style: restore rebalance formatting

Merge branch 'main' into cxymds/fix-1933-heal-progress-model
Signed-off-by: houseme <housemecn@gmail.com>
2026-08-23 04:39:04 +00:00 · 2026-08-23 12:34:48 +08:00 · 2026-08-23 12:31:09 +08:00 · 2026-08-23 12:29:52 +08:00 · 2026-08-23 12:09:06 +08:00 · 2026-08-23 12:07:25 +08:00
141 changed files with 12728 additions and 1942 deletions
@@ -1,2 +1,2 @@
-sha256-darwin=b4ae71aa894e5c7795ae3eb8116f1777a7601d0f5db3898be2e48faf3329bd9b
-sha256-linux=433debd9d9defa832986269abdf0f1d131597b2d7a417ce930e17c1fd47d85ba
+sha256-darwin=9f767b37ed8b1c82da62ea441462d75487785c8086e56f08fb6f6cd89c6e2e52
+sha256-linux=fbdaf42b220958d4b1e8880e0f8b5a7992d38e21051bb60596dd4538424757d6
@@ -36,6 +36,7 @@ script-tests: ## Run shell script tests
 	./scripts/test_manual_transition_runbooks.sh
 	./scripts/check_embedded_secrets.sh --self-test
 	python3 ./scripts/check_test_wiring.py --self-test
+	python3 ./scripts/check_scheduled_validation_freshness.py --self-test
 	python3 ./scripts/s3-tests/test_report_compat.py
 	bash -n ./scripts/validate_object_data_cache_cold_stampede.sh
 	python3 ./scripts/check_object_data_cache_follower_samples.py --self-test
@@ -14,9 +14,10 @@

 name: "Schedule Failure Issue"
 description: >-
-  Open (or update) a tracking issue when a scheduled workflow run fails.
+  Open (or update) a tracking issue when a scheduled workflow run fails or
+  does not complete normally.
  Dedupes by workflow name: if an open issue titled
-  "[scheduled-failure] <workflow name>" already exists, the failure is
+  "[scheduled-failure] <workflow name>" already exists, the result is
  appended as a comment; otherwise a new issue is created. This is the
  single alerting mechanism for all scheduled pipelines (backlog#1149 ci-8).

@@ -38,6 +39,30 @@ inputs:
      Set to an empty string to skip labeling.
    required: false
    default: "infrastructure"
+  source-run-id:
+    description: "Run ID to report. Defaults to the current workflow run."
+    required: false
+    default: ${{ github.run_id }}
+  source-run-attempt:
+    description: "Run attempt to report. Defaults to the current attempt."
+    required: false
+    default: ${{ github.run_attempt }}
+  source-event:
+    description: "Trigger event of the run being reported."
+    required: false
+    default: ${{ github.event_name }}
+  source-ref-name:
+    description: "Ref name of the run being reported."
+    required: false
+    default: ${{ github.ref_name }}
+  source-sha:
+    description: "Commit SHA of the run being reported."
+    required: false
+    default: ${{ github.sha }}
+  details-file:
+    description: "Optional Markdown file appended to the issue body."
+    required: false
+    default: ""

 runs:
  using: "composite"
@@ -48,17 +73,22 @@ runs:
        GH_TOKEN: ${{ inputs.github-token }}
        WORKFLOW_NAME: ${{ inputs.workflow-name }}
        ISSUE_LABEL: ${{ inputs.label }}
+        SOURCE_RUN_ID: ${{ inputs.source-run-id }}
+        SOURCE_RUN_ATTEMPT: ${{ inputs.source-run-attempt }}
+        SOURCE_EVENT: ${{ inputs.source-event }}
+        SOURCE_REF_NAME: ${{ inputs.source-ref-name }}
+        SOURCE_SHA: ${{ inputs.source-sha }}
+        DETAILS_FILE: ${{ inputs.details-file }}
      run: |
        set -euo pipefail

        title="[scheduled-failure] ${WORKFLOW_NAME}"
-        run_url="${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/actions/runs/${GITHUB_RUN_ID}"
+        run_url="${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/actions/runs/${SOURCE_RUN_ID}"

-        # Failed job names for this run attempt. The alert job runs while the
-        # run as a whole is still in progress, so inspect the jobs that have
-        # already completed with a non-success conclusion.
+        # Inspect the reported run attempt. It can be the current in-workflow
+        # failure or a completed run observed by the external watchdog.
        failed_jobs="$(gh api \
-          "repos/${GITHUB_REPOSITORY}/actions/runs/${GITHUB_RUN_ID}/attempts/${GITHUB_RUN_ATTEMPT}/jobs" \
+          "repos/${GITHUB_REPOSITORY}/actions/runs/${SOURCE_RUN_ID}/attempts/${SOURCE_RUN_ATTEMPT}/jobs" \
          --paginate \
          --jq '.jobs[]
                | select(.conclusion == "failure" or .conclusion == "timed_out" or .conclusion == "cancelled")
@@ -67,15 +97,26 @@ runs:
          failed_jobs="- (failed job not recorded yet — see the run page)"
        fi

+        details=""
+        if [ -n "${DETAILS_FILE}" ]; then
+          if [ -f "${DETAILS_FILE}" ]; then
+            details="$(cat "${DETAILS_FILE}")"
+          else
+            details="Details file was not available: \`${DETAILS_FILE}\`"
+          fi
+        fi
+
        body="$(cat <<EOF
-        Scheduled run of **${WORKFLOW_NAME}** failed.
+        Run of **${WORKFLOW_NAME}** did not complete successfully.

-        - Run: ${run_url} (attempt ${GITHUB_RUN_ATTEMPT})
-        - Event: \`${GITHUB_EVENT_NAME}\`
-        - Ref: \`${GITHUB_REF_NAME}\` @ \`${GITHUB_SHA}\`
+        - Run: ${run_url} (attempt ${SOURCE_RUN_ATTEMPT})
+        - Event: \`${SOURCE_EVENT}\`
+        - Ref: \`${SOURCE_REF_NAME}\` @ \`${SOURCE_SHA}\`

-        Failed jobs:
+        Non-success jobs:
        ${failed_jobs}
+
+        ${details}
        EOF
        )"

@@ -0,0 +1,14 @@
+[
+  { "workflow": ".github/workflows/audit.yml", "max_age_hours": 36 },
+  { "workflow": ".github/workflows/build.yml", "max_age_hours": 192 },
+  { "workflow": ".github/workflows/ci.yml", "max_age_hours": 192 },
+  { "workflow": ".github/workflows/coverage.yml", "max_age_hours": 192 },
+  { "workflow": ".github/workflows/e2e-replication-nightly.yml", "max_age_hours": 36 },
+  { "workflow": ".github/workflows/e2e-s3tests.yml", "max_age_hours": 192 },
+  { "workflow": ".github/workflows/fuzz.yml", "max_age_hours": 36 },
+  { "workflow": ".github/workflows/mint.yml", "max_age_hours": 192 },
+  { "workflow": ".github/workflows/minio-interop.yml", "max_age_hours": 36 },
+  { "workflow": ".github/workflows/nightly-gnu.yml", "max_age_hours": 36 },
+  { "workflow": ".github/workflows/performance-ab.yml", "max_age_hours": 36 },
+  { "workflow": ".github/workflows/runner-hygiene.yml", "max_age_hours": 792 }
+]
@@ -46,7 +46,7 @@ on:
    # advisory could sit unnoticed for seven days. The check list is unchanged —
    # splitting it into a light daily advisories-only run and a weekly full run
    # would create runs where sources/bans/licenses go unverified.
-    - cron: '0 3 * * *' # Daily 03:00 UTC (staggered after the midnight ci/build crons)
+    - cron: '23 3 * * *' # Daily 03:23 UTC
  workflow_dispatch:

 permissions:
@@ -52,7 +52,7 @@ on:
      - ".dockerignore"
      - "flake.lock"
  schedule:
-    - cron: "0 1 * * 0" # Weekly on Sunday 01:00 UTC (staggered after the ci.yml midnight cron)
+    - cron: "13 1 * * 0" # Weekly on Sunday 01:13 UTC
  workflow_dispatch:
    inputs:
      build_docker:
@@ -1032,3 +1032,23 @@ jobs:

          echo "🎉 Released $TAG successfully!"
          echo "📄 Release URL: ${{ needs.create-release.outputs.release_url }}"
+
+  alert-on-failure:
+    name: Alert on scheduled failure
+    needs: [build-check, prepare-platform-matrix, build-rustfs, build-summary]
+    if: >-
+      always() && github.event_name == 'schedule' &&
+      (contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled'))
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+      issues: write
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+      - name: Open or update failure-tracking issue
+        uses: ./.github/actions/schedule-failure-issue
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -94,6 +94,9 @@ concurrency:

 env:
  CARGO_TERM_COLOR: always
+  # Swatinem/rust-cache hashes every RUST* variable. Keep this aligned with
+  # ci.yml or the writer and readers use disjoint cache keys.
+  RUST_BACKTRACE: 1

 jobs:
  # Readers: test-and-lint, test-ilm-integration-serial, build-rustfs-debug-binary,
@@ -101,7 +104,7 @@ jobs:
  warm-ci-dev:
    name: Warm ci-dev
    runs-on: sm-standard-4
-    timeout-minutes: 90
+    timeout-minutes: 120
    env:
      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
    steps:
@@ -191,7 +194,7 @@ jobs:
  warm-ci-feat-rio:
    name: Warm ci-feat-rio
    runs-on: sm-standard-4
-    timeout-minutes: 90
+    timeout-minutes: 120
    env:
      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
    steps:
@@ -219,7 +222,7 @@ jobs:
  warm-ci-feat-proto:
    name: Warm ci-feat-proto
    runs-on: sm-standard-4
-    timeout-minutes: 90
+    timeout-minutes: 120
    env:
      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
    steps:
@@ -126,7 +126,10 @@ jobs:
        run: ./scripts/check_embedded_secrets.sh

      - name: Check test wiring
-        run: python3 ./scripts/check_test_wiring.py
+        run: |
+          python3 ./scripts/check_test_wiring.py --self-test
+          python3 ./scripts/check_scheduled_validation_freshness.py --self-test
+          python3 ./scripts/check_test_wiring.py

      - name: Check no planning docs committed
        run: ./scripts/check_no_planning_docs.sh
@@ -59,7 +59,7 @@ on:
  merge_group:
    types: [ checks_requested ]
  schedule:
-    - cron: "0 0 * * 0" # Weekly on Sunday at midnight UTC
+    - cron: "11 0 * * 0" # Weekly on Sunday 00:11 UTC
  workflow_dispatch:

 permissions:
@@ -161,7 +161,10 @@ jobs:
        run: ./scripts/check_embedded_secrets.sh

      - name: Check test wiring
-        run: python3 ./scripts/check_test_wiring.py
+        run: |
+          python3 ./scripts/check_test_wiring.py --self-test
+          python3 ./scripts/check_scheduled_validation_freshness.py --self-test
+          python3 ./scripts/check_test_wiring.py

      - name: Check no planning docs committed
        run: ./scripts/check_no_planning_docs.sh
@@ -1032,3 +1035,37 @@ jobs:
          path: artifacts/s3tests-single/**
          if-no-files-found: ignore
          retention-days: 3
+
+  alert-on-failure:
+    name: Alert on scheduled failure
+    needs:
+      - typos
+      - quick-checks
+      - test-and-lint
+      - test-ilm-integration-serial
+      - test-and-lint-rio-v2
+      - test-and-lint-protocols
+      - build-rustfs-debug-binary
+      - build-rustfs-debug-binary-rio-v2
+      - uring-integration
+      - e2e-tests
+      - e2e-full
+      - e2e-tests-rio-v2
+      - s3-implemented-tests
+      - s3-lifecycle-behavior-tests
+    if: >-
+      always() && github.event_name == 'schedule' &&
+      (contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled'))
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+      issues: write
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+      - name: Open or update failure-tracking issue
+        uses: ./.github/actions/schedule-failure-issue
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -37,7 +37,7 @@ on:
    # build (01:00), e2e-s3tests (02:00), audit (03:00), nix-flake-update
    # (05:00), mint (06:00), and the daily fuzz (02:00), minio-interop (03:17),
    # e2e-replication-nightly (04:00) and performance-ab (06:00) lanes.
-    - cron: "0 7 * * 0"
+    - cron: "43 7 * * 0"

 # Only alert-on-failure needs more than read access; it declares its own
 # job-level `issues: write`.
@@ -40,7 +40,7 @@ on:
  schedule:
    # 04:00 UTC nightly — staggered clear of fuzz/e2e-s3tests (02:00),
    # stale (01:30) and performance-ab (06:00).
-    - cron: "0 4 * * *"
+    - cron: "29 4 * * *"

 # Only alert-on-failure needs more than read access; it declares its own
 # job-level `issues: write`.
@@ -196,6 +196,9 @@ jobs:
          cache-save-if: 'false'
          install-build-packaging-tools: 'false'

+      - name: Verify protocol socket oracle
+        run: ss -tn state CLOSE-WAIT >/dev/null
+
      # The suite owns fixed protocol ports and serializes its internal cases.
      - name: Verify protocol e2e membership
        env:
@@ -90,10 +90,14 @@ on:
        description: "Optional pytest -m expression"
        required: false
        default: ""
+      testexpr:
+        description: "Optional pytest -k expression"
+        required: false
+        default: ""
  schedule:
    # Weekly full sweep (Sunday 02:00 UTC): full suite, run against BOTH the
    # single-node and the 4-node distributed topologies (matrix below).
-    - cron: "0 2 * * 0"
+    - cron: "19 2 * * 0"

 env:
  # main user
@@ -116,6 +120,7 @@ env:
  XDIST: ${{ github.event.inputs.xdist || '4' }}
  MAXFAIL: ${{ github.event.inputs.maxfail || '0' }}
  MARKEXPR: ${{ github.event.inputs.markexpr || '' }}
+  TESTEXPR: ${{ github.event.inputs.testexpr || '' }}
  S3_SHARD_COUNT: ${{ github.event_name == 'schedule' && '4' || github.event.inputs.shard-count || '1' }}
  TEST_TIMEOUT: "300"

@@ -269,14 +274,20 @@ jobs:
          EOF

          cat > haproxy.cfg <<'EOF'
+          global
+            log stdout format raw local0 info
+
          defaults
            mode http
+            log global
+            log-format '%ci:%cp [%tr] %ft %b/%s %TR/%Tw/%Tc/%Tr/%Ta %ST %B %tsc %HM %HP'
            timeout connect 5s
            timeout client  30s
            timeout server  30s

          frontend fe_s3
            bind *:9000
+            option http-buffer-request
            default_backend be_s3

          backend be_s3
@@ -314,6 +325,7 @@ jobs:
          XDIST="${XDIST}" \
          MAXFAIL="${MAXFAIL}" \
          MARKEXPR="${MARKEXPR}" \
+          TESTEXPR="${TESTEXPR}" \
          ./scripts/s3-tests/run.sh

      - name: Publish compatibility report
@@ -30,7 +30,7 @@ on:
      - "Cargo.lock"
      - ".github/workflows/fuzz.yml"
  schedule:
-    - cron: "0 2 * * *"
+    - cron: "17 2 * * *"
  workflow_dispatch:
    inputs:
      profile:
@@ -121,3 +121,21 @@ jobs:
          cargo nextest run --run-ignored ignored-only --no-tests=fail \
            -p "$INTEROP_PACKAGE" --features "$INTEROP_FEATURES" \
            -E "$INTEROP_FILTER"
+
+  alert-on-failure:
+    name: Alert on scheduled failure
+    needs: [minio-interop]
+    if: always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure')
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+      issues: write
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+      - name: Open or update failure-tracking issue
+        uses: ./.github/actions/schedule-failure-issue
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -45,13 +45,6 @@
 # docker-capable self-hosted `dind-sm-standard-2` label was the alternative but
 # has fewer cores and reintroduces fleet-state risk for no reliability gain.

-# DISABLED. This workflow is switched off in the repository's Actions settings
-# (state: disabled_manually) and does not run on any trigger, including its cron
-# and workflow_dispatch. That state lives in GitHub's UI and is invisible when
-# reading this file, which has already misled at least one audit — hence this
-# banner. Re-enabling is a UI action; anyone doing so should first check that the
-# workflow still matches the current CI layout. See rustfs/backlog#1603.
-#
 name: mint

 on:
@@ -70,13 +63,13 @@ on:
          - core
          - full
      mint-image:
-        description: "Mint image reference"
+        description: "Mint image reference (empty = pinned default)"
        required: false
-        default: "minio/mint:edge"
+        default: ""
  schedule:
    # Weekly, after the Sunday s3-tests full sweep (starts 02:00 UTC, up to
    # 3h) has finished, so the two never contend for the same runner pool.
-    - cron: "0 6 * * 0"
+    - cron: "41 6 * * 0"

 env:
  S3_ACCESS_KEY: rustfsadmin-ci
@@ -16,7 +16,7 @@ name: Nightly GNU Build

 on:
  schedule:
-    - cron: "0 0 * * *"
+    - cron: "7 0 * * *"
      timezone: "Asia/Shanghai"
  workflow_dispatch:

@@ -194,3 +194,23 @@ jobs:

      - name: Run HA leader failover live checks (three-node Raft cluster in Docker)
        run: bash scripts/test/vault_ha_kms_live.sh
+
+  alert-on-failure:
+    name: Alert on scheduled failure
+    needs: [build, kms-vault-lane, kms-vault-ha-failover]
+    if: >-
+      always() && github.event_name == 'schedule' &&
+      (contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled'))
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+      issues: write
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+      - name: Open or update failure-tracking issue
+        uses: ./.github/actions/schedule-failure-issue
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -22,22 +22,15 @@
 # correctness cost (e.g. the #4221 fsync durability fix) is recorded, not
 # blocked (rustfs/backlog#935 correction 1).

-# DISABLED. This workflow is switched off in the repository's Actions settings
-# (state: disabled_manually) and does not run on any trigger, including its cron
-# and workflow_dispatch. That state lives in GitHub's UI and is invisible when
-# reading this file, which has already misled at least one audit — hence this
-# banner. Re-enabling is a UI action; anyone doing so should first check that the
-# workflow still matches the current CI layout. See rustfs/backlog#1603.
-#
 name: Performance A/B

 on:
  schedule:
-    - cron: "0 6 * * *" # 06:00 UTC nightly, against main
+    - cron: "31 6 * * *" # 06:31 UTC nightly, against main
  workflow_dispatch:
    inputs:
      duration:
-        description: "warp duration per round (short by default to fit the double-build budget)"
+        description: "warp duration per round"
        required: false
        default: "12s"
        type: string
@@ -46,12 +39,8 @@ on:
        required: false
        default: false
        type: boolean
-  push:
-    # Every main commit pre-builds and caches its release binary (perf-3) so the
-    # nightly A/B restores a ready baseline instead of paying the double build.
-    branches: [main]
-
 permissions:
+  actions: read
  contents: read

 env:
@@ -59,83 +48,19 @@ env:
  RUST_BACKTRACE: 1

 jobs:
-  # perf-3: on every push to main, build the release binary once and cache it
-  # keyed by commit SHA (rustfs-baseline-<sha>). The warp-ab measurements
-  # restore this instead of paying the ~32min-per-side source
-  # build. That double build is what pushed the expanded 24-cell nightly past its
-  # ceiling — 2026-07-11..07-14 all cancelled on the 120min timeout. Incremental
-  # builds off the shared cargo cache keep each push cheap, and building on the
-  # same sm-standard-2 runner the A/B measures on guarantees the cached binary is
-  # ABI-identical. Do NOT source this from build.yml's per-merge artifact: those
-  # are cancelled ~7/8 of the time and are not a reliable baseline.
-  build-baseline-cache:
-    name: Build + cache baseline binary
-    if: github.event_name == 'push'
-    runs-on: sm-standard-2
-    # Latest-wins: consumers only ever restore the binary for the *current*
-    # origin/main tip, so when pushes land faster than the ~65min build, a
-    # superseded build's output is dead weight — cancel it instead of stacking
-    # hour-long jobs on the shared runner pool. A skipped intermediate SHA at
-    # most costs one same-commit self-heal in the A/B job.
-    concurrency:
-      group: perf-baseline-build-main
-      cancel-in-progress: true
-    # #4806 put thin LTO + codegen-units=1 on [profile.release], pushing a
-    # single release build past 60min on this runner — every cache build on
-    # 2026-07-15 died on the old 60min ceiling ("exceeded the maximum execution
-    # time of 1h0m0s") and the cache never populated. The measured binary must
-    # keep the production profile, so the budget absorbs the build instead.
-    timeout-minutes: 100
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
-        with:
-          persist-credentials: false
-
-      - name: Setup Rust environment
-        uses: ./.github/actions/setup
-        with:
-          rust-version: stable
-          cache-shared-key: warp-ab-${{ hashFiles('**/Cargo.lock') }}
-          cache-save-if: ${{ github.ref == 'refs/heads/main' }}
-
-      - name: Build release rustfs
-        run: cargo build --release --bin rustfs
-
-      - name: Stage binary for cache
-        run: |
-          set -euo pipefail
-          mkdir -p baseline-bin
-          cp target/release/rustfs baseline-bin/rustfs
-
-      - name: Cache baseline binary by SHA
-        uses: actions/cache/save@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6
-        with:
-          path: baseline-bin/rustfs
-          key: rustfs-baseline-${{ github.sha }}
-
  warp-ab:
    name: Warp A/B budget gate
-    # Always run on schedule / manual dispatch. Never on push — that event only
-    # feeds build-baseline-cache above.
-    if: >-
-      github.event_name == 'schedule' ||
-      github.event_name == 'workflow_dispatch'
    runs-on: sm-standard-2
-    # With perf-3's cached baseline binary the common (cache-hit) nightly is
-    # measurement-only and finishes well under 50min. This ceiling stays
-    # generous only to absorb the same-commit cache-miss self-heal (~65min
-    # single build with the post-#4806 LTO profile + measurement). A timeout
-    # surfaces via the alert-on-failure job (it fires on cancelled/timed-out,
-    # not just failure). perf-6 recalibrates the budget once the noise study
-    # lands.
-    timeout-minutes: 120
+    # A normal nightly restores the last successful binary and builds only the
+    # candidate; daily access keeps that cache warm. A cache miss may build both
+    # and needs room for the A/B run plus artifact and cache publication.
+    timeout-minutes: 180
    steps:
      - name: Checkout repository
        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
        with:
          persist-credentials: false
-          fetch-depth: 0 # baseline is built from origin/main
+          fetch-depth: 0 # baseline may be an earlier successful scheduled head

      - name: Setup Rust environment
        uses: ./.github/actions/setup
@@ -163,24 +88,55 @@ jobs:
          fi
          echo "allow_regression=$allow" >> "$GITHUB_OUTPUT"

-      # perf-3: resolve the commits so the cache can be keyed by SHA. The
-      # baseline is origin/main; the candidate is the checked-out ref. On the
-      # nightly (checkout == main) they are the same commit, so one cached binary
-      # serves both phases and the run does zero source builds.
+      # A failed regression run must keep comparing against the last known-good
+      # scheduled head. Otherwise the next nightly would absorb the regression
+      # into its baseline and turn green without a fix.
+      - name: Find last successful scheduled baseline
+        id: scheduled_baseline
+        if: github.event_name == 'schedule'
+        uses: actions/github-script@ed597411d8f924073f98dfc5c65a23a2325f34cd # v8
+        with:
+          result-encoding: string
+          script: |
+            const { data } = await github.rest.actions.listWorkflowRuns({
+              owner: context.repo.owner,
+              repo: context.repo.repo,
+              workflow_id: "performance-ab.yml",
+              event: "schedule",
+              status: "success",
+              per_page: 1,
+            });
+            return data.workflow_runs[0]?.head_sha ?? "";
+
+      # Manual runs compare a selected ref with current main. Scheduled runs
+      # compare current main with the last successful scheduled head. With no
+      # history, the first run measures the candidate against itself and seeds
+      # that head only if the complete rig succeeds.
      - name: Resolve baseline / candidate commits
        id: commits
+        env:
+          SCHEDULED_BASELINE_SHA: ${{ steps.scheduled_baseline.outputs.result }}
        run: |
          set -euo pipefail
-          baseline_sha="$(git rev-parse origin/main)"
          candidate_sha="$(git rev-parse HEAD)"
+          if [[ "${{ github.event_name }}" == "schedule" ]]; then
+            baseline_sha="${SCHEDULED_BASELINE_SHA:-$candidate_sha}"
+            if ! git merge-base --is-ancestor "$baseline_sha" "$candidate_sha"; then
+              echo "::error::scheduled baseline $baseline_sha is not an ancestor of candidate $candidate_sha" >&2
+              exit 1
+            fi
+          else
+            baseline_sha="$(git rev-parse origin/main)"
+          fi
+          git cat-file -e "${baseline_sha}^{commit}"
          echo "baseline_sha=$baseline_sha" >> "$GITHUB_OUTPUT"
          echo "candidate_sha=$candidate_sha" >> "$GITHUB_OUTPUT"
          echo "baseline commit: $baseline_sha"
          echo "candidate commit: $candidate_sha"

-      # Exact-key restore of the baseline binary built by build-baseline-cache
-      # when origin/main last landed. A miss (binary evicted or not built yet)
-      # leaves cache-hit unset and the rig falls back to a source build.
+      # Exact-key restore of the candidate binary saved by its successful
+      # scheduled run. A miss leaves cache-hit unset and falls back to a source
+      # build of that known-good head.
      - name: Restore cached baseline binary
        id: baseline_cache
        uses: actions/cache/restore@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6
@@ -270,11 +226,11 @@ jobs:
            elif [[ "$selfheal_built" == "true" ]]; then
              base_src="source build (cache self-heal, saved as rustfs-baseline-$baseline_sha)"
            else
-              base_src="isolated origin/main source build (saved as rustfs-baseline-$baseline_sha)"
+              base_src="isolated baseline source build (saved as rustfs-baseline-$baseline_sha)"
            fi
            if [[ "$candidate_sha" == "$baseline_sha" ]]; then
-              # Nightly on main: the candidate is the same commit as the baseline,
-              # so reuse the one binary for both phases and skip all builds.
+              # No commits landed since the last successful baseline, so reuse
+              # the one binary for both phases and measure only rig drift.
              args+=(--candidate-bin "$base_bin")
              cand_src="same binary as baseline (same commit)"
            elif [[ "$candidate_built" == "true" ]]; then
@@ -362,6 +318,23 @@ jobs:
            fi
          } >> "$GITHUB_STEP_SUMMARY"

+      - name: Stage successful candidate baseline
+        if: >-
+          steps.ab.outputs.status == '0' &&
+          steps.commits.outputs.baseline_sha != steps.commits.outputs.candidate_sha
+        run: |
+          set -euo pipefail
+          cp candidate-bin/rustfs baseline-bin/rustfs
+
+      - name: Cache successful candidate baseline
+        if: >-
+          steps.ab.outputs.status == '0' &&
+          steps.commits.outputs.baseline_sha != steps.commits.outputs.candidate_sha
+        uses: actions/cache/save@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6
+        with:
+          path: baseline-bin/rustfs
+          key: rustfs-baseline-${{ steps.commits.outputs.candidate_sha }}
+
      # Scheduled failure alerting is handled by the alert-on-failure job below
      # (perf-2 consuming ci-8's schedule-failure-issue composite action).

@@ -30,7 +30,7 @@ name: Runner Hygiene

 on:
  schedule:
-    - cron: "0 6 1 * *" # Monthly, 1st at 06:00 UTC (after the daily audit cron)
+    - cron: "37 6 1 * *" # Monthly, 1st at 06:37 UTC
  workflow_dispatch:

 permissions:
@@ -0,0 +1,57 @@
+# Copyright 2024 RustFS Team
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: Scheduled Validation Freshness
+
+on:
+  schedule:
+    - cron: "47 23 * * *"
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+concurrency:
+  group: scheduled-validation-freshness
+  cancel-in-progress: false
+
+jobs:
+  check-freshness:
+    name: Check scheduled validation freshness
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      actions: read
+      contents: read
+      issues: write
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+      - name: Check latest scheduled runs
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          set +e
+          python3 scripts/check_scheduled_validation_freshness.py \
+            --report "${RUNNER_TEMP}/scheduled-validation-freshness.md"
+          status=$?
+          cat "${RUNNER_TEMP}/scheduled-validation-freshness.md" >> "${GITHUB_STEP_SUMMARY}"
+          exit "${status}"
+      - name: Open or update freshness issue
+        if: failure()
+        uses: ./.github/actions/schedule-failure-issue
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          details-file: ${{ runner.temp }}/scheduled-validation-freshness.md
@@ -0,0 +1,63 @@
+# Copyright 2024 RustFS Team
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: Scheduled Validation Watchdog
+
+on:
+  workflow_run:
+    workflows:
+      - "Security Audit"
+      - "Build and Release"
+      - "Continuous Integration"
+      - "coverage"
+      - "e2e-nightly"
+      - "e2e-s3tests"
+      - "Fuzz"
+      - "mint"
+      - "minio-interop"
+      - "Nightly GNU Build"
+      - "Performance A/B"
+      - "Runner Hygiene"
+    types: [completed]
+
+permissions:
+  contents: read
+
+jobs:
+  alert-on-incomplete-run:
+    name: Alert on incomplete scheduled run
+    if: >-
+      github.event.workflow_run.event == 'schedule' &&
+      github.event.workflow_run.conclusion != 'success' &&
+      github.event.workflow_run.conclusion != 'failure'
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      actions: read
+      contents: read
+      issues: write
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+      - name: Open or update incomplete-run issue
+        uses: ./.github/actions/schedule-failure-issue
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}
+          workflow-name: ${{ github.event.workflow_run.name }}
+          source-run-id: ${{ github.event.workflow_run.id }}
+          source-run-attempt: ${{ github.event.workflow_run.run_attempt }}
+          source-event: ${{ github.event.workflow_run.event }}
+          source-ref-name: ${{ github.event.workflow_run.head_branch }}
+          source-sha: ${{ github.event.workflow_run.head_sha }}
@@ -30,7 +30,8 @@ make build-docker BUILD_OS=ubuntu22.04
 - Crate membership: `Cargo.toml` `[workspace].members`
 - Architecture, layering, crate map: [ARCHITECTURE.md](ARCHITECTURE.md)
 - Migration guardrails & readiness contracts: [docs/architecture/](docs/architecture/README.md)
- CI gates: `.github/workflows/ci.yml` (source of truth; never copy its steps into docs)
+- CI workflow steps: `.github/workflows/`; event, timeout, and required-status
+  matrix: [docs/testing/ci-gates.md](docs/testing/ci-gates.md)
 - Test-layer taxonomy, per-layer entry commands, serial/nextest rules, flake
  policy: [docs/testing/README.md](docs/testing/README.md)
 - Tier/ILM transition debugging (xl.meta inspection, versionId tracing):
@@ -70,6 +70,8 @@ make pre-pr

 > For the full test-layer taxonomy (unit / ecstore black-box / e2e / s3s-e2e / S3 compatibility / chaos / fuzz / bench), each layer's entry command, the naming conventions the migration gate depends on, and the serial/nextest rules, see [docs/testing/README.md](docs/testing/README.md).

+> For the event, timeout, required-status, and local reproduction matrix, see [docs/testing/ci-gates.md](docs/testing/ci-gates.md).
+
 ### 🔒 Automated Pre-commit Hooks
 #### What `make pre-commit` and `make pre-pr` actually run

@@ -1858,9 +1858,9 @@ dependencies = [

 [[package]]
 name = "cc"
-version = "1.4.3"
+version = "1.4.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "509591b7bcd67f4ef775afad7662703b4935daaa6ec0e5605cfb1090b32a2b6d"
+checksum = "0ad534f4357a5264cce5019c989cf66a4f0dc4e0d1b1d15f8aacec0ff7360273"
 dependencies = [
 "find-msvc-tools",
 "jobserver",
@@ -2522,12 +2522,6 @@ dependencies = [
 "subtle",
 ]

-[[package]]
-name = "cty"
-version = "0.2.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b365fabc795046672053e29c954733ec3b05e4be654ab130fe8f1f94d7051f35"
-
 [[package]]
 name = "curve25519-dalek"
 version = "4.1.3"
@@ -5988,15 +5982,6 @@ version = "0.2.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b6d2cec3eae94f9f509c767b45932f1ada8350c4bdb85af2fcab4a3c14807981"

-[[package]]
-name = "libmimalloc-sys"
-version = "0.1.49"
-source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11#6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11"
-dependencies = [
- "cc",
- "cty",
-]
-
 [[package]]
 name = "libredox"
 version = "0.1.20"
@@ -6397,14 +6382,6 @@ dependencies = [
 "synstructure 0.13.2",
 ]

-[[package]]
-name = "mimalloc"
-version = "0.1.52"
-source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11#6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11"
-dependencies = [
- "libmimalloc-sys",
-]
-
 [[package]]
 name = "mime"
 version = "0.3.17"
@@ -9162,13 +9139,11 @@ dependencies = [
 "insta",
 "jiff",
 "libc",
- "libmimalloc-sys",
 "libsystemd",
 "matchit 0.9.2",
 "md-5 0.11.0",
 "metrics",
 "metrics-util",
- "mimalloc",
 "mime_guess",
 "opentelemetry",
 "opentelemetry_sdk",
@@ -9204,6 +9179,8 @@ dependencies = [
 "rustfs-lock",
 "rustfs-log-analyzer",
 "rustfs-madmin",
+ "rustfs-mimalloc",
+ "rustfs-mimalloc-sys",
 "rustfs-notify",
 "rustfs-object-capacity",
 "rustfs-object-data-cache",
@@ -9875,6 +9852,24 @@ dependencies = [
 "tokio",
 ]

+[[package]]
+name = "rustfs-mimalloc"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a406f4aa07084301d485beec873af6dccc8e3f8762da244743df92038b1db1a6"
+dependencies = [
+ "rustfs-mimalloc-sys",
+]
+
+[[package]]
+name = "rustfs-mimalloc-sys"
+version = "0.5.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c3051b819175f58445d4c369a72f0ab88149f3885ba8bea2aff3be01f53fe7cd"
+dependencies = [
+ "cc",
+]
+
 [[package]]
 name = "rustfs-notify"
 version = "1.0.0-rc.3"
@@ -12688,6 +12683,7 @@ dependencies = [
 "js-sys",
 "rand 0.10.2",
 "serde_core",
+ "sha1_smol",
 "wasm-bindgen",
 ]

@@ -350,8 +350,8 @@ russh-sftp = "2.4.0"
 dav-server = "0.11.0"

 # Performance Analysis and Memory Profiling
-mimalloc = { version = "0.1.52", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11" }
-libmimalloc-sys = { version = "0.1.49", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11", features = ["extended"] }
+rustfs-mimalloc = { version = "0.5.0" }
+rustfs-mimalloc-sys = { version = "0.5.0" }
 hotpath = { version = "0.23.3", default-features = false }
 # Snapshot testing for output format regression detection
 insta = { version = "1.48" }
@@ -901,6 +901,10 @@ pub struct Metrics {
    scanner_cycle_max_duration_millis: AtomicU64,
    scanner_cycle_max_objects: AtomicU64,
    scanner_cycle_max_directories: AtomicU64,
+    scanner_cycle_timeout_total: AtomicU64,
+    scanner_cycle_recovery_required_total: AtomicU64,
+    scanner_cycle_last_progress_age_seconds: AtomicU64,
+    scanner_leader_lease_without_progress: AtomicBool,
    scanner_bitrot_cycle_enabled: AtomicBool,
    scanner_bitrot_cycle_millis: AtomicU64,
    scanner_checkpoint: Mutex<Option<ScannerCheckpointReport>>,
@@ -1370,6 +1374,14 @@ pub struct ScannerMetricsReport {
    #[serde(default)]
    pub cycle_max_directories: u64,
    #[serde(default)]
+    pub cycle_timeout_total: u64,
+    #[serde(default)]
+    pub cycle_recovery_required_total: u64,
+    #[serde(default)]
+    pub cycle_last_progress_age: u64,
+    #[serde(default)]
+    pub leader_lease_without_progress: bool,
+    #[serde(default)]
    pub bitrot_cycle_enabled: bool,
    #[serde(default)]
    pub bitrot_cycle_seconds: f64,
@@ -1430,6 +1442,9 @@ const OTEL_SCANNER_BUCKETS_SCANNED: &str = "rustfs_scanner_buckets_scanned_total
 const OTEL_SCANNER_CYCLES: &str = "rustfs_scanner_cycles_total";
 const OTEL_SCANNER_CYCLE_DURATION_SECONDS: &str = "rustfs_scanner_cycle_duration_seconds";
 const OTEL_SCANNER_BUCKET_DRIVE_DURATION_SECONDS: &str = "rustfs_scanner_bucket_drive_duration_seconds";
+const OTEL_SCANNER_CYCLE_TIMEOUT_TOTAL: &str = "rustfs_scanner_cycle_timeout_total";
+const OTEL_SCANNER_CYCLE_LAST_PROGRESS_AGE: &str = "rustfs_scanner_cycle_last_progress_age";
+const OTEL_SCANNER_LEADER_LEASE_WITHOUT_PROGRESS: &str = "rustfs_scanner_leader_lease_without_progress";

 fn scan_cycle_result_label(result: u8) -> &'static str {
    match result {
@@ -1913,6 +1928,10 @@ impl Metrics {
            scanner_cycle_max_duration_millis: AtomicU64::new(0),
            scanner_cycle_max_objects: AtomicU64::new(0),
            scanner_cycle_max_directories: AtomicU64::new(0),
+            scanner_cycle_timeout_total: AtomicU64::new(0),
+            scanner_cycle_recovery_required_total: AtomicU64::new(0),
+            scanner_cycle_last_progress_age_seconds: AtomicU64::new(0),
+            scanner_leader_lease_without_progress: AtomicBool::new(false),
            scanner_bitrot_cycle_enabled: AtomicBool::new(false),
            scanner_bitrot_cycle_millis: AtomicU64::new(0),
            scanner_checkpoint: Mutex::new(None),
@@ -2412,12 +2431,29 @@ impl Metrics {
            .store(cycle_max_objects.unwrap_or_default(), Ordering::Relaxed);
        self.scanner_cycle_max_directories
            .store(cycle_max_directories.unwrap_or_default(), Ordering::Relaxed);
+        self.scanner_leader_lease_without_progress.store(false, Ordering::Relaxed);
+        self.scanner_cycle_last_progress_age_seconds.store(0, Ordering::Relaxed);
+        metrics::gauge!(OTEL_SCANNER_LEADER_LEASE_WITHOUT_PROGRESS).set(0.0);
+        metrics::gauge!(OTEL_SCANNER_CYCLE_LAST_PROGRESS_AGE).set(0.0);
        self.scanner_bitrot_cycle_enabled
            .store(bitrot_cycle.is_some(), Ordering::Relaxed);
        self.scanner_bitrot_cycle_millis
            .store(bitrot_cycle.map(duration_millis_saturated).unwrap_or_default(), Ordering::Relaxed);
    }

+    pub fn record_scanner_cycle_timeout(&self, recovery_required: bool, progress_age: Duration) {
+        self.scanner_cycle_timeout_total.fetch_add(1, Ordering::Relaxed);
+        if recovery_required {
+            self.scanner_cycle_recovery_required_total.fetch_add(1, Ordering::Relaxed);
+        }
+        self.scanner_cycle_last_progress_age_seconds
+            .store(progress_age.as_secs(), Ordering::Relaxed);
+        self.scanner_leader_lease_without_progress.store(true, Ordering::Relaxed);
+        metrics::counter!(OTEL_SCANNER_CYCLE_TIMEOUT_TOTAL).increment(1);
+        metrics::gauge!(OTEL_SCANNER_CYCLE_LAST_PROGRESS_AGE).set(progress_age.as_secs_f64());
+        metrics::gauge!(OTEL_SCANNER_LEADER_LEASE_WITHOUT_PROGRESS).set(1.0);
+    }
+
    pub fn record_scanner_set_scan_state(&self, concurrency_limit: Option<usize>, queued: Option<usize>, active: Option<usize>) {
        if let Some(concurrency_limit) = concurrency_limit {
            self.scanner_set_scan_concurrency_limit
@@ -3265,6 +3301,10 @@ impl Metrics {
        m.cycle_max_duration_seconds = self.scanner_cycle_max_duration_millis.load(Ordering::Relaxed) as f64 / 1000.0;
        m.cycle_max_objects = self.scanner_cycle_max_objects.load(Ordering::Relaxed);
        m.cycle_max_directories = self.scanner_cycle_max_directories.load(Ordering::Relaxed);
+        m.cycle_timeout_total = self.scanner_cycle_timeout_total.load(Ordering::Relaxed);
+        m.cycle_recovery_required_total = self.scanner_cycle_recovery_required_total.load(Ordering::Relaxed);
+        m.cycle_last_progress_age = self.scanner_cycle_last_progress_age_seconds.load(Ordering::Relaxed);
+        m.leader_lease_without_progress = self.scanner_leader_lease_without_progress.load(Ordering::Relaxed);
        m.bitrot_cycle_enabled = self.scanner_bitrot_cycle_enabled.load(Ordering::Relaxed);
        m.bitrot_cycle_seconds = self.scanner_bitrot_cycle_millis.load(Ordering::Relaxed) as f64 / 1000.0;
        m.scan_checkpoint = match self.scanner_checkpoint.lock() {
@@ -4926,4 +4966,20 @@ mod tests {
        assert!(!report.bitrot_cycle_enabled);
        assert_eq!(report.bitrot_cycle_seconds, 0.0);
    }
+
+    #[tokio::test]
+    async fn scanner_cycle_timeout_metrics_reset_for_a_new_cycle() {
+        let metrics = Metrics::new();
+        metrics.record_scanner_cycle_timeout(true, Duration::from_secs(17));
+        let timed_out = metrics.report().await;
+        assert_eq!(timed_out.cycle_timeout_total, 1);
+        assert_eq!(timed_out.cycle_last_progress_age, 17);
+        assert!(timed_out.leader_lease_without_progress);
+
+        metrics.record_scanner_cycle_config(Duration::from_secs(60), None, Some(Duration::from_secs(1)), None, None);
+        let current = metrics.report().await;
+        assert_eq!(current.cycle_timeout_total, 1);
+        assert_eq!(current.cycle_last_progress_age, 0);
+        assert!(!current.leader_lease_without_progress);
+    }
 }
@@ -84,6 +84,12 @@ Current guidance:
 - `RUSTFS_SCANNER_CYCLE_MAX_OBJECTS` (canonical)
 - `RUSTFS_SCANNER_CYCLE_MAX_DIRECTORIES` (canonical)

+Scanner cycle budget controls:
+
+- When `RUSTFS_SCANNER_CYCLE_MAX_DURATION_SECS` is unset, the finite default is 1800 seconds (30 minutes), matching the scanner benchmark guidance.
+- An explicit `0` preserves the compatibility behavior of an unbounded runtime budget. Object and directory budgets likewise remain unbounded when explicitly set to `0`.
+- A timed-out cycle cancels cooperative scanner work, then fences its leader epoch before releasing the lease. An uncooperative I/O operation is dropped after the bounded shutdown window; its cursor is not claimed to be durable and the scanner reports `recovery-required` when the worker cannot stop cooperatively, the cycle state was not confirmed durable, or epoch fencing cannot be persisted.
+
 ## Mmap read environment aliases

 - `RUSTFS_OBJECT_MMAP_READ_ENABLE` (canonical)
@@ -143,9 +143,12 @@ pub const ENV_SCANNER_MAX_WAIT_SECS: &str = "RUSTFS_SCANNER_MAX_WAIT_SECS";
 /// Default scanner speed preset.
 pub const DEFAULT_SCANNER_SPEED: &str = "default";

-/// Default scanner cycle runtime budget.
-/// `0` keeps the existing unbounded per-cycle behavior.
-pub const DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS: u64 = 0;
+/// Default scanner cycle runtime budget when no override is configured.
+///
+/// An explicit `0` remains the compatibility escape hatch for an unbounded
+/// cycle. Keeping the unset default finite prevents a stalled scanner I/O
+/// operation from holding the leader lease forever.
+pub const DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS: u64 = 30 * 60;

 /// Default scanner per-cycle object budget.
 /// `0` keeps the existing unbounded per-cycle behavior.
@@ -585,9 +585,12 @@ impl VersionsHistogram {
    }
 }

-/// Replication statistics for a single target
-#[derive(Debug, Default, Clone, Serialize, Deserialize)]
-pub struct ReplicationStats {
+/// Replication statistics for a single target.
+///
+/// Renamed from `ReplicationStats`; serde field names are preserved
+/// byte-identically to maintain wire compatibility with existing snapshots.
+#[derive(Debug, Default, Clone, Serialize, Deserialize, PartialEq, Eq)]
+pub struct ReplicationTargetUsage {
    pub pending_size: u64,
    pub replicated_size: u64,
    pub failed_size: u64,
@@ -600,7 +603,7 @@ pub struct ReplicationStats {
    pub replicated_count: u64,
 }

-impl ReplicationStats {
+impl ReplicationTargetUsage {
    pub fn is_empty(&self) -> bool {
        let Self {
            pending_size,
@@ -636,7 +639,7 @@ impl ReplicationStats {
 /// Replication statistics for all targets
 #[derive(Debug, Default, Clone, Serialize, Deserialize)]
 pub struct ReplicationAllStats {
-    pub targets: HashMap<String, ReplicationStats>,
+    pub targets: HashMap<String, ReplicationTargetUsage>,
    pub replica_size: u64,
    pub replica_count: u64,
 }
@@ -649,7 +652,7 @@ impl ReplicationAllStats {
            targets,
        } = self;

-        *replica_size == 0 && *replica_count == 0 && targets.values().all(ReplicationStats::is_empty)
+        *replica_size == 0 && *replica_count == 0 && targets.values().all(ReplicationTargetUsage::is_empty)
    }

    #[deprecated(note = "use is_empty instead")]
@@ -2466,7 +2469,7 @@ mod tests {

    #[test]
    fn replication_stats_empty_checks_every_field() {
-        type SetField = fn(&mut ReplicationStats);
+        type SetField = fn(&mut ReplicationTargetUsage);

        let cases: [(&str, SetField); 10] = [
            ("pending_size", |stats| stats.pending_size = 1),
@@ -2481,9 +2484,9 @@ mod tests {
            ("replicated_count", |stats| stats.replicated_count = 1),
        ];

-        assert!(ReplicationStats::default().is_empty());
+        assert!(ReplicationTargetUsage::default().is_empty());
        for (field, set_nonzero) in cases {
-            let mut stats = ReplicationStats::default();
+            let mut stats = ReplicationTargetUsage::default();
            set_nonzero(&mut stats);
            assert!(!stats.is_empty(), "{field} must make replication stats non-empty");
        }
@@ -2514,17 +2517,17 @@ mod tests {
        }

        let empty_targets = ReplicationAllStats {
-            targets: HashMap::from([("arn:test:empty".to_string(), ReplicationStats::default())]),
+            targets: HashMap::from([("arn:test:empty".to_string(), ReplicationTargetUsage::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:empty".to_string(), ReplicationTargetUsage::default()),
                (
                    "arn:test:non-empty".to_string(),
-                    ReplicationStats {
+                    ReplicationTargetUsage {
                        pending_count: 1,
                        ..Default::default()
                    },
@@ -2565,7 +2568,7 @@ mod tests {
                replication_stats: Some(ReplicationAllStats {
                    targets: HashMap::from([(
                        "arn:test:pending".to_string(),
-                        ReplicationStats {
+                        ReplicationTargetUsage {
                            pending_count: 1,
                            ..Default::default()
                        },
@@ -2714,7 +2717,7 @@ mod tests {
                targets: HashMap::from([
                    (
                        "arn:self-only".to_string(),
-                        ReplicationStats {
+                        ReplicationTargetUsage {
                            pending_size: 7,
                            pending_count: 1,
                            ..Default::default()
@@ -2722,7 +2725,7 @@ mod tests {
                    ),
                    (
                        "arn:shared".to_string(),
-                        ReplicationStats {
+                        ReplicationTargetUsage {
                            failed_size: 3,
                            failed_count: 1,
                            missed_threshold_size: 2,
@@ -2741,7 +2744,7 @@ mod tests {
                targets: HashMap::from([
                    (
                        "arn:shared".to_string(),
-                        ReplicationStats {
+                        ReplicationTargetUsage {
                            failed_size: 5,
                            failed_count: 2,
                            after_threshold_size: 4,
@@ -2751,7 +2754,7 @@ mod tests {
                    ),
                    (
                        "arn:other-only".to_string(),
-                        ReplicationStats {
+                        ReplicationTargetUsage {
                            replicated_size: 11,
                            replicated_count: 3,
                            ..Default::default()
@@ -2993,7 +2996,9 @@ mod tests {
    fn replication_target_deserialization_preserves_large_historical_maps() {
        let mut stats = ReplicationAllStats::default();
        for index in 0..=1024 {
-            stats.targets.insert(format!("target-{index}"), ReplicationStats::default());
+            stats
+                .targets
+                .insert(format!("target-{index}"), ReplicationTargetUsage::default());
        }
        let encoded = rmp_serde::to_vec_named(&stats).expect("large replication target fixture should encode");
        let decoded = rmp_serde::from_slice::<ReplicationAllStats>(&encoded)
@@ -3002,6 +3007,47 @@ mod tests {
        assert_eq!(decoded.targets.len(), stats.targets.len());
    }

+    /// Round-trip test: encoding a [`ReplicationTargetUsage`] and decoding it back
+    /// must produce the exact same value.  This guards against accidental serde
+    /// field-name drift during the `ReplicationStats` -> `ReplicationTargetUsage`
+    /// rename.  Wire-level field names are the serialized Rust field identifiers,
+    /// which must remain byte-identical.
+    #[test]
+    fn replication_target_usage_rmp_round_trip() {
+        let original = ReplicationTargetUsage {
+            pending_size: 100,
+            replicated_size: 2_000,
+            failed_size: 50,
+            failed_count: 3,
+            pending_count: 7,
+            missed_threshold_size: 11,
+            after_threshold_size: 22,
+            missed_threshold_count: 1,
+            after_threshold_count: 2,
+            replicated_count: 99,
+        };
+
+        let buf = rmp_serde::to_vec_named(&original).expect("encode ReplicationTargetUsage to msgpack");
+        let decoded: ReplicationTargetUsage = rmp_serde::from_slice(&buf).expect("decode ReplicationTargetUsage from msgpack");
+        assert_eq!(original, decoded, "round-trip through rmp must preserve every field");
+
+        // Also verify that encoding as an unnamed sequence and then decoding
+        // with named fields produces the correct mapping (this catches reordering).
+        let named_buf = rmp_serde::to_vec_named(&original).expect("re-encode for field-name pinning");
+        // Spot-check that known field names appear in the named encoding.
+        let named_str = String::from_utf8_lossy(&named_buf);
+        assert!(named_str.contains("pending_size"), "field 'pending_size' must survive the rename");
+        assert!(named_str.contains("replicated_size"), "field 'replicated_size' must survive the rename");
+        assert!(
+            named_str.contains("missed_threshold_size"),
+            "field 'missed_threshold_size' must survive the rename"
+        );
+        assert!(
+            named_str.contains("after_threshold_count"),
+            "field 'after_threshold_count' must survive the rename"
+        );
+    }
+
    #[test]
    fn checked_merge_rejects_noncanonical_histograms_without_mutation() {
        let mut entry = DataUsageEntry {
@@ -380,10 +380,24 @@ mod tests {
        cluster.start_node(1).await?;

        let status_url = format!("{}/rustfs/admin/v3/background-heal/status", cluster.nodes[0].url);
-        let status_body = signed_admin_post(&status_url, None, &cluster.access_key, &cluster.secret_key).await?;
-        assert!(
-            !status_body.contains("MissingContentLength"),
-            "background heal status should not fail without an explicit Content-Length: {status_body}"
+        let mut recovered = serde_json::Value::Null;
+        for _ in 0..60 {
+            let status_body = signed_admin_post(&status_url, None, &cluster.access_key, &cluster.secret_key).await?;
+            assert!(
+                !status_body.contains("MissingContentLength"),
+                "background heal status should not fail without an explicit Content-Length: {status_body}"
+            );
+            recovered = serde_json::from_str(&status_body)
+                .map_err(|err| format!("background heal status is not JSON ({err}): {status_body}"))?;
+            if recovered["clusterStatusComplete"] == serde_json::Value::Bool(true) {
+                break;
+            }
+            sleep(Duration::from_secs(1)).await;
+        }
+        assert_eq!(
+            recovered["clusterStatusComplete"],
+            serde_json::Value::Bool(true),
+            "cluster heal status should recover before root heal starts: {recovered}"
        );

        let heal_body = r#"{"recursive":true,"dryRun":false,"remove":false,"recreate":true,"scanMode":2,"updateParity":false,"nolock":false}"#;
@@ -37,7 +37,7 @@ async fn test_bucket_default_sse_s3_put_object() -> Result<(), Box<dyn std::erro

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -159,7 +159,7 @@ async fn test_bucket_default_sse_kms_put_object() -> Result<(), Box<dyn std::err

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -278,7 +278,7 @@ async fn test_bucket_default_sse_kms_multipart_crc32() -> Result<(), Box<dyn std

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -475,7 +475,7 @@ async fn test_explicit_encryption_overrides_bucket_default() -> Result<(), Box<d

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -570,7 +570,7 @@ async fn test_sse_kms_without_key_id_populates_default() -> Result<(), Box<dyn s

