupported, true),
+            ("r2", true, true, ConditionalCreateCapability::IfNoneMatchStar, true),
+            ("wasabi", false, false, ConditionalCreateCapability::Unsupported, true),
+            ("aliyun", false, false, ConditionalCreateCapability::Unsupported, true),
+            ("tencent", false, false, ConditionalCreateCapability::Unsupported, true),
+            ("huaweicloud", false, false, ConditionalCreateCapability::Unsupported, true),
+            ("gcs", false, false, ConditionalCreateCapability::GenerationMatchZero, false),
+            ("azure", false, false, ConditionalCreateCapability::Unsupported, false),
+            ("unsupported", false, false, ConditionalCreateCapability::Unsupported, false),
+        ] {
+            let capabilities = ProviderVersionCapabilities::for_tier_type(tier_type);
+            assert_eq!(capabilities.bucket_versioning_state, state, "{tier_type} versioning state");
+            assert_eq!(capabilities.list_object_versions, list, "{tier_type} version listing");
+            assert_eq!(capabilities.conditional_create, conditional_create, "{tier_type} conditional create");
+            assert_eq!(capabilities.exact_get_delete, exact_get_delete, "{tier_type} exact routing");
+        }
+    }
+
+    #[test]
+    fn remote_version_states_preserve_unknown_disabled_suspended_and_exact() {
+        let capabilities = ProviderVersionCapabilities::for_tier_type("s3");
+        let empty = HeaderMap::new();
+        let mut null = HeaderMap::new();
+        null.insert("x-amz-version-id", HeaderValue::from_static("null"));
+        let mut exact = HeaderMap::new();
+        exact.insert("x-amz-version-id", HeaderValue::from_static("opaque.generation-7"));
+
+        for (headers, state, expected) in [
+            (&empty, BucketVersioningState::Unknown, RemoteVersion::Unknown),
+            (&empty, BucketVersioningState::Disabled, RemoteVersion::Disabled),
+            (&empty, BucketVersioningState::Suspended, RemoteVersion::Unknown),
+            (&empty, BucketVersioningState::Enabled, RemoteVersion::Unknown),
+            (&null, BucketVersioningState::Suspended, RemoteVersion::SuspendedNull),
+            (
+                &exact,
+                BucketVersioningState::Enabled,
+                RemoteVersion::Exact("opaque.generation-7".to_string()),
+            ),
+        ] {
+            assert_eq!(
+                capabilities
+                    .remote_version(headers, state)
+                    .expect("version state should normalize"),
+                expected
+            );
+        }
+    }
+
+    #[test]
+    fn exact_request_routing_fails_closed_for_unknown_versions() {
+        for (version, expected) in [
+            (RemoteVersion::Disabled, None),
+            (RemoteVersion::SuspendedNull, Some("null")),
+            (RemoteVersion::Exact("opaque-v1".to_string()), Some("opaque-v1")),
+        ] {
+            assert_eq!(version.exact_request_id().expect("known version state"), expected);
+        }
+        assert!(RemoteVersion::Unknown.exact_request_id().is_err());
+    }
+
    #[test]
    fn provider_version_rejects_empty_or_oversized_headers() {
        let oversized = "v".repeat(1025);
@@ -31,7 +31,7 @@ use crate::client::{
    },
    constants::{UNSIGNED_PAYLOAD, UNSIGNED_PAYLOAD_TRAILER},
    credentials::{CredContext, Credentials, SignatureType, Static},
-    provider_versions::{BucketVersioningState, ProviderVersionCapabilities},
+    provider_versions::{BucketVersioningState, ProviderVersionCapabilities, RemoteVersion},
    signer_error,
 };
 use crate::{client::checksum::ChecksumMode, object_api::GetObjectReader};
@@ -332,6 +332,22 @@ impl TransitionClient {
        self.provider_version_capabilities().raw_version_id(headers)
    }

+    pub(crate) fn remote_version(
+        &self,
+        headers: &HeaderMap,
+        versioning: BucketVersioningState,
+    ) -> Result<RemoteVersion, std::io::Error> {
+        self.provider_version_capabilities().remote_version(headers, versioning)
+    }
+
+    pub(crate) fn legacy_remote_version_id(&self, headers: &HeaderMap) -> Result<String, std::io::Error> {
+        Ok(self
+            .remote_version(headers, BucketVersioningState::Unknown)?
+            .exact_id()
+            .unwrap_or_default()
+            .to_string())
+    }
+
    fn trace_errors_only_off(&self) {
        if let Ok(mut trace_errors_only) = self.trace_errors_only.lock() {
            *trace_errors_only = false;
@@ -1095,6 +1111,16 @@ impl Default for ObjectInfo {
    }
 }

+impl ObjectInfo {
+    pub(crate) fn remote_version(
+        &self,
+        capabilities: ProviderVersionCapabilities,
+        versioning: BucketVersioningState,
+    ) -> Result<RemoteVersion, std::io::Error> {
+        capabilities.remote_version(&self.metadata, versioning)
+    }
+}
+
 #[derive(Serialize, Deserialize, Debug, Clone)]
 pub struct RestoreInfo {
    ongoing_restore: bool,
@@ -1414,7 +1440,7 @@ mod tests {
        MAX_S3_CLIENT_RESPONSE_SIZE, MAX_S3_ERROR_RESPONSE_SIZE, SignatureType, build_tls_config, collect_response_body,
        signer_error_to_io_error, to_object_info_for_provider, validate_header_values, with_rustls_init_guard,
    };
-    use crate::client::provider_versions::ProviderVersionCapabilities;
+    use crate::client::provider_versions::{BucketVersioningState, ProviderVersionCapabilities, RemoteVersion};
    use http::{HeaderMap, HeaderValue};
    use http_body_util::Full;
    use hyper::body::Bytes;
@@ -1539,6 +1565,11 @@ mod tests {
                .expect("opaque provider version should parse");

        assert_eq!(info.version_id, None);
+        assert_eq!(
+            info.remote_version(ProviderVersionCapabilities::for_tier_type("tencent"), BucketVersioningState::Enabled,)
+                .expect("opaque response version should remain available"),
+            RemoteVersion::Exact("opaque.version_01".to_string())
+        );
        assert_eq!(
            info.metadata.get("x-cos-version-id").and_then(|value| value.to_str().ok()),
            Some("opaque.version_01")
@@ -12,11 +12,20 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

-use crate::cluster::rpc::http_auth::RPC_CONTENT_SHA256_HEADER;
-use crate::cluster::rpc::{gen_tonic_signature_headers, normalize_tonic_rpc_audience};
-use crate::disk::error::{DiskError, Error as DiskErrorType};
+#[cfg(test)]
+use crate::cluster::rpc::http_auth::RPC_REPLAY_SCOPE_VERSION_HEADER;
+use crate::cluster::rpc::http_auth::{
+    RPC_AUTH_VERSION_HEADER, RPC_AUTH_VERSION_V2, RPC_BOOT_EPOCH_CHALLENGE_HEADER, RPC_BOOT_EPOCH_HEADER,
+    RPC_BOOT_EPOCH_PROOF_HEADER, RPC_CONTENT_SHA256_HEADER, TIMESTAMP_HEADER,
+};
+use crate::cluster::rpc::{
+    gen_tonic_replay_scope_headers, gen_tonic_signature_headers, normalize_tonic_rpc_audience, verify_tonic_boot_epoch_response,
+};
+#[cfg(test)]
+use crate::cluster::rpc::{tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers};
+use crate::disk::error::{DiskError, Error as DiskErrorType, RpcStatusError};
 use crate::runtime::sources as runtime_sources;
-use http::Uri;
+use http::{Request as HttpRequest, Response as HttpResponse, Uri};
 use rustfs_protos::{
    ChannelClass, create_new_channel, get_channel_for_class,
    proto_gen::node_service::{
@@ -24,9 +33,19 @@ use rustfs_protos::{
        tier_mutation_control_service_client::TierMutationControlServiceClient,
    },
 };
-use std::{error::Error, io::ErrorKind};
+use std::{
+    collections::HashMap,
+    error::Error,
+    future::Future,
+    io::ErrorKind,
+    pin::Pin,
+    sync::{LazyLock, Mutex},
+    task::{Context, Poll},
+};
 use tonic::{service::interceptor::InterceptedService, transport::Channel};
+use tower::Service;
 use tracing::debug;
+use uuid::Uuid;

 use super::context_propagation::{inject_request_id_into_metadata, inject_trace_context_into_metadata};

@@ -35,7 +54,7 @@ use super::context_propagation::{inject_request_id_into_metadata, inject_trace_c
 pub async fn node_service_time_out_client(
    addr: &String,
    interceptor: TonicInterceptor,
-) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>, Box<dyn Error>> {
+) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>, Box<dyn Error>> {
    // Default to the latency-sensitive control channel; bulk `bytes` RPCs opt in via the
    // `_for_class` variant below (grpc-optimization P1).
    node_service_time_out_client_for_class(addr, interceptor, ChannelClass::Control).await
@@ -44,13 +63,14 @@ pub async fn node_service_time_out_client(
 pub async fn heal_control_time_out_client(
    addr: &str,
    interceptor: TonicInterceptor,
-) -> Result<HealControlServiceClient<InterceptedService<Channel, TonicInterceptor>>, Box<dyn Error>> {
+) -> Result<HealControlServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>, Box<dyn Error>> {
    let interceptor = interceptor.with_rpc_audience(addr)?;
    let channel = match runtime_sources::cached_node_channel(addr).await {
        Some(channel) => channel,
        None => create_new_channel(addr).await?,
    };
    let max_message_size = rustfs_protos::HEAL_CONTROL_RPC_MAX_MESSAGE_SIZE;
+    let channel = ReplayScopeChannel::new(channel, interceptor.replay_scope_audience());
    Ok(HealControlServiceClient::with_interceptor(channel, interceptor)
        .max_decoding_message_size(max_message_size)
        .max_encoding_message_size(max_message_size))
@@ -59,13 +79,14 @@ pub async fn heal_control_time_out_client(
 pub async fn tier_mutation_control_time_out_client(
    addr: &str,
    interceptor: TonicInterceptor,
-) -> Result<TierMutationControlServiceClient<InterceptedService<Channel, TonicInterceptor>>, Box<dyn Error>> {
+) -> Result<TierMutationControlServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>, Box<dyn Error>> {
    let interceptor = interceptor.with_rpc_audience(addr)?;
    let channel = match runtime_sources::cached_node_channel(addr).await {
        Some(channel) => channel,
        None => create_new_channel(addr).await?,
    };
    let max_message_size = rustfs_protos::TIER_MUTATION_RPC_MAX_MESSAGE_SIZE;
+    let channel = ReplayScopeChannel::new(channel, interceptor.replay_scope_audience());
    Ok(TierMutationControlServiceClient::with_interceptor(channel, interceptor)
        .max_decoding_message_size(max_message_size)
        .max_encoding_message_size(max_message_size))
@@ -81,7 +102,7 @@ pub async fn node_service_time_out_client_for_class(
    addr: &String,
    interceptor: TonicInterceptor,
    class: ChannelClass,
-) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>, Box<dyn Error>> {
+) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>, Box<dyn Error>> {
    let interceptor = interceptor.with_rpc_audience(addr)?;
    let channel = match class {
        ChannelClass::Control => match runtime_sources::cached_node_channel(addr).await {
@@ -96,6 +117,7 @@ pub async fn node_service_time_out_client_for_class(
    };

    let max_message_size = rustfs_protos::internode_rpc_max_message_size();
+    let channel = ReplayScopeChannel::new(channel, interceptor.replay_scope_audience());
    Ok(NodeServiceClient::with_interceptor(channel, interceptor)
        .max_decoding_message_size(max_message_size)
        .max_encoding_message_size(max_message_size))
@@ -103,14 +125,83 @@ pub async fn node_service_time_out_client_for_class(

 pub async fn node_service_time_out_client_no_auth(
    addr: &String,
-) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>, Box<dyn Error>> {
+) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>, Box<dyn Error>> {
    node_service_time_out_client(addr, TonicInterceptor::NoOp(NoOpInterceptor)).await
 }

+/// The typed `tonic::Status` an internode RPC failure was converted from, if
+/// this error carries one.
+pub(crate) fn embedded_tonic_status(io_err: &std::io::Error) -> Option<&tonic::Status> {
+    io_err.get_ref()?.downcast_ref::<RpcStatusError>().map(RpcStatusError::status)
+}
+
+/// Decide whether a gRPC status reports a peer we cannot currently reach,
+/// rather than an application outcome from a live peer.
+///
+/// `Unavailable` is the one code that means "no service behind this channel":
+/// the client transport raises it when the connection is broken, and the
+/// server's own not-ready gates use it deliberately.
+///
+/// `Unknown` is the client transport's escape hatch for a cause it could not
+/// map to a code — tower's "Service was not ready: <cause>", an h2 error with
+/// no gRPC mapping. Our handlers never return it, so there its message is the
+/// only evidence available and the anchored needles decide.
+///
+/// Every other code is an answer from a live peer and is never a transport
+/// failure, whatever its message says. That distinction is the point of
+/// classifying by code: a peer relaying its own downstream trouble as
+/// `Internal("connection refused ...")`, or a handler interpolating a local
+/// `io::Error` into `Status::internal`, answered us perfectly well. Marking it
+/// offline over that text is the bug this classification replaces. Likewise a
+/// `Cancelled` "Timeout expired" from the per-RPC channel deadline means the
+/// peer is slow, not gone; gating it would turn load into a partition.
+pub(crate) fn is_network_like_status(status: &tonic::Status) -> bool {
+    match status.code() {
+        tonic::Code::Unavailable => true,
+        tonic::Code::Unknown => message_has_network_needle(&status.to_string()),
+        _ => false,
+    }
+}
+
+/// Substring fallback for failures that only exist as text: dial errors
+/// wrapped by `get_client`, remote `error_info` payloads, and statuses
+/// flattened through `format!`. Needles must stay anchored to transport
+/// context — a bare word like "unavailable" also matches application text
+/// (e.g. a bucket named "unavailable-logs") and would take a healthy peer
+/// offline.
+pub(crate) fn message_has_network_needle(message: &str) -> bool {
+    let message = message.to_ascii_lowercase();
+    [
+        "temporarily offline",
+        "transport error",
+        // tonic >= 0.14 renders Code::Unavailable as
+        // `code: 'The service is currently unavailable'`.
+        "code: 'the service is currently unavailable'",
+        // RUSTFS_COMPAT_TODO(tonic-013-status-render): releases up to 1.0.0-alpha.38 shipped tonic 0.13, which rendered the same status as `status: Unavailable`, and peers relay that text in error_info. Remove after the minimum supported RustFS peer version ships tonic >= 0.14.
+        "status: unavailable",
+        "error trying to connect",
+        "connection refused",
+        "connection reset",
+        "broken pipe",
+        "not connected",
+        "unexpected eof",
+        "timed out",
+        "deadline has elapsed",
+        "connection closed",
+        "connection aborted",
+        "tcp connect error",
+    ]
+    .iter()
+    .any(|needle| message.contains(needle))
+}
+
 pub(crate) fn is_network_like_disk_error(err: &DiskErrorType) -> bool {
    match err {
        DiskError::Timeout => true,
        DiskError::Io(io_err) => {
+            if let Some(status) = embedded_tonic_status(io_err) {
+                return is_network_like_status(status);
+            }
            if matches!(
                io_err.kind(),
                ErrorKind::TimedOut
@@ -124,29 +215,110 @@ pub(crate) fn is_network_like_disk_error(err: &DiskErrorType) -> bool {
                return true;
            }

-            let message = io_err.to_string().to_ascii_lowercase();
-            [
-                "transport error",
-                "unavailable",
-                "error trying to connect",
-                "connection refused",
-                "connection reset",
-                "broken pipe",
-                "not connected",
-                "unexpected eof",
-                "timed out",
-                "deadline has elapsed",
-                "connection closed",
-                "connection aborted",
-                "tcp connect error",
-            ]
-            .iter()
-            .any(|needle| message.contains(needle))
+            message_has_network_needle(&io_err.to_string())
        }
        _ => false,
    }
 }

+/// The transport service that learns an authenticated peer boot epoch and adds the replay-scoped
+/// signature only after one has been observed. The v1/v2 interceptor stays inside this wrapper so
+/// old servers continue receiving precisely the metadata they understand.
+#[derive(Clone, Debug)]
+pub struct ReplayScopeChannel<S> {
+    inner: S,
+    audience: Option<String>,
+}
+
+/// The channel type used by internode clients after v2 authentication and replay-scope handling.
+pub type AuthenticatedChannel = ReplayScopeChannel<Channel>;
+
+static PEER_BOOT_EPOCHS: LazyLock<Mutex<HashMap<String, Uuid>>> = LazyLock::new(|| Mutex::new(HashMap::new()));
+
+impl<S> ReplayScopeChannel<S> {
+    fn new(inner: S, audience: Option<String>) -> Self {
+        Self { inner, audience }
+    }
+}
+
+fn cached_peer_boot_epoch(audience: &str) -> Option<Uuid> {
+    PEER_BOOT_EPOCHS.lock().ok().and_then(|epochs| epochs.get(audience).copied())
+}
+
+fn remember_peer_boot_epoch(audience: String, epoch: Uuid) {
+    if let Ok(mut epochs) = PEER_BOOT_EPOCHS.lock() {
+        epochs.insert(audience, epoch);
+    }
+}
+
+impl<S, ReqBody, ResBody> Service<HttpRequest<ReqBody>> for ReplayScopeChannel<S>
+where
+    S: Service<HttpRequest<ReqBody>, Response = HttpResponse<ResBody>>,
+    S::Error: Send + 'static,
+    S::Future: Send + 'static,
+    ReqBody: Send + 'static,
+    ResBody: Send + 'static,
+{
+    type Response = HttpResponse<ResBody>;
+    type Error = S::Error;
+    type Future = Pin<Box<dyn Future<Output = Result<Self::Response, Self::Error>> + Send>>;
+
+    fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
+        self.inner.poll_ready(cx)
+    }
+
+    fn call(&mut self, mut request: HttpRequest<ReqBody>) -> Self::Future {
+        let authenticated = self.audience.as_ref().is_some_and(|_| {
+            request
+                .headers()
+                .get(RPC_AUTH_VERSION_HEADER)
+                .and_then(|value| value.to_str().ok())
+                == Some(RPC_AUTH_VERSION_V2)
+        });
+        let challenge = authenticated.then(Uuid::new_v4);
+        if let (Some(audience), Some(challenge)) = (self.audience.as_deref(), challenge) {
+            // The challenge is independently HMAC-authenticated by the response proof. It is not
+            // part of v2 so old peers ignore it, while a new peer can safely advertise its epoch.
+            request.headers_mut().insert(
+                RPC_BOOT_EPOCH_CHALLENGE_HEADER,
+                challenge.to_string().parse().expect("UUID must be a valid header value"),
+            );
+            if let (Some(boot_epoch), Some(timestamp), Some(content_sha256)) = (
+                cached_peer_boot_epoch(audience),
+                request.headers().get(TIMESTAMP_HEADER).and_then(|value| value.to_str().ok()),
+                request
+                    .headers()
+                    .get(RPC_CONTENT_SHA256_HEADER)
+                    .and_then(|value| value.to_str().ok()),
+            ) {
+                match gen_tonic_replay_scope_headers(audience, request.uri().path(), timestamp, content_sha256, boot_epoch) {
+                    Ok(headers) => request.headers_mut().extend(headers),
+                    Err(error) => debug!(error = %error, "could not attach replay-scoped RPC signature"),
+                }
+            }
+        }
+
+        let audience = self.audience.clone();
+        let future = self.inner.call(request);
+        Box::pin(async move {
+            let response = future.await?;
+            if let (Some(audience), Some(challenge)) = (audience, challenge) {
+                match verify_tonic_boot_epoch_response(&audience, challenge, response.headers()) {
+                    Ok(epoch) => remember_peer_boot_epoch(audience, epoch),
+                    Err(error)
+                        if response.headers().contains_key(RPC_BOOT_EPOCH_HEADER)
+                            || response.headers().contains_key(RPC_BOOT_EPOCH_PROOF_HEADER) =>
+                    {
+                        debug!(error = %error, "peer boot epoch response proof was rejected")
+                    }
+                    Err(_) => {}
+                }
+            }
+            Ok(response)
+        })
+    }
+}
+
 pub struct TonicSignatureInterceptor {
    audience: Option<String>,
 }
@@ -205,6 +377,13 @@ impl TonicInterceptor {
        }
        Ok(self)
    }
+
+    fn replay_scope_audience(&self) -> Option<String> {
+        match self {
+            Self::Signature(interceptor) => interceptor.audience.clone(),
+            Self::NoOp(_) => None,
+        }
+    }
 }

 impl tonic::service::Interceptor for TonicInterceptor {
@@ -227,6 +406,38 @@ mod tests {
    use tracing_opentelemetry::OpenTelemetrySpanExt;
    use tracing_subscriber::{Registry, layer::SubscriberExt};

+    #[derive(Clone)]
+    struct EpochProofService {
+        audience: String,
+        seen_headers: std::sync::Arc<Mutex<Vec<http::HeaderMap>>>,
+    }
+
+    impl Service<HttpRequest<()>> for EpochProofService {
+        type Response = HttpResponse<()>;
+        type Error = std::convert::Infallible;
+        type Future = std::future::Ready<Result<Self::Response, Self::Error>>;
+
+        fn poll_ready(&mut self, _cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
+            Poll::Ready(Ok(()))
+        }
+
+        fn call(&mut self, request: HttpRequest<()>) -> Self::Future {
+            self.seen_headers
+                .lock()
+                .expect("test header capture lock must not be poisoned")
+                .push(request.headers().clone());
+            let challenge = tonic_boot_epoch_challenge(request.headers())
+                .expect("client challenge must be syntactically valid")
+                .expect("authenticated client request must carry a boot epoch challenge");
+            let mut response = HttpResponse::new(());
+            response.headers_mut().extend(
+                tonic_boot_epoch_response_headers(&self.audience, challenge)
+                    .expect("test server must be able to sign an epoch proof"),
+            );
+            std::future::ready(Ok(response))
+        }
+    }
+
    fn ensure_test_rpc_secret() {
        runtime_sources::ensure_test_rpc_secret();
    }
@@ -269,6 +480,70 @@ mod tests {
        let _ = provider.shutdown();
    }

+    #[test]
+    fn network_like_disk_error_uses_typed_status_code() {
+        // Transport-level Unavailable statuses justify retry/eviction.
+        assert!(is_network_like_disk_error(&DiskError::from(tonic::Status::unavailable(
+            "storage layer is not initialized"
+        ))));
+        // Application statuses from a live peer must not look network-like,
+        // even when their message contains transport-sounding words.
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::internal(
+            "failed to heal bucket \"unavailable-logs\""
+        ))));
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::unauthenticated(
+            "No valid auth token"
+        ))));
+        // A slow peer that blew the per-RPC deadline is still answering.
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::cancelled("Timeout expired"))));
+    }
+
+    #[test]
+    fn embedded_tonic_status_is_recovered_across_error_conversions() {
+        // DiskError and StorageError share one wrapper, so a status keeps its
+        // typed classification whichever error it was converted into first.
+        let from_storage: DiskErrorType = crate::error::Error::from(tonic::Status::unavailable("peer gone")).into();
+        let DiskError::Io(io_err) = &from_storage else {
+            panic!("status-derived disk error should stay an Io error");
+        };
+        assert_eq!(embedded_tonic_status(io_err).map(|status| status.code()), Some(tonic::Code::Unavailable));
+
+        let from_disk = crate::error::Error::from(DiskError::from(tonic::Status::unavailable("peer gone")));
+        let crate::error::Error::Io(io_err) = &from_disk else {
+            panic!("status-derived storage error should stay an Io error");
+        };
+        assert_eq!(embedded_tonic_status(io_err).map(|status| status.code()), Some(tonic::Code::Unavailable));
+    }
+
+    #[test]
+    fn network_like_disk_error_ignores_transport_words_in_application_statuses() {
+        // Same contract as the peer client: a status the peer answered with
+        // is not a transport failure, so it must not drive a reconnect even
+        // when its message describes one.
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::internal(
+            "connection refused while dialing downstream backend"
+        ))));
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::unauthenticated(
+            "connection reset while validating token"
+        ))));
+    }
+
+    #[test]
+    fn network_like_disk_error_requires_anchored_unavailable_needle() {
+        // Regression: a bare "unavailable" needle used to match application
+        // text such as a bucket name.
+        assert!(!is_network_like_disk_error(&DiskError::other("bucket \"unavailable-logs\" not found")));
+        // Anchored renderings of a flattened Unavailable status still match.
+        assert!(is_network_like_disk_error(&DiskError::other(
+            "code: 'The service is currently unavailable', message: \"peer gone\""
+        )));
+        assert!(is_network_like_disk_error(&DiskError::other(
+            "status: Unavailable, message: \"peer gone\""
+        )));
+        assert!(is_network_like_disk_error(&DiskError::other("connection refused")));
+        assert!(!is_network_like_disk_error(&DiskError::FileNotFound));
+    }
+
    #[test]
    fn test_signature_interceptor_keeps_auth_headers() {
        ensure_test_rpc_secret();
@@ -304,6 +579,52 @@ mod tests {
        assert_eq!(interceptor.audience.as_deref(), Some("node-a:9000"));
    }

+    #[test]
+    fn replay_scope_channel_uses_epoch_proof_before_sending_v3() {
+        ensure_test_rpc_secret();
+        let audience = "replay-scope-client-test:9000";
+        PEER_BOOT_EPOCHS
+            .lock()
+            .expect("peer epoch cache lock must not be poisoned")
+            .remove(audience);
+        let seen_headers = std::sync::Arc::new(Mutex::new(Vec::new()));
+        let service = EpochProofService {
+            audience: audience.to_string(),
+            seen_headers: seen_headers.clone(),
+        };
+        let mut channel = ReplayScopeChannel::new(service, Some(audience.to_string()));
+        let make_request = || {
+            let mut request = HttpRequest::builder()
+                .uri("/node_service.NodeService/Ping")
+                .body(())
+                .expect("test RPC request must build");
+            request.headers_mut().extend(
+                gen_tonic_signature_headers(audience, "node_service.NodeService", "Ping", None)
+                    .expect("v2 test headers must mint"),
+            );
+            request
+        };
+
+        futures::executor::block_on(channel.call(make_request())).expect("first request must complete");
+        futures::executor::block_on(channel.call(make_request())).expect("second request must complete");
+
+        let headers = seen_headers.lock().expect("test header capture lock must not be poisoned");
+        assert_eq!(headers.len(), 2);
+        assert!(headers[0].contains_key(RPC_BOOT_EPOCH_CHALLENGE_HEADER));
+        assert!(
+            !headers[0].contains_key(RPC_REPLAY_SCOPE_VERSION_HEADER),
+            "the first request must remain v2-compatible until the peer proves its epoch"
+        );
+        assert!(
+            headers[1].contains_key(RPC_REPLAY_SCOPE_VERSION_HEADER),
+            "the second request must carry the replay-scoped v3 signature"
+        );
+        PEER_BOOT_EPOCHS
+            .lock()
+            .expect("peer epoch cache lock must not be poisoned")
+            .remove(audience);
+    }
+
    #[test]
    fn test_signature_interceptor_requires_generated_method_metadata() {
        ensure_test_rpc_secret();
@@ -20,8 +20,8 @@
 //! rustfs/rustfs#4402) is anchored by the `ghsa_r5qv_*` tests in the module
 //! below, plus the broader negative-signature suite. The advisory class is: a
 //! node must never accept an RPC whose auth is missing, malformed, or signed
-//! with the default/empty shared secret. Body-bound v2 requests additionally
-//! receive process-local replay protection. See
+//! with the default/empty shared secret. Body-bound v2 requests and all replay-scoped v3
+//! requests additionally receive process-local replay protection. See
 //! `docs/testing/security-regressions.md` for the full advisory -> test map.
 //!
 //! Advisory: <https://github.com/rustfs/rustfs/security/advisories/GHSA-r5qv-rc46-hv8q>
@@ -50,13 +50,22 @@ use uuid::Uuid;
 type HmacSha256 = Hmac<Sha256>;

 const SIGNATURE_HEADER: &str = "x-rustfs-signature";
-const TIMESTAMP_HEADER: &str = "x-rustfs-timestamp";
-const RPC_AUTH_VERSION_HEADER: &str = "x-rustfs-rpc-auth-version";
+pub(crate) const TIMESTAMP_HEADER: &str = "x-rustfs-timestamp";
+pub(crate) const RPC_AUTH_VERSION_HEADER: &str = "x-rustfs-rpc-auth-version";
 const RPC_SIGNATURE_V2_HEADER: &str = "x-rustfs-rpc-signature-v2";
 const RPC_NONCE_HEADER: &str = "x-rustfs-rpc-nonce";
 pub(crate) const RPC_CONTENT_SHA256_HEADER: &str = "x-rustfs-content-sha256";
-const RPC_AUTH_VERSION_V2: &str = "2";
+pub(crate) const RPC_AUTH_VERSION_V2: &str = "2";
+pub const RPC_REPLAY_SCOPE_VERSION_HEADER: &str = "x-rustfs-rpc-replay-scope-version";
+pub const RPC_REPLAY_SCOPE_SIGNATURE_HEADER: &str = "x-rustfs-rpc-signature-v3";
+pub const RPC_REPLAY_SCOPE_NONCE_HEADER: &str = "x-rustfs-rpc-replay-nonce";
+pub const RPC_BOOT_EPOCH_HEADER: &str = "x-rustfs-rpc-boot-epoch";
+pub const RPC_BOOT_EPOCH_CHALLENGE_HEADER: &str = "x-rustfs-rpc-boot-epoch-challenge";
+pub const RPC_BOOT_EPOCH_PROOF_HEADER: &str = "x-rustfs-rpc-boot-epoch-proof";
+const RPC_REPLAY_SCOPE_VERSION_V3: &str = "3";
 const RPC_RESPONSE_PROOF_DOMAIN: &[u8] = b"rustfs-rpc-response-proof-v1\0";
+const RPC_REPLAY_SCOPE_DOMAIN: &[u8] = b"rustfs-rpc-replay-scope-v3\0";
+const RPC_BOOT_EPOCH_PROOF_DOMAIN: &[u8] = b"rustfs-rpc-boot-epoch-proof-v1\0";
 const UNSIGNED_PAYLOAD: &str = "UNSIGNED-PAYLOAD";
 const UNSIGNED_PAYLOAD_NONCE: &str = "unsigned";
 const SIGNATURE_VALID_DURATION: i64 = 300; // 5 minutes
@@ -75,9 +84,15 @@ static INTERNODE_RPC_BODY_DIGEST_STRICT: LazyLock<bool> = LazyLock::new(|| {
        rustfs_config::DEFAULT_INTERNODE_RPC_BODY_DIGEST_STRICT,
    )
 });
-// Sized for peak legitimate body-bound mutation RPS x the retention window; overflow fails closed
-// and increments the replay-cache overflow counter. Clamped to at least 1 so a misconfigured zero
-// cannot disable replay protection by rejecting every body-bound request.
+static INTERNODE_RPC_REPLAY_SCOPE_STRICT: LazyLock<bool> = LazyLock::new(|| {
+    get_env_bool(
+        rustfs_config::ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT,
+        rustfs_config::DEFAULT_INTERNODE_RPC_REPLAY_SCOPE_STRICT,
+    )
+});
+// Sized for peak legitimate authenticated RPC RPS x the retention window once replay scope is
+// active; overflow fails closed and increments the replay-cache overflow counter. Clamped to at
+// least 1 so a misconfigured zero cannot disable replay protection by rejecting every request.
 static REPLAY_CACHE_CAPACITY: LazyLock<usize> = LazyLock::new(|| {
    rustfs_utils::get_env_usize(
        rustfs_config::ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY,
@@ -86,6 +101,7 @@ static REPLAY_CACHE_CAPACITY: LazyLock<usize> = LazyLock::new(|| {
    .max(1)
 });
 static RPC_SECRET_RESOLUTION_LOG_ONCE: Once = Once::new();
+static RPC_BOOT_EPOCH: LazyLock<Uuid> = LazyLock::new(Uuid::new_v4);

 #[derive(Default)]
 struct RpcNonceCache {
@@ -313,6 +329,189 @@ fn verify_signature_v2(secret: &str, scope: SignatureV2Scope<'_>, signature: &st
    mac.verify_slice(&signature).is_ok()
 }

+#[derive(Clone, Copy)]
+struct ReplayScope<'a> {
+    audience: &'a str,
+    path: &'a str,
+    timestamp: &'a str,
+    nonce: Uuid,
+    content_sha256: &'a str,
+    boot_epoch: Uuid,
+}
+
+fn update_replay_scope(mac: &mut HmacSha256, scope: ReplayScope<'_>) {
+    mac.update(RPC_REPLAY_SCOPE_DOMAIN);
+    for part in [
+        scope.audience.as_bytes(),
+        b"|",
+        scope.path.as_bytes(),
+        b"|POST|",
+        scope.timestamp.as_bytes(),
+        b"|",
+        scope.nonce.as_bytes(),
+        b"|",
+        scope.content_sha256.as_bytes(),
+        b"|",
+        scope.boot_epoch.as_bytes(),
+    ] {
+        mac.update(part);
+    }
+}
+
+fn generate_replay_scope_signature(secret: &str, scope: ReplayScope<'_>) -> std::io::Result<String> {
+    let mut mac =
+        <HmacSha256 as KeyInit>::new_from_slice(secret.as_bytes()).map_err(|_| std::io::Error::other("Invalid RPC HMAC key"))?;
+    update_replay_scope(&mut mac, scope);
+    Ok(general_purpose::STANDARD.encode(mac.finalize().into_bytes()))
+}
+
+fn verify_replay_scope_signature(secret: &str, scope: ReplayScope<'_>, signature: &str) -> bool {
+    let Ok(signature) = general_purpose::STANDARD.decode(signature) else {
+        return false;
+    };
+    let Ok(mut mac) = <HmacSha256 as KeyInit>::new_from_slice(secret.as_bytes()) else {
+        return false;
+    };
+    update_replay_scope(&mut mac, scope);
+    mac.verify_slice(&signature).is_ok()
+}
+
+fn update_boot_epoch_proof(mac: &mut HmacSha256, audience: &str, challenge: Uuid, boot_epoch: Uuid) {
+    mac.update(RPC_BOOT_EPOCH_PROOF_DOMAIN);
+    mac.update(audience.as_bytes());
+    mac.update(b"|");
+    mac.update(challenge.as_bytes());
+    mac.update(boot_epoch.as_bytes());
+}
+
+fn generate_boot_epoch_proof(secret: &str, audience: &str, challenge: Uuid, boot_epoch: Uuid) -> std::io::Result<String> {
+    if audience.is_empty() || challenge.is_nil() || boot_epoch.is_nil() {
+        return Err(std::io::Error::other("Invalid RPC boot epoch proof scope"));
+    }
+    let mut mac =
+        <HmacSha256 as KeyInit>::new_from_slice(secret.as_bytes()).map_err(|_| std::io::Error::other("Invalid RPC HMAC key"))?;
+    update_boot_epoch_proof(&mut mac, audience, challenge, boot_epoch);
+    Ok(general_purpose::STANDARD.encode(mac.finalize().into_bytes()))
+}
+
+fn verify_boot_epoch_proof(secret: &str, audience: &str, challenge: Uuid, boot_epoch: Uuid, proof: &str) -> std::io::Result<()> {
+    if audience.is_empty() || challenge.is_nil() || boot_epoch.is_nil() {
+        return Err(std::io::Error::other("Invalid RPC boot epoch proof scope"));
+    }
+    let proof = general_purpose::STANDARD
+        .decode(proof)
+        .map_err(|_| std::io::Error::other("Invalid RPC boot epoch proof"))?;
+    let mut mac =
+        <HmacSha256 as KeyInit>::new_from_slice(secret.as_bytes()).map_err(|_| std::io::Error::other("Invalid RPC HMAC key"))?;
+    update_boot_epoch_proof(&mut mac, audience, challenge, boot_epoch);
+    mac.verify_slice(&proof)
+        .map_err(|_| std::io::Error::new(std::io::ErrorKind::PermissionDenied, "Invalid RPC boot epoch proof"))
+}
+
+fn non_nil_uuid(value: &str, name: &str) -> std::io::Result<Uuid> {
+    let value = Uuid::parse_str(value).map_err(|_| std::io::Error::other(format!("Invalid {name}")))?;
+    (!value.is_nil())
+        .then_some(value)
+        .ok_or_else(|| std::io::Error::other(format!("Invalid {name}")))
+}
+
+fn parse_tonic_rpc_path(path: &str) -> std::io::Result<(&str, &str)> {
+    path.strip_prefix('/')
+        .and_then(|path| path.split_once('/'))
+        .filter(|(service, rpc_method)| !service.is_empty() && !rpc_method.is_empty() && !rpc_method.contains('/'))
+        .ok_or_else(|| std::io::Error::other("Invalid RPC request path"))
+}
+
+/// The process-unique epoch included in every replay-scoped server verification.
+///
+/// A fresh process gets a fresh value, so a signature captured before a server restart cannot be
+/// admitted even though the bounded in-memory nonce cache necessarily starts empty again.
+pub fn tonic_rpc_boot_epoch() -> Uuid {
+    *RPC_BOOT_EPOCH
+}
+
+/// Build the additive replay-scope headers for a request that already carries rolling-upgrade-safe
+/// v1/v2 metadata. `timestamp` and `content_sha256` are deliberately reused from the v2 scope so
+/// old servers can continue validating the same request unchanged.
+pub fn gen_tonic_replay_scope_headers(
+    audience: &str,
+    path: &str,
+    timestamp: &str,
+    content_sha256: &str,
+    boot_epoch: Uuid,
+) -> std::io::Result<HeaderMap> {
+    if audience.is_empty() || !path.starts_with('/') || !valid_content_sha256(content_sha256) || boot_epoch.is_nil() {
+        return Err(std::io::Error::other("Invalid replay-scoped RPC signing scope"));
+    }
+    parse_tonic_rpc_path(path)?;
+    timestamp
+        .parse::<i64>()
+        .map_err(|_| std::io::Error::other("Invalid timestamp format"))?;
+
+    let nonce = Uuid::new_v4();
+    let signature = generate_replay_scope_signature(
+        &get_shared_secret()?,
+        ReplayScope {
+            audience,
+            path,
+            timestamp,
+            nonce,
+            content_sha256,
+            boot_epoch,
+        },
+    )?;
+    let mut headers = HeaderMap::new();
+    headers.insert(RPC_REPLAY_SCOPE_VERSION_HEADER, HeaderValue::from_static(RPC_REPLAY_SCOPE_VERSION_V3));
+    headers.insert(
+        RPC_REPLAY_SCOPE_SIGNATURE_HEADER,
+        header_value(&signature, RPC_REPLAY_SCOPE_SIGNATURE_HEADER)?,
+    );
+    headers.insert(
+        RPC_REPLAY_SCOPE_NONCE_HEADER,
+        header_value(&nonce.to_string(), RPC_REPLAY_SCOPE_NONCE_HEADER)?,
+    );
+    headers.insert(RPC_BOOT_EPOCH_HEADER, header_value(&boot_epoch.to_string(), RPC_BOOT_EPOCH_HEADER)?);
+    Ok(headers)
+}
+
+/// Parse the optional client challenge used to authenticate a server boot-epoch advertisement.
+pub fn tonic_boot_epoch_challenge(headers: &HeaderMap) -> std::io::Result<Option<Uuid>> {
+    headers
+        .get(RPC_BOOT_EPOCH_CHALLENGE_HEADER)
+        .map(|value| {
+            value
+                .to_str()
+                .map_err(|_| std::io::Error::other("Invalid RPC boot epoch challenge"))
+                .and_then(|value| non_nil_uuid(value, "RPC boot epoch challenge"))
+        })
+        .transpose()
+}
+
+/// Build the authenticated response headers for a client boot-epoch challenge.
+pub fn tonic_boot_epoch_response_headers(audience: &str, challenge: Uuid) -> std::io::Result<HeaderMap> {
+    let boot_epoch = tonic_rpc_boot_epoch();
+    let proof = generate_boot_epoch_proof(&get_shared_secret()?, audience, challenge, boot_epoch)?;
+    let mut headers = HeaderMap::new();
+    headers.insert(RPC_BOOT_EPOCH_HEADER, header_value(&boot_epoch.to_string(), RPC_BOOT_EPOCH_HEADER)?);
+    headers.insert(RPC_BOOT_EPOCH_PROOF_HEADER, header_value(&proof, RPC_BOOT_EPOCH_PROOF_HEADER)?);
+    Ok(headers)
+}
+
+/// Verify the server boot-epoch response for a challenge generated by this client.
+pub fn verify_tonic_boot_epoch_response(audience: &str, challenge: Uuid, headers: &HeaderMap) -> std::io::Result<Uuid> {
+    let boot_epoch = headers
+        .get(RPC_BOOT_EPOCH_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC boot epoch"))
+        .and_then(|value| non_nil_uuid(value, "RPC boot epoch"))?;
+    let proof = headers
+        .get(RPC_BOOT_EPOCH_PROOF_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC boot epoch proof"))?;
+    verify_boot_epoch_proof(&get_shared_secret()?, audience, challenge, boot_epoch, proof)?;
+    Ok(boot_epoch)
+}
+
 fn valid_content_sha256(value: &str) -> bool {
    value == UNSIGNED_PAYLOAD
        || (value.len() == 64
@@ -443,6 +642,16 @@ pub fn set_tonic_canonical_body_digest<T>(request: &mut tonic::Request<T>, canon
    Ok(())
 }

+pub fn set_tonic_mutation_body_digest<T: rustfs_protos::CanonicalMutationBody>(
+    request: &mut tonic::Request<T>,
+) -> std::io::Result<()> {
+    let canonical_body = request
+        .get_ref()
+        .canonical_body()
+        .map_err(|_| std::io::Error::other("RPC mutation body length cannot be represented"))?;
+    set_tonic_canonical_body_digest(request, &canonical_body)
+}
+
 pub fn verify_tonic_canonical_body_digest<T>(request: &tonic::Request<T>, canonical_body: &[u8]) -> std::io::Result<()> {
    let version = request
        .metadata()
@@ -466,7 +675,7 @@ pub fn verify_tonic_canonical_body_digest<T>(request: &tonic::Request<T>, canoni
    Ok(())
 }

-/// Verify a mutating disk RPC's canonical body digest with a rolling-upgrade fallback.
+/// Verify a mutating RPC's canonical body digest with a rolling-upgrade fallback.
 ///
 /// When the request carries a real (non-`UNSIGNED-PAYLOAD`) content SHA-256 it is verified exactly
 /// like [`verify_tonic_canonical_body_digest`]. The digest value is a member of the signed v2
@@ -497,7 +706,7 @@ fn verify_tonic_mutation_body_digest_with_strictness<T>(
        Some(digest) if digest != UNSIGNED_PAYLOAD => verify_tonic_canonical_body_digest(request, canonical_body),
        _ => {
            // RUSTFS_COMPAT_TODO(disk-mutation-body-digest): accept digestless peers during rolling upgrades. Remove after the
-            // minimum supported RustFS peer version body-binds every mutating disk RPC.
+            // minimum supported RustFS peer version body-binds every mutating RPC.
            if strict {
                return Err(std::io::Error::other("RPC mutation requires a body-bound v2 signature"));
            }
@@ -521,6 +730,17 @@ fn has_v2_auth_headers(headers: &HeaderMap) -> bool {
    .any(|name| headers.contains_key(*name))
 }

+fn has_replay_scope_headers(headers: &HeaderMap) -> bool {
+    [
+        RPC_REPLAY_SCOPE_VERSION_HEADER,
+        RPC_REPLAY_SCOPE_SIGNATURE_HEADER,
+        RPC_REPLAY_SCOPE_NONCE_HEADER,
+        RPC_BOOT_EPOCH_HEADER,
+    ]
+    .iter()
+    .any(|name| headers.contains_key(*name))
+}
+
 /// Whether the server requires target-bound v2 authentication on every internode gRPC request,
 /// rejecting the legacy constant-target fallback instead of accepting it. Default-off rollout
 /// lever gated on the v1-fallback counter reading zero fleet-wide; see
@@ -530,9 +750,127 @@ fn internode_rpc_signature_strict() -> bool {
    *INTERNODE_RPC_SIGNATURE_STRICT
 }

