fix(storage): preserve retryable publication guard errors

Merge branch 'main' into release
* main: fix(rebalance): align activation locks and preserve retryable causes (#7551)
2026-09-09 05:36:24 +00:00 · 2026-09-09 13:24:39 +08:00 · 2026-09-09 12:32:55 +08:00 · 2026-09-09 12:08:06 +08:00 · 2026-09-09 10:26:59 +08:00 · 2026-09-09 10:22:13 +08:00
135 changed files with 17533 additions and 709 deletions
@@ -1,2 +1,2 @@
-sha256-linux=4696a43b167ac608b3b8677027c9fe9fdac3396d37c8cca11dce531c720ac6d2
-sha256-darwin=9785867929047dfd8c6f768e0d2b1e0a8fdba85216f4a4139093b1619d03ff07
+sha256-linux=563bff8f1171d6dbe166ff8440310dbe98430e466aa3ecd8dc39e3c872b320f7
+sha256-darwin=563bff8f1171d6dbe166ff8440310dbe98430e466aa3ecd8dc39e3c872b320f7
@@ -1,2 +1,2 @@
-sha256-darwin=f0c78fdb93471575d9a64c5c46eae6c806bdd0bc10a6e33d7fb574aabd8db5a3
-sha256-linux=03ed7016cab672de9320e31375a0358eceacb4408b0e79cf063614fa7c878b87
+sha256-darwin=874c881d7b45f12378a5817c7f42c95c4981960a2ec9ce12dcf4af239ae1f9d5
+sha256-linux=9515861be899ceb10e2e0ef93c34208bb7a7a8a7f8067a02db4cfba23270ebd6
@@ -1,2 +1,2 @@
-sha256-darwin=364f2329a7b72eb9f1608dbe1a3af37af4095354014f3cbe23ca448492d89961
-sha256-linux=60983f1ebe7068cf660d473c5f76c76a650410ccc99d71934ddca7fd67607987
+sha256-darwin=83a7dcaffd5a789517ae9f02a224f66a9713937885cff96fca2ad7e216f197ae
+sha256-linux=626c10f8c964507ff987b6c86069e9019dc6d2ae7fb02db9be5df5aa8cc5145b
@@ -32,6 +32,8 @@ script-tests: ## Run shell script tests
 	./scripts/test_hotpath_warp_ab_gate.sh
 	./scripts/test_hotpath_warp_abba.sh
 	./scripts/test_scanner_validation_harness.sh
+	./scripts/test_scanner_heal_w13_mrf_evidence.sh
+	./scripts/test_scanner_heal_w16_recovery_evidence.sh
 	./scripts/test_exact_1mib_handoff_abba.sh
 	./scripts/test_pinned_paired_abba_bench.sh
 	./scripts/test_manual_transition_runbooks.sh
@@ -197,12 +197,6 @@ test-group = 'e2e-cluster-nightly'
 filter = 'package(e2e_test) & (test(/^kms::kms_vault_test::/) | test(/^kms::kms_rekey_sweep_test::/) | test(/^kms::configured_roundtrip_test::test_configured_vault_kms_admin_and_versioned_cleanup$/))'
 test-group = 'e2e-vault'

-# This four-disk, 65-member rollback probe already drives up to 32 concurrent
-# durable deletions. Reserve this nextest run's capacity for its progress oracle.
-[[profile.default.overrides]]
-filter = 'package(rustfs-ecstore) & test(=store::init::tests::dispatch_manifest_rollback_bounded_concurrency_reaches_tail_behind_slow_member)'
-threads-required = "num-test-threads"
-
 # ---------------------------------------------------------------------------
 # ci profile — the strict CI gate (ci.yml `cargo nextest run --profile ci`)
 # ---------------------------------------------------------------------------
@@ -1,5 +1,5 @@
 {
-  "schema": 1,
+  "schema": 2,
  "cases": {
    "background-target-restart": {
      "gate": "G14",
@@ -13,6 +13,7 @@
      "min_objects": 9,
      "max_objects": 65,
      "topology": {"nodes": 4, "drives_per_node": 1},
+      "erasure": {"data_blocks": 2, "parity_blocks": 2},
      "scope": "Target process restart, exact unversioned S3 bodies and replacement-disk shards; not power loss or EC8+4."
    },
    "background-target-crash": {
@@ -27,30 +28,325 @@
      "min_objects": 9,
      "max_objects": 65,
      "topology": {"nodes": 4, "drives_per_node": 1},
+      "erasure": {"data_blocks": 2, "parity_blocks": 2},
      "scope": "Target process killed during partial background rebuild, real unclean-shutdown marker, exact unversioned S3 bodies and replacement-disk shards; not power loss or EC8+4."
+    },
+    "ec84-target-drive-restart": {
+      "gate": "G14",
+      "task": "W20/W21",
+      "lane": "e2e-distributed",
+      "suite": "e2e_test",
+      "name": "distributed::heal_test::three_node_four_drive_ec8_4_root_heal_rebuilds_replaced_drive_after_restart",
+      "oracle": "ec84-target-drive-restart.json",
+      "evidence": "process-restart",
+      "unclean_shutdown_marker": false,
+      "min_objects": 5,
+      "max_objects": 5,
+      "topology": {"nodes": 3, "drives_per_node": 4},
+      "erasure": {"data_blocks": 8, "parity_blocks": 4},
+      "erasure_set_drive_count": 12,
+      "scope": "3-node x 4-drive single-set EC8+4, graceful target restart, preformatted replacement drive, exact unversioned S3 bodies and physical target shards; not mixed-version, multi-pool or long-window ABBA."
+    },
+    "background-target-restart-ec8-4": {
+      "gate": "G14",
+      "task": "W21",
+      "lane": "e2e-nightly",
+      "suite": "e2e_test",
+      "name": "heal_erasure_disk_rebuild_test::tests::test_cluster_root_heal_recovers_ec84_shards_after_background_target_restart",
+      "oracle": "background-target-restart-ec8-4.json",
+      "evidence": "process-restart",
+      "unclean_shutdown_marker": false,
+      "min_objects": 9,
+      "max_objects": 65,
+      "topology": {"nodes": 3, "drives_per_node": 4},
+      "erasure": {"data_blocks": 8, "parity_blocks": 4},
+      "erasure_set_drive_count": 12,
+      "scope": "Target process restart during partial background rebuild on a single 3x4 EC8+4 set; exact unversioned S3 bodies and replacement-drive shards; not power loss, multi-set, or multi-pool."
+    },
+    "background-target-crash-ec8-4": {
+      "gate": "G14",
+      "task": "W21",
+      "lane": "e2e-nightly",
+      "suite": "e2e_test",
+      "name": "heal_erasure_disk_rebuild_test::tests::test_cluster_root_heal_recovers_ec84_shards_after_background_target_crash",
+      "oracle": "background-target-crash-ec8-4.json",
+      "evidence": "process-crash-restart",
+      "unclean_shutdown_marker": true,
+      "min_objects": 9,
+      "max_objects": 65,
+      "topology": {"nodes": 3, "drives_per_node": 4},
+      "erasure": {"data_blocks": 8, "parity_blocks": 4},
+      "erasure_set_drive_count": 12,
+      "scope": "Target process killed during partial background rebuild on a single 3x4 EC8+4 set; real unclean-shutdown marker, exact unversioned S3 bodies and replacement-drive shards; not power loss, multi-set, or multi-pool."
    }
  },
-  "release_pending": {
-    "G01": "W02/W04 complete root and quota authority coverage",
-    "G02": "W03 bounded checkpoint progress and independent version inventory",
-    "G03": "W17/W18 exact scoped ACK with durable publication and mixed peers",
-    "G04": "W03/W15/W16 crash at every cache/root/floor/intent boundary",
-    "G05": "W06/W07 per-object outcomes and bounded terminal retention",
-    "G06": "W06/W08/W23 concurrent status, legacy clients and truncation",
-    "G07": "W12/W13/W14 durable MRF responsibility at every commit boundary",
-    "G08": "W12/W13/W14 MRF capacity, disk-full and replica-loss matrix",
-    "G09": "W13/W18/W23 actual mixed-version reader/writer and rollback payloads",
-    "G10": "W05/W09/W10/W11 bounded scheduling and pressure recovery",
-    "G11": "W04/W19/W24 maintenance and complete producer coverage",
-    "G12": "W02/W15/W16 both quota paths during reset and settlement",
-    "G13": "W07/W14 quorum-minus-one, unknown disks, remount, Object Lock, dry-run, grace and commit tail",
-    "G14": "W20/W21 same-window field evidence; 3x4 EC8+4 and multi-set/pool coverage",
-    "P1": "W20 measured cold-walk share and foreground latency/throughput",
-    "P2": "W20/W24 measured post-stop convergence and cold segment reuse",
-    "P3": "W20 measured two-hour pressure/heal capacity and recovery window",
-    "P4": "W20 measured MRF scale and replay cost with retained responsibility",
-    "R-E": "W03/W05 fixed-budget real process restart through enumeration and classification",
-    "R-D": "W07/W14 manager-to-event-to-ledger exact disposition, including grace",
-    "R-L": "W13/W14 legacy source conflicts, migration gaps and crash-safe source retirement"
-  }
+  "release_lanes": {
+    "single-set-restart": {
+      "status": "implemented",
+      "cases": ["background-target-restart", "background-target-crash"],
+      "covers": ["four-node one-drive topology", "unversioned objects", "target restart/crash"]
+    },
+    "authority-coverage": {
+      "status": "pending",
+      "gates": ["G01", "G12"],
+      "requires": ["root authority coverage", "quota authority coverage"]
+    },
+    "checkpoint-and-crash": {
+      "status": "pending",
+      "gates": ["G02", "G04", "R-E"],
+      "requires": ["bounded checkpoint progress", "boundary crash matrix", "fixed-budget restart evidence"]
+    },
+    "status-and-outcome": {
+      "status": "pending",
+      "gates": ["G05", "G06", "R-D"],
+      "requires": ["per-object outcomes", "legacy status clients", "manager/event/ledger disposition"]
+    },
+    "mrf-responsibility": {
+      "status": "pending",
+      "gates": ["G07", "G08", "P4"],
+      "requires": ["durable MRF responsibility", "disk-full and replica-loss matrix", "MRF replay cost"]
+    },
+    "mixed-version-rollback": {
+      "status": "pending",
+      "gates": ["G03", "G09", "R-L"],
+      "requires": ["mixed-version peers", "rollback payloads", "crash-safe source retirement"]
+    },
+    "scheduler-pressure": {
+      "status": "pending",
+      "gates": ["G10", "P1", "P2", "P3"],
+      "requires": ["bounded scheduling", "foreground latency and throughput", "two-hour pressure evidence"]
+    },
+    "maintenance-producers": {
+      "status": "pending",
+      "gates": ["G11", "G13"],
+      "requires": ["complete producer coverage", "quorum-minus-one and remount matrix"]
+    },
+    "ec8-4-multiset": {
+      "status": "pending",
+      "gates": ["G14"],
+      "requires": ["3x4 EC8+4 topology", "multi-set coverage", "multi-pool coverage"]
+    }
+  },
+  "release_requirements": [
+    {
+      "gate": "G01",
+      "task": "W02/W04",
+      "lane": "authority-coverage",
+      "status": "pending",
+      "description": "Complete root and quota authority coverage",
+      "requires": ["root authority evidence", "quota authority evidence"]
+    },
+    {
+      "gate": "G02",
+      "task": "W03",
+      "lane": "checkpoint-and-crash",
+      "status": "pending",
+      "description": "Bounded checkpoint progress and independent version inventory",
+      "requires": ["bounded checkpoint oracle", "independent version inventory"]
+    },
+    {
+      "gate": "G03",
+      "task": "W17/W18",
+      "lane": "mixed-version-rollback",
+      "status": "pending",
+      "description": "Exact scoped ACK with durable publication and mixed peers",
+      "requires": ["durable scoped ACK publication", "mixed-peer evidence"],
+      "evidence_fields": [
+        "durable_root_publication_proof",
+        "scoped_ack_request_identity",
+        "participating_peer_capability_snapshot",
+        "mixed_peer_ack_fallback_oracle"
+      ]
+    },
+    {
+      "gate": "G04",
+      "task": "W03/W15/W16",
+      "lane": "checkpoint-and-crash",
+      "status": "pending",
+      "description": "Crash at every cache, root, floor and intent boundary",
+      "requires": ["cache boundary crash evidence", "root/floor/intent crash evidence"]
+    },
+    {
+      "gate": "G05",
+      "task": "W06/W07",
+      "lane": "status-and-outcome",
+      "status": "pending",
+      "description": "Per-object outcomes and bounded terminal retention",
+      "requires": ["per-object outcome oracle", "terminal retention bounds"]
+    },
+    {
+      "gate": "G06",
+      "task": "W06/W08/W23",
+      "lane": "status-and-outcome",
+      "status": "pending",
+      "description": "Concurrent status, legacy clients and truncation",
+      "requires": ["concurrent status evidence", "legacy client compatibility", "truncation behavior"]
+    },
+    {
+      "gate": "G07",
+      "task": "W12/W13/W14",
+      "lane": "mrf-responsibility",
+      "status": "pending",
+      "description": "Durable MRF responsibility at every commit boundary",
+      "requires": ["MRF responsibility oracle", "commit-boundary crash matrix"],
+      "evidence_fields": [
+        "mrf_responsibility_oracle",
+        "commit_boundary_crash_matrix"
+      ]
+    },
+    {
+      "gate": "G08",
+      "task": "W12/W13/W14",
+      "lane": "mrf-responsibility",
+      "status": "pending",
+      "description": "MRF capacity, disk-full and replica-loss matrix",
+      "requires": ["MRF capacity evidence", "disk-full matrix", "replica-loss matrix"],
+      "evidence_fields": [
+        "mrf_capacity_evidence",
+        "disk_full_matrix",
+        "replica_loss_matrix"
+      ]
+    },
+    {
+      "gate": "G09",
+      "task": "W13/W18/W23",
+      "lane": "mixed-version-rollback",
+      "status": "pending",
+      "description": "Actual mixed-version reader/writer and rollback payloads",
+      "requires": ["mixed-version reader evidence", "mixed-version writer evidence", "rollback payload evidence"],
+      "evidence_fields": [
+        "mixed_version_reader_evidence",
+        "mixed_version_writer_evidence",
+        "rollback_payload_evidence"
+      ]
+    },
+    {
+      "gate": "G10",
+      "task": "W05/W09/W10/W11",
+      "lane": "scheduler-pressure",
+      "status": "pending",
+      "description": "Bounded scheduling and pressure recovery",
+      "requires": ["scheduler bound evidence", "pressure recovery evidence"]
+    },
+    {
+      "gate": "G11",
+      "task": "W04/W19/W24",
+      "lane": "maintenance-producers",
+      "status": "pending",
+      "description": "Maintenance and complete producer coverage",
+      "requires": [
+        "maintenance producer matrix",
+        "complete producer inventory",
+        "segment activation preflight"
+      ],
+      "evidence_fields": [
+        "maintenance_producer_matrix",
+        "complete_producer_inventory",
+        "segment_activation_preflight"
+      ]
+    },
+    {
+      "gate": "G12",
+      "task": "W02/W15/W16",
+      "lane": "authority-coverage",
+      "status": "pending",
+      "description": "Both quota paths during reset and settlement",
+      "requires": ["reset quota-path evidence", "settlement quota-path evidence"]
+    },
+    {
+      "gate": "G13",
+      "task": "W07/W14",
+      "lane": "maintenance-producers",
+      "status": "pending",
+      "description": "Quorum-minus-one, unknown disks, remount, Object Lock, dry-run, grace and commit tail",
+      "requires": ["quorum-minus-one matrix", "unknown-disk/remount matrix", "Object Lock dry-run grace evidence"]
+    },
+    {
+      "gate": "G14",
+      "task": "W20/W21",
+      "lane": "ec8-4-multiset",
+      "status": "pending",
+      "description": "Same-window field evidence with 3x4 EC8+4 and multi-set/pool coverage",
+      "requires": [
+        "same-window field evidence",
+        "3x4 EC8+4 evidence",
+        "multi-set evidence",
+        "multi-pool evidence",
+        "distributed segment invalidation evidence"
+      ],
+      "evidence_fields": [
+        "same_window_field_evidence",
+        "ec8_4_evidence",
+        "multi_set_evidence",
+        "multi_pool_evidence",
+        "distributed_segment_invalidation_evidence"
+      ]
+    },
+    {
+      "gate": "P1",
+      "task": "W20",
+      "lane": "scheduler-pressure",
+      "status": "pending",
+      "description": "Measured cold-walk share and foreground latency/throughput",
+      "requires": ["cold-walk share measurement", "foreground latency/throughput measurement"]
+    },
+    {
+      "gate": "P2",
+      "task": "W20/W24",
+      "lane": "scheduler-pressure",
+      "status": "pending",
+      "description": "Measured post-stop convergence and cold segment reuse",
+      "requires": ["post-stop convergence measurement", "cold segment reuse measurement"],
+      "evidence_fields": [
+        "post_stop_convergence_measurement",
+        "cold_segment_reuse_measurement"
+      ]
+    },
+    {
+      "gate": "P3",
+      "task": "W20",
+      "lane": "scheduler-pressure",
+      "status": "pending",
+      "description": "Measured two-hour pressure/heal capacity and recovery window",
+      "requires": ["two-hour pressure measurement", "heal capacity measurement", "recovery-window measurement"]
+    },
+    {
+      "gate": "P4",
+      "task": "W20",
+      "lane": "mrf-responsibility",
+      "status": "pending",
+      "description": "Measured MRF scale and replay cost with retained responsibility",
+      "requires": ["MRF scale measurement", "MRF replay-cost measurement", "retained responsibility evidence", "cleanup/GC soak evidence"],
+      "evidence_fields": [
+        "mrf_scale_measurement",
+        "mrf_replay_cost_measurement",
+        "retained_responsibility_evidence",
+        "mrf_cleanup_gc_soak_evidence"
+      ]
+    },
+    {
+      "gate": "R-E",
+      "task": "W03/W05",
+      "lane": "checkpoint-and-crash",
+      "status": "pending",
+      "description": "Fixed-budget real process restart through enumeration and classification",
+      "requires": ["fixed-budget restart evidence", "enumeration evidence", "classification evidence"]
+    },
+    {
+      "gate": "R-D",
+      "task": "W07/W14",
+      "lane": "status-and-outcome",
+      "status": "pending",
+      "description": "Manager-to-event-to-ledger exact disposition, including grace",
+      "requires": ["manager disposition evidence", "event disposition evidence", "ledger disposition evidence", "grace handling"]
+    },
+    {
+      "gate": "R-L",
+      "task": "W13/W14",
+      "lane": "mixed-version-rollback",
+      "status": "pending",
+      "description": "Legacy source conflicts, migration gaps and crash-safe source retirement",
+      "requires": ["legacy source-conflict evidence", "migration-gap evidence", "crash-safe source retirement evidence"]
+    }
+  ]
 }
@@ -16,7 +16,7 @@ name: Continuous Integration

 on:
  push:
-    branches: [ main ]
+    branches: [ main, release ]
    paths-ignore:
      - "**.md"
      - "docs/**"
@@ -36,7 +36,7 @@ on:
      - "flake.lock"
  pull_request:
    types: [ opened, synchronize, reopened, closed ]
-    branches: [ main ]
+    branches: [ main, release ]
  merge_group:
    types: [ checks_requested ]
  schedule:
@@ -840,13 +840,14 @@ jobs:
    # Merge gate only (backlog#1149 ci-5): the never-automated user-visible
    # suites — KMS, object_lock, multipart_auth, quota, checksum, encryption,
    # security-boundary, ... — via the e2e-full nextest profile. Too heavy for
-    # every PR, so it is gated to main pushes, the merge queue, and manual
+    # every PR, so it is gated to main/release pushes, the merge queue, and manual
    # dispatch. protocols / the 7 cluster suites / replication / #[ignore] are
    # owned by other lanes (see .config/nextest.toml profile.e2e-full).
    if: >-
      github.event_name == 'workflow_dispatch' ||
      github.event_name == 'merge_group' ||
-      (github.event_name == 'push' && github.ref == 'refs/heads/main')
+      (github.event_name == 'push' &&
+        (github.ref == 'refs/heads/main' || github.ref == 'refs/heads/release'))
    needs: [ build-rustfs-debug-binary ]
    runs-on: sm-standard-2
    timeout-minutes: 55
@@ -129,11 +129,22 @@ jobs:
          : > target/debug/rustfs.features

      - name: Run upgrade compatibility test
+        env:
+          RUSTFS_SCANNER_HEAL_G09_EVIDENCE_DIR: ${{ runner.temp }}/rustfs-upgrade-g09-evidence/${{ matrix.artifact }}
        run: |
          cargo test --locked -p e2e_test \
            "upgrade_compatibility_test::${{ matrix.test }}" \
            -- --ignored --exact --nocapture

+      - name: Upload scanner/heal G09 evidence
+        if: always()
+        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
+        with:
+          name: ${{ matrix.artifact }}-g09-evidence-${{ github.run_number }}
+          path: ${{ runner.temp }}/rustfs-upgrade-g09-evidence/${{ matrix.artifact }}
+          if-no-files-found: ignore
+          retention-days: 14
+
      - name: Upload server logs
        if: always()
        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
@@ -23,6 +23,14 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ### Fixed
 - **Fresh multi-pool bootstrap with distinct format creators**: a new deployment whose pools have their first endpoint on different nodes (for example two single-node pools) could never publish its initial `pool.bin`: each node held fresh-bootstrap proof only for the pool it formatted, the deployment-wide proof collapsed to none, and every node died with `pool metadata recovery required: no durable bootstrap identity or pool.bin replica is available` after the startup retry budget. The first pool's creator now mints the pending cluster identity on its own pool, every other creator copies that nonce-bound identity onto the pool it formatted first-hand, and the elected writer publishes `pool.bin` once every pool replica carries the same pending identity. Corrupt or disagreeing replicas, pools that merely have a format, expansion pools joining an initialized deployment, and restarts without first-hand proof still fail closed. Non-elected nodes that start before `pool.bin` exists, and the elected writer while it waits for the other creators, no longer latch their pool-metadata write gate for the life of the process. Refs rustfs/backlog#2338, rustfs/backlog#2375.
 - **Lock RPC timeout storms** (#7363): the remote lock client no longer evicts and re-dials the shared internode HTTP/2 channel on every request deadline. A timeout evicts only when the peer has not completed any lock RPC for two deadlines, evictions and transport-failure re-dials are rate limited per peer (`RUSTFS_OBJECT_LOCK_RPC_EVICTION_COOLDOWN_MS`, default 5 s), and a timed-out request is left running instead of being reset (bounded per peer by `RUSTFS_OBJECT_LOCK_RPC_DETACHED_LIMIT`, default 256), so a slow lock endpoint can no longer drive the `RST_STREAM`/`GOAWAY too_many_resets`/reconnect loop. A lock granted after its caller timed out is released immediately, and unlocks that fail the quick retries continue on a deferred 1/2/4/8/16 s schedule before the server lease reclaims them. New `rustfs_remote_lock_*` metrics cover timeouts, evictions, suppressed evictions, detached streams, late completions and late releases per peer. Operator guide at `docs/operations/lock-rpc-storm-protection.md`.
+- **KMS failures on the S3 data path carry an actionable status**: only "key not found" and a backend outage were classified; every other KMS failure — a disabled or pending-deletion key, a denied KMS grant, an encryption-context mismatch, an unsupported algorithm, a credential or timeout failure, a capability the backend does not have — collapsed onto `500 InternalError`. SDKs therefore applied exponential backoff to configuration errors that no retry can fix, and monitoring filed every one of them as a server fault. Unusable-key and request-side failures now return `400`, a denied grant `403`, transient backend failures `503` — including a key store the backend could not read, so an outage stays distinguishable from a missing key all the way to the client — and a missing backend capability `501`. Damaged or unreadable key material still returns `500`, which is what it is.
+- **SSE-C on buckets with default encryption**: a `PutObject` carrying a valid SSE-C header triple on a bucket that has default encryption configured no longer fails with `400 InvalidArgument` ("The SSE-C and managed server-side encryption headers cannot be used together"). PUT and the POST-object/extract path resolved the bucket default with a hard-coded "no explicit SSE-C" flag, so the default was layered onto the request and then tripped the request's own mutual-exclusion check; an SSE-C request now suppresses the bucket default on all three write paths, matching COPY and AWS S3. Every bucket with default encryption previously refused SSE-C single PUTs outright, while `CreateMultipartUpload` on the same bucket succeeded.
+- **Explicit SSE-S3 on SSE-KMS-default buckets**: `x-amz-server-side-encryption: AES256` against a bucket whose default is `aws:kms` no longer fails with `400 InvalidArgument`. The bucket default's KMS key id was inherited independently of the effective algorithm, producing a self-contradictory `AES256` + key-id pair; the key id is now inherited only when the effective algorithm is `aws:kms`. `PutBucketEncryption` fills in a default key id automatically, so this affected nearly every SSE-KMS-default bucket.
+- **Restore of encrypted or compressed multipart objects (silent data corruption)**: restoring a multipart object from a remote tier addressed the tier in *plaintext* coordinates while the copy-back reads the *stored* representation. Every part received a misaligned slice of the remote object whose length still satisfied the range, the hash reader and the completion size check, so the restore reported success and replaced the object's bytes. Restore now accumulates stored part sizes, passes the stored length to the hash reader alongside the plaintext length, and validates against the stored size. Objects restored by an affected release must be re-restored from the tier or recovered from a backup — this release does not detect or repair them retroactively.
+- **Restore no longer drifts the object ETag**: the copy-back digests stored (encrypted or compressed) bytes, so the recomputed MD5 is not the object's public ETag. Single-part and multipart restores now preserve the original object ETag, and each restored part keeps its own recorded part ETag.
+- **ILM archive no longer forwards encryption metadata to the tier**: transition requests carried the object's SSE headers and the RustFS-wrapped data key as request headers. Any S3 target rejected an SSE-C archive outright (`400`, no key supplied), an SSE-KMS archive asked the target to encrypt a second time under a key id it does not own, and the wrapped DEK left the cluster. The archive request now strips every SSE header and encryption marker using the same predicate the replication path uses; the local `xl.meta` keeps all of it, so read-through and restore are unaffected.
+- **KMS reload is no longer a no-op on a node whose KMS failed to start**: `POST /rustfs/admin/v3/kms/reload` short-circuited whenever the persisted configuration matched the in-memory one byte for byte. A node whose KMS failed to start (for example Vault briefly unreachable during a rolling restart) keeps that configuration and sits in `Error`, so the documented recovery call returned "reloaded successfully" while leaving the node down — and did the same on every peer through the reload broadcast. Reload now short-circuits only for a service that is actually running, and otherwise reconfigures, which starts the service.
+- **AWS KMS capability reporting**: the AWS backend no longer advertises `versioning` support through `GET /rustfs/admin/v3/kms/status`. AWS KMS key versions are not enumerable through this backend, as the backend documentation already stated.
 - **Multipart admission queue**: an `UploadPart` waiting for a foreground write permit now waits at most 10 s by default (`RUSTFS_PUT_MULTIPART_FOREGROUND_ADMISSION_WAIT_TIMEOUT_MS`, previously 30 s), so a queued part returns S3 `SlowDown` before the client's socket write timeout drops the connection. Separately, the API listener no longer forces a 4 MiB `SO_RCVBUF` on every accepted socket (kernel autotuning applies; `RUSTFS_HTTP_SOCKET_RECV_BUFFER_BYTES` restores a fixed size), so a queued part no longer lets up to 8 MiB of unread body accumulate in kernel memory per connection, which is what throttled whole nodes under SDK-default multipart concurrency. Fixes #7385.
 - **Helm Ingress**: `customAnnotations` are now merged with class-specific annotations (nginx/traefik) instead of being ignored when `ingress.className` is set.
 - **Per-pool erasure parity**: Erasure parity (STANDARD and reduced-redundancy) is now resolved independently for every pool instead of reusing the first pool's value. A heterogeneous topology — for example a 4-drive pool plus a 2-drive pool created during expansion — previously inherited the first pool's parity and could resolve to zero data shards in the smaller pool, panicking Reed-Solomon construction on write. Automatic parity now resolves per pool (for example `2+2` in the 4-drive pool and `1+1` in the 2-drive pool). Fixes #4801.
@@ -64,6 +72,18 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
  - Four-layer regression-prevention tests guard against silent feature deletion: compile-time module assertion, module-presence unit test, cross-module `Protocol` enum assertion, end-to-end SSH banner test against the running binary

 ### Changed
+- **Encryption and KMS work merged since `1.0.0-rc.5`** (entries were missing from this section):
+  - **Persisted KMS configuration secrets** are sealed field-by-field with `RUSTFS_KMS_CONFIG_SECRET`. **When the variable is unset the secrets are persisted in cleartext and the server only warns** (`persisted KMS configuration carries cleartext secrets`); it never refuses the write. Set it, identically, on every node, and re-save the configuration to seal an existing one.
+  - **New v2 ciphertext frame format** with per-frame index binding and final-frame authentication. Its write switch `RUSTFS_ENCRYPTION_FRAME_V2` is **off by default**: v2 frames are unreadable by nodes without v2 read support, and encrypted ciphertext travels verbatim through transition, decommission and SSE-C replication passthrough, so turn it on only after every node — and every RustFS warm/replication target that receives raw ciphertext — runs a release with v2 read support. Reading v2 objects needs no switch.
+  - **Per-key SSE-KMS authorization** (`RUSTFS_KMS_ENFORCE_SSE_KEY_POLICY`, default `false`). With it on, anonymous callers hold no KMS grants, so **a public bucket serving SSE-KMS objects is an incompatible combination** and those reads return `AccessDenied`.
+  - Envelope context binding as KMS AAD (`ENV_KMS_ENVELOPE_AAD`, off by default; a node that predates the field cannot open bound envelopes).
+  - Vault custom CA and mutual TLS; object-level DEK rewrap plus a batch rekey admin API; a backend-locality runtime signal on `kms/status`.
+  - Single-pass decryption for encrypted GET, and encrypted single-part closed-range seek — the latter is now **on by default** (`RUSTFS_ENCRYPTED_RANGE_SEEK`, default `true`; the switch remains as a kill switch).
+- **Vault static tokens are now tracked and renewed**: with `Token` authentication RustFS hard-coded "this token has no lease", so the renewal task never started and no remaining-TTL gauge was published. `vault token create` grants a 768-hour TTL by default, which turned a healthy-looking cluster into one where every KMS call returned 403 about a month later, with no self-healing short of a restart or reconfigure. RustFS now calls `auth/token/lookup-self` at login and adopts what Vault reports: a non-expiring token behaves exactly as before, an expiring renewable one is renewed at half TTL like the other auth methods, and an expiring non-renewable one logs `vault_static_token_not_renewable` and publishes its remaining TTL. The probe never fails the login: a token whose policy omits `lookup-self` (Vault's `default` policy grants it), or a Vault that is unreachable at that moment, logs `vault_static_token_lookup_failed` and falls back to the previous no-lease behaviour, so no deployment that works today stops working.
+- **SSE-C over a plaintext transport is reported**: AWS S3 and MinIO refuse an SSE-C request that did not arrive over TLS, because the customer key travels in a request header. RustFS accepted them on any transport and still does by default — flipping to a rejection inside a release window would break plaintext staging and test deployments. Each such request now increments `rustfs_ssec_plaintext_requests_total` and logs one `ssec_request_without_tls` warning per process, and `RUSTFS_SSE_C_REQUIRE_TLS=true` opts into the AWS `400` now. The default is expected to flip in a later release; confirm the counter reads zero first. The verdict is per connection: a TLS listener satisfies it, and so does an `https` protocol forwarded by a proxy the trusted-proxy configuration accepts.
+- **Local KMS backend on a distributed deployment says what actually breaks**: the backend keeps key material and its Argon2id salt on each node's own disk, so two nodes derive different keys from the same `master_key` and an object encrypted on one node cannot be decrypted on another — intermittent 500s behind a load balancer. Configuring it while the deployment is distributed now logs `kms_node_local_backend_in_distributed_deployment` and appends that consequence to the `kms/configure` response, instead of only the generic "development only" positioning warning. It remains a warning, not a gate.
+- **SSE-KMS is refused when no KMS is running (breaking)**: a write requesting `x-amz-server-side-encryption: aws:kms` on a node with no KMS service no longer succeeds. Earlier releases wrapped the data key with the node-local `RUSTFS_SSE_S3_MASTER_KEY` while still writing `aws:kms` and the requested key id into the object metadata — metadata that claimed a KMS protection the object never had, under a key that was never consulted. Such a request now returns `400 InvalidRequest` when KMS was never configured and `503` when a configured service is not running; the refusal is evaluated after the per-key authorization gate, so an unauthorized caller still receives `403 AccessDenied`. **Upgrade note:** a deployment that relied on this write succeeding will start receiving 4xx/503. Either configure a KMS, or request `AES256` and keep the documented SSE-S3 local-master-key fallback, which is unchanged. Objects already written this way remain readable.
+- **Legacy ciphertext nonce layouts are now locked per segment**: while decrypting a v1 segment, the reader locks onto whichever of the three historical nonce layouts decoded the segment's first non-zero-index frame and rejects any later frame that needs a different one. Because a frame encrypted at block index zero authenticates under the pre-`1.0.0-alpha.91` reused-part-nonce layout at any position, an attacker able to rewrite the underlying shards could previously replay it and have the forged plaintext returned with `200`. New `RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK` (default `true`) drops that third layout entirely when set to `false`, which closes the residual case of a stream built purely from repeats of frame zero. Turn it off only after migrating pre-alpha.91 encrypted objects (rewrite in place with CopyObject); see [KMS backend security properties](docs/operations/kms-backend-security.md) for what the v1 frame layout does and does not authenticate.
 - **HTTP Server Stack**: Integrated `KeystoneAuthLayer` middleware from `rustfs-keystone` crate into service stack (positioned after ReadinessGateLayer)
 - **Storage-class validation on startup (upgrade note)**: A persisted explicit storage class (`RUSTFS_STORAGE_CLASS_STANDARD` / `RUSTFS_STORAGE_CLASS_RRS`, for example `EC:2`) is now validated against the actual per-pool drive counts at startup and rejected when a pool cannot satisfy it. This is fail-closed and correct, but a cluster that persisted a storage class larger than a small or heterogeneous pool can hold (for example `EC:2` alongside a 2-drive pool), which earlier releases accepted and silently resolved to an invalid layout, will now refuse to start after upgrade. To recover, unset `RUSTFS_STORAGE_CLASS_STANDARD` so the server derives a valid per-pool default automatically, or set it to a value every pool can satisfy.
 - **IAMAuth**: Enhanced `get_secret_key()` to return empty secret for Keystone credentials (bypasses signature validation)
@@ -9928,6 +9928,7 @@ dependencies = [
 "async-trait",
 "base64-simd",
 "bytes",
+ "chrono",
 "crc-fast",
 "futures",
 "hotpath",
@@ -10578,6 +10579,7 @@ dependencies = [
 "serde_json",
 "sha1 0.11.0",
 "sha2 0.11.0",
+ "temp-env",
 "thiserror 2.0.20",
 "tokio",
 "tokio-test",
@@ -46,14 +46,31 @@ RustFS 是一个基于 Rust 构建的高性能分布式对象存储系统。Rust
 - **完全开源**：采用 Apache 2.0 许可证，鼓励社区贡献和商业使用。
 - **简单易用**：设计简洁，易于部署和管理。

-| 功能               | 状态    | 功能                    | 状态      |
-| :----------------- | :------ | :---------------------- | :-------- |
-| **S3 核心功能**    | ✅ 可用 | **Bitrot (防数据腐烂)** | ✅ 可用   |
-| **上传 / 下载**    | ✅ 可用 | **单机模式**            | ✅ 可用   |
-| **版本控制**       | ✅ 可用 | **存储桶复制**          | ✅ 可用   |
-| **日志功能**       | ✅ 可用 | **生命周期管理**        | 🚧 测试中 |
-| **事件通知**       | ✅ 可用 | **分布式模式**          | 🚧 测试中 |
-| **K8s Helm Chart** | ✅ 可用 | **OPA (策略引擎)**      | 🚧 测试中 |
+状态说明：✅ 可用 —— 已发布并有 CI 门禁覆盖；🧪 预览 —— 已发布但需显式开关，或兼容性承诺有边界。
+
+| 功能                        | 状态    | 功能                     | 状态    |
+| :-------------------------- | :------ | :----------------------- | :------ |
+| **S3 核心功能**             | ✅ 可用 | **分布式模式**           | ✅ 可用 |
+| **上传 / 下载**             | ✅ 可用 | **单机模式**             | ✅ 可用 |
+| **版本控制**                | ✅ 可用 | **Bitrot (防数据腐烂)**  | ✅ 可用 |
+| **对象锁定 (WORM)**         | ✅ 可用 | **修复与扫描器**         | ✅ 可用 |
+| **服务端加密 (SSE)**        | ✅ 可用 | **存储池扩容 / 下线**    | ✅ 可用 |
+| **RustFS KMS**              | ✅ 可用 | **存储桶复制**           | ✅ 可用 |
+| **生命周期管理 (ILM)**      | ✅ 可用 | **站点复制**             | ✅ 可用 |
+| **ILM 分层 (远端 S3)**      | ✅ 可用 | **存储桶配额**           | ✅ 可用 |
+| **S3 Select**               | ✅ 可用 | **事件通知**             | ✅ 可用 |
+| **S3 Tables (Iceberg REST)**| 🧪 预览 | **审计日志**             | ✅ 可用 |
+| **IAM / 策略**              | ✅ 可用 | **日志与可观测性**       | ✅ 可用 |
+| **OIDC / SSO**              | ✅ 可用 | **Web 控制台**           | ✅ 可用 |
+| **Keystone 认证**           | ✅ 可用 | **K8s Helm Chart**       | ✅ 可用 |
+| **Swift API**               | ✅ 可用 | **FTPS / WebDAV**        | ✅ 可用 |
+| **多租户**                  | ✅ 可用 | **SFTP**                 | ✅ 可用 |
+| **MinIO 磁盘格式兼容**      | 🧪 预览 |                          |         |
+
+说明：
+
+- **服务端加密**：支持 SSE-C、SSE-S3 与 SSE-KMS。SSE-KMS 必须先配置 KMS 服务；未配置 KMS 时请求 `aws:kms` 会被拒绝，不会降级到本地主密钥。
+- **RustFS KMS**：生产环境支持 Vault（KV2 / Transit）与 AWS KMS 后端；`Local` 与 `Static` 后端仅供开发与测试使用，详见 [KMS 后端安全属性](docs/operations/kms-backend-security.md)。

 ## RustFS vs MinIO 性能对比

@@ -145,6 +145,38 @@ pub fn consume_verified_mrf_repair_events(anchors: &mut Vec<MrfDurableRepairAnch
    before.saturating_sub(anchors.len())
 }

+/// Consume recorded verified repairs for one bucket without draining
+/// unrelated or still-unmatched proofs. If the caller crashes before
+/// persisting the retained anchor set, the proof may be replayed by repair
+/// instead of silently deleting the old responsibility.
+pub fn consume_recorded_verified_mrf_repair_events_for(bucket: &str, anchors: &mut Vec<MrfDurableRepairAnchor>) -> usize {
+    let Some(registry) = MRF_VERIFIED_REPAIR_EVENTS.get() else {
+        return 0;
+    };
+    let Ok(mut events) = registry.lock() else {
+        return 0;
+    };
+    let before = anchors.len();
+    let mut retained = std::collections::VecDeque::with_capacity(events.len());
+    while let Some(event) = events.pop_front() {
+        if event.bucket.as_ref() != bucket {
+            retained.push_back(event);
+            continue;
+        }
+        let mut matched = false;
+        anchors.retain(|anchor| {
+            let proven = anchor.is_proven_by(&event);
+            matched |= proven;
+            !proven
+        });
+        if !matched {
+            retained.push_back(event);
+        }
+    }
+    *events = retained;
+    before.saturating_sub(anchors.len())
+}
+
 #[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
 pub struct MrfScope {
    pub pool_index: u32,
@@ -821,6 +853,71 @@ mod tests {
        assert!(retained.is_empty());
    }

+    #[test]
+    fn recorded_verified_repair_consumer_retains_unmatched_proofs() {
+        let bucket = Arc::<str>::from(format!("recorded-proof-{}", Uuid::new_v4()));
+        let other_bucket = Arc::<str>::from(format!("recorded-proof-other-{}", Uuid::new_v4()));
+        let incarnation = Uuid::new_v4();
+        let lease = MrfIngressLease::new(21);
+        let retained_anchor = MrfDurableRepairAnchor {
+            kind: MrfKind::PartialWrite,
+            bucket: bucket.clone(),
+            object: Arc::from("retained"),
+            version_id: Some([7; 16]),
+            scope: Some(MrfScope {
+                pool_index: 1,
+                set_index: 2,
+            }),
+            lease,
+            bucket_incarnation_id: incarnation,
+        };
+        let waiting_anchor = MrfDurableRepairAnchor {
+            object: Arc::from("waiting"),
+            lease: MrfIngressLease::new(22),
+            ..retained_anchor.clone()
+        };
+        let matched_event = MrfVerifiedRepairEvent {
+            kind: retained_anchor.kind,
+            bucket: bucket.clone(),
+            object: retained_anchor.object.clone(),
+            version_id: retained_anchor.version_id,
+            scope: retained_anchor.scope,
+            lease: Some(retained_anchor.lease),
+            bucket_incarnation_id: retained_anchor.bucket_incarnation_id,
+            disposition: MrfVerifiedRepairDisposition::Repaired,
+        };
+        let same_bucket_unmatched = MrfVerifiedRepairEvent {
+            object: Arc::from("future"),
+            lease: Some(MrfIngressLease::new(23)),
+            ..matched_event.clone()
+        };
+        let other_bucket_event = MrfVerifiedRepairEvent {
+            bucket: other_bucket.clone(),
+            ..matched_event.clone()
+        };
+
+        note_mrf_verified_repair(matched_event);
+        note_mrf_verified_repair(same_bucket_unmatched.clone());
+        note_mrf_verified_repair(other_bucket_event.clone());
+
+        let mut anchors = vec![retained_anchor, waiting_anchor.clone()];
+        assert_eq!(consume_recorded_verified_mrf_repair_events_for(&bucket, &mut anchors), 1);
+        assert_eq!(anchors, vec![waiting_anchor]);
+
+        let remaining_bucket_events = take_mrf_verified_repair_events_for(&bucket);
+        assert_eq!(
+            remaining_bucket_events,
+            vec![same_bucket_unmatched],
+            "same-bucket proofs without a retained anchor must remain available"
+        );
+        let remaining_other_events = take_mrf_verified_repair_events_for(&other_bucket);
+        assert_eq!(
+            remaining_other_events,
+            vec![other_bucket_event],
+            "proofs for other buckets must not be drained by this consumer"
+        );
+    }
+
    #[tokio::test]
    async fn try_send_delivers_and_respects_capacity() {
        let mut receiver = init_mrf_channel().expect("first initialization should succeed");
@@ -593,6 +593,20 @@ pub struct DataUsageSnapshotIdentity {
    pub scanner_epoch: Option<u64>,
 }

+#[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
+pub struct DataUsageSegmentInvalidationProof {
+    #[serde(default)]
+    pub process_epoch: String,
+    #[serde(default)]
+    pub generation_start: u64,
+    #[serde(default)]
+    pub generation_end: u64,
+    #[serde(default)]
+    pub producer_identity_coverage_complete: bool,
+    #[serde(default)]
+    pub cold_zero_walk_oracle: bool,
+}
+
 #[derive(Clone, Debug, Default, Serialize, Deserialize, PartialEq, Eq)]
 pub struct DataUsageSnapshotSetState {
    pub pool_index: u64,
@@ -607,6 +621,8 @@ pub struct DataUsageSnapshotSetState {
    pub complete: bool,
    #[serde(default)]
    pub tombstone: bool,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub segment_invalidation_proof: Option<DataUsageSegmentInvalidationProof>,
 }

 impl DataUsageInfo {
@@ -3073,6 +3089,7 @@ mod tests {
            scan_plan_digest: Some([1; 32]),
            complete: false,
            tombstone: false,
+            segment_invalidation_proof: None,
        }];
        assert!(observed_data_usage_is_newer(&partial, &authoritative));
    }
@@ -3095,6 +3112,7 @@ mod tests {
                    scan_plan_digest: Some([1; 32]),
                    complete: true,
                    tombstone: false,
+                    segment_invalidation_proof: None,
                },
                DataUsageSnapshotSetState {
                    pool_index: 1,
@@ -3104,6 +3122,7 @@ mod tests {
                    scan_plan_digest: Some([2; 32]),
                    complete: false,
                    tombstone: false,
+                    segment_invalidation_proof: None,
                },
            ],
            ..Default::default()
@@ -3113,6 +3132,42 @@ mod tests {
        assert!(partial.is_valid_partial_snapshot());
    }

+    #[test]
+    fn set_state_segment_invalidation_proof_is_additive() {
+        #[derive(Deserialize)]
+        struct LegacySetState {
+            pool_index: u64,
+            set_index: u64,
+            complete: bool,
+        }
+
+        let proof = DataUsageSegmentInvalidationProof {
+            process_epoch: "scanner-process".to_string(),
+            generation_start: 3,
+            generation_end: 5,
+            producer_identity_coverage_complete: true,
+            cold_zero_walk_oracle: true,
+        };
+        let state = DataUsageSnapshotSetState {
+            pool_index: 1,
+            set_index: 2,
+            scanner_cycle: Some(9),
+            scanner_epoch: Some(4),
+            scan_plan_digest: Some([7; 32]),
+            complete: true,
+            tombstone: false,
+            segment_invalidation_proof: Some(proof.clone()),
+        };
+        let encoded = rmp_serde::to_vec_named(&state).expect("set state should encode with additive proof");
+        let legacy: LegacySetState = rmp_serde::from_slice(&encoded).expect("legacy readers should ignore proof metadata");
+        assert_eq!(legacy.pool_index, 1);
+        assert_eq!(legacy.set_index, 2);
+        assert!(legacy.complete);
+
+        let decoded: DataUsageSnapshotSetState = rmp_serde::from_slice(&encoded).expect("new readers should restore proof");
+        assert_eq!(decoded.segment_invalidation_proof, Some(proof));
+    }
+
    #[test]
    fn completeness_marker_requires_a_snapshot_timestamp() {
        let untimestamped = DataUsageInfo {
@@ -51,7 +51,7 @@ fn main() {
        }
    }
    let revision = git(&root, &["rev-parse", "HEAD"]).unwrap_or_else(|| "unknown".to_owned());
-    let dirty = git(&root, &["status", "--porcelain", "--untracked-files=normal"]).is_none_or(|status| !status.is_empty());
+    let dirty = git(&root, &["status", "--porcelain", "--untracked-files=no"]).is_none_or(|status| !status.is_empty());
    let lock = git(&root, &["hash-object", "Cargo.lock"]).unwrap_or_else(|| "unknown".to_owned());
    let mut features = std::env::vars()
        .filter_map(|(key, _)| {
@@ -268,6 +268,7 @@ async fn test_bucket_cors_write_is_visible_on_peer_before_response() -> Result<(
    let rule = CorsRule::builder()
        .allowed_methods("GET")
        .allowed_origins("https://example.com")
+        .allowed_headers("*")
        .build()?;
    let configuration = CorsConfiguration::builder().cors_rules(rule).build()?;

@@ -288,6 +289,60 @@ async fn test_bucket_cors_write_is_visible_on_peer_before_response() -> Result<(
    assert_eq!(rules[0].allowed_methods(), ["GET"]);
    assert_eq!(rules[0].allowed_origins(), ["https://example.com"]);

+    let http = reqwest::Client::builder().no_proxy().build()?;
+    let url = format!("http://{}/{}", cluster.nodes[1].address, BUCKET_METADATA_RELOAD_BUCKET);
+    let without_headers = http
+        .request(reqwest::Method::OPTIONS, &url)
+        .header("Origin", "https://example.com")
+        .header("Access-Control-Request-Method", "GET")
+        .send()
+        .await?;
+    assert!(without_headers.status().is_success());
+    assert!(!without_headers.headers().contains_key("access-control-allow-headers"));
+    assert!(
+        without_headers
+            .headers()
+            .get("vary")
+            .and_then(|value| value.to_str().ok())
+            .is_some_and(|value| value.contains("Access-Control-Request-Headers")),
+        "a cached header-free preflight must not suppress a later requested header grant"
+    );
+    let preflight = http
+        .request(reqwest::Method::OPTIONS, &url)
+        .header("Origin", "https://example.com")
+        .header("Access-Control-Request-Method", "GET")
+        .header("Access-Control-Request-Headers", "X-Another-Header, x-could-be-anything")
+        .send()
+        .await?;
+    assert!(preflight.status().is_success(), "peer preflight should succeed: {preflight:?}");
+    assert_eq!(
+        preflight
+            .headers()
+            .get("access-control-allow-headers")
+            .and_then(|value| value.to_str().ok()),
+        Some("x-another-header,x-could-be-anything"),
+        "a wildcard rule must return only the headers requested by this preflight"
+    );
+    assert!(
+        preflight
+            .headers()
+            .get("vary")
+            .and_then(|value| value.to_str().ok())
+            .is_some_and(|value| value.contains("Access-Control-Request-Headers")),
+        "preflight caches must distinguish the requested header list"
+    );
+    let denied = http
+        .request(reqwest::Method::OPTIONS, &url)
+        .header("Origin", "https://disallowed.example.com")
+        .header("Access-Control-Request-Method", "GET")
+        .header("Access-Control-Request-Headers", "x-another-header")
+        .send()
+        .await?;
+    assert!(
+        !denied.headers().contains_key("access-control-allow-headers"),
+        "a rejected origin must not receive the requested header grant"
+    );
+
    writer
        .delete_bucket_cors()
        .bucket(BUCKET_METADATA_RELOAD_BUCKET)
@@ -188,6 +188,15 @@ fn write_next_test_port(port: u16) -> Result<(), Box<dyn std::error::Error + Sen
    Ok(())
 }

+fn no_available_test_port_error(config: TestPortAllocatorConfig, attempts: u16, last_error: Option<&std::io::Error>) -> String {
+    let max_inclusive = config.max_exclusive() - 1;
+    let detail = last_error.map(|err| format!("; last bind error: {err}")).unwrap_or_default();
+    format!(
+        "no available E2E test port found in {}..={} after {} attempts{}",
+        config.min, max_inclusive, attempts, detail
+    )
+}
+
 pub(crate) fn capture_command_logs(
    command: &mut Command,
    log_path: Option<&str>,
@@ -694,19 +703,23 @@ impl RustFSTestEnvironment {
        let _guard = PortAllocatorGuard::acquire().await?;
        let config = test_port_allocator_config()?;
        let mut next_port = read_next_test_port(config);
+        let mut last_error = None;

        for _ in 0..config.range {
            let port = next_port;
            next_port = advance_test_port(next_port, config);
            write_next_test_port(next_port)?;

-            if let Ok(listener) = TcpListener::bind(("127.0.0.1", port)) {
-                drop(listener);
-                return Ok(port);
+            match TcpListener::bind(("127.0.0.1", port)) {
+                Ok(listener) => {
+                    drop(listener);
+                    return Ok(port);
+                }
+                Err(err) => last_error = Some(err),
            }
        }

-        Err("no available E2E test port found".into())
+        Err(no_available_test_port_error(config, config.range, last_error.as_ref()).into())
    }

    /// Kill any existing RustFS processes
@@ -2201,6 +2214,19 @@ mod tests {
        assert!(parse_test_port_allocator_config(Some("not-a-port"), Some("128")).is_err());
    }

+    #[test]
+    fn e2e_port_allocator_reports_attempt_window_and_last_bind_error() {
+        let config = TestPortAllocatorConfig { min: 41000, range: 3 };
+        let error = std::io::Error::from(ErrorKind::PermissionDenied);
+
+        let message = no_available_test_port_error(config, config.range, Some(&error));
+
+        assert!(message.contains("41000..=41002"));
+        assert!(message.contains("after 3 attempts"));
+        assert!(message.contains("last bind error"));
+        assert!(message.contains("permission denied"));
+    }
+
    #[test]
    fn resolves_rustfs_binary_in_configured_cargo_target_directory() {
        let workspace = Path::new("workspace");
@@ -59,6 +59,9 @@ const POOL_META_V3_ENV: [(&str, &str); 2] = [

 #[derive(Clone, Copy, Debug)]
 pub(crate) enum DistLayout {
+    /// 3 nodes × 4 drives, one erasure pool. With `EC:4` this is the
+    /// release-evidence EC8+4 geometry.
+    ThreeByFourEc84,
    /// 4 nodes × 4 drives, one erasure pool spanning every endpoint.
    FourByFour,
    /// 4 nodes × 1 drive, one erasure pool (minimum 4-node 4-disk layout).
@@ -94,6 +97,7 @@ impl DistCluster {

    pub async fn new_stopped_with_env(layout: DistLayout, extra_env: &[(&str, &str)]) -> TestResult<Self> {
        let topology = match layout {
+            DistLayout::ThreeByFourEc84 => ClusterTopology::single_pool_multidrive(3, DRIVES_PER_NODE),
            DistLayout::FourByFour => ClusterTopology::single_pool_multidrive(NODE_COUNT, DRIVES_PER_NODE),
            DistLayout::FourNodeFourDisk => ClusterTopology::single_pool(NODE_COUNT),
            DistLayout::SingleNodeFourDrive => ClusterTopology::per_node_pools(DRIVES_PER_NODE, vec![vec![0]]),
@@ -101,7 +105,7 @@ impl DistCluster {
        let mut cluster = RustFSTestClusterEnvironment::with_topology(topology).await?;
        let pool_storage_roots = match layout {
            DistLayout::SingleNodeFourDrive => Some(configured_pool_storage_roots()?),
-            DistLayout::FourByFour | DistLayout::FourNodeFourDisk => None,
+            DistLayout::ThreeByFourEc84 | DistLayout::FourByFour | DistLayout::FourNodeFourDisk => None,
        };
        let mut owned_pool_dirs = Vec::new();
        if let Some(roots) = pool_storage_roots.as_deref() {
@@ -958,27 +962,6 @@ pub(crate) async fn wait_for_rebalance_active(
    }
 }

-pub(crate) async fn wait_for_rebalance_running_with_progress(
-    cluster: &RustFSTestClusterEnvironment,
-    expected_id: &str,
-    timeout: Duration,
-) -> TestResult {
-    let deadline = Instant::now() + timeout;
-    loop {
-        let status = rebalance_status_json(cluster).await?;
-        if rebalance_running_with_progress(&status, expected_id)? {
-            return Ok(());
-        }
-        if Instant::now() >= deadline {
-            return Err(format!(
-                "rebalance did not become active with non-zero progress within {timeout:?}; last status: {status}"
-            )
-            .into());
-        }
-        sleep(Duration::from_millis(100)).await;
-    }
-}
-
 pub(crate) async fn wait_for_rebalance_complete(
    cluster: &RustFSTestClusterEnvironment,
    expected_id: &str,
@@ -0,0 +1,413 @@
+// Copyright 2026 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use super::harness::{
+    DistCluster, DistLayout, TestResult, assert_inventory, get_object_bytes, payload_for, put_object, sha256_hex, unique_bucket,
+    wait_until,
+};
+use crate::chaos::{VersionShardCensus, census_object_version_on_disk, signed_admin_post};
+use crate::common::init_logging;
+use aws_sdk_s3::Client;
+use aws_sdk_s3::primitives::ByteStream;
+use serde_json::Value;
+use sha2::{Digest, Sha256};
+use std::collections::{BTreeMap, HashSet};
+use std::io::{Read, Write};
+use std::path::{Path, PathBuf};
+use std::time::Duration;
+use tokio::time::{Instant, sleep};
+
+const EC84_NODE_COUNT: usize = 3;
+const EC84_DRIVES_PER_NODE: usize = 4;
+const EC84_DATA_BLOCKS: usize = 8;
+const EC84_PARITY_BLOCKS: usize = 4;
+const EC84_TARGET_DRIVE_RESTART_CASE: &str = "ec84-target-drive-restart";
+const EC84_TARGET_DRIVE_RESTART_ORACLE: &str = "ec84-target-drive-restart.json";
+const EC84_HEAL_CONTROL_READY_TIMEOUT: Duration = Duration::from_secs(45);
+const EC84_HEAL_CONTROL_RETRY_DELAY: Duration = Duration::from_millis(250);
+
+#[derive(Clone)]
+struct ExpectedShard {
+    key: String,
+    body: Vec<u8>,
+    baseline: VersionShardCensus,
+}
+
+struct ScannerHealEvidenceContext {
+    directory: PathBuf,
+    run: Value,
+}
+
+struct ScannerHealEvidencePayload<'a> {
+    dist: &'a DistCluster,
+    bucket: &'a str,
+    expected: &'a [ExpectedShard],
+    outage_key: &'a str,
+    outage_body: &'a [u8],
+    replaced_drive: &'a Path,
+    pid_before: u32,
+    pid_after: u32,
+    node_listings: Vec<Vec<String>>,
+}
+
+fn file_sha256(path: &Path) -> TestResult<String> {
+    let mut file = std::fs::File::open(path)?;
+    let mut digest = Sha256::new();
+    let mut buffer = [0_u8; 64 * 1024];
+    loop {
+        let read = file.read(&mut buffer)?;
+        if read == 0 {
+            break;
+        }
+        digest.update(&buffer[..read]);
+    }
+    Ok(digest.finalize().iter().map(|byte| format!("{byte:02x}")).collect())
+}
+
+fn compiled_test_identity() -> Value {
+    serde_json::json!({
+        "source_revision": env!("RUSTFS_E2E_BUILD_COMMIT"),
+        "dirty": env!("RUSTFS_E2E_BUILD_DIRTY") != "false",
+        "lock_blob": env!("RUSTFS_E2E_BUILD_LOCK"),
+        "features": env!("RUSTFS_E2E_BUILD_FEATURES"),
+        "target": env!("RUSTFS_E2E_BUILD_TARGET"),
+        "profile": env!("RUSTFS_E2E_BUILD_PROFILE"),
+        "rustflags_hex": env!("RUSTFS_E2E_BUILD_RUSTFLAGS_HEX"),
+    })
+}
+
+fn string_field<'a>(value: &'a Value, path: &str) -> TestResult<&'a str> {
+    let mut current = value;
+    for segment in path.split('.') {
+        current = current
+            .get(segment)
+            .ok_or_else(|| format!("scanner/heal run receipt missing {path}"))?;
+    }
+    current
+        .as_str()
+        .filter(|text| !text.is_empty())
+        .ok_or_else(|| format!("scanner/heal run receipt has invalid {path}").into())
+}
+
+fn scanner_heal_evidence_context() -> TestResult<Option<ScannerHealEvidenceContext>> {
+    let Some(directory) = std::env::var_os("RUSTFS_SCANNER_HEAL_RUN_DIR") else {
+        return Ok(None);
+    };
+    let directory = PathBuf::from(directory);
+    let receipt = directory.join("run.json");
+    if receipt.metadata()?.len() > 1024 * 1024 {
+        return Err("oversized scanner/heal execution receipt".into());
+    }
+    let run: Value = serde_json::from_slice(&std::fs::read(receipt)?)?;
+    let built = compiled_test_identity();
+    for key in ["source_revision", "dirty", "lock_blob", "features"] {
+        if built[key] != run["test_build"][key] {
+            return Err(format!("compiled test identity differs for {key}").into());
+        }
+    }
+    let binary_path = PathBuf::from(string_field(&run, "binary.path")?);
+    if file_sha256(&binary_path)? != string_field(&run, "binary.sha256")? {
+        return Err("server binary must match the run receipt".into());
+    }
+    if file_sha256(&std::env::current_exe()?)? != string_field(&run, "test_binary.sha256")? {
+        return Err("test executable must match the run receipt".into());
+    }
+    if directory.join(EC84_TARGET_DRIVE_RESTART_ORACLE).exists() {
+        return Err("scanner/heal oracle already exists; create a new execution receipt".into());
+    }
+    Ok(Some(ScannerHealEvidenceContext { directory, run }))
+}
+
+fn assert_ec84_geometry(census: &VersionShardCensus, key: &str) -> TestResult {
+    if census.data_blocks != Some(EC84_DATA_BLOCKS) || census.parity_blocks != Some(EC84_PARITY_BLOCKS) {
+        return Err(format!("object {key} did not use EC8+4 geometry: {census:?}").into());
+    }
+    let erasure_index = census
+        .erasure_index
+        .ok_or_else(|| format!("object {key} did not record an erasure index: {census:?}"))?;
+    if !(1..=EC84_DATA_BLOCKS + EC84_PARITY_BLOCKS).contains(&erasure_index) {
+        return Err(format!("object {key} has out-of-range erasure index {erasure_index}: {census:?}").into());
+    }
+    if !census.is_complete() || census.expected_part_numbers.is_empty() {
+        return Err(format!("object {key} does not have complete physical shard evidence: {census:?}").into());
+    }
+    Ok(())
+}
+
+async fn write_scanner_heal_evidence(context: ScannerHealEvidenceContext, payload: ScannerHealEvidencePayload<'_>) -> TestResult {
+    let verifier = payload.dist.client(0)?;
+    let mut objects = Vec::new();
+    for item in payload.expected {
+        let actual = get_object_bytes(&verifier, payload.bucket, &item.key).await?;
+        let physical = census_object_version_on_disk(payload.replaced_drive, payload.bucket, &item.key, None)?;
+        objects.push(serde_json::json!({
+            "key": item.key,
+            "version_id": null,
+            "expected_bytes": item.body.len(),
+            "actual_bytes": actual.len(),
+            "expected_sha256": sha256_hex(&item.body),
+            "actual_sha256": sha256_hex(&actual),
+            "expected_physical": item.baseline,
+            "physical": physical,
+        }));
+    }
+    let actual = get_object_bytes(&verifier, payload.bucket, payload.outage_key).await?;
+    let physical = census_object_version_on_disk(payload.replaced_drive, payload.bucket, payload.outage_key, None)?;
+    objects.push(serde_json::json!({
+        "key": payload.outage_key,
+        "version_id": null,
+        "expected_bytes": payload.outage_body.len(),
+        "actual_bytes": actual.len(),
+        "expected_sha256": sha256_hex(payload.outage_body),
+        "actual_sha256": sha256_hex(&actual),
+        "expected_physical": null,
+        "physical": physical,
+    }));
+
+    let evidence = serde_json::json!({
+        "schema": 1,
+        "case": EC84_TARGET_DRIVE_RESTART_CASE,
+        "evidence": "process-restart",
+        "run_id": string_field(&context.run, "run_id")?,
+        "source_revision": string_field(&context.run, "source_revision")?,
+        "test_build": compiled_test_identity(),
+        "binary_sha256": string_field(&context.run, "binary.sha256")?,
+        "test_binary_sha256": string_field(&context.run, "test_binary.sha256")?,
+        "topology": {"nodes": EC84_NODE_COUNT, "drives_per_node": EC84_DRIVES_PER_NODE},
+        "pid_before": payload.pid_before,
+        "pid_after": payload.pid_after,
+        "unclean_shutdown_marker": false,
+        "objects": objects,
+        "node_listings": payload.node_listings,
+    });
+    let data = serde_json::to_vec(&evidence)?;
+    if data.len() > 1024 * 1024 {
+        return Err("scanner/heal oracle exceeds the 1 MiB artifact budget".into());
+    }
+    let mut output = std::fs::OpenOptions::new()
+        .write(true)
+        .create_new(true)
+        .open(context.directory.join(EC84_TARGET_DRIVE_RESTART_ORACLE))?;
+    output.write_all(&data)?;
+    output.sync_all()?;
+    Ok(())
+}
+
+fn assert_replaced_drive_empty(drive: &Path, bucket: &str, keys: &[String]) -> TestResult {
+    for key in keys {
+        let census = census_object_version_on_disk(drive, bucket, key, None)?;
+        if census.has_xl_meta {
+            return Err(format!("replacement drive unexpectedly retained {bucket}/{key}: {census:?}").into());
+        }
+    }
+    Ok(())
+}
+
+fn is_cluster_heal_coordination_unavailable(error: &(dyn std::error::Error + Send + Sync)) -> bool {
+    let message = error.to_string();
+    message.contains("500 Internal Server Error") && message.contains("cluster heal coordination unavailable")
+}
+
+async fn start_ec84_root_heal_when_control_ready(
+    heal_url: &str,
+    heal_body: &str,
+    access_key: &str,
+    secret_key: &str,
+) -> TestResult {
+    let deadline = Instant::now() + EC84_HEAL_CONTROL_READY_TIMEOUT;
+    loop {
+        match signed_admin_post(heal_url, Some(heal_body), access_key, secret_key).await {
+            Ok(_) => return Ok(()),
+            Err(error) if is_cluster_heal_coordination_unavailable(error.as_ref()) && Instant::now() < deadline => {
+                sleep(EC84_HEAL_CONTROL_RETRY_DELAY).await;
+            }
+            Err(error) => return Err(error),
+        }
+    }
+}
+
+async fn put_large_inventory(client: &Client, bucket: &str) -> TestResult<Vec<ExpectedShard>> {
+    let mut expected = Vec::new();
+    for index in 0..4 {
+        let key = format!("ec84/prefix-{}/object-{index:04}.bin", index % 2);
+        let body = payload_for(&key, 10 * 1024 * 1024);
+        put_object(client, bucket, &key, body.clone()).await?;
+        expected.push(ExpectedShard {
+            key,
+            body,
+            baseline: VersionShardCensus {
+                version_id: None,
+                has_xl_meta: false,
+                data_dir: None,
+                erasure_index: None,
+                data_blocks: None,
+                parity_blocks: None,
+                expected_part_numbers: Default::default(),
+                present_part_fingerprints: Default::default(),
+                inline_data_fingerprint: None,
+            },
+        });
+    }
+    Ok(expected)
+}
+
+#[tokio::test]
+async fn three_node_four_drive_ec8_4_root_heal_rebuilds_replaced_drive_after_restart() -> TestResult {
+    init_logging();
+    let evidence_context = scanner_heal_evidence_context()?;
+    let mut dist = DistCluster::start_with_env(
+        DistLayout::ThreeByFourEc84,
+        &[
+            ("RUSTFS_STORAGE_CLASS_STANDARD", "EC:4"),
+            ("RUSTFS_HEAL_ENABLED", "true"),
+            ("RUSTFS_HEAL_AUTO_HEAL_ENABLE", "false"),
+            ("RUSTFS_HEAL_MRF_ENABLE", "false"),
+            ("RUSTFS_SCANNER_ENABLED", "false"),
+        ],
+    )
+    .await?;
+    assert_eq!(dist.cluster.nodes.len(), EC84_NODE_COUNT);
+    assert_eq!(dist.cluster.topology.drives_per_node, EC84_DRIVES_PER_NODE);
+
+    let bucket = unique_bucket("healec84");
+    dist.create_bucket(&bucket).await?;
+    let writer = dist.client(0)?;
+    let mut expected = put_large_inventory(&writer, &bucket).await?;
+    let replaced_node = 1;
+    let replaced_drive_index = 2;
+    let replaced_drive = PathBuf::from(&dist.cluster.nodes[replaced_node].data_dirs[replaced_drive_index]);
+
+    for item in &mut expected {
+        item.baseline = census_object_version_on_disk(&replaced_drive, &bucket, &item.key, None)?;
+        assert_ec84_geometry(&item.baseline, &item.key)?;
+    }
+
+    let format_path = replaced_drive.join(".rustfs.sys").join("format.json");
+    let format_json = std::fs::read(&format_path)?;
+    let target_pid_before = dist.cluster.nodes[replaced_node]
+        .process
+        .as_ref()
+        .ok_or("target process is absent before graceful restart")?
+        .id();
+    dist.cluster.stop_node_gracefully(replaced_node).await?;
+    let retired_drive = PathBuf::from(format!("{}.retired", replaced_drive.display()));
+    std::fs::rename(&replaced_drive, &retired_drive)?;
+    std::fs::create_dir_all(format_path.parent().ok_or("replacement format path has no parent")?)?;
+    std::fs::write(&format_path, format_json)?;
+    assert_replaced_drive_empty(
+        &replaced_drive,
+        &bucket,
+        &expected.iter().map(|item| item.key.clone()).collect::<Vec<_>>(),
+    )?;
+
+    let outage_key = "ec84/written-while-node-restarting.bin";
+    let outage_body = payload_for(outage_key, 10 * 1024 * 1024);
+    writer
+        .put_object()
+        .bucket(&bucket)
+        .key(outage_key)
+        .body(ByteStream::from(outage_body.clone()))
+        .send()
+        .await?;
+
+    dist.cluster.start_node(replaced_node).await?;
+    let target_pid_after = dist.cluster.nodes[replaced_node]
+        .process
+        .as_ref()
+        .ok_or("target process is absent after restart")?
+        .id();
+    let heal_body =
+        r#"{"recursive":true,"dryRun":false,"remove":false,"recreate":true,"scanMode":2,"updateParity":false,"nolock":false}"#;
+    let heal_url = format!("{}/rustfs/admin/v3/heal/{bucket}?forceStart=true", dist.cluster.nodes[0].url);
+    start_ec84_root_heal_when_control_ready(&heal_url, heal_body, &dist.cluster.access_key, &dist.cluster.secret_key).await?;
+
+    wait_until(
+        Duration::from_secs(120),
+        || async {
+            for item in &expected {
+                let current = census_object_version_on_disk(&replaced_drive, &bucket, &item.key, None)?;
+                if !current.matches_manifest(&item.baseline) {
+                    return Ok(false);
+                }
+            }
+            let outage = census_object_version_on_disk(&replaced_drive, &bucket, outage_key, None)?;
+            Ok(outage.is_complete()
+                && outage.data_blocks == Some(EC84_DATA_BLOCKS)
+                && outage.parity_blocks == Some(EC84_PARITY_BLOCKS))
+        },
+        "EC8+4 replacement drive rebuilt baseline and outage shards",
+    )
+    .await?;
+
+    let inventory = expected
+        .iter()
+        .map(|item| (item.key.clone(), item.body.clone()))
+        .chain(std::iter::once((outage_key.to_string(), outage_body.clone())))
+        .collect::<BTreeMap<_, _>>();
+    let expected_keys = inventory.keys().cloned().collect::<HashSet<_>>();
+    let mut node_listings = Vec::new();
+    for node_index in 0..dist.cluster.nodes.len() {
+        let client = dist.client(node_index)?;
+        assert_inventory(&client, &bucket, &inventory).await?;
+        let listing = client.list_objects_v2().bucket(&bucket).send().await?;
+        let observed = listing
+            .contents()
+            .iter()
+            .filter_map(|object| object.key().map(str::to_owned))
+            .collect::<HashSet<_>>();
+        assert_eq!(observed, expected_keys, "node {node_index} listing diverged after EC8+4 heal");
+        let mut observed = observed.into_iter().collect::<Vec<_>>();
+        observed.sort();
+        node_listings.push(observed);
+    }
+    if let Some(context) = evidence_context {
+        write_scanner_heal_evidence(
+            context,
+            ScannerHealEvidencePayload {
+                dist: &dist,
+                bucket: &bucket,
+                expected: &expected,
+                outage_key,
+                outage_body: &outage_body,
+                replaced_drive: &replaced_drive,
+                pid_before: target_pid_before,
+                pid_after: target_pid_after,
+                node_listings,
+            },
+        )
+        .await?;
+    }
+
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn cluster_heal_coordination_retry_is_exact() {
+        let retryable: Box<dyn std::error::Error + Send + Sync> =
+            "admin POST failed: 500 Internal Server Error cluster heal coordination unavailable".into();
+        assert!(is_cluster_heal_coordination_unavailable(retryable.as_ref()));
+
+        let other_internal: Box<dyn std::error::Error + Send + Sync> =
+            "admin POST failed: 500 Internal Server Error unrelated".into();
+        assert!(!is_cluster_heal_coordination_unavailable(other_internal.as_ref()));
+
+        let wrong_status: Box<dyn std::error::Error + Send + Sync> =
+            "admin POST failed: 503 Service Unavailable cluster heal coordination unavailable".into();
+        assert!(!is_cluster_heal_coordination_unavailable(wrong_status.as_ref()));
+    }
+}
@@ -25,6 +25,7 @@ mod data_integrity_movement_test;
 mod expand_decommission_rebalance_test;
 mod extra_test;
 mod harness;
+mod heal_test;
 mod object_lock_test;
 mod observability_test;
 mod replication_quota_test;
@@ -14,9 +14,9 @@

 use super::harness::{
    DECOMMISSION_POOL_ID, DistCluster, DistLayout, TestResult, assert_inventory, decommission_running_with_progress,
-    decommission_status_json, put_inventory_retrying, rebalance_running_with_progress, rebalance_status_json,
-    retrying_get_equals, retrying_put, start_decommission, start_rebalance, unique_bucket, wait_for_decommission_complete,
-    wait_for_decommission_running_with_progress, wait_for_rebalance_complete, wait_for_rebalance_running_with_progress,
+    decommission_status_json, put_inventory_retrying, rebalance_active, rebalance_status_json, retrying_get_equals, retrying_put,
+    start_decommission, start_rebalance, unique_bucket, wait_for_decommission_complete,
+    wait_for_decommission_running_with_progress, wait_for_rebalance_active, wait_for_rebalance_complete,
 };
 use crate::common::init_logging;
 use std::time::Duration;
@@ -67,7 +67,10 @@ async fn s3_put_get_list_succeed_during_decommission_and_rebalance() -> TestResu
    assert_inventory(&live, &bucket, &inventory).await?;

    let rebalance_id = start_rebalance(&dist.cluster).await?;
-    wait_for_rebalance_running_with_progress(&dist.cluster, &rebalance_id, Duration::from_secs(30)).await?;
+    // The status API reads persisted progress, whose first periodic save is
+    // after 30 seconds. A shorter run can remain at zero until completion.
+    // Require Started around the S3 operations and nonzero progress at completion.
+    wait_for_rebalance_active(&dist.cluster, &rebalance_id, Duration::from_secs(30)).await?;
    retrying_put(
        &live,
        &bucket,
@@ -84,11 +87,26 @@ async fn s3_put_get_list_succeed_during_decommission_and_rebalance() -> TestResu
        Duration::from_secs(30),
    )
    .await?;
+    let listed = live.list_objects_v2().bucket(&bucket).send().await?;
+    assert!(
+        listed
+            .contents()
+            .iter()
+            .any(|object| object.key() == Some("during-rebalance.bin")),
+        "list during rebalance missed the newly written key"
+    );
    let status = rebalance_status_json(&dist.cluster).await?;
-    if !rebalance_running_with_progress(&status, &rebalance_id)? {
+    if !rebalance_active(&status, &rebalance_id)? {
        return Err(format!("rebalance did not remain active across the S3 operations: {status}").into());
    }
    wait_for_rebalance_complete(&dist.cluster, &rebalance_id, Duration::from_secs(180)).await?;
-    assert_inventory(&dist.client(1)?, &bucket, &inventory).await?;
+    let after = dist.client(1)?;
+    assert_inventory(&after, &bucket, &inventory).await?;
+    for (key, body) in [
+        ("during-decommission.bin", b"written-while-decommissioning".as_slice()),
+        ("during-rebalance.bin", b"written-while-rebalancing".as_slice()),
+    ] {
+        retrying_get_equals(&after, &bucket, key, body, Duration::from_secs(30)).await?;
+    }
    Ok(())
 }
@@ -18,8 +18,8 @@
 mod tests {
    use crate::chaos::{VersionShardCensus, census_object_version_on_disk, sha256_hex, signed_admin_post};
    use crate::common::{
-        FAST_DATA_USAGE_SCANNER_ENV, RustFSTestClusterEnvironment, RustFSTestEnvironment, admin_request, init_logging,
-        rustfs_binary_path,
+        ClusterTopology, FAST_DATA_USAGE_SCANNER_ENV, RustFSTestClusterEnvironment, RustFSTestEnvironment, admin_request,
+        init_logging, rustfs_binary_path,
    };
    use crate::storage_api::RUSTFS_META_BUCKET;
    use aws_sdk_s3::primitives::ByteStream;
@@ -60,6 +60,29 @@ mod tests {
        oracle: &'static str,
        evidence: &'static str,
        unclean_shutdown_marker: bool,
+        topology: EvidenceTopology,
+        storage_class_standard: Option<&'static str>,
+        erasure_set_drive_count: Option<&'static str>,
+    }
+
+    #[derive(Clone, Copy)]
+    struct EvidenceTopology {
+        nodes: usize,
+        drives_per_node: usize,
+    }
+
+    impl EvidenceTopology {
+        const fn new(nodes: usize, drives_per_node: usize) -> Self {
+            Self { nodes, drives_per_node }
+        }
+
+        fn total_drives(self) -> usize {
+            self.nodes * self.drives_per_node
+        }
+
+        fn cluster_topology(self) -> ClusterTopology {
+            ClusterTopology::single_pool_multidrive(self.nodes, self.drives_per_node)
+        }
    }

    const BACKGROUND_TARGET_RESTART_EVIDENCE: ScannerHealEvidenceCase = ScannerHealEvidenceCase {
@@ -67,6 +90,9 @@ mod tests {
        oracle: "background-target-restart.json",
        evidence: "process-restart",
        unclean_shutdown_marker: false,
+        topology: EvidenceTopology::new(4, 1),
+        storage_class_standard: None,
+        erasure_set_drive_count: None,
    };

    const BACKGROUND_TARGET_CRASH_EVIDENCE: ScannerHealEvidenceCase = ScannerHealEvidenceCase {
@@ -74,6 +100,29 @@ mod tests {
        oracle: "background-target-crash.json",
        evidence: "process-crash-restart",
        unclean_shutdown_marker: true,
+        topology: EvidenceTopology::new(4, 1),
+        storage_class_standard: None,
+        erasure_set_drive_count: None,
+    };
+
+    const BACKGROUND_TARGET_RESTART_EC84_EVIDENCE: ScannerHealEvidenceCase = ScannerHealEvidenceCase {
+        id: "background-target-restart-ec8-4",
+        oracle: "background-target-restart-ec8-4.json",
+        evidence: "process-restart",
+        unclean_shutdown_marker: false,
+        topology: EvidenceTopology::new(3, 4),
+        storage_class_standard: Some("EC:4"),
+        erasure_set_drive_count: Some("12"),
+    };
+
+    const BACKGROUND_TARGET_CRASH_EC84_EVIDENCE: ScannerHealEvidenceCase = ScannerHealEvidenceCase {
+        id: "background-target-crash-ec8-4",
+        oracle: "background-target-crash-ec8-4.json",
+        evidence: "process-crash-restart",
+        unclean_shutdown_marker: true,
+        topology: EvidenceTopology::new(3, 4),
+        storage_class_standard: Some("EC:4"),
+        erasure_set_drive_count: Some("12"),
    };

    struct RestartEvidenceContext {
@@ -332,11 +381,10 @@ mod tests {

    // Healing may rewrite non-identity bookkeeping in xl.meta. The census
    // therefore compares the canonical selected metadata fields plus every
-    // physical shard, while the payload seed makes object mix-ups observable.
+    // physical shard.
    #[derive(Debug)]
    struct PhysicalObjectManifest {
        key: String,
-        payload_seed: u8,
        shard_census: VersionShardCensus,
    }