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -189,34 +189,121 @@ pub async fn wait_for_kms_ready(
    access_key: &str,
    secret_key: &str,
 ) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
-    let total_deadline = Duration::from_secs(5);
+    wait_for_kms_ready_with_timeout(base_url, access_key, secret_key, Duration::from_secs(5)).await
+}
+
+async fn wait_for_kms_ready_with_timeout(
+    base_url: &str,
+    access_key: &str,
+    secret_key: &str,
+    total_deadline: Duration,
+) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
    let start = tokio::time::Instant::now();
+    let deadline = start + total_deadline;
    let mut backoff = Duration::from_millis(200);
    let max_backoff = Duration::from_secs(1);
-    let mut first_attempt = true;

    loop {
-        if !first_attempt {
-            if start.elapsed() >= total_deadline {
-                return Err("KMS failed to become ready within 5 seconds".into());
-            }
-            sleep(backoff).await;
-            backoff = (backoff * 2).min(max_backoff);
-        }
-        first_attempt = false;
-
-        match get_kms_status(base_url, access_key, secret_key).await {
-            Ok(status) => {
-                info!("KMS is ready (status: {})", status);
-                return Ok(());
-            }
-            Err(e) => {
-                if start.elapsed() >= total_deadline {
-                    return Err(format!("KMS did not become ready within 5 s: last error: {e}").into());
+        match tokio::time::timeout_at(deadline, get_kms_status(base_url, access_key, secret_key)).await {
+            Ok(Ok(status)) => {
+                let backend_status = serde_json::from_str::<serde_json::Value>(&status)
+                    .ok()
+                    .and_then(|value| value.get("backend_status")?.as_str().map(str::to_owned));
+                if backend_status.as_deref() == Some("healthy") {
+                    info!("KMS is ready (status: {})", status);
+                    return Ok(());
                }
-                warn!(error = %e, elapsed_ms = start.elapsed().as_millis() as u64, "KMS not ready yet, retrying…");
+                warn!(
+                    backend_status = backend_status.as_deref().unwrap_or("missing"),
+                    elapsed_ms = u64::try_from(start.elapsed().as_millis()).unwrap_or(u64::MAX),
+                    "KMS not ready yet, retrying…"
+                );
            }
+            Ok(Err(e)) => {
+                let elapsed_ms = u64::try_from(start.elapsed().as_millis()).unwrap_or(u64::MAX);
+                warn!(error = %e, elapsed_ms, "KMS not ready yet, retrying…");
+            }
+            Err(_) => return Err(format!("KMS failed to become ready within {} ms", total_deadline.as_millis()).into()),
        }
+
+        let now = tokio::time::Instant::now();
+        if now >= deadline {
+            return Err(format!("KMS failed to become ready within {} ms", total_deadline.as_millis()).into());
+        }
+        sleep((now + backoff).min(deadline) - now).await;
+        backoff = (backoff * 2).min(max_backoff);
+    }
+}
+
+#[cfg(test)]
+mod readiness_tests {
+    use super::{wait_for_kms_ready, wait_for_kms_ready_with_timeout};
+    use std::sync::{
+        Arc,
+        atomic::{AtomicUsize, Ordering},
+    };
+    use std::time::Duration;
+    use tokio::io::{AsyncReadExt, AsyncWriteExt};
+    use tokio::net::TcpListener;
+
+    #[tokio::test]
+    async fn kms_readiness_retries_http_success_until_backend_is_healthy() {
+        let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind readiness test server");
+        let address = listener.local_addr().expect("read readiness test server address");
+        let requests = Arc::new(AtomicUsize::new(0));
+        let server_requests = Arc::clone(&requests);
+        let server = tokio::spawn(async move {
+            for backend_status in ["error", "healthy"] {
+                let (mut socket, _) = listener.accept().await.expect("accept readiness request");
+                let mut request = Vec::new();
+                let mut chunk = [0_u8; 1024];
+                while !request.windows(4).any(|window| window == b"\r\n\r\n") {
+                    let read = socket.read(&mut chunk).await.expect("read readiness request");
+                    if read == 0 {
+                        break;
+                    }
+                    request.extend_from_slice(&chunk[..read]);
+                }
+                server_requests.fetch_add(1, Ordering::SeqCst);
+
+                let body = format!(r#"{{"backend_status":"{backend_status}"}}"#);
+                let response = format!(
+                    "HTTP/1.1 200 OK\r\nContent-Type: application/json\r\nContent-Length: {}\r\nConnection: close\r\n\r\n{body}",
+                    body.len()
+                );
+                socket.write_all(response.as_bytes()).await.expect("write readiness response");
+            }
+        });
+
+        wait_for_kms_ready(&format!("http://{address}"), "access-key", "secret-key")
+            .await
+            .expect("KMS should become ready after the healthy response");
+
+        let observed_requests = requests.load(Ordering::SeqCst);
+        server.abort();
+        assert_eq!(observed_requests, 2, "an HTTP 200 unhealthy status must be retried");
+    }
+
+    #[tokio::test]
+    async fn kms_readiness_deadline_covers_a_stalled_status_request() {
+        let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind readiness test server");
+        let address = listener.local_addr().expect("read readiness test server address");
+        let server = tokio::spawn(async move {
+            let (mut socket, _) = listener.accept().await.expect("accept readiness request");
+            let mut request = [0_u8; 1024];
+            let _ = socket.read(&mut request).await.expect("read readiness request");
+            std::future::pending::<()>().await;
+        });
+
+        let result = tokio::time::timeout(
+            Duration::from_secs(1),
+            wait_for_kms_ready_with_timeout(&format!("http://{address}"), "access-key", "secret-key", Duration::from_millis(50)),
+        )
+        .await
+        .expect("readiness helper must enforce its own deadline");
+
+        server.abort();
+        assert!(result.is_err(), "a stalled status request must not outlive the readiness deadline");
    }
 }

@@ -61,7 +61,7 @@ async fn test_metadata_replace_self_copy_of_sse_object_stays_decryptable() {
        )
        .await
        .expect("failed to start RustFS with local KMS");
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");

    let client = kms_env.base_env.create_s3_client();
    // Deliberately an UNVERSIONED bucket: that is the branch where the store layer can service
@@ -160,7 +160,7 @@ async fn test_metadata_replace_self_copy_dropping_sse_rewrites_plaintext() {
        )
        .await
        .expect("failed to start RustFS with local KMS");
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");

    let client = kms_env.base_env.create_s3_client();
    // Unversioned, and deliberately WITHOUT a bucket default-encryption rule, so the copy below
@@ -256,7 +256,7 @@ async fn test_metadata_replace_self_copy_under_bucket_default_sse_stays_decrypta
        )
        .await
        .expect("failed to start RustFS with local KMS");
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");

    let client = kms_env.base_env.create_s3_client();
    let bucket = "copy-object-self-copy-bucket-default-sse-test";
@@ -56,7 +56,7 @@ async fn test_self_copy_of_historical_sse_s3_version_is_readable() {
        )
        .await
        .expect("failed to start RustFS with local KMS");
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");

    let client = kms_env.base_env.create_s3_client();
    let bucket = "copy-object-version-restore-sse-test";
@@ -87,7 +87,7 @@ async fn test_head_reports_managed_metadata_for_sse_s3() -> Result<(), Box<dyn s

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -147,7 +147,7 @@ async fn test_head_reports_managed_metadata_for_sse_kms_and_copy() -> Result<(),

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -250,7 +250,7 @@ async fn test_multipart_upload_writes_encrypted_data() -> Result<(), Box<dyn std

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -24,7 +24,6 @@ use super::common::{
    test_sse_kms_encryption, test_sse_s3_encryption,
 };
 use crate::common::{TEST_BUCKET, init_logging};
-use tokio::time::{Duration, sleep};
 use tracing::info;

 /// Comprehensive test: Full KMS workflow with all encryption types
@@ -35,7 +34,7 @@ async fn test_comprehensive_kms_full_workflow() -> Result<(), Box<dyn std::error

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -103,7 +102,7 @@ async fn test_comprehensive_stress_test() -> Result<(), Box<dyn std::error::Erro

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -137,7 +136,7 @@ async fn test_comprehensive_key_isolation() -> Result<(), Box<dyn std::error::Er

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -208,7 +207,7 @@ async fn test_comprehensive_concurrent_operations() -> Result<(), Box<dyn std::e

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -253,7 +252,7 @@ async fn test_comprehensive_performance_benchmark() -> Result<(), Box<dyn std::e

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -44,7 +44,7 @@ async fn test_kms_zero_byte_file_encryption() -> Result<(), Box<dyn std::error::

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -117,7 +117,7 @@ async fn test_kms_single_byte_file_encryption() -> Result<(), Box<dyn std::error

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -209,7 +209,7 @@ async fn test_kms_multipart_boundary_conditions() -> Result<(), Box<dyn std::err

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -284,7 +284,7 @@ async fn test_kms_invalid_key_scenarios() -> Result<(), Box<dyn std::error::Erro

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -371,7 +371,7 @@ async fn test_kms_concurrent_encryption() -> Result<(), Box<dyn std::error::Erro

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = Arc::new(kms_env.base_env.create_s3_client());
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -478,7 +478,7 @@ async fn test_kms_key_validation_security() -> Result<(), Box<dyn std::error::Er

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -37,7 +37,7 @@ async fn test_kms_key_directory_unavailable() -> Result<(), Box<dyn std::error::

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -127,7 +127,7 @@ async fn test_kms_corrupted_key_files() -> Result<(), Box<dyn std::error::Error

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -218,7 +218,7 @@ async fn test_kms_multipart_upload_interruption() -> Result<(), Box<dyn std::err

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -401,7 +401,7 @@ async fn test_kms_resource_constraints() -> Result<(), Box<dyn std::error::Error

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -46,7 +46,7 @@ async fn test_local_kms_end_to_end() -> Result<(), Box<dyn std::error::Error + S
        .expect("Failed to start RustFS with Local KMS");

    // Wait a moment for RustFS to fully start up and initialize KMS
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    info!("RustFS started with KMS auto-configuration, default_key_id: {}", default_key_id);

@@ -127,7 +127,7 @@ async fn test_local_kms_key_isolation() {
        .expect("Failed to start RustFS with Local KMS");

    // Wait a moment for RustFS to fully start up and initialize KMS
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");

    info!("RustFS started with KMS auto-configuration, default_key_id: {}", default_key_id);

@@ -227,7 +227,7 @@ async fn test_local_kms_large_file() {
        .expect("Failed to start RustFS with Local KMS");

    // Wait a moment for RustFS to fully start up and initialize KMS
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");

    info!("RustFS started with KMS auto-configuration, default_key_id: {}", default_key_id);

@@ -309,7 +309,7 @@ async fn test_local_kms_multipart_upload() {
        .expect("Failed to start RustFS with Local KMS");

    // Wait for KMS initialization
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await.expect("KMS ready");

    info!("RustFS started with KMS auto-configuration, default_key_id: {}", default_key_id);

@@ -19,7 +19,6 @@
 //! multipart upload behaviour.

 use crate::common::{TEST_BUCKET, init_logging};
-use tokio::time::{Duration, sleep};
 use tracing::{error, info};

 use super::common::{
@@ -45,8 +44,8 @@ impl VaultKmsTestContext {

        start_kms(&env.base_env.url, &env.base_env.access_key, &env.base_env.secret_key).await?;

-        // Allow Vault to finish initialising token auth and transit engine.
-        sleep(Duration::from_secs(2)).await;
+        // Wait for KMS to finish initialising.
+        super::common::wait_for_kms_ready(&env.base_env.url, &env.base_env.access_key, &env.base_env.secret_key).await?;

        Ok(Self { env })
    }
@@ -39,9 +39,6 @@ mod kms_edge_cases_test;
 #[cfg(test)]
 mod kms_fault_recovery_test;

-#[cfg(test)]
-mod test_runner;
-
 #[cfg(test)]
 mod bucket_default_encryption_test;

@@ -33,7 +33,7 @@ async fn test_step1_basic_single_file_encryption() -> Result<(), Box<dyn std::er

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -89,7 +89,7 @@ async fn test_step2_basic_multipart_upload_without_encryption() -> Result<(), Bo

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -187,7 +187,7 @@ async fn test_step3_multipart_upload_with_sse_s3() -> Result<(), Box<dyn std::er

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -310,7 +310,7 @@ async fn test_step4_large_multipart_upload_with_encryption() -> Result<(), Box<d

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -435,7 +435,7 @@ async fn test_step5_all_encryption_types_multipart() -> Result<(), Box<dyn std::

    let mut kms_env = LocalKMSTestEnvironment::new().await?;
    let _default_key_id = kms_env.start_rustfs_for_local_kms().await?;
-    tokio::time::sleep(tokio::time::Duration::from_secs(3)).await;
+    kms_env.wait_for_kms_ready().await?;

    let s3_client = kms_env.base_env.create_s3_client();
    kms_env.base_env.create_test_bucket(TEST_BUCKET).await?;
@@ -1,499 +0,0 @@
-// Copyright 2024 RustFS Team
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-#![allow(dead_code)]
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-//! Unified KMS test suite runner
-//!
-//! This module provides a unified interface for running KMS tests with categorization,
-//! filtering, and comprehensive reporting capabilities.
-
-use crate::common::init_logging;
-use std::time::Instant;
-use tokio::time::{Duration, sleep};
-use tracing::{debug, error, info, warn};
-
-/// Test category for organization and filtering
-#[derive(Debug, Clone, PartialEq, Eq, Hash)]
-pub enum TestCategory {
-    CoreFunctionality,
-    MultipartEncryption,
-    EdgeCases,
-    FaultRecovery,
-    Comprehensive,
-    Performance,
-}
-
-impl TestCategory {
-    pub fn as_str(&self) -> &'static str {
-        match self {
-            TestCategory::CoreFunctionality => "core-functionality",
-            TestCategory::MultipartEncryption => "multipart-encryption",
-            TestCategory::EdgeCases => "edge-cases",
-            TestCategory::FaultRecovery => "fault-recovery",
-            TestCategory::Comprehensive => "comprehensive",
-            TestCategory::Performance => "performance",
-        }
-    }
-}
-
-/// Test definition with metadata
-#[derive(Debug, Clone)]
-pub struct TestDefinition {
-    pub name: String,
-    pub description: String,
-    pub category: TestCategory,
-    pub estimated_duration: Duration,
-    pub is_critical: bool,
-}
-
-impl TestDefinition {
-    pub fn new(
-        name: impl Into<String>,
-        description: impl Into<String>,
-        category: TestCategory,
-        estimated_duration: Duration,
-        is_critical: bool,
-    ) -> Self {
-        Self {
-            name: name.into(),
-            description: description.into(),
-            category,
-            estimated_duration,
-            is_critical,
-        }
-    }
-}
-
-/// Test execution result
-#[derive(Debug, Clone)]
-pub struct TestResult {
-    pub test_name: String,
-    pub category: TestCategory,
-    pub success: bool,
-    pub duration: Duration,
-    pub error_message: Option<String>,
-}
-
-impl TestResult {
-    pub fn success(test_name: String, category: TestCategory, duration: Duration) -> Self {
-        Self {
-            test_name,
-            category,
-            success: true,
-            duration,
-            error_message: None,
-        }
-    }
-
-    pub fn failure(test_name: String, category: TestCategory, duration: Duration, error: String) -> Self {
-        Self {
-            test_name,
-            category,
-            success: false,
-            duration,
-            error_message: Some(error),
-        }
-    }
-}
-
-/// Comprehensive test suite configuration
-#[derive(Debug, Clone)]
-pub struct TestSuiteConfig {
-    pub categories: Vec<TestCategory>,
-    pub include_critical_only: bool,
-    pub max_duration: Option<Duration>,
-    pub parallel_execution: bool,
-}
-
-impl Default for TestSuiteConfig {
-    fn default() -> Self {
-        Self {
-            categories: vec![
-                TestCategory::CoreFunctionality,
-                TestCategory::MultipartEncryption,
-                TestCategory::EdgeCases,
-                TestCategory::FaultRecovery,
-                TestCategory::Comprehensive,
-            ],
-            include_critical_only: false,
-            max_duration: None,
-            parallel_execution: false,
-        }
-    }
-}
-
-/// Unified KMS test suite runner
-pub struct KMSTestSuite {
-    tests: Vec<TestDefinition>,
-    config: TestSuiteConfig,
-}
-
-impl KMSTestSuite {
-    /// Create a new test suite with default configuration
-    pub fn new() -> Self {
-        let tests = vec![
-            // Core Functionality Tests
-            TestDefinition::new(
-                "test_local_kms_end_to_end",
-                "End-to-end KMS test with all encryption types",
-                TestCategory::CoreFunctionality,
-                Duration::from_secs(60),
-                true,
-            ),
-            TestDefinition::new(
-                "test_local_kms_key_isolation",
-                "Test KMS key isolation and security",
-                TestCategory::CoreFunctionality,
-                Duration::from_secs(45),
-                true,
-            ),
-            // Multipart Encryption Tests
-            TestDefinition::new(
-                "test_local_kms_multipart_upload",
-                "Test large file multipart upload with encryption",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(120),
-                true,
-            ),
-            TestDefinition::new(
-                "test_step1_basic_single_file_encryption",
-                "Basic single file encryption test",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(30),
-                false,
-            ),
-            TestDefinition::new(
-                "test_step2_basic_multipart_upload_without_encryption",
-                "Basic multipart upload without encryption",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(45),
-                false,
-            ),
-            TestDefinition::new(
-                "test_step3_multipart_upload_with_sse_s3",
-                "Multipart upload with SSE-S3 encryption",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(60),
-                true,
-            ),
-            TestDefinition::new(
-                "test_step4_large_multipart_upload_with_encryption",
-                "Large file multipart upload with encryption",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(90),
-                false,
-            ),
-            TestDefinition::new(
-                "test_step5_all_encryption_types_multipart",
-                "All encryption types multipart test",
-                TestCategory::MultipartEncryption,
-                Duration::from_secs(120),
-                true,
-            ),
-            // Edge Cases Tests
-            TestDefinition::new(
-                "test_kms_zero_byte_file_encryption",
-                "Test encryption of zero-byte files",
-                TestCategory::EdgeCases,
-                Duration::from_secs(20),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_single_byte_file_encryption",
-                "Test encryption of single-byte files",
-                TestCategory::EdgeCases,
-                Duration::from_secs(20),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_multipart_boundary_conditions",
-                "Test multipart upload boundary conditions",
-                TestCategory::EdgeCases,
-                Duration::from_secs(45),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_invalid_key_scenarios",
-                "Test invalid key scenarios",
-                TestCategory::EdgeCases,
-                Duration::from_secs(30),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_concurrent_encryption",
-                "Test concurrent encryption operations",
-                TestCategory::EdgeCases,
-                Duration::from_secs(60),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_key_validation_security",
-                "Test key validation security",
-                TestCategory::EdgeCases,
-                Duration::from_secs(30),
-                false,
-            ),
-            // Fault Recovery Tests
-            TestDefinition::new(
-                "test_kms_key_directory_unavailable",
-                "Test KMS when key directory is unavailable",
-                TestCategory::FaultRecovery,
-                Duration::from_secs(45),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_corrupted_key_files",
-                "Test KMS with corrupted key files",
-                TestCategory::FaultRecovery,
-                Duration::from_secs(30),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_multipart_upload_interruption",
-                "Test multipart upload interruption recovery",
-                TestCategory::FaultRecovery,
-                Duration::from_secs(60),
-                false,
-            ),
-            TestDefinition::new(
-                "test_kms_resource_constraints",
-                "Test KMS under resource constraints",
-                TestCategory::FaultRecovery,
-                Duration::from_secs(90),
-                false,
-            ),
-            // Comprehensive Tests
-            TestDefinition::new(
-                "test_comprehensive_kms_full_workflow",
-                "Full KMS workflow comprehensive test",
-                TestCategory::Comprehensive,
-                Duration::from_secs(300),
-                true,
-            ),
-            TestDefinition::new(
-                "test_comprehensive_stress_test",
-                "KMS stress test with large datasets",
-                TestCategory::Comprehensive,
-                Duration::from_secs(400),
-                false,
-            ),
-            TestDefinition::new(
-                "test_comprehensive_key_isolation",
-                "Comprehensive key isolation test",
-                TestCategory::Comprehensive,
-                Duration::from_secs(180),
-                false,
-            ),
-            TestDefinition::new(
-                "test_comprehensive_concurrent_operations",
-                "Comprehensive concurrent operations test",
-                TestCategory::Comprehensive,
-                Duration::from_secs(240),
-                false,
-            ),
-            TestDefinition::new(
-                "test_comprehensive_performance_benchmark",
-                "KMS performance benchmark test",
-                TestCategory::Comprehensive,
-                Duration::from_secs(360),
-                false,
-            ),
-        ];
-
-        Self {
-            tests,
-            config: TestSuiteConfig::default(),
-        }
-    }
-
-    /// Configure the test suite
-    pub fn with_config(mut self, config: TestSuiteConfig) -> Self {
-        self.config = config;
-        self
-    }
-
-    /// Filter tests based on category
-    pub fn filter_by_category(&self, category: &TestCategory) -> Vec<&TestDefinition> {
-        self.tests.iter().filter(|test| &test.category == category).collect()
-    }
-
-    /// Filter tests based on criticality
-    pub fn filter_critical_tests(&self) -> Vec<&TestDefinition> {
-        self.tests.iter().filter(|test| test.is_critical).collect()
-    }
-
-    /// Get test summary by category
-    pub fn get_category_summary(&self) -> std::collections::HashMap<TestCategory, Vec<&TestDefinition>> {
-        let mut summary = std::collections::HashMap::new();
-        for test in &self.tests {
-            summary.entry(test.category.clone()).or_insert_with(Vec::new).push(test);
-        }
-        summary
-    }
-
-    /// Run the complete test suite
-    pub async fn run_test_suite(&self) -> Vec<TestResult> {
-        init_logging();
-        info!("🚀 Starting unified KMS test suite");
-
-        let start_time = Instant::now();
-        let mut results = Vec::new();
-
-        // Filter tests based on configuration
-        let tests_to_run: Vec<&TestDefinition> = self
-            .tests
-            .iter()
-            .filter(|test| self.config.categories.contains(&test.category))
-            .filter(|test| !self.config.include_critical_only || test.is_critical)
-            .collect();
-
-        info!("📊 Test plan: {} test(s) scheduled", tests_to_run.len());
-        for (i, test) in tests_to_run.iter().enumerate() {
-            info!("  {}. {} ({})", i + 1, test.name, test.category.as_str());
-        }
-
-        // Execute tests
-        for (i, test_def) in tests_to_run.iter().enumerate() {
-            info!("🧪 Running test {}/{}: {}", i + 1, tests_to_run.len(), test_def.name);
-            info!("   📝 Description: {}", test_def.description);
-            info!("   🏷️  Category: {}", test_def.category.as_str());
-            info!("   ⏱️  Estimated duration: {:?}", test_def.estimated_duration);
-
-            let test_start = Instant::now();
-            let result = self.run_single_test(test_def).await;
-            let test_duration = test_start.elapsed();
-
-            match result {
-                Ok(_) => {
-                    info!("✅ Test passed: {} ({:.2}s)", test_def.name, test_duration.as_secs_f64());
-                    results.push(TestResult::success(test_def.name.clone(), test_def.category.clone(), test_duration));
-                }
-                Err(e) => {
-                    error!("❌ Test failed: {} ({:.2}s): {}", test_def.name, test_duration.as_secs_f64(), e);
-                    results.push(TestResult::failure(
-                        test_def.name.clone(),
-                        test_def.category.clone(),
-                        test_duration,
-                        e.to_string(),
-                    ));
-                }
-            }
-
-            // Add delay between tests to avoid resource conflicts
-            if i < tests_to_run.len() - 1 {
-                debug!("⏸️  Waiting two seconds before the next test...");
-                sleep(Duration::from_secs(2)).await;
-            }
-        }
-
-        let total_duration = start_time.elapsed();
-        self.print_test_summary(&results, total_duration);
-
-        results
-    }
-
-    /// Run a single test by dispatching to the appropriate test function
-    async fn run_single_test(&self, test_def: &TestDefinition) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
-        // This is a placeholder for test dispatch logic
-        // In a real implementation, this would dispatch to actual test functions
-        warn!("⚠️  Test '{}' is not implemented in the unified runner; skipping", test_def.name);
-        Ok(())
-    }
-
-    /// Print comprehensive test summary
-    fn print_test_summary(&self, results: &[TestResult], total_duration: Duration) {
-        info!("📊 KMS test suite summary");
-        info!("⏱️  Total duration: {:.2} seconds", total_duration.as_secs_f64());
-        info!("📈 Total tests: {}", results.len());
-
-        let passed = results.iter().filter(|r| r.success).count();
-        let failed = results.iter().filter(|r| !r.success).count();
-
-        info!("✅ Passed: {}", passed);
-        info!("❌ Failed: {}", failed);
-        info!("📊 Success rate: {:.1}%", (passed as f64 / results.len() as f64) * 100.0);
-
-        // Summary by category
-        let mut category_summary: std::collections::HashMap<TestCategory, (usize, usize)> = std::collections::HashMap::new();
-        for result in results {
-            let (total, passed_count) = category_summary.entry(result.category.clone()).or_insert((0, 0));
-            *total += 1;
-            if result.success {
-                *passed_count += 1;
-            }
-        }
-
-        info!("📊 Category summary:");
-        for (category, (total, passed_count)) in category_summary {
-            info!(
-                "  🏷️  {}: {}/{} ({:.1}%)",
-                category.as_str(),
-                passed_count,
-                total,
-                (passed_count as f64 / total as f64) * 100.0
-            );
-        }
-
-        // List failed tests
-        if failed > 0 {
-            warn!("❌ Failing tests:");
-            for result in results.iter().filter(|r| !r.success) {
-                warn!("  - {}: {}", result.test_name, result.error_message.as_deref().unwrap_or("Unknown error"));
-            }
-        }
-    }
-}
-
-/// Quick test suite for critical tests only
-#[tokio::test]
-async fn test_kms_critical_suite() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
-    let config = TestSuiteConfig {
-        categories: vec![TestCategory::CoreFunctionality, TestCategory::MultipartEncryption],
-        include_critical_only: true,
-        max_duration: Some(Duration::from_secs(600)), // 10 minutes max
-        parallel_execution: false,
-    };
-
-    let suite = KMSTestSuite::new().with_config(config);
-    let results = suite.run_test_suite().await;
-
-    let failed_count = results.iter().filter(|r| !r.success).count();
-    if failed_count > 0 {
-        return Err(format!("Critical test suite failed: {failed_count} tests failed").into());
-    }
-
-    info!("✅ All critical tests passed");
-    Ok(())
-}
-
-/// Full comprehensive test suite
-#[tokio::test]
-async fn test_kms_full_suite() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
-    let suite = KMSTestSuite::new();
-    let results = suite.run_test_suite().await;
-
-    let total_tests = results.len();
-    let failed_count = results.iter().filter(|r| !r.success).count();
-    let success_rate = ((total_tests - failed_count) as f64 / total_tests as f64) * 100.0;
-
-    info!("📊 Full suite success rate: {:.1}%", success_rate);
-
-    // Allow up to 10% failure rate for non-critical tests
-    if success_rate < 90.0 {
-        return Err(format!("Test suite success rate too low: {success_rate:.1}%").into());
-    }
-
-    info!("✅ Full test suite succeeded");
-    Ok(())
-}
@@ -16,6 +16,7 @@ use crate::common::{RustFSTestClusterEnvironment, RustFSTestEnvironment, init_lo
 use aws_sdk_s3::primitives::ByteStream;
 use http::header::{CONTENT_TYPE, HOST};
 use reqwest::StatusCode;
+use rustfs_config::{ENV_DRIVE_ACTIVE_CHECK_INTERVAL_SECS, ENV_NOTIFY_ENABLE};
 use rustfs_signer::pre_sign_v4;
 use rustfs_utils::egress::ENV_OUTBOUND_ALLOW_ORIGINS;
 use s3s::Body;
@@ -976,7 +977,8 @@ async fn test_get_object_lambda_rejects_disabled_target() -> Result<(), Box<dyn
    init_logging();

    let mut env = RustFSTestEnvironment::new().await?;
-    env.start_rustfs_server(vec![]).await?;
+    env.start_rustfs_server_with_env(vec![], &[(ENV_NOTIFY_ENABLE, "true")])
+        .await?;

    let bucket = "object-lambda-e2e-disabled-target";
    let key = "input.txt";
@@ -992,17 +994,24 @@ async fn test_get_object_lambda_rejects_disabled_target() -> Result<(), Box<dyn
        .send()
        .await?;

-    configure_webhook_target_with_key_values(
-        &env,
-        "transformer",
-        vec![
-            ("endpoint", "http://127.0.0.1:9/transform".to_string()),
-            ("auth_token", "secret-token".to_string()),
-            ("enable", "off".to_string()),
-        ],
+    let queue_dir = format!("{}/disabled-target-queue", env.temp_dir);
+    tokio::fs::create_dir_all(&queue_dir).await?;
+    let config_url = format!("{}/rustfs/admin/v3/set-config-kv", env.url);
+    let directive = format!(
+        "notify_webhook:transformer enable=off endpoint=\"http://127.0.0.1:9/transform\" auth_token=\"secret-token\" queue_dir=\"{queue_dir}\""
+    );
+    let disable_response = signed_request(
+        http::Method::PUT,
+        &config_url,
+        &env.access_key,
+        &env.secret_key,
+        Some(directive.into_bytes()),
+        Some("text/plain"),
    )
    .await?;
-    wait_for_target_visibility(&env, "transformer").await?;
+    let disable_status = disable_response.status();
+    let disable_body = disable_response.text().await?;
+    assert_eq!(disable_status, StatusCode::OK, "failed to disable target: {disable_body}");

    let lambda_url = format!("{}/{}/{}?lambdaArn={}", env.url, bucket, key, urlencoding::encode(lambda_arn));
    let response = signed_request(http::Method::GET, &lambda_url, &env.access_key, &env.secret_key, None, None).await?;
@@ -1021,7 +1030,8 @@ async fn test_configure_object_lambda_target_rejects_invalid_endpoint() -> Resul
    init_logging();

    let mut env = RustFSTestEnvironment::new().await?;
-    env.start_rustfs_server(vec![]).await?;
+    env.start_rustfs_server_with_env(vec![], &[(ENV_NOTIFY_ENABLE, "true")])
+        .await?;

    let bucket = "object-lambda-e2e-invalid-endpoint";

@@ -1064,7 +1074,8 @@ async fn test_configure_object_lambda_notify_webhook_rejects_response_header_tim
    init_logging();

    let mut env = RustFSTestEnvironment::new().await?;
-    env.start_rustfs_server(vec![]).await?;
+    env.start_rustfs_server_with_env(vec![], &[(ENV_NOTIFY_ENABLE, "true")])
+        .await?;

    let response = send_configure_webhook_target_request(
        &env,
@@ -1173,6 +1184,8 @@ async fn test_listen_notification_fans_in_remote_node_events() -> Result<(), Box
    init_logging();

    let mut cluster = RustFSTestClusterEnvironment::new(2).await?;
+    cluster.set_env(ENV_NOTIFY_ENABLE, "true");
+    cluster.set_env(ENV_DRIVE_ACTIVE_CHECK_INTERVAL_SECS, "1");
    cluster.start().await?;

    let bucket = "listen-notification-cluster";
@@ -15,7 +15,6 @@
 use crate::common::{RustFSTestClusterEnvironment, init_logging};
 use aws_sdk_s3::error::SdkError;
 use aws_sdk_s3::primitives::ByteStream;
-use aws_sdk_s3::types::CompletedMultipartUpload;
 use tokio::time::{Duration, sleep};
 use tracing::info;
 use uuid::Uuid;
@@ -43,32 +42,18 @@ async fn list_parts_reports_missing_upload(
    }
 }

-async fn complete_reports_missing_upload(
+async fn multipart_listing_reports_missing_upload(
    client: &aws_sdk_s3::Client,
    bucket: &str,
    key: &str,
    upload_id: &str,
 ) -> Result<bool, Box<dyn std::error::Error + Send + Sync>> {
-    let result = client
-        .complete_multipart_upload()
-        .bucket(bucket)
-        .key(key)
-        .upload_id(upload_id)
-        .multipart_upload(CompletedMultipartUpload::builder().build())
-        .send()
-        .await;
-    match result {
-        Ok(_) => Ok(false),
-        Err(SdkError::ServiceError(err)) => {
-            let code = err.err().meta().code().unwrap_or("");
-            if code == "NoSuchUpload" {
-                Ok(true)
-            } else {
-                Err(format!("unexpected complete_multipart_upload service error: code={code}, err={err:?}").into())
-            }
-        }
-        Err(err) => Err(format!("unexpected complete_multipart_upload error: {err:?}").into()),
-    }
+    let result = client.list_multipart_uploads().bucket(bucket).prefix(key).send().await?;
+
+    Ok(!result
+        .uploads()
+        .iter()
+        .any(|upload| upload.key() == Some(key) && upload.upload_id() == Some(upload_id)))
 }

 async fn wait_for_cleanup_on_all_nodes(
@@ -81,8 +66,8 @@ async fn wait_for_cleanup_on_all_nodes(
        let mut all_cleaned = true;
        for (idx, client) in clients.iter().enumerate() {
            let list_parts_missing = list_parts_reports_missing_upload(client, bucket, key, upload_id).await?;
-            let complete_missing = complete_reports_missing_upload(client, bucket, key, upload_id).await?;
-            if !(list_parts_missing && complete_missing) {
+            let listing_missing = multipart_listing_reports_missing_upload(client, bucket, key, upload_id).await?;
+            if !(list_parts_missing && listing_missing) {
                info!("stale multipart still visible on node {} at attempt {}", idx, attempt + 1);
                all_cleaned = false;
                break;
@@ -146,6 +131,10 @@ async fn test_stale_multipart_cleanup_removes_incomplete_upload_across_cluster()
        1,
        "multipart upload should be visible before background cleanup"
    );
+    assert!(
+        !multipart_listing_reports_missing_upload(&clients[2], CLEANUP_BUCKET, &key, &upload_id).await?,
+        "multipart upload listing should contain the upload before background cleanup"
+    );

    wait_for_cleanup_on_all_nodes(&clients, CLEANUP_BUCKET, &key, &upload_id).await?;

@@ -1089,7 +1089,9 @@ impl PeerRestClient {
                    .await?
                    .max_decoding_message_size(BACKGROUND_HEAL_STATUS_MAX_MESSAGE_SIZE);
                let response = match client
-                    .background_heal_status(Request::new(BackgroundHealStatusRequest::default()))
+                    .background_heal_status(Request::new(BackgroundHealStatusRequest {
+                        protocol_version: rustfs_protos::BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION,
+                    }))
                    .await
                {
                    Ok(response) => response.into_inner(),
@@ -42,8 +42,9 @@ use futures::lock::Mutex;
 use metrics::counter;
 use rustfs_filemeta::{FileInfo, ObjectPartInfo, RawFileInfo};
 use rustfs_io_metrics::internode_metrics::{
-    INTERNODE_STAGE_READ_VERSION_REQUEST_ENCODE, INTERNODE_STAGE_READ_VERSION_RESPONSE_DECODE,
-    INTERNODE_STAGE_READ_VERSION_RPC_ROUNDTRIP,
+    INTERNODE_STAGE_BATCH_READ_VERSION_REQUEST_ENCODE, INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_DECODE,
+    INTERNODE_STAGE_BATCH_READ_VERSION_RPC_ROUNDTRIP, INTERNODE_STAGE_READ_VERSION_REQUEST_ENCODE,
+    INTERNODE_STAGE_READ_VERSION_RESPONSE_DECODE, INTERNODE_STAGE_READ_VERSION_RPC_ROUNDTRIP,
 };
 use rustfs_protos::ChannelClass;
 use rustfs_protos::evict_failed_connection;
@@ -98,6 +99,7 @@ const NS_SCANNER_CAPABILITY_PROBE_TIMEOUT: Duration = Duration::from_secs(5);
 const REMOTE_DISK_READ_RETRY_BASE_BACKOFF: Duration = Duration::from_millis(50);
 const ENV_RUSTFS_METADATA_BATCH_READ: &str = "RUSTFS_METADATA_BATCH_READ";
 const LEGACY_ENV_RUSTFS_BATCH_METADATA_RPC: &str = "RUSTFS_BATCH_METADATA_RPC";
+const ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE: &str = "RUSTFS_GET_METADATA_READ_VERSION_COALESCE";
 const BATCH_METADATA_RPC_OFF: &str = "off";
 const BATCH_METADATA_RPC_AUTO: &str = "auto";
 const BATCH_METADATA_RPC_ON: &str = "on";
@@ -202,7 +204,8 @@ fn parse_batch_metadata_rpc_mode(raw: &str) -> BatchMetadataRpcMode {
 }

 fn batch_metadata_rpc_mode_from_env() -> BatchMetadataRpcMode {
-    rustfs_utils::get_env_opt_str(ENV_RUSTFS_METADATA_BATCH_READ)
+    rustfs_utils::get_env_opt_str(ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE)
+        .or_else(|| rustfs_utils::get_env_opt_str(ENV_RUSTFS_METADATA_BATCH_READ))
        .or_else(|| rustfs_utils::get_env_opt_str(LEGACY_ENV_RUSTFS_BATCH_METADATA_RPC))
        .as_deref()
        .map(parse_batch_metadata_rpc_mode)
@@ -1826,6 +1829,12 @@ fn record_read_version_stage(stage: &'static str, started_at: Option<Instant>) {
    }
 }

+fn record_batch_read_version_stage(stage: &'static str, started_at: Option<Instant>) {
+    if let Some(started_at) = started_at {
+        crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_stage(stage, started_at.elapsed());
+    }
+}
+
 /// Aggregate encoded size (bytes) of a `ReadMultiple` response, preferring the msgpack payloads
 /// and falling back to the JSON compatibility strings. Used to size the RPC for the payload
 /// histogram / large-payload alerting (grpc-optimization P0 instrumentation).
@@ -1936,6 +1945,27 @@ fn decode_batch_read_version_response_items(
    Ok(batch_read_version_resps)
 }

+fn batch_read_version_request_payload_len(req: &BatchReadVersionReq, req_json: &str, req_bin: &[u8]) -> usize {
+    req.items
+        .iter()
+        .fold(req_json.len().saturating_add(req_bin.len()), |total, item| {
+            total
+                .saturating_add(item.org_volume.len())
+                .saturating_add(item.volume.len())
+                .saturating_add(item.path.len())
+                .saturating_add(item.version_id.len())
+        })
+}
+
+fn batch_read_version_response_payload_len(response: &BatchReadVersionResponse) -> usize {
+    response
+        .batch_read_version_resps
+        .iter()
+        .map(String::len)
+        .sum::<usize>()
+        .saturating_add(response.batch_read_version_resps_bin.iter().map(Bytes::len).sum::<usize>())
+}
+
 fn validate_decoded_file_info(file_info: &FileInfo) -> Result<()> {
    file_info.validate_for_metadata_read().map_err(Into::into)
 }
@@ -2837,14 +2867,19 @@ impl DiskAPI for RemoteDisk {
            state = "started",
            "Remote disk RPC started"
        );
+        let batch_read_version_attribution_enabled = rustfs_io_metrics::get_stage_metrics_enabled();
+        let encode_started = read_version_stage_timer(batch_read_version_attribution_enabled);
        let batch_read_version_req = compat_json(&req)?;
        let batch_read_version_req_bin = encode_msgpack(&req)?;
-
+        record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_REQUEST_ENCODE, encode_started);
+        let request_payload_bytes = batch_read_version_attribution_enabled
+            .then(|| batch_read_version_request_payload_len(&req, &batch_read_version_req, &batch_read_version_req_bin));
        let batch_result = self
            .execute_with_timeout_for_op(
                "batch_read_version",
                move || async move {
                    let disk = self.disk_ref().await;
+                    let disk_len = disk.len();
                    let mut client = self
                        .get_bulk_client()
                        .await
@@ -2855,9 +2890,20 @@ impl DiskAPI for RemoteDisk {
                        batch_read_version_req_bin: batch_read_version_req_bin.into(),
                    });

+                    crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_request();
+                    if let Some(request_payload_bytes) = request_payload_bytes {
+                        crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_sent_bytes(
+                            request_payload_bytes.saturating_add(disk_len),
+                        );
+                    }
+                    let rpc_started = read_version_stage_timer(batch_read_version_attribution_enabled);
                    let response = match client.batch_read_version(request).await {
-                        Ok(response) => response.into_inner(),
+                        Ok(response) => {
+                            record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_RPC_ROUNDTRIP, rpc_started);
+                            response.into_inner()
+                        }
                        Err(status) if status.code() == Code::Unimplemented => {
+                            record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_RPC_ROUNDTRIP, rpc_started);
                            if mode.should_fallback_on_unimplemented() {
                                record_batch_read_version_gate_decision(mode, BATCH_READ_VERSION_GATE_FALLBACK_UNIMPLEMENTED);
                                warn!(
@@ -2874,6 +2920,7 @@ impl DiskAPI for RemoteDisk {
                            }

                            record_batch_read_version_gate_decision(mode, BATCH_READ_VERSION_GATE_UNSUPPORTED_NO_FALLBACK);
+                            crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_error();
                            warn!(
                                event = EVENT_REMOTE_DISK_RPC,
                                component = LOG_COMPONENT_ECSTORE,
@@ -2886,14 +2933,33 @@ impl DiskAPI for RemoteDisk {
                            );
                            return Err(Error::from(status));
                        }
-                        Err(status) => return Err(Error::from(status)),
+                        Err(status) => {
+                            record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_RPC_ROUNDTRIP, rpc_started);
+                            crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_error();
+                            return Err(Error::from(status));
+                        }
                    };

                    if !response.success {
+                        crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_error();
                        return Err(response.error.unwrap_or_default().into());
                    }

-                    decode_batch_read_version_response_items(response, &self.endpoint).map(Some)
+                    crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_recv_bytes(
+                        batch_read_version_response_payload_len(&response),
+                    );
+                    let decode_started = read_version_stage_timer(batch_read_version_attribution_enabled);
+                    match decode_batch_read_version_response_items(response, &self.endpoint) {
+                        Ok(batch_read_version_resps) => {
+                            record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_DECODE, decode_started);
+                            Ok(Some(batch_read_version_resps))
+                        }
+                        Err(err) => {
+                            record_batch_read_version_stage(INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_DECODE, decode_started);
+                            crate::cluster::rpc::runtime_sources::record_remote_disk_grpc_batch_read_version_error();
+                            Err(err)
+                        }
+                    }
                },
                get_max_timeout_duration(),
            )
@@ -4621,6 +4687,7 @@ mod tests {
            } else {
                "file version not found".to_string()
            },
+            error_code: if success { 0 } else { DiskError::FileVersionNotFound.to_u32() },
        }
    }

@@ -4740,6 +4807,7 @@ mod tests {
    fn batch_metadata_rpc_mode_uses_documented_env_before_legacy_alias() {
        temp_env::with_vars(
            [
+                (ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE, None::<&str>),
                (ENV_RUSTFS_METADATA_BATCH_READ, Some("auto")),
                (LEGACY_ENV_RUSTFS_BATCH_METADATA_RPC, Some("on")),
            ],
@@ -4749,10 +4817,25 @@ mod tests {
        );
    }

+    #[test]
+    fn batch_metadata_rpc_mode_uses_get_coalescer_env_before_batch_env() {
+        temp_env::with_vars(
+            [
+                (ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE, Some("on")),
+                (ENV_RUSTFS_METADATA_BATCH_READ, Some("off")),
+                (LEGACY_ENV_RUSTFS_BATCH_METADATA_RPC, Some("off")),
+            ],
+            || {
+                assert_eq!(batch_metadata_rpc_mode_from_env(), BatchMetadataRpcMode::On);
+            },
+        );
+    }
+
    #[test]
    fn batch_metadata_rpc_mode_falls_back_to_legacy_env_alias() {
        temp_env::with_vars(
            [
+                (ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE, None::<&str>),
                (ENV_RUSTFS_METADATA_BATCH_READ, None::<&str>),
                (LEGACY_ENV_RUSTFS_BATCH_METADATA_RPC, Some("on")),
            ],
@@ -14,9 +14,10 @@

 use rustfs_io_metrics::internode_metrics::{
    INTERNODE_MSGPACK_CODEC_JSON, INTERNODE_MSGPACK_CODEC_MSGPACK, INTERNODE_MSGPACK_DIRECTION_RESPONSE,
-    INTERNODE_OPERATION_GRPC_READ_ALL, INTERNODE_OPERATION_GRPC_READ_MULTIPLE, INTERNODE_OPERATION_GRPC_READ_VERSION,
-    INTERNODE_OPERATION_GRPC_WRITE_ALL, INTERNODE_OPERATION_PUT_FILE_STREAM, INTERNODE_OPERATION_READ_FILE_STREAM,
-    INTERNODE_TRANSPORT_BACKEND_GRPC, INTERNODE_TRANSPORT_BACKEND_TCP_HTTP, global_internode_metrics,
+    INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION, INTERNODE_OPERATION_GRPC_READ_ALL, INTERNODE_OPERATION_GRPC_READ_MULTIPLE,
+    INTERNODE_OPERATION_GRPC_READ_VERSION, INTERNODE_OPERATION_GRPC_WRITE_ALL, INTERNODE_OPERATION_PUT_FILE_STREAM,
+    INTERNODE_OPERATION_READ_FILE_STREAM, INTERNODE_TRANSPORT_BACKEND_GRPC, INTERNODE_TRANSPORT_BACKEND_TCP_HTTP,
+    global_internode_metrics,
 };
 use std::time::Duration;

@@ -93,6 +94,59 @@ pub(crate) fn record_remote_disk_grpc_read_version_request() {
    );
 }

+pub(crate) fn record_remote_disk_grpc_batch_read_version_request() {
+    if !rustfs_io_metrics::get_stage_metrics_enabled() {
+        return;
+    }
+    global_internode_metrics().record_outgoing_request_for_operation_and_backend(
+        INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION,
+        INTERNODE_TRANSPORT_BACKEND_GRPC,
+    );
+}
+
+pub(crate) fn record_remote_disk_grpc_batch_read_version_stage(stage: &'static str, duration: Duration) {
+    if !rustfs_io_metrics::get_stage_metrics_enabled() {
+        return;
+    }
+    global_internode_metrics().record_stage_duration_for_operation_and_backend(
+        INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION,
+        INTERNODE_TRANSPORT_BACKEND_GRPC,
+        stage,
+        duration,
+    );
+}
+
+pub(crate) fn record_remote_disk_grpc_batch_read_version_error() {
+    if !rustfs_io_metrics::get_stage_metrics_enabled() {
+        return;
+    }
+    global_internode_metrics()
+        .record_error_for_operation_and_backend(INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION, INTERNODE_TRANSPORT_BACKEND_GRPC);
+}
+
+pub(crate) fn record_remote_disk_grpc_batch_read_version_sent_bytes(bytes: usize) {
+    if !rustfs_io_metrics::get_stage_metrics_enabled() {
+        return;
+    }
+    global_internode_metrics().record_sent_bytes_for_operation_and_backend(
+        INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION,
+        INTERNODE_TRANSPORT_BACKEND_GRPC,
+        bytes,
+    );
+}
+
+pub(crate) fn record_remote_disk_grpc_batch_read_version_recv_bytes(bytes: usize) {
+    if !rustfs_io_metrics::get_stage_metrics_enabled() {
+        return;
+    }
+    global_internode_metrics().record_recv_bytes_for_operation_and_backend(
+        INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION,
+        INTERNODE_TRANSPORT_BACKEND_GRPC,
+        bytes,
+    );
+    record_grpc_payload_size(INTERNODE_OPERATION_GRPC_BATCH_READ_VERSION, bytes);
+}
+
 pub(crate) fn record_remote_disk_grpc_read_version_error() {
    if !rustfs_io_metrics::get_stage_metrics_enabled() {
        return;
@@ -248,10 +248,13 @@ impl Config {
        let shard_size = shard_size as usize;
        // Keep the historical two-data-shard object budget while preventing
        // wider EC layouts from multiplying the maximum inline object size.
+        // Use div_ceil to match the shard_file_size calculation (which also uses
+        // div_ceil), avoiding a 1-byte rounding discrepancy that prevents inline
+        // for objects right at the threshold.
        let inline_block = if self.initialized && self.inline_block_explicit {
            self.inline_block
        } else {
-            (DEFAULT_INLINE_OBJECT_BUDGET / data_shards).min(DEFAULT_INLINE_BLOCK)
+            DEFAULT_INLINE_OBJECT_BUDGET.div_ceil(data_shards).min(DEFAULT_INLINE_BLOCK)
        };

        if versioned {
@@ -988,14 +988,11 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for Sets {
    }
 }

-#[async_trait::async_trait]
-impl crate::storage_api_contracts::heal::HealOperations for Sets {
-    type Error = Error;
-    type HealResultItem = HealResultItem;
-    type HealOptions = HealOpts;
-
-    #[tracing::instrument(skip(self))]
-    async fn heal_format(&self, dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
+impl Sets {
+    pub(crate) async fn heal_format_with_fence<F>(&self, dry_run: bool, fence_lost: F) -> Result<(HealResultItem, Option<Error>)>
+    where
+        F: Fn() -> bool + Send + Sync,
+    {
        let (disks, init_errs) = init_storage_disks_with_errors(
            &self.endpoints.endpoints,
            &DiskOption {
@@ -1068,6 +1065,9 @@ impl crate::storage_api_contracts::heal::HealOperations for Sets {
            // Save new formats `format.json` on unformatted disks.
            for (index, (fm, disk)) in tmp_new_formats.iter_mut().zip(disks.iter()).enumerate() {
                if fm.is_some() && disk.is_some() {
+                    if fence_lost() {
+                        return Ok((res, Some(StorageError::SlowDown)));
+                    }
                    if let Err(err) = save_format_file(disk, fm).await {
                        if let Some(disk) = disk.as_ref() {
                            let _ = disk.close().await;
@@ -1101,6 +1101,18 @@ impl crate::storage_api_contracts::heal::HealOperations for Sets {
        }
        Ok((res, None))
    }
+}
+
+#[async_trait::async_trait]
+impl crate::storage_api_contracts::heal::HealOperations for Sets {
+    type Error = Error;
+    type HealResultItem = HealResultItem;
+    type HealOptions = HealOpts;
+
+    #[tracing::instrument(skip(self))]
+    async fn heal_format(&self, dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
+        self.heal_format_with_fence(dry_run, || false).await
+    }
    #[tracing::instrument(skip(self))]
    async fn heal_bucket(&self, bucket: &str, opts: &HealOpts) -> Result<HealResultItem> {
        let mut result = HealResultItem {
@@ -26,7 +26,7 @@ use crate::storage_api_contracts::{
    namespace::NamespaceLocking as _,
    object::{HTTPPreconditions, ObjectOperations as _},
 };
-use crate::store::ECStore;
+use crate::store::{ECStore, ObjectLockDiagGuard, SourceCleanupMutationFence};
 use bytes::Bytes;
 use rustfs_filemeta::{FileInfo, FileInfoVersions, ObjectPartInfo};
 use rustfs_rio::{EtagResolvable, HashReader, HashReaderDetector, Index, TryGetIndex};
@@ -856,7 +856,6 @@ fn is_equivalent_data_movement_object(source: &ObjectInfo, target: &ObjectInfo)
 fn is_superseding_unversioned_data_movement_object(source: &ObjectInfo, target: &ObjectInfo) -> bool {
    is_unversioned_data_movement_object(source)
        && is_unversioned_data_movement_object(target)
-        && !target.delete_marker
        && source
            .mod_time
            .zip(target.mod_time)
@@ -1028,6 +1027,7 @@ pub(crate) enum SourceCleanupError {
 pub(crate) struct SourceCleanupBucketFence<'a> {
    pub(crate) expected_incarnation_id: Option<uuid::Uuid>,
    pub(crate) lifecycle_guard: Option<&'a rustfs_lock::NamespaceLockGuard>,
+    pub(crate) object_mutation_fence: Option<&'a SourceCleanupMutationFence>,
 }

 fn ensure_source_cleanup_versions_match(
@@ -1065,7 +1065,9 @@ pub(crate) async fn ensure_source_cleanup_versions_unchanged(
 struct SourceCleanupDeleteBarrierState {
    bucket: String,
    object: String,
+    fence_pending: tokio::sync::Notify,
    arrived: tokio::sync::Notify,
+    is_paused: AtomicBool,
    release: tokio::sync::Notify,
 }

@@ -1079,7 +1081,7 @@ pub(crate) struct SourceCleanupDeleteBarrier {
 }

 #[cfg(test)]
-static SOURCE_CLEANUP_DELETE_BARRIER: std::sync::OnceLock<std::sync::Mutex<Option<Arc<SourceCleanupDeleteBarrierState>>>> =
+static SOURCE_CLEANUP_DELETE_BARRIERS: std::sync::OnceLock<std::sync::Mutex<Vec<Arc<SourceCleanupDeleteBarrierState>>>> =
    std::sync::OnceLock::new();

 #[cfg(test)]
@@ -1092,15 +1094,22 @@ impl SourceCleanupDeleteBarrier {
        let state = Arc::new(SourceCleanupDeleteBarrierState {
            bucket: bucket.to_string(),
            object: object.to_string(),
+            fence_pending: tokio::sync::Notify::new(),
            arrived: tokio::sync::Notify::new(),
+            is_paused: AtomicBool::new(false),
            release: tokio::sync::Notify::new(),
        });
-        let mut slot = SOURCE_CLEANUP_DELETE_BARRIER
-            .get_or_init(|| std::sync::Mutex::new(None))
+        let mut barriers = SOURCE_CLEANUP_DELETE_BARRIERS
+            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
            .lock()
            .expect("source cleanup delete barrier mutex should not poison");
-        assert!(slot.is_none(), "source cleanup delete barrier must be unique");
-        *slot = Some(Arc::clone(&state));
+        assert!(
+            !barriers
+                .iter()
+                .any(|barrier| barrier.bucket == bucket && barrier.object == object),
+            "source cleanup delete barrier must be unique per object"
+        );
+        barriers.push(Arc::clone(&state));
        Self { state }
    }

@@ -1110,35 +1119,58 @@ impl SourceCleanupDeleteBarrier {
            .expect("source cleanup should reach the pre-delete barrier");
    }

+    pub(crate) async fn wait_until_fence_pending(&self) {
+        tokio::time::timeout(StdDuration::from_secs(30), self.state.fence_pending.notified())
+            .await
+            .expect("source cleanup should attempt the fixed mutation fence");
+    }
+
+    pub(crate) fn is_paused(&self) -> bool {
+        self.state.is_paused.load(Ordering::Acquire)
+    }
+
    pub(crate) fn release(&self) {
        self.state.release.notify_one();
    }
 }

+#[cfg(test)]
+pub(crate) fn notify_source_cleanup_mutation_fence_pending(bucket: &str, object: &str) {
+    let barrier = SOURCE_CLEANUP_DELETE_BARRIERS
+        .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+        .lock()
+        .expect("source cleanup delete barrier mutex should not poison")
+        .iter()
+        .find(|barrier| barrier.bucket == bucket && barrier.object == object)
+        .cloned();
+    if let Some(barrier) = barrier {
+        barrier.fence_pending.notify_one();
+    }
+}
+
 #[cfg(test)]
 impl Drop for SourceCleanupDeleteBarrier {
    fn drop(&mut self) {
        self.state.release.notify_one();
-        let mut slot = SOURCE_CLEANUP_DELETE_BARRIER
-            .get_or_init(|| std::sync::Mutex::new(None))
+        let mut barriers = SOURCE_CLEANUP_DELETE_BARRIERS
+            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
            .lock()
            .expect("source cleanup delete barrier mutex should not poison");
-        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
-            *slot = None;
-        }
+        barriers.retain(|state| !Arc::ptr_eq(state, &self.state));
    }
 }

 #[cfg(test)]
 async fn pause_source_cleanup_before_delete(bucket: &str, object: &str) {
-    let barrier = SOURCE_CLEANUP_DELETE_BARRIER
-        .get_or_init(|| std::sync::Mutex::new(None))
+    let barrier = SOURCE_CLEANUP_DELETE_BARRIERS
+        .get_or_init(|| std::sync::Mutex::new(Vec::new()))
        .lock()
        .expect("source cleanup delete barrier mutex should not poison")
-        .as_ref()
-        .filter(|barrier| barrier.bucket == bucket && barrier.object == object)
+        .iter()
+        .find(|barrier| barrier.bucket == bucket && barrier.object == object)
        .cloned();
    if let Some(barrier) = barrier {
+        barrier.is_paused.store(true, Ordering::Release);
        barrier.arrived.notify_one();
        barrier.release.notified().await;
    }
@@ -1154,11 +1186,20 @@ pub(crate) async fn cleanup_source_entry_if_unchanged(
    op_label: &str,
 ) -> std::result::Result<ObjectInfo, SourceCleanupError> {
    let cleanup_key = encode_dir_object(object);
-    let ns_lock = set.new_ns_lock(bucket, cleanup_key.as_str()).await?;
-    let _guard = ns_lock
-        .get_write_lock(get_lock_acquire_timeout())
-        .await
-        .map_err(Error::from)?;
+    let source_guard = if bucket_fence
+        .object_mutation_fence
+        .is_some_and(SourceCleanupMutationFence::source_lock_covered)
+    {
+        None
+    } else {
+        let ns_lock = set.new_ns_lock(bucket, cleanup_key.as_str()).await?;
+        Some(
+            ns_lock
+                .get_write_lock(get_lock_acquire_timeout())
+                .await
+                .map_err(Error::from)?,
+        )
+    };

    if bucket_fence
        .lifecycle_guard
@@ -1168,6 +1209,14 @@ pub(crate) async fn cleanup_source_entry_if_unchanged(
            "{op_label}: bucket incarnation fence was lost before source cleanup"
        ))));
    }
+    if bucket_fence
+        .object_mutation_fence
+        .is_some_and(SourceCleanupMutationFence::is_lock_lost)
+    {
+        return Err(SourceCleanupError::Storage(Error::other(format!(
+            "{op_label}: object mutation fence was lost before source cleanup"
+        ))));
+    }

    ensure_source_cleanup_versions_unchanged(set.clone(), bucket, object, expected, allowed_missing, op_label).await?;

