fix(heal): preserve decode repair signals and retry startup quorum (#8067 )

* fix(startup): retry transient bucket metadata quorum failures * fix(ecstore): retain recovered shard damage for read repair * fix(ecstore): reconnect missing read sources before GET * test(heal): prove exact marker replay preserves history * fix(ecstore): fence reconnect retries at the deadline Reject expired reconnect waiters before dispatch and start cooldown on sweep completion, including timeout and cancellation. Cover exact-deadline admission independently of cooldown and preserve the original strict RPC-count regression.
fix(ci): publish version-tagged preview Docker images (#8068 )
2026-10-30 15:55:55 +00:00 · 2026-09-22 14:58:31 +00:00 · 2026-09-22 13:06:23 +00:00 · 2026-09-22 12:16:40 +00:00 · 2026-09-22 12:15:20 +00:00 · 2026-09-22 19:23:41 +08:00
142 changed files with 11412 additions and 1066 deletions
@@ -101,6 +101,7 @@ Update this file only when an advisory adds or changes a reusable lesson, affect
 ### SSE and on-disk storage invariants

 - `GHSA-xrrf-67jm-3c2r`: SSE metadata reported encryption while reader composition bypassed `EncryptReader` and stored plaintext. Lesson: test actual bytes on disk and wrapper order, not only API metadata.
+- `GHSA-wqmc-vjgv-jrpw`: SSE-C persisted and listed the plaintext MD5 as the ETag, allowing list-only users to confirm or brute-force low-entropy content and correlate equal plaintext across keys. Lesson: encrypted-object ETags must not expose deterministic plaintext fingerprints; protect them with per-object key material and keep listing, conditional, copy, and multipart semantics consistent.

 ### Object Lock and retention invariants

@@ -126,7 +127,7 @@ rg -n "normalize_extract_entry_key|Snowball|auto-extract|PathBuf::join|canonical
 rg -n "DEFAULT_SECRET|DEFAULT_ACCESS|TEST_PRIVATE_KEY|rustfs rpc|RUSTFS_RPC_SECRET" rustfs crates
 rg -n "TONIC_RPC_PREFIX|verify_rpc_signature|check_auth|NodeServiceServer|x-rustfs-signature" rustfs crates
 rg -n "debug!|trace!|info!|error!|\\?resp|\\?merged_config|session_token|secret_key" rustfs crates
-rg -n "HashReader|EncryptReader|SSE|server-side encryption|Access-Control-Allow-Credentials|Origin" rustfs crates
+rg -n "HashReader|EncryptReader|try_resolve_etag|ETag|SSE|server-side encryption|Access-Control-Allow-Credentials|Origin" rustfs crates
 rg -n "ObjectLock|object_lock|retention|COMPLIANCE|GOVERNANCE|delete_prefix|lifecycle|scanner" rustfs crates
 rg -n "deny_unknown_fields|serde.default|as u32|as usize|as i32" rustfs crates
 ```
@@ -145,5 +146,5 @@ rg -n "deny_unknown_fields|serde.default|as u32|as usize|as i32" rustfs crates
 - IAM export fixes: assert exported archives omit plaintext user and service-account secrets unless the format deliberately encrypts or seals them.
 - RPC auth fixes: include captured metadata replay across two concrete methods, stale timestamps, wrong path, wrong method surrogate, wrong secret, and valid same-method calls.
 - Browser/CORS fixes: assert no credentials on reflected/default origins, correct behavior for explicit allowlists, and no same-origin script execution for previewed object content.
- SSE fixes: inspect stored bytes and verify API metadata, read-back behavior, and on-disk ciphertext together.
+- SSE fixes: inspect stored bytes and verify API metadata, read-back behavior, and on-disk ciphertext together; for ETags, upload equal plaintext under different keys and assert list-only callers cannot derive or correlate the plaintext digest, including multipart objects.
 - Object Lock fixes: include unreadable metadata, fabricated metadata defaults, unparsable config, confirmed absent config, COMPLIANCE/GOVERNANCE retention, lifecycle expiry, scanner sweeps, and force-delete paths.
@@ -29,6 +29,7 @@ script-tests: ## Run shell script tests
 	./scripts/test_entrypoint_credentials.sh
 	./scripts/test_internode_grpc_ab_bench.sh
 	./scripts/test_object_batch_bench_enhanced.sh
+	./scripts/test_package_service_scripts.sh
 	./scripts/test_hotpath_warp_ab_gate.sh
 	./scripts/test_hotpath_warp_abba.sh
 	./scripts/test_scanner_validation_harness.sh
@@ -24,8 +24,11 @@ on:
      - '.github/actions/**'
      - '.github/workflows/**'
      - 'scripts/release/create_or_update_release.sh'
+      - 'scripts/release/package_service_scripts.sh'
      - 'scripts/release/package_versions.sh'
+      - 'scripts/test_package_service_scripts.sh'
      - 'scripts/test_package_versions.sh'
+      - 'scripts/test_docker_workflow.py'
      - 'scripts/security/check_performance_ab_workflow.sh'
      - 'scripts/security/check_preview_release_workflow.sh'
      - 'scripts/security/check_tier_artifact_workflow.sh'
@@ -40,8 +43,11 @@ on:
      - '.github/actions/**'
      - '.github/workflows/**'
      - 'scripts/release/create_or_update_release.sh'
+      - 'scripts/release/package_service_scripts.sh'
      - 'scripts/release/package_versions.sh'
+      - 'scripts/test_package_service_scripts.sh'
      - 'scripts/test_package_versions.sh'
+      - 'scripts/test_docker_workflow.py'
      - 'scripts/security/check_performance_ab_workflow.sh'
      - 'scripts/security/check_preview_release_workflow.sh'
      - 'scripts/security/check_tier_artifact_workflow.sh'
@@ -158,6 +164,9 @@ jobs:
      - name: Check package version contract
        run: ./scripts/test_package_versions.sh

+      - name: Check package service lifecycle contract
+        run: ./scripts/test_package_service_scripts.sh
+
  dependency-review:
    name: Dependency Review
    runs-on: ubuntu-latest
@@ -818,9 +818,7 @@ jobs:

          echo ""
          echo "🐳 Docker Images:"
-          if [[ "$BUILD_TYPE" == "preview" ]]; then
-            echo "⏭️  Preview tags do not publish Docker images"
-          elif [[ "$INPUT_BUILD_DOCKER" == "false" ]]; then
+          if [[ "$INPUT_BUILD_DOCKER" == "false" ]]; then
            echo "⏭️  Docker image build was skipped (binary only build)"
          elif [[ "$BUILD_STATUS" == "success" ]]; then
            echo "🔄 Docker images will be built and pushed automatically via workflow_run event"
@@ -47,7 +47,7 @@ on:
        default: true
        type: boolean
      version:
-        description: "Version to build (latest for stable release, or specific version like v1.0.0, v1.0.0-alpha1)"
+        description: "Version to build (latest, v1.0.0, v1.0.0-alpha1, or v1.0.1-preview.1)"
        required: false
        default: "latest"
        type: string
@@ -82,8 +82,7 @@ jobs:
      github.event_name == 'workflow_dispatch' ||
      (github.event.workflow_run.conclusion == 'success' &&
       github.event.workflow_run.event == 'push' &&
-       github.event.workflow_run.head_branch != 'main' &&
-       !contains(github.event.workflow_run.head_branch, '-preview'))
+       github.event.workflow_run.head_branch != 'main')
    runs-on: ubuntu-latest
    timeout-minutes: 30
    outputs:
@@ -197,7 +196,14 @@ jobs:
                  version="${version#v}"
                fi

-                if [[ "$version" == *"alpha"* ]] || [[ "$version" == *"beta"* ]] || [[ "$version" == *"rc"* ]]; then
+                if [[ "$version" =~ -preview\.[0-9]+$ ]]; then
+                  build_type="preview"
+                  is_prerelease=true
+                  echo "🔍 Building Docker image for preview: $version (version tags only)"
+                elif [[ "$version" == *"-preview"* ]]; then
+                  echo "❌ Invalid preview tag: $version (expected suffix: -preview.<number>)" >&2
+                  exit 1
+                elif [[ "$version" == *"alpha"* ]] || [[ "$version" == *"beta"* ]] || [[ "$version" == *"rc"* ]]; then
                  build_type="prerelease"
                  is_prerelease=true
                  # Pre-GA policy: prereleases update latest until the first stable tag exists.
@@ -249,11 +255,14 @@ jobs:
                echo "🚀 Building with latest stable release version"
                ;;
              *-preview*)
+                if [[ ! "$input_version" =~ ^v?[0-9]+\.[0-9]+\.[0-9]+(-[0-9A-Za-z.-]+)?-preview\.[0-9]+$ ]]; then
+                  echo "❌ Invalid preview version: $input_version" >&2
+                  exit 1
+                fi
                build_type="preview"
                is_prerelease=true
-                should_build=false
-                should_push=false
-                echo "⏭️  Preview tags do not publish Docker images"
+                version="${input_version#v}"
+                echo "🔍 Building with preview version: $version (version tags only)"
                ;;
              # Prerelease versions (must match first, more specific)
              v*alpha*|v*beta*|v*rc*|*alpha*|*beta*|*rc*)
@@ -542,6 +551,54 @@ jobs:
          if-no-files-found: error
          retention-days: 30

+  # Docker Hub's overview is a separate `full_description` field that
+  # `docker push` never touches, so it drifted from README.md and kept
+  # linking to pages that no longer exist (#7995). Republish it from the
+  # commit the images were built from, only when images were pushed.
+  sync-dockerhub-description:
+    name: Sync Docker Hub Description
+    needs: [ build-check, build-docker ]
+    if: needs.build-check.outputs.should_build == 'true' && needs.build-check.outputs.should_push == 'true'
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+          ref: ${{ needs.build-check.outputs.source_ref }}
+          sparse-checkout: README.md
+          sparse-checkout-cone-mode: false
+
+      - name: Sync Docker Hub description
+        uses: peter-evans/dockerhub-description@1b9a80c056b620d92cedb9d9b5a223409c68ddfa # v5.0.0
+        with:
+          username: ${{ env.DOCKERHUB_USERNAME }}
+          password: ${{ secrets.DOCKERHUB_TOKEN }}
+          repository: ${{ env.REGISTRY_DOCKERHUB }}
+          readme-filepath: ./README.md
+          # README.md links to docs/ and CONTRIBUTING.md by relative path;
+          # rewrite those to github.com URLs so they resolve on Docker Hub.
+          enable-url-completion: true
+
+      # Docker Hub caps `full_description` at 25,000 bytes and the action
+      # truncates to fit with only a warning, which would publish a README
+      # cut off mid-sentence. README.md is already near the cap, so read the
+      # published overview back and fail if it hit the limit.
+      - name: Check published description was not truncated
+        env:
+          REPOSITORY: ${{ env.REGISTRY_DOCKERHUB }}
+        run: |
+          limit=25000
+          published=$(curl -sSf "https://hub.docker.com/v2/repositories/${REPOSITORY}/" \
+            | jq -j '.full_description' | wc -c | tr -d ' ')
+          echo "Published overview: ${published} bytes (Docker Hub limit: ${limit})"
+          # Truncation lands within one UTF-8 character of the cap.
+          if [ "$published" -ge $((limit - 4)) ]; then
+            echo "::error file=README.md::Docker Hub overview was truncated at ${limit} bytes; trim README.md before the next release."
+            exit 1
+          fi
+
  # Docker build summary
  docker-summary:
    name: Docker Build Summary
@@ -563,6 +620,10 @@ jobs:
          echo ""

          case "$BUILD_TYPE" in
+            "preview")
+              echo "🔍 Preview Docker image has been built with ${VERSION} tags"
+              echo "⏭️  Preview images do not update latest or prerelease channels"
+              ;;
            "release")
              echo "🚀 Release Docker image has been built with ${VERSION} tags"
              echo "✅ This image is ready for production use"
@@ -117,10 +117,15 @@ jobs:

          cat > "${PKG_DIR}/etc/default/rustfs" << 'ENVEOF'
          # RustFS Environment Configuration
-          # See https://rustfs.com/docs/ for more information
+          # See https://docs.rustfs.com for more information
+          # RUSTFS_ACCESS_KEY=<changeme>
+          # RUSTFS_SECRET_KEY=<changeme>
          # RUSTFS_VOLUMES=""
-          # RUSTFS_ROOT_USER=""
-          # RUSTFS_ROOT_PASSWORD=""
+          # RUSTFS_ADDRESS=":9000"
+          # RUSTFS_CONSOLE_ADDRESS=":9001"
+          # RUSTFS_CONSOLE_ENABLE=true
+          # RUSTFS_OBS_LOGGER_LEVEL=error
+          # RUSTFS_OBS_LOG_DIRECTORY="/var/log/rustfs/"
          ENVEOF

          # dpkg versions must start with a digit and cannot contain hyphens;
@@ -146,37 +151,13 @@ jobs:
          /etc/default/rustfs
          CONFFILES

-          cat > "${PKG_DIR}/DEBIAN/postinst" << 'POSTINST'
-          #!/bin/bash
-          set -e
-          if ! getent passwd rustfs > /dev/null 2>&1; then
-              useradd -r -s /bin/false -d /opt/rustfs rustfs
-          fi
-          mkdir -p /opt/rustfs /data/rustfs /var/log/rustfs
-          chown rustfs:rustfs /opt/rustfs /data/rustfs /var/log/rustfs
-          if [ -d /run/systemd/system ]; then
-              systemctl daemon-reload
-          fi
-          echo "RustFS installed. Configure /etc/default/rustfs then: systemctl start rustfs"
-          POSTINST
+          ./scripts/release/package_service_scripts.sh after-install > "${PKG_DIR}/DEBIAN/postinst"
          chmod 755 "${PKG_DIR}/DEBIAN/postinst"

-          cat > "${PKG_DIR}/DEBIAN/prerm" << 'PRERM'
-          #!/bin/bash
-          set -e
-          if [ -d /run/systemd/system ] && systemctl is-active --quiet rustfs; then
-              systemctl stop rustfs
-          fi
-          PRERM
+          ./scripts/release/package_service_scripts.sh before-remove > "${PKG_DIR}/DEBIAN/prerm"
          chmod 755 "${PKG_DIR}/DEBIAN/prerm"

-          cat > "${PKG_DIR}/DEBIAN/postrm" << 'POSTRM'
-          #!/bin/bash
-          set -e
-          if [ -d /run/systemd/system ]; then
-              systemctl daemon-reload
-          fi
-          POSTRM
+          ./scripts/release/package_service_scripts.sh after-remove > "${PKG_DIR}/DEBIAN/postrm"
          chmod 755 "${PKG_DIR}/DEBIAN/postrm"

          cp LICENSE "${PKG_DIR}/usr/share/doc/rustfs/"
@@ -214,10 +195,15 @@ jobs:
          mkdir -p ./tmp-pkg/etc/default
          cat > ./tmp-pkg/etc/default/rustfs << 'ENVEOF'
          # RustFS Environment Configuration
-          # See https://rustfs.com/docs/ for more information
+          # See https://docs.rustfs.com for more information
+          # RUSTFS_ACCESS_KEY=<changeme>
+          # RUSTFS_SECRET_KEY=<changeme>
          # RUSTFS_VOLUMES=""
-          # RUSTFS_ROOT_USER=""
-          # RUSTFS_ROOT_PASSWORD=""
+          # RUSTFS_ADDRESS=":9000"
+          # RUSTFS_CONSOLE_ADDRESS=":9001"
+          # RUSTFS_CONSOLE_ENABLE=true
+          # RUSTFS_OBS_LOGGER_LEVEL=error
+          # RUSTFS_OBS_LOG_DIRECTORY="/var/log/rustfs/"
          ENVEOF

          fpm -s dir -t rpm \
@@ -231,35 +217,13 @@ jobs:
            --description "High-performance distributed object storage" \
            --url "https://rustfs.com" \
            --license "Apache-2.0" \
-            --after-install <(cat << 'POSTINST'
-          #!/bin/bash
-          set -e
-          if ! getent passwd rustfs > /dev/null 2>&1; then
-              useradd -r -s /bin/false -d /opt/rustfs rustfs
-          fi
-          mkdir -p /opt/rustfs /data/rustfs /var/log/rustfs
-          chown rustfs:rustfs /opt/rustfs /data/rustfs /var/log/rustfs
-          if [ -d /run/systemd/system ]; then
-              systemctl daemon-reload
-          fi
-          POSTINST
-            ) \
-            --before-remove <(cat << 'PRERM'
-          #!/bin/bash
-          set -e
-          if [ -d /run/systemd/system ] && systemctl is-active --quiet rustfs; then
-              systemctl stop rustfs
-          fi
-          PRERM
-            ) \
-            --after-remove <(cat << 'POSTRM'
-          #!/bin/bash
-          set -e
-          if [ -d /run/systemd/system ]; then
-              systemctl daemon-reload
-          fi
-          POSTRM
-            ) \
+            --before-install <(./scripts/release/package_service_scripts.sh rpm-before-install) \
+            --before-upgrade <(./scripts/release/package_service_scripts.sh rpm-before-upgrade) \
+            --after-install <(./scripts/release/package_service_scripts.sh after-install) \
+            --after-upgrade <(./scripts/release/package_service_scripts.sh after-install) \
+            --before-remove <(./scripts/release/package_service_scripts.sh before-remove) \
+            --after-remove <(./scripts/release/package_service_scripts.sh after-remove) \
+            --rpm-posttrans <(./scripts/release/package_service_scripts.sh rpm-posttrans) \
            --config-files /etc/default/rustfs \
            "rustfs-nightly-${DEB_DATE}/usr/bin/rustfs=/usr/bin/rustfs" \
            ./tmp-pkg/etc/default/rustfs=/etc/default/rustfs \
@@ -328,10 +328,15 @@ jobs:

          cat > "${PKG_DIR}/etc/default/rustfs" << 'ENVEOF'
          # RustFS Environment Configuration
-          # See https://rustfs.com/docs/ for more information
+          # See https://docs.rustfs.com for more information
+          # RUSTFS_ACCESS_KEY=<changeme>
+          # RUSTFS_SECRET_KEY=<changeme>
          # RUSTFS_VOLUMES=""
-          # RUSTFS_ROOT_USER=""
-          # RUSTFS_ROOT_PASSWORD=""
+          # RUSTFS_ADDRESS=":9000"
+          # RUSTFS_CONSOLE_ADDRESS=":9001"
+          # RUSTFS_CONSOLE_ENABLE=true
+          # RUSTFS_OBS_LOGGER_LEVEL=error
+          # RUSTFS_OBS_LOG_DIRECTORY="/var/log/rustfs/"
          ENVEOF

          cat > "${PKG_DIR}/DEBIAN/control" << EOF
@@ -354,37 +359,13 @@ jobs:
          /etc/default/rustfs
          CONFFILES

-          cat > "${PKG_DIR}/DEBIAN/postinst" << 'POSTINST'
-          #!/bin/bash
-          set -e
-          if ! getent passwd rustfs > /dev/null 2>&1; then
-              useradd -r -s /bin/false -d /opt/rustfs rustfs
-          fi
-          mkdir -p /opt/rustfs /data/rustfs /var/log/rustfs
-          chown rustfs:rustfs /opt/rustfs /data/rustfs /var/log/rustfs
-          if [ -d /run/systemd/system ]; then
-              systemctl daemon-reload
-          fi
-          echo "RustFS installed. Configure /etc/default/rustfs then: systemctl start rustfs"
-          POSTINST
+          ./scripts/release/package_service_scripts.sh after-install > "${PKG_DIR}/DEBIAN/postinst"
          chmod 755 "${PKG_DIR}/DEBIAN/postinst"

-          cat > "${PKG_DIR}/DEBIAN/prerm" << 'PRERM'
-          #!/bin/bash
-          set -e
-          if [ -d /run/systemd/system ] && systemctl is-active --quiet rustfs; then
-              systemctl stop rustfs
-          fi
-          PRERM
+          ./scripts/release/package_service_scripts.sh before-remove > "${PKG_DIR}/DEBIAN/prerm"
          chmod 755 "${PKG_DIR}/DEBIAN/prerm"

-          cat > "${PKG_DIR}/DEBIAN/postrm" << 'POSTRM'
-          #!/bin/bash
-          set -e
-          if [ -d /run/systemd/system ]; then
-              systemctl daemon-reload
-          fi
-          POSTRM
+          ./scripts/release/package_service_scripts.sh after-remove > "${PKG_DIR}/DEBIAN/postrm"
          chmod 755 "${PKG_DIR}/DEBIAN/postrm"

          cp LICENSE "${PKG_DIR}/usr/share/doc/rustfs/"
@@ -422,10 +403,15 @@ jobs:
          mkdir -p ./tmp-pkg/etc/default
          cat > ./tmp-pkg/etc/default/rustfs << 'ENVEOF'
          # RustFS Environment Configuration
-          # See https://rustfs.com/docs/ for more information
+          # See https://docs.rustfs.com for more information
+          # RUSTFS_ACCESS_KEY=<changeme>
+          # RUSTFS_SECRET_KEY=<changeme>
          # RUSTFS_VOLUMES=""
-          # RUSTFS_ROOT_USER=""
-          # RUSTFS_ROOT_PASSWORD=""
+          # RUSTFS_ADDRESS=":9000"
+          # RUSTFS_CONSOLE_ADDRESS=":9001"
+          # RUSTFS_CONSOLE_ENABLE=true
+          # RUSTFS_OBS_LOGGER_LEVEL=error
+          # RUSTFS_OBS_LOG_DIRECTORY="/var/log/rustfs/"
          ENVEOF

          fpm -s dir -t rpm \
@@ -439,35 +425,13 @@ jobs:
            --description "High-performance distributed object storage" \
            --url "https://rustfs.com" \
            --license "Apache-2.0" \
-            --after-install <(cat <<'POSTINST'
-          #!/bin/bash
-          set -e
-          if ! getent passwd rustfs > /dev/null 2>&1; then
-              useradd -r -s /bin/false -d /opt/rustfs rustfs
-          fi
-          mkdir -p /opt/rustfs /data/rustfs /var/log/rustfs
-          chown rustfs:rustfs /opt/rustfs /data/rustfs /var/log/rustfs
-          if [ -d /run/systemd/system ]; then
-              systemctl daemon-reload
-          fi
-          POSTINST
-            ) \
-            --before-remove <(cat <<'PRERM'
-          #!/bin/bash
-          set -e
-          if [ -d /run/systemd/system ] && systemctl is-active --quiet rustfs; then
-              systemctl stop rustfs
-          fi
-          PRERM
-            ) \
-            --after-remove <(cat <<'POSTRM'
-          #!/bin/bash
-          set -e
-          if [ -d /run/systemd/system ]; then
-              systemctl daemon-reload
-          fi
-          POSTRM
-            ) \
+            --before-install <(./scripts/release/package_service_scripts.sh rpm-before-install) \
+            --before-upgrade <(./scripts/release/package_service_scripts.sh rpm-before-upgrade) \
+            --after-install <(./scripts/release/package_service_scripts.sh after-install) \
+            --after-upgrade <(./scripts/release/package_service_scripts.sh after-install) \
+            --before-remove <(./scripts/release/package_service_scripts.sh before-remove) \
+            --after-remove <(./scripts/release/package_service_scripts.sh after-remove) \
+            --rpm-posttrans <(./scripts/release/package_service_scripts.sh rpm-posttrans) \
            --config-files /etc/default/rustfs \
            ./bin/rustfs=/usr/bin/rustfs \
            ./tmp-pkg/etc/default/rustfs=/etc/default/rustfs \
@@ -0,0 +1,221 @@
+# RustFS Fault-Tolerance (degradation) Matrix Sweep
+#
+# Exhaustive topology x EC outage sweep, complementary to the scenario suite
+# (rustfs-fault-tolerance-test.yml). For each topology (nodes x drives per
+# node) and each EC level P in 1..N*drives/2 the runner installs that exact
+# cluster shape and drives progressive node outages plus recovery:
+#
+#   M-<topo>EC<P>-full   full-capacity reads/writes must succeed
+#   M-<topo>EC<P>-n<k>   with k nodes online: reads succeed iff the online
+#                        drives cover the read quorum (N-P), writes iff they
+#                        cover the write quorum (N-P+1); probes below quorum
+#                        assert the expected outage
+#   M-<topo>EC<P>-rec    after restoring every node, reads/writes succeed
+#
+# 10 topologies x EC levels = 38 combinations, 320 cases. The full sweep runs
+# 3-5 hours and is dispatched on demand (e.g. after a nightly chain goes
+# green), not nightly. Case ids and expectations come from product source
+# (default_parity_count, erasure set sizing) via rustfs/auto-testing.
+
+name: RustFS Fault-Tolerance Matrix
+
+on:
+  workflow_dispatch:
+    inputs:
+      package_url:
+        description: 'Direct .deb URL. Required unless the nightly default is wanted.'
+        required: false
+        type: string
+      auto_testing_ref:
+        description: 'auto-testing branch, tag, or SHA (defaults to main)'
+        required: false
+        default: main
+        type: string
+      matrix_topo:
+        description: 'Topology filter like 4x2; empty sweeps all ten topologies'
+        required: false
+        default: ''
+        type: string
+      matrix_ec:
+        description: 'EC/parity filter; empty sweeps every EC level'
+        required: false
+        default: ''
+        type: string
+      cleanup_before:
+        description: 'Reset the nodes before the test (DESTROYS existing data/config)'
+        type: boolean
+        default: true
+      cleanup_after:
+        description: 'Reset the nodes after the test (DESTROYS test data/config)'
+        type: boolean
+        default: true
+
+permissions:
+  contents: read
+
+# The suite stops services and revokes drive access on the shared fleet; only one
+# functional suite may touch the environment at a time.
+concurrency:
+  group: rustfs-shared-functional-tests-v2
+  cancel-in-progress: false
+
+defaults:
+  run:
+    shell: bash
+
+env:
+  RUSTFS_ACCESS_KEY: ${{ secrets.RUSTFS_ACCESS_KEY }}
+  RUSTFS_SECRET_KEY: ${{ secrets.RUSTFS_SECRET_KEY }}
+  RUSTFS_NODES: ${{ secrets.RUSTFS_NODES || vars.RUSTFS_NODES }}
+  RUSTFS_SSH_USER: ${{ secrets.RUSTFS_SSH_USER || vars.RUSTFS_SSH_USER }}
+  PF_TESTING_GH_TOKEN: ${{ secrets.PF_TESTING_GH_TOKEN }}
+  RUSTFS_NIGHTLY_PACKAGE_URL: ${{ vars.RUSTFS_NIGHTLY_PACKAGE_URL || 'https://dl.rustfs.com/artifacts/rustfs/packages/nightly/rustfs-nightly-latest.deb' }}
+
+jobs:
+  fault-tolerance-matrix:
+    runs-on: smoke-testing
+    # The full 38-combination sweep takes 3-5h; 360 is the runner job ceiling.
+    timeout-minutes: 360
+    steps:
+      - name: Initialize functional evidence
+        id: evidence
+        run: |
+          set -euo pipefail
+          umask 077
+          FUNCTIONAL_ARTIFACTS_DIR="${RUNNER_TEMP}/rustfs-fault-tolerance-matrix-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}"
+          mkdir -- "${FUNCTIONAL_ARTIFACTS_DIR}" "${FUNCTIONAL_ARTIFACTS_DIR}-scratch" "${FUNCTIONAL_ARTIFACTS_DIR}/evidence"
+          {
+            printf 'FUNCTIONAL_ARTIFACTS_DIR=%s\n' "${FUNCTIONAL_ARTIFACTS_DIR}"
+            printf 'LOG_FILE=%s/suite.log\n' "${FUNCTIONAL_ARTIFACTS_DIR}"
+            printf 'REPORT_FILE=%s/report.md\n' "${FUNCTIONAL_ARTIFACTS_DIR}"
+            printf 'RESULTS_FILE=%s/results.json\n' "${FUNCTIONAL_ARTIFACTS_DIR}"
+            printf 'EVIDENCE_DIR=%s/evidence\n' "${FUNCTIONAL_ARTIFACTS_DIR}"
+            printf 'TMPDIR=%s-scratch\n' "${FUNCTIONAL_ARTIFACTS_DIR}"
+          } >> "${GITHUB_ENV}"
+
+      - name: Checkout auto-testing scripts
+        timeout-minutes: 5
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          repository: rustfs/auto-testing
+          ref: ${{ inputs.auto_testing_ref }}
+          token: ${{ secrets.PF_TESTING_GH_TOKEN }}
+          path: auto-testing
+          persist-credentials: false
+
+      - name: Record auto-testing revision
+        run: |
+          set -euo pipefail
+          AUTO_TESTING_SHA="$(git -C auto-testing rev-parse HEAD)"
+          printf 'AUTO_TESTING_SHA=%s\n' "${AUTO_TESTING_SHA}" >> "${GITHUB_ENV}"
+          echo "auto-testing revision: ${AUTO_TESTING_SHA}"
+
+      - name: Show environment
+        run: |
+          uname -a
+          jq --version
+          aws --version
+          echo "auto-testing: ${AUTO_TESTING_SHA}"
+          df -h /data | tail -1
+
+      - name: Cleanup environment (before)
+        timeout-minutes: 5
+        if: ${{ inputs.cleanup_before != 'false' }}
+        run: |
+          ./auto-testing/rustfs-fault-tolerance-test.sh --cleanup -y --log-file "${LOG_FILE}"
+
+      - name: Run matrix sweep (topology x EC outage combinations)
+        timeout-minutes: 350
+        id: test
+        # No continue-on-error: an UNEXPECTED verdict (a probe that should
+        # fail by quorum math but passes, or vice versa) is the finding this
+        # sweep exists to surface, and the red run is the signal until matrix
+        # runs get backlog integration like the scenario suite has.
+        env:
+          MATRIX_EC: ${{ inputs.matrix_ec }}
+        run: |
+          set -euo pipefail
+          ARGS=(--matrix -y --package-url "${{ inputs.package_url || env.RUSTFS_NIGHTLY_PACKAGE_URL }}" --log-file "${LOG_FILE}" --results-file "${RESULTS_FILE}")
+          if [ -n "${{ inputs.matrix_topo }}" ]; then
+            # Trailing value form: consumed as `--matrix-topo <topo> [ec]`.
+            ARGS+=(--matrix-topo "${{ inputs.matrix_topo }}")
+            if [ -n "${MATRIX_EC}" ]; then
+              ARGS+=("${MATRIX_EC}")
+            fi
+          fi
+          ./auto-testing/rustfs-fault-tolerance-test.sh "${ARGS[@]}"
+
+      - name: Generate report
+        timeout-minutes: 2
+        if: ${{ always() && steps.evidence.outcome == 'success' }}
+        run: |
+          set -euo pipefail
+          {
+            echo "# RustFS fault-tolerance matrix report"
+            echo ""
+            echo "- Run: ${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/actions/runs/${GITHUB_RUN_ID}"
+            echo "- Package: ${{ inputs.package_url || 'nightly (R2 latest)' }}"
+            echo "- auto-testing: \`${AUTO_TESTING_SHA}\`"
+            echo "- Topology filter: ${{ inputs.matrix_topo || 'all (10 topologies)' }}"
+            echo "- EC filter: ${{ inputs.matrix_ec || 'all levels' }}"
+            echo ""
+            echo "## Per-probe results"
+            echo ""
+            echo '```'
+            grep -E '^FT-(CASE|COMPLETE|SUMMARY|REPORT)' "${LOG_FILE}" || echo "(no FT-CASE lines found)"
+            echo '```'
+            echo ""
+            echo "## Health snapshots"
+            echo ""
+            for f in "${FUNCTIONAL_ARTIFACTS_DIR}"/evidence/*.code; do
+              [ -e "${f}" ] || continue
+              printf '%s -> %s\n' "$(basename "${f}" .code)" "$(cat "${f}")"
+            done
+          } > "${REPORT_FILE}"
+          cat "${REPORT_FILE}" >> "${GITHUB_STEP_SUMMARY}"
+
+          # An unfiltered sweep must land exactly 320 cases (38 topology x EC
+          # combinations at 2*(nodes+1) cases each); filtered sweeps are held
+          # to internal consistency. Missing/harness-error rows always fail.
+          EXPECTED_TOTAL=""
+          if [ -z "${{ inputs.matrix_topo }}" ] && [ -z "${MATRIX_EC}" ]; then
+            EXPECTED_TOTAL=320
+          fi
+          jq -e --arg testing_sha "${AUTO_TESTING_SHA}" --argjson expected_total "${EXPECTED_TOTAL:-null}" '
+            .schema_version == 1 and
+            .auto_testing_sha == $testing_sha and
+            .complete == true and
+            .counts.expected == ($expected_total // .counts.expected) and
+            .counts.expected == .counts.seen and
+            .counts.harness_errors == 0 and
+            (.expected_cases | length) == .counts.expected and
+            (.expected_cases | unique | length) == .counts.expected and
+            (.seen_cases | length) == .counts.seen and
+            (.missing_cases | length) == 0 and
+            (.harness_errors | length) == 0 and
+            (.cases | length) == .counts.seen and
+            ([.cases[].case_id] | unique | length) == .counts.seen and
+            ([.cases[].case_id] | sort) == (.expected_cases | sort) and
+            all(.cases[]; (.case_id | startswith("M-")) and (.verdict == "pass" or .verdict == "known-divergence" or .verdict == "UNEXPECTED")) and
+            .counts.unexpected == ([.cases[] | select(.verdict == "UNEXPECTED")] | length) and
+            .counts.known_divergence == ([.cases[] | select(.verdict == "known-divergence")] | length)
+          ' "${RESULTS_FILE}" >/dev/null
+
+      - name: Upload test logs & evidence
+        timeout-minutes: 2
+        if: ${{ always() && steps.evidence.outcome == 'success' }}
+        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
+        with:
+          name: rustfs-fault-tolerance-matrix-${{ github.run_id }}-${{ github.run_attempt }}
+          path: |
+            ${{ env.FUNCTIONAL_ARTIFACTS_DIR }}/report.md
+            ${{ env.FUNCTIONAL_ARTIFACTS_DIR }}/results.json
+            ${{ env.FUNCTIONAL_ARTIFACTS_DIR }}/suite.log
+            ${{ env.FUNCTIONAL_ARTIFACTS_DIR }}/evidence/
+          if-no-files-found: warn
+
+      - name: Cleanup environment (after)
+        timeout-minutes: 5
+        if: ${{ always() && inputs.cleanup_after != 'false' }}
+        run: |
+          ./auto-testing/rustfs-fault-tolerance-test.sh --cleanup -y --log-file "${LOG_FILE}" || true
@@ -149,7 +149,7 @@ jobs:
          CHAIN_MANIFEST: ${{ inputs.chain_manifest }}
        run: python3 scripts/prepare_functional_package.py prepare

-      - name: Run fault-tolerance scenarios (A, B, C, C2, D)
+      - name: Run fault-tolerance scenarios (A, B, C, C2, D, E)
        timeout-minutes: 45
        id: test
        # Standalone case failures are reported by the backlog manager.
@@ -194,21 +194,21 @@ jobs:
          jq -e --arg testing_sha "${AUTO_TESTING_SHA}" '
            .schema_version == 1 and
            .auto_testing_sha == $testing_sha and
-            .selected_scenarios == ["A", "B", "C", "C2", "D"] and
+            .selected_scenarios == ["A", "B", "C", "C2", "D", "E"] and
            .complete == true and
-            .counts.expected == 38 and
-            .counts.seen == 38 and
+            .counts.expected == 44 and
+            .counts.seen == 44 and
            .counts.harness_errors == 0 and
-            (.expected_cases | length) == 38 and
-            (.expected_cases | unique | length) == 38 and
-            (.seen_cases | length) == 38 and
-            (.seen_cases | unique | length) == 38 and
+            (.expected_cases | length) == 44 and
+            (.expected_cases | unique | length) == 44 and
+            (.seen_cases | length) == 44 and
+            (.seen_cases | unique | length) == 44 and
            (.seen_cases | sort) == (.expected_cases | sort) and
            (.missing_cases | length) == 0 and
            (.harness_errors | length) == 0 and
-            (.cases | length) == 38 and
+            (.cases | length) == 44 and
            ([.cases[].case_id] | sort) == (.expected_cases | sort) and
-            ([.cases[].case_id] | unique | length) == 38 and
+            ([.cases[].case_id] | unique | length) == 44 and
            all(.cases[]; .verdict == "pass" or .verdict == "known-divergence" or .verdict == "UNEXPECTED") and
            .counts.unexpected == ([.cases[] | select(.verdict == "UNEXPECTED")] | length) and
            .counts.known_divergence == ([.cases[] | select(.verdict == "known-divergence")] | length)
@@ -54,7 +54,10 @@ jobs:
      - name: Resolve published candidate
        id: candidate
        env:
-          GH_TOKEN: ${{ github.token }}
+          # The staleness probe in resolve_functional_candidate.py reads the
+          # private rustfs/auto-testing repo; github.token cannot, and a probe
+          # failure silently keeps a stale testing_sha pinned for the chain.
+          GH_TOKEN: ${{ secrets.PF_TESTING_GH_TOKEN }}
          BUILD_RUN_ID: ${{ inputs.build_run_id }}
          BUILD_RUN_ATTEMPT: ${{ inputs.build_run_attempt }}
          CHAIN_OUTPUT: ${{ runner.temp }}/chain-candidate.json
@@ -1984,9 +1984,9 @@ dependencies = [

 [[package]]
 name = "cc"
-version = "1.4.6"
+version = "1.4.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a3eb0f42d6c360dc3f8a821f6bf2fdea7f72bfd36b3076eb0e6d1e9e0752fff4"
+checksum = "54413ede23c2daf518f35156dfde027feb2374004d63bd497f983c8db9c0e313"
 dependencies = [
 "find-msvc-tools",
 "jobserver",
@@ -3792,9 +3792,9 @@ dependencies = [

 [[package]]
 name = "dial9"
-version = "0.5.0"
+version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "49b11bce55173f77d9003cce8f129b4af889a4639005c53c3498cef933bc895c"
+checksum = "bab33c9f10e909fb3f1f6af2a9d976398ea793b90d1d286b4e23647a7d131d53"
 dependencies = [
 "bon",
 "dial9-core",
@@ -3809,9 +3809,9 @@ dependencies = [

 [[package]]
 name = "dial9-core"
-version = "0.5.0"
+version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9e8cbbc8955394be626249a3b52ddd6bf664373661eaeae70d87eb12bf6f20b6"
+checksum = "ffd7231006b074b02765ae929bded4757a414224b06c7db324d3e6e902f182ee"
 dependencies = [
 "arc-swap",
 "bon",
@@ -3842,9 +3842,9 @@ dependencies = [

 [[package]]
 name = "dial9-metrique"
-version = "0.5.0"
+version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d42fe0e3e14e8780c4ecf9ee14c183dbcb9b1ae3612cc9a689e7dc83b13adef1"
+checksum = "f987e56dbafcab5998acc8e77eb91aaf66fa5441a691b38e99c8058c70725f58"
 dependencies = [
 "dial9-core",
 "dial9-trace-format",
@@ -3856,9 +3856,9 @@ dependencies = [

 [[package]]
 name = "dial9-perf-self-profile"
-version = "0.5.0"
+version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a8f65948455c504bf08576c7b5cfe93bfcaa486d661dc4ee85c2a6309ab89629"
+checksum = "d81f74ada03919614ca5b1ba767271d00980376a5a0e0fc43f8c265bf90413d4"
 dependencies = [
 "blazesym",
 "bon",
@@ -3875,9 +3875,9 @@ dependencies = [

 [[package]]
 name = "dial9-tokio-telemetry"
-version = "0.5.0"
+version = "0.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "af1244091367c805b5d98a6a590f8ab992efda06e6d2560a6967ef898ffd30a2"
+checksum = "934c027f628c59ae4c9664ca4dce758843c3da027dee5052d358f86a5d292ae6"
 dependencies = [
 "bon",
 "bytes",
@@ -4410,9 +4410,9 @@ dependencies = [

 [[package]]
 name = "find-msvc-tools"
-version = "0.1.12"
+version = "0.1.13"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3e0f1c7c3a72c66fd80abe965175f7523475c0489a87d3ff9d6e8c87d87a9d2d"
+checksum = "ef25905e51abafe4dcea6c15fec58c57b601cdbd0ee53d22ea1d3016c587d39b"

 [[package]]
 name = "findshlibs"
@@ -10408,6 +10408,8 @@ dependencies = [
 "opentelemetry_sdk",
 "percent-encoding",
 "pyroscope",
+ "rcgen",
+ "reqwest",
 "rustfs-audit",
 "rustfs-common",
 "rustfs-config",
@@ -10421,6 +10423,8 @@ dependencies = [
 "rustfs-storage-api",
 "rustfs-utils",
 "rustix",
+ "rustls",
+ "rustls-pki-types",
 "serde",
 "serde_json",
 "sysinfo",
@@ -10428,6 +10432,7 @@ dependencies = [
 "tempfile",
 "thiserror 2.0.20",
 "tokio",
+ "tokio-rustls",
 "tokio-util",
 "tracing",
 "tracing-appender",
@@ -11258,9 +11263,8 @@ checksum = "9774ba4a74de5f7b1c1451ed6cd5285a32eddb5cccb8cc655a4e50009e06477f"

 [[package]]
 name = "s3s"
-version = "0.16.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "87d22507a8ebd0ea64972c76dfedbc477418dc70afd98bfce928263894067d2f"
+version = "0.16.0"
+source = "git+https://github.com/s3s-project/s3s.git?rev=50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6#50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6"
 dependencies = [
 "arc-swap",
 "arrayvec",
@@ -11278,7 +11282,6 @@ dependencies = [
 "http 1.5.0",
 "http-body 1.1.0",
 "http-body-util",
- "httparse",
 "hyper",
 "itoa",
 "jiff",
@@ -11291,6 +11294,7 @@ dependencies = [
 "pin-project-lite",
 "quick-xml",
 "regex",
+ "s3s-multipart",
 "s3s-rfc2047",
 "s3s-sigv2",
 "s3s-sigv4",
@@ -11315,11 +11319,23 @@ dependencies = [
 "zeroize",
 ]

+[[package]]
+name = "s3s-multipart"
+version = "0.17.0-alpha.1"
+source = "git+https://github.com/s3s-project/s3s.git?rev=50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6#50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6"
+dependencies = [
+ "arrayvec",
+ "bytes",
+ "futures-core",
+ "httparse",
+ "memchr",
+ "thiserror 2.0.20",
+]
+
 [[package]]
 name = "s3s-rfc2047"
-version = "0.16.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f9d3e6b36be7cca3da377b4cabcd668ffe5b16c3b73a83f7a8f7b63ff6a178ee"
+version = "0.16.0"
+source = "git+https://github.com/s3s-project/s3s.git?rev=50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6#50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6"
 dependencies = [
 "base64-simd",
 "thiserror 2.0.20",
@@ -11327,9 +11343,8 @@ dependencies = [

 [[package]]
 name = "s3s-sigv2"
-version = "0.16.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d96880b5ad8d639c9224155126741708aeb62ad3682e0476d2fd14290af808ef"
+version = "0.16.0"
+source = "git+https://github.com/s3s-project/s3s.git?rev=50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6#50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6"
 dependencies = [
 "base64-simd",
 "hmac 0.13.0",
@@ -11341,9 +11356,8 @@ dependencies = [

 [[package]]
 name = "s3s-sigv4"
-version = "0.16.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fed605cbaaa1cf4d82e8ee15098fb3e42d49c92bb6215eb1359f63bfc0c6b020"
+version = "0.16.0"
+source = "git+https://github.com/s3s-project/s3s.git?rev=50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6#50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6"
 dependencies = [
 "arrayvec",
 "base64-simd",
@@ -313,8 +313,8 @@ rustify = { version = "0.7", default-features = false }
 rustix = { version = "1.1.5" }
 rust-embed = { version = "8.12.0" }
 rustc-hash = { version = "2.1.3" }
-s3s = { version = "0.16.1", features = ["minio"] }
-s3s-sigv4 = { version = "0.16.1" }
+s3s = { git = "https://github.com/s3s-project/s3s.git", rev = "50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6", features = ["minio"] }
+s3s-sigv4 = { git = "https://github.com/s3s-project/s3s.git", rev = "50c94aeb1e5ea9ef8e7393ef6bbc87b72f6891c6" }
 serial_test = "4.0.1"
 shadow-rs = { default-features = false, version = "2.0.0" }
 siphasher = "1.0.3"
@@ -351,7 +351,7 @@ zstd = "0.14.0"
 # Observability and Metrics
 metrics = "0.24.6"
 metrics-util = "0.20"
-dial9 = { version = "0.5.0", default-features = false }
+dial9 = { version = "0.5.1", default-features = false }
 opentelemetry = { version = "0.32.0" }
 opentelemetry-appender-tracing = { version = "0.32.0" }
 opentelemetry-otlp = { version = "0.32.0" }
@@ -30,6 +30,9 @@ pub const ENV_OBS_ENDPOINT_TIMEOUT_MILLIS: &str = "RUSTFS_OBS_ENDPOINT_TIMEOUT_M
 pub const ENV_OBS_ENDPOINT_TRACES_TIMEOUT_MILLIS: &str = "RUSTFS_OBS_ENDPOINT_TRACES_TIMEOUT_MILLIS";
 pub const ENV_OBS_ENDPOINT_METRICS_TIMEOUT_MILLIS: &str = "RUSTFS_OBS_ENDPOINT_METRICS_TIMEOUT_MILLIS";
 pub const ENV_OBS_ENDPOINT_LOGS_TIMEOUT_MILLIS: &str = "RUSTFS_OBS_ENDPOINT_LOGS_TIMEOUT_MILLIS";
+/// Absolute path to a PEM bundle whose certificates are added to the system
+/// trust roots used by OTLP/HTTP exporters.
+pub const ENV_OBS_TLS_CA_FILE: &str = "RUSTFS_OBS_TLS_CA_FILE";
 pub const ENV_OBS_PROFILING_ENDPOINT: &str = "RUSTFS_OBS_PROFILING_ENDPOINT";
 pub const ENV_OBS_USE_STDOUT: &str = "RUSTFS_OBS_USE_STDOUT";
 pub const ENV_OBS_SAMPLE_RATIO: &str = "RUSTFS_OBS_SAMPLE_RATIO";
@@ -55,6 +58,13 @@ pub const ENV_OBS_LOG_FILENAME: &str = "RUSTFS_OBS_LOG_FILENAME";
 pub const ENV_OBS_LOG_ROTATION_TIME: &str = "RUSTFS_OBS_LOG_ROTATION_TIME";
 pub const ENV_OBS_LOG_KEEP_FILES: &str = "RUSTFS_OBS_LOG_KEEP_FILES";

+/// Span lifecycle events emitted by the JSON log layers (`none`|`close`|`full`).
+///
+/// Defaults to `none`: a record per span lifecycle multiplies log volume on
+/// object hot paths, and emitting the ancestor span list on every record
+/// produced multi-megabyte lines that journald truncated (backlog#2642).
+pub const ENV_OBS_LOG_SPAN_EVENTS: &str = "RUSTFS_OBS_LOG_SPAN_EVENTS";
+
 /// Log cleanup related configurations
 pub const ENV_OBS_LOG_MAX_TOTAL_SIZE_BYTES: &str = "RUSTFS_OBS_LOG_MAX_TOTAL_SIZE_BYTES";
 pub const ENV_OBS_LOG_MAX_SINGLE_FILE_SIZE_BYTES: &str = "RUSTFS_OBS_LOG_MAX_SINGLE_FILE_SIZE_BYTES";
@@ -99,6 +109,8 @@ pub const DEFAULT_OBS_LOG_DRY_RUN: bool = false;
 pub const DEFAULT_OBS_LOG_MATCH_MODE_PREFIX: &str = "prefix";
 pub const DEFAULT_OBS_LOG_MATCH_MODE: &str = "suffix";

+pub const DEFAULT_OBS_LOG_SPAN_EVENTS: &str = "none";
+
 /// Default values for observability configuration
 // ### Supported Environment Values
 // - `production` - Secure file-only logging
@@ -141,6 +153,7 @@ mod tests {
        assert_eq!(ENV_OBS_LOG_FILENAME, "RUSTFS_OBS_LOG_FILENAME");
        assert_eq!(ENV_OBS_LOG_ROTATION_TIME, "RUSTFS_OBS_LOG_ROTATION_TIME");
        assert_eq!(ENV_OBS_LOG_KEEP_FILES, "RUSTFS_OBS_LOG_KEEP_FILES");
+        assert_eq!(ENV_OBS_LOG_SPAN_EVENTS, "RUSTFS_OBS_LOG_SPAN_EVENTS");
        assert_eq!(ENV_OBS_TRACES_EXPORT_ENABLED, "RUSTFS_OBS_TRACES_EXPORT_ENABLED");
        assert_eq!(ENV_OBS_METRICS_EXPORT_ENABLED, "RUSTFS_OBS_METRICS_EXPORT_ENABLED");
        assert_eq!(ENV_OBS_LOGS_EXPORT_ENABLED, "RUSTFS_OBS_LOGS_EXPORT_ENABLED");
@@ -178,6 +191,7 @@ mod tests {
        assert_eq!(DEFAULT_OBS_LOG_COMPRESSION_ALGORITHM_GZIP, "gzip");
        assert_eq!(DEFAULT_OBS_LOG_COMPRESSION_ALGORITHM_ZSTD, "zstd");
        assert_eq!(DEFAULT_OBS_LOG_MATCH_MODE_PREFIX, "prefix");
+        assert_eq!(DEFAULT_OBS_LOG_SPAN_EVENTS, "none");
        assert_eq!(DEFAULT_OBS_LOG_MATCH_MODE, "suffix");
        assert_eq!(DEFAULT_OBS_LOG_COMPRESSION_ALGORITHM, "zstd");
    }
@@ -184,7 +184,7 @@ async fn four_node_health_inventory_metrics_and_audit_delivery_are_consistent()
        let (status, metrics_body) = admin_request(
            &node.url,
            Method::GET,
-            "/rustfs/admin/v3/metrics?n=1&by-host=true&by-disk=true",
+            "/rustfs/admin/v3/realtime?n=1&by-host=true&by-disk=true",
            None,
            &dist.cluster.access_key,
            &dist.cluster.secret_key,
@@ -257,7 +257,7 @@ async fn node_admin_body(dist: &DistCluster, node: usize, path: &str) -> TestRes
 }

 async fn http_put_counts(dist: &DistCluster, node: usize) -> TestResult<[u64; 2]> {
-    let body = node_admin_body(dist, node, "/rustfs/admin/v3/metrics?types=512&by-host=true&n=1").await?;
+    let body = node_admin_body(dist, node, "/rustfs/admin/v3/realtime?types=512&by-host=true&n=1").await?;
    let sample: RealtimeMetrics = serde_json::from_str(body.lines().next().ok_or("empty HTTP metrics stream")?)?;
    assert!(sample.errors.is_empty(), "HTTP metrics returned errors: {:?}", sample.errors);
    let http = sample.aggregated.http.ok_or("HTTP metrics missing at WARN log level")?;
@@ -504,6 +504,13 @@ pub mod notification {
    };
 }

+pub mod integrity {
+    pub use crate::services::integrity::{
+        IntegrityError, InventoryItem, InventoryPage, ItemRequest, ItemResult, ItemState, Job, JobMode, JobRequest, JobState,
+        Protection, Readiness, control_job, create_job, get_job, inventory, readiness, resume_job,
+    };
+}
+
 pub mod object {
    pub use crate::object_api::{
        BLOCK_SIZE_V2, ERASURE_ALGORITHM, EncryptionResolutionError, EncryptionResolutionErrorKind, GetObjectBodyCacheHook,
@@ -35,6 +35,7 @@ use crate::bucket::lifecycle::tier_sweeper::{
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::{Error, Result as EcstoreResult};
 use crate::object_api::{ObjectInfo, ObjectOptions};
+use crate::services::notification_sys::acquire_remote_version_state_writer_fleet_proof;
 use crate::services::tier::{tier::TierConfigMgr, warm_backend::TransitionCandidateProbe};
 use crate::storage_api_contracts::{
    list::ListOperations as _,
@@ -1237,6 +1238,12 @@ pub async fn delete_transition_candidate_for_operator(
    lease
        .validate_remote_version_id(remote_version_id)
        .map_err(TransitionOperatorError::Remote)?;
+    if remote_version_requires_fleet_proof(remote_version_id) && acquire_remote_version_state_writer_fleet_proof().is_none() {
+        return Err(TransitionOperatorError::Remote(std::io::Error::new(
+            std::io::ErrorKind::Unsupported,
+            "opaque remote tier version cleanup requires the operator-attested live fleet capability gate",
+        )));
+    }
    let before_delete_probe = lease
        .probe_transition_candidate_for(&transaction.remote_object, transaction.transaction_id)
        .await
@@ -1845,6 +1852,11 @@ async fn recover_unknown_upload_outcome(
            ))
        }
        TransitionCandidateProbe::VersionedPresent(version_id) => {
+            if remote_version_requires_fleet_proof(&version_id) && acquire_remote_version_state_writer_fleet_proof().is_none() {
+                return Ok(TransitionTransactionRecoveryOutcome::RetainedAmbiguous(
+                    IlmRecoveryErrorCode::RemoteVersionUnknown,
+                ));
+            }
            cleanup_recovered_unknown_upload_candidate(api, transaction, TransitionRemoteVersion::versioned(version_id)).await
        }
        TransitionCandidateProbe::Ambiguous => Ok(TransitionTransactionRecoveryOutcome::RetainedAmbiguous(
@@ -1940,6 +1952,14 @@ fn transition_source_lookup_options(transaction: &TransitionTransaction) -> Obje
 async fn delete_transition_remote_candidate(api: Arc<ECStore>, transaction: &TransitionTransaction) -> EcstoreResult<()> {
    let version_id = transaction.remote_version.tier_delete_version_id().unwrap_or_default();
    let version_id_exact = transaction.remote_version.kind == TransitionRemoteVersionKind::Versioned;
+    if version_id_exact
+        && remote_version_requires_fleet_proof(version_id)
+        && acquire_remote_version_state_writer_fleet_proof().is_none()
+    {
+        return Err(Error::other(
+            "opaque remote tier version cleanup requires the operator-attested live fleet capability gate",
+        ));
+    }
    delete_object_from_remote_tier_idempotent_with_manager_and_identity(
        &transaction.remote_object,
        version_id,
@@ -1953,6 +1973,10 @@ async fn delete_transition_remote_candidate(api: Arc<ECStore>, transaction: &Tra
    .map_err(Error::other)
 }

+fn remote_version_requires_fleet_proof(version_id: &str) -> bool {
+    !version_id.is_empty() && Uuid::parse_str(version_id).is_err()
+}
+
 pub async fn recover_transition_transaction_records(
    api: Arc<ECStore>,
    limit: usize,
@@ -2439,6 +2463,14 @@ mod tests {
        }
    }

+    #[test]
+    fn opaque_remote_versions_require_the_fleet_proof_boundary() {
+        assert!(!remote_version_requires_fleet_proof(""));
+        assert!(!remote_version_requires_fleet_proof(&Uuid::new_v4().to_string()));
+        assert!(remote_version_requires_fleet_proof("null"));
+        assert!(remote_version_requires_fleet_proof("b2-opaque-version"));
+    }
+
    #[test]
    fn remote_version_distinguishes_unknown_unversioned_and_versioned() {
        assert_eq!(TransitionRemoteVersion::known_from_put_response("").tier_delete_version_id(), None);
@@ -1464,9 +1464,13 @@ pub(crate) async fn load_bucket_metadata_parse_with_presence(
                "bucket incarnation sidecar is missing for new-format metadata: {bucket}"
            )));
        }
-    } else if incarnation.is_some() {
-        return Err(Error::other("bucket incarnation sidecar exists without bucket metadata"));
    }
+    // A sidecar without `.metadata.bin` is the window between the two legacy
+    // migration writes (rustfs/rustfs#8003): the sidecar lands first, and a
+    // crash or lost namespace lease before the metadata write leaves the
+    // bucket in this state on every node. Report it as not persisted so the
+    // migration runs again; the writers adopt the stored incarnation instead
+    // of minting a new one, so the sidecar keeps its authority.

    bm.default_timestamps();

@@ -1291,6 +1291,28 @@ pub(crate) async fn get_object_lock_config_and_incarnation_from_disk_in(
    }
 }

+/// Inspect all migration-relevant settings while the caller holds an Object Lock
+/// snapshot's lifecycle and metadata transaction guards. Cached settings are not
+/// sufficient to authorize a direct storage writer that cannot apply S3 defaults.
+pub(crate) async fn integrity_migration_metadata_in(
+    ctx: &crate::runtime::instance::InstanceContext,
+    bucket: &str,
+) -> Result<Arc<BucketMetadata>> {
+    let sys = bucket_metadata_sys_of(ctx)?.read().await.clone();
+    match sys
+        .read_authoritative_metadata_from_disk_under_transaction_lock(bucket)
+        .await?
+    {
+        BucketMetadataAuthority::Authoritative(metadata)
+            if metadata.bucket_incarnation_sidecar && !metadata.bucket_incarnation_id.is_nil() =>
+        {
+            Ok(metadata)
+        }
+        BucketMetadataAuthority::MissingBucket => Err(Error::BucketNotFound(bucket.to_string())),
+        _ => Err(Error::other("migration requires authoritative bucket metadata")),
+    }
+}
+
 /// Re-read the quota configuration and bucket incarnation from the same
 /// authoritative metadata blob while the caller holds the bucket metadata
 /// transaction read lock.
@@ -2504,6 +2526,16 @@ impl BucketMetadataSys {
        if !persisted {
            metadata = BucketMetadata::new(bucket);
            metadata.created = bucket_info.created.unwrap_or(OffsetDateTime::UNIX_EPOCH);
+            // An interrupted migration may already have published this
+            // bucket's incarnation; re-read it under the transaction lock so
+            // the retry never replaces an identity other nodes fenced on. A
+            // retired incarnation is residue of a deleted bucket and must not
+            // come back, or heal would reclaim the bucket's new objects.
+            if let Some(stored) = load_bucket_incarnation(self.object_store(), bucket).await?
+                && !crate::bucket::retirement::is_retired(self.object_store(), bucket, stored).await?
+            {
+                metadata.bucket_incarnation_id = stored;
+            }
        } else if metadata.bucket_incarnation_id.is_nil() {
            metadata.bucket_incarnation_id = Uuid::new_v4();
        }
@@ -3556,6 +3588,98 @@ mod tests {
        assert!(err.to_string().contains("sidecar is missing"));
    }

+    /// rustfs/rustfs#8003: the legacy migration writes the incarnation sidecar
+    /// before `.metadata.bin`. A crash or lost namespace lease between the two
+    /// left every pre-existing bucket answering 500 on every request, with no
+    /// path back. The sidecar-only state must load as a legacy bucket, and the
+    /// retried migration must keep the stored incarnation.
+    #[tokio::test]
+    async fn issue_8003_sidecar_without_metadata_loads_as_legacy_and_migrates_with_its_incarnation() {
+        let (dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let sys = BucketMetadataSys::new(ecstore.clone());
+        let bucket = "issue-8003-sidecar-only";
+        for dir in &dirs {
+            std::fs::create_dir_all(dir.path().join(bucket)).unwrap();
+        }
+        let stored = Uuid::new_v4();
+        save_bucket_incarnation(ecstore.clone(), bucket, stored)
+            .await
+            .expect("simulate the interrupted migration's first write");
+
+        let (fabricated, persisted) = load_bucket_metadata_parse_with_presence(ecstore.clone(), bucket, true)
+            .await
+            .expect("a sidecar without metadata must read as a legacy bucket, not fail closed");
+        assert!(!persisted);
+        assert!(!fabricated.bucket_incarnation_sidecar);
+
+        let (_, fabricated_read) = sys
+            .get_config(bucket)
+            .await
+            .expect("request-path metadata reads must not fail closed on the sidecar-only state");
+        assert!(fabricated_read);
+
+        let migrated = sys
+            .get_authoritative_metadata(bucket)
+            .await
+            .expect("the retried legacy migration must complete");
+        assert!(migrated.bucket_incarnation_sidecar);
+        assert_eq!(
+            migrated.bucket_incarnation_id, stored,
+            "the retry must adopt the published incarnation instead of minting a new one"
+        );
+        assert_eq!(sys.get_bucket_incarnation_id(bucket).await.unwrap(), stored);
+
+        let on_disk = sys.get_config_from_disk(bucket).await.expect("metadata is now persisted");
+        assert!(on_disk.bucket_incarnation_sidecar);
+        assert_eq!(on_disk.bucket_incarnation_id, stored);
+        assert_eq!(load_bucket_incarnation(ecstore, bucket).await.unwrap(), Some(stored));
+    }
+
+    /// A sidecar left behind by a deleted bucket names a retired incarnation.
+    /// The migration must mint a new identity for a same-name volume instead
+    /// of re-publishing the retired one, or heal would treat the bucket's new
+    /// objects as reclaimable residue.
+    #[tokio::test]
+    async fn issue_8003_sidecar_only_migration_does_not_adopt_a_retired_incarnation() {
+        let (dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let sys = BucketMetadataSys::new(ecstore.clone());
+        let bucket = "issue-8003-retired-sidecar";
+        for dir in &dirs {
+            std::fs::create_dir_all(dir.path().join(bucket)).unwrap();
+        }
+        let retired = Uuid::new_v4();
+        save_bucket_incarnation(ecstore.clone(), bucket, retired)
+            .await
+            .expect("residual sidecar");
+        crate::bucket::retirement::commit_retirement(
+            ecstore.clone(),
+            bucket,
+            retired,
+            &ObjectOptions {
+                max_parity: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("retirement record");
+
+        let migrated = sys
+            .get_authoritative_metadata(bucket)
+            .await
+            .expect("the migration must still complete for the same-name volume");
+        assert!(migrated.bucket_incarnation_sidecar);
+        assert_ne!(
+            migrated.bucket_incarnation_id, retired,
+            "a retired incarnation must never be re-published"
+        );
+        assert!(!migrated.bucket_incarnation_id.is_nil());
+        assert_eq!(
+            load_bucket_incarnation(ecstore, bucket).await.unwrap(),
+            Some(migrated.bucket_incarnation_id),
+            "the sidecar must be rewritten to the new incarnation"
+        );
+    }
+
    /// Concurrent cache misses for one bucket must collapse into a single disk
    /// load.
    ///
@@ -199,6 +199,38 @@ fn has_raw_status(err: &SdkError<HeadObjectError>, status: u16) -> bool {
    err.raw_response().is_some_and(|r| r.status().as_u16() == status)
 }

+fn raw_header_value(err: &SdkError<HeadObjectError>, name: &str) -> Option<String> {
+    err.raw_response()
+        .and_then(|response| response.headers().get(name))
+        .filter(|value| !value.is_empty())
+        .map(ToOwned::to_owned)
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum DeleteMarkerHeadOutcome {
+    Missing,
+    Exists,
+    Ambiguous,
+}
+
+fn classify_delete_marker_head_error(err: &SdkError<HeadObjectError>) -> DeleteMarkerHeadOutcome {
+    let (is_not_found, code) = err
+        .as_service_error()
+        .map(|service_err| (service_err.is_not_found(), service_err.code()))
+        .unwrap_or((false, None));
+
+    // Prefer the raw HTTP status when present. HEAD responses commonly have no
+    // body, and a contradictory XML code must not turn a 403/5xx into success.