@@ -973,6 +1021,26 @@ mod tests {
        .await?
    }

+    #[tokio::test(flavor = "multi_thread")]
+    async fn test_cluster_root_heal_recovers_ec84_shards_after_background_target_restart()
+    -> Result<(), Box<dyn Error + Send + Sync>> {
+        timeout(
+            Duration::from_secs(420),
+            run_cluster_root_heal_interruption(InterruptionScenario::BackgroundTargetRestartEc84),
+        )
+        .await?
+    }
+
+    #[tokio::test(flavor = "multi_thread")]
+    async fn test_cluster_root_heal_recovers_ec84_shards_after_background_target_crash()
+    -> Result<(), Box<dyn Error + Send + Sync>> {
+        timeout(
+            Duration::from_secs(420),
+            run_cluster_root_heal_interruption(InterruptionScenario::BackgroundTargetCrashEc84),
+        )
+        .await?
+    }
+
    #[tokio::test(flavor = "multi_thread")]
    async fn test_cluster_root_heal_recovers_remote_shards_after_coordinator_restart() -> Result<(), Box<dyn Error + Send + Sync>>
    {
@@ -1010,29 +1078,38 @@ mod tests {
        IsolatedTargetRestart,
        BackgroundTargetRestart,
        BackgroundTargetCrash,
+        BackgroundTargetRestartEc84,
+        BackgroundTargetCrashEc84,
        BackgroundCoordinatorRestart,
        TargetEndpointBlackhole,
    }

    async fn run_cluster_root_heal_interruption(scenario: InterruptionScenario) -> Result<(), Box<dyn Error + Send + Sync>> {
        let server_binary = rustfs_binary_path();
-        let evidence_run = match scenario {
-            InterruptionScenario::BackgroundTargetRestart => {
-                restart_evidence_run(&server_binary, BACKGROUND_TARGET_RESTART_EVIDENCE)?
-            }
-            InterruptionScenario::BackgroundTargetCrash => {
-                restart_evidence_run(&server_binary, BACKGROUND_TARGET_CRASH_EVIDENCE)?
-            }
+        let evidence_case = match scenario {
+            InterruptionScenario::BackgroundTargetRestart => Some(BACKGROUND_TARGET_RESTART_EVIDENCE),
+            InterruptionScenario::BackgroundTargetCrash => Some(BACKGROUND_TARGET_CRASH_EVIDENCE),
+            InterruptionScenario::BackgroundTargetRestartEc84 => Some(BACKGROUND_TARGET_RESTART_EC84_EVIDENCE),
+            InterruptionScenario::BackgroundTargetCrashEc84 => Some(BACKGROUND_TARGET_CRASH_EC84_EVIDENCE),
            _ => None,
        };
+        let evidence_run = match evidence_case {
+            Some(case) => restart_evidence_run(&server_binary, case)?,
+            None => None,
+        };
        let mut evidence_objects = Vec::new();
        let (background_enabled, interruption_node, interruption_kind) = match scenario {
            InterruptionScenario::IsolatedTargetRestart => (false, 1, "target_restart"),
            InterruptionScenario::BackgroundTargetRestart => (true, 1, "background_target_restart"),
            InterruptionScenario::BackgroundTargetCrash => (true, 1, "background_target_crash"),
+            InterruptionScenario::BackgroundTargetRestartEc84 => (true, 1, "background_target_restart_ec8_4"),
+            InterruptionScenario::BackgroundTargetCrashEc84 => (true, 1, "background_target_crash_ec8_4"),
            InterruptionScenario::BackgroundCoordinatorRestart => (true, 0, "coordinator_restart"),
            InterruptionScenario::TargetEndpointBlackhole => (false, 1, "target_endpoint_blackhole"),
        };
+        let topology = evidence_case
+            .map(|case| case.topology)
+            .unwrap_or_else(|| EvidenceTopology::new(4, 1));
        init_logging();
        info!(
            event = "heal_interruption_started",
@@ -1044,9 +1121,15 @@ mod tests {
            "Starting root-heal interruption test"
        );

-        let mut cluster = RustFSTestClusterEnvironment::new(4).await?;
+        let mut cluster = RustFSTestClusterEnvironment::with_topology(topology.cluster_topology()).await?;
        cluster.set_env("RUSTFS_UNSAFE_BYPASS_DISK_CHECK", "true");
        cluster.set_env("RUSTFS_HEAL_ENABLED", "true");
+        if let Some(storage_class) = evidence_case.and_then(|case| case.storage_class_standard) {
+            cluster.set_env("RUSTFS_STORAGE_CLASS_STANDARD", storage_class);
+        }
+        if let Some(erasure_set_drive_count) = evidence_case.and_then(|case| case.erasure_set_drive_count) {
+            cluster.set_env("RUSTFS_ERASURE_SET_DRIVE_COUNT", erasure_set_drive_count);
+        }
        // Heal control uses the first lexicographically sorted grid host.
        // Keep that coordinator distinct from the remote target at index 1.
        cluster.nodes.sort_by(|left, right| left.url.cmp(&right.url));
@@ -1095,11 +1178,22 @@ mod tests {
            .and_then(|value| value.parse::<usize>().ok())
            .unwrap_or(4 * 1024 * 1024)
            .clamp(1024 * 1024, 16 * 1024 * 1024);
+        let mut created_online_objects = Vec::with_capacity(online_object_count);
        let mut expected_manifests = Vec::with_capacity(online_object_count);
        let mut unclean_shutdown_marker_observed = None;
-        for index in 0..online_object_count {
-            let key = format!("cluster/online/object-{index:04}.bin");
-            let payload_seed = u8::try_from(index + 1).expect("clamped object count must fit in u8");
+        let mut attempt_count = 0usize;
+        let max_online_attempts = online_object_count.saturating_mul(topology.total_drives().max(1));
+        while expected_manifests.len() < online_object_count {
+            if attempt_count >= max_online_attempts {
+                return Err(format!(
+                    "target replacement drive held only {}/{} baseline object shards after {attempt_count} writes",
+                    expected_manifests.len(),
+                    online_object_count
+                )
+                .into());
+            }
+            let key = format!("cluster/online/object-{attempt_count:04}.bin");
+            let payload_seed = ((attempt_count % 251) + 1) as u8;
            timeout(
                Duration::from_secs(30),
                clients[0]
@@ -1111,6 +1205,11 @@ mod tests {
            )
            .await??;
            let shard_census = census_object_version_on_disk(&replaced_disk, bucket, &key, None)?;
+            if !shard_census.has_xl_meta {
+                timeout(Duration::from_secs(30), clients[0].delete_object().bucket(bucket).key(&key).send()).await??;
+                attempt_count += 1;
+                continue;
+            }
            assert!(
                shard_census.is_complete(),
                "node 1 should hold a complete baseline shard for {key}: {shard_census:?}"
@@ -1119,11 +1218,9 @@ mod tests {
                !shard_census.expected_part_numbers.is_empty(),
                "chaos objects must use physical part shards rather than inline data: {shard_census:?}"
            );
-            expected_manifests.push(PhysicalObjectManifest {
-                key,
-                payload_seed,
-                shard_census,
-            });
+            created_online_objects.push((key.clone(), payload_seed));
+            expected_manifests.push(PhysicalObjectManifest { key, shard_census });
+            attempt_count += 1;
        }

        let expected_pool_metadata = if background_enabled {
@@ -1169,31 +1266,38 @@ mod tests {
            if node_index == 1 {
                continue;
            }
-            let census = census_object_version_on_disk(Path::new(&node.data_dir), bucket, outage_key, None)?;
-            assert!(
-                census.is_complete(),
-                "online node {node_index} must hold a complete outage-object shard: {census:?}"
-            );
-            let erasure_index = census
-                .erasure_index
-                .ok_or_else(|| format!("online node {node_index} outage-object shard has no erasure index: {census:?}"))?;
-            assert!(
-                (1..=cluster.nodes.len()).contains(&erasure_index),
-                "online node {node_index} outage-object erasure index is out of range: {census:?}"
-            );
-            assert!(
-                outage_peer_erasure_indices.insert(erasure_index),
-                "outage-object erasure index {erasure_index} is duplicated across online nodes"
-            );
+            for (drive_index, drive) in node.data_dirs.iter().enumerate() {
+                let census = census_object_version_on_disk(Path::new(drive), bucket, outage_key, None)?;
+                assert!(
+                    census.is_complete(),
+                    "online node {node_index} drive {drive_index} must hold a complete outage-object shard: {census:?}"
+                );
+                let erasure_index = census.erasure_index.ok_or_else(|| {
+                    format!("online node {node_index} drive {drive_index} outage-object shard has no erasure index: {census:?}")
+                })?;
+                assert!(
+                    (1..=topology.total_drives()).contains(&erasure_index),
+                    "online node {node_index} drive {drive_index} outage-object erasure index is out of range: {census:?}"
+                );
+                assert!(
+                    outage_peer_erasure_indices.insert(erasure_index),
+                    "outage-object erasure index {erasure_index} is duplicated across online drives"
+                );
+            }
        }
        assert_eq!(
            outage_peer_erasure_indices.len(),
-            cluster.nodes.len().saturating_sub(1),
-            "every online node must contribute one unique outage-object erasure index"
+            topology.total_drives().saturating_sub(cluster.nodes[1].data_dirs.len()),
+            "every online drive must contribute one unique outage-object erasure index"
+        );
+        let missing_outage_erasure_indices = (1..=topology.total_drives())
+            .filter(|index| !outage_peer_erasure_indices.contains(index))
+            .collect::<HashSet<_>>();
+        assert_eq!(
+            missing_outage_erasure_indices.len(),
+            cluster.nodes[1].data_dirs.len(),
+            "the stopped node must account for every missing outage-object erasure index"
        );
-        let expected_outage_target_erasure_index = (1..=cluster.nodes.len())
-            .find(|index| !outage_peer_erasure_indices.contains(index))
-            .ok_or("online outage-object shards leave no erasure index for the replacement target")?;

        let heal_body = r#"{"recursive":true,"dryRun":false,"remove":false,"recreate":true,"scanMode":2,"updateParity":false,"nolock":false}"#;
        if !background_enabled {
@@ -1463,7 +1567,12 @@ mod tests {
                "Restored target endpoint forwarding"
            );
        } else {
-            if scenario == InterruptionScenario::BackgroundTargetRestart {
+            if matches!(
+                scenario,
+                InterruptionScenario::BackgroundTargetRestart
+                    | InterruptionScenario::BackgroundTargetRestartEc84
+                    | InterruptionScenario::BackgroundCoordinatorRestart
+            ) {
                cluster.stop_node_gracefully(interruption_node).await?;
            } else {
                cluster.stop_node(interruption_node)?;
@@ -1485,7 +1594,10 @@ mod tests {
            if background_enabled {
                let marker_exists = unclean_shutdown_marker.is_file();
                unclean_shutdown_marker_observed = Some(marker_exists);
-                let expected_marker = !matches!(scenario, InterruptionScenario::BackgroundTargetRestart);
+                let expected_marker = matches!(
+                    scenario,
+                    InterruptionScenario::BackgroundTargetCrash | InterruptionScenario::BackgroundTargetCrashEc84
+                );
                assert!(
                    marker_exists == expected_marker,
                    "background restart/crash lane observed unexpected unclean-shutdown marker state"
@@ -1586,29 +1698,35 @@ mod tests {
            "outage object must have a complete target shard: {outage_census:?}"
        );
        assert_eq!(
+            outage_census
+                .erasure_index
+                .filter(|index| missing_outage_erasure_indices.contains(index)),
            outage_census.erasure_index,
-            Some(expected_outage_target_erasure_index),
-            "the outage object must be rebuilt into its own missing erasure slot"
+            "the outage object must be rebuilt into one of the stopped node's missing erasure slots"
        );

        if let Some(cycle_end) = scanner_cycle_floor {
            wait_for_scanner_cycle_after(&cluster, cycle_end).await?;
        }

-        let mut expected_keys = expected_manifests
+        let mut expected_keys = created_online_objects
            .iter()
-            .map(|manifest| manifest.key.clone())
+            .map(|(key, _)| key.clone())
            .collect::<HashSet<_>>();
        assert!(expected_keys.insert(outage_key.to_string()));
        let node_listings = assert_all_nodes_list_exact_keys(&clients, bucket, &expected_keys).await?;

        let target_client = cluster.create_s3_client(1)?;
-        for expected in &expected_manifests {
-            let response = target_client.get_object().bucket(bucket).key(&expected.key).send().await?;
+        for (key, payload_seed) in &created_online_objects {
+            let response = target_client.get_object().bucket(bucket).key(key).send().await?;
            let actual = response.body.collect().await?.into_bytes();
-            let expected_body = deterministic_object_body(object_size_bytes, expected.payload_seed);
-            assert_eq!(actual.as_ref(), expected_body.as_slice(), "object body changed for {}", expected.key);
-            if evidence_run.is_some() {
+            let expected_body = deterministic_object_body(object_size_bytes, *payload_seed);
+            assert_eq!(actual.as_ref(), expected_body.as_slice(), "object body changed for {key}");
+            if evidence_run.is_some()
+                && let Some(expected) = expected_manifests
+                    .iter()
+                    .find(|manifest| manifest.key.as_str() == key.as_str())
+            {
                evidence_objects.push(serde_json::json!({
                    "key": expected.key, "version_id": expected.shard_census.version_id,
                    "expected_bytes": expected_body.len(), "actual_bytes": actual.len(),
@@ -1651,33 +1769,22 @@ mod tests {
        let task_status_body = signed_admin_post(&task_status_url, None, &cluster.access_key, &cluster.secret_key).await?;
        let task_status: serde_json::Value = serde_json::from_str(&task_status_body)
            .map_err(|err| format!("heal task status is not JSON ({err}): {task_status_body}"))?;
+        if task_status["summary"].as_str() != Some("finished") {
+            return Err(format!("heal data rebuilt but task did not finish successfully: {task_status}").into());
+        }
        if interruption_node == 0 {
-            // Admin tasks are process-local. Physical and queue convergence
-            // above establish recovery; a lost task must not report success.
-            assert_eq!(
-                task_status["summary"].as_str(),
-                Some("notFound"),
-                "interrupted task status: {task_status}"
-            );
-            assert_eq!(
-                task_status["detail"].as_str(),
-                Some("heal task not found or expired"),
-                "interrupted admin task must be explicitly unavailable: {task_status}"
-            );
+            // Restart recovery must finish the original durable root request.
            info!(
                event = "heal_interruption_recovered",
                component = "e2e_test",
                subsystem = "heal",
                interruption_node,
                interruption_kind,
-                task_state = "not_found",
-                "Physical recovery completed after coordinator restart"
+                task_state = "finished",
+                "Original root heal completed after coordinator restart"
            );
            return Ok(());
        }
-        if task_status["summary"].as_str() != Some("finished") {
-            return Err(format!("heal data rebuilt but task did not finish successfully: {task_status}").into());
-        }

        if let Some(evidence_context) = evidence_run {
            let restarted_pid = cluster.nodes[1].process.as_ref().ok_or("restarted target is absent")?.id();
@@ -109,10 +109,10 @@ async fn test_kms_key_directory_unavailable() -> Result<(), Box<dyn std::error::
        .await;

    let unavailable_error = put_result2.expect_err("a missing Local KMS key directory must reject encrypted writes");
-    assert_eq!(unavailable_error.raw_response().map(|response| response.status().as_u16()), Some(500));
+    assert_eq!(unavailable_error.raw_response().map(|response| response.status().as_u16()), Some(503));
    assert_eq!(
        unavailable_error.as_service_error().and_then(ProvideErrorMetadata::code),
-        Some("InternalError")
+        Some("ServiceUnavailable")
    );
    let unavailable_absence = s3_client
        .get_object()
@@ -15,6 +15,7 @@
 //! Regression coverage for anonymous access on multipart control APIs.

 use crate::common::{RustFSTestEnvironment, init_logging, local_http_client};
+use crate::kms::common::LocalKMSTestEnvironment;
 use async_compression::tokio::write::{BzEncoder, Lz4Encoder, XzEncoder};
 use aws_sdk_s3::error::{ProvideErrorMetadata, SdkError};
 use aws_sdk_s3::operation::head_object::HeadObjectOutput;
@@ -1465,10 +1466,10 @@ async fn test_anonymous_post_object_uses_bucket_default_sse_s3() -> Result<(), B
 async fn test_anonymous_post_object_uses_bucket_default_sse_kms() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
    init_logging();

-    let mut env = RustFSTestEnvironment::new().await?;
-    let master_key = local_sse_master_key_value();
-    env.start_rustfs_server_with_env(vec![], &[(LOCAL_SSE_MASTER_KEY_ENV, master_key.as_str())])
-        .await?;
+    let mut kms_env = LocalKMSTestEnvironment::new().await?;
+    let default_key_id = kms_env.start_rustfs_for_local_kms().await?;
+    kms_env.wait_for_kms_ready().await?;
+    let env = &kms_env.base_env;

    let bucket = "anon-post-default-sse-kms";
    let object_key = "post-default-sse-kms-object.txt";
@@ -1484,7 +1485,7 @@ async fn test_anonymous_post_object_uses_bucket_default_sse_kms() -> Result<(),
                .apply_server_side_encryption_by_default(
                    ServerSideEncryptionByDefault::builder()
                        .sse_algorithm(ServerSideEncryption::AwsKms)
-                        .kms_master_key_id("test-key")
+                        .kms_master_key_id(default_key_id)
                        .build()
                        .expect("default encryption rule should build"),
                )
@@ -25,6 +25,123 @@ const KEY: &str = "thumb/79/concurrent-overwrite.jpg";

 type TestResult = Result<(), Box<dyn std::error::Error + Send + Sync>>;

+async fn assert_degraded_cluster_publication_guard_errors_are_retryable() -> TestResult {
+    let mut cluster = RustFSTestClusterEnvironment::new(4).await?;
+    cluster.set_env("RUSTFS_STORAGE_CLASS_STANDARD", "EC:2");
+    cluster.set_env("RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE", "false");
+    cluster.set_env("RUSTFS_OBS_METRICS_EXPORT_ENABLED", "false");
+    cluster.set_env("RUST_LOG", "warn");
+    cluster.start().await?;
+    cluster.create_test_bucket(BUCKET).await?;
+    let clients: Vec<_> = cluster
+        .create_all_clients()?
+        .into_iter()
+        .map(|client| {
+            Client::from_conf(
+                client
+                    .config()
+                    .to_builder()
+                    .retry_config(aws_sdk_s3::config::retry::RetryConfig::standard().with_max_attempts(1))
+                    .build(),
+            )
+        })
+        .collect();
+
+    for alive in (1..=4).rev() {
+        if alive < 4 {
+            cluster.stop_node(alive)?;
+        }
+        for (node, client) in clients.iter().take(alive).enumerate() {
+            let key = format!("publication-put-{alive}-{node}");
+            let put = client
+                .put_object()
+                .bucket(BUCKET)
+                .key(&key)
+                .body(Bytes::from_static(b"publication guard regression").into())
+                .send()
+                .await;
+            if alive >= 3 {
+                put?;
+            } else {
+                let err = put.expect_err("PUT must reject writes without a write quorum");
+                assert_eq!(
+                    err.raw_response().map(|response| response.status().as_u16()),
+                    Some(503),
+                    "PUT with {alive} nodes alive, requested through node {node}: {err:?}"
+                );
+                assert_eq!(
+                    err.as_service_error().and_then(|error| error.meta().code()),
+                    Some("ServiceUnavailable"),
+                    "PUT with {alive} nodes alive, requested through node {node}: {err:?}"
+                );
+            }
+
+            let multipart_key = format!("publication-multipart-{alive}-{node}");
+            let multipart = client
+                .create_multipart_upload()
+                .bucket(BUCKET)
+                .key(&multipart_key)
+                .send()
+                .await;
+            if alive >= 3 {
+                let upload = multipart?;
+                let upload_id = upload
+                    .upload_id()
+                    .expect("successful multipart initialization must return an upload ID");
+                client
+                    .abort_multipart_upload()
+                    .bucket(BUCKET)
+                    .key(&multipart_key)
+                    .upload_id(upload_id)
+                    .send()
+                    .await?;
+            } else {
+                let err = multipart.expect_err("multipart initialization must reject writes without a write quorum");
+                assert_eq!(
+                    err.raw_response().map(|response| response.status().as_u16()),
+                    Some(503),
+                    "CreateMultipartUpload with {alive} nodes alive, requested through node {node}: {err:?}"
+                );
+                assert_eq!(
+                    err.as_service_error().and_then(|error| error.meta().code()),
+                    Some("ServiceUnavailable"),
+                    "CreateMultipartUpload with {alive} nodes alive, requested through node {node}: {err:?}"
+                );
+            }
+        }
+    }
+
+    cluster.stop();
+    cluster.start().await?;
+    for client in &clients {
+        for alive in [1, 3, 4] {
+            for node in 0..alive {
+                let key = format!("publication-put-{alive}-{node}");
+                let get = client.get_object().bucket(BUCKET).key(key).send().await;
+                if alive >= 3 {
+                    assert_eq!(
+                        get?.body.collect().await?.into_bytes().as_ref(),
+                        b"publication guard regression",
+                        "acknowledged writes must survive restart"
+                    );
+                } else {
+                    let err = get.expect_err("a rejected publication guard must not publish an object");
+                    assert_eq!(err.as_service_error().and_then(|error| error.meta().code()), Some("NoSuchKey"));
+                }
+            }
+        }
+        let uploads = client.list_multipart_uploads().bucket(BUCKET).send().await?;
+        assert!(
+            uploads
+                .uploads()
+                .iter()
+                .all(|upload| upload.key() != Some("publication-multipart-1-0")),
+            "a rejected publication guard must not publish a multipart upload"
+        );
+    }
+    Ok(())
+}
+
 async fn put_object(client: Client, payload: Vec<u8>, writer_id: usize) -> Result<(), String> {
    client
        .put_object()
@@ -125,6 +242,8 @@ async fn test_concurrent_cluster_overwrites_do_not_fail_namespace_lock_quorum()
 /// Before the fix, `map_namespace_lock_error` wrapped lock timeout/conflict errors as
 /// `StorageError::other(...)` → `StorageError::Io(...)`, which fell through to
 /// `S3ErrorCode::InternalError` (500) in the error mapping.
+/// Also checks PUT and multipart initialization when node failures prevent
+/// acquiring a table publication guard.
 #[tokio::test]
 async fn test_concurrent_put_same_key_never_returns_500() -> TestResult {
    crate::common::init_logging();
@@ -227,5 +346,6 @@ async fn test_concurrent_put_same_key_never_returns_500() -> TestResult {
    );

    clients[0].delete_object().bucket(BUCKET).key(KEY).send().await?;
-    Ok(())
+    cluster.stop();
+    assert_degraded_cluster_publication_guard_errors_are_retryable().await
 }
@@ -32,8 +32,10 @@ use aws_sdk_s3::types::{
    VersioningConfiguration,
 };
 use http::{Method, StatusCode};
+use serde_json::Value;
+use std::io::Write;
 use std::path::{Path, PathBuf};
-use std::time::Duration;
+use std::time::{Duration, SystemTime, UNIX_EPOCH};
 use tokio::task::JoinSet;
 use tokio::time::{Instant, sleep};

@@ -41,6 +43,7 @@ type TestResult = Result<(), Box<dyn std::error::Error + Send + Sync>>;
 type BoxError = Box<dyn std::error::Error + Send + Sync>;

 const SOURCE_BINARY_ENV: &str = "RUSTFS_UPGRADE_SOURCE_BINARY";
+const G09_EVIDENCE_DIR_ENV: &str = "RUSTFS_SCANNER_HEAL_G09_EVIDENCE_DIR";
 const RC5_COMMIT: &str = "40a2470feb567201165a5b809b7598bb4b1f68f5";
 const SSE_MASTER_KEY_ENV: &str = "RUSTFS_SSE_S3_MASTER_KEY";
 const SSE_MASTER_KEY: &str = "QkJCQkJCQkJCQkJCQkJCQkJCQkJCQkJCQkJCQkJCQkI=";
@@ -81,6 +84,14 @@ const QUOTA_READINESS_TIMEOUT: Duration = Duration::from_secs(30);
 // of treating it as an upgrade failure.
 const QUOTA_ADMISSION_WARMUP_TIMEOUT: Duration = Duration::from_secs(90);

+struct G09EvidenceContext {
+    directory: PathBuf,
+    current_revision: String,
+    previous_revision: String,
+    run_id: String,
+    measurement_window_id: String,
+}
+
 fn source_binary() -> Result<PathBuf, Box<dyn std::error::Error + Send + Sync>> {
    let path = std::env::var_os(SOURCE_BINARY_ENV)
        .map(PathBuf::from)
@@ -91,6 +102,155 @@ fn source_binary() -> Result<PathBuf, Box<dyn std::error::Error + Send + Sync>>
    Ok(path)
 }

+fn is_lower_hex_revision(value: &str) -> bool {
+    value.len() == 40
+        && value
+            .bytes()
+            .all(|byte| byte.is_ascii_hexdigit() && !byte.is_ascii_uppercase())
+}
+
+fn current_source_revision() -> Result<String, BoxError> {
+    let revision = env!("RUSTFS_E2E_BUILD_COMMIT");
+    if !is_lower_hex_revision(revision) {
+        return Err(format!("current test binary has invalid source revision {revision}").into());
+    }
+    Ok(revision.to_string())
+}
+
+async fn binary_source_revision(binary: &Path) -> Result<String, BoxError> {
+    let output = tokio::process::Command::new(binary).arg("--version").output().await?;
+    if !output.status.success() {
+        return Err(format!("{} --version failed with {}", binary.display(), output.status).into());
+    }
+    let version = String::from_utf8(output.stdout)?;
+    source_revision_from_version_output(&version, binary)
+}
+
+fn source_revision_from_version_output(version: &str, binary: &Path) -> Result<String, BoxError> {
+    version
+        .split(|ch: char| !ch.is_ascii_hexdigit())
+        .find(|token| is_lower_hex_revision(token))
+        .map(str::to_string)
+        .ok_or_else(|| format!("{} --version did not expose a 40-byte source revision", binary.display()).into())
+}
+
+async fn g09_evidence_context(previous_binary: &Path) -> Result<Option<G09EvidenceContext>, BoxError> {
+    let Some(directory) = std::env::var_os(G09_EVIDENCE_DIR_ENV) else {
+        return Ok(None);
+    };
+    let directory = PathBuf::from(directory);
+    std::fs::create_dir_all(&directory)?;
+    let current_revision = current_source_revision()?;
+    let previous_revision = binary_source_revision(previous_binary).await?;
+    if current_revision == previous_revision {
+        return Err("G09 mixed-version evidence requires distinct current and previous source revisions".into());
+    }
+    let now = SystemTime::now().duration_since(UNIX_EPOCH)?.as_nanos();
+    Ok(Some(G09EvidenceContext {
+        directory,
+        current_revision,
+        previous_revision,
+        run_id: format!("g09-upgrade-{}-{now}", std::process::id()),
+        measurement_window_id: format!("g09-mixed-version-window-{now}"),
+    }))
+}
+
+fn write_g09_evidence(
+    context: &G09EvidenceContext,
+    field: &str,
+    role: &str,
+    cases: &[&str],
+    test: &str,
+    details: Value,
+) -> TestResult {
+    let path = context.directory.join(format!("G09-{field}.json"));
+    let mut evidence = serde_json::json!({
+        "schema": 1,
+        "evidence_type": "measured",
+        "artifact_kind": "upgrade-compatibility-e2e",
+        "source_revision": context.current_revision,
+        "run_id": context.run_id,
+        "measurement_window_id": context.measurement_window_id,
+        "gate": "G09",
+        "field": field,
+        "versions": [context.previous_revision, context.current_revision],
+        "current_revision": context.current_revision,
+        "previous_revision": context.previous_revision,
+        "mixed_version_role": role,
+        "mixed_version_cases": cases,
+        "test": test,
+        "details": details,
+    });
+    if field == "rollback_payload_evidence" {
+        evidence["rollback_payload_replayed"] = Value::Bool(true);
+    }
+    let data = serde_json::to_vec_pretty(&evidence)?;
+    if data.len() > 1024 * 1024 {
+        return Err("G09 mixed-version evidence exceeds the 1 MiB artifact budget".into());
+    }
+    let mut output = std::fs::OpenOptions::new().write(true).create_new(true).open(path)?;
+    output.write_all(&data)?;
+    output.write_all(b"\n")?;
+    output.sync_all()?;
+    Ok(())
+}
+
+#[cfg(test)]
+mod g09_evidence_tests {
+    use super::*;
+    use tempfile::TempDir;
+
+    #[test]
+    fn source_revision_parser_requires_lowercase_forty_byte_sha() -> TestResult {
+        let binary = Path::new("rustfs");
+        assert_eq!(
+            source_revision_from_version_output("rustfs 1.0.0 abcdef0123456789abcdef0123456789abcdef01 clean", binary)?,
+            "abcdef0123456789abcdef0123456789abcdef01"
+        );
+        assert!(source_revision_from_version_output("rustfs ABCDEF0123456789ABCDEF0123456789ABCDEF01", binary).is_err());
+        assert!(source_revision_from_version_output("rustfs abcdef", binary).is_err());
+        Ok(())
+    }
+
+    #[test]
+    fn g09_evidence_writer_records_measured_role_and_refuses_overwrite() -> TestResult {
+        let directory = TempDir::new()?;
+        let context = G09EvidenceContext {
+            directory: directory.path().to_path_buf(),
+            current_revision: "b".repeat(40),
+            previous_revision: "a".repeat(40),
+            run_id: "g09-upgrade-test-run".to_string(),
+            measurement_window_id: "g09-upgrade-test-window".to_string(),
+        };
+        write_g09_evidence(
+            &context,
+            "mixed_version_reader_evidence",
+            "mixed-version-reader",
+            &["old-writer-new-reader", "new-writer-old-reader"],
+            "unit",
+            serde_json::json!({"assertions": ["reader evidence"]}),
+        )?;
+        let path = directory.path().join("G09-mixed_version_reader_evidence.json");
+        let evidence: Value = serde_json::from_slice(&std::fs::read(path)?)?;
+        assert_eq!(evidence["evidence_type"], "measured");
+        assert_eq!(evidence["gate"], "G09");
+        assert_eq!(evidence["field"], "mixed_version_reader_evidence");
+        assert_eq!(evidence["mixed_version_role"], "mixed-version-reader");
+        assert_eq!(evidence["versions"], serde_json::json!(["a".repeat(40), "b".repeat(40)]));
+
+        let overwrite = write_g09_evidence(
+            &context,
+            "mixed_version_reader_evidence",
+            "mixed-version-reader",
+            &["old-writer-new-reader", "new-writer-old-reader"],
+            "unit",
+            serde_json::json!({}),
+        );
+        assert!(overwrite.is_err(), "G09 evidence must not overwrite an existing artifact");
+        Ok(())
+    }
+}
+
 async fn enable_versioning(client: &Client, bucket: &str) -> TestResult {
    let configuration = VersioningConfiguration::builder()
        .status(BucketVersioningStatus::Enabled)
@@ -523,6 +683,7 @@ async fn direct_upgrade_from_rc2_preserves_object_contracts() -> TestResult {
 async fn rolling_upgrade_from_rc2_preserves_mixed_version_contracts() -> TestResult {
    init_logging();
    let previous_binary = source_binary()?;
+    let evidence_context = g09_evidence_context(&previous_binary).await?;
    let current_binary = rustfs_binary_path();
    let mut cluster = RustFSTestClusterEnvironment::new(MIXED_NODE_COUNT).await?;
    cluster.set_env("RUST_LOG", "rustfs=warn,rustfs_notify=warn");
@@ -555,6 +716,45 @@ async fn rolling_upgrade_from_rc2_preserves_mixed_version_contracts() -> TestRes
        }
    }

+    if let Some(context) = evidence_context.as_ref() {
+        let phases = ["one-current-node", "one-previous-node"];
+        let objects_per_phase = MULTIPART_WORKERS * MULTIPART_UPLOADS_PER_WORKER + 2;
+        write_g09_evidence(
+            context,
+            "mixed_version_reader_evidence",
+            "mixed-version-reader",
+            &["old-writer-new-reader", "new-writer-old-reader"],
+            "upgrade_compatibility_test::rolling_upgrade_from_rc2_preserves_mixed_version_contracts",
+            serde_json::json!({
+                "bucket": MIXED_BUCKET,
+                "phases": phases,
+                "objects_per_phase": objects_per_phase,
+                "assertions": [
+                    "current node reads objects written through previous-release client",
+                    "previous-release node reads objects written through current client",
+                    "all nodes list every mixed-version object after homogeneous-current convergence"
+                ],
+            }),
+        )?;
+        write_g09_evidence(
+            context,
+            "mixed_version_writer_evidence",
+            "mixed-version-writer",
+            &["old-reader-new-writer", "new-reader-old-writer"],
+            "upgrade_compatibility_test::rolling_upgrade_from_rc2_preserves_mixed_version_contracts",
+            serde_json::json!({
+                "bucket": MIXED_BUCKET,
+                "phases": phases,
+                "objects_per_phase": objects_per_phase,
+                "assertions": [
+                    "current writer publishes objects readable by previous-release node",
+                    "previous-release writer publishes objects readable by current node",
+                    "multipart writers continue under one-current-node and one-previous-node layouts"
+                ],
+            }),
+        )?;
+    }
+
    Ok(())
 }

@@ -1136,6 +1336,7 @@ async fn direct_upgrade_from_previous_release_preserves_bucket_configuration() -
 async fn rollback_to_previous_release_reads_current_bucket_metadata() -> TestResult {
    init_logging();
    let previous_binary = source_binary()?;
+    let evidence_context = g09_evidence_context(&previous_binary).await?;

    let replication_target = FakeS3Target::start().await?;
    replication_target.create_bucket(ROLLBACK_REPLICA_BUCKET);
@@ -1204,6 +1405,54 @@ async fn rollback_to_previous_release_reads_current_bucket_metadata() -> TestRes
    );
    assert_eq!(body, post_rollback_bytes);

+    env.restart_server_preserving_data(vec![], &server_env).await?;
+    let current_again = env.create_s3_client();
+    assert_versioning_enabled(&current_again, ROLLBACK_BUCKET, "after rolling forward again").await?;
+    assert_default_sse_s3_encryption(&current_again, ROLLBACK_BUCKET, "after rolling forward again").await?;
+    assert_bucket_tag(&current_again, ROLLBACK_BUCKET, "after rolling forward again").await?;
+    assert_remote_target_preserved(&env, ROLLBACK_BUCKET, &target_arn, "after rolling forward again").await?;
+    assert_eq!(
+        read_object(&current_again, ROLLBACK_BUCKET, single_key, Some(&single_version))
+            .await?
+            .1,
+        single_bytes
+    );
+    assert_eq!(
+        read_object(&current_again, ROLLBACK_BUCKET, multipart_key, None).await?.1,
+        multipart_bytes
+    );
+    assert_eq!(
+        read_object(&current_again, ROLLBACK_BUCKET, post_rollback_key, None).await?.1,
+        post_rollback_bytes
+    );
+
+    if let Some(context) = evidence_context.as_ref() {
+        write_g09_evidence(
+            context,
+            "rollback_payload_evidence",
+            "rollback-payload",
+            &["rollback-to-old", "rollback-to-new", "unknown-field-retained"],
+            "upgrade_compatibility_test::rollback_to_previous_release_reads_current_bucket_metadata",
+            serde_json::json!({
+                "bucket": ROLLBACK_BUCKET,
+                "cases": {
+                    "rollback-to-old": [
+                        "previous-release binary reads current-build versioning, SSE-S3, tags, replication target and objects",
+                        "previous-release writer honors the decoded current-build encryption configuration"
+                    ],
+                    "rollback-to-new": [
+                        "current build reads the object written by the rolled-back previous release",
+                        "current build reads the current-build single-part and multipart objects after rolling forward again"
+                    ],
+                    "unknown-field-retained": [
+                        "previous release skips current-build bucket metadata extension fields without dropping known bucket configuration",
+                        "current build reads the retained bucket configuration after the previous-release round trip"
+                    ]
+                },
+            }),
+        )?;
+    }
+
    replication_target.shutdown().await;
    Ok(())
 }
@@ -563,7 +563,7 @@ pub mod store_list {
 }

 pub mod storage {
-    pub use crate::core::pools::HealLifecycleExpiryContext;
+    pub use crate::core::pools::{HealLifecycleExpiryContext, POOL_META_NAME};
    pub use crate::store::HealWalkVersion;
    pub use crate::store::{
        BootstrapLocalTarget, ECStore, SCANNER_PUBLICATION_LEASE_TTL_MS, ScannerDataMovementPauseStatus, all_local_disk,
@@ -5165,6 +5165,9 @@ pub async fn put_restore_opts(
        user_defined: meta,
        version_id: oi.version_id.map(|e| e.to_string()),
        mod_time: oi.mod_time,
+        // Restore writes stored (possibly encrypted) bytes, so the writer's
+        // computed MD5 is not the object's public plaintext ETag.
+        preserve_etag: oi.etag.clone(),
        //expires:           oi.expires,
        ..Default::default()
    })
@@ -285,6 +285,20 @@ fn peer_replay_state(audience: &str) -> PeerReplayState {
        .unwrap_or_default()
 }

+pub(crate) fn clear_peer_replay_state_for_addr(addr: &str) -> std::io::Result<()> {
+    let uri = addr
+        .parse::<Uri>()
+        .map_err(|_| std::io::Error::other("Invalid gRPC peer URI"))?;
+    let audience = uri
+        .authority()
+        .map(|authority| normalize_tonic_rpc_audience(authority.as_str()))
+        .ok_or_else(|| std::io::Error::other("Missing gRPC peer authority"))??;
+    if let Ok(mut states) = PEER_REPLAY_STATES.lock() {
+        states.remove(&audience);
+    }
+    Ok(())
+}
+
 fn apply_peer_replay_response(
    audience: String,
    sent_state: PeerReplayState,
@@ -619,6 +633,13 @@ mod tests {
            .remove(audience);
    }

+    fn set_peer_capability(audience: &str, state: PeerReplayState) {
+        PEER_REPLAY_STATES
+            .lock()
+            .expect("peer capability cache lock must not be poisoned")
+            .insert(audience.to_string(), state);
+    }
+
    fn rolling_mutation_request(method: &'static str) -> tonic::Request<()> {
        let mut request = tonic::Request::new(rustfs_protos::proto_gen::node_service::GenerallyLockRequest {
            args: "canonical mutation request".to_string(),
@@ -1090,6 +1111,23 @@ mod tests {
        clear_peer_capability(audience);
    }

+    #[test]
+    fn clear_peer_replay_state_for_addr_removes_normalized_audience() {
+        let audience = "clear-peer-replay-state-test:9000";
+        let boot_epoch = Uuid::new_v4();
+        set_peer_capability(
+            audience,
+            PeerReplayState {
+                boot_epoch: Some(boot_epoch),
+                cache_capability: Some(PeerReplayCapability::Capable { boot_epoch }),
+            },
+        );
+
+        clear_peer_replay_state_for_addr("http://clear-peer-replay-state-test:9000").expect("peer URI should clear replay state");
+
+        assert_eq!(peer_replay_state(audience), PeerReplayState::default());
+    }
+
    #[test]
    fn interceptor_snapshot_prevents_delayed_legacy_response_from_revoking_capability() {
        ensure_test_rpc_secret();
@@ -13,8 +13,9 @@
 // limitations under the License.

 use crate::cluster::rpc::client::{
-    AuthenticatedChannel, TonicInterceptor, embedded_tonic_status, gen_tonic_signature_interceptor, heal_control_time_out_client,
-    is_network_like_status, message_has_network_needle, node_service_time_out_client, tier_mutation_control_time_out_client,
+    AuthenticatedChannel, TonicInterceptor, clear_peer_replay_state_for_addr, embedded_tonic_status,
+    gen_tonic_signature_interceptor, heal_control_time_out_client, is_network_like_status, message_has_network_needle,
+    node_service_time_out_client, tier_mutation_control_time_out_client,
 };
 use crate::cluster::rpc::{set_tonic_canonical_body_digest, set_tonic_mutation_body_digest, verify_tonic_rpc_response_proof};
 use crate::error::{Error, Result};
@@ -544,6 +545,16 @@ fn validate_heal_control_response_proof(canonical_response: &[u8], proof: &[u8])
        .map_err(|_| Error::other("peer returned an invalid heal control response proof"))
 }

+fn heal_control_auth_may_need_replay_scope_refresh(err: &Error) -> bool {
+    matches!(
+        err,
+        Error::Io(io_err)
+            if embedded_tonic_status(io_err).is_some_and(|status| {
+                status.code() == tonic::Code::Unauthenticated && status.message() == "No valid auth token"
+            })
+    )
+}
+
 fn decode_remote_version_state_capability(expected_member: &str, result: &[u8]) -> Result<Uuid> {
    let (topology_member, process_epoch) = rustfs_protos::decode_remote_version_state_capability(result).map_err(Error::other)?;
    if topology_member != expected_member {
@@ -1720,45 +1731,72 @@ impl PeerRestClient {
            return Err(Error::other("heal control command exceeds size limit"));
        }
        let capability_probe = rustfs_protos::is_heal_control_capability_probe(&command);
-        self.finalize_result(
-            async {
-                let mut client = self
-                    .get_heal_control_client()
-                    .await?
-                    .max_encoding_message_size(rustfs_protos::HEAL_CONTROL_RPC_MAX_MESSAGE_SIZE)
-                    .max_decoding_message_size(rustfs_protos::HEAL_CONTROL_RPC_MAX_MESSAGE_SIZE);
-                let canonical_body = rustfs_protos::canonical_heal_control_request_body(version, &topology_fingerprint, &command)
-                    .map_err(|_| Error::other("heal control request length cannot be represented"))?;
-                let mut request = Request::new(HealControlRequest {
-                    version,
-                    topology_fingerprint: topology_fingerprint.clone(),
-                    command: command.clone().into(),
-                });
-                request.set_timeout(rustfs_protos::heal_control_execution_timeout());
-                set_tonic_canonical_body_digest(&mut request, &canonical_body)?;
-                let response = client.heal_control(request).await?.into_inner();
-                if !response.success {
-                    return Err(Error::other(
-                        response
-                            .error_info
-                            .unwrap_or_else(|| "peer heal control failed without an error".to_string()),
-                    ));
-                }
-                if !capability_probe {
-                    let canonical_response = rustfs_protos::canonical_heal_control_response_body(
-                        version,
-                        &topology_fingerprint,
-                        &command,
-                        &response.result,
-                    )
+        let result = self
+            .heal_control_once(version, &topology_fingerprint, &command, capability_probe)
+            .await;
+        if result
+            .as_ref()
+            .err()
+            .is_some_and(heal_control_auth_may_need_replay_scope_refresh)
+        {
+            self.prepare_heal_control_auth_retry().await;
+            return self
+                .finalize_result(
+                    self.heal_control_once(version, &topology_fingerprint, &command, capability_probe)
+                        .await,
+                )
+                .await;
+        }
+        self.finalize_result(result).await
+    }
+
+    async fn prepare_heal_control_auth_retry(&self) {
+        if let Err(err) = clear_peer_replay_state_for_addr(&self.grid_host) {
+            debug!(
+                peer = %self.grid_host,
+                error = %err,
+                "could not clear heal control replay state before retry"
+            );
+        }
+        self.evict_connection().await;
+    }
+
+    async fn heal_control_once(
+        &self,
+        version: u32,
+        topology_fingerprint: &str,
+        command: &[u8],
+        capability_probe: bool,
+    ) -> Result<Vec<u8>> {
+        let mut client = self
+            .get_heal_control_client()
+            .await?
+            .max_encoding_message_size(rustfs_protos::HEAL_CONTROL_RPC_MAX_MESSAGE_SIZE)
+            .max_decoding_message_size(rustfs_protos::HEAL_CONTROL_RPC_MAX_MESSAGE_SIZE);
+        let canonical_body = rustfs_protos::canonical_heal_control_request_body(version, topology_fingerprint, command)
+            .map_err(|_| Error::other("heal control request length cannot be represented"))?;
+        let mut request = Request::new(HealControlRequest {
+            version,
+            topology_fingerprint: topology_fingerprint.to_string(),
+            command: command.to_vec().into(),
+        });
+        request.set_timeout(rustfs_protos::heal_control_execution_timeout());
+        set_tonic_canonical_body_digest(&mut request, &canonical_body)?;
+        let response = client.heal_control(request).await?.into_inner();
+        if !response.success {
+            return Err(Error::other(
+                response
+                    .error_info
+                    .unwrap_or_else(|| "peer heal control failed without an error".to_string()),
+            ));
+        }
+        if !capability_probe {
+            let canonical_response =
+                rustfs_protos::canonical_heal_control_response_body(version, topology_fingerprint, command, &response.result)
                    .map_err(|_| Error::other("heal control response length cannot be represented"))?;
-                    validate_heal_control_response_proof(&canonical_response, &response.response_proof)?;
-                }
-                Ok(response.result.to_vec())
-            }
-            .await,
-        )
-        .await
+            validate_heal_control_response_proof(&canonical_response, &response.response_proof)?;
+        }
+        Ok(response.result.to_vec())
    }

    /// Confirms that a peer supports the current heal-control coordination
@@ -3728,6 +3766,22 @@ mod tests {
        }
    }

+    #[test]
+    fn heal_control_auth_retry_is_limited_to_transport_auth_rejection() {
+        assert!(heal_control_auth_may_need_replay_scope_refresh(&Error::from(
+            tonic::Status::unauthenticated("No valid auth token")
+        )));
+        assert!(!heal_control_auth_may_need_replay_scope_refresh(&Error::from(
+            tonic::Status::permission_denied("bad signature")
+        )));
+        assert!(!heal_control_auth_may_need_replay_scope_refresh(&Error::from(
+            tonic::Status::unauthenticated("application rejected heal control")
+        )));
+        assert!(!heal_control_auth_may_need_replay_scope_refresh(&Error::other(
+            "Io error: code: 'Unauthenticated', message: \"No valid auth token\""
+        )));
+    }
+
    #[test]
    fn peer_rest_client_network_classifier_keeps_slow_peers_online() {
        // The per-RPC channel deadline (RUSTFS_INTERNODE_RPC_TIMEOUT, 30s)
@@ -2606,6 +2606,24 @@ where
    usize::try_from(size).unwrap_or_default()
 }

+fn is_decommission_set_local_usage_cache(bucket: &str, object: &str) -> bool {
+    if bucket != RUSTFS_META_BUCKET {
+        return false;
+    }
+    let Some(path) = object
+        .strip_prefix(BUCKET_META_PREFIX)
+        .and_then(|path| path.strip_prefix('/'))
+    else {
+        return false;
+    };
+    let name = match path.rsplit_once('/') {
+        Some((bucket, name)) if !bucket.is_empty() && !bucket.contains('/') && bucket != "." && bucket != ".." => name,
+        Some(_) => return false,
+        None => path,
+    };
+    name.strip_suffix(".bkp").unwrap_or(name) == DATA_USAGE_CACHE_NAME
+}
+
 fn with_decommission_entry_context<E: Display>(stage: &str, bucket: &str, object: &str, err: E) -> Error {
    Error::other(format!("decommission entry {stage} failed for bucket {bucket} object {object}: {err}"))
 }
@@ -13294,6 +13312,12 @@ impl ECStore {
            );
            return Ok(DecommissionEntryAttemptOutcome::Complete);
        }
+        // Scanner caches describe their own erasure set and are rebuilt there.
+        // Copying one onto another set can overwrite unrelated cache contents or
+        // leave an unresolvable target-capacity intent after a conditional PUT.
+        if is_decommission_set_local_usage_cache(&bucket, &entry.name) {
+            return Ok(DecommissionEntryAttemptOutcome::Complete);
+        }
        let durable_ilm_record = if bucket == RUSTFS_META_BUCKET {
            classify_durable_ilm_record(&entry.name)
                .map_err(|err| with_decommission_entry_context("durable_ilm_namespace", &bucket, &entry.name, err))?
@@ -13851,7 +13875,7 @@ impl ECStore {

                let bucket = bucket.clone();

-                let rd = match set
+                let read_result = set
                    .get_object_reader(
                        bucket.as_str(),
                        &encode_dir_object(&version.name),
@@ -13859,8 +13883,11 @@ impl ECStore {
                        HeaderMap::new(),
                        &decommission_object_migration_read_opts(version_id.clone()),
                    )
-                    .await
-                {
+                    .await;
+                #[cfg(test)]
+                let read_result =
+                    decommission_test_wrap_result("object_read", &bucket, &version.name, version_attempt, read_result);
+                let rd = match read_result {
                    Ok(rd) => rd,
                    Err(err) => {
                        if is_err_object_not_found(&err) || is_err_version_not_found(&err) {
@@ -13869,15 +13896,6 @@ impl ECStore {
                            break;
                        }

-                        if !ignore {
-                            //
-                            if bucket == RUSTFS_META_BUCKET && version.name.contains(DATA_USAGE_CACHE_NAME) {
-                                ignore = true;
-                                error!("decommission_pool: ignore data usage cache {}", &version.name);
-                                break;
-                            }
-                        }
-
                        failure = true;
                        if version_attempt == DECOMMISSION_VERSION_COPY_ATTEMPTS {
                            error!(
@@ -16845,7 +16863,7 @@ impl ECStore {
                            return;
                        }

-                        if bucket_name == RUSTFS_META_BUCKET && entry.name.contains(DATA_USAGE_CACHE_NAME) {
+                        if is_decommission_set_local_usage_cache(&bucket_name, &entry.name) {
                            return;
                        }

@@ -17163,6 +17181,361 @@ mod tests {
    use crate::storage_api_contracts::multipart::MultipartOperations as _;
    use serde::Serialize;

+    #[test]
+    fn decommission_set_local_usage_cache_classification_is_exact() {
+        for object in [
+            "buckets/.usage-cache.bin",
+            "buckets/.usage-cache.bin.bkp",
+            "buckets/photos/.usage-cache.bin",
+            "buckets/photos/.usage-cache.bin.bkp",
+        ] {
+            assert!(is_decommission_set_local_usage_cache(RUSTFS_META_BUCKET, object), "{object}");
+            assert!(!is_decommission_set_local_usage_cache("user-bucket", object), "{object}");
+        }
+        for object in [
+            "buckets/.usage.v2.json",
+            "buckets/.usage.v2.json.bkp",
+            "buckets/.usage-cache.bin.extra",
+            "buckets/.usage-cache.bin.bkp.extra",
+            "buckets/prefix.usage-cache.bin",
+            "buckets/photos/.usage-cache.bin.bkp.bkp",
+            "buckets/photos/nested/.usage-cache.bin",
+            "buckets//.usage-cache.bin",
+            "buckets/../.usage-cache.bin",
+            "buckets/./.usage-cache.bin",
+            "buckets/.usage-cache.bin/child",
+            "config/.usage-cache.bin",
+            "buckets-other/.usage-cache.bin",
+            ".usage-cache.bin",
+        ] {
+            assert!(!is_decommission_set_local_usage_cache(RUSTFS_META_BUCKET, object), "{object}");
+        }
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn decommission_keeps_set_local_usage_caches_out_of_target_capacity() {
+        use crate::object_api::PutObjReader;
+        use tokio::io::AsyncReadExt as _;
+
+        // Keep the scenario's large setup and migration futures off the test
+        // future so ordinary metadata I/O retains the default thread stack.
+        let (_temp_dirs, store, _other_store) =
+            Box::pin(crate::services::rebalance::test_two_pool_stores_with_isolated_node_contexts(None)).await;
+        let user_bucket = "decommission-usage-cache-control";
+        Box::pin(store.make_bucket(user_bucket, &MakeBucketOptions::default()))
+            .await
+            .expect("create the ordinary-object control bucket");
+        let incarnation = Box::pin(store.bucket_incarnation_id(user_bucket))
+            .await
+            .expect("control bucket incarnation");
+        let source_time = OffsetDateTime::now_utc();
+        let cache_objects = [
+            "buckets/.usage-cache.bin",
+            "buckets/.usage-cache.bin.bkp",
+            "buckets/photos/.usage-cache.bin",
+            "buckets/photos/.usage-cache.bin.bkp",
+        ];
+        let conflict_object = "buckets/.usage-cache.bin.conflict";
+        let source_read_failure_object = "buckets/.usage-cache.bin.read-error";
+        let source_body = b"source set cache";
+        let target_body = b"independent older target set cache";
+        for object in cache_objects.into_iter().chain([conflict_object]) {
+            for (pool_index, body, mod_time) in [
+                (0, source_body.as_slice(), source_time),
+                (1, target_body.as_slice(), source_time - Duration::seconds(1)),
+            ] {
+                // Scanner cache persistence writes directly to its own set.
+                store.pools[pool_index]
+                    .get_disks_by_key(object)
+                    .put_object(
+                        RUSTFS_META_BUCKET,
+                        object,
+                        &mut PutObjReader::from_vec(body.to_vec()),
+                        &ObjectOptions {
+                            mod_time: Some(mod_time),
+                            ..Default::default()
+                        },
+                    )
+                    .await
+                    .expect("seed distinct native set-local objects");
+            }
+        }
+        let controls = [
+            (RUSTFS_META_BUCKET, "buckets/.usage.v2.json"),
+            (RUSTFS_META_BUCKET, "buckets/photos/.usage-cache.bin.extra"),
+            (user_bucket, "ordinary-object"),
+            (user_bucket, "buckets/.usage-cache.bin"),
+        ];
+        for (bucket, object) in controls {
+            store.pools[0]
+                .put_object(
+                    bucket,
+                    object,
+                    &mut PutObjReader::from_vec(b"ordinary object contents".to_vec()),
+                    &ObjectOptions {
+                        expected_bucket_incarnation_id: (bucket == user_bucket).then_some(incarnation),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("seed a control that must migrate");
+        }
+        store.pools[0]
+            .put_object(
+                RUSTFS_META_BUCKET,
+                source_read_failure_object,
+                &mut PutObjReader::from_vec(source_body.to_vec()),
+                &ObjectOptions::default(),
+            )
+            .await
+            .expect("seed a similarly named object whose source read will fail");
+
+        let layout = DecommissionErasureLayout { data: 1, parity: 0 };
+        set_decommission_capacity_info_overrides_for_test(
+            store.id,
+            vec![vec![
+                DecommissionPoolCapacityInfo::for_test(0, layout, 0, 16_384, 16_384),
+                DecommissionPoolCapacityInfo::for_test(1, layout, 131_072, 131_072, 0),
+            ]],
+        );
+        Box::pin(store.save_current_pool_meta_for_decommission_start(&[0], Vec::new()))
+            .await
+            .expect("activate the decommission capacity reservation");
+
+        for object in cache_objects {
+            Box::pin(store.decommission_entry_for_test(
+                0,
+                MetaCacheEntry {
+                    name: object.to_string(),
+                    ..Default::default()
+                },
+                RUSTFS_META_BUCKET.to_string(),
+                store.pools[0].get_disks_by_key(object),
+            ))
+            .await
+            .expect("set-local cache must not enter cross-pool migration");
+            for (pool_index, expected) in [(0, source_body.as_slice()), (1, target_body.as_slice())] {
+                let mut reader = store.pools[pool_index]
+                    .get_disks_by_key(object)
+                    .get_object_reader(RUSTFS_META_BUCKET, object, None, HeaderMap::new(), &ObjectOptions::default())
+                    .await
+                    .expect("each set must retain its own cache");
+                let mut actual = Vec::new();
+                reader.stream.read_to_end(&mut actual).await.expect("read retained cache");
+                assert_eq!(actual, expected, "pool {pool_index}, {object}");
+            }
+            let meta = store.pool_meta.read().await;
+            let info = meta.pools[0].decommission.as_ref().expect("decommission progress");
+            assert_eq!((info.items_decommissioned, info.items_decommission_failed), (0, 0));
+            assert_eq!((info.bytes_done, info.bytes_failed), (0, 0));
+            let reservation = info.capacity_reservation.as_ref().expect("capacity reservation");
+            assert_eq!(reservation.pending_target_physical_bytes, 0);
+            assert_eq!(reservation.consumed_target_physical_bytes, 0);
+            assert!(reservation.targets.iter().all(|target| target.pending_mutation_id.is_none()));
+        }
+        let mut persisted = PoolMeta::default();
+        Box::pin(persisted.load_no_lock_from_replicas(store.pools.clone()))
+            .await
+            .expect("reload durable capacity intents after cache entries");
+        let reservation = persisted.pools[0]
+            .decommission
+            .as_ref()
+            .and_then(|info| info.capacity_reservation.as_ref())
+            .expect("durable reservation");
+        assert_eq!(reservation.pending_target_physical_bytes, 0);
+        assert!(reservation.targets.iter().all(|target| target.pending_mutation_id.is_none()));
+
+        let injected_reads = Arc::new(AtomicUsize::new(0));
+        let observed_reads = Arc::clone(&injected_reads);
+        let read_fault = DecommissionTestFaultGuard::install(Arc::new(move |stage, bucket, object, _, success| {
+            if stage == "object_read" && bucket == RUSTFS_META_BUCKET && object == source_read_failure_object && success {
+                observed_reads.fetch_add(1, Ordering::SeqCst);
+                return true;
+            }
+            false
+        }));
+        Box::pin(store.decommission_entry_for_test(
+            0,
+            MetaCacheEntry {
+                name: source_read_failure_object.to_string(),
+                ..Default::default()
+            },
+            RUSTFS_META_BUCKET.to_string(),
+            store.pools[0].get_disks_by_key(source_read_failure_object),
+        ))
+        .await
+        .expect("entry must record the non-NotFound source read failure");
+        drop(read_fault);
+        assert_eq!(injected_reads.load(Ordering::SeqCst), DECOMMISSION_VERSION_COPY_ATTEMPTS);
+        {
+            let meta = store.pool_meta.read().await;
+            let info = meta.pools[0].decommission.as_ref().expect("source read failure progress");
+            assert_eq!((info.items_decommissioned, info.items_decommission_failed), (0, 1));
+            assert_eq!(info.bytes_failed, source_body.len());
+            assert_eq!(
+                info.capacity_reservation
+                    .as_ref()
+                    .expect("reservation")
+                    .pending_target_physical_bytes,
+                0
+            );
+        }
+        let mut retained = store.pools[0]
+            .get_object_reader(
+                RUSTFS_META_BUCKET,
+                source_read_failure_object,
+                None,
+                HeaderMap::new(),
+                &ObjectOptions::default(),
+            )
+            .await
+            .expect("source read failure must retain the source");
+        let mut retained_body = Vec::new();
+        retained
+            .stream
+            .read_to_end(&mut retained_body)
+            .await
+            .expect("read retained source");
+        assert_eq!(retained_body, source_body);
+        drop(retained);
+        let target_err = store.pools[1]
+            .get_object_info(RUSTFS_META_BUCKET, source_read_failure_object, &ObjectOptions::default())
+            .await
+            .expect_err("failed source read must not create a target object");
+        assert!(is_err_object_not_found(&target_err), "unexpected target state: {target_err:?}");
+
+        for (bucket, object) in controls {
+            Box::pin(store.decommission_entry_for_test(
+                0,
+                MetaCacheEntry {
+                    name: object.to_string(),
+                    ..Default::default()
+                },
+                bucket.to_string(),
+                store.pools[0].get_disks_by_key(object),
+            ))
+            .await
+            .expect("ordinary and similarly named objects must migrate");
+            let mut reader = store.pools[1]
+                .get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default())
+                .await
+                .expect("control must exist on the target");
+            let mut actual = Vec::new();
+            reader.stream.read_to_end(&mut actual).await.expect("read migrated control");
+            assert_eq!(actual, b"ordinary object contents", "{bucket}/{object}");
+            let err = store.pools[0]
+                .get_object_info(bucket, object, &ObjectOptions::default())
+                .await
+                .expect_err("migrated control must be removed from the source");
+            assert!(is_err_object_not_found(&err), "{bucket}/{object}: {err:?}");
+        }
+
+        Box::pin(store.decommission_entry_for_test(
+            0,
+            MetaCacheEntry {
+                name: conflict_object.to_string(),
+                ..Default::default()
+            },
+            RUSTFS_META_BUCKET.to_string(),
+            store.pools[0].get_disks_by_key(conflict_object),
+        ))
+        .await
+        .expect("entry must record a real conditional-copy failure");
+        let meta = store.pool_meta.read().await;
+        let info = meta.pools[0].decommission.as_ref().expect("final progress");
+        assert_eq!(info.items_decommissioned, controls.len());
+        assert_eq!(
+            info.items_decommission_failed, 2,
+            "similar names must not hide read or migration failures"
+        );
+        assert_eq!(info.bytes_failed, source_body.len() * 2);
+        drop(meta);
+        for (pool_index, expected) in [(0, source_body.as_slice()), (1, target_body.as_slice())] {
+            let mut reader = store.pools[pool_index]
+                .get_object_reader(RUSTFS_META_BUCKET, conflict_object, None, HeaderMap::new(), &ObjectOptions::default())
+                .await
+                .expect("failed migration must preserve both objects");
+            let mut actual = Vec::new();
+            reader.stream.read_to_end(&mut actual).await.expect("read conflict object");
+            assert_eq!(actual, expected);
+        }
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn decommission_final_sweep_excludes_only_set_local_usage_caches() {
+        use crate::object_api::PutObjReader;
+
+        let (_temp_dirs, store, _other_store) =
+            crate::services::rebalance::test_two_pool_stores_with_isolated_node_contexts(None).await;
+        for object in [
+            "buckets/.usage-cache.bin",
+            "buckets/.usage-cache.bin.bkp",
+            "buckets/photos/.usage-cache.bin",
+            "buckets/photos/.usage-cache.bin.bkp",
+        ] {
+            store.pools[0]
+                .get_disks_by_key(object)
+                .put_object(
+                    RUSTFS_META_BUCKET,
+                    object,
+                    &mut PutObjReader::from_vec(b"set-local cache".to_vec()),
+                    &ObjectOptions::default(),
+                )
+                .await
+                .expect("seed each supported set-local cache path");
+        }
+        let layout = DecommissionErasureLayout { data: 1, parity: 0 };
+        set_decommission_capacity_info_overrides_for_test(
+            store.id,
+            vec![vec![
+                DecommissionPoolCapacityInfo::for_test(0, layout, 0, 16_384, 16_384),
+                DecommissionPoolCapacityInfo::for_test(1, layout, 131_072, 131_072, 0),
+            ]],
+        );
+        store
+            .save_current_pool_meta_for_decommission_start(&[0], Vec::new())
+            .await
+            .expect("activate the final-sweep generation");
+        let generation = store.active_decommission_generation(0).await.expect("active generation");
+        store
+            .check_after_decommission(0, &CancellationToken::new(), generation)
+            .await
+            .expect("the four set-local cache forms must not block the final sweep");
+
+        for object in [
+            "buckets/.usage-cache.bin.extra",
+            "buckets/photos/.usage-cache.bin.bkp.extra",
+            "buckets/.usage.v2.json",
+        ] {
+            let source_set = store.pools[0].get_disks_by_key(object);
+            source_set
+                .put_object(
+                    RUSTFS_META_BUCKET,
+                    object,
+                    &mut PutObjReader::from_vec(b"unmigrated ordinary metadata".to_vec()),
+                    &ObjectOptions::default(),
+                )
+                .await
+                .expect("seed ordinary metadata that must prevent completion");
+            let err = store
+                .check_after_decommission(0, &CancellationToken::new(), generation)
+                .await
+                .expect_err("a remaining similar name or global usage snapshot must block completion");
+            assert!(err.to_string().contains("after decommissioning"), "unexpected final-sweep error: {err:?}");
+            assert!(err.to_string().contains(object), "the final sweep must identify {object}: {err:?}");
+            source_set
+                .delete_object(RUSTFS_META_BUCKET, object, ObjectOptions::default())
+                .await
+                .expect("remove only the ordinary-metadata control before the next sweep");
+        }
+        store
+            .check_after_decommission(0, &CancellationToken::new(), generation)
+            .await
+            .expect("only the four set-local caches remain after removing the controls");
+    }
+
    #[test]
    fn pool_activation_fleet_proof_error_classifier_matches_only_retryable_proof_failures() {
        assert!(is_pool_activation_fleet_proof_error(&Error::other(POOL_ACTIVATION_FLEET_PROOF_REQUIRED)));
@@ -19874,6 +20247,55 @@ mod tests {
        assert!(!is_decommission_copy_cleanup_safe_error(&wrap(Error::SlowDown)));
    }

