fix(storage): queue multipart parts for foreground write permits (#7337)

Multipart parts shared the 250 ms direct-PutObject wait on the foreground
write permit pool, so SDK-default concurrency (many parts per upload in
flight at once) was rejected wholesale with SlowDown at stock settings.
Keep the pool that bounds in-flight bodies, but let parts wait in a
bounded queue with their own timeout before body ingest, report the
queue depth in the ForegroundWrite admission snapshot, and document the
foreground write admission environment variables.

.config/e2e-distributed-selection.txt

@@ -0,0 +1,2 @@
+sha256-linux=4696a43b167ac608b3b8677027c9fe9fdac3396d37c8cca11dce531c720ac6d2
+sha256-darwin=9785867929047dfd8c6f768e0d2b1e0a8fdba85216f4a4139093b1619d03ff07

.config/e2e-full-selection.txt

@@ -1,2 +1,2 @@
-sha256-darwin=a881fd7d3f5cb94654221ca85b8b30cce1b95e608824a55a15339cbc294e6d34
-sha256-linux=a2933d83dfe74ffa03410a0959333a1c48288b8469ca9f17273d449d7510c24b
+sha256-darwin=53b05ac745905809d3828c6994bdd8ecf9d20b2b61a8a9d80fe15eb62f932193
+sha256-linux=7c892afa4b9d1591b46bd79c976b647109a277284fddb3b98edced4b0297eda2

.config/e2e-smoke-selection.txt

@@ -1 +1 @@
-sha256=a2542dc86bbff56b2177efc621785c56fa7e8d813b209b7d935e1e41a9f0ad15
+sha256=5db88c6fec94d4f269c7d9cfc128bd2adc27b3d7021127e2fa0b1daccc5f900f

.config/ecstore-required-tests.json

@@ -40,6 +40,21 @@
       "invariant": "corrupt-part-arrays",
       "suite": "rustfs-filemeta",
       "name": "filemeta::test::crc_valid_but_part_arrays_corrupt_into_fileinfo_errors_not_panics"
+    },
+    {
+      "invariant": "odm-source-contract-s3",
+      "suite": "rustfs",
+      "name": "on_demand_migration::source_client::tests::s3_backend_satisfies_the_shared_backend_contract"
+    },
+    {
+      "invariant": "odm-source-contract-azure",
+      "suite": "rustfs",
+      "name": "on_demand_migration::azure::tests::azure_backend_satisfies_the_shared_backend_contract"
+    },
+    {
+      "invariant": "odm-source-contract-gcs",
+      "suite": "rustfs",
+      "name": "on_demand_migration::gcs::tests::gcs_native_backend_satisfies_the_shared_backend_contract"
     }
   ],
   "fixtures": [

.config/make/tests.mak

@@ -31,6 +31,7 @@ script-tests: ## Run shell script tests
 	./scripts/test_object_batch_bench_enhanced.sh
 	./scripts/test_hotpath_warp_ab_gate.sh
 	./scripts/test_hotpath_warp_abba.sh
+	./scripts/test_scanner_validation_harness.sh
 	./scripts/test_exact_1mib_handoff_abba.sh
 	./scripts/test_pinned_paired_abba_bench.sh
 	./scripts/test_manual_transition_runbooks.sh

.config/migration-gate-floor.txt

@@ -1,10 +1,11 @@
 # Committed floor for the number of tests selected by the migration-critical
 # CI gate (see scripts/check_migration_gate_count.sh, backlog#1153 infra-12).
 #
-# The floor equals the exact count of rustfs-ecstore --lib tests matching the
-# gate filter (name substrings: data_movement, rebalance, decommission,
-# source_cleanup, delete_marker) at the time this file was last updated.
+# The floor equals the exact count of rustfs-ecstore --lib tests, with the
+# test-util feature enabled, matching the gate filter (name substrings:
+# data_movement, rebalance, decommission, source_cleanup, delete_marker) at
+# the time this file was last updated.
 # CI fails if the selected count drops below this number, so renames or
 # removals that thin the gate must update this file in the same PR.
 # Adding tests does not require a bump, but bumping keeps the guard tight.
-571
+946

.config/nextest.toml

@@ -183,6 +183,13 @@ test-group = 'e2e-reliability'
 filter = 'package(e2e_test) & test(/^inline_fast_path_cluster_test::/)'
 test-group = 'e2e-inline-boundaries'
 
+# 4-node 4-drive distributed Actions suite: each case starts four rustfs
+# processes and up to sixteen data directories. Serialize across nextest's
+# process boundary so several 4x4 clusters never overlap.
+[[profile.default.overrides]]
+filter = 'package(e2e_test) & test(/^distributed::/)'
+test-group = 'e2e-cluster-nightly'
+
 # Vault KMS tests share the fixed dev-server port 8200. serial_test's #[serial]
 # does not cross nextest process boundaries, so keep every Vault-backed test in
 # one group.
@@ -526,6 +533,27 @@ path = "junit.xml"
 filter = 'package(e2e_test)'
 test-group = 'e2e-cluster-nightly'
 
+# ---------------------------------------------------------------------------
+# e2e-distributed profile — 4-node 4-disk Actions suite
+# ---------------------------------------------------------------------------
+# Storage-sensitive PR / nightly / dispatch lane owned by
+# .github/workflows/e2e-distributed.yml.
+# Each case starts four rustfs processes (and for site replication, two
+# clusters). Upgrade cases also require RUSTFS_UPGRADE_SOURCE_BINARY.
+# Serialized via e2e-cluster-nightly with no retries.
+[profile.e2e-distributed]
+default-filter = 'package(e2e_test) & test(/^distributed::/)'
+fail-fast = false
+# Decommission / rebalance cases poll for up to 180s with little stdout.
+slow-timeout = { period = "120s", terminate-after = 6 }
+
+[profile.e2e-distributed.junit]
+path = "junit.xml"
+
+[[profile.e2e-distributed.overrides]]
+filter = 'package(e2e_test)'
+test-group = 'e2e-cluster-nightly'
+
 # ---------------------------------------------------------------------------
 # e2e-odm-interop profile — on-demand migration provider interop lane (ODM-20)
 # ---------------------------------------------------------------------------
@@ -586,6 +614,10 @@ path = "junit.xml"
 #     cluster-fault lane. heal_erasure_disk_rebuild is intentionally not
 #     excluded here because backlog#2213 promotes core heal rebuild coverage to
 #     this merge/main lane while retaining nightly coverage.
+#   * distributed:: — 4-node 4-disk Actions suite (S3, lock, versioning,
+#     replication, quota, observability, expand/decommission/rebalance, site
+#     replication, chaos, upgrade history/IAM). Owns [profile.e2e-distributed] and
+#     .github/workflows/e2e-distributed.yml.
 #   * on_demand_migration::interop_test — the ODM-20 provider interoperability
 #     cases, which are meaningless without a source: they run in the dedicated
 #     [profile.e2e-odm-interop] lane below, where the workflow points them at a
@@ -607,6 +639,7 @@ default-filter = """
     package(e2e_test)
     & !test(/^protocols::/)
     & !test(/^(admin_timeout_regression_test|cluster_concurrency_test|cluster_multidrive_pool_test|degraded_listing_availability_test|namespace_lock_quorum_test|object_lambda_test|stale_multipart_cleanup_cluster_test)::/)
+    & !test(/^distributed::/)
     & !test(/^replication_extension_test::/)
     & !test(/^replication_target_matrix_test::/)
     & !test(/^on_demand_migration::(concurrency_test|fault_test|interop_test|real_source_test)::/)

.config/scanner-heal-required-tests.json

@@ -0,0 +1,40 @@
+{
+  "schema": 1,
+  "cases": {
+    "background-target-restart": {
+      "gate": "G14",
+      "task": "W21",
+      "lane": "e2e-nightly",
+      "suite": "e2e_test",
+      "name": "heal_erasure_disk_rebuild_test::tests::test_cluster_root_heal_recovers_remote_shards_after_background_target_restart",
+      "oracle": "background-target-restart.json",
+      "min_objects": 9,
+      "max_objects": 65,
+      "topology": {"nodes": 4, "drives_per_node": 1},
+      "scope": "Target process restart, exact unversioned S3 bodies and replacement-disk shards; not power loss or EC8+4."
+    }
+  },
+  "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"
+  }
+}

.github/scheduled-validations.json

@@ -4,6 +4,11 @@
   { "workflow": ".github/workflows/ci.yml", "max_age_hours": 192 },
   { "workflow": ".github/workflows/coverage.yml", "max_age_hours": 192 },
   { "workflow": ".github/workflows/e2e-replication-nightly.yml", "max_age_hours": 36 },
+  {
+    "workflow": ".github/workflows/e2e-distributed.yml",
+    "max_age_hours": 36,
+    "never_ran_grace_until": "2026-09-18T00:00:00Z"
+  },
   { "workflow": ".github/workflows/e2e-s3tests.yml", "max_age_hours": 192 },
   { "workflow": ".github/workflows/fuzz.yml", "max_age_hours": 36 },
   { "workflow": ".github/workflows/mint.yml", "max_age_hours": 192 },

.github/workflows/ci.yml

@@ -850,6 +850,17 @@ jobs:
           cache-save-if: 'false'
           install-build-packaging-tools: 'false'
 
+      - name: Install network fault-injection tools
+        run: |
+          sudo apt-get install -y iptables
+          sudo -n iptables --version
+          # The endpoint-blackhole heal scenario needs CAP_NET_ADMIN. Containerised
+          # runners can run iptables but not touch the rule set; the test then logs
+          # a skip instead of failing, so surface that here where it is visible.
+          if ! sudo -n iptables -w 5 -S OUTPUT >/dev/null 2>&1; then
+            echo "::warning::iptables cannot read the OUTPUT chain on this runner (no CAP_NET_ADMIN); the endpoint-blackhole heal scenario will be skipped"
+          fi
+
       - name: Set up Python
         uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6.3.0
         with:

.github/workflows/docker.yml

@@ -121,6 +121,22 @@ jobs:
           create_latest=false
           source_ref="$GITHUB_SHA"
 
+          # Pre-GA policy: until the first stable (vX.Y.Z) tag exists, every
+          # prerelease (alpha/beta/rc) also moves `latest`, so users pulling
+          # `latest` get the newest test build. Once a stable tag is published
+          # this returns false and `latest` follows stable releases only.
+          prerelease_moves_latest() {
+            local stable_tags
+            stable_tags=$(git ls-remote --tags --refs origin 2>/dev/null \
+              | awk '{print $2}' \
+              | grep -E '^refs/tags/v?[0-9]+\.[0-9]+\.[0-9]+$' || true)
+            if [[ -z "$stable_tags" ]]; then
+              return 0
+            fi
+            echo "ℹ️  Stable release tag(s) already exist; prereleases no longer update latest"
+            return 1
+          }
+
           if [[ "${{ github.event_name }}" == "workflow_run" ]]; then
             # Triggered by build workflow completion
             echo "🔗 Triggered by build workflow completion"
@@ -184,8 +200,8 @@ jobs:
                 if [[ "$version" == *"alpha"* ]] || [[ "$version" == *"beta"* ]] || [[ "$version" == *"rc"* ]]; then
                   build_type="prerelease"
                   is_prerelease=true
-                  # Current policy: create latest tags for stable releases and selected prereleases (alpha/beta).
-                  if [[ "$version" == *"alpha"* ]] || [[ "$version" == *"beta"* ]]; then
+                  # Pre-GA policy: prereleases update latest until the first stable tag exists.
+                  if prerelease_moves_latest; then
                     create_latest=true
                     echo "🧪 Building Docker image for prerelease: $version (creating latest tag)"
                   else
@@ -243,8 +259,8 @@ jobs:
               v*alpha*|v*beta*|v*rc*|*alpha*|*beta*|*rc*)
                 build_type="prerelease"
                 is_prerelease=true
-                # Current policy: create latest tags for stable releases and selected prereleases (alpha/beta).
-                if [[ "$version" == *"alpha"* ]] || [[ "$version" == *"beta"* ]]; then
+                # Pre-GA policy: prereleases update latest until the first stable tag exists.
+                if prerelease_moves_latest; then
                   create_latest=true
                   echo "🧪 Building with prerelease version: $input_version (creating latest tag)"
                 else
@@ -394,11 +410,13 @@ jobs:
           TAG_BASE="${VERSION}${VARIANT_SUFFIX}"
           TAGS="${{ env.REGISTRY_DOCKERHUB }}:$TAG_BASE,${{ env.REGISTRY_GHCR }}:$TAG_BASE,${{ env.REGISTRY_QUAY }}:$TAG_BASE"
 
-          # Add channel tags for prereleases and latest for stable
+          # Add latest when requested (stable releases, and prereleases before GA)
           if [[ "$CREATE_LATEST" == "true" ]]; then
-            # Create latest tags for stable releases and selected prereleases when CREATE_LATEST=true.
             TAGS="$TAGS,${{ env.REGISTRY_DOCKERHUB }}:latest${VARIANT_SUFFIX},${{ env.REGISTRY_GHCR }}:latest${VARIANT_SUFFIX},${{ env.REGISTRY_QUAY }}:latest${VARIANT_SUFFIX}"
-          elif [[ "$BUILD_TYPE" == "prerelease" ]]; then
+          fi
+
+          # Always add the channel tag for prereleases, independent of latest
+          if [[ "$BUILD_TYPE" == "prerelease" ]]; then
             # Prerelease channel tags (alpha, beta, rc)
             if [[ "$VERSION" == *"alpha"* ]]; then
               CHANNEL="alpha"
@@ -555,7 +573,7 @@ jobs:
             "prerelease")
               echo "🧪 Prerelease Docker image has been built with ${VERSION} tags"
               echo "⚠️  This is a prerelease image - use with caution"
-              # Create latest tags for stable releases and selected prereleases when CREATE_LATEST=true.
+              # Prereleases move latest until the first stable tag exists (pre-GA policy).
               if [[ "$CREATE_LATEST" == "true" ]]; then
                 echo "🏷️  Latest tag has been created for prerelease: $VERSION"
               else

.github/workflows/e2e-distributed.yml

@@ -0,0 +1,216 @@
+# 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/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.
+
+# 4-node 4-disk distributed e2e lane.
+#
+# Each selected test starts a real localhost cluster via
+# `RustFSTestClusterEnvironment` (4 processes; 4 drives per node unless the
+# case is a two-site 4-node 1-drive pair or a 4-node upgrade). Membership is
+# `[profile.e2e-distributed]` in `.config/nextest.toml`. Storage-sensitive PRs,
+# nightly runs, and manual dispatches all execute the same fail-closed suite.
+# Upgrade cases download the same pinned previous release as e2e-upgrade.yml.
+#
+# Isolated pool filesystems: expand/decommission/rebalance cases require
+# independent `statfs` capacity. This job runs on GitHub-hosted
+# `ubuntu-latest` because the self-hosted `sm-standard-4` ARC pods cannot
+# create filesystems: `mount -o loop` fails with ENOENT (no
+# `/dev/loop-control`), and `mount -t tmpfs` fails with "cannot mount tmpfs
+# read-only" (no `CAP_SYS_ADMIN` in the initial namespace). The same reason
+# `uring-integration` and `e2e-s3tests.yml` left that label. The prepare
+# step mounts four 1 GiB tmpfs instances and exports `RUSTFS_E2E_POOL_ROOTS`.
+
+name: e2e-distributed
+
+on:
+  pull_request:
+    paths:
+      - "Cargo.lock"
+      - "Cargo.toml"
+      - ".config/nextest.toml"
+      - ".github/workflows/e2e-distributed.yml"
+      - "crates/audit/**"
+      - "crates/common/**"
+      - "crates/config/**"
+      - "crates/e2e_test/**"
+      - "crates/ecstore/**"
+      - "crates/filemeta/**"
+      - "crates/heal/**"
+      - "crates/iam/**"
+      - "crates/lock/**"
+      - "crates/madmin/**"
+      - "crates/notify/**"
+      - "crates/replication/**"
+      - "crates/s3-client/**"
+      - "crates/s3-ops/**"
+      - "crates/s3-types/**"
+      - "crates/scanner/**"
+      - "crates/storage-api/**"
+      - "crates/utils/**"
+      - "rustfs/**"
+  workflow_dispatch:
+    inputs:
+      filter:
+        description: "Optional nextest -E filter (default: the whole e2e-distributed profile)"
+        required: false
+        default: ""
+  schedule:
+    # 05:53 UTC nightly — clear of e2e-nightly (04:29) and ODM interop (05:23).
+    - cron: "53 5 * * *"
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: ${{ github.event_name != 'schedule' }}
+
+jobs:
+  distributed:
+    name: Distributed 4-node 4-disk e2e
+    # GitHub-hosted VM: loop and tmpfs mounts work here. sm-standard-4 is an
+    # ARC pod and rejects both (`mount -o loop` ENOENT, tmpfs "read-only").
+    runs-on: ubuntu-latest
+    timeout-minutes: 180
+    env:
+      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
+      NO_PROXY: 127.0.0.1,localhost
+      HTTP_PROXY: ""
+      HTTPS_PROXY: ""
+      # Pinned previous release used by distributed::upgrade_test (same pin as e2e-upgrade.yml).
+      UPGRADE_SOURCE_VERSION: 1.0.0-rc.2
+      UPGRADE_SOURCE_ASSET: rustfs-linux-x86_64-gnu-v1.0.0-rc.2.zip
+      UPGRADE_SOURCE_SHA256: 7c789386bf85278f865b8e0d359bf4edb84d5aa408cc3fa54a18c25ca74cd6e7
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+
+      - name: Setup Rust environment
+        uses: ./.github/actions/setup
+        with:
+          rust-version: stable
+          # Dedicated key: ubuntu-latest and sm-standard-4 share runner.os, so
+          # a shared key would mix VM and ARC pod target/ artifacts.
+          cache-shared-key: ci-e2e-distributed-hosted
+          cache-save-if: ${{ github.ref == 'refs/heads/main' }}
+          install-build-packaging-tools: 'false'
+
+      - name: Prepare isolated filesystems for pool movement
+        run: |
+          set -euo pipefail
+          mount_base="${RUNNER_TEMP}/rustfs-e2e-pools"
+          mkdir -p "${mount_base}"
+          roots=()
+          for pool in 0 1 2 3; do
+            mountpoint="${mount_base}/pool-${pool}"
+            mkdir -p "${mountpoint}"
+            # Sized tmpfs reports a distinct st_dev and independent 1G
+            # statfs capacity. Requires a VM runner (ubuntu-latest).
+            if ! sudo mount -t tmpfs -o size=1G,nosuid,nodev,mode=1777 tmpfs "${mountpoint}"; then
+              echo "tmpfs mount failed on $(uname -a)" >&2
+              findmnt || true
+              grep Cap /proc/self/status || true
+              exit 1
+            fi
+            sudo chmod 1777 "${mountpoint}"
+            roots+=("${mountpoint}")
+          done
+          printf -v joined_roots '%s:' "${roots[@]}"
+          echo "RUSTFS_E2E_POOL_ROOTS=${joined_roots%:}" >> "${GITHUB_ENV}"
+          findmnt --noheadings --output TARGET,SOURCE,FSTYPE,SIZE --target "${roots[0]}"
+          findmnt --noheadings --output TARGET,SOURCE,FSTYPE,SIZE --target "${roots[1]}"
+          findmnt --noheadings --output TARGET,SOURCE,FSTYPE,SIZE --target "${roots[2]}"
+          findmnt --noheadings --output TARGET,SOURCE,FSTYPE,SIZE --target "${roots[3]}"
+
+      - name: Download pinned previous release
+        env:
+          SOURCE_DIR: ${{ runner.temp }}/rustfs-upgrade-source
+        run: |
+          set -euo pipefail
+          mkdir -p "$SOURCE_DIR"
+          archive="$SOURCE_DIR/$UPGRADE_SOURCE_ASSET"
+          curl --fail --location --retry 3 --output "$archive" \
+            "https://github.com/${GITHUB_REPOSITORY}/releases/download/${UPGRADE_SOURCE_VERSION}/${UPGRADE_SOURCE_ASSET}"
+          echo "$UPGRADE_SOURCE_SHA256  $archive" | sha256sum --check --strict
+          unzip -q "$archive" -d "$SOURCE_DIR"
+          chmod +x "$SOURCE_DIR/rustfs"
+          test -x "$SOURCE_DIR/rustfs"
+          echo "RUSTFS_UPGRADE_SOURCE_BINARY=$SOURCE_DIR/rustfs" >> "$GITHUB_ENV"
+
+      - name: Build rustfs binary
+        run: |
+          cargo build -p rustfs --bins
+          : > target/debug/rustfs.features
+
+      - name: Verify distributed e2e membership
+        env:
+          NEXTEST_LISTING: ${{ runner.temp }}/rustfs-e2e-distributed-list.json
+        run: |
+          cargo nextest list --profile e2e-distributed -p e2e_test --message-format json > "${NEXTEST_LISTING}"
+          python3 ./scripts/check_test_wiring.py --check-profile e2e-distributed "${NEXTEST_LISTING}"
+
+      - name: Run distributed 4-node e2e suite
+        env:
+          RUSTFS_E2E_LOG_DIR: ${{ runner.temp }}/rustfs-e2e-distributed-logs
+          FILTER: ${{ inputs.filter }}
+        run: |
+          set -euo pipefail
+          if [ -n "${FILTER}" ]; then
+            cargo nextest run --profile e2e-distributed -p e2e_test -E "${FILTER}"
+          else
+            cargo nextest run --profile e2e-distributed -p e2e_test --no-tests=fail
+          fi
+
+      - name: Upload distributed e2e diagnostics
+        if: always()
+        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
+        with:
+          name: e2e-distributed-${{ github.run_number }}
+          path: |
+            target/nextest/e2e-distributed/junit.xml
+            ${{ runner.temp }}/rustfs-e2e-distributed-list.json
+            ${{ runner.temp }}/rustfs-e2e-distributed-logs/
+          retention-days: 7
+          if-no-files-found: warn
+
+      - name: Unmount isolated pool filesystems
+        if: always()
+        run: |
+          set -euo pipefail
+          mount_base="${RUNNER_TEMP}/rustfs-e2e-pools"
+          for pool in 0 1 2 3; do
+            mountpoint="${mount_base}/pool-${pool}"
+            if mountpoint --quiet "${mountpoint}"; then
+              sudo umount "${mountpoint}"
+            fi
+          done
+
+  alert-on-failure:
+    name: Alert on scheduled failure
+    needs: [distributed]
+    if: always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure')
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+      issues: write
+    steps:
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+      - name: Open or update failure-tracking issue
+        uses: ./.github/actions/schedule-failure-issue
+        with:
+          github-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/scheduled-validation-watchdog.yml

@@ -22,6 +22,7 @@ on:
       - "Continuous Integration"
       - "coverage"
       - "e2e-nightly"
+      - "e2e-distributed"
       - "e2e-s3tests"
       - "Fuzz"
       - "mint"

Cargo.lock

@@ -2527,18 +2527,18 @@ checksum = "790eea4361631c5e7d22598ecd5723ff611904e3344ce8720784c93e3d83d40b"
 
 [[package]]
 name = "crossbeam-channel"
-version = "0.5.16"
+version = "0.5.17"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d85363c37faeca707aef026efa9f3b34d077bce547e48f770770625c6013679e"
+checksum = "98b0cc327b5bc766e7fda9c9260cc0fa81b43a8e240440422dff70788e3f9ef1"
 dependencies = [
  "crossbeam-utils",
 ]
 
 [[package]]
 name = "crossbeam-deque"
-version = "0.8.7"
+version = "0.8.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5181e0de7b61eb03a81e347d6dd8797bae9da5146707b51077e2d71a54ec0ceb"
+checksum = "622f3fc73690be383c7214310406f28a90e6edeadc3cea882f9d71e495b9711a"
 dependencies = [
  "crossbeam-epoch",
  "crossbeam-utils",
@@ -2546,27 +2546,27 @@ dependencies = [
 
 [[package]]
 name = "crossbeam-epoch"
-version = "0.9.20"
+version = "0.9.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2d6914041f254d6e9176c01941b21115dcfb7089e55135a35411081bd106ef3f"
+checksum = "dc74980687109a3b14c72fd458107bf0baa1da1a1a805e178d15501ba9b86d9d"
 dependencies = [
  "crossbeam-utils",
 ]
 
 [[package]]
 name = "crossbeam-queue"
-version = "0.3.13"
+version = "0.3.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "803d13fb3b09d88be9f4dbc29062c66b19bf7170867ceb746d2a8689bf6c7a26"
+checksum = "03e8bd762f7479489c70ed6c768ddca99d7296857de437a68dcb2a94365b3fae"
 dependencies = [
  "crossbeam-utils",
 ]
 
 [[package]]
 name = "crossbeam-utils"
-version = "0.8.22"
+version = "0.8.23"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "61803da095bee82a81bb1a452ecc25d3b2f1416d1897eb86430c6159ef717c17"
+checksum = "a31eee39dddec8330830986fcd7625edb5a24ec90ea038215273bbc3adb08ac6"
 
 [[package]]
 name = "crunchy"
@@ -3673,9 +3673,9 @@ dependencies = [
 
 [[package]]
 name = "der"
-version = "0.8.1"
+version = "0.8.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a69dedd701da44b0536442edf09c81a64b0ab97a7a4a5e3d1971f00027cbc63d"
+checksum = "a878c850e9e421b20262e9b41f9c860e4785fa07541c266b62ff9d1ef998a80a"
 dependencies = [
  "const-oid 0.10.2",
  "pem-rfc7468 1.0.0",
@@ -3946,7 +3946,7 @@ dependencies = [
  "libc",
  "option-ext",
  "redox_users 0.5.2",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -4090,7 +4090,7 @@ version = "0.17.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c0681a4fc24c767085329728d8dfba959af91228aa4610cca4f8ce317ba46ae0"
 dependencies = [
- "der 0.8.1",
+ "der 0.8.2",
  "digest 0.11.3",
  "elliptic-curve 0.14.1",
  "rfc6979 0.6.0",
@@ -4295,7 +4295,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
 dependencies = [
  "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -5693,7 +5693,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "20fd6de4ccfcc187e38bc21cfa543cb5a302cb86a8b114eb7f0bf0dc9f8ac00f"
 dependencies = [
  "io-lifetimes 3.0.1",
- "windows-sys 0.60.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -5734,9 +5734,9 @@ dependencies = [
 
 [[package]]
 name = "ipnet"
-version = "2.12.1"
+version = "2.12.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6a756c3fac73139e83f14c2d742155dd2b78d3ee56597b419a0579b7bdd6dd78"
+checksum = "791930b43c0d5973160d90a8f3894509f2b273430f5c5c73b668636d0287c5c0"
 dependencies = [
  "serde",
 ]
@@ -5758,7 +5758,7 @@ checksum = "3640c1c38b8e4e43584d8df18be5fc6b0aa314ce6ebf51b53313d4306cca8e46"
 dependencies = [
  "hermit-abi",
  "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -6139,9 +6139,9 @@ checksum = "3eaf3ede3fee6db1a4c2ee091bf8a8b4dccdc6d17f656fb07896ee72867612f2"
 
 [[package]]
 name = "libflate"
-version = "2.3.1"
+version = "2.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a4da9b700e758e57152a1fd1c52cbdc5727c1aa6d8743dc1acda917398f1d76c"
+checksum = "561a8da1a50e1428d3c51321dafeca849df992a5bb67720c386131234caba82e"
 dependencies = [
  "adler32",
  "crc32fast",
@@ -6955,7 +6955,7 @@ version = "0.50.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7957b9740744892f114936ab4a57b3f487491bbeafaf8083688b16841a4240e5"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -7933,7 +7933,7 @@ version = "0.8.0-rc.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "986d2e952779af96ea048f160fd9194e1751b4faea78bcf3ceb456efe008088e"
 dependencies = [
- "der 0.8.1",
+ "der 0.8.2",
  "spki 0.8.0",
 ]
 
@@ -7976,7 +7976,7 @@ dependencies = [
  "aes 0.9.3",
  "aes-gcm",
  "cbc 0.2.1",
- "der 0.8.1",
+ "der 0.8.2",
  "pbkdf2 0.13.0",
  "rand_core 0.10.1",
  "scrypt 0.12.0",
@@ -8000,7 +8000,7 @@ version = "0.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "451913da69c775a56034ea8d9003d27ee8948e12443eae7c038ba100a4f21cb7"
 dependencies = [
- "der 0.8.1",
+ "der 0.8.2",
  "pkcs5 0.8.1",
  "rand_core 0.10.1",
  "spki 0.8.0",
@@ -8706,7 +8706,7 @@ dependencies = [
  "once_cell",
  "socket2",
  "tracing",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -8942,9 +8942,9 @@ dependencies = [
 
 [[package]]
 name = "redis"
-version = "1.6.0"
+version = "1.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e37a4ca5c6ca42aa3e6df2fd32b987a65d32a4c2159a6f3fe0fd1df306a2658f"
+checksum = "2acbc41a996f7652b2ddd9dfd98cc4ff602cfd742ae35382f07f608405ab50ed"
 dependencies = [
  "arc-swap",
  "arcstr",
@@ -9326,7 +9326,7 @@ dependencies = [
  "curve25519-dalek 5.0.0",
  "data-encoding",
  "delegate",
- "der 0.8.1",
+ "der 0.8.2",
  "digest 0.11.3",
  "ecdsa 0.17.0",
  "ed25519-dalek 3.0.0",
@@ -10745,6 +10745,7 @@ dependencies = [
  "rustfs-data-usage",
  "rustfs-ecstore",
  "rustfs-filemeta",
+ "rustfs-heal",
  "rustfs-heal-contracts",
  "rustfs-lifecycle",
  "rustfs-lock",
@@ -10942,9 +10943,9 @@ dependencies = [
 
 [[package]]
 name = "rustfs-uring"
-version = "0.2.1"
+version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0486e62d0efe25db95c00aeacb2da84368adcba299216cda99fcb11328061c84"
+checksum = "b29bc57b4bd62a73f4fae408b536adf578332e50e464797d09dc2382c7cb68c2"
 dependencies = [
  "io-uring",
  "libc",
@@ -11065,7 +11066,7 @@ dependencies = [
  "errno",
  "libc",
  "linux-raw-sys",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -11148,7 +11149,7 @@ dependencies = [
  "security-framework",
  "security-framework-sys",
  "webpki-root-certs",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -11419,7 +11420,7 @@ checksum = "d56d437c2f19203ce5f7122e507831de96f3d2d4d3be5af44a0b0a09d8a80e4d"
 dependencies = [
  "base16ct 1.0.0",
  "ctutils",
- "der 0.8.1",
+ "der 0.8.2",
  "hybrid-array",
  "subtle",
  "zeroize",
@@ -11993,7 +11994,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "1d9efca8738c78ee9484207732f728b1ef517bbb1833d6fc0879ca898a522f6f"
 dependencies = [
  "base64ct",
- "der 0.8.1",
+ "der 0.8.2",
 ]
 
 [[package]]
@@ -12408,7 +12409,7 @@ dependencies = [
  "getrandom 0.4.3",
  "once_cell",
  "rustix",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -13527,7 +13528,7 @@ version = "0.1.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]

Cargo.toml

@@ -256,10 +256,10 @@ clap = { version = "4.6.6" }
 const-str = { version = "1.1.0" }
 convert_case = "0.12.0"
 criterion = { version = "0.8" }
-crossbeam-queue = "0.3.13"
-crossbeam-channel = "0.5.16"
-crossbeam-deque = "0.8.7"
-crossbeam-utils = "0.8.22"
+crossbeam-queue = "0.3.14"
+crossbeam-channel = "0.5.17"
+crossbeam-deque = "0.8.8"
+crossbeam-utils = "0.8.23"
 datafusion = { default-features = false, version = "55.0.0" }
 derive_builder = "0.20.2"
 enumset = "1.1.14"
@@ -306,7 +306,7 @@ rustfs-erasure-codec = { version = "8.0.2" }
 reed-solomon-simd = "3.1.0"
 regex = { version = "1.13.1" }
 rumqttc = { package = "rumqttc-next", version = "0.34.0" }
-redis = { version = "1.6.0" }
+redis = { version = "1.7.0" }
 rustify = { version = "0.7", default-features = false }
 rustix = { version = "1.1.4" }
 rust-embed = { version = "8.12.0" }

crates/common/src/mrf_channel.rs

@@ -422,9 +422,9 @@ fn unix_now_ms() -> u64 {
         .unwrap_or(0)
 }
 
-/// A repair the MRF consumer landed, fanned out so retry ledgers can drop
-/// entries the journal no longer tracks (backlog#1894 axis B). The payload
-/// mirrors the intent identity so consumers match without re-parsing.
+/// Legacy, unverified repair notice. Its identity lacks kind, set scope,
+/// bucket incarnation and responsibility generation. Consumers must not use
+/// it to discharge persisted repair responsibility.
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct MrfRepairedEvent {
     pub bucket: Arc<str>,
@@ -439,8 +439,8 @@ const MRF_REPAIRED_EVENT_CAP: usize = 4096;
 
 static MRF_REPAIRED_EVENTS: OnceLock<std::sync::Mutex<std::collections::VecDeque<MrfRepairedEvent>>> = OnceLock::new();
 
-/// Record that the MRF consumer landed a repair. Never blocks: the critical
-/// section is a deque push under a std mutex.
+/// Record a legacy notification for compatibility. This is not an
+/// acknowledgement of storage verification or durable repair completion.
 pub fn note_mrf_repaired(bucket: &str, object: &str, version_id: Option<[u8; 16]>) {
     let registry = MRF_REPAIRED_EVENTS.get_or_init(|| std::sync::Mutex::new(std::collections::VecDeque::new()));
     let Ok(mut events) = registry.lock() else {
@@ -515,6 +515,9 @@ mod tests {
         }
         coalescer_release(&key, Some(lease));
         let retry_lease = coalescer_admit(key.clone()).expect("released identity must admit a retry");
+        assert_ne!(lease, retry_lease);
+        coalescer_release(&key, Some(lease));
+        assert_eq!(coalescer_admit(key.clone()), Err(MrfIngressResult::Coalesced));
         coalescer_release(&key, Some(retry_lease));
     }
 

crates/config/README.md

@@ -130,6 +130,37 @@ Scanner cycle budget controls:
   - timeout returns S3 `SlowDown`, so clients should use normal SDK retry handling.
   - this is not a fdatasync or group-commit switch. Track fdatasync batching separately with `rustfs_s3_put_object_rename_fdatasync_batch_files`.
 
+## Foreground write admission environment variables
+
+Large direct `PutObject` requests and multipart `UploadPart` requests share one
+per-process permit pool that bounds how many bodies are ingested and written
+concurrently. Small direct PUTs stay on the legacy path.
+
+- `RUSTFS_PUT_LARGE_FOREGROUND_ADMISSION_ENABLE`
+  - enables the default-on pool; `false` keeps only the soft request counter.
+  - default is `true`.
+- `RUSTFS_PUT_LARGE_FOREGROUND_ADMISSION_LIMIT`
+  - permits in the pool; `0` derives half of `RUSTFS_OBJECT_MAX_CONCURRENT_DISK_READS`, clamped to `32`.
+  - default is `0` (32 permits at stock settings).
+- `RUSTFS_PUT_LARGE_FOREGROUND_ADMISSION_MIN_SIZE_BYTES`
+  - smallest direct `PutObject` that takes a permit; unknown-size requests always do.
+  - default is `33554432` (32 MiB).
+- `RUSTFS_PUT_LARGE_FOREGROUND_ADMISSION_WAIT_TIMEOUT_MS`
+  - how long a direct `PutObject` waits for a permit before returning S3 `SlowDown`.
+  - default is `250`.
+- `RUSTFS_PUT_MULTIPART_FOREGROUND_ADMISSION_MIN_SIZE_BYTES`
+  - smallest `UploadPart` that takes a permit; `0` gates every part.
+  - default is `0`.
+- `RUSTFS_PUT_MULTIPART_FOREGROUND_ADMISSION_WAIT_TIMEOUT_MS`
+  - how long an `UploadPart` waits in the bounded queue for a permit before returning S3 `SlowDown`; `0` rejects immediately when the pool is full.
+  - default is `30000`. Parts wait before body ingest, so SDK-default clients that send every part of an upload concurrently drain through the pool instead of failing.
+- `RUSTFS_PUT_MULTIPART_FOREGROUND_ADMISSION_MAX_PENDING`
+  - maximum `UploadPart` requests waiting for a permit at once; parts beyond it return `SlowDown` without waiting.
+  - default is `0`, which derives 16 times the permit limit (512 at stock settings).
+- `RUSTFS_PUT_FOREGROUND_ADMISSION_ENABLE`, `RUSTFS_PUT_FOREGROUND_ADMISSION_LIMIT`, `RUSTFS_PUT_FOREGROUND_ADMISSION_WAIT_TIMEOUT_MS`
+  - experimental strict gate that applies to every foreground write regardless of size and replaces the pool above when enabled.
+  - default is disabled; enabling it with limit `0` disables foreground write admission entirely.
+
 ## Remote tier timeout environment variables
 
 - `RUSTFS_TIER_REMOTE_CONNECT_TIMEOUT_SECS`
@@ -172,6 +203,14 @@ Drive timeout profile preset:
   - Then `RUSTFS_DRIVE_MAX_TIMEOUT_DURATION` legacy fallback.
   - Then the profile-derived default (`default` or `high_latency`).
 
+## Admin peer probe timeout
+
+- `RUSTFS_ADMIN_PEER_PROBE_TIMEOUT_SECS`
+  - total per-peer budget for the `server_info`/`storage_info` admin probe round; `server_info` may reconnect once and `storage_info` remains a single attempt.
+  - default is `10` seconds, preserving the previous two-attempt worst-case budget.
+  - values must be positive; `0` or an invalid value falls back to the default, and values above `60` are clamped to `60`.
+  - the setting is read by the aggregating node only; it does not change the internode RPC wire contract. Any retry shares one round deadline rather than receiving a fresh timeout.
+
 ## Startup filesystem boundary policy
 
 - `RUSTFS_UNSUPPORTED_FS_POLICY` controls startup behavior when RustFS detects local endpoint filesystems that are outside the supported production boundary.

crates/config/src/constants/internode.rs

@@ -39,6 +39,15 @@ pub const DEFAULT_INTERNODE_HTTP2_KEEPALIVE_TIMEOUT_SECS: u64 = 20;
 pub const ENV_INTERNODE_RPC_TIMEOUT_SECS: &str = "RUSTFS_INTERNODE_RPC_TIMEOUT_SECS";
 pub const DEFAULT_INTERNODE_RPC_TIMEOUT_SECS: u64 = 30;
 
+/// Total budget for one admin peer probe round, including any reconnect retry.
+///
+/// This is intentionally separate from the transport-level RPC timeout: admin
+/// probes may retry once, but the retry must consume the same round budget.
+pub const ENV_ADMIN_PEER_PROBE_TIMEOUT_SECS: &str = "RUSTFS_ADMIN_PEER_PROBE_TIMEOUT_SECS";
+pub const DEFAULT_ADMIN_PEER_PROBE_TIMEOUT_SECS: u64 = 10;
+pub const MAX_ADMIN_PEER_PROBE_TIMEOUT_SECS: u64 = 60;
+const _: () = assert!(DEFAULT_ADMIN_PEER_PROBE_TIMEOUT_SECS <= MAX_ADMIN_PEER_PROBE_TIMEOUT_SECS);
+
 // ── Client-side internode gRPC channel tuning (P0) ──
 // These mirror the server-side HTTP/2 transport tuning in `rustfs/src/server/http.rs`
 // on the *client* `tonic` `Endpoint` used for internode control-plane RPCs. Prior to
@@ -312,6 +321,7 @@ mod tests {
         assert_eq!(DEFAULT_INTERNODE_HTTP2_KEEPALIVE_INTERVAL_SECS, 5);
         assert_eq!(DEFAULT_INTERNODE_HTTP2_KEEPALIVE_TIMEOUT_SECS, 20);
         assert_eq!(DEFAULT_INTERNODE_RPC_TIMEOUT_SECS, 30);
+        assert_eq!(DEFAULT_ADMIN_PEER_PROBE_TIMEOUT_SECS, 10);
         assert_eq!(DEFAULT_INTERNODE_HTTP_TUNING_PROFILE, "legacy");
     }
 
@@ -412,6 +422,7 @@ mod tests {
             "RUSTFS_INTERNODE_HTTP2_KEEPALIVE_TIMEOUT_SECS"
         );
         assert_eq!(ENV_INTERNODE_RPC_TIMEOUT_SECS, "RUSTFS_INTERNODE_RPC_TIMEOUT_SECS");
+        assert_eq!(ENV_ADMIN_PEER_PROBE_TIMEOUT_SECS, "RUSTFS_ADMIN_PEER_PROBE_TIMEOUT_SECS");
         assert_eq!(ENV_INTERNODE_HTTP_TUNING_PROFILE, "RUSTFS_INTERNODE_HTTP_TUNING_PROFILE");
         assert_eq!(ENV_INTERNODE_HTTP_POOL_MAX_IDLE_PER_HOST, "RUSTFS_INTERNODE_HTTP_POOL_MAX_IDLE_PER_HOST");
         assert_eq!(ENV_INTERNODE_HTTP_POOL_IDLE_TIMEOUT_SECS, "RUSTFS_INTERNODE_HTTP_POOL_IDLE_TIMEOUT_SECS");

crates/config/src/constants/object.rs

@@ -365,13 +365,36 @@ pub const ENV_PUT_MULTIPART_FOREGROUND_ADMISSION_MIN_SIZE_BYTES: &str =
     "RUSTFS_PUT_MULTIPART_FOREGROUND_ADMISSION_MIN_SIZE_BYTES";
 pub const DEFAULT_PUT_MULTIPART_FOREGROUND_ADMISSION_MIN_SIZE_BYTES: usize = 0;
 
-/// Time in milliseconds an automatic foreground write waits for a permit.
+/// Time in milliseconds an automatic foreground direct PutObject waits for a permit.
 ///
 /// A short wait smooths transient bursts while still returning S3
 /// `SlowDown`/503 before body ingest when the node is already saturated.
 pub const ENV_PUT_LARGE_FOREGROUND_ADMISSION_WAIT_TIMEOUT_MS: &str = "RUSTFS_PUT_LARGE_FOREGROUND_ADMISSION_WAIT_TIMEOUT_MS";
 pub const DEFAULT_PUT_LARGE_FOREGROUND_ADMISSION_WAIT_TIMEOUT_MS: u64 = 250;
 
+/// Time in milliseconds a multipart UploadPart waits for a foreground write permit.
+///
+/// SDK-default multipart clients send every part of an upload concurrently, so
+/// a single node routinely sees several times more parts in flight than the
+/// permit pool allows. Those parts have not ingested a body yet, so queueing
+/// them costs a connection rather than memory or internode streams; the pool
+/// still bounds the number of parts being written. The wait is long enough for
+/// an ordinary queue to drain on modest hardware, and a part that cannot get a
+/// permit within it fails with S3 `SlowDown`/503 for the client to retry.
+/// `0` rejects immediately when the pool is full.
+pub const ENV_PUT_MULTIPART_FOREGROUND_ADMISSION_WAIT_TIMEOUT_MS: &str =
+    "RUSTFS_PUT_MULTIPART_FOREGROUND_ADMISSION_WAIT_TIMEOUT_MS";
+pub const DEFAULT_PUT_MULTIPART_FOREGROUND_ADMISSION_WAIT_TIMEOUT_MS: u64 = 30_000;
+
+/// Maximum multipart UploadPart requests waiting for a foreground write permit per process.
+///
+/// Parts beyond this queue depth are rejected with S3 `SlowDown`/503 without
+/// waiting, so a genuinely saturated node still fails fast instead of holding
+/// an unbounded set of connections open for the whole wait timeout.
+/// `0` derives the depth from the permit limit.
+pub const ENV_PUT_MULTIPART_FOREGROUND_ADMISSION_MAX_PENDING: &str = "RUSTFS_PUT_MULTIPART_FOREGROUND_ADMISSION_MAX_PENDING";
+pub const DEFAULT_PUT_MULTIPART_FOREGROUND_ADMISSION_MAX_PENDING: usize = 0;
+
 const _: () = assert!(DEFAULT_PUT_LARGE_FOREGROUND_ADMISSION_ENABLE);
 
 /// Environment variable for minimum GetObject timeout in seconds.

crates/e2e_test/README.md

@@ -26,6 +26,7 @@ Registered in [`src/lib.rs`](src/lib.rs). Grouped by concern:
 | **protocols** | [`src/protocols/`](src/protocols) | FTPS, WebDAV, SFTP compliance. Fixed ports, own guide: [`src/protocols/README.md`](src/protocols/README.md) |
 | **reliant** | [`src/reliant/`](src/reliant) | Tests that reuse an **externally started** server (SQL/select, conditional writes, lifecycle, deleted-object reads, node-interact). Run via [`scripts/run_e2e_tests.sh`](../../scripts/run_e2e_tests.sh); see [`src/reliant/README.md`](src/reliant/README.md) |
 | **cluster** | `cluster_concurrency_test`, `stale_multipart_cleanup_cluster_test`, `namespace_lock_quorum_test`, `admin_timeout_regression_test`, `object_lambda_test`, `replication_extension_test`, `tier_stats_cluster_test` | Multi-node scenarios via `RustFSTestClusterEnvironment` |
+| **distributed 4×4** | [`src/distributed/`](src/distributed) | Storage-sensitive PR and nightly `e2e-distributed` lane: S3, object lock/WORM, versioning, bucket/site replication, quota, expand/decommission/rebalance, concurrency, chaos, 4-node upgrade of historical data and IAM AK/SK. Map: [`docs/testing/distributed-e2e.md`](../../docs/testing/distributed-e2e.md) |
 | **chaos / reliability** | [`src/chaos.rs`](src/chaos.rs), `reliability_disk_fault_test`, `heal_erasure_disk_rebuild_test`, `server_startup_failfast_test` | Disk offline/replace/corrupt, EC rebuild, heal, fail-fast startup |
 | **upgrade compatibility** | `upgrade_compatibility_test` | Pinned previous-release writes followed by current-build reads on the same data directory |
 
@@ -171,6 +172,7 @@ the same profile for membership and execution with one nightly worker.
 | KMS suite | `e2e-full` job, merge queue + main | **Active** |
 | Direct and mixed-version rolling upgrades from pinned previous release | `e2e-upgrade.yml`, storage-sensitive PRs + release tags + weekly | **Active** |
 | Cluster faults (`e2e-nightly` profile) | consolidated nightly workflow | **Active** (backlog#1149 ci-7) |
+| Distributed 4-node 4-disk (`e2e-distributed` profile) | `.github/workflows/e2e-distributed.yml` | **Active** (storage-sensitive PR / nightly / dispatch) |
 | Protocols (FTPS/WebDAV/SFTP) | consolidated nightly workflow, serial | **Active** (backlog#1149 ci-7) |
 | Replication (fast subset) | `e2e-smoke` profile, `e2e-tests` job, every PR | **Active** (backlog#1147 repl-1) |
 | Replication (slow + multi-node) | `e2e-repl-nightly` profile, consolidated nightly workflow | **Active** (backlog#1147 repl-1) |
@@ -182,6 +184,8 @@ the wiring source of truth. Committed test-ID digests under
 