+fn internode_rpc_replay_scope_strict() -> bool {
+    *INTERNODE_RPC_REPLAY_SCOPE_STRICT
+}
+
+fn verify_tonic_replay_scope_signature(audience: &str, path: &str, headers: &HeaderMap) -> std::io::Result<()> {
+    if audience.is_empty() {
+        return Err(std::io::Error::other("Missing RPC audience"));
+    }
+    parse_tonic_rpc_path(path)?;
+
+    let version = headers
+        .get(RPC_REPLAY_SCOPE_VERSION_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC replay scope version"))?;
+    if version != RPC_REPLAY_SCOPE_VERSION_V3 {
+        return Err(std::io::Error::other("Unsupported RPC replay scope version"));
+    }
+    let signature = headers
+        .get(RPC_REPLAY_SCOPE_SIGNATURE_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC replay scope signature"))?;
+    let timestamp = headers
+        .get(TIMESTAMP_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing timestamp header"))?;
+    let signed_at = timestamp
+        .parse::<i64>()
+        .map_err(|_| std::io::Error::other("Invalid timestamp format"))?;
+    check_timestamp(signed_at)?;
+    let nonce = headers
+        .get(RPC_REPLAY_SCOPE_NONCE_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC replay scope nonce"))
+        .and_then(|value| non_nil_uuid(value, "RPC replay scope nonce"))?;
+    let content_sha256 = headers
+        .get(RPC_CONTENT_SHA256_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC content SHA-256"))?;
+    if !valid_content_sha256(content_sha256) {
+        return Err(std::io::Error::other("Invalid RPC content SHA-256"));
+    }
+    let boot_epoch = headers
+        .get(RPC_BOOT_EPOCH_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC boot epoch"))
+        .and_then(|value| non_nil_uuid(value, "RPC boot epoch"))?;
+    let secret = get_shared_secret()?;
+    if !verify_replay_scope_signature(
+        &secret,
+        ReplayScope {
+            audience,
+            path,
+            timestamp,
+            nonce,
+            content_sha256,
+            boot_epoch,
+        },
+        signature,
+    ) {
+        return Err(std::io::Error::other("Invalid RPC replay scope signature"));
+    }
+    if boot_epoch != tonic_rpc_boot_epoch() {
+        return Err(std::io::Error::other("RPC boot epoch is stale"));
+    }
+    check_and_record_nonce(nonce, signed_at)
+}
+
 /// Verify gRPC authentication, preferring v2 without downgrade on malformed v2 metadata.
 pub fn verify_tonic_rpc_signature(audience: &str, path: &str, headers: &HeaderMap) -> std::io::Result<()> {
-    verify_tonic_rpc_signature_with_strictness(audience, path, headers, internode_rpc_signature_strict())
+    verify_tonic_rpc_signature_with_policy(
+        audience,
+        path,
+        headers,
+        internode_rpc_signature_strict(),
+        internode_rpc_replay_scope_strict(),
+        false,
+    )
+}
+
+/// Verify gRPC authentication while allowing the narrowly scoped v2 `Ping` bootstrap used to
+/// obtain an authenticated server boot epoch when replay-scope strictness is enabled.
+pub fn verify_tonic_rpc_signature_with_bootstrap(
+    audience: &str,
+    path: &str,
+    headers: &HeaderMap,
+    allow_replay_scope_bootstrap: bool,
+) -> std::io::Result<()> {
+    verify_tonic_rpc_signature_with_policy(
+        audience,
+        path,
+        headers,
+        internode_rpc_signature_strict(),
+        internode_rpc_replay_scope_strict(),
+        allow_replay_scope_bootstrap,
+    )
+}
+
+fn verify_tonic_rpc_signature_with_policy(
+    audience: &str,
+    path: &str,
+    headers: &HeaderMap,
+    signature_strict: bool,
+    replay_scope_strict: bool,
+    allow_replay_scope_bootstrap: bool,
+) -> std::io::Result<()> {
+    if has_replay_scope_headers(headers) {
+        return verify_tonic_replay_scope_signature(audience, path, headers);
+    }
+
+    // Only a method-bound v2 Ping with a syntactically valid challenge may bootstrap a strict
+    // client after its peer restarts. Legacy metadata never gets this exception.
+    let bootstrap = allow_replay_scope_bootstrap
+        && has_v2_auth_headers(headers)
+        && tonic_boot_epoch_challenge(headers).is_ok_and(|challenge| challenge.is_some());
+    if replay_scope_strict && !bootstrap {
+        return Err(std::io::Error::other("RPC replay-scoped authentication required"));
+    }
+
+    verify_tonic_rpc_signature_with_strictness(audience, path, headers, signature_strict)?;
+    global_internode_metrics().record_replay_scope_fallback();
+    Ok(())
 }

 /// [`verify_tonic_rpc_signature`] with the strict gate injected as a parameter, so both rollout
@@ -677,11 +1015,28 @@ mod tests {
    use crate::cluster::rpc::context_propagation::REQUEST_ID_HEADER;
    use crate::runtime::sources as runtime_sources;
    use http::{HeaderMap, Method};
+    use rustfs_protos::{
+        CanonicalMutationBody as _, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS,
+        proto_gen::node_service::{Mss, SignalServiceRequest},
+    };
+    use std::collections::HashMap;
    use std::io::{self, Write};
    use std::sync::{Arc, Mutex};
    use time::OffsetDateTime;
    use tracing_subscriber::fmt::MakeWriter;

+    fn signal_service_request(signal: &str, sub_system: &str, dry_run: &str) -> SignalServiceRequest {
+        SignalServiceRequest {
+            vars: Some(Mss {
+                value: HashMap::from([
+                    (PEER_RESTSIGNAL.to_string(), signal.to_string()),
+                    (PEER_RESTSUB_SYS.to_string(), sub_system.to_string()),
+                    (PEER_RESTDRY_RUN.to_string(), dry_run.to_string()),
+                ]),
+            }),
+        }
+    }
+
    #[derive(Clone, Default)]
    struct CapturedLogs {
        buffer: Arc<Mutex<Vec<u8>>>,
@@ -1246,6 +1601,104 @@ mod tests {
        assert_eq!(error.to_string(), "Invalid RPC v2 signature");
    }

+    #[test]
+    fn replay_scope_binds_path_epoch_and_random_nonce() {
+        ensure_test_rpc_secret();
+        let path = "/node_service.NodeService/Ping";
+        let mut headers = gen_tonic_signature_headers("node-a:9000", "node_service.NodeService", "Ping", None)
+            .expect("v2 compatibility headers should build");
+        let timestamp = headers
+            .get(TIMESTAMP_HEADER)
+            .and_then(|value| value.to_str().ok())
+            .expect("v2 timestamp")
+            .to_string();
+        let content_sha256 = headers
+            .get(RPC_CONTENT_SHA256_HEADER)
+            .and_then(|value| value.to_str().ok())
+            .expect("v2 content digest")
+            .to_string();
+        headers.extend(
+            gen_tonic_replay_scope_headers("node-a:9000", path, &timestamp, &content_sha256, tonic_rpc_boot_epoch())
+                .expect("replay-scope headers should build"),
+        );
+
+        assert!(
+            verify_tonic_rpc_signature_with_policy("node-a:9000", path, &headers, false, false, false).is_ok(),
+            "the first replay-scoped request must be accepted"
+        );
+        let replay = verify_tonic_rpc_signature_with_policy("node-a:9000", path, &headers, false, false, false)
+            .expect_err("the random replay-scope nonce must be single-use");
+        assert_eq!(replay.to_string(), "RPC request replay detected");
+
+        let path_error = verify_tonic_replay_scope_signature("node-a:9000", "/node_service.NodeService/SignalService", &headers)
+            .expect_err("a replay-scoped signature must not move to another method");
+        assert_eq!(path_error.to_string(), "Invalid RPC replay scope signature");
+    }
+
+    #[test]
+    fn replay_scope_rejects_partial_metadata_and_stale_epoch_without_fallback() {
+        ensure_test_rpc_secret();
+        let path = "/node_service.NodeService/Ping";
+        let mut partial = gen_tonic_signature_headers("node-a:9000", "node_service.NodeService", "Ping", None)
+            .expect("v2 compatibility headers should build");
+        partial.insert(RPC_REPLAY_SCOPE_VERSION_HEADER, HeaderValue::from_static(RPC_REPLAY_SCOPE_VERSION_V3));
+        let error = verify_tonic_rpc_signature_with_policy("node-a:9000", path, &partial, false, false, false)
+            .expect_err("partial replay-scope metadata must never downgrade to v2");
+        assert_eq!(error.to_string(), "Missing RPC replay scope signature");
+
+        let timestamp = partial
+            .get(TIMESTAMP_HEADER)
+            .and_then(|value| value.to_str().ok())
+            .expect("v2 timestamp")
+            .to_string();
+        let content_sha256 = partial
+            .get(RPC_CONTENT_SHA256_HEADER)
+            .and_then(|value| value.to_str().ok())
+            .expect("v2 content digest")
+            .to_string();
+        let stale_epoch = Uuid::new_v4();
+        partial.extend(
+            gen_tonic_replay_scope_headers("node-a:9000", path, &timestamp, &content_sha256, stale_epoch)
+                .expect("replay-scope headers should build"),
+        );
+        let stale = verify_tonic_rpc_signature_with_policy("node-a:9000", path, &partial, false, false, false)
+            .expect_err("a signature from a prior server boot epoch must be rejected");
+        assert_eq!(stale.to_string(), "RPC boot epoch is stale");
+    }
+
+    #[test]
+    fn replay_scope_strictness_allows_only_authenticated_ping_bootstrap() {
+        ensure_test_rpc_secret();
+        let mut headers = gen_tonic_signature_headers("node-a:9000", "node_service.NodeService", "Ping", None)
+            .expect("v2 compatibility headers should build");
+        let rejected =
+            verify_tonic_rpc_signature_with_policy("node-a:9000", "/node_service.NodeService/Ping", &headers, false, true, false)
+                .expect_err("strict replay scope must reject stripped new metadata");
+        assert_eq!(rejected.to_string(), "RPC replay-scoped authentication required");
+
+        headers.insert(
+            RPC_BOOT_EPOCH_CHALLENGE_HEADER,
+            HeaderValue::from_str(&Uuid::new_v4().to_string()).expect("UUID header"),
+        );
+        assert!(
+            verify_tonic_rpc_signature_with_policy("node-a:9000", "/node_service.NodeService/Ping", &headers, false, true, true,)
+                .is_ok(),
+            "only the signed Ping bootstrap may obtain a new server epoch in strict mode"
+        );
+    }
+
+    #[test]
+    fn boot_epoch_response_proof_binds_audience_challenge_and_epoch() {
+        ensure_test_rpc_secret();
+        let challenge = Uuid::new_v4();
+        let headers = tonic_boot_epoch_response_headers("node-a:9000", challenge).expect("proof headers should build");
+        let epoch =
+            verify_tonic_boot_epoch_response("node-a:9000", challenge, &headers).expect("matching proof headers should verify");
+        assert_eq!(epoch, tonic_rpc_boot_epoch());
+        assert!(verify_tonic_boot_epoch_response("node-b:9000", challenge, &headers).is_err());
+        assert!(verify_tonic_boot_epoch_response("node-a:9000", Uuid::new_v4(), &headers).is_err());
+    }
+
    #[test]
    fn malformed_v2_auth_does_not_downgrade_to_valid_legacy_signature() {
        ensure_test_rpc_secret();
@@ -1596,6 +2049,72 @@ mod tests {
        assert_eq!(stripped.to_string(), "RPC content SHA-256 mismatch");
    }

+    #[test]
+    fn signal_service_mutation_contract_rejects_tampering_and_replay() {
+        ensure_test_rpc_secret();
+        let body = signal_service_request("2", "scanner", "false")
+            .canonical_body()
+            .expect("small signal request should encode");
+        let mut request = tonic::Request::new(());
+        set_tonic_canonical_body_digest(&mut request, &body).expect("canonical body digest should be attached");
+        let content_sha256 = request
+            .metadata()
+            .get(RPC_CONTENT_SHA256_HEADER)
+            .and_then(|value| value.to_str().ok());
+        let headers = gen_tonic_signature_headers("node-a:9000", "node_service.NodeService", "SignalService", content_sha256)
+            .expect("body-bound auth headers should build");
+        request.metadata_mut().as_mut().extend(headers.clone());
+
+        assert!(
+            verify_tonic_rpc_signature("node-a:9000", "/node_service.NodeService/SignalService", &headers).is_ok(),
+            "the first body-bound signal request must authenticate"
+        );
+        assert!(verify_tonic_mutation_body_digest(&request, &body).is_ok());
+
+        let tampered = signal_service_request("1", "scanner", "false")
+            .canonical_body()
+            .expect("small signal request should encode");
+        let error = verify_tonic_mutation_body_digest(&request, &tampered)
+            .expect_err("changing the signal must invalidate the signed digest");
+        assert_eq!(error.to_string(), "RPC content SHA-256 mismatch");
+
+        let replay = verify_tonic_rpc_signature("node-a:9000", "/node_service.NodeService/SignalService", &headers)
+            .expect_err("reusing the signal nonce must fail");
+        assert_eq!(replay.to_string(), "RPC request replay detected");
+    }
+
+    #[test]
+    #[serial_test::serial(rpc_body_digest_fallback_counter)]
+    fn signal_service_mutation_contract_preserves_rollout_fallback_and_strictness() {
+        let body = signal_service_request("2", "scanner", "false")
+            .canonical_body()
+            .expect("small signal request should encode");
+        let before = global_internode_metrics().snapshot().body_digest_fallback_total;
+        let digestless = tonic::Request::new(());
+
+        assert!(
+            verify_tonic_mutation_body_digest_with_strictness(&digestless, &body, false).is_ok(),
+            "old peers must remain compatible while the rollout gate is open"
+        );
+        assert_eq!(
+            global_internode_metrics().snapshot().body_digest_fallback_total,
+            before + 1,
+            "accepted digestless signal requests must be visible in the fallback metric"
+        );
+
+        let error = verify_tonic_mutation_body_digest_with_strictness(&digestless, &body, true)
+            .expect_err("strict mode must reject a digestless signal request");
+        assert_eq!(error.to_string(), "RPC mutation requires a body-bound v2 signature");
+
+        let mut bound = tonic::Request::new(());
+        set_tonic_canonical_body_digest(&mut bound, &body).expect("canonical body digest should be attached");
+        bound
+            .metadata_mut()
+            .as_mut()
+            .insert(RPC_AUTH_VERSION_HEADER, HeaderValue::from_static(RPC_AUTH_VERSION_V2));
+        assert!(verify_tonic_mutation_body_digest_with_strictness(&bound, &body, true).is_ok());
+    }
+
    #[test]
    fn nonce_cache_rejects_replay_after_wall_clock_regression() {
        let now = Instant::now();
@@ -26,13 +26,18 @@ pub(crate) mod runtime_sources;
 pub use background_monitor::shutdown_background_monitors;
 pub(crate) use background_monitor::spawn_background_monitor;
 pub use client::{
-    TonicInterceptor, gen_tonic_signature_interceptor, node_service_time_out_client, node_service_time_out_client_no_auth,
+    AuthenticatedChannel, TonicInterceptor, gen_tonic_signature_interceptor, node_service_time_out_client,
+    node_service_time_out_client_no_auth,
 };
+// Re-exported through `api::rpc`; not every item is consumed inside this crate.
+#[allow(unused_imports)]
 pub use http_auth::{
-    TONIC_RPC_PREFIX, build_auth_headers, gen_signature_headers, gen_tonic_signature_headers, normalize_tonic_rpc_audience,
-    set_tonic_canonical_body_digest, sign_ns_scanner_capability, sign_tonic_rpc_response_proof, verify_ns_scanner_capability,
-    verify_rpc_signature, verify_tonic_canonical_body_digest, verify_tonic_mutation_body_digest, verify_tonic_rpc_response_proof,
-    verify_tonic_rpc_signature,
+    TONIC_RPC_PREFIX, build_auth_headers, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
+    normalize_tonic_rpc_audience, set_tonic_canonical_body_digest, set_tonic_mutation_body_digest, sign_ns_scanner_capability,
+    sign_tonic_rpc_response_proof, tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers, verify_ns_scanner_capability,
+    verify_rpc_signature, verify_tonic_boot_epoch_response, verify_tonic_canonical_body_digest,
+    verify_tonic_mutation_body_digest, verify_tonic_rpc_response_proof, verify_tonic_rpc_signature,
+    verify_tonic_rpc_signature_with_bootstrap,
 };
 #[cfg(test)]
 pub(crate) use internode_data_transport::TcpHttpInternodeDataTransport;
@@ -13,10 +13,10 @@
 // limitations under the License.

 use crate::cluster::rpc::client::{
-    TonicInterceptor, gen_tonic_signature_interceptor, heal_control_time_out_client, node_service_time_out_client,
-    tier_mutation_control_time_out_client,
+    AuthenticatedChannel, TonicInterceptor, embedded_tonic_status, gen_tonic_signature_interceptor, heal_control_time_out_client,
+    is_network_like_status, message_has_network_needle, node_service_time_out_client, tier_mutation_control_time_out_client,
 };
-use crate::cluster::rpc::{set_tonic_canonical_body_digest, verify_tonic_rpc_response_proof};
+use crate::cluster::rpc::{set_tonic_canonical_body_digest, set_tonic_mutation_body_digest, verify_tonic_rpc_response_proof};
 use crate::error::{Error, Result};
 use crate::storage_api_contracts::internode::{
    SCANNER_ACTIVITY_LEGACY_PROTOCOL_VERSION, SCANNER_ACTIVITY_PREVIOUS_PROTOCOL_VERSION, SCANNER_ACTIVITY_PROTOCOL_VERSION,
@@ -50,6 +50,7 @@ use rustfs_protos::proto_gen::node_service::{
    TierMutationPeerState, TierMutationPrepareRequest, node_service_client::NodeServiceClient,
    tier_mutation_control_service_client::TierMutationControlServiceClient,
 };
+pub use rustfs_protos::{PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS};
 use rustfs_protos::{TierMutationRpcPhase, evict_failed_connection};
 use rustfs_utils::XHost;
 use serde::{Deserialize, Serialize as _};
@@ -65,13 +66,9 @@ use std::{
 use tokio::{net::TcpStream, time::Duration};
 use tonic::Request;
 use tonic::service::interceptor::InterceptedService;
-use tonic::transport::Channel;
 use tracing::{debug, info, warn};
 use uuid::Uuid;

-pub const PEER_RESTSIGNAL: &str = "signal";
-pub const PEER_RESTSUB_SYS: &str = "sub-sys";
-pub const PEER_RESTDRY_RUN: &str = "dry-run";
 pub const SERVICE_SIGNAL_REFRESH_CONFIG: u64 = 1;
 pub const SERVICE_SIGNAL_RELOAD_DYNAMIC: u64 = 2;
 const BACKGROUND_HEAL_STATUS_MAX_MESSAGE_SIZE: usize = 64 * 1024;
@@ -224,6 +221,21 @@ fn validate_heal_control_response_proof(canonical_response: &[u8], proof: &[u8])
        .map_err(|_| Error::other("peer returned an invalid heal control response proof"))
 }

+fn decode_remote_version_state_capability(expected_member: &str, result: &[u8]) -> Result<Uuid> {
+    let (topology_member, process_epoch) = rustfs_protos::decode_remote_version_state_capability(result).map_err(Error::other)?;
+    if topology_member != expected_member {
+        return Err(Error::other(
+            "peer returned a remote version state capability for a different topology member",
+        ));
+    }
+    let server_epoch =
+        Uuid::from_slice(process_epoch).map_err(|_| Error::other("peer returned an invalid remote version state epoch"))?;
+    if server_epoch.is_nil() {
+        return Err(Error::other("peer returned a nil remote version state epoch"));
+    }
+    Ok(server_epoch)
+}
+
 #[derive(Clone, Debug)]
 pub struct PeerLiveEventsBatch {
    pub events: Vec<u8>,
@@ -235,6 +247,7 @@ pub struct PeerLiveEventsBatch {
 pub struct PeerRestClient {
    pub host: XHost,
    pub grid_host: String,
+    topology_member: String,
    offline: Arc<AtomicBool>,
    recovery_running: Arc<AtomicBool>,
 }
@@ -327,9 +340,11 @@ impl PeerRestClient {
    }

    pub fn new(host: XHost, grid_host: String) -> Self {
+        let topology_member = host.to_string();
        Self {
            host,
            grid_host,
+            topology_member,
            offline: Arc::new(AtomicBool::new(false)),
            recovery_running: Arc::new(AtomicBool::new(false)),
        }
@@ -349,7 +364,11 @@ impl PeerRestClient {

            let client = match grid_host {
                Some(grid_host) => match XHost::try_from(peer_host_port.clone()) {
-                    Ok(host) => Some(PeerRestClient::new(host, grid_host)),
+                    Ok(host) => {
+                        let mut client = PeerRestClient::new(host, grid_host);
+                        client.topology_member = peer_host_port.clone();
+                        Some(client)
+                    }
                    Err(err) => {
                        warn!(peer = %peer_host_port, "Xhost parse failed while constructing peer client: {err:?}");
                        None
@@ -392,7 +411,7 @@ impl PeerRestClient {
        (remote, all, remote_topology_hosts)
    }

-    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
        if self.offline.load(Ordering::Acquire) {
            self.mark_offline_and_spawn_recovery();
            return Err(Error::other(format!("peer {} is temporarily offline", self.grid_host)));
@@ -413,7 +432,7 @@ impl PeerRestClient {
        &self,
    ) -> Result<
        rustfs_protos::proto_gen::node_service::heal_control_service_client::HealControlServiceClient<
-            InterceptedService<Channel, TonicInterceptor>,
+            InterceptedService<AuthenticatedChannel, TonicInterceptor>,
        >,
    > {
        if self.offline.load(Ordering::Acquire) {
@@ -434,7 +453,7 @@ impl PeerRestClient {

    async fn get_tier_mutation_control_client(
        &self,
-    ) -> Result<TierMutationControlServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    ) -> Result<TierMutationControlServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
        if self.offline.load(Ordering::Acquire) {
            self.mark_offline_and_spawn_recovery();
            return Err(Error::other(format!("peer {} is temporarily offline", self.grid_host)));
@@ -471,26 +490,21 @@ impl PeerRestClient {
        self.offline.store(false, Ordering::Release);
    }

+    /// Whether this failure means the peer is unreachable, so it should be
+    /// gated offline and its connection evicted.
+    ///
+    /// RPC failures are classified by their typed gRPC code first
+    /// (`is_network_like_status`); an application error from a live peer must
+    /// never take it offline no matter what its message says. The substring
+    /// fallback only covers failures that exist purely as text, such as the
+    /// dial errors `get_client` wraps.
    fn is_network_like_error(err: &Error) -> bool {
-        let message = err.to_string().to_ascii_lowercase();
-        [
-            "temporarily offline",
-            "transport error",
-            "unavailable",
-            "error trying to connect",
-            "connection refused",
-            "connection reset",
-            "broken pipe",
-            "not connected",
-            "unexpected eof",
-            "timed out",
-            "deadline has elapsed",
-            "connection closed",
-            "connection aborted",
-            "tcp connect error",
-        ]
-        .iter()
-        .any(|needle| message.contains(needle))
+        if let Error::Io(io_err) = err
+            && let Some(status) = embedded_tonic_status(io_err)
+        {
+            return is_network_like_status(status);
+        }
+        message_has_network_needle(&err.to_string())
    }

    fn mark_offline_and_spawn_recovery(&self) {
@@ -1161,14 +1175,24 @@ impl PeerRestClient {
        validate_heal_control_capability_proof(&canonical_ack, &proof)
    }

+    pub async fn probe_remote_version_state(&self, topology_fingerprint: String) -> Result<(String, Uuid)> {
+        let probe = rustfs_protos::remote_version_state_capability_probe(Uuid::new_v4().as_bytes());
+        let result = self
+            .heal_control(rustfs_protos::HEAL_CONTROL_PROTOCOL_VERSION, topology_fingerprint, probe)
+            .await?;
+        let epoch = decode_remote_version_state_capability(&self.topology_member, &result)?;
+        Ok((self.topology_member.clone(), epoch))
+    }
+
    pub async fn load_bucket_metadata(&self, bucket: &str, scanner_maintenance_change: bool) -> Result<()> {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(LoadBucketMetadataRequest {
+                let mut request = Request::new(LoadBucketMetadataRequest {
                    bucket: bucket.to_string(),
                    scanner_maintenance_change,
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.load_bucket_metadata(request).await?.into_inner();
                if !response.success {
@@ -1188,9 +1212,10 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(DeleteBucketMetadataRequest {
+                let mut request = Request::new(DeleteBucketMetadataRequest {
                    bucket: bucket.to_string(),
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.delete_bucket_metadata(request).await?.into_inner();
                if !response.success {
@@ -1210,9 +1235,10 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(DeletePolicyRequest {
+                let mut request = Request::new(DeletePolicyRequest {
                    policy_name: policy.to_string(),
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.delete_policy(request).await?.into_inner();
                if !response.success {
@@ -1232,9 +1258,10 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(LoadPolicyRequest {
+                let mut request = Request::new(LoadPolicyRequest {
                    policy_name: policy.to_string(),
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.load_policy(request).await?.into_inner();
                if !response.success {
@@ -1254,11 +1281,12 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(LoadPolicyMappingRequest {
+                let mut request = Request::new(LoadPolicyMappingRequest {
                    user_or_group: user_or_group.to_string(),
                    user_type,
                    is_group,
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.load_policy_mapping(request).await?.into_inner();
                if !response.success {
@@ -1278,9 +1306,10 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(DeleteUserRequest {
+                let mut request = Request::new(DeleteUserRequest {
                    access_key: access_key.to_string(),
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.delete_user(request).await?.into_inner();
                if !response.success {
@@ -1300,9 +1329,10 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(DeleteServiceAccountRequest {
+                let mut request = Request::new(DeleteServiceAccountRequest {
                    access_key: access_key.to_string(),
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.delete_service_account(request).await?.into_inner();
                if !response.success {
@@ -1322,10 +1352,11 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(LoadUserRequest {
+                let mut request = Request::new(LoadUserRequest {
                    access_key: access_key.to_string(),
                    temp,
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.load_user(request).await?.into_inner();
                if !response.success {
@@ -1345,9 +1376,10 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(LoadServiceAccountRequest {
+                let mut request = Request::new(LoadServiceAccountRequest {
                    access_key: access_key.to_string(),
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.load_service_account(request).await?.into_inner();
                if !response.success {
@@ -1367,9 +1399,10 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(LoadGroupRequest {
+                let mut request = Request::new(LoadGroupRequest {
                    group: group.to_string(),
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.load_group(request).await?.into_inner();
                if !response.success {
@@ -1389,7 +1422,8 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(ReloadSiteReplicationConfigRequest {});
+                let mut request = Request::new(ReloadSiteReplicationConfigRequest {});
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.reload_site_replication_config(request).await?.into_inner();
                if !response.success {
@@ -1413,9 +1447,10 @@ impl PeerRestClient {
                vars.insert(PEER_RESTSIGNAL.to_string(), sig.to_string());
                vars.insert(PEER_RESTSUB_SYS.to_string(), sub_sys.to_string());
                vars.insert(PEER_RESTDRY_RUN.to_string(), dry_run.to_string());
-                let request = Request::new(SignalServiceRequest {
+                let mut request = Request::new(SignalServiceRequest {
                    vars: Some(Mss { value: vars }),
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.signal_service(request).await?.into_inner();
                if !response.success {
@@ -1484,7 +1519,8 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(ReloadPoolMetaRequest {});
+                let mut request = Request::new(ReloadPoolMetaRequest {});
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.reload_pool_meta(request).await?.into_inner();
                if !response.success {
@@ -1505,9 +1541,10 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(StopRebalanceRequest {
+                let mut request = Request::new(StopRebalanceRequest {
                    expected_rebalance_id: expected_rebalance_id.unwrap_or_default().to_string(),
                });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.stop_rebalance(request).await?.into_inner();
                if !response.success {
@@ -1528,7 +1565,8 @@ impl PeerRestClient {
        self.finalize_result(
            async {
                let mut client = self.get_client().await?;
-                let request = Request::new(LoadRebalanceMetaRequest { start_rebalance });
+                let mut request = Request::new(LoadRebalanceMetaRequest { start_rebalance });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.load_rebalance_meta(request).await?.into_inner();

@@ -1567,7 +1605,8 @@ impl PeerRestClient {
                    })
                    .collect::<Result<Vec<_>>>()?;
                let mut client = self.get_client().await?;
-                let request = Request::new(StartDecommissionRequest { pool_indices });
+                let mut request = Request::new(StartDecommissionRequest { pool_indices });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.start_decommission(request).await?.into_inner();
                if !response.success {
@@ -1590,7 +1629,8 @@ impl PeerRestClient {
                let pool_index = u32::try_from(pool_index)
                    .map_err(|_| Error::other(format!("decommission pool index {pool_index} exceeds RPC range")))?;
                let mut client = self.get_client().await?;
-                let request = Request::new(CancelDecommissionRequest { pool_index });
+                let mut request = Request::new(CancelDecommissionRequest { pool_index });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.cancel_decommission(request).await?.into_inner();
                if !response.success {
@@ -1613,7 +1653,8 @@ impl PeerRestClient {
                let pool_index = u32::try_from(pool_index)
                    .map_err(|_| Error::other(format!("decommission pool index {pool_index} exceeds RPC range")))?;
                let mut client = self.get_client().await?;
-                let request = Request::new(ClearDecommissionRequest { pool_index });
+                let mut request = Request::new(ClearDecommissionRequest { pool_index });
+                set_tonic_mutation_body_digest(&mut request)?;

                let response = client.clear_decommission(request).await?.into_inner();
                if !response.success {
@@ -1633,10 +1674,14 @@ impl PeerRestClient {
    pub async fn load_transition_tier_config(&self) -> Result<()> {
        match self.load_transition_tier_config_outcome().await {
            TierConfigReloadOutcome::Success => Ok(()),
-            TierConfigReloadOutcome::TransientReconnect(err) | TierConfigReloadOutcome::TransientRetrySameChannel(err) => {
-                self.finalize_result(Err(err)).await
-            }
-            TierConfigReloadOutcome::Terminal(err) => Err(err),
+            // Only a reconnect-class failure says anything about the channel.
+            // `finalize_result` marks the peer offline and evicts its connection
+            // whenever the message looks network-like, and a peer that answered
+            // and rejected the apply can easily report one ("release RPC failed:
+            // transport error"). Routing those through here would gate a healthy,
+            // responding peer out of every unrelated RPC.
+            TierConfigReloadOutcome::TransientReconnect(err) => self.finalize_result(Err(err)).await,
+            TierConfigReloadOutcome::TransientRetrySameChannel(err) | TierConfigReloadOutcome::Terminal(err) => Err(err),
        }
    }

@@ -1662,6 +1707,9 @@ impl PeerRestClient {
            Err(err) => return tier_config_reload_connection_outcome(err),
        };
        let mut request = Request::new(LoadTransitionTierConfigRequest {});
+        if let Err(err) = set_tonic_mutation_body_digest(&mut request) {
+            return TierConfigReloadOutcome::Terminal(Error::other(err));
+        }
        request.set_timeout(rustfs_protos::heal_control_execution_timeout());

        let response = match client.load_transition_tier_config(request).await {
@@ -1702,39 +1750,37 @@ fn tier_config_reload_connection_outcome(err: Error) -> TierConfigReloadOutcome
 }

 fn is_tier_config_reload_connection_failure(err: &Error) -> bool {
-    let message = err.to_string().to_ascii_lowercase();
+    let message = err.to_string();
+    // A bare "unavailable" is only trusted inside the local dial-failure
+    // wrapper from `get_client`, never in application text.
    if message
+        .to_ascii_lowercase()
        .split_once("can not get client, err:")
        .is_some_and(|(_, local_error)| local_error.contains("unavailable"))
    {
        return true;
    }
-    [
-        "temporarily offline",
-        "transport error",
-        "error trying to connect",
-        "connection refused",
-        "connection reset",
-        "connection closed",
-        "connection aborted",
-        "broken pipe",
-        "not connected",
-        "unexpected eof",
-        "timed out",
-        "deadline has elapsed",
-        "tcp connect error",
-    ]
-    .iter()
-    .any(|needle| message.contains(needle))
+    message_has_network_needle(&message)
 }

+/// Classifies a reload the peer answered but refused to apply.
+///
+/// The peer replied, so the channel is healthy and only the remote apply
+/// failed. Those failures are transient by nature: the reload reads the tier
+/// mutation intents and takes the distributed tier-config lock, both of which
+/// fail while any other node is restarting or while the lock quorum is briefly
+/// disturbed. Retiring the worker on the first such rejection leaves that peer
+/// pinned to the old configuration with nothing left to heal it, so it answers
+/// `TierNotFound` for a tier the rest of the cluster already committed until a
+/// second admin mutation happens to spawn a fresh worker.
+///
+/// Convergence is the whole point of this path, so a rejection is retried on
+/// the same channel. The worker's exponential backoff caps the cost at one
+/// reload every `TIER_CONFIG_RELOAD_RETRY_CAP`, and `Terminal` stays reachable
+/// for transport and gRPC status failures, which is where a genuinely
+/// unrecoverable peer surfaces.
 fn tier_config_reload_remote_failure(error_info: Option<String>) -> TierConfigReloadOutcome {
-    let error_info = error_info.unwrap_or_default();
-    if matches!(error_info.as_str(), "errServerNotInitialized" | "ServerNotInitialized") {
-        TierConfigReloadOutcome::TransientRetrySameChannel(Error::other(error_info))
-    } else {
-        TierConfigReloadOutcome::Terminal(Error::other(error_info))
-    }
+    TierConfigReloadOutcome::TransientRetrySameChannel(Error::other(error_info.unwrap_or_default()))
 }

 fn tier_config_reload_status_outcome(status: tonic::Status) -> TierConfigReloadOutcome {
@@ -1744,6 +1790,14 @@ fn tier_config_reload_status_outcome(status: tonic::Status) -> TierConfigReloadO
        TierConfigReloadOutcome::TransientReconnect(status.into())
    } else if status.code() == Code::Unknown && status.message().starts_with("Service was not ready:") {
        TierConfigReloadOutcome::TransientRetrySameChannel(status.into())
+    } else if status.code() == Code::Unknown
+        && is_tier_config_reload_connection_failure(&Error::other(status.message().to_string()))
+    {
+        // tonic reports a connection dropped mid-call as `Unknown` carrying the
+        // transport error text rather than as `Unavailable`, which is what a peer
+        // restarting under an active mutation produces. Reconnect and retry, so
+        // the restart does not permanently retire this peer's reload worker.
+        TierConfigReloadOutcome::TransientReconnect(status.into())
    } else {
        TierConfigReloadOutcome::Terminal(status.into())
    }
@@ -2151,6 +2205,126 @@ mod tests {
        assert!(!PeerRestClient::is_network_like_error(&Error::NotImplemented));
    }

+    #[test]
+    fn peer_rest_client_network_classifier_uses_typed_status_code() {
+        // The one code that means "nothing is answering on this channel".
+        assert!(PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unavailable(
+            "storage layer is not initialized"
+        ))));
+        // Application statuses from a live peer must not mark it offline,
+        // even when their message contains transport-sounding words.
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::internal(
+            "failed to reload metadata for bucket \"unavailable-logs\""
+        ))));
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unauthenticated(
+            "No valid auth token"
+        ))));
+        // A request-budget expiry answered by a live peer is an application
+        // outcome, not a transport failure.
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::deadline_exceeded(
+            "heal control request expired"
+        ))));
+        // Unknown is the transport's escape hatch for a cause it could not
+        // map, and our handlers never return it, so there the text decides.
+        assert!(PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unknown(
+            "Service was not ready: transport error"
+        ))));
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unknown(
+            "peer response unknown"
+        ))));
+    }
+
+    #[test]
+    fn peer_rest_client_network_classifier_ignores_transport_words_in_application_statuses() {
+        // The reason classification reads the code rather than the text: a
+        // peer that answers is reachable, even when what it says describes a
+        // connection failure of its own. A handler interpolating a local
+        // io::Error into Status::internal, or relaying trouble with its own
+        // downstream, must not cost us the channel to a healthy peer.
+        for status in [
+            tonic::Status::internal("connection refused while dialing downstream backend"),
+            tonic::Status::internal("write failed: broken pipe"),
+            tonic::Status::unauthenticated("connection reset while validating token"),
+            tonic::Status::failed_precondition("scanner lease timed out"),
+            tonic::Status::deadline_exceeded("heal control request timed out"),
+        ] {
+            let rendered = status.to_string();
+            assert!(
+                !PeerRestClient::is_network_like_error(&Error::from(status)),
+                "an answered application status must not mark the peer offline: {rendered}"
+            );
+        }
+    }
+
+    #[test]
+    fn peer_rest_client_network_classifier_keeps_slow_peers_online() {
+        // The per-RPC channel deadline (RUSTFS_INTERNODE_RPC_TIMEOUT, 30s)
+        // surfaces as Cancelled "Timeout expired" carrying the transport
+        // cause as its source. A peer that is merely slow must stay online:
+        // gating it would spend a full recovery cycle fast-failing every RPC
+        // to a host that is still answering, turning load into a partition.
+        let timeout_status = tonic::Status::cancelled("Timeout expired");
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(timeout_status)));
+
+        let sourced = tonic::Status::from_error(Box::new(std::io::Error::other("Timeout expired")));
+        assert!(
+            std::error::Error::source(&sourced).is_some(),
+            "the transport builds this status through Status::from_error, which attaches the cause"
+        );
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(sourced)));
+    }
+
+    #[test]
+    fn rpc_status_errors_keep_their_rendering_and_hide_peer_metadata() {
+        let err = Error::from(tonic::Status::unavailable("peer gone"));
+        assert_eq!(
+            err.to_string(),
+            "Io error: code: 'The service is currently unavailable', message: \"peer gone\""
+        );
+
+        // tonic's own Debug prints the MetadataMap, i.e. every response header
+        // the peer sent; those must not reach a log through this error.
+        let mut status = tonic::Status::unavailable("peer gone");
+        status
+            .metadata_mut()
+            .insert("authorization", "Bearer secret".parse().expect("valid header value"));
+        let rendered = format!("{:?}", Error::from(status));
+        assert!(!rendered.contains("Bearer secret"), "{rendered}");
+        assert!(!rendered.contains("MetadataMap"), "{rendered}");
+    }
+
+    #[test]
+    fn peer_rest_client_network_classifier_ignores_application_text_containing_unavailable() {
+        // Regression: a bare "unavailable" needle used to match application
+        // strings like these and take a healthy peer offline.
+        assert!(!PeerRestClient::is_network_like_error(&Error::other(
+            "peer replication statistics provider is unavailable"
+        )));
+        assert!(!PeerRestClient::is_network_like_error(&Error::other(
+            "bucket \"unavailable-logs\" not found"
+        )));
+        // Anchored renderings of a flattened Unavailable status still match:
+        // tonic >= 0.14 form ...
+        assert!(PeerRestClient::is_network_like_error(&Error::other(
+            "peer tier mutation commit RPC failed: code: 'The service is currently unavailable', message: \"peer gone\""
+        )));
+        // ... which is only anchored as long as tonic renders Unavailable this
+        // way. A tonic bump that reworded it leaves the typed path correct but
+        // this needle stale, so pin the coupling rather than discover it in a
+        // partition.
+        assert!(
+            tonic::Status::unavailable("peer gone")
+                .to_string()
+                .to_ascii_lowercase()
+                .contains("code: 'the service is currently unavailable'"),
+            "tonic reworded Code::Unavailable; update the anchored needle"
+        );
+        // ... and the tonic <= 0.13 form peers may relay in error_info.
+        assert!(PeerRestClient::is_network_like_error(&Error::other(
+            "peer tier mutation commit RPC failed: status: Unavailable, message: \"peer gone\""
+        )));
+    }
+
    #[test]
    fn tier_config_reload_outcome_keeps_tonic_and_remote_errors_typed() {
        assert!(matches!(
@@ -2177,9 +2351,12 @@ mod tests {
            tier_config_reload_status_outcome(tonic::Status::cancelled("request cancelled")),
            TierConfigReloadOutcome::Terminal(_)
        ));
+        // A peer that answered and then refused the apply is retried rather than
+        // retired: the channel is healthy, so the rejection reflects remote state
+        // that the next attempt can find healed.
        assert!(matches!(
            tier_config_reload_remote_failure(Some("backend unavailable".to_string())),
-            TierConfigReloadOutcome::Terminal(_)
+            TierConfigReloadOutcome::TransientRetrySameChannel(_)
        ));
        assert!(matches!(
            tier_config_reload_remote_failure(Some("errServerNotInitialized".to_string())),
@@ -2193,6 +2370,58 @@ mod tests {
            tier_config_reload_connection_outcome(Error::other("can not get client, err: connection unavailable")),
            TierConfigReloadOutcome::TransientReconnect(_)
        ));
+        // The bare word is trusted only to the right of the dial-failure
+        // prefix, not anywhere in the message.
+        assert!(matches!(
+            tier_config_reload_connection_outcome(Error::other(
+                "bucket unavailable-logs rejected it, then: can not get client, err: some other reason"
+            )),
+            TierConfigReloadOutcome::Terminal(_)
+        ));
+    }
+
+    /// A tier mutation issued while another node restarts must still converge on
+    /// the nodes that stayed up. Those peers answer the reload RPC and reject the
+    /// apply, because reloading reads the tier mutation intents and takes the
+    /// distributed tier-config lock while the lock quorum is still disturbed.
+    /// Classifying those rejections as terminal retired the reload worker on its
+    /// first attempt and pinned the peer to the previous configuration, so it
+    /// served `TierNotFound` for an already-committed tier until an unrelated
+    /// second admin mutation spawned a new worker.
+    #[test]
+    fn tier_config_reload_retries_peers_that_reject_the_apply_mid_restart() {
+        for error_info in [
+            "Lock acquisition timeout for resource '.rustfs.sys/config/tier-config.bin.lock' after 5s",
+            "Resource '.rustfs.sys/config/tier-config.bin.lock' is already locked by node-3",
+            "Internal error: release RPC failed: transport error",
+            "save_config_with_opts: err: PreconditionFailed",
+            "erasure read quorum",
+        ] {
+            assert!(
+                matches!(
+                    tier_config_reload_remote_failure(Some(error_info.to_string())),
+                    TierConfigReloadOutcome::TransientRetrySameChannel(_)
+                ),
+                "a peer that rejected the apply must stay retryable so it converges: {error_info}"
+            );
+        }
+
+        // An absent error message is still a rejection, not a reason to stop.
+        assert!(matches!(
+            tier_config_reload_remote_failure(None),
+            TierConfigReloadOutcome::TransientRetrySameChannel(_)
+        ));
+
+        // tonic surfaces a connection dropped mid-call as `Unknown`, not `Unavailable`.
+        assert!(matches!(
+            tier_config_reload_status_outcome(tonic::Status::unknown("transport error")),
+            TierConfigReloadOutcome::TransientReconnect(_)
+        ));
+        // An `Unknown` that is not transport-shaped stays terminal.
+        assert!(matches!(
+            tier_config_reload_status_outcome(tonic::Status::unknown("peer response unknown")),
+            TierConfigReloadOutcome::Terminal(_)
+        ));
    }

    #[tokio::test]
@@ -2271,6 +2500,22 @@ mod tests {
        }
    }

+    #[test]
+    fn remote_version_state_capability_decoder_fails_closed() {
+        let epoch = Uuid::new_v4();
+        let result = rustfs_protos::encode_remote_version_state_capability("node-a:9000", epoch.as_bytes())
+            .expect("small capability response should encode");
+        assert_eq!(
+            decode_remote_version_state_capability("node-a:9000", &result).expect("valid epoch should decode"),
+            epoch
+        );
+        assert!(decode_remote_version_state_capability("node-b:9000", &result).is_err());
+        assert!(decode_remote_version_state_capability("node-a:9000", &result[..result.len() - 1]).is_err());
+        let nil = rustfs_protos::encode_remote_version_state_capability("node-a:9000", Uuid::nil().as_bytes())
+            .expect("small capability response should encode");
+        assert!(decode_remote_version_state_capability("node-a:9000", &nil).is_err());
+    }
+
    struct TierMutationResponseFixture<'a> {
        version: u32,
        phase: TierMutationRpcPhase,
@@ -2495,6 +2740,39 @@ mod tests {
        assert!(!client.offline.load(Ordering::Acquire));
    }