+    match err.raw_response().map(|response| response.status().as_u16()) {
+        Some(404) => DeleteMarkerHeadOutcome::Missing,
+        Some(405) => DeleteMarkerHeadOutcome::Exists,
+        Some(_) => DeleteMarkerHeadOutcome::Ambiguous,
+        None if is_not_found => DeleteMarkerHeadOutcome::Missing,
+        None if !is_retryable_delete_replication_head_error(false, code) => DeleteMarkerHeadOutcome::Exists,
+        None => DeleteMarkerHeadOutcome::Ambiguous,
+    }
+}
+
 fn metadata_requires_existing_target(op_type: ReplicationType, object_info: &ObjectInfo) -> bool {
    op_type == ReplicationType::Metadata
        && object_info
@@ -1809,32 +1841,14 @@ async fn verify_resync_head_result(
            (roi.size, None)
        }
        Err(err) if roi.delete_marker => {
-            // Verifying a replicated delete marker: only a
-            // definitive 404/NoSuchKey or 405/MethodNotAllowed
-            // confirms the marker propagated. Any other
-            // (retryable/ambiguous) HEAD error leaves the outcome
-            // unverified, so it must count as failed — not as a
-            // blanket success (backlog#862 / #799 B13).
-            // A HEAD carries no body, so the SDK only synthesizes an error
-            // code for 404; a target answering the delete marker's version
-            // with 405 (RustFS, MinIO, AWS) leaves `code()` empty and the
-            // raw status is the only evidence (rustfs/backlog#2479).
-            // A marker *version purge* (`version_purge_status` set) asks
-            // for the opposite outcome: the marker must be gone from the
-            // target, so only absence verifies it and a 405 proves the
-            // marker is still there.
-            let retryable = {
-                let (is_not_found, code) = err
-                    .as_service_error()
-                    .map(|se| (se.is_not_found(), se.code()))
-                    .unwrap_or((false, None));
-                let not_found = is_not_found || has_raw_status(&err, 404);
-                if roi.version_purge_status.is_empty() {
-                    let code = code.or_else(|| has_raw_status(&err, 405).then_some("405"));
-                    is_retryable_delete_replication_head_error(not_found, code)
-                } else {
-                    !not_found
-                }
+            // A marker creation is verified by either absence (404) or
+            // presence (405). A marker-version purge has the opposite goal:
+            // only absence proves completion. All other outcomes remain
+            // failed (backlog#862 / #799 B13).
+            let retryable = if roi.version_purge_status.is_empty() {
+                matches!(classify_delete_marker_head_error(&err), DeleteMarkerHeadOutcome::Ambiguous)
+            } else {
+                !matches!(classify_delete_marker_head_error(&err), DeleteMarkerHeadOutcome::Missing)
            };
            if retryable {
                st.failed_count += 1;
@@ -3553,7 +3567,7 @@ async fn replicate_delete_to_target(
        }
    }

-    if dobj.delete_object.delete_marker && dobj.delete_object.delete_marker_version_id.is_some() {
+    if !is_version_purge && dobj.delete_object.delete_marker && dobj.delete_object.delete_marker_version_id.is_some() {
        match head_object_for_worker(
            tgt_client.as_ref(),
            &tgt_client.bucket,
@@ -3562,20 +3576,48 @@ async fn replicate_delete_to_target(
        )
        .await
        {
-            Ok(_) => {}
+            Ok(head) if head.delete_marker == Some(true) => {
+                // Some targets expose an existing marker as a successful HEAD
+                // with the marker bit instead of the usual 405. Treat that as
+                // positive convergence evidence and retain the target version
+                // id for any later marker purge.
+                rinfo.target_delete_marker_version_id = head.version_id.filter(|version_id| !version_id.is_empty());
+                rinfo.replication_status = ReplicationStatusType::Completed;
+                return rinfo;
+            }
+            Ok(_) => {
+                // A successful HEAD without the delete-marker bit proves that
+                // the addressed version is a regular object (or that the
+                // target omitted the marker contract). Creating a marker in
+                // that state would hide a live version, so fail closed.
+                rinfo.replication_status = ReplicationStatusType::Failed;
+                rinfo.error = Some("target version is not a delete marker".to_string());
+                return rinfo;
+            }
            Err(e) => {
-                let non_retryable = matches!(
-                    e.as_ref(),
-                    SdkError::ServiceError(service_err)
-                        if is_retryable_delete_replication_head_error(
-                            service_err.err().is_not_found(),
-                            service_err.err().code(),
-                        )
-                );
-                if non_retryable {
-                    rinfo.replication_status = ReplicationStatusType::Failed;
-                    rinfo.error = Some(e.to_string());
-                    return rinfo;
+                match classify_delete_marker_head_error(e.as_ref()) {
+                    DeleteMarkerHeadOutcome::Exists => {
+                        // An explicit HEAD of a delete-marker version is expected
+                        // to answer 405 (often without an SDK error code). That
+                        // is positive evidence the marker already exists, so the
+                        // creation is complete and must not issue another
+                        // versionless DELETE that would mint a second marker.
+                        if let Some(version_id) = raw_header_value(e.as_ref(), "x-amz-version-id") {
+                            rinfo.target_delete_marker_version_id = Some(version_id);
+                        }
+                        rinfo.replication_status = ReplicationStatusType::Completed;
+                        return rinfo;
+                    }
+                    DeleteMarkerHeadOutcome::Missing => {}
+                    DeleteMarkerHeadOutcome::Ambiguous => {
+                        // Only a definitive not-found means the marker is absent
+                        // and may be created. Keep transport and ambiguous
+                        // failures fail-closed instead of creating blindly.
+                        rinfo.replication_status = ReplicationStatusType::Failed;
+                        rinfo.error = Some(e.to_string());
+                        mark_replication_target_offline_if_needed(&tgt_client, &e).await;
+                        return rinfo;
+                    }
                }
            }
        }
@@ -6529,6 +6571,110 @@ mod tests {
        }
    }

+    fn marker_creation_dobj(target_arn: &str) -> DeletedObjectReplicationInfo {
+        let mut state = ReplicationState::default();
+        state.targets.insert(target_arn.to_string(), ReplicationStatusType::Pending);
+        DeletedObjectReplicationInfo {
+            bucket: "source".to_string(),
+            target_arn: target_arn.to_string(),
+            op_type: ReplicationType::Heal,
+            delete_object: ReplicationDeletedObject {
+                object_name: "gone.txt".to_string(),
+                delete_marker: true,
+                delete_marker_version_id: Some(Uuid::new_v4()),
+                replication_state: Some(state),
+                ..Default::default()
+            },
+            ..Default::default()
+        }
+    }
+
+    /// A target that already has the marker answers a versioned HEAD with a
+    /// bodiless 405. The live worker must settle the marker as completed and
+    /// avoid minting another marker with a versionless DELETE.
+    #[tokio::test]
+    async fn live_delete_marker_worker_accepts_bodiless_405_without_delete() {
+        let (endpoint, server) = spawn_scripted_target_server(1, |_| {
+            "HTTP/1.1 405 Method Not Allowed\r\nX-Amz-Version-Id: target-marker\r\nContent-Length: 0\r\nConnection: close\r\n\r\n"
+                .to_string()
+        });
+        let target = test_target_client(endpoint);
+        register_test_target(&target).await;
+        let dobj = marker_creation_dobj(&target.arn);
+
+        let rinfo = replicate_delete_to_target(&dobj, target.clone(), None).await;
+
+        assert_eq!(rinfo.replication_status, ReplicationStatusType::Completed, "{rinfo:?}");
+        assert!(rinfo.error.is_none(), "{rinfo:?}");
+        assert_eq!(rinfo.target_delete_marker_version_id.as_deref(), Some("target-marker"));
+        let seen = server.join().expect("test HTTP server should finish");
+        assert_eq!(seen.len(), 1, "an existing marker must not be re-created: {seen:?}");
+        assert!(seen[0].starts_with("HEAD "), "{seen:?}");
+    }
+
+    /// A successful HEAD that does not identify a delete marker is not
+    /// convergence evidence: creating a marker would hide a live version.
+    #[tokio::test]
+    async fn live_delete_marker_worker_rejects_successful_regular_head_without_delete() {
+        let (endpoint, server) = spawn_scripted_target_server(1, |_| empty_response("200 OK"));
+        let target = test_target_client(endpoint);
+        register_test_target(&target).await;
+        let dobj = marker_creation_dobj(&target.arn);
+
+        let rinfo = replicate_delete_to_target(&dobj, target.clone(), None).await;
+
+        assert_eq!(rinfo.replication_status, ReplicationStatusType::Failed, "{rinfo:?}");
+        assert_eq!(rinfo.error.as_deref(), Some("target version is not a delete marker"));
+        let seen = server.join().expect("test HTTP server should finish");
+        assert_eq!(seen.len(), 1, "a regular object must not be hidden by a new marker: {seen:?}");
+        assert!(seen[0].starts_with("HEAD "), "{seen:?}");
+    }
+
+    /// A definitive 404 means the marker is absent, so exactly one marker
+    /// creation DELETE is still required.
+    #[tokio::test]
+    async fn live_delete_marker_worker_creates_marker_after_404() {
+        let (endpoint, server) = spawn_scripted_target_server(2, |line| {
+            if line.starts_with("HEAD ") {
+                empty_response("404 Not Found")
+            } else if line.starts_with("DELETE ") {
+                empty_response("204 No Content")
+            } else {
+                empty_response("500 Unexpected")
+            }
+        });
+        let target = test_target_client(endpoint);
+        register_test_target(&target).await;
+        let dobj = marker_creation_dobj(&target.arn);
+
+        let rinfo = replicate_delete_to_target(&dobj, target.clone(), None).await;
+
+        assert_eq!(rinfo.replication_status, ReplicationStatusType::Completed, "{rinfo:?}");
+        let seen = server.join().expect("test HTTP server should finish");
+        assert_eq!(seen.len(), 2, "a missing marker should be created once: {seen:?}");
+        assert!(seen[0].starts_with("HEAD "), "{seen:?}");
+        assert!(seen[1].starts_with("DELETE "), "{seen:?}");
+        assert!(!seen[1].contains("versionId="), "marker creation must remain versionless: {seen:?}");
+    }
+
+    /// Ambiguous service errors must remain failed and must not fall through
+    /// to a blind marker-creation DELETE.
+    #[tokio::test]
+    async fn live_delete_marker_worker_keeps_ambiguous_head_failure_closed() {
+        let (endpoint, server) = spawn_scripted_target_server(1, |_| empty_response("503 Service Unavailable"));
+        let target = test_target_client(endpoint);
+        register_test_target(&target).await;
+        let dobj = marker_creation_dobj(&target.arn);
+
+        let rinfo = replicate_delete_to_target(&dobj, target.clone(), None).await;
+
+        assert_eq!(rinfo.replication_status, ReplicationStatusType::Failed, "{rinfo:?}");
+        assert!(rinfo.error.is_some(), "{rinfo:?}");
+        let seen = server.join().expect("test HTTP server should finish");
+        assert_eq!(seen.len(), 1, "an ambiguous HEAD must not trigger DELETE: {seen:?}");
+        assert!(seen[0].starts_with("HEAD "), "{seen:?}");
+    }
+
    /// A target answers `HEAD ?versionId=<delete marker>` with a bodiless
    /// 405, so the SDK reports no error code. That is the marker having
    /// propagated, not a failure (rustfs/backlog#2479).
@@ -23,7 +23,7 @@ use std::{
    future::Future,
    io::ErrorKind,
    pin::Pin,
-    sync::{Arc, OnceLock},
+    sync::{Arc, OnceLock, atomic::AtomicBool},
    task::{Context, Poll},
    time::Duration,
 };
@@ -234,6 +234,8 @@ pub struct ListPathRawOptions {
    pub skip_walkdir_total_timeout: bool,
    pub walkdir_timeout: Option<Duration>,
    pub walkdir_stall_timeout: Option<Duration>,
+    /// Shared terminal state for bounded producer walks.
+    pub producer_limit_reached: Option<Arc<AtomicBool>>,
    pub agreed: Option<AgreedFn>,
    pub partial: Option<PartialFn>,
    pub finished: Option<FinishedFn>,
@@ -267,6 +269,7 @@ impl Clone for ListPathRawOptions {
            skip_walkdir_total_timeout: self.skip_walkdir_total_timeout,
            walkdir_timeout: self.walkdir_timeout,
            walkdir_stall_timeout: self.walkdir_stall_timeout,
+            producer_limit_reached: self.producer_limit_reached.clone(),
            #[cfg(test)]
            test_reader_behaviors: self.test_reader_behaviors.clone(),
            #[cfg(test)]
@@ -292,6 +295,7 @@ fn walk_dir_options(opts: &ListPathRawOptions) -> WalkDirOptions {
        skip_total_timeout: opts.skip_walkdir_total_timeout,
        timeout_ms: opts.walkdir_timeout.map(duration_millis),
        stall_timeout_ms: opts.walkdir_stall_timeout.map(duration_millis),
+        producer_limit_reached: opts.producer_limit_reached.clone(),
        ..Default::default()
    }
 }
@@ -2141,7 +2141,7 @@ impl RemoteDisk {
                let file_info_bin = encode_file_info_msgpack(fi)?;
                let mut client = self.get_client().await?;
                let mut request = Request::new(RenameDataRequest {
-                    bucket_incarnation_id: crate::store::bucket_heal_scope(dst_volume)
+                    bucket_incarnation_id: crate::store::bucket_heal_scope_for_object(dst_volume, dst_path)
                        .map(|scope| scope.incarnation.as_bytes().to_vec().into())
                        .unwrap_or_default(),
                    disk: self.endpoint.to_string(),
@@ -4308,6 +4308,13 @@ mod tests {
        object_read_all_disks: Arc<StdMutex<Vec<String>>>,
        format_data: Bytes,
        read_all_data: Bytes,
+        next_format_read_gate: Arc<StdMutex<Option<Arc<TestFormatReadGate>>>>,
+    }
+
+    #[derive(Debug, Default)]
+    struct TestFormatReadGate {
+        entered: tokio::sync::Notify,
+        release: CancellationToken,
    }

    impl AuthenticatedReadPeer {
@@ -4319,9 +4326,16 @@ mod tests {
                object_read_all_disks: Arc::default(),
                format_data,
                read_all_data,
+                next_format_read_gate: Arc::default(),
            }
        }

+        fn pause_next_format_read(&self) -> Arc<TestFormatReadGate> {
+            let gate = Arc::new(TestFormatReadGate::default());
+            *self.next_format_read_gate.lock().expect("format gate lock poisoned") = Some(Arc::clone(&gate));
+            gate
+        }
+
        fn disk_info_calls(&self) -> u32 {
            self.disk_info_calls.load(Ordering::Acquire)
        }
@@ -4429,6 +4443,11 @@ mod tests {
                                let disk = request.disk;
                                peer.read_all_calls.fetch_add(1, Ordering::AcqRel);
                                let data = if is_format_read {
+                                    let gate = peer.next_format_read_gate.lock().expect("format gate lock poisoned").take();
+                                    if let Some(gate) = gate {
+                                        gate.entered.notify_one();
+                                        gate.release.cancelled().await;
+                                    }
                                    peer.format_data.clone()
                                } else {
                                    peer.object_read_all_disks
@@ -7131,6 +7150,319 @@ mod tests {
        peer.stop().await;
    }

+    async fn missing_read_set(peer: &TestGrpcPeer, format: crate::layout::format::FormatV3) -> Arc<crate::set_disk::SetDisks> {
+        let count = format.erasure.sets[0].len();
+        let endpoints = (0..count)
+            .map(|index| Endpoint {
+                url: url::Url::parse(&format!("{}/data/rustfs{index}", peer.addr)).expect("endpoint should parse"),
+                is_local: false,
+                pool_idx: 0,
+                set_idx: 0,
+                disk_idx: i32::try_from(index).expect("fixture index fits"),
+            })
+            .collect();
+        crate::set_disk::SetDisks::new(
+            "missing-read-test".to_string(),
+            Arc::new(tokio::sync::RwLock::new(vec![None; count])),
+            count,
+            0,
+            0,
+            0,
+            endpoints,
+            format,
+            Vec::new(),
+        )
+        .await
+    }
+
+    #[tokio::test]
+    #[serial(remote_disk_recovery_probe)]
+    async fn missing_read_disks_reconnect_before_periodic_monitor_and_coalesce() {
+        runtime_sources::ensure_test_rpc_secret();
+        let mut format = crate::layout::format::FormatV3::new(1, 1);
+        let disk_id = format.erasure.sets[0][0];
+        format.erasure.this = disk_id;
+        let Some(peer) = TestGrpcPeer::spawn(
+            Bytes::from(format.to_json().expect("format should serialize")),
+            Bytes::from_static(b"startup-recovered-data"),
+        )
+        .await
+        else {
+            panic!("authenticated regression peer is required");
+        };
+        let set = missing_read_set(&peer, format).await;
+        let snapshots = futures::future::join_all((0..8).map(|_| set.get_disks_for_data_read())).await;
+        let disk = snapshots[0][0]
+            .as_ref()
+            .expect("a ready peer must be usable before the periodic reconnect");
+        for snapshot in &snapshots {
+            assert!(Arc::ptr_eq(
+                disk,
+                snapshot[0]
+                    .as_ref()
+                    .expect("every waiting read should observe the same handle")
+            ));
+        }
+        assert_eq!(peer.peer.read_all_calls(), 1, "concurrent requests must share the format probe");
+        assert_eq!(disk.get_disk_id().await.expect("disk identity should be available"), Some(disk_id));
+        let raw = disk
+            .read_all("bucket", "object")
+            .await
+            .expect("reconnected handle should read data");
+        assert_eq!(raw, Bytes::from_static(b"startup-recovered-data"));
+        assert_eq!(peer.peer.object_read_all_disks(), vec![disk_id.to_string()]);
+        let healthy = set.get_disks_for_data_read().await;
+        assert!(Arc::ptr_eq(disk, healthy[0].as_ref().expect("healthy handle must remain registered")));
+        assert_eq!(peer.peer.read_all_calls(), 2, "healthy reads must not re-probe the format");
+        disk.close().await.expect("reconnected disk should close");
+        peer.stop().await;
+    }
+
+    #[tokio::test]
+    #[serial(remote_disk_recovery_probe)]
+    async fn missing_read_disks_reject_foreign_format_and_back_off() {
+        runtime_sources::ensure_test_rpc_secret();
+        let reference = crate::layout::format::FormatV3::new(1, 1);
+        let mut foreign = reference.clone();
+        foreign.id = Uuid::new_v4();
+        foreign.erasure.this = reference.erasure.sets[0][0];
+        let Some(peer) = TestGrpcPeer::spawn(
+            Bytes::from(foreign.to_json().expect("foreign format should serialize")),
+            Bytes::from_static(b"foreign-data"),
+        )
+        .await
+        else {
+            panic!("authenticated regression peer is required");
+        };
+        let set = missing_read_set(&peer, reference).await;
+        assert!(
+            set.get_disks_for_data_read().await[0].is_none(),
+            "a matching disk ID cannot authorize a foreign deployment"
+        );
+        assert_eq!(peer.peer.read_all_calls(), 1, "the foreign format must actually be inspected");
+        assert!(set.get_disks_for_data_read().await[0].is_none());
+        assert_eq!(peer.peer.read_all_calls(), 1, "failed probes must not repeat on every degraded GET");
+        assert!(peer.peer.object_read_all_disks().is_empty(), "foreign object data must never be read");
+        peer.stop().await;
+    }
+
+    #[tokio::test]
+    #[serial(remote_disk_recovery_probe)]
+    async fn missing_read_disks_reject_wrong_slot_and_invalid_formats() {
+        runtime_sources::ensure_test_rpc_secret();
+        let mut reference = crate::layout::format::FormatV3::new(1, 2);
+        reference.erasure.this = reference.erasure.sets[0][0];
+        let mut nil = reference.clone();
+        nil.erasure.this = Uuid::nil();
+        let mut unknown = reference.clone();
+        unknown.erasure.this = Uuid::new_v4();
+        let mut future = reference.clone();
+        future.erasure.version = crate::layout::format::FormatErasureVersion::Unknown;
+        for (name, payload, first_accepted) in [
+            ("duplicate slot zero", reference.to_json().expect("format should serialize"), true),
+            ("nil drive", nil.to_json().expect("nil format should serialize"), false),
+            ("unknown drive", unknown.to_json().expect("unknown format should serialize"), false),
+            ("future version", future.to_json().expect("future format should serialize"), false),
+            ("malformed", "not-json".to_string(), false),
+        ] {
+            let peer = TestGrpcPeer::spawn(Bytes::from(payload), Bytes::new())
+                .await
+                .expect("authenticated regression peer is required");
+            let set = missing_read_set(&peer, reference.clone()).await;
+            let disks = set.get_disks_for_data_read().await;
+            assert_eq!(disks[0].is_some(), first_accepted, "{name}: slot zero identity decision");
+            assert!(disks[1].is_none(), "{name}: a peer cannot claim another topology slot");
+            assert_eq!(peer.peer.read_all_calls(), 2, "{name}: both missing slots must be probed");
+            assert!(peer.peer.object_read_all_disks().is_empty(), "{name}: no object payload may be read");
+            for disk in disks.into_iter().flatten() {
+                disk.close().await.expect("accepted disk should close");
+            }
+            peer.stop().await;
+        }
+    }
+
+    #[tokio::test]
+    #[serial(remote_disk_recovery_probe)]
+    async fn missing_read_disks_do_not_claim_local_or_misconfigured_endpoints() {
+        runtime_sources::ensure_test_rpc_secret();
+        let mut reference = crate::layout::format::FormatV3::new(1, 1);
+        reference.erasure.this = reference.erasure.sets[0][0];
+        let peer = TestGrpcPeer::spawn(Bytes::from(reference.to_json().expect("format should serialize")), Bytes::new())
+            .await
+            .expect("authenticated regression peer is required");
+        for (local, pool, set_index, disk_index) in [(true, 0, 0, 0), (false, -1, 0, 0), (false, 0, 1, 0), (false, 0, 0, 1)] {
+            let mut set = missing_read_set(&peer, reference.clone()).await;
+            let endpoint = &mut Arc::get_mut(&mut set)
+                .expect("fixture set must be uniquely owned")
+                .set_endpoints[0];
+            endpoint.is_local = local;
+            endpoint.pool_idx = pool;
+            endpoint.set_idx = set_index;
+            endpoint.disk_idx = disk_index;
+            assert!(set.get_disks_for_data_read().await[0].is_none());
+        }
+        assert_eq!(peer.peer.read_all_calls(), 0, "invalid endpoint ownership must fail before I/O");
+        peer.stop().await;
+    }
+
+    #[tokio::test]
+    #[serial(remote_disk_recovery_probe)]
+    async fn missing_read_disks_preserve_concurrently_published_handles() {
+        runtime_sources::ensure_test_rpc_secret();
+        let mut reference = crate::layout::format::FormatV3::new(1, 1);
+        reference.erasure.this = reference.erasure.sets[0][0];
+        for reader_first in [false, true] {
+            let peer = TestGrpcPeer::spawn(Bytes::from(reference.to_json().expect("format should serialize")), Bytes::new())
+                .await
+                .expect("authenticated regression peer is required");
+            let set = missing_read_set(&peer, reference.clone()).await;
+            let gate = peer.peer.pause_next_format_read();
+            let pending_set = Arc::clone(&set);
+            let pending = tokio::spawn(async move {
+                if reader_first {
+                    pending_set.get_disks_for_data_read().await;
+                } else {
+                    pending_set.renew_disk(&pending_set.set_endpoints[0]).await;
+                }
+            });
+            time::timeout(Duration::from_secs(2), gate.entered.notified())
+                .await
+                .expect("the first reconnect must reach the format barrier");
+            if reader_first {
+                set.renew_disk(&set.set_endpoints[0]).await;
+            } else {
+                set.get_disks_for_data_read().await;
+            }
+            let published = set.disks.read().await[0]
+                .clone()
+                .expect("the competing reconnect must publish");
+            gate.release.cancel();
+            pending.await.expect("paused reconnect must complete");
+            assert!(
+                Arc::ptr_eq(&published, set.disks.read().await[0].as_ref().expect("published slot must remain")),
+                "neither reconnect may replace a handle published while its format I/O was pending"
+            );
+            published.close().await.expect("published disk should close");
+            peer.stop().await;
+        }
+    }
+
+    #[tokio::test]
+    #[serial(remote_disk_recovery_probe, internode_metrics)]
+    async fn missing_read_disks_expired_waiter_does_not_start_another_probe() {
+        // Disable cooldown so this independently proves deadline admission,
+        // rather than passing only because the leader has just finished.
+        temp_env::async_with_vars([(rustfs_config::ENV_DRIVE_RETURNING_PROBE_INTERVAL_SECS, Some("0"))], async {
+            runtime_sources::ensure_test_rpc_secret();
+            let mut reference = crate::layout::format::FormatV3::new(1, 1);
+            reference.erasure.this = reference.erasure.sets[0][0];
+            let peer = TestGrpcPeer::spawn(Bytes::from(reference.to_json().expect("format should serialize")), Bytes::new())
+                .await
+                .expect("authenticated regression peer is required");
+            let set = missing_read_set(&peer, reference).await;
+            let gate = peer.peer.pause_next_format_read();
+            let mut leader = Box::pin(set.get_disks_for_data_read());
+            time::timeout(Duration::from_secs(2), async {
+                tokio::select! {
+                    _ = gate.entered.notified() => {}
+                    _ = &mut leader => panic!("leader must remain blocked in its format RPC"),
+                }
+            })
+            .await
+            .expect("leader must reach the format barrier");
+            time::pause();
+            let mut waiter = Box::pin(set.get_disks_for_data_read());
+            assert!(futures::poll!(&mut waiter).is_pending(), "waiter must queue behind the leader");
+            let requests_before =
+                crate::cluster::rpc::runtime_sources::internode_metrics_snapshot_for_test().outgoing_requests_total;
+
+            time::advance(crate::disk::disk_store::get_drive_active_check_timeout()).await;
+            assert!(leader.await[0].is_none(), "expired leader must not publish a disk");
+            assert!(waiter.await[0].is_none(), "expired waiter must not publish a disk");
+            time::resume();
+
+            // Count client dispatches, not server arrivals: a cancelled RPC can still
+            // be buffered in the transport when the expired waiter returns.
+            assert_eq!(
+                crate::cluster::rpc::runtime_sources::internode_metrics_snapshot_for_test().outgoing_requests_total,
+                requests_before,
+                "an expired waiter must not dispatch another format RPC after the leader releases the lock"
+            );
+            gate.release.cancel();
+            peer.stop().await;
+        })
+        .await;
+    }
+
+    #[tokio::test]
+    #[serial(remote_disk_recovery_probe)]
+    async fn missing_read_disks_timeout_releases_singleflight_and_allows_retry() {
+        runtime_sources::ensure_test_rpc_secret();
+        let mut reference = crate::layout::format::FormatV3::new(1, 1);
+        reference.erasure.this = reference.erasure.sets[0][0];
+        let peer = TestGrpcPeer::spawn(Bytes::from(reference.to_json().expect("format should serialize")), Bytes::new())
+            .await
+            .expect("authenticated regression peer is required");
+        let set = missing_read_set(&peer, reference).await;
+        let gate = peer.peer.pause_next_format_read();
+        let deadline = crate::disk::disk_store::get_drive_active_check_timeout();
+        let snapshots = time::timeout(
+            deadline + Duration::from_secs(2),
+            futures::future::join_all((0..4).map(|_| set.get_disks_for_data_read())),
+        )
+        .await
+        .expect("all waiting requests must respect the bounded probe deadline");
+        assert!(snapshots.iter().all(|snapshot| snapshot[0].is_none()));
+        assert_eq!(peer.peer.read_all_calls(), 1, "only one format RPC may remain stalled");
+        gate.release.cancel();
+        time::pause();
+        time::advance(crate::disk::health_state::get_drive_returning_probe_interval()).await;
+        time::resume();
+        let disks = set.get_disks_for_data_read().await;
+        let disk = disks[0]
+            .as_ref()
+            .expect("a timed-out probe must not prevent a later read from reconnecting");
+        assert_eq!(peer.peer.read_all_calls(), 2, "retry must validate the format again");
+        disk.close().await.expect("reconnected disk should close");
+        peer.stop().await;
+    }
+
+    #[tokio::test]
+    #[serial(remote_disk_recovery_probe)]
+    async fn missing_read_disks_cancellation_leaves_no_published_handle() {
+        runtime_sources::ensure_test_rpc_secret();
+        let mut reference = crate::layout::format::FormatV3::new(1, 1);
+        reference.erasure.this = reference.erasure.sets[0][0];
+        let peer = TestGrpcPeer::spawn(Bytes::from(reference.to_json().expect("format should serialize")), Bytes::new())
+            .await
+            .expect("authenticated regression peer is required");
+        let set = missing_read_set(&peer, reference).await;
+        let gate = peer.peer.pause_next_format_read();
+        let pending_set = Arc::clone(&set);
+        let pending = tokio::spawn(async move { pending_set.get_disks_for_data_read().await });
+        time::timeout(Duration::from_secs(2), gate.entered.notified())
+            .await
+            .expect("probe must reach the format barrier");
+        // Cancellation can happen after the start-based cooldown has elapsed.
+        // It must still cool down before another caller probes the same peer.
+        time::pause();
+        time::advance(crate::disk::health_state::get_drive_returning_probe_interval()).await;
+        time::resume();
+        pending.abort();
+        assert!(pending.await.expect_err("probe must be cancelled").is_cancelled());
+        assert!(
+            set.disks.read().await[0].is_none(),
+            "cancelled format I/O must not publish an unvalidated handle"
+        );
+        let waiting = time::timeout(Duration::from_secs(1), set.get_disks_for_data_read())
+            .await
+            .expect("cancelled probe must release the singleflight lock");
+        assert!(waiting[0].is_none(), "retry cooldown must still apply after cancellation");
+        assert_eq!(peer.peer.read_all_calls(), 1);
+        gate.release.cancel();
+        peer.stop().await;
+    }
+
    #[tokio::test]
    async fn test_copy_stream_with_buffer_copies_full_payload() {
        let payload = b"walk-dir-stream".repeat(1024);
@@ -4253,6 +4253,13 @@ mod decommission_lock_order_tests {
                panic!("cancellation probe finished before observing target contention: {result:?}");
            }
        }
+        // The mutation absorbs the gate contention inline before the outer wait
+        // loop ever sees it, and that inline budget is bounded.
+        assert_eq!(
+            retry_observer.target_gate_inline_retries(),
+            crate::core::pools::DECOMMISSION_MUTATION_GATE_MAX_INLINE_ATTEMPTS - 1,
+            "inline target-gate retries must stop at their bounded attempt budget"
+        );
        tokio::time::sleep(Duration::from_millis(800)).await;
        assert_eq!(
            retry_observer.target_gate_exact_reloads(),
@@ -1493,12 +1493,12 @@ impl Sets {
        object: &str,
        version_id: &str,
        opts: &HealOpts,
-        retirement: Option<&crate::bucket::retirement::MarkerRetirementContext<'_>>,
+        proof: crate::set_disk::AbsenceProofRequest<'_>,
    ) -> Result<(HealResultItem, Option<Error>, Option<crate::set_disk::HealedObjectAbsence>)> {
        let mut absence = None;
        let (item, error) = self
            .get_disks_for_heal_object(object, opts)?
-            .heal_object_with_retirement(bucket, object, version_id, opts, &mut absence, retirement)
+            .heal_object_with_retirement(bucket, object, version_id, opts, &mut absence, proof)
            .await?;
        // A caller-owned lock does not expose its lease to this boundary.
        // Keep cleanup unverified when that lease cannot be checked here.
@@ -244,13 +244,16 @@ impl StdError for ConditionalFileNotCommittedError {
    }
 }

-fn classify_internode_missing_error(error: &InternodeHttpError) -> Option<DiskError> {
+fn classify_internode_disk_error(error: &InternodeHttpError) -> Option<DiskError> {
    if error.is_remote_file_not_found() {
        return Some(DiskError::FileNotFound);
    }
    if error.is_remote_volume_not_found() {
        return Some(DiskError::VolumeNotFound);
    }
+    if error.is_remote_file_corrupt() {
+        return Some(DiskError::FileCorrupt);
+    }
    None
 }

@@ -526,13 +529,10 @@ fn io_error_chain_contains_kind(io_error: &std::io::Error, kind: std::io::ErrorK

 impl From<std::io::Error> for DiskError {
    fn from(e: std::io::Error) -> Self {
-        if let Some(error) = e.get_ref().and_then(|source| source.downcast_ref::<InternodeHttpError>()) {
-            if error.is_remote_file_not_found() {
-                return DiskError::FileNotFound;
-            }
-            if error.is_remote_volume_not_found() {
-                return DiskError::VolumeNotFound;
-            }
+        if let Some(error) = e.get_ref().and_then(|source| source.downcast_ref::<InternodeHttpError>())
+            && let Some(classified) = classify_internode_disk_error(error)
+        {
+            return classified;
        }
        let e = match e.downcast::<TerminalReadError>() {
            Ok(terminal_error) => {
@@ -541,7 +541,7 @@ impl From<std::io::Error> for DiskError {
                    && let Some(internode_error) = io_error
                        .get_ref()
                        .and_then(|source| source.downcast_ref::<InternodeHttpError>())
-                    && let Some(classified) = classify_internode_missing_error(internode_error)
+                    && let Some(classified) = classify_internode_disk_error(internode_error)
                {
                    return classified;
                }
@@ -963,6 +963,7 @@ mod tests {
        for (remote_error, expected) in [
            (rustfs_rio::new_test_remote_file_not_found_http_io_error(), DiskError::FileNotFound),
            (rustfs_rio::new_test_remote_volume_not_found_http_io_error(), DiskError::VolumeNotFound),
+            (rustfs_rio::new_test_remote_file_corrupt_http_io_error(), DiskError::FileCorrupt),
        ] {
            let wrapped = terminal_read_error_to_io(DiskError::Io(remote_error));
            assert_eq!(DiskError::from(wrapped), expected);
@@ -1533,25 +1534,23 @@ mod tests {
    }

    #[test]
-    fn test_internode_missing_errors_preserve_disk_error_types() {
+    fn test_internode_disk_errors_preserve_disk_error_types() {
        let file_missing = DiskError::from(rustfs_rio::new_test_remote_file_not_found_http_io_error());
        let volume_missing = DiskError::from(rustfs_rio::new_test_remote_volume_not_found_http_io_error());
+        let file_corrupt = DiskError::from(rustfs_rio::new_test_remote_file_corrupt_http_io_error());
        let unmarked_server_error = DiskError::from(rustfs_rio::new_test_internode_http_io_error(
            rustfs_rio::InternodeHttpErrorKind::HttpStatus(http::StatusCode::INTERNAL_SERVER_ERROR),
        ));

        assert_eq!(file_missing, DiskError::FileNotFound);
        assert_eq!(volume_missing, DiskError::VolumeNotFound);
+        assert_eq!(file_corrupt, DiskError::FileCorrupt);
        assert!(matches!(unmarked_server_error, DiskError::Io(_)));
-        for missing in [file_missing, volume_missing] {
-            assert_eq!(missing.clone(), missing);
+        for error in [file_missing, volume_missing, file_corrupt] {
+            assert_eq!(error.clone(), error);
            assert_eq!(
-                crate::disk::error_reduce::reduce_write_quorum_errs(
-                    &[Some(missing.clone()), Some(missing.clone()), None],
-                    &[],
-                    2
-                ),
-                Some(missing)
+                crate::disk::error_reduce::reduce_write_quorum_errs(&[Some(error.clone()), Some(error.clone()), None], &[], 2),
+                Some(error)
            );
        }
    }
@@ -1899,6 +1899,61 @@ impl AsyncWrite for DirectWriter {
        }
    }

+    fn poll_write_vectored(
+        self: std::pin::Pin<&mut Self>,
+        cx: &mut std::task::Context<'_>,
+        bufs: &[std::io::IoSlice<'_>],
+    ) -> std::task::Poll<std::io::Result<usize>> {
+        // Bitrot writes arrive as [hash, data]. Coalesce both slices into the
+        // aligned bounce buffer so the O_DIRECT path has the same byte-stream
+        // contract as the buffered writer instead of relying on the trait's
+        // first-slice fallback.
+        let this = self.get_mut();
+        loop {
+            match &mut this.state {
+                DirectWriteState::Busy(_) => {
+                    std::task::ready!(this.poll_drive_busy(cx))?;
+                }
+                DirectWriteState::Idle(inner_opt) => {
+                    let inner = inner_opt.as_mut().expect("idle direct writer must hold inner state");
+                    let capacity = inner.buf.len;
+                    let space = capacity - inner.filled;
+                    let mut remaining = space;
+                    let mut written = 0;
+                    for src in bufs {
+                        if remaining == 0 {
+                            break;
+                        }
+                        let take = src.len().min(remaining);
+                        let start = inner.filled + written;
+                        inner.buf.as_mut_slice()[start..start + take].copy_from_slice(&src[..take]);
+                        written += take;
+                        remaining -= take;
+                    }
+
+                    if written == 0 {
+                        return std::task::Poll::Ready(Ok(0));
+                    }
+                    inner.filled += written;
+
+                    if inner.filled == capacity {
+                        let mut inner = inner_opt.take().expect("idle direct writer must hold inner state");
+                        let handle = tokio::task::spawn_blocking(move || {
+                            let res = inner.flush_batch();
+                            (inner, res)
+                        });
+                        this.state = DirectWriteState::Busy(handle);
+                    }
+                    return std::task::Poll::Ready(Ok(written));
+                }
+            }
+        }
+    }
+
+    fn is_write_vectored(&self) -> bool {
+        true
+    }
+
    fn poll_flush(self: std::pin::Pin<&mut Self>, cx: &mut std::task::Context<'_>) -> std::task::Poll<std::io::Result<()>> {
        // Only drive an in-flight batch to completion. Sub-alignment staged
        // bytes cannot be flushed mid-stream (they would misalign the next
@@ -7556,7 +7611,7 @@ impl LocalDisk {
        objs_returned: &mut i32,
        skip_current_dir_object: bool,
        multipart_dir_to_skip: Option<HashSet<String>>,
-    ) -> Result<()>
+    ) -> Result<bool>
    where
        W: AsyncWrite + Unpin + Send,
    {
@@ -7574,7 +7629,7 @@ impl LocalDisk {
        };

        if opts.limit > 0 && *objs_returned >= opts.limit {
-            return Ok(());
+            return Ok(true);
        }

        // TODO(backlog): add directory listing lock to prevent concurrent enumeration
@@ -7635,12 +7690,12 @@ impl LocalDisk {
                    return Err(DiskError::FileNotFound);
                }

-                return Ok(());
+                return Ok(false);
            }
        };

        if entries.is_empty() {
-            return Ok(());
+            return Ok(false);
        }

        current = current.trim_matches('/').to_owned();
@@ -7654,7 +7709,7 @@ impl LocalDisk {
            let entry = item.clone();
            // check limit
            if opts.limit > 0 && *objs_returned >= opts.limit {
-                return Ok(());
+                return Ok(true);
            }
            // check multipart dir
            if skip_current_dir_object
@@ -7759,14 +7814,14 @@ impl LocalDisk {
        prefix = "".to_owned();

        for entry in entries.iter() {
-            if opts.limit > 0 && *objs_returned >= opts.limit {
-                return Ok(());
-            }
-
            if entry.is_empty() {
                continue;
            }

+            if opts.limit > 0 && *objs_returned >= opts.limit {
+                return Ok(true);
+            }
+
            let name = path_join_buf(&[current.as_str(), entry.as_str()]);

            while let Some((last_name, _, _, _)) = dir_stack.last()
@@ -7780,7 +7835,7 @@ impl LocalDisk {
                // moment the limit is reached, same as the check below this
                // loop guards against for the current entry itself.
                if opts.limit > 0 && *objs_returned >= opts.limit {
-                    return Ok(());
+                    return Ok(true);
                }

                let (pop, skip_object, dir_to_skip, scan_required) = dir_stack.pop().expect("operation should succeed");
@@ -7794,23 +7849,25 @@ impl LocalDisk {
                .await?;

                let scan_path = pop.clone();
-                if opts.recursive
-                    && scan_required
-                    && let Err(er) =
-                        Box::pin(self.scan_dir(pop, prefix.clone(), opts, out, objs_returned, skip_object, dir_to_skip)).await
-                {
-                    if !er.is_metacache_output_stream_closed() {
-                        error!(
-                            event = EVENT_DISK_LOCAL_SCAN_FAILED,
-                            component = LOG_COMPONENT_ECSTORE,
-                            subsystem = LOG_SUBSYSTEM_DISK_LOCAL,
-                            path = %scan_path,
-                            operation = "scan_dir",
-                            error = ?er,
-                            "Disk local scan failed"
-                        );
+                if opts.recursive && scan_required {
+                    match Box::pin(self.scan_dir(pop, prefix.clone(), opts, out, objs_returned, skip_object, dir_to_skip)).await {
+                        Ok(true) => return Ok(true),
+                        Ok(false) => {}
+                        Err(er) => {
+                            if !er.is_metacache_output_stream_closed() {
+                                error!(
+                                    event = EVENT_DISK_LOCAL_SCAN_FAILED,
+                                    component = LOG_COMPONENT_ECSTORE,
+                                    subsystem = LOG_SUBSYSTEM_DISK_LOCAL,
+                                    path = %scan_path,
+                                    operation = "scan_dir",
+                                    error = ?er,
+                                    "Disk local scan failed"
+                                );
+                            }
+                            return Err(er);
+                        }
                    }
-                    return Err(er);
                }
            }

@@ -7821,7 +7878,7 @@ impl LocalDisk {
            // tail, permanently skipping those keys on the next page instead
            // of just deferring them to it.
            if opts.limit > 0 && *objs_returned >= opts.limit {
-                return Ok(());
+                return Ok(true);
            }

            let mut meta = MetaCacheEntry {
@@ -7964,7 +8021,7 @@ impl LocalDisk {

        while let Some((dir, skip_object, dir_to_skip, scan_required)) = dir_stack.pop() {
            if opts.limit > 0 && *objs_returned >= opts.limit {
-                return Ok(());
+                return Ok(true);
            }

            write_metacache_obj(
@@ -7977,27 +8034,29 @@ impl LocalDisk {
            .await?;

            let scan_path = dir.clone();
-            if opts.recursive
-                && scan_required
-                && let Err(er) =
-                    Box::pin(self.scan_dir(dir, prefix.clone(), opts, out, objs_returned, skip_object, dir_to_skip)).await
-            {
-                if !er.is_metacache_output_stream_closed() {
-                    error!(
-                        event = EVENT_DISK_LOCAL_SCAN_FAILED,
-                        component = LOG_COMPONENT_ECSTORE,
-                        subsystem = LOG_SUBSYSTEM_DISK_LOCAL,
-                        path = %scan_path,
-                        operation = "scan_dir",
-                        error = ?er,
-                        "Disk local recursive scan failed"
-                    );
+            if opts.recursive && scan_required {
+                match Box::pin(self.scan_dir(dir, prefix.clone(), opts, out, objs_returned, skip_object, dir_to_skip)).await {
+                    Ok(true) => return Ok(true),
+                    Ok(false) => {}
+                    Err(er) => {
+                        if !er.is_metacache_output_stream_closed() {
+                            error!(
+                                event = EVENT_DISK_LOCAL_SCAN_FAILED,
+                                component = LOG_COMPONENT_ECSTORE,
+                                subsystem = LOG_SUBSYSTEM_DISK_LOCAL,
+                                path = %scan_path,
+                                operation = "scan_dir",
+                                error = ?er,
+                                "Disk local recursive scan failed"
+                            );
+                        }
+                        return Err(er);
+                    }
                }
-                return Err(er);
            }
        }

-        Ok(())
+        Ok(false)
    }

    /// Whether the backing directory of plain object `object_name` also holds
@@ -10121,21 +10180,25 @@ impl DiskAPI for LocalDisk {
            }
        }

-        self.scan_dir(
-            opts.base_dir.clone(),
-            opts.filter_prefix.clone().unwrap_or_default(),
-            &opts,
-            &mut out,
-            &mut objs_returned,
-            skip_current_dir_object,
-            if multipart_dir_to_skip.is_empty() {
-                None
-            } else {
-                Some(multipart_dir_to_skip)
-            },
-        )
-        .await?;
+        let limit_reached = self
+            .scan_dir(
+                opts.base_dir.clone(),
+                opts.filter_prefix.clone().unwrap_or_default(),
+                &opts,
+                &mut out,
+                &mut objs_returned,
+                skip_current_dir_object,
+                if multipart_dir_to_skip.is_empty() {
+                    None
+                } else {
+                    Some(multipart_dir_to_skip)
+                },
+            )
+            .await?;

+        if let Some(flag) = opts.producer_limit_reached.as_ref() {
+            flag.store(limit_reached, std::sync::atomic::Ordering::Release);
+        }
        out.close().await?;
        Ok(())
    }
@@ -22134,7 +22197,7 @@ mod test {
    #[cfg(target_os = "linux")]
    #[tokio::test]
    async fn direct_writer_state_machine_round_trips_over_plain_file() {
-        use std::io::Read;
+        use std::io::{IoSlice, Read};
        use tempfile::tempdir;

        let dir = tempdir().expect("tempdir");
@@ -22153,6 +22216,16 @@ mod test {

            let mut writer = DirectWriter::from_std_file_for_test(file, align, capacity);
            let mut off = 0;
+            if content.len() >= 2 {
+                let split = content.len().min(300);
+                let first = split / 2;
+                let written = writer
+                    .write_vectored(&[IoSlice::new(&content[..first]), IoSlice::new(&content[first..split])])
+                    .await
+                    .expect("vectored write");
+                assert_eq!(written, split, "vectored write must consume both hash/data-like slices");
+                off = split;
+            }
            while off < content.len() {
                let end = (off + 300).min(content.len());
                writer.write_all(&content[off..end]).await.expect("write_all");
@@ -70,7 +70,12 @@ use rustfs_filemeta::{FileInfo, ObjectPartInfo, RawFileInfo};
 use rustfs_madmin::info_commands::DiskMetrics;
 use rustfs_rio::ChunkReaderBox;
 use serde::{Deserialize, Serialize};
-use std::{fmt::Debug, path::PathBuf, sync::Arc, time::Duration};
+use std::{
+    fmt::Debug,
+    path::PathBuf,
+    sync::{Arc, atomic::AtomicBool},
+    time::Duration,
+};
 use time::OffsetDateTime;
 use tokio::io::{AsyncRead, AsyncWrite};
 use uuid::Uuid;
@@ -820,7 +825,7 @@ impl Disk {
        dst_volume: &str,
        dst_path: &str,
    ) -> Result<RenameDataResp> {
-        let Some(scope) = crate::store::bucket_heal_scope(dst_volume) else {
+        let Some(scope) = crate::store::bucket_heal_scope_for_object(dst_volume, dst_path) else {
            return self
                .rename_data_borrowed_with_fence(src_volume, src_path, fi, dst_volume, dst_path, None)
                .await;
@@ -1479,6 +1484,12 @@ pub struct WalkDirOptions {
    // Override the remote stream stall timeout for long background walks.
    #[serde(default)]
    pub stall_timeout_ms: Option<u64>,
+
+    /// In-process completion state for bounded local walks. This is skipped
+    /// from RPC serialization; remote peers retain the legacy natural-EOF
+    /// behavior until they support an explicit capability.
+    #[serde(skip)]
+    pub producer_limit_reached: Option<Arc<AtomicBool>>,
 }

 impl WalkDirOptions {
@@ -1802,6 +1813,7 @@ mod tests {
            skip_total_timeout: false,
            timeout_ms: Some(10_000),
            stall_timeout_ms: Some(20_000),
+            producer_limit_reached: None,
        };

        assert_eq!(opts.bucket, "test-bucket");
@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

+use crate::disk::error::Error as DiskError;
 use crate::erasure::codec::buffer_pool::{get_ec_buffer, return_ec_buffer};
 use pin_project_lite::pin_project;
 use rustfs_utils::HashAlgorithm;
@@ -239,7 +240,7 @@ fn short_shard_read(got: usize, want: usize) -> std::io::Error {
        want,
        "short shard read: got {got} of {want} bytes"
    );
-    std::io::Error::new(std::io::ErrorKind::UnexpectedEof, format!("short shard read: got {got} of {want} bytes"))
+    std::io::Error::new(std::io::ErrorKind::UnexpectedEof, DiskError::FileCorrupt)
 }

 /// Split a `[hash][data]` block, verify the hash (unless `skip_verify`), and
@@ -265,7 +266,7 @@ fn split_and_verify<'a>(hash_algo: &HashAlgorithm, skip_verify: bool, block: &'a
            data_len = data.len(),
            "bitrot hash mismatch"
        );
-        return Err(std::io::Error::new(std::io::ErrorKind::InvalidData, "bitrot hash mismatch"));
+        return Err(std::io::Error::new(std::io::ErrorKind::InvalidData, DiskError::FileCorrupt));
    }
    Ok((data, verify))
 }
@@ -434,7 +435,7 @@ where
                            data_len = want,
                            "bitrot hash mismatch"
                        );
-                        return Err(std::io::Error::new(std::io::ErrorKind::InvalidData, "bitrot hash mismatch"));
+                        return Err(std::io::Error::new(std::io::ErrorKind::InvalidData, DiskError::FileCorrupt));
                    }
                    hash_offset += take;
                    remaining -= take;
@@ -1141,6 +1142,26 @@ mod tests {
    use std::time::Duration;
    use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt, ReadBuf};

+    #[test]
+    fn bitrot_damage_preserves_typed_repair_evidence() {
+        use crate::disk::error::Error as DiskError;
+        let algo = HashAlgorithm::HighwayHash256;
+        let data = b"repair evidence";
+        let mut frame = algo.hash_encode(data).as_ref().to_vec();
+        frame.extend_from_slice(data);
+        *frame.last_mut().expect("test frame has payload") ^= 1;
+        let error = super::split_and_verify(&algo, false, &frame).expect_err("a hash mismatch must reject the shard");
+        assert_eq!(error.kind(), io::ErrorKind::InvalidData);
+        assert_eq!(DiskError::from(error), DiskError::FileCorrupt);
+        let error = super::short_shard_read(3, 4);
+        assert_eq!(error.kind(), io::ErrorKind::UnexpectedEof);
+        assert_eq!(DiskError::from(error), DiskError::FileCorrupt);
+        assert!(matches!(
+            DiskError::from(io::Error::new(io::ErrorKind::InvalidData, "unrelated invalid data")),
+            DiskError::Io(_)
+        ));
+    }
+
    struct FragmentedSource {
        chunks: VecDeque<Bytes>,
    }
@@ -26,7 +26,7 @@ use crate::erasure::codec::workspace::ShardBufferPool;
 use crate::erasure::coding::{BitrotReader, Erasure};
 use crate::io_support::bitrot::DeferredReaderStripeHandle;
 use crate::set_disk::shard_source::{
-    INLINE_SHARD_SLOTS, ShardBuffers, ShardErrors, ShardReadCost, ShardStripeSource, StripeReadState,
+    INLINE_SHARD_SLOTS, ShardBuffers, ShardErrors, ShardReadCost, ShardReadRepair, ShardStripeSource, StripeReadState,
 };
 use futures::FutureExt;
 use futures::stream::{FuturesUnordered, StreamExt};
@@ -46,7 +46,13 @@ type ShardReadFuture<'a> = Pin<Box<dyn Future<Output = (usize, ShardReadCost, Re
 type OwnedShardReadFuture<'a, R> =
    Pin<Box<dyn Future<Output = (usize, ShardReadCost, Result<Vec<u8>, Error>, Option<BitrotReader<R>>, bool)> + Send + 'a>>;
 pub(crate) type DeferredReaderReopener<R> = Arc<dyn Fn(usize) -> Option<BitrotReader<R>> + Send + Sync>;
-pub(crate) type DecodeOutcome = (usize, Option<std::io::Error>, bool);
+#[derive(Debug, Default)]
+pub(crate) struct DecodeOutcome {
+    pub(crate) written: usize,
+    pub(crate) error: Option<std::io::Error>,
+    pub(crate) exact_quorum: bool,
+    pub(crate) repair: ShardReadRepair,
+}

 type ShardIndexes = SmallVec<[usize; INLINE_SHARD_SLOTS]>;
 type ActiveReaders = SmallVec<[bool; INLINE_SHARD_SLOTS]>;
@@ -596,6 +602,7 @@ pub(crate) struct ParallelReader<R> {
    hedged_readers: ActiveReaders,
    stripe_index: usize,
    integrity: Option<std::sync::Arc<crate::io_support::shard_integrity::PartProofReader>>,
+    repair_evidence: Option<Arc<ShardReadRepair>>,
 }
 }

@@ -808,9 +815,15 @@ where
            deferred_reopeners: Vec::new(),
            hedged_readers,
            stripe_index: 0,
+            repair_evidence: None,
        }
    }

+    pub(crate) fn with_repair_evidence(mut self, evidence: Arc<ShardReadRepair>) -> Self {
+        self.repair_evidence = Some(evidence);
+        self
+    }
+
    /// Attach the per-slot deferred stripe handles produced during bitrot
    /// reader setup. Only parity slots are consulted: a handle lets the
    /// lockstep path open a parity shard aligned to the stripe where a data
@@ -2027,6 +2040,9 @@ where
            .take()
            .unwrap_or_else(|| Box::new(StripeReadState::with_slot_count(self.readers.len(), self.data_shards)));
        self.read_into_state(&mut state).await;
+        if let Some(evidence) = &self.repair_evidence {
+            evidence.observe(state.parts_mut().1);
+        }
        if state.can_decode()
            && let Some(proof) = &self.integrity
        {
@@ -2231,10 +2247,10 @@ impl Erasure {
        W: AsyncWrite + Send + Sync + Unpin,
        R: crate::erasure::coding::ShardSource,
    {
-        let (written, error, _) = self
+        let outcome = self
            .decode_inner(writer, readers, offset, length, total_length, None, Vec::new(), Vec::new())
            .await;
-        (written, error)
+        (outcome.written, outcome.error)
    }

    #[allow(dead_code, reason = "read-cost decode path asserted by this file's tests (backlog#1823)")]
@@ -2251,10 +2267,10 @@ impl Erasure {
        W: AsyncWrite + Send + Sync + Unpin,
        R: crate::erasure::coding::ShardSource,
    {
-        let (written, error, _) = self
+        let outcome = self
            .decode_inner(writer, readers, offset, length, total_length, Some(read_costs), Vec::new(), Vec::new())
            .await;
-        (written, error)
+        (outcome.written, outcome.error)
    }

    /// GET decode entry point that also carries the deferred-parity stripe
@@ -2311,7 +2327,7 @@ impl Erasure {
        W: AsyncWrite + Send + Sync + Unpin,
        R: crate::erasure::coding::ShardSource,
    {
-        let (written, error, _) = self
+        let outcome = self
            .decode_inner(
                writer,
                readers,
@@ -2323,7 +2339,7 @@ impl Erasure {
                deferred_reopeners,
            )
            .await;
-        (written, error)
+        (outcome.written, outcome.error)
    }

    #[allow(clippy::too_many_arguments)]
@@ -2374,6 +2390,7 @@ impl Erasure {
        block_length: usize,
        written: &mut usize,
        ret_err: &mut Option<std::io::Error>,
+        repair: &ShardReadRepair,
        stage_metrics_enabled: bool,
        require_surplus_source: bool,
        integrity: Option<&std::sync::Arc<crate::io_support::shard_integrity::PartProofReader>>,
@@ -2382,6 +2399,7 @@ impl Erasure {
    where
        W: AsyncWrite + Send + Sync + Unpin,
    {
+        repair.observe(errs);
        if ret_err.is_none()
            && let (_, Some(err)) = reduce_errs(errs, &[])
            && (err == Error::FileNotFound || err == Error::FileCorrupt)
@@ -2516,41 +2534,42 @@ impl Erasure {
    {
        if readers.len() != self.data_shards + self.parity_shards {
            record_get_object_pipeline_failure(GET_STAGE_RANGE, GetObjectFailureReason::RangeOrLengthInvalid);
-            return (0, Some(io::Error::new(ErrorKind::InvalidInput, "Invalid number of readers")), false);
+            return DecodeOutcome {
+                error: Some(io::Error::new(ErrorKind::InvalidInput, "Invalid number of readers")),
+                ..Default::default()
+            };
        }

        // block_size/data_shards come from on-disk metadata; a corrupt FileInfo with a
        // zero here must surface as an error, not a divide-by-zero panic on every GET.
        if self.block_size == 0 || self.data_shards == 0 {
            record_get_object_pipeline_failure(GET_STAGE_RANGE, GetObjectFailureReason::RangeOrLengthInvalid);
-            return (
-                0,
-                Some(io::Error::new(ErrorKind::InvalidInput, "Invalid erasure coding parameters")),
-                false,
-            );
+            return DecodeOutcome {
+                error: Some(io::Error::new(ErrorKind::InvalidInput, "Invalid erasure coding parameters")),
+                ..Default::default()
+            };
        }

        let Some(end_offset) = offset.checked_add(length) else {
            record_get_object_pipeline_failure(GET_STAGE_RANGE, GetObjectFailureReason::RangeOrLengthInvalid);
-            return (
-                0,
-                Some(io::Error::new(ErrorKind::InvalidInput, "offset + length exceeds total length")),
-                false,
-            );
+            return DecodeOutcome {
+                error: Some(io::Error::new(ErrorKind::InvalidInput, "offset + length exceeds total length")),
+                ..Default::default()
+            };
        };
        if end_offset > total_length {
            record_get_object_pipeline_failure(GET_STAGE_RANGE, GetObjectFailureReason::RangeOrLengthInvalid);
-            return (
-                0,
-                Some(io::Error::new(ErrorKind::InvalidInput, "offset + length exceeds total length")),
-                false,
-            );
+            return DecodeOutcome {
+                error: Some(io::Error::new(ErrorKind::InvalidInput, "offset + length exceeds total length")),
+                ..Default::default()
+            };
        }

        let mut ret_err = None;
+        let repair = ShardReadRepair::default();

        if length == 0 {
-            return (0, ret_err, false);
+            return DecodeOutcome::default();
        }

        let mut written = 0;
@@ -2573,6 +2592,11 @@ impl Erasure {
        .with_deferred_parity_reopeners(deferred_reopeners);

        let integrity = reader.integrity.clone();
+        // Copy sources use demand-bound scheduling for backpressure, but retain
+        // the ordinary GET read quorum. Only the opt-in GET rollout requires a
+        // surplus source; available surplus shards and integrity proofs are
+        // still verified before any reconstructed copy bytes are emitted.
+        let require_surplus_source = reader.demand_bound_lockstep && matches!(decode_read_policy(), DecodeReadPolicy::Default);

        let start = offset / self.block_size;
        let end = end_offset.saturating_sub(1) / self.block_size;
@@ -2667,7 +2691,6 @@ impl Erasure {
                        // `shards` are borrowed again below. In the `Stop` case that
                        // drop is what cancels the still-in-flight prefetch read.
                        let (flow, next): (Option<StripeFlow>, Option<StripeReadOutput>) = {
-                            let require_surplus_source = reader.demand_bound_lockstep;
                            let read_fut = read_stripe_timed(&mut reader, stage_metrics_enabled);
                            let emit_fut = self.emit_decoded_stripe(
                                writer,
@@ -2677,6 +2700,7 @@ impl Erasure {
                                block_length,
                                &mut written,
                                &mut ret_err,
+                                &repair,
                                stage_metrics_enabled,
                                require_surplus_source,
                                integrity.as_ref(),
@@ -2727,8 +2751,9 @@ impl Erasure {
                                block_length,
                                &mut written,
                                &mut ret_err,
+                                &repair,
                                stage_metrics_enabled,
-                                reader.demand_bound_lockstep,
+                                require_surplus_source,
                                integrity.as_ref(),
                                offset,
                            )
@@ -2770,8 +2795,9 @@ impl Erasure {
                        block_length,
                        &mut written,
                        &mut ret_err,
+                        &repair,
                        stage_metrics_enabled,
-                        reader.demand_bound_lockstep,
+                        require_surplus_source,
                        integrity.as_ref(),
                        offset,
                    )
@@ -2785,15 +2811,16 @@ impl Erasure {
            }
        }

-        if ret_err.is_some() {
-            return (written, ret_err, exact_quorum);
-        }
-
-        if written < length {
+        if ret_err.is_none() && written < length {
            ret_err = Some(Error::LessData.into());
        }

-        (written, ret_err, exact_quorum)
+        DecodeOutcome {
+            written,
+            error: ret_err,
+            exact_quorum,
+            repair,
+        }
    }
 }

@@ -3449,6 +3476,150 @@ mod tests {
        }
    }

+    async fn copy_source_bitrot_shards(erasure: &Erasure, payload: &[u8], corrupt_parity: bool) -> Vec<Vec<u8>> {
+        let mut writers: Vec<_> = (0..erasure.data_shards + erasure.parity_shards)
+            .map(|_| BitrotWriter::new(Cursor::new(Vec::new()), erasure.shard_size(), HashAlgorithm::HighwayHash256))
+            .collect();
+        for block in payload.chunks(erasure.block_size) {
+            let shards = erasure.encode_data(block).expect("copy source should encode");
+            for (index, shard) in shards.iter().enumerate() {
+                let mut bytes = shard.to_vec();
+                if corrupt_parity && index == erasure.data_shards {
+                    // A stale source can have a valid per-shard bitrot checksum.
+                    // Reframe the changed parity to exercise reconstruction verification.