@@ -1182,7 +1231,12 @@ pub(crate) async fn cleanup_source_entry_if_unchanged(
        expected_bucket_incarnation_id: bucket_fence.expected_incarnation_id,
        ..Default::default()
    };
-    opts.add_namespace_lock_guard(&_guard);
+    if let Some(source_guard) = source_guard.as_ref() {
+        opts.add_namespace_lock_guard(source_guard);
+    }
+    if let Some(object_mutation_fence) = bucket_fence.object_mutation_fence {
+        object_mutation_fence.add_namespace_lock_fence(&mut opts);
+    }
    if let Some(bucket_lifecycle_guard) = bucket_fence.lifecycle_guard {
        opts.add_bucket_lifecycle_lock_guard(bucket_lifecycle_guard);
    }
@@ -1330,6 +1384,37 @@ fn data_movement_part_upload_failure_stage(err: &Error) -> &'static str {
    }
 }

+pub(crate) async fn migrate_decommission_object(
+    store: Arc<ECStore>,
+    pool_idx: usize,
+    bucket: String,
+    rd: GetObjectReader,
+    source_bucket_incarnation_id: Option<uuid::Uuid>,
+    op_label: &str,
+) -> Result<()> {
+    let source = rd.object_info.clone();
+    let _mutation_fence = store
+        .acquire_decommission_object_mutation_fence(&bucket, &source.name)
+        .await?;
+    let current = find_data_movement_target_info(store.as_ref(), pool_idx, &bucket, &source)
+        .await?
+        .ok_or(Error::FileNotFound)?;
+    if !is_equivalent_data_movement_object_identity(&source, &current, true, false) {
+        return Err(Error::FileNotFound);
+    }
+
+    migrate_object_inner(
+        store,
+        pool_idx,
+        bucket,
+        rd,
+        source_bucket_incarnation_id,
+        op_label,
+        Some(&_mutation_fence),
+    )
+    .await
+}
+
 pub(crate) async fn migrate_object(
    store: Arc<ECStore>,
    pool_idx: usize,
@@ -1337,6 +1422,18 @@ pub(crate) async fn migrate_object(
    rd: GetObjectReader,
    source_bucket_incarnation_id: Option<uuid::Uuid>,
    op_label: &str,
+) -> Result<()> {
+    migrate_object_inner(store, pool_idx, bucket, rd, source_bucket_incarnation_id, op_label, None).await
+}
+
+async fn migrate_object_inner(
+    store: Arc<ECStore>,
+    pool_idx: usize,
+    bucket: String,
+    rd: GetObjectReader,
+    source_bucket_incarnation_id: Option<uuid::Uuid>,
+    op_label: &str,
+    mutation_fence: Option<&ObjectLockDiagGuard>,
 ) -> Result<()> {
    let object_info = rd.object_info.clone();
    let has_part_checksums = object_info
@@ -1350,7 +1447,7 @@ pub(crate) async fn migrate_object(
        let mut new_multipart_opts = data_movement_new_multipart_opts(&object_info, pool_idx);
        new_multipart_opts.expected_bucket_incarnation_id = source_bucket_incarnation_id;
        let (res, target_pool_idx, expected_bucket_incarnation_id) = match store
-            .handle_new_multipart_upload_with_pool_idx(&bucket, &object_info.name, &new_multipart_opts)
+            .handle_new_multipart_upload_with_pool_idx(&bucket, &object_info.name, &new_multipart_opts, mutation_fence)
            .await
        {
            Ok(res) => res,
@@ -1448,7 +1545,7 @@ pub(crate) async fn migrate_object(
            if let Err(err) = store
                .clone()
                .complete_multipart_upload_for_data_movement(
-                    target_pool_idx,
+                    (target_pool_idx, mutation_fence),
                    &bucket,
                    &object_info.name,
                    &res.upload_id,
@@ -1609,7 +1706,7 @@ pub(crate) async fn migrate_object(
    let mut put_opts = data_movement_put_object_opts(&object_info, pool_idx);
    put_opts.expected_bucket_incarnation_id = source_bucket_incarnation_id;
    let (target_pool_idx, put_result) = store
-        .put_object_for_data_movement(&bucket, &object_info.name, &mut data, &put_opts)
+        .put_object_for_data_movement(&bucket, &object_info.name, &mut data, &put_opts, mutation_fence)
        .await
        .map_err(|err| data_movement_stage_error(op_label, "prepare_put_object", &bucket, &object_info.name, err))?;
    if let Err(err) = put_result {
@@ -3541,25 +3638,47 @@ mod tests {
    }

    #[test]
-    fn test_precondition_conflict_rejects_newer_delete_marker() {
-        let source = ObjectInfo {
-            size: 128,
-            etag: Some("etag-source".to_string()),
-            mod_time: Some(OffsetDateTime::UNIX_EPOCH),
-            ..Default::default()
-        };
-        let target = ObjectInfo {
-            delete_marker: true,
-            etag: None,
-            mod_time: OffsetDateTime::UNIX_EPOCH.checked_add(time::Duration::SECOND),
-            ..source.clone()
-        };
+    fn test_precondition_conflict_accepts_only_newer_null_delete_marker() {
+        for version_id in [None, Some(Uuid::nil())] {
+            let source = ObjectInfo {
+                version_id,
+                size: 128,
+                etag: Some("etag-source".to_string()),
+                mod_time: Some(OffsetDateTime::UNIX_EPOCH),
+                ..Default::default()
+            };
+            let target = ObjectInfo {
+                delete_marker: true,
+                etag: None,
+                mod_time: OffsetDateTime::UNIX_EPOCH.checked_add(time::Duration::SECOND),
+                ..source.clone()
+            };

-        let should_resume =
-            resolve_data_movement_overwrite_resume_result(&Error::PreconditionFailed, Ok(Some(target)), &source, 0, 1)
-                .expect("delete marker conflict should be evaluated");
+            assert!(
+                resolve_data_movement_overwrite_resume_result(
+                    &Error::PreconditionFailed,
+                    Ok(Some(target.clone())),
+                    &source,
+                    0,
+                    1,
+                )
+                .expect("newer null delete marker should be evaluated")
+            );

-        assert!(!should_resume);
+            let mut same_time = target.clone();
+            same_time.mod_time = source.mod_time;
+            assert!(
+                !resolve_data_movement_overwrite_resume_result(&Error::PreconditionFailed, Ok(Some(same_time)), &source, 0, 1,)
+                    .expect("same-generation null delete marker should be rejected")
+            );
+
+            let mut versioned = target;
+            versioned.version_id = Some(Uuid::new_v4());
+            assert!(
+                !resolve_data_movement_overwrite_resume_result(&Error::PreconditionFailed, Ok(Some(versioned)), &source, 0, 1,)
+                    .expect("a UUID delete marker must not erase a null source version")
+            );
+        }
    }

    #[test]
@@ -858,6 +858,7 @@ const EVENT_DISK_LOCAL_DIRECT_IO_FALLBACK: &str = "disk_local_direct_io_fallback
 #[cfg(target_os = "linux")]
 const EVENT_DISK_LOCAL_URING_LATCH_OFF: &str = "disk_local_uring_latch_off";
 const EVENT_DISK_LOCAL_DELETE_FAILED: &str = "disk_local_delete_failed";
+const EVENT_DISK_LOCAL_DELETE_ROLLBACK_FAILED: &str = "disk_local_delete_rollback_failed";
 const EVENT_DISK_LOCAL_CHECK_PARTS: &str = "disk_local_check_parts";
 const EVENT_DISK_LOCAL_ACCESS_FAILED: &str = "disk_local_access_failed";
 const EVENT_DISK_LOCAL_VOLUME_SETUP_FAILED: &str = "disk_local_volume_setup_failed";
@@ -6106,6 +6107,43 @@ impl LocalDisk {
        Ok((bytes, modtime))
    }

+    async fn write_missing_delete_marker(
+        &self,
+        volume: &str,
+        path: &str,
+        fi: FileInfo,
+        object_dir: &Path,
+        xl_path: &Path,
+        rollback_dir: Option<Uuid>,
+    ) -> Result<()> {
+        if let Some(rollback_dir) = rollback_dir {
+            let rollback_path = object_dir.join(rollback_dir.to_string());
+            fs::create_dir_all(&rollback_path).await.map_err(to_file_error)?;
+            fs::write(rollback_path.join(DELETE_MARKER_ROLLBACK_FILE), [])
+                .await
+                .map_err(to_file_error)?;
+        }
+        if let Err(err) = self.write_metadata("", volume, path, fi).await {
+            if let Some(rollback_dir) = rollback_dir
+                && let Err(restore_err) = restore_delete_rollback(object_dir, xl_path, rollback_dir, &self.publication_root).await
+            {
+                warn!(
+                    event = EVENT_DISK_LOCAL_DELETE_ROLLBACK_FAILED,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_DISK_LOCAL,
+                    result = "failed",
+                    volume,
+                    path,
+                    rollback_dir = %rollback_dir,
+                    error = ?restore_err,
+                    "Disk local delete rollback failed"
+                );
+            }
+            return Err(err);
+        }
+        Ok(())
+    }
+
    async fn delete_versions_internal(&self, volume: &str, path: &str, fis: &[FileInfo], opts: &DeleteOptions) -> Result<()> {
        let volume_dir = self.io_get_bucket_path(volume)?;
        let xlpath = self.io_get_object_path(volume, format!("{path}/{STORAGE_FORMAT_FILE}").as_str())?;
@@ -6123,7 +6161,20 @@ impl LocalDisk {
            return restore_metadata_backup(object_dir, &xlpath, rollback_dir, &self.publication_root).await;
        }

-        let (data, _) = self.read_all_data_with_dmtime(volume, volume_dir.as_path(), &xlpath).await?;
+        let (data, _) = match self.read_all_data_with_dmtime(volume, volume_dir.as_path(), &xlpath).await {
+            Ok(data) => data,
+            Err(DiskError::FileNotFound) => {
+                // `deleted` alone can be an explicit marker purge; only
+                // `mark_deleted` may create metadata that was not present.
+                let Some(delete_marker) = fis.iter().find(|fi| fi.deleted && fi.mark_deleted).cloned() else {
+                    return Err(DiskError::FileNotFound);
+                };
+                return self
+                    .write_missing_delete_marker(volume, path, delete_marker, object_dir, &xlpath, opts.old_data_dir)
+                    .await;
+            }
+            Err(err) => return Err(err),
+        };

        if data.is_empty() {
            return Err(DiskError::FileNotFound);
@@ -10422,29 +10473,9 @@ impl DiskAPI for LocalDisk {
                }

                if fi.deleted && force_del_marker {
-                    if let Some(rollback_dir) = rollback_dir {
-                        let rollback_path = file_path.join(rollback_dir.to_string());
-                        fs::create_dir_all(&rollback_path).await.map_err(to_file_error)?;
-                        fs::write(rollback_path.join(DELETE_MARKER_ROLLBACK_FILE), [])
-                            .await
-                            .map_err(to_file_error)?;
-                    }
-                    if let Err(err) = self.write_metadata("", volume, path, fi).await {
-                        if let Some(rollback_dir) = rollback_dir
-                            && let Err(restore_err) =
-                                restore_delete_rollback(file_path.as_path(), &xl_path, rollback_dir, &self.publication_root).await
-                        {
-                            warn!(
-                                volume,
-                                path,
-                                rollback_dir = %rollback_dir,
-                                error = ?restore_err,
-                                "failed to restore metadata after delete marker commit error"
-                            );
-                        }
-                        return Err(err);
-                    }
-                    return Ok(());
+                    return self
+                        .write_missing_delete_marker(volume, path, fi, file_path.as_path(), &xl_path, rollback_dir)
+                        .await;
                }

                return if fi.version_id.is_some() {
@@ -44,6 +44,8 @@ pub const PART_TRANSACTION_ROLLBACK: &str = "rollback";
 const LOG_COMPONENT_ECSTORE: &str = "ecstore";
 const LOG_SUBSYSTEM_DISK: &str = "disk";
 const EVENT_DISK_PART_ERR_UNCLASSIFIED: &str = "disk_part_err_unclassified";
+const ENV_BATCH_READ_VERSION_SERVER_PARALLELISM: &str = "RUSTFS_BATCH_READ_VERSION_SERVER_PARALLELISM";
+const BATCH_READ_VERSION_SERVER_PARALLELISM: usize = 4;

 pub fn part_transaction_path(part_path: &str) -> String {
    match part_path.rsplit_once('/') {
@@ -62,6 +64,7 @@ use bytes::Bytes;
 use endpoint::Endpoint;
 use error::DiskError;
 use error::{Error, Result};
+use futures::stream::{self, StreamExt};
 use local::LocalDisk;
 use rustfs_filemeta::{FileInfo, ObjectPartInfo, RawFileInfo};
 use rustfs_madmin::info_commands::DiskMetrics;
@@ -417,6 +420,14 @@ impl DiskAPI for Disk {
        }
    }

+    #[tracing::instrument(level = "trace", skip_all)]
+    async fn batch_read_version(&self, req: BatchReadVersionReq) -> Result<Vec<BatchReadVersionResp>> {
+        match self {
+            Disk::Local(local_disk) => local_disk.batch_read_version(req).await,
+            Disk::Remote(remote_disk) => remote_disk.batch_read_version(req).await,
+        }
+    }
+
    #[tracing::instrument(level = "trace", skip_all)]
    async fn read_xl(&self, volume: &str, path: &str, read_data: bool) -> Result<RawFileInfo> {
        match self {
@@ -1028,36 +1039,47 @@ where
    D: DiskAPI + ?Sized,
 {
    validate_batch_read_version_item_count(req.items.len())?;
+    let parallelism = batch_read_version_server_parallelism();

-    let mut responses = Vec::with_capacity(req.items.len());
-    for (index, item) in req.items.iter().enumerate() {
-        let response = match disk
-            .read_version(&item.org_volume, &item.volume, &item.path, &item.version_id, &req.opts)
-            .await
-        {
-            Ok(file_info) => BatchReadVersionResp {
-                index,
-                path: item.path.clone(),
-                version_id: item.version_id.clone(),
-                success: true,
-                file_info,
-                error: String::new(),
-            },
-            Err(err) => BatchReadVersionResp {
-                index,
-                path: item.path.clone(),
-                version_id: item.version_id.clone(),
-                success: false,
-                file_info: FileInfo::default(),
-                error: err.to_string(),
-            },
-        };
-        responses.push(response);
-    }
+    let mut responses = stream::iter(req.items.into_iter().enumerate())
+        .map(|(index, item)| async move {
+            match disk
+                .read_version(&item.org_volume, &item.volume, &item.path, &item.version_id, &req.opts)
+                .await
+            {
+                Ok(file_info) => BatchReadVersionResp {
+                    index,
+                    path: item.path,
+                    version_id: item.version_id,
+                    success: true,
+                    file_info,
+                    error: String::new(),
+                    error_code: 0,
+                },
+                Err(err) => BatchReadVersionResp {
+                    index,
+                    path: item.path,
+                    version_id: item.version_id,
+                    success: false,
+                    file_info: FileInfo::default(),
+                    error: err.to_string(),
+                    error_code: err.to_u32(),
+                },
+            }
+        })
+        .buffer_unordered(parallelism)
+        .collect::<Vec<_>>()
+        .await;
+    responses.sort_unstable_by_key(|response| response.index);

    Ok(responses)
 }

+fn batch_read_version_server_parallelism() -> usize {
+    rustfs_utils::get_env_usize(ENV_BATCH_READ_VERSION_SERVER_PARALLELISM, BATCH_READ_VERSION_SERVER_PARALLELISM)
+        .clamp(1, BATCH_READ_VERSION_MAX_ITEMS)
+}
+
 #[derive(Debug, Default, Serialize, Deserialize)]
 pub struct CheckPartsResp {
    pub results: Vec<usize>,
@@ -1322,6 +1344,8 @@ pub struct BatchReadVersionResp {
    pub success: bool,
    pub file_info: FileInfo,
    pub error: String,
+    #[serde(default)]
+    pub error_code: u32,
 }

 pub fn validate_batch_read_version_item_count(item_count: usize) -> Result<()> {
@@ -1417,6 +1441,26 @@ mod tests {
        assert!(!partial_valid_location.valid());
    }

+    #[test]
+    fn batch_read_version_server_parallelism_defaults_to_conservative_four() {
+        temp_env::with_var(ENV_BATCH_READ_VERSION_SERVER_PARALLELISM, None::<&str>, || {
+            assert_eq!(batch_read_version_server_parallelism(), 4);
+        });
+    }
+
+    #[test]
+    fn batch_read_version_server_parallelism_honors_env_with_bounds() {
+        temp_env::with_var(ENV_BATCH_READ_VERSION_SERVER_PARALLELISM, Some("8"), || {
+            assert_eq!(batch_read_version_server_parallelism(), 8);
+        });
+        temp_env::with_var(ENV_BATCH_READ_VERSION_SERVER_PARALLELISM, Some("0"), || {
+            assert_eq!(batch_read_version_server_parallelism(), 1);
+        });
+        temp_env::with_var(ENV_BATCH_READ_VERSION_SERVER_PARALLELISM, Some("9999"), || {
+            assert_eq!(batch_read_version_server_parallelism(), BATCH_READ_VERSION_MAX_ITEMS);
+        });
+    }
+
    /// Test FileInfoVersions find_version_index
    #[test]
    fn test_file_info_versions_find_version_index() {
@@ -784,6 +784,24 @@ pub(crate) fn create_deferred_bitrot_reader_with_stripe_handle(
 ///
 /// # Returns
 /// A Result containing the BitrotWriterWrapper or an error
+/// Size hint handed to `DiskAPI::create_file` for a bitrot-wrapped shard.
+///
+/// A known length is grown by one checksum per shard so the on-disk file size
+/// matches what the bitrot writer emits. A negative length is the
+/// unknown-size sentinel (`HashReader::SIZE_PRESERVE_LAYER`, used by SSE and
+/// compression) and must be preserved: `RemoteDisk::create_file` forwards it
+/// in the `put_file_stream` query, and the receiver only treats `size > 0` as
+/// a fixed body length when locating the authenticated trailer. Clamping it
+/// to `0` would claim an empty body and misframe the stream. `0` stays `0`
+/// because a genuinely empty object still means an empty body.
+fn bitrot_create_file_size(length: i64, shard_size: usize, checksum_algo: &HashAlgorithm) -> i64 {
+    if length <= 0 {
+        return length;
+    }
+    let length = length as usize;
+    (length.div_ceil(shard_size) * checksum_algo.size() + length) as i64
+}
+
 pub async fn create_bitrot_writer(
    is_inline_buffer: bool,
    disk: Option<&DiskStore>,
@@ -796,12 +814,7 @@ pub async fn create_bitrot_writer(
    let writer = if is_inline_buffer {
        CustomWriter::new_inline_buffer()
    } else if let Some(disk) = disk {
-        let length = if length > 0 {
-            let length = length as usize;
-            (length.div_ceil(shard_size) * checksum_algo.size() + length) as i64
-        } else {
-            0
-        };
+        let length = bitrot_create_file_size(length, shard_size, &checksum_algo);

        let file = disk.create_file("", volume, path, length).await?;
        #[cfg(feature = "hotpath")]
@@ -820,6 +833,25 @@ mod tests {
    use rustfs_rio::ChunkReader;
    use std::collections::VecDeque;

+    #[test]
+    fn bitrot_create_file_size_grows_known_length_by_checksums() {
+        // 10 bytes over 4-byte shards = 3 shards, each followed by a 32-byte hash.
+        assert_eq!(bitrot_create_file_size(10, 4, &HashAlgorithm::HighwayHash256), 10 + 3 * 32);
+        assert_eq!(bitrot_create_file_size(10, 4, &HashAlgorithm::None), 10);
+    }
+
+    #[test]
+    fn bitrot_create_file_size_keeps_empty_and_unknown_distinct() {
+        assert_eq!(bitrot_create_file_size(0, 4, &HashAlgorithm::HighwayHash256), 0);
+        // SSE/compression streams advertise SIZE_PRESERVE_LAYER (-1); the remote
+        // put_file_stream receiver relies on a non-positive size to parse the auth
+        // trailer from the stream tail, so the sentinel must survive untouched.
+        assert_eq!(
+            bitrot_create_file_size(rustfs_rio::HashReader::SIZE_PRESERVE_LAYER, 4, &HashAlgorithm::HighwayHash256),
+            rustfs_rio::HashReader::SIZE_PRESERVE_LAYER
+        );
+    }
+
    struct TestChunkReader {
        chunks: VecDeque<Bytes>,
    }
@@ -81,6 +81,14 @@ pub fn shutdown_background_monitors() {
    cluster::rpc::shutdown_background_monitors();
 }

+/// Publish that the process is ready to serve user-object GET traffic.
+///
+/// Experimental metadata coalescing is allowed to run only after this point so
+/// startup and internal metadata reads keep the original per-disk path.
+pub fn mark_get_metadata_read_version_coalescing_service_ready() {
+    runtime::global::mark_get_metadata_read_version_coalescing_service_ready();
+}
+
 #[cfg(test)]
 mod rio_tests {
    #[test]
@@ -24,7 +24,7 @@ use crate::storage_api_contracts::{
 pub struct NamespaceLockFence {
    signals: Arc<Vec<Arc<rustfs_lock::distributed_lock::LockLostSignal>>>,
    #[cfg(test)]
-    forced_lost: Arc<std::sync::atomic::AtomicBool>,
+    forced_lost: Arc<Vec<Arc<std::sync::atomic::AtomicBool>>>,
 }

 impl Debug for NamespaceLockFence {
@@ -40,13 +40,17 @@ impl NamespaceLockFence {
        Self {
            signals: Arc::default(),
            #[cfg(test)]
-            forced_lost: Arc::new(std::sync::atomic::AtomicBool::new(false)),
+            forced_lost: Arc::new(vec![Arc::new(std::sync::atomic::AtomicBool::new(false))]),
        }
    }

    pub(crate) fn is_lock_lost(&self) -> bool {
        #[cfg(test)]
-        if self.forced_lost.load(std::sync::atomic::Ordering::Acquire) {
+        if self
+            .forced_lost
+            .iter()
+            .any(|lost| lost.load(std::sync::atomic::Ordering::Acquire))
+        {
            return true;
        }
        self.signals.iter().any(|signal| signal.is_lost())
@@ -57,27 +61,26 @@ impl NamespaceLockFence {
    }

    fn extend(&mut self, other: &Self) {
-        if Arc::ptr_eq(&self.signals, &other.signals) {
-            return;
+        if !Arc::ptr_eq(&self.signals, &other.signals) {
+            Arc::make_mut(&mut self.signals).extend(other.signals.iter().cloned());
        }
-        Arc::make_mut(&mut self.signals).extend(other.signals.iter().cloned());
        #[cfg(test)]
-        if other.forced_lost.load(std::sync::atomic::Ordering::Acquire) {
-            self.forced_lost.store(true, std::sync::atomic::Ordering::Release);
+        if !Arc::ptr_eq(&self.forced_lost, &other.forced_lost) {
+            Arc::make_mut(&mut self.forced_lost).extend(other.forced_lost.iter().cloned());
        }
    }

    #[cfg(test)]
    pub(crate) fn lost_for_test() -> Self {
        let fence = Self::new();
-        fence.forced_lost.store(true, std::sync::atomic::Ordering::Release);
+        fence.forced_lost[0].store(true, std::sync::atomic::Ordering::Release);
        fence
    }

    #[cfg(test)]
    pub(crate) fn loss_handle_for_test() -> (Self, Arc<std::sync::atomic::AtomicBool>) {
        let fence = Self::new();
-        (fence.clone(), Arc::clone(&fence.forced_lost))
+        (fence.clone(), Arc::clone(&fence.forced_lost[0]))
    }
 }

@@ -411,6 +414,13 @@ impl ObjectOptions {
        self.namespace_lock_fence.get_or_insert_with(NamespaceLockFence::new);
    }

+    #[cfg(test)]
+    pub(crate) fn add_namespace_lock_fence_for_test(&mut self, fence: &NamespaceLockFence) {
+        self.namespace_lock_fence
+            .get_or_insert_with(NamespaceLockFence::new)
+            .extend(fence);
+    }
+
    pub(crate) fn ensure_lifecycle_delete_all_journal(&mut self) {
        self.lifecycle_delete_all_journal
            .get_or_insert_with(|| Arc::new(parking_lot::Mutex::new(LifecycleDeleteAllJournalState::default())));
@@ -25,7 +25,10 @@ use lazy_static::lazy_static;
 use rustfs_lock::client::LockClient;
 use std::{
    collections::HashMap,
-    sync::{Arc, OnceLock},
+    sync::{
+        Arc, OnceLock,
+        atomic::{AtomicBool, Ordering},
+    },
    time::SystemTime,
 };
 use tokio::sync::{OnceCell, RwLock};
@@ -37,6 +40,16 @@ pub const DISK_MIN_INODES: u64 = 1000;
 pub const DISK_FILL_FRACTION: f64 = 0.99;
 pub const DISK_RESERVE_FRACTION: f64 = 0.15;

+static GET_METADATA_READ_VERSION_COALESCING_SERVICE_READY: AtomicBool = AtomicBool::new(false);
+
+pub(crate) fn mark_get_metadata_read_version_coalescing_service_ready() {
+    GET_METADATA_READ_VERSION_COALESCING_SERVICE_READY.store(true, Ordering::Release);
+}
+
+pub(crate) fn get_metadata_read_version_coalescing_service_ready() -> bool {
+    GET_METADATA_READ_VERSION_COALESCING_SERVICE_READY.load(Ordering::Acquire)
+}
+
 // Global singletons for backward compatibility with MinIO port.
 // These should be migrated to AppContext over time.
 // See issue #730 for migration plan.
@@ -160,6 +160,10 @@ pub struct InstanceContext {
    /// workers (scanner/heal/tier/lifecycle) without touching another instance.
    /// Replaces the process-global cancel-token static.
    background_cancel_token: OnceLock<CancellationToken>,
+    /// Serializes decommission data-movement operations with cancellation and
+    /// a subsequent restart. Readers are held across one object side effect;
+    /// the transition path takes the writer after cancelling the routine.
+    decommission_operation_gate: Arc<RwLock<()>>,
    /// Resolves object-encryption material at the application boundary.
    object_encryption_resolver: OnceLock<Arc<dyn ObjectEncryptionResolver>>,
    tier_delete_journal_recovery_stores: std::sync::Mutex<HashSet<Uuid>>,
@@ -200,6 +204,7 @@ impl InstanceContext {
            local_disk_set_drives: Arc::new(RwLock::new(Vec::new())),
            bucket_metadata_sys: std::sync::Mutex::new(None),
            background_cancel_token: OnceLock::new(),
+            decommission_operation_gate: Arc::new(RwLock::new(())),
            object_encryption_resolver: OnceLock::new(),
            tier_delete_journal_recovery_stores: std::sync::Mutex::new(HashSet::new()),
            transition_transaction_recovery_stores: std::sync::Mutex::new(HashSet::new()),
@@ -218,6 +223,10 @@ impl InstanceContext {
        self.lock_manager.clone()
    }

+    pub(crate) fn decommission_operation_gate(&self) -> Arc<RwLock<()>> {
+        Arc::clone(&self.decommission_operation_gate)
+    }
+
    /// Install the application-owned object-encryption resolver once.
    pub fn set_object_encryption_resolver(
        &self,
@@ -256,6 +256,10 @@ fn to_madmin_scanner_metrics(metrics: rustfs_common::metrics::ScannerMetricsRepo
        cycle_max_duration_seconds: metrics.cycle_max_duration_seconds,
        cycle_max_objects: metrics.cycle_max_objects,
        cycle_max_directories: metrics.cycle_max_directories,
+        cycle_timeout_total: metrics.cycle_timeout_total,
+        cycle_recovery_required_total: metrics.cycle_recovery_required_total,
+        cycle_last_progress_age: metrics.cycle_last_progress_age,
+        leader_lease_without_progress: metrics.leader_lease_without_progress,
        bitrot_cycle_enabled: metrics.bitrot_cycle_enabled,
        bitrot_cycle_seconds: metrics.bitrot_cycle_seconds,
        scan_checkpoint: metrics.scan_checkpoint.map(|checkpoint| MadminScannerCheckpointReport {
@@ -611,6 +615,10 @@ mod test {
            current_started: chrono_to_jiff_timestamp(current_started),
            last_cycle_partial_source: "usage".to_string(),
            last_cycle_partial_source_code: 1,
+            cycle_timeout_total: 3,
+            cycle_recovery_required_total: 2,
+            cycle_last_progress_age: 17,
+            leader_lease_without_progress: true,
            partial_cycles_by_source: vec![rustfs_common::metrics::ScannerSourceCycleSnapshot {
                source: "usage".to_string(),
                cycles: 2,
@@ -622,6 +630,10 @@ mod test {
        assert_eq!(scanner.current_started, chrono_to_jiff_timestamp(current_started));
        assert_eq!(scanner.last_cycle_partial_source, "usage");
        assert_eq!(scanner.last_cycle_partial_source_code, 1);
+        assert_eq!(scanner.cycle_timeout_total, 3);
+        assert_eq!(scanner.cycle_recovery_required_total, 2);
+        assert_eq!(scanner.cycle_last_progress_age, 17);
+        assert!(scanner.leader_lease_without_progress);
        let usage = scanner
            .partial_cycles_by_source
            .iter()
@@ -334,6 +334,7 @@ impl ECStore {
                            lifecycle_guard: bucket_incarnation_fence
                                .as_ref()
                                .and_then(|guard| guard.namespace_lock_guard()),
+                            ..Default::default()
                        },
                        "rebalance",
                    ),
@@ -53,11 +53,12 @@ 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::disk_store::DiskStoreRenameDataExt;
+use crate::disk::disk_store::{DiskStoreRenameDataExt, get_drive_metadata_timeout};
 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, STORAGE_FORMAT_FILE_BACKUP, part_transaction_path,
+    BATCH_READ_VERSION_MAX_ITEMS, BatchReadVersionItem, BatchReadVersionReq, BatchReadVersionResp, DataDirDeleteStatus, Disk,
+    OldCurrentSize, PART_TRANSACTION_NEW_META, PART_TRANSACTION_OLD_META, PART_TRANSACTION_ROLLBACK, PartTransactionAction,
+    STORAGE_FORMAT_FILE_BACKUP, part_transaction_path,
 };
 use crate::erasure::coding::BitrotReader;
 use crate::io_support::bitrot::ShardReader;
@@ -75,7 +76,7 @@ use std::{
    future::Future,
    pin::Pin,
    sync::{
-        OnceLock,
+        Arc, OnceLock,
        atomic::{AtomicUsize, Ordering},
    },
    task::{Context, Poll},
@@ -94,6 +95,242 @@ fn metadata_distribution_key(bucket: &str, object: &str) -> String {
    [bucket, object].join("/")
 }

+fn read_version_coalescing_enabled() -> bool {
+    let enabled = || {
+        rustfs_utils::get_env_opt_str(ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE)
+            .is_some_and(|value| value.eq_ignore_ascii_case("auto") || value.eq_ignore_ascii_case("on"))
+    };
+
+    #[cfg(test)]
+    {
+        enabled()
+    }
+
+    #[cfg(not(test))]
+    {
+        static ENABLED: OnceLock<bool> = OnceLock::new();
+        *ENABLED.get_or_init(enabled)
+    }
+}
+
+fn read_version_coalescing_delay() -> Duration {
+    #[cfg(test)]
+    {
+        let micros = rustfs_utils::get_env_u64(
+            ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS,
+            DEFAULT_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS,
+        );
+        Duration::from_micros(micros)
+    }
+
+    #[cfg(not(test))]
+    {
+        static DELAY: OnceLock<Duration> = OnceLock::new();
+        *DELAY.get_or_init(|| {
+            Duration::from_micros(rustfs_utils::get_env_u64(
+                ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS,
+                DEFAULT_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS,
+            ))
+        })
+    }
+}
+
+struct CoalescedReadVersionRequest {
+    item: BatchReadVersionItem,
+    tx: oneshot::Sender<disk::error::Result<FileInfo>>,
+}
+
+#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
+struct ReadVersionCoalescerKey {
+    disk: usize,
+    incl_free_versions: bool,
+    read_data: bool,
+    healing: bool,
+}
+
+impl ReadVersionCoalescerKey {
+    fn new(disk: &DiskStore, opts: &ReadOptions) -> Self {
+        Self {
+            disk: Arc::as_ptr(disk) as usize,
+            incl_free_versions: opts.incl_free_versions,
+            read_data: opts.read_data,
+            healing: opts.healing,
+        }
+    }
+}
+
+#[derive(Default)]
+struct ReadVersionCoalescer {
+    lanes: HashMap<ReadVersionCoalescerKey, Vec<CoalescedReadVersionRequest>>,
+}
+
+fn read_version_coalescer() -> &'static Mutex<ReadVersionCoalescer> {
+    static COALESCER: OnceLock<Mutex<ReadVersionCoalescer>> = OnceLock::new();
+    COALESCER.get_or_init(|| Mutex::new(ReadVersionCoalescer::default()))
+}
+
+fn record_read_version_coalescer_event(event: &'static str, item_count: usize) {
+    counter!(
+        METRIC_GET_METADATA_READ_VERSION_COALESCER_TOTAL,
+        "event" => event,
+        "item_count" => item_count.to_string()
+    )
+    .increment(1);
+}
+
+async fn read_version_via_coalescer(
+    disk: DiskStore,
+    org_bucket: &str,
+    bucket: &str,
+    object: &str,
+    version_id: &str,
+    opts: &ReadOptions,
+    allow_coalescing: bool,
+) -> disk::error::Result<FileInfo> {
+    if !allow_coalescing || !read_version_coalescing_enabled() {
+        return disk.read_version(org_bucket, bucket, object, version_id, opts).await;
+    }
+    if !matches!(disk.as_ref(), Disk::Remote(_)) {
+        record_read_version_coalescer_event("bypass_non_remote", 1);
+        return disk.read_version(org_bucket, bucket, object, version_id, opts).await;
+    }
+
+    let (tx, rx) = oneshot::channel();
+    let item = BatchReadVersionItem {
+        org_volume: org_bucket.to_string(),
+        volume: bucket.to_string(),
+        path: object.to_string(),
+        version_id: version_id.to_string(),
+    };
+    let lane_key = ReadVersionCoalescerKey::new(&disk, opts);
+    let pending = {
+        let mut coalescer = read_version_coalescer().lock().await;
+        let lane = coalescer.lanes.entry(lane_key).or_default();
+        let schedule_delayed_flush = lane.is_empty();
+        lane.push(CoalescedReadVersionRequest { item, tx });
+        if lane.len() >= BATCH_READ_VERSION_MAX_ITEMS {
+            coalescer.lanes.remove(&lane_key)
+        } else if schedule_delayed_flush {
+            let disk = disk.clone();
+            let task_opts = *opts;
+            tokio::spawn(async move {
+                tokio::time::sleep(read_version_coalescing_delay()).await;
+                flush_read_version_coalescer_lane(lane_key, disk, task_opts).await;
+            });
+            None
+        } else {
+            None
+        }
+    };
+
+    if let Some(pending) = pending {
+        flush_read_version_coalescer_pending(lane_key, disk, *opts, pending).await;
+    }
+
+    rx.await
+        .unwrap_or_else(|_| Err(DiskError::other("coalesced read_version response channel closed")))
+}
+
+async fn flush_read_version_coalescer_lane(lane_key: ReadVersionCoalescerKey, disk: DiskStore, opts: ReadOptions) {
+    let pending = {
+        let mut coalescer = read_version_coalescer().lock().await;
+        coalescer.lanes.remove(&lane_key).unwrap_or_default()
+    };
+    flush_read_version_coalescer_pending(lane_key, disk, opts, pending).await;
+}
+
+async fn flush_read_version_coalescer_pending(
+    lane_key: ReadVersionCoalescerKey,
+    disk: DiskStore,
+    opts: ReadOptions,
+    pending: Vec<CoalescedReadVersionRequest>,
+) {
+    if pending.is_empty() {
+        return;
+    }
+
+    #[cfg(test)]
+    {
+        let mut observed_paths = HashSet::new();
+        for request in &pending {
+            if observed_paths.insert(request.item.path.as_str()) {
+                disk_call_counters::record(&request.item.path, disk_call_counters::KIND_BATCH_READ_VERSION, lane_key.disk);
+            }
+        }
+    }
+
+    let mut senders = Vec::with_capacity(pending.len());
+    let mut items = Vec::with_capacity(pending.len());
+    for request in pending {
+        senders.push(request.tx);
+        items.push(request.item);
+    }
+
+    let expected_items = items.clone();
+    record_read_version_coalescer_event("attempted_batch", items.len());
+    let result =
+        match tokio::time::timeout(get_drive_metadata_timeout(), disk.batch_read_version(BatchReadVersionReq { items, opts }))
+            .await
+        {
+            Ok(result) => result,
+            Err(_) => Err(DiskError::Timeout),
+        };
+    match result {
+        Ok(responses) => {
+            let results = map_batch_read_version_responses(&expected_items, responses);
+            for (tx, result) in senders.into_iter().zip(results) {
+                let _ = tx.send(result);
+            }
+        }
+        Err(err) => {
+            let message = err.to_string();
+            for tx in senders {
+                let _ = tx.send(Err(DiskError::other(message.clone())));
+            }
+        }
+    }
+}
+
+fn map_batch_read_version_responses(
+    expected_items: &[BatchReadVersionItem],
+    responses: Vec<BatchReadVersionResp>,
+) -> Vec<crate::disk::error::Result<FileInfo>> {
+    let mut results = (0..expected_items.len())
+        .map(|_| Err(DiskError::other("coalesced read_version response missing")))
+        .collect::<Vec<_>>();
+    let mut seen = vec![false; expected_items.len()];
+    for response in responses {
+        let Some(expected) = expected_items.get(response.index) else {
+            continue;
+        };
+        let Some(slot) = results.get_mut(response.index) else {
+            continue;
+        };
+        if seen[response.index] {
+            *slot = Err(DiskError::other("coalesced read_version response duplicate index"));
+            continue;
+        }
+        seen[response.index] = true;
+        if response.path != expected.path || response.version_id != expected.version_id {
+            *slot = Err(DiskError::other("coalesced read_version response identity mismatch"));
+        } else {
+            *slot = if response.success {
+                Ok(response.file_info)
+            } else {
+                Err(batch_read_version_response_error(response.error_code, response.error))
+            };
+        }
+    }
+    results
+}
+
+fn batch_read_version_response_error(error_code: u32, error: String) -> DiskError {
+    match DiskError::from_u32(error_code) {
+        Some(DiskError::Io(_)) | None => DiskError::other(error),
+        Some(error) => error,
+    }
+}
+
 pub(in crate::set_disk) fn bounded_metadata_fanout_order(
    bucket: &str,
    object: &str,
@@ -133,11 +370,15 @@ pub(in crate::set_disk) fn bounded_metadata_fanout_order(
    order
 }
 use tokio::io::{AsyncRead, ReadBuf};
-use tokio::sync::RwLock;
+use tokio::sync::{Mutex, RwLock, oneshot};
 use tokio::task::JoinSet;

 pub(in crate::set_disk) const EVENT_SET_DISK_READ: &str = "set_disk_read";
 pub(in crate::set_disk) const ENV_RUSTFS_GET_DATA_BLOCKS_FIRST_READER_SETUP: &str = "RUSTFS_GET_DATA_BLOCKS_FIRST_READER_SETUP";
+const ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE: &str = "RUSTFS_GET_METADATA_READ_VERSION_COALESCE";
+const ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS: &str = "RUSTFS_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS";
+const DEFAULT_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS: u64 = 200;
+const METRIC_GET_METADATA_READ_VERSION_COALESCER_TOTAL: &str = "rustfs_get_metadata_read_version_coalescer_total";
 pub(in crate::set_disk) const ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE: &str = "RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE";
 /// Default reader-setup strategy for the GET read path (rustfs/backlog#1215,
 /// #1159, #923).
@@ -2356,6 +2597,7 @@ impl SetDisks {
            false,
            true,
            0,
+            false,
        )
        .await?;
        Ok((ress, errors))
@@ -2386,6 +2628,36 @@ impl SetDisks {
            true,
            caller_allows_early_stop,
            default_parity_count,
+            false,
+        )
+        .await
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    pub(in crate::set_disk) async fn read_all_fileinfo_observed_for_get_object(
+        disks: &[Option<DiskStore>],
+        org_bucket: &str,
+        bucket: &str,
+        object: &str,
+        version_id: &str,
+        read_data: bool,
+        incl_free_versions: bool,
+        caller_allows_early_stop: bool,
+        default_parity_count: usize,
+    ) -> disk::error::Result<(Vec<FileInfo>, Vec<Option<DiskError>>, MetadataFanoutDiagnostics)> {
+        Self::read_all_fileinfo_inner(
+            disks,
+            org_bucket,
+            bucket,
+            object,
+            version_id,
+            read_data,
+            false,
+            incl_free_versions,
+            true,
+            caller_allows_early_stop,
+            default_parity_count,
+            true,
        )
        .await
    }
@@ -2408,6 +2680,7 @@ impl SetDisks {
        // subset would fail write quorum (backlog#872 regression).
        caller_allows_early_stop: bool,
        default_parity_count: usize,
+        allow_coalescing: bool,
    ) -> disk::error::Result<(Vec<FileInfo>, Vec<Option<DiskError>>, MetadataFanoutDiagnostics)> {
        let early_stop_enabled =
            caller_allows_early_stop && observe && (is_get_metadata_early_stop_enabled() || is_version_early_stop_enabled());
@@ -2424,6 +2697,7 @@ impl SetDisks {
                healing,
                incl_free_versions,
                default_parity_count,
+                allow_coalescing,
            )
            .await;
        }
@@ -2446,6 +2720,7 @@ impl SetDisks {
            healing,
            incl_free_versions,
            observe,
+            allow_coalescing,
        )
        .await
    }
@@ -2461,6 +2736,7 @@ impl SetDisks {
        healing: bool,
        incl_free_versions: bool,
        observe: bool,
+        allow_coalescing: bool,
    ) -> disk::error::Result<(Vec<FileInfo>, Vec<Option<DiskError>>, MetadataFanoutDiagnostics)> {
        let fanout_start = observe.then(Instant::now);
        let mut ress = Vec::with_capacity(disks.len());
@@ -2492,7 +2768,7 @@ impl SetDisks {
                    if let Some(delay) = slowtail_fault.as_ref().and_then(|fault| fault.delay_for_disk(disk_index)) {
                        tokio::time::sleep(delay).await;
                    }
-                    disk.read_version(&org_bucket, &bucket, &object, &version_id, &task_opts)
+                    read_version_via_coalescer(disk, &org_bucket, &bucket, &object, &version_id, &task_opts, allow_coalescing)
                        .await
                } else {
                    Err(DiskError::DiskNotFound)
@@ -2559,6 +2835,7 @@ impl SetDisks {
        healing: bool,
        incl_free_versions: bool,
        default_parity_count: usize,
+        allow_coalescing: bool,
    ) -> disk::error::Result<(Vec<FileInfo>, Vec<Option<DiskError>>, MetadataFanoutDiagnostics)> {
        let fanout_start = Instant::now();
        let mut ress = vec![FileInfo::default(); disks.len()];
@@ -2607,7 +2884,7 @@ impl SetDisks {
                        if let Some(delay) = slowtail_fault.as_ref().and_then(|fault| fault.delay_for_disk(index)) {
                            tokio::time::sleep(delay).await;
                        }
-                        disk.read_version(&org_bucket, &bucket, &object, &version_id, &task_opts)
+                        read_version_via_coalescer(disk, &org_bucket, &bucket, &object, &version_id, &task_opts, allow_coalescing)
                            .await
                    } else {
                        Err(DiskError::DiskNotFound)
@@ -5737,6 +6014,7 @@ pub(crate) mod disk_call_counters {

    /// Kind label for the per-disk `read_version` metadata RPC.
    pub const KIND_READ_VERSION: &str = "read_version";
+    pub const KIND_BATCH_READ_VERSION: &str = "batch_read_version";

    /// Registry key: (object, kind, disk_index).
    type CountKey = (String, String, usize);
@@ -6460,6 +6738,286 @@ mod tests {
        drop(dirs);
    }

+    #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+    async fn metadata_read_version_coalescer_bypasses_local_disks() {
+        const DISKS: usize = 4;
+        let bucket = "coalesced-read-version-local-bypass-bucket";
+        let object_a = "coalesced-local-object-a";
+        let object_b = "coalesced-local-object-b";
+        let (dirs, disks) = call_counter_local_disks(bucket, DISKS).await;
+        install_metadata_fanout_fileinfo(&disks, bucket, object_a, None).await;
+        install_metadata_fanout_fileinfo(&disks, bucket, object_b, None).await;
+
+        temp_env::async_with_vars(
+            [
+                (ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE, Some("auto")),
+                (ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE_DELAY_MICROS, Some("5000")),
+            ],
+            async {
+                let calls = disk_call_counters::observe(object_a);
+                let disks_a = disks.clone();
+                let disks_b = disks.clone();
+                let read_a = tokio::spawn(async move {
+                    SetDisks::read_all_fileinfo_observed_for_get_object(
+                        &disks_a, "", bucket, object_a, "", false, false, false, 2,
+                    )
+                    .await
+                    .map(|(file_infos, errors, _)| (file_infos, errors))
+                });
+                tokio::task::yield_now().await;
+                let read_b = tokio::spawn(async move {
+                    SetDisks::read_all_fileinfo_observed_for_get_object(
+                        &disks_b, "", bucket, object_b, "", false, false, false, 2,
+                    )
+                    .await
+                    .map(|(file_infos, errors, _)| (file_infos, errors))
+                });
+
+                let (metadata_a, errs_a) = read_a
+                    .await
+                    .expect("first read task should not panic")
+                    .expect("first coalesced read should resolve");
+                let (metadata_b, errs_b) = read_b
+                    .await
+                    .expect("second read task should not panic")
+                    .expect("second coalesced read should resolve");
+
+                assert_eq!(metadata_a.iter().filter(|fi| fi.name == object_a).count(), DISKS);
+                assert_eq!(metadata_b.iter().filter(|fi| fi.name == object_b).count(), DISKS);
+                assert!(errs_a.iter().all(Option::is_none));
+                assert!(errs_b.iter().all(Option::is_none));
+                assert_eq!(
+                    calls.total(disk_call_counters::KIND_READ_VERSION),
+                    DISKS as u64,
+                    "local disks still execute the ordinary per-disk read_version path"
+                );
+                assert_eq!(
+                    calls.total(disk_call_counters::KIND_BATCH_READ_VERSION),
+                    0,
+                    "GET coalescing targets internode RPC count only and must not batch local disk reads"
+                );
+            },
+        )
+        .await;
+
+        drop(dirs);
+    }
+
+    #[tokio::test]
+    async fn metadata_read_version_coalescer_requires_get_object_intent() {
+        const DISKS: usize = 4;
+        let bucket = "coalesced-read-version-default-bypass-bucket";
+        let object = "default-bypass-object";
+        let (dirs, disks) = call_counter_local_disks(bucket, DISKS).await;
+        install_metadata_fanout_fileinfo(&disks, bucket, object, None).await;
+
+        temp_env::async_with_vars([(ENV_RUSTFS_GET_METADATA_READ_VERSION_COALESCE, Some("auto"))], async {
+            let calls = disk_call_counters::observe(object);
+            let (metadata, errs) = SetDisks::read_all_fileinfo(&disks, "", bucket, object, "", false, false, false)
+                .await
+                .expect("default metadata read should resolve");
+
+            assert_eq!(metadata.iter().filter(|fi| fi.name == object).count(), DISKS);
+            assert!(errs.iter().all(Option::is_none));
+            assert_eq!(calls.total(disk_call_counters::KIND_READ_VERSION), DISKS as u64);
+            assert_eq!(
+                calls.total(disk_call_counters::KIND_BATCH_READ_VERSION),
+                0,
+                "non-GET metadata paths must bypass coalescer even when the env gate is enabled"
+            );
+        })
+        .await;
+
+        drop(dirs);
+    }
+
+    #[test]
+    fn batch_read_version_response_mapping_preserves_index_and_errors() {
+        let expected_items = vec![
+            BatchReadVersionItem {
+                org_volume: String::new(),
+                volume: "bucket".to_string(),
+                path: "object-a".to_string(),
+                version_id: "v-a".to_string(),
+            },
+            BatchReadVersionItem {
+                org_volume: String::new(),
+                volume: "bucket".to_string(),
+                path: "object-b".to_string(),
+                version_id: "v-b".to_string(),
+            },
+            BatchReadVersionItem {
+                org_volume: String::new(),
+                volume: "bucket".to_string(),
+                path: "object-c".to_string(),
+                version_id: "v-c".to_string(),
+            },
+        ];
+        let ok_file_info = FileInfo {
+            name: "object-a".to_string(),
+            ..Default::default()
+        };
+        let responses = vec![
+            BatchReadVersionResp {
+                index: 2,
+                path: "object-c".to_string(),
+                version_id: "v-c".to_string(),
+                success: false,
+                file_info: FileInfo::default(),
+                error: "disk read failed".to_string(),
+                error_code: 0,
+            },
+            BatchReadVersionResp {
+                index: 0,
+                path: "object-a".to_string(),
+                version_id: "v-a".to_string(),
+                success: true,
+                file_info: ok_file_info,
+                error: String::new(),
+                error_code: 0,
+            },
+        ];
+
+        let mut results = map_batch_read_version_responses(&expected_items, responses).into_iter();
+        let first = results
+            .next()
+            .expect("slot 0 should exist")
+            .expect("slot 0 should map the success response by index");
+        assert_eq!(first.name, "object-a");
+
+        let missing = results
+            .next()
+            .expect("slot 1 should exist")
+            .expect_err("slot 1 should stay missing");
+        assert!(
+            missing.to_string().contains("response missing"),
+            "unexpected missing response error: {missing}"
+        );
+
+        let failed = results
+            .next()
+            .expect("slot 2 should exist")
+            .expect_err("slot 2 should map the response error");
+        assert!(failed.to_string().contains("disk read failed"), "unexpected per-item error: {failed}");
+        assert!(results.next().is_none());
+    }
+
+    #[test]
+    fn batch_read_version_response_mapping_preserves_typed_not_found_errors() {
+        let expected_items = vec![
+            BatchReadVersionItem {
+                org_volume: String::new(),
+                volume: "bucket".to_string(),
+                path: "object-a".to_string(),
+                version_id: "v-a".to_string(),
+            },
+            BatchReadVersionItem {
+                org_volume: String::new(),
+                volume: "bucket".to_string(),
+                path: "object-b".to_string(),
+                version_id: "v-b".to_string(),
+            },
+        ];
+        let results = map_batch_read_version_responses(
+            &expected_items,
+            vec![
+                BatchReadVersionResp {
+                    index: 0,
+                    path: "object-a".to_string(),
+                    version_id: "v-a".to_string(),
+                    success: false,
+                    file_info: FileInfo::default(),
+                    error: DiskError::FileNotFound.to_string(),
+                    error_code: DiskError::FileNotFound.to_u32(),
+                },
+                BatchReadVersionResp {
+                    index: 1,
+                    path: "object-b".to_string(),
+                    version_id: "v-b".to_string(),
+                    success: false,
+                    file_info: FileInfo::default(),
+                    error: DiskError::FileVersionNotFound.to_string(),
+                    error_code: DiskError::FileVersionNotFound.to_u32(),
+                },
+            ],
+        );
+
+        assert!(matches!(results.first().expect("slot 0 should exist"), Err(DiskError::FileNotFound)));
+        assert!(matches!(
+            results.get(1).expect("slot 1 should exist"),
+            Err(DiskError::FileVersionNotFound)
+        ));
+    }
+
+    #[test]
+    fn batch_read_version_response_mapping_rejects_identity_mismatch_and_duplicate_index() {
+        let expected_items = vec![BatchReadVersionItem {
+            org_volume: String::new(),
+            volume: "bucket".to_string(),
+            path: "object-a".to_string(),
+            version_id: "v-a".to_string(),
+        }];
+        let mismatched = map_batch_read_version_responses(
+            &expected_items,
+            vec![BatchReadVersionResp {
+                index: 0,
+                path: "object-b".to_string(),
+                version_id: "v-a".to_string(),
+                success: true,
+                file_info: FileInfo {
+                    name: "object-b".to_string(),
+                    ..Default::default()
+                },
+                error: String::new(),
+                error_code: 0,
+            }],
+        )
+        .pop()
+        .expect("slot 0 should exist")
+        .expect_err("identity mismatch should fail closed");
+        assert!(
+            mismatched.to_string().contains("identity mismatch"),
+            "unexpected mismatch error: {mismatched}"
+        );
+
+        let duplicate = map_batch_read_version_responses(
+            &expected_items,
+            vec![
+                BatchReadVersionResp {
+                    index: 0,
+                    path: "object-a".to_string(),
+                    version_id: "v-a".to_string(),
+                    success: true,
+                    file_info: FileInfo {
+                        name: "object-a".to_string(),
+                        ..Default::default()
+                    },
+                    error: String::new(),
+                    error_code: 0,
+                },
+                BatchReadVersionResp {
+                    index: 0,
+                    path: "object-a".to_string(),
+                    version_id: "v-a".to_string(),
+                    success: true,
+                    file_info: FileInfo {
+                        name: "object-a".to_string(),
+                        ..Default::default()
+                    },
+                    error: String::new(),
+                    error_code: 0,
+                },
+            ],
+        )
+        .pop()
+        .expect("slot 0 should exist")
+        .expect_err("duplicate response index should fail closed");
+        assert!(
+            duplicate.to_string().contains("duplicate index"),
+            "unexpected duplicate error: {duplicate}"
+        );
+    }
+
    /// Isolation guard: unobserved objects record nothing (so parallel tests do
    /// not inflate one another), and a scope clears its own counts on drop.
    #[tokio::test]
@@ -735,8 +735,12 @@ pub(crate) use core::io_primitives::disk_call_counters;
 mod ctx;
 mod metadata;
 mod ops;
+#[cfg(test)]
+pub(crate) use ops::multipart::NewMultipartUploadCommitObservation;
 #[cfg(any(test, feature = "test-util"))]
 pub use ops::multipart::{MultipartCommitBarrier, MultipartCommitPause};
+#[cfg(test)]
+pub(crate) use ops::object::DeleteObjectCommitBarrier;
 #[cfg(feature = "test-util")]
 pub(crate) use ops::object::TransitionCleanupStoreBarrier as SetDiskTransitionCleanupStoreBarrier;
 pub(crate) use ops::object::body_cache_plaintext_len;
@@ -3025,6 +3029,16 @@ pub struct SetDisks {
    storage_class_config_override: Arc<std::sync::RwLock<Option<Arc<storageclass::Config>>>>,
 }