+    #[tokio::test]
+    async fn peer_rest_client_finalize_result_keeps_online_for_app_errors_mentioning_unavailable() {
+        // Regression: application error text containing "unavailable" (a
+        // remote error_info payload, or a bucket named "unavailable-logs" in
+        // a typed application status) must not take a healthy peer offline.
+        let client = test_peer_client();
+        let err = client
+            .finalize_result::<()>(Err(Error::other("peer replication statistics provider is unavailable")))
+            .await
+            .expect_err("application error should still be returned");
+        assert!(err.to_string().contains("provider is unavailable"));
+        assert!(!client.offline.load(Ordering::Acquire));
+
+        let err = client
+            .finalize_result::<()>(Err(Error::from(tonic::Status::internal(
+                "failed to reload metadata for bucket \"unavailable-logs\"",
+            ))))
+            .await
+            .expect_err("application status should still be returned");
+        assert!(err.to_string().contains("unavailable-logs"));
+        assert!(!client.offline.load(Ordering::Acquire));
+    }
+
+    #[tokio::test]
+    async fn peer_rest_client_finalize_result_marks_offline_for_typed_unavailable_status() {
+        let client = test_peer_client();
+        client
+            .finalize_result::<()>(Err(Error::from(tonic::Status::unavailable("storage layer is not initialized"))))
+            .await
+            .expect_err("network error should still be returned");
+        assert!(client.offline.load(Ordering::Acquire));
+    }
+
    #[tokio::test(flavor = "current_thread")]
    async fn peer_rest_recovery_probe_logs_keep_request_id_span_context() {
        let logs = CapturedLogs::default();
@@ -2509,6 +2787,11 @@ mod tests {
                .with_span_list(true),
        );
        let _guard = tracing::subscriber::set_default(subscriber);
+        // The `recovery-monitor` callsite is shared with the production
+        // `mark_offline_and_spawn_recovery` path that sibling tests exercise from
+        // subscriber-less threads; without this the span can be cached as
+        // `Interest::never()` and silently degrade to `Span::none()`.
+        let _callsite_pin = crate::test_tracing::pin_callsite_interest_for_test();

        let client = test_peer_client();
        let span = tracing::info_span!("request-span", request_id = "req-peer-rest");
@@ -14,8 +14,10 @@

 use crate::bucket::metadata_sys;
 use crate::cluster::rpc::client::{
-    TonicInterceptor, gen_tonic_signature_interceptor, is_network_like_disk_error, node_service_time_out_client,
+    AuthenticatedChannel, TonicInterceptor, gen_tonic_signature_interceptor, is_network_like_disk_error,
+    node_service_time_out_client,
 };
+use crate::cluster::rpc::set_tonic_mutation_body_digest;
 use crate::disk::error::DiskError;
 use crate::disk::error::{Error, Result};
 use crate::disk::error_reduce::{BUCKET_OP_IGNORED_ERRS, is_all_buckets_not_found, reduce_write_quorum_errs};
@@ -39,16 +41,84 @@ use rustfs_protos::proto_gen::node_service::node_service_client::NodeServiceClie
 use rustfs_protos::proto_gen::node_service::{
    DeleteBucketRequest, GetBucketInfoRequest, HealBucketRequest, ListBucketRequest, MakeBucketRequest,
 };
+#[cfg(test)]
+use std::sync::{
+    Mutex as StdMutex,
+    atomic::{AtomicBool, Ordering},
+};
 use std::{collections::HashMap, fmt::Debug, sync::Arc, time::Duration};
+#[cfg(test)]
+use tokio::sync::Notify;
 use tokio::{net::TcpStream, sync::RwLock, time};
 use tokio_util::sync::CancellationToken;
 use tonic::Request;
 use tonic::service::interceptor::InterceptedService;
-use tonic::transport::Channel;
 use tracing::{debug, info, warn};

 type Client = Arc<Box<dyn PeerS3Client>>;

+#[cfg(test)]
+#[derive(Default)]
+pub(crate) struct DeleteBucketEmptyScanBarrier {
+    arrived: AtomicBool,
+    arrived_notify: Notify,
+    released: AtomicBool,
+    release_notify: Notify,
+}
+
+#[cfg(test)]
+impl DeleteBucketEmptyScanBarrier {
+    pub(crate) async fn wait_until_paused(&self) {
+        loop {
+            let notified = self.arrived_notify.notified();
+            if self.arrived.load(Ordering::Acquire) {
+                return;
+            }
+            notified.await;
+        }
+    }
+
+    pub(crate) fn release(&self) {
+        self.released.store(true, Ordering::Release);
+        self.release_notify.notify_waiters();
+    }
+
+    async fn pause(&self) {
+        self.arrived.store(true, Ordering::Release);
+        self.arrived_notify.notify_waiters();
+        loop {
+            let notified = self.release_notify.notified();
+            if self.released.load(Ordering::Acquire) {
+                return;
+            }
+            notified.await;
+        }
+    }
+}
+
+#[cfg(test)]
+static DELETE_BUCKET_EMPTY_SCAN_BARRIER: StdMutex<Option<Arc<DeleteBucketEmptyScanBarrier>>> = StdMutex::new(None);
+
+#[cfg(test)]
+pub(crate) fn install_delete_bucket_empty_scan_barrier() -> Arc<DeleteBucketEmptyScanBarrier> {
+    let barrier = Arc::new(DeleteBucketEmptyScanBarrier::default());
+    *DELETE_BUCKET_EMPTY_SCAN_BARRIER
+        .lock()
+        .expect("empty scan barrier lock should not be poisoned") = Some(barrier.clone());
+    barrier
+}
+
+#[cfg(test)]
+async fn pause_after_delete_bucket_empty_scan() {
+    let barrier = DELETE_BUCKET_EMPTY_SCAN_BARRIER
+        .lock()
+        .expect("empty scan barrier lock should not be poisoned")
+        .take();
+    if let Some(barrier) = barrier {
+        barrier.pause().await;
+    }
+}
+
 #[derive(Clone, Debug)]
 pub struct ScannerBucketListing {
    pub buckets: Vec<BucketInfo>,
@@ -89,6 +159,22 @@ fn reduce_pool_write_quorum_errs(per_pool_errs: &[Option<Error>]) -> Option<Erro
    reduce_write_quorum_errs(per_pool_errs, BUCKET_OP_IGNORED_ERRS, pool_write_quorum(per_pool_errs.len()))
 }

+fn resolve_heal_bucket_mode(opts: &mut HealOpts, pool_errs: &[Option<Error>]) -> Result<()> {
+    if opts.recreate {
+        return Ok(());
+    }
+    if let Some(err) = pool_errs
+        .iter()
+        .flatten()
+        .find(|err| **err != Error::DiskNotFound && **err != Error::VolumeNotFound)
+    {
+        return Err(err.clone());
+    }
+    opts.remove = is_all_buckets_not_found(pool_errs);
+    opts.recreate = !opts.remove;
+    Ok(())
+}
+
 #[async_trait]
 pub trait PeerS3Client: Debug + Sync + Send + 'static {
    async fn heal_bucket(&self, bucket: &str, opts: &HealOpts) -> Result<HealResultItem>;
@@ -159,10 +245,7 @@ impl S3PeerSys {
            pool_errs.push(reduce_pool_write_quorum_errs(&per_pool_errs));
        }

-        if !opts.recreate {
-            opts.remove = is_all_buckets_not_found(&pool_errs);
-            opts.recreate = !opts.remove;
-        }
+        resolve_heal_bucket_mode(&mut opts, &pool_errs)?;

        let mut futures = Vec::new();
        let heal_bucket_results = Arc::new(RwLock::new(vec![HealResultItem::default(); self.clients.len()]));
@@ -636,24 +719,22 @@ impl PeerS3Client for LocalPeerS3Client {
            return Err(Error::ErasureWriteQuorum);
        }

-        let force = if opts.force_if_empty && !opts.force {
+        if opts.force_if_empty && !opts.force {
            for disk in local_disks.iter() {
                if has_xlmeta_files(&disk.path().join(bucket)).await.map_err(Error::Io)? {
                    return Err(Error::VolumeNotEmpty);
                }
            }
-            true
-        } else {
-            opts.force
-        };
+            #[cfg(test)]
+            pause_after_delete_bucket_empty_scan().await;
+        }

        let mut futures = Vec::with_capacity(local_disks.len());

        for disk in local_disks.iter() {
-            // Non-force delete refuses a non-empty bucket (VolumeNotEmpty), which
-            // the recreate loop below turns into BucketNotEmpty; only an explicit
-            // force delete removes recursively (backlog#799 B1).
-            futures.push(disk.delete_volume(bucket, force));
+            // `force_if_empty` is validation-only. Passing it as force would let
+            // a PutObject committed after the scan be removed recursively.
+            futures.push(disk.delete_volume(bucket, opts.force));
        }

        let results = join_all(futures).await;
@@ -716,6 +797,15 @@ pub struct RemotePeerS3Client {
 }

 impl RemotePeerS3Client {
+    fn encode_delete_bucket_options(opts: &DeleteBucketOptions) -> Result<String> {
+        let mut remote_opts = opts.clone();
+        // Older peers promote `force_if_empty` to recursive force after their
+        // metadata scan. Keep this coordinator-only hint off the wire so a
+        // mixed-version delete fails closed on non-empty directory remnants.
+        remote_opts.force_if_empty = false;
+        serde_json::to_string(&remote_opts).map_err(Into::into)
+    }
+
    fn recovery_monitor_span(addr: &str) -> tracing::Span {
        tracing::info_span!(
            "recovery-monitor",
@@ -742,7 +832,7 @@ impl RemotePeerS3Client {
        client
    }

-    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
        node_service_time_out_client(&self.addr, TonicInterceptor::Signature(gen_tonic_signature_interceptor()))
            .await
            .map_err(|err| Error::other(format!("can not get client, err: {err}")))
@@ -917,10 +1007,11 @@ impl PeerS3Client for RemotePeerS3Client {
            || async {
                let options: String = serde_json::to_string(opts)?;
                let mut client = self.get_client().await?;
-                let request = Request::new(HealBucketRequest {
+                let mut request = Request::new(HealBucketRequest {
                    bucket: bucket.to_string(),
                    options,
                });
+                set_tonic_mutation_body_digest(&mut request)?;
                let response = client.heal_bucket(request).await?.into_inner();
                if !response.success {
                    return if let Some(err) = response.error {
@@ -973,10 +1064,11 @@ impl PeerS3Client for RemotePeerS3Client {
            || async {
                let options = serde_json::to_string(opts)?;
                let mut client = self.get_client().await?;
-                let request = Request::new(MakeBucketRequest {
+                let mut request = Request::new(MakeBucketRequest {
                    name: bucket.to_string(),
                    options,
                });
+                set_tonic_mutation_body_digest(&mut request)?;
                let response = client.make_bucket(request).await?.into_inner();

                if !response.success {
@@ -1024,13 +1116,14 @@ impl PeerS3Client for RemotePeerS3Client {
    async fn delete_bucket(&self, bucket: &str, opts: &DeleteBucketOptions) -> Result<()> {
        self.execute_with_timeout(
            || async {
-                let options = serde_json::to_string(opts)?;
+                let options = Self::encode_delete_bucket_options(opts)?;
                let mut client = self.get_client().await?;

-                let request = Request::new(DeleteBucketRequest {
+                let mut request = Request::new(DeleteBucketRequest {
                    bucket: bucket.to_string(),
                    options,
                });
+                set_tonic_mutation_body_digest(&mut request)?;
                let response = client.delete_bucket(request).await?.into_inner();
                if !response.success {
                    return if let Some(err) = response.error {
@@ -1397,6 +1490,49 @@ mod tests {
        }
    }

+    #[test]
+    fn remote_delete_bucket_options_fail_closed_for_legacy_peers() {
+        let encoded = RemotePeerS3Client::encode_delete_bucket_options(&DeleteBucketOptions {
+            no_lock: true,
+            no_recreate: true,
+            force_if_empty: true,
+            ..Default::default()
+        })
+        .expect("remote delete options should serialize");
+        let legacy_opts: DeleteBucketOptions =
+            serde_json::from_str(&encoded).expect("legacy peer should decode remote delete options");
+
+        assert!(legacy_opts.no_lock);
+        assert!(legacy_opts.no_recreate);
+        assert!(!legacy_opts.force);
+        assert!(!legacy_opts.force_if_empty);
+
+        let legacy_recursive_force = if legacy_opts.force_if_empty && !legacy_opts.force {
+            true
+        } else {
+            legacy_opts.force
+        };
+        assert!(
+            !legacy_recursive_force,
+            "legacy peer must not upgrade empty-only delete to recursive force"
+        );
+    }
+
+    #[test]
+    fn remote_delete_bucket_options_preserve_explicit_force() {
+        let encoded = RemotePeerS3Client::encode_delete_bucket_options(&DeleteBucketOptions {
+            force: true,
+            force_if_empty: true,
+            ..Default::default()
+        })
+        .expect("remote force-delete options should serialize");
+        let remote_opts: DeleteBucketOptions =
+            serde_json::from_str(&encoded).expect("remote peer should decode force-delete options");
+
+        assert!(remote_opts.force);
+        assert!(!remote_opts.force_if_empty);
+    }
+
    #[tokio::test]
    async fn test_execute_with_timeout_marks_remote_peer_faulty_on_network_like_error() {
        let client = test_remote_peer("http://peer-network-error:9000");
@@ -1554,6 +1690,54 @@ mod tests {
        reset_local_disk_test_state().await;
    }

+    #[tokio::test]
+    #[serial]
+    async fn local_peer_force_if_empty_preserves_unclassified_file_in_selected_pool() {
+        reset_local_disk_test_state().await;
+
+        let temp_dir = TempDir::new().expect("create temp dir for empty-only delete regression");
+        let disks = init_test_local_disks_for_pools(
+            &temp_dir,
+            &[(0, 1), (1, 1)],
+            "local-peer-force-if-empty-preserves-unclassified-file",
+        )
+        .await;
+        let bucket = "empty-only-delete-bucket";
+        let marker = "object/commit-marker";
+        let data = bytes::Bytes::from_static(b"committed object data");
+
+        disks[1]
+            .make_volume(bucket)
+            .await
+            .expect("bucket should be created in the selected pool");
+        disks[1]
+            .write_all(bucket, marker, data.clone())
+            .await
+            .expect("unclassified committed file should be written");
+
+        let err = LocalPeerS3Client::new_with_local_disks(None, Some(vec![1]), disks.clone())
+            .delete_bucket(
+                bucket,
+                &DeleteBucketOptions {
+                    force_if_empty: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect_err("empty-only delete must not recursively remove an unclassified file");
+
+        assert_eq!(err, Error::VolumeNotEmpty);
+        assert_eq!(
+            disks[1]
+                .read_all(bucket, marker)
+                .await
+                .expect("unclassified committed file should be preserved"),
+            data
+        );
+
+        reset_local_disk_test_state().await;
+    }
+
    #[tokio::test]
    #[serial]
    async fn heal_bucket_local_recreates_missing_bucket_volumes() {
@@ -1618,6 +1802,30 @@ mod tests {
        assert_eq!(err, Error::VolumeExists);
    }

+    #[test]
+    fn heal_bucket_mode_fails_closed_on_incomplete_topology() {
+        let mut opts = HealOpts::default();
+        assert_eq!(
+            resolve_heal_bucket_mode(&mut opts, &[Some(Error::ErasureWriteQuorum)]),
+            Err(Error::ErasureWriteQuorum)
+        );
+        assert!(!opts.recreate);
+        assert!(!opts.remove);
+    }
+
+    #[test]
+    fn heal_bucket_mode_distinguishes_deleted_and_partial_buckets() {
+        let mut deleted = HealOpts::default();
+        resolve_heal_bucket_mode(&mut deleted, &[Some(Error::VolumeNotFound)]).unwrap();
+        assert!(deleted.remove);
+        assert!(!deleted.recreate);
+
+        let mut partial = HealOpts::default();
+        resolve_heal_bucket_mode(&mut partial, &[None, Some(Error::VolumeNotFound)]).unwrap();
+        assert!(!partial.remove);
+        assert!(partial.recreate);
+    }
+
    #[tokio::test]
    async fn test_make_bucket_reduces_quorum_by_pool_participants() {
        let peer_sys = S3PeerSys {
@@ -13,8 +13,8 @@
 // limitations under the License.

 use crate::cluster::rpc::client::{
-    TonicInterceptor, gen_tonic_signature_interceptor, is_network_like_disk_error, node_service_time_out_client,
-    node_service_time_out_client_for_class, node_service_time_out_client_no_auth,
+    AuthenticatedChannel, TonicInterceptor, gen_tonic_signature_interceptor, is_network_like_disk_error,
+    node_service_time_out_client, node_service_time_out_client_for_class, node_service_time_out_client_no_auth,
 };
 use crate::cluster::rpc::http_auth::set_tonic_canonical_body_digest;
 use crate::cluster::rpc::internode_data_transport::{
@@ -24,8 +24,8 @@ use crate::cluster::rpc::internode_data_transport::{
 use crate::disk::error::{Error, Result};
 use crate::disk::{
    BatchReadVersionReq, BatchReadVersionResp, CheckPartsResp, DeleteOptions, DiskAPI, DiskInfo, DiskInfoOptions, DiskLocation,
-    DiskOption, FileInfoVersions, FileReader, FileWriter, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp,
-    UpdateMetadataOpts, VolumeInfo, WalkDirOptions, batch_read_version_one_by_one,
+    DiskOption, FileInfoVersions, FileReader, FileWriter, PartTransactionAction, ReadMultipleReq, ReadMultipleResp, ReadOptions,
+    RenameDataResp, SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, batch_read_version_one_by_one,
    disk_store::{
        DEFAULT_RUSTFS_DRIVE_ACTIVE_MONITORING, ENV_RUSTFS_DRIVE_ACTIVE_MONITORING, SKIP_IF_SUCCESS_BEFORE,
        get_drive_active_check_interval, get_drive_active_check_timeout, get_drive_disk_info_timeout, get_drive_list_dir_timeout,
@@ -48,9 +48,11 @@ use rustfs_protos::proto_gen::node_service::RenamePartRequest;
 use rustfs_protos::proto_gen::node_service::{
    BatchReadVersionRequest, BatchReadVersionResponse, CheckPartsRequest, DeletePathsRequest, DeleteRequest,
    DeleteVersionRequest, DeleteVersionsRequest, DeleteVolumeRequest, DiskInfoRequest, ListDirRequest, ListVolumesRequest,
-    MakeVolumeRequest, MakeVolumesRequest, ReadAllRequest, ReadMetadataRequest, ReadMultipleRequest, ReadMultipleResponse,
-    ReadPartsRequest, ReadVersionRequest, ReadXlRequest, RenameDataRequest, RenameFileRequest, StatVolumeRequest,
-    UpdateMetadataRequest, VerifyFileRequest, WriteAllRequest, WriteMetadataRequest, node_service_client::NodeServiceClient,
+    MakeVolumeRequest, MakeVolumesRequest, PreparePartTransactionRequest, ReadAllRequest, ReadMetadataRequest,
+    ReadMultipleRequest, ReadMultipleResponse, ReadPartsRequest, ReadVersionRequest, ReadXlRequest, RenameDataRequest,
+    RenameFileRequest, SettlePartTransactionRequest, SnapshotLeaseReleaseRequest, SnapshotLeaseRenewRequest,
+    SnapshotLeaseRequest, SnapshotLeaseResponse, StatVolumeRequest, UpdateMetadataRequest, VerifyFileRequest, WriteAllRequest,
+    WriteMetadataRequest, node_service_client::NodeServiceClient,
 };
 use serde::{Serialize, de::DeserializeOwned};
 use std::{
@@ -69,7 +71,7 @@ use tokio::{
    time::timeout,
 };
 use tokio_util::sync::CancellationToken;
-use tonic::{Code, Request, service::interceptor::InterceptedService, transport::Channel};
+use tonic::{Code, Request, service::interceptor::InterceptedService};
 use tracing::{debug, trace, warn};
 use uuid::Uuid;

@@ -99,6 +101,18 @@ const LOG_COMPONENT_ECSTORE: &str = "ecstore";
 const LOG_SUBSYSTEM_REMOTE_DISK: &str = "remote_disk";
 const EVENT_REMOTE_DISK_HEALTH: &str = "remote_disk_health";
 const EVENT_REMOTE_DISK_RPC: &str = "remote_disk_rpc";
+const SNAPSHOT_LEASE_PROTOCOL_VERSION: u32 = 1;
+pub const REMOTE_SNAPSHOT_LEASE_TTL: Duration = Duration::from_secs(60);
+
+fn snapshot_lease_token_from_response(response: SnapshotLeaseResponse) -> Result<SnapshotLeaseToken> {
+    if !response.success {
+        return Err(response.error.unwrap_or_default().into());
+    }
+    if response.protocol_version != SNAPSHOT_LEASE_PROTOCOL_VERSION {
+        return Err(Error::other("remote snapshot lease protocol is incompatible"));
+    }
+    SnapshotLeaseToken::from_slice(&response.token)
+}

 /// Bind a mutating disk RPC to its canonical body: the digest lands in the request metadata, and
 /// the signing interceptor folds it (plus a replay-protected nonce) into the v2 signature scope
@@ -1069,7 +1083,7 @@ impl RemoteDisk {
        internode_offline_bypass_reason(&self.addr).map(Error::other)
    }

-    async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
        if let Some(err) = self.offline_bypass_error() {
            return Err(err);
        }
@@ -1082,7 +1096,7 @@ impl RemoteDisk {
    /// Routes onto the isolated bulk channel pool so large transfers cannot head-of-line block
    /// lock/health RPCs (grpc-optimization P1). Falls back to the control channel when isolation
    /// is disabled.
-    async fn get_bulk_client(&self) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    async fn get_bulk_client(&self) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
        if let Some(err) = self.offline_bypass_error() {
            return Err(err);
        }
@@ -1783,6 +1797,81 @@ impl DiskAPI for RemoteDisk {
        .await
    }

+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(SnapshotLeaseRequest {
+                    disk: self.endpoint.to_string(),
+                    volume: volume.to_string(),
+                    path: path.to_string(),
+                    ttl_ms: u64::try_from(REMOTE_SNAPSHOT_LEASE_TTL.as_millis())
+                        .map_err(|_| Error::other("snapshot lease TTL cannot be represented"))?,
+                });
+                let canonical_body = rustfs_protos::canonical_snapshot_lease_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "acquire_snapshot_lease")?;
+                let response = client.acquire_snapshot_lease(request).await?.into_inner();
+                snapshot_lease_token_from_response(response)
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(SnapshotLeaseRenewRequest {
+                    disk: self.endpoint.to_string(),
+                    volume: volume.to_string(),
+                    path: path.to_string(),
+                    token: token.as_bytes().to_vec().into(),
+                    ttl_ms: u64::try_from(REMOTE_SNAPSHOT_LEASE_TTL.as_millis())
+                        .map_err(|_| Error::other("snapshot lease TTL cannot be represented"))?,
+                });
+                let canonical_body = rustfs_protos::canonical_snapshot_lease_renew_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "renew_snapshot_lease")?;
+                let response = client.renew_snapshot_lease(request).await?.into_inner();
+                snapshot_lease_token_from_response(response)
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<()> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(SnapshotLeaseReleaseRequest {
+                    disk: self.endpoint.to_string(),
+                    volume: volume.to_string(),
+                    path: path.to_string(),
+                    token: token.as_bytes().to_vec().into(),
+                });
+                let canonical_body = rustfs_protos::canonical_snapshot_lease_release_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "release_snapshot_lease")?;
+                let response = client.release_snapshot_lease(request).await?.into_inner();
+                if !response.success {
+                    return Err(response.error.unwrap_or_default().into());
+                }
+                Ok(())
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
    #[tracing::instrument(level = "trace", skip_all)]
    async fn write_metadata(&self, _org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()> {
        trace!(
@@ -2481,6 +2570,71 @@ impl DiskAPI for RemoteDisk {
        .await
    }

+    #[tracing::instrument(level = "trace", skip_all)]
+    async fn prepare_part_transaction(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        dst_volume: &str,
+        dst_path: &str,
+        meta: Bytes,
+    ) -> Result<()> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(PreparePartTransactionRequest {
+                    disk: self.endpoint.to_string(),
+                    src_volume: src_volume.to_string(),
+                    src_path: src_path.to_string(),
+                    dst_volume: dst_volume.to_string(),
+                    dst_path: dst_path.to_string(),
+                    meta,
+                });
+                let canonical_body = rustfs_protos::canonical_prepare_part_transaction_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "prepare_part_transaction")?;
+
+                let response = client.prepare_part_transaction(request).await?.into_inner();
+                if !response.success {
+                    return Err(response.error.unwrap_or_default().into());
+                }
+                Ok(())
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    #[tracing::instrument(level = "trace", skip_all)]
+    async fn settle_part_transaction(&self, volume: &str, path: &str, action: PartTransactionAction) -> Result<()> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(SettlePartTransactionRequest {
+                    disk: self.endpoint.to_string(),
+                    volume: volume.to_string(),
+                    path: path.to_string(),
+                    rollback: action == PartTransactionAction::Rollback,
+                });
+                let canonical_body = rustfs_protos::canonical_settle_part_transaction_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "settle_part_transaction")?;
+
+                let response = client.settle_part_transaction(request).await?.into_inner();
+                if !response.success {
+                    return Err(response.error.unwrap_or_default().into());
+                }
+                Ok(())
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
    #[tracing::instrument(level = "trace", skip_all)]
    async fn delete(&self, volume: &str, path: &str, opt: DeleteOptions) -> Result<()> {
        trace!(
@@ -2851,6 +3005,7 @@ mod tests {
    use crate::cluster::rpc::internode_data_transport::{InternodeDataTransportCapabilities, TcpHttpInternodeDataTransport};
    use crate::runtime::sources as runtime_sources;
    use serde_json::Value;
+    use serial_test::serial;
    use std::io::{self as std_io, Write};
    use std::pin::Pin;
    use std::sync::{Arc, Mutex, Mutex as StdMutex, Once};
@@ -2864,6 +3019,48 @@ mod tests {

    static INIT: Once = Once::new();

+    // `#[serial(internode_metrics)]` marks every test that observes
+    // `global_internode_metrics()`. Those counters are a process-wide singleton:
+    // some of these tests snapshot a counter, run one decode, and assert on the
+    // delta, while others deliberately record decode errors or call
+    // `reset_internode_metrics_for_test()`. Run concurrently in one process they
+    // corrupt each other's deltas — a sibling's error bumps the "no decode error"
+    // assertion off zero, and a sibling's reset can drive an `after > before`
+    // assertion backwards.
+    //
+    // The marker only takes effect under the `cargo test` fallback; nextest
+    // already isolates each test in its own process, so every test there gets its
+    // own copy of the counters (see `docs/testing/README.md`). Any new test that
+    // reads or mutates the global internode metrics belongs in this group.
+
+    #[test]
+    fn snapshot_lease_response_requires_current_protocol_and_valid_token() {
+        let token = SnapshotLeaseToken::new();
+        let response = SnapshotLeaseResponse {
+            success: true,
+            token: token.as_bytes().to_vec().into(),
+            protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+            error: None,
+        };
+        assert_eq!(snapshot_lease_token_from_response(response).unwrap(), token);
+
+        let incompatible = SnapshotLeaseResponse {
+            success: true,
+            token: token.as_bytes().to_vec().into(),
+            protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION + 1,
+            error: None,
+        };
+        assert!(snapshot_lease_token_from_response(incompatible).is_err());
+
+        let malformed = SnapshotLeaseResponse {
+            success: true,
+            token: Bytes::from_static(b"not-a-uuid"),
+            protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+            error: None,
+        };
+        assert!(snapshot_lease_token_from_response(malformed).is_err());
+    }
+
    #[test]
    fn list_volumes_decode_rejects_a_malformed_entry() {
        let valid = serde_json::to_string(&VolumeInfo {
@@ -3124,6 +3321,7 @@ mod tests {
    }

    #[test]
+    #[serial(internode_metrics)]
    fn read_multiple_response_decode_prefers_msgpack_payloads() {
        let endpoint = sample_remote_endpoint();
        let msgpack_resp = sample_read_multiple_resp("msgpack", b"binary");
@@ -3146,6 +3344,7 @@ mod tests {
    }

    #[test]
+    #[serial(internode_metrics)]
    fn read_multiple_response_decode_falls_back_to_json_payloads() {
        let endpoint = sample_remote_endpoint();
        let json_resp = sample_read_multiple_resp("json", b"fallback");
@@ -3167,6 +3366,7 @@ mod tests {
    }

    #[test]
+    #[serial(internode_metrics)]
    fn rename_data_response_accepts_legacy_json_without_decode_error() {
        crate::cluster::rpc::runtime_sources::reset_internode_metrics_for_test();
        let response = RenameDataResp {
@@ -3318,6 +3518,7 @@ mod tests {
    }

    #[test]
+    #[serial(internode_metrics)]
    fn read_multiple_response_decode_reports_corrupt_msgpack_item() {
        let endpoint = sample_remote_endpoint();
        let response = ReadMultipleResponse {
@@ -3343,6 +3544,7 @@ mod tests {
    }

    #[test]
+    #[serial(internode_metrics)]
    fn read_multiple_response_decode_reports_corrupt_json_item() {
        let endpoint = sample_remote_endpoint();
        let response = ReadMultipleResponse {
@@ -3383,6 +3585,7 @@ mod tests {
    }

    #[test]
+    #[serial(internode_metrics)]
    fn batch_read_version_response_decode_prefers_msgpack_payloads() {
        let endpoint = sample_remote_endpoint();
        let msgpack_resp = sample_batch_read_version_resp(7, "msgpack-object", true);
@@ -3403,6 +3606,7 @@ mod tests {
    }

    #[test]
+    #[serial(internode_metrics)]
    fn batch_read_version_response_rejects_invalid_success_metadata() {
        let endpoint = sample_remote_endpoint();
        let mut response_item = sample_batch_read_version_resp(0, "invalid-object", true);
@@ -3422,6 +3626,7 @@ mod tests {
    }

    #[test]
+    #[serial(internode_metrics)]
    fn batch_read_version_response_decode_reports_corrupt_msgpack_item() {
        let endpoint = sample_remote_endpoint();
        let response = BatchReadVersionResponse {
@@ -3448,6 +3653,7 @@ mod tests {
    }

    #[test]
+    #[serial(internode_metrics)]
    fn batch_read_version_response_decode_reports_corrupt_json_item() {
        let endpoint = sample_remote_endpoint();
        let response = BatchReadVersionResponse {
@@ -4237,6 +4443,7 @@ mod tests {
    }

    #[tokio::test]
+    #[serial(internode_metrics)]
    async fn test_remote_disk_create_file_retries_once_on_retryable_open_write_error() {
        let transport = RetryingOpenWriteInternodeDataTransport::with_steps(vec![
            OpenWriteTestStep::Error(DiskError::from(rustfs_rio::new_test_internode_http_io_error(
@@ -4275,6 +4482,7 @@ mod tests {
    }

    #[tokio::test]
+    #[serial(internode_metrics)]
    async fn test_remote_disk_read_file_stream_retries_once_on_retryable_open_read_error() {
        // A transient reset-by-peer on a shard read during the read-after-write window must be
        // absorbed by one re-dial rather than eroding read quorum (issue #2761).
@@ -5124,6 +5332,11 @@ mod tests {
                .with_span_list(true),
        );
        let _guard = tracing::subscriber::set_default(subscriber);
+        // The `recovery-monitor` span and the monitor's own log events are
+        // production callsites that sibling tests exercise from subscriber-less
+        // threads; without this they can be cached as `Interest::never()` and go
+        // silently missing here.
+        let _callsite_pin = crate::test_tracing::pin_callsite_interest_for_test();

        let endpoint = Endpoint {
            url: url::Url::parse("http://127.0.0.1:59996/data").expect("endpoint URL should parse"),
@@ -5178,6 +5391,11 @@ mod tests {
                .with_span_list(true),
        );
        let _guard = tracing::subscriber::set_default(subscriber);
+        // The `recovery-monitor` span and the monitor's own log events are
+        // production callsites that sibling tests exercise from subscriber-less
+        // threads; without this they can be cached as `Interest::never()` and go
+        // silently missing here.
+        let _callsite_pin = crate::test_tracing::pin_callsite_interest_for_test();

        let addr = "http://127.0.0.1:59997".to_string();
        let endpoint = Endpoint {
@@ -12,7 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

-use crate::cluster::rpc::client::{TonicInterceptor, gen_tonic_signature_interceptor, node_service_time_out_client};
+use crate::cluster::rpc::client::{
+    AuthenticatedChannel, TonicInterceptor, gen_tonic_signature_interceptor, node_service_time_out_client,
+};
+use crate::cluster::rpc::set_tonic_mutation_body_digest;
 use async_trait::async_trait;
 use bytes::Bytes;
 use rustfs_lock::{
@@ -28,7 +31,6 @@ use std::time::Duration;
 use tokio::time::timeout;
 use tonic::Request;
 use tonic::service::interceptor::InterceptedService;
-use tonic::transport::Channel;
 use tracing::{debug, info, warn};

 /// Remote lock client implementation
@@ -77,7 +79,7 @@ impl RemoteClient {
        }
    }

-    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
        // P3-2 offline bypass (now covering the lock path too): fast-fail a peer already marked
        // offline instead of paying the connect timeout, so dsync reaches quorum sooner. Does not
        // change quorum; the self-healing re-probe keeps the peer recoverable.
@@ -313,10 +315,11 @@ impl LockClient for RemoteClient {
        info!("remote acquire_exclusive for {}", request.resource);
        let mut client = self.get_client().await?;
        let resource_summary = request.resource.to_string();
-        let req = Request::new(GenerallyLockRequest {
+        let mut req = Request::new(GenerallyLockRequest {
            args: serde_json::to_string(&request)
                .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?,
        });
+        set_tonic_mutation_body_digest(&mut req)?;

        let resp = match self.execute_rpc("lock", &resource_summary, client.lock(req)).await {
            Ok(resp) => resp.into_inner(),
@@ -347,7 +350,7 @@ impl LockClient for RemoteClient {

        let mut client = self.get_client().await?;
        let resource_summary = Self::summarize_resources(requests);
-        let req = Request::new(BatchGenerallyLockRequest {
+        let mut req = Request::new(BatchGenerallyLockRequest {
            args: requests
                .iter()
                .map(|request| {
@@ -355,6 +358,7 @@ impl LockClient for RemoteClient {
                })
                .collect::<Result<Vec<_>>>()?,
        });
+        set_tonic_mutation_body_digest(&mut req)?;

        let resp = match self
            .execute_rpc("lock_batch", &resource_summary, client.lock_batch(req))
@@ -395,7 +399,8 @@ impl LockClient for RemoteClient {
            .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?;
        let mut client = self.get_client().await?;
        let resource_summary = unlock_request.resource.to_string();
-        let req = Request::new(GenerallyLockRequest { args: request_string });
+        let mut req = Request::new(GenerallyLockRequest { args: request_string });
+        set_tonic_mutation_body_digest(&mut req)?;
        let resp = self
            .execute_rpc("release", &resource_summary, client.un_lock(req))
            .await?
@@ -414,7 +419,7 @@ impl LockClient for RemoteClient {
        let unlock_requests = lock_ids.iter().map(Self::create_unlock_request).collect::<Vec<_>>();
        let mut client = self.get_client().await?;
        let resource_summary = Self::summarize_resources(&unlock_requests);
-        let req = Request::new(BatchGenerallyLockRequest {
+        let mut req = Request::new(BatchGenerallyLockRequest {
            args: unlock_requests
                .iter()
                .map(|request| {
@@ -422,6 +427,7 @@ impl LockClient for RemoteClient {
                })
                .collect::<Result<Vec<_>>>()?,
        });
+        set_tonic_mutation_body_digest(&mut req)?;

        let resp = self
            .execute_rpc("release_batch", &resource_summary, client.un_lock_batch(req))
@@ -440,10 +446,11 @@ impl LockClient for RemoteClient {
        let refresh_request = Self::create_unlock_request(lock_id);
        let mut client = self.get_client().await?;
        let resource_summary = refresh_request.resource.to_string();
-        let req = Request::new(GenerallyLockRequest {
+        let mut req = Request::new(GenerallyLockRequest {
            args: serde_json::to_string(&refresh_request)
                .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?,
        });
+        set_tonic_mutation_body_digest(&mut req)?;
        let resp = self
            .execute_rpc("refresh", &resource_summary, client.refresh(req))
            .await?
@@ -459,10 +466,11 @@ impl LockClient for RemoteClient {
        let force_request = Self::create_unlock_request(lock_id);
        let mut client = self.get_client().await?;
        let resource_summary = force_request.resource.to_string();
-        let req = Request::new(GenerallyLockRequest {
+        let mut req = Request::new(GenerallyLockRequest {
            args: serde_json::to_string(&force_request)
                .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?,
        });
+        set_tonic_mutation_body_digest(&mut req)?;
        let resp = self
            .execute_rpc("force_release", &resource_summary, client.force_un_lock(req))
            .await?
@@ -483,10 +491,11 @@ impl LockClient for RemoteClient {
        let mut client = self.get_client().await?;

        // Try to acquire a very short-lived lock to test availability
-        let req = Request::new(GenerallyLockRequest {
+        let mut req = Request::new(GenerallyLockRequest {
            args: serde_json::to_string(&status_request)
                .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?,
        });
+        set_tonic_mutation_body_digest(&mut req)?;

        // Try exclusive lock first with very short timeout
        let resp = match self.execute_rpc("check_status", &resource_summary, client.lock(req)).await {
@@ -497,10 +506,11 @@ impl LockClient for RemoteClient {
        if resp.success {
            // If we successfully acquired the lock, the resource was free.
            // Immediately release it on a best-effort basis.
-            let release_req = Request::new(GenerallyLockRequest {
+            let mut release_req = Request::new(GenerallyLockRequest {
                args: serde_json::to_string(&status_request)
                    .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?,
            });
+            set_tonic_mutation_body_digest(&mut release_req)?;
            let _ = self
                .execute_rpc("check_status_release", &resource_summary, client.un_lock(release_req))
                .await;
@@ -2543,6 +2543,7 @@ mod tests {
                data_dir: None,
                delete_marker: false,
                transitioned_object: Default::default(),
+                transition_version_state: Default::default(),
                restore_ongoing: false,
                restore_expires: None,
                user_tags: Arc::new(String::new()),
@@ -16,6 +16,7 @@
 use crate::disk::error_reduce::count_errs;
 use crate::error::{Error, Result};
 use crate::layout::set_heal::{formats_to_drives_info, new_heal_format_sets};
+use crate::multipart_listing::paginate_multipart_listing;
 use crate::storage_api_contracts::{
    bucket::{BucketInfo, BucketOperations, BucketOptions, DeleteBucketOptions, MakeBucketOptions},
    list::{StorageListObjectVersionsInfo, StorageListObjectsV2Info, StorageObjectInfoOrErr, StorageWalkOptions},
@@ -49,7 +50,10 @@ use rustfs_filemeta::FileInfo;
 use rustfs_lock::NamespaceLockWrapper;
 use rustfs_madmin::heal_commands::HealResultItem;
 use rustfs_utils::{crc_hash, path::path_join_buf, sip_hash};
-use std::{collections::HashMap, sync::Arc};
+use std::{
+    collections::{HashMap, HashSet},
+    sync::Arc,
+};
 use tokio::sync::RwLock;
 use tokio::sync::broadcast::{Receiver, Sender};
 use tokio::time::Duration;
@@ -63,6 +67,8 @@ type ListObjectVersionsInfo = StorageListObjectVersionsInfo<ObjectInfo>;
 type ObjectInfoOrErr = StorageObjectInfoOrErr<ObjectInfo, Error>;
 type WalkOptions = StorageWalkOptions<fn(&FileInfo) -> bool>;

+const LIST_MULTIPART_SETS_CONCURRENCY: usize = 4;
+
 #[derive(Debug, Clone)]
 pub struct Sets {
    pub id: Uuid,
@@ -799,9 +805,52 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for Sets {
        delimiter: Option<String>,
        max_uploads: usize,
    ) -> Result<ListMultipartsInfo> {
-        self.get_disks_by_key(prefix)
-            .list_multipart_uploads(bucket, prefix, key_marker, upload_id_marker, delimiter, max_uploads)
-            .await
+        let per_set_limit = max_uploads.saturating_add(1);
+        let results = futures::stream::iter(self.disk_set.iter().cloned())
+            .map(|set| {
+                let key_marker = key_marker.clone();
+                let upload_id_marker = upload_id_marker.clone();
+                let delimiter = delimiter.clone();
+                async move {
+                    set.list_multipart_uploads(bucket, prefix, key_marker, upload_id_marker, delimiter, per_set_limit)
+                        .await
+                }
+            })
+            .buffer_unordered(LIST_MULTIPART_SETS_CONCURRENCY)
+            .collect::<Vec<_>>()
+            .await;
+
+        let mut uploads = Vec::new();
+        let mut common_prefixes = HashSet::new();
+        let mut source_truncated = false;
+        for result in results {
+            let page = result?;
+            uploads.extend(page.uploads);
+            common_prefixes.extend(page.common_prefixes);
+            source_truncated |= page.is_truncated;
+        }
+
+        let page = paginate_multipart_listing(
+            uploads,
+            common_prefixes.into_iter().collect(),
+            key_marker.as_deref(),
+            key_marker.as_ref().and(upload_id_marker.as_deref()),
+            max_uploads,
+            source_truncated,
+        );
+
+        Ok(ListMultipartsInfo {
+            key_marker,
+            upload_id_marker,
+            next_key_marker: page.next_key_marker,
+            next_upload_id_marker: page.next_upload_id_marker,
+            max_uploads,
+            is_truncated: page.is_truncated,
+            uploads: page.uploads,
+            common_prefixes: page.common_prefixes,
+            prefix: prefix.to_owned(),
+            delimiter,
+        })
    }
    #[tracing::instrument(skip(self))]
    async fn new_multipart_upload(&self, bucket: &str, object: &str, opts: &ObjectOptions) -> Result<MultipartUploadResult> {
@@ -1111,6 +1160,7 @@ mod tests {
    use crate::layout::endpoints::SetupType;
    use crate::storage_api_contracts::heal::HealOperations as _;
    use crate::storage_api_contracts::list::ListOperations as _;
+    use crate::storage_api_contracts::multipart::MultipartOperations as _;
    use rustfs_lock::client::local::LocalClient;
    use serial_test::serial;

@@ -1248,6 +1298,194 @@ mod tests {
        assert_eq!(result, (Some(3), Some(1), Some(0)));
    }