+    #[test]
+    fn decommission_target_gate_retry_recognizes_multipart_part_errors() {
+        let wrap = |inner: Error| {
+            data_movement::data_movement_part_stage_error_for_test(
+                "decommission_object",
+                "put_object_part",
+                "bucket-a",
+                "object-a",
+                1,
+                inner,
+            )
+        };
+        let gate_busy_message =
+            format!("{DECOMMISSION_CAPACITY_TARGET_GATE_BUSY_PREFIX}7{DECOMMISSION_CAPACITY_TARGET_GATE_BUSY_SUFFIX}");
+        let wrapped = wrap(decommission_capacity_blocked_error(&gate_busy_message));
+        assert!(is_decommission_capacity_target_gate_busy(&wrapped));
+        assert_eq!(decommission_capacity_target_gate_busy_index(&wrapped), Some(7));
+        assert_eq!(
+            wrapped.to_string(),
+            format!(
+                "Io error: decommission_object: put_object_part failed for bucket-a/object-a part 1: {}",
+                decommission_capacity_blocked_error(&gate_busy_message)
+            )
+        );
+
+        for unrelated in [
+            Error::SlowDown,
+            Error::DiskFull,
+            decommission_capacity_blocked_error("target capacity is exhausted"),
+            Error::other(gate_busy_message),
+        ] {
+            let wrapped = wrap(unrelated);
+            assert!(!is_decommission_capacity_target_gate_busy(&wrapped));
+            assert_eq!(decommission_capacity_target_gate_busy_index(&wrapped), None);
+        }
+        for missing_target in [
+            Error::FileNotFound,
+            Error::ObjectNotFound("bucket-a".to_string(), "object-a".to_string()),
+            Error::VersionNotFound("bucket-a".to_string(), "object-a".to_string(), "version-a".to_string()),
+        ] {
+            assert!(is_decommission_copy_cleanup_safe_error(&missing_target));
+            assert!(
+                !is_decommission_copy_cleanup_safe_error(&wrap(missing_target)),
+                "a missing target part must never authorize source cleanup"
+            );
+        }
+        assert!(is_decommission_target_capacity_error(&wrap(Error::DiskFull)));
+    }
+
    #[test]
    fn decommission_target_capacity_error_accepts_wrapped_capacity_errors() {
        let disk_full = Error::other(format!("decommission_object: put_object failed for bucket/object: {}", Error::DiskFull));
@@ -1521,9 +1521,27 @@ fn data_movement_part_stage_error(
    bucket: &str,
    object: &str,
    part_number: usize,
-    err: impl std::fmt::Display,
+    err: Error,
 ) -> Error {
-    Error::other(format!("{op_label}: {stage} failed for {bucket}/{object} part {part_number}: {err}"))
+    let rendered = format!("{op_label}: {stage} failed for {bucket}/{object} part {part_number}: {err}");
+    if matches!(&err, Error::DecommissionCapacityBlocked { .. }) {
+        return data_movement_context_error(rendered, err);
+    }
+    // A missing target part is not evidence that the source can be deleted.
+    // Keep other part errors opaque to the source-cleanup classifiers.
+    Error::other(rendered)
+}
+
+#[cfg(test)]
+pub(crate) fn data_movement_part_stage_error_for_test(
+    op_label: &str,
+    stage: &str,
+    bucket: &str,
+    object: &str,
+    part_number: usize,
+    err: Error,
+) -> Error {
+    data_movement_part_stage_error(op_label, stage, bucket, object, part_number, err)
 }

 fn is_data_movement_part_read_error(err: &Error) -> bool {
@@ -2428,8 +2446,15 @@ mod tests {
        let err =
            data_movement_part_stage_error("rebalance_object", "put_object_part", "bucket-a", "object-a", 7, Error::SlowDown);
        let message = err.to_string();
-        assert!(message.contains("rebalance_object: put_object_part failed for bucket-a/object-a part 7"));
-        assert!(message.contains(Error::SlowDown.to_string().as_str()));
+        assert_eq!(
+            message,
+            Error::other(format!(
+                "rebalance_object: put_object_part failed for bucket-a/object-a part 7: {}",
+                Error::SlowDown
+            ))
+            .to_string()
+        );
+        assert!(data_movement_stage_source(&err).is_none());
    }

    #[test]
@@ -3397,6 +3397,7 @@ mod tests {
            scan_plan_digest: Some([1; 32]),
            complete: false,
            tombstone: false,
+            segment_invalidation_proof: None,
        }];
        partial.buckets_usage.insert(
            "bucket".to_string(),
@@ -3469,6 +3470,7 @@ mod tests {
                scan_plan_digest: Some([1; 32]),
                complete: true,
                tombstone: false,
+                segment_invalidation_proof: None,
            }],
            ..Default::default()
        };
@@ -22463,6 +22463,192 @@ mod test {
        );
    }

+    #[cfg(unix)]
+    #[tokio::test]
+    async fn conditional_mrf_manifest_dir_fsync_failure_keeps_recovery_anchors() {
+        use tempfile::tempdir;
+
+        const MRF_COMMIT_MANIFEST_SLOT_0: &str = ".heal-mrf-commit.0.bin";
+        const MRF_COMMIT_MANIFEST_SLOT_1: &str = ".heal-mrf-commit.1.bin";
+        const MRF_SCOPED_JOURNAL_PATH: &str = "buckets/.heal/mrf/journal-scoped.bin";
+
+        let _mode = durability_mode_override::set(DurabilityMode::Relaxed);
+        let dir = tempdir().expect("temp dir should be created");
+        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
+        let disk = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+        let previous_manifest = Bytes::from_static(b"mrf-committed-manifest-v1");
+        let successor_manifest = Bytes::from_static(b"mrf-committed-manifest-v2");
+        let legacy_journal = Bytes::from_static(b"legacy-mrf-journal-records");
+
+        assert_eq!(
+            disk.compare_and_update_file(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_0, None, Some(previous_manifest.clone()),)
+                .await
+                .expect("previous MRF manifest should commit"),
+            ConditionalFileUpdate::Updated
+        );
+        disk.write_all(RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH, legacy_journal.clone())
+            .await
+            .expect("legacy MRF journal should be retained");
+
+        let manifest_path = disk
+            .get_object_path(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_0)
+            .expect("MRF manifest path should resolve");
+        let parent = manifest_path.parent().expect("MRF manifest path should have a parent");
+        assert!(
+            os::fsync_dir_recorder::was_fsynced(parent),
+            "system metadata MRF manifest publication must fsync the metadata directory even under relaxed durability"
+        );
+        os::fsync_dir_recorder::set_failure(parent, ErrorKind::Other);
+
+        let err = disk
+            .compare_and_update_file(
+                RUSTFS_META_BUCKET,
+                MRF_COMMIT_MANIFEST_SLOT_0,
+                Some(previous_manifest.clone()),
+                Some(successor_manifest),
+            )
+            .await
+            .expect_err("directory fsync failure must fail the MRF manifest successor commit");
+        assert!(matches!(err, DiskError::Io(ref err) if err.kind() == ErrorKind::Other));
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_0)
+                .await
+                .expect("previous committed MRF manifest should remain readable after rollback"),
+            previous_manifest
+        );
+
+        os::fsync_dir_recorder::set_failure(parent, ErrorKind::Other);
+        let err = disk
+            .compare_and_update_file(
+                RUSTFS_META_BUCKET,
+                MRF_COMMIT_MANIFEST_SLOT_1,
+                None,
+                Some(Bytes::from_static(b"first-successor-manifest")),
+            )
+            .await
+            .expect_err("directory fsync failure must fail first MRF manifest commit");
+        assert!(matches!(err, DiskError::Io(ref err) if err.kind() == ErrorKind::Other));
+        assert!(
+            matches!(
+                disk.read_all(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_1).await,
+                Err(DiskError::FileNotFound)
+            ),
+            "uncommitted first MRF manifest must be removed when no committed anchor exists"
+        );
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH)
+                .await
+                .expect("legacy MRF journal should remain readable after failed manifest publication"),
+            legacy_journal
+        );
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn conditional_mrf_manifest_storage_full_keeps_recovery_anchors() {
+        use tempfile::tempdir;
+
+        const MRF_COMMIT_MANIFEST_SLOT_0: &str = ".heal-mrf-commit.0.bin";
+        const MRF_SCOPED_JOURNAL_PATH: &str = "buckets/.heal/mrf/journal-scoped.bin";
+
+        let _mode = durability_mode_override::set(DurabilityMode::Relaxed);
+        let dir = tempdir().expect("temp dir should be created");
+        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
+        let disk = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+        let previous_manifest = Bytes::from_static(b"mrf-committed-manifest-v1");
+        let successor_manifest = Bytes::from_static(b"mrf-committed-manifest-v2");
+        let legacy_journal = Bytes::from_static(b"legacy-mrf-journal-records");
+
+        assert_eq!(
+            disk.compare_and_update_file(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_0, None, Some(previous_manifest.clone()),)
+                .await
+                .expect("previous MRF manifest should commit"),
+            ConditionalFileUpdate::Updated
+        );
+        disk.write_all(RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH, legacy_journal.clone())
+            .await
+            .expect("legacy MRF journal should be retained");
+
+        let manifest_path = disk
+            .get_object_path(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_0)
+            .expect("MRF manifest path should resolve");
+        let parent = manifest_path.parent().expect("MRF manifest path should have a parent");
+        os::fsync_dir_recorder::set_failure(parent, ErrorKind::StorageFull);
+
+        let err = disk
+            .compare_and_update_file(
+                RUSTFS_META_BUCKET,
+                MRF_COMMIT_MANIFEST_SLOT_0,
+                Some(previous_manifest.clone()),
+                Some(successor_manifest),
+            )
+            .await
+            .expect_err("storage-full fsync failure must fail the MRF manifest successor commit");
+        assert!(matches!(err, DiskError::Io(ref err) if err.kind() == ErrorKind::StorageFull));
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_0)
+                .await
+                .expect("previous committed MRF manifest should remain readable after storage-full rollback"),
+            previous_manifest
+        );
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH)
+                .await
+                .expect("legacy MRF journal should remain readable after storage-full manifest publication failure"),
+            legacy_journal
+        );
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn conditional_mrf_manifest_storage_full_delete_keeps_recovery_anchors() {
+        use tempfile::tempdir;
+
+        const MRF_COMMIT_MANIFEST_SLOT_0: &str = ".heal-mrf-commit.0.bin";
+        const MRF_SCOPED_JOURNAL_PATH: &str = "buckets/.heal/mrf/journal-scoped.bin";
+
+        let _mode = durability_mode_override::set(DurabilityMode::Relaxed);
+        let dir = tempdir().expect("temp dir should be created");
+        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
+        let disk = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+        let committed_manifest = Bytes::from_static(b"mrf-committed-manifest-v1");
+        let legacy_journal = Bytes::from_static(b"legacy-mrf-journal-records");
+
+        assert_eq!(
+            disk.compare_and_update_file(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_0, None, Some(committed_manifest.clone()),)
+                .await
+                .expect("committed MRF manifest should publish"),
+            ConditionalFileUpdate::Updated
+        );
+        disk.write_all(RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH, legacy_journal.clone())
+            .await
+            .expect("legacy MRF journal should be retained");
+
+        let manifest_path = disk
+            .get_object_path(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_0)
+            .expect("MRF manifest path should resolve");
+        let parent = manifest_path.parent().expect("MRF manifest path should have a parent");
+        os::fsync_dir_recorder::set_failure(parent, ErrorKind::StorageFull);
+
+        let err = disk
+            .compare_and_update_file(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_0, Some(committed_manifest.clone()), None)
+            .await
+            .expect_err("storage-full fsync failure must fail the MRF manifest cleanup delete");
+        assert!(matches!(err, DiskError::Io(ref err) if err.kind() == ErrorKind::StorageFull));
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, MRF_COMMIT_MANIFEST_SLOT_0)
+                .await
+                .expect("committed MRF manifest should be restored after failed cleanup delete"),
+            committed_manifest
+        );
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH)
+                .await
+                .expect("legacy MRF journal should remain readable after failed cleanup delete"),
+            legacy_journal
+        );
+    }
+
    #[cfg(unix)]
    #[tokio::test]
    async fn conditional_file_update_dir_fsync_failure_removes_new_file_without_anchor() {
@@ -9209,6 +9209,11 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
        let dest_obj = transaction.remote_object.clone();
        let mut transition_meta = (*oi.user_defined).clone();
        rustfs_utils::http::remove_str(&mut transition_meta, rustfs_utils::http::SUFFIX_PART_CHECKSUMS);
+        // The tier holds opaque stored bytes. Its metadata must not be treated
+        // as a second object header set: forwarding SSE intent or wrapped DEKs
+        // would request a second encryption pass and disclose local envelope
+        // material to the remote provider.
+        transition_meta.retain(|key, _| !rustfs_utils::http::is_replication_stripped_encryption_key(key));
        transition_meta.insert("name".to_string(), object.to_string());
        rustfs_utils::http::metadata_compat::insert_str(
            &mut transition_meta,
@@ -9765,13 +9770,16 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                    part_opts.part_number = Some(part_info.number);
                    #[cfg(test)]
                    fail_restore_multipart_at(RestoreMultipartFailurePoint::InvalidPartSize)?;
-                    if part_info.actual_size <= 0 {
-                        return Err(Error::other(format!("invalid multipart restore part size {}", part_info.actual_size)));
+                    if part_info.size == 0 {
+                        return Err(Error::other(format!("invalid multipart restore stored part size {}", part_info.size)));
                    }
+                    let stored_part_size = i64::try_from(part_info.size).map_err(|_| {
+                        Error::other(format!("multipart restore stored part size exceeds i64: {}", part_info.size))
+                    })?;
                    #[cfg(test)]
                    fail_restore_multipart_at(RestoreMultipartFailurePoint::RangeOverflow)?;
                    let part_end = part_offset
-                        .checked_add(part_info.actual_size - 1)
+                        .checked_add(stored_part_size - 1)
                        .ok_or_else(|| Error::other("multipart restore part range overflow".to_string()))?;
                    let rs = Some(HTTPRangeSpec {
                        is_suffix_length: false,
@@ -9799,13 +9807,19 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                    #[cfg(test)]
                    fail_restore_multipart_at(RestoreMultipartFailurePoint::HashReader)?;
                    let hash_reader =
-                        HashReader::from_stream(reader, part_info.actual_size, part_info.actual_size, None, None, false)?;
+                        HashReader::from_stream(reader, stored_part_size, part_info.actual_size, None, None, false)?;
                    let mut p_reader = PutObjReader::new(hash_reader);
                    #[cfg(test)]
                    fail_restore_multipart_at(RestoreMultipartFailurePoint::PutPart)?;
+                    // `ropts` carries the object's ETag so the single-part copy-back
+                    // keeps it (the writer only ever sees stored bytes). A part write
+                    // must not inherit that object-level value, or every restored part
+                    // would be recorded under the same ETag; each part keeps its own.
+                    let mut part_write_opts = ropts.clone();
+                    part_write_opts.preserve_etag = Some(part_info.etag.clone()).filter(|etag| !etag.is_empty());
                    let p_info = self_
                        .clone()
-                        .put_object_part(bucket, object, &res.upload_id, part_info.number, &mut p_reader, &ropts)
+                        .put_object_part(bucket, object, &res.upload_id, part_info.number, &mut p_reader, &part_write_opts)
                        .await?;
                    #[cfg(test)]
                    let p_info = if restore_multipart_failure_is(RestoreMultipartFailurePoint::SizeMismatch) {
@@ -9815,7 +9829,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                    } else {
                        p_info
                    };
-                    if p_info.size as i64 != part_info.actual_size {
+                    if p_info.size as i64 != stored_part_size {
                        return Err(Error::other(ObjectApiError::InvalidObjectState(GenericError {
                            bucket: bucket.to_string(),
                            object: object.to_string(),
@@ -9847,6 +9861,11 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                    user_defined: restore_commit_metadata,
                    no_lock: false,
                    decommission_capacity_admission: opts.decommission_capacity_admission.clone(),
+                    // The composite ETag would otherwise be recomputed from the
+                    // parts as they were written back, which for an encrypted or
+                    // compressed object digests stored bytes rather than the
+                    // object's public ETag.
+                    preserve_etag: oi.etag.clone(),
                    ..Default::default()
                };
                self_
@@ -13488,6 +13507,303 @@ mod transition_commit_failure_tests {
        );
    }

+    /// backlog#2368 B5: the tier stores opaque bytes, so the archive request
+    /// must not carry the object's encryption headers. Forwarding them made
+    /// every S3 target reject an SSE-C archive outright, asked the target to
+    /// encrypt an SSE-KMS object a second time under a key id it does not own,
+    /// and handed the wrapped DEK to a third-party provider.
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn transition_does_not_forward_encryption_metadata_to_the_tier() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "transition-encryption-metadata-bucket";
+        let object = "object.bin";
+        for disk in &disk_stores {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+
+        let encryption_metadata = [
+            ("x-amz-server-side-encryption", "aws:kms"),
+            ("x-amz-server-side-encryption-aws-kms-key-id", "arn:aws:kms:us-east-1:123:key/abc"),
+            ("x-amz-server-side-encryption-customer-algorithm", "AES256"),
+            (rustfs_utils::http::INTERNAL_ENCRYPTION_KEY_HEADER, "d3JhcHBlZC1kZWs="),
+            (rustfs_utils::http::INTERNAL_ENCRYPTION_IV_HEADER, "AAAAAAAAAAAAAAAA"),
+            (rustfs_utils::http::INTERNAL_ENCRYPTION_ALGORITHM_HEADER, "AES256"),
+        ];
+        let mut user_defined: HashMap<String, String> = encryption_metadata
+            .iter()
+            .map(|(key, value)| ((*key).to_string(), (*value).to_string()))
+            .collect();
+        user_defined.insert("x-amz-meta-owner".to_string(), "finance".to_string());
+
+        let mut reader = PutObjReader::from_vec(b"stored bytes the tier keeps opaque ".repeat(64));
+        set_disks
+            .put_object(
+                bucket,
+                object,
+                &mut reader,
+                &ObjectOptions {
+                    user_defined,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("the encrypted source object should be written");
+        let original = set_disks
+            .get_object_info(bucket, object, &ObjectOptions::default())
+            .await
+            .expect("the source object should be readable");
+
+        let tier_name = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let backend = register_mock_tier(&runtime_sources::global_tier_config_mgr(), &tier_name).await;
+        set_disks
+            .transition_object(
+                bucket,
+                object,
+                &ObjectOptions {
+                    no_lock: true,
+                    transition: TransitionOptions {
+                        status: TRANSITION_PENDING.to_string(),
+                        tier: tier_name,
+                        etag: original.etag.clone().unwrap_or_default(),
+                        ..Default::default()
+                    },
+                    version_id: original.version_id.map(|version| version.to_string()),
+                    mod_time: original.mod_time,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("the encrypted object should transition");
+
+        let transitioned = set_disks
+            .get_object_info(bucket, object, &ObjectOptions::default())
+            .await
+            .expect("the transitioned object should be readable");
+        let remote_metadata = backend
+            .metadata(&transitioned.transitioned_object.name)
+            .await
+            .expect("the tier must have received the object");
+
+        for (key, _) in encryption_metadata {
+            assert!(
+                !remote_metadata.keys().any(|stored| stored.eq_ignore_ascii_case(key)),
+                "transition must not forward {key} to the tier: {remote_metadata:?}"
+            );
+        }
+        assert!(
+            remote_metadata
+                .iter()
+                .any(|(key, value)| key.eq_ignore_ascii_case("x-amz-meta-owner") && value == "finance"),
+            "ordinary user metadata must still travel to the tier: {remote_metadata:?}"
+        );
+
+        // Read-through and restore both resolve encryption locally, so the
+        // stripped keys must survive untouched in the local metadata.
+        for (key, value) in encryption_metadata {
+            assert_eq!(
+                transitioned.user_defined.get(key).map(String::as_str),
+                Some(value),
+                "the local copy must keep {key}"
+            );
+        }
+    }
+
+    /// Deterministic bytes that do not repeat with a short period, so a slice
+    /// taken at the wrong offset cannot coincidentally compare equal.
+    fn stored_representation_bytes(seed: u32, len: usize) -> Vec<u8> {
+        (0..len)
+            .map(|index| {
+                let mixed = (index as u32).wrapping_add(seed).wrapping_mul(2_654_435_761);
+                (mixed >> 13) as u8
+            })
+            .collect()
+    }
+
+    /// Compares two stored representations without dumping megabytes of bytes
+    /// into the failure output.
+    fn assert_stored_representation_eq(actual: &[u8], expected: &[u8], what: &str) {
+        assert_eq!(actual.len(), expected.len(), "{what}: stored length differs");
+        if let Some(offset) = actual.iter().zip(expected).position(|(left, right)| left != right) {
+            panic!(
+                "{what}: stored bytes differ at offset {offset} (found {:#04x}, expected {:#04x})",
+                actual[offset], expected[offset]
+            );
+        }
+    }
+
+    async fn read_stored_representation(set_disks: &Arc<SetDisks>, bucket: &str, object: &str) -> Vec<u8> {
+        let mut reader = set_disks
+            .get_object_reader(
+                bucket,
+                object,
+                None,
+                HeaderMap::new(),
+                &ObjectOptions {
+                    no_lock: true,
+                    raw_data_movement_read: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("stored-representation reader should open");
+        let mut body = Vec::new();
+        reader.stream.read_to_end(&mut body).await.expect("stored body should drain");
+        body
+    }
+
+    /// backlog#2368 B3: the multipart restore loop addresses the tier in STORED
+    /// coordinates. Accumulating each part's PLAINTEXT length instead handed
+    /// every part a misaligned slice of the remote object whose length still
+    /// satisfied the range, the `HashReader` and the completion size check, so
+    /// the copy-back reported success while silently replacing the bytes.
+    ///
+    /// The fixture reproduces the encrypted geometry — a stored form LONGER
+    /// than the plaintext it encodes — because that is what keeps a
+    /// plaintext-coordinate range inside the tier object and makes the
+    /// corruption silent rather than a short read.
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn multipart_restore_copies_the_stored_representation_back_verbatim() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "restore-multipart-stored-coordinates-bucket";
+        let object = "object.bin";
+        for disk in &disk_stores {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+
+        // The minimum-part-size gate reads the PLAINTEXT length, so part one
+        // clears 5 MiB there while its stored form carries encoding overhead.
+        let part_shapes = [(6 * 1024 * 1024_usize, 9_216_usize), (256 * 1024_usize, 512_usize)];
+        let mut user_defined = HashMap::new();
+        user_defined.insert(rustfs_utils::http::INTERNAL_ENCRYPTION_ALGORITHM_HEADER.to_string(), "AES256".to_string());
+        user_defined.insert(
+            rustfs_utils::http::INTERNAL_ENCRYPTION_IV_HEADER.to_string(),
+            "AAAAAAAAAAAAAAAA".to_string(),
+        );
+
+        let upload = set_disks
+            .new_multipart_upload(
+                bucket,
+                object,
+                &ObjectOptions {
+                    user_defined: user_defined.clone(),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("multipart upload should be created");
+
+        let mut uploaded_parts = Vec::new();
+        let mut expected_stored = Vec::new();
+        for (index, (plaintext_len, overhead)) in part_shapes.iter().enumerate() {
+            let stored = stored_representation_bytes(index as u32 * 7 + 1, plaintext_len + overhead);
+            expected_stored.extend_from_slice(&stored);
+            let stored_len = stored.len() as i64;
+            let hash_reader =
+                HashReader::from_stream(std::io::Cursor::new(stored), stored_len, *plaintext_len as i64, None, None, false)
+                    .expect("hash reader over the stored representation");
+            let mut reader = PutObjReader::new(hash_reader);
+            let info = set_disks
+                .put_object_part(bucket, object, &upload.upload_id, index + 1, &mut reader, &ObjectOptions::default())
+                .await
+                .expect("stored part should be staged");
+            assert_eq!(info.size as i64, stored_len, "a part is stored in its encoded length");
+            uploaded_parts.push(CompletePart {
+                part_num: info.part_num,
+                etag: info.etag,
+                ..Default::default()
+            });
+        }
+
+        let original = set_disks
+            .clone()
+            .complete_multipart_upload(bucket, object, &upload.upload_id, uploaded_parts, &ObjectOptions::default())
+            .await
+            .expect("source multipart upload should complete");
+        let original_parts: Vec<(usize, usize, i64, String)> = original
+            .parts
+            .iter()
+            .map(|part| (part.number, part.size, part.actual_size, part.etag.clone()))
+            .collect();
+        for (_, size, actual_size, _) in &original_parts {
+            assert!(
+                *size as i64 > *actual_size,
+                "the fixture must keep the two coordinate systems apart: stored {size} vs plaintext {actual_size}"
+            );
+        }
+        let stored_before = read_stored_representation(&set_disks, bucket, object).await;
+        assert_stored_representation_eq(&stored_before, &expected_stored, "the fixture must store its encoded bytes verbatim");
+
+        let tier_name = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        register_mock_tier(&runtime_sources::global_tier_config_mgr(), &tier_name).await;
+        set_disks
+            .transition_object(
+                bucket,
+                object,
+                &ObjectOptions {
+                    no_lock: true,
+                    transition: TransitionOptions {
+                        status: TRANSITION_PENDING.to_string(),
+                        tier: tier_name,
+                        etag: original.etag.clone().unwrap_or_default(),
+                        ..Default::default()
+                    },
+                    version_id: original.version_id.map(|version| version.to_string()),
+                    mod_time: original.mod_time,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("multipart source should transition before restore");
+
+        let operation_id = Uuid::new_v4();
+        set_disks
+            .put_object_metadata(
+                bucket,
+                object,
+                &ObjectOptions {
+                    eval_metadata: Some(restore_metadata(operation_id, true)),
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("the restore generation should be installed");
+        let mut restore_opts = ObjectOptions::default();
+        restore_opts.transition.restore_request.days = Some(1);
+        restore_opts.user_defined = restore_operation_id_metadata(operation_id);
+        set_disks
+            .clone()
+            .restore_transitioned_object(bucket, object, &restore_opts)
+            .await
+            .expect("multipart restore should complete");
+
+        let stored_after = read_stored_representation(&set_disks, bucket, object).await;
+        assert_stored_representation_eq(
+            &stored_after,
+            &expected_stored,
+            "a multipart restore must copy the stored representation back verbatim",
+        );
+
+        let restored = set_disks
+            .get_object_info(bucket, object, &ObjectOptions::default())
+            .await
+            .expect("the restored object should be readable");
+        let restored_parts: Vec<(usize, usize, i64, String)> = restored
+            .parts
+            .iter()
+            .map(|part| (part.number, part.size, part.actual_size, part.etag.clone()))
+            .collect();
+        assert_eq!(
+            restored_parts, original_parts,
+            "restore must rebuild the same part layout, sizes and part ETags"
+        );
+        assert_eq!(restored.size, original.size, "restore must keep the stored object size");
+        // backlog#2369 P7.1: the copy-back digests stored bytes, so the object's
+        // public ETag has to be carried over rather than recomputed.
+        assert_eq!(restored.etag, original.etag, "restore must preserve the object ETag");
+    }
+
    #[tokio::test]
    #[serial_test::serial]
    async fn restore_failure_after_snapshot_cleans_exact_generation_and_returns_primary_error() {
@@ -164,6 +164,16 @@ pub fn max_keys_plus_one(max_keys: i32, add_one: bool) -> i32 {
    max_keys
 }

+fn list_versions_scan_limit(max_keys: i32, has_version_marker: bool) -> i32 {
+    if max_keys <= 0 {
+        return 0;
+    }
+
+    // The marker object's versions may all be filtered out after gathering.
+    // Reserve its raw entry in addition to the next-page lookahead entry.
+    max_keys_plus_one(max_keys, true) + i32::from(has_version_marker)
+}
+
 #[derive(Debug, Clone, Copy, Eq, PartialEq)]
 enum GatherResultsState {
    LimitReached,
@@ -2139,15 +2149,19 @@ fn build_list_versions_next_marker(
        // here; advertise it as the literal `null` marker so a resumed listing
        // parses it back to `VersionMarker::Null` instead of a nil UUID that
        // `find_version_index` can never match (issue #6745).
-        (
-            Some(append_list_cache_id_to_marker(last.name.clone(), cache_id)),
+        let version_marker = if last.is_dir && last.mod_time.is_none() {
+            // A CommonPrefix has no version to resume; a version marker would
+            // make the next page include this same prefix again.
+            None
+        } else {
            Some(
                last.version_id
                    .filter(|v| !v.is_nil())
                    .map(|v| v.to_string())
                    .unwrap_or_else(|| "null".to_string()),
-            ),
-        )
+            )
+        };
+        (Some(append_list_cache_id_to_marker(last.name.clone(), cache_id)), version_marker)
    } else if let Some(last_prefix) = prefixes.last() {
        (Some(append_list_cache_id_to_marker(last_prefix.clone(), cache_id)), None)
    } else {
@@ -2866,6 +2880,20 @@ fn listing_entries_supplement_target(
        return None;
    }

+    if let Some(directory) = entries.0.iter().flatten().find(|entry| entry.is_dir()) {
+        let directory_copies = entries
+            .0
+            .iter()
+            .flatten()
+            .filter(|entry| entry.is_dir() && entry.name == directory.name)
+            .count();
+        // A committed child may have some of its directory copies only on
+        // fallback disks, just like object metadata in a partial primary sample.
+        if directory_copies < resolver.dir_quorum {
+            return Some(directory.name.clone());
+        }
+    }
+
    for (idx, entry) in entries.0.iter().enumerate() {
        let Some(entry) = entry.as_ref().filter(|entry| entry.is_object()) else {
            continue;
@@ -4018,8 +4046,7 @@ impl ECStore {
            None
        };

-        let effective_max_keys = if max_keys <= 0 { 0 } else { max_keys_plus_one(max_keys, true) };
-        // Always request max_keys + 1 to detect if there are more results
+        let effective_max_keys = list_versions_scan_limit(max_keys, has_version_marker);
        let mut opts = ListPathOptions {
            bucket: bucket.to_owned(),
            prefix: prefix.to_owned(),
@@ -5325,7 +5352,7 @@ impl Sets {
            None
        };

-        let effective_max_keys = if max_keys <= 0 { 0 } else { max_keys_plus_one(max_keys, true) };
+        let effective_max_keys = list_versions_scan_limit(max_keys, has_version_marker);
        let mut opts = ListPathOptions {
            bucket: bucket.to_owned(),
            prefix: prefix.to_owned(),
@@ -6034,7 +6061,7 @@ impl SetDisks {

        let has_version_marker = version_marker.is_some();
        let version_marker = version_marker.map(parse_version_marker).transpose()?;
-        let effective_max_keys = if max_keys <= 0 { 0 } else { max_keys_plus_one(max_keys, true) };
+        let effective_max_keys = list_versions_scan_limit(max_keys, has_version_marker);
        let mut opts = ListPathOptions {
            bucket: bucket.to_owned(),
            prefix: prefix.to_owned(),
@@ -6248,7 +6275,7 @@ impl SetDisks {
            None
        };

-        let effective_max_keys = if max_keys <= 0 { 0 } else { max_keys_plus_one(max_keys, true) };
+        let effective_max_keys = list_versions_scan_limit(max_keys, has_version_marker);
        let mut opts = ListPathOptions {
            bucket: bucket.to_owned(),
            prefix: prefix.to_owned(),
@@ -7441,6 +7468,153 @@ mod test {
        assert!(cancel.is_cancelled());
    }

+    #[test]
+    fn list_versions_pagination_scan_limit_boundaries() {
+        for has_version_marker in [false, true] {
+            assert_eq!(super::list_versions_scan_limit(-1, has_version_marker), 0);
+            assert_eq!(super::list_versions_scan_limit(0, has_version_marker), 0);
+            let marker_slot = i32::from(has_version_marker);
+            assert_eq!(super::list_versions_scan_limit(1, has_version_marker), 2 + marker_slot);
+            assert_eq!(super::list_versions_scan_limit(MAX_OBJECT_LIST, has_version_marker), 1001 + marker_slot);
+            assert_eq!(super::list_versions_scan_limit(i32::MAX, has_version_marker), 1001 + marker_slot);
+        }
+    }
+
+    #[tokio::test]
+    async fn list_versions_pagination_does_not_require_an_empty_final_page() {
+        use crate::bucket::metadata_sys::{init_bucket_metadata_sys, test_support::isolated_store_over_temp_disks};
+        use crate::storage_api_contracts::bucket::{BucketOperations as _, MakeBucketOptions};
+
+        let (dirs, store) = isolated_store_over_temp_disks().await;
+        let bucket = "version-pagination-bucket";
+        init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("pagination bucket should be created");
+        let mod_time = time::OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
+
+        for kind in ["objects", "deletes", "null", "mixed", "delimiter"] {
+            let count = if kind == "mixed" { 5 } else { 10 };
+            let mut expected = Vec::new();
+            for index in 0..count {
+                let name = if kind == "delimiter" && index % 2 == 1 {
+                    format!("{kind}/testobject-{index:02}/child")
+                } else {
+                    format!("{kind}/testobject-{index:02}")
+                };
+                let entry = match kind {
+                    "deletes" => test_delete_marker_meta_entry(&name, mod_time),
+                    "null" => test_object_meta_entry(&name),
+                    "mixed" => test_object_with_delete_marker_meta_entry(&name, mod_time, mod_time + time::Duration::SECOND),
+                    _ => test_object_meta_entry_with_erasure_versions(&name, &[(mod_time, "etag", 2, 2)]),
+                };
+                for dir in &dirs {
+                    let object_dir = dir.path().join(bucket).join(&name);
+                    tokio::fs::create_dir_all(&object_dir)
+                        .await
+                        .expect("pagination object directory should be created");
+                    tokio::fs::write(object_dir.join(STORAGE_FORMAT_FILE), &entry.metadata)
+                        .await
+                        .expect("pagination metadata should be written");
+                }
+                if kind == "delimiter" && index % 2 == 1 {
+                    expected.push((name.trim_end_matches("child").to_owned(), None, false));
+                } else {
+                    let versions = entry.file_info_versions(bucket).expect("fixture versions should decode");
+                    expected.extend(
+                        versions
+                            .versions
+                            .iter()
+                            .map(|version| (name.clone(), version.version_id, version.deleted)),
+                    );
+                }
+            }
+            let prefix = format!("{kind}/");
+            let delimiter = (kind == "delimiter").then(|| "/".to_owned());
+            // Exercise each public/internal entry point with the reported page size.
+            // The store entry point also covers exact and one-over limit boundaries.
+            for (layer, max_keys) in [(0, 0), (0, 1), (0, 5), (0, 9), (0, 10), (0, 11), (1, 5), (2, 5), (3, 5)] {
+                if layer == 3 && delimiter.is_some() {
+                    continue;
+                }
+                let mut marker = None;
+                let mut version_marker = None;
+                let expected_pages = if max_keys == 0 {
+                    1
+                } else {
+                    10usize.div_ceil(usize::try_from(max_keys).expect("positive page size"))
+                };
+                let mut actual = Vec::new();
+                for page in 0..expected_pages {
+                    let result = match layer {
+                        0 => {
+                            store
+                                .clone()
+                                .inner_list_object_versions(bucket, &prefix, marker, version_marker, delimiter.clone(), max_keys)
+                                .await
+                        }
+                        1 => {
+                            store.pools[0]
+                                .clone()
+                                .inner_list_object_versions(bucket, &prefix, marker, version_marker, delimiter.clone(), max_keys)
+                                .await
+                        }
+                        2 => {
+                            store.pools[0].disk_set[0]
+                                .clone()
+                                .inner_list_object_versions(bucket, &prefix, marker, version_marker, delimiter.clone(), max_keys)
+                                .await
+                        }
+                        _ => {
+                            store.pools[0].disk_set[0]
+                                .clone()
+                                .inner_list_object_versions_for_recursive_delete(
+                                    bucket,
+                                    &prefix,
+                                    marker,
+                                    version_marker,
+                                    max_keys,
+                                )
+                                .await
+                        }
+                    }
+                    .expect("version page should list successfully");
+                    let page_size = usize::try_from(max_keys).expect("nonnegative page size");
+                    assert_eq!(result.objects.len() + result.prefixes.len(), (10 - page * page_size).min(page_size));
+                    let has_more = page + 1 < expected_pages;
+                    assert_eq!(result.is_truncated, has_more, "{kind}, layer {layer}, max_keys {max_keys}, page {page}");
+                    assert_eq!(
+                        result.next_marker.is_some(),
+                        has_more,
+                        "key marker must exist only when another page exists"
+                    );
+                    if !has_more {
+                        assert!(
+                            result.next_version_idmarker.is_none(),
+                            "the final page must not advertise a version marker"
+                        );
+                    }
+                    actual.extend(
+                        result
+                            .objects
+                            .into_iter()
+                            .map(|object| (object.name, object.version_id, object.delete_marker)),
+                    );
+                    actual.extend(result.prefixes.into_iter().map(|prefix| (prefix, None, false)));
+                    marker = result.next_marker;
+                    version_marker = result.next_version_idmarker;
+                }
+                // Objects and CommonPrefixes are serialized separately; compare their
+                // identities without relying on their relative position in the response.
+                actual.sort();
+                let mut expected = if max_keys == 0 { Vec::new() } else { expected.clone() };
+                expected.sort();
+                assert_eq!(actual, expected, "{kind}, layer {layer}, max_keys {max_keys}");
+            }
+        }
+    }
+
    #[test]
    fn version_marker_is_applied_only_when_key_marker_entry_is_present() {
        let version_marker = Some(VersionMarker::Null);
@@ -9448,6 +9622,71 @@ mod test {
        assert!(supplemented.is_latest_delete_marker());
    }

+    #[tokio::test]
+    async fn latest_listing_supplement_checks_fallback_disks_for_common_prefix_quorum() {
+        let mut fallback_disks = Vec::new();
+        let mut fallback_tempdirs = Vec::new();
+        for index in 0..4 {
+            let tempdir = tempfile::tempdir().expect("fallback tempdir should be created");
+            let endpoint = Endpoint::try_from(tempdir.path().to_str().expect("fallback path should be utf8"))
+                .expect("fallback endpoint should parse");
+            let disk = new_disk(
+                &endpoint,
+                &DiskOption {
+                    cleanup: false,
+                    health_check: false,
+                },
+            )
+            .await
+            .expect("fallback disk should be created");
+            disk.make_volume("bucket").await.expect("fallback bucket should be created");
+            for copies in [3, 4] {
+                if index < copies {
+                    let object = format!("quux-{copies}/thud");
+                    let entry = test_object_meta_entry(&object);
+                    disk.write_all("bucket", &format!("{object}/{STORAGE_FORMAT_FILE}"), bytes::Bytes::from(entry.metadata))
+                        .await
+                        .expect("fallback child metadata should be written");
+                }
+            }
+            fallback_disks.push(disk);
+            fallback_tempdirs.push(tempdir);
+        }
+        let supplement = ListingSupplement::new(
+            ListingSupplementOptions {
+                bucket: "bucket".to_owned(),
+                path: String::new(),
+                recursive: false,
+                incl_deleted: false,
+                skip_hidden_prefix_check: false,
+                filter_prefix: None,
+                forward_to: None,
+                per_disk_limit: 100,
+                skip_total_timeout: true,
+                walkdir_timeout: None,
+                walkdir_stall_timeout: None,
+            },
+            Arc::new(fallback_disks),
+            FallbackClaimTracker::default(),
+        );
+        // A 16-drive EC:4 set asks 12 primary disks. A committed write may
+        // exist on eight primary disks and all four remaining fallback disks.
+        let resolver = list_metadata_resolution_params("bucket".to_owned(), 4, 12, false, 0);
+        for fallback_copies in [3, 4] {
+            let prefix = format!("quux-{fallback_copies}/");
+            let mut primary = vec![Some(test_dir_meta_entry(&prefix)); 8];
+            primary.extend([None, None, None, None]);
+            let entry =
+                resolve_listing_entries_with_supplement(MetaCacheEntries(primary), resolver.clone(), true, supplement.clone())
+                    .await;
+            assert_eq!(
+                entry.map(|entry| entry.name),
+                (fallback_copies == 4).then_some(prefix),
+                "the common prefix needs all twelve copies, including fallback disks"
+            );
+        }
+    }
+
    #[test]
    fn latest_listing_supplement_keeps_a_subquorum_delete_marker_hidden() {
        let object_mod_time = time::OffsetDateTime::from_unix_timestamp(1_705_312_300).expect("valid timestamp");
@@ -104,6 +104,7 @@ walkdir = { workspace = true }
 http = { workspace = true }
 temp-env = { workspace = true, features = ["async_closure"] }
 tokio = { workspace = true, features = ["test-util", "fs", "rt-multi-thread"] }
+chrono = { workspace = true }

 [lib]
 doctest = false
@@ -34,7 +34,7 @@ use std::time::{Duration, UNIX_EPOCH};
 use tokio::sync::{RwLock, Semaphore};
 use tracing::{debug, error, warn};

-use super::{DiskStore, EcstoreError};
+use super::{DiskStore, EcstoreError, POOL_META_NAME, RUSTFS_META_BUCKET};

 /// Outcome of classifying an error returned by [`HealStorageAPI::heal_object`].
 enum HealObjectOutcome {
@@ -68,6 +68,13 @@ struct PageConcurrencyGuard {
    set_label: String,
 }

+struct ErasureSetPassCounters<'a> {
+    processed_objects: &'a mut u64,
+    successful_objects: &'a mut u64,
+    failed_objects: &'a mut u64,
+    skipped_objects: &'a mut u64,
+}
+
 impl PageConcurrencyGuard {
    fn new(in_flight: Arc<AtomicUsize>, set_label: String) -> Self {
        let current = in_flight.fetch_add(1, Ordering::SeqCst) + 1;
@@ -106,6 +113,7 @@ pub struct ErasureSetHealer {
    heal_opts: HealOpts,
    source: HealRequestSource,
    target_endpoints: Arc<[String]>,
+    pool_metadata_target_endpoints: Arc<[String]>,
    replacement_task_id: Option<String>,
    replacement_target_identities: Option<Arc<[ReplacementTargetIdentity]>>,
    mainline_pacer: Option<Arc<super::pacing::MainlinePacer>>,
@@ -355,6 +363,7 @@ impl ErasureSetHealer {
            heal_opts,
            source,
            target_endpoints: Vec::new().into(),
+            pool_metadata_target_endpoints: Vec::new().into(),
            replacement_task_id: None,
            replacement_target_identities: None,
            mainline_pacer: None,
@@ -378,6 +387,13 @@ impl ErasureSetHealer {
        self
    }

+    pub(crate) fn with_pool_metadata_targets(mut self, mut target_endpoints: Vec<String>) -> Self {
+        target_endpoints.sort_unstable();
+        target_endpoints.dedup();
+        self.pool_metadata_target_endpoints = target_endpoints.into();
+        self
+    }
+
    pub(crate) fn with_replacement_identity_fence(
        mut self,
        replacement_target_identities: Option<Vec<ReplacementTargetIdentity>>,
@@ -834,6 +850,21 @@ impl ErasureSetHealer {
            current_object_index = 0;
        }

+        if failed_objects == 0 && skipped_objects == 0 && failed_buckets == 0 {
+            self.heal_replacement_pool_metadata(
+                set_disk_id,
+                &mut ErasureSetPassCounters {
+                    processed_objects: &mut processed_objects,
+                    successful_objects: &mut successful_objects,
+                    failed_objects: &mut failed_objects,
+                    skipped_objects: &mut skipped_objects,
+                },
+                resume_manager,
+                checkpoint_manager,
+            )
+            .await?;
+        }
+
        // 5. finalize. Only declare the set healed when nothing failed AND
        // nothing was transiently skipped — otherwise the resume/checkpoint
        // state must survive so the failed/skipped versions are retried instead
@@ -919,6 +950,207 @@ impl ErasureSetHealer {
        Ok(())
    }

+    async fn heal_replacement_pool_metadata(
+        &self,
+        set_disk_id: &str,
+        counters: &mut ErasureSetPassCounters<'_>,
+        resume_manager: &ResumeManager,
+        checkpoint_manager: &CheckpointManager,
+    ) -> Result<()> {
+        let target_endpoints = if self.pool_metadata_target_endpoints.is_empty() {
+            self.target_endpoints.as_ref()
+        } else {
+            self.pool_metadata_target_endpoints.as_ref()
+        };
+        let target_scoped_recreate = !self.heal_opts.dry_run && self.heal_opts.recreate && !target_endpoints.is_empty();
+        if self.replacement_task_id.is_none() && !target_scoped_recreate {
+            return Ok(());
+        }
+        if target_endpoints.is_empty() {
+            return Err(Error::TaskExecutionFailed {
+                message: "Replacement pool metadata heal requires target endpoints".to_string(),
+            });
+        }
+
+        let object_key = format!("{RUSTFS_META_BUCKET}/{POOL_META_NAME}");
+        let checkpoint_key = compose_key(&object_key, None);
+        let checkpoint = checkpoint_manager.get_checkpoint().await;
+        if checkpoint.processed_objects.contains(&checkpoint_key)
+            || checkpoint.failed_objects.contains(&checkpoint_key)
+            || checkpoint.skipped_objects.contains(&checkpoint_key)
+        {
+            return Ok(());
+        }
+        drop(checkpoint);
+
+        self.verify_replacement_identity_fence("pool metadata").await?;
+        resume_manager
+            .set_current_item(Some(RUSTFS_META_BUCKET.to_string()), Some(POOL_META_NAME.to_string()))
+            .await?;
+
+        let result = match self
+            .storage
+            .heal_object(RUSTFS_META_BUCKET, POOL_META_NAME, None, &self.heal_opts)
+            .await
+        {
+            Ok((result, None)) if target_outcomes_complete(&result, target_endpoints) => {
+                let object_size = result_object_size_u64(&result);
+                match self
+                    .storage
+                    .replacement_targets_have_version(RUSTFS_META_BUCKET, POOL_META_NAME, None, &self.heal_opts, target_endpoints)
+                    .await
+                {
+                    Ok(true) => (object_size, Ok(())),
+                    Ok(false) => (
+                        object_size,
+                        Err(Error::transient_skip(
+                            "Skipped replacement pool metadata heal because target readback did not confirm the committed version",
+                        )),
+                    ),
+                    Err(err) => (
+                        object_size,
+                        Err(Error::transient_skip(format!(
+                            "Skipped replacement pool metadata heal because target readback failed: {err}"
+                        ))),
+                    ),
+                }
+            }
+            Ok((result, None)) => (
+                result_object_size_u64(&result),
+                Err(Error::transient_skip(
+                    "Skipped replacement pool metadata heal because a replacement target was not committed",
+                )),
+            ),
+            Ok((result, Some(err))) => {
+                let object_size = result_object_size_u64(&result);
+                match Self::classify_heal_object_error(&err) {
+                    HealObjectOutcome::Absent | HealObjectOutcome::Transient => (
+                        object_size,
+                        Err(Error::transient_skip(format!(
+                            "Skipped replacement pool metadata heal due to transient error: {err}"
+                        ))),
+                    ),
+                    HealObjectOutcome::Failed => (object_size, Err(err)),
+                }
+            }
+            Err(err @ Error::TaskCancelled) | Err(err @ Error::TaskTimeout) => return Err(err),
+            Err(err) => match Self::classify_heal_object_error(&err) {
+                HealObjectOutcome::Absent | HealObjectOutcome::Transient => (
+                    0,
+                    Err(Error::transient_skip(format!(
+                        "Skipped replacement pool metadata heal due to transient error: {err}"
+                    ))),
+                ),
+                HealObjectOutcome::Failed => (0, Err(err)),
+            },
+        };
+
+        let (object_size, result) = result;
+        let mut bytes_processed = self.progress.read().await.bytes_processed;
+        let mut telemetry_unknown = false;
+        let checkpoint_outcome = match result {
+            Ok(()) => {
+                telemetry_unknown |= !increment_counter(counters.successful_objects);
+                telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
+                debug!(
+                    target: "rustfs::heal::erasure_healer",
+                    event = EVENT_HEAL_ERASURE_OBJECT_STATE,
+                    component = LOG_COMPONENT_HEAL,
+                    subsystem = LOG_SUBSYSTEM_ERASURE_HEALER,
+                    set_disk_id,
+                    bucket = RUSTFS_META_BUCKET,
+                    object = POOL_META_NAME,
+                    state = "healed",
+                    "Replacement pool metadata healed"
+                );
+                CheckpointObjectOutcome::Processed
+            }
+            Err(Error::TransientSkip { message }) => {
+                telemetry_unknown |= !increment_counter(counters.skipped_objects);
+                telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
+                warn!(
+                    target: "rustfs::heal::erasure_healer",
+                    event = EVENT_HEAL_ERASURE_OBJECT_STATE,
+                    component = LOG_COMPONENT_HEAL,
+                    subsystem = LOG_SUBSYSTEM_ERASURE_HEALER,
+                    set_disk_id,
+                    bucket = RUSTFS_META_BUCKET,
+                    object = POOL_META_NAME,
+                    state = "transient_skip",
+                    error = %message,
+                    "Replacement pool metadata heal skipped due to transient error"
+                );
+                CheckpointObjectOutcome::Skipped
+            }
+            Err(err) => {
+                telemetry_unknown |= !increment_counter(counters.failed_objects);
+                telemetry_unknown |= !add_bytes(&mut bytes_processed, object_size);
+                warn!(
+                    target: "rustfs::heal::erasure_healer",
+                    event = EVENT_HEAL_ERASURE_OBJECT_STATE,
+                    component = LOG_COMPONENT_HEAL,
+                    subsystem = LOG_SUBSYSTEM_ERASURE_HEALER,
+                    set_disk_id,
+                    bucket = RUSTFS_META_BUCKET,
+                    object = POOL_META_NAME,
+                    state = "failed",
+                    error = %err,
+                    "Replacement pool metadata heal failed"
+                );
+                CheckpointObjectOutcome::Failed
+            }
+        };
+
+        telemetry_unknown |= !increment_counter(counters.processed_objects);
+        let (outcome_record, counter_unknown, skipped_new_versions, skipped_ilm_expired) = {
+            let mut progress = self.progress.write().await;
+            progress.set_current_object(Some(object_key.clone()));
+            progress.update_object_progress(
+                *counters.processed_objects,
+                *counters.successful_objects,
+                *counters.failed_objects,
+                *counters.skipped_objects,
+                bytes_processed,
+            );
+            if telemetry_unknown {
+                progress.mark_unknown();
+            }
+            (
+                CheckpointObjectOutcomeRecord {
+                    object: checkpoint_key,
+                    outcome: checkpoint_outcome,
+                    successful: progress.objects_healed,
+                    failed: progress.objects_failed,
+                    skipped: progress.skipped_objects,
+                    bytes: progress.bytes_processed,
+                    skipped_new_versions: progress.skipped_new_versions,
+                    skipped_ilm_expired: progress.skipped_ilm_expired,
+                    counter_unknown: progress.counter_unknown,
+                },
+                progress.counter_unknown,
+                progress.skipped_new_versions,
+                progress.skipped_ilm_expired,
+            )
+        };
+        checkpoint_manager.record_object_outcome(outcome_record).await?;
+        resume_manager
+            .update_progress_with_bytes(
+                *counters.processed_objects,
+                *counters.successful_objects,
+                *counters.failed_objects,
+                *counters.skipped_objects,
+                bytes_processed,
+            )
+            .await?;
+        resume_manager
+            .set_skipped_version_counts(skipped_new_versions, skipped_ilm_expired)
+            .await?;
+        if counter_unknown {
+            resume_manager.mark_counter_unknown().await?;
+        }
+        Ok(())
+    }
+
    /// heal single bucket with resume
    #[allow(clippy::too_many_arguments)]
    #[tracing::instrument(skip(self, current_object_index, processed_objects, successful_objects, failed_objects, skipped_objects, resume_manager, checkpoint_manager), fields(bucket = %bucket, bucket_index = bucket_index))]
@@ -1679,7 +1911,8 @@ mod resume_loop_tests {
    use crate::heal::storage::{HealLifecycleExpiryContext, HealListItem, HealObjectInfo, HealStorageAPI};
    use crate::heal::storage_api::status::BucketInfo;
    use crate::heal::{
-        BUCKET_META_PREFIX, DiskOption, DiskStore, EcstoreError, Endpoint, HealDiskExt as _, RUSTFS_META_BUCKET, new_disk,
+        BUCKET_META_PREFIX, DiskOption, DiskStore, EcstoreError, Endpoint, HealDiskExt as _, POOL_META_NAME, RUSTFS_META_BUCKET,
+        new_disk,
    };
    use crate::{Error, Result};
    use rustfs_heal_contracts::heal_channel::{HealOpts, HealRequestSource};
@@ -1773,6 +2006,28 @@ mod resume_loop_tests {
        assert!(!target_outcomes_complete(&duplicate, &["replacement-a".to_string()]));
    }

+    fn replacement_target_ok_result(endpoint: &str, object: &str) -> HealResultItem {
+        HealResultItem {
+            object: object.to_string(),
+            object_size: 1024,
+            before: Infos {
+                drives: vec![HealDriveInfo {
+                    endpoint: endpoint.to_string(),
+                    state: "missing".to_string(),
+                    ..Default::default()
+                }],
+            },
+            after: Infos {
+                drives: vec![HealDriveInfo {
+                    endpoint: endpoint.to_string(),
+                    state: "ok".to_string(),
+                    ..Default::default()
+                }],
+            },
+            ..Default::default()
+        }
+    }
+
    #[derive(Clone)]
    struct Page {
        items: Vec<HealListItem>,
@@ -2444,6 +2699,8 @@ mod resume_loop_tests {
            HealRequestSource::AutoHeal,
        )
        .with_replacement_targets(vec!["replacement-a".to_string()], Some(replacement_task_id.clone()));
+        env.storage
+            .set_result(POOL_META_NAME, None, replacement_target_ok_result("replacement-a", POOL_META_NAME));

        healer
            .heal_erasure_set(&["b".to_string()], "pool_0_set_0")
@@ -2461,9 +2718,121 @@ mod resume_loop_tests {
            CheckpointManager::has_checkpoint(&env.healer.disk, &replacement_task_id).await,
            "the checkpoint must survive until the caller clears the healing marker"
        );
+        assert_eq!(env.storage.calls(), vec![(POOL_META_NAME.to_string(), None)]);
        drop(checkpoint);
    }

+    #[tokio::test]
+    async fn replacement_pool_metadata_readback_failure_schedules_retry() {
+        let env = make_env_with_targets(vec!["replacement-a".to_string()]).await;
+        let replacement_task_id = ResumeUtils::generate_task_id();
+        ResumeManager::new_replacement_intent(
+            env.healer.disk.clone(),
+            replacement_task_id.clone(),
+            "pool_0_set_0".to_string(),
+            vec!["b".to_string()],
+            vec!["replacement-a".to_string()],
+            vec![crate::heal::resume::ReplacementTargetIdentity {
+                endpoint: "replacement-a".to_string(),
+                canonical_path: "/mnt/replacement-a".to_string(),
+                physical_device_ids: vec!["device-a".to_string()],
+                filesystem_identity: "1:2:3".to_string(),
+            }],
+        )
+        .await
+        .expect("replacement intent should persist");
+        env.storage
+            .set_result(POOL_META_NAME, None, replacement_target_ok_result("replacement-a", POOL_META_NAME));
+        env.storage.set_replacement_commit_evidence(POOL_META_NAME, None, false);
+        let healer = ErasureSetHealer::new(
+            env.storage.clone(),
+            Arc::new(RwLock::new(HealProgress::new())),
+            CancellationToken::new(),
+            env.healer.disk.clone(),
+            HealOpts::default(),
+            HealRequestSource::AutoHeal,
+        )
+        .with_replacement_targets(vec!["replacement-a".to_string()], Some(replacement_task_id.clone()));
+
+        let error = healer
+            .heal_erasure_set(&["b".to_string()], "pool_0_set_0")
+            .await
+            .expect_err("unconfirmed pool metadata readback must keep the replacement incomplete");
+
+        assert!(error.to_string().contains("Replacement erasure set heal incomplete"));
+        let state = ResumeManager::load_replacement_intent(env.healer.disk.clone(), &replacement_task_id)
+            .await
+            .expect("replacement retry state must remain")
+            .get_state()
+            .await;
+        assert!(!state.completed);
+        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)]);
+    }
+
+    #[tokio::test]
+    async fn admin_recreate_target_heals_pool_metadata_before_completion() {
+        let env = make_env_with_targets(vec!["replacement-a".to_string()]).await;
+        let healer = ErasureSetHealer::new(
+            env.storage.clone(),
+            Arc::new(RwLock::new(HealProgress::new())),
+            CancellationToken::new(),
+            env.healer.disk.clone(),
+            HealOpts {
+                recreate: true,
+                pool: Some(0),
+                set: Some(0),
+                ..Default::default()
+            },
+            HealRequestSource::Admin,
+        )
+        .with_pool_metadata_targets(vec!["replacement-a".to_string()]);
+        env.storage
+            .set_result(POOL_META_NAME, None, replacement_target_ok_result("replacement-a", POOL_META_NAME));
+
+        healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &env.resume, &env.checkpoint)
+            .await
+            .expect("admin recreate should heal and verify pool metadata");
+
+        assert!(env.resume.get_state().await.completed);
+        assert_eq!(env.storage.calls(), vec![(POOL_META_NAME.to_string(), None)]);
+    }
+
+    #[tokio::test]
+    async fn admin_recreate_pool_metadata_readback_failure_keeps_resume_state() {
+        let env = make_env_with_targets(vec!["replacement-a".to_string()]).await;
+        let healer = ErasureSetHealer::new(
+            env.storage.clone(),
+            Arc::new(RwLock::new(HealProgress::new())),
+            CancellationToken::new(),
+            env.healer.disk.clone(),
+            HealOpts {
+                recreate: true,
+                pool: Some(0),
+                set: Some(0),
+                ..Default::default()
+            },
+            HealRequestSource::Admin,
+        )
+        .with_pool_metadata_targets(vec!["replacement-a".to_string()]);
+        env.storage
+            .set_result(POOL_META_NAME, None, replacement_target_ok_result("replacement-a", POOL_META_NAME));
+        env.storage.set_replacement_commit_evidence(POOL_META_NAME, None, false);
+
+        let error = healer
+            .execute_heal_with_resume(&["b".to_string()], "pool_0_set_0", &env.resume, &env.checkpoint)
+            .await
+            .expect_err("unconfirmed admin recreate pool metadata must keep the set incomplete");
+
+        assert!(error.to_string().contains("Erasure set heal incomplete"));
+        let state = env.resume.get_state().await;
+        assert!(!state.completed);
+        assert_eq!(state.retry_count, 1);
+        assert_eq!(env.storage.calls(), vec![(POOL_META_NAME.to_string(), None)]);
+    }
+
    #[tokio::test]
    async fn retry_exhaustion_keeps_resume_artifacts_for_recovery() {
        let env = make_env().await;
@@ -506,6 +506,25 @@ fn active_heal_for_dedup_key(active_heals: &HashMap<String, Arc<HealTask>>, key:
        .map(|(task_id, task)| (task_id.clone(), task.heal_type.clone()))
 }

+fn request_matches_task(request: &HealRequest, task: &HealTask) -> bool {
+    request.heal_type == task.heal_type
+        && request.options == task.options
+        && request.priority == task.priority
+        && request.source == task.source
+        && request.retry_attempts == task.retry_attempts
+        && request.heal_endpoints == task.heal_endpoints
+}
+
+fn request_matches_request(request: &HealRequest, existing: &HealRequest) -> bool {
+    request.heal_type == existing.heal_type
+        && request.options == existing.options
+        && request.priority == existing.priority
+        && request.source == existing.source
+        && request.force_start == existing.force_start
+        && request.retry_attempts == existing.retry_attempts
+        && request.heal_endpoints == existing.heal_endpoints
+}
+
 fn retrying_heal_for_dedup_key(retrying_heals: &HashMap<String, RetryingHeal>, key: &str) -> Option<(String, HealType)> {
    retrying_heals
        .iter()
@@ -822,6 +841,10 @@ pub struct HealManager {
    replacement_recovery_anchors: Arc<std::sync::Mutex<HashMap<String, String>>>,
    /// Set IDs whose durable replacement metadata is corrupt or conflicting.
    replacement_recovery_blocked_sets: Arc<std::sync::Mutex<HashSet<String>>>,
+    /// Durable handoff of interrupted administrator root traversals.
+    root_recovery: Arc<root_recovery::RootHealRecovery>,
+    /// Keep forceStart's cancellation side effects inside the shutdown fence.
+    force_start_shutdown: Mutex<()>,
    /// Storage layer interface
    storage: Arc<dyn HealStorageAPI>,
    /// Cancel token
@@ -857,6 +880,7 @@ struct HealQueueContext<'a> {
    retrying_heals: &'a Arc<Mutex<HashMap<String, RetryingHeal>>>,
    mrf_repair_notice_targets: &'a Arc<StdMutex<HashMap<String, Vec<MrfRepairNoticeTarget>>>>,
    replacement_recovery_anchors: &'a Arc<std::sync::Mutex<HashMap<String, String>>>,
+    root_recovery: &'a Arc<root_recovery::RootHealRecovery>,
    config: &'a Arc<RwLock<HealConfig>>,
    statistics: &'a Arc<RwLock<HealStatistics>>,
    storage: &'a Arc<dyn HealStorageAPI>,
@@ -1358,6 +1382,8 @@ impl HealManager {
            mrf_repair_notice_targets: Arc::new(StdMutex::new(HashMap::new())),
            replacement_recovery_anchors: Arc::new(std::sync::Mutex::new(HashMap::new())),
            replacement_recovery_blocked_sets: Arc::new(std::sync::Mutex::new(HashSet::new())),
+            root_recovery: Arc::new(root_recovery::RootHealRecovery::default()),
+            force_start_shutdown: Mutex::new(()),
            storage,
            cancel_token: CancellationToken::new(),
            statistics: Arc::new(RwLock::new(HealStatistics::new())),
@@ -1393,6 +1419,23 @@ impl HealManager {
            "Heal manager starting"
        );

+        // Restore graceful-shutdown root responsibilities before automatic
+        // repair can admit overlapping work.
+        if let Err(error) = self.replay_root_heals().await {
+            // A missing owner or invalid root record must not block existing
+            // replacement recovery. Keep its file for a later restart after
+            // the owner is readable or the record has been repaired.
+            warn!(
+                target: "rustfs::heal::manager",
+                event = EVENT_HEAL_MANAGER_STATE,
+                component = LOG_COMPONENT_HEAL,
+                subsystem = LOG_SUBSYSTEM_MANAGER,
+                state = "root_recovery_deferred",
+                error = %error,
+                "Root heal restart recovery deferred"
+            );
+        }
+
        // start scheduler
        self.start_scheduler().await?;

@@ -1430,6 +1473,7 @@ impl HealManager {

    /// Stop HealManager
    pub async fn stop(&self) -> Result<()> {
+        let _force_start_guard = self.force_start_shutdown.lock().await;
        info!(
            target: "rustfs::heal::manager",
            event = EVENT_HEAL_MANAGER_STATE,
@@ -1439,11 +1483,39 @@ impl HealManager {
            "Heal manager stopping"
        );

-        // cancel all tasks
-        self.cancel_token.cancel();
-
-        // wait for all tasks to complete
+        // Keep scheduler, cancellation, and retry ownership stable until every
+        // unfinished root traversal has a durable successor. A failed write
+        // must leave the manager running and the shutdown marker unclean.
        let mut active_heals = self.active_heals.lock().await;
+        let queue = self.heal_queue.lock().await;
+        let retrying = self.retrying_heals.lock().await;
+        for task in active_heals.values() {
+            if root_recovery::is_root_heal(&task.heal_type, task.source) {
+                if task.get_status().await == HealTaskStatus::Completed {
+                    self.root_recovery.remove(&task.id, &task.heal_type, task.source).await?;
+                } else {
+                    let mut request = match task.retry_request_with_remaining_timeout().await {
+                        Ok(request) => request,
+                        Err(Error::TaskTimeout) => {
+                            let mut request = task.retry_request();
+                            request.options.timeout = Some(Duration::ZERO);
+                            request
+                        }
+                        Err(error) => return Err(error),
+                    };
+                    request.retry_attempts = task.retry_attempts;
+                    self.root_recovery.persist(&request).await?;
+                }
+            }
+        }
+        for request in queue.requests().chain(retrying.values().map(|retrying| &retrying.request)) {
+            self.root_recovery.persist(request).await?;
+        }
+        self.cancel_token.cancel();
+        drop(retrying);
+        drop(queue);
+
+        // cancel active workers after the durable handoff
        for task in active_heals.values() {
            if let Err(e) = task.cancel().await {
                warn!(
@@ -1508,7 +1580,7 @@ impl HealManager {
        request: HealRequest,
        preserve_alias: bool,
    ) -> Result<HealAdmissionReceipt> {
-        self.submit_heal_request_with_receipt_alias_and_mrf_notice(request, preserve_alias, None)
+        self.submit_heal_request_with_receipt_alias_and_mrf_notice(request, preserve_alias, true, None)
            .await
    }

@@ -1544,19 +1616,43 @@ impl HealManager {
        request: HealRequest,
        mrf_notice_target: MrfRepairNoticeTarget,
    ) -> Result<HealAdmissionReceipt> {
-        self.submit_heal_request_with_receipt_alias_and_mrf_notice(request, true, Some(mrf_notice_target))
+        self.submit_heal_request_with_receipt_alias_and_mrf_notice(request, true, true, Some(mrf_notice_target))
            .await
    }