+// DistributedLock sends the raw ObjectKey to its clients; LockRegistry clones
+// each endpoint's canonical Arc, so an exact Arc set identifies the lock domain.
+pub(crate) fn same_distributed_lock_domain(left: &[Arc<dyn LockClient>], right: &[Arc<dyn LockClient>]) -> bool {
+    left.iter()
+        .all(|left_client| right.iter().any(|right_client| Arc::ptr_eq(left_client, right_client)))
+        && right
+            .iter()
+            .all(|right_client| left.iter().any(|left_client| Arc::ptr_eq(left_client, right_client)))
+}
+
 const ERASURE_CACHE_MAX_ENTRIES: usize = 32;

 #[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
@@ -3600,6 +3614,15 @@ impl SetDisks {
        &self.ctx
    }

+    /// Whether both sets' namespace-lock implementations cover the same object key.
+    pub(crate) async fn shares_namespace_lock_domain(&self, other: &Self) -> bool {
+        match (self.ctx.is_dist_erasure().await, other.ctx.is_dist_erasure().await) {
+            (false, false) => Arc::ptr_eq(&self.local_lock_manager, &other.local_lock_manager),
+            (true, true) => same_distributed_lock_domain(&self.lockers, &other.lockers),
+            _ => false,
+        }
+    }
+
    /// The lock manager this set actually uses (test-only; Phase 5 Slice 3).
    #[cfg(test)]
    pub(crate) fn local_lock_manager_for_test(&self) -> &Arc<rustfs_lock::GlobalLockManager> {
@@ -4584,11 +4607,11 @@ fn should_preserve_delete_replication_state(opts: &ObjectOptions) -> bool {
 }

 fn should_force_delete_marker_for_missing_version(opts: &ObjectOptions) -> bool {
-    opts.delete_marker || (opts.versioned && opts.version_id.is_none() && !opts.data_movement)
+    opts.delete_marker || ((opts.versioned || opts.version_suspended) && opts.version_id.is_none() && !opts.data_movement)
 }

 fn resolve_delete_version_state(opts: &ObjectOptions, goi: &ObjectInfo, version_found: bool) -> (bool, bool) {
-    let mut mark_delete = goi.version_id.is_some() || (opts.versioned && opts.version_id.is_none());
+    let mut mark_delete = goi.version_id.is_some() || ((opts.versioned || opts.version_suspended) && opts.version_id.is_none());
    let mut delete_marker = opts.versioned;

    if opts.version_id.is_some() {
@@ -32,6 +32,8 @@ use crate::crash_inject::{self, CrashPoint};
 use crate::multipart_listing::paginate_multipart_listing;
 use futures::{StreamExt, stream};
 use std::future::Future;
+#[cfg(test)]
+use std::sync::atomic::AtomicBool;
 #[cfg(any(test, feature = "test-util"))]
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::time::Duration;
@@ -65,6 +67,7 @@ impl StaleMultipartCleanupGuard {
 #[cfg(any(test, feature = "test-util"))]
 #[derive(Clone, Copy, PartialEq, Eq)]
 pub enum MultipartCommitPause {
+    NewUploadBeforeLockLost,
    PutPartBeforeLockAcquire,
    PutPartBeforeLockLost,
    PutPartAfterRename,
@@ -156,6 +159,72 @@ impl Drop for MultipartCommitBarrier {
    }
 }

+#[cfg(test)]
+struct NewMultipartUploadCommitObservationState {
+    bucket: String,
+    object: String,
+    committed: AtomicBool,
+}
+
+#[cfg(test)]
+pub(crate) struct NewMultipartUploadCommitObservation {
+    state: Arc<NewMultipartUploadCommitObservationState>,
+}
+
+#[cfg(test)]
+static NEW_MULTIPART_UPLOAD_COMMIT_OBSERVATION: std::sync::OnceLock<
+    std::sync::Mutex<Option<Arc<NewMultipartUploadCommitObservationState>>>,
+> = std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl NewMultipartUploadCommitObservation {
+    pub(crate) fn install(bucket: &str, object: &str) -> Self {
+        let state = Arc::new(NewMultipartUploadCommitObservationState {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            committed: AtomicBool::new(false),
+        });
+        let mut slot = NEW_MULTIPART_UPLOAD_COMMIT_OBSERVATION
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("new multipart upload commit observation mutex should not poison");
+        assert!(slot.is_none(), "new multipart upload commit observation must be unique");
+        *slot = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) fn committed(&self) -> bool {
+        self.state.committed.load(Ordering::Acquire)
+    }
+}
+
+#[cfg(test)]
+impl Drop for NewMultipartUploadCommitObservation {
+    fn drop(&mut self) {
+        let mut slot = NEW_MULTIPART_UPLOAD_COMMIT_OBSERVATION
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("new multipart upload commit observation mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+fn observe_new_multipart_upload_commit(bucket: &str, object: &str) {
+    let state = NEW_MULTIPART_UPLOAD_COMMIT_OBSERVATION
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("new multipart upload commit observation mutex should not poison")
+        .as_ref()
+        .filter(|state| state.bucket == bucket && state.object == object)
+        .cloned();
+    if let Some(state) = state {
+        state.committed.store(true, Ordering::Release);
+    }
+}
+
 #[cfg(any(test, feature = "test-util"))]
 async fn pause_multipart_commit(bucket: &str, object: &str, pause: MultipartCommitPause) {
    let barrier = {
@@ -1615,6 +1684,30 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {

        let upload_path = Self::get_multipart_upload_dir(bucket, object, upload_uuid.as_str(), opts.data_movement);

+        #[cfg(any(test, feature = "test-util"))]
+        pause_multipart_commit(bucket, object, MultipartCommitPause::NewUploadBeforeLockLost).await;
+        if _object_lock_guard.as_ref().is_some_and(|guard| guard.is_lock_lost()) {
+            return Err(StorageError::NamespaceLockQuorumUnavailable {
+                mode: "new_multipart_upload_commit",
+                bucket: bucket.to_string(),
+                object: object.to_string(),
+                required: 1,
+                achieved: 0,
+            });
+        }
+        if opts
+            .namespace_lock_fence
+            .as_ref()
+            .is_some_and(NamespaceLockFence::is_lock_lost)
+        {
+            return Err(StorageError::NamespaceLockQuorumUnavailable {
+                mode: "new_multipart_upload_outer_lock",
+                bucket: bucket.to_string(),
+                object: object.to_string(),
+                required: 1,
+                achieved: 0,
+            });
+        }
        ensure_multipart_bucket_lifecycle_lock_held(bucket, object, opts)?;
        Self::write_unique_file_info(
            &shuffle_disks,
@@ -1626,6 +1719,8 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
        )
        .await
        .map_err(|e| to_object_err(e.into(), vec![bucket, object]))?;
+        #[cfg(test)]
+        observe_new_multipart_upload_commit(bucket, object);

        // evalDisks

@@ -1294,7 +1294,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
            (prepared.snapshot, prepared.object_info)
        } else {
            match self
-                .get_object_fileinfo(
+                .get_object_fileinfo_for_get_object_reader(
                    bucket,
                    object,
                    opts,
@@ -2123,11 +2123,11 @@ impl SetDisks {
            let erasure = Arc::new(erasure_from_file_info(&fi, false)?);

            let put_object_size = known_put_object_storage_size(data.size());
-            let is_inline_buffer =
-                storage_class_config.should_inline(erasure.shard_file_size(put_object_size), erasure.data_shards, opts.versioned);
+            let shard_file_size_raw = erasure.shard_file_size(put_object_size);
+            let is_inline_buffer = storage_class_config.should_inline(shard_file_size_raw, erasure.data_shards, opts.versioned);

            let collect_stage_timing = rustfs_io_metrics::put_stage_metrics_enabled() || issue3031_diag_enabled();
-            let shard_file_size = erasure.shard_file_size(put_object_size);
+            let shard_file_size = shard_file_size_raw;
            let shard_size = erasure.shard_size();
            let write_path = classify_put_write_path(is_inline_buffer, put_object_size, fi.erasure.block_size);
            let direct_inline_commit = matches!(write_path, SmallWritePath::Inline);
@@ -2484,6 +2484,7 @@ impl SetDisks {
                        })
                        .await?,
                    );
+                    notify_put_object_commit_namespace_acquired(bucket, object);
                }
                #[cfg(not(any(test, feature = "test-util")))]
                {
@@ -4631,6 +4632,7 @@ struct PutObjectCommitBarrierState {
    arrived: tokio::sync::Notify,
    release: tokio::sync::Notify,
    namespace_pending: tokio::sync::Notify,
+    namespace_acquired: std::sync::atomic::AtomicBool,
 }

 #[cfg(any(test, feature = "test-util"))]
@@ -4652,6 +4654,7 @@ impl PutObjectCommitBarrier {
            arrived: tokio::sync::Notify::new(),
            release: tokio::sync::Notify::new(),
            namespace_pending: tokio::sync::Notify::new(),
+            namespace_acquired: std::sync::atomic::AtomicBool::new(false),
        });
        let mut slot = PUT_OBJECT_COMMIT_BARRIER
            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
@@ -4686,6 +4689,10 @@ impl PutObjectCommitBarrier {
            .await
            .expect("put object should wait for the namespace lock after leaving the commit barrier");
    }
+
+    pub fn namespace_acquired(&self) -> bool {
+        self.state.namespace_acquired.load(std::sync::atomic::Ordering::Acquire)
+    }
 }

 #[cfg(any(test, feature = "test-util"))]
@@ -4742,6 +4749,22 @@ fn notify_put_object_commit_namespace_pending(bucket: &str, object: &str) {
    }
 }

+#[cfg(any(test, feature = "test-util"))]
+fn notify_put_object_commit_namespace_acquired(bucket: &str, object: &str) {
+    let barrier = PUT_OBJECT_COMMIT_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(Vec::new()))
+        .lock()
+        .expect("put object commit barrier mutex should not poison")
+        .iter()
+        .find(|barrier| {
+            barrier.bucket == bucket && barrier.object == object && barrier.pause == PutObjectCommitPause::BeforeNamespace
+        })
+        .cloned();
+    if let Some(barrier) = barrier {
+        barrier.namespace_acquired.store(true, std::sync::atomic::Ordering::Release);
+    }
+}
+
 #[cfg(test)]
 struct DeleteObjectCommitBarrierState {
    bucket: String,
@@ -4751,7 +4774,7 @@ struct DeleteObjectCommitBarrierState {
 }

 #[cfg(test)]
-struct DeleteObjectCommitBarrier {
+pub(crate) struct DeleteObjectCommitBarrier {
    state: Arc<DeleteObjectCommitBarrierState>,
 }

@@ -4761,7 +4784,7 @@ static DELETE_OBJECT_COMMIT_BARRIER: std::sync::OnceLock<std::sync::Mutex<Option

 #[cfg(test)]
 impl DeleteObjectCommitBarrier {
-    fn install(bucket: &str, object: &str) -> Self {
+    pub(crate) fn install(bucket: &str, object: &str) -> Self {
        let state = Arc::new(DeleteObjectCommitBarrierState {
            bucket: bucket.to_string(),
            object: object.to_string(),
@@ -4777,13 +4800,13 @@ impl DeleteObjectCommitBarrier {
        Self { state }
    }

-    async fn wait_until_paused(&self) {
+    pub(crate) async fn wait_until_paused(&self) {
        tokio::time::timeout(Duration::from_secs(30), self.state.arrived.notified())
            .await
            .expect("delete object should reach the deterministic commit barrier");
    }

-    fn release(&self) {
+    pub(crate) fn release(&self) {
        self.state.release.notify_one();
    }
 }
@@ -5905,6 +5928,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
            if dobj.version_id.is_none() && (version_suspended || versioned) {
                vr.mod_time = Some(OffsetDateTime::now_utc());
                vr.deleted = true;
+                vr.mark_deleted = true;
                if versioned {
                    vr.version_id = Some(Uuid::new_v4());
                }
@@ -11791,6 +11815,7 @@ mod transition_upload_integrity_tests {
                crate::data_movement::SourceCleanupBucketFence {
                    expected_incarnation_id: None,
                    lifecycle_guard: Some(&bucket_guard),
+                    ..Default::default()
                },
                "test_data_movement",
            )
@@ -259,10 +259,33 @@ impl SetDisks {
        read_data: bool,
        caller_allows_early_stop: bool,
    ) -> Result<GetObjectFileInfo> {
-        self.get_object_fileinfo_gated(bucket, object, opts, read_data, caller_allows_early_stop)
+        self.get_object_fileinfo_gated_inner(bucket, object, opts, read_data, caller_allows_early_stop, false)
            .await
    }

+    #[tracing::instrument(level = "debug", skip(self))]
+    #[hotpath::measure(impl_type = "SetDisks")]
+    pub(super) async fn get_object_fileinfo_for_get_object_reader(
+        &self,
+        bucket: &str,
+        object: &str,
+        opts: &ObjectOptions,
+        read_data: bool,
+        caller_allows_early_stop: bool,
+    ) -> Result<GetObjectFileInfo> {
+        let allow_read_version_coalescing = !crate::bucket::utils::is_meta_bucketname(bucket)
+            && crate::runtime::global::get_metadata_read_version_coalescing_service_ready();
+        self.get_object_fileinfo_gated_inner(
+            bucket,
+            object,
+            opts,
+            read_data,
+            caller_allows_early_stop,
+            allow_read_version_coalescing,
+        )
+        .await
+    }
+
    /// Like `get_object_fileinfo`, but `allow_early_stop=false` forces the full
    /// quorum fanout. Read-before-write callers (object tagging) must use this:
    /// the returned online-disk set is the write target, and the early-stop
@@ -275,6 +298,20 @@ impl SetDisks {
        opts: &ObjectOptions,
        read_data: bool,
        allow_early_stop: bool,
+    ) -> Result<GetObjectFileInfo> {
+        self.get_object_fileinfo_gated_inner(bucket, object, opts, read_data, allow_early_stop, false)
+            .await
+    }
+
+    #[allow(clippy::too_many_arguments)]
+    async fn get_object_fileinfo_gated_inner(
+        &self,
+        bucket: &str,
+        object: &str,
+        opts: &ObjectOptions,
+        read_data: bool,
+        allow_early_stop: bool,
+        allow_read_version_coalescing: bool,
    ) -> Result<GetObjectFileInfo> {
        let vid = opts.version_id.clone().unwrap_or_default();
        let stage_metrics_enabled = rustfs_io_metrics::get_stage_metrics_enabled();
@@ -337,19 +374,34 @@ impl SetDisks {
        // read_all_fileinfo_observed (see read_all_fileinfo_early_stop in
        // core/io_primitives.rs); unsafe requests and callers that opt out
        // (allow_early_stop=false) fall back to full-wait.
-        let (mut parts_metadata, errs, metadata_fanout_diagnostics) = Self::read_all_fileinfo_observed(
-            &disks,
-            "",
-            bucket,
-            object,
-            vid.as_str(),
-            read_data,
-            false,
-            opts.incl_free_versions,
-            allow_early_stop,
-            self.default_parity_count,
-        )
-        .await?;
+        let (mut parts_metadata, errs, metadata_fanout_diagnostics) = if allow_read_version_coalescing {
+            Self::read_all_fileinfo_observed_for_get_object(
+                &disks,
+                "",
+                bucket,
+                object,
+                vid.as_str(),
+                read_data,
+                opts.incl_free_versions,
+                allow_early_stop,
+                self.default_parity_count,
+            )
+            .await?
+        } else {
+            Self::read_all_fileinfo_observed(
+                &disks,
+                "",
+                bucket,
+                object,
+                vid.as_str(),
+                read_data,
+                false,
+                opts.incl_free_versions,
+                allow_early_stop,
+                self.default_parity_count,
+            )
+            .await?
+        };
        let metadata_metrics_path = if crate::bucket::utils::is_meta_bucketname(bucket) {
            GET_OBJECT_PATH_INTERNAL_META
        } else {
@@ -13,7 +13,12 @@
 // limitations under the License.

 use super::*;
+use crate::core::pools::POOL_META_NAME;
+use crate::services::rebalance::{REBAL_META_NAME, RebalStatus};
+use crate::set_disk::get_lock_acquire_timeout;
 use crate::storage_api_contracts::heal::HealOperations as _;
+use crate::storage_api_contracts::namespace::NamespaceLocking as _;
+use rustfs_lock::NamespaceLockGuard;
 use tracing::trace;

 const LOG_COMPONENT_ECSTORE: &str = "ecstore";
@@ -30,7 +35,119 @@ fn invalid_heal_pool_index(pool_idx: usize, pool_count: usize) -> Error {
    )
 }

+#[derive(Debug, Clone, Copy)]
+enum HealFormatPoolSkip {
+    Completed,
+    Retryable,
+}
+
+fn classify_heal_format_pool(
+    pool_idx: usize,
+    pool_cmd_line: &str,
+    pool_meta: &PoolMeta,
+    rebalance_meta: Option<&RebalanceMeta>,
+) -> Option<HealFormatPoolSkip> {
+    let Some(pool) = pool_meta.pools.get(pool_idx) else {
+        return Some(HealFormatPoolSkip::Retryable);
+    };
+
+    if pool.id != pool_idx || pool_cmd_line.is_empty() || pool.cmd_line.is_empty() || pool.cmd_line != pool_cmd_line {
+        return Some(HealFormatPoolSkip::Retryable);
+    }
+
+    if let Some(decommission) = pool.decommission.as_ref() {
+        if decommission.complete {
+            return Some(HealFormatPoolSkip::Completed);
+        }
+        if decommission.failed || decommission.canceled || decommission.queued || pool_meta.is_suspended(pool_idx) {
+            return Some(HealFormatPoolSkip::Retryable);
+        }
+    }
+
+    if let Some(meta) = rebalance_meta {
+        let Some(pool_stats) = meta.pool_stats.get(pool_idx) else {
+            return Some(HealFormatPoolSkip::Retryable);
+        };
+        if pool_stats.info.stopping || (pool_stats.participating && pool_stats.info.status == RebalStatus::Started) {
+            return Some(HealFormatPoolSkip::Retryable);
+        }
+    }
+
+    None
+}
+
+fn heal_format_pool_skip_error(skip: HealFormatPoolSkip) -> Error {
+    match skip {
+        HealFormatPoolSkip::Completed => StorageError::NoHealRequired,
+        HealFormatPoolSkip::Retryable => StorageError::SlowDown,
+    }
+}
+
+fn heal_format_fence_lost_error() -> Error {
+    StorageError::SlowDown
+}
+
 impl ECStore {
+    async fn acquire_heal_format_fence(
+        &self,
+    ) -> Result<(NamespaceLockGuard, NamespaceLockGuard, PoolMeta, Option<RebalanceMeta>)> {
+        let metadata_pool = self
+            .pools
+            .first()
+            .cloned()
+            .ok_or_else(|| Error::other("heal format requires at least one storage pool"))?;
+
+        // Metadata fence order is part of the decommission/rebalance protocol:
+        // pool.bin must always be acquired before rebalance.bin.
+        let pool_lock = metadata_pool.new_ns_lock(RUSTFS_META_BUCKET, POOL_META_NAME).await?;
+        let pool_guard = pool_lock.get_write_lock(get_lock_acquire_timeout()).await?;
+        let rebalance_lock = metadata_pool.new_ns_lock(RUSTFS_META_BUCKET, REBAL_META_NAME).await?;
+        let rebalance_guard = rebalance_lock.get_write_lock(get_lock_acquire_timeout()).await?;
+
+        if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        let mut pool_meta = PoolMeta::default();
+        pool_meta.load_no_lock(metadata_pool.clone()).await?;
+        if pool_meta.pools.len() != self.pools.len()
+            || pool_meta.pools.iter().enumerate().any(|(pool_idx, pool)| {
+                pool.id != pool_idx || pool.cmd_line.is_empty() || pool.cmd_line != self.pools[pool_idx].endpoints.cmd_line
+            })
+        {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        let mut rebalance_meta = RebalanceMeta::new();
+        let rebalance_meta = match rebalance_meta
+            .load_with_opts(
+                metadata_pool,
+                ObjectOptions {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+        {
+            Ok(()) => Some(rebalance_meta),
+            Err(Error::ConfigNotFound) => None,
+            Err(err) => return Err(err),
+        };
+
+        if rebalance_meta
+            .as_ref()
+            .is_some_and(|meta| meta.pool_stats.len() != self.pools.len())
+        {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+            return Err(heal_format_fence_lost_error());
+        }
+
+        Ok((pool_guard, rebalance_guard, pool_meta, rebalance_meta))
+    }
+
    fn get_pools_for_heal_object(&self, opts: &HealOpts) -> Result<Vec<Arc<Sets>>> {
        match opts.pool {
            Some(pool_idx) => Ok(vec![
@@ -52,9 +169,26 @@ impl ECStore {
        };

        let mut count_no_heal = 0;
+        let mut count_completed = 0;
        let mut first_error = None;
-        for pool in self.pools.iter() {
-            let (mut result, err) = pool.heal_format(dry_run).await?;
+        for (pool_idx, pool) in self.pools.iter().enumerate() {
+            let (pool_guard, rebalance_guard, pool_meta, rebalance_meta) = self.acquire_heal_format_fence().await?;
+            if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+                first_error.get_or_insert(heal_format_fence_lost_error());
+                break;
+            }
+            if let Some(skip) = classify_heal_format_pool(pool_idx, &pool.endpoints.cmd_line, &pool_meta, rebalance_meta.as_ref())
+            {
+                if matches!(skip, HealFormatPoolSkip::Completed) {
+                    count_completed += 1;
+                } else {
+                    first_error.get_or_insert(heal_format_pool_skip_error(skip));
+                }
+                continue;
+            }
+
+            let fence_lost = || pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost();
+            let (mut result, err) = pool.heal_format_with_fence(dry_run, fence_lost).await?;
            if let Some(err) = err {
                match err {
                    StorageError::NoHealRequired => {
@@ -69,11 +203,18 @@ impl ECStore {
            r.set_count += result.set_count;
            r.before.drives.append(&mut result.before.drives);
            r.after.drives.append(&mut result.after.drives);
+
+            // A lease can be lost after the final write; fail closed before
+            // reporting the pool as successfully healed.
+            if pool_guard.is_lock_lost() || rebalance_guard.is_lock_lost() {
+                first_error.get_or_insert(heal_format_fence_lost_error());
+                break;
+            }
        }
        if let Some(err) = first_error {
            return Ok((r, Some(err)));
        }
-        if count_no_heal == self.pools.len() {
+        if count_no_heal + count_completed == self.pools.len() {
            info!(
                event = EVENT_HEAL_FORMAT_COMPLETED,
                component = LOG_COMPONENT_ECSTORE,
@@ -302,6 +443,7 @@ mod tests {
    use crate::disk::{DeleteOptions, DiskOption, format::FormatV3, new_disk};
    use crate::layout::endpoints::{EndpointServerPools, Endpoints, PoolEndpoints};
    use crate::runtime::instance::InstanceContext;
+    use crate::services::rebalance::{RebalanceInfo, RebalanceStats};
    use crate::storage_api_contracts::bucket::{BucketOperations, MakeBucketOptions};
    use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations};
    use crate::store::init_format::{load_format_erasure, save_format_file};
@@ -353,6 +495,164 @@ mod tests {
        }
    }

+    fn pool_meta_with_decommission(info: PoolDecommissionInfo) -> PoolMeta {
+        PoolMeta {
+            pools: vec![PoolStatus {
+                id: 0,
+                cmd_line: "pool-0".to_string(),
+                last_update: OffsetDateTime::UNIX_EPOCH,
+                decommission: Some(info),
+            }],
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn heal_format_pool_state_barriers_are_classified() {
+        let active = pool_meta_with_decommission(PoolDecommissionInfo {
+            start_time: Some(OffsetDateTime::UNIX_EPOCH),
+            ..Default::default()
+        });
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &active, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        for info in [
+            PoolDecommissionInfo {
+                failed: true,
+                ..Default::default()
+            },
+            PoolDecommissionInfo {
+                canceled: true,
+                ..Default::default()
+            },
+        ] {
+            assert!(matches!(
+                classify_heal_format_pool(0, "pool-0", &pool_meta_with_decommission(info), None),
+                Some(HealFormatPoolSkip::Retryable)
+            ));
+        }
+
+        let completed = pool_meta_with_decommission(PoolDecommissionInfo {
+            complete: true,
+            ..Default::default()
+        });
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &completed, None),
+            Some(HealFormatPoolSkip::Completed)
+        ));
+    }
+
+    #[test]
+    fn heal_format_pool_rebalance_barriers_and_identity_are_fail_closed() {
+        let identity_meta = pool_meta_with_decommission(PoolDecommissionInfo::default());
+        let rebalance = RebalanceMeta {
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Started,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&rebalance)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let stopping = RebalanceMeta {
+            pool_stats: vec![RebalanceStats {
+                info: RebalanceInfo {
+                    stopping: true,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopping)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let identity = pool_meta_with_decommission(PoolDecommissionInfo::default());
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-new", &identity, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let identity_without_decommission = PoolMeta {
+            pools: vec![PoolStatus {
+                id: 0,
+                cmd_line: "pool-0".to_string(),
+                last_update: OffsetDateTime::UNIX_EPOCH,
+                decommission: None,
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-new", &identity_without_decommission, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        assert!(matches!(
+            classify_heal_format_pool(0, "", &identity_meta, None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &PoolMeta::default(), None),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+
+        let stopped = RebalanceMeta {
+            stopped_at: Some(OffsetDateTime::UNIX_EPOCH),
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Stopped,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopped)).is_none());
+
+        let stopping_after_stop = RebalanceMeta {
+            stopped_at: Some(OffsetDateTime::UNIX_EPOCH),
+            pool_stats: vec![RebalanceStats {
+                participating: true,
+                info: RebalanceInfo {
+                    status: RebalStatus::Started,
+                    stopping: true,
+                    ..Default::default()
+                },
+                ..Default::default()
+            }],
+            ..Default::default()
+        };
+        assert!(matches!(
+            classify_heal_format_pool(0, "pool-0", &identity_meta, Some(&stopping_after_stop)),
+            Some(HealFormatPoolSkip::Retryable)
+        ));
+    }
+
+    #[test]
+    fn skipped_heal_format_pool_is_never_reported_as_success() {
+        assert!(matches!(
+            heal_format_pool_skip_error(HealFormatPoolSkip::Retryable),
+            StorageError::SlowDown
+        ));
+        assert!(matches!(
+            heal_format_pool_skip_error(HealFormatPoolSkip::Completed),
+            StorageError::NoHealRequired
+        ));
+    }
+
    async fn multi_pool_heal_store() -> (tempfile::TempDir, Arc<ECStore>, CancellationToken) {
        let temp_dir = tempfile::tempdir().expect("multi-pool heal test directory should be created");
        let mut pool_endpoints = Vec::new();
@@ -889,6 +1189,18 @@ mod tests {
            bucket_fence_registry: std::sync::Arc::default(),
        };

+        let err = store
+            .handle_heal_format(false)
+            .await
+            .expect_err("missing pool metadata must fail closed before format writes");
+        assert!(matches!(err, StorageError::SlowDown));
+
+        let pool_meta = PoolMeta::new(&store.pools, &PoolMeta::default());
+        pool_meta
+            .save(store.pools.clone())
+            .await
+            .expect("pool metadata should be persisted before format heal");
+
        let (result, err) = store
            .handle_heal_format(false)
            .await
@@ -902,5 +1214,22 @@ mod tests {
            .await
            .expect("the later pool should be healed despite the first pool error");
        assert_eq!(healed.erasure.this, recoverable_format.erasure.sets[0][2]);
+
+        let mut completed_meta = PoolMeta::new(&store.pools, &PoolMeta::default());
+        for status in &mut completed_meta.pools {
+            status.decommission = Some(PoolDecommissionInfo {
+                complete: true,
+                ..Default::default()
+            });
+        }
+        completed_meta
+            .save(store.pools.clone())
+            .await
+            .expect("completed pool metadata should be persisted");
+        let (_, err) = store
+            .handle_heal_format(false)
+            .await
+            .expect("completed pools should be reported as a no-op");
+        assert!(matches!(err, Some(StorageError::NoHealRequired)));
    }
 }
@@ -33,7 +33,7 @@ use crate::bucket::utils::check_put_object_part_args;
 use crate::bucket::utils::{check_valid_bucket_name, check_valid_bucket_name_strict, is_meta_bucketname};
 use crate::cluster::rpc::{RemoteClient, S3PeerSys};
 use crate::config::storageclass;
-use crate::core::pools::PoolMeta;
+use crate::core::pools::{DecommissionCanceler, PoolMeta};
 use crate::disk::endpoint::{Endpoint, EndpointType};
 use crate::disk::{DiskAPI, DiskInfo, DiskInfoOptions};
 use crate::error::{Error, Result};
@@ -151,7 +151,7 @@ pub(crate) mod init_format;
 pub(crate) mod list_objects;
 mod multipart;
 mod object;
-pub(crate) use object::ObjectLockDiagGuard;
+pub(crate) use object::{ObjectLockDiagGuard, SourceCleanupMutationFence};
 pub use object::{
    PrepareSelectObjectSnapshotError, PreparedGetObjectReader, SelectObjectSnapshot, SelectObjectSnapshotReadError,
    SnapshotConsistencyError,
@@ -176,7 +176,7 @@ pub struct ECStore {
    // pub local_disks: Vec<DiskStore>,
    pub pool_meta: RwLock<PoolMeta>,
    pub rebalance_meta: RwLock<Option<RebalanceMeta>>,
-    pub decommission_cancelers: RwLock<Vec<Option<CancellationToken>>>,
+    pub decommission_cancelers: RwLock<Vec<Option<DecommissionCanceler>>>,
    /// Serializes rebalance/decommission start transitions.
    ///
    /// Lock order: acquire `start_gate` before `pool_meta`, `rebalance_meta`,
@@ -400,7 +400,7 @@ impl ECStore {
        object: &str,
        opts: &ObjectOptions,
    ) -> Result<MultipartUploadResult> {
-        self.handle_new_multipart_upload_with_pool_idx(bucket, object, opts)
+        self.handle_new_multipart_upload_with_pool_idx(bucket, object, opts, None)
            .await
            .map(|(res, _, _)| res)
    }
@@ -410,20 +410,22 @@ impl ECStore {
        bucket: &str,
        object: &str,
        opts: &ObjectOptions,
+        mutation_fence: Option<&ObjectLockDiagGuard>,
    ) -> Result<(MultipartUploadResult, usize, Option<Uuid>)> {
        check_new_multipart_args(bucket, object)?;
-        let (opts, _bucket_lifecycle_guard) = self.guard_multipart_bucket_incarnation(bucket, opts).await?;
-        let opts = &opts;
+        let (mut opts, _bucket_lifecycle_guard) = self.guard_multipart_bucket_incarnation(bucket, opts).await?;

        if self.single_pool() {
+            self.apply_decommission_target_mutation_fence(0, object, &mut opts, mutation_fence)
+                .await;
            return self.pools[0]
-                .new_multipart_upload(bucket, object, opts)
+                .new_multipart_upload(bucket, object, &opts)
                .await
                .map(|res| (res, 0, opts.expected_bucket_incarnation_id));
        }

        if opts.data_movement && opts.version_id.is_some() {
-            let idx = self.select_data_movement_pool_idx(bucket, object, -1, opts, false).await?;
+            let idx = self.select_data_movement_pool_idx(bucket, object, -1, &opts, false).await?;
            if idx == opts.src_pool_idx {
                return Err(StorageError::DataMovementOverwriteErr(
                    bucket.to_owned(),
@@ -431,7 +433,9 @@ impl ECStore {
                    opts.version_id.clone().unwrap_or_default(),
                ));
            }
-            let res = self.pools[idx].new_multipart_upload(bucket, object, opts).await?;
+            self.apply_decommission_target_mutation_fence(idx, object, &mut opts, mutation_fence)
+                .await;
+            let res = self.pools[idx].new_multipart_upload(bucket, object, &opts).await?;
            return Ok((res, idx, opts.expected_bucket_incarnation_id));
        }

@@ -454,7 +458,9 @@ impl ECStore {
            .await?;

            if !res.uploads.is_empty() {
-                let res = self.pools[idx].new_multipart_upload(bucket, object, opts).await?;
+                self.apply_decommission_target_mutation_fence(idx, object, &mut opts, mutation_fence)
+                    .await;
+                let res = self.pools[idx].new_multipart_upload(bucket, object, &opts).await?;
                return Ok((res, idx, opts.expected_bucket_incarnation_id));
            }
        }
@@ -467,7 +473,9 @@ impl ECStore {
            ));
        }

-        let res = self.pools[idx].new_multipart_upload(bucket, object, opts).await?;
+        self.apply_decommission_target_mutation_fence(idx, object, &mut opts, mutation_fence)
+            .await;
+        let res = self.pools[idx].new_multipart_upload(bucket, object, &opts).await?;
        Ok((res, idx, opts.expected_bucket_incarnation_id))
    }

@@ -704,13 +712,14 @@ impl ECStore {

    pub(crate) async fn complete_multipart_upload_for_data_movement(
        self: Arc<Self>,
-        target_pool_idx: usize,
+        target: (usize, Option<&ObjectLockDiagGuard>),
        bucket: &str,
        object: &str,
        upload_id: &str,
        uploaded_parts: Vec<CompletePart>,
        opts: &ObjectOptions,
    ) -> Result<ObjectInfo> {
+        let (target_pool_idx, mutation_fence) = target;
        check_complete_multipart_args(bucket, object, upload_id)?;
        if !opts.data_movement {
            return Err(Error::other("targeted multipart completion requires data_movement options"));
@@ -739,6 +748,8 @@ impl ECStore {
            snapshot.add_lock_fences(&mut opts);
            opts.object_lock_config_snapshot = Some(snapshot);
        }
+        self.apply_decommission_target_mutation_fence(target_pool_idx, object, &mut opts, mutation_fence)
+            .await;
        #[cfg(test)]
        pause_data_movement_multipart_before_selected_completion(bucket).await;
        let pool = self
@@ -32,12 +32,13 @@ use crate::bucket::metadata_sys::{
 use crate::bucket::object_lock::objectlock_sys::{
    check_object_lock_for_deletion_with_state, ensure_recursive_force_delete_allowed_for_state,
 };
-use crate::bucket::replication::ReplicationObjectBridge;
+use crate::bucket::replication::{DeleteReplicationConfigSnapshot, ReplicationObjectBridge};
+use crate::bucket::versioning::VersioningApi;
 use crate::disk::OldCurrentSize;
 use crate::object_api::{NamespaceLockFence, ObjectLockConfigSnapshot};
 use crate::set_disk::{
-    get_lock_acquire_timeout, get_object_lock_diag_slow_acquire_threshold, get_object_lock_diag_slow_hold_threshold,
-    is_lock_optimization_enabled, is_object_lock_diag_enabled,
+    SetDisks, get_lock_acquire_timeout, get_object_lock_diag_slow_acquire_threshold, get_object_lock_diag_slow_hold_threshold,
+    is_lock_optimization_enabled, is_object_lock_diag_enabled, same_distributed_lock_domain,
 };
 use crate::storage_api_contracts::{
    namespace::NamespaceLocking as _,
@@ -352,6 +353,8 @@ impl fmt::Display for ObjectLockDiagMode {

 pub(crate) struct ObjectLockDiagGuard {
    guard: rustfs_lock::NamespaceLockGuard,
+    #[cfg(test)]
+    test_namespace_lock_fence: Option<NamespaceLockFence>,
    enabled: bool,
    op: &'static str,
    bucket: Option<String>,
@@ -373,6 +376,8 @@ impl ObjectLockDiagGuard {
    ) -> Self {
        Self {
            guard,
+            #[cfg(test)]
+            test_namespace_lock_fence: None,
            enabled,
            op,
            bucket,
@@ -393,6 +398,115 @@ impl ObjectLockDiagGuard {
    pub(crate) fn is_lock_lost(&self) -> bool {
        self.guard.is_lock_lost()
    }
+
+    pub(crate) fn add_namespace_lock_fence(&self, opts: &mut ObjectOptions) {
+        opts.ensure_namespace_lock_fence();
+        if let Some(signal) = self.lock_lost_signal() {
+            opts.add_namespace_lock_lost_signal(signal);
+        }
+        #[cfg(test)]
+        if let Some(fence) = self.test_namespace_lock_fence.as_ref() {
+            opts.add_namespace_lock_fence_for_test(fence);
+        }
+    }
+}
+
+#[cfg(test)]
+#[derive(Clone, Copy, PartialEq, Eq)]
+pub(crate) enum DecommissionMutationFenceTestPhase {
+    Migration,
+    SourceCleanup,
+}
+
+#[cfg(test)]
+struct DecommissionMutationFenceLossState {
+    bucket: String,
+    object: String,
+    phase: DecommissionMutationFenceTestPhase,
+    fence: NamespaceLockFence,
+    loss_handle: Arc<std::sync::atomic::AtomicBool>,
+}
+
+#[cfg(test)]
+pub(crate) struct DecommissionMutationFenceLossHook {
+    state: Arc<DecommissionMutationFenceLossState>,
+}
+
+#[cfg(test)]
+static DECOMMISSION_MUTATION_FENCE_LOSS_HOOK: std::sync::OnceLock<
+    std::sync::Mutex<Option<Arc<DecommissionMutationFenceLossState>>>,
+> = std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl DecommissionMutationFenceLossHook {
+    pub(crate) fn install(bucket: &str, object: &str, phase: DecommissionMutationFenceTestPhase) -> Self {
+        let (fence, loss_handle) = NamespaceLockFence::loss_handle_for_test();
+        let state = Arc::new(DecommissionMutationFenceLossState {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            phase,
+            fence,
+            loss_handle,
+        });
+        let mut slot = DECOMMISSION_MUTATION_FENCE_LOSS_HOOK
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("decommission mutation fence loss hooks should not poison");
+        assert!(slot.is_none(), "decommission mutation fence loss hook must be unique");
+        *slot = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) fn mark_lost(&self) {
+        self.state.loss_handle.store(true, Ordering::Release);
+    }
+}
+
+#[cfg(test)]
+impl Drop for DecommissionMutationFenceLossHook {
+    fn drop(&mut self) {
+        let mut slot = DECOMMISSION_MUTATION_FENCE_LOSS_HOOK
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("decommission mutation fence loss hooks should not poison");
+        if slot.as_ref().is_some_and(|hook| Arc::ptr_eq(hook, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+fn decommission_mutation_fence_for_test(
+    bucket: &str,
+    object: &str,
+    phase: DecommissionMutationFenceTestPhase,
+) -> Option<NamespaceLockFence> {
+    DECOMMISSION_MUTATION_FENCE_LOSS_HOOK
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("decommission mutation fence loss hooks should not poison")
+        .as_ref()
+        .filter(|hook| hook.bucket == bucket && hook.object == object && hook.phase == phase)
+        .map(|hook| hook.fence.clone())
+}
+
+pub(crate) struct SourceCleanupMutationFence {
+    guard: ObjectLockDiagGuard,
+    source_lock_covered: bool,
+}
+
+impl SourceCleanupMutationFence {
+    pub(crate) fn source_lock_covered(&self) -> bool {
+        self.source_lock_covered
+    }
+
+    pub(crate) fn is_lock_lost(&self) -> bool {
+        self.guard.is_lock_lost()
+    }
+
+    pub(crate) fn add_namespace_lock_fence(&self, opts: &mut ObjectOptions) {
+        self.guard.add_namespace_lock_fence(opts);
+    }
 }

 /// Opaque write-lock guard for the RestoreObject accept path; see
@@ -410,10 +524,7 @@ impl RestoreAcceptGuard {
    }

    pub fn add_namespace_lock_fence(&self, opts: &mut ObjectOptions) {
-        opts.ensure_namespace_lock_fence();
-        if let Some(signal) = self.0.lock_lost_signal() {
-            opts.add_namespace_lock_lost_signal(signal);
-        }
+        self.0.add_namespace_lock_fence(opts);
    }
 }

@@ -690,16 +801,6 @@ impl SelectObjectSnapshotLockLossWake {
    }
 }

-// LockRegistry clones its canonical client Arc for each endpoint host, so an
-// exact Arc set identifies one distributed namespace-lock quorum domain.
-fn same_distributed_lock_domain(left: &[Arc<dyn rustfs_lock::LockClient>], right: &[Arc<dyn rustfs_lock::LockClient>]) -> bool {
-    left.iter()
-        .all(|left_client| right.iter().any(|right_client| Arc::ptr_eq(left_client, right_client)))
-        && right
-            .iter()
-            .all(|right_client| left.iter().any(|left_client| Arc::ptr_eq(left_client, right_client)))
-}
-
 impl AsyncRead for SelectObjectSnapshotReader {
    fn poll_read(mut self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &mut ReadBuf<'_>) -> Poll<std::io::Result<()>> {
        if self.lock_loss_wake.poll_lost(cx) || self.lease.is_lost() {
@@ -805,7 +906,7 @@ fn resolve_latest_object_access(
 }

 fn should_create_delete_marker_for_missing_object(opts: &ObjectOptions) -> bool {
-    opts.versioned && opts.version_id.is_none() && !opts.delete_marker && !opts.data_movement
+    (opts.versioned || opts.version_suspended) && opts.version_id.is_none() && !opts.delete_marker && !opts.data_movement
 }

 #[cfg(test)]
@@ -813,6 +914,8 @@ struct DeleteAfterObjectLockSnapshotBarrierState {
    bucket: String,
    arrived: tokio::sync::Notify,
    release: tokio::sync::Notify,
+    namespace_pending: tokio::sync::Notify,
+    namespace_acquired: AtomicBool,
 }

 #[cfg(test)]
@@ -832,6 +935,8 @@ impl DeleteAfterObjectLockSnapshotBarrier {
            bucket: bucket.to_string(),
            arrived: tokio::sync::Notify::new(),
            release: tokio::sync::Notify::new(),
+            namespace_pending: tokio::sync::Notify::new(),
+            namespace_acquired: AtomicBool::new(false),
        });
        let mut slot = DELETE_AFTER_OBJECT_LOCK_SNAPSHOT_BARRIER
            .get_or_init(|| std::sync::Mutex::new(None))
@@ -849,6 +954,18 @@ impl DeleteAfterObjectLockSnapshotBarrier {
    pub(crate) fn release(&self) {
        self.state.release.notify_one();
    }
+
+    pub(crate) async fn release_and_wait_until_namespace_pending(&self) {
+        let namespace_pending = self.state.namespace_pending.notified();
+        self.release();
+        tokio::time::timeout(Duration::from_secs(5), namespace_pending)
+            .await
+            .expect("delete should proceed to its namespace lock after leaving the snapshot barrier");
+    }
+
+    pub(crate) fn namespace_acquired(&self) -> bool {
+        self.state.namespace_acquired.load(Ordering::Acquire)
+    }
 }

 #[cfg(test)]
@@ -873,6 +990,97 @@ async fn pause_delete_after_object_lock_snapshot(bucket: &str) {
        .as_ref()
        .filter(|state| state.bucket == bucket)
        .cloned();
+    if let Some(state) = state {
+        state.arrived.notify_one();
+        state.release.notified().await;
+        state.namespace_pending.notify_one();
+    }
+}
+
+#[cfg(test)]
+fn notify_delete_namespace_acquired(bucket: &str) {
+    let state = DELETE_AFTER_OBJECT_LOCK_SNAPSHOT_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("delete snapshot barrier mutex should not poison")
+        .as_ref()
+        .filter(|state| state.bucket == bucket)
+        .cloned();
+    if let Some(state) = state {
+        state.namespace_acquired.store(true, Ordering::Release);
+    }
+}
+
+#[cfg(test)]
+struct VersionedDeleteMarkerCommitBarrierState {
+    bucket: String,
+    object: String,
+    arrived: tokio::sync::Notify,
+    release: tokio::sync::Notify,
+}
+
+#[cfg(test)]
+pub(crate) struct VersionedDeleteMarkerCommitBarrier {
+    state: Arc<VersionedDeleteMarkerCommitBarrierState>,
+}
+
+#[cfg(test)]
+static VERSIONED_DELETE_MARKER_COMMIT_BARRIER: std::sync::OnceLock<
+    std::sync::Mutex<Option<Arc<VersionedDeleteMarkerCommitBarrierState>>>,
+> = std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl VersionedDeleteMarkerCommitBarrier {
+    pub(crate) fn install(bucket: &str, object: &str) -> Self {
+        let state = Arc::new(VersionedDeleteMarkerCommitBarrierState {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            arrived: tokio::sync::Notify::new(),
+            release: tokio::sync::Notify::new(),
+        });
+        let mut slot = VERSIONED_DELETE_MARKER_COMMIT_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("versioned delete-marker commit barrier mutex should not poison");
+        assert!(slot.is_none(), "versioned delete-marker commit barrier must be unique");
+        *slot = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) async fn wait_until_paused(&self) {
+        tokio::time::timeout(Duration::from_secs(30), self.state.arrived.notified())
+            .await
+            .expect("versioned DELETE should reach the post-marker-commit barrier");
+    }
+
+    pub(crate) fn release(&self) {
+        self.state.release.notify_one();
+    }
+}
+
+#[cfg(test)]
+impl Drop for VersionedDeleteMarkerCommitBarrier {
+    fn drop(&mut self) {
+        self.state.release.notify_one();
+        let mut slot = VERSIONED_DELETE_MARKER_COMMIT_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("versioned delete-marker commit barrier mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+async fn pause_versioned_delete_marker_after_commit(bucket: &str, object: &str) {
+    let state = VERSIONED_DELETE_MARKER_COMMIT_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("versioned delete-marker commit barrier mutex should not poison")
+        .as_ref()
+        .filter(|state| state.bucket == bucket && state.object == object)
+        .cloned();
    if let Some(state) = state {
        state.arrived.notify_one();
        state.release.notified().await;
@@ -913,6 +1121,160 @@ fn writer_pool_lookup_opts(opts: &ObjectOptions, no_lock: bool) -> ObjectOptions
    lookup_opts
 }

+fn delete_pool_lookup_opts(opts: &ObjectOptions, no_lock: bool) -> ObjectOptions {
+    let mut lookup_opts = writer_pool_lookup_opts(opts, no_lock);
+    lookup_opts.skip_decommissioned = opts.data_movement;
+    lookup_opts
+}
+
+fn should_delete_from_all_pools(opts: &ObjectOptions, pool_count: usize) -> bool {
+    pool_count > 0 && (!opts.versioned && !opts.version_suspended || opts.version_id.is_some())
+}
+
+fn batch_delete_creates_latest_marker(object: &ObjectToDelete, delete_config_snapshot: &DeleteReplicationConfigSnapshot) -> bool {
+    if object.version_id.is_some() {
+        return false;
+    }
+
+    let object_name = decode_dir_object(&object.object_name);
+    let (versioned, version_suspended) = delete_config_snapshot.versioning_config().delete_state(&object_name);
+    versioned || version_suspended
+}
+
+fn batch_delete_targets_pool(creates_latest_marker: bool, marker_target_pool_idx: Option<usize>, pool_idx: usize) -> bool {
+    !creates_latest_marker || marker_target_pool_idx == Some(pool_idx)
+}
+
+#[cfg(test)]
+struct BatchDeletePoolErrorInjectionState {
+    bucket: String,
+    pool_idx: usize,
+    errors: std::collections::HashMap<String, Error>,
+    observed: std::sync::atomic::AtomicUsize,
+}
+
+#[cfg(test)]
+pub(crate) struct BatchDeletePoolErrorInjection {
+    state: Arc<BatchDeletePoolErrorInjectionState>,
+}
+
+#[cfg(test)]
+static BATCH_DELETE_POOL_ERROR_INJECTION: std::sync::OnceLock<std::sync::Mutex<Option<Arc<BatchDeletePoolErrorInjectionState>>>> =
+    std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl BatchDeletePoolErrorInjection {
+    pub(crate) fn install(bucket: &str, pool_idx: usize, errors: Vec<(String, Error)>) -> Self {
+        let state = Arc::new(BatchDeletePoolErrorInjectionState {
+            bucket: bucket.to_string(),
+            pool_idx,
+            errors: errors.into_iter().collect(),
+            observed: std::sync::atomic::AtomicUsize::new(0),
+        });
+        let mut slot = BATCH_DELETE_POOL_ERROR_INJECTION
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("batch delete pool error injection mutex should not poison");
+        assert!(slot.is_none(), "batch delete pool error injection must be unique");
+        *slot = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) fn observed(&self) -> usize {
+        self.state.observed.load(Ordering::Acquire)
+    }
+}
+
+#[cfg(test)]
+impl Drop for BatchDeletePoolErrorInjection {
+    fn drop(&mut self) {
+        let mut slot = BATCH_DELETE_POOL_ERROR_INJECTION
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("batch delete pool error injection mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+fn inject_batch_delete_pool_errors(
+    bucket: &str,
+    pool_idx: usize,
+    object_names: &[String],
+    result: &mut (Vec<DeletedObject>, Vec<Option<Error>>),
+) {
+    let state = BATCH_DELETE_POOL_ERROR_INJECTION
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("batch delete pool error injection mutex should not poison")
+        .as_ref()
+        .filter(|state| state.bucket == bucket && state.pool_idx == pool_idx)
+        .cloned();
+    let Some(state) = state else {
+        return;
+    };
+
+    for (idx, object_name) in object_names.iter().enumerate() {
+        let Some(error) = state.errors.get(object_name) else {
+            continue;
+        };
+        if result.1[idx].is_none() && result.0[idx].found {
+            result.1[idx] = Some(error.clone());
+            state.observed.fetch_add(1, Ordering::AcqRel);
+        }
+    }
+}
+
+fn resolve_batch_delete_pool_results<'a>(
+    initial_error: Option<Error>,
+    pool_results: impl IntoIterator<Item = (&'a DeletedObject, &'a Option<Error>)>,
+) -> (Option<DeletedObject>, Option<Error>, bool) {
+    let mut failure = initial_error.map(|err| (None, err));
+    let mut deleted = None;
+    let mut fallback: Option<(DeletedObject, Option<Error>)> = None;
+    let mut attempted = false;
+
+    for (pool_delete, pool_error) in pool_results {
+        attempted = true;
+        match pool_error {
+            Some(err) if is_err_object_not_found(err) || is_err_version_not_found(err) => {
+                if fallback.as_ref().is_none_or(|(_, error)| error.is_none()) {
+                    fallback = Some(((*pool_delete).clone(), Some(err.clone())));
+                }
+            }
+            Some(err) => {
+                if failure.is_none() {
+                    failure = Some((Some((*pool_delete).clone()), err.clone()));
+                }
+            }
+            None if pool_delete.found => {
+                if deleted.is_none() {
+                    deleted = Some((*pool_delete).clone());
+                }
+            }
+            None => {
+                if fallback.is_none() {
+                    fallback = Some(((*pool_delete).clone(), None));
+                }
+            }
+        }
+    }
+
+    if let Some((failed_delete, err)) = failure {
+        return (failed_delete, Some(err), attempted);
+    }
+    if let Some(deleted) = deleted {
+        return (Some(deleted), None, attempted);
+    }
+    if let Some((deleted, err)) = fallback {
+        return (Some(deleted), err, attempted);
+    }
+
+    (None, None, attempted)
+}
+
 fn transition_restore_pool_opts(opts: &ObjectOptions) -> ObjectOptions {
    let mut lookup_opts = opts.clone();
    lookup_opts.skip_decommissioned = true;
@@ -1541,6 +1903,89 @@ impl ECStore {
        )))
    }

+    pub(crate) async fn acquire_decommission_object_mutation_fence(
+        &self,
+        bucket: &str,
+        object: &str,
+    ) -> Result<ObjectLockDiagGuard> {
+        if self.ctx.lock_manager().is_disabled() {
+            return Err(Error::other("decommission object migration requires namespace locking"));
+        }
+
+        #[cfg(test)]
+        let test_namespace_lock_fence =
+            decommission_mutation_fence_for_test(bucket, object, DecommissionMutationFenceTestPhase::Migration);
+        let object = encode_dir_object(object);
+        let mut opts = ObjectOptions::default();
+        let guard = self
+            .acquire_object_read_lock_if_needed("decommission_object", bucket, &object, &mut opts)
+            .await?
+            .ok_or_else(|| Error::other("decommission object migration failed to acquire its namespace fence"))?;
+        #[cfg(test)]
+        let guard = {
+            let mut guard = guard;
+            guard.test_namespace_lock_fence = test_namespace_lock_fence;
+            guard
+        };
+        Ok(guard)
+    }
+
+    pub(super) async fn apply_decommission_target_mutation_fence(
+        &self,
+        target_pool_idx: usize,
+        object: &str,
+        opts: &mut ObjectOptions,
+        mutation_fence: Option<&ObjectLockDiagGuard>,
+    ) {
+        let Some(mutation_fence) = mutation_fence else {
+            return;
+        };
+
+        mutation_fence.add_namespace_lock_fence(opts);
+        let fixed_set = self.pools.first().and_then(|pool| pool.disk_set.first());
+        let target_set = self.pools.get(target_pool_idx).map(|pool| pool.get_disks_by_key(object));
+        opts.no_lock = match (fixed_set, target_set) {
+            (Some(fixed), Some(target)) => fixed.shares_namespace_lock_domain(&target).await,
+            _ => false,
+        };
+    }
+
+    pub(crate) async fn acquire_decommission_source_cleanup_fence(
+        &self,
+        bucket: &str,
+        object: &str,
+        source_set: &SetDisks,
+    ) -> Result<SourceCleanupMutationFence> {
+        if self.ctx.lock_manager().is_disabled() {
+            return Err(Error::other("decommission source cleanup requires namespace locking"));
+        }
+
+        #[cfg(test)]
+        crate::data_movement::notify_source_cleanup_mutation_fence_pending(bucket, object);
+        #[cfg(test)]
+        let test_namespace_lock_fence =
+            decommission_mutation_fence_for_test(bucket, object, DecommissionMutationFenceTestPhase::SourceCleanup);
+        let object = encode_dir_object(object);
+        let fixed_set = Arc::clone(&self.pools[0].disk_set[0]);
+        let source_lock_covered = fixed_set.shares_namespace_lock_domain(source_set).await;
+        // Lock order: fixed store mutation domain first; source cleanup takes its
+        // hashed source-domain lock second only when this guard does not cover it.
+        let guard = self
+            .acquire_object_write_lock("decommission_source_cleanup", bucket, &object)
+            .await?;
+        #[cfg(test)]
+        let guard = {
+            let mut guard = guard;
+            guard.test_namespace_lock_fence = test_namespace_lock_fence;
+            guard
+        };
+
+        Ok(SourceCleanupMutationFence {
+            guard,
+            source_lock_covered,
+        })
+    }
+
    pub(crate) async fn acquire_all_object_read_locks(
        &self,
        op: &'static str,
@@ -1994,14 +2439,17 @@ impl ECStore {
        object: &str,
        data: &mut PutObjReader,
        opts: &ObjectOptions,
+        mutation_fence: Option<&ObjectLockDiagGuard>,
    ) -> Result<(usize, Result<ObjectInfo>)> {
        if !opts.data_movement {
            return Err(Error::other("data movement PUT requires data_movement options"));
        }
-        let (object, opts) = self.prepare_put_object(bucket, object, opts).await?;
+        let (object, mut opts) = self.prepare_put_object(bucket, object, opts).await?;
        let idx = self
            .select_put_object_pool_idx(bucket, object.as_str(), data.size(), &opts)
            .await?;
+        self.apply_decommission_target_mutation_fence(idx, object.as_str(), &mut opts, mutation_fence)
+            .await;
        let result = self.pools[idx]
            .put_object_with_old_current_size(bucket, &object, data, &opts)
            .await
@@ -2470,6 +2918,10 @@ impl ECStore {
        } else {
            None
        };
+        #[cfg(test)]
+        if _object_lock_guard.is_some() {
+            notify_delete_namespace_acquired(bucket);
+        }
        if let Some(trigger) = opts.lifecycle_delete_all.as_ref() {
            let configs = delete_all_configs.as_ref().ok_or(StorageError::PreconditionFailed)?;
            let expected_bucket_incarnation_id = opts.expected_bucket_incarnation_id.ok_or(StorageError::PreconditionFailed)?;
@@ -2503,7 +2955,7 @@ impl ECStore {
            return Ok(ObjectInfo::default());
        }