+    async fn multipart_listing_test_sets() -> (Vec<tempfile::TempDir>, Arc<Sets>) {
+        let format = FormatV3::new(2, 2);
+        let mut temp_dirs = Vec::new();
+        let mut all_endpoints = Vec::new();
+        let mut disk_sets = Vec::new();
+
+        for set_index in 0..2 {
+            let mut endpoints = Vec::new();
+            let mut disks = Vec::new();
+            for disk_index in 0..2 {
+                let temp_dir = tempfile::tempdir().expect("tempdir should be created");
+                let mut endpoint = Endpoint::try_from(temp_dir.path().to_str().expect("tempdir path should be utf8"))
+                    .expect("endpoint should parse");
+                endpoint.set_pool_index(0);
+                endpoint.set_set_index(set_index);
+                endpoint.set_disk_index(disk_index);
+                let disk = new_disk(
+                    &endpoint,
+                    &DiskOption {
+                        cleanup: false,
+                        health_check: false,
+                    },
+                )
+                .await
+                .expect("disk should be created");
+                let mut disk_format = format.clone();
+                disk_format.erasure.this = format.erasure.sets[set_index][disk_index];
+                save_format_file(&Some(disk.clone()), &Some(disk_format))
+                    .await
+                    .expect("format should be saved");
+                temp_dirs.push(temp_dir);
+                all_endpoints.push(endpoint.clone());
+                endpoints.push(endpoint);
+                disks.push(Some(disk));
+            }
+            disk_sets.push(
+                SetDisks::new(
+                    "test-owner".to_string(),
+                    Arc::new(RwLock::new(disks)),
+                    2,
+                    1,
+                    0,
+                    set_index,
+                    endpoints,
+                    format.clone(),
+                    vec![Arc::new(LocalClient::new()), Arc::new(LocalClient::new())],
+                )
+                .await,
+            );
+        }
+
+        let sets = Arc::new(Sets {
+            id: format.id,
+            disk_set: disk_sets,
+            pool_idx: 0,
+            endpoints: PoolEndpoints {
+                legacy: false,
+                set_count: 2,
+                drives_per_set: 2,
+                endpoints: Endpoints::from(all_endpoints),
+                cmd_line: String::new(),
+                platform: String::new(),
+            },
+            format,
+            parity_count: 1,
+            set_count: 2,
+            set_drive_count: 2,
+            default_parity_count: 1,
+            distribution_algo: DistributionAlgoVersion::V1,
+            exit_signal: None,
+            ctx: bootstrap_ctx(),
+        });
+        (temp_dirs, sets)
+    }
+
+    #[tokio::test(flavor = "multi_thread")]
+    #[serial]
+    async fn list_multipart_uploads_merges_all_sets_without_pagination_loss() {
+        let _setup_type_guard = SetupTypeGuard::switch_to(SetupType::Erasure).await;
+        let (_temp_dirs, sets) = multipart_listing_test_sets().await;
+        let bucket = format!("multipart-list-{}", Uuid::new_v4().simple());
+        sets.make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("bucket should be created");
+
+        let mut keys_by_set = [Vec::new(), Vec::new()];
+        for index in 0..100 {
+            let key = format!("logs/{index:03}.bin");
+            let set_index = sets.get_hashed_set_index(&key);
+            if keys_by_set[set_index].len() < 2 {
+                keys_by_set[set_index].push(key);
+            }
+            if keys_by_set.iter().all(|keys| keys.len() == 2) {
+                break;
+            }
+        }
+        assert!(keys_by_set.iter().all(|keys| keys.len() == 2), "test keys must span both sets");
+
+        let repeated_key = keys_by_set[0][0].clone();
+        let mut expected = Vec::new();
+        for key in keys_by_set.iter().flatten() {
+            let upload = sets
+                .new_multipart_upload(&bucket, key, &ObjectOptions::default())
+                .await
+                .expect("multipart upload should be created");
+            expected.push((key.clone(), upload.upload_id));
+        }
+        let second = sets
+            .new_multipart_upload(&bucket, &repeated_key, &ObjectOptions::default())
+            .await
+            .expect("second upload for the same key should be created");
+        expected.push((repeated_key, second.upload_id));
+        expected.sort();
+
+        let mut actual = Vec::new();
+        let mut key_marker = None;
+        let mut upload_id_marker = None;
+        for _ in 0..expected.len() + 1 {
+            let page = sets
+                .list_multipart_uploads(&bucket, "logs/", key_marker.clone(), upload_id_marker.clone(), None, 2)
+                .await
+                .expect("multipart page should list across every set");
+            assert!(page.uploads.len() <= 2);
+            actual.extend(
+                page.uploads
+                    .iter()
+                    .map(|upload| (upload.object.clone(), upload.upload_id.clone())),
+            );
+            if !page.is_truncated {
+                break;
+            }
+            key_marker = page.next_key_marker;
+            upload_id_marker = page.next_upload_id_marker;
+        }
+
+        assert_eq!(actual, expected, "set-level merge must return every upload exactly once");
+        let mut deduped = actual.clone();
+        deduped.dedup();
+        assert_eq!(deduped.len(), actual.len(), "set-level pagination must not duplicate uploads");
+
+        let mut nested_by_set = [None, None];
+        for index in 0..100 {
+            let key = format!("nested/group-{index:03}/file.bin");
+            let set_index = sets.get_hashed_set_index(&key);
+            nested_by_set[set_index].get_or_insert(key);
+            if nested_by_set.iter().all(Option::is_some) {
+                break;
+            }
+        }
+        for key in nested_by_set.iter().flatten() {
+            sets.new_multipart_upload(&bucket, key, &ObjectOptions::default())
+                .await
+                .expect("nested multipart upload should be created");
+        }
+        let mut expected_prefixes = nested_by_set
+            .iter()
+            .flatten()
+            .map(|key| {
+                key.rsplit_once('/')
+                    .expect("nested key should contain a delimiter")
+                    .0
+                    .to_string()
+                    + "/"
+            })
+            .collect::<Vec<_>>();
+        expected_prefixes.sort();
+
+        let first = sets
+            .list_multipart_uploads(&bucket, "nested/", None, None, Some("/".to_string()), 1)
+            .await
+            .expect("first delimiter page should list across every set");
+        assert!(first.is_truncated);
+        assert_eq!(first.common_prefixes, expected_prefixes[..1]);
+        let second = sets
+            .list_multipart_uploads(
+                &bucket,
+                "nested/",
+                first.next_key_marker,
+                first.next_upload_id_marker,
+                Some("/".to_string()),
+                1,
+            )
+            .await
+            .expect("second delimiter page should list across every set");
+        assert!(!second.is_truncated);
+        assert_eq!(second.common_prefixes, expected_prefixes[1..]);
+    }
+
    #[tokio::test(flavor = "multi_thread")]
    #[serial]
    async fn sets_list_objects_v2_lists_objects_within_the_pool() {
@@ -13,9 +13,9 @@
 // limitations under the License.

 use crate::disk::{
-    CheckPartsResp, DeleteOptions, DiskAPI, DiskError, DiskInfo, DiskInfoOptions, DiskLocation, Endpoint, Error,
-    FileInfoVersions, MmapCopyStageMetrics, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp, Result,
-    UpdateMetadataOpts, VolumeInfo, WalkDirOptions,
+    CheckPartsResp, DataDirDeleteStatus, DeleteOptions, DiskAPI, DiskError, DiskInfo, DiskInfoOptions, DiskLocation, Endpoint,
+    Error, FileInfoVersions, MmapCopyStageMetrics, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp, Result,
+    SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions,
    health_state::{
        RuntimeDriveHealthState, classify_drive_recovery, get_drive_returning_probe_interval,
        get_drive_returning_success_threshold, get_drive_suspect_failure_threshold, record_drive_offline_duration,
@@ -35,7 +35,7 @@ use std::{
        Arc,
        atomic::{AtomicI64, AtomicU32, AtomicU64, Ordering},
    },
-    time::Duration,
+    time::{Duration, SystemTime, UNIX_EPOCH},
 };
 use tokio::{sync::RwLock, time};
 use tokio_util::sync::CancellationToken;
@@ -339,21 +339,19 @@ impl Drop for DiskHealthWaitingGuard<'_> {
 impl DiskHealthTracker {
    /// Create a new disk health tracker
    pub fn new() -> Self {
-        let now = std::time::SystemTime::now()
-            .duration_since(std::time::UNIX_EPOCH)
-            .unwrap()
-            .as_nanos() as i64;
+        let now = current_unix_time();
+        let now_nanos = unix_nanos(now);

        Self {
-            last_success: AtomicI64::new(now),
-            last_started: AtomicI64::new(now),
+            last_success: AtomicI64::new(now_nanos),
+            last_started: AtomicI64::new(now_nanos),
            status: AtomicU32::new(DISK_HEALTH_OK),
            waiting: AtomicU32::new(0),
            runtime_state: AtomicU32::new(RuntimeDriveHealthState::Online as u32),
            consecutive_failures: AtomicU32::new(0),
            consecutive_successes: AtomicU32::new(0),
            offline_since_unix_secs: AtomicI64::new(0),
-            last_transition_unix_secs: AtomicI64::new(now / 1_000_000_000),
+            last_transition_unix_secs: AtomicI64::new(unix_secs_i64(now)),
            last_capacity_total: AtomicU64::new(0),
            last_capacity_used: AtomicU64::new(0),
            last_capacity_free: AtomicU64::new(0),
@@ -363,11 +361,7 @@ impl DiskHealthTracker {

    /// Log a successful operation
    pub fn log_success(&self) {
-        let now = std::time::SystemTime::now()
-            .duration_since(std::time::UNIX_EPOCH)
-            .unwrap()
-            .as_nanos() as i64;
-        self.last_success.store(now, Ordering::Relaxed);
+        self.last_success.store(current_unix_nanos(), Ordering::Relaxed);
    }

    pub fn record_capacity_probe(&self, total: u64, used: u64, free: u64) {
@@ -429,11 +423,14 @@ impl DiskHealthTracker {
    }

    pub fn offline_duration(&self) -> Option<Duration> {
+        self.offline_duration_at(current_unix_secs())
+    }
+
+    fn offline_duration_at(&self, now: u64) -> Option<Duration> {
        let offline_since = self.offline_since_unix_secs.load(Ordering::Acquire);
        if offline_since <= 0 {
            return None;
        }
-        let now = current_unix_secs();
        Some(Duration::from_secs(now.saturating_sub(offline_since as u64)))
    }

@@ -492,6 +489,12 @@ impl DiskHealthTracker {
    /// Remote disks are marked faulty on timeout/network errors; the init loop retries with the
    /// same [`DiskStore`] handles, which would otherwise fail immediately at `is_faulty()`.
    pub fn reset_for_store_init_retry(&self, endpoint: &Endpoint) {
+        self.reset_for_store_init_retry_at(endpoint, current_unix_time());
+    }
+
+    fn reset_for_store_init_retry_at(&self, endpoint: &Endpoint, now: Duration) {
+        let now_nanos = unix_nanos(now);
+        let now_secs = unix_secs_i64(now);
        self.status.store(DISK_HEALTH_OK, Ordering::Release);
        self.runtime_state
            .store(RuntimeDriveHealthState::Online as u32, Ordering::Release);
@@ -499,11 +502,9 @@ impl DiskHealthTracker {
        self.consecutive_successes.store(0, Ordering::Release);
        self.offline_since_unix_secs.store(0, Ordering::Release);
        self.waiting.store(0, Ordering::Release);
-        let now = std::time::SystemTime::now().duration_since(std::time::UNIX_EPOCH).unwrap();
-        let now_nanos = now.as_nanos() as i64;
        self.last_success.store(now_nanos, Ordering::Relaxed);
        self.last_started.store(now_nanos, Ordering::Relaxed);
-        self.last_transition_unix_secs.store(now.as_secs() as i64, Ordering::Release);
+        self.last_transition_unix_secs.store(now_secs, Ordering::Release);
        record_drive_runtime_state(endpoint, RuntimeDriveHealthState::Online);
    }

@@ -612,10 +613,33 @@ impl DiskHealthTracker {
 }

 fn current_unix_secs() -> u64 {
-    std::time::SystemTime::now()
-        .duration_since(std::time::UNIX_EPOCH)
-        .unwrap()
-        .as_secs()
+    // Zero is reserved as "not recorded" by health timestamp atomics.
+    current_unix_time().as_secs().max(1)
+}
+
+fn current_unix_nanos() -> i64 {
+    unix_nanos(current_unix_time())
+}
+
+fn current_unix_time() -> Duration {
+    unix_time_since_epoch(SystemTime::now())
+}
+
+fn unix_time_since_epoch(time: SystemTime) -> Duration {
+    time.duration_since(UNIX_EPOCH).unwrap_or(Duration::ZERO)
+}
+
+fn unix_nanos(time: Duration) -> i64 {
+    i64::try_from(time.as_nanos()).unwrap_or(i64::MAX)
+}
+
+fn unix_secs_i64(time: Duration) -> i64 {
+    i64::try_from(time.as_secs()).unwrap_or(i64::MAX)
+}
+
+fn elapsed_since(last_nanos: i64, now_nanos: i64) -> Duration {
+    let elapsed_nanos = now_nanos.saturating_sub(last_nanos).max(0);
+    Duration::from_nanos(u64::try_from(elapsed_nanos).unwrap_or(u64::MAX))
 }

 impl Default for DiskHealthTracker {
@@ -635,11 +659,7 @@ struct HealthDiskCtxValue {

 impl HealthDiskCtxValue {
    fn log_success(&self) {
-        let now = std::time::SystemTime::now()
-            .duration_since(std::time::UNIX_EPOCH)
-            .unwrap()
-            .as_nanos() as i64;
-        self.last_success.store(now, Ordering::Relaxed);
+        self.last_success.store(current_unix_nanos(), Ordering::Relaxed);
    }
 }

@@ -774,12 +794,7 @@ impl LocalDiskWrapper {
                    }

                    let last_success_nanos = health.last_success.load(Ordering::Relaxed);
-                    let elapsed = Duration::from_nanos(
-                        (std::time::SystemTime::now()
-                            .duration_since(std::time::UNIX_EPOCH)
-                            .unwrap()
-                            .as_nanos() as i64 - last_success_nanos) as u64
-                    );
+                    let elapsed = elapsed_since(last_success_nanos, current_unix_nanos());

                    if elapsed < SKIP_IF_SUCCESS_BEFORE {
                        continue;
@@ -1091,11 +1106,7 @@ impl LocalDiskWrapper {
        self.check_disk_stale().await?;

        // Record operation start
-        let now = std::time::SystemTime::now()
-            .duration_since(std::time::UNIX_EPOCH)
-            .unwrap()
-            .as_nanos() as i64;
-        self.health.last_started.store(now, Ordering::Relaxed);
+        self.health.last_started.store(current_unix_nanos(), Ordering::Relaxed);
        let _waiting_guard = self.health.waiting_guard();

        if timeout_duration == Duration::ZERO {
@@ -1317,11 +1328,7 @@ impl DiskAPI for LocalDiskWrapper {
        }

        // Record operation start
-        let now = std::time::SystemTime::now()
-            .duration_since(std::time::UNIX_EPOCH)
-            .unwrap()
-            .as_nanos() as i64;
-        self.health.last_started.store(now, Ordering::Relaxed);
+        self.health.last_started.store(current_unix_nanos(), Ordering::Relaxed);
        self.health.increment_waiting();

        // Execute the operation
@@ -1342,6 +1349,38 @@ impl DiskAPI for LocalDiskWrapper {
            .await
    }

+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
+        self.track_disk_health(
+            || async { self.disk.acquire_snapshot_lease(volume, path).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<()> {
+        self.track_disk_health(
+            || async { self.disk.release_snapshot_lease(volume, path, token).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        self.track_disk_health(
+            || async { self.disk.renew_snapshot_lease(volume, path, token).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
+        self.track_disk_health(
+            || async { self.disk.delete_data_dir(volume, path, opts).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
    async fn write_metadata(&self, org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()> {
        self.track_disk_health(
            || async { self.disk.write_metadata(org_volume, volume, path, fi).await },
@@ -1473,6 +1512,33 @@ impl DiskAPI for LocalDiskWrapper {
        .await
    }

+    async fn prepare_part_transaction(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        dst_volume: &str,
+        dst_path: &str,
+        meta: Bytes,
+    ) -> Result<()> {
+        self.track_disk_health(
+            || async {
+                self.disk
+                    .prepare_part_transaction(src_volume, src_path, dst_volume, dst_path, meta)
+                    .await
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn settle_part_transaction(&self, volume: &str, path: &str, action: crate::disk::PartTransactionAction) -> Result<()> {
+        self.track_disk_health(
+            || async { self.disk.settle_part_transaction(volume, path, action).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
    async fn delete(&self, volume: &str, path: &str, opt: DeleteOptions) -> Result<()> {
        self.track_disk_health(|| async { self.disk.delete(volume, path, opt).await }, get_max_timeout_duration())
            .await
@@ -2243,4 +2309,53 @@ mod tests {
        assert!(health.mark_offline(&endpoint, "again"));
        assert!(health.is_faulty());
    }
+
+    #[test]
+    fn unix_time_clamps_epoch_and_pre_epoch_to_zero() {
+        let before_epoch = UNIX_EPOCH
+            .checked_sub(Duration::from_nanos(1))
+            .expect("one nanosecond before the Unix epoch should be representable");
+
+        assert_eq!(unix_time_since_epoch(UNIX_EPOCH), Duration::ZERO);
+        assert_eq!(unix_time_since_epoch(before_epoch), Duration::ZERO);
+    }
+
+    #[test]
+    fn elapsed_and_offline_duration_saturate_on_clock_rollback() {
+        let health = DiskHealthTracker::new();
+        health.offline_since_unix_secs.store(10, Ordering::Release);
+
+        assert_eq!(elapsed_since(10, 12), Duration::from_nanos(2));
+        assert_eq!(elapsed_since(10, 9), Duration::ZERO);
+        assert_eq!(health.offline_duration_at(9), Some(Duration::ZERO));
+    }
+
+    #[test]
+    fn pre_epoch_retry_reset_updates_the_complete_health_state() {
+        let endpoint = Endpoint::try_from("/tmp/reset-store-init-retry-pre-epoch").expect("endpoint should parse");
+        let health = DiskHealthTracker::new();
+        health.status.store(DISK_HEALTH_FAULTY, Ordering::Release);
+        health
+            .runtime_state
+            .store(RuntimeDriveHealthState::Offline as u32, Ordering::Release);
+        health.consecutive_failures.store(3, Ordering::Release);
+        health.consecutive_successes.store(2, Ordering::Release);
+        health.offline_since_unix_secs.store(11, Ordering::Release);
+        health.waiting.store(4, Ordering::Release);
+        health.last_success.store(12, Ordering::Release);
+        health.last_started.store(13, Ordering::Release);
+        health.last_transition_unix_secs.store(14, Ordering::Release);
+
+        health.reset_for_store_init_retry_at(&endpoint, unix_time_since_epoch(UNIX_EPOCH - Duration::from_nanos(1)));
+
+        assert_eq!(health.status.load(Ordering::Acquire), DISK_HEALTH_OK);
+        assert_eq!(health.runtime_state(), RuntimeDriveHealthState::Online);
+        assert_eq!(health.consecutive_failures.load(Ordering::Acquire), 0);
+        assert_eq!(health.consecutive_successes.load(Ordering::Acquire), 0);
+        assert_eq!(health.offline_since_unix_secs.load(Ordering::Acquire), 0);
+        assert_eq!(health.waiting.load(Ordering::Acquire), 0);
+        assert_eq!(health.last_success.load(Ordering::Acquire), 0);
+        assert_eq!(health.last_started.load(Ordering::Acquire), 0);
+        assert_eq!(health.last_transition_unix_secs.load(Ordering::Acquire), 0);
+    }
 }
@@ -337,9 +337,54 @@ impl From<DiskError> for std::io::Error {
    }
 }

+/// The single in-band representation of a failed internode RPC: it carries the
+/// typed `tonic::Status` so failure classifiers can read the gRPC code instead
+/// of substring-matching the rendered message (see `is_network_like_status`).
+/// Both `DiskError` and `StorageError` wrap statuses in this type, so one
+/// downcast recovers the status regardless of which error the status was
+/// converted into first.
+pub(crate) struct RpcStatusError(tonic::Status);
+
+impl RpcStatusError {
+    pub(crate) fn status(&self) -> &tonic::Status {
+        &self.0
+    }
+}
+
+impl From<tonic::Status> for RpcStatusError {
+    fn from(status: tonic::Status) -> Self {
+        Self(status)
+    }
+}
+
+impl std::fmt::Display for RpcStatusError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        std::fmt::Display::fmt(&self.0, f)
+    }
+}
+
+/// `tonic::Status`'s own `Debug` prints its `MetadataMap`, i.e. every response
+/// header and trailer the peer sent. Those are remote-controlled and can carry
+/// credentials injected by a proxy in front of the peer, so keep them out of
+/// anything that reaches a log.
+impl std::fmt::Debug for RpcStatusError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("RpcStatusError")
+            .field("code", &self.0.code())
+            .field("message", &self.0.message())
+            .finish()
+    }
+}
+
+impl StdError for RpcStatusError {
+    fn source(&self) -> Option<&(dyn StdError + 'static)> {
+        Some(&self.0)
+    }
+}
+
 impl From<tonic::Status> for DiskError {
    fn from(e: tonic::Status) -> Self {
-        DiskError::other(e.message().to_string())
+        DiskError::Io(io::Error::other(RpcStatusError(e)))
    }
 }

@@ -38,6 +38,16 @@ pub const FORMAT_CONFIG_FILE: &str = "format.json";
 pub const HEALING_MARKER_PATH: &str = "healing.bin";
 pub const STORAGE_FORMAT_FILE: &str = "xl.meta";
 pub const STORAGE_FORMAT_FILE_BACKUP: &str = "xl.meta.bkp";
+pub const PART_TRANSACTION_NEW_META: &str = "new.meta";
+pub const PART_TRANSACTION_OLD_META: &str = "old.meta";
+pub const PART_TRANSACTION_ROLLBACK: &str = "rollback";
+
+pub fn part_transaction_path(part_path: &str) -> String {
+    match part_path.rsplit_once('/') {
+        Some((parent, name)) => format!("{parent}/.{name}.rustfs-txn"),
+        None => format!(".{part_path}.rustfs-txn"),
+    }
+}

 use crate::cluster::rpc::RemoteDisk;
 use crate::cluster::rpc::build_internode_data_transport_from_env;
@@ -62,6 +72,45 @@ pub type DiskStore = Arc<Disk>;
 pub type FileReader = Box<dyn AsyncRead + Send + Sync + Unpin>;
 pub type FileWriter = Box<dyn AsyncWrite + Send + Sync + Unpin>;

+#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq, Serialize, Deserialize)]
+pub struct SnapshotLeaseToken(Uuid);
+
+impl SnapshotLeaseToken {
+    pub fn new() -> Self {
+        Self(Uuid::new_v4())
+    }
+
+    pub fn from_slice(bytes: &[u8]) -> Result<Self> {
+        let uuid = Uuid::from_slice(bytes).map_err(|_| Error::other("invalid snapshot lease token"))?;
+        if uuid.is_nil() {
+            return Err(Error::other("invalid snapshot lease token"));
+        }
+        Ok(Self(uuid))
+    }
+
+    pub fn as_bytes(&self) -> &[u8; 16] {
+        self.0.as_bytes()
+    }
+}
+
+impl Default for SnapshotLeaseToken {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum DataDirDeleteStatus {
+    Deleted,
+    Deferred,
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum PartTransactionAction {
+    Commit,
+    Rollback,
+}
+
 #[derive(Clone, Copy, Debug)]
 pub struct MmapCopyStageMetrics {
    pub(crate) path: &'static str,
@@ -233,6 +282,34 @@ impl DiskAPI for Disk {
        }
    }

+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
+        match self {
+            Disk::Local(local_disk) => local_disk.acquire_snapshot_lease(volume, path).await,
+            Disk::Remote(remote_disk) => remote_disk.acquire_snapshot_lease(volume, path).await,
+        }
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<()> {
+        match self {
+            Disk::Local(local_disk) => local_disk.release_snapshot_lease(volume, path, token).await,
+            Disk::Remote(remote_disk) => remote_disk.release_snapshot_lease(volume, path, token).await,
+        }
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        match self {
+            Disk::Local(local_disk) => local_disk.renew_snapshot_lease(volume, path, token).await,
+            Disk::Remote(remote_disk) => remote_disk.renew_snapshot_lease(volume, path, token).await,
+        }
+    }
+
+    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
+        match self {
+            Disk::Local(local_disk) => local_disk.delete_data_dir(volume, path, opts).await,
+            Disk::Remote(remote_disk) => remote_disk.delete_data_dir(volume, path, opts).await,
+        }
+    }
+
    #[tracing::instrument(level = "trace", skip_all)]
    async fn write_metadata(&self, _org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()> {
        match self {
@@ -381,6 +458,35 @@ impl DiskAPI for Disk {
        }
    }

+    async fn prepare_part_transaction(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        dst_volume: &str,
+        dst_path: &str,
+        meta: Bytes,
+    ) -> Result<()> {
+        match self {
+            Disk::Local(local_disk) => {
+                local_disk
+                    .prepare_part_transaction(src_volume, src_path, dst_volume, dst_path, meta)
+                    .await
+            }
+            Disk::Remote(remote_disk) => {
+                remote_disk
+                    .prepare_part_transaction(src_volume, src_path, dst_volume, dst_path, meta)
+                    .await
+            }
+        }
+    }
+
+    async fn settle_part_transaction(&self, volume: &str, path: &str, action: PartTransactionAction) -> Result<()> {
+        match self {
+            Disk::Local(local_disk) => local_disk.settle_part_transaction(volume, path, action).await,
+            Disk::Remote(remote_disk) => remote_disk.settle_part_transaction(volume, path, action).await,
+        }
+    }
+
    #[tracing::instrument(level = "trace", skip_all)]
    async fn delete(&self, volume: &str, path: &str, opt: DeleteOptions) -> Result<()> {
        match self {
@@ -601,6 +707,19 @@ pub trait DiskAPI: Debug + Send + Sync + 'static {
    ) -> Result<()>;
    async fn delete_versions(&self, volume: &str, versions: Vec<FileInfoVersions>, opts: DeleteOptions) -> Vec<Option<Error>>;
    async fn delete_paths(&self, volume: &str, paths: &[String]) -> Result<()>;
+    async fn acquire_snapshot_lease(&self, _volume: &str, _path: &str) -> Result<SnapshotLeaseToken> {
+        Err(Error::other("snapshot leases are not supported by this disk"))
+    }
+    async fn release_snapshot_lease(&self, _volume: &str, _path: &str, _token: SnapshotLeaseToken) -> Result<()> {
+        Err(Error::other("snapshot leases are not supported by this disk"))
+    }
+    async fn renew_snapshot_lease(&self, _volume: &str, _path: &str, _token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        Err(Error::other("snapshot leases are not supported by this disk"))
+    }
+    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
+        self.delete(volume, path, opts).await?;
+        Ok(DataDirDeleteStatus::Deleted)
+    }
    async fn write_metadata(&self, org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()>;
    async fn update_metadata(&self, volume: &str, path: &str, fi: FileInfo, opts: &UpdateMetadataOpts) -> Result<()>;
    async fn read_version(
@@ -659,6 +778,19 @@ pub trait DiskAPI: Debug + Send + Sync + 'static {
    // ReadFileStream
    async fn rename_file(&self, src_volume: &str, src_path: &str, dst_volume: &str, dst_path: &str) -> Result<()>;
    async fn rename_part(&self, src_volume: &str, src_path: &str, dst_volume: &str, dst_path: &str, meta: Bytes) -> Result<()>;
+    async fn prepare_part_transaction(
+        &self,
+        _src_volume: &str,
+        _src_path: &str,
+        _dst_volume: &str,
+        _dst_path: &str,
+        _meta: Bytes,
+    ) -> Result<()> {
+        Err(DiskError::MethodNotAllowed)
+    }
+    async fn settle_part_transaction(&self, _volume: &str, _path: &str, _action: PartTransactionAction) -> Result<()> {
+        Err(DiskError::MethodNotAllowed)
+    }
    async fn delete(&self, volume: &str, path: &str, opt: DeleteOptions) -> Result<()>;
    // VerifyFile
    async fn verify_file(&self, volume: &str, path: &str, fi: &FileInfo) -> Result<CheckPartsResp>;
@@ -25,7 +25,7 @@ use std::{
    sync::{Arc, LazyLock, Weak},
 };
 use tokio::fs;
-use tokio::sync::{OwnedSemaphorePermit, Semaphore, SemaphorePermit};
+use tokio::sync::{OwnedSemaphorePermit, RwLock, Semaphore, SemaphorePermit};
 use tracing::warn;

 /// Check path length according to OS limits.
@@ -123,6 +123,8 @@ const TEST_GLOBAL_FILE_SYNCS: usize = 64;

 static FILE_SYNC_PERMITS: LazyLock<Semaphore> = LazyLock::new(|| Semaphore::new(global_file_sync_limit()));
 static DISK_FILE_SYNC_LIMITERS: LazyLock<Mutex<HashMap<PathBuf, Weak<Semaphore>>>> = LazyLock::new(|| Mutex::new(HashMap::new()));
+static DISK_VOLUME_MUTATION_LOCKS: LazyLock<Mutex<HashMap<PathBuf, Weak<RwLock<()>>>>> =
+    LazyLock::new(|| Mutex::new(HashMap::new()));

 fn default_global_file_sync_limit(cpu_count: usize, max_blocking_threads: usize) -> usize {
    let cpu_scaled = cpu_count
@@ -157,6 +159,24 @@ pub(crate) fn disk_file_sync_limiter(root: &Path) -> Arc<Semaphore> {
    limiter
 }

+/// Serialize a bucket's local metadata commits with physical bucket removal.
+///
+/// The key includes the canonical disk root, so independently reconnected
+/// [`LocalDisk`](super::local::LocalDisk) instances share the same lock while
+/// disconnected disks do not keep the registry alive.
+pub(crate) fn disk_volume_mutation_lock(root: &Path, volume: &str) -> Arc<RwLock<()>> {
+    let key = root.join(volume);
+    let mut locks = DISK_VOLUME_MUTATION_LOCKS.lock();
+    locks.retain(|_, lock| lock.strong_count() > 0);
+    if let Some(lock) = locks.get(&key).and_then(Weak::upgrade) {
+        return lock;
+    }
+
+    let lock = Arc::new(RwLock::new(()));
+    locks.insert(key, Arc::downgrade(&lock));
+    lock
+}
+
 /// Always acquire the per-disk permit before the process-wide permit. Keeping
 /// this order uniform prevents one slow disk from reserving global capacity
 /// while it waits for its own concurrency slot.
@@ -572,15 +592,14 @@ async fn reliable_rename_inner(
    base_dir: impl AsRef<Path>,
    warn_on_missing_source: bool,
 ) -> io::Result<()> {
-    if let Some(parent) = dst_file_path.as_ref().parent()
-        && !file_exists(parent)
-    {
-        reliable_mkdir_all(parent, base_dir.as_ref()).await?;
-    }
+    let parent_guard = match dst_file_path.as_ref().parent() {
+        Some(parent) => Some(mkdir_all_below_existing_base(parent, base_dir.as_ref()).await?),
+        None => None,
+    };

    let mut i = 0;
    loop {
-        if let Err(e) = super::fs::rename_std(src_file_path.as_ref(), dst_file_path.as_ref()) {
+        if let Err(e) = rename_into_existing_parent(src_file_path.as_ref(), dst_file_path.as_ref(), parent_guard.as_ref()) {
            if should_retry_rename(&e, i) {
                i += 1;
                continue;
@@ -597,6 +616,159 @@ async fn reliable_rename_inner(
    Ok(())
 }

+#[cfg(unix)]
+fn rename_into_existing_parent(
+    src_file_path: &Path,
+    dst_file_path: &Path,
+    parent_guard: Option<&ExistingBaseDirectoryGuard>,
+) -> io::Result<()> {
+    use rustix::fs::{Mode, OFlags, open, renameat};
+
+    let Some(parent_guard) = parent_guard else {
+        return super::fs::rename_std(src_file_path, dst_file_path);
+    };
+    let src_parent = src_file_path
+        .parent()
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "rename source must have a parent directory"))?;
+    let src_name = src_file_path
+        .file_name()
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "rename source must have a file name"))?;
+    let dst_name = dst_file_path
+        .file_name()
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "rename destination must have a file name"))?;
+    let src_parent = open(
+        src_parent,
+        OFlags::RDONLY | OFlags::DIRECTORY | OFlags::NOFOLLOW | OFlags::CLOEXEC,
+        Mode::empty(),
+    )
+    .map_err(io::Error::from)?;
+    let dst_parent = parent_guard
+        .last()
+        .ok_or_else(|| io::Error::other("rename destination parent guard is empty"))?;
+
+    renameat(&src_parent, src_name, dst_parent, dst_name).map_err(io::Error::from)
+}
+
+#[cfg(not(unix))]
+fn rename_into_existing_parent(
+    src_file_path: &Path,
+    dst_file_path: &Path,
+    _parent_guard: Option<&ExistingBaseDirectoryGuard>,
+) -> io::Result<()> {
+    super::fs::rename_std(src_file_path, dst_file_path)
+}
+
+async fn mkdir_all_below_existing_base(dir_path: &Path, base_dir: &Path) -> io::Result<ExistingBaseDirectoryGuard> {
+    let dir_path = dir_path.to_path_buf();
+    let base_dir = base_dir.to_path_buf();
+
+    tokio::task::spawn_blocking(move || mkdir_all_below_existing_base_std(&dir_path, &base_dir)).await?
+}
+
+#[cfg(windows)]
+pub(crate) type ExistingBaseDirectoryGuard = Vec<winapi_util::Handle>;
+
+#[cfg(unix)]
+pub(crate) type ExistingBaseDirectoryGuard = Vec<std::os::fd::OwnedFd>;
+
+#[cfg(all(not(unix), not(windows)))]
+pub(crate) type ExistingBaseDirectoryGuard = ();
+
+#[cfg(windows)]
+fn lock_windows_directory(path: &Path) -> io::Result<winapi_util::Handle> {
+    use std::os::windows::fs::OpenOptionsExt;
+
+    const FILE_ATTRIBUTE_DIRECTORY: u64 = 0x10;
+    const FILE_ATTRIBUTE_REPARSE_POINT: u32 = 0x400;
+    const FILE_FLAG_BACKUP_SEMANTICS: u32 = 0x0200_0000;
+    const FILE_FLAG_OPEN_REPARSE_POINT: u32 = 0x0020_0000;
+    const FILE_SHARE_READ: u32 = 0x1;
+
+    let file = std::fs::OpenOptions::new()
+        .read(true)
+        .share_mode(FILE_SHARE_READ)
+        .custom_flags(FILE_FLAG_BACKUP_SEMANTICS | FILE_FLAG_OPEN_REPARSE_POINT)
+        .open(path)?;
+    let handle = winapi_util::Handle::from_file(file);
+    let info = winapi_util::file::information(&handle)?;
+    if info.file_attributes() & FILE_ATTRIBUTE_DIRECTORY == 0
+        || info.file_attributes() & u64::from(FILE_ATTRIBUTE_REPARSE_POINT) != 0
+    {
+        return Err(io::Error::from(io::ErrorKind::NotADirectory));
+    }
+    Ok(handle)
+}
+
+pub(crate) fn mkdir_all_below_existing_base_std(dir_path: &Path, base_dir: &Path) -> io::Result<ExistingBaseDirectoryGuard> {
+    let relative = dir_path
+        .strip_prefix(base_dir)
+        .map_err(|_| io::Error::new(io::ErrorKind::InvalidInput, "rename destination must remain below its base directory"))?;
+    for component in relative.components() {
+        if !matches!(component, Component::Normal(_) | Component::CurDir) {
+            return Err(io::Error::new(
+                io::ErrorKind::InvalidInput,
+                "rename destination contains an invalid path component",
+            ));
+        }
+    }
+
+    #[cfg(unix)]
+    {
+        use rustix::fs::{Mode, OFlags, mkdirat, open, openat};
+        use rustix::io::Errno;
+
+        let flags = OFlags::RDONLY | OFlags::DIRECTORY | OFlags::NOFOLLOW | OFlags::CLOEXEC;
+        let mode = Mode::RWXU | Mode::RWXG | Mode::RWXO;
+        let mut parents = vec![open(base_dir, flags, Mode::empty()).map_err(io::Error::from)?];
+
+        for component in relative.components() {
+            let Component::Normal(component) = component else {
+                continue;
+            };
+            let parent = parents
+                .last()
+                .expect("base directory guard should contain the base directory");
+            match mkdirat(parent, component, mode) {
+                Ok(()) => {}
+                Err(Errno::EXIST) => {}
+                Err(err) => return Err(err.into()),
+            }
+            parents.push(openat(parent, component, flags, Mode::empty()).map_err(io::Error::from)?);
+        }
+
+        Ok(parents)
+    }
+
+    #[cfg(windows)]
+    {
+        let mut handles = vec![lock_windows_directory(base_dir)?];
+        let mut current = base_dir.to_path_buf();
+        for component in relative.components() {
+            let Component::Normal(component) = component else {
+                continue;
+            };
+            current.push(component);
+            match std::fs::create_dir(&current) {
+                Ok(()) => {}
+                Err(err) if err.kind() == io::ErrorKind::AlreadyExists => {}
+                Err(err) => return Err(err),
+            }
+            handles.push(lock_windows_directory(&current)?);
+        }
+
+        Ok(handles)
+    }
+
+    #[cfg(all(not(unix), not(windows)))]
+    {
+        let _ = relative;
+        Err(io::Error::new(
+            io::ErrorKind::Unsupported,
+            "safe recursive directory creation is unavailable on this platform",
+        ))
+    }
+}
+
 fn warn_reliable_rename_failure(src_file_path: &Path, dst_file_path: &Path, base_dir: &Path, err: &io::Error) {
    warn!(
        "reliable_rename failed. src_file_path: {:?}, dst_file_path: {:?}, base_dir: {:?}, err: {:?}",
@@ -727,6 +899,20 @@ mod tests {
        Arc::new(Semaphore::new(MAX_PARALLEL_FILE_SYNCS))
    }

+    #[tokio::test]
+    async fn disk_volume_mutation_lock_is_shared_per_root_and_volume() {
+        let temp_dir = tempdir().expect("create temp dir");
+        let first = disk_volume_mutation_lock(temp_dir.path(), "bucket");
+        let second = disk_volume_mutation_lock(temp_dir.path(), "bucket");
+        let other = disk_volume_mutation_lock(temp_dir.path(), "other-bucket");
+
+        assert!(Arc::ptr_eq(&first, &second), "reconnected disks must share a bucket mutation lock");
+        assert!(!Arc::ptr_eq(&first, &other), "different buckets must not serialize each other");
+
+        let _write_guard = first.write().await;
+        assert!(second.try_read().is_err(), "a bucket delete lock must exclude local commits");
+    }
+
    #[derive(Clone, Default)]
    struct CapturedLogs {
        buffer: Arc<Mutex<Vec<u8>>>,
@@ -771,9 +957,26 @@ mod tests {
        }
    }

+    /// Holds a `warn_capture()` capture alive: the thread-local subscriber, plus
+    /// the pin that keeps tracing's process-global callsite-interest cache from
+    /// being decided by some other test's thread.
+    struct WarnCaptureGuard {
+        _subscriber: tracing::subscriber::DefaultGuard,
+        _callsite_pin: tracing::Dispatch,
+    }
+
    /// Capture WARN-level output on the current thread; tokio tests here run on
    /// the current-thread runtime, so the guard covers the whole test body.
-    fn warn_capture() -> (CapturedLogs, tracing::subscriber::DefaultGuard) {
+    ///
+    /// The callsite pin matters because `warn_reliable_rename_failure` is a
+    /// single production callsite shared with tests that call `rename_all`
+    /// *without* installing a subscriber — `rename_all_missing_source_returns_file_not_found`
+    /// is one. Whichever thread reaches it first fixes its `Interest`
+    /// process-wide, so without the pin that sibling can cache
+    /// `Interest::never()` and the WARN never fires here at all, leaving the
+    /// "must keep the WARN" assertions staring at empty output. See
+    /// [`crate::test_tracing::pin_callsite_interest_for_test`].
+    fn warn_capture() -> (CapturedLogs, WarnCaptureGuard) {
        let logs = CapturedLogs::default();
        let subscriber = tracing_subscriber::fmt()
            .with_max_level(tracing::Level::WARN)
@@ -781,7 +984,10 @@ mod tests {
            .with_ansi(false)
            .without_time()
            .finish();
-        let guard = tracing::subscriber::set_default(subscriber);
+        let guard = WarnCaptureGuard {
+            _subscriber: tracing::subscriber::set_default(subscriber),
+            _callsite_pin: crate::test_tracing::pin_callsite_interest_for_test(),
+        };
        (logs, guard)
    }

@@ -968,6 +1174,112 @@ mod tests {
        assert_eq!(std::fs::read(dst.join("nested").join("part.1")).expect("read moved part"), b"payload");
    }

+    #[tokio::test]
+    async fn rename_all_does_not_recreate_missing_base_directory() {
+        let temp_dir = tempdir().expect("create temp dir");
+        let base = temp_dir.path().join("bucket");
+        std::fs::create_dir(&base).expect("create destination base");
+        let src = temp_dir.path().join("staged-object");
+        std::fs::write(&src, b"payload").expect("write staged object");
+        let dst = base.join("object").join("xl.meta");
+        std::fs::remove_dir(&base).expect("delete destination base before commit");
+
+        let err = rename_all(&src, &dst, &base)
+            .await
+            .expect_err("rename must not recreate a deleted destination base");
+
+        assert!(matches!(err, DiskError::FileNotFound));
+        assert!(src.exists(), "failed commit must preserve the staged source");
+        assert!(!base.exists(), "failed commit must not recreate the deleted bucket");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn rename_all_rejects_a_replaced_base_with_an_existing_parent() {
+        use std::os::unix::fs::symlink;
+
+        let temp_dir = tempdir().expect("create temp dir");
+        let base = temp_dir.path().join("bucket");
+        let outside = temp_dir.path().join("outside");
+        std::fs::create_dir(&base).expect("create destination base");
+        std::fs::create_dir_all(outside.join("object")).expect("create outside destination parent");
+        let src = temp_dir.path().join("staged-object");
+        std::fs::write(&src, b"payload").expect("write staged object");
+        let dst = base.join("object").join("xl.meta");
+
+        std::fs::remove_dir(&base).expect("remove destination base before replacement");
+        symlink(&outside, &base).expect("replace destination base with a symlink");
+
+        rename_all(&src, &dst, &base)
+            .await
+            .expect_err("rename must reject an existing destination parent below a replaced base");
+
+        assert!(src.exists(), "rejected rename must preserve the staged source");
+        assert!(
+            !outside.join("object/xl.meta").exists(),
+            "rename must not publish through the replacement symlink"
+        );
+    }
+
+    #[cfg(windows)]
+    #[test]
+    fn windows_parent_guard_blocks_base_and_intermediate_replacement() {
+        let temp_dir = tempdir().expect("create temp dir");
+        let base = temp_dir.path().join("bucket");
+        std::fs::create_dir(&base).expect("create destination base");
+        let parent = base.join("object").join("nested");
+        let guard = mkdir_all_below_existing_base_std(&parent, &base).expect("create and lock destination parents");
+
+        std::fs::rename(&base, temp_dir.path().join("replacement-base"))
+            .expect_err("the locked base must not be replaceable before commit");
+        std::fs::rename(base.join("object"), base.join("replacement-object"))
+            .expect_err("a locked intermediate directory must not be replaceable before commit");
+
+        drop(guard);
+        std::fs::rename(base.join("object"), base.join("replacement-object"))
+            .expect("replacement should succeed after the commit guard is released");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn rename_parent_creation_rejects_symlinked_base() {
+        use std::os::unix::fs::symlink;
+
+        let temp_dir = tempdir().expect("create temp dir");
+        let outside = temp_dir.path().join("outside");
+        std::fs::create_dir(&outside).expect("create outside directory");
+        let base = temp_dir.path().join("bucket");
+        symlink(&outside, &base).expect("create symlinked base");
+
+        mkdir_all_below_existing_base(&base.join("object"), &base)
+            .await
+            .expect_err("symlinked base must be rejected");
+
+        assert!(!outside.join("object").exists(), "parent creation must remain confined to the base");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn rename_parent_creation_rejects_symlink_below_base() {
+        use std::os::unix::fs::symlink;
+
+        let temp_dir = tempdir().expect("create temp dir");
+        let base = temp_dir.path().join("bucket");
+        let outside = temp_dir.path().join("outside");
+        std::fs::create_dir(&base).expect("create destination base");
+        std::fs::create_dir(&outside).expect("create outside directory");
+        symlink(&outside, base.join("linked")).expect("create symlink below base");
+
+        mkdir_all_below_existing_base(&base.join("linked/object"), &base)
+            .await
+            .expect_err("symlink below base must be rejected");
+
+        assert!(
+            !outside.join("object").exists(),
+            "parent creation must not follow a symlink outside the base"
+        );
+    }
+
    #[tokio::test]
    async fn fsync_dir_succeeds_on_directory() {
        let temp_dir = tempdir().expect("create temp dir");
@@ -443,7 +443,9 @@ pub fn bitrot_shard_file_size(size: usize, shard_size: usize, algo: HashAlgorith
    size.div_ceil(shard_size) * algo.size() + size
 }

-/// Verify an interleaved per-block bitrot shard file.
+/// Verify an interleaved per-block bitrot shard file and consume the reader
+/// through EOF. Bytes beyond the encoded length are corruption, even when every
+/// expected block has a valid hash.
 ///
 /// The read loop below assumes every block on disk is `[hash][data]` (streaming
 /// bitrot). It is therefore only valid for the streaming Highway variants, whose
@@ -487,6 +489,11 @@ pub async fn bitrot_verify<R: AsyncRead + Unpin + Send>(
        left -= read;
    }

+    let mut trailing = [0u8; 1];
+    if r.read(&mut trailing).await? != 0 {
+        return Err(std::io::Error::other("bitrot shard file has trailing data"));
+    }
+
    Ok(())
 }

@@ -860,12 +867,19 @@ mod tests {
            .await
            .expect("valid bitrot shard file should verify");

+        let mut truncated = written.clone();
+        truncated.pop();
+        let err = bitrot_verify(Cursor::new(truncated), written.len(), data.len(), algo.clone(), shard_size)
+            .await
+            .expect_err("one-byte-short shard file must be rejected while reading");
+        assert_eq!(err.kind(), std::io::ErrorKind::UnexpectedEof);
+
        let err = bitrot_verify(Cursor::new(written.clone()), written.len() - 1, data.len(), algo.clone(), shard_size)
            .await
            .expect_err("wrong file size must be rejected before reading data");
        assert!(err.to_string().contains("size mismatch"));