+    pub(crate) async fn durable_mrf_repair_anchor(
+        &self,
+        intent: &rustfs_common::mrf_channel::MrfIntent,
+    ) -> Option<rustfs_common::mrf_channel::MrfDurableRepairAnchor> {
+        match self.storage.mrf_bucket_incarnation_id(intent.bucket.as_ref()).await {
+            Ok(Some(bucket_incarnation_id)) => {
+                rustfs_common::mrf_channel::MrfDurableRepairAnchor::from_intent(intent, bucket_incarnation_id)
+            }
+            Ok(None) | Err(_) => None,
+        }
+    }
+
    async fn submit_heal_request_with_receipt_alias_and_mrf_notice(
        &self,
        request: HealRequest,
        preserve_alias: bool,
+        accept_same_request_id_replay: bool,
        mrf_notice_target: Option<MrfRepairNoticeTarget>,
    ) -> Result<HealAdmissionReceipt> {
        let admission_start = Instant::now();
        let source = request.source;
        let force_start = request.force_start;
+        // A forceStart must not retire an old durable owner if shutdown will
+        // reject its replacement. Hold the same gate through final admission.
+        let _force_start_guard = if source == HealRequestSource::Admin && force_start {
+            let guard = self.force_start_shutdown.lock().await;
+            if self.cancel_token.is_cancelled() {
+                return Err(Error::Other("Heal manager is stopping".to_string()));
+            }
+            Some(guard)
+        } else {
+            None
+        };
        // HS-06 forceStart semantics (admin only): MinIO stops the old task
        // first and then starts the new one. Cancel any active admin task
        // overlapping this request's path before entering admission, so the
@@ -1564,7 +1660,9 @@ impl HealManager {
        if request.source == HealRequestSource::Admin && request.force_start {
            let overlapping: Vec<String> = {
                let active_heals = self.active_heals.lock().await;
-                active_heals
+                let queue = self.heal_queue.lock().await;
+                let retrying = self.retrying_heals.lock().await;
+                let mut ids = active_heals
                    .iter()
                    .filter(|(task_id, task)| {
                        task.source == HealRequestSource::Admin
@@ -1572,7 +1670,17 @@ impl HealManager {
                            && *task_id != &request.id
                    })
                    .map(|(task_id, _)| task_id.clone())
-                    .collect()
+                    .collect::<Vec<_>>();
+                ids.extend(
+                    queue
+                        .requests()
+                        .chain(retrying.values().map(|retrying| &retrying.request))
+                        .filter(|pending| {
+                            root_recovery::is_root_heal(&pending.heal_type, pending.source) && pending.id != request.id
+                        })
+                        .map(|pending| pending.id.clone()),
+                );
+                ids
            };
            for task_id in overlapping {
                match self.cancel_task(&task_id).await {
@@ -1586,17 +1694,14 @@ impl HealManager {
                        result = "force_start_cancelled_overlap",
                        "Admin forceStart cancelled an overlapping heal task"
                    ),
-                    Err(err) => warn!(
-                        target: "rustfs::heal::manager",
-                        event = EVENT_HEAL_QUEUE_ADMISSION,
-                        component = LOG_COMPONENT_HEAL,
-                        subsystem = LOG_SUBSYSTEM_MANAGER,
-                        request_id = %request.id,
-                        cancelled_task_id = %task_id,
-                        error = %err,
-                        result = "force_start_cancel_failed",
-                        "Admin forceStart failed to cancel an overlapping heal task"
-                    ),
+                    Err(err) => return Err(err),
+                }
+            }
+            // A failed or timed-out replay may have only its durable owner
+            // left. Root responsibility overlaps every administrator path.
+            for pending in self.root_recovery.pending().await? {
+                if pending.id != request.id {
+                    self.cancel_task(&pending.id).await?;
                }
            }
        }
@@ -1609,10 +1714,75 @@ impl HealManager {
        // active -> retrying transitions can slip between duplicate checks.
        let lock_phase_start = Instant::now();
        let active_heals = self.active_heals.lock().await;
+        if self.cancel_token.is_cancelled() {
+            return Err(Error::Other("Heal manager is stopping".to_string()));
+        }
        #[cfg(test)]
        pause_duplicate_admission_after_active_lock(&request.id).await;
        let mut queue = self.heal_queue.lock().await;
        let retrying_heals = self.retrying_heals.lock().await;
+
+        let request_id_admission = active_heals
+            .get(&request.id)
+            .map(|task| (request_matches_task(&request, task), "active"))
+            .or_else(|| {
+                queue
+                    .requests()
+                    .find(|queued| queued.id == request.id)
+                    .map(|queued| (request_matches_request(&request, queued), "queued"))
+            })
+            .or_else(|| {
+                retrying_heals
+                    .get(&request.id)
+                    .map(|retrying| (request_matches_request(&request, &retrying.request), "retrying"))
+            });
+        if let Some((matches_existing, duplicate_state)) = request_id_admission {
+            let admission = if matches_existing {
+                if accept_same_request_id_replay {
+                    HealAdmissionResult::Accepted
+                } else {
+                    Self::duplicate_admission_for_request(&request, &config)
+                }
+            } else {
+                HealAdmissionResult::Dropped(HealAdmissionDropReason::AlreadyRunning)
+            };
+            if matches!(admission, HealAdmissionResult::Accepted | HealAdmissionResult::Merged)
+                && let Some(target) = mrf_notice_target
+            {
+                let mut targets = lock_mrf_repair_notice_targets(&self.mrf_repair_notice_targets);
+                Self::insert_mrf_repair_notice_target(&mut targets, &request.id, target);
+            }
+            drop(retrying_heals);
+            drop(queue);
+            drop(active_heals);
+            let lock_phase = lock_phase_start.elapsed();
+            Self::record_admission_metric(request.source, admission, "duplicate");
+            self.record_admission_observation(HealAdmissionObservation {
+                source,
+                result: admission,
+                context: "duplicate",
+                force_start,
+                displaced: false,
+                start_duration: admission_start.elapsed(),
+                lock_phase,
+            });
+            debug!(
+                target: "rustfs::heal::manager",
+                event = EVENT_HEAL_QUEUE_ADMISSION,
+                component = LOG_COMPONENT_HEAL,
+                subsystem = LOG_SUBSYSTEM_MANAGER,
+                request_id = %request.id,
+                duplicate_state,
+                result = admission.result_label(),
+                reason = admission.reason_label(),
+                "Heal queue admission reused an existing request id"
+            );
+            return Ok(HealAdmissionReceipt {
+                result: admission,
+                task_id: request.id,
+            });
+        }
+
        let duplicate = (!request.force_start).then(|| {
            active_heal_for_dedup_key(&active_heals, &dedup_key)
                .map(|(task_id, _)| (task_id, "active"))
@@ -1819,7 +1989,10 @@ impl HealManager {

    /// Submit heal request.
    pub async fn submit_heal_request(&self, request: HealRequest) -> Result<HealAdmissionResult> {
-        Ok(self.submit_heal_request_with_receipt_and_alias(request, true).await?.result)
+        Ok(self
+            .submit_heal_request_with_receipt_alias_and_mrf_notice(request, true, false, None)
+            .await?
+            .result)
    }

    /// Get task status
@@ -2048,6 +2221,7 @@ impl HealManager {
        {
            let mut active_heals = self.active_heals.lock().await;
            if let Some(task) = active_heals.get(&canonical_task_id) {
+                self.root_recovery.remove(&task.id, &task.heal_type, task.source).await?;
                task.cancel().await?;
                let completed = CompletedHealStatus::snapshot(task, HealTaskStatus::Cancelled).await;
                publish_completed_heal(&self.completed_heals, &self.task_aliases, &canonical_task_id, completed, true).await;
@@ -2071,6 +2245,11 @@ impl HealManager {

        {
            let mut retrying_heals = self.retrying_heals.lock().await;
+            if let Some(retrying) = retrying_heals.get(&canonical_task_id) {
+                self.root_recovery
+                    .remove(&canonical_task_id, &retrying.request.heal_type, retrying.request.source)
+                    .await?;
+            }
            if let Some(retrying) = retrying_heals.remove(&canonical_task_id) {
                retrying.cancel_token.cancel();
                drop(retrying_heals);
@@ -2091,6 +2270,11 @@ impl HealManager {
        }

        let mut queue = self.heal_queue.lock().await;
+        if let Some(request) = queue.requests().find(|request| request.id == canonical_task_id) {
+            self.root_recovery
+                .remove(&request.id, &request.heal_type, request.source)
+                .await?;
+        }
        if queue.remove_request_id(&canonical_task_id).is_some() {
            publish_heal_queue_length(&queue);
            info!(
@@ -2108,6 +2292,10 @@ impl HealManager {
            return Ok(());
        }

+        drop(queue);
+        if self.root_recovery.cancel_pending(&canonical_task_id).await? {
+            return Ok(());
+        }
        Err(Error::TaskNotFound {
            task_id: task_id.to_string(),
        })
@@ -2126,6 +2314,7 @@ impl HealManager {

            for task_id in &task_ids {
                if let Some(task) = active_heals.get(task_id) {
+                    self.root_recovery.remove(&task.id, &task.heal_type, task.source).await?;
                    task.cancel().await?;
                    let completed = CompletedHealStatus::snapshot(task, HealTaskStatus::Cancelled).await;
                    publish_completed_heal(&self.completed_heals, &self.task_aliases, task_id, completed, true).await;
@@ -2154,6 +2343,11 @@ impl HealManager {
                .collect::<Vec<_>>();

            for task_id in &task_ids {
+                if let Some(retrying) = retrying_heals.get(task_id) {
+                    self.root_recovery
+                        .remove(task_id, &retrying.request.heal_type, retrying.request.source)
+                        .await?;
+                }
                if let Some(retrying) = retrying_heals.remove(task_id) {
                    retrying.cancel_token.cancel();
                    cancelled += 1;
@@ -2176,6 +2370,14 @@ impl HealManager {
        }

        let mut queue = self.heal_queue.lock().await;
+        for request in queue
+            .requests()
+            .filter(|request| heal_type_matches_path(&request.heal_type, heal_path))
+        {
+            self.root_recovery
+                .remove(&request.id, &request.heal_type, request.source)
+                .await?;
+        }
        let queued_cancelled = queue.remove_matching(|request| heal_type_matches_path(&request.heal_type, heal_path));
        if !queued_cancelled.is_empty() {
            publish_heal_queue_length(&queue);
@@ -2187,6 +2389,13 @@ impl HealManager {
            self.remove_mrf_repair_notice_targets_for_task(&request.id);
        }

+        if heal_type_matches_path(&HealType::Cluster, heal_path) {
+            for pending in self.root_recovery.pending().await? {
+                if self.root_recovery.cancel_pending(&pending.id).await? {
+                    cancelled += 1;
+                }
+            }
+        }
        if cancelled == 0 {
            return Err(Error::TaskNotFound {
                task_id: heal_path.to_string(),
@@ -2298,6 +2507,7 @@ impl std::fmt::Debug for HealManager {

 mod auto_scan;
 mod queue;
+mod root_recovery;
 mod scheduler;
 mod unclean_shutdown;

@@ -0,0 +1,342 @@
+// Copyright 2024 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//! Graceful-shutdown handoff for administrator root heals. This namespace is
+//! separate from erasure-set checkpoints and replacement generations, which
+//! cannot represent a cluster traversal. One coordinator disk owns each
+//! record; never create a fallback copy after an uncertain write or deletion.
+
+use super::*;
+use crate::heal::storage_api::owner::{EcstoreConditionalFileUpdate, EcstoreDiskAPI, EcstoreDiskBytes};
+use crate::heal::{DiskStore, RUSTFS_META_BUCKET};
+use serde::{Deserialize, Serialize};
+
+// The metadata bucket already exists and its parent is durable. Creating a
+// nested journal directory here would also require syncing every ancestor.
+const ROOT_RECOVERY_PREFIX: &str = "root-heal-";
+const ROOT_RECOVERY_SCHEMA: u32 = 1;
+
+#[derive(Debug, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+struct RootHealIntent {
+    schema: u32,
+    task_id: String,
+    #[serde(deserialize_with = "decode_options")]
+    options: HealOptions,
+    priority: HealPriority,
+    retry_attempts: u32,
+    created_at: SystemTime,
+}
+
+impl RootHealIntent {
+    fn from_request(request: &HealRequest) -> Self {
+        Self {
+            schema: ROOT_RECOVERY_SCHEMA,
+            task_id: request.id.clone(),
+            options: request.options.clone(),
+            priority: request.priority,
+            retry_attempts: request.retry_attempts,
+            created_at: request.created_at,
+        }
+    }
+
+    fn into_request(self) -> HealRequest {
+        let mut request = HealRequest::new(HealType::Cluster, self.options, self.priority);
+        request.id = self.task_id;
+        request.source = HealRequestSource::Admin;
+        request.retry_attempts = self.retry_attempts;
+        request.created_at = self.created_at;
+        request
+    }
+}
+
+#[derive(Default)]
+pub(super) struct RootHealRecovery {
+    mutation: Mutex<()>,
+    #[cfg(test)]
+    disks: Option<Vec<DiskStore>>,
+}
+
+pub(super) fn is_root_heal(heal_type: &HealType, source: HealRequestSource) -> bool {
+    source == HealRequestSource::Admin && matches!(heal_type, HealType::Cluster)
+}
+
+fn decode_options<'de, D: serde::Deserializer<'de>>(deserializer: D) -> std::result::Result<HealOptions, D::Error> {
+    let value = serde_json::Value::deserialize(deserializer)?;
+    let object = value
+        .as_object()
+        .ok_or_else(|| serde::de::Error::custom("root heal options must be an object"))?;
+    const FIELDS: &[&str] = &[
+        "scan_mode",
+        "remove_corrupted",
+        "recreate_missing",
+        "update_parity",
+        "recursive",
+        "dry_run",
+        "no_lock",
+        "timeout",
+        "pool_index",
+        "set_index",
+    ];
+    if object.keys().any(|key| !FIELDS.contains(&key.as_str())) {
+        return Err(serde::de::Error::custom("unknown root heal recovery option"));
+    }
+    let options: HealOptions = serde_json::from_value(value).map_err(serde::de::Error::custom)?;
+    if options.no_lock {
+        return Err(serde::de::Error::custom("administrator root heal cannot skip namespace locking"));
+    }
+    Ok(options)
+}
+
+fn intent_path(task_id: &str) -> Result<String> {
+    let parsed = uuid::Uuid::parse_str(task_id).map_err(|_| Error::Other("Invalid root heal recovery task id".to_string()))?;
+    if parsed.to_string() != task_id {
+        return Err(Error::Other("Noncanonical root heal recovery task id".to_string()));
+    }
+    Ok(format!("{ROOT_RECOVERY_PREFIX}{task_id}.json"))
+}
+
+fn decode_intent(task_id: &str, bytes: &[u8]) -> Result<RootHealIntent> {
+    let _ = intent_path(task_id)?;
+    let intent: RootHealIntent = serde_json::from_slice(bytes)
+        .map_err(|error| Error::Other(format!("Invalid root heal recovery record {task_id}: {error}")))?;
+    if intent.schema != ROOT_RECOVERY_SCHEMA || intent.task_id != task_id {
+        return Err(Error::Other(format!("Unsupported or mismatched root heal recovery record {task_id}")));
+    }
+    Ok(intent)
+}
+
+impl RootHealRecovery {
+    #[cfg(test)]
+    pub(super) fn with_disks(disks: Vec<DiskStore>) -> Self {
+        Self {
+            mutation: Mutex::new(()),
+            disks: Some(disks),
+        }
+    }
+
+    async fn disks(&self) -> Result<Vec<DiskStore>> {
+        #[cfg(test)]
+        if let Some(disks) = &self.disks {
+            return Ok(disks.clone());
+        }
+        let map = local_disk_map_read().await;
+        if map.values().any(Option::is_none) {
+            return Err(Error::Other("Root heal recovery owner may be on an unavailable local disk".to_string()));
+        }
+        let mut disks = map.values().flatten().cloned().collect::<Vec<_>>();
+        disks.sort_by_key(|disk| EcstoreDiskAPI::endpoint(disk.as_ref()).to_string());
+        Ok(disks)
+    }
+
+    async fn find(disks: &[DiskStore], task_id: &str) -> Result<Option<(DiskStore, EcstoreDiskBytes)>> {
+        let path = intent_path(task_id)?;
+        let mut found = None;
+        for disk in disks {
+            // read_all reports FileNotFound even when the whole metadata
+            // volume is absent; that is an unknown owner, not empty state.
+            EcstoreDiskAPI::stat_volume(disk.as_ref(), RUSTFS_META_BUCKET).await?;
+            match EcstoreDiskAPI::read_all(disk.as_ref(), RUSTFS_META_BUCKET, &path).await {
+                Ok(bytes) => {
+                    decode_intent(task_id, &bytes)?;
+                    if found.is_some() {
+                        return Err(Error::Other(format!("Multiple root heal recovery owners for {task_id}")));
+                    }
+                    found = Some((disk.clone(), bytes));
+                }
+                Err(DiskError::FileNotFound) => {}
+                Err(error) => return Err(Error::Disk(error)),
+            }
+        }
+        Ok(found)
+    }
+
+    pub(super) async fn persist(&self, request: &HealRequest) -> Result<()> {
+        if !is_root_heal(&request.heal_type, request.source) {
+            return Ok(());
+        }
+        let _guard = self.mutation.lock().await;
+        let disks = self.disks().await?;
+        let existing = Self::find(&disks, &request.id).await?;
+        let (disk, expected) = match existing {
+            Some((disk, bytes)) => (disk, Some(bytes)),
+            None => {
+                let disk = disks
+                    .first()
+                    .cloned()
+                    .ok_or_else(|| Error::Other("No local disk available for root heal shutdown recovery".to_string()))?;
+                (disk, None)
+            }
+        };
+        if request.options.no_lock {
+            return Err(Error::Other("Administrator root heal cannot skip namespace locking".to_string()));
+        }
+        let bytes = serde_json::to_vec(&RootHealIntent::from_request(request))
+            .map_err(|error| Error::Other(format!("Serialize root heal recovery record: {error}")))?;
+        match EcstoreDiskAPI::compare_and_update_file(
+            disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            &intent_path(&request.id)?,
+            expected,
+            Some(bytes.into()),
+        )
+        .await?
+        {
+            EcstoreConditionalFileUpdate::Updated => Ok(()),
+            _ => Err(Error::Other(format!("Root heal recovery record changed for {}", request.id))),
+        }
+    }
+
+    pub(super) async fn remove(&self, task_id: &str, heal_type: &HealType, source: HealRequestSource) -> Result<bool> {
+        if !is_root_heal(heal_type, source) {
+            return Ok(false);
+        }
+        let _guard = self.mutation.lock().await;
+        let Some((disk, bytes)) = Self::find(&self.disks().await?, task_id).await? else {
+            return Ok(false);
+        };
+        match EcstoreDiskAPI::compare_and_update_file(
+            disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            &intent_path(task_id)?,
+            Some(bytes),
+            None,
+        )
+        .await?
+        {
+            EcstoreConditionalFileUpdate::Updated => Ok(true),
+            _ => Err(Error::Other(format!("Root heal recovery record changed while retiring {task_id}"))),
+        }
+    }
+
+    pub(super) async fn checkpoint_failed_execution(&self, task: &HealTask) -> Result<()> {
+        if !is_root_heal(&task.heal_type, task.source) {
+            return Ok(());
+        }
+        let remaining = match task.retry_request_with_remaining_timeout().await {
+            Ok(request) => request.options.timeout,
+            Err(Error::TaskTimeout) => Some(Duration::ZERO),
+            Err(error) => return Err(error),
+        };
+        let _guard = self.mutation.lock().await;
+        let Some((disk, expected)) = Self::find(&self.disks().await?, &task.id).await? else {
+            // A first execution that failed has no restart handoff to update.
+            return Ok(());
+        };
+        let mut intent = decode_intent(&task.id, &expected)?;
+        let mut expected_options = intent.options.clone();
+        expected_options.timeout = task.options.timeout;
+        if intent.created_at != task.created_at || intent.priority != task.priority || expected_options != task.options {
+            return Err(Error::Other(format!("Root heal recovery owner changed for {}", task.id)));
+        }
+        // A terminal timeout leaves no runtime owner for stop() to snapshot.
+        // Checkpoint its consumed budget before publishing terminal status;
+        // never refund time if an earlier checkpoint is already stricter.
+        intent.options.timeout = match (intent.options.timeout, remaining) {
+            (Some(previous), Some(remaining)) => Some(previous.min(remaining)),
+            (previous, remaining) => previous.or(remaining),
+        };
+        intent.retry_attempts = intent.retry_attempts.max(task.retry_attempts);
+        let bytes = serde_json::to_vec(&intent)
+            .map_err(|error| Error::Other(format!("Serialize root heal recovery checkpoint: {error}")))?;
+        match EcstoreDiskAPI::compare_and_update_file(
+            disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            &intent_path(&task.id)?,
+            Some(expected),
+            Some(bytes.into()),
+        )
+        .await?
+        {
+            EcstoreConditionalFileUpdate::Updated => Ok(()),
+            _ => Err(Error::Other(format!("Root heal recovery record changed while checkpointing {}", task.id))),
+        }
+    }
+
+    pub(super) async fn cancel_pending(&self, task_id: &str) -> Result<bool> {
+        if intent_path(task_id).is_err() {
+            return Ok(false);
+        }
+        self.remove(task_id, &HealType::Cluster, HealRequestSource::Admin).await
+    }
+
+    pub(super) async fn pending(&self) -> Result<Vec<HealRequest>> {
+        let _guard = self.mutation.lock().await;
+        let disks = self.disks().await?;
+        let mut ids = HashSet::new();
+        for disk in &disks {
+            EcstoreDiskAPI::stat_volume(disk.as_ref(), RUSTFS_META_BUCKET).await?;
+            let entries = match EcstoreDiskAPI::list_dir(disk.as_ref(), "", RUSTFS_META_BUCKET, "", -1).await {
+                Ok(entries) => entries,
+                Err(DiskError::FileNotFound) => continue,
+                Err(error) => return Err(Error::Disk(error)),
+            };
+            for entry in entries {
+                let Some(task_id) = entry
+                    .strip_prefix(ROOT_RECOVERY_PREFIX)
+                    .and_then(|entry| entry.strip_suffix(".json"))
+                else {
+                    continue;
+                };
+                let _ = intent_path(task_id)?;
+                ids.insert(task_id.to_string());
+            }
+        }
+        let mut requests = Vec::new();
+        for task_id in ids {
+            if let Some((_, bytes)) = Self::find(&disks, &task_id).await? {
+                requests.push(decode_intent(&task_id, &bytes)?.into_request());
+            }
+        }
+        requests.sort_by(|left, right| left.created_at.cmp(&right.created_at).then_with(|| left.id.cmp(&right.id)));
+        Ok(requests)
+    }
+}
+
+impl HealManager {
+    pub(super) async fn replay_root_heals(&self) -> Result<()> {
+        // Decode every record before admitting anything. These are already
+        // accepted responsibilities, so restore distinct IDs even when their
+        // paths overlap or the configured admission capacity has changed.
+        let requests = self.root_recovery.pending().await?;
+        let active = self.active_heals.lock().await;
+        let mut queue = self.heal_queue.lock().await;
+        let retrying = self.retrying_heals.lock().await;
+        for mut request in requests {
+            request.force_start = true;
+            let existing = active
+                .get(&request.id)
+                .map(|task| request_matches_task(&request, task))
+                .or_else(|| {
+                    queue
+                        .requests()
+                        .find(|queued| queued.id == request.id)
+                        .map(|queued| request_matches_request(&request, queued))
+                })
+                .or_else(|| {
+                    retrying
+                        .get(&request.id)
+                        .map(|retrying| request_matches_request(&request, &retrying.request))
+                });
+            match existing {
+                Some(true) => continue,
+                Some(false) => return Err(Error::Other(format!("Conflicting root heal recovery task {}", request.id))),
+                None => {}
+            }
+            queue.push(request);
+        }
+        publish_heal_queue_length(&queue);
+        Ok(())
+    }
+}
@@ -26,6 +26,7 @@ impl HealManager {
        let retrying_heals = self.retrying_heals.clone();
        let mrf_repair_notice_targets = self.mrf_repair_notice_targets.clone();
        let replacement_recovery_anchors = self.replacement_recovery_anchors.clone();
+        let root_recovery = self.root_recovery.clone();
        let cancel_token = self.cancel_token.clone();
        let statistics = self.statistics.clone();
        let storage = self.storage.clone();
@@ -59,6 +60,7 @@ impl HealManager {
                            retrying_heals: &retrying_heals,
                            mrf_repair_notice_targets: &mrf_repair_notice_targets,
                            replacement_recovery_anchors: &replacement_recovery_anchors,
+                            root_recovery: &root_recovery,
                            config: &config,
                            statistics: &statistics,
                            storage: &storage,
@@ -78,6 +80,7 @@ impl HealManager {
                            retrying_heals: &retrying_heals,
                            mrf_repair_notice_targets: &mrf_repair_notice_targets,
                            replacement_recovery_anchors: &replacement_recovery_anchors,
+                            root_recovery: &root_recovery,
                            config: &config,
                            statistics: &statistics,
                            storage: &storage,
@@ -106,6 +109,7 @@ impl HealManager {
            retrying_heals,
            mrf_repair_notice_targets,
            replacement_recovery_anchors,
+            root_recovery,
            config,
            statistics,
            storage,
@@ -117,6 +121,9 @@ impl HealManager {
        let config = config.read().await;
        let mainline_pressure = Self::mainline_throttle_active(&config, workload_provider);
        let mut active_heals_guard = active_heals.lock().await;
+        if cancel_token.is_cancelled() {
+            return;
+        }
        publish_active_heal_count(&active_heals_guard);

        // Check if new heal tasks can be started
@@ -206,6 +213,7 @@ impl HealManager {
                let replacement_recovery_anchors_clone = replacement_recovery_anchors.clone();
                let statistics_clone = statistics.clone();
                let notify_clone = notify.clone();
+                let root_recovery_clone = root_recovery.clone();
                let manager_cancel_token = cancel_token.clone();
                let task_type_label_for_spawn = task_type_label.clone();
                let task_set_label_for_spawn = task_set_label.clone();
@@ -294,6 +302,38 @@ impl HealManager {
                    tests::pause_completed_retention_before_publish(&task_id, &completed_status).await;
                    let mut active_heals_guard = active_heals_clone.lock().await;
                    let owns_completion = active_heals_guard.contains_key(&task_id);
+                    if owns_completion
+                        && result.is_ok()
+                        && let Err(error) = root_recovery_clone.remove(&task_id, &task.heal_type, task.source).await
+                    {
+                        // Keep the durable responsibility if retirement fails.
+                        // Replaying a completed traversal is idempotent.
+                        warn!(
+                            target: "rustfs::heal::manager",
+                            event = EVENT_HEAL_SCHEDULER_STATE,
+                            component = LOG_COMPONENT_HEAL,
+                            subsystem = LOG_SUBSYSTEM_MANAGER,
+                            task_id,
+                            state = "root_recovery_retirement_failed",
+                            error = %error,
+                            "Failed to retire root heal recovery record"
+                        );
+                    }
+                    if owns_completion
+                        && result.is_err()
+                        && let Err(error) = root_recovery_clone.checkpoint_failed_execution(&task).await
+                    {
+                        warn!(
+                            target: "rustfs::heal::manager",
+                            event = EVENT_HEAL_SCHEDULER_STATE,
+                            component = LOG_COMPONENT_HEAL,
+                            subsystem = LOG_SUBSYSTEM_MANAGER,
+                            task_id,
+                            state = "root_recovery_checkpoint_failed",
+                            error = %error,
+                            "Failed to checkpoint root heal recovery execution budget"
+                        );
+                    }
                    let cancelled_completion = if owns_completion {
                        false
                    } else {
@@ -733,23 +773,30 @@ pub(super) fn mrf_verified_repair_event_for_target(
        HealObjectDisposition::AuthoritativelyAbsent => MrfVerifiedRepairDisposition::AuthoritativelyAbsent,
        _ => return None,
    };
-    if target.kind != MrfKind::PartialWrite {
-        return None;
-    }
-    let expected_kind = HealObjectKind::Object;
+    let expected_kind = match target.kind {
+        MrfKind::DecodeFailure => HealObjectKind::Decode,
+        MrfKind::MetadataCorruption => HealObjectKind::Metadata,
+        MrfKind::PartialWrite => HealObjectKind::Object,
+    };
    if outcome.identity.kind != expected_kind
        || outcome.identity.bucket.as_str() != target.bucket.as_ref()
        || outcome.identity.object.as_str() != target.object.as_ref()
    {
        return None;
    }
-    let version_id = target.version_id.filter(|bytes| *bytes != [0; 16]);
+    let version_id = (!matches!(target.kind, MrfKind::MetadataCorruption))
+        .then_some(target.version_id)
+        .flatten()
+        .filter(|bytes| *bytes != [0; 16]);
    let expected_version = version_id.map(|bytes| uuid::Uuid::from_bytes(bytes).to_string());
    if outcome.identity.version_id != expected_version {
        return None;
    }
-    let expected_pool = target.scope.and_then(|scope| usize::try_from(scope.pool_index).ok());
-    let expected_set = target.scope.and_then(|scope| usize::try_from(scope.set_index).ok());
+    let scope = (!matches!(target.kind, MrfKind::MetadataCorruption))
+        .then_some(target.scope)
+        .flatten();
+    let expected_pool = scope.and_then(|scope| usize::try_from(scope.pool_index).ok());
+    let expected_set = scope.and_then(|scope| usize::try_from(scope.set_index).ok());
    if outcome.identity.pool_index != expected_pool || outcome.identity.set_index != expected_set {
        return None;
    }
@@ -759,7 +806,7 @@ pub(super) fn mrf_verified_repair_event_for_target(
        bucket: target.bucket.clone(),
        object: target.object.clone(),
        version_id,
-        scope: target.scope,
+        scope,
        lease: target.lease,
        bucket_incarnation_id,
        disposition,
@@ -26,6 +26,7 @@ use rustfs_madmin::heal_commands::HealResultItem;
 use std::sync::Mutex as StdMutex;
 use tempfile::TempDir;

+mod root_recovery;
 mod running_mainline;

 use super::super::{DiskOption, DiskStore, Endpoint, new_disk, storage_api::status::BucketInfo};
@@ -94,6 +95,7 @@ async fn process_manager_queue_once(manager: &HealManager) {
        retrying_heals: &manager.retrying_heals,
        mrf_repair_notice_targets: &manager.mrf_repair_notice_targets,
        replacement_recovery_anchors: &manager.replacement_recovery_anchors,
+        root_recovery: &manager.root_recovery,
        config: &manager.config,
        statistics: &manager.statistics,
        storage: &manager.storage,
@@ -1142,18 +1144,48 @@ fn mrf_verified_repair_event_requires_positive_exact_identity() {
    assert_eq!(event.bucket_incarnation_id, incarnation);
    assert_eq!(event.disposition, MrfVerifiedRepairDisposition::Repaired);

-    assert!(
-        mrf_verified_repair_event_for_target(
-            &MrfRepairNoticeTarget {
-                kind: MrfKind::DecodeFailure,
-                ..target.clone()
-            },
-            &matching
-        )
-        .is_none(),
-        "only receipt-producing partial-write object heals can publish verified events today"
+    let decode_target = MrfRepairNoticeTarget {
+        kind: MrfKind::DecodeFailure,
+        ..target.clone()
+    };
+    let decode_outcome = HealObjectOutcome {
+        identity: HealObjectIdentity {
+            kind: HealObjectKind::Decode,
+            ..matching.identity.clone()
+        },
+        ..matching.clone()
+    };
+    let decode_event =
+        mrf_verified_repair_event_for_target(&decode_target, &decode_outcome).expect("decode repairs publish exact proofs");
+    assert_eq!(decode_event.kind, MrfKind::DecodeFailure);
+    assert_eq!(
+        decode_event.scope,
+        Some(MrfScope {
+            pool_index: 1,
+            set_index: 2
+        })
    );

+    let metadata_target = MrfRepairNoticeTarget {
+        kind: MrfKind::MetadataCorruption,
+        ..target.clone()
+    };
+    let metadata_outcome = HealObjectOutcome {
+        identity: HealObjectIdentity {
+            kind: HealObjectKind::Metadata,
+            version_id: None,
+            pool_index: None,
+            set_index: None,
+            ..matching.identity.clone()
+        },
+        ..matching.clone()
+    };
+    let metadata_event =
+        mrf_verified_repair_event_for_target(&metadata_target, &metadata_outcome).expect("metadata repairs publish exact proofs");
+    assert_eq!(metadata_event.kind, MrfKind::MetadataCorruption);
+    assert_eq!(metadata_event.version_id, None);
+    assert_eq!(metadata_event.scope, None);
+
    for rejected in [
        HealObjectOutcome {
            disposition: HealObjectDisposition::Unknown,
@@ -4537,6 +4569,48 @@ async fn test_force_start_marks_dedup_key_for_future_duplicates() {
    );
 }

+#[tokio::test]
+async fn same_request_id_replay_reuses_existing_task_without_force_start_duplication() {
+    let storage: Arc<dyn HealStorageAPI> = Arc::new(MockStorage);
+    let manager = HealManager::new(storage, None);
+
+    let mut original = admin_prefix_request("bucket", "logs/");
+    original.force_start = true;
+    let original_id = original.id.clone();
+    let accepted = manager
+        .submit_heal_request_with_receipt(original.clone())
+        .await
+        .expect("original forceStart request should queue");
+    assert_eq!(accepted.result, HealAdmissionResult::Accepted);
+    assert_eq!(accepted.task_id, original_id);
+
+    let replayed = manager
+        .submit_heal_request_with_receipt(original.clone())
+        .await
+        .expect("same request id and payload should reuse the existing task");
+    assert_eq!(replayed.result, HealAdmissionResult::Accepted);
+    assert_eq!(replayed.task_id, original_id);
+    assert_eq!(
+        manager.get_queue_length().await,
+        1,
+        "exact forceStart replay must not create a second queued task"
+    );
+
+    let mut changed = original;
+    changed.options.remove_corrupted = true;
+    let changed = manager
+        .submit_heal_request_with_receipt(changed)
+        .await
+        .expect("same request id with a changed payload should fail closed");
+    assert_eq!(changed.result, HealAdmissionResult::Dropped(HealAdmissionDropReason::AlreadyRunning));
+    assert_eq!(changed.task_id, original_id);
+    assert_eq!(
+        manager.get_queue_length().await,
+        1,
+        "same-id conflict must not displace or duplicate the original task"
+    );
+}
+
 #[test]
 fn test_running_heal_set_counts_groups_set_scoped_tasks() {
    let storage: Arc<dyn HealStorageAPI> = Arc::new(MockStorage);
@@ -0,0 +1,465 @@
+// Copyright 2024 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use super::super::root_recovery::RootHealRecovery;
+use super::*;
+use crate::heal::RUSTFS_META_BUCKET;
+
+async fn recovery_disk() -> (TempDir, DiskStore) {
+    let temp = TempDir::new().expect("temporary root recovery disk");
+    let endpoint = Endpoint::try_from(temp.path().to_string_lossy().as_ref()).expect("disk endpoint");
+    let disk = new_disk(
+        &endpoint,
+        &DiskOption {
+            cleanup: false,
+            health_check: false,
+        },
+    )
+    .await
+    .expect("local recovery disk");
+    match disk.make_volume(RUSTFS_META_BUCKET).await {
+        Ok(()) | Err(DiskError::VolumeExists) => {}
+        Err(error) => panic!("metadata volume: {error}"),
+    }
+    (temp, disk)
+}
+
+fn recovery_manager(disks: Vec<DiskStore>) -> HealManager {
+    let mut manager = HealManager::new(
+        Arc::new(MockStorage),
+        Some(HealConfig {
+            enable_auto_heal: false,
+            ..Default::default()
+        }),
+    );
+    manager.root_recovery = Arc::new(RootHealRecovery::with_disks(disks));
+    manager
+}
+
+fn root_request() -> HealRequest {
+    let mut request = HealRequest::new(HealType::Cluster, HealOptions::default(), HealPriority::High);
+    request.source = HealRequestSource::Admin;
+    request
+}
+
+async fn active_root(manager: &HealManager, request: HealRequest) -> Arc<HealTask> {
+    let task = Arc::new(HealTask::from_request(request, manager.storage.clone()));
+    *task.status.write().await = HealTaskStatus::Running;
+    task.progress.write().await.update_object_progress(1, 1, 0, 0, 128);
+    manager.active_heals.lock().await.insert(task.id.clone(), task.clone());
+    task
+}
+
+#[tokio::test]
+async fn root_recovery_shutdown_restart_replays_same_id_and_success_retires_intent() {
+    let (_temp, disk) = recovery_disk().await;
+    let manager = recovery_manager(vec![disk.clone()]);
+    let mut request = root_request();
+    request.options.recursive = true;
+    let task = active_root(&manager, request.clone()).await;
+    manager.stop().await.expect("durable shutdown handoff");
+    assert!(task.cancel_token.is_cancelled());
+    drop(manager);
+
+    let restarted = recovery_manager(vec![disk]);
+    restarted.replay_root_heals().await.expect("replay durable root");
+    restarted.replay_root_heals().await.expect("replay is idempotent");
+    assert_eq!(restarted.get_queue_length().await, 1);
+    let restored = restarted
+        .heal_queue
+        .lock()
+        .await
+        .requests()
+        .next()
+        .cloned()
+        .expect("restored request");
+    assert_eq!(restored.id, request.id);
+    assert_eq!(restored.options, request.options);
+    assert_eq!(restored.priority, request.priority);
+    assert_eq!(restored.retry_attempts, request.retry_attempts);
+    assert_eq!(restored.created_at, request.created_at);
+
+    process_manager_queue_once(&restarted).await;
+    tokio::time::timeout(Duration::from_secs(5), async {
+        loop {
+            if matches!(restarted.get_task_status(&request.id).await, Ok(HealTaskStatus::Completed))
+                && !restarted.active_heals.lock().await.contains_key(&request.id)
+            {
+                break;
+            }
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .expect("restored root executes successfully");
+    assert!(restarted.root_recovery.pending().await.expect("read completion").is_empty());
+}
+
+#[tokio::test]
+async fn root_recovery_explicit_cancel_covers_active_queued_retrying_and_durable_only() {
+    for state in ["active", "queued", "retrying", "durable_only", "root_path"] {
+        let (_temp, disk) = recovery_disk().await;
+        let manager = recovery_manager(vec![disk.clone()]);
+        let request = root_request();
+        manager.root_recovery.persist(&request).await.expect("durable responsibility");
+        match state {
+            "active" => {
+                active_root(&manager, request.clone()).await;
+            }
+            "queued" => {
+                manager.replay_root_heals().await.expect("queued recovery");
+            }
+            "retrying" => {
+                insert_retrying_request(&manager, request.clone()).await;
+            }
+            _ => {}
+        }
+        if state == "root_path" {
+            assert_eq!(manager.cancel_tasks_for_path("").await.expect("cancel durable root path"), 1);
+        } else {
+            manager.cancel_task(&request.id).await.expect("cancel root responsibility");
+        }
+        drop(manager);
+        let restarted = recovery_manager(vec![disk]);
+        restarted
+            .replay_root_heals()
+            .await
+            .expect("restart after explicit cancellation");
+        assert_eq!(restarted.get_queue_length().await, 0, "state={state}");
+    }
+}
+
+#[tokio::test]
+async fn root_recovery_force_start_cancels_durable_only_responsibility() {
+    let (_temp, disk) = recovery_disk().await;
+    let manager = recovery_manager(vec![disk.clone()]);
+    let old = root_request();
+    manager
+        .root_recovery
+        .persist(&old)
+        .await
+        .expect("old terminal responsibility");
+    let mut new = root_request();
+    new.force_start = true;
+    assert_eq!(
+        manager
+            .submit_heal_request(new.clone())
+            .await
+            .expect("force start replacement"),
+        HealAdmissionResult::Accepted
+    );
+    assert!(manager.root_recovery.pending().await.expect("old owner retired").is_empty());
+    manager.stop().await.expect("persist new root only");
+    let restarted = recovery_manager(vec![disk]);
+    restarted.replay_root_heals().await.expect("restart replacement");
+    let ids = restarted
+        .heal_queue
+        .lock()
+        .await
+        .requests()
+        .map(|request| request.id.clone())
+        .collect::<Vec<_>>();
+    assert_eq!(ids, [new.id]);
+}
+
+#[tokio::test]
+async fn root_recovery_force_start_replaces_fresh_queued_and_retrying_admin_roots() {
+    for retrying in [false, true] {
+        let (_temp, disk) = recovery_disk().await;
+        let manager = recovery_manager(vec![disk.clone()]);
+        let old = root_request();
+        if retrying {
+            insert_retrying_request(&manager, old.clone()).await;
+        } else {
+            manager.submit_heal_request(old.clone()).await.expect("queue original root");
+        }
+        assert!(manager.root_recovery.pending().await.expect("not handed off yet").is_empty());
+        let mut new = root_request();
+        new.force_start = true;
+        assert_eq!(
+            manager.submit_heal_request(new.clone()).await.expect("force replacement"),
+            HealAdmissionResult::Accepted
+        );
+        manager.stop().await.expect("handoff only the new responsibility");
+        let restarted = recovery_manager(vec![disk]);
+        restarted.replay_root_heals().await.expect("restart after forceStart");
+        let ids = restarted
+            .heal_queue
+            .lock()
+            .await
+            .requests()
+            .map(|request| request.id.clone())
+            .collect::<Vec<_>>();
+        assert_eq!(ids, [new.id], "retrying={retrying}; old={}", old.id);
+    }
+}
+
+#[tokio::test]
+async fn root_recovery_invalid_records_are_retained_without_partial_replay() {
+    for kind in ["truncated", "schema", "identity", "option", "no_lock"] {
+        let (_temp, disk) = recovery_disk().await;
+        let manager = recovery_manager(vec![disk.clone()]);
+        let valid = root_request();
+        let invalid = root_request();
+        manager.root_recovery.persist(&valid).await.expect("valid root record");
+        manager
+            .root_recovery
+            .persist(&invalid)
+            .await
+            .expect("record before corruption");
+        let path = format!("root-heal-{}.json", invalid.id);
+        let original = disk.read_all(RUSTFS_META_BUCKET, &path).await.expect("read root record");
+        let mut value: serde_json::Value = serde_json::from_slice(&original).expect("record JSON");
+        match kind {
+            "schema" => value["schema"] = 2.into(),
+            "identity" => value["task_id"] = valid.id.clone().into(),
+            "option" => value["options"]["future_delete_mode"] = true.into(),
+            "no_lock" => value["options"]["no_lock"] = true.into(),
+            _ => {}
+        }
+        let bytes = if kind == "truncated" {
+            b"{".to_vec()
+        } else {
+            serde_json::to_vec(&value).expect("modified record")
+        };
+        disk.write_all(RUSTFS_META_BUCKET, &path, bytes.clone().into())
+            .await
+            .expect("inject bad record");
+        assert!(manager.replay_root_heals().await.is_err(), "kind={kind}");
+        assert_eq!(manager.get_queue_length().await, 0, "no partial admission for {kind}");
+        assert_eq!(
+            disk.read_all(RUSTFS_META_BUCKET, &path)
+                .await
+                .expect("bad record retained")
+                .as_ref(),
+            bytes
+        );
+        let mut forced = root_request();
+        forced.force_start = true;
+        assert!(
+            manager.submit_heal_request(forced).await.is_err(),
+            "forceStart must not discard unknown state"
+        );
+    }
+}
+
+#[tokio::test]
+async fn root_recovery_failed_handoff_keeps_runtime_owner_and_does_not_try_another_disk() {
+    let (_temp, disk) = recovery_disk().await;
+    let (unavailable_temp, unavailable) = recovery_disk().await;
+    std::fs::remove_dir_all(unavailable_temp.path().join(RUSTFS_META_BUCKET)).expect("make owner volume unavailable");
+    let manager = recovery_manager(vec![unavailable, disk.clone()]);
+    let task = active_root(&manager, root_request()).await;
+    assert!(manager.stop().await.is_err());
+    assert!(
+        manager.cancel_task(&task.id).await.is_err(),
+        "missing owner cannot acknowledge cancellation"
+    );
+    assert!(!manager.cancel_token.is_cancelled());
+    assert!(!task.cancel_token.is_cancelled());
+    assert!(manager.active_heals.lock().await.contains_key(&task.id));
+    assert!(
+        RootHealRecovery::with_disks(vec![disk])
+            .pending()
+            .await
+            .expect("other disk remains empty")
+            .is_empty()
+    );
+}
+
+#[tokio::test]
+async fn root_recovery_shutdown_fences_new_admission_and_preserves_later_cancellation() {
+    for operation_kind in ["submit", "force_start", "cancel"] {
+        let cancel = operation_kind == "cancel";
+        let (_temp, disk) = recovery_disk().await;
+        let manager = Arc::new(recovery_manager(vec![disk.clone()]));
+        let request = root_request();
+        active_root(&manager, request.clone()).await;
+        let queue = manager.heal_queue.lock().await;
+        let stopping = manager.clone();
+        let stop = tokio::spawn(async move { stopping.stop().await });
+        tokio::time::timeout(Duration::from_secs(5), async {
+            loop {
+                if manager.active_heals.try_lock().is_err() {
+                    break;
+                }
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("shutdown owns active lock while waiting for queue");
+        let concurrent = manager.clone();
+        let operation = tokio::spawn(async move {
+            if cancel {
+                concurrent.cancel_task(&request.id).await
+            } else {
+                let mut new = root_request();
+                new.force_start = operation_kind == "force_start";
+                concurrent.submit_heal_request(new).await.map(|_| ())
+            }
+        });
+        drop(queue);
+        stop.await.expect("shutdown task").expect("durable shutdown");
+        let result = operation.await.expect("concurrent operation");
+        assert_eq!(result.is_ok(), cancel, "operation={operation_kind}");
+        let restarted = recovery_manager(vec![disk]);
+        restarted.replay_root_heals().await.expect("read final responsibility");
+        assert_eq!(restarted.get_queue_length().await, usize::from(!cancel));
+    }
+}
+
+#[tokio::test]
+async fn root_recovery_exhausted_timeout_is_not_reset_by_restart() {
+    let (_temp, disk) = recovery_disk().await;
+    let manager = recovery_manager(vec![disk.clone()]);
+    let mut request = root_request();
+    request.options.timeout = Some(Duration::from_secs(10));
+    let task = active_root(&manager, request.clone()).await;
+    task.set_execution_elapsed_for_test(Duration::from_secs(11)).await;
+    manager.stop().await.expect("persist exhausted execution budget");
+    let restarted = recovery_manager(vec![disk]);
+    restarted.replay_root_heals().await.expect("restore bounded request");
+    assert_eq!(
+        restarted
+            .heal_queue
+            .lock()
+            .await
+            .requests()
+            .next()
+            .expect("restored root")
+            .options
+            .timeout,
+        Some(Duration::ZERO)
+    );
+    process_manager_queue_once(&restarted).await;
+    tokio::time::timeout(Duration::from_secs(5), async {
+        loop {
+            if matches!(restarted.get_task_status(&request.id).await, Ok(HealTaskStatus::Timeout)) {
+                break;
+            }
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .expect("exhausted request stays timed out");
+    restarted
+        .cancel_task(&request.id)
+        .await
+        .expect("timeout responsibility remains cancellable");
+    assert!(restarted.root_recovery.pending().await.expect("retired timeout").is_empty());
+}
+
+#[tokio::test]
+async fn root_recovery_shutdown_preserves_remaining_execution_budget() {
+    let (_temp, disk) = recovery_disk().await;
+    let manager = recovery_manager(vec![disk.clone()]);
+    let mut request = root_request();
+    request.options.timeout = Some(Duration::from_secs(60));
+    let task = active_root(&manager, request).await;
+    task.set_execution_elapsed_for_test(Duration::from_secs(20)).await;
+    manager.stop().await.expect("handoff with consumed execution time");
+    let restarted = recovery_manager(vec![disk]);
+    restarted.replay_root_heals().await.expect("restore remaining budget");
+    let queue = restarted.heal_queue.lock().await;
+    let remaining = queue
+        .requests()
+        .next()
+        .expect("restored root")
+        .options
+        .timeout
+        .expect("remaining timeout");
+    assert!(remaining <= Duration::from_secs(40), "elapsed execution must not be refunded");
+    assert!(
+        remaining >= Duration::from_secs(30),
+        "shutdown fixture should retain most of its remaining budget"
+    );
+}
+
+#[tokio::test]
+async fn root_recovery_force_start_after_shutdown_does_not_retire_original_owner() {
+    let (_temp, disk) = recovery_disk().await;
+    let manager = recovery_manager(vec![disk.clone()]);
+    let old = root_request();
+    active_root(&manager, old.clone()).await;
+    manager.stop().await.expect("handoff original root");
+    let mut new = root_request();
+    new.force_start = true;
+    assert!(manager.submit_heal_request(new).await.is_err());
+    let restarted = recovery_manager(vec![disk]);
+    restarted.replay_root_heals().await.expect("original responsibility remains");
+    let ids = restarted
+        .heal_queue
+        .lock()
+        .await
+        .requests()
+        .map(|request| request.id.clone())
+        .collect::<Vec<_>>();
+    assert_eq!(ids, [old.id]);
+}
+
+#[tokio::test]
+async fn root_recovery_terminal_timeout_updates_only_existing_journal_before_second_restart() {
+    for durable in [false, true] {
+        let (_temp, disk) = recovery_disk().await;
+        let manager = recovery_manager(vec![disk.clone()]);
+        let mut request = root_request();
+        request.options.timeout = Some(Duration::from_nanos(1));
+        if durable {
+            manager
+                .root_recovery
+                .persist(&request)
+                .await
+                .expect("persist nonzero execution budget");
+            manager.replay_root_heals().await.expect("first restart");
+        } else {
+            manager
+                .submit_heal_request(request.clone())
+                .await
+                .expect("first root execution");
+        }
+        process_manager_queue_once(&manager).await;
+        tokio::time::timeout(Duration::from_secs(5), async {
+            loop {
+                if matches!(manager.get_task_status(&request.id).await, Ok(HealTaskStatus::Timeout))
+                    && !manager.active_heals.lock().await.contains_key(&request.id)
+                {
+                    break;
+                }
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("real execution exhausts a nonzero budget");
+        assert!(!manager.active_heals.lock().await.contains_key(&request.id));
+        let restarted = recovery_manager(vec![disk]);
+        restarted
+            .replay_root_heals()
+            .await
+            .expect("second restart after terminal timeout");
+        let queue = restarted.heal_queue.lock().await;
+        if durable {
+            assert_eq!(
+                queue
+                    .requests()
+                    .next()
+                    .expect("remaining timeout responsibility")
+                    .options
+                    .timeout,
+                Some(Duration::ZERO)
+            );
+        } else {
+            assert!(queue.is_empty(), "terminal failure must not create a new durable responsibility");
+        }
+    }
+}
@@ -27,10 +27,10 @@ pub mod task;
 pub mod utils;

 use storage_api::owner::{
-    ECSTORE_BUCKET_META_PREFIX, ECSTORE_DATA_USAGE_CACHE_NAME, ECSTORE_HEALING_MARKER_PATH, ECSTORE_RUSTFS_META_BUCKET,
-    EcstoreConditionalFileUpdate, EcstoreDeleteOptions, EcstoreDiskAPI, EcstoreDiskBytes, EcstoreDiskError, EcstoreDiskOption,
-    EcstoreDiskResult, EcstoreDiskStore, EcstoreEndpoint, EcstoreErrorType, EcstoreStorageError, EcstoreStore, ObjectIO,
-    ObjectOperations, ecstore_local_disk_map_read, ecstore_new_disk,
+    ECSTORE_BUCKET_META_PREFIX, ECSTORE_DATA_USAGE_CACHE_NAME, ECSTORE_HEALING_MARKER_PATH, ECSTORE_POOL_META_NAME,
+    ECSTORE_RUSTFS_META_BUCKET, EcstoreConditionalFileUpdate, EcstoreDeleteOptions, EcstoreDiskAPI, EcstoreDiskBytes,
+    EcstoreDiskError, EcstoreDiskOption, EcstoreDiskResult, EcstoreDiskStore, EcstoreEndpoint, EcstoreErrorType,
+    EcstoreStorageError, EcstoreStore, ObjectIO, ObjectOperations, ecstore_local_disk_map_read, ecstore_new_disk,
 };

 pub use erasure_healer::ErasureSetHealer;
@@ -41,6 +41,7 @@ pub use task::{HealOptions, HealPriority, HealRequest, HealTask, HealType};
 pub(crate) const DATA_USAGE_CACHE_NAME: &str = ECSTORE_DATA_USAGE_CACHE_NAME;
 pub(crate) const BUCKET_META_PREFIX: &str = ECSTORE_BUCKET_META_PREFIX;
 pub(crate) const RUSTFS_META_BUCKET: &str = ECSTORE_RUSTFS_META_BUCKET;
+pub(crate) const POOL_META_NAME: &str = ECSTORE_POOL_META_NAME;

 /// Marker written to every local disk while the process runs; removed by
 /// [`clear_unclean_shutdown_markers`] on graceful shutdown. Finding it at
@@ -110,7 +110,8 @@ impl HealObjectReceipt {
            && self.identity.version_id == expected.version_id
            && self.identity.pool_index == expected.pool_index
            && self.identity.set_index == expected.set_index
-            && self.identity.bucket_incarnation_id.is_some()
+            && self.identity.bucket_incarnation_id == expected.bucket_incarnation_id
+            && expected.bucket_incarnation_id.is_some()
    }
 }

@@ -496,21 +497,31 @@ mod canonical_outcome_tests {

    #[test]
    fn positive_receipt_requires_exact_identity_and_bucket_incarnation() {
-        let expected = item(HealObjectDisposition::Unknown).identity;
+        let incarnation = Uuid::new_v4();
+        let expected = HealObjectIdentity {
+            bucket_incarnation_id: Some(incarnation),
+            ..item(HealObjectDisposition::Unknown).identity
+        };
        let mut receipt = HealObjectReceipt {
            identity: expected.clone(),
            disposition: HealObjectDisposition::Repaired,
        };

+        receipt.identity.bucket_incarnation_id = None;
        assert!(
            !receipt.verified_for(&expected),
            "a positive storage receipt without bucket incarnation must remain untrusted"
        );

-        let incarnation = Uuid::new_v4();
        receipt.identity.bucket_incarnation_id = Some(incarnation);
        assert!(receipt.verified_for(&expected));

+        receipt.identity.bucket_incarnation_id = Some(Uuid::new_v4());
+        assert!(
+            !receipt.verified_for(&expected),
+            "a storage receipt for a different bucket incarnation must not clear the requested responsibility"
+        );
+
        receipt.identity.version_id = Some("older-version".to_string());
        assert!(
            !receipt.verified_for(&expected),
@@ -20,6 +20,7 @@ use rustfs_madmin::heal_commands::HealResultItem;
 use serde::{Deserialize, Serialize};
 use std::sync::Arc;
 use tracing::{debug, error, warn};
+use uuid::Uuid;

 use super::outcome::{HealObjectDisposition, HealObjectIdentity, HealObjectKind, HealObjectReceipt};
 use super::progress::stable_generation;
@@ -355,6 +356,14 @@ pub trait HealStorageAPI: Send + Sync {
    /// Get bucket info
    async fn get_bucket_info(&self, bucket: &str) -> Result<Option<BucketInfo>>;

+    /// Return the current bucket incarnation for exact MRF durable proof
+    /// matching. Alternate backends that cannot expose this must return
+    /// `None`, leaving replay anchors retained instead of acknowledged with an
+    /// incomplete identity.
+    async fn mrf_bucket_incarnation_id(&self, _bucket: &str) -> Result<Option<Uuid>> {
+        Ok(None)
+    }
+
    /// Aggregate usage-cache baselines for the requested buckets.
    async fn erasure_set_usage_baseline(&self, _buckets: &[String]) -> Result<Option<HealBucketUsageBaseline>> {
        Ok(None)
@@ -383,6 +392,11 @@ pub trait HealStorageAPI: Send + Sync {
    /// Check object exists
    async fn object_exists(&self, bucket: &str, object: &str) -> Result<bool>;

+    /// Stable bucket incarnation observed before an object heal starts.
+    async fn bucket_incarnation_id(&self, _bucket: &str) -> Result<Option<Uuid>> {
+        Ok(None)
+    }
+
    /// Heal object using ecstore
    async fn heal_object(
        &self,
@@ -809,6 +823,14 @@ impl HealStorageAPI for ECStoreHealStorage {
        }
    }

+    async fn mrf_bucket_incarnation_id(&self, bucket: &str) -> Result<Option<Uuid>> {
+        self.ecstore
+            .bucket_incarnation_id(bucket)
+            .await
+            .map(Some)
+            .map_err(Error::Storage)
+    }
+
    async fn erasure_set_usage_baseline(&self, buckets: &[String]) -> Result<Option<HealBucketUsageBaseline>> {
        if buckets.is_empty() {
            return Ok(None);
@@ -1028,6 +1050,14 @@ impl HealStorageAPI for ECStoreHealStorage {
        }
    }

+    async fn bucket_incarnation_id(&self, bucket: &str) -> Result<Option<Uuid>> {
+        self.ecstore
+            .bucket_incarnation_id(bucket)
+            .await
+            .map(Some)
+            .map_err(Error::Storage)
+    }
+
    async fn heal_object(
        &self,
        bucket: &str,
@@ -29,6 +29,7 @@ pub(crate) use rustfs_ecstore::api::error::{Error as EcstoreErrorType, StorageEr
 pub(crate) use rustfs_ecstore::api::runtime::local_disk_map_read as ecstore_local_disk_map_read;
 pub(crate) use rustfs_ecstore::api::storage::{
    ECStore as EcstoreStore, HealLifecycleExpiryContext as EcstoreHealLifecycleExpiryContext,
+    POOL_META_NAME as ECSTORE_POOL_META_NAME,
 };
 use rustfs_storage_api as storage_contracts;

@@ -36,10 +37,11 @@ pub(crate) mod owner {
    pub(crate) use super::storage_contracts::{ObjectIO, ObjectOperations};

    pub(crate) use super::{
-        ECSTORE_BUCKET_META_PREFIX, ECSTORE_DATA_USAGE_CACHE_NAME, ECSTORE_HEALING_MARKER_PATH, ECSTORE_RUSTFS_META_BUCKET,
-        EcstoreConditionalFileUpdate, EcstoreDeleteOptions, EcstoreDiskAPI, EcstoreDiskBytes, EcstoreDiskError,
-        EcstoreDiskResult, EcstoreDiskStore, EcstoreEndpoint, EcstoreErrorType, EcstoreHealLifecycleExpiryContext,
-        EcstoreStorageError, EcstoreStore, ecstore_load_admin_data_usage_from_backend_cached, ecstore_local_disk_map_read,
+        ECSTORE_BUCKET_META_PREFIX, ECSTORE_DATA_USAGE_CACHE_NAME, ECSTORE_HEALING_MARKER_PATH, ECSTORE_POOL_META_NAME,
+        ECSTORE_RUSTFS_META_BUCKET, EcstoreConditionalFileUpdate, EcstoreDeleteOptions, EcstoreDiskAPI, EcstoreDiskBytes,
+        EcstoreDiskError, EcstoreDiskResult, EcstoreDiskStore, EcstoreEndpoint, EcstoreErrorType,
+        EcstoreHealLifecycleExpiryContext, EcstoreStorageError, EcstoreStore, ecstore_load_admin_data_usage_from_backend_cached,
+        ecstore_local_disk_map_read,
    };

    pub(crate) use super::{EcstoreDiskOption, ecstore_new_disk};
@@ -45,7 +45,7 @@ use tokio::sync::RwLock;
 use tracing::{debug, error, info, warn};
 use uuid::Uuid;

-use super::{BUCKET_META_PREFIX, DATA_USAGE_CACHE_NAME, RUSTFS_META_BUCKET};
+use super::{BUCKET_META_PREFIX, DATA_USAGE_CACHE_NAME, POOL_META_NAME, RUSTFS_META_BUCKET};

 #[cfg(test)]
 pub(crate) struct OutcomeFinishTestHook {
@@ -106,7 +106,7 @@ const EVENT_HEAL_ERASURE_SET_STAGE: &str = "heal_erasure_set_stage";
 const EVENT_HEAL_ERASURE_SET_RESULT: &str = "heal_erasure_set_result";

 /// Heal type
-#[derive(Debug, Clone)]
+#[derive(Debug, Clone, PartialEq, Eq)]
 pub enum HealType {
    /// Cluster heal
    Cluster,
@@ -209,7 +209,7 @@ impl HealPriority {
 }

 /// Heal options
-#[derive(Debug, Clone, Serialize, Deserialize)]
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
 pub struct HealOptions {
    /// Scan mode
    pub scan_mode: HealScanMode,
@@ -513,6 +513,11 @@ impl HealTask {
        }
    }

+    #[cfg(test)]
+    pub(crate) async fn set_execution_elapsed_for_test(&self, elapsed: Duration) {
+        *self.task_start_instant.write().await = Some(Instant::now() - elapsed);
+    }
+
    pub(crate) async fn retry_request_with_remaining_timeout(&self) -> Result<HealRequest> {
        let mut request = self.retry_request();
        if self.options.timeout.is_some() {
@@ -574,6 +579,18 @@ impl HealTask {
        }
    }

+    pub(super) async fn outcome_bucket_incarnation_id(&self, bucket: &str, dry_run: bool) -> Result<Option<Uuid>> {
+        if dry_run {
+            return Ok(None);
+        }
+        match self.await_with_control(self.storage.bucket_incarnation_id(bucket)).await {
+            Ok(incarnation_id) => Ok(incarnation_id),
+            Err(Error::TaskCancelled) => Err(Error::TaskCancelled),
+            Err(Error::TaskTimeout) => Err(Error::TaskTimeout),
+            Err(_) => Ok(None),
+        }
+    }
+
    fn single_object_identity(&self) -> Option<HealObjectIdentity> {
        let (bucket, object, version) = match &self.heal_type {
            HealType::Object {
@@ -597,6 +614,9 @@ impl HealTask {
        expected: HealObjectIdentity,
        receipt: Option<HealObjectReceipt>,
    ) -> bool {
+        if self.options.dry_run || self.cancel_token.is_cancelled() {
+            return false;
+        }
        let Some(receipt) = receipt else {
            return false;
        };
@@ -340,9 +340,88 @@ impl HealTask {
            return Err(self.record_batch_failure(failure).await);
        }

+        if self.options.recreate_missing && !self.options.dry_run {
+            self.heal_cluster_pool_metadata().await?;
+        }
+
        Ok(())
    }

+    async fn heal_cluster_pool_metadata(&self) -> Result<()> {
+        let heal_opts = HealOpts {
+            recursive: false,
+            dry_run: self.options.dry_run,
+            remove: false,
+            recreate: self.options.recreate_missing,
+            scan_mode: self.options.scan_mode,
+            update_parity: self.options.update_parity,
+            no_lock: self.options.no_lock,
+            read_repair: false,
+            pool: self.options.pool_index,
+            set: self.options.set_index,
+        };
+
+        let heal_result = self
+            .await_with_control(self.storage.heal_object(RUSTFS_META_BUCKET, POOL_META_NAME, None, &heal_opts))
+            .await;
+        match heal_result {
+            Ok((result, None)) => {
+                debug!(
+                    target: "rustfs::heal::task",
+                    event = EVENT_HEAL_BUCKET_RESULT,
+                    component = LOG_COMPONENT_HEAL,
+                    subsystem = LOG_SUBSYSTEM_TASK,
+                    task_id = %self.id,
+                    bucket = RUSTFS_META_BUCKET,
+                    object = POOL_META_NAME,
+                    drives_healed = result.drives_healed(),
+                    drives_total = result.drives_reported(),
+                    result = "pool_metadata_ok",
+                    "Heal cluster pool metadata repaired"
+                );
+                self.record_result_item(result).await;
+                Ok(())
+            }
+            Ok((result, Some(err))) => {
+                self.record_result_item(result).await;
+                warn!(
+                    target: "rustfs::heal::task",
+                    event = EVENT_HEAL_BUCKET_RESULT,
+                    component = LOG_COMPONENT_HEAL,
+                    subsystem = LOG_SUBSYSTEM_TASK,
+                    task_id = %self.id,
+                    bucket = RUSTFS_META_BUCKET,
+                    object = POOL_META_NAME,
+                    result = "pool_metadata_failed",
+                    error = %err,
+                    "Heal cluster pool metadata failed"
+                );
+                Err(Error::TaskExecutionFailed {
+                    message: format!("Failed to heal cluster pool metadata: {err}"),
+                })
+            }
+            Err(Error::TaskCancelled) => Err(Error::TaskCancelled),
+            Err(Error::TaskTimeout) => Err(Error::TaskTimeout),
+            Err(err) => {
+                warn!(
+                    target: "rustfs::heal::task",
+                    event = EVENT_HEAL_BUCKET_RESULT,
+                    component = LOG_COMPONENT_HEAL,
+                    subsystem = LOG_SUBSYSTEM_TASK,
+                    task_id = %self.id,
+                    bucket = RUSTFS_META_BUCKET,
+                    object = POOL_META_NAME,
+                    result = "pool_metadata_failed",
+                    error = %err,
+                    "Heal cluster pool metadata failed"
+                );
+                Err(Error::TaskExecutionFailed {
+                    message: format!("Failed to heal cluster pool metadata: {err}"),
+                })
+            }
+        }
+    }
+
    pub(super) async fn heal_prefix(&self, bucket: &str, prefix: &str) -> Result<()> {
        debug!(
            target: "rustfs::heal::task",
@@ -403,6 +482,7 @@ impl HealTask {
        };

        for (set_disk_id, heal_opts) in listing_scopes {
+            let bucket_incarnation_id = self.outcome_bucket_incarnation_id(bucket, heal_opts.dry_run).await?;
            let mut continuation_token: Option<String> = None;
            let mut deferred = DeferredWindow::default();
            let mut inline_retry: Option<DeferredObject> = None;
@@ -501,13 +581,15 @@ impl HealTask {
                    let retry_attempt = item.attempt;
                    let mut telemetry_unknown = false;
                    let object = item.name.as_str();
-                    let identity =
+                    let mut identity =
                        self.outcome_identity(bucket, object, item.version_id.as_deref(), heal_opts.pool, heal_opts.set);
+                    identity.bucket_incarnation_id = bucket_incarnation_id;
                    let mut disposition = if heal_opts.dry_run {
                        HealObjectDisposition::DryRunObserved
                    } else {
                        HealObjectDisposition::Unknown
                    };
+                    let mut recorded_authoritative_outcome = false;
                    let mut detail = None;
                    {
                        let mut progress = self.progress.write().await;
@@ -520,23 +602,37 @@ impl HealTask {
                        Some(Error::other("heal object retry age exhausted"))
                    } else {
                        match self
-                            .await_with_control(
-                                self.storage
-                                    .heal_object(bucket, object, item.version_id.as_deref(), &heal_opts),
-                            )
+                            .await_with_control(self.storage.heal_object_with_receipt(
+                                bucket,
+                                object,
+                                item.version_id.as_deref(),
+                                &heal_opts,
+                            ))
                            .await
                        {
-                            Ok((result, None)) => match unavailable_recreate_error(&result, &heal_opts) {
-                                Some(error) => Some(error),
-                                None => {
-                                    telemetry_unknown |= !increment_counter(&mut healed);
-                                    telemetry_unknown |=
-                                        !add_bytes(&mut bytes, u64::try_from(result.object_size).unwrap_or(u64::MAX));
-                                    self.record_result_item(result).await;
-                                    None
+                            Ok(storage_result) if storage_result.error.is_none() => {
+                                match unavailable_recreate_error(&storage_result.item, &heal_opts) {
+                                    Some(error) => Some(error),
+                                    None => {
+                                        telemetry_unknown |= !increment_counter(&mut healed);
+                                        telemetry_unknown |= !add_bytes(
+                                            &mut bytes,
+                                            u64::try_from(storage_result.item.object_size).unwrap_or(u64::MAX),
+                                        );
+                                        recorded_authoritative_outcome = self
+                                            .record_verified_storage_receipt(identity.clone(), storage_result.receipt)
+                                            .await;
+                                        self.record_result_item(storage_result.item).await;
+                                        None
+                                    }
                                }
-                            },
-                            Ok((_, Some(err))) if is_missing_object_dir_heal_result(object, &err) => {
+                            }
+                            Ok(storage_result)
+                                if storage_result
+                                    .error
+                                    .as_ref()
+                                    .is_some_and(|err| is_missing_object_dir_heal_result(object, err)) =>
+                            {
                                telemetry_unknown |= !increment_counter(&mut healed);
                                debug!(
                                    target: "rustfs::heal::task",
@@ -551,7 +647,8 @@ impl HealTask {
                                );
                                None
                            }
-                            Ok((_, Some(err))) | Err(err) => Some(err),
+                            Ok(storage_result) => storage_result.error,
+                            Err(err) => Some(err),
                        }
                    };

@@ -674,11 +771,13 @@ impl HealTask {
                        continue;
                    }

-                    self.outcome.write().await.record(HealObjectOutcome {
-                        identity,
-                        disposition,
-                        detail,
-                    });
+                    if !recorded_authoritative_outcome {
+                        self.outcome.write().await.record(HealObjectOutcome {
+                            identity,
+                            disposition,
+                            detail,
+                        });
+                    }

                    let mut progress = self.progress.write().await;
                    progress.update_object_progress(
@@ -451,6 +451,11 @@ impl HealTask {
            self.source,
        )
        .with_replacement_targets(replacement_targets, is_auto_replacement.then(|| self.id.clone()))
+        .with_pool_metadata_targets(if self.options.recreate_missing && !self.options.dry_run {
+            self.heal_endpoints.clone()
+        } else {
+            Vec::new()
+        })
        .with_replacement_identity_fence(replacement_target_identities.clone())
        .with_mainline_pacer(self.mainline_pacer.clone());

@@ -162,6 +162,9 @@ impl HealTask {
            pool: self.options.pool_index,
            set: self.options.set_index,
        };
+        let mut expected_identity =
+            self.outcome_identity(bucket, object, version_id, self.options.pool_index, self.options.set_index);
+        expected_identity.bucket_incarnation_id = self.outcome_bucket_incarnation_id(bucket, self.options.dry_run).await?;

        let heal_fut = self.storage.heal_object_with_receipt(bucket, object, version_id, &heal_opts);
        let heal_result = if self.source == HealRequestSource::ReadRepair {
@@ -266,8 +269,6 @@ impl HealTask {
                    let mut progress = self.progress.write().await;
                    progress.update_object_progress(1, 1, 0, 0, object_size);
                }
-                let expected_identity =
-                    self.outcome_identity(bucket, object, version_id, self.options.pool_index, self.options.set_index);
                self.record_verified_storage_receipt(expected_identity, storage_result.receipt)
                    .await;
                self.record_result_item(result).await;
@@ -66,7 +66,7 @@ mod canonical_outcome {
        assert_eq!(task.get_progress().await.objects_scanned, 2);
        assert_eq!(
            storage.heal_object_calls.lock().expect("object calls").as_slice(),
-            ["object-a", "object-b"]
+            ["object-a", "object-b", POOL_META_NAME]
        );
        assert_eq!(
            storage.listing_tokens.lock().expect("listing tokens").as_slice(),
@@ -154,6 +154,126 @@ mod canonical_outcome {
        );
    }