-        let gopts = writer_pool_lookup_opts(&opts, true);
+        let gopts = delete_pool_lookup_opts(&opts, true);

        if opts.data_movement {
            let existing_pool_info = self.get_pool_info_existing_with_opts(bucket, object, &gopts).await;
@@ -2608,6 +3060,8 @@ impl ECStore {
            Err(err) if is_err_object_not_found(&err) && should_create_delete_marker_for_missing_object(&opts) => {
                let target_pool_idx = self.get_pool_idx_no_lock(bucket, object, 0).await?;
                let mut obj = self.pools[target_pool_idx].delete_object(bucket, object, opts).await?;
+                #[cfg(test)]
+                pause_versioned_delete_marker_after_commit(bucket, object).await;
                obj.name = decode_dir_object(object);
                return Ok(obj);
            }
@@ -2646,7 +3100,7 @@ impl ECStore {
            None
        };

-        if !errs.is_empty() && !opts.versioned && !opts.version_suspended {
+        if should_delete_from_all_pools(&opts, errs.len()) {
            let mut obj = match self.delete_object_from_all_pools(bucket, object, &opts, errs).await {
                Ok(obj) => obj,
                Err(err) => {
@@ -2670,6 +3124,8 @@ impl ECStore {

            match pool.delete_object(bucket, object, opts.clone()).await {
                Ok(res) => {
+                    #[cfg(test)]
+                    pause_versioned_delete_marker_after_commit(bucket, object).await;
                    if let (Some(api), Some(je)) = (tier_journal_api.as_ref(), journal_entry.as_ref()) {
                        commit_prepared_tier_delete_journal_entry(api, je).await;
                    }
@@ -2738,7 +3194,7 @@ impl ECStore {

        // Default return value
        let mut del_objects = vec![DeletedObject::default(); objects.len()];
-        let mut accounting = vec![None; objects.len()];
+        let accounting = vec![None; objects.len()];

        let mut del_errs = Vec::with_capacity(objects.len());
        for _ in 0..objects.len() {
@@ -2817,32 +3273,104 @@ impl ECStore {
            Ok(guards) => guards,
            Err(err) => return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err),
        };
+        #[cfg(test)]
+        if !_object_lock_guards.is_empty() {
+            notify_delete_namespace_acquired(bucket);
+        }
+
+        let delete_config_snapshot = opts
+            .delete_replication_config_snapshot
+            .as_deref()
+            .expect("batch delete replication config snapshot should be loaded");
+        let latest_marker_objects = objects
+            .iter()
+            .map(|object| batch_delete_creates_latest_marker(object, delete_config_snapshot))
+            .collect::<Vec<_>>();
+        let marker_target_results = join_all(objects.iter().zip(&latest_marker_objects).map(
+            |(object, creates_marker)| async move {
+                if *creates_marker {
+                    Some(self.get_pool_idx_no_lock(bucket, &object.object_name, 0).await)
+                } else {
+                    None
+                }
+            },
+        ))
+        .await;
+        let mut marker_target_pool_indices = Vec::with_capacity(objects.len());
+        for (idx, target_result) in marker_target_results.into_iter().enumerate() {
+            match target_result {
+                Some(Ok(pool_idx)) => marker_target_pool_indices.push(Some(pool_idx)),
+                Some(Err(err)) => {
+                    del_errs[idx] = Some(err);
+                    marker_target_pool_indices.push(None);
+                }
+                None => marker_target_pool_indices.push(None),
+            }
+        }

        let mut futures = Vec::with_capacity(self.pools.len());
-
        for pool in self.pools.iter() {
            if self.is_pool_rebalancing(pool.pool_idx).await {
                continue;
            }
-            futures.push(pool.delete_objects_with_accounting(bucket, objects.clone(), opts.clone()));
+
+            let (object_indices, pool_objects): (Vec<_>, Vec<_>) = objects
+                .iter()
+                .enumerate()
+                .filter(|(idx, _)| {
+                    batch_delete_targets_pool(latest_marker_objects[*idx], marker_target_pool_indices[*idx], pool.pool_idx)
+                })
+                .map(|(idx, object)| (idx, object.clone()))
+                .unzip();
+            if pool_objects.is_empty() {
+                continue;
+            }
+
+            let pool_opts = opts.clone();
+            futures.push(async move {
+                #[cfg(test)]
+                let pool_object_names = pool_objects
+                    .iter()
+                    .map(|object| object.object_name.clone())
+                    .collect::<Vec<_>>();
+                let result = pool.delete_objects(bucket, pool_objects, pool_opts).await;
+                #[cfg(test)]
+                let result = {
+                    let mut result = result;
+                    inject_batch_delete_pool_errors(bucket, pool.pool_idx, &pool_object_names, &mut result);
+                    result
+                };
+                (object_indices, result)
+            });
        }

        let results = join_all(futures).await;

        for idx in 0..del_objects.len() {
-            for (dels, errs, pool_accounting) in results.iter() {
-                if errs[idx].is_none() && dels[idx].found {
-                    del_errs[idx] = None;
-                    del_objects[idx] = dels[idx].clone();
-                    accounting[idx] = pool_accounting[idx].clone();
-                    break;
-                }
+            let pool_results = results.iter().filter_map(|(object_indices, (dels, errs))| {
+                let pool_object_idx = object_indices.binary_search(&idx).ok()?;
+                Some((&dels[pool_object_idx], &errs[pool_object_idx]))
+            });
+            let (deleted, error, attempted) = resolve_batch_delete_pool_results(del_errs[idx].take(), pool_results);
+            if let Some(deleted) = deleted {
+                del_objects[idx] = deleted;
+            }
+            del_errs[idx] = error;

-                if del_errs[idx].is_none() {
-                    del_errs[idx] = errs[idx].clone();
-                    del_objects[idx] = dels[idx].clone();
-                    accounting[idx] = pool_accounting[idx].clone();
-                }
+            if !attempted && del_errs[idx].is_none() && latest_marker_objects[idx] {
+                del_objects[idx] = DeletedObject {
+                    object_name: objects[idx].object_name.clone(),
+                    version_id: objects[idx].version_id,
+                    ..Default::default()
+                };
+                del_errs[idx] = Some(StorageError::ObjectNotFound(bucket.to_owned(), objects[idx].object_name.clone()));
+            }
+        }
+
+        #[cfg(test)]
+        for (idx, object) in objects.iter().enumerate() {
+            if del_errs[idx].is_none() && del_objects[idx].delete_marker {
+                pause_versioned_delete_marker_after_commit(bucket, &object.object_name).await;
            }
        }

@@ -3417,6 +3945,80 @@ mod tests {
        assert!(!same_distributed_lock_domain(&[first, second], &[other]));
    }

+    #[tokio::test]
+    async fn decommission_fence_covers_dist_sets_with_same_clients_despite_different_namespaces() {
+        let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let (_dirs, original_sets) = make_local_two_set_sets_with_ctx(Arc::clone(&ctx)).await;
+        let mut second_set = (*original_sets.disk_set[1]).clone();
+        second_set.lockers = original_sets.disk_set[0].lockers.clone();
+        let mut sets = (*original_sets).clone();
+        sets.disk_set[1] = Arc::new(second_set);
+        let sets = Arc::new(sets);
+        ctx.update_erasure_type(SetupType::DistErasure).await;
+
+        assert!(
+            sets.disk_set[0]
+                .lockers
+                .iter()
+                .zip(&sets.disk_set[1].lockers)
+                .all(|(fixed, hashed)| Arc::ptr_eq(fixed, hashed)),
+            "the regression requires identical distributed lock clients"
+        );
+        assert_ne!(sets.disk_set[0].set_index, sets.disk_set[1].set_index);
+
+        let pool_config = sets.endpoints.clone();
+        let store = new_prepared_reader_test_store_from_pools(vec![Arc::clone(&sets)], vec![pool_config], ctx);
+        let object = (0..1_000)
+            .map(|index| format!("decommission-dist-domain-{index}.bin"))
+            .find(|candidate| Arc::ptr_eq(&sets.get_disks_by_key(candidate), &sets.disk_set[1]))
+            .expect("a key should hash to the second set namespace");
+        let mutation_fence = store
+            .acquire_decommission_object_mutation_fence("bucket", &object)
+            .await
+            .expect("the fixed distributed mutation fence should be acquired");
+        let target_lock = sets.disk_set[1]
+            .new_ns_lock("bucket", &object)
+            .await
+            .expect("the hashed-set namespace lock should be created");
+        let target_err = target_lock
+            .get_write_lock(Duration::from_millis(50))
+            .await
+            .expect_err("the fixed read fence must conflict through the shared clients");
+        assert!(matches!(target_err, rustfs_lock::LockError::Timeout { .. }));
+
+        let mut put_opts = ObjectOptions::default();
+        store
+            .apply_decommission_target_mutation_fence(0, &object, &mut put_opts, Some(&mutation_fence))
+            .await;
+        assert!(put_opts.no_lock, "migration target PUT must reuse the covering fixed fence");
+
+        let mut multipart_opts = ObjectOptions::default();
+        store
+            .apply_decommission_target_mutation_fence(0, &object, &mut multipart_opts, Some(&mutation_fence))
+            .await;
+        assert!(multipart_opts.no_lock, "migration target multipart must reuse the covering fixed fence");
+        drop(mutation_fence);
+
+        let cleanup_object = (0..1_000)
+            .map(|index| format!("decommission-dist-cleanup-{index}.bin"))
+            .find(|candidate| Arc::ptr_eq(&sets.get_disks_by_key(candidate), &sets.disk_set[1]))
+            .expect("a cleanup key should hash to the second set namespace");
+        let source_fence = store
+            .acquire_decommission_source_cleanup_fence("bucket", &cleanup_object, sets.disk_set[1].as_ref())
+            .await
+            .expect("the fixed distributed cleanup fence should be acquired");
+        assert!(source_fence.source_lock_covered(), "source cleanup must reuse the covering fixed fence");
+        let source_lock = sets.disk_set[1]
+            .new_ns_lock("bucket", &cleanup_object)
+            .await
+            .expect("the source-set namespace lock should be created");
+        let source_err = source_lock
+            .get_read_lock(Duration::from_millis(50))
+            .await
+            .expect_err("the fixed write fence must conflict through the shared clients");
+        assert!(matches!(source_err, rustfs_lock::LockError::Timeout { .. }));
+    }
+
    #[test]
    fn select_snapshot_version_matching_normalizes_null_and_uuid_forms() {
        let nil = Uuid::nil();
@@ -4476,6 +5078,159 @@ mod tests {
        assert_eq!(lookup_opts.version_id.as_deref(), Some("vid-1"));
    }

+    #[test]
+    fn ordinary_delete_lookup_includes_decommission_source_and_skips_rebalance_source() {
+        let lookup_opts = delete_pool_lookup_opts(&ObjectOptions::default(), true);
+
+        assert!(lookup_opts.no_lock);
+        assert!(!lookup_opts.skip_decommissioned);
+        assert!(lookup_opts.skip_rebalancing);
+
+        let explicit_version = delete_pool_lookup_opts(
+            &ObjectOptions {
+                versioned: true,
+                version_id: Some(uuid::Uuid::new_v4().to_string()),
+                ..Default::default()
+            },
+            true,
+        );
+        assert!(!explicit_version.skip_decommissioned);
+    }
+
+    #[test]
+    fn delete_fans_out_for_unversioned_and_explicit_version_mutations() {
+        assert!(should_delete_from_all_pools(&ObjectOptions::default(), 1));
+        assert!(should_delete_from_all_pools(
+            &ObjectOptions {
+                versioned: true,
+                version_id: Some(uuid::Uuid::new_v4().to_string()),
+                ..Default::default()
+            },
+            2,
+        ));
+        assert!(!should_delete_from_all_pools(
+            &ObjectOptions {
+                versioned: true,
+                ..Default::default()
+            },
+            1,
+        ));
+        assert!(!should_delete_from_all_pools(&ObjectOptions::default(), 0));
+    }
+
+    #[test]
+    fn batch_delete_identifies_only_latest_versioned_markers() {
+        let versioned = DeleteReplicationConfigSnapshot::from_configs_for_test(
+            s3s::dto::VersioningConfiguration {
+                status: Some(s3s::dto::BucketVersioningStatus::from_static(s3s::dto::BucketVersioningStatus::ENABLED)),
+                ..Default::default()
+            },
+            None,
+        );
+        let latest = ObjectToDelete {
+            object_name: "latest".to_string(),
+            ..Default::default()
+        };
+        assert!(batch_delete_creates_latest_marker(&latest, &versioned));
+        assert!(!batch_delete_targets_pool(true, Some(1), 0));
+        assert!(batch_delete_targets_pool(true, Some(1), 1));
+        assert!(!batch_delete_targets_pool(true, Some(1), 2));
+
+        let explicit = ObjectToDelete {
+            object_name: "explicit".to_string(),
+            version_id: Some(uuid::Uuid::new_v4()),
+            ..Default::default()
+        };
+        assert!(!batch_delete_creates_latest_marker(&explicit, &versioned));
+        assert!(batch_delete_targets_pool(false, Some(1), 0));
+
+        let unversioned = DeleteReplicationConfigSnapshot::default();
+        assert!(!batch_delete_creates_latest_marker(&latest, &unversioned));
+        assert!(batch_delete_targets_pool(false, None, 0));
+    }
+
+    #[test]
+    fn batch_delete_pool_failures_override_success_in_any_pool_order() {
+        let success = DeletedObject {
+            object_name: "object".to_string(),
+            found: true,
+            ..Default::default()
+        };
+        let source_errors = [
+            StorageError::ErasureWriteQuorum,
+            StorageError::NamespaceLockQuorumUnavailable {
+                mode: "delete_objects_commit",
+                bucket: "bucket".to_string(),
+                object: "object".to_string(),
+                required: 1,
+                achieved: 0,
+            },
+        ];
+
+        for source_error in source_errors {
+            for source_first in [true, false] {
+                let failed = (DeletedObject::default(), Some(source_error.clone()));
+                let succeeded = (success.clone(), None);
+                let pool_results = if source_first {
+                    vec![failed, succeeded]
+                } else {
+                    vec![succeeded, failed]
+                };
+
+                let (_, error, attempted) =
+                    resolve_batch_delete_pool_results(None, pool_results.iter().map(|(deleted, error)| (deleted, error)));
+
+                assert!(attempted);
+                assert_eq!(error, Some(source_error.clone()));
+            }
+        }
+    }
+
+    #[test]
+    fn batch_delete_ignores_missing_pool_only_after_another_pool_succeeds() {
+        let success = DeletedObject {
+            object_name: "object".to_string(),
+            found: true,
+            ..Default::default()
+        };
+        let missing_errors = [
+            StorageError::ObjectNotFound("bucket".to_string(), "object".to_string()),
+            StorageError::VersionNotFound("bucket".to_string(), "object".to_string(), "version".to_string()),
+        ];
+
+        for missing_error in missing_errors {
+            let missing = (DeletedObject::default(), Some(missing_error.clone()));
+            for missing_first in [true, false] {
+                let succeeded = (success.clone(), None);
+                let pool_results = if missing_first {
+                    vec![missing.clone(), succeeded]
+                } else {
+                    vec![succeeded, missing.clone()]
+                };
+                let (deleted, error, attempted) =
+                    resolve_batch_delete_pool_results(None, pool_results.iter().map(|(deleted, error)| (deleted, error)));
+
+                assert!(attempted);
+                let deleted = deleted.expect("successful pool result should be retained");
+                assert!(deleted.found);
+                assert_eq!(deleted.object_name, success.object_name.as_str());
+                assert!(error.is_none());
+            }
+
+            let missing_only = [missing];
+            let (_, error, attempted) =
+                resolve_batch_delete_pool_results(None, missing_only.iter().map(|(deleted, error)| (deleted, error)));
+            assert!(attempted);
+            assert_eq!(error, Some(missing_error));
+        }
+
+        let silent_missing = [(DeletedObject::default(), None)];
+        let (_, error, attempted) =
+            resolve_batch_delete_pool_results(None, silent_missing.iter().map(|(deleted, error)| (deleted, error)));
+        assert!(attempted);
+        assert!(error.is_none());
+    }
+
    #[test]
    fn data_movement_pool_lookup_opts_keeps_no_lock_for_tiered_moves() {
        let lookup_opts = data_movement_pool_lookup_opts(
@@ -859,6 +859,7 @@ fn lifecycle_delete_all_test_failure(phase: crate::object_api::LifecycleDeleteAl
 #[cfg(test)]
 mod tests {
    use super::*;
+    use crate::bucket::replication::{ReplicationStatusType, VersionPurgeStatusType};
    use crate::config::storageclass::{CLASS_RRS, CLASS_STANDARD, lookup_config_for_pools_without_env};
    use crate::disk::error::DiskError;
    use crate::layout::endpoint::Endpoint;
@@ -1423,6 +1424,14 @@ mod tests {
        }
    }

+    fn object_info_with_identity(unix_ts: i64, delete_marker: bool, version_id: Uuid, etag: Option<String>) -> ObjectInfo {
+        ObjectInfo {
+            version_id: Some(version_id),
+            etag,
+            ..object_info_with_mod_time(unix_ts, delete_marker)
+        }
+    }
+
    #[test]
    fn resolve_latest_object_info_candidates_returns_latest_delete_marker() {
        let candidates = vec![
@@ -1446,7 +1455,7 @@ mod tests {
    }

    #[test]
-    fn resolve_latest_object_info_candidates_prefers_higher_pool_idx_on_equal_mod_time() {
+    fn resolve_latest_object_info_candidates_prefers_higher_pool_idx_on_equal_mod_time_for_equivalent_candidates() {
        let candidates = vec![
            LatestObjectInfoCandidate {
                info: Some(object_info_with_mod_time(10, false)),
@@ -1466,6 +1475,382 @@ mod tests {
        assert_eq!(idx, 1);
    }

+    #[test]
+    fn resolve_latest_object_info_candidates_keeps_index_fallback_for_fully_equivalent_identities() {
+        let candidates = vec![
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()))),
+                idx: 2,
+                err: None,
+            },
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()))),
+                idx: 7,
+                err: None,
+            },
+        ];
+
+        let (info, idx) = resolve_latest_object_info_candidates(candidates, "bucket", "object", &ObjectOptions::default())
+            .expect("equivalent replicas must resolve deterministically");
+
+        assert_eq!(idx, 7);
+        assert_eq!(info.version_id, Some(Uuid::from_u128(1)));
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_equal_time_version_id_conflict() {
+        let candidates = vec![
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()))),
+                idx: 0,
+                err: None,
+            },
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(2), Some("etag-a".to_string()))),
+                idx: 1,
+                err: None,
+            },
+        ];
+
+        let err = resolve_latest_object_info_candidates(candidates, "bucket", "object", &ObjectOptions::default())
+            .expect_err("divergent version ids must not silently resolve to the higher pool index");
+
+        assert_eq!(err, Error::ErasureReadQuorum);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_equal_time_etag_conflict() {
+        let candidates = vec![
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-old".to_string()))),
+                idx: 0,
+                err: None,
+            },
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-new".to_string()))),
+                idx: 1,
+                err: None,
+            },
+        ];
+
+        let err = resolve_latest_object_info_candidates(candidates, "bucket", "object", &ObjectOptions::default())
+            .expect_err("divergent etags must not silently resolve to the higher pool index");
+
+        assert_eq!(err, Error::ErasureReadQuorum);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_equal_time_delete_marker_conflict() {
+        let candidates = vec![
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), None)),
+                idx: 0,
+                err: None,
+            },
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, true, Uuid::from_u128(1), Some("etag-a".to_string()))),
+                idx: 1,
+                err: None,
+            },
+        ];
+
+        let err = resolve_latest_object_info_candidates(candidates, "bucket", "object", &ObjectOptions::default())
+            .expect_err("a delete marker tied with a live version must not be masked by the pool index");
+
+        assert_eq!(err, Error::ErasureReadQuorum);
+    }
+
+    fn assert_equal_time_identity_conflict(left: ObjectInfo, right: ObjectInfo) {
+        let err = resolve_latest_object_info_candidates(
+            vec![
+                LatestObjectInfoCandidate {
+                    info: Some(left),
+                    idx: 0,
+                    err: None,
+                },
+                LatestObjectInfoCandidate {
+                    info: Some(right),
+                    idx: 1,
+                    err: None,
+                },
+            ],
+            "bucket",
+            "object",
+            &ObjectOptions::default(),
+        )
+        .expect_err("equal-time identity divergence must fail closed");
+
+        assert_eq!(err, Error::ErasureReadQuorum);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_equal_time_payload_identity_conflicts() {
+        let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+
+        let mut data_dir = base.clone();
+        data_dir.data_dir = Some(Uuid::from_u128(2));
+        assert_equal_time_identity_conflict(base.clone(), data_dir);
+
+        let mut size = base.clone();
+        size.size = 1;
+        assert_equal_time_identity_conflict(base.clone(), size);
+
+        let mut actual_size = base.clone();
+        actual_size.actual_size = 1;
+        assert_equal_time_identity_conflict(base.clone(), actual_size);
+
+        let mut checksum = base.clone();
+        checksum.checksum = Some(bytes::Bytes::from_static(b"checksum"));
+        assert_equal_time_identity_conflict(base.clone(), checksum);
+
+        let mut parts = base.clone();
+        parts.parts = std::sync::Arc::new(vec![rustfs_filemeta::ObjectPartInfo {
+            etag: "part-etag".to_string(),
+            number: 1,
+            size: 1,
+            ..Default::default()
+        }]);
+        assert_equal_time_identity_conflict(base.clone(), parts);
+
+        let mut transition = base;
+        transition.transitioned_object.tier = "tier-a".to_string();
+        assert_equal_time_identity_conflict(
+            object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string())),
+            transition,
+        );
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_accepts_internal_metadata_aliases() {
+        let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+        let mut rustfs_alias = base.clone();
+        rustfs_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+            "x-rustfs-internal-compression".to_string(),
+            "zstd".to_string(),
+        )]));
+        let mut minio_alias = base.clone();
+        minio_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+            "X-MINIO-INTERNAL-COMPRESSION".to_string(),
+            "zstd".to_string(),
+        )]));
+
+        let (_, idx) = resolve_latest_object_info_candidates(
+            vec![
+                LatestObjectInfoCandidate {
+                    info: Some(rustfs_alias),
+                    idx: 0,
+                    err: None,
+                },
+                LatestObjectInfoCandidate {
+                    info: Some(minio_alias),
+                    idx: 1,
+                    err: None,
+                },
+            ],
+            "bucket",
+            "object",
+            &ObjectOptions::default(),
+        )
+        .expect("same-value internal aliases should resolve");
+        assert_eq!(idx, 1);
+
+        let mut dual_alias = base.clone();
+        dual_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([
+            ("x-rustfs-internal-compression".to_string(), "zstd".to_string()),
+            ("x-minio-internal-compression".to_string(), "zstd".to_string()),
+        ]));
+        let mut single_alias = base;
+        single_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+            "x-rustfs-internal-compression".to_string(),
+            "zstd".to_string(),
+        )]));
+
+        let (_, idx) = resolve_latest_object_info_candidates(
+            vec![
+                LatestObjectInfoCandidate {
+                    info: Some(dual_alias),
+                    idx: 0,
+                    err: None,
+                },
+                LatestObjectInfoCandidate {
+                    info: Some(single_alias),
+                    idx: 1,
+                    err: None,
+                },
+            ],
+            "bucket",
+            "object",
+            &ObjectOptions::default(),
+        )
+        .expect("dual-key and single-key internal metadata should resolve");
+        assert_eq!(idx, 1);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_different_internal_metadata_alias_values() {
+        let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+        let mut rustfs_alias = base.clone();
+        rustfs_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+            "x-rustfs-internal-compression".to_string(),
+            "zstd".to_string(),
+        )]));
+        let mut minio_alias = base;
+        minio_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+            "x-minio-internal-compression".to_string(),
+            "snappy".to_string(),
+        )]));
+
+        assert_equal_time_identity_conflict(rustfs_alias, minio_alias);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_preserves_dynamic_internal_metadata_identity_case() {
+        for suffix_prefix in ["replication-reset-", "replication-delete-marker-version-"] {
+            let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+            let mut rustfs_alias = base.clone();
+            rustfs_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+                format!(
+                    "X-RUSTFS-INTERNAL-{}{suffix}",
+                    suffix_prefix.to_uppercase(),
+                    suffix = "arn:aws:s3:::Bucket"
+                ),
+                "value".to_string(),
+            )]));
+            let mut minio_alias = base.clone();
+            minio_alias.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+                format!("x-minio-internal-{suffix_prefix}arn:aws:s3:::Bucket"),
+                "value".to_string(),
+            )]));
+
+            let (_, idx) = resolve_latest_object_info_candidates(
+                vec![
+                    LatestObjectInfoCandidate {
+                        info: Some(rustfs_alias.clone()),
+                        idx: 0,
+                        err: None,
+                    },
+                    LatestObjectInfoCandidate {
+                        info: Some(minio_alias),
+                        idx: 1,
+                        err: None,
+                    },
+                ],
+                "bucket",
+                "object",
+                &ObjectOptions::default(),
+            )
+            .expect("dynamic internal aliases with the same target should resolve");
+            assert_eq!(idx, 1);
+
+            let mut different_target_case = base;
+            different_target_case.user_defined = std::sync::Arc::new(std::collections::HashMap::from([(
+                format!("x-minio-internal-{suffix_prefix}arn:aws:s3:::bucket"),
+                "value".to_string(),
+            )]));
+
+            assert_equal_time_identity_conflict(rustfs_alias, different_target_case);
+        }
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_conflicting_internal_metadata_aliases_in_one_candidate() {
+        let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+        let mut first = base.clone();
+        first.user_defined = std::sync::Arc::new(std::collections::HashMap::from([
+            ("x-rustfs-internal-compression".to_string(), "zstd".to_string()),
+            ("x-minio-internal-compression".to_string(), "snappy".to_string()),
+        ]));
+        let mut second = base;
+        second.user_defined = first.user_defined.clone();
+
+        assert_equal_time_identity_conflict(first, second);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_replication_identity_conflict() {
+        let base = object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+
+        let mut replication = base.clone();
+        replication.replication_status_internal = Some("PENDING".to_string());
+        replication.replication_status = ReplicationStatusType::Pending;
+        assert_equal_time_identity_conflict(base.clone(), replication);
+
+        let mut purge = base.clone();
+        purge.version_purge_status_internal = Some("PENDING".to_string());
+        purge.version_purge_status = VersionPurgeStatusType::Pending;
+        assert_equal_time_identity_conflict(base.clone(), purge);
+
+        let mut decision = base;
+        decision.replication_decision = "replicate".to_string();
+        assert_equal_time_identity_conflict(
+            object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string())),
+            decision,
+        );
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_rejects_none_vs_unix_epoch_mod_time() {
+        let mut without_mod_time = object_info_with_identity(0, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+        without_mod_time.mod_time = None;
+        let with_unix_epoch = object_info_with_identity(0, false, Uuid::from_u128(1), Some("etag-a".to_string()));
+
+        assert_equal_time_identity_conflict(without_mod_time, with_unix_epoch);
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_ignores_older_identity_conflicts() {
+        let latest = object_info_with_identity(20, false, Uuid::from_u128(1), Some("etag-latest".to_string()));
+        let mut older = object_info_with_identity(10, true, Uuid::from_u128(2), Some("etag-old".to_string()));
+        older.data_dir = Some(Uuid::from_u128(2));
+
+        let (info, idx) = resolve_latest_object_info_candidates(
+            vec![
+                LatestObjectInfoCandidate {
+                    info: Some(latest),
+                    idx: 0,
+                    err: None,
+                },
+                LatestObjectInfoCandidate {
+                    info: Some(older),
+                    idx: 9,
+                    err: None,
+                },
+            ],
+            "bucket",
+            "object",
+            &ObjectOptions::default(),
+        )
+        .expect("older identity divergence must not affect the latest candidate");
+
+        assert_eq!(idx, 0);
+        assert_eq!(
+            info.mod_time,
+            Some(OffsetDateTime::from_unix_timestamp(20).expect("operation should succeed"))
+        );
+    }
+
+    #[test]
+    fn resolve_latest_object_info_candidates_ignores_not_found_pools_when_resolving() {
+        let candidates = vec![
+            LatestObjectInfoCandidate {
+                info: Some(object_info_with_identity(10, false, Uuid::from_u128(1), Some("etag-a".to_string()))),
+                idx: 0,
+                err: None,
+            },
+            LatestObjectInfoCandidate {
+                info: None,
+                idx: 1,
+                err: Some(Error::ObjectNotFound("bucket".to_string(), "object".to_string())),
+            },
+        ];
+
+        let (info, idx) = resolve_latest_object_info_candidates(candidates, "bucket", "object", &ObjectOptions::default())
+            .expect("not-found pools must not block resolution of found candidates");
+
+        assert_eq!(idx, 0);
+        assert_eq!(info.version_id, Some(Uuid::from_u128(1)));
+    }
+
    #[test]
    fn resolve_latest_object_info_candidates_returns_non_not_found_error() {
        let err = resolve_latest_object_info_candidates(
@@ -12,10 +12,14 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

-use std::cmp::Ordering;
+use std::collections::HashMap;

 use crate::error::{Error, Result, StorageError, is_err_object_not_found, is_err_version_not_found};
 use crate::object_api::{ObjectInfo, ObjectOptions};
+use rustfs_utils::http::metadata_compat::{
+    SUFFIX_REPLICATION_DELETE_MARKER_VERSION_ARN_PREFIX, SUFFIX_REPLICATION_RESET_ARN_PREFIX,
+    strip_internal_prefix_preserving_case,
+};
 use rustfs_utils::path::decode_dir_object;
 use time::OffsetDateTime;

@@ -73,7 +77,7 @@ pub(super) fn resolve_rebalance_delete_from_all_pools_result(
    object: &str,
 ) -> Result<ObjectInfo> {
    result.map_err(|err| {
-        if err == Error::PreconditionFailed {
+        if matches!(&err, Error::PreconditionFailed | Error::PrefixAccessDenied(_, _)) {
            err
        } else {
            Error::other(format!("failed to delete rebalance source object {bucket}/{object}: {err}"))
@@ -86,7 +90,7 @@ fn is_ignorable_rebalance_delete_error(err: &Error) -> bool {
 }

 fn rebalance_delete_pool_error(pool_idx: usize, bucket: &str, object: &str, err: Error) -> Error {
-    if err == Error::PreconditionFailed {
+    if matches!(&err, Error::PreconditionFailed | Error::PrefixAccessDenied(_, _)) {
        err
    } else {
        Error::other(format!("pool {pool_idx} delete failed for {bucket}/{object}: {err}"))
@@ -137,37 +141,158 @@ pub(super) fn rebalance_disk_set_lookup_error(pool_idx: usize, set_idx: usize, p
    ))
 }

+fn latest_candidate_mod_time(candidate: &LatestObjectInfoCandidate) -> Option<OffsetDateTime> {
+    candidate
+        .info
+        .as_ref()
+        .map(|info| info.mod_time.unwrap_or(OffsetDateTime::UNIX_EPOCH))
+}
+
+fn same_transition_identity(left: &ObjectInfo, right: &ObjectInfo) -> bool {
+    left.transition_version_state == right.transition_version_state
+        && left.transitioned_object.name == right.transitioned_object.name
+        && left.transitioned_object.version_id == right.transitioned_object.version_id
+        && left.transitioned_object.tier == right.transitioned_object.tier
+        && left.transitioned_object.free_version == right.transitioned_object.free_version
+        && left.transitioned_object.status == right.transitioned_object.status
+}
+
+#[derive(PartialEq, Eq)]
+struct LatestUserDefinedIdentity {
+    internal: HashMap<String, String>,
+    other: HashMap<String, String>,
+}
+
+fn normalize_internal_identity_suffix(key: &str) -> Option<String> {
+    let suffix = strip_internal_prefix_preserving_case(key)?;
+
+    for dynamic_prefix in [
+        SUFFIX_REPLICATION_RESET_ARN_PREFIX,
+        SUFFIX_REPLICATION_DELETE_MARKER_VERSION_ARN_PREFIX,
+    ] {
+        let prefix_len = dynamic_prefix.len();
+        if let (Some(prefix), Some(remainder)) = (suffix.get(..prefix_len), suffix.get(prefix_len..))
+            && prefix.eq_ignore_ascii_case(dynamic_prefix)
+        {
+            return Some(format!("{dynamic_prefix}{remainder}"));
+        }
+    }
+
+    Some(suffix.to_lowercase())
+}
+
+fn normalize_user_defined_identity(user_defined: &HashMap<String, String>) -> Option<LatestUserDefinedIdentity> {
+    let mut identity = LatestUserDefinedIdentity {
+        internal: HashMap::with_capacity(user_defined.len()),
+        other: HashMap::with_capacity(user_defined.len()),
+    };
+
+    for (key, value) in user_defined {
+        if let Some(suffix) = normalize_internal_identity_suffix(key) {
+            if identity
+                .internal
+                .insert(suffix, value.clone())
+                .is_some_and(|previous| previous != *value)
+            {
+                return None;
+            }
+        } else {
+            identity.other.insert(key.clone(), value.clone());
+        }
+    }
+
+    Some(identity)
+}
+
+fn same_user_defined_identity(left: &ObjectInfo, right: &ObjectInfo) -> bool {
+    match (
+        normalize_user_defined_identity(&left.user_defined),
+        normalize_user_defined_identity(&right.user_defined),
+    ) {
+        (Some(left), Some(right)) => left == right,
+        _ => false,
+    }
+}
+
+/// Pool-specific erasure geometry is intentionally excluded: `get_object_info`
+/// returns each pool's own `data_blocks`/`parity_blocks`, so those values can
+/// differ for the same object version while the selected winner still carries
+/// the chosen pool's layout. `put_object_reader` is also intentionally
+/// excluded because it is a transient request handle that `ObjectInfo::clone`
+/// drops. Every other ObjectInfo field is part of the production-visible
+/// identity and must agree before the pool index can provide a deterministic
+/// tie-break.
+fn same_latest_object_info_identity(left: &ObjectInfo, right: &ObjectInfo) -> bool {
+    left.bucket == right.bucket
+        && left.name == right.name
+        && left.storage_class == right.storage_class
+        && left.mod_time == right.mod_time
+        && left.size == right.size
+        && left.actual_size == right.actual_size
+        && left.is_dir == right.is_dir
+        && same_user_defined_identity(left, right)
+        && left.user_tags == right.user_tags
+        && left.version_id == right.version_id
+        && left.data_dir == right.data_dir
+        && left.delete_marker == right.delete_marker
+        && same_transition_identity(left, right)
+        && left.restore_ongoing == right.restore_ongoing
+        && left.restore_expires == right.restore_expires
+        && left.parts == right.parts
+        && left.is_latest == right.is_latest
+        && left.content_type == right.content_type
+        && left.content_encoding == right.content_encoding
+        && left.expires == right.expires
+        && left.num_versions == right.num_versions
+        && left.successor_mod_time == right.successor_mod_time
+        && left.etag == right.etag
+        && left.inlined == right.inlined
+        && left.metadata_only == right.metadata_only
+        && left.version_only == right.version_only
+        && left.replication_status_internal == right.replication_status_internal
+        && left.replication_status == right.replication_status
+        && left.version_purge_status_internal == right.version_purge_status_internal
+        && left.version_purge_status == right.version_purge_status
+        && left.replication_decision == right.replication_decision
+        && left.checksum == right.checksum
+}
+
 pub(super) fn resolve_latest_object_info_candidates(
-    mut candidates: Vec<LatestObjectInfoCandidate>,
+    candidates: Vec<LatestObjectInfoCandidate>,
    bucket: &str,
    object: &str,
    opts: &ObjectOptions,
 ) -> Result<(ObjectInfo, usize)> {
-    candidates.sort_by(|a, b| {
-        let a_mod = if let Some(info) = &a.info {
-            info.mod_time.unwrap_or(OffsetDateTime::UNIX_EPOCH)
-        } else {
-            OffsetDateTime::UNIX_EPOCH
+    let latest_mod_time = candidates.iter().filter_map(latest_candidate_mod_time).max();
+
+    if let Some(latest_mod_time) = latest_mod_time {
+        let mut latest_candidates = candidates
+            .into_iter()
+            .filter(|candidate| latest_candidate_mod_time(candidate) == Some(latest_mod_time))
+            .collect::<Vec<_>>();
+
+        latest_candidates.sort_by_key(|candidate| std::cmp::Reverse(candidate.idx));
+
+        let Some(winner) = latest_candidates.first() else {
+            return Err(Error::ErasureReadQuorum);
+        };
+        let Some(winner_info) = winner.info.as_ref() else {
+            return Err(Error::ErasureReadQuorum);
        };

-        let b_mod = if let Some(info) = &b.info {
-            info.mod_time.unwrap_or(OffsetDateTime::UNIX_EPOCH)
-        } else {
-            OffsetDateTime::UNIX_EPOCH
-        };
-
-        if a_mod == b_mod {
-            return if a.idx < b.idx { Ordering::Greater } else { Ordering::Less };
+        if latest_candidates.iter().skip(1).any(|candidate| {
+            candidate
+                .info
+                .as_ref()
+                .is_none_or(|info| !same_latest_object_info_identity(winner_info, info))
+        }) {
+            return Err(Error::ErasureReadQuorum);
        }

-        b_mod.cmp(&a_mod)
-    });
+        return Ok((winner_info.clone(), winner.idx));
+    }

    for candidate in candidates {
-        if let Some(info) = candidate.info {
-            return Ok((info, candidate.idx));
-        }
-
        if let Some(err) = candidate.err
            && !is_err_object_not_found(&err)
            && !is_err_version_not_found(&err)
@@ -191,6 +316,18 @@ mod tests {
        assert_eq!(err, Error::PreconditionFailed);
    }

+    #[test]
+    fn rebalance_delete_result_preserves_prefix_access_denied() {
+        let err = resolve_rebalance_delete_from_all_pools_result(
+            Err(Error::PrefixAccessDenied("bucket".to_owned(), "object".to_owned())),
+            "bucket",
+            "object",
+        )
+        .expect_err("prefix access denial should remain structured");
+
+        assert_eq!(err, Error::PrefixAccessDenied("bucket".to_owned(), "object".to_owned()));
+    }
+
    #[test]
    fn rebalance_delete_pool_result_preserves_precondition_failed() {
        let err = resolve_rebalance_delete_from_all_pools_results(
@@ -205,4 +342,19 @@ mod tests {

        assert_eq!(err, Error::PreconditionFailed);
    }
+
+    #[test]
+    fn rebalance_delete_pool_result_preserves_prefix_access_denied() {
+        let err = resolve_rebalance_delete_from_all_pools_results(
+            vec![RebalanceDeletePoolResult {
+                pool_idx: 0,
+                result: Err(Error::PrefixAccessDenied("bucket".to_owned(), "object".to_owned())),
+            }],
+            "bucket",
+            "object",
+        )
+        .expect_err("prefix access denial should remain structured");
+
+        assert_eq!(err, Error::PrefixAccessDenied("bucket".to_owned(), "object".to_owned()));
+    }
 }
@@ -13,10 +13,10 @@
 // limitations under the License.

 use crate::heal::{
-    progress::HealProgress,
+    progress::{HealProgress, add_bytes, increment_counter},
    resume::{
-        CheckpointManager, ReplacementTargetIdentity, ResumeManager, ResumeUtils, compose_key,
-        replacement_target_identities_match,
+        CheckpointManager, CheckpointObjectOutcome, CheckpointObjectOutcomeRecord, ReplacementTargetIdentity, ResumeManager,
+        ResumeUtils, compose_key, replacement_target_identities_match,
    },
    storage::{HealStorageAPI, next_heal_listing_token},
    task::{demote_to_debug_when, is_missing_object_dir_heal_result, take_failure_log_sample},
@@ -410,6 +410,9 @@ impl ErasureSetHealer {
                && state.successful_objects == 0
                && state.failed_objects == 0
                && state.skipped_objects == 0
+                && state.skipped_new_versions == 0
+                && state.skipped_ilm_expired == 0
+                && state.processed_bytes == 0
            {
                // schedule_retry persists the authoritative resume reset before
                // resetting the checkpoint. Reapply the checkpoint reset after
@@ -474,6 +477,23 @@ impl ErasureSetHealer {

        // 2. initialize progress
        self.initialize_progress(buckets, &state).await;
+        let (baseline_known, baseline_count, baseline_size, baseline_generation) = {
+            let baseline = self.progress.read().await;
+            (
+                baseline.baseline_known,
+                baseline.objects_total_count,
+                baseline.objects_total_size,
+                baseline.baseline_generation,
+            )
+        };
+        if baseline_known {
+            resume_manager
+                .set_progress_baseline(baseline_count, baseline_size, baseline_generation)
+                .await?;
+            checkpoint_manager
+                .set_progress_baseline(baseline_count, baseline_size, baseline_generation)
+                .await?;
+        }

        // 3. continue from checkpoint
        let current_bucket_index = checkpoint.current_bucket_index;
@@ -483,12 +503,66 @@ impl ErasureSetHealer {
        let mut successful_objects = state.successful_objects;
        let mut failed_objects = state.failed_objects;
        let mut skipped_objects = state.skipped_objects;
+        let checkpoint_has_progress = checkpoint.baseline_known
+            || checkpoint.successful_objects > 0
+            || checkpoint.failed_object_count > 0
+            || checkpoint.skipped_object_count > 0
+            || checkpoint.skipped_new_versions > 0
+            || checkpoint.skipped_ilm_expired > 0
+            || checkpoint.processed_bytes > 0
+            || checkpoint.total_objects > 0
+            || checkpoint.total_bytes > 0
+            || checkpoint.baseline_generation.is_some()
+            || checkpoint.counter_unknown;
+        let checkpoint_generation_mismatch = checkpoint.baseline_known && checkpoint.baseline_generation != baseline_generation;
+        let mut restored_counter_unknown = state.counter_unknown || checkpoint.counter_unknown;
+        if checkpoint_has_progress {
+            successful_objects = checkpoint.successful_objects;
+            failed_objects = checkpoint.failed_object_count;
+            skipped_objects = checkpoint.skipped_object_count;
+            let restored_processed_objects = successful_objects
+                .checked_add(failed_objects)
+                .and_then(|value| value.checked_add(skipped_objects))
+                .and_then(|value| value.checked_add(checkpoint.skipped_new_versions))
+                .and_then(|value| value.checked_add(checkpoint.skipped_ilm_expired));
+            let checkpoint_counter_overflow = restored_processed_objects.is_none();
+            restored_counter_unknown |= checkpoint_counter_overflow;
+            processed_objects = restored_processed_objects.unwrap_or(u64::MAX);
+            let mut progress = self.progress.write().await;
+            progress.objects_scanned = processed_objects;
+            progress.objects_healed = successful_objects;
+            progress.objects_failed = failed_objects;
+            progress.skipped_objects = skipped_objects;
+            progress.skipped_new_versions = checkpoint.skipped_new_versions;
+            progress.skipped_ilm_expired = checkpoint.skipped_ilm_expired;
+            if checkpoint.baseline_known && !checkpoint_generation_mismatch {
+                progress.objects_total_count = checkpoint.total_objects;
+                progress.objects_total_size = checkpoint.total_bytes;
+                progress.baseline_generation = checkpoint.baseline_generation;
+                progress.baseline_known = true;
+            }
+            progress.bytes_processed = checkpoint.processed_bytes;
+            progress.counter_unknown = state.counter_unknown || checkpoint.counter_unknown;
+            progress.refresh_progress_percentage();
+            if checkpoint_generation_mismatch || checkpoint_counter_overflow || progress.counter_unknown {
+                progress.mark_unknown();
+            }
+        }
+        if checkpoint_generation_mismatch {
+            restored_counter_unknown = true;
+        }
+        if restored_counter_unknown {
+            checkpoint_manager.mark_counter_unknown().await?;
+            resume_manager.mark_counter_unknown().await?;
+        }
        let mut failed_buckets = 0u64;

        // 4. process remaining buckets
        for (bucket_idx, bucket) in buckets.iter().enumerate().skip(current_bucket_index) {
            // check if completed
            if state.completed_buckets.contains(bucket) {
+                checkpoint_manager.complete_bucket(bucket_idx.saturating_add(1)).await?;
+                current_object_index = 0;
                continue;
            }

@@ -516,13 +590,42 @@ impl ErasureSetHealer {
                return bucket_result;
            }

-            // update checkpoint position
-            checkpoint_manager.update_position(bucket_idx, current_object_index).await?;
-
            // update progress
-            resume_manager
-                .update_progress(processed_objects, successful_objects, failed_objects, skipped_objects)
+            let progress_snapshot = self.progress.read().await;
+            let bytes_processed = progress_snapshot.bytes_processed;
+            let skipped_new_versions = progress_snapshot.skipped_new_versions;
+            let skipped_ilm_expired = progress_snapshot.skipped_ilm_expired;
+            let counter_unknown = progress_snapshot.counter_unknown;
+            drop(progress_snapshot);
+            // The checkpoint is the recovery authority for object progress.
+            // Publish its counters and fence before the resume summary so a
+            // crash between the two stores cannot make recovery select newer
+            // summary bytes with an older checkpoint ledger.
+            if counter_unknown {
+                checkpoint_manager.mark_counter_unknown().await?;
+            }
+            checkpoint_manager
+                .update_progress(successful_objects, failed_objects, skipped_objects, bytes_processed)
                .await?;
+            checkpoint_manager
+                .set_skipped_version_counts(skipped_new_versions, skipped_ilm_expired)
+                .await?;
+            checkpoint_manager.update_position(bucket_idx, current_object_index).await?;
+            resume_manager
+                .update_progress_with_bytes(
+                    processed_objects,
+                    successful_objects,
+                    failed_objects,
+                    skipped_objects,
+                    bytes_processed,
+                )
+                .await?;
+            resume_manager
+                .set_skipped_version_counts(skipped_new_versions, skipped_ilm_expired)
+                .await?;
+            if counter_unknown {
+                resume_manager.mark_counter_unknown().await?;
+            }

            // check cancel status
            if self.cancel_token.is_cancelled() {
@@ -542,6 +645,7 @@ impl ErasureSetHealer {
            match bucket_result {
                Ok(_) => {
                    resume_manager.complete_bucket(bucket).await?;
+                    checkpoint_manager.complete_bucket(bucket_idx.saturating_add(1)).await?;
                    debug!(
                        target: "rustfs::heal::erasure_healer",
                        event = EVENT_HEAL_ERASURE_BUCKET_STATE,
@@ -567,7 +671,9 @@ impl ErasureSetHealer {
                        error = %e,
                        "Erasure set bucket heal failed"
                    );
-                    // continue to next bucket, do not interrupt the whole process
+                    // A single durable cursor and ledger cannot safely preserve
+                    // this bucket while processing a later one.
+                    break;
                }
            }

@@ -775,20 +881,49 @@ impl ErasureSetHealer {

                // Per-version dedup identity — the single canonical key.
                let key = compose_key(&item.name, item.version_id.as_deref());
-                if checkpoint.processed_objects.contains(&key) || checkpoint.skipped_objects.contains(&key) {
+                if checkpoint.processed_objects.contains(&key)
+                    || checkpoint.failed_objects.contains(&key)
+                    || checkpoint.skipped_objects.contains(&key)
+                {
                    continue;
                }

                if should_skip_new_version(item.mod_time_unix_nanos, started_at_secs) {
-                    checkpoint_manager.add_processed_object(key).await?;
-                    *processed_objects = processed_objects.saturating_add(1);
+                    let counter_ok = increment_counter(processed_objects);
                    completed_in_page = completed_in_page.saturating_add(1);
                    counter!("rustfs_heal_skipped_new_versions_total").increment(1);
-                    {
+                    let (outcome_record, counter_unknown) = {
                        let mut progress = self.progress.write().await;
                        progress.record_skipped_new_version();
                        progress.set_current_object(Some(format!("skipped_new: {bucket}/{}", item.name)));
-                        progress.update_progress(*processed_objects, *successful_objects, *failed_objects, bytes_processed);
+                        progress.update_object_progress(
+                            *processed_objects,
+                            *successful_objects,
+                            *failed_objects,
+                            *skipped_objects,
+                            bytes_processed,
+                        );
+                        if !counter_ok {
+                            progress.mark_unknown();
+                        }
+                        (
+                            CheckpointObjectOutcomeRecord {
+                                object: key,
+                                outcome: CheckpointObjectOutcome::Processed,
+                                successful: progress.objects_healed,
+                                failed: progress.objects_failed,
+                                skipped: progress.skipped_objects,
+                                bytes: progress.bytes_processed,
+                                skipped_new_versions: progress.skipped_new_versions,
+                                skipped_ilm_expired: progress.skipped_ilm_expired,
+                                counter_unknown: progress.counter_unknown,
+                            },
+                            progress.counter_unknown,
+                        )
+                    };
+                    checkpoint_manager.record_object_outcome(outcome_record).await?;
+                    if counter_unknown {
+                        resume_manager.mark_counter_unknown().await?;
                    }
                    debug!(
                        target: "rustfs::heal::erasure_healer",
@@ -820,15 +955,41 @@ impl ErasureSetHealer {
                        )
                        .await?
                {
-                    checkpoint_manager.add_processed_object(key).await?;
-                    *processed_objects = processed_objects.saturating_add(1);
+                    let counter_ok = increment_counter(processed_objects);
                    completed_in_page = completed_in_page.saturating_add(1);
                    counter!("rustfs_heal_skipped_ilm_expired_total").increment(1);
-                    {
+                    let (outcome_record, counter_unknown) = {
                        let mut progress = self.progress.write().await;
                        progress.record_skipped_ilm_expired();
                        progress.set_current_object(Some(format!("skipped_ilm: {bucket}/{}", item.name)));
-                        progress.update_progress(*processed_objects, *successful_objects, *failed_objects, bytes_processed);
+                        progress.update_object_progress(
+                            *processed_objects,
+                            *successful_objects,
+                            *failed_objects,
+                            *skipped_objects,
+                            bytes_processed,
+                        );
+                        if !counter_ok {
+                            progress.mark_unknown();
+                        }
+                        (
+                            CheckpointObjectOutcomeRecord {
+                                object: key,
+                                outcome: CheckpointObjectOutcome::Processed,
+                                successful: progress.objects_healed,
+                                failed: progress.objects_failed,
+                                skipped: progress.skipped_objects,
+                                bytes: progress.bytes_processed,
+                                skipped_new_versions: progress.skipped_new_versions,
+                                skipped_ilm_expired: progress.skipped_ilm_expired,
+                                counter_unknown: progress.counter_unknown,
+                            },
+                            progress.counter_unknown,
+                        )
+                    };
+                    checkpoint_manager.record_object_outcome(outcome_record).await?;
+                    if counter_unknown {
+                        resume_manager.mark_counter_unknown().await?;
                    }
                    debug!(
                        target: "rustfs::heal::erasure_healer",
@@ -954,11 +1115,11 @@ impl ErasureSetHealer {

            while let Some((key, object, version_id, result)) = page_tasks.next().await {
                let (object_size, result) = result;
-                match result {
+                let mut telemetry_unknown = false;
+                let checkpoint_outcome = match result {
                    Ok(true) => {
-                        *successful_objects += 1;
-                        bytes_processed = bytes_processed.saturating_add(object_size);
-                        checkpoint_manager.add_processed_object(key).await?;
+                        telemetry_unknown |= !increment_counter(successful_objects);
+                        telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
                        debug!(
                            target: "rustfs::heal::erasure_healer",
                            event = EVENT_HEAL_ERASURE_OBJECT_STATE,
@@ -971,11 +1132,11 @@ impl ErasureSetHealer {
                            state = "healed",
                            "Erasure set object healed"
                        );
+                        CheckpointObjectOutcome::Processed
                    }
                    Ok(false) => {
-                        checkpoint_manager.add_processed_object(key).await?;
-                        *successful_objects += 1;
-                        bytes_processed = bytes_processed.saturating_add(object_size);
+                        telemetry_unknown |= !increment_counter(successful_objects);
+                        telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
                        debug!(
                            target: "rustfs::heal::erasure_healer",
                            event = EVENT_HEAL_ERASURE_OBJECT_STATE,
@@ -988,12 +1149,12 @@ impl ErasureSetHealer {
                            state = "missing_treated_as_ok",
                            "Erasure set missing object treated as ok"
                        );
+                        CheckpointObjectOutcome::Processed
                    }
                    Err(err @ Error::TaskCancelled) | Err(err @ Error::TaskTimeout) => return Err(err),
                    Err(Error::TransientSkip { message }) => {
-                        *skipped_objects += 1;
-                        bytes_processed = bytes_processed.saturating_add(object_size);
-                        checkpoint_manager.add_skipped_object(key).await?;
+                        telemetry_unknown |= !increment_counter(skipped_objects);
+                        telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
                        demote_to_debug_when!(!take_failure_log_sample(&mut transient_skip_samples_logged), warn, target: "rustfs::heal::erasure_healer", {
                            event = EVENT_HEAL_ERASURE_OBJECT_STATE,
                            component = LOG_COMPONENT_HEAL,
@@ -1006,11 +1167,11 @@ impl ErasureSetHealer {
                            error = %message,
                            "Erasure set object heal skipped due to transient error"
                        });
+                        CheckpointObjectOutcome::Skipped
                    }
                    Err(err) => {
-                        *failed_objects += 1;
-                        bytes_processed = bytes_processed.saturating_add(object_size);
-                        checkpoint_manager.add_failed_object(key).await?;
+                        telemetry_unknown |= !increment_counter(failed_objects);
+                        telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
                        demote_to_debug_when!(!take_failure_log_sample(&mut failure_samples_logged), warn, target: "rustfs::heal::erasure_healer", {
                            event = EVENT_HEAL_ERASURE_OBJECT_STATE,
                            component = LOG_COMPONENT_HEAL,
@@ -1023,15 +1184,43 @@ impl ErasureSetHealer {
                            error = %err,
                            "Erasure set object heal failed"
                        });
+                        CheckpointObjectOutcome::Failed
                    }
-                }
+                };

-                *processed_objects += 1;
+                telemetry_unknown |= !increment_counter(processed_objects);
                completed_in_page += 1;
-                {
+                let (outcome_record, counter_unknown) = {
                    let mut progress = self.progress.write().await;
                    progress.set_current_object(Some(format!("{bucket}/{object}")));
-                    progress.update_progress(*processed_objects, *successful_objects, *failed_objects, bytes_processed);
+                    progress.update_object_progress(
+                        *processed_objects,
+                        *successful_objects,
+                        *failed_objects,
+                        *skipped_objects,
+                        bytes_processed,
+                    );
+                    if telemetry_unknown {
+                        progress.mark_unknown();
+                    }
+                    (
+                        CheckpointObjectOutcomeRecord {
+                            object: key,
+                            outcome: checkpoint_outcome,
+                            successful: progress.objects_healed,
+                            failed: progress.objects_failed,
+                            skipped: progress.skipped_objects,
+                            bytes: progress.bytes_processed,
+                            skipped_new_versions: progress.skipped_new_versions,
+                            skipped_ilm_expired: progress.skipped_ilm_expired,
+                            counter_unknown: progress.counter_unknown,
+                        },
+                        progress.counter_unknown,
+                    )
+                };
+                checkpoint_manager.record_object_outcome(outcome_record).await?;
+                if counter_unknown {
+                    resume_manager.mark_counter_unknown().await?;
                }

                if completed_in_page.is_multiple_of(100) {
@@ -1041,16 +1230,22 @@ impl ErasureSetHealer {