-        let mut corrupt = written;
+        let mut corrupt = written.clone();
        let last = corrupt.len() - 1;
        corrupt[last] ^= 0x80;
        let err = bitrot_verify(
@@ -878,6 +892,21 @@ mod tests {
        .await
        .expect_err("hash mismatch must reject corrupted data");
        assert!(err.to_string().contains("hash mismatch"));
+
+        for trailing in [vec![0xa5], vec![0xa5; 17]] {
+            let mut oversized = written.clone();
+            oversized.extend_from_slice(&trailing);
+            let err = bitrot_verify(
+                Cursor::new(oversized),
+                written.len(),
+                data.len(),
+                HashAlgorithm::HighwayHash256S,
+                shard_size,
+            )
+            .await
+            .expect_err("trailing bytes after a valid encoded shard must be rejected");
+            assert!(err.to_string().contains("trailing data"));
+        }
    }

    #[tokio::test]
@@ -909,6 +938,22 @@ mod tests {
        assert!(err.to_string().contains("hash mismatch"));
    }

+    #[tokio::test]
+    async fn bitrot_verify_accepts_exact_legacy_streaming_layout() {
+        let data = b"legacy streaming bitrot";
+        let shard_size = 8;
+        let algo = HashAlgorithm::HighwayHash256SLegacy;
+        let mut writer = BitrotWriter::new(Cursor::new(Vec::new()), shard_size, algo.clone());
+        for chunk in data.chunks(shard_size) {
+            writer.write(chunk).await.expect("legacy streaming shard should encode");
+        }
+        let written = writer.into_inner().into_inner();
+
+        bitrot_verify(Cursor::new(written.clone()), written.len(), data.len(), algo, shard_size)
+            .await
+            .expect("exact legacy streaming shard should remain valid");
+    }
+
    #[tokio::test]
    async fn write_all_vectored_retries_partial_hash_and_data_writes_and_rejects_zero_write() {
        let mut writer = LimitedVectoredWriter {
@@ -818,7 +818,11 @@ impl From<s3s::xml::DeError> for Error {

 impl From<tonic::Status> for Error {
    fn from(e: tonic::Status) -> Self {
-        Error::other(e.to_string())
+        // Keep the typed status as the io::Error payload instead of a
+        // flattened string so RPC failure classifiers can read the gRPC code
+        // via downcast (see `PeerRestClient::is_network_like_error`).
+        // `RpcStatusError` renders exactly as `e.to_string()` did.
+        Error::other(crate::disk::error::RpcStatusError::from(e))
    }
 }

@@ -93,3 +93,6 @@ pub(crate) mod ecstore_validation_blackbox;

 #[cfg(test)]
 pub(crate) mod test_metrics;
+
+#[cfg(test)]
+pub(crate) mod test_tracing;
@@ -28,7 +28,6 @@ use md5::{Digest, Md5};
 use rustfs_kms::{KmsUnavailableError, is_data_key_envelope, types::ObjectEncryptionContext};
 use rustfs_utils::http::{SSEC_ALGORITHM_HEADER, SSEC_KEY_HEADER, SSEC_KEY_MD5_HEADER};
 use rustfs_utils::path::path_join_buf;
-#[cfg(feature = "rio-v2")]
 use serde::Deserialize;
 #[cfg(feature = "rio-v2")]
 use sha2::Sha256;
@@ -44,6 +43,7 @@ const INTERNAL_ENCRYPTION_IV_HEADER: &str = "x-rustfs-encryption-iv";
 const INTERNAL_ENCRYPTION_ORIGINAL_SIZE_HEADER: &str = "x-rustfs-encryption-original-size";
 const SSEC_ORIGINAL_SIZE_HEADER: &str = "x-amz-server-side-encryption-customer-original-size";
 const DEFAULT_SSE_ALGORITHM: &str = "AES256";
+const LOCAL_SSE_DEK_FORMAT_VERSION: u8 = 1;
 #[cfg(feature = "rio-v2")]
 const DARE_PAYLOAD_SIZE: i64 = 64 * 1024;
 #[cfg(feature = "rio-v2")]
@@ -1716,16 +1716,39 @@ fn decrypt_local_sse_dek(encrypted_dek: &[u8], _kms_key_id: &str, object_context

 fn decrypt_rustfs_local_sse_dek(encrypted_dek: &[u8]) -> Result<[u8; 32]> {
    let encrypted_dek = std::str::from_utf8(encrypted_dek).map_err(|_| Error::other("managed DEK is not valid UTF-8"))?;
-    let parts: Vec<&str> = encrypted_dek.split(':').collect();
-    if parts.len() != 2 {
-        return Err(Error::other("invalid managed DEK format"));
+    #[derive(Deserialize)]
+    #[serde(deny_unknown_fields)]
+    struct LocalSseDekEnvelope<'a> {
+        version: u8,
+        nonce: &'a str,
+        ciphertext: &'a str,
    }

+    let (nonce, ciphertext) = match serde_json::from_str::<LocalSseDekEnvelope<'_>>(encrypted_dek) {
+        Ok(envelope) => {
+            if envelope.version != LOCAL_SSE_DEK_FORMAT_VERSION {
+                return Err(Error::other(format!("unsupported managed DEK format version: {}", envelope.version)));
+            }
+            (envelope.nonce, envelope.ciphertext)
+        }
+        Err(_) => {
+            // DEPRECATED: read-only compatibility for persisted colon-delimited DEKs.
+            // RUSTFS_COMPAT_TODO(sse-local-dek-json-v1): Remove after all supported upgrades have rewritten legacy DEKs.
+            let Some((nonce, ciphertext)) = encrypted_dek.split_once(':') else {
+                return Err(Error::other("invalid managed DEK format"));
+            };
+            if ciphertext.contains(':') {
+                return Err(Error::other("invalid managed DEK format"));
+            }
+            (nonce, ciphertext)
+        }
+    };
+
    let nonce_vec = BASE64_STANDARD
-        .decode(parts[0])
+        .decode(nonce)
        .map_err(|_| Error::other("invalid managed DEK nonce"))?;
    let ciphertext = BASE64_STANDARD
-        .decode(parts[1])
+        .decode(ciphertext)
        .map_err(|_| Error::other("invalid managed DEK ciphertext"))?;

    let nonce_array: [u8; 12] = nonce_vec
@@ -2324,9 +2347,36 @@ mod tests {
        let cipher = Aes256Gcm::new(&key);
        let nonce = Nonce::from([0u8; 12]);
        let ciphertext = cipher.encrypt(&nonce, dek.as_slice()).expect("encrypt managed dek");
+        serde_json::json!({
+            "version": LOCAL_SSE_DEK_FORMAT_VERSION,
+            "nonce": BASE64_STANDARD.encode(nonce),
+            "ciphertext": BASE64_STANDARD.encode(ciphertext),
+        })
+        .to_string()
+    }
+
+    fn encrypt_legacy_managed_dek_for_test(dek: [u8; 32], master_key: [u8; 32]) -> String {
+        let key = Key::<Aes256Gcm>::from(master_key);
+        let cipher = Aes256Gcm::new(&key);
+        let nonce = Nonce::from([0u8; 12]);
+        let ciphertext = cipher.encrypt(&nonce, dek.as_slice()).expect("encrypt legacy managed dek");
        format!("{}:{}", BASE64_STANDARD.encode(nonce), BASE64_STANDARD.encode(ciphertext))
    }

+    #[test]
+    fn decrypt_rustfs_local_sse_dek_rejects_unknown_json_version() {
+        let envelope = serde_json::json!({
+            "version": LOCAL_SSE_DEK_FORMAT_VERSION + 1,
+            "nonce": BASE64_STANDARD.encode([0u8; 12]),
+            "ciphertext": BASE64_STANDARD.encode([0u8; 48]),
+        })
+        .to_string();
+
+        let error =
+            decrypt_rustfs_local_sse_dek(envelope.as_bytes()).expect_err("unknown local SSE DEK versions must fail closed");
+        assert!(error.to_string().contains("unsupported managed DEK format version"));
+    }
+
    #[cfg(feature = "rio-v2")]
    fn seal_managed_s3_object_key_for_test(
        bucket: &str,
@@ -2433,7 +2483,7 @@ mod tests {
        async_with_vars([("__RUSTFS_SSE_SIMPLE_CMK", Some(BASE64_STANDARD.encode([7u8; 32])))], async {
            let data_key = [0x24; 32];
            let base_nonce = [0x14; 12];
-            let encrypted_dek = encrypt_managed_dek_for_test(data_key, [7u8; 32]);
+            let encrypted_dek = encrypt_legacy_managed_dek_for_test(data_key, [7u8; 32]);
            let metadata = HashMap::from([
                (
                    INTERNAL_ENCRYPTION_KEY_HEADER.to_string(),
@@ -180,6 +180,7 @@ pub struct ObjectInfo {
    pub data_dir: Option<Uuid>,
    pub delete_marker: bool,
    pub transitioned_object: TransitionedObject,
+    pub transition_version_state: rustfs_filemeta::TransitionVersionState,
    pub restore_ongoing: bool,
    pub restore_expires: Option<OffsetDateTime>,
    pub user_tags: Arc<String>,
@@ -220,6 +221,7 @@ impl Clone for ObjectInfo {
            data_dir: self.data_dir,
            delete_marker: self.delete_marker,
            transitioned_object: self.transitioned_object.clone(),
+            transition_version_state: self.transition_version_state,
            restore_ongoing: self.restore_ongoing,
            restore_expires: self.restore_expires,
            user_tags: self.user_tags.clone(),
@@ -464,11 +466,11 @@ impl ObjectInfo {

        let transitioned_object = TransitionedObject {
            name: fi.transitioned_objname.clone(),
-            version_id: if let Some(transition_version_id) = fi.transition_version_id {
-                transition_version_id.to_string()
-            } else {
-                "".to_string()
-            },
+            version_id: fi
+                .transition_version
+                .clone()
+                .or_else(|| fi.transition_version_id.map(|version_id| version_id.to_string()))
+                .unwrap_or_default(),
            status: fi.transition_status.clone(),
            free_version: fi.tier_free_version(),
            tier: fi.transition_tier.clone(),
@@ -537,6 +539,7 @@ impl ObjectInfo {
            inlined,
            user_defined: Arc::new(metadata),
            transitioned_object,
+            transition_version_state: fi.transition_version_state,
            checksum: fi.checksum.clone(),
            storage_class,
            restore_ongoing,
@@ -207,10 +207,6 @@ pub(crate) async fn ensure_boot_time() {
    GLOBAL_BOOT_TIME.get_or_init(|| async { SystemTime::now() }).await;
 }

-pub(crate) async fn scanner_init_time() -> Option<chrono::DateTime<chrono::Utc>> {
-    rustfs_common::get_global_init_time().await
-}
-
 pub(crate) async fn root_disk_threshold_for_erasure_disk() -> Option<u64> {
    if is_erasure_sd().await {
        None
@@ -71,6 +71,7 @@ fn to_madmin_scanner_metrics(metrics: rustfs_common::metrics::ScannerMetricsRepo
    MadminScannerMetrics {
        collected_at: metrics.collected_at,
        current_cycle: metrics.current_cycle,
+        current_cycle_active: Some(metrics.current_cycle_active),
        current_started: metrics.current_started,
        cycles_completed_at: metrics.cycles_completed_at,
        ongoing_buckets: metrics.ongoing_buckets,
@@ -398,10 +399,7 @@ pub async fn collect_local_metrics(types: MetricType, opts: &CollectMetricsOpts)

    if types.contains(&MetricType::SCANNER) {
        debug!("start get scanner metrics");
-        let mut metrics = global_metrics().report().await;
-        if let Some(init_time) = runtime_sources::scanner_init_time().await {
-            metrics.current_started = init_time;
-        }
+        let metrics = global_metrics().report().await;
        real_time_metrics.aggregated.scanner = Some(to_madmin_scanner_metrics(metrics));
    }

@@ -540,7 +538,9 @@ async fn collect_local_disks_metrics(disks: &HashSet<String>) -> HashMap<String,
 #[cfg(test)]
 mod test {
    use super::*;
+    use rustfs_common::metrics::CurrentCycle;
    use rustfs_io_metrics::internode_metrics::global_internode_metrics;
+    use serial_test::serial;
    use std::time::Duration;

    #[test]
@@ -588,7 +588,10 @@ mod test {

    #[test]
    fn scanner_metrics_mapping_preserves_partial_source_status() {
+        let current_started = Utc::now() - chrono::Duration::seconds(5);
        let scanner = to_madmin_scanner_metrics(rustfs_common::metrics::ScannerMetricsReport {
+            current_cycle_active: true,
+            current_started,
            last_cycle_partial_source: "usage".to_string(),
            last_cycle_partial_source_code: 1,
            partial_cycles_by_source: vec![rustfs_common::metrics::ScannerSourceCycleSnapshot {
@@ -598,6 +601,8 @@ mod test {
            ..Default::default()
        });

+        assert_eq!(scanner.current_cycle_active, Some(true));
+        assert_eq!(scanner.current_started, current_started);
        assert_eq!(scanner.last_cycle_partial_source, "usage");
        assert_eq!(scanner.last_cycle_partial_source_code, 1);
        let usage = scanner
@@ -608,6 +613,39 @@ mod test {
        assert_eq!(usage.cycles, 2);
    }

+    #[tokio::test]
+    #[serial]
+    async fn collect_local_metrics_preserves_scanner_cycle_started_time() {
+        let previous_init_time = *rustfs_common::globals::GLOBAL_INIT_TIME.read().await;
+        let previous_cycle = global_metrics().get_cycle().await;
+        let init_time = Utc::now() - chrono::Duration::hours(1);
+        let cycle_started = Utc::now() - chrono::Duration::seconds(5);
+        *rustfs_common::globals::GLOBAL_INIT_TIME.write().await = Some(init_time);
+        let cycle = CurrentCycle {
+            current: 0,
+            next: 1,
+            started: cycle_started,
+            ..Default::default()
+        };
+        let cycle_start = global_metrics().start_scan_cycle_work_with_cycle(cycle).await;
+
+        let realtime = collect_local_metrics(MetricType::SCANNER, &CollectMetricsOpts::default()).await;
+
+        global_metrics()
+            .finish_scan_cycle_work_with_cycle(cycle_start, previous_cycle.clone().unwrap_or_default())
+            .await;
+        global_metrics().set_cycle(previous_cycle).await;
+        *rustfs_common::globals::GLOBAL_INIT_TIME.write().await = previous_init_time;
+
+        let encoded = rmp_serde::to_vec_named(&realtime).expect("realtime metrics should encode");
+        let decoded: RealtimeMetrics = rmp_serde::from_slice(&encoded).expect("realtime metrics should decode");
+        let mut aggregated = RealtimeMetrics::default();
+        aggregated.merge(decoded);
+        let scanner = aggregated.aggregated.scanner.expect("scanner metrics");
+        assert_eq!(scanner.current_cycle_active, Some(true));
+        assert_eq!(scanner.current_started, cycle_started);
+    }
+
    #[test]
    fn scanner_metrics_mapping_preserves_pacing_pressure() {
        let scanner = to_madmin_scanner_metrics(rustfs_common::metrics::ScannerMetricsReport {
@@ -29,11 +29,11 @@ use rustfs_madmin::metrics::RealtimeMetrics;
 use rustfs_madmin::net::NetInfo;
 use rustfs_madmin::{ItemState, ServerProperties, StorageInfo};
 use rustfs_utils::XHost;
-use std::collections::{HashMap, hash_map::DefaultHasher};
+use std::collections::{BTreeMap, HashMap, hash_map::DefaultHasher};
 use std::future::Future;
 use std::hash::{Hash, Hasher};
 use std::sync::{Arc, Mutex, OnceLock};
-use std::time::{Duration, SystemTime};
+use std::time::{Duration, Instant, SystemTime};
 use tokio::time::{sleep, timeout};
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, error, info, warn};
@@ -47,6 +47,9 @@ const EVENT_NOTIFICATION_PEER_PROPAGATION: &str = "notification_peer_propagation
 const SCANNER_ACTIVITY_PROBE_TIMEOUT: Duration = Duration::from_secs(5);
 const TIER_CONFIG_RELOAD_RETRY_BASE: Duration = Duration::from_millis(100);
 const TIER_CONFIG_RELOAD_RETRY_CAP: Duration = Duration::from_secs(5);
+const REMOTE_VERSION_STATE_PROBE_INTERVAL: Duration = Duration::from_secs(10);
+const REMOTE_VERSION_STATE_PROBE_TIMEOUT: Duration = Duration::from_secs(5);
+const REMOTE_VERSION_STATE_PROOF_TTL: Duration = Duration::from_secs(30);

 /// Cached result from the last successful admin call to a peer.
 struct PeerAdminCache {
@@ -91,6 +94,180 @@ lazy_static! {
    pub static ref GLOBAL_NOTIFICATION_SYS: OnceLock<Arc<NotificationSys>> = OnceLock::new();
 }

+#[derive(Clone)]
+struct RemoteVersionStateFleetProof {
+    topology_fingerprint: String,
+    peer_epochs: Arc<BTreeMap<String, Uuid>>,
+    expires_at: Instant,
+}
+
+impl RemoteVersionStateFleetProof {
+    fn token(&self) -> RemoteVersionStateFleetProofToken {
+        RemoteVersionStateFleetProofToken {
+            topology_fingerprint: self.topology_fingerprint.clone(),
+            peer_epochs: self.peer_epochs.clone(),
+        }
+    }
+}
+
+#[derive(Clone, PartialEq, Eq)]
+pub(crate) struct RemoteVersionStateFleetProofToken {
+    topology_fingerprint: String,
+    peer_epochs: Arc<BTreeMap<String, Uuid>>,
+}
+
+#[derive(Default)]
+struct RemoteVersionStateFleetProofState {
+    proof: Option<RemoteVersionStateFleetProof>,
+    topology_conflict: bool,
+}
+
+static REMOTE_VERSION_STATE_FLEET_PROOF: OnceLock<std::sync::RwLock<RemoteVersionStateFleetProofState>> = OnceLock::new();
+static REMOTE_VERSION_STATE_PROBE_TOPOLOGY: OnceLock<String> = OnceLock::new();
+
+fn remote_version_state_fleet_proof_slot() -> &'static std::sync::RwLock<RemoteVersionStateFleetProofState> {
+    REMOTE_VERSION_STATE_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(RemoteVersionStateFleetProofState::default()))
+}
+
+fn replace_remote_version_state_fleet_proof(proof: Option<RemoteVersionStateFleetProof>) {
+    replace_remote_version_state_fleet_proof_in(remote_version_state_fleet_proof_slot(), proof);
+}
+
+fn replace_remote_version_state_fleet_proof_in(
+    slot: &std::sync::RwLock<RemoteVersionStateFleetProofState>,
+    proof: Option<RemoteVersionStateFleetProof>,
+) {
+    slot.write().unwrap_or_else(std::sync::PoisonError::into_inner).proof = proof;
+}
+
+fn publish_remote_version_state_probe_result(
+    slot: &std::sync::RwLock<RemoteVersionStateFleetProofState>,
+    topology_fingerprint: &str,
+    result: Result<BTreeMap<String, Uuid>>,
+    observed_at: Instant,
+) -> Option<Error> {
+    match result {
+        Ok(peer_epochs) => {
+            let mut state = slot.write().unwrap_or_else(std::sync::PoisonError::into_inner);
+            let peer_epochs = state
+                .proof
+                .as_ref()
+                .filter(|proof| proof.topology_fingerprint == topology_fingerprint && proof.peer_epochs.as_ref() == &peer_epochs)
+                .map(|proof| Arc::clone(&proof.peer_epochs))
+                .unwrap_or_else(|| Arc::new(peer_epochs));
+            state.proof = Some(RemoteVersionStateFleetProof {
+                topology_fingerprint: topology_fingerprint.to_string(),
+                peer_epochs,
+                expires_at: observed_at + REMOTE_VERSION_STATE_PROOF_TTL,
+            });
+            None
+        }
+        Err(err) => {
+            replace_remote_version_state_fleet_proof_in(slot, None);
+            Some(err)
+        }
+    }
+}
+
+pub(crate) fn acquire_remote_version_state_fleet_proof() -> Option<RemoteVersionStateFleetProofToken> {
+    let expected_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get()?;
+    let state = remote_version_state_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    acquire_remote_version_state_fleet_proof_from(&state, expected_topology, Instant::now())
+}
+
+fn acquire_remote_version_state_fleet_proof_from(
+    state: &RemoteVersionStateFleetProofState,
+    expected_topology: &str,
+    now: Instant,
+) -> Option<RemoteVersionStateFleetProofToken> {
+    if state.topology_conflict || !remote_version_state_fleet_proof_valid_at(state.proof.as_ref(), expected_topology, now) {
+        return None;
+    }
+    state.proof.as_ref().map(RemoteVersionStateFleetProof::token)
+}
+
+pub(crate) fn remote_version_state_fleet_proof_matches(proof: &RemoteVersionStateFleetProofToken) -> bool {
+    let Some(expected_topology) = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() else {
+        return false;
+    };
+    let state = remote_version_state_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    if state.topology_conflict {
+        return false;
+    }
+    state.proof.as_ref().is_some_and(|current| {
+        current.topology_fingerprint == *expected_topology
+            && current.topology_fingerprint == proof.topology_fingerprint
+            && Arc::ptr_eq(&current.peer_epochs, &proof.peer_epochs)
+            && Instant::now() < current.expires_at
+    })
+}
+
+fn remote_version_state_fleet_proof_valid_at(
+    proof: Option<&RemoteVersionStateFleetProof>,
+    expected_topology: &str,
+    now: Instant,
+) -> bool {
+    proof.is_some_and(|proof| proof.topology_fingerprint == expected_topology && now < proof.expires_at)
+}
+
+fn insert_remote_version_state_peer(peer_epochs: &mut BTreeMap<String, Uuid>, peer: String, epoch: Uuid) -> Result<()> {
+    if epoch.is_nil() || peer_epochs.values().any(|existing| *existing == epoch) || peer_epochs.insert(peer, epoch).is_some() {
+        return Err(Error::other("remote version state capability peer identity is invalid"));
+    }
+    Ok(())
+}
+
+pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
+    if REMOTE_VERSION_STATE_PROBE_TOPOLOGY.set(topology_fingerprint.clone()).is_err() {
+        if REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() != Some(&topology_fingerprint) {
+            let mut state = remote_version_state_fleet_proof_slot()
+                .write()
+                .unwrap_or_else(std::sync::PoisonError::into_inner);
+            state.topology_conflict = true;
+            state.proof = None;
+        }
+        return;
+    }
+
+    tokio::spawn(async move {
+        loop {
+            let result = match get_global_notification_sys() {
+                Some(notification_sys) => {
+                    match timeout(
+                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                        notification_sys.probe_remote_version_state_fleet(&topology_fingerprint),
+                    )
+                    .await
+                    {
+                        Ok(result) => result,
+                        Err(_) => Err(Error::other("remote version state fleet capability probe timed out")),
+                    }
+                }
+                None => Err(Error::other("remote version state fleet capability notification system is unavailable")),
+            };
+            let topology_conflict = remote_version_state_fleet_proof_slot()
+                .read()
+                .unwrap_or_else(std::sync::PoisonError::into_inner)
+                .topology_conflict;
+            if topology_conflict {
+                replace_remote_version_state_fleet_proof(None);
+            } else if let Some(err) = publish_remote_version_state_probe_result(
+                remote_version_state_fleet_proof_slot(),
+                &topology_fingerprint,
+                result,
+                Instant::now(),
+            ) {
+                debug!(error = %err, "remote version state fleet capability probe failed closed");
+            }
+            sleep(REMOTE_VERSION_STATE_PROBE_INTERVAL).await;
+        }
+    });
+}
+
 pub async fn new_global_notification_sys(eps: EndpointServerPools) -> Result<()> {
    let _ = GLOBAL_NOTIFICATION_SYS
        .set(Arc::new(NotificationSys::new(eps).await))
@@ -115,7 +292,17 @@ pub struct NotificationSys {

 impl NotificationSys {
    pub async fn new(eps: EndpointServerPools) -> Self {
+        let expected_remote_hosts = eps
+            .peer_grid_host_slots_sorted()
+            .into_iter()
+            .filter_map(|(peer, _, is_local)| (!is_local).then_some(peer))
+            .collect::<Vec<_>>();
        let (peer_clients, all_peer_clients, peer_topology_hosts) = PeerRestClient::new_clients_with_topology(eps).await;
+        let peer_topology_hosts = if peer_topology_hosts.is_empty() {
+            expected_remote_hosts
+        } else {
+            peer_topology_hosts
+        };
        let peer_admin_caches = (0..peer_clients.len()).map(|_| Mutex::new(PeerAdminCache::new())).collect();
        Self {
            peer_clients,
@@ -125,6 +312,24 @@ impl NotificationSys {
            tier_config_reload_workers: Default::default(),
        }
    }
+
+    async fn probe_remote_version_state_fleet(&self, topology_fingerprint: &str) -> Result<BTreeMap<String, Uuid>> {
+        if self.peer_clients.len() != self.peer_topology_hosts.len() {
+            return Err(Error::other("remote version state capability fleet membership is incomplete"));
+        }
+        let probes = self.peer_clients.iter().map(|client| async {
+            let client = client
+                .as_ref()
+                .ok_or_else(|| Error::other("remote version state capability peer is unreachable"))?;
+            client.probe_remote_version_state(topology_fingerprint.to_string()).await
+        });
+        let mut peer_epochs = BTreeMap::new();
+        for result in join_all(probes).await {
+            let (peer, epoch) = result?;
+            insert_remote_version_state_peer(&mut peer_epochs, peer, epoch)?;
+        }
+        Ok(peer_epochs)
+    }
 }

 pub struct NotificationPeerErr {
@@ -1344,8 +1549,11 @@ async fn run_tier_config_reload_worker<F, Fut>(
                }
                TierConfigReloadFinish::Pending => retry_attempt = 0,
            },
-            TierConfigReloadOutcome::Terminal(_) => match sys.finish_tier_config_reload_worker(&host) {
+            TierConfigReloadOutcome::Terminal(err) => match sys.finish_tier_config_reload_worker(&host) {
                TierConfigReloadFinish::Completed => {
+                    // This peer keeps the previous tier configuration for good, so record
+                    // why. Dropping the error here hides the only evidence of a divergent
+                    // node behind an outcome label that cannot be acted on.
                    warn!(
                        event = EVENT_NOTIFICATION_PEER_PROPAGATION,
                        component = LOG_COMPONENT_ECSTORE,
@@ -1353,6 +1561,7 @@ async fn run_tier_config_reload_worker<F, Fut>(
                        action = "reload_transition_tier_config",
                        host,
                        outcome = "terminal",
+                        error = ?err,
                        "tier configuration reload stopped after a terminal outcome"
                    );
                    return;
@@ -1886,6 +2095,183 @@ fn aggregate_scanner_dirty_usage_acknowledgement_results(
 mod tests {
    use super::*;

+    #[test]
+    fn remote_version_state_fleet_proof_rejects_stale_or_mismatched_membership() {
+        let now = Instant::now();
+        let mut peer_epochs = BTreeMap::new();
+        peer_epochs.insert("peer-a".to_string(), Uuid::new_v4());
+        let proof = RemoteVersionStateFleetProof {
+            topology_fingerprint: "topology-a".to_string(),
+            peer_epochs: Arc::new(peer_epochs),
+            expires_at: now + Duration::from_secs(1),
+        };
+
+        assert!(remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-a", now));
+        assert!(!remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-b", now));
+        assert!(!remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-a", proof.expires_at));
+        assert!(!remote_version_state_fleet_proof_valid_at(None, "topology-a", now));
+    }
+
+    #[test]
+    fn remote_version_state_fleet_proof_rejects_nil_process_epoch() {
+        let mut peer_epochs = BTreeMap::new();
+
+        assert!(insert_remote_version_state_peer(&mut peer_epochs, "peer-a".to_string(), Uuid::nil()).is_err());
+        assert!(peer_epochs.is_empty());
+    }
+
+    #[test]
+    fn remote_version_state_fleet_proof_accepts_single_node_membership() {
+        let now = Instant::now();
+        let proof = RemoteVersionStateFleetProof {
+            topology_fingerprint: "topology-a".to_string(),
+            peer_epochs: Arc::new(BTreeMap::new()),
+            expires_at: now + Duration::from_secs(1),
+        };
+
+        assert!(remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-a", now));
+    }
+
+    #[test]
+    fn remote_version_state_fleet_proof_token_changes_with_process_epoch() {
+        let now = Instant::now();
+        let proof = RemoteVersionStateFleetProof {
+            topology_fingerprint: "topology-a".to_string(),
+            peer_epochs: Arc::new(BTreeMap::from([("peer-a".to_string(), Uuid::new_v4())])),
+            expires_at: now + Duration::from_secs(1),
+        };
+        let captured = proof.token();
+        let restarted = RemoteVersionStateFleetProof {
+            topology_fingerprint: proof.topology_fingerprint.clone(),
+            peer_epochs: Arc::new(BTreeMap::from([("peer-a".to_string(), Uuid::new_v4())])),
+            expires_at: proof.expires_at,
+        };
+
+        assert!(captured != restarted.token());
+    }
+
+    #[test]
+    fn remote_version_state_fleet_proof_renewal_preserves_only_same_epoch_token() {
+        let slot = std::sync::RwLock::new(RemoteVersionStateFleetProofState::default());
+        let now = Instant::now();
+        let epoch = Uuid::new_v4();
+        let peers = BTreeMap::from([("peer-a".to_string(), epoch)]);
+        assert!(publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peers.clone()), now).is_none());
+        let original = slot
+            .read()
+            .expect("proof slot should not poison")
+            .proof
+            .as_ref()
+            .expect("successful probe should publish proof")
+            .token();
+
+        assert!(
+            publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peers), now + Duration::from_millis(1)).is_none()
+        );
+        let renewed = slot
+            .read()
+            .expect("proof slot should not poison")
+            .proof
+            .as_ref()
+            .expect("renewal should retain proof")
+            .token();
+        assert!(Arc::ptr_eq(&original.peer_epochs, &renewed.peer_epochs));
+
+        let restarted = BTreeMap::from([("peer-a".to_string(), Uuid::new_v4())]);
+        assert!(
+            publish_remote_version_state_probe_result(&slot, "topology-a", Ok(restarted), now + Duration::from_millis(2))
+                .is_none()
+        );
+        let replaced = slot
+            .read()
+            .expect("proof slot should not poison")
+            .proof
+            .as_ref()
+            .expect("restarted peer should publish a new proof")
+            .token();
+        assert!(!Arc::ptr_eq(&original.peer_epochs, &replaced.peer_epochs));
+    }
+
+    #[test]
+    fn remote_version_state_fleet_proof_conflict_revokes_atomic_snapshot() {
+        let now = Instant::now();
+        let mut state = RemoteVersionStateFleetProofState {
+            proof: Some(RemoteVersionStateFleetProof {
+                topology_fingerprint: "topology-a".to_string(),
+                peer_epochs: Arc::new(BTreeMap::new()),
+                expires_at: now + Duration::from_secs(1),
+            }),
+            topology_conflict: false,
+        };
+        assert!(acquire_remote_version_state_fleet_proof_from(&state, "topology-a", now).is_some());
+
+        state.topology_conflict = true;
+        assert!(acquire_remote_version_state_fleet_proof_from(&state, "topology-a", now).is_none());
+    }
+
+    #[test]
+    fn remote_version_state_fleet_probe_rejects_duplicate_member_or_process_epoch() {
+        let epoch = Uuid::new_v4();
+        let mut peer_epochs = BTreeMap::new();
+        insert_remote_version_state_peer(&mut peer_epochs, "node-a:9000".to_string(), epoch)
+            .expect("first member should be admitted");
+        assert!(insert_remote_version_state_peer(&mut peer_epochs, "node-b:9000".to_string(), epoch).is_err());
+        assert!(insert_remote_version_state_peer(&mut peer_epochs, "node-a:9000".to_string(), Uuid::new_v4()).is_err());
+        assert!(insert_remote_version_state_peer(&mut peer_epochs, "node-c:9000".to_string(), Uuid::nil()).is_err());
+    }
+
+    #[test]
+    fn remote_version_state_fleet_probe_failure_revokes_previous_proof() {
+        let slot = std::sync::RwLock::new(RemoteVersionStateFleetProofState::default());
+        let now = Instant::now();
+        let peer_epochs = BTreeMap::from([("node-a:9000".to_string(), Uuid::new_v4())]);
+        assert!(publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peer_epochs), now).is_none());
+        assert!(slot.read().expect("proof slot should not poison").proof.is_some());
+
+        assert!(
+            publish_remote_version_state_probe_result(&slot, "topology-a", Err(Error::other("peer unavailable")), now,).is_some()
+        );
+        assert!(slot.read().expect("proof slot should not poison").proof.is_none());
+
+        let peer_epochs = BTreeMap::from([("node-a:9000".to_string(), Uuid::new_v4())]);
+        assert!(publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peer_epochs), now).is_none());
+        assert!(slot.read().expect("proof slot should not poison").proof.is_some());
+    }
+
+    #[tokio::test]
+    async fn remote_version_state_fleet_probe_rejects_unreachable_member() {
+        let notification_sys = NotificationSys {
+            peer_clients: vec![None],
+            all_peer_clients: vec![None, None],
+            peer_topology_hosts: vec!["peer-a".to_string()],
+            peer_admin_caches: vec![Mutex::new(PeerAdminCache::new())],
+            tier_config_reload_workers: Default::default(),
+        };
+
+        let err = notification_sys
+            .probe_remote_version_state_fleet("topology-a")
+            .await
+            .expect_err("an unreachable configured member must fail the fleet proof");
+        assert!(err.to_string().contains("unreachable"));
+    }
+
+    #[tokio::test]
+    async fn remote_version_state_fleet_probe_rejects_missing_member_slot() {
+        let notification_sys = NotificationSys {
+            peer_clients: Vec::new(),
+            all_peer_clients: vec![None],
+            peer_topology_hosts: vec!["peer-a".to_string()],
+            peer_admin_caches: Vec::new(),
+            tier_config_reload_workers: Default::default(),
+        };
+
+        let err = notification_sys
+            .probe_remote_version_state_fleet("topology-a")
+            .await
+            .expect_err("a missing configured member slot must fail the fleet proof");
+        assert!(err.to_string().contains("incomplete"));
+    }
+
    fn build_props(endpoint: &str) -> ServerProperties {
        ServerProperties {
            endpoint: endpoint.to_string(),
@@ -931,7 +931,10 @@ pub async fn read_transition_meta(disk_path: &Path, bucket: &str, object: &str)
        status: fi.transition_status.clone(),
        tier: fi.transition_tier.clone(),
        remote_object: fi.transitioned_objname.clone(),
-        remote_version_id: fi.transition_version_id.map(|id| id.to_string()),
+        remote_version_id: fi
+            .transition_version
+            .clone()
+            .or_else(|| fi.transition_version_id.map(|id| id.to_string())),
        free_version_count,
    })
 }
@@ -236,6 +236,7 @@ struct TierPublishTransition {

 struct PreparedTierDriver {
    tier_name: String,
+    tier_config: TierConfig,
    config_fingerprint: TierDriverFingerprint,
    backend_identity: TierDestinationId,
    exact_get_delete: bool,
@@ -1342,6 +1343,7 @@ fn tier_exact_get_delete(config: &TierConfig) -> bool {

 struct TierDriverGeneration {
    tier_name: Arc<str>,
+    tier_config: TierConfig,
    generation: DriverRevision,
    // Process-local only: this may reflect credential changes and must never be persisted or logged.
    config_fingerprint: TierDriverFingerprint,
@@ -1349,6 +1351,9 @@ struct TierDriverGeneration {
    backend_identity: TierDestinationId,
    exact_get_delete: bool,
    driver: SharedWarmBackend,
+    reconciler: tokio::sync::OnceCell<
+        Option<Arc<dyn crate::services::tier::warm_backend::TransitionCandidateReconciler + Send + Sync + 'static>>,
+    >,
    accepting: AtomicBool,
    active_leases: AtomicUsize,
    drained: tokio::sync::Notify,
@@ -1473,6 +1478,35 @@ impl TierOperationLease {
        self.inner.backend_identity
    }

+    pub(crate) async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        transaction_id: uuid::Uuid,
+    ) -> io::Result<TransitionCandidateProbe> {
+        let Some(reconciler) = self
+            .inner
+            .reconciler
+            .get_or_try_init(|| async {
+                crate::services::tier::warm_backend::new_transition_candidate_reconciler(&self.inner.tier_config)
+                    .await
+                    .map(|reconciler| reconciler.map(Arc::from))
+            })
+            .await
+            .map_err(|err| io::Error::other(err.message))?
+        else {
+            return self.inner.driver.probe_transition_candidate(object).await;
+        };
+        reconciler
+            .probe_transition_candidate_for(
+                object,
+                crate::services::tier::warm_backend::TransitionCandidateIdentity {
+                    transaction_id,
+                    destination_id: self.backend_identity(),
+                },
+            )
+            .await
+    }
+
    pub(crate) fn validate_remote_version_id(&self, remote_version_id: &str) -> io::Result<()> {
        self.inner.driver.validate_remote_version_id(remote_version_id)?;
        if !remote_version_id.is_empty() && !self.inner.exact_get_delete {
@@ -2833,6 +2867,7 @@ impl TierConfigMgr {
                    let exact_get_delete = tier_exact_get_delete(config);
                    Some(PreparedTierDriver {
                        tier_name: tier_name.to_string(),
+                        tier_config: config.clone(),
                        config_fingerprint,
                        backend_identity,
                        exact_get_delete,
@@ -2861,11 +2896,13 @@ impl TierConfigMgr {
            })?;
            let entry = Arc::new(TierDriverGeneration {
                tier_name: Arc::from(prepared.tier_name.as_str()),
+                tier_config: prepared.tier_config.clone(),
                generation,
                config_fingerprint: prepared.config_fingerprint,
                backend_identity: prepared.backend_identity,
                exact_get_delete: prepared.exact_get_delete,
                driver: prepared.driver.clone(),
+                reconciler: tokio::sync::OnceCell::new(),
                accepting: AtomicBool::new(true),
                active_leases: AtomicUsize::new(0),
                drained: tokio::sync::Notify::new(),
@@ -3609,11 +3646,13 @@ impl TierConfigMgr {
        let driver: SharedWarmBackend = Arc::from(driver);
        let entry = Arc::new(TierDriverGeneration {
            tier_name: Arc::from(tier_name),
+            tier_config: config.clone(),
            generation,
            config_fingerprint,
            backend_identity,
            exact_get_delete,
            driver: driver.clone(),
+            reconciler: tokio::sync::OnceCell::new(),
            accepting: AtomicBool::new(true),
            active_leases: AtomicUsize::new(0),
            drained: tokio::sync::Notify::new(),
@@ -9274,7 +9313,8 @@ mod tests {
            version_id: "v1".to_string(),
            tier_name: "COLD-A".to_string(),
            backend_identity: Some(current_identity),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
        };
        journal_store
            .insert_config_object(
@@ -73,6 +73,21 @@ pub enum TransitionCandidateProbe {
    Unsupported,
 }

+#[derive(Clone, Copy)]
+pub(crate) struct TransitionCandidateIdentity {
+    pub transaction_id: uuid::Uuid,
+    pub destination_id: [u8; 32],
+}
+
+#[async_trait::async_trait]
+pub(crate) trait TransitionCandidateReconciler {
+    async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        identity: TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error>;
+}
+
 #[async_trait::async_trait]
 pub trait WarmBackend {
    async fn validate(&self) -> Result<(), std::io::Error> {
@@ -189,6 +204,20 @@ pub fn build_transition_put_options(storage_class: String, mut metadata: HashMap
        metadata.remove(key);
    }

+    for suffix in [
+        rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+        rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
+    ] {
+        for key in [
+            rustfs_utils::http::metadata_compat::internal_key_rustfs(suffix),
+            format!("{}{}", rustfs_utils::http::metadata_compat::MINIO_INTERNAL_PREFIX, suffix),
+        ] {
+            if let Some(value) = metadata.remove(&key) {
+                metadata.insert(format!("x-amz-meta-{key}"), value);
+            }
+        }
+    }
+
    opts.user_metadata = metadata;
    opts
 }
@@ -446,6 +475,51 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
    Ok(d)
 }

+pub(crate) async fn new_transition_candidate_reconciler(
+    tier: &TierConfig,
+) -> Result<Option<Box<dyn TransitionCandidateReconciler + Send + Sync + 'static>>, AdminError> {
+    let reconciler: Box<dyn TransitionCandidateReconciler + Send + Sync + 'static> = match tier.tier_type {
+        TierType::S3 => Box::new(
+            WarmBackendS3::new(tier.s3.as_ref().ok_or_else(|| ERR_TIER_INVALID_CONFIG.clone())?, &tier.name)
+                .await
+                .map_err(|err| {
+                    let mut admin_err = ERR_TIER_INVALID_CONFIG.clone();
+                    admin_err.message = err.to_string();
+                    admin_err
+                })?,
+        ),
+        TierType::MinIO => Box::new(
+            WarmBackendMinIO::new(tier.minio.as_ref().ok_or_else(|| ERR_TIER_INVALID_CONFIG.clone())?, &tier.name)
+                .await
+                .map_err(|err| {
+                    let mut admin_err = ERR_TIER_INVALID_CONFIG.clone();
+                    admin_err.message = err.to_string();
+                    admin_err
+                })?,
+        ),
+        TierType::RustFS => Box::new(
+            WarmBackendRustFS::new(tier.rustfs.as_ref().ok_or_else(|| ERR_TIER_INVALID_CONFIG.clone())?, &tier.name)
+                .await
+                .map_err(|err| {
+                    let mut admin_err = ERR_TIER_INVALID_CONFIG.clone();
+                    admin_err.message = err.to_string();
+                    admin_err
+                })?,
+        ),
+        TierType::R2 => Box::new(
+            WarmBackendR2::new(tier.r2.as_ref().ok_or_else(|| ERR_TIER_INVALID_CONFIG.clone())?, &tier.name)
+                .await
+                .map_err(|err| {
+                    let mut admin_err = ERR_TIER_INVALID_CONFIG.clone();
+                    admin_err.message = err.to_string();
+                    admin_err
+                })?,
+        ),
+        _ => return Ok(None),
+    };
+    Ok(Some(reconciler))
+}
+
 #[cfg(test)]
 mod tests {
    use super::*;
@@ -775,6 +849,37 @@ mod tests {
        assert!(!opts.user_metadata.contains_key(X_AMZ_REPLICATION_STATUS.as_str()));
    }

+    #[test]
+    fn build_transition_put_options_persists_both_candidate_identity_keys_as_s3_metadata() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            "aaaaaaaa-aaaa-4aaa-8aaa-aaaaaaaaaaaa".to_string(),
+        );
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
+            "5a".repeat(32),
+        );
+
+        let opts = build_transition_put_options("COLD".to_string(), metadata);
+
+        for suffix in [
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
+        ] {
+            assert!(opts.user_metadata.contains_key(&format!(
+                "x-amz-meta-{}",
+                rustfs_utils::http::metadata_compat::internal_key_rustfs(suffix)
+            )));
+            assert!(opts.user_metadata.contains_key(&format!(
+                "x-amz-meta-{}{suffix}",
+                rustfs_utils::http::metadata_compat::MINIO_INTERNAL_PREFIX
+            )));
+        }
+    }
+
    #[test]
    fn build_transition_put_options_requests_no_checksum_and_content_md5() {
        // Regression for rustfs/rustfs#4811: transition uploads must leave the
@@ -19,6 +19,7 @@
 #![allow(clippy::all)]

 use std::collections::HashMap;
+use std::io::{Error, ErrorKind};
 use std::sync::Arc;

 use bytes::Bytes;
@@ -45,6 +46,19 @@ const MAX_PARTS_COUNT: i64 = 10000;
 const _MAX_PART_SIZE: i64 = 1024 * 1024 * 1024 * 5;
 const MIN_PART_SIZE: i64 = 1024 * 1024 * 128;

+fn parse_generation(remote_version: &str) -> Result<Option<i64>, Error> {
+    if remote_version.is_empty() {
+        return Ok(None);
+    }
+    let generation = remote_version
+        .parse::<i64>()
+        .map_err(|_| Error::new(ErrorKind::InvalidData, "GCS remote version is not a valid generation"))?;
+    if generation <= 0 {
+        return Err(Error::new(ErrorKind::InvalidData, "GCS remote version generation must be positive"));
+    }
+    Ok(Some(generation))
+}
+
 pub struct WarmBackendGCS {
    pub client: Arc<Storage>,
    pub control: Arc<StorageControl>,
@@ -105,6 +119,10 @@ impl WarmBackendGCS {