 ## Troubleshooting
 
+**Endpoint blackhole scenario skipped** — `heal_erasure_disk_rebuild_test::tests::test_cluster_root_heal_recovers_after_target_endpoint_blackhole` installs a loopback `iptables` DROP rule and therefore needs `CAP_NET_ADMIN` (root or passwordless `sudo -n iptables`). A host where `iptables` is missing or cannot read the OUTPUT chain (typical inside an unprivileged container, where the nf_tables backend reports "Permission denied" even under `sudo`) logs a `heal_interruption_skipped` warning and returns without exercising heal. Set `RUSTFS_E2E_REQUIRE_NET_FAULT_INJECTION=1` on lanes that do provision the capability so a broken runner fails instead of skipping.
+
 **Reproduce a CI failure locally** — run the exact profile/lane:
 
 ```bash
@@ -191,6 +195,9 @@ cargo nextest run --profile e2e-smoke -p e2e_test
 cargo nextest run --profile e2e-full -p e2e_test
 # Cluster fault nightly lane
 cargo nextest run --profile e2e-nightly -p e2e_test
+# 4-node 4-disk distributed lane (S3 / lock / versioning / replication / decommission / chaos / upgrade)
+# Upgrade cases need RUSTFS_UPGRADE_SOURCE_BINARY; without it they fail closed.
+cargo nextest run --profile e2e-distributed -p e2e_test
 # Replication nightly lane; awscurl is required for STS paths
 cargo nextest run --profile e2e-repl-nightly -p e2e_test
 # Fixed-port protocol nightly lane

crates/e2e_test/build.rs

@@ -0,0 +1,74 @@
+// Copyright 2024 RustFS Team
+// Licensed under the Apache License, Version 2.0.
+
+use std::path::Path;
+use std::process::Command;
+
+fn git(root: &Path, args: &[&str]) -> Option<String> {
+    let output = Command::new("git").args(args).current_dir(root).output().ok()?;
+    output
+        .status
+        .success()
+        .then(|| String::from_utf8_lossy(&output.stdout).trim().to_owned())
+}
+
+fn emit(name: &str, value: &str) {
+    let value = if value.contains(['\n', '\r']) { "unknown" } else { value };
+    println!("cargo:rustc-env=RUSTFS_E2E_BUILD_{name}={value}");
+}
+
+fn main() {
+    let manifest = std::env::var_os("CARGO_MANIFEST_DIR").unwrap_or_default();
+    let root = Path::new(&manifest).join("../..");
+    // Cover dependency/common sources as well as this crate. HEAD/ref/index
+    // changes must refresh identity even when no Rust source mtime changes.
+    for path in [
+        "crates",
+        "rustfs",
+        "Cargo.toml",
+        "Cargo.lock",
+        "rust-toolchain.toml",
+        ".cargo",
+        ".config",
+    ] {
+        println!("cargo:rerun-if-changed={}", root.join(path).display());
+    }
+    let mut git_paths = vec!["HEAD".to_owned(), "index".to_owned(), "packed-refs".to_owned()];
+    if let Some(reference) = git(&root, &["symbolic-ref", "-q", "HEAD"]) {
+        git_paths.push(reference);
+    }
+    for path in git_paths {
+        if let Some(path) = git(&root, &["rev-parse", "--git-path", &path]) {
+            let path = Path::new(&path);
+            let path = if path.is_absolute() {
+                path.to_owned()
+            } else {
+                root.join(path)
+            };
+            if path.exists() {
+                println!("cargo:rerun-if-changed={}", path.display());
+            }
+        }
+    }
+    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 lock = git(&root, &["hash-object", "Cargo.lock"]).unwrap_or_else(|| "unknown".to_owned());
+    let mut features = std::env::vars()
+        .filter_map(|(key, _)| {
+            key.strip_prefix("CARGO_FEATURE_")
+                .map(|name| name.to_ascii_lowercase().replace('_', "-"))
+        })
+        .collect::<Vec<_>>();
+    features.sort();
+    emit("COMMIT", &revision);
+    emit("DIRTY", if dirty { "true" } else { "false" });
+    emit("LOCK", &lock);
+    emit("FEATURES", &features.join(","));
+    for name in ["TARGET", "PROFILE"] {
+        emit(name, &std::env::var(name).unwrap_or_else(|_| "unknown".to_owned()));
+    }
+    println!("cargo:rerun-if-env-changed=CARGO_ENCODED_RUSTFLAGS");
+    let flags = std::env::var("CARGO_ENCODED_RUSTFLAGS").unwrap_or_default();
+    let flags: String = flags.as_bytes().iter().map(|byte| format!("{byte:02x}")).collect();
+    emit("RUSTFLAGS_HEX", &flags);
+}

crates/e2e_test/src/chaos.rs

@@ -55,18 +55,20 @@ type ChaosResult<T> = Result<T, Box<dyn Error + Send + Sync>>;
 /// A successful S3 GET only proves that a quorum can serve an object. Replacement
 /// tests need this lower-level record to prove that the rebuilt target holds the
 /// `xl.meta` selected for a specific version and every `part.N` it declares.
-#[derive(Clone, Debug, Eq, PartialEq)]
+#[derive(Clone, Debug, Eq, PartialEq, serde::Serialize)]
 pub(crate) struct VersionShardCensus {
     pub version_id: Option<String>,
     pub has_xl_meta: bool,
     pub data_dir: Option<String>,
     pub erasure_index: Option<usize>,
+    pub data_blocks: Option<usize>,
+    pub parity_blocks: Option<usize>,
     pub expected_part_numbers: BTreeSet<usize>,
     pub present_part_fingerprints: BTreeMap<usize, PartShardFingerprint>,
     pub inline_data_fingerprint: Option<PartShardFingerprint>,
 }
 
-#[derive(Clone, Debug, Eq, PartialEq)]
+#[derive(Clone, Debug, Eq, PartialEq, serde::Serialize)]
 pub(crate) struct PartShardFingerprint {
     pub size: u64,
     pub sha256: String,
@@ -88,13 +90,15 @@ impl VersionShardCensus {
             && manifest.is_complete()
             && self.data_dir == manifest.data_dir
             && self.erasure_index == manifest.erasure_index
+            && self.data_blocks == manifest.data_blocks
+            && self.parity_blocks == manifest.parity_blocks
             && self.expected_part_numbers == manifest.expected_part_numbers
             && self.present_part_fingerprints == manifest.present_part_fingerprints
             && self.inline_data_fingerprint == manifest.inline_data_fingerprint
     }
 }
 
-fn sha256_hex(data: &[u8]) -> String {
+pub(crate) fn sha256_hex(data: &[u8]) -> String {
     let digest = Sha256::digest(data);
     digest.iter().map(|byte| format!("{byte:02x}")).collect()
 }
@@ -313,6 +317,8 @@ pub(crate) fn census_object_version_on_disk(
             has_xl_meta: false,
             data_dir: None,
             erasure_index: None,
+            data_blocks: None,
+            parity_blocks: None,
             expected_part_numbers: BTreeSet::new(),
             present_part_fingerprints: BTreeMap::new(),
             inline_data_fingerprint: None,
@@ -360,6 +366,8 @@ pub(crate) fn census_object_version_on_disk(
         has_xl_meta: true,
         data_dir,
         erasure_index,
+        data_blocks: Some(file_info.erasure.data_blocks),
+        parity_blocks: Some(file_info.erasure.parity_blocks),
         expected_part_numbers,
         present_part_fingerprints,
         inline_data_fingerprint,
@@ -413,6 +421,8 @@ mod tests {
             has_xl_meta: true,
             data_dir: Some("data-dir".to_string()),
             erasure_index: Some(3),
+            data_blocks: Some(2),
+            parity_blocks: Some(2),
             expected_part_numbers: BTreeSet::from([1]),
             present_part_fingerprints: BTreeMap::from([(1, shard_fingerprint(b"part").unwrap())]),
             inline_data_fingerprint: None,

crates/e2e_test/src/common.rs

@@ -1700,6 +1700,69 @@ impl RustFSTestClusterEnvironment {
         Ok(())
     }
 
+    /// Append a new single-node erasure pool to a stopped multi-pool cluster.
+    ///
+    /// Used to simulate pool expansion on localhost: every pool already owns
+    /// exactly one node with `drives_per_node >= 2` (the only multi-pool layout
+    /// the single-host `RUSTFS_VOLUMES` syntax can express). The new node is
+    /// allocated a fresh port and empty drive directories; callers must
+    /// [`Self::start`] afterwards so every process picks up the extended
+    /// volumes argument. Existing data directories are left untouched.
+    pub async fn append_single_node_pool(&mut self) -> Result<usize, Box<dyn std::error::Error + Send + Sync>> {
+        if self.nodes.iter().any(|node| node.process.is_some()) {
+            return Err("stop the cluster before appending a pool".into());
+        }
+        if self.topology.drives_per_node < 2 {
+            return Err(
+                "append_single_node_pool requires drives_per_node >= 2 (the server parser rejects a single-drive ellipses pool)"
+                    .into(),
+            );
+        }
+
+        let mut pools = self.topology.normalized_pools();
+        for (pool_idx, nodes) in pools.iter().enumerate() {
+            if nodes.len() != 1 {
+                return Err(format!(
+                    "pool {pool_idx} spans {} nodes; append_single_node_pool requires one node per pool",
+                    nodes.len()
+                )
+                .into());
+            }
+        }
+
+        let new_idx = self.nodes.len();
+        let port = RustFSTestEnvironment::find_available_port().await?;
+        let address = format!("127.0.0.1:{port}");
+        let data_dirs: Vec<String> = (0..self.topology.drives_per_node)
+            .map(|drive| format!("{}/node{}/drive{}", self.temp_dir, new_idx, drive))
+            .collect();
+        for dir in &data_dirs {
+            fs::create_dir_all(dir).await?;
+        }
+
+        self.nodes.push(ClusterNode {
+            url: format!("http://{address}"),
+            address,
+            data_dir: data_dirs[0].clone(),
+            data_dirs,
+            pool_idx: pools.len(),
+            process: None,
+        });
+        pools.push(vec![new_idx]);
+        self.topology.node_count = self.nodes.len();
+        self.topology.pools = pools;
+        self.node_extra_env.push(Vec::new());
+        self.node_capture_log_paths.push(None);
+        self.volume_proxy_addresses.push(None);
+
+        if !self.extra_env.iter().any(|(key, _)| key == "RUSTFS_UNSAFE_BYPASS_DISK_CHECK") {
+            self.extra_env
+                .push(("RUSTFS_UNSAFE_BYPASS_DISK_CHECK".to_string(), "true".to_string()));
+        }
+
+        Ok(new_idx)
+    }
+
     /// Gracefully stop one cluster node and wait for its process to exit.
     ///
     /// This is intentionally separate from [`Self::stop_node`]: the latter is

crates/e2e_test/src/data_usage_test.rs

@@ -35,11 +35,15 @@ where
 {
     let mut last_usage = DataUsageInfo::default();
     let mut last_query_error = None;
-    for _ in 0..45 {
+    for _ in 0..90 {
         match get_data_usage_info(env).await {
             Ok(usage) => {
                 last_query_error = None;
-                if usage.buckets_usage.contains_key(bucket) && predicate(&usage) {
+                if usage.is_complete_bucket_usage_snapshot()
+                    && usage.usage_snapshot_converged != Some(false)
+                    && usage.buckets_usage.contains_key(bucket)
+                    && predicate(&usage)
+                {
                     return Ok(usage);
                 }
                 last_usage = usage;

crates/e2e_test/src/distributed/chaos_test.rs

@@ -0,0 +1,222 @@
+// 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/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_object_bytes, payload_for, put_object, retrying_get_equals, unique_bucket,
+    wait_for_ready, wait_until,
+};
+use crate::chaos::{census_object_version_on_disk, signed_admin_post};
+use crate::common::{build_test_s3_config, init_logging};
+use crate::fault_proxy::FaultMode;
+use aws_sdk_s3::Client;
+use std::path::PathBuf;
+use std::sync::Arc;
+use std::time::Duration;
+use tokio::sync::{Barrier, mpsc};
+use tokio::time::timeout;
+
+#[tokio::test]
+async fn kill_and_restart_node_preserves_objects() -> TestResult {
+    init_logging();
+    let mut dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let bucket = unique_bucket("killnode");
+    dist.create_bucket(&bucket).await?;
+    let body = vec![0x11u8; 128 * 1024];
+    put_object(&dist.client(0)?, &bucket, "keep.bin", body.clone()).await?;
+
+    dist.cluster.stop_node(3)?;
+    retrying_get_equals(&dist.client(0)?, &bucket, "keep.bin", &body, Duration::from_secs(20)).await?;
+
+    dist.cluster.start_node(3).await?;
+    wait_for_ready(&dist.cluster).await?;
+    assert_object_bytes(&dist.client(3)?, &bucket, "keep.bin", &body).await?;
+    Ok(())
+}
+
+#[tokio::test]
+async fn full_cluster_restart_preserves_objects() -> TestResult {
+    init_logging();
+    let mut dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let bucket = unique_bucket("pwr");
+    dist.create_bucket(&bucket).await?;
+    let body = vec![0x44u8; 64 * 1024];
+    put_object(&dist.client(1)?, &bucket, "survive.bin", body.clone()).await?;
+
+    dist.cluster.stop();
+    dist.cluster.start().await?;
+    wait_for_ready(&dist.cluster).await?;
+    for node_idx in 0..dist.cluster.nodes.len() {
+        assert_object_bytes(&dist.client(node_idx)?, &bucket, "survive.bin", &body).await?;
+    }
+    Ok(())
+}
+
+#[tokio::test]
+async fn fresh_drive_replacement_is_physically_healed_without_data_change() -> TestResult {
+    init_logging();
+    let mut dist = DistCluster::start_with_env(DistLayout::FourByFour, &[("RUSTFS_HEAL_ENABLED", "true")]).await?;
+    let bucket = unique_bucket("baddrive");
+    dist.create_bucket(&bucket).await?;
+    let body = payload_for("fresh-drive/durable.bin", 8 * 1024 * 1024);
+    put_object(&dist.client(1)?, &bucket, "durable.bin", body.clone()).await?;
+
+    let replaced_drive = PathBuf::from(&dist.cluster.nodes[0].data_dirs[0]);
+    let baseline = census_object_version_on_disk(&replaced_drive, &bucket, "durable.bin", None)?;
+    assert!(
+        baseline.is_complete(),
+        "replacement target did not hold a complete baseline shard: {baseline:?}"
+    );
+    assert!(
+        !baseline.expected_part_numbers.is_empty(),
+        "replacement witness must use physical part shards: {baseline:?}"
+    );
+
+    dist.cluster.stop_node(0)?;
+    let format_path = replaced_drive.join(".rustfs.sys/format.json");
+    let format = std::fs::read(&format_path)?;
+    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 omitted parent")?)?;
+    std::fs::write(&format_path, format)?;
+    let empty = census_object_version_on_disk(&replaced_drive, &bucket, "durable.bin", None)?;
+    assert!(!empty.has_xl_meta, "fresh replacement unexpectedly retained object metadata: {empty:?}");
+
+    dist.cluster.start_node(0).await?;
+    wait_for_ready(&dist.cluster).await?;
+    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[1].url);
+    signed_admin_post(&heal_url, Some(heal_body), &dist.cluster.access_key, &dist.cluster.secret_key).await?;
+    wait_until(
+        Duration::from_secs(90),
+        || async {
+            let healed = census_object_version_on_disk(&replaced_drive, &bucket, "durable.bin", None)?;
+            Ok(healed.matches_manifest(&baseline))
+        },
+        "fresh replacement contains the original complete shard manifest",
+    )
+    .await?;
+
+    for node_idx in 0..dist.cluster.nodes.len() {
+        assert_object_bytes(&dist.client(node_idx)?, &bucket, "durable.bin", &body).await?;
+    }
+    Ok(())
+}
+
+#[tokio::test]
+async fn concurrent_gets_survive_peer_node_kill() -> TestResult {
+    init_logging();
+    let mut dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let bucket = unique_bucket("getkill");
+    dist.create_bucket(&bucket).await?;
+    let body = payload_for("inflight/steady.bin", 8 * 1024 * 1024);
+    put_object(&dist.client(0)?, &bucket, "steady.bin", body.clone()).await?;
+
+    let live: Vec<_> = (0..3).map(|idx| dist.client(idx)).collect::<Result<Vec<_>, _>>()?;
+    let worker_count = 12;
+    let release = Arc::new(Barrier::new(worker_count + 1));
+    let (started_tx, mut started_rx) = mpsc::unbounded_channel();
+    let mut handles = Vec::new();
+    for idx in 0..worker_count {
+        let client = live[idx % live.len()].clone();
+        let bucket = bucket.clone();
+        let body = body.clone();
+        let release = release.clone();
+        let started_tx = started_tx.clone();
+        handles.push(tokio::spawn(async move {
+            let response = client.get_object().bucket(&bucket).key("steady.bin").send().await?;
+            if response.content_length() != Some(body.len() as i64) {
+                return Err::<(), Box<dyn std::error::Error + Send + Sync>>(
+                    format!("worker {idx} received a wrong content length").into(),
+                );
+            }
+            started_tx.send(idx)?;
+            release.wait().await;
+            let actual = response.body.collect().await?.into_bytes();
+            if actual.as_ref() != body.as_slice() {
+                return Err(format!("worker {idx} received corrupted bytes after peer kill").into());
+            }
+            Ok(())
+        }));
+    }
+    drop(started_tx);
+    for _ in 0..worker_count {
+        timeout(Duration::from_secs(30), started_rx.recv())
+            .await?
+            .ok_or("a streaming GET exited before reaching the kill barrier")?;
+    }
+
+    dist.cluster.stop_node(3)?;
+    release.wait().await;
+    for handle in handles {
+        handle.await??;
+    }
+
+    dist.cluster.start_node(3).await?;
+    wait_for_ready(&dist.cluster).await?;
+    assert_object_bytes(&dist.client(3)?, &bucket, "steady.bin", &body).await?;
+    Ok(())
+}
+
+#[tokio::test]
+async fn blackholed_node_client_network_preserves_cluster_availability_and_recovers() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let proxy = crate::fault_proxy::FaultProxy::start(dist.cluster.nodes[3].address.parse()?).await?;
+    let proxied_url = format!("http://{}", proxy.local_addr());
+    let proxied_client = Client::from_conf(build_test_s3_config(
+        &proxied_url,
+        &dist.cluster.access_key,
+        &dist.cluster.secret_key,
+        None,
+        "distributed-network-chaos",
+    ));
+
+    let result: TestResult = async {
+        let bucket = unique_bucket("netfault");
+        dist.create_bucket(&bucket).await?;
+        let baseline = payload_for("network/baseline.bin", 1024 * 1024);
+        put_object(&dist.client(0)?, &bucket, "baseline.bin", baseline.clone()).await?;
+        assert_object_bytes(&proxied_client, &bucket, "baseline.bin", &baseline).await?;
+
+        proxy.set_mode(FaultMode::Blackhole);
+        assert_eq!(proxy.mode(), FaultMode::Blackhole);
+        if let Ok(Ok(_)) = timeout(
+            Duration::from_secs(5),
+            proxied_client.get_object().bucket(&bucket).key("baseline.bin").send(),
+        )
+        .await
+        {
+            return Err("blackholed node endpoint unexpectedly completed a GET".into());
+        }
+
+        let during = payload_for("network/during.bin", 1024 * 1024);
+        timeout(Duration::from_secs(30), async {
+            put_object(&dist.client(1)?, &bucket, "during-blackhole.bin", during.clone()).await?;
+            assert_object_bytes(&dist.client(2)?, &bucket, "baseline.bin", &baseline).await?;
+            assert_object_bytes(&dist.client(0)?, &bucket, "during-blackhole.bin", &during).await?;
+            Ok::<_, Box<dyn std::error::Error + Send + Sync>>(())
+        })
+        .await??;
+
+        proxy.set_mode(FaultMode::Pass);
+        retrying_get_equals(&proxied_client, &bucket, "during-blackhole.bin", &during, Duration::from_secs(30)).await?;
+        Ok(())
+    }
+    .await;
+
+    proxy.set_mode(FaultMode::Pass);
+    proxy.shutdown().await;
+    result
+}

crates/e2e_test/src/distributed/concurrency_stability_test.rs

@@ -0,0 +1,98 @@
+// 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/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_object_bytes, payload_for, put_object, unique_bucket};
+use crate::common::init_logging;
+use std::collections::BTreeSet;
+use std::sync::Arc;
+use tokio::sync::Barrier;
+
+#[tokio::test]
+async fn four_node_high_concurrency_mixed_workload_is_consistent_on_every_node() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let bucket = unique_bucket("conc");
+    dist.create_bucket(&bucket).await?;
+    let clients = Arc::new(dist.clients()?);
+    let worker_count = 24;
+    let rounds = 4;
+    let barrier = Arc::new(Barrier::new(worker_count));
+
+    let mut handles = Vec::new();
+    for idx in 0..worker_count {
+        let clients = clients.clone();
+        let barrier = barrier.clone();
+        let bucket = bucket.clone();
+        handles.push(tokio::spawn(async move {
+            barrier.wait().await;
+            let writer = &clients[idx % clients.len()];
+            let reader = &clients[(idx + 1) % clients.len()];
+            let copier = &clients[(idx + 2) % clients.len()];
+            let mut retained = Vec::with_capacity(rounds);
+            for round in 0..rounds {
+                let key = format!("source/worker-{idx:02}-round-{round}.bin");
+                let copy_key = format!("retained/worker-{idx:02}-round-{round}.bin");
+                let body = payload_for(&key, 64 * 1024);
+                put_object(writer, &bucket, &key, body.clone()).await?;
+
+                let head = reader.head_object().bucket(&bucket).key(&key).send().await?;
+                if head.content_length() != Some(body.len() as i64) {
+                    return Err(format!("HEAD returned the wrong size for {key}: {head:?}").into());
+                }
+                assert_object_bytes(reader, &bucket, &key, &body).await?;
+
+                copier
+                    .copy_object()
+                    .bucket(&bucket)
+                    .key(&copy_key)
+                    .copy_source(format!("{bucket}/{key}"))
+                    .send()
+                    .await?;
+                assert_object_bytes(writer, &bucket, &copy_key, &body).await?;
+
+                writer.delete_object().bucket(&bucket).key(&key).send().await?;
+                let missing = reader
+                    .head_object()
+                    .bucket(&bucket)
+                    .key(&key)
+                    .send()
+                    .await
+                    .expect_err("deleted source key must not remain visible");
+                if missing.raw_response().map(|response| response.status().as_u16()) != Some(404) {
+                    return Err(format!("deleted source {key} returned an unexpected result: {missing:?}").into());
+                }
+                retained.push((copy_key, body));
+            }
+            Ok::<_, Box<dyn std::error::Error + Send + Sync>>(retained)
+        }));
+    }
+
+    let mut inventory = Vec::new();
+    for handle in handles {
+        inventory.extend(handle.await??);
+    }
+
+    let expected_keys: BTreeSet<_> = inventory.iter().map(|(key, _)| key.as_str()).collect();
+    for (node_idx, client) in clients.iter().enumerate() {
+        let listed = client.list_objects_v2().bucket(&bucket).prefix("retained/").send().await?;
+        let listed_keys: BTreeSet<_> = listed.contents().iter().filter_map(|object| object.key()).collect();
+        assert_eq!(listed_keys, expected_keys, "node {node_idx} returned a divergent retained-key listing");
+        for (key, body) in &inventory {
+            assert_object_bytes(client, &bucket, key, body)
+                .await
+                .map_err(|error| format!("node {node_idx} failed to read {key}: {error}"))?;
+        }
+    }
+    Ok(())
+}

crates/e2e_test/src/distributed/concurrent_data_movement_test.rs

@@ -0,0 +1,74 @@
+// 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/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::{
+    DECOMMISSION_POOL_ID, DistCluster, DistLayout, TestResult, assert_inventory, decommission_running_with_progress,
+    decommission_status_json, payload_for, put_inventory_retrying, retrying_get_equals, retrying_put, start_decommission,
+    unique_bucket, wait_for_decommission_complete, wait_for_decommission_running_with_progress,
+};
+use crate::common::init_logging;
+use std::sync::Arc;
+use std::time::Duration;
+use tokio::sync::Barrier;
+
+#[tokio::test]
+async fn concurrent_puts_during_decommission_do_not_lose_baseline_or_new_objects() -> TestResult {
+    init_logging();
+    let mut dist = DistCluster::start(DistLayout::SingleNodeFourDrive).await?;
+    let bucket = unique_bucket("concdecom");
+    dist.create_bucket(&bucket).await?;
+    let baseline_client = dist.client(0)?;
+    let inventory = put_inventory_retrying(&baseline_client, &bucket, 96, 256 * 1024, Duration::from_secs(30)).await?;
+    dist.expand_to_four_pools().await?;
+
+    start_decommission(&dist.cluster, DECOMMISSION_POOL_ID).await?;
+
+    let clients = Arc::new(dist.clients()?);
+    let barrier = Arc::new(Barrier::new(17));
+    let mut handles = Vec::new();
+    for idx in 0..16 {
+        let clients = clients.clone();
+        let barrier = barrier.clone();
+        let bucket = bucket.clone();
+        handles.push(tokio::spawn(async move {
+            barrier.wait().await;
+            let client = &clients[idx % clients.len()];
+            let key = format!("live/{idx:02}.bin");
+            let body = payload_for(&key, 8 * 1024);
+            retrying_put(client, &bucket, &key, body.clone(), Duration::from_secs(45)).await?;
+            Ok::<_, Box<dyn std::error::Error + Send + Sync>>((key, body))
+        }));
+    }
+
+    wait_for_decommission_running_with_progress(&dist.cluster, DECOMMISSION_POOL_ID, Duration::from_secs(30)).await?;
+    barrier.wait().await;
+
+    let mut live_objects = Vec::new();
+    for handle in handles {
+        live_objects.push(handle.await??);
+    }
+    let status = decommission_status_json(&dist.cluster).await?;
+    if !decommission_running_with_progress(&status, DECOMMISSION_POOL_ID)? {
+        return Err(format!("decommission did not remain active across concurrent PUTs: {status}").into());
+    }
+
+    wait_for_decommission_complete(&dist.cluster, DECOMMISSION_POOL_ID, Duration::from_secs(180)).await?;
+
+    let checker = dist.client(2)?;
+    assert_inventory(&checker, &bucket, &inventory).await?;
+    for (key, body) in live_objects {
+        retrying_get_equals(&checker, &bucket, &key, &body, Duration::from_secs(30)).await?;
+    }
+    Ok(())
+}

crates/e2e_test/src/distributed/data_integrity_movement_test.rs

@@ -0,0 +1,156 @@
+// 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/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::{
+    DECOMMISSION_POOL_ID, DistCluster, DistLayout, TestResult, assert_inventory, enable_versioning, put_inventory_retrying,
+    sha256_hex, start_decommission, unique_bucket, wait_for_decommission_active, wait_for_decommission_complete,
+};
+use crate::common::init_logging;
+use aws_sdk_s3::primitives::ByteStream;
+use aws_sdk_s3::types::{CompletedMultipartUpload, CompletedPart};
+use std::time::Duration;
+
+#[tokio::test]
+async fn decommission_does_not_alter_object_sha256_across_pools() -> TestResult {
+    init_logging();
+    let mut dist = DistCluster::start(DistLayout::SingleNodeFourDrive).await?;
+    let bucket = unique_bucket("integrity");
+    dist.create_bucket(&bucket).await?;
+    let client = dist.client(0)?;
+    enable_versioning(&client, &bucket).await?;
+    let inventory = put_inventory_retrying(&client, &bucket, 96, 256 * 1024, Duration::from_secs(30)).await?;
+    let before: Vec<(String, String)> = inventory.iter().map(|(key, body)| (key.clone(), sha256_hex(body))).collect();
+
+    let versioned_key = "history/versioned.bin";
+    let version_one = b"historical bytes before data movement".to_vec();
+    let version_two = b"current bytes before data movement".to_vec();
+    let version_one_id = client
+        .put_object()
+        .bucket(&bucket)
+        .key(versioned_key)
+        .body(ByteStream::from(version_one.clone()))
+        .send()
+        .await?
+        .version_id()
+        .ok_or("historical PUT omitted version ID")?
+        .to_string();
+    let version_two_id = client
+        .put_object()
+        .bucket(&bucket)
+        .key(versioned_key)
+        .body(ByteStream::from(version_two.clone()))
+        .send()
+        .await?
+        .version_id()
+        .ok_or("current PUT omitted version ID")?
+        .to_string();
+
+    let multipart_key = "multipart/moved.bin";
+    let first_part = vec![0x31; 5 * 1024 * 1024];
+    let second_part = vec![0x72; 1024 * 1024];
+    let upload = client
+        .create_multipart_upload()
+        .bucket(&bucket)
+        .key(multipart_key)
+        .send()
+        .await?;
+    let upload_id = upload.upload_id().ok_or("movement multipart upload omitted upload ID")?;
+    let uploaded_one = client
+        .upload_part()
+        .bucket(&bucket)
+        .key(multipart_key)
+        .upload_id(upload_id)
+        .part_number(1)
+        .body(ByteStream::from(first_part.clone()))
+        .send()
+        .await?;
+    let uploaded_two = client
+        .upload_part()
+        .bucket(&bucket)
+        .key(multipart_key)
+        .upload_id(upload_id)
+        .part_number(2)
+        .body(ByteStream::from(second_part.clone()))
+        .send()
+        .await?;
+    client
+        .complete_multipart_upload()
+        .bucket(&bucket)
+        .key(multipart_key)
+        .upload_id(upload_id)
+        .multipart_upload(
+            CompletedMultipartUpload::builder()
+                .parts(
+                    CompletedPart::builder()
+                        .part_number(1)
+                        .e_tag(uploaded_one.e_tag().ok_or("movement part 1 omitted ETag")?)
+                        .build(),
+                )
+                .parts(
+                    CompletedPart::builder()
+                        .part_number(2)
+                        .e_tag(uploaded_two.e_tag().ok_or("movement part 2 omitted ETag")?)
+                        .build(),
+                )
+                .build(),
+        )
+        .send()
+        .await?;
+
+    dist.expand_to_four_pools().await?;
+
+    start_decommission(&dist.cluster, DECOMMISSION_POOL_ID).await?;
+    wait_for_decommission_active(&dist.cluster, DECOMMISSION_POOL_ID, Duration::from_secs(30)).await?;
+    wait_for_decommission_complete(&dist.cluster, DECOMMISSION_POOL_ID, Duration::from_secs(180)).await?;
+
+    let after_client = dist.client(2)?;
+    assert_inventory(&after_client, &bucket, &inventory).await?;
+    for (key, expected_hash) in before {
+        let got = after_client.get_object().bucket(&bucket).key(&key).send().await?;
+        let body = got.body.collect().await?.into_bytes();
+        assert_eq!(sha256_hex(body.as_ref()), expected_hash, "checksum changed for {key} after decommission");
+    }
+    for (version_id, expected) in [(&version_one_id, &version_one), (&version_two_id, &version_two)] {
+        let got = after_client
+            .get_object()
+            .bucket(&bucket)
+            .key(versioned_key)
+            .version_id(version_id)
+            .send()
+            .await?
+            .body
+            .collect()
+            .await?
+            .into_bytes();
+        assert_eq!(got.as_ref(), expected.as_slice(), "version {version_id} changed after decommission");
+    }
+    let mut expected_multipart = first_part;
+    expected_multipart.extend_from_slice(&second_part);
+    let got_multipart = after_client
+        .get_object()
+        .bucket(&bucket)
+        .key(multipart_key)
+        .send()
+        .await?
+        .body
+        .collect()
+        .await?
+        .into_bytes();
+    assert_eq!(
+        sha256_hex(got_multipart.as_ref()),
+        sha256_hex(&expected_multipart),
+        "multipart checksum changed after decommission"
+    );
+    Ok(())
+}

crates/e2e_test/src/distributed/expand_decommission_rebalance_test.rs

@@ -0,0 +1,81 @@
+// 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/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::{
+    DECOMMISSION_POOL_ID, DistCluster, DistLayout, TestResult, assert_inventory, list_pools_json, put_inventory,
+    put_inventory_retrying, start_decommission, start_rebalance, unique_bucket, wait_for_decommission_active,
+    wait_for_decommission_complete, wait_for_rebalance_active, wait_for_rebalance_complete,
+};
+use crate::common::init_logging;
+use std::time::Duration;
+
+#[tokio::test]
+async fn four_node_pool_expand_preserves_objects_then_rebalance() -> TestResult {
+    init_logging();
+    let mut dist = DistCluster::start(DistLayout::SingleNodeFourDrive).await?;
+    let bucket = unique_bucket("expand");
+    dist.create_bucket(&bucket).await?;
+    let client = dist.client(0)?;
+    let inventory = put_inventory(&client, &bucket, 64, 256 * 1024).await?;
+    assert_inventory(&client, &bucket, &inventory).await?;
+
+    for expected_nodes in 2..=4 {
+        let new_node = dist.append_pool_and_restart().await?;
+        assert_eq!(new_node + 1, expected_nodes);
+        assert_inventory(&dist.client(new_node)?, &bucket, &inventory).await?;
+    }
+    assert_eq!(dist.cluster.nodes.len(), 4);
+
+    // Prove that the expanded pool map is durable, and clear any recovery
+    // latch raised while the newly-added pool replicas converged.
+    dist.restart_current_binary_gracefully().await?;
+
+    let after_expand = dist.client(0)?;
+    assert_inventory(&after_expand, &bucket, &inventory).await?;
+    let peer = dist.client(3)?;
+    assert_inventory(&peer, &bucket, &inventory).await?;
+
+    let rebalance_id = start_rebalance(&dist.cluster).await?;
+    wait_for_rebalance_active(&dist.cluster, &rebalance_id, Duration::from_secs(30)).await?;
+    wait_for_rebalance_complete(&dist.cluster, &rebalance_id, Duration::from_secs(180)).await?;
+    assert_inventory(&peer, &bucket, &inventory).await?;
+    Ok(())
+}
+
+#[tokio::test]
+async fn four_pool_decommission_moves_objects_without_loss() -> TestResult {
+    init_logging();
+    let mut dist = DistCluster::start(DistLayout::SingleNodeFourDrive).await?;
+    let bucket = unique_bucket("decom");
+    dist.create_bucket(&bucket).await?;
+    let client = dist.client(0)?;
+    let inventory = put_inventory_retrying(&client, &bucket, 96, 128 * 1024, Duration::from_secs(30)).await?;
+    dist.expand_to_four_pools().await?;
+
+    let pools_before = list_pools_json(&dist.cluster).await?;
+    let pool_count = pools_before
+        .as_array()
+        .map(Vec::len)
+        .or_else(|| pools_before.get("pools").and_then(serde_json::Value::as_array).map(Vec::len))
+        .ok_or_else(|| format!("pool list omitted an array: {pools_before}"))?;
+    assert_eq!(pool_count, 4, "expected exactly four pools before decommission: {pools_before}");
+
+    start_decommission(&dist.cluster, DECOMMISSION_POOL_ID).await?;
+    wait_for_decommission_active(&dist.cluster, DECOMMISSION_POOL_ID, Duration::from_secs(30)).await?;
+    wait_for_decommission_complete(&dist.cluster, DECOMMISSION_POOL_ID, Duration::from_secs(180)).await?;
+
+    let after = dist.client(2)?;
+    assert_inventory(&after, &bucket, &inventory).await?;
+    Ok(())
+}

crates/e2e_test/src/distributed/extra_test.rs

@@ -0,0 +1,149 @@
+// 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/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_object_bytes, get_object_bytes, put_object, unique_bucket, wait_until,
+};
+use crate::common::init_logging;
+use aws_sdk_s3::primitives::ByteStream;
+use aws_sdk_s3::types::{CompletedMultipartUpload, CompletedPart};
+use std::time::Duration;
+
+#[tokio::test]
+async fn four_node_four_drive_multipart_and_cross_node_listing_agree() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let bucket = unique_bucket("extra");
+    dist.create_bucket(&bucket).await?;
+    let client = dist.client(0)?;
+
+    let key = "multipart.bin";
+    let part1 = vec![0x41u8; 5 * 1024 * 1024];
+    let part2 = vec![0x42u8; 5 * 1024 * 1024];
+    let upload = client.create_multipart_upload().bucket(&bucket).key(key).send().await?;
+    let upload_id = upload.upload_id().ok_or("missing upload id")?.to_string();
+
+    let uploaded1 = client
+        .upload_part()
+        .bucket(&bucket)
+        .key(key)
+        .upload_id(&upload_id)
+        .part_number(1)
+        .body(ByteStream::from(part1.clone()))
+        .send()
+        .await?;
+    let uploaded2 = client
+        .upload_part()
+        .bucket(&bucket)
+        .key(key)
+        .upload_id(&upload_id)
+        .part_number(2)
+        .body(ByteStream::from(part2.clone()))
+        .send()
+        .await?;
+
+    client
+        .complete_multipart_upload()
+        .bucket(&bucket)
+        .key(key)
+        .upload_id(&upload_id)
+        .multipart_upload(
+            CompletedMultipartUpload::builder()
+                .parts(
+                    CompletedPart::builder()
+                        .part_number(1)
+                        .e_tag(uploaded1.e_tag().unwrap_or_default())
+                        .build(),
+                )
+                .parts(
+                    CompletedPart::builder()
+                        .part_number(2)
+                        .e_tag(uploaded2.e_tag().unwrap_or_default())
+                        .build(),
+                )
+                .build(),
+        )
+        .send()
+        .await?;
+
+    let mut expected = part1;
+    expected.extend_from_slice(&part2);
+    for node_idx in 0..dist.cluster.nodes.len() {
+        assert_object_bytes(&dist.client(node_idx)?, &bucket, key, &expected).await?;
+    }
+
+    put_object(&client, &bucket, "list/a", b"a".to_vec()).await?;
+    put_object(&dist.client(2)?, &bucket, "list/b", b"b".to_vec()).await?;
+    let mut seen = Vec::new();
+    for node_idx in 0..dist.cluster.nodes.len() {
+        let listed = dist
+            .client(node_idx)?
+            .list_objects_v2()
+            .bucket(&bucket)
+            .prefix("list/")
+            .send()
+            .await?;
+        let keys: Vec<String> = listed
+            .contents()
+            .iter()
+            .filter_map(|object| object.key().map(str::to_string))
+            .collect();
+        seen.push(keys);
+    }
+    for keys in &seen[1..] {
+        assert_eq!(&seen[0], keys, "list results diverged across nodes: {seen:?}");
+    }
+
+    let got = get_object_bytes(&dist.client(3)?, &bucket, "list/a").await?;
+    assert_eq!(got, b"a");
+    Ok(())
+}
+
+#[tokio::test]
+async fn four_node_list_buckets_agree_across_all_nodes() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let bucket = unique_bucket("listed");
+    dist.create_bucket(&bucket).await?;
+    put_object(&dist.client(0)?, &bucket, "seed.bin", b"seed".to_vec()).await?;
+
+    for node_idx in 0..dist.cluster.nodes.len() {
+        let client = dist.client(node_idx)?;
+        let name = bucket.clone();
+        wait_until(
+            Duration::from_secs(20),
+            || {
+                let client = client.clone();
+                let name = name.clone();
+                async move {
+                    let listed = client.list_buckets().send().await?;
+                    Ok(listed.buckets().iter().any(|entry| entry.name() == Some(name.as_str())))
+                }
+            },
+            &format!("node {node_idx} lists {bucket}"),
+        )
+        .await?;
+        wait_until(
+            Duration::from_secs(20),
+            || {
+                let client = dist.client(node_idx).expect("client");
+                let name = bucket.clone();
+                async move { Ok(get_object_bytes(&client, &name, "seed.bin").await.ok() == Some(b"seed".to_vec())) }
+            },
+            &format!("node {node_idx} reads seed.bin"),
+        )
+        .await?;
+    }
+    Ok(())
+}

crates/e2e_test/src/distributed/mod.rs

@@ -0,0 +1,35 @@
+// 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.
+
+//! 4-node 4-drive distributed e2e coverage.
+//!
+//! Selected by `[profile.e2e-distributed]` and run from
+//! `.github/workflows/e2e-distributed.yml`. Excluded from `e2e-full` because
+//! each case starts four real `rustfs` processes.
+
+mod chaos_test;
+mod concurrency_stability_test;
+mod concurrent_data_movement_test;
+mod data_integrity_movement_test;
+mod expand_decommission_rebalance_test;
+mod extra_test;
+mod harness;
+mod object_lock_test;
+mod observability_test;
+mod replication_quota_test;
+mod s3_basic_test;
+mod s3_during_data_movement_test;
+mod site_replication_test;
+mod upgrade_test;
+mod versioning_test;

crates/e2e_test/src/distributed/object_lock_test.rs

@@ -0,0 +1,219 @@
+// 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, unique_bucket};
+use crate::common::init_logging;
+use crate::object_lock::common::{
+    delete_object_with_bypass, put_object_lock_configuration, put_object_with_legal_hold, put_object_with_retention,
+};
+use aws_sdk_s3::Client;
+use aws_sdk_s3::error::ProvideErrorMetadata;
+use aws_sdk_s3::error::SdkError;
+use aws_sdk_s3::operation::delete_object::DeleteObjectError;
+use aws_sdk_s3::primitives::ByteStream;
+use aws_sdk_s3::types::{
+    DefaultRetention, ObjectLockConfiguration, ObjectLockEnabled, ObjectLockLegalHoldStatus, ObjectLockRetentionMode,
+    ObjectLockRule,
+};
+use chrono::{Duration as ChronoDuration, Utc};
+
+fn delete_denied(error: &SdkError<DeleteObjectError>, context: &str) -> TestResult {
+    let code = error.as_service_error().and_then(ProvideErrorMetadata::code);
+    if code == Some("AccessDenied") {
+        Ok(())
+    } else {
+        Err(format!("{context}: expected AccessDenied, got {error:?}").into())
+    }
+}
+
+async fn expect_versioned_delete_denied(
+    client: &Client,
+    bucket: &str,
+    key: &str,
+    version_id: &str,
+    bypass: bool,
+    context: &str,
+) -> TestResult {
+    match delete_object_with_bypass(client, bucket, key, Some(version_id), bypass).await {
+        Ok(_) => Err(format!("{context}: DeleteObject of retained version must be denied").into()),
+        Err(error) => delete_denied(error.as_ref(), context),
+    }
+}
+
+#[tokio::test]
+async fn four_node_four_drive_object_lock_worm_blocks_delete() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let client = dist.client(0)?;
+    let peer = dist.client(2)?;
+    let bucket = unique_bucket("objlock");
+
+    client
+        .create_bucket()
+        .bucket(&bucket)
+        .object_lock_enabled_for_bucket(true)
+        .send()
+        .await?;
+
+    let retain_until = Utc::now() + ChronoDuration::days(1);
+
+    let compliance_key = "compliance.bin";
+    let compliance_version = put_object_with_retention(
+        &client,
+        &bucket,
+        compliance_key,
+        b"locked-compliance",
+        ObjectLockRetentionMode::Compliance,
+        retain_until,
+    )
+    .await?;
+
+    // Unversioned DELETE is allowed: it only creates a delete marker. WORM
+    // applies to a specific version id.
+    let marker = peer.delete_object().bucket(&bucket).key(compliance_key).send().await?;
+    assert_eq!(
+        marker.delete_marker(),
+        Some(true),
+        "unversioned DELETE on a locked object must create a delete marker"
+    );
+
+    expect_versioned_delete_denied(&peer, &bucket, compliance_key, &compliance_version, false, "COMPLIANCE without bypass")
+        .await?;
+    expect_versioned_delete_denied(&peer, &bucket, compliance_key, &compliance_version, true, "COMPLIANCE with bypass").await?;
+
+    let governance_key = "governance.bin";
+    let governance_version = put_object_with_retention(
+        &client,
+        &bucket,
+        governance_key,
+        b"locked-governance",
+        ObjectLockRetentionMode::Governance,
+        retain_until,
+    )
+    .await?;
+
+    expect_versioned_delete_denied(&peer, &bucket, governance_key, &governance_version, false, "GOVERNANCE without bypass")
+        .await?;
+    delete_object_with_bypass(&peer, &bucket, governance_key, Some(&governance_version), true).await?;
+    let deleted_governance = peer
+        .head_object()
+        .bucket(&bucket)
+        .key(governance_key)
+        .version_id(&governance_version)
+        .send()
+        .await
+        .expect_err("GOVERNANCE bypass must remove the retained version");
+    assert_eq!(
+        deleted_governance.raw_response().map(|response| response.status().as_u16()),
+        Some(404),
+        "deleted GOVERNANCE version returned an unexpected HEAD result: {deleted_governance:?}"
+    );
+
+    let hold_key = "legal-hold.bin";
+    let hold_version =
+        put_object_with_legal_hold(&client, &bucket, hold_key, b"legal-hold", ObjectLockLegalHoldStatus::On).await?;
+    expect_versioned_delete_denied(&peer, &bucket, hold_key, &hold_version, false, "legal hold without bypass").await?;
+    expect_versioned_delete_denied(&peer, &bucket, hold_key, &hold_version, true, "legal hold with bypass").await?;
+
+    Ok(())
+}
+
+#[tokio::test]
+async fn four_node_default_retention_is_visible_and_non_lock_bucket_rejects_configuration() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let writer = dist.client(0)?;
+    let reader = dist.client(3)?;
+    let bucket = unique_bucket("default-lock");
+
+    writer
+        .create_bucket()
+        .bucket(&bucket)
+        .object_lock_enabled_for_bucket(true)
+        .send()
+        .await?;
+    put_object_lock_configuration(&writer, &bucket, ObjectLockRetentionMode::Governance, Some(1), None).await?;
+
+    let key = "default-governance.bin";
+    let put = writer
+        .put_object()
+        .bucket(&bucket)
+        .key(key)
+        .body(ByteStream::from_static(b"default retention payload"))
+        .send()
+        .await?;
+    let version_id = put.version_id().ok_or("default-retained PUT omitted version ID")?;
+
+    let config = reader.get_object_lock_configuration().bucket(&bucket).send().await?;
+    let default_retention = config
+        .object_lock_configuration()
+        .and_then(|configuration| configuration.rule())
+        .and_then(|rule| rule.default_retention())
+        .ok_or("GetObjectLockConfiguration omitted default retention")?;
+    assert_eq!(default_retention.mode().map(|mode| mode.as_str()), Some("GOVERNANCE"));
+    assert_eq!(default_retention.days(), Some(1));
+
+    let retention = reader
+        .get_object_retention()
+        .bucket(&bucket)
+        .key(key)
+        .version_id(version_id)
+        .send()
+        .await?;
+    let retention = retention.retention().ok_or("GetObjectRetention omitted applied retention")?;
+    assert_eq!(retention.mode().map(|mode| mode.as_str()), Some("GOVERNANCE"));
+    let retain_until = retention
+        .retain_until_date()
+        .ok_or("default retention omitted retain-until date")?;
+    assert!(retain_until.secs() > Utc::now().timestamp(), "default retention is not in the future");
+
+    let versioning = reader.get_bucket_versioning().bucket(&bucket).send().await?;
+    assert_eq!(versioning.status().map(|status| status.as_str()), Some("Enabled"));
+    expect_versioned_delete_denied(&reader, &bucket, key, version_id, false, "default GOVERNANCE retention without bypass")
+        .await?;
+
+    let plain_bucket = unique_bucket("no-lock");
+    dist.create_bucket(&plain_bucket).await?;
+    let configuration = ObjectLockConfiguration::builder()
+        .object_lock_enabled(ObjectLockEnabled::Enabled)
+        .rule(
+            ObjectLockRule::builder()
+                .default_retention(
+                    DefaultRetention::builder()
+                        .mode(ObjectLockRetentionMode::Governance)
+                        .days(1)
+                        .build(),
+                )
+                .build(),
+        )
+        .build();
+    let error = writer
+        .put_object_lock_configuration()
+        .bucket(&plain_bucket)
+        .object_lock_configuration(configuration)
+        .send()
+        .await
+        .expect_err("an unversioned bucket must reject Object Lock enablement");
+    let service_error = error
+        .as_service_error()
+        .ok_or("non-lock bucket rejection was not an S3 service error")?;
+    assert_eq!(service_error.code(), Some("InvalidBucketState"), "unexpected error: {error:?}");
+    assert_eq!(
+        service_error.message(),
+        Some("Object Lock configuration cannot be enabled on existing buckets"),
+        "unexpected error: {error:?}"
+    );
+
+    Ok(())
+}

crates/e2e_test/src/distributed/observability_test.rs

@@ -0,0 +1,236 @@
+// 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, cluster_admin_ok, unique_bucket, wait_for_ready};
+use crate::common::{admin_request, init_logging, local_http_client};
+use aws_sdk_s3::operation::RequestId;
+use aws_sdk_s3::primitives::ByteStream;
+use bytes::Bytes;
+use http::Method;
+use http_body_util::{BodyExt, Empty};
+use hyper::body::Incoming;
+use hyper::service::service_fn;
+use hyper::{Request, Response};
+use hyper_util::rt::TokioIo;
+use local_ip_address::local_ip;
+use rustfs_madmin::metrics::RealtimeMetrics;
+use rustfs_utils::egress::ENV_OUTBOUND_ALLOW_ORIGINS;
+use serde_json::Value;
+use std::convert::Infallible;
+use std::time::Duration;
+use tokio::net::TcpListener;
+use tokio::sync::mpsc;
+use tokio::task::JoinHandle;
+use tokio::time::{Instant, timeout};
+
+async fn spawn_audit_collector() -> TestResult<(String, mpsc::UnboundedReceiver<Value>, JoinHandle<()>)> {
+    let listener = TcpListener::bind("0.0.0.0:0").await?;
+    let endpoint = format!("http://{}/audit", std::net::SocketAddr::new(local_ip()?, listener.local_addr()?.port()));
+    let (tx, rx) = mpsc::unbounded_channel();
+    let handle = tokio::spawn(async move {
+        loop {
+            let Ok((stream, _)) = listener.accept().await else {
+                return;
+            };
+            let tx = tx.clone();
+            tokio::spawn(async move {
+                let service = service_fn(move |request: Request<Incoming>| {
+                    let tx = tx.clone();
+                    async move {
+                        let method = request.method().clone();
+                        if let Ok(body) = request.into_body().collect().await
+                            && method == Method::POST
+                            && let Ok(payload) = serde_json::from_slice::<Value>(&body.to_bytes())
+                        {
+                            if let Some(records) = payload["Records"].as_array() {
+                                for entry in records {
+                                    let _ = tx.send(entry.clone());
+                                }
+                            } else {
+                                let _ = tx.send(payload);
+                            }
+                        }
+                        Ok::<_, Infallible>(Response::new(Empty::<Bytes>::new()))
+                    }
+                });
+                let _ = hyper::server::conn::http1::Builder::new()
+                    .serve_connection(TokioIo::new(stream), service)
+                    .await;
+            });
+        }
+    });
+    Ok((endpoint, rx, handle))
+}
+
+async fn wait_for_audit_entry(
+    rx: &mut mpsc::UnboundedReceiver<Value>,
+    bucket: &str,
+    key: &str,
+    request_id: &str,
+) -> TestResult<Value> {
+    let deadline = Instant::now() + Duration::from_secs(30);
+    let mut seen = Vec::new();
+    loop {
+        let remaining = deadline.saturating_duration_since(Instant::now());
+        if remaining.is_zero() {
+            return Err(format!(
+                "audit webhook did not receive PutObject for {bucket}/{key}; received {} other records: {seen:?}",
+                seen.len()
+            )
+            .into());
+        }
+        let entry = match timeout(remaining, rx.recv()).await {
+            Ok(Some(entry)) => entry,
+            Ok(None) => return Err("audit collector stopped before the expected entry arrived".into()),
+            Err(_) => {
+                return Err(format!(
+                    "audit webhook did not receive PutObject for {bucket}/{key}; received {} other records: {seen:?}",
+                    seen.len()
+                )
+                .into());
+            }
+        };
+        if entry["api"]["name"].as_str() == Some("s3:PutObject")
+            && entry["api"]["bucket"].as_str() == Some(bucket)
+            && entry["api"]["object"].as_str() == Some(key)
+            && entry["requestID"].as_str() == Some(request_id)
+        {
+            return Ok(entry);
+        }
+        if seen.len() < 8 {
+            seen.push(format!(
+                "api={:?} bucket={:?} object={:?} requestID={:?}",
+                entry["api"]["name"].as_str(),
+                entry["api"]["bucket"].as_str(),
+                entry["api"]["object"].as_str(),
+                entry["requestID"].as_str()
+            ));
+        }
+    }
+}
+
+#[tokio::test]
+async fn four_node_health_inventory_metrics_and_audit_delivery_are_consistent() -> TestResult {
+    init_logging();
+    let (audit_endpoint, mut audit_entries, collector) = spawn_audit_collector().await?;
+    let audit_origin = reqwest::Url::parse(&audit_endpoint)?.origin().ascii_serialization();
+    let audit_env = [
+        ("RUSTFS_AUDIT_ENABLE", "true"),
+        ("RUSTFS_AUDIT_WEBHOOK_ENABLE_DISTRIBUTED", "on"),
+        ("RUSTFS_AUDIT_WEBHOOK_ENDPOINT_DISTRIBUTED", audit_endpoint.as_str()),
+        (ENV_OUTBOUND_ALLOW_ORIGINS, audit_origin.as_str()),
+    ];
+    let mut dist = DistCluster::new_stopped_with_env(DistLayout::FourByFour, &audit_env).await?;
+    for node_idx in 0..dist.cluster.nodes.len() {
+        let queue_dir = format!("{}/audit-queue-node-{node_idx}", dist.cluster.temp_dir);
+        tokio::fs::create_dir_all(&queue_dir).await?;
+        dist.cluster
+            .set_node_env(node_idx, "RUSTFS_AUDIT_WEBHOOK_QUEUE_DIR_DISTRIBUTED", queue_dir)?;
+    }
+    dist.cluster.start().await?;
+    wait_for_ready(&dist.cluster).await?;
+
+    let http = local_http_client();
+    for node in &dist.cluster.nodes {
+        for probe in ["ready", "live"] {
+            let response = http.get(format!("{}/health/{probe}", node.url)).send().await?;
+            assert!(
+                response.status().is_success(),
+                "node {} {probe} probe failed: {}",
+                node.address,
+                response.status()
+            );
+        }
+    }
+
+    let info_body = cluster_admin_ok(&dist.cluster, Method::GET, "/rustfs/admin/v3/info", None).await?;
+    let info: Value = serde_json::from_str(&info_body)?;
+    let servers = info["info"]["servers"]
+        .as_array()
+        .ok_or_else(|| format!("admin info omitted servers: {info}"))?;
+    assert_eq!(servers.len(), 4, "admin info did not report all four nodes: {info}");
+
+    let storage_body = cluster_admin_ok(&dist.cluster, Method::GET, "/rustfs/admin/v3/storageinfo", None).await?;
+    let storage: Value = serde_json::from_str(&storage_body)?;
+    let disks = storage["info"]["disks"]
+        .as_array()
+        .ok_or_else(|| format!("storageinfo omitted disks: {storage}"))?;
+    assert_eq!(disks.len(), 16, "storageinfo did not report all sixteen drives: {storage}");
+    assert!(
+        disks.iter().all(|disk| {
+            disk["state"].as_str().is_some_and(|state| state.eq_ignore_ascii_case("ok"))
+                && disk["runtimeState"]
+                    .as_str()
+                    .is_some_and(|state| state.eq_ignore_ascii_case("online"))
+        }),
+        "storageinfo reported a drive that was not healthy and online: {storage}"
+    );
+
+    for (node_idx, node) in dist.cluster.nodes.iter().enumerate() {
+        let (status, metrics_body) = admin_request(
+            &node.url,
+            Method::GET,
+            "/rustfs/admin/v3/metrics?n=1&by-host=true&by-disk=true",
+            None,
+            &dist.cluster.access_key,
+            &dist.cluster.secret_key,
+        )
+        .await?;
+        assert!(status.is_success(), "node {node_idx} metrics failed: {status} {metrics_body}");
+        let sample: RealtimeMetrics = serde_json::from_str(
+            metrics_body
+                .lines()
+                .next()
+                .ok_or_else(|| format!("node {node_idx} returned empty metrics"))?,
+        )?;
+        assert!(sample.finally, "node {node_idx} metrics sample was not terminal");
+        assert!(sample.errors.is_empty(), "node {node_idx} metrics reported errors: {:?}", sample.errors);
+        assert!(!sample.hosts.is_empty(), "node {node_idx} metrics omitted hosts");
+    }
+
+    let targets_body = cluster_admin_ok(&dist.cluster, Method::GET, "/rustfs/admin/v3/audit/target/list", None).await?;
+    let targets: Value = serde_json::from_str(&targets_body)?;
+    let configured = targets["audit_endpoints"]
+        .as_array()
+        .ok_or_else(|| format!("audit target list omitted audit_endpoints: {targets}"))?
+        .iter()
+        .any(|target| target["account_id"].as_str() == Some("distributed") && target["service"].as_str() == Some("webhook"));
+    assert!(configured, "configured audit webhook was missing: {targets}");
+
+    let bucket = unique_bucket("audit");
+    dist.create_bucket(&bucket).await?;
+    let key = "correlated/audit-object.bin";
+    let put = dist
+        .client(2)?
+        .put_object()
+        .bucket(&bucket)
+        .key(key)
+        .body(ByteStream::from_static(b"distributed audit payload"))
+        .send()
+        .await?;
+    let request_id = put.request_id().ok_or("PutObject response omitted request ID")?;
+    let audit = wait_for_audit_entry(&mut audit_entries, &bucket, key, request_id).await?;
+    assert_eq!(
+        audit["api"]["status_code"].as_i64(),
+        Some(200),
+        "audit entry did not report success: {audit}"
+    );
+    assert!(
+        !audit.to_string().contains(&dist.cluster.secret_key),
+        "audit entry leaked the root secret key"
+    );
+
+    collector.abort();
+    Ok(())
+}

crates/e2e_test/src/distributed/replication_quota_test.rs

@@ -0,0 +1,191 @@
+// 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/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, enable_versioning, put_bucket_replication, put_object, retrying_put, set_bucket_quota,
+    set_remote_target, unique_bucket, wait_for_ready, wait_for_replicated_bytes, wait_until,
+};
+use crate::common::{FAST_DATA_USAGE_SCANNER_ENV, init_logging};
+use aws_sdk_s3::error::ProvideErrorMetadata;
+use aws_sdk_s3::primitives::ByteStream;
+use http::Method;
+use std::time::Duration;
+
+async fn wait_for_replication_status(
+    client: &aws_sdk_s3::Client,
+    bucket: &str,
+    key: &str,
+    expected: &[&str],
+    timeout: Duration,
+) -> TestResult {
+    wait_until(
+        timeout,
+        || async {
+            let head = client.head_object().bucket(bucket).key(key).send().await?;
+            Ok(head
+                .replication_status()
+                .is_some_and(|status| expected.contains(&status.as_str())))
+        },
+        &format!("replication status for {bucket}/{key} in {expected:?}"),
+    )
+    .await
+}
+
+#[tokio::test]
+async fn four_node_bucket_replication_converges_to_peer_cluster() -> TestResult {
+    init_logging();
+    let (source, mut target) = DistCluster::start_replication_pair().await?;
+    let source_bucket = unique_bucket("replsrc");
+    let target_bucket = unique_bucket("repldst");
+    source.create_bucket(&source_bucket).await?;
+    target.create_bucket(&target_bucket).await?;
+
+    let source_client = source.client(0)?;
+    let target_client = target.client(0)?;
+    enable_versioning(&source_client, &source_bucket).await?;
+    enable_versioning(&target_client, &target_bucket).await?;
+
+    let arn = set_remote_target(&source.cluster, &source_bucket, &target.cluster, &target_bucket).await?;
+    put_bucket_replication(&source.cluster, &source_bucket, &arn).await?;
+
+    let key = "replicated/metadata-and-tags.bin";
+    let body = b"distributed-bucket-replication".to_vec();
+    source_client
+        .put_object()
+        .bucket(&source_bucket)
+        .key(key)
+        .metadata("origin", "four-node-source")
+        .tagging("suite=distributed&shape=metadata")
+        .body(ByteStream::from(body.clone()))
+        .send()
+        .await?;
+    wait_for_replicated_bytes(&target_client, &target_bucket, key, &body, Duration::from_secs(45)).await?;
+    wait_for_replication_status(&source_client, &source_bucket, key, &["COMPLETED"], Duration::from_secs(30)).await?;
+
+    let peer_read = target.client(3)?;
+    wait_for_replicated_bytes(&peer_read, &target_bucket, key, &body, Duration::from_secs(15)).await?;
+    let replica_head = peer_read.head_object().bucket(&target_bucket).key(key).send().await?;
+    assert_eq!(
+        replica_head
+            .metadata()
+            .and_then(|metadata| metadata.get("origin"))
+            .map(String::as_str),
+        Some("four-node-source")
+    );
+    assert_eq!(replica_head.replication_status().map(|status| status.as_str()), Some("REPLICA"));
+    let replica_tags = peer_read.get_object_tagging().bucket(&target_bucket).key(key).send().await?;
+    let tags: std::collections::BTreeMap<_, _> = replica_tags.tag_set().iter().map(|tag| (tag.key(), tag.value())).collect();
+    assert_eq!(tags.get("suite"), Some(&"distributed"));
+    assert_eq!(tags.get("shape"), Some(&"metadata"));
+
+    target.cluster.stop();
+    let outage_key = "replicated/queued-during-target-outage.bin";
+    let outage_body = b"retry-after-target-restart".to_vec();
+    put_object(&source_client, &source_bucket, outage_key, outage_body.clone()).await?;
+    wait_for_replication_status(
+        &source_client,
+        &source_bucket,
+        outage_key,
+        &["PENDING", "FAILED"],
+        Duration::from_secs(30),
+    )
+    .await?;
+
+    target.cluster.start().await?;
+    wait_for_ready(&target.cluster).await?;
+    wait_for_replicated_bytes(&target.client(2)?, &target_bucket, outage_key, &outage_body, Duration::from_secs(90)).await?;
+    wait_for_replication_status(&source_client, &source_bucket, outage_key, &["COMPLETED"], Duration::from_secs(45)).await?;
+    Ok(())
+}
+
+#[tokio::test]
+async fn four_node_four_drive_hard_quota_rejects_over_limit_put() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start_with_env(DistLayout::FourByFour, FAST_DATA_USAGE_SCANNER_ENV).await?;
+    let bucket = unique_bucket("quota");
+    dist.create_bucket(&bucket).await?;
+    set_bucket_quota(&dist.cluster, &bucket, 8 * 1024).await?;
+
+    let client = dist.client(1)?;
+    retrying_put(&client, &bucket, "small.bin", vec![0u8; 1024], Duration::from_secs(30)).await?;
+    wait_until(
+        Duration::from_secs(30),
+        || async {
+            let (status, body) = super::harness::cluster_admin(
+                &dist.cluster,
+                Method::GET,
+                &format!("/rustfs/admin/v3/quota-stats/{bucket}"),
+                None,
+            )
+            .await?;
+            if !status.is_success() {
+                return Ok(false);
+            }
+            let stats: serde_json::Value =
+                serde_json::from_str(&body).map_err(|error| format!("quota stats returned invalid JSON: {error}: {body}"))?;
+            let usage = stats
+                .get("current_usage")
+                .and_then(serde_json::Value::as_u64)
+                .ok_or_else(|| format!("quota stats omitted current_usage: {stats}"))?;
+            Ok(usage >= 1024)
+        },
+        "quota stats observe small object",
+    )
+    .await?;
+
+    let oversized_key = "too-big.bin";
+    let error = client
+        .put_object()
+        .bucket(&bucket)
+        .key(oversized_key)
+        .body(vec![0u8; 16 * 1024].into())
+        .send()
+        .await
+        .expect_err("hard quota must reject the oversized PUT");
+    let service_error = error
+        .as_service_error()
+        .ok_or("quota rejection was not an S3 service error")?;
+    assert_eq!(
+        error.raw_response().map(|response| response.status().as_u16()),
+        Some(400),
+        "quota rejection must be HTTP 400: {error:?}"
+    );
+    assert_eq!(service_error.code(), Some("InvalidRequest"), "unexpected quota error: {error:?}");
+    assert!(
+        service_error
+            .message()
+            .is_some_and(|message| message.starts_with("Bucket quota exceeded")),
+        "PUT must fail specifically at quota admission: {error:?}"
+    );
+
+    let missing = client
+        .head_object()
+        .bucket(&bucket)
+        .key(oversized_key)
+        .send()
+        .await
+        .expect_err("an object rejected by quota must not become visible");
+    assert_eq!(
+        missing.raw_response().map(|response| response.status().as_u16()),
+        Some(404),
+        "quota-rejected object returned an unexpected HEAD result: {missing:?}"
+    );
+
+    let listed = client.list_objects_v2().bucket(&bucket).send().await?;
+    assert!(
+        listed.contents().iter().all(|object| object.key() != Some(oversized_key)),
+        "quota-rejected key leaked into ListObjectsV2"
+    );
+    Ok(())
+}

crates/e2e_test/src/distributed/s3_basic_test.rs

@@ -0,0 +1,258 @@
+// 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_object_bytes, get_object_bytes, put_object, unique_bucket};
+use crate::common::{init_logging, local_http_client};
+use aws_sdk_s3::error::ProvideErrorMetadata;
+use aws_sdk_s3::presigning::PresigningConfig;
+use aws_sdk_s3::primitives::ByteStream;
+use aws_sdk_s3::types::{Delete, MetadataDirective, ObjectIdentifier};
+use std::time::Duration;
+
+#[tokio::test]
+async fn four_node_four_drive_s3_put_get_head_list_copy_rename_delete_and_presign() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let bucket = unique_bucket("s3basic");
+    dist.create_bucket(&bucket).await?;
+
+    let writer = dist.client(0)?;
+    let reader = dist.client(3)?;
+    let key = "dir/object.bin";
+    let body = vec![0xA5u8; 256 * 1024];
+    put_object(&writer, &bucket, key, body.clone()).await?;
+
+    let head = reader.head_object().bucket(&bucket).key(key).send().await?;
+    assert_eq!(head.content_length(), Some(body.len() as i64));
+    assert_object_bytes(&reader, &bucket, key, &body).await?;
+
+    let ranged = reader
+        .get_object()
+        .bucket(&bucket)
+        .key(key)
+        .range("bytes=0-15")
+        .send()
+        .await?;
+    let ranged_body = ranged.body.collect().await?.into_bytes();
+    assert_eq!(ranged_body.as_ref(), &body[..16]);
+
+    let listed = reader.list_objects_v2().bucket(&bucket).prefix("dir/").send().await?;
+    let keys: Vec<_> = listed.contents().iter().filter_map(|object| object.key()).collect();
+    assert_eq!(keys, vec![key]);
+
+    let copy_key = "dir/object-copy.bin";
+    reader
+        .copy_object()
+        .bucket(&bucket)
+        .key(copy_key)
+        .copy_source(format!("{bucket}/{key}"))
+        .metadata_directive(MetadataDirective::Copy)
+        .send()
+        .await?;
+    assert_object_bytes(&writer, &bucket, copy_key, &body).await?;
+
+    let moved_key = "dir/object-moved.bin";
+    writer
+        .copy_object()
+        .bucket(&bucket)
+        .key(moved_key)
+        .copy_source(format!("{bucket}/{copy_key}"))
+        .send()
+        .await?;
+    writer.delete_object().bucket(&bucket).key(copy_key).send().await?;
+    match writer.head_object().bucket(&bucket).key(copy_key).send().await {
+        Ok(_) => return Err("copied source still present after rename delete".into()),
+        Err(error) if error.as_service_error().is_some_and(|err| err.is_not_found()) => {}
+        Err(error) => return Err(error.into()),
+    }
+    assert_object_bytes(&reader, &bucket, moved_key, &body).await?;
+
+    let presigned = writer
+        .get_object()
+        .bucket(&bucket)
+        .key(key)
+        .presigned(PresigningConfig::expires_in(Duration::from_secs(120))?)
+        .await?;
+    let response = local_http_client().get(presigned.uri().to_string()).send().await?;
+    assert!(response.status().is_success(), "presigned GET failed: {}", response.status());
+    let presigned_body = response.bytes().await?;
+    assert_eq!(presigned_body.as_ref(), body.as_slice());
+
+    let empty_key = "empty";
+    put_object(&writer, &bucket, empty_key, Vec::new()).await?;
+    let empty = get_object_bytes(&reader, &bucket, empty_key).await?;
+    assert!(empty.is_empty());
+
+    let deleted = writer
+        .delete_objects()
+        .bucket(&bucket)
+        .delete(
+            Delete::builder()
+                .objects(ObjectIdentifier::builder().key(key).build()?)
+                .objects(ObjectIdentifier::builder().key(moved_key).build()?)
+                .objects(ObjectIdentifier::builder().key(empty_key).build()?)
+                .build()?,
+        )
+        .send()
+        .await?;
+    assert!(deleted.errors().is_empty(), "DeleteObjects reported failures: {deleted:?}");
+    assert_eq!(deleted.deleted().len(), 3, "DeleteObjects did not acknowledge every key");
+
+    let remaining = reader.list_objects_v2().bucket(&bucket).send().await?;
+    assert!(remaining.contents().is_empty(), "bucket still has objects after delete");
+    Ok(())
+}
+
+#[tokio::test]
+async fn four_node_s3_metadata_tags_special_keys_pagination_and_multipart_abort() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let bucket = unique_bucket("s3matrix");
+    dist.create_bucket(&bucket).await?;
+    let writer = dist.client(0)?;
+    let reader = dist.client(3)?;
+
+    let special_key = "unicode/测试 space+percent%25.txt";
+    let special_body = b"metadata and tagging survive distributed routing".to_vec();
+    let put = writer
+        .put_object()
+        .bucket(&bucket)
+        .key(special_key)
+        .metadata("test-meta", "distributed")
+        .tagging("purpose=compatibility&scope=four-by-four")
+        .body(ByteStream::from(special_body.clone()))
+        .send()
+        .await?;
+    let etag = put.e_tag().ok_or("PutObject omitted ETag")?.to_string();
+
+    let head = reader.head_object().bucket(&bucket).key(special_key).send().await?;
+    assert_eq!(
+        head.metadata()
+            .and_then(|metadata| metadata.get("test-meta"))
+            .map(String::as_str),
+        Some("distributed")
+    );
+    assert_eq!(head.e_tag(), Some(etag.as_str()));
+    let tags = reader.get_object_tagging().bucket(&bucket).key(special_key).send().await?;
+    let actual_tags: std::collections::BTreeMap<_, _> = tags
+        .tag_set()
+        .iter()
+        .map(|tag| (tag.key().to_string(), tag.value().to_string()))
+        .collect();
+    assert_eq!(actual_tags.get("purpose").map(String::as_str), Some("compatibility"));
+    assert_eq!(actual_tags.get("scope").map(String::as_str), Some("four-by-four"));
+
+    let conditional = reader
+        .get_object()
+        .bucket(&bucket)
+        .key(special_key)
+        .if_match(&etag)
+        .send()
+        .await?;
+    assert_eq!(conditional.body.collect().await?.into_bytes().as_ref(), special_body.as_slice());
+    let invalid_range = reader
+        .get_object()
+        .bucket(&bucket)
+        .key(special_key)
+        .range("bytes=999999-1000000")
+        .send()
+        .await
+        .expect_err("an unsatisfiable range must fail");
+    assert_eq!(
+        invalid_range.as_service_error().and_then(ProvideErrorMetadata::code),
+        Some("InvalidRange"),
+        "unexpected invalid-range error: {invalid_range:?}"
+    );
+
+    let upload_key = "multipart/aborted.bin";
+    let upload = writer
+        .create_multipart_upload()
+        .bucket(&bucket)
+        .key(upload_key)
+        .send()
+        .await?;
+    let upload_id = upload.upload_id().ok_or("CreateMultipartUpload omitted upload ID")?;
+    writer
+        .upload_part()
+        .bucket(&bucket)
+        .key(upload_key)
+        .upload_id(upload_id)
+        .part_number(1)
+        .body(ByteStream::from(vec![0x5Au8; 5 * 1024 * 1024]))
+        .send()
+        .await?;
+    let pending = reader
+        .list_multipart_uploads()
+        .bucket(&bucket)
+        .prefix("multipart/")
+        .send()
+        .await?;
+    assert!(pending.uploads().iter().any(|entry| entry.upload_id() == Some(upload_id)));
+    writer
+        .abort_multipart_upload()
+        .bucket(&bucket)
+        .key(upload_key)
+        .upload_id(upload_id)
+        .send()
+        .await?;
+    let after_abort = reader
+        .list_multipart_uploads()
+        .bucket(&bucket)
+        .prefix("multipart/")
+        .send()
+        .await?;
+    assert!(after_abort.uploads().iter().all(|entry| entry.upload_id() != Some(upload_id)));
+    let aborted_head = reader
+        .head_object()
+        .bucket(&bucket)
+        .key(upload_key)
+        .send()
+        .await
+        .expect_err("aborted multipart upload must not create an object");
+    assert_eq!(
+        aborted_head.raw_response().map(|response| response.status().as_u16()),
+        Some(404),
+        "aborted multipart object returned an unexpected HEAD result: {aborted_head:?}"
+    );
+
+    for index in 0..113 {
+        let key = format!("page/{index:04}.txt");
+        put_object(&writer, &bucket, &key, format!("page-{index}").into_bytes()).await?;
+    }
+    let mut token = None;
+    let mut paged_keys = Vec::new();
+    loop {
+        let page = reader
+            .list_objects_v2()
+            .bucket(&bucket)
+            .prefix("page/")
+            .max_keys(37)
+            .set_continuation_token(token.take())
+            .send()
+            .await?;
+        paged_keys.extend(page.contents().iter().filter_map(|object| object.key().map(str::to_string)));
+        if page.is_truncated() != Some(true) {
+            break;
+        }
+        token = Some(
+            page.next_continuation_token()
+                .ok_or("truncated ListObjectsV2 page omitted next continuation token")?
+                .to_string(),
+        );
+    }
+    assert_eq!(paged_keys.len(), 113);
+    let expected: Vec<_> = (0..113).map(|index| format!("page/{index:04}.txt")).collect();
+    assert_eq!(paged_keys, expected, "pagination lost, duplicated, or reordered keys");
+    Ok(())
+}

crates/e2e_test/src/distributed/s3_during_data_movement_test.rs

@@ -0,0 +1,94 @@
+// 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/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::{
+    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,
+};
+use crate::common::init_logging;
+use std::time::Duration;
+
+#[tokio::test]
+async fn s3_put_get_list_succeed_during_decommission_and_rebalance() -> TestResult {
+    init_logging();
+    let mut dist = DistCluster::start(DistLayout::SingleNodeFourDrive).await?;
+    let bucket = unique_bucket("s3move");
+    dist.create_bucket(&bucket).await?;
+    let client = dist.client(0)?;
+    let inventory = put_inventory_retrying(&client, &bucket, 96, 256 * 1024, Duration::from_secs(30)).await?;
+    dist.expand_to_four_pools().await?;
+
+    start_decommission(&dist.cluster, DECOMMISSION_POOL_ID).await?;
+    wait_for_decommission_running_with_progress(&dist.cluster, DECOMMISSION_POOL_ID, Duration::from_secs(30)).await?;
+    let live = dist.client(2)?;
+    retrying_put(
+        &live,
+        &bucket,
+        "during-decommission.bin",
+        b"written-while-decommissioning".to_vec(),
+        Duration::from_secs(30),
+    )
+    .await?;
+    retrying_get_equals(
+        &live,
+        &bucket,
+        "during-decommission.bin",
+        b"written-while-decommissioning",
+        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-decommission.bin")),
+        "list during decommission missed the newly written key"
+    );
+    let status = decommission_status_json(&dist.cluster).await?;
+    if !decommission_running_with_progress(&status, DECOMMISSION_POOL_ID)? {
+        return Err(format!("decommission did not remain active across the S3 operations: {status}").into());
+    }
+
+    wait_for_decommission_complete(&dist.cluster, DECOMMISSION_POOL_ID, Duration::from_secs(180)).await?;
+    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?;
+    retrying_put(
+        &live,
+        &bucket,
+        "during-rebalance.bin",
+        b"written-while-rebalancing".to_vec(),
+        Duration::from_secs(30),
+    )
+    .await?;
+    retrying_get_equals(
+        &live,
+        &bucket,
+        "during-rebalance.bin",
+        b"written-while-rebalancing",
+        Duration::from_secs(30),
+    )
+    .await?;
+    let status = rebalance_status_json(&dist.cluster).await?;
+    if !rebalance_running_with_progress(&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?;
+    Ok(())
+}

crates/e2e_test/src/distributed/site_replication_test.rs

@@ -0,0 +1,128 @@
+// 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/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, TestResult, cluster_admin_ok, enable_versioning, put_object, unique_bucket, wait_for_replicated_bytes,
+    wait_until,
+};
+use crate::common::{init_logging, signed_request};
+use http::{Method, StatusCode};
+use rustfs_madmin::{PeerSite, ReplicateAddStatus, SiteReplicationInfo, SyncStatus};
+use std::time::Duration;
+
+async fn site_replication_add(
+    cluster: &crate::common::RustFSTestClusterEnvironment,
+    sites: &[PeerSite],
+) -> TestResult<ReplicateAddStatus> {
+    let url = format!("{}/rustfs/admin/v3/site-replication/add?replicateILMExpiry=false", cluster.nodes[0].url);
+    let response = signed_request(
+        Method::PUT,
+        &url,
+        &cluster.access_key,
+        &cluster.secret_key,
+        Some(serde_json::to_vec(sites)?),
+        Some("application/json"),
+    )
+    .await?;
+    if response.status() != StatusCode::OK {
+        let status = response.status();
+        let body = response.text().await.unwrap_or_default();
+        return Err(format!("site replication add failed: {status} {body}").into());
+    }
+    Ok(serde_json::from_slice(&response.bytes().await?)?)
+}
+
+async fn site_replication_info(cluster: &crate::common::RustFSTestClusterEnvironment) -> TestResult<SiteReplicationInfo> {
+    let body = cluster_admin_ok(cluster, Method::GET, "/rustfs/admin/v3/site-replication/info", None).await?;
+    Ok(serde_json::from_str(&body)?)
+}
+
+async fn wait_for_site_replication_enabled(cluster: &crate::common::RustFSTestClusterEnvironment) -> TestResult {
+    wait_until(
+        Duration::from_secs(30),
+        || async {
+            let info = site_replication_info(cluster).await?;
+            Ok(info.enabled && info.sites.len() == 2 && info.sites.iter().all(|site| site.sync_state == SyncStatus::Enable))
+        },
+        "site replication enabled with two synchronized sites",
+    )
+    .await
+}
+
+#[tokio::test]
+async fn four_node_site_replication_replicates_object_to_peer_site() -> TestResult {
+    init_logging();
+    let (site_a, site_b) = DistCluster::start_replication_pair().await?;
+    let bucket = unique_bucket("siterepl");
+    site_a.create_bucket(&bucket).await?;
+    site_b.create_bucket(&bucket).await?;
+
+    let client_a = site_a.client(0)?;
+    let client_b = site_b.client(0)?;
+    enable_versioning(&client_a, &bucket).await?;
+    enable_versioning(&client_b, &bucket).await?;
+
+    let sites = vec![
+        PeerSite {
+            name: "site-a".to_string(),
+            endpoint: site_a.cluster.nodes[0].url.clone(),
+            access_key: site_a.cluster.access_key.clone(),
+            secret_key: site_a.cluster.secret_key.clone(),
+            ..Default::default()
+        },
+        PeerSite {
+            name: "site-b".to_string(),
+            endpoint: site_b.cluster.nodes[0].url.clone(),
+            access_key: site_b.cluster.access_key.clone(),
+            secret_key: site_b.cluster.secret_key.clone(),
+            ..Default::default()
+        },
+    ];
+    let add_status = site_replication_add(&site_a.cluster, &sites).await?;
+    assert!(
+        add_status.success && add_status.err_detail.is_empty() && add_status.initial_sync_error_message.is_empty(),
+        "site replication add reported failure: {add_status:?}"
+    );
+    wait_for_site_replication_enabled(&site_a.cluster).await?;
+    wait_for_site_replication_enabled(&site_b.cluster).await?;
+
+    let info_a = site_replication_info(&site_a.cluster).await?;
+    let remote = info_a
+        .sites
+        .iter()
+        .find(|site| site.name == "site-b")
+        .ok_or_else(|| format!("site A info omitted the configured site-b peer: {info_a:?}"))?;
+    assert_eq!(remote.endpoint, site_b.cluster.nodes[0].url);
+    let deployment_ids: std::collections::BTreeSet<_> = info_a.sites.iter().map(|site| site.deployment_id.as_str()).collect();
+    assert!(
+        deployment_ids.iter().all(|deployment_id| !deployment_id.is_empty()) && deployment_ids.len() == 2,
+        "site peers must have two distinct non-empty deployment IDs: {info_a:?}"
+    );
+    assert!(info_a.retry_stats.is_none(), "site A has pending replication retries: {info_a:?}");
+    assert!(info_a.pending_operation.is_none(), "site A has a pending operation: {info_a:?}");
+
+    let key = "site-object.bin";
+    let body = b"four-node-site-replication".to_vec();
+    put_object(&client_a, &bucket, key, body.clone()).await?;
+    wait_for_replicated_bytes(&client_b, &bucket, key, &body, Duration::from_secs(60)).await?;
+
+    let peer_b = site_b.client(3)?;
+    wait_for_replicated_bytes(&peer_b, &bucket, key, &body, Duration::from_secs(20)).await?;
+
+    let reverse_key = "reverse/site-object.bin";
+    let reverse_body = b"site-b-to-site-a".to_vec();
+    put_object(&site_b.client(2)?, &bucket, reverse_key, reverse_body.clone()).await?;
+    wait_for_replicated_bytes(&site_a.client(3)?, &bucket, reverse_key, &reverse_body, Duration::from_secs(60)).await?;
+    Ok(())
+}

crates/e2e_test/src/distributed/upgrade_test.rs

@@ -0,0 +1,345 @@
+// 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/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.
+
+//! 4-node upgrade coverage for historical objects and IAM AK/SK.
+//!
+//! Complements `upgrade_compatibility_test` (single-node SSE/multipart and
+//! mixed-version listing). This module pins the distributed contract the
+//! hardware upgrade chain is meant to catch: after a 4-node upgrade, objects
+//! written on the previous release still read back, and IAM user credentials
+//! created before the upgrade still authenticate.
+//!
+//! Requires `RUSTFS_UPGRADE_SOURCE_BINARY` pointing at the pinned previous
+//! release. The `e2e-distributed` workflow downloads that binary; a local run
+//! without it fails closed rather than skipping.
+
+use super::harness::{
+    DistCluster, DistLayout, TestResult, assert_object_bytes, cluster_admin_ok, enable_versioning, get_object_bytes, put_object,
+    unique_bucket, wait_until,
+};
+use crate::common::{
+    AdminTransport, admin_add_canned_policy_via, admin_attach_user_policy_via, admin_create_user_via, init_logging,
+};
+use aws_sdk_s3::Client;
+use aws_sdk_s3::error::ProvideErrorMetadata;
+use std::path::{Path, PathBuf};
+use std::time::Duration;
+use uuid::Uuid;
+
+const SOURCE_BINARY_ENV: &str = "RUSTFS_UPGRADE_SOURCE_BINARY";
+const IAM_SECRET: &str = "UpgradeTestSecretKey1";
+const WRONG_SECRET: &str = "WrongSecretKey000000";
+const CREDENTIAL_TIMEOUT: Duration = Duration::from_secs(30);
+
+struct UpgradeSeed {
+    history_bucket: String,
+    history_key: &'static str,
+    history_body: Vec<u8>,
+    versioned_bucket: String,
+    versioned_key: &'static str,
+    version1: String,
+    version1_body: Vec<u8>,
+    version2: String,
+    version2_body: Vec<u8>,
+    iam_bucket: String,
+    iam_key: &'static str,
+    iam_body: Vec<u8>,
+    iam_user: String,
+    iam_secret: &'static str,
+}
+
+fn source_binary() -> TestResult<PathBuf> {
+    let path = std::env::var_os(SOURCE_BINARY_ENV).map(PathBuf::from).ok_or_else(|| {
+        format!(
+            "{SOURCE_BINARY_ENV} must point to the pinned previous release binary (the e2e-distributed workflow downloads it)"
+        )
+    })?;
+    if !path.is_file() {
+        return Err(format!("upgrade source binary does not exist: {}", path.display()).into());
+    }
+    Ok(path)
+}
+
+fn capture_upgrade_logs(cluster: &mut DistCluster, label: &str) -> TestResult {
+    let Some(log_dir) = std::env::var_os("RUSTFS_E2E_LOG_DIR") else {
+        return Ok(());
+    };
+    std::fs::create_dir_all(&log_dir)?;
+    for node_idx in 0..cluster.cluster.nodes.len() {
+        let path = Path::new(&log_dir).join(format!("{label}-node-{node_idx}.log"));
+        cluster
+            .cluster
+            .set_node_capture_log_path(node_idx, path.to_string_lossy().into_owned())?;
+    }
+    Ok(())
+}
+
+fn iam_rw_policy(bucket: &str) -> String {
+    serde_json::json!({
+        "Version": "2012-10-17",
+        "Statement": [{
+            "Effect": "Allow",
+            "Action": ["s3:*"],
+            "Resource": [
+                format!("arn:aws:s3:::{bucket}"),
+                format!("arn:aws:s3:::{bucket}/*")
+            ]
+        }]
+    })
+    .to_string()
+}
+
+async fn create_iam_user(dist: &DistCluster, user: &str, secret: &str, policy_name: &str, bucket: &str) -> TestResult {
+    let url = &dist.cluster.nodes[0].url;
+    let access = &dist.cluster.access_key;
+    let admin_secret = &dist.cluster.secret_key;
+    admin_create_user_via(AdminTransport::Signed, url, access, admin_secret, user, secret).await?;
+    admin_add_canned_policy_via(AdminTransport::Signed, url, access, admin_secret, policy_name, &iam_rw_policy(bucket)).await?;
+    admin_attach_user_policy_via(AdminTransport::Signed, url, access, admin_secret, policy_name, user).await?;
+    Ok(())
+}
+
+async fn wait_for_put(client: &Client, bucket: &str, key: &str, body: Vec<u8>, label: &str) -> TestResult {
+    wait_until(
+        CREDENTIAL_TIMEOUT,
+        || {
+            let client = client.clone();
+            let bucket = bucket.to_string();
+            let key = key.to_string();
+            let body = body.clone();
+            async move {
+                put_object(&client, &bucket, &key, body).await?;
+                Ok(true)
+            }
+        },
+        label,
+    )
+    .await
+}
+
+async fn wait_for_bytes(client: &Client, bucket: &str, key: &str, expected: &[u8], label: &str) -> TestResult {
+    wait_until(
+        CREDENTIAL_TIMEOUT,
+        || {
+            let client = client.clone();
+            let bucket = bucket.to_string();
+            let key = key.to_string();
+            let expected = expected.to_vec();
+            async move {
+                let got = get_object_bytes(&client, &bucket, &key).await?;
+                Ok(got == expected)
+            }
+        },
+        label,
+    )
+    .await
+}
+
+async fn seed_history_and_iam(dist: &DistCluster) -> TestResult<UpgradeSeed> {
+    let history_bucket = unique_bucket("upg-hist");
+    let versioned_bucket = unique_bucket("upg-ver");
+    let iam_bucket = unique_bucket("upg-iam");
+    dist.create_bucket(&history_bucket).await?;
+    dist.create_bucket(&versioned_bucket).await?;
+    dist.create_bucket(&iam_bucket).await?;
+
+    let root = dist.client(0)?;
+    enable_versioning(&root, &versioned_bucket).await?;
+
+    let history_key = "plain-history.bin";
+    let history_body = b"written by the previous 4-node release".to_vec();
+    put_object(&root, &history_bucket, history_key, history_body.clone()).await?;
+
+    let versioned_key = "versioned-history.txt";
+    let version1_body = b"version-one-before-upgrade".to_vec();
+    let version1 = root
+        .put_object()
+        .bucket(&versioned_bucket)
+        .key(versioned_key)
+        .body(aws_sdk_s3::primitives::ByteStream::from(version1_body.clone()))
+        .send()
+        .await?
+        .version_id()
+        .ok_or("first versioned PUT omitted version ID")?
+        .to_string();
+    let version2_body = b"version-two-before-upgrade".to_vec();
+    let version2 = root
+        .put_object()
+        .bucket(&versioned_bucket)
+        .key(versioned_key)
+        .body(aws_sdk_s3::primitives::ByteStream::from(version2_body.clone()))
+        .send()
+        .await?
+        .version_id()
+        .ok_or("second versioned PUT omitted version ID")?
+        .to_string();
+
+    let iam_user = format!("upg{}", &Uuid::new_v4().simple().to_string()[..8]);
+    let policy_name = format!("upgpol{}", &Uuid::new_v4().simple().to_string()[..8]);
+    create_iam_user(dist, &iam_user, IAM_SECRET, &policy_name, &iam_bucket).await?;
+