            *current_object_index = global_obj_idx;

-            // Persist the authoritative cursor FIRST (points at the next page
-            // boundary), then prune the per-version dedup sets. Both are
-            // idempotent under crash: heal_object re-heals safely.
-            let next_cursor = if is_truncated { next_token.clone() } else { None };
-            resume_manager.set_resume_cursor(next_cursor.clone()).await?;
-            checkpoint_manager.complete_page(bucket_index, *current_object_index).await?;
+            // Persist the checkpoint ledger and page position before exposing
+            // the next resume cursor. A crash before cursor publication keeps
+            // the page identities available for exact-once replay.
+            checkpoint_manager.advance_page(bucket_index, *current_object_index).await?;
            // Check if there are more pages
            if !is_truncated {
                break;
            }
+            continuation_token = next_heal_listing_token(bucket, "", next_token, is_truncated)?;
+            if continuation_token.is_none() {
+                // A truncated page without a continuation token is terminal.
+                // Retain its ledger until bucket completion is durable.
+                break;
+            }
+            resume_manager.set_resume_cursor(continuation_token.clone()).await?;
+            checkpoint_manager.prune_completed_page().await?;

            // Anti-loop guard: an empty page reported as truncated cannot advance
            // the cursor (there is no last identity to move past), so treat it as a
@@ -1069,12 +1264,6 @@ impl ErasureSetHealer {
                )));
            }
            previous_page_last = page_last;
-
-            continuation_token = next_heal_listing_token(bucket, "", next_token, is_truncated)?;
-            if continuation_token.is_none() {
-                // Truncated but no continuation token: treat as end of listing.
-                break;
-            }
        }

        Ok(())
@@ -1083,10 +1272,66 @@ impl ErasureSetHealer {
    /// initialize progress tracking
    async fn initialize_progress(&self, _buckets: &[String], state: &crate::heal::resume::ResumeState) {
        let mut progress = self.progress.write().await;
-        progress.objects_scanned = state.total_objects;
+        let existing_baseline = (
+            progress.objects_total_count,
+            progress.objects_total_size,
+            progress.baseline_generation,
+            progress.progress_state,
+            progress.baseline_known,
+        );
+        let baseline_generation_mismatch =
+            state.baseline_known && existing_baseline.4 && state.baseline_generation != existing_baseline.2;
+        let use_persisted_baseline = state.baseline_known && !baseline_generation_mismatch;
+        progress.objects_scanned = state.processed_objects;
        progress.objects_healed = state.successful_objects;
        progress.objects_failed = state.failed_objects;
-        progress.bytes_processed = 0; // Resume state tracks object counts, not byte counters.
+        progress.skipped_objects = state.skipped_objects;
+        progress.skipped_new_versions = state.skipped_new_versions;
+        progress.skipped_ilm_expired = state.skipped_ilm_expired;
+        progress.bytes_processed = state.processed_bytes;
+        progress.counter_unknown = state.counter_unknown;
+        if use_persisted_baseline
+            || existing_baseline.0 > 0
+            || existing_baseline.1 > 0
+            || existing_baseline.2.is_some()
+            || existing_baseline.4
+        {
+            progress.objects_total_count = if use_persisted_baseline {
+                state.total_objects
+            } else {
+                existing_baseline.0
+            };
+            progress.objects_total_size = if use_persisted_baseline {
+                state.total_bytes
+            } else {
+                existing_baseline.1
+            };
+            progress.baseline_generation = if use_persisted_baseline {
+                state.baseline_generation
+            } else {
+                existing_baseline.2
+            };
+            progress.baseline_known = use_persisted_baseline
+                || existing_baseline.0 > 0
+                || existing_baseline.1 > 0
+                || existing_baseline.2.is_some()
+                || existing_baseline.4;
+        }
+        progress.progress_state = if use_persisted_baseline
+            || existing_baseline.0 > 0
+            || existing_baseline.1 > 0
+            || existing_baseline.2.is_some()
+            || existing_baseline.4
+        {
+            crate::heal::progress::HealProgressState::Running
+        } else {
+            crate::heal::progress::HealProgressState::Indeterminate
+        };
+        if baseline_generation_mismatch || state.counter_unknown {
+            progress.mark_unknown();
+        }
+        progress.ledger_complete = false;
+        progress.refresh_progress_percentage();
        progress.start_time = UNIX_EPOCH.checked_add(Duration::from_secs(state.start_time));
        progress.last_update_time = UNIX_EPOCH.checked_add(Duration::from_secs(state.last_update));
        progress.set_current_object(state.current_object.clone());
@@ -1264,8 +1509,8 @@ mod resume_loop_tests {
    };
    use crate::heal::progress::HealProgress;
    use crate::heal::resume::{
-        CheckpointManager, RESUME_CHECKPOINT_FILE, ReplacementTargetIdentity, ResumeDeleteFailure, ResumeManager, ResumeUtils,
-        compose_key,
+        CheckpointManager, CheckpointObjectOutcome, CheckpointObjectOutcomeRecord, RESUME_CHECKPOINT_FILE,
+        ReplacementTargetIdentity, ResumeDeleteFailure, ResumeManager, ResumeUtils, compose_key,
    };
    use crate::heal::storage::{HealLifecycleExpiryContext, HealListItem, HealObjectInfo, HealStorageAPI};
    use crate::heal::storage_api::status::BucketInfo;
@@ -1405,6 +1650,7 @@ mod resume_loop_tests {
        list_include_lifecycle_object_info: Mutex<Vec<bool>>,
        replacement_target_identity_sequences: Mutex<VecDeque<Vec<ReplacementTargetIdentity>>>,
        fail_listing: AtomicBool,
+        fail_listing_buckets: Mutex<HashSet<String>>,
    }

    impl FakeStorage {
@@ -1441,6 +1687,9 @@ mod resume_loop_tests {
        fn fail_listing(&self) {
            self.fail_listing.store(true, Ordering::SeqCst);
        }
+        fn fail_bucket_listing(&self, bucket: &str) {
+            self.fail_listing_buckets.lock().unwrap().insert(bucket.to_string());
+        }
    }

    #[async_trait::async_trait]
@@ -1531,7 +1780,7 @@ mod resume_loop_tests {
        }
        async fn list_objects_for_heal_page(
            &self,
-            _bucket: &str,
+            bucket: &str,
            _prefix: &str,
            continuation_token: Option<&str>,
            include_lifecycle_object_info: bool,
@@ -1540,7 +1789,7 @@ mod resume_loop_tests {
                .lock()
                .unwrap()
                .push(include_lifecycle_object_info);
-            if self.fail_listing.load(Ordering::SeqCst) {
+            if self.fail_listing.load(Ordering::SeqCst) || self.fail_listing_buckets.lock().unwrap().contains(bucket) {
                return Err(Error::other("injected listing failure"));
            }
            let key = continuation_token.map(str::to_string);
@@ -1918,6 +2167,49 @@ mod resume_loop_tests {
        assert!(state.completed_buckets.is_empty(), "the failed bucket must remain resumable");
    }

+    #[tokio::test]
+    async fn bucket_failure_stops_before_a_later_bucket_checkpoint() {
+        let env = make_env().await;
+        let task_id = ResumeUtils::generate_task_id();
+        let buckets = vec!["a".to_string(), "b".to_string()];
+        let resume = ResumeManager::new(
+            env.healer.disk.clone(),
+            task_id.clone(),
+            "erasure_set".to_string(),
+            "pool_0_set_0".to_string(),
+            buckets.clone(),
+        )
+        .await
+        .unwrap();
+        let checkpoint = CheckpointManager::new(env.healer.disk.clone(), task_id.clone())
+            .await
+            .unwrap();
+        env.storage.fail_bucket_listing("a");
+        for _ in 0..3 {
+            assert!(resume.schedule_retry().await.unwrap());
+        }
+
+        env.healer
+            .execute_heal_with_resume(&buckets, "pool_0_set_0", &resume, &checkpoint)
+            .await
+            .expect_err("the first bucket failure must keep the pass incomplete");
+        let persisted = checkpoint.get_checkpoint().await;
+        assert_eq!(persisted.current_bucket_index, 0);
+        assert!(resume.get_state().await.completed_buckets.is_empty());
+
+        let resumed = ResumeManager::load_from_disk(env.healer.disk.clone(), &task_id)
+            .await
+            .unwrap();
+        let checkpoint = CheckpointManager::load_from_disk(env.healer.disk.clone(), &task_id)
+            .await
+            .unwrap();
+        env.healer
+            .execute_heal_with_resume(&buckets, "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect_err("recovery must retry the earlier failed bucket");
+        assert!(!resumed.get_state().await.completed);
+    }
+
    #[tokio::test]
    async fn completed_resume_state_is_not_selected_for_a_new_heal() {
        let env = make_env().await;
@@ -2107,8 +2399,175 @@ mod resume_loop_tests {
        let mut names: Vec<String> = env.storage.calls().into_iter().map(|(n, _)| n).collect();
        names.sort();
        assert_eq!(names, vec!["a", "b", "c", "d"], "every object exactly once, none dropped/doubled");
-        // Final page not truncated => cursor cleared.
-        assert_eq!(env.resume.resume_cursor().await, None);
+        // Keep the final page cursor until the outer loop durably completes the
+        // bucket, so a crash can replay only this page against its identities.
+        assert_eq!(env.resume.resume_cursor().await, Some("t1".to_string()));
+    }
+
+    #[tokio::test]
+    async fn persisted_failure_waits_for_the_bounded_retry_after_page_replay() {
+        let env = make_env().await;
+        env.storage.set_page(
+            None,
+            Page {
+                items: vec![item("object", Some("v1"), false)],
+                next: None,
+                truncated: false,
+            },
+        );
+        env.checkpoint
+            .record_object_outcome(CheckpointObjectOutcomeRecord {
+                object: compose_key("object", Some("v1")),
+                outcome: CheckpointObjectOutcome::Failed,
+                successful: 0,
+                failed: 1,
+                skipped: 0,
+                bytes: 0,
+                skipped_new_versions: 0,
+                skipped_ilm_expired: 0,
+                counter_unknown: false,
+            })
+            .await
+            .unwrap();
+        env.checkpoint.advance_page(0, 1).await.unwrap();
+
+        let resumed = ResumeManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        let checkpoint = CheckpointManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+
+        env.healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect_err("the persisted failure must schedule a bounded retry");
+        assert!(
+            env.storage.calls().is_empty(),
+            "the failed identity must not be repeated in the same pass"
+        );
+
+        env.healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect("the bounded retry must heal the object");
+        assert_eq!(env.storage.calls(), vec![("object".to_string(), Some("v1".to_string()))]);
+        let state = resumed.get_state().await;
+        assert_eq!(state.successful_objects, 1);
+        assert_eq!(state.failed_objects, 0);
+    }
+
+    #[tokio::test]
+    async fn final_page_crash_replays_only_the_retained_page_identities() {
+        let env = make_env().await;
+        env.storage.set_page(
+            None,
+            Page {
+                items: vec![item("first", Some("v1"), false)],
+                next: Some("final-page".to_string()),
+                truncated: true,
+            },
+        );
+        env.storage.set_page(
+            Some("final-page"),
+            Page {
+                items: vec![item("last", Some("v1"), false)],
+                next: None,
+                truncated: false,
+            },
+        );
+
+        let (processed, successful, failed, skipped, result) = run(&env).await;
+        result.expect("the bucket pass must finish before the simulated crash");
+        assert_eq!((processed, successful, failed, skipped), (2, 2, 0, 0));
+
+        let resumed = ResumeManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        let checkpoint = CheckpointManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        env.healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect("the retained final-page ledger must make recovery exact");
+
+        assert_eq!(
+            env.storage.calls(),
+            vec![
+                ("first".to_string(), Some("v1".to_string())),
+                ("last".to_string(), Some("v1".to_string()))
+            ]
+        );
+        let state = resumed.get_state().await;
+        assert_eq!(state.successful_objects, 2);
+        assert_eq!(state.processed_objects, 2);
+    }
+
+    #[tokio::test]
+    async fn truncated_page_without_token_retains_its_replay_ledger() {
+        let env = make_env().await;
+        env.storage.set_page(
+            None,
+            Page {
+                items: vec![item("object", Some("v1"), false)],
+                next: None,
+                truncated: true,
+            },
+        );
+
+        let (processed, successful, failed, skipped, result) = run(&env).await;
+        result.expect("the tokenless truncated page is a terminal page");
+        assert_eq!((processed, successful, failed, skipped), (1, 1, 0, 0));
+
+        let resumed = ResumeManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        let checkpoint = CheckpointManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        env.healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect("terminal-page recovery must not replay a durable identity");
+
+        assert_eq!(env.storage.calls(), vec![("object".to_string(), Some("v1".to_string()))]);
+        let state = resumed.get_state().await;
+        assert_eq!(state.successful_objects, 1);
+        assert_eq!(state.processed_objects, 1);
+    }
+
+    #[tokio::test]
+    async fn completed_bucket_reconciles_its_final_page_checkpoint_after_crash() {
+        let env = make_env().await;
+        env.storage.set_page(
+            None,
+            Page {
+                items: vec![item("object", Some("v1"), false)],
+                next: None,
+                truncated: false,
+            },
+        );
+
+        let (_, _, _, _, result) = run(&env).await;
+        result.expect("the bucket pass must finish before the simulated crash");
+        env.resume.complete_bucket("b").await.unwrap();
+
+        let resumed = ResumeManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        let checkpoint = CheckpointManager::load_from_disk(env.healer.disk.clone(), &env.task_id)
+            .await
+            .unwrap();
+        env.healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &resumed, &checkpoint)
+            .await
+            .expect("recovery must finish the checkpoint transition without replaying the bucket");
+
+        assert_eq!(env.storage.calls(), vec![("object".to_string(), Some("v1".to_string()))]);
+        let checkpoint = checkpoint.get_checkpoint().await;
+        assert_eq!(checkpoint.current_bucket_index, 1);
+        assert!(checkpoint.processed_objects.is_empty());
    }

    #[tokio::test]
@@ -2011,34 +2011,11 @@ impl HealManager {
            return None;
        }

-        let mut snapshot = HealProgress::default();
+        let mut progresses = Vec::with_capacity(active_tasks.len());
        for task in active_tasks {
-            let progress = task.get_progress().await;
-            snapshot.objects_scanned = snapshot.objects_scanned.saturating_add(progress.objects_scanned);
-            snapshot.objects_healed = snapshot.objects_healed.saturating_add(progress.objects_healed);
-            snapshot.objects_failed = snapshot.objects_failed.saturating_add(progress.objects_failed);
-            snapshot.skipped_new_versions = snapshot.skipped_new_versions.saturating_add(progress.skipped_new_versions);
-            snapshot.skipped_ilm_expired = snapshot.skipped_ilm_expired.saturating_add(progress.skipped_ilm_expired);
-            snapshot.objects_total_count = snapshot.objects_total_count.saturating_add(progress.objects_total_count);
-            snapshot.objects_total_size = snapshot.objects_total_size.saturating_add(progress.objects_total_size);
-            snapshot.bytes_processed = snapshot.bytes_processed.saturating_add(progress.bytes_processed);
-            snapshot.start_time = match (snapshot.start_time, progress.start_time) {
-                (Some(current), Some(next)) => Some(current.min(next)),
-                (None, next) => next,
-                (current, None) => current,
-            };
-            snapshot.last_update_time = match (snapshot.last_update_time, progress.last_update_time) {
-                (Some(current), Some(next)) => Some(current.max(next)),
-                (None, next) => next,
-                (current, None) => current,
-            };
-            if progress.current_object.is_some() {
-                snapshot.current_object = progress.current_object;
-            }
+            progresses.push(task.get_progress().await);
        }
-        snapshot.refresh_progress_percentage();
-        snapshot.refresh_estimated_completion_time();
-        Some(snapshot)
+        crate::heal::progress::aggregate_heal_progress(progresses)
    }
 }

@@ -15,15 +15,91 @@
 use serde::{Deserialize, Serialize};
 use std::time::{Duration, SystemTime};

-#[derive(Debug, Default, Clone, Serialize, Deserialize)]
+pub(crate) fn stable_generation(parts: &[&[u8]]) -> u64 {
+    let mut hash = 0xcbf29ce484222325u64;
+    for part in parts {
+        for byte in (part.len() as u64).to_be_bytes().into_iter().chain(part.iter().copied()) {
+            hash ^= u64::from(byte);
+            hash = hash.wrapping_mul(0x100000001b3);
+        }
+    }
+    hash
+}
+
+#[cfg(test)]
+mod stable_generation_tests {
+    use super::stable_generation;
+
+    #[test]
+    fn stable_generation_has_a_fixed_vector() {
+        assert_eq!(stable_generation(&[b"rustfs", b"heal", b"42"]), 11_007_672_338_488_385_056);
+    }
+}
+
+pub(crate) fn increment_counter(counter: &mut u64) -> bool {
+    match counter.checked_add(1) {
+        Some(next) => {
+            *counter = next;
+            true
+        }
+        None => {
+            *counter = u64::MAX;
+            false
+        }
+    }
+}
+
+pub(crate) fn add_bytes(total: &mut u64, amount: u64) -> bool {
+    match total.checked_add(amount) {
+        Some(next) => {
+            *total = next;
+            true
+        }
+        None => {
+            *total = u64::MAX;
+            false
+        }
+    }
+}
+
+#[derive(Debug, Default, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(rename_all = "camelCase")]
+pub enum HealProgressKind {
+    #[default]
+    Unknown,
+    Stage,
+    ObjectSweep,
+}
+
+/// Whether the object ledger can produce a meaningful percentage.
+///
+/// A zero-valued baseline is not a completed scan: it means that no complete
+/// usage snapshot was available.  Keep this state explicit so callers do not
+/// mistake the legacy `0.0` wire value for a measured zero-percent result.
+#[derive(Debug, Default, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub enum HealProgressState {
+    #[default]
+    Unknown,
+    Indeterminate,
+    Running,
+    Completed,
+}
+
+#[derive(Debug, Default, Clone, PartialEq, Serialize, Deserialize)]
+#[serde(default, rename_all = "camelCase")]
 pub struct HealProgress {
+    #[serde(default)]
+    pub kind: HealProgressKind,
    /// Objects scanned
    pub objects_scanned: u64,
    /// Objects healed
    pub objects_healed: u64,
    /// Objects failed
    pub objects_failed: u64,
+    /// Versions deferred for a later retry pass.
+    #[serde(default)]
+    pub skipped_objects: u64,
    /// Versions skipped because they were written after this heal started
    pub skipped_new_versions: u64,
    /// Versions skipped because lifecycle already selected them for expiry
@@ -44,11 +120,38 @@ pub struct HealProgress {
    pub last_update_time: Option<SystemTime>,
    /// Estimated completion time
    pub estimated_completion_time: Option<SystemTime>,
+    /// Current stage number. Stage updates are intentionally independent from
+    /// the object ledger below.
+    #[serde(default)]
+    pub stage_current: u64,
+    /// Number of stages in the current task.
+    #[serde(default)]
+    pub stage_total: u64,
+    /// Explicitly distinguishes a missing usage baseline from measured 0%.
+    #[serde(default)]
+    pub progress_state: HealProgressState,
+    /// True only after the task's durable completion ledger was committed.
+    #[serde(default)]
+    pub ledger_complete: bool,
+    /// Generation of the usage snapshot used for the baseline, if available.
+    #[serde(default)]
+    pub baseline_generation: Option<u64>,
+    /// Whether the baseline was explicitly observed.  This is separate from
+    /// the counters so a known empty scope (0 objects, 0 bytes) is not
+    /// confused with a legacy snapshot that omitted the baseline fields.
+    #[serde(default)]
+    pub baseline_known: bool,
+    /// Internal telemetry fence set when an aggregate counter overflows or
+    /// becomes inconsistent.  It prevents a later refresh from fabricating a
+    /// percentage from the poisoned values.
+    #[serde(default)]
+    pub counter_unknown: bool,
 }

 impl HealProgress {
    pub fn new() -> Self {
        Self {
+            kind: HealProgressKind::Unknown,
            start_time: Some(SystemTime::now()),
            last_update_time: Some(SystemTime::now()),
            ..Default::default()
@@ -56,12 +159,87 @@ impl HealProgress {
    }

    pub fn update_progress(&mut self, scanned: u64, healed: u64, failed: u64, bytes: u64) {
+        self.update_object_sweep_progress(scanned, healed, failed, bytes);
+    }
+
+    pub fn update_object_sweep_progress(&mut self, scanned: u64, healed: u64, failed: u64, bytes: u64) {
+        self.kind = HealProgressKind::ObjectSweep;
        self.objects_scanned = scanned;
        self.objects_healed = healed;
        self.objects_failed = failed;
        self.bytes_processed = bytes;
        self.last_update_time = Some(SystemTime::now());

+        let explicit_skipped = match self.skipped_new_versions.checked_add(self.skipped_ilm_expired) {
+            Some(value) => value,
+            None => {
+                self.mark_unknown();
+                0
+            }
+        };
+        let skipped = healed
+            .checked_add(failed)
+            .and_then(|value| value.checked_add(explicit_skipped))
+            .and_then(|value| scanned.checked_sub(value))
+            .unwrap_or(0);
+        self.update_object_progress(scanned, healed, failed, skipped, bytes);
+    }
+
+    /// Update task stage progress without modifying object counters.
+    pub fn update_stage(&mut self, current: u64, total: u64) {
+        let object_sweep_active = matches!(self.kind, HealProgressKind::ObjectSweep);
+        if !object_sweep_active {
+            self.kind = HealProgressKind::Stage;
+        }
+        self.ledger_complete = false;
+        self.stage_current = current.min(total);
+        self.stage_total = total;
+        if object_sweep_active {
+            self.last_update_time = Some(SystemTime::now());
+            self.refresh_progress_percentage();
+            return;
+        }
+        self.progress_state = if total == 0 {
+            HealProgressState::Indeterminate
+        } else {
+            HealProgressState::Running
+        };
+        self.progress_percentage = if total == 0 {
+            0.0
+        } else {
+            (current as f64 / total as f64 * 100.0).min(100.0)
+        };
+        self.last_update_time = Some(SystemTime::now());
+    }
+
+    /// Update the disjoint object ledger. `scanned` is the number of terminal
+    /// object outcomes and must equal healed + failed + deferred skipped plus
+    /// the two terminal skip classes. Overflow is a corrupt/unknown counter
+    /// state, not a reason to abort a completed heal.
+    pub fn update_object_progress(&mut self, scanned: u64, healed: u64, failed: u64, skipped: u64, bytes: u64) {
+        self.kind = HealProgressKind::ObjectSweep;
+        // `skipped` is the transient/deferred class.  The two explicit skip
+        // counters are terminal classifications too, so include them in the
+        // same ledger without making callers maintain a second aggregate.
+        let outcomes = healed
+            .checked_add(failed)
+            .and_then(|value| value.checked_add(skipped))
+            .and_then(|value| value.checked_add(self.skipped_new_versions))
+            .and_then(|value| value.checked_add(self.skipped_ilm_expired));
+        self.objects_scanned = scanned;
+        self.objects_healed = healed;
+        self.objects_failed = failed;
+        self.skipped_objects = skipped;
+        self.bytes_processed = bytes;
+        self.last_update_time = Some(SystemTime::now());
+        self.ledger_complete = false;
+        if outcomes != Some(scanned) {
+            // Telemetry corruption must not abort a heal.  Preserve the
+            // counters for diagnostics, but do not derive a percentage from a
+            // double-counted or overflowing ledger.
+            self.mark_unknown();
+            return;
+        }
        self.refresh_progress_percentage();
        self.refresh_estimated_completion_time();
    }
@@ -69,50 +247,88 @@ impl HealProgress {
    pub fn set_total_baseline(&mut self, objects_total_count: u64, objects_total_size: u64) {
        self.objects_total_count = objects_total_count;
        self.objects_total_size = objects_total_size;
+        self.baseline_known = true;
        self.last_update_time = Some(SystemTime::now());
        self.refresh_progress_percentage();
        self.refresh_estimated_completion_time();
    }

+    pub fn set_total_baseline_with_generation(&mut self, objects_total_count: u64, objects_total_size: u64, generation: u64) {
+        self.baseline_generation = Some(generation);
+        self.set_total_baseline(objects_total_count, objects_total_size);
+    }
+
    pub fn record_skipped_new_version(&mut self) {
-        self.skipped_new_versions = self.skipped_new_versions.saturating_add(1);
+        let Some(next) = self.skipped_new_versions.checked_add(1) else {
+            self.mark_unknown();
+            return;
+        };
+        self.skipped_new_versions = next;
        self.last_update_time = Some(SystemTime::now());
        self.refresh_progress_percentage();
        self.refresh_estimated_completion_time();
    }

    pub fn record_skipped_ilm_expired(&mut self) {
-        self.skipped_ilm_expired = self.skipped_ilm_expired.saturating_add(1);
+        let Some(next) = self.skipped_ilm_expired.checked_add(1) else {
+            self.mark_unknown();
+            return;
+        };
+        self.skipped_ilm_expired = next;
        self.last_update_time = Some(SystemTime::now());
        self.refresh_progress_percentage();
        self.refresh_estimated_completion_time();
    }

-    fn completed_for_baseline(&self) -> u64 {
+    fn completed_for_baseline(&self) -> Option<u64> {
        self.objects_healed
-            .saturating_add(self.objects_failed)
-            .saturating_add(self.skipped_new_versions)
-            .saturating_add(self.skipped_ilm_expired)
+            .checked_add(self.objects_failed)?
+            .checked_add(self.skipped_objects)?
+            .checked_add(self.skipped_new_versions)?
+            .checked_add(self.skipped_ilm_expired)
    }

    pub(crate) fn refresh_progress_percentage(&mut self) {
+        if self.ledger_complete {
+            self.progress_state = HealProgressState::Completed;
+            self.progress_percentage = 100.0;
+            return;
+        }
+        if self.counter_unknown {
+            self.progress_state = HealProgressState::Unknown;
+            self.progress_percentage = 0.0;
+            return;
+        }
+        if !self.baseline_known {
+            self.progress_state = HealProgressState::Indeterminate;
+            self.progress_percentage = 0.0;
+            self.estimated_completion_time = None;
+            return;
+        }
        if self.objects_total_size > 0 {
            self.progress_percentage = ((self.bytes_processed as f64 / self.objects_total_size as f64) * 100.0).min(100.0);
+            self.progress_percentage = self.progress_percentage.min(99.999);
+            self.progress_state = HealProgressState::Running;
            return;
        }
        if self.objects_total_count > 0 {
-            let completed = self.completed_for_baseline();
+            let Some(completed) = self.completed_for_baseline() else {
+                self.progress_state = HealProgressState::Unknown;
+                self.progress_percentage = 0.0;
+                return;
+            };
            self.progress_percentage = ((completed as f64 / self.objects_total_count as f64) * 100.0).min(100.0);
+            self.progress_percentage = self.progress_percentage.min(99.999);
+            self.progress_state = HealProgressState::Running;
            return;
        }
-
-        let total = self
-            .objects_scanned
-            .saturating_add(self.objects_healed)
-            .saturating_add(self.objects_failed);
-        if total > 0 {
-            self.progress_percentage = (self.objects_healed as f64 / total as f64) * 100.0;
+        if self.baseline_known {
+            self.progress_state = HealProgressState::Running;
+            self.progress_percentage = 0.0;
+            return;
        }
+        self.progress_state = HealProgressState::Indeterminate;
+        self.progress_percentage = 0.0;
    }

    pub fn set_current_object(&mut self, object: Option<String>) {
@@ -125,7 +341,11 @@ impl HealProgress {
            self.estimated_completion_time = None;
            return;
        };
-        if self.is_completed() || !(0.0..100.0).contains(&self.progress_percentage) || self.bytes_processed == 0 {
+        if self.is_completed()
+            || self.progress_percentage <= 0.0
+            || self.progress_percentage >= 100.0
+            || self.bytes_processed == 0
+        {
            self.estimated_completion_time = None;
            return;
        }
@@ -142,18 +362,39 @@ impl HealProgress {
    }

    pub fn is_completed(&self) -> bool {
-        if self.progress_percentage >= 100.0 {
-            return true;
-        }
-        if self.objects_total_count > 0 || self.objects_total_size > 0 {
-            return false;
-        }
+        self.ledger_complete
+    }

-        self.objects_scanned > 0 && self.objects_healed.saturating_add(self.objects_failed) >= self.objects_scanned
+    /// Mark telemetry unknown while allowing the underlying heal operation to
+    /// continue.  This is used for corrupt/overflowing counters at the
+    /// observability boundary; it must never turn a successful heal into an
+    /// execution error.
+    pub fn mark_unknown(&mut self) {
+        self.counter_unknown = true;
+        self.progress_state = HealProgressState::Unknown;
+        self.ledger_complete = false;
+        self.progress_percentage = 0.0;
+        self.estimated_completion_time = None;
+        self.last_update_time = Some(SystemTime::now());
+    }
+
+    /// Mark the object ledger terminal only after the enclosing task has
+    /// committed all durable resume state and cleanup fences.
+    pub fn mark_completed(&mut self) {
+        let telemetry_unknown = self.counter_unknown || self.progress_state == HealProgressState::Unknown;
+        self.ledger_complete = true;
+        if !telemetry_unknown {
+            self.progress_state = HealProgressState::Completed;
+        }
+        self.progress_percentage = 100.0;
+        self.last_update_time = Some(SystemTime::now());
+        self.estimated_completion_time = None;
    }

    pub fn get_success_rate(&self) -> f64 {
-        let total = self.objects_healed + self.objects_failed;
+        let Some(total) = self.objects_healed.checked_add(self.objects_failed) else {
+            return 0.0;
+        };
        if total > 0 {
            (self.objects_healed as f64 / total as f64) * 100.0
        } else {
@@ -162,6 +403,101 @@ impl HealProgress {
    }
 }

+pub fn aggregate_heal_progress(progresses: impl IntoIterator<Item = HealProgress>) -> Option<HealProgress> {
+    let mut snapshot = HealProgress::default();
+    let mut found = false;
+    let mut has_object_sweep = false;
+    let mut all_object_baselines_known = true;
+    let mut baseline_generation = None;
+    let mut baseline_generation_consistent = true;
+    let mut all_ledgers_complete = true;
+    let mut counter_overflow = false;
+
+    for progress in progresses {
+        found = true;
+        let object_sweep = matches!(progress.kind, HealProgressKind::ObjectSweep);
+        has_object_sweep |= object_sweep;
+        all_ledgers_complete &= progress.ledger_complete;
+        if object_sweep {
+            all_object_baselines_known &= progress.baseline_known;
+            match baseline_generation {
+                None => baseline_generation = Some(progress.baseline_generation),
+                Some(generation) => baseline_generation_consistent &= generation == progress.baseline_generation,
+            }
+        }
+        counter_overflow |= progress.counter_unknown || matches!(progress.progress_state, HealProgressState::Unknown);
+        for (target, value) in [
+            (&mut snapshot.objects_scanned, progress.objects_scanned),
+            (&mut snapshot.objects_healed, progress.objects_healed),
+            (&mut snapshot.objects_failed, progress.objects_failed),
+            (&mut snapshot.skipped_objects, progress.skipped_objects),
+            (&mut snapshot.skipped_new_versions, progress.skipped_new_versions),
+            (&mut snapshot.skipped_ilm_expired, progress.skipped_ilm_expired),
+            (&mut snapshot.objects_total_count, progress.objects_total_count),
+            (&mut snapshot.objects_total_size, progress.objects_total_size),
+            (&mut snapshot.bytes_processed, progress.bytes_processed),
+            (&mut snapshot.stage_current, progress.stage_current),
+            (&mut snapshot.stage_total, progress.stage_total),
+        ] {
+            match target.checked_add(value) {
+                Some(sum) => *target = sum,
+                None => {
+                    *target = u64::MAX;
+                    counter_overflow = true;
+                }
+            }
+        }
+        snapshot.start_time = match (snapshot.start_time, progress.start_time) {
+            (Some(current), Some(next)) => Some(current.min(next)),
+            (None, next) => next,
+            (current, None) => current,
+        };
+        snapshot.last_update_time = match (snapshot.last_update_time, progress.last_update_time) {
+            (Some(current), Some(next)) => Some(current.max(next)),
+            (None, next) => next,
+            (current, None) => current,
+        };
+        if progress.current_object.is_some() {
+            snapshot.current_object = progress.current_object;
+        }
+    }
+
+    if !found {
+        return None;
+    }
+
+    snapshot.kind = if has_object_sweep {
+        HealProgressKind::ObjectSweep
+    } else {
+        HealProgressKind::Stage
+    };
+    snapshot.baseline_known = has_object_sweep && all_object_baselines_known && baseline_generation_consistent;
+    snapshot.baseline_generation = if snapshot.baseline_known && baseline_generation_consistent {
+        baseline_generation.flatten()
+    } else {
+        None
+    };
+    snapshot.ledger_complete = all_ledgers_complete;
+    snapshot.counter_unknown = counter_overflow;
+    if counter_overflow {
+        snapshot.progress_state = HealProgressState::Unknown;
+        snapshot.progress_percentage = if snapshot.ledger_complete { 100.0 } else { 0.0 };
+    } else if snapshot.ledger_complete {
+        snapshot.progress_state = HealProgressState::Completed;
+        snapshot.progress_percentage = 100.0;
+    } else if has_object_sweep {
+        snapshot.refresh_progress_percentage();
+    } else if snapshot.stage_total == 0 {
+        snapshot.progress_state = HealProgressState::Indeterminate;
+        snapshot.progress_percentage = 0.0;
+    } else {
+        snapshot.progress_state = HealProgressState::Running;
+        snapshot.progress_percentage = ((snapshot.stage_current as f64 / snapshot.stage_total as f64) * 100.0).min(99.999);
+    }
+    snapshot.refresh_estimated_completion_time();
+    Some(snapshot)
+}
+
 #[derive(Debug, Clone, Serialize, Deserialize)]
 pub struct HealStatistics {
    /// Total heal tasks
@@ -230,6 +566,7 @@ mod tests {
        assert_eq!(progress.objects_scanned, 0);
        assert_eq!(progress.objects_healed, 0);
        assert_eq!(progress.objects_failed, 0);
+        assert_eq!(progress.skipped_objects, 0);
        assert_eq!(progress.skipped_new_versions, 0);
        assert_eq!(progress.skipped_ilm_expired, 0);
        assert_eq!(progress.objects_total_count, 0);
@@ -250,10 +587,8 @@ mod tests {
        assert_eq!(progress.objects_healed, 8);
        assert_eq!(progress.objects_failed, 2);
        assert_eq!(progress.bytes_processed, 1024);
-        // Progress percentage should be calculated based on healed/total
-        // total = scanned + healed + failed = 10 + 8 + 2 = 20
-        // healed/total = 8/20 = 0.4 = 40%
-        assert!((progress.progress_percentage - 40.0).abs() < 0.001);
+        assert_eq!(progress.progress_state, HealProgressState::Indeterminate);
+        assert_eq!(progress.progress_percentage, 0.0);
        assert!(progress.last_update_time.is_some());
    }

@@ -262,7 +597,8 @@ mod tests {
        let mut progress = HealProgress::new();
        progress.start_time = Some(SystemTime::now() - Duration::from_secs(10));

-        progress.update_progress(100, 25, 0, 4096);
+        progress.set_total_baseline(100, 16384);
+        progress.update_progress(25, 25, 0, 4096);

        let eta = progress
            .estimated_completion_time
@@ -275,7 +611,7 @@ mod tests {
        let mut progress = HealProgress::new();
        progress.set_total_baseline(10, 8192);

-        progress.update_progress(100, 25, 0, 4096);
+        progress.update_progress(25, 25, 0, 4096);

        assert!((progress.progress_percentage - 50.0).abs() < 0.001);
    }
@@ -285,7 +621,7 @@ mod tests {
        let mut progress = HealProgress::new();
        progress.set_total_baseline(10, 0);

-        progress.update_progress(100, 3, 2, 0);
+        progress.update_progress(5, 3, 2, 0);

        assert!((progress.progress_percentage - 50.0).abs() < 0.001);
    }
@@ -295,7 +631,7 @@ mod tests {
        let mut progress = HealProgress::new();
        progress.set_total_baseline(10, 0);

-        progress.update_progress(100, 3, 2, 0);
+        progress.update_progress(5, 3, 2, 0);
        progress.record_skipped_new_version();

        assert_eq!(progress.skipped_new_versions, 1);
@@ -336,7 +672,8 @@ mod tests {
    fn test_heal_progress_update_progress_all_healed() {
        let mut progress = HealProgress::new();
        // When scanned=0, healed=10, failed=0: total=10, progress = 10/10 = 100%
-        progress.update_progress(0, 10, 0, 2048);
+        progress.update_progress(10, 10, 0, 2048);
+        progress.mark_completed();

        // All healed, should be 100%
        assert!((progress.progress_percentage - 100.0).abs() < 0.001);
@@ -394,6 +731,7 @@ mod tests {
        assert_eq!(json["objectsScanned"], 10);
        assert_eq!(json["objectsHealed"], 8);
        assert_eq!(json["objectsFailed"], 2);
+        assert_eq!(json["skippedObjects"], 0);
        assert_eq!(json["skippedNewVersions"], 0);
        assert_eq!(json["skippedIlmExpired"], 0);
        assert_eq!(json["bytesProcessed"], 1024);
@@ -405,6 +743,7 @@ mod tests {
    fn test_heal_progress_is_completed_by_percentage() {
        let mut progress = HealProgress::new();
        progress.update_progress(10, 10, 0, 1024);
+        progress.mark_completed();

        assert!(progress.is_completed());
    }
@@ -415,7 +754,7 @@ mod tests {
        progress.objects_scanned = 10;
        progress.objects_healed = 8;
        progress.objects_failed = 2;
-        // healed + failed = 8 + 2 = 10 >= scanned = 10
+        progress.mark_completed();
        assert!(progress.is_completed());
    }

@@ -455,6 +794,108 @@ mod tests {
        assert!((progress.get_success_rate() - 100.0).abs() < 0.001);
    }

+    #[test]
+    fn single_object_progress_reaches_terminal_100() {
+        let mut progress = HealProgress::new();
+        progress.update_object_progress(1, 1, 0, 0, 128);
+        assert!(!progress.is_completed());
+        progress.mark_completed();
+        assert!(progress.is_completed());
+        assert_eq!(progress.progress_percentage, 100.0);
+    }
+
+    #[test]
+    fn progress_without_baseline_is_indeterminate() {
+        let mut progress = HealProgress::new();
+        progress.update_object_progress(1, 1, 0, 0, 128);
+        assert_eq!(progress.progress_state, HealProgressState::Indeterminate);
+        assert_eq!(progress.progress_percentage, 0.0);
+        assert!(progress.estimated_completion_time.is_none());
+    }
+
+    #[test]
+    fn progress_retry_is_exactly_once() {
+        let mut progress = HealProgress::new();
+        progress.set_total_baseline(1, 128);
+        progress.update_object_progress(1, 1, 0, 0, 128);
+        progress.update_object_progress(1, 1, 0, 0, 128);
+        assert_eq!(progress.objects_scanned, 1);
+        assert_eq!(progress.objects_healed, 1);
+        assert_eq!(progress.bytes_processed, 128);
+    }
+
+    #[test]
+    fn progress_never_triggers_cleanup_before_terminal_ledger_empty() {
+        let mut progress = HealProgress::new();
+        progress.progress_percentage = 100.0;
+        assert!(!progress.is_completed());
+        progress.mark_completed();
+        assert!(progress.is_completed());
+    }
+
+    #[test]
+    fn progress_counter_overflow_is_marked_unknown_without_aborting_completed_heal() {
+        let mut progress = HealProgress::new();
+        progress.update_object_progress(u64::MAX, u64::MAX, 1, 0, 0);
+        assert_eq!(progress.progress_state, HealProgressState::Unknown);
+        progress.mark_completed();
+        assert!(progress.is_completed());
+        assert_eq!(progress.progress_state, HealProgressState::Unknown);
+
+        let aggregate = aggregate_heal_progress([progress]).expect("progress should aggregate");
+        assert!(aggregate.ledger_complete);
+        assert!(aggregate.counter_unknown);
+        assert_eq!(aggregate.progress_state, HealProgressState::Unknown);
+        assert_eq!(aggregate.progress_percentage, 100.0);
+    }
+
+    #[test]
+    fn aggregate_rejects_mixed_baseline_generations() {
+        let progress = |generation| HealProgress {
+            kind: HealProgressKind::ObjectSweep,
+            objects_scanned: 5,
+            objects_total_count: 10,
+            progress_state: HealProgressState::Running,
+            baseline_generation: Some(generation),
+            baseline_known: true,
+            ..Default::default()
+        };
+
+        let aggregate = aggregate_heal_progress([progress(1), progress(2)]).expect("progress should aggregate");
+        assert!(!aggregate.baseline_known);
+        assert_eq!(aggregate.baseline_generation, None);
+        assert_eq!(aggregate.progress_state, HealProgressState::Indeterminate);
+        assert_eq!(aggregate.progress_percentage, 0.0);
+    }
+
+    #[test]
+    fn aggregate_accepts_multiple_sets_from_one_snapshot_generation() {
+        let progress = |objects_scanned| HealProgress {
+            kind: HealProgressKind::ObjectSweep,
+            objects_scanned,
+            objects_total_count: 10,
+            progress_state: HealProgressState::Running,
+            baseline_generation: Some(7),
+            baseline_known: true,
+            ..Default::default()
+        };
+
+        let aggregate = aggregate_heal_progress([progress(5), progress(3)]).expect("progress should aggregate");
+        assert!(aggregate.baseline_known);
+        assert_eq!(aggregate.baseline_generation, Some(7));
+    }
+
+    #[test]
+    fn stage_updates_do_not_double_count_object_outcomes() {
+        let mut progress = HealProgress::new();
+        progress.update_object_progress(2, 1, 0, 1, 256);
+        progress.update_stage(3, 4);
+        assert_eq!(progress.kind, HealProgressKind::ObjectSweep);
+        assert_eq!(progress.objects_scanned, 2);
+        assert_eq!(progress.objects_healed, 1);
+        assert_eq!(progress.skipped_objects, 1);
+    }
+
    #[test]
    fn test_heal_statistics_new() {
        let stats = HealStatistics::new();
@@ -31,7 +31,7 @@ mod checkpoint;
 mod replacement;
 mod utils;

-pub use checkpoint::{CheckpointManager, ResumeCheckpoint};
+pub use checkpoint::{CheckpointManager, CheckpointObjectOutcome, CheckpointObjectOutcomeRecord, ResumeCheckpoint};
 pub(crate) use replacement::replacement_target_identities_match;
 use replacement::replacement_targets_match_identities;
 pub use replacement::{
@@ -340,6 +340,12 @@ pub struct ResumeState {
    pub failed_objects: u64,
    /// skipped objects
    pub skipped_objects: u64,
+    /// Terminal versions skipped because they were newer than the heal start.
+    #[serde(default)]
+    pub skipped_new_versions: u64,
+    /// Terminal versions handed to lifecycle expiry.
+    #[serde(default)]
+    pub skipped_ilm_expired: u64,
    /// current bucket
    pub current_bucket: Option<String>,
    /// current object
@@ -354,6 +360,24 @@ pub struct ResumeState {
    pub retry_count: u32,
    /// max retries
    pub max_retries: u32,
+    /// Bytes accounted by the object ledger; additive for old snapshots.
+    #[serde(default)]
+    pub processed_bytes: u64,
+    /// Total bytes from a complete usage snapshot, when available.
+    #[serde(default)]
+    pub total_bytes: u64,
+    /// Generation of the usage snapshot used for the baseline.
+    #[serde(default)]
+    pub baseline_generation: Option<u64>,
+    /// Whether the usage baseline is known. Missing in old snapshots means
+    /// indeterminate rather than a measured zero baseline.
+    #[serde(default)]
+    pub baseline_known: bool,
+    /// Persistent telemetry fence for counter/byte overflow or corruption.
+    /// It must survive a restart so a saturated snapshot is never presented as
+    /// a measured percentage on the next resume.
+    #[serde(default)]
+    pub counter_unknown: bool,
 }

 impl ResumeState {
@@ -377,6 +401,8 @@ impl ResumeState {
            successful_objects: 0,
            failed_objects: 0,
            skipped_objects: 0,
+            skipped_new_versions: 0,
+            skipped_ilm_expired: 0,
            current_bucket: None,
            current_object: None,
            completed_buckets: Vec::new(),
@@ -384,6 +410,11 @@ impl ResumeState {
            error_message: None,
            retry_count: 0,
            max_retries: 3,
+            processed_bytes: 0,
+            total_bytes: 0,
+            baseline_generation: None,
+            baseline_known: false,
+            counter_unknown: false,
        }
    }

@@ -412,6 +443,39 @@ impl ResumeState {
        self.last_update = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
    }

+    pub fn update_progress_with_bytes(
+        &mut self,
+        processed: u64,
+        successful: u64,
+        failed: u64,
+        skipped: u64,
+        processed_bytes: u64,
+    ) {
+        self.update_progress(processed, successful, failed, skipped);
+        self.processed_bytes = processed_bytes;
+    }
+
+    pub fn set_skipped_version_counts(&mut self, new_versions: u64, ilm_expired: u64) {
+        self.skipped_new_versions = new_versions;
+        self.skipped_ilm_expired = ilm_expired;
+        self.last_update = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
+    }
+
+    pub fn set_progress_baseline(&mut self, total_objects: u64, total_bytes: u64, generation: Option<u64>) {
+        self.total_objects = total_objects;
+        self.total_bytes = total_bytes;
+        self.baseline_generation = generation;
+        // This method is called only after a complete usage snapshot has been
+        // validated.  A complete but empty snapshot is still a known baseline.
+        self.baseline_known = true;
+        self.last_update = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
+    }
+
+    pub fn mark_counter_unknown(&mut self) {
+        self.counter_unknown = true;
+        self.last_update = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
+    }
+
    pub fn set_current_item(&mut self, bucket: Option<String>, object: Option<String>) {
        self.current_bucket = bucket;
        self.current_object = object;
@@ -437,6 +501,7 @@ impl ResumeState {
        if let Some(pos) = self.pending_buckets.iter().position(|b| b == bucket) {
            self.pending_buckets.remove(pos);
        }
+        self.resume_cursor = None;
        self.last_update = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
    }

@@ -454,6 +519,10 @@ impl ResumeState {
        self.successful_objects = 0;
        self.failed_objects = 0;
        self.skipped_objects = 0;
+        self.skipped_new_versions = 0;
+        self.skipped_ilm_expired = 0;
+        self.processed_bytes = 0;
+        self.counter_unknown = false;
        self.completed = false;
        // A retry re-scans every bucket from the beginning, so the version
        // cursor must be cleared too — otherwise the retry would resume mid-scan.
@@ -476,14 +545,28 @@ impl ResumeState {
    }

    pub fn get_progress_percentage(&self) -> f64 {
+        if self.completed {
+            return 100.0;
+        }
+        if self.counter_unknown {
+            return 0.0;
+        }
+        if !self.baseline_known {
+            return 0.0;
+        }
+        if self.total_bytes > 0 {
+            return ((self.processed_bytes as f64 / self.total_bytes as f64) * 100.0).min(99.999);
+        }
        if self.total_objects == 0 {
            return 0.0;
        }
-        (self.processed_objects as f64 / self.total_objects as f64) * 100.0
+        ((self.processed_objects as f64 / self.total_objects as f64) * 100.0).min(99.999)
    }

    pub fn get_success_rate(&self) -> f64 {
-        let total = self.successful_objects + self.failed_objects;
+        let Some(total) = self.successful_objects.checked_add(self.failed_objects) else {
+            return 0.0;
+        };
        if total == 0 {
            return 0.0;
        }
@@ -754,6 +837,14 @@ impl ResumeManager {
            state.successful_objects = 0;
            state.failed_objects = 0;
            state.skipped_objects = 0;
+            state.skipped_new_versions = 0;
+            state.skipped_ilm_expired = 0;
+            state.processed_bytes = 0;
+            state.total_objects = 0;
+            state.total_bytes = 0;
+            state.baseline_generation = None;
+            state.baseline_known = false;
+            state.counter_unknown = false;
            state.completed = false;
            state.completed_buckets.clear();
            state.schema_version = CURRENT_RESUME_SCHEMA;
@@ -838,6 +929,41 @@ impl ResumeManager {
        self.save_state_throttled().await
    }

+    pub async fn update_progress_with_bytes(
+        &self,
+        processed: u64,
+        successful: u64,
+        failed: u64,
+        skipped: u64,
+        processed_bytes: u64,
+    ) -> Result<()> {
+        let mut state = self.state.write().await;
+        state.update_progress_with_bytes(processed, successful, failed, skipped, processed_bytes);
+        drop(state);
+        self.save_state_throttled().await
+    }
+
+    pub async fn set_progress_baseline(&self, total_objects: u64, total_bytes: u64, generation: Option<u64>) -> Result<()> {
+        let mut state = self.state.write().await;
+        state.set_progress_baseline(total_objects, total_bytes, generation);
+        drop(state);
+        self.save_state_throttled().await
+    }
+
+    pub async fn mark_counter_unknown(&self) -> Result<()> {
+        let mut state = self.state.write().await;
+        state.mark_counter_unknown();
+        drop(state);
+        self.save_state().await
+    }
+
+    pub async fn set_skipped_version_counts(&self, new_versions: u64, ilm_expired: u64) -> Result<()> {
+        let mut state = self.state.write().await;
+        state.set_skipped_version_counts(new_versions, ilm_expired);
+        drop(state);
+        self.save_state_throttled().await
+    }
+
    /// Set current item. Called once per healed object, so persistence is
    /// throttled: the in-memory state always updates, but the snapshot is only
    /// written every `PERSIST_EVERY_MUTATIONS` calls or `PERSIST_INTERVAL`.
@@ -882,7 +1008,7 @@ impl ResumeManager {
        let mut state = self.state.write().await;
        state.complete_bucket(bucket);
        drop(state);
-        self.save_state_throttled().await
+        self.save_state().await
    }

    /// mark task completed
@@ -29,10 +29,30 @@ use super::{

 const EVENT_HEAL_CHECKPOINT_STATE: &str = "heal_checkpoint_state";

-/// Current on-disk schema version for `ResumeCheckpoint`. Same rationale as
-/// `CURRENT_RESUME_SCHEMA`: pre-per-version dedup identities are not comparable
-/// to the new `compose_key` identities, so a stale checkpoint is discarded.
-pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 5;
+/// Current on-disk schema version for `ResumeCheckpoint`. Schema 5 could
+/// persist dedup identities without the aggregate counters needed to restore
+/// them safely, so stale checkpoints are discarded and replayed.
+pub(super) const CURRENT_CHECKPOINT_SCHEMA: u32 = 6;
+
+#[derive(Debug, Clone, Copy)]
+pub enum CheckpointObjectOutcome {
+    Processed,
+    Failed,
+    Skipped,
+}
+
+#[derive(Debug)]
+pub struct CheckpointObjectOutcomeRecord {
+    pub object: String,
+    pub outcome: CheckpointObjectOutcome,
+    pub successful: u64,
+    pub failed: u64,
+    pub skipped: u64,
+    pub bytes: u64,
+    pub skipped_new_versions: u64,
+    pub skipped_ilm_expired: u64,
+    pub counter_unknown: bool,
+}

 /// resume checkpoint
 #[derive(Debug, Clone, Serialize, Deserialize)]
@@ -57,6 +77,30 @@ pub struct ResumeCheckpoint {
    pub failed_objects: HashSet<String>,
    /// skipped objects
    pub skipped_objects: HashSet<String>,
+    /// Aggregate object ledger counters restored alongside the dedup sets.
+    #[serde(default)]
+    pub successful_objects: u64,
+    #[serde(default)]
+    pub failed_object_count: u64,
+    #[serde(default)]
+    pub skipped_object_count: u64,
+    #[serde(default)]
+    pub skipped_new_versions: u64,
+    #[serde(default)]
+    pub skipped_ilm_expired: u64,
+    #[serde(default)]
+    pub processed_bytes: u64,
+    #[serde(default)]
+    pub total_objects: u64,
+    #[serde(default)]
+    pub total_bytes: u64,
+    #[serde(default)]
+    pub baseline_generation: Option<u64>,
+    #[serde(default)]
+    pub baseline_known: bool,
+    /// Persistent telemetry fence for counter/byte overflow or corruption.
+    #[serde(default)]
+    pub counter_unknown: bool,
 }

 impl ResumeCheckpoint {
@@ -70,6 +114,17 @@ impl ResumeCheckpoint {
            processed_objects: HashSet::new(),
            failed_objects: HashSet::new(),
            skipped_objects: HashSet::new(),
+            successful_objects: 0,
+            failed_object_count: 0,
+            skipped_object_count: 0,
+            skipped_new_versions: 0,
+            skipped_ilm_expired: 0,
+            processed_bytes: 0,
+            total_objects: 0,
+            total_bytes: 0,
+            baseline_generation: None,
+            baseline_known: false,
+            counter_unknown: false,
        }
    }

@@ -91,6 +146,34 @@ impl ResumeCheckpoint {
        self.skipped_objects.insert(object);
    }

+    pub fn update_progress(&mut self, successful: u64, failed: u64, skipped: u64, bytes: u64) {
+        self.successful_objects = successful;
+        self.failed_object_count = failed;
+        self.skipped_object_count = skipped;
+        self.processed_bytes = bytes;
+    }
+
+    pub fn set_progress_baseline(&mut self, total_objects: u64, total_bytes: u64, generation: Option<u64>) {
+        self.total_objects = total_objects;
+        self.total_bytes = total_bytes;
+        self.baseline_generation = generation;
+        // The caller has already validated that this is a complete snapshot;
+        // preserve the distinction between a known empty scope and an old
+        // checkpoint that omitted all baseline fields.
+        self.baseline_known = true;
+    }
+
+    pub fn mark_counter_unknown(&mut self) {
+        self.counter_unknown = true;
+        self.checkpoint_time = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
+    }
+
+    pub fn set_skipped_version_counts(&mut self, new_versions: u64, ilm_expired: u64) {
+        self.skipped_new_versions = new_versions;
+        self.skipped_ilm_expired = ilm_expired;
+        self.checkpoint_time = SystemTime::now().duration_since(UNIX_EPOCH).unwrap_or_default().as_secs();
+    }
+
    /// Advance past a fully-processed page: objects below `object_index` are
    /// skipped by position on resume, so the per-object sets no longer need
    /// their entries and would otherwise grow with the whole bucket.
@@ -107,6 +190,17 @@ impl ResumeCheckpoint {
        self.update_position(0, 0);
        self.processed_objects.clear();
        self.skipped_objects.clear();
+        self.successful_objects = 0;
+        self.failed_object_count = 0;
+        self.skipped_object_count = 0;
+        self.skipped_new_versions = 0;
+        self.skipped_ilm_expired = 0;
+        self.processed_bytes = 0;
+        self.total_objects = 0;
+        self.total_bytes = 0;
+        self.baseline_generation = None;
+        self.baseline_known = false;
+        self.counter_unknown = false;
        self.failed_objects.clear();
    }
 }
@@ -158,10 +252,9 @@ impl CheckpointManager {
            });
        }