+                    bytes[0] ^= 0x80;
+                }
+                writers[index].write(&bytes).await.expect("copy source should frame");
+            }
+        }
+        writers.into_iter().map(|writer| writer.into_inner().into_inner()).collect()
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn copy_source_reads_exact_quorum_across_stripes_and_ranges() {
+        const BLOCK_SIZE: usize = 768;
+        let payload: Vec<_> = (0..BLOCK_SIZE * 3 + 37)
+            .map(|index| (index.wrapping_mul(31) % 251) as u8)
+            .collect();
+
+        for uses_legacy in [false, true] {
+            let erasure = Erasure::new_with_options(12, 4, BLOCK_SIZE, uses_legacy);
+            let buffers = copy_source_bitrot_shards(&erasure, &payload, false).await;
+            for offline in [[0, 1, 2, 3], [0, 5, 12, 15]] {
+                for (offset, length) in [(0, payload.len()), (BLOCK_SIZE + 17, payload.len() - BLOCK_SIZE - 17)] {
+                    let frame_offset = offset / BLOCK_SIZE * (erasure.shard_size() + HashAlgorithm::HighwayHash256.size());
+                    let positioned: Vec<_> = buffers.iter().map(|bytes| bytes[frame_offset..].to_vec()).collect();
+                    let (mut readers, handles) =
+                        readers_with_deferred_parity(&positioned, 12, erasure.shard_size(), &HashAlgorithm::HighwayHash256, &[]);
+                    for index in offline {
+                        readers[index] = None;
+                    }
+                    let mut output = Vec::new();
+                    let (written, error) = crate::set_disk::with_get_object_read_policy(
+                        crate::set_disk::GetObjectReadPolicy::CopySource,
+                        erasure.decode_with_stripe_handles(&mut output, readers, offset, length, payload.len(), None, handles),
+                    )
+                    .await;
+
+                    assert!(error.is_none(), "legacy={uses_legacy}, offline={offline:?}, offset={offset}: {error:?}");
+                    assert_eq!(written, length);
+                    assert_eq!(output, payload[offset..offset + length]);
+                }
+            }
+        }
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn copy_source_recovers_exact_quorum_after_first_stripe() {
+        const BLOCK_SIZE: usize = 768;
+        let payload: Vec<_> = (0..BLOCK_SIZE * 3 + 37)
+            .map(|index| (index.wrapping_mul(31) % 251) as u8)
+            .collect();
+        for uses_legacy in [false, true] {
+            let erasure = Erasure::new_with_options(12, 4, BLOCK_SIZE, uses_legacy);
+            let buffers = copy_source_bitrot_shards(&erasure, &payload, false).await;
+            let first_frame = erasure.shard_size() + HashAlgorithm::HighwayHash256.size();
+            let (readers, handles) = readers_with_deferred_parity(
+                &buffers,
+                12,
+                erasure.shard_size(),
+                &HashAlgorithm::HighwayHash256,
+                &[(0, first_frame), (1, first_frame), (2, first_frame), (3, first_frame)],
+            );
+            let mut output = Vec::new();
+            let (written, error) = crate::set_disk::with_get_object_read_policy(
+                crate::set_disk::GetObjectReadPolicy::CopySource,
+                erasure.decode_with_stripe_handles(&mut output, readers, 0, payload.len(), payload.len(), None, handles),
+            )
+            .await;
+
+            assert!(error.is_none(), "legacy={uses_legacy}: late shard loss should recover: {error:?}");
+            assert_eq!(written, payload.len());
+            assert_eq!(output, payload);
+        }
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn copy_source_below_read_quorum_fails_before_output() {
+        const BLOCK_SIZE: usize = 768;
+        let payload = vec![0x5a; BLOCK_SIZE * 2 + 37];
+        for uses_legacy in [false, true] {
+            let erasure = Erasure::new_with_options(12, 4, BLOCK_SIZE, uses_legacy);
+            let buffers = copy_source_bitrot_shards(&erasure, &payload, false).await;
+            let (mut readers, handles) =
+                readers_with_deferred_parity(&buffers, 12, erasure.shard_size(), &HashAlgorithm::HighwayHash256, &[]);
+            for reader in readers.iter_mut().take(5) {
+                *reader = None;
+            }
+            let mut output = Vec::new();
+            let (written, error) = crate::set_disk::with_get_object_read_policy(
+                crate::set_disk::GetObjectReadPolicy::CopySource,
+                erasure.decode_with_stripe_handles(&mut output, readers, 0, payload.len(), payload.len(), None, handles),
+            )
+            .await;
+
+            assert_eq!(written, 0);
+            assert!(output.is_empty());
+            assert_eq!(
+                error
+                    .expect("eleven shards cannot reconstruct twelve data shards")
+                    .to_string(),
+                Error::ErasureReadQuorum.to_string()
+            );
+        }
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn copy_source_rejects_corrupt_surplus_parity_before_output() {
+        const BLOCK_SIZE: usize = 768;
+        let payload = vec![0x5a; BLOCK_SIZE * 2 + 37];
+        for uses_legacy in [false, true] {
+            let erasure = Erasure::new_with_options(12, 4, BLOCK_SIZE, uses_legacy);
+            let buffers = copy_source_bitrot_shards(&erasure, &payload, true).await;
+            let (mut readers, handles) =
+                readers_with_deferred_parity(&buffers, 12, erasure.shard_size(), &HashAlgorithm::HighwayHash256, &[]);
+            for reader in readers.iter_mut().take(3) {
+                *reader = None;
+            }
+            let mut output = Vec::new();
+            let (written, error) = crate::set_disk::with_get_object_read_policy(
+                crate::set_disk::GetObjectReadPolicy::CopySource,
+                erasure.decode_with_stripe_handles(&mut output, readers, 0, payload.len(), payload.len(), None, handles),
+            )
+            .await;
+
+            assert_eq!(written, 0);
+            assert!(output.is_empty());
+            let error = error.expect("copy must reject inconsistent surplus parity despite valid bitrot framing");
+            assert_eq!(error.kind(), ErrorKind::InvalidData);
+            assert!(error.to_string().contains("inconsistent read source shards"));
+        }
+    }
+
    /// Guards the shard-buffer reuse in `ParallelReader`: a multi-stripe
    /// `Erasure::decode` that reconstructs missing data shards on every stripe
    /// must still return byte-exact output. Reconstructed and parity buffers are
@@ -3834,10 +4005,45 @@ mod tests {
            temp_env::async_with_vars(vars, async {
                let readers = build_prefetch_readers(&shard_bufs, shard_size, &hash_algo, BLOCK_SIZE, 0, &[], &[1]);
                let mut output = Vec::new();
-                let (written, err) = erasure.decode(&mut output, readers, 0, total_len, total_len).await;
-                assert!(err.is_none(), "{label}: corrupt-shard read errored {err:?}");
-                assert_eq!(written, total_len, "{label}: corrupt-shard short write");
+                let outcome = erasure
+                    .decode_with_stripe_handles_and_reopeners_with_diagnostics(
+                        &mut output,
+                        readers,
+                        0,
+                        total_len,
+                        total_len,
+                        None,
+                        Vec::new(),
+                        Vec::new(),
+                    )
+                    .await;
+                assert!(outcome.error.is_none(), "{label}: corrupt-shard read errored {:?}", outcome.error);
+                assert_eq!(outcome.written, total_len, "{label}: corrupt-shard short write");
                assert_eq!(output, total_data, "{label}: corrupt bytes leaked into output");
+                assert!(
+                    outcome.repair.needed(false),
+                    "{label}: minority corruption must survive successful reconstruction"
+                );
+                let readers = build_prefetch_readers(&shard_bufs, shard_size, &hash_algo, BLOCK_SIZE, 0, &[], &[]);
+                let mut healthy_output = Vec::new();
+                let healthy = erasure
+                    .decode_with_stripe_handles_and_reopeners_with_diagnostics(
+                        &mut healthy_output,
+                        readers,
+                        0,
+                        total_len,
+                        total_len,
+                        None,
+                        Vec::new(),
+                        Vec::new(),
+                    )
+                    .await;
+                assert!(healthy.error.is_none());
+                assert_eq!(healthy_output, total_data);
+                assert!(
+                    !healthy.repair.needed(true),
+                    "{label}: healthy/deferred shards cannot create repair evidence"
+                );
            })
            .await;
        }
@@ -4635,40 +4841,38 @@ mod tests {
                (ENV_RUSTFS_GET_SHARD_LOCALITY_PREFERENCE_ENABLE, None::<&str>),
            ],
            async {
-            const NUM_SHARDS: usize = 1;
-            const BLOCK_SIZE: usize = 64;
-            const DATA_SHARDS: usize = 4;
-            const PARITY_SHARDS: usize = 2;
-            const SHARD_SIZE: usize = BLOCK_SIZE / DATA_SHARDS;
+                const NUM_SHARDS: usize = 1;
+                const BLOCK_SIZE: usize = 64;
+                const DATA_SHARDS: usize = 4;
+                const PARITY_SHARDS: usize = 2;
+                const SHARD_SIZE: usize = BLOCK_SIZE / DATA_SHARDS;

-            let hash_algo = HashAlgorithm::HighwayHash256;
-            let readers = make_test_readers(DATA_SHARDS + PARITY_SHARDS, SHARD_SIZE, NUM_SHARDS, &hash_algo, &[], &[1]).await;
-            let read_costs = vec![
-                ShardReadCost::Local,
-                ShardReadCost::Local,
-                ShardReadCost::Local,
-                ShardReadCost::Local,
-                ShardReadCost::Remote,
-                ShardReadCost::Remote,
-            ];
-            let erasure = Erasure::new(DATA_SHARDS, PARITY_SHARDS, BLOCK_SIZE);
-            let mut parallel_reader = ParallelReader::new_with_metrics_path_and_read_costs(
-                readers,
-                erasure,
-                0,
-                NUM_SHARDS * BLOCK_SIZE,
-                None,
-                read_costs,
-            );
+                let hash_algo = HashAlgorithm::HighwayHash256;
+                let readers = make_test_readers(DATA_SHARDS + PARITY_SHARDS, SHARD_SIZE, NUM_SHARDS, &hash_algo, &[], &[1]).await;
+                let read_costs = vec![
+                    ShardReadCost::Local,
+                    ShardReadCost::Local,
+                    ShardReadCost::Local,
+                    ShardReadCost::Local,
+                    ShardReadCost::Remote,
+                    ShardReadCost::Remote,
+                ];
+                let erasure = Erasure::new(DATA_SHARDS, PARITY_SHARDS, BLOCK_SIZE);
+                let mut parallel_reader = ParallelReader::new_with_metrics_path_and_read_costs(
+                    readers,
+                    erasure,
+                    0,
+                    NUM_SHARDS * BLOCK_SIZE,
+                    None,
+                    read_costs,
+                );

-            let (bufs, errs) = parallel_reader.read().await;
+                let (bufs, errs) = parallel_reader.read().await;

-            assert_eq!(DATA_SHARDS, bufs.iter().filter(|buf| buf.is_some()).count());
-            assert!(
-                matches!(&errs[1], Some(DiskError::Io(err)) if err.kind() == ErrorKind::InvalidData && err.to_string().contains("bitrot"))
-            );
-            assert_eq!(bufs[4].as_deref(), Some(&[4u8; SHARD_SIZE][..]));
-            assert!(bufs[5].is_none());
+                assert_eq!(DATA_SHARDS, bufs.iter().filter(|buf| buf.is_some()).count());
+                assert_eq!(errs[1], Some(DiskError::FileCorrupt));
+                assert_eq!(bufs[4].as_deref(), Some(&[4u8; SHARD_SIZE][..]));
+                assert!(bufs[5].is_none());
            },
        )
        .await;
@@ -4814,16 +5018,7 @@ mod tests {
            assert_eq!(DATA_SHARDS, bufs.iter().filter(|buf| buf.is_some()).count());
            assert_eq!(
                BITROT_DISKS,
-                errs.iter()
-                    .filter(|err| {
-                        match err {
-                            Some(DiskError::Io(err)) => {
-                                err.kind() == std::io::ErrorKind::InvalidData && err.to_string().contains("bitrot")
-                            }
-                            _ => false,
-                        }
-                    })
-                    .count()
+                errs.iter().filter(|err| matches!(err, Some(DiskError::FileCorrupt))).count()
            );
        }
    }
@@ -587,7 +587,9 @@ pub(crate) async fn verify_deep_parts(
            return Err(DiskError::FileCorrupt);
        }
        let part_status = statuses.get_mut(&part_index).ok_or(DiskError::FileCorrupt)?;
-        let length = erasure.shard_file_offset(0, part.size, part.size);
+        // Deep verification must consume every encoded stripe, including a
+        // partial final stripe, before certifying the shard's integrity.
+        let length = usize::try_from(erasure.shard_file_size(part.size as i64)).map_err(|_| DiskError::FileCorrupt)?;
        let mut readers = Vec::with_capacity(disks.len());
        for (index, disk) in disks.iter().enumerate() {
            if part_status[index] != CHECK_PART_UNKNOWN {
@@ -631,6 +633,9 @@ pub(crate) async fn verify_deep_parts(
                    part_status[index] = CHECK_PART_FILE_CORRUPT;
                }
                Err(error) => {
+                    if !matches!(error, DiskError::FileNotFound | DiskError::FileVersionNotFound | DiskError::FileCorrupt) {
+                        return Err(error);
+                    }
                    readers.push(None);
                    part_status[index] = conv_part_err_to_int(&Some(error));
                }
@@ -935,6 +935,8 @@ pub struct ObjectOptions {
    /// Persisted bucket incarnation observed before authorization.
    pub expected_bucket_incarnation_id: Option<Uuid>,
    pub no_lock: bool,
+    /// Internal read-only inspection must not enqueue metadata or payload repairs.
+    pub suppress_read_repair: bool,
    /// Control-plane writers that immediately read or CAS the same namespace
    /// key use TailDrained without changing namespace lock ownership.
    #[doc(hidden)]
@@ -0,0 +1,290 @@
+// Copyright 2024 RustFS Team
+// Licensed under the Apache License, Version 2.0.
+
+//! Explicit, bounded integrity operations. This service does not participate in ordinary GET or Heal.
+//! Lock order: cluster worker lock, bucket incarnation fence, object commit lock.
+//! Control requests use checkpoint CAS without taking the worker lock; a worker that loses CAS stops.
+
+mod model;
+mod runner;
+#[cfg(test)]
+mod tests;
+
+pub use model::{
+    InventoryItem, InventoryPage, ItemRequest, ItemResult, ItemState, Job, JobMode, JobRequest, JobState, Protection, Readiness,
+    readiness,
+};
+
+use crate::config::com::{read_config_limited_preserve_empty_with_metadata, save_config_with_opts_quiet};
+use crate::disk::RUSTFS_META_BUCKET;
+use crate::error::Error as StorageError;
+use crate::object_api::{ObjectOptions, WriteCompletion};
+use crate::storage_api_contracts::{
+    list::ListOperations,
+    namespace::NamespaceLocking,
+    object::{HTTPPreconditions, ObjectOperations},
+};
+use crate::store::ECStore;
+use model::InventoryItem as Item;
+use std::sync::Arc;
+use std::time::Duration;
+use uuid::Uuid;
+
+type Result<T> = std::result::Result<T, IntegrityError>;
+
+#[derive(Debug, thiserror::Error)]
+pub enum IntegrityError {
+    #[error("invalid integrity request: {0}")]
+    Invalid(&'static str),
+    #[error("integrity job or target changed concurrently")]
+    Conflict,
+    #[error("an integrity worker is already active")]
+    Busy,
+    #[error("integrity job not found")]
+    NotFound,
+    #[error("integrity operation was stopped")]
+    Stopped,
+    #[error("protected writes are not enabled on this coordinator")]
+    NotActivated,
+    #[error("unsupported migration bucket configuration")]
+    UnsupportedBucket,
+    #[error("integrity storage operation failed: {0}")]
+    Storage(#[from] StorageError),
+    #[error("integrity stream operation failed: {0}")]
+    Io(#[from] std::io::Error),
+    #[error("integrity checkpoint is malformed: {0}")]
+    Json(#[from] serde_json::Error),
+}
+
+fn check_bucket(bucket: &str) -> Result<()> {
+    crate::bucket::utils::check_valid_bucket_name_strict(bucket).map_err(|_| IntegrityError::Invalid("invalid source bucket"))
+}
+
+fn path(bucket: &str, id: Uuid) -> String {
+    format!("buckets/{bucket}/integrity-jobs/{id}.json")
+}
+
+pub async fn inventory(
+    api: Arc<ECStore>,
+    bucket: &str,
+    prefix: &str,
+    key_marker: Option<String>,
+    version_marker: Option<String>,
+    limit: i32,
+) -> Result<InventoryPage> {
+    check_bucket(bucket)?;
+    if !(1..=100).contains(&limit) || prefix.len() > 1024 || (version_marker.is_some() && key_marker.is_none()) {
+        return Err(IntegrityError::Invalid("invalid inventory page"));
+    }
+    let incarnation = api.bucket_incarnation_id_from_disk(bucket).await?;
+    let page = api
+        .clone()
+        .list_object_versions(bucket, prefix, key_marker.clone(), version_marker.clone(), None, limit)
+        .await?;
+    if page.is_truncated && (page.next_marker.clone(), page.next_version_idmarker.clone()) == (key_marker, version_marker) {
+        return Err(IntegrityError::Invalid("inventory cursor did not advance"));
+    }
+    let mut items = Vec::with_capacity(page.objects.len());
+    for listed in page.objects {
+        if listed.delete_marker {
+            items.push(Item::from_info(&listed));
+            continue;
+        }
+        let opts = ObjectOptions {
+            version_id: Some(listed.version_id.unwrap_or_else(Uuid::nil).to_string()),
+            include_part_checksums: true,
+            suppress_read_repair: true,
+            ..Default::default()
+        };
+        match api.get_object_info(bucket, &listed.name, &opts).await {
+            Ok(info) => items.push(Item::from_info(&info)),
+            Err(_) => {
+                let mut item = Item::from_info(&listed);
+                item.observation_error = Some("metadata_unavailable_or_changed".into());
+                item.protection = Protection::Unknown;
+                item.audit_unsupported = Some("metadata_unavailable_or_changed".into());
+                items.push(item);
+            }
+        }
+    }
+    if api.bucket_incarnation_id_from_disk(bucket).await? != incarnation {
+        return Err(IntegrityError::Conflict);
+    }
+    Ok(InventoryPage {
+        observed_at: time::OffsetDateTime::now_utc().to_string(),
+        bucket_incarnation: incarnation,
+        items,
+        is_truncated: page.is_truncated,
+        next_key_marker: page.next_marker,
+        next_version_marker: page.next_version_idmarker,
+    })
+}
+
+pub async fn create_job(api: Arc<ECStore>, bucket: &str, request: JobRequest) -> Result<Job> {
+    check_bucket(bucket)?;
+    request.validate()?;
+    if request.mode == JobMode::Migrate && !readiness().new_writes_enabled_here {
+        return Err(IntegrityError::NotActivated);
+    }
+    let incarnation = api.bucket_incarnation_id_from_disk(bucket).await?;
+    let job = Job {
+        format_version: 1,
+        id: Uuid::new_v4(),
+        bucket: bucket.into(),
+        bucket_incarnation: incarnation,
+        results: vec![ItemResult::default(); request.items.len()],
+        request,
+        state: JobState::Paused,
+        revision: 0,
+        last_error: None,
+    };
+    save(&api, &job, None, None).await?;
+    Ok(job)
+}
+
+pub async fn get_job(api: Arc<ECStore>, bucket: &str, id: Uuid) -> Result<Job> {
+    Ok(load(&api, bucket, id).await?.0)
+}
+
+async fn load(api: &Arc<ECStore>, bucket: &str, id: Uuid) -> Result<(Job, String)> {
+    check_bucket(bucket)?;
+    let (bytes, info) = match read_config_limited_preserve_empty_with_metadata(api.clone(), &path(bucket, id), 1024 * 1024).await
+    {
+        Ok(value) => value,
+        Err(StorageError::ConfigNotFound | StorageError::FileNotFound) => return Err(IntegrityError::NotFound),
+        Err(error) => return Err(error.into()),
+    };
+    let job: Job = serde_json::from_slice(&bytes)?;
+    job.request.validate()?;
+    if job.format_version != 1
+        || job.bucket != bucket
+        || job.id != id
+        || job.results.len() != job.request.items.len()
+        || api.bucket_incarnation_id_from_disk(bucket).await? != job.bucket_incarnation
+    {
+        return Err(IntegrityError::Conflict);
+    }
+    Ok((job, info.etag.ok_or(IntegrityError::Conflict)?))
+}
+
+async fn save(
+    api: &Arc<ECStore>,
+    job: &Job,
+    etag: Option<&str>,
+    worker: Option<&rustfs_lock::NamespaceLockGuard>,
+) -> Result<String> {
+    let fence = api
+        .acquire_bucket_incarnation_fence(&job.bucket, job.bucket_incarnation)
+        .await?;
+    let mut opts = ObjectOptions {
+        max_parity: true,
+        write_completion: WriteCompletion::TailDrained,
+        http_preconditions: Some(match etag {
+            Some(value) => HTTPPreconditions {
+                if_match: Some(value.into()),
+                ..Default::default()
+            },
+            None => HTTPPreconditions {
+                if_none_match: Some("*".into()),
+                ..Default::default()
+            },
+        }),
+        ..Default::default()
+    };
+    fence.attach_to_object_options(&mut opts);
+    if let Some(guard) = worker {
+        opts.add_namespace_lock_guard(guard);
+    }
+    let api = api.clone();
+    let job = job.clone();
+    // Keep the lifecycle guard through the drained commit even if the HTTP waiter is cancelled.
+    tokio::spawn(async move {
+        save_config_with_opts_quiet(api.clone(), &path(&job.bucket, job.id), serde_json::to_vec(&job)?, &opts).await?;
+        let (stored, next) = load(&api, &job.bucket, job.id).await?;
+        if stored.revision != job.revision {
+            return Err(IntegrityError::Conflict);
+        }
+        drop(fence);
+        Ok(next)
+    })
+    .await
+    .map_err(|_| IntegrityError::Invalid("checkpoint commit task failed"))?
+}
+
+/// Pause/cancel acknowledge intent. An in-flight conditional publication is allowed to drain.
+pub async fn control_job(api: Arc<ECStore>, bucket: &str, id: Uuid, cancel: bool) -> Result<Job> {
+    let (mut job, etag) = load(&api, bucket, id).await?;
+    match job.state {
+        JobState::Complete | JobState::Cancelled | JobState::CancelRequested => return Ok(job),
+        JobState::Paused | JobState::Failed => job.state = if cancel { JobState::Cancelled } else { JobState::Paused },
+        _ => {
+            job.state = if cancel {
+                JobState::CancelRequested
+            } else {
+                JobState::PauseRequested
+            }
+        }
+    }
+    job.revision = job.revision.checked_add(1).ok_or(IntegrityError::Conflict)?;
+    save(&api, &job, Some(&etag), None).await?;
+    Ok(job)
+}
+
+/// Resume is explicit after restart. The distributed lock excludes another worker, including another node.
+pub async fn resume_job(api: Arc<ECStore>, bucket: &str, id: Uuid) -> Result<Job> {
+    check_bucket(bucket)?;
+    let lock = api.new_ns_lock(RUSTFS_META_BUCKET, "integrity-jobs/worker.lock").await?;
+    let guard = lock
+        .get_write_lock_quiet(Duration::from_secs(1))
+        .await
+        .map_err(|_| IntegrityError::Busy)?;
+    let (mut job, etag) = load(&api, bucket, id).await?;
+    if matches!(job.state, JobState::Complete | JobState::Cancelled) {
+        return Ok(job);
+    }
+    if job.state == JobState::CancelRequested {
+        job.state = JobState::Cancelled;
+        job.revision = job.revision.checked_add(1).ok_or(IntegrityError::Conflict)?;
+        save(&api, &job, Some(&etag), Some(&guard)).await?;
+        return Ok(job);
+    }
+    if job.request.mode == JobMode::Migrate && !readiness().new_writes_enabled_here {
+        return Err(IntegrityError::NotActivated);
+    }
+    job.state = JobState::Running;
+    job.last_error = None;
+    job.revision = job.revision.checked_add(1).ok_or(IntegrityError::Conflict)?;
+    let etag = save(&api, &job, Some(&etag), Some(&guard)).await?;
+    let response = job.clone();
+    tokio::spawn(async move {
+        let result = runner::run(&api, &mut job, etag, &guard).await;
+        if let Err(error) = result {
+            // A failed write/CAS stops publication. Never overwrite a concurrent control request.
+            if let Ok((mut latest, etag)) = load(&api, &job.bucket, job.id).await {
+                match latest.state {
+                    JobState::PauseRequested => latest.state = JobState::Paused,
+                    JobState::CancelRequested => latest.state = JobState::Cancelled,
+                    JobState::Running if latest.revision <= job.revision => {
+                        latest.state = JobState::Failed;
+                        latest.last_error = Some(
+                            match error {
+                                IntegrityError::Conflict => "concurrent_change",
+                                IntegrityError::Stopped => "worker_stopped",
+                                IntegrityError::UnsupportedBucket => "unsupported_destination_bucket_configuration",
+                                IntegrityError::NotActivated => "protected_writes_not_activated",
+                                _ => "operation_failed_retry_requires_resume",
+                            }
+                            .into(),
+                        );
+                    }
+                    _ => return,
+                }
+                if let Some(revision) = latest.revision.checked_add(1) {
+                    latest.revision = revision;
+                    let _ = save(&api, &latest, Some(&etag), Some(&guard)).await;
+                }
+            }
+        }
+    });
+    Ok(response)
+}
@@ -0,0 +1,343 @@
+// Copyright 2024 RustFS Team
+// Licensed under the Apache License, Version 2.0.
+
+use super::{IntegrityError, Result};
+use crate::object_api::ObjectInfo;
+use rustfs_rio::{Checksum, read_checksums};
+use serde::{Deserialize, Serialize};
+use sha2::{Digest, Sha256};
+use std::collections::{BTreeMap, HashSet};
+use uuid::Uuid;
+
+pub const MAX_ITEMS: usize = 64;
+pub const MAX_OBJECT_BYTES: u64 = 5 * 1024 * 1024 * 1024;
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct ItemRequest {
+    pub key: String,
+    pub version_id: Option<String>,
+    /// Base64 SHA-256 supplied by the administrator from an independent source.
+    pub expected_sha256: Option<String>,
+    pub target_key: Option<String>,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum JobMode {
+    Audit,
+    Migrate,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct JobRequest {
+    pub mode: JobMode,
+    pub items: Vec<ItemRequest>,
+    /// Bounds each source read, destination PUT, and verification stream; one worker runs per cluster.
+    pub bytes_per_second: u64,
+    pub max_object_bytes: u64,
+}
+
+impl JobRequest {
+    pub(crate) fn validate(&self) -> Result<()> {
+        if self.items.is_empty() || self.items.len() > MAX_ITEMS {
+            return Err(IntegrityError::Invalid("a job requires 1..=64 explicit source versions"));
+        }
+        if !(64 * 1024..=1024 * 1024 * 1024).contains(&self.bytes_per_second)
+            || self.max_object_bytes == 0
+            || self.max_object_bytes > MAX_OBJECT_BYTES
+        {
+            return Err(IntegrityError::Invalid("invalid byte rate or staging limit"));
+        }
+        let mut targets = HashSet::new();
+        let sources: HashSet<_> = self.items.iter().map(|i| i.key.as_str()).collect();
+        for item in &self.items {
+            validate_key(&item.key)?;
+            if let Some(version) = &item.version_id
+                && version != "null"
+                && Uuid::parse_str(version).ok().is_none_or(|id| id.is_nil())
+            {
+                return Err(IntegrityError::Invalid("invalid source version"));
+            }
+            if let Some(value) = &item.expected_sha256 {
+                sha256_bytes(value)?;
+            }
+            match (self.mode, &item.target_key) {
+                (JobMode::Audit, None) => {}
+                (JobMode::Migrate, Some(target)) => {
+                    validate_key(target)?;
+                    if sources.contains(target.as_str()) || !targets.insert(target) {
+                        return Err(IntegrityError::Invalid("targets must be distinct from every source and target"));
+                    }
+                }
+                _ => return Err(IntegrityError::Invalid("only migrations require an explicit target key")),
+            }
+        }
+        Ok(())
+    }
+}
+
+fn validate_key(key: &str) -> Result<()> {
+    if key.is_empty() || key.len() > 1024 || key.ends_with('/') || key.contains('\0') {
+        return Err(IntegrityError::Invalid("invalid object key"));
+    }
+    Ok(())
+}
+
+pub(crate) fn sha256_bytes(value: &str) -> Result<Vec<u8>> {
+    let raw = base64_simd::STANDARD
+        .decode_to_vec(value)
+        .map_err(|_| IntegrityError::Invalid("SHA256 must be canonical Base64"))?;
+    if raw.len() != 32 || base64_simd::STANDARD.encode_to_string(&raw) != value {
+        return Err(IntegrityError::Invalid("SHA256 must be canonical Base64 of 32 bytes"));
+    }
+    Ok(raw)
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum Protection {
+    Unknown,
+    Legacy,
+    IndependentCommitment,
+    InvalidDeclaration,
+    NoPayload,
+}
+
+pub(crate) fn protection(info: &ObjectInfo) -> Protection {
+    if info.delete_marker {
+        return Protection::NoPayload;
+    }
+    match rustfs_filemeta::shard_integrity::descriptor_from_metadata(&info.user_defined) {
+        Ok(None) if info.parts.iter().all(|p| p.integrity.is_none()) => Protection::Legacy,
+        Ok(Some(parts))
+            if parts.len() == info.parts.len()
+                && !parts.is_empty()
+                && parts
+                    .iter()
+                    .zip(info.parts.iter())
+                    .all(|(declared, part)| part.integrity.as_ref() == Some(declared)) =>
+        {
+            Protection::IndependentCommitment
+        }
+        _ => Protection::InvalidDeclaration,
+    }
+}
+
+/// Checksums shown in inventory are not automatically accepted as migration evidence.
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct InventoryItem {
+    pub key: String,
+    /// S3 selector, including the literal "null" for the null version slot.
+    pub version_id: String,
+    pub data_dir: Option<Uuid>,
+    pub size: i64,
+    pub protection: Protection,
+    pub checksums: BTreeMap<String, String>,
+    pub checksum_is_multipart: bool,
+    pub audit_unsupported: Option<String>,
+    pub source_fingerprint: String,
+    pub observation_error: Option<String>,
+}
+
+impl InventoryItem {
+    pub(crate) fn from_info(info: &ObjectInfo) -> Self {
+        let (checksums, checksum_is_multipart) = info.checksum.as_ref().map_or_else(Default::default, |b| read_checksums(b, 0));
+        Self {
+            key: info.name.clone(),
+            version_id: info
+                .version_id
+                .filter(|id| !id.is_nil())
+                .map_or_else(|| "null".into(), |id| id.to_string()),
+            data_dir: info.data_dir,
+            size: info.size,
+            protection: protection(info),
+            checksums: checksums.into_iter().collect(),
+            checksum_is_multipart,
+            audit_unsupported: unsupported(info).map(str::to_owned),
+            source_fingerprint: fingerprint(info),
+            observation_error: None,
+        }
+    }
+}
+
+pub(crate) fn unsupported(info: &ObjectInfo) -> Option<&'static str> {
+    if info.delete_marker {
+        Some("delete_marker")
+    } else if !info.transitioned_object.status.is_empty() {
+        Some("transitioned_object")
+    } else if info.is_encrypted() || info.is_compressed() {
+        Some("transformed_object")
+    } else if info.is_multipart() {
+        Some("multipart_object")
+    } else if protection(info) == Protection::InvalidDeclaration {
+        Some("invalid_protection_declaration")
+    } else {
+        None
+    }
+}
+
+pub(crate) fn stored_sha256(info: &ObjectInfo) -> Option<String> {
+    let bytes = info.checksum.as_ref()?;
+    let (sums, multipart) = read_checksums(bytes, 0);
+    let value = sums.get("SHA256")?;
+    let canonical = Checksum::new_from_string("SHA256", value)?.to_bytes(&[]);
+    // The general display decoder tolerates truncated suffixes. Eligibility does not.
+    (!multipart && canonical.as_ref() == bytes.as_ref()).then(|| value.clone())
+}
+
+pub(crate) fn fingerprint(info: &ObjectInfo) -> String {
+    let mut hash = Sha256::new();
+    // Ordered metadata includes transformation and protection declarations as well as user metadata.
+    for value in [
+        info.bucket.clone(),
+        info.name.clone(),
+        format!("{:?}", info.version_id),
+        format!("{:?}", info.data_dir),
+        format!("{:?}", info.mod_time),
+        info.size.to_string(),
+        info.actual_size.to_string(),
+        format!("{:?}", info.etag),
+        format!("{:?}", info.checksum),
+        format!("{:?}", info.transitioned_object),
+        info.user_tags.to_string(),
+        format!("{:?}", info.expires),
+        format!("{:?}", info.content_type),
+        format!("{:?}", info.content_encoding),
+        format!("{:?}", info.storage_class),
+    ] {
+        hash.update(value.len().to_le_bytes());
+        hash.update(value.as_bytes());
+    }
+    for part in info.parts.iter() {
+        let values = format!(
+            "{}:{}:{}:{:?}:{:?}:{:?}:{:?}",
+            part.number, part.size, part.actual_size, part.etag, part.mod_time, part.index, part.integrity
+        );
+        hash.update(values.len().to_le_bytes());
+        hash.update(values.as_bytes());
+        if let Some(checksums) = &part.checksums {
+            let mut checksums: Vec<_> = checksums.iter().collect();
+            checksums.sort_unstable_by_key(|(key, _)| *key);
+            for (key, value) in checksums {
+                hash.update(key.len().to_le_bytes());
+                hash.update(key.as_bytes());
+                hash.update(value.len().to_le_bytes());
+                hash.update(value.as_bytes());
+            }
+        }
+    }
+    let mut metadata: Vec<_> = info.user_defined.iter().collect();
+    metadata.sort_unstable_by_key(|(key, _)| *key);
+    for (key, value) in metadata {
+        hash.update(key.len().to_le_bytes());
+        hash.update(key.as_bytes());
+        hash.update(value.len().to_le_bytes());
+        hash.update(value.as_bytes());
+    }
+    hex_simd::encode_to_string(hash.finalize(), hex_simd::AsciiCase::Lower)
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum JobState {
+    Paused,
+    Running,
+    PauseRequested,
+    CancelRequested,
+    Cancelled,
+    Complete,
+    Failed,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(rename_all = "snake_case")]
+pub enum ItemState {
+    Pending,
+    Prepared,
+    Verified,
+    Migrated,
+    Unsupported,
+    Unavailable,
+    Mismatch,
+    Stale,
+    Conflict,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct ItemResult {
+    pub state: ItemState,
+    pub source_fingerprint: Option<String>,
+    pub expected_sha256: Option<String>,
+    pub evidence_source: Option<String>,
+    pub source_size: Option<u64>,
+    pub target_version_id: Option<Uuid>,
+    pub detail: Option<String>,
+}
+
+impl Default for ItemResult {
+    fn default() -> Self {
+        Self {
+            state: ItemState::Pending,
+            source_fingerprint: None,
+            expected_sha256: None,
+            evidence_source: None,
+            source_size: None,
+            target_version_id: None,
+            detail: None,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct Job {
+    pub format_version: u32,
+    pub id: Uuid,
+    pub bucket: String,
+    pub bucket_incarnation: Uuid,
+    pub request: JobRequest,
+    pub state: JobState,
+    pub revision: u64,
+    pub results: Vec<ItemResult>,
+    pub last_error: Option<String>,
+}
+
+#[derive(Debug, Serialize)]
+pub struct InventoryPage {
+    pub observed_at: String,
+    pub bucket_incarnation: Uuid,
+    pub items: Vec<InventoryItem>,
+    pub is_truncated: bool,
+    pub next_key_marker: Option<String>,
+    pub next_version_marker: Option<String>,
+}
+
+#[derive(Debug, Serialize)]
+pub struct Readiness {
+    pub local_reader_supported: bool,
+    pub write_requested: bool,
+    pub fleet_operator_attested: bool,
+    pub new_writes_enabled_here: bool,
+    pub fleet_capability_verified: bool,
+    pub unverified_requirements: [&'static str; 3],
+}
+
+pub fn readiness() -> Readiness {
+    let write = rustfs_utils::get_env_bool(rustfs_config::ENV_SHARD_INTEGRITY_WRITE, false);
+    let fleet = rustfs_utils::get_env_bool(rustfs_config::ENV_SHARD_INTEGRITY_FLEET_CONFIRMED, false);
+    Readiness {
+        local_reader_supported: true,
+        write_requested: write,
+        fleet_operator_attested: fleet,
+        new_writes_enabled_here: write && fleet,
+        fleet_capability_verified: false,
+        unverified_requirements: [
+            "peer_reader_writer_and_repair_capabilities",
+            "old_process_storage_admission",
+            "deployment_failure_and_performance_qualification",
+        ],
+    }
+}
@@ -0,0 +1,468 @@
+// Copyright 2024 RustFS Team
+// Licensed under the Apache License, Version 2.0.
+
+use super::model::{Protection, fingerprint, protection, sha256_bytes, stored_sha256, unsupported};
+use super::{IntegrityError, ItemState, Job, JobMode, JobState, Result, load, readiness, save};
+use crate::object_api::{
+    ObjectInfo, ObjectOptions, PutObjReader, ShardIntegrityWriteMode, WriteCompletion, without_get_object_body_cache_hook,
+};
+use crate::storage_api_contracts::object::{HTTPPreconditions, ObjectIO, ObjectOperations};
+use crate::store::ECStore;
+use rustfs_rio::{Checksum, HashReader};
+use sha2::{Digest, Sha256};
+use std::{collections::HashMap, sync::Arc, time::Duration};
+use tokio::io::{AsyncReadExt, AsyncSeekExt, AsyncWriteExt};
+
+const MARKER: &str = "integrity-migration-job-v1";
+
+pub(super) async fn run(
+    api: &Arc<ECStore>,
+    job: &mut Job,
+    mut etag: String,
+    guard: &rustfs_lock::NamespaceLockGuard,
+) -> Result<()> {
+    without_get_object_body_cache_hook(async {
+        for index in 0..job.results.len() {
+            ensure_running(api, job, guard).await?;
+            if !matches!(
+                job.results[index].state,
+                ItemState::Pending | ItemState::Prepared | ItemState::Unavailable
+            ) {
+                continue;
+            }
+            if job.request.mode == JobMode::Migrate {
+                validate_destination(api, job).await?;
+            }
+            if job.results[index].state == ItemState::Prepared
+                && job.request.mode == JobMode::Migrate
+                && reconcile_target(api, job, index, guard).await?
+            {
+                etag = checkpoint(api, job, &etag, guard).await?;
+                continue;
+            }
+            let request = job.request.items[index].clone();
+            let opts = ObjectOptions {
+                version_id: request.version_id.as_ref().map(|version| {
+                    if version == "null" {
+                        uuid::Uuid::nil().to_string()
+                    } else {
+                        version.clone()
+                    }
+                }),
+                include_part_checksums: true,
+                suppress_read_repair: true,
+                ..Default::default()
+            };
+            let info = match api.get_object_info(&job.bucket, &request.key, &opts).await {
+                Ok(info) => info,
+                Err(_) => {
+                    job.results[index].state = ItemState::Unavailable;
+                    job.results[index].detail = Some("source_metadata_unavailable".into());
+                    etag = checkpoint(api, job, &etag, guard).await?;
+                    continue;
+                }
+            };
+            if let Some(reason) = unsupported(&info) {
+                job.results[index].state = ItemState::Unsupported;
+                job.results[index].detail = Some(reason.into());
+                etag = checkpoint(api, job, &etag, guard).await?;
+                continue;
+            }
+            let identity = fingerprint(&info);
+            if job.results[index]
+                .source_fingerprint
+                .as_ref()
+                .is_some_and(|old| old != &identity)
+            {
+                job.results[index].state = ItemState::Stale;
+                etag = checkpoint(api, job, &etag, guard).await?;
+                continue;
+            }
+            let expected = request.expected_sha256.clone().or_else(|| stored_sha256(&info));
+            let Some(expected) = expected else {
+                job.results[index].state = ItemState::Unsupported;
+                job.results[index].detail = Some("no_canonical_single_part_sha256".into());
+                etag = checkpoint(api, job, &etag, guard).await?;
+                continue;
+            };
+            let size = u64::try_from(info.size).map_err(|_| IntegrityError::Invalid("negative source size"))?;
+            if size > job.request.max_object_bytes {
+                job.results[index].state = ItemState::Unsupported;
+                job.results[index].detail = Some("object_exceeds_staging_budget".into());
+                etag = checkpoint(api, job, &etag, guard).await?;
+                continue;
+            }
+            let result = &mut job.results[index];
+            result.source_fingerprint = Some(identity);
+            result.expected_sha256 = Some(expected.clone());
+            result.evidence_source = Some(
+                if request.expected_sha256.is_some() {
+                    "administrator_supplied_sha256"
+                } else {
+                    "stored_sha256"
+                }
+                .into(),
+            );
+            result.source_size = Some(size);
+            result.state = ItemState::Prepared;
+            result.detail = None;
+            etag = checkpoint(api, job, &etag, guard).await?;
+            let mut file = match stage(api, job, index, &info, &opts, guard).await {
+                Ok(file) => file,
+                Err(IntegrityError::Conflict) => {
+                    job.results[index].state = ItemState::Stale;
+                    etag = checkpoint(api, job, &etag, guard).await?;
+                    continue;
+                }
+                Err(IntegrityError::Invalid("content_sha256_mismatch")) => {
+                    job.results[index].state = ItemState::Mismatch;
+                    etag = checkpoint(api, job, &etag, guard).await?;
+                    continue;
+                }
+                Err(error) => return Err(error),
+            };
+            ensure_running(api, job, guard).await?;
+            if job.request.mode == JobMode::Audit {
+                job.results[index].state = ItemState::Verified;
+            } else {
+                let snapshot = validate_destination(api, job).await?;
+                file.rewind().await?;
+                let raw = sha256_bytes(&expected)?;
+                let expected_hex = hex_simd::encode_to_string(raw, hex_simd::AsciiCase::Lower);
+                let mut reader = HashReader::from_stream(
+                    ThrottledFile::new(file, job.request.bytes_per_second),
+                    info.size,
+                    info.size,
+                    None,
+                    Some(expected_hex),
+                    false,
+                )?;
+                reader.add_non_trailing_checksum(Checksum::new_from_string("SHA256", &expected), false)?;
+                let mut body = PutObjReader::new(reader);
+                let mut metadata: HashMap<String, String> = info
+                    .user_defined
+                    .iter()
+                    .filter(|(key, _)| {
+                        key.starts_with("x-amz-meta-")
+                            || matches!(key.as_str(), "cache-control" | "content-disposition" | "content-language" | "expires")
+                    })
+                    .map(|(key, value)| (key.clone(), value.clone()))
+                    .collect();
+                if let Some(content_type) = &info.content_type {
+                    metadata.insert("content-type".into(), content_type.clone());
+                }
+                if let Some(content_encoding) = &info.content_encoding {
+                    metadata.insert("content-encoding".into(), content_encoding.clone());
+                }
+                if let Some(expires) = info.expires {
+                    metadata.insert(
+                        "expires".into(),
+                        expires
+                            .format(&time::format_description::well_known::Rfc3339)
+                            .map_err(|_| IntegrityError::Invalid("invalid source expiration"))?,
+                    );
+                }
+                if let Some(storage_class) = &info.storage_class {
+                    metadata.insert("x-amz-storage-class".into(), storage_class.clone());
+                }
+                if !info.user_tags.is_empty() {
+                    metadata.insert(rustfs_utils::http::headers::AMZ_OBJECT_TAGGING.into(), info.user_tags.to_string());
+                }
+                rustfs_utils::http::insert_str(&mut metadata, MARKER, marker(job, index)?);
+                let mut write = ObjectOptions {
+                    user_defined: metadata,
+                    preserve_delete_marker: true,
+                    shard_integrity_write_mode: Some(ShardIntegrityWriteMode::Protected),
+                    write_completion: WriteCompletion::TailDrained,
+                    http_preconditions: Some(HTTPPreconditions {
+                        if_none_match: Some("*".into()),
+                        ..Default::default()
+                    }),
+                    ..Default::default()
+                };
+                write.add_namespace_lock_guard(guard);
+                write.expected_bucket_incarnation_id = Some(job.bucket_incarnation);
+                snapshot.add_lock_fences(&mut write);
+                write.object_lock_config_snapshot = Some(snapshot.clone());
+                let target = request
+                    .target_key
+                    .as_deref()
+                    .ok_or(IntegrityError::Invalid("missing target"))?;
+                // Even an error may follow a committed PUT. Reconcile before retrying; never overwrite.
+                let put = api.put_object(&job.bucket, target, &mut body, &write).await;
+                drop(write);
+                drop(snapshot);
+                if !reconcile_target(api, job, index, guard).await? {
+                    put?;
+                    return Err(IntegrityError::Conflict);
+                }
+            }
+            etag = checkpoint(api, job, &etag, guard).await?;
+        }
+        job.state = if job.results.iter().any(|item| item.state == ItemState::Unavailable) {
+            job.last_error = Some("source_unavailable_retry_requires_resume".into());
+            JobState::Failed
+        } else {
+            JobState::Complete
+        };
+        checkpoint(api, job, &etag, guard).await?;
+        Ok(())
+    })
+    .await
+}
+
+async fn checkpoint(api: &Arc<ECStore>, job: &mut Job, etag: &str, guard: &rustfs_lock::NamespaceLockGuard) -> Result<String> {
+    ensure_running(api, job, guard).await?;
+    job.revision = job.revision.checked_add(1).ok_or(IntegrityError::Conflict)?;
+    save(api, job, Some(etag), Some(guard)).await
+}
+
+async fn ensure_running(api: &Arc<ECStore>, job: &Job, guard: &rustfs_lock::NamespaceLockGuard) -> Result<()> {
+    if guard.lock_lost_signal().is_some_and(|signal| signal.is_lost()) {
+        return Err(IntegrityError::Stopped);
+    }
+    let (current, _) = load(api, &job.bucket, job.id).await?;
+    if current.state != JobState::Running {
+        return Err(IntegrityError::Stopped);
+    }
+    if current.revision != job.revision {
+        return Err(IntegrityError::Conflict);
+    }
+    Ok(())
+}
+
+async fn stage(
+    api: &Arc<ECStore>,
+    job: &Job,
+    index: usize,
+    info: &ObjectInfo,
+    opts: &ObjectOptions,
+    guard: &rustfs_lock::NamespaceLockGuard,
+) -> Result<tokio::fs::File> {
+    let mut reader = api
+        .get_object_reader(&job.bucket, &job.request.items[index].key, None, Default::default(), opts)
+        .await?;
+    if fingerprint(&reader.object_info) != fingerprint(info) {
+        return Err(IntegrityError::Conflict);
+    }
+    let std_file = tokio::task::spawn_blocking(tempfile::tempfile)
+        .await
+        .map_err(|_| IntegrityError::Invalid("staging task failed"))??;
+    let mut file = tokio::fs::File::from_std(std_file);
+    let expected = job.results[index]
+        .expected_sha256
+        .as_deref()
+        .ok_or(IntegrityError::Invalid("missing expected digest"))?;
+    copy_verified(api, job, &mut reader.stream, Some(&mut file), info.size, expected, guard).await?;
+    file.flush().await?;
+    let after = api.get_object_info(&job.bucket, &job.request.items[index].key, opts).await?;
+    if fingerprint(&after) != fingerprint(info) {
+        return Err(IntegrityError::Conflict);
+    }
+    Ok(file)
+}
+
+async fn copy_verified(
+    api: &Arc<ECStore>,
+    job: &Job,
+    input: &mut (dyn tokio::io::AsyncRead + Unpin + Send + Sync),
+    mut output: Option<&mut tokio::fs::File>,
+    size: i64,
+    expected: &str,
+    guard: &rustfs_lock::NamespaceLockGuard,
+) -> Result<()> {
+    let expected = sha256_bytes(expected)?;
+    let expected_size = u64::try_from(size).map_err(|_| IntegrityError::Invalid("negative object size"))?;
+    let mut buffer = vec![0u8; 64 * 1024];
+    let mut count = 0u64;
+    let mut hash = Sha256::new();
+    let started = tokio::time::Instant::now();
+    let mut checked = started;
+    loop {
+        if checked.elapsed() >= Duration::from_secs(1) {
+            ensure_running(api, job, guard).await?;
+            checked = tokio::time::Instant::now();
+        }
+        let n = tokio::time::timeout(Duration::from_secs(30), input.read(&mut buffer))
+            .await
+            .map_err(|_| std::io::Error::new(std::io::ErrorKind::TimedOut, "integrity read timed out"))??;
+        if n == 0 {
+            break;
+        }
+        count = count
+            .checked_add(u64::try_from(n).map_err(|_| IntegrityError::Invalid("read size overflow"))?)
+            .ok_or(IntegrityError::Invalid("read size overflow"))?;
+        if count > expected_size || count > job.request.max_object_bytes {
+            return Err(IntegrityError::Invalid("content_sha256_mismatch"));
+        }
+        hash.update(&buffer[..n]);
+        if let Some(file) = output.as_deref_mut() {
+            file.write_all(&buffer[..n]).await?;
+        }
+        let delay = Duration::from_secs_f64(count as f64 / job.request.bytes_per_second as f64);
+        tokio::time::sleep_until(started + delay).await;
+    }
+    if count != expected_size || hash.finalize().as_slice() != expected {
+        return Err(IntegrityError::Invalid("content_sha256_mismatch"));
+    }
+    Ok(())
+}
+
+fn marker(job: &Job, index: usize) -> Result<String> {
+    let result = &job.results[index];
+    Ok(format!(
+        "{}:{}:{}:{}",
+        job.id,
+        index,
+        result
+            .source_fingerprint
+            .as_deref()
+            .ok_or(IntegrityError::Invalid("missing source identity"))?,
+        result
+            .expected_sha256
+            .as_deref()
+            .ok_or(IntegrityError::Invalid("missing source digest"))?
+    ))
+}
+
+async fn reconcile_target(
+    api: &Arc<ECStore>,
+    job: &mut Job,
+    index: usize,
+    guard: &rustfs_lock::NamespaceLockGuard,
+) -> Result<bool> {
+    let key = job.request.items[index]
+        .target_key
+        .as_deref()
+        .ok_or(IntegrityError::Invalid("missing target"))?;
+    let opts = ObjectOptions {
+        include_part_checksums: true,
+        suppress_read_repair: true,
+        ..Default::default()
+    };
+    let target = match api.get_object_info(&job.bucket, key, &opts).await {
+        Ok(info) => info,
+        Err(crate::error::Error::FileNotFound | crate::error::Error::ObjectNotFound(_, _)) => return Ok(false),
+        Err(error) => return Err(error.into()),
+    };
+    let expected_marker = marker(job, index)?;
+    if rustfs_utils::http::get_consistent_str(&target.user_defined, MARKER) != Some(expected_marker.as_str()) {
+        job.results[index].state = ItemState::Conflict;
+        job.results[index].detail = Some("target_owned_by_another_operation".into());
+        return Ok(true);
+    }
+    if protection(&target) != Protection::IndependentCommitment {
+        return Err(IntegrityError::Invalid("target_not_protected"));
+    }
+    let mut reader = api
+        .get_object_reader(&job.bucket, key, None, Default::default(), &opts)
+        .await?;
+    if fingerprint(&target) != fingerprint(&reader.object_info) {
+        return Err(IntegrityError::Conflict);
+    }
+    let size = job.results[index]
+        .source_size
+        .ok_or(IntegrityError::Invalid("missing source size"))?;
+    let expected = job.results[index]
+        .expected_sha256
+        .clone()
+        .ok_or(IntegrityError::Invalid("missing source digest"))?;
+    copy_verified(
+        api,
+        job,
+        &mut reader.stream,
+        None,
+        i64::try_from(size).map_err(|_| IntegrityError::Invalid("source size overflow"))?,
+        &expected,
+        guard,
+    )
+    .await?;
+    let after = api.get_object_info(&job.bucket, key, &opts).await?;
+    if fingerprint(&after) != fingerprint(&target) {
+        return Err(IntegrityError::Conflict);
+    }
+    job.results[index].state = ItemState::Migrated;
+    job.results[index].target_version_id = target.version_id;
+    Ok(true)
+}
+
+async fn validate_destination(api: &Arc<ECStore>, job: &Job) -> Result<Arc<crate::object_api::ObjectLockConfigSnapshot>> {
+    if !readiness().new_writes_enabled_here {
+        return Err(IntegrityError::NotActivated);
+    }
+    let snapshot = api.object_lock_config_snapshot(&job.bucket).await?;
+    if snapshot
+        .metadata_transaction_guard_for(api.id, &job.bucket, Some(job.bucket_incarnation))
+        .is_none()
+    {
+        return Err(IntegrityError::Conflict);
+    }
+    let metadata = crate::bucket::metadata_sys::integrity_migration_metadata_in(&api.ctx, &job.bucket).await?;
+    // The direct storage writer is not the S3 orchestration layer. Reject configurations it cannot preserve.
+    if metadata.bucket_incarnation_id != job.bucket_incarnation {
+        return Err(IntegrityError::Conflict);
+    }
+    if !metadata.versioning_config_xml.is_empty()
+        || !metadata.encryption_config_xml.is_empty()
+        || !metadata.object_lock_config_xml.is_empty()
+        || !metadata.replication_config_xml.is_empty()
+        || !metadata.notification_config_xml.is_empty()
+        || !metadata.quota_config_json.is_empty()
+        || !metadata.on_demand_migration_config_json.is_empty()
+        || !metadata.lifecycle_config_xml.is_empty()
+        || !metadata.table_bucket_config_json.is_empty()
+        || !metadata.bucket_acl_config_json.is_empty()
+        || !metadata.logging_config_xml.is_empty()
+    {
+        return Err(IntegrityError::UnsupportedBucket);
+    }
+    Ok(snapshot)
+}
+
+/// Bound publication throughput as well as the source and destination verification reads.
+struct ThrottledFile {
+    file: tokio::fs::File,
+    started: tokio::time::Instant,
+    count: u64,
+    rate: u64,
+    sleep: std::pin::Pin<Box<tokio::time::Sleep>>,
+}
+
+impl ThrottledFile {
+    fn new(file: tokio::fs::File, rate: u64) -> Self {
+        let started = tokio::time::Instant::now();
+        Self {
+            file,
+            started,
+            count: 0,
+            rate,
+            sleep: Box::pin(tokio::time::sleep_until(started)),
+        }
+    }
+}
+
+impl tokio::io::AsyncRead for ThrottledFile {
+    fn poll_read(
+        mut self: std::pin::Pin<&mut Self>,
+        cx: &mut std::task::Context<'_>,
+        output: &mut tokio::io::ReadBuf<'_>,
+    ) -> std::task::Poll<std::io::Result<()>> {
+        use std::future::Future;
+        if self.sleep.as_mut().poll(cx).is_pending() {
+            return std::task::Poll::Pending;
+        }
+        let capacity = output.remaining().min(64 * 1024);
+        let mut buffer = tokio::io::ReadBuf::new(output.initialize_unfilled_to(capacity));
+        match std::pin::Pin::new(&mut self.file).poll_read(cx, &mut buffer) {
+            std::task::Poll::Ready(Ok(())) => {
+                let count = buffer.filled().len();
+                output.advance(count);
+                self.count += count as u64;
+                let deadline = self.started + Duration::from_secs_f64(self.count as f64 / self.rate as f64);
+                self.sleep.as_mut().reset(deadline);
+                std::task::Poll::Ready(Ok(()))
+            }
+            result => result,
+        }
+    }
+}
@@ -0,0 +1,387 @@
+// Copyright 2024 RustFS Team
+// Licensed under the Apache License, Version 2.0.
+
+use super::*;
+use crate::object_api::{ObjectOptions, PutObjReader, ShardIntegrityWriteMode};
+use crate::storage_api_contracts::{
+    bucket::{BucketOperations, MakeBucketOptions},
+    object::{ObjectIO, ObjectOperations},
+};
+use tokio::io::AsyncReadExt;
+
+async fn fixture() -> (Vec<tempfile::TempDir>, Arc<crate::store::ECStore>, String) {
+    let (dirs, store) = crate::services::rebalance::test_store_with_persisted_rebalance_meta(Default::default()).await;
+    crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+    let bucket = format!("integrity-{}", Uuid::new_v4());
+    store
+        .make_bucket(&bucket, &MakeBucketOptions::default())
+        .await
+        .expect("fixture bucket");
+    (dirs, store, bucket)
+}
+
+fn request(mode: JobMode, digest: String) -> JobRequest {
+    JobRequest {
+        mode,
+        items: vec![ItemRequest {
+            key: "source".into(),
+            version_id: Some("null".into()),
+            expected_sha256: Some(digest),
+            target_key: (mode == JobMode::Migrate).then(|| "target".into()),
+        }],
+        bytes_per_second: 1024 * 1024,
+        max_object_bytes: 16 * 1024 * 1024,
+    }
+}
+
+async fn finish(store: Arc<crate::store::ECStore>, bucket: &str, id: Uuid) -> Job {
+    tokio::time::timeout(std::time::Duration::from_secs(60), async {
+        loop {
+            let job = get_job(store.clone(), bucket, id).await.expect("durable job");
+            if job.state != JobState::Running {
+                return job;
+            }
+            tokio::time::sleep(std::time::Duration::from_millis(20)).await;
+        }
+    })
+    .await
+    .expect("job must finish")
+}
+
+#[tokio::test]
+#[serial_test::serial(shard_integrity_rollout)]
+async fn audit_migration_and_ack_recovery_preserve_source_and_reject_conflicts() {
+    temp_env::async_with_vars(
+        [
+            (rustfs_config::ENV_SHARD_INTEGRITY_WRITE, Some("true")),
+            (rustfs_config::ENV_SHARD_INTEGRITY_FLEET_CONFIRMED, Some("true")),
+        ],
+        async {
+            let (_dirs, store, bucket) = fixture().await;
+            let payload = vec![42; 256 * 1024 + 17];
+            let expected = rustfs_rio::Checksum::new_from_data(rustfs_rio::ChecksumType::SHA256, &payload)
+                .expect("digest")
+                .encoded;
+            store
+                .put_object(
+                    &bucket,
+                    "source",
+                    &mut PutObjReader::from_vec(payload.clone()),
+                    &ObjectOptions {
+                        user_defined: [
+                            ("cache-control".into(), "max-age=300".into()),
+                            ("content-disposition".into(), "attachment; filename=sample.bin".into()),
+                            ("expires".into(), "2030-01-01T00:00:00Z".into()),
+                            (rustfs_utils::http::headers::AMZ_OBJECT_TAGGING.into(), "kind=archive".into()),
+                        ]
+                        .into(),
+                        shard_integrity_write_mode: Some(ShardIntegrityWriteMode::Legacy),
+                        write_completion: crate::object_api::WriteCompletion::TailDrained,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("source");
+            let source = store
+                .get_object_info(&bucket, "source", &ObjectOptions::default())
+                .await
+                .expect("source metadata");
+            assert_eq!(source.user_tags.as_str(), "kind=archive", "source fixture tags");
+            let before = model::fingerprint(&source);
+            let page = inventory(store.clone(), &bucket, "", None, None, 1).await.expect("inventory");
+            assert_eq!(page.items.len(), 1);
+            assert_eq!(page.items[0].protection, Protection::Legacy);
+            assert_eq!(page.items[0].version_id, "null", "inventory must round-trip an explicit null selector");
+            let audit = create_job(store.clone(), &bucket, request(JobMode::Audit, expected.clone()))
+                .await
+                .expect("create audit");
+            resume_job(store.clone(), &bucket, audit.id).await.expect("resume audit");
+            let audited = finish(store.clone(), &bucket, audit.id).await;
+            assert_eq!(audited.state, JobState::Complete, "{audited:?}");
+            assert_eq!(audited.results[0].state, ItemState::Verified);
+
+            let migration = create_job(store.clone(), &bucket, request(JobMode::Migrate, expected.clone()))
+                .await
+                .expect("create migration");
+            resume_job(store.clone(), &bucket, migration.id)
+                .await
+                .expect("resume migration");
+            let migrated = finish(store.clone(), &bucket, migration.id).await;
+            assert_eq!(migrated.state, JobState::Complete, "{migrated:?}");
+            assert_eq!(migrated.results[0].state, ItemState::Migrated);
+            let mut reader = store
+                .get_object_reader(&bucket, "target", None, Default::default(), &ObjectOptions::default())
+                .await
+                .expect("target");
+            assert_eq!(model::protection(&reader.object_info), Protection::IndependentCommitment);
+            assert_eq!(reader.object_info.user_tags.as_str(), "kind=archive");
+            assert_eq!(reader.object_info.expires, source.expires);
+            assert_eq!(
+                reader.object_info.user_defined.get("cache-control"),
+                source.user_defined.get("cache-control")
+            );
+            assert_eq!(
+                reader.object_info.user_defined.get("content-disposition"),
+                source.user_defined.get("content-disposition")
+            );
+            let mut actual = Vec::new();
+            reader.stream.read_to_end(&mut actual).await.expect("target content");
+            assert_eq!(actual, payload);
+            let after = store
+                .get_object_info(&bucket, "source", &ObjectOptions::default())
+                .await
+                .expect("source after");
+            assert_eq!(model::fingerprint(&after), before);
+
+            // Simulate a lost completion checkpoint after the create-only PUT committed.
+            let (mut replay, etag) = load(&store, &bucket, migration.id).await.expect("load receipt");
+            replay.state = JobState::Paused;
+            replay.results[0].state = ItemState::Prepared;
+            replay.revision += 1;
+            save(&store, &replay, Some(&etag), None)
+                .await
+                .expect("lost acknowledgement fixture");
+            resume_job(store.clone(), &bucket, replay.id)
+                .await
+                .expect("reconcile after restart");
+            let reconciled = finish(store.clone(), &bucket, replay.id).await;
+            assert_eq!(reconciled.results[0].state, ItemState::Migrated, "{reconciled:?}");
+
+            let conflict = create_job(store.clone(), &bucket, request(JobMode::Migrate, expected))
+                .await
+                .expect("conflicting job");
+            resume_job(store.clone(), &bucket, conflict.id)
+                .await
+                .expect("resume conflict");
+            let conflict = finish(store.clone(), &bucket, conflict.id).await;
+            assert_eq!(conflict.results[0].state, ItemState::Conflict, "{conflict:?}");
+            let target = store
+                .get_object_info(&bucket, "target", &ObjectOptions::default())
+                .await
+                .expect("unchanged target");
+            assert_eq!(target.data_dir, reader.object_info.data_dir);
+        },
+    )
+    .await;
+}
+
+#[tokio::test]
+async fn mismatch_never_publishes_and_paused_job_survives_reload() {
+    let (_dirs, store, bucket) = fixture().await;
+    store
+        .put_object(
+            &bucket,
+            "source",
+            &mut PutObjReader::from_vec(b"actual".to_vec()),
+            &ObjectOptions {
+                shard_integrity_write_mode: Some(ShardIntegrityWriteMode::Legacy),
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("source");
+    let wrong = rustfs_rio::Checksum::new_from_data(rustfs_rio::ChecksumType::SHA256, b"different")
+        .expect("digest")
+        .encoded;
+    let job = create_job(store.clone(), &bucket, request(JobMode::Audit, wrong))
+        .await
+        .expect("job");
+    assert_eq!(get_job(store.clone(), &bucket, job.id).await.expect("reload").state, JobState::Paused);
+    resume_job(store.clone(), &bucket, job.id).await.expect("resume");
+    let job = finish(store.clone(), &bucket, job.id).await;
+    assert_eq!(job.results[0].state, ItemState::Mismatch, "{job:?}");
+    assert!(
+        store
+            .get_object_info(&bucket, "target", &ObjectOptions::default())
+            .await
+            .is_err()
+    );
+}
+
+#[test]
+fn fingerprint_is_independent_of_metadata_insertion_order() {
+    let mut a = crate::object_api::ObjectInfo::default();
+    let mut first = std::collections::HashMap::new();
+    first.insert("a".into(), "one".into());
+    first.insert("b".into(), "two".into());
+    a.user_defined = Arc::new(first);
+    let mut b = a.clone();
+    let mut second = std::collections::HashMap::new();
+    second.insert("b".into(), "two".into());
+    second.insert("a".into(), "one".into());
+    b.user_defined = Arc::new(second);
+    assert_eq!(model::fingerprint(&a), model::fingerprint(&b));
+    b.data_dir = Some(Uuid::new_v4());
+    assert_ne!(model::fingerprint(&a), model::fingerprint(&b));
+}
+
+#[test]
+fn fingerprint_is_independent_of_part_checksum_insertion_order() {
+    let mut checksums = [
+        ("SHA256", "sha256"),
+        ("SHA1", "sha1"),
+        ("CRC32", "crc32"),
+        ("CRC32C", "crc32c"),
+    ];
+    let a = crate::object_api::ObjectInfo {
+        parts: Arc::new(vec![rustfs_filemeta::ObjectPartInfo {
+            checksums: Some(checksums.map(|(key, value)| (key.to_string(), value.to_string())).into()),
+            ..Default::default()
+        }]),
+        ..Default::default()
+    };
+    let mut b = a.clone();
+    let expected = model::fingerprint(&a);
+    for _ in 0..checksums.len() {
+        checksums.rotate_left(1);
+        Arc::make_mut(&mut b.parts)[0].checksums =
+            Some(checksums.map(|(key, value)| (key.to_string(), value.to_string())).into());
+        assert_eq!(expected, model::fingerprint(&b));
+    }
+    Arc::make_mut(&mut b.parts)[0]
+        .checksums
+        .as_mut()
+        .expect("part checksums")
+        .insert("SHA256".into(), "changed".into());
+    assert_ne!(expected, model::fingerprint(&b));
+}
+
+#[test]
+fn rejects_migration_into_a_source_key() {
+    let request = JobRequest {
+        mode: JobMode::Migrate,
+        items: vec![ItemRequest {
+            key: "source".into(),
+            version_id: None,
+            expected_sha256: None,
+            target_key: Some("source".into()),
+        }],
+        bytes_per_second: 1024 * 1024,
+        max_object_bytes: 1024,
+    };
+    assert!(request.validate().is_err());
+}
+
+#[test]
+fn rejects_truncated_checksum_suffix_as_migration_evidence() {
+    let checksum = rustfs_rio::Checksum::new_from_data(rustfs_rio::ChecksumType::SHA256, b"payload").expect("checksum");
+    let mut info = crate::object_api::ObjectInfo {
+        checksum: Some(checksum.to_bytes(&[])),
+        ..Default::default()
+    };
+    assert_eq!(model::stored_sha256(&info), Some(checksum.encoded));
+    let mut bytes = info.checksum.take().expect("bytes").to_vec();
+    bytes.push(0x80);
+    info.checksum = Some(bytes.into());
+    assert!(model::stored_sha256(&info).is_none());
+}
+
+#[tokio::test]
+async fn unavailable_source_can_resume_and_cancel_intent_cannot_restart() {
+    let (_dirs, store, bucket) = fixture().await;
+    let payload = b"recovered".to_vec();
+    let digest = rustfs_rio::Checksum::new_from_data(rustfs_rio::ChecksumType::SHA256, &payload)
+        .expect("digest")
+        .encoded;
+    let job = create_job(store.clone(), &bucket, request(JobMode::Audit, digest))
+        .await
+        .expect("create");
+    resume_job(store.clone(), &bucket, job.id).await.expect("start");
+    let failed = finish(store.clone(), &bucket, job.id).await;
+    assert_eq!(failed.state, JobState::Failed);
+    assert_eq!(failed.results[0].state, ItemState::Unavailable);
+    store
+        .put_object(
+            &bucket,
+            "source",
+            &mut PutObjReader::from_vec(payload),
+            &ObjectOptions {
+                shard_integrity_write_mode: Some(ShardIntegrityWriteMode::Legacy),
+                write_completion: WriteCompletion::TailDrained,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("recovered source");
+    resume_job(store.clone(), &bucket, job.id).await.expect("retry");
+    let completed = finish(store.clone(), &bucket, job.id).await;
+    assert_eq!(completed.state, JobState::Complete, "{completed:?}");
+    assert_eq!(completed.results[0].state, ItemState::Verified);
+
+    // A process restart may leave the durable cancel intent without a live worker.