+    let iam_key = "iam-history.bin";
+    let iam_body = b"written with pre-upgrade IAM AK/SK".to_vec();
+    let iam_client = dist.client_with_credentials(1, &iam_user, IAM_SECRET)?;
+    wait_for_put(&iam_client, &iam_bucket, iam_key, iam_body.clone(), "IAM user PUT before upgrade").await?;
+
+    Ok(UpgradeSeed {
+        history_bucket,
+        history_key,
+        history_body,
+        versioned_bucket,
+        versioned_key,
+        version1,
+        version1_body,
+        version2,
+        version2_body,
+        iam_bucket,
+        iam_key,
+        iam_body,
+        iam_user,
+        iam_secret: IAM_SECRET,
+    })
+}
+
+async fn assert_history_and_iam(dist: &DistCluster, seed: &UpgradeSeed, context: &str) -> TestResult {
+    let root_a = dist.client(0)?;
+    let root_b = dist.client(3)?;
+    wait_for_bytes(
+        &root_b,
+        &seed.history_bucket,
+        seed.history_key,
+        &seed.history_body,
+        &format!("{context}: root GET historical object"),
+    )
+    .await?;
+    assert_object_bytes(&root_a, &seed.history_bucket, seed.history_key, &seed.history_body).await?;
+
+    let v1 = root_b
+        .get_object()
+        .bucket(&seed.versioned_bucket)
+        .key(seed.versioned_key)
+        .version_id(&seed.version1)
+        .send()
+        .await?;
+    let v1_body = v1.body.collect().await?.into_bytes();
+    if v1_body.as_ref() != seed.version1_body.as_slice() {
+        return Err(format!("{context}: version 1 bytes changed after upgrade").into());
+    }
+    let v2 = root_a
+        .get_object()
+        .bucket(&seed.versioned_bucket)
+        .key(seed.versioned_key)
+        .version_id(&seed.version2)
+        .send()
+        .await?;
+    let v2_body = v2.body.collect().await?.into_bytes();
+    if v2_body.as_ref() != seed.version2_body.as_slice() {
+        return Err(format!("{context}: version 2 bytes changed after upgrade").into());
+    }
+
+    let users = cluster_admin_ok(&dist.cluster, http::Method::GET, "/rustfs/admin/v3/list-users", None).await?;
+    if !users.contains(&seed.iam_user) {
+        return Err(format!("{context}: list-users lost IAM user {}: {users}", seed.iam_user).into());
+    }
+
+    let iam_on_upgraded = dist.client_with_credentials(0, &seed.iam_user, seed.iam_secret)?;
+    let iam_on_peer = dist.client_with_credentials(3, &seed.iam_user, seed.iam_secret)?;
+    wait_for_bytes(
+        &iam_on_upgraded,
+        &seed.iam_bucket,
+        seed.iam_key,
+        &seed.iam_body,
+        &format!("{context}: IAM GET historical object on node 0"),
+    )
+    .await?;
+    wait_for_bytes(
+        &iam_on_peer,
+        &seed.iam_bucket,
+        seed.iam_key,
+        &seed.iam_body,
+        &format!("{context}: IAM GET historical object on node 3"),
+    )
+    .await?;
+
+    let post_key = format!("after-upgrade-{context}.txt");
+    let post_body = format!("{context}: written with the same IAM AK/SK after upgrade").into_bytes();
+    wait_for_put(
+        &iam_on_peer,
+        &seed.iam_bucket,
+        &post_key,
+        post_body.clone(),
+        &format!("{context}: IAM PUT after upgrade"),
+    )
+    .await?;
+    assert_object_bytes(&iam_on_upgraded, &seed.iam_bucket, &post_key, &post_body).await?;
+
+    let bad = dist.client_with_credentials(1, &seed.iam_user, WRONG_SECRET)?;
+    match bad.get_object().bucket(&seed.iam_bucket).key(seed.iam_key).send().await {
+        Ok(_) => return Err(format!("{context}: wrong secret must not read the IAM object").into()),
+        Err(error) => {
+            let code = error.as_service_error().and_then(ProvideErrorMetadata::code);
+            let rejected = code == Some("SignatureDoesNotMatch")
+                || code == Some("InvalidAccessKeyId")
+                || code == Some("AccessDenied")
+                || code == Some("InvalidArgument")
+                || error.raw_response().is_some_and(|response| response.status().as_u16() == 403);
+            if !rejected {
+                return Err(format!("{context}: wrong secret failed with unexpected error {error:?}").into());
+            }
+        }
+    }
+
+    let post_root_key = format!("root-after-{context}.bin");
+    let post_root_body = format!("{context}: root write after upgrade").into_bytes();
+    put_object(&root_a, &seed.history_bucket, &post_root_key, post_root_body.clone()).await?;
+    assert_object_bytes(&root_b, &seed.history_bucket, &post_root_key, &post_root_body).await?;
+    Ok(())
+}
+
+#[tokio::test]
+async fn four_node_direct_upgrade_preserves_history_and_iam_credentials() -> TestResult {
+    init_logging();
+    let previous = source_binary()?;
+    let mut dist = DistCluster::new_stopped(DistLayout::FourNodeFourDisk).await?;
+    capture_upgrade_logs(&mut dist, "direct-upgrade")?;
+    dist.start_from_binary(&previous).await?;
+
+    let seed = seed_history_and_iam(&dist).await?;
+    dist.restart_with_current_binary().await?;
+    assert_history_and_iam(&dist, &seed, "direct").await?;
+    Ok(())
+}
+
+#[tokio::test]
+async fn four_node_rolling_upgrade_preserves_history_and_iam_credentials() -> TestResult {
+    init_logging();
+    let previous = source_binary()?;
+    let mut dist = DistCluster::new_stopped(DistLayout::FourNodeFourDisk).await?;
+    capture_upgrade_logs(&mut dist, "rolling-upgrade")?;
+    dist.start_from_binary(&previous).await?;
+
+    let seed = seed_history_and_iam(&dist).await?;
+
+    dist.replace_node_with_current_binary(0).await?;
+    assert_history_and_iam(&dist, &seed, "one-current-node").await?;
+
+    for node_idx in [1, 2] {
+        dist.replace_node_with_current_binary(node_idx).await?;
+    }
+    assert_history_and_iam(&dist, &seed, "one-previous-node").await?;
+
+    dist.replace_node_with_current_binary(3).await?;
+    assert_history_and_iam(&dist, &seed, "homogeneous-current").await?;
+    Ok(())
+}

crates/e2e_test/src/distributed/versioning_test.rs

@@ -0,0 +1,188 @@
+// 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/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, enable_versioning, get_object_bytes, put_object, unique_bucket};
+use crate::common::init_logging;
+use aws_sdk_s3::error::ProvideErrorMetadata;
+use aws_sdk_s3::types::{BucketVersioningStatus, VersioningConfiguration};
+
+#[tokio::test]
+async fn four_node_four_drive_versioning_put_list_get_delete_marker() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let bucket = unique_bucket("version");
+    dist.create_bucket(&bucket).await?;
+    let writer = dist.client(0)?;
+    let reader = dist.client(3)?;
+    enable_versioning(&writer, &bucket).await?;
+
+    let key = "versioned.txt";
+    let v1_id = writer
+        .put_object()
+        .bucket(&bucket)
+        .key(key)
+        .body(b"v1".to_vec().into())
+        .send()
+        .await?
+        .version_id()
+        .ok_or("v1 PUT omitted version ID")?
+        .to_string();
+    let v2_id = writer
+        .put_object()
+        .bucket(&bucket)
+        .key(key)
+        .body(b"v2".to_vec().into())
+        .send()
+        .await?
+        .version_id()
+        .ok_or("v2 PUT omitted version ID")?
+        .to_string();
+
+    let versions = reader.list_object_versions().bucket(&bucket).prefix(key).send().await?;
+    let matching_versions: Vec<_> = versions
+        .versions()
+        .iter()
+        .filter(|version| version.key() == Some(key))
+        .collect();
+    assert_eq!(matching_versions.len(), 2, "fresh key must have exactly two versions: {versions:?}");
+    assert!(versions.delete_markers().is_empty(), "fresh key unexpectedly has a delete marker");
+    assert!(
+        matching_versions
+            .iter()
+            .any(|version| version.version_id() == Some(v1_id.as_str()) && version.is_latest() != Some(true)),
+        "v1 was not the historical version: {versions:?}"
+    );
+    assert!(
+        matching_versions
+            .iter()
+            .any(|version| version.version_id() == Some(v2_id.as_str()) && version.is_latest() == Some(true)),
+        "v2 was not the latest version: {versions:?}"
+    );
+
+    let latest = get_object_bytes(&reader, &bucket, key).await?;
+    assert_eq!(latest, b"v2");
+
+    let older = reader.get_object().bucket(&bucket).key(key).version_id(&v1_id).send().await?;
+    let older_body = older.body.collect().await?.into_bytes();
+    assert_eq!(older_body.as_ref(), b"v1");
+
+    let deleted = writer.delete_object().bucket(&bucket).key(key).send().await?;
+    assert_eq!(deleted.delete_marker(), Some(true));
+    let marker_id = deleted.version_id().ok_or("DeleteObject omitted delete-marker version ID")?;
+    let after_delete = reader.list_object_versions().bucket(&bucket).prefix(key).send().await?;
+    let matching_markers: Vec<_> = after_delete
+        .delete_markers()
+        .iter()
+        .filter(|marker| marker.key() == Some(key))
+        .collect();
+    assert_eq!(
+        matching_markers.len(),
+        1,
+        "delete marker missing or duplicated after current-version delete: {after_delete:?}"
+    );
+    assert!(
+        matching_markers[0].version_id() == Some(marker_id) && matching_markers[0].is_latest() == Some(true),
+        "DeleteObject response and ListObjectVersions disagree about the marker: {after_delete:?}"
+    );
+
+    let latest_after_delete = reader.get_object().bucket(&bucket).key(key).send().await;
+    match latest_after_delete {
+        Ok(_) => return Err("current version should be a delete marker".into()),
+        Err(error)
+            if error
+                .as_service_error()
+                .and_then(ProvideErrorMetadata::code)
+                .is_some_and(|code| code == "NoSuchKey" || code == "NotFound") => {}
+        Err(error) => return Err(error.into()),
+    }
+
+    let restored = reader.get_object().bucket(&bucket).key(key).version_id(&v1_id).send().await?;
+    let restored_body = restored.body.collect().await?.into_bytes();
+    assert_eq!(restored_body.as_ref(), b"v1");
+
+    writer
+        .delete_object()
+        .bucket(&bucket)
+        .key(key)
+        .version_id(marker_id)
+        .send()
+        .await?;
+    assert_eq!(get_object_bytes(&reader, &bucket, key).await?, b"v2");
+    Ok(())
+}
+
+#[tokio::test]
+async fn four_node_versioning_suspension_keeps_one_null_version_and_history() -> TestResult {
+    init_logging();
+    let dist = DistCluster::start(DistLayout::FourByFour).await?;
+    let bucket = unique_bucket("suspend");
+    dist.create_bucket(&bucket).await?;
+    let writer = dist.client(0)?;
+    let reader = dist.client(3)?;
+    enable_versioning(&writer, &bucket).await?;
+
+    let key = "suspended.txt";
+    let original = writer
+        .put_object()
+        .bucket(&bucket)
+        .key(key)
+        .body(b"enabled-history".to_vec().into())
+        .send()
+        .await?
+        .version_id()
+        .ok_or("enabled PUT omitted version ID")?
+        .to_string();
+    writer
+        .put_bucket_versioning()
+        .bucket(&bucket)
+        .versioning_configuration(
+            VersioningConfiguration::builder()
+                .status(BucketVersioningStatus::Suspended)
+                .build(),
+        )
+        .send()
+        .await?;
+
+    put_object(&writer, &bucket, key, b"null-one".to_vec()).await?;
+    put_object(&writer, &bucket, key, b"null-two".to_vec()).await?;
+    assert_eq!(get_object_bytes(&reader, &bucket, key).await?, b"null-two");
+
+    let versions = reader.list_object_versions().bucket(&bucket).prefix(key).send().await?;
+    let matching: Vec<_> = versions
+        .versions()
+        .iter()
+        .filter(|version| version.key() == Some(key))
+        .collect();
+    assert!(matching.iter().any(|version| version.version_id() == Some(original.as_str())));
+    let null_version_count = matching
+        .iter()
+        .filter(|version| {
+            matches!(
+                version.version_id(),
+                None | Some("") | Some("null") | Some("00000000-0000-0000-0000-000000000000")
+            )
+        })
+        .count();
+    assert_eq!(null_version_count, 1, "suspended overwrites must keep one null version: {versions:?}");
+
+    let historical = reader
+        .get_object()
+        .bucket(&bucket)
+        .key(key)
+        .version_id(&original)
+        .send()
+        .await?;
+    assert_eq!(historical.body.collect().await?.into_bytes().as_ref(), b"enabled-history");
+    Ok(())
+}

crates/e2e_test/src/fake_s3_target/mod.rs

@@ -451,10 +451,40 @@ impl JournaledHeaders {
 struct ControlState {
     scripts: HashMap<Operation, VecDeque<FaultAction>>,
     keyed_scripts: HashMap<(Operation, String), VecDeque<FaultAction>>,
+    held_get: Option<HeldGetObject>,
     requests: VecDeque<RequestRecord>,
     next_sequence: u64,
 }
 
+#[derive(Clone)]
+struct HeldGetObject {
+    bucket: String,
+    key: String,
+    entered: watch::Sender<usize>,
+    released: watch::Receiver<bool>,
+}
+
+/// Holds every GET of one object, including retries, until this guard is dropped.
+#[must_use = "dropping the guard releases the held GET requests"]
+pub struct GetObjectGate {
+    control: Arc<Mutex<ControlState>>,
+    entered: watch::Receiver<usize>,
+    released: watch::Sender<bool>,
+}
+
+impl GetObjectGate {
+    pub async fn wait_until_entered(&mut self) -> Result<(), watch::error::RecvError> {
+        self.entered.wait_for(|count| *count > 0).await.map(|_| ())
+    }
+}
+
+impl Drop for GetObjectGate {
+    fn drop(&mut self) {
+        lock(&self.control).held_get = None;
+        self.released.send_replace(true);
+    }
+}
+
 #[derive(Default)]
 struct StoreState {
     assign_own_version_ids: bool,
@@ -936,6 +966,30 @@ impl FakeS3Target {
             .extend(std::iter::repeat_n(action, times));
     }
 
+    /// Hold one exact bucket/key before any GET response can reach the client.
+    /// The fixture supports one live gate; request and connection deadlines still apply.
+    pub fn hold_get_object(&self, bucket: &str, key: &str) -> GetObjectGate {
+        assert!(
+            bucket.len() <= MAX_RETAINED_IDENTIFIER_BYTES && key.len() <= MAX_RETAINED_IDENTIFIER_BYTES,
+            "held GET identifiers exceed the fixture limit"
+        );
+        let mut state = lock(&self.control);
+        assert!(state.held_get.is_none(), "fake target already holds a GET gate");
+        let (entered, entered_rx) = watch::channel(0);
+        let (released, released_rx) = watch::channel(false);
+        state.held_get = Some(HeldGetObject {
+            bucket: bucket.to_string(),
+            key: key.to_string(),
+            entered,
+            released: released_rx,
+        });
+        GetObjectGate {
+            control: Arc::clone(&self.control),
+            entered: entered_rx,
+            released,
+        }
+    }
+
     pub fn clear_faults(&self) {
         let mut state = lock(&self.control);
         state.scripts.clear();
@@ -2243,6 +2297,17 @@ impl S3 for FakeBackend {
         let fault = request_fault(&req);
         apply_non_body_fault(fault.as_ref(), &self.control).await?;
         let input = req.input;
+        let held_get = lock(&self.control)
+            .held_get
+            .as_ref()
+            .filter(|held| held.bucket == input.bucket && held.key == input.key)
+            .cloned();
+        if let Some(mut held) = held_get {
+            held.entered.send_modify(|count| *count += 1);
+            // Keep the gate installed when a request is cancelled or times out:
+            // a retry must cross the same boundary before returning any bytes.
+            let _ = held.released.wait_for(|released| *released).await;
+        }
         let (version, versioned) = {
             let state = lock(&self.store);
             (
@@ -2894,6 +2959,81 @@ mod tests {
         aws_sdk_s3::primitives::DateTime::from_secs(4_102_444_800)
     }
 
+    #[tokio::test]
+    async fn get_object_gate_holds_retries_and_releases_on_drop() -> Result<(), BoxError> {
+        let target = FakeS3Target::start().await?;
+        let bucket = "gated-target";
+        target.create_bucket(bucket);
+        for key in ["held", "unrelated"] {
+            target.put_seed_object(bucket, key, Bytes::from_static(b"payload"), &SeedMetadata::default());
+        }
+        {
+            let gate = target.hold_get_object(bucket, "held");
+            let request = || S3Request {
+                input: GetObjectInput {
+                    bucket: bucket.to_string(),
+                    key: "held".to_string(),
+                    ..Default::default()
+                },
+                method: Method::GET,
+                uri: Uri::from_static("/gated-target/held"),
+                headers: HeaderMap::new(),
+                extensions: http::Extensions::new(),
+                credentials: None,
+                region: None,
+                service: None,
+                trailing_headers: None,
+            };
+            // Without a fault, only the gate can suspend this backend method.
+            let mut first = target.backend.get_object(request());
+            assert!(futures::poll!(first.as_mut()).is_pending(), "the first GET must wait at the gate");
+            drop(first);
+            let mut retry = target.backend.get_object(request());
+            assert!(futures::poll!(retry.as_mut()).is_pending(), "a cancelled GET must not consume the gate");
+            drop(gate);
+            let std::task::Poll::Ready(response) = futures::poll!(retry.as_mut()) else {
+                panic!("dropping the gate must release the waiting GET");
+            };
+            let mut body = response?.output.body.expect("released GET body");
+            assert_eq!(body.next().await.transpose()?, Some(Bytes::from_static(b"payload")));
+            assert!(body.next().await.is_none(), "released GET body must be complete");
+        }
+        let client = client(&target);
+        let mut gate = target.hold_get_object(bucket, "held");
+        let mut requests = tokio::task::JoinSet::new();
+        let first = client.clone();
+        requests.spawn(async move { get_bytes(&first, bucket, "held", None).await });
+        timeout(Duration::from_secs(2), gate.wait_until_entered()).await??;
+        requests.abort_all();
+        assert!(
+            requests
+                .join_next()
+                .await
+                .expect("first GET task")
+                .expect_err("cancel the first GET attempt")
+                .is_cancelled()
+        );
+
+        let retry = client.clone();
+        requests.spawn(async move { get_bytes(&retry, bucket, "held", None).await });
+        timeout(Duration::from_secs(2), gate.entered.wait_for(|count| *count == 2)).await??;
+        assert_eq!(
+            timeout(Duration::from_secs(2), get_bytes(&client, bucket, "unrelated", None)).await??,
+            Bytes::from_static(b"payload")
+        );
+        assert!(requests.try_join_next().is_none(), "the retry must remain behind the gate");
+        drop(gate);
+        assert_eq!(
+            timeout(Duration::from_secs(2), requests.join_next())
+                .await?
+                .expect("retried GET task")??,
+            Bytes::from_static(b"payload")
+        );
+        assert_eq!(get_bytes(&client, bucket, "held", None).await?, Bytes::from_static(b"payload"));
+        assert_eq!(target.count_requests(Operation::GetObject, "held"), 3);
+        Ok(())
+    }
+
     #[tokio::test]
     async fn object_lock_target_requires_a_checksum_on_locked_puts() -> Result<(), BoxError> {
         use aws_sdk_s3::error::ProvideErrorMetadata;

crates/e2e_test/src/heal_erasure_disk_rebuild_test.rs

@@ -16,24 +16,99 @@
 
 #[cfg(test)]
 mod tests {
-    use crate::chaos::{VersionShardCensus, census_object_version_on_disk, signed_admin_post};
+    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,
     };
     use crate::storage_api::RUSTFS_META_BUCKET;
     use aws_sdk_s3::primitives::ByteStream;
     use http::Method;
+    use sha2::{Digest, Sha256};
     use std::collections::HashSet;
     use std::error::Error;
+    use std::io::{Read, Write};
     use std::net::SocketAddr;
     use std::path::{Path, PathBuf};
     use std::process::Command;
     use tokio::net::TcpStream;
     use tokio::time::{Duration, Instant, sleep, timeout};
     use tracing::info;
+    #[cfg(target_os = "linux")]
+    use tracing::warn;
 
     const POOL_METADATA_OBJECT: &str = "pool.bin";
 
+    #[derive(serde::Deserialize)]
+    struct EvidenceBuild {
+        sha256: String,
+    }
+
+    #[derive(serde::Deserialize)]
+    struct RestartEvidenceRun {
+        schema: u32,
+        run_id: String,
+        source_revision: String,
+        test_build: serde_json::Value,
+        binary: EvidenceBuild,
+        test_binary: EvidenceBuild,
+    }
+
+    fn file_sha256(path: &Path) -> Result<String, Box<dyn Error + Send + Sync>> {
+        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 restart_evidence_run(binary: &Path) -> Result<Option<(PathBuf, RestartEvidenceRun)>, Box<dyn Error + Send + Sync>> {
+        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: RestartEvidenceRun = serde_json::from_slice(&std::fs::read(receipt)?)?;
+        if run.schema != 1 || run.run_id.len() != 32 || run.source_revision.len() != 40 {
+            return Err("invalid scanner/heal execution identity".into());
+        }
+        let built = compiled_test_identity();
+        for key in ["source_revision", "dirty", "lock_blob", "features"] {
+            assert_eq!(built[key], run.test_build[key], "compiled test identity differs for {key}");
+        }
+        assert_eq!(file_sha256(binary)?, run.binary.sha256, "server binary must match the run receipt");
+        assert_eq!(
+            file_sha256(&std::env::current_exe()?)?,
+            run.test_binary.sha256,
+            "test executable must match the run receipt"
+        );
+        if directory.join("background-target-restart.json").exists() {
+            return Err("scanner/heal oracle already exists; create a new execution receipt".into());
+        }
+        Ok(Some((directory, run)))
+    }
+
+    fn compiled_test_identity() -> serde_json::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"),
+        })
+    }
+
     struct TcpPortBlackhole {
         port: u16,
         comment: String,
@@ -42,6 +117,49 @@ mod tests {
     }
 
     impl TcpPortBlackhole {
+        /// Environment flag that turns an unusable fault-injection host into a
+        /// hard failure instead of a logged skip. Lanes that provision
+        /// `CAP_NET_ADMIN` set it so a broken runner cannot pass silently.
+        #[cfg(target_os = "linux")]
+        const REQUIRE_ENV: &str = "RUSTFS_E2E_REQUIRE_NET_FAULT_INJECTION";
+
+        /// Probe whether this host can manipulate the OUTPUT chain at all.
+        ///
+        /// Returns `Ok(Some(reason))` when `iptables` is missing or lacks
+        /// `CAP_NET_ADMIN` (the nf_tables backend reports "Permission denied"
+        /// even under `sudo` inside an unprivileged container) and the lane did
+        /// not demand fault injection; returns an error when the lane demands
+        /// it; returns `Ok(None)` when the blackhole can be installed.
+        #[cfg(target_os = "linux")]
+        fn unavailable_reason() -> Result<Option<String>, Box<dyn Error + Send + Sync>> {
+            let id = Command::new("id").arg("-u").output()?;
+            if !id.status.success() {
+                return Err(format!("failed to determine the test process uid: {}", String::from_utf8_lossy(&id.stderr)).into());
+            }
+            let use_sudo = String::from_utf8_lossy(&id.stdout).trim() != "0";
+            let mut command = if use_sudo {
+                let mut command = Command::new("sudo");
+                command.args(["-n", "iptables"]);
+                command
+            } else {
+                Command::new("iptables")
+            };
+            let probe = command.args(["-w", "5", "-S", "OUTPUT"]).output();
+            let reason = match probe {
+                Ok(output) if output.status.success() => return Ok(None),
+                Ok(output) => format!(
+                    "iptables cannot read the OUTPUT chain (status {}): {}",
+                    output.status,
+                    String::from_utf8_lossy(&output.stderr).trim()
+                ),
+                Err(err) => format!("iptables is not runnable: {err}"),
+            };
+            if std::env::var_os(Self::REQUIRE_ENV).is_some() {
+                return Err(format!("{} is set but network fault injection is unavailable: {reason}", Self::REQUIRE_ENV).into());
+            }
+            Ok(Some(reason))
+        }
+
         fn install(address: &str) -> Result<Self, Box<dyn Error + Send + Sync>> {
             let address = address.parse::<SocketAddr>()?;
             if !address.ip().is_loopback() {
@@ -126,6 +244,27 @@ mod tests {
         }
     }
 
+    /// Remove a disk directory underneath a running server. Background writers
+    /// (scanner, usage cache, heal markers) can recreate entries between the
+    /// recursive listing and the final `rmdir`, which surfaces as
+    /// `DirectoryNotEmpty` on macOS; retry briefly so the wipe reflects the
+    /// operator action rather than a listing race.
+    fn wipe_directory_while_server_runs(disk: &Path) -> std::io::Result<()> {
+        let mut last_err = None;
+        for _ in 0..20 {
+            match std::fs::remove_dir_all(disk) {
+                Ok(()) => return Ok(()),
+                Err(err) if err.kind() == std::io::ErrorKind::NotFound => return Ok(()),
+                Err(err) if err.kind() == std::io::ErrorKind::DirectoryNotEmpty => {
+                    last_err = Some(err);
+                    std::thread::sleep(std::time::Duration::from_millis(100));
+                }
+                Err(err) => return Err(err),
+            }
+        }
+        Err(last_err.expect("retry loop only exits without success after recording an error"))
+    }
+
     fn has_file_under(path: &Path) -> bool {
         let Ok(entries) = std::fs::read_dir(path) else {
             return false;
@@ -195,8 +334,9 @@ mod tests {
         clients: &[aws_sdk_s3::Client],
         bucket: &str,
         expected_keys: &HashSet<String>,
-    ) -> Result<(), Box<dyn Error + Send + Sync>> {
+    ) -> Result<Vec<Vec<String>>, Box<dyn Error + Send + Sync>> {
         const PAGE_SIZE: i32 = 10;
+        let mut node_listings = Vec::with_capacity(clients.len());
         for (node_index, client) in clients.iter().enumerate() {
             let mut listed_keys = Vec::new();
             let mut continuation_token = None;
@@ -243,8 +383,10 @@ mod tests {
                 &listed_key_set, expected_keys,
                 "node {node_index} did not expose the complete recovered namespace"
             );
+            listed_keys.sort();
+            node_listings.push(listed_keys);
         }
-        Ok(())
+        Ok(node_listings)
     }
 
     fn heal_task_status_diagnostic(body: &str) -> String {
@@ -405,7 +547,7 @@ mod tests {
             );
         }
 
-        std::fs::remove_dir_all(&disk0).expect("disk0 wipe should succeed while server is running");
+        wipe_directory_while_server_runs(&disk0).expect("disk0 wipe should succeed while server is running");
         std::fs::create_dir_all(&disk0).expect("disk0 should be recreated empty while server is running");
         assert!(!has_file_under(&disk0), "disk0 must be empty immediately after runtime wipe");
 
@@ -792,6 +934,18 @@ mod tests {
     #[cfg(target_os = "linux")]
     #[tokio::test(flavor = "multi_thread")]
     async fn test_cluster_root_heal_recovers_after_target_endpoint_blackhole() -> Result<(), Box<dyn Error + Send + Sync>> {
+        if let Some(reason) = TcpPortBlackhole::unavailable_reason()? {
+            init_logging();
+            warn!(
+                event = "heal_interruption_skipped",
+                component = "e2e_test",
+                subsystem = "heal",
+                interruption_kind = "target_endpoint_blackhole",
+                reason,
+                "Skipping endpoint blackhole scenario: network fault injection is unavailable on this host"
+            );
+            return Ok(());
+        }
         timeout(
             Duration::from_secs(420),
             run_cluster_root_heal_interruption(InterruptionScenario::TargetEndpointBlackhole),
@@ -808,6 +962,13 @@ mod tests {
     }
 
     async fn run_cluster_root_heal_interruption(scenario: InterruptionScenario) -> Result<(), Box<dyn Error + Send + Sync>> {
+        let server_binary = rustfs_binary_path();
+        let evidence_run = if scenario == InterruptionScenario::BackgroundTargetRestart {
+            restart_evidence_run(&server_binary)?
+        } else {
+            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"),
@@ -855,7 +1016,7 @@ mod tests {
         for node_index in 0..cluster.nodes.len() {
             cluster.set_node_capture_log_path(node_index, format!("{log_dir}/node{node_index}.log"))?;
         }
-        cluster.start().await?;
+        cluster.start_with_binary(&server_binary).await?;
         let clients = cluster.create_all_clients()?;
 
         let bucket = "heal-restart-during-rebuild";
@@ -996,7 +1157,7 @@ mod tests {
             }
         }
 
-        cluster.start_node(1).await?;
+        cluster.start_node_from_binary(1, &server_binary).await?;
 
         let status_url = format!("{}/rustfs/admin/v3/background-heal/status", cluster.nodes[0].url);
         let recovery_deadline = Instant::now() + Duration::from_secs(60);
@@ -1274,7 +1435,7 @@ mod tests {
                     }
                 }
             }
-            cluster.start_node(interruption_node).await?;
+            cluster.start_node_from_binary(interruption_node, &server_binary).await?;
             if interruption_node == 0 {
                 let target = cluster.nodes[1]
                     .process
@@ -1373,7 +1534,7 @@ mod tests {
             .map(|manifest| manifest.key.clone())
             .collect::<HashSet<_>>();
         assert!(expected_keys.insert(outage_key.to_string()));
-        assert_all_nodes_list_exact_keys(&clients, bucket, &expected_keys).await?;
+        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 {
@@ -1381,11 +1542,31 @@ mod tests {
             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() {
+                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(),
+                    "expected_sha256": sha256_hex(&expected_body),
+                    "actual_sha256": sha256_hex(&actual),
+                    "expected_physical": expected.shard_census,
+                    "physical": census_object_version_on_disk(&replaced_disk, bucket, &expected.key, None)?,
+                }));
+            }
         }
         let response = target_client.get_object().bucket(bucket).key(outage_key).send().await?;
         let actual = response.body.collect().await?.into_bytes();
         let expected_outage_body = deterministic_object_body(object_size_bytes, outage_payload_seed);
         assert_eq!(actual.as_ref(), expected_outage_body.as_slice(), "object body changed for {outage_key}");
+        if evidence_run.is_some() {
+            evidence_objects.push(serde_json::json!({
+                "key": outage_key, "version_id": null,
+                "expected_bytes": expected_outage_body.len(), "actual_bytes": actual.len(),
+                "expected_sha256": sha256_hex(&expected_outage_body),
+                "actual_sha256": sha256_hex(&actual),
+                "expected_physical": null,
+                "physical": census_object_version_on_disk(&replaced_disk, bucket, outage_key, None)?,
+            }));
+        }
 
         let terminal_deadline = Instant::now() + Duration::from_secs(30);
         loop {
@@ -1432,6 +1613,31 @@ mod tests {
             return Err(format!("heal data rebuilt but task did not finish successfully: {task_status}").into());
         }
 
+        if let Some((directory, run)) = evidence_run {
+            let restarted_pid = cluster.nodes[1].process.as_ref().ok_or("restarted target is absent")?.id();
+            assert_ne!(target_pid, restarted_pid, "target must be a new process");
+            assert_eq!(file_sha256(&server_binary)?, run.binary.sha256, "server build changed during restart");
+            let evidence = serde_json::json!({
+                "schema": 1, "case": "background-target-restart", "evidence": "process-restart",
+                "run_id": run.run_id, "source_revision": run.source_revision,
+                "test_build": compiled_test_identity(),
+                "binary_sha256": run.binary.sha256, "test_binary_sha256": run.test_binary.sha256,
+                "topology": {"nodes": cluster.nodes.len(), "drives_per_node": cluster.nodes[0].data_dirs.len()},
+                "pid_before": target_pid, "pid_after": restarted_pid,
+                "objects": evidence_objects, "node_listings": 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(directory.join("background-target-restart.json"))?;
+            output.write_all(&data)?;
+            output.sync_all()?;
+        }
+
         Ok(())
     }
 

crates/e2e_test/src/lib.rs

@@ -378,6 +378,11 @@ mod bucket_stats_regression_test;
 #[cfg(test)]
 mod distributed_startup_regression_test;
 
+// 4-node / 4-disk distributed Actions suite (S3, lock, versioning, replication,
+// quota, observability, expand/decommission/rebalance, site replication, chaos).
+#[cfg(test)]
+mod distributed;
+
 // P1 regression: tier/ILM transition (rustfs#5218, #5130, #5011, #4826, #5024)
 #[cfg(test)]
 mod tier_transition_regression_test;

crates/e2e_test/src/on_demand_migration/get_basic_test.rs

@@ -21,7 +21,7 @@
 use super::common::{BoxError, OdmSourceSpec, OdmTestEnv, SeedObject};
 use crate::fake_s3_target::{BucketMode, Operation};
 use aws_sdk_s3::error::ProvideErrorMetadata;
-use aws_sdk_s3::types::{BucketVersioningStatus, VersioningConfiguration};
+use aws_sdk_s3::types::{BucketVersioningStatus, ObjectAttributes, VersioningConfiguration};
 use bytes::Bytes;
 use std::time::Duration;
 
@@ -103,11 +103,19 @@ async fn get_miss_pulls_inline_and_serves_locally_afterwards() -> TestResult {
 
 #[tokio::test]
 async fn get_large_object_streams_through_and_backfills_in_background() -> TestResult {
+    const PART_SIZE: usize = 5 * 1024 * 1024;
     let bucket = "odm-get-large";
-    let env = configured_env(bucket, |spec| spec.policy.inline_max_bytes = 4096).await?;
+    let env = configured_env(bucket, |spec| {
+        spec.policy.inline_max_bytes = 4096;
+        spec.policy.multipart_part_size_bytes = u64::try_from(PART_SIZE).expect("part size fits in u64");
+    })
+    .await?;
     let key = "large/archive.bin";
-    let body = payload(512 * 1024);
-    env.seed_source(SOURCE_BUCKET, &[SeedObject::new(key, body.clone())]);
+    let body = payload(PART_SIZE + 4096);
+    let etag = env
+        .seed_source(SOURCE_BUCKET, &[SeedObject::new(key, body.clone())])
+        .remove(0);
+    assert_eq!(etag.len(), 32, "the source fixture has a plain MD5 ETag");
 
     let response = env.raw_get(bucket, key).await?;
     assert_eq!(response.status, 200, "{}", String::from_utf8_lossy(&response.body));
@@ -125,6 +133,68 @@ async fn get_large_object_streams_through_and_backfills_in_background() -> TestR
         vec![None, None],
         "one passthrough GET plus one background pull, both unranged"
     );
+
+    let source_requests = env.source.requests().len();
+    let second_part = env.client.get_object().bucket(bucket).key(key).part_number(2).send().await?;
+    assert_eq!(second_part.content_length(), Some(4096), "the completed second part is the tail");
+    assert_eq!(
+        second_part.content_range(),
+        Some(format!("bytes {PART_SIZE}-{}/{}", body.len() - 1, body.len()).as_str()),
+        "partNumber reads the stored multipart boundary"
+    );
+    assert_eq!(
+        second_part.body.collect().await?.into_bytes(),
+        body.slice(PART_SIZE..),
+        "the local second part contains the exact source tail"
+    );
+    let third_part = env
+        .client
+        .get_object()
+        .bucket(bucket)
+        .key(key)
+        .part_number(3)
+        .send()
+        .await
+        .expect_err("the completed object has exactly two parts");
+    assert_eq!(third_part.code(), Some("InvalidPart"));
+
+    let mut part_marker = None;
+    for (part_number, part_size) in [(1, PART_SIZE), (2, 4096)] {
+        let attributes = env
+            .client
+            .get_object_attributes()
+            .bucket(bucket)
+            .key(key)
+            .object_attributes(ObjectAttributes::ObjectParts)
+            .object_attributes(ObjectAttributes::Etag)
+            .max_parts(1)
+            .set_part_number_marker(part_marker.clone())
+            .send()
+            .await?;
+        assert_eq!(
+            attributes.e_tag().map(|value| value.trim_matches('"')),
+            Some(etag.as_str()),
+            "multipart write-back preserves the source MD5 ETag"
+        );
+        let parts = attributes
+            .object_parts()
+            .expect("RustFS must expose the stored multipart layout");
+        assert_eq!(parts.total_parts_count(), Some(2));
+        assert_eq!(parts.max_parts(), Some(1));
+        assert_eq!(parts.is_truncated(), Some(part_number == 1));
+        assert_eq!(parts.parts().len(), 1, "RustFS returns one stored part per requested page");
+        assert_eq!(parts.parts()[0].part_number(), Some(part_number));
+        assert_eq!(parts.parts()[0].size(), Some(i64::try_from(part_size).expect("part size fits in i64")));
+        part_marker = parts.next_part_number_marker().map(str::to_owned);
+        if part_number == 1 {
+            assert_eq!(part_marker.as_deref(), Some("1"), "the next request continues after the first part");
+        }
+    }
+    assert_eq!(
+        env.source.requests().len(),
+        source_requests,
+        "local part reads must not consult the source"
+    );
     Ok(())
 }
 

crates/e2e_test/src/on_demand_migration/interaction_test.rs

@@ -22,17 +22,21 @@
 //! local object and what the source was asked for.
 
 use super::common::{
-    AdminResponse, BoxError, OdmEnvOptions, OdmSourceSpec, OdmTestEnv, SeedObject, start_configured_env,
-    start_configured_env_with,
+    ALLOW_LOOPBACK_SOURCE_ENV, AdminResponse, BackfillOp, BackfillRequest, BoxError, ODM_MODULE_SWITCH_ENV, ODM_SERVER_ENV,
+    OdmEnvOptions, OdmSourceSpec, OdmTestEnv, SeedObject, start_configured_env, start_configured_env_with, start_source_rustfs,
 };
 use crate::common::{RustFSTestEnvironment, replication_fast_env, signed_request};
 use crate::fake_s3_target::{BucketMode, FAKE_ACCESS_KEY, FAKE_SECRET_KEY, FakeS3Target, Operation};
 use crate::object_lock::common::put_object_lock_configuration;
+use crate::replication_extension_test::{
+    ReplicationTargetOptions, enable_bucket_versioning, set_replication_target_with_options,
+};
 use aws_sdk_s3::error::ProvideErrorMetadata;
 use aws_sdk_s3::types::{
     BucketVersioningStatus, Event, FilterRule, FilterRuleName, NotificationConfiguration, NotificationConfigurationFilter,
-    ObjectLockRetentionMode, QueueConfiguration, S3KeyFilter, ServerSideEncryption, ServerSideEncryptionByDefault,
-    ServerSideEncryptionConfiguration, ServerSideEncryptionRule, VersioningConfiguration,
+    ObjectAttributes, ObjectLockRetentionMode, QueueConfiguration, S3KeyFilter, ServerSideEncryption,
+    ServerSideEncryptionByDefault, ServerSideEncryptionConfiguration, ServerSideEncryptionRule, Tag, Tagging,
+    VersioningConfiguration,
 };
 use bytes::Bytes;
 use local_ip_address::local_ip;
@@ -580,6 +584,130 @@ async fn test_odm_pulled_object_replicates_and_target_as_source_is_rejected() ->
         "a bucket may not migrate from its own replication target: {}",
         rejected.body
     );
+    Box::pin(assert_odm_multipart_replicates_to_rustfs(&env, bucket)).await?;
+    Ok(())
+}
+
+async fn assert_odm_multipart_replicates_to_rustfs(env: &OdmTestEnv, bucket: &str) -> TestResult {
+    const PART_SIZE: usize = 5 * 1024 * 1024;
+    let replica = start_source_rustfs().await?;
+    let replica_bucket = "odm-real-replica";
+    replica.create_test_bucket(replica_bucket).await?;
+    enable_bucket_versioning(&replica, replica_bucket).await?;
+    let arn = set_replication_target_with_options(
+        &env.rustfs,
+        bucket,
+        ReplicationTargetOptions {
+            endpoint: &replica.address,
+            access_key: &replica.access_key,
+            secret_key: &replica.secret_key,
+            target_bucket: replica_bucket,
+            secure: false,
+            skip_tls_verify: false,
+            ca_cert_pem: None,
+        },
+    )
+    .await?;
+    put_bucket_replication(&env.rustfs, bucket, &arn).await?;
+    let mut spec = env.fake_source_spec(SOURCE_BUCKET);
+    // Below the 16 MiB inline default the pull is one tee'd PUT with a single
+    // part; force the passthrough + background multipart write-back instead.
+    spec.policy.inline_max_bytes = 4096;
+    spec.policy.multipart_part_size_bytes = PART_SIZE as u64;
+    spec.policy.preserve_etag = true;
+    env.configure_and_wait(bucket, &spec).await?;
+
+    let key = "replicated/preserved-md5-multipart.bin";
+    let body = payload(PART_SIZE + 4096);
+    let source_put = env
+        .source_client()
+        .put_object()
+        .bucket(SOURCE_BUCKET)
+        .key(key)
+        .body(aws_sdk_s3::primitives::ByteStream::from(body.clone()))
+        .send()
+        .await?;
+    let etag = source_put.e_tag().ok_or("source PUT omitted its ETag")?.trim_matches('"');
+    assert_eq!(etag.len(), 32, "the source must retain a single-PUT MD5 ETag");
+    assert!(etag.bytes().all(|byte| byte.is_ascii_hexdigit()));
+    let pulled = env.raw_get(bucket, key).await?;
+    assert_eq!(pulled.status, 200, "{}", String::from_utf8_lossy(&pulled.body));
+    assert_eq!(pulled.body, body);
+    assert!(env.wait_local_listed(bucket, key, SETTLE).await?, "the multipart pull must persist");
+
+    let deadline = Instant::now() + SETTLE;
+    let source_head = loop {
+        let head = env.client.head_object().bucket(bucket).key(key).send().await?;
+        match head.replication_status().map(|status| status.as_str()) {
+            Some("COMPLETED") => break head,
+            Some("FAILED") => return Err("the ODM multipart copy failed replication to RustFS".into()),
+            _ => {
+                assert!(Instant::now() < deadline, "the ODM multipart copy never completed replication to RustFS");
+                tokio::time::sleep(Duration::from_millis(200)).await;
+            }
+        }
+    };
+    let version = source_head
+        .version_id()
+        .ok_or("the versioned ODM copy omitted its version id")?;
+    assert_ne!(version, "null");
+    let replica_client = replica.create_s3_client();
+    for (client, object_bucket) in [(&env.client, bucket), (&replica_client, replica_bucket)] {
+        let attributes = client
+            .get_object_attributes()
+            .bucket(object_bucket)
+            .key(key)
+            .version_id(version)
+            .object_attributes(ObjectAttributes::Etag)
+            .object_attributes(ObjectAttributes::ObjectParts)
+            .send()
+            .await?;
+        assert_eq!(attributes.e_tag().map(|value| value.trim_matches('"')), Some(etag));
+        let parts = attributes
+            .object_parts()
+            .ok_or("the local copy and replica must both expose two parts")?;
+        assert_eq!(parts.total_parts_count(), Some(2));
+        assert_eq!(
+            parts
+                .parts()
+                .iter()
+                .map(|part| (part.part_number(), part.size()))
+                .collect::<Vec<_>>(),
+            [(Some(1), Some(PART_SIZE as i64)), (Some(2), Some(4096))]
+        );
+    }
+    // REPLICA status surfaces on HEAD, like the other inbound-replica checks.
+    let replica_head = replica_client
+        .head_object()
+        .bucket(replica_bucket)
+        .key(key)
+        .version_id(version)
+        .send()
+        .await?;
+    assert_eq!(replica_head.replication_status().map(|status| status.as_str()), Some("REPLICA"));
+    let replica_get = replica_client
+        .get_object()
+        .bucket(replica_bucket)
+        .key(key)
+        .version_id(version)
+        .send()
+        .await?;
+    assert_eq!(replica_get.version_id(), Some(version));
+    assert_eq!(replica_get.body.collect().await?.into_bytes(), body);
+    let boundary = replica_client
+        .get_object()
+        .bucket(replica_bucket)
+        .key(key)
+        .version_id(version)
+        .range(format!("bytes={}-{}", PART_SIZE - 32, PART_SIZE + 31))
+        .send()
+        .await?;
+    assert_eq!(boundary.body.collect().await?.into_bytes(), body.slice(PART_SIZE - 32..PART_SIZE + 32));
+    assert_eq!(
+        env.source.count_requests(Operation::GetObject, key),
+        2,
+        "one passthrough GET plus one background pull; replication and local reads must not fetch the migration source again"
+    );
     Ok(())
 }
 
@@ -733,6 +861,242 @@ async fn test_odm_disable_keeps_pulled_objects_and_stops_source_traffic() -> Tes
     Ok(())
 }
 
+/// The process switch preserves configured buckets and unfinished jobs while
+/// restoring local-only S3 behavior, including after an ordinary metadata write.
+#[tokio::test]
+async fn test_odm_global_disable_preserves_data_config_and_backfill_across_restarts() -> TestResult {
+    let bucket = "odm-global-disable";
+    let mut env = start_configured_env(bucket, SOURCE_BUCKET, |spec| spec.policy.list_through = true).await?;
+    let pulled_key = "migrated/pulled.bin";
+    let remote_key = "remote/untouched.bin";
+    let pending_key = "backfill/pending.bin";
+    let local_key = "local/kept.bin";
+    let source_body = Bytes::from_static(b"source payload");
+    let local_body = Bytes::from_static(b"client payload");
+    env.seed_source(
+        SOURCE_BUCKET,
+        &[
+            SeedObject::new(pulled_key, source_body.clone()),
+            SeedObject::new(remote_key, source_body.clone()),
+            SeedObject::new(pending_key, source_body.clone()),
+        ],
+    );
+    env.client
+        .put_object()
+        .bucket(bucket)
+        .key(local_key)
+        .body(local_body.clone().into())
+        .send()
+        .await?;
+    let pulled = env.raw_get(bucket, pulled_key).await?;
+    assert_eq!(pulled.status, 200);
+    assert_eq!(pulled.header(ODM_RESPONSE_HEADER), Some("source"));
+    assert_eq!(pulled.body, source_body);
+    let stored = env.raw_get(bucket, pulled_key).await?;
+    assert_eq!(stored.status, 200);
+    assert_eq!(stored.header(ODM_RESPONSE_HEADER), None, "the inline pull has committed locally");
+    assert_eq!(stored.body, source_body);
+    let config = env.get_config(bucket).await?;
+    assert_eq!(config.status, 200, "{}", config.body);
+    let config = config.json()?;
+
+    // Hold every attempt until the process has exited, so retries cannot commit
+    // the only backfill object before the crash. The start checkpoint exists.
+    let mut pending_get = env.source.hold_get_object(SOURCE_BUCKET, pending_key);
+    let started = env
+        .start_backfill(
+            bucket,
+            BackfillRequest {
+                prefix: Some("backfill/".to_string()),
+                ..BackfillRequest::default()
+            },
+        )
+        .await?;
+    assert_eq!(started.status, 200, "{}", started.body);
+    let job_id = started.json()?["job"]["job_id"].as_str().ok_or("missing job ID")?.to_string();
+    tokio::time::timeout(Duration::from_secs(10), pending_get.wait_until_entered())
+        .await
+        .expect("backfill never reached the held source GET")?;
+    let process = env.rustfs.process.as_mut().ok_or("missing RustFS process before crash")?;
+    assert!(process.try_wait()?.is_none(), "RustFS exited before the controlled crash");
+    process.kill()?;
+    let stopped = process.wait()?;
+    assert!(!stopped.success(), "the interrupted process must exit after being killed");
+    drop(env.rustfs.process.take());
+    drop(pending_get);
+    env.source.take_requests();
+    env.rustfs
+        .restart_server_preserving_data(vec![], &[(ODM_MODULE_SWITCH_ENV, "false"), (ALLOW_LOOPBACK_SOURCE_ENV, "true")])
+        .await?;
+
+    let off_config = env.get_config(bucket).await?;
+    assert_eq!(off_config.status, 200, "{}", off_config.body);
+    assert_eq!(off_config.json()?, config, "the saved configuration and timestamp survive disabling");
+    let status = env.status_json(bucket).await?;
+    assert_eq!(status["configured"], true, "{status}");
+    assert_eq!(status["enabled"], true, "the bucket remains configured as enabled: {status}");
+    assert_eq!(status["module_enabled"], false, "{status}");
+    assert_eq!(status["counters"], Value::Null, "no bucket runtime is installed: {status}");
+    let checkpoint = env.backfill_job(bucket).await?.ok_or("disabled module lost the checkpoint")?;
+    assert_eq!(checkpoint["job_id"], job_id);
+    assert_eq!(checkpoint["state"], "running", "the interrupted job is retained: {checkpoint}");
+
+    for (key, body) in [(local_key, &local_body), (pulled_key, &source_body)] {
+        let get = env.raw_get(bucket, key).await?;
+        assert_eq!(get.status, 200);
+        assert_eq!(&get.body, body);
+        assert_eq!(get.header(ODM_RESPONSE_HEADER), None);
+        let head = env.client.head_object().bucket(bucket).key(key).send().await?;
+        assert_eq!(head.content_length(), Some(i64::try_from(body.len())?));
+    }
+    for key in [remote_key, pending_key] {
+        let get = env.raw_get(bucket, key).await?;
+        assert_eq!(get.status, 404, "disabled source GET {key}: {}", String::from_utf8_lossy(&get.body));
+        let head = env.client.head_object().bucket(bucket).key(key).send().await;
+        let err = head.expect_err("a source-only object must remain absent locally");
+        assert_eq!(err.raw_response().map(|response| response.status().as_u16()), Some(404));
+    }
+
+    let replacement = Bytes::from_static(b"written while the module is off");
+    for key in [local_key, "local/deleted.bin"] {
+        env.client
+            .put_object()
+            .bucket(bucket)
+            .key(key)
+            .body(replacement.clone().into())
+            .send()
+            .await?;
+    }
+    env.client
+        .delete_object()
+        .bucket(bucket)
+        .key("local/deleted.bin")
+        .send()
+        .await?;
+    assert_eq!(env.raw_get(bucket, "local/deleted.bin").await?.status, 404);
+    assert_eq!(env.raw_get(bucket, local_key).await?.body, replacement);
+
+    // Both wire protocols must finish their local pages even though the saved
+    // configuration still requests list-through.
+    for use_v2 in [false, true] {
+        let mut cursor = None;
+        let mut listed = Vec::new();
+        for page_number in 0..2 {
+            let (keys, truncated, next) = if use_v2 {
+                let page = env
+                    .client
+                    .list_objects_v2()
+                    .bucket(bucket)
+                    .max_keys(1)
+                    .set_continuation_token(cursor)
+                    .send()
+                    .await?;
+                (
+                    page.contents()
+                        .iter()
+                        .map(|object| object.key().expect("listed key").to_string())
+                        .collect::<Vec<_>>(),
+                    page.is_truncated(),
+                    page.next_continuation_token().map(str::to_string),
+                )
+            } else {
+                let page = env
+                    .client
+                    .list_objects()
+                    .bucket(bucket)
+                    .max_keys(1)
+                    .set_marker(cursor)
+                    .send()
+                    .await?;
+                // V1 may omit NextMarker without a delimiter; clients then
+                // continue from the last returned key.
+                let next = page.next_marker().or_else(|| {
+                    if page.is_truncated() == Some(true) {
+                        page.contents().last().and_then(|object| object.key())
+                    } else {
+                        None
+                    }
+                });
+                (
+                    page.contents()
+                        .iter()
+                        .map(|object| object.key().expect("listed key").to_string())
+                        .collect::<Vec<_>>(),
+                    page.is_truncated(),
+                    next.map(str::to_string),
+                )
+            };
+            assert_eq!(keys.len(), 1, "one local key per page, V2={use_v2}");
+            assert_eq!(truncated, Some(page_number == 0), "local pagination must terminate, V2={use_v2}");
+            if page_number == 0 {
+                assert!(next.as_ref().is_some_and(|value| !value.is_empty()), "missing local cursor, V2={use_v2}");
+            }
+            cursor = next;
+            listed.extend(keys);
+        }
+        assert_eq!(listed, [local_key, pulled_key], "source-only keys must stay absent, V2={use_v2}");
+    }
+
+    let spec = env.fake_source_spec(SOURCE_BUCKET);
+    for response in [
+        env.configure_source(bucket, &spec).await?,
+        env.validate_source(bucket, &spec).await?,
+        env.backfill(bucket, BackfillOp::Start(BackfillRequest::default())).await?,
+    ] {
+        assert_eq!(response.status, 400, "{}", response.body);
+        assert!(response.body.contains("OnDemandMigrationDisabled"), "{}", response.body);
+    }
+    let tagging = Tagging::builder()
+        .tag_set(Tag::builder().key("module").value("disabled").build()?)
+        .build()?;
+    env.client
+        .put_bucket_tagging()
+        .bucket(bucket)
+        .tagging(tagging.clone())
+        .send()
+        .await?;
+    assert_eq!(env.get_config(bucket).await?.json()?, config, "an unrelated metadata write preserves ODM");
+    assert_eq!(
+        env.backfill_job(bucket).await?,
+        Some(checkpoint),
+        "no recovery or checkpoint update while disabled"
+    );
+    assert!(
+        env.source.requests().is_empty(),
+        "disabled startup and all requests must leave the source untouched"
+    );
+
+    env.rustfs.restart_server_preserving_data(vec![], ODM_SERVER_ENV).await?;
+    env.wait_until_source_consulted(bucket).await?;
+    assert_eq!(
+        env.get_config(bucket).await?.json()?,
+        config,
+        "reenabling uses the persisted configuration"
+    );
+    let tags = env.client.get_bucket_tagging().bucket(bucket).send().await?;
+    assert_eq!(tags.tag_set(), tagging.tag_set(), "the ordinary metadata write also persists");
+    let resumed = env.raw_get(bucket, remote_key).await?;
+    assert_eq!(resumed.status, 200);
+    assert_eq!(resumed.header(ODM_RESPONSE_HEADER), Some("source"));
+    assert_eq!(resumed.body, source_body, "stored credentials still authenticate without reconfiguration");
+    let completed = env
+        .wait_for_backfill(bucket, SETTLE, |job| job["state"] == "completed")
+        .await?;
+    assert_eq!(completed["job_id"], job_id, "the interrupted job resumes without a new start");
+    assert_eq!(completed["failed"], 0, "{completed}");
+    for (key, body) in [
+        (local_key, &replacement),
+        (pulled_key, &source_body),
+        (pending_key, &source_body),
+    ] {
+        let get = env.raw_get(bucket, key).await?;
+        assert_eq!(get.status, 200);
+        assert_eq!(&get.body, body);
+        assert_eq!(get.header(ODM_RESPONSE_HEADER), None, "{key} remains stored locally");
+    }
+    Ok(())
+}
+
 /// Case 19: the admin surface an operator sees — the configuration read back
 /// without its secret, and a status document whose counters match the source
 /// journal exactly.

crates/e2e_test/src/replication_extension_test.rs

@@ -368,23 +368,23 @@ impl Drop for SlowReplicationTargetGuard {
 // Mirrors madmin-go `ResyncTargetsInfo`/`ResyncTarget` json tags — the same
 // shape `mc replicate resync status` decodes.
 #[derive(Debug, Clone, serde::Deserialize)]
-struct ReplicationResetStatusResponse {
+pub(crate) struct ReplicationResetStatusResponse {
     #[serde(rename = "target", default)]
-    targets: Vec<ReplicationResetStatusTarget>,
+    pub(crate) targets: Vec<ReplicationResetStatusTarget>,
 }
 
 #[derive(Debug, Clone, serde::Deserialize)]
-struct ReplicationResetStatusTarget {
+pub(crate) struct ReplicationResetStatusTarget {
     #[serde(rename = "arn", default)]
-    arn: String,
+    pub(crate) arn: String,
     #[serde(rename = "resetid", default)]
-    reset_id: String,
+    pub(crate) reset_id: String,
     #[serde(rename = "resyncStatus", default)]
-    status: String,
+    pub(crate) status: String,
     #[serde(rename = "replicationCount", default)]
-    replicated_count: i64,
+    pub(crate) replicated_count: i64,
     #[serde(rename = "object", default)]
-    object: String,
+    pub(crate) object: String,
 }
 
 fn extract_xml_tag(xml: &str, tag: &str) -> Option<String> {
@@ -2294,7 +2294,7 @@ async fn site_replication_state_edit(
 /// return the target `(arn, reset_id)`, asserting the response carries the
 /// madmin `ResyncTargetsInfo` shape (`target[0].arn` / `target[0].resetid`)
 /// that `mc replicate resync start` decodes.
-async fn start_bucket_replication_reset(
+pub(crate) async fn start_bucket_replication_reset(
     env: &RustFSTestEnvironment,
     bucket: &str,
 ) -> Result<(String, String), Box<dyn Error + Send + Sync>> {
@@ -2314,7 +2314,7 @@ async fn start_bucket_replication_reset(
     Ok((arn, reset_id))
 }
 
-async fn get_replication_reset_status(
+pub(crate) async fn get_replication_reset_status(
     env: &RustFSTestEnvironment,
     bucket: &str,
     arn: &str,
@@ -3837,6 +3837,244 @@ async fn test_bucket_replication_converges_delete_marker_and_version_purge() ->
     Ok(())
 }
 
+/// Regression for rustfs/backlog#2340 (not Wasabi specific): a directory
+/// marker (`prefix/` with a body) in a versioned bucket is stored as the null
+/// version, like MinIO (`putOpts`: "for directory objects skip creating new
+/// versions"), and must still replicate to completion instead of staying
+/// `PENDING`.
+#[tokio::test]
+async fn test_bucket_replication_replicates_directory_marker_in_versioned_bucket() -> TestResult {
+    init_logging();
+
+    let mut source_env = RustFSTestEnvironment::new().await?;
+    let mut source_env_vars = replication_fast_env();
+    source_env_vars.extend_from_slice(LOOPBACK_REPLICATION_TARGET_ENV);
+    source_env.start_rustfs_server_with_env(vec![], &source_env_vars).await?;
+
+    let mut target_env = RustFSTestEnvironment::new().await?;
+    target_env.start_rustfs_server_without_cleanup(vec![]).await?;
+
+    let source_bucket = "replication-dir-marker-src";
+    let target_bucket = "replication-dir-marker-dst";
+    let source_client = source_env.create_s3_client();
+    let target_client = target_env.create_s3_client();
+
+    source_client.create_bucket().bucket(source_bucket).send().await?;
+    target_client.create_bucket().bucket(target_bucket).send().await?;
+    enable_bucket_versioning(&source_env, source_bucket).await?;
+    enable_bucket_versioning(&target_env, target_bucket).await?;
+    let target_arn = set_replication_target(&source_env, source_bucket, &target_env, target_bucket).await?;
+    put_bucket_replication(&source_env, source_bucket, &target_arn).await?;
+
+    let marker_key = "dir/trailing/";
+    let body = b"directory marker body";
+    let put = source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(marker_key)
+        .body(ByteStream::from_static(body))
+        .send()
+        .await?;
+    assert!(
+        put.version_id()
+            .is_none_or(|id| id == "null" || id == uuid::Uuid::nil().to_string()),
+        "a directory marker is the null version even in a versioned bucket: {:?}",
+        put.version_id()
+    );
+
+    wait_for_source_replication_status(&source_client, source_bucket, marker_key, "COMPLETED", false).await?;
+
+    let replica = target_client
+        .get_object()
+        .bucket(target_bucket)
+        .key(marker_key)
+        .send()
+        .await?;
+    assert_eq!(replica.body.collect().await?.into_bytes().as_ref(), body);
+    let listed = target_client
+        .list_object_versions()
+        .bucket(target_bucket)
+        .prefix(marker_key)
+        .send()
+        .await?;
+    let marker_versions: Vec<_> = listed.versions().iter().filter(|v| v.key() == Some(marker_key)).collect();
+    assert_eq!(marker_versions.len(), 1, "the marker must land exactly once: {marker_versions:?}");
+    assert_eq!(
+        marker_versions[0].version_id(),
+        Some("null"),
+        "the replica keeps the null version identity"
+    );
+
+    Ok(())
+}
+
+/// Regression for rustfs/backlog#2340 (not Wasabi specific): permanently
+/// deleting a version whose payload lives in a data dir must leave the source
+/// clean once the purge replicates. Managed-SSE objects are never inlined and a
+/// plain object above the inline threshold takes the same layout. The version
+/// retained with a pending purge used to lose its data dir, so the purge state
+/// could never be applied (`VersionNotFound` on every retry) and the bucket
+/// stayed `BucketNotEmpty` while `ListObjectVersions` was already empty.
+#[tokio::test]
+async fn test_bucket_replication_version_purge_of_non_inline_object_releases_source_bucket() -> TestResult {
+    init_logging();
+
+    let (source_env, target_env, source_bucket, target_bucket) = build_sse_replication_pair("purge-datadir", true, true).await?;
+    let target_arn = wait_for_remote_target_arn(&source_env, &source_bucket).await?;
+    put_bucket_replication_with_delete_statuses(&source_env, &source_bucket, &target_arn, "Enabled", Some("Enabled")).await?;
+    let source_client = source_env.create_s3_client();
+    let target_client = target_env.create_s3_client();
+
+    let sse_key = "sse-object.bin";
+    let large_key = "large-object.bin";
+    let sse_put = source_client
+        .put_object()
+        .bucket(&source_bucket)
+        .key(sse_key)
+        .body(ByteStream::from_static(b"encrypted source payload"))
+        .server_side_encryption(ServerSideEncryption::Aes256)
+        .send()
+        .await?;
+    let large_put = source_client
+        .put_object()
+        .bucket(&source_bucket)
+        .key(large_key)
+        .body(ByteStream::from(vec![0x5a; 2 * 1024 * 1024]))
+        .send()
+        .await?;
+    let purged = [
+        (sse_key, sse_put.version_id().ok_or("SSE PUT omitted version ID")?.to_string()),
+        (large_key, large_put.version_id().ok_or("large PUT omitted version ID")?.to_string()),
+    ];
+    assert_replication_converged(&source_client, &source_bucket, &target_client, &target_bucket).await?;
+
+    for (key, version_id) in &purged {
+        source_client
+            .delete_object()
+            .bucket(&source_bucket)
+            .key(*key)
+            .version_id(version_id)
+            .send()
+            .await?;
+    }
+    assert_replication_converged(&source_client, &source_bucket, &target_client, &target_bucket).await?;
+    let target_state = list_replication_state(&target_client, &target_bucket).await?;
+    assert!(target_state.is_empty(), "target retained an explicitly purged version: {target_state:?}");
+
+    // The purge state is applied on the source asynchronously after the target
+    // acknowledges the delete; only then does the retained version go away and
+    // the bucket become deletable. A listing that is empty while DeleteBucket
+    // keeps answering BucketNotEmpty is exactly the regression.
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(60);
+    loop {
+        let listing = source_client.list_object_versions().bucket(&source_bucket).send().await?;
+        let listed = listing.versions().len() + listing.delete_markers().len();
+        match source_client.delete_bucket().bucket(&source_bucket).send().await {
+            Ok(_) => break,
+            Err(err) if err.code() == Some("BucketNotEmpty") => {
+                if tokio::time::Instant::now() >= deadline {
+                    return Err(format!(
+                        "source bucket stayed BucketNotEmpty after the version purge replicated; \
+                         ListObjectVersions shows {listed} entries"
+                    )
+                    .into());
+                }
+                sleep(Duration::from_millis(500)).await;
+            }
+            Err(err) => return Err(err.into()),
+        }
+    }
+
+    Ok(())
+}
+
+/// Regression for rustfs/backlog#2340 (not Wasabi specific): a single-part
+/// object uploaded with `x-amz-checksum-*` must reach the target with the same
+/// checksum. The outbound options keyed the stored record by algorithm name,
+/// which the target client sent as `x-amz-meta-*` user metadata, so a replica
+/// never carried a checksum although the source HEAD returned one.
+#[tokio::test]
+async fn test_bucket_replication_forwards_single_part_object_checksums() -> TestResult {
+    init_logging();
+
+    let mut source_env = RustFSTestEnvironment::new().await?;
+    let mut source_env_vars = replication_fast_env();
+    source_env_vars.extend_from_slice(LOOPBACK_REPLICATION_TARGET_ENV);
+    source_env.start_rustfs_server_with_env(vec![], &source_env_vars).await?;
+
+    let mut target_env = RustFSTestEnvironment::new().await?;
+    target_env.start_rustfs_server_without_cleanup(vec![]).await?;
+
+    let source_bucket = "replication-checksum-src";
+    let target_bucket = "replication-checksum-dst";
+    let source_client = source_env.create_s3_client();
+    let target_client = target_env.create_s3_client();
+
+    source_client.create_bucket().bucket(source_bucket).send().await?;
+    target_client.create_bucket().bucket(target_bucket).send().await?;
+    enable_bucket_versioning(&source_env, source_bucket).await?;
+    enable_bucket_versioning(&target_env, target_bucket).await?;
+    let target_arn = set_replication_target(&source_env, source_bucket, &target_env, target_bucket).await?;
+    put_bucket_replication(&source_env, source_bucket, &target_arn).await?;
+
+    let body = b"123456789";
+    let crc32_key = "checksum-crc32.txt";
+    let sha256_key = "checksum-sha256.txt";
+    let crc32_put = source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(crc32_key)
+        .body(ByteStream::from_static(body))
+        .checksum_algorithm(aws_sdk_s3::types::ChecksumAlgorithm::Crc32)
+        .send()
+        .await?;
+    let expected_crc32 = crc32_put.checksum_crc32().ok_or("source PUT omitted CRC32")?.to_string();
+    let sha256_put = source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(sha256_key)
+        .body(ByteStream::from_static(body))
+        .checksum_algorithm(aws_sdk_s3::types::ChecksumAlgorithm::Sha256)
+        .send()
+        .await?;
+    let expected_sha256 = sha256_put.checksum_sha256().ok_or("source PUT omitted SHA256")?.to_string();
+
+    for key in [crc32_key, sha256_key] {
+        wait_for_source_replication_status(&source_client, source_bucket, key, "COMPLETED", false).await?;
+    }
+
+    let replica = target_client
+        .head_object()
+        .bucket(target_bucket)
+        .key(crc32_key)
+        .checksum_mode(aws_sdk_s3::types::ChecksumMode::Enabled)
+        .send()
+        .await?;
+    assert_eq!(replica.checksum_crc32(), Some(expected_crc32.as_str()), "replica lost the CRC32 checksum");
+    let replica = target_client
+        .head_object()
+        .bucket(target_bucket)
+        .key(sha256_key)
+        .checksum_mode(aws_sdk_s3::types::ChecksumMode::Enabled)
+        .send()
+        .await?;
+    assert_eq!(
+        replica.checksum_sha256(),
+        Some(expected_sha256.as_str()),
+        "replica lost the SHA256 checksum"
+    );
+    // The bare algorithm name must not leak as user metadata either.
+    assert!(
+        replica
+            .metadata()
+            .is_none_or(|meta| !meta.keys().any(|k| k.eq_ignore_ascii_case("sha256"))),
+        "replica carries the checksum as user metadata: {:?}",
+        replica.metadata()
+    );
+
+    Ok(())
+}
+
 #[tokio::test]
 async fn test_bucket_replication_disabled_delete_marker_does_not_propagate() -> TestResult {
     init_logging();

crates/e2e_test/src/replication_target_matrix_test.rs

@@ -31,17 +31,20 @@
 //! Adding a target behavior the fleet has shown: add the mode to the fake
 //! target, add a row here, and record any cell that is red before the fix.
 
-use crate::common::{RustFSTestEnvironment, init_logging, replication_fast_env};
-use crate::fake_s3_target::{FAKE_ACCESS_KEY, FAKE_SECRET_KEY};
-use crate::fake_s3_target::{FakeS3Target, Operation as FakeTargetOperation, RequestRecord};
-use crate::on_demand_migration::common::fake_source_client;
+use crate::common::{init_logging, replication_fast_env};
+use crate::fake_s3_target::{BucketMode, FAKE_ACCESS_KEY, FAKE_SECRET_KEY};
+use crate::fake_s3_target::{FakeS3Target, FaultAction as FakeTargetFault, Operation as FakeTargetOperation, RequestRecord};
+use crate::on_demand_migration::common::{OdmEnvOptions, OdmTestEnv, fake_source_client};
 use crate::replication_extension_test::{
-    LOOPBACK_REPLICATION_TARGET_ENV, ReplicationTargetOptions, enable_bucket_versioning, put_bucket_replication,
-    set_replication_target_with_options,
+    LOOPBACK_REPLICATION_TARGET_ENV, ReplicationTargetOptions, enable_bucket_versioning, get_replication_reset_status,
+    put_bucket_replication, set_replication_target_with_options, start_bucket_replication_reset,
 };
 use aws_sdk_s3::Client;
 use aws_sdk_s3::primitives::{ByteStream, DateTime};
-use aws_sdk_s3::types::{CompletedMultipartUpload, CompletedPart, ObjectLockLegalHoldStatus, ObjectLockMode};
+use aws_sdk_s3::types::{
+    Checksum, ChecksumAlgorithm, CompletedMultipartUpload, CompletedPart, ObjectAttributes, ObjectLockLegalHoldStatus,
+    ObjectLockMode,
+};
 use bytes::Bytes;
 use std::error::Error;
 use std::time::{SystemTime, UNIX_EPOCH};
@@ -110,16 +113,23 @@ enum ObjectShape {
     /// Two-part multipart upload with a GOVERNANCE retention period; the
     /// lock headers travel on CreateMultipartUpload, which has no body.
     LockedMultipart,
+    /// ODM stores two local parts while preserving a single-PUT source's MD5 ETag.
+    OdmPreservedMd5Multipart,
+    /// Single-part object uploaded with `x-amz-checksum-sha256`; the replica
+    /// must carry the same header (rustfs/backlog#2340).
+    Checksummed,
 }
 
 impl ObjectShape {
-    const ALL: [ObjectShape; 6] = [
+    const ALL: [ObjectShape; 8] = [
         ObjectShape::Empty,
         ObjectShape::Plain,
         ObjectShape::Retention,
         ObjectShape::LegalHold,
         ObjectShape::Multipart,
         ObjectShape::LockedMultipart,
+        ObjectShape::OdmPreservedMd5Multipart,
+        ObjectShape::Checksummed,
     ];
 
     fn key(self) -> &'static str {
@@ -130,6 +140,17 @@ impl ObjectShape {
             ObjectShape::LegalHold => "matrix/legal-hold.bin",
             ObjectShape::Multipart => "matrix/multipart.bin",
             ObjectShape::LockedMultipart => "matrix/locked-multipart.bin",
+            ObjectShape::OdmPreservedMd5Multipart => "matrix/odm-preserved-md5.bin",
+            ObjectShape::Checksummed => "matrix/checksummed.bin",
+        }
+    }
+
+    /// The `x-amz-checksum-*` header the source stored and every upload of
+    /// the replica must repeat.
+    fn forwarded_checksum_header(self) -> Option<&'static str> {
+        match self {
+            ObjectShape::Checksummed => Some("x-amz-checksum-sha256"),
+            _ => None,
         }
     }
 
@@ -139,7 +160,8 @@ impl ObjectShape {
 
     /// Upload the shape to the source and return the bytes the target must
     /// end up holding.
-    async fn put(self, client: &Client, bucket: &str) -> Result<Bytes, Box<dyn Error + Send + Sync>> {
+    async fn put(self, env: &OdmTestEnv, bucket: &str) -> Result<Bytes, Box<dyn Error + Send + Sync>> {
+        let client = &env.client;
         let key = self.key();
         match self {
             ObjectShape::Empty => {
@@ -190,6 +212,19 @@ impl ObjectShape {
             }
             ObjectShape::Multipart => multipart_put(client, bucket, key, 0x44, false).await,
             ObjectShape::LockedMultipart => multipart_put(client, bucket, key, 0x55, true).await,
+            ObjectShape::OdmPreservedMd5Multipart => odm_preserved_md5_multipart(env, bucket, key).await,
+            ObjectShape::Checksummed => {
+                let body = payload(40 * 1024, 0x66);
+                client
+                    .put_object()
+                    .bucket(bucket)
+                    .key(key)
+                    .body(ByteStream::from(body.clone()))
+                    .checksum_algorithm(ChecksumAlgorithm::Sha256)
+                    .send()
+                    .await?;
+                Ok(body)
+            }
         }
     }
 }
@@ -219,6 +254,170 @@ fn expectation(mode: TargetMode, shape: ObjectShape) -> Expectation {
         .unwrap_or(Expectation::Completed)
 }
 
+/// rustfs/backlog#2340: a target that mints its own version ids (Wasabi,
+/// AWS S3) answers 404 to a HEAD by the source uuid, which the worker used to
+/// read as "replica missing" and re-drive the PUT — one more target version
+/// per heal, MRF retry or resync. Two re-drive shapes, both must converge on
+/// the single version the first PUT created:
+/// - the first PUT lands but its response is lost, so the object is FAILED
+///   and the scanner heal pass re-drives it;
+/// - an existing-object resync re-drives a COMPLETED object unconditionally.
+#[tokio::test]
+async fn matrix_mint_own_version_ids_redrive_does_not_duplicate() -> TestResult {
+    init_logging();
+
+    let target = FakeS3Target::start().await?;
+    let target_bucket = "matrix-mint-own-redrive-dst".to_string();
+    target.create_bucket_with_object_lock(target_bucket.clone());
+    target.assign_own_version_ids(true);
+
+    let mut env_vars = replication_fast_env();
+    env_vars.extend_from_slice(LOOPBACK_REPLICATION_TARGET_ENV);
+    env_vars.extend_from_slice(&[
+        ("NO_PROXY", "127.0.0.1,localhost"),
+        ("HTTP_PROXY", ""),
+        ("HTTPS_PROXY", ""),
+        // The scanner heal pass is what re-drives a FAILED object.
+        ("RUSTFS_SCANNER_CYCLE", "1"),
+        ("RUSTFS_SCANNER_START_DELAY_SECS", "1"),
+    ]);
+    let env = OdmTestEnv::start_with(OdmEnvOptions {
+        env: env_vars,
+        ..OdmEnvOptions::default()
+    })
+    .await?;
+    let source_env = &env.rustfs;
+
+    let source_bucket = "matrix-mint-own-redrive-src";
+    let source_client = source_env.create_s3_client();
+    source_client
+        .create_bucket()
+        .bucket(source_bucket)
+        .object_lock_enabled_for_bucket(true)
+        .send()
+        .await?;
+    enable_bucket_versioning(source_env, source_bucket).await?;
+    let target_arn = set_replication_target_with_options(
+        source_env,
+        source_bucket,
+        ReplicationTargetOptions {
+            endpoint: &target.address(),
+            access_key: FAKE_ACCESS_KEY,
+            secret_key: FAKE_SECRET_KEY,
+            target_bucket: &target_bucket,
+            secure: false,
+            skip_tls_verify: false,
+            ca_cert_pem: None,
+        },
+    )
+    .await?;
+    put_bucket_replication(source_env, source_bucket, &target_arn).await?;
+
+    // Teach the worker the target's identity contract with one ordinary
+    // write, exactly as production learns it (the PUT response carries the
+    // minted id).
+    let probe_key = "redrive/identity-probe.bin";
+    source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(probe_key)
+        .body(ByteStream::from(payload(4 * 1024, 0x01)))
+        .send()
+        .await?;
+    assert_eq!(
+        wait_for_terminal_replication_status(&source_client, source_bucket, probe_key).await?,
+        "COMPLETED"
+    );
+
+    // Shape 1: the PUT is stored, its response never arrives, heal re-drives.
+    let heal_key = "redrive/heal.bin";
+    target.inject_for_key(FakeTargetOperation::PutObject, heal_key, FakeTargetFault::DisconnectAfterResponse, 1);
+    source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(heal_key)
+        .body(ByteStream::from(payload(8 * 1024, 0x02)))
+        .send()
+        .await?;
+    wait_for_replication_status_and_single_version(&source_client, source_bucket, &target, &target_bucket, heal_key).await?;
+
+    // Shape 2: an existing-object resync re-drives a COMPLETED object.
+    let resync_key = "redrive/resync.bin";
+    source_client
+        .put_object()
+        .bucket(source_bucket)
+        .key(resync_key)
+        .body(ByteStream::from(payload(8 * 1024, 0x03)))
+        .send()
+        .await?;
+    assert_eq!(
+        wait_for_terminal_replication_status(&source_client, source_bucket, resync_key).await?,
+        "COMPLETED"
+    );
+    let (reset_arn, _reset_id) = start_bucket_replication_reset(source_env, source_bucket).await?;
+    assert_eq!(reset_arn, target_arn);
+    let resync = async {
+        loop {
+            let status = get_replication_reset_status(source_env, source_bucket, &target_arn).await?;
+            if let Some(entry) = status.targets.iter().find(|entry| entry.arn == target_arn)
+                && entry.status == "Completed"
+            {
+                return Ok::<_, Box<dyn Error + Send + Sync>>(entry.replicated_count);
+            }
+            sleep(Duration::from_millis(250)).await;
+        }
+    };
+    let replicated = timeout(Duration::from_secs(90), resync)
+        .await
+        .map_err(|_| "existing-object resync did not complete within 90 seconds")??;
+    assert!(replicated >= 3, "resync must count the located replicas as replicated, got {replicated}");
+    for key in [probe_key, heal_key, resync_key] {
+        let versions = target.stored_versions(&target_bucket, key);
+        assert_eq!(
+            versions.len(),
+            1,
+            "{key}: a re-drive against a target that mints its own version ids must not mint another one: {versions:?}"
+        );
+    }
+
+    target.shutdown().await;
+    Ok(())
+}
+
+/// Wait until `key` is COMPLETED on the source and, for the observation
+/// window after that, the target still holds exactly one live version of it.
+async fn wait_for_replication_status_and_single_version(
+    source_client: &Client,
+    source_bucket: &str,
+    target: &FakeS3Target,
+    target_bucket: &str,
+    key: &str,
+) -> TestResult {
+    // The lost PUT response first settles the object FAILED; only the next
+    // scanner heal pass can turn that into COMPLETED, so FAILED is transient
+    // here and the wait is for COMPLETED alone.
+    let converged = async {
+        loop {
+            let head = source_client.head_object().bucket(source_bucket).key(key).send().await?;
+            if head.replication_status().is_some_and(|status| status.as_str() == "COMPLETED") {
+                return Ok::<_, Box<dyn Error + Send + Sync>>(());
+            }
+            sleep(Duration::from_millis(250)).await;
+        }
+    };
+    timeout(Duration::from_secs(90), converged)
+        .await
+        .map_err(|_| format!("{key}: heal re-drive did not converge to COMPLETED within 90 seconds"))??;
+    // The heal pass keeps visiting the key for a few scanner cycles; a
+    // duplicate would show up here as a second stored version.
+    for _ in 0..12 {
+        let versions = target.stored_versions(target_bucket, key);
+        assert_eq!(versions.len(), 1, "{key}: target minted another version on re-drive: {versions:?}");
+        sleep(Duration::from_millis(500)).await;
+    }
+    Ok(())
+}
+
 #[tokio::test]
 async fn matrix_baseline_target() -> TestResult {
     run_row(TargetMode::Baseline).await
@@ -270,11 +469,15 @@ async fn run_row(mode: TargetMode) -> TestResult {
     target.create_bucket_with_object_lock(target_bucket.clone());
     mode.apply(&target);
 
-    let mut source_env = RustFSTestEnvironment::new().await?;
     let mut env_vars = replication_fast_env();
     env_vars.extend_from_slice(LOOPBACK_REPLICATION_TARGET_ENV);
     env_vars.extend_from_slice(&[("NO_PROXY", "127.0.0.1,localhost"), ("HTTP_PROXY", ""), ("HTTPS_PROXY", "")]);
-    source_env.start_rustfs_server_with_env(vec![], &env_vars).await?;
+    let env = OdmTestEnv::start_with(OdmEnvOptions {
+        env: env_vars,
+        ..OdmEnvOptions::default()
+    })
+    .await?;
+    let source_env = &env.rustfs;
 
     let source_bucket = format!("matrix-{}-src", mode.slug());
     let source_client = source_env.create_s3_client();
@@ -284,9 +487,9 @@ async fn run_row(mode: TargetMode) -> TestResult {
         .object_lock_enabled_for_bucket(true)
         .send()
         .await?;
-    enable_bucket_versioning(&source_env, &source_bucket).await?;
+    enable_bucket_versioning(source_env, &source_bucket).await?;
     let target_arn = set_replication_target_with_options(
-        &source_env,
+        source_env,
         &source_bucket,
         ReplicationTargetOptions {
             endpoint: &target.address(),
@@ -299,14 +502,21 @@ async fn run_row(mode: TargetMode) -> TestResult {
         },
     )
     .await?;
-    put_bucket_replication(&source_env, &source_bucket, &target_arn).await?;
+    put_bucket_replication(source_env, &source_bucket, &target_arn).await?;
 
     let target_client = fake_source_client(&target);
     let mut failures = Vec::new();
     for shape in ObjectShape::ALL {
         let cell = format!("{}/{:?}", mode.slug(), shape);
-        let expected_body = shape.put(&source_client, &source_bucket).await?;
+        let expected_body = shape.put(&env, &source_bucket).await?;
         let status = wait_for_terminal_replication_status(&source_client, &source_bucket, shape.key()).await?;
+        if shape == ObjectShape::OdmPreservedMd5Multipart {
+            assert_eq!(
+                env.source.count_requests(FakeTargetOperation::GetObject, shape.key()),
+                2,
+                "one passthrough GET plus one background pull; replication must read the persisted local parts"
+            );
+        }
         let journal = target.requests();
         let outcome = match expectation(mode, shape) {
             Expectation::Completed => {
@@ -379,6 +589,36 @@ async fn check_completed_cell(
     if uploads.is_empty() {
         return Err("no upload reached the target although the source reports COMPLETED".into());
     }
+    if shape == ObjectShape::OdmPreservedMd5Multipart {
+        let key_requests: Vec<_> = journal
+            .iter()
+            .filter(|record| record.key.as_deref() == Some(shape.key()))
+            .collect();
+        for operation in [
+            FakeTargetOperation::CreateMultipartUpload,
+            FakeTargetOperation::CompleteMultipartUpload,
+        ] {
+            if !key_requests.iter().any(|record| record.operation == operation) {
+                return Err(format!("preserved-MD5 multipart object did not use {operation:?}").into());
+            }
+        }
+        if key_requests
+            .iter()
+            .any(|record| record.operation == FakeTargetOperation::PutObject)
+        {
+            return Err("preserved-MD5 multipart object used a single PutObject".into());
+        }
+        let mut part_numbers: Vec<_> = key_requests
+            .iter()
+            .filter(|record| record.operation == FakeTargetOperation::UploadPart)
+            .map(|record| record.part_number)
+            .collect();
+        part_numbers.sort_unstable();
+        part_numbers.dedup();
+        if part_numbers != [Some(1), Some(2)] {
+            return Err(format!("preserved-MD5 multipart object uploaded unexpected parts: {part_numbers:?}").into());
+        }
+    }
     if let Some(framed) = uploads.iter().find(|record| record.transport.aws_chunked) {
         return Err(format!("{cell}: an upload went out aws-chunked (rustfs#6853 framing): {framed:?}").into());
     }
@@ -401,6 +641,19 @@ async fn check_completed_cell(
     }) {
         return Err(format!("a locked PutObject went out without any integrity header (rustfs#7082): {bare:?}").into());
     }
+    // rustfs/backlog#2340 contract: a source checksum reaches the target as
+    // the `x-amz-checksum-*` header, not as user metadata; every PutObject of
+    // the shape carries it.
+    if let Some(header) = shape.forwarded_checksum_header()
+        && let Some(missing) = uploads.iter().find(|record| {
+            record.operation == FakeTargetOperation::PutObject
+                && !record.transport.checksum_headers.iter().any(|name| name == header)
+        })
+    {
+        return Err(
+            format!("a PutObject went out without the source's {header} header (rustfs/backlog#2340): {missing:?}").into(),
+        );
+    }
     Ok(())
 }
 
@@ -455,6 +708,71 @@ async fn wait_for_terminal_replication_status(
     }
 }
 
+async fn odm_preserved_md5_multipart(env: &OdmTestEnv, bucket: &str, key: &str) -> Result<Bytes, Box<dyn Error + Send + Sync>> {
+    const PART_SIZE: usize = 5 * 1024 * 1024;
+    let origin_bucket = format!("{bucket}-origin");
+    env.source.create_bucket_with_mode(&origin_bucket, BucketMode::Unversioned);
+    let mut spec = env.fake_source_spec(&origin_bucket);
+    // Below the 16 MiB inline default the pull is one tee'd PUT with a single
+    // part; force the passthrough + background multipart write-back instead.
+    spec.policy.inline_max_bytes = 4096;
+    spec.policy.multipart_part_size_bytes = PART_SIZE as u64;
+    spec.policy.preserve_etag = true;
+    env.configure_and_wait(bucket, &spec).await?;
+
+    // A normal source PUT produces the MD5 ETag; only ODM chooses the local parts.
+    let body = payload(PART_SIZE + 4096, 0x66);
+    let source_put = env
+        .source_client()
+        .put_object()
+        .bucket(&origin_bucket)
+        .key(key)
+        .body(ByteStream::from(body.clone()))
+        .send()
+        .await?;
+    let source_etag = source_put.e_tag().ok_or("source PUT omitted its ETag")?.trim_matches('"');
+    assert_eq!(source_etag.len(), 32, "source fixture must have a single-PUT MD5 ETag");
+    assert!(source_etag.bytes().all(|byte| byte.is_ascii_hexdigit()));
+
+    let pulled = env.raw_get(bucket, key).await?;
+    assert_eq!(pulled.status, 200, "{}", String::from_utf8_lossy(&pulled.body));
+    assert_eq!(pulled.body, body);
+    assert!(
+        env.wait_local_listed(bucket, key, Duration::from_secs(30)).await?,
+        "ODM must persist the object"
+    );
+    let attributes = env
+        .client
+        .get_object_attributes()
+        .bucket(bucket)
+        .key(key)
+        .object_attributes(ObjectAttributes::Etag)
+        .object_attributes(ObjectAttributes::ObjectParts)
+        .object_attributes(ObjectAttributes::Checksum)
+        .send()
+        .await?;
+    assert_eq!(attributes.e_tag().map(|etag| etag.trim_matches('"')), Some(source_etag));
+    let parts = attributes
+        .object_parts()
+        .ok_or("the ODM copy must expose its two local parts")?;
+    assert_eq!(parts.total_parts_count(), Some(2));
+    assert_eq!(
+        parts
+            .parts()
+            .iter()
+            .map(|part| (part.part_number(), part.size()))
+            .collect::<Vec<_>>(),
+        [(Some(1), Some(PART_SIZE as i64)), (Some(2), Some(4096))]
+    );
+    assert!(
+        attributes
+            .checksum()
+            .is_none_or(|checksum| checksum == &Checksum::builder().build()),
+        "multipart routing must work without an object checksum record"
+    );
+    Ok(body)
+}
+
 async fn multipart_put(
     client: &Client,
     bucket: &str,

crates/e2e_test/src/upgrade_compatibility_test.rs

@@ -14,9 +14,10 @@
 
 use crate::common::{
     RustFSTestClusterEnvironment, RustFSTestEnvironment, admin_request, init_logging, replication_fast_env, rustfs_binary_path,
+    signed_request,
 };
-use crate::fake_s3_target::{BucketMode, FAKE_ACCESS_KEY, FAKE_SECRET_KEY, FakeS3Target};