+    #[tokio::test]
+    async fn bucket_heal_records_matching_positive_storage_receipts() {
+        let incarnation = Uuid::new_v4();
+        let storage = Arc::new(MockStorage {
+            heal_object_receipts: Mutex::new(HashMap::from([
+                (
+                    "object-a".to_string(),
+                    VecDeque::from([object_receipt("object-a", None, HealObjectDisposition::Repaired, incarnation)]),
+                ),
+                (
+                    "object-b".to_string(),
+                    VecDeque::from([object_receipt("object-b", None, HealObjectDisposition::Repaired, incarnation)]),
+                ),
+            ])),
+            bucket_incarnation_id: Mutex::new(Some(incarnation)),
+            ..Default::default()
+        });
+        let task = bucket_task(storage);
+
+        task.execute()
+            .await
+            .expect("bucket heal should record verified object receipts");
+
+        let outcome = task.get_outcome().await;
+        assert_eq!(outcome.execution, HealExecutionOutcome::Completed);
+        assert_eq!(outcome.counters.healed, 2);
+        assert_eq!(outcome.counters.unknown, 0);
+        assert_eq!(outcome.objects.len(), 2);
+        assert!(outcome.objects.iter().all(|item| {
+            item.identity.bucket_incarnation_id == Some(incarnation) && item.disposition == HealObjectDisposition::Repaired
+        }));
+    }
+
+    #[tokio::test]
+    async fn bucket_heal_keeps_repairing_when_bucket_incarnation_is_unavailable() {
+        let storage = Arc::new(MockStorage {
+            heal_object_receipts: Mutex::new(HashMap::from([(
+                "object-a".to_string(),
+                VecDeque::from([object_receipt(
+                    "object-a",
+                    None,
+                    HealObjectDisposition::Repaired,
+                    Uuid::new_v4(),
+                )]),
+            )])),
+            bucket_incarnation_unavailable: Mutex::new(true),
+            ..Default::default()
+        });
+        let task = bucket_task(storage.clone());
+
+        task.execute()
+            .await
+            .expect("bucket heal should continue when only proof ownership is unavailable");
+
+        let outcome = task.get_outcome().await;
+        assert_eq!(outcome.execution, HealExecutionOutcome::Completed);
+        assert_eq!(outcome.counters.healed, 0);
+        assert_eq!(outcome.counters.unknown, 2);
+        assert!(
+            outcome
+                .objects
+                .iter()
+                .all(|item| item.disposition == HealObjectDisposition::Unknown)
+        );
+        assert_eq!(storage.healed_objects.lock().expect("healed objects").len(), 2);
+    }
+
+    #[tokio::test]
+    async fn bucket_heal_rejects_stale_receipts_without_double_recording() {
+        let expected_incarnation = Uuid::new_v4();
+        let storage = Arc::new(MockStorage {
+            heal_object_receipts: Mutex::new(HashMap::from([
+                (
+                    "object-a".to_string(),
+                    VecDeque::from([object_receipt(
+                        "object-a",
+                        None,
+                        HealObjectDisposition::Repaired,
+                        Uuid::new_v4(),
+                    )]),
+                ),
+                (
+                    "object-b".to_string(),
+                    VecDeque::from([object_receipt(
+                        "object-b",
+                        None,
+                        HealObjectDisposition::Repaired,
+                        expected_incarnation,
+                    )]),
+                ),
+            ])),
+            bucket_incarnation_id: Mutex::new(Some(expected_incarnation)),
+            ..Default::default()
+        });
+        let task = bucket_task(storage);
+
+        task.execute()
+            .await
+            .expect("stale bucket receipt should not fail the legacy heal");
+
+        let outcome = task.get_outcome().await;
+        assert_eq!(outcome.execution, HealExecutionOutcome::Completed);
+        assert_eq!(outcome.counters.healed, 1);
+        assert_eq!(outcome.counters.unknown, 1);
+        assert_eq!(outcome.objects.len(), 2);
+        let object_a = outcome
+            .objects
+            .iter()
+            .find(|item| item.identity.object == "object-a")
+            .expect("stale receipt object outcome");
+        assert_eq!(object_a.disposition, HealObjectDisposition::Unknown);
+        let object_b = outcome
+            .objects
+            .iter()
+            .find(|item| item.identity.object == "object-b")
+            .expect("matching receipt object outcome");
+        assert_eq!(object_b.disposition, HealObjectDisposition::Repaired);
+        assert_eq!(object_b.identity.bucket_incarnation_id, Some(expected_incarnation));
+    }
+
    #[tokio::test(start_paused = true)]
    async fn exhausted_object_does_not_abort_other_objects_or_erase_counts() {
        let storage = Arc::new(MockStorage::default());
@@ -313,6 +433,96 @@ mod canonical_outcome {
        );
    }

+    #[tokio::test]
+    async fn mixed_grace_and_repaired_receipt_transfer_only_repaired_responsibility() {
+        let incarnation = Uuid::new_v4();
+        let storage = Arc::new(MockStorage {
+            heal_object_outcomes: Mutex::new(HashMap::from([(
+                "object-a".to_string(),
+                VecDeque::from([MockHealObjectOutcome::DanglingGraceDeferred]),
+            )])),
+            heal_object_receipts: Mutex::new(HashMap::from([(
+                "object-b".to_string(),
+                VecDeque::from([object_receipt("object-b", None, HealObjectDisposition::Repaired, incarnation)]),
+            )])),
+            bucket_incarnation_id: Mutex::new(Some(incarnation)),
+            ..Default::default()
+        });
+        let task = bucket_task(storage);
+
+        task.execute()
+            .await
+            .expect("mixed grace and repaired receipt should complete");
+
+        let outcome = task.get_outcome().await;
+        assert_eq!(outcome.coverage, HealTraversalCoverage::Complete);
+        assert_eq!(outcome.counters.processed, 2);
+        assert_eq!(outcome.counters.healed, 1);
+        assert_eq!(outcome.counters.skipped, 1);
+        let deferred = outcome
+            .objects
+            .iter()
+            .find(|item| item.identity.object == "object-a")
+            .expect("grace object should remain recorded");
+        assert!(matches!(
+            deferred.disposition,
+            HealObjectDisposition::Deferred {
+                reason: HealDeferredReason::DanglingDeleteGrace,
+                ..
+            }
+        ));
+        assert_ne!(deferred.disposition, HealObjectDisposition::Repaired);
+        let repaired = outcome
+            .objects
+            .iter()
+            .find(|item| item.identity.object == "object-b")
+            .expect("receipt-backed object should be recorded");
+        assert_eq!(repaired.identity.bucket_incarnation_id, Some(incarnation));
+        assert_eq!(repaired.disposition, HealObjectDisposition::Repaired);
+    }
+
+    #[tokio::test]
+    async fn bucket_heal_records_matching_positive_storage_receipt() {
+        let incarnation = Uuid::new_v4();
+        let storage = Arc::new(MockStorage {
+            heal_object_receipts: Mutex::new(HashMap::from([(
+                "object-a".to_string(),
+                VecDeque::from([object_receipt("object-a", None, HealObjectDisposition::Repaired, incarnation)]),
+            )])),
+            bucket_incarnation_id: Mutex::new(Some(incarnation)),
+            ..Default::default()
+        });
+        let task = bucket_task(storage);
+
+        task.execute().await.expect("bucket heal should complete");
+
+        let outcome = task.get_outcome().await;
+        assert_eq!(outcome.counters.processed, 2);
+        assert_eq!(outcome.counters.healed, 1);
+        assert_eq!(outcome.counters.unknown, 1);
+        assert_eq!(
+            outcome
+                .objects
+                .iter()
+                .filter(|item| item.identity.object == "object-a")
+                .count(),
+            1
+        );
+        let repaired = outcome
+            .objects
+            .iter()
+            .find(|item| item.identity.object == "object-a")
+            .expect("receipt-backed bucket object should be recorded");
+        assert_eq!(repaired.identity.bucket_incarnation_id, Some(incarnation));
+        assert_eq!(repaired.disposition, HealObjectDisposition::Repaired);
+        let legacy = outcome
+            .objects
+            .iter()
+            .find(|item| item.identity.object == "object-b")
+            .expect("legacy bucket object should still be recorded");
+        assert_eq!(legacy.disposition, HealObjectDisposition::Unknown);
+    }
+
    #[tokio::test]
    async fn grace_single_object_is_completed_but_deferred() {
        let storage = Arc::new(MockStorage {
@@ -344,6 +554,18 @@ mod canonical_outcome {
                    .lock()
                    .expect("existence fixture")
                    .insert("object".to_string(), MockObjectExists::TransientSkip("retry later"));
+            } else {
+                let incarnation = Uuid::new_v4();
+                *storage.bucket_incarnation_id.lock().expect("bucket incarnation") = Some(incarnation);
+                storage.heal_object_receipts.lock().expect("receipt fixture").insert(
+                    "object".to_string(),
+                    VecDeque::from([object_receipt(
+                        "object",
+                        None,
+                        HealObjectDisposition::VerifiedHealthy,
+                        incarnation,
+                    )]),
+                );
            }
            let mut request = HealRequest::object("bucket-a".to_string(), "object".to_string(), None);
            request.options.dry_run = !transient;
@@ -1116,6 +1338,9 @@ struct MockStorage {
    heal_object_outcome: Mutex<Option<MockHealObjectOutcome>>,
    heal_object_outcomes: Mutex<HashMap<String, VecDeque<MockHealObjectOutcome>>>,
    heal_object_receipts: Mutex<HashMap<String, VecDeque<HealObjectReceipt>>>,
+    bucket_incarnation_id: Mutex<Option<Uuid>>,
+    bucket_incarnation_after_object_heal: Mutex<Option<Uuid>>,
+    bucket_incarnation_unavailable: Mutex<bool>,
    format_no_heal_required: Mutex<bool>,
    format_error: Mutex<Option<Error>>,
    global_format_calls: Mutex<u32>,
@@ -1219,14 +1444,19 @@ async fn execute_emits_heal_trace_task_state() {
    assert_eq!(trace_attr_string(&completed, "state").as_deref(), Some("completed"));
 }

-fn object_receipt(object: &str, version_id: Option<&str>, disposition: HealObjectDisposition) -> HealObjectReceipt {
+fn object_receipt(
+    object: &str,
+    version_id: Option<&str>,
+    disposition: HealObjectDisposition,
+    bucket_incarnation_id: Uuid,
+) -> HealObjectReceipt {
    HealObjectReceipt {
        identity: HealObjectIdentity {
            kind: HealObjectKind::Object,
            bucket: "bucket-a".to_string(),
            object: object.to_string(),
            version_id: version_id.map(ToOwned::to_owned),
-            bucket_incarnation_id: Some(Uuid::new_v4()),
+            bucket_incarnation_id: Some(bucket_incarnation_id),
            pool_index: None,
            set_index: None,
        },
@@ -1236,11 +1466,18 @@ fn object_receipt(object: &str, version_id: Option<&str>, disposition: HealObjec

 #[tokio::test]
 async fn object_heal_records_matching_positive_storage_receipt() {
+    let incarnation = Uuid::new_v4();
    let storage = Arc::new(MockStorage {
        heal_object_receipts: Mutex::new(HashMap::from([(
            "object-a".to_string(),
-            VecDeque::from([object_receipt("object-a", Some("version-a"), HealObjectDisposition::Repaired)]),
+            VecDeque::from([object_receipt(
+                "object-a",
+                Some("version-a"),
+                HealObjectDisposition::Repaired,
+                incarnation,
+            )]),
        )])),
+        bucket_incarnation_id: Mutex::new(Some(incarnation)),
        ..Default::default()
    });
    let task = HealTask::from_request(
@@ -1261,16 +1498,165 @@ async fn object_heal_records_matching_positive_storage_receipt() {
 }

 #[tokio::test]
-async fn object_heal_rejects_mismatched_or_legacy_storage_receipts() {
+async fn cancelled_object_heal_rejects_matching_positive_storage_receipt() {
+    let incarnation = Uuid::new_v4();
+    let storage = Arc::new(MockStorage::default());
+    let task = HealTask::from_request(
+        HealRequest::object("bucket-a".to_string(), "object-a".to_string(), Some("version-a".to_string())),
+        storage,
+    );
+    task.cancel().await.expect("task cancellation should succeed");
+
+    let expected = HealObjectIdentity {
+        kind: HealObjectKind::Object,
+        bucket: "bucket-a".to_string(),
+        object: "object-a".to_string(),
+        version_id: Some("version-a".to_string()),
+        bucket_incarnation_id: Some(incarnation),
+        pool_index: None,
+        set_index: None,
+    };
+    let accepted = task
+        .record_verified_storage_receipt(
+            expected,
+            Some(object_receipt(
+                "object-a",
+                Some("version-a"),
+                HealObjectDisposition::Repaired,
+                incarnation,
+            )),
+        )
+        .await;
+
+    let outcome = task.get_outcome().await;
+    assert!(!accepted);
+    assert_eq!(outcome.counters.healed, 0);
+    assert!(outcome.objects.is_empty());
+}
+
+#[tokio::test]
+async fn failed_object_heal_rejects_matching_positive_storage_receipt() {
+    let incarnation = Uuid::new_v4();
+    let storage = Arc::new(MockStorage {
+        heal_object_outcome: Mutex::new(Some(MockHealObjectOutcome::OkWithOtherError("commit failed"))),
+        heal_object_receipts: Mutex::new(HashMap::from([(
+            "object-a".to_string(),
+            VecDeque::from([object_receipt(
+                "object-a",
+                Some("version-a"),
+                HealObjectDisposition::Repaired,
+                incarnation,
+            )]),
+        )])),
+        bucket_incarnation_id: Mutex::new(Some(incarnation)),
+        ..Default::default()
+    });
+    let task = HealTask::from_request(
+        HealRequest::object("bucket-a".to_string(), "object-a".to_string(), Some("version-a".to_string())),
+        storage,
+    );
+
+    let result = task.execute().await;
+
+    let outcome = task.get_outcome().await;
+    assert!(result.is_err());
+    assert_eq!(outcome.counters.healed, 0);
+    assert_eq!(outcome.counters.unchanged, 0);
+    assert!(outcome.objects.iter().all(|object| {
+        !matches!(
+            object.disposition,
+            HealObjectDisposition::Repaired
+                | HealObjectDisposition::VerifiedHealthy
+                | HealObjectDisposition::AuthoritativelyAbsent
+        )
+    }));
+}
+
+#[tokio::test]
+async fn transient_quorum_object_heal_rejects_matching_positive_storage_receipt() {
+    let incarnation = Uuid::new_v4();
+    let storage = Arc::new(MockStorage {
+        heal_object_outcome: Mutex::new(Some(MockHealObjectOutcome::OkWithReadQuorum)),
+        heal_object_receipts: Mutex::new(HashMap::from([(
+            "object-a".to_string(),
+            VecDeque::from([object_receipt(
+                "object-a",
+                Some("version-a"),
+                HealObjectDisposition::Repaired,
+                incarnation,
+            )]),
+        )])),
+        bucket_incarnation_id: Mutex::new(Some(incarnation)),
+        ..Default::default()
+    });
+    let task = HealTask::from_request(
+        HealRequest::object("bucket-a".to_string(), "object-a".to_string(), Some("version-a".to_string())),
+        storage,
+    );
+
+    let result = task.execute().await;
+
+    let outcome = task.get_outcome().await;
+    assert!(result.is_err());
+    assert_eq!(outcome.counters.healed, 0);
+    assert_eq!(outcome.counters.unchanged, 0);
+    assert!(outcome.objects.iter().all(|object| {
+        !matches!(
+            object.disposition,
+            HealObjectDisposition::Repaired
+                | HealObjectDisposition::VerifiedHealthy
+                | HealObjectDisposition::AuthoritativelyAbsent
+        )
+    }));
+}
+
+#[tokio::test]
+async fn object_heal_latches_expected_incarnation_before_repair() {
+    let original_incarnation = Uuid::new_v4();
+    let successor_incarnation = Uuid::new_v4();
    let storage = Arc::new(MockStorage {
        heal_object_receipts: Mutex::new(HashMap::from([(
            "object-a".to_string(),
            VecDeque::from([object_receipt(
                "object-a",
-                Some("old-version"),
-                HealObjectDisposition::Repaired,
+                Some("version-a"),
+                HealObjectDisposition::VerifiedHealthy,
+                original_incarnation,
            )]),
        )])),
+        bucket_incarnation_id: Mutex::new(Some(original_incarnation)),
+        bucket_incarnation_after_object_heal: Mutex::new(Some(successor_incarnation)),
+        ..Default::default()
+    });
+    let task = HealTask::from_request(
+        HealRequest::object("bucket-a".to_string(), "object-a".to_string(), Some("version-a".to_string())),
+        storage,
+    );
+
+    task.execute().await.expect("mock object heal should complete");
+
+    let outcome = task.get_outcome().await;
+    assert_eq!(outcome.counters.unchanged, 1);
+    assert_eq!(outcome.counters.unknown, 0);
+    let object = outcome.objects.front().expect("latched receipt should be recorded");
+    assert_eq!(object.identity.bucket_incarnation_id, Some(original_incarnation));
+    assert_eq!(object.disposition, HealObjectDisposition::VerifiedHealthy);
+}
+
+#[tokio::test]
+async fn object_heal_rejects_mismatched_or_legacy_storage_receipts() {
+    let expected_incarnation = Uuid::new_v4();
+    let storage = Arc::new(MockStorage {
+        heal_object_receipts: Mutex::new(HashMap::from([(
+            "object-a".to_string(),
+            VecDeque::from([object_receipt(
+                "object-a",
+                Some("version-a"),
+                HealObjectDisposition::Repaired,
+                Uuid::new_v4(),
+            )]),
+        )])),
+        bucket_incarnation_id: Mutex::new(Some(expected_incarnation)),
        ..Default::default()
    });
    let task = HealTask::from_request(
@@ -1361,6 +1747,7 @@ enum MockHealObjectOutcome {
    RetryableLock,
    RetryableLockTimeout,
    OkWithOtherError(&'static str),
+    OkWithReadQuorum,
    ErrOther(&'static str),
    DanglingGraceDeferred,
    UnavailableDrive(DriveState),
@@ -1456,6 +1843,13 @@ impl HealStorageAPI for MockStorage {
        Ok(self.object_exists.lock().unwrap().unwrap_or(true))
    }

+    async fn bucket_incarnation_id(&self, _bucket: &str) -> Result<Option<Uuid>> {
+        if *self.bucket_incarnation_unavailable.lock().unwrap() {
+            return Err(Error::Other("bucket incarnation unavailable".to_string()));
+        }
+        Ok(*self.bucket_incarnation_id.lock().unwrap())
+    }
+
    async fn heal_object(
        &self,
        bucket: &str,
@@ -1513,6 +1907,13 @@ impl HealStorageAPI for MockStorage {
                MockHealObjectOutcome::RetryableSlowDown => {
                    Ok((HealResultItem::default(), Some(Error::Storage(EcstoreError::SlowDown))))
                }
+                MockHealObjectOutcome::OkWithReadQuorum => Ok((
+                    HealResultItem::default(),
+                    Some(Error::Storage(EcstoreError::InsufficientReadQuorum(
+                        bucket.to_string(),
+                        object.to_string(),
+                    ))),
+                )),
                MockHealObjectOutcome::PermanentOther(message) => Err(Error::other(message)),
                MockHealObjectOutcome::OkWithOtherError(message) => Ok((HealResultItem::default(), Some(Error::other(message)))),
                MockHealObjectOutcome::ErrOther(message) => Err(Error::other(message)),
@@ -1552,6 +1953,13 @@ impl HealStorageAPI for MockStorage {
                MockHealObjectOutcome::RetryableSlowDown => {
                    Ok((HealResultItem::default(), Some(Error::Storage(EcstoreError::SlowDown))))
                }
+                MockHealObjectOutcome::OkWithReadQuorum => Ok((
+                    HealResultItem::default(),
+                    Some(Error::Storage(EcstoreError::InsufficientReadQuorum(
+                        bucket.to_string(),
+                        object.to_string(),
+                    ))),
+                )),
            };
        }
        if bucket == RUSTFS_META_BUCKET && object == format!("{BUCKET_META_PREFIX}/{DATA_USAGE_CACHE_NAME}") {
@@ -1566,6 +1974,9 @@ impl HealStorageAPI for MockStorage {
            return Ok((HealResultItem::default(), Some(Error::Disk(DiskError::FileNotFound))));
        }
        self.healed_objects.lock().unwrap().push(object.to_string());
+        if let Some(bucket_incarnation_id) = self.bucket_incarnation_after_object_heal.lock().unwrap().take() {
+            *self.bucket_incarnation_id.lock().unwrap() = Some(bucket_incarnation_id);
+        }
        Ok((
            HealResultItem {
                object_size: 1,
@@ -2171,11 +2582,95 @@ async fn test_cluster_heal_visits_bucket_objects() {

    assert_eq!(
        storage.healed_objects.lock().unwrap().as_slice(),
-        ["object-a".to_string(), "object-b".to_string()]
+        ["object-a".to_string(), "object-b".to_string(), POOL_META_NAME.to_string()]
    );
    assert!(matches!(task.get_status().await, HealTaskStatus::Completed));
 }

+#[tokio::test]
+async fn cluster_recreate_heals_pool_metadata_after_user_buckets() {
+    let storage = Arc::new(MockStorage::default());
+    let request = HealRequest::new(
+        HealType::Cluster,
+        HealOptions {
+            recursive: true,
+            recreate_missing: true,
+            timeout: None,
+            ..Default::default()
+        },
+        HealPriority::Normal,
+    );
+    let task = HealTask::from_request(request, storage.clone());
+
+    task.execute()
+        .await
+        .expect("cluster recreate heal should include pool metadata");
+
+    assert_eq!(
+        storage.heal_object_calls.lock().expect("object calls").as_slice(),
+        ["object-a".to_string(), "object-b".to_string(), POOL_META_NAME.to_string()]
+    );
+    let opts = storage.object_heal_opts.lock().expect("object opts");
+    assert!(opts.last().expect("pool metadata opts").recreate);
+}
+
+#[tokio::test]
+async fn cluster_recreate_fails_when_pool_metadata_heal_fails() {
+    let storage = Arc::new(MockStorage::default());
+    storage.heal_object_outcomes.lock().expect("object outcomes").insert(
+        POOL_META_NAME.to_string(),
+        VecDeque::from([MockHealObjectOutcome::ErrOther("pool metadata missing")]),
+    );
+    let request = HealRequest::new(
+        HealType::Cluster,
+        HealOptions {
+            recursive: true,
+            recreate_missing: true,
+            timeout: None,
+            ..Default::default()
+        },
+        HealPriority::Normal,
+    );
+    let task = HealTask::from_request(request, storage.clone());
+
+    let err = task
+        .execute()
+        .await
+        .expect_err("cluster recreate heal must not hide pool metadata failure");
+
+    assert!(matches!(err, Error::TaskExecutionFailed { .. }));
+    assert_eq!(
+        storage.heal_object_calls.lock().expect("object calls").as_slice(),
+        ["object-a".to_string(), "object-b".to_string(), POOL_META_NAME.to_string()]
+    );
+}
+
+#[tokio::test]
+async fn cluster_dry_run_does_not_heal_pool_metadata() {
+    let storage = Arc::new(MockStorage::default());
+    let request = HealRequest::new(
+        HealType::Cluster,
+        HealOptions {
+            recursive: true,
+            dry_run: true,
+            recreate_missing: true,
+            timeout: None,
+            ..Default::default()
+        },
+        HealPriority::Normal,
+    );
+    let task = HealTask::from_request(request, storage.clone());
+
+    task.execute()
+        .await
+        .expect("dry-run cluster heal should preserve existing coverage");
+
+    assert_eq!(
+        storage.heal_object_calls.lock().expect("object calls").as_slice(),
+        ["object-a".to_string(), "object-b".to_string()]
+    );
+}
+
 #[tokio::test]
 async fn object_heal_skips_dangling_delete_grace_without_failing_task() {
    let storage = Arc::new(MockStorage {
@@ -61,10 +61,14 @@ pub fn create_ahm_services_cancel_token() -> CancellationToken {
 }

 /// Shutdown all heal services gracefully
-pub fn shutdown_ahm_services() {
+pub async fn shutdown_ahm_services() -> Result<()> {
+    if let Some(manager) = get_heal_manager() {
+        manager.stop().await?;
+    }
    if let Some(cancel_token) = GLOBAL_AHM_SERVICES_CANCEL_TOKEN.get() {
        cancel_token.cancel();
    }
+    Ok(())
 }

 struct HealRuntime {
@@ -22,13 +22,14 @@
 //! Under `cargo nextest` each test runs in its own process, which keeps the
 //! process-global MRF channel singleton safe.

-use rustfs_common::mrf_channel::{self, MrfKind};
+use rustfs_common::mrf_channel::{self, MrfIngressResult, MrfKind, MrfScope};
 use rustfs_heal::heal::{
    manager::{HealConfig, HealManager},
    mrf_queue,
    storage::{ECStoreHealStorage, HealStorageAPI},
 };
 use serial_test::serial;
+use sha2::{Digest, Sha256};
 #[cfg(unix)]
 use std::{
    fs::{File, OpenOptions},
@@ -48,6 +49,12 @@ use storage_api::endpoint_index::{Endpoint, EndpointServerPools, Endpoints, Pool
 const META_BUCKET: &str = ".rustfs.sys";
 const JOURNAL_REL: &str = "buckets/.heal/mrf/journal.bin";
 const SCOPED_JOURNAL_REL: &str = "buckets/.heal/mrf/journal-scoped.bin";
+const COMMITTED_PAYLOAD_REL: &str = ".heal-mrf-snapshot.0.bin";
+const COMMITTED_MANIFEST_REL: &str = ".heal-mrf-commit.0.bin";
+const COMMITTED_PAYLOAD_RELS: [&str; 2] = [".heal-mrf-snapshot.0.bin", ".heal-mrf-snapshot.1.bin"];
+const COMMITTED_MANIFEST_RELS: [&str; 2] = [".heal-mrf-commit.0.bin", ".heal-mrf-commit.1.bin"];
+const COMMITTED_MAGIC: &[u8; 8] = b"RFMRFC01";
+const COMMITTED_MANIFEST_LEN: usize = 8 + 1 + 16 + 8 + 8 + 32 + 32;

 async fn heal_env() -> (Vec<std::path::PathBuf>, Arc<dyn HealStorageAPI>) {
    heal_env_at(None).await
@@ -195,6 +202,29 @@ fn write_journal_to_disks(disk_paths: &[std::path::PathBuf], data: &[u8]) {
    write_journal_path_to_disks(disk_paths, JOURNAL_REL, data);
 }

+fn committed_manifest(owner: uuid::Uuid, sequence: u64, payload: &[u8]) -> Vec<u8> {
+    let mut manifest = Vec::with_capacity(COMMITTED_MANIFEST_LEN);
+    manifest.extend_from_slice(COMMITTED_MAGIC);
+    manifest.push(1);
+    manifest.extend_from_slice(owner.as_bytes());
+    manifest.extend_from_slice(&sequence.to_le_bytes());
+    manifest.extend_from_slice(
+        &u64::try_from(payload.len())
+            .expect("fixture payload length fits")
+            .to_le_bytes(),
+    );
+    manifest.extend_from_slice(&Sha256::digest(payload));
+    manifest.extend_from_slice(&Sha256::digest(&manifest));
+    assert_eq!(manifest.len(), COMMITTED_MANIFEST_LEN, "committed fixture manifest length");
+    manifest
+}
+
+fn write_committed_snapshot_to_disks(disk_paths: &[std::path::PathBuf], sequence: u64, payload: &[u8]) {
+    let manifest = committed_manifest(uuid::Uuid::new_v4(), sequence, payload);
+    write_journal_path_to_disks(disk_paths, COMMITTED_PAYLOAD_REL, payload);
+    write_journal_path_to_disks(disk_paths, COMMITTED_MANIFEST_REL, &manifest);
+}
+
 fn journal_exists_on_all_disks(disk_paths: &[std::path::PathBuf], relative_path: &str) -> bool {
    disk_paths
        .iter()
@@ -207,6 +237,71 @@ fn journal_matches_on_all_disks(disk_paths: &[PathBuf], relative_path: &str, exp
        .all(|path| std::fs::read(path.join(META_BUCKET).join(relative_path)).is_ok_and(|actual| actual == expected))
 }

+fn journal_contains_on_all_disks(disk_paths: &[PathBuf], relative_path: &str, needle: &[u8]) -> bool {
+    disk_paths.iter().all(|path| {
+        std::fs::read(path.join(META_BUCKET).join(relative_path))
+            .is_ok_and(|actual| actual.windows(needle.len()).any(|window| window == needle))
+    })
+}
+
+fn journal_contains_on_any_disk(disk_paths: &[PathBuf], relative_path: &str, needle: &[u8]) -> bool {
+    disk_paths.iter().any(|path| {
+        std::fs::read(path.join(META_BUCKET).join(relative_path))
+            .is_ok_and(|actual| actual.windows(needle.len()).any(|window| window == needle))
+    })
+}
+
+fn committed_payload_contains_on_all_disks(disk_paths: &[PathBuf], needles: &[&[u8]]) -> bool {
+    disk_paths.iter().all(|path| {
+        let root = path.join(META_BUCKET);
+        COMMITTED_PAYLOAD_RELS.into_iter().any(|payload_rel| {
+            std::fs::read(root.join(payload_rel)).is_ok_and(|payload| {
+                needles
+                    .iter()
+                    .all(|needle| payload.windows(needle.len()).any(|window| window == *needle))
+            })
+        })
+    })
+}
+
+fn committed_checkpoint_matches_on_all_disks(disk_paths: &[PathBuf], sequence: u64, expected_payload: &[u8]) -> bool {
+    disk_paths.iter().all(|path| {
+        let root = path.join(META_BUCKET);
+        COMMITTED_PAYLOAD_RELS
+            .into_iter()
+            .zip(COMMITTED_MANIFEST_RELS)
+            .any(|(payload_rel, manifest_rel)| {
+                let Ok(payload) = std::fs::read(root.join(payload_rel)) else {
+                    return false;
+                };
+                if payload != expected_payload {
+                    return false;
+                }
+                let Ok(manifest) = std::fs::read(root.join(manifest_rel)) else {
+                    return false;
+                };
+                if manifest.len() != COMMITTED_MANIFEST_LEN || &manifest[..8] != COMMITTED_MAGIC || manifest[8] != 1 {
+                    return false;
+                }
+                let Ok(recorded_sequence) = <[u8; 8]>::try_from(&manifest[25..33]).map(u64::from_le_bytes) else {
+                    return false;
+                };
+                let Ok(recorded_len) = <[u8; 8]>::try_from(&manifest[33..41]).map(u64::from_le_bytes) else {
+                    return false;
+                };
+                let Ok(expected_len) = u64::try_from(expected_payload.len()) else {
+                    return false;
+                };
+                if recorded_sequence != sequence || recorded_len != expected_len {
+                    return false;
+                }
+                let payload_digest: [u8; 32] = Sha256::digest(expected_payload).into();
+                let manifest_digest: [u8; 32] = Sha256::digest(&manifest[..COMMITTED_MANIFEST_LEN - 32]).into();
+                payload_digest.as_slice() == &manifest[41..73] && manifest_digest.as_slice() == &manifest[73..]
+            })
+    })
+}
+
 async fn wait_until<F, Fut>(deadline: Duration, mut probe: F) -> bool
 where
    F: FnMut() -> Fut,
@@ -255,11 +350,12 @@ async fn decode_failure_intent_maps_to_urgent_mrf_heal_request() {
 }

 /// A journal left behind by a previous process must be replayed into the
-/// manager queue and then removed, and a torn tail must not block replay of
-/// the intact records.
+/// manager queue, and a torn tail must not block replay of the intact records.
+/// The partial-write record keeps the legacy journal as the durable anchor
+/// until an exact verified repair proof can discharge it.
 #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
 #[serial]
-async fn journal_replay_arms_intents_and_deletes_the_file() {
+async fn journal_replay_arms_intents_and_retains_unproven_partial_write_anchor() {
    let (disk_paths, storage) = heal_env().await;

    // The journal reader resolves disks through the process-local disk map;
@@ -285,14 +381,14 @@ async fn journal_replay_arms_intents_and_deletes_the_file() {
    assert!(
        disk_paths
            .iter()
-            .all(|path| !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()),
-        "the journal file must be removed after a successful replay"
+            .all(|path| Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()),
+        "partial-write replay must retain the legacy journal until durable proof"
    );
    assert!(
        disk_paths
            .iter()
            .all(|path| !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()),
-        "the authoritative journal file must also be removed after replay"
+        "missing authoritative journal remains absent"
    );

    let snapshot = manager.operations_snapshot().await;
@@ -300,6 +396,76 @@ async fn journal_replay_arms_intents_and_deletes_the_file() {
    assert!(snapshot.queued_by_priority.normal >= 1, "the partial-write record must replay as Normal");
 }

+/// A committed checkpoint published by the new two-slot writer is the
+/// authoritative startup snapshot. Legacy mirrors are fallback-only and must
+/// not be merged with or preferred over the committed epoch.
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+#[serial]
+async fn committed_snapshot_replay_takes_precedence_over_stale_legacy_mirror() {
+    let (disk_paths, storage) = heal_env().await;
+    register_local_disks(&disk_paths, "mrf-committed-replay-test").await;
+
+    let committed = scoped_journal_record(3, "committed-bucket", "committed-object", Some([9u8; 16]), 0, 0, 0);
+    let stale_legacy = journal_record(1, "legacy-bucket", "legacy-object", None, 0);
+    write_committed_snapshot_to_disks(&disk_paths, 7, &committed);
+    write_journal_path_to_disks(&disk_paths, SCOPED_JOURNAL_REL, &stale_legacy);
+    write_journal_path_to_disks(&disk_paths, JOURNAL_REL, &stale_legacy);
+
+    let manager = make_manager(storage);
+    let replayed = mrf_queue::replay_journal_once(&manager).await;
+    assert_eq!(replayed, 1, "only the committed snapshot epoch may replay");
+
+    let snapshot = manager.operations_snapshot().await;
+    assert_eq!(snapshot.queued_by_source.mrf, 1);
+    assert_eq!(
+        snapshot.queued_by_priority.normal, 1,
+        "the committed partial-write record must replay instead of the stale legacy decode-failure"
+    );
+    assert_eq!(
+        snapshot.queued_by_priority.urgent, 0,
+        "stale legacy decode-failure records must not be mixed into committed replay"
+    );
+    assert!(
+        journal_exists_on_all_disks(&disk_paths, COMMITTED_MANIFEST_REL),
+        "the committed checkpoint remains until the accepted partial-write has proof"
+    );
+}
+
+/// A damaged committed checkpoint is ambiguous: replay must not fall back to
+/// older legacy bytes or delete any recovery anchor until another process can
+/// publish a valid successor.
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+#[serial]
+async fn damaged_committed_snapshot_blocks_legacy_fallback_and_retains_anchors() {
+    let (disk_paths, storage) = heal_env().await;
+    register_local_disks(&disk_paths, "mrf-damaged-committed-replay-test").await;
+
+    let committed = scoped_journal_record(3, "damaged-committed-bucket", "committed-object", Some([8u8; 16]), 0, 0, 0);
+    let stale_legacy = journal_record(1, "damaged-legacy-bucket", "legacy-object", None, 0);
+    write_journal_path_to_disks(&disk_paths, COMMITTED_PAYLOAD_REL, &committed);
+    let mut manifest = committed_manifest(uuid::Uuid::new_v4(), 9, &committed);
+    manifest[25] ^= 1;
+    write_journal_path_to_disks(&disk_paths, COMMITTED_MANIFEST_REL, &manifest);
+    write_journal_path_to_disks(&disk_paths, SCOPED_JOURNAL_REL, &stale_legacy);
+    write_journal_path_to_disks(&disk_paths, JOURNAL_REL, &stale_legacy);
+
+    let manager = make_manager(storage);
+    let replayed = mrf_queue::replay_journal_once(&manager).await;
+    assert_eq!(replayed, 0, "damaged committed state must fail closed");
+    assert_eq!(
+        manager.operations_snapshot().await.queued_by_source.mrf,
+        0,
+        "stale legacy bytes must not be replayed when committed state is ambiguous"
+    );
+    assert!(
+        journal_exists_on_all_disks(&disk_paths, COMMITTED_MANIFEST_REL)
+            && journal_exists_on_all_disks(&disk_paths, COMMITTED_PAYLOAD_REL)
+            && journal_matches_on_all_disks(&disk_paths, SCOPED_JOURNAL_REL, &stale_legacy)
+            && journal_matches_on_all_disks(&disk_paths, JOURNAL_REL, &stale_legacy),
+        "all recovery anchors must remain after a fail-closed committed read"
+    );
+}
+
 /// A canonical snapshot and its compatibility mirror may differ after a
 /// partial flush. Replay must choose the complete canonical epoch instead of
 /// combining records that never coexisted in memory.
@@ -322,21 +488,24 @@ async fn authoritative_journal_is_not_merged_with_legacy_mirror() {
    assert_eq!(snapshot.queued_by_source.mrf, 1);
    assert!(
        disk_paths.iter().all(|path| {
-            !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
-                && !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
+            Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
+                && Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
        }),
-        "replay cleanup must remove both journal paths"
+        "accepted replay responsibilities remain anchored until a verified repair proof"
    );

    // A scoped-only snapshot is valid during a rollout where no legacy
    // compatibility mirror was written. Missing legacy files must not leave
    // the runtime in a permanent cleanup-retry state.
+    let (disk_paths, storage) = heal_env().await;
+    register_local_disks(&disk_paths, "mrf-scoped-authoritative-test").await;
+    let manager = make_manager(storage);
    let scoped_only = journal_record(1, "scoped-only-bucket", "scoped-only-object", None, 0);
    write_journal_path_to_disks(&disk_paths, SCOPED_JOURNAL_REL, &scoped_only);
    assert_eq!(mrf_queue::replay_journal_once(&manager).await, 1);
    assert!(disk_paths.iter().all(|path| {
        !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
-            && !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
+            && Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
    }));

    let scoped_v2 = scoped_journal_record(1, "scoped-v2-bucket", "scoped-v2-object", None, 0, 3, 7);
@@ -350,12 +519,12 @@ async fn authoritative_journal_is_not_merged_with_legacy_mirror() {
    );
    assert_eq!(
        manager.operations_snapshot().await.queued_by_source.mrf,
-        3,
-        "only the three authoritative/scoped-only epochs should have reached the manager"
+        2,
+        "only the scoped-only and scoped-v2 authoritative epochs should have reached the manager"
    );
    assert!(disk_paths.iter().all(|path| {
-        !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
-            && !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
+        Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
+            && Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
    }));
 }

@@ -400,10 +569,13 @@ async fn authoritative_journal_replay_preserves_kind_and_scope_identity() {
        snapshot.queued_by_priority.urgent, 1,
        "decode-failure repair must not merge with object repair responsibility"
    );
-    assert!(disk_paths.iter().all(|path| {
-        !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
-            && !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
-    }));
+    assert!(
+        disk_paths.iter().all(|path| {
+            Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
+                && Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
+        }),
+        "partial-write responsibilities keep both replay anchors until proof"
+    );
 }

 /// If replay reaches a full heal-manager queue, the old journal remains the
@@ -470,6 +642,73 @@ async fn journal_replay_retains_file_when_manager_is_full() {
    );
 }

+/// Rollback mirrors are for v1 readers only: the committed and scoped
+/// snapshots remain authoritative, while the legacy journal omits scoped-only
+/// records that an older binary cannot represent safely.
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+#[serial]
+async fn rollback_legacy_mirror_persists_only_v1_compatible_records() {
+    let (disk_paths, storage) = heal_env().await;
+    register_local_disks(&disk_paths, "mrf-rollback-mirror-test").await;
+
+    let manager = Arc::new(HealManager::new(
+        storage,
+        Some(HealConfig {
+            queue_size: 0,
+            heal_interval: Duration::from_secs(3600),
+            enable_auto_heal: false,
+            ..Default::default()
+        }),
+    ));
+    mrf_queue::spawn_mrf_consumer(manager.clone());
+
+    let scoped_only = b"rollback-scoped-only-object";
+    let v1_compatible = b"rollback-v1-compatible-object";
+    assert_eq!(
+        mrf_channel::try_send_mrf_intent_typed(
+            MrfKind::PartialWrite,
+            "rollback-bucket",
+            std::str::from_utf8(scoped_only).expect("fixture object is UTF-8"),
+            None,
+            Some(MrfScope {
+                pool_index: 3,
+                set_index: 7,
+            }),
+        ),
+        MrfIngressResult::Enqueued,
+        "scoped-only intent should be accepted by the live consumer"
+    );
+    assert_eq!(
+        mrf_channel::try_send_mrf_intent_typed(
+            MrfKind::PartialWrite,
+            "rollback-bucket",
+            std::str::from_utf8(v1_compatible).expect("fixture object is UTF-8"),
+            None,
+            None,
+        ),
+        MrfIngressResult::Enqueued,
+        "v1-compatible intent should be accepted by the live consumer"
+    );
+
+    let flushed = wait_until(Duration::from_secs(10), || async {
+        committed_payload_contains_on_all_disks(&disk_paths, &[scoped_only, v1_compatible])
+            && journal_contains_on_all_disks(&disk_paths, SCOPED_JOURNAL_REL, scoped_only)
+            && journal_contains_on_all_disks(&disk_paths, SCOPED_JOURNAL_REL, v1_compatible)
+            && journal_contains_on_all_disks(&disk_paths, JOURNAL_REL, v1_compatible)
+            && !journal_contains_on_any_disk(&disk_paths, JOURNAL_REL, scoped_only)
+    })
+    .await;
+    assert!(
+        flushed,
+        "runtime flush must persist rollback-safe mirrors without leaking scoped-only records into the legacy journal"
+    );
+    assert_eq!(
+        manager.operations_snapshot().await.queued_by_source.mrf,
+        0,
+        "zero-capacity manager keeps both intents in the MRF runtime so the persisted snapshot is observable"
+    );
+}
+
 #[test]
 fn mrf_journal_child_process_fixture() {
    let Ok(root) = std::env::var("RUSTFS_MRF_REPLAY_CHILD_ROOT") else {
@@ -524,13 +763,14 @@ fn mrf_successor_flush_child_process_fixture() {
        let expected_successor = journal_record(1, "successor-bucket", "second-object", None, 2);
        let flushed = wait_until(Duration::from_secs(10), || async {
            manager.operations_snapshot().await.queued_by_source.mrf == 1
+                && committed_checkpoint_matches_on_all_disks(&disk_paths, 2, &expected_successor)
                && journal_matches_on_all_disks(&disk_paths, SCOPED_JOURNAL_REL, &expected_successor)
                && journal_matches_on_all_disks(&disk_paths, JOURNAL_REL, &expected_successor)
        })
        .await;
        assert!(
            flushed,
-            "child process must publish the pending successor snapshot before the delete phase"
+            "child process must publish the committed pending successor before the delete phase"
        );
    });
    std::process::exit(78);
@@ -571,13 +811,14 @@ fn mrf_successor_flush_waiting_child_process_fixture() {
        let expected_successor = journal_record(1, "service-kill-bucket", "second-object", None, 2);
        let flushed = wait_until(Duration::from_secs(10), || async {
            manager.operations_snapshot().await.queued_by_source.mrf == 1
+                && committed_checkpoint_matches_on_all_disks(&disk_paths, 2, &expected_successor)
                && journal_matches_on_all_disks(&disk_paths, SCOPED_JOURNAL_REL, &expected_successor)
                && journal_matches_on_all_disks(&disk_paths, JOURNAL_REL, &expected_successor)
        })
        .await;
        assert!(
            flushed,
-            "child process must publish the pending successor snapshot before it can be killed"
+            "child process must publish the committed pending successor before it can be killed"
        );
        std::fs::write(&ready_path, b"ready").expect("write ready marker");
        loop {
@@ -698,10 +939,10 @@ async fn journal_replay_survives_successor_flush_before_delete() {
    );
    assert!(
        disk_paths.iter().all(|path| {
-            !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
-                && !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
+            Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
+                && Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
        }),
-        "a fully consumed successor snapshot may be deleted after restart replay"
+        "the accepted successor remains anchored until a verified repair proof"
    );
 }

@@ -751,10 +992,10 @@ async fn journal_replay_survives_service_kill_after_successor_flush() {
    );
    assert!(
        disk_paths.iter().all(|path| {
-            !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
-                && !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
+            Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
+                && Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
        }),
-        "a fully consumed successor snapshot may be deleted after service-kill restart replay"
+        "the accepted successor remains anchored until a verified repair proof after service-kill restart"
    );
 }

@@ -814,9 +1055,9 @@ async fn journal_replay_survives_sigkill_after_authoritative_successor_fsync_bef
    );
    assert!(
        disk_paths.iter().all(|path| {
-            !Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
-                && !Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
+            Path::new(path).join(META_BUCKET).join(JOURNAL_REL).exists()
+                && Path::new(path).join(META_BUCKET).join(SCOPED_JOURNAL_REL).exists()
        }),
-        "a fully consumed authoritative successor may clean both epochs after restart replay"
+        "the accepted authoritative successor remains anchored until a verified repair proof"
    );
 }
@@ -819,7 +819,9 @@ impl KmsBackend for AwsKmsBackend {
            .with_rotate(true)
            .with_enable_disable(true)
            .with_schedule_deletion(true)
-            .with_versioning(true)
+            // AWS KMS exposes rotation state but does not enumerate key
+            // versions through this backend's API contract.
+            .with_versioning(false)
            .with_physical_delete(false)
            .with_production_supported(true)
    }
@@ -74,6 +74,39 @@ impl ScriptedResponse {
    }
 }

+/// The `auth/token/lookup-self` answer every scripted Vault serves for free.
+///
+/// A Vault client now probes its token's remaining lifetime at login
+/// (backlog#2369 P3), which is credential plumbing rather than the protocol any
+/// of these tests is scripting. Answering it out of band keeps every existing
+/// script meaningful: `ttl` 0 is Vault's "this token does not expire", so the
+/// probe changes nothing about how a scripted test behaves.
+pub(crate) fn token_lookup_self_response() -> String {
+    serde_json::json!({
+        "data": {
+            "accessor": "scripted-accessor",
+            "creation_time": 1_700_000_000u64,
+            "creation_ttl": 0,
+            "display_name": "token",
+            "entity_id": "",
+            "explicit_max_ttl": 0,
+            "id": "scripted-token",
+            "num_uses": 0,
+            "orphan": true,
+            "path": "auth/token/create",
+            "policies": ["default"],
+            "renewable": false,
+            "ttl": 0
+        }
+    })
+    .to_string()
+}
+
+/// Whether a recorded request line addresses the token self-lookup.
+pub(crate) fn is_token_lookup_self(request_line: &str) -> bool {
+    request_line.contains("/v1/auth/token/lookup-self")
+}
+
 /// A scripted stand-in Vault listening on a loopback port.
 pub(crate) struct ScriptedVault {
    /// Base address (`http://127.0.0.1:port`) to point a Vault client at.
@@ -102,6 +135,19 @@ impl ScriptedVault {
                let Some((request_line, body, mut stream)) = read_request(stream).await else {
                    continue;
                };
+                if is_token_lookup_self(&request_line) {
+                    // Served out of band so the credential probe does not
+                    // consume a scripted response meant for the protocol under
+                    // test, and is not recorded as one of its requests.
+                    let body = token_lookup_self_response();
+                    let payload = format!(
+                        "HTTP/1.1 200 Scripted\r\ncontent-type: application/json\r\ncontent-length: {}\r\nconnection: close\r\n\r\n{body}",
+                        body.len(),
+                    );
+                    let _ = stream.write_all(payload.as_bytes()).await;
+                    let _ = stream.shutdown().await;
+                    continue;
+                }
                recorded
                    .lock()
                    .expect("scripted vault request log poisoned")
@@ -155,6 +201,19 @@ impl ScriptedVault {
                    let Some((request_line, body, stream)) = read_request(stream).await else {
                        return;
                    };
+                    if is_token_lookup_self(&request_line) {
+                        // Credential plumbing, not part of the KV2 protocol
+                        // this responder models; see token_lookup_self_response.
+                        write_response(
+                            stream,
+                            ScriptedResponse::Http {
+                                status: 200,
+                                body: token_lookup_self_response(),
+                            },
+                        )
+                        .await;
+                        return;
+                    }
                    recorded
                        .lock()
                        .expect("scripted vault request log poisoned")
@@ -13,5 +13,5 @@ expression: capabilities_snapshot(backend.capabilities())
  "rotate": true,
  "schedule_deletion": true,
  "update_key_metadata": false,
-  "versioning": true
+  "versioning": false
 }
@@ -114,7 +114,7 @@ impl fmt::Debug for SecretString {
 }

 /// Expiry attributes of a lease-bound token.
-#[derive(Debug, Clone, Copy)]
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub(crate) struct LeaseInfo {
    /// Time-to-live granted at issue or renewal.
    pub(crate) ttl: Duration,
@@ -207,31 +207,102 @@ pub(crate) trait TokenSource: fmt::Debug + Send + Sync {
 }

 /// Token source for [`VaultAuthMethod::Token`]: always yields the token fixed
-/// at configuration time. The token carries no lease, so it is never renewed
-/// and never expires from the provider's point of view.
+/// at configuration time.
+///
+/// The token itself is never re-issued, but it usually still expires:
+/// `vault token create` defaults to a 768-hour TTL. Hard-coding "no lease"
+/// here left the renewal task unstarted and published no remaining-TTL gauge,
+/// so a healthy-looking cluster turned every KMS call into a 403 a month later
+/// and could only be recovered by a restart or a reconfigure (backlog#2369 P3).
+/// The source therefore asks Vault what it is holding, once per client
+/// generation, and lets the existing renewal loop take over whenever the answer
+/// carries a TTL.
+/// Map a `lookup-self` answer onto a lease.
+///
+/// A zero TTL is Vault's answer for a token that never expires (root and
+/// periodic-root tokens), which keeps the pre-probe behaviour exactly: no
+/// lease, no renewal task, no expiry gate. A response that omits `renewable`
+/// is treated as not renewable, so the renewal loop falls back to re-reading
+/// the remaining TTL instead of assuming it can extend it.
+fn static_token_lease(ttl_secs: u64, renewable: Option<bool>) -> Option<LeaseInfo> {
+    (ttl_secs > 0).then_some(LeaseInfo {
+        ttl: Duration::from_secs(ttl_secs),
+        renewable: renewable.unwrap_or(false),
+    })
+}
+
 pub(crate) struct StaticToken {
    token: TokenLease,
+    /// Client authenticated with the configured token, used only for
+    /// `lookup-self`. Per-generation renewals use the generation's own client.
+    lookup_client: VaultClient,
 }

 impl StaticToken {
-    pub(crate) fn new(token: String) -> Self {
-        Self {
+    pub(crate) fn new(settings: &VaultConnectionSettings, token: String) -> Result<Self> {
+        let lookup_client = settings.build_client(&token)?;
+        Ok(Self {
            token: TokenLease::new(token, None),
-        }
+            lookup_client,
+        })
    }
 }

 #[async_trait]
 impl TokenSource for StaticToken {
    async fn acquire(&self) -> AttemptResult<TokenLease> {
-        Ok(self.token.clone())
+        // A lookup failure must not fail the login. The token itself may well
+        // be valid: a policy can omit `lookup-self`, and Vault may simply be
+        // unreachable for the moment. Failing here would take down deployments
+        // that work today, so the probe degrades to the pre-probe behaviour —
+        // no lease, no renewal — and says so loudly instead.
+        let lease = match vaultrs::token::lookup_self(&self.lookup_client).await {
+            Ok(lookup) => static_token_lease(lookup.ttl, lookup.renewable),
+            Err(error) => {
+                warn!(
+                    event = "vault_static_token_lookup_failed",
+                    error = %error,
+                    "Could not read the configured Vault token's remaining lifetime, so it will not be \
+                     renewed and its expiry will not be tracked. Grant the token `lookup-self` (Vault's \
+                     default policy does) or switch to AppRole, Kubernetes or an agent-managed token file"
+                );
+                None
+            }
+        };
+
+        if let Some(lease) = lease
+            && !lease.renewable
+        {
+            warn!(
+                event = "vault_static_token_not_renewable",
+                ttl_secs = lease.ttl.as_secs(),
+                "The configured Vault token expires and cannot be renewed; RustFS will fail closed as it \
+                 approaches expiry. Switch to AppRole, Kubernetes or an agent-managed token file, or \
+                 reconfigure with a fresh token before it lapses"
+            );
+        }
+
+        Ok(TokenLease::new(self.token.expose().to_string(), lease))
+    }
+
+    async fn renew(&self, client: &VaultClient) -> AttemptResult<TokenLease> {
+        // Vault refuses renew-self on a non-renewable token; the renewal loop
+        // then falls back to `acquire`, which re-reads the remaining TTL and
+        // keeps the gauge honest until the fail-closed window is reached.
+        let auth = vaultrs::token::renew_self(client, None)
+            .await
+            .map_err(|error| attempt_error("token renewal", error))?;
+        Ok(TokenLease::from_auth(auth))
    }
 }

 impl fmt::Debug for StaticToken {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        // TokenLease::fmt already redacts the token value.
-        f.debug_struct("StaticToken").field("token", &self.token).finish()
+        // TokenLease::fmt already redacts the token value; VaultClient embeds
+        // its settings, including the token, so it must stay out of Debug.
+        f.debug_struct("StaticToken")
+            .field("token", &self.token)
+            .finish_non_exhaustive()
    }
 }

@@ -541,7 +612,7 @@ pub(crate) fn token_source_for(
    settings: &VaultConnectionSettings,
 ) -> Result<Box<dyn TokenSource>> {
    match auth_method {
-        VaultAuthMethod::Token { token } => Ok(Box::new(StaticToken::new(token.clone()))),
+        VaultAuthMethod::Token { token } => Ok(Box::new(StaticToken::new(settings, token.clone())?)),
        VaultAuthMethod::AppRole {
            role_id,
            secret_id,
@@ -1201,14 +1272,22 @@ mod tests {
        (Arc::new(provider), state)
    }

+    /// A provider whose token reports no expiry, which is what `lookup-self`
+    /// answers for a root or periodic-root token. Scripted rather than backed
+    /// by [`StaticToken`] because the real source now asks Vault what it holds.
    async fn static_provider() -> VaultCredentialProvider {
        VaultCredentialProvider::new(
            test_settings(),
-            Box::new(StaticToken::new(TEST_TOKEN.to_string())),
+            Box::new(ScriptedSource {
+                state: Arc::new(ScriptedState::default()),
+                ttl: Duration::ZERO,
+                renewable: false,
+                login_delay: Duration::ZERO,
+            }),
            test_policy(Duration::from_secs(10), Duration::from_secs(5)),
        )
        .await
-        .expect("static provider must build without a live Vault")
+        .expect("a token without an expiry must build without a live Vault")
    }

    #[tokio::test]
@@ -1231,20 +1310,57 @@ mod tests {
        assert!(provider.spawn_renewal_task().is_none(), "a token without a lease has nothing to renew");
    }

-    #[tokio::test]
-    async fn test_static_token_source_yields_configured_token() {
-        let settings = test_settings();
-        let source = token_source_for(
+    #[test]
+    fn test_static_token_source_builds_without_contacting_vault() {
+        token_source_for(
            &VaultAuthMethod::Token {
                token: TEST_TOKEN.to_string(),
            },
-            &settings,
+            &test_settings(),
        )
        .expect("token auth must map to a source");
+    }

-        let lease = source.acquire().await.expect("static acquire cannot fail");
-        assert_eq!(lease.expose(), TEST_TOKEN);
-        assert!(lease.lease_info().is_none(), "static tokens must not carry a lease");
+    /// backlog#2369 P3: `vault token create` defaults to a 768-hour TTL, so
+    /// hard-coding "no lease" for token auth left the renewal task unstarted
+    /// and turned a healthy cluster into one that answers 403 a month later.
+    /// The lease now comes from what Vault reports.
+    #[test]
+    fn static_token_lease_follows_what_vault_reports() {
+        assert_eq!(
+            static_token_lease(0, Some(true)),
+            None,
+            "a token Vault reports as non-expiring must keep behaving as one"
+        );
+        assert_eq!(
+            static_token_lease(0, None),
+            None,
+            "a non-expiring token stays non-expiring whatever renewable says"
+        );
+        assert_eq!(
+            static_token_lease(2_764_800, Some(true)),
+            Some(LeaseInfo {
+                ttl: Duration::from_secs(2_764_800),
+                renewable: true,
+            }),
+            "the default 768-hour token must be tracked and renewed"
+        );
+        assert_eq!(
+            static_token_lease(3_600, Some(false)),
+            Some(LeaseInfo {
+                ttl: Duration::from_secs(3_600),
+                renewable: false,
+            }),
+            "an expiring token that cannot be renewed still needs its expiry tracked"
+        );
+        assert_eq!(
+            static_token_lease(3_600, None),
+            Some(LeaseInfo {
+                ttl: Duration::from_secs(3_600),
+                renewable: false,
+            }),
+            "an omitted renewable flag must not be read as renewable"
+        );
    }

    #[tokio::test]
@@ -1759,7 +1875,7 @@ mod tests {
                renewable: true,
            }),
        );
-        let static_source = StaticToken::new(TEST_TOKEN.to_string());
+        let static_source = StaticToken::new(&test_settings(), TEST_TOKEN.to_string()).expect("static source");
        let approle_source = AppRoleLogin::new(
            &test_settings(),
            "approle".to_string(),
@@ -263,6 +263,14 @@ pub struct ScannerUsageRecoveryIntentResponse {
    pub mode: String,
    pub intent_id: String,
    pub state: String,
+    #[serde(default)]
+    pub actor_sha256: Option<String>,
+    #[serde(default)]
+    pub idempotency_key_sha256: Option<String>,
+    #[serde(default)]
+    pub request_sha256: Option<String>,
+    #[serde(default)]
+    pub accepted_at_unix_secs: Option<u64>,
    #[serde(flatten)]
    pub extra: serde_json::Map<String, serde_json::Value>,
 }
@@ -869,6 +877,10 @@ mod tests {
            "mode": "full-rebuild",
            "intent_id": "0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef",
            "state": "accepted",
+            "actor_sha256": "1111111111111111111111111111111111111111111111111111111111111111",
+            "idempotency_key_sha256": "2222222222222222222222222222222222222222222222222222222222222222",
+            "request_sha256": "3333333333333333333333333333333333333333333333333333333333333333",
+            "accepted_at_unix_secs": 7,
            "future": {"worker": "pending"}
        }))
        .unwrap();
@@ -877,7 +889,33 @@ mod tests {
        assert_eq!(intent.mode, "full-rebuild");
        assert_eq!(intent.intent_id, "0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef");
        assert_eq!(intent.state, "accepted");
+        assert_eq!(
+            intent.actor_sha256.as_deref(),
+            Some("1111111111111111111111111111111111111111111111111111111111111111")
+        );
+        assert_eq!(
+            intent.idempotency_key_sha256.as_deref(),
+            Some("2222222222222222222222222222222222222222222222222222222222222222")
+        );
+        assert_eq!(
+            intent.request_sha256.as_deref(),
+            Some("3333333333333333333333333333333333333333333333333333333333333333")
+        );
+        assert_eq!(intent.accepted_at_unix_secs, Some(7));
        assert_eq!(intent.extra["future"]["worker"], "pending");
+
+        let legacy_intent: ScannerUsageRecoveryIntentResponse = serde_json::from_value(json!({
+            "status": "accepted",
+            "action": "usage-full-rebuild",
+            "mode": "full-rebuild",
+            "intent_id": "0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef",
+            "state": "accepted"
+        }))
+        .unwrap();
+        assert!(legacy_intent.actor_sha256.is_none());
+        assert!(legacy_intent.idempotency_key_sha256.is_none());
+        assert!(legacy_intent.request_sha256.is_none());
+        assert!(legacy_intent.accepted_at_unix_secs.is_none());
    }

    #[test]
@@ -971,7 +1009,7 @@ mod tests {
    #[tokio::test]
    async fn scanner_usage_async_reset_posts_explicit_intent_contract() {
        let server = TestServer::spawn(
-            r#"{"status":"accepted","action":"usage-full-rebuild","mode":"full-rebuild","intent_id":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa","state":"accepted"}"#,
+            r#"{"status":"accepted","action":"usage-full-rebuild","mode":"full-rebuild","intent_id":"aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa","state":"accepted","actor_sha256":"1111111111111111111111111111111111111111111111111111111111111111","idempotency_key_sha256":"2222222222222222222222222222222222222222222222222222222222222222","request_sha256":"3333333333333333333333333333333333333333333333333333333333333333","accepted_at_unix_secs":7}"#,
            202,
        )
        .await;
@@ -985,6 +1023,7 @@ mod tests {
        assert_eq!(accepted.status, "accepted");
        assert_eq!(accepted.mode, "full-rebuild");
        assert_eq!(accepted.state, "accepted");
+        assert_eq!(accepted.accepted_at_unix_secs, Some(7));
        let request = server.recorded();
        assert_eq!(request.method, "POST");
        assert_eq!(request.path, "/rustfs/admin/v3/scanner/usage-state/reset");
@@ -90,6 +90,7 @@ s3s = { workspace = true, features = ["minio"] }
 hex-simd.workspace = true

 [dev-dependencies]
+temp-env = { workspace = true, features = ["async_closure"] }
 tokio = { workspace = true, features = ["test-util"] }
 tokio-test = { workspace = true }
 criterion = { workspace = true, features = ["html_reports"] }
@@ -327,6 +327,55 @@ where
    }
 }

+/// Read-side switch for the pre-`1.0.0-alpha.91` nonce layout, in which a whole
+/// v1 segment reused the part nonce for every block.
+///
+/// On by default, because turning it off refuses to decrypt objects written
+/// before that release. Block zero's derived nonce equals that base nonce, so
+/// the layout also lets a frame encrypted at index zero authenticate anywhere
+/// in its segment; the in-segment layout lock catches that as soon as a later
+/// frame disagrees, but a stream that is nothing but repeats of frame zero has
+/// no such later frame. A deployment with no pre-alpha.91 objects should set
+/// this to `false` to remove that surface outright (backlog#2369 P2).
+///
+// RUSTFS_COMPAT_TODO(backlog-2369-legacy-nonce-fallback): Remove after the
+// minimum supported direct-upgrade release and after migration tooling has
+// rewritten every pre-alpha.91 encrypted object.
+pub const ENV_RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK: &str = "RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK";
+const DEFAULT_RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK: bool = true;
+
+fn legacy_nonce_fallback_enabled() -> bool {
+    #[cfg(test)]
+    {
+        rustfs_utils::get_env_bool(
+            ENV_RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK,
+            DEFAULT_RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK,
+        )
+    }
+    #[cfg(not(test))]
+    {
+        static CACHED: std::sync::OnceLock<bool> = std::sync::OnceLock::new();
+        *CACHED.get_or_init(|| {
+            rustfs_utils::get_env_bool(
+                ENV_RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK,
+                DEFAULT_RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK,
+            )
+        })
+    }
+}
+
+/// The nonce layout selected while decoding a legacy v1 segment.
+///
+/// A historical writer used one of these layouts consistently for every
+/// block in a segment. Once a non-zero block identifies that layout, accepting
+/// another layout would let an attacker replay a block encrypted at index zero.
+#[derive(Clone, Copy, PartialEq, Eq)]
+enum V1NonceLayout {
+    Current,
+    LegacyBlock,
+    ReusedPart,
+}
+
 pin_project! {
    /// A reader wrapper that decrypts data on the fly using AES-256-GCM.
    /// This is a demonstration. For production, use a secure and audited crypto library.
@@ -358,6 +407,8 @@ pin_project! {
        segment_frames: usize,
        stream_saw_v2: bool,
        segments_completed: usize,
+        v1_nonce_layout: Option<V1NonceLayout>,
+        legacy_nonce_fallback: bool,
    }
 }

@@ -391,6 +442,8 @@ where
            segment_frames: 0,
            stream_saw_v2: false,
            segments_completed: 0,
+            v1_nonce_layout: None,
+            legacy_nonce_fallback: legacy_nonce_fallback_enabled(),
        }
    }

@@ -441,6 +494,8 @@ where
            segment_frames: 0,
            stream_saw_v2: false,
            segments_completed: 0,
+            v1_nonce_layout: None,
+            legacy_nonce_fallback: legacy_nonce_fallback_enabled(),
        }
    }
 }
@@ -547,6 +602,7 @@ where
                    *this.segment_frame_version = None;
                    *this.saw_final_frame = false;
                    *this.segment_frames = 0;
+                    *this.v1_nonce_layout = None;

                    if *this.multipart_mode {
                        let next_part = if *this.current_part_index + 1 < this.multipart_parts.len() {
@@ -696,26 +752,46 @@ where
                    *this.base_nonce
                };
                let legacy_block_nonce = derive_block_nonce(&legacy_part_nonce, *this.block_index);
-                match this.cipher.decrypt(&nonce, ciphertext) {
-                    Ok(plaintext) => plaintext,
-                    Err(primary_err) => {
-                        let legacy_nonce =
-                            Nonce::try_from(legacy_block_nonce.as_slice()).map_err(|_| Error::other("invalid nonce length"))?;
-
-                        match this.cipher.decrypt(&legacy_nonce, ciphertext) {
-                            Ok(plaintext) => plaintext,
-                            Err(_) => {
-                                // Accept previously written streams that reused the part nonce
-                                // for every block inside a segment.
-                                let legacy_part_nonce = Nonce::try_from(legacy_part_nonce.as_slice())
-                                    .map_err(|_| Error::other("invalid nonce length"))?;
-                                this.cipher
-                                    .decrypt(&legacy_part_nonce, ciphertext)
-                                    .map_err(|_| Error::other(format!("decrypt error: {primary_err}")))?
-                            }
+                let legacy_part_nonce =
+                    Nonce::try_from(legacy_part_nonce.as_slice()).map_err(|_| Error::other("invalid nonce length"))?;
+                let legacy_block_nonce =
+                    Nonce::try_from(legacy_block_nonce.as_slice()).map_err(|_| Error::other("invalid nonce length"))?;
+                let layouts = [
+                    (V1NonceLayout::Current, &nonce),
+                    (V1NonceLayout::LegacyBlock, &legacy_block_nonce),
+                    (V1NonceLayout::ReusedPart, &legacy_part_nonce),
+                ];
+                let selected = if *this.block_index == 0 { None } else { *this.v1_nonce_layout };
+                let mut plaintext = None;
+                let mut last_error = None;
+                for (layout, candidate_nonce) in layouts {
+                    if selected.is_some_and(|expected| expected != layout) {
+                        continue;
+                    }
+                    if layout == V1NonceLayout::ReusedPart && !*this.legacy_nonce_fallback {
+                        continue;
+                    }
+                    match this.cipher.decrypt(candidate_nonce, ciphertext) {
+                        Ok(value) => {
+                            plaintext = Some((value, layout));
+                            break;
                        }
+                        Err(error) => last_error = Some(error),
                    }
                }
+                let (plaintext, layout) = plaintext.ok_or_else(|| {
+                    Error::new(
+                        std::io::ErrorKind::InvalidData,
+                        format!(
+                            "decrypt error: {}",
+                            last_error.map_or_else(|| "nonce layout rejected".to_string(), |error| error.to_string())
+                        ),
+                    )
+                })?;
+                if *this.block_index > 0 && this.v1_nonce_layout.is_none() {
+                    *this.v1_nonce_layout = Some(layout);
+                }
+                plaintext
            };
            if *this.current_frame_type == FRAME_TYPE_V2_FINAL {
                *this.saw_final_frame = true;
@@ -1003,6 +1079,154 @@ mod tests {
        assert_eq!(&decrypted, data);
    }