+    let (mut interrupted, etag) = load(&store, &bucket, job.id).await.expect("load");
+    interrupted.state = JobState::CancelRequested;
+    interrupted.revision += 1;
+    save(&store, &interrupted, Some(&etag), None).await.expect("persist intent");
+    assert_eq!(
+        control_job(store.clone(), &bucket, job.id, false).await.expect("pause").state,
+        JobState::CancelRequested
+    );
+    assert_eq!(
+        resume_job(store.clone(), &bucket, job.id)
+            .await
+            .expect("recover cancel")
+            .state,
+        JobState::Cancelled
+    );
+}
+
+#[tokio::test]
+async fn pause_and_cancel_are_durable_during_content_read() {
+    let (_dirs, store, bucket) = fixture().await;
+    let payload = vec![7; 512 * 1024];
+    let digest = rustfs_rio::Checksum::new_from_data(rustfs_rio::ChecksumType::SHA256, &payload)
+        .expect("digest")
+        .encoded;
+    store
+        .put_object(
+            &bucket,
+            "source",
+            &mut PutObjReader::from_vec(payload),
+            &ObjectOptions {
+                shard_integrity_write_mode: Some(ShardIntegrityWriteMode::Legacy),
+                write_completion: WriteCompletion::TailDrained,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("source");
+    let mut req = request(JobMode::Audit, digest);
+    req.bytes_per_second = 64 * 1024;
+    let job = create_job(store.clone(), &bucket, req).await.expect("create");
+    resume_job(store.clone(), &bucket, job.id).await.expect("start");
+    tokio::time::timeout(std::time::Duration::from_secs(20), async {
+        loop {
+            if get_job(store.clone(), &bucket, job.id).await.expect("status").results[0].state == ItemState::Prepared {
+                break;
+            }
+            tokio::time::sleep(std::time::Duration::from_millis(20)).await;
+        }
+    })
+    .await
+    .expect("prepared");
+    let intent = control_job(store.clone(), &bucket, job.id, false).await.expect("pause");
+    assert_eq!(intent.state, JobState::PauseRequested);
+    tokio::time::timeout(std::time::Duration::from_secs(20), async {
+        loop {
+            let paused = get_job(store.clone(), &bucket, job.id).await.expect("status");
+            if paused.state == JobState::Paused {
+                break;
+            }
+            tokio::time::sleep(std::time::Duration::from_millis(20)).await;
+        }
+    })
+    .await
+    .expect("pause acknowledged");
+    assert_eq!(
+        control_job(store.clone(), &bucket, job.id, true).await.expect("cancel").state,
+        JobState::Cancelled
+    );
+    assert_eq!(
+        resume_job(store.clone(), &bucket, job.id)
+            .await
+            .expect("terminal resume")
+            .state,
+        JobState::Cancelled
+    );
+}
@@ -16,6 +16,7 @@

 pub(crate) mod batch_processor;
 pub(crate) mod event_notification;
+pub(crate) mod integrity;
 pub(crate) mod metrics_realtime;
 pub(crate) mod notification_sys;
 pub(crate) mod rebalance;
@@ -42,6 +42,7 @@ use std::sync::{
    atomic::{AtomicBool, AtomicUsize, Ordering},
 };
 use std::time::{Duration, Instant, SystemTime};
+use tokio::sync::Notify;
 use tokio::time::{sleep, timeout};
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, error, info, warn};
@@ -61,6 +62,14 @@ const TIER_CONFIG_RELOAD_RETRY_CAP: Duration = Duration::from_secs(5);
 const REMOTE_VERSION_STATE_PROBE_INTERVAL: Duration = Duration::from_secs(10);
 const REMOTE_VERSION_STATE_PROBE_TIMEOUT: Duration = Duration::from_secs(5);
 const REMOTE_VERSION_STATE_PROOF_TTL: Duration = Duration::from_secs(30);
+/// Fallback poll cadence while startup has not yet published the notification
+/// system. IAM finalization starts the probe before `init_notification_runtime`
+/// runs, so the first pass always fails closed; waiting a full probe interval
+/// there left a single node without its durable quota capability for ten
+/// seconds after `/health` reported ok (rustfs/rustfs#8014). Publication wakes
+/// the probe immediately through `NOTIFICATION_SYS_PUBLISHED`; this bound only
+/// covers a wakeup that races the availability check.
+const REMOTE_VERSION_STATE_PROBE_BOOTSTRAP_INTERVAL: Duration = Duration::from_millis(100);
 const CROSS_POOL_FENCE_SUPPORTED_VERSION: u32 = 2;
 const TIER_DELETE_JOURNAL_POLICY_SUPPORTED_VERSION: u32 = 3;
 const DECOMMISSION_TARGET_FENCE_POLICY_SUPPORTED_VERSION: u32 = 4;
@@ -338,6 +347,9 @@ static LEGACY_TRANSITION_STATE_RECONCILE_FLEET_PROOF: OnceLock<std::sync::RwLock
 static ILM_RECOVERY_EXPORT_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static TRANSITION_TRANSACTION_COMPACTION_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static REMOTE_VERSION_STATE_PROBE_TOPOLOGY: OnceLock<String> = OnceLock::new();
+/// Signalled once `GLOBAL_NOTIFICATION_SYS` is published so the fleet probe can
+/// run its first real pass without waiting for the bootstrap poll.
+static NOTIFICATION_SYS_PUBLISHED: Notify = Notify::const_new();
 static ILM_RECOVERY_EXPORT_LOCAL_PROCESS_EPOCH: LazyLock<Uuid> = LazyLock::new(Uuid::new_v4);

 fn cross_pool_fence_fleet_proof_slot() -> &'static std::sync::RwLock<FleetCapabilityProofState> {
@@ -368,6 +380,24 @@ fn transition_transaction_compaction_fleet_proof_slot() -> &'static std::sync::R
    TRANSITION_TRANSACTION_COMPACTION_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
 }

+fn all_fleet_capability_proof_slots() -> [&'static std::sync::RwLock<FleetCapabilityProofState>; 7] {
+    [
+        remote_version_state_fleet_proof_slot(),
+        cross_pool_fence_fleet_proof_slot(),
+        tier_delete_journal_fleet_proof_slot(),
+        decommission_target_fence_fleet_proof_slot(),
+        legacy_transition_state_reconcile_fleet_proof_slot(),
+        ilm_recovery_export_fleet_proof_slot(),
+        transition_transaction_compaction_fleet_proof_slot(),
+    ]
+}
+
+fn revoke_all_fleet_capability_proofs() {
+    for slot in all_fleet_capability_proof_slots() {
+        revoke_fleet_capability_proof(slot);
+    }
+}
+
 fn revoke_fleet_capability_proof_state(state: &mut FleetCapabilityProofState) {
    if let Some(proof) = state.proof.take() {
        proof.generation.revoke();
@@ -453,6 +483,20 @@ pub(crate) fn acquire_remote_version_state_fleet_proof() -> Option<RemoteVersion
    acquire_fleet_capability_proof_from(&state, expected_topology, Instant::now()).map(RemoteVersionStateFleetProofToken)
 }

+pub(crate) fn acquire_remote_version_state_writer_fleet_proof() -> Option<RemoteVersionStateFleetProofToken> {
+    let requested = rustfs_utils::get_env_bool(
+        rustfs_config::ENV_TIER_REMOTE_VERSION_STATE_WRITE,
+        rustfs_config::DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE,
+    );
+    let fleet_confirmed = rustfs_utils::get_env_bool(
+        rustfs_config::ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED,
+        rustfs_config::DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED,
+    );
+    (requested && fleet_confirmed)
+        .then(acquire_remote_version_state_fleet_proof)
+        .flatten()
+}
+
 fn acquire_fleet_capability_proof_from(
    state: &FleetCapabilityProofState,
    expected_topology: &str,
@@ -1212,15 +1256,7 @@ fn insert_remote_version_state_peer(peer_epochs: &mut BTreeMap<String, Uuid>, pe
 pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
    if REMOTE_VERSION_STATE_PROBE_TOPOLOGY.set(topology_fingerprint.clone()).is_err() {
        if REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() != Some(&topology_fingerprint) {
-            for slot in [
-                remote_version_state_fleet_proof_slot(),
-                cross_pool_fence_fleet_proof_slot(),
-                tier_delete_journal_fleet_proof_slot(),
-                decommission_target_fence_fleet_proof_slot(),
-                legacy_transition_state_reconcile_fleet_proof_slot(),
-                ilm_recovery_export_fleet_proof_slot(),
-                transition_transaction_compaction_fleet_proof_slot(),
-            ] {
+            for slot in all_fleet_capability_proof_slots() {
                mark_fleet_capability_topology_conflict(slot);
            }
        }
@@ -1228,53 +1264,59 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
    }

    tokio::spawn(async move {
+        let mut notification_sys_unavailable_logged = false;
        loop {
-            let notification_sys = get_global_notification_sys();
-            let remote_version_state_probe = async {
-                match notification_sys.as_ref() {
-                    Some(notification_sys) => timeout(
-                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
-                        notification_sys.probe_remote_version_state_fleet(&topology_fingerprint),
-                    )
-                    .await
-                    .unwrap_or_else(|_| Err(Error::other("remote version state fleet capability probe timed out"))),
-                    None => Err(Error::other("remote version state fleet capability notification system is unavailable")),
+            let Some(notification_sys) = get_global_notification_sys() else {
+                // Startup publishes the notification system after this probe
+                // is started. Stay failed closed, but retry on the bootstrap
+                // cadence so a single node gains its capabilities as soon as
+                // the system appears instead of one full probe interval later.
+                revoke_all_fleet_capability_proofs();
+                if !notification_sys_unavailable_logged {
+                    notification_sys_unavailable_logged = true;
+                    debug!(
+                        event = EVENT_NOTIFICATION_CAPABILITY_PROBE,
+                        component = LOG_COMPONENT_ECSTORE,
+                        subsystem = LOG_SUBSYSTEM_NOTIFICATION,
+                        state = "waiting_for_notification_system",
+                        "notification capability probe"
+                    );
                }
+                let _ = timeout(REMOTE_VERSION_STATE_PROBE_BOOTSTRAP_INTERVAL, NOTIFICATION_SYS_PUBLISHED.notified()).await;
+                continue;
+            };
+            notification_sys_unavailable_logged = false;
+            let remote_version_state_probe = async {
+                timeout(
+                    REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                    notification_sys.probe_remote_version_state_fleet(&topology_fingerprint),
+                )
+                .await
+                .unwrap_or_else(|_| Err(Error::other("remote version state fleet capability probe timed out")))
            };
            let cross_pool_fence_probe = async {
-                match notification_sys.as_ref() {
-                    Some(notification_sys) => timeout(
-                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
-                        notification_sys.probe_cross_pool_fence_fleet(&topology_fingerprint),
-                    )
-                    .await
-                    .unwrap_or_else(|_| Err(Error::other("cross-pool fence fleet capability probe timed out"))),
-                    None => Err(Error::other("cross-pool fence fleet capability notification system is unavailable")),
-                }
+                timeout(
+                    REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                    notification_sys.probe_cross_pool_fence_fleet(&topology_fingerprint),
+                )
+                .await
+                .unwrap_or_else(|_| Err(Error::other("cross-pool fence fleet capability probe timed out")))
            };
            let recovery_export_probe = async {
-                match notification_sys.as_ref() {
-                    Some(notification_sys) => timeout(
-                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
-                        notification_sys.probe_ilm_recovery_export_fleet(&topology_fingerprint),
-                    )
-                    .await
-                    .unwrap_or_else(|_| Err(Error::other("ILM recovery export fleet capability probe timed out"))),
-                    None => Err(Error::other("ILM recovery export fleet capability notification system is unavailable")),
-                }
+                timeout(
+                    REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                    notification_sys.probe_ilm_recovery_export_fleet(&topology_fingerprint),
+                )
+                .await
+                .unwrap_or_else(|_| Err(Error::other("ILM recovery export fleet capability probe timed out")))
            };
            let transition_transaction_compaction_probe = async {
-                match notification_sys.as_ref() {
-                    Some(notification_sys) => timeout(
-                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
-                        notification_sys.probe_transition_transaction_compaction_fleet(&topology_fingerprint),
-                    )
-                    .await
-                    .unwrap_or_else(|_| Err(Error::other("transition transaction compaction fleet capability probe timed out"))),
-                    None => Err(Error::other(
-                        "transition transaction compaction fleet capability notification system is unavailable",
-                    )),
-                }
+                timeout(
+                    REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                    notification_sys.probe_transition_transaction_compaction_fleet(&topology_fingerprint),
+                )
+                .await
+                .unwrap_or_else(|_| Err(Error::other("transition transaction compaction fleet capability probe timed out")))
            };
            let (result, fence_probe, recovery_export_result, transition_transaction_compaction_result) = tokio::join!(
                remote_version_state_probe,
@@ -1299,13 +1341,7 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
                .unwrap_or_else(std::sync::PoisonError::into_inner)
                .topology_conflict;
            if topology_conflict {
-                revoke_fleet_capability_proof(remote_version_state_fleet_proof_slot());
-                revoke_fleet_capability_proof(cross_pool_fence_fleet_proof_slot());
-                revoke_fleet_capability_proof(tier_delete_journal_fleet_proof_slot());
-                revoke_fleet_capability_proof(decommission_target_fence_fleet_proof_slot());
-                revoke_fleet_capability_proof(legacy_transition_state_reconcile_fleet_proof_slot());
-                revoke_fleet_capability_proof(ilm_recovery_export_fleet_proof_slot());
-                revoke_fleet_capability_proof(transition_transaction_compaction_fleet_proof_slot());
+                revoke_all_fleet_capability_proofs();
            } else if let Some(err) = publish_fleet_capability_probe_result(
                remote_version_state_fleet_proof_slot(),
                &topology_fingerprint,
@@ -1428,9 +1464,11 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
 }

 pub async fn new_global_notification_sys(eps: EndpointServerPools) -> Result<()> {
-    let _ = GLOBAL_NOTIFICATION_SYS
-        .set(Arc::new(NotificationSys::new(eps).await))
-        .map_err(|_| Error::other("init notification_sys fail"));
+    if GLOBAL_NOTIFICATION_SYS.set(Arc::new(NotificationSys::new(eps).await)).is_ok() {
+        // `notify_one` stores a permit, so a probe that checks availability
+        // just before this publication still wakes without losing the signal.
+        NOTIFICATION_SYS_PUBLISHED.notify_one();
+    }
    Ok(())
 }

@@ -3905,6 +3943,49 @@ mod tests {
        stats
    }

+    /// rustfs/rustfs#8014: IAM finalization starts the fleet probe before the
+    /// notification system is published. The first probe therefore fails
+    /// closed, and a single node used to wait a full probe interval before the
+    /// durable quota capability appeared, even though `/health` was already
+    /// reporting ok. The probe must retry promptly while it waits for startup
+    /// to publish the notification system.
+    #[tokio::test]
+    async fn fleet_probe_publishes_promptly_once_the_notification_system_appears() {
+        if REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get().is_some() || GLOBAL_NOTIFICATION_SYS.get().is_some() {
+            // Process-wide state was bound by another test in this binary. The
+            // startup ordering under test needs a fresh process; nextest
+            // (the authoritative runner) always provides one.
+            eprintln!("skipping: fleet probe globals already bound in this process");
+            return;
+        }
+
+        // Share the fingerprint every other proof-installing test in this crate
+        // binds, so the plain `cargo test` fallback cannot see two topologies.
+        start_remote_version_state_fleet_probe("object-transaction-fencing-test".to_string());
+        sleep(Duration::from_millis(300)).await;
+        assert!(
+            acquire_cross_pool_fence_fleet_proof().is_none(),
+            "no capability proof may exist before the notification system is published"
+        );
+
+        let published_at = Instant::now();
+        new_global_notification_sys(EndpointServerPools::default())
+            .await
+            .expect("single-node notification system initializes");
+
+        let observed = timeout(Duration::from_secs(2), async {
+            while acquire_cross_pool_fence_fleet_proof().is_none() {
+                sleep(Duration::from_millis(20)).await;
+            }
+        })
+        .await;
+        assert!(
+            observed.is_ok(),
+            "single-node cross-pool fence proof should publish within 2s of the notification system, waited {:?}",
+            published_at.elapsed()
+        );
+    }
+
    #[test]
    fn merging_peer_rings_sums_each_tier_without_double_counting() {
        let mut cluster = DailyAllTierStats::from([("WARM".to_string(), ring(10))]);
@@ -7,8 +7,9 @@ use super::meta::{
    validate_start_rebalance_state,
 };
 use super::worker::{
-    resolve_rebalance_bucket_result, resolve_rebalance_meta_save_result, resolve_rebalance_save_task_result,
-    resolve_rebalance_terminal_error, send_rebalance_done_signal,
+    rebalance_max_attempts, resolve_rebalance_bucket_result, resolve_rebalance_meta_save_result,
+    resolve_rebalance_save_task_result, resolve_rebalance_terminal_error, retry_rebalance_metadata_access,
+    send_rebalance_done_signal,
 };
 use super::{
    EVENT_REBALANCE_BUCKET, EVENT_REBALANCE_STATE, LOG_COMPONENT_ECSTORE, LOG_SUBSYSTEM_REBALANCE,
@@ -449,13 +450,14 @@ impl ECStore {
                        };

                        if terminal_state_present {
-                            if let Err(err) = store
-                                .save_rebalance_stats_inner(
+                            if let Err(err) = retry_rebalance_metadata_access(None, rebalance_max_attempts(), || {
+                                store.save_rebalance_stats_inner(
                                    pool_index,
                                    RebalSaveOpt::Stats,
                                    Some(save_rebalance_id.as_ref()),
                                )
-                                .await
+                            })
+                            .await
                            {
                                let mut rebalance_meta = store.rebalance_meta.write().await;
                                *rebalance_meta = previous_meta;
@@ -475,9 +477,10 @@ impl ECStore {
                }

                if !terminal_state_saved
-                    && let Err(err) = store
-                        .save_rebalance_stats_for_id(pool_index, RebalSaveOpt::Stats, save_rebalance_id.as_ref())
-                        .await
+                    && let Err(err) = retry_rebalance_metadata_access(None, rebalance_max_attempts(), || {
+                        store.save_rebalance_stats_for_id(pool_index, RebalSaveOpt::Stats, save_rebalance_id.as_ref())
+                    })
+                    .await
                {
                    let wrapped = Error::other(format!("rebalance save_task stats save failed for pool {pool_index}: {err}"));
                    error!("{} err: {:?}", msg, wrapped);
@@ -109,7 +109,13 @@ pub(super) fn resolve_rebalance_save_task_result(
 }

 pub(super) fn resolve_rebalance_meta_save_result(result: Result<()>, stage: &str) -> Result<()> {
-    result.map_err(|err| Error::other(format!("rebalance meta save failed during {stage}: {err}")))
+    // Keep the source error reachable: the metadata retry policy classifies
+    // transient lock timeouts by inspecting the source chain, so collapsing the
+    // failure into a plain string here would make that retry a no-op.
+    result.map_err(|err| {
+        let rendered = format!("rebalance meta save failed during {stage}: {err}");
+        crate::data_movement::data_movement_context_error(rendered, err)
+    })
 }

 pub(super) fn rebalance_meta_lock_error(err: rustfs_lock::LockError, mode: &'static str) -> Error {
@@ -739,6 +745,29 @@ mod error_source_tests {
        }
    }

+    #[tokio::test]
+    async fn rebalance_metadata_retry_engages_for_wrapped_meta_save_lock_timeout() {
+        let mut attempts = 0;
+        let result = retry_rebalance_metadata_access(None, 3, || {
+            attempts += 1;
+            std::future::ready(resolve_rebalance_meta_save_result(
+                Err(rebalance_meta_lock_error(
+                    rustfs_lock::LockError::timeout(".rustfs.sys/rebalance.bin@latest", Duration::from_secs(5)),
+                    "write",
+                )),
+                "save_rebalance_stats for pool 0 opt Stats",
+            ))
+        })
+        .await;
+
+        assert_eq!(attempts, 3, "a wrapped meta save lock timeout must stay retryable");
+        let err = result.expect_err("persistent lock contention must not become success");
+        assert!(matches!(
+            rebalance_error_source(&err),
+            Error::Lock(rustfs_lock::LockError::Timeout { .. })
+        ));
+    }
+
    #[tokio::test]
    async fn rebalance_metadata_retry_cancels_a_pending_attempt() {
        struct DropProbe(Arc<std::sync::atomic::AtomicUsize>);
@@ -37,7 +37,7 @@ use rustfs_s3_client::{
    api_s3_datatypes::ListVersionsResult,
    credentials::{Credentials, SignatureType, Static, Value},
    provider_versions::validate_remote_version_id,
-    transition_api::{BucketLookupType, Options, TransitionClient, TransitionCore},
+    transition_api::{BucketLookupType, ObjectInfo, Options, TransitionClient, TransitionCore},
    transition_api::{ReadCloser, ReaderImpl},
 };
 use rustfs_utils::egress::validate_outbound_url;
@@ -278,19 +278,7 @@ impl WarmBackendS3 {
                let mut stat_opts = GetObjectOptions::default();
                stat_opts.version_id.clone_from(&version.version_id);
                let info = self.client.stat_object(&self.bucket, &remote_object, &stat_opts).await?;
-                let mut metadata = info.user_metadata;
-                for (name, value) in &info.metadata {
-                    if (name
-                        .as_str()
-                        .starts_with(rustfs_utils::http::metadata_compat::RUSTFS_INTERNAL_PREFIX)
-                        || name
-                            .as_str()
-                            .starts_with(rustfs_utils::http::metadata_compat::MINIO_INTERNAL_PREFIX))
-                        && let Ok(value) = value.to_str()
-                    {
-                        metadata.insert(name.as_str().to_string(), value.to_string());
-                    }
-                }
+                let metadata = transition_candidate_metadata(info);
                if transition_candidate_metadata_matches(&metadata, identity)? {
                    if matched_version.is_some() {
                        return Ok(TransitionCandidateProbe::Ambiguous);
@@ -313,6 +301,53 @@ impl WarmBackendS3 {
            advance_version_markers(&mut key_marker, &mut version_id_marker, &versions)?;
        }
    }
+
+    async fn probe_unversioned_transition_candidate_identity(
+        &self,
+        object: &str,
+        identity: TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        let remote_object = self.get_dest(object);
+        match self
+            .client
+            .stat_object(&self.bucket, &remote_object, &GetObjectOptions::default())
+            .await
+        {
+            Ok(info) => {
+                let metadata = transition_candidate_metadata(info);
+                if transition_candidate_metadata_matches(&metadata, identity)? {
+                    Ok(TransitionCandidateProbe::UnversionedPresent)
+                } else {
+                    Ok(TransitionCandidateProbe::Unsupported)
+                }
+            }
+            Err(err) => {
+                let response = to_error_response(&err);
+                if response.code == S3ErrorCode::NoSuchKey {
+                    Ok(TransitionCandidateProbe::Missing)
+                } else {
+                    Err(err)
+                }
+            }
+        }
+    }
+}
+
+fn transition_candidate_metadata(info: ObjectInfo) -> HashMap<String, String> {
+    let mut metadata = info.user_metadata;
+    for (name, value) in &info.metadata {
+        if (name
+            .as_str()
+            .starts_with(rustfs_utils::http::metadata_compat::RUSTFS_INTERNAL_PREFIX)
+            || name
+                .as_str()
+                .starts_with(rustfs_utils::http::metadata_compat::MINIO_INTERNAL_PREFIX))
+            && let Ok(value) = value.to_str()
+        {
+            metadata.insert(name.as_str().to_string(), value.to_string());
+        }
+    }
+    metadata
 }

 fn transition_candidate_metadata_matches(
@@ -525,6 +560,13 @@ mod tests {
    }

    async fn scripted_probe_fixture(responses: Vec<String>) -> Option<(WarmBackendS3, tokio::task::JoinHandle<Vec<String>>)> {
+        scripted_probe_fixture_for_tier(responses, "s3").await
+    }
+
+    async fn scripted_probe_fixture_for_tier(
+        responses: Vec<String>,
+        tier_type: &str,
+    ) -> Option<(WarmBackendS3, tokio::task::JoinHandle<Vec<String>>)> {
        let listener = match tokio::net::TcpListener::bind("127.0.0.1:0").await {
            Ok(listener) => listener,
            Err(err) if err.kind() == std::io::ErrorKind::PermissionDenied => return None,
@@ -571,7 +613,7 @@ mod tests {
                    max_retries: 1,
                    ..Default::default()
                },
-                "s3",
+                tier_type,
            )
            .await
            .expect("fixture client should build"),
@@ -675,6 +717,37 @@ mod tests {
        assert!(requests[8].to_ascii_lowercase().contains("?versionid=historical-version"));
    }

+    #[tokio::test]
+    async fn unversioned_candidate_reconciler_uses_head_without_version_apis() {
+        let identity = candidate_identity();
+        let transaction_id = identity.transaction_id;
+        let destination_id = rustfs_utils::crypto::hex(identity.destination_id);
+        let response = format!(
+            "HTTP/1.1 200 OK\r\nContent-Length: 0\r\nx-amz-meta-x-rustfs-internal-transition-transaction-id: {transaction_id}\r\nx-amz-meta-x-rustfs-internal-transition-tier-destination-id: {destination_id}\r\nConnection: close\r\n\r\n"
+        );
+        let Some((backend, fixture)) = scripted_probe_fixture_for_tier(vec![response], "r2").await else {
+            return;
+        };
+
+        assert_eq!(
+            crate::services::tier::warm_backend::TransitionCandidateReconciler::probe_transition_candidate_for(
+                &backend,
+                "archive/object",
+                identity,
+            )
+            .await
+            .expect("R2 candidate probe should use the current unversioned object"),
+            TransitionCandidateProbe::UnversionedPresent
+        );
+        let requests = fixture.await.expect("R2 candidate fixture should finish");
+        assert_eq!(requests.len(), 1);
+        let request = requests[0].to_ascii_lowercase();
+        assert!(request.starts_with("head /bucket/archive/object"));
+        assert!(!request.contains("versioning"));
+        assert!(!request.contains("versions"));
+        assert!(!request.contains("versionid="));
+    }
+
    fn legacy_probe_xml_response(body: &str) -> String {
        format!(
            "HTTP/1.1 200 OK\r\nContent-Type: application/xml\r\nContent-Length: {}\r\nConnection: close\r\n\r\n{body}",
@@ -1089,6 +1162,10 @@ impl TransitionCandidateReconciler for WarmBackendS3 {
        object: &str,
        identity: TransitionCandidateIdentity,
    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        let capabilities = self.client.provider_version_capabilities();
+        if !capabilities.bucket_versioning_state && !capabilities.list_object_versions {
+            return self.probe_unversioned_transition_candidate_identity(object, identity).await;
+        }
        let bucket_versioning = self.remote_bucket_versioning().await?;
        self.probe_transition_candidate_identity(object, identity, bucket_versioning)
            .await
@@ -1163,13 +1163,10 @@ pub(in crate::set_disk) struct ReadRepairHealSubmission<'a> {
    pub(in crate::set_disk) set_index: usize,
    pub(in crate::set_disk) part_number: Option<usize>,
    pub(in crate::set_disk) reason: &'static str,
-    /// Durable MRF journal intent to file alongside the read-repair request
-    /// (backlog#1894 axis A): the intent kind plus its native `Uuid`
-    /// version id (the submission's string form stays display-only). Bound
-    /// to the reservation — the intent is only delivered when this sighting
-    /// wins the dedup TTL, so a burst of reads failing on the same object
-    /// books exactly one journal record instead of one per retry. `None`
-    /// keeps the historical no-intent behavior.
+    /// Best-effort ingress for the durable MRF consumer. Its identity coalescer
+    /// is independent of the heal reservation: an earlier metadata repair must
+    /// not suppress newly observed payload damage, and rejected ingress must
+    /// remain retryable. Channel admission does not prove checkpoint commit.
    pub(in crate::set_disk) mrf_intent: Option<(rustfs_common::mrf_channel::MrfKind, Option<uuid::Uuid>)>,
 }

@@ -1224,6 +1221,27 @@ pub(in crate::set_disk) async fn submit_read_repair_heal_with_submitter(
        mrf_intent,
    } = submission;

+    if let Some((kind, version_uuid)) = mrf_intent
+        && let (Ok(pool_index), Ok(set_index)) = (u32::try_from(pool_index), u32::try_from(set_index))
+    {
+        let scope = rustfs_common::mrf_channel::MrfScope { pool_index, set_index };
+        let ingress = rustfs_common::mrf_channel::try_send_mrf_intent_typed(kind, bucket, object, version_uuid, Some(scope));
+        debug!(
+            event = EVENT_SET_DISK_READ,
+            component = LOG_COMPONENT_ECSTORE,
+            subsystem = LOG_SUBSYSTEM_SET_DISK,
+            state = "mrf_ingress",
+            mrf_kind = ?kind,
+            bucket,
+            object,
+            version_id,
+            pool_index,
+            set_index,
+            result = ?ingress,
+            "Read-repair MRF ingress result"
+        );
+    }
+
    let Some(dedup_key) = reserve_read_repair_heal(bucket, object, version_id, pool_index, set_index).await else {
        record_read_repair_dedup("duplicate");
        debug!(
@@ -1233,15 +1251,6 @@ pub(in crate::set_disk) async fn submit_read_repair_heal_with_submitter(
        return;
    };

-    // Reservation won: this sighting owns the repair records for the object,
-    // including the durable journal intent when the caller asked for one.
-    if let Some((kind, version_uuid)) = mrf_intent
-        && let (Ok(pool_index), Ok(set_index)) = (u32::try_from(pool_index), u32::try_from(set_index))
-    {
-        let scope = rustfs_common::mrf_channel::MrfScope { pool_index, set_index };
-        let _ = rustfs_common::mrf_channel::try_send_mrf_intent_typed(kind, bucket, object, version_uuid, Some(scope));
-    }
-
    let mut request = rustfs_heal_contracts::heal_channel::create_heal_request_with_options(
        bucket.to_string(),
        Some(object.to_string()),
@@ -12961,7 +12970,7 @@ mod tests {

    #[tokio::test]
    #[serial_test::serial]
-    async fn mrf_intent_is_filed_once_per_read_repair_reservation() {
+    async fn mrf_ingress_is_coalesced_independently_of_read_repair_reservations() {
        // Serial: owns the process-global MRF channel for this test binary
        // (same key as the other channel-owning tests above).
        let bucket = format!("mrf-intent-bucket-{}", Uuid::new_v4());
@@ -12982,8 +12991,13 @@ mod tests {
            }
        }

-        // First sighting wins the reservation: the journal intent is filed
-        // synchronously before the admission task is spawned.
+        // An earlier metadata/missing-shard admission has no payload MRF intent.
+        let mut metadata_submission = intent_submission(&bucket, &object);
+        metadata_submission.mrf_intent = None;
+        submit_read_repair_heal_with_submitter(metadata_submission, accepted_read_repair_submitter).await;
+        assert!(receiver.try_recv().is_err());
+
+        // Newly observed corruption must reach MRF despite that reservation.
        submit_read_repair_heal_with_submitter(intent_submission(&bucket, &object), accepted_read_repair_submitter).await;
        let first = receiver.try_recv().expect("first sighting must file exactly one MRF intent");
        assert_eq!(*first.bucket, bucket);
@@ -12993,6 +13007,20 @@ mod tests {
        // second journal record.
        submit_read_repair_heal_with_submitter(intent_submission(&bucket, &object), accepted_read_repair_submitter).await;
        assert!(receiver.try_recv().is_err(), "duplicate sighting must not file another MRF intent");
+
+        let retry_object = format!("retry-object-{}", Uuid::new_v4());
+        rustfs_common::mrf_channel::set_mrf_delivery_enabled(false);
+        submit_read_repair_heal_with_submitter(intent_submission(&bucket, &retry_object), accepted_read_repair_submitter).await;
+        assert!(receiver.try_recv().is_err());
+        rustfs_common::mrf_channel::set_mrf_delivery_enabled(true);
+        submit_read_repair_heal_with_submitter(intent_submission(&bucket, &retry_object), accepted_read_repair_submitter).await;
+        let retried = receiver
+            .try_recv()
+            .expect("a heal reservation cannot suppress a previously rejected MRF intent");
+        assert_eq!(*retried.object, retry_object);
+        rustfs_common::mrf_channel::release_mrf_intent(&first);
+        rustfs_common::mrf_channel::release_mrf_intent(&retried);
+        rustfs_common::mrf_channel::set_mrf_delivery_enabled(false);
    }

    #[tokio::test]
@@ -874,6 +874,7 @@ mod ctx;
 mod metadata;
 mod ops;
 pub(crate) use ops::bucket::BucketInfoQuorum;
+pub(crate) use ops::heal::AbsenceProofRequest;
 #[cfg(test)]
 pub(crate) use ops::heal::DanglingDeleteFailure;
 pub(crate) use ops::heal::HealedObjectAbsence;
@@ -3876,6 +3877,9 @@ pub struct SetDisks {
    pub format: FormatV3,
    #[allow(dead_code, reason = "asserted by this file's tests (backlog#1823)")]
    disk_health_cache: Arc<RwLock<Vec<Option<DiskHealthEntry>>>>,
+    /// Coalesce bounded GET probes of missing remote slots. Shared by set clones;
+    /// the timestamp also limits repeated probes while a peer remains absent.
+    read_reconnect: Arc<tokio::sync::Mutex<Option<tokio::time::Instant>>>,
    get_object_metadata_cache: moka::future::Cache<GetObjectMetadataCacheKey, Arc<GetObjectMetadataCacheEntry>>,
    get_object_metadata_cache_hash_builder: std::collections::hash_map::RandomState,
    get_object_metadata_cache_generations: Arc<[AtomicU64]>,
@@ -3914,6 +3918,72 @@ pub struct SetDisks {
    >,
 }

+fn marker_purge_receipt_path(
+    bucket: &str,
+    object: &str,
+    version: Uuid,
+    purge: &rustfs_common::mrf_channel::MrfDeleteMarkerPurge,
+) -> String {
+    let object_hash = rustfs_utils::crypto::hex(Sha256::digest(object.as_bytes()));
+    let marker_hash = rustfs_utils::crypto::hex(purge.marker_identity);
+    format!(
+        "marker-purge-receipts/{}/{}/{}-{}-{}.json",
+        bucket, purge.bucket_incarnation_id, version, object_hash, marker_hash
+    )
+}
+
+impl SetDisks {
+    /// Persist a quorum DELETE receipt outside the object marker itself. The
+    /// receipt is written to the metadata bucket so it survives a node restart
+    /// while an inaccessible member still holds the stale marker.
+    pub(crate) async fn persist_marker_purge_receipt(
+        &self,
+        bucket: &str,
+        object: &str,
+        version: Uuid,
+        purge: &rustfs_common::mrf_channel::MrfDeleteMarkerPurge,
+    ) -> bool {
+        let path = marker_purge_receipt_path(bucket, object, version, purge);
+        let api = Arc::new(self.clone());
+        crate::config::com::save_config_with_opts(
+            api,
+            &path,
+            b"rustfs-marker-purge-receipt-v1".to_vec(),
+            &ObjectOptions {
+                max_parity: true,
+                no_lock: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .is_ok()
+    }
+
+    pub(crate) async fn has_marker_purge_receipt(
+        &self,
+        bucket: &str,
+        object: &str,
+        version: Uuid,
+        purge: &rustfs_common::mrf_channel::MrfDeleteMarkerPurge,
+    ) -> bool {
+        let path = marker_purge_receipt_path(bucket, object, version, purge);
+        crate::config::com::read_config_limited_preserve_empty(Arc::new(self.clone()), &path, 128)
+            .await
+            .is_ok()
+    }
+
+    pub(crate) async fn consume_marker_purge_receipt(
+        &self,
+        bucket: &str,
+        object: &str,
+        version: Uuid,
+        purge: &rustfs_common::mrf_channel::MrfDeleteMarkerPurge,
+    ) {
+        let path = marker_purge_receipt_path(bucket, object, version, purge);
+        let _ = crate::config::com::delete_config_no_lock(Arc::new(self.clone()), &path).await;
+    }
+}
+
 /// Read every physical copy before selecting a version quorum. A minority
 /// legacy record is still evidence and must not disappear behind a majority
 /// not-found result. Only an explicit file/volume absence produces `None`;
@@ -4676,6 +4746,7 @@ impl SetDisks {
            format,
            set_endpoints,
            disk_health_cache: Arc::new(RwLock::new(Vec::new())),
+            read_reconnect: Arc::new(tokio::sync::Mutex::new(None)),
            get_object_metadata_cache: moka::future::Cache::builder()
                .max_capacity(get_object_metadata_cache_max_entries() as u64)
                .time_to_live(GET_OBJECT_METADATA_CACHE_TTL)
@@ -6406,6 +6477,28 @@ async fn verify_inline_part_bitrot(meta: &FileInfo) -> disk::error::Result<()> {
        .map_err(|_| DiskError::FileCorrupt)
 }

+fn validate_deep_scan_results(results: &[usize], expected_parts: usize) -> disk::error::Result<()> {
+    if results.len() != expected_parts {
+        return Err(DiskError::other(format!(
+            "incomplete deep scan result: expected {expected_parts} parts, received {}",
+            results.len()
+        )));
+    }
+    for (part, status) in results.iter().enumerate() {
+        match *status {
+            CHECK_PART_SUCCESS | CHECK_PART_FILE_NOT_FOUND | CHECK_PART_FILE_CORRUPT => {}
+            CHECK_PART_DISK_NOT_FOUND => return Err(DiskError::DiskNotFound),
+            crate::disk::CHECK_PART_VOLUME_NOT_FOUND => return Err(DiskError::VolumeNotFound),
+            _ => {
+                return Err(DiskError::other(format!(
+                    "incomplete deep scan result: part {part} has unverified status {status}"
+                )));
+            }
+        }
+    }
+    Ok(())
+}
+
 /// disks_with_all_partsv2 is a corrected version based on Go implementation.
 /// It sets partsMetadata and onlineDisks when xl.meta is inexistant/corrupted or outdated.
 /// It also checks if the status of each part (corrupted, missing, ok) in each drive.
@@ -6626,9 +6719,13 @@ async fn disks_with_all_parts(
            // it needs healing too.
            match disk.verify_file(bucket, object, meta).await {
                Ok(v) => {
+                    validate_deep_scan_results(&v.results, latest_meta.parts.len())?;
                    verify_resp = v;
                }
                Err(err) => {
+                    if !matches!(err, DiskError::FileNotFound | DiskError::FileVersionNotFound | DiskError::FileCorrupt) {
+                        return Err(err);
+                    }
                    debug!(
                        event = EVENT_SET_DISK_HEAL,
                        component = LOG_COMPONENT_ECSTORE,
@@ -11042,6 +11139,50 @@ mod tests {
        assert_eq!(conv_part_err_to_int(&Some(other_err)), CHECK_PART_UNKNOWN); // Other errors should return UNKNOWN, not SUCCESS
    }

+    #[test]
+    fn deep_scan_results_accept_only_complete_verified_or_repairable_parts() {
+        for (results, expected_parts) in [
+            (vec![], 0),
+            (vec![CHECK_PART_SUCCESS], 1),
+            (vec![CHECK_PART_FILE_NOT_FOUND], 1),
+            (vec![CHECK_PART_FILE_CORRUPT], 1),
+            (vec![CHECK_PART_SUCCESS, CHECK_PART_FILE_NOT_FOUND, CHECK_PART_FILE_CORRUPT], 3),
+        ] {
+            validate_deep_scan_results(&results, expected_parts)
+                .expect("complete results must allow healthy or repairable parts");
+        }
+    }
+
+    #[test]
+    fn deep_scan_results_reject_unknown_and_incomplete_observations() {
+        for (results, expected_parts) in [
+            (vec![CHECK_PART_UNKNOWN], 1),
+            (vec![usize::MAX], 1),
+            (vec![CHECK_PART_SUCCESS, CHECK_PART_UNKNOWN], 2),
+            (vec![], 1),
+            (vec![CHECK_PART_SUCCESS], 2),
+            (vec![CHECK_PART_SUCCESS, CHECK_PART_SUCCESS], 1),
+            (vec![CHECK_PART_SUCCESS], 0),
+        ] {
+            let error =
+                validate_deep_scan_results(&results, expected_parts).expect_err("unverified parts must not become healthy");
+            assert!(matches!(error, DiskError::Io(_)), "an incomplete observation is not proven corruption");
+            assert!(error.to_string().contains("incomplete deep scan result"));
+        }
+    }
+
+    #[test]
+    fn deep_scan_results_preserve_disk_and_volume_failures() {
+        for (status, expected_error) in [
+            (CHECK_PART_DISK_NOT_FOUND, DiskError::DiskNotFound),
+            (CHECK_PART_VOLUME_NOT_FOUND, DiskError::VolumeNotFound),
+        ] {
+            let error = validate_deep_scan_results(&[CHECK_PART_SUCCESS, status], 2)
+                .expect_err("an unavailable part cannot certify a healthy disk");
+            assert_eq!(error, expected_error);
+        }
+    }
+
    #[test]
    fn test_has_part_err() {
        // Test checking for part errors
@@ -13561,6 +13702,7 @@ mod tests {
                true,
                false,
                false,
+                false,
                GET_OBJECT_PATH_LEGACY_DUPLEX,
                GET_CODEC_STREAMING_OBJECT_CLASS_PLAIN_SINGLE_PART,
                metrics_size_bucket,
@@ -13674,6 +13816,7 @@ mod tests {
                true,
                false,
                false,
+                false,
                GET_OBJECT_PATH_LEGACY_DUPLEX,
                GET_CODEC_STREAMING_OBJECT_CLASS_PLAIN_SINGLE_PART,
                metrics_size_bucket,
@@ -141,6 +141,7 @@ pub(in crate::set_disk::ops) async fn verify_written_bitrot_shards(
 mod tests {
    use super::super::object::hermetic_set_disks_support::hermetic_set_disks_for_pool_with_default_parity;
    use super::*;
+    use crate::storage_api_contracts::object::ObjectIO as _;

    async fn encode_streaming_shard(data: &[u8], shard_size: usize) -> Bytes {
        let mut writer = coding::BitrotWriter::new(Cursor::new(Vec::new()), shard_size, HashAlgorithm::HighwayHash256S);
@@ -295,4 +296,19 @@ mod tests {
            assert!(err.to_string().contains("trailing data"));
        }
    }
+
+    #[tokio::test]
+    async fn no_parity_put_round_trip_large_stream() {
+        let (_temp_dirs, disks, set_disks) = hermetic_set_disks_for_pool_with_default_parity(1, 0, 0).await;
+        let bucket = "no-parity-put-round-trip";
+        for disk in &disks {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+        let payload = vec![0x5a; 13 * 1024 * 1024 + 727_213];
+        let mut reader = crate::set_disk::PutObjReader::from_vec(payload);
+        set_disks
+            .put_object(bucket, "object", &mut reader, &crate::set_disk::ObjectOptions::default())
+            .await
+            .expect("single-disk no-parity PUT should pass self-verify");
+    }
 }
@@ -38,6 +38,7 @@ use crate::disk::DiskAPI;
 use crate::disk::local::{DELETE_DATA_DIR_MARKER_PREFIX, metadata_less_part_file};
 use crate::io_support::bitrot::object_mmap_read_enabled;
 use crate::storage_api_contracts::namespace::NamespaceLocking as _;
+use rustfs_common::mrf_channel::MrfDeleteMarkerPurge;
 use rustfs_common::trace_bus::{TraceEvent, TraceFunc, TraceKind, trace_emit};
 use tracing::trace;

@@ -55,6 +56,13 @@ pub(crate) struct HealedObjectAbsence {
    pub removed: bool,
 }

+/// Conditions under which a missing object may become a storage-owned proof.
+#[derive(Clone, Copy)]
+pub(crate) struct AbsenceProofRequest<'a> {
+    pub retirement: Option<&'a MarkerRetirementContext<'a>>,
+    pub allow_unversioned: bool,
+}
+
 fn is_unavailable_heal_rpc(error: &DiskError) -> bool {
    let DiskError::Io(error) = error else {
        return false;
@@ -846,7 +854,7 @@ impl SetDisks {
        );

        let disks = self.get_disks_internal().await;
-        let bucket_heal_scope = crate::store::bucket_heal_scope(bucket);
+        let bucket_heal_scope = crate::store::bucket_heal_scope_for_object(bucket, object);
        if let Some(scope) = &bucket_heal_scope {
            scope.check()?;
        }
@@ -971,12 +979,25 @@ impl SetDisks {
                result.parity_blocks = result.disk_count - read_quorum as usize;
                result.data_blocks = read_quorum as usize;

-                let ((mut online_disks, quorum_mod_time, quorum_etag), disk_len) = {
+                let ((quorum_disks, quorum_mod_time, quorum_etag), disk_len) = {
                    let disks = self.disks.read().await;
                    let disk_len = disks.len();
                    (Self::list_online_disks(&disks, &parts_metadata, &errs, read_quorum as usize), disk_len)
                };

+                // A deep heal must inspect every reachable member, not only the
+                // quorum that agreed on the canonical metadata. A returning disk
+                // can legitimately carry an older xl.meta (or an explicit-version
+                // absence) while still being the exact stale target that needs
+                // repair. Restricting the repair set to quorum members silently
+                // turns that stale member into `verified_healthy`/`unchanged`
+                // (backlog#2612, #2596). Normal scans retain the quorum fast path.
+                let mut online_disks = if opts.scan_mode == HealScanMode::Deep {
+                    self.get_disks_internal().await
+                } else {
+                    quorum_disks
+                };
+
                trace!(
                    event = EVENT_SET_DISK_HEAL,
                    component = LOG_COMPONENT_ECSTORE,
@@ -1449,7 +1470,12 @@ impl SetDisks {
                                    })
                                    .transpose()
                                    .map_err(DiskError::from)?;
-                                let till_offset = erasure.shard_file_offset(0, part.size, part.size);
+                                // This reader covers the whole encoded shard. A range
+                                // offset can exclude the final encoded byte when the
+                                // logical part ends inside a data block, allowing a
+                                // one-byte tail truncation to evade deep Heal.
+                                let till_offset = usize::try_from(erasure.shard_file_size(part.size as i64))
+                                    .map_err(|_| DiskError::FileCorrupt)?;
                                let use_mmap_read = object_mmap_read_enabled();

                                let mut readers = Vec::with_capacity(latest_disks.len());
@@ -1887,13 +1913,26 @@ impl SetDisks {
                        .zip(&errs)
                        .any(|(info, error)| error.is_none() && info.is_canonical_delete_marker())
                {
-                    let cleanup = match Self::retired_marker_candidate(&parts_metadata, &errs) {
-                        Ok(marker) => {
-                            self.remove_retired_marker(retirement, bucket, object, marker, opts, &disks)
-                                .await
-                        }
-                        Err(error) => Err(error),
-                    };
+                    // A current-generation delete marker can be left behind on
+                    // an inaccessible member after an acknowledged S3 DELETE.
+                    // That case has no bucket-retirement record, so evaluate its
+                    // stricter same-incarnation proof before the retired-marker
+                    // path. The proof requires an exact marker identity, at
+                    // least one replica already absent, and every target online.
+                    let cleanup =
+                        match Self::same_incarnation_marker_candidate(&parts_metadata, &errs, retirement.current_incarnation) {
+                            Ok(marker) => {
+                                self.remove_current_marker(retirement, bucket, object, marker, opts, &disks)
+                                    .await
+                            }
+                            Err(_) => match Self::retired_marker_candidate(&parts_metadata, &errs) {
+                                Ok(marker) => {
+                                    self.remove_retired_marker(retirement, bucket, object, marker, opts, &disks)
+                                        .await
+                                }
+                                Err(error) => Err(error),
+                            },
+                        };
                    let absent = cleanup.is_ok();
                    let item = self
                        .dangling_heal_result(FileInfo::default(), &errs, bucket, object, version_id, absent)
@@ -2007,6 +2046,41 @@ impl SetDisks {
        candidate.ok_or_else(|| DiskError::retired_marker_deferred("no exact marker candidate"))
    }

+    fn same_incarnation_marker_candidate(
+        metadata: &[FileInfo],
+        errors: &[Option<DiskError>],
+        current_incarnation: Option<Uuid>,
+    ) -> disk::error::Result<rustfs_filemeta::MetaDeleteMarker> {
+        let defer = DiskError::retired_marker_deferred;
+        let Some(current) = current_incarnation.filter(|id| !id.is_nil()) else {
+            return Err(defer("current bucket incarnation is unavailable"));
+        };
+        let mut candidate = None;
+        let mut missing_replica = false;
+        for (info, error) in metadata.iter().zip(errors) {
+            match error {
+                Some(DiskError::FileNotFound | DiskError::FileVersionNotFound) => {
+                    missing_replica = true;
+                    continue;
+                }
+                Some(_) => return Err(defer("not every replica is readable")),
+                None => {}
+            }
+            if !info.is_canonical_delete_marker() || info.delete_marker_incarnation() != Some(current) {
+                return Err(defer("marker is not a canonical current-generation marker"));
+            }
+            let marker = rustfs_filemeta::MetaDeleteMarker::from(info.clone());
+            if candidate.as_ref().is_some_and(|previous| previous != &marker) {
+                return Err(defer("surviving marker identities conflict"));
+            }
+            candidate = Some(marker);
+        }
+        if !missing_replica {
+            return Err(defer("marker is present on every replica"));
+        }
+        candidate.ok_or_else(|| defer("no exact current-generation marker candidate"))
+    }
+
    async fn remove_retired_marker(
        &self,
        context: &MarkerRetirementContext<'_>,
@@ -2045,6 +2119,72 @@ impl SetDisks {
        if context.lifecycle_guard.is_lock_lost() {
            return Err(defer("bucket lifecycle fence was lost"));
        }
+        let result = self
+            .remove_marker_exact_under_fence(bucket, object, marker, version, disks)
+            .await;
+        if context.lifecycle_guard.is_lock_lost() {
+            return Err(defer("bucket lifecycle fence was lost"));
+        }
+        result
+    }
+
+    async fn remove_current_marker(
+        &self,
+        context: &MarkerRetirementContext<'_>,
+        bucket: &str,
+        object: &str,
+        marker: rustfs_filemeta::MetaDeleteMarker,
+        opts: &HealOpts,
+        disks: &[Option<DiskStore>],
+    ) -> disk::error::Result<()> {
+        let defer = DiskError::retired_marker_deferred;
+        if opts.dry_run || !opts.remove {
+            return Err(defer("cleanup requires remove=true and dry_run=false"));
+        }
+        if disks.len() != self.set_drive_count || disks.iter().any(Option::is_none) {
+            return Err(defer("every target disk must be online"));
+        }
+        let Some(current) = context.current_incarnation.filter(|id| !id.is_nil()) else {
+            return Err(defer("current bucket incarnation is unavailable"));
+        };
+        if context.lifecycle_guard.is_lock_lost() {
+            return Err(defer("bucket lifecycle fence was lost"));
+        }
+        let incarnation = marker.into_fileinfo(bucket, object, false)?.delete_marker_incarnation();
+        if incarnation != Some(current) {
+            return Err(defer("marker is not from the current bucket incarnation"));
+        }
+        let Some(version) = marker.version_id.filter(|id| !id.is_nil()) else {
+            return Err(defer("cleanup requires an explicit non-null version"));
+        };
+        let marker_identity = marker.stable_identity();
+        let marker_bytes = marker.marshal_msg().map_err(DiskError::other)?;
+        let Some(purge) = MrfDeleteMarkerPurge::new(current, current, marker_identity, marker_bytes) else {
+            return Err(defer("marker purge proof payload is invalid"));
+        };
+        if !self.has_marker_purge_receipt(bucket, object, version, &purge).await {
+            return Err(defer("same-incarnation marker lacks an acknowledged delete receipt"));
+        }
+        let result = self
+            .remove_marker_exact_under_fence(bucket, object, marker, version, disks)
+            .await;
+        if context.lifecycle_guard.is_lock_lost() {
+            return Err(defer("bucket lifecycle fence was lost"));
+        }
+        if result.is_ok() {
+            self.consume_marker_purge_receipt(bucket, object, version, &purge).await;
+        }
+        result
+    }
+
+    async fn remove_marker_exact_under_fence(
+        &self,
+        bucket: &str,
+        object: &str,
+        marker: rustfs_filemeta::MetaDeleteMarker,
+        version: Uuid,
+        disks: &[Option<DiskStore>],
+    ) -> disk::error::Result<()> {
        let request = FileInfo {
            volume: bucket.to_owned(),
            name: object.to_owned(),
@@ -2085,13 +2225,14 @@ impl SetDisks {
        let same_targets = selected.len() == disks.len() && selected.iter().zip(disks).all(|(current, original)| {
            matches!((current, original), (Some(current), Some(original)) if std::sync::Arc::ptr_eq(current, original))
        });
-        if context.lifecycle_guard.is_lock_lost()
-            || !same_targets
+        if !same_targets
            || !errors
                .iter()
                .all(|error| matches!(error, Some(DiskError::FileNotFound | DiskError::FileVersionNotFound)))
        {
-            return Err(defer("complete marker absence could not be verified under the original fence"));
+            return Err(DiskError::retired_marker_deferred(
+                "complete marker absence could not be verified under the original target set",
+            ));
        }
        self.invalidate_get_object_metadata_cache(bucket, object).await;
        Ok(())
@@ -3012,8 +3153,18 @@ impl crate::storage_api_contracts::heal::HealOperations for SetDisks {
        version_id: &str,
        opts: &HealOpts,
    ) -> Result<(HealResultItem, Option<Error>)> {
-        self.heal_object_with_absence(bucket, object, version_id, opts, &mut None)
-            .await
+        self.heal_object_with_absence(
+            bucket,
+            object,
+            version_id,
+            opts,
+            &mut None,
+            AbsenceProofRequest {
+                retirement: None,
+                allow_unversioned: false,
+            },
+        )
+        .await
    }

    #[tracing::instrument(skip(self))]
@@ -3098,8 +3249,9 @@ impl SetDisks {
        version_id: &str,
        opts: &HealOpts,
        absence: &mut Option<HealedObjectAbsence>,
+        proof: AbsenceProofRequest<'_>,
    ) -> Result<(HealResultItem, Option<Error>)> {
-        self.heal_object_with_retirement(bucket, object, version_id, opts, absence, None)
+        self.heal_object_with_retirement(bucket, object, version_id, opts, absence, proof)
            .await
    }

@@ -3110,7 +3262,7 @@ impl SetDisks {
        version_id: &str,
        opts: &HealOpts,
        absence: &mut Option<HealedObjectAbsence>,
-        retirement: Option<&MarkerRetirementContext<'_>>,
+        proof: AbsenceProofRequest<'_>,
    ) -> Result<(HealResultItem, Option<Error>)> {
        *absence = None;
        let _write_lock_guard = if !opts.no_lock {
@@ -3176,7 +3328,7 @@ impl SetDisks {
                .await;
            // Check the lease after the final await before publishing the proof.
            if !opts.dry_run
-                && !version_id.is_empty()
+                && (proof.allow_unversioned || !version_id.is_empty())
                && !disks.is_empty()
                && disks.iter().all(Option::is_some)
                && errs
@@ -3205,7 +3357,7 @@ impl SetDisks {
                ExplicitVersionHeal {
                    opts: &inner_opts,
                    allow_regeneration: true,
-                    retirement,
+                    retirement: proof.retirement,
                },
                absence,
            )
@@ -3225,7 +3377,7 @@ impl SetDisks {
                            ExplicitVersionHeal {
                                opts: &inner_opts,
                                allow_regeneration: true,
-                                retirement,
+                                retirement: proof.retirement,
                            },
                            absence,
                        )
@@ -3277,7 +3429,7 @@ impl SetDisks {
 #[cfg(test)]
 mod heal_result_report_tests {
    use super::{
-        DanglingCheckPartsFailure, DanglingDeleteFailure, DanglingDeleteSafety, ReadRepairCommitFingerprint,
+        AbsenceProofRequest, DanglingCheckPartsFailure, DanglingDeleteFailure, DanglingDeleteSafety, ReadRepairCommitFingerprint,
        ReadRepairPauseScope, SetDisks, heal_writer_error_summary,
    };
    use super::{HEAL_RENAME_INCOMPLETE, HealRenameFailureScope, HealWriterFailureScope};
@@ -3288,7 +3440,9 @@ mod heal_result_report_tests {
    use crate::disk::{DiskAPI as _, DiskOption, DiskStore, RUSTFS_META_TMP_BUCKET, ReadOptions, STORAGE_FORMAT_FILE, new_disk};
    use crate::error::Error;
    use crate::object_api::{ObjectOptions, PutObjReader};
-    use crate::set_disk::ops::object::hermetic_set_disks_support::hermetic_set_disks_isolated;
+    use crate::set_disk::ops::object::hermetic_set_disks_support::{
+        hermetic_set_disks_for_pool_with_default_parity_isolated, hermetic_set_disks_isolated,
+    };
    use crate::storage_api_contracts::bucket::{BucketOperations as _, DeleteBucketOptions, MakeBucketOptions};
    use crate::storage_api_contracts::heal::HealOperations as _;
    use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations as _};
@@ -4069,6 +4223,634 @@ mod heal_result_report_tests {
        );
    }

+    #[tokio::test]
+    async fn deep_heal_rebuilds_a_stale_current_delete_marker_metadata_replica() {
+        let (_temp_dirs, disks, set) = hermetic_set_disks_isolated(4).await;
+        let bucket = "deep-heal-stale-current-marker";
+        let object = "object.bin";
+        for disk in &disks {
+            disk.make_volume(bucket).await.expect("test bucket should be created");
+        }
+
+        let versioned = ObjectOptions {
+            no_lock: true,
+            versioned: true,
+            ..Default::default()
+        };
+        let original = set
+            .put_object(bucket, object, &mut PutObjReader::from_vec(vec![0x41; 1024]), &versioned)
+            .await
+            .expect("versioned source object should be written");
+        let stale_metadata = disks[0]
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("pre-marker metadata should be readable");
+
+        let marker = set
+            .delete_object(bucket, object, versioned)
+            .await
+            .expect("versioned delete should create a current marker");
+        let marker_version = marker.version_id.expect("delete marker should have a version id");
+
+        // Simulate a member that was offline while the acknowledged marker was
+        // committed on the other members: its physical root still contains the
+        // pre-delete metadata and the original data version.