-use crate::on_demand_migration::common::{ODM_SERVER_ENV, OdmTestEnv, SeedObject};
+use crate::fake_s3_target::{BucketMode, FAKE_ACCESS_KEY, FAKE_SECRET_KEY, FakeS3Target, Operation as FakeTargetOperation};
+use crate::on_demand_migration::common::{ODM_SERVER_ENV, OdmTestEnv, SeedObject, fake_source_client};
 use crate::replication_extension_test::{
     LOOPBACK_REPLICATION_TARGET_ENV, ReplicationTargetOptions, put_bucket_replication, set_replication_target_with_options,
 };
@@ -25,9 +26,10 @@ use aws_sdk_s3::error::ProvideErrorMetadata;
 use aws_sdk_s3::primitives::ByteStream;
 use aws_sdk_s3::types::{
     BucketLifecycleConfiguration, BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, DefaultRetention,
-    ExpirationStatus, LifecycleExpiration, LifecycleRule, LifecycleRuleFilter, ObjectLockConfiguration, ObjectLockEnabled,
-    ObjectLockRetentionMode, ObjectLockRule, PublicAccessBlockConfiguration, ServerSideEncryption, ServerSideEncryptionByDefault,
-    ServerSideEncryptionConfiguration, ServerSideEncryptionRule, Tag, Tagging, VersioningConfiguration,
+    ExpirationStatus, LifecycleExpiration, LifecycleRule, LifecycleRuleFilter, ObjectAttributes, ObjectLockConfiguration,
+    ObjectLockEnabled, ObjectLockRetentionMode, ObjectLockRule, PublicAccessBlockConfiguration, ServerSideEncryption,
+    ServerSideEncryptionByDefault, ServerSideEncryptionConfiguration, ServerSideEncryptionRule, Tag, Tagging,
+    VersioningConfiguration,
 };
 use http::{Method, StatusCode};
 use std::path::{Path, PathBuf};
@@ -1205,3 +1207,724 @@ async fn rollback_to_previous_release_reads_current_bucket_metadata() -> TestRes
     replication_target.shutdown().await;
     Ok(())
 }
+
+// ---------------------------------------------------------------------------
+// rc.5 multipart layouts under the current build (backlog#2147 follow-up to
+// rustfs#7305)
+// ---------------------------------------------------------------------------
+//
+// rustfs#7305 changed `ObjectInfo::is_multipart` to consult the stored part
+// list before the ETag shape. Every earlier check of that change used
+// synthetic metadata; this scenario writes the layouts with the published
+// rc.5 binary and then reads, describes, and replicates them with the
+// current build on the same data directory.
+
+const LAYOUT_PLAIN_BUCKET: &str = "upgrade-layout-plain";
+const LAYOUT_ENCRYPTED_BUCKET: &str = "upgrade-layout-encrypted";
+const LAYOUT_REPLICA_BUCKET: &str = "upgrade-layout-replica";
+const LAYOUT_PART_SIZE: usize = 5 * 1024 * 1024;
+const LAYOUT_TAIL_SIZE: usize = 1024 * 1024 + 4096;
+const LAYOUT_SSEC_KEY: &str = "0123456789abcdef0123456789abcdef";
+const LAYOUT_REPLICATION_TIMEOUT: Duration = Duration::from_secs(180);
+
+struct LayoutCase {
+    bucket: &'static str,
+    key: &'static str,
+    /// Empty for a single PUT.
+    part_sizes: Vec<usize>,
+    body: Vec<u8>,
+    ssec: bool,
+    /// `false` for layouts whose replication is a known pre-existing failure;
+    /// their outcome is logged, not asserted.
+    assert_replication: bool,
+    /// Recorded from the rc.5 writer.
+    rc5_etag: String,
+    /// Whether rc.5 reported `ObjectParts` for the object.
+    rc5_reported_parts: Option<usize>,
+}
+
+impl LayoutCase {
+    fn is_multipart_layout(&self) -> bool {
+        self.part_sizes.len() > 1
+    }
+
+    fn label(&self) -> String {
+        format!("{}/{}", self.bucket, self.key)
+    }
+}
+
+fn layout_noise(len: usize, seed: u64) -> Vec<u8> {
+    let mut state = seed ^ 0x9E37_79B9_7F4A_7C15;
+    (0..len)
+        .map(|_| {
+            state ^= state << 13;
+            state ^= state >> 7;
+            state ^= state << 17;
+            (state >> 24) as u8
+        })
+        .collect()
+}
+
+fn layout_text(len: usize, seed: u64) -> Vec<u8> {
+    let mut out = Vec::with_capacity(len + 64);
+    let mut line = 0u64;
+    while out.len() < len {
+        out.extend_from_slice(format!("rc5 legacy layout seed={seed} line={line} lorem ipsum dolor sit amet\n").as_bytes());
+        line += 1;
+    }
+    out.truncate(len);
+    out
+}
+
+fn layout_ssec_key_md5() -> String {
+    use md5::{Digest as _, Md5};
+    let mut hasher = Md5::new();
+    hasher.update(LAYOUT_SSEC_KEY.as_bytes());
+    base64_simd::STANDARD.encode_to_string(hasher.finalize())
+}
+
+fn layout_ssec_key() -> String {
+    base64_simd::STANDARD.encode_to_string(LAYOUT_SSEC_KEY)
+}
+
+async fn layout_head(
+    client: &Client,
+    case: &LayoutCase,
+) -> Result<aws_sdk_s3::operation::head_object::HeadObjectOutput, BoxError> {
+    let request = client.head_object().bucket(case.bucket).key(case.key);
+    let request = if case.ssec {
+        request
+            .sse_customer_algorithm("AES256")
+            .sse_customer_key(layout_ssec_key())
+            .sse_customer_key_md5(layout_ssec_key_md5())
+    } else {
+        request
+    };
+    Ok(request.send().await?)
+}
+
+async fn layout_get(
+    client: &Client,
+    case: &LayoutCase,
+    range: Option<String>,
+    part_number: Option<i32>,
+) -> Result<aws_sdk_s3::operation::get_object::GetObjectOutput, BoxError> {
+    let request = client.get_object().bucket(case.bucket).key(case.key);
+    let request = if case.ssec {
+        request
+            .sse_customer_algorithm("AES256")
+            .sse_customer_key(layout_ssec_key())
+            .sse_customer_key_md5(layout_ssec_key_md5())
+    } else {
+        request
+    };
+    let request = request.set_range(range).set_part_number(part_number);
+    Ok(request.send().await?)
+}
+
+async fn layout_attributes(
+    client: &Client,
+    case: &LayoutCase,
+) -> Result<aws_sdk_s3::operation::get_object_attributes::GetObjectAttributesOutput, BoxError> {
+    let request = client
+        .get_object_attributes()
+        .bucket(case.bucket)
+        .key(case.key)
+        .object_attributes(ObjectAttributes::Etag)
+        .object_attributes(ObjectAttributes::ObjectParts)
+        .object_attributes(ObjectAttributes::ObjectSize)
+        .max_parts(100);
+    let request = if case.ssec {
+        request
+            .sse_customer_algorithm("AES256")
+            .sse_customer_key(layout_ssec_key())
+            .sse_customer_key_md5(layout_ssec_key_md5())
+    } else {
+        request
+    };
+    Ok(request.send().await?)
+}
+
+/// Write `case` with the rc.5 client; single PUT when `part_sizes` is empty.
+async fn layout_write(client: &Client, case: &LayoutCase) -> Result<(), BoxError> {
+    let content_type = "text/plain";
+    if case.part_sizes.is_empty() {
+        let request = client
+            .put_object()
+            .bucket(case.bucket)
+            .key(case.key)
+            .content_type(content_type)
+            .body(ByteStream::from(case.body.clone()));
+        let request = if case.ssec {
+            request
+                .sse_customer_algorithm("AES256")
+                .sse_customer_key(layout_ssec_key())
+                .sse_customer_key_md5(layout_ssec_key_md5())
+        } else {
+            request
+        };
+        request.send().await?;
+        return Ok(());
+    }
+    let create = client
+        .create_multipart_upload()
+        .bucket(case.bucket)
+        .key(case.key)
+        .content_type(content_type);
+    let create = if case.ssec {
+        create
+            .sse_customer_algorithm("AES256")
+            .sse_customer_key(layout_ssec_key())
+            .sse_customer_key_md5(layout_ssec_key_md5())
+    } else {
+        create
+    };
+    let created = create.send().await?;
+    let upload_id = created.upload_id().ok_or("CreateMultipartUpload omitted upload ID")?;
+    let mut completed = Vec::with_capacity(case.part_sizes.len());
+    let mut offset = 0usize;
+    for (index, size) in case.part_sizes.iter().enumerate() {
+        let part_number = i32::try_from(index + 1)?;
+        let chunk = case.body[offset..offset + size].to_vec();
+        offset += size;
+        let upload = client
+            .upload_part()
+            .bucket(case.bucket)
+            .key(case.key)
+            .upload_id(upload_id)
+            .part_number(part_number)
+            .body(ByteStream::from(chunk));
+        let upload = if case.ssec {
+            upload
+                .sse_customer_algorithm("AES256")
+                .sse_customer_key(layout_ssec_key())
+                .sse_customer_key_md5(layout_ssec_key_md5())
+        } else {
+            upload
+        };
+        let uploaded = upload.send().await?;
+        completed.push(
+            CompletedPart::builder()
+                .part_number(part_number)
+                .e_tag(uploaded.e_tag().ok_or("UploadPart omitted ETag")?)
+                .build(),
+        );
+    }
+    client
+        .complete_multipart_upload()
+        .bucket(case.bucket)
+        .key(case.key)
+        .upload_id(upload_id)
+        .multipart_upload(CompletedMultipartUpload::builder().set_parts(Some(completed)).build())
+        .send()
+        .await?;
+    Ok(())
+}
+
+fn layout_cases() -> Vec<LayoutCase> {
+    let two = vec![LAYOUT_PART_SIZE, LAYOUT_TAIL_SIZE];
+    let three = vec![LAYOUT_PART_SIZE, LAYOUT_PART_SIZE, 4096];
+    let total = |sizes: &[usize]| sizes.iter().sum::<usize>();
+    let case = |bucket, key, part_sizes: Vec<usize>, body: Vec<u8>, ssec| LayoutCase {
+        bucket,
+        key,
+        part_sizes,
+        body,
+        ssec,
+        assert_replication: true,
+        rc5_etag: String::new(),
+        rc5_reported_parts: None,
+    };
+    vec![
+        case(LAYOUT_PLAIN_BUCKET, "plain/single.bin", vec![], layout_noise(1024 * 1024 + 17, 1), false),
+        case(
+            LAYOUT_PLAIN_BUCKET,
+            "plain/multipart-2.bin",
+            two.clone(),
+            layout_noise(total(&two), 2),
+            false,
+        ),
+        case(
+            LAYOUT_PLAIN_BUCKET,
+            "plain/multipart-3.bin",
+            three.clone(),
+            layout_noise(total(&three), 3),
+            false,
+        ),
+        case(
+            LAYOUT_PLAIN_BUCKET,
+            "plain/compressed-single.txt",
+            vec![],
+            layout_text(1024 * 1024 + 17, 4),
+            false,
+        ),
+        case(
+            LAYOUT_PLAIN_BUCKET,
+            "plain/compressed-multipart-2.txt",
+            two.clone(),
+            layout_text(total(&two), 5),
+            false,
+        ),
+        case(
+            LAYOUT_PLAIN_BUCKET,
+            "plain/ssec-multipart-2.bin",
+            two.clone(),
+            layout_noise(total(&two), 6),
+            true,
+        ),
+        // SSE-C passthrough replicates the stored ciphertext part by part; a
+        // compressible first part is stored well below 5 MiB and a standard
+        // target rejects it with EntityTooSmall. rc.5 fails the same way (see
+        // `rc5_baseline_replicates_multipart_layouts`), so the outcome is
+        // recorded rather than asserted here; tracked as rustfs/backlog#2363.
+        LayoutCase {
+            assert_replication: false,
+            ..case(
+                LAYOUT_PLAIN_BUCKET,
+                "plain/ssec-compressed-multipart-2.txt",
+                two.clone(),
+                layout_text(total(&two), 7),
+                true,
+            )
+        },
+        case(
+            LAYOUT_ENCRYPTED_BUCKET,
+            "encrypted/single.bin",
+            vec![],
+            layout_noise(1024 * 1024 + 17, 8),
+            false,
+        ),
+        case(
+            LAYOUT_ENCRYPTED_BUCKET,
+            "encrypted/multipart-2.bin",
+            two.clone(),
+            layout_noise(total(&two), 9),
+            false,
+        ),
+        case(
+            LAYOUT_ENCRYPTED_BUCKET,
+            "encrypted/multipart-3.bin",
+            three.clone(),
+            layout_noise(total(&three), 10),
+            false,
+        ),
+        case(
+            LAYOUT_ENCRYPTED_BUCKET,
+            "encrypted/compressed-multipart-2.txt",
+            two.clone(),
+            layout_text(total(&two), 11),
+            false,
+        ),
+    ]
+}
+
+fn layout_server_env() -> Vec<(&'static str, &'static str)> {
+    let mut env = bucket_config_server_env();
+    env.push(("RUSTFS_COMPRESSION_ENABLED", "true"));
+    env.push(("RUSTFS_COMPRESSION_MULTIPART_ENABLED", "true"));
+    env
+}
+
+fn layout_reported_parts(attributes: &aws_sdk_s3::operation::get_object_attributes::GetObjectAttributesOutput) -> Option<usize> {
+    attributes.object_parts().map(|parts| parts.parts().len())
+}
+
+async fn assert_layout_readable(client: &Client, case: &LayoutCase, context: &str) -> TestResult {
+    let label = case.label();
+    let head = layout_head(client, case).await?;
+    assert_eq!(
+        head.e_tag().map(|etag| etag.trim_matches('"')),
+        Some(case.rc5_etag.as_str()),
+        "{context}: {label}: the ETag written by rc.5 must be reported unchanged"
+    );
+    assert_eq!(
+        head.content_length(),
+        Some(i64::try_from(case.body.len())?),
+        "{context}: {label}: HEAD content length"
+    );
+
+    let full = layout_get(client, case, None, None).await?.body.collect().await?.into_bytes();
+    assert_eq!(full.len(), case.body.len(), "{context}: {label}: full GET length");
+    assert!(full == case.body, "{context}: {label}: full GET body must equal the rc.5 upload");
+
+    if case.is_multipart_layout() {
+        let first = case.part_sizes[0];
+        let range = format!("bytes={}-{}", first - 32, first + 31);
+        let crossing = layout_get(client, case, Some(range), None)
+            .await?
+            .body
+            .collect()
+            .await?
+            .into_bytes();
+        assert!(
+            crossing == case.body[first - 32..first + 32],
+            "{context}: {label}: range across the first part boundary"
+        );
+        let tail_start: usize = case.part_sizes[..case.part_sizes.len() - 1].iter().sum();
+        let last_number = i32::try_from(case.part_sizes.len())?;
+        let last = layout_get(client, case, None, Some(last_number)).await?;
+        assert_eq!(
+            last.content_length(),
+            Some(i64::try_from(case.part_sizes[case.part_sizes.len() - 1])?),
+            "{context}: {label}: partNumber={last_number} length"
+        );
+        let last_body = last.body.collect().await?.into_bytes();
+        assert!(
+            last_body == case.body[tail_start..],
+            "{context}: {label}: partNumber={last_number} body must be the stored last part"
+        );
+    }
+    Ok(())
+}
+
+async fn assert_layout_attributes(client: &Client, case: &LayoutCase, context: &str) -> TestResult {
+    let label = case.label();
+    let attributes = layout_attributes(client, case).await?;
+    assert_eq!(
+        attributes.e_tag().map(|etag| etag.trim_matches('"')),
+        Some(case.rc5_etag.as_str()),
+        "{context}: {label}: attributes ETag"
+    );
+    assert_eq!(
+        attributes.object_size(),
+        Some(i64::try_from(case.body.len())?),
+        "{context}: {label}: attributes ObjectSize"
+    );
+    if case.is_multipart_layout() {
+        let parts = attributes
+            .object_parts()
+            .ok_or_else(|| format!("{context}: {label}: multipart layout must expose ObjectParts"))?;
+        assert_eq!(
+            parts.total_parts_count(),
+            Some(i32::try_from(case.part_sizes.len())?),
+            "{context}: {label}: TotalPartsCount"
+        );
+        let observed: Vec<(Option<i32>, Option<i64>)> =
+            parts.parts().iter().map(|part| (part.part_number(), part.size())).collect();
+        let expected: Vec<(Option<i32>, Option<i64>)> = case
+            .part_sizes
+            .iter()
+            .enumerate()
+            .map(|(index, size)| (Some(index as i32 + 1), Some(*size as i64)))
+            .collect();
+        assert_eq!(
+            observed, expected,
+            "{context}: {label}: ObjectParts must report the plaintext part layout"
+        );
+    } else {
+        assert!(
+            attributes.object_parts().is_none_or(|parts| parts.parts().is_empty()),
+            "{context}: {label}: a single PUT must not report stored parts"
+        );
+    }
+    Ok(())
+}
+
+async fn put_layout_replication_rule(env: &RustFSTestEnvironment, bucket: &str, arn: &str) -> TestResult {
+    let body = format!(
+        r#"<ReplicationConfiguration xmlns="http://s3.amazonaws.com/doc/2006-03-01/">
+  <Role></Role>
+  <Rule>
+    <ID>legacy-layouts</ID>
+    <Priority>1</Priority>
+    <Status>Enabled</Status>
+    <Filter><Prefix></Prefix></Filter>
+    <DeleteMarkerReplication><Status>Enabled</Status></DeleteMarkerReplication>
+    <DeleteReplication><Status>Enabled</Status></DeleteReplication>
+    <ExistingObjectReplication><Status>Enabled</Status></ExistingObjectReplication>
+    <Destination><Bucket>{arn}</Bucket></Destination>
+  </Rule>
+</ReplicationConfiguration>"#
+    );
+    let url = format!("{}/{bucket}?replication", env.url);
+    let response = signed_request(
+        Method::PUT,
+        &url,
+        &env.access_key,
+        &env.secret_key,
+        Some(body.into_bytes()),
+        Some("application/xml"),
+    )
+    .await?;
+    if response.status() != StatusCode::OK {
+        let status = response.status();
+        let body = response.text().await.unwrap_or_default();
+        return Err(format!("put replication rule on {bucket} failed: {status} {body}").into());
+    }
+    Ok(())
+}
+
+/// Wait for the existing-object replication of `case` to reach a terminal
+/// status and return it (`COMPLETED` or `FAILED`).
+async fn wait_layout_replication_terminal(client: &Client, case: &LayoutCase) -> Result<String, BoxError> {
+    let deadline = Instant::now() + LAYOUT_REPLICATION_TIMEOUT;
+    loop {
+        let head = layout_head(client, case).await?;
+        let status = head.replication_status().map(|status| status.as_str().to_string());
+        if matches!(status.as_deref(), Some("COMPLETED") | Some("FAILED")) {
+            return Ok(status.unwrap_or_default());
+        }
+        if Instant::now() >= deadline {
+            return Err(format!(
+                "{}: existing-object replication never reached a terminal status; last {status:?}",
+                case.label()
+            )
+            .into());
+        }
+        sleep(Duration::from_millis(500)).await;
+    }
+}
+
+#[derive(Debug)]
+struct LayoutTransport {
+    status: String,
+    uploaded_parts: Vec<i32>,
+    single_puts: usize,
+    completes: usize,
+    /// Raw per-key journal in target order: (sequence, operation, part number,
+    /// upload id), so duplicate drives can be told apart from retries.
+    journal: Vec<(u64, String, Option<i32>, Option<String>)>,
+}
+
+/// Configure every layout bucket to replicate its existing objects to a fresh
+/// fake target, wait for each case to settle, and report the transport the
+/// target observed per case.
+async fn replicate_layouts(
+    env: &RustFSTestEnvironment,
+    client: &Client,
+    cases: &[LayoutCase],
+) -> Result<(FakeS3Target, Vec<LayoutTransport>), BoxError> {
+    let target = FakeS3Target::start().await?;
+    target.create_bucket(LAYOUT_REPLICA_BUCKET);
+    for bucket in [LAYOUT_PLAIN_BUCKET, LAYOUT_ENCRYPTED_BUCKET] {
+        let arn = set_replication_target_with_options(
+            env,
+            bucket,
+            ReplicationTargetOptions {
+                endpoint: &target.address(),
+                access_key: FAKE_ACCESS_KEY,
+                secret_key: FAKE_SECRET_KEY,
+                target_bucket: LAYOUT_REPLICA_BUCKET,
+                secure: false,
+                skip_tls_verify: false,
+                ca_cert_pem: None,
+            },
+        )
+        .await?;
+        put_layout_replication_rule(env, bucket, &arn).await?;
+    }
+    let mut statuses = Vec::with_capacity(cases.len());
+    for case in cases {
+        statuses.push(wait_layout_replication_terminal(client, case).await?);
+    }
+    let journal = target.requests();
+    let mut transports = Vec::with_capacity(cases.len());
+    for (case, status) in cases.iter().zip(statuses) {
+        let key_requests: Vec<_> = journal
+            .iter()
+            .filter(|record| record.key.as_deref() == Some(case.key))
+            .collect();
+        let mut uploaded_parts: Vec<i32> = key_requests
+            .iter()
+            .filter(|record| record.operation == FakeTargetOperation::UploadPart)
+            .filter_map(|record| record.part_number)
+            .collect();
+        uploaded_parts.sort_unstable();
+        uploaded_parts.dedup();
+        let transport = LayoutTransport {
+            status,
+            uploaded_parts,
+            single_puts: key_requests
+                .iter()
+                .filter(|record| record.operation == FakeTargetOperation::PutObject)
+                .count(),
+            completes: key_requests
+                .iter()
+                .filter(|record| record.operation == FakeTargetOperation::CompleteMultipartUpload)
+                .count(),
+            journal: key_requests
+                .iter()
+                .map(|record| {
+                    (
+                        record.sequence,
+                        format!("{:?}", record.operation),
+                        record.part_number,
+                        record.upload_id.as_ref().map(|id| id.chars().take(12).collect()),
+                    )
+                })
+                .collect(),
+        };
+        tracing::info!(
+            target: "e2e_test::upgrade_compatibility_test",
+            object = %case.label(),
+            ?transport,
+            "replication transport observed on the target"
+        );
+        transports.push(transport);
+    }
+    Ok((target, transports))
+}
+
+/// rc.5 writes single-PUT, multipart, compressed, SSE-C and SSE-S3 layouts;
+/// the current build must read every byte, expose the stored part layout
+/// through GetObjectAttributes and partNumber reads, and replicate the objects
+/// with the transport that matches their stored parts.
+#[tokio::test]
+#[ignore = "requires the pinned 1.0.0-rc.5 release binary"]
+async fn direct_upgrade_from_rc5_preserves_multipart_layouts() -> TestResult {
+    init_logging();
+    let previous_binary = source_binary()?;
+    let server_env = layout_server_env();
+
+    let mut env = RustFSTestEnvironment::new().await?;
+    env.start_rustfs_server_from_binary(&previous_binary, vec![], &server_env)
+        .await?;
+    let old_client = env.create_s3_client();
+    env.create_test_bucket(LAYOUT_PLAIN_BUCKET).await?;
+    env.create_test_bucket(LAYOUT_ENCRYPTED_BUCKET).await?;
+    enable_versioning(&old_client, LAYOUT_PLAIN_BUCKET).await?;
+    enable_versioning(&old_client, LAYOUT_ENCRYPTED_BUCKET).await?;
+    put_default_sse_s3_encryption(&old_client, LAYOUT_ENCRYPTED_BUCKET).await?;
+    assert_default_sse_s3_encryption(&old_client, LAYOUT_ENCRYPTED_BUCKET, "rc.5").await?;
+
+    let mut cases = layout_cases();
+    for case in cases.iter_mut() {
+        layout_write(&old_client, case).await?;
+        let head = layout_head(&old_client, case).await?;
+        case.rc5_etag = head
+            .e_tag()
+            .ok_or_else(|| format!("{}: rc.5 HEAD omitted the ETag", case.label()))?
+            .trim_matches('"')
+            .to_string();
+        case.rc5_reported_parts = layout_attributes(&old_client, case)
+            .await
+            .ok()
+            .and_then(|a| layout_reported_parts(&a));
+        tracing::info!(
+            target: "e2e_test::upgrade_compatibility_test",
+            object = %case.label(),
+            parts = case.part_sizes.len(),
+            etag = %case.rc5_etag,
+            rc5_reported_parts = ?case.rc5_reported_parts,
+            "rc.5 wrote a legacy layout"
+        );
+    }
+    // The rc.5 writer must itself still read what it wrote, so a later
+    // failure is attributable to the upgrade rather than to the fixture.
+    for case in &cases {
+        assert_layout_readable(&old_client, case, "rc.5").await?;
+    }
+
+    // Upgrade in place.
+    env.restart_server_preserving_data(vec![], &server_env).await?;
+    let client = env.create_s3_client();
+    for case in &cases {
+        assert_layout_readable(&client, case, "upgraded").await?;
+        assert_layout_attributes(&client, case, "upgraded").await?;
+    }
+
+    // Replicate the pre-existing objects with the current build.
+    let (target, transports) = replicate_layouts(&env, &client, &cases).await?;
+    let replica_client = fake_source_client(&target);
+    for (case, transport) in cases.iter().zip(&transports) {
+        let label = case.label();
+        if !case.assert_replication {
+            continue;
+        }
+        assert_eq!(transport.status, "COMPLETED", "{label}: existing-object replication must complete");
+        if case.is_multipart_layout() {
+            let expected: Vec<i32> = (1..=i32::try_from(case.part_sizes.len())?).collect();
+            assert_eq!(
+                transport.uploaded_parts, expected,
+                "{label}: stored parts must replicate as the same multipart layout"
+            );
+            // The current build can drive an existing object twice (two
+            // full CreateMultipartUpload/UploadPart/Complete rounds with
+            // distinct upload ids) while its status is still PENDING; the
+            // rc.5 baseline drives once. That is a scheduling difference,
+            // not a layout one, tracked as rustfs/backlog#2362.
+            assert!(transport.completes >= 1, "{label}: at least one CompleteMultipartUpload");
+            if transport.completes > 1 {
+                tracing::warn!(
+                    target: "e2e_test::upgrade_compatibility_test",
+                    object = %label,
+                    completes = transport.completes,
+                    journal = ?transport.journal,
+                    "existing-object replication drove the same object more than once (rustfs/backlog#2362)"
+                );
+            }
+            assert_eq!(
+                transport.single_puts, 0,
+                "{label}: a multipart layout must not go out as a single PutObject"
+            );
+        } else {
+            assert!(transport.single_puts >= 1, "{label}: a single PUT replicates as PutObject");
+            assert!(transport.uploaded_parts.is_empty(), "{label}: a single PUT must not go out as multipart");
+        }
+        if !case.ssec {
+            let replica = replica_client
+                .get_object()
+                .bucket(LAYOUT_REPLICA_BUCKET)
+                .key(case.key)
+                .send()
+                .await
+                .map_err(|err| format!("{label}: replica missing on the target: {err}"))?
+                .body
+                .collect()
+                .await?
+                .into_bytes();
+            assert_eq!(replica.len(), case.body.len(), "{label}: replica length");
+            assert!(replica == case.body, "{label}: replica body must equal the rc.5 upload");
+        }
+    }
+    Ok(())
+}
+
+/// The same layouts replicated by rc.5 itself, without an upgrade. This is the
+/// baseline that tells a pre-existing transport failure apart from one the
+/// current build introduced; it records the outcome per layout and only fails
+/// when the fixture cannot run.
+#[tokio::test]
+#[ignore = "requires the pinned 1.0.0-rc.5 release binary"]
+async fn rc5_baseline_replicates_multipart_layouts() -> TestResult {
+    init_logging();
+    let previous_binary = source_binary()?;
+    let server_env = layout_server_env();
+
+    let mut env = RustFSTestEnvironment::new().await?;
+    env.start_rustfs_server_from_binary(&previous_binary, vec![], &server_env)
+        .await?;
+    let client = env.create_s3_client();
+    env.create_test_bucket(LAYOUT_PLAIN_BUCKET).await?;
+    env.create_test_bucket(LAYOUT_ENCRYPTED_BUCKET).await?;
+    enable_versioning(&client, LAYOUT_PLAIN_BUCKET).await?;
+    enable_versioning(&client, LAYOUT_ENCRYPTED_BUCKET).await?;
+    put_default_sse_s3_encryption(&client, LAYOUT_ENCRYPTED_BUCKET).await?;
+
+    let mut cases = layout_cases();
+    for case in cases.iter_mut() {
+        layout_write(&client, case).await?;
+        let head = layout_head(&client, case).await?;
+        case.rc5_etag = head
+            .e_tag()
+            .ok_or_else(|| format!("{}: rc.5 HEAD omitted the ETag", case.label()))?
+            .trim_matches('"')
+            .to_string();
+    }
+    let (_target, transports) = replicate_layouts(&env, &client, &cases).await?;
+    let summary: Vec<String> = cases
+        .iter()
+        .zip(&transports)
+        .map(|(case, transport)| {
+            format!(
+                "{}: {} parts={:?} puts={}",
+                case.label(),
+                transport.status,
+                transport.uploaded_parts,
+                transport.single_puts
+            )
+        })
+        .collect();
+    tracing::info!(target: "e2e_test::upgrade_compatibility_test", ?summary, "rc.5 baseline replication outcomes");
+    Ok(())
+}

crates/ecstore/Cargo.toml

@@ -226,7 +226,7 @@ metrics = { workspace = true }
 # crates.io. The guard scripts/check_no_tokio_io_uring.sh allows an explicit
 # io-uring integration; only the tokio "io-uring" runtime feature is banned.
 [target.'cfg(target_os = "linux")'.dependencies]
-rustfs-uring = "0.2.1"
+rustfs-uring = "0.2.2"
 
 [target.'cfg(windows)'.dependencies]
 winapi-util.workspace = true

crates/ecstore/src/api/mod.rs

@@ -32,7 +32,7 @@ pub mod bucket {
     pub mod bucket_target_sys {
         pub use crate::bucket::bucket_target_sys::{
             AdvancedPutOptions, BucketTargetError, BucketTargetSys, PutObjectOptions, RemoveObjectOptions, S3ClientError,
-            SsecPassthroughCapability, TargetClient, append_version_id_query,
+            SsecPassthroughCapability, TargetClient, VersionIdentityCapability, append_version_id_query,
         };
     }
 
@@ -69,11 +69,33 @@ pub mod bucket {
             };
         }
 
+        pub mod recovery_control {
+            pub use crate::bucket::lifecycle::recovery_control::{
+                IlmRecoveryClassification, IlmRecoveryControlPage, IlmRecoveryControlView, IlmRecoveryProtocol,
+                inspect_recovery_control, list_recovery_controls,
+            };
+        }
+
+        pub mod recovery_disposition {
+            pub use crate::bucket::lifecycle::recovery_disposition::{
+                IlmRecoveryDispositionExecutionOutcome, IlmRecoveryDispositionReasonCode, IlmRecoveryDispositionState,
+                dry_run_recovery_disposition, execute_recovery_disposition,
+            };
+        }
+
+        pub mod recovery_export {
+            pub use crate::bucket::lifecycle::recovery_export::{
+                IlmRecoveryExportCreated, IlmRecoveryExportObservation, create_recovery_export,
+                inspect_recovery_export_observation, load_recovery_export,
+            };
+        }
+
         pub mod transition_transaction {
             pub use crate::bucket::lifecycle::transition_transaction::{
                 TransitionOperatorDeleteResult, TransitionOperatorError, TransitionOperatorProbe, TransitionOperatorStatus,
-                delete_transition_candidate_for_operator, finalize_missing_transition_transaction_for_operator,
-                inspect_transition_transaction_for_operator,
+                TransitionRecoveryRetryResult, TransitionRecoveryRetryStatus, delete_transition_candidate_for_operator,
+                finalize_missing_transition_transaction_for_operator, inspect_transition_recovery_retry_for_operator,
+                inspect_transition_transaction_for_operator, retry_transition_recovery_for_operator,
             };
             #[cfg(feature = "test-util")]
             pub use crate::bucket::lifecycle::transition_transaction::{
@@ -89,8 +111,9 @@ pub mod bucket {
         #[allow(clippy::module_inception)]
         pub mod lifecycle {
             pub use crate::bucket::lifecycle::lifecycle::{
-                Event, ExpirationOptions, IlmAction, Lifecycle, LifecycleCalculate, ObjectOpts, RuleValidate,
-                TRANSITION_COMPLETE, TRANSITION_PENDING, TransitionOptions, expected_expiry_time, object_opts_from_object_info,
+                Event, ExpirationOptions, IlmAction, LIFECYCLE_MALFORMED_XML_ERROR_KIND, Lifecycle, LifecycleCalculate,
+                ObjectOpts, RuleValidate, TRANSITION_COMPLETE, TRANSITION_PENDING, TransitionOptions, expected_expiry_time,
+                object_opts_from_object_info,
             };
         }
 
@@ -151,15 +174,17 @@ pub mod bucket {
             BUCKET_CONFIG_PUBLISH_HOOK, BucketConfigPublishHook, BucketMetadataMutationGuard, BucketMetadataSys,
             ObjectLockConfigState, acquire_bucket_metadata_transaction_lock,
             acquire_bucket_metadata_transaction_lock_for_incarnation, acquire_scanner_bucket_incarnation_fence,
-            capture_bucket_metadata_incarnation, delete, delete_if_incarnation, delete_under_transaction_lock, get,
-            get_accelerate_config, get_bucket_policy, get_bucket_policy_raw, get_bucket_targets_config, get_config_from_disk,
-            get_cors_config, get_durability_config, get_global_bucket_metadata_sys, get_lifecycle_config, get_logging_config,
-            get_notification_config, get_object_lock_config, get_object_lock_config_state, get_on_demand_migration_config,
+            capture_bucket_metadata_incarnation, delete, delete_if_incarnation, delete_if_incarnation_at,
+            delete_under_transaction_lock, get, get_accelerate_config, get_bucket_policy, get_bucket_policy_raw,
+            get_bucket_targets_config, get_config_from_disk, get_cors_config, get_durability_config,
+            get_global_bucket_metadata_sys, get_lifecycle_config, get_logging_config, get_notification_config,
+            get_object_lock_config, get_object_lock_config_state, get_on_demand_migration_config,
             get_on_demand_migration_config_in, get_public_access_block_config, get_quota_config, get_replication_config,
             get_request_payment_config, get_sse_config, get_tagging_config, get_versioning_config, get_website_config,
             init_bucket_metadata_sys, list_bucket_targets, reload_bucket_metadata, remove_bucket_metadata, set_bucket_metadata,
             update, update_bucket_targets_under_transaction_lock, update_config_with, update_if_incarnation,
-            update_quota_if_incarnation, update_under_transaction_lock,
+            update_if_incarnation_at, update_quota_if_incarnation, update_quota_if_incarnation_at, update_under_transaction_lock,
+            update_under_transaction_lock_at,
         };
         #[cfg(feature = "test-util")]
         pub use crate::bucket::metadata_sys::{ConfigWriteLockProbe, test_support};
@@ -283,7 +308,7 @@ pub mod cache {
 pub mod capacity {
     pub use crate::core::pools::{
         DecommissionUnresolvedEntry, PoolDecommissionInfo, PoolStatus, get_total_usable_capacity, get_total_usable_capacity_free,
-        path2_bucket_object, path2_bucket_object_with_base_path,
+        is_pool_activation_fleet_proof_error, path2_bucket_object, path2_bucket_object_with_base_path,
     };
     pub use crate::store::utils::is_reserved_or_invalid_bucket;
 }
@@ -436,9 +461,12 @@ pub mod notification {
     #[cfg(any(test, feature = "test-util"))]
     pub use crate::services::notification_sys::rotate_cross_pool_fence_fleet_proof_for_test;
     pub use crate::services::notification_sys::{
-        ClusterTierDailyStats, CrossPoolFenceFleetProofToken, LegacyTransitionStateReconcileFleetProofToken, NotificationPeerErr,
-        NotificationSys, ScannerPublicationLeaseGrant, acquire_cross_pool_fence_fleet_proof,
+        ClusterTierDailyStats, CrossPoolFenceFleetProofToken, IlmRecoveryExportFleetProofToken,
+        LegacyTransitionStateReconcileFleetProofToken, NotificationPeerErr, NotificationSys, ScannerPublicationLeaseGrant,
+        acquire_cross_pool_fence_fleet_proof, acquire_ilm_recovery_export_fleet_proof,
         acquire_legacy_transition_state_reconcile_fleet_proof, cross_pool_fence_fleet_proof_matches, get_global_notification_sys,
+        ilm_recovery_export_fleet_proof_matches, ilm_recovery_export_local_process_epoch,
+        ilm_recovery_export_member_epochs_sha256, ilm_recovery_export_topology_generation,
         legacy_transition_state_reconcile_fleet_proof_matches, new_global_notification_sys,
         scanner_peer_transport_error_message_is_retryable, start_remote_version_state_fleet_probe,
     };
@@ -491,7 +519,8 @@ pub mod rpc {
     pub use crate::cluster::rpc::{
         AuthenticatedChannel, KMS_SIGNAL_SUBSYSTEM, LocalPeerS3Client, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS,
         PeerRestClient, PeerS3Client, S3PeerSys, SERVICE_SIGNAL_REFRESH_CONFIG, SERVICE_SIGNAL_RELOAD_DYNAMIC,
-        ScannerBucketListing, ScannerPeerActivity, ScannerPeerDirtyUsageSnapshot, ScannerPublicationLease, TONIC_RPC_PREFIX,
+        ScannerBucketListing, ScannerDirtyUsageAcknowledgement, ScannerPeerActivity, ScannerPeerDirtyUsageBucket,
+        ScannerPeerDirtyUsageSnapshot, ScannerPublicationLease, ScannerScopedDirtyUsageAckEntry, TONIC_RPC_PREFIX,
         TonicInterceptor, build_put_file_auth_trailer, check_and_record_signed_rpc_nonce, decode_heal_bucket_rpc_options,
         encode_heal_bucket_rpc_options, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
         gen_tonic_signature_interceptor, node_service_time_out_client, node_service_time_out_client_no_auth,

crates/ecstore/src/bucket/bucket_target_sys.rs

@@ -18,7 +18,7 @@ use crate::bucket::metadata_sys::get_replication_config;
 use crate::bucket::remote_s3_client::{
     PathStyle, REPLICATION_TARGET_RETRY_POLICY, RemoteCredentials, RemoteS3EndpointSpec, build_remote_s3_client,
 };
-use crate::bucket::replication::{ObjectLockIntegrity, object_lock_put_integrity};
+use crate::bucket::replication::{ObjectLockIntegrity, object_lock_put_integrity, replication_etags_match};
 use crate::bucket::replication::{ReplicationStatusType, ReplicationTargetConfigBridge};
 use crate::bucket::target::ARN;
 use crate::bucket::target::BucketTargetType;
@@ -126,6 +126,22 @@ impl From<&BucketTarget> for RemoteS3EndpointSpec {
 }
 
 pub type HeadObjectSdkError = Box<SdkError<HeadObjectError>>;
+
+/// Whether an edited bucket target still addresses the same remote service
+/// (endpoint, bucket, path style, TLS and identity), so a verdict learned
+/// about that service stays valid across the edit.
+fn same_replication_service(edited: &BucketTarget, previous: &BucketTarget) -> bool {
+    let access_key = |target: &BucketTarget| target.credentials.as_ref().map(|credentials| credentials.access_key.clone());
+    edited.endpoint == previous.endpoint
+        && edited.target_bucket == previous.target_bucket
+        && edited.secure == previous.secure
+        && edited.path == previous.path
+        && access_key(edited) == access_key(previous)
+}
+
+/// Page size and page budget for [`TargetClient::find_version_by_etag`].
+const FIND_VERSION_BY_ETAG_PAGE_SIZE: i32 = 1000;
+const FIND_VERSION_BY_ETAG_MAX_PAGES: usize = 8;
 pub type GetObjectSdkError = Box<SdkError<GetObjectError>>;
 pub type GetObjectTaggingSdkError = Box<SdkError<GetObjectTaggingError>>;
 pub type PutObjectTaggingSdkError = Box<SdkError<PutObjectTaggingError>>;
@@ -349,6 +365,13 @@ struct TargetClientBuildProbe {
 /// their import path while the verdict vocabulary lives with the
 /// replication decision logic.
 pub use crate::bucket::replication::SsecPassthroughCapability;
+/// Version-identity verdicts (see the enum's own docs in
+/// `rustfs-replication`) are cached here per target ARN and follow the same
+/// `arn_remotes_map` lifecycle. They carry no TTL: the verdict is refreshed
+/// by every replication write's response, so it can only go stale on a
+/// target that receives no writes — and a stale `MintsOwn` costs one extra
+/// content-identity lookup before a PUT, never a lost replica.
+pub use crate::bucket::replication::VersionIdentityCapability;
 
 /// How long an audited SSE-C passthrough verdict stays authoritative.
 ///
@@ -375,6 +398,11 @@ pub struct BucketTargetSys {
     /// SSE-C passthrough capability verdicts keyed by target ARN. See
     /// [`SsecPassthroughCapability`]; reset alongside `arn_remotes_map`.
     ssec_passthrough_map: Arc<RwLock<HashMap<String, SsecPassthroughRecord>>>,
+    /// Version-identity verdicts keyed by target ARN. See
+    /// [`VersionIdentityCapability`]; reset alongside `arn_remotes_map`. A std
+    /// lock (never held across an await) so the replication worker can record
+    /// a verdict from inside its synchronous PUT-response audit.
+    version_identity_map: Arc<std::sync::RwLock<HashMap<String, VersionIdentityCapability>>>,
     pub targets_map: Arc<RwLock<HashMap<String, Vec<BucketTarget>>>>,
     /// Buckets whose persisted `bucket-targets.json` exists but cannot be
     /// decoded (rustfs/backlog#2282). Written under the bucket's update mutex
@@ -423,6 +451,7 @@ impl BucketTargetSys {
         Self {
             arn_remotes_map: Arc::new(RwLock::new(HashMap::new())),
             ssec_passthrough_map: Arc::new(RwLock::new(HashMap::new())),
+            version_identity_map: Arc::new(std::sync::RwLock::new(HashMap::new())),
             targets_map: Arc::new(RwLock::new(HashMap::new())),
             unreadable_targets: Arc::new(RwLock::new(HashSet::new())),
             h_mutex: Arc::new(RwLock::new(HashMap::new())),
@@ -746,10 +775,40 @@ impl BucketTargetSys {
                 arn_remotes_map.remove(&target.arn);
                 health_map.remove(&target.arn);
                 ssec_map.remove(&target.arn);
+                self.forget_version_identity_capability(&target.arn);
             }
         }
     }
 
+    /// Cached version-identity verdict for a target ARN; `Unknown` until a
+    /// replication write or a replication-check VersionFidelity probe judged
+    /// it since the target was built.
+    pub fn version_identity_capability(&self, arn: &str) -> VersionIdentityCapability {
+        self.version_identity_map
+            .read()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .get(arn)
+            .copied()
+            .unwrap_or_default()
+    }
+
+    /// Record a version-identity verdict for a target ARN. Written by the
+    /// replication worker after every PutObject / CompleteMultipartUpload
+    /// response and by the replication-check VersionFidelity phase.
+    pub fn record_version_identity_capability(&self, arn: &str, capability: VersionIdentityCapability) {
+        self.version_identity_map
+            .write()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .insert(arn.to_string(), capability);
+    }
+
+    fn forget_version_identity_capability(&self, arn: &str) {
+        self.version_identity_map
+            .write()
+            .unwrap_or_else(|poisoned| poisoned.into_inner())
+            .remove(arn);
+    }
+
     /// Cached SSE-C passthrough capability for a target ARN, plus whether the
     /// verdict is older than [`SSEC_PASSTHROUGH_CAPABILITY_TTL`]. `(Unknown,
     /// false)` when no verdict has been recorded since the target was built.
@@ -794,17 +853,24 @@ impl BucketTargetSys {
     ) -> Result<BucketTargets, BucketTargetError> {
         self.validate_target(bucket, target).await?;
 
-        let mut bucket_targets = match self.list_bucket_targets(bucket).await {
-            Ok(targets) => targets,
-            Err(BucketTargetError::BucketRemoteTargetNotFound { .. }) => BucketTargets::default(),
-            Err(err) => return Err(err),
-        };
+        let mut bucket_targets = self.targets_base_for_write(bucket).await?;
 
         Self::upsert_target_entry(&mut bucket_targets.targets, target, update)?;
 
         Ok(bucket_targets)
     }
 
+    /// Ordinary writes must not turn an unreadable cached snapshot into an
+    /// empty configuration. Explicit repair belongs to the metadata transaction
+    /// that can inspect the current persisted state.
+    async fn targets_base_for_write(&self, bucket: &str) -> Result<BucketTargets, BucketTargetError> {
+        match self.list_bucket_targets(bucket).await {
+            Ok(targets) => Ok(targets),
+            Err(BucketTargetError::BucketRemoteTargetNotFound { .. }) => Ok(BucketTargets::default()),
+            Err(err) => Err(err),
+        }
+    }
+
     pub async fn validate_target(&self, bucket: &str, target: &BucketTarget) -> Result<(), BucketTargetError> {
         if !target.target_type.is_valid() {
             return Err(BucketTargetError::BucketRemoteArnTypeInvalid {
@@ -863,7 +929,9 @@ impl BucketTargetSys {
         Ok(())
     }
 
-    fn upsert_target_entry(
+    /// Merge a validated target into a caller-owned snapshot. The caller must
+    /// protect that snapshot through persistence.
+    pub fn upsert_target_entry(
         bucket_targets: &mut Vec<BucketTarget>,
         target: &BucketTarget,
         update: bool,
@@ -1153,12 +1221,32 @@ impl BucketTargetSys {
         // Remove existing targets
         if let Some(existing_targets) = targets_map.remove(bucket) {
             let mut ssec_map = self.ssec_passthrough_map.write().await;
+            let unchanged_service: HashMap<&str, &BucketTarget> = targets
+                .map(|new_targets| {
+                    new_targets
+                        .targets
+                        .iter()
+                        .map(|target| (target.arn.as_str(), target))
+                        .collect()
+                })
+                .unwrap_or_default();
             for target in existing_targets {
                 arn_remotes_map.remove(&target.arn);
                 health_map.remove(&target.arn);
                 // A rebuilt/edited target may point at a different service:
                 // the SSE-C passthrough verdict must be re-audited from Unknown.
                 ssec_map.remove(&target.arn);
+                // The version-identity verdict survives an edit that keeps the
+                // same remote service (a resync start or a bandwidth change
+                // rewrites the entry in place): forgetting it there would make
+                // the very resync that follows re-drive every object as a
+                // duplicate on a target that mints its own version ids.
+                if unchanged_service
+                    .get(target.arn.as_str())
+                    .is_none_or(|edited| !same_replication_service(edited, &target))
+                {
+                    self.forget_version_identity_capability(&target.arn);
+                }
                 self.update_bandwidth_limit(bucket, &target.arn, 0);
             }
         }
@@ -1227,25 +1315,27 @@ impl BucketTargetSys {
             return (String::new(), false);
         };
 
-        {
-            let targets_map = self.targets_map.read().await;
-            if let Some(targets) = targets_map.get(bucket) {
-                for tgt in targets {
-                    if tgt.target_type == target.target_type
-                        && tgt.target_bucket == target.target_bucket
-                        && target.endpoint == tgt.endpoint
-                        && tgt
-                            .credentials
-                            .as_ref()
-                            .map(|c| {
-                                let default_creds = Credentials::default();
-                                c.access_key == target.credentials.as_ref().unwrap_or(&default_creds).access_key
-                            })
-                            .unwrap_or(false)
-                    {
-                        return (tgt.arn.clone(), true);
-                    }
-                }
+        let targets_map = self.targets_map.read().await;
+        let targets = targets_map.get(bucket).map(Vec::as_slice).unwrap_or_default();
+        Self::remote_arn_for_targets(targets, target, depl_id)
+    }
+
+    /// Resolve create idempotency against the snapshot the caller will persist.
+    pub fn remote_arn_for_targets(targets: &[BucketTarget], target: &BucketTarget, depl_id: &str) -> (String, bool) {
+        for tgt in targets {
+            if tgt.target_type == target.target_type
+                && tgt.target_bucket == target.target_bucket
+                && target.endpoint == tgt.endpoint
+                && tgt
+                    .credentials
+                    .as_ref()
+                    .map(|c| {
+                        let default_creds = Credentials::default();
+                        c.access_key == target.credentials.as_ref().unwrap_or(&default_creds).access_key
+                    })
+                    .unwrap_or(false)
+            {
+                return (tgt.arn.clone(), true);
             }
         }
 
@@ -1881,6 +1971,62 @@ impl TargetClient {
             .map_err(Box::new)
     }
 
+    /// Locate a replica by content identity on a target that mints its own
+    /// version ids: page `ListObjectVersions` under the exact key and return
+    /// the newest live version whose ETag matches `source_etag`. Delete
+    /// markers and prefix siblings never match. Bounded to
+    /// [`FIND_VERSION_BY_ETAG_MAX_PAGES`] pages so a key with a very deep
+    /// history cannot turn one convergence check into an unbounded scan; a
+    /// replica beyond that window reads as missing, which only costs a
+    /// re-PUT (today's behaviour), never a lost object.
+    pub async fn find_version_by_etag(
+        &self,
+        bucket: &str,
+        object: &str,
+        source_etag: &str,
+    ) -> Result<Option<String>, Box<SdkError<aws_sdk_s3::operation::list_object_versions::ListObjectVersionsError>>> {
+        let mut key_marker: Option<String> = None;
+        let mut version_id_marker: Option<String> = None;
+        for _ in 0..FIND_VERSION_BY_ETAG_MAX_PAGES {
+            let page = self
+                .client
+                .list_object_versions()
+                .bucket(bucket)
+                .prefix(object)
+                .max_keys(FIND_VERSION_BY_ETAG_PAGE_SIZE)
+                .set_key_marker(key_marker.take())
+                .set_version_id_marker(version_id_marker.take())
+                .send()
+                .await
+                .map_err(Box::new)?;
+            if let Some(version) = page.versions().iter().find(|version| {
+                version.key() == Some(object)
+                    && version.version_id().is_some_and(|id| !id.is_empty())
+                    && replication_etags_match(Some(source_etag), version.e_tag())
+            }) {
+                return Ok(version.version_id().map(str::to_string));
+            }
+            // Every listed key is >= the prefix; once the listing moved past
+            // the exact key there is nothing left to find.
+            if page
+                .versions()
+                .iter()
+                .any(|version| version.key().is_some_and(|key| key > object))
+            {
+                return Ok(None);
+            }
+            if !page.is_truncated().unwrap_or(false) {
+                return Ok(None);
+            }
+            key_marker = page.next_key_marker().map(str::to_string);
+            version_id_marker = page.next_version_id_marker().map(str::to_string);
+            if key_marker.is_none() {
+                return Ok(None);
+            }
+        }
+        Ok(None)
+    }
+
     /// HEAD used by the read-proxy path (GET/HEAD of an object not yet
     /// replicated locally, MinIO `proxyHeadToRepTarget`).
     ///
@@ -2053,7 +2199,15 @@ impl TargetClient {
             }
         }
 
-        match builder
+        // A forwarded source checksum is this PUT's integrity header. In
+        // streaming-checksum mode (`RUSTFS_REPLICATION_STREAMING_CHECKSUMS`)
+        // the SDK would still add its default CRC32 trailer, and a target that
+        // receives both keeps the trailer's algorithm: a forwarded SHA256
+        // vanished from the replica while the source reported COMPLETED. Pin
+        // this request to WhenRequired so nothing is sent beside the source's
+        // own checksum.
+        let forwards_source_checksum = headers.keys().any(|name| name.as_str().starts_with("x-amz-checksum-"));
+        let mut operation = builder
             .bucket(bucket)
             .key(object)
             .content_length(size)
@@ -2073,10 +2227,14 @@ impl TargetClient {
                 }
 
                 Result::<_, aws_smithy_types::error::operation::BuildError>::Ok(req)
-            })
-            .send()
-            .await
-        {
+            });
+        if forwards_source_checksum {
+            operation = operation.config_override(
+                aws_sdk_s3::config::Builder::new()
+                    .request_checksum_calculation(aws_sdk_s3::config::RequestChecksumCalculation::WhenRequired),
+            );
+        }
+        match operation.send().await {
             Ok(output) => {
                 // Under SSE-KMS/DSSE or SSE-C the target's ETag is not the MD5
                 // of the stored plaintext, so it cannot be compared against the
@@ -2496,13 +2654,21 @@ mod tests {
     }
 
     fn header_recording_target_client(response_headers: Vec<(String, String)>) -> (TargetClient, RecordedHeaders) {
+        header_recording_target_client_with_checksums(response_headers, replication_request_checksum_calculation())
+    }
+
+    fn header_recording_target_client_with_checksums(
+        response_headers: Vec<(String, String)>,
+        checksums: RequestChecksumCalculation,
+    ) -> (TargetClient, RecordedHeaders) {
         let request_headers: RecordedHeaders = Arc::new(std::sync::Mutex::new(Vec::new()));
         let connector = SharedHttpConnector::new(RecordingHeaderConnector {
             request_headers: Arc::clone(&request_headers),
             response_headers,
         });
         let http_client = http_client_fn(move |_settings, _components| connector.clone());
-        let client = s3_client_for_test(443, Some(http_client));
+        let client =
+            s3_client_for_endpoint_test_with_checksums("https://localhost:443".to_string(), Some(http_client), checksums);
         (
             TargetClient {
                 endpoint: "https://localhost:443".to_string(),
@@ -2669,6 +2835,47 @@ mod tests {
         }
     }
 
+    /// With streaming checksums enabled the SDK adds a CRC32 trailer to every
+    /// upload. A PUT that forwards the source's checksum must not get that
+    /// second algorithm: a target that receives both keeps the trailer's and
+    /// the forwarded SHA256 never reaches the replica (rustfs/backlog#2340).
+    #[tokio::test]
+    async fn streaming_put_object_with_forwarded_checksum_sends_no_sdk_checksum() {
+        let (client, recorded) =
+            header_recording_target_client_with_checksums(Vec::new(), RequestChecksumCalculation::WhenSupported);
+        let mut forwarded = PutObjectOptions::default();
+        forwarded.user_metadata.insert(
+            "x-amz-checksum-sha256".to_string(),
+            "OoJ3yNhRwv3wwtZoGqEIPrPX9xwTnfLl+ka0wStN1g0=".to_string(),
+        );
+        client
+            .put_object("target-bucket", "object", 4, streaming_test_body(b"data"), &forwarded)
+            .await
+            .expect("recorded put_object should succeed");
+        client
+            .put_object("target-bucket", "object", 4, streaming_test_body(b"data"), &PutObjectOptions::default())
+            .await
+            .expect("recorded put_object should succeed");
+        let recorded = recorded.lock().expect("recorded header lock should not be poisoned");
+        let with_forwarded = &recorded[0];
+        assert_eq!(
+            recorded_header(with_forwarded, "x-amz-checksum-sha256"),
+            Some("OoJ3yNhRwv3wwtZoGqEIPrPX9xwTnfLl+ka0wStN1g0=")
+        );
+        assert_eq!(
+            recorded_header(with_forwarded, "x-amz-trailer"),
+            None,
+            "the SDK must not add a trailer checksum"
+        );
+        assert_eq!(recorded_header(with_forwarded, "x-amz-sdk-checksum-algorithm"), None);
+        // Control: the same client still streams a trailer when nothing is forwarded.
+        let without_forwarded = &recorded[1];
+        assert!(
+            recorded_header(without_forwarded, "x-amz-trailer").is_some(),
+            "streaming mode must still apply to uploads without a forwarded checksum: {without_forwarded:?}"
+        );
+    }
+
     /// A forwarded source checksum already satisfies the rule; nothing is added.
     #[tokio::test]
     async fn locked_put_object_keeps_a_forwarded_source_checksum() {
@@ -3034,6 +3241,14 @@ mod tests {
     }
 
     fn s3_client_for_endpoint_test(endpoint: String, http_client: Option<SharedHttpClient>) -> S3Client {
+        s3_client_for_endpoint_test_with_checksums(endpoint, http_client, replication_request_checksum_calculation())
+    }
+
+    fn s3_client_for_endpoint_test_with_checksums(
+        endpoint: String,
+        http_client: Option<SharedHttpClient>,
+        checksums: RequestChecksumCalculation,
+    ) -> S3Client {
         let credentials = SdkCredentials::builder()
             .access_key_id("test-access")
             .secret_access_key("test-secret")
@@ -3047,7 +3262,7 @@ mod tests {
             .behavior_version(aws_sdk_s3::config::BehaviorVersion::latest())
             // Mirror the production remote-target builder so recorded requests
             // exercise the same checksum/framing behavior (#6853).
-            .request_checksum_calculation(replication_request_checksum_calculation());
+            .request_checksum_calculation(checksums);
         if let Some(http_client) = http_client {
             config = config.http_client(http_client);
         }
@@ -3210,6 +3425,64 @@ mod tests {
         assert!(message.contains("connection refused"));
     }
 
+    #[test]
+    fn same_replication_service_ignores_resync_and_bandwidth_edits() {
+        let base = BucketTarget {
+            endpoint: "target.example:9000".to_string(),
+            target_bucket: "replica".to_string(),
+            secure: true,
+            path: "on".to_string(),
+            arn: "arn:rustfs:replication:us-east-1:bucket:same".to_string(),
+            credentials: Some(Credentials {
+                access_key: "access".to_string(),
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let resync_edit = BucketTarget {
+            reset_id: "reset-1".to_string(),
+            bandwidth_limit: 1024,
+            ..base.clone()
+        };
+        assert!(same_replication_service(&resync_edit, &base));
+        for moved in [
+            BucketTarget {
+                endpoint: "other.example:9000".to_string(),
+                ..base.clone()
+            },
+            BucketTarget {
+                target_bucket: "other".to_string(),
+                ..base.clone()
+            },
+            BucketTarget {
+                secure: false,
+                ..base.clone()
+            },
+            BucketTarget {
+                credentials: Some(Credentials {
+                    access_key: "rotated".to_string(),
+                    ..Default::default()
+                }),
+                ..base.clone()
+            },
+        ] {
+            assert!(!same_replication_service(&moved, &base));
+        }
+    }
+
+    #[test]
+    fn version_identity_verdict_is_per_arn_and_forgotten_with_the_target() {
+        let sys = BucketTargetSys::default();
+        let arn = "arn:rustfs:replication:us-east-1:bucket:identity";
+        assert_eq!(sys.version_identity_capability(arn), VersionIdentityCapability::Unknown);
+        sys.record_version_identity_capability(arn, VersionIdentityCapability::MintsOwn);
+        assert_eq!(sys.version_identity_capability(arn), VersionIdentityCapability::MintsOwn);
+        assert_eq!(sys.version_identity_capability("other"), VersionIdentityCapability::Unknown);
+        // A rebuilt target may point at a different service.
+        sys.forget_version_identity_capability(arn);
+        assert_eq!(sys.version_identity_capability(arn), VersionIdentityCapability::Unknown);
+    }
+
     #[test]
     fn endpoint_health_key_preserves_explicit_port() {
         let url = Url::parse("https://remote.example:9443").expect("url should parse");
@@ -4313,4 +4586,41 @@ mod tests {
         let window = LastMinuteLatency::new();
         assert_eq!(window.get_total().avg, Duration::from_secs(0));
     }
+
+    fn repair_target(bucket: &str, id: &str) -> BucketTarget {
+        BucketTarget {
+            source_bucket: bucket.to_string(),
+            endpoint: "remote.example.com".to_string(),
+            target_bucket: "remote".to_string(),
+            arn: format!("arn:rustfs:replication:us-east-1:{bucket}:{id}"),
+            target_type: BucketTargetType::ReplicationService,
+            region: "us-east-1".to_string(),
+            ..Default::default()
+        }
+    }
+
+    #[tokio::test]
+    async fn an_unreadable_target_set_refuses_cached_writes() {
+        let sys = BucketTargetSys::default();
+        let bucket = "targets-repair-opt-in";
+        sys.mark_targets_unreadable(bucket).await;
+        assert!(matches!(
+            sys.targets_base_for_write(bucket).await,
+            Err(BucketTargetError::BucketRemoteTargetsUnreadable { .. })
+        ));
+    }
+
+    #[tokio::test]
+    async fn a_readable_target_set_remains_the_write_base() {
+        let sys = BucketTargetSys::default();
+        let bucket = "targets-repair-readable";
+        let existing = repair_target(bucket, "keep");
+        sys.targets_map
+            .write()
+            .await
+            .insert(bucket.to_string(), vec![existing.clone()]);
+        let base = sys.targets_base_for_write(bucket).await.expect("read targets");
+        assert_eq!(base.targets.len(), 1);
+        assert_eq!(base.targets[0].arn, existing.arn);
+    }
 }

crates/ecstore/src/bucket/lifecycle/bucket_lifecycle_ops.rs

@@ -32,6 +32,7 @@ use crate::bucket::lifecycle::manual_transition_job::{
     record_manual_transition_worker_result_with_reason, renew_manual_transition_job_lease_if_owned,
     save_manual_transition_job_record_if_current, save_manual_transition_task_if_absent, update_manual_transition_job_record,
 };
+use crate::bucket::lifecycle::recovery_disposition_runtime::run_recovery_disposition_maintenance_loop;
 use crate::bucket::lifecycle::replication_sink;
 use crate::bucket::lifecycle::replication_sink::{
     DeleteReplicationConfigSnapshot, ReplicationObjectBridge, ReplicationStatusType, replication_state_to_filemeta,
@@ -149,6 +150,19 @@ pub type ExpiryOpType = Box<dyn ExpiryOp + Send + Sync + 'static>;
 static XXHASH_SEED: u64 = 0;
 static TIER_FREE_VERSION_RECOVERY_STARTED: OnceLock<()> = OnceLock::new();
 static MANUAL_TRANSITION_JOB_RECOVERY_STARTED: OnceLock<()> = OnceLock::new();
+static RECOVERY_DISPOSITION_MAINTENANCE_STARTED: OnceLock<()> = OnceLock::new();
+
+#[cfg(test)]
+#[derive(Default)]
+struct FreeVersionPostRemoteDeleteTestBarrier {
+    arrived: Notify,
+    release: Notify,
+}
+
+#[cfg(test)]
+tokio::task_local! {
+    static FREE_VERSION_POST_REMOTE_DELETE_TEST_BARRIER: Arc<FreeVersionPostRemoteDeleteTestBarrier>;
+}
 
 pub const AMZ_OBJECT_TAGGING: &str = "X-Amz-Tagging";
 #[allow(
@@ -910,6 +924,11 @@ async fn cleanup_free_version_exact(api: Arc<ECStore>, oi: &ObjectInfo, cancel:
             })??;
         }
     }
+    #[cfg(test)]
+    if let Ok(barrier) = FREE_VERSION_POST_REMOTE_DELETE_TEST_BARRIER.try_with(Arc::clone) {
+        barrier.arrived.notify_one();
+        barrier.release.notified().await;
+    }
     if !free_version_cleanup_fences_current(&topology_generation, &api, &bucket_guard, &object_guards, &lease, cancel, deadline) {
         // Remote DELETE is idempotent, but a changed fence makes the local
         // outcome ambiguous. Keep every marker for a fully fenced retry.
@@ -2381,9 +2400,20 @@ pub async fn init_background_expiry(api: Arc<ECStore>) {
     let _ = spawn_tier_free_version_recovery_once(api.clone(), &TIER_FREE_VERSION_RECOVERY_STARTED);
     spawn_tier_delete_journal_recovery_once(api.clone());
     spawn_transition_transaction_recovery_once(api.clone());
+    spawn_recovery_disposition_maintenance_once(api.clone());
     spawn_manual_transition_job_recovery_once(api);
 }
 
+fn spawn_recovery_disposition_maintenance_once(api: Arc<ECStore>) -> Option<JoinHandle<()>> {
+    let cancel_token = api.ctx.background_cancel_token()?;
+    if RECOVERY_DISPOSITION_MAINTENANCE_STARTED.set(()).is_err() {
+        return None;
+    }
+    Some(tokio::spawn(async move {
+        run_recovery_disposition_maintenance_loop(api, cancel_token).await;
+    }))
+}
+
 fn spawn_manual_transition_job_recovery_once(api: Arc<ECStore>) -> Option<JoinHandle<()>> {
     if MANUAL_TRANSITION_JOB_RECOVERY_STARTED.set(()).is_err() {
         return None;
@@ -5831,7 +5861,7 @@ mod tests {
     #[cfg(feature = "test-util")]
     use crate::services::tier::test_util::register_mock_tier;
     #[cfg(feature = "test-util")]
-    use crate::services::tier::tier::TierConfigMgr;
+    use crate::services::tier::tier::{TIER_DRIVER_TEST_FACTORY, TierConfigMgr, TierDriverTestFactory};
     #[cfg(feature = "test-util")]
     use crate::services::tier::warm_backend::{TransitionCandidateProbe, WarmBackend as _};
     use crate::set_disk::{MultipartCommitBarrier, MultipartCommitPause};
@@ -7830,6 +7860,119 @@ mod tests {
         }
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial]
+    async fn tier_remove_waits_for_inflight_free_version_local_commit() {
+        let (disk_paths, ecstore) = setup_test_env().await;
+        let bucket = format!("tier-remove-free-version-{}", Uuid::new_v4());
+        let object = "free-version";
+        create_test_bucket(&ecstore, &bucket).await;
+        let (backend, identity_hex) = register_recovery_mock_tier(&ecstore).await;
+        let tier_manager = ecstore.tier_config_mgr();
+        {
+            let manager = tier_manager.read().await;
+            manager
+                .save_tiering_config(Arc::clone(&ecstore))
+                .await
+                .expect("mock tier configuration should persist before removal");
+        }
+        seed_recoverable_free_version(&disk_paths, &bucket, object, None, Some(identity_hex)).await;
+        let page = list_tier_free_versions(Arc::clone(&ecstore), 1, None, None, CancellationToken::new())
+            .await
+            .expect("seeded free version should be listed");
+        let oi = page
+            .items
+            .into_iter()
+            .next()
+            .expect("seeded free version should be recoverable");
+
+        backend
+            .set_put_remote_version(Some(oi.transitioned_object.version_id.clone()))
+            .await;
+        let seed_lease = TierConfigMgr::acquire_operation_lease(&tier_manager, "WARM")
+            .await
+            .expect("mock tier lease should be available");
+        seed_lease
+            .put(&oi.transitioned_object.name, ReaderImpl::Body(Bytes::from_static(b"body")), 4)
+            .await
+            .expect("remote free-version tuple should be seeded");
+        drop(seed_lease);
+
+        let barrier = Arc::new(super::FreeVersionPostRemoteDeleteTestBarrier::default());
+        let cleanup_barrier = Arc::clone(&barrier);
+        let cleanup_store = Arc::clone(&ecstore);
+        let cleanup_oi = oi.clone();
+        let cleanup = tokio::spawn(async move {
+            super::FREE_VERSION_POST_REMOTE_DELETE_TEST_BARRIER
+                .scope(cleanup_barrier, async move {
+                    super::cleanup_free_version_exact(cleanup_store, &cleanup_oi, &CancellationToken::new()).await
+                })
+                .await
+        });
+        tokio::time::timeout(StdDuration::from_secs(30), barrier.arrived.notified())
+            .await
+            .expect("free-version cleanup should pause after the remote delete");
+        assert!(!backend.contains(&oi.transitioned_object.name).await);
+
+        let remove_manager = Arc::clone(&tier_manager);
+        let remove_store = Arc::clone(&ecstore);
+        let remove_backend = backend.clone();
+        let remove_driver_factory: TierDriverTestFactory = Arc::new(move |_| Ok(Box::new(remove_backend.clone())));
+        let mut remove = tokio::spawn(async move {
+            TIER_DRIVER_TEST_FACTORY
+                .scope(
+                    remove_driver_factory,
+                    TierConfigMgr::remove_and_save(&remove_manager, remove_store, "WARM", true),
+                )
+                .await
+        });
+        let prepared = tokio::time::timeout(StdDuration::from_secs(30), async {
+            loop {
+                match TierConfigMgr::acquire_operation_lease(&tier_manager, "WARM").await {
+                    Ok(lease) => drop(lease),
+                    Err(err) if TierConfigMgr::operation_lease_blocked_by_mutation(&err) => break,
+                    Err(err) => panic!("tier remove should only block new operations while cleanup is paused: {err}"),
+                }
+                tokio::task::yield_now().await;
+            }
+        });
+        tokio::select! {
+            prepared = prepared => {
+                prepared.expect("tier remove should install its prepared admission fence");
+            }
+            result = &mut remove => {
+                panic!("tier remove finished before installing its prepared admission fence: {result:?}");
+            }
+        }
+        assert!(!remove.is_finished(), "tier remove must wait for the leased local cleanup commit");
+
+        barrier.release.notify_one();
+        tokio::time::timeout(StdDuration::from_secs(30), cleanup)
+            .await
+            .expect("free-version cleanup should finish after release")
+            .expect("free-version cleanup task should join")
+            .expect("free-version cleanup should keep its generation current");
+        tokio::time::timeout(StdDuration::from_secs(30), remove)
+            .await
+            .expect("tier remove should finish after local cleanup")
+            .expect("tier remove task should join")
+            .expect("tier remove should pass its fresh authoritative proof");
+
+        assert!(!tier_manager.read().await.is_tier_valid("WARM"));
+        for disk_path in &disk_paths {
+            assert!(
+                !fs::try_exists(disk_path.join(&bucket).join(object))
+                    .await
+                    .expect("post-removal free-version path check should succeed")
+            );
+        }
+        ecstore
+            .delete_bucket(&bucket, &DeleteBucketOptions::default())
+            .await
+            .expect("empty free-version test bucket should be removed");
+    }
+
     #[cfg(feature = "test-util")]
     #[tokio::test]
     #[serial]

crates/ecstore/src/bucket/lifecycle/config_boundary.rs

@@ -41,6 +41,21 @@ where
     com::read_config(api, file).await
 }
 
+pub(crate) async fn read_config_limited_preserve_empty<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<Vec<u8>>
+where
+    S: ObjectIO<
+            Error = Error,
+            RangeSpec = HTTPRangeSpec,
+            HeaderMap = HeaderMap,
+            ObjectOptions = ObjectOptions,
+            ObjectInfo = ObjectInfo,
+            GetObjectReader = GetObjectReader,
+            PutObjectReader = PutObjReader,
+        >,
+{
+    com::read_config_limited_preserve_empty(api, file, max_bytes).await
+}
+
 pub(crate) async fn read_config_with_metadata<S>(api: Arc<S>, file: &str, opts: &ObjectOptions) -> Result<(Vec<u8>, ObjectInfo)>
 where
     S: ObjectIO<
@@ -56,6 +71,26 @@ where
     com::read_config_with_metadata(api, file, opts).await
 }
 
+pub(crate) async fn read_config_limited_preserve_empty_with_metadata<S>(
+    api: Arc<S>,
+    file: &str,
+    opts: &ObjectOptions,
+    max_bytes: usize,
+) -> Result<(Vec<u8>, ObjectInfo)>
+where
+    S: ObjectIO<
+            Error = Error,
+            RangeSpec = HTTPRangeSpec,
+            HeaderMap = HeaderMap,
+            ObjectOptions = ObjectOptions,
+            ObjectInfo = ObjectInfo,
+            GetObjectReader = GetObjectReader,
+            PutObjectReader = PutObjReader,
+        >,
+{
+    com::read_config_limited_preserve_empty_with_metadata_opts(api, file, opts, max_bytes).await
+}
+
 pub(crate) async fn save_config<S>(api: Arc<S>, file: &str, data: Vec<u8>) -> Result<()>
 where
     S: ObjectIO<
@@ -126,20 +161,30 @@ where
             DeletedObject = DeletedObject,
         >,
 {
-    match api
-        .delete_object(
-            RUSTFS_META_BUCKET,
-            file,
-            ObjectOptions {
-                http_preconditions: Some(HTTPPreconditions {
-                    if_match: Some(etag.to_string()),
-                    ..Default::default()
-                }),
-                ..Default::default()
-            },
-        )
-        .await
-    {
+    delete_config_if_match_with_opts(api, file, etag, ObjectOptions::default()).await
+}
+
+pub(crate) async fn delete_config_if_match_with_opts<S>(
+    api: Arc<S>,
+    file: &str,
+    etag: &str,
+    mut options: ObjectOptions,
+) -> Result<()>
+where
+    S: ObjectOperations<
+            Error = Error,
+            ObjectInfo = ObjectInfo,
+            ObjectOptions = ObjectOptions,
+            FileInfo = FileInfo,
+            ObjectToDelete = ObjectToDelete,
+            DeletedObject = DeletedObject,
+        >,
+{
+    options.http_preconditions = Some(HTTPPreconditions {
+        if_match: Some(etag.to_string()),
+        ..Default::default()
+    });
+    match api.delete_object(RUSTFS_META_BUCKET, file, options).await {
         Ok(_) => Ok(()),
         Err(err) => {
             if err == Error::FileNotFound || matches!(err, Error::ObjectNotFound(_, _)) {

crates/ecstore/src/bucket/lifecycle/core.rs

@@ -15,9 +15,9 @@
 use crate::object_api::ObjectInfo;
 
 pub use rustfs_lifecycle::{
-    Event, ExpirationOptions, IlmAction, Lifecycle, LifecycleCalculate, ObjectOpts, RuleValidate, TRANSITION_COMPLETE,
-    TRANSITION_PENDING, TransitionOptions, abort_incomplete_multipart_upload_due, expected_expiry_time,
-    expiration_action_has_valid_target,
+    Event, ExpirationOptions, IlmAction, LIFECYCLE_MALFORMED_XML_ERROR_KIND, Lifecycle, LifecycleCalculate, ObjectOpts,
+    RuleValidate, TRANSITION_COMPLETE, TRANSITION_PENDING, TransitionOptions, abort_incomplete_multipart_upload_due,
+    expected_expiry_time, expiration_action_has_valid_target,
 };
 
 pub fn object_opts_from_object_info(oi: &ObjectInfo) -> ObjectOpts {

crates/ecstore/src/bucket/lifecycle/durable_namespace.rs

@@ -22,7 +22,7 @@ use super::{
     bucket_lifecycle_ops::{
         ManualTransitionQueueSnapshot, ManualTransitionRunReport, decode_manual_transition_continuation_token,
     },
-    manual_transition_job, tier_delete_journal, transition_transaction,
+    manual_transition_job, recovery_control, recovery_disposition, recovery_export, tier_delete_journal, transition_transaction,
 };
 use crate::error::{Error, Result};
 use crate::services::tier::tier_probe_intent;
@@ -41,6 +41,9 @@ pub(crate) enum DurableIlmRecordKind {
     ManualTransitionScope,
     ManualTransitionTask,
     ManualTransitionWorkerResult,
+    RecoveryControl,
+    RecoveryExport,
+    RecoveryDisposition,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
@@ -105,8 +108,26 @@ pub(crate) const MANUAL_TRANSITION_WORKER_RESULT_NAMESPACE: DurableIlmNamespace
     max_record_size: manual_transition_job::MAX_MANUAL_TRANSITION_WORKER_RESULT_RECORD_SIZE,
     kind: DurableIlmRecordKind::ManualTransitionWorkerResult,
 };
+pub(crate) const RECOVERY_CONTROL_NAMESPACE: DurableIlmNamespace = DurableIlmNamespace {
+    name: "recovery-control",
+    prefix: recovery_control::ILM_RECOVERY_CONTROL_PREFIX,
+    max_record_size: recovery_control::MAX_ILM_RECOVERY_CONTROL_SIZE,
+    kind: DurableIlmRecordKind::RecoveryControl,
+};
+pub(crate) const RECOVERY_EXPORT_NAMESPACE: DurableIlmNamespace = DurableIlmNamespace {
+    name: "recovery-export",
+    prefix: recovery_export::ILM_RECOVERY_EXPORT_PREFIX,
+    max_record_size: recovery_export::MAX_ILM_RECOVERY_EXPORT_SIZE,
+    kind: DurableIlmRecordKind::RecoveryExport,
+};
+pub(crate) const RECOVERY_DISPOSITION_NAMESPACE: DurableIlmNamespace = DurableIlmNamespace {
+    name: "recovery-disposition",
+    prefix: recovery_disposition::ILM_RECOVERY_DISPOSITION_PREFIX,
+    max_record_size: recovery_disposition::MAX_ILM_RECOVERY_DISPOSITION_SIZE,
+    kind: DurableIlmRecordKind::RecoveryDisposition,
+};
 
-pub(crate) const DURABLE_ILM_NAMESPACES: [DurableIlmNamespace; 9] = [
+pub(crate) const DURABLE_ILM_NAMESPACES: [DurableIlmNamespace; 12] = [
     TIER_DELETE_JOURNAL_NAMESPACE,
     TIER_DELETE_JOURNAL_V6_NAMESPACE,
     TIER_DELETE_DISPATCH_MANIFEST_NAMESPACE,
@@ -116,6 +137,9 @@ pub(crate) const DURABLE_ILM_NAMESPACES: [DurableIlmNamespace; 9] = [
     MANUAL_TRANSITION_SCOPE_NAMESPACE,
     MANUAL_TRANSITION_TASK_NAMESPACE,
     MANUAL_TRANSITION_WORKER_RESULT_NAMESPACE,
+    RECOVERY_CONTROL_NAMESPACE,
+    RECOVERY_EXPORT_NAMESPACE,
+    RECOVERY_DISPOSITION_NAMESPACE,
 ];
 
 #[derive(Debug, Clone, PartialEq, Eq)]
@@ -241,6 +265,40 @@ pub(crate) enum DurableIlmRecordCheckpoint {
     ManualTransitionWorkerResult {
         content_sha256: String,
     },
+    RecoveryControl {
+        content_sha256: String,
+        identity_sha256: String,
+        source_generation_sha256: String,
+        first_seen_at_unix_nanos: i64,
+        revision: u64,
+        classification: recovery_control::IlmRecoveryClassification,
+        attempt_count: u64,
+        consecutive_failure_count: u32,
+        #[serde(default, skip_serializing_if = "Option::is_none")]
+        owner_fence_sha256: Option<String>,
+    },
+    RecoveryExport {
+        content_sha256: String,
+        source_generation_sha256: String,
+        topology_generation: String,
+        member_epochs_sha256: String,
+        creator_sha256: String,
+        retain_until_unix_nanos: i64,
+    },
+    RecoveryDisposition {
+        content_sha256: String,
+        identity_sha256: String,
+        copy_manifest_sha256: String,
+        copy_manifest_count: usize,
+        created_at_unix_nanos: i64,
+        revision: u64,
+        state: recovery_disposition::IlmRecoveryDispositionState,
+        owner_fence_sha256: Option<String>,
+        owner_lease_acquired_at_unix_nanos: Option<i64>,
+        owner_lease_expires_at_unix_nanos: Option<i64>,
+        confirmed_absent_sha256: Vec<String>,
+        retain_until_unix_nanos: i64,
+    },
 }
 
 impl DurableIlmRecordCheckpoint {
@@ -254,7 +312,10 @@ impl DurableIlmRecordCheckpoint {
             | Self::ManualTransitionJob { content_sha256, .. }
             | Self::ManualTransitionScope { content_sha256, .. }
             | Self::ManualTransitionTask { content_sha256 }
-            | Self::ManualTransitionWorkerResult { content_sha256 } => content_sha256,
+            | Self::ManualTransitionWorkerResult { content_sha256 }
+            | Self::RecoveryControl { content_sha256, .. }
+            | Self::RecoveryExport { content_sha256, .. }
+            | Self::RecoveryDisposition { content_sha256, .. } => content_sha256,
         }
     }
 
@@ -295,6 +356,9 @@ impl DurableIlmRecordCheckpoint {
             {
                 return Err(Error::other("durable ILM tier delete journal checkpoint is invalid"));
             }
+            if !recovery_disposition_checkpoint_is_valid(checkpoint) {
+                return Err(Error::other("durable ILM recovery disposition checkpoint is invalid"));
+            }
         }
         if self == next {
             if let Self::ManualTransitionJob {
@@ -435,9 +499,7 @@ impl DurableIlmRecordCheckpoint {
                 },
             ) => {
                 previous_identity == next_identity
-                    && transition_state_distance(*previous_state, *next_state)
-                        .and_then(|distance| previous_revision.checked_add(distance))
-                        .is_some_and(|expected_revision| *next_revision == expected_revision)
+                    && transition_state_revision_is_successor(*previous_state, *previous_revision, *next_state, *next_revision)
                     && (!previous_remote_version_known || previous_remote_version == next_remote_version)
             }
             (
@@ -528,6 +590,128 @@ impl DurableIlmRecordCheckpoint {
                     ..
                 },
             ) => previous_identity == next_identity && next_updated_at > previous_updated_at,
+            (
+                Self::RecoveryControl {
+                    identity_sha256: previous_identity,
+                    source_generation_sha256: previous_generation,
+                    first_seen_at_unix_nanos: previous_first_seen,
+                    revision: previous_revision,
+                    classification: previous_classification,
+                    attempt_count: previous_attempts,
+                    consecutive_failure_count: previous_failures,
+                    owner_fence_sha256: previous_owner,
+                    ..
+                },
+                Self::RecoveryControl {
+                    identity_sha256: next_identity,
+                    source_generation_sha256: next_generation,
+                    first_seen_at_unix_nanos: next_first_seen,
+                    revision: next_revision,
+                    classification: next_classification,
+                    attempt_count: next_attempts,
+                    consecutive_failure_count: next_failures,
+                    owner_fence_sha256: next_owner,
+                    ..
+                },
+            ) => {
+                let adjacent = previous_identity == next_identity
+                    && previous_first_seen == next_first_seen
+                    && previous_revision.checked_add(1) == Some(*next_revision);
+                let claim = next_owner.is_some()
+                    && *previous_classification == recovery_control::IlmRecoveryClassification::Retrying
+                    && *next_classification == recovery_control::IlmRecoveryClassification::Retrying
+                    && previous_attempts.checked_add(1) == Some(*next_attempts)
+                    && previous_failures == next_failures;
+                let source_refresh = previous_owner.is_some()
+                    && previous_owner == next_owner
+                    && *previous_classification == recovery_control::IlmRecoveryClassification::Retrying
+                    && *next_classification == recovery_control::IlmRecoveryClassification::Retrying
+                    && previous_attempts == next_attempts
+                    && previous_failures == next_failures
+                    && previous_generation != next_generation;
+                let completion = previous_owner.is_some()
+                    && next_owner.is_none()
+                    && previous_generation == next_generation
+                    && previous_attempts == next_attempts;
+                adjacent && (claim || source_refresh || completion)
+            }
+            (
+                Self::RecoveryDisposition {
+                    identity_sha256: previous_identity,
+                    copy_manifest_sha256: previous_manifest,
+                    copy_manifest_count: previous_manifest_count,
+                    created_at_unix_nanos: previous_created_at,
+                    revision: previous_revision,
+                    state: previous_state,
+                    owner_fence_sha256: previous_owner,
+                    owner_lease_acquired_at_unix_nanos: previous_owner_acquired,
+                    owner_lease_expires_at_unix_nanos: previous_owner_expires,
+                    confirmed_absent_sha256: previous_confirmed,
+                    retain_until_unix_nanos: previous_retain_until,
+                    ..
+                },
+                Self::RecoveryDisposition {
+                    identity_sha256: next_identity,
+                    copy_manifest_sha256: next_manifest,
+                    copy_manifest_count: next_manifest_count,
+                    created_at_unix_nanos: next_created_at,
+                    revision: next_revision,
+                    state: next_state,
+                    owner_fence_sha256: next_owner,
+                    owner_lease_acquired_at_unix_nanos: next_owner_acquired,
+                    owner_lease_expires_at_unix_nanos: next_owner_expires,
+                    confirmed_absent_sha256: next_confirmed,
+                    retain_until_unix_nanos: next_retain_until,
+                    ..
+                },
+            ) => {
+                use recovery_disposition::IlmRecoveryDispositionState::{Applying, Completed, Prepared};
+
+                let immutable_identity_matches = previous_identity == next_identity
+                    && previous_manifest == next_manifest
+                    && previous_manifest_count == next_manifest_count
+                    && previous_created_at == next_created_at
+                    && previous_retain_until == next_retain_until;
+                let adjacent = previous_revision.checked_add(1) == Some(*next_revision);
+                let progress_is_monotonic = sorted_sha256_set_is_subset(previous_confirmed, next_confirmed);
+                let legal_edge = match (previous_state, next_state) {
+                    (Prepared, Prepared) => {
+                        let claim = previous_owner.is_none() && next_owner.is_some();
+                        let takeover = previous_owner.is_some()
+                            && previous_owner != next_owner
+                            && previous_owner_expires
+                                .zip(*next_owner_acquired)
+                                .is_some_and(|(expires, acquired)| acquired >= expires);
+                        previous_confirmed == next_confirmed && (claim || takeover)
+                    }
+                    (Prepared, Applying) => {
+                        previous_confirmed == next_confirmed
+                            && previous_owner.is_some()
+                            && previous_owner == next_owner
+                            && previous_owner_acquired == next_owner_acquired
+                            && previous_owner_expires == next_owner_expires
+                    }
+                    (Applying, Applying) => {
+                        let progress = previous_owner == next_owner
+                            && previous_owner_acquired == next_owner_acquired
+                            && previous_owner_expires == next_owner_expires
+                            && previous_confirmed.len().checked_add(1) == Some(next_confirmed.len());
+                        let takeover = previous_owner.is_some()
+                            && previous_owner != next_owner
+                            && previous_confirmed == next_confirmed
+                            && previous_owner_expires
+                                .zip(*next_owner_acquired)
+                                .is_some_and(|(expires, acquired)| acquired >= expires);
+                        progress || takeover
+                    }
+                    (Applying, Completed) => {
+                        previous_owner.is_some() && next_owner.is_none() && previous_confirmed == next_confirmed
+                    }
+                    _ => false,
+                };
+
+                immutable_identity_matches && adjacent && progress_is_monotonic && legal_edge
+            }
             _ => false,
         };
 
@@ -545,6 +729,11 @@ impl DurableIlmRecordCheckpoint {
     /// after the exact terminal ETag and terminal receipt were committed, to
     /// purge older object versions exposed by that deletion.
     pub(crate) fn is_predecessor_of_terminal(&self, terminal: &Self) -> bool {
+        for checkpoint in [self, terminal] {
+            if !recovery_disposition_checkpoint_is_valid(checkpoint) {
+                return false;
+            }
+        }
         if let Self::TierProbeIntent { state, .. } = terminal
             && !matches!(
                 state,
@@ -553,6 +742,19 @@ impl DurableIlmRecordCheckpoint {
         {
             return false;
         }