+    /// Encrypts `block_count` full v1 blocks, then overwrites frame one with a
+    /// verbatim copy of frame zero. Every frame is the same length, so the
+    /// stream keeps its original size and the forgery is invisible to any
+    /// length check.
+    async fn v1_stream_with_frame_zero_replayed_at_index_one(key: [u8; 32], nonce: [u8; 12], block_count: usize) -> Vec<u8> {
+        assert!(block_count >= 2, "a replay needs at least two frames");
+        let mut data = Vec::with_capacity(ENCRYPTION_BLOCK_SIZE * block_count);
+        for index in 0..block_count {
+            data.extend(std::iter::repeat_n(0xA1u8.wrapping_add(index as u8 * 17), ENCRYPTION_BLOCK_SIZE));
+        }
+
+        let mut encrypt_reader = EncryptReader::new(Cursor::new(data), key, nonce);
+        let mut encrypted = Vec::new();
+        encrypt_reader.read_to_end(&mut encrypted).await.expect("encrypt v1 frames");
+
+        // Header layout: [type][len:24][crc:32]; `len` counts the payload plus
+        // its own 4-byte CRC field, so the frame occupies 8 + (len - 4) bytes.
+        let declared_len = (encrypted[1] as usize) | ((encrypted[2] as usize) << 8) | ((encrypted[3] as usize) << 16);
+        let frame_len = 8 + declared_len - 4;
+        let replayed_first = encrypted[..frame_len].to_vec();
+        encrypted[frame_len..frame_len * 2].copy_from_slice(&replayed_first);
+        encrypted
+    }
+
+    #[tokio::test]
+    async fn decrypt_reader_rejects_a_replayed_first_v1_frame() {
+        let key = [0x11; 32];
+        let nonce = [0x22; 12];
+        let encrypted = v1_stream_with_frame_zero_replayed_at_index_one(key, nonce, 3).await;
+
+        let mut decrypt_reader = DecryptReader::new(Cursor::new(encrypted), key, nonce);
+        let error = decrypt_reader
+            .read_to_end(&mut Vec::new())
+            .await
+            .expect_err("a repeated index-zero frame must not authenticate at index one");
+        assert_eq!(error.kind(), std::io::ErrorKind::InvalidData);
+    }
+
+    /// The in-segment layout lock must not cost compatibility: every legacy v1
+    /// shape the fallback chain exists for still decrypts under the default.
+    #[tokio::test]
+    async fn legacy_v1_streams_still_decrypt_under_the_default_fallback() {
+        temp_env::async_with_vars([(ENV_RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK, None::<&str>)], async {
+            assert!(legacy_nonce_fallback_enabled(), "the legacy nonce fallback must stay on by default");
+
+            let mut key = [0u8; 32];
+            let mut nonce = [0u8; 12];
+            rand::rng().fill_bytes(&mut key);
+            rand::rng().fill_bytes(&mut nonce);
+            let mut data = vec![0u8; ENCRYPTION_BLOCK_SIZE * 3 + 17];
+            rand::rng().fill(&mut data[..]);
+
+            // Modern single-part stream.
+            let mut encrypted = Vec::new();
+            EncryptReader::new(Cursor::new(data.clone()), key, nonce)
+                .read_to_end(&mut encrypted)
+                .await
+                .expect("modern v1 stream should encrypt");
+            let mut decrypted = Vec::new();
+            DecryptReader::new(Cursor::new(encrypted), key, nonce)
+                .read_to_end(&mut decrypted)
+                .await
+                .expect("modern v1 stream should decrypt");
+            assert_eq!(decrypted, data);
+
+            // Pre-alpha.91 stream that reused the part nonce for every block.
+            let legacy = encrypt_with_legacy_nonce_reuse(&data, key, nonce);
+            let mut decrypted = Vec::new();
+            DecryptReader::new(Cursor::new(legacy), key, nonce)
+                .read_to_end(&mut decrypted)
+                .await
+                .expect("a reused-nonce legacy stream should still decrypt");
+            assert_eq!(decrypted, data);
+        })
+        .await;
+    }
+
+    /// The residual after the layout lock: a stream that is nothing but repeats
+    /// of frame zero has no later frame to disagree with the reused-part
+    /// layout, so only turning the fallback off rejects it.
+    #[tokio::test]
+    async fn a_two_frame_replay_is_closed_only_by_disabling_the_legacy_fallback() {
+        let key = [0x33; 32];
+        let nonce = [0x44; 12];
+        let encrypted = v1_stream_with_frame_zero_replayed_at_index_one(key, nonce, 2).await;
+
+        temp_env::async_with_vars([(ENV_RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK, None::<&str>)], async {
+            let mut forged = Vec::new();
+            DecryptReader::new(Cursor::new(encrypted.clone()), key, nonce)
+                .read_to_end(&mut forged)
+                .await
+                .expect("with the fallback on this forgery is still accepted");
+            assert_eq!(forged.len(), ENCRYPTION_BLOCK_SIZE * 2);
+            assert_eq!(
+                &forged[..ENCRYPTION_BLOCK_SIZE],
+                &forged[ENCRYPTION_BLOCK_SIZE..],
+                "the accepted forgery is frame zero's plaintext twice over"
+            );
+        })
+        .await;
+
+        temp_env::async_with_vars([(ENV_RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK, Some("false"))], async {
+            let error = DecryptReader::new(Cursor::new(encrypted.clone()), key, nonce)
+                .read_to_end(&mut Vec::new())
+                .await
+                .expect_err("with the fallback off the replayed frame must not authenticate");
+            assert_eq!(error.kind(), std::io::ErrorKind::InvalidData);
+        })
+        .await;
+    }
+
+    /// Turning the fallback off removes exactly the third layout: modern v1
+    /// streams keep decrypting, pre-alpha.91 reused-nonce streams stop.
+    #[tokio::test]
+    async fn disabling_the_legacy_nonce_fallback_refuses_only_reused_part_nonces() {
+        let mut key = [0u8; 32];
+        let mut nonce = [0u8; 12];
+        rand::rng().fill_bytes(&mut key);
+        rand::rng().fill_bytes(&mut nonce);
+        let mut data = vec![0u8; ENCRYPTION_BLOCK_SIZE * 3 + 17];
+        rand::rng().fill(&mut data[..]);
+
+        let mut modern = Vec::new();
+        EncryptReader::new(Cursor::new(data.clone()), key, nonce)
+            .read_to_end(&mut modern)
+            .await
+            .expect("modern v1 stream should encrypt");
+        let legacy = encrypt_with_legacy_nonce_reuse(&data, key, nonce);
+
+        temp_env::async_with_vars([(ENV_RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK, Some("false"))], async {
+            assert!(!legacy_nonce_fallback_enabled(), "the switch must be observed");
+
+            let mut decrypted = Vec::new();
+            DecryptReader::new(Cursor::new(modern), key, nonce)
+                .read_to_end(&mut decrypted)
+                .await
+                .expect("modern v1 streams must keep decrypting with the fallback off");
+            assert_eq!(decrypted, data);
+
+            let error = DecryptReader::new(Cursor::new(legacy), key, nonce)
+                .read_to_end(&mut Vec::new())
+                .await
+                .expect_err("the third layout must be gone when the fallback is off");
+            assert_eq!(error.kind(), std::io::ErrorKind::InvalidData);
+        })
+        .await;
+    }
+
    #[tokio::test]
    async fn test_decrypt_reader_only() {
        // Encrypt some data first
@@ -0,0 +1,432 @@
+// Copyright 2024 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+// http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use bytes::Bytes;
+use datafusion::object_store::{Error, Result};
+use futures::{Stream, StreamExt, stream::BoxStream};
+use transform_stream::AsyncTryStream;
+
+use crate::SelectError;
+
+/// Arrow accepts byte-sized CSV controls. Unicode quotes need streaming normalization.
+pub fn csv_input_requires_normalization(quote: Option<&str>, escape: Option<&str>) -> bool {
+    quote.is_some_and(|quote| quote.len() > 1) || escape.is_some_and(|escape| escape.len() > 1)
+}
+
+/// CSV syntax independent of request headers, serialization formats, or S3 DTOs.
+#[derive(Default)]
+pub(crate) struct CsvSyntax<'a> {
+    pub quote: Option<&'a str>,
+    pub escape: Option<&'a str>,
+    pub field: Option<&'a str>,
+    pub record: Option<&'a str>,
+    pub comment: Option<u8>,
+}
+
+#[derive(Clone, Copy, PartialEq, Eq)]
+enum State {
+    FieldStart,
+    Unquoted,
+    Quoted,
+    AfterQuote,
+    Escaped,
+    Comment,
+}
+
+/// Emits ordinary CSV with every field quoted. This avoids reserving a sentinel
+/// byte that might also appear in a UTF-8 field. Only a partial control token is
+/// retained between chunks; neither records nor objects are buffered.
+struct CsvInputNormalizer {
+    quote: Vec<u8>,
+    escape: Vec<u8>,
+    field: Vec<u8>,
+    record: Vec<u8>,
+    comment: Option<u8>,
+    default_records: bool,
+    state: State,
+    record_start: bool,
+    carry: Vec<u8>,
+    token_size: usize,
+}
+
+impl CsvInputNormalizer {
+    fn new(csv: &CsvSyntax<'_>) -> Self {
+        let quote = csv
+            .quote
+            .filter(|value| !value.is_empty())
+            .unwrap_or("\"")
+            .as_bytes()
+            .to_vec();
+        let escape = csv
+            .escape
+            .filter(|value| !value.is_empty())
+            .unwrap_or("\"")
+            .as_bytes()
+            .to_vec();
+        let field = csv.field.filter(|value| !value.is_empty()).unwrap_or(",").as_bytes().to_vec();
+        let record = csv
+            .record
+            .filter(|value| !value.is_empty())
+            .unwrap_or("\n")
+            .as_bytes()
+            .to_vec();
+        let token_size = quote.len().max(escape.len()).max(field.len()).max(record.len()).max(2);
+        Self {
+            quote,
+            escape,
+            field,
+            record,
+            comment: csv.comment,
+            default_records: csv.record.is_none(),
+            state: State::FieldStart,
+            record_start: true,
+            carry: Vec::new(),
+            token_size,
+        }
+    }
+
+    fn record_len(&self, bytes: &[u8]) -> usize {
+        if self.default_records && bytes.starts_with(b"\r\n") {
+            2
+        } else if self.default_records && bytes.starts_with(b"\r") {
+            1
+        } else if bytes.starts_with(&self.record) {
+            self.record.len()
+        } else {
+            0
+        }
+    }
+
+    fn push_value(output: &mut Vec<u8>, bytes: &[u8]) {
+        for byte in bytes {
+            if *byte == b'"' {
+                output.push(b'"');
+            }
+            output.push(*byte);
+        }
+    }
+
+    fn convert(&mut self, chunk: &[u8], last: bool) -> std::result::Result<Vec<u8>, SelectError> {
+        let mut bytes = std::mem::take(&mut self.carry);
+        bytes.extend_from_slice(chunk);
+        let end = if last {
+            bytes.len()
+        } else {
+            bytes.len().saturating_sub(self.token_size - 1)
+        };
+        let mut output = Vec::with_capacity(bytes.len());
+        let mut pos = 0;
+        while pos < end {
+            let rest = &bytes[pos..];
+            let record_len = self.record_len(rest);
+            let field = rest.starts_with(&self.field) && self.field.len() > record_len;
+            match self.state {
+                State::Comment => {
+                    if record_len > 0 {
+                        self.state = State::FieldStart;
+                        pos += record_len;
+                    } else {
+                        pos += 1;
+                    }
+                }
+                State::Escaped => {
+                    if record_len > 0 {
+                        return Err(SelectError::CsvParsingError);
+                    }
+                    Self::push_value(&mut output, &rest[..1]);
+                    self.state = State::Quoted;
+                    pos += 1;
+                }
+                State::Quoted if rest.starts_with(&self.quote) => {
+                    self.state = State::AfterQuote;
+                    pos += self.quote.len();
+                }
+                State::Quoted if rest.starts_with(&self.escape) => {
+                    self.state = State::Escaped;
+                    pos += self.escape.len();
+                }
+                State::Quoted => {
+                    if record_len > 0 {
+                        return Err(SelectError::CsvParsingError);
+                    }
+                    Self::push_value(&mut output, &rest[..1]);
+                    pos += 1;
+                }
+                State::AfterQuote if rest.starts_with(&self.quote) => {
+                    Self::push_value(&mut output, &self.quote);
+                    self.state = State::Quoted;
+                    pos += self.quote.len();
+                }
+                State::FieldStart if self.record_start && self.comment == Some(rest[0]) => {
+                    self.state = State::Comment;
+                    pos += 1;
+                }
+                State::FieldStart if rest.starts_with(&self.quote) => {
+                    output.push(b'"');
+                    self.state = State::Quoted;
+                    self.record_start = false;
+                    pos += self.quote.len();
+                }
+                _ if field || record_len > 0 => {
+                    if self.state == State::FieldStart {
+                        if field || !self.record_start {
+                            output.extend_from_slice(b"\"\"");
+                        }
+                    } else {
+                        output.push(b'"');
+                    }
+                    output.push(if field { b',' } else { b'\n' });
+                    self.state = State::FieldStart;
+                    self.record_start = !field;
+                    pos += if field { self.field.len() } else { record_len };
+                }
+                _ => {
+                    if self.state == State::FieldStart {
+                        output.push(b'"');
+                    }
+                    self.state = State::Unquoted;
+                    self.record_start = false;
+                    Self::push_value(&mut output, &rest[..1]);
+                    pos += 1;
+                }
+            }
+        }
+        self.carry.extend_from_slice(&bytes[pos..]);
+        if last {
+            match self.state {
+                State::Quoted | State::Escaped => return Err(SelectError::CsvParsingError),
+                State::Unquoted | State::AfterQuote => output.push(b'"'),
+                State::FieldStart if !self.record_start => output.extend_from_slice(b"\"\""),
+                State::FieldStart | State::Comment => {}
+            }
+        }
+        Ok(output)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn normalize_chunks(csv: &CsvSyntax<'_>, input: &[u8], chunk_size: usize) -> Vec<u8> {
+        let mut normalizer = CsvInputNormalizer::new(csv);
+        let mut output = Vec::new();
+        for chunk in input.chunks(chunk_size) {
+            output.extend(normalizer.convert(chunk, false).expect("normalize complete CSV input"));
+            assert!(normalizer.carry.len() < normalizer.token_size, "only a partial token may be retained");
+        }
+        output.extend(normalizer.convert(&[], true).expect("finish complete CSV input"));
+        output
+    }
+
+    #[test]
+    fn unicode_csv_quotes_preserve_values_at_every_chunk_boundary() {
+        let cases = [
+            ("ع", "\"", "عcol1ع,عcol2ع,عcol3ع\n", "\"col1\",\"col2\",\"col3\"\n"),
+            ("ع", "\"", "\"left\",tail\n", "\"\"\"left\"\"\",\"tail\"\n"),
+            ("ع", "\"", "عA,Bع,plain\n", "\"A,B\",\"plain\"\n"),
+            ("ع", "\"", "عAععBع,tail\n", "\"AعB\",\"tail\"\n"),
+            ("ع", "\"", "عA\"عBع,tail\n", "\"AعB\",\"tail\"\n"),
+            ("ع", "\\", "عA\\\"Bع,tail\n", "\"A\"\"B\",\"tail\"\n"),
+            ("\"", "界", "\"A界\"B\",\"C\"\n", "\"A\"\"B\",\"C\"\n"),
+            ("🦀", "🦀", "🦀A🦀🦀B🦀,C\n", "\"A🦀B\",\"C\"\n"),
+            ("ع", "\"", "a\0b,عc\0dع\n", "\"a\0b\",\"c\0d\"\n"),
+            ("ع", "\"", "AعB,tail\n", "\"AعB\",\"tail\"\n"),
+            ("ع", "\"", "عaعsuffix,tail\n", "\"asuffix\",\"tail\"\n"),
+            ("ع", "\"", ",\n", "\"\",\"\"\n"),
+            ("ع", "\"", "a,", "\"a\",\"\""),
+            ("ع", "\"", "عع", "\"\""),
+            ("ع", "\"", "\n", "\n"),
+            ("ع", "\"", "", ""),
+        ];
+        for (quote, escape, input, expected) in cases {
+            let csv = CsvSyntax {
+                quote: Some(quote),
+                escape: Some(escape),
+                record: Some("\n"),
+                ..Default::default()
+            };
+            for chunk_size in 1..=input.len().max(1) {
+                assert_eq!(
+                    normalize_chunks(&csv, input.as_bytes(), chunk_size),
+                    expected.as_bytes(),
+                    "input={input:?}, chunk_size={chunk_size}"
+                );
+            }
+        }
+    }
+
+    #[test]
+    fn unicode_csv_quotes_keep_custom_delimiters_and_comments_out_of_values() {
+        let csv = CsvSyntax {
+            quote: Some("ع"),
+            escape: Some("\\"),
+            field: Some("界"),
+            record: Some("^Y"),
+            comment: Some(b'#'),
+        };
+        let input = "#skipع界^Yعa界bع界\"literal\"^Yعline\nbreakع界end^Y";
+        let expected = "\"a界b\",\"\"\"literal\"\"\"\n\"line\nbreak\",\"end\"\n";
+        for chunk_size in 1..=input.len() {
+            assert_eq!(normalize_chunks(&csv, input.as_bytes(), chunk_size), expected.as_bytes());
+        }
+    }
+
+    #[test]
+    fn unicode_csv_quotes_reject_unterminated_fields_and_quoted_record_delimiters() {
+        for input in ["عunfinished", "عescape\\", "عline\nbreakع\n", "عline\\\nbreakع\n"] {
+            let csv = CsvSyntax {
+                quote: Some("ع"),
+                escape: Some("\\"),
+                ..Default::default()
+            };
+            let mut normalizer = CsvInputNormalizer::new(&csv);
+            assert_eq!(normalizer.convert(input.as_bytes(), true), Err(SelectError::CsvParsingError));
+        }
+    }
+
+    #[test]
+    fn unicode_csv_quotes_preserve_omitted_syntax_defaults() {
+        assert!(!csv_input_requires_normalization(None, None));
+        assert!(!csv_input_requires_normalization(Some("\""), Some("\\")));
+        assert!(csv_input_requires_normalization(Some("ع"), None));
+        assert!(csv_input_requires_normalization(None, Some("界")));
+
+        let quote_only = CsvSyntax {
+            quote: Some("ع"),
+            ..Default::default()
+        };
+        assert_eq!(
+            normalize_chunks(&quote_only, "عA\"عBع,tail\r\n".as_bytes(), 1),
+            "\"AعB\",\"tail\"\n".as_bytes()
+        );
+        let escape_only = CsvSyntax {
+            escape: Some("界"),
+            ..Default::default()
+        };
+        assert_eq!(
+            normalize_chunks(&escape_only, "\"A界\"B\",tail\r\n".as_bytes(), 1),
+            b"\"A\"\"B\",\"tail\"\n"
+        );
+    }
+
+    #[test]
+    fn unicode_csv_quotes_stream_large_fields_without_retaining_records() {
+        let csv = CsvSyntax {
+            quote: Some("ع"),
+            ..Default::default()
+        };
+        let mut normalizer = CsvInputNormalizer::new(&csv);
+        let chunk = vec![b'x'; 64 * 1024];
+        let mut output_len = normalizer.convert("ع".as_bytes(), false).expect("opening quote").len();
+        for _ in 0..64 {
+            let output = normalizer.convert(&chunk, false).expect("stream field chunk");
+            assert!(output.len() >= chunk.len() - 3, "field data must be emitted before its closing quote");
+            assert!(normalizer.carry.len() < 4);
+            output_len += output.len();
+        }
+        output_len += normalizer.convert("ع\n".as_bytes(), true).expect("close field").len();
+        assert_eq!(output_len, chunk.len() * 64 + 3);
+    }
+}
+
+pub(crate) fn normalize_csv_stream<S>(stream: S, csv: &CsvSyntax<'_>) -> BoxStream<'static, Result<Bytes>>
+where
+    S: Stream<Item = Result<Bytes>> + Send + 'static,
+{
+    let mut normalizer = CsvInputNormalizer::new(csv);
+    AsyncTryStream::<Bytes, Error, _>::new(|mut y| async move {
+        futures::pin_mut!(stream);
+        while let Some(chunk) = stream.next().await {
+            let converted = normalizer.convert(&chunk?, false).map_err(|source| Error::Generic {
+                store: "EcObjectStore",
+                source: Box::new(source),
+            })?;
+            if !converted.is_empty() {
+                y.yield_ok(Bytes::from(converted)).await;
+            }
+        }
+        let converted = normalizer.convert(&[], true).map_err(|source| Error::Generic {
+            store: "EcObjectStore",
+            source: Box::new(source),
+        })?;
+        if !converted.is_empty() {
+            y.yield_ok(Bytes::from(converted)).await;
+        }
+        Ok(())
+    })
+    .boxed()
+}
+
+#[cfg(test)]
+mod stream_tests {
+    use super::*;
+    use std::sync::{
+        Arc,
+        atomic::{AtomicBool, AtomicUsize, Ordering},
+    };
+
+    struct DropProbe(Arc<AtomicBool>);
+
+    impl Drop for DropProbe {
+        fn drop(&mut self) {
+            self.0.store(true, Ordering::SeqCst);
+        }
+    }
+
+    #[tokio::test]
+    async fn unicode_csv_quotes_drop_the_source_without_reading_ahead() {
+        let polls = Arc::new(AtomicUsize::new(0));
+        let dropped = Arc::new(AtomicBool::new(false));
+        let source =
+            futures::stream::unfold((DropProbe(Arc::clone(&dropped)), Arc::clone(&polls)), |(guard, polls)| async move {
+                polls.fetch_add(1, Ordering::SeqCst);
+                Some((Ok(Bytes::from_static("عvalueع\n".as_bytes())), (guard, polls)))
+            });
+        let csv = CsvSyntax {
+            quote: Some("ع"),
+            ..Default::default()
+        };
+        let mut stream = normalize_csv_stream(source, &csv);
+        assert!(!stream.next().await.expect("first output").expect("valid CSV").is_empty());
+        assert_eq!(polls.load(Ordering::SeqCst), 1);
+        drop(stream);
+        assert!(dropped.load(Ordering::SeqCst), "cancellation must release the source reader");
+        assert_eq!(polls.load(Ordering::SeqCst), 1);
+    }
+
+    #[tokio::test]
+    async fn unicode_csv_quotes_preserve_source_errors_after_partial_output() {
+        let source = futures::stream::iter([
+            Ok(Bytes::from_static("عvalueع\n".as_bytes())),
+            Err(Error::Generic {
+                store: "fixture",
+                source: std::io::Error::other("source read failed").into(),
+            }),
+        ]);
+        let csv = CsvSyntax {
+            quote: Some("ع"),
+            ..Default::default()
+        };
+        let mut stream = normalize_csv_stream(source, &csv);
+        assert!(!stream.next().await.expect("partial output").expect("valid prefix").is_empty());
+        let error = stream
+            .next()
+            .await
+            .expect("source failure must remain visible")
+            .expect_err("must not return a successful tail");
+        assert!(error.to_string().contains("source read failed"));
+        assert!(stream.next().await.is_none());
+    }
+}
@@ -23,12 +23,14 @@ use datafusion::{
 use std::{error::Error as StdError, fmt::Display};
 use thiserror::Error;

+mod csv_input;
 mod input_stream;
 mod metrics;
 pub mod object_store;
 pub mod query;
 pub mod server;
 mod storage_api;
+pub use csv_input::csv_input_requires_normalization;
 pub use metrics::{SelectInputMetrics, SelectInputMetricsSnapshot};
 pub use storage_api::SelectObjectSnapshot;

@@ -64,6 +64,7 @@ use tokio::{io::AsyncReadExt, sync::OnceCell};
 use tokio_util::io::ReaderStream;
 use transform_stream::AsyncTryStream;

+use crate::csv_input::{CsvSyntax, csv_input_requires_normalization, normalize_csv_stream};
 use crate::storage_api::object_store::HTTPRangeSpec;

 mod json_document;
@@ -345,6 +346,29 @@ impl EcObjectStore {
        (self.need_convert || (delimiter.len() == 2 && delimiter != NORMALIZED_RECORD_DELIMITER)).then_some(delimiter)
    }

+    fn convert_csv_stream<S>(&self, stream: S) -> BoxStream<'static, Result<Bytes>>
+    where
+        S: Stream<Item = Result<Bytes>> + Send + 'static,
+    {
+        if let Some(csv) = self.input.request.input_serialization.csv.as_ref()
+            && csv_input_requires_normalization(csv.quote_character.as_deref(), csv.quote_escape_character.as_deref())
+        {
+            let syntax = CsvSyntax {
+                quote: csv.quote_character.as_deref(),
+                escape: csv.quote_escape_character.as_deref(),
+                field: csv.field_delimiter.as_deref(),
+                record: csv.record_delimiter.as_deref(),
+                comment: csv.comments.as_ref().and_then(|comment| comment.as_bytes().first().copied()),
+            };
+            return normalize_csv_stream(stream, &syntax);
+        }
+        convert_csv_delimiter_stream(
+            stream,
+            self.record_delimiter_for_conversion(),
+            self.need_convert.then(|| self.delimiter.clone()),
+        )
+    }
+
    fn csv_has_header(&self) -> bool {
        self.input
            .request
@@ -820,7 +844,6 @@ impl ObjectStore for EcObjectStore {
            });
        }

-        let record_delimiter = self.record_delimiter_for_conversion();
        let needs_scan_context = options.range.is_none() && has_effective_request_range;
        let scan_context = if needs_scan_context {
            if let Some(scan_range) = self.scan_range(original_size)? {
@@ -883,8 +906,7 @@ impl ObjectStore for EcObjectStore {
                    max_processed_bytes,
                    query_guard,
                )?;
-                let stream =
-                    convert_csv_delimiter_stream(stream, record_delimiter, self.need_convert.then(|| self.delimiter.clone()));
+                let stream = self.convert_csv_stream(stream);
                GetResultPayload::Stream(stream)
            }
        } else if options.range.is_some() {
@@ -937,8 +959,7 @@ impl ObjectStore for EcObjectStore {
            } else {
                stream
            };
-            let stream =
-                convert_csv_delimiter_stream(stream, record_delimiter, self.need_convert.then(|| self.delimiter.clone()));
+            let stream = self.convert_csv_stream(stream);
            GetResultPayload::Stream(stream)
        } else {
            let stream_size = usize::try_from(original_size).map_err(|err| o_Error::Generic {
@@ -948,8 +969,7 @@ impl ObjectStore for EcObjectStore {
            let stream = bytes_stream(ReaderStream::with_capacity(reader.stream, SELECT_DEFAULT_READ_BUFFER_SIZE), stream_size);
            if meter_input {
                let stream = meter_uncompressed_input_stream(stream, Arc::clone(&self.input_metrics));
-                let stream =
-                    convert_csv_delimiter_stream(stream, record_delimiter, self.need_convert.then(|| self.delimiter.clone()));
+                let stream = self.convert_csv_stream(stream);
                GetResultPayload::Stream(stream)
            } else {
                GetResultPayload::Stream(stream.boxed())
@@ -2866,6 +2886,88 @@ mod test {
        assert_eq!(input_metrics.snapshot().bytes_processed, 2);
    }

+    #[tokio::test]
+    async fn unicode_csv_quotes_preserve_raw_offsets_and_metrics() {
+        const BUCKET: &str = "s3select-unicode-csv-stream";
+        const HEADER: &str = "عnameع,عkindع\n";
+        const SKIP: &str = "عskipع,عzeroع\n";
+        const ROW: &str = "عA,Bع,عAععBع\n";
+        let data = format!("{HEADER}{SKIP}{ROW}");
+        let env = crate::storage_api::select_test_ecstore_env().await;
+        env.make_bucket(BUCKET, false).await;
+        for (object, compression, range_offset) in [
+            ("plain.csv", None, None),
+            ("range.csv", None, Some(0)),
+            ("range-mid-character.csv", None, Some(1)),
+            ("gzip.csv", Some(CompressionFormat::Gzip), None),
+            ("bzip.csv", Some(CompressionFormat::Bzip2), None),
+        ] {
+            let bytes = match compression {
+                Some(format) => encode_compressed_fixture(format, data.as_bytes()).await,
+                None => data.as_bytes().to_vec(),
+            };
+            let raw_size = bytes.len();
+            let mut reader = SelectPutObjReader::from_vec(bytes);
+            env.ecstore
+                .put_object(BUCKET, object, &mut reader, &Default::default())
+                .await
+                .expect("write Unicode CSV fixture");
+            let mut input = (*csv_input(BUCKET, object)).clone();
+            let csv = input.request.input_serialization.csv.as_mut().expect("CSV input");
+            csv.file_header_info = Some(FileHeaderInfo::from_static(FileHeaderInfo::USE));
+            csv.quote_character = Some("ع".to_owned());
+            csv.quote_escape_character = Some("\\".to_owned());
+            csv.record_delimiter = Some("\n".to_owned());
+            input.request.input_serialization.compression_type = compression.map(|format| {
+                CompressionType::from_static(match format {
+                    CompressionFormat::Gzip => CompressionType::GZIP,
+                    CompressionFormat::Bzip2 => CompressionType::BZIP2,
+                })
+            });
+            let start = HEADER.len() + SKIP.len();
+            if let Some(range_offset) = range_offset {
+                let offset = i64::try_from(start + range_offset).expect("fixture offset");
+                input.request.scan_range = Some(ScanRange {
+                    start: Some(offset),
+                    end: Some(offset),
+                });
+            }
+            let metrics = Arc::new(SelectInputMetrics::default());
+            let store = EcObjectStore::build_with_snapshot(
+                Arc::new(input),
+                Arc::new(GreedyMemoryPool::new(1024 * 1024)),
+                None,
+                Arc::clone(&metrics),
+                prepare_test_snapshot(BUCKET, object).await,
+                JsonSource::default(),
+            )
+            .expect("snapshot store");
+            let result = store
+                .get_opts(&Path::from(object), GetOptions::default())
+                .await
+                .expect("open Unicode CSV stream");
+            let GetResultPayload::Stream(stream) = result.payload else { panic!("CSV must remain streaming") };
+            let output = stream.try_collect::<Vec<_>>().await.expect("normalize CSV stream").concat();
+            let expected = match range_offset {
+                Some(0) => "\"name\",\"kind\"\n\"A,B\",\"AعB\"\n",
+                Some(_) => "\"name\",\"kind\"\n",
+                None => "\"name\",\"kind\"\n\"skip\",\"zero\"\n\"A,B\",\"AعB\"\n",
+            };
+            assert_eq!(output, expected.as_bytes(), "object={object}");
+            let measured = metrics.snapshot();
+            if let Some(range_offset) = range_offset {
+                // The range reader includes one byte of delimiter context and a
+                // separate header read; offsets always refer to the original CSV.
+                let processed = u64::try_from(ROW.len() + 1 - range_offset + HEADER.len()).expect("raw range length");
+                assert_eq!(measured.bytes_scanned, processed);
+                assert_eq!(measured.bytes_processed, processed);
+            } else {
+                assert_eq!(measured.bytes_scanned, u64::try_from(raw_size).expect("raw length"));
+                assert_eq!(measured.bytes_processed, u64::try_from(data.len()).expect("decoded length"));
+            }
+        }
+    }
+
    #[tokio::test]
    async fn compressed_object_uses_one_full_stream_and_rejects_internal_ranges() {
        const BUCKET: &str = "s3select-compressed-object";
@@ -456,7 +456,12 @@ impl SessionCtxFactory {
            .is_some_and(|compression| compression.as_str() != CompressionType::NONE);
        let metered_input_requires_single_file_scan =
            input_metrics.is_some() && context.input.request.input_serialization.parquet.is_none();
-        let config = if custom_two_byte_record_delimiter
+        let normalized_csv_requires_single_file_scan =
+            context.input.request.input_serialization.csv.as_ref().is_some_and(|csv| {
+                crate::csv_input_requires_normalization(csv.quote_character.as_deref(), csv.quote_escape_character.as_deref())
+            });
+        let config = if normalized_csv_requires_single_file_scan
+            || custom_two_byte_record_delimiter
            || scan_range_requires_single_file_scan
            || json_document_requires_single_file_scan
            || compressed_input_requires_single_file_scan
@@ -906,6 +911,25 @@ mod tests {
        assert!(!session.inner().config().options().optimizer.repartition_file_scans);
    }

+    #[tokio::test]
+    async fn unicode_csv_quotes_disable_file_scan_repartition() {
+        let mut context = test_context();
+        Arc::get_mut(&mut context.input)
+            .expect("unique context")
+            .request
+            .input_serialization
+            .csv
+            .as_mut()
+            .expect("CSV input")
+            .quote_character = Some("ع".to_owned());
+        let session = SessionCtxFactory::new(true)
+            .with_target_partitions(4)
+            .create_session_ctx(&context)
+            .await
+            .expect("Unicode CSV session");
+        assert!(!session.inner().config().options().optimizer.repartition_file_scans);
+    }
+
    #[tokio::test]
    async fn two_byte_csv_record_delimiter_disables_file_scan_repartition() {
        let mut context = test_context();
@@ -53,7 +53,7 @@ use rustfs_s3select_api::{
        },
    },
 };
-use s3s::dto::{CompressionType, FileHeaderInfo, JSONType, SelectObjectContentInput};
+use s3s::dto::{FileHeaderInfo, JSONType, SelectObjectContentInput};
 use std::sync::LazyLock;
 use tokio::{
    sync::Semaphore,
@@ -430,13 +430,6 @@ impl SimpleQueryDispatcher {

        let path = format!("s3://{}/{}", self.input.bucket, self.input.key);
        let table_path = ListingTableUrl::parse(path)?;
-        let compressed_input = self
-            .input
-            .request
-            .input_serialization
-            .compression_type
-            .as_ref()
-            .is_some_and(|compression| compression.as_str() != CompressionType::NONE);
        let (listing_options, need_rename_volume_name, need_ignore_volume_name) =
            if let Some(csv) = self.input.request.input_serialization.csv.as_ref() {
                let mut need_rename_volume_name = false;
@@ -485,28 +478,27 @@ impl SimpleQueryDispatcher {
                if let Some(quote) = csv.quote_character.as_ref() {
                    file_format = file_format.with_quote(quote.as_bytes().first().copied().unwrap_or_default());
                }
+                if rustfs_s3select_api::csv_input_requires_normalization(
+                    csv.quote_character.as_deref(),
+                    csv.quote_escape_character.as_deref(),
+                ) {
+                    file_format = file_format
+                        .with_quote(b'"')
+                        .with_escape(None)
+                        .with_delimiter(b',')
+                        .with_terminator(Some(b'\n'))
+                        .with_comment(None)
+                        .with_newlines_in_values(true);
+                }
                (
-                    ListingOptions::new(Arc::new(file_format)).with_file_extension(if compressed_input {
-                        EXACT_OBJECT_FILE_EXTENSION
-                    } else {
-                        ".csv"
-                    }),
+                    ListingOptions::new(Arc::new(file_format)).with_file_extension(EXACT_OBJECT_FILE_EXTENSION),
                    need_rename_volume_name,
                    need_ignore_volume_name,
                )
            } else if self.input.request.input_serialization.json.is_some() {
                let file_format = JsonFormat::default();
-                let file_extension = if compressed_input {
-                    EXACT_OBJECT_FILE_EXTENSION.to_string()
-                } else {
-                    std::path::Path::new(&self.input.key)
-                        .extension()
-                        .and_then(|extension| extension.to_str())
-                        .map(|extension| format!(".{extension}"))
-                        .unwrap_or_else(|| ".json".to_string())
-                };
                (
-                    ListingOptions::new(Arc::new(file_format)).with_file_extension(file_extension),
+                    ListingOptions::new(Arc::new(file_format)).with_file_extension(EXACT_OBJECT_FILE_EXTENSION),
                    false,
                    false,
                )
@@ -1531,6 +1523,130 @@ mod tests {
        assert_eq!(error.select_error(), SelectError::InvalidDataSource);
    }

+    #[tokio::test]
+    async fn unicode_csv_quotes_reach_arrow_without_changing_field_values() {
+        let cases = [
+            ("ع", "\"", ",", "\n", "عcol1ع,عcol2ع,عcol3ع\n", vec![vec!["col1", "col2", "col3"]]),
+            (
+                "ع",
+                "\\",
+                ",",
+                "\n",
+                "\"literal\",عA\\\"Bع,عAععBع\n",
+                vec![vec!["\"literal\"", "A\"B", "AعB"]],
+            ),
+            ("\"", "界", ",", "\n", "\"A界\"B\",🦀\n", vec![vec!["A\"B", "🦀"]]),
+            ("ع", "\\", "界", "^Y", "عa界bع界عline\nbreakع^Y", vec![vec!["a界b", "line\nbreak"]]),
+        ];
+        let env = snapshot_test_env().await;
+        for (index, (quote, escape, field, record, data, expected)) in cases.into_iter().enumerate() {
+            let mut input = test_input();
+            input.bucket = format!("select-unicode-quotes-{index}");
+            input.key = "records".to_owned();
+            let csv = input.request.input_serialization.csv.as_mut().expect("CSV input");
+            csv.file_header_info = Some(FileHeaderInfo::from_static(FileHeaderInfo::NONE));
+            csv.quote_character = Some(quote.to_owned());
+            csv.quote_escape_character = Some(escape.to_owned());
+            csv.field_delimiter = Some(field.to_owned());
+            csv.record_delimiter = Some(record.to_owned());
+            env.make_bucket(&input.bucket, false).await;
+            env.put_object_bytes(&input.bucket, &input.key, data.as_bytes().to_vec())
+                .await;
+            let snapshot = env.prepare_select_object_snapshot(&input.bucket, &input.key).await;
+            let input = Arc::new(input);
+            let dispatcher = production_dispatcher(Arc::clone(&input));
+            let query = Query::new_with_snapshot(QueryContext { input }, "SELECT * FROM S3Object".to_owned(), snapshot);
+            let output = dispatcher.execute_query(&query).await.expect("execute Unicode CSV query");
+            let mut stream = output.into_record_batch_stream().expect("record stream");
+            let mut rows = Vec::new();
+            while let Some(batch) = stream.next().await {
+                let batch = batch.expect("Arrow must receive valid UTF-8 fields");
+                for row in 0..batch.num_rows() {
+                    rows.push(
+                        batch
+                            .columns()
+                            .iter()
+                            .map(|column| {
+                                column
+                                    .as_any()
+                                    .downcast_ref::<StringArray>()
+                                    .expect("CSV string column")
+                                    .value(row)
+                                    .to_owned()
+                            })
+                            .collect::<Vec<_>>(),
+                    );
+                }
+            }
+            assert_eq!(rows, expected, "fixture={index}");
+        }
+    }
+
+    #[tokio::test]
+    async fn select_uses_input_serialization_independently_of_object_extension() {
+        for (key, json) in [
+            ("records", false),
+            ("records.bin", false),
+            ("records", true),
+            ("records.csv", true),
+        ] {
+            let mut input = test_input();
+            input.key = key.to_owned();
+            let data = if json {
+                input.request.input_serialization.csv = None;
+                input.request.input_serialization.json = Some(s3s::dto::JSONInput {
+                    type_: Some(JSONType::from_static(JSONType::LINES)),
+                });
+                b"{\"value\":\"selected\"}\n".as_slice()
+            } else {
+                b"value\nselected\n".as_slice()
+            };
+            let input = Arc::new(input);
+            let optimizer = Arc::new(CascadeOptimizerBuilder::default().build());
+            let dispatcher = test_dispatcher_for_input(
+                Arc::clone(&input),
+                Arc::new(Semaphore::new(1)),
+                Duration::from_secs(30),
+                Arc::new(SqlQueryExecutionFactory::new(optimizer, Arc::new(LocalScheduler {}))),
+            );
+            let query = Query::new(QueryContext { input }, "SELECT * FROM S3Object".to_owned());
+            let machine = dispatcher.build_query_state_machine(query).await.expect("build query state");
+            let store_url = ObjectStoreUrl::parse("s3://test-bucket").expect("test store URL");
+            let store = machine
+                .session
+                .inner()
+                .runtime_env()
+                .object_store(&store_url)
+                .expect("test store");
+            store.put(&Path::from(key), data.into()).await.expect("write selected object");
+            store
+                .put(&Path::from(format!("{key}.other")), b"unrelated\nwrong\n".as_slice().into())
+                .await
+                .expect("write neighboring object");
+            let plan = dispatcher
+                .build_logical_plan(Arc::clone(&machine))
+                .await
+                .expect("infer schema without an extension filter")
+                .expect("select plan");
+            let output = dispatcher
+                .execute_logical_plan(plan, machine)
+                .await
+                .expect("execute selected object");
+            let mut stream = output.into_record_batch_stream().expect("record stream");
+            let mut values = Vec::new();
+            while let Some(batch) = stream.next().await {
+                let batch = batch.expect("selected batch");
+                let column = batch
+                    .column(0)
+                    .as_any()
+                    .downcast_ref::<datafusion::arrow::array::StringArray>()
+                    .expect("string column");
+                values.extend(column.iter().map(|value| value.expect("selected value").to_owned()));
+            }
+            assert_eq!(values, ["selected"], "key={key}, json={json}");
+        }
+    }
+
    #[tokio::test]
    async fn csv_query_uses_custom_record_delimiter_across_file_partitions() {
        const ROW_COUNT: usize = 200_000;
@@ -28,10 +28,10 @@ use metrics::{counter, describe_counter, describe_histogram, histogram};
 use rustfs_config::ENV_SCANNER_CACHE_SAVE_TIMEOUT_SECS;
 pub use rustfs_data_usage::{
    AllTierStats, BucketTargetUsageInfo, BucketUsageInfo, DATA_USAGE_OBJECT_NAME, DATA_USAGE_OBSERVED_OBJECT_NAME,
-    DataUsageEntry, DataUsageHash, DataUsageHashMap, DataUsageInfo, DataUsageSnapshotSetState, LEGACY_DATA_USAGE_OBJECT_NAME,
-    PrefixUsageEntry, PrefixUsageQuery, PrefixUsageSummary, ReplTargetSizeSummary, SizeReconciliationEntry,
-    SizeReconciliationScope, SizeSummary, TierAccountingProof, TierStats, UNKNOWN_TIER, UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP,
-    UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP, UnknownTierStats, hash_path, prefix_usage_in_cache,
+    DataUsageEntry, DataUsageHash, DataUsageHashMap, DataUsageInfo, DataUsageSegmentInvalidationProof, DataUsageSnapshotSetState,
+    LEGACY_DATA_USAGE_OBJECT_NAME, PrefixUsageEntry, PrefixUsageQuery, PrefixUsageSummary, ReplTargetSizeSummary,
+    SizeReconciliationEntry, SizeReconciliationScope, SizeSummary, TierAccountingProof, TierStats, UNKNOWN_TIER,
+    UNKNOWN_TIER_DIAGNOSTIC_BYTE_CAP, UNKNOWN_TIER_DIAGNOSTIC_ENTRY_CAP, UnknownTierStats, hash_path, prefix_usage_in_cache,
 };
 use rustfs_heal_contracts::heal_channel::HealScanMode;
 use rustfs_utils::path::{SLASH_SEPARATOR, path_join_buf};
@@ -657,6 +657,11 @@ pub struct DataUsageCacheInfo {
    /// structural plan remains reusable across ordinary bucket writes.
    #[serde(default)]
    pub scan_execution_digest: Option<DataUsageScanPlanDigest>,
+    /// Process-epoch and generation window that produced a complete set cache
+    /// with all known segment invalidation producers wired. This proof is
+    /// additive compatibility metadata; absence keeps segment reuse disabled.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub segment_invalidation_proof: Option<DataUsageSegmentInvalidationProof>,
    /// Durable bucket incarnations captured for a complete set aggregate.
    /// Missing or nil entries are legacy/unproven and cannot authorize
    /// skipping an unselected bucket in a later scoped set scan.
@@ -686,6 +691,7 @@ impl Serialize for DataUsageCacheInfo {
            + usize::from(self.lkg_leader_epoch.is_some())
            + usize::from(self.lkg_scan_plan_digest.is_some())
            + usize::from(self.scan_execution_digest.is_some())
+            + usize::from(self.segment_invalidation_proof.is_some())
            + usize::from(!self.scan_bucket_incarnations.is_empty());
        let mut state = serializer.serialize_map(Some(field_count))?;
        state.serialize_entry("name", &self.name)?;
@@ -746,6 +752,9 @@ impl Serialize for DataUsageCacheInfo {
        if let Some(scan_execution_digest) = self.scan_execution_digest {
            state.serialize_entry("scan_execution_digest", &scan_execution_digest)?;
        }
+        if let Some(proof) = &self.segment_invalidation_proof {
+            state.serialize_entry("segment_invalidation_proof", proof)?;
+        }
        if !self.scan_bucket_incarnations.is_empty() {
            state.serialize_entry("scan_bucket_incarnations", &self.scan_bucket_incarnations)?;
        }
@@ -1095,6 +1095,7 @@ fn test_data_usage_cache_info_deserialize_defaults_scan_resume_after() {
    assert!(!decoded.snapshot_complete);
    assert!(decoded.scan_plan_digest.is_none());
    assert!(decoded.scan_execution_digest.is_none());
+    assert!(decoded.segment_invalidation_proof.is_none());
    assert_eq!(decoded.cache_key_format, 0);
 }

@@ -1183,6 +1184,13 @@ fn test_new_data_usage_cache_msgpack_round_trips_and_supports_old_reader() {
            snapshot_complete: true,
            scan_plan_digest: Some(TEST_PLAN_DIGEST),
            scan_execution_digest: Some(DataUsageScanPlanDigest([42; 32])),
+            segment_invalidation_proof: Some(DataUsageSegmentInvalidationProof {
+                process_epoch: "scanner-process".to_string(),
+                generation_start: 7,
+                generation_end: 9,
+                producer_identity_coverage_complete: true,
+                cold_zero_walk_oracle: true,
+            }),
            cache_key_format: DATA_USAGE_CACHE_KEY_FORMAT,
            ..Default::default()
        },
@@ -1212,6 +1220,16 @@ fn test_new_data_usage_cache_msgpack_round_trips_and_supports_old_reader() {
    assert!(current.info.snapshot_complete);
    assert_eq!(current.info.scan_plan_digest, Some(TEST_PLAN_DIGEST));
    assert_eq!(current.info.scan_execution_digest, Some(DataUsageScanPlanDigest([42; 32])));
+    assert_eq!(
+        current.info.segment_invalidation_proof,
+        Some(DataUsageSegmentInvalidationProof {
+            process_epoch: "scanner-process".to_string(),
+            generation_start: 7,
+            generation_end: 9,
+            producer_identity_coverage_complete: true,
+            cold_zero_walk_oracle: true,
+        })
+    );
    assert_eq!(current.info.cache_key_format, DATA_USAGE_CACHE_KEY_FORMAT);
    assert_eq!(current.find("bucket").map(|entry| entry.objects), Some(3));

@@ -97,9 +97,11 @@ pub use scanner::{
 pub use scanner_io::{
    ScannerDirtyUsageAckError, ScannerDirtyUsageBucket, ScannerDirtyUsageSnapshot, ScannerDirtyUsageState,
    acknowledge_dirty_usage_generation, acknowledge_scoped_dirty_usage, clear_dirty_usage_bucket, record_dirty_usage_bucket,
-    record_dirty_usage_object, record_scanner_maintenance_change, scanner_activity_epoch, scanner_dirty_usage_snapshot,
-    scanner_dirty_usage_state, scanner_maintenance_generation,
+    record_dirty_usage_bucket_from_producer, record_dirty_usage_bucket_from_producers, record_dirty_usage_object,
+    record_dirty_usage_object_from_producer, record_scanner_maintenance_change, scanner_activity_epoch,
+    scanner_dirty_usage_snapshot, scanner_dirty_usage_state, scanner_maintenance_generation,
 };
+pub use segment_invalidation::SegmentInvalidationProducerIdentity;
 pub use sleeper::{DynamicSleeper, SCANNER_IDLE_MODE, SCANNER_SLEEPER};
 use std::sync::atomic::{AtomicU64, Ordering};
 pub use storage_api::ScannerReplicationConfig as ReplicationConfig;
@@ -964,9 +964,10 @@ async fn run_scanner_usage_recovery_intents_for_startup(
    Ok(attempted)
 }

-/// Start normal scanning when enabled, or one resume-only cleanup attempt.
+/// Start normal scanning when enabled, or one bounded recovery attempt.
 /// The disabled branch returns a finite task for the startup owner to join;
-/// it never enables ordinary namespace scanning or accepts a new reset intent.
+/// it never enables ordinary namespace scanning while it replays durable reset
+/// intents and cleanup markers.
 pub async fn init_scanner_with_recovery(
    ctx: CancellationToken,
    storeapi: Arc<ECStore>,
@@ -988,6 +989,17 @@ pub async fn init_scanner_with_recovery(
        return None;
    }
    Some(tokio::spawn(async move {
+        if let Err(error) = run_scanner_usage_recovery_intents_for_startup(ctx.clone(), storeapi.clone()).await {
+            warn!(
+                target: "rustfs::scanner",
+                event = EVENT_SCANNER_PERSIST_STATE,
+                component = LOG_COMPONENT_SCANNER,
+                subsystem = LOG_SUBSYSTEM_RUNTIME,
+                state = "recovery_intent_startup_discovery_failed",
+                error = %error,
+                "Scanner recovery intent startup discovery failed"
+            );
+        }
        if let Err(error) = resume_scanner_cycle_cleanup(ctx, storeapi).await {
            warn!(
                target: "rustfs::scanner",
@@ -128,6 +128,79 @@ pub(super) mod cleanup_io_fault {
    }
 }

+#[cfg(test)]
+pub(super) mod recovery_intent_accept_fault {
+    use super::*;
+
+    enum Fault {
+        Corrupt,
+        Running,
+    }
+
+    static NEXT_ACCEPT_READBACK_FAULT: StdMutex<Option<Fault>> = StdMutex::new(None);
+
+    pub(in crate::scanner) struct Guard;
+
+    impl Drop for Guard {
+        fn drop(&mut self) {
+            *NEXT_ACCEPT_READBACK_FAULT
+                .lock()
+                .unwrap_or_else(|poisoned| poisoned.into_inner()) = None;
+        }
+    }
+
+    pub(in crate::scanner) fn corrupt_next_accept_readback() -> Guard {
+        install(Fault::Corrupt)
+    }
+
+    pub(in crate::scanner) fn advance_next_accept_readback_to_running() -> Guard {
+        install(Fault::Running)
+    }
+
+    fn install(fault: Fault) -> Guard {
+        let mut slot = NEXT_ACCEPT_READBACK_FAULT
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner());
+        assert!(slot.is_none(), "only one recovery intent accept readback fault may be installed");
+        *slot = Some(fault);
+        Guard
+    }
+
+    pub(super) async fn maybe_apply<S>(storeapi: Arc<S>, path: &str) -> Result<(), ScannerError>
+    where
+        S: ScannerObjectIO,
+    {
+        let Some(fault) = NEXT_ACCEPT_READBACK_FAULT
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .take()
+        else {
+            return Ok(());
+        };
+        match fault {
+            Fault::Corrupt => save_config(storeapi, path, b"{corrupt".to_vec()).await.map_err(|err| {
+                ScannerError::Other(format!("failed to inject scanner recovery intent accept readback fault: {err}"))
+            }),
+            Fault::Running => {
+                let mut record = read_recovery_intent_record(storeapi.clone(), path).await?.ok_or_else(|| {
+                    ScannerError::Other("scanner recovery intent disappeared before fault injection".to_string())
+                })?;
+                record.state = SCANNER_RECOVERY_INTENT_STATE_RUNNING.to_string();
+                save_config(
+                    storeapi,
+                    path,
+                    serde_json::to_vec(&record)
+                        .map_err(|err| ScannerError::Other(format!("failed to encode scanner recovery intent fault: {err}")))?,
+                )
+                .await
+                .map_err(|err| {
+                    ScannerError::Other(format!("failed to inject scanner recovery intent accept readback fault: {err}"))
+                })
+            }
+        }
+    }
+}
+
 #[derive(Clone, Debug, Default, Serialize)]
 pub struct ScannerCycleRecoveryStatus {
    /// The immutable primary object whose revision is being guarded.
@@ -564,6 +637,17 @@ fn compare_recovery_intent(
    }
 }

+fn confirm_recovery_intent_acceptance(
+    expected: ScannerRecoveryIntentRecord,
+    persisted: ScannerRecoveryIntentRecord,
+) -> ScannerRecoveryIntentAcceptResult {
+    if persisted == expected {
+        ScannerRecoveryIntentAcceptResult::Accepted { record: persisted }
+    } else {
+        compare_recovery_intent(&expected, persisted)
+    }
+}
+
 async fn read_recovery_intent_record(
    storeapi: Arc<impl ScannerObjectIO>,
    path: &str,
@@ -757,7 +841,20 @@ pub async fn accept_scanner_usage_recovery_intent(
        .map_err(|err| ScannerError::Other(format!("failed to encode scanner recovery intent: {err}")))?;
    match save_config_with_preconditions(storeapi.clone(), &path, encoded, DataUsageCacheRevision::Missing.preconditions()).await
    {
-        Ok(_) => Ok(ScannerRecoveryIntentAcceptResult::Accepted { record: candidate }),
+        Ok(_) => {
+            #[cfg(test)]
+            recovery_intent_accept_fault::maybe_apply(storeapi.clone(), &path).await?;
+            let persisted = match read_recovery_intent_record(storeapi.clone(), &path).await {
+                Ok(Some(record)) => record,
+                Ok(None) => {
+                    return Err(ScannerError::Other(
+                        "scanner recovery intent disappeared before acceptance confirmation".to_string(),
+                    ));
+                }
+                Err(error) => return Err(error),
+            };
+            Ok(confirm_recovery_intent_acceptance(candidate, persisted))
+        }
        Err(EcstoreError::PreconditionFailed) => {
            let existing = read_recovery_intent_record(storeapi, &path).await?;
            let Some(existing) = existing else {
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

-use super::super::cycle_state::cleanup_io_fault;
+use super::super::cycle_state::{cleanup_io_fault, recovery_intent_accept_fault};
 use super::*;
 use crate::storage_api::owner::{EcstoreRebalStatus, EcstoreRebalanceInfo, EcstoreRebalanceMeta, EcstoreRebalanceStats};

@@ -256,6 +256,100 @@ async fn scanner_recovery_intent_accept_is_durable_and_idempotent() {
    assert_eq!(replay, ScannerRecoveryIntentAcceptResult::Replayed { record });
 }

+#[tokio::test]
+#[serial]
+async fn scanner_recovery_intent_accept_requires_confirmed_readback() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    let _fault = recovery_intent_accept_fault::corrupt_next_accept_readback();
+    let error =
+        accept_scanner_usage_recovery_intent(store.clone(), recovery_intent_request("intent-key-0001-readback", "operator-a"))
+            .await
+            .expect_err("accept must fail when the just-written intent cannot be confirmed");
+    assert!(error.to_string().contains("scanner recovery intent is invalid"), "{error}");
+
+    let restarted = restart_scanner_cycle_store_from(&store).await;
+    let error = scanner_usage_recovery_intents_for_startup(&CancellationToken::new(), restarted)
+        .await
+        .expect_err("unconfirmed corrupt intent must remain a fail-closed startup error");
+    assert!(error.to_string().contains("scanner recovery intent is invalid"), "{error}");
+}
+
+#[tokio::test]
+#[serial]
+async fn scanner_recovery_intent_accept_replays_if_execution_advances_before_readback() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    let request = recovery_intent_request("intent-key-0001-running", "operator-a");
+    let _fault = recovery_intent_accept_fault::advance_next_accept_readback_to_running();
+    let replay = accept_scanner_usage_recovery_intent(store.clone(), request.clone())
+        .await
+        .expect("same request advanced by execution remains idempotent");
+    let record = match replay {
+        ScannerRecoveryIntentAcceptResult::Replayed { record } => record,
+        other => panic!("advanced same-request record must replay instead of conflict: {other:?}"),
+    };
+    assert_eq!(record.state, "running");
+
+    let retry = accept_scanner_usage_recovery_intent(store, request)
+        .await
+        .expect("lost response retry observes the running durable record");
+    assert_eq!(retry, ScannerRecoveryIntentAcceptResult::Replayed { record });
+}
+
+#[tokio::test]
+#[serial]
+async fn concurrent_scanner_recovery_intent_acceptance_uses_one_durable_record() {
+    let (_dir, store) = setup_scanner_cycle_store().await;
+    let request = recovery_intent_request("intent-key-0001-concurrent", "operator-a");
+    let mut tasks = Vec::new();
+    for _ in 0..12 {
+        let store = store.clone();
+        let request = request.clone();
+        tasks.push(tokio::spawn(async move {
+            accept_scanner_usage_recovery_intent(store, request)
+                .await
+                .expect("concurrent same-key accept should converge")
+        }));
+    }
+
+    let mut accepted = 0usize;
+    let mut replayed = 0usize;
+    let mut records = Vec::new();
+    for task in tasks {
+        match task.await.expect("accept task should not panic") {
+            ScannerRecoveryIntentAcceptResult::Accepted { record } => {
+                accepted += 1;
+                records.push(record);
+            }
+            ScannerRecoveryIntentAcceptResult::Replayed { record } => {
+                replayed += 1;
+                records.push(record);
+            }
+            other => panic!("same-key accepts must not conflict: {other:?}"),
+        }
+    }
+    assert_eq!(accepted, 1, "exactly one request may win the missing-record CAS");
+    assert_eq!(replayed, 11, "all other same-key requests must replay the durable winner");
+    assert!(
+        records.windows(2).all(|pair| pair[0] == pair[1]),
+        "all accepts must return the same durable identity"
+    );
+
+    let restarted = restart_scanner_cycle_store_from(&store).await;
+    let replayable = scanner_usage_recovery_intents_for_startup(&CancellationToken::new(), restarted.clone())
+        .await
+        .expect("startup should rediscover the single indexed intent");
+    assert_eq!(replayable, vec![records[0].intent_id.clone()]);
+    let replay = accept_scanner_usage_recovery_intent(restarted, request)
+        .await
+        .expect("lost response after restart should replay the same record");
+    assert_eq!(
+        replay,
+        ScannerRecoveryIntentAcceptResult::Replayed {
+            record: records[0].clone()
+        }
+    );
+}
+
 #[tokio::test]
 #[serial]
 async fn scanner_recovery_intent_executor_persists_completed_progress() {
@@ -457,7 +551,7 @@ async fn scanner_recovery_intent_startup_rejects_corrupt_pending_record() {

 #[tokio::test]
 #[serial]
-async fn scanner_recovery_intent_disabled_startup_preserves_non_terminal_intent() {
+async fn scanner_recovery_intent_disabled_startup_replays_non_terminal_intent() {
    let (_dir, store) = setup_scanner_cycle_store().await;
    let record = match accept_scanner_usage_recovery_intent(
        store.clone(),
@@ -473,12 +567,12 @@ async fn scanner_recovery_intent_disabled_startup_preserves_non_terminal_intent(
    let restarted = restart_scanner_cycle_store_from(&store).await;
    run_disabled_startup(CancellationToken::new(), restarted.clone()).await;

-    let preserved = get_scanner_usage_recovery_intent(restarted, &record.intent_id)
+    let completed = get_scanner_usage_recovery_intent(restarted, &record.intent_id)
        .await
-        .expect("startup-skipped intent should read")
-        .expect("startup-skipped intent should remain durable");
-    assert_eq!(preserved.state, "accepted");
-    assert_eq!(preserved.intent_id, record.intent_id);
+        .expect("startup-replayed intent should read")
+        .expect("startup-replayed intent should remain durable");
+    assert_eq!(completed.state, "completed");
+    assert_eq!(completed.intent_id, record.intent_id);
 }

 #[tokio::test]
@@ -547,6 +641,17 @@ async fn scanner_recovery_intent_query_rejects_corrupt_or_unknown_records() {
        .await
        .expect_err("corrupt intent must not decode as absent");
    assert!(error.to_string().contains("scanner recovery intent is invalid"));
+
+    let mut future = serde_json::to_value(&record).expect("record value");
+    future["future_writer_capability"] = serde_json::json!("durable-accept-v2");
+    save_config(store.clone(), &path, serde_json::to_vec(&future).expect("future record should encode"))
+        .await
+        .expect("future durable record");
+    let error = get_scanner_usage_recovery_intent(store.clone(), &record.intent_id)
+        .await
+        .expect_err("future writer payload must not decode as a known terminal state");
+    assert!(error.to_string().contains("scanner recovery intent is invalid"));
+
    let unknown = get_scanner_usage_recovery_intent(store, &scanner_recovery_actor_sha256("missing"))
        .await
        .expect("missing intent should read as absent");
@@ -3,7 +3,8 @@
 use super::*;
 use crate::segment_invalidation::{
    MAX_SEGMENT_INVALIDATION_BYTES, MAX_SEGMENT_INVALIDATION_ENTRIES, SegmentInvalidationDomain, SegmentInvalidationEnvelope,
-    SegmentInvalidationError, SegmentInvalidationProducer, SegmentInvalidationProof, admit_segment_invalidation,
+    SegmentInvalidationError, SegmentInvalidationProducer, SegmentInvalidationProducerIdentity, SegmentInvalidationProof,
+    admit_segment_invalidation, complete_segment_invalidation_producers,
 };
 use std::collections::BTreeSet;

@@ -11,7 +12,8 @@ const MAX_WALK_SAMPLES: usize = 32;
 const MAX_WALK_BYTES: usize = 1024;

 fn segment_producers() -> BTreeSet<SegmentInvalidationProducer> {
-    SegmentInvalidationProducer::REQUIRED.into_iter().collect()
+    complete_segment_invalidation_producers(SegmentInvalidationProducerIdentity::REQUIRED_PRODUCTION)
+        .expect("fixture should enumerate the complete production producer matrix")
 }

 fn segment_envelope() -> SegmentInvalidationEnvelope {
@@ -39,7 +39,7 @@ use s3s::dto::{
    BucketLifecycleConfiguration, ObjectLockConfiguration, ObjectLockEnabled, ReplicationConfiguration, VersioningConfiguration,
 };
 use sha2::{Digest as _, Sha256};
-use std::collections::{HashMap, HashSet};
+use std::collections::{BTreeMap, BTreeSet, HashMap, HashSet};
 use std::future::Future;
 use std::path::Path;
 use std::pin::Pin;
@@ -98,6 +98,24 @@ const METRIC_SCANNER_SET_SCANS_QUEUED: &str = "rustfs_scanner_set_scans_queued";
 const METRIC_SCANNER_DISK_BUCKET_SCANS_ACTIVE: &str = "rustfs_scanner_disk_bucket_scans_active";
 const METRIC_SCANNER_DISK_BUCKET_SCANS_QUEUED: &str = "rustfs_scanner_disk_bucket_scans_queued";

+pub(crate) const SCANNER_SEGMENT_ACTIVATION_PROOF_INPUTS: [&str; 7] = [
+    "source",
+    "bucket_incarnation",
+    "key_format",
+    "baseline_scan_plan_digest",
+    "process_epoch",
+    "generation_window",
+    "producer_identities",
+];
+pub(crate) const SCANNER_SEGMENT_ACTIVATION_FAIL_CLOSED_CHECKS: [&str; 6] = [
+    "missing_producer_identity",
+    "restart_gap",
+    "generation_gap",
+    "overflow",
+    "missing_cold_zero_walk_oracle",
+    "distributed_without_peer_invalidation",
+];
+
 pub type DirtyUsageBuckets = HashMap<String, u64>;

 #[derive(Clone, Debug)]
@@ -163,11 +181,52 @@ struct ScannerPeerDirtyUsageExpectation {
 struct VerifiedRemoteDirtyUsage {
    dirty_buckets: HashSet<String>,
    acknowledgements: Vec<crate::scanner::ScannerDirtyUsageAcknowledgement>,
+    peer_count: usize,
+    dirty_peer_count: usize,
 }

 struct ScannerBucketScopeResolutionResult {
    scope: ScannerBucketScanScope,
    remote_dirty_usage_acknowledgements: Vec<crate::scanner::ScannerDirtyUsageAcknowledgement>,
+    distributed_segment_invalidation_evidence: Option<DistributedSegmentInvalidationEvidence>,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) struct DistributedSegmentInvalidationEvidence {
+    pub(crate) invalidation_domain: crate::segment_invalidation::SegmentInvalidationDomain,
+    pub(crate) distributed_ec_invalidation: bool,
+    pub(crate) peer_count: usize,
+    pub(crate) dirty_peer_count: usize,
+    pub(crate) same_window_remote_proof: bool,
+    pub(crate) all_peers_bound_to_generation_window: bool,
+}
+
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub(crate) struct ScannerSegmentReuseActivationProof {
+    pub(crate) production_activation: bool,
+    pub(crate) producer_identity_coverage_complete: bool,
+    pub(crate) durable_producer_identity: bool,
+    pub(crate) restart_gap_absent: bool,
+    pub(crate) generation_window_bound: bool,
+    pub(crate) overflow_absent: bool,
+    pub(crate) cold_zero_walk_oracle: bool,
+    pub(crate) distributed_peer_invalidation: bool,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) struct ScannerSegmentReuseActivationPreflight {
+    pub(crate) production_activation: bool,
+    pub(crate) scanner_segment_reuse_activated: bool,
+    pub(crate) proof_inputs: &'static [&'static str],
+    pub(crate) fail_closed_checks: &'static [&'static str],
+    pub(crate) fail_closed_blockers: [Option<&'static str>; 6],
+}
+
+impl ScannerSegmentReuseActivationPreflight {
+    #[cfg(test)]
+    pub(crate) fn fail_closed_blockers(&self) -> impl Iterator<Item = &'static str> + '_ {
+        self.fail_closed_blockers.iter().filter_map(|blocker| *blocker)
+    }
 }

 fn verified_remote_dirty_usage(
@@ -191,6 +250,7 @@ fn verified_remote_dirty_usage(
            || !snapshot.complete
            || snapshot.pending_bucket_count != u64::try_from(snapshot.buckets.len()).unwrap_or(u64::MAX)
            || (expected.pending && snapshot.pending_bucket_count == 0)
+            || (!expected.pending && snapshot.pending_bucket_count != 0)
        {
            return None;
        }
@@ -216,9 +276,13 @@ fn verified_remote_dirty_usage(
        }
    }

+    let peer_count = received_peers.len();
+    let dirty_peer_count = acknowledgements.len();
    (received_peers.len() == expected_peers.len()).then_some(VerifiedRemoteDirtyUsage {
        dirty_buckets,
        acknowledgements,
+        peer_count,
+        dirty_peer_count,
    })
 }

@@ -244,8 +308,11 @@ fn resolve_remote_dirty_usage_scope(
    let default_result = |scope: ScannerBucketScanScope| ScannerBucketScopeResolutionResult {
        scope,
        remote_dirty_usage_acknowledgements: Vec::new(),
+        distributed_segment_invalidation_evidence: None,
    };

+    let peer_count = remote_dirty_usage.peer_count;
+    let dirty_peer_count = remote_dirty_usage.dirty_peer_count;
    dirty_buckets.extend(remote_dirty_usage.dirty_buckets);
    // Peer snapshots contribute bucket names only; the local prefix scopes
    // would narrow a bucket a peer dirtied elsewhere, so the merged scope
@@ -283,10 +350,21 @@ fn resolve_remote_dirty_usage_scope(
    if scanner_scoped_dirty_usage_ack_exceeds_cost_threshold(&scoped_acknowledgements) {
        return default_result(ScannerBucketScanScope::default());
    }
+    let has_scoped_acknowledgements = !scoped_acknowledgements.is_empty();

    ScannerBucketScopeResolutionResult {
        scope,
        remote_dirty_usage_acknowledgements: scoped_acknowledgements,
+        distributed_segment_invalidation_evidence: (dirty_peer_count > 0 && has_scoped_acknowledgements).then_some(
+            DistributedSegmentInvalidationEvidence {
+                invalidation_domain: crate::segment_invalidation::SegmentInvalidationDomain::DistributedEc,
+                distributed_ec_invalidation: true,
+                peer_count,
+                dirty_peer_count,
+                same_window_remote_proof: true,
+                all_peers_bound_to_generation_window: true,
+            },
+        ),
    }
 }

@@ -358,6 +436,60 @@ fn complete_scanner_cache_baseline_plan_digest(proof: ScannerCacheBaselineProof<
    complete_scanner_cache_snapshot_plan_digest(&observed, proof, false)
 }

+fn scanner_segment_reuse_baseline_producer_evidence(
+    dirty_usage_snapshot: &DirtyUsageSnapshot,
+    baseline_proof: ScannerCacheBaselineProof<'_>,
+) -> (DirtyUsageProducerEvidence, bool) {
+    let mut evidence = dirty_usage_producer_evidence(dirty_usage_snapshot);
+    if !evidence.generation_window_bound || !evidence.producer_identity_coverage_complete {
+        return (evidence, false);
+    }
+
+    let Some(authoritative_data) = baseline_proof.authoritative_data else {
+        return (evidence, false);
+    };
+    let Ok(authoritative) = serde_json::from_slice::<DataUsageInfo>(authoritative_data) else {
+        return (evidence, false);
+    };
+    if complete_scanner_cache_snapshot_plan_digest(&authoritative, baseline_proof, true).is_none()
+        || !scanner_snapshot_set_states_have_segment_reuse_activation_proof(
+            &authoritative,
+            &evidence,
+            baseline_proof.expected_sources,
+        )
+    {
+        return (evidence, false);
+    }
+
+    evidence.durable_producer_identity = true;
+    evidence.restart_gap_absent = true;
+    (evidence, true)
+}
+
+fn scanner_snapshot_set_states_have_segment_reuse_activation_proof(
+    snapshot: &DataUsageInfo,
+    evidence: &DirtyUsageProducerEvidence,
+    expected_sources: &HashSet<DataUsageCacheSource>,
+) -> bool {
+    let mut covered_sources = HashSet::with_capacity(expected_sources.len());
+    let all_sets_proved = snapshot.usage_snapshot_set_states.iter().all(|state| {
+        let (Ok(pool_index), Ok(set_index)) = (usize::try_from(state.pool_index), usize::try_from(state.set_index)) else {
+            return false;
+        };
+        let source = DataUsageCacheSource::new(pool_index, set_index);
+        state.complete
+            && !state.tombstone
+            && expected_sources.contains(&source)
+            && covered_sources.insert(source)
+            && scanner_segment_invalidation_proof_matches(state.segment_invalidation_proof.as_ref(), evidence)
+            && state
+                .segment_invalidation_proof
+                .as_ref()
+                .is_some_and(|proof| proof.cold_zero_walk_oracle)
+    });
+    all_sets_proved && covered_sources.len() == expected_sources.len()
+}
+
 fn scoped_scan_scope_from_dirty_buckets(
    requested_scope: ScannerBucketScanScope,
    dirty_buckets: HashSet<String>,
@@ -406,10 +538,152 @@ fn scoped_scan_scope_from_dirty_buckets(
    ScannerBucketScanScope::from_dirty_buckets(selected_buckets, selected_bucket_prefixes, baseline_scan_plan_digest)
 }