-        // A checkpoint from an older schema stored latest-only dedup identities
-        // that are not comparable to the new per-version `compose_key`
-        // identities. Discard the stale sets and position, then stamp the
-        // current schema so the scan restarts cleanly.
+        // Older checkpoints can contain identities that are not comparable to
+        // the current keys or lack their corresponding aggregate counters.
+        // Discard the stale sets and position so the scan restarts cleanly.
        if checkpoint.schema_version > CURRENT_CHECKPOINT_SCHEMA {
            return Err(Error::TaskExecutionFailed {
                message: format!(
@@ -185,6 +278,17 @@ impl CheckpointManager {
            checkpoint.processed_objects.clear();
            checkpoint.failed_objects.clear();
            checkpoint.skipped_objects.clear();
+            checkpoint.successful_objects = 0;
+            checkpoint.failed_object_count = 0;
+            checkpoint.skipped_object_count = 0;
+            checkpoint.skipped_new_versions = 0;
+            checkpoint.skipped_ilm_expired = 0;
+            checkpoint.processed_bytes = 0;
+            checkpoint.total_objects = 0;
+            checkpoint.total_bytes = 0;
+            checkpoint.baseline_generation = None;
+            checkpoint.baseline_known = false;
+            checkpoint.counter_unknown = false;
            checkpoint.current_bucket_index = 0;
            checkpoint.current_object_index = 0;
            checkpoint.schema_version = CURRENT_CHECKPOINT_SCHEMA;
@@ -225,7 +329,7 @@ impl CheckpointManager {
        self.save_checkpoint_throttled().await
    }

-    /// Advance past a completed page and prune the per-object sets, then persist.
+    /// Persist a completed page position while retaining its identities.
    pub async fn complete_page(&self, bucket_index: usize, object_index: usize) -> Result<()> {
        let mut checkpoint = self.checkpoint.write().await;
        checkpoint.complete_page(bucket_index, object_index);
@@ -233,6 +337,35 @@ impl CheckpointManager {
        self.save_checkpoint_throttled().await
    }

+    /// Persist the page position while retaining identities until the resume
+    /// cursor is durable.
+    pub async fn advance_page(&self, bucket_index: usize, object_index: usize) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.update_position(bucket_index, object_index);
+        drop(checkpoint);
+        self.save_checkpoint().await
+    }
+
+    /// Remove the previous page's dedup identities only after its resume cursor
+    /// has been durably exposed.
+    pub async fn prune_completed_page(&self) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.processed_objects.clear();
+        checkpoint.skipped_objects.clear();
+        checkpoint.failed_objects.clear();
+        drop(checkpoint);
+        self.save_checkpoint().await
+    }
+
+    /// Advance to the next bucket and clear the final page identities after the
+    /// resume state has durably recorded the completed bucket.
+    pub async fn complete_bucket(&self, next_bucket_index: usize) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.complete_page(next_bucket_index, 0);
+        drop(checkpoint);
+        self.save_checkpoint().await
+    }
+
    /// Reset the checkpoint to the start of the scan for a retry, then persist.
    pub async fn reset_for_retry(&self) -> Result<()> {
        let mut checkpoint = self.checkpoint.write().await;
@@ -267,6 +400,62 @@ impl CheckpointManager {
        self.save_checkpoint_if_due().await
    }

+    /// Atomically persist an object's dedup identity with its aggregate result.
+    pub async fn record_object_outcome(&self, record: CheckpointObjectOutcomeRecord) -> Result<()> {
+        let CheckpointObjectOutcomeRecord {
+            object,
+            outcome,
+            successful,
+            failed,
+            skipped,
+            bytes,
+            skipped_new_versions,
+            skipped_ilm_expired,
+            counter_unknown,
+        } = record;
+        let mut checkpoint = self.checkpoint.write().await;
+        match outcome {
+            CheckpointObjectOutcome::Processed => checkpoint.add_processed_object(object),
+            CheckpointObjectOutcome::Failed => checkpoint.add_failed_object(object),
+            CheckpointObjectOutcome::Skipped => checkpoint.add_skipped_object(object),
+        }
+        checkpoint.update_progress(successful, failed, skipped, bytes);
+        checkpoint.set_skipped_version_counts(skipped_new_versions, skipped_ilm_expired);
+        if counter_unknown {
+            checkpoint.mark_counter_unknown();
+        }
+        drop(checkpoint);
+        self.save_checkpoint_if_due().await
+    }
+
+    pub async fn update_progress(&self, successful: u64, failed: u64, skipped: u64, bytes: u64) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.update_progress(successful, failed, skipped, bytes);
+        drop(checkpoint);
+        self.save_checkpoint_if_due().await
+    }
+
+    pub async fn set_progress_baseline(&self, total_objects: u64, total_bytes: u64, generation: Option<u64>) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.set_progress_baseline(total_objects, total_bytes, generation);
+        drop(checkpoint);
+        self.save_checkpoint_throttled().await
+    }
+
+    pub async fn mark_counter_unknown(&self) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.mark_counter_unknown();
+        drop(checkpoint);
+        self.save_checkpoint().await
+    }
+
+    pub async fn set_skipped_version_counts(&self, new_versions: u64, ilm_expired: u64) -> Result<()> {
+        let mut checkpoint = self.checkpoint.write().await;
+        checkpoint.set_skipped_version_counts(new_versions, ilm_expired);
+        drop(checkpoint);
+        self.save_checkpoint_throttled().await
+    }
+
    async fn save_checkpoint_if_due(&self) -> Result<()> {
        let should_save = self.throttle.lock().map(|mut throttle| throttle.record()).unwrap_or(true);
        if !should_save {
@@ -1296,6 +1296,7 @@ async fn test_resume_state_progress() {
    assert_eq!(progress, 0.0); // total_objects is 0

    state.total_objects = 100;
+    state.baseline_known = true;
    let progress = state.get_progress_percentage();
    assert_eq!(progress, 10.0);
 }
@@ -1475,6 +1476,40 @@ fn test_checkpoint_object_sets_dedupe_and_prune() {
    assert!(checkpoint.failed_objects.is_empty());
 }

+#[tokio::test]
+async fn checkpoint_page_commit_keeps_ledger_until_cursor_is_durable() {
+    let (_temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let checkpoint = CheckpointManager::new(disk.clone(), task_id.clone()).await.unwrap();
+
+    checkpoint
+        .record_object_outcome(CheckpointObjectOutcomeRecord {
+            object: "bucket/object:v1".to_string(),
+            outcome: CheckpointObjectOutcome::Processed,
+            successful: 1,
+            failed: 0,
+            skipped: 0,
+            bytes: 128,
+            skipped_new_versions: 0,
+            skipped_ilm_expired: 0,
+            counter_unknown: false,
+        })
+        .await
+        .unwrap();
+    checkpoint.advance_page(0, 1).await.unwrap();
+
+    let reloaded = CheckpointManager::load_from_disk(disk.clone(), &task_id).await.unwrap();
+    let snapshot = reloaded.get_checkpoint().await;
+    assert_eq!(snapshot.current_object_index, 1);
+    assert_eq!(snapshot.successful_objects, 1);
+    assert_eq!(snapshot.processed_bytes, 128);
+    assert!(snapshot.processed_objects.contains("bucket/object:v1"));
+
+    checkpoint.prune_completed_page().await.unwrap();
+    let reloaded = CheckpointManager::load_from_disk(disk, &task_id).await.unwrap();
+    assert!(reloaded.get_checkpoint().await.processed_objects.is_empty());
+}
+
 #[test]
 fn test_checkpoint_loads_legacy_vec_format() {
    // Checkpoints written before the HashSet migration stored the object
@@ -1568,14 +1603,14 @@ async fn test_resumestate_schema_v0_discarded_on_load() {
 }

 #[tokio::test]
-async fn test_checkpoint_schema_v4_discarded_on_load() {
+async fn test_checkpoint_schema_v5_discarded_on_load() {
    let (temp_dir, disk) = schema_test_disk().await;

-    // The previous checkpoint schema is unsafe once its paired resume
-    // state is discarded: retaining either position would skip work.
+    // Schema v5 can persist failed identities without the aggregate counters
+    // that make those identities safe to deduplicate after an upgrade.
    let task_id = "00000000-0000-4000-8000-000000000002";
    let legacy = r#"{
-            "schema_version": 4,
+            "schema_version": 5,
            "task_id": "00000000-0000-4000-8000-000000000002",
            "checkpoint_time": 1700000000,
            "current_bucket_index": 2,
@@ -1639,6 +1674,120 @@ async fn current_normal_resume_schema_preserves_progress() {
    temp_dir.close().expect("remove schema test directory");
 }

+#[test]
+fn progress_checkpoint_restores_bytes_and_generation() {
+    let mut checkpoint = ResumeCheckpoint::new("progress-checkpoint".to_string());
+    checkpoint.set_progress_baseline(9, 4096, Some(77));
+    checkpoint.update_progress(4, 1, 2, 2048);
+    checkpoint.set_skipped_version_counts(3, 1);
+    checkpoint.mark_counter_unknown();
+
+    let restored: ResumeCheckpoint =
+        serde_json::from_slice(&serde_json::to_vec(&checkpoint).expect("serialize checkpoint")).expect("deserialize checkpoint");
+    assert_eq!(restored.processed_bytes, 2048);
+    assert_eq!(restored.total_objects, 9);
+    assert_eq!(restored.total_bytes, 4096);
+    assert_eq!(restored.baseline_generation, Some(77));
+    assert!(restored.baseline_known);
+    assert_eq!(restored.skipped_new_versions, 3);
+    assert_eq!(restored.skipped_ilm_expired, 1);
+    assert!(restored.counter_unknown);
+}
+
+#[test]
+fn old_progress_schema_migrates_missing_fields_to_unknown() {
+    let state = ResumeState::new(
+        "legacy-progress".to_string(),
+        "erasure_set".to_string(),
+        "pool_0_set_0".to_string(),
+        Vec::new(),
+    );
+    let mut value = serde_json::to_value(state).expect("serialize legacy-compatible state");
+    let object = value.as_object_mut().expect("state must be an object");
+    for field in [
+        "processed_bytes",
+        "total_bytes",
+        "baseline_generation",
+        "baseline_known",
+        "skipped_new_versions",
+        "skipped_ilm_expired",
+    ] {
+        object.remove(field);
+    }
+    object.insert("total_objects".to_string(), serde_json::json!(10));
+    object.insert("processed_objects".to_string(), serde_json::json!(5));
+    let restored: ResumeState = serde_json::from_value(value).expect("deserialize old progress state");
+    assert_eq!(restored.processed_bytes, 0);
+    assert_eq!(restored.total_bytes, 0);
+    assert_eq!(restored.baseline_generation, None);
+    assert!(!restored.baseline_known, "missing baseline must remain unknown");
+    assert_eq!(restored.get_progress_percentage(), 0.0);
+    assert_eq!(restored.skipped_new_versions, 0);
+    assert_eq!(restored.skipped_ilm_expired, 0);
+}
+
+#[test]
+fn progress_counter_unknown_survives_resume_round_trip() {
+    let mut state = ResumeState::new(
+        "overflow-progress".to_string(),
+        "erasure_set".to_string(),
+        "pool_0_set_0".to_string(),
+        Vec::new(),
+    );
+    state.mark_counter_unknown();
+
+    let restored: ResumeState =
+        serde_json::from_slice(&serde_json::to_vec(&state).expect("serialize resume state")).expect("deserialize resume state");
+    assert!(restored.counter_unknown);
+}
+
+#[tokio::test]
+async fn checkpoint_progress_survives_a_torn_resume_summary_write() {
+    let (_temp_dir, disk) = schema_test_disk().await;
+    let task_id = ResumeUtils::generate_task_id();
+    let _resume = ResumeManager::new(
+        disk.clone(),
+        task_id.clone(),
+        "erasure_set".to_string(),
+        "pool_0_set_0".to_string(),
+        vec!["bucket".to_string()],
+    )
+    .await
+    .expect("resume state should persist");
+    let checkpoint = CheckpointManager::new(disk.clone(), task_id.clone())
+        .await
+        .expect("checkpoint should persist");
+
+    // This is the ordering used by the erasure-set loop: the checkpoint is
+    // durable before the summary write.  Stop here to model a crash in the
+    // inter-store window and verify that the recovery authority retains the
+    // telemetry fence and bytes.
+    checkpoint
+        .update_progress(3, 0, 0, 1024)
+        .await
+        .expect("checkpoint progress should persist");
+    checkpoint.mark_counter_unknown().await.expect("unknown fence should persist");
+    checkpoint
+        .update_position(0, 3)
+        .await
+        .expect("checkpoint position should persist");
+
+    let restored_checkpoint = CheckpointManager::load_from_disk(disk.clone(), &task_id)
+        .await
+        .expect("checkpoint should reload")
+        .get_checkpoint()
+        .await;
+    let restored_resume = ResumeManager::load_from_disk(disk, &task_id)
+        .await
+        .expect("resume summary should reload")
+        .get_state()
+        .await;
+    assert!(restored_checkpoint.counter_unknown);
+    assert_eq!(restored_checkpoint.processed_bytes, 1024);
+    assert_eq!(restored_checkpoint.current_object_index, 3);
+    assert!(!restored_resume.counter_unknown, "summary is intentionally the torn/older store");
+}
+
 #[tokio::test]
 async fn future_resume_and_checkpoint_schemas_are_rejected() {
    let (temp_dir, disk) = schema_test_disk().await;
@@ -22,6 +22,7 @@ use serde::{Deserialize, Serialize};
 use std::sync::Arc;
 use tracing::{debug, error, warn};

+use super::progress::stable_generation;
 use super::storage_api::owner::{EcstoreHealLifecycleExpiryContext, ecstore_load_admin_data_usage_from_backend_cached};
 use super::storage_api::storage::{
    BucketInfo, BucketOperations, DiskSetSelector, HealOperations as _, ListOperations as _, ObjectIO as _,
@@ -34,6 +35,9 @@ pub use super::{HealObjectInfo, HealObjectOptions, HealPutObjReader};
 pub struct HealBucketUsageBaseline {
    pub objects_count: u64,
    pub bytes: u64,
+    /// Stable identity of the validated usage snapshot and selected scope.
+    /// `None` is retained for test/legacy providers that cannot expose one.
+    pub generation: Option<u64>,
 }

 pub struct HealLifecycleExpiryContext {
@@ -785,11 +789,52 @@ impl HealStorageAPI for ECStoreHealStorage {
        let mut baseline = HealBucketUsageBaseline::default();
        for bucket in buckets {
            if let Some(usage) = info.buckets_usage.get(bucket) {
-                baseline.objects_count = baseline.objects_count.saturating_add(usage.objects_count);
-                baseline.bytes = baseline.bytes.saturating_add(usage.size);
+                baseline.objects_count = match baseline.objects_count.checked_add(usage.objects_count) {
+                    Some(total) => total,
+                    // A corrupt/overflowing usage snapshot is not a usable
+                    // denominator.  Leave progress indeterminate instead of
+                    // turning saturation into a plausible percentage.
+                    None => return Ok(None),
+                };
+                baseline.bytes = match baseline.bytes.checked_add(usage.size) {
+                    Some(total) => total,
+                    None => return Ok(None),
+                };
            }
        }

+        let identity = info.snapshot_identity();
+        let mut canonical = Vec::new();
+        match identity.last_update {
+            Some(last_update) => {
+                canonical.push(1);
+                canonical.extend_from_slice(
+                    &last_update
+                        .duration_since(std::time::UNIX_EPOCH)
+                        .unwrap_or_default()
+                        .as_nanos()
+                        .to_be_bytes(),
+                );
+            }
+            None => canonical.push(0),
+        }
+        for value in [identity.scanner_cycle, identity.scanner_epoch] {
+            match value {
+                Some(value) => {
+                    canonical.push(1);
+                    canonical.extend_from_slice(&value.to_be_bytes());
+                }
+                None => canonical.push(0),
+            }
+        }
+        let mut scope = buckets.to_vec();
+        scope.sort_unstable();
+        for bucket in scope {
+            canonical.extend_from_slice(&(bucket.len() as u64).to_be_bytes());
+            canonical.extend_from_slice(bucket.as_bytes());
+        }
+        baseline.generation = Some(stable_generation(&[&canonical]));
+
        Ok(Some(baseline))
    }

@@ -649,7 +649,7 @@ impl HealTask {

        let mut progress = self.progress.write().await;
        progress.set_current_object(Some(format!("skipped: {bucket}/{object}")));
-        progress.update_progress(0, 1, 0, 0);
+        progress.update_stage(1, 1);
        Ok(())
    }

@@ -733,7 +733,7 @@ impl HealTask {
            "Heal object skipped for data usage cache after transient error"
        );
        let mut progress = self.progress.write().await;
-        progress.update_progress(3, 3, 0, 0);
+        progress.update_stage(3, 3);
        true
    }

@@ -757,7 +757,7 @@ impl HealTask {
        );
        let mut progress = self.progress.write().await;
        progress.set_current_object(Some(format!("skipped: {bucket}/{object}")));
-        progress.update_progress(4, 4, 0, 0);
+        progress.update_stage(4, 4);
        true
    }

@@ -831,6 +831,10 @@ impl HealTask {

        match &result {
            Ok(_) => {
+                // A stage can reach its final step before the durable resume
+                // ledger and cleanup fences commit. Publish terminal 100 only
+                // after the enclosing operation has returned success.
+                self.progress.write().await.mark_completed();
                let mut status = self.status.write().await;
                *status = HealTaskStatus::Completed;
                demote_to_debug_when!(self.heal_type.is_per_object(), info, target: "rustfs::heal::task", {
@@ -13,6 +13,7 @@
 // limitations under the License.
 /// bucket/cluster/prefix heal: the recursive bucket-objects sweep and the erasure-set usage baseline
 use super::*;
+use crate::heal::progress::{add_bytes, increment_counter, stable_generation};

 impl HealTask {
    pub(super) async fn heal_bucket(&self, bucket: &str) -> Result<()> {
@@ -32,7 +33,7 @@ impl HealTask {
        {
            let mut progress = self.progress.write().await;
            progress.set_current_object(Some(format!("bucket: {bucket}")));
-            progress.update_progress(0, 3, 0, 0);
+            progress.update_stage(0, 3);
        }

        // Step 1: Check if bucket exists
@@ -66,7 +67,7 @@ impl HealTask {

        {
            let mut progress = self.progress.write().await;
-            progress.update_progress(1, 3, 0, 0);
+            progress.update_stage(1, 3);
        }

        // Step 2: Perform bucket heal using ecstore
@@ -122,7 +123,7 @@ impl HealTask {

                if !self.options.recursive {
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                }
                Ok(())
            }
@@ -142,7 +143,7 @@ impl HealTask {
                );
                {
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                }
                Err(Error::TaskExecutionFailed {
                    message: format!("Failed to heal bucket {bucket}: {e}"),
@@ -245,6 +246,7 @@ impl HealTask {
        let mut scanned = 0u64;
        let mut healed = 0u64;
        let mut failed = 0u64;
+        let mut skipped = 0u64;
        let mut retryable_failed = 0u64;
        let mut permanent_failed = 0u64;
        let mut bytes = 0u64;
@@ -286,16 +288,14 @@ impl HealTask {
                let mut retry = Vec::with_capacity(pending.len());
                for item in pending {
                    self.check_control_flags().await?;
+                    let mut telemetry_unknown = false;
                    let object = item.name.as_str();
-                    if retry_attempt == 0 {
-                        scanned = scanned.saturating_add(1);
-                    }
                    {
                        let mut progress = self.progress.write().await;
                        progress.set_current_object(Some(format!("{bucket}/{object}")));
-                        progress.update_progress(scanned, healed, failed, bytes);
                    }

+                    let mut terminal_outcome = true;
                    let error = match self
                        .await_with_control(
                            self.storage
@@ -304,13 +304,13 @@ impl HealTask {
                        .await
                    {
                        Ok((result, None)) => {
-                            healed = healed.saturating_add(1);
-                            bytes = bytes.saturating_add(u64::try_from(result.object_size).unwrap_or_default());
+                            telemetry_unknown |= !increment_counter(&mut healed);
+                            telemetry_unknown |= !add_bytes(&mut bytes, u64::try_from(result.object_size).unwrap_or(u64::MAX));
                            self.record_result_item(result).await;
                            None
                        }
                        Ok((_, Some(err))) if is_missing_object_dir_heal_result(object, &err) => {
-                            healed = healed.saturating_add(1);
+                            telemetry_unknown |= !increment_counter(&mut healed);
                            debug!(
                                target: "rustfs::heal::task",
                                event = EVENT_HEAL_BUCKET_RESULT,
@@ -329,6 +329,7 @@ impl HealTask {

                    if let Some(err) = error {
                        if Self::should_skip_data_usage_cache_heal_error(bucket, object, &err) {
+                            telemetry_unknown |= !increment_counter(&mut skipped);
                            warn!(
                                target: "rustfs::heal::task",
                                event = EVENT_HEAL_BUCKET_RESULT,
@@ -342,6 +343,7 @@ impl HealTask {
                                "Heal bucket object repair skipped due to transient metadata error"
                            );
                        } else if err.is_recoverable_heal() && retry_attempt < MAX_BUCKET_OBJECT_HEAL_RETRIES {
+                            terminal_outcome = false;
                            debug!(
                                target: "rustfs::heal::task",
                                event = EVENT_HEAL_BUCKET_RESULT,
@@ -357,7 +359,7 @@ impl HealTask {
                            );
                            retry.push(item);
                        } else {
-                            failed = failed.saturating_add(1);
+                            telemetry_unknown |= !increment_counter(&mut failed);
                            if err.is_recoverable_heal() {
                                retryable_failed = retryable_failed.saturating_add(1);
                            } else {
@@ -383,8 +385,19 @@ impl HealTask {
                        }
                    }

+                    if terminal_outcome {
+                        telemetry_unknown |= !increment_counter(&mut scanned);
+                    }
+
+                    if !terminal_outcome {
+                        continue;
+                    }
+
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(scanned, healed, failed, bytes);
+                    progress.update_object_progress(scanned, healed, failed, skipped, bytes);
+                    if telemetry_unknown {
+                        progress.mark_unknown();
+                    }
                }
                pending = retry;
                retry_attempt = retry_attempt.saturating_add(1);
@@ -431,7 +444,10 @@ impl HealTask {
        Ok(())
    }

-    pub(super) async fn apply_erasure_set_usage_baseline(&self, buckets: &[String]) -> Result<()> {
+    pub(super) async fn apply_erasure_set_usage_baseline(&self, buckets: &[String], _set_disk_id: &str) -> Result<()> {
+        if matches!(self.options.scan_mode, HealScanMode::Deep) || matches!(self.source, HealRequestSource::AutoHeal) {
+            return Ok(());
+        }
        let baseline = match self
            .await_with_control(self.storage.erasure_set_usage_baseline(buckets))
            .await
@@ -442,9 +458,18 @@ impl HealTask {
            Err(_) => return Ok(()),
        };

-        let HealBucketUsageBaseline { objects_count, bytes } = baseline;
+        let HealBucketUsageBaseline {
+            objects_count,
+            bytes,
+            generation,
+        } = baseline;
+        let generation = generation.map(|snapshot_generation| stable_generation(&[&snapshot_generation.to_be_bytes()]));
        let mut progress = self.progress.write().await;
-        progress.set_total_baseline(objects_count, bytes);
+        if let Some(generation) = generation {
+            progress.set_total_baseline_with_generation(objects_count, bytes, generation);
+        } else {
+            progress.set_total_baseline(objects_count, bytes);
+        }
        Ok(())
    }
 }
@@ -32,7 +32,7 @@ impl HealTask {
        {
            let mut progress = self.progress.write().await;
            progress.set_current_object(Some(format!("erasure_set: {} ({} buckets)", set_disk_id, buckets.len())));
-            progress.update_progress(0, 4, 0, 0);
+            progress.update_stage(0, 4);
        }

        let is_auto_replacement = matches!(self.source, HealRequestSource::AutoHeal) && !self.heal_endpoints.is_empty();
@@ -158,7 +158,7 @@ impl HealTask {
            None
        };

-        self.apply_erasure_set_usage_baseline(&buckets).await?;
+        self.apply_erasure_set_usage_baseline(&buckets, &set_disk_id).await?;

        let healing_marker = format!("{set_disk_id}:{}", self.id);
        if let Some((disk, resume_manager, _)) = replacement_resume.as_ref() {
@@ -231,6 +231,10 @@ impl HealTask {
                            "Heal erasure set format repair skipped because no format heal was required"
                        );
                    } else {
+                        let error = e;
+                        if error.is_recoverable_heal() {
+                            return Err(error);
+                        }
                        error!(
                            target: "rustfs::heal::task",
                            event = EVENT_HEAL_ERASURE_SET_RESULT,
@@ -239,15 +243,15 @@ impl HealTask {
                            task_id = %self.id,
                            set_disk_id,
                            result = "format_failed",
-                            error = %e,
+                            error = %error,
                            "Heal erasure set failed"
                        );
                        {
                            let mut progress = self.progress.write().await;
-                            progress.update_progress(4, 4, 0, 0);
+                            progress.update_stage(4, 4);
                        }
                        return Err(Error::TaskExecutionFailed {
-                            message: format!("Failed to heal disk format for {set_disk_id}: {e}"),
+                            message: format!("Failed to heal disk format for {set_disk_id}: {error}"),
                        });
                    }
                } else {
@@ -284,6 +288,9 @@ impl HealTask {
            Err(Error::TaskCancelled) => return Err(Error::TaskCancelled),
            Err(Error::TaskTimeout) => return Err(Error::TaskTimeout),
            Err(e) => {
+                if e.is_recoverable_heal() {
+                    return Err(e);
+                }
                error!(
                    target: "rustfs::heal::task",
                    event = EVENT_HEAL_ERASURE_SET_RESULT,
@@ -297,7 +304,7 @@ impl HealTask {
                );
                {
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(4, 4, 0, 0);
+                    progress.update_stage(4, 4);
                }
                return Err(Error::TaskExecutionFailed {
                    message: format!("Failed to heal disk format for {set_disk_id}: {e}"),
@@ -307,7 +314,7 @@ impl HealTask {

        {
            let mut progress = self.progress.write().await;
-            progress.update_progress(1, 4, 0, 0);
+            progress.update_stage(1, 4);
        }

        // The rebuilt disks are formatted now: mark them as healing so
@@ -336,7 +343,7 @@ impl HealTask {

        {
            let mut progress = self.progress.write().await;
-            progress.update_progress(2, 4, 0, 0);
+            progress.update_stage(2, 4);
        }

        // Step 3: Heal bucket structure
@@ -420,7 +427,7 @@ impl HealTask {

        {
            let mut progress = self.progress.write().await;
-            progress.update_progress(3, 4, 0, 0);
+            progress.update_stage(3, 4);
        }

        // Step 4: Execute erasure set heal with resume
@@ -463,9 +470,7 @@ impl HealTask {
        };

        {
-            let mut progress = self.progress.write().await;
-            let bytes_processed = progress.bytes_processed;
-            progress.update_progress(4, 4, 0, bytes_processed);
+            self.progress.write().await.update_stage(4, 4);
        }

        match result {
@@ -32,7 +32,7 @@ impl HealTask {
        {
            let mut progress = self.progress.write().await;
            progress.set_current_object(Some(format!("metadata: {bucket}/{object}")));
-            progress.update_progress(0, 3, 0, 0);
+            progress.update_stage(0, 3);
        }

        // Step 1: Check if object exists
@@ -74,7 +74,7 @@ impl HealTask {

        {
            let mut progress = self.progress.write().await;
-            progress.update_progress(1, 3, 0, 0);
+            progress.update_stage(1, 3);
        }

        // Step 2: Perform metadata heal using ecstore
@@ -122,7 +122,7 @@ impl HealTask {
                    );
                    {
                        let mut progress = self.progress.write().await;
-                        progress.update_progress(3, 3, 0, 0);
+                        progress.update_stage(3, 3);
                    }
                    return Err(Error::TaskExecutionFailed {
                        message: format!("Failed to heal metadata {bucket}/{object}: {e}"),
@@ -145,7 +145,7 @@ impl HealTask {

                {
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                }
                self.record_result_item(result).await;
                Ok(())
@@ -167,7 +167,7 @@ impl HealTask {
                );
                {
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                }
                Err(Error::TaskExecutionFailed {
                    message: format!("Failed to heal metadata {bucket}/{object}: {e}"),
@@ -194,7 +194,7 @@ impl HealTask {
        {
            let mut progress = self.progress.write().await;
            progress.set_current_object(Some(format!("ec_decode: {bucket}/{object}")));
-            progress.update_progress(0, 3, 0, 0);
+            progress.update_stage(0, 3);
        }

        // Step 1: Check if object exists
@@ -236,7 +236,7 @@ impl HealTask {

        {
            let mut progress = self.progress.write().await;
-            progress.update_progress(1, 3, 0, 0);
+            progress.update_stage(1, 3);
        }

        // Step 2: Perform EC decode heal using ecstore
@@ -284,7 +284,7 @@ impl HealTask {
                    );
                    {
                        let mut progress = self.progress.write().await;
-                        progress.update_progress(3, 3, 0, 0);
+                        progress.update_stage(3, 3);
                    }
                    return Err(Error::TaskExecutionFailed {
                        message: format!("Failed to heal EC decode {bucket}/{object}: {e}"),
@@ -309,7 +309,7 @@ impl HealTask {

                {
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, object_size);
+                    progress.update_object_progress(1, 1, 0, 0, object_size);
                }
                self.record_result_item(result).await;
                Ok(())
@@ -331,7 +331,7 @@ impl HealTask {
                );
                {
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                }
                Err(Error::TaskExecutionFailed {
                    message: format!("Failed to heal EC decode {bucket}/{object}: {e}"),
@@ -36,7 +36,7 @@ impl HealTask {
        {
            let mut progress = self.progress.write().await;
            progress.set_current_object(Some(format!("{bucket}/{object}")));
-            progress.update_progress(0, 4, 0, 0);
+            progress.update_stage(0, 4);
        }

        // Step 1: Check if object exists and get metadata
@@ -132,7 +132,7 @@ impl HealTask {

        {
            let mut progress = self.progress.write().await;
-            progress.update_progress(1, 3, 0, 0);
+            progress.update_stage(1, 3);
        }

        // Step 2: directly call ecstore to perform heal
@@ -187,7 +187,7 @@ impl HealTask {
                        );
                        {
                            let mut progress = self.progress.write().await;
-                            progress.update_progress(3, 3, 0, 0);
+                            progress.update_stage(3, 3);
                        }
                        return Ok(());
                    }
@@ -207,7 +207,7 @@ impl HealTask {

                    {
                        let mut progress = self.progress.write().await;
-                        progress.update_progress(3, 3, 0, 0);
+                        progress.update_stage(3, 3);
                    }

                    if Self::should_return_typed_heal_error(&e) {
@@ -249,7 +249,7 @@ impl HealTask {

                {
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, object_size);
+                    progress.update_object_progress(1, 1, 0, 0, object_size);
                }
                self.record_result_item(result).await;
                Ok(())
@@ -275,7 +275,7 @@ impl HealTask {
                    );
                    {
                        let mut progress = self.progress.write().await;
-                        progress.update_progress(3, 3, 0, 0);
+                        progress.update_stage(3, 3);
                    }
                    return Ok(());
                }
@@ -295,7 +295,7 @@ impl HealTask {

                {
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(3, 3, 0, 0);
+                    progress.update_stage(3, 3);
                }

                if Self::should_return_typed_heal_error(&e) {
@@ -414,7 +414,7 @@ impl HealTask {

                {
                    let mut progress = self.progress.write().await;
-                    progress.update_progress(4, 4, 0, object_size);
+                    progress.update_object_progress(1, 1, 0, 0, object_size);
                }
                self.record_result_item(result).await;
                Ok(())
@@ -22,6 +22,7 @@ use std::sync::Mutex;
 use tempfile::TempDir;

 use super::super::storage_api::status::BucketInfo;
+use crate::heal::progress::HealProgressState;

 #[tokio::test]
 async fn retry_request_carries_remaining_timeout_budget() {
@@ -547,6 +548,7 @@ struct MockStorage {
    heal_object_outcome: Mutex<Option<MockHealObjectOutcome>>,
    heal_object_outcomes: Mutex<HashMap<String, VecDeque<MockHealObjectOutcome>>>,
    format_no_heal_required: Mutex<bool>,
+    format_error: Mutex<Option<Error>>,
    global_format_calls: Mutex<u32>,
    replacement_format_calls: Mutex<Vec<(usize, usize, Vec<String>)>>,
    replacement_targets_ready: Mutex<bool>,
@@ -867,6 +869,9 @@ impl HealStorageAPI for MockStorage {

    async fn heal_format(&self, _dry_run: bool) -> Result<(HealResultItem, Option<Error>)> {
        *self.global_format_calls.lock().unwrap() += 1;
+        if let Some(error) = self.format_error.lock().unwrap().take() {
+            return Err(error);
+        }
        let no_heal_required = *self.format_no_heal_required.lock().unwrap();
        if no_heal_required {
            Ok((HealResultItem::default(), Some(Error::Storage(EcstoreError::NoHealRequired))))
@@ -2052,6 +2057,30 @@ async fn test_erasure_set_heal_continues_after_format_no_heal_required() {
    );
 }

+#[tokio::test]
+async fn erasure_set_format_slowdown_is_propagated() {
+    let storage = Arc::new(MockStorage {
+        format_error: Mutex::new(Some(Error::Storage(EcstoreError::SlowDown))),
+        ..Default::default()
+    });
+    let request = HealRequest::new(
+        HealType::ErasureSet {
+            buckets: Vec::new(),
+            set_disk_id: "pool_0_set_0".to_string(),
+        },
+        HealOptions::default(),
+        HealPriority::Normal,
+    );
+    let task = HealTask::from_request(request, storage);
+
+    let error = task
+        .execute()
+        .await
+        .expect_err("format SlowDown must remain recoverable for the task manager");
+
+    assert!(matches!(error, Error::Storage(EcstoreError::SlowDown)));
+}
+
 #[tokio::test]
 async fn erasure_set_bucket_prepass_failure_stops_before_object_heal() {
    let temp = TempDir::new().expect("temporary directory should be created");
@@ -2096,6 +2125,7 @@ async fn erasure_set_heal_applies_usage_baseline_to_progress() {
        usage_baseline: Mutex::new(Some(HealBucketUsageBaseline {
            objects_count: 10,
            bytes: 8,
+            generation: Some(1),
        })),
        ..Default::default()
    });
@@ -2119,10 +2149,55 @@ async fn erasure_set_heal_applies_usage_baseline_to_progress() {
    let progress = task.get_progress().await;
    assert_eq!(progress.objects_total_count, 10);
    assert_eq!(progress.objects_total_size, 8);
+    assert!(progress.baseline_generation.is_some());
+    assert!(progress.baseline_known);
    assert_eq!(progress.bytes_processed, 2);
    assert!((progress.progress_percentage - 25.0).abs() < 0.001);
 }

+#[tokio::test]
+async fn erasure_set_disk_walk_keeps_cluster_usage_baseline_indeterminate() {
+    for (scan_mode, source) in [
+        (HealScanMode::Deep, HealRequestSource::Admin),
+        (HealScanMode::Normal, HealRequestSource::AutoHeal),
+    ] {
+        let temp = TempDir::new().expect("temporary directory should be created");
+        let disk = make_resume_disk(&temp).await;
+        let storage = Arc::new(MockStorage {
+            resume_disk: Mutex::new(Some(disk)),
+            usage_baseline: Mutex::new(Some(HealBucketUsageBaseline {
+                objects_count: 10,
+                bytes: 8,
+                generation: Some(1),
+            })),
+            ..Default::default()
+        });
+        let mut request = HealRequest::new(
+            HealType::ErasureSet {
+                buckets: vec!["bucket-a".to_string()],
+                set_disk_id: "pool_0_set_0".to_string(),
+            },
+            HealOptions {
+                scan_mode,
+                timeout: None,
+                ..Default::default()
+            },
+            HealPriority::Normal,
+        );
+        request.source = source;
+        let task = HealTask::from_request(request, storage);
+
+        task.heal_erasure_set(vec!["bucket-a".to_string()], "pool_0_set_0".to_string())
+            .await
+            .expect("erasure set heal should complete");
+
+        let progress = task.get_progress().await;
+        assert!(!progress.baseline_known);
+        assert_eq!(progress.baseline_generation, None);
+        assert_eq!(progress.progress_state, HealProgressState::Indeterminate);
+    }
+}
+
 #[tokio::test]
 async fn erasure_set_heal_ignores_usage_baseline_errors() {
    let temp = TempDir::new().expect("temporary directory should be created");
@@ -19,7 +19,7 @@ pub use error::{Error, Result};
 pub use heal::{
    HealManager, HealOperationsSnapshot, HealOptions, HealPriority, HealPriorityCounts, HealRequest, HealSourceCounts, HealType,
    channel::HealChannelProcessor,
-    progress::HealProgress,
+    progress::{HealProgress, aggregate_heal_progress},
    resume::{ReplacementRecoveryRecord, ReplacementRecoveryState, ResumeUtils},
 };
 use rustfs_concurrency::WorkloadAdmissionSnapshotProvider;
@@ -54,6 +54,13 @@ pub const INTERNODE_STAGE_READ_VERSION_RESPONSE_JSON_ENCODE: &str = "read_versio
 pub const INTERNODE_STAGE_READ_VERSION_RESPONSE_MSGPACK_ENCODE: &str = "read_version_response_msgpack_encode";
 pub const INTERNODE_STAGE_READ_VERSION_RPC_ROUNDTRIP: &str = "read_version_rpc_roundtrip";
 pub const INTERNODE_STAGE_READ_VERSION_RESPONSE_DECODE: &str = "read_version_response_decode";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_REQUEST_ENCODE: &str = "batch_read_version_request_encode";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_REQUEST_DECODE: &str = "batch_read_version_request_decode";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_DISK_READ: &str = "batch_read_version_disk_read";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_JSON_ENCODE: &str = "batch_read_version_response_json_encode";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_MSGPACK_ENCODE: &str = "batch_read_version_response_msgpack_encode";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_RPC_ROUNDTRIP: &str = "batch_read_version_rpc_roundtrip";
+pub const INTERNODE_STAGE_BATCH_READ_VERSION_RESPONSE_DECODE: &str = "batch_read_version_response_decode";

 const OPERATION_LABEL: &str = "operation";
 const BACKEND_LABEL: &str = "backend";
@@ -689,6 +689,14 @@ pub struct ScannerMetrics {
    pub cycle_max_objects: u64,
    #[serde(rename = "cycle_max_directories", default)]
    pub cycle_max_directories: u64,
+    #[serde(rename = "cycle_timeout_total", default)]
+    pub cycle_timeout_total: u64,
+    #[serde(rename = "cycle_recovery_required_total", default)]
+    pub cycle_recovery_required_total: u64,
+    #[serde(rename = "cycle_last_progress_age", default)]
+    pub cycle_last_progress_age: u64,
+    #[serde(rename = "leader_lease_without_progress", default)]
+    pub leader_lease_without_progress: bool,
    #[serde(rename = "bitrot_cycle_enabled", default)]
    pub bitrot_cycle_enabled: bool,
    #[serde(rename = "bitrot_cycle_seconds", default)]
@@ -764,6 +772,8 @@ impl ScannerMetrics {
            self.cycle_max_duration_seconds = other.cycle_max_duration_seconds;
            self.cycle_max_objects = other.cycle_max_objects;
            self.cycle_max_directories = other.cycle_max_directories;
+            self.cycle_last_progress_age = other.cycle_last_progress_age;
+            self.leader_lease_without_progress = other.leader_lease_without_progress;
            self.bitrot_cycle_enabled = other.bitrot_cycle_enabled;
            self.bitrot_cycle_seconds = other.bitrot_cycle_seconds;
        }
@@ -857,6 +867,12 @@ impl ScannerMetrics {
            .saturating_add(other.last_cycle_replication_checks);
        self.last_cycle_usage_saves = self.last_cycle_usage_saves.saturating_add(other.last_cycle_usage_saves);
        self.failed_cycles = self.failed_cycles.saturating_add(other.failed_cycles);
+        self.cycle_timeout_total = self.cycle_timeout_total.saturating_add(other.cycle_timeout_total);
+        self.cycle_recovery_required_total = self
+            .cycle_recovery_required_total
+            .saturating_add(other.cycle_recovery_required_total);
+        self.cycle_last_progress_age = self.cycle_last_progress_age.max(other.cycle_last_progress_age);
+        self.leader_lease_without_progress |= other.leader_lease_without_progress;
        self.superseded_cycles = self.superseded_cycles.saturating_add(other.superseded_cycles);
        self.partial_cycles_unknown = self.partial_cycles_unknown.saturating_add(other.partial_cycles_unknown);
        self.partial_cycles_runtime = self.partial_cycles_runtime.saturating_add(other.partial_cycles_runtime);
@@ -195,7 +195,8 @@ pub struct BucketPolicyArgs<'a> {
 #[derive(Serialize, Deserialize, Clone, Default, Debug)]
 #[serde(deny_unknown_fields)]
 pub struct BucketPolicy {
-    #[serde(default, rename = "Id", skip_serializing_if = "ID::is_empty")]
+    // RUSTFS_COMPAT_TODO(rustfs-6339): accept bucket policies persisted with the legacy "ID" key. Remove after migration tooling rewrites every retained legacy bucket policy.
+    #[serde(default, rename = "Id", alias = "ID", skip_serializing_if = "ID::is_empty")]
    pub id: ID,
    #[serde(rename = "Version")]
    pub version: String,
@@ -2786,7 +2787,7 @@ mod test {
        let parsed: serde_json::Value = serde_json::from_str(&json).expect("Should parse");

        // Verify empty fields are omitted
-        assert!(!parsed.as_object().unwrap().contains_key("ID"), "Empty ID should be omitted");
+        assert!(parsed.get("Id").is_none(), "Empty ID should be omitted");

        let statement = &parsed["Statement"][0];
        assert!(!statement.as_object().unwrap().contains_key("Sid"), "Empty Sid should be omitted");
@@ -2809,6 +2810,43 @@ mod test {
        assert_eq!(statement["Principal"]["AWS"], "*");
    }

+    #[test]
+    fn test_bucket_policy_deserializes_legacy_id() {
+        let legacy_policy = br#"{"ID":"","Version":"2012-10-17","Statement":[{"Sid":"","Effect":"Allow","Principal":{"AWS":["*"]},"Action":["s3:GetObject"],"NotAction":[],"Resource":["arn:aws:s3:::bucket/*"],"NotResource":[],"Condition":{}}]}"#;
+
+        let policy: BucketPolicy =
+            serde_json::from_slice(legacy_policy).expect("bucket policy with legacy ID should deserialize");
+        assert!(policy.id.is_empty());
+        policy.is_valid().expect("legacy bucket policy should remain valid");
+
+        let policy: BucketPolicy = serde_json::from_str(r#"{"ID":"legacy-policy","Version":"2012-10-17","Statement":[]}"#)
+            .expect("non-empty legacy ID should deserialize");
+        assert_eq!(policy.id.0, "legacy-policy");
+
+        let serialized = serde_json::to_value(&policy).expect("bucket policy should serialize");
+        assert_eq!(serialized["Id"], "legacy-policy");
+        assert!(serialized.get("ID").is_none(), "legacy ID spelling should not be serialized");
+    }
+
+    #[test]
+    fn test_bucket_policy_legacy_id_alias_remains_strict() {
+        let unknown_field = r#"{"Version":"2012-10-17","Statement":[],"Unexpected":true}"#;
+        let error =
+            serde_json::from_str::<BucketPolicy>(unknown_field).expect_err("unrelated unknown fields should remain rejected");
+        assert!(
+            error.to_string().contains("unknown field `Unexpected`"),
+            "unexpected deserialization error: {error}"
+        );
+
+        let duplicate_id = r#"{"Id":"current-policy","ID":"legacy-policy","Version":"2012-10-17","Statement":[]}"#;
+        let error = serde_json::from_str::<BucketPolicy>(duplicate_id)
+            .expect_err("canonical and legacy ID fields should not be accepted together");
+        assert!(
+            error.to_string().contains("duplicate field `Id`"),
+            "unexpected deserialization error: {error}"
+        );
+    }
+
    #[test]
    fn test_existing_object_tag_condition_helpers() {
        let identity_policy = Policy::parse_config(
@@ -1215,7 +1215,10 @@ pub struct ScannerActivityResponse {
    pub dirty_usage_pending: bool,
 }
 #[derive(Clone, Copy, PartialEq, Eq, Hash, ::prost::Message)]
-pub struct BackgroundHealStatusRequest {}
+pub struct BackgroundHealStatusRequest {
+    #[prost(uint32, tag = "1")]
+    pub protocol_version: u32,
+}
 #[derive(Clone, PartialEq, Eq, Hash, ::prost::Message)]
 pub struct BackgroundHealStatusResponse {
    #[prost(bool, tag = "1")]
@@ -170,6 +170,7 @@ pub fn internode_rpc_max_message_size() -> usize {
 pub const HEAL_CONTROL_RPC_MAX_MESSAGE_SIZE: usize = heal_control::RESULT_MAX_SIZE + 1024;
 pub const HEAL_CONTROL_PROTOCOL_VERSION: u32 = 3;
 pub const DYNAMIC_CONFIG_PROTOCOL_VERSION: u32 = 1;
+pub const BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION: u32 = 2;
 pub const HEAL_CONTROL_CAPABILITY_PROBE_PREFIX: &[u8] = b"rustfs-heal-control-capability-v3\0";
 pub const REMOTE_VERSION_STATE_CAPABILITY_PROBE_PREFIX: &[u8] = b"rustfs-tier-remote-version-state-capability-v1\0";
 pub const CROSS_POOL_FENCE_CAPABILITY_PROBE_PREFIX: &[u8] = b"rustfs-cross-pool-fence-capability-v1\0";
@@ -2346,10 +2347,28 @@ pub async fn evict_failed_connection_with_log_level(addr: &str, log_level: Conne
 #[cfg(test)]
 mod tests {
    use super::*;
+    use prost::Message as _;
    use std::sync::Mutex;

    static INTERNODE_RPC_MSGPACK_ONLY_ENV_LOCK: Mutex<()> = Mutex::new(());

+    #[derive(Clone, PartialEq, prost::Message)]
+    struct BackgroundHealStatusRequestV1 {}
+
+    #[test]
+    fn background_heal_status_request_remains_rolling_upgrade_compatible() {
+        let current = proto_gen::node_service::BackgroundHealStatusRequest {
+            protocol_version: BACKGROUND_HEAL_STATUS_PROTOCOL_VERSION,
+        };
+        let encoded = current.encode_to_vec();
+        BackgroundHealStatusRequestV1::decode(encoded.as_slice()).expect("v1 server should ignore the version field");
+
+        let encoded = BackgroundHealStatusRequestV1 {}.encode_to_vec();
+        let decoded = proto_gen::node_service::BackgroundHealStatusRequest::decode(encoded.as_slice())
+            .expect("v2 server should accept a v1 request");
+        assert_eq!(decoded.protocol_version, 0);
+    }
+
    #[derive(Clone, Copy, Debug, Eq, Ord, PartialEq, PartialOrd)]
    struct CompatPayloadField {
        message: &'static str,
@@ -846,7 +846,9 @@ message ScannerActivityResponse {
  bool dirty_usage_pending = 9;
 }

-message BackgroundHealStatusRequest {}
+message BackgroundHealStatusRequest {
+  uint32 protocol_version = 1;
+}

 message BackgroundHealStatusResponse {
  bool success = 1;
@@ -16,7 +16,7 @@ use super::persistence::DataUsageCacheLoadAttempt;
 use super::*;
 use crate::storage_api::scanner_io::{HTTPRangeSpec, ObjectIO};
 use crate::{ScannerGetObjectReader, ScannerPutObjReader};
-use rustfs_data_usage::{ReplicationAllStats, ReplicationStats};
+use rustfs_data_usage::{ReplicationAllStats, ReplicationTargetUsage};
 use serde_json::Value;
 use std::io::Cursor;
 use std::pin::Pin;
@@ -1636,7 +1636,7 @@ fn size_recursive_prunes_empty_and_preserves_threshold_replication_stats() {
            replication_stats: Some(ReplicationAllStats {
                targets: HashMap::from([(
                    "arn:test:threshold".to_string(),
-                    ReplicationStats {
+                    ReplicationTargetUsage {
                        after_threshold_count: 1,
                        ..Default::default()
                    },
@@ -125,7 +125,10 @@ impl Default for ScannerRuntimeConfig {
            cycle_interval_source: ScannerRuntimeConfigSource::Default,
            bitrot_cycle: Some(Duration::from_secs(DEFAULT_HEAL_BITROT_CYCLE_SECS)),
            bitrot_cycle_source: ScannerRuntimeConfigSource::Default,
-            cycle_budget: ScannerCycleBudgetConfig::default(),
+            cycle_budget: ScannerCycleBudgetConfig {
+                max_duration: Some(Duration::from_secs(DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS)),
+                ..Default::default()
+            },
            cycle_max_duration_source: ScannerRuntimeConfigSource::Default,
            cycle_max_objects_source: ScannerRuntimeConfigSource::Default,
            cycle_max_directories_source: ScannerRuntimeConfigSource::Default,
@@ -374,7 +377,10 @@ fn validate_persisted_scanner_runtime_config(config: &ServerConfig) -> Result<()
    }
    validate_optional_config_u64(scanner_kvs, SCANNER_START_DELAY, "")?;
    validate_optional_config_u64(scanner_kvs, SCANNER_CYCLE, "")?;
-    validate_optional_config_u64(scanner_kvs, SCANNER_CYCLE_MAX_DURATION, DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS)?;
+    if let Some(value) = config_value(scanner_kvs, SCANNER_CYCLE_MAX_DURATION, DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS) {
+        let secs = parse_config_u64(SCANNER_CYCLE_MAX_DURATION, value)?;
+        cycle_duration_from_secs(SCANNER_CYCLE_MAX_DURATION, secs)?;
+    }
    validate_optional_config_u64(scanner_kvs, SCANNER_CYCLE_MAX_OBJECTS, DEFAULT_SCANNER_CYCLE_MAX_OBJECTS)?;
    validate_optional_config_u64(scanner_kvs, SCANNER_CYCLE_MAX_DIRECTORIES, DEFAULT_SCANNER_CYCLE_MAX_DIRECTORIES)?;
    if let Some(value) = config_value(heal_kvs, HEAL_BITROT_CYCLE, DEFAULT_HEAL_BITROT_CYCLE_SECS) {
@@ -436,19 +442,46 @@ fn lookup_max_wait(
    Ok((speed.max_sleep(), speed_source))
 }

-fn lookup_optional_seconds(
-    kvs: Option<&KVS>,
-    key: &'static str,
-    env_key: &'static str,
-    default: u64,
-) -> Result<(Option<Duration>, ScannerRuntimeConfigSource), ScannerRuntimeConfigError> {
-    if let Some(secs) = rustfs_utils::get_env_opt_u64(env_key) {
-        return Ok((Some(Duration::from_secs(secs)), ScannerRuntimeConfigSource::Env));
+fn lookup_cycle_duration(kvs: Option<&KVS>) -> Result<(Option<Duration>, ScannerRuntimeConfigSource), ScannerRuntimeConfigError> {
+    match rustfs_utils::get_env_parse_outcome::<u64>(ENV_SCANNER_CYCLE_MAX_DURATION_SECS) {
+        rustfs_utils::EnvParseOutcome::Parsed(secs) => {
+            return cycle_duration_from_secs(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, secs)
+                .map(|duration| (duration, ScannerRuntimeConfigSource::Env));
+        }
+        rustfs_utils::EnvParseOutcome::Invalid => {
+            // Do not include the raw environment value in the typed error:
+            // deployments occasionally put sensitive material in inherited
+            // environment snapshots. The key still identifies the control.
+            return Err(invalid_value(
+                ENV_SCANNER_CYCLE_MAX_DURATION_SECS,
+                "<invalid>",
+                "expected unsigned integer seconds",
+            ));
+        }
+        rustfs_utils::EnvParseOutcome::Absent => {}
    }
-    if let Some(value) = config_value(kvs, key, default) {
-        return parse_config_u64(key, value).map(|secs| (Some(Duration::from_secs(secs)), ScannerRuntimeConfigSource::Config));
+
+    if let Some(value) = config_value(kvs, SCANNER_CYCLE_MAX_DURATION, DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS) {
+        let secs = parse_config_u64(SCANNER_CYCLE_MAX_DURATION, value)?;
+        return cycle_duration_from_secs(SCANNER_CYCLE_MAX_DURATION, secs)
+            .map(|duration| (duration, ScannerRuntimeConfigSource::Config));
    }
-    Ok((None, ScannerRuntimeConfigSource::Default))
+
+    Ok((
+        Some(Duration::from_secs(DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS)),
+        ScannerRuntimeConfigSource::Default,
+    ))
+}
+
+fn cycle_duration_from_secs(key: &'static str, secs: u64) -> Result<Option<Duration>, ScannerRuntimeConfigError> {
+    if secs == 0 {
+        return Ok(None);
+    }
+    let duration = Duration::from_secs(secs);
+    if std::time::Instant::now().checked_add(duration).is_none() {
+        return Err(invalid_value(key, "<overflow>", "duration exceeds the timer range"));
+    }
+    Ok(Some(duration))
 }

 fn lookup_start_delay(kvs: Option<&KVS>) -> Result<(Option<Duration>, ScannerRuntimeConfigSource), ScannerRuntimeConfigError> {
@@ -553,12 +586,7 @@ pub(crate) fn lookup_scanner_runtime_config(
        (speed.cycle_interval(), speed_source)
    };

-    let (cycle_max_duration, cycle_max_duration_source) = lookup_optional_seconds(
-        scanner_kvs,
-        SCANNER_CYCLE_MAX_DURATION,
-        ENV_SCANNER_CYCLE_MAX_DURATION_SECS,
-        DEFAULT_SCANNER_CYCLE_MAX_DURATION_SECS,
-    )?;
+    let (cycle_max_duration, cycle_max_duration_source) = lookup_cycle_duration(scanner_kvs)?;
    let (cycle_max_objects, cycle_max_objects_source) = lookup_count_budget(
        scanner_kvs,
        SCANNER_CYCLE_MAX_OBJECTS,
@@ -863,10 +891,10 @@ mod tests {
    use rustfs_config::server_config::{Config as ServerConfig, KVS};
    use rustfs_config::{
        DEFAULT_DELIMITER, DEFAULT_HEAL_BITROT_CYCLE_SECS, ENV_SCANNER_BITROT_CYCLE_SECS, ENV_SCANNER_CACHE_SAVE_TIMEOUT_SECS,
-        ENV_SCANNER_CYCLE, ENV_SCANNER_CYCLE_MAX_OBJECTS, ENV_SCANNER_DELAY, ENV_SCANNER_MAX_WAIT_SECS, ENV_SCANNER_SPEED,
-        HEAL_BITROT_CYCLE, HEAL_SUB_SYS, SCANNER_BITROT_CYCLE, SCANNER_CACHE_SAVE_TIMEOUT, SCANNER_CYCLE,
-        SCANNER_CYCLE_MAX_DIRECTORIES, SCANNER_CYCLE_MAX_DURATION, SCANNER_CYCLE_MAX_OBJECTS, SCANNER_DELAY, SCANNER_IDLE_MODE,
-        SCANNER_SPEED, SCANNER_SUB_SYS, ScannerSpeed,
+        ENV_SCANNER_CYCLE, ENV_SCANNER_CYCLE_MAX_DURATION_SECS, ENV_SCANNER_CYCLE_MAX_OBJECTS, ENV_SCANNER_DELAY,
+        ENV_SCANNER_MAX_WAIT_SECS, ENV_SCANNER_SPEED, HEAL_BITROT_CYCLE, HEAL_SUB_SYS, SCANNER_BITROT_CYCLE,
+        SCANNER_CACHE_SAVE_TIMEOUT, SCANNER_CYCLE, SCANNER_CYCLE_MAX_DIRECTORIES, SCANNER_CYCLE_MAX_DURATION,
+        SCANNER_CYCLE_MAX_OBJECTS, SCANNER_DELAY, SCANNER_IDLE_MODE, SCANNER_SPEED, SCANNER_SUB_SYS, ScannerSpeed,
    };
    use std::collections::HashMap;
    use std::time::Duration;
@@ -941,6 +969,50 @@ mod tests {
        });
    }