+        if let Self::RecoveryControl { classification, .. } = terminal
+            && !matches!(
+                classification,
+                recovery_control::IlmRecoveryClassification::Terminal | recovery_control::IlmRecoveryClassification::Abandoned
+            )
+        {
+            return false;
+        }
+        if let Self::RecoveryDisposition { state, .. } = terminal
+            && state != &recovery_disposition::IlmRecoveryDispositionState::Completed
+        {
+            return false;
+        }
         if self == terminal || self.validate_successor(terminal).is_ok() {
             return true;
         }
@@ -652,11 +854,147 @@ impl DurableIlmRecordCheckpoint {
                         .is_some_and(|distance| tier_probe_state_reaches(*previous_state, *terminal_state, distance))
                     && (!previous_remote_version_known || previous_remote_version == terminal_remote_version)
             }
+            (
+                Self::RecoveryControl {
+                    identity_sha256: previous_identity,
+                    source_generation_sha256: previous_generation,
+                    first_seen_at_unix_nanos: previous_first_seen,
+                    revision: previous_revision,
+                    attempt_count: previous_attempts,
+                    ..
+                },
+                Self::RecoveryControl {
+                    identity_sha256: terminal_identity,
+                    source_generation_sha256: terminal_generation,
+                    first_seen_at_unix_nanos: terminal_first_seen,
+                    revision: terminal_revision,
+                    attempt_count: terminal_attempts,
+                    classification:
+                        recovery_control::IlmRecoveryClassification::Terminal | recovery_control::IlmRecoveryClassification::Abandoned,
+                    ..
+                },
+            ) => {
+                previous_identity == terminal_identity
+                    && (previous_generation == terminal_generation || terminal_attempts > previous_attempts)
+                    && previous_first_seen == terminal_first_seen
+                    && terminal_revision > previous_revision
+                    && terminal_attempts >= previous_attempts
+            }
+            (
+                Self::RecoveryDisposition {
+                    identity_sha256: previous_identity,
+                    copy_manifest_sha256: previous_manifest,
+                    copy_manifest_count: previous_manifest_count,
+                    created_at_unix_nanos: previous_created_at,
+                    revision: previous_revision,
+                    state: previous_state,
+                    owner_fence_sha256: previous_owner,
+                    confirmed_absent_sha256: previous_confirmed,
+                    retain_until_unix_nanos: previous_retain_until,
+                    ..
+                },
+                Self::RecoveryDisposition {
+                    identity_sha256: terminal_identity,
+                    copy_manifest_sha256: terminal_manifest,
+                    copy_manifest_count: terminal_manifest_count,
+                    created_at_unix_nanos: terminal_created_at,
+                    revision: terminal_revision,
+                    state: recovery_disposition::IlmRecoveryDispositionState::Completed,
+                    confirmed_absent_sha256: terminal_confirmed,
+                    retain_until_unix_nanos: terminal_retain_until,
+                    ..
+                },
+            ) => {
+                matches!(
+                    previous_state,
+                    recovery_disposition::IlmRecoveryDispositionState::Prepared
+                        | recovery_disposition::IlmRecoveryDispositionState::Applying
+                ) && previous_identity == terminal_identity
+                    && previous_manifest == terminal_manifest
+                    && previous_manifest_count == terminal_manifest_count
+                    && previous_created_at == terminal_created_at
+                    && previous_retain_until == terminal_retain_until
+                    && terminal_revision.checked_sub(*previous_revision).is_some_and(|distance| {
+                        let minimum_distance = match previous_state {
+                            recovery_disposition::IlmRecoveryDispositionState::Prepared if previous_owner.is_some() => 3,
+                            recovery_disposition::IlmRecoveryDispositionState::Prepared => 4,
+                            recovery_disposition::IlmRecoveryDispositionState::Applying
+                                if previous_confirmed.len() == *previous_manifest_count =>
+                            {
+                                1
+                            }
+                            recovery_disposition::IlmRecoveryDispositionState::Applying => 2,
+                            recovery_disposition::IlmRecoveryDispositionState::Completed => u64::MAX,
+                        };
+                        distance >= minimum_distance
+                    })
+                    && sorted_sha256_set_is_subset(previous_confirmed, terminal_confirmed)
+            }
             _ => false,
         }
     }
 }
 
+fn recovery_disposition_checkpoint_is_valid(checkpoint: &DurableIlmRecordCheckpoint) -> bool {
+    use recovery_disposition::IlmRecoveryDispositionState::{Applying, Completed, Prepared};
+
+    let DurableIlmRecordCheckpoint::RecoveryDisposition {
+        content_sha256,
+        identity_sha256,
+        copy_manifest_sha256,
+        copy_manifest_count,
+        created_at_unix_nanos,
+        revision,
+        state,
+        owner_fence_sha256,
+        owner_lease_acquired_at_unix_nanos,
+        owner_lease_expires_at_unix_nanos,
+        confirmed_absent_sha256,
+        retain_until_unix_nanos,
+    } = checkpoint
+    else {
+        return true;
+    };
+    let owner_fence_sha256 = owner_fence_sha256.as_deref();
+
+    is_canonical_sha256(content_sha256)
+        && is_canonical_sha256(identity_sha256)
+        && is_canonical_sha256(copy_manifest_sha256)
+        && *copy_manifest_count > 0
+        && *created_at_unix_nanos > 0
+        && *revision > 0
+        && *retain_until_unix_nanos > 0
+        && owner_fence_sha256.is_none_or(is_canonical_sha256)
+        && match (
+            owner_fence_sha256,
+            *owner_lease_acquired_at_unix_nanos,
+            *owner_lease_expires_at_unix_nanos,
+        ) {
+            (None, None, None) => true,
+            (Some(_), Some(acquired), Some(expires)) => acquired > 0 && expires > acquired,
+            _ => false,
+        }
+        && confirmed_absent_sha256.len() <= *copy_manifest_count
+        && confirmed_absent_sha256.iter().all(|digest| is_canonical_sha256(digest))
+        && confirmed_absent_sha256.windows(2).all(|pair| pair[0] < pair[1])
+        && match *state {
+            Prepared => confirmed_absent_sha256.is_empty(),
+            Applying => owner_fence_sha256.is_some(),
+            Completed => owner_fence_sha256.is_none() && confirmed_absent_sha256.len() == *copy_manifest_count,
+        }
+}
+
+fn is_canonical_sha256(value: &str) -> bool {
+    is_sha256_checksum(value)
+        && !value
+            .bytes()
+            .any(|byte| byte.is_ascii_hexdigit() && byte.is_ascii_uppercase())
+}
+
+fn sorted_sha256_set_is_subset(subset: &[String], superset: &[String]) -> bool {
+    subset.iter().all(|candidate| superset.binary_search(candidate).is_ok())
+}
+
 fn tier_delete_dispatch_parent_progress_delta(
     previous_sequence: u64,
     previous_completed_journals: u64,
@@ -690,6 +1028,22 @@ fn transition_state_distance(
     }
 }
 
+fn transition_state_revision_is_successor(
+    from: transition_transaction::TransitionTransactionState,
+    from_revision: u64,
+    to: transition_transaction::TransitionTransactionState,
+    to_revision: u64,
+) -> bool {
+    use transition_transaction::TransitionTransactionState::{LocalCommitStarted, UploadOutcomeUnknown};
+
+    if from == UploadOutcomeUnknown && from_revision == 1 && to == LocalCommitStarted {
+        return to_revision == 2;
+    }
+    transition_state_distance(from, to)
+        .and_then(|distance| from_revision.checked_add(distance))
+        .is_some_and(|expected_revision| to_revision == expected_revision)
+}
+
 fn tier_probe_state_reaches(
     from: tier_probe_intent::TierProbeIntentState,
     to: tier_probe_intent::TierProbeIntentState,
@@ -1219,6 +1573,84 @@ pub(crate) fn validate_durable_ilm_record(path: &str, data: &[u8]) -> Result<Val
                 },
             )
         }
+        DurableIlmRecordKind::RecoveryControl => {
+            let (protocol, control_id) = recovery_control::recovery_control_id_from_record_object_name(path)
+                .map_err(|err| Error::other(err.to_string()))?;
+            let control =
+                recovery_control::IlmRecoveryControl::decode(&control_id, data).map_err(|err| Error::other(err.to_string()))?;
+            let canonical = recovery_control::recovery_control_record_object_name(protocol, &control_id)
+                .map_err(|err| Error::other(err.to_string()))?;
+            if canonical != path || control.identity.protocol != protocol {
+                return Err(Error::other("ILM recovery control path is not canonical"));
+            }
+            let identity_sha256 = checkpoint_hash(&control.identity)?;
+            let source_generation_sha256 = checkpoint_hash(&control.observed_source_generation)?;
+            let owner_fence_sha256 = control.owner.as_ref().map(checkpoint_hash).transpose()?;
+            (
+                "control_id",
+                control_id,
+                DurableIlmRecordCheckpoint::RecoveryControl {
+                    content_sha256,
+                    identity_sha256,
+                    source_generation_sha256,
+                    first_seen_at_unix_nanos: control.first_seen_at_unix_nanos,
+                    revision: control.revision,
+                    classification: control.classification,
+                    attempt_count: control.attempt_count,
+                    consecutive_failure_count: control.consecutive_failure_count,
+                    owner_fence_sha256,
+                },
+            )
+        }
+        DurableIlmRecordKind::RecoveryExport => {
+            let (protocol, export_id) = recovery_export::recovery_export_id_from_record_object_name(path)?;
+            let export = recovery_export::IlmRecoveryExport::decode(&export_id, data)?;
+            let canonical = recovery_export::recovery_export_record_object_name(protocol, &export_id)?;
+            if canonical != path || export.protocol != protocol {
+                return Err(Error::other("ILM recovery export path is not canonical"));
+            }
+            let source_generation_sha256 = checkpoint_hash(&export.source_generation)?;
+            (
+                "export_id",
+                export_id,
+                DurableIlmRecordCheckpoint::RecoveryExport {
+                    content_sha256,
+                    source_generation_sha256,
+                    topology_generation: export.topology_generation,
+                    member_epochs_sha256: export.member_epochs_sha256,
+                    creator_sha256: export.creator_sha256,
+                    retain_until_unix_nanos: export.retain_until_unix_nanos,
+                },
+            )
+        }
+        DurableIlmRecordKind::RecoveryDisposition => {
+            // The disposition module owns strict schema, checksum, canonical
+            // path, immutable-manifest, and state-specific validation. Keep
+            // this boundary limited to decommission identity/checkpoint
+            // projection so the two readers cannot accept different records.
+            let disposition = recovery_disposition::decode_recovery_disposition_checkpoint(path, data)?;
+            if disposition.content_sha256 != content_sha256 {
+                return Err(Error::other("ILM recovery disposition checkpoint content digest is invalid"));
+            }
+            (
+                "disposition_id",
+                disposition.disposition_id,
+                DurableIlmRecordCheckpoint::RecoveryDisposition {
+                    content_sha256: disposition.content_sha256,
+                    identity_sha256: disposition.identity_sha256,
+                    copy_manifest_sha256: disposition.copy_manifest_sha256,
+                    copy_manifest_count: disposition.copy_manifest_count,
+                    created_at_unix_nanos: disposition.created_at_unix_nanos,
+                    revision: disposition.revision,
+                    state: disposition.state,
+                    owner_fence_sha256: disposition.owner_fence_sha256,
+                    owner_lease_acquired_at_unix_nanos: disposition.owner_lease_acquired_at_unix_nanos,
+                    owner_lease_expires_at_unix_nanos: disposition.owner_lease_expires_at_unix_nanos,
+                    confirmed_absent_sha256: disposition.confirmed_absent_sha256,
+                    retain_until_unix_nanos: disposition.retain_until_unix_nanos,
+                },
+            )
+        }
         DurableIlmRecordKind::ManualTransitionJob => {
             let job_id = manual_transition_job::manual_transition_job_id_from_record_object_name(path)
                 .map_err(|err| Error::other(err.to_string()))?;
@@ -1374,6 +1806,203 @@ mod tests {
         }
     }
 
+    fn recovery_disposition_checkpoint(
+        revision: u64,
+        state: recovery_disposition::IlmRecoveryDispositionState,
+        owner_fence: Option<&str>,
+        confirmed_absent_sha256: Vec<String>,
+    ) -> DurableIlmRecordCheckpoint {
+        let (owner_lease_acquired_at_unix_nanos, owner_lease_expires_at_unix_nanos) = match owner_fence {
+            Some("f") => (Some(10), Some(20)),
+            Some(_) => (Some(1), Some(10)),
+            None => (None, None),
+        };
+        DurableIlmRecordCheckpoint::RecoveryDisposition {
+            content_sha256: format!("{revision:064x}"),
+            identity_sha256: "a".repeat(64),
+            copy_manifest_sha256: "d".repeat(64),
+            copy_manifest_count: 2,
+            created_at_unix_nanos: 1_700_000_000_000_000_000,
+            revision,
+            state,
+            owner_fence_sha256: owner_fence.map(|digest| digest.repeat(64)),
+            owner_lease_acquired_at_unix_nanos,
+            owner_lease_expires_at_unix_nanos,
+            confirmed_absent_sha256,
+            retain_until_unix_nanos: 1_820_000_000_000_000_000,
+        }
+    }
+
+    #[test]
+    fn recovery_disposition_namespace_is_registered_without_shadowing_its_root() {
+        let disposition_id = "a".repeat(64);
+        let path = format!(
+            "{}/tier_delete_journal/{}/{}/{}.json",
+            recovery_disposition::ILM_RECOVERY_DISPOSITION_PREFIX,
+            &disposition_id[..2],
+            &disposition_id[2..4],
+            disposition_id
+        );
+        let namespace = classify_durable_ilm_record(&path)
+            .expect("recovery disposition path should classify")
+            .expect("recovery disposition should be durable");
+
+        assert_eq!(namespace, &RECOVERY_DISPOSITION_NAMESPACE);
+        assert!(classify_durable_ilm_record(recovery_disposition::ILM_RECOVERY_DISPOSITION_PREFIX).is_err());
+    }
+
+    #[test]
+    fn recovery_disposition_checkpoint_accepts_only_monotonic_progress() {
+        use recovery_disposition::IlmRecoveryDispositionState::{Applying, Completed, Prepared};
+
+        let first_copy = "b".repeat(64);
+        let second_copy = "c".repeat(64);
+        let prepared = recovery_disposition_checkpoint(1, Prepared, None, Vec::new());
+        let claimed = recovery_disposition_checkpoint(2, Prepared, Some("e"), Vec::new());
+        let applying = recovery_disposition_checkpoint(3, Applying, Some("e"), Vec::new());
+        let first_absent = recovery_disposition_checkpoint(4, Applying, Some("e"), vec![first_copy.clone()]);
+        let taken_over = recovery_disposition_checkpoint(5, Applying, Some("f"), vec![first_copy.clone()]);
+        let all_absent = recovery_disposition_checkpoint(6, Applying, Some("f"), vec![first_copy.clone(), second_copy.clone()]);
+        let completed = recovery_disposition_checkpoint(7, Completed, None, vec![first_copy.clone(), second_copy.clone()]);
+
+        prepared
+            .validate_successor(&claimed)
+            .expect("Prepared should record an owner claim without absence progress");
+        claimed
+            .validate_successor(&applying)
+            .expect("Prepared should advance to Applying without folding in deletion progress");
+        applying
+            .validate_successor(&first_absent)
+            .expect("Applying should append newly confirmed absent copies");
+        first_absent
+            .validate_successor(&taken_over)
+            .expect("Applying should record a fenced owner takeover without losing progress");
+        taken_over
+            .validate_successor(&all_absent)
+            .expect("Applying should preserve every earlier confirmation while making progress");
+        all_absent
+            .validate_successor(&completed)
+            .expect("a fully confirmed manifest should advance to Completed");
+
+        assert!(
+            prepared.validate_successor(&completed).is_err(),
+            "adjacent receipt updates must not skip Applying"
+        );
+        assert!(
+            first_absent
+                .validate_successor(&recovery_disposition_checkpoint(5, Applying, Some("e"), Vec::new()))
+                .is_err(),
+            "confirmed-absent progress must not move backwards"
+        );
+        assert!(
+            applying
+                .validate_successor(&recovery_disposition_checkpoint(4, Completed, None, vec![first_copy.clone()]))
+                .is_err(),
+            "Completed must cover the complete immutable copy manifest"
+        );
+        assert!(
+            completed
+                .validate_successor(&recovery_disposition_checkpoint(7, Applying, Some("e"), vec![second_copy]))
+                .is_err(),
+            "Completed is terminal"
+        );
+        assert!(
+            first_absent
+                .validate_successor(&recovery_disposition_checkpoint(5, Applying, Some("e"), vec![first_copy.clone()]))
+                .is_err(),
+            "a same-state revision bump must change the owner fence or absence progress"
+        );
+        assert!(
+            applying
+                .validate_successor(&recovery_disposition_checkpoint(4, Applying, None, vec![first_copy]))
+                .is_err(),
+            "Applying must retain a fenced owner"
+        );
+
+        let mut noncanonical_identity = claimed.clone();
+        if let DurableIlmRecordCheckpoint::RecoveryDisposition { identity_sha256, .. } = &mut noncanonical_identity {
+            *identity_sha256 = "A".repeat(64);
+        }
+        assert!(prepared.validate_successor(&noncanonical_identity).is_err());
+        let mut changed_created_at = claimed;
+        if let DurableIlmRecordCheckpoint::RecoveryDisposition {
+            created_at_unix_nanos, ..
+        } = &mut changed_created_at
+        {
+            *created_at_unix_nanos += 1;
+        }
+        assert!(prepared.validate_successor(&changed_created_at).is_err());
+
+        let mut early_takeover = taken_over;
+        if let DurableIlmRecordCheckpoint::RecoveryDisposition {
+            owner_lease_acquired_at_unix_nanos,
+            ..
+        } = &mut early_takeover
+        {
+            *owner_lease_acquired_at_unix_nanos = Some(9);
+        }
+        assert!(first_absent.validate_successor(&early_takeover).is_err());
+        assert!(
+            applying
+                .validate_successor(&recovery_disposition_checkpoint(
+                    4,
+                    Applying,
+                    Some("e"),
+                    vec!["c".repeat(64), "b".repeat(64)],
+                ))
+                .is_err(),
+            "confirmed-absent entries must be a canonical sorted set"
+        );
+    }
+
+    #[test]
+    fn recovery_disposition_terminal_predecessor_requires_exact_identity_and_full_manifest() {
+        use recovery_disposition::IlmRecoveryDispositionState::{Applying, Completed, Prepared};
+
+        let first_copy = "b".repeat(64);
+        let second_copy = "c".repeat(64);
+        let prepared = recovery_disposition_checkpoint(1, Prepared, None, Vec::new());
+        let applying = recovery_disposition_checkpoint(3, Applying, Some("e"), vec![first_copy.clone()]);
+        let completed = recovery_disposition_checkpoint(5, Completed, None, vec![first_copy.clone(), second_copy]);
+
+        assert!(prepared.is_predecessor_of_terminal(&completed));
+        assert!(applying.is_predecessor_of_terminal(&completed));
+        assert!(
+            !prepared.is_predecessor_of_terminal(&recovery_disposition_checkpoint(2, Applying, Some("e"), Vec::new())),
+            "a nonterminal disposition must not authorize terminal cleanup"
+        );
+        assert!(
+            !prepared.is_predecessor_of_terminal(&recovery_disposition_checkpoint(
+                4,
+                Completed,
+                None,
+                vec![first_copy.clone(), "c".repeat(64)],
+            )),
+            "terminal proof must leave enough revisions for claim, apply, progress, and completion"
+        );
+        assert!(
+            !applying.is_predecessor_of_terminal(&recovery_disposition_checkpoint(
+                4,
+                Completed,
+                None,
+                vec![first_copy.clone(), "c".repeat(64)],
+            )),
+            "an incomplete Applying checkpoint cannot complete without a progress generation"
+        );
+
+        let mut other_identity = completed;
+        if let DurableIlmRecordCheckpoint::RecoveryDisposition { identity_sha256, .. } = &mut other_identity {
+            *identity_sha256 = "e".repeat(64);
+        }
+        assert!(!prepared.is_predecessor_of_terminal(&other_identity));
+
+        let incomplete_terminal = recovery_disposition_checkpoint(4, Completed, None, vec![first_copy]);
+        assert!(
+            !prepared.is_predecessor_of_terminal(&incomplete_terminal),
+            "a partial confirmed-absent set must not become terminal proof"
+        );
+    }
+
     fn tier_probe_intent_fixture() -> tier_probe_intent::TierProbeIntent {
         let probe_id = Uuid::parse_str("36e2220e-9ad2-495b-b3bc-c4d2caf70a31").expect("fixture uuid should parse");
         tier_probe_intent::TierProbeIntent {
@@ -1412,6 +2041,94 @@ mod tests {
             .checkpoint
     }
 
+    fn recovery_control_fixture() -> recovery_control::IlmRecoveryControl {
+        let source_path = "ilm/transition-transactions/records/12/34/1234567890abcdef1234567890abcdef.json";
+        let generation = recovery_control::IlmRecoverySourceGeneration::new(
+            transition_transaction::TRANSITION_TRANSACTION_SCHEMA,
+            "source-etag",
+            "a".repeat(64),
+            vec![recovery_control::IlmRecoverySourceCopy {
+                authority: "pool-0/set-0".to_string(),
+                canonical_path: source_path.to_string(),
+                etag: "source-etag".to_string(),
+                encoded_len: 128,
+                content_sha256: "a".repeat(64),
+            }],
+        )
+        .expect("source generation should build");
+        recovery_control::IlmRecoveryControl::new(
+            recovery_control::IlmRecoveryControlIdentity {
+                protocol: recovery_control::IlmRecoveryProtocol::TransitionTransaction,
+                canonical_source_path: source_path.to_string(),
+                stable_operation_identity: "12345678-90ab-cdef-1234-567890abcdef".to_string(),
+                record_class: "transition_transaction_v1".to_string(),
+            },
+            generation,
+            recovery_control::IlmRecoveryClassification::Retrying,
+            1_000_000_000,
+            recovery_control::IlmRecoveryErrorCode::None,
+        )
+        .expect("recovery control should build")
+    }
+
+    fn recovery_control_checkpoint(control: &recovery_control::IlmRecoveryControl) -> DurableIlmRecordCheckpoint {
+        let control_id = control.identity.source_operation_digest().expect("control id should derive");
+        let path = recovery_control::recovery_control_record_object_name(control.identity.protocol, &control_id)
+            .expect("control path should build");
+        let encoded = control.encode().expect("control should encode");
+        let namespace = classify_durable_ilm_record(&path)
+            .expect("recovery control namespace should classify")
+            .expect("recovery control should be durable");
+        assert_eq!(namespace, &RECOVERY_CONTROL_NAMESPACE);
+        validate_durable_ilm_record(&path, &encoded)
+            .expect("recovery control should validate")
+            .checkpoint
+    }
+
+    #[test]
+    fn recovery_control_checkpoint_tracks_claim_retry_and_terminal_generations() {
+        let initial_control = recovery_control_fixture();
+        let initial = recovery_control_checkpoint(&initial_control);
+
+        let mut claimed_control = initial_control;
+        let mut advanced_generation = claimed_control.observed_source_generation.clone();
+        advanced_generation.source_schema = "rustfs-transition-transaction-v2".to_string();
+        claimed_control
+            .claim_for_source_generation("node-a", Uuid::new_v4(), 2_000_000_000, 300_000_000_000, advanced_generation)
+            .expect("control should claim");
+        let claimed = recovery_control_checkpoint(&claimed_control);
+        initial.validate_successor(&claimed).expect("claim should advance receipt");
+
+        let mut retry_control = claimed_control;
+        retry_control
+            .record_retryable_failure(3_000_000_000, recovery_control::IlmRecoveryErrorCode::BackendTimeout)
+            .expect("retry should persist");
+        let retry = recovery_control_checkpoint(&retry_control);
+        claimed.validate_successor(&retry).expect("retry should advance receipt");
+
+        let ready_at = retry_control
+            .next_attempt_at_unix_nanos
+            .expect("retry deadline should persist");
+        let mut terminal_control = retry_control;
+        terminal_control
+            .claim("node-b", Uuid::new_v4(), ready_at, 300_000_000_000)
+            .expect("retry should claim");
+        let reclaimed = recovery_control_checkpoint(&terminal_control);
+        retry.validate_successor(&reclaimed).expect("reclaim should advance receipt");
+        terminal_control
+            .finish_attempt(
+                recovery_control::IlmRecoveryClassification::Terminal,
+                recovery_control::IlmRecoveryErrorCode::None,
+            )
+            .expect("control should terminate");
+        let terminal = recovery_control_checkpoint(&terminal_control);
+        reclaimed
+            .validate_successor(&terminal)
+            .expect("terminal state should advance receipt");
+        assert!(initial.is_predecessor_of_terminal(&terminal));
+        assert!(!initial.is_predecessor_of_terminal(&retry));
+    }
+
     #[test]
     fn tier_probe_intent_checkpoint_tracks_exact_monotonic_generations() {
         let initial_intent = tier_probe_intent_fixture();
@@ -1470,6 +2187,64 @@ mod tests {
         );
     }
 
+    #[test]
+    fn transition_checkpoint_accepts_only_the_distinguishable_compact_edge() {
+        let identity_sha256 = "a".repeat(64);
+        let unknown_remote_sha256 = "b".repeat(64);
+        let known_remote_sha256 = "c".repeat(64);
+        let checkpoint = |revision, state, remote_version_sha256: String, remote_version_known| {
+            DurableIlmRecordCheckpoint::TransitionTransaction {
+                content_sha256: format!("{revision:064x}"),
+                identity_sha256: identity_sha256.clone(),
+                remote_version_sha256,
+                remote_version_known,
+                revision,
+                state,
+            }
+        };
+        let compact_unknown = checkpoint(
+            1,
+            transition_transaction::TransitionTransactionState::UploadOutcomeUnknown,
+            unknown_remote_sha256.clone(),
+            false,
+        );
+        let compact_local_commit = checkpoint(
+            2,
+            transition_transaction::TransitionTransactionState::LocalCommitStarted,
+            known_remote_sha256.clone(),
+            true,
+        );
+        compact_unknown
+            .validate_successor(&compact_local_commit)
+            .expect("compact pre-upload fence should advance directly to the exact local-commit fence");
+
+        let legacy_unknown = checkpoint(
+            2,
+            transition_transaction::TransitionTransactionState::UploadOutcomeUnknown,
+            unknown_remote_sha256,
+            false,
+        );
+        let invalid_legacy_skip = checkpoint(
+            3,
+            transition_transaction::TransitionTransactionState::LocalCommitStarted,
+            known_remote_sha256.clone(),
+            true,
+        );
+        assert!(
+            legacy_unknown.validate_successor(&invalid_legacy_skip).is_err(),
+            "legacy UploadOutcomeUnknown@2 must not masquerade as the compact edge"
+        );
+        let valid_legacy_skip = checkpoint(
+            4,
+            transition_transaction::TransitionTransactionState::LocalCommitStarted,
+            known_remote_sha256,
+            true,
+        );
+        legacy_unknown
+            .validate_successor(&valid_legacy_skip)
+            .expect("legacy receipts may still observe the existing two-edge state advance");
+    }
+
     #[test]
     fn tier_delete_dispatch_manifest_namespace_validates_monotonic_branches() {
         use tier_delete_journal::TierDeleteDispatchManifestState::{Aborted, Aborting, Completed, DispatchAuthorized, Preparing};

crates/ecstore/src/bucket/lifecycle/mod.rs

@@ -24,6 +24,10 @@ pub(crate) use metadata_boundary::{LifecycleExpiryConfigs, get_expiry_configs, g
 mod object_handlers_common;
 mod object_lock_boundary;
 pub use self::core as lifecycle;
+pub mod recovery_control;
+pub mod recovery_disposition;
+pub(crate) mod recovery_disposition_runtime;
+pub mod recovery_export;
 mod replication_sink;
 pub mod rule;
 mod runtime_boundary;

crates/ecstore/src/bucket/lifecycle/recovery_export.rs

@@ -0,0 +1,840 @@
+// 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 std::{collections::HashSet, sync::Arc};
+
+use rustfs_utils::crypto::{hex_sha256, is_sha256_checksum};
+use serde::{Deserialize, Serialize};
+
+use super::config_boundary;
+use super::recovery_control::{
+    IlmRecoveryClassification, IlmRecoveryControl, IlmRecoveryProtocol, IlmRecoverySourceCopy, IlmRecoverySourceGeneration,
+    MAX_ILM_RECOVERY_CONTROL_SIZE, ObservedIlmRecoveryControl, ObservedIlmRecoverySource, recovery_control_record_object_name,
+};
+use super::tier_delete_journal::{
+    TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA, TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA, validate_legacy_tier_delete_recovery_source,
+};
+use crate::disk::RUSTFS_META_BUCKET;
+use crate::error::{Error, Result};
+use crate::object_api::{ObjectOptions, WriteCompletion};
+use crate::services::notification_sys::{
+    acquire_ilm_recovery_export_fleet_proof, ilm_recovery_export_fleet_proof_matches, ilm_recovery_export_member_epochs_sha256,
+    ilm_recovery_export_topology_generation,
+};
+use crate::storage_api_contracts::{list::ListOperations as _, namespace::NamespaceLocking as _, object::HTTPPreconditions};
+use crate::store::ECStore;
+
+pub const ILM_RECOVERY_EXPORT_SCHEMA: &str = "rustfs-ilm-recovery-export-v1";
+pub const ILM_RECOVERY_EXPORT_PREFIX: &str = "ilm/recovery-exports";
+pub const MAX_ILM_RECOVERY_EXPORT_SIZE: usize = 128 * 1024;
+const MAX_ILM_RECOVERY_EXPORTS: usize = 10_000;
+const MAX_ILM_RECOVERY_EXPORT_BYTES: u64 = 1024 * 1024 * 1024;
+const MAX_ACTOR_EXPORTS_PER_MINUTE: usize = 10;
+const MAX_CLUSTER_EXPORTS_PER_MINUTE: usize = 100;
+const EXPORT_RETENTION_NANOS: i64 = 90 * 24 * 60 * 60 * 1_000_000_000;
+const EXPORT_ADMISSION_LOCK: &str = "ilm/recovery-admission/export.lock";
+const MAX_LEGACY_TIER_DELETE_SOURCE_SIZE: usize = 64 * 1024;
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct IlmRecoveryExportObservation {
+    pub control_id: String,
+    pub protocol: IlmRecoveryProtocol,
+    pub control_etag: String,
+    pub control_revision: u64,
+    pub classification: IlmRecoveryClassification,
+    pub canonical_source_path: String,
+    pub source_generation: IlmRecoverySourceGeneration,
+    pub topology_generation: String,
+    pub member_epochs_sha256: String,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct IlmRecoveryExport {
+    pub export_id: String,
+    pub control_id: String,
+    pub protocol: IlmRecoveryProtocol,
+    pub control_etag: String,
+    pub control_revision: u64,
+    pub classification: IlmRecoveryClassification,
+    pub canonical_source_path: String,
+    pub source_generation: IlmRecoverySourceGeneration,
+    pub topology_generation: String,
+    pub member_epochs_sha256: String,
+    pub creator_sha256: String,
+    pub created_at_unix_nanos: i64,
+    pub retain_until_unix_nanos: i64,
+    pub source_bytes_base64: String,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+struct PersistedIlmRecoveryExport {
+    schema: String,
+    content_sha256: String,
+    export: IlmRecoveryExport,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+pub struct IlmRecoveryExportCreated {
+    pub export_id: String,
+    pub content_sha256: String,
+    pub encoded: Vec<u8>,
+    pub replayed: bool,
+}
+
+impl IlmRecoveryExport {
+    fn validate(&self) -> Result<()> {
+        self.source_generation.validate().map_err(Error::other)?;
+        validate_sha256(&self.export_id, "ILM recovery export ID is invalid")?;
+        validate_sha256(&self.control_id, "ILM recovery export control ID is invalid")?;
+        validate_sha256(&self.topology_generation, "ILM recovery export topology generation is invalid")?;
+        validate_sha256(&self.member_epochs_sha256, "ILM recovery export member epoch digest is invalid")?;
+        validate_sha256(&self.creator_sha256, "ILM recovery export creator digest is invalid")?;
+        if self.protocol != IlmRecoveryProtocol::TierDeleteJournal
+            || self.classification != IlmRecoveryClassification::RetainedAmbiguous
+            || !is_legacy_export_schema(&self.source_generation.source_schema)
+        {
+            return Err(Error::other("ILM recovery export source is not an exportable legacy journal"));
+        }
+        if self.control_etag.trim().is_empty() || self.control_revision == 0 {
+            return Err(Error::other("ILM recovery export control generation is invalid"));
+        }
+        if self.canonical_source_path.is_empty()
+            || self.canonical_source_path.starts_with('/')
+            || self.canonical_source_path.ends_with('/')
+            || self.canonical_source_path.split('/').any(str::is_empty)
+        {
+            return Err(Error::other("ILM recovery export source path is invalid"));
+        }
+        if self.created_at_unix_nanos <= 0
+            || self.retain_until_unix_nanos < self.created_at_unix_nanos.saturating_add(EXPORT_RETENTION_NANOS)
+        {
+            return Err(Error::other("ILM recovery export retention is invalid"));
+        }
+        let source = base64_simd::STANDARD
+            .decode_to_vec(self.source_bytes_base64.as_bytes())
+            .map_err(|_| Error::other("ILM recovery export source encoding is invalid"))?;
+        validate_legacy_tier_delete_recovery_source(&self.canonical_source_path, &self.source_generation.source_schema, &source)?;
+        let encoded_len = u64::try_from(source.len()).map_err(|_| Error::other("ILM recovery export source length overflow"))?;
+        if source.is_empty()
+            || source.len() > MAX_LEGACY_TIER_DELETE_SOURCE_SIZE
+            || hex_sha256(&source, ToOwned::to_owned) != self.source_generation.content_sha256
+            || self.source_generation.copies.iter().any(|copy| {
+                copy.canonical_path != self.canonical_source_path
+                    || copy.etag != self.source_generation.source_etag
+                    || copy.content_sha256 != self.source_generation.content_sha256
+                    || copy.encoded_len != encoded_len
+            })
+        {
+            return Err(Error::other("ILM recovery export source bytes do not match the observed generation"));
+        }
+        if recovery_export_id(&self.control_id, &self.source_generation)? != self.export_id {
+            return Err(Error::other("ILM recovery export ID does not match its source generation"));
+        }
+        Ok(())
+    }
+
+    pub fn encode(&self) -> Result<Vec<u8>> {
+        self.validate()?;
+        let export_bytes = serde_json::to_vec(self).map_err(Error::other)?;
+        let persisted = PersistedIlmRecoveryExport {
+            schema: ILM_RECOVERY_EXPORT_SCHEMA.to_string(),
+            content_sha256: hex_sha256(&export_bytes, ToOwned::to_owned),
+            export: self.clone(),
+        };
+        let encoded = serde_json::to_vec(&persisted).map_err(Error::other)?;
+        if encoded.len() > MAX_ILM_RECOVERY_EXPORT_SIZE {
+            return Err(Error::other("encoded ILM recovery export exceeds maximum size"));
+        }
+        Ok(encoded)
+    }
+
+    pub fn decode(expected_export_id: &str, data: &[u8]) -> Result<Self> {
+        validate_sha256(expected_export_id, "ILM recovery export ID is invalid")?;
+        if data.len() > MAX_ILM_RECOVERY_EXPORT_SIZE {
+            return Err(Error::other("encoded ILM recovery export exceeds maximum size"));
+        }
+        let persisted: PersistedIlmRecoveryExport = serde_json::from_slice(data).map_err(Error::other)?;
+        if persisted.schema != ILM_RECOVERY_EXPORT_SCHEMA {
+            return Err(Error::other("ILM recovery export schema is unsupported"));
+        }
+        validate_sha256(&persisted.content_sha256, "ILM recovery export checksum is invalid")?;
+        let export_bytes = serde_json::to_vec(&persisted.export).map_err(Error::other)?;
+        if hex_sha256(&export_bytes, ToOwned::to_owned) != persisted.content_sha256 {
+            return Err(Error::other("ILM recovery export checksum mismatch"));
+        }
+        persisted.export.validate()?;
+        if persisted.export.export_id != expected_export_id {
+            return Err(Error::other("ILM recovery export ID does not match record key"));
+        }
+        Ok(persisted.export)
+    }
+}
+
+pub fn recovery_export_record_object_name(protocol: IlmRecoveryProtocol, export_id: &str) -> Result<String> {
+    validate_sha256(export_id, "ILM recovery export ID is invalid")?;
+    Ok(format!(
+        "{}/{}/{}/{}/{}.json",
+        ILM_RECOVERY_EXPORT_PREFIX,
+        protocol.as_str(),
+        &export_id[..2],
+        &export_id[2..4],
+        export_id
+    ))
+}
+
+pub fn recovery_export_id_from_record_object_name(object: &str) -> Result<(IlmRecoveryProtocol, String)> {
+    let suffix = object
+        .strip_prefix(ILM_RECOVERY_EXPORT_PREFIX)
+        .and_then(|suffix| suffix.strip_prefix('/'))
+        .ok_or_else(|| Error::other("ILM recovery export path has wrong prefix"))?;
+    let mut parts = suffix.split('/');
+    let protocol = match parts.next() {
+        Some("tier_delete_journal") => IlmRecoveryProtocol::TierDeleteJournal,
+        _ => return Err(Error::other("ILM recovery export protocol is invalid")),
+    };
+    let shard_a = parts
+        .next()
+        .ok_or_else(|| Error::other("ILM recovery export path is incomplete"))?;
+    let shard_b = parts
+        .next()
+        .ok_or_else(|| Error::other("ILM recovery export path is incomplete"))?;
+    let export_id = parts
+        .next()
+        .and_then(|name| name.strip_suffix(".json"))
+        .ok_or_else(|| Error::other("ILM recovery export suffix is invalid"))?;
+    if parts.next().is_some() {
+        return Err(Error::other("ILM recovery export path is not canonical"));
+    }
+    validate_sha256(export_id, "ILM recovery export ID is invalid")?;
+    if shard_a != &export_id[..2] || shard_b != &export_id[2..4] {
+        return Err(Error::other("ILM recovery export shard does not match export ID"));
+    }
+    Ok((protocol, export_id.to_string()))
+}
+
+pub async fn inspect_recovery_export_observation(api: Arc<ECStore>, control_id: &str) -> Result<IlmRecoveryExportObservation> {
+    let proof = acquire_ilm_recovery_export_fleet_proof()
+        .await
+        .ok_or_else(|| Error::other("ILM recovery export fleet proof is unavailable"))?;
+    let observed_control = load_exportable_control(api.clone(), control_id).await?;
+    let observed_source = observe_export_source(
+        api,
+        &observed_control.control.identity.canonical_source_path,
+        &observed_control.control.observed_source_generation.source_schema,
+    )
+    .await?;
+    if !observed_source.is_consistent()
+        || observed_source.generation != observed_control.control.observed_source_generation
+        || !ilm_recovery_export_fleet_proof_matches(&proof).await
+    {
+        return Err(Error::other("ILM recovery export observation changed or is incomplete"));
+    }
+    Ok(IlmRecoveryExportObservation {
+        control_id: control_id.to_string(),
+        protocol: observed_control.control.identity.protocol,
+        control_etag: observed_control.etag,
+        control_revision: observed_control.control.revision,
+        classification: observed_control.control.classification,
+        canonical_source_path: observed_control.control.identity.canonical_source_path,
+        source_generation: observed_source.generation,
+        topology_generation: ilm_recovery_export_topology_generation(&proof),
+        member_epochs_sha256: ilm_recovery_export_member_epochs_sha256(&proof),
+    })
+}
+
+pub async fn create_recovery_export(
+    api: Arc<ECStore>,
+    observation: &IlmRecoveryExportObservation,
+    creator_sha256: &str,
+) -> Result<IlmRecoveryExportCreated> {
+    validate_sha256(creator_sha256, "ILM recovery export creator digest is invalid")?;
+    let lock = api.new_ns_lock(RUSTFS_META_BUCKET, EXPORT_ADMISSION_LOCK).await?;
+    let admission_guard = lock.get_write_lock(crate::set_disk::get_lock_acquire_timeout()).await?;
+
+    let proof = acquire_ilm_recovery_export_fleet_proof()
+        .await
+        .ok_or_else(|| Error::other("ILM recovery export fleet proof is unavailable"))?;
+    if ilm_recovery_export_topology_generation(&proof) != observation.topology_generation
+        || ilm_recovery_export_member_epochs_sha256(&proof) != observation.member_epochs_sha256
+    {
+        return Err(Error::PreconditionFailed);
+    }
+    let control_object = recovery_control_record_object_name(IlmRecoveryProtocol::TierDeleteJournal, &observation.control_id)
+        .map_err(Error::other)?;
+    let control_lock = api.new_ns_lock(RUSTFS_META_BUCKET, &control_object).await?;
+    let control_guard = control_lock
+        .get_read_lock(crate::set_disk::get_lock_acquire_timeout())
+        .await?;
+    let source_lock = api
+        .new_ns_lock(RUSTFS_META_BUCKET, &observation.canonical_source_path)
+        .await?;
+    let source_guard = source_lock.get_read_lock(crate::set_disk::get_lock_acquire_timeout()).await?;
+    let locks_current = || !admission_guard.is_lock_lost() && !control_guard.is_lock_lost() && !source_guard.is_lock_lost();
+    let (current, current_source_bytes) = current_observation_under_proof_no_lock(api.clone(), observation, &proof).await?;
+    if &current != observation || !locks_current() {
+        return Err(Error::PreconditionFailed);
+    }
+    let current_source_base64 = base64_simd::STANDARD.encode_to_string(current_source_bytes);
+    let candidate_export_id = recovery_export_id(&current.control_id, &current.source_generation)?;
+    let object = recovery_export_record_object_name(current.protocol, &candidate_export_id)?;
+    match load_recovery_export_decoded(api.clone(), &candidate_export_id).await {
+        Ok((existing, export)) if export_matches_observation(&export, observation) => {
+            if !locks_current() || !ilm_recovery_export_fleet_proof_matches(&proof).await {
+                return Err(Error::PreconditionFailed);
+            }
+            api.record_durable_ilm_decommission_progress(&object, &existing.encoded)
+                .await?;
+            if !locks_current() {
+                return Err(Error::PreconditionFailed);
+            }
+            return Ok(existing.with_replayed());
+        }
+        Ok(_) => return Err(Error::PreconditionFailed),
+        Err(Error::ConfigNotFound) => {}
+        Err(err) => return Err(err),
+    }
+    let inventory = collect_export_inventory(api.clone()).await?;
+    if !locks_current() || !ilm_recovery_export_fleet_proof_matches(&proof).await {
+        return Err(Error::PreconditionFailed);
+    }
+    let created_at_unix_nanos = now_unix_nanos()?;
+    let export = build_export_from_source(&current, creator_sha256, created_at_unix_nanos, &current_source_base64)?;
+    let encoded = export.encode()?;
+    inventory.check(creator_sha256, encoded.len(), created_at_unix_nanos)?;
+
+    let mut write_options = ObjectOptions {
+        max_parity: true,
+        write_completion: WriteCompletion::TailDrained,
+        http_preconditions: Some(HTTPPreconditions {
+            if_none_match: Some("*".to_string()),
+            ..Default::default()
+        }),
+        ..Default::default()
+    };
+    write_options.add_namespace_lock_guard(&admission_guard);
+    write_options.add_namespace_lock_guard(&control_guard);
+    write_options.add_namespace_lock_guard(&source_guard);
+    if !locks_current() {
+        return Err(Error::PreconditionFailed);
+    }
+    let write_result = config_boundary::save_config_with_opts(api.clone(), &object, encoded.clone(), &write_options).await;
+    let stored = match load_recovery_export(api.clone(), &export.export_id).await {
+        Ok(stored) if stored.encoded == encoded => stored,
+        Ok(_) => return Err(Error::PreconditionFailed),
+        Err(read_err) => return Err(write_result.err().unwrap_or(read_err)),
+    };
+    if !locks_current() || !ilm_recovery_export_fleet_proof_matches(&proof).await {
+        return Err(Error::PreconditionFailed);
+    }
+    api.record_durable_ilm_decommission_progress(&object, &encoded).await?;
+    if !locks_current() {
+        return Err(Error::PreconditionFailed);
+    }
+    Ok(stored)
+}
+
+pub async fn load_recovery_export(api: Arc<ECStore>, export_id: &str) -> Result<IlmRecoveryExportCreated> {
+    let (created, _) = load_recovery_export_decoded(api, export_id).await?;
+    Ok(created)
+}
+
+async fn load_recovery_export_decoded(
+    api: Arc<ECStore>,
+    export_id: &str,
+) -> Result<(IlmRecoveryExportCreated, IlmRecoveryExport)> {
+    let object = recovery_export_record_object_name(IlmRecoveryProtocol::TierDeleteJournal, export_id)?;
+    let encoded = config_boundary::read_config_limited_preserve_empty(api, &object, MAX_ILM_RECOVERY_EXPORT_SIZE).await?;
+    let export = IlmRecoveryExport::decode(export_id, &encoded)?;
+    let content_sha256 = hex_sha256(&encoded, ToOwned::to_owned);
+    Ok((
+        IlmRecoveryExportCreated {
+            export_id: export.export_id.clone(),
+            content_sha256,
+            encoded,
+            replayed: false,
+        },
+        export,
+    ))
+}
+
+impl IlmRecoveryExportCreated {
+    fn with_replayed(mut self) -> Self {
+        self.replayed = true;
+        self
+    }
+}
+
+async fn load_exportable_control(api: Arc<ECStore>, control_id: &str) -> Result<ObservedIlmRecoveryControl> {
+    load_exportable_control_with_options(api, control_id, &ObjectOptions::default()).await
+}
+
+async fn load_exportable_control_no_lock(api: Arc<ECStore>, control_id: &str) -> Result<ObservedIlmRecoveryControl> {
+    load_exportable_control_with_options(
+        api,
+        control_id,
+        &ObjectOptions {
+            no_lock: true,
+            ..Default::default()
+        },
+    )
+    .await
+}
+
+async fn load_exportable_control_with_options(
+    api: Arc<ECStore>,
+    control_id: &str,
+    options: &ObjectOptions,
+) -> Result<ObservedIlmRecoveryControl> {
+    let object = recovery_control_record_object_name(IlmRecoveryProtocol::TierDeleteJournal, control_id).map_err(Error::other)?;
+    let (data, metadata) =
+        config_boundary::read_config_limited_preserve_empty_with_metadata(api, &object, options, MAX_ILM_RECOVERY_CONTROL_SIZE)
+            .await?;
+    let etag = metadata
+        .etag
+        .filter(|etag| !etag.trim().is_empty())
+        .ok_or_else(|| Error::other("ILM recovery control is missing an ETag"))?;
+    let control = IlmRecoveryControl::decode(control_id, &data).map_err(Error::other)?;
+    if control.identity.protocol != IlmRecoveryProtocol::TierDeleteJournal
+        || control.classification != IlmRecoveryClassification::RetainedAmbiguous
+        || !is_legacy_export_schema(&control.observed_source_generation.source_schema)
+    {
+        return Err(Error::other("ILM recovery control is not exportable"));
+    }
+    Ok(ObservedIlmRecoveryControl { control, etag })
+}
+
+async fn current_observation_under_proof_no_lock(
+    api: Arc<ECStore>,
+    expected: &IlmRecoveryExportObservation,
+    proof: &crate::services::notification_sys::IlmRecoveryExportFleetProofToken,
+) -> Result<(IlmRecoveryExportObservation, Vec<u8>)> {
+    let observed_control = load_exportable_control_no_lock(api.clone(), &expected.control_id).await?;
+    let observed_source = observe_export_source_no_lock(
+        api,
+        &observed_control.control.identity.canonical_source_path,
+        &observed_control.control.observed_source_generation.source_schema,
+    )
+    .await?;
+    let source_bytes = observed_source
+        .canonical_data
+        .clone()
+        .ok_or_else(|| Error::other("ILM recovery export source copies diverge"))?;
+    if !observed_source.is_consistent()
+        || observed_source.generation != observed_control.control.observed_source_generation
+        || !ilm_recovery_export_fleet_proof_matches(proof).await
+    {
+        return Err(Error::PreconditionFailed);
+    }
+    Ok((
+        IlmRecoveryExportObservation {
+            control_id: expected.control_id.clone(),
+            protocol: observed_control.control.identity.protocol,
+            control_etag: observed_control.etag,
+            control_revision: observed_control.control.revision,
+            classification: observed_control.control.classification,
+            canonical_source_path: observed_control.control.identity.canonical_source_path,
+            source_generation: observed_source.generation,
+            topology_generation: ilm_recovery_export_topology_generation(proof),
+            member_epochs_sha256: ilm_recovery_export_member_epochs_sha256(proof),
+        },
+        source_bytes,
+    ))
+}
+
+async fn observe_export_source(
+    api: Arc<ECStore>,
+    canonical_path: &str,
+    source_schema: &str,
+) -> Result<ObservedIlmRecoverySource> {
+    if canonical_path.is_empty()
+        || canonical_path.starts_with('/')
+        || canonical_path.ends_with('/')
+        || canonical_path.split('/').any(str::is_empty)
+        || !is_legacy_export_schema(source_schema)
+    {
+        return Err(Error::other("ILM recovery export source identity is invalid"));
+    }
+    let lock = api.new_ns_lock(RUSTFS_META_BUCKET, canonical_path).await?;
+    let _guard = lock.get_read_lock(crate::set_disk::get_lock_acquire_timeout()).await?;
+    observe_export_source_no_lock(api, canonical_path, source_schema).await
+}
+
+async fn observe_export_source_no_lock(
+    api: Arc<ECStore>,
+    canonical_path: &str,
+    source_schema: &str,
+) -> Result<ObservedIlmRecoverySource> {
+    let mut copies = Vec::new();
+    let mut canonical: Option<(String, String, Vec<u8>)> = None;
+    let mut consistent = true;
+    for set in api.all_set_disks() {
+        let authority = format!("pool-{}/set-{}", set.pool_index, set.set_index);
+        let result = config_boundary::read_config_limited_preserve_empty_with_metadata(
+            set,
+            canonical_path,
+            &ObjectOptions {
+                no_lock: true,
+                ..Default::default()
+            },
+            MAX_LEGACY_TIER_DELETE_SOURCE_SIZE,
+        )
+        .await;
+        match result {
+            Ok((data, metadata)) => {
+                if data.is_empty() || data.len() > MAX_LEGACY_TIER_DELETE_SOURCE_SIZE {
+                    return Err(Error::other("ILM recovery export source exceeds its protocol size limit"));
+                }
+                validate_legacy_tier_delete_recovery_source(canonical_path, source_schema, &data)?;
+                let etag = metadata
+                    .etag
+                    .filter(|etag| !etag.trim().is_empty())
+                    .ok_or_else(|| Error::other("ILM recovery export source copy is missing an ETag"))?;
+                let content_sha256 = hex_sha256(&data, ToOwned::to_owned);
+                let encoded_len =
+                    u64::try_from(data.len()).map_err(|_| Error::other("ILM recovery export source length does not fit u64"))?;
+                copies.push(IlmRecoverySourceCopy {
+                    authority,
+                    canonical_path: canonical_path.to_string(),
+                    etag: etag.clone(),
+                    encoded_len,
+                    content_sha256: content_sha256.clone(),
+                });
+                match canonical.as_ref() {
+                    Some((first_etag, first_digest, first_data)) => {
+                        consistent &= first_etag == &etag && first_digest == &content_sha256 && first_data == &data;
+                    }
+                    None => canonical = Some((etag, content_sha256, data)),
+                }
+            }
+            Err(err) if export_source_is_missing(&err) => {}
+            Err(err) => return Err(err),
+        }
+    }
+    let Some((source_etag, content_sha256, source_bytes)) = canonical else {
+        return Err(Error::ConfigNotFound);
+    };
+    let generation =
+        IlmRecoverySourceGeneration::new(source_schema, source_etag, content_sha256, copies).map_err(Error::other)?;
+    Ok(ObservedIlmRecoverySource {
+        generation,
+        canonical_data: consistent.then_some(source_bytes),
+    })
+}
+
+fn export_source_is_missing(err: &Error) -> bool {
+    matches!(
+        err,
+        Error::ConfigNotFound | Error::FileNotFound | Error::ObjectNotFound(_, _) | Error::VersionNotFound(_, _, _)
+    )
+}
+
+fn build_export_from_source(
+    observation: &IlmRecoveryExportObservation,
+    creator_sha256: &str,
+    created_at_unix_nanos: i64,
+    source_bytes_base64: &str,
+) -> Result<IlmRecoveryExport> {
+    let retain_until_unix_nanos = created_at_unix_nanos
+        .checked_add(EXPORT_RETENTION_NANOS)
+        .ok_or_else(|| Error::other("ILM recovery export retention timestamp overflow"))?;
+    let export = IlmRecoveryExport {
+        export_id: recovery_export_id(&observation.control_id, &observation.source_generation)?,
+        control_id: observation.control_id.clone(),
+        protocol: observation.protocol,
+        control_etag: observation.control_etag.clone(),
+        control_revision: observation.control_revision,
+        classification: observation.classification,
+        canonical_source_path: observation.canonical_source_path.clone(),
+        source_generation: observation.source_generation.clone(),
+        topology_generation: observation.topology_generation.clone(),
+        member_epochs_sha256: observation.member_epochs_sha256.clone(),
+        creator_sha256: creator_sha256.to_string(),
+        created_at_unix_nanos,
+        retain_until_unix_nanos,
+        source_bytes_base64: source_bytes_base64.to_string(),
+    };
+    export.validate()?;
+    Ok(export)
+}
+
+pub(crate) fn recovery_export_id(control_id: &str, generation: &IlmRecoverySourceGeneration) -> Result<String> {
+    validate_sha256(control_id, "ILM recovery export control ID is invalid")?;
+    validate_sha256(&generation.content_sha256, "ILM recovery export source checksum is invalid")?;
+    validate_sha256(&generation.copy_set_sha256, "ILM recovery export copy-set checksum is invalid")?;
+    let mut data = Vec::new();
+    for part in [control_id, &generation.content_sha256, &generation.copy_set_sha256] {
+        data.extend_from_slice(&(part.len() as u64).to_be_bytes());
+        data.extend_from_slice(part.as_bytes());
+    }
+    Ok(hex_sha256(&data, ToOwned::to_owned))
+}
+
+fn export_matches_observation(export: &IlmRecoveryExport, observation: &IlmRecoveryExportObservation) -> bool {
+    export.control_id == observation.control_id
+        && export.protocol == observation.protocol
+        && export.classification == observation.classification
+        && export.canonical_source_path == observation.canonical_source_path
+        && export.source_generation == observation.source_generation
+}
+
+#[derive(Debug, Default)]
+struct IlmRecoveryExportInventory {
+    count: usize,
+    bytes: u64,
+    creations: Vec<(i64, String)>,
+}
+
+impl IlmRecoveryExportInventory {
+    fn check(&self, creator_sha256: &str, candidate_len: usize, now: i64) -> Result<()> {
+        let recent_after = now.saturating_sub(60 * 1_000_000_000);
+        let cluster_recent = self
+            .creations
+            .iter()
+            .filter(|(created_at, _)| *created_at > recent_after)
+            .count();
+        let actor_recent = self
+            .creations
+            .iter()
+            .filter(|(created_at, creator)| *created_at > recent_after && creator == creator_sha256)
+            .count();
+        check_export_admission(self.count, self.bytes, actor_recent, cluster_recent, candidate_len)
+    }
+}
+
+async fn collect_export_inventory(api: Arc<ECStore>) -> Result<IlmRecoveryExportInventory> {
+    let mut marker = None;
+    let mut seen_markers = HashSet::new();
+    let mut inventory = IlmRecoveryExportInventory::default();
+    loop {
+        let page = api
+            .clone()
+            .list_objects_v2(
+                RUSTFS_META_BUCKET,
+                &format!("{ILM_RECOVERY_EXPORT_PREFIX}/"),
+                marker.clone(),
+                None,
+                1_000,
+                false,
+                None,
+                false,
+            )
+            .await?;
+        for object in page.objects {
+            let (_, export_id) = recovery_export_id_from_record_object_name(&object.name)?;
+            let (stored, export) = load_recovery_export_decoded(api.clone(), &export_id).await?;
+            inventory.count = inventory
+                .count
+                .checked_add(1)
+                .ok_or_else(|| Error::other("ILM recovery export count overflow"))?;
+            inventory.bytes = inventory
+                .bytes
+                .checked_add(u64::try_from(stored.encoded.len()).map_err(|_| Error::other("ILM recovery export size overflow"))?)
+                .ok_or_else(|| Error::other("ILM recovery export byte total overflow"))?;
+            inventory
+                .creations
+                .push((export.created_at_unix_nanos, export.creator_sha256));
+        }
+        if !page.is_truncated {
+            break;
+        }
+        let next = page
+            .next_continuation_token
+            .ok_or_else(|| Error::other("ILM recovery export inventory omitted its continuation marker"))?;
+        marker = Some(record_export_inventory_marker(&mut seen_markers, next)?);
+    }
+    Ok(inventory)
+}
+
+fn record_export_inventory_marker(seen_markers: &mut HashSet<String>, next: String) -> Result<String> {
+    if !seen_markers.insert(next.clone()) {
+        return Err(Error::other("ILM recovery export inventory repeated its continuation marker"));
+    }
+    Ok(next)
+}
+
+fn check_export_admission(
+    count: usize,
+    bytes: u64,
+    actor_recent: usize,
+    cluster_recent: usize,
+    candidate_len: usize,
+) -> Result<()> {
+    let candidate_len = u64::try_from(candidate_len).map_err(|_| Error::other("ILM recovery export size does not fit u64"))?;
+    if count >= MAX_ILM_RECOVERY_EXPORTS
+        || bytes
+            .checked_add(candidate_len)
+            .is_none_or(|total| total > MAX_ILM_RECOVERY_EXPORT_BYTES)
+        || actor_recent >= MAX_ACTOR_EXPORTS_PER_MINUTE
+        || cluster_recent >= MAX_CLUSTER_EXPORTS_PER_MINUTE
+    {
+        return Err(Error::SlowDown);
+    }
+    Ok(())
+}
+
+fn is_legacy_export_schema(schema: &str) -> bool {
+    matches!(schema, TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA | TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA)
+}
+
+fn validate_sha256(value: &str, message: &'static str) -> Result<()> {
+    if !is_sha256_checksum(value)
+        || value
+            .bytes()
+            .any(|byte| byte.is_ascii_hexdigit() && byte.is_ascii_uppercase())
+    {
+        return Err(Error::other(message));
+    }
+    Ok(())
+}
+
+fn now_unix_nanos() -> Result<i64> {
+    i64::try_from(time::OffsetDateTime::now_utc().unix_timestamp_nanos())
+        .map_err(|_| Error::other("ILM recovery export timestamp does not fit i64"))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::bucket::lifecycle::recovery_control::IlmRecoverySourceCopy;
+
+    const PINNED_V1_EXPORT: &[u8] = br#"{"schema":"rustfs-ilm-recovery-export-v1","content_sha256":"3dfb3ec3892256e909de1211c1a963ca7008963ff32b3a869f7161a7b9b44028","export":{"export_id":"2b78e7a825bfc2edbf7f773d0b6ed3bf93e360ff1702d73a449109c11bfaa105","control_id":"0fcd568a5cb9bdb4677b69354b11ee415af8f784519cff3da49a26f84eaee7f2","protocol":"tier_delete_journal","control_etag":"control-etag","control_revision":1,"classification":"retained_ambiguous","canonical_source_path":"ilm/tier-delete-journal/872072554f66ab326f10ce7adbae11422b7a4b0663aa7112d6061a8f6ed41b94.json","source_generation":{"source_schema":"rustfs-tier-delete-journal-v1","source_etag":"etag-a","content_sha256":"0e0b010ebdeeb7b41473fe8575e989d6bb1303c0ca551dd984e9400f0ae306bd","copy_set_sha256":"5a7406115b6c3923ffe79dcd1f43ccae7beed786e557163f019dd10ec409a653","copies":[{"authority":"pool-0/set-0","canonical_path":"ilm/tier-delete-journal/872072554f66ab326f10ce7adbae11422b7a4b0663aa7112d6061a8f6ed41b94.json","etag":"etag-a","encoded_len":81,"content_sha256":"0e0b010ebdeeb7b41473fe8575e989d6bb1303c0ca551dd984e9400f0ae306bd"}]},"topology_generation":"e6e2b826e31fca5c36125c48f130dcb6f961e698ff8a8776a1f290cf0892e8e6","member_epochs_sha256":"612dd8a861161819a4ad8f6f3e2a0567602877c043a2353ca933a13c78dc0ed4","creator_sha256":"50c9c4aeb40b5b206b6d98f516f8b8c0efd29ce2e56a76b345fb9240c225a1b7","created_at_unix_nanos":1000000000,"retain_until_unix_nanos":7776001000000000,"source_bytes_base64":"eyJ2ZXJzaW9uIjoxLCJvYmpfbmFtZSI6ImxlZ2FjeS9yZW1vdGUiLCJ2ZXJzaW9uX2lkIjoib3BhcXVlIiwidGllcl9uYW1lIjoiV0FSTSJ9"}}"#;
+
+    fn legacy_source() -> Vec<u8> {
+        br#"{"version":1,"obj_name":"legacy/remote","version_id":"opaque","tier_name":"WARM"}"#.to_vec()
+    }
+
+    fn observation() -> IlmRecoveryExportObservation {
+        let source = legacy_source();
+        let source_path = super::super::tier_delete_journal::tier_delete_journal_object_name(
+            &super::super::tier_delete_journal::decode_tier_delete_journal_entry(&source).expect("legacy fixture should decode"),
+        );
+        let source_sha256 = hex_sha256(&source, ToOwned::to_owned);
+        let generation = IlmRecoverySourceGeneration::new(
+            TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA,
+            "etag-a",
+            source_sha256.clone(),
+            vec![IlmRecoverySourceCopy {
+                authority: "pool-0/set-0".to_string(),
+                canonical_path: source_path.clone(),
+                etag: "etag-a".to_string(),
+                encoded_len: source.len() as u64,
+                content_sha256: source_sha256,
+            }],
+        )
+        .expect("generation should be valid");
+        IlmRecoveryExportObservation {
+            control_id: hex_sha256(b"control", ToOwned::to_owned),
+            protocol: IlmRecoveryProtocol::TierDeleteJournal,
+            control_etag: "control-etag".to_string(),
+            control_revision: 1,
+            classification: IlmRecoveryClassification::RetainedAmbiguous,
+            canonical_source_path: source_path,
+            source_generation: generation,
+            topology_generation: hex_sha256(b"topology", ToOwned::to_owned),
+            member_epochs_sha256: hex_sha256(b"epochs", ToOwned::to_owned),
+        }
+    }
+
+    #[test]
+    fn recovery_export_round_trip_is_strict_and_deterministic() {
+        let observed = observation();
+        let creator = hex_sha256(b"actor", ToOwned::to_owned);
+        let export = build_export_from_source(
+            &observed,
+            &creator,
+            1_000_000_000,
+            &base64_simd::STANDARD.encode_to_string(legacy_source()),
+        )
+        .expect("export should be valid");
+        assert_eq!(
+            export.export_id,
+            recovery_export_id(&observed.control_id, &observed.source_generation).unwrap()
+        );
+        let encoded = export.encode().expect("export should encode");
+        assert_eq!(encoded, PINNED_V1_EXPORT, "v1 export wire format must remain pinned");
+        assert_eq!(IlmRecoveryExport::decode(&export.export_id, &encoded).unwrap(), export);
+        assert_eq!(
+            IlmRecoveryExport::decode("2b78e7a825bfc2edbf7f773d0b6ed3bf93e360ff1702d73a449109c11bfaa105", PINNED_V1_EXPORT)
+                .unwrap(),
+            export,
+        );
+
+        let path = recovery_export_record_object_name(export.protocol, &export.export_id).unwrap();
+        let durable = super::super::durable_namespace::validate_durable_ilm_record(&path, &encoded)
+            .expect("export should be registered as a durable ILM record");
+        assert_eq!(durable.namespace, "recovery-export");
+        assert_eq!(durable.id_kind, "export_id");
+        assert_eq!(durable.id, export.export_id);
+
+        let mut wrong_source = export.clone();
+        wrong_source.source_bytes_base64 = base64_simd::STANDARD.encode_to_string(b"changed");
+        assert!(wrong_source.encode().is_err());
+
+        let mut persisted: serde_json::Value = serde_json::from_slice(&encoded).unwrap();
+        persisted["unknown"] = serde_json::json!(true);
+        assert!(IlmRecoveryExport::decode(&export.export_id, &serde_json::to_vec(&persisted).unwrap()).is_err());
+    }
+
+    #[test]
+    fn export_inventory_rejects_non_adjacent_continuation_cycles() {
+        let mut seen = HashSet::new();
+        assert_eq!(record_export_inventory_marker(&mut seen, "a".to_string()).unwrap(), "a");
+        assert_eq!(record_export_inventory_marker(&mut seen, "b".to_string()).unwrap(), "b");
+        record_export_inventory_marker(&mut seen, "a".to_string())
+            .expect_err("a non-adjacent continuation marker cycle must fail closed");
+    }
+
+    #[test]
+    fn recovery_export_path_rejects_noncanonical_shards() {
+        let id = hex_sha256(b"export", ToOwned::to_owned);
+        let path = recovery_export_record_object_name(IlmRecoveryProtocol::TierDeleteJournal, &id).unwrap();
+        assert_eq!(recovery_export_id_from_record_object_name(&path).unwrap().1, id);
+        let wrong_shard = path.replacen(&format!("/{}/", &id[..2]), "/zz/", 1);
+        assert!(recovery_export_id_from_record_object_name(&wrong_shard).is_err());
+    }
+
+    #[test]
+    fn canonical_replay_survives_fleet_rotation_but_not_source_change() {
+        let observed = observation();
+        let creator = hex_sha256(b"actor", ToOwned::to_owned);
+        let export = build_export_from_source(
+            &observed,
+            &creator,
+            1_000_000_000,
+            &base64_simd::STANDARD.encode_to_string(legacy_source()),
+        )
+        .unwrap();
+        let mut rotated = observed;
+        rotated.control_etag = "new-control-etag".to_string();
+        rotated.control_revision += 1;
+        rotated.topology_generation = hex_sha256(b"new-topology", ToOwned::to_owned);
+        rotated.member_epochs_sha256 = hex_sha256(b"new-members", ToOwned::to_owned);
+        assert!(export_matches_observation(&export, &rotated));
+
+        rotated.source_generation.content_sha256 = hex_sha256(b"changed", ToOwned::to_owned);
+        assert!(!export_matches_observation(&export, &rotated));
+    }
+
+    #[test]
+    fn export_admission_enforces_exact_count_byte_and_rate_boundaries() {
+        assert!(check_export_admission(9_999, MAX_ILM_RECOVERY_EXPORT_BYTES - 1, 9, 99, 1).is_ok());
+        assert!(check_export_admission(10_000, 0, 0, 0, 1).is_err());
+        assert!(check_export_admission(0, MAX_ILM_RECOVERY_EXPORT_BYTES, 0, 0, 1).is_err());
+        assert!(check_export_admission(0, 0, 10, 0, 1).is_err());
+        assert!(check_export_admission(0, 0, 0, 100, 1).is_err());
+    }
+}

crates/ecstore/src/bucket/lifecycle/tier_delete_journal.rs

@@ -35,6 +35,10 @@ use crate::bucket::lifecycle::config_boundary;
 use crate::bucket::lifecycle::durable_namespace::{
     TIER_DELETE_JOURNAL_NAMESPACE, TIER_DELETE_JOURNAL_V6_NAMESPACE, validate_durable_ilm_record,
 };
+use crate::bucket::lifecycle::recovery_control::{
+    IlmRecoveryClassification, IlmRecoveryControl, IlmRecoveryControlIdentity, IlmRecoveryErrorCode, IlmRecoveryProtocol,
+    load_recovery_control, observe_recovery_source, save_recovery_control_if_absent,
+};
 use crate::bucket::lifecycle::runtime_boundary;
 use crate::bucket::lifecycle::tier_sweeper::{
     Jentry, TierDeleteDispatchBinding, TierDeleteJournalState, TierDeleteSourceIdentity,
@@ -78,6 +82,13 @@ const TIER_DELETE_DISPATCH_MEMBER_DELETE_CONCURRENCY: usize = 32;
 const TIER_DELETE_DISPATCH_PREPARE_CONCURRENCY: usize = 16;
 const TIER_DELETE_DISPATCH_CAS_CONCURRENCY: usize = 32;
 const TIER_DELETE_JOURNAL_VERSION: u8 = 2;
+pub(crate) const TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA: &str = "rustfs-tier-delete-journal-v1";
+pub(crate) const TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA: &str = "rustfs-tier-delete-journal-v2";
+const TIER_DELETE_JOURNAL_UNKNOWN_RECOVERY_SCHEMA: &str = "rustfs-tier-delete-journal-unknown";
+const TIER_DELETE_JOURNAL_V1_RECOVERY_CLASS: &str = "tier_delete_journal_v1";
+const TIER_DELETE_JOURNAL_V2_RECOVERY_CLASS: &str = "tier_delete_journal_v2";
+const TIER_DELETE_JOURNAL_CORRUPT_RECOVERY_CLASS: &str = "tier_delete_journal_corrupt";
+const CORRUPT_TIER_DELETE_JOURNAL_IDENTITY: &str = "corrupt";
 const TIER_DELETE_JOURNAL_EXACT_VERSION: u8 = 3;
 const TIER_DELETE_JOURNAL_STATE_VERSION: u8 = 4;
 const TIER_DELETE_JOURNAL_TRANSACTION_VERSION: u8 = 5;
@@ -873,6 +884,23 @@ struct PersistedTierDeleteJournalEntry {
 }
 
 impl PersistedTierDeleteJournalEntry {
+    fn validate_legacy_recovery_shape(&self) -> Result<()> {
+        let has_later_version_fields = self.version_id_exact.is_some()
+            || self.version_state.is_some()
+            || self.state.is_some()
+            || self.source.is_some()
+            || self.dispatch.is_some();
+        match self.version {
+            1 if self.backend_identity.is_none() && !has_later_version_fields => Ok(()),
+            TIER_DELETE_JOURNAL_VERSION if self.backend_identity.is_some() && !has_later_version_fields => Ok(()),
+            1 => Err(Error::other("tier delete journal v1 entry contains fields from a later version")),
+            TIER_DELETE_JOURNAL_VERSION => Err(Error::other(
+                "tier delete journal v2 entry is missing its identity or contains fields from a later version",
+            )),
+            _ => Err(Error::other("tier delete journal is not an exportable legacy version")),
+        }
+    }
+
     fn from_jentry(je: &Jentry) -> Result<Self> {
         validate_version_state(je.version_state, &je.version_id, je.version_id_exact)?;
         let legacy_unknown = je.version_state == rustfs_filemeta::TransitionVersionState::Unknown;
@@ -5509,6 +5537,146 @@ enum TierDeleteJournalEntryRecoveryOutcome {
     Failed,
 }
 
+fn canonical_legacy_tier_delete_journal_identity(object_name: &str) -> Option<&str> {
+    let identity = object_name
+        .strip_prefix(TIER_DELETE_JOURNAL_LEGACY_PREFIX)?
+        .strip_suffix(".json")?;
+    (rustfs_utils::crypto::is_sha256_checksum(identity)
+        && !identity
+            .bytes()
+            .any(|byte| byte.is_ascii_hexdigit() && byte.is_ascii_uppercase()))
+    .then_some(identity)
+}
+
+pub(crate) fn validate_legacy_tier_delete_recovery_path(object_name: &str) -> Result<()> {
+    canonical_legacy_tier_delete_journal_identity(object_name)
+        .map(|_| ())
+        .ok_or_else(|| Error::other("legacy tier delete journal path is not canonical"))
+}
+
+fn legacy_tier_delete_recovery_descriptor(entry: &Jentry) -> Option<(&'static str, &'static str)> {
+    match entry.persisted_version {
+        1 => Some((TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA, TIER_DELETE_JOURNAL_V1_RECOVERY_CLASS)),
+        TIER_DELETE_JOURNAL_VERSION => Some((TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA, TIER_DELETE_JOURNAL_V2_RECOVERY_CLASS)),
+        _ => None,
+    }
+}
+
+pub(crate) fn validate_legacy_tier_delete_recovery_source(object_name: &str, source_schema: &str, data: &[u8]) -> Result<()> {
+    validate_legacy_tier_delete_recovery_path(object_name)?;
+    let persisted: PersistedTierDeleteJournalEntry =
+        serde_json::from_slice(data).map_err(|err| Error::other_with_context("decode tier delete journal failed", err))?;
+    persisted.validate_legacy_recovery_shape()?;
+    let entry = persisted.into_jentry()?;
+    let Some((decoded_schema, _)) = legacy_tier_delete_recovery_descriptor(&entry) else {
+        return Err(Error::other("tier delete journal is not an exportable legacy version"));
+    };
+    if decoded_schema != source_schema || tier_delete_journal_object_name(&entry) != object_name {
+        return Err(Error::other("legacy tier delete journal identity does not match its recovery source"));
+    }
+    Ok(())
+}
+
+fn legacy_tier_delete_control_matches(
+    control: &IlmRecoveryControl,
+    identity: &IlmRecoveryControlIdentity,
+    generation: &crate::bucket::lifecycle::recovery_control::IlmRecoverySourceGeneration,
+    classification: IlmRecoveryClassification,
+    error_code: IlmRecoveryErrorCode,
+) -> bool {
+    control.identity == *identity
+        && control.observed_source_generation == *generation
+        && control.classification == classification
+        && control.last_error_code == error_code
+        && control.owner.is_none()
+        && control.attempt_count == 0
+        && control.consecutive_failure_count == 0
+}
+
+fn legacy_tier_delete_control_is_scheduler_fence(control: &IlmRecoveryControl, identity: &IlmRecoveryControlIdentity) -> bool {
+    control.identity == *identity && control.owner.is_none() && !control.classification.permits_automatic_attempt()
+}
+
+async fn persist_legacy_tier_delete_recovery_control(
+    api: Arc<ECStore>,
+    object_name: &str,
+    observed_data: &[u8],
+    stable_operation_identity: String,
+    (source_schema, record_class): (&'static str, &'static str),
+    intended_classification: IlmRecoveryClassification,
+    intended_error_code: IlmRecoveryErrorCode,
+) -> Result<()> {
+    let identity = IlmRecoveryControlIdentity {
+        protocol: IlmRecoveryProtocol::TierDeleteJournal,
+        canonical_source_path: object_name.to_string(),
+        stable_operation_identity,
+        record_class: record_class.to_string(),
+    };
+    let control_id = identity.source_operation_digest().map_err(Error::other)?;
+    match load_recovery_control(api.clone(), IlmRecoveryProtocol::TierDeleteJournal, &control_id).await {
+        Ok(observed) if legacy_tier_delete_control_is_scheduler_fence(&observed.control, &identity) => return Ok(()),
+        Ok(_) => return Err(Error::PreconditionFailed),
+        Err(Error::ConfigNotFound) => {}
+        Err(err) => return Err(err),
+    }
+
+    let source = observe_recovery_source(api.clone(), object_name, source_schema).await?;
+    let exact_source = source.is_consistent() && source.canonical_data.as_deref() == Some(observed_data);
+    let (classification, error_code) = if exact_source {
+        (intended_classification, intended_error_code)
+    } else {
+        (IlmRecoveryClassification::Corrupt, IlmRecoveryErrorCode::SourceDivergent)
+    };
+    let candidate = IlmRecoveryControl::new(
+        identity.clone(),
+        source.generation.clone(),
+        classification,
+        i64::try_from(time::OffsetDateTime::now_utc().unix_timestamp_nanos())
+            .map_err(|_| Error::other("tier delete journal recovery timestamp does not fit i64"))?,
+        error_code,
+    )
+    .map_err(Error::other)?;
+
+    match save_recovery_control_if_absent(api.clone(), &candidate).await {
+        Ok(()) | Err(Error::PreconditionFailed) => {}
+        Err(save_error) => match load_recovery_control(api.clone(), IlmRecoveryProtocol::TierDeleteJournal, &control_id).await {
+            Ok(observed)
+                if legacy_tier_delete_control_matches(
+                    &observed.control,
+                    &identity,
+                    &source.generation,
+                    classification,
+                    error_code,
+                ) =>
+            {
+                return Ok(());
+            }
+            Ok(_) | Err(_) => return Err(save_error),
+        },
+    }
+
+    let observed = load_recovery_control(api, IlmRecoveryProtocol::TierDeleteJournal, &control_id).await?;
+    if !legacy_tier_delete_control_matches(&observed.control, &identity, &source.generation, classification, error_code) {
+        return Err(Error::PreconditionFailed);
+    }
+    Ok(())
+}
+
+async fn retain_corrupt_legacy_tier_delete_journal(api: Arc<ECStore>, object_name: &str, data: &[u8]) -> Result<()> {
+    canonical_legacy_tier_delete_journal_identity(object_name)
+        .ok_or_else(|| Error::other("tier delete journal path is not canonical"))?;
+    persist_legacy_tier_delete_recovery_control(
+        api,
+        object_name,
+        data,
+        CORRUPT_TIER_DELETE_JOURNAL_IDENTITY.to_string(),
+        (TIER_DELETE_JOURNAL_UNKNOWN_RECOVERY_SCHEMA, TIER_DELETE_JOURNAL_CORRUPT_RECOVERY_CLASS),
+        IlmRecoveryClassification::Corrupt,
+        IlmRecoveryErrorCode::SourceCorrupt,
+    )
+    .await
+}
+
 async fn recover_tier_delete_journal_entry(api: Arc<ECStore>, object_name: String) -> TierDeleteJournalEntryRecoveryOutcome {
     let data = match config_boundary::read_config(api.clone(), &object_name).await {
         Ok(data) => data,
@@ -5529,6 +5697,22 @@ async fn recover_tier_delete_journal_entry(api: Arc<ECStore>, object_name: Strin
     let je = match decode_tier_delete_journal_entry(&data) {
         Ok(je) => je,
         Err(err) => {
+            if canonical_legacy_tier_delete_journal_identity(&object_name).is_some() {
+                return match retain_corrupt_legacy_tier_delete_journal(api, &object_name, &data).await {
+                    Ok(()) => TierDeleteJournalEntryRecoveryOutcome::Retained,
+                    Err(control_error) => {
+                        warn!(
+                            event = EVENT_LIFECYCLE_TIER_DELETE_JOURNAL,
+                            component = LOG_COMPONENT_ECSTORE,
+                            subsystem = LOG_SUBSYSTEM_LIFECYCLE,
+                            journal_object = %object_name,
+                            error = ?control_error,
+                            "Failed to retain corrupt tier delete journal recovery control"
+                        );
+                        TierDeleteJournalEntryRecoveryOutcome::Failed
+                    }
+                };
+            }
             warn!(
                 event = EVENT_LIFECYCLE_TIER_DELETE_JOURNAL,
                 component = LOG_COMPONENT_ECSTORE,
@@ -5542,6 +5726,22 @@ async fn recover_tier_delete_journal_entry(api: Arc<ECStore>, object_name: Strin
     };
 
     if tier_delete_journal_object_name(&je) != object_name {
+        if canonical_legacy_tier_delete_journal_identity(&object_name).is_some() {
+            return match retain_corrupt_legacy_tier_delete_journal(api, &object_name, &data).await {
+                Ok(()) => TierDeleteJournalEntryRecoveryOutcome::Retained,
+                Err(err) => {
+                    warn!(
+                        event = EVENT_LIFECYCLE_TIER_DELETE_JOURNAL,
+                        component = LOG_COMPONENT_ECSTORE,
+                        subsystem = LOG_SUBSYSTEM_LIFECYCLE,
+                        journal_object = %object_name,
+                        error = ?err,
+                        "Failed to retain mismatched tier delete journal recovery control"
+                    );
+                    TierDeleteJournalEntryRecoveryOutcome::Failed
+                }
+            };
+        }
         warn!(
             event = EVENT_LIFECYCLE_TIER_DELETE_JOURNAL,
             component = LOG_COMPONENT_ECSTORE,
@@ -5552,6 +5752,36 @@ async fn recover_tier_delete_journal_entry(api: Arc<ECStore>, object_name: Strin
         return TierDeleteJournalEntryRecoveryOutcome::Failed;
     }
 
+    if let Some((source_schema, record_class)) = legacy_tier_delete_recovery_descriptor(&je) {
+        let stable_operation_identity = canonical_legacy_tier_delete_journal_identity(&object_name)
+            .expect("decoded legacy journal path was validated against its canonical object name")
+            .to_string();
+        return match persist_legacy_tier_delete_recovery_control(
+            api,
+            &object_name,
+            &data,
+            stable_operation_identity,
+            (source_schema, record_class),
+            IlmRecoveryClassification::RetainedAmbiguous,
+            IlmRecoveryErrorCode::RemoteVersionUnknown,
+        )
+        .await
+        {
+            Ok(()) => TierDeleteJournalEntryRecoveryOutcome::Retained,
+            Err(err) => {
+                warn!(
+                    event = EVENT_LIFECYCLE_TIER_DELETE_JOURNAL,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_LIFECYCLE,
+                    journal_object = %object_name,
+                    error = ?err,
+                    "Failed to retain legacy tier delete journal recovery control"
+                );
+                TierDeleteJournalEntryRecoveryOutcome::Failed
+            }
+        };
+    }
+
     match api
         .durable_ilm_terminal_receipt_covers_active_source(&object_name, &data)
         .await
@@ -5948,17 +6178,18 @@ where
 #[cfg(test)]
 mod tests {
     use super::{
-        TIER_DELETE_DISPATCH_MANIFEST_VERSION, TIER_DELETE_DISPATCH_PARENT_RECORD_TYPE, TIER_DELETE_DISPATCH_PARENT_VERSION,
-        TIER_DELETE_JOURNAL_EXACT_VERSION, TIER_DELETE_JOURNAL_LEGACY_PREFIX, TIER_DELETE_JOURNAL_SOLE_OWNER_VERSION,
-        TIER_DELETE_JOURNAL_STATE_VERSION, TIER_DELETE_JOURNAL_TRANSACTION_VERSION, TIER_DELETE_JOURNAL_V6_PREFIX,
-        TierDeleteDispatchChunkBinding, TierDeleteDispatchManifest, TierDeleteDispatchManifestState, TierDeleteDispatchParent,
-        TierDeleteDispatchParentState, TierDeleteDispatchRecord, await_tier_delete_journal_recovery,
+        PersistedTierDeleteJournalEntry, TIER_DELETE_DISPATCH_MANIFEST_VERSION, TIER_DELETE_DISPATCH_PARENT_RECORD_TYPE,
+        TIER_DELETE_DISPATCH_PARENT_VERSION, TIER_DELETE_JOURNAL_EXACT_VERSION, TIER_DELETE_JOURNAL_LEGACY_PREFIX,
+        TIER_DELETE_JOURNAL_SOLE_OWNER_VERSION, TIER_DELETE_JOURNAL_STATE_VERSION, TIER_DELETE_JOURNAL_TRANSACTION_VERSION,
+        TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA, TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA, TIER_DELETE_JOURNAL_V6_PREFIX,
+        TIER_DELETE_JOURNAL_VERSION, TierDeleteDispatchChunkBinding, TierDeleteDispatchManifest, TierDeleteDispatchManifestState,
+        TierDeleteDispatchParent, TierDeleteDispatchParentState, TierDeleteDispatchRecord, await_tier_delete_journal_recovery,
         decode_tier_delete_dispatch_record, decode_tier_delete_journal_entry, encode_tier_delete_dispatch_manifest,
         encode_tier_delete_dispatch_parent, encode_tier_delete_journal_entry, object_info_references_tier_delete,
         record_tier_delete_journal_backend_identity, same_tier_delete_authorization_identity, same_tier_delete_journal_identity,
         tier_delete_dispatch_child_matches_parent, tier_delete_dispatch_chunk_manifest_object_name,
         tier_delete_dispatch_journal_set_digest, tier_delete_dispatch_manifest_object_name, tier_delete_journal_object_name,
-        tier_delete_source_matches_dispatch_scope,
+        tier_delete_source_matches_dispatch_scope, validate_legacy_tier_delete_recovery_source,
     };
     use crate::bucket::lifecycle::tier_sweeper::{
         Jentry, TierDeleteDispatchBinding, TierDeleteJournalState, TierDeleteSourceIdentity,
@@ -6417,6 +6648,72 @@ mod tests {
         }
     }
 
+    #[test]
+    fn legacy_recovery_export_rejects_fields_from_later_journal_versions() {
+        let later = bound_v6_journal_entry(TierDeleteJournalState::Prepared);
+        let v1 = PersistedTierDeleteJournalEntry {
+            version: 1,
+            obj_name: "remote/object".to_string(),
+            version_id: "opaque".to_string(),
+            tier_name: "WARM".to_string(),
+            backend_identity: None,
+            version_id_exact: None,
+            version_state: None,
+            state: None,
+            source: None,
+            dispatch: None,
+        };
+        let mut v2 = v1.clone();
+        v2.version = TIER_DELETE_JOURNAL_VERSION;
+        v2.backend_identity = Some([7; 32]);
+
+        let assert_rejected = |persisted: PersistedTierDeleteJournalEntry, schema: &str| {
+            let normalized = persisted
+                .clone()
+                .into_jentry()
+                .expect("the generic compatibility decoder should demonstrate the discarded field");
+            let object_name = tier_delete_journal_object_name(&normalized);
+            let encoded = serde_json::to_vec(&persisted).expect("mixed-version journal fixture should encode");
+            let err = validate_legacy_tier_delete_recovery_source(&object_name, schema, &encoded)
+                .expect_err("legacy recovery export must reject fields from later versions");
+            assert!(err.to_string().contains("later version"));
+        };
+
+        let mut invalid_v1 = Vec::new();
+        let mut with_backend = v1.clone();
+        with_backend.backend_identity = Some([7; 32]);
+        invalid_v1.push(with_backend);
+        for persisted in [&v1, &v2] {
+            let schema = if persisted.version == 1 {
+                TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA
+            } else {
+                TIER_DELETE_JOURNAL_V2_RECOVERY_SCHEMA
+            };
+            let mut invalid = Vec::new();
+            let mut with_exact = persisted.clone();
+            with_exact.version_id_exact = Some(false);
+            invalid.push(with_exact);
+            let mut with_version_state = persisted.clone();
+            with_version_state.version_state = Some(rustfs_filemeta::TransitionVersionState::Unknown);
+            invalid.push(with_version_state);
+            let mut with_state = persisted.clone();
+            with_state.state = Some(TierDeleteJournalState::Committed);
+            invalid.push(with_state);
+            let mut with_source = persisted.clone();
+            with_source.source = later.source.clone();
+            invalid.push(with_source);
+            let mut with_dispatch = persisted.clone();
+            with_dispatch.dispatch = later.dispatch.clone();
+            invalid.push(with_dispatch);
+            for record in invalid {