+fn scanner_segment_reuse_activation_preflight() -> ScannerSegmentReuseActivationPreflight {
+    scanner_segment_reuse_activation_preflight_from_proof(ScannerSegmentReuseActivationProof {
+        production_activation: true,
+        ..Default::default()
+    })
+}
+
+fn scanner_segment_reuse_activation_preflight_from_proof(
+    proof: ScannerSegmentReuseActivationProof,
+) -> ScannerSegmentReuseActivationPreflight {
+    ScannerSegmentReuseActivationPreflight {
+        production_activation: proof.production_activation,
+        scanner_segment_reuse_activated: proof.production_activation
+            && proof.producer_identity_coverage_complete
+            && proof.durable_producer_identity
+            && proof.restart_gap_absent
+            && proof.generation_window_bound
+            && proof.overflow_absent
+            && proof.cold_zero_walk_oracle
+            && proof.distributed_peer_invalidation,
+        proof_inputs: &SCANNER_SEGMENT_ACTIVATION_PROOF_INPUTS,
+        fail_closed_checks: &SCANNER_SEGMENT_ACTIVATION_FAIL_CLOSED_CHECKS,
+        fail_closed_blockers: [
+            (!proof.producer_identity_coverage_complete || !proof.durable_producer_identity)
+                .then_some("missing_producer_identity"),
+            (!proof.restart_gap_absent).then_some("restart_gap"),
+            (!proof.generation_window_bound).then_some("generation_gap"),
+            (!proof.overflow_absent).then_some("overflow"),
+            (!proof.cold_zero_walk_oracle).then_some("missing_cold_zero_walk_oracle"),
+            (!proof.distributed_peer_invalidation).then_some("distributed_without_peer_invalidation"),
+        ],
+    }
+}
+
+fn scanner_segment_reuse_activation_preflight_for_cycle(
+    dirty_usage_snapshot: &DirtyUsageSnapshot,
+    dirty_usage_producer_evidence: DirtyUsageProducerEvidence,
+    distributed: bool,
+    distributed_segment_invalidation_evidence: Option<DistributedSegmentInvalidationEvidence>,
+    cold_zero_walk_oracle: bool,
+) -> ScannerSegmentReuseActivationPreflight {
+    scanner_segment_reuse_activation_preflight_from_proof(ScannerSegmentReuseActivationProof {
+        production_activation: true,
+        producer_identity_coverage_complete: dirty_usage_producer_evidence.producer_identity_coverage_complete,
+        durable_producer_identity: dirty_usage_producer_evidence.durable_producer_identity,
+        restart_gap_absent: dirty_usage_producer_evidence.restart_gap_absent,
+        generation_window_bound: dirty_usage_snapshot.covers_all_pending
+            && dirty_usage_snapshot.generation != 0
+            && dirty_usage_snapshot.generation != u64::MAX
+            && dirty_usage_producer_evidence.generation_window_bound,
+        overflow_absent: dirty_usage_snapshot.covers_all_pending,
+        cold_zero_walk_oracle,
+        distributed_peer_invalidation: scanner_distributed_segment_invalidation_admitted(
+            distributed,
+            distributed_segment_invalidation_evidence,
+        ),
+    })
+}
+
+fn scanner_segment_reuse_activation_preflight_for_baseline(
+    dirty_usage_snapshot: &DirtyUsageSnapshot,
+    distributed: bool,
+    baseline_proof: ScannerCacheBaselineProof<'_>,
+) -> ScannerSegmentReuseActivationPreflight {
+    let (dirty_usage_producer_evidence, cold_zero_walk_oracle) =
+        scanner_segment_reuse_baseline_producer_evidence(dirty_usage_snapshot, baseline_proof);
+    scanner_segment_reuse_activation_preflight_for_cycle(
+        dirty_usage_snapshot,
+        dirty_usage_producer_evidence,
+        distributed,
+        None,
+        cold_zero_walk_oracle,
+    )
+}
+
+fn scanner_distributed_segment_invalidation_admitted(
+    distributed: bool,
+    evidence: Option<DistributedSegmentInvalidationEvidence>,
+) -> bool {
+    if !distributed {
+        return true;
+    }
+    evidence.is_some_and(|evidence| {
+        evidence.invalidation_domain == crate::segment_invalidation::SegmentInvalidationDomain::DistributedEc
+            && evidence.distributed_ec_invalidation
+            && evidence.same_window_remote_proof
+            && evidence.all_peers_bound_to_generation_window
+            && evidence.dirty_peer_count > 0
+            && evidence.dirty_peer_count <= evidence.peer_count
+    })
+}
+
+fn scanner_durable_segment_invalidation_evidence(
+    dirty_usage_snapshot: &DirtyUsageSnapshot,
+    results: &[DataUsageCache],
+    expected_sources: &HashSet<DataUsageCacheSource>,
+) -> DirtyUsageProducerEvidence {
+    let mut evidence = dirty_usage_producer_evidence(dirty_usage_snapshot);
+    if !evidence.generation_window_bound
+        || !evidence.producer_identity_coverage_complete
+        || !scanner_results_form_complete_snapshot(results, expected_sources)
+    {
+        return evidence;
+    }
+
+    let mut covered_sources = HashSet::with_capacity(expected_sources.len());
+    let all_sets_proved = results.iter().all(|result| {
+        let Some(source) = result.info.source else {
+            return false;
+        };
+        expected_sources.contains(&source)
+            && covered_sources.insert(source)
+            && scanner_segment_invalidation_proof_matches(result.info.segment_invalidation_proof.as_ref(), &evidence)
+    });
+    if all_sets_proved && covered_sources.len() == expected_sources.len() {
+        evidence.durable_producer_identity = true;
+        evidence.restart_gap_absent = true;
+    }
+    evidence
+}
+
+fn scanner_segment_invalidation_proof_matches(
+    proof: Option<&crate::DataUsageSegmentInvalidationProof>,
+    evidence: &DirtyUsageProducerEvidence,
+) -> bool {
+    proof.is_some_and(|proof| {
+        proof.process_epoch == scanner_activity_epoch()
+            && proof.generation_start == evidence.generation_start
+            && proof.generation_end == evidence.generation_end
+            && proof.producer_identity_coverage_complete
+    })
+}
+
+fn scanner_completed_set_segment_invalidation_proof(
+    proof: &Option<crate::DataUsageSegmentInvalidationProof>,
+    cold_zero_walk_reuse_candidate: bool,
+) -> Option<crate::DataUsageSegmentInvalidationProof> {
+    proof.clone().map(|mut proof| {
+        proof.cold_zero_walk_oracle = cold_zero_walk_reuse_candidate;
+        proof
+    })
+}
+
+#[cfg(test)]
 fn scanner_segment_reuse_activated() -> bool {
-    // Production segment reuse stays disabled until a durable mutation-stream
-    // proof satisfies the segment invalidation contract.
-    false
+    scanner_segment_reuse_activation_preflight().scanner_segment_reuse_activated
 }

 pub(crate) fn is_scanner_metadata_corrupt_error(err: &StorageError) -> bool {
@@ -471,6 +745,8 @@ pub struct ScannerBucketScanPlan {
    bucket_failures: ScannerBucketFailureState,
    pending_maintenance_work: Arc<AtomicBool>,
    cache_cycle_floor: Arc<AtomicU64>,
+    cold_zero_walk_reuse_observed: Arc<AtomicBool>,
+    segment_invalidation_proof: Option<crate::DataUsageSegmentInvalidationProof>,
 }

 #[derive(Clone, Default)]
@@ -603,6 +879,25 @@ fn scanner_bucket_scan_status(has_failed: bool, has_partial: bool, has_namespace
    }
 }

+fn scanner_cycle_cold_zero_walk_oracle(
+    scan_scope: &ScannerBucketScanScope,
+    all_buckets: &[BucketInfo],
+    completed_all_sets: bool,
+    scan_scope_matches: bool,
+    bucket_scan_status: ScannerBucketScanStatus,
+    cold_zero_walk_reuse_observed: bool,
+) -> bool {
+    let Some(selected_buckets) = scan_scope.selected_buckets.as_deref() else {
+        return false;
+    };
+    cold_zero_walk_reuse_observed
+        && !selected_buckets.is_empty()
+        && completed_all_sets
+        && scan_scope_matches
+        && bucket_scan_status == ScannerBucketScanStatus::Complete
+        && all_buckets.iter().any(|bucket| !selected_buckets.contains(&bucket.name))
+}
+
 fn classify_nsscanner_cycle(
    completed_all_sets: bool,
    budget_elapsed: bool,
@@ -1056,6 +1351,8 @@ pub(crate) struct ScannerCycleResult {
    observational_snapshot_published: bool,
    dirty_usage_clear: Option<DirtyUsageBuckets>,
    remote_dirty_usage_acknowledgements: Vec<crate::scanner::ScannerDirtyUsageAcknowledgement>,
+    distributed_segment_invalidation_evidence: Option<DistributedSegmentInvalidationEvidence>,
+    segment_reuse_activation_preflight: ScannerSegmentReuseActivationPreflight,
    remote_publication_lease_targets: Vec<(String, String, u64)>,
    failed_dirty_usage: bool,
    pending_maintenance_work: bool,
@@ -1072,6 +1369,8 @@ impl ScannerCycleResult {
            observational_snapshot_published: false,
            dirty_usage_clear,
            remote_dirty_usage_acknowledgements: Vec::new(),
+            distributed_segment_invalidation_evidence: None,
+            segment_reuse_activation_preflight: scanner_segment_reuse_activation_preflight(),
            remote_publication_lease_targets: Vec::new(),
            failed_dirty_usage: false,
            pending_maintenance_work: false,
@@ -1137,6 +1436,21 @@ impl ScannerCycleResult {
        self
    }

+    fn with_distributed_segment_invalidation_evidence(
+        mut self,
+        evidence: Option<DistributedSegmentInvalidationEvidence>,
+    ) -> Self {
+        self.publication_expectation = None;
+        self.distributed_segment_invalidation_evidence = evidence;
+        self
+    }
+
+    fn with_segment_reuse_activation_preflight(mut self, preflight: ScannerSegmentReuseActivationPreflight) -> Self {
+        self.publication_expectation = None;
+        self.segment_reuse_activation_preflight = preflight;
+        self
+    }
+
    pub(crate) fn with_remote_publication_lease_targets(mut self, targets: Vec<(String, String, u64)>) -> Self {
        self.publication_expectation = None;
        self.remote_publication_lease_targets = targets;
@@ -1224,8 +1538,9 @@ pub(crate) use cache::{
 pub use dirty_usage::{
    ScannerDirtyUsageAckError, ScannerDirtyUsageBucket, ScannerDirtyUsageSnapshot, ScannerDirtyUsageState,
    acknowledge_dirty_usage_generation, acknowledge_scoped_dirty_usage, clear_dirty_usage_bucket, record_dirty_usage_bucket,
-    record_dirty_usage_object, record_scanner_maintenance_change, scanner_activity_epoch, scanner_dirty_usage_snapshot,
-    scanner_dirty_usage_state, scanner_maintenance_generation,
+    record_dirty_usage_bucket_from_producer, record_dirty_usage_bucket_from_producers, record_dirty_usage_object,
+    record_dirty_usage_object_from_producer, record_scanner_maintenance_change, scanner_activity_epoch,
+    scanner_dirty_usage_snapshot, scanner_dirty_usage_state, scanner_maintenance_generation,
 };
 #[cfg(test)]
 pub(crate) use dirty_usage::{clear_dirty_usage_buckets_for_tests, dirty_usage_buckets_for_tests};
@@ -464,6 +464,7 @@ pub(super) fn completed_usage_candidate(
                scan_plan_digest: Some(result.info.scan_plan_digest?.0),
                complete: true,
                tombstone: false,
+                segment_invalidation_proof: result.info.segment_invalidation_proof.clone(),
            })
        })
        .collect::<Option<Vec<_>>>()?;
@@ -642,13 +643,14 @@ pub(super) fn observational_data_usage_info(
        let current_snapshot = current.is_some();
        let selected = current.or(lkg);
        if let Some(selected) = selected {
-            let (cycle, epoch, digest, last_update, complete) = if current_snapshot {
+            let (cycle, epoch, digest, last_update, complete, segment_invalidation_proof) = if current_snapshot {
                (
                    Some(selected.info.next_cycle),
                    Some(selected.info.leader_epoch),
                    selected.info.scan_plan_digest.map(|digest| digest.0),
                    selected.info.last_update,
                    true,
+                    selected.info.segment_invalidation_proof.clone(),
                )
            } else {
                (
@@ -657,6 +659,7 @@ pub(super) fn observational_data_usage_info(
                    selected.info.lkg_scan_plan_digest.map(|digest| digest.0),
                    selected.info.lkg_last_update,
                    false,
+                    None,
                )
            };
            set_states.push(DataUsageSnapshotSetState {
@@ -667,6 +670,7 @@ pub(super) fn observational_data_usage_info(
                scan_plan_digest: digest,
                complete,
                tombstone: false,
+                segment_invalidation_proof,
            });
            usable.push((selected, last_update));
        } else {
@@ -678,6 +682,7 @@ pub(super) fn observational_data_usage_info(
                scan_plan_digest: Some(expected_plan_digest.0),
                complete: false,
                tombstone: false,
+                segment_invalidation_proof: None,
            });
        }
    }
@@ -16,12 +16,16 @@ use super::*;

 pub(super) static DIRTY_USAGE_BUCKET_GENERATION: AtomicU64 = AtomicU64::new(0);
 pub(super) static DIRTY_USAGE_BUCKETS: LazyLock<StdMutex<DirtyUsageBuckets>> = LazyLock::new(|| StdMutex::new(HashMap::new()));
-// Lock order when both dirty maps are needed is `DIRTY_USAGE_BUCKETS` followed
-// by `DIRTY_USAGE_BUCKET_SCOPES`. Both are held only for synchronous map
-// updates, so no scanner task can observe a bucket generation without its
-// matching scope.
+// Lock order when dirty usage state is updated is `DIRTY_USAGE_BUCKETS`,
+// `DIRTY_USAGE_BUCKET_SCOPES`, then `DIRTY_USAGE_PRODUCER_IDENTITIES`. All
+// guards are held only for synchronous map updates, so no scanner task can
+// observe a bucket generation without its matching scope and producer evidence.
 pub(super) static DIRTY_USAGE_BUCKET_SCOPES: LazyLock<StdMutex<DirtyUsageBucketScopes>> =
    LazyLock::new(|| StdMutex::new(HashMap::new()));
+// Non-authoritative process-local producer coverage. Any future segment reuse
+// activation must bind this to the exact generation window and durable proof.
+pub(super) static DIRTY_USAGE_PRODUCER_IDENTITIES: LazyLock<StdMutex<DirtyUsageProducerIdentities>> =
+    LazyLock::new(|| StdMutex::new(BTreeMap::new()));
 pub(super) static DIRTY_USAGE_BUCKET_NOTIFY: LazyLock<Notify> = LazyLock::new(Notify::new);
 pub(super) static SCANNER_ACTIVITY_EPOCH: LazyLock<String> = LazyLock::new(|| format!("{:032x}", rand::random::<u128>()));
 pub(super) static SCANNER_MAINTENANCE_GENERATION: AtomicU64 = AtomicU64::new(0);
@@ -54,6 +58,39 @@ pub(super) type DirtyUsageBucketScopes = HashMap<String, DirtyUsageBucketScope>;

 const MAX_DIRTY_USAGE_TOP_LEVEL_ENTRIES_PER_BUCKET: usize = 128;

+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(super) struct DirtyUsageProducerIdentityState {
+    first_generation: u64,
+    last_generation: u64,
+}
+
+pub(super) type DirtyUsageProducerIdentities =
+    BTreeMap<crate::segment_invalidation::SegmentInvalidationProducerIdentity, DirtyUsageProducerIdentityState>;
+
+#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
+pub(super) struct DirtyUsageProducerEvidence {
+    pub(super) producer_identity_coverage_complete: bool,
+    pub(super) durable_producer_identity: bool,
+    pub(super) restart_gap_absent: bool,
+    pub(super) generation_window_bound: bool,
+    pub(super) generation_start: u64,
+    pub(super) generation_end: u64,
+}
+
+impl DirtyUsageProducerEvidence {
+    pub(super) fn segment_invalidation_proof(self) -> Option<crate::DataUsageSegmentInvalidationProof> {
+        (self.generation_window_bound && self.producer_identity_coverage_complete).then(|| {
+            crate::DataUsageSegmentInvalidationProof {
+                process_epoch: scanner_activity_epoch().to_string(),
+                generation_start: self.generation_start,
+                generation_end: self.generation_end,
+                producer_identity_coverage_complete: true,
+                cold_zero_walk_oracle: false,
+            }
+        })
+    }
+}
+
 /// A point-in-time view of the local dirty bucket generations.
 ///
 /// `complete == false` is an all-or-nothing overflow signal: `buckets` is
@@ -153,6 +190,7 @@ fn apply_scoped_dirty_usage_ack(
 #[cfg(test)]
 mod scoped_dirty_usage_tests {
    use super::*;
+    use crate::segment_invalidation::SegmentInvalidationProducerIdentity;

    #[test]
    fn scoped_dirty_usage_preserves_uncovered_newer_and_replayed_generations() {
@@ -213,6 +251,39 @@ mod scoped_dirty_usage_tests {
            assert_eq!(scopes, original_scopes);
        }
    }
+
+    #[test]
+    fn dirty_usage_tracks_known_segment_producer_identities_without_authorizing_unknown_sources() {
+        clear_dirty_usage_buckets_for_tests();
+        record_dirty_usage_object_from_producer("photos", "hot/object", SegmentInvalidationProducerIdentity::PutObject);
+        record_dirty_usage_object_from_producer("photos", "archive/object", SegmentInvalidationProducerIdentity::DeleteObject);
+        record_dirty_usage_bucket_from_producer("photos", SegmentInvalidationProducerIdentity::Unknown);
+        record_dirty_usage_bucket_from_producers(
+            "photos",
+            [
+                SegmentInvalidationProducerIdentity::DeleteMarker,
+                SegmentInvalidationProducerIdentity::AbortMultipartUpload,
+            ],
+        );
+
+        assert_eq!(
+            dirty_usage_producer_identities_for_tests(),
+            BTreeSet::from([
+                SegmentInvalidationProducerIdentity::PutObject,
+                SegmentInvalidationProducerIdentity::DeleteObject,
+                SegmentInvalidationProducerIdentity::DeleteMarker,
+                SegmentInvalidationProducerIdentity::AbortMultipartUpload
+            ])
+        );
+        assert_eq!(
+            dirty_usage_bucket_scopes_for_tests().get("photos"),
+            Some(&DirtyUsageBucketScope::WholeBucket),
+            "an unknown producer keeps the bucket dirty but must not count as producer coverage"
+        );
+
+        clear_dirty_usage_buckets_for_tests();
+        assert!(dirty_usage_producer_identities_for_tests().is_empty());
+    }
 }

 pub(super) fn dirty_usage_buckets() -> MutexGuard<'static, DirtyUsageBuckets> {
@@ -225,6 +296,12 @@ fn dirty_usage_bucket_scopes() -> MutexGuard<'static, DirtyUsageBucketScopes> {
        .unwrap_or_else(|poisoned| poisoned.into_inner())
 }

+fn dirty_usage_producer_identities() -> MutexGuard<'static, DirtyUsageProducerIdentities> {
+    DIRTY_USAGE_PRODUCER_IDENTITIES
+        .lock()
+        .unwrap_or_else(|poisoned| poisoned.into_inner())
+}
+
 pub(super) fn usize_to_u64_saturated(value: usize) -> u64 {
    u64::try_from(value).unwrap_or(u64::MAX)
 }
@@ -240,12 +317,43 @@ pub fn record_dirty_usage_bucket(bucket: &str) {
        return;
    }

+    record_dirty_usage_bucket_inner(bucket, std::iter::empty());
+}
+
+pub fn record_dirty_usage_bucket_from_producer(
+    bucket: &str,
+    producer: crate::segment_invalidation::SegmentInvalidationProducerIdentity,
+) {
+    if bucket.is_empty() {
+        return;
+    }
+
+    record_dirty_usage_bucket_inner(bucket, [producer]);
+}
+
+pub fn record_dirty_usage_bucket_from_producers<I>(bucket: &str, producers: I)
+where
+    I: IntoIterator<Item = crate::segment_invalidation::SegmentInvalidationProducerIdentity>,
+{
+    if bucket.is_empty() {
+        return;
+    }
+
+    record_dirty_usage_bucket_inner(bucket, producers);
+}
+
+fn record_dirty_usage_bucket_inner<I>(bucket: &str, producers: I)
+where
+    I: IntoIterator<Item = crate::segment_invalidation::SegmentInvalidationProducerIdentity>,
+{
    let pending_buckets = {
        let mut dirty_buckets = dirty_usage_buckets();
        let mut dirty_scopes = dirty_usage_bucket_scopes();
+        let mut producer_identities = dirty_usage_producer_identities();
        let generation = advance_generation(&DIRTY_USAGE_BUCKET_GENERATION);
        dirty_buckets.insert(bucket.to_string(), generation);
        dirty_scopes.insert(bucket.to_string(), DirtyUsageBucketScope::WholeBucket);
+        record_segment_invalidation_producer_identities_for_generation(&mut producer_identities, generation, producers);
        dirty_buckets.len()
    };
    global_metrics().record_scanner_dirty_usage_pending(usize_to_u64_saturated(pending_buckets));
@@ -263,8 +371,28 @@ pub fn record_dirty_usage_bucket(bucket: &str) {
 /// local: after restart or any unverified distributed path the scanner falls
 /// back to its ordinary bucket scan.
 pub fn record_dirty_usage_object(bucket: &str, object: &str) {
+    record_dirty_usage_object_inner(bucket, object, std::iter::empty());
+}
+
+pub fn record_dirty_usage_object_from_producer(
+    bucket: &str,
+    object: &str,
+    producer: crate::segment_invalidation::SegmentInvalidationProducerIdentity,
+) {
+    if bucket.is_empty() {
+        return;
+    }
+
+    record_dirty_usage_object_inner(bucket, object, [producer]);
+}
+
+fn record_dirty_usage_object_inner<I>(bucket: &str, object: &str, producers: I)
+where
+    I: IntoIterator<Item = crate::segment_invalidation::SegmentInvalidationProducerIdentity>,
+{
+    let producers = producers.into_iter().collect::<Vec<_>>();
    let Some(top_level_entry) = dirty_usage_top_level_entry(object) else {
-        record_dirty_usage_bucket(bucket);
+        record_dirty_usage_bucket_inner(bucket, producers);
        return;
    };
    if bucket.is_empty() {
@@ -274,6 +402,7 @@ pub fn record_dirty_usage_object(bucket: &str, object: &str) {
    let pending_buckets = {
        let mut dirty_buckets = dirty_usage_buckets();
        let mut dirty_scopes = dirty_usage_bucket_scopes();
+        let mut producer_identities = dirty_usage_producer_identities();
        let generation = advance_generation(&DIRTY_USAGE_BUCKET_GENERATION);
        dirty_buckets.insert(bucket.to_string(), generation);
        let scope = dirty_scopes
@@ -289,6 +418,7 @@ pub fn record_dirty_usage_object(bucket: &str, object: &str) {
        if overflowed {
            *scope = DirtyUsageBucketScope::WholeBucket;
        }
+        record_segment_invalidation_producer_identities_for_generation(&mut producer_identities, generation, producers);
        dirty_buckets.len()
    };
    global_metrics().record_scanner_dirty_usage_pending(usize_to_u64_saturated(pending_buckets));
@@ -296,6 +426,31 @@ pub fn record_dirty_usage_object(bucket: &str, object: &str) {
    DIRTY_USAGE_BUCKET_NOTIFY.notify_one();
 }

+fn record_segment_invalidation_producer_identities_for_generation<I>(
+    identities: &mut DirtyUsageProducerIdentities,
+    generation: u64,
+    producers: I,
+) where
+    I: IntoIterator<Item = crate::segment_invalidation::SegmentInvalidationProducerIdentity>,
+{
+    for producer in producers {
+        if producer.producer().is_some() {
+            identities
+                .entry(producer)
+                .and_modify(|state| state.last_generation = state.last_generation.max(generation))
+                .or_insert(DirtyUsageProducerIdentityState {
+                    first_generation: generation,
+                    last_generation: generation,
+                });
+        }
+    }
+}
+
+#[cfg(test)]
+fn dirty_usage_producer_identities_for_tests() -> BTreeSet<crate::segment_invalidation::SegmentInvalidationProducerIdentity> {
+    dirty_usage_producer_identities().keys().copied().collect()
+}
+
 fn dirty_usage_top_level_entry(object: &str) -> Option<String> {
    let (top_level_entry, _) = object.split_once('/').unwrap_or((object, ""));
    (!top_level_entry.is_empty()
@@ -568,6 +723,31 @@ pub(super) fn dirty_usage_snapshot_status(snapshot: &DirtyUsageSnapshot) -> Dirt
    }
 }

+pub(super) fn dirty_usage_producer_evidence(snapshot: &DirtyUsageSnapshot) -> DirtyUsageProducerEvidence {
+    let generation_window_bound = dirty_usage_snapshot_status(snapshot) == DirtyUsageSnapshotStatus::Current
+        && snapshot.generation != 0
+        && snapshot.generation != u64::MAX;
+    let identities = dirty_usage_producer_identities()
+        .iter()
+        .filter(|(_, state)| state.first_generation <= snapshot.generation)
+        .map(|(identity, _)| *identity)
+        .collect::<BTreeSet<_>>();
+    let producer_identity_coverage_complete =
+        generation_window_bound && crate::segment_invalidation::complete_segment_invalidation_producers(identities).is_ok();
+
+    DirtyUsageProducerEvidence {
+        producer_identity_coverage_complete,
+        // The current producer journal is still process-local. Keep the
+        // durable/restart gates closed until the mutation evidence is persisted
+        // and replayable across scanner restarts.
+        durable_producer_identity: false,
+        restart_gap_absent: false,
+        generation_window_bound,
+        generation_start: snapshot.generation,
+        generation_end: snapshot.generation,
+    }
+}
+
 #[cfg(test)]
 pub(super) fn dirty_usage_bucket_count() -> usize {
    dirty_usage_buckets().len()
@@ -577,6 +757,7 @@ pub(super) fn dirty_usage_bucket_count() -> usize {
 pub(crate) fn clear_dirty_usage_buckets_for_tests() {
    dirty_usage_buckets().clear();
    dirty_usage_bucket_scopes().clear();
+    dirty_usage_producer_identities().clear();
 }

 #[cfg(test)]
@@ -172,6 +172,8 @@ impl ScannerIOCache for SetDisks {
            bucket_failures,
            pending_maintenance_work,
            cache_cycle_floor,
+            cold_zero_walk_reuse_observed,
+            segment_invalidation_proof,
        } = scan_plan;
        let scan_plan_digest = scanner_bucket_work_digest(scan_plan_digest, scan_mode, requires_full_scan);
        let bucket_work_digest = scanner_bucket_work_digest(bucket_coverage_digest, scan_mode, requires_full_scan);
@@ -219,6 +221,9 @@ impl ScannerIOCache for SetDisks {
            },
            current_bucket_incarnations.as_ref(),
        );
+        let cold_zero_walk_reuse_candidate = scoped_scan.as_ref().is_some_and(|prepared| {
+            old_cache.info.next_cycle < want_cycle && !prepared.buckets.is_empty() && prepared.buckets.len() < all_buckets.len()
+        });
        let mut scoped_cache = scoped_scan.map(|mut prepared| {
            buckets = prepared.buckets;
            prepared.cache.info.scan_coverage_digest = Some(bucket_coverage_digest);
@@ -226,6 +231,8 @@ impl ScannerIOCache for SetDisks {
        });
        if buckets.is_empty() {
            let now = SystemTime::now();
+            let completed_segment_invalidation_proof =
+                scanner_completed_set_segment_invalidation_proof(&segment_invalidation_proof, cold_zero_walk_reuse_candidate);
            let mut cache = match scoped_cache.take() {
                Some(cache) => cache,
                None => {
@@ -239,6 +246,7 @@ impl ScannerIOCache for SetDisks {
                            scan_plan_digest: Some(scan_plan_digest),
                            scan_coverage_digest: Some(bucket_coverage_digest),
                            cache_key_format: DATA_USAGE_CACHE_KEY_FORMAT,
+                            segment_invalidation_proof: completed_segment_invalidation_proof.clone(),
                            scan_bucket_incarnations: current_bucket_incarnations.clone().unwrap_or_default(),
                            ..Default::default()
                        },
@@ -254,6 +262,7 @@ impl ScannerIOCache for SetDisks {
            cache.info.last_update = Some(now);
            cache.info.snapshot_complete = true;
            cache.info.scan_execution_digest = Some(execution_digest);
+            cache.info.segment_invalidation_proof = completed_segment_invalidation_proof;
            cache.info.lkg_snapshot_complete = false;
            cache.info.lkg_next_cycle = None;
            cache.info.lkg_last_update = None;
@@ -535,6 +544,7 @@ impl ScannerIOCache for SetDisks {
                    lkg_last_update: old_cache.info.lkg_last_update,
                    lkg_leader_epoch: old_cache.info.lkg_leader_epoch,
                    lkg_scan_plan_digest: old_cache.info.lkg_scan_plan_digest,
+                    segment_invalidation_proof: None,
                    scan_bucket_incarnations: current_bucket_incarnations.clone().unwrap_or_default(),
                    ..Default::default()
                },
@@ -1455,17 +1465,23 @@ impl ScannerIOCache for SetDisks {

        let completed_count = completed_bucket_count.load(Ordering::Relaxed);
        if should_publish_completed_snapshot(completed_count, buckets.len(), budget.budget_elapsed(), ctx.is_cancelled()) {
+            let completed_segment_invalidation_proof =
+                scanner_completed_set_segment_invalidation_proof(&segment_invalidation_proof, cold_zero_walk_reuse_candidate);
            let cache_snapshot = {
                let mut cache = cache_mutex.lock().await;
                cache.info.next_cycle = want_cycle;
                cache.info.last_update.get_or_insert_with(SystemTime::now);
                cache.info.snapshot_complete = true;
                cache.info.scan_execution_digest = Some(execution_digest);
+                cache.info.segment_invalidation_proof = completed_segment_invalidation_proof;
                cache.info.lkg_snapshot_complete = false;
                cache.info.lkg_next_cycle = None;
                cache.info.lkg_last_update = None;
                cache.info.lkg_leader_epoch = None;
                cache.info.lkg_scan_plan_digest = None;
+                if cold_zero_walk_reuse_candidate {
+                    cold_zero_walk_reuse_observed.store(true, Ordering::Release);
+                }
                cache.clone()
            };
            let _ = persist_and_publish_cache_snapshot(
@@ -1486,6 +1502,7 @@ impl ScannerIOCache for SetDisks {
            incomplete_scope.info.tier_registry_generation = Some(tier_registry_generation);
            incomplete_scope.info.source = Some(source);
            incomplete_scope.info.snapshot_complete = false;
+            incomplete_scope.info.segment_invalidation_proof = None;
            incomplete_scope.info.scan_plan_digest = Some(scan_plan_digest);
            incomplete_scope.info.cache_key_format = DATA_USAGE_CACHE_KEY_FORMAT;
            if let Err(e) = updates.send(incomplete_scope).await {
@@ -122,6 +122,7 @@ where
    let default_result = |scope: ScannerBucketScanScope| ScannerBucketScopeResolutionResult {
        scope,
        remote_dirty_usage_acknowledgements: Vec::new(),
+        distributed_segment_invalidation_evidence: None,
    };
    if resolution.requires_full_scan {
        return default_result(ScannerBucketScanScope::default());
@@ -240,12 +241,17 @@ where
        return remote_resolution;
    }

+    let segment_reuse_activation_preflight = scanner_segment_reuse_activation_preflight_for_baseline(
+        resolution.dirty_usage_snapshot,
+        distributed,
+        resolution.baseline_proof,
+    );
    default_result(scoped_scan_scope_from_dirty_buckets(
        resolution.requested_scope,
        dirty_buckets,
        (!distributed).then_some(resolution.dirty_usage_snapshot.scopes.as_ref()),
        true,
-        scanner_segment_reuse_activated(),
+        segment_reuse_activation_preflight.scanner_segment_reuse_activated,
        resolution.all_buckets,
        resolution.baseline_proof,
    ))
@@ -408,7 +414,9 @@ where
    )
    .await;
    let remote_dirty_usage_acknowledgements = scope_resolution.remote_dirty_usage_acknowledgements;
+    let distributed_segment_invalidation_evidence = scope_resolution.distributed_segment_invalidation_evidence;
    let scan_scope = scope_resolution.scope;
+    let segment_invalidation_proof = dirty_usage_producer_evidence(&dirty_usage_snapshot).segment_invalidation_proof();
    #[cfg(test)]
    if let Some(observer) = resolved_scope_observer {
        let _ = observer.send(scan_scope.clone());
@@ -465,12 +473,20 @@ where
        } else {
            Vec::new()
        };
+        let segment_reuse_activation_preflight = scanner_segment_reuse_activation_preflight_for_cycle(
+            &dirty_usage_snapshot,
+            dirty_usage_producer_evidence(&dirty_usage_snapshot),
+            distributed,
+            None,
+            false,
+        );
        return Ok(ScannerCycleResult::new(status, dirty_usage_clear)
            .with_publication_epoch(publication_epoch)
            .with_activity_digest(activity_digest)
            .with_observational_snapshot_published(observational_snapshot_published)
            .with_remote_publication_lease_targets(remote_publication_lease_targets)
            .with_remote_dirty_usage_acknowledgements(remote_dirty_usage_acknowledgements)
+            .with_segment_reuse_activation_preflight(segment_reuse_activation_preflight)
            .with_publication_expectation(publication_expectation));
    }

@@ -495,6 +511,7 @@ where
    );
    let bucket_failures = ScannerBucketFailureState::default();
    let pending_maintenance_work = Arc::new(AtomicBool::new(false));
+    let cold_zero_walk_reuse_observed = Arc::new(AtomicBool::new(false));
    record_set_scan_concurrency_limit(set_scan_limit);
    debug!(
        target: "rustfs::scanner::io",
@@ -588,6 +605,8 @@ where
            bucket_failures: bucket_failures.clone(),
            pending_maintenance_work: pending_maintenance_work.clone(),
            cache_cycle_floor: cache_cycle_floor.clone(),
+            cold_zero_walk_reuse_observed: cold_zero_walk_reuse_observed.clone(),
+            segment_invalidation_proof: segment_invalidation_proof.clone(),
        };
        // Spawn task to run the scanner
        let scanner_fut = tokio::spawn(async move {
@@ -691,6 +710,21 @@ where
        scan_scope_matches && !partial_buckets.is_empty(),
        scan_scope_matches && !namespace_not_found_buckets.is_empty(),
    );
+    let cold_zero_walk_oracle = scanner_cycle_cold_zero_walk_oracle(
+        &scan_scope,
+        &all_buckets,
+        completed_all_sets,
+        scan_scope_matches,
+        bucket_scan_status,
+        cold_zero_walk_reuse_observed.load(Ordering::Acquire),
+    );
+    let segment_reuse_activation_preflight = scanner_segment_reuse_activation_preflight_for_cycle(
+        &dirty_usage_snapshot,
+        scanner_durable_segment_invalidation_evidence(&dirty_usage_snapshot, &results, &expected_sources),
+        distributed,
+        distributed_segment_invalidation_evidence,
+        cold_zero_walk_oracle,
+    );
    let pending_maintenance_work = pending_maintenance_work_for_cycle(&pending_maintenance_work, &results);
    let observed_cycle_floor = cache_cycle_floor.load(Ordering::Acquire);
    let required_cycle_floor = (observed_cycle_floor > want_cycle).then_some(observed_cycle_floor);
@@ -783,6 +817,8 @@ where
        .with_observational_snapshot_published(observational_snapshot_published)
        .with_remote_publication_lease_targets(remote_publication_lease_targets)
        .with_remote_dirty_usage_acknowledgements(remote_dirty_usage_acknowledgements)
+        .with_distributed_segment_invalidation_evidence(distributed_segment_invalidation_evidence)
+        .with_segment_reuse_activation_preflight(segment_reuse_activation_preflight)
        .with_failed_dirty_usage(!failed_buckets.is_empty())
        .with_pending_maintenance_work(pending_maintenance_work)
        .with_required_cycle_floor(required_cycle_floor)
@@ -13,7 +13,7 @@
 // limitations under the License.

 use super::*;
-use crate::data_usage_define::{UNKNOWN_TIER, UnknownTierStats, hash_path};
+use crate::data_usage_define::{DataUsageSegmentInvalidationProof, UNKNOWN_TIER, UnknownTierStats, hash_path};
 use rustfs_data_usage::{ReplicationAllStats, ReplicationTargetUsage, TierAccountingProof};

 const TEST_PLAN_DIGEST: DataUsageScanPlanDigest = DataUsageScanPlanDigest([7; 32]);
@@ -176,6 +176,27 @@ fn completed_data_usage_info_rejects_duplicate_bucket_inventory() {
    assert!(completed_data_usage_info_for_test(&[set], &buckets, false, false).is_none());
 }

+#[test]
+fn completed_data_usage_info_carries_segment_invalidation_proof_to_set_state() {
+    let source = DataUsageCacheSource::new(0, 0);
+    let proof = DataUsageSegmentInvalidationProof {
+        process_epoch: "scanner-process".to_string(),
+        generation_start: 5,
+        generation_end: 8,
+        producer_identity_coverage_complete: true,
+        cold_zero_walk_oracle: true,
+    };
+    let mut set = completed_root_cache("bucket", 2, 10, source);
+    set.info.segment_invalidation_proof = Some(proof.clone());
+
+    let (usage, _) =
+        completed_usage_for_scope(&[set], &HashSet::from([source]), &["bucket".to_string()], &[], true, false, false)
+            .expect("complete set should publish root usage");
+
+    assert_eq!(usage.usage_snapshot_set_states.len(), 1);
+    assert_eq!(usage.usage_snapshot_set_states[0].segment_invalidation_proof, Some(proof));
+}
+
 #[test]
 fn completed_data_usage_info_rejects_extra_or_detached_bucket_data() {
    let buckets = vec!["bucket".to_string()];
@@ -350,6 +371,7 @@ fn set_membership_add_remove_uses_generation_and_tombstone() {
        scan_plan_digest: Some(TEST_PLAN_DIGEST.0),
        complete: false,
        tombstone: true,
+        segment_invalidation_proof: None,
    };
    let encoded = serde_json::to_vec(&state).expect("set state should serialize");
    let decoded: DataUsageSnapshotSetState = serde_json::from_slice(&encoded).expect("set state should deserialize");
@@ -371,6 +393,7 @@ fn set_membership_add_remove_uses_generation_and_tombstone() {
                scan_plan_digest: Some(TEST_PLAN_DIGEST.0),
                complete: true,
                tombstone: false,
+                segment_invalidation_proof: None,
            },
            state,
        ],
@@ -85,6 +85,351 @@ fn scanner_activity_preflight_defers_a_temporarily_offline_peer() {
    }
 }

+#[test]
+fn scanner_segment_reuse_activation_preflight_reports_release_gate_inputs() {
+    let preflight = scanner_segment_reuse_activation_preflight();
+
+    assert!(preflight.production_activation);
+    assert!(!preflight.scanner_segment_reuse_activated);
+    assert!(!scanner_segment_reuse_activated());
+    assert_eq!(preflight.proof_inputs, SCANNER_SEGMENT_ACTIVATION_PROOF_INPUTS);
+    assert_eq!(preflight.fail_closed_checks, SCANNER_SEGMENT_ACTIVATION_FAIL_CLOSED_CHECKS);
+    assert_eq!(
+        preflight.fail_closed_blockers().collect::<Vec<_>>(),
+        SCANNER_SEGMENT_ACTIVATION_FAIL_CLOSED_CHECKS
+    );
+}
+
+#[test]
+fn scanner_segment_reuse_activation_requires_every_preflight_proof() {
+    let complete_proof = ScannerSegmentReuseActivationProof {
+        production_activation: true,
+        producer_identity_coverage_complete: true,
+        durable_producer_identity: true,
+        restart_gap_absent: true,
+        generation_window_bound: true,
+        overflow_absent: true,
+        cold_zero_walk_oracle: true,
+        distributed_peer_invalidation: true,
+    };
+
+    let mut production_disabled = complete_proof;
+    production_disabled.production_activation = false;
+    let preflight = scanner_segment_reuse_activation_preflight_from_proof(production_disabled);
+    assert!(!preflight.production_activation);
+    assert!(!preflight.scanner_segment_reuse_activated);
+    assert_eq!(preflight.fail_closed_blockers().collect::<Vec<_>>(), Vec::<&str>::new());
+
+    let preflight = scanner_segment_reuse_activation_preflight_from_proof(complete_proof);
+    assert!(preflight.production_activation);
+    assert!(preflight.scanner_segment_reuse_activated);
+    assert_eq!(preflight.fail_closed_blockers().collect::<Vec<_>>(), Vec::<&str>::new());
+
+    let mut missing_identity = complete_proof;
+    missing_identity.producer_identity_coverage_complete = false;
+    assert_segment_reuse_activation_blocked_by(missing_identity, "missing_producer_identity");
+
+    let mut non_durable_identity = complete_proof;
+    non_durable_identity.durable_producer_identity = false;
+    assert_segment_reuse_activation_blocked_by(non_durable_identity, "missing_producer_identity");
+
+    let mut restart_gap = complete_proof;
+    restart_gap.restart_gap_absent = false;
+    assert_segment_reuse_activation_blocked_by(restart_gap, "restart_gap");
+
+    let mut generation_gap = complete_proof;
+    generation_gap.generation_window_bound = false;
+    assert_segment_reuse_activation_blocked_by(generation_gap, "generation_gap");
+
+    let mut overflow = complete_proof;
+    overflow.overflow_absent = false;
+    assert_segment_reuse_activation_blocked_by(overflow, "overflow");
+
+    let mut missing_cold_oracle = complete_proof;
+    missing_cold_oracle.cold_zero_walk_oracle = false;
+    assert_segment_reuse_activation_blocked_by(missing_cold_oracle, "missing_cold_zero_walk_oracle");
+
+    let mut missing_distributed_invalidation = complete_proof;
+    missing_distributed_invalidation.distributed_peer_invalidation = false;
+    assert_segment_reuse_activation_blocked_by(missing_distributed_invalidation, "distributed_without_peer_invalidation");
+}
+
+#[test]
+fn scanner_segment_reuse_activation_preflight_for_cycle_reports_cycle_inputs_without_activation() {
+    let dirty_usage_snapshot = DirtyUsageSnapshot {
+        buckets: Arc::new(DirtyUsageBuckets::from([("photos".to_string(), 7)])),
+        scopes: Arc::new(DirtyUsageBucketScopes::default()),
+        generation: 7,
+        covers_all_pending: true,
+    };
+    let distributed_evidence = DistributedSegmentInvalidationEvidence {
+        invalidation_domain: crate::segment_invalidation::SegmentInvalidationDomain::DistributedEc,
+        distributed_ec_invalidation: true,
+        peer_count: 2,
+        dirty_peer_count: 1,
+        same_window_remote_proof: true,
+        all_peers_bound_to_generation_window: true,
+    };
+
+    let preflight = scanner_segment_reuse_activation_preflight_for_cycle(
+        &dirty_usage_snapshot,
+        complete_process_local_producer_evidence(),
+        true,
+        Some(distributed_evidence),
+        true,
+    );
+
+    assert!(preflight.production_activation);
+    assert!(!preflight.scanner_segment_reuse_activated);
+    assert_eq!(
+        preflight.fail_closed_blockers().collect::<Vec<_>>(),
+        vec!["missing_producer_identity", "restart_gap"]
+    );
+}
+
+#[test]
+fn scanner_segment_reuse_activation_preflight_for_cycle_blocks_unbounded_inputs() {
+    let dirty_usage_snapshot = DirtyUsageSnapshot {
+        buckets: Arc::new(DirtyUsageBuckets::default()),
+        scopes: Arc::new(DirtyUsageBucketScopes::default()),
+        generation: u64::MAX,
+        covers_all_pending: false,
+    };
+
+    let preflight = scanner_segment_reuse_activation_preflight_for_cycle(
+        &dirty_usage_snapshot,
+        DirtyUsageProducerEvidence::default(),
+        true,
+        None,
+        false,
+    );
+
+    assert!(preflight.production_activation);
+    assert!(!preflight.scanner_segment_reuse_activated);
+    assert_eq!(
+        preflight.fail_closed_blockers().collect::<Vec<_>>(),
+        SCANNER_SEGMENT_ACTIVATION_FAIL_CLOSED_CHECKS
+    );
+}
+
+#[test]
+fn scanner_segment_reuse_activation_preflight_for_cycle_skips_distributed_blocker_for_local_scan() {
+    let dirty_usage_snapshot = DirtyUsageSnapshot {
+        buckets: Arc::new(DirtyUsageBuckets::from([("photos".to_string(), 7)])),
+        scopes: Arc::new(DirtyUsageBucketScopes::default()),
+        generation: 7,
+        covers_all_pending: true,
+    };
+
+    let preflight = scanner_segment_reuse_activation_preflight_for_cycle(
+        &dirty_usage_snapshot,
+        complete_process_local_producer_evidence(),
+        false,
+        None,
+        true,
+    );
+
+    assert!(preflight.production_activation);
+    assert!(!preflight.scanner_segment_reuse_activated);
+    assert_eq!(
+        preflight.fail_closed_blockers().collect::<Vec<_>>(),
+        vec!["missing_producer_identity", "restart_gap"]
+    );
+}
+
+#[test]
+#[serial]
+fn scanner_durable_segment_invalidation_evidence_requires_matching_complete_set_proofs() {
+    use crate::segment_invalidation::SegmentInvalidationProducerIdentity;
+
+    clear_dirty_usage_buckets_for_tests();
+    record_dirty_usage_bucket_from_producers("photos", SegmentInvalidationProducerIdentity::REQUIRED_PRODUCTION);
+    let dirty_usage_snapshot = snapshot_dirty_usage_buckets(&[bucket_info("photos")], dirty_usage_generation());
+    let process_proof = dirty_usage_producer_evidence(&dirty_usage_snapshot)
+        .segment_invalidation_proof()
+        .expect("complete process-local producer coverage should produce proof metadata");
+    let expected_sources = HashSet::from([DataUsageCacheSource::new(0, 0), DataUsageCacheSource::new(0, 1)]);
+    let results = vec![
+        complete_set_cache_with_segment_proof(DataUsageCacheSource::new(0, 0), process_proof.clone()),
+        complete_set_cache_with_segment_proof(DataUsageCacheSource::new(0, 1), process_proof),
+    ];
+
+    let durable_evidence = scanner_durable_segment_invalidation_evidence(&dirty_usage_snapshot, &results, &expected_sources);
+
+    assert!(durable_evidence.producer_identity_coverage_complete);
+    assert!(durable_evidence.durable_producer_identity);
+    assert!(durable_evidence.restart_gap_absent);
+
+    let mut stale_epoch = results.clone();
+    stale_epoch[0]
+        .info
+        .segment_invalidation_proof
+        .as_mut()
+        .expect("proof fixture should exist")
+        .process_epoch = "stale-process".to_string();
+    let stale_evidence = scanner_durable_segment_invalidation_evidence(&dirty_usage_snapshot, &stale_epoch, &expected_sources);
+    assert!(stale_evidence.producer_identity_coverage_complete);
+    assert!(!stale_evidence.durable_producer_identity);
+    assert!(!stale_evidence.restart_gap_absent);
+
+    record_dirty_usage_bucket("videos");
+    let changed_evidence = scanner_durable_segment_invalidation_evidence(&dirty_usage_snapshot, &results, &expected_sources);
+    assert!(!changed_evidence.producer_identity_coverage_complete);
+    assert!(!changed_evidence.durable_producer_identity);
+    assert!(!changed_evidence.restart_gap_absent);
+    clear_dirty_usage_buckets_for_tests();
+}
+
+#[test]
+#[serial]
+fn scanner_segment_reuse_activation_replays_cold_durable_baseline() {
+    use crate::segment_invalidation::SegmentInvalidationProducerIdentity;
+
+    clear_dirty_usage_buckets_for_tests();
+    for producer in SegmentInvalidationProducerIdentity::REQUIRED_PRODUCTION {
+        record_dirty_usage_object_from_producer("photos", "2026/object", producer);
+    }
+    let dirty_usage_snapshot =
+        snapshot_dirty_usage_buckets(&[bucket_info("photos"), bucket_info("archive")], dirty_usage_generation());
+    let mut segment_proof = dirty_usage_producer_evidence(&dirty_usage_snapshot)
+        .segment_invalidation_proof()
+        .expect("complete process-local producer coverage should produce proof metadata");
+    segment_proof.cold_zero_walk_oracle = true;
+    let scan_plan_digest = DataUsageScanPlanDigest([6; 32]);
+    let expected_sources = HashSet::from([DataUsageCacheSource::new(0, 0), DataUsageCacheSource::new(0, 1)]);
+    let baseline = DataUsageInfo {
+        last_update: Some(SystemTime::UNIX_EPOCH + Duration::from_secs(10)),
+        scanner_cycle: Some(7),
+        scanner_epoch: Some(11),
+        buckets_count: 2,
+        buckets_usage: HashMap::from([
+            ("photos".to_string(), Default::default()),
+            ("archive".to_string(), Default::default()),
+        ]),
+        usage_snapshot_complete: true,
+        usage_snapshot_converged: Some(true),
+        usage_snapshot_set_states: expected_sources
+            .iter()
+            .map(|source| DataUsageSnapshotSetState {
+                pool_index: u64::try_from(source.pool_index).expect("test pool index should fit"),
+                set_index: u64::try_from(source.set_index).expect("test set index should fit"),
+                scanner_cycle: Some(7),
+                scanner_epoch: Some(11),
+                scan_plan_digest: Some(scan_plan_digest.0),
+                complete: true,
+                tombstone: false,
+                segment_invalidation_proof: Some(segment_proof.clone()),
+            })
+            .collect(),
+        ..Default::default()
+    };
+    let baseline = Bytes::from(serde_json::to_vec(&baseline).expect("baseline should encode"));
+
+    let preflight = scanner_segment_reuse_activation_preflight_for_baseline(
+        &dirty_usage_snapshot,
+        false,
+        ScannerCacheBaselineProof {
+            authoritative_data: Some(&baseline),
+            observed_candidate_data: None,
+            expected_sources: &expected_sources,
+            leader_epoch: 11,
+            want_cycle: 8,
+            scan_plan_digest,
+        },
+    );
+
+    assert!(preflight.production_activation);
+    assert!(preflight.scanner_segment_reuse_activated);
+    assert_eq!(preflight.fail_closed_blockers().collect::<Vec<_>>(), Vec::<&str>::new());
+
+    let mut missing_cold_baseline =
+        serde_json::from_slice::<DataUsageInfo>(&baseline).expect("baseline should decode for negative case");
+    missing_cold_baseline.usage_snapshot_set_states[0]
+        .segment_invalidation_proof
+        .as_mut()
+        .expect("proof should exist")
+        .cold_zero_walk_oracle = false;
+    let missing_cold_baseline = Bytes::from(serde_json::to_vec(&missing_cold_baseline).expect("negative baseline should encode"));
+    let preflight = scanner_segment_reuse_activation_preflight_for_baseline(
+        &dirty_usage_snapshot,
+        false,
+        ScannerCacheBaselineProof {
+            authoritative_data: Some(&missing_cold_baseline),
+            observed_candidate_data: None,
+            expected_sources: &expected_sources,
+            leader_epoch: 11,
+            want_cycle: 8,
+            scan_plan_digest,
+        },
+    );
+
+    assert!(preflight.production_activation);
+    assert!(!preflight.scanner_segment_reuse_activated);
+    assert_eq!(
+        preflight.fail_closed_blockers().collect::<Vec<_>>(),
+        vec!["missing_producer_identity", "restart_gap", "missing_cold_zero_walk_oracle"]
+    );
+    clear_dirty_usage_buckets_for_tests();
+}
+
+#[test]
+fn scanner_cycle_result_returns_segment_reuse_activation_preflight() {
+    let proof = ScannerSegmentReuseActivationProof {
+        production_activation: true,
+        producer_identity_coverage_complete: true,
+        durable_producer_identity: true,
+        restart_gap_absent: true,
+        generation_window_bound: true,
+        overflow_absent: true,
+        cold_zero_walk_oracle: true,
+        distributed_peer_invalidation: true,
+    };
+    let preflight = scanner_segment_reuse_activation_preflight_from_proof(proof);
+
+    let result = ScannerCycleResult::new(ScannerCycleStatus::Complete, None).with_segment_reuse_activation_preflight(preflight);
+
+    assert_eq!(result.segment_reuse_activation_preflight, preflight);
+}
+
+fn assert_segment_reuse_activation_blocked_by(proof: ScannerSegmentReuseActivationProof, blocker: &'static str) {
+    let preflight = scanner_segment_reuse_activation_preflight_from_proof(proof);
+
+    assert!(preflight.production_activation);
+    assert!(!preflight.scanner_segment_reuse_activated);
+    assert_eq!(preflight.fail_closed_blockers().collect::<Vec<_>>(), vec![blocker]);
+}
+
+fn complete_process_local_producer_evidence() -> DirtyUsageProducerEvidence {
+    DirtyUsageProducerEvidence {
+        producer_identity_coverage_complete: true,
+        durable_producer_identity: false,
+        restart_gap_absent: false,
+        generation_window_bound: true,
+        generation_start: 7,
+        generation_end: 7,
+    }
+}
+
+fn complete_set_cache_with_segment_proof(
+    source: DataUsageCacheSource,
+    proof: crate::DataUsageSegmentInvalidationProof,
+) -> DataUsageCache {
+    DataUsageCache {
+        info: DataUsageCacheInfo {
+            name: DATA_USAGE_ROOT.to_string(),
+            next_cycle: 7,
+            last_update: Some(SystemTime::UNIX_EPOCH),
+            leader_epoch: 11,
+            source: Some(source),
+            snapshot_complete: true,
+            scan_plan_digest: Some(DataUsageScanPlanDigest([3; 32])),
+            segment_invalidation_proof: Some(proof),
+            ..Default::default()
+        },
+        cache: HashMap::new(),
+    }
+}
+
 async fn setup_two_pool_scanner_store() -> (tempfile::TempDir, Arc<ECStore>) {
    init_ecstore_config_for_scanner_tests();
    let temp_dir = tempfile::tempdir().expect("multi-pool scanner test directory should be created");
@@ -1039,6 +1384,29 @@ fn dirty_usage_snapshot_detects_uncovered_generation() {
    clear_dirty_usage_buckets_for_tests();
 }

+#[test]
+#[serial]
+fn dirty_usage_producer_evidence_tracks_process_local_coverage_without_durable_restart_authority() {
+    use crate::segment_invalidation::SegmentInvalidationProducerIdentity;
+
+    clear_dirty_usage_buckets_for_tests();
+    record_dirty_usage_bucket_from_producers("photos", SegmentInvalidationProducerIdentity::REQUIRED_PRODUCTION);
+    let snapshot = snapshot_dirty_usage_buckets(&[bucket_info("photos")], dirty_usage_generation());
+
+    let evidence = dirty_usage_producer_evidence(&snapshot);
+
+    assert!(evidence.generation_window_bound);
+    assert!(evidence.producer_identity_coverage_complete);
+    assert!(!evidence.durable_producer_identity);
+    assert!(!evidence.restart_gap_absent);
+
+    record_dirty_usage_bucket_from_producer("videos", SegmentInvalidationProducerIdentity::PutObject);
+    let stale_evidence = dirty_usage_producer_evidence(&snapshot);
+    assert!(!stale_evidence.generation_window_bound);
+    assert!(!stale_evidence.producer_identity_coverage_complete);
+    clear_dirty_usage_buckets_for_tests();
+}
+
 #[test]
 fn generation_saturates_instead_of_wrapping() {
    let generation = AtomicU64::new(u64::MAX - 1);
@@ -1346,6 +1714,13 @@ async fn set_snapshot_reuse_requires_execution_identity_and_fences_stale_writers

    let ctx = CancellationToken::new();
    let empty_execution = DataUsageScanPlanDigest([5; 32]);
+    let segment_invalidation_proof = crate::DataUsageSegmentInvalidationProof {
+        process_epoch: scanner_activity_epoch().to_string(),
+        generation_start: 8,
+        generation_end: 8,
+        producer_identity_coverage_complete: true,
+        cold_zero_walk_oracle: false,
+    };
    set.nsscanner_cache(
        ctx.clone(),
        ScannerCycleBudget::new(&ctx, ScannerCycleBudgetConfig::default()),
@@ -1365,6 +1740,8 @@ async fn set_snapshot_reuse_requires_execution_identity_and_fences_stale_writers
            bucket_failures: ScannerBucketFailureState::default(),
            pending_maintenance_work: Arc::new(AtomicBool::new(false)),
            cache_cycle_floor: Arc::new(AtomicU64::new(8)),
+            cold_zero_walk_reuse_observed: Arc::new(AtomicBool::new(false)),
+            segment_invalidation_proof: Some(segment_invalidation_proof.clone()),
        },
        tx,
        8,
@@ -1374,6 +1751,7 @@ async fn set_snapshot_reuse_requires_execution_identity_and_fences_stale_writers
    .expect("empty set scope should replace its prior nonempty cache");
    let empty = rx.try_recv().expect("empty set snapshot should be published");
    assert_eq!(empty.info.scan_execution_digest, Some(empty_execution));
+    assert_eq!(empty.info.segment_invalidation_proof, Some(segment_invalidation_proof));
    assert!(empty.info.snapshot_complete);
    let root = empty.checked_flatten(DATA_USAGE_ROOT).expect("complete empty root");
    assert_eq!((root.size, root.objects), (0, 0));
@@ -1401,6 +1779,7 @@ fn complete_usage_baseline(
            scan_plan_digest: Some(scan_plan_digest.0),
            complete: true,
            tombstone: false,
+            segment_invalidation_proof: None,
        }],
        ..Default::default()
    };
@@ -1781,6 +2160,18 @@ fn remote_dirty_usage_invalidates_local_prefix_hints_until_distributed_proof_exi
        "peer dirty state is not a distributed segment invalidation proof"
    );
    assert_eq!(distributed.remote_dirty_usage_acknowledgements.len(), 1);
+    let evidence = distributed
+        .distributed_segment_invalidation_evidence
+        .expect("same-window peer snapshot and scoped ACK capability form distributed evidence");
+    assert_eq!(evidence.peer_count, 1);
+    assert_eq!(evidence.dirty_peer_count, 1);
+    assert_eq!(
+        evidence.invalidation_domain,
+        crate::segment_invalidation::SegmentInvalidationDomain::DistributedEc
+    );
+    assert!(evidence.distributed_ec_invalidation);
+    assert!(evidence.same_window_remote_proof);
+    assert!(evidence.all_peers_bound_to_generation_window);
 }

 fn peer_dirty_usage_snapshot(
@@ -1827,7 +2218,7 @@ fn verified_remote_dirty_usage_buckets_merges_only_complete_current_snapshots()
            ScannerPeerDirtyUsageExpectation {
                instance_id: "instance-b".to_string(),
                generation: 3,
-                pending: false,
+                pending: true,
            },
        ),
    ]);
@@ -1874,10 +2265,36 @@ fn verified_remote_dirty_usage_buckets_merges_only_complete_current_snapshots()
                    },
                },
            ],
+            peer_count: 2,
+            dirty_peer_count: 2,
        })
    );
 }

+#[test]
+fn verified_remote_dirty_usage_rejects_peer_snapshot_that_contradicts_activity_pending_state() {
+    let expected_peers = HashMap::from([(
+        "node-a:9000".to_string(),
+        ScannerPeerDirtyUsageExpectation {
+            instance_id: "instance-a".to_string(),
+            generation: 7,
+            pending: false,
+        },
+    )]);
+
+    assert!(
+        verified_remote_dirty_usage(
+            &expected_peers,
+            vec![(
+                "node-a:9000".to_string(),
+                peer_dirty_usage_snapshot("instance-a", 7, true, &[("photos", 7)]),
+            )],
+        )
+        .is_none(),
+        "a clean activity window cannot authorize a dirty peer snapshot or scoped ACK"
+    );
+}
+
 #[test]
 fn scanner_scoped_dirty_usage_ack_cost_threshold_is_single_protocol_batch() {
    let acknowledgement = |entry_count: usize| crate::scanner::ScannerDirtyUsageAcknowledgement {
@@ -1951,6 +2368,7 @@ fn remote_dirty_usage_scope_resolution_falls_back_when_ack_batch_exceeds_thresho
        result.remote_dirty_usage_acknowledgements.is_empty(),
        "full-scan fallback must not send a scoped ACK that peers would reject or split"
    );
+    assert!(result.distributed_segment_invalidation_evidence.is_none());
 }

 #[test]
@@ -2092,6 +2510,11 @@ async fn distributed_scoped_scan_falls_back_when_remote_scoped_ack_capability_is

        assert_eq!(result.scope.selected_buckets.as_deref(), expected_buckets.as_ref());
        assert_eq!(result.remote_dirty_usage_acknowledgements.len(), expected_ack_count);
+        assert_eq!(
+            result.distributed_segment_invalidation_evidence.is_some(),
+            capability,
+            "distributed evidence requires an authenticated scoped ACK capability probe"
+        );
    }
 }

@@ -84,7 +84,14 @@ async fn persist_baseline(store: &Arc<ECStore>, baseline: &DataUsageInfo) {

 // Every invocation uses the production default scope. Once durable bucket
 // incarnations are present, the expected walker set follows the resolved scope.
-async fn run_entry(store: &Arc<ECStore>, cycle: u64, selected: Option<&str>, expect_walks: bool) -> DataUsageInfo {
+async fn run_entry(
+    store: &Arc<ECStore>,
+    cycle: u64,
+    selected: Option<&str>,
+    expect_walks: bool,
+    expect_activation: bool,
+    expect_prefix_scope: bool,
+) -> DataUsageInfo {
    let drives = drive_identities(store).await;
    let inventory = store
        .list_bucket_for_scanner(&BucketOptions::default())
@@ -138,11 +145,19 @@ async fn run_entry(store: &Arc<ECStore>, cycle: u64, selected: Option<&str>, exp
    .expect("entry cycle should finish within the fixture deadline")
    .expect("entry cycle should succeed");
    assert_eq!(result.status, ScannerCycleStatus::Complete);
+    let activation_preflight = result.segment_reuse_activation_preflight;
    let scope = observed.await.expect("production resolver should report its decision");
    assert_eq!(
        scope.selected_buckets.as_deref(),
        selected.map(|name| HashSet::from([name.to_string()])).as_ref()
    );
+    if let Some(selected) = selected {
+        assert_eq!(
+            scope.prefix_scope_for(selected).is_some(),
+            expect_prefix_scope,
+            "resolved prefix scope must match activation replay for cycle {cycle}"
+        );
+    }
    let usage = receiver.recv().await.expect("one candidate should be delivered");
    assert!(receiver.recv().await.is_none(), "there must be exactly one terminal candidate");
    assert!(usage.usage_snapshot_complete);
@@ -174,6 +189,22 @@ async fn run_entry(store: &Arc<ECStore>, cycle: u64, selected: Option<&str>, exp
        actual, expected_walks,
        "each listed source/bucket must have exactly the expected real walks"
    );
+    assert!(activation_preflight.production_activation);
+    assert_eq!(activation_preflight.scanner_segment_reuse_activated, expect_activation);
+    let activation_blockers = activation_preflight.fail_closed_blockers().collect::<Vec<_>>();
+    if expect_activation {
+        assert_eq!(activation_blockers, Vec::<&str>::new());
+    } else if selected.is_some() && expect_walks {
+        assert!(
+            !activation_blockers.contains(&"missing_cold_zero_walk_oracle"),
+            "a complete scoped reuse cycle must carry the cold zero-walk oracle: cycle={cycle} selected={selected:?} blockers={activation_blockers:?}"
+        );
+    } else {
+        assert!(
+            activation_blockers.contains(&"missing_cold_zero_walk_oracle"),
+            "unscoped or same-cycle cache reuse must not claim the cold zero-walk oracle: cycle={cycle} selected={selected:?} expect_walks={expect_walks} blockers={activation_blockers:?}"
+        );
+    }
    assert_eq!(
        read_config_with_revision(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str())
            .await
@@ -189,32 +220,78 @@ async fn run_entry(store: &Arc<ECStore>, cycle: u64, selected: Option<&str>, exp
    usage
 }

-#[tokio::test]
+fn record_segment_dirty_usage(bucket: &str) {
+    for producer in crate::segment_invalidation::SegmentInvalidationProducerIdentity::REQUIRED_PRODUCTION {
+        record_dirty_usage_object_from_producer(bucket, "hot-segment/object", producer);
+    }
+}
+
+// The scoped fallback fixture keeps two EC pools and several scan futures live
+// at once. Run the async cases on a dedicated stack so Linux libtest defaults
+// exercise the assertions instead of aborting before the oracle finishes.
+fn run_scoped_entry_fallback_test<F, Fut>(thread_name: &'static str, test_fn: F)
+where
+    F: FnOnce() -> Fut + Send + 'static,
+    Fut: std::future::Future<Output = ()> + 'static,
+{
+    let handle = std::thread::Builder::new()
+        .name(thread_name.to_string())
+        .stack_size(32 * 1024 * 1024)
+        .spawn(move || {
+            let runtime = tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+                .expect("scoped entry fallback runtime should build");
+            runtime.block_on(test_fn());
+        })
+        .expect("scoped entry fallback test thread should spawn");
+    if let Err(payload) = handle.join() {
+        std::panic::resume_unwind(payload);
+    }
+}
+
+#[test]
 #[serial]
-async fn scoped_entry_fallback_distinguishes_planned_scope_from_real_cold_walks() {
+fn scoped_entry_fallback_distinguishes_planned_scope_from_real_cold_walks() {
+    run_scoped_entry_fallback_test(
+        "scanner-scoped-entry-planned-scope",
+        scoped_entry_fallback_distinguishes_planned_scope_from_real_cold_walks_case,
+    );
+}
+
+async fn scoped_entry_fallback_distinguishes_planned_scope_from_real_cold_walks_case() {
    let (_dir, store) = setup_two_pool_scanner_store().await;
    clear_dirty_usage_buckets_for_tests();
    let hot = format!("hot-{}", Uuid::new_v4().simple());
    let cold = format!("cold-{}", Uuid::new_v4().simple());
    create_bucket(&store, &hot).await;
    create_bucket(&store, &cold).await;
-    record_dirty_usage_bucket(&hot);
-    let baseline = run_entry(&store, 1, None, true).await;
+    record_segment_dirty_usage(&hot);
+    let baseline = run_entry(&store, 1, None, true, false, false).await;
    persist_baseline(&store, &baseline).await;

    // Same-cycle Current remains a retry. The later cycle may skip the cold
    // bucket only after the prior complete set cache has durable incarnations.
-    run_entry(&store, 1, Some(&hot), false).await;
-    let usage = run_entry(&store, 2, Some(&hot), true).await;
+    run_entry(&store, 1, Some(&hot), false, false, false).await;
+    let usage = run_entry(&store, 2, Some(&hot), true, true, false).await;
+    persist_baseline(&store, &usage).await;
+    let usage = run_entry(&store, 3, Some(&hot), true, true, true).await;
    assert_eq!(usage.buckets_usage[&hot].objects_count, 1);
    assert_eq!(usage.buckets_usage[&cold].objects_count, 1);
    assert_eq!(usage.objects_total_count, 2);
    clear_dirty_usage_buckets_for_tests();
 }

-#[tokio::test]
+#[test]
 #[serial]
-async fn scoped_entry_fallback_rejects_invalid_persisted_baseline_at_the_walker() {
+fn scoped_entry_fallback_rejects_invalid_persisted_baseline_at_the_walker() {
+    run_scoped_entry_fallback_test(
+        "scanner-scoped-entry-invalid-baseline",
+        scoped_entry_fallback_rejects_invalid_persisted_baseline_at_the_walker_case,
+    );
+}
+
+async fn scoped_entry_fallback_rejects_invalid_persisted_baseline_at_the_walker_case() {
    let (_dir, store) = setup_two_pool_scanner_store().await;
    clear_dirty_usage_buckets_for_tests();
    let hot = format!("hot-{}", Uuid::new_v4().simple());
@@ -223,7 +300,7 @@ async fn scoped_entry_fallback_rejects_invalid_persisted_baseline_at_the_walker(
    create_bucket(&store, &cold).await;
    record_dirty_usage_bucket(&hot);
    // The first real scan is also the missing persisted-baseline case.
-    let baseline = run_entry(&store, 1, None, true).await;
+    let baseline = run_entry(&store, 1, None, true, false, false).await;
    for (index, kind) in [
        "malformed",
        "unconverged",
@@ -256,28 +333,43 @@ async fn scoped_entry_fallback_rejects_invalid_persisted_baseline_at_the_walker(
        crate::save_config(store.clone(), DATA_USAGE_OBJ_NAME_PATH.as_str(), bytes)
            .await
            .expect("negative baseline should persist");
-        let usage = run_entry(&store, u64::try_from(index).expect("fixture cycle index should fit") + 2, None, true).await;
+        let usage = run_entry(
+            &store,
+            u64::try_from(index).expect("fixture cycle index should fit") + 2,
+            None,
+            true,
+            false,
+            false,
+        )
+        .await;
        assert_eq!(usage.objects_total_count, 2, "{kind}");
        assert_eq!(usage.buckets_usage[&cold].objects_count, 1, "{kind}");
    }
    clear_dirty_usage_buckets_for_tests();
 }

-#[tokio::test]
+#[test]
 #[serial]
-async fn scoped_entry_fallback_covers_overflow_and_new_bucket_inventory() {
+fn scoped_entry_fallback_covers_overflow_and_new_bucket_inventory() {
+    run_scoped_entry_fallback_test(
+        "scanner-scoped-entry-overflow-inventory",
+        scoped_entry_fallback_covers_overflow_and_new_bucket_inventory_case,
+    );
+}
+
+async fn scoped_entry_fallback_covers_overflow_and_new_bucket_inventory_case() {
    let (_dir, store) = setup_two_pool_scanner_store().await;
    clear_dirty_usage_buckets_for_tests();
    let hot = format!("hot-{}", Uuid::new_v4().simple());
    create_bucket(&store, &hot).await;
    record_dirty_usage_bucket(&hot);
-    let baseline = run_entry(&store, 1, None, true).await;
+    let baseline = run_entry(&store, 1, None, true, false, false).await;
    persist_baseline(&store, &baseline).await;
    for index in 0..=crate::SCANNER_DIRTY_USAGE_SNAPSHOT_MAX_ENTRIES {
        record_dirty_usage_bucket(&format!("overflow-{index}"));
    }
    assert!(dirty_usage_buckets_for_tests().len() > crate::SCANNER_DIRTY_USAGE_SNAPSHOT_MAX_ENTRIES);
-    let usage = run_entry(&store, 2, None, true).await;
+    let usage = run_entry(&store, 2, None, true, false, false).await;
    assert_eq!(usage.objects_total_count, 1);

    clear_dirty_usage_buckets_for_tests();
@@ -285,7 +377,7 @@ async fn scoped_entry_fallback_covers_overflow_and_new_bucket_inventory() {
    let new_bucket = format!("new-{}", Uuid::new_v4().simple());
    create_bucket(&store, &new_bucket).await;
    // Even a previously valid baseline cannot cover the changed inventory.
-    let usage = run_entry(&store, 3, None, true).await;
+    let usage = run_entry(&store, 3, None, true, false, false).await;
    assert_eq!(usage.objects_total_count, 2);
    assert_eq!(usage.buckets_usage[&new_bucket].objects_count, 1);
    clear_dirty_usage_buckets_for_tests();
@@ -16,23 +16,29 @@ use super::*;
 use crate::data_usage_define::{DATA_USAGE_OBJ_NAME_PATH, read_config_with_revision};

 async fn create_cohort_bucket(store: &ECStore, bucket: &str) {
+    create_cohort_bucket_objects(store, bucket, 1).await;
+}
+
+async fn create_cohort_bucket_objects(store: &ECStore, bucket: &str, objects: usize) {
    store
        .make_bucket(bucket, &MakeBucketOptions::default())
        .await
        .expect("fixture bucket");
    for set in store.all_set_disks() {
-        let mut reader = ScannerPutObjReader::from_vec(b"cohort".to_vec());
-        set.put_object(
-            bucket,
-            "initial",
-            &mut reader,
-            &ScannerObjectOptions {
-                no_lock: true,
-                ..Default::default()
-            },
-        )
-        .await
-        .expect("fixture object and all rename tails should persist");
+        for index in 0..objects {
+            let mut reader = ScannerPutObjReader::from_vec(b"cohort".to_vec());
+            set.put_object(
+                bucket,
+                &format!("object-{index:04}"),
+                &mut reader,
+                &ScannerObjectOptions {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("fixture object and all rename tails should persist");
+        }
    }
 }

@@ -168,6 +174,92 @@ async fn service_cohort_production_dispatch_services_waiters_across_sources() {
    clear_dirty_usage_buckets_for_tests();
 }