 #[async_trait::async_trait]
 impl WarmBackend for WarmBackendGCS {
+    fn validate_remote_version_id(&self, remote_version_id: &str) -> Result<(), std::io::Error> {
+        parse_generation(remote_version_id).map(|_| ())
+    }
+
    async fn put_with_meta(
        &self,
        object: &str,
@@ -135,6 +153,9 @@ impl WarmBackend for WarmBackendGCS {

    async fn get(&self, object: &str, rv: &str, opts: WarmBackendGetOpts) -> Result<ReadCloser, std::io::Error> {
        let mut req = self.client.read_object(&self.bucket, &self.get_dest(object));
+        if let Some(generation) = parse_generation(rv)? {
+            req = req.set_generation(generation);
+        }

        // Honor the requested byte range so Range GETs on tiered objects return the exact
        // interval instead of the whole object (matches the s3/s3sdk/rustfs warm backends).
@@ -164,13 +185,15 @@ impl WarmBackend for WarmBackendGCS {
        // gRPC v2 DeleteObject requires the bucket in resource-name form. Without this the
        // deleted tiered object was never removed from GCS (empty impl returned Ok), leaking
        // remote data forever.
-        self.control
+        let mut req = self
+            .control
            .delete_object()
            .set_bucket(format!("projects/_/buckets/{}", self.bucket))
-            .set_object(self.get_dest(object))
-            .send()
-            .await
-            .map_err(|e| std::io::Error::other(e.to_string()))?;
+            .set_object(self.get_dest(object));
+        if let Some(generation) = parse_generation(rv)? {
+            req = req.set_generation(generation);
+        }
+        req.send().await.map_err(|e| std::io::Error::other(e.to_string()))?;
        Ok(())
    }

@@ -191,6 +214,30 @@ impl WarmBackend for WarmBackendGCS {
    }
 }

+#[cfg(test)]
+mod tests {
+    use super::parse_generation;
+    use std::io::ErrorKind;
+
+    #[test]
+    fn generation_parser_preserves_exact_numeric_versions() {
+        assert_eq!(parse_generation("").expect("empty generation means no version condition"), None);
+        assert_eq!(parse_generation("1").expect("minimum generation should parse"), Some(1));
+        assert_eq!(
+            parse_generation(&i64::MAX.to_string()).expect("maximum generation should parse"),
+            Some(i64::MAX)
+        );
+    }
+
+    #[test]
+    fn generation_parser_rejects_unknown_or_non_positive_versions() {
+        for value in ["unknown", "1.0", "-1", "0", "9223372036854775808"] {
+            let err = parse_generation(value).expect_err("unknown generation must fail closed");
+            assert_eq!(err.kind(), ErrorKind::InvalidData, "{value}");
+        }
+    }
+}
+
 /*fn gcs_to_object_error(err: Error, params: Vec<String>) -> Option<Error> {
  if err == nil {
    return nil
@@ -135,6 +135,20 @@ impl WarmBackend for WarmBackendMinIO {
    }
 }

+#[async_trait::async_trait]
+impl crate::services::tier::warm_backend::TransitionCandidateReconciler for WarmBackendMinIO {
+    async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        identity: crate::services::tier::warm_backend::TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        crate::services::tier::warm_backend::TransitionCandidateReconciler::probe_transition_candidate_for(
+            &self.0, object, identity,
+        )
+        .await
+    }
+}
+
 fn optimal_part_size(object_size: i64) -> Result<i64, std::io::Error> {
    let mut object_size = object_size;
    if object_size == -1 {
@@ -135,6 +135,20 @@ impl WarmBackend for WarmBackendR2 {
    }
 }

+#[async_trait::async_trait]
+impl crate::services::tier::warm_backend::TransitionCandidateReconciler for WarmBackendR2 {
+    async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        identity: crate::services::tier::warm_backend::TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        crate::services::tier::warm_backend::TransitionCandidateReconciler::probe_transition_candidate_for(
+            &self.0, object, identity,
+        )
+        .await
+    }
+}
+
 fn optimal_part_size(object_size: i64) -> Result<i64, std::io::Error> {
    let mut object_size = object_size;
    if object_size == -1 {
@@ -132,6 +132,20 @@ impl WarmBackend for WarmBackendRustFS {
    }
 }

+#[async_trait::async_trait]
+impl crate::services::tier::warm_backend::TransitionCandidateReconciler for WarmBackendRustFS {
+    async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        identity: crate::services::tier::warm_backend::TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        crate::services::tier::warm_backend::TransitionCandidateReconciler::probe_transition_candidate_for(
+            &self.0, object, identity,
+        )
+        .await
+    }
+}
+
 fn optimal_part_size(object_size: i64) -> Result<i64, std::io::Error> {
    let mut object_size = object_size;
    if object_size == -1 {
@@ -29,6 +29,7 @@ use crate::client::{
    api_remove::{RemoveObjectOptions, RemoveObjectResult},
    api_s3_datatypes::ListVersionsResult,
    credentials::{Credentials, SignatureType, Static, Value},
+    provider_versions::validate_remote_version_id,
    transition_api::{BucketLookupType, Options, TransitionClient, TransitionCore},
    transition_api::{ReadCloser, ReaderImpl},
 };
@@ -36,7 +37,10 @@ use crate::error::ErrorResponse;
 use crate::error::error_resp_to_object_err;
 use crate::services::tier::{
    tier_config::TierS3,
-    warm_backend::{TransitionCandidateProbe, WarmBackend, WarmBackendGetOpts, build_transition_put_options},
+    warm_backend::{
+        TransitionCandidateIdentity, TransitionCandidateProbe, TransitionCandidateReconciler, WarmBackend, WarmBackendGetOpts,
+        build_transition_put_options,
+    },
 };
 use http::HeaderMap;
 use rustfs_utils::egress::validate_outbound_url;
@@ -189,44 +193,188 @@ impl WarmBackendS3 {
        remote_bucket_versioning_from_status(config.status.as_ref().map(|status| status.as_str()))
    }

-    async fn list_transition_candidate_versions(&self, object: &str) -> Result<ListVersionsResult, std::io::Error> {
+    async fn probe_transition_candidate_versions(
+        &self,
+        object: &str,
+        bucket_versioning: RemoteBucketVersioning,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        let remote_object = self.get_dest(object);
        let mut opts = ListObjectsOptions::default();
-        opts.set("prefix", &self.get_dest(object));
-        opts.set("max-keys", "2");
-        self.client.list_object_versions_query(&self.bucket, &opts, "", "", "").await
+        opts.set("prefix", &remote_object);
+        opts.set("max-keys", "1000");
+
+        let mut key_marker = String::new();
+        let mut version_id_marker = String::new();
+        let mut candidates = TransitionCandidateVersions::default();
+        loop {
+            let versions = self
+                .client
+                .list_object_versions_query(&self.bucket, &opts, &key_marker, &version_id_marker, "")
+                .await?;
+            candidates.extend(&remote_object, &versions);
+            if candidates.is_ambiguous() {
+                return Ok(TransitionCandidateProbe::Ambiguous);
+            }
+            if !versions.is_truncated {
+                return classify_transition_candidates(candidates, bucket_versioning);
+            }
+
+            advance_version_markers(&mut key_marker, &mut version_id_marker, &versions)?;
+        }
+    }
+
+    async fn probe_transition_candidate_identity(
+        &self,
+        object: &str,
+        identity: TransitionCandidateIdentity,
+        bucket_versioning: RemoteBucketVersioning,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        let remote_object = self.get_dest(object);
+        let mut opts = ListObjectsOptions::default();
+        opts.set("prefix", &remote_object);
+        opts.set("max-keys", "1000");
+        let mut key_marker = String::new();
+        let mut version_id_marker = String::new();
+        let mut matched_version = None;
+        let mut saw_unproven_candidate = false;
+
+        loop {
+            let versions = self
+                .client
+                .list_object_versions_query(&self.bucket, &opts, &key_marker, &version_id_marker, "")
+                .await?;
+            for version in versions.versions.iter().filter(|version| version.key == remote_object) {
+                let mut stat_opts = GetObjectOptions::default();
+                stat_opts.version_id.clone_from(&version.version_id);
+                let info = self.client.stat_object(&self.bucket, &remote_object, &stat_opts).await?;
+                let mut metadata = info.user_metadata;
+                for (name, value) in &info.metadata {
+                    if (name
+                        .as_str()
+                        .starts_with(rustfs_utils::http::metadata_compat::RUSTFS_INTERNAL_PREFIX)
+                        || name
+                            .as_str()
+                            .starts_with(rustfs_utils::http::metadata_compat::MINIO_INTERNAL_PREFIX))
+                        && let Ok(value) = value.to_str()
+                    {
+                        metadata.insert(name.as_str().to_string(), value.to_string());
+                    }
+                }
+                if transition_candidate_metadata_matches(&metadata, identity)? {
+                    if matched_version.is_some() {
+                        return Ok(TransitionCandidateProbe::Ambiguous);
+                    }
+                    matched_version = Some(version.version_id.clone());
+                } else {
+                    saw_unproven_candidate = true;
+                }
+            }
+            if !versions.is_truncated {
+                if matched_version.is_none() && saw_unproven_candidate {
+                    return Ok(TransitionCandidateProbe::Unsupported);
+                }
+                let candidates = TransitionCandidateVersions {
+                    version_id: matched_version,
+                    ambiguous: false,
+                };
+                return classify_transition_candidates(candidates, bucket_versioning);
+            }
+            advance_version_markers(&mut key_marker, &mut version_id_marker, &versions)?;
+        }
    }
 }

-fn classify_transition_candidate_versions(
-    remote_object: &str,
-    bucket_versioning: RemoteBucketVersioning,
-    versions: &ListVersionsResult,
-) -> TransitionCandidateProbe {
-    if versions.is_truncated {
-        return TransitionCandidateProbe::Ambiguous;
-    }
-
-    if versions.delete_markers.iter().any(|marker| marker.key == remote_object) {
-        return TransitionCandidateProbe::Ambiguous;
-    }
-
-    let mut exact_versions = versions.versions.iter().filter(|version| version.key == remote_object);
-    let Some(version) = exact_versions.next() else {
-        return TransitionCandidateProbe::Missing;
+fn transition_candidate_metadata_matches(
+    metadata: &HashMap<String, String>,
+    identity: TransitionCandidateIdentity,
+) -> Result<bool, std::io::Error> {
+    use rustfs_utils::http::metadata_compat::{
+        SUFFIX_TRANSITION_TIER_DESTINATION_ID, SUFFIX_TRANSITION_TRANSACTION_ID, contains_key_str, get_consistent_str,
    };
-    if exact_versions.next().is_some() {
-        return TransitionCandidateProbe::Ambiguous;
+
+    let transaction_id = get_consistent_str(metadata, SUFFIX_TRANSITION_TRANSACTION_ID);
+    let destination_id = get_consistent_str(metadata, SUFFIX_TRANSITION_TIER_DESTINATION_ID);
+    if transaction_id.is_none() || destination_id.is_none() {
+        if contains_key_str(metadata, SUFFIX_TRANSITION_TRANSACTION_ID)
+            || contains_key_str(metadata, SUFFIX_TRANSITION_TIER_DESTINATION_ID)
+        {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidData,
+                "transition candidate identity metadata is empty or conflicting",
+            ));
+        }
+        return Ok(false);
+    }
+    let expected_transaction_id = identity.transaction_id.to_string();
+    let expected_destination_id = rustfs_utils::crypto::hex(identity.destination_id);
+    Ok(transaction_id == Some(expected_transaction_id.as_str()) && destination_id == Some(expected_destination_id.as_str()))
+}
+
+fn classify_transition_candidates(
+    candidates: TransitionCandidateVersions,
+    bucket_versioning: RemoteBucketVersioning,
+) -> Result<TransitionCandidateProbe, std::io::Error> {
+    let probe = candidates.classify(bucket_versioning);
+    if let TransitionCandidateProbe::VersionedPresent(version_id) = &probe {
+        validate_remote_version_id(version_id)?;
+    }
+    Ok(probe)
+}
+
+fn advance_version_markers(
+    key_marker: &mut String,
+    version_id_marker: &mut String,
+    versions: &ListVersionsResult,
+) -> Result<(), std::io::Error> {
+    let next_markers = (&versions.next_key_marker, &versions.next_version_id_marker);
+    if next_markers == (&*key_marker, &*version_id_marker) {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "ListObjectVersions pagination markers did not advance",
+        ));
+    }
+    key_marker.clone_from(&versions.next_key_marker);
+    version_id_marker.clone_from(&versions.next_version_id_marker);
+    Ok(())
+}
+
+#[derive(Default)]
+struct TransitionCandidateVersions {
+    version_id: Option<String>,
+    ambiguous: bool,
+}
+
+impl TransitionCandidateVersions {
+    fn extend(&mut self, remote_object: &str, versions: &ListVersionsResult) {
+        for version in versions.versions.iter().filter(|version| version.key == remote_object) {
+            if self.version_id.is_some() {
+                self.ambiguous = true;
+                return;
+            }
+            self.version_id = Some(version.version_id.clone());
+        }
    }

-    match bucket_versioning {
-        RemoteBucketVersioning::Disabled => TransitionCandidateProbe::UnversionedPresent,
-        RemoteBucketVersioning::Suspended if version.version_id == "null" => {
-            TransitionCandidateProbe::VersionedPresent(version.version_id.clone())
+    fn is_ambiguous(&self) -> bool {
+        self.ambiguous
+    }
+
+    fn classify(self, bucket_versioning: RemoteBucketVersioning) -> TransitionCandidateProbe {
+        if self.ambiguous {
+            return TransitionCandidateProbe::Ambiguous;
        }
-        RemoteBucketVersioning::Suspended | RemoteBucketVersioning::Enabled if !version.version_id.is_empty() => {
-            TransitionCandidateProbe::VersionedPresent(version.version_id.clone())
+        let Some(version_id) = self.version_id else {
+            return TransitionCandidateProbe::Missing;
+        };
+
+        match bucket_versioning {
+            RemoteBucketVersioning::Disabled => TransitionCandidateProbe::UnversionedPresent,
+            RemoteBucketVersioning::Suspended if version_id == "null" => TransitionCandidateProbe::VersionedPresent(version_id),
+            RemoteBucketVersioning::Suspended | RemoteBucketVersioning::Enabled if !version_id.is_empty() => {
+                TransitionCandidateProbe::VersionedPresent(version_id)
+            }
+            RemoteBucketVersioning::Suspended | RemoteBucketVersioning::Enabled => TransitionCandidateProbe::Ambiguous,
        }
-        RemoteBucketVersioning::Suspended | RemoteBucketVersioning::Enabled => TransitionCandidateProbe::Ambiguous,
    }
 }

@@ -275,74 +423,163 @@ mod tests {
        }
    }

+    fn classify_pages(bucket_versioning: RemoteBucketVersioning, pages: &[ListVersionsResult]) -> TransitionCandidateProbe {
+        let mut candidates = TransitionCandidateVersions::default();
+        for page in pages {
+            candidates.extend("archive/object", page);
+        }
+        candidates.classify(bucket_versioning)
+    }
+
+    fn candidate_identity() -> TransitionCandidateIdentity {
+        TransitionCandidateIdentity {
+            transaction_id: uuid::Uuid::parse_str("aaaaaaaa-aaaa-4aaa-8aaa-aaaaaaaaaaaa").unwrap(),
+            destination_id: [0x5a; 32],
+        }
+    }
+
+    fn candidate_metadata(identity: TransitionCandidateIdentity) -> HashMap<String, String> {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            identity.transaction_id.to_string(),
+        );
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
+            rustfs_utils::crypto::hex(identity.destination_id),
+        );
+        metadata
+    }
+
+    #[test]
+    fn transition_candidate_identity_requires_exact_compatible_metadata() {
+        let identity = candidate_identity();
+        let metadata = candidate_metadata(identity);
+        assert!(transition_candidate_metadata_matches(&metadata, identity).unwrap());
+
+        let mut adjacent = metadata;
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut adjacent,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            uuid::Uuid::parse_str("bbbbbbbb-bbbb-4bbb-8bbb-bbbbbbbbbbbb")
+                .unwrap()
+                .to_string(),
+        );
+        assert!(!transition_candidate_metadata_matches(&adjacent, identity).unwrap());
+
+        assert!(!transition_candidate_metadata_matches(&HashMap::new(), identity).unwrap());
+    }
+
+    #[test]
+    fn transition_candidate_identity_rejects_conflicting_compatibility_keys() {
+        let identity = candidate_identity();
+        let mut metadata = candidate_metadata(identity);
+        metadata.insert(
+            rustfs_utils::http::metadata_compat::internal_key_rustfs(
+                rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            ),
+            uuid::Uuid::new_v4().to_string(),
+        );
+        assert!(transition_candidate_metadata_matches(&metadata, identity).is_err());
+    }
+
    #[test]
    fn transition_candidate_probe_classifier_is_fail_closed() {
        assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
-                RemoteBucketVersioning::Disabled,
-                &list_versions(&[], &[], false),
-            ),
+            classify_pages(RemoteBucketVersioning::Disabled, &[list_versions(&[], &[], false)],),
            TransitionCandidateProbe::Missing
        );
        assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
-                RemoteBucketVersioning::Disabled,
-                &list_versions(&[("archive/object", "")], &[], false),
-            ),
+            classify_pages(RemoteBucketVersioning::Disabled, &[list_versions(&[("archive/object", "")], &[], false)],),
            TransitionCandidateProbe::UnversionedPresent
        );
        assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
+            classify_pages(
                RemoteBucketVersioning::Enabled,
-                &list_versions(&[("archive/object", "version-a")], &[], false),
+                &[list_versions(&[("archive/object", "version-a")], &[], false)],
            ),
            TransitionCandidateProbe::VersionedPresent("version-a".to_string())
        );
        assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
+            classify_pages(
                RemoteBucketVersioning::Suspended,
-                &list_versions(&[("archive/object", "null")], &[], false),
+                &[list_versions(&[("archive/object", "null")], &[], false)],
            ),
            TransitionCandidateProbe::VersionedPresent("null".to_string())
        );
        assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
-                RemoteBucketVersioning::Enabled,
-                &list_versions(&[("archive/object", "")], &[], false),
-            ),
+            classify_pages(RemoteBucketVersioning::Enabled, &[list_versions(&[("archive/object", "")], &[], false)],),
            TransitionCandidateProbe::Ambiguous
        );
        assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
+            classify_pages(
                RemoteBucketVersioning::Enabled,
-                &list_versions(&[("archive/object", "version-a"), ("archive/object", "version-b")], &[], false),
+                &[list_versions(
+                    &[("archive/object", "version-a"), ("archive/object", "version-b")],
+                    &[],
+                    false,
+                )],
            ),
            TransitionCandidateProbe::Ambiguous
        );
+    }
+
+    #[test]
+    fn transition_candidate_probe_reconciles_all_pages_and_ignores_delete_markers() {
+        assert_eq!(
+            classify_pages(
+                RemoteBucketVersioning::Enabled,
+                &[
+                    list_versions(&[], &[("archive/object", "marker-a")], true),
+                    list_versions(&[("archive/object", "version-a"), ("archive/object-adjacent", "unrelated"),], &[], false,),
+                ],
+            ),
+            TransitionCandidateProbe::VersionedPresent("version-a".to_string())
+        );
        assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
+            classify_pages(
                RemoteBucketVersioning::Enabled,
-                &list_versions(&[("archive/object", "version-a")], &[("archive/object", "marker-a")], false),
-            ),
-            TransitionCandidateProbe::Ambiguous
-        );
-        assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
-                RemoteBucketVersioning::Enabled,
-                &list_versions(&[("archive/object", "version-a")], &[], true),
+                &[
+                    list_versions(&[("archive/object", "version-a")], &[], true),
+                    list_versions(&[("archive/object", "version-b")], &[], false),
+                ],
            ),
            TransitionCandidateProbe::Ambiguous
        );
    }

+    #[test]
+    fn transition_candidate_pagination_advances_both_markers() {
+        let mut key_marker = "old-key".to_string();
+        let mut version_id_marker = "old-version".to_string();
+        let page = ListVersionsResult {
+            next_key_marker: "next-key".to_string(),
+            next_version_id_marker: "next-version".to_string(),
+            ..Default::default()
+        };
+
+        advance_version_markers(&mut key_marker, &mut version_id_marker, &page)
+            .expect("new ListObjectVersions markers should advance pagination");
+        assert_eq!(key_marker, "next-key");
+        assert_eq!(version_id_marker, "next-version");
+
+        let err = advance_version_markers(&mut key_marker, &mut version_id_marker, &page)
+            .expect_err("repeated ListObjectVersions markers must fail closed");
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+    }
+
+    #[test]
+    fn transition_candidate_probe_rejects_untrusted_version_ids() {
+        let mut candidates = TransitionCandidateVersions::default();
+        candidates.extend("archive/object", &list_versions(&[("archive/object", "version\ninjection")], &[], false));
+
+        let err = classify_transition_candidates(candidates, RemoteBucketVersioning::Enabled)
+            .expect_err("control characters in listed version IDs must fail closed");
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+    }
+
    #[test]
    fn remote_bucket_versioning_status_parser_fails_closed() {
        assert_eq!(
@@ -397,12 +634,7 @@ impl WarmBackend for WarmBackendS3 {

    async fn probe_transition_candidate(&self, object: &str) -> Result<TransitionCandidateProbe, std::io::Error> {
        let bucket_versioning = self.remote_bucket_versioning().await?;
-        let versions = self.list_transition_candidate_versions(object).await?;
-        Ok(classify_transition_candidate_versions(
-            &self.get_dest(object),
-            bucket_versioning,
-            &versions,
-        ))
+        self.probe_transition_candidate_versions(object, bucket_versioning).await
    }

    async fn in_use(&self) -> Result<bool, std::io::Error> {
@@ -414,3 +646,16 @@ impl WarmBackend for WarmBackendS3 {
        Ok(result.common_prefixes.len() > 0 || result.contents.len() > 0)
    }
 }
+
+#[async_trait::async_trait]
+impl TransitionCandidateReconciler for WarmBackendS3 {
+    async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        identity: TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        let bucket_versioning = self.remote_bucket_versioning().await?;
+        self.probe_transition_candidate_identity(object, identity, bucket_versioning)
+            .await
+    }
+}
@@ -46,7 +46,11 @@ use crate::diagnostics::get::{
    GetObjectFailureReason, classify_disk_error, get_stage_timer_if_enabled, record_get_object_pipeline_failure,
    record_get_object_pipeline_failure_for_path, record_get_stage_duration_if_enabled,
 };
-use crate::disk::OldCurrentSize;
+use crate::disk::local::DELETE_DATA_DIR_MARKER_PREFIX;
+use crate::disk::{
+    DataDirDeleteStatus, OldCurrentSize, PART_TRANSACTION_NEW_META, PART_TRANSACTION_OLD_META, PART_TRANSACTION_ROLLBACK,
+    PartTransactionAction, part_transaction_path,
+};
 use crate::erasure::coding::BitrotReader;
 use crate::io_support::bitrot::ShardReader;
 use crate::io_support::bitrot::{
@@ -544,6 +548,8 @@ pub(in crate::set_disk) fn metadata_early_stop_candidate_matches(left: &FileInfo
        && left.transitioned_objname == right.transitioned_objname
        && left.transition_tier == right.transition_tier
        && left.transition_version_id == right.transition_version_id
+        && left.transition_version == right.transition_version
+        && left.transition_version_state == right.transition_version_state
        && left.expire_restored == right.expire_restored
        && left.size == right.size
        && left.mod_time == right.mod_time
@@ -2982,29 +2988,48 @@ impl SetDisks {
                Self::rename_fanout_barrier(&object_for_fault, idx, rename_fanout_barrier_phase::CLEANUP).await;

                if let Some(err) = Self::cleanup_injected_error(&object_for_fault, idx) {
-                    return Some(err);
+                    return (false, Some(err));
                }
                if let Some(disk) = disk {
-                    disk.delete(
-                        &bucket,
-                        &file_path,
-                        DeleteOptions {
-                            recursive: true,
-                            ..Default::default()
-                        },
-                    )
-                    .await
-                    .err()
+                    match disk
+                        .delete_data_dir(
+                            &bucket,
+                            &file_path,
+                            DeleteOptions {
+                                recursive: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                    {
+                        Ok(DataDirDeleteStatus::Deleted) => (false, None),
+                        Ok(DataDirDeleteStatus::Deferred) => (true, None),
+                        Err(err) => (false, Some(err)),
+                    }
                } else {
                    // `None` slot: ignored placeholder. It is not `attempted`, so
                    // classification excludes it from residue regardless.
-                    Some(DiskError::DiskNotFound)
+                    (false, Some(DiskError::DiskNotFound))
                }
            })
        });
-        let errs: Vec<Option<DiskError>> = join_all(futures).await.into_iter().map(map_cleanup_join_result).collect();
+        let mut deferred = 0usize;
+        let errs: Vec<Option<DiskError>> = join_all(futures)
+            .await
+            .into_iter()
+            .map(|result| match result {
+                Ok((was_deferred, err)) => {
+                    deferred += usize::from(was_deferred);
+                    err
+                }
+                Err(join_err) => Some(DiskError::other(format!("old data dir cleanup task failed: {join_err}"))),
+            })
+            .collect();

-        classify_old_data_dir_cleanup(&errs, &attempted, write_quorum)
+        let mut cleanup = classify_old_data_dir_cleanup(&errs, &attempted, write_quorum);
+        cleanup.deferred = deferred;
+        cleanup.reclaimed = cleanup.reclaimed.saturating_sub(deferred);
+        cleanup
    }

    /// Test-only fault-injection seam for the old-data-dir cleanup path
@@ -3095,6 +3120,20 @@ impl SetDisks {

        rustfs_io_metrics::record_old_data_dir_cleanup(c.attempted, c.reclaimed, c.unreclaimed_disks.len(), c.below_quorum);

+        if c.deferred > 0 {
+            debug!(
+                event = EVENT_SET_DISK_WRITE,
+                component = LOG_COMPONENT_ECSTORE,
+                subsystem = LOG_SUBSYSTEM_SET_DISK,
+                bucket = %bucket,
+                object = %object,
+                old_data_dir = %old_dir,
+                deferred = c.deferred,
+                state = "old_data_cleanup_deferred",
+                "Old data directory cleanup deferred for active snapshot leases"
+            );
+        }
+
        if actions.warn {
            warn!(
                component = LOG_COMPONENT_ECSTORE,
@@ -3174,6 +3213,155 @@ impl SetDisks {
        }
    }

+    async fn recover_part_transaction(&self, dst_object: &str, write_quorum: usize) -> disk::error::Result<bool> {
+        let disks = self.get_disks_internal().await;
+        let transaction_path = part_transaction_path(dst_object);
+        let transaction_meta_path = format!("{transaction_path}/{PART_TRANSACTION_NEW_META}");
+        let rollback_path = format!("{transaction_path}/{PART_TRANSACTION_ROLLBACK}");
+        let current_meta_path = format!("{dst_object}.meta");
+
+        let reads = disks.iter().map(|disk| {
+            let disk = disk.clone();
+            let transaction_meta_path = transaction_meta_path.clone();
+            let rollback_path = rollback_path.clone();
+            let current_meta_path = current_meta_path.clone();
+            async move {
+                let Some(disk) = disk else {
+                    return Ok((None, None, false));
+                };
+                let transaction_meta = match disk.read_all(RUSTFS_META_MULTIPART_BUCKET, &transaction_meta_path).await {
+                    Ok(meta) => Some(meta),
+                    Err(DiskError::FileNotFound) => None,
+                    Err(err) => return Err(err),
+                };
+                let rollback = match disk.read_all(RUSTFS_META_MULTIPART_BUCKET, &rollback_path).await {
+                    Ok(_) => true,
+                    Err(DiskError::FileNotFound) => false,
+                    Err(err) => return Err(err),
+                };
+                let current_meta = match disk.read_all(RUSTFS_META_MULTIPART_BUCKET, &current_meta_path).await {
+                    Ok(meta) => Some(meta),
+                    Err(DiskError::FileNotFound | DiskError::DiskNotFound) => None,
+                    Err(_) => None,
+                };
+                Ok((transaction_meta, current_meta, rollback))
+            }
+        });
+        let observations = join_all(reads).await.into_iter().collect::<disk::error::Result<Vec<_>>>()?;
+        if observations.iter().all(|(transaction, _, _)| transaction.is_none()) {
+            return Ok(false);
+        }
+
+        let mut current_counts: HashMap<Bytes, usize> = HashMap::new();
+        for (_, current, _) in &observations {
+            if let Some(current) = current {
+                *current_counts.entry(current.clone()).or_default() += 1;
+            }
+        }
+        let current_quorum = current_counts
+            .into_iter()
+            .find_map(|(meta, count)| (count >= write_quorum).then_some(meta));
+
+        let old_meta_path = format!("{transaction_path}/{PART_TRANSACTION_OLD_META}");
+        let old_meta_absent_path = format!("{transaction_path}/old.meta.absent");
+        let decisions = observations
+            .iter()
+            .enumerate()
+            .filter_map(|(index, (transaction_meta, _, rollback))| {
+                transaction_meta.as_ref().map(|meta| (index, meta.clone(), *rollback))
+            })
+            .map(|(index, transaction_meta, rollback)| {
+                let disk = disks[index].clone();
+                let old_meta_path = old_meta_path.clone();
+                let old_meta_absent_path = old_meta_absent_path.clone();
+                let current_quorum = current_quorum.clone();
+                async move {
+                    let Some(disk) = disk else {
+                        return Err(DiskError::DiskNotFound);
+                    };
+                    let action = if rollback {
+                        PartTransactionAction::Rollback
+                    } else if current_quorum.as_ref() == Some(&transaction_meta) {
+                        PartTransactionAction::Commit
+                    } else if let Some(current_quorum) = current_quorum {
+                        match disk.read_all(RUSTFS_META_MULTIPART_BUCKET, &old_meta_path).await {
+                            Ok(old_meta) if old_meta == current_quorum => PartTransactionAction::Rollback,
+                            Ok(_) => PartTransactionAction::Commit,
+                            Err(DiskError::FileNotFound) => {
+                                match disk.read_all(RUSTFS_META_MULTIPART_BUCKET, &old_meta_absent_path).await {
+                                    Ok(_) => PartTransactionAction::Commit,
+                                    Err(_) => return Err(DiskError::FileCorrupt),
+                                }
+                            }
+                            Err(err) => return Err(err),
+                        }
+                    } else {
+                        PartTransactionAction::Rollback
+                    };
+                    disk.settle_part_transaction(RUSTFS_META_MULTIPART_BUCKET, dst_object, action)
+                        .await?;
+                    Ok(action == PartTransactionAction::Commit)
+                }
+            });
+
+        let results = join_all(decisions).await;
+        if let Some(err) = results.iter().find_map(|result| result.as_ref().err()) {
+            return Err(err.clone());
+        }
+        Ok(results.iter().any(|result| matches!(result, Ok(true))))
+    }
+
+    pub(in crate::set_disk) async fn recover_part_transactions(
+        &self,
+        part_path: &str,
+        read_quorum: usize,
+        write_quorum: usize,
+    ) -> disk::error::Result<()> {
+        let disks = self.get_disks_internal().await;
+        let listings = join_all(disks.iter().map(|disk| {
+            let disk = disk.clone();
+            async move {
+                let Some(disk) = disk else {
+                    return Err(DiskError::DiskNotFound);
+                };
+                disk.list_dir(RUSTFS_META_MULTIPART_BUCKET, RUSTFS_META_MULTIPART_BUCKET, part_path, -1)
+                    .await
+            }
+        }))
+        .await;
+
+        let mut transaction_parts = HashSet::new();
+        let mut errs = Vec::with_capacity(listings.len());
+        for listing in listings {
+            match listing {
+                Ok(entries) => {
+                    errs.push(None);
+                    for entry in entries {
+                        let name = entry.trim_end_matches('/');
+                        let Some(part_number) = name
+                            .strip_prefix(".part.")
+                            .and_then(|name| name.strip_suffix(".rustfs-txn"))
+                            .and_then(|number| number.parse::<usize>().ok())
+                        else {
+                            continue;
+                        };
+                        transaction_parts.insert(part_number);
+                    }
+                }
+                Err(err) => errs.push(Some(err)),
+            }
+        }
+        if let Some(err) = reduce_read_quorum_errs(&errs, OBJECT_OP_IGNORED_ERRS, read_quorum) {
+            return Err(err);
+        }
+
+        for part_number in transaction_parts {
+            self.recover_part_transaction(&format!("{part_path}part.{part_number}"), write_quorum)
+                .await?;
+        }
+        Ok(())
+    }
+
    #[tracing::instrument(skip(disks, meta))]
    #[allow(clippy::too_many_arguments)]
    pub(in crate::set_disk) async fn rename_part(
@@ -3187,23 +3375,46 @@ 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 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());
        let dst_object = Arc::new(dst_object.to_string());

-        // Do NOT pre-delete the destination part before renaming: the per-disk
-        // `rename_part` replaces `part.N` atomically (std::fs::rename) and rewrites
-        // `part.N.meta`, so the pre-delete is redundant — and destructive. It
-        // opened a window where an already-committed (ACKed) part was removed on
-        // every disk before the new rename landed, so a re-upload that then failed
-        // quorum destroyed the committed part outright (backlog#853 / #799 B4).
-        // The atomic rename overwrites in place; on quorum failure below we roll
-        // the destination back.
+        let prepare_tasks = disks.iter().map(|disk| {
+            let disk = disk.clone();
+            let src_bucket = src_bucket.clone();
+            let src_object = src_object.clone();
+            let dst_bucket = dst_bucket.clone();
+            let dst_object = dst_object.clone();
+            let meta = meta.clone();
+            async move {
+                let disk = disk?;
+                Some(
+                    disk.prepare_part_transaction(&src_bucket, &src_object, &dst_bucket, &dst_object, meta)
+                        .await,
+                )
+            }
+        });
+        let prepare_results = join_all(prepare_tasks).await;
+        let prepare_errs = prepare_results
+            .into_iter()
+            .map(|result| match result {
+                Some(Ok(())) => None,
+                Some(Err(err)) => Some(err),
+                None => Some(DiskError::DiskNotFound),
+            })
+            .collect::<Vec<_>>();
+        let prepared_disks = Self::eval_disks(disks, &prepare_errs);
+        if reduce_write_quorum_errs(&prepare_errs, OBJECT_OP_IGNORED_ERRS, write_quorum).is_some() {
+            self.recover_part_transaction(&dst_object, write_quorum).await?;
+            return Err(DiskError::ErasureWriteQuorum);
+        }

        let mut errs = Vec::with_capacity(disks.len());

-        let futures = disks.iter().map(|disk| {
+        let futures = prepared_disks.iter().map(|disk| {
            let disk = disk.clone();
            let meta = meta.clone();
            let src_bucket = src_bucket.clone();
@@ -3253,19 +3464,42 @@ impl SetDisks {
            );
        }

-        if let Some(err) = reduce_write_quorum_errs(&errs, OBJECT_OP_IGNORED_ERRS, write_quorum) {
+        let reduced_err = reduce_write_quorum_errs(&errs, OBJECT_OP_IGNORED_ERRS, write_quorum);
+        if let Some(err) = reduced_err {
+            let rollbacks = prepared_disks.iter().filter_map(|disk| {
+                disk.clone().map(|disk| {
+                    let dst_object = dst_object.clone();
+                    async move {
+                        disk.settle_part_transaction(RUSTFS_META_MULTIPART_BUCKET, &dst_object, PartTransactionAction::Rollback)
+                            .await
+                    }
+                })
+            });
+            let rollback_results = join_all(rollbacks).await;
+            self.recover_part_transaction(&dst_object, write_quorum).await?;
+            if let Some(rollback_err) = rollback_results.iter().find_map(|result| result.as_ref().err()) {
+                warn!(error = %rollback_err, "rename_part rollback did not settle on every prepared disk");
+            }
            if let Some(context) = quorum_context {
                log_multipart_write_quorum_failure(context, &errs, write_quorum, &err);
            } else {
                warn!("rename_part errs {:?}", &errs);
            }
-            self.cleanup_multipart_path(&[dst_object.to_string(), format!("{dst_object}.meta")])
-                .await;
            return Err(err);
        }

-        let disks = Self::eval_disks(disks, &errs);
-        Ok(disks)
+        let committed = self.recover_part_transaction(&dst_object, write_quorum).await?;
+        if !committed {
+            let err = DiskError::ErasureWriteQuorum;
+            if let Some(context) = quorum_context {
+                log_multipart_write_quorum_failure(context, &errs, write_quorum, &err);
+            } else {
+                warn!("rename_part errs {:?}", &errs);
+            }
+            return Err(err);
+        }
+
+        Ok(Self::eval_disks(&prepared_disks, &errs))
    }

    pub(in crate::set_disk) fn eval_disks(disks: &[Option<DiskStore>], errs: &[Option<DiskError>]) -> Vec<Option<DiskStore>> {
@@ -3719,9 +3953,9 @@ impl SetDisks {
    /// * The set of referenced data dirs is the UNION of `get_data_dirs()` across
    ///   every online disk's `xl.meta`, so a dir named by *any* replica is kept.
    /// * If a disk holds the object directory but its `xl.meta` is missing or
-    ///   unparsable, the object is treated as degraded and NOTHING is removed —
-    ///   the unreadable copy could be the only one naming a live data dir, and a
-    ///   heal must run first.
+    ///   unparsable, the object is treated as degraded and unmarked data dirs are
+    ///   never removed. A data dir carrying a committed delete-transaction marker
+    ///   remains reclaimable after a downgrade/re-upgrade cleanup interruption.
    /// * Only subdirectories whose names parse as a UUID are ever considered;
    ///   removal is non-recursive-safe via a recursive delete of the full stray
    ///   data-dir path only.
@@ -3736,7 +3970,7 @@ impl SetDisks {
        // physical UUID subdirectories present on each disk. Abort on any degraded
        // copy so a healable object is never stripped of a referenced data dir.
        let mut referenced: HashSet<Uuid> = HashSet::new();
-        let mut per_disk_dirs: Vec<(usize, Vec<Uuid>)> = Vec::new();
+        let mut per_disk_dirs: Vec<(usize, Vec<(Uuid, bool)>)> = Vec::new();
        let mut healthy_metas = 0usize;

        for (i, disk) in disks.iter().enumerate() {
@@ -3772,6 +4006,22 @@ impl SetDisks {
                        // to the orphan-dir / dangling-object heal paths.
                        continue;
                    }
+                    let mut committed = Vec::with_capacity(physical.len());
+                    for dir in physical {
+                        let data_dir = format!("{object}/{dir}");
+                        let committed_delete = disk.list_dir("", bucket, &data_dir, 0).await.is_ok_and(|entries| {
+                            entries.iter().any(|entry| {
+                                entry
+                                    .strip_prefix(DELETE_DATA_DIR_MARKER_PREFIX)
+                                    .is_some_and(|transaction| Uuid::parse_str(transaction).is_ok())
+                            })
+                        });
+                        committed.push((dir, committed_delete));
+                    }
+                    if committed.iter().all(|(_, committed_delete)| *committed_delete) {
+                        per_disk_dirs.push((i, committed));
+                        continue;
+                    }
                    warn!(
                        target: "rustfs_ecstore::set_disk",
                        bucket, object,
@@ -3808,22 +4058,16 @@ impl SetDisks {

            healthy_metas += 1;
            if !physical.is_empty() {
-                per_disk_dirs.push((i, physical));
+                per_disk_dirs.push((i, physical.into_iter().map(|dir| (dir, false)).collect()));
            }
        }

-        // No healthy metadata anywhere: this is not a live object, so surplus dirs
-        // (if any) belong to the dangling-object heal path, not here.
-        if healthy_metas == 0 {
-            return Ok(0);
-        }
-
        // Phase 2: delete every physical data dir not referenced by the union.
        let mut removed = 0usize;
        for (i, physical) in per_disk_dirs {
            let Some(disk) = disks[i].as_ref() else { continue };
-            for dir in physical {
-                if referenced.contains(&dir) {
+            for (dir, committed_delete) in physical {
+                if referenced.contains(&dir) || (healthy_metas == 0 && !committed_delete) {
                    continue;
                }
                let stray = format!("{object}/{dir}");
@@ -3931,6 +4175,9 @@ pub(in crate::set_disk) struct OldDataDirCleanup {
    /// Number of attempted disks that returned `Ok` or a not-found variant
    /// (a missing dir == already reclaimed).
    pub reclaimed: usize,
+    /// Number of attempted disks that retained the directory for an active
+    /// snapshot lease and registered it for deletion after the final release.
+    pub deferred: usize,
    /// Indices of attempted disks that failed with a non-ignored, non-not-found
    /// error (including task panic/cancel). This is the residue that actually
    /// leaks and drives the leak metric + heal enqueue.
@@ -3993,6 +4240,7 @@ fn classify_old_data_dir_cleanup(errs: &[Option<DiskError>], attempted: &[bool],
    OldDataDirCleanup {
        attempted: attempted_count,
        reclaimed,
+        deferred: 0,
        unreclaimed_disks,
        below_quorum,
    }
@@ -4933,6 +5181,40 @@ mod tests {
        drop((disk1, disk2));
    }

+    #[tokio::test]
+    async fn commit_cleanup_reports_and_releases_deferred_snapshot_data_dirs() {
+        let bucket = "cleanup-lease-bucket";
+        let object = "cleanup-lease-object";
+        let old_data_dir = "11111111-1111-1111-1111-111111111111";
+        let committed_data_dir = "22222222-2222-2222-2222-222222222222";
+        let data_dir_path = format!("{object}/{old_data_dir}");
+        let shard_path = format!("{data_dir_path}/part.1");
+        let (_dir1, disk1) = read_multiple_test_disk(bucket, &[(&shard_path, b"one".as_slice())]).await;
+        let set = io_primitives_test_set(vec![Some(disk1.clone())], 0).await;
+        let lease = disk1
+            .acquire_snapshot_lease(bucket, &data_dir_path)
+            .await
+            .expect("snapshot lease should be acquired before cleanup");
+
+        let cleanup = set
+            .commit_rename_data_dir(&[Some(disk1.clone())], bucket, object, old_data_dir, committed_data_dir, 1)
+            .await;
+        assert_eq!(cleanup.attempted, 1);
+        assert_eq!(cleanup.reclaimed, 0);
+        assert_eq!(cleanup.deferred, 1);
+        assert!(cleanup.unreclaimed_disks.is_empty());
+        disk1
+            .read_all(bucket, &shard_path)
+            .await
+            .expect("deferred cleanup must leave later shard opens available");
+
+        disk1
+            .release_snapshot_lease(bucket, &data_dir_path, lease)
+            .await
+            .expect("final lease release should reclaim the old data directory");
+        assert!(matches!(disk1.read_all(bucket, &shard_path).await, Err(DiskError::FileNotFound)));
+    }
+
    /// Isolation guard: an armed barrier / observed object only affects its own
    /// object. A fan-out for a different (unobserved, unarmed) object must not be
    /// paused and must not accrue any tracked task count — so concurrent tests
@@ -555,7 +555,7 @@ impl SetDisks {
        }
    }

-    fn file_info_quorum_hash(meta: &FileInfo) -> [u8; 32] {
+    pub(super) fn file_info_quorum_hash(meta: &FileInfo) -> [u8; 32] {
        let mut hasher = Sha256::new();
        Self::update_file_info_quorum_hash(&mut hasher, meta);
        let digest = hasher.finalize();
@@ -578,6 +578,13 @@ impl SetDisks {
        Self::update_hash_str(hasher, &meta.transition_tier);
        Self::update_hash_str(hasher, &meta.transitioned_objname);
        Self::update_hash_optional_uuid(hasher, meta.transition_version_id);
+        Self::update_hash_optional_str(hasher, meta.transition_version.as_deref());
+        hasher.update([match meta.transition_version_state {
+            rustfs_filemeta::TransitionVersionState::Unknown => 0,
+            rustfs_filemeta::TransitionVersionState::KnownDisabled => 1,
+            rustfs_filemeta::TransitionVersionState::SuspendedNull => 2,
+            rustfs_filemeta::TransitionVersionState::Exact => 3,
+        }]);
        Self::update_hash_optional_u32(hasher, meta.mode);
        Self::update_hash_optional_u64(hasher, meta.written_by_version);

@@ -199,6 +199,32 @@ mod tests {
        .expect("healthy temp shard should verify");
        assert_eq!(verified, 1);