+                assert_rejected(record, schema);
+            }
+        }
+        for record in invalid_v1 {
+            assert_rejected(record, TIER_DELETE_JOURNAL_V1_RECOVERY_SCHEMA);
+        }
+    }
+
     #[test]
     fn tier_delete_journal_path_is_stable_and_sanitized() {
         let je = journal_entry();

crates/ecstore/src/bucket/lifecycle/tier_sweeper.rs

@@ -575,7 +575,7 @@ pub(crate) async fn delete_confirmed_transition_candidate_exact_with_lease_idemp
 #[cfg(test)]
 static CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES: std::sync::atomic::AtomicUsize = std::sync::atomic::AtomicUsize::new(0);
 
-#[cfg(test)]
+#[cfg(all(test, feature = "test-util"))]
 pub(crate) async fn delete_confirmed_transition_candidate_exact_with_manager_and_identity(
     obj_name: &str,
     rv_id: &str,
@@ -706,15 +706,16 @@ pub(crate) fn transitioned_delete_journal_entry_for_source(
 
 #[cfg(test)]
 mod test {
+    #[cfg(feature = "test-util")]
+    use super::delete_confirmed_transition_candidate_exact_with_manager_and_identity;
     use rustfs_s3_client::signer_error::invalid_utf8_header_error;
 
     use super::{
         CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES, ERR_REMOTE_DELETE_BREAKER_OPEN, ERR_REMOTE_DELETE_LIMITER_CLOSED, Jentry,
         RemoteDeleteBreaker, RemoteTierDeleteOutcome, TierDeleteJournalState, TierDeleteSourceIdentity,
-        delete_confirmed_transition_candidate_exact_with_manager_and_identity, delete_object_from_remote_tier_idempotent,
-        delete_object_from_remote_tier_idempotent_with_manager_and_identity, is_remote_tier_not_found_error,
-        is_signer_header_error, lifecycle, set_remote_tier_delete_test_hook, should_record_remote_delete_failure,
-        transitioned_delete_journal_entry, transitioned_force_delete_journal_entry,
+        delete_object_from_remote_tier_idempotent, delete_object_from_remote_tier_idempotent_with_manager_and_identity,
+        is_remote_tier_not_found_error, is_signer_header_error, lifecycle, set_remote_tier_delete_test_hook,
+        should_record_remote_delete_failure, transitioned_delete_journal_entry, transitioned_force_delete_journal_entry,
     };
     use crate::storage_api_contracts::lifecycle::TransitionedObject;
     use rustfs_filemeta::TransitionVersionState;

crates/ecstore/src/bucket/metadata.rs

@@ -791,9 +791,22 @@ impl BucketMetadata {
         }
     }
 
+    /// Replace one config payload and stamp its `*_config_updated_at` with the
+    /// local clock. This is the entry for edits that originate here: the
+    /// local write time is the edit's source time.
     pub fn update_config(&mut self, config_file: &str, data: Vec<u8>) -> Result<OffsetDateTime> {
-        let updated = OffsetDateTime::now_utc();
+        self.update_config_at(config_file, data, OffsetDateTime::now_utc())
+    }
 
+    /// [`Self::update_config`] with an explicit `updated_at` stamp.
+    ///
+    /// For a config replicated from another site the edit's source time is
+    /// the peer's `updated_at`, not the moment it lands here: staleness of
+    /// the next incoming item is judged against the stored stamp, so stamping
+    /// the local apply time would reject a newer source edit that was merely
+    /// delivered late (backlog#2292). Only replication receivers should pass
+    /// a foreign time; local edits keep [`Self::update_config`].
+    pub fn update_config_at(&mut self, config_file: &str, data: Vec<u8>, updated: OffsetDateTime) -> Result<OffsetDateTime> {
         match config_file {
             BUCKET_POLICY_CONFIG => {
                 self.policy_config_json = data;
@@ -1525,6 +1538,39 @@ mod test {
         assert_eq!(metadata.bucket_incarnation_id, incarnation);
     }
 
+    /// backlog#2292: a replicated config is stamped with the source
+    /// `updated_at` it was given, not the local clock, while the plain
+    /// `update_config` entry keeps stamping the local clock.
+    #[test]
+    fn update_config_at_stamps_the_given_time_and_update_config_stamps_now() {
+        let source_time = OffsetDateTime::now_utc() - time::Duration::hours(3);
+        let mut metadata = BucketMetadata::new("source-stamped");
+
+        let stamped = metadata
+            .update_config_at(BUCKET_POLICY_CONFIG, br#"{"Version":"2012-10-17","Statement":[]}"#.to_vec(), source_time)
+            .unwrap();
+        assert_eq!(stamped, source_time);
+        assert_eq!(metadata.policy_config_updated_at, source_time);
+
+        let tagging = b"<Tagging><TagSet><Tag><Key>k</Key><Value>v</Value></Tag></TagSet></Tagging>".to_vec();
+        let stamped = metadata
+            .update_config_at(BUCKET_TAGGING_CONFIG, tagging, source_time)
+            .unwrap();
+        assert_eq!(stamped, source_time);
+        assert_eq!(metadata.tagging_config_updated_at, source_time);
+
+        let before = OffsetDateTime::now_utc();
+        let stamped = metadata
+            .update_config(BUCKET_POLICY_CONFIG, br#"{"Version":"2012-10-17","Statement":[]}"#.to_vec())
+            .unwrap();
+        assert!(stamped >= before, "a local edit is stamped with the local clock");
+        assert_eq!(metadata.policy_config_updated_at, stamped);
+        assert_eq!(
+            metadata.tagging_config_updated_at, source_time,
+            "restamping one config must not move another config's stamp"
+        );
+    }
+
     #[test]
     fn object_locking_requires_lock_metadata_not_plain_versioning() {
         use s3s::dto::ObjectLockEnabled;
@@ -1611,6 +1657,34 @@ mod test {
         assert!(bm.bucket_target_config.is_none());
     }
 
+    /// rustfs/backlog#2309: the MinIO-origin `.metadata.bin` this repository
+    /// already carries as a compatibility fixture stores
+    /// `BucketTargetsConfigJSON` as a bare JSON array, which `BucketTargets`
+    /// (a `{"targets":[…]}` struct with no array fallback) cannot decode. The
+    /// bytes below are the exact payload the fixture in
+    /// `metadata_test.rs::TEST_BUCKET_METADATA_HEX` decodes to, so if RustFS
+    /// ever grows the array-shaped compatibility parse, this test is where the
+    /// upgrade break is pinned and where the decision has to be recorded.
+    #[test]
+    fn minio_array_shaped_bucket_targets_are_unreadable() {
+        let minio_array = br#"[{"endpoint":"http://target.example.com","targetBucket":"tb","region":"us-east-1"}]"#.to_vec();
+        let mut bm = BucketMetadata::new("minio-array-targets");
+        bm.bucket_targets_config_json = minio_array.clone();
+
+        bm.parse_all_configs()
+            .expect("a MinIO-shaped targets blob must not fail the whole metadata load");
+
+        assert!(
+            bm.bucket_targets_unreadable(),
+            "an array-shaped MinIO targets blob is unreadable, not an empty target set"
+        );
+        assert!(bm.bucket_target_config.is_none());
+        assert_eq!(
+            bm.bucket_targets_config_json, minio_array,
+            "the raw MinIO bytes must survive so the configuration stays recoverable"
+        );
+    }
+
     /// The invariant every branch of `parse_all_configs` shares: a stored but
     /// undecodable payload keeps its raw bytes and leaves the typed field
     /// `None`, so no branch fabricates a value. What a reader may then do with

crates/ecstore/src/bucket/metadata_sys.rs

@@ -567,6 +567,32 @@ pub async fn update_if_incarnation(
         config_file,
         data,
         Some(expected_incarnation_id),
+        None,
+    ))
+    .await
+}
+
+/// [`update_if_incarnation`] stamping the config with `updated_at` instead of
+/// the local clock.
+///
+/// For a site-replication receiver the edit's source time is the peer's
+/// `updated_at`; persisting it keeps the stored `*_config_updated_at` on the
+/// source clock so the next item's staleness is judged source-time against
+/// source-time (backlog#2292). See [`BucketMetadata::update_config_at`].
+pub async fn update_if_incarnation_at(
+    bucket: &str,
+    config_file: &str,
+    data: Vec<u8>,
+    expected_incarnation_id: Uuid,
+    updated_at: OffsetDateTime,
+) -> Result<OffsetDateTime> {
+    Box::pin(update_with_sys_expected(
+        get_bucket_metadata_sys()?,
+        bucket,
+        config_file,
+        data,
+        Some(expected_incarnation_id),
+        Some(updated_at),
     ))
     .await
 }
@@ -577,6 +603,30 @@ pub async fn delete_if_incarnation(bucket: &str, config_file: &str, expected_inc
         bucket,
         config_file,
         Some(expected_incarnation_id),
+        None,
+    ))
+    .await
+}
+
+/// [`delete_if_incarnation`] stamping the cleared config with `updated_at`
+/// (a replicated deletion's source time) instead of the local clock.
+///
+/// The stamp survives the deletion as the config's `*_config_updated_at`, and
+/// that is what the next incoming item is judged against: a local stamp on
+/// the delete would reject a newer source re-create that was merely delivered
+/// later (backlog#2292). See [`update_if_incarnation_at`].
+pub async fn delete_if_incarnation_at(
+    bucket: &str,
+    config_file: &str,
+    expected_incarnation_id: Uuid,
+    updated_at: OffsetDateTime,
+) -> Result<OffsetDateTime> {
+    Box::pin(delete_with_sys_expected(
+        get_bucket_metadata_sys()?,
+        bucket,
+        config_file,
+        Some(expected_incarnation_id),
+        Some(updated_at),
     ))
     .await
 }
@@ -598,34 +648,41 @@ async fn update_with_sys(
     config_file: &str,
     data: Vec<u8>,
 ) -> Result<OffsetDateTime> {
-    update_with_sys_expected(sys, bucket, config_file, data, None).await
+    update_with_sys_expected(sys, bucket, config_file, data, None, None).await
 }
 
+/// `updated_at` is the stamp persisted on the config; `None` uses the local
+/// clock (the edit originates here), `Some` carries a replicated edit's
+/// source time (backlog#2292).
 async fn update_with_sys_expected(
     sys: Arc<RwLock<BucketMetadataSys>>,
     bucket: &str,
     config_file: &str,
     data: Vec<u8>,
     expected_incarnation_id: Option<Uuid>,
+    updated_at: Option<OffsetDateTime>,
 ) -> Result<OffsetDateTime> {
     let guard = acquire_config_write_guard_for_incarnation(sys.clone(), bucket, expected_incarnation_id).await?;
-    update_under_config_write_guard(sys, &guard, config_file, data).await
+    update_under_config_write_guard(sys, &guard, config_file, data, updated_at).await
 }
 
 /// [`delete`] against an explicitly supplied metadata system. See
 /// [`update_with_sys`].
 async fn delete_with_sys(sys: Arc<RwLock<BucketMetadataSys>>, bucket: &str, config_file: &str) -> Result<OffsetDateTime> {
-    delete_with_sys_expected(sys, bucket, config_file, None).await
+    delete_with_sys_expected(sys, bucket, config_file, None, None).await
 }
 
+/// `updated_at`: `None` stamps the local clock; `Some` persists a replicated
+/// deletion's source time (backlog#2292).
 async fn delete_with_sys_expected(
     sys: Arc<RwLock<BucketMetadataSys>>,
     bucket: &str,
     config_file: &str,
     expected_incarnation_id: Option<Uuid>,
+    updated_at: Option<OffsetDateTime>,
 ) -> Result<OffsetDateTime> {
     let guard = acquire_config_write_guard_for_incarnation(sys.clone(), bucket, expected_incarnation_id).await?;
-    delete_under_config_write_guard(sys, &guard, config_file).await
+    delete_under_config_write_guard(sys, &guard, config_file, updated_at).await
 }
 
 /// Owns the complete bucket-config mutation fence.
@@ -772,7 +829,21 @@ pub async fn update_under_transaction_lock(
     data: Vec<u8>,
 ) -> Result<OffsetDateTime> {
     guard.ensure_valid(bucket)?;
-    update_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file, data).await
+    update_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file, data, None).await
+}
+
+/// [`update_under_transaction_lock`] stamping the config with `updated_at`
+/// (a replicated edit's source time) instead of the local clock; see
+/// [`update_if_incarnation_at`] (backlog#2292).
+pub async fn update_under_transaction_lock_at(
+    guard: &BucketMetadataMutationGuard,
+    bucket: &str,
+    config_file: &str,
+    data: Vec<u8>,
+    updated_at: OffsetDateTime,
+) -> Result<OffsetDateTime> {
+    guard.ensure_valid(bucket)?;
+    update_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file, data, Some(updated_at)).await
 }
 
 /// Clear one config file while the caller holds this bucket's transaction lock.
@@ -782,7 +853,7 @@ pub async fn delete_under_transaction_lock(
     config_file: &str,
 ) -> Result<OffsetDateTime> {
     guard.ensure_valid(bucket)?;
-    delete_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file).await
+    delete_under_config_write_guard(get_bucket_metadata_sys()?, guard, config_file, None).await
 }
 
 pub async fn update_quota_if_incarnation(
@@ -790,6 +861,29 @@ pub async fn update_quota_if_incarnation(
     data: Vec<u8>,
     expected_incarnation_id: Uuid,
     proof: &crate::services::notification_sys::CrossPoolFenceFleetProofToken,
+) -> Result<OffsetDateTime> {
+    update_quota_if_incarnation_stamped(bucket, data, expected_incarnation_id, proof, None).await
+}
+
+/// [`update_quota_if_incarnation`] stamping the quota config with
+/// `updated_at` (a replicated edit's source time) instead of the local
+/// clock; see [`update_if_incarnation_at`] (backlog#2292).
+pub async fn update_quota_if_incarnation_at(
+    bucket: &str,
+    data: Vec<u8>,
+    expected_incarnation_id: Uuid,
+    proof: &crate::services::notification_sys::CrossPoolFenceFleetProofToken,
+    updated_at: OffsetDateTime,
+) -> Result<OffsetDateTime> {
+    update_quota_if_incarnation_stamped(bucket, data, expected_incarnation_id, proof, Some(updated_at)).await
+}
+
+async fn update_quota_if_incarnation_stamped(
+    bucket: &str,
+    data: Vec<u8>,
+    expected_incarnation_id: Uuid,
+    proof: &crate::services::notification_sys::CrossPoolFenceFleetProofToken,
+    updated_at: Option<OffsetDateTime>,
 ) -> Result<OffsetDateTime> {
     let sys = get_bucket_metadata_sys()?;
     let guard = Box::pin(acquire_config_write_guard_for_incarnation(
@@ -807,7 +901,7 @@ pub async fn update_quota_if_incarnation(
             achieved: 0,
         });
     }
-    update_under_config_write_guard(sys, &guard, rustfs_config::QUOTA_CONFIG_FILE, data).await
+    update_under_config_write_guard(sys, &guard, rustfs_config::QUOTA_CONFIG_FILE, data, updated_at).await
 }
 
 pub async fn update_bucket_targets_under_transaction_lock(
@@ -823,6 +917,7 @@ async fn update_under_config_write_guard(
     guard: &BucketMetadataMutationGuard,
     config_file: &str,
     data: Vec<u8>,
+    updated_at: Option<OffsetDateTime>,
 ) -> Result<OffsetDateTime> {
     guard.ensure_valid(&guard.bucket)?;
     let metadata_sys = sys.read().await.clone();
@@ -834,7 +929,7 @@ async fn update_under_config_write_guard(
             Some(&guard.transaction_guard),
             &guard.bucket,
             "bucket config transaction",
-            metadata_sys.update_checked(&guard.bucket, config_file, data, true, guard.incarnation_id),
+            metadata_sys.update_checked(&guard.bucket, config_file, data, true, guard.incarnation_id, updated_at),
         ),
     )
     .await?;
@@ -846,6 +941,7 @@ async fn delete_under_config_write_guard(
     sys: Arc<RwLock<BucketMetadataSys>>,
     guard: &BucketMetadataMutationGuard,
     config_file: &str,
+    updated_at: Option<OffsetDateTime>,
 ) -> Result<OffsetDateTime> {
     guard.ensure_valid(&guard.bucket)?;
     let metadata_sys = sys.read().await.clone();
@@ -857,7 +953,7 @@ async fn delete_under_config_write_guard(
             Some(&guard.transaction_guard),
             &guard.bucket,
             "bucket config deletion transaction",
-            metadata_sys.update_checked(&guard.bucket, config_file, Vec::new(), false, guard.incarnation_id),
+            metadata_sys.update_checked(&guard.bucket, config_file, Vec::new(), false, guard.incarnation_id, updated_at),
         ),
     )
     .await?;
@@ -1762,15 +1858,17 @@ impl BucketMetadataSys {
     /// `update` and the config read alone). Keep these boxed.
     pub async fn update(&self, bucket: &str, config_file: &str, data: Vec<u8>) -> Result<OffsetDateTime> {
         let incarnation_id = Box::pin(self.get_bucket_incarnation_id(bucket)).await?;
-        Box::pin(self.update_checked(bucket, config_file, data, true, incarnation_id)).await
+        Box::pin(self.update_checked(bucket, config_file, data, true, incarnation_id, None)).await
     }
 
     pub async fn delete(&self, bucket: &str, config_file: &str) -> Result<OffsetDateTime> {
         let incarnation_id = self.get_bucket_incarnation_id(bucket).await?;
-        self.update_checked(bucket, config_file, Vec::new(), false, incarnation_id)
+        self.update_checked(bucket, config_file, Vec::new(), false, incarnation_id, None)
             .await
     }
 
+    /// `updated_at`: `None` stamps the local clock; `Some` persists a
+    /// replicated edit's source time (backlog#2292).
     async fn update_checked(
         &self,
         bucket: &str,
@@ -1778,6 +1876,7 @@ impl BucketMetadataSys {
         data: Vec<u8>,
         parse: bool,
         expected_incarnation_id: Uuid,
+        updated_at: Option<OffsetDateTime>,
     ) -> Result<OffsetDateTime> {
         // Load through this system's own store, the one `save` persists to
         // (backlog#1052 S7). Reading from the ambient handle instead made the
@@ -1788,7 +1887,10 @@ impl BucketMetadataSys {
             return Err(Error::BucketNotFound(bucket.to_string()));
         }
 
-        let updated = bm.update_config(config_file, data)?;
+        let updated = match updated_at {
+            Some(updated_at) => bm.update_config_at(config_file, data, updated_at)?,
+            None => bm.update_config(config_file, data)?,
+        };
 
         Box::pin(self.save(bm)).await?;
 
@@ -3755,6 +3857,106 @@ mod tests {
         );
     }
 
+    /// backlog#2292: the explicit-stamp write path persists the given source
+    /// time as the config's `*_config_updated_at` — through the incarnation
+    /// path and through an already-held transaction guard — and survives a
+    /// reload from disk, while the plain path keeps stamping the local clock.
+    #[tokio::test]
+    async fn explicit_updated_at_is_persisted_as_the_config_stamp() {
+        let (dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let bucket = "source-stamped-config";
+        for dir in &dirs {
+            std::fs::create_dir_all(dir.path().join(bucket)).expect("bucket volume should be created");
+        }
+        let sys = Arc::new(RwLock::new(BucketMetadataSys::new(ecstore)));
+        let source_time = OffsetDateTime::now_utc() - Duration::from_secs(3 * 3600);
+        let policy = br#"{"Version":"2012-10-17","Statement":[]}"#.to_vec();
+        let tagging = b"<Tagging><TagSet><Tag><Key>k</Key><Value>v</Value></Tag></TagSet></Tagging>".to_vec();
+
+        // Incarnation path (`update_if_incarnation_at` minus the ambient lookup).
+        let stamped =
+            update_with_sys_expected(sys.clone(), bucket, BUCKET_POLICY_CONFIG, policy.clone(), None, Some(source_time))
+                .await
+                .expect("source-stamped policy write should persist");
+        assert_eq!(stamped, source_time);
+
+        // Held-guard path (`update_under_transaction_lock_at` minus the ambient lookup).
+        let guard = acquire_config_write_guard(sys.clone(), bucket).await.expect("write guard");
+        let stamped = update_under_config_write_guard(sys.clone(), &guard, BUCKET_TAGGING_CONFIG, tagging, Some(source_time))
+            .await
+            .expect("source-stamped tagging write should persist");
+        drop(guard);
+        assert_eq!(stamped, source_time);
+
+        let metadata_sys = sys.read().await.clone();
+        metadata_sys.metadata_map.write().await.clear();
+        let reloaded = metadata_sys.get_config_from_disk(bucket).await.expect("reload from disk");
+        assert_eq!(reloaded.policy_config_updated_at, source_time);
+        assert_eq!(reloaded.tagging_config_updated_at, source_time);
+
+        // The plain path is unchanged: a local edit is stamped with the local clock.
+        let before = OffsetDateTime::now_utc();
+        let stamped = update_with_sys(sys.clone(), bucket, BUCKET_POLICY_CONFIG, policy)
+            .await
+            .expect("locally stamped policy write should persist");
+        assert!(stamped >= before, "the plain write path must keep stamping the local clock");
+        let reloaded = metadata_sys.get_config_from_disk(bucket).await.expect("reload from disk");
+        assert_eq!(reloaded.policy_config_updated_at, stamped);
+        assert_eq!(
+            reloaded.tagging_config_updated_at, source_time,
+            "an unrelated config keeps its source stamp"
+        );
+    }
+
+    /// backlog#2292: a replicated delete persists the source time as the
+    /// cleared config's `*_config_updated_at`, so the receive-side gate
+    /// (source time against stored stamp) lets a newer source re-create land
+    /// even when the delete was applied later than the re-create's source
+    /// time; the plain delete keeps stamping the local clock.
+    #[tokio::test]
+    async fn explicit_updated_at_is_persisted_by_a_delete() {
+        let (dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let bucket = "source-stamped-delete";
+        for dir in &dirs {
+            std::fs::create_dir_all(dir.path().join(bucket)).expect("bucket volume should be created");
+        }
+        let sys = Arc::new(RwLock::new(BucketMetadataSys::new(ecstore)));
+        let policy = br#"{"Version":"2012-10-17","Statement":[]}"#.to_vec();
+        let created_at = OffsetDateTime::now_utc() - Duration::from_secs(3 * 3600);
+        let deleted_at = created_at + Duration::from_secs(60);
+        let recreated_at = deleted_at + Duration::from_secs(60);
+
+        update_with_sys_expected(sys.clone(), bucket, BUCKET_POLICY_CONFIG, policy.clone(), None, Some(created_at))
+            .await
+            .expect("source-stamped policy write should persist");
+        let stamped = delete_with_sys_expected(sys.clone(), bucket, BUCKET_POLICY_CONFIG, None, Some(deleted_at))
+            .await
+            .expect("source-stamped policy delete should persist");
+        assert_eq!(stamped, deleted_at);
+
+        let metadata_sys = sys.read().await.clone();
+        metadata_sys.metadata_map.write().await.clear();
+        let reloaded = metadata_sys.get_config_from_disk(bucket).await.expect("reload from disk");
+        assert!(reloaded.policy_config_json.is_empty(), "the delete cleared the payload");
+        assert_eq!(reloaded.policy_config_updated_at, deleted_at, "the delete kept the source stamp");
+        assert!(
+            recreated_at >= reloaded.policy_config_updated_at,
+            "a re-create newer than the delete's source time is not stale against the stored stamp"
+        );
+
+        // The plain delete path is unchanged: stamped with the local clock.
+        update_with_sys_expected(sys.clone(), bucket, BUCKET_POLICY_CONFIG, policy, None, Some(recreated_at))
+            .await
+            .expect("re-create should persist");
+        let before = OffsetDateTime::now_utc();
+        let stamped = delete_with_sys_expected(sys.clone(), bucket, BUCKET_POLICY_CONFIG, None, None)
+            .await
+            .expect("locally stamped delete should persist");
+        assert!(stamped >= before, "the plain delete path must keep stamping the local clock");
+        let reloaded = metadata_sys.get_config_from_disk(bucket).await.expect("reload from disk");
+        assert_eq!(reloaded.policy_config_updated_at, stamped);
+    }
+
     /// The load and the persisted write share one write guard, so concurrent
     /// rewrites of the same config compose instead of clobbering each other.
     /// Moving the load outside that guard loses all but the last tag.
@@ -3971,10 +4173,16 @@ mod tests {
         let new_incarnation = store.bucket_incarnation_id_from_disk(bucket).await.unwrap();
         assert_ne!(old_incarnation, new_incarnation);
 
-        let err =
-            update_with_sys_expected(sys.clone(), bucket, BUCKET_TAGGING_CONFIG, b"<Tagging/>".to_vec(), Some(old_incarnation))
-                .await
-                .expect_err("a request authorized for the deleted incarnation must fail closed");
+        let err = update_with_sys_expected(
+            sys.clone(),
+            bucket,
+            BUCKET_TAGGING_CONFIG,
+            b"<Tagging/>".to_vec(),
+            Some(old_incarnation),
+            None,
+        )
+        .await
+        .expect_err("a request authorized for the deleted incarnation must fail closed");
         assert!(matches!(err, Error::BucketNotFound(name) if name == bucket));
 
         let persisted = sys.read().await.get_config_from_disk(bucket).await.unwrap();
@@ -4009,7 +4217,7 @@ mod tests {
             }],
         })
         .unwrap();
-        update_under_config_write_guard(sys, &guard, BUCKET_TAGGING_CONFIG, tagging)
+        update_under_config_write_guard(sys, &guard, BUCKET_TAGGING_CONFIG, tagging, None)
             .await
             .unwrap();
         assert!(!delete.is_finished());

crates/ecstore/src/bucket/replication/mod.rs

@@ -66,6 +66,7 @@ pub(crate) use replication_lifecycle_bridge::ReplicationLifecycleBridge;
 pub(crate) use replication_migration_bridge::ReplicationMigrationBridge;
 pub use replication_object_bridge::ReplicationObjectBridge;
 pub use replication_object_config::{DeleteReplicationConfigSnapshot, ReplicationConfig};
+pub(crate) use replication_object_decision_boundary::replication_etags_match;
 pub use replication_object_decision_boundary::{
     MustReplicateOptions, ReplicationDeleteScheduleInput, ReplicationDeleteStateSource, delete_replication_state_from_config,
     delete_replication_version_id, should_schedule_delete_replication, should_use_existing_delete_replication_info,
@@ -88,5 +89,6 @@ pub use replication_state::{ReplicationStats, RuntimeReplicationTargetBacklog};
 pub use replication_stats_boundary::{BucketReplicationStat, BucketReplicationStats, BucketStats, InQueueMetric, XferStats};
 pub use replication_storage_boundary::{ReplicationObjectIO, ReplicationStorage};
 pub use replication_target_boundary::SsecPassthroughCapability;
+pub use replication_target_boundary::VersionIdentityCapability;
 pub use replication_target_boundary::{ObjectLockIntegrity, object_lock_put_integrity};
 pub(crate) use replication_target_config_bridge::ReplicationTargetConfigBridge;

crates/ecstore/src/bucket/replication/replication_filemeta_boundary.rs

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#[cfg(test)]
+pub(crate) use rustfs_filemeta::ObjectPartInfo;
 pub use rustfs_replication::{MrfOpKind, MrfReplicateEntry};
 pub(crate) use rustfs_replication::{
     REPLICATE_EXISTING, REPLICATE_HEAL_DELETE, ReplicateTargetDecision, ReplicatedInfos, ReplicatedTargetInfo, ReplicationAction,

crates/ecstore/src/bucket/replication/replication_object_decision_boundary.rs

@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#[cfg(test)]
+pub(crate) use rustfs_replication::ReplicationMultipartPlanError;
 pub use rustfs_replication::{
     MustReplicateOptions, ReplicationDeleteScheduleInput, ReplicationDeleteStateSource, delete_replication_state_from_config,
     delete_replication_version_id, should_schedule_delete_replication, should_use_existing_delete_replication_info,
@@ -20,9 +22,9 @@ pub use rustfs_replication::{
 pub(crate) use rustfs_replication::{
     ReplicationDeleteSource, ReplicationMultipartPartInput, ReplicationResyncTargetObject, delete_marker_purge_mrf_entry,
     delete_marker_purge_version_id, delete_replication_creates_marker, delete_replication_missing_source_decision,
-    delete_replication_object_opts, heal_uses_delete_replication_path, is_object_lock_denied_delete,
-    is_retryable_delete_replication_head_error, is_version_delete_replication, replicate_delete_outcome, replication_etags_match,
-    replication_multipart_complete_actual_size, replication_multipart_part_plan, replication_single_put_size_error,
-    resync_existing_delete_replication_info, resync_target_for_object, should_retry_delete_marker_purge,
-    single_part_replica_etag_mismatch, target_delete_version_id,
+    delete_replication_object_opts, delete_replication_target_version_id, heal_uses_delete_replication_path,
+    is_object_lock_denied_delete, is_retryable_delete_replication_head_error, is_version_delete_replication,
+    replicate_delete_outcome, replication_etags_match, replication_multipart_complete_actual_size,
+    replication_multipart_part_plan, replication_single_put_size_error, resync_existing_delete_replication_info,
+    resync_target_for_object, should_retry_delete_marker_purge, single_part_replica_etag_mismatch,
 };

crates/ecstore/src/bucket/replication/replication_pool.rs

@@ -882,6 +882,20 @@ fn reconstructed_heal_delete_info(
 ) -> DeletedObjectReplicationInfo {
     let mut rstate = oi.replication_state();
     rstate.replicate_decision_str = dsc.to_string();
+    // The caller hands us a blank ObjectInfo (the source marker may already be
+    // gone), so the state above carries no target-assigned marker version ids.
+    // Restore them from the journal: `delete_marker_purge_version_id` must hit
+    // the id the target reported, not fall back to the source marker id, which
+    // a target that mints its own ids answers with an idempotent 204 that would
+    // acknowledge the intent while the real marker stays behind (backlog#2290).
+    // The corrupt flag rides along so a refusal stays a refusal after restart.
+    for (arn, version_id) in &entry.target_delete_marker_version_ids {
+        rstate
+            .target_delete_marker_version_ids
+            .entry(arn.clone())
+            .or_insert_with(|| version_id.clone());
+    }
+    rstate.target_delete_marker_version_ids_corrupt |= entry.target_delete_marker_version_ids_corrupt;
 
     let delete_marker_mtime = entry
         .delete_marker_mtime
@@ -6601,4 +6615,87 @@ mod tests {
             replacement_data
         );
     }
+
+    /// backlog#2290: a delete-marker purge intent that survives a restart
+    /// through the MRF journal addresses the marker version the TARGET
+    /// assigned, exactly as the live watcher does (see the
+    /// `requires_delayed_purge` spawn). The journal carries the per-ARN ids
+    /// (`targetDeleteMarkerVersionIDs`) and replay restores them into the
+    /// reconstructed replication state; without that the replay would fall
+    /// back to the source marker id, which a target that mints its own ids
+    /// answers with an idempotent 204 — the entry would be acknowledged while
+    /// the real marker stayed behind.
+    #[test]
+    fn mrf_delete_marker_purge_replay_preserves_target_assigned_marker_version() {
+        use super::super::replication_object_decision_boundary::{delete_marker_purge_mrf_entry, delete_marker_purge_version_id};
+
+        let arn = "arn:minio:replication::generic-target:photos".to_string();
+        let source_marker = uuid::Uuid::new_v4();
+        let remote_marker = "remote-assigned-marker-version".to_string();
+
+        let live_oi = ObjectInfo {
+            bucket: "photos".to_string(),
+            name: "obj".to_string(),
+            version_id: Some(source_marker),
+            delete_marker: true,
+            ..Default::default()
+        };
+        let mut live_state = live_oi.replication_state();
+        live_state.replicate_decision_str = replicate_decision_for_admitted_targets(std::slice::from_ref(&arn)).to_string();
+        live_state
+            .target_delete_marker_version_ids
+            .insert(arn.clone(), remote_marker.clone());
+        let live = DeletedObjectReplicationInfo {
+            delete_object: ReplicationDeletedObject {
+                object_name: "obj".to_string(),
+                delete_marker: true,
+                delete_marker_version_id: Some(source_marker),
+                replication_state: Some(live_state),
+                ..Default::default()
+            },
+            bucket: "photos".to_string(),
+            ..Default::default()
+        };
+        assert_eq!(
+            delete_marker_purge_version_id(live.delete_object.replication_state.as_ref(), &arn, source_marker),
+            Some(Some(remote_marker.clone())),
+            "the live purge addresses the recorded target version"
+        );
+
+        // Watch window exhausted: persist the intent, restart, replay it.
+        let entry = delete_marker_purge_mrf_entry(&live, vec![arn.clone()]);
+        let replay_oi = ObjectInfo {
+            bucket: entry.bucket.clone(),
+            name: entry.object.clone(),
+            version_id: entry.version_id,
+            delete_marker: entry.delete_marker,
+            ..Default::default()
+        };
+        let dsc = replicate_decision_for_admitted_targets(&entry.target_arns);
+        let replayed = reconstructed_heal_delete_info(&entry, &replay_oi, &dsc);
+
+        assert_eq!(
+            delete_marker_purge_version_id(replayed.delete_object.replication_state.as_ref(), &arn, source_marker),
+            Some(Some(remote_marker)),
+            "the MRF replay must address the target-assigned marker version, not source marker {source_marker}"
+        );
+
+        // A refusal (inconsistent recorded ids) must stay a refusal across the
+        // journal round trip instead of degrading into the source-id fallback.
+        let mut refused = live;
+        refused
+            .delete_object
+            .replication_state
+            .as_mut()
+            .expect("state was set above")
+            .target_delete_marker_version_ids_corrupt = true;
+        let entry = delete_marker_purge_mrf_entry(&refused, vec![arn.clone()]);
+        assert!(entry.target_delete_marker_version_ids_corrupt);
+        let replayed = reconstructed_heal_delete_info(&entry, &replay_oi, &dsc);
+        assert_eq!(
+            delete_marker_purge_version_id(replayed.delete_object.replication_state.as_ref(), &arn, source_marker),
+            None,
+            "the MRF replay must keep refusing to guess when the recorded ids were inconsistent"
+        );
+    }
 }

crates/ecstore/src/bucket/replication/replication_target_boundary.rs

@@ -48,6 +48,7 @@ pub use rustfs_replication::{ObjectLockIntegrity, object_lock_put_integrity};
 pub(crate) use rustfs_replication::{
     SsecPassthroughGate, is_replication_target_offline_error, ssec_passthrough_gate, version_identity_drifted,
 };
+pub use rustfs_replication::{VersionIdentityCapability, version_identity_capability_from_put};
 
 use super::replication_config_store::ReplicationConfigStore;
 use super::replication_error_boundary::{Error, Result};
@@ -192,6 +193,14 @@ impl ReplicationTargetStore {
             .await
     }
 
+    pub(crate) fn version_identity_capability(arn: &str) -> VersionIdentityCapability {
+        BucketTargetSys::get().version_identity_capability(arn)
+    }
+
+    pub(crate) fn record_version_identity_capability(arn: &str, capability: VersionIdentityCapability) {
+        BucketTargetSys::get().record_version_identity_capability(arn, capability)
+    }
+
     #[cfg(test)]
     pub(crate) async fn register_test_target(target_client: &Arc<TargetClient>) {
         BucketTargetSys::get().arn_remotes_map.write().await.insert(
@@ -248,7 +257,16 @@ pub(crate) fn replication_put_object_options(sc: &str, object_info: &ObjectInfo)
         meta.insert(AMZ_SERVER_SIDE_ENCRYPTION.to_string(), "aws:kms".to_string());
     }
 
-    let mut is_multipart = object_info.is_multipart();
+    // Older transformed objects can have physical parts without logical part
+    // lengths. Keep their existing whole-object transport: physical sizes are
+    // not plaintext boundaries for a multipart replication read.
+    let legacy_single_put = object_info.etag.as_deref().is_none_or(|etag| etag.len() == 32);
+    let base_is_multipart = object_info.is_multipart()
+        && !(legacy_single_put
+            && object_info.parts.len() > 1
+            && (object_info.is_compressed() || object_info.is_encrypted())
+            && object_info.parts.iter().any(|part| part.actual_size <= 0));
+    let mut is_multipart = base_is_multipart;
 
     if let Some(checksum_data) = &object_info.checksum
         && !checksum_data.is_empty()
@@ -259,8 +277,8 @@ pub(crate) fn replication_put_object_options(sc: &str, object_info: &ObjectInfo)
         } else if object_info.is_encrypted() {
             // Encrypted checksums cannot be exposed as plaintext headers, and
             // decrypt_checksums reports is_multipart=false for them (a value
-            // the response path relies on). Keep the object's own multipart
-            // flag so encrypted objects stay on the multipart route.
+            // the response path relies on). Keep the transport selected from
+            // the object's layout and readable part boundaries.
         } else {
             let (checksum_meta, checksum_record_is_multipart) = object_info.decrypt_checksums(0, &HeaderMap::new())?;
             // The checksum record describes how the *checksum* is composed,
@@ -268,23 +286,37 @@ pub(crate) fn replication_put_object_options(sc: &str, object_info: &ObjectInfo)
             // MULTIPART flag even on a multipart upload, so trusting it here
             // routed a 768-part object through a single PutObject and the
             // target rejected the 6 GiB body with EntityTooLarge
-            // (rustfs#6825). The object's own shape is the authority: the
+            // (rustfs#6825). The usable part layout is the authority: the
             // record may only add multipart-ness, never take it away.
-            is_multipart = object_info.is_multipart() || checksum_record_is_multipart;
+            is_multipart = base_is_multipart || checksum_record_is_multipart;
 
-            for (key, value) in checksum_meta.iter() {
-                if key != AMZ_CHECKSUM_TYPE {
-                    meta.insert(key.clone(), value.clone());
-                }
-            }
-
-            if !object_info.is_multipart()
+            if !base_is_multipart
                 && checksum_meta
                     .get(AMZ_CHECKSUM_TYPE)
                     .is_some_and(|value| value == AMZ_CHECKSUM_TYPE_FULL_OBJECT)
             {
                 is_multipart = false;
             }
+
+            // The record keys each checksum by algorithm name ("CRC32"); the
+            // target only reads `x-amz-checksum-<algorithm>`. Inserting the bare
+            // name here made `PutObjectOptions::header()` send it as user
+            // metadata (`x-amz-meta-crc32`), so no replica ever carried the
+            // source checksum (rustfs/backlog#2340). The object-level record
+            // describes one PUT body: a multipart replica is rebuilt part by
+            // part, and its CreateMultipartUpload must not announce a checksum
+            // the parts do not carry, so the record is forwarded on the
+            // single-PUT route only (MinIO `getCRCMeta` parity).
+            if !is_multipart {
+                for (key, value) in checksum_meta.iter() {
+                    if key == AMZ_CHECKSUM_TYPE {
+                        continue;
+                    }
+                    if let Some(header) = rustfs_rio::ChecksumType::from_string(key).key() {
+                        meta.insert(header.to_string(), value.clone());
+                    }
+                }
+            }
         }
     }
 
@@ -516,6 +548,7 @@ fn is_standard_header(key: &str) -> bool {
 
 #[cfg(test)]
 mod tests {
+    use super::super::replication_filemeta_boundary::ObjectPartInfo;
     use super::*;
     use aws_smithy_types::DateTime;
     use rustfs_replication::content_matches_by_etag;
@@ -550,6 +583,109 @@ mod tests {
         checksum.to_bytes(&combined)
     }
 
+    fn replication_route_metadata() -> [(&'static str, Arc<HashMap<String, String>>); 4] {
+        let mut compressed = HashMap::new();
+        rustfs_utils::http::insert_str(&mut compressed, rustfs_utils::http::SUFFIX_COMPRESSION, "zstd".to_string());
+        [
+            ("plain", Arc::new(HashMap::new())),
+            ("compressed", Arc::new(compressed)),
+            (
+                "encrypted",
+                Arc::new(HashMap::from([(AMZ_SERVER_SIDE_ENCRYPTION.to_string(), "AES256".to_string())])),
+            ),
+            (
+                "ssec",
+                Arc::new(HashMap::from([(SSEC_ALGORITHM_HEADER.to_string(), "AES256".to_string())])),
+            ),
+        ]
+    }
+
+    fn replication_route_object(
+        etag: Option<&str>,
+        actual_sizes: [i64; 3],
+        metadata: Arc<HashMap<String, String>>,
+    ) -> ObjectInfo {
+        ObjectInfo {
+            etag: etag.map(str::to_string),
+            size: 48,
+            actual_size: 12,
+            user_defined: metadata,
+            parts: Arc::new(
+                actual_sizes
+                    .into_iter()
+                    .enumerate()
+                    .map(|(index, actual_size)| ObjectPartInfo {
+                        number: index + 1,
+                        size: 16,
+                        actual_size,
+                        ..Default::default()
+                    })
+                    .collect(),
+            ),
+            ..Default::default()
+        }
+    }
+
+    #[test]
+    fn legacy_transformed_single_put_parts_keep_the_previous_replication_route() {
+        let [_, (_, compressed), (_, encrypted), (_, ssec)] = replication_route_metadata();
+        let cases = [
+            (
+                "compressed middle zero",
+                compressed.clone(),
+                Some("0123456789abcdef0123456789abcdef"),
+                [4, 0, 4],
+            ),
+            ("compressed tail unknown", compressed, None, [4, 4, -1]),
+            (
+                "encrypted middle unknown",
+                encrypted,
+                Some("gggggggggggggggggggggggggggggggg"),
+                [4, -1, 4],
+            ),
+            ("ssec tail zero", ssec.clone(), None, [4, 4, 0]),
+            ("ssec middle unknown", ssec, Some("gggggggggggggggggggggggggggggggg"), [4, -1, 4]),
+        ];
+        for (name, metadata, etag, actual_sizes) in cases {
+            for checksum in [None, Some(full_object_multipart_checksum_record())] {
+                let mut object_info = replication_route_object(etag, actual_sizes, metadata.clone());
+                object_info.checksum = checksum;
+                assert!(object_info.is_multipart(), "{name}: physical parts remain visible to metadata APIs");
+                assert!(object_info.is_compressed() || object_info.is_encrypted());
+
+                let (options, is_multipart) =
+                    replication_put_object_options("STANDARD", &object_info).expect("legacy transformed put options");
+                assert!(
+                    !is_multipart,
+                    "{name}: unknown logical part sizes must preserve the old whole-object route"
+                );
+                assert_eq!(options.internal.source_etag, etag.unwrap_or_default());
+                if metadata.contains_key(SSEC_ALGORITHM_HEADER) {
+                    assert_eq!(
+                        get_header_map(&options.user_metadata, SUFFIX_REPLICATION_SSEC_CRC).is_some(),
+                        object_info.checksum.is_some(),
+                        "SSE-C checksums retain their raw passthrough transport"
+                    );
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn positive_part_sizes_and_legacy_multipart_etags_keep_the_replication_route() {
+        for (name, metadata) in replication_route_metadata() {
+            for (etag, actual_sizes) in [
+                ("0123456789abcdef0123456789abcdef", [4, 4, 4]),
+                ("0123456789abcdef0123456789abcdef-3", [4, 0, -1]),
+            ] {
+                let mut object_info = replication_route_object(Some(etag), actual_sizes, metadata.clone());
+                object_info.checksum = Some(full_object_multipart_checksum_record());
+                let (_, is_multipart) = replication_put_object_options("STANDARD", &object_info).expect("multipart put options");
+                assert!(is_multipart, "{name}/{etag}: usable sizes and old multipart ETags must retain MPU");
+            }
+        }
+    }
+
     #[test]
     fn multipart_object_with_full_object_checksum_keeps_the_multipart_route() {
         // rustfs#6825: a 768-part upload was replicated with a single
@@ -582,6 +718,36 @@ mod tests {
         );
     }
 
+    #[test]
+    fn stored_multipart_parts_keep_the_replication_route_without_a_multipart_etag() {
+        for etag in [Some("0123456789abcdef0123456789abcdef"), None] {
+            for checksum in [None, Some(full_object_multipart_checksum_record())] {
+                let object_info = ObjectInfo {
+                    etag: etag.map(str::to_string),
+                    checksum,
+                    parts: Arc::new(
+                        (1..=2)
+                            .map(|number| ObjectPartInfo {
+                                number,
+                                ..Default::default()
+                            })
+                            .collect(),
+                    ),
+                    ..Default::default()
+                };
+                let (options, is_multipart) =
+                    replication_put_object_options("STANDARD", &object_info).expect("build put options");
+
+                assert!(
+                    is_multipart,
+                    "stored parts must retain multipart routing: etag={etag:?}, checksum={:?}",
+                    object_info.checksum
+                );
+                assert_eq!(options.internal.source_etag, etag.unwrap_or_default());
+            }
+        }
+    }
+
     #[test]
     fn checksum_record_never_changes_the_transport_a_single_part_object_needs() {
         // The mirror of the rustfs#6825 guard: an object stored as one PUT
@@ -592,6 +758,10 @@ mod tests {
         let object_info = ObjectInfo {
             etag: Some("0123456789abcdef0123456789abcdef".to_string()),
             checksum: Some(checksum.to_bytes(&[])),
+            parts: Arc::new(vec![ObjectPartInfo {
+                number: 1,
+                ..Default::default()
+            }]),
             ..Default::default()
         };
 
@@ -628,6 +798,19 @@ mod tests {
         let (_, is_multipart) = replication_put_object_options("STANDARD", &object_info).expect("build put options");
 
         assert!(is_multipart, "a composite-checksum multipart object must stay on the multipart transport");
+
+        for (name, metadata) in replication_route_metadata() {
+            let mut legacy = replication_route_object(Some("0123456789abcdef0123456789abcdef"), [4, 0, 4], metadata);
+            legacy.checksum = Some(checksum.to_bytes(&combined));
+            let (_, record_is_multipart) = legacy.decrypt_checksums(0, &HeaderMap::new()).expect("decode checksum");
+            let (_, is_multipart) = replication_put_object_options("STANDARD", &legacy).expect("legacy checksum put options");
+            if legacy.is_encrypted() {
+                assert!(!is_multipart, "{name}: encrypted checksum records must not change the old transport");
+            } else {
+                assert!(record_is_multipart, "the composite checksum must carry its own multipart signal");
+                assert!(is_multipart, "{name}: a composite record can still promote the legacy route to MPU");
+            }
+        }
     }
 
     #[test]
@@ -1308,12 +1491,63 @@ mod tests {
                 ..Default::default()
             };
 
-            let (opts, _is_multipart) = replication_put_object_options("", &object_info).expect("build replication put options");
+            let (opts, is_multipart) = replication_put_object_options("", &object_info).expect("build replication put options");
 
+            assert!(!is_multipart, "{name}: a single-part checksum record must keep the single-PUT route");
+            let header = ty.key().expect("every forwarded algorithm has an x-amz-checksum header");
             assert_eq!(
-                opts.user_metadata.get(name),
+                opts.user_metadata.get(header),
                 Some(&checksum.encoded),
-                "replication must forward the {name} checksum into user_metadata identically to the classic algorithms"
+                "replication must forward the {name} checksum as the {header} header"
+            );
+            assert!(
+                !opts.user_metadata.contains_key(name),
+                "{name}: the bare algorithm name would leave as x-amz-meta user metadata"
+            );
+        }
+    }
+
+    /// The object-level record of a multipart upload (composite or full-object)
+    /// must not become a PutObject checksum header: the replica is rebuilt
+    /// through CreateMultipartUpload/UploadPart, and a checksum announced there
+    /// that the parts do not carry would be rejected by the target.
+    #[test]
+    fn replication_put_object_options_keeps_multipart_checksum_records_off_the_wire() {
+        let mut composite_type = rustfs_rio::ChecksumType::from_string("crc32");
+        composite_type
+            .merge(rustfs_rio::ChecksumType::MULTIPART)
+            .merge(rustfs_rio::ChecksumType::INCLUDES_MULTIPART);
+        let mut combined = Vec::new();
+        for part in [b"part-one".as_slice(), b"part-two".as_slice()] {
+            let part_checksum =
+                rustfs_rio::Checksum::new_from_data(rustfs_rio::ChecksumType::from_string("crc32"), part).expect("part checksum");
+            combined.extend_from_slice(part_checksum.raw.as_slice());
+        }
+        let composite = rustfs_rio::Checksum::new_from_data(composite_type, &combined)
+            .expect("composite checksum")
+            .to_bytes(&combined);
+
+        for (label, checksum, etag) in [
+            ("composite", composite, "0123456789abcdef0123456789abcdef-2"),
+            (
+                "full-object",
+                full_object_multipart_checksum_record(),
+                "0123456789abcdef0123456789abcdef-3",
+            ),
+        ] {
+            let object_info = ObjectInfo {
+                etag: Some(etag.to_string()),
+                checksum: Some(checksum),
+                ..Default::default()
+            };
+            let (opts, is_multipart) = replication_put_object_options("", &object_info).expect("build replication put options");
+            assert!(is_multipart, "{label}: a multipart object must keep the multipart route");
+            assert!(
+                opts.user_metadata
+                    .keys()
+                    .all(|key| !key.starts_with("x-amz-checksum-") && key != "CRC32"),
+                "{label}: no object-level checksum may reach the target's CreateMultipartUpload: {:?}",
+                opts.user_metadata
             );
         }
     }

crates/ecstore/src/cluster/rpc/mod.rs

@@ -48,7 +48,8 @@ pub use internode_data_transport::build_internode_data_transport_from_env;
 pub(crate) use peer_rest_client::TierConfigReloadOutcome;
 pub use peer_rest_client::{
     KMS_SIGNAL_SUBSYSTEM, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS, PeerRestClient, SERVICE_SIGNAL_REFRESH_CONFIG,
-    SERVICE_SIGNAL_RELOAD_DYNAMIC, ScannerPeerActivity, ScannerPeerDirtyUsageSnapshot, ScannerPublicationLease,
+    SERVICE_SIGNAL_RELOAD_DYNAMIC, ScannerDirtyUsageAcknowledgement, ScannerPeerActivity, ScannerPeerDirtyUsageBucket,
+    ScannerPeerDirtyUsageSnapshot, ScannerPublicationLease, ScannerScopedDirtyUsageAckEntry,
 };
 pub(crate) use peer_s3_client::heal_bucket_local_on_disks;
 pub use peer_s3_client::{

crates/ecstore/src/cluster/rpc/peer_rest_client.rs

@@ -49,10 +49,11 @@ use rustfs_protos::proto_gen::node_service::{
     LoadRebalanceMetaRequest, LoadServiceAccountRequest, LoadTransitionTierConfigRequest, LoadUserRequest,
     LocalStorageInfoRequest, Mss, ReloadPoolMetaRequest, ReloadSiteReplicationConfigRequest, ReplacementRecoveryStatusRequest,
     ScannerActivityRequest, ScannerActivityResponse, ScannerDirtyUsageSnapshotRequest, ScannerDirtyUsageSnapshotResponse,
-    ScannerPublicationLeaseReleaseRequest, ScannerPublicationLeaseRequest, ScannerPublicationLeaseResponse, ServerInfoRequest,
-    SignalServiceRequest, SignalServiceResponse, StartDecommissionRequest, StartProfilingRequest, StopRebalanceRequest,
-    TierDailyStatsRequest, TierMutationAbortRequest, TierMutationCommitRequest, TierMutationControlResponse,
-    TierMutationFailureClass, TierMutationPeerState, TierMutationPrepareRequest, node_service_client::NodeServiceClient,
+    ScannerPublicationLeaseReleaseRequest, ScannerPublicationLeaseRequest, ScannerPublicationLeaseResponse,
+    ScannerScopedDirtyUsageAckRequest, ScannerScopedDirtyUsageEntry, ServerInfoRequest, SignalServiceRequest,
+    SignalServiceResponse, StartDecommissionRequest, StartProfilingRequest, StopRebalanceRequest, TierDailyStatsRequest,
+    TierMutationAbortRequest, TierMutationCommitRequest, TierMutationControlResponse, TierMutationFailureClass,
+    TierMutationPeerState, TierMutationPrepareRequest, node_service_client::NodeServiceClient,
     tier_mutation_control_service_client::TierMutationControlServiceClient,
 };
 pub use rustfs_protos::{PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS};
@@ -68,7 +69,10 @@ use std::{
     },
     time::SystemTime,
 };
-use tokio::{net::TcpStream, time::Duration};
+use tokio::{
+    net::TcpStream,
+    time::{Duration, timeout},
+};
 use tonic::Request;
 use tonic::service::interceptor::InterceptedService;
 use tracing::{debug, info, warn};
@@ -89,6 +93,7 @@ const HEAL_CONTROL_PAYLOAD_MAX_SIZE: usize = 64 * 1024;
 const PEER_REST_RECOVERY_MAX_ATTEMPTS: u32 = 60;
 const PEER_REST_RECOVERY_MAX_BACKOFF: Duration = Duration::from_secs(30);
 const SCANNER_ACTIVITY_MAX_MESSAGE_SIZE: usize = 1024;
+const SCANNER_SCOPED_DIRTY_USAGE_STAGE_TIMEOUT: Duration = Duration::from_secs(5);
 /// Reserve time for the acquire response's network/clock uncertainty.  The
 /// server owns the real expiry; this local deadline is intentionally earlier
 /// so a coordinator never starts a bounded persistence operation at the edge
@@ -189,12 +194,102 @@ pub struct ScannerPeerActivity {
 
 #[derive(Clone, Debug, PartialEq, Eq)]
 pub struct ScannerPeerDirtyUsageSnapshot {
+    pub owner_id: String,
     pub instance_id: String,
     pub generation: u64,
     pub pending_bucket_count: u64,
     pub protocol_version: u32,
     pub complete: bool,
-    pub buckets: BTreeMap<String, u64>,
+    pub buckets: BTreeMap<String, ScannerPeerDirtyUsageBucket>,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub struct ScannerPeerDirtyUsageBucket {
+    pub bucket_incarnation: Uuid,
+    pub generation: u64,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct ScannerScopedDirtyUsageAckEntry {
+    pub bucket: String,
+    pub bucket_incarnation: Uuid,
+    pub generation: u64,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub enum ScannerDirtyUsageAcknowledgement {
+    Generation {
+        host: String,
+        instance_id: String,
+        generation: u64,
+    },
+    Scoped {
+        host: String,
+        owner_id: String,
+        instance_id: String,
+        entries: Vec<ScannerScopedDirtyUsageAckEntry>,
+    },
+}
+
+fn scanner_scoped_dirty_usage_ack_payloads(
+    owner_id: String,
+    instance_id: String,
+    probe_only: bool,
+    entries: Vec<ScannerScopedDirtyUsageAckEntry>,
+) -> Result<Vec<ScannerScopedDirtyUsageAckRequest>> {
+    use rustfs_protos::scoped_dirty_usage::*;
+
+    if entries.is_empty() {
+        return Err(Error::other("scoped dirty usage acknowledgement entries must be nonempty"));
+    }
+
+    let mut payloads = Vec::with_capacity(entries.len().div_ceil(SCOPED_DIRTY_USAGE_MAX_ENTRIES as usize));
+    let mut batch = Vec::with_capacity(SCOPED_DIRTY_USAGE_MAX_ENTRIES as usize);
+    for entry in entries {
+        batch.push(ScannerScopedDirtyUsageEntry {
+            bucket: entry.bucket,
+            bucket_incarnation: entry.bucket_incarnation.as_bytes().to_vec().into(),
+            generation: entry.generation,
+        });
+        if batch.len() == SCOPED_DIRTY_USAGE_MAX_ENTRIES as usize {
+            payloads.push(scanner_scoped_dirty_usage_ack_payload(
+                &owner_id,
+                &instance_id,
+                probe_only,
+                std::mem::take(&mut batch),
+            )?);
+        }
+    }
+    if !batch.is_empty() {
+        payloads.push(scanner_scoped_dirty_usage_ack_payload(&owner_id, &instance_id, probe_only, batch)?);
+    }
+
+    Ok(payloads)
+}
+
+fn scanner_scoped_dirty_usage_ack_payload(
+    owner_id: &str,
+    instance_id: &str,
+    probe_only: bool,
+    entries: Vec<ScannerScopedDirtyUsageEntry>,
+) -> Result<ScannerScopedDirtyUsageAckRequest> {
+    use rustfs_protos::scoped_dirty_usage::*;
+
+    let payload = ScannerScopedDirtyUsageAckRequest {
+        challenge: Uuid::new_v4().as_bytes().to_vec().into(),
+        protocol_version: SCOPED_DIRTY_USAGE_PROTOCOL_VERSION,
+        owner_id: owner_id.to_string(),
+        instance_id: instance_id.to_string(),
+        scope: SCOPED_DIRTY_USAGE_BUCKET_SCOPE,
+        probe_only,
+        entries,
+    };
+    canonical_scoped_dirty_usage_request(&payload).map_err(|err| Error::other(err.to_string()))?;
+    Ok(payload)
+}
+
+fn scanner_scoped_dirty_usage_ack_reconciled(activity: &ScannerPeerActivity, expected_instance_id: &str) -> bool {
+    activity.instance_id == expected_instance_id && activity.dirty_usage_pending == Some(false)
 }
 
 fn scanner_instance_id_is_valid(instance_id: &str) -> bool {
@@ -348,6 +443,11 @@ fn decode_scanner_dirty_usage_snapshot_with_verifier(
     if !scanner_instance_id_is_valid(&response.instance_id) {
         return Err(Error::other("peer returned an invalid scanner dirty usage snapshot instance ID"));
     }
+    let owner_id = Uuid::parse_str(&response.owner_id)
+        .ok()
+        .filter(|owner_id| !owner_id.is_nil())
+        .map(|owner_id| owner_id.to_string())
+        .ok_or_else(|| Error::other("peer returned an invalid scanner dirty usage snapshot owner"))?;
     if response.generation == u64::MAX {
         return Err(Error::other("peer scanner dirty usage snapshot exhausted its generation"));
     }
@@ -383,9 +483,14 @@ fn decode_scanner_dirty_usage_snapshot_with_verifier(
         if bucket.generation == 0 || bucket.generation > response.generation {
             return Err(Error::other("peer scanner dirty usage snapshot contains an invalid bucket generation"));
         }
+        Uuid::from_slice(bucket.bucket_incarnation.as_ref())
+            .ok()
+            .filter(|bucket_incarnation| !bucket_incarnation.is_nil())
+            .ok_or_else(|| Error::other("peer scanner dirty usage snapshot contains an invalid bucket incarnation"))?;
     }
 
     Ok(ScannerPeerDirtyUsageSnapshot {
+        owner_id,
         instance_id: response.instance_id,
         generation: response.generation,
         pending_bucket_count: response.pending_bucket_count,
@@ -394,7 +499,16 @@ fn decode_scanner_dirty_usage_snapshot_with_verifier(
         buckets: response
             .buckets
             .into_iter()
-            .map(|bucket| (bucket.bucket, bucket.generation))
+            .map(|bucket| {
+                (
+                    bucket.bucket,
+                    ScannerPeerDirtyUsageBucket {
+                        bucket_incarnation: Uuid::from_slice(bucket.bucket_incarnation.as_ref())
+                            .expect("bucket incarnation was validated"),
+                        generation: bucket.generation,
+                    },
+                )
+            })
             .collect(),
     })
 }
@@ -874,6 +988,16 @@ impl PeerRestClient {
         self.offline.store(false, Ordering::Release);
     }
 
+    /// Prepare a retry without allowing connection-cache cleanup to extend the
+    /// caller's probe deadline. The offline gate is cleared even when eviction
+    /// times out so a cancelled cleanup cannot strand the peer in fast-fail
+    /// mode; a later request can perform a fresh eviction if needed.
+    pub async fn prepare_retry_with_timeout(&self, timeout_duration: Duration) -> bool {
+        let evicted = timeout(timeout_duration, self.evict_connection()).await.is_ok();
+        self.offline.store(false, Ordering::Release);
+        evicted
+    }
+
     /// Whether this failure means the peer is unreachable, so it should be
     /// gated offline and its connection evicted.
     ///
@@ -1675,6 +1799,24 @@ impl PeerRestClient {
         Ok((self.topology_member.clone(), supported_version, epoch))
     }
 
+    pub async fn probe_ilm_recovery_export(&self, topology_fingerprint: String) -> Result<(String, Uuid)> {
+        let probe = rustfs_protos::ilm_recovery_export_capability_probe(Uuid::new_v4().as_bytes());
+        let result = self
+            .heal_control(rustfs_protos::HEAL_CONTROL_PROTOCOL_VERSION, topology_fingerprint, probe)
+            .await?;
+        let epoch = decode_remote_version_state_capability(&self.topology_member, &result)?;
+        Ok((self.topology_member.clone(), epoch))
+    }
+
+    pub async fn probe_transition_transaction_compaction(&self, topology_fingerprint: String) -> Result<(String, Uuid)> {
+        let probe = rustfs_protos::transition_transaction_compaction_capability_probe(Uuid::new_v4().as_bytes());
+        let result = self
+            .heal_control(rustfs_protos::HEAL_CONTROL_PROTOCOL_VERSION, topology_fingerprint, probe)
+            .await?;
+        let epoch = decode_remote_version_state_capability(&self.topology_member, &result)?;
+        Ok((self.topology_member.clone(), epoch))
+    }
+
     pub async fn load_bucket_metadata(&self, bucket: &str, scanner_maintenance_change: bool) -> Result<()> {
         let result = tokio::time::timeout(BUCKET_METADATA_RELOAD_TIMEOUT, async {
             let result = self.load_bucket_metadata_once(bucket, scanner_maintenance_change).await;
@@ -2055,19 +2197,10 @@ impl PeerRestClient {
         &self,
         owner_id: String,
         instance_id: String,
-        entries: Vec<rustfs_protos::proto_gen::node_service::ScannerScopedDirtyUsageEntry>,
+        entries: Vec<ScannerScopedDirtyUsageAckEntry>,
     ) -> Result<bool> {
         use rustfs_protos::scoped_dirty_usage::*;
-        let payload = rustfs_protos::proto_gen::node_service::ScannerScopedDirtyUsageAckRequest {
-            challenge: Uuid::new_v4().as_bytes().to_vec().into(),
-            protocol_version: SCOPED_DIRTY_USAGE_PROTOCOL_VERSION,
-            owner_id,
-            instance_id,
-            scope: SCOPED_DIRTY_USAGE_BUCKET_SCOPE,
-            probe_only: true,
-            entries,
-        };
-        let canonical = canonical_scoped_dirty_usage_request(&payload).map_err(|err| Error::other(err.to_string()))?;
+        let payloads = scanner_scoped_dirty_usage_ack_payloads(owner_id, instance_id, true, entries)?;
         self.finalize_result(
             async {
                 let mut client = super::client::scanner_control_time_out_client(
@@ -2075,26 +2208,106 @@ impl PeerRestClient {
                     TonicInterceptor::Signature(gen_tonic_signature_interceptor()),
                 )
                 .await?;
+                for payload in payloads {
+                    let canonical =
+                        canonical_scoped_dirty_usage_request(&payload).map_err(|err| Error::other(err.to_string()))?;
+                    let mut request = Request::new(payload.clone());
+                    set_tonic_canonical_body_digest(&mut request, &canonical)?;
+                    let response = client.scanner_scoped_dirty_usage_ack(request).await?.into_inner();
+                    let body = canonical_scoped_dirty_usage_response(&canonical, &response)
+                        .map_err(|_| Error::other("scoped dirty usage capability response is too large"))?;
+                    verify_tonic_rpc_response_proof(&body, response.response_proof.as_ref())?;
+                    if response.protocol_version != SCOPED_DIRTY_USAGE_PROTOCOL_VERSION
+                        || response.owner_id != payload.owner_id
+                        || response.instance_id != payload.instance_id
+                        || response.max_entries != SCOPED_DIRTY_USAGE_MAX_ENTRIES
+                        || response.max_request_bytes != SCOPED_DIRTY_USAGE_MAX_REQUEST_BYTES
+                        || response.cleared != 0
+                    {
+                        return Err(Error::other("scoped dirty usage capability response does not match request"));
+                    }
+                    if !response.supported {
+                        return Ok(false);
+                    }
+                }
+                Ok(true)
+            }
+            .await,
+        )
+        .await
+    }
+
+    pub async fn acknowledge_scanner_scoped_dirty_usage(
+        &self,
+        owner_id: String,
+        instance_id: String,
+        entries: Vec<ScannerScopedDirtyUsageAckEntry>,
+    ) -> Result<ScannerPeerActivity> {
+        use rustfs_protos::scoped_dirty_usage::*;
+        let payloads = scanner_scoped_dirty_usage_ack_payloads(owner_id, instance_id.clone(), false, entries)?;
+        let ack_attempt = async {
+            let mut client = super::client::scanner_control_time_out_client(
+                &self.grid_host,
+                TonicInterceptor::Signature(gen_tonic_signature_interceptor()),
+            )
+            .await?;
+            for payload in payloads {
+                let canonical = canonical_scoped_dirty_usage_request(&payload).map_err(|err| Error::other(err.to_string()))?;
                 let mut request = Request::new(payload.clone());
                 set_tonic_canonical_body_digest(&mut request, &canonical)?;
                 let response = client.scanner_scoped_dirty_usage_ack(request).await?.into_inner();
                 let body = canonical_scoped_dirty_usage_response(&canonical, &response)
-                    .map_err(|_| Error::other("scoped dirty usage capability response is too large"))?;
+                    .map_err(|_| Error::other("scoped dirty usage acknowledgement response is too large"))?;
                 verify_tonic_rpc_response_proof(&body, response.response_proof.as_ref())?;
                 if response.protocol_version != SCOPED_DIRTY_USAGE_PROTOCOL_VERSION
                     || response.owner_id != payload.owner_id
                     || response.instance_id != payload.instance_id
                     || response.max_entries != SCOPED_DIRTY_USAGE_MAX_ENTRIES
                     || response.max_request_bytes != SCOPED_DIRTY_USAGE_MAX_REQUEST_BYTES
-                    || response.cleared != 0
+                    || !response.supported
                 {
-                    return Err(Error::other("scoped dirty usage capability response does not match request"));
+                    return Err(Error::other("scoped dirty usage acknowledgement response does not match request"));
                 }
-                Ok(response.supported)
             }
-            .await,
-        )
-        .await
+            Ok(())
+        };
+        let result = match timeout(SCANNER_SCOPED_DIRTY_USAGE_STAGE_TIMEOUT, ack_attempt).await {
+            Ok(result) => self.finalize_result(result).await,
+            Err(_) => {
+                self.prepare_retry_with_timeout(SCANNER_SCOPED_DIRTY_USAGE_STAGE_TIMEOUT)
+                    .await;
+                Err(Error::other("scoped dirty usage acknowledgement deadline elapsed"))
+            }
+        };
+
+        match result {
+            Ok(()) => {
+                let activity = self.scanner_scoped_dirty_usage_activity_confirmation().await?;
+                if activity.instance_id == instance_id {
+                    Ok(activity)
+                } else {
+                    Err(Error::other(
+                        "scoped dirty usage acknowledgement peer restarted before activity confirmation",
+                    ))
+                }
+            }
+            Err(err) => {
+                if Self::is_network_like_error(&err) {
+                    self.prepare_retry_with_timeout(SCANNER_SCOPED_DIRTY_USAGE_STAGE_TIMEOUT)
+                        .await;
+                }
+                match self.scanner_scoped_dirty_usage_activity_confirmation().await {
+                    Ok(activity) if scanner_scoped_dirty_usage_ack_reconciled(&activity, &instance_id) => Ok(activity),
+                    _ => Err(err),
+                }
+            }
+        }
+    }
+
+    async fn scanner_scoped_dirty_usage_activity_confirmation(&self) -> Result<ScannerPeerActivity> {
+        timeout(SCANNER_SCOPED_DIRTY_USAGE_STAGE_TIMEOUT, self.scanner_activity())
+            .await
+            .map_err(|_| Error::other("scoped dirty usage activity confirmation timed out"))?
     }
 
     pub async fn acknowledge_scanner_dirty_usage(&self, instance_id: String, generation: u64) -> Result<ScannerPeerActivity> {
@@ -2823,16 +3036,79 @@ mod tests {
                 rustfs_protos::proto_gen::node_service::ScannerDirtyUsageBucket {
                     bucket: "archive".to_string(),
                     generation: 3,
+                    bucket_incarnation: Uuid::from_u128(0x11111111111111111111111111111111).as_bytes().to_vec().into(),
                 },
                 rustfs_protos::proto_gen::node_service::ScannerDirtyUsageBucket {
                     bucket: "photos".to_string(),
                     generation: 7,
+                    bucket_incarnation: Uuid::from_u128(0x22222222222222222222222222222222).as_bytes().to_vec().into(),
                 },
             ],
             response_proof: b"proof".to_vec().into(),
+            owner_id: "33333333-3333-3333-3333-333333333333".to_string(),
         }
     }
 
+    #[test]
+    fn scanner_scoped_dirty_usage_ack_payloads_split_at_protocol_limit() {
+        use rustfs_protos::scoped_dirty_usage::{SCOPED_DIRTY_USAGE_MAX_ENTRIES, canonical_scoped_dirty_usage_request};
+
+        let entries = (0..=SCOPED_DIRTY_USAGE_MAX_ENTRIES)
+            .map(|index| ScannerScopedDirtyUsageAckEntry {
+                bucket: format!("bucket-{index:02}"),
+                bucket_incarnation: Uuid::from_u128(0x11111111111111111111111111111111),
+                generation: 9,
+            })
+            .collect::<Vec<_>>();
+
+        let payloads = scanner_scoped_dirty_usage_ack_payloads(
+            "33333333-3333-3333-3333-333333333333".to_string(),
+            "0123456789abcdef0123456789abcdef".to_string(),
+            false,
+            entries,
+        )
+        .expect("33 entries should split into valid scoped dirty usage requests");
+
+        assert_eq!(payloads.len(), 2);
+        assert_eq!(payloads[0].entries.len(), SCOPED_DIRTY_USAGE_MAX_ENTRIES as usize);
+        assert_eq!(payloads[1].entries.len(), 1);
+        assert_eq!(payloads[0].entries.first().map(|entry| entry.bucket.as_str()), Some("bucket-00"));
+        assert_eq!(payloads[0].entries.last().map(|entry| entry.bucket.as_str()), Some("bucket-31"));
+        assert_eq!(payloads[1].entries.first().map(|entry| entry.bucket.as_str()), Some("bucket-32"));
+        for payload in payloads {
+            canonical_scoped_dirty_usage_request(&payload).expect("each split scoped ACK payload should be canonical");
+        }
+    }
+
+    #[test]
+    fn scanner_scoped_dirty_usage_ack_reconciliation_requires_same_clean_instance() {
+        let activity = |instance_id: &str, pending| ScannerPeerActivity {
+            instance_id: instance_id.to_string(),
+            namespace_generation: 1,
+            maintenance_generation: 1,
+            protocol_version: SCANNER_ACTIVITY_PROTOCOL_VERSION,
+            topology_digest: Some([1; 32]),
+            data_movement_active: Some(false),
+            dirty_usage_generation: Some(9),
+            dirty_usage_pending: pending,
+            movement_generation: Some(1),
+            publication_blocked: Some(false),
+        };
+
+        assert!(scanner_scoped_dirty_usage_ack_reconciled(
+            &activity("0123456789abcdef0123456789abcdef", Some(false)),
+            "0123456789abcdef0123456789abcdef"
+        ));
+        assert!(!scanner_scoped_dirty_usage_ack_reconciled(
+            &activity("0123456789abcdef0123456789abcdef", Some(true)),
+            "0123456789abcdef0123456789abcdef"
+        ));
+        assert!(!scanner_scoped_dirty_usage_ack_reconciled(
+            &activity("fedcba9876543210fedcba9876543210", Some(false)),
+            "0123456789abcdef0123456789abcdef"
+        ));
+    }
+
     #[test]
     fn scanner_dirty_usage_snapshot_requires_a_complete_authenticated_ordered_view() {
         let decoded = decode_test_scanner_dirty_usage_snapshot(test_scanner_dirty_usage_snapshot_response())
@@ -2841,9 +3117,18 @@ mod tests {
         assert_eq!(decoded.generation, 7);
         assert_eq!(decoded.pending_bucket_count, 2);
         assert_eq!(decoded.protocol_version, SCANNER_DIRTY_USAGE_SNAPSHOT_PROTOCOL_VERSION);
+        assert_eq!(decoded.owner_id, "33333333-3333-3333-3333-333333333333");
         assert!(decoded.complete);
-        assert_eq!(decoded.buckets.get("archive"), Some(&3));
-        assert_eq!(decoded.buckets.get("photos"), Some(&7));
+        assert_eq!(
+            decoded.buckets.get("archive").map(|bucket| bucket.bucket_incarnation),
+            Some(Uuid::from_u128(0x11111111111111111111111111111111))
+        );
+        assert_eq!(decoded.buckets.get("archive").map(|bucket| bucket.generation), Some(3));
+        assert_eq!(
+            decoded.buckets.get("photos").map(|bucket| bucket.bucket_incarnation),
+            Some(Uuid::from_u128(0x22222222222222222222222222222222))
+        );
+        assert_eq!(decoded.buckets.get("photos").map(|bucket| bucket.generation), Some(7));
 
         let overflow_count =
             u64::try_from(SCANNER_DIRTY_USAGE_SNAPSHOT_MAX_ENTRIES + 1).expect("the test snapshot entry limit should fit in u64");
@@ -2894,6 +3179,14 @@ mod tests {
         empty_bucket.buckets[0].bucket.clear();
         cases.push((empty_bucket, "empty bucket name"));
 
+        let mut invalid_owner = test_scanner_dirty_usage_snapshot_response();
+        invalid_owner.owner_id.clear();
+        cases.push((invalid_owner, "snapshot owner"));
+
+        let mut invalid_incarnation = test_scanner_dirty_usage_snapshot_response();
+        invalid_incarnation.buckets[0].bucket_incarnation = Uuid::nil().as_bytes().to_vec().into();
+        cases.push((invalid_incarnation, "bucket incarnation"));
+
         let mut partial = test_scanner_dirty_usage_snapshot_response();
         partial.complete = false;
         cases.push((partial, "entry-limit overflow"));
@@ -2906,6 +3199,7 @@ mod tests {
                 .map(|index| rustfs_protos::proto_gen::node_service::ScannerDirtyUsageBucket {
                     bucket: format!("bucket-{index:04}"),
                     generation: 1,
+                    bucket_incarnation: Uuid::from_u128(0x11111111111111111111111111111111).as_bytes().to_vec().into(),
                 })
                 .collect(),
             ..test_scanner_dirty_usage_snapshot_response()

crates/ecstore/src/config/com.rs

@@ -449,14 +449,6 @@ where
     Ok(data)
 }
 
-pub(crate) async fn read_config_limited<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<Vec<u8>>
-where
-    S: EcstoreObjectIO,
-{
-    let (data, _obj) = read_config_with_metadata_inner(api, file, &ObjectOptions::default(), false, Some(max_bytes)).await?;
-    Ok(data)
-}
-
 pub(crate) async fn read_config_limited_preserve_empty<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<Vec<u8>>
 where
     S: EcstoreObjectIO,
@@ -465,6 +457,14 @@ where
     Ok(data)
 }
 
+pub(crate) async fn read_config_limited<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<Vec<u8>>
+where
+    S: EcstoreObjectIO,
+{
+    let (data, _obj) = read_config_with_metadata_inner(api, file, &ObjectOptions::default(), false, Some(max_bytes)).await?;
+    Ok(data)
+}
+
 pub(crate) async fn read_config_limited_preserve_empty_with_metadata<S>(
     api: Arc<S>,
     file: &str,
@@ -476,6 +476,18 @@ where
     read_config_with_metadata_inner(api, file, &ObjectOptions::default(), true, Some(max_bytes)).await
 }
 
+pub(crate) async fn read_config_limited_preserve_empty_with_metadata_opts<S>(
+    api: Arc<S>,
+    file: &str,
+    opts: &ObjectOptions,
+    max_bytes: usize,
+) -> Result<(Vec<u8>, ObjectInfo)>
+where
+    S: EcstoreObjectIO,
+{
+    read_config_with_metadata_inner(api, file, opts, true, Some(max_bytes)).await
+}
+
 /// Read an existing config object without treating an empty payload as absent.
 /// Callers that validate their own payload format need to distinguish corruption
 /// from `ConfigNotFound`.

crates/ecstore/src/core/pools.rs

@@ -3013,7 +3013,7 @@ fn parse_decommission_durable_ilm_receipt_path(path: &str) -> Result<Decommissio
         .ok_or_else(|| Error::other(format!("durable ILM receipt path `{path}` is missing its record id")))?;
     let id_kind = parts
         .next()
-        .filter(|id_kind| matches!(*id_kind, "operation_id" | "transaction_id" | "job_id"))
+        .filter(|id_kind| matches!(*id_kind, "operation_id" | "transaction_id" | "job_id" | "control_id"))
         .ok_or_else(|| Error::other(format!("durable ILM receipt path `{path}` has an invalid id kind")))?;
     let source_path = parts
         .next()
@@ -3023,8 +3023,9 @@ fn parse_decommission_durable_ilm_receipt_path(path: &str) -> Result<Decommissio
         return Err(Error::other(format!("durable ILM receipt path `{path}` has an invalid run token")));
     }
     match id_kind {
-        "operation_id" if !is_sha256_checksum(id) => {
-            return Err(Error::other(format!("durable ILM receipt path `{path}` has an invalid operation id")));
+        "operation_id" | "control_id" if !is_sha256_checksum(id) => {
+            let id_label = id_kind.trim_end_matches("_id");
+            return Err(Error::other(format!("durable ILM receipt path `{path}` has an invalid {id_label} id")));
         }
         "transaction_id" | "job_id" if uuid::Uuid::parse_str(id).is_err() => {
             return Err(Error::other(format!("durable ILM receipt path `{path}` has an invalid UUID")));
@@ -3384,7 +3385,7 @@ pub(crate) async fn acquire_pool_activation_fleet_proof(
         .ok_or_else(|| Error::other(POOL_ACTIVATION_FLEET_PROOF_REQUIRED))
 }
 
-pub(crate) fn is_pool_activation_fleet_proof_error(err: &Error) -> bool {
+pub fn is_pool_activation_fleet_proof_error(err: &Error) -> bool {
     // Save-stage helpers add context by formatting the original error, so the
     // marker may be nested in the display string. Restrict matching to the
     // `Error::other` I/O shape used by this activation path.
@@ -19300,8 +19301,8 @@ mod pools_tests {
         load_decommission_entry_versions, local_decommission_queue_prefix, mark_decommission_bucket_done,
         merge_decommission_durable_ilm_receipts, merge_pool_meta_updates_for_save, merge_pool_status_refresh,
         missing_decommission_worker_prefix, next_decommission_capacity_generation, observe_decommission_terminal_reload_result,
-        pool_meta_has_active_decommission, publish_pool_meta_updates, read_pool_meta_replica,
-        reconcile_decommission_meta_buckets, reconcile_decommission_unresolved_entries_for_completion,
+        parse_decommission_durable_ilm_receipt_path, pool_meta_has_active_decommission, publish_pool_meta_updates,
+        read_pool_meta_replica, reconcile_decommission_meta_buckets, reconcile_decommission_unresolved_entries_for_completion,
         record_decommission_unresolved_entry, recover_decommission_capacity_reservations,
         renew_decommission_capacity_reservation, require_decommission_store, reserve_decommission_start_cancelers,
         reserve_decommission_start_target_capacity, resolve_decommission_bucket_state,
@@ -20169,6 +20170,26 @@ mod pools_tests {
         assert!(!old_receipt.starts_with(&decommission_durable_ilm_receipt_run_prefix(&second_token)));
     }
 
+    #[test]
+    fn decommission_recovery_control_receipt_path_round_trips() {
+        let run_token = "b".repeat(64);
+        let control_id = "a".repeat(64);
+        let source_path = format!(
+            "ilm/recovery-controls/transition_transaction/{}/{}/{}.json",
+            &control_id[..2],
+            &control_id[2..4],
+            control_id
+        );
+        let path = decommission_durable_ilm_receipt_path(&run_token, &source_path, "control_id", &control_id);
+
+        let locator = parse_decommission_durable_ilm_receipt_path(&path).expect("recovery control receipt path should parse");
+
+        assert_eq!(locator.run_token, run_token);
+        assert_eq!(locator.source_path, source_path);
+        assert_eq!(locator.id_kind, "control_id");
+        assert_eq!(locator.id, control_id);
+    }
+
     #[test]
     fn decommission_receipt_merge_preserves_terminal_proof() {
         let operation_id = "a".repeat(64);

crates/ecstore/src/data_movement/mod.rs

@@ -2673,7 +2673,7 @@ mod tests {
             ]),
             ..Default::default()
         };
-        assert!(!object_info.is_multipart());
+        assert!(object_info.is_multipart());
         assert!(should_use_multipart_data_movement(&object_info, false));
 
         let single_nonstandard_part = ObjectInfo {
@@ -3050,7 +3050,7 @@ mod tests {
             ..Default::default()
         };
 
-        assert!(!object_info.is_multipart());
+        assert!(object_info.is_multipart());
         assert!(object_info.parts.iter().any(|part| part.checksums.is_some()));
         let opts = data_movement_put_object_opts(&object_info, 0);
         assert!(!rustfs_utils::http::contains_key_str(&opts.user_defined, SUFFIX_PART_CHECKSUMS));

crates/ecstore/src/diagnostics/get.rs

@@ -99,11 +99,17 @@ pub(crate) const GET_STAGE_READER_OPEN_MMAP_COPY_FALLBACK: &str = "reader_open_m
 pub(crate) const GET_STAGE_READER_OPEN_MMAP_COPY_SUCCESS: &str = "reader_open_mmap_copy_success";
 pub(crate) const GET_STAGE_READER_OPEN_STREAM: &str = "reader_open_stream";
 pub(crate) const GET_STAGE_READER_MMAP_ACCESS_CHECK: &str = "reader_mmap_access_check";
+#[cfg(unix)]
 pub(crate) const GET_STAGE_READER_MMAP_BLOCKING_TASK: &str = "reader_mmap_blocking_task";
+#[cfg(unix)]
 pub(crate) const GET_STAGE_READER_MMAP_BLOCKING_WAIT: &str = "reader_mmap_blocking_wait";
+#[cfg(unix)]
 pub(crate) const GET_STAGE_READER_MMAP_COPY_BUFFER: &str = "reader_mmap_copy_buffer";
+#[cfg(unix)]
 pub(crate) const GET_STAGE_READER_MMAP_DIRECT_READ_COPY: &str = "reader_mmap_direct_read_copy";
+#[cfg(unix)]
 pub(crate) const GET_STAGE_READER_MMAP_FILE_OPEN: &str = "reader_mmap_file_open";
+#[cfg(unix)]
 pub(crate) const GET_STAGE_READER_MMAP_MAP: &str = "reader_mmap_map";
 pub(crate) const GET_STAGE_READER_MMAP_METADATA_LOOKUP: &str = "reader_mmap_metadata_lookup";
 pub(crate) const GET_STAGE_READER_MMAP_METADATA_VALIDATE: &str = "reader_mmap_metadata_validate";

crates/ecstore/src/disk/disk_store.rs

@@ -324,6 +324,70 @@ impl DiskStoreRenameDataExt for LocalDiskWrapper {
 }
 
 impl LocalDiskWrapper {
+    pub(in crate::disk) async fn delete_version_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        fi: FileInfo,
+        force_del_marker: bool,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        self.track_disk_health_mutation(
+            "delete_version",
+            DiskMetricMutation::Delete,
+            || async {
+                Box::pin(
+                    self.disk
+                        .delete_version_with_namespace_owner(volume, path, fi, force_del_marker, opts, namespace_owner),
+                )
+                .await
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    pub(in crate::disk) async fn delete_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        self.track_disk_health_mutation(
+            "delete",
+            DiskMetricMutation::Delete,
+            || async { Box::pin(self.disk.delete_with_namespace_owner(volume, path, opts, namespace_owner)).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    pub(in crate::disk) async fn undo_write_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        fi: FileInfo,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        self.track_disk_health_mutation(
+            "delete_version",
+            DiskMetricMutation::Delete,
+            || async {
+                // Preserve the old DiskAPI future's boxing boundary.
+                Box::pin(
+                    self.disk
+                        .undo_write_with_namespace_owner(volume, path, fi, opts, namespace_owner),
+                )
+                .await
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
     pub(in crate::disk) async fn rename_data_observed(
         &self,
         src_volume: &str,
@@ -333,6 +397,34 @@ impl LocalDiskWrapper {
         dst_path: &str,
         external_guard: Option<Arc<dyn Send + Sync>>,
     ) -> super::RenameDataObservation {
+        self.rename_data_observed_with_guards(
+            src_volume,
+            src_path,
+            fi,
+            dst_volume,
+            dst_path,
+            super::RenameDataGuards {
+                external_guard,
+                ..Default::default()
+            },
+        )
+        .await
+    }
+
+    pub(in crate::disk) async fn rename_data_observed_with_guards(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        fi: &FileInfo,
+        dst_volume: &str,
+        dst_path: &str,
+        guards: super::RenameDataGuards,
+    ) -> super::RenameDataObservation {
+        let super::RenameDataGuards {
+            external_guard,
+            namespace_owner,
+            ..
+        } = guards;
         let operation = self.clone();
         let src_volume = src_volume.to_owned();
         let src_path = src_path.to_owned();
@@ -357,13 +449,15 @@ impl LocalDiskWrapper {
                     DiskMetricMutation::Write,
                     || async {
                         // Preserve the former DiskAPI future's single boxing boundary.
-                        let observed =
-                            Box::pin(
-                                operation
-                                    .disk
-                                    .rename_data_observed(&src_volume, &src_path, &fi, &dst_volume, &dst_path),
-                            )
-                            .await;
+                        let observed = Box::pin(operation.disk.rename_data_observed(
+                            &src_volume,
+                            &src_path,
+                            &fi,
+                            &dst_volume,
+                            &dst_path,
+                            namespace_owner,
+                        ))
+                        .await;
                         preflight_rejection = observed.preflight_rejection;
                         observed.result
                     },
@@ -1301,6 +1395,7 @@ impl LocalDiskWrapper {
         self.disk.get_object_path(volume, path)
     }
 
+    #[cfg(unix)]
     pub(crate) fn get_object_path_for_io(&self, volume: &str, path: &str) -> crate::disk::error::Result<std::path::PathBuf> {
         self.disk.get_object_path_for_io(volume, path)
     }

crates/ecstore/src/disk/fs.rs

@@ -218,10 +218,12 @@ pub async fn rename(from: impl AsRef<Path>, to: impl AsRef<Path>) -> io::Result<
     fs::rename(from, to).await
 }
 
+#[cfg(any(not(windows), test))]
 pub fn rename_std(from: impl AsRef<Path>, to: impl AsRef<Path>) -> io::Result<()> {
     std::fs::rename(from, to)
 }
 
+#[cfg(any(not(windows), test))]
 #[tracing::instrument(level = "debug", skip_all)]
 pub async fn read_file(path: impl AsRef<Path>) -> io::Result<Vec<u8>> {
     fs::read(path.as_ref()).await

crates/ecstore/src/disk/local/commit.rs

@@ -17,13 +17,14 @@
 
 #[cfg(all(test, windows))]
 use super::run_destination_commit_directory_preparation;
+#[cfg(any(not(windows), test))]
+use super::should_fail_local_inline_rollback_hardlink;
 use super::{
     EVENT_DISK_LOCAL_ACCESS_FAILED, EVENT_DISK_LOCAL_HEAL_PURGE_FAILED, EVENT_DISK_LOCAL_RENAME_REJECTED, LOG_COMPONENT_ECSTORE,