+        let mut stale_metadata = stale_metadata;
+        stale_metadata.fresh = true;
+        disks[0]
+            .write_metadata("", bucket, object, stale_metadata)
+            .await
+            .expect("stale member metadata should be restored for the fixture");
+        assert_eq!(
+            disks[0]
+                .read_version("", bucket, object, &marker_version.to_string(), &ReadOptions::default())
+                .await
+                .expect_err("stale member must not contain the current marker")
+                .to_string(),
+            "file version not found"
+        );
+
+        let (result, error) = set
+            .heal_object(
+                bucket,
+                object,
+                &marker_version.to_string(),
+                &HealOpts {
+                    no_lock: true,
+                    scan_mode: HealScanMode::Deep,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("deep heal should report the stale marker result");
+
+        assert!(error.is_none(), "deep heal should repair stale marker metadata: {error:?}");
+        assert_eq!(result.drives_healed(), Some(1));
+        let repaired = disks[0]
+            .read_version("", bucket, object, &marker_version.to_string(), &ReadOptions::default())
+            .await
+            .expect("rejoined member should contain the marker");
+        assert!(repaired.deleted, "repaired metadata must remain a delete marker");
+        assert_eq!(repaired.version_id, Some(marker_version));
+        assert!(original.version_id.is_some(), "source object should have a version id");
+    }
+
+    #[tokio::test]
+    async fn deep_heal_rebuilds_near_tail_truncated_part() {
+        let (temp_dirs, disks, set) = hermetic_set_disks_isolated(4).await;
+        let bucket = "deep-heal-near-tail-truncation";
+        let object = "object.bin";
+        for disk in &disks {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+
+        let expected_payload = vec![0x6d; 1024 * 1024];
+        set.put_object(
+            bucket,
+            object,
+            &mut PutObjReader::from_vec(expected_payload.clone()),
+            &ObjectOptions {
+                no_lock: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("source object should be written before shard truncation");
+        let source = disks[2]
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("source metadata should be readable");
+        let data_dir = source.data_dir.expect("non-inline source should have a data directory");
+        let truncated_part = temp_dirs[1]
+            .path()
+            .join(bucket)
+            .join(object)
+            .join(data_dir.to_string())
+            .join("part.1");
+        let original_len = tokio::fs::metadata(&truncated_part)
+            .await
+            .expect("target shard should exist")
+            .len();
+        assert!(original_len > 1, "test shard must be large enough to truncate");
+        let file = tokio::fs::OpenOptions::new()
+            .write(true)
+            .open(&truncated_part)
+            .await
+            .expect("target shard should be writable");
+        file.set_len(original_len - 1)
+            .await
+            .expect("target shard should be truncated");
+
+        let mut reader = set
+            .get_object_reader(bucket, object, None, Default::default(), &ObjectOptions::default())
+            .await
+            .expect("GET should remain readable after a one-byte shard truncation");
+        let mut read_back = Vec::new();
+        tokio::io::copy(&mut reader, &mut read_back)
+            .await
+            .expect("GET should reconstruct the truncated shard through EC");
+        assert_eq!(read_back, expected_payload, "EC GET must preserve the object bytes");
+
+        let (result, error) = set
+            .heal_object(
+                bucket,
+                object,
+                "",
+                &HealOpts {
+                    no_lock: true,
+                    scan_mode: HealScanMode::Deep,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("deep heal should finish after a one-byte tail truncation");
+
+        assert!(error.is_none(), "deep heal should recover the truncated shard: {error:?}");
+        assert_eq!(result.drives_healed(), Some(1));
+        assert_eq!(result.before.drives[1].state, DriveState::Corrupt.to_string());
+        assert_eq!(
+            tokio::fs::metadata(&truncated_part)
+                .await
+                .expect("repaired shard should exist")
+                .len(),
+            original_len,
+            "deep heal must restore the complete shard length"
+        );
+    }
+
+    #[derive(Debug)]
+    struct FailedIntegrityReadTransport {
+        error: DiskError,
+        reads: std::sync::atomic::AtomicUsize,
+    }
+
+    #[async_trait::async_trait]
+    impl crate::cluster::rpc::internode_data_transport::InternodeDataTransport for FailedIntegrityReadTransport {
+        async fn open_read(
+            &self,
+            _request: crate::cluster::rpc::internode_data_transport::ReadStreamRequest,
+        ) -> crate::disk::error::Result<crate::disk::FileReader> {
+            self.reads.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+            let error = match &self.error {
+                DiskError::FileCorrupt => DiskError::from(rustfs_rio::new_test_remote_file_corrupt_http_io_error()),
+                DiskError::FileNotFound => DiskError::from(rustfs_rio::new_test_remote_file_not_found_http_io_error()),
+                DiskError::VolumeNotFound => DiskError::from(rustfs_rio::new_test_remote_volume_not_found_http_io_error()),
+                error => error.clone(),
+            };
+            Err(error)
+        }
+
+        async fn open_write(
+            &self,
+            _request: crate::cluster::rpc::internode_data_transport::WriteStreamRequest,
+        ) -> crate::disk::error::Result<crate::disk::FileWriter> {
+            panic!("an integrity scan must not write remote data");
+        }
+
+        async fn open_walk_dir(
+            &self,
+            _request: crate::cluster::rpc::internode_data_transport::WalkDirStreamRequest,
+        ) -> crate::disk::error::Result<crate::disk::FileReader> {
+            panic!("an object integrity scan must not walk remote directories");
+        }
+
+        fn name(&self) -> &'static str {
+            "failed-integrity-read"
+        }
+
+        fn capabilities(&self) -> crate::cluster::rpc::internode_data_transport::InternodeDataTransportCapabilities {
+            crate::cluster::rpc::internode_data_transport::InternodeDataTransportCapabilities::tcp_http()
+        }
+    }
+
+    async fn assert_remote_integrity_read_failure(expected_error: DiskError) {
+        use crate::disk::{CHECK_PART_FILE_CORRUPT, CHECK_PART_FILE_NOT_FOUND, CHECK_PART_SUCCESS, CHECK_PART_UNKNOWN};
+        use crate::object_api::ShardIntegrityWriteMode;
+
+        let (dirs, disks, set) = hermetic_set_disks_for_pool_with_default_parity_isolated(4, 0, 2).await;
+        let bucket = "remote-integrity-read-failure";
+        let object = "object.bin";
+        for disk in &disks {
+            disk.make_volume(bucket).await.expect("create fixture bucket");
+        }
+        set.put_object(
+            bucket,
+            object,
+            &mut PutObjReader::from_vec(vec![0x6d; 1024 * 1024 + 123]),
+            &ObjectOptions {
+                shard_integrity_write_mode: Some(ShardIntegrityWriteMode::Protected),
+                no_lock: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("write protected external fixture");
+
+        let mut files = Vec::new();
+        let mut snapshots = Vec::new();
+        for (dir, disk) in dirs.iter().zip(&disks) {
+            let file = disk
+                .read_version("", bucket, object, "", &ReadOptions::default())
+                .await
+                .expect("read fixture metadata");
+            assert!(file.data.is_none(), "the remote test must open an external shard");
+            let object_dir = dir.path().join(bucket).join(object);
+            let part_path = object_dir
+                .join(file.data_dir.expect("external shard directory").to_string())
+                .join("part.1");
+            for path in [part_path, object_dir.join("xl.meta")] {
+                let bytes = tokio::fs::read(&path).await.expect("snapshot fixture before scan");
+                snapshots.push((path, bytes));
+            }
+            files.push(file);
+        }
+        let transport = Arc::new(FailedIntegrityReadTransport {
+            error: expected_error.clone(),
+            reads: std::sync::atomic::AtomicUsize::new(0),
+        });
+        let endpoint = Endpoint {
+            url: url::Url::parse("http://remote-integrity.invalid:9000/data/disk0").expect("fixture endpoint"),
+            is_local: false,
+            pool_idx: 0,
+            set_idx: 0,
+            disk_idx: 0,
+        };
+        let remote = crate::cluster::rpc::RemoteDisk::new(
+            &endpoint,
+            &DiskOption {
+                cleanup: false,
+                health_check: false,
+            },
+            transport.clone(),
+        )
+        .await
+        .expect("construct remote disk with failing data transport");
+        let mut verification_disks: Vec<_> = disks.iter().cloned().map(Some).collect();
+        verification_disks[0] = Some(Arc::new(crate::disk::Disk::Remote(Box::new(remote))));
+        let mut statuses = [(0, vec![CHECK_PART_UNKNOWN; disks.len()])].into_iter().collect();
+        let result = crate::io_support::shard_integrity::verify_deep_parts(
+            &files,
+            &verification_disks,
+            &files[0],
+            bucket,
+            object,
+            &mut statuses,
+        )
+        .await;
+        assert!(transport.reads.load(std::sync::atomic::Ordering::Relaxed) > 0);
+        if matches!(expected_error, DiskError::FileCorrupt | DiskError::FileNotFound) {
+            result.expect("typed remote payload damage must identify a repairable shard");
+            let expected_status = if expected_error == DiskError::FileCorrupt {
+                CHECK_PART_FILE_CORRUPT
+            } else {
+                CHECK_PART_FILE_NOT_FOUND
+            };
+            assert_eq!(statuses[&0][0], expected_status);
+            assert!(statuses[&0][1..].iter().all(|status| *status == CHECK_PART_SUCCESS));
+            let (heal, metadata, reason) = super::should_heal_object_on_disk(&None, &[statuses[&0][0]], &files[0], &files[0]);
+            assert!(heal, "remote payload damage must schedule repair: {expected_error:?}");
+            assert!(!metadata);
+            let expected_reason = if expected_error == DiskError::FileCorrupt {
+                DiskError::FileCorrupt
+            } else {
+                DiskError::PartMissingOrCorrupt
+            };
+            assert_eq!(reason, Some(expected_reason));
+        } else {
+            let error = result.expect_err("unavailable payload storage must abort health verification");
+            if expected_error.is_internode_http_status(500) {
+                assert!(error.is_internode_http_status(500), "preserve the original transport error: {error:?}");
+            } else {
+                assert_eq!(error, expected_error, "preserve the original storage failure");
+            }
+            assert_eq!(statuses[&0][0], CHECK_PART_UNKNOWN, "failed verification cannot certify the shard");
+        }
+        for (path, before) in snapshots {
+            assert_eq!(tokio::fs::read(path).await.expect("read fixture after scan"), before);
+        }
+    }
+
+    #[tokio::test]
+    async fn deep_scan_remote_payload_damage_schedules_repair() {
+        for error in [DiskError::FileCorrupt, DiskError::FileNotFound] {
+            assert_remote_integrity_read_failure(error).await;
+        }
+    }
+
+    #[tokio::test]
+    async fn deep_scan_remote_storage_failure_cannot_certify_healthy() {
+        for error in [
+            DiskError::from(rustfs_rio::new_test_internode_http_io_error(
+                rustfs_rio::InternodeHttpErrorKind::HttpStatus(reqwest::StatusCode::INTERNAL_SERVER_ERROR),
+            )),
+            DiskError::DiskNotFound,
+            DiskError::VolumeNotFound,
+            DiskError::FileAccessDenied,
+        ] {
+            assert_remote_integrity_read_failure(error).await;
+        }
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn deep_heal_legacy_verification_error_cannot_certify_healthy() {
+        use crate::object_api::ShardIntegrityWriteMode;
+
+        let (dirs, disks, set) = hermetic_set_disks_isolated(4).await;
+        let bucket = "legacy-deep-scan-verification-error";
+        let object = "object.bin";
+        for disk in &disks {
+            disk.make_volume(bucket).await.expect("create legacy fixture bucket");
+        }
+        set.put_object(
+            bucket,
+            object,
+            &mut PutObjReader::from_vec(vec![0x6d; 1024 * 1024 + 123]),
+            &ObjectOptions {
+                shard_integrity_write_mode: Some(ShardIntegrityWriteMode::Legacy),
+                no_lock: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("write external legacy fixture");
+
+        let mut files = Vec::new();
+        let mut part_paths = Vec::new();
+        let mut preserved = Vec::new();
+        for (index, (dir, disk)) in dirs.iter().zip(&disks).enumerate() {
+            let file = disk
+                .read_version("", bucket, object, "", &ReadOptions::default())
+                .await
+                .expect("read legacy fixture metadata");
+            assert!(file.data.is_none(), "legacy verification must open the physical shard");
+            assert!(file.parts.iter().all(|part| part.integrity.is_none()));
+            let object_dir = dir.path().join(bucket).join(object);
+            let part = object_dir
+                .join(file.data_dir.expect("external legacy shard directory").to_string())
+                .join("part.1");
+            let metadata = object_dir.join("xl.meta");
+            let bytes = tokio::fs::read(&metadata).await.expect("snapshot legacy metadata");
+            preserved.push((metadata, bytes));
+            if index != 0 {
+                let bytes = tokio::fs::read(&part).await.expect("snapshot healthy legacy shard");
+                preserved.push((part.clone(), bytes));
+            }
+            part_paths.push(part);
+            files.push(file);
+        }
+        let broken_part = &part_paths[0];
+        tokio::fs::remove_file(broken_part)
+            .await
+            .expect("remove fixture shard before symlink fault");
+        std::os::unix::fs::symlink("part.1", broken_part).expect("create self-referential shard symlink");
+        let open_error = tokio::fs::File::open(broken_part)
+            .await
+            .expect_err("self-referential symlink must fail to open");
+        assert_eq!(open_error.raw_os_error(), Some(libc::ELOOP));
+
+        let verification_error = disks[0]
+            .verify_file(bucket, object, &files[0])
+            .await
+            .expect_err("legacy disk verifier must reject the symlink path");
+        assert_eq!(verification_error, DiskError::InvalidPath);
+        let error = set
+            .heal_object(
+                bucket,
+                object,
+                "",
+                &HealOpts {
+                    no_lock: true,
+                    scan_mode: HealScanMode::Deep,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect_err("an unverified legacy part must fail the complete heal operation");
+        assert_eq!(error, DiskError::InvalidPath, "complete heal must preserve the verifier's path error");
+        assert_eq!(
+            tokio::fs::read_link(broken_part)
+                .await
+                .expect("failed heal must preserve the symlink fault"),
+            std::path::Path::new("part.1")
+        );
+        for (path, before) in preserved {
+            assert_eq!(tokio::fs::read(path).await.expect("read unchanged legacy fixture"), before);
+        }
+    }
+
+    async fn assert_h2_truncated_shards(
+        mode: crate::object_api::ShardIntegrityWriteMode,
+        coding_indexes: &[usize],
+        retained_len: fn(usize, usize) -> usize,
+    ) {
+        use crate::object_api::ShardIntegrityWriteMode;
+
+        let (temp_dirs, disks, set) = hermetic_set_disks_for_pool_with_default_parity_isolated(16, 0, 4).await;
+        let bucket = "deep-heal-h2-truncation";
+        let object = "object.bin";
+        for disk in &disks {
+            disk.make_volume(bucket).await.expect("create bucket volume");
+        }
+        let expected_payload: Vec<_> = (0..5 * 1024 * 1024 + 123)
+            .map(|index| u8::try_from((index / 4096 + index) % 251).expect("payload byte fits"))
+            .collect();
+        set.put_object(
+            bucket,
+            object,
+            &mut PutObjReader::from_vec(expected_payload.clone()),
+            &ObjectOptions {
+                shard_integrity_write_mode: Some(mode),
+                no_lock: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("commit all source shards");
+
+        let mut files = Vec::new();
+        let mut paths = Vec::new();
+        let mut original_shards = Vec::new();
+        let mut original_metadata = Vec::new();
+        let mut damaged = Vec::new();
+        for (slot, disk) in disks.iter().enumerate() {
+            let source = disk
+                .read_version("", bucket, object, "", &ReadOptions::default())
+                .await
+                .expect("read source metadata");
+            assert_eq!((source.erasure.data_blocks, source.erasure.parity_blocks), (12, 4));
+            assert_eq!(source.parts[0].integrity.is_some(), mode == ShardIntegrityWriteMode::Protected);
+            let directory = temp_dirs[slot].path().join(bucket).join(object);
+            let path = directory
+                .join(source.data_dir.expect("external shard data directory").to_string())
+                .join("part.1");
+            let original = tokio::fs::read(&path).await.expect("read original shard");
+            original_metadata.push(tokio::fs::read(directory.join("xl.meta")).await.expect("read xl.meta"));
+            if coding_indexes.contains(&source.erasure.index) {
+                let erasure = crate::erasure::coding::Erasure::try_new_with_options(
+                    12,
+                    4,
+                    source.erasure.block_size,
+                    source.uses_legacy_checksum,
+                )
+                .expect("valid erasure layout");
+                let stripe_size = erasure.shard_size() + source.erasure.get_checksum_info(1).algorithm.size();
+                let keep = retained_len(original.len(), stripe_size);
+                assert!(keep < original.len(), "fixture must truncate the shard");
+                tokio::fs::OpenOptions::new()
+                    .write(true)
+                    .open(&path)
+                    .await
+                    .expect("open target shard")
+                    .set_len(u64::try_from(keep).expect("shard length fits"))
+                    .await
+                    .expect("truncate target shard");
+                damaged.push((slot, keep));
+            }
+            paths.push(path);
+            original_shards.push(original);
+            files.push(source);
+        }
+        assert_eq!(damaged.len(), coding_indexes.len());
+        if mode == ShardIntegrityWriteMode::Protected {
+            let mut statuses = [(0, vec![crate::disk::CHECK_PART_UNKNOWN; disks.len()])]
+                .into_iter()
+                .collect();
+            let online_disks: Vec<_> = disks.iter().cloned().map(Some).collect();
+            crate::io_support::shard_integrity::verify_deep_parts(
+                &files,
+                &online_disks,
+                &files[0],
+                bucket,
+                object,
+                &mut statuses,
+            )
+            .await
+            .expect("verify protected shards directly");
+            for (slot, status) in statuses[&0].iter().enumerate() {
+                let expected = if damaged.iter().any(|(index, _)| *index == slot) {
+                    crate::disk::CHECK_PART_FILE_CORRUPT
+                } else {
+                    crate::disk::CHECK_PART_SUCCESS
+                };
+                assert_eq!(*status, expected, "protected shard {slot}");
+            }
+        }
+
+        let opts = HealOpts {
+            no_lock: true,
+            scan_mode: HealScanMode::Deep,
+            ..Default::default()
+        };
+        let (dry_run, error) = set
+            .heal_object(bucket, object, "", &HealOpts { dry_run: true, ..opts })
+            .await
+            .expect("dry-run should report corruption");
+        assert!(error.is_none(), "dry-run error: {error:?}");
+        assert_eq!(dry_run.drives_healed(), Some(0));
+        for (slot, keep) in &damaged {
+            assert_eq!(dry_run.after.drives[*slot].state, DriveState::Corrupt.to_string());
+            assert_eq!(
+                tokio::fs::read(&paths[*slot]).await.expect("dry-run preserved shard"),
+                original_shards[*slot][..*keep]
+            );
+        }
+
+        let recoverable = damaged.len() <= 4;
+        if recoverable {
+            let mut reader = set
+                .get_object_reader(bucket, object, None, Default::default(), &ObjectOptions::default())
+                .await
+                .expect("degraded GET should remain readable");
+            let mut read_back = Vec::new();
+            tokio::io::copy(&mut reader, &mut read_back)
+                .await
+                .expect("read complete degraded GET");
+            assert_eq!(read_back, expected_payload);
+        }
+        let (result, error) = set
+            .heal_object(bucket, object, "", &opts)
+            .await
+            .expect("heal should report outcome");
+        if recoverable {
+            assert!(error.is_none(), "heal error: {error:?}");
+            assert_eq!(result.drives_healed(), Some(damaged.len()));
+            for (slot, _) in &damaged {
+                assert_eq!(result.before.drives[*slot].state, DriveState::Corrupt.to_string());
+            }
+        } else {
+            assert_eq!(error, Some(DiskError::ErasureReadQuorum));
+            assert_eq!(result.drives_healed(), Some(0));
+        }
+        for (slot, path) in paths.iter().enumerate() {
+            let expected = if !recoverable && let Some((_, keep)) = damaged.iter().find(|(index, _)| *index == slot) {
+                &original_shards[slot][..*keep]
+            } else {
+                original_shards[slot].as_slice()
+            };
+            assert_eq!(
+                tokio::fs::read(path).await.expect("read shard after heal"),
+                expected,
+                "physical shard {slot}"
+            );
+            if !recoverable {
+                let metadata_path = temp_dirs[slot].path().join(bucket).join(object).join("xl.meta");
+                assert_eq!(tokio::fs::read(metadata_path).await.expect("preserved metadata"), original_metadata[slot]);
+            }
+        }
+        if recoverable {
+            let mut reader = set
+                .get_object_reader(bucket, object, None, Default::default(), &ObjectOptions::default())
+                .await
+                .expect("GET after repair");
+            let mut read_back = Vec::new();
+            tokio::io::copy(&mut reader, &mut read_back)
+                .await
+                .expect("read repaired object");
+            assert_eq!(read_back, expected_payload);
+            let (healthy, error) = set.heal_object(bucket, object, "", &opts).await.expect("second deep scan");
+            assert!(error.is_none());
+            assert_eq!(healthy.drives_healed(), Some(0), "repair must converge");
+            assert_eq!(healthy.integrity_verified, mode == ShardIntegrityWriteMode::Protected);
+        }
+    }
+
+    #[tokio::test]
+    async fn deep_heal_rebuilds_h2_near_tail_truncated_part() {
+        use crate::object_api::ShardIntegrityWriteMode::{Legacy, Protected};
+        for mode in [Legacy, Protected] {
+            for coding_index in [1, 13] {
+                assert_h2_truncated_shards(mode, &[coding_index], |length, _| length - 1).await;
+            }
+        }
+    }
+
+    #[tokio::test]
+    async fn deep_heal_rebuilds_h2_half_truncated_part() {
+        use crate::object_api::ShardIntegrityWriteMode::{Legacy, Protected};
+        for mode in [Legacy, Protected] {
+            for coding_index in [1, 13] {
+                assert_h2_truncated_shards(mode, &[coding_index], |length, _| length / 2).await;
+            }
+        }
+    }
+
+    #[tokio::test]
+    async fn deep_heal_rebuilds_h2_stripe_boundary_truncated_parts_at_read_quorum() {
+        use crate::object_api::ShardIntegrityWriteMode::{Legacy, Protected};
+        for mode in [Legacy, Protected] {
+            assert_h2_truncated_shards(mode, &[1, 2, 13, 16], |_, stripe| stripe).await;
+        }
+    }
+
+    #[tokio::test]
+    async fn deep_heal_rejects_h2_truncated_parts_below_read_quorum() {
+        use crate::object_api::ShardIntegrityWriteMode::{Legacy, Protected};
+        for mode in [Legacy, Protected] {
+            assert_h2_truncated_shards(mode, &[1, 2, 3, 13, 16], |length, _| length / 2).await;
+        }
+    }
+
    #[tokio::test]
    async fn deep_heal_inline_bitrot_rebuilds_physical_data_and_parity_shards() {
        use crate::storage_api_contracts::range::HTTPRangeSpec;
@@ -5385,7 +6167,17 @@ mod heal_result_report_tests {
            let failure = DanglingDeleteFailure::install(bucket, object, 0, DiskError::FaultyDisk);
            let mut proof = None;
            let (_, error) = set
-                .heal_object_with_absence(bucket, object, &version.to_string(), &opts, &mut proof)
+                .heal_object_with_absence(
+                    bucket,
+                    object,
+                    &version.to_string(),
+                    &opts,
+                    &mut proof,
+                    AbsenceProofRequest {
+                        retirement: None,
+                        allow_unversioned: false,
+                    },
+                )
                .await
                .expect("heal should return a per-object failure");
            assert!(
@@ -5401,7 +6193,17 @@ mod heal_result_report_tests {
            );
            drop(failure);
            let (result, error) = set
-                .heal_object_with_absence(bucket, object, &version.to_string(), &opts, &mut proof)
+                .heal_object_with_absence(
+                    bucket,
+                    object,
+                    &version.to_string(),
+                    &opts,
+                    &mut proof,
+                    AbsenceProofRequest {
+                        retirement: None,
+                        allow_unversioned: false,
+                    },
+                )
                .await
                .expect("retry should execute cleanup");
            assert!(error.is_none(), "retry should complete: {error:?}");
@@ -5416,7 +6218,17 @@ mod heal_result_report_tests {
                    .all(|drive| drive.state == DriveState::Missing.to_string())
            );
            let (_, _) = set
-                .heal_object_with_absence(bucket, object, &version.to_string(), &opts, &mut proof)
+                .heal_object_with_absence(
+                    bucket,
+                    object,
+                    &version.to_string(),
+                    &opts,
+                    &mut proof,
+                    AbsenceProofRequest {
+                        retirement: None,
+                        allow_unversioned: false,
+                    },
+                )
                .await
                .expect("already absent replay should execute");
            assert!(!proof.expect("exact already-absent replay must remain provable").removed);
@@ -341,14 +341,17 @@ impl SetDisks {
    pub async fn renew_disk(&self, ep: &Endpoint) {
        debug!("renew_disk: start {:?}", ep);

-        let previous_health = {
+        let previous_disk = {
            let disks = self.disks.read().await;
            disks
                .iter()
                .filter_map(|disk| disk.as_ref())
                .find(|disk| disk.endpoint() == *ep)
-                .and_then(|disk| disk.local_health_tracker_epoch_for_reconnect())
+                .cloned()
        };
+        let previous_health = previous_disk
+            .as_ref()
+            .and_then(|disk| disk.local_health_tracker_epoch_for_reconnect());

        let (new_disk, fm) = match Self::connect_endpoint(ep, previous_health).await {
            Ok(res) => res,
@@ -406,6 +409,25 @@ impl SetDisks {
        }

        let _ = new_disk.set_disk_id(Some(fm.erasure.this)).await;
+
+        if !new_disk.is_local() {
+            let mut disks = self.disks.write().await;
+            let slot = &mut disks[disk_idx];
+            let unchanged = match (slot.as_ref(), previous_disk.as_ref()) {
+                (None, None) => true,
+                (Some(current), Some(previous)) => Arc::ptr_eq(current, previous),
+                _ => false,
+            };
+            // A GET probe or another reconnect may have filled this slot while
+            // format I/O was pending. Do not replace its handle or start a
+            // second health monitor for an unpublished handle.
+            if unchanged {
+                new_disk.enable_health_check();
+                *slot = Some(new_disk);
+            }
+            return;
+        }
+
        new_disk.enable_health_check();

        if new_disk.is_local() {
@@ -2776,6 +2776,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
                    self.set_index,
                    self.pool_index,
                    opts.skip_verify_bitrot,
+                    opts.suppress_read_repair,
                    true,
                    true,
                    GET_OBJECT_PATH_LEGACY_DUPLEX,
@@ -2805,6 +2806,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
                        self.set_index,
                        self.pool_index,
                        opts.skip_verify_bitrot,
+                        opts.suppress_read_repair,
                        true,
                        false,
                        GET_OBJECT_PATH_LEGACY_DUPLEX,
@@ -2893,6 +2895,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
                self.set_index,
                self.pool_index,
                opts.skip_verify_bitrot,
+                opts.suppress_read_repair,
                true,
                false,
                GET_OBJECT_PATH_DIRECT_MEMORY,
@@ -2949,6 +2952,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
                self.set_index,
                self.pool_index,
                opts.skip_verify_bitrot,
+                opts.suppress_read_repair,
                object_class.as_str(),
                size_bucket,
                false,
@@ -2995,6 +2999,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
                    self.set_index,
                    self.pool_index,
                    opts.skip_verify_bitrot,
+                    opts.suppress_read_repair,
                    object_class.as_str(),
                    size_bucket,
                    codec_streaming_gate.prefer_data_blocks_first_reader_setup,
@@ -3063,6 +3068,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
        let set_index = self.set_index;
        let pool_index = self.pool_index;
        let skip_verify = opts.skip_verify_bitrot;
+        let suppress_read_repair = opts.suppress_read_repair;
        // The producer runs in a separate Tokio task, so carry the caller's
        // read policy across the task boundary explicitly. Tokio task-local
        // values are not inherited by spawned tasks.
@@ -3093,6 +3099,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
                        set_index,
                        pool_index,
                        skip_verify,
+                        suppress_read_repair,
                        false,
                        false,
                        GET_OBJECT_PATH_LEGACY_DUPLEX,
@@ -6715,7 +6722,7 @@ fn remote_version_state_writer_enabled() -> bool {
 }

 fn remote_version_state_writer_fleet_proof() -> Option<RemoteVersionStateFleetProofToken> {
-    transaction_fencing_fleet_proof(remote_version_state_writer_requested())
+    crate::services::notification_sys::acquire_remote_version_state_writer_fleet_proof()
 }

 fn remote_version_state_writer_requested() -> bool {
@@ -7571,7 +7578,11 @@ impl SetDisks {
        let fi = &transported;
        let disks = self.disk_inventory().await;
        let namespace_owner = (!is_meta_bucketname(bucket)).then(|| self.ctx.begin_namespace_commit());
-        let write_quorum = disks.len() / 2 + 1;
+        // Quorum is a property of the configured set, not of the current
+        // online snapshot. A decommission/offline refresh may temporarily
+        // shorten the snapshot; deriving quorum from it would turn a
+        // quorum-minus-one delete into an apparent success.
+        let write_quorum = self.set_drive_count / 2 + 1;
        let rollback_dir = Uuid::new_v4();

        let mut futures = Vec::with_capacity(disks.len());
@@ -7699,10 +7710,11 @@ impl SetDisks {
        drop(namespace_owner);
        if quorum_result.is_ok()
            && errs.iter().any(Option::is_some)
-            && let Some(purge) = delete_marker_purge
+            && let Some(purge) = delete_marker_purge.as_ref()
            && let Some(version) = fi.version_id.filter(|version| !version.is_nil())
        {
-            let _ = self.persist_delete_marker_purge(bucket, object, version, purge).await;
+            let _ = self.persist_marker_purge_receipt(bucket, object, version, purge).await;
+            let _ = self.persist_delete_marker_purge(bucket, object, version, purge.clone()).await;
        }
        // An explicit purge can carry deleted=true for the existing marker.
        // It must not create a repair intent that could reintroduce that marker.
@@ -9488,6 +9500,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
        let set_index = self.set_index;
        let pool_index = self.pool_index;
        let skip_verify = opts.skip_verify_bitrot;
+        let suppress_read_repair = opts.suppress_read_repair;
        let metrics_size_bucket = rustfs_io_metrics::get_object_size_bucket(cloned_fi.size);
        let erasure_cache = Arc::clone(&self.erasure_cache);
        let producer = async move {
@@ -9505,6 +9518,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                set_index,
                pool_index,
                skip_verify,
+                suppress_read_repair,
                false,
                false,
                GET_OBJECT_PATH_LEGACY_DUPLEX,
@@ -18,6 +18,7 @@ use crate::diagnostics::get::{
 use crate::disk::error::Error;
 use crate::layout::disks_layout::MAX_ERASURE_SET_DRIVE_COUNT;
 use smallvec::SmallVec;
+use std::sync::atomic::{AtomicU8, Ordering};

 /// Generic codec callers may exceed the production set limit; `SmallVec` then
 /// spills without changing slot semantics.
@@ -25,6 +26,35 @@ pub(crate) const INLINE_SHARD_SLOTS: usize = MAX_ERASURE_SET_DRIVE_COUNT;
 pub(crate) type ShardBuffers = SmallVec<[Option<Vec<u8>>; INLINE_SHARD_SLOTS]>;
 pub(crate) type ShardErrors = SmallVec<[Option<Error>; INLINE_SHARD_SLOTS]>;

+/// Repair evidence is independent of reconstruction quorum. Successful shards
+/// must not vote away a minority's observed corruption. Streaming decoders
+/// share this small accumulator with the reader that completes the response.
+#[derive(Debug, Default)]
+pub(crate) struct ShardReadRepair {
+    observed: AtomicU8,
+}
+
+impl ShardReadRepair {
+    pub(crate) fn observe(&self, errors: &[Option<Error>]) {
+        let flags = errors.iter().flatten().fold(0, |flags, error| {
+            flags
+                | match error {
+                    Error::FileCorrupt => 1,
+                    Error::FileNotFound => 2,
+                    _ => 0,
+                }
+        });
+        if flags != 0 {
+            self.observed.fetch_or(flags, Ordering::Relaxed);
+        }
+    }
+
+    pub(crate) fn needed(&self, allow_missing: bool) -> bool {
+        let mask = if allow_missing { 3 } else { 1 };
+        self.observed.load(Ordering::Relaxed) & mask != 0
+    }
+}
+
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub(crate) enum ShardReadCost {
    Local,
@@ -164,6 +194,19 @@ mod tests {
    use super::*;
    use std::mem::size_of;

+    #[test]
+    fn repair_evidence_survives_healthy_stripes_and_excludes_unproven_failures() {
+        let evidence = ShardReadRepair::default();
+        evidence.observe(&[None, Some(Error::Timeout), Some(Error::DiskNotFound)]);
+        assert!(!evidence.needed(true));
+        evidence.observe(&[None, None, Some(Error::FileNotFound)]);
+        assert!(!evidence.needed(false), "unattempted reader slots do not prove a missing shard");
+        assert!(evidence.needed(true));
+        evidence.observe(&[None, None, Some(Error::FileCorrupt)]);
+        evidence.observe(&[None, None, None]);
+        assert!(evidence.needed(false), "a later healthy stripe must not erase observed corruption");
+    }
+
    #[test]
    fn stripe_scratch_capacity_matches_the_production_set_limit() {
        type OversizedShardBuffers = SmallVec<[Option<Vec<u8>>; 32]>;
@@ -645,6 +645,17 @@ impl ECStore {
                    .as_ref()
                    .and_then(|info| info.created)
                    .unwrap_or(OffsetDateTime::UNIX_EPOCH);
+                // A sidecar left by an interrupted legacy migration is the
+                // bucket's published incarnation; keep it rather than minting
+                // one that would invalidate fences taken on the stored value.
+                // A retired incarnation is residue of a deleted bucket and
+                // must not come back.
+                let store = metadata_sys::object_store_in(&self.ctx).await?;
+                if let Some(stored) = crate::bucket::metadata::load_bucket_incarnation(store.clone(), bucket).await?
+                    && !crate::bucket::retirement::is_retired(store, bucket, stored).await?
+                {
+                    metadata.bucket_incarnation_id = stored;
+                }
            } else if !metadata.bucket_incarnation_sidecar && !metadata.bucket_incarnation_id.is_nil() {
                return Err(Error::other(format!(
                    "bucket incarnation sidecar is missing for new-format metadata: {bucket}"
@@ -2881,6 +2892,57 @@ mod tests {
        );
    }

+    /// rustfs/rustfs#8003: a force-create (site replication replay, admin
+    /// import) against a bucket whose legacy migration stopped after the
+    /// sidecar write must persist `.metadata.bin` under the stored incarnation
+    /// rather than fail closed or replace the published identity.
+    #[tokio::test(flavor = "multi_thread")]
+    #[serial]
+    async fn issue_8003_force_create_adopts_existing_incarnation_sidecar() {
+        let (disk_paths, ecstore) = setup_bucket_delete_test_env().await;
+        let bucket = format!("bucket-8003-sidecar-{}", Uuid::new_v4().simple());
+
+        ecstore
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("bucket should be created");
+        let stored = metadata_sys::get_in(&ecstore.ctx, &bucket)
+            .await
+            .expect("metadata should load")
+            .bucket_incarnation_id;
+        assert!(!stored.is_nil());
+
+        let metadata_path = format!("{BUCKET_META_PREFIX}/{bucket}/{}", crate::bucket::metadata::BUCKET_METADATA_FILE);
+        crate::config::com::delete_config(ecstore.clone(), &metadata_path)
+            .await
+            .expect("simulate the interrupted migration by removing only .metadata.bin");
+        assert!(!any_disk_path_exists(&disk_paths, format!("{RUSTFS_META_BUCKET}/{metadata_path}")).await);
+        metadata_sys::remove_bucket_metadata_in(&ecstore.ctx, &bucket)
+            .await
+            .expect("drop the cached copy so the next read hits disk");
+
+        ecstore
+            .make_bucket(
+                &bucket,
+                &MakeBucketOptions {
+                    force_create: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("force create must repair the sidecar-only bucket");
+
+        let (repaired, persisted) = metadata_sys::get_config_from_disk_with_presence_in(&ecstore.ctx, &bucket)
+            .await
+            .expect("repaired metadata should load");
+        assert!(persisted, "force create must persist .metadata.bin again");
+        assert!(repaired.bucket_incarnation_sidecar);
+        assert_eq!(
+            repaired.bucket_incarnation_id, stored,
+            "the repair must keep the incarnation the sidecar already published"
+        );
+    }
+
    #[tokio::test(flavor = "multi_thread")]
    #[serial]
    async fn make_bucket_seeds_new_bucket_durability_override() {
@@ -58,6 +58,26 @@ pub(crate) fn bucket_heal_scope(bucket: &str) -> Option<Arc<BucketHealScope>> {
        .flatten()
 }

+/// Resolve only the two canonical bucket records; other internal paths carry no bucket authority.
+pub(super) fn bucket_metadata_owner(object: &str) -> Option<&str> {
+    use crate::bucket::metadata::{BUCKET_INCARNATION_FILE, BUCKET_METADATA_FILE};
+    let (prefix, rest) = object.split_once('/')?;
+    let (bucket, file) = rest.split_once('/')?;
+    (prefix == disk::BUCKET_META_PREFIX
+        && matches!(file, BUCKET_METADATA_FILE | BUCKET_INCARNATION_FILE)
+        && check_valid_bucket_name_strict(bucket).is_ok())
+    .then_some(bucket)
+}
+
+pub(crate) fn bucket_heal_scope_for_object(bucket: &str, object: &str) -> Option<Arc<BucketHealScope>> {
+    let owner = if bucket == RUSTFS_META_BUCKET {
+        bucket_metadata_owner(object)?
+    } else {
+        bucket
+    };
+    bucket_heal_scope(owner)
+}
+
 /// Storage-owned proof for the exact version and every selected erasure location.
 /// This is an in-process result, never reconstructed from admin drive telemetry.
 #[derive(Debug)]
@@ -296,6 +316,24 @@ impl ECStore {
        .await
    }

+    pub async fn heal_mrf_object_at_incarnation(
+        self: &Arc<Self>,
+        bucket: &str,
+        object: &str,
+        version_id: &str,
+        expected: uuid::Uuid,
+        opts: &HealOpts,
+    ) -> Result<HealObjectStorageResult> {
+        let object = object.to_owned();
+        let version_id = version_id.to_owned();
+        self.run_bucket_heal_at_incarnation(bucket, expected, opts, |store, bucket, opts| async move {
+            store
+                .heal_object_with_authoritative_absence_proof(&bucket, &object, &version_id, &opts)
+                .await
+        })
+        .await
+    }
+
    async fn acquire_heal_format_fence(
        &self,
    ) -> Result<(
@@ -404,6 +442,97 @@ impl ECStore {
        Ok(Arc::ptr_eq(&target_set, &pool.get_disks_by_key(POOL_META_NAME)))
    }

+    /// Restore the bucket configuration and incarnation before retiring a replacement marker.
+    pub async fn heal_replacement_bucket_metadata(
+        self: &Arc<Self>,
+        bucket: &str,
+        opts: &HealOpts,
+        targets: &[String],
+    ) -> Result<()> {
+        use crate::bucket::metadata::{BUCKET_INCARNATION_FILE, BUCKET_METADATA_FILE};
+        use crate::bucket::metadata_sys::acquire_bucket_metadata_transaction_read_lock_in;
+
+        let (Some(pool_index), Some(set_index)) = (opts.pool, opts.set) else {
+            return Err(Error::PreconditionFailed);
+        };
+        if targets.is_empty() || opts.dry_run || opts.no_lock || check_valid_bucket_name_strict(bucket).is_err() {
+            return Err(Error::PreconditionFailed);
+        }
+        let pool = self
+            .pools
+            .get(pool_index)
+            .ok_or_else(|| invalid_heal_pool_index(pool_index, self.pools.len()))?;
+        let target_set = pool.get_disks_for_heal_object(BUCKET_METADATA_FILE, opts)?;
+        if targets.iter().any(|target| {
+            target_set
+                .set_endpoints
+                .iter()
+                .filter(|endpoint| endpoint.to_string() == *target)
+                .count()
+                != 1
+        }) {
+            return Err(Error::PreconditionFailed);
+        }
+        let objects: Vec<_> = [BUCKET_METADATA_FILE, BUCKET_INCARNATION_FILE]
+            .into_iter()
+            .map(|file| format!("{}/{bucket}/{file}", disk::BUCKET_META_PREFIX))
+            .filter(|object| Arc::ptr_eq(&target_set, &pool.get_disks_by_key(object)))
+            .collect();
+        if objects.is_empty() {
+            return Ok(());
+        }
+
+        // Read the persisted identity without lazy migration: a replacement
+        // must never turn lost configuration into newly fabricated defaults.
+        let incarnation = self.bucket_incarnation_id_from_disk(bucket).await?;
+        let targets = targets.to_vec();
+        self.run_bucket_heal_at_incarnation(bucket, incarnation, opts, move |store, bucket, opts| async move {
+            // Match config writers: lifecycle, metadata transaction, object locks.
+            // Keep the transaction stable through target readback, including when
+            // the caller is cancelled and the storage owner finishes its write.
+            let transaction = acquire_bucket_metadata_transaction_read_lock_in(&store.ctx, &bucket).await?;
+            for object in objects {
+                let scope = bucket_heal_scope(&bucket).ok_or(Error::PreconditionFailed)?;
+                scope.check()?;
+                if transaction.is_lock_lost() {
+                    return Err(Error::PreconditionFailed);
+                }
+                // Config objects occupy one pool, unlike pool.bin. Require a
+                // successful all-pool lookup before accepting another pool's ownership.
+                if !store.single_pool() {
+                    let (owner, _) = store
+                        .get_pool_info_for_delete_marker(RUSTFS_META_BUCKET, &object, &ObjectOptions::default())
+                        .await?;
+                    if owner.index != pool_index {
+                        continue;
+                    }
+                }
+                let metadata_opts = HealOpts { remove: false, ..opts };
+                let (result, error) = store.heal_object(RUSTFS_META_BUCKET, &object, "", &metadata_opts).await?;
+                if let Some(error) = error {
+                    return Err(error);
+                }
+                for target in &targets {
+                    let mut drives = result.after.drives.iter().filter(|drive| &drive.endpoint == target);
+                    if !drives.next().is_some_and(|drive| drive.state == "ok") || drives.next().is_some() {
+                        return Err(Error::PreconditionFailed);
+                    }
+                }
+                if !store
+                    .replacement_targets_have_version(RUSTFS_META_BUCKET, &object, "", pool_index, set_index, &targets)
+                    .await?
+                {
+                    return Err(Error::PreconditionFailed);
+                }
+            }
+            if transaction.is_lock_lost() {
+                return Err(Error::PreconditionFailed);
+            }
+            Ok(())
+        })
+        .await
+    }
+
    /// Return every live erasure set selected by an object-heal scope.
    pub async fn heal_erasure_set_scopes(&self, opts: &HealOpts) -> Result<Vec<(usize, usize)>> {
        let pools = self.get_pools_for_heal_object(opts)?;
@@ -684,8 +813,18 @@ impl ECStore {
        version_id: &str,
        opts: &HealOpts,
    ) -> Result<(HealResultItem, Option<Error>)> {
-        self.handle_heal_object_with_absence(bucket, object, version_id, opts, &mut None, None)
-            .await
+        self.handle_heal_object_with_absence(
+            bucket,
+            object,
+            version_id,
+            opts,
+            &mut None,
+            crate::set_disk::AbsenceProofRequest {
+                retirement: None,
+                allow_unversioned: false,
+            },
+        )
+        .await
    }

    pub async fn heal_object_with_proof(
@@ -695,7 +834,34 @@ impl ECStore {
        version_id: &str,
        opts: &HealOpts,
    ) -> Result<HealObjectStorageResult> {
-        if opts.dry_run || opts.no_lock || version_id.is_empty() || super::utils::is_reserved_or_invalid_bucket(bucket, false) {
+        self.heal_object_with_absence_proof(bucket, object, version_id, opts, false)
+            .await
+    }
+
+    async fn heal_object_with_authoritative_absence_proof(
+        &self,
+        bucket: &str,
+        object: &str,
+        version_id: &str,
+        opts: &HealOpts,
+    ) -> Result<HealObjectStorageResult> {
+        self.heal_object_with_absence_proof(bucket, object, version_id, opts, true)
+            .await
+    }
+
+    async fn heal_object_with_absence_proof(
+        &self,
+        bucket: &str,
+        object: &str,
+        version_id: &str,
+        opts: &HealOpts,
+        allow_unversioned_absence: bool,
+    ) -> Result<HealObjectStorageResult> {
+        if opts.dry_run
+            || opts.no_lock
+            || (!allow_unversioned_absence && version_id.is_empty())
+            || super::utils::is_reserved_or_invalid_bucket(bucket, false)
+        {
            let (item, error) = self.handle_heal_object(bucket, object, version_id, opts).await?;
            return Ok(HealObjectStorageResult {
                item,
@@ -738,7 +904,17 @@ impl ECStore {
        };
        let mut proofs = None;
        let (item, mut error) = self
-            .handle_heal_object_with_absence(bucket, object, version_id, opts, &mut proofs, Some(&retirement))
+            .handle_heal_object_with_absence(
+                bucket,
+                object,
+                version_id,
+                opts,
+                &mut proofs,
+                crate::set_disk::AbsenceProofRequest {
+                    retirement: Some(&retirement),
+                    allow_unversioned: allow_unversioned_absence,
+                },
+            )
            .await?;
        // Read the authoritative incarnation only for an absence candidate.
        // The lifecycle guard has pinned it throughout the storage operation.
@@ -780,7 +956,7 @@ impl ECStore {
        version_id: &str,
        opts: &HealOpts,
        absence: &mut Option<Vec<crate::set_disk::HealedObjectAbsence>>,
-        retirement: Option<&crate::bucket::retirement::MarkerRetirementContext<'_>>,
+        proof: crate::set_disk::AbsenceProofRequest<'_>,
    ) -> Result<(HealResultItem, Option<Error>)> {
        trace!(
            event = EVENT_HEAL_OBJECT_STARTED,
@@ -871,7 +1047,7 @@ impl ECStore {
                            }
                            #[cfg(test)]
                            crate::core::pools::notify_decommission_external_heal_operation_started(store_id);
-                            pool.heal_object_with_absence(bucket, &pool_object, version_id, &opts, retirement)
+                            pool.heal_object_with_absence(bucket, &pool_object, version_id, &opts, proof)
                                .await
                        }
                    });
@@ -895,7 +1071,7 @@ impl ECStore {
                    move |opts| async move {
                        #[cfg(test)]
                        crate::core::pools::notify_decommission_external_heal_operation_started(store_id);
-                        pool.heal_object_with_absence(bucket, &pool_object, version_id, &opts, retirement)
+                        pool.heal_object_with_absence(bucket, &pool_object, version_id, &opts, proof)
                            .await
                    },
                ));
@@ -1280,6 +1456,16 @@ mod tests {
                .await
                .is_err()
        );
+        for file in [".metadata.bin", ".bucket-incarnation"] {
+            let metadata = format!("buckets/{bucket}/{file}");
+            assert!(
+                store
+                    .rename_local_data_at_incarnation(&disk_ref, (&bucket, object), &fi, (RUSTFS_META_BUCKET, &metadata), old)
+                    .await
+                    .is_err(),
+                "stale remote heal must not replace the successor's {file}"
+            );
+        }
        assert!(store.heal_bucket_at_incarnation(&bucket, old, &opts).await.is_err());
        assert!(
            store
@@ -1678,6 +1864,194 @@ mod tests {
        (temp_dir, store, shutdown)
    }

+    #[test]
+    fn bucket_metadata_heal_owner_rejects_path_aliases_and_unrelated_records() {
+        for file in [".metadata.bin", ".bucket-incarnation"] {
+            assert_eq!(bucket_metadata_owner(&format!("buckets/example/{file}")), Some("example"));
+        }
+        for path in [
+            "/buckets/example/.metadata.bin",
+            "buckets/../.metadata.bin",
+            "buckets/example/../.metadata.bin",
+            "buckets/example/.metadata.bin/extra",
+            "buckets/example//.metadata.bin",
+            "buckets/example/.metadata.bin.old",
+            "config/iam/example/.metadata.bin",
+            "buckets/example/usage.json",
+        ] {
+            assert_eq!(bucket_metadata_owner(path), None, "unexpected bucket authority for {path}");
+        }
+    }
+
+    #[tokio::test]
+    async fn replacement_bucket_metadata_respects_pool_ownership_and_required_records() {
+        let (root, store, shutdown) = multi_pool_heal_store().await;
+        let bucket = "replacement-bucket-metadata";
+        for invalid_bucket in ["MixedCase", "../example", RUSTFS_META_BUCKET] {
+            assert!(matches!(
+                store
+                    .heal_replacement_bucket_metadata(
+                        invalid_bucket,
+                        &HealOpts {
+                            pool: Some(0),
+                            set: Some(0),
+                            ..Default::default()
+                        },
+                        &[root.path().join("pool0-disk0").to_string_lossy().into_owned()],
+                    )
+                    .await,
+                Err(Error::PreconditionFailed)
+            ));
+        }
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create bucket and its persisted records");
+        let mut owned = 0;
+        for pool in 0..2 {
+            let target = root.path().join(format!("pool{pool}-disk0"));
+            let records: Vec<_> = [".metadata.bin", ".bucket-incarnation"]
+                .into_iter()
+                .map(|file| {
+                    let path = target.join(RUSTFS_META_BUCKET).join("buckets").join(bucket).join(file);
+                    let exists = path.join("xl.meta").exists();
+                    if exists {
+                        std::fs::remove_dir_all(&path).expect("remove replacement metadata shard");
+                        owned += 1;
+                    }
+                    (path, exists)
+                })
+                .collect();
+            store
+                .heal_replacement_bucket_metadata(
+                    bucket,
+                    &HealOpts {
+                        pool: Some(pool),
+                        set: Some(0),
+                        recreate: true,
+                        ..Default::default()
+                    },
+                    &[target.to_string_lossy().into_owned()],
+                )
+                .await
+                .expect("repair owned records and accept authoritative ownership in the other pool");
+            for (path, existed) in records {
+                assert_eq!(path.join("xl.meta").exists(), existed, "wrong placement for {}", path.display());
+            }
+        }
+        assert_eq!(owned, 2, "each required record must have exactly one owning pool");
+        let object = format!("buckets/{bucket}/.metadata.bin");
+        let owner = store
+            .get_pool_idx_existing_with_opts(RUSTFS_META_BUCKET, &object, &ObjectOptions::default())
+            .await
+            .expect("persisted metadata owner");
+        let disk = store.pools[owner].disk_set[0].disks.read().await[0]
+            .clone()
+            .expect("metadata target disk");
+        let fi = disk
+            .read_version(
+                "",
+                RUSTFS_META_BUCKET,
+                &object,
+                "",
+                &disk::ReadOptions {
+                    read_data: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("inline metadata to repair through target-side admission");
+        assert!(fi.data.is_some(), "fixture must contain an inline metadata shard");
+        let target_path = root.path().join(format!("pool{owner}-disk0"));
+        std::fs::remove_dir_all(target_path.join(RUSTFS_META_BUCKET).join(&object)).expect("remove remote target shard");
+        let incarnation = store
+            .bucket_incarnation_id_from_disk(bucket)
+            .await
+            .expect("current incarnation");
+        store
+            .rename_local_data_at_incarnation(
+                &disk.endpoint().to_string(),
+                (disk::RUSTFS_META_TMP_BUCKET, "metadata-heal-control"),
+                &fi,
+                (RUSTFS_META_BUCKET, &object),
+                incarnation,
+            )
+            .await
+            .expect("the target-side RPC admission must accept current bucket metadata repair");
+        assert!(target_path.join(RUSTFS_META_BUCKET).join(&object).join("xl.meta").exists());
+        let (configuration, _) = read_config_no_lock_preserve_empty_with_metadata(store.clone(), &object)
+            .await
+            .expect("read the current bucket configuration");
+        save_config(store.clone(), &object, configuration)
+            .await
+            .expect("publish a newer config revision in the same bucket incarnation");
+        let committed =
+            std::fs::read(target_path.join(RUSTFS_META_BUCKET).join(&object).join("xl.meta")).expect("new target configuration");
+        assert!(
+            store
+                .rename_local_data_at_incarnation(
+                    &disk.endpoint().to_string(),
+                    (disk::RUSTFS_META_TMP_BUCKET, "delayed-metadata-heal"),
+                    &fi,
+                    (RUSTFS_META_BUCKET, &object),
+                    incarnation,
+                )
+                .await
+                .is_err(),
+            "a delayed repair must not overwrite a newer config in the same incarnation"
+        );
+        assert_eq!(
+            std::fs::read(target_path.join(RUSTFS_META_BUCKET).join(&object).join("xl.meta")).expect("retained configuration"),
+            committed,
+            "rejected repair must leave the newer configuration unchanged"
+        );
+        for pool in 0..2 {
+            for disk in 0..4 {
+                let path = root
+                    .path()
+                    .join(format!("pool{pool}-disk{disk}"))
+                    .join(RUSTFS_META_BUCKET)
+                    .join("buckets")
+                    .join(bucket)
+                    .join(".metadata.bin");
+                if path.exists() {
+                    std::fs::remove_dir_all(path).expect("remove every copy of required bucket configuration");
+                }
+            }
+        }
+        metadata_sys::remove_bucket_metadata_in(&store.ctx, bucket)
+            .await
+            .expect("remove the cached configuration so repair cannot rely on it");
+        assert!(
+            store
+                .heal_replacement_bucket_metadata(
+                    bucket,
+                    &HealOpts {
+                        pool: Some(0),
+                        set: Some(0),
+                        recreate: true,
+                        ..Default::default()
+                    },
+                    &[root.path().join("pool0-disk0").to_string_lossy().into_owned()],
+                )
+                .await
+                .is_err(),
+            "total metadata loss must not be accepted as ownership in another pool"
+        );
+        for pool in 0..2 {
+            assert!(
+                !root
+                    .path()
+                    .join(format!("pool{pool}-disk0"))
+                    .join(RUSTFS_META_BUCKET)
+                    .join(&object)
+                    .exists(),
+                "missing required configuration must not be recreated with defaults"
+            );
+        }
+        shutdown.cancel();
+    }
+
    #[tokio::test]
    #[serial_test::serial]
    async fn absence_proof_requires_every_selected_pool() {
@@ -10052,9 +10052,18 @@ mod tests {
                recovered_transition_version.clone(),
            )))
            .await;
-        let transaction_stats = recover_transition_transaction_records(store.clone(), 100, None)
+        let transaction_stats = {
+            let _proof = crate::services::notification_sys::install_current_remote_version_state_fleet_proof_for_test();
+            temp_env::async_with_vars(
+                [
+                    (rustfs_config::ENV_TIER_REMOTE_VERSION_STATE_WRITE, Some("true")),
+                    (rustfs_config::ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED, Some("true")),
+                ],
+                recover_transition_transaction_records(store.clone(), 100, None),
+            )
            .await
-            .expect("transition recovery should advance and consume the migrated transaction before completion");
+            .expect("transition recovery should advance and consume the migrated transaction before completion")
+        };
        backend.set_transition_candidate_probe_override(None).await;
        assert_eq!(
            (
@@ -21938,9 +21947,18 @@ mod tests {
            expected_removes.push((transaction.remote_object, put_version));
        }

-        let stats = recover_transition_transaction_records(store.clone(), 100, None)
+        let stats = {
+            let _proof = crate::services::notification_sys::install_current_remote_version_state_fleet_proof_for_test();
+            temp_env::async_with_vars(
+                [
+                    (rustfs_config::ENV_TIER_REMOTE_VERSION_STATE_WRITE, Some("true")),
+                    (rustfs_config::ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED, Some("true")),
+                ],
+                recover_transition_transaction_records(store.clone(), 100, None),
+            )
            .await
-            .expect("transition transaction recovery should run");
+            .expect("transition transaction recovery should run")
+        };

        assert_eq!((stats.scanned, stats.recovered, stats.retained, stats.failed), (3, 3, 0, 0));
        assert_eq!(transition_transaction_record_count(store.clone()).await, 0);
@@ -66,7 +66,7 @@ use std::future::Future;
 use std::path::{Path, PathBuf};
 use std::sync::{
    Arc,
-    atomic::{AtomicU64, Ordering},
+    atomic::{AtomicBool, AtomicU64, Ordering},
 };
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
 use tokio::io::duplex;
@@ -341,6 +341,18 @@ pub struct ListPathOptions {
    pub cursor_generation: Option<String>,
    pub walkdir_timeout: Option<Duration>,
    pub walkdir_stall_timeout: Option<Duration>,
+    /// Shared by the collector and raw producers to preserve bounded-walk
+    /// completion when filtering removes all entries from a batch.
+    pub producer_limit_reached: Option<Arc<AtomicBool>>,
+}
+
+fn ensure_producer_limit_state(options: &mut ListPathOptions) -> Arc<AtomicBool> {
+    let state = options
+        .producer_limit_reached
+        .clone()
+        .unwrap_or_else(|| Arc::new(AtomicBool::new(false)));
+    options.producer_limit_reached = Some(state.clone());
+    state
 }

 async fn can_skip_hidden_prefix_check(options: &ListPathOptions) -> bool {
@@ -4304,6 +4316,7 @@ impl ECStore {
        // cancel channel
        let cancel = CancellationToken::new();
        let _cancel_guard = cancel.clone().drop_guard();
+        ensure_producer_limit_state(&mut o);

        let (err_tx, mut err_rx) = broadcast::channel::<Arc<Error>>(1);

@@ -5047,6 +5060,15 @@ async fn gather_results(
        }
    }

+    // A producer can close its stream after exhausting its own scan budget
+    // before this collector reaches its output limit.  In that case the input
+    // channel closing is not authoritative EOF: the caller must advertise a
+    // continuation page even when every remaining entry was filtered out.
+    let producer_limit_reached = opts
+        .producer_limit_reached
+        .as_ref()
+        .is_some_and(|reached| reached.load(Ordering::Acquire));
+
    // finish not full, return eof
    let filtered = scanned_entries.saturating_sub(candidate_entries);
    if let Some(started) = gather_started {
@@ -5083,7 +5105,7 @@ async fn gather_results(
                o: MetaCacheEntries(entries),
                ..Default::default()
            }),
-            err: Some(rustfs_filemeta::Error::Unexpected),
+            err: (!producer_limit_reached).then_some(rustfs_filemeta::Error::Unexpected),
        })
        .await
        .is_err()
@@ -5734,6 +5756,7 @@ impl Sets {
        let (err_tx, mut err_rx) = broadcast::channel::<Arc<Error>>(1);
        let (sender, recv) = mpsc::channel(o.limit as usize);

+        ensure_producer_limit_state(&mut o);
        let sets = self.clone();
        let opts = o.clone();
        let cancel_rx1 = cancel.clone();
@@ -6777,6 +6800,7 @@ impl SetDisks {
        let (err_tx, mut err_rx) = broadcast::channel::<Arc<Error>>(1);
        let (sender, recv) = mpsc::channel(o.limit as usize);

+        ensure_producer_limit_state(&mut o);
        let set = self.clone();
        let opts = o.clone();
        let cancel_rx1 = cancel.clone();
@@ -7098,6 +7122,7 @@ impl SetDisks {
                forward_to: opts.marker,
                min_disks: raw_min_disks,
                per_disk_limit: limit,
+                producer_limit_reached: opts.producer_limit_reached.clone(),
                // A foreground listing is bounded by lack of drive progress (the walk
                // stall timeout) and by the page limit, never by how long a healthy
                // walk takes — a large prefix on slow media is not a fault (#4644).
@@ -7328,6 +7353,7 @@ mod test {
    };
    use std::collections::{HashMap, HashSet};
    use std::path::PathBuf;
+    use std::sync::atomic::AtomicBool;
    use std::sync::{Arc, Mutex};
    use std::time::Duration;
    use tokio::sync::mpsc;
@@ -8347,6 +8373,45 @@ mod test {
        assert!(!cancel.is_cancelled());
    }

+    #[tokio::test]
+    async fn list_path_gather_results_preserves_bounded_producer_after_filtering() {
+        let (entry_tx, entry_rx) = mpsc::channel(4);
+        let (result_tx, mut result_rx) = mpsc::channel(1);
+        let cancel = CancellationToken::new();
+        let producer_limit_reached = Arc::new(AtomicBool::new(true));
+
+        entry_tx
+            .send(test_meta_entry("outside-prefix"))
+            .await
+            .expect("filtered test entry should be queued");
+        drop(entry_tx);
+
+        let handle = tokio::spawn(gather_results(
+            cancel,
+            ListPathOptions {
+                bucket: "bucket".to_owned(),
+                prefix: "requested/".to_owned(),
+                limit: 8,
+                incl_deleted: true,
+                producer_limit_reached: Some(producer_limit_reached),
+                ..Default::default()
+            },
+            entry_rx,
+            result_tx,
+        ));
+
+        let result = result_rx.recv().await.expect("bounded producer result should be delivered");
+        assert!(result.entries.expect("entries should be present").entries().is_empty());
+        assert!(result.err.is_none(), "bounded producer must not be reported as EOF");
+        assert_eq!(
+            handle
+                .await
+                .expect("gather task should not panic")
+                .expect("gather should succeed"),
+            GatherResultsState::InputClosed
+        );
+    }
+
    /// A-1 guard (rustfs/backlog#1306): pin that a *successful* send is never
    /// misclassified as `ConsumerGone`. With the receiver alive, gather_results
    /// must return the real drain state and deliver the payload carrying the
@@ -426,8 +426,8 @@ mod bucket_fence;
 pub(crate) use bucket::await_bucket_namespace_operation;
 pub use bucket_fence::BucketIncarnationFenceGuard;
 mod heal;
-pub(crate) use heal::bucket_heal_scope;
 pub use heal::{HealObjectAbsenceProof, HealObjectStorageResult};
+pub(crate) use heal::{bucket_heal_scope, bucket_heal_scope_for_object};
 mod heal_walk;
 pub use heal_walk::HealWalkVersion;
 mod init;
@@ -9124,6 +9124,168 @@ mod tests {
        assert_eq!(dirs.len(), 16);
    }

+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    async fn current_marker_delete_ack_removes_stale_rejoined_copy() {
+        let bucket = "current-marker-stale-rejoin";
+        let object = "marker.bin";
+        let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let (_dirs, original) = make_local_set_disks_with_ctx(16, 4, ctx.clone()).await;
+        let store = Arc::new(new_prepared_reader_test_store_with_ctx(&[original], ctx).await);
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        store
+            .handle_make_bucket(
+                bucket,
+                &MakeBucketOptions {
+                    versioning_enabled: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("create versioned bucket");
+        let mut history = Vec::new();
+        for value in 1..=3_u8 {
+            let payload = vec![value; 4097];
+            let version = store
+                .put_object(
+                    bucket,
+                    object,
+                    &mut PutObjReader::from_vec(payload.clone()),
+                    &ObjectOptions {
+                        versioned: true,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("write the historical data versions")
+                .version_id
+                .expect("historical version must have an explicit identity");
+            history.push((version, payload));
+        }
+        let marker = store
+            .delete_object(
+                bucket,
+                object,
+                ObjectOptions {
+                    versioned: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("create current marker")
+            .version_id
+            .expect("marker has explicit version");
+        let current = store
+            .bucket_incarnation_id_from_disk(bucket)
+            .await
+            .expect("bucket incarnation");
+        let set = &store.pools[0].disk_set[0];
+        let before = set.disks.read().await.clone();
+        for slot in 12..16 {
+            set.disks.write().await[slot] = None;
+        }
+        store
+            .delete_object(
+                bucket,
+                object,
+                ObjectOptions {
+                    versioned: true,
+                    version_id: Some(marker.to_string()),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("acknowledged delete should remove the marker from the online quorum");
+        *set.disks.write().await = before.clone();
+
+        for disk in before.iter().take(12).flatten() {
+            assert!(matches!(
+                disk.read_version("", bucket, object, &marker.to_string(), &crate::disk::ReadOptions::default())
+                    .await,
+                Err(crate::disk::error::DiskError::FileVersionNotFound | crate::disk::error::DiskError::FileNotFound)
+            ));
+        }
+        for disk in before.iter().skip(12).flatten() {
+            assert!(
+                disk.read_version("", bucket, object, &marker.to_string(), &crate::disk::ReadOptions::default())
+                    .await
+                    .is_ok(),
+                "offline member must retain the stale current marker"
+            );
+        }
+        let healed = store
+            .heal_object_at_incarnation(
+                bucket,
+                object,
+                &marker.to_string(),
+                current,
+                &rustfs_heal_contracts::heal_channel::HealOpts {
+                    remove: true,
+                    scan_mode: rustfs_heal_contracts::heal_channel::HealScanMode::Deep,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("stale current marker heal should complete");
+        assert!(healed.error.is_none(), "stale current marker heal: {:?}", healed.error);
+        assert!(healed.absence.as_ref().is_some_and(|proof| proof.removed));
+        for disk in before.iter().flatten() {
+            assert!(matches!(
+                disk.read_version("", bucket, object, &marker.to_string(), &crate::disk::ReadOptions::default())
+                    .await,
+                Err(crate::disk::error::DiskError::FileVersionNotFound | crate::disk::error::DiskError::FileNotFound)
+            ));
+        }
+        let replay = store
+            .heal_object_at_incarnation(
+                bucket,
+                object,
+                &marker.to_string(),
+                current,
+                &rustfs_heal_contracts::heal_channel::HealOpts {
+                    remove: true,
+                    scan_mode: rustfs_heal_contracts::heal_channel::HealScanMode::Deep,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("replay the exact removed marker identity");
+        assert!(replay.error.is_none(), "exact marker replay: {:?}", replay.error);
+        let absence = replay.absence.expect("exact replay must prove authoritative absence");
+        assert!(!absence.removed, "replay must not claim another physical repair");
+        assert_eq!(absence.bucket_incarnation_id, current);
+        assert_eq!(absence.version_id, marker.to_string());
+        let listed = store
+            .clone()
+            .inner_list_object_versions(bucket, "", None, None, None, 100)
+            .await
+            .expect("list surviving versions after marker cleanup");
+        assert_eq!(listed.objects.len(), history.len(), "an absent marker is not enumerable in a new walk");
+        assert!(listed.objects.iter().all(|info| !info.delete_marker));
+        for (version, expected) in history {
+            assert!(listed.objects.iter().any(|info| info.version_id == Some(version)));
+            let mut reader = store
+                .handle_get_object_reader(
+                    bucket,
+                    object,
+                    None,
+                    HeaderMap::new(),
+                    &ObjectOptions {
+                        version_id: Some(version.to_string()),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("historical data must remain readable after cleanup and replay");
+            let mut actual = Vec::new();
+            reader
+                .stream
+                .read_to_end(&mut actual)
+                .await
+                .expect("read all historical bytes");
+            assert_eq!(actual, expected);
+        }
+    }
+
    #[tokio::test]
    async fn multipool_delete_marker_stays_with_existing_versions() {
        let bucket = "multipool-marker-routing";
@@ -161,14 +161,52 @@ impl ECStore {
                "incarnation-bound rename requires namespace locking and a non-nil identity",
            ));
        }
+        let bucket = if destination.0 == RUSTFS_META_BUCKET {
+            super::heal::bucket_metadata_owner(destination.1).ok_or(DiskError::FileAccessDenied)?
+        } else {
+            destination.0
+        };
        let guard = Arc::new(
-            self.acquire_bucket_incarnation_fence(destination.0, expected)
+            self.acquire_bucket_incarnation_fence(bucket, expected)
                .await
                .map_err(|error| DiskError::other(error.to_string()))?,
        );
        if guard.is_lock_lost() {
            return Err(DiskError::other("bucket heal incarnation fence was lost"));
        }
+        let transaction = if destination.0 == RUSTFS_META_BUCKET {
+            // The receiver must retain the config fence after an RPC timeout.