+        for trailing in [vec![0xa5], vec![0xa5; 17]] {
+            let mut oversized = encoded.to_vec();
+            oversized.extend_from_slice(&trailing);
+            disk.write_all(RUSTFS_META_TMP_BUCKET, path, Bytes::from(oversized))
+                .await
+                .expect("oversized temp shard should be staged");
+            let err = verify_written_bitrot_shards(
+                &[Some(disk.clone())],
+                None,
+                BitrotSelfVerifyTarget {
+                    operation: "put_object",
+                    bucket: "bucket",
+                    object: "object",
+                    part_number: None,
+                    volume: RUSTFS_META_TMP_BUCKET,
+                    path,
+                    logical_shard_size,
+                    shard_size,
+                    write_quorum: 1,
+                },
+            )
+            .await
+            .expect_err("disk shard with trailing bytes must not be committed");
+            assert!(err.to_string().contains("trailing data"));
+        }
+
        let mut corrupt = encoded.to_vec();
        let last = corrupt.len() - 1;
        corrupt[last] ^= 0x01;
@@ -225,4 +251,40 @@ mod tests {
        .expect_err("corrupt no-parity temp shard must not be committed");
        assert!(err.to_string().contains("bitrot self-verify failed"));
    }
+
+    #[tokio::test]
+    async fn no_parity_inline_self_verify_rejects_trailing_bytes() {
+        let (_temp_dirs, disks, _set_disks) = hermetic_set_disks_for_pool_with_default_parity(1, 0, 0).await;
+        let shard_size = 16usize;
+        let payload = b"inline bitrot payload";
+        let encoded = encode_streaming_shard(payload, shard_size).await;
+        let disks = disks.into_iter().map(Some).collect::<Vec<_>>();
+
+        for trailing in [vec![0xa5], vec![0xa5; 17]] {
+            let mut oversized = encoded.to_vec();
+            oversized.extend_from_slice(&trailing);
+            let parts = [FileInfo {
+                data: Some(Bytes::from(oversized)),
+                ..Default::default()
+            }];
+            let err = verify_written_bitrot_shards(
+                &disks,
+                Some(&parts),
+                BitrotSelfVerifyTarget {
+                    operation: "put_object",
+                    bucket: "bucket",
+                    object: "inline-object",
+                    part_number: None,
+                    volume: RUSTFS_META_TMP_BUCKET,
+                    path: "unused-for-inline",
+                    logical_shard_size: payload.len(),
+                    shard_size,
+                    write_quorum: 1,
+                },
+            )
+            .await
+            .expect_err("inline shard with trailing bytes must not be committed");
+            assert!(err.to_string().contains("trailing data"));
+        }
+    }
 }
@@ -69,6 +69,13 @@ struct HealWalkCollector {
 }

 impl HealWalkCollector {
+    fn lock_objects(&self) -> disk::error::Result<std::sync::MutexGuard<'_, Vec<HealWalkObject>>> {
+        self.objects.lock().map_err(|_| {
+            self.cancel.cancel();
+            DiskError::FileCorrupt
+        })
+    }
+
    /// Expand one resolved entry into its versions and record it. Cancels the
    /// walk once EITHER page bound (distinct object names OR expanded versions)
    /// is met — always at a sorted object-key boundary so a heavily-versioned
@@ -105,7 +112,9 @@ impl HealWalkCollector {

        let added = versions.len();
        let (objs_len, ver_total) = {
-            let mut objects = self.objects.lock().unwrap();
+            let Ok(mut objects) = self.lock_objects() else {
+                return;
+            };
            objects.push(HealWalkObject {
                name: entry.name,
                versions,
@@ -239,7 +248,10 @@ impl SetDisks {
            Err(err) => return Err(err),
        }

-        let objects = std::mem::take(&mut *collector.objects.lock().unwrap());
+        let objects = {
+            let mut objects = collector.lock_objects()?;
+            std::mem::take(&mut *objects)
+        };
        let truncated = collector.truncated.load(Ordering::SeqCst);
        Ok(finalize_heal_walk_page(objects, forward_to, truncated))
    }
@@ -310,4 +322,28 @@ mod tests {
        assert_eq!(next_forward, None, "a complete final page must not carry a resume cursor");
        assert!(!truncated);
    }
+
+    #[test]
+    fn poisoned_collector_state_cancels_the_walk() {
+        let collector = Arc::new(HealWalkCollector {
+            bucket: "bucket".to_string(),
+            batch_objects: 2,
+            version_budget: 2,
+            objects: Mutex::new(Vec::new()),
+            version_total: AtomicUsize::new(0),
+            truncated: AtomicBool::new(false),
+            cancel: CancellationToken::new(),
+        });
+        let poison_target = Arc::clone(&collector);
+        let _ = std::thread::spawn(move || {
+            let _guard = poison_target.objects.lock().expect("fresh mutex should lock");
+            panic!("poison heal collector");
+        })
+        .join();
+
+        let error = collector.lock_objects().expect_err("poisoned collection must fail closed");
+        assert_eq!(error, DiskError::FileCorrupt);
+        assert!(collector.cancel.is_cancelled(), "a poisoned page collector must cancel its walk");
+        assert!(collector.objects.lock().is_err(), "poisoned state must remain fail-closed");
+    }
 }
@@ -29,6 +29,12 @@ impl SetDisks {
    pub async fn delete_all(&self, bucket: &str, prefix: &str) -> Result<()> {
        ListOperations::new(self.ctx()).delete_all(bucket, prefix).await
    }
+
+    pub(crate) async fn delete_all_with_quorum(&self, bucket: &str, prefix: &str, write_quorum: usize) -> Result<()> {
+        ListOperations::new(self.ctx())
+            .delete_all_with_quorum(bucket, prefix, write_quorum)
+            .await
+    }
 }

 /// List/prefix maintenance operations, borrowing the `SetDisks` core state
@@ -48,6 +54,14 @@ impl<'a> ListOperations<'a> {
    }

    pub(crate) async fn delete_all(&self, bucket: &str, prefix: &str) -> Result<()> {
+        self.delete_all_inner(bucket, prefix, None).await
+    }
+
+    async fn delete_all_with_quorum(&self, bucket: &str, prefix: &str, write_quorum: usize) -> Result<()> {
+        self.delete_all_inner(bucket, prefix, Some(write_quorum)).await
+    }
+
+    async fn delete_all_inner(&self, bucket: &str, prefix: &str, write_quorum: Option<usize>) -> Result<()> {
        let disks = self.ctx.disks().read().await;

        let disks = disks.clone();
@@ -79,6 +93,9 @@ impl<'a> ListOperations<'a> {
                Ok(_) => {
                    errors.push(None);
                }
+                Err(DiskError::FileNotFound | DiskError::PathNotFound | DiskError::VolumeNotFound) => {
+                    errors.push(None);
+                }
                Err(e) => {
                    errors.push(Some(e));
                }
@@ -97,6 +114,12 @@ impl<'a> ListOperations<'a> {
            );
        }

+        if let Some(write_quorum) = write_quorum
+            && let Some(err) = reduce_write_quorum_errs(&errors, OBJECT_OP_IGNORED_ERRS, write_quorum)
+        {
+            return Err(err.into());
+        }
+
        Ok(())
    }
 }
@@ -207,11 +207,9 @@ impl SetDisks {
        version_id: &str,
        opts: &ReadOptions,
    ) -> Result<Vec<FileInfo>> {
-        // Use existing disk selection logic
-        let disks = self.disks.read().await;
-        let required_reads = self.format.erasure.sets.len();
+        let disks = self.disks.read().await.clone();
+        let required_reads = self.default_read_quorum();

-        // Clone parameters outside the closure to avoid lifetime issues
        let bucket = bucket.to_string();
        let object = object.to_string();
        let version_id = version_id.to_string();
@@ -220,7 +218,6 @@ impl SetDisks {
        let processor = runtime_sources::batch_processors().read_processor();
        let tasks: Vec<_> = disks
            .iter()
-            .take(required_reads + 2) // Read a few extra for reliability
            .filter_map(|disk| {
                disk.as_ref().map(|d| {
                    let disk = d.clone();
@@ -234,10 +231,10 @@ impl SetDisks {
            })
            .collect();

-        match processor.execute_batch_with_quorum(tasks, required_reads).await {
-            Ok(results) => Ok(results),
-            Err(_) => Err(DiskError::FileNotFound.into()), // Use existing error type
-        }
+        processor
+            .execute_batch_with_quorum(tasks, required_reads)
+            .await
+            .map_err(Into::into)
    }

    #[tracing::instrument(level = "debug", skip(self))]
@@ -1117,10 +1114,6 @@ impl SetDisks {

            total_read += part_length;
            part_offset = 0;
-            #[cfg(test)]
-            if current_part < last_part_index {
-                get_part_boundary_barrier::checkpoint(bucket, object, current_part).await;
-            }
        }

        // debug!("read end");
@@ -1824,79 +1817,6 @@ fn is_get_object_metadata_cache_request_eligible(bucket: &str, opts: &ObjectOpti
    get_object_metadata_cache_request_bypass_reason(bucket, opts, read_data).is_none()
 }

-#[cfg(test)]
-pub(in crate::set_disk) mod get_part_boundary_barrier {
-    use std::collections::HashMap;
-    use std::sync::{Arc, Mutex, OnceLock};
-    use tokio::sync::Notify;
-
-    struct Armed {
-        part_index: usize,
-        arrived: Arc<Notify>,
-        release: Arc<Notify>,
-    }
-
-    fn registry() -> &'static Mutex<HashMap<(String, String), Armed>> {
-        static REGISTRY: OnceLock<Mutex<HashMap<(String, String), Armed>>> = OnceLock::new();
-        REGISTRY.get_or_init(|| Mutex::new(HashMap::new()))
-    }
-
-    pub struct BarrierHandle {
-        key: (String, String),
-        arrived: Arc<Notify>,
-        release: Arc<Notify>,
-    }
-
-    pub fn arm(bucket: &str, object: &str, part_index: usize) -> BarrierHandle {
-        let key = (bucket.to_string(), object.to_string());
-        let arrived = Arc::new(Notify::new());
-        let release = Arc::new(Notify::new());
-        let previous = registry().lock().expect("GET part barrier registry poisoned").insert(
-            key.clone(),
-            Armed {
-                part_index,
-                arrived: Arc::clone(&arrived),
-                release: Arc::clone(&release),
-            },
-        );
-        assert!(previous.is_none(), "GET part barrier already armed for object");
-        BarrierHandle { key, arrived, release }
-    }
-
-    impl BarrierHandle {
-        pub async fn wait_until_paused(&self) {
-            self.arrived.notified().await;
-        }
-
-        pub fn release(&self) {
-            self.release.notify_one();
-        }
-    }
-
-    impl Drop for BarrierHandle {
-        fn drop(&mut self) {
-            self.release.notify_one();
-            registry()
-                .lock()
-                .expect("GET part barrier registry poisoned")
-                .remove(&self.key);
-        }
-    }
-
-    pub async fn checkpoint(bucket: &str, object: &str, part_index: usize) {
-        let state = registry()
-            .lock()
-            .expect("GET part barrier registry poisoned")
-            .get(&(bucket.to_string(), object.to_string()))
-            .filter(|armed| armed.part_index == part_index)
-            .map(|armed| (Arc::clone(&armed.arrived), Arc::clone(&armed.release)));
-        if let Some((arrived, release)) = state {
-            arrived.notify_one();
-            release.notified().await;
-        }
-    }
-}
-
 #[cfg(test)]
 mod metadata_cache_tests {
    use super::*;
@@ -2037,6 +1957,42 @@ mod metadata_cache_tests {
        (dir, disk)
    }

+    async fn read_version_quorum_test_set(
+        bucket: &str,
+        object: &str,
+        pool_set_count: usize,
+        readable_disks: usize,
+    ) -> (Vec<tempfile::TempDir>, Arc<SetDisks>) {
+        let mut dirs = Vec::new();
+        let mut disks = Vec::new();
+        for disk_index in 0..4 {
+            let (dir, disk) = new_read_version_test_disk(bucket).await;
+            if disk_index < readable_disks {
+                let mut fi = valid_test_fileinfo(object);
+                fi.mod_time = Some(OffsetDateTime::now_utc());
+                fi.erasure.index = fi.erasure.distribution[disk_index];
+                disk.write_metadata(bucket, bucket, object, fi)
+                    .await
+                    .expect("metadata should be written before quorum read");
+            }
+            dirs.push(dir);
+            disks.push(Some(disk));
+        }
+        let set = SetDisks::new(
+            "read-version-quorum-test".to_string(),
+            Arc::new(RwLock::new(disks)),
+            4,
+            2,
+            0,
+            0,
+            Vec::new(),
+            FormatV3::new(pool_set_count, 4),
+            Vec::new(),
+        )
+        .await;
+        (dirs, set)
+    }
+
    #[test]
    #[serial]
    fn get_object_metadata_cache_capacity_uses_default_and_env_override() {
@@ -2060,9 +2016,40 @@ mod metadata_cache_tests {
        fi.size = 1;
        fi.erasure.index = 1;
        fi.metadata.insert("etag".to_string(), "etag-1".to_string());
+        fi.add_object_part(1, "part-etag".to_string(), 1, fi.mod_time, 1, None, None);
        fi
    }

+    #[tokio::test]
+    async fn get_object_with_fileinfo_rejects_positive_size_without_parts() {
+        let mut fi = valid_test_fileinfo("object");
+        fi.parts.clear();
+
+        let mut output = Vec::new();
+        let err = SetDisks::get_object_with_fileinfo(
+            "bucket",
+            "object",
+            0,
+            1,
+            &mut output,
+            fi,
+            Vec::new(),
+            &[],
+            0,
+            0,
+            false,
+            false,
+            GET_OBJECT_PATH_SET_DISK,
+            "plain",
+            "small",
+        )
+        .await
+        .expect_err("positive-size metadata without parts must fail without panicking");
+
+        assert_eq!(err, Error::FileCorrupt);
+        assert!(output.is_empty());
+    }
+
    #[tokio::test]
    async fn get_object_with_fileinfo_rejects_invalid_ranges_before_reader_setup() {
        let bucket = "bucket";
@@ -2164,6 +2151,38 @@ mod metadata_cache_tests {
        assert!(output.is_empty());
    }

+    #[tokio::test]
+    async fn get_object_with_fileinfo_accepts_zero_size_without_parts() {
+        let bucket = "bucket";
+        let object = "empty";
+        let mut fi = valid_test_fileinfo(object);
+        fi.size = 0;
+        fi.parts.clear();
+
+        let mut output = Vec::new();
+        SetDisks::get_object_with_fileinfo(
+            bucket,
+            object,
+            0,
+            0,
+            &mut output,
+            fi,
+            Vec::new(),
+            &[],
+            0,
+            0,
+            false,
+            false,
+            GET_OBJECT_PATH_SET_DISK,
+            "plain",
+            "empty",
+        )
+        .await
+        .expect("zero-byte object without parts must remain readable");
+
+        assert!(output.is_empty());
+    }
+
    #[tokio::test]
    async fn get_object_with_fileinfo_fails_closed_without_read_quorum() {
        let bucket = "bucket";
@@ -2171,12 +2190,6 @@ mod metadata_cache_tests {
        let mut fi = valid_test_fileinfo(object);
        fi.erasure.block_size = 1;
        fi.erasure.distribution = vec![1, 2, 3, 4];
-        fi.parts.push(ObjectPartInfo {
-            number: 1,
-            size: 1,
-            actual_size: 1,
-            ..Default::default()
-        });

        let mut output = Vec::new();
        let err = SetDisks::get_object_with_fileinfo(
@@ -2210,6 +2223,7 @@ mod metadata_cache_tests {
        let object = "object";
        let (_dir, disk) = new_read_version_test_disk(bucket).await;
        let mut fi = valid_test_fileinfo(object);
+        fi.size = 0;
        fi.mod_time = Some(OffsetDateTime::now_utc());
        disk.write_metadata(bucket, bucket, object, fi.clone())
            .await
@@ -2242,11 +2256,70 @@ mod metadata_cache_tests {
            .await
            .expect_err("empty disk set must fail closed");
        assert!(
-            missing_quorum.to_string().to_ascii_lowercase().contains("file"),
-            "optimized read failure should map to file-not-found style error: {missing_quorum}"
+            missing_quorum.to_string().contains("Insufficient successful results"),
+            "optimized read failure should preserve the batch quorum diagnostic: {missing_quorum}"
        );
    }

+    #[tokio::test]
+    async fn read_version_optimized_uses_current_set_drive_quorum_not_pool_set_count() {
+        let bucket = "read-version-layout-quorum-bucket";
+        let object = "object";
+
+        let (_single_set_dirs, single_set) = read_version_quorum_test_set(bucket, object, 1, 1).await;
+        single_set
+            .read_version_optimized(bucket, object, "", &ReadOptions::default())
+            .await
+            .expect_err("one metadata copy must not satisfy a four-drive 2+2 set");
+
+        let (_multi_set_dirs, multi_set) = read_version_quorum_test_set(bucket, object, 3, 2).await;
+        let versions = multi_set
+            .read_version_optimized(bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("two metadata copies should satisfy the current set's read quorum");
+        assert_eq!(versions.len(), 2);
+        assert!(versions.iter().all(|version| version.name == object));
+    }
+
+    #[tokio::test]
+    async fn read_version_optimized_counts_only_valid_metadata_toward_quorum() {
+        let bucket = "read-version-corrupt-quorum-bucket";
+        let object = "object";
+
+        let (quorum_minus_one_dirs, quorum_minus_one) = read_version_quorum_test_set(bucket, object, 1, 2).await;
+        tokio::fs::write(
+            quorum_minus_one_dirs[1]
+                .path()
+                .join(bucket)
+                .join(object)
+                .join(crate::disk::STORAGE_FORMAT_FILE),
+            b"corrupt metadata",
+        )
+        .await
+        .expect("metadata should be corrupted for the test");
+        quorum_minus_one
+            .read_version_optimized(bucket, object, "", &ReadOptions::default())
+            .await
+            .expect_err("one valid and one corrupt metadata copy must not satisfy read quorum");
+
+        let (quorum_dirs, quorum) = read_version_quorum_test_set(bucket, object, 1, 3).await;
+        tokio::fs::write(
+            quorum_dirs[2]
+                .path()
+                .join(bucket)
+                .join(object)
+                .join(crate::disk::STORAGE_FORMAT_FILE),
+            b"corrupt metadata",
+        )
+        .await
+        .expect("metadata should be corrupted for the test");
+        let versions = quorum
+            .read_version_optimized(bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("two valid metadata copies must satisfy read quorum despite one corrupt copy");
+        assert_eq!(versions.len(), 2);
+    }
+
    #[tokio::test]
    async fn get_object_info_and_quorum_maps_delete_marker_and_purge_states() {
        let bucket = "get-object-info-marker-bucket";
@@ -13,7 +13,10 @@
 // limitations under the License.

 use super::*;
+use crate::bucket::lifecycle::lifecycle;
+use rustfs_filemeta::RestoreStatusOps;
 use rustfs_utils::http::headers::{AMZ_RESTORE_EXPIRY_DAYS, AMZ_RESTORE_REQUEST_DATE};
+use s3s::dto::{RestoreStatus, Timestamp};

 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 struct RestoreCleanupIdentity {
@@ -43,6 +46,69 @@ impl RestoreCleanupIdentity {
 }

 impl SetDisks {
+    pub(super) async fn finalize_restore_metadata(
+        &self,
+        bucket: &str,
+        object: &str,
+        obj_info: &ObjectInfo,
+        opts: &ObjectOptions,
+    ) -> Result<ObjectInfo> {
+        let expected = RestoreCleanupIdentity::from_object_info(obj_info);
+        let expected_operation_id = restore_operation_id_from_metadata(&opts.user_defined)?;
+        let expected_etag = obj_info
+            .etag
+            .clone()
+            .unwrap_or_else(|| get_raw_etag(obj_info.user_defined.as_ref()));
+        let version_id = expected.version_id.map(|v| v.to_string());
+        let _lock_guard = if !opts.no_lock {
+            Some(
+                self.acquire_write_lock_diag("restore_finalize_metadata", bucket, object)
+                    .await?,
+            )
+        } else {
+            None
+        };
+        let read_opts = ObjectOptions {
+            version_id,
+            versioned: opts.versioned,
+            version_suspended: opts.version_suspended,
+            ..Default::default()
+        };
+        let (mut fi, _, disks) = self
+            .get_object_fileinfo_gated(bucket, object, &read_opts, false, false)
+            .await?;
+        if let Some(expected_operation_id) = expected_operation_id {
+            require_restore_operation_id(&fi.metadata, expected_operation_id)?;
+        }
+        if !expected.matches_file_info(&fi, &expected_etag) {
+            return Err(Error::other("restored object changed before restore metadata finalization"));
+        }
+        let restore_expiry =
+            lifecycle::expected_expiry_time(OffsetDateTime::now_utc(), opts.transition.restore_request.days.unwrap_or(1));
+        fi.metadata.insert(
+            X_AMZ_RESTORE.as_str().to_string(),
+            RestoreStatus {
+                is_restore_in_progress: Some(false),
+                restore_expiry_date: Some(Timestamp::from(restore_expiry)),
+            }
+            .to_string(),
+        );
+        self.invalidate_get_object_metadata_cache(bucket, object).await;
+        self.update_object_meta_with_opts(
+            bucket,
+            object,
+            fi.clone(),
+            disks.as_slice(),
+            &UpdateMetadataOpts {
+                replace_user_metadata: true,
+                ..Default::default()
+            },
+        )
+        .await?;
+        self.invalidate_get_object_metadata_cache(bucket, object).await;
+        Ok(ObjectInfo::from_file_info(&fi, bucket, object, opts.versioned || opts.version_suspended))
+    }
+
    pub async fn update_restore_metadata(
        &self,
        bucket: &str,
@@ -13,10 +13,11 @@
 // limitations under the License.

 use super::*;
-use crate::bucket::lifecycle::lifecycle::{TRANSITION_COMPLETE, TRANSITION_PENDING, TransitionOptions};
+use crate::bucket::lifecycle::lifecycle::{TRANSITION_COMPLETE, TRANSITION_PENDING, TransitionOptions, expected_expiry_time};
 use crate::ecstore_validation_blackbox::make_local_set_disks;
 use crate::services::tier::test_util::register_mock_tier;
 use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations as _};
+use rustfs_filemeta::{RestoreStatusOps as _, parse_restore_obj_status};
 use tokio::io::AsyncReadExt;

 async fn prime_metadata_generation(set_disks: &SetDisks, bucket: &str, object: &str) -> GetObjectMetadataCacheKey {
@@ -83,13 +84,57 @@ async fn transition_and_restore_reclaim_prior_metadata_generations() {
    let transitioned_generation = prime_metadata_generation(&set_disks, bucket, object).await;
    let mut restore_opts = ObjectOptions::default();
    restore_opts.transition.restore_request.days = Some(1);
-    Arc::clone(&set_disks)
-        .restore_transitioned_object(bucket, object, &restore_opts)
-        .await
-        .expect("restore should succeed");
+    let restore_started = OffsetDateTime::now_utc();
+    let expiry_from_restore_start = temp_env::async_with_vars(
+        [
+            ("RUSTFS_ILM_DEBUG_DAY_SECS", Some("1")),
+            ("RUSTFS_ILM_PROCESS_TIME", Some("1")),
+        ],
+        async {
+            let expiry_from_restore_start = expected_expiry_time(restore_started, 1);
+            let get_barrier = backend.arm_get_barrier().await;
+            let restore_set = Arc::clone(&set_disks);
+            let restore =
+                tokio::spawn(async move { restore_set.restore_transitioned_object(bucket, object, &restore_opts).await });
+            get_barrier.wait_until_paused().await;
+            tokio::time::timeout(Duration::from_secs(5), async {
+                loop {
+                    if expected_expiry_time(OffsetDateTime::now_utc(), 1) > expiry_from_restore_start {
+                        break;
+                    }
+                    tokio::time::sleep(Duration::from_millis(10)).await;
+                }
+            })
+            .await
+            .expect("test clock should cross the next accelerated lifecycle boundary");
+            get_barrier.release();
+            restore
+                .await
+                .expect("restore task should join")
+                .expect("restore should succeed");
+            expiry_from_restore_start
+        },
+    )
+    .await;
    assert_generation_reclaimed(&set_disks, &transitioned_generation).await;
    assert_eq!(backend.get_count().await, 1, "restore should read the remote candidate exactly once");

+    let restored_info = set_disks
+        .get_object_info(bucket, object, &ObjectOptions::default())
+        .await
+        .expect("restored object metadata should be readable");
+    let restore_status = parse_restore_obj_status(
+        restored_info
+            .user_defined
+            .get(s3s::header::X_AMZ_RESTORE.as_str())
+            .expect("completed restore header should be present"),
+    )
+    .expect("completed restore header should parse");
+    assert!(
+        restore_status.expiry().expect("completed restore should have an expiry") > expiry_from_restore_start,
+        "restore expiry must be based on completion, not the time the remote copy started"
+    );
+
    let mut restored = Vec::new();
    set_disks
        .get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default())
@@ -87,7 +87,7 @@ fn bucket_deleted_marker_volume(bucket: &str) -> String {
    format!("{RUSTFS_META_BUCKET}/{}", bucket_deleted_marker_prefix(bucket))
 }

-async fn await_bucket_namespace_operation<T, F>(
+pub(crate) async fn await_bucket_namespace_operation<T, F>(
    guard: Option<&rustfs_lock::NamespaceLockGuard>,
    bucket: &str,
    operation: &'static str,
@@ -310,12 +310,13 @@ impl ECStore {
        meta.set_created(opts.created_at);

        if opts.lock_enabled {
-            meta.object_lock_config_xml = crate::bucket::utils::serialize::<ObjectLockConfiguration>(&enableObjcetLockConfig)?;
-            meta.versioning_config_xml = crate::bucket::utils::serialize::<VersioningConfiguration>(&enableVersioningConfig)?;
+            meta.object_lock_config_xml =
+                crate::bucket::utils::serialize::<ObjectLockConfiguration>(&ENABLED_OBJECT_LOCK_CONFIG)?;
+            meta.versioning_config_xml = crate::bucket::utils::serialize::<VersioningConfiguration>(&ENABLED_VERSIONING_CONFIG)?;
        }

        if opts.versioning_enabled {
-            meta.versioning_config_xml = crate::bucket::utils::serialize::<VersioningConfiguration>(&enableVersioningConfig)?;
+            meta.versioning_config_xml = crate::bucket::utils::serialize::<VersioningConfiguration>(&ENABLED_VERSIONING_CONFIG)?;
        }

        await_bucket_namespace_operation(
@@ -568,6 +569,7 @@ mod tests {
    };
    use crate::bucket::metadata::table_bucket_catalog_metadata_prefix;
    use crate::bucket::metadata_sys;
+    use crate::cluster::rpc::peer_s3_client::install_delete_bucket_empty_scan_barrier;
    use crate::disk::{BUCKET_META_PREFIX, RUSTFS_META_BUCKET};
    use crate::error::StorageError;
    use crate::object_api::{ObjectOptions, PutObjReader};
@@ -589,6 +591,7 @@ mod tests {
    use std::sync::atomic::{AtomicBool, Ordering};
    use std::time::{Duration, SystemTime};
    use time::OffsetDateTime;
+    use tokio::io::AsyncReadExt;
    use tokio::sync::{Notify, OnceCell};
    use tokio_util::sync::CancellationToken;
    use uuid::Uuid;
@@ -1048,7 +1051,7 @@ mod tests {
            .await
            .expect("bucket should be created");
        for disk_path in &disk_paths {
-            tokio::fs::create_dir_all(disk_path.join(&bucket).join("empty-directory"))
+            tokio::fs::create_dir_all(disk_path.join(&bucket).join("empty-directory/nested/leaf"))
                .await
                .expect("empty directory remnant should be created");
        }
@@ -1257,6 +1260,55 @@ mod tests {
        );
    }

+    #[tokio::test]
+    #[serial]
+    async fn bucket_delete_preserves_put_committed_after_empty_scan() {
+        let (_, ecstore) = setup_bucket_delete_test_env().await;
+        let bucket = format!("bucket-delete-empty-scan-race-{}", Uuid::new_v4().simple());
+        let object = "committed-after-empty-scan";
+        let payload = b"object committed after DeleteBucket empty scan".to_vec();
+
+        ecstore
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("bucket should be created");
+
+        let barrier = install_delete_bucket_empty_scan_barrier();
+        let delete_store = ecstore.clone();
+        let delete_bucket = bucket.clone();
+        let delete = tokio::spawn(async move {
+            delete_store
+                .delete_bucket(&delete_bucket, &DeleteBucketOptions::default())
+                .await
+        });
+        barrier.wait_until_paused().await;
+
+        let mut put_reader = PutObjReader::from_vec(payload.clone());
+        ecstore
+            .put_object(&bucket, object, &mut put_reader, &ObjectOptions::default())
+            .await
+            .expect("PUT should commit while DeleteBucket is paused after its empty scan");
+
+        barrier.release();
+        let err = delete
+            .await
+            .expect("DeleteBucket task should join")
+            .expect_err("DeleteBucket must reject a PUT committed after its empty scan");
+        assert!(matches!(err, StorageError::BucketNotEmpty(name) if name == bucket));
+
+        let mut reader = ecstore
+            .get_object_reader(&bucket, object, None, http::HeaderMap::new(), &ObjectOptions::default())
+            .await
+            .expect("committed object should remain readable after DeleteBucket fails");
+        let mut restored = Vec::new();
+        reader
+            .stream
+            .read_to_end(&mut restored)
+            .await
+            .expect("object body should remain readable");
+        assert_eq!(restored, payload);
+    }
+
    #[tokio::test]
    #[serial]
    async fn bucket_recreation_does_not_publish_unverified_usage() {
@@ -562,10 +562,12 @@ mod tests {
            },
            tier_sweeper::Jentry,
            transition_transaction::{
-                TRANSITION_TRANSACTION_RECORD_PREFIX, TransitionCleanupDecision, TransitionCleanupProof, TransitionRemoteVersion,
-                TransitionSourceIdentity, TransitionSourceVersionMode, TransitionTransaction, TransitionTransactionInit,
-                TransitionTransactionState, load_transition_transaction_record, recover_transition_transaction_records,
-                save_transition_transaction_record,
+                TRANSITION_TRANSACTION_RECORD_PREFIX, TransitionCleanupDecision, TransitionCleanupProof, TransitionOperatorError,
+                TransitionOperatorProbe, TransitionRemoteVersion, TransitionSourceIdentity, TransitionSourceVersionMode,
+                TransitionTransaction, TransitionTransactionInit, TransitionTransactionState,
+                delete_transition_candidate_for_operator, finalize_missing_transition_transaction_for_operator,
+                inspect_transition_transaction_for_operator, load_transition_transaction_record,
+                recover_transition_transaction_records, save_transition_transaction_record,
            },
        },
        client::transition_api::ReaderImpl,
@@ -575,6 +577,7 @@ mod tests {
        services::tier::{
            test_util::{MockWarmBackend, MockWarmOp, TransitionCleanupStoreBarrier, register_mock_tier},
            tier::{TIER_CONFIG_FILE, TierConfigMgr},
+            tier_config::{TierConfig, TierType, TierWasabi},
            tier_mutation_intent::{
                TIER_MUTATION_INTENT_RECORD_PREFIX, TierMutationIntent, TierMutationIntentKind, TierMutationIntentState,
                TierMutationIntentTarget, advance_tier_mutation_intent_record_idempotent, delete_tier_mutation_intent_record,
@@ -1163,6 +1166,37 @@ mod tests {
            .len()
    }

+    #[cfg(feature = "test-util")]
+    async fn register_transition_reconcile_test_tier(
+        handle: &Arc<tokio::sync::RwLock<TierConfigMgr>>,
+        tier_name: &str,
+    ) -> MockWarmBackend {
+        let backend = MockWarmBackend::new();
+        let mut manager = handle.write().await;
+        manager.tiers.insert(
+            tier_name.to_string(),
+            TierConfig {
+                version: "v1".to_string(),
+                tier_type: TierType::Wasabi,
+                name: tier_name.to_string(),
+                wasabi: Some(TierWasabi {
+                    name: tier_name.to_string(),
+                    endpoint: "https://s3.wasabisys.com".to_string(),
+                    access_key: "test-access-key".to_string(),
+                    secret_key: "test-secret-key".to_string(),
+                    bucket: "mock-tier".to_string(),
+                    prefix: format!("mock/{}/", uuid::Uuid::new_v4()),
+                    region: "us-east-1".to_string(),
+                }),
+                ..Default::default()
+            },
+        );
+        manager
+            .install_test_driver(tier_name, Box::new(backend.clone()))
+            .expect("mock fallback tier driver should install");
+        backend
+    }
+
    #[cfg(feature = "test-util")]
    async fn wait_for_tier_delete_journal_recovery(
        store: Arc<crate::store::ECStore>,
@@ -1311,14 +1345,16 @@ mod tests {
            version_id: "version-a".to_string(),
            tier_name: tier_a.to_string(),
            backend_identity: Some(identity_a),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
        };
        let entry_b = Jentry {
            obj_name: "remote-b".to_string(),
            version_id: "version-b".to_string(),
            tier_name: tier_b.to_string(),
            backend_identity: Some(identity_b),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
        };
        let remove_a = backend_a.arm_failing_remove_barrier().await;
        persist_tier_delete_journal_entry(store_a.clone(), &entry_a)
@@ -2720,10 +2756,12 @@ mod tests {
    #[serial_test::serial(storage_class_env)]
    async fn transition_transaction_recovery_deletes_provider_recovered_unknown_upload() {
        let versioned_remote = uuid::Uuid::new_v4().to_string();
+        let nil_remote = uuid::Uuid::nil().to_string();
        for (case, tier_name, remote_version) in [
            ("missing", "TXPROBEMISSING", None),
            ("unversioned", "TXPROBEUNVERSIONED", Some(String::new())),
            ("versioned", "TXPROBEVERSIONED", Some(versioned_remote)),
+            ("nil-version", "TXPROBENILVERSION", Some(nil_remote)),
        ] {
            let temp_dir = tempfile::tempdir().expect("create temp store dir");
            let (ctx, store, _shutdown) = without_storage_class_env(build_isolated_test_store(
@@ -2734,7 +2772,7 @@ mod tests {
            .await;
            crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;

-            let backend = register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
+            let backend = register_transition_reconcile_test_tier(&ctx.tier_config_mgr(), tier_name).await;
            let backend_identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
                .await
                .expect("tier lease should resolve")
@@ -2805,6 +2843,157 @@ mod tests {
        }
    }

+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn operator_reconcile_deletes_exact_candidate_before_finalizing_record() {
+        let temp_dir = tempfile::tempdir().expect("create temp store dir");
+        let (ctx, store, _shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "transition-operator-reconcile", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+        let tier_name = "TXOPERATOR";
+        let backend = register_transition_reconcile_test_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let backend_identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
+            .await
+            .expect("tier lease should resolve")
+            .backend_identity();
+        let mut transaction = TransitionTransaction::new(TransitionTransactionInit {
+            deployment_id: ctx.deployment_id().expect("test store should initialize deployment id"),
+            transaction_id: uuid::Uuid::new_v4(),
+            owner_epoch: uuid::Uuid::new_v4(),
+            write_id: uuid::Uuid::new_v4(),
+            source: TransitionSourceIdentity {
+                bucket: "source-bucket".to_string(),
+                object: "source-object".to_string(),
+                version_id: None,
+                data_dir: uuid::Uuid::new_v4(),
+                mod_time_unix_nanos: 1_770_000_000_000_000_000,
+                size: 42,
+                etag: "source-etag".to_string(),
+                version_mode: TransitionSourceVersionMode::Unversioned,
+            },
+            tier_name: tier_name.to_string(),
+            backend_fingerprint: backend_identity,
+            not_after_unix_nanos: 1,
+        })
+        .expect("transaction should build");
+        transaction
+            .advance(transaction.fence(), TransitionTransactionState::UploadOutcomeUnknown, None)
+            .expect("transaction should enter unknown upload outcome state");
+
+        let remote_version = uuid::Uuid::new_v4().to_string();
+        backend.set_put_remote_version(Some(remote_version.clone())).await;
+        let candidate = bytes::Bytes::from_static(b"operator-confirmed transition candidate");
+        backend
+            .put(
+                &transaction.remote_object,
+                ReaderImpl::Body(candidate.clone()),
+                i64::try_from(candidate.len()).expect("test candidate length should fit i64"),
+            )
+            .await
+            .expect("mock backend should accept candidate");
+        save_transition_transaction_record(store.clone(), &transaction)
+            .await
+            .expect("transaction record should persist");
+
+        let status = inspect_transition_transaction_for_operator(store.clone(), transaction.transaction_id)
+            .await
+            .expect("operator inspection should probe the candidate");
+        assert_eq!(status.probe, TransitionOperatorProbe::VersionedPresent(remote_version.clone()));
+
+        let wrong_version = uuid::Uuid::new_v4().to_string();
+        let err = delete_transition_candidate_for_operator(store.clone(), transaction.transaction_id, &wrong_version)
+            .await
+            .expect_err("a mismatched exact version must fail before deleting a candidate");
+        assert!(matches!(
+            err,
+            TransitionOperatorError::CandidateVersionMismatch {
+                expected,
+                actual: TransitionOperatorProbe::VersionedPresent(ref observed),
+            } if expected == wrong_version && observed == &remote_version
+        ));
+        assert!(backend.contains(&transaction.remote_object).await);
+        assert_eq!(backend.exact_remove_count(), 0);
+        load_transition_transaction_record(store.clone(), transaction.transaction_id)
+            .await
+            .expect("an incorrect exact version must retain the transaction journal");
+
+        let result = delete_transition_candidate_for_operator(store.clone(), transaction.transaction_id, &remote_version)
+            .await
+            .expect("operator-confirmed exact candidate should be deleted");
+        assert_eq!(result.status.probe, TransitionOperatorProbe::Missing);
+        assert!(result.journal_observed_after_delete);
+        assert_eq!(backend.exact_remove_count(), 1);
+        assert_eq!(backend.remove_versions().await, vec![(transaction.remote_object.clone(), remote_version)]);
+        load_transition_transaction_record(store.clone(), transaction.transaction_id)
+            .await
+            .expect("candidate deletion must retain the transaction journal");
+
+        finalize_missing_transition_transaction_for_operator(store.clone(), transaction.transaction_id)
+            .await
+            .expect("a separately confirmed missing candidate should permit finalization");
+        assert!(matches!(
+            load_transition_transaction_record(store, transaction.transaction_id).await,
+            Err(Error::ConfigNotFound)
+        ));
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn operator_finalize_retains_record_without_missing_proof() {
+        let temp_dir = tempfile::tempdir().expect("create temp store dir");
+        let (ctx, store, _shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "transition-operator-fail-closed", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+        let tier_name = "TXOPERATORFAIL";
+        let backend = register_transition_reconcile_test_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let backend_identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
+            .await
+            .expect("tier lease should resolve")
+            .backend_identity();
+        let mut transaction = TransitionTransaction::new(TransitionTransactionInit {
+            deployment_id: ctx.deployment_id().expect("test store should initialize deployment id"),
+            transaction_id: uuid::Uuid::new_v4(),
+            owner_epoch: uuid::Uuid::new_v4(),
+            write_id: uuid::Uuid::new_v4(),
+            source: TransitionSourceIdentity {
+                bucket: "source-bucket".to_string(),
+                object: "source-object".to_string(),
+                version_id: None,
+                data_dir: uuid::Uuid::new_v4(),
+                mod_time_unix_nanos: 1_770_000_000_000_000_000,
+                size: 42,
+                etag: "source-etag".to_string(),
+                version_mode: TransitionSourceVersionMode::Unversioned,
+            },
+            tier_name: tier_name.to_string(),
+            backend_fingerprint: backend_identity,
+            not_after_unix_nanos: 1,
+        })
+        .expect("transaction should build");
+        transaction
+            .advance(transaction.fence(), TransitionTransactionState::UploadOutcomeUnknown, None)
+            .expect("transaction should enter unknown upload outcome state");
+        save_transition_transaction_record(store.clone(), &transaction)
+            .await
+            .expect("transaction record should persist");
+        backend
+            .set_transition_candidate_probe_override(Some(TransitionCandidateProbe::Unsupported))
+            .await;
+
+        assert!(matches!(
+            finalize_missing_transition_transaction_for_operator(store.clone(), transaction.transaction_id).await,
+            Err(TransitionOperatorError::CandidateNotMissing(TransitionOperatorProbe::Unsupported))
+        ));
+        load_transition_transaction_record(store, transaction.transaction_id)
+            .await
+            .expect("an unsupported probe must retain the transaction journal");
+        assert_eq!(backend.remove_count().await, 0);
+    }
+
    #[cfg(feature = "test-util")]
    #[tokio::test]
    #[serial_test::serial(storage_class_env)]
@@ -2815,7 +3004,7 @@ mod tests {
        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;

        let tier_name = "TXRESPONSELOSS";
-        let backend = register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let backend = register_transition_reconcile_test_tier(&ctx.tier_config_mgr(), tier_name).await;
        let bucket = "transition-response-loss-bucket";
        let object = "source.bin";
        store
@@ -6696,7 +6696,7 @@ mod test {
    use crate::object_api::ObjectInfo;
    use rustfs_filemeta::{
        FileInfo, FileMeta, FileMetaVersion, MetaCacheEntries, MetaCacheEntriesSorted, MetaCacheEntry, MetaDeleteMarker,
-        VersionType,
+        ObjectPartInfo, VersionType,
    };
    use std::collections::{HashMap, HashSet};
    use std::path::PathBuf;
@@ -6895,10 +6895,17 @@ mod test {
        )
        .await
        .expect("local disk should be created");
+        disk.make_volume("bucket").await.expect("test bucket should be created");
        let mut fi = FileInfo::new("object", 1, 1);
        fi.volume = "bucket".to_owned();
        fi.name = "object".to_owned();
        fi.size = 1;
+        fi.parts = vec![ObjectPartInfo {
+            number: 1,
+            size: 1,
+            actual_size: 1,
+            ..Default::default()
+        }];
        fi.fresh = true;
        fi.erasure.index = 1;
        fi.mod_time = Some(time::OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp"));
@@ -6954,6 +6961,12 @@ mod test {
            fi.version_id = Some(Uuid::from_u128(version_idx));
            fi.versioned = true;
            fi.size = 1;
+            fi.parts = vec![ObjectPartInfo {
+                number: 1,
+                size: 1,
+                actual_size: 1,
+                ..Default::default()
+            }];
            fi.mod_time = Some(*mod_time);
            fi.metadata = metadata;

@@ -141,6 +141,7 @@ fn should_enqueue_transition_immediately(oi: &ObjectInfo) -> bool {
 const MAX_UPLOADS_LIST: usize = 10000;

 mod bucket;
+pub(crate) use bucket::await_bucket_namespace_operation;
 mod heal;
 mod heal_walk;
 pub use heal_walk::HealWalkVersion;
@@ -427,11 +428,11 @@ impl ECStore {
 }

 lazy_static! {
-    static ref enableObjcetLockConfig: ObjectLockConfiguration = ObjectLockConfiguration {
+    static ref ENABLED_OBJECT_LOCK_CONFIG: ObjectLockConfiguration = ObjectLockConfiguration {
        object_lock_enabled: Some(ObjectLockEnabled::from_static(ObjectLockEnabled::ENABLED)),
        ..Default::default()
    };
-    static ref enableVersioningConfig: VersioningConfiguration = VersioningConfiguration {
+    static ref ENABLED_VERSIONING_CONFIG: VersioningConfiguration = VersioningConfiguration {
        status: Some(BucketVersioningStatus::from_static(BucketVersioningStatus::ENABLED)),
        ..Default::default()
    };
@@ -16,7 +16,7 @@ use super::*;
 use crate::disk::OldCurrentSize;
 use crate::set_disk::{
    get_lock_acquire_timeout, get_object_lock_diag_slow_acquire_threshold, get_object_lock_diag_slow_hold_threshold,
-    is_object_lock_diag_enabled,
+    is_lock_optimization_enabled, is_object_lock_diag_enabled,
 };
 use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations as _};
 use rustfs_io_metrics::{
@@ -89,7 +89,6 @@ impl PreparedGetObjectReader {
 struct LockGuardedReader {
    inner: Box<dyn AsyncRead + Unpin + Send + Sync>,
    guard: Option<ObjectLockDiagGuard>,
-    remaining: Option<usize>,
 }

 impl AsyncRead for LockGuardedReader {
@@ -97,23 +96,8 @@ impl AsyncRead for LockGuardedReader {
        let had_capacity = buf.remaining() > 0;
        let filled_before = buf.filled().len();
        let poll = Pin::new(&mut self.inner).poll_read(cx, buf);
-        match &poll {
-            Poll::Ready(Ok(())) if buf.filled().len() > filled_before => {
-                let produced = buf.filled().len() - filled_before;
-                if let Some(remaining) = self.remaining.as_mut() {
-                    *remaining = remaining.saturating_sub(produced);
-                    if *remaining == 0 {
-                        self.guard.take();
-                    }
-                }
-            }
-            Poll::Ready(Ok(())) if had_capacity => {
-                self.guard.take();
-            }
-            Poll::Ready(Err(_)) => {
-                self.guard.take();
-            }
-            _ => {}
+        if had_capacity && matches!(poll, Poll::Ready(Ok(()))) && buf.filled().len() == filled_before {
+            self.guard.take();
        }
        poll
    }
@@ -657,16 +641,14 @@ impl ECStore {
    }

    fn attach_read_lock_guard(mut reader: GetObjectReader, guard: Option<ObjectLockDiagGuard>) -> GetObjectReader {
-        if reader.buffered_body.is_some() || reader.object_info.size == 0 {
+        if is_lock_optimization_enabled() || reader.buffered_body.is_some() {
            return reader;
        }

        if let Some(guard) = guard {
-            let remaining = usize::try_from(reader.object_info.size).ok();
            reader.stream = Box::new(LockGuardedReader {
                inner: reader.stream,
                guard: Some(guard),
-                remaining,
            });
        }