     LOG_SUBSYSTEM_DISK_LOCAL, LocalDisk, SyncMode, effective_durability, inline_metadata_rollback_dir, observe_old_current_size,
     remove_dir_all_if_exists, remove_dst_base_before_commit, remove_file_if_exists, rename_data_versions_signature,
     run_inline_preparation_before_backup, should_fail_after_metadata_commit, should_fail_before_old_metadata_backup,
-    should_fail_commit_rename, should_fail_local_inline_rollback_hardlink, should_remove_staged_meta_before_commit,
-    skip_access_checks,
+    should_fail_commit_rename, should_remove_staged_meta_before_commit, skip_access_checks,
 };
 #[cfg(test)]
 use super::{run_inline_before_file_sync_admission, run_owned_file_write_before_open, run_rename_data_after_first_publication};
@@ -33,7 +34,7 @@ use crate::disk::{
     error::{DiskError, Result},
     error_conv::{to_access_error, to_file_error},
     os,
-    os::{check_path_length, rename_all},
+    os::check_path_length,
 };
 use bytes::Bytes;
 use rustfs_filemeta::{FileInfo, FileMeta};
@@ -73,6 +74,8 @@ fn rollback_inline_metadata_commit_std(
     rollback_data_dir: Option<Uuid>,
     local_rollback_path: Option<&Path>,
 ) -> std::io::Result<()> {
+    #[cfg(all(test, not(windows)))]
+    os::prepared_publication_test_hooks::run(os::prepared_publication_test_hooks::Stage::Rollback, dst_file_path);
     if let Some(backup_path) = local_rollback_path {
         // The commit immediately before this rollback renamed the staged
         // xl.meta from the same directory as `backup_path` onto
@@ -86,6 +89,7 @@ fn rollback_inline_metadata_commit_std(
     Ok(())
 }
 
+#[cfg(any(not(windows), test))]
 pub(super) fn create_local_inline_rollback_backup(
     dst_file_path: &Path,
     staging_file_path: &Path,
@@ -231,6 +235,12 @@ async fn restore_published_data_source(
 #[derive(Debug)]
 pub(in crate::disk) struct LocalRenamePreflightRejection(());
 
+#[derive(Default)]
+pub(super) struct RenameDataState {
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+    preflight_rejection: Option<LocalRenamePreflightRejection>,
+}
+
 impl LocalDisk {
     #[tracing::instrument(name = "rename_data", target = "rustfs_ecstore::disk::local", level = "trace", skip_all)]
     pub(super) async fn rename_data_inner(
@@ -240,7 +250,7 @@ impl LocalDisk {
         fi: FileInfo,
         dst_volume: &str,
         dst_path: &str,
-        preflight_rejection: &mut Option<LocalRenamePreflightRejection>,
+        state: &mut RenameDataState,
     ) -> Result<RenameDataResp> {
         crate::hp_guard!("LocalDisk::rename_data");
         let mut fi = fi;
@@ -269,7 +279,13 @@ impl LocalDisk {
             Some(token) => Some(self.claim_quota_mutation_fence(dst_volume, dst_path, token).await?),
             None => None,
         };
-        let mutation_lease = os::acquire_rename_data_mutation_lease(&self.root, dst_volume, &destination_object_path).await;
+        let mutation_lease = os::acquire_rename_data_mutation_lease_with_owner(
+            &self.root,
+            dst_volume,
+            &destination_object_path,
+            state.namespace_owner.take(),
+        )
+        .await;
         if let Some(claim) = quota_fence_claim {
             mutation_lease.attach_external_guard(claim);
         }
@@ -302,7 +318,7 @@ impl LocalDisk {
                 error = %e,
                 "Disk local access check failed"
             );
-            *preflight_rejection = Some(LocalRenamePreflightRejection(()));
+            state.preflight_rejection = Some(LocalRenamePreflightRejection(()));
             return Err(to_access_error(e, DiskError::VolumeAccessDenied).into());
         }
 
@@ -320,7 +336,7 @@ impl LocalDisk {
                 error = %e,
                 "Disk local access check failed"
             );
-            *preflight_rejection = Some(LocalRenamePreflightRejection(()));
+            state.preflight_rejection = Some(LocalRenamePreflightRejection(()));
             return Err(to_access_error(e, DiskError::VolumeAccessDenied).into());
         }
 
@@ -528,7 +544,9 @@ impl LocalDisk {
             // rename below.
             if fi_healing
                 && let Some((_, dst_data_path)) = has_data_dir_path.as_ref()
-                && let Err(err) = self.move_to_trash(dst_data_path, true, false).await
+                && let Err(err) = self
+                    .move_to_trash_with_namespace_owner(dst_data_path, true, false, Some(mutation_lease.clone()))
+                    .await
             {
                 warn!(
                     target: "rustfs_ecstore::disk::local",
@@ -755,7 +773,7 @@ impl LocalDisk {
                 && let Some(parent) = dst_file_path.parent()
             {
                 let fsync_started = rustfs_io_metrics::put_stage_timer();
-                if let Err(err) = os::fsync_dst_dir_group_commit(parent).await {
+                if let Err(err) = os::fsync_dst_dir_group_commit(parent, Some(mutation_lease.clone())).await {
                     rustfs_io_metrics::record_put_object_stage_duration_from(
                         rustfs_io_metrics::PUT_STAGE_SET_DISK_RENAME_DST_DIR_FSYNC,
                         fsync_started,
@@ -793,7 +811,7 @@ impl LocalDisk {
                         break;
                     }
                     let fsync_started = rustfs_io_metrics::put_stage_timer();
-                    if let Err(err) = os::fsync_dir(dir).await {
+                    if let Err(err) = os::fsync_dir_with_owner(dir, Some(mutation_lease.clone())).await {
                         rustfs_io_metrics::record_put_object_stage_duration_from(
                             rustfs_io_metrics::PUT_STAGE_SET_DISK_RENAME_ANCESTOR_DIR_FSYNC,
                             fsync_started,
@@ -1024,7 +1042,15 @@ impl LocalDisk {
                 // rename_all acquires the backup path's namespace lease. Do not
                 // hold a disk admission while acquiring another namespace lock.
                 drop(file_sync_admission.take());
-                if let Err(err) = rename_all(staged_backup, &backup_path, &dst_volume_dir, &self.publication_root).await {
+                if let Err(err) = os::rename_all_with_owner(
+                    staged_backup,
+                    &backup_path,
+                    &dst_volume_dir,
+                    &self.publication_root,
+                    Some(mutation_lease.clone()),
+                )
+                .await
+                {
                     let _ = remove_file_if_exists(staged_backup);
                     return Err(err);
                 }
@@ -1220,14 +1246,18 @@ impl LocalDisk {
         fi: &FileInfo,
         dst_volume: &str,
         dst_path: &str,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
     ) -> super::super::RenameDataObservation {
-        let mut preflight_rejection = None;
+        let mut state = RenameDataState {
+            namespace_owner,
+            ..Default::default()
+        };
         let result = self
-            .rename_data_inner(src_volume, src_path, fi.clone(), dst_volume, dst_path, &mut preflight_rejection)
+            .rename_data_inner(src_volume, src_path, fi.clone(), dst_volume, dst_path, &mut state)
             .await;
         super::super::RenameDataObservation {
             result,
-            preflight_rejection,
+            preflight_rejection: state.preflight_rejection,
         }
     }
 }

crates/ecstore/src/disk/mod.rs

@@ -75,6 +75,14 @@ use time::OffsetDateTime;
 use tokio::io::{AsyncRead, AsyncWrite};
 use uuid::Uuid;
 
+/// Independent admission and physical ownership for one disk rename.
+#[derive(Default)]
+pub(crate) struct RenameDataGuards {
+    pub(crate) scanner_publication_lease_token: Option<Uuid>,
+    pub(crate) external_guard: Option<Arc<dyn Send + Sync>>,
+    pub(crate) namespace_owner: Option<Arc<dyn Send + Sync>>,
+}
+
 /// Local preflight evidence stays outside DiskAPI and the RPC response format.
 pub(crate) struct RenameDataObservation {
     pub(crate) result: Result<RenameDataResp>,
@@ -190,11 +198,17 @@ pub struct MmapCopyStageMetrics {
     pub(crate) path_resolve_stage: &'static str,
     pub(crate) metadata_lookup_stage: &'static str,
     pub(crate) metadata_validate_stage: &'static str,
+    #[cfg(unix)]
     pub(crate) blocking_wait_stage: &'static str,
+    #[cfg(unix)]
     pub(crate) blocking_task_stage: &'static str,
+    #[cfg(unix)]
     pub(crate) file_open_stage: &'static str,
+    #[cfg(unix)]
     pub(crate) mmap_map_stage: &'static str,
+    #[cfg(unix)]
     pub(crate) mmap_copy_stage: &'static str,
+    #[cfg(unix)]
     pub(crate) direct_read_copy_stage: &'static str,
 }
 
@@ -718,6 +732,65 @@ impl Disk {
         }
     }
 
+    pub(crate) async fn delete_version_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        fi: FileInfo,
+        force_del_marker: bool,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        match self {
+            Self::Local(disk) => {
+                disk.delete_version_with_namespace_owner(volume, path, fi, force_del_marker, opts, namespace_owner)
+                    .await
+            }
+            Self::Remote(disk) => {
+                let result = disk.delete_version(volume, path, fi, force_del_marker, opts).await;
+                // This is sender lifetime only, not proof of a remote physical drain.
+                drop(namespace_owner);
+                result
+            }
+        }
+    }
+
+    pub(crate) async fn delete_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        match self {
+            Self::Local(disk) => disk.delete_with_namespace_owner(volume, path, opts, namespace_owner).await,
+            Self::Remote(disk) => {
+                let result = disk.delete(volume, path, opts).await;
+                drop(namespace_owner);
+                result
+            }
+        }
+    }
+
+    /// Keep local undo publication owned independently of the wrapper deadline.
+    /// Remote undo retains its existing RPC contract; this is not a remote drain proof.
+    pub(crate) async fn undo_write_with_namespace_owner(
+        &self,
+        volume: &str,
+        path: &str,
+        fi: FileInfo,
+        opts: DeleteOptions,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
+    ) -> Result<()> {
+        match self {
+            Self::Local(disk) => {
+                disk.undo_write_with_namespace_owner(volume, path, fi, opts, namespace_owner)
+                    .await
+            }
+            Self::Remote(disk) => disk.delete_version(volume, path, fi, false, opts).await,
+        }
+    }
+
     pub(crate) async fn rename_data_borrowed(
         &self,
         src_volume: &str,
@@ -737,12 +810,12 @@ impl Disk {
         fi: &FileInfo,
         dst_volume: &str,
         dst_path: &str,
-        scanner_publication_lease_token: Option<Uuid>,
+        guards: RenameDataGuards,
     ) -> RenameDataObservation {
         match self {
             Disk::Local(local_disk) => {
                 local_disk
-                    .rename_data_observed(src_volume, src_path, fi, dst_volume, dst_path, None)
+                    .rename_data_observed_with_guards(src_volume, src_path, fi, dst_volume, dst_path, guards)
                     .await
             }
             Disk::Remote(remote_disk) => RenameDataObservation::unknown(
@@ -753,7 +826,7 @@ impl Disk {
                         fi,
                         dst_volume,
                         dst_path,
-                        scanner_publication_lease_token,
+                        guards.scanner_publication_lease_token,
                     )
                     .await,
             ),
@@ -922,6 +995,7 @@ impl Disk {
         }
     }
 
+    #[cfg(unix)]
     pub(crate) fn get_object_path_for_io_if_local(
         &self,
         volume: &str,

crates/ecstore/src/disk/os.rs

@@ -91,6 +91,7 @@ pub(crate) mod fsync_dir_recorder {
     static RECORDED: Mutex<Vec<PathBuf>> = Mutex::new(Vec::new());
     static LIMITED: Mutex<Vec<PathBuf>> = Mutex::new(Vec::new());
     static GROUPED: Mutex<Vec<(PathBuf, usize)>> = Mutex::new(Vec::new());
+    #[cfg(unix)]
     static BEFORE_LIMITED: std::sync::LazyLock<Mutex<HashMap<PathBuf, Hook>>> =
         std::sync::LazyLock::new(|| Mutex::new(HashMap::new()));
     static BEFORE_GROUP_BATCH: std::sync::LazyLock<Mutex<HashMap<PathBuf, Hook>>> =
@@ -150,6 +151,7 @@ pub(crate) mod fsync_dir_recorder {
         contains_path(&RECORDED.lock().expect("fsync dir recorder poisoned"), dir)
     }
 
+    #[cfg(unix)]
     pub(crate) fn record_limited(dir: &Path) {
         record_path(&LIMITED, dir, "limited fsync dir recorder");
         let hook = remove_hook(&BEFORE_LIMITED, dir, "limited fsync hook poisoned");
@@ -162,6 +164,7 @@ pub(crate) mod fsync_dir_recorder {
         contains_path(&LIMITED.lock().expect("limited fsync dir recorder poisoned"), dir)
     }
 
+    #[cfg(unix)]
     pub(crate) fn set_before_limited(dir: &Path, hook: impl FnOnce() + Send + 'static) {
         BEFORE_LIMITED
             .lock()
@@ -237,11 +240,96 @@ pub(crate) mod fsync_dir_recorder {
             .insert(dir.to_path_buf(), kind);
     }
 
+    #[cfg(unix)]
     pub(crate) fn take_grouped_failure(dir: &Path) -> Option<io::ErrorKind> {
         remove_path_keyed(&GROUPED_FAILURES, dir, "grouped fsync failure hook poisoned")
     }
 }
 
+/// Pause a real namespace mutation inside its physical executor.
+#[cfg(all(test, not(windows)))]
+pub(crate) mod prepared_publication_test_hooks {
+    use super::*;
+
+    #[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
+    pub(crate) enum Stage {
+        PreparedRename,
+        Rename,
+        Remove,
+        Rollback,
+        DirFsync,
+    }
+
+    type Hook = Box<dyn FnOnce() + Send>;
+    type Key = (Stage, PathBuf);
+    static BEFORE_PUBLICATION: LazyLock<Mutex<HashMap<Key, Hook>>> = LazyLock::new(|| Mutex::new(HashMap::new()));
+
+    pub(crate) struct Guard(Key);
+
+    impl Drop for Guard {
+        fn drop(&mut self) {
+            BEFORE_PUBLICATION.lock().remove(&self.0);
+        }
+    }
+
+    pub(crate) fn install(path: &Path, hook: impl FnOnce() + Send + 'static) -> Guard {
+        install_at(Stage::PreparedRename, path, hook)
+    }
+
+    pub(crate) fn install_at(stage: Stage, path: &Path, hook: impl FnOnce() + Send + 'static) -> Guard {
+        let key = (stage, path.to_path_buf());
+        assert!(BEFORE_PUBLICATION.lock().insert(key.clone(), Box::new(hook)).is_none());
+        Guard(key)
+    }
+
+    pub(crate) fn run(stage: Stage, path: &Path) {
+        let hook = BEFORE_PUBLICATION.lock().remove(&(stage, path.to_path_buf()));
+        if let Some(hook) = hook {
+            hook();
+        }
+    }
+
+    #[cfg(test)]
+    type RenameDestinationHook = Box<dyn FnOnce(&Path) + Send>;
+    #[cfg(test)]
+    static RENAME_DESTINATIONS: LazyLock<Mutex<HashMap<PathBuf, RenameDestinationHook>>> =
+        LazyLock::new(|| Mutex::new(HashMap::new()));
+
+    #[cfg(test)]
+    pub(crate) struct RenameDestinationGuard(PathBuf);
+
+    #[cfg(test)]
+    impl Drop for RenameDestinationGuard {
+        fn drop(&mut self) {
+            RENAME_DESTINATIONS.lock().remove(&self.0);
+        }
+    }
+
+    #[cfg(test)]
+    pub(crate) fn observe_rename_destination(source: &Path, hook: impl FnOnce(&Path) + Send + 'static) -> RenameDestinationGuard {
+        assert!(
+            RENAME_DESTINATIONS
+                .lock()
+                .insert(source.to_path_buf(), Box::new(hook))
+                .is_none()
+        );
+        RenameDestinationGuard(source.to_path_buf())
+    }
+
+    #[cfg(test)]
+    pub(crate) async fn drain_namespace_key(path: &Path) {
+        drop(super::acquire_namespace_mutation_lease(path).await);
+    }
+
+    #[cfg(test)]
+    pub(super) fn run_rename_destination(source: &Path, destination: &Path) {
+        let hook = RENAME_DESTINATIONS.lock().remove(source);
+        if let Some(hook) = hook {
+            hook(destination);
+        }
+    }
+}
+
 #[cfg(all(test, windows))]
 pub(crate) mod windows_rename_test_hooks {
     use super::*;
@@ -576,6 +664,7 @@ impl OpenedDstDirFsyncGroup {
 }
 
 struct DstDirFsyncWaiter {
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
     result_tx: oneshot::Sender<SharedDstDirFsyncResult>,
 }
 
@@ -634,6 +723,7 @@ impl DstDirFsyncGroupCommit {
     fn enqueue_opened(
         &self,
         opened: OpenedDstDirFsyncGroup,
+        namespace_owner: Option<Arc<dyn Send + Sync>>,
     ) -> io::Result<(oneshot::Receiver<SharedDstDirFsyncResult>, Option<Arc<DstDirFsyncGroup>>)> {
         let (result_tx, result_rx) = oneshot::channel();
         let mut registry = self.inner.lock();
@@ -664,7 +754,10 @@ impl DstDirFsyncGroupCommit {
             group
         };
         let mut group_state = group.inner.lock();
-        group_state.pending.push_back(DstDirFsyncWaiter { result_tx });
+        group_state.pending.push_back(DstDirFsyncWaiter {
+            result_tx,
+            namespace_owner,
+        });
         let start_worker = !group_state.worker_running;
         if start_worker {
             group_state.worker_running = true;
@@ -686,7 +779,13 @@ impl DstDirFsyncGroupCommit {
     fn remove_idle_group(&self, group: &Arc<DstDirFsyncGroup>) {
         let mut registry = self.inner.lock();
         let group_state = group.inner.lock();
-        if !group_state.worker_running && group_state.pending.is_empty() {
+        if !group_state.worker_running
+            && group_state.pending.is_empty()
+            && registry
+                .groups
+                .get(&group.key)
+                .is_some_and(|registered| Arc::ptr_eq(registered, group))
+        {
             registry.groups.remove(&group.key);
         }
     }
@@ -709,16 +808,20 @@ impl DstDirFsyncGroupCommit {
         &self,
         dir: &Path,
     ) -> io::Result<(oneshot::Receiver<SharedDstDirFsyncResult>, Option<Arc<DstDirFsyncGroup>>)> {
-        self.enqueue_opened(OpenedDstDirFsyncGroup::open(dir)?)
+        self.enqueue_opened(OpenedDstDirFsyncGroup::open(dir)?, None)
     }
 }
 
 #[cfg(unix)]
-async fn fsync_open_dst_dir_group(group: &DstDirFsyncGroup) -> io::Result<()> {
+async fn fsync_open_dst_dir_group(group: &DstDirFsyncGroup, namespace_owners: Vec<Arc<dyn Send + Sync>>) -> io::Result<()> {
     #[cfg(test)]
     let dir = group.dir.clone();
     let dir_file = group.dir_file.clone();
     fsync_spawn_blocking(move || {
+        // The batch worker may be cancelled while this syscall is still running.
+        let _namespace_owners = namespace_owners;
+        #[cfg(all(test, not(windows)))]
+        prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::DirFsync, &dir);
         #[cfg(test)]
         {
             if let Some(kind) = fsync_dir_recorder::take_grouped_failure(&dir) {
@@ -733,66 +836,118 @@ async fn fsync_open_dst_dir_group(group: &DstDirFsyncGroup) -> io::Result<()> {
 }
 
 #[cfg(not(unix))]
-async fn fsync_open_dst_dir_group(group: &DstDirFsyncGroup) -> io::Result<()> {
+async fn fsync_open_dst_dir_group(group: &DstDirFsyncGroup, namespace_owners: Vec<Arc<dyn Send + Sync>>) -> io::Result<()> {
+    let _namespace_owners = namespace_owners;
     fsync_dir(&group.dir).await
 }
 
-async fn run_dst_dir_fsync_group_worker(group: Arc<DstDirFsyncGroup>) {
-    loop {
-        #[cfg(test)]
-        fsync_dir_recorder::run_before_group_batch(&group.dir);
-        tokio::task::yield_now().await;
-        let batch: Vec<DstDirFsyncWaiter> = {
-            let mut group_state = group.inner.lock();
-            group_state.pending.drain(..).collect()
-        };
-        if batch.is_empty() {
-            let mut group_state = group.inner.lock();
+struct DstDirFsyncWorkerGuard {
+    group: Arc<DstDirFsyncGroup>,
+    in_flight: usize,
+    armed: bool,
+}
+
+impl Drop for DstDirFsyncWorkerGuard {
+    fn drop(&mut self) {
+        if !self.armed {
+            return;
+        }
+        // Cancellation must release queued owners, but the physical batch keeps
+        // its own owners until its blocking syscall returns.
+        let pending = {
+            let mut registry = DST_DIR_FSYNC_GROUP_COMMIT.inner.lock();
+            let mut group_state = self.group.inner.lock();
+            let pending = std::mem::take(&mut group_state.pending);
             group_state.worker_running = false;
-            drop(group_state);
-            DST_DIR_FSYNC_GROUP_COMMIT.remove_idle_group(&group);
-            return;
-        }
-
-        #[cfg(test)]
-        fsync_dir_recorder::record_grouped(&group.dir, batch.len());
-        let result = fsync_open_dst_dir_group(&group)
-            .await
-            .map_err(SharedDstDirFsyncError::from_error);
-        let batch_len = batch.len();
-        DST_DIR_FSYNC_GROUP_COMMIT.complete_batch(batch_len);
-
-        let should_stop = {
-            let mut group_state = group.inner.lock();
-            if group_state.pending.is_empty() {
-                group_state.worker_running = false;
-                true
-            } else {
-                false
+            if registry
+                .groups
+                .get(&self.group.key)
+                .is_some_and(|group| Arc::ptr_eq(group, &self.group))
+            {
+                registry.total_waiters = registry.total_waiters.saturating_sub(pending.len() + self.in_flight);
+                registry.groups.remove(&self.group.key);
             }
+            pending
         };
-        if should_stop {
-            DST_DIR_FSYNC_GROUP_COMMIT.remove_idle_group(&group);
-        }
-        for waiter in batch {
-            let _ = waiter.result_tx.send(result.clone());
-        }
-        if should_stop {
-            return;
+        // Lease and channel destructors must run outside the registry locks.
+        drop(pending);
+    }
+}
+
+fn run_dst_dir_fsync_group_worker(group: Arc<DstDirFsyncGroup>) -> impl std::future::Future<Output = ()> {
+    // Capture before spawning: shutdown may drop the future without polling it.
+    let worker_guard = DstDirFsyncWorkerGuard {
+        group: group.clone(),
+        in_flight: 0,
+        armed: true,
+    };
+    async move {
+        let mut worker_guard = worker_guard;
+        loop {
+            #[cfg(test)]
+            fsync_dir_recorder::run_before_group_batch(&group.dir);
+            tokio::task::yield_now().await;
+            let mut batch: Vec<DstDirFsyncWaiter> = {
+                let mut group_state = group.inner.lock();
+                group_state.pending.drain(..).collect()
+            };
+            if batch.is_empty() {
+                let mut group_state = group.inner.lock();
+                worker_guard.armed = false;
+                group_state.worker_running = false;
+                drop(group_state);
+                DST_DIR_FSYNC_GROUP_COMMIT.remove_idle_group(&group);
+                return;
+            }
+            worker_guard.in_flight = batch.len();
+
+            #[cfg(test)]
+            fsync_dir_recorder::record_grouped(&group.dir, batch.len());
+            let namespace_owners = batch.iter_mut().filter_map(|waiter| waiter.namespace_owner.take()).collect();
+            let result = fsync_open_dst_dir_group(&group, namespace_owners)
+                .await
+                .map_err(SharedDstDirFsyncError::from_error);
+            let batch_len = batch.len();
+            DST_DIR_FSYNC_GROUP_COMMIT.complete_batch(batch_len);
+            worker_guard.in_flight = 0;
+
+            let should_stop = {
+                let mut group_state = group.inner.lock();
+                if group_state.pending.is_empty() {
+                    worker_guard.armed = false;
+                    group_state.worker_running = false;
+                    true
+                } else {
+                    false
+                }
+            };
+            if should_stop {
+                DST_DIR_FSYNC_GROUP_COMMIT.remove_idle_group(&group);
+            }
+            for waiter in batch {
+                let _ = waiter.result_tx.send(result.clone());
+            }
+            if should_stop {
+                return;
+            }
         }
     }
 }
 
-async fn fsync_dst_dir_group_commit_with_enabled(dir: impl AsRef<Path>, enabled: bool) -> io::Result<()> {
+async fn fsync_dst_dir_group_commit_with_enabled(
+    dir: impl AsRef<Path>,
+    enabled: bool,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> io::Result<()> {
     if !enabled {
-        return fsync_dir(dir).await;
+        return fsync_dir_with_owner(dir.as_ref(), namespace_owner).await;
     }
 
     let dir = dir.as_ref().to_path_buf();
     let opened = tokio::task::spawn_blocking(move || OpenedDstDirFsyncGroup::open(&dir))
         .await
         .map_err(|err| io::Error::other(format!("blocking dst dir group open failed: {err}")))??;
-    let (result_rx, worker) = DST_DIR_FSYNC_GROUP_COMMIT.enqueue_opened(opened)?;
+    let (result_rx, worker) = DST_DIR_FSYNC_GROUP_COMMIT.enqueue_opened(opened, namespace_owner)?;
     if let Some(group) = worker {
         tokio::spawn(run_dst_dir_fsync_group_worker(group));
     }
@@ -804,8 +959,11 @@ async fn fsync_dst_dir_group_commit_with_enabled(dir: impl AsRef<Path>, enabled:
     }
 }
 
-pub(crate) async fn fsync_dst_dir_group_commit(dir: impl AsRef<Path>) -> io::Result<()> {
-    fsync_dst_dir_group_commit_with_enabled(dir, dst_dir_fsync_group_commit_enabled()).await
+pub(crate) async fn fsync_dst_dir_group_commit(
+    dir: impl AsRef<Path>,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> io::Result<()> {
+    fsync_dst_dir_group_commit_with_enabled(dir, dst_dir_fsync_group_commit_enabled(), namespace_owner).await
 }
 
 pub(crate) async fn fsync_dst_dir_group_commit_or_namespace_file_sync_limit(
@@ -814,7 +972,7 @@ pub(crate) async fn fsync_dst_dir_group_commit_or_namespace_file_sync_limit(
     admission: &FileSyncAdmission,
 ) -> io::Result<()> {
     if dst_dir_fsync_group_commit_enabled() {
-        fsync_dst_dir_group_commit_with_enabled(dir, true).await
+        fsync_dst_dir_group_commit_with_enabled(dir, true, Some(lease)).await
     } else {
         fsync_dir_with_namespace_file_sync_limit(dir, lease, admission).await
     }
@@ -822,7 +980,7 @@ pub(crate) async fn fsync_dst_dir_group_commit_or_namespace_file_sync_limit(
 
 #[cfg(test)]
 pub(crate) async fn fsync_dst_dir_group_commit_for_test(dir: impl AsRef<Path>, enabled: bool) -> io::Result<()> {
-    fsync_dst_dir_group_commit_with_enabled(dir, enabled).await
+    fsync_dst_dir_group_commit_with_enabled(dir, enabled, None).await
 }
 
 #[cfg(test)]
@@ -1229,6 +1387,8 @@ pub(crate) struct NamespaceMutationLease {
     _namespace_guard: OwnedMutexGuard<()>,
     _volume_guard: Option<OwnedRwLockReadGuard<()>>,
     external_guard: Mutex<Option<Arc<dyn Send + Sync>>>,
+    // Independent of the quota claim; both survive cancellation of the waiter.
+    _namespace_owner: Option<Arc<dyn Send + Sync>>,
 }
 
 impl NamespaceMutationLease {
@@ -1238,10 +1398,18 @@ impl NamespaceMutationLease {
 }
 
 async fn acquire_namespace_mutation_lease(path: &Path) -> Arc<NamespaceMutationLease> {
+    acquire_namespace_mutation_lease_with_owner(path, None).await
+}
+
+pub(in crate::disk) async fn acquire_namespace_mutation_lease_with_owner(
+    path: &Path,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> Arc<NamespaceMutationLease> {
     Arc::new(NamespaceMutationLease {
         _namespace_guard: disk_namespace_mutation_lock(path).lock_owned().await,
         _volume_guard: None,
         external_guard: Mutex::new(None),
+        _namespace_owner: namespace_owner,
     })
 }
 
@@ -1251,6 +1419,15 @@ pub(crate) async fn acquire_rename_data_mutation_lease(
     root: &Path,
     volume: &str,
     destination_object: &Path,
+) -> Arc<NamespaceMutationLease> {
+    acquire_rename_data_mutation_lease_with_owner(root, volume, destination_object, None).await
+}
+
+pub(crate) async fn acquire_rename_data_mutation_lease_with_owner(
+    root: &Path,
+    volume: &str,
+    destination_object: &Path,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
 ) -> Arc<NamespaceMutationLease> {
     let namespace_guard = disk_namespace_mutation_lock(destination_object).lock_owned().await;
     let volume_guard = disk_volume_mutation_lock(root, volume).read_owned().await;
@@ -1258,6 +1435,7 @@ pub(crate) async fn acquire_rename_data_mutation_lease(
         _namespace_guard: namespace_guard,
         _volume_guard: Some(volume_guard),
         external_guard: Mutex::new(None),
+        _namespace_owner: namespace_owner,
     })
 }
 
@@ -1747,6 +1925,105 @@ pub async fn rename_all(
     Ok(())
 }
 
+pub(crate) async fn fsync_dir_with_owner(path: &Path, namespace_owner: Option<Arc<dyn Send + Sync>>) -> io::Result<()> {
+    #[cfg(unix)]
+    {
+        if namespace_owner.is_none() {
+            return fsync_dir(path).await;
+        }
+        let path = path.to_path_buf();
+        fsync_spawn_blocking(move || {
+            let _namespace_owner = namespace_owner;
+            fsync_dir_std(path)
+        })
+        .await?
+    }
+    #[cfg(not(unix))]
+    {
+        let _ = namespace_owner;
+        fsync_dir(path).await
+    }
+}
+
+/// Retain namespace ownership in the actual filesystem executor after timeout.
+pub(crate) async fn remove_file_with_owner(
+    path: impl AsRef<Path>,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> io::Result<()> {
+    if namespace_owner.is_none() {
+        return tokio::fs::remove_file(path).await;
+    }
+    let path = path.as_ref().to_path_buf();
+    let lease = acquire_namespace_mutation_lease_with_owner(&path, namespace_owner).await;
+    run_blocking_namespace_operation(lease, move || {
+        #[cfg(all(test, not(windows)))]
+        prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Remove, &path);
+        std::fs::remove_file(path)
+    })
+    .await
+}
+
+/// Retain namespace ownership in the actual filesystem executor after timeout.
+pub(crate) async fn remove_dir_with_owner(
+    path: impl AsRef<Path>,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> io::Result<()> {
+    if namespace_owner.is_none() {
+        return tokio::fs::remove_dir(path).await;
+    }
+    let path = path.as_ref().to_path_buf();
+    let lease = acquire_namespace_mutation_lease_with_owner(&path, namespace_owner).await;
+    run_blocking_namespace_operation(lease, move || std::fs::remove_dir(path)).await
+}
+
+/// Preserve raw rename semantics while retaining a counted owner in the syscall.
+/// Unlike reliable rename, this never creates parents or retries a missing source.
+pub(in crate::disk) async fn rename_with_namespace_owner(
+    src: &Path,
+    dst: &Path,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> io::Result<()> {
+    if namespace_owner.is_none() {
+        return tokio::fs::rename(src, dst).await;
+    }
+    let src = src.to_path_buf();
+    let dst = dst.to_path_buf();
+    let lease = acquire_namespace_mutation_lease_with_owner(&dst, namespace_owner).await;
+    run_blocking_namespace_operation(lease, move || {
+        #[cfg(all(test, not(windows)))]
+        {
+            prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Rename, &src);
+            prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Rename, &dst);
+        }
+        std::fs::rename(src, dst)
+    })
+    .await
+}
+
+pub(in crate::disk) async fn create_dir_all_with_namespace_owner(
+    path: &Path,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> io::Result<()> {
+    if namespace_owner.is_none() {
+        return tokio::fs::create_dir_all(path).await;
+    }
+    let path = path.to_path_buf();
+    let lease = acquire_namespace_mutation_lease_with_owner(&path, namespace_owner).await;
+    run_blocking_namespace_operation(lease, move || std::fs::create_dir_all(path)).await
+}
+
+#[tracing::instrument(name = "rename_all", level = "debug", skip_all)]
+pub(crate) async fn rename_all_with_owner(
+    src_file_path: impl AsRef<Path>,
+    dst_file_path: impl AsRef<Path>,
+    base_dir: impl AsRef<Path>,
+    publication_root: &PublicationRoot,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> Result<()> {
+    let lease = acquire_namespace_mutation_lease_with_owner(dst_file_path.as_ref(), namespace_owner).await;
+    rename_all_with_lease(src_file_path, dst_file_path, base_dir, publication_root, lease).await
+}
+
 pub(crate) async fn rename_all_with_lease(
     src_file_path: impl AsRef<Path>,
     dst_file_path: impl AsRef<Path>,
@@ -1939,6 +2216,8 @@ pub(crate) async fn rename_all_with_prepared_source(
         move || {
             validate_prepared_rename_source(&prepared_source, &src_file_path)?;
             let preparation = prepare_rename_with_retry(&src_file_path, &dst_file_path, &base_dir, &publication_root)?;
+            #[cfg(test)]
+            prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::PreparedRename, &dst_file_path);
             rename_prepared(&src_file_path, &dst_file_path, &preparation)
         }
     };
@@ -1977,6 +2256,32 @@ pub async fn rename_all_ignore_missing_source(
     }
 }
 
+#[tracing::instrument(name = "rename_all_ignore_missing_source", level = "debug", skip_all)]
+pub(crate) async fn rename_all_ignore_missing_source_with_owner(
+    src_file_path: impl AsRef<Path>,
+    dst_file_path: impl AsRef<Path>,
+    base_dir: impl AsRef<Path>,
+    publication_root: &PublicationRoot,
+    namespace_owner: Option<Arc<dyn Send + Sync>>,
+) -> Result<()> {
+    let src_file_path = src_file_path.as_ref();
+    let lease = acquire_namespace_mutation_lease_with_owner(dst_file_path.as_ref(), namespace_owner).await;
+    match reliable_rename_inner_with_lease(
+        src_file_path.to_path_buf(),
+        dst_file_path.as_ref().to_path_buf(),
+        base_dir.as_ref().to_path_buf(),
+        publication_root.clone(),
+        false,
+        lease,
+    )
+    .await
+    {
+        Ok(()) => Ok(()),
+        Err(err) if err.kind() == io::ErrorKind::NotFound && rename_source_is_missing(src_file_path, publication_root) => Ok(()),
+        Err(err) => Err(to_file_error(err).into()),
+    }
+}
+
 #[cfg(windows)]
 pub(crate) fn rename_source_is_missing(src_file_path: &Path, publication_root: &PublicationRoot) -> bool {
     let Some(source_parent) = src_file_path.parent() else {
@@ -2042,6 +2347,13 @@ async fn reliable_rename_inner_with_lease(
         let base_dir = base_dir.clone();
         move || {
             let preparation = prepare_rename_with_retry(&src_file_path, &dst_file_path, &base_dir, &publication_root)?;
+            #[cfg(all(test, not(windows)))]
+            prepared_publication_test_hooks::run_rename_destination(&src_file_path, &dst_file_path);
+            #[cfg(all(test, not(windows)))]
+            {
+                prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Rename, &src_file_path);
+                prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Rename, &dst_file_path);
+            }
             rename_prepared(&src_file_path, &dst_file_path, &preparation)
         }
     };
@@ -6136,6 +6448,245 @@ mod tests {
         wait_for_dst_dir_fsync_group_commit_idle().await;
     }
 
+    #[cfg(unix)]
+    #[tokio::test]
+    #[serial_test::serial(dst_dir_fsync_group_commit)]
+    async fn grouped_fsync_physical_batch_keeps_all_owners_after_worker_cancellation() {
+        let temp_dir = tempdir().expect("fixture directory");
+        let dir = temp_dir.path().canonicalize().expect("canonical fsync path");
+        let first_ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let second_ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let first_owner = first_ctx.begin_namespace_commit();
+        let second_owner = second_ctx.begin_namespace_commit();
+        let first_probe = Arc::downgrade(&first_owner);
+        let second_probe = Arc::downgrade(&second_owner);
+        let (first_rx, group) = DST_DIR_FSYNC_GROUP_COMMIT
+            .enqueue_opened(
+                OpenedDstDirFsyncGroup::open(&dir).expect("open first waiter directory"),
+                Some(first_owner),
+            )
+            .expect("queue first real waiter");
+        let group = group.expect("first waiter starts the group");
+        let (second_rx, second_worker) = DST_DIR_FSYNC_GROUP_COMMIT
+            .enqueue_opened(
+                OpenedDstDirFsyncGroup::open(&dir).expect("open second waiter directory"),
+                Some(second_owner),
+            )
+            .expect("queue second real waiter");
+        assert!(second_worker.is_none(), "same directory must join the same batch");
+        assert_eq!(group.inner.lock().pending.len(), 2);
+        let (entered_tx, entered_rx) = tokio::sync::oneshot::channel();
+        let (release_tx, release_rx) = std::sync::mpsc::channel::<()>();
+        let _hook =
+            prepared_publication_test_hooks::install_at(prepared_publication_test_hooks::Stage::DirFsync, &dir, move || {
+                let _ = entered_tx.send(());
+                let _ = release_rx.recv();
+            });
+        let worker = tokio::spawn(run_dst_dir_fsync_group_worker(group.clone()));
+        tokio::time::timeout(Duration::from_secs(5), entered_rx)
+            .await
+            .expect("batch must reach its physical fsync")
+            .expect("physical fsync entry");
+        assert_eq!(fsync_dir_recorder::grouped_batch_sizes(&dir), vec![2]);
+        assert!(
+            group.inner.lock().pending.is_empty(),
+            "both waiters were transferred into the physical batch"
+        );
+        let queued_ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let queued_owner = queued_ctx.begin_namespace_commit();
+        let queued_probe = Arc::downgrade(&queued_owner);
+        let queued_generation = queued_ctx.namespace_commit_generation();
+        let (queued_rx, queued_worker) = DST_DIR_FSYNC_GROUP_COMMIT
+            .enqueue_opened(
+                OpenedDstDirFsyncGroup::open(&dir).expect("open queued waiter directory"),
+                Some(queued_owner),
+            )
+            .expect("queue a waiter after the physical batch was frozen");
+        assert!(queued_worker.is_none());
+        assert_eq!(group.inner.lock().pending.len(), 1);
+        drop((first_rx, second_rx));
+        worker.abort();
+        assert!(worker.await.expect_err("cancel the async batch owner").is_cancelled());
+        assert!(queued_rx.await.is_err(), "an undispatched waiter must observe worker cancellation");
+        assert!(queued_probe.upgrade().is_none());
+        assert!(!queued_ctx.namespace_commits_pending());
+        assert!(queued_ctx.namespace_commit_generation() > queued_generation);
+        assert!(group.inner.lock().pending.is_empty());
+        assert!(!group.inner.lock().worker_running);
+        assert_eq!(DST_DIR_FSYNC_GROUP_COMMIT.counts_for_test(), (0, 0));
+        let first_pending = first_ctx.namespace_commits_pending() && first_probe.upgrade().is_some();
+        let second_pending = second_ctx.namespace_commits_pending() && second_probe.upgrade().is_some();
+        let generations = (first_ctx.namespace_commit_generation(), second_ctx.namespace_commit_generation());
+        drop(release_tx);
+        tokio::time::timeout(Duration::from_secs(5), async {
+            while Arc::strong_count(&group.dir_file) != 1
+                || first_probe.upgrade().is_some()
+                || second_probe.upgrade().is_some()
+                || first_ctx.namespace_commits_pending()
+                || second_ctx.namespace_commits_pending()
+            {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("physical fsync must release every batch owner");
+        assert!(fsync_dir_recorder::was_fsynced(&dir), "the detached syscall must really execute");
+        assert!(
+            first_pending && second_pending,
+            "one physical batch must preserve both independent namespace owners"
+        );
+        assert!(!first_ctx.namespace_commits_pending());
+        assert!(!second_ctx.namespace_commits_pending());
+        assert!(first_ctx.namespace_commit_generation() > generations.0);
+        assert!(second_ctx.namespace_commit_generation() > generations.1);
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    #[serial_test::serial(dst_dir_fsync_group_commit)]
+    async fn grouped_fsync_unpolled_worker_releases_queued_owner() {
+        let temp_dir = tempdir().expect("fixture directory");
+        let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let owner = ctx.begin_namespace_commit();
+        let probe = Arc::downgrade(&owner);
+        let generation = ctx.namespace_commit_generation();
+        let (rx, group) = DST_DIR_FSYNC_GROUP_COMMIT
+            .enqueue_opened(
+                OpenedDstDirFsyncGroup::open(temp_dir.path()).expect("open queued waiter directory"),
+                Some(owner),
+            )
+            .expect("queue a real waiter");
+        let group = group.expect("first waiter starts the group");
+        let worker = run_dst_dir_fsync_group_worker(group.clone());
+        assert!(ctx.namespace_commits_pending());
+        drop(worker);
+        assert!(rx.await.is_err(), "shutdown before first poll must release the waiter");
+        assert!(probe.upgrade().is_none());
+        assert!(!ctx.namespace_commits_pending());
+        assert!(ctx.namespace_commit_generation() > generation);
+        assert!(group.inner.lock().pending.is_empty());
+        assert!(!group.inner.lock().worker_running);
+        assert_eq!(DST_DIR_FSYNC_GROUP_COMMIT.counts_for_test(), (0, 0));
+        assert!(
+            fsync_dir_recorder::grouped_batch_sizes(temp_dir.path()).is_empty(),
+            "the dropped future must not dispatch a physical batch"
+        );
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn stale_idle_group_cleanup_preserves_successor_registration() {
+        let temp_dir = tempdir().expect("fixture directory");
+        let registry = DstDirFsyncGroupCommit::default();
+        let (mut first_rx, first_worker) = registry.enqueue_for_test(temp_dir.path()).expect("enqueue first worker");
+        let old_group = first_worker.expect("first waiter starts a worker");
+        // W1 has completed its batch and marked G idle, but has not cleaned G up.
+        let first_waiter = old_group.inner.lock().pending.pop_front().expect("first batch waiter");
+        registry.complete_batch(1);
+        old_group.inner.lock().worker_running = false;
+
+        let (mut second_rx, second_worker) = registry.enqueue_for_test(temp_dir.path()).expect("enqueue second worker");
+        let reused_group = second_worker.expect("idle G starts another worker");
+        assert!(Arc::ptr_eq(&old_group, &reused_group));
+        let second_waiter = reused_group.inner.lock().pending.pop_front().expect("second batch waiter");
+        registry.complete_batch(1);
+        reused_group.inner.lock().worker_running = false;
+        registry.remove_idle_group(&reused_group);
+        assert_eq!(registry.counts_for_test(), (0, 0), "normal idle cleanup must remove G");
+        assert!(second_waiter.result_tx.send(Ok(())).is_ok());
+        assert!(second_rx.try_recv().expect("second worker reports completion").is_ok());
+
+        let (mut successor_rx, successor_worker) = registry.enqueue_for_test(temp_dir.path()).expect("enqueue successor");
+        let successor = successor_worker.expect("successor starts a new group");
+        assert!(!Arc::ptr_eq(&old_group, &successor));
+        assert_eq!(registry.counts_for_test(), (1, 1));
+        // W1 resumes with its old Arc after W2 removed G and W3 installed G2.
+        registry.remove_idle_group(&old_group);
+        assert!(first_waiter.result_tx.send(Ok(())).is_ok());
+        assert!(first_rx.try_recv().expect("first worker reports completion").is_ok());
+        assert!(
+            registry
+                .inner
+                .lock()
+                .groups
+                .get(&successor.key)
+                .is_some_and(|registered| Arc::ptr_eq(registered, &successor)),
+            "stale cleanup must retain the exact successor Arc"
+        );
+        assert_eq!(registry.counts_for_test(), (1, 1));
+        assert!(successor.inner.lock().worker_running);
+        assert_eq!(successor.inner.lock().pending.len(), 1);
+        assert!(matches!(successor_rx.try_recv(), Err(oneshot::error::TryRecvError::Empty)));
+
+        let (_joined_rx, new_worker) = registry.enqueue_for_test(temp_dir.path()).expect("join successor");
+        assert!(new_worker.is_none(), "a later waiter must join G2 instead of creating G3");
+        assert_eq!(successor.inner.lock().pending.len(), 2);
+        assert_eq!(registry.counts_for_test(), (1, 2));
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    #[serial_test::serial(dst_dir_fsync_group_commit)]
+    async fn stale_idle_cleanup_then_unpolled_worker_drop_releases_waiter_budget() {
+        wait_for_dst_dir_fsync_group_commit_idle().await;
+        let temp_dir = tempdir().expect("fixture directory");
+        let (old_rx, old_worker) = DST_DIR_FSYNC_GROUP_COMMIT
+            .enqueue_for_test(temp_dir.path())
+            .expect("enqueue old group");
+        let old_group = old_worker.expect("old group starts a worker");
+        tokio::time::timeout(Duration::from_secs(5), run_dst_dir_fsync_group_worker(old_group.clone()))
+            .await
+            .expect("old worker must finish its actual fsync");
+        assert!(old_rx.await.expect("old worker reports completion").is_ok());
+        assert!(fsync_dir_recorder::was_fsynced(temp_dir.path()));
+        assert_eq!(fsync_dir_recorder::grouped_batch_sizes(temp_dir.path()), vec![1]);
+        assert_eq!(DST_DIR_FSYNC_GROUP_COMMIT.counts_for_test(), (0, 0));
+
+        let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let owner = ctx.begin_namespace_commit();
+        let probe = Arc::downgrade(&owner);
+        let generation = ctx.namespace_commit_generation();
+        let (rx, successor_worker) = DST_DIR_FSYNC_GROUP_COMMIT
+            .enqueue_opened(
+                OpenedDstDirFsyncGroup::open(temp_dir.path()).expect("open successor directory"),
+                Some(owner),
+            )
+            .expect("enqueue successor owner");
+        let successor = successor_worker.expect("successor starts a new group");
+        assert!(!Arc::ptr_eq(&old_group, &successor));
+        let worker = run_dst_dir_fsync_group_worker(successor.clone());
+        // The stale Arc represents W1 resuming after another worker removed G.
+        DST_DIR_FSYNC_GROUP_COMMIT.remove_idle_group(&old_group);
+        assert!(ctx.namespace_commits_pending());
+        assert!(probe.upgrade().is_some());
+        drop(worker);
+        let channel_closed = tokio::time::timeout(Duration::from_secs(5), rx)
+            .await
+            .expect("dropping the unpolled worker must release its channel")
+            .is_err();
+        let counts_after_drop = DST_DIR_FSYNC_GROUP_COMMIT.counts_for_test();
+        let owner_released = probe.upgrade().is_none();
+        let namespace_pending = ctx.namespace_commits_pending();
+        let generation_after_drop = ctx.namespace_commit_generation();
+        let successor_pending = successor.inner.lock().pending.len();
+        let worker_running = successor.inner.lock().worker_running;
+        // Preserve the observed result before cleanup, so a RED run cannot leak
+        // its phantom count into unrelated tests in the same process.
+        clear_dst_dir_fsync_group_commit_for_test();
+        assert!(channel_closed);
+        assert!(owner_released);
+        assert!(!namespace_pending);
+        assert!(generation_after_drop > generation);
+        assert_eq!(successor_pending, 0);
+        assert!(!worker_running);
+        assert_eq!(
+            fsync_dir_recorder::grouped_batch_sizes(temp_dir.path()),
+            vec![1],
+            "dropping the successor before its first poll must not dispatch another fsync"
+        );
+        assert_eq!(counts_after_drop, (0, 0), "stale cleanup must not strand a phantom waiter");
+    }
+
     #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
     #[serial_test::serial(dst_dir_fsync_group_commit)]
     async fn dst_dir_fsync_group_commit_cancellation_releases_waiter_state() {

crates/ecstore/src/io_support/bitrot.rs

@@ -13,12 +13,15 @@
 // limitations under the License.
 
 use crate::diagnostics::get::{
-    GET_STAGE_READER_MMAP_ACCESS_CHECK, GET_STAGE_READER_MMAP_BLOCKING_TASK, GET_STAGE_READER_MMAP_BLOCKING_WAIT,
-    GET_STAGE_READER_MMAP_COPY_BUFFER, GET_STAGE_READER_MMAP_DIRECT_READ_COPY, GET_STAGE_READER_MMAP_FILE_OPEN,
-    GET_STAGE_READER_MMAP_MAP, GET_STAGE_READER_MMAP_METADATA_LOOKUP, GET_STAGE_READER_MMAP_METADATA_VALIDATE,
+    GET_STAGE_READER_MMAP_ACCESS_CHECK, GET_STAGE_READER_MMAP_METADATA_LOOKUP, GET_STAGE_READER_MMAP_METADATA_VALIDATE,
     GET_STAGE_READER_MMAP_PATH_RESOLVE, GET_STAGE_READER_OPEN_MMAP_COPY_FALLBACK, GET_STAGE_READER_OPEN_MMAP_COPY_SUCCESS,
     GET_STAGE_READER_OPEN_STREAM, GET_STAGE_READER_STREAM_FIRST_READ, record_get_stage_duration_if_enabled,
 };
+#[cfg(unix)]
+use crate::diagnostics::get::{
+    GET_STAGE_READER_MMAP_BLOCKING_TASK, GET_STAGE_READER_MMAP_BLOCKING_WAIT, GET_STAGE_READER_MMAP_COPY_BUFFER,
+    GET_STAGE_READER_MMAP_DIRECT_READ_COPY, GET_STAGE_READER_MMAP_FILE_OPEN, GET_STAGE_READER_MMAP_MAP,
+};
 #[cfg(feature = "hotpath")]
 use crate::disk::FileWriter;
 use crate::disk::{self, DiskAPI as _, DiskStore, FileReader, MmapCopyStageMetrics, error::DiskError};
@@ -406,11 +409,17 @@ async fn open_disk_reader(
             path_resolve_stage: GET_STAGE_READER_MMAP_PATH_RESOLVE,
             metadata_lookup_stage: GET_STAGE_READER_MMAP_METADATA_LOOKUP,
             metadata_validate_stage: GET_STAGE_READER_MMAP_METADATA_VALIDATE,
+            #[cfg(unix)]
             blocking_wait_stage: GET_STAGE_READER_MMAP_BLOCKING_WAIT,
+            #[cfg(unix)]
             blocking_task_stage: GET_STAGE_READER_MMAP_BLOCKING_TASK,
+            #[cfg(unix)]
             file_open_stage: GET_STAGE_READER_MMAP_FILE_OPEN,
+            #[cfg(unix)]
             mmap_map_stage: GET_STAGE_READER_MMAP_MAP,
+            #[cfg(unix)]
             mmap_copy_stage: GET_STAGE_READER_MMAP_COPY_BUFFER,
+            #[cfg(unix)]
             direct_read_copy_stage: GET_STAGE_READER_MMAP_DIRECT_READ_COPY,
         });
         let mmap_result = {

crates/ecstore/src/object_api/readers.rs

@@ -2278,6 +2278,121 @@ mod tests {
         assert_eq!(read, fixture.plaintext, "SSE-C + compression full GET must reassemble all parts");
     }
 
+    #[tokio::test]
+    async fn multipart_empty_tail_full_reads_preserve_plaintext() {
+        let key = [0x6Eu8; 32];
+        let part_sizes = [5 * 1024 * 1024, 0];
+        let encrypted = build_legacy_ssec_multipart_fixture(key, &part_sizes).await;
+        for (kind, mut fixture, headers) in [
+            (
+                "encrypted",
+                CompressedMultipartFixture {
+                    object_info: encrypted.object_info,
+                    stored: encrypted.ciphertext,
+                    plaintext: encrypted.plaintext,
+                },
+                ssec_headers_from_key(key),
+            ),
+            ("compressed", compressed_multipart_fixture(&part_sizes).await, HeaderMap::new()),
+            (
+                "compressed and encrypted",
+                compressed_encrypted_multipart_fixture(key, &part_sizes).await,
+                ssec_headers_from_key(key),
+            ),
+        ] {
+            fixture.object_info.etag = Some(faster_hex::hex_string(Md5::digest(&fixture.plaintext).as_ref()));
+            assert_eq!(fixture.object_info.etag.as_ref().expect("source ETag").len(), 32);
+            assert_eq!(fixture.object_info.parts.len(), 2);
+            let tail = &fixture.object_info.parts[1];
+            assert_eq!(tail.actual_size, 0, "{kind}: final part has no plaintext");
+            if kind == "compressed" {
+                assert_eq!(tail.size, 0, "unpadded compression emits no bytes for an empty part");
+            } else {
+                assert!(tail.size > 0, "{kind}: the empty part still has a stored frame");
+            }
+            let stored_size = i64::try_from(fixture.stored.len()).expect("fixture size fits i64");
+            let (mut reader, offset, length) = GetObjectReader::new(
+                Box::new(Cursor::new(fixture.stored)),
+                None,
+                &fixture.object_info,
+                &ObjectOptions::default(),
+                &headers,
+            )
+            .await
+            .expect("full transformed read must include the empty tail");
+            assert_eq!((offset, length), (0, stored_size), "{kind}: full read includes all stored parts");
+            let mut body = Vec::new();
+            reader
+                .stream
+                .read_to_end(&mut body)
+                .await
+                .expect("read through the complete decoder EOF");
+            assert_eq!(body, fixture.plaintext, "{kind}: no plaintext is added or lost by the empty tail");
+        }
+    }
+
+    #[tokio::test]
+    async fn multipart_empty_tail_full_read_authenticates_v2_final_frame() {
+        let key = [0x6Eu8; 32];
+        let plaintext = legacy_fixture_part_plaintext(1, 5 * 1024 * 1024);
+        let mut ciphertext = Vec::new();
+        let mut parts = Vec::new();
+        for (number, body) in [(1, plaintext.as_slice()), (2, b"".as_slice())] {
+            let start = ciphertext.len();
+            rustfs_rio::EncryptReader::new_multipart_v2(Cursor::new(body), key, LEGACY_FIXTURE_BASE_NONCE, number)
+                .read_to_end(&mut ciphertext)
+                .await
+                .expect("encrypt a v2 fixture part with an authenticated final frame");
+            parts.push(ObjectPartInfo {
+                number,
+                size: ciphertext.len() - start,
+                actual_size: i64::try_from(body.len()).expect("fixture plaintext size fits"),
+                ..Default::default()
+            });
+        }
+        let tail_start = parts[0].size;
+        assert_eq!(parts[1].actual_size, 0);
+        assert!(parts[1].size > 8, "the empty final frame carries more than an END marker");
+        let object_info = ObjectInfo {
+            bucket: "bucket".to_string(),
+            name: "v2-empty-tail".to_string(),
+            size: i64::try_from(ciphertext.len()).expect("fixture ciphertext size fits"),
+            etag: Some(faster_hex::hex_string(Md5::digest(&plaintext).as_ref())),
+            parts: Arc::new(parts),
+            user_defined: Arc::new(legacy_ssec_multipart_metadata(key, plaintext.len())),
+            ..Default::default()
+        };
+        for corrupt_tail in [false, true] {
+            let mut stored = ciphertext.clone();
+            if corrupt_tail {
+                // The v2 header is authenticated associated data, including
+                // the header of a final frame containing zero plaintext.
+                stored[tail_start + 5] ^= 1;
+            }
+            let (mut reader, offset, length) = GetObjectReader::new(
+                Box::new(Cursor::new(stored)),
+                None,
+                &object_info,
+                &ObjectOptions::default(),
+                &ssec_headers_from_key(key),
+            )
+            .await
+            .expect("construct the full reader before consuming the final frame");
+            assert_eq!((offset, length), (0, object_info.size));
+            let result = tokio::io::copy(&mut reader.stream, &mut tokio::io::sink()).await;
+            if corrupt_tail {
+                let err = result.expect_err("EOF must authenticate the empty final frame after all plaintext is returned");
+                assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+                assert_eq!(err.to_string(), "v2 encrypted frame failed authentication");
+            } else {
+                assert_eq!(
+                    result.expect("valid empty final frame must reach EOF"),
+                    u64::try_from(plaintext.len()).expect("plaintext length fits")
+                );
+            }
+        }
+    }
+
     #[tokio::test]
     async fn compressed_encrypted_multipart_range_crosses_part_boundary() {
         let key_bytes = [0x6Eu8; 32];
@@ -3656,6 +3771,61 @@ mod tests {
         .await;
     }
 
+    #[tokio::test]
+    async fn multipart_full_read_preserves_legacy_zero_and_negative_part_sizes() {
+        let key = [0x77; 32];
+        let part_sizes = [5 * 1024 * 1024, 1024 * 1024];
+        let encrypted = build_legacy_ssec_multipart_fixture(key, &part_sizes).await;
+        // The encrypted case supplies the fixture key explicitly. This covers
+        // full decrypted reads, not managed-key acquisition.
+        for (kind, fixture, headers) in [
+            ("compressed", compressed_multipart_fixture(&part_sizes).await, HeaderMap::new()),
+            (
+                "encrypted with supplied key",
+                CompressedMultipartFixture {
+                    object_info: encrypted.object_info,
+                    stored: encrypted.ciphertext,
+                    plaintext: encrypted.plaintext,
+                },
+                ssec_headers_from_key(key),
+            ),
+        ] {
+            let source_etag = faster_hex::hex_string(Md5::digest(&fixture.plaintext).as_ref());
+            assert_eq!(source_etag.len(), 32);
+            assert_eq!(fixture.plaintext.len(), 6 * 1024 * 1024);
+            for part_index in 0..part_sizes.len() {
+                assert!(fixture.object_info.parts[part_index].actual_size > 0, "the selected part is nonempty");
+                for actual_size in [0, -1] {
+                    let mut object_info = fixture.object_info.clone();
+                    object_info.etag = Some(source_etag.clone());
+                    Arc::make_mut(&mut object_info.parts)[part_index].actual_size = actual_size;
+                    let (mut reader, offset, length) = GetObjectReader::new(
+                        Box::new(Cursor::new(fixture.stored.clone())),
+                        None,
+                        &object_info,
+                        &ObjectOptions::default(),
+                        &headers,
+                    )
+                    .await
+                    .expect("the authoritative total size must keep full legacy reads available");
+                    assert_eq!(offset, 0);
+                    assert_eq!(length, i64::try_from(fixture.stored.len()).expect("stored size fits"));
+                    let mut body = Vec::new();
+                    reader
+                        .stream
+                        .read_to_end(&mut body)
+                        .await
+                        .expect("full read must reach EOF despite an unspecified per-part logical size");
+                    assert_eq!(
+                        body, fixture.plaintext,
+                        "{kind}: part {part_index} with actual_size={actual_size} must not lose readable data"
+                    );
+                    assert_eq!(reader.object_info.etag.as_deref(), Some(source_etag.as_str()));
+                }
+            }
+        }
+    }
+
     /// The physical part sizes must add up to `oi.size` for a seek to be safe;
     /// inconsistent metadata must fall back to the previous full-object read
     /// instead of scheduling an erasure read past the object end.

crates/ecstore/src/object_api/types.rs

@@ -1597,7 +1597,7 @@ impl ObjectInfo {
     }
 
     pub fn is_multipart(&self) -> bool {
-        self.etag.as_ref().is_some_and(|v| v.len() != 32)
+        self.parts.len() > 1 || self.etag.as_ref().is_some_and(|v| v.len() != 32)
     }
 
     pub fn is_encrypted(&self) -> bool {
@@ -2235,6 +2235,35 @@ mod tests {
     }
     use rustfs_filemeta::{FileInfo, FileMeta, MetaCacheEntry, TRANSITION_COMPLETE};
 
+    #[test]
+    fn multipart_identity_uses_stored_parts_and_preserves_the_etag_fallback() {
+        let plain_etag = "0123456789abcdef0123456789abcdef";
+        let multipart_etag = "0123456789abcdef0123456789abcdef-1";
+        for (case, part_count, etag, expected) in [
+            ("preserved source ETag", 2, Some(plain_etag), true),
+            ("missing ETag", 2, None, true),
+            ("ordinary PUT", 1, Some(plain_etag), false),
+            ("ordinary PUT without ETag", 1, None, false),
+            ("single-part MPU", 1, Some(multipart_etag), true),
+            ("legacy MPU without parts", 0, Some(multipart_etag), true),
+        ] {
+            let object = ObjectInfo {
+                etag: etag.map(str::to_string),
+                parts: Arc::new(
+                    (1..=part_count)
+                        .map(|number| ObjectPartInfo {
+                            number,
+                            ..Default::default()
+                        })
+                        .collect(),
+                ),
+                ..Default::default()
+            };
+
+            assert_eq!(object.is_multipart(), expected, "{case}");
+        }
+    }
+
     fn inline_fast_path_object(size: i64, versioned: bool) -> ObjectInfo {
         ObjectInfo {
             size,

crates/ecstore/src/services/notification_sys.rs

@@ -14,7 +14,8 @@
 
 use crate::bucket::lifecycle::tier_last_day_stats::DailyAllTierStats;
 use crate::cluster::rpc::{
-    PeerRestClient, ScannerPeerActivity, ScannerPeerDirtyUsageSnapshot, ScannerPublicationLease, TierConfigReloadOutcome,
+    PeerRestClient, ScannerDirtyUsageAcknowledgement, ScannerPeerActivity, ScannerPeerDirtyUsageSnapshot,
+    ScannerPublicationLease, TierConfigReloadOutcome,
 };
 use crate::diagnostics::admin_server_info::get_commit_id;
 use crate::disk::DiskAPI;
@@ -33,11 +34,11 @@ use rustfs_madmin::net::NetInfo;
 use rustfs_madmin::{ItemState, ServerProperties, StorageInfo};
 use rustfs_utils::XHost;
 use sha2::{Digest, Sha256};
-use std::collections::{BTreeMap, HashMap, hash_map::DefaultHasher};
+use std::collections::{BTreeMap, BTreeSet, HashMap, hash_map::DefaultHasher};
 use std::future::Future;
 use std::hash::{Hash, Hasher};
 use std::sync::{
-    Arc, Mutex, OnceLock,
+    Arc, LazyLock, Mutex, OnceLock,
     atomic::{AtomicBool, AtomicUsize, Ordering},
 };
 use std::time::{Duration, Instant, SystemTime};
@@ -72,6 +73,29 @@ const LOCAL_CROSS_POOL_FENCE_POLICY_SUPPORTED_VERSION: u32 = 4;
 /// service must not advertise this version until the conditional writer from
 /// rustfs/backlog#684 is available.
 const LEGACY_TRANSITION_STATE_RECONCILE_POLICY_SUPPORTED_VERSION: u32 = 5;
+
+fn resolve_admin_peer_probe_timeout_secs(configured: Option<u64>) -> u64 {
+    configured
+        .filter(|seconds| *seconds > 0)
+        .unwrap_or(rustfs_config::DEFAULT_ADMIN_PEER_PROBE_TIMEOUT_SECS)
+        .min(rustfs_config::MAX_ADMIN_PEER_PROBE_TIMEOUT_SECS)
+}
+
+fn admin_peer_probe_timeout() -> Duration {
+    let configured = rustfs_utils::get_env_opt_u64_with_aliases(rustfs_config::ENV_ADMIN_PEER_PROBE_TIMEOUT_SECS, &[]);
+    let seconds = resolve_admin_peer_probe_timeout_secs(configured);
+    Duration::from_secs(seconds)
+}
+
+fn remaining_admin_peer_probe_timeout(deadline: Instant) -> Option<Duration> {
+    remaining_admin_peer_probe_timeout_at(deadline, Instant::now())
+}
+
+fn remaining_admin_peer_probe_timeout_at(deadline: Instant, now: Instant) -> Option<Duration> {
+    let remaining = deadline.saturating_duration_since(now);
+    (!remaining.is_zero()).then_some(remaining)
+}
+
 type CrossPoolFencePolicyResult = Result<BTreeMap<String, Uuid>>;
 
 fn cross_pool_fence_policy_results(
@@ -288,12 +312,29 @@ pub struct LegacyTransitionStateReconcileFleetProofToken {
     _permit: FleetCapabilityProofPermit,
 }
 
+/// Effect-window authority for one immutable ILM recovery export.
+pub struct IlmRecoveryExportFleetProofToken {
+    token: FleetCapabilityProofToken,
+    _permit: FleetCapabilityProofPermit,
+}
+
+/// Effect-window authority for emitting the compact transition-transaction
+/// state sequence. The generation permit prevents a successor proof from
+/// being published until the admitted writer has finished.
+pub(crate) struct TransitionTransactionCompactionFleetProofToken {
+    token: FleetCapabilityProofToken,
+    _permit: FleetCapabilityProofPermit,
+}
+
 static REMOTE_VERSION_STATE_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static CROSS_POOL_FENCE_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static TIER_DELETE_JOURNAL_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static DECOMMISSION_TARGET_FENCE_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static LEGACY_TRANSITION_STATE_RECONCILE_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
+static ILM_RECOVERY_EXPORT_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
+static TRANSITION_TRANSACTION_COMPACTION_FLEET_PROOF: OnceLock<std::sync::RwLock<FleetCapabilityProofState>> = OnceLock::new();
 static REMOTE_VERSION_STATE_PROBE_TOPOLOGY: OnceLock<String> = OnceLock::new();
+static ILM_RECOVERY_EXPORT_LOCAL_PROCESS_EPOCH: LazyLock<Uuid> = LazyLock::new(Uuid::new_v4);
 
 fn cross_pool_fence_fleet_proof_slot() -> &'static std::sync::RwLock<FleetCapabilityProofState> {
     CROSS_POOL_FENCE_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
@@ -315,6 +356,14 @@ fn legacy_transition_state_reconcile_fleet_proof_slot() -> &'static std::sync::R
     LEGACY_TRANSITION_STATE_RECONCILE_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
 }
 
+fn ilm_recovery_export_fleet_proof_slot() -> &'static std::sync::RwLock<FleetCapabilityProofState> {
+    ILM_RECOVERY_EXPORT_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
+}
+
+fn transition_transaction_compaction_fleet_proof_slot() -> &'static std::sync::RwLock<FleetCapabilityProofState> {
+    TRANSITION_TRANSACTION_COMPACTION_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(FleetCapabilityProofState::default()))
+}
+
 fn revoke_fleet_capability_proof_state(state: &mut FleetCapabilityProofState) {
     if let Some(proof) = state.proof.take() {
         proof.generation.revoke();
@@ -415,6 +464,33 @@ pub(crate) fn remote_version_state_fleet_proof_matches(proof: &RemoteVersionStat
     fleet_capability_proof_matches(remote_version_state_fleet_proof_slot(), &proof.0)
 }
 
+pub(crate) fn acquire_transition_transaction_compaction_fleet_proof() -> Option<TransitionTransactionCompactionFleetProofToken> {
+    let expected_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get()?;
+    let state = transition_transaction_compaction_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    let token = acquire_fleet_capability_proof_from(&state, expected_topology, Instant::now())?;
+    let permit = state.proof.as_ref()?.generation.try_acquire()?;
+    Some(TransitionTransactionCompactionFleetProofToken { token, _permit: permit })
+}
+
+pub(crate) fn transition_transaction_compaction_fleet_proof_matches(
+    proof: &TransitionTransactionCompactionFleetProofToken,
+) -> bool {
+    let Some(expected_topology) = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() else {
+        return false;
+    };
+    let state = transition_transaction_compaction_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    proof._permit.generation.is_accepting()
+        && fleet_capability_proof_matches_at(&state, &proof.token, expected_topology, Instant::now())
+        && state
+            .proof
+            .as_ref()
+            .is_some_and(|current| Arc::ptr_eq(&current.generation, &proof._permit.generation))
+}
+
 pub fn acquire_cross_pool_fence_fleet_proof() -> Option<CrossPoolFenceFleetProofToken> {
     let expected_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get()?;
     let state = cross_pool_fence_fleet_proof_slot()
@@ -550,6 +626,117 @@ pub async fn legacy_transition_state_reconcile_fleet_proof_matches(
     .await
 }
 
+pub async fn acquire_ilm_recovery_export_fleet_proof() -> Option<IlmRecoveryExportFleetProofToken> {
+    let expected_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get()?;
+    let proof = {
+        let state = ilm_recovery_export_fleet_proof_slot()
+            .read()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        acquire_ilm_recovery_export_fleet_proof_from(&state, expected_topology, Instant::now())?
+    };
+    let observed = observe_ilm_recovery_export_fleet(expected_topology).await?;
+    let state = ilm_recovery_export_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    ilm_recovery_export_fleet_proof_matches_observation_at(&state, &proof, expected_topology, &observed, Instant::now())
+        .then_some(proof)
+}
+
+fn acquire_ilm_recovery_export_fleet_proof_from(
+    state: &FleetCapabilityProofState,
+    expected_topology: &str,
+    now: Instant,
+) -> Option<IlmRecoveryExportFleetProofToken> {
+    let token = acquire_fleet_capability_proof_from(state, expected_topology, now)?;
+    let permit = state.proof.as_ref()?.generation.try_acquire()?;
+    Some(IlmRecoveryExportFleetProofToken { token, _permit: permit })
+}
+
+pub async fn ilm_recovery_export_fleet_proof_matches(proof: &IlmRecoveryExportFleetProofToken) -> bool {
+    let Some(expected_topology) = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() else {
+        return false;
+    };
+    {
+        let state = ilm_recovery_export_fleet_proof_slot()
+            .read()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        if !ilm_recovery_export_fleet_proof_matches_at(&state, proof, expected_topology, Instant::now()) {
+            return false;
+        }
+    }
+    let Some(observed) = observe_ilm_recovery_export_fleet(expected_topology).await else {
+        return false;
+    };
+    let state = ilm_recovery_export_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    ilm_recovery_export_fleet_proof_matches_observation_at(&state, proof, expected_topology, &observed, Instant::now())
+}
+
+pub fn ilm_recovery_export_topology_generation(proof: &IlmRecoveryExportFleetProofToken) -> String {
+    let mut hasher = Sha256::new();
+    hasher.update(b"rustfs-ilm-recovery-export-topology-v1\0");
+    hasher.update(proof.token.topology_fingerprint.as_bytes());
+    rustfs_utils::crypto::hex(hasher.finalize().as_slice())
+}
+
+pub fn ilm_recovery_export_member_epochs_sha256(proof: &IlmRecoveryExportFleetProofToken) -> String {
+    let encoded = serde_json::to_vec(proof.token.peer_epochs.as_ref()).expect("member epoch map is JSON encodable");
+    let mut hasher = Sha256::new();
+    hasher.update(b"rustfs-ilm-recovery-export-members-v1\0");
+    hasher.update(encoded);
+    rustfs_utils::crypto::hex(hasher.finalize().as_slice())
+}
+
+pub fn ilm_recovery_export_local_process_epoch() -> Uuid {
+    *ILM_RECOVERY_EXPORT_LOCAL_PROCESS_EPOCH
+}
+
+fn ilm_recovery_export_fleet_proof_matches_at(
+    state: &FleetCapabilityProofState,
+    proof: &IlmRecoveryExportFleetProofToken,
+    expected_topology: &str,
+    now: Instant,
+) -> bool {
+    proof._permit.generation.is_accepting()
+        && fleet_capability_proof_matches_at(state, &proof.token, expected_topology, now)
+        && state
+            .proof
+            .as_ref()
+            .is_some_and(|current| Arc::ptr_eq(&current.generation, &proof._permit.generation))
+}
+
+fn ilm_recovery_export_fleet_proof_matches_observation_at(
+    state: &FleetCapabilityProofState,
+    proof: &IlmRecoveryExportFleetProofToken,
+    expected_topology: &str,
+    observed: &BTreeMap<String, Uuid>,
+    now: Instant,
+) -> bool {
+    ilm_recovery_export_fleet_proof_matches_at(state, proof, expected_topology, now)
+        && proof.token.peer_epochs.as_ref() == observed
+}
+
+async fn observe_ilm_recovery_export_fleet(expected_topology: &str) -> Option<BTreeMap<String, Uuid>> {
+    #[cfg(test)]
+    {
+        let state = ilm_recovery_export_fleet_proof_slot()
+            .read()
+            .unwrap_or_else(std::sync::PoisonError::into_inner);
+        if fleet_capability_proof_valid_at(state.proof.as_ref(), expected_topology, Instant::now()) {
+            return state.proof.as_ref().map(|proof| proof.peer_epochs.as_ref().clone());
+        }
+    }
+    let notification_sys = get_global_notification_sys()?;
+    timeout(
+        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+        notification_sys.probe_ilm_recovery_export_fleet(expected_topology),
+    )
+    .await
+    .ok()?
+    .ok()
+}
+
 async fn legacy_transition_state_reconcile_fleet_proof_matches_with_observer<F, Fut>(
     slot: &std::sync::RwLock<FleetCapabilityProofState>,
     proof: &LegacyTransitionStateReconcileFleetProofToken,
@@ -638,7 +825,7 @@ pub(crate) fn install_cross_pool_fence_fleet_proof_for_test() {
         state.proof.clone()
     } else {
         Some(FleetCapabilityProof::new(
-            topology,
+            topology.clone(),
             Arc::new(BTreeMap::new()),
             now + Duration::from_secs(60 * 60),
         ))
@@ -671,6 +858,21 @@ pub(crate) fn install_cross_pool_fence_fleet_proof_for_test() {
     decommission_state.topology_conflict = false;
     decommission_state.draining_generation = None;
     decommission_state.proof = proof.as_ref().map(FleetCapabilityProof::with_fresh_generation);
+    drop(decommission_state);
+    let mut export_state = ilm_recovery_export_fleet_proof_slot()
+        .write()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    if !fleet_capability_proof_valid_at(export_state.proof.as_ref(), &topology, now) {
+        debug_assert!(
+            export_state
+                .proof
+                .as_ref()
+                .is_none_or(|current| current.generation.is_drained())
+        );
+        export_state.topology_conflict = false;
+        export_state.draining_generation = None;
+        export_state.proof = proof.as_ref().map(FleetCapabilityProof::with_fresh_generation);
+    }
 }
 
 #[cfg(test)]
@@ -911,6 +1113,35 @@ pub(crate) fn install_remote_version_state_fleet_proof_for_test(topology_fingerp
     RemoteVersionStateFleetProofGuard
 }
 
+#[cfg(all(test, feature = "test-util"))]
+pub(crate) struct TransitionTransactionCompactionFleetProofGuard;
+
+#[cfg(all(test, feature = "test-util"))]
+impl Drop for TransitionTransactionCompactionFleetProofGuard {
+    fn drop(&mut self) {
+        revoke_fleet_capability_proof(transition_transaction_compaction_fleet_proof_slot());
+    }
+}
+
+#[cfg(all(test, feature = "test-util"))]
+pub(crate) fn install_transition_transaction_compaction_fleet_proof_for_test(
+    topology_fingerprint: &str,
+) -> TransitionTransactionCompactionFleetProofGuard {
+    let _ = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.set(topology_fingerprint.to_string());
+    let effective_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY
+        .get()
+        .expect("transition transaction compaction test topology should be initialized");
+    if let Some(err) = publish_fleet_capability_probe_result(
+        transition_transaction_compaction_fleet_proof_slot(),
+        effective_topology,
+        Ok(BTreeMap::new()),
+        Instant::now(),
+    ) {
+        panic!("test proof installation must not fail: {err}");
+    }
+    TransitionTransactionCompactionFleetProofGuard
+}
+
 fn insert_remote_version_state_peer(peer_epochs: &mut BTreeMap<String, Uuid>, peer: String, epoch: Uuid) -> Result<()> {
     if epoch.is_nil() || peer_epochs.values().any(|existing| *existing == epoch) || peer_epochs.insert(peer, epoch).is_some() {
         return Err(Error::other("remote version state capability peer identity is invalid"));
@@ -927,6 +1158,8 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
                 tier_delete_journal_fleet_proof_slot(),
                 decommission_target_fence_fleet_proof_slot(),
                 legacy_transition_state_reconcile_fleet_proof_slot(),
+                ilm_recovery_export_fleet_proof_slot(),
+                transition_transaction_compaction_fleet_proof_slot(),
             ] {
                 mark_fleet_capability_topology_conflict(slot);
             }
@@ -936,29 +1169,59 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
 
     tokio::spawn(async move {
         loop {
-            let result = match get_global_notification_sys() {
-                Some(notification_sys) => {
-                    match timeout(
+            let notification_sys = get_global_notification_sys();
+            let remote_version_state_probe = async {
+                match notification_sys.as_ref() {
+                    Some(notification_sys) => timeout(
                         REMOTE_VERSION_STATE_PROBE_TIMEOUT,
                         notification_sys.probe_remote_version_state_fleet(&topology_fingerprint),
                     )
                     .await
-                    {
-                        Ok(result) => result,
-                        Err(_) => Err(Error::other("remote version state fleet capability probe timed out")),
-                    }
+                    .unwrap_or_else(|_| Err(Error::other("remote version state fleet capability probe timed out"))),
+                    None => Err(Error::other("remote version state fleet capability notification system is unavailable")),
                 }
-                None => Err(Error::other("remote version state fleet capability notification system is unavailable")),
             };
-            let fence_probe = match get_global_notification_sys() {
-                Some(notification_sys) => timeout(
-                    REMOTE_VERSION_STATE_PROBE_TIMEOUT,
-                    notification_sys.probe_cross_pool_fence_fleet(&topology_fingerprint),
-                )
-                .await
-                .unwrap_or_else(|_| Err(Error::other("cross-pool fence fleet capability probe timed out"))),
-                None => Err(Error::other("cross-pool fence fleet capability notification system is unavailable")),
+            let cross_pool_fence_probe = async {
+                match notification_sys.as_ref() {
+                    Some(notification_sys) => timeout(
+                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                        notification_sys.probe_cross_pool_fence_fleet(&topology_fingerprint),
+                    )
+                    .await
+                    .unwrap_or_else(|_| Err(Error::other("cross-pool fence fleet capability probe timed out"))),
+                    None => Err(Error::other("cross-pool fence fleet capability notification system is unavailable")),
+                }
             };
+            let recovery_export_probe = async {
+                match notification_sys.as_ref() {
+                    Some(notification_sys) => timeout(
+                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                        notification_sys.probe_ilm_recovery_export_fleet(&topology_fingerprint),
+                    )
+                    .await
+                    .unwrap_or_else(|_| Err(Error::other("ILM recovery export fleet capability probe timed out"))),
+                    None => Err(Error::other("ILM recovery export fleet capability notification system is unavailable")),
+                }
+            };
+            let transition_transaction_compaction_probe = async {
+                match notification_sys.as_ref() {
+                    Some(notification_sys) => timeout(
+                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                        notification_sys.probe_transition_transaction_compaction_fleet(&topology_fingerprint),
+                    )
+                    .await
+                    .unwrap_or_else(|_| Err(Error::other("transition transaction compaction fleet capability probe timed out"))),
+                    None => Err(Error::other(
+                        "transition transaction compaction fleet capability notification system is unavailable",
+                    )),
+                }
+            };
+            let (result, fence_probe, recovery_export_result, transition_transaction_compaction_result) = tokio::join!(
+                remote_version_state_probe,
+                cross_pool_fence_probe,
+                recovery_export_probe,
+                transition_transaction_compaction_probe
+            );
             let (fence_result, journal_result, decommission_target_fence_result, reconcile_result) = match fence_probe {
                 Ok((peer_epochs, minimum_version)) => cross_pool_fence_policy_results(peer_epochs, minimum_version),
                 Err(err) => {
@@ -981,6 +1244,8 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
                 revoke_fleet_capability_proof(tier_delete_journal_fleet_proof_slot());
                 revoke_fleet_capability_proof(decommission_target_fence_fleet_proof_slot());
                 revoke_fleet_capability_proof(legacy_transition_state_reconcile_fleet_proof_slot());
+                revoke_fleet_capability_proof(ilm_recovery_export_fleet_proof_slot());
+                revoke_fleet_capability_proof(transition_transaction_compaction_fleet_proof_slot());
             } else if let Some(err) = publish_fleet_capability_probe_result(
                 remote_version_state_fleet_proof_slot(),
                 &topology_fingerprint,
@@ -1007,6 +1272,42 @@ pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
                     "notification capability probe"
                 );
             }
+            if !topology_conflict
+                && let Some(err) = publish_fleet_capability_probe_result(
+                    ilm_recovery_export_fleet_proof_slot(),
+                    &topology_fingerprint,
+                    recovery_export_result,
+                    Instant::now(),
+                )
+            {
+                debug!(
+                    event = EVENT_NOTIFICATION_CAPABILITY_PROBE,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_NOTIFICATION,
+                    capability = "ilm_recovery_export_v1",
+                    state = "failed_closed",
+                    error = %err,
+                    "notification capability probe"
+                );
+            }
+            if !topology_conflict
+                && let Some(err) = publish_fleet_capability_probe_result(
+                    transition_transaction_compaction_fleet_proof_slot(),
+                    &topology_fingerprint,
+                    transition_transaction_compaction_result,
+                    Instant::now(),
+                )
+            {
+                debug!(
+                    event = EVENT_NOTIFICATION_CAPABILITY_PROBE,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_NOTIFICATION,
+                    capability = "transition_transaction_compaction_v1",
+                    state = "failed_closed",
+                    error = %err,
+                    "notification capability probe"
+                );
+            }
             if !topology_conflict
                 && let Some(err) = publish_fleet_capability_probe_result(
                     tier_delete_journal_fleet_proof_slot(),
@@ -1129,6 +1430,28 @@ impl NotificationSys {
         Ok(peer_epochs)
     }
 
+    async fn probe_transition_transaction_compaction_fleet(&self, topology_fingerprint: &str) -> Result<BTreeMap<String, Uuid>> {
+        if self.peer_clients.len() != self.peer_topology_hosts.len() {
+            return Err(Error::other(
+                "transition transaction compaction capability fleet membership is incomplete",
+            ));
+        }
+        let probes = self.peer_clients.iter().map(|client| async {
+            let client = client
+                .as_ref()
+                .ok_or_else(|| Error::other("transition transaction compaction capability peer is unreachable"))?;
+            client
+                .probe_transition_transaction_compaction(topology_fingerprint.to_string())
+                .await
+        });
+        let mut peer_epochs = BTreeMap::new();
+        for result in join_all(probes).await {
+            let (peer, epoch) = result?;
+            insert_remote_version_state_peer(&mut peer_epochs, peer, epoch)?;
+        }
+        Ok(peer_epochs)
+    }
+
     async fn probe_cross_pool_fence_fleet(&self, topology_fingerprint: &str) -> Result<(BTreeMap<String, Uuid>, u32)> {
         if self.peer_clients.len() != self.peer_topology_hosts.len() {
             return Err(Error::other("cross-pool fence capability fleet membership is incomplete"));
@@ -1151,6 +1474,46 @@ impl NotificationSys {
         }
         Ok((peer_epochs, minimum_version))
     }
+
+    async fn probe_ilm_recovery_export_fleet(&self, topology_fingerprint: &str) -> Result<BTreeMap<String, Uuid>> {
+        if self.peer_clients.len() != self.peer_topology_hosts.len() {
+            return Err(Error::other("ILM recovery export capability fleet membership is incomplete"));
+        }
+        let local_member = runtime_sources::local_node_name().await;
+        if local_member.trim().is_empty() {
+            return Err(Error::other("ILM recovery export local member identity is unavailable"));
+        }
+        let mut peer_epochs = BTreeMap::new();
+        insert_remote_version_state_peer(&mut peer_epochs, local_member.clone(), ilm_recovery_export_local_process_epoch())?;
+        let probes = self.peer_clients.iter().map(|client| async {
+            let client = client
+                .as_ref()
+                .ok_or_else(|| Error::other("ILM recovery export capability peer is unreachable"))?;
+            client.probe_ilm_recovery_export(topology_fingerprint.to_string()).await
+        });
+        for result in join_all(probes).await {
+            let (peer, epoch) = result?;
+            insert_remote_version_state_peer(&mut peer_epochs, peer, epoch)?;
+        }
+        validate_ilm_recovery_export_members(&self.peer_topology_hosts, &local_member, &peer_epochs)?;
+        Ok(peer_epochs)
+    }
+}
+
+fn validate_ilm_recovery_export_members(
+    expected_remote_members: &[String],
+    local_member: &str,
+    observed: &BTreeMap<String, Uuid>,
+) -> Result<()> {
+    let expected = expected_remote_members
+        .iter()
+        .cloned()
+        .chain(std::iter::once(local_member.to_string()))
+        .collect::<BTreeSet<_>>();
+    if expected.len() != expected_remote_members.len().saturating_add(1) || observed.keys().ne(expected.iter()) {
+        return Err(Error::other("ILM recovery export capability fleet membership does not match topology"));
+    }
+    Ok(())
 }
 
 /// Rolling tier activity summed over every cluster member that answered, with
@@ -1538,7 +1901,7 @@ impl NotificationSys {
     {
         let mut futures = Vec::with_capacity(self.peer_clients.len());
         let endpoints = runtime_sources::endpoint_pools().unwrap_or_else(|| Vec::new().into());
-        let peer_timeout = Duration::from_secs(5);
+        let peer_timeout = admin_peer_probe_timeout();
 
         for (idx, client) in self.peer_clients.iter().enumerate() {
             let endpoints = endpoints.clone();
@@ -1546,7 +1909,9 @@ impl NotificationSys {
             futures.push(async move {
                 if let Some(client) = client {
                     let host = client.host.to_string();
-                    match timeout(peer_timeout, client.local_storage_info()).await {
+                    let deadline = Instant::now() + peer_timeout;
+                    let probe_timeout = remaining_admin_peer_probe_timeout(deadline).unwrap_or_default();
+                    match timeout(probe_timeout, client.local_storage_info()).await {
                         Ok(Ok(mut info)) => {
                             normalize_and_cache_peer_storage_info(cache, &host, &mut info);
                             Some(info)
@@ -1557,7 +1922,6 @@ impl NotificationSys {
                         }
                         Err(_) => {
                             warn!("peer {} storage_info timed out after {:?}", host, peer_timeout);
-                            client.evict_connection().await;
                             handle_peer_failure(cache, &host, &endpoints)
                         }
                     }
@@ -1583,7 +1947,7 @@ impl NotificationSys {
     pub async fn server_info(&self) -> Vec<ServerProperties> {
         let mut futures = Vec::with_capacity(self.peer_clients.len());
         let endpoints = runtime_sources::endpoint_pools().unwrap_or_else(|| Vec::new().into());
-        let peer_timeout = Duration::from_secs(5);