+            // A delayed request must also reject a newer config in the same
+            // bucket incarnation before publishing its older repair payload.
+            let transaction = crate::bucket::metadata_sys::acquire_bucket_metadata_transaction_read_lock_in(&self.ctx, bucket)
+                .await
+                .map_err(|error| DiskError::other(error.to_string()))?;
+            let (current, _) = self
+                .get_pool_info_for_delete_marker(
+                    destination.0,
+                    destination.1,
+                    &ObjectOptions {
+                        no_lock: true,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .map_err(|error| DiskError::other(error.to_string()))?;
+            let current = current.object_info;
+            if transaction.is_lock_lost()
+                || guard.is_lock_lost()
+                || fi.mod_time.is_none()
+                || current.mod_time != fi.mod_time
+                || current.data_dir != fi.data_dir
+                || current.size != fi.size
+                || current.etag != fi.get_etag()
+            {
+                return Err(DiskError::FileAccessDenied);
+            }
+            Some(transaction)
+        } else {
+            None
+        };
        rename_local_data_with_ctx(
            &self.ctx,
            disk_ref,
@@ -176,7 +214,7 @@ impl ECStore {
            fi,
            destination,
            RenameDataGuards {
-                external_guard: Some(guard),
+                external_guard: Some(Arc::new((guard, transaction))),
                ..Default::default()
            },
        )
@@ -907,6 +907,21 @@ impl ErasureSetHealer {
        }

        if failed_objects == 0 && skipped_objects == 0 && failed_buckets == 0 {
+            let targets = if self.pool_metadata_target_endpoints.is_empty() {
+                self.target_endpoints.as_ref()
+            } else {
+                self.pool_metadata_target_endpoints.as_ref()
+            };
+            if self.replacement_task_id.is_some() || (!self.heal_opts.dry_run && self.heal_opts.recreate && !targets.is_empty()) {
+                self.verify_replacement_identity_fence("bucket metadata").await?;
+                // Recheck even resumed buckets: their user-object cursor does not
+                // prove that the replacement holds the internal bucket records.
+                for bucket in buckets {
+                    self.storage
+                        .heal_replacement_bucket_metadata(bucket, &self.heal_opts, targets)
+                        .await?;
+                }
+            }
            self.heal_replacement_pool_metadata(
                set_disk_id,
                &mut ErasureSetPassCounters {
@@ -2267,6 +2282,8 @@ mod resume_loop_tests {
        list_include_lifecycle_object_info: Mutex<Vec<bool>>,
        replacement_target_identity_sequences: Mutex<VecDeque<Vec<ReplacementTargetIdentity>>>,
        pool_metadata_placement: Mutex<Option<ReplacementCommitEvidence>>,
+        bucket_metadata_calls: Mutex<Vec<String>>,
+        bucket_metadata_failure: AtomicBool,
        fail_listing: AtomicBool,
        fail_listing_buckets: Mutex<HashSet<String>>,
    }
@@ -2462,6 +2479,13 @@ mod resume_loop_tests {
                None => Ok(true),
            }
        }
+        async fn heal_replacement_bucket_metadata(&self, bucket: &str, _opts: &HealOpts, _targets: &[String]) -> Result<()> {
+            self.bucket_metadata_calls.lock().unwrap().push(bucket.to_owned());
+            if self.bucket_metadata_failure.load(Ordering::SeqCst) {
+                return Err(Error::Storage(EcstoreError::PreconditionFailed));
+            }
+            Ok(())
+        }
        async fn replacement_targets_have_version(
            &self,
            _bucket: &str,
@@ -3058,6 +3082,18 @@ mod resume_loop_tests {
        )
        .with_replacement_targets(vec!["replacement-a".to_string()], Some(replacement_task_id.clone()));

+        env.storage.bucket_metadata_failure.store(true, Ordering::SeqCst);
+        assert!(healer.heal_erasure_set(&["b".to_string()], "pool_0_set_0").await.is_err());
+        let incomplete = ResumeManager::load_replacement_intent(env.healer.disk.clone(), &replacement_task_id)
+            .await
+            .expect("failed metadata repair must retain replacement intent")
+            .get_state()
+            .await;
+        assert!(!incomplete.completed);
+        assert_ne!(incomplete.replacement_phase, crate::heal::resume::ReplacementPhase::Verified);
+        assert!(env.storage.calls().is_empty(), "metadata failure must stop completion before pool.bin");
+        env.storage.bucket_metadata_failure.store(false, Ordering::SeqCst);
+
        let error = healer
            .heal_erasure_set(&["b".to_string()], "pool_0_set_0")
            .await
@@ -3073,6 +3109,11 @@ mod resume_loop_tests {
        assert_eq!(state.replacement_phase, crate::heal::resume::ReplacementPhase::Intent);
        assert_eq!(state.retry_count, 1);
        assert_eq!(env.storage.calls(), vec![(POOL_META_NAME.to_string(), None)]);
+        assert_eq!(
+            *env.storage.bucket_metadata_calls.lock().unwrap(),
+            vec!["b", "b"],
+            "resuming a completed user scan must retry bucket metadata"
+        );
    }

    #[tokio::test]
@@ -532,6 +532,19 @@ pub trait HealStorageAPI: Send + Sync {
        Err(Error::other("storage does not support incarnation-bound object healing"))
    }

+    /// Durable MRF repair may request an authoritative unversioned absence proof.
+    async fn heal_mrf_object_at_incarnation(
+        &self,
+        bucket: &str,
+        object: &str,
+        version_id: Option<&str>,
+        expected: Uuid,
+        opts: &HealOpts,
+    ) -> Result<HealStorageObjectResult> {
+        self.heal_object_at_incarnation(bucket, object, version_id, expected, opts)
+            .await
+    }
+
    /// Heal object using ecstore
    async fn heal_object(
        &self,
@@ -578,6 +591,11 @@ pub trait HealStorageAPI: Send + Sync {
        Err(Error::other("replacement pool metadata placement is unsupported"))
    }

+    /// Heal and physically verify bucket configuration on its replacement targets.
+    async fn heal_replacement_bucket_metadata(&self, _bucket: &str, _opts: &HealOpts, _targets: &[String]) -> Result<()> {
+        Err(Error::Storage(StorageError::PreconditionFailed))
+    }
+
    /// Read target-specific physical evidence for one replacement version.
    ///
    /// This is only used by automatic replacement healing after the normal
@@ -967,6 +985,24 @@ impl HealStorageAPI for ECStoreHealStorage {
            .await)
    }

+    async fn heal_mrf_object_at_incarnation(
+        &self,
+        bucket: &str,
+        object: &str,
+        version_id: Option<&str>,
+        expected: Uuid,
+        opts: &HealOpts,
+    ) -> Result<HealStorageObjectResult> {
+        let result = self
+            .ecstore
+            .heal_mrf_object_at_incarnation(bucket, object, version_id.unwrap_or_default(), expected, opts)
+            .await
+            .map_err(|error| incarnation_storage_error(bucket, expected, error))?;
+        Ok(self
+            .object_result_with_receipt(bucket, object, version_id, opts, result, Some(expected))
+            .await)
+    }
+
    async fn get_object_meta(&self, bucket: &str, object: &str) -> Result<Option<HealObjectInfo>> {
        debug!(
            target: "rustfs::heal::storage",
@@ -1564,6 +1600,13 @@ impl HealStorageAPI for ECStoreHealStorage {
            .map_err(Error::Storage)
    }

+    async fn heal_replacement_bucket_metadata(&self, bucket: &str, opts: &HealOpts, targets: &[String]) -> Result<()> {
+        self.ecstore
+            .heal_replacement_bucket_metadata(bucket, opts, targets)
+            .await
+            .map_err(Error::Storage)
+    }
+
    async fn replacement_targets_have_version(
        &self,
        bucket: &str,
@@ -298,6 +298,18 @@ impl HealTask {
                    "Heal object stage entered"
                );
                let object_size = result.object_size as u64;
+                let ok_drive_state = DriveState::Ok.to_string();
+                if self.source == HealRequestSource::Admin
+                    && !self.options.dry_run
+                    && !result.after.drives.is_empty()
+                    && result.after.drives.iter().any(|drive| drive.state != ok_drive_state)
+                {
+                    return Err(Error::TaskExecutionFailed {
+                        message: format!(
+                            "Heal left one or more drives unhealthy for {bucket}/{object}; retry after the missing drives are restored"
+                        ),
+                    });
+                }
                debug!(
                    target: "rustfs::heal::task",
                    event = EVENT_HEAL_OBJECT_RESULT,
@@ -417,6 +429,10 @@ impl HealTask {

    /// Recreate missing object (for EC decode scenarios)
    async fn recreate_missing_object(&self, bucket: &str, object: &str, version_id: Option<&str>) -> Result<()> {
+        if self.source == HealRequestSource::Mrf {
+            return self.recreate_missing_mrf_object(bucket, object, version_id).await;
+        }
+
        debug!(
            target: "rustfs::heal::task",
            event = EVENT_HEAL_OBJECT_STAGE,
@@ -517,4 +533,65 @@ impl HealTask {
            }
        }
    }
+
+    /// Durable MRF responsibilities may complete only with an exact storage proof.
+    async fn recreate_missing_mrf_object(&self, bucket: &str, object: &str, version_id: Option<&str>) -> Result<()> {
+        let heal_opts = HealOpts {
+            recursive: false,
+            dry_run: self.options.dry_run,
+            remove: false,
+            recreate: true,
+            scan_mode: HealScanMode::Deep,
+            update_parity: true,
+            no_lock: self.options.no_lock,
+            read_repair: false,
+            pool: self.options.pool_index,
+            set: self.options.set_index,
+        };
+        let mut expected = self.outcome_identity(bucket, object, version_id, self.options.pool_index, self.options.set_index);
+        let bucket_incarnation_id = self
+            .outcome_bucket_incarnation_id(bucket, self.options.dry_run)
+            .await?
+            .ok_or_else(|| Error::TaskExecutionFailed {
+                message: format!("Missing bucket incarnation for durable MRF repair {bucket}/{object}"),
+            })?;
+        expected.bucket_incarnation_id = Some(bucket_incarnation_id);
+
+        let storage_result = self
+            .await_with_control(self.storage.heal_mrf_object_at_incarnation(
+                bucket,
+                object,
+                version_id,
+                bucket_incarnation_id,
+                &heal_opts,
+            ))
+            .await?;
+        if let Some(error) = storage_result.error {
+            return Err(Error::TaskExecutionFailed {
+                message: format!("Failed to recreate missing object {bucket}/{object}: {error}"),
+            });
+        }
+
+        let object_size = storage_result.item.object_size as u64;
+        let authoritatively_absent = matches!(
+            storage_result.receipt.as_ref().map(|receipt| &receipt.disposition),
+            Some(HealObjectDisposition::AuthoritativelyAbsent)
+        );
+        if !self.record_verified_storage_receipt(expected, storage_result.receipt).await {
+            return Err(Error::TaskExecutionFailed {
+                message: format!("Missing exact storage proof for durable MRF repair {bucket}/{object}"),
+            });
+        }
+
+        {
+            let mut progress = self.progress.write().await;
+            if authoritatively_absent {
+                progress.update_object_progress(1, 0, 0, 1, 0);
+            } else {
+                progress.update_object_progress(1, 1, 0, 0, object_size);
+            }
+        }
+        self.record_result_item(storage_result.item).await;
+        Ok(())
+    }
 }
@@ -3471,6 +3471,40 @@ async fn test_prefix_heal_lists_and_repairs_objects_under_prefix() {
    );
 }

+#[tokio::test]
+async fn admin_object_heal_fails_when_drive_remains_unhealthy() {
+    let storage = Arc::new(MockStorage::default());
+    storage.heal_object_outcomes.lock().expect("outcomes").insert(
+        "object-a".to_string(),
+        VecDeque::from([MockHealObjectOutcome::UnavailableDrive(DriveState::Missing)]),
+    );
+    let mut request = HealRequest::new(
+        HealType::Object {
+            bucket: "bucket-a".to_string(),
+            object: "object-a".to_string(),
+            version_id: None,
+        },
+        HealOptions {
+            recreate_missing: false,
+            timeout: None,
+            ..Default::default()
+        },
+        HealPriority::Normal,
+    );
+    request.source = HealRequestSource::Admin;
+    let task = HealTask::from_request(request, storage);
+
+    let error = task
+        .execute()
+        .await
+        .expect_err("admin heal must not finish with an unhealthy drive");
+    assert!(error.to_string().contains("left one or more drives unhealthy"));
+    let outcome = task.get_outcome().await;
+    assert_eq!(outcome.counters.failed, 1);
+    assert_eq!(outcome.counters.unknown, 0);
+    assert_eq!(outcome.execution, HealExecutionOutcome::CompletedWithErrors);
+}
+
 #[tokio::test]
 async fn test_data_usage_cache_lock_timeout_does_not_fail_object_heal() {
    let storage = Arc::new(MockStorage::default());
@@ -3855,6 +3889,67 @@ async fn test_heal_recreate_scanner_non_dir_not_found_fails() {
    assert!(matches!(task.get_status().await, HealTaskStatus::Failed { .. }));
 }

+#[tokio::test]
+async fn mrf_recreate_missing_object_records_exact_absence_receipt_with_scope() {
+    let incarnation = Uuid::new_v4();
+    let storage = Arc::new(MockStorage {
+        object_exists: Mutex::new(Some(false)),
+        bucket_incarnation_id: Mutex::new(Some(incarnation)),
+        heal_object_receipts: Mutex::new(HashMap::from([(
+            "deleted.bin".to_string(),
+            VecDeque::from([HealObjectReceipt {
+                identity: HealObjectIdentity {
+                    kind: HealObjectKind::Object,
+                    bucket: "bucket-a".to_string(),
+                    object: "deleted.bin".to_string(),
+                    version_id: None,
+                    bucket_incarnation_id: Some(incarnation),
+                    pool_index: Some(2),
+                    set_index: Some(3),
+                },
+                disposition: HealObjectDisposition::AuthoritativelyAbsent,
+            }]),
+        )])),
+        ..Default::default()
+    });
+    let mut request = HealRequest::new(
+        HealType::Object {
+            bucket: "bucket-a".to_string(),
+            object: "deleted.bin".to_string(),
+            version_id: None,
+        },
+        HealOptions {
+            recreate_missing: true,
+            pool_index: Some(2),
+            set_index: Some(3),
+            timeout: None,
+            ..Default::default()
+        },
+        HealPriority::Normal,
+    );
+    request.source = HealRequestSource::Mrf;
+    let task = HealTask::from_request(request, storage.clone());
+
+    task.execute()
+        .await
+        .expect("a complete MRF absence receipt must complete the task");
+
+    let opts = storage.object_heal_opts.lock().unwrap()[0];
+    assert_eq!(opts.pool, Some(2));
+    assert_eq!(opts.set, Some(3));
+    assert!(opts.recreate);
+    assert_eq!(opts.scan_mode, HealScanMode::Deep);
+    let outcome = task.get_outcome().await;
+    assert_eq!(outcome.counters.unchanged, 1);
+    assert_eq!(outcome.counters.unknown, 0);
+    assert_eq!(outcome.objects.len(), 1);
+    assert_eq!(outcome.objects[0].disposition, HealObjectDisposition::AuthoritativelyAbsent);
+    assert_eq!(
+        (outcome.objects[0].identity.pool_index, outcome.objects[0].identity.set_index),
+        (Some(2), Some(3))
+    );
+}
+
 #[tokio::test]
 async fn test_heal_scanner_missing_object_without_recreate_probes_storage() {
    let storage = Arc::new(MockStorage {
@@ -36,6 +36,7 @@ use rustfs_heal::heal::{
 };
 use rustfs_heal_contracts::heal_channel::{HealOpts, HealScanMode};
 use serial_test::serial;
+use sha2::{Digest, Sha256};
 use std::{
    path::{Path, PathBuf},
    sync::Arc,
@@ -722,6 +723,91 @@ mod serial_tests {
        );
    }

+    #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+    #[serial]
+    async fn test_recursive_deep_heal_reports_stale_delete_marker_repair() {
+        let (disk_paths, ecstore, storage) = heal_env().await;
+        let manager = HealManager::new(
+            storage.clone(),
+            Some(HealConfig {
+                heal_interval: Duration::from_millis(1),
+                ..Default::default()
+            }),
+        );
+        manager.start().await.expect("heal manager should start");
+
+        let bucket = "b5-stale-delete-marker-recursive";
+        let object = "obj.bin";
+        create_versioned_bucket(&ecstore, bucket).await;
+        put_unversioned(&ecstore, bucket, "control/object.bin", &versioned_test_data(42)).await;
+        let historical_data = versioned_test_data(43);
+        let historical = put_versioned(&ecstore, bucket, object, &historical_data).await;
+        let target = &disk_paths[1];
+        let target_meta = xl_meta_path(&object_dir(target, bucket, object));
+        let stale_meta = std::fs::read(&target_meta).expect("target xl.meta must exist before marker creation");
+        let stale_digest = Sha256::digest(&stale_meta);
+        let marker = put_delete_marker(&ecstore, bucket, object).await;
+        std::fs::write(&target_meta, &stale_meta).expect("restore stale target xl.meta");
+
+        let request = HealRequest::new(
+            HealType::Bucket {
+                bucket: bucket.to_string(),
+            },
+            HealOptions {
+                recursive: true,
+                scan_mode: HealScanMode::Deep,
+                pool_index: Some(0),
+                set_index: Some(0),
+                ..Default::default()
+            },
+            HealPriority::Normal,
+        );
+        let task_id = request.id.clone();
+        assert!(
+            manager
+                .submit_heal_request(request)
+                .await
+                .expect("submit recursive heal")
+                .is_admitted()
+        );
+        wait_for_task(&manager, &task_id, Duration::from_secs(60)).await;
+
+        let report = manager.get_task_report(&task_id).await.expect("completed task report");
+        let outcome = report.outcome.expect("recursive task must have an outcome");
+        assert_eq!(outcome.counters.failed, 0);
+        assert_eq!(outcome.counters.unknown, 0);
+        assert_eq!(outcome.counters.processed, 3);
+        assert_eq!(outcome.counters.healed, 1, "delete marker repair must be counted as healed");
+        assert_eq!(outcome.counters.unchanged, 2);
+
+        let repaired_meta = std::fs::read(&target_meta).expect("repaired target xl.meta must exist");
+        assert_ne!(
+            Sha256::digest(&repaired_meta),
+            stale_digest,
+            "deep heal must replace stale rejoined metadata"
+        );
+
+        let marker_info = physical_version(target, bucket, object, &marker);
+        assert!(
+            marker_info.deleted && marker_info.is_latest,
+            "target metadata must converge to delete-marker latest"
+        );
+        let historical_info = physical_version(target, bucket, object, &historical);
+        assert!(!historical_info.is_latest, "historical version must no longer be marked latest");
+        assert_eq!(
+            read_version(&ecstore, bucket, object, &historical).await,
+            historical_data,
+            "historical version must remain readable after marker convergence"
+        );
+        let marker_receipt = outcome
+            .objects
+            .iter()
+            .find(|item| item.identity.version_id.as_deref() == Some(marker.as_str()))
+            .expect("deep heal must emit a receipt for the repaired delete marker");
+        assert_eq!(marker_receipt.disposition, HealObjectDisposition::Repaired);
+        manager.stop().await.expect("heal manager should stop");
+    }
+
    /// A missing xl.meta must produce an exact marker repair receipt, while a
    /// healthy replay must prove metadata health without claiming payload integrity.
    #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
@@ -717,6 +717,49 @@ mod absence_receipt_regressions {
        assert_versions(&store, bucket, &old, &current).await;
    }

+    #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+    #[serial]
+    async fn unversioned_missing_object_produces_fenced_absence_receipt() {
+        let bucket = "absence-receipt-unversioned";
+        let object = "deleted.bin";
+        let (_paths, store, storage) = heal_env_n(16).await;
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create unversioned bucket");
+        let incarnation = storage
+            .admit_bucket_incarnation(bucket)
+            .await
+            .expect("admit bucket incarnation for the durable proof");
+
+        let result = storage
+            .heal_mrf_object_at_incarnation(
+                bucket,
+                object,
+                None,
+                incarnation,
+                &HealOpts {
+                    pool: Some(0),
+                    set: Some(0),
+                    ..deep_heal_opts()
+                },
+            )
+            .await
+            .expect("fenced absence probe should return a storage result");
+        assert!(
+            result.error.is_none(),
+            "complete unversioned absence must not remain an error: {:?}",
+            result.error
+        );
+        let receipt = result.receipt.expect("complete unversioned absence needs a receipt");
+        assert_eq!(receipt.disposition, HealObjectDisposition::AuthoritativelyAbsent);
+        assert_eq!(receipt.identity.bucket, bucket);
+        assert_eq!(receipt.identity.object, object);
+        assert!(receipt.identity.version_id.is_none());
+        assert_eq!((receipt.identity.pool_index, receipt.identity.set_index), (Some(0), Some(0)));
+        assert_eq!(receipt.identity.bucket_incarnation_id, Some(incarnation));
+    }
+
    #[test]
    #[serial]
    fn historical_absence_receipt_bucket_outcome_matches_c06() {
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

-//! Regression for replacement healing requiring pool.bin on a non-owning set.
+//! Replacement healing must restore internal records only on their owning sets.
 #![recursion_limit = "256"]

 use http::HeaderMap;
@@ -101,6 +101,31 @@ async fn replacement_pool_metadata_follows_real_two_set_placement() {
            storage.replacement_pool_metadata_required(&opts).is_err(),
            "invalid scope must fail closed"
        );
+        assert!(
+            storage
+                .heal_replacement_bucket_metadata(bucket, &opts, &[paths[0].to_string_lossy().into_owned()])
+                .await
+                .is_err(),
+            "bucket metadata repair must reject an incomplete or invalid replacement scope"
+        );
+    }
+
+    for targets in [vec![], vec![paths[4].to_string_lossy().into_owned()]] {
+        assert!(
+            storage
+                .heal_replacement_bucket_metadata(
+                    bucket,
+                    &HealOpts {
+                        pool: Some(0),
+                        set: Some(0),
+                        ..Default::default()
+                    },
+                    &targets,
+                )
+                .await
+                .is_err(),
+            "empty or wrong-set targets must be rejected before repairing metadata"
+        );
    }

    let mut owning_set = None;
@@ -120,6 +145,17 @@ async fn replacement_pool_metadata_follows_real_two_set_placement() {
            assert!(owning_set.replace(set).is_none(), "only one set owns pool.bin");
            std::fs::remove_dir_all(&metadata_path).unwrap();
        }
+        let bucket_records: Vec<_> = [".metadata.bin", ".bucket-incarnation"]
+            .into_iter()
+            .map(|file| {
+                let path = target_path.join(RUSTFS_META_BUCKET).join("buckets").join(bucket).join(file);
+                let existed = path.join("xl.meta").exists();
+                if existed {
+                    std::fs::remove_dir_all(&path).expect("remove bucket metadata shard from replacement");
+                }
+                (path, existed)
+            })
+            .collect();
        // Select a user key using the real placement algorithm, independently
        // of the metadata-scope decision under test.
        let key = (0..1000)
@@ -191,6 +227,14 @@ async fn replacement_pool_metadata_follows_real_two_set_placement() {
            "completed repair must clear its healing marker"
        );
        assert_eq!(metadata_path.join("xl.meta").exists(), owns_metadata);
+        for (path, owned) in bucket_records {
+            assert_eq!(
+                path.join("xl.meta").exists(),
+                owned,
+                "replacement must restore each owned bucket record without creating a wrong-set copy: {}",
+                path.display()
+            );
+        }
        for (relative, bytes) in original_parts {
            assert_eq!(std::fs::read(object_path.join(relative)).unwrap(), bytes, "reconstructed shard differs");
        }
@@ -90,6 +90,82 @@ async fn partial_write_persistence_failure_is_reported_and_retained_for_retry()
    );
 }

+#[test]
+fn unversioned_deleted_partial_write_is_discharged_by_an_absence_proof() {
+    const STACK_SIZE: usize = 8 * 1024 * 1024;
+    std::thread::Builder::new()
+        .name("mrf-partial-write-absence".to_owned())
+        .stack_size(STACK_SIZE)
+        .spawn(|| {
+            let runtime = tokio::runtime::Builder::new_current_thread()
+                .thread_stack_size(STACK_SIZE)
+                .enable_all()
+                .build()
+                .expect("partial-write absence runtime should build");
+            runtime.block_on(unversioned_deleted_partial_write_is_discharged_by_an_absence_proof_inner());
+        })
+        .expect("partial-write absence test thread should spawn")
+        .join()
+        .expect("partial-write absence test thread should finish");
+}
+
+async fn unversioned_deleted_partial_write_is_discharged_by_an_absence_proof_inner() {
+    use rustfs_common::mrf_channel::{MrfScope, persist_partial_write_intent};
+
+    temp_env::async_with_vars([("RUSTFS_HEAL_MRF_ENABLE", Some("true"))], async {
+        let root = tempfile::tempdir().expect("partial-write absence fixture directory");
+        let env = TestECStoreEnv::builder().disk_count(16).base_dir(root.path()).build().await;
+        env.make_bucket("partial-absence", false).await;
+        let mut coordinator_pool = env.endpoint_pools.as_ref()[0].clone();
+        let mut endpoints = coordinator_pool.endpoints.as_ref().to_vec();
+        for endpoint in endpoints.iter_mut().skip(4) {
+            endpoint.is_local = false;
+        }
+        coordinator_pool.endpoints = Endpoints::from(endpoints);
+        init_local_disks(EndpointServerPools::from(vec![coordinator_pool]))
+            .await
+            .expect("coordinator journal disks");
+
+        let manager = manager(&env);
+        mrf_queue::spawn_mrf_consumer(manager.clone());
+        for (object, version_id) in [
+            ("deleted-unversioned.bin", None),
+            ("deleted-versioned.bin", Some(uuid::Uuid::new_v4())),
+        ] {
+            persist_partial_write_intent(
+                "partial-absence",
+                object,
+                version_id,
+                MrfScope {
+                    pool_index: 0,
+                    set_index: 0,
+                },
+            )
+            .await
+            .expect("durable partial-write responsibility must commit before scheduling");
+            assert!(snapshot_contains(object).await, "committed responsibility must exist before repair runs");
+        }
+
+        manager.start().await.expect("MRF scheduler should start");
+        for object in ["deleted-unversioned.bin", "deleted-versioned.bin"] {
+            assert!(
+                wait_until(|| async { !snapshot_contains(object).await }).await,
+                "complete absence proof must discharge the durable responsibility"
+            );
+        }
+        assert!(
+            wait_until(|| async {
+                let snapshot = manager.operations_snapshot().await;
+                snapshot.queue_length == 0 && snapshot.active_tasks == 0
+            })
+            .await,
+            "discharged absence repair must leave no queued work"
+        );
+        manager.stop().await.expect("absence manager should stop");
+    })
+    .await;
+}
+
 async fn wait_until<F: FnMut() -> Fut, Fut: Future<Output = bool>>(mut probe: F) -> bool {
    tokio::time::timeout(Duration::from_secs(60), async {
        loop {
@@ -241,7 +241,9 @@ impl RuleValidate for LifecycleRule {
            return Err(std::io::Error::other(ERR_LIFECYCLE_INVALID_DEL_MARKER_EXPIRATION_DAYS));
        }
        // Rule must have at least one action
-        let has_expiration = self.expiration.is_some();
+        let has_expiration = self.expiration.as_ref().is_some_and(|expiration| {
+            expiration.days.is_some() || expiration.date.is_some() || expiration.expired_object_delete_marker.is_some()
+        });
        let has_transition = self.transitions.as_ref().is_some_and(|t| !t.is_empty());
        // `NewerNoncurrentVersions` on its own is a MinIO extension, not an AWS
        // form: it keeps the newest N noncurrent versions and expires the rest
@@ -4455,6 +4457,19 @@ mod tests {
            .expect_err("a negative retention count on a transition must be rejected");
    }

+    #[tokio::test]
+    async fn validate_rejects_empty_expiration_action() {
+        let mut rule = rule_with_filter(LifecycleRuleFilter::default());
+        rule.expiration = Some(LifecycleExpiration::default());
+
+        let err = config_with_rules(vec![rule])
+            .validate(&ObjectLockConfiguration::default())
+            .await
+            .expect_err("an empty Expiration object must not count as an action");
+
+        assert_eq!(err.to_string(), ERR_LIFECYCLE_RULE_MUST_HAVE_ACTION);
+    }
+
    #[tokio::test]
    async fn zero_newer_noncurrent_versions_means_no_count_constraint() {
        // Zero carries no constraint, matching how the batch limit path has
@@ -139,6 +139,7 @@ opentelemetry-stdout = { workspace = true }
 opentelemetry-otlp = { workspace = true, features = ["gzip-http", "reqwest-rustls"] }
 opentelemetry-semantic-conventions = { workspace = true, features = ["semconv_experimental"] }
 percent-encoding = { workspace = true }
+reqwest.workspace = true
 serde = { workspace = true, features = ["derive"] }
 serde_json = { workspace = true, features = ["raw_value"] }
 tracing = { workspace = true, features = ["std", "attributes"] }
@@ -167,3 +168,8 @@ tempfile = { workspace = true }
 temp-env = { workspace = true }
 log.workspace = true
 metrics-util = { workspace = true, features = ["debugging"] }
+rcgen.workspace = true
+rustls = { workspace = true, default-features = false, features = ["aws-lc-rs", "std"] }
+rustls-pki-types.workspace = true
+tokio = { workspace = true, features = ["io-util", "net"] }
+tokio-rustls = { workspace = true, default-features = false, features = ["aws-lc-rs"] }
@@ -115,6 +115,7 @@ All configuration is read from environment variables at startup.
 | `RUSTFS_OBS_TRACE_ENDPOINT`           | _(empty)_ | Dedicated trace endpoint (overrides root + `/v1/traces`)   |
 | `RUSTFS_OBS_METRIC_ENDPOINT`          | _(empty)_ | Dedicated metrics endpoint                                 |
 | `RUSTFS_OBS_LOG_ENDPOINT`             | _(empty)_ | Dedicated log endpoint                                     |
+| `RUSTFS_OBS_TLS_CA_FILE`              | _(empty)_ | Absolute PEM CA bundle added to system trust for OTLP/HTTP |
 | `RUSTFS_OBS_PROFILING_ENDPOINT`       | _(empty)_ | Dedicated profiling endpoint (e.g. Pyroscope)              |
 | `RUSTFS_OBS_TRACES_EXPORT_ENABLED`    | `true`    | Toggle trace export                                        |
 | `RUSTFS_OBS_METRICS_EXPORT_ENABLED`   | `true`    | Toggle metrics export                                      |
@@ -124,6 +125,8 @@ All configuration is read from environment variables at startup.
 | `RUSTFS_OBS_SAMPLE_RATIO`             | `0.1`     | Trace sampling ratio `0.0`–`1.0`                           |
 | `RUSTFS_OBS_METER_INTERVAL`           | `30`      | Metrics export interval (seconds)                          |

+`RUSTFS_OBS_TLS_CA_FILE` is read only while OTLP exporters are initialized. Its PEM certificates are added to, not substituted for, the platform trust roots, and TLS chain, expiry, and hostname checks remain enabled. Restart RustFS after rotating the bundle. A missing, unreadable, non-absolute, empty, or invalid bundle fails OTLP initialization.
+
 The export interval is separate from application metric collection. Without interval overrides, node/disk metrics refresh every 60 seconds (`RUSTFS_METRICS_NODE_INTERVAL_SEC`), cluster metrics every 60 seconds (`RUSTFS_METRICS_CLUSTER_INTERVAL_SEC`), and per-bucket metrics every 300 seconds (`RUSTFS_METRICS_BUCKET_INTERVAL_SEC`). See the [collection defaults](src/metrics/config.rs) and [interval configuration](src/metrics/scheduler.rs) for collector-specific and global overrides.

 With OTLP → Collector → Prometheus, the time from a failure to an alert also includes underlying health detection, collection and export work, transport and Collector batching, Prometheus `scrape_interval` and `evaluation_interval`, and any alert rule `for` duration. A shorter scrape interval cannot refresh a value that RustFS has not yet collected and exported. These periodic intervals do not guarantee a fixed detection or alert latency; see the [Prometheus alert rule documentation](https://prometheus.io/docs/prometheus/latest/configuration/alerting_rules/) for the role of `for`.
@@ -40,7 +40,8 @@ use rustfs_config::observability::{
    ENV_OBS_LOG_STDOUT_ENABLED, ENV_OBS_LOG_ZSTD_COMPRESSION_LEVEL, ENV_OBS_LOG_ZSTD_FALLBACK_TO_GZIP, ENV_OBS_LOG_ZSTD_WORKERS,
    ENV_OBS_LOGGER_LEVEL, ENV_OBS_LOGS_EXPORT_ENABLED, ENV_OBS_METER_INTERVAL, ENV_OBS_METRIC_ENDPOINT,
    ENV_OBS_METRICS_EXPORT_ENABLED, ENV_OBS_PROFILING_ENDPOINT, ENV_OBS_PROFILING_EXPORT_ENABLED, ENV_OBS_SAMPLE_RATIO,
-    ENV_OBS_SERVICE_NAME, ENV_OBS_SERVICE_VERSION, ENV_OBS_TRACE_ENDPOINT, ENV_OBS_TRACES_EXPORT_ENABLED, ENV_OBS_USE_STDOUT,
+    ENV_OBS_SERVICE_NAME, ENV_OBS_SERVICE_VERSION, ENV_OBS_TLS_CA_FILE, ENV_OBS_TRACE_ENDPOINT, ENV_OBS_TRACES_EXPORT_ENABLED,
+    ENV_OBS_USE_STDOUT,
 };
 use rustfs_config::{
    APP_NAME, DEFAULT_LOG_KEEP_FILES, DEFAULT_LOG_LEVEL, DEFAULT_LOG_ROTATION_TIME, DEFAULT_OBS_LOG_FILENAME,
@@ -139,6 +140,9 @@ pub struct OtelConfig {
    pub metric_timeout_millis: Option<u64>,
    /// Timeout (milliseconds) for log OTLP HTTP export.
    pub log_timeout_millis: Option<u64>,
+    /// Absolute path to a PEM bundle whose certificates are added to the
+    /// system trust roots used by OTLP/HTTP exporters.
+    pub tls_ca_file: Option<String>,
    /// Dedicated profiling endpoint.
    pub profiling_endpoint: Option<String>,
    /// Whether to export distributed traces (default: `true`).
@@ -292,6 +296,9 @@ impl OtelConfig {
            trace_timeout_millis: get_env_opt_u64(ENV_OBS_ENDPOINT_TRACES_TIMEOUT_MILLIS),
            metric_timeout_millis: get_env_opt_u64(ENV_OBS_ENDPOINT_METRICS_TIMEOUT_MILLIS),
            log_timeout_millis: get_env_opt_u64(ENV_OBS_ENDPOINT_LOGS_TIMEOUT_MILLIS),
+            tls_ca_file: get_env_opt_str(ENV_OBS_TLS_CA_FILE)
+                .map(|path| path.trim().to_string())
+                .filter(|path| !path.is_empty()),
            profiling_endpoint: get_env_opt_str(ENV_OBS_PROFILING_ENDPOINT),
            traces_export_enabled: Some(get_env_bool(ENV_OBS_TRACES_EXPORT_ENABLED, DEFAULT_OBS_TRACES_EXPORT_ENABLED)),
            metrics_export_enabled: Some(get_env_bool(ENV_OBS_METRICS_EXPORT_ENABLED, DEFAULT_OBS_METRICS_EXPORT_ENABLED)),
@@ -509,4 +516,22 @@ mod tests {
            });
        });
    }
+
+    #[test]
+    fn otlp_tls_ca_file_environment_preserves_unset_empty_and_configured_values() {
+        with_profiling_env_lock(|| {
+            temp_env::with_var_unset(ENV_OBS_TLS_CA_FILE, || {
+                assert_eq!(OtelConfig::extract_otel_config_from_env(None).tls_ca_file, None);
+            });
+            temp_env::with_var(ENV_OBS_TLS_CA_FILE, Some(""), || {
+                assert_eq!(OtelConfig::extract_otel_config_from_env(None).tls_ca_file, None);
+            });
+            temp_env::with_var(ENV_OBS_TLS_CA_FILE, Some("/etc/rustfs/otlp-ca.pem"), || {
+                assert_eq!(
+                    OtelConfig::extract_otel_config_from_env(None).tls_ca_file.as_deref(),
+                    Some("/etc/rustfs/otlp-ca.pem")
+                );
+            });
+        });
+    }
 }
@@ -60,6 +60,18 @@ pub enum TelemetryError {
    BuildMetricExporter(String),
    #[error("Log exporter build failed: {0}")]
    BuildLogExporter(String),
+    #[error("RUSTFS_OBS_TLS_CA_FILE must be an absolute path")]
+    OtlpTlsCaPathNotAbsolute,
+    #[error("Failed to read RUSTFS_OBS_TLS_CA_FILE")]
+    OtlpTlsCaRead(#[source] std::io::Error),
+    #[error("RUSTFS_OBS_TLS_CA_FILE does not contain a PEM certificate")]
+    OtlpTlsCaEmpty,
+    #[error("RUSTFS_OBS_TLS_CA_FILE does not contain a valid PEM certificate")]
+    OtlpTlsCaInvalid,
+    #[error("Failed to parse RUSTFS_OBS_TLS_CA_FILE as a PEM certificate bundle")]
+    OtlpTlsCaParse(#[source] reqwest::Error),
+    #[error("Failed to build OTLP HTTP client")]
+    BuildOtlpHttpClient(#[source] reqwest::Error),
    #[error("Install metrics recorder failed: {0}")]
    InstallMetricsRecorder(String),
    #[error("Tracing subscriber init failed: {0}")]
@@ -89,6 +89,18 @@ fn rust_log_requests_verbose(rust_log: &str) -> bool {
    })
 }

+/// Returns `true` when the effective logging configuration selects a verbose
+/// (`debug`/`trace`) level.
+///
+/// Follows the same precedence as [`build_env_filter`]: `RUST_LOG` overrides
+/// `logger_level` when it is set.
+pub(super) fn effective_verbose_logging(logger_level: &str) -> bool {
+    match env::var("RUST_LOG").ok().filter(|value| !value.trim().is_empty()) {
+        Some(rust_log) => rust_log_requests_verbose(&rust_log),
+        None => is_verbose_level(logger_level),
+    }
+}
+
 fn should_suppress_noisy_crates(logger_level: &str, default_level: Option<&str>, rust_log: Option<&str>) -> bool {
    if let Some(level) = default_level {
        return !is_verbose_level(level);
@@ -36,7 +36,7 @@ use crate::cleaner::LogCleaner;
 use crate::cleaner::types::{CompressionAlgorithm, FileMatchMode};
 use crate::config::OtelConfig;
 use crate::global::{METRIC_LOG_CLEANER_RUN_FAILURES_TOTAL, METRIC_LOG_CLEANER_RUNS_TOTAL, set_observability_metric_enabled};
-use crate::telemetry::filter::{build_env_filter, pyroscope_log_filter};
+use crate::telemetry::filter::{build_env_filter, effective_verbose_logging, pyroscope_log_filter};
 use crate::telemetry::rolling::{RollingAppender, Rotation};
 use metrics::counter;
 use rustfs_config::observability::{
@@ -44,10 +44,11 @@ use rustfs_config::observability::{
    DEFAULT_OBS_LOG_COMPRESSION_ALGORITHM, DEFAULT_OBS_LOG_DELETE_EMPTY_FILES, DEFAULT_OBS_LOG_DRY_RUN,
    DEFAULT_OBS_LOG_GZIP_COMPRESSION_LEVEL, DEFAULT_OBS_LOG_MATCH_MODE, DEFAULT_OBS_LOG_MAX_SINGLE_FILE_SIZE_BYTES,
    DEFAULT_OBS_LOG_MAX_TOTAL_SIZE_BYTES, DEFAULT_OBS_LOG_MIN_FILE_AGE_SECONDS, DEFAULT_OBS_LOG_PARALLEL_COMPRESS,
-    DEFAULT_OBS_LOG_PARALLEL_WORKERS, DEFAULT_OBS_LOG_ZSTD_COMPRESSION_LEVEL, DEFAULT_OBS_LOG_ZSTD_FALLBACK_TO_GZIP,
-    DEFAULT_OBS_LOG_ZSTD_WORKERS,
+    DEFAULT_OBS_LOG_PARALLEL_WORKERS, DEFAULT_OBS_LOG_SPAN_EVENTS, DEFAULT_OBS_LOG_ZSTD_COMPRESSION_LEVEL,
+    DEFAULT_OBS_LOG_ZSTD_FALLBACK_TO_GZIP, DEFAULT_OBS_LOG_ZSTD_WORKERS, ENV_OBS_LOG_SPAN_EVENTS,
 };
 use rustfs_config::{APP_NAME, DEFAULT_LOG_KEEP_FILES, DEFAULT_LOG_ROTATION_TIME};
+use rustfs_utils::get_env_str;
 use serde_json::Value as JsonValue;
 use std::sync::Arc;
 use std::{collections::BTreeMap, fmt};
@@ -78,6 +79,37 @@ pub(super) fn resolve_file_stdout_mirror(configured: Option<bool>, is_production
    configured.unwrap_or(!is_production)
 }

+/// Resolves the span lifecycle events emitted by the JSON log layers.
+///
+/// Defaults to [`FmtSpan::NONE`]. A record per span lifecycle multiplies log
+/// volume on object hot paths and, combined with the ancestor span list,
+/// produced multi-megabyte lines that journald truncated (backlog#2642).
+/// Operators opt back in with `RUSTFS_OBS_LOG_SPAN_EVENTS=close|full`.
+pub(super) fn resolve_span_events() -> FmtSpan {
+    match get_env_str(ENV_OBS_LOG_SPAN_EVENTS, DEFAULT_OBS_LOG_SPAN_EVENTS)
+        .trim()
+        .to_ascii_lowercase()
+        .as_str()
+    {
+        "close" => FmtSpan::CLOSE,
+        "full" => FmtSpan::FULL,
+        "none" | "" => FmtSpan::NONE,
+        unknown => {
+            // An unrecognised value must not silently re-enable span events.
+            eprintln!("{STDERR_WARNING_PREFIX} obs: unknown {ENV_OBS_LOG_SPAN_EVENTS}={unknown:?}; falling back to none");
+            FmtSpan::NONE
+        }
+    }
+}
+
+/// Whether JSON log records embed the full ancestor span list.
+///
+/// `with_span_list(true)` re-serializes every ancestor span on each record, so
+/// it is reserved for verbose levels where that context earns its volume.
+pub(super) fn resolve_span_list(logger_level: &str) -> bool {
+    effective_verbose_logging(logger_level)
+}
+
 #[cfg(unix)]
 pub(super) fn validate_stdout_sink(file_appender: &RollingAppender) -> Result<(), TelemetryError> {
    use std::os::fd::AsFd;
@@ -222,7 +254,12 @@ where
    request_id
 }

-pub(super) fn build_json_log_layer<S, W>(writer: W, enable_ansi: bool, span_events: FmtSpan) -> impl tracing_subscriber::Layer<S>
+pub(super) fn build_json_log_layer<S, W>(
+    writer: W,
+    enable_ansi: bool,
+    span_events: FmtSpan,
+    span_list: bool,
+) -> impl tracing_subscriber::Layer<S>
 where
    S: Subscriber + for<'span> LookupSpan<'span>,
    W: for<'writer> tracing_subscriber::fmt::MakeWriter<'writer> + Send + Sync + 'static,
@@ -239,7 +276,7 @@ where
        .json()
        .flatten_event(true)
        .with_current_span(true)
-        .with_span_list(true)
+        .with_span_list(span_list)
        .with_span_events(span_events)
        .map_event_format(RequestIdJsonFormat::new)
 }
@@ -311,11 +348,12 @@ pub(super) fn init_local_logging(
 fn init_stdout_only(_config: &OtelConfig, logger_level: &str, is_production: bool) -> Result<OtelGuard, TelemetryError> {
    let env_filter = build_env_filter(logger_level, None);
    let (nb, guard) = tracing_appender::non_blocking(std::io::stdout());
-    let span_events = if is_production { FmtSpan::CLOSE } else { FmtSpan::FULL };
+    let span_events = resolve_span_events();
+    let span_list = resolve_span_list(logger_level);

    // Keep stdout formatting JSON-shaped even in local-only mode so operators
    // can ship the same log schema to external collectors if needed.
-    let fmt_layer = build_json_log_layer(nb, std::io::stdout().is_terminal(), span_events);
+    let fmt_layer = build_json_log_layer(nb, std::io::stdout().is_terminal(), span_events, span_list);

    tracing_subscriber::registry()
        .with(env_filter)
@@ -413,11 +451,12 @@ fn init_file_logging_internal(

    // ── 4. Build subscriber layers ────────────────────────────────────────────
    let env_filter = build_env_filter(logger_level, None);
-    let span_events = if is_production { FmtSpan::CLOSE } else { FmtSpan::FULL };
+    let span_events = resolve_span_events();
+    let span_list = resolve_span_list(logger_level);

    // File output stays machine-readable and free of ANSI sequences so the
    // resulting files are safe to parse or ship to log processors.
-    let file_layer = build_json_log_layer(non_blocking, false, span_events.clone());
+    let file_layer = build_json_log_layer(non_blocking, false, span_events.clone(), span_list);

    // Optional stdout mirror: enabled explicitly via `log_stdout_enabled`, or
    // unconditionally in non-production environments so developers still see
@@ -425,7 +464,10 @@ fn init_file_logging_internal(
    let (stdout_layer, stdout_guard) = if resolve_file_stdout_mirror(config.log_stdout_enabled, is_production) {
        let (stdout_nb, stdout_guard) = tracing_appender::non_blocking(std::io::stdout());
        let enable_color = std::io::stdout().is_terminal();
-        (Some(build_json_log_layer(stdout_nb, enable_color, span_events)), Some(stdout_guard))
+        (
+            Some(build_json_log_layer(stdout_nb, enable_color, span_events, span_list)),
+            Some(stdout_guard),
+        )
    } else {
        (None, None)
    };
@@ -773,7 +815,7 @@ mod tests {

    fn render_json_log(enable_ansi: bool) -> (String, Value) {
        let writer = SharedWriter::default();
-        let layer = build_json_log_layer(writer.clone(), enable_ansi, FmtSpan::NONE);
+        let layer = build_json_log_layer(writer.clone(), enable_ansi, FmtSpan::NONE, false);
        let subscriber = Registry::default().with(layer);

        tracing::subscriber::with_default(subscriber, || {
@@ -793,9 +835,32 @@ mod tests {
        (raw, parsed)
    }

+    fn render_event_in_nested_spans(depth: usize, span_list: bool) -> String {
+        let writer = SharedWriter::default();
+        let layer = build_json_log_layer(writer.clone(), false, FmtSpan::NONE, span_list);
+        let subscriber = Registry::default().with(layer);
+
+        tracing::subscriber::with_default(subscriber, || {
+            let spans = (0..depth)
+                .map(|level| tracing::info_span!("layer", depth = level, payload = "x".repeat(1024)))
+                .collect::<Vec<_>>();
+            let _guards = spans.iter().map(tracing::Span::enter).collect::<Vec<_>>();
+            info!(component = "obs_contract_test", "nested event");
+        });
+
+        String::from_utf8(
+            writer
+                .inner
+                .lock()
+                .expect("shared writer lock should not be poisoned")
+                .clone(),
+        )
+        .expect("log output should be valid UTF-8")
+    }
+
    fn render_json_log_with_request_span() -> Value {
        let writer = SharedWriter::default();
-        let layer = build_json_log_layer(writer.clone(), false, FmtSpan::NONE);
+        let layer = build_json_log_layer(writer.clone(), false, FmtSpan::NONE, false);
        let subscriber = Registry::default().with(layer);

        tracing::subscriber::with_default(subscriber, || {
@@ -818,7 +883,7 @@ mod tests {

    fn render_json_log_with_recovery_monitor_child_span() -> Value {
        let writer = SharedWriter::default();
-        let layer = build_json_log_layer(writer.clone(), false, FmtSpan::NONE);
+        let layer = build_json_log_layer(writer.clone(), false, FmtSpan::NONE, true);
        let subscriber = Registry::default().with(layer);

        tracing::subscriber::with_default(subscriber, || {
@@ -1075,6 +1140,10 @@ mod tests {
        assert_eq!(parsed["request-id"], Value::String("req-123".to_string()));
        assert_eq!(parsed["message"], Value::String("inside request span".to_string()));
        assert_eq!(parsed["span"]["request_id"], Value::String("req-123".to_string()));
+        assert!(
+            parsed.get("spans").is_none(),
+            "ancestor span list must stay out of records when span_list is disabled: {parsed}"
+        );
    }

    #[test]
@@ -1112,4 +1181,65 @@ mod tests {

        assert_eq!(resolve_log_cleanup_file_pattern(&config, "rustfs.log"), "rustfs.log");
    }
+
+    /// Regression for backlog#2642: records must not embed the ancestor span
+    /// list, so their size stays flat as the enclosing span chain deepens.
+    #[test]
+    fn record_size_does_not_grow_with_span_depth_when_span_list_is_disabled() {
+        let shallow = render_event_in_nested_spans(1, false);
+        let deep = render_event_in_nested_spans(32, false);
+
+        assert!(
+            deep.len() <= shallow.len() * 2,
+            "ancestor spans leaked into the record: shallow={} deep={}",
+            shallow.len(),
+            deep.len()
+        );
+
+        // The flag is what keeps it bounded: re-enabling the list must show the
+        // growth the default avoids.
+        let deep_with_list = render_event_in_nested_spans(32, true);
+        assert!(
+            deep_with_list.len() > deep.len(),
+            "enabling the span list must reintroduce per-ancestor growth"
+        );
+    }
+
+    #[test]
+    fn span_events_default_to_none_and_parse_overrides() {
+        temp_env::with_var(ENV_OBS_LOG_SPAN_EVENTS, None::<&str>, || {
+            assert_eq!(resolve_span_events(), FmtSpan::NONE);
+        });
+        temp_env::with_var(ENV_OBS_LOG_SPAN_EVENTS, Some("close"), || {
+            assert_eq!(resolve_span_events(), FmtSpan::CLOSE);
+        });
+        temp_env::with_var(ENV_OBS_LOG_SPAN_EVENTS, Some("FULL"), || {
+            assert_eq!(resolve_span_events(), FmtSpan::FULL);
+        });
+        temp_env::with_var(ENV_OBS_LOG_SPAN_EVENTS, Some("verbose"), || {
+            assert_eq!(
+                resolve_span_events(),
+                FmtSpan::NONE,
+                "an unrecognised value must not re-enable span events"
+            );
+        });
+    }
+
+    #[test]
+    fn span_list_is_reserved_for_verbose_levels() {
+        temp_env::with_var("RUST_LOG", None::<&str>, || {
+            assert!(!resolve_span_list("info"), "info records must not embed the ancestor span list");
+            assert!(!resolve_span_list("warn"));
+            assert!(!resolve_span_list("error"));
+            assert!(resolve_span_list("debug"));
+            assert!(resolve_span_list("trace"));
+        });
+
+        temp_env::with_var("RUST_LOG", Some("debug"), || {
+            assert!(
+                resolve_span_list("info"),
+                "RUST_LOG=debug overrides the base level and should keep span context"
+            );
+        });
+    }
 }
@@ -64,11 +64,11 @@ use rustfs_config::{
    DEFAULT_OBS_METRICS_EXPORT_ENABLED, DEFAULT_OBS_TRACES_EXPORT_ENABLED, METER_INTERVAL, SAMPLE_RATIO,
 };
 use std::collections::HashMap;
-use std::{fs, io::IsTerminal, time::Duration};
+use std::{fs, io::IsTerminal, path::Path, time::Duration};
 use tracing::{info, warn};
 use tracing_error::ErrorLayer;
 use tracing_opentelemetry::{MetricsLayer, OpenTelemetryLayer};
-use tracing_subscriber::{fmt::format::FmtSpan, layer::SubscriberExt, util::SubscriberInitExt};
+use tracing_subscriber::{layer::SubscriberExt, util::SubscriberInitExt};

 const GET_OBJECT_DURATION_HISTOGRAM_METRICS: &[&str] = &[
    "rustfs_io_get_object_request_duration_seconds",
@@ -83,6 +83,12 @@ const GET_OBJECT_DURATION_HISTOGRAM_BUCKETS: &[f64] = &[
    0.1, 0.25, 0.5, 1.0, 2.5, 5.0, 10.0,
 ];

+const DEFAULT_OTLP_HTTP_TIMEOUT: Duration = Duration::from_secs(10);
+const OTEL_EXPORTER_OTLP_TIMEOUT: &str = "OTEL_EXPORTER_OTLP_TIMEOUT";
+const OTEL_EXPORTER_OTLP_TRACES_TIMEOUT: &str = "OTEL_EXPORTER_OTLP_TRACES_TIMEOUT";
+const OTEL_EXPORTER_OTLP_METRICS_TIMEOUT: &str = "OTEL_EXPORTER_OTLP_METRICS_TIMEOUT";
+const OTEL_EXPORTER_OTLP_LOGS_TIMEOUT: &str = "OTEL_EXPORTER_OTLP_LOGS_TIMEOUT";
+
 #[cfg(all(
    feature = "pyroscope",
    any(target_os = "macos", all(target_os = "linux", target_env = "gnu", target_arch = "x86_64"))
@@ -130,6 +136,8 @@ pub(super) fn init_observability_http(
    logger_level: &str,
    is_production: bool,
 ) -> Result<OtelGuard, TelemetryError> {
+    let otlp_tls_ca_bundle = OtlpTlsCaBundle::load(config)?;
+
    // ── Resource & sampling ──────────────────────────────────────────────────
    // Build the common resource once so all enabled signals report the same
    // service identity and deployment metadata.
@@ -153,10 +161,12 @@ pub(super) fn init_observability_http(
    let log_ep = resolve_signal_endpoint(config.log_endpoint.as_deref(), root_ep, "/v1/logs");

    // ── Tracer provider (HTTP) ────────────────────────────────────────────────
-    let tracer_provider = build_tracer_provider(&trace_ep, config, res.clone(), sampler, use_stdout)?;
+    let tracer_provider =
+        build_tracer_provider(&trace_ep, config, otlp_tls_ca_bundle.as_ref(), res.clone(), sampler, use_stdout)?;

    // ── Meter provider (HTTP) ─────────────────────────────────────────────────
-    let meter_provider = build_meter_provider(&metric_ep, config, res.clone(), &service_name, use_stdout)?;
+    let meter_provider =
+        build_meter_provider(&metric_ep, config, otlp_tls_ca_bundle.as_ref(), res.clone(), &service_name, use_stdout)?;

    // ── Logger Logic ──────────────────────────────────────────────────────────
    // Logging is the only signal that may intentionally route to either OTLP
@@ -175,7 +185,7 @@ pub(super) fn init_observability_http(
        // Init OTLP logger logic.
        // We initialize the OTLP collector and honor the configured stdout setting
        // (e.g. via RUSTFS_OBS_USE_STDOUT / config.use_stdout) when building the provider.
-        logger_provider = build_logger_provider(&log_ep, config, res, use_stdout)?;
+        logger_provider = build_logger_provider(&log_ep, config, otlp_tls_ca_bundle.as_ref(), res, use_stdout)?;

        // Build bridge to capture `tracing` events.
        otel_bridge = logger_provider.as_ref().map(OpenTelemetryTracingBridge::new);
@@ -184,7 +194,8 @@ pub(super) fn init_observability_http(
        // active, the OpenTelemetry bridge is the authoritative sink for
        // `tracing` events unless local file logging is needed as a fallback.
    }
-    let span_events = if is_production { FmtSpan::CLOSE } else { FmtSpan::FULL };
+    let span_events = crate::telemetry::local::resolve_span_events();
+    let span_list = crate::telemetry::local::resolve_span_list(logger_level);
    // ── Case 2: File Logging
    // If a log directory is configured and OTLP log export is unavailable, use
    // the same rolling-file behavior as the local-only telemetry backend.