@@ -2619,10 +2601,10 @@ mod tests {
    // (backlog#1304): restore entry no longer serializes on the object lock.
    // The replacement semantics — non-blocking reads during the copy-back and
    // fast rejection of a concurrent restore — are covered end-to-end by
-    // `restore_object_usecase_reports_ongoing_conflict_and_completion`
-    // (rustfs/src/app/lifecycle_transition_api_test.rs) and at the lock level
-    // by the accept-guard test below; restore-vs-reader data protection lives
-    // in the inner put_object/complete_multipart_upload commit locks.
+    // `restore_object_usecase_reports_ongoing_conflict`
+    // (rustfs/src/app/lifecycle_transition_api_test.rs), while the SetDisks
+    // transition matrix covers the final local commit. Restore-vs-reader data
+    // protection lives in the inner put_object/complete_multipart_upload locks.
    #[tokio::test]
    #[serial_test::serial]
    async fn restore_accept_guard_serializes_concurrent_accepts() {
@@ -2667,11 +2649,8 @@ mod tests {
                ObjectLockDiagMode::Read,
            );
            let reader = GetObjectReader {
-                stream: Box::new(Cursor::new(vec![1])),
-                object_info: ObjectInfo {
-                    size: 1,
-                    ..Default::default()
-                },
+                stream: Box::new(Cursor::new(Vec::<u8>::new())),
+                object_info: ObjectInfo::default(),
                buffered_body: None,
                body_source: Default::default(),
            };
@@ -2691,7 +2670,7 @@ mod tests {

    #[tokio::test]
    #[serial_test::serial]
-    async fn reader_lock_is_held_for_stream_when_optimization_is_enabled() {
+    async fn reader_lock_is_not_held_for_stream_when_optimization_is_enabled() {
        temp_env::async_with_vars([(rustfs_config::ENV_OBJECT_LOCK_OPTIMIZATION_ENABLE, Some("true"))], async {
            let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
            let lock = rustfs_lock::NamespaceLock::with_local_manager("test".to_string(), manager);
@@ -2711,23 +2690,17 @@ mod tests {
            );
            let reader = GetObjectReader {
                stream: Box::new(Cursor::new(vec![1, 2, 3])),
-                object_info: ObjectInfo {
-                    size: 3,
-                    ..Default::default()
-                },
+                object_info: ObjectInfo::default(),
                buffered_body: None,
                body_source: Default::default(),
            };

            let reader = ECStore::attach_read_lock_guard(reader, Some(read_guard));

-            lock.get_write_lock(key.clone(), "writer", Duration::from_millis(20))
-                .await
-                .expect_err("streaming readers must retain the read lock when lock optimization is enabled");
-            drop(reader);
            lock.get_write_lock(key, "writer", Duration::from_secs(1))
                .await
-                .expect("cancelling the reader should release the read lock");
+                .expect("lock optimization should release the read lock before returning the stream");
+            drop(reader);
        })
        .await;
    }
@@ -2771,50 +2744,41 @@ mod tests {

    #[tokio::test]
    #[serial_test::serial]
-    async fn reader_lock_is_released_after_exact_content_length_without_extra_eof_poll() {
-        for enabled in ["false", "true"] {
-            temp_env::async_with_vars([(rustfs_config::ENV_OBJECT_LOCK_OPTIMIZATION_ENABLE, Some(enabled))], async {
-                let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
-                let lock = rustfs_lock::NamespaceLock::with_local_manager("test".to_string(), manager);
-                let key = rustfs_lock::ObjectKey::new("bucket", "object");
-                let read_guard = lock
-                    .get_read_lock(key.clone(), "reader", Duration::from_secs(1))
-                    .await
-                    .expect("read lock should be acquired");
-                let read_guard = ObjectLockDiagGuard::new(
-                    read_guard,
-                    true,
-                    "test_get_object",
-                    Some("bucket".to_string()),
-                    Some("object".to_string()),
-                    Some("reader".to_string()),
-                    ObjectLockDiagMode::Read,
-                );
-                let reader = GetObjectReader {
-                    stream: Box::new(Cursor::new(vec![1, 2, 3])),
-                    object_info: ObjectInfo {
-                        size: 3,
-                        ..Default::default()
-                    },
-                    buffered_body: None,
-                    body_source: Default::default(),
-                };
+    async fn reader_lock_is_released_after_stream_eof() {
+        temp_env::async_with_vars([(rustfs_config::ENV_OBJECT_LOCK_OPTIMIZATION_ENABLE, Some("false"))], async {
+            let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
+            let lock = rustfs_lock::NamespaceLock::with_local_manager("test".to_string(), manager);
+            let key = rustfs_lock::ObjectKey::new("bucket", "object");
+            let read_guard = lock
+                .get_read_lock(key.clone(), "reader", Duration::from_secs(1))
+                .await
+                .expect("read lock should be acquired");
+            let read_guard = ObjectLockDiagGuard::new(
+                read_guard,
+                true,
+                "test_get_object",
+                Some("bucket".to_string()),
+                Some("object".to_string()),
+                Some("reader".to_string()),
+                ObjectLockDiagMode::Read,
+            );
+            let reader = GetObjectReader {
+                stream: Box::new(Cursor::new(vec![1, 2, 3])),
+                object_info: ObjectInfo::default(),
+                buffered_body: None,
+                body_source: Default::default(),
+            };

-                let mut reader = ECStore::attach_read_lock_guard(reader, Some(read_guard));
-                let mut output = [0_u8; 3];
-                reader
-                    .stream
-                    .read_exact(&mut output)
-                    .await
-                    .expect("reader should deliver the exact content length");
-                assert_eq!(output, [1, 2, 3]);
+            let mut reader = ECStore::attach_read_lock_guard(reader, Some(read_guard));
+            let mut output = Vec::new();
+            reader.stream.read_to_end(&mut output).await.expect("reader should reach EOF");
+            assert_eq!(output, vec![1, 2, 3]);

-                lock.get_write_lock(key, "writer", Duration::from_secs(1))
-                    .await
-                    .expect("the final response byte should release the read lock without an extra EOF poll");
-                drop(reader);
-            })
-            .await;
-        }
+            lock.get_write_lock(key, "writer", Duration::from_secs(1))
+                .await
+                .expect("EOF should release the read lock before the reader is dropped");
+            drop(reader);
+        })
+        .await;
    }
 }
@@ -0,0 +1,43 @@
+//! Test-only guard against tracing's process-global callsite-interest cache.
+//!
+//! Any test that installs its own subscriber and then asserts on span or event
+//! context is asserting on process-global state. See
+//! [`pin_callsite_interest_for_test`] for what goes wrong and why holding its
+//! guard fixes it.
+
+/// Stop a sibling test thread from poisoning tracing's process-global callsite
+/// interest cache while this test asserts on span or event context.
+///
+/// `tracing` caches every callsite's `Interest` in **process-global** state, and
+/// the first thread to reach a callsite fixes that value. While at most one
+/// dispatcher is registered, tracing-core takes a fast path
+/// (`Dispatchers::rebuilder` -> `Rebuilder::JustOne`) that derives a
+/// newly-registered callsite's interest from whichever subscriber is current
+/// *on the registering thread*, and registration happens exactly once (guarded
+/// by a CAS in `DefaultCallsite::register`).
+///
+/// In a multi-threaded `cargo test` binary a sibling test therefore routinely
+/// reaches a production callsite first, from a thread with no subscriber
+/// installed: the interest is derived from that thread's `NoSubscriber` and
+/// cached as `Interest::never()` for the whole process. From then on the
+/// callsite is dead for *every* later caller, including a test that installed
+/// its own subscriber — an `info_span!` silently evaluates to `Span::none()`, and
+/// a `warn!`/`info!` event never fires at all.
+///
+/// Registering a second, inert dispatcher closes both halves of that race:
+///
+/// * constructing it rebuilds every *already-registered* callsite's interest
+///   against the live dispatcher set — which includes the caller's subscriber —
+///   repairing whatever a sibling may already have poisoned; and
+/// * holding it alive keeps tracing-core off the single-dispatcher fast path, so
+///   a callsite registered *later* by any thread is resolved against that live
+///   set instead of the registering thread's `NoSubscriber`.
+///
+/// Call this **after** installing the test's subscriber, and hold the returned
+/// guard for the rest of the test. Only the `cargo test` fallback needs it:
+/// nextest runs each test in its own process, where there is no sibling thread
+/// to lose the race to (see `docs/testing/README.md`).
+#[must_use = "callsite interest is only pinned while the returned guard is held"]
+pub(crate) fn pin_callsite_interest_for_test() -> tracing::Dispatch {
+    tracing::Dispatch::new(tracing_core::subscriber::NoSubscriber::default())
+}
@@ -20,5 +20,5 @@ Example:
 ```powershell
 $env:RUSTFS_MINIO_FIXTURE_ROOT = '.\rustfs\tmp\minio-fixture-lab-local-key'
 $env:RUSTFS_MINIO_STATIC_KMS_KEY_B64 = '<base64-32-byte-local-minio-kms-key>'
-cargo +1.96.0 test -p rustfs-ecstore --features rio-v2 --test minio_generated_read_test -- --ignored
+cargo +1.97.1 test -p rustfs-ecstore --features rio-v2 --test minio_generated_read_test -- --ignored
 ```
@@ -219,6 +219,16 @@ impl ErasureInfo {
 }

 // #[derive(Debug, Clone)]
+#[derive(Serialize, Deserialize, Debug, PartialEq, Eq, Clone, Copy, Default)]
+#[serde(rename_all = "kebab-case")]
+pub enum TransitionVersionState {
+    #[default]
+    Unknown,
+    KnownDisabled,
+    SuspendedNull,
+    Exact,
+}
+
 #[derive(Serialize, Deserialize, Debug, PartialEq, Clone, Default)]
 pub struct FileInfo {
    pub volume: String,
@@ -230,6 +240,10 @@ pub struct FileInfo {
    pub transitioned_objname: String,
    pub transition_tier: String,
    pub transition_version_id: Option<Uuid>,
+    #[serde(default)]
+    pub transition_version: Option<String>,
+    #[serde(default)]
+    pub transition_version_state: TransitionVersionState,
    pub expire_restored: bool,
    pub data_dir: Option<Uuid>,
    pub mod_time: Option<OffsetDateTime>,
@@ -415,6 +429,10 @@ impl FileInfo {
    }

    fn validate_collection_contents(&self, layout: Option<&ValidatedErasureLayout>) -> Result<()> {
+        if layout.is_some() && self.size > 0 && self.parts.is_empty() {
+            return Err(Error::FileCorrupt);
+        }
+
        let mut part_numbers = [0u64; FILEINFO_PART_BITMAP_WORDS];
        for part in &self.parts {
            if let Some(layout) = layout {
@@ -455,6 +473,10 @@ impl FileInfo {
        if self.mod_time.is_none_or(|mod_time| mod_time <= OffsetDateTime::UNIX_EPOCH)
            || (!allow_nil_version_id && self.version_id.is_some_and(|version_id| version_id.is_nil()))
            || self.transition_version_id.is_some_and(|version_id| version_id.is_nil())
+            || self
+                .transition_version
+                .as_ref()
+                .is_some_and(|version_id| version_id.is_empty())
            || self.size != 0
            || self.data_dir.is_some()
            || self.mode.is_some()
@@ -488,6 +510,7 @@ impl FileInfo {
            || !self.transitioned_objname.is_empty()
            || !self.transition_tier.is_empty()
            || self.transition_version_id.is_some()
+            || self.transition_version.is_some()
            || self.expire_restored
            || self.size != 0
            || self.data_dir.is_some()
@@ -532,6 +555,25 @@ impl FileInfo {
    /// return `None`.
    pub fn validate(&self, mode: ValidationMode) -> Result<Option<ValidatedErasureLayout>> {
        self.validate_collection_bounds()?;
+        if let (Some(version), Some(version_id)) = (&self.transition_version, self.transition_version_id)
+            && Uuid::parse_str(version).ok() != Some(version_id)
+        {
+            return Err(Error::FileCorrupt);
+        }
+        let transition_state_valid = match self.transition_version_state {
+            TransitionVersionState::Unknown => true,
+            TransitionVersionState::KnownDisabled => self.transition_version.is_none() && self.transition_version_id.is_none(),
+            TransitionVersionState::SuspendedNull => {
+                self.transition_version.as_deref() == Some("null") && self.transition_version_id.is_none()
+            }
+            TransitionVersionState::Exact => self
+                .transition_version
+                .as_deref()
+                .is_some_and(|version| version != "null" && !version.is_empty()),
+        };
+        if !transition_state_valid {
+            return Err(Error::FileCorrupt);
+        }

        let erasure_layout = match mode {
            ValidationMode::RequireErasure => Some(self.validate_erasure_geometry()?),
@@ -692,6 +734,10 @@ impl FileInfo {

    // to_part_offset gets the part index where offset is located, returns part index and offset
    pub fn to_part_offset(&self, offset: usize) -> Result<(usize, usize)> {
+        if self.size > 0 && self.parts.is_empty() {
+            return Err(Error::FileCorrupt);
+        }
+
        if offset == 0 {
            return Ok((0, 0));
        }
@@ -824,6 +870,8 @@ impl FileInfo {
            && self.transition_tier == other.transition_tier
            && self.transitioned_objname == other.transitioned_objname
            && self.transition_version_id == other.transition_version_id
+            && self.transition_version == other.transition_version
+            && self.transition_version_state == other.transition_version_state
    }

    /// Check if metadata maps are equal
@@ -1185,6 +1233,21 @@ mod tests {
        assert_eq!(layout.expect("strict layout must be present").data_blocks, 16);
    }

+    #[test]
+    fn validate_require_erasure_rejects_positive_size_without_parts() {
+        let mut fi = FileInfo::new("bucket/object", 4, 2);
+        fi.erasure.index = 1;
+        fi.size = 1;
+
+        assert_eq!(fi.validate_for_metadata_read(), Err(Error::FileCorrupt));
+        assert_eq!(fi.to_part_offset(0), Err(Error::FileCorrupt));
+
+        fi.size = 0;
+        fi.validate_for_metadata_read()
+            .expect("zero-byte object metadata may omit parts");
+        assert_eq!(fi.to_part_offset(0), Ok((0, 0)));
+    }
+
    #[test]
    fn validate_for_erasure_write_only_relaxes_pending_shard_index() {
        let mut fi = validation_test_fileinfo();
@@ -1328,6 +1391,15 @@ mod tests {
        assert_file_corrupt(&fi, ValidationMode::DeleteOnly);
    }

+    #[test]
+    fn metadata_read_validation_rejects_conflicting_transition_versions() {
+        let mut fi = one_shard_validation_fileinfo(1);
+        fi.transition_version_id = Some(Uuid::new_v4());
+        fi.transition_version = Some(Uuid::new_v4().to_string());
+
+        assert_file_corrupt(&fi, ValidationMode::RequireErasure);
+    }
+
    #[test]
    fn metadata_read_validation_requires_canonical_delete_marker_shape() {
        let marker = FileInfo {
@@ -1699,6 +1771,12 @@ mod tests {
                    transitioned_objname,
                    transition_tier,
                    transition_version_id,
+                    transition_version: transition_version_id.map(|version_id| version_id.to_string()),
+                    transition_version_state: if transition_version_id.is_some() {
+                        TransitionVersionState::Exact
+                    } else {
+                        TransitionVersionState::Unknown
+                    },
                    expire_restored,
                    data_dir,
                    mod_time,
@@ -182,14 +182,9 @@ impl FileMeta {
        // TODO: use old buf
        let meta_buf = ver.marshal_msg()?;

-        let pre_mod_time = self.versions[idx].header.mod_time;
-
        self.versions[idx].header = ver.header();
        self.versions[idx].meta = meta_buf;
-
-        if pre_mod_time != self.versions[idx].header.mod_time {
-            self.sort_by_mod_time();
-        }
+        self.sort_by_mod_time();

        Ok(())
    }
@@ -356,16 +351,19 @@ impl FileMeta {
            return self.set_idx(fidx, version);
        }

-        let mod_time = version.get_mod_time();
        let new_shallow = FileMetaShallowVersion::try_from(version)?;
-        let insert_pos = match mod_time {
-            Some(nm) => self.versions.partition_point(|exist| match exist.header.mod_time {
-                Some(em) => em > nm,
-                None => false,
-            }),
-            None => self.versions.partition_point(|exist| exist.header.mod_time.is_some()),
-        };
+        let insert_pos = self
+            .versions
+            .partition_point(|existing| existing.header.sorts_before(&new_shallow.header));
        self.versions.insert(insert_pos, new_shallow);
+        // `partition_point` only returns the canonical slot when `versions` is
+        // already canonically ordered, and nothing establishes that: `FileMeta::load`
+        // replays the on-disk order verbatim, and metadata written before the
+        // canonical-order fix ordered equal-`mod_time` ties by insertion rather than
+        // by `sorts_before`. Re-sort so an insert leaves the list canonical either
+        // way, matching what `set_idx` already does on the replace path. Cheap: after
+        // a correctly placed insert the slice is a single sorted run.
+        self.sort_by_mod_time();
        Ok(())

        // if !ver.valid() {
@@ -1112,6 +1110,170 @@ mod test {
        assert!(fm.is_sorted_by_mod_time());
    }

+    fn version_for_ordering(
+        version_type: VersionType,
+        version_id: Uuid,
+        mod_time: OffsetDateTime,
+        marker: u8,
+    ) -> FileMetaVersion {
+        match version_type {
+            VersionType::Object => FileMetaVersion {
+                version_type,
+                object: Some(MetaObject {
+                    version_id: Some(version_id),
+                    erasure_algorithm: ErasureAlgo::ReedSolomon,
+                    erasure_m: 2,
+                    erasure_n: 2,
+                    erasure_block_size: 1 << 20,
+                    bitrot_checksum_algo: ChecksumAlgo::HighwayHash,
+                    mod_time: Some(mod_time),
+                    meta_sys: HashMap::from([("ordering-marker".to_string(), vec![marker])]),
+                    ..Default::default()
+                }),
+                ..Default::default()
+            },
+            VersionType::Delete => FileMetaVersion {
+                version_type,
+                delete_marker: Some(MetaDeleteMarker {
+                    version_id: Some(version_id),
+                    mod_time: Some(mod_time),
+                    meta_sys: HashMap::from([("ordering-marker".to_string(), vec![marker])]),
+                }),
+                ..Default::default()
+            },
+            _ => unreachable!("ordering regression only constructs object and delete versions"),
+        }
+    }
+
+    #[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");
+        let object = version_for_ordering(VersionType::Object, Uuid::from_u128(1), mod_time, 1);
+        let delete_marker = version_for_ordering(VersionType::Delete, Uuid::from_u128(2), mod_time, 2);
+
+        for versions in [[object.clone(), delete_marker.clone()], [delete_marker, object]] {
+            let mut fm = FileMeta::new();
+            for version in versions {
+                fm.add_version_filemata(version).expect("add equal-time version");
+            }
+
+            assert_eq!(fm.versions[0].header.version_type, VersionType::Object);
+            assert_eq!(fm.versions[1].header.version_type, VersionType::Delete);
+            assert!(fm.versions[0].header.sorts_before(&fm.versions[1].header));
+        }
+    }
+
+    #[test]
+    fn add_version_filemata_preserves_tie_breaks_after_reload() {
+        let mod_time = OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid test timestamp");
+        let versions = [
+            version_for_ordering(VersionType::Object, Uuid::from_u128(10), mod_time, 10),
+            version_for_ordering(VersionType::Object, Uuid::from_u128(20), mod_time, 20),
+            version_for_ordering(VersionType::Delete, Uuid::from_u128(30), mod_time, 30),
+        ];
+        let mut expected = versions.iter().map(FileMetaVersion::header).collect::<Vec<_>>();
+        expected.sort_by(|a, b| {
+            if a.sorts_before(b) {
+                Ordering::Less
+            } else if b.sorts_before(a) {
+                Ordering::Greater
+            } else {
+                Ordering::Equal
+            }
+        });
+
+        let mut fm = FileMeta::new();
+        for version in versions.into_iter().rev() {
+            fm.add_version_filemata(version).expect("add equal-time version");
+        }
+        let loaded = FileMeta::load(&fm.marshal_msg().expect("serialize file metadata")).expect("reload file metadata");
+        let actual = loaded
+            .versions
+            .iter()
+            .map(|version| version.header.clone())
+            .collect::<Vec<_>>();
+
+        assert_ne!(expected[0].signature, expected[1].signature, "fixture must exercise signature ordering");
+        assert_eq!(actual, expected);
+        assert!(actual.windows(2).all(|pair| pair[0].sorts_before(&pair[1])));
+    }
+
+    #[test]
+    fn add_version_filemata_reorders_equal_time_replacement() {
+        let mod_time = OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid test timestamp");
+        let first = version_for_ordering(VersionType::Object, Uuid::from_u128(40), mod_time, 40);
+        let second = version_for_ordering(VersionType::Object, Uuid::from_u128(50), mod_time, 50);
+        let (target, peer) = if first.header().sorts_before(&second.header()) {
+            (first, second)
+        } else {
+            (second, first)
+        };
+        let target_id = target.header().version_id.expect("test version id");
+
+        let mut fm = FileMeta::new();
+        fm.add_version_filemata(peer).expect("add peer object");
+        fm.add_version_filemata(target).expect("add target object");
+        assert_eq!(fm.versions[0].header.version_id, Some(target_id));
+
+        fm.add_version_filemata(version_for_ordering(VersionType::Delete, target_id, mod_time, 60))
+            .expect("replace target with equal-time delete marker");
+
+        assert_eq!(fm.versions[0].header.version_type, VersionType::Object);
+        assert_eq!(fm.versions[1].header.version_type, VersionType::Delete);
+        assert!(fm.versions[0].header.sorts_before(&fm.versions[1].header));
+    }
+
+    /// `add_version_filemata` positions an inserted version with
+    /// `partition_point(sorts_before)`, which only yields the canonical slot when
+    /// `versions` is already canonically ordered. Nothing establishes that
+    /// precondition on load: `FileMeta::unmarshal_msg` pushes versions in file
+    /// order, and metadata written before the canonical-order fix ordered
+    /// equal-`mod_time` ties by insertion instead of by `sorts_before`. An insert
+    /// into such a list must still leave it canonical, the same way `set_idx`
+    /// already re-sorts on the replace path.
+    #[test]
+    fn add_version_filemata_canonicalizes_versions_loaded_out_of_order() {
+        let mod_time = OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid test timestamp");
+        let first = version_for_ordering(VersionType::Object, Uuid::from_u128(70), mod_time, 70);
+        let second = version_for_ordering(VersionType::Object, Uuid::from_u128(80), mod_time, 80);
+        let (early, late) = if first.header().sorts_before(&second.header()) {
+            (first, second)
+        } else {
+            (second, first)
+        };
+
+        // Persist the pair the wrong way round to emulate a pre-canonical-order
+        // xl.meta, bypassing add_version_filemata so the bad order really lands on
+        // the wire.
+        let mut legacy = FileMeta::new();
+        legacy
+            .versions
+            .push(FileMetaShallowVersion::try_from(late).expect("shallow later version"));
+        legacy
+            .versions
+            .push(FileMetaShallowVersion::try_from(early).expect("shallow earlier version"));
+
+        let mut loaded =
+            FileMeta::load(&legacy.marshal_msg().expect("serialize legacy metadata")).expect("reload legacy metadata");
+        assert!(
+            !loaded.versions[0].header.sorts_before(&loaded.versions[1].header),
+            "fixture must reach add_version_filemata non-canonically ordered"
+        );
+
+        loaded
+            .add_version_filemata(version_for_ordering(VersionType::Delete, Uuid::from_u128(90), mod_time, 90))
+            .expect("insert equal-time delete marker");
+
+        assert_eq!(loaded.versions.len(), 3);
+        assert!(
+            loaded
+                .versions
+                .windows(2)
+                .all(|pair| pair[0].header.sorts_before(&pair[1].header)),
+            "insert must leave the version list canonically ordered"
+        );
+    }
+
    /// Regression for backlog#799 B16: replication reset state must be
    /// persisted under both internal prefixes, never as a bare ARN. A bare-ARN
    /// key (produced by `ObjectInfo::replication_state`) has no internal prefix,
@@ -26,13 +26,14 @@ use super::msgp_decode::{
    PrependByteReader, prealloc_hint, read_exact_vec, read_nil_or_array_len, read_nil_or_map_len, skip_msgp_value,
 };
 use super::*;
-use crate::ChecksumInfo;
+use crate::{ChecksumInfo, TransitionVersionState};
 use rustfs_utils::HashAlgorithm;
 use rustfs_utils::http::{
    RUSTFS_INTERNAL_PREFIX, SUFFIX_CRC, SUFFIX_FREE_VERSION, SUFFIX_INLINE_DATA, SUFFIX_PURGESTATUS, SUFFIX_TIER_FV_ID,
    SUFFIX_TIER_FV_MARKER, SUFFIX_TRANSITION_STATUS, SUFFIX_TRANSITION_TIER, SUFFIX_TRANSITION_TIER_DESTINATION_ID,
-    SUFFIX_TRANSITIONED_OBJECTNAME, SUFFIX_TRANSITIONED_VERSION_ID, contains_key_bytes, get_bytes, get_consistent_bytes, get_str,
-    has_internal_suffix, insert_bytes, is_internal_key, remove_bytes, strip_internal_prefix,
+    SUFFIX_TRANSITIONED_OBJECTNAME, SUFFIX_TRANSITIONED_VERSION_ID, SUFFIX_TRANSITIONED_VERSION_STATE, contains_key_bytes,
+    get_bytes, get_consistent_bytes, get_str, has_internal_suffix, insert_bytes, is_internal_key, remove_bytes,
+    strip_internal_prefix,
 };

 const MSGPACK_EXT8: u8 = 0xc7;
@@ -43,6 +44,7 @@ const MSGPACK_FIXEXT8: u8 = 0xd7;
 const MSGPACK_TIME_EXT_LEGACY: i8 = 5;
 const MSGPACK_TIME_EXT_OFFICIAL: i8 = -1;
 const MSGPACK_TIME_LEN: u8 = 12;
+const MAX_TRANSITION_VERSION_LEN: usize = 1024;

 /// Sentinel signature returned when a version has no computable body (invalid /
 /// missing inner object). Mirrors MinIO's `signatureErr` so such versions never
@@ -251,23 +253,93 @@ fn parse_legacy_uuid_bytes(bytes: &[u8], field: &str) -> Result<Option<Uuid>> {

 /// Decode a stored transitioned-version-id from a version's `meta_sys`.
 ///
-/// RustFS writes it as 16 raw UUID bytes; MinIO-migrated tiered objects store
-/// the remote tier's version id as a UUID *string*. Accept both, and treat any
-/// absent / nil / otherwise-unparseable value as "no tier version" (matching the
-/// tolerant pre-hardening behavior) rather than failing the whole object read —
-/// a malformed tier id must not make an otherwise-readable object unreadable.
-fn transitioned_version_id_from_meta_sys(meta_sys: &HashMap<String, Vec<u8>>) -> Option<Uuid> {
-    let value = get_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_ID)?;
+/// Legacy RustFS writes used 16 raw UUID bytes. New writes and MinIO-migrated
+/// records use the provider's exact UTF-8 version text. Empty, nil UUID, and
+/// malformed bytes are not usable remote versions.
+fn transitioned_version_from_meta_sys(meta_sys: &HashMap<String, Vec<u8>>) -> Result<Option<String>> {
+    if !contains_key_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_ID) {
+        return Ok(None);
+    }
+    let Some(value) = get_consistent_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_ID) else {
+        return Ok(None);
+    };
+    let value = value.to_vec();
    if value.is_empty() {
-        return None;
+        return Ok(None);
    }
    if let Ok(id) = Uuid::from_slice(&value) {
-        return (!id.is_nil()).then_some(id);
+        return Ok((!id.is_nil()).then(|| id.to_string()));
    }
-    std::str::from_utf8(&value)
+    let Ok(value) = String::from_utf8(value) else {
+        return Ok(None);
+    };
+    if value.is_empty()
+        || value.len() > MAX_TRANSITION_VERSION_LEN
+        || value.chars().any(char::is_control)
+        || Uuid::parse_str(&value).is_ok_and(|id| id.is_nil())
+    {
+        Ok(None)
+    } else {
+        Ok(Some(value))
+    }
+}
+
+fn transition_version_state_from_meta_sys(
+    meta_sys: &HashMap<String, Vec<u8>>,
+    version: Option<&str>,
+) -> Result<TransitionVersionState> {
+    if !contains_key_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE) {
+        return Ok(TransitionVersionState::Unknown);
+    }
+    let value = get_consistent_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE).ok_or(Error::FileCorrupt)?;
+    let state = match value {
+        b"known-disabled" => TransitionVersionState::KnownDisabled,
+        b"suspended-null" => TransitionVersionState::SuspendedNull,
+        b"exact" => TransitionVersionState::Exact,
+        b"unknown" => TransitionVersionState::Unknown,
+        _ => return Err(Error::FileCorrupt),
+    };
+    let valid = match state {
+        TransitionVersionState::Unknown | TransitionVersionState::KnownDisabled => version.is_none(),
+        TransitionVersionState::SuspendedNull => version == Some("null"),
+        TransitionVersionState::Exact => version.is_some_and(|value| value != "null"),
+    };
+    valid.then_some(state).ok_or(Error::FileCorrupt)
+}
+
+fn transition_version_state_bytes(state: TransitionVersionState) -> &'static [u8] {
+    match state {
+        TransitionVersionState::Unknown => b"unknown",
+        TransitionVersionState::KnownDisabled => b"known-disabled",
+        TransitionVersionState::SuspendedNull => b"suspended-null",
+        TransitionVersionState::Exact => b"exact",
+    }
+}
+
+fn set_transition_version_state(meta_sys: &mut HashMap<String, Vec<u8>>, state: TransitionVersionState) {
+    if state == TransitionVersionState::Unknown {
+        remove_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE);
+    } else {
+        insert_bytes(
+            meta_sys,
+            SUFFIX_TRANSITIONED_VERSION_STATE,
+            transition_version_state_bytes(state).to_vec(),
+        );
+    }
+}
+
+fn legacy_transitioned_version_id_from_meta_sys(meta_sys: &HashMap<String, Vec<u8>>) -> Option<Uuid> {
+    transitioned_version_from_meta_sys(meta_sys)
        .ok()
-        .and_then(|s| Uuid::parse_str(s.trim()).ok())
-        .filter(|id| !id.is_nil())
+        .flatten()
+        .and_then(|value| Uuid::parse_str(&value).ok())
+}
+
+fn transitioned_version_bytes(fi: &FileInfo) -> Option<Vec<u8>> {
+    fi.transition_version
+        .as_ref()
+        .map(|version| version.as_bytes().to_vec())
+        .or_else(|| fi.transition_version_id.map(|version_id| version_id.as_bytes().to_vec()))
 }

 fn parse_legacy_erasure_algo(value: &str) -> ErasureAlgo {
@@ -2398,7 +2470,9 @@ impl MetaObject {
        let transitioned_objname = get_bytes(&self.meta_sys, SUFFIX_TRANSITIONED_OBJECTNAME)
            .map(|v| String::from_utf8_lossy(&v).to_string())
            .unwrap_or_default();
-        let transition_version_id = transitioned_version_id_from_meta_sys(&self.meta_sys);
+        let transition_version = transitioned_version_from_meta_sys(&self.meta_sys)?;
+        let transition_version_state = transition_version_state_from_meta_sys(&self.meta_sys, transition_version.as_deref())?;
+        let transition_version_id = transition_version.as_deref().and_then(|value| Uuid::parse_str(value).ok());
        let transition_tier = get_bytes(&self.meta_sys, SUFFIX_TRANSITION_TIER)
            .map(|v| String::from_utf8_lossy(&v).to_string())
            .unwrap_or_default();
@@ -2419,6 +2493,8 @@ impl MetaObject {
            transition_status,
            transitioned_objname,
            transition_version_id,
+            transition_version,
+            transition_version_state,
            transition_tier,
            ..Default::default()
        })
@@ -2431,13 +2507,12 @@ impl MetaObject {
            SUFFIX_TRANSITIONED_OBJECTNAME,
            fi.transitioned_objname.as_bytes().to_vec(),
        );
-        if let Some(transition_version_id) = fi.transition_version_id.as_ref() {
-            insert_bytes(
-                &mut self.meta_sys,
-                SUFFIX_TRANSITIONED_VERSION_ID,
-                transition_version_id.as_bytes().to_vec(),
-            );
+        if let Some(transition_version) = transitioned_version_bytes(fi) {
+            insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
+        } else {
+            remove_bytes(&mut self.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID);
        }
+        set_transition_version_state(&mut self.meta_sys, fi.transition_version_state);
        insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITION_TIER, fi.transition_tier.as_bytes().to_vec());
        if let Some(destination_id) = get_str(&fi.metadata, SUFFIX_TRANSITION_TIER_DESTINATION_ID) {
            insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITION_TIER_DESTINATION_ID, destination_id.into_bytes());
@@ -2501,6 +2576,7 @@ impl MetaObject {
                SUFFIX_TRANSITION_TIER,
                SUFFIX_TRANSITIONED_OBJECTNAME,
                SUFFIX_TRANSITIONED_VERSION_ID,
+                SUFFIX_TRANSITIONED_VERSION_STATE,
            ] {
                if let Some(v) = get_bytes(&self.meta_sys, suffix) {
                    insert_bytes(&mut delete_marker.meta_sys, suffix, v);
@@ -2562,8 +2638,11 @@ impl From<FileInfo> for MetaObject {
            );
        }

-        if let Some(vid) = &value.transition_version_id {
-            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, vid.as_bytes().to_vec());
+        if let Some(transition_version) = transitioned_version_bytes(&value) {
+            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
+        }
+        if !value.transition_status.is_empty() {
+            set_transition_version_state(&mut meta_sys, value.transition_version_state);
        }

        if !value.transition_tier.is_empty() {
@@ -2706,7 +2785,11 @@ impl MetaDeleteMarker {
                .map(|v| String::from_utf8_lossy(&v).to_string())
                .unwrap_or_default();

-            fi.transition_version_id = transitioned_version_id_from_meta_sys(&self.meta_sys);
+            fi.transition_version = transitioned_version_from_meta_sys(&self.meta_sys).ok().flatten();
+            fi.transition_version_id = legacy_transitioned_version_id_from_meta_sys(&self.meta_sys);
+            fi.transition_version_state =
+                transition_version_state_from_meta_sys(&self.meta_sys, fi.transition_version.as_deref())
+                    .unwrap_or(TransitionVersionState::Unknown);
        }

        fi
@@ -2859,8 +2942,11 @@ impl From<FileInfo> for MetaDeleteMarker {
                value.transitioned_objname.as_bytes().to_vec(),
            );
        }
-        if let Some(version_id) = value.transition_version_id {
-            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, version_id.as_bytes().to_vec());
+        if let Some(transition_version) = transitioned_version_bytes(&value) {
+            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
+        }
+        if !value.transition_status.is_empty() || value.tier_free_version() {
+            set_transition_version_state(&mut meta_sys, value.transition_version_state);
        }
        if !value.transition_tier.is_empty() {
            insert_bytes(&mut meta_sys, SUFFIX_TRANSITION_TIER, value.transition_tier.as_bytes().to_vec());
@@ -3412,7 +3498,7 @@ mod tests {
            .insert("x-rustfs-internal-healing".to_string(), "true".to_string());
        marker.metadata.insert("content-type".to_string(), "text/plain".to_string());
        let remote_version_id = Uuid::new_v4();
-        marker.transition_version_id = Some(remote_version_id);
+        marker.transition_version = Some(remote_version_id.to_string());

        let converted = MetaDeleteMarker::from(marker);

@@ -3420,7 +3506,19 @@ mod tests {
        assert_eq!(converted.meta_sys.get("x-minio-internal-purgestatus"), Some(&b"pending".to_vec()));
        assert_eq!(
            get_bytes(&converted.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID),
-            Some(remote_version_id.as_bytes().to_vec())
+            Some(remote_version_id.to_string().into_bytes())
+        );
+        assert_eq!(
+            converted
+                .meta_sys
+                .get(&format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_ID}")),
+            Some(&remote_version_id.to_string().into_bytes())
+        );
+        assert_eq!(
+            converted
+                .meta_sys
+                .get(&format!("{}{SUFFIX_TRANSITIONED_VERSION_ID}", rustfs_utils::http::MINIO_INTERNAL_PREFIX)),
+            Some(&remote_version_id.to_string().into_bytes())
        );
        assert!(!converted.meta_sys.contains_key("x-rustfs-internal-healing"));
        assert!(!converted.meta_sys.contains_key("content-type"));
@@ -4097,19 +4195,129 @@ mod tests {
            .into_fileinfo("b", "k", false)
            .expect("into_fileinfo");
        assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
+        assert_eq!(fi.transition_version_state, TransitionVersionState::Unknown);
    }

    #[test]
-    fn meta_object_transition_version_id_unparseable_stays_readable_as_none() {
-        // A non-UUID / non-16-byte tier version id must NOT make the object
-        // unreadable; it is tolerated as "no tier version" (compat with
-        // pre-hardening behavior and foreign/edge metadata).
+    fn meta_object_transition_version_id_opaque_text_is_preserved() {
        let mut sys = HashMap::new();
-        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"not-a-uuid".to_vec());
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"opaque-generation-42".to_vec());
        let fi = make_meta_object_with_sys(sys)
            .into_fileinfo("b", "k", false)
-            .expect("unparseable transition version id must not fail the object read");
+            .expect("opaque transition version id must decode");
        assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version.as_deref(), Some("opaque-generation-42"));
+        assert_eq!(fi.transition_version_state, TransitionVersionState::Unknown);
+    }
+
+    #[test]
+    fn meta_object_transition_version_state_exact_round_trips_dual_keys() {
+        let id = sample_version_id();
+        let expected_version = id.to_string();
+        let fi = FileInfo {
+            transition_status: "complete".to_string(),
+            transition_version: Some(expected_version.clone()),
+            transition_version_state: TransitionVersionState::Exact,
+            ..Default::default()
+        };
+
+        let object = MetaObject::from(fi);
+        assert_eq!(
+            object
+                .meta_sys
+                .get(&format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_STATE}"))
+                .map(Vec::as_slice),
+            Some(b"exact".as_slice())
+        );
+        assert_eq!(
+            object
+                .meta_sys
+                .get(&format!(
+                    "{}{SUFFIX_TRANSITIONED_VERSION_STATE}",
+                    rustfs_utils::http::MINIO_INTERNAL_PREFIX
+                ))
+                .map(Vec::as_slice),
+            Some(b"exact".as_slice())
+        );
+        assert_eq!(
+            legacy_transitioned_version_id_from_meta_sys(&object.meta_sys),
+            Some(id),
+            "UUID exact writes must remain readable by the legacy UUID consumer"
+        );
+        let decoded = object.into_fileinfo("b", "k", false).expect("exact state should round trip");
+        assert_eq!(decoded.transition_version_state, TransitionVersionState::Exact);
+        assert_eq!(decoded.transition_version.as_deref(), Some(expected_version.as_str()));
+    }
+
+    #[test]
+    fn set_transition_known_disabled_removes_stale_version_dual_keys() {
+        let mut meta_sys = HashMap::new();
+        insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, b"stale-legacy-version".to_vec());
+        let mut object = make_meta_object_with_sys(meta_sys);
+        object.set_transition(&FileInfo {
+            transition_status: TRANSITION_COMPLETE.to_string(),
+            transitioned_objname: "remote/object".to_string(),
+            transition_version_state: TransitionVersionState::KnownDisabled,
+            transition_tier: "WARM".to_string(),
+            ..Default::default()
+        });
+
+        assert_eq!(get_bytes(&object.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID), None);
+        assert!(
+            !object
+                .meta_sys
+                .contains_key(&format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_ID}"))
+        );
+        assert!(
+            !object
+                .meta_sys
+                .contains_key(&format!("{}{SUFFIX_TRANSITIONED_VERSION_ID}", rustfs_utils::http::MINIO_INTERNAL_PREFIX))
+        );
+        let decoded = object
+            .into_fileinfo("b", "k", false)
+            .expect("known-disabled transition must remain readable after replacing stale metadata");
+        assert_eq!(decoded.transition_version, None);
+        assert_eq!(decoded.transition_version_state, TransitionVersionState::KnownDisabled);
+    }
+
+    #[test]
+    fn meta_object_transition_version_state_conflict_fails_closed() {
+        let mut sys = HashMap::new();
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, sample_version_id().as_bytes().to_vec());
+        sys.insert(format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_STATE}"), b"exact".to_vec());
+        sys.insert(
+            format!("{}{SUFFIX_TRANSITIONED_VERSION_STATE}", rustfs_utils::http::MINIO_INTERNAL_PREFIX),
+            b"known-disabled".to_vec(),
+        );
+
+        make_meta_object_with_sys(sys)
+            .into_fileinfo("b", "k", false)
+            .expect_err("conflicting state keys must fail closed");
+    }
+
+    #[test]
+    fn meta_object_transition_version_id_invalid_utf8_yields_none() {
+        let mut sys = HashMap::new();
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, vec![0xff]);
+        let fi = make_meta_object_with_sys(sys)
+            .into_fileinfo("b", "k", false)
+            .expect("invalid transition version bytes must not fail the object read");
+        assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version, None);
+    }
+
+    #[test]
+    fn meta_object_transition_version_id_unsafe_text_yields_none() {
+        for value in [b"opaque\0version".to_vec(), vec![b'x'; MAX_TRANSITION_VERSION_LEN + 1]] {
+            let mut sys = HashMap::new();
+            insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, value);
+            let fi = make_meta_object_with_sys(sys)
+                .into_fileinfo("b", "k", false)
+                .expect("unsafe transition version text must not fail the object read");
+            assert_eq!(fi.transition_version_id, None);
+            assert_eq!(fi.transition_version, None);
+        }
    }

    #[test]
@@ -4123,6 +4331,7 @@ mod tests {
            .into_fileinfo("b", "k", false)
            .expect("string-form transition version id must decode");
        assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
    }

    #[test]
@@ -4152,16 +4361,14 @@ mod tests {
        }
        .into_fileinfo("b", "k", false);
        assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
    }

    #[test]
-    fn delete_marker_free_version_transition_version_id_unparseable_stays_readable() {
-        // A malformed tier version id must not make a free-version record corrupt:
-        // it decodes to None and stays readable. Otherwise free-version expiry
-        // fails and the remote-tier object leaks.
+    fn delete_marker_free_version_transition_version_id_opaque_text_is_preserved() {
        let mut sys = HashMap::new();
        insert_bytes(&mut sys, SUFFIX_FREE_VERSION, vec![]);
-        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"not-a-uuid".to_vec());
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"opaque-generation-42".to_vec());
        insert_bytes(&mut sys, SUFFIX_TRANSITION_TIER, b"WARM".to_vec());
        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_OBJECTNAME, b"remote-object".to_vec());
        let fi = MetaDeleteMarker {
@@ -4172,8 +4379,9 @@ mod tests {
        .into_fileinfo("b", "k", false);

        assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version.as_deref(), Some("opaque-generation-42"));
        fi.validate_for_metadata_read()
-            .expect("free-version record with an unparseable tier id must remain readable");
+            .expect("free-version record with an opaque tier id must remain readable");
    }

    #[test]
@@ -4193,6 +4401,7 @@ mod tests {
        .into_fileinfo("b", "k", false);

        assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
    }

    #[test]
--- a/Show More
+++ b/Show More