+        let peer_timeout = admin_peer_probe_timeout();
 
         for (idx, client) in self.peer_clients.iter().enumerate() {
             let host = self
@@ -1600,12 +1964,23 @@ impl NotificationSys {
                     };
                 };
 
+                let deadline = Instant::now() + peer_timeout;
+                let Some(first_timeout) = remaining_admin_peer_probe_timeout(deadline) else {
+                    let health = peer_disk_health_with_deadline(&host, deadline).await;
+                    return PeerServerInfoProbe {
+                        host,
+                        result: Err(PeerServerInfoProbeFailure::Rpc { health }),
+                    };
+                };
+
                 // First attempt. A single evicted or half-open internode channel
                 // is enough to fail one probe and, before retrying, would drop
-                // the member to unknown/offline for this whole snapshot. So on any
-                // first-attempt failure we evict the channel and re-dial once
-                // before falling back (rustfs/backlog#1049, P1-B).
-                match timeout(peer_timeout, client.server_info()).await {
+                // the member to unknown/offline for this whole snapshot. On a
+                // quick failure we evict the channel and re-dial once before
+                // falling back (rustfs/backlog#1049, P1-B). A slow attempt
+                // consumes the round budget and therefore does not trigger a
+                // second full wait or an asynchronous eviction side effect.
+                match timeout(first_timeout, client.server_info()).await {
                     Ok(Ok(info)) => {
                         return PeerServerInfoProbe { host, result: Ok(info) };
                     }
@@ -1619,14 +1994,37 @@ impl NotificationSys {
                 // `evict_connection` would leave that gate up and the retry would
                 // fast-fail with "temporarily offline" instead of reconnecting
                 // (rustfs/backlog#1049 P1-B).
-                client.prepare_retry().await;
+                let Some(retry_budget) = remaining_admin_peer_probe_timeout(deadline) else {
+                    let health = peer_disk_health_with_deadline(&host, deadline).await;
+                    return PeerServerInfoProbe {
+                        host,
+                        result: Err(PeerServerInfoProbeFailure::Rpc { health }),
+                    };
+                };
+                // Bound connection-cache cleanup too. The helper clears the offline gate even
+                // when eviction itself times out, so cancellation cannot strand this peer in
+                // fast-fail mode.
+                if !client.prepare_retry_with_timeout(retry_budget).await {
+                    let health = peer_disk_health_with_deadline(&host, deadline).await;
+                    return PeerServerInfoProbe {
+                        host,
+                        result: Err(PeerServerInfoProbeFailure::Rpc { health }),
+                    };
+                }
 
                 // Second and final attempt on the fresh channel.
-                match timeout(peer_timeout, client.server_info()).await {
+                let Some(retry_timeout) = remaining_admin_peer_probe_timeout(deadline) else {
+                    let health = peer_disk_health_with_deadline(&host, deadline).await;
+                    return PeerServerInfoProbe {
+                        host,
+                        result: Err(PeerServerInfoProbeFailure::Rpc { health }),
+                    };
+                };
+                match timeout(retry_timeout, client.server_info()).await {
                     Ok(Ok(info)) => PeerServerInfoProbe { host, result: Ok(info) },
                     Ok(Err(err)) => {
                         warn!("peer {host} server_info failed after retry: {err}");
-                        let health = peer_disk_health(&host).await;
+                        let health = peer_disk_health_with_deadline(&host, deadline).await;
                         PeerServerInfoProbe {
                             host,
                             result: Err(PeerServerInfoProbeFailure::Rpc { health }),
@@ -1634,8 +2032,7 @@ impl NotificationSys {
                     }
                     Err(_) => {
                         warn!("peer {host} server_info timed out after retry ({peer_timeout:?})");
-                        client.evict_connection().await;
-                        let health = peer_disk_health(&host).await;
+                        let health = peer_disk_health_with_deadline(&host, deadline).await;
                         PeerServerInfoProbe {
                             host,
                             result: Err(PeerServerInfoProbeFailure::Rpc { health }),
@@ -2284,11 +2681,70 @@ impl NotificationSys {
         Ok(snapshots)
     }
 
-    pub async fn acknowledge_scanner_dirty_usage(&self, acknowledgements: Vec<(String, String, u64)>) -> Result<bool> {
+    pub async fn scanner_scoped_dirty_usage_capabilities(
+        &self,
+        acknowledgements: Vec<ScannerDirtyUsageAcknowledgement>,
+    ) -> Result<bool> {
         let mut by_host = HashMap::with_capacity(acknowledgements.len());
-        for (host, instance_id, generation) in acknowledgements {
-            if by_host.insert(host.clone(), (instance_id, generation)).is_some() {
-                return Err(Error::other(format!("duplicate scanner dirty usage acknowledgement target: {host}")));
+        for acknowledgement in acknowledgements {
+            let host = match &acknowledgement {
+                ScannerDirtyUsageAcknowledgement::Scoped { host, .. } => host.clone(),
+                ScannerDirtyUsageAcknowledgement::Generation { .. } => {
+                    return Err(Error::other("scanner scoped dirty usage capability requires scoped acknowledgements"));
+                }
+            };
+            if by_host.insert(host.clone(), acknowledgement).is_some() {
+                return Err(Error::other("duplicate scanner dirty usage acknowledgement target"));
+            }
+        }
+
+        let clients = self
+            .peer_clients
+            .iter()
+            .flatten()
+            .map(|client| (client.grid_host.clone(), client.clone()))
+            .collect::<HashMap<_, _>>();
+        let mut futures = Vec::with_capacity(by_host.len());
+        for (host, acknowledgement) in by_host {
+            let Some(client) = clients.get(&host).cloned() else {
+                return Err(Error::other("scanner scoped dirty usage capability failed: peer is not reachable"));
+            };
+            futures.push(async move {
+                let ScannerDirtyUsageAcknowledgement::Scoped {
+                    owner_id,
+                    instance_id,
+                    entries,
+                    ..
+                } = acknowledgement
+                else {
+                    unreachable!("scoped acknowledgement was validated before probing");
+                };
+                timeout(
+                    SCANNER_ACTIVITY_PROBE_TIMEOUT,
+                    client.scanner_scoped_dirty_usage_capability(owner_id, instance_id, entries),
+                )
+                .await
+                .map_err(|_| Error::other("scanner scoped dirty usage capability timed out"))?
+            });
+        }
+
+        for result in join_all(futures).await {
+            if !result? {
+                return Ok(false);
+            }
+        }
+        Ok(true)
+    }
+
+    pub async fn acknowledge_scanner_dirty_usage(&self, acknowledgements: Vec<ScannerDirtyUsageAcknowledgement>) -> Result<bool> {
+        let mut by_host = HashMap::with_capacity(acknowledgements.len());
+        for acknowledgement in acknowledgements {
+            let host = match &acknowledgement {
+                ScannerDirtyUsageAcknowledgement::Generation { host, .. }
+                | ScannerDirtyUsageAcknowledgement::Scoped { host, .. } => host.clone(),
+            };
+            if by_host.insert(host.clone(), acknowledgement).is_some() {
+                return Err(Error::other("duplicate scanner dirty usage acknowledgement target"));
             }
         }
 
@@ -2300,18 +2756,34 @@ impl NotificationSys {
             .collect::<HashMap<_, _>>();
         let mut failures = Vec::new();
         let mut futures = Vec::with_capacity(by_host.len());
-        for (host, (instance_id, generation)) in by_host {
+        for (host, acknowledgement) in by_host {
             let Some(client) = clients.get(&host).cloned() else {
                 failures.push(format!("peer {host} scanner dirty usage acknowledgement failed: peer is not reachable"));
                 continue;
             };
             futures.push(async move {
-                let result = scanner_activity_with_timeout(
-                    SCANNER_ACTIVITY_PROBE_TIMEOUT,
-                    &host,
-                    client.acknowledge_scanner_dirty_usage(instance_id, generation),
-                )
-                .await;
+                let result = match acknowledgement {
+                    ScannerDirtyUsageAcknowledgement::Generation {
+                        instance_id, generation, ..
+                    } => {
+                        scanner_activity_with_timeout(
+                            SCANNER_ACTIVITY_PROBE_TIMEOUT,
+                            &host,
+                            client.acknowledge_scanner_dirty_usage(instance_id, generation),
+                        )
+                        .await
+                    }
+                    ScannerDirtyUsageAcknowledgement::Scoped {
+                        owner_id,
+                        instance_id,
+                        entries,
+                        ..
+                    } => {
+                        client
+                            .acknowledge_scanner_scoped_dirty_usage(owner_id, instance_id, entries)
+                            .await
+                    }
+                };
                 (host, result)
             });
         }
@@ -3023,6 +3495,11 @@ async fn peer_disk_health(host: &str) -> Option<PeerDiskHealth> {
     }
 }
 
+async fn peer_disk_health_with_deadline(host: &str, deadline: Instant) -> Option<PeerDiskHealth> {
+    let remaining = remaining_admin_peer_probe_timeout(deadline)?;
+    timeout(remaining, peer_disk_health(host)).await.ok().flatten()
+}
+
 /// Handle a peer failure for server_info: return cached data if available, or
 /// classify the member as `unknown` / `degraded` / `offline` depending on how
 /// many consecutive probes have failed and whether the peer's drives are still
@@ -3445,6 +3922,81 @@ mod tests {
         assert!(captured != restarted.token());
     }
 
+    #[test]
+    fn ilm_recovery_export_member_digest_is_order_independent_and_epoch_bound() {
+        let now = Instant::now();
+        let local_epoch = ilm_recovery_export_local_process_epoch();
+        assert!(!local_epoch.is_nil());
+        assert_eq!(local_epoch, ilm_recovery_export_local_process_epoch());
+        let remote_epoch = Uuid::new_v4();
+        let slot = std::sync::RwLock::new(FleetCapabilityProofState::default());
+        let peers = BTreeMap::from([("node-b".to_string(), remote_epoch), ("node-a".to_string(), local_epoch)]);
+        assert!(publish_fleet_capability_probe_result(&slot, "topology-a", Ok(peers), now).is_none());
+        let proof = {
+            let state = slot.read().expect("export proof slot should not poison");
+            acquire_ilm_recovery_export_fleet_proof_from(&state, "topology-a", now).expect("complete fleet should admit export")
+        };
+        let digest = ilm_recovery_export_member_epochs_sha256(&proof);
+
+        let changed_slot = std::sync::RwLock::new(FleetCapabilityProofState::default());
+        let changed = BTreeMap::from([("node-a".to_string(), local_epoch), ("node-b".to_string(), Uuid::new_v4())]);
+        assert!(publish_fleet_capability_probe_result(&changed_slot, "topology-a", Ok(changed), now).is_none());
+        let changed_proof = {
+            let state = changed_slot.read().expect("export proof slot should not poison");
+            acquire_ilm_recovery_export_fleet_proof_from(&state, "topology-a", now).expect("complete fleet should admit export")
+        };
+        assert_ne!(digest, ilm_recovery_export_member_epochs_sha256(&changed_proof));
+    }
+
+    #[test]
+    fn ilm_recovery_export_members_must_match_the_exact_topology() {
+        let expected_remote = vec!["node-b".to_string()];
+        let local = "node-a";
+        let complete = BTreeMap::from([
+            (local.to_string(), Uuid::new_v4()),
+            (expected_remote[0].clone(), Uuid::new_v4()),
+        ]);
+        assert!(validate_ilm_recovery_export_members(&expected_remote, local, &complete).is_ok());
+
+        let unexpected = BTreeMap::from([(local.to_string(), Uuid::new_v4()), ("node-c".to_string(), Uuid::new_v4())]);
+        assert!(validate_ilm_recovery_export_members(&expected_remote, local, &unexpected).is_err());
+        assert!(
+            validate_ilm_recovery_export_members(&[local.to_string()], local, &complete).is_err(),
+            "the configured remote set cannot repeat the local member"
+        );
+    }
+
+    #[test]
+    fn ilm_recovery_export_restart_revokes_authority_until_permit_drains() {
+        let slot = std::sync::RwLock::new(FleetCapabilityProofState::default());
+        let now = Instant::now();
+        let original = BTreeMap::from([("node-a".to_string(), Uuid::new_v4())]);
+        assert!(publish_fleet_capability_probe_result(&slot, "topology-a", Ok(original), now).is_none());
+        let admitted = {
+            let state = slot.read().expect("export proof slot should not poison");
+            acquire_ilm_recovery_export_fleet_proof_from(&state, "topology-a", now).expect("fresh fleet should admit export")
+        };
+
+        let restarted = BTreeMap::from([("node-a".to_string(), Uuid::new_v4())]);
+        let draining = publish_fleet_capability_probe_result(&slot, "topology-a", Ok(restarted.clone()), now)
+            .expect("restart must wait for the admitted export effect window");
+        assert!(draining.to_string().contains("previous generation to drain"));
+        {
+            let state = slot.read().expect("export proof slot should not poison");
+            assert!(!ilm_recovery_export_fleet_proof_matches_at(&state, &admitted, "topology-a", now));
+            assert!(
+                acquire_ilm_recovery_export_fleet_proof_from(&state, "topology-a", now).is_none(),
+                "successor authority must wait for the old effect window to drain"
+            );
+        }
+        drop(admitted);
+        assert!(
+            publish_fleet_capability_probe_result(&slot, "topology-a", Ok(restarted), now + Duration::from_millis(1)).is_none()
+        );
+        let state = slot.read().expect("export proof slot should not poison");
+        assert!(acquire_ilm_recovery_export_fleet_proof_from(&state, "topology-a", now).is_some());
+    }
+
     #[test]
     fn tier_delete_journal_generation_is_stable_across_members_and_process_restarts() {
         let topology = "topology-a";
@@ -4017,6 +4569,37 @@ mod tests {
         }
     }
 
+    #[test]
+    fn admin_peer_probe_timeout_rejects_zero_and_caps_large_values() {
+        assert_eq!(
+            resolve_admin_peer_probe_timeout_secs(None),
+            rustfs_config::DEFAULT_ADMIN_PEER_PROBE_TIMEOUT_SECS
+        );
+        assert_eq!(
+            resolve_admin_peer_probe_timeout_secs(Some(0)),
+            rustfs_config::DEFAULT_ADMIN_PEER_PROBE_TIMEOUT_SECS
+        );
+        assert_eq!(
+            resolve_admin_peer_probe_timeout_secs(Some(rustfs_config::MAX_ADMIN_PEER_PROBE_TIMEOUT_SECS + 1)),
+            rustfs_config::MAX_ADMIN_PEER_PROBE_TIMEOUT_SECS
+        );
+        assert_eq!(resolve_admin_peer_probe_timeout_secs(Some(7)), 7);
+    }
+
+    #[tokio::test]
+    async fn admin_peer_probe_health_fallback_respects_expired_deadline() {
+        let deadline = Instant::now();
+        assert!(peer_disk_health_with_deadline("peer-1", deadline).await.is_none());
+    }
+
+    #[test]
+    fn admin_peer_probe_deadline_is_shared_across_attempts() {
+        let start = Instant::now();
+        let deadline = start + Duration::from_secs(10);
+        assert!(remaining_admin_peer_probe_timeout_at(deadline, start + Duration::from_secs(6)).is_some());
+        assert!(remaining_admin_peer_probe_timeout_at(deadline, start + Duration::from_secs(10)).is_none());
+    }
+
     #[tokio::test]
     async fn call_peer_with_timeout_returns_value_when_fast() {
         let result = call_peer_with_timeout(
@@ -4365,15 +4948,28 @@ mod tests {
             peer_topology_hosts: Vec::new(),
         };
         let missing = sys
-            .acknowledge_scanner_dirty_usage(vec![("peer-1".to_string(), "0123456789abcdef0123456789abcdef".to_string(), 7)])
+            .acknowledge_scanner_dirty_usage(vec![ScannerDirtyUsageAcknowledgement::Generation {
+                host: "peer-1".to_string(),
+                instance_id: "0123456789abcdef0123456789abcdef".to_string(),
+                generation: 7,
+            }])
             .await
             .expect_err("a missing acknowledgement target must remain pending");
         assert!(missing.to_string().contains("peer is not reachable"));
 
         let duplicate = sys
             .acknowledge_scanner_dirty_usage(vec![
-                ("peer-1".to_string(), "0123456789abcdef0123456789abcdef".to_string(), 7),
-                ("peer-1".to_string(), "0123456789abcdef0123456789abcdef".to_string(), 7),
+                ScannerDirtyUsageAcknowledgement::Generation {
+                    host: "peer-1".to_string(),
+                    instance_id: "0123456789abcdef0123456789abcdef".to_string(),
+                    generation: 7,
+                },
+                ScannerDirtyUsageAcknowledgement::Scoped {
+                    host: "peer-1".to_string(),
+                    owner_id: "11111111-1111-1111-1111-111111111111".to_string(),
+                    instance_id: "0123456789abcdef0123456789abcdef".to_string(),
+                    entries: Vec::new(),
+                },
             ])
             .await
             .expect_err("duplicate acknowledgement targets must be rejected");

crates/ecstore/src/services/tier/mod.rs

@@ -14,6 +14,7 @@
 
 #[cfg(any(test, feature = "test-util"))]
 pub mod test_util;
+#[allow(clippy::module_inception, reason = "preserve the public services::tier::tier path")]
 pub mod tier;
 pub mod tier_admin;
 pub mod tier_config;

crates/ecstore/src/services/tier/test_util.rs

@@ -56,6 +56,7 @@
 
 use std::collections::HashMap;
 use std::io::Cursor;
+#[cfg(feature = "test-util")]
 use std::path::Path;
 use std::sync::{
     Arc,
@@ -68,21 +69,28 @@ use tokio::io::AsyncReadExt;
 use tokio::sync::{Mutex, Notify, RwLock};
 use uuid::Uuid;
 
+#[cfg(feature = "test-util")]
 use crate::disk::endpoint::Endpoint;
+#[cfg(feature = "test-util")]
 use crate::disk::format::FormatV3;
+#[cfg(feature = "test-util")]
 use crate::disk::{DiskAPI, DiskOption, FORMAT_CONFIG_FILE, RUSTFS_META_BUCKET, STORAGE_FORMAT_FILE, new_disk};
 use crate::services::tier::tier::TierConfigMgr;
 use crate::services::tier::tier_config::{TierConfig, TierMinIO, TierType};
 use crate::services::tier::warm_backend::{
     TransitionCandidateProbe, WarmBackend, WarmBackendGetOpts, build_transition_put_options,
 };
+#[cfg(feature = "test-util")]
 use rustfs_filemeta::FileMeta;
 use rustfs_s3_client::transition_api::{ReadCloser, ReaderImpl};
+#[cfg(feature = "test-util")]
 use rustfs_utils::path::path_join_buf;
 
 /// One-shot barrier before rejected transition cleanup resolves its ECStore.
+#[cfg(feature = "test-util")]
 pub struct TransitionCleanupStoreBarrier(crate::set_disk::SetDiskTransitionCleanupStoreBarrier);
 
+#[cfg(feature = "test-util")]
 impl TransitionCleanupStoreBarrier {
     /// Install the barrier for the next rejected transition cleanup.
     pub fn install() -> Self {
@@ -96,6 +104,7 @@ impl TransitionCleanupStoreBarrier {
 }
 
 /// Default polling cadence used by the `wait_for_*` helpers.
+#[cfg(feature = "test-util")]
 const POLL_INTERVAL: Duration = Duration::from_millis(50);
 
 /// A fault to inject into [`MockWarmBackend`] operations.
@@ -208,10 +217,12 @@ impl Drop for MockRemoveOperationGuard {
 }
 
 /// One-shot barrier that pauses a mock tier PUT after storing its remote body.
+#[cfg(feature = "test-util")]
 pub struct MockPutBarrier {
     state: Arc<MockPutBarrierState>,
 }
 
+#[cfg(feature = "test-util")]
 impl MockPutBarrier {
     /// Wait until the remote body is stored and the PUT is paused before returning.
     pub async fn wait_until_paused(&self) {
@@ -226,6 +237,7 @@ impl MockPutBarrier {
     }
 }
 
+#[cfg(feature = "test-util")]
 impl Drop for MockPutBarrier {
     fn drop(&mut self) {
         self.state.release.notify_one();
@@ -258,10 +270,12 @@ impl Drop for MockGetBarrier {
 }
 
 /// One-shot barrier that pauses and then fails a mock tier DELETE.
+#[cfg(feature = "test-util")]
 pub struct MockRemoveBarrier {
     state: Arc<MockRemoveBarrierState>,
 }
 
+#[cfg(feature = "test-util")]
 impl MockRemoveBarrier {
     /// Wait until DELETE reaches the deterministic failure point.
     pub async fn wait_until_paused(&self) {
@@ -283,6 +297,7 @@ impl MockRemoveBarrier {
     }
 }
 
+#[cfg(feature = "test-util")]
 impl Drop for MockRemoveBarrier {
     fn drop(&mut self) {
         self.state.release.notify_one();
@@ -306,6 +321,7 @@ impl MockWarmBackend {
     }
 
     /// Arm a one-shot pause after the next tier PUT stores its remote body.
+    #[cfg(feature = "test-util")]
     pub async fn arm_put_barrier(&self) -> MockPutBarrier {
         let state = Arc::new(MockPutBarrierState::default());
         *self.inner.put_barrier.lock().await = Some(Arc::clone(&state));
@@ -313,6 +329,7 @@ impl MockWarmBackend {
     }
 
     /// Pause and then fail the next DELETE after it reaches the backend.
+    #[cfg(feature = "test-util")]
     pub async fn arm_failing_remove_barrier(&self) -> MockRemoveBarrier {
         let state = Arc::new(MockRemoveBarrierState::default());
         let mut barrier = self.inner.remove_barrier.lock().await;
@@ -323,6 +340,7 @@ impl MockWarmBackend {
 
     /// Arm a one-shot pause before the next tier GET, then return an error
     /// after the test releases it.
+    #[cfg(feature = "test-util")]
     pub async fn arm_failing_get_barrier(&self) -> MockGetBarrier {
         let state = Arc::new(MockGetBarrierState {
             fail_after_release: true,
@@ -343,6 +361,7 @@ impl MockWarmBackend {
     // ---- fault injection -------------------------------------------------
 
     /// Replace the entire fault configuration.
+    #[cfg(feature = "test-util")]
     pub async fn set_faults(&self, faults: FaultConfig) {
         *self.inner.faults.lock().await = faults;
     }
@@ -353,6 +372,7 @@ impl MockWarmBackend {
     }
 
     /// Toggle "HTTP 5xx" server errors on every operation.
+    #[cfg(feature = "test-util")]
     pub async fn set_server_error(&self, server_error: bool) {
         self.inner.faults.lock().await.server_error = server_error;
     }
@@ -363,11 +383,13 @@ impl MockWarmBackend {
     }
 
     /// Set (or clear, with `None`) injected latency applied before each op.
+    #[cfg(feature = "test-util")]
     pub async fn set_latency(&self, latency: Option<Duration>) {
         self.inner.faults.lock().await.latency = latency;
     }
 
     /// Clear all injected faults, restoring healthy behaviour.
+    #[cfg(feature = "test-util")]
     pub async fn clear_faults(&self) {
         *self.inner.faults.lock().await = FaultConfig::default();
     }
@@ -375,6 +397,7 @@ impl MockWarmBackend {
     /// Limit how many body bytes a successful mock PUT consumes. `None` drains
     /// the complete body. This models a backend that incorrectly accepts a
     /// truncated stream while still returning success.
+    #[cfg(feature = "test-util")]
     pub async fn set_put_read_limit(&self, limit: Option<usize>) {
         *self.inner.put_read_limit.lock().await = limit;
     }
@@ -395,12 +418,14 @@ impl MockWarmBackend {
     }
 
     /// Reject non-empty remote versions before transition metadata is committed.
+    #[cfg(feature = "test-util")]
     pub fn set_reject_non_empty_remote_versions(&self, reject: bool) {
         self.inner.reject_non_empty_remote_versions.store(reject, Ordering::Release);
     }
 
     /// Reject the next non-empty remote version validation without changing
     /// subsequent exact-version backend cleanup behavior.
+    #[cfg(feature = "test-util")]
     pub fn reject_next_non_empty_remote_version_validation(&self) {
         self.inner
             .reject_non_empty_remote_version_validations
@@ -438,6 +463,7 @@ impl MockWarmBackend {
     }
 
     /// Clear the operation log without touching stored objects or faults.
+    #[cfg(feature = "test-util")]
     pub async fn clear_op_log(&self) {
         self.inner.op_log.lock().await.clear();
     }
@@ -459,11 +485,13 @@ impl MockWarmBackend {
     }
 
     /// Return the exact object/version pairs produced by successful tier PUTs.
+    #[cfg(feature = "test-util")]
     pub async fn put_versions(&self) -> Vec<(String, String)> {
         self.inner.put_versions.lock().await.clone()
     }
 
     /// Return the exact object/version pairs passed to successful tier removes.
+    #[cfg(feature = "test-util")]
     pub async fn remove_versions(&self) -> Vec<(String, String)> {
         self.inner.remove_versions.lock().await.clone()
     }
@@ -475,6 +503,7 @@ impl MockWarmBackend {
 
     /// Number of `get` calls recorded — useful to assert restore reads hit the
     /// local copy rather than the remote tier.
+    #[cfg(feature = "test-util")]
     pub async fn get_count(&self) -> usize {
         self.inner
             .op_log
@@ -486,6 +515,7 @@ impl MockWarmBackend {
     }
 
     /// Number of `put` calls recorded.
+    #[cfg(feature = "test-util")]
     pub async fn put_count(&self) -> usize {
         self.inner
             .op_log
@@ -499,6 +529,7 @@ impl MockWarmBackend {
     // ---- storage inspection ---------------------------------------------
 
     /// Whether the backend currently stores `object`.
+    #[cfg(feature = "test-util")]
     pub async fn contains(&self, object: &str) -> bool {
         self.inner.objects.lock().await.contains_key(object)
     }
@@ -509,11 +540,13 @@ impl MockWarmBackend {
     }
 
     /// A clone of the stored object, if present.
+    #[cfg(feature = "test-util")]
     pub async fn stored(&self, object: &str) -> Option<MockStoredObject> {
         self.inner.objects.lock().await.get(object).cloned()
     }
 
     /// A clone of the raw bytes stored for `object`, if present.
+    #[cfg(feature = "test-util")]
     pub async fn bytes(&self, object: &str) -> Option<Vec<u8>> {
         self.inner.objects.lock().await.get(object).map(|o| o.bytes.clone())
     }
@@ -538,6 +571,7 @@ impl MockWarmBackend {
 
     /// Poll until `object` is absent from the backend, or `timeout` elapses.
     /// Returns `true` if the object disappeared within the budget.
+    #[cfg(feature = "test-util")]
     pub async fn wait_for_remote_absence(&self, object: &str, timeout: Duration) -> bool {
         let deadline = tokio::time::Instant::now() + timeout;
         loop {
@@ -553,6 +587,7 @@ impl MockWarmBackend {
 
     /// Poll until the backend holds exactly `expected` objects, or `timeout`
     /// elapses. Returns `true` if the count was reached within the budget.
+    #[cfg(feature = "test-util")]
     pub async fn wait_for_object_count(&self, expected: usize, timeout: Duration) -> bool {
         let deadline = tokio::time::Instant::now() + timeout;
         loop {
@@ -847,6 +882,7 @@ pub async fn register_mock_tier_backend(handle: &Arc<RwLock<TierConfigMgr>>, tie
 /// The transition-state tuple read from an on-disk `xl.meta`, plus the object's
 /// free-version count.
 #[derive(Clone, Debug, PartialEq, Eq)]
+#[cfg(feature = "test-util")]
 pub struct TransitionMeta {
     /// `transition_status` (e.g. `"complete"`), empty when not transitioned.
     pub status: String,
@@ -860,6 +896,7 @@ pub struct TransitionMeta {
     pub free_version_count: usize,
 }
 
+#[cfg(feature = "test-util")]
 async fn open_disk(disk_path: &Path) -> Option<crate::disk::DiskStore> {
     // `LocalDisk::new` rejects an endpoint whose (set_idx, disk_idx) disagrees
     // with the position recorded in the disk's own format.json, so derive the
@@ -890,6 +927,7 @@ async fn open_disk(disk_path: &Path) -> Option<crate::disk::DiskStore> {
 /// The free-version metadata removal lands asynchronously after the remote
 /// object disappears, so callers typically poll via
 /// [`wait_for_free_version_absence`] instead of asserting a single read.
+#[cfg(feature = "test-util")]
 pub async fn free_version_count(disk_path: &Path, bucket: &str, object: &str) -> usize {
     let Some(disk) = open_disk(disk_path).await else {
         return 0;
@@ -914,6 +952,7 @@ pub async fn free_version_count(disk_path: &Path, bucket: &str, object: &str) ->
 /// fields are taken from the newest version that carries a transition record;
 /// if no version is transitioned, they are taken from the current version (and
 /// will be empty).
+#[cfg(feature = "test-util")]
 pub async fn read_transition_meta(disk_path: &Path, bucket: &str, object: &str) -> Option<TransitionMeta> {
     let disk = open_disk(disk_path).await?;
     let data = disk
@@ -947,6 +986,7 @@ pub async fn read_transition_meta(disk_path: &Path, bucket: &str, object: &str)
 /// disk is missing the object or disagrees — this is the shard-consistency
 /// check required by ilm-6 (the `(status, tier, remote key, remote version id)`
 /// four-tuple plus free-version count must match across all erasure shards).
+#[cfg(feature = "test-util")]
 pub async fn assert_transition_meta_consistent<P: AsRef<Path>>(disk_paths: &[P], bucket: &str, object: &str) -> TransitionMeta {
     assert!(!disk_paths.is_empty(), "assert_transition_meta_consistent needs at least one disk");
 
@@ -972,6 +1012,7 @@ pub async fn assert_transition_meta_consistent<P: AsRef<Path>>(disk_paths: &[P],
 
 /// Poll until `object` retains no free versions on `disk_path`, or `timeout`
 /// elapses. Returns `true` if the free versions drained within the budget.
+#[cfg(feature = "test-util")]
 pub async fn wait_for_free_version_absence(disk_path: &Path, bucket: &str, object: &str, timeout: Duration) -> bool {
     let deadline = tokio::time::Instant::now() + timeout;
     loop {
@@ -1040,6 +1081,44 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn mock_metadata_survives_put_and_external_delete_is_distinct() {
+        let backend = MockWarmBackend::new();
+        let metadata = HashMap::from([
+            ("content-type".to_string(), "text/plain".to_string()),
+            ("project".to_string(), "archive".to_string()),
+        ]);
+        let version = backend
+            .put_with_meta("object", ReaderImpl::Body(Bytes::from_static(b"body")), 4, metadata.clone())
+            .await
+            .expect("mock PUT should preserve remote metadata");
+        assert_eq!(backend.metadata("object").await, Some(metadata));
+        assert_eq!(
+            backend
+                .probe_transition_candidate_state("object")
+                .await
+                .expect("probe stored object"),
+            TransitionCandidateProbe::VersionedPresent(version)
+        );
+
+        backend.external_remove("object").await;
+        assert_eq!(backend.metadata("object").await, None);
+        assert_eq!(
+            backend
+                .probe_transition_candidate_state("object")
+                .await
+                .expect("probe removed object"),
+            TransitionCandidateProbe::Missing
+        );
+        let operations = backend.op_log().await;
+        assert!(
+            operations
+                .iter()
+                .any(|op| matches!(op, MockWarmOp::ExternalRemove { object } if object == "object"))
+        );
+        assert!(!operations.iter().any(|op| matches!(op, MockWarmOp::Remove { .. })));
+    }
+
     #[tokio::test]
     async fn mock_probe_preserves_fault_fail_closed_behavior() {
         let backend = MockWarmBackend::new();

crates/ecstore/src/services/tier/warm_backend.rs

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use crate::error::is_err_bucket_not_found;
 #[cfg(feature = "gcs")]
@@ -719,17 +717,7 @@ async fn check_warm_backend_with_deadlines(
     if !matches!(cleanup_result, Ok(Ok(()))) {
         return Err(probe_cleanup_incomplete_error());
     }
-    if let Err(err) = read_result {
-        //if is_err_bucket_not_found(&err) {
-        //    return Err(ERR_TIER_BUCKET_NOT_FOUND);
-        //}
-        /*else if is_err_signature_does_not_match(err) {
-            return Err(ERR_TIER_MISSING_CREDENTIALS);
-        }*/
-        //else {
-        return Err(err);
-        //}
-    }
+    read_result?;
     Ok(())
 }
 
@@ -759,7 +747,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                     warn!("{}", err);
                     return Err(AdminError {
                         code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                         status_code: StatusCode::BAD_REQUEST,
                     });
                 }
@@ -800,7 +788,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                     warn!("{}", err);
                     return Err(AdminError {
                         code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                         status_code: StatusCode::BAD_REQUEST,
                     });
                 }
@@ -820,7 +808,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                     warn!("{}", err);
                     return Err(AdminError {
                         code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                         status_code: StatusCode::BAD_REQUEST,
                     });
                 }
@@ -840,7 +828,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                     warn!("{}", err);
                     return Err(AdminError {
                         code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                         status_code: StatusCode::BAD_REQUEST,
                     });
                 }
@@ -860,7 +848,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                     warn!("{}", err);
                     return Err(AdminError {
                         code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                         status_code: StatusCode::BAD_REQUEST,
                     });
                 }
@@ -880,7 +868,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                     warn!("{}", err);
                     return Err(AdminError {
                         code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                         status_code: StatusCode::BAD_REQUEST,
                     });
                 }
@@ -900,7 +888,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                     warn!("{}", err);
                     return Err(AdminError {
                         code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                         status_code: StatusCode::BAD_REQUEST,
                     });
                 }
@@ -929,7 +917,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                     warn!("{}", err);
                     return Err(AdminError {
                         code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                         status_code: StatusCode::BAD_REQUEST,
                     });
                 }
@@ -949,7 +937,7 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
                     warn!("{}", err);
                     return Err(AdminError {
                         code: "XRustFSAdminTierInvalidConfig".to_string(),
-                        message: format!("Unable to setup remote tier, check tier configuration: {}", err.to_string()),
+                        message: format!("Unable to setup remote tier, check tier configuration: {err}"),
                         status_code: StatusCode::BAD_REQUEST,
                     });
                 }

crates/ecstore/src/services/tier/warm_backend_s3.rs

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use std::collections::HashMap;
 use std::sync::Arc;
@@ -106,39 +104,35 @@ impl WarmBackendS3 {
         };
         validate_outbound_url(&u).map_err(|err| std::io::Error::other(format!("tier endpoint is not allowed: {err}")))?;
 
-        if conf.aws_role_web_identity_token_file == "" && conf.aws_role_arn != ""
-            || conf.aws_role_web_identity_token_file != "" && conf.aws_role_arn == ""
-        {
+        let has_web_identity_token_file = !conf.aws_role_web_identity_token_file.is_empty();
+        let has_role_arn = !conf.aws_role_arn.is_empty();
+        let has_access_key = !conf.access_key.is_empty();
+        let has_secret_key = !conf.secret_key.is_empty();
+
+        if has_web_identity_token_file != has_role_arn {
             return Err(std::io::Error::other("both the token file and the role ARN are required"));
-        } else if conf.access_key == "" && conf.secret_key != "" || conf.access_key != "" && conf.secret_key == "" {
+        } else if has_access_key != has_secret_key {
             return Err(std::io::Error::other("both the access and secret keys are required"));
-        } else if conf.aws_role
-            && (conf.aws_role_web_identity_token_file != ""
-                || conf.aws_role_arn != ""
-                || conf.access_key != ""
-                || conf.secret_key != "")
-        {
+        } else if conf.aws_role && (has_web_identity_token_file || has_role_arn || has_access_key || has_secret_key) {
             return Err(std::io::Error::other(
                 "AWS Role cannot be activated with static credentials or the web identity token file",
             ));
-        } else if conf.bucket == "" {
+        } else if conf.bucket.is_empty() {
             return Err(std::io::Error::other("no bucket name was provided"));
         }
 
-        let creds: Credentials<Static>;
-
-        if conf.access_key != "" && conf.secret_key != "" {
+        let creds = if has_access_key && has_secret_key {
             //creds = Credentials::new_static_v4(conf.access_key, conf.secret_key, "");
-            creds = Credentials::new(Static(Value {
+            Credentials::new(Static(Value {
                 access_key_id: conf.access_key.clone(),
                 secret_access_key: conf.secret_key.clone(),
                 session_token: "".to_string(),
                 signer_type: SignatureType::SignatureV4,
                 ..Default::default()
-            }));
+            }))
         } else {
             return Err(std::io::Error::other("insufficient parameters for S3 backend authentication"));
-        }
+        };
         let timeouts = transition_client_timeouts_from_env();
         let opts = Options {
             creds,
@@ -162,11 +156,11 @@ impl WarmBackendS3 {
     }
 
     pub fn get_dest(&self, object: &str) -> String {
-        let mut dest_obj = object.to_string();
-        if self.prefix != "" {
-            dest_obj = format!("{}/{}", &self.prefix, object);
+        if self.prefix.is_empty() {
+            object.to_string()
+        } else {
+            format!("{}/{}", self.prefix, object)
         }
-        return dest_obj;
     }
 
     pub(crate) async fn remove_with_result(&self, object: &str, rv: &str) -> Result<RemoveObjectResult, std::io::Error> {
@@ -413,6 +407,10 @@ impl TransitionCandidateVersions {
 }
 
 #[cfg(test)]
+#[allow(
+    clippy::items_after_test_module,
+    reason = "keep parsing tests adjacent to the helpers they cover"
+)]
 mod tests {
     use super::*;
     use rustfs_s3_client::api_s3_datatypes::{ListVersionsResult, Version};
@@ -917,7 +915,7 @@ impl WarmBackend for WarmBackendS3 {
             .list_objects_v2(&self.bucket, &self.prefix, "", "", SLASH_SEPARATOR, 1)
             .await?;
 
-        Ok(result.common_prefixes.len() > 0 || result.contents.len() > 0)
+        Ok(!result.common_prefixes.is_empty() || !result.contents.is_empty())
     }
 }
 

crates/ecstore/src/set_disk/core/io_primitives.rs

@@ -111,10 +111,11 @@ use crate::disk::{
 use crate::erasure::coding::BitrotReader;
 use crate::io_support::bitrot::ShardReader;
 use crate::io_support::bitrot::{
-    BitrotReaderStageMetrics, DeferredReaderStripeHandle, adjust_shard_read_params,
-    create_bitrot_reader_from_bytes_with_stage_metrics, create_deferred_bitrot_reader_with_stripe_handle,
-    object_mmap_read_max_length,
+    BitrotReaderStageMetrics, DeferredReaderStripeHandle, create_bitrot_reader_from_bytes_with_stage_metrics,
+    create_deferred_bitrot_reader_with_stripe_handle,
 };
+#[cfg(unix)]
+use crate::io_support::bitrot::{adjust_shard_read_params, object_mmap_read_max_length};
 use crate::set_disk::runtime_sources;
 use crate::set_disk::shard_source::ShardReadCost;
 use crate::storage_api_contracts::object::ObjectOperations;
@@ -3662,22 +3663,22 @@ async fn rollback_failed_rename(
             let object = object.to_string();
             let disk_namespace_commit_guard = namespace_commit_guard.clone();
             let task = tokio::spawn(async move {
-                let _namespace_commit_guard = disk_namespace_commit_guard;
+                let _namespace_commit_guard = disk_namespace_commit_guard.clone();
                 #[allow(clippy::let_unit_value)]
                 let _task_guard = SetDisks::rename_fanout_task_guard(&object);
                 SetDisks::rename_fanout_barrier(&object, disk_index, rename_fanout_barrier_phase::ROLLBACK).await;
                 #[cfg(test)]
                 rollback_fault_injection::before_undo(&object, disk_index)?;
-                disk.delete_version(
+                disk.undo_write_with_namespace_owner(
                     &bucket,
                     &object,
                     fi,
-                    false,
                     DeleteOptions {
                         undo_write: true,
                         old_data_dir: rollback_dir,
                         ..Default::default()
                     },
+                    disk_namespace_commit_guard.map(|owner| owner as Arc<dyn Send + Sync>),
                 )
                 .await
             });
@@ -4237,7 +4238,7 @@ impl SetDisks {
                     let successful_rename_completion_rank = successful_rename_completion_rank.clone();
                     let namespace_commit_guard = namespace_commit_guard.clone();
                     tasks.spawn(async move {
-                        let _namespace_commit_guard = namespace_commit_guard;
+                        let _namespace_commit_guard = namespace_commit_guard.clone();
                         let mut dispatch_state = RenameDispatchState::NotDispatched;
                         let result = std::panic::AssertUnwindSafe(async {
                             #[allow(clippy::let_unit_value)]
@@ -4272,7 +4273,13 @@ impl SetDisks {
                                     &file_info,
                                     &dst_bucket,
                                     &dst_object,
-                                    scanner_publication_lease_token,
+                                    crate::disk::RenameDataGuards {
+                                        scanner_publication_lease_token,
+                                        namespace_owner: namespace_commit_guard
+                                            .clone()
+                                            .map(|owner| owner as Arc<dyn Send + Sync>),
+                                        ..Default::default()
+                                    },
                                 )
                                 .await;
                             let rejected_before_publication = observed.rejected_before_publication();
@@ -4601,7 +4608,7 @@ impl SetDisks {
             // Keep the storage-owned movement permit attached to the actual
             // fan-out owner, even if the caller future is cancelled.
             let _fanout_publication_scope = fanout_publication_scope;
-            let _namespace_commit_guard = fanout_namespace_commit_guard;
+            let _namespace_commit_guard = fanout_namespace_commit_guard.clone();
             let successful_rename_completion_rank =
                 rustfs_io_metrics::put_stage_metrics_enabled().then(|| Arc::new(AtomicUsize::new(0)));
             let futures = fanout_disks
@@ -4616,6 +4623,7 @@ impl SetDisks {
                     let dst_bucket = fanout_dst_bucket.clone();
                     let successful_rename_completion_rank = successful_rename_completion_rank.clone();
                     let publication_scope = scanner_publication_commit_scope.clone();
+                    let namespace_commit_guard = fanout_namespace_commit_guard.clone();
 
                     async move {
                         let mut dispatch_state = RenameDispatchState::NotDispatched;
@@ -4668,7 +4676,13 @@ impl SetDisks {
                                     file_info,
                                     &dst_bucket,
                                     &dst_object,
-                                    scanner_publication_lease_token,
+                                    crate::disk::RenameDataGuards {
+                                        scanner_publication_lease_token,
+                                        namespace_owner: namespace_commit_guard
+                                            .clone()
+                                            .map(|owner| owner as Arc<dyn Send + Sync>),
+                                        ..Default::default()
+                                    },
                                 )
                                 .await;
                             let rejected_before_publication = observed.rejected_before_publication();
@@ -10859,6 +10873,358 @@ mod tests {
         .await;
     }
 
+    #[cfg(not(windows))]
+    async fn assert_namespace_owner_survives_physical_publication_timeout(allow_early_ack: bool) {
+        use crate::disk::os;
+        use futures::FutureExt;
+
+        temp_env::async_with_vars(
+            [
+                (rustfs_config::ENV_DRIVE_MAX_TIMEOUT_DURATION, Some("60")),
+                (ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true")),
+            ],
+            async {
+                const DISKS: usize = 4;
+                let bucket = "namespace-physical-tail";
+                let object = "inline-overwrite";
+                let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+                let (dirs, disks) = call_counter_local_disks(bucket, DISKS).await;
+                prepare_rename_source_dirs(&dirs, &disks, "source").await;
+                let mut old = metadata_test_fileinfo(object);
+                old.mod_time = Some(OffsetDateTime::now_utc());
+                old.size = 15;
+                old.parts.clear();
+                old.add_object_part(1, "old-etag".to_string(), 15, None, 15, None, None);
+                old.data = Some(Bytes::from_static(b"old-inline-body"));
+                old.set_inline_data();
+                old.metadata.insert("etag".to_string(), "old-etag".to_string());
+                let mut infos = rename_commit_fileinfos(object, DISKS, "new-etag");
+                let mut hooks = Vec::new();
+                let mut entered = Vec::new();
+                let mut releases = Vec::new();
+                let mut publication_paths = Vec::new();
+                for (disk, info) in disks.iter().flatten().zip(&mut infos) {
+                    disk.write_metadata(bucket, bucket, object, old.clone())
+                        .await
+                        .expect("the old inline version must be readable before overwrite");
+                    info.size = 11;
+                    info.parts.clear();
+                    info.add_object_part(1, "new-etag".to_string(), 11, None, 11, None, None);
+                    let crate::disk::Disk::Local(local) = disk.as_ref() else {
+                        panic!("physical publication fixture requires local disks");
+                    };
+                    // Linux IO paths use a mount FD, which is also the namespace lock key.
+                    let destination = local
+                        .get_disk()
+                        .get_object_path_for_io(bucket, object)
+                        .expect("the publication path must resolve through the disk's mount lease");
+                    let (entered_tx, entered_rx) = tokio::sync::oneshot::channel();
+                    let (release_tx, release_rx) = std::sync::mpsc::channel::<()>();
+                    hooks.push(os::prepared_publication_test_hooks::install(
+                        &destination.join(STORAGE_FORMAT_FILE),
+                        move || {
+                            let _ = entered_tx.send(());
+                            // Sender drop also releases the syscall when an earlier assertion fails.
+                            let _ = release_rx.recv();
+                        },
+                    ));
+                    entered.push(entered_rx);
+                    releases.push(release_tx);
+                    publication_paths.push(destination);
+                }
+                let namespace_owner = ctx.begin_namespace_commit();
+                let namespace_probe = Arc::downgrade(&namespace_owner);
+                let receipt = RenameRollbackReceipt::default();
+                let mut rename = Box::pin(SetDisks::rename_data_owned_with_fence(
+                    &disks,
+                    (RUSTFS_META_TMP_BUCKET, "source"),
+                    infos,
+                    (bucket, object),
+                    allow_early_ack,
+                    RenameDataFenceOptions::new(3, None)
+                        .with_rollback_receipt(receipt.clone())
+                        .with_namespace_commit_guard(Some(namespace_owner)),
+                ));
+                tokio::time::timeout(Duration::from_secs(10), async {
+                    tokio::select! {
+                        signals = join_all(entered) => {
+                            assert!(signals.into_iter().all(|signal| signal.is_ok()), "all physical publishers must enter");
+                        }
+                        _ = rename.as_mut() => panic!("rename must not finish before physical publication is paused"),
+                    }
+                })
+                .await
+                .expect("all four prepared metadata renames must reach their blocking syscall");
+                assert!(ctx.namespace_commits_pending());
+                assert_eq!(ctx.namespace_commit_generation(), 1);
+
+                // Every wrapper timer exists before advancing; the physical closures stay blocked.
+                tokio::time::pause();
+                tokio::time::advance(Duration::from_secs(61)).await;
+                tokio::time::resume();
+                let result = tokio::time::timeout(Duration::from_secs(5), rename)
+                    .await
+                    .expect("ordinary disk timeout must not wait for the physical rename");
+                assert!(result.is_err(), "four timed-out disks cannot satisfy write quorum");
+                let report = receipt.0.get().expect("failed fanout must finish rollback accounting");
+                assert_eq!(report.disks.len(), DISKS);
+                assert!(
+                    report
+                        .disks
+                        .iter()
+                        .all(|disk| matches!(disk.outcome, RenameRollbackOutcome::Indeterminate(DiskError::Timeout)))
+                );
+                let pending_before_release = ctx.namespace_commits_pending();
+                let owner_alive_before_release = namespace_probe.upgrade().is_some();
+                let old_snapshot_generation = ctx.namespace_commit_generation();
+                for (disk, destination) in disks.iter().flatten().zip(&publication_paths) {
+                    let root = disk.path();
+                    assert!(
+                        os::acquire_rename_data_mutation_lease(&root, bucket, destination)
+                            .now_or_never()
+                            .is_none(),
+                        "the physical publication must still own object serialization after the async timeout"
+                    );
+                    assert!(
+                        root.join(RUSTFS_META_TMP_BUCKET)
+                            .join("source")
+                            .join(STORAGE_FORMAT_FILE)
+                            .exists()
+                    );
+                    let stored = disk
+                        .read_version(
+                            "",
+                            bucket,
+                            object,
+                            "",
+                            &ReadOptions {
+                                read_data: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .expect("a scanner can still read the complete old metadata while publication is paused");
+                    assert_eq!(stored.size, 15);
+                    assert_eq!(stored.data.as_deref(), Some(b"old-inline-body".as_slice()));
+                }
+                assert_eq!(ctx.namespace_commit_generation(), old_snapshot_generation);
+
+                // Drain real syscalls before checking the regression, including on the RED run.
+                drop(releases);
+                for (disk, destination) in disks.iter().flatten().zip(&publication_paths) {
+                    let root = disk.path();
+                    let lease = tokio::time::timeout(
+                        Duration::from_secs(5),
+                        os::acquire_rename_data_mutation_lease(&root, bucket, destination),
+                    )
+                    .await
+                    .expect("released physical publishers must drain");
+                    drop(lease);
+                }
+                for dir in &dirs {
+                    let reopened = reopen_local_disk(dir).await;
+                    let stored = reopened
+                        .read_version(
+                            "",
+                            bucket,
+                            object,
+                            "",
+                            &ReadOptions {
+                                read_data: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .expect("the detached prepared rename must actually publish after timeout");
+                    assert_eq!(stored.size, 11);
+                    assert_eq!(stored.data.as_deref(), Some(b"inline-body".as_slice()));
+                }
+                // The lease releases its locks before dropping the namespace owner, and the
+                // owner's `Drop` runs after its `Weak` probe stops upgrading, so wait for the
+                // pending counter itself instead of asserting it right after the drain.
+                tokio::time::timeout(Duration::from_secs(5), async {
+                    while ctx.namespace_commits_pending() || namespace_probe.upgrade().is_some() {
+                        tokio::task::yield_now().await;
+                    }
+                })
+                .await
+                .expect("released physical publishers must release namespace ownership");
+                let generation_after_publication = ctx.namespace_commit_generation();
+                assert!(!ctx.namespace_commits_pending());
+                assert!(namespace_probe.upgrade().is_none());
+                assert!(receipt.is_incomplete(), "late publication must not erase failed-write recovery evidence");
+                assert!(
+                    pending_before_release && owner_alive_before_release,
+                    "physical publication outlived namespace accounting: early_ack={allow_early_ack}, \
+                     pending={pending_before_release}, owner_alive={owner_alive_before_release}, \
+                     old_snapshot_generation={old_snapshot_generation}, after_late_publication={generation_after_publication}"
+                );
+                assert!(
+                    generation_after_publication > old_snapshot_generation,
+                    "physical completion must invalidate the scanner's old metadata snapshot"
+                );
+            },
+        )
+        .await;
+    }
+
+    #[cfg(not(windows))]
+    #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
+    async fn rename_full_wait_timeout_keeps_namespace_owner_until_physical_publication() {
+        assert_namespace_owner_survives_physical_publication_timeout(false).await;
+    }
+
+    #[cfg(not(windows))]
+    #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
+    async fn rename_early_ack_timeout_keeps_namespace_owner_until_physical_publication() {
+        assert_namespace_owner_survives_physical_publication_timeout(true).await;
+    }
+
+    #[cfg(not(windows))]
+    #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
+    async fn successful_rename_ack_keeps_physical_tail_owner_after_caller_cancellation() {
+        use crate::disk::os;
+        temp_env::async_with_vars(
+            [
+                (rustfs_config::ENV_DRIVE_MAX_TIMEOUT_DURATION, Some("60")),
+                (ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true")),
+            ],
+            async {
+                let bucket = "physical-ack-tail";
+                let object = "ack-object";
+                let (dirs, disks) = call_counter_local_disks(bucket, 4).await;
+                prepare_rename_source_dirs(&dirs, &disks, "source").await;
+                let mut infos = rename_commit_fileinfos(object, 4, "new-etag");
+                for info in &mut infos {
+                    info.size = 11;
+                    info.parts.clear();
+                    info.add_object_part(1, "new-etag".to_string(), 11, None, 11, None, None);
+                }
+                let disk = disks[3].as_ref().expect("tail disk");
+                let crate::disk::Disk::Local(local) = disk.as_ref() else {
+                    panic!("local fixture");
+                };
+                let destination = local.get_disk().get_object_path_for_io(bucket, object).expect("tail IO path");
+                let (entered_tx, entered_rx) = tokio::sync::oneshot::channel();
+                let (release_tx, release_rx) = std::sync::mpsc::channel::<()>();
+                let _hook = os::prepared_publication_test_hooks::install(&destination.join(STORAGE_FORMAT_FILE), move || {
+                    let _ = entered_tx.send(());
+                    let _ = release_rx.recv();
+                });
+                let ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+                let owner = ctx.begin_namespace_commit();
+                let owner_probe = Arc::downgrade(&owner);
+                let receipt = RenameRollbackReceipt::default();
+                let caller_receipt = receipt.clone();
+                let caller_disks = disks.clone();
+                let (ack_tx, ack_rx) = tokio::sync::oneshot::channel();
+                let caller = tokio::spawn(async move {
+                    let commit = SetDisks::rename_data_owned_with_fence(
+                        &caller_disks,
+                        (RUSTFS_META_TMP_BUCKET, "source"),
+                        infos,
+                        (bucket, object),
+                        true,
+                        RenameDataFenceOptions::new(3, None)
+                            .with_namespace_commit_guard(Some(owner))
+                            .with_rollback_receipt(caller_receipt),
+                    )
+                    .await
+                    .expect("three real disk publications must produce a successful ACK");
+                    assert!(ack_tx.send(commit).is_ok(), "deliver successful ACK");
+                    std::future::pending::<()>().await;
+                });
+                let mut commit = tokio::time::timeout(Duration::from_secs(10), async {
+                    entered_rx.await.expect("physical tail entry");
+                    ack_rx
+                        .await
+                        .expect("ACK must arrive while the fourth disk is physically paused")
+                })
+                .await
+                .expect("successful quorum ACK must not wait for its physical tail");
+                assert_eq!(commit.online_disks.iter().flatten().count(), 3);
+                assert!(!destination.join(STORAGE_FORMAT_FILE).exists(), "tail has not published at ACK");
+                let tail_drain = commit.tail_drain.take().expect("early ACK transfers a real tail handle");
+                drop(commit);
+                caller.abort();
+                assert!(caller.await.expect_err("cancel caller after it delivered ACK").is_cancelled());
+                tokio::time::pause();
+                tokio::time::advance(Duration::from_secs(61)).await;
+                tokio::time::resume();
+                let tail = tokio::time::timeout(Duration::from_secs(5), tail_drain)
+                    .await
+                    .expect("ordinary tail timeout stays bounded after ACK")
+                    .expect("tail owner must not panic")
+                    .expect("successful ACK keeps its convergence result");
+                assert_eq!(tail.convergence, RenameConvergence::PartialCommit);
+                assert!(receipt.0.get().is_none(), "an acknowledged write must never enter rollback");
+                let pending = ctx.namespace_commits_pending();
+                let alive = owner_probe.upgrade().is_some();
+                let generation = ctx.namespace_commit_generation();
+                for disk in disks.iter().flatten().take(3) {
+                    let stored = disk
+                        .read_version(
+                            "",
+                            bucket,
+                            object,
+                            "",
+                            &ReadOptions {
+                                read_data: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .expect("all ACK voters keep the new object after caller cancellation");
+                    assert_eq!(stored.data.as_deref(), Some(b"inline-body".as_slice()));
+                }
+                drop(release_tx);
+                let lease = tokio::time::timeout(
+                    Duration::from_secs(5),
+                    os::acquire_rename_data_mutation_lease(&disk.path(), bucket, &destination),
+                )
+                .await
+                .expect("late physical tail drains");
+                drop(lease);
+                tokio::time::timeout(Duration::from_secs(5), async {
+                    while ctx.namespace_commits_pending() || owner_probe.upgrade().is_some() {
+                        tokio::task::yield_now().await;
+                    }
+                })
+                .await
+                .expect("late physical tail must release namespace ownership");
+                for dir in &dirs {
+                    let stored = reopen_local_disk(dir)
+                        .await
+                        .read_version(
+                            "",
+                            bucket,
+                            object,
+                            "",
+                            &ReadOptions {
+                                read_data: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .expect("successful ACK remains committed on every disk after late publication");
+                    assert_eq!(stored.data.as_deref(), Some(b"inline-body".as_slice()));
+                }
+                assert!(
+                    pending && alive,
+                    "physical ACK tail must retain namespace ownership after the coordinator exits"
+                );
+                assert!(!ctx.namespace_commits_pending());
+                assert!(owner_probe.upgrade().is_none());
+                assert!(ctx.namespace_commit_generation() > generation);
+                assert!(receipt.0.get().is_none(), "late publication cannot change success into rollback");
+            },
+        )
+        .await;
+    }
+
     #[tokio::test]
     #[serial_test::serial(capacity_dirty_scope)]
     async fn rename_rollback_incomplete_receipt_waits_for_undo_barrier() {

crates/ecstore/src/set_disk/mod.rs

@@ -874,7 +874,7 @@ pub(crate) use ops::multipart::NewMultipartUploadCommitObservation;
 pub use ops::multipart::{MultipartCommitBarrier, MultipartCommitPause};
 #[cfg(test)]
 pub(crate) use ops::object::DeleteObjectCommitBarrier;
-#[cfg(any(test, feature = "test-util"))]
+#[cfg(feature = "test-util")]
 pub(crate) use ops::object::TransitionCleanupStoreBarrier as SetDiskTransitionCleanupStoreBarrier;
 #[cfg(all(test, feature = "test-util"))]
 pub(crate) use ops::object::TransitionUploadedCommitBarrier as SetDiskTransitionUploadedCommitBarrier;
@@ -883,6 +883,17 @@ pub(crate) use ops::object::body_cache_plaintext_len;
 pub(crate) use ops::object::cleanup_rejected_transition_upload_durably;
 #[cfg(any(test, feature = "test-util"))]
 pub use ops::object::{PutObjectCommitBarrier, PutObjectCommitPause};
+#[cfg(all(test, feature = "test-util"))]
+pub(crate) use ops::object::{
+    TransitionTransactionKillPoint as SetDiskTransitionTransactionKillPoint,
+    TransitionTransactionKillPointBarrier as SetDiskTransitionTransactionKillPointBarrier,
+};
+#[cfg(all(test, feature = "test-util"))]
+pub(crate) use ops::object::{
+    TransitionTransactionMutationKind as SetDiskTransitionTransactionMutationKind,
+    TransitionTransactionMutationObservation as SetDiskTransitionTransactionMutationObservation,
+    TransitionTransactionMutationProbe as SetDiskTransitionTransactionMutationProbe,
+};
 mod read;
 mod replication;
 pub(crate) mod shard_source;
@@ -4498,6 +4509,12 @@ impl SetDisks {
         &self.ctx
     }
 
+    /// Read the persisted bucket identity through this set's metadata owner.
+    /// Missing or non-authoritative legacy identities remain errors.
+    pub async fn bucket_incarnation_id_from_disk(&self, bucket: &str) -> Result<Uuid> {
+        metadata_sys::get_bucket_incarnation_id_in(&self.ctx, bucket).await
+    }
+
     /// Admit one short scanner cache publication under this set's instance
     /// movement fence. The caller must hold the returned guard through its
     /// final conditional cache write; no scan-round work belongs under it.

crates/ecstore/src/set_disk/ops/multipart.rs

@@ -3301,13 +3301,18 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
             // (rustfs/backlog#1009): CompleteMultipartUpload keeps its pre-commit
             // `get_object_info` lookup, so the backfill has no consumer here yet.
             Self::assign_rename_data_indexes(&mut parts_metadatas);
-            let mut rename_result = SetDisks::rename_data_owned(
+            // Disk deadlines can expire before physical publication or failure undo drains.
+            let mut rename_result = SetDisks::rename_data_owned_with_fence(
                 &commit_disks,
                 (RUSTFS_META_MULTIPART_BUCKET, &commit_upload_id_path),
                 parts_metadatas,
                 (&commit_bucket, &commit_object),
-                write_quorum,
                 commit_allows_early_ack,
+                crate::set_disk::core::io_primitives::RenameDataFenceOptions::new(write_quorum, None)
+                    .with_namespace_commit_guard(
+                        (!crate::bucket::utils::is_meta_bucketname(&commit_bucket))
+                            .then(|| commit_set.ctx.begin_namespace_commit()),
+                    ),
             )
             .await;
             if let Ok(rename_commit) = rename_result.as_mut() {
@@ -6758,6 +6763,301 @@ mod tests {
         .await;
     }
 
+    #[tokio::test]
+    #[serial(capacity_dirty_scope)]
+    async fn complete_multipart_advances_namespace_generation_after_commit() {
+        let (dirs, disks, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "multipart-namespace-commit";
+        let object = "completed-object";
+        let body = vec![0x65; 4096];
+        make_bucket_on_all(&disks, bucket).await;
+        let before = set_disks.ctx.namespace_commit_generation();
+        let (upload_id, parts) =
+            stage_upload_with_create_opts(&set_disks, bucket, object, &body, &ObjectOptions::default()).await;
+        assert_eq!(set_disks.ctx.namespace_commit_generation(), before, "staging is not publication");
+        assert!(!set_disks.ctx.namespace_commits_pending());
+        tokio::time::timeout(
+            Duration::from_secs(10),
+            set_disks
+                .clone()
+                .complete_multipart_upload(bucket, object, &upload_id, parts, &ObjectOptions::default()),
+        )
+        .await
+        .expect("completion must finish")
+        .expect("the four real shards must commit");
+        let mut reader = tokio::time::timeout(
+            Duration::from_secs(5),
+            set_disks.get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default()),
+        )
+        .await
+        .expect("GET after completion must finish")
+        .expect("a successful completion must be immediately readable");
+        let mut actual = Vec::new();
+        tokio::time::timeout(Duration::from_secs(5), reader.stream.read_to_end(&mut actual))
+            .await
+            .expect("the completed body stream must finish")
+            .expect("read completed body");
+        assert_eq!(actual, body);
+        let upload_path = SetDisks::get_upload_id_dir(bucket, object, &upload_id);
+        for dir in &dirs {
+            assert!(!dir.path().join(RUSTFS_META_MULTIPART_BUCKET).join(&upload_path).exists());
+        }
+        assert!(matches!(
+            set_disks.check_upload_id_exists(bucket, object, &upload_id, false).await,
+            Err(StorageError::InvalidUploadID(..))
+        ));
+        assert!(!set_disks.ctx.namespace_commits_pending());
+        assert_eq!(
+            set_disks.ctx.namespace_commit_generation(),
+            before + 2,
+            "one completed MPU must invalidate snapshots at namespace admission and physical retirement"
+        );
+    }
+
+    #[cfg(not(windows))]
+    async fn assert_complete_multipart_physical_namespace_owner(undo: bool) {
+        use crate::disk::os::{self, prepared_publication_test_hooks as hooks};
+        use crate::set_disk::core::io_primitives::rename_fault_injection;
+        use futures::FutureExt;
+
+        temp_env::async_with_vars([(rustfs_config::ENV_DRIVE_MAX_TIMEOUT_DURATION, Some("60"))], async {
+            let (dirs, disks, set_disks) = hermetic_set_disks(4).await;
+            let bucket = "multipart-physical-namespace";
+            let object = if undo { "undo-tail" } else { "publication-tail" };
+            let old_body = vec![0x41; 1024];
+            let new_body = vec![0x62; 4096];
+            make_bucket_on_all(&disks, bucket).await;
+            let old = set_disks
+                .put_object(
+                    bucket,
+                    object,
+                    &mut PutObjReader::from_vec(old_body.clone()),
+                    &ObjectOptions {
+                        write_completion: crate::object_api::WriteCompletion::TailDrained,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("seed a real readable old version");
+            let old_etag = old.etag.expect("the committed old object must have an ETag");
+            let before = set_disks.ctx.namespace_commit_generation();
+            assert!(!set_disks.ctx.namespace_commits_pending());
+            let (upload_id, parts) =
+                stage_upload_with_create_opts(&set_disks, bucket, object, &new_body, &ObjectOptions::default()).await;
+            let new_etag = get_complete_multipart_md5(&parts);
+            let upload_path = SetDisks::get_upload_id_dir(bucket, object, &upload_id);
+            for dir in &dirs {
+                assert!(dir.path().join(RUSTFS_META_MULTIPART_BUCKET).join(&upload_path).exists());
+            }
+            assert_eq!(set_disks.ctx.namespace_commit_generation(), before);
+            let _fault = undo.then(|| rename_fault_injection::fail_rename_on(object, &[2, 3]));
+            let stage = if undo {
+                hooks::Stage::Rename
+            } else {
+                hooks::Stage::PreparedRename
+            };
+            let (entered_tx, mut entered_rx) = tokio::sync::mpsc::unbounded_channel();
+            let mut hooks = Vec::new();
+            let mut releases = Vec::new();
+            let mut mutation_paths = Vec::new();
+            for (index, disk) in disks.iter().enumerate() {
+                let disk::Disk::Local(local) = disk.as_ref() else {
+                    panic!("physical MPU fixture requires local disks");
+                };
+                let destination = local
+                    .get_disk()
+                    .get_object_path_for_io(bucket, object)
+                    .expect("leased IO path");
+                let entered_tx = entered_tx.clone();
+                let (release_tx, release_rx) = std::sync::mpsc::channel::<()>();
+                hooks.push(hooks::install_at(stage, &destination.join(STORAGE_FORMAT_FILE), move || {
+                    let _ = entered_tx.send(index);
+                    // Dropping senders releases every syscall on assertion failure, too.
+                    let _ = release_rx.recv();
+                }));
+                releases.push(release_tx);
+                // Canonical rename serializes the object directory; backup restore
+                // serializes its xl.meta destination. Drain the actual executor key.
+                mutation_paths.push(if undo {
+                    destination.join(STORAGE_FORMAT_FILE)
+                } else {
+                    destination
+                });
+            }
+            drop(entered_tx);
+            let complete_set = set_disks.clone();
+            let complete_upload = upload_id.clone();
+            let mut complete = tokio::spawn(async move {
+                complete_set
+                    .complete_multipart_upload(bucket, object, &complete_upload, parts, &ObjectOptions::default())
+                    .await
+            });
+            let mut complete_joined = false;
+            let mut observed_counts = None;
+            let observations = std::panic::AssertUnwindSafe(async {
+                let expected_publishers = if undo { 2 } else { 4 };
+                let entered = tokio::time::timeout(Duration::from_secs(10), async {
+                    let mut entered = HashSet::new();
+                    while entered.len() < expected_publishers {
+                        tokio::select! {
+                            index = entered_rx.recv() => {
+                                assert!(entered.insert(index.expect("physical publisher must signal entry")));
+                            }
+                            result = &mut complete => {
+                                complete_joined = true;
+                                panic!("completion returned before physical entry: {result:?}");
+                            }
+                        }
+                    }
+                    entered
+                })
+                .await
+                .expect("all expected physical metadata operations must enter");
+                let pending_at_entry = set_disks.ctx.namespace_commits_pending();
+                let generation_at_entry = set_disks.ctx.namespace_commit_generation();
+                for &index in &entered {
+                    let metadata = tokio::time::timeout(
+                        Duration::from_secs(5),
+                        disks[index].read_version("", bucket, object, "", &ReadOptions::default()),
+                    )
+                    .await
+                    .expect("metadata observation must finish while publication is paused")
+                    .expect("metadata before the paused physical action must be readable");
+                    assert_eq!(
+                        metadata.metadata.get("etag"),
+                        Some(if undo { &new_etag } else { &old_etag }),
+                        "undo must follow actual publication; prepared rename must precede publication"
+                    );
+                }
+
+                // The entry signals run inside the real blocking closures, after each
+                // wrapper installed its normal deadline. No quota/external guard disables it.
+                tokio::time::pause();
+                tokio::time::advance(Duration::from_secs(61)).await;
+                tokio::time::resume();
+                let joined = tokio::time::timeout(Duration::from_secs(5), &mut complete).await;
+                complete_joined = joined.is_ok();
+                let result = joined
+                    .expect("ordinary MPU disk/undo deadlines must still return before physical drain")
+                    .expect("completion task must not panic");
+                assert!(result.is_err(), "a timed-out or two-shard commit cannot acknowledge success");
+                let pending_after_timeout = set_disks.ctx.namespace_commits_pending();
+                let generation_after_timeout = set_disks.ctx.namespace_commit_generation();
+                observed_counts = Some((pending_at_entry, generation_at_entry, pending_after_timeout, generation_after_timeout));
+                for &index in &entered {
+                    assert!(
+                        os::acquire_rename_data_mutation_lease(&disks[index].path(), bucket, &mutation_paths[index])
+                            .now_or_never()
+                            .is_none(),
+                        "timed-out physical work must still own its object serialization"
+                    );
+                }
+                for dir in &dirs {
+                    assert!(
+                        dir.path().join(RUSTFS_META_MULTIPART_BUCKET).join(&upload_path).exists(),
+                        "failed completion must not clean the upload staging"
+                    );
+                }
+            })
+            .catch_unwind()
+            .await;
+
+            // Release even after a failed observation, then finish dispatch before
+            // draining every physical key. No per-disk assertion may skip a later drain.
+            drop(releases);
+            drop(hooks);
+            let coordinator_drained = complete_joined
+                || tokio::time::timeout(Duration::from_secs(10), &mut complete).await.is_ok();
+            if !coordinator_drained {
+                complete.abort();
+                let _ = tokio::time::timeout(Duration::from_secs(5), &mut complete).await;
+            }
+            let drains = futures::future::join_all(disks.iter().zip(&mutation_paths).map(|(disk, destination)| async move {
+                tokio::time::timeout(
+                    Duration::from_secs(5),
+                    os::acquire_rename_data_mutation_lease(&disk.path(), bucket, destination),
+                )
+                .await
+                .map(drop)
+            }))
+            .await;
+            let owner_drained = tokio::time::timeout(Duration::from_secs(5), async {
+                while set_disks.ctx.namespace_commits_pending() {
+                    tokio::task::yield_now().await;
+                }
+            })
+            .await;
+            let physical_drained = drains.iter().all(|drain| drain.is_ok());
+            if !coordinator_drained || !physical_drained || owner_drained.is_err() {
+                // A bounded cleanup failure cannot justify deleting roots that a
+                // detached executor might still use. Keep them for diagnosis.
+                let retained = dirs.into_iter().map(TempDir::keep).collect::<Vec<_>>();
+                eprintln!("MPU cleanup incomplete: coordinator={coordinator_drained}, physical={physical_drained}, retained={retained:?}");
+                if let Err(panic) = observations {
+                    std::panic::resume_unwind(panic);
+                }
+                panic!("MPU cleanup did not drain: coordinator={coordinator_drained}, physical={physical_drained}, retained={retained:?}");
+            }
+            if let Err(panic) = observations {
+                std::panic::resume_unwind(panic);
+            }
+            // Preserve the original drain checks after collecting every result.
+            for drained in drains {
+                drained.expect("released physical MPU work must drain");
+            }
+            owner_drained.expect("physical retirement must finish its namespace counter decrement");
+            for disk in &disks {
+                let metadata = disk
+                    .read_version("", bucket, object, "", &ReadOptions::default())
+                    .await
+                    .expect("all disks must expose the expected final metadata");
+                assert_eq!(metadata.metadata.get("etag"), Some(if undo { &old_etag } else { &new_etag }));
+            }
+            let (pending_at_entry, generation_at_entry, pending_after_timeout, generation_after_timeout) =
+                observed_counts.expect("successful observations must record namespace counters");
+            let mut reader = tokio::time::timeout(
+                Duration::from_secs(5),
+                set_disks.get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default()),
+            )
+            .await
+            .expect("GET after physical drain must finish")
+            .expect("the real final object must be readable");
+            let mut actual = Vec::new();
+            tokio::time::timeout(Duration::from_secs(5), reader.stream.read_to_end(&mut actual))
+                .await
+                .expect("the final object stream must finish")
+                .expect("read final object bytes");
+            assert_eq!(actual, if undo { old_body } else { new_body });
+            assert!(
+                pending_at_entry && pending_after_timeout,
+                "physical MPU work outlived namespace accounting: undo={undo}"
+            );
+            assert_eq!(generation_at_entry, before + 1);
+            assert_eq!(
+                generation_after_timeout, generation_at_entry,
+                "the blocked physical owner cannot retire early"
+            );
+            assert_eq!(set_disks.ctx.namespace_commit_generation(), before + 2);
+            assert!(!set_disks.ctx.namespace_commits_pending());
+
+        })
+        .await;
+    }
+
+    #[cfg(not(windows))]
+    #[tokio::test]
+    #[serial(capacity_dirty_scope)]
+    async fn complete_multipart_timeout_keeps_namespace_owner_until_physical_publication() {
+        assert_complete_multipart_physical_namespace_owner(false).await;
+    }
+
+    #[cfg(not(windows))]
+    #[tokio::test]
+    #[serial(capacity_dirty_scope)]
+    async fn complete_multipart_failed_quorum_keeps_namespace_owner_until_physical_undo() {
+        assert_complete_multipart_physical_namespace_owner(true).await;
+    }
+
     #[tokio::test(flavor = "multi_thread")]
     #[serial]
     async fn complete_multipart_releases_disk_snapshot_before_cleanup() {

crates/ecstore/src/store/multipart.rs

@@ -238,7 +238,7 @@ async fn list_pool_multipart_uploads_for_incarnation(
 }
 
 impl ECStore {
-    #[cfg(test)]
+    #[cfg(all(test, feature = "test-util"))]
     pub(crate) fn reset_data_movement_multipart_discovery_count_for_test(&self) {
         data_movement_multipart_discovery_counts()
             .lock()
@@ -246,7 +246,7 @@ impl ECStore {
             .insert(self.id, 0);
     }
 
-    #[cfg(test)]
+    #[cfg(all(test, feature = "test-util"))]
     pub(crate) fn data_movement_multipart_discovery_count_for_test(&self) -> usize {
         data_movement_multipart_discovery_counts()
             .lock()

crates/filemeta/src/filemeta.rs

@@ -692,10 +692,15 @@ impl FileMeta {
                             }
                         }
 
-                        let old_dir = v.object.as_ref().map(|v| v.data_dir).unwrap_or_default();
+                        // The version stays on disk while the purge replicates
+                        // (status PENDING/FAILED); its data dir must stay with
+                        // it. Returning the dir here made the disk layer delete
+                        // it, which turned every non-inline retained version
+                        // into an unreadable zombie: the purge state could never
+                        // be applied and the bucket could never be deleted.
                         self.set_idx(i, v)?;
 
-                        return Ok(old_dir);
+                        return Ok(None);
                     }
                     found_index = Some(i);
                 }
@@ -2702,6 +2707,58 @@ mod test {
         );
     }
 
+    /// Regression for rustfs/backlog#2340: a version purge that still awaits
+    /// the replication target keeps the object version on disk with a pending
+    /// purge status. Its data dir must be retained with it; handing the dir
+    /// back here made the disk layer delete it, leaving every non-inline
+    /// retained version unreadable. The dir is released only once the purge
+    /// completes and the version itself goes away.
+    #[test]
+    fn delete_version_pending_version_purge_retains_object_data_dir() {
+        let version_id = Uuid::new_v4();
+        let data_dir = Uuid::new_v4();
+        let mut fm = FileMeta::new();
+        let mut fi = FileInfo::new("object", 2, 2);
+        fi.version_id = Some(version_id);
+        fi.data_dir = Some(data_dir);
+        fi.mod_time = Some(OffsetDateTime::now_utc());
+        fm.add_version(fi).unwrap();
+
+        let pending_purge = FileInfo {
+            name: "object".to_string(),
+            version_id: Some(version_id),
+            mark_deleted: true,
+            replication_state_internal: Some(ReplicationState {
+                version_purge_status_internal: Some("target=PENDING;".to_string()),
+                purge_targets: version_purge_statuses_map("target=PENDING;"),
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let freed = fm.delete_version(&pending_purge).unwrap();
+        assert_eq!(freed, None, "a pending purge must not release the retained version's data dir");
+        assert_eq!(fm.versions.len(), 1, "the version must stay until the purge replicates");
+        let retained = fm
+            .into_fileinfo("vol", "object", &version_id.to_string(), false, false, true)
+            .unwrap();
+        assert_eq!(retained.data_dir, Some(data_dir));
+        assert_eq!(retained.version_purge_status(), VersionPurgeStatusType::Pending);
+
+        let completed_purge = FileInfo {
+            name: "object".to_string(),
+            version_id: Some(version_id),
+            replication_state_internal: Some(ReplicationState {
+                version_purge_status_internal: Some("target=COMPLETE;".to_string()),
+                purge_targets: version_purge_statuses_map("target=COMPLETE;"),
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let freed = fm.delete_version(&completed_purge).unwrap();
+        assert_eq!(freed, Some(data_dir), "a completed purge removes the version and releases its data dir");
+        assert!(fm.versions.is_empty());
+    }
+
     #[test]
     fn delete_version_accepts_delete_only_marker_and_free_version_paths() {
         let marker_version_id = Uuid::new_v4();

crates/heal/src/error.rs

@@ -54,6 +54,15 @@ pub enum Error {
     #[error("Heal task execution failed: {message}")]
     TaskExecutionFailed { message: String },
 
+    /// The current page already exhausted its local retry budget. Retrying
+    /// the enclosing bucket would replay pages whose results were counted.
+    #[error("Heal listing failed for bucket {bucket}: {source}")]
+    HealListingFailed {
+        bucket: String,
+        #[source]
+        source: Box<Error>,
+    },
+
     #[error("Invalid heal type: {heal_type}")]
     InvalidHealType { heal_type: String },