@@ -221,7 +232,7 @@ pub(super) fn init_observability_http(
            crate::telemetry::local::validate_stdout_sink(&file_appender)?;

            let (non_blocking, guard) = tracing_appender::non_blocking(file_appender);
-            let file_layer = build_json_log_layer(non_blocking, false, span_events.clone());
+            let file_layer = build_json_log_layer(non_blocking, false, span_events.clone(), span_list);
            let cleanup_handle = spawn_cleanup_task(config, log_directory, log_filename, keep_files);
            Ok((file_layer, guard, cleanup_handle, rotation_str))
        })();
@@ -267,7 +278,7 @@ pub(super) fn init_observability_http(
    if force_stdout_logging || crate::telemetry::local::resolve_file_stdout_mirror(config.log_stdout_enabled, is_production) {
        let (stdout_nb, stdout_g) = tracing_appender::non_blocking(std::io::stdout());
        stdout_guard = Some(stdout_g);
-        stdout_layer_opt = Some(build_json_log_layer(stdout_nb, std::io::stdout().is_terminal(), span_events));
+        stdout_layer_opt = Some(build_json_log_layer(stdout_nb, std::io::stdout().is_terminal(), span_events, span_list));
    }
    let local_file_fallback_enabled = file_layer_opt.is_some();
    let stdout_mirror_enabled = stdout_guard.is_some();
@@ -320,6 +331,7 @@ pub(super) fn init_observability_http(
 fn build_tracer_provider(
    trace_ep: &str,
    config: &OtelConfig,
+    otlp_tls_ca_bundle: Option<&OtlpTlsCaBundle>,
    res: opentelemetry_sdk::Resource,
    sampler: Sampler,
    use_stdout: bool,
@@ -337,9 +349,13 @@ fn build_tracer_provider(
    if !trace_headers.is_empty() {
        exporter_builder = exporter_builder.with_headers(trace_headers);
    }
-    if let Some(timeout) = resolve_signal_timeout(config.endpoint_timeout_millis, config.trace_timeout_millis) {
+    let timeout = resolve_signal_timeout(config.endpoint_timeout_millis, config.trace_timeout_millis);
+    if let Some(timeout) = timeout {
        exporter_builder = exporter_builder.with_timeout(timeout);
    }
+    if let Some(http_client) = build_otlp_http_client(otlp_tls_ca_bundle, timeout, OTEL_EXPORTER_OTLP_TRACES_TIMEOUT)? {
+        exporter_builder = exporter_builder.with_http_client(http_client);
+    }
    let exporter = exporter_builder
        .build()
        .map_err(|e| TelemetryError::BuildSpanExporter(e.to_string()))?;
@@ -408,6 +424,7 @@ fn resolve_signal_endpoint(signal_endpoint: Option<&str>, root_ep: &str, suffix:
 fn build_meter_provider(
    metric_ep: &str,
    config: &OtelConfig,
+    otlp_tls_ca_bundle: Option<&OtlpTlsCaBundle>,
    res: opentelemetry_sdk::Resource,
    service_name: &str,
    use_stdout: bool,
@@ -426,9 +443,13 @@ fn build_meter_provider(
    if !metric_headers.is_empty() {
        exporter_builder = exporter_builder.with_headers(metric_headers);
    }
-    if let Some(timeout) = resolve_signal_timeout(config.endpoint_timeout_millis, config.metric_timeout_millis) {
+    let timeout = resolve_signal_timeout(config.endpoint_timeout_millis, config.metric_timeout_millis);
+    if let Some(timeout) = timeout {
        exporter_builder = exporter_builder.with_timeout(timeout);
    }
+    if let Some(http_client) = build_otlp_http_client(otlp_tls_ca_bundle, timeout, OTEL_EXPORTER_OTLP_METRICS_TIMEOUT)? {
+        exporter_builder = exporter_builder.with_http_client(http_client);
+    }
    let exporter = exporter_builder
        .build()
        .map_err(|e| TelemetryError::BuildMetricExporter(e.to_string()))?;
@@ -484,6 +505,7 @@ fn is_get_object_duration_histogram_metric(name: &str) -> bool {
 fn build_logger_provider(
    log_ep: &str,
    config: &OtelConfig,
+    otlp_tls_ca_bundle: Option<&OtlpTlsCaBundle>,
    res: opentelemetry_sdk::Resource,
    use_stdout: bool,
 ) -> Result<Option<SdkLoggerProvider>, TelemetryError> {
@@ -500,9 +522,13 @@ fn build_logger_provider(
    if !log_headers.is_empty() {
        exporter_builder = exporter_builder.with_headers(log_headers);
    }
-    if let Some(timeout) = resolve_signal_timeout(config.endpoint_timeout_millis, config.log_timeout_millis) {
+    let timeout = resolve_signal_timeout(config.endpoint_timeout_millis, config.log_timeout_millis);
+    if let Some(timeout) = timeout {
        exporter_builder = exporter_builder.with_timeout(timeout);
    }
+    if let Some(http_client) = build_otlp_http_client(otlp_tls_ca_bundle, timeout, OTEL_EXPORTER_OTLP_LOGS_TIMEOUT)? {
+        exporter_builder = exporter_builder.with_http_client(http_client);
+    }
    let exporter = exporter_builder
        .build()
        .map_err(|e| TelemetryError::BuildLogExporter(e.to_string()))?;
@@ -659,16 +685,239 @@ fn resolve_signal_timeout(common_timeout_millis: Option<u64>, signal_timeout_mil
        .map(Duration::from_millis)
 }

+struct OtlpTlsCaBundle {
+    certificates: Vec<reqwest::Certificate>,
+}
+
+impl OtlpTlsCaBundle {
+    fn load(config: &OtelConfig) -> Result<Option<Self>, TelemetryError> {
+        let Some(ca_file) = config.tls_ca_file.as_deref() else {
+            return Ok(None);
+        };
+        let path = Path::new(ca_file);
+        if !path.is_absolute() {
+            return Err(TelemetryError::OtlpTlsCaPathNotAbsolute);
+        }
+
+        let pem_bundle = fs::read(path).map_err(TelemetryError::OtlpTlsCaRead)?;
+        if pem_bundle.iter().all(u8::is_ascii_whitespace) {
+            return Err(TelemetryError::OtlpTlsCaEmpty);
+        }
+        let certificates = reqwest::Certificate::from_pem_bundle(&pem_bundle).map_err(TelemetryError::OtlpTlsCaParse)?;
+        if certificates.is_empty() {
+            return Err(TelemetryError::OtlpTlsCaInvalid);
+        }
+        Ok(Some(Self { certificates }))
+    }
+
+    fn build_http_client(
+        &self,
+        configured_timeout: Option<Duration>,
+        otel_signal_timeout_env: &str,
+    ) -> Result<reqwest::Client, TelemetryError> {
+        let timeout = configured_timeout.unwrap_or_else(|| resolve_otlp_http_timeout(otel_signal_timeout_env));
+        reqwest::Client::builder()
+            .timeout(timeout)
+            .tls_certs_merge(self.certificates.clone())
+            .build()
+            .map_err(TelemetryError::BuildOtlpHttpClient)
+    }
+}
+
+fn build_otlp_http_client(
+    otlp_tls_ca_bundle: Option<&OtlpTlsCaBundle>,
+    configured_timeout: Option<Duration>,
+    otel_signal_timeout_env: &str,
+) -> Result<Option<reqwest::Client>, TelemetryError> {
+    otlp_tls_ca_bundle
+        .map(|bundle| bundle.build_http_client(configured_timeout, otel_signal_timeout_env))
+        .transpose()
+}
+
+fn resolve_otlp_http_timeout(signal_timeout_env: &str) -> Duration {
+    [signal_timeout_env, OTEL_EXPORTER_OTLP_TIMEOUT]
+        .into_iter()
+        .find_map(|name| std::env::var(name).ok().and_then(|value| value.parse::<u64>().ok()))
+        .map(Duration::from_millis)
+        .unwrap_or(DEFAULT_OTLP_HTTP_TIMEOUT)
+}
+
 #[cfg(test)]
 mod tests {
    use super::*;
    use std::env;
    use std::io::{self, Write};
+    use std::path::Path;
    use std::process::Command;
    use std::sync::{Arc, Mutex};
+    use tokio::io::{AsyncReadExt, AsyncWriteExt};
+    use tokio::net::TcpListener;

    const LOG_TRACER_CHILD_ENV: &str = "RUSTFS_OBS_LOG_TRACER_CHILD";

+    fn config_with_tls_ca_file(path: &Path) -> OtelConfig {
+        OtelConfig {
+            tls_ca_file: Some(path.display().to_string()),
+            ..OtelConfig::default()
+        }
+    }
+
+    fn load_tls_ca_bundle(path: &Path) -> OtlpTlsCaBundle {
+        OtlpTlsCaBundle::load(&config_with_tls_ca_file(path))
+            .expect("load CA bundle")
+            .expect("configured CA file should produce a bundle")
+    }
+
+    async fn spawn_test_tls_server() -> (String, String, tokio::task::JoinHandle<bool>) {
+        let _ = rustls::crypto::aws_lc_rs::default_provider().install_default();
+        let certified =
+            rcgen::generate_simple_self_signed(vec!["127.0.0.1".to_string()]).expect("generate TLS server certificate");
+        let ca_pem = certified.cert.pem();
+        let private_key = rustls_pki_types::PrivateKeyDer::try_from(certified.signing_key.serialize_der())
+            .expect("convert TLS server private key");
+        let config = rustls::ServerConfig::builder()
+            .with_no_client_auth()
+            .with_single_cert(vec![certified.cert.der().clone()], private_key)
+            .expect("build TLS server config");
+        let acceptor = tokio_rustls::TlsAcceptor::from(Arc::new(config));
+        let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind TLS test server");
+        let endpoint = format!("https://{}", listener.local_addr().expect("read TLS test server address"));
+        let server = tokio::spawn(async move {
+            let Ok((stream, _)) = listener.accept().await else {
+                return false;
+            };
+            let Ok(mut stream) = acceptor.accept(stream).await else {
+                return false;
+            };
+            let mut request = Vec::new();
+            let mut buffer = [0_u8; 1_024];
+            loop {
+                let Ok(read) = stream.read(&mut buffer).await else {
+                    return false;
+                };
+                if read == 0 {
+                    return false;
+                }
+                request.extend_from_slice(&buffer[..read]);
+                if request.windows(4).any(|window| window == b"\r\n\r\n") {
+                    break;
+                }
+            }
+            stream
+                .write_all(b"HTTP/1.1 200 OK\r\ncontent-length: 2\r\nconnection: close\r\n\r\nok")
+                .await
+                .is_ok()
+                && stream.shutdown().await.is_ok()
+        });
+        (endpoint, ca_pem, server)
+    }
+
+    #[tokio::test]
+    async fn otlp_http_client_trusts_every_certificate_in_a_pem_bundle() {
+        let (first_endpoint, first_ca_pem, first_server) = spawn_test_tls_server().await;
+        let (second_endpoint, second_ca_pem, second_server) = spawn_test_tls_server().await;
+        let file = tempfile::NamedTempFile::new().expect("create CA bundle file");
+        std::fs::write(file.path(), format!("{first_ca_pem}\n{second_ca_pem}")).expect("write CA bundle");
+
+        let bundle = load_tls_ca_bundle(file.path());
+        let client = build_otlp_http_client(Some(&bundle), Some(Duration::from_millis(250)), OTEL_EXPORTER_OTLP_TRACES_TIMEOUT)
+            .expect("build client with CA bundle");
+        let client = client.expect("custom CA should build a client");
+        assert_eq!(
+            client
+                .get(first_endpoint)
+                .send()
+                .await
+                .expect("first bundled CA should complete TLS handshake")
+                .status(),
+            reqwest::StatusCode::OK
+        );
+        assert_eq!(
+            client
+                .get(second_endpoint)
+                .send()
+                .await
+                .expect("second bundled CA should complete TLS handshake")
+                .status(),
+            reqwest::StatusCode::OK
+        );
+        assert!(first_server.await.expect("first bundled TLS server task"));
+        assert!(second_server.await.expect("second bundled TLS server task"));
+    }
+
+    #[test]
+    fn otlp_http_client_rejects_relative_missing_empty_and_invalid_ca_files() {
+        let relative = OtelConfig {
+            tls_ca_file: Some("otlp-ca.pem".to_string()),
+            ..OtelConfig::default()
+        };
+        assert!(matches!(OtlpTlsCaBundle::load(&relative), Err(TelemetryError::OtlpTlsCaPathNotAbsolute)));
+
+        let missing_directory = tempfile::tempdir().expect("create missing CA directory");
+        let missing = OtelConfig {
+            tls_ca_file: Some(missing_directory.path().join("otlp-ca.pem").display().to_string()),
+            ..OtelConfig::default()
+        };
+        assert!(matches!(OtlpTlsCaBundle::load(&missing), Err(TelemetryError::OtlpTlsCaRead(_))));
+
+        let empty = tempfile::NamedTempFile::new().expect("create empty CA file");
+        assert!(matches!(
+            OtlpTlsCaBundle::load(&config_with_tls_ca_file(empty.path())),
+            Err(TelemetryError::OtlpTlsCaEmpty)
+        ));
+
+        let invalid = tempfile::NamedTempFile::new().expect("create invalid CA file");
+        std::fs::write(invalid.path(), b"not a PEM certificate").expect("write invalid CA file");
+        assert!(matches!(
+            OtlpTlsCaBundle::load(&config_with_tls_ca_file(invalid.path())),
+            Err(TelemetryError::OtlpTlsCaInvalid)
+        ));
+    }
+
+    #[tokio::test]
+    async fn otlp_http_client_trusts_the_configured_ca_and_rejects_another_ca() {
+        let (endpoint, ca_pem, trusted_server) = spawn_test_tls_server().await;
+        let trusted_file = tempfile::NamedTempFile::new().expect("create trusted CA file");
+        std::fs::write(trusted_file.path(), ca_pem).expect("write trusted CA file");
+        let trusted_bundle = load_tls_ca_bundle(trusted_file.path());
+        let trusted_client =
+            build_otlp_http_client(Some(&trusted_bundle), Some(Duration::from_secs(1)), OTEL_EXPORTER_OTLP_TRACES_TIMEOUT)
+                .expect("build trusted client")
+                .expect("custom CA should build a client");
+        assert_eq!(
+            trusted_client
+                .get(&endpoint)
+                .send()
+                .await
+                .expect("configured CA should complete TLS handshake")
+                .status(),
+            reqwest::StatusCode::OK
+        );
+        assert!(trusted_server.await.expect("trusted TLS server task"));
+
+        let (endpoint, _ca_pem, untrusted_server) = spawn_test_tls_server().await;
+        let wrong_ca =
+            rcgen::generate_simple_self_signed(vec!["wrong.test".to_string()]).expect("generate unrelated CA certificate");
+        let untrusted_file = tempfile::NamedTempFile::new().expect("create unrelated CA file");
+        std::fs::write(untrusted_file.path(), wrong_ca.cert.pem()).expect("write unrelated CA file");
+        let untrusted_bundle = load_tls_ca_bundle(untrusted_file.path());
+        let untrusted_client =
+            build_otlp_http_client(Some(&untrusted_bundle), Some(Duration::from_secs(1)), OTEL_EXPORTER_OTLP_TRACES_TIMEOUT)
+                .expect("build untrusted client")
+                .expect("custom CA should build a client");
+        assert!(untrusted_client.get(&endpoint).send().await.is_err());
+        assert!(!untrusted_server.await.expect("untrusted TLS server task"));
+    }
+
+    #[test]
+    fn otlp_http_client_is_not_built_without_a_custom_ca_file() {
+        assert!(
+            build_otlp_http_client(None, None, OTEL_EXPORTER_OTLP_TRACES_TIMEOUT)
+                .expect("no CA file keeps the default client")
+                .is_none()
+        );
+    }
+
    #[derive(Clone)]
    struct TestWriter(Arc<Mutex<Vec<u8>>>);

@@ -133,6 +133,8 @@ pub struct DeleteObjectCall {
 }

 struct Inner {
+    capture_upload_bodies: bool,
+    upload_bodies: Vec<Vec<u8>>,
    // Response queues. Each method pops from its own queue. Empty queue
    // plus no default means a configured-miss error.
    get_object: VecDeque<Result<GetObjectOutput, DummyError>>,
@@ -190,6 +192,8 @@ struct Inner {
 impl Inner {
    fn new() -> Self {
        Self {
+            capture_upload_bodies: false,
+            upload_bodies: Vec::new(),
            get_object: VecDeque::new(),
            get_object_range: VecDeque::new(),
            put_object: VecDeque::new(),
@@ -265,6 +269,30 @@ impl DummyBackend {
        }
    }

+    /// Opt in to consuming upload bodies for byte-for-byte protocol tests.
+    pub fn capture_upload_bodies(&self) {
+        self.inner.lock().expect("lock").capture_upload_bodies = true;
+    }
+
+    pub fn upload_bodies(&self) -> Vec<Vec<u8>> {
+        self.inner.lock().expect("lock").upload_bodies.clone()
+    }
+
+    async fn record_upload_body(&self, body: &mut Option<StreamingBlob>) -> Result<(), DummyError> {
+        use futures_util::TryStreamExt;
+        if !self.inner.lock().expect("lock").capture_upload_bodies {
+            return Ok(());
+        }
+        let mut bytes = Vec::new();
+        if let Some(mut stream) = body.take() {
+            while let Some(chunk) = stream.try_next().await.map_err(|e| DummyError::Injected(e.to_string()))? {
+                bytes.extend_from_slice(&chunk);
+            }
+        }
+        self.inner.lock().expect("lock").upload_bodies.push(bytes);
+        Ok(())
+    }
+
    // Queue-configuration helpers. Each test stages the responses it
    // expects in order. The method pops in FIFO order.

@@ -629,7 +657,8 @@ impl StorageBackend for DummyBackend {
        }
    }

-    async fn put_object(&self, input: PutObjectInput, _credentials: &Credentials) -> Result<PutObjectOutput, Self::Error> {
+    async fn put_object(&self, mut input: PutObjectInput, _credentials: &Credentials) -> Result<PutObjectOutput, Self::Error> {
+        self.record_upload_body(&mut input.body).await?;
        // Decide control flow while holding the lock. Release before
        // awaiting so the stall path does not hold the Mutex across
        // an await point.
@@ -789,7 +818,8 @@ impl StorageBackend for DummyBackend {
        }
    }

-    async fn upload_part(&self, input: UploadPartInput, _credentials: &Credentials) -> Result<UploadPartOutput, Self::Error> {
+    async fn upload_part(&self, mut input: UploadPartInput, _credentials: &Credentials) -> Result<UploadPartOutput, Self::Error> {
+        self.record_upload_body(&mut input.body).await?;
        // Record the call and decide the control flow while holding the
        // lock. Release the lock before awaiting so the stall path does
        // not hold the Mutex across an await point.
@@ -150,10 +150,10 @@ pub fn is_operation_supported(protocol: super::session::Protocol, action: &S3Act
            S3Action::HeadObject => true,   // SIZE command

            // Multipart operations
-            S3Action::CreateMultipartUpload => false,
-            S3Action::UploadPart => false,
-            S3Action::CompleteMultipartUpload => false,
-            S3Action::AbortMultipartUpload => false,
+            S3Action::CreateMultipartUpload => true,
+            S3Action::UploadPart => true,
+            S3Action::CompleteMultipartUpload => true,
+            S3Action::AbortMultipartUpload => true,
            S3Action::ListMultipartUploads => false,
            S3Action::ListParts => false,

@@ -16,12 +16,12 @@ use crate::common::client::s3::StorageBackend as S3StorageBackend;
 use crate::common::gateway::S3Action;
 use crate::common::gateway::authorize_operation;
 use async_trait::async_trait;
-use futures_util::stream;
 use rustfs_utils::MaskedAccessKey;
 use rustfs_utils::path;
 use s3s::dto::*;
 use std::fmt::Debug;
 use std::path::{Path, PathBuf};
+use std::sync::Arc;
 use tokio::io::AsyncRead;
 use tracing::{debug, error};
 use unftp_core::storage::{Error, ErrorKind, Fileinfo, Metadata, Result, StorageBackend};
@@ -94,12 +94,12 @@ impl Metadata for FtpsMetadata {
 /// FTPS storage driver implementation
 pub struct FtpsDriver<S> {
    /// Storage backend for S3 operations
-    storage: S,
+    storage: Arc<S>,
 }

 impl<S> Debug for FtpsDriver<S>
 where
-    S: S3StorageBackend + Debug,
+    S: S3StorageBackend + Debug + 'static,
 {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        f.debug_struct("FtpsDriver").field("storage", &"StorageBackend").finish()
@@ -108,11 +108,13 @@ where

 impl<S> FtpsDriver<S>
 where
-    S: S3StorageBackend + Debug,
+    S: S3StorageBackend + Debug + 'static,
 {
    /// Create a new FTPS driver with the given storage backend
    pub fn new(storage: S) -> Self {
-        Self { storage }
+        Self {
+            storage: Arc::new(storage),
+        }
    }

    /// List all buckets (for root path)
@@ -232,7 +234,7 @@ where
 #[async_trait]
 impl<S> StorageBackend<super::server::FtpsUser> for FtpsDriver<S>
 where
-    S: S3StorageBackend + Debug,
+    S: S3StorageBackend + Debug + 'static,
 {
    type Metadata = FtpsMetadata;

@@ -544,55 +546,18 @@ where
            .await
            .map_err(|_| Error::new(ErrorKind::PermanentFileNotAvailable, "Access denied"))?;

-        // Convert AsyncRead to bytes
-        let bytes_vec = {
-            let mut buffer = Vec::new();
-            let mut reader = bytes;
-            tokio::io::copy(&mut reader, &mut buffer)
-                .await
-                .map_err(|e| Error::new(ErrorKind::TransientFileNotAvailable, e.to_string()))?;
-            buffer
-        };
-
-        let file_size = bytes_vec.len();
-
-        let mut put_builder = PutObjectInput::builder();
-        put_builder.set_bucket(bucket.clone());
-        put_builder.set_key(key.clone());
-        put_builder.set_content_length(Some(file_size as i64));
-
-        // Create StreamingBlob with known size
-        let data_bytes = bytes::Bytes::from(bytes_vec);
-        let stream = stream::once(async move { Ok::<bytes::Bytes, std::io::Error>(data_bytes) });
-        let streaming_blob = s3s::dto::StreamingBlob::wrap(stream);
-        put_builder.set_body(Some(streaming_blob));
-        let put_input = put_builder
-            .build()
-            .map_err(|_| Error::new(ErrorKind::PermanentFileNotAvailable, "Failed to build PutObjectInput"))?;
-
-        match self.storage.put_object(put_input, session_context.credentials()).await {
-            Ok(_output) => {
-                Ok(file_size as u64) // Return the size of the uploaded object
-            }
-            Err(e) => {
+        upload::upload(Arc::clone(&self.storage), session_context, bytes, &bucket, &key)
+            .await
+            .inspect_err(|e| {
                error!(
                    event = EVENT_FTPS_OBJECT_PUT_FAILED,
                    component = LOG_COMPONENT_PROTOCOLS,
                    subsystem = LOG_SUBSYSTEM_FTPS_DRIVER,
                    username = %masked_username,
-                    path = %path_str,
-                    bucket = %bucket,
-                    object = %key,
-                    file_size,
-                    error = ?e,
+                    error = %e,
                    "ftps object put failed"
                );
-                Err(Error::new(
-                    ErrorKind::PermanentFileNotAvailable,
-                    format!("Failed to upload object: {:?}", e),
-                ))
-            }
-        }
+            })
    }

    async fn del<P: AsRef<Path> + Send>(&self, user: &super::server::FtpsUser, path: P) -> Result<()> {
@@ -927,3 +892,517 @@ mod tests {
        assert!(parse_s3_path("/bucket/__XLDIR__").is_err());
    }
 }
+
+mod upload {
+    use crate::common::{
+        client::s3::StorageBackend,
+        gateway::{S3Action, authorize_operation},
+        session::SessionContext,
+    };
+    use bytes::Bytes;
+    use s3s::dto::*;
+    use std::{
+        sync::{Arc, LazyLock},
+        time::Duration,
+    };
+    use tokio::{
+        io::{AsyncRead, AsyncReadExt},
+        sync::Semaphore,
+    };
+    use unftp_core::storage::{Error, ErrorKind, Result};
+
+    // One in-flight part per upload; never buffer the next part while the backend consumes this one.
+    const PART_SIZE: usize = 16 * 1024 * 1024;
+    const MAX_PARTS: i32 = 10_000;
+    const ABORT_TIMEOUT: Duration = Duration::from_secs(30);
+    static ABORT_PERMITS: LazyLock<Arc<Semaphore>> = LazyLock::new(|| Arc::new(Semaphore::new(32)));
+    const EVENT_FTPS_UPLOAD_CLEANUP: &str = "ftps_upload_cleanup";
+    const LOG_COMPONENT_PROTOCOLS: &str = "protocols";
+    const LOG_SUBSYSTEM_FTPS_UPLOAD: &str = "ftps_upload";
+
+    fn upload_error(error: impl std::fmt::Display) -> Error {
+        Error::new(ErrorKind::TransientFileNotAvailable, error.to_string())
+    }
+
+    async fn authorize(session: &SessionContext, action: S3Action, bucket: &str, key: &str) -> Result<()> {
+        authorize_operation(session, &action, bucket, Some(key))
+            .await
+            .map_err(|_| Error::new(ErrorKind::PermanentFileNotAvailable, "Access denied"))
+    }
+
+    fn next_part_number(completed: usize) -> Result<i32> {
+        let completed = i32::try_from(completed).map_err(upload_error)?;
+        if completed >= MAX_PARTS {
+            return Err(Error::new(
+                ErrorKind::ExceededStorageAllocationError,
+                "FTPS multipart upload exceeds 10000 parts",
+            ));
+        }
+        Ok(completed + 1)
+    }
+
+    async fn read_part(reader: &mut (impl AsyncRead + Unpin)) -> Result<Vec<u8>> {
+        let mut part = Vec::with_capacity(PART_SIZE);
+        reader
+            .take(u64::try_from(PART_SIZE).map_err(upload_error)?)
+            .read_to_end(&mut part)
+            .await
+            .map_err(upload_error)?;
+        Ok(part)
+    }
+
+    fn body(part: Vec<u8>) -> StreamingBlob {
+        StreamingBlob::from_bytes(Bytes::from(part))
+    }
+
+    // Keep the upload ID alive across all cancellable awaits after CreateMultipartUpload.
+    // A process/runtime crash still requires an AbortIncompleteMultipartUpload lifecycle rule.
+    struct PendingUpload<S: StorageBackend + 'static> {
+        storage: Arc<S>,
+        session: SessionContext,
+        input: Option<AbortMultipartUploadInput>,
+    }
+
+    impl<S: StorageBackend + 'static> PendingUpload<S> {
+        async fn abort(&mut self) {
+            if let Some(input) = self.input.as_ref() {
+                abort_upload(&*self.storage, &self.session, input.clone()).await;
+                self.input = None;
+            }
+        }
+    }
+
+    async fn abort_upload<S: StorageBackend>(storage: &S, session: &SessionContext, input: AbortMultipartUploadInput) {
+        let cleanup = async {
+            authorize(session, S3Action::AbortMultipartUpload, &input.bucket, &input.key).await?;
+            storage
+                .abort_multipart_upload(input, session.credentials())
+                .await
+                .map_err(upload_error)?;
+            Ok::<_, Error>(())
+        };
+        if !matches!(tokio::time::timeout(ABORT_TIMEOUT, cleanup).await, Ok(Ok(()))) {
+            tracing::warn!(
+                event = EVENT_FTPS_UPLOAD_CLEANUP,
+                component = LOG_COMPONENT_PROTOCOLS,
+                subsystem = LOG_SUBSYSTEM_FTPS_UPLOAD,
+                result = "abort_failed",
+                "FTPS incomplete upload requires lifecycle cleanup"
+            );
+        }
+    }
+
+    impl<S: StorageBackend + 'static> Drop for PendingUpload<S> {
+        fn drop(&mut self) {
+            let Some(input) = self.input.take() else {
+                return;
+            };
+            let (Ok(runtime), Ok(permit)) =
+                (tokio::runtime::Handle::try_current(), Arc::clone(&ABORT_PERMITS).try_acquire_owned())
+            else {
+                tracing::warn!(
+                    event = EVENT_FTPS_UPLOAD_CLEANUP,
+                    component = LOG_COMPONENT_PROTOCOLS,
+                    subsystem = LOG_SUBSYSTEM_FTPS_UPLOAD,
+                    result = "abort_unavailable",
+                    "FTPS incomplete upload requires lifecycle cleanup"
+                );
+                return;
+            };
+            let storage = Arc::clone(&self.storage);
+            let session = self.session.clone();
+            runtime.spawn(async move {
+                let _permit = permit;
+                abort_upload(&*storage, &session, input).await;
+            });
+        }
+    }
+
+    pub(super) async fn upload<S: StorageBackend + 'static>(
+        storage: Arc<S>,
+        session: &SessionContext,
+        mut reader: impl AsyncRead + Unpin,
+        bucket: &str,
+        key: &str,
+    ) -> Result<u64> {
+        let mut part = read_part(&mut reader).await?;
+        if part.len() < PART_SIZE {
+            let size = u64::try_from(part.len()).map_err(upload_error)?;
+            authorize(session, S3Action::PutObject, bucket, key).await?;
+            let input = PutObjectInput::builder()
+                .bucket(bucket.to_owned())
+                .key(key.to_owned())
+                .content_length(Some(i64::try_from(size).map_err(upload_error)?))
+                .body(Some(body(part)))
+                .build()
+                .map_err(upload_error)?;
+            storage.put_object(input, session.credentials()).await.map_err(upload_error)?;
+            return Ok(size);
+        }
+
+        authorize(session, S3Action::CreateMultipartUpload, bucket, key).await?;
+        let input = CreateMultipartUploadInput::builder()
+            .bucket(bucket.to_owned())
+            .key(key.to_owned())
+            .build()
+            .map_err(upload_error)?;
+        let output = storage
+            .create_multipart_upload(input, session.credentials())
+            .await
+            .map_err(upload_error)?;
+        let upload_id = output
+            .upload_id
+            .filter(|id| !id.is_empty())
+            .ok_or_else(|| upload_error("CreateMultipartUpload returned no upload ID"))?;
+        let abort = AbortMultipartUploadInput::builder()
+            .bucket(bucket.to_owned())
+            .key(key.to_owned())
+            .upload_id(upload_id.clone())
+            .build()
+            .map_err(upload_error)?;
+        let mut pending = PendingUpload {
+            storage: Arc::clone(&storage),
+            session: session.clone(),
+            input: Some(abort),
+        };
+
+        let result = async {
+            let mut parts = Vec::new();
+            let mut total = 0u64;
+            while !part.is_empty() {
+                let number = next_part_number(parts.len())?;
+                let length = i64::try_from(part.len()).map_err(upload_error)?;
+                total = total
+                    .checked_add(u64::try_from(part.len()).map_err(upload_error)?)
+                    .ok_or_else(|| upload_error("FTPS upload size overflow"))?;
+                authorize(session, S3Action::UploadPart, bucket, key).await?;
+                let input = UploadPartInput::builder()
+                    .bucket(bucket.to_owned())
+                    .key(key.to_owned())
+                    .upload_id(upload_id.clone())
+                    .part_number(number)
+                    .content_length(Some(length))
+                    .body(Some(body(part)))
+                    .build()
+                    .map_err(upload_error)?;
+                let output = storage
+                    .upload_part(input, session.credentials())
+                    .await
+                    .map_err(upload_error)?;
+                let etag = output.e_tag.ok_or_else(|| upload_error("UploadPart returned no ETag"))?;
+                parts.push(CompletedPart {
+                    e_tag: Some(etag),
+                    part_number: Some(number),
+                    ..Default::default()
+                });
+                part = read_part(&mut reader).await?;
+            }
+            authorize(session, S3Action::CompleteMultipartUpload, bucket, key).await?;
+            let input = CompleteMultipartUploadInput::builder()
+                .bucket(bucket.to_owned())
+                .key(key.to_owned())
+                .upload_id(upload_id)
+                .multipart_upload(Some(CompletedMultipartUpload { parts: Some(parts) }))
+                .build()
+                .map_err(upload_error)?;
+            storage
+                .complete_multipart_upload(input, session.credentials())
+                .await
+                .map_err(upload_error)?;
+            // No await between a successful commit and disarming cancellation cleanup.
+            pending.input = None;
+            Ok(total)
+        }
+        .await;
+        if result.is_err() {
+            pending.abort().await;
+        }
+        result
+    }
+
+    #[cfg(test)]
+    mod tests {
+        use super::*;
+        use crate::{
+            common::{
+                dummy_storage::{DummyBackend, DummyError},
+                gateway::{is_operation_supported, with_test_auth_override},
+                session::{Protocol, test_session},
+            },
+            ftps::{driver::FtpsDriver, server::FtpsUser},
+        };
+        use std::{
+            pin::Pin,
+            task::{Context, Poll},
+        };
+        use tokio::{io::ReadBuf, sync::Notify};
+        use unftp_core::storage::StorageBackend as _;
+
+        fn user() -> FtpsUser {
+            FtpsUser {
+                username: "upload-test".into(),
+                name: None,
+                session_context: test_session(Protocol::Ftps),
+            }
+        }
+
+        fn multipart_backend() -> DummyBackend {
+            let backend = DummyBackend::new();
+            backend.queue_create_multipart_upload_ok("upload-1");
+            backend.queue_complete_multipart_upload_ok();
+            backend
+        }
+
+        // Refuse to produce part N+1 before part N has reached the backend. The old
+        // read-to-EOF implementation fails this without allocating a huge fixture.
+        struct PacedReader {
+            backend: DummyBackend,
+            position: usize,
+            length: usize,
+            fail_at_end: bool,
+        }
+        impl AsyncRead for PacedReader {
+            fn poll_read(mut self: Pin<&mut Self>, _: &mut Context<'_>, buf: &mut ReadBuf<'_>) -> Poll<std::io::Result<()>> {
+                if self.position / PART_SIZE > self.backend.upload_part_calls().len() {
+                    return Poll::Ready(Err(std::io::Error::other("read ahead of uploaded part")));
+                }
+                if self.position == self.length && self.fail_at_end {
+                    return Poll::Ready(Err(std::io::Error::other("interrupted source")));
+                }
+                let count = buf.remaining().min(self.length - self.position).min(8191);
+                let value = u8::try_from(self.position / PART_SIZE).expect("small test part number");
+                // Do not cross a part boundary in a single read, so each part has a distinct byte.
+                let count = count.min(PART_SIZE - self.position % PART_SIZE);
+                buf.put_slice(&vec![value; count]);
+                self.position += count;
+                Poll::Ready(Ok(()))
+            }
+        }
+
+        #[test]
+        fn ftps_part_numbers_stop_at_s3_limit() {
+            assert_eq!(next_part_number(0).expect("first part"), 1);
+            assert_eq!(next_part_number(9_999).expect("last part"), 10_000);
+            assert!(next_part_number(10_000).is_err());
+            assert!(next_part_number(usize::MAX).is_err());
+        }
+
+        #[tokio::test]
+        async fn ftps_part_buffer_does_not_grow_to_detect_eof() {
+            let mut reader = tokio::io::repeat(0).take(u64::try_from(PART_SIZE).expect("size"));
+            let part = read_part(&mut reader).await.expect("read full part");
+            assert_eq!(part.len(), PART_SIZE);
+            assert_eq!(part.capacity(), PART_SIZE);
+            assert!(read_part(&mut reader).await.expect("EOF").is_empty());
+        }
+
+        #[tokio::test]
+        async fn ftps_small_and_empty_uploads_preserve_bytes() {
+            for payload in [Vec::new(), b"small upload\x00\xff".to_vec()] {
+                let backend = DummyBackend::new();
+                backend.capture_upload_bodies();
+                let driver = FtpsDriver::new(backend.clone());
+                let result = with_test_auth_override(
+                    |_, _, _| true,
+                    driver.put(&user(), std::io::Cursor::new(payload.clone()), "/bucket/key", 0),
+                )
+                .await
+                .expect("STOR");
+                assert_eq!(result, u64::try_from(payload.len()).expect("size"));
+                assert_eq!(backend.upload_bodies(), vec![payload]);
+                assert!(backend.create_multipart_calls().is_empty());
+            }
+        }
+
+        #[tokio::test]
+        async fn ftps_upload_streams_before_eof_and_preserves_part_bytes() {
+            for length in [PART_SIZE - 1, PART_SIZE, PART_SIZE + 1, 2 * PART_SIZE, 2 * PART_SIZE + 7] {
+                let backend = multipart_backend();
+                backend.capture_upload_bodies();
+                for n in 1..=3 {
+                    backend.queue_upload_part_ok(format!("etag-{n}"));
+                }
+                let driver = FtpsDriver::new(backend.clone());
+                let reader = PacedReader {
+                    backend: backend.clone(),
+                    position: 0,
+                    length,
+                    fail_at_end: false,
+                };
+                let size = with_test_auth_override(|_, _, _| true, driver.put(&user(), reader, "/bucket/key", 0))
+                    .await
+                    .expect("bounded STOR");
+                assert_eq!(size, u64::try_from(length).expect("size"));
+                let bodies = backend.upload_bodies();
+                assert_eq!(bodies.iter().map(Vec::len).sum::<usize>(), length);
+                for (index, part) in bodies.iter().enumerate() {
+                    assert!(part.len() <= PART_SIZE);
+                    assert!(part.iter().all(|b| *b == u8::try_from(index).expect("index")));
+                }
+                if length >= PART_SIZE {
+                    assert_eq!(backend.complete_multipart_calls()[0].part_count, length.div_ceil(PART_SIZE));
+                    for (index, call) in backend.upload_part_calls().iter().enumerate() {
+                        assert_eq!(call.part_number, i32::try_from(index + 1).expect("part"));
+                        assert_eq!(call.content_length, Some(i64::try_from(bodies[index].len()).expect("length")));
+                    }
+                }
+                assert!(backend.abort_multipart_calls().is_empty());
+            }
+        }
+
+        #[tokio::test]
+        async fn ftps_upload_aborts_on_source_part_or_complete_failure() {
+            for failure in ["source", "part", "etag", "complete"] {
+                let backend = DummyBackend::new();
+                backend.queue_create_multipart_upload_ok("upload-1");
+                match failure {
+                    "part" => backend.queue_upload_part_err(DummyError::Injected("part failed".into())),
+                    "etag" => backend.queue_upload_part_ok_without_etag(),
+                    _ => backend.queue_upload_part_ok("etag-1"),
+                }
+                backend.queue_complete_multipart_upload_err(DummyError::Injected("complete failed".into()));
+                let driver = FtpsDriver::new(backend.clone());
+                let reader = PacedReader {
+                    backend: backend.clone(),
+                    position: 0,
+                    length: PART_SIZE,
+                    fail_at_end: failure == "source",
+                };
+                let result = with_test_auth_override(|_, _, _| true, driver.put(&user(), reader, "/bucket/key", 0)).await;
+                assert!(result.is_err(), "{failure} must not report a committed object");
+                assert_eq!(backend.abort_multipart_calls().len(), 1, "{failure}");
+                assert_eq!(backend.abort_multipart_calls()[0].upload_id, "upload-1");
+                assert_eq!(backend.complete_multipart_calls().len(), usize::from(failure == "complete"));
+            }
+        }
+
+        #[tokio::test]
+        async fn ftps_upload_enforces_authorization_at_each_mutation() {
+            for denied in [
+                S3Action::PutObject,
+                S3Action::CreateMultipartUpload,
+                S3Action::UploadPart,
+                S3Action::CompleteMultipartUpload,
+            ] {
+                assert!(is_operation_supported(Protocol::Ftps, &denied));
+                let backend = multipart_backend();
+                backend.queue_upload_part_ok("etag-1");
+                let driver = FtpsDriver::new(backend.clone());
+                let reader = tokio::io::repeat(0).take(u64::try_from(PART_SIZE).expect("size"));
+                let denied_action = denied.clone();
+                let result = with_test_auth_override(
+                    move |action, _, _| action != &denied_action,
+                    driver.put(&user(), reader, "/bucket/key", 0),
+                )
+                .await;
+                assert!(result.is_err());
+                match denied {
+                    S3Action::PutObject | S3Action::CreateMultipartUpload => assert!(backend.create_multipart_calls().is_empty()),
+                    S3Action::UploadPart => assert!(backend.upload_part_calls().is_empty()),
+                    _ => assert!(backend.complete_multipart_calls().is_empty()),
+                }
+            }
+        }
+
+        #[tokio::test]
+        async fn ftps_upload_cleanup_does_not_bypass_abort_permission() {
+            let backend = multipart_backend();
+            backend.queue_upload_part_err(DummyError::Injected("part failed".into()));
+            let driver = FtpsDriver::new(backend.clone());
+            let reader = tokio::io::repeat(0).take(u64::try_from(PART_SIZE).expect("size"));
+            let result = with_test_auth_override(
+                |action, _, _| action != &S3Action::AbortMultipartUpload,
+                driver.put(&user(), reader, "/bucket/key", 0),
+            )
+            .await;
+            assert!(result.is_err());
+            assert!(backend.abort_multipart_calls().is_empty());
+        }
+
+        #[tokio::test]
+        async fn ftps_cancelled_upload_aborts_pending_parts() {
+            let backend = multipart_backend();
+            let entered = Arc::new(Notify::new());
+            backend.stall_upload_part(Arc::clone(&entered));
+            let driver = FtpsDriver::new(backend.clone());
+            let user = user();
+            with_test_auth_override(|_, _, _| true, async {
+                let reader = tokio::io::repeat(0).take(u64::try_from(PART_SIZE * 2).expect("size"));
+                let mut put = Box::pin(driver.put(&user, reader, "/bucket/key", 0));
+                tokio::select! {
+                    _ = entered.notified() => {},
+                    result = &mut put => panic!("upload should stall: {result:?}"),
+                }
+                drop(put);
+                tokio::time::timeout(Duration::from_secs(5), async {
+                    while backend.abort_multipart_calls().is_empty() {
+                        tokio::task::yield_now().await;
+                    }
+                })
+                .await
+                .expect("cancellation cleanup");
+            })
+            .await;
+            assert_eq!(backend.abort_multipart_calls().len(), 1);
+            assert!(backend.complete_multipart_calls().is_empty());
+        }
+        // Yield on each read so the memory probe keeps all upload buffers live at once.
+        struct YieldingReader {
+            remaining: u64,
+            yielded: bool,
+        }
+
+        impl AsyncRead for YieldingReader {
+            fn poll_read(mut self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &mut ReadBuf<'_>) -> Poll<std::io::Result<()>> {
+                if self.remaining == 0 {
+                    return Poll::Ready(Ok(()));
+                }
+                if !self.yielded {
+                    self.yielded = true;
+                    cx.waker().wake_by_ref();
+                    return Poll::Pending;
+                }
+                self.yielded = false;
+                let count = buf
+                    .remaining()
+                    .min(1024 * 1024)
+                    .min(usize::try_from(self.remaining).unwrap_or(usize::MAX));
+                buf.initialize_unfilled_to(count).fill(0x5a);
+                buf.advance(count);
+                self.remaining -= u64::try_from(count).expect("read size");
+                Poll::Ready(Ok(()))
+            }
+        }
+
+        #[tokio::test]
+        #[ignore = "manual synthetic concurrent-upload memory measurement"]
+        async fn ftps_large_upload_memory_probe() {
+            let bytes: u64 = std::env::var("RUSTFS_FTPS_TEST_BYTES")
+                .unwrap_or_else(|_| "67108864".into())
+                .parse()
+                .expect("byte count");
+            let concurrency: usize = std::env::var("RUSTFS_FTPS_TEST_CONCURRENCY")
+                .unwrap_or_else(|_| "10".into())
+                .parse()
+                .expect("concurrency");
+            assert!(bytes > 0 && concurrency > 0);
+            let operations = (0..concurrency).map(|_| async {
+                let backend = multipart_backend();
+                for n in 0..bytes.div_ceil(u64::try_from(PART_SIZE).expect("part size")) {
+                    backend.queue_upload_part_ok(format!("part-{n}"));
+                }
+                let driver = FtpsDriver::new(backend);
+                let reader = YieldingReader {
+                    remaining: bytes,
+                    yielded: false,
+                };
+                driver.put(&user(), reader, "/bucket/key", 0).await
+            });
+            let results = with_test_auth_override(|_, _, _| true, futures_util::future::join_all(operations)).await;
+            for result in results {
+                assert_eq!(result.expect("upload"), bytes);
+            }
+            println!("synthetic_upload_bytes={bytes} concurrent_uploads={concurrency}");
+        }
+    }
+}
@@ -57,6 +57,7 @@ const EXCESSIVE_EMPTY_CHUNKS_ERROR: &str = "HTTP body returned too many empty ch
 pub const INTERNODE_DISK_ERROR_HEADER: &str = "x-rustfs-disk-error";
 pub const INTERNODE_FILE_NOT_FOUND: &str = "file-not-found";
 pub const INTERNODE_VOLUME_NOT_FOUND: &str = "volume-not-found";
+pub const INTERNODE_FILE_CORRUPT: &str = "file-corrupt";

 #[derive(Debug, Clone, Copy, Eq, PartialEq)]
 pub enum InternodeHttpErrorKind {
@@ -185,6 +186,7 @@ pub struct InternodeHttpError {
 enum RemoteDiskErrorKind {
    FileNotFound,
    VolumeNotFound,
+    FileCorrupt,
 }

 impl std::fmt::Debug for InternodeHttpError {
@@ -215,6 +217,10 @@ impl InternodeHttpError {
        self.remote_disk_error == Some(RemoteDiskErrorKind::VolumeNotFound)
    }

+    pub fn is_remote_file_corrupt(&self) -> bool {
+        self.remote_disk_error == Some(RemoteDiskErrorKind::FileCorrupt)
+    }
+
    fn new(kind: InternodeHttpErrorKind, context: InternodeHttpRequestContext) -> Self {
        Self {
            kind,
@@ -328,6 +334,20 @@ pub fn new_test_remote_volume_not_found_http_io_error() -> io::Error {
    .into_io_error()
 }

+#[doc(hidden)]
+pub fn new_test_remote_file_corrupt_http_io_error() -> io::Error {
+    InternodeHttpError::with_remote_disk_error(
+        InternodeHttpErrorKind::HttpStatus(reqwest::StatusCode::INTERNAL_SERVER_ERROR),
+        InternodeHttpRequestContext {
+            method: "GET".to_string(),
+            target: READ_FILE_STREAM_PATH.to_string(),
+            operation: Some(INTERNODE_OPERATION_READ_FILE_STREAM),
+        },
+        RemoteDiskErrorKind::FileCorrupt,
+    )
+    .into_io_error()
+}
+
 fn add_root_certificates_from_der(builder: reqwest::ClientBuilder, certs_der: &[Vec<u8>]) -> reqwest::ClientBuilder {
    let mut b = builder;
    for der in certs_der {
@@ -850,6 +870,7 @@ fn classify_http_response(
    let remote_disk_error = match headers.get(INTERNODE_DISK_ERROR_HEADER).and_then(|value| value.to_str().ok()) {
        Some(INTERNODE_FILE_NOT_FOUND) => Some(RemoteDiskErrorKind::FileNotFound),
        Some(INTERNODE_VOLUME_NOT_FOUND) => Some(RemoteDiskErrorKind::VolumeNotFound),
+        Some(INTERNODE_FILE_CORRUPT) => Some(RemoteDiskErrorKind::FileCorrupt),
        _ => None,
    };
    ClassifiedHttpResponse { kind, remote_disk_error }
@@ -2047,6 +2068,7 @@ mod tests {
        assert_eq!(INTERNODE_DISK_ERROR_HEADER, "x-rustfs-disk-error");
        assert_eq!(INTERNODE_FILE_NOT_FOUND, "file-not-found");
        assert_eq!(INTERNODE_VOLUME_NOT_FOUND, "volume-not-found");
+        assert_eq!(INTERNODE_FILE_CORRUPT, "file-corrupt");
    }

    #[test]
@@ -2090,6 +2112,48 @@ mod tests {
        assert!(wrong_operation.remote_disk_error.is_none());
    }

+    #[test]
+    fn classify_http_response_scopes_corrupt_disk_error_to_read_failures() {
+        let mut headers = HeaderMap::new();
+        headers.insert(INTERNODE_DISK_ERROR_HEADER, INTERNODE_FILE_CORRUPT.parse().expect("valid error token"));
+        let classified = classify_http_response(
+            reqwest::StatusCode::INTERNAL_SERVER_ERROR,
+            &headers,
+            Some(INTERNODE_OPERATION_READ_FILE_STREAM),
+        );
+        assert_eq!(classified.remote_disk_error, Some(RemoteDiskErrorKind::FileCorrupt));
+        for status in [reqwest::StatusCode::NOT_FOUND, reqwest::StatusCode::SERVICE_UNAVAILABLE] {
+            assert!(
+                classify_http_response(status, &headers, Some(INTERNODE_OPERATION_READ_FILE_STREAM))
+                    .remote_disk_error
+                    .is_none()
+            );
+        }
+        for operation in [
+            None,
+            Some(INTERNODE_OPERATION_WALK_DIR),
+            Some(INTERNODE_OPERATION_PUT_FILE_STREAM),
+        ] {
+            assert!(
+                classify_http_response(reqwest::StatusCode::INTERNAL_SERVER_ERROR, &headers, operation)
+                    .remote_disk_error
+                    .is_none()
+            );
+        }
+        headers.insert(INTERNODE_DISK_ERROR_HEADER, "unknown-disk-error".parse().expect("valid unknown token"));
+        for headers in [&headers, &HeaderMap::new()] {
+            assert!(
+                classify_http_response(
+                    reqwest::StatusCode::INTERNAL_SERVER_ERROR,
+                    headers,
+                    Some(INTERNODE_OPERATION_READ_FILE_STREAM),
+                )
+                .remote_disk_error
+                .is_none()
+            );
+        }
+    }
+
    #[derive(Clone, Default)]
    struct TestState {
        head_count: Arc<AtomicUsize>,
@@ -2158,6 +2222,16 @@ mod tests {
        let addr = listener.local_addr().expect("listener local address should be available");
        let app = Router::new()
            .route("/stream", get(get_stream).head(reject_head).put(accept_put))
+            .route(
+                READ_FILE_STREAM_PATH,
+                get(|| async {
+                    (
+                        StatusCode::INTERNAL_SERVER_ERROR,
+                        [(INTERNODE_DISK_ERROR_HEADER, INTERNODE_FILE_CORRUPT)],
+                        "read file err file corrupt",
+                    )
+                }),
+            )
            .route(WALK_DIR_PATH, get(get_stream))
            .route("/reject-put", get(get_stream).put(reject_put))
            .route("/stall", get(get_stalling_stream))
@@ -2172,6 +2246,44 @@ mod tests {
        Some((format!("http://{addr}/stream"), handle))
    }

+    #[tokio::test]
+    async fn http_readers_preserve_remote_file_corruption() {
+        let (url, server) = start_test_server(TestState::default())
+            .await
+            .expect("corruption regression server must bind");
+        let url = format!("{}{READ_FILE_STREAM_PATH}", url.strip_suffix("/stream").expect("test server URL suffix"));
+        let byte_error = HttpReader::new(url.clone(), Method::GET, HeaderMap::new(), None)
+            .await
+            .err()
+            .expect("byte reader must reject corrupt shard response");
+        let chunk_error = HttpChunkReader::new_with_stall_timeout(url, Method::GET, HeaderMap::new(), None, None)
+            .await
+            .err()
+            .expect("chunk reader must reject corrupt shard response");
+        for error in [byte_error, chunk_error] {
+            let source = error
+                .get_ref()
+                .and_then(|source| source.downcast_ref::<InternodeHttpError>())
+                .expect("HTTP error must retain typed internode source");
+            assert!(source.is_remote_file_corrupt());
+            assert!(!source.is_remote_file_not_found());
+            assert!(!source.is_remote_volume_not_found());
+            assert_eq!(
+                source.kind(),
+                InternodeHttpErrorKind::HttpStatus(reqwest::StatusCode::INTERNAL_SERVER_ERROR)
+            );
+            let cloned = clone_internode_http_io_error(&error).expect("typed transport error must be cloneable");
+            assert!(
+                cloned
+                    .get_ref()
+                    .and_then(|source| source.downcast_ref::<InternodeHttpError>())
+                    .expect("clone must retain typed internode source")
+                    .is_remote_file_corrupt()
+            );
+        }
+        server.abort();
+    }
+
    struct BlockedH2Server {
        url: String,
        shutdown: tokio::sync::oneshot::Sender<()>,
@@ -83,13 +83,11 @@ impl ProviderVersionCapabilities {
        if tier_type.eq_ignore_ascii_case("s3")
            || tier_type.eq_ignore_ascii_case("rustfs")
            || tier_type.eq_ignore_ascii_case("minio")
-            || tier_type.eq_ignore_ascii_case("r2")
            || tier_type.eq_ignore_ascii_case("wasabi")
        {
            let list_object_versions = tier_type.eq_ignore_ascii_case("s3")
                || tier_type.eq_ignore_ascii_case("rustfs")
-                || tier_type.eq_ignore_ascii_case("minio")
-                || tier_type.eq_ignore_ascii_case("r2");
+                || tier_type.eq_ignore_ascii_case("minio");
            Self {
                raw_version_header: Some(X_AMZ_VERSION_ID),
                bucket_versioning_state: list_object_versions,
@@ -101,6 +99,17 @@ impl ProviderVersionCapabilities {
                },
                exact_get_delete: true,
            }
+        } else if tier_type.eq_ignore_ascii_case("r2") {
+            // R2 exposes an x-amz-version-id on some PUT responses, but it does
+            // not provide routable S3 object versions. GET/DELETE must therefore
+            // use the unversioned object path and must not call version APIs.
+            Self {
+                raw_version_header: None,
+                bucket_versioning_state: false,
+                list_object_versions: false,
+                conditional_create: ConditionalCreateCapability::IfNoneMatchStar,
+                exact_get_delete: false,
+            }
        } else if tier_type.eq_ignore_ascii_case("aliyun") {
            Self {
                raw_version_header: Some(X_OSS_VERSION_ID),
@@ -210,8 +219,6 @@ mod tests {
            ("RustFS", "x-amz-version-id"),
            ("minio", "x-amz-version-id"),
            ("MinIO", "x-amz-version-id"),
-            ("r2", "x-amz-version-id"),
-            ("R2", "x-amz-version-id"),
            ("wasabi", "x-amz-version-id"),
            ("Wasabi", "x-amz-version-id"),
            ("aliyun", "x-oss-version-id"),
@@ -255,6 +262,26 @@ mod tests {
        );
    }

+    #[test]
+    fn r2_ignores_non_routable_put_version_header() {
+        let mut headers = HeaderMap::new();
+        headers.insert("x-amz-version-id", HeaderValue::from_static("r2-response-token"));
+        let capabilities = ProviderVersionCapabilities::for_tier_type("r2");
+
+        assert_eq!(
+            capabilities
+                .raw_version_id(&headers)
+                .expect("R2 header parsing should succeed"),
+            None
+        );
+        assert_eq!(
+            capabilities
+                .remote_version(&headers, BucketVersioningState::Disabled)
+                .expect("R2 uses unversioned routing"),
+            RemoteVersion::Disabled
+        );
+    }
+
    #[test]
    fn provider_version_missing_header_is_unknown_until_bucket_state_is_known() {
        let headers = HeaderMap::new();
@@ -280,7 +307,7 @@ mod tests {
            ("s3", true, true, ConditionalCreateCapability::IfNoneMatchStar, true),
            ("rustfs", true, true, ConditionalCreateCapability::Unsupported, true),
            ("minio", true, true, ConditionalCreateCapability::Unsupported, true),
-            ("r2", true, true, ConditionalCreateCapability::IfNoneMatchStar, true),
+            ("r2", false, false, ConditionalCreateCapability::IfNoneMatchStar, false),
            ("wasabi", false, false, ConditionalCreateCapability::Unsupported, true),
            ("aliyun", false, false, ConditionalCreateCapability::Unsupported, true),
            ("tencent", false, false, ConditionalCreateCapability::Unsupported, true),
@@ -16,8 +16,8 @@
 use crate::RUSTFS_META_BUCKET;
 use crate::scanner_budget::{ScannerCycleBudget, ScannerCycleBudgetConfig};
 use crate::scanner_io::{
-    DataUsageCacheReuseOptions, DataUsageCacheScanState, ScannerCheckpointPersistResult, ScannerDiskScanOptions,
-    ScannerDiskScanOutcome, ScannerIODisk, acquire_scanner_cache_locks, cache_root_entry_info,
+    DataUsageCacheReuseOptions, DataUsageCacheScanState, ScannerCheckpointPersistContext, ScannerCheckpointPersistResult,
+    ScannerDiskScanOptions, ScannerDiskScanOutcome, ScannerIODisk, acquire_scanner_cache_locks, cache_root_entry_info,
    current_cache_root_or_prepare_with_generation, persist_scanner_checkpoint, scanner_set_disk_inventory,
 };
 use crate::storage_api::owner::NS_SCANNER_PROTOCOL_VERSION;
@@ -840,12 +840,15 @@ async fn scan_and_persist_local_bucket(
                }
                match persist_scanner_checkpoint(
                    set.clone(),
+                    ScannerCheckpointPersistContext {
+                        ctx: &scan_ctx,
+                        expected_publication_epoch,
+                        cycle: next_cycle,
+                        leader_epoch,
+                    },
                    &cache_name,
                    &checkpoint,
                    &mut revisions,
-                    expected_publication_epoch,
-                    next_cycle,
-                    leader_epoch,
                )
                .await
                {
@@ -1778,9 +1778,7 @@ where
            mark_scan_cycle_idle(cycle_info, &mut cycle_metrics_guard).await;
            return ScannerCycleOutcome::Failed;
        }
-        BackgroundHealInfoReadStatus::ErasureSd
-        | BackgroundHealInfoReadStatus::Loaded
-        | BackgroundHealInfoReadStatus::Missing => {}
+        BackgroundHealInfoReadStatus::Loaded | BackgroundHealInfoReadStatus::Missing => {}
    }
    let mut background_heal_info = background_heal_read.info;
    let background_heal_epoch = background_heal_read.expected_epoch;
@@ -28,7 +28,6 @@ pub struct BackgroundHealInfo {

 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 pub(super) enum BackgroundHealInfoReadStatus {
-    ErasureSd,
    Loaded,
    Missing,
    Blocked,
@@ -66,15 +65,6 @@ pub(super) async fn read_background_heal_info_with_epoch<S>(storeapi: Arc<S>) ->
 where
    S: ScannerStorage,
 {
-    // Skip for ErasureSD setup
-    if storeapi.setup_is_erasure_sd().await {
-        return BackgroundHealInfoRead {
-            info: BackgroundHealInfo::default(),
-            expected_epoch: None,
-            status: BackgroundHealInfoReadStatus::ErasureSd,
-        };
-    }
-
    let expected_epoch = scanner_publication_epoch(storeapi.clone()).await;
    if expected_epoch.is_none() {
        return BackgroundHealInfoRead {
@@ -146,11 +136,6 @@ pub(super) async fn save_background_heal_info_for_epoch<S>(
 ) where
    S: ScannerStorage,
 {
-    // Skip for ErasureSD setup
-    if storeapi.setup_is_erasure_sd().await {
-        return;
-    }
-
    // Serialize to JSON
    let data = match serde_json::to_vec(&info) {
        Ok(data) => data,
@@ -1806,6 +1806,14 @@ impl FolderScanner {
                    continue;
                }

+                // Do not start another metadata read while foreground work is
+                // active. The post-object timer protects the next request only
+                // after the read has already been dispatched; this admission
+                // point keeps the scanner from extending a single-disk I/O
+                // burst across foreground requests.
+                if crate::workload_admission::foreground_workload_activity() > 0 {
+                    self.sleeper.sleep_folder().await;
+                }
                let timer = self.sleeper.timer();

                let heal_enabled = this_hash.mod_alt(
@@ -1609,8 +1609,8 @@ use dirty_usage::*;
 use guards::*;

 pub(crate) use cache::{
-    DataUsageCacheReuseOptions, DataUsageCacheScanState, ScannerCheckpointPersistResult, acquire_scanner_cache_locks,
-    current_cache_root_or_prepare_with_generation, persist_scanner_checkpoint,
+    DataUsageCacheReuseOptions, DataUsageCacheScanState, ScannerCheckpointPersistContext, ScannerCheckpointPersistResult,
+    acquire_scanner_cache_locks, current_cache_root_or_prepare_with_generation, persist_scanner_checkpoint,
 };
 pub use dirty_usage::{
    ScannerDirtyUsageAckError, ScannerDirtyUsageBucket, ScannerDirtyUsageClearObserver, ScannerDirtyUsageMutationObserver,
@@ -31,6 +31,41 @@ pub(crate) enum ScannerCheckpointPersistResult {
    Failed(StorageError),
 }

+pub(crate) struct ScannerCheckpointPersistContext<'a> {
+    pub(crate) ctx: &'a CancellationToken,
+    pub(crate) expected_publication_epoch: u64,
+    pub(crate) cycle: u64,
+    pub(crate) leader_epoch: u64,
+}
+
+const CHECKPOINT_FOREGROUND_QUIET_WAIT: Duration = Duration::from_secs(1);
+
+async fn wait_for_checkpoint_foreground_quiet(ctx: &CancellationToken) -> bool {
+    let deadline = tokio::time::Instant::now() + CHECKPOINT_FOREGROUND_QUIET_WAIT;
+    loop {
+        if crate::workload_admission::foreground_workload_activity() == 0 {
+            return true;
+        }
+
+        let remaining = deadline.saturating_duration_since(tokio::time::Instant::now());
+        if remaining.is_zero() {
+            return false;
+        }
+
+        let backoff = Duration::from_millis(
+            crate::workload_admission::foreground_workload_activity()
+                .saturating_mul(10)
+                .min(250),
+        )
+        .max(Duration::from_millis(10))
+        .min(remaining);
+        tokio::select! {
+            _ = ctx.cancelled() => return false,
+            _ = tokio::time::sleep(backoff) => {}
+        }
+    }
+}
+
 /// Persist one bounded checkpoint and refresh its CAS revisions.
 ///
 /// Local and remote workers share the same publication/leader fencing and
@@ -38,23 +73,37 @@ pub(crate) enum ScannerCheckpointPersistResult {
 /// at the caller because those guards have different concrete types.
 pub(crate) async fn persist_scanner_checkpoint<S>(
    store: Arc<S>,
+    context: ScannerCheckpointPersistContext<'_>,
    cache_name: &str,
    checkpoint: &DataUsageCache,
    revisions: &mut DataUsageCacheRevisions,
-    expected_publication_epoch: u64,
-    cycle: u64,
-    leader_epoch: u64,
 ) -> ScannerCheckpointPersistResult
 where
    S: ScannerObjectIO + ScannerConfigObjectDelete,
 {
-    if crate::remote_scanner::validate_remote_scanner_request_fence_with_store(cycle, leader_epoch, store.clone())
+    let foreground_quiet = wait_for_checkpoint_foreground_quiet(context.ctx).await;
+    if !foreground_quiet && context.ctx.is_cancelled() {
+        return ScannerCheckpointPersistResult::FenceChanged;
+    }
+    if !foreground_quiet {
+        debug!(
+            target: "rustfs::scanner::io",
+            event = EVENT_SCANNER_CACHE_PERSIST_STATE,
+            component = LOG_COMPONENT_SCANNER,
+            subsystem = LOG_SUBSYSTEM_IO,
+            cache_name,
+            state = "checkpoint_foreground_wait_expired",
+            "Scanner checkpoint foreground quiet wait expired; preserving bounded progress"
+        );
+    }
+
+    if crate::remote_scanner::validate_remote_scanner_request_fence_with_store(context.cycle, context.leader_epoch, store.clone())
        .await
        .is_err()
    {
        return ScannerCheckpointPersistResult::FenceChanged;
    }
-    if scanner_publication_admission_for_epoch(store.clone(), expected_publication_epoch)
+    if scanner_publication_admission_for_epoch(store.clone(), context.expected_publication_epoch)
        .await
        .is_none()
    {
@@ -62,16 +111,16 @@ where
    }

    if let Err(error) = checkpoint
-        .save_with_revisions_for_epoch(store.clone(), cache_name, revisions, expected_publication_epoch)
+        .save_with_revisions_for_epoch(store.clone(), cache_name, revisions, context.expected_publication_epoch)
        .await
    {
        return ScannerCheckpointPersistResult::Failed(error);
    }

-    if crate::remote_scanner::validate_remote_scanner_request_fence_with_store(cycle, leader_epoch, store.clone())
+    if crate::remote_scanner::validate_remote_scanner_request_fence_with_store(context.cycle, context.leader_epoch, store.clone())
        .await
        .is_err()
-        || scanner_publication_admission_for_epoch(store.clone(), expected_publication_epoch)
+        || scanner_publication_admission_for_epoch(store.clone(), context.expected_publication_epoch)
            .await
            .is_none()
    {
@@ -1191,12 +1191,15 @@ impl ScannerIOCache for SetDisks {
                                }
                                match persist_scanner_checkpoint(
                                    store_clone_clone.clone(),
+                                    ScannerCheckpointPersistContext {
+                                        ctx: &ctx_clone,
+                                        expected_publication_epoch: expected_publication_epoch_clone,
+                                        cycle: want_cycle,
+                                        leader_epoch,
+                                    },
                                    cache_name.as_str(),
                                    &checkpoint,
                                    &mut revisions,
-                                    expected_publication_epoch_clone,
-                                    want_cycle,
-                                    leader_epoch,
                                )
                                .await
                                {
@@ -30,8 +30,8 @@ use crate::{
    store::{Key, Store},
    target::{
        ChannelTargetType, EntityTarget, QueuedPayload, QueuedPayloadMeta, TargetDeliveryCounters, TargetDeliverySnapshot,
-        TargetTlsState, TargetType, build_queued_payload_with_records, build_target_tls_fingerprint, is_connectivity_error,
-        open_target_queue_store, persist_queued_payload_to_store,
+        TargetTlsState, TargetType, build_queued_payload, build_queued_payload_with_records, build_target_tls_fingerprint,
+        is_connectivity_error, open_target_queue_store, persist_queued_payload_to_store,
    },
 };
 use async_trait::async_trait;
@@ -404,7 +404,10 @@ where
    }

    fn build_queued_payload(&self, event: &EntityTarget<E>) -> Result<QueuedPayload, TargetError> {
-        build_queued_payload_with_records(event, vec![event.clone()])
+        match self.args.target_type {
+            TargetType::NotifyEvent => build_queued_payload(event),
+            TargetType::AuditLog => build_queued_payload_with_records(event, vec![event.clone()]),
+        }
    }

    async fn get_or_connect(&self) -> Result<Arc<AMQPConnection>, TargetError> {
@@ -676,6 +679,7 @@ mod tests {
    use serde_json::json;
    use std::path::PathBuf;
    use std::sync::Arc;
+    use std::time::{SystemTime, UNIX_EPOCH};
    use uuid::Uuid;

    fn valid_args() -> AMQPArgs {
@@ -713,6 +717,87 @@ mod tests {
        })
    }

+    fn notification_event() -> EntityTarget<serde_json::Value> {
+        EntityTarget {
+            object_name: "incoming%2Fclip+%252F.mp4".to_string(),
+            bucket_name: "example-bucket".to_string(),
+            event_name: EventName::ObjectCreatedPut,
+            data: json!({
+                "eventVersion": "2.0",
+                "eventSource": "aws:s3",
+                "eventName": "s3:ObjectCreated:Put",
+                "s3": {
+                    "bucket": {"name": "example-bucket"},
+                    "object": {"key": "incoming%2Fclip+%252F.mp4", "size": 42,
+                               "eTag": "example-etag", "versionId": "example-version"}
+                }
+            }),
+        }
+    }
+
+    #[test]
+    fn notification_records_contain_the_event_directly() {
+        let target = AMQPTarget::new("notification".to_string(), valid_args()).unwrap();
+        let event = notification_event();
+        let queued = target.build_queued_payload(&event).unwrap();
+        let payload: serde_json::Value = serde_json::from_slice(&queued.body).unwrap();
+
+        assert_eq!(payload["Records"], json!([event.data]));
+        assert_eq!(payload["Key"], "example-bucket/incoming/clip %2F.mp4");
+        assert_eq!(payload["EventName"], "s3:ObjectCreated:Put");
+        assert_eq!(payload["Records"][0]["s3"]["object"]["key"], event.object_name);
+    }
+
+    #[test]
+    fn audit_records_preserve_the_entity_envelope() {
+        let mut args = valid_args();
+        args.target_type = TargetType::AuditLog;
+        let target = AMQPTarget::new("audit".to_string(), args).unwrap();
+        let event = test_event();
+        let queued = target.build_queued_payload(&event).unwrap();
+        let payload: serde_json::Value = serde_json::from_slice(&queued.body).unwrap();
+
+        assert_eq!(payload["Records"], json!([event.as_ref()]));
+    }
+
+    #[tokio::test]
+    async fn queued_notification_preserves_the_flat_record_and_metadata() {
+        let mut args = unreachable_args();
+        args.queue_dir = temp_store_dir("record-shape").to_string_lossy().to_string();
+        let target = AMQPTarget::new("notification".to_string(), args.clone()).unwrap();
+        let event = notification_event();
+        let expected = target.build_queued_payload(&event).unwrap();
+        let unix_time_ms = || {
+            u64::try_from(
+                SystemTime::now()
+                    .duration_since(UNIX_EPOCH)
+                    .expect("system clock should be after the Unix epoch")
+                    .as_millis(),
+            )
+            .expect("current Unix timestamp should fit in u64")
+        };
+        let before_save = unix_time_ms();
+        target.save(Arc::new(event.clone())).await.unwrap();
+        let after_save = unix_time_ms();
+        let store = target.store().unwrap();
+        let keys = store.list();
+        assert_eq!(keys.len(), 1);
+        let raw = store.get_raw(&keys[0]).unwrap();
+        let queued = QueuedPayload::decode(&raw).unwrap();
+        assert_eq!(queued.body, expected.body);
+        assert_eq!(queued.meta.event_name, expected.meta.event_name);
+        assert_eq!(queued.meta.bucket_name, expected.meta.bucket_name);
+        assert_eq!(queued.meta.object_name, expected.meta.object_name);
+        assert_eq!(queued.meta.content_type, expected.meta.content_type);
+        assert_eq!(queued.meta.payload_len, expected.meta.payload_len);
+        assert_eq!(queued.meta.dedup_id, expected.meta.dedup_id);
+        assert_eq!(queued.meta.failure, expected.meta.failure);
+        assert!((before_save..=after_save).contains(&queued.meta.queued_at_unix_ms));
+        let payload: serde_json::Value = serde_json::from_slice(&queued.body).unwrap();
+        assert_eq!(payload["Records"], json!([event.data]));
+        std::fs::remove_dir_all(args.queue_dir).unwrap();
+    }
+
    fn temp_store_dir(name: &str) -> PathBuf {
        std::env::temp_dir().join(format!("rustfs-amqp-target-{name}-{}", Uuid::new_v4()))
    }
@@ -32,9 +32,9 @@ use lapin::{
 use rustfs_s3_types::EventName;
 use rustfs_targets::Target;
 use rustfs_targets::check_amqp_broker_available;
-use rustfs_targets::target::EntityTarget;
 use rustfs_targets::target::TargetType;
 use rustfs_targets::target::amqp::{AMQPArgs, AMQPTarget};
+use rustfs_targets::target::{EntityTarget, QueuedPayload, QueuedPayloadMeta};
 use serde_json::Value;
 use std::sync::Arc;
 use uuid::Uuid;
@@ -67,7 +67,12 @@ fn entity_for(bucket: &str, object: &str) -> Arc<EntityTarget<serde_json::Value>
        bucket_name: bucket.to_string(),
        object_name: object.to_string(),
        event_name: EventName::ObjectCreatedPut,
-        data: serde_json::json!({"bucket": bucket, "object": object}),
+        data: serde_json::json!({
+            "eventVersion": "2.0",
+            "eventSource": "aws:s3",
+            "eventName": "s3:ObjectCreated:Put",
+            "s3": {"bucket": {"name": bucket}, "object": {"key": object, "size": 42}}
+        }),
    })
 }

@@ -102,7 +107,7 @@ async fn bind_queue(queue: &str, routing_key: &str) -> lapin::Channel {
    channel
 }

-async fn read_one(channel: &lapin::Channel, queue: &str) -> (Value, BasicProperties) {
+async fn read_one_raw(channel: &lapin::Channel, queue: &str) -> (Vec<u8>, BasicProperties) {
    let msg = tokio::time::timeout(std::time::Duration::from_secs(5), async {
        loop {
            if let Some(msg) = channel
@@ -120,8 +125,12 @@ async fn read_one(channel: &lapin::Channel, queue: &str) -> (Value, BasicPropert
    .expect("message should arrive");

    let properties = msg.properties.clone();
-    let payload = serde_json::from_slice(&msg.data).expect("message payload should be JSON");
-    (payload, properties)
+    (msg.data.clone(), properties)
+}
+
+async fn read_one(channel: &lapin::Channel, queue: &str) -> (Value, BasicProperties) {
+    let (body, properties) = read_one_raw(channel, queue).await;
+    (serde_json::from_slice(&body).expect("message payload should be JSON"), properties)
 }

 #[tokio::test]
@@ -147,7 +156,7 @@ async fn test_direct_publish_delivers_json_payload() {

    let (payload, properties) = read_one(&channel, &queue).await;
    assert_eq!(payload["Key"], "bucket1/object-A");
-    assert_eq!(payload["Records"][0]["data"]["bucket"], "bucket1");
+    assert_eq!(payload["Records"], serde_json::json!([entity_for("bucket1", "object-A").data]));
    assert_eq!(properties.content_type().as_ref().map(|s| s.as_str()), Some("application/json"));
    assert_eq!(*properties.delivery_mode(), Some(2));

@@ -209,6 +218,7 @@ async fn test_queue_replay_delivers_and_removes_stored_payload() {

    let (payload, properties) = read_one(&channel, &queue).await;
    assert_eq!(payload["Key"], "bucket1/object-B");
+    assert_eq!(payload["Records"], serde_json::json!([entity_for("bucket1", "object-B").data]));
    assert_eq!(properties.content_type().as_ref().map(|s| s.as_str()), Some("application/json"));
    assert_eq!(*properties.delivery_mode(), Some(2));
    assert_eq!(target.delivery_snapshot().queue_length, 0);
@@ -219,3 +229,42 @@ async fn test_queue_replay_delivers_and_removes_stored_payload() {
        .expect("delete queue");
    let _ = std::fs::remove_dir_all(args.queue_dir);
 }
+
+#[tokio::test]
+#[ignore = "requires running RabbitMQ-compatible AMQP broker"]
+async fn test_legacy_queued_notification_replays_original_bytes() {
+    let routing_key = format!("rustfs.legacy.{}", Uuid::new_v4().simple());
+    let queue = format!("rustfs-test-{}", Uuid::new_v4().simple());
+    let channel = bind_queue(&queue, &routing_key).await;
+    let queue_dir = std::env::temp_dir().join(format!("rustfs-amqp-legacy-{}", Uuid::new_v4()));
+    let mut args = test_args(&routing_key);
+    args.queue_dir = queue_dir.to_string_lossy().to_string();
+    let target = AMQPTarget::<Value>::new("legacy".to_string(), args).expect("construct AMQP target");
+    let event = entity_for("example-bucket", "old-object");
+    // Model the pre-upgrade envelope, including whitespace that reserialization would change.