+#[tokio::test]
+#[serial]
+async fn service_cohort_flat_bucket_budget_does_not_publish_unscanned_small_bucket() {
+    let (_dir, store) = setup_two_pool_scanner_store().await;
+    clear_dirty_usage_buckets_for_tests();
+    let flat = format!("a-flat-{}", Uuid::new_v4().simple());
+    let small = format!("z-small-{}", Uuid::new_v4().simple());
+    create_cohort_bucket_objects(&store, &flat, 6).await;
+    create_cohort_bucket(&store, &small).await;
+    let cohort = Arc::new(StdMutex::new(ScannerServiceCohort::default()));
+
+    let expected_flat = store
+        .all_set_disks()
+        .iter()
+        .map(|set| (DataUsageCacheSource::new(set.pool_index, set.set_index), flat.clone()))
+        .collect::<HashSet<_>>();
+    let expected_small = store
+        .all_set_disks()
+        .iter()
+        .map(|set| (DataUsageCacheSource::new(set.pool_index, set.set_index), small.clone()))
+        .collect::<HashSet<_>>();
+
+    let ctx = CancellationToken::new();
+    let budget = ScannerCycleBudget::new_with_progress_tracking(
+        &ctx,
+        ScannerCycleBudgetConfig {
+            max_objects: Some(1),
+            ..Default::default()
+        },
+    );
+    let (result, usage) = run_cohort_cycle(&store, cohort.clone(), 1, budget.clone()).await;
+    assert_eq!(result.status, ScannerCycleStatus::Incomplete);
+    assert!(usage.is_none(), "wide-bucket budget exhaustion must not publish a partial aggregate");
+    assert!(budget.budget_elapsed());
+    let first_round_admitted = cohort
+        .lock()
+        .expect("cohort lock")
+        .admitted_members()
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert!(
+        !first_round_admitted.is_disjoint(&expected_flat),
+        "the first fixed budget round should exercise the wide flat bucket"
+    );
+    assert!(
+        first_round_admitted.is_disjoint(&expected_small),
+        "a small bucket not yet reached by the real scanner must not be marked admitted"
+    );
+
+    for cycle in 2..=4 {
+        let ctx = CancellationToken::new();
+        let (result, usage) = run_cohort_cycle(
+            &store,
+            cohort.clone(),
+            cycle,
+            ScannerCycleBudget::new_with_progress_tracking(
+                &ctx,
+                ScannerCycleBudgetConfig {
+                    max_objects: Some(1),
+                    ..Default::default()
+                },
+            ),
+        )
+        .await;
+        assert_eq!(result.status, ScannerCycleStatus::Incomplete);
+        assert!(usage.is_none(), "mixed partial coverage still cannot publish the set root");
+    }
+    let admitted_after_budgeted_rounds = cohort
+        .lock()
+        .expect("cohort lock")
+        .admitted_members()
+        .into_iter()
+        .collect::<HashSet<_>>();
+    assert!(
+        expected_small.is_subset(&admitted_after_budgeted_rounds),
+        "tracked small buckets must receive real execution opportunities within their fixed service-round bound"
+    );
+
+    let ctx = CancellationToken::new();
+    let (result, usage) =
+        run_cohort_cycle(&store, cohort, 5, ScannerCycleBudget::new(&ctx, ScannerCycleBudgetConfig::default())).await;
+    assert_eq!(result.status, ScannerCycleStatus::Complete);
+    assert_eq!(usage.expect("final complete aggregate").objects_total_count, 14);
+    clear_dirty_usage_buckets_for_tests();
+}
+
 #[tokio::test]
 #[serial]
 async fn service_cohort_fresh_complete_aggregate_preserves_reordered_sources() {
@@ -25,6 +25,7 @@ pub enum SegmentInvalidationError {
    ByteLimit,
    InvalidProof,
    InvalidKey,
+    UnknownProducer,
 }

 #[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord)]
@@ -50,6 +51,82 @@ impl SegmentInvalidationProducer {
    ];
 }

+#[derive(Clone, Copy, Debug, PartialEq, Eq, PartialOrd, Ord)]
+pub enum SegmentInvalidationProducerIdentity {
+    PutObject,
+    DeleteObject,
+    DeleteMarker,
+    CompleteMultipartUpload,
+    AbortMultipartUpload,
+    ObjectMetadata,
+    BucketMetadata,
+    Replication,
+    TierTransition,
+    TierExpiration,
+    DirectoryObject,
+    Unknown,
+    TestFixture,
+}
+
+impl SegmentInvalidationProducerIdentity {
+    pub const REQUIRED_PRODUCTION: [Self; 11] = [
+        Self::PutObject,
+        Self::DeleteObject,
+        Self::DeleteMarker,
+        Self::CompleteMultipartUpload,
+        Self::AbortMultipartUpload,
+        Self::ObjectMetadata,
+        Self::BucketMetadata,
+        Self::Replication,
+        Self::TierTransition,
+        Self::TierExpiration,
+        Self::DirectoryObject,
+    ];
+
+    pub fn producer(self) -> Option<SegmentInvalidationProducer> {
+        match self {
+            Self::PutObject => Some(SegmentInvalidationProducer::Put),
+            Self::DeleteObject => Some(SegmentInvalidationProducer::Delete),
+            Self::DeleteMarker => Some(SegmentInvalidationProducer::DeleteMarker),
+            Self::CompleteMultipartUpload | Self::AbortMultipartUpload => Some(SegmentInvalidationProducer::Multipart),
+            Self::ObjectMetadata => Some(SegmentInvalidationProducer::Put),
+            Self::BucketMetadata => Some(SegmentInvalidationProducer::DirectoryObject),
+            Self::Replication => Some(SegmentInvalidationProducer::Replication),
+            Self::TierTransition | Self::TierExpiration => Some(SegmentInvalidationProducer::Tier),
+            Self::DirectoryObject => Some(SegmentInvalidationProducer::DirectoryObject),
+            Self::Unknown | Self::TestFixture => None,
+        }
+    }
+}
+
+pub fn complete_segment_invalidation_producers<I>(
+    identities: I,
+) -> Result<BTreeSet<SegmentInvalidationProducer>, SegmentInvalidationError>
+where
+    I: IntoIterator<Item = SegmentInvalidationProducerIdentity>,
+{
+    let mut covered_identities = BTreeSet::new();
+    let mut producers = BTreeSet::new();
+    for identity in identities {
+        let Some(producer) = identity.producer() else {
+            return Err(SegmentInvalidationError::UnknownProducer);
+        };
+        covered_identities.insert(identity);
+        producers.insert(producer);
+    }
+    if SegmentInvalidationProducerIdentity::REQUIRED_PRODUCTION
+        .iter()
+        .all(|identity| covered_identities.contains(identity))
+        && SegmentInvalidationProducer::REQUIRED
+            .iter()
+            .all(|producer| producers.contains(producer))
+    {
+        Ok(producers)
+    } else {
+        Err(SegmentInvalidationError::InvalidProof)
+    }
+}
+
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
 pub enum SegmentInvalidationDomain {
    LocalSingleSet,
@@ -170,7 +247,8 @@ mod tests {
    use super::*;

    fn producers() -> BTreeSet<SegmentInvalidationProducer> {
-        SegmentInvalidationProducer::REQUIRED.into_iter().collect()
+        complete_segment_invalidation_producers(SegmentInvalidationProducerIdentity::REQUIRED_PRODUCTION)
+            .expect("production producer matrix should be complete")
    }

    fn envelope() -> SegmentInvalidationEnvelope {
@@ -320,6 +398,75 @@ mod tests {
        );
    }

+    #[test]
+    fn segment_invalidation_producer_identities_must_be_known_and_complete() {
+        assert_eq!(
+            complete_segment_invalidation_producers(SegmentInvalidationProducerIdentity::REQUIRED_PRODUCTION),
+            Ok(SegmentInvalidationProducer::REQUIRED.into_iter().collect())
+        );
+        assert_eq!(
+            complete_segment_invalidation_producers([
+                SegmentInvalidationProducerIdentity::PutObject,
+                SegmentInvalidationProducerIdentity::DeleteObject,
+                SegmentInvalidationProducerIdentity::DeleteMarker,
+                SegmentInvalidationProducerIdentity::CompleteMultipartUpload,
+                SegmentInvalidationProducerIdentity::AbortMultipartUpload,
+                SegmentInvalidationProducerIdentity::ObjectMetadata,
+                SegmentInvalidationProducerIdentity::BucketMetadata,
+                SegmentInvalidationProducerIdentity::Replication,
+                SegmentInvalidationProducerIdentity::TierTransition,
+                SegmentInvalidationProducerIdentity::TierExpiration,
+                SegmentInvalidationProducerIdentity::DirectoryObject,
+                SegmentInvalidationProducerIdentity::Unknown,
+            ]),
+            Err(SegmentInvalidationError::UnknownProducer)
+        );
+        assert_eq!(
+            complete_segment_invalidation_producers([
+                SegmentInvalidationProducerIdentity::PutObject,
+                SegmentInvalidationProducerIdentity::DeleteObject,
+                SegmentInvalidationProducerIdentity::DeleteMarker,
+                SegmentInvalidationProducerIdentity::CompleteMultipartUpload,
+                SegmentInvalidationProducerIdentity::AbortMultipartUpload,
+                SegmentInvalidationProducerIdentity::ObjectMetadata,
+                SegmentInvalidationProducerIdentity::BucketMetadata,
+                SegmentInvalidationProducerIdentity::Replication,
+                SegmentInvalidationProducerIdentity::TierTransition,
+                SegmentInvalidationProducerIdentity::TierExpiration,
+                SegmentInvalidationProducerIdentity::DirectoryObject,
+                SegmentInvalidationProducerIdentity::TestFixture,
+            ]),
+            Err(SegmentInvalidationError::UnknownProducer)
+        );
+        assert_eq!(
+            complete_segment_invalidation_producers([
+                SegmentInvalidationProducerIdentity::PutObject,
+                SegmentInvalidationProducerIdentity::DeleteObject,
+                SegmentInvalidationProducerIdentity::DeleteMarker,
+                SegmentInvalidationProducerIdentity::CompleteMultipartUpload,
+                SegmentInvalidationProducerIdentity::Replication,
+                SegmentInvalidationProducerIdentity::TierTransition,
+                SegmentInvalidationProducerIdentity::DirectoryObject,
+            ]),
+            Err(SegmentInvalidationError::InvalidProof)
+        );
+        assert_eq!(
+            complete_segment_invalidation_producers([
+                SegmentInvalidationProducerIdentity::PutObject,
+                SegmentInvalidationProducerIdentity::DeleteObject,
+                SegmentInvalidationProducerIdentity::DeleteMarker,
+                SegmentInvalidationProducerIdentity::CompleteMultipartUpload,
+                SegmentInvalidationProducerIdentity::AbortMultipartUpload,
+                SegmentInvalidationProducerIdentity::ObjectMetadata,
+                SegmentInvalidationProducerIdentity::BucketMetadata,
+                SegmentInvalidationProducerIdentity::Replication,
+                SegmentInvalidationProducerIdentity::TierTransition,
+                SegmentInvalidationProducerIdentity::DirectoryObject,
+            ]),
+            Err(SegmentInvalidationError::InvalidProof)
+        );
+    }
+
    #[test]
    fn segment_invalidation_entries_are_bounded_and_key_checked() {
        let envelope = envelope();
@@ -38,9 +38,10 @@
 - `put-file-auth-epoch-strict` internode put_file epoch compatibility: rc.2 peers can cache a remote put_file capability before that remote node restarts, then continue sending v1 authenticated uploads with the old server epoch; those peers cannot recover from the 409 conflict used by newer clients to trigger a re-probe. Servers temporarily accept signed, non-nil stale put_file epochs while legacy put_file auth remains non-strict so mixed-version rolling upgrades can finish multipart/object writes. Remove the stale-epoch fallback after the minimum supported RustFS peer version re-probes put_file capability after server-epoch conflicts and legacy put_file auth is no longer accepted.
 - `disk-mutation-body-digest` internode mutating disk RPCs: servers temporarily accept mutating disk RPCs (RenameData, DeleteVersion, DeleteVersions, WriteMetadata, UpdateMetadata, WriteAll, Delete, DeletePaths, RenameFile, RenamePart, DeleteVolume, MakeVolume, MakeVolumes) that carry no signature-bound canonical body digest, so peers from releases that predate body-digest signing remain available during rolling upgrades. Accepted digestless mutations increment the internode body-digest fallback counter; that counter must read zero fleet-wide across a release window before RUSTFS_INTERNODE_RPC_BODY_DIGEST_STRICT is enabled. Because body-bound requests now consume replay-cache nonces on the receiver, deploy the raised RUSTFS_INTERNODE_RPC_REPLAY_CACHE_CAPACITY default fleet-wide before enabling strict mode, and watch the internode replay-cache overflow counter for undersized capacity during the rollout. Remove the digestless fallback after the minimum supported RustFS peer version body-binds every mutating disk RPC.
 - `heal-status-rpc-v1` node heal status capability: new peers treat an unimplemented BackgroundHealStatus RPC as an explicitly incomplete rolling-upgrade response. Remove the fallback after the minimum supported RustFS peer version implements BackgroundHealStatus.
- `backlog-1316` legacy encrypted multipart range seek: the feature remains opt-in until every server that can initiate, write, or complete multipart uploads supports the candidate-to-final marker protocol and uploadId commit lock, and pre-upgrade multipart uploads have drained. Remove the RUSTFS_ENCRYPTED_RANGE_SEEK switch after the minimum supported release does so; keep the quorum marker and malformed-layout full-read guards permanently.
+- `backlog-1316` legacy encrypted multipart range seek: the feature is on by default (RUSTFS_ENCRYPTED_RANGE_SEEK, default true) and the switch remains only as a kill switch until every server that can initiate, write, or complete multipart uploads supports the candidate-to-final marker protocol and uploadId commit lock, and pre-upgrade multipart uploads have drained. Remove the RUSTFS_ENCRYPTED_RANGE_SEEK switch after the minimum supported release does so; keep the quorum marker and malformed-layout full-read guards permanently.
 - `tonic-013-status-render` peer RPC failure classification: internode failures that reach a node only as text (a peer's error_info payload, a status flattened through format!) are classified by matching the rendering of an Unavailable gRPC status. Releases up to 1.0.0-alpha.38 shipped tonic 0.13, which rendered that status as "status: Unavailable, message: ..."; tonic 0.14 renders it as "code: 'The service is currently unavailable', message: ...". Both forms are matched so an older peer's relayed text still marks an unreachable peer offline. Remove the tonic 0.13 form after the minimum supported RustFS peer version ships tonic 0.14 or later.
 - `rustfs-5063` pre-beta.9 Local KMS recovery: persisted Local KMS configs from beta.8 and earlier predate the explicit insecure-development flag, and encrypted key files use the legacy SHA-256 KDF. Remove the config fallback after supported upgrades have rewritten or explicitly resaved all pre-beta.9 configs with the development-default field, and remove the legacy KDF after supported upgrades have rewritten all pre-beta.9 Local KMS key files with explicit at-rest protection.
+- `backlog-2369-legacy-nonce-fallback` pre-alpha.91 v1 segment nonce layout: releases before `1.0.0-alpha.91` reused a segment's part nonce for every block inside it, so the decrypt reader keeps that layout as its third and last v1 fallback. Because block zero's derived nonce equals that base nonce, the layout also lets a frame encrypted at index zero authenticate anywhere in its segment. Reads now lock a segment to whichever layout decoded its first non-zero block, which rejects a replay as soon as a later frame disagrees; a stream that is nothing but repeats of frame zero has no such later frame, so a deployment holding no pre-alpha.91 objects should set RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK=false to drop the layout outright. Remove the fallback (and the switch) after the minimum supported direct-upgrade release, and after migration tooling has rewritten every pre-alpha.91 encrypted object.
 - `sse-local-dek-json-v1` legacy local SSE DEK decoding: releases before the JSON envelope wrote wrapped DEKs as `base64(nonce):base64(ciphertext)`, so readers retain that decoder while all new writes use the versioned JSON envelope. Remove the colon decoder after the minimum supported direct-upgrade release writes JSON envelopes and migration tooling has rewritten every retained legacy object.
 - `not-initialized-error-code-v1` typed control-plane not-initialized wire code: control-plane RPC responses historically signaled an uninitialized peer only through the literal error_info string "errServerNotInitialized" (one drift site says "storage layer not initialized"). Responses now dual-carry a typed ControlPlaneErrorCode beside the legacy string, and clients prefer the code; the string stays populated and the client substring fallback (is_err_not_initialized, control_plane_failure) stays in place so mixed-version clusters keep classifying older peers' responses. Remove the substring fallback (and stop populating error_info for this case) after the minimum supported RustFS peer version always sends error_code.
 - `backlog-2097-tier-mutation-v4-error-text` tier-mutation Prepare rejection classification: a v3 server rejects a v4 request before store/runtime dispatch with the FailedPrecondition status and an authenticated, byte-exact unsupported-version message. A v4 coordinator recognizes only that exact code/message/requested-version tuple as definitely not persisted and fails the mutation without sending that peer an incompatible Abort; Unimplemented, near-text, missing/unknown failure classes, timeouts, and every other transport outcome remain ambiguous and stay in identity-bound Abort fanout. There is no automatic v3 retry. New servers retain v3 request/proof decoding for older coordinators, while operators must pause tier edit/remove/clear during a mixed v3/v4 rollout. Remove the text classifier after the minimum supported RustFS peer version returns the signed v4 PreDispatchRejected failure class.
@@ -67,6 +67,12 @@ Heal-side invariants that hold regardless of the caller:
 - Read-repair's local TTL reservation dedups only its own source and does not block heals from other sources; the namespace lock is the backstop.
 - The healing flag is never persisted, so there is no reverse risk of a leftover marker making a later commit yield incorrectly.

+## Graceful root-heal restart recovery
+
+Before a graceful shutdown cancels administrator cluster-wide heals, the manager saves unfinished requests on one coordinator disk as `.rustfs.sys/root-heal-<task-id>.json`. Startup replays the same task IDs and remaining execution budgets. Completion, cancellation, and replacement by `force_start` retire the record conditionally. An uncertain write or deletion does not create a fallback copy; an unsuccessful handoff retains the unclean-shutdown marker. Invalid or unsupported records remain on disk and defer root recovery without blocking the existing replacement-recovery path.
+
+This handoff covers the same coordinator and storage topology while the record disk remains configured and readable. It does not migrate records when a pool is retired or provide failover after loss of that disk. Older versions do not understand these records; cancellation while downgraded cannot retire a newer version's pending record. If a terminal budget checkpoint cannot be written, the previous record is retained and a warning is logged; the remaining-budget guarantee requires that write to succeed. The format is separate from object metadata and erasure-set checkpoints.
+
 ## Regression tests

 Both live in the test module of `crates/ecstore/src/set_disk/ops/heal.rs`:
@@ -38,6 +38,12 @@ Counts ignore blank lines and comments; compute them from the files. The lifecyc

 "Supported" for the SSE row means RustFS encrypts and decrypts its own objects. MinIO SSE objects (SSE-S3, SSE-KMS, SSE-C) are not readable in default builds; see [minio-file-format-compat.md Part C](minio-file-format-compat.md#part-c--server-side-encryption-sse) for the `rio-v2` migration build.

+### Client metadata expectations
+
+`CopyObject` with `MetadataDirective=REPLACE` clears standard metadata fields that the request omits, including `Content-Type`; it does not retain the source type or infer a default. Clients requiring a MIME type on the copied object must send `Content-Type` with the replacement metadata. This contract is covered by `crates/e2e_test/src/copy_object_metadata_test.rs`. A client test that expects an implicit `application/octet-stream` does not match this behavior.
+
+The MinIO-style `metadata=true` listing extension returns user metadata names without the HTTP `x-amz-meta-` prefix. It is not the standard S3 `ListObjectsV2` response. Clients that expect canonical HTTP header names in `UserMetadata` must normalize the names at that boundary; ordinary HEAD/GET metadata is unaffected. See `rustfs/src/app/bucket_usecase.rs` and its serialization tests.
+
 ## Replication Support Boundary

 Site replication and bucket replication are not the same compatibility claim.
@@ -15,6 +15,16 @@ RustFS ships several KMS backends. They differ not only in deployment effort but
 | Vault Transit | `VaultTransit` | Key-encryption keys never leave Vault; only Transit ciphertext is visible outside | Vault Transit engine (cryptographic isolation) | Delegated to Vault storage | Via Vault Transit key versioning | Deployments that need key material to be unreadable through storage APIs |
 | AWS KMS | `AWS` (alias `AwsKms`) | Key material never leaves AWS KMS; RustFS mirrors no key state | AWS KMS (cryptographic isolation) + IAM | Delegated to AWS | On-demand `RotateKeyOnDemand`; prior backing keys stay usable for decryption | Deployments rooted in AWS IAM — read [AWS KMS: deviations from the shared backend contract](#aws-kms-deviations-from-the-shared-backend-contract) first |

+## No KMS configured: the SSE-S3 local master key
+
+A deployment that never configures a KMS can still serve **SSE-S3** by setting `RUSTFS_SSE_S3_MASTER_KEY` to a base64-encoded 32-byte key. Data keys are then wrapped with that key using AES-256-GCM, on the node that serves the write. Understand three consequences before relying on it:
+
+- **The key is the whole confidentiality boundary.** It lives in the process environment of every node, with no ACL, no audit trail and no policy engine in front of it.
+- **Objects written this way can never be rotated.** There is no key record to rotate and no rewrap path; changing the value makes every object written under the old one unreadable. Migrating to a KMS later means rewriting those objects (for example with CopyObject), not reconfiguring.
+- **It does not serve SSE-KMS.** A request for `x-amz-server-side-encryption: aws:kms` on a node with no running KMS is refused — `400 InvalidRequest` when KMS was never configured, `503` when a configured service is not running. Earlier releases silently wrapped the data key with the local master key while still stamping `aws:kms` and the requested key id into the object metadata; that metadata claimed a KMS protection the object never had. If a deployment depended on that, either configure a KMS or ask for `AES256`.
+
+The value is unset by default, and a deployment that neither configures a KMS nor sets it simply cannot serve SSE-S3 (the write is refused, never silently downgraded to plaintext).
+
 ## Migrating from MinIO: encrypted objects do not carry over

 > **Warning: default RustFS builds fail closed on objects that MinIO encrypted.** This applies to SSE-S3, SSE-KMS, and SSE-C, whichever KMS backend you configure; configuring `Static` with MinIO's key material does not make them readable. Such objects list and HEAD normally (their `xl.meta` parses), and only the payload read fails — with S3 `InvalidObjectState`, never plaintext. Read a sample of encrypted objects, not just their listings, before decommissioning the MinIO deployment.
@@ -117,6 +127,37 @@ Decryption loads exactly the version recorded in the envelope and fails closed w

 Do not rotate any key until **every** RustFS node runs a build that understands the `master_key_version` envelope field. Older binaries ignore the field and always decrypt with the current material: harmless while nothing has been rotated, but after a rotation they fail to decrypt every object wrapped by an earlier key version. Complete the rolling upgrade of the entire cluster first, then rotate. The rest of this constraint class is collected in [Mixed-version clusters during a rolling upgrade](#mixed-version-clusters-during-a-rolling-upgrade).

+## SSE-C requires a secure transport
+
+An SSE-C request carries the customer's AES key in a request header, so AWS S3 and MinIO both refuse one that did not arrive over TLS. A plaintext hop hands that key to anyone on the path, and because the object cannot be read without the same key, the exposure lasts as long as the object does.
+
+This release reports rather than refuses, because flipping straight to a rejection would break every plaintext staging and test deployment inside a release window:
+
+- Every SSE-C request on a plaintext transport increments `rustfs_ssec_plaintext_requests_total` and logs one `ssec_request_without_tls` warning per process.
+- `RUSTFS_SSE_C_REQUIRE_TLS=true` (default `false`) refuses those requests now, with the same `400 InvalidRequest` wording AWS uses. Confirm the counter reads zero before enabling it.
+- The default is expected to flip in a later release.
+
+The verdict is per connection: a listener that terminates TLS satisfies it, and so does an `https` protocol forwarded by a proxy the trusted-proxy configuration accepts. A direct plaintext client asserts nothing, and a forwarded protocol from an untrusted peer is not consulted.
+
+## Object ciphertext format: what the v1 frame layout does and does not authenticate
+
+Every object RustFS writes today uses the **v1** frame layout (the v2 layout exists and is read automatically, but its write switch `RUSTFS_ENCRYPTION_FRAME_V2` is off by default). Each frame is authenticated with AES-256-GCM under a nonce derived from the object's base nonce and the frame's index. Three properties do **not** follow from that, and an operator's threat model has to account for them:
+
+- **No frame-index binding.** A frame's index is not part of its associated data. Authentication proves a frame was produced under this object's key; it does not by itself prove the frame belongs at the position it occupies.
+- **No final-frame authentication.** Nothing in a v1 stream marks the last frame, so a stream that has been cut short is not distinguishable from a shorter object by cryptographic means.
+- **Truncation is not detected server-side.** A full GET is cut off by the length gate mid-stream and surfaces as `IncompleteBody` — after the response headers have already gone out. A ranged GET that ends early looks like an ordinary EOF and is not reported at all. A client that needs a truncation signal must compare the delivered length against `Content-Length` itself.
+
+These matter only to an attacker who can already rewrite the underlying shards. Shard integrity uses a keyed-hash-free checksum (HighwayHash), which such an attacker can recompute, so it is not a barrier.
+
+Two historical shapes additionally reuse a GCM nonce and cannot be repaired by any read-side change:
+
+| Shape | Written by | Consequence | Migration |
+| --- | --- | --- | --- |
+| Multipart objects written before `1.0.0-alpha.91` | The pre-alpha.91 writer reused a segment's part nonce for every block in it | The whole segment shares one nonce; a frame from index zero authenticates anywhere in that segment | Rewrite in place with CopyObject; then set `RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK=false` |
+| SSE-C objects written before `1.0.0-beta.9` that carry no stored IV | The nonce was derived deterministically from bucket and key | Historical versions of the same key share a nonce, which affects confidentiality as well as forgeability | Rewrite in place with CopyObject |
+
+The decrypt reader locks each segment to whichever nonce layout decoded its first non-zero-index frame, so a replayed frame is rejected as soon as any later frame disagrees. A stream that is nothing but repeats of frame zero has no later frame to disagree, so a deployment that holds no pre-alpha.91 objects should set `RUSTFS_ENCRYPTION_LEGACY_NONCE_FALLBACK=false` (default `true`) to drop that layout entirely. Turning it off refuses to decrypt pre-alpha.91 objects, so migrate first.
+
 ## Mixed-version clusters during a rolling upgrade

 During a rolling upgrade KMS state is shared three ways: **Vault** holds key records and Transit metadata, **cluster storage** holds the persisted KMS configuration, and **each node's process memory** holds caches and the live backend instance. Nodes on different builds agree on the first, may disagree on the third, and can disagree on configuration for as long as the operator leaves them running, because the reload broadcast that converges configuration is one of the things an older build rejects. This section is written for the KV2 and Transit backends; the Local backend is unsupported for multi-node deployments regardless of version (see the [deployment support matrix](#deployment-support-matrix)).
@@ -237,6 +278,7 @@ The Local backend stores one JSON record per key (`<key_id>.key`) plus an Argon2
 - `Local` is the default backend (`kms_backend` defaults to `local`) and is a development, testing and demo backend; it is not supported for production. Activating a backend whose capabilities report `production_supported: false` logs a `kms_backend_positioning` warning on every start, restart and reconfigure, and the `kms/status` capability matrix carries the same flag. The positioning is a warning, not a gate.
 - Configuration validation enforces stricter rules outside explicit development mode: a master key is required and `key_dir` must not live under the process temp directory.
 - The RustFS Kubernetes operator places the key directory on a PersistentVolumeClaim, so keys survive pod rescheduling.
+- **A multi-node deployment cannot share it.** Key material lives on each node's own disk and the Argon2id salt is generated per node, so two nodes derive different keys from the same `master_key`. An object encrypted on node A cannot be decrypted on node B; behind a load balancer that appears as intermittent 500s on reads that succeeded a moment earlier. Configuring `Local` while the deployment is distributed logs `kms_node_local_backend_in_distributed_deployment` and appends the same warning to the `kms/configure` response. This stays a warning, not a gate.
 - Production multi-node deployments should use the Vault Transit backend.

 ### Deployment support matrix
@@ -9,12 +9,14 @@ The drill rehearses the complete loop — back up, lose the persistence layer, p

 The drill covers the **Local** backend, the only backend RustFS produces a full-material bundle for. The responsibility split is described in `crates/kms/src/backup/capability.rs`:

-| Backend | What a RustFS bundle carries | What restores it |
+| Backend | RustFS bundle export | What restores it |
 | --- | --- | --- |
 | Local | Key records, all stored versions, the KDF salt, sanitized configuration | The RustFS restore in this runbook |
-| Static | Non-sensitive references only | The operator re-supplies the secret out of band |
-| Vault KV2 + Transit | KV metadata and Transit ciphertext references | Vault's native snapshot restore, then the RustFS orchestration |
-| Vault Transit | Metadata, configuration references, verification data | Vault's native snapshot restore, then the RustFS orchestration |
+| Static | Refused with `501`; RustFS holds no material to export | The operator re-supplies the secret out of band |
+| Vault KV2 | Refused with `501` | Vault's native snapshot restore, then the RustFS orchestration |
+| Vault Transit | Refused with `501` | Vault's native snapshot restore, then the RustFS orchestration |
+
+The `501` is not a gap in this runbook: `POST /rustfs/admin/v3/kms/backup` refuses any backend other than `Local` (`rustfs/src/admin/handlers/kms_backup.rs`, `execute_backup`), so no RustFS bundle exists to plan around for the other three. Note that `capability.rs` still *declares* `FullMaterial` responsibility for Vault KV2 in storage-only mode; no export path implements it, so treat the declaration as a reservation, not a capability.

 For the Vault backends there is no RustFS-side export: the cryptographic root is non-exportable and comes back through Vault's own disaster-recovery flow. RustFS owns the refusal to proceed before that has happened and the ordering of everything after it — see the Vault section below.

@@ -225,6 +225,10 @@ KMS configured through the admin API is persisted to cluster storage and restore

 To recover from `load_failed` — or from any state where the server runs but its in-memory KMS lags the persisted configuration — call `POST /rustfs/admin/v3/kms/reload` (`kms:ServiceControl`). It re-reads the persisted configuration from cluster storage and reconfigures the service without resubmitting secrets, then broadcasts the reload to peer nodes. If reload keeps failing, check cluster storage health first (the read needs quorum), then `RUSTFS_KMS_CONFIG_SECRET`: an unseal error means the secret is missing or differs from the one that sealed the persisted copy — it must be identical on every node.

+Reload short-circuits only when this node is **already running** the persisted configuration. A node whose KMS failed to start keeps that configuration and sits in `Error`, so reload reconfigures it — which starts the service — rather than reporting success while the node stays down. The same holds on every peer, which reaches the same path through the reload broadcast.
+
+**Cluster-wide versus node-local routes.** `configure`, `reconfigure` and `reload` are cluster operations: the node that serves the request broadcasts to its peers. `start` and `stop` are node-local and are **not** broadcast. Calling `POST /rustfs/admin/v3/kms/stop` through a load balancer therefore stops whichever node answered and leaves the cluster in a mixed state; address a specific node directly when you mean node-local semantics, and expect a later cluster-wide `reload` to start a stopped node again.
+
 A separate event, `kms_config_load_skipped` with `reason="storage_uninitialized"`, comes from the ambient loader used by the peer-reload RPC path; seeing it outside a peer reload indicates a request arrived before storage initialization finished.

 ## Threshold calibration
@@ -39,8 +39,8 @@ The `scanner` and `heal` subsystems are served by `GetConfigKVHandler` (`rustfs/
 The `--abba` mode runs five independent scenario cells: `cold-hot`, `fresh-hot`,
 `multi-hot-new`, `running-heal`, and `mrf-replay`. Each scenario runs at least
 three A1/B1/B2/A2 groups for both baseline/candidate with background work on,
-and candidate-only background off/on. A measured leg lasts at least 900
-seconds; the minimum matrix contains 120 legs (30 hours before setup/oracles).
+and candidate-only background off/on. A measured release leg lasts at least 7200
+seconds; the minimum matrix contains 120 legs (240 hours before setup/oracles).
 The existing `performance-ab.yml` supplies the pattern for immutable build
 provenance and failure propagation, but its short Warp workload is not this
 scanner gate. No scheduled workflow starts this matrix automatically.
@@ -63,9 +63,10 @@ The manifest has the following JSON contract (all fields are required):
 | Field | Value |
 |---|---|
 | `schema`, `evidence` | `1`, and `measured` or `synthetic`. |
-| `rounds`, `duration_seconds`, `min_free_bytes` | 3..10 groups, 900..86400 seconds for measured runs, and the independently estimated free-space reservation in bytes. Synthetic runs may use 1 second. |
+| `rounds`, `duration_seconds`, `min_free_bytes` | 3..10 groups, 7200..86400 seconds for measured release runs, and the independently estimated free-space reservation in bytes. Synthetic runs may use 1 second. |
 | `baseline`, `candidate` | Each contains executable `binary`, full 40-character `revision`, and verified `sha256`. The runner rehashes binaries before every leg. |
 | `fixed` | `config_sha256`, `dataset_sha256`, `release_flags`, `durability`, `disk_type`, `cache_state`, `load_command`, `resource_isolation`, `topology` (`EC8+4`), and positive `offered_load_ops`. Hashes use 64 lowercase hexadecimal characters. |
+| `release_evidence` | Required for `measured` runs. It binds the 3x4 EC8+4 topology, multi-pool/multi-set coverage, per-node metrics endpoints, same-window distributed sampling, process restart and crash-restart fault modes, mixed-version reader/writer/rollback participation, and allocation/flamegraph/RSS/save-frequency profile artifact requirements. Synthetic runs do not need this field and still cannot approve release evidence. |
 | `oracles` | A map with all five scenario names. Each value contains positive integer `objects`, `versions`, `bytes`, and `sha256` of the independently prepared canonical object/version/content manifest. |
 | `expected_healed_objects` | A map with all five scenario names and independently seeded repair counts. Running-heal and MRF-replay require a positive count. |

@@ -75,6 +76,14 @@ object/version/content result. Fix the foreground arrival rate (offered load),
 cache preparation procedure, configuration, and hardware across every leg.
 Do not include credentials in the manifest, adapter output, or saved commands;
 the collector reads `RUSTFS_ACCESS_KEY` and `RUSTFS_SECRET_KEY` from its environment.
+The adapter must echo the measured run's `release_evidence` object in every
+measurement response. A mismatch fails the cell because it means the deployment,
+mixed-version set, crash mode, or profiler contract no longer matches the
+operator-reviewed manifest. This echo is provenance binding only; it does not
+replace the independent correctness oracle, distributed metrics samples, profile
+artifacts, or ABBA comparison thresholds. The summary tool revalidates the same
+manifest contract before it can print a measured PASS result, so hand-built or
+trimmed reports without this provenance fail closed.

 #### Deployment Adapter Contract

@@ -150,6 +159,14 @@ delay counts are both retained so an operator can reject unrelated or
 process-lifetime counter contamination. Correct repair oracles and the existing
 regression limits still apply in every case.

+The `running-heal` build comparison also records a `w11` section for the bounded
+retry-window evidence. `status=observed` requires same-window healthy-page
+latency improvement, reduced heal lock-wait p99, bounded candidate RSS growth,
+and a candidate attempt-cost value. `rss_regression` means latency and lock-wait
+improved but RSS exceeded the allowed growth limit; `no_measured_benefit` means
+attempt-cost evidence exists without the full W11 benefit; `pending` means the
+attempt-cost evidence needed for the comparison is missing.
+
 For P2, `measure.convergence` contains booleans `writes_stopped`,
 `last_mutation_observed`, `first_complete_publication`; numeric
 `last_mutation_time`, `last_mutation_observed_time`, `writes_stopped_time`, `window_start`, `window_end`,
@@ -204,6 +221,17 @@ The command prints only `PASS scanner_heal_perf ...` for measured passing ABBA
 evidence, otherwise `FAIL scanner_heal_perf ...`. The JSON and Markdown outputs
 carry the key p99/throughput/P1/P2/cache-cost fields and artifact provenance
 hashes; raw per-cell logs remain in the original artifact tree for audit.
+Failed or interrupted ABBA reports that contain only `status`, `performance`,
+`completed_cells`, and `error` also summarize as `FAIL`; they do not become
+performance evidence, and a missing comparison matrix is accepted only for a
+non-passing report.
+Measured passing reports must also retain the W10/W11 foreground-pressure,
+heal-lock-wait, and heal-attempt-cost fields emitted by the ABBA evaluator. If
+those fields are removed, empty, malformed, or length-mismatched, the quiet
+summary fails closed instead of treating the report as performance evidence. For
+`running-heal` build comparisons, the summary additionally requires the emitted
+W11 section to be `observed` and to retain the RSS-growth, lock-wait,
+healthy-page-latency, and candidate attempt-cost values.

 They cover the complete 120-cell schedule, data isolation, missing builds and
 oracles, zero samples/requests, swallowed request errors, offered-load drift,
@@ -8,11 +8,21 @@

 | Method | Config tag | Credential lifetime | Background renewal | Recommended for |
 | --- | --- | --- | --- | --- |
-| Static token | `Token` | Whatever the operator provisioned; RustFS never renews it | None | Development; short-lived experiments |
+| Static token | `Token` | Whatever the operator provisioned; read from Vault at login | Renewed at half TTL when Vault reports the token as renewable | Development; short-lived experiments |
 | AppRole | `AppRole` | Lease-bound token obtained by login; renewed by RustFS | Renew at half TTL, re-login on failure | Production without a Vault Agent sidecar |
 | Kubernetes | `Kubernetes` | Lease-bound token obtained by login; renewed by RustFS | Renew at half TTL, re-login on failure | Production on Kubernetes, with no credential to distribute |
 | Agent token file | `TokenFile` | Owned by Vault Agent; RustFS only re-reads the sink file | File re-read once per poll interval | Production with a Vault Agent (or equivalent) managing auth |

+### Static token: what RustFS now knows about it
+
+`vault token create` grants a 768-hour TTL by default, so a static token normally *does* expire. At login RustFS calls `auth/token/lookup-self` and adopts whatever Vault reports:
+
+- **No expiry** (a root or periodic-root token, `ttl` 0): unchanged — no lease is tracked, no renewal task runs, and the token is never refused locally.
+- **Expiring and renewable:** the ordinary renewal loop takes over, renewing at half the remaining TTL and publishing the remaining-TTL gauge.
+- **Expiring but not renewable:** a `vault_static_token_not_renewable` warning is logged with the remaining TTL, the gauge is published, and requests fail closed inside the safety window rather than lapsing mid-flight against Vault. Rotate to a fresh token, or move to AppRole, Kubernetes, or an agent-managed token file.
+
+The probe never fails the login. A token whose policy omits `lookup-self` (Vault's `default` policy grants it), or a Vault that is unreachable at that moment, logs `vault_static_token_lookup_failed` and falls back to the previous behaviour — no lease tracked, no renewal — so a deployment that works today keeps working. That fallback lasts for the life of the client generation, so treat the warning as something to fix rather than tolerate.
+
 Exactly one method must be configured. Setting `RUSTFS_KMS_VAULT_TOKEN_FILE` together with any other method, or `RUSTFS_KMS_VAULT_KUBERNETES_ROLE` together with `RUSTFS_KMS_VAULT_APPROLE_ROLE_ID`, is rejected at startup with a configuration error because the effective identity would be ambiguous. A leftover `RUSTFS_KMS_VAULT_TOKEN` alongside a configured login method is tolerated and ignored, so a stale variable cannot silently downgrade the identity.

 All of these are read the same way whether the service is started with `RUSTFS_KMS_ENABLE=true` or configured later through `POST /rustfs/admin/v3/kms/configure`.
@@ -135,26 +135,39 @@ evidence registered in `.config/scanner-heal-required-tests.json`. It records
 already-built binaries and checks existing nextest output; it does not build,
 run tests, deploy servers, inject faults, or start another CI lane.

-The initial case is `background-target-restart`, emitted by
-`heal_erasure_disk_rebuild_test::tests::test_cluster_root_heal_recovers_remote_shards_after_background_target_restart`.
-That test already runs in `e2e-nightly`. When `RUSTFS_SCANNER_HEAL_RUN_DIR` is set,
-it checks the actual server and test-executable hashes against `run.json`, pins
-the same server binary for all node starts, and writes its oracle only after
-the real assertions pass. The artifact contains the actual pre/post target
+The registered cases are emitted by existing E2E tests. The original
+`background-target-restart` / `background-target-crash` cases run in
+`e2e-nightly` on a four-node, one-drive-per-node topology. The
+`ec84-target-drive-restart` case runs in `e2e-distributed` on a three-node,
+four-drive EC8+4 topology. When `RUSTFS_SCANNER_HEAL_RUN_DIR` is set, the
+producer checks the actual server and test-executable hashes against `run.json`,
+pins the same server binary for all node starts, and writes its oracle only
+after the real assertions pass. The artifact contains the actual pre/post target
 PIDs, per-node S3 listings, expected and downloaded complete-body hashes/lengths,
 and target-disk `VersionShardCensus` fingerprints. Existing baseline objects
 must match their pre-fault physical manifests; the object created during the
 outage has no pre-fault target shard and is checked for complete physical parts
 and exact S3 content.

-This case is a **four-node, one-drive-per-node process-restart test**. It is not
-power-loss validation, a 3x4 EC8+4 experiment, an all-version inventory, or proof
-of scanner enumeration, exact MRF disposition, legacy migration, or rollback.
-The registry keeps all G01-G14/P1-P4 and R-E/R-D/R-L release requirements pending
-until their actual feature-specific oracles and required topologies exist.
-Missing cases cannot be supplied by synthetic W20 results. W20's bounded JSON
-and file-hash helpers are reused; its ABBA performance contracts remain in
+These cases are still restart-focused evidence slices. They are not power-loss
+validation, an all-version inventory, or proof of scanner enumeration, exact MRF
+disposition, legacy migration, multi-pool/multi-set release coverage, or
+rollback.
+The schema 2 registry separates the implemented single-set restart lane from
+structured release lanes for authority coverage, checkpoint/crash, status and
+outcome, MRF responsibility, mixed-version rollback, scheduler pressure,
+maintenance producers, and EC8+4 multi-set coverage. All G01-G14/P1-P4 and
+R-E/R-D/R-L release requirements stay `pending` until their actual
+feature-specific oracles, measurements and required topologies exist. Missing
+cases cannot be supplied by synthetic W20 results. W20's bounded JSON and
+file-hash helpers are reused; its ABBA performance contracts remain in
 `docs/operations/scanner-benchmark-runbook.md`.
+Measured ABBA manifests must also carry the runbook's `release_evidence`
+contract. The runner rejects reports that cannot bind the exact 3x4 EC8+4
+topology, multi-pool/multi-set shape, distributed same-window metrics endpoints,
+restart/crash modes, mixed-version reader/writer/rollback participation, and
+allocation/flamegraph/RSS/save-frequency profile artifact plan. Synthetic runs
+and manifests missing that contract remain harness-only evidence.

 ### Recording One Case

@@ -170,27 +183,12 @@ The producer checks this compiled identity against the receipt; it does not
 copy a current source revision into an older test binary's identity. The E2E
 uses its existing temporary cluster directories and cleanup. `CARGO_TARGET_DIR`
 controls compilation output; nextest's default report store remains the
-workspace's `target/nextest`. Execute the existing selected case as follows:
+workspace's `target/nextest`. Prefer the registry-aware runner for concrete
+cases:

 ```bash
-CASE=background-target-restart
-FILTER='test(test_cluster_root_heal_recovers_remote_shards_after_background_target_restart)'
-RUN_DIR="$PWD/artifacts/scanner-heal-run"
-export RUSTFS_E2E_EXPECTED_FEATURES=default
-scripts/python_bin.sh scripts/check_test_wiring.py \
-  --begin-scanner-heal "$RUN_DIR" "$SERVER_BINARY" "$TEST_BINARY"
-export RUSTFS_SCANNER_HEAL_RUN_DIR="$RUN_DIR"
-export CARGO_BIN_EXE_rustfs="$SERVER_BINARY"
-cargo nextest list --profile e2e-nightly -p e2e_test -E "$FILTER" \
-  --message-format json > "$RUN_DIR/listing.json"
-rm -f target/nextest/e2e-nightly/junit.xml
-set +e
-cargo nextest run --profile e2e-nightly -p e2e_test -E "$FILTER"
-test_exit=$?
-set -e
-cp target/nextest/e2e-nightly/junit.xml "$RUN_DIR/junit.xml"
-scripts/python_bin.sh scripts/check_test_wiring.py --finish-scanner-heal "$RUN_DIR" "$test_exit"
-scripts/python_bin.sh scripts/check_test_wiring.py --check-scanner-heal "$RUN_DIR" "$CASE"
+scripts/run_scanner_heal_evidence_case.sh --case background-target-restart
+scripts/run_scanner_heal_evidence_case.sh --case ec84-target-drive-restart
 ```

 Set `RUSTFS_E2E_EXPECTED_FEATURES` to the actual intended e2e crate feature set,
@@ -234,15 +232,147 @@ For automation, `--check-scanner-heal-release "$RUN_DIR"` emits one compact
 JSON decision and exits nonzero while blocked. `verified_cases` contains only
 cases that pass the complete receipt, build provenance, nextest/JUnit and real
 oracle checks; `rejected_cases` names registered cases that do not, and
-`pending_gates` names the unimplemented release requirements. Approval requires
-every registered case to verify, `pending_gates` to be empty, and a future
-registry schema capable of representing the complete release matrix. Schema 1
-is deliberately marked `release_schema_capable: false`: it models only the
-single-version, unversioned-object restart/crash cases and cannot represent
-mixed-version, rollback, EC8+4 or performance evidence. A focused run,
-synthetic harness, compile-only result, skipped/retried test, ordinary CI
-success, or removal of pending text therefore cannot become a release approval.
+`pending_gates` names the unimplemented release requirements and
+`pending_lanes` names the structured release lanes that still need real
+evidence. Schema 1 is deliberately marked `release_schema_capable: false`
+because it models only the single-version, unversioned-object restart/crash
+cases. Schema 2 can describe the wider release matrix, but approval still
+requires every registered case to verify and every required gate to leave
+`pending` only after a future checker can bind it to real feature-specific
+evidence. The current checker hard-rejects missing structured requirements and
+pending gates mapped to an implemented lane, so clearing pending text cannot
+become approval. A focused run, synthetic harness, compile-only result,
+skipped/retried test, ordinary CI success, or unregistered mixed-version,
+rollback, EC8+4 or performance claim therefore cannot become a release approval.
+For high-risk rollback gates, `evidence_fields` records the specific proof
+fields that a future real-evidence checker must bind before a pending gate can
+move out of the blocked set. G03 keeps scoped ACK tied to durable root
+publication, ACK request identity, participating peer capability snapshots, and
+mixed-peer fallback oracles; G09 keeps mixed-version reader, writer, and rollback
+payload evidence explicit. These fields are part of the release contract, not
+evidence by themselves.
+
+The upgrade compatibility E2E can emit raw G09 JSON artifacts when
+`RUSTFS_SCANNER_HEAL_G09_EVIDENCE_DIR` points at a fresh, task-owned directory.
+The rolling mixed-version test writes `G09-mixed_version_reader_evidence.json`
+and `G09-mixed_version_writer_evidence.json` after the old/new reader and writer
+assertions pass. The bucket-metadata rollback test writes
+`G09-rollback_payload_evidence.json` after the current -> previous -> current
+round trip has read back the known bucket configuration and objects. These
+artifacts are measured inputs for a later release bundle; the bundle must still
+record their relative paths, hashes, command provenance, timestamps, roles,
+participating revisions, and case lists before
+`--check-scanner-heal-release-bundle` can validate them.
+
+For a release-candidate or PR-head Linux x86_64 host, run the full raw G09
+artifact pass with:
+
+```bash
+scripts/run_scanner_heal_g09_upgrade_evidence.sh
+```
+
+The script mirrors the pinned previous-release asset used by the upgrade
+workflow, builds the current checkout, runs the mixed-version and rollback E2E
+lanes, and fails unless all three raw G09 artifacts are measured, revision-bound,
+and role-bound. Use `--source-binary` for a custom previous-release binary on
+another platform, or `--test mixed-version|rollback` while narrowing a failure.
+It performs a free-space preflight before building so a saturated validation
+host fails before producing partial evidence.
+
+The W16 recovery-intent and quota-authority lanes can emit raw G04/G12 JSON
+artifacts with:
+
+```bash
+scripts/run_scanner_heal_w16_recovery_evidence.sh
+```
+
+The runner builds the current checkout, runs the scanner recovery-intent and
+disabled-startup crash-boundary tests, runs the scanner quota reset-preservation
+tests, and runs the distributed hard-quota admission E2E. A full run writes
+`release-bundle-w16.json` and validates the G04 and G12 gates with
+`--check-scanner-heal-release-bundle-gate`. Use `--test g04|g12` while narrowing
+a failure; a single gate descriptor still does not approve the complete release
+bundle.
+
+The W13 durable MRF replay lanes can emit raw G07/G08/P4 JSON artifacts with:
+
+```bash
+scripts/run_scanner_heal_w13_mrf_evidence.sh
+```
+
+The runner builds the current checkout, runs the ignored MRF evidence test, and
+writes `release-bundle-w13.json` for `--check-scanner-heal-release-bundle-gate`.
+Use `--test g07|g08|p4` while narrowing a failure. G08 disk-full evidence must
+run against a real fillable filesystem: on Linux as root the runner mounts a
+small tmpfs automatically, otherwise pass `--enospc-root` pointing at a
+pre-mounted small filesystem. P4 is release evidence only when it completes the
+default two-hour soak; `--allow-short-soak` is diagnostic and skips P4 bundle
+gate validation.
+
+When the real release lanes have produced their dedicated artifacts, validate
+the complete hard-gate bundle with:
+
+```bash
+scripts/python_bin.sh scripts/check_test_wiring.py \
+  --check-scanner-heal-release-bundle /path/to/release-evidence.json
+```
+
+The bundle checker is intentionally stricter than the case checker. It requires
+schema 2 registry metadata, `evidence: measured`, the current checkout revision,
+all G01-G14/P1-P4/R-E/R-D/R-L gates, per-gate `status: pass`, lane identity,
+relative artifact paths, matching SHA256 hashes, and non-empty summaries. It
+also binds each evidence field to its own run provenance: `source_revision`,
+`run_id`, `measurement_window_id`, timezone-qualified `started_at` and
+`finished_at`, command arguments, and artifact format. The field
+`source_revision` must match the bundle revision, and measured performance
+duration cannot exceed the recorded run window.
+When an evidence or profile artifact declares a JSON format, the checker also
+opens that artifact and requires its payload to repeat the same measured
+`source_revision`, `run_id`, `measurement_window_id`, gate and field identity;
+profile sub-artifacts must additionally name their artifact kind. Updating only
+the outer bundle hash cannot turn a stale JSON summary into current release
+evidence.
+
+The hard evidence shape remains claim-specific: mixed-version gates must name at
+least two participating versions, crash/durable replay gates must include
+crash-boundary evidence, G14 must record EC8+4 with at least three nodes and four
+drives per node plus multi-set and multi-pool evidence, performance gates need
+measured durations, P3's pressure run needs at least two hours, and P1 needs a
+symbolized profile summary with resolved samples. Every G14 field and every
+performance gate's fields must also share one `measurement_window_id`, so EC8+4,
+multi-set/multi-pool, ABBA, throughput, and profiling artifacts cannot be
+stitched together from unrelated runs. P1 `profile_evidence` must bind every
+required profile artifact kind (`allocation-profile`, `flamegraph`,
+`rss-samples`, and `save-frequency`) with a relative path, artifact format,
+non-empty file, matching SHA256, and the same measurement window when a
+per-artifact window is declared. Missing, synthetic, stale, tampered, undersized,
+or topology-mismatched evidence returns a compact blocked or invalid JSON result
+and a nonzero exit.
+
+The scheduler-pressure lane must also carry the numbers needed to close W09,
+W10, and W11: bounded deferred item/byte/age limits, zero duplicate tasks,
+pressure pacing engagement, recovery and lock-hold timings, fixed offered load,
+foreground p95/p99 latency, throughput, error count, attempt-cost samples, and
+completed heal object counts.
+
+This command validates the evidence package; it does not create evidence. A
+handwritten JSON file, a synthetic harness pass, a single focused case, or a
+local unit fixture still cannot satisfy the distributed, mixed-version,
+crash-restart, durable MRF replay, EC8+4, ABBA, or profiling gates.

 Run parser/receipt regressions with
 `scripts/python_bin.sh scripts/check_test_wiring.py --self-test`. Those fixtures
 validate the checker only and produce no runtime or performance evidence.
+
+For local bundle-shape dry runs, generate a task-owned fixture directory with:
+
+```bash
+scripts/python_bin.sh scripts/check_test_wiring.py \
+  --write-scanner-heal-release-bundle-fixture /path/to/fixture-dir
+```
+
+The generated file is marked `fixture_only` and is intentionally rejected by the
+release bundle checker. Use it to rehearse field names, artifact paths, hashes,
+profile artifact membership, mixed-version roles, and same-window provenance
+before copying the shape into a real measured bundle. It is not ABBA, profile,
+mixed-version, crash-restart, or release approval evidence.
@@ -77,4 +77,30 @@ cargo nextest run --profile e2e-distributed -p e2e_test -E 'not test(/^distribut

 The upgrade topology is `ClusterTopology::single_pool(4)` (4 nodes × 1 drive). That matches the proven mixed-version fixture in `upgrade_compatibility_test`; 4×4 localhost drives are rejected by the previous release's same-device disk check.

+Scanner/Heal G09 release evidence can be collected on Linux x86_64 with the
+same pinned previous-release binary used by `e2e-upgrade.yml`:
+
+```bash
+scripts/run_scanner_heal_g09_upgrade_evidence.sh
+```
+
+The runner builds the current `rustfs` binary, downloads and verifies the
+pinned previous release, runs the mixed-version rolling-upgrade and rollback
+tests, and leaves the required raw G09 artifacts under
+`target/scanner-heal-g09-upgrade-evidence/<timestamp>/`. These artifacts are
+inputs for the Scanner/Heal release bundle gate; the runner does not mark the
+full release matrix complete by itself.
+
+The distributed Scanner/Heal EC8+4 restart case is registered as
+`ec84-target-drive-restart` and selected by this profile:
+
+```bash
+scripts/run_scanner_heal_evidence_case.sh --case ec84-target-drive-restart
+```
+
+That command records the current build, runs exactly the registered
+`distributed::heal_test` case, validates the JUnit/listing/oracle receipt, and
+keeps the wider release gate blocked until the remaining release evidence lanes
+have measured artifacts.
+
 Membership is pinned by `.config/e2e-distributed-selection.txt`. Update the Linux and Darwin entries with `python3 ./scripts/check_test_wiring.py --update-profile e2e-distributed <listing.json> <platform>` after adding or renaming a case.
@@ -160,7 +160,7 @@ behavior.
 | config.rustfs.kms.vault.vault_backend | string | `""`| The vault backend, `vault-kv2` or `vault-transit`. |
 | config.rustfs.kms.vault.vault_address | string | `""`| The vault address. |
 | config.rustfs.kms.vault.vault_token | string | `""`| The vault token. Rendered into a dedicated Secret (`<fullname>-kms-secret`), never into the ConfigMap. |
-| config.rustfs.kms.vault.vault_mount_path | string | `"transit"`| The vault mount path, only works if `vault_backend` equals `vault-transit` . |
+| config.rustfs.kms.vault.vault_mount_path | string | `"transit"`| The vault mount path. Rendered as `RUSTFS_KMS_VAULT_MOUNT_PATH` for `vault-transit`, and as `RUSTFS_KMS_VAULT_KV_MOUNT` for `vault-kv2` (only when set; unset keeps the `secret` default). |
 | config.rustfs.kms.vault.default_key | string | `"transit"`| The master key id for RustFS. |
 | extraEnv | list | `[]` | Extra environment variables for the RustFS container. An explicit `RUSTFS_LOCAL_ENDPOINT_HOST` or `RUSTFS_VOLUMES`, or a bounded, dynamic, or unrecognized startup mode, disables generated anchor injection. `POD_NAME` and `RUSTFS_ADDRESS` remain independent overrides. |
 | extraVolumes | list | `[]` | Extra volumes to add to the pod spec. Supported in both standalone (Deployment) and distributed (StatefulSet) modes. |
@@ -23,6 +23,7 @@ use crate::admin::runtime_sources::{
    current_or_init_kms_runtime_service_manager,
 };
 use crate::admin::storage_api::config::{read_admin_config, save_admin_config};
+use crate::admin::storage_api::ecstore_topology::is_dist_erasure;
 use crate::admin::storage_api::error::StorageError;
 use crate::admin::storage_api::runtime::ECStore;
 use crate::admin::storage_api::s3::{S3ErrorCode, error as admin_s3_error};
@@ -507,6 +508,21 @@ pub async fn reload_persisted_kms_config() -> Result<(), String> {
    reload_persisted_kms_config_from_store(store, kms_service_manager_from_context(), "peer_reload").await
 }

+/// Whether a reload may return early because this node is already serving
+/// exactly the persisted configuration.
+///
+/// Byte-identical configuration is not sufficient on its own. A node whose KMS
+/// failed to start keeps its configuration and sits in `Error`, so comparing
+/// only the bytes turned the documented recovery call
+/// (`POST /rustfs/admin/v3/kms/reload`) into a no-op that reported success and
+/// left the node down — including on every peer, which reaches this same
+/// function through the reload broadcast (backlog#2369 P1). Any state other
+/// than `Running` falls through to `reconfigure`, which starts the service when
+/// none is running.
+fn kms_reload_is_already_current(status: rustfs_kms::KmsServiceStatus, config_is_unchanged: bool) -> bool {
+    matches!(status, rustfs_kms::KmsServiceStatus::Running) && config_is_unchanged
+}
+
 async fn reload_persisted_kms_config_from_store(
    store: Arc<ECStore>,
    service_manager: Arc<rustfs_kms::KmsServiceManager>,
@@ -525,11 +541,11 @@ async fn reload_persisted_kms_config_from_store(
        return Err("no persisted KMS configuration is available".to_string());
    };

-    if service_manager
+    let config_is_unchanged = service_manager
        .get_config()
        .await
-        .is_some_and(|current| kms_config_is_unchanged(&current, &config))
-    {
+        .is_some_and(|current| kms_config_is_unchanged(&current, &config));
+    if kms_reload_is_already_current(service_manager.get_status().await, config_is_unchanged) {
        info!(
            event = "kms_service_state",
            component = LOG_COMPONENT_ADMIN,
@@ -624,6 +640,51 @@ fn local_success_with_peer_report(message: &str, unconverged: &[String]) -> (boo
    )
 }

+/// What a node-local KMS backend means for a multi-node deployment
+/// (backlog#2369 P7.4).
+///
+/// The Local backend keeps key material on each node's own disk and generates
+/// its KDF salt per node, so two nodes derive different keys from the same
+/// `master_key`. An object encrypted on node A cannot be decrypted on node B:
+/// behind a load balancer that shows up as intermittent 500s on reads that
+/// worked a moment earlier. The product decision to warn rather than refuse
+/// stands; the generic "development only" warning simply never said what
+/// actually goes wrong, so an operator had no way to connect the symptom to
+/// the cause.
+///
+/// Returns the sentence to append to the configure response, or `None` when the
+/// combination does not apply.
+async fn node_local_backend_warning(backend: &rustfs_kms::KmsBackend) -> Option<&'static str> {
+    if !matches!(backend, rustfs_kms::KmsBackend::Local) || !is_dist_erasure().await {
+        return None;
+    }
+
+    warn!(
+        component = LOG_COMPONENT_ADMIN,
+        subsystem = LOG_SUBSYSTEM_KMS,
+        event = "kms_node_local_backend_in_distributed_deployment",
+        backend = rustfs_kms::KmsBackend::Local.as_str(),
+        "The Local KMS backend stores key material on each node's own disk with a per-node salt, so objects \
+         encrypted on one node cannot be decrypted on another. In a distributed deployment this surfaces as \
+         intermittent 500s on reads behind a load balancer. Use Vault Transit, Vault KV2 or AWS KMS for a \
+         multi-node deployment"
+    );
+
+    Some(
+        "Warning: the Local KMS backend is node-local. Key material and its salt live on each node's own disk, so \
+         objects encrypted on one node cannot be decrypted on another and reads behind a load balancer will fail \
+         intermittently. Use Vault Transit, Vault KV2 or AWS KMS for a distributed deployment",
+    )
+}
+
+/// Append the node-local backend warning to a successful configure message.
+fn with_node_local_backend_warning(message: String, warning: Option<&'static str>) -> String {
+    match warning {
+        Some(warning) => format!("{message}. {warning}"),
+        None => message,
+    }
+}
+
 pub fn register_kms_dynamic_route(r: &mut S3Router<AdminOperation>) -> std::io::Result<()> {
    r.insert(
        Method::POST,
@@ -746,6 +807,7 @@ impl Operation for ConfigureKmsHandler {
        let kms_config = configure_request.to_kms_config();

        let persisted_config = kms_config.clone();
+        let node_local_warning = node_local_backend_warning(&kms_config.backend).await;
        let (success, message, status) = match service_manager
            .configure_with_persistence(kms_config, || async move {
                save_kms_config(&persisted_config)
@@ -768,7 +830,7 @@ impl Operation for ConfigureKmsHandler {
                let unconverged = broadcast_kms_config_reload().await;
                let (success, message) = local_success_with_peer_report("KMS configured successfully", &unconverged);
                audit.finish(KmsAdminOperation::Configure, None, None);
-                (success, message, status)
+                (success, with_node_local_backend_warning(message, node_local_warning), status)
            }
            Err(e) => {
                let error_msg = format!("Failed to configure KMS: {e}");
@@ -1345,6 +1407,7 @@ impl Operation for ReconfigureKmsHandler {
        let kms_config = configure_request.to_kms_config();

        let persisted_config = kms_config.clone();
+        let node_local_warning = node_local_backend_warning(&kms_config.backend).await;
        let (success, message, status) = match service_manager
            .reconfigure_with_persistence(kms_config, || async move {
                save_kms_config(&persisted_config)
@@ -1368,7 +1431,7 @@ impl Operation for ReconfigureKmsHandler {
                let (success, message) =
                    local_success_with_peer_report("KMS reconfigured and restarted successfully", &unconverged);
                audit.finish(KmsAdminOperation::Reconfigure, None, None);
-                (success, message, status)
+                (success, with_node_local_backend_warning(message, node_local_warning), status)
            }
            Err(e) => {
                let error_msg = format!("Failed to reconfigure KMS: {e}");
@@ -1420,9 +1483,10 @@ impl Operation for ReconfigureKmsHandler {
 mod tests {
    use super::{
        KmsConfigLoadError, decode_persisted_kms_config, ensure_kms_config_persistable, ensure_kms_request_persistable,
-        kms_config_fingerprint, kms_config_is_unchanged, kms_configure_actions, kms_service_control_actions,
-        load_kms_config_with, local_success_with_peer_report, normalize_configure_request_secrets, open_persisted_kms_config,
-        redacted_canonical_config, register_kms_dynamic_route, seal_persisted_kms_config,
+        kms_config_fingerprint, kms_config_is_unchanged, kms_configure_actions, kms_reload_is_already_current,
+        kms_service_control_actions, load_kms_config_with, local_success_with_peer_report, normalize_configure_request_secrets,
+        open_persisted_kms_config, redacted_canonical_config, register_kms_dynamic_route, seal_persisted_kms_config,
+        with_node_local_backend_warning,
    };
    use crate::admin::router::{AdminOperation, S3Router};
    use crate::admin::storage_api::error::StorageError;
@@ -1432,6 +1496,49 @@ mod tests {
    use std::path::PathBuf;
    use tempfile::TempDir;

+    /// backlog#2369 P1: a node whose KMS failed to start keeps its persisted
+    /// configuration, so an unchanged-bytes comparison made the documented
+    /// recovery call a no-op that still reported success.
+    #[test]
+    fn kms_reload_only_short_circuits_for_a_running_service() {
+        use rustfs_kms::KmsServiceStatus;
+
+        assert!(
+            kms_reload_is_already_current(KmsServiceStatus::Running, true),
+            "a running service on identical configuration has nothing to apply"
+        );
+        assert!(
+            !kms_reload_is_already_current(KmsServiceStatus::Running, false),
+            "changed configuration must always be applied"
+        );
+
+        for status in [
+            KmsServiceStatus::NotConfigured,
+            KmsServiceStatus::Configured,
+            KmsServiceStatus::Error("vault unreachable at startup".to_string()),
+        ] {
+            assert!(
+                !kms_reload_is_already_current(status.clone(), true),
+                "reload must reconfigure instead of reporting success from {status:?}"
+            );
+        }
+    }
+
+    /// backlog#2369 P7.4: the operator has to learn the consequence from the
+    /// response, not just from a log line the configuring client never sees.
+    #[test]
+    fn a_node_local_backend_warning_reaches_the_configure_response() {
+        let plain = with_node_local_backend_warning("KMS configured successfully".to_string(), None);
+        assert_eq!(plain, "KMS configured successfully");
+
+        let warned = with_node_local_backend_warning(
+            "KMS configured successfully".to_string(),
+            Some("Warning: the Local KMS backend is node-local"),
+        );
+        assert!(warned.starts_with("KMS configured successfully."), "{warned}");
+        assert!(warned.contains("node-local"), "{warned}");
+    }
+
    fn assert_has_action(actions: &[Action], action: Action) {
        assert!(actions.contains(&action), "expected action list to contain {action:?}");
    }
@@ -75,6 +75,10 @@ struct ScannerRecoveryIntentResponse {
    mode: String,
    intent_id: String,
    state: String,
+    actor_sha256: String,
+    idempotency_key_sha256: String,
+    request_sha256: String,
+    accepted_at_unix_secs: u64,
 }

 #[derive(Debug, Serialize)]
@@ -332,6 +336,10 @@ fn scanner_recovery_intent_record_response(
        mode: record.mode,
        intent_id: record.intent_id,
        state: record.state,
+        actor_sha256: record.actor_sha256,
+        idempotency_key_sha256: record.idempotency_key_sha256,
+        request_sha256: record.request_sha256,
+        accepted_at_unix_secs: record.accepted_at_unix_secs,
    };
    let body = serde_json::to_vec(&response).map_err(|err| {
        S3Error::with_message(
@@ -371,10 +379,7 @@ fn scanner_recovery_intent_accept_response(

 fn scanner_recovery_intent_executor_id(result: &rustfs_scanner::ScannerRecoveryIntentAcceptResult) -> Option<String> {
    match result {
-        rustfs_scanner::ScannerRecoveryIntentAcceptResult::Accepted { record }
-        | rustfs_scanner::ScannerRecoveryIntentAcceptResult::Replayed { record }
-            if matches!(record.state.as_str(), "accepted" | "running") =>
-        {
+        rustfs_scanner::ScannerRecoveryIntentAcceptResult::Accepted { record } if record.state == "accepted" => {
            Some(record.intent_id.clone())
        }
        _ => None,
@@ -673,7 +678,7 @@ mod tests {
    }

    #[test]
-    fn scanner_recovery_intent_executor_only_starts_non_terminal_work() {
+    fn scanner_recovery_intent_executor_starts_only_newly_accepted_work() {
        let mut record = rustfs_scanner::ScannerRecoveryIntentRecord {
            schema_version: 1,
            intent_id: "0".repeat(64),
@@ -699,7 +704,16 @@ mod tests {
                record: record.clone(),
            })
            .as_deref(),
-            Some(record.intent_id.as_str())
+            None,
+            "a lost-response retry must not start a duplicate executor"
+        );
+        record.state = "accepted".to_string();
+        assert!(
+            scanner_recovery_intent_executor_id(&rustfs_scanner::ScannerRecoveryIntentAcceptResult::Replayed {
+                record: record.clone(),
+            })
+            .is_none(),
+            "replayed accepted records remain durable for startup/control recovery instead of duplicating work"
        );
        record.state = "completed".to_string();
        assert!(
@@ -41,6 +41,12 @@ pub(crate) mod ecstore_cluster {
    };
 }

+/// Deployment topology. The KMS configure path needs it to tell an operator
+/// that a node-local backend cannot serve a multi-node deployment.
+pub(crate) mod ecstore_topology {
+    pub(crate) use crate::storage::storage_api::is_dist_erasure;
+}
+
 mod ecstore_config {
    pub(crate) use crate::storage::storage_api::ecstore_config::{com, init, storageclass};
 }
@@ -775,6 +775,7 @@ mod tests {
            scan_plan_digest: Some([1; 32]),
            complete: true,
            tombstone: false,
+            segment_invalidation_proof: None,
        }];
        DefaultAdminUsecase::narrow_data_usage_snapshot_to_measured_buckets(&mut info, ["bucket-a".to_string()]);
        assert_eq!(info.usage_snapshot_converged, Some(false));
@@ -1389,7 +1389,10 @@ impl DefaultBucketUsecase {
        counter!("rustfs_create_bucket_total").increment(1);
        let result = Ok(S3Response::new(output));
        let _ = helper.complete(&result);
-        rustfs_scanner::record_dirty_usage_bucket(&bucket);
+        rustfs_scanner::record_dirty_usage_bucket_from_producer(
+            &bucket,
+            rustfs_scanner::SegmentInvalidationProducerIdentity::BucketMetadata,
+        );
        result
    }

@@ -1781,7 +1784,10 @@ impl DefaultBucketUsecase {
            warn!(bucket = %bucket, error = ?err, "site replication bucket tagging delete hook failed");
        }

-        rustfs_scanner::record_dirty_usage_bucket(&bucket);
+        rustfs_scanner::record_dirty_usage_bucket_from_producer(
+            &bucket,
+            rustfs_scanner::SegmentInvalidationProducerIdentity::BucketMetadata,
+        );
        Ok(S3Response::new(DeleteBucketTaggingOutput {}))
    }

@@ -2698,7 +2704,10 @@ impl DefaultBucketUsecase {
            warn!(bucket = %bucket, error = ?err, "site replication bucket tagging hook failed");
        }

-        rustfs_scanner::record_dirty_usage_bucket(&bucket);
+        rustfs_scanner::record_dirty_usage_bucket_from_producer(
+            &bucket,
+            rustfs_scanner::SegmentInvalidationProducerIdentity::BucketMetadata,
+        );
        Ok(S3Response::new(PutBucketTaggingOutput::default()))
    }

@@ -2733,7 +2742,10 @@ impl DefaultBucketUsecase {
            warn!(bucket = %bucket, error = ?err, "site replication bucket versioning hook failed");
        }

-        rustfs_scanner::record_dirty_usage_bucket(&bucket);
+        rustfs_scanner::record_dirty_usage_bucket_from_producer(
+            &bucket,
+            rustfs_scanner::SegmentInvalidationProducerIdentity::BucketMetadata,
+        );
        Ok(S3Response::new(PutBucketVersioningOutput {}))
    }

--- a/Show More
+++ b/Show More