+    #[test]
+    fn scanner_unset_budget_uses_safe_default_but_explicit_zero_is_unbounded() {
+        let config = server_config_with_scanner(&[]);
+        with_var_unset(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, || {
+            let resolved = lookup_scanner_runtime_config(Some(&config)).expect("scanner runtime config");
+            assert_eq!(resolved.cycle_budget.max_duration, Some(Duration::from_secs(1800)));
+            assert_eq!(resolved.cycle_max_duration_source, ScannerRuntimeConfigSource::Default);
+        });
+
+        let config = server_config_with_scanner(&[(SCANNER_CYCLE_MAX_DURATION, "0")]);
+        with_var_unset(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, || {
+            let resolved = lookup_scanner_runtime_config(Some(&config)).expect("scanner runtime config");
+            assert_eq!(resolved.cycle_budget.max_duration, None);
+            assert_eq!(resolved.cycle_max_duration_source, ScannerRuntimeConfigSource::Config);
+        });
+    }
+
+    #[test]
+    fn cycle_budget_invalid_or_overflow_config_is_rejected() {
+        with_var(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, Some("invalid"), || {
+            let error = lookup_scanner_runtime_config(None).expect_err("invalid duration env must be rejected");
+            assert!(error.to_string().contains(ENV_SCANNER_CYCLE_MAX_DURATION_SECS));
+            assert!(error.to_string().contains("<invalid>"));
+            assert!(!error.to_string().contains(": invalid ("));
+        });
+        with_var(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, Some("18446744073709551616"), || {
+            assert!(lookup_scanner_runtime_config(None).is_err());
+        });
+        with_var(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, Some("18446744073709551615"), || {
+            assert!(lookup_scanner_runtime_config(None).is_err());
+        });
+        let config = server_config_with_scanner(&[(SCANNER_CYCLE_MAX_DURATION, "not-a-duration")]);
+        assert!(lookup_scanner_runtime_config(Some(&config)).is_err());
+    }
+
+    #[test]
+    fn scanner_runtime_config_validation_rejects_overflow_persisted_duration() {
+        let config = server_config_with_scanner(&[(SCANNER_CYCLE_MAX_DURATION, "18446744073709551615")]);
+
+        let error = validate_scanner_runtime_config(&config)
+            .expect_err("persisted duration that exceeds the timer range must be rejected");
+        assert!(error.to_string().contains(SCANNER_CYCLE_MAX_DURATION));
+    }
+
    #[test]
    fn scanner_runtime_config_normalizes_persisted_default_speed() {
        let config = server_config_with_scanner(&[(SCANNER_SPEED, "default")]);
@@ -52,6 +52,7 @@ use rustfs_config::{
 };
 use rustfs_config::{ENV_SCANNER_CYCLE, ENV_SCANNER_SPEED, ENV_SCANNER_START_DELAY_SECS};
 use rustfs_data_usage::observed_data_usage_is_newer;
+use rustfs_lock::NamespaceLockGuard;
 use serde::{Deserialize, Serialize};
 use sha2::{Digest as _, Sha256};
 use tokio::sync::{Notify, mpsc};
@@ -1037,20 +1038,116 @@ fn data_usage_persist_timeout() -> Duration {
    DataUsageCache::persistence_timeout()
 }

+#[cfg(not(test))]
+const SCANNER_CYCLE_EPOCH_FENCE_TIMEOUT: Duration = Duration::from_secs(30);
+#[cfg(test)]
+const SCANNER_CYCLE_EPOCH_FENCE_TIMEOUT: Duration = Duration::from_millis(50);
+
+async fn fence_scanner_epoch_after_cycle_timeout<Store, LockLost>(
+    ctx: &CancellationToken,
+    storeapi: Arc<Store>,
+    cycle_info: &mut CurrentCycle,
+    cycle_revision: &mut DataUsageCacheRevision,
+    leader_epoch: &mut u64,
+    lock_lost: LockLost,
+) -> bool
+where
+    Store: ScannerObjectIO,
+    LockLost: Future<Output = ()>,
+{
+    let fence_ctx = ctx.child_token();
+    let claim = claim_scanner_leadership(&fence_ctx, storeapi, cycle_info, cycle_revision, leader_epoch);
+    tokio::pin!(claim);
+    tokio::pin!(lock_lost);
+    tokio::select! {
+        biased;
+        _ = &mut lock_lost => {
+            fence_ctx.cancel();
+            false
+        }
+        result = tokio::time::timeout(SCANNER_CYCLE_EPOCH_FENCE_TIMEOUT, &mut claim) => {
+            result.unwrap_or(false) && !fence_ctx.is_cancelled()
+        }
+    }
+}
+
+struct ScannerCycleDeadlineState<'a> {
+    cycle_info: &'a mut CurrentCycle,
+    cycle_revision: &'a mut DataUsageCacheRevision,
+    leader_epoch: &'a mut u64,
+    cycle_budget: &'a ScannerCycleBudget,
+}
+
+fn cycle_timeout_requires_recovery(worker_stopped: bool, cycle_state_persisted: bool, generation_fenced: bool) -> bool {
+    !worker_stopped || !cycle_state_persisted || !generation_fenced
+}
+
+async fn handle_scanner_cycle_deadline<Store>(
+    ctx: &CancellationToken,
+    storeapi: Arc<Store>,
+    state: ScannerCycleDeadlineState<'_>,
+    worker_stopped: bool,
+    guard: &mut NamespaceLockGuard,
+) where
+    Store: ScannerObjectIO,
+{
+    let fenced = fence_scanner_epoch_after_cycle_timeout(
+        ctx,
+        storeapi,
+        state.cycle_info,
+        state.cycle_revision,
+        state.leader_epoch,
+        guard.lock_lost_notified(),
+    )
+    .await;
+    let cycle_state_persisted = state.cycle_budget.cycle_state_persisted();
+    let recovery_required = cycle_timeout_requires_recovery(worker_stopped, cycle_state_persisted, fenced);
+    warn!(
+        target: "rustfs::scanner",
+        event = EVENT_SCANNER_CYCLE_STATE,
+        component = LOG_COMPONENT_SCANNER,
+        subsystem = LOG_SUBSYSTEM_RUNTIME,
+        state = "cycle_timeout",
+        worker_stopped,
+        cycle_state_persisted,
+        generation_fenced = fenced,
+        recovery_required,
+        "Scanner cycle deadline expired; durable cursor/generation fencing completed when possible"
+    );
+    global_metrics().record_scanner_cycle_timeout(recovery_required, state.cycle_budget.progress_age());
+    // Stop renewing before releasing the lease. A new leader can then claim the
+    // higher persisted generation instead of inheriting the expired worker.
+    guard.release();
+    global_metrics().set_cycle(None).await;
+}
+
 async fn mark_scan_cycle_idle(cycle_info: &mut CurrentCycle, cycle_metrics_guard: &mut ScannerCycleMetricsGuard) {
    cycle_info.current = 0;
    global_metrics().clear_current_scan_mode();
    cycle_metrics_guard.finish(cycle_info.clone()).await;
 }

-#[instrument(skip_all)]
-#[hotpath::measure]
+#[cfg(test)]
 async fn run_data_scanner_cycle(
    ctx: &CancellationToken,
    storeapi: &Arc<ECStore>,
    cycle_info: &mut CurrentCycle,
    cycle_revision: &mut DataUsageCacheRevision,
    leader_epoch: u64,
+) -> ScannerCycleOutcome {
+    let cycle_budget = ScannerCycleBudget::new(ctx, scanner_cycle_budget_config());
+    run_data_scanner_cycle_with_budget(ctx, storeapi, cycle_info, cycle_revision, leader_epoch, cycle_budget).await
+}
+
+#[instrument(skip_all)]
+#[hotpath::measure]
+async fn run_data_scanner_cycle_with_budget(
+    ctx: &CancellationToken,
+    storeapi: &Arc<ECStore>,
+    cycle_info: &mut CurrentCycle,
+    cycle_revision: &mut DataUsageCacheRevision,
+    leader_epoch: u64,
+    cycle_budget: Arc<ScannerCycleBudget>,
 ) -> ScannerCycleOutcome {
    let _activity_guard = ScannerActivityGuard::new();
    if let Err(err) = refresh_scanner_runtime_config_from_global() {
@@ -1066,7 +1163,11 @@ async fn run_data_scanner_cycle(
    }
    let configured_cycle_interval = scanner_cycle_interval();
    let configured_bitrot_cycle = scanner_bitrot_cycle();
-    let cycle_budget_config = scanner_cycle_budget_config();
+    let cycle_budget_config = ScannerCycleBudgetConfig {
+        max_duration: cycle_budget.max_duration(),
+        max_objects: cycle_budget.max_objects(),
+        max_directories: cycle_budget.max_directories(),
+    };
    let usage_persist_timeout = data_usage_persist_timeout();
    global_metrics().record_scanner_cycle_config(
        configured_cycle_interval,
@@ -1137,7 +1238,6 @@ async fn run_data_scanner_cycle(
    let (sender, receiver) = mpsc::channel::<DataUsageInfo>(1);

    let done_cycle = Metrics::time(Metric::ScanCycle);
-    let cycle_budget = ScannerCycleBudget::new(ctx, cycle_budget_config);
    let scan_result = storeapi
        .clone()
        .nsscanner_with_status(
@@ -1277,7 +1377,7 @@ async fn run_data_scanner_cycle(
                "Scanner cycle is recovering to a newer durable cache generation"
            );
            emit_scan_cycle_partial_with_source(cycle_start.elapsed(), ScanCyclePartialReason::Unknown, None);
-            return if persist_required_scanner_cycle_floor(
+            let persisted = persist_required_scanner_cycle_floor(
                ctx,
                storeapi.clone(),
                cycle_info,
@@ -1286,8 +1386,9 @@ async fn run_data_scanner_cycle(
                required_cycle,
                &mut cycle_metrics_guard,
            )
-            .await
-            {
+            .await;
+            return if persisted {
+                cycle_budget.mark_cycle_state_persisted();
                ScannerCycleOutcome::Partial
            } else {
                ScannerCycleOutcome::Failed
@@ -1345,7 +1446,7 @@ async fn run_data_scanner_cycle(
            scan_cycle_partial_reason(budget_reason),
            scan_cycle_partial_source(budget_reason),
        );
-        return if finalize_partial_scan_cycle(
+        let persisted = finalize_partial_scan_cycle(
            ctx,
            storeapi.clone(),
            cycle_info,
@@ -1353,8 +1454,9 @@ async fn run_data_scanner_cycle(
            leader_epoch,
            &mut cycle_metrics_guard,
        )
-        .await
-        {
+        .await;
+        return if persisted {
+            cycle_budget.mark_cycle_state_persisted();
            ScannerCycleOutcome::Partial
        } else {
            ScannerCycleOutcome::Failed
@@ -1429,7 +1531,7 @@ async fn run_data_scanner_cycle(
                );
            }
            emit_scan_cycle_partial_with_source(cycle_start.elapsed(), ScanCyclePartialReason::Unknown, None);
-            return if finalize_partial_scan_cycle(
+            let persisted = finalize_partial_scan_cycle(
                ctx,
                storeapi.clone(),
                cycle_info,
@@ -1437,8 +1539,9 @@ async fn run_data_scanner_cycle(
                leader_epoch,
                &mut cycle_metrics_guard,
            )
-            .await
-            {
+            .await;
+            return if persisted {
+                cycle_budget.mark_cycle_state_persisted();
                ScannerCycleOutcome::Partial
            } else {
                ScannerCycleOutcome::Failed
@@ -1479,6 +1582,7 @@ async fn run_data_scanner_cycle(
            )
            .await
            {
+                cycle_budget.mark_cycle_state_persisted();
                emit_scan_cycle_superseded(cycle_start.elapsed());
                return ScannerCycleOutcome::Superseded;
            }
@@ -1511,6 +1615,7 @@ async fn run_data_scanner_cycle(
        emit_scan_cycle_complete(false, cycle_start.elapsed());
        return ScannerCycleOutcome::Failed;
    }
+    cycle_budget.mark_cycle_state_persisted();

    done_cycle();
    emit_scan_cycle_complete(true, cycle_start.elapsed());
@@ -1575,7 +1680,7 @@ async fn run_data_scanner_with_maintenance_state(
 ) -> Result<(), ScannerError> {
    reset_scanner_cycle_schedule();
    // Acquire leader lock (write lock) to ensure only one scanner runs
-    let guard = match storeapi.new_ns_lock(RUSTFS_META_BUCKET, "leader.lock").await {
+    let mut guard = match storeapi.new_ns_lock(RUSTFS_META_BUCKET, "leader.lock").await {
        Ok(ns_lock) => match ns_lock.get_write_lock_quiet(get_lock_acquire_timeout()).await {
            Ok(guard) => {
                record_scanner_leader_lock_state("acquired");
@@ -1740,13 +1845,49 @@ async fn run_data_scanner_with_maintenance_state(
            return Ok(());
        }
        let cycle_ctx = ctx.child_token();
-        let initial_outcome = await_scanner_cycle_with_lock_fence(
+        let cycle_budget = ScannerCycleBudget::new_with_runtime_progress_tracking(&cycle_ctx, scanner_cycle_budget_config());
+        let initial_outcome = match await_scanner_cycle_with_budget_fence(
            &cycle_ctx,
-            run_data_scanner_cycle(&cycle_ctx, &storeapi, &mut cycle_info, &mut cycle_revision, leader_epoch),
+            &cycle_budget,
+            run_data_scanner_cycle_with_budget(
+                &cycle_ctx,
+                &storeapi,
+                &mut cycle_info,
+                &mut cycle_revision,
+                leader_epoch,
+                cycle_budget.clone(),
+            ),
            guard.lock_lost_notified(),
        )
        .await
-        .unwrap_or(ScannerCycleOutcome::Failed);
+        {
+            ScannerCycleWaitOutcome::Completed(outcome) => outcome,
+            ScannerCycleWaitOutcome::LockLost => {
+                record_scanner_leader_lock_lost("Scanner leader lock lost during the initial cycle").await;
+                global_metrics().set_cycle(None).await;
+                return Ok(());
+            }
+            ScannerCycleWaitOutcome::Cancelled => {
+                global_metrics().set_cycle(None).await;
+                return Ok(());
+            }
+            ScannerCycleWaitOutcome::Deadline { worker_stopped } => {
+                handle_scanner_cycle_deadline(
+                    &ctx,
+                    storeapi.clone(),
+                    ScannerCycleDeadlineState {
+                        cycle_info: &mut cycle_info,
+                        cycle_revision: &mut cycle_revision,
+                        leader_epoch: &mut leader_epoch,
+                        cycle_budget: &cycle_budget,
+                    },
+                    worker_stopped,
+                    &mut guard,
+                )
+                .await;
+                return Ok(());
+            }
+        };
        superseded_backoff.record_retryable_cycle(initial_outcome == ScannerCycleOutcome::Superseded);
        deferred_backoff.record_retryable_cycle(matches!(initial_outcome, ScannerCycleOutcome::Deferred(_)));
        dirty_usage_generation_seen = dirty_generation_before_cycle;
@@ -1952,13 +2093,49 @@ async fn run_data_scanner_with_maintenance_state(
        }
        let dirty_generation_before_cycle = dirty_usage_generation();
        let cycle_ctx = ctx.child_token();
-        let outcome = await_scanner_cycle_with_lock_fence(
+        let cycle_budget = ScannerCycleBudget::new_with_runtime_progress_tracking(&cycle_ctx, scanner_cycle_budget_config());
+        let outcome = match await_scanner_cycle_with_budget_fence(
            &cycle_ctx,
-            run_data_scanner_cycle(&cycle_ctx, &storeapi, &mut cycle_info, &mut cycle_revision, leader_epoch),
+            &cycle_budget,
+            run_data_scanner_cycle_with_budget(
+                &cycle_ctx,
+                &storeapi,
+                &mut cycle_info,
+                &mut cycle_revision,
+                leader_epoch,
+                cycle_budget.clone(),
+            ),
            guard.lock_lost_notified(),
        )
        .await
-        .unwrap_or(ScannerCycleOutcome::Failed);
+        {
+            ScannerCycleWaitOutcome::Completed(outcome) => outcome,
+            ScannerCycleWaitOutcome::LockLost => {
+                record_scanner_leader_lock_lost("Scanner leader lock lost during a scanner cycle").await;
+                global_metrics().set_cycle(None).await;
+                return Ok(());
+            }
+            ScannerCycleWaitOutcome::Cancelled => {
+                global_metrics().set_cycle(None).await;
+                return Ok(());
+            }
+            ScannerCycleWaitOutcome::Deadline { worker_stopped } => {
+                handle_scanner_cycle_deadline(
+                    &ctx,
+                    storeapi.clone(),
+                    ScannerCycleDeadlineState {
+                        cycle_info: &mut cycle_info,
+                        cycle_revision: &mut cycle_revision,
+                        leader_epoch: &mut leader_epoch,
+                        cycle_budget: &cycle_budget,
+                    },
+                    worker_stopped,
+                    &mut guard,
+                )
+                .await;
+                return Ok(());
+            }
+        };
        superseded_backoff.record_retryable_cycle(outcome == ScannerCycleOutcome::Superseded);
        deferred_backoff.record_retryable_cycle(matches!(outcome, ScannerCycleOutcome::Deferred(_)));
        dirty_usage_generation_seen = dirty_generation_before_cycle;
@@ -1581,3 +1581,63 @@ where
        output = &mut cycle => Some(output),
    }
 }
+
+#[derive(Debug, PartialEq, Eq)]
+pub(super) enum ScannerCycleWaitOutcome<T> {
+    Completed(T),
+    LockLost,
+    Cancelled,
+    Deadline { worker_stopped: bool },
+}
+
+pub(super) async fn await_scanner_cycle_with_budget_fence<Cycle, LockLost>(
+    cycle_ctx: &CancellationToken,
+    budget: &ScannerCycleBudget,
+    cycle: Cycle,
+    lock_lost: LockLost,
+) -> ScannerCycleWaitOutcome<Cycle::Output>
+where
+    Cycle: Future,
+    LockLost: Future<Output = ()>,
+{
+    tokio::pin!(cycle);
+    tokio::pin!(lock_lost);
+    let deadline = async {
+        if let Some(deadline) = budget.deadline() {
+            tokio::time::sleep_until(deadline).await;
+        } else {
+            std::future::pending::<()>().await;
+        }
+    };
+    tokio::pin!(deadline);
+    tokio::select! {
+        biased;
+        _ = &mut lock_lost => {
+            cycle_ctx.cancel();
+            let _ = tokio::time::timeout(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT, &mut cycle).await;
+            ScannerCycleWaitOutcome::LockLost
+        }
+        _ = &mut deadline => {
+            budget.cancel_for_runtime();
+            // Let the budget cancellation reach the scanner first so it can
+            // persist a partial cursor. Only an uncooperative worker gets the
+            // parent cancellation, and it is dropped after the bounded window;
+            // the caller fences its epoch next.
+            let worker_stopped = if tokio::time::timeout(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT, &mut cycle)
+                .await
+                .is_ok()
+            {
+                true
+            } else {
+                cycle_ctx.cancel();
+                false
+            };
+            ScannerCycleWaitOutcome::Deadline { worker_stopped }
+        }
+        _ = cycle_ctx.cancelled() => {
+            let _ = tokio::time::timeout(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT, &mut cycle).await;
+            ScannerCycleWaitOutcome::Cancelled
+        }
+        output = &mut cycle => ScannerCycleWaitOutcome::Completed(output),
+    }
+}
@@ -26,6 +26,7 @@ use std::task::Poll;
 use temp_env::{with_var, with_var_unset};
 use tokio::io::AsyncReadExt;
 use tokio::sync::Mutex;
+use tokio::time::{Duration, advance};

 const TEST_DEFAULT_SCANNER_CYCLE_SECS: u64 = 24 * 60 * 60;

@@ -118,6 +119,178 @@ async fn scanner_cycle_lock_fence_bounds_uncooperative_shutdown() {
    assert!(cycle_ctx.is_cancelled());
 }

+#[tokio::test(start_paused = true)]
+async fn cycle_budget_fences_late_writer_after_timeout() {
+    let cycle_ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(
+        &cycle_ctx,
+        ScannerCycleBudgetConfig {
+            max_duration: Some(Duration::from_secs(5)),
+            ..Default::default()
+        },
+    );
+    let outcome = {
+        let cycle = std::future::pending::<()>();
+        let lock_lost = std::future::pending::<()>();
+        let waiter = await_scanner_cycle_with_budget_fence(&cycle_ctx, &budget, cycle, lock_lost);
+        tokio::pin!(waiter);
+        tokio::task::yield_now().await;
+        advance(Duration::from_secs(5)).await;
+        tokio::task::yield_now().await;
+        advance(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT).await;
+        waiter.await
+    };
+    assert_eq!(outcome, ScannerCycleWaitOutcome::Deadline { worker_stopped: false });
+    assert!(cycle_ctx.is_cancelled());
+    assert_eq!(budget.reason(), Some(ScannerCycleBudgetReason::Runtime));
+
+    // A newer leadership epoch is the durable fence that rejects a late
+    // writer after the timed-out future has been dropped.
+    let store = Arc::new(MemoryConfigStore::default());
+    let mut revision = DataUsageCacheRevision::Missing;
+    let mut cycle = CurrentCycle {
+        current: 0,
+        next: 12,
+        ..Default::default()
+    };
+    let persist_ctx = CancellationToken::new();
+    assert!(persist_scanner_cycle_state(&persist_ctx, store.clone(), &mut cycle, &mut revision, 1).await);
+    let newer = encode_scanner_cycle_state(&cycle, 2).expect("new epoch fence should encode");
+    let key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    store.interleaving_puts.lock().await.insert(key, (2, newer));
+    let mut late_cycle = CurrentCycle { next: 13, ..cycle };
+    assert!(!persist_scanner_cycle_state(&persist_ctx, store, &mut late_cycle, &mut revision, 1).await);
+}
+
+#[tokio::test(start_paused = true)]
+async fn cycle_budget_parent_cancellation_is_not_reported_as_timeout() {
+    let cycle_ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(
+        &cycle_ctx,
+        ScannerCycleBudgetConfig {
+            max_duration: Some(Duration::from_secs(5)),
+            ..Default::default()
+        },
+    );
+    let waiter = await_scanner_cycle_with_budget_fence(&cycle_ctx, &budget, std::future::pending::<()>(), std::future::pending());
+    tokio::pin!(waiter);
+    tokio::task::yield_now().await;
+    cycle_ctx.cancel();
+    tokio::task::yield_now().await;
+    advance(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT).await;
+    assert_eq!(waiter.await, ScannerCycleWaitOutcome::Cancelled);
+}
+
+#[tokio::test(start_paused = true)]
+async fn cycle_budget_deadline_wins_same_tick_as_parent_cancellation() {
+    let cycle_ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new(
+        &cycle_ctx,
+        ScannerCycleBudgetConfig {
+            max_duration: Some(Duration::from_secs(5)),
+            ..Default::default()
+        },
+    );
+    let waiter = await_scanner_cycle_with_budget_fence(&cycle_ctx, &budget, std::future::pending::<()>(), std::future::pending());
+    tokio::pin!(waiter);
+    tokio::task::yield_now().await;
+    advance(Duration::from_secs(5)).await;
+    cycle_ctx.cancel();
+    tokio::task::yield_now().await;
+    advance(SCANNER_LOCK_LOSS_SHUTDOWN_TIMEOUT).await;
+
+    assert_eq!(waiter.await, ScannerCycleWaitOutcome::Deadline { worker_stopped: false });
+    assert_eq!(budget.reason(), Some(ScannerCycleBudgetReason::Runtime));
+}
+
+#[tokio::test]
+async fn cycle_budget_persist_cursor_failure_is_recovery_required() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_BLOOM_NAME_PATH.as_str());
+    store.fail_put_number.lock().await.insert(key, 1);
+
+    let ctx = CancellationToken::new();
+    let mut revision = DataUsageCacheRevision::Missing;
+    let mut cycle = CurrentCycle {
+        current: 12,
+        next: 12,
+        ..Default::default()
+    };
+    let mut leader_epoch = 1;
+    let fenced = fence_scanner_epoch_after_cycle_timeout(
+        &ctx,
+        store,
+        &mut cycle,
+        &mut revision,
+        &mut leader_epoch,
+        std::future::pending(),
+    )
+    .await;
+    assert!(!fenced, "a failed cursor/generation write must require recovery");
+    let budget = ScannerCycleBudget::new(&ctx, ScannerCycleBudgetConfig::default());
+    assert!(cycle_timeout_requires_recovery(true, budget.cycle_state_persisted(), fenced));
+
+    let metrics = Metrics::new();
+    metrics.record_scanner_cycle_timeout(!fenced, Duration::from_secs(17));
+    let report = metrics.report().await;
+    assert_eq!(report.cycle_timeout_total, 1);
+    assert_eq!(report.cycle_recovery_required_total, 1);
+    assert_eq!(report.cycle_last_progress_age, 17);
+    assert!(report.leader_lease_without_progress);
+}
+
+#[tokio::test]
+async fn cycle_budget_deadline_handler_fences_and_releases_guard() {
+    let (_temp_dir, store) = setup_scanner_cycle_store().await;
+    let lock = store
+        .new_ns_lock(RUSTFS_META_BUCKET, "leader.lock")
+        .await
+        .expect("scanner leader lock should be created");
+    let mut guard = lock
+        .get_write_lock(Duration::from_secs(1))
+        .await
+        .expect("scanner leader lock should be acquired");
+
+    let ctx = CancellationToken::new();
+    let mut cycle_info = CurrentCycle {
+        current: 12,
+        next: 12,
+        ..Default::default()
+    };
+    let mut cycle_revision = DataUsageCacheRevision::Missing;
+    let mut leader_epoch = 1;
+    let budget = ScannerCycleBudget::new(
+        &ctx,
+        ScannerCycleBudgetConfig {
+            max_duration: Some(Duration::from_secs(60)),
+            ..Default::default()
+        },
+    );
+    budget.mark_cycle_state_persisted();
+
+    handle_scanner_cycle_deadline(
+        &ctx,
+        store.clone(),
+        ScannerCycleDeadlineState {
+            cycle_info: &mut cycle_info,
+            cycle_revision: &mut cycle_revision,
+            leader_epoch: &mut leader_epoch,
+            cycle_budget: &budget,
+        },
+        true,
+        &mut guard,
+    )
+    .await;
+
+    assert!(guard.is_released());
+    let persisted = read_config(store, &DATA_USAGE_BLOOM_NAME_PATH)
+        .await
+        .expect("deadline handler should persist a fenced cursor");
+    let (_, persisted_epoch) = decode_scanner_cycle_state(&persisted).expect("fenced cursor should decode");
+    assert_eq!(persisted_epoch, 2);
+    global_metrics().set_cycle(None).await;
+}
+
 #[tokio::test]
 async fn scanner_cycle_recovery_wake_survives_wait_registration_race() {
    notify_scanner_cycle_recovery_wake();
@@ -428,13 +601,6 @@ fn test_scanner_cycle_max_duration_uses_env() {
    });
 }

-#[test]
-fn test_scanner_cycle_max_duration_default_is_disabled() {
-    with_var_unset(ENV_SCANNER_CYCLE_MAX_DURATION_SECS, || {
-        assert_eq!(scanner_cycle_max_duration(), None);
-    });
-}
-
 #[tokio::test]
 async fn test_scanner_cycle_budget_cancels_after_duration() {
    let parent = CancellationToken::new();
@@ -2242,7 +2408,7 @@ async fn test_leadership_claim_usage_fence_rejects_old_inflight_writer() {
 }

 #[tokio::test]
-async fn test_successful_old_epoch_commit_is_fenced_after_cancellation() {
+async fn cycle_budget_lease_takeover_rejects_old_generation() {
    let store = Arc::new(MemoryConfigStore::default());
    let ctx = CancellationToken::new();
    let mut revision = DataUsageCacheRevision::Missing;
@@ -2287,12 +2453,17 @@ async fn test_successful_old_epoch_commit_is_fenced_after_cancellation() {
        .await
    );

-    let state = read_config(store, &DATA_USAGE_BLOOM_NAME_PATH)
+    let state = read_config(store.clone(), &DATA_USAGE_BLOOM_NAME_PATH)
        .await
        .expect("replacement leadership claim should persist");
    let (claimed_cycle, claimed_epoch) = decode_scanner_cycle_state(&state).expect("replacement cycle state should decode");
    assert_eq!(claimed_cycle.next, 14);
    assert_eq!(claimed_epoch, 2);
+
+    let mut stale_cycle = CurrentCycle { next: 15, ..cycle };
+    let mut stale_revision = DataUsageCacheRevision::Etag("memory-2".to_string());
+    let stale_ctx = CancellationToken::new();
+    assert!(!persist_scanner_cycle_state(&stale_ctx, store, &mut stale_cycle, &mut stale_revision, 1,).await);
 }

 #[tokio::test]
@@ -14,17 +14,16 @@

 use std::sync::{
    Arc,
-    atomic::{AtomicU8, AtomicU64, Ordering},
+    atomic::{AtomicBool, AtomicU8, AtomicU64, Ordering},
 };
-use std::time::Instant;
-
-use tokio::time::Duration;
+use tokio::time::{Duration, Instant};
 use tokio_util::sync::CancellationToken;

 const BUDGET_REASON_NONE: u8 = 0;
 const BUDGET_REASON_RUNTIME: u8 = 1;
 const BUDGET_REASON_OBJECTS: u8 = 2;
 const BUDGET_REASON_DIRECTORIES: u8 = 3;
+const PROGRESS_CLOCK_SAMPLE_INTERVAL: u64 = 128;

 #[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
 pub(crate) struct ScannerCycleBudgetConfig {
@@ -63,29 +62,51 @@ pub struct ScannerCycleBudget {
    token: CancellationToken,
    reason: Arc<AtomicU8>,
    started_at: Instant,
+    deadline: Option<Instant>,
    max_duration: Option<Duration>,
    max_objects: Option<u64>,
    max_directories: Option<u64>,
    track_progress: bool,
+    track_unbounded_counts: bool,
    objects_scanned: AtomicU64,
    directories_started: AtomicU64,
    entries_visited: AtomicU64,
+    last_progress_millis: AtomicU64,
+    cycle_state_persisted: AtomicBool,
 }

 impl ScannerCycleBudget {
+    #[cfg(test)]
    pub(crate) fn new(parent: &CancellationToken, config: ScannerCycleBudgetConfig) -> Arc<Self> {
-        Self::new_inner(parent, config, false)
+        Self::new_inner(parent, config, false, false)
    }

    pub(crate) fn new_with_progress_tracking(parent: &CancellationToken, config: ScannerCycleBudgetConfig) -> Arc<Self> {
-        Self::new_inner(parent, config, true)
+        Self::new_inner(parent, config, true, true)
    }

-    fn new_inner(parent: &CancellationToken, config: ScannerCycleBudgetConfig, track_progress: bool) -> Arc<Self> {
+    pub(crate) fn new_with_runtime_progress_tracking(parent: &CancellationToken, config: ScannerCycleBudgetConfig) -> Arc<Self> {
+        let track_progress = config.max_duration.is_some();
+        Self::new_inner(parent, config, track_progress, false)
+    }
+
+    fn new_inner(
+        parent: &CancellationToken,
+        config: ScannerCycleBudgetConfig,
+        track_progress: bool,
+        track_unbounded_counts: bool,
+    ) -> Arc<Self> {
        let token = parent.child_token();
        let reason = Arc::new(AtomicU8::new(BUDGET_REASON_NONE));
+        let started_at = Instant::now();
+        let deadline = config.max_duration.map(|duration| match started_at.checked_add(duration) {
+            Some(deadline) => deadline,
+            // Runtime config rejects this range, but keep programmatic callers
+            // fail-closed instead of panicking or silently disabling the wall clock.
+            None => started_at,
+        });

-        if let Some(duration) = config.max_duration {
+        if let Some(deadline) = deadline {
            let parent = parent.clone();
            let token_wait = token.clone();
            let token_cancel = token.clone();
@@ -94,7 +115,7 @@ impl ScannerCycleBudget {
                tokio::select! {
                    _ = parent.cancelled() => {}
                    _ = token_wait.cancelled() => {}
-                    _ = tokio::time::sleep(duration) => {
+                    _ = tokio::time::sleep_until(deadline) => {
                        Self::cancel_for_reason(&reason, &token_cancel, ScannerCycleBudgetReason::Runtime);
                    }
                }
@@ -104,14 +125,18 @@ impl ScannerCycleBudget {
        Arc::new(Self {
            token,
            reason,
-            started_at: Instant::now(),
+            started_at,
+            deadline,
            max_duration: config.max_duration,
            max_objects: config.max_objects,
            max_directories: config.max_directories,
            track_progress,
+            track_unbounded_counts,
            objects_scanned: AtomicU64::new(0),
            directories_started: AtomicU64::new(0),
            entries_visited: AtomicU64::new(0),
+            last_progress_millis: AtomicU64::new(0),
+            cycle_state_persisted: AtomicBool::new(false),
        })
    }

@@ -131,6 +156,14 @@ impl ScannerCycleBudget {
        self.max_duration
    }

+    pub(crate) fn deadline(&self) -> Option<Instant> {
+        self.deadline
+    }
+
+    pub(crate) fn cancel_for_runtime(&self) {
+        self.cancel_for(ScannerCycleBudgetReason::Runtime);
+    }
+
    pub(crate) fn max_objects(&self) -> Option<u64> {
        self.max_objects
    }
@@ -173,15 +206,43 @@ impl ScannerCycleBudget {
        self.entries_visited.load(Ordering::Relaxed)
    }

+    pub(crate) fn mark_cycle_state_persisted(&self) {
+        self.cycle_state_persisted.store(true, Ordering::Release);
+    }
+
+    pub(crate) fn cycle_state_persisted(&self) -> bool {
+        self.cycle_state_persisted.load(Ordering::Acquire)
+    }
+
+    pub(crate) fn progress_age(&self) -> Duration {
+        let elapsed_millis = u64::try_from(self.started_at.elapsed().as_millis()).unwrap_or(u64::MAX);
+        let last_progress = self.last_progress_millis.load(Ordering::Relaxed);
+        Duration::from_millis(elapsed_millis.saturating_sub(last_progress))
+    }
+
+    fn record_progress_sample(&self, event: u64) {
+        // Clock reads are sampled at batch/count boundaries; the scanner's
+        // per-object path does not add a second progress atomic.
+        if event == 0 || (event != 1 && !event.is_multiple_of(PROGRESS_CLOCK_SAMPLE_INTERVAL)) {
+            return;
+        }
+        let elapsed_millis = u64::try_from(self.started_at.elapsed().as_millis()).unwrap_or(u64::MAX);
+        self.last_progress_millis.store(elapsed_millis, Ordering::Relaxed);
+    }
+
    pub(crate) fn record_entries_visited(&self, entries_visited: u64) {
        if self.track_progress {
-            saturating_fetch_add(&self.entries_visited, entries_visited);
+            let entries = saturating_fetch_add(&self.entries_visited, entries_visited);
+            self.record_progress_sample(entries);
        }
    }

    pub(crate) fn record_remote_progress(&self, objects_scanned: u64, directories_started: u64) {
        if self.track_progress || self.max_objects.is_some() {
            let objects = saturating_fetch_add(&self.objects_scanned, objects_scanned);
+            if self.track_progress {
+                self.record_progress_sample(objects);
+            }
            if self.max_objects.is_some_and(|max_objects| objects >= max_objects) {
                self.cancel_for(ScannerCycleBudgetReason::Objects);
            }
@@ -189,9 +250,12 @@ impl ScannerCycleBudget {

        if self.track_progress || self.max_directories.is_some() {
            let directories = saturating_fetch_add(&self.directories_started, directories_started);
+            if self.track_progress {
+                self.record_progress_sample(directories);
+            }
            if self
                .max_directories
-                .is_some_and(|max_directories| directories > max_directories)
+                .is_some_and(|max_directories| directory_budget_exhausted(directories, max_directories))
            {
                self.cancel_for(ScannerCycleBudgetReason::Directories);
            }
@@ -207,14 +271,17 @@ impl ScannerCycleBudget {
    }

    pub(crate) fn try_start_directory(&self) -> bool {
-        if !self.track_progress && self.max_directories.is_none() {
+        if self.max_directories.is_none() && !self.track_unbounded_counts {
            return true;
        }

        let directories = saturating_fetch_add(&self.directories_started, 1);
+        if self.track_progress {
+            self.record_progress_sample(directories);
+        }
        if self
            .max_directories
-            .is_some_and(|max_directories| directories > max_directories)
+            .is_some_and(|max_directories| directory_budget_exhausted(directories, max_directories))
        {
            self.cancel_for(ScannerCycleBudgetReason::Directories);
            return false;
@@ -224,11 +291,14 @@ impl ScannerCycleBudget {
    }

    pub(crate) fn record_object_scanned(&self) {
-        if !self.track_progress && self.max_objects.is_none() {
+        if self.max_objects.is_none() && !self.track_unbounded_counts {
            return;
        }

        let objects = saturating_fetch_add(&self.objects_scanned, 1);
+        if self.track_progress {
+            self.record_progress_sample(objects);
+        }
        if self.max_objects.is_some_and(|max_objects| objects >= max_objects) {
            self.cancel_for(ScannerCycleBudgetReason::Objects);
        }
@@ -259,6 +329,13 @@ fn saturating_fetch_add(value: &AtomicU64, delta: u64) -> u64 {
    }
 }

+fn directory_budget_exhausted(directories: u64, max_directories: u64) -> bool {
+    // Saturation hides a remote max+1 update when the configured limit is the
+    // largest representable counter. Treat that boundary as exhausted rather
+    // than allowing work to continue indefinitely.
+    directories > max_directories || (directories == u64::MAX && max_directories == u64::MAX)
+}
+
 impl Drop for ScannerCycleBudget {
    fn drop(&mut self) {
        self.token.cancel();
@@ -401,6 +478,35 @@ mod tests {
        assert_eq!(directory_budget.reason(), Some(ScannerCycleBudgetReason::Directories));
    }

+    #[test]
+    fn directory_budget_fails_closed_when_progress_saturates() {
+        let parent = CancellationToken::new();
+        let budget = ScannerCycleBudget::new(
+            &parent,
+            ScannerCycleBudgetConfig {
+                max_directories: Some(u64::MAX),
+                ..Default::default()
+            },
+        );
+
+        budget.record_remote_progress(0, u64::MAX);
+
+        assert_eq!(budget.reason(), Some(ScannerCycleBudgetReason::Directories));
+        assert!(budget.token().is_cancelled());
+
+        let local_budget = ScannerCycleBudget::new(
+            &parent,
+            ScannerCycleBudgetConfig {
+                max_directories: Some(u64::MAX),
+                ..Default::default()
+            },
+        );
+        local_budget.record_remote_progress(0, u64::MAX - 1);
+        assert!(!local_budget.budget_elapsed());
+        assert!(!local_budget.try_start_directory());
+        assert_eq!(local_budget.reason(), Some(ScannerCycleBudgetReason::Directories));
+    }
+
    #[test]
    fn explicit_progress_tracking_counts_unbounded_remote_work_without_cancelling() {
        let parent = CancellationToken::new();
@@ -461,4 +567,29 @@ mod tests {
        assert!(object_limited.requires_serial_progress_accounting());
        assert!(directory_limited.requires_serial_progress_accounting());
    }
+
+    #[tokio::test(start_paused = true)]
+    async fn progress_age_uses_virtual_time_and_sampled_progress() {
+        let parent = CancellationToken::new();
+        let budget = ScannerCycleBudget::new_with_runtime_progress_tracking(
+            &parent,
+            ScannerCycleBudgetConfig {
+                max_duration: Some(Duration::from_secs(60)),
+                ..Default::default()
+            },
+        );
+
+        tokio::time::advance(Duration::from_secs(5)).await;
+        assert_eq!(budget.progress_age(), Duration::from_secs(5));
+        budget.record_entries_visited(1);
+        assert_eq!(budget.progress_age(), Duration::ZERO);
+
+        tokio::time::advance(Duration::from_secs(2)).await;
+        for _ in 0..126 {
+            budget.record_entries_visited(1);
+        }
+        assert_eq!(budget.progress_age(), Duration::from_secs(2));
+        budget.record_entries_visited(1);
+        assert_eq!(budget.progress_age(), Duration::ZERO);
+    }
 }
@@ -48,6 +48,7 @@ use time::OffsetDateTime;
 use tokio::sync::{Mutex, Notify, Semaphore, mpsc};
 use tokio::time::Duration;
 use tokio_util::sync::CancellationToken;
+use tokio_util::task::AbortOnDropHandle;
 use tracing::{debug, error, warn};

 use crate::ScannerObjectInfo as ObjectInfo;
@@ -314,7 +314,7 @@ impl ScannerIOCache for SetDisks {
        let ctx_clone = ctx.clone();
        let completed_bucket_count = Arc::new(AtomicUsize::new(0));
        let completed_bucket_count_clone = completed_bucket_count.clone();
-        let collect_bucket_results_fut = tokio::spawn(async move {
+        let collect_bucket_results_fut = AbortOnDropHandle::new(tokio::spawn(async move {
            let mut cancelled = false;

            loop {
@@ -333,7 +333,7 @@ impl ScannerIOCache for SetDisks {
                    }
                }
            }
-        });
+        }));

        let mut futs = Vec::new();

@@ -365,7 +365,7 @@ impl ScannerIOCache for SetDisks {
                NamespaceScannerWorkerMode::RemoteV4(server_epoch) => Some(server_epoch),
                NamespaceScannerWorkerMode::Coordinator => None,
            };
-            futs.push(tokio::spawn(async move {
+            futs.push(AbortOnDropHandle::new(tokio::spawn(async move {
                let remote_session_id = uuid::Uuid::new_v4();
                let mut remote_session_sequence = 0_u64;
                loop {
@@ -1038,7 +1038,7 @@ impl ScannerIOCache for SetDisks {
                        );
                    }
                }
-            }));
+            })));
        }
        drop(bucket_tx);
        drop(bucket_result_tx);
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
马登山	7e7a442129	style: restore rebalance formatting	2026-08-23 12:34:48 +08:00
houseme	ef1986a381	Merge branch 'main' into cxymds/fix-1933-heal-progress-model Signed-off-by: houseme <housemecn@gmail.com>	2026-08-23 12:31:09 +08:00
唐小鸭	5f72209446	fix(ecstore): keep unknown-size sentinel in create_bitrot_writer (#6380 ) SSE and compression wrap the payload so its length is unknown and advertise HashReader::SIZE_PRESERVE_LAYER (-1). Every layer preserved that sentinel except create_bitrot_writer, which clamped it to 0 before calling DiskAPI::create_file. RemoteDisk forwards that size verbatim in the put_file_stream query, so remote peers were told the body was empty. Since the authenticated put-file trailer (#5868) the receiver used the declared size to split body from trailer, turning the clamp into a fatal "auth trailer has trailing data" failure for every SSE PUT on multi-node deployments (rc.2). #6320 relaxed the receiver to only trust size > 0; this change fixes the sender so the sentinel survives end to end and the wire no longer conflates empty objects with unknown-length streams. Refs #6331	2026-08-23 12:29:52 +08:00
Zhengchao An	b6ba89d9e4	docs(testing): document CI gate matrix (#6412 )	2026-08-23 12:09:06 +08:00
houseme	648d5166e2	feat(allocator): replace mimalloc/libmimalloc-sys with rustfs-mimalloc/rustfs-mimalloc-sys (#6404 ) Replace the upstream xonatius/mimalloc_rust.git fork (mimalloc + libmimalloc-sys) with the published rustfs-mimalloc (v0.5.0) and rustfs-mimalloc-sys (v0.5.0) crates from crates.io. The new crates are based on mimalloc V3 (v3.5.0) and provide: - MiMalloc global allocator with safe API (collect, stats_json, process_info) - Heap management and arena operations (heap module) - Full FFI bindings to mimalloc V3 Changes: - Workspace deps: mimalloc + libmimalloc-sys (git) → rustfs-mimalloc + rustfs-mimalloc-sys (crates.io) - allocator_reclaim.rs: libmimalloc_sys::mi_collect → rustfs_mimalloc::MiMalloc::collect - memory_observability.rs: raw FFI mi_stats_get_json → MiMalloc::stats_json() - main.rs: heap ownership tests use Heap::contains() (V3 API) - deny.toml: remove xonatius/mimalloc_rust.git from allow-git Co-authored-by: heihutu <heihutu@gmail.com>	2026-08-23 12:07:25 +08:00
houseme	84eb5aebef	fix(ecstore): remove inline write debug noise (#6408 ) * fix(ecstore): remove inline write debug noise Co-Authored-By: heihutu <heihutu@gmail.com> * fix(ecstore): satisfy warning-as-error lints Co-Authored-By: heihutu <heihutu@gmail.com> --------- Co-authored-by: heihutu <heihutu@gmail.com>	2026-08-23 12:07:20 +08:00
马登山	9837cc7c7b	fix(heal): stabilize progress generations	2026-08-23 09:56:48 +08:00
overtrue	e63033fbc7	fix(heal): preserve progress status across nodes	2026-08-23 09:01:44 +08:00
马登山	37c5fef252	fix(heal): satisfy checkpoint outcome lint	2026-08-23 08:14:07 +08:00
马登山	276d8e156f	merge main and fix swift lint	2026-08-23 05:21:32 +08:00
cxymds	20d1266496	fix(heal): fence format repair during pool transitions (#6342 ) * fix(heal): fence format repair during pool transitions * fix(heal): fence format writes during transitions	2026-08-23 04:24:48 +08:00
唐小鸭	f7003dfddd	fix(admin): four site-replication interop correctness fixes (B5-rc T2) (#6399 ) * fix(admin): send versioningEnabled on site replication make-bucket ops The outbound make-with-versioning bucket-op query only carried operation/createdAt/lockEnabled. MinIO's own create-bucket hook sends versioningEnabled=true on this op, so align the outbound query with MinIO's site-replication make-bucket wire contract. Route both outbound builders (bootstrap plan and create-bucket hook) through one shared builder that always appends versioningEnabled=true. RustFS's own inbound handler force-enables versioning either way, so RustFS-to-RustFS behavior is unchanged; the MinIO release verified against (RELEASE.2025-09-07) also force-enables versioning regardless of the flag, so this aligns the wire contract rather than changing observable behavior there. * fix(admin): propagate purge-deleted-bucket errors in site replication The purge-deleted-bucket branch of the peer bucket-ops handler dropped the delete_bucket error and answered 200, so a peer-driven purge that failed (disk full, quorum loss) was reported as success while the bucket survived on this site. Tolerate only bucket-not-found (the purge raced an earlier replay or a local delete) and propagate every other error through ApiError like the sibling delete branches do. * fix(admin): derive fallback site deployment ID with UUIDv5 deployment_id_for_endpoint used DefaultHasher, whose algorithm is not guaranteed stable across Rust releases. The fallback fires when a peer response carries an empty deploymentID; the result is persisted in site-replication state, used for collision disambiguation, and broadcast to peers, so a toolchain bump could re-derive a different ID for the same endpoint. Note that the add preflight currently rejects that case upstream of this fallback. Derive UUIDv5 (NAMESPACE_URL) over the canonical endpoint instead, and log a structured warn when a peer metainfo response arrives without a deploymentID. Already persisted fallback IDs are non-empty and therefore never re-derived, so existing state is unaffected. * fix(admin): stream site replication devnull body without 1MB cap The site-replication devnull endpoint buffered the request body through read_plain_admin_body, which enforces the 1MB admin body cap. MinIO peers stream multi-megabyte probe bodies to this endpoint during site netperf link checks and expect an unbounded discard, so any larger probe got a 400 and was misreported as a broken link. Stream and discard the body chunk by chunk with no size cap instead, mirroring MinIO's io.Discard drain. The response stays 204 with an empty body.	2026-08-23 04:24:04 +08:00
overtrue	629e477da2	fix(heal): preserve resumable bucket checkpoints	2026-08-23 03:58:39 +08:00
overtrue	4fd3d7b643	fix(heal): make resume handoff crash safe	2026-08-23 03:19:18 +08:00
Zhengchao An	2d7120460b	test(e2e): remove fake KMS suite results (#6401 )	2026-08-23 01:45:15 +08:00
Zhengchao An	b91845c98c	ci: align cache writer and reader keys (#6398 )	2026-08-23 01:44:50 +08:00
Zhengchao An	7ba6f8cb33	test(e2e): fix cluster nightly oracles (#6397 )	2026-08-23 01:44:35 +08:00
Zhengchao An	f44b30c61a	ci(perf): fix nightly regression baseline (#6389 )	2026-08-23 01:43:52 +08:00
Zhengchao An	c6590182ed	ci(mint): pin manual image default (#6387 )	2026-08-23 01:43:27 +08:00
Zhengchao An	6d85a9c6a8	ci(s3tests): stabilize HAProxy request handling (#6386 )	2026-08-23 01:42:58 +08:00
Zhengchao An	26e6508b64	fix(ecstore): reject equal-time latest identity conflicts before index fallback (#6374 )	2026-08-23 01:42:11 +08:00
GatewayJ	51e369be6c	fix(policy): accept legacy bucket policy ID field (#6362 )	2026-08-23 01:40:44 +08:00
Zhengchao An	ddc4120c82	ci: detect incomplete and stale scheduled validations (#6357 )	2026-08-23 01:40:28 +08:00
cxymds	87235ffd28	perf(ecstore): bound decommission entry workers (#6360 )	2026-08-23 01:40:03 +08:00
cxymds	62c465ecef	fix(heal): supervise scheduler task panics (#6351 )	2026-08-23 01:39:06 +08:00
houseme	f1b92af4a3	feat(ecstore): coalesce GET ReadVersion RPCs (#6395 )	2026-08-23 01:15:38 +08:00
Zhengchao An	7c1a76dfd9	fix(ecstore): fence deletes against decommission commits (#6363 ) * fix(ecstore): fence deletes against decommission commits * fix(ecstore): preserve decommission target write locks * fix(ecstore): preserve delete markers during source cleanup * fix(ecstore): route batch delete markers to active pools * fix(ecstore): preserve batch delete pool errors * fix(ecstore): retain source-set lock during cleanup * test(ecstore): exercise decommission delete fences * test(ecstore): finish decommission delete fence scenario * fix(ecstore): reuse fixed fence for reverse decommission * fix(ecstore): fence decommission commit loss * fix(ecstore): annotate batch delete fallback * test(ecstore): fix decommission fence fixtures * fix(ecstore): unblock decommission delete fences * fix(ecstore): preserve distributed decommission set locks * fix(ecstore): match decommission lock backend domain * test(ecstore): align decommission fence barriers * fix(ecstore): satisfy delete fence lint checks * fix(rebalance): preserve access-denied delete errors	2026-08-23 00:40:38 +08:00
Zhengchao An	3ddf1a81ac	test(kms): replace 33 hard-coded startup sleeps with readiness probe (#6349 ) * refactor(e2e/kms): replace fixed startup sleeps with KMS readiness probe Replace 33 hard-coded sleep(3s) / sleep(2s) startup waits in KMS e2e tests with an active readiness probe (wait_for_kms_ready) that polls the KMS status endpoint with exponential backoff (200ms→1s, 5s budget). This cuts per-test startup latency from a fixed 3s to ~200-500ms while remaining robust against slow CI machines. Non-startup sleeps (ILM polling loops, fault-recovery detection delays, test-runner inter-test pauses) are left untouched. * style: cargo fmt * fix(kms): use .expect() instead of ? in test functions that return () 7 call sites of wait_for_kms_ready() used ? in async test functions that return () instead of Result. Changed to .expect("KMS ready"). * fix(kms): enforce readiness probe deadline * fix(kms): validate readiness backend status	2026-08-22 16:35:57 +00:00
马登山	79a8371479	fix(heal): preserve terminal progress counters	2026-08-23 00:25:27 +08:00
Zhengchao An	cc412914d5	feat(connect): emit durable heartbeats (#6383 )	2026-08-22 16:03:23 +00:00
cxymds	2fccfdeabe	fix(scanner): fence timed out scan cycles (#6352 ) * fix(scanner): fence timed out scan cycles * fix(scanner): cancel scan workers with cycle scope * fix(scanner): reject persisted timer overflow * fix(scanner): reject terminal leadership epochs * fix(scanner): reject trailing cycle state bytes	2026-08-22 14:38:18 +00:00
Zhengchao An	9815694301	fix(ecstore): supervise decommission worker cleanup (#6372 )	2026-08-22 14:33:24 +00:00
houseme	0e79106c2f	fix(storageclass): use div_ceil for inline threshold to match shard size calc (#6390 ) The inline_block threshold used floor division (DEFAULT_INLINE_OBJECT_BUDGET / data_shards) while shard_file_size uses ceiling division (div_ceil). For EC 12:4 with 256KiB objects, this caused a 1-byte discrepancy: - inline_block = 262144 / 12 = 21845 (floor) - shard_file_size = 262144.div_ceil(12) = 21846 (ceil) - should_inline(21846, 12, false) = false (wrong!) Fix by using div_ceil for the inline_block calculation, so both sides use the same rounding and the inline path is correctly triggered. Co-authored-by: heihutu <heihutu@gmail.com>	2026-08-22 22:05:20 +08:00
Zhengchao An	12a9e654b5	refactor(data-usage): rename ReplicationStats to ReplicationTargetUsage (#6345 ) * refactor(data-usage): ReplicationStats -> ReplicationTargetUsage Rename the data-usage crate's ReplicationStats to ReplicationTargetUsage. Serde field names are byte-identical (only the Rust type name changed; field identifiers that rmp encodes are untouched). An rmp round-trip test guards against future drift. Scanner test imports updated to match. * style: cargo fmt	2026-08-22 13:56:58 +00:00
Zhengchao An	6b5e0feef6	test(e2e): wait for heal peers after node rejoin (#6359 )	2026-08-22 13:52:08 +00:00
马登山	69ae9ef671	fix(heal): atomically persist page progress	2026-08-22 20:53:55 +08:00
马登山	db709aee57	Merge remote-tracking branch 'origin/main' into cxymds/fix-1933-heal-progress-model	2026-08-22 20:50:34 +08:00
马登山	16c56fe0bd	fix(heal): correct progress accounting	2026-08-22 17:58:55 +08:00