+    let body = serde_json::to_vec_pretty(&serde_json::json!({
+        "EventName": event.event_name,
+        "Key": "example-bucket/old-object",
+        "Records": [event.as_ref()]
+    }))
+    .unwrap();
+    let meta = QueuedPayloadMeta::new(
+        event.event_name,
+        event.bucket_name.clone(),
+        event.object_name.clone(),
+        "application/json",
+        body.len(),
+    );
+    let encoded = QueuedPayload::new(meta, body.clone()).encode().unwrap();
+    let store = target.store().expect("store configured");
+    let key = store.put_raw(&encoded).expect("persist pre-upgrade payload");
+    target.send_from_store(key).await.expect("replay legacy payload");
+    let (received, _) = read_one_raw(&channel, &queue).await;
+    assert_eq!(received, body);
+    assert!(store.list().is_empty());
+    channel
+        .queue_delete(queue.into(), QueueDeleteOptions::default())
+        .await
+        .expect("delete queue");
+    std::fs::remove_dir_all(queue_dir).unwrap();
+}
@@ -39,7 +39,7 @@ allow-git = [
    # Official s3s repository. Temporarily pinned to the merged generic REST
    # SigV4 payload-checksum fix until it is available in a crates.io release.
    # owner: marshawcoco  review: 2026-10
-    # "https://github.com/s3s-project/s3s.git",
+    "https://github.com/s3s-project/s3s.git",
    # RustFS fork carrying presigned expiry and constant-time authentication fixes.
    # owner: rustfs-maintainers  review: 2026-10
    # "https://github.com/rustfs/s3s.git",
@@ -8,6 +8,8 @@ collection:

 ## Operations

+For legacy protection assessment and bounded protected copies, see [Object integrity inventory, audit, and migration](operations/shard-integrity-audit.md).
+
 Operational runbooks live under [`operations/`](operations/). Replication
 operators should start with:

@@ -18,6 +20,11 @@ operators should start with:
 | [Replication object size limits](operations/replication-object-size-limits.md) | Multipart routing, large-object limits, and retry characteristics. |
 | [Replication outbound transport](operations/replication-outbound-transport.md) | Integrity headers, generic target behavior, and transport knobs. |

+For the erasure-coded cluster lifecycle (planning, parity and `EC:0`,
+expansion, rebalance, decommission, heal, drive replacement, restart
+recovery, and the `rc` CLI mapping), start with
+[Cluster and erasure-coding lifecycle operations](operations/cluster-lifecycle-operations.md).
+
 For persisted administrator bucket tasks and bucket recreation, see
 [Bucket heal recovery](operations/bucket-heal-recovery.md).

@@ -92,7 +92,21 @@ When decommission metadata is present, `decommissionInfo` includes:
 - progress counters: `objectsDecommissioned`, `objectsDecommissionedFailed`, `bytesDecommissioned`, and `bytesDecommissionedFailed`;
 - current location: `bucket`, `prefix`, and `object`;
 - queue/history lists: `queuedBuckets` and `decommissionedBuckets`;
- `waitingReason`: `queued` for queued entries and `waiting_for_worker` when metadata exists but no worker has started.
+- `waitingReason`: `capacity` while the pool is paused on target capacity,
+  `queued` for queued entries, and `waiting_for_worker` when metadata exists but
+  no worker has started;
+- `capacityBlockedReason`: the persisted detail for the active capacity pause,
+  absent once the pause clears.
+
+A capacity pause is reported ahead of the worker states so operators can tell
+"contending but progressing" from "genuinely out of target capacity". The existing `pool.bin` layout is unchanged;
+cumulative pause history requires a separately versioned persistence contract.
+
+The object-attempt metrics count entry passes and repeated version-copy attempts
+within a listing and its deferred replay. Inline gate waits use the gate-retry
+counter. A new listing after a durable pause starts a new per-object observation;
+the maximum is the highest observation in the process, not a persisted lifetime
+attempt count.

 This makes queued pools and stalled metadata visible without requiring operators to inspect pool metadata files directly.

@@ -12,7 +12,7 @@ The broad `rustfs_ecstore::api` facade is a compatibility boundary, not an archi
 | `storage`, `layout`, `error`, `runtime`, `cluster`, `rpc` | Compatibility spine for storage, topology, runtime handles, cluster control, and internode calls. | Keep until replacement contracts compile in downstream boundary files. |
 | `bucket` | Domain facade consumed through owner-local `storage_api` boundaries; explicit submodules and symbol lists, never whole bucket owner modules. | Keep lists aligned with boundary consumers; never restore whole-module passthroughs. |
 | `config`, `disk`, `tier` | Compatibility paths with explicit nested submodules and symbol lists. | Same as `bucket`. |
-| `data_usage`, `capacity`, `notification`, `metrics`, `rebalance` | Domain and service facades consumed through owner-local boundaries. | Narrow one group at a time after explicit aliases or wrappers exist. |
+| `data_usage`, `capacity`, `notification`, `metrics`, `rebalance`, `integrity` | Domain and service facades consumed through owner-local boundaries. | Narrow one group at a time after explicit aliases or wrappers exist. |
 | `set_disk`, `object`, `object_api_utils`, `rio`, `bitrot`, `erasure`, `compression`, `cache`, `store_list` | Low-level object IO, reader, erasure, cache, and migration helper compatibility. | Keep stable while `SetDisks` remains the shared state carrier. |
 | `admin`, `event`, `global` | Admin, event hook, and bootstrap-global compatibility. | `global` is limited to bootstrap writes and lifecycle controls; read-only runtime access goes through `runtime`. |

@@ -28,7 +28,7 @@ External `rustfs_ecstore::api` imports stay in these local boundary files:

 | Boundary file | Facade families consumed |
 |---|---|
-| `rustfs/src/storage/storage_api.rs` | Broad storage-owner bridge: admin, bucket submodules, capacity, compression, cluster, config, data usage, disk, error, event, global bootstrap controls, runtime getters, layout, metrics, notification, rebalance, rio, rpc, set disk, storage, tier. Replication pool/stat handles are projected into RustFS-local wrapper types here. |
+| `rustfs/src/storage/storage_api.rs` | Broad storage-owner bridge: admin, bucket submodules, capacity, compression, cluster, config, data usage, disk, error, event, global bootstrap controls, runtime getters, layout, metrics, notification, rebalance, integrity, rio, rpc, set disk, storage, tier. Replication pool/stat handles are projected into RustFS-local wrapper types here. |
 | `rustfs/src/storage_api.rs`, `rustfs/src/admin/storage_api.rs`, `rustfs/src/app/storage_api.rs` | Root, admin, and app owner boundaries: explicit aliases only, no `metadata`, `metadata_sys`, `quota`, `com`, or bare `init` module passthroughs; object and error aliases anchor on storage-api associated types and a local `StorageError`. |
 | `crates/scanner/src/storage_api.rs` | Bucket lifecycle, replication, metadata, capacity, config, data usage, disk, error, runtime, set disk, storage, tier. Replication queue config, admission, and heal object DTOs are projected into scanner-local types. |
 | `crates/obs/src/metrics/storage_api.rs` | Bucket bandwidth, lifecycle, replication, quota, capacity, data usage, error, runtime, storage; data usage is consumed as a local DTO projection. |
@@ -17,6 +17,7 @@ This document describes the algorithm as implemented on `main`. The *invariant*
  - [placement-repair-invariants.md](placement-repair-invariants.md) — owns object-to-**set** placement (which erasure set a key lands in), per-set readiness/lock quorum, scanner/heal admission, and the behavior-change gates.
  - [ecstore-layout-boundary.md](ecstore-layout-boundary.md) — owns the ecstore module ownership map, `FormatV3` set-ordering and disk-UUID-position invariants, and where the erasure engine physically lives.
  - [decommission-compatibility.md](decommission-compatibility.md) — owns moving encoded objects between pools (decommission/rebalance) and its persisted `PoolMeta` contract.
+  - [../operations/cluster-lifecycle-operations.md](../operations/cluster-lifecycle-operations.md) — owns the operator runbook (planning, `EC:0` consequences, expansion, rebalance, decommission, heal, drive replacement, restart recovery); this spec stays normative for the rules it cites.
  - [../operations/tier-ilm-debugging.md](../operations/tier-ilm-debugging.md) — owns the ILM/tier runtime runbook, the dual internal-metadata-key table, the defensive binary-UUID read pattern, and the `xl.meta` inspection tooling (`dump_fileinfo` / `dump_versions`).
  - [AGENTS.md](../../AGENTS.md) "Cross-Cutting Domain Invariants" — dual internal metadata keys, defensive UUID reads, unversioned tier buckets.
  - [../../ARCHITECTURE.md](../../ARCHITECTURE.md) — crate roles (`ecstore`, `filemeta`, the `rustfs-erasure-codec` codec fork, `heal`, `scanner`).
@@ -41,7 +41,7 @@ Every route asserted in `rustfs/src/admin/route_registration_test.rs` is registe
 | Pools decommission / cancel / clear | 部分兼容 | `rustfs/src/admin/handlers/pools.rs` returns `NotImplemented` when endpoints are not initialized (single-pool or uninitialized clusters). |
 | `/v3/top/drives`, `/v3/top/net` | 缺失 | Only `/v3/top/locks` is registered (`rustfs/src/admin/handlers/diagnostics.rs`). |
 | Bucket / site replication per-object diff | 缺失 | `/v3/replicationmetrics` and site-replication status exist; no diff endpoint. |
-| MRF (most-recent-failures) replication metrics breakdown | 缺失 | Only the generic `/v3/metrics` stream and replication metrics wire (`rustfs/src/admin/replication_metrics_wire.rs`). |
+| MRF (most-recent-failures) replication metrics breakdown | 缺失 | Only the generic `/v3/realtime` stream and replication metrics wire (`rustfs/src/admin/replication_metrics_wire.rs`). |

 Formerly-missing families that are now registered and therefore not exceptions: `/v3/healthinfo`, `/v3/obdinfo`, `/v3/force-unlock`, `/v3/top/locks`, `/v3/speedtest*`, `/v3/log`, `/v3/trace`, `/v3/profile`, `/v3/profiling/*`, `/v3/idp/{ldap|openid}/*`, `/v3/idp-config/*`.

@@ -0,0 +1,51 @@
+# AMQP notification record layout
+
+New AMQP bucket notifications place the S3 event directly in `Records`, matching
+Kafka and webhook notification targets. For example, a consumer reads the bucket
+from `Records[0].s3.bucket.name` and the object key from
+`Records[0].s3.object.key`. Previously these fields were nested under
+`Records[0].data` in an internal target envelope.
+
+This follows the record layout described in the
+[AWS S3 notification message structure](https://docs.aws.amazon.com/AmazonS3/latest/userguide/notification-content-structure.html).
+It changes the record envelope only; the event's existing fields and values are
+preserved. The outer `EventName` and `Key` fields are retained. The outer `Key`
+decodes the object name once, while `Records[0].s3.object.key` retains the event's
+encoded key. AMQP audit log payloads retain their existing envelope.
+
+## Upgrading consumers
+
+Consumers can read event fields at the standard S3 record paths. Consumers built
+around the previous AMQP envelope must be updated before upgrading producers. During rolling upgrades or queue draining, consumers
+may receive both layouts. A temporary compatibility path can unwrap `record.data`
+when present and otherwise use `record` directly.
+
+Existing disk-queued messages are replayed byte for byte. Upgrading does not
+rewrite their bodies. Newly queued notifications use the same flat record as
+direct publishing. Retain dual-layout handling until old producers are upgraded
+and old queued messages have drained. If rolling back a producer, retain that
+handling while any flat messages remain queued or in the broker.
+
+## Verification
+
+The regression `notification_records_contain_the_event_directly` fails against
+the old serializer because `Records[0]` contains the target envelope. It passes
+with the notification serializer using the shared event-record builder. Unit
+tests also cover encoded keys, preserved metadata in the queue, and the unchanged
+audit envelope:
+
+```bash
+cargo test -p rustfs-targets --lib target::amqp::tests
+```
+
+Broker integration tests exercise direct publishing, disk-queue replay, reconnect,
+and byte-for-byte replay of a pre-upgrade envelope. They use synthetic events and
+require an isolated RabbitMQ-compatible broker:
+
+```bash
+RUSTFS_TEST_AMQP_URL='amqp://guest:guest@127.0.0.1:5672/%2f' \
+  cargo test -p rustfs-targets --test amqp_integration -- --ignored
+```
+
+These checks cover the target's serialization and delivery boundary. They do not
+exercise the full bucket-notification pipeline or every downstream S3 client.
@@ -0,0 +1,290 @@
+# Cluster and erasure-coding lifecycle operations
+
+**Use this when:** planning, starting, expanding, rebalancing, decommissioning, healing, replacing drives in, or restarting an erasure-coded RustFS deployment, or deciding whether an `EC:<parity>` setting (including `EC:0`) is safe for a workload.
+**Source of truth:** `crates/ecstore/src/layout/disks_layout.rs` (`DisksLayout::from_volumes`, `SET_SIZES`, `RUSTFS_ERASURE_SET_DRIVE_COUNT`), `crates/ecstore/src/config/storageclass.rs` (`default_parity_count`, `validate_parity_inner`, `Config::should_inline`, `Config::effective_inline_block`), `crates/ecstore/src/set_disk/mod.rs` (`resolve_write_layout`, `WriteLayout::from_parity`), `crates/ecstore/src/set_disk/metadata.rs` (`object_quorum_from_meta`, `common_parity`), `crates/ecstore/src/set_disk/ops/heal.rs` (no-parity `cannot_heal` classification), `crates/ecstore/src/store/init_format.rs` (`check_format_erasure_value_for_topology`), `rustfs/src/admin/handlers/pools.rs`, `rustfs/src/admin/handlers/rebalance.rs`, `rustfs/src/admin/handlers/heal.rs`, and `rustfs/src/admin/route_policy.rs` (admin surface).
+
+This runbook is the operator entry point. It restates only what an operator needs to act and links the normative contracts for everything else: [erasure-coding.md](../architecture/erasure-coding.md) (algorithm, quorum, on-disk format), [ecstore-layout-boundary.md](../architecture/ecstore-layout-boundary.md) (immutable layout), [decommission-compatibility.md](../architecture/decommission-compatibility.md) (decommission and rebalance contract), [heal-concurrency-model.md](../architecture/heal-concurrency-model.md) (heal versus write locking), [rolling-restart.md](rolling-restart.md) (restarts), [replacement-generation-recovery.md](replacement-generation-recovery.md) (drive replacement recovery), and [scanner-runtime-controls.md](scanner-runtime-controls.md) (heal and scanner knobs). Admin routes below require an authenticated request holding the matching `AdminAction`; the route to action mapping is owned by `rustfs/src/admin/route_policy.rs` (see [admin-route-action-snapshot.md](../architecture/admin-route-action-snapshot.md)).
+
+## 1. Immutable layout boundaries
+
+A pool's endpoint count, set count, drives per set, position in the pool list, and the disk UUID positions in `FormatV3` are persisted layout. On startup `check_format_erasure_value_for_topology` (`crates/ecstore/src/store/init_format.rs`) compares the stored format with the configured drive count and set width and fails permanently with `PoolTopologyMismatch` (or `UnsupportedSnsdExpansion` for a single-drive deployment) instead of migrating. Restore the original endpoints and set width; do not delete `format.json` to get past the error. Details and the regression matrix are in [pool-layout-compatibility.md](../testing/pool-layout-compatibility.md).
+
+Never do these:
+
+- overwrite an existing pool with more or fewer endpoints, or change `RUSTFS_ERASURE_SET_DRIVE_COUNT` for an initialized pool;
+- reorder volume arguments or move a drive directory to another slot;
+- delete, copy, or rename object shards, `xl.meta`, or anything under `.rustfs.sys` (`RUSTFS_META_BUCKET` in `crates/ecstore/src/disk/mod.rs`) by hand;
+- treat a manual file copy as data movement. Rebalance and decommission move objects under object locks with version re-checks and source cleanup; nothing else does.
+
+Runtime roles:
+
+- **Pool**: one endpoint list with persisted layout. Multi-drive sets contain 2 to 16 drives (`SET_SIZES`); a single local path is the only single-drive layout (`is_single_drive_layout`).
+- **Pool leader**: the first endpoint of a pool. Decommission start, cancel, and clear may be sent to any node; when the target pool's first endpoint is remote the handler forwards over authenticated internode RPC (`decommission_peer_target` in `rustfs/src/admin/handlers/pools.rs`), and the leader still enforces `ensure_local_decommission_pool_leader` before mutating state.
+- **Cluster-wide state**: rebalance and root heal propagate to peers and recover from persisted metadata on restart (§9).
+
+## 2. Planning and first start
+
+1. Plan endpoints by node, rack, power, and network failure domain. Topology admission is not a high-availability guarantee: a single-node pool loses every shard with its host, and at startup `log_storage_pool_layout` (`rustfs/src/startup_storage.rs`) warns `host_failure_data_unavailable` for every multi-drive pool.
+2. Pick each pool's set width. Pools may differ in width, but each pool must split evenly into sets of one width from `SET_SIZES`; `RUSTFS_ERASURE_SET_DRIVE_COUNT` may pin a width that is already a symmetric divisor, nothing else.
+3. Pick STANDARD and RRS parity per §4, then validate against the narrowest pool, because an explicit parity must fit every pool.
+4. Record the full launch command, endpoint order, pool order, set width, and storage-class variables. They must be identical on every node and on every restart.
+5. Give RustFS exclusive ownership of every endpoint path.
+
+Volumes come from positional arguments or `RUSTFS_VOLUMES`; `rustfs` with no arguments behaves as `rustfs server` with volumes from the environment (`rustfs/src/config/cli.rs`). Ellipsis expansion follows MinIO:
+
+```bash
+# one pool, four nodes, four drives each: sixteen endpoints
+rustfs server http://node{1...4}:9000/data{1...4}
+
+# two pools; every argument must carry an ellipsis once there is more than one
+rustfs server http://node{1...4}:9000/data{1...4} http://node{5...8}:9000/data{1...4}
+```
+
+Rules enforced by `DisksLayout::from_volumes`: with more than one argument every argument needs an ellipsis; each pool must expand to at least two distinct drive endpoints; a single-drive pool cannot join a multi-pool deployment; a singleton range such as `{3...3}` does not bypass the minimum.
+
+Startup order (`init_startup_storage_foundation` in `rustfs/src/startup_storage.rs`): parse endpoints, validate storage classes against every pool (`ECStore::validate_startup_storage_class`), enforce the unsupported-filesystem policy, initialize local disks, initialize lock clients, then load or create the format quorum and build the store. A storage-class error stops startup before any disk is touched.
+
+## 3. Startup configuration
+
+| Variable | Effect | Boundary |
+|---|---|---|
+| `RUSTFS_ERASURE_SET_DRIVE_COUNT` | Pin the set width | Must be a valid symmetric divisor of the pool, at most `MAX_ERASURE_SET_DRIVE_COUNT` (16); cannot change an initialized pool |
+| `RUSTFS_STORAGE_CLASS_STANDARD` | STANDARD parity as `EC:<n>` | Validated per pool; affects new writes only |
+| `RUSTFS_STORAGE_CLASS_RRS` | RRS parity as `EC:<n>` | Validated per pool; an explicit value fails on a single drive, the persisted default `EC:1` resolves to `0` there |
+| `RUSTFS_STORAGE_CLASS_OPTIMIZE` | Accepted for MinIO compatibility | Stored on `Config`, but `Config::capacity_optimized` has no caller: it changes nothing at runtime today |
+| `RUSTFS_STORAGE_CLASS_INLINE_BLOCK` | Fixed per-shard inline limit (bytesize syntax such as `128KiB`) | Replaces the scaled default in §4.3; values above 128 KiB log `storage_class_inline_block_large` |
+
+An empty `RUSTFS_STORAGE_CLASS_STANDARD` or `RUSTFS_STORAGE_CLASS_RRS` restores the automatic policy and overrides a persisted value (`standard_policy` / `rrs_policy`). `MINIO_`-prefixed spellings of these keys, of `ERASURE_SET_DRIVE_COUNT`, and of `VOLUMES` are mapped onto the `RUSTFS_` names when the `RUSTFS_` name is unset (`apply_external_env_compat` in `crates/utils/src/envs.rs`); when both are set the `RUSTFS_` value wins and the conflict is reported at startup.
+
+Heal knobs (`RUSTFS_HEAL_*`) and their defaults are documented once, in [scanner-runtime-controls.md](scanner-runtime-controls.md#heal-runtime-controls); the deep bitrot cycle is the admin config key `heal.bitrot_cycle` (environment `RUSTFS_SCANNER_BITROT_CYCLE_SECS`), not a `RUSTFS_HEAL_*` variable. Data-movement knobs: `RUSTFS_REBALANCE_MAX_ATTEMPTS` (default `DEFAULT_REBALANCE_MAX_ATTEMPTS` = 3, `crates/ecstore/src/services/rebalance/mod.rs`), `RUSTFS_DECOMMISSION_BUCKET_CONCURRENCY` (default cap 4) and `RUSTFS_DECOMMISSION_ENTRY_CONCURRENCY` (default cap 8, hard cap 64) in `crates/ecstore/src/core/pools.rs`, and `RUSTFS_LIST_OBJECTS_QUORUM` (default `optimal`; `disk`, `reduced`, `optimal`, `auto`, anything else `strict`, per `normalize_list_quorum` in `crates/ecstore/src/store/list_objects.rs`). None of them relaxes object quorum, object locks, or version checks.
+
+## 4. Storage classes, parity, capacity, quorum
+
+### 4.1 Geometry
+
+Each object is Reed-Solomon coded across the `N` drives of one set into `K` data and `M` parity shards, `N = K + M`; any `K` shards rebuild it. The geometry is stored per object in `xl.meta`, so changing the default later rewrites nothing. Objects hash to a set (`get_hashed_set_index`, [placement-repair-invariants.md](../architecture/placement-repair-invariants.md)) and shards rotate within it; sets never share a failure budget.
+
+Default STANDARD parity (`default_parity_count`): 0 for one drive, 1 for 2 to 3, 2 for 4 to 5, 3 for 6 to 7, 4 for 8 to 16. Default RRS parity is 1 (0 on a single drive).
+
+Validation (`validate_parity_inner`): each parity is at most `N/2` for every pool, and STANDARD parity is at least RRS parity **only when both are non-zero**. There is no minimum, so `EC:0` passes on any width (§5). Only `STANDARD` and `REDUCED_REDUNDANCY` are accepted on PUT, CopyObject, and CreateMultipartUpload; AWS labels return `InvalidStorageClass`.
+
+### 4.2 Quorum and capacity
+
+```text
+usable capacity  ≈ raw × K / N
+read quorum      = K            (metadata vote and shard decode)
+write quorum     = K, or K + 1 when K == M
+delete markers   = N/2 + 1      (majority, both write and vote)
+```
+
+Sixteen drives at `EC:4`: `K = 12`, about 75 percent usable, four failures tolerated, twelve valid shards needed to read. Two drives at `EC:1`: `K = M = 1`, so one drive down still reads but no longer writes. Internal metadata under `.rustfs.sys` is written with `max_parity`, meaning parity `N/2`, regardless of the configured class. Contract: [erasure-coding.md §2.2 and §10](../architecture/erasure-coding.md#22-parity-selection).
+
+### 4.3 Inline objects
+
+An object is stored inline in `xl.meta` when every shard is at most the per-shard budget (`Config::should_inline`). The default budget is `DEFAULT_INLINE_OBJECT_BUDGET / K` capped at `DEFAULT_INLINE_BLOCK` (256 KiB / K, at most 128 KiB per shard), so a wider set does not raise the maximum inline object size; on a versioned bucket the budget is divided by 8. `RUSTFS_STORAGE_CLASS_INLINE_BLOCK` replaces the scaled default with a fixed per-shard limit. Each drive keeps only its own shard inline, so inline data has the same loss semantics as external shards. Compressed or encrypted streams inline only when the stored size is known and within budget.
+
+### 4.4 Write and read paths
+
+Writes resolve `WriteLayout` per pool (`resolve_write_layout`), encode with `rs-vandermonde` in 1 MiB blocks (`BLOCK_SIZE_V2`), write bitrot-protected shards, and commit only after write quorum. A write below quorum is never reported as success; rollback is best effort and residue is reconciled by heal or the scanner ([erasure-coding.md §7](../architecture/erasure-coding.md#7-write-path-and-write-quorum)). Reads pick the authoritative metadata at read quorum, verify bitrot, decode from at least `K` shards, and fail closed below that; a read that succeeds with missing shards enqueues read-repair ([§8](../architecture/erasure-coding.md#8-read-path-and-read-quorum)).
+
+Multipart parts use the object's geometry and write quorum; `CompleteMultipartUpload` re-checks metadata and commit quorum, so an uploaded part is not a committed object.
+
+## 5. `EC:0`: zero-parity mode
+
+`EC:0` is accepted for STANDARD, for RRS, or for both, on any set width. It gives raw capacity and no redundancy. Every statement here is what the code does, not a guideline.
+
+**Configuration.**
+
+- Explicit `EC:0` produces no startup warning. The `storage_class_zero_redundancy` warning (`publish_storage_class_config` in `crates/ecstore/src/config/mod.rs`) fires only for automatic zero parity, meaning a single-drive pool.
+- `STANDARD=EC:2` with `RRS=EC:0` is valid because the ordering rule is skipped when either side is zero. Only objects written with `x-amz-storage-class: REDUCED_REDUNDANCY` then carry the semantics below; STANDARD objects keep parity 2.
+- The set's default parity (`SetDisks.default_parity_count`) stays the topology default from `ec_drives_no_config`, not 0. Quorum arithmetic for a multi-drive set therefore takes the parity-present branch of `object_quorum_from_meta`; the `default_parity_count == 0` branch is the single-drive case.
+
+**Writes.** `WriteLayout::from_parity(N, 0)` yields `K = N`, `M = 0`, write quorum `N`. If any drive of the set is unwritable, `put_object` returns `ErasureWriteQuorum` before or after encoding; nothing is committed. Because a zero-parity shard cannot be rebuilt later, the write path verifies every shard it just wrote against its bitrot hash before commit (`verify_written_bitrot_shards` in `put_object` and `put_object_part`, gated on `parity_blocks == 0`) and refuses to commit a shard that already fails.
+
+**Reads.** `common_parity` classifies a parity value of 0 like a delete marker for vote selection (majority of replies), but the resulting read quorum is `data_blocks = N`. `find_file_info_in_quorum` (GET and HEAD) and `pick_valid_fileinfo` (listing cache) then need `N` matching metadata copies, and decode needs `N` shards. With any drive of the set offline, HEAD and GET of every zero-parity object in that set fail with `ErasureReadQuorum`. There is no degraded read.
+
+**Deletes.** Delete markers and version deletes use majority quorum (`disks.len() / 2 + 1`), so a delete can succeed on a set whose reads and writes are failing. A successful delete is not evidence that the set is healthy.
+
+**Heal.** `heal_object_with_explicit_version_regen` computes `cannot_heal` when the metadata copies to heal exceed `parity_blocks` (0) or any part has more missing or corrupt shards than `parity_blocks` (0). The result is `no-parity object is unrecoverable` with `FileCorrupt` for a bitrot failure and `ErasureReadQuorum` for a missing shard. Scanner and MRF heal requests repeat that outcome; they never repair. A replaced or wiped drive therefore loses every zero-parity object that had a shard on it, permanently, unless a copy exists outside the set. Recovery boundary and evidence handling: [no-parity-bitrot-recovery.md](no-parity-bitrot-recovery.md).
+
+**What still works with one drive down.** Bucket operations and `.rustfs.sys` metadata (written at parity `N/2`), delete markers, and listings that satisfy the list quorum. Everything else in the set is unavailable, including rebalance and decommission of its objects, which must read all `N` shards.
+
+**Operational consequences.**
+
+- Any maintenance that takes a drive or its node offline is a read and write outage for ordinary objects in every set that spans it, for the whole window. Rolling restarts are not transparent.
+- Drive replacement is data loss for that set, not a repair.
+- Raising parity later rewrites nothing; re-upload or server-side copy each object after the change to re-protect it.
+- Use `EC:0` only where an external layer (RAID, replication, backup, or a re-creatable source) owns durability and availability, and only with an explicit sign-off that one drive loss equals set loss. Production deployments should use `EC:1` or higher; `EC:2` or more per set is the usual floor once a set spans more than one node.
+
+## 6. Post-start verification
+
+Read-only checks:
+
+```text
+GET /rustfs/admin/v3/pools/list
+GET /rustfs/admin/v3/pools/status?pool=<pool>[&by-id=true]
+GET /rustfs/admin/v3/decommission/status[?pool=<pool>&by-id=true]
+GET /rustfs/admin/v4/cluster/snapshot
+GET /rustfs/admin/v4/runtime/capabilities
+```
+
+`pool` is a pool command line, or a zero-based index with `by-id=true`. Status queries ignore unknown parameters; mutation queries reject unknown or duplicate parameters and a `by-id` value other than `true` or `false` (`parse_pool_query`). Per-pool status fields, including `decommissionInfo`, are listed in [decommission-compatibility.md](../architecture/decommission-compatibility.md#status-response-shape); the `storage_classes` payload of runtime capabilities is owned by [runtime-capability-contracts.md](../architecture/runtime-capability-contracts.md); `pool_meta_write_gate` in the cluster snapshot is explained in [s3-write-failure-diagnostics.md](s3-write-failure-diagnostics.md).
+
+Confirm: pool count, endpoints per pool, set count and width per pool, local versus remote endpoints, no legacy flag on a deployment that will ever decommission, peer readiness, and no topology mismatch.
+
+Data-plane checks: write one STANDARD and one RRS object; GET, HEAD, range GET, CopyObject, and a multipart upload; take one drive or node offline within parity and confirm reads continue and writes behave per §4.2; bring it back and confirm the heal queue drains and bitrot errors stop; confirm every node reports the same pool metadata and cluster snapshot.
+
+## 7. Expansion
+
+An existing pool cannot grow in place. Add a pool:
+
+1. Plan its nodes, drives, and failure domain; each new pool needs at least two drive endpoints and its own valid set layout.
+2. Prepare empty, RustFS-exclusive paths on every node.
+3. Append the pool to the launch arguments on **every** node, keeping existing pools, their order, endpoints, and set width unchanged. A single-node single-drive deployment cannot expand; migrate through S3 instead.
+4. Restart per [rolling-restart.md](rolling-restart.md) and verify per §6.
+5. New writes are placed across available pools; old objects stay where they are. To move them, start a rebalance (§8). A restart never rebalances.
+
+Adding a pool does not add parity to existing objects and does not copy anything. Cross-pool movement is rebalance or decommission only.
+
+## 8. Rebalance
+
+Rebalance redistributes objects across pools by usage. Preconditions checked by `RebalanceStart` (`rustfs/src/admin/handlers/rebalance.rs`): more than one pool (a single pool returns `NotImplemented`), no decommission running or conflicting (`InvalidRequest`), and no query string on `start` or `stop` (`InvalidArgument`).
+
+```text
+POST /rustfs/admin/v3/rebalance/start     → {"id": "<uuid>"}
+GET  /rustfs/admin/v3/rebalance/status
+POST /rustfs/admin/v3/rebalance/stop
+```
+
+Status (`RebalanceAdminStatus`): `id`; `pools[]` with `id`, `status`, `stopping`, `used`, `lastError`, `cleanupWarnings`, and `progress` (`objects`, `versions`, `bytes`, `remainingBuckets`, `bucket`, `object`, `elapsed`, `eta`); `stoppedAt`; and `stopPropagation` (`lastAttemptAt`, `failedPeers`, `terminalReloadAttemptAt`, `terminalReloadFailedPeers`, `pendingTerminalReload`).
+
+Start propagates an admission fence and worker state to peers; a propagation failure rolls back to a terminal state and reports any peer that did not roll back. Stop closes admission, stops local and remote workers, and persists the stopped state. Treat a rebalance as finished only when no pool is active or stopping, `lastError` is empty or explained, `cleanupWarnings` are understood, `pendingTerminalReload` is false, and migrated objects read from their new pool. Per-pool state lives in `rebalance.bin` (`REBAL_META_NAME`); never delete it to unblock a run. If the deployment ran rebalance on a release listed in [rebalance-stored-representation-impact.md](rebalance-stored-representation-impact.md), run that assessment first.
+
+## 9. Decommission
+
+Contract: [decommission-compatibility.md](../architecture/decommission-compatibility.md). Operator summary:
+
+- Supported only on multi-pool, non-legacy (ellipsis) deployments; anything else returns `NotImplemented`. A rebalance in progress blocks start.
+- Targets are validated as a batch before anything is persisted: unknown, duplicate, active, queued, or completed targets are rejected; failed or canceled targets must be cleared first unless they hold unresolved listing entries, in which case a new start is the retry path (`DecommissionStartPoolState::Retryable` versus `Blocked` in `crates/ecstore/src/core/pools.rs`).
+- Requests go to any node and are forwarded to the pool leader (§1).
+
+```text
+POST /rustfs/admin/v3/pools/decommission?pool=<a>[,<b>]      # by command line
+POST /rustfs/admin/v3/pools/decommission?pool=1&by-id=true   # by index
+GET  /rustfs/admin/v3/decommission/status[?pool=1&by-id=true]
+POST /rustfs/admin/v3/pools/cancel?pool=1&by-id=true
+POST /rustfs/admin/v3/pools/clear?pool=1&by-id=true
+```
+
+Entry states persisted in `pool.bin`: `queued` → `active` → `completed`, `failed`, or `canceled`. One entry owns a worker at a time; startup resumes non-terminal entries and skips terminal predecessors. A source pool in any decommission state, terminal or not, rejects new ordinary PUTs and new multipart uploads (a staged PUT gets `SlowDown`); existing multipart uploads may drain while the source is non-terminal. `clear` removes failed or canceled metadata only and never moves data back. `queuedBuckets` is an inventory, not proof of a live worker.
+
+Remove a pool from the launch arguments only after its entry is `completed`, free versions and tier ownership have converged, and every node shows the same status. Never detach source drives before that, and never substitute a directory delete for the configuration change.
+
+## 10. Heal and drive replacement
+
+### 10.1 Admin heal
+
+```text
+POST /rustfs/admin/v3/heal/                    # root: cluster, or one erasure set with pool+set
+POST /rustfs/admin/v3/heal/<bucket>
+POST /rustfs/admin/v3/heal/<bucket>/<prefix>
+POST /rustfs/admin/v3/background-heal/status
+GET  /rustfs/admin/v4/heal/replacement-recovery
+```
+
+The JSON body carries `recursive`, `dryRun`, `remove`, `recreate`, `scanMode`, `updateParity`, `nolock`, `readRepair`, `pool`, and `set`. Handler rules (`rustfs/src/admin/handlers/heal.rs`):
+
+- a root heal must set `recursive=true` or name both `pool` and `set`; a bucket heal without a prefix is always recursive, and a prefix heal honors `recursive`;
+- `readRepair=true` is rejected with `InvalidArgument`;
+- `nolock` is accepted on the wire but forced to `false` (`build_heal_channel_request`); admin heals always take the object namespace lock, so there is no lock bypass to misuse;
+- `clientToken` correlates start, status, and stop; `forceStart` and `forceStop` are validated against it and against each other, and unknown or duplicate query keys are rejected;
+- overlapping starts follow `RUSTFS_HEAL_OVERLAP_POLICY`.
+
+Heal selects the authoritative version from quorum metadata, refuses when more shards are missing than parity can rebuild or when the stored geometry does not match the set (`heal-concurrency-model.md`, [erasure-coding.md §9](../architecture/erasure-coding.md#9-healing)), and shares the `(bucket, object)` lock with PUT, delete, multipart complete, and data movement. Retained outcomes are described in [heal-terminal-reports.md](../architecture/heal-terminal-reports.md).
+
+### 10.2 Drive replacement
+
+1. Confirm the affected set still meets read quorum for its objects (impossible at `EC:0`, see §5) and record pool index, set index, slot, and the failed disk's UUID from the pool status.
+2. Replace the drive at the **same** endpoint and mount path with an empty filesystem owned by RustFS. Do not copy the old drive's directories; a copied `format.json` claims an identity the set already tracks.
+3. Bring the node back. The empty drive is not a format-quorum member, so `retain_format_quorum_members` keeps its slot offline; the endpoint monitor (`monitor_and_connect_endpoints_task` in `crates/ecstore/src/core/sets.rs`) retries the connection every 15 seconds. The erasure-set heal task (`crates/heal/src/heal/task/heal_erasure_set.rs`) writes the format into the original slot, through `heal_replacement_format` when an automatic replacement intent exists (schema 7 records under `.rustfs.sys/buckets/ahm-replacement/`, created when startup or the disk scanner observes a changed mount identity) or through `heal_format` for an admin erasure-set heal (`POST /rustfs/admin/v3/heal/` with `pool` and `set`), and then rebuilds the shards. Legacy schema 5/6 intents follow [replacement-generation-recovery.md](replacement-generation-recovery.md). If the drive held the node's pool metadata, follow [pool-metadata-recovery.md](pool-metadata-recovery.md#disk-replacement-and-metadata-erasure) first.
+4. Watch `GET /rustfs/admin/v4/heal/replacement-recovery`, background heal status, disk health, and bitrot counters. Completion criteria are in [scanner-runtime-controls.md](scanner-runtime-controls.md#replacement-recovery-completion): an `idle` queue or a readable object is not proof.
+5. The replacement is done when recovery reports `completed` for that instance and sampled objects have `xl.meta` and parts on the new drive.
+
+If a set has lost more drives than its parity, heal cannot rebuild; recover from replication, a remote tier, or backup.
+
+## 11. Restart and recovery
+
+- **Ordinary restart.** No topology change in flight, identical arguments on every node, one node at a time per [rolling-restart.md](rolling-restart.md). Startup reloads pool metadata, `rebalance.bin`, the decommission queue, and heal recovery records, and resumes workers by leader and owner rules.
+- **Rebalance.** Read `/rebalance/status` first. Handle active or stopping pools, `lastError`, cleanup warnings, and stop propagation before deciding to stop or restart. Do not delete `rebalance.bin`.
+- **Decommission.** Non-terminal entries resume; failed and canceled entries survive restarts until cleared or retried through the allowed path (§9).
+- **Root heal.** A graceful shutdown persists an unfinished cluster heal as `root-heal-<task-id>.json` in `.rustfs.sys` (`RootHealRecovery` in `crates/heal/src/heal/manager/root_recovery.rs`, 64 KiB bound) and replays it with the same task id. Terminal records use `terminal-root-heal-`; invalid, oversized, or unsupported records move to `quarantined-root-heal-intent-` or `quarantined-root-heal-terminal-` and are skipped with a log line until an operator repairs them. Do not delete quarantine markers or reuse a task id.
+- **Unclean shutdown.** When the previous run left its marker behind, the heal manager's `process_unclean_shutdown` (`crates/heal/src/heal/manager/unclean_shutdown.rs`) enqueues a full erasure-set heal for every local set; graceful shutdown clears the marker. See [durability-modes.md](durability-modes.md) for which sync tier can leave what behind.
+
+## 12. Mutual exclusion and prohibited actions
+
+| Operation | With rebalance | With decommission | Notes |
+|---|---|---|---|
+| PUT, DELETE, multipart complete | Object lock | Source pool rejects new publication | Never bypass the pool fence |
+| Decommission | Rejected while rebalance runs | Serial queue | Leader-only mutation |
+| Rebalance | Multi-pool only | Rejected while decommission runs or conflicts | No query parameters |
+| Admin heal | Object lock | Repairs data, never changes decommission state | `nolock` is ignored |
+| Drive replacement | Cannot change `FormatV3` layout | Avoid on a source pool mid-decommission | Same slot, empty drive |
+
+Do not: change launch endpoints while rebalance or decommission runs; delete `.rustfs.sys`, `rebalance.bin`, `pool.bin`, or root-heal records; run manual heal, file copies, and drive replacement on one pool at the same time; treat an HTTP 200 on a start call as completion.
+
+## 13. Checklists
+
+After start or expansion: identical endpoints and pool order on every node; topology equals `FormatV3`; peers ready and the pool metadata write gate open; new endpoints unique and RustFS-exclusive; GET, HEAD, range GET, and multipart verified.
+
+After rebalance: no pool active or stopping; `lastError` empty or explained; `cleanupWarnings` handled; `pendingTerminalReload` false; migrated objects readable.
+
+After decommission: status `completed`; other entries handled by policy; source pool rejects new writes; free versions, tier ownership, and listing entries converged; launch arguments updated on every node before drives are detached.
+
+After heal or replacement: replacement recovery `completed`; slot and UUID correct; heal queue drained; no growing bitrot count; sampled current and historical versions readable; no leftover temporary directories.
+
+## 14. `rc` CLI mapping
+
+`rc` ships from the separate `rustfs/cli` repository; its admin reference is https://github.com/rustfs/cli/blob/main/docs/reference/rc/admin.md, and its `expand` group is an alias of `rebalance` (there is no server-side expand API). Flags below are the ones that reference documents; anything else, check `rc admin <group> --help` for the installed version.
+
+```bash
+rc alias set local http://node1:9000 <ACCESS_KEY> <SECRET_KEY>
+rc ping local
+rc ready local --timeout 2
+rc admin info cluster local
+rc admin info disk local --offline
+rc admin pool list local
+rc admin pool status local 0 --by-id
+rc admin rebalance start local
+rc admin rebalance status local
+rc admin rebalance stop local
+rc admin decommission start local '/data/pool1/disk{1...4}'
+rc admin decommission status local 1 --by-id
+rc admin decommission cancel local 1 --by-id
+rc admin decommission clear local 1 --by-id
+rc admin heal start local --scan-mode deep
+rc admin heal start local --bucket logs --prefix 2026/ --scan-mode deep
+rc admin heal status local --client-token <TOKEN_FROM_START>
+rc admin heal stop local --client-token <TOKEN_FROM_START>
+```
+
+| `rc` command | Server route |
+|---|---|
+| `rc ping` / `rc ready` | `GET /health` / `GET /health/ready` |
+| `rc admin pool list` / `status` | `GET /rustfs/admin/v3/pools/list` / `GET /rustfs/admin/v3/pools/status` |
+| `rc admin rebalance …` (alias `expand`) | `/rustfs/admin/v3/rebalance/*` |
+| `rc admin decommission start` / `cancel` / `clear` | `POST /rustfs/admin/v3/pools/decommission` / `cancel` / `clear` |
+| `rc admin decommission status` | `GET /rustfs/admin/v3/decommission/status` |
+| `rc admin heal start` / `status` / `stop` | `/rustfs/admin/v3/heal/*`, `POST /rustfs/admin/v3/background-heal/status` |
+
+`rc admin heal start` exposes `--bucket`, `--prefix`, `--scan-mode`, `--remove`, `--recreate`, and `--dry-run`; it does not expose `nolock`, `forceStart`, `forceStop`, `pool`, or `set`. There is no `rc` command for replacement recovery; call `GET /rustfs/admin/v4/heal/replacement-recovery` directly.
+
+## 15. Source and contract index
+
+- [erasure-coding.md](../architecture/erasure-coding.md): algorithm, quorum, on-disk format, change procedure.
+- [ecstore-layout-boundary.md](../architecture/ecstore-layout-boundary.md): `FormatV3` ordering and disk UUID positions.
+- [placement-repair-invariants.md](../architecture/placement-repair-invariants.md): object to set placement, scanner and heal admission.
+- [decommission-compatibility.md](../architecture/decommission-compatibility.md): decommission and rebalance contract, `PoolMeta`, status shape.
+- [heal-concurrency-model.md](../architecture/heal-concurrency-model.md) and [heal-terminal-reports.md](../architecture/heal-terminal-reports.md): heal locking and retained reports.
+- [pool-layout-compatibility.md](../testing/pool-layout-compatibility.md): expansion admission rules and regression matrix.
+- [no-parity-bitrot-recovery.md](no-parity-bitrot-recovery.md), [durability-modes.md](durability-modes.md), [rolling-restart.md](rolling-restart.md), [pool-metadata-recovery.md](pool-metadata-recovery.md), [replacement-generation-recovery.md](replacement-generation-recovery.md), [scanner-runtime-controls.md](scanner-runtime-controls.md).
+- `rustfs/src/startup_storage.rs`, `crates/ecstore/src/layout/disks_layout.rs`, `crates/ecstore/src/config/storageclass.rs`, `crates/ecstore/src/set_disk/mod.rs`, `crates/ecstore/src/set_disk/metadata.rs`, `crates/ecstore/src/set_disk/ops/heal.rs`, `crates/ecstore/src/core/pools.rs`, `crates/ecstore/src/services/rebalance/mod.rs`, `rustfs/src/admin/handlers/pools.rs`, `rustfs/src/admin/handlers/rebalance.rs`, `rustfs/src/admin/handlers/heal.rs`, `crates/heal/src/heal/manager/root_recovery.rs`, `crates/config/src/constants/heal.rs`.
@@ -0,0 +1,25 @@
+# FTPS uploads
+
+FTPS `STOR` buffers at most one 16 MiB payload chunk per active upload in the
+protocol driver. It waits for the storage backend to consume each chunk before
+reading the next. Backend, TLS, connection, and allocator overhead are additional;
+this is not a process-wide memory limit. Concurrent uploads each have their own
+buffer.
+
+Files smaller than 16 MiB, including empty files, use `PutObject`. Files at or
+above that threshold use sequential S3 multipart uploads, so they can exceed the
+5 GiB single-`PutObject` limit. Multipart ETags differ from single-PUT ETags and
+must not be interpreted as a whole-file MD5 checksum. An upload is successful only
+after the multipart completion succeeds. The fixed part size and S3's 10,000-part
+limit allow up to 156.25 GiB per FTPS upload; a larger input fails and cleanup is
+attempted. Resuming or appending with a nonzero offset remains unsupported.
+
+Each write operation requires `s3:PutObject`. Failed or cancelled transfers also
+attempt `s3:AbortMultipartUpload` using the authenticated user's permissions;
+cleanup does not bypass IAM. Grant that permission on the upload prefix to allow
+immediate cleanup. Cancellation cleanup has a bounded task count and timeout.
+Configure an `AbortIncompleteMultipartUpload` bucket lifecycle rule as a fallback
+for denied/failed cleanup, process crashes, or losing the upload ID while upload
+initiation is in flight. Before completion, received parts do not replace an existing completed object.
+A lost or failed completion response may have an ambiguous outcome; clients
+should verify the destination before retrying.
@@ -11,7 +11,7 @@ Every stage is env-gated, so "before" and "after" run the *same binary* with dif
 | --- | --- |
 | RPC secret | Internode RPC fails closed: remote endpoints with default credentials and no `RUSTFS_RPC_SECRET` abort startup with `store init aborted: endpoints include remote nodes but ...` (`crates/ecstore/src/store/init.rs`). Set a non-default `RUSTFS_RPC_SECRET`, identical on every node. |
 | systemd start timeout | `deploy/build/rustfs.service` is `Type=notify` and ships `TimeoutStartSec=120s`; READY fires only after quorum. If freshly purged disks need longer, raise it in a drop-in rather than lowering it. |
-| Metrics export | RustFS has no Prometheus pull endpoint (`/admin/v3/metrics` is NDJSON, see `rustfs/src/admin/handlers/metrics.rs`); it pushes OTLP. Run an otel-collector (OTLP receiver → Prometheus exporter) and set `RUSTFS_OBS_ENDPOINT`, `RUSTFS_OBS_METRICS_EXPORT_ENABLED=true`, `RUSTFS_OBS_METER_INTERVAL=5`. For lock p99 also set `RUSTFS_OBJECT_LOCK_DIAG_ENABLE=true` (default off). |
+| Metrics export | RustFS has no Prometheus pull endpoint (`/admin/v3/realtime` is NDJSON, see `rustfs/src/admin/handlers/realtime.rs`); it pushes OTLP. Run an otel-collector (OTLP receiver → Prometheus exporter) and set `RUSTFS_OBS_ENDPOINT`, `RUSTFS_OBS_METRICS_EXPORT_ENABLED=true`, `RUSTFS_OBS_METER_INTERVAL=5`. For lock p99 also set `RUSTFS_OBJECT_LOCK_DIAG_ENABLE=true` (default off). |
 | Server env | `RUSTFS_INTERNODE_*` are **server** env. For p0/p1/p2, source the emitted `server-env.sh` on every node and restart before the run; the driver cannot mutate a running server. |

 ## Driver
@@ -20,7 +20,7 @@ sum by (instance) (rate(rustfs_s3_http_requests_total{method="PUT",outcome=~"[1-

 Inspect `service_error` and `cancelled` separately; neither implies a received HTTP status. A zero denominator or absent series means no observed traffic, not proof of health. Use `rate` or reset-aware deltas because counters restart with the process. The older `rustfs_s3_operations_total` counter measures handler entries and excludes pre-dispatch rejections; it is not this HTTP denominator.

-The authenticated admin metrics endpoint exposes the same counters through the optional `http` field in `aggregated` and `by_host`. Request `/rustfs/admin/v3/metrics?types=512&by-host=true&n=1` on each node for HTTP-only data. The default type selection also includes HTTP outcomes. This endpoint remains an NDJSON stream and does not become a cluster-wide peer fanout. `http.requests` contains `method`, `operation`, `outcome`, and `total`; `http.collected` timestamps collection. Compare consecutive samples from the same host. Concurrent snapshots are not atomic across series.
+The authenticated admin metrics endpoint exposes the same counters through the optional `http` field in `aggregated` and `by_host`. Request `/rustfs/admin/v3/realtime?types=512&by-host=true&n=1` on each node for HTTP-only data. The default type selection also includes HTTP outcomes. This endpoint remains an NDJSON stream and does not become a cluster-wide peer fanout. `http.requests` contains `method`, `operation`, `outcome`, and `total`; `http.collected` timestamps collection. Compare consecutive samples from the same host. Concurrent snapshots are not atomic across series.

 A missing `http` field means an older or non-reporting node, not zero failures. Old map-encoded RPC readers ignore the additive field; new readers accept older snapshots. In mixed-version deployments, check reporting coverage before aggregating a fleet-wide ratio.

@@ -317,7 +317,7 @@ Use the harness around a post-start bucket creation workload to cover the timing
 1. Start RustFS from an empty data path.
 2. Start the harness before creating buckets.
 3. Create a bucket, upload objects, and keep the harness running until at least one usage save is observed.
-4. Compare `scanner-summary.csv` with `/rustfs/admin/v3/metrics?types=1&n=1` bucket metrics.
+4. Compare `scanner-summary.csv` with `/rustfs/admin/v3/realtime?types=1&n=1` bucket metrics.

 Expected evidence: dirty usage is marked, `life_time_scan_cycle` or `life_time_scan_bucket_drive` advances, `life_time_scan_object` advances for object workloads, and `life_time_save_usage` plus `usage_last_save_result=success` appear before non-zero bucket usage metrics are accepted as fresh.

@@ -481,7 +481,7 @@ Read this snapshot before changing scanner controls: `blocked_source` with `life

 ## Reading Distributed Metrics

-`/rustfs/admin/v3/scanner/status` and `/rustfs/admin/v3/metrics` report the node that handles the HTTP request; the metrics endpoint does not fan out to peers. In distributed deployments, query every node explicitly and keep `by-host=true` so each response includes that node's host view:
+`/rustfs/admin/v3/scanner/status` and `/rustfs/admin/v3/realtime` report the node that handles the HTTP request; the metrics endpoint does not fan out to peers. In distributed deployments, query every node explicitly and keep `by-host=true` so each response includes that node's host view:

 ```bash
 for endpoint in http://node-a:9000 http://node-b:9000 http://node-c:9000; do
@@ -493,7 +493,7 @@ for endpoint in http://node-a:9000 http://node-b:9000 http://node-c:9000; do
    --access_key "$RUSTFS_ACCESS_KEY" \
    --secret_key "$RUSTFS_SECRET_KEY" \
    --request GET \
-    "${endpoint}/rustfs/admin/v3/metrics?types=1&by-host=true&n=1" \
+    "${endpoint}/rustfs/admin/v3/realtime?types=1&by-host=true&n=1" \
    > "artifacts/scanner-metrics.${node}.$(date -u +%Y%m%dT%H%M%SZ).ndjson"
 done
 ```
@@ -0,0 +1,72 @@
+# Object integrity inventory, audit, and migration
+
+**Use this when:** assessing legacy object protection, checking a bounded manifest against an existing SHA256, or creating protected copies without overwriting the source.
+
+## Protection and activation
+
+Legacy shard-local checksums can detect accidental corruption, but do not independently bind a complete valid donor shard to its original object. A successful legacy read or Heal is not proof of original content. Inventory reports the stored protection declaration; it does not verify payload bytes. `independent_commitment` requires consistent object and part declarations. `unknown` means metadata could not be observed reliably.
+
+Both `RUSTFS_SHARD_INTEGRITY_WRITE` and `RUSTFS_SHARD_INTEGRITY_FLEET_CONFIRMED` remain false by default. Enabling both permits protected new writes on that process. The second flag is an operator attestation, not a peer capability probe or a storage fencing mechanism. Before enabling, qualify all readers, writers, repair processes, rollback binaries, and replacement nodes against the protected format. Prevent old processes from accessing those drives. Measure full GET, Range GET, PUT, degraded read, and repair latency and throughput on the deployment's erasure layout. The readiness endpoint reports these requirements as unverified; it does not make the deployment safe automatically.
+
+Disabling the write flags stops protection for new writes without an inherited mode; it does not remove existing commitments or make an old binary safe. Keep compatible readers for already migrated objects. Normal S3 requests and existing legacy reads do not depend on the integrity job worker being available.
+
+## Administrative API
+
+Use an authenticated, signed admin client. Paths use `/rustfs/admin/v3/integrity` and normal RustFS admin request authentication and body conventions.
+
+| Method and suffix | Permission | Behavior |
+|---|---|---|
+| `GET /readiness` | `admin:ServerInfo` | Local support, local flags, operator attestation, and unverified fleet requirements |
+| `GET /{bucket}/inventory` | `admin:InspectData` | One page of version observations, no payload scan |
+| `POST /{bucket}/jobs` | `admin:StartBatchJob` | Persist a paused manifest and return its ID |
+| `GET /{bucket}/jobs/{job_id}` | `admin:DescribeBatchJob` | Durable status and per-item outcomes |
+| `POST /{bucket}/jobs/{job_id}/control` | `admin:StartBatchJob` | Explicit `resume`, `pause`, or `cancel` |
+
+Inventory accepts `prefix`, `key-marker`, `version-marker`, and `limit` (1–100). Continue with both returned markers only after storing the entire page. Each inventory item returns a reusable string version selector, including `"null"` for the null version. Pages are observations, not a bucket snapshot: concurrent creates and deletes can affect enumeration. A recreated bucket has a different incarnation; do not combine observations across incarnations. Inventory and audit suppress their own read-repair submissions and bypass the body cache for content verification; they do not stop unrelated normal reads or background maintenance.
+
+Example job body:
+
+```json
+{
+  "mode": "audit",
+  "items": [
+    {
+      "key": "archive/object.bin",
+      "version_id": null,
+      "expected_sha256": "ungWv48Bz+pBQUDeXa4iI7ADYaOWF3qctBD/YfIAFa0="
+    }
+  ],
+  "bytes_per_second": 10485760,
+  "max_object_bytes": 1073741824
+}
+```
+
+The example digest is illustrative; supply the expected digest of your own object from an independent trusted record. `version_id` omitted or JSON null selects the current version; the string `"null"` selects the null version. A non-null version must be a UUID. The job records and rechecks the observed source identity. For `mode: "migrate"`, each item also requires a distinct `target_key` in the same bucket. No target may equal any source key in the manifest.
+
+`expected_sha256` is canonical Base64 of 32 bytes, not hex or an ETag. When omitted, only a canonical stored single-part SHA256 is eligible. A stored checksum verifies consistency with stored metadata; it does not establish an external historical provenance. The service never derives a new expected digest from an unchecked read and calls that historical evidence.
+
+Start the persisted job with:
+
+```json
+{"operation":"resume"}
+```
+
+Use the same control endpoint with `pause` or `cancel`. Poll status by ID and retain the ID externally; there is no job-list endpoint.
+
+## Supported scope and resource limits
+
+Audit accepts local, untransformed single-part objects with an expected SHA256. Multipart, encrypted, compressed, tiered objects, delete markers, and invalid protection declarations are reported as unsupported. Migration additionally requires a plain unversioned bucket without configured encryption, Object Lock, replication, notification, quota, lifecycle, table namespace, ACL, access logging, or on-demand migration. Disk commits continue to use the bucket’s existing durability policy. These checks hold the bucket configuration fence through publication so the direct storage writer cannot bypass those policies during a race.
+
+Each manifest contains 1–64 items. `max_object_bytes` is 1 byte–5 GiB and limits the anonymous local temporary file used for one object. Ensure that amount of space is available in the process temporary directory. One cluster-wide worker runs at a time. `bytes_per_second` is 64 KiB/s–1 GiB/s and bounds each sequential source read, destination PUT stream, and verification read. It is not a physical disk or network aggregate ceiling: erasure redundancy and internal buffering still apply. Metadata requests are not byte throttled. Lower rates extend the time a source read or destination publication holds its normal object and bucket configuration locks; choose the rate and per-object budget with concurrent traffic in mind.
+
+Migration uploads the same staged bytes that passed SHA256 verification, rechecks that checksum during PUT, and publishes with a create-only precondition. It preserves user metadata, content type/encoding/language, disposition, cache control, expiration, storage class, and tags. The new key has a new modification time and storage identity; consumers switch keys separately. The source is neither overwritten nor deleted. No original-object metadata is backfilled in place.
+
+## Recovery and interpretation
+
+`complete` means every item reached a final outcome, not that every item passed. Inspect `verified`, `migrated`, `mismatch`, `unsupported`, `stale`, and `conflict` individually. A source that is temporarily unavailable leaves the job `failed`; `resume` retries unavailable or prepared items while preserving completed results. A mismatch never publishes a target.
+
+Pause and cancel first persist intent. An in-flight conditional target publication may finish; terminal acknowledgement follows after the worker stops. Cancellation does not delete a successfully published target. After a coordinator restart, `running` is the last durable state and does not prove a worker is alive. Explicitly resume to reacquire the distributed worker lock and recover. A persisted cancel intent is completed rather than restarted. A busy response means another worker still owns the lock; retry later.
+
+A prepared checkpoint precedes destination publication. If the PUT acknowledgement or completion checkpoint is lost, resume first checks the destination's internal operation identity, protection declarations, size, and uncached full-content SHA256. Only a matching receipt is accepted as this job's completed copy. An unrelated destination is a conflict and is never overwritten. Source changes before verification finishes produce `stale`; create a new manifest after investigating the change. Bucket deletion/recreation invalidates the old job incarnation.
+
+This workflow does not automatically repair damaged historical content, certify all historical shards, enable protected writes fleet-wide, clean MRF queues, or authorize rolling back to an incompatible binary.
@@ -65,6 +65,8 @@ get_bytes(&self.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID)

 `transition_version_id == None` means only that no usable legacy UUID projection exists; it does not prove the remote bucket's versioning model. Only an explicit `KnownDisabled` state authorizes ordinary GET/DELETE to omit `versionId`. A missing state with an absent or empty version key remains `Unknown` and requires the bounded compatibility probe or the approved reconcile workflow; it never directly authorizes cleanup. A nil UUID (`00000000-...`) sent as `?versionId=` causes `NoSuchVersion`. Do not use `Uuid::from_slice(..).unwrap_or_default()` here: it converts an empty metadata value into `Uuid::nil()`, which is exactly that failure.

+Opaque remote version IDs, such as those returned by Backblaze B2, are treated as exact remote identifiers only after the operator-attested live fleet capability gate is active. The gate requires both `RUSTFS_TIER_REMOTE_VERSION_STATE_WRITE=true` and `RUSTFS_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED=true`, plus a current fleet proof. Recovery and operator cleanup retain an ambiguous candidate while that proof is absent or stale; they never issue a guessed opaque-version delete. Existing persisted exact opaque IDs remain routable, but new persistence and recovery cleanup fail closed until the gate is current.
+
 ## Inspect xl.meta directly

 ```bash
--- a/Show More
+++ b/Show More