feat(fips): route internode RPC HMAC through aws-lc-rs FIPS (#9137 )

* feat(fips): route internode RPC HMAC through aws-lc-rs FIPS * Revise KMS cryptographic compliance details Updated the KMS cryptographic compliance documentation to reflect changes in HMAC authentication and added details about the internode RPC HMAC authentication. Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> Signed-off-by: Hauser Zhi <housemecn@gmail.com> --------- Signed-off-by: Hauser Zhi <housemecn@gmail.com> Co-authored-by: Hauser Zhi <housemecn@gmail.com> Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
fix(heal): preserve lock, format, and repair identity (#9132 )
2026-11-03 09:33:46 +00:00 · 2026-11-03 06:56:56 +00:00 · 2026-11-03 04:59:06 +00:00 · 2026-11-03 02:08:32 +00:00 · 2026-11-03 06:07:41 +08:00 · 2026-11-02 19:07:37 +00:00
91 changed files with 5284 additions and 780 deletions
@@ -1,2 +1,2 @@
-sha256-linux=686dcc075e41a75dde3d37e5bd0393d016469f02f83aa3508f92bfc17f26a4d4
-sha256-darwin=686dcc075e41a75dde3d37e5bd0393d016469f02f83aa3508f92bfc17f26a4d4
+sha256-linux=2f97cbbeb2dbf52919b704a3d07336c162a41f95079d941d74251f9279c494e6
+sha256-darwin=2f97cbbeb2dbf52919b704a3d07336c162a41f95079d941d74251f9279c494e6
@@ -827,7 +827,7 @@ slow-timeout = { period = "120s", terminate-after = 12, grace-period = "10s" }
 # ---------------------------------------------------------------------------
 # e2e-distributed profile — 4-node 4-disk Actions suite
 # ---------------------------------------------------------------------------
-# Storage-sensitive PR / nightly / dispatch lane owned by
+# Nightly and manual-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.
@@ -32,6 +32,8 @@ updates:
      - "overtrue"
      - "majinghe"
    ignore:
+      - dependency-name: "pulsar"
+        versions: [ "6.9.x" ]
      - dependency-name: "object_store"
        versions: [ "0.13.x" ]
      - dependency-name: "libunftp"
@@ -769,12 +769,10 @@
      "workflow": ".github/workflows/e2e-distributed.yml",
      "name": "e2e-distributed",
      "roles": [
-        "pr",
        "periodic",
        "manual"
      ],
      "events": [
-        "pull_request",
        "workflow_dispatch",
        "schedule"
      ],
@@ -849,12 +847,11 @@
      "workflow": ".github/workflows/e2e-upgrade.yml",
      "name": "Upgrade Compatibility",
      "roles": [
-        "pr",
        "periodic",
-        "manual"
+        "manual",
+        "release"
      ],
      "events": [
-        "pull_request",
        "push",
        "schedule",
        "workflow_dispatch"
@@ -113,7 +113,7 @@ jobs:
            printf '%s\n' 'mode=full' >> "$GITHUB_OUTPUT"
            echo "Base CI policy unavailable; running the full matrix."
          fi
-          if ! grep -q '^plan=' "$GITHUB_OUTPUT" || ! grep -q '^plan_contract=2$' "$GITHUB_OUTPUT"; then
+          if ! grep -q '^plan=' "$GITHUB_OUTPUT" || ! grep -q '^plan_contract=3$' "$GITHUB_OUTPUT"; then
            printf '%s\n' 'mode=full' 'connect=true' 'full_validation=true' 'rehearse_shards=false' 'packages=' 'plan=' >> "$GITHUB_OUTPUT"
            echo "Base policy does not cover this workflow contract; retaining the full matrix."
          fi
@@ -497,6 +497,63 @@ jobs:
      - name: Check offline enrollment E2E root boundary
        run: ./scripts/check_offline_enrollment_e2e.sh

+  fips-provider:
+    name: FIPS provider boundaries
+    if: needs.classify-changes.outputs.mode == 'full' && contains(fromJSON('["schedule", "workflow_dispatch"]'), github.event_name)
+    needs: [ quick-checks, classify-changes ]
+    # This builds the RustFS library test target, which exceeds hosted-runner disk capacity.
+    runs-on: sm-standard-4
+    timeout-minutes: 90
+    strategy:
+      fail-fast: false
+      matrix:
+        mode: [default, fips]
+    env:
+      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
+      CARGO_BUILD_JOBS: "2"
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          persist-credentials: false
+
+      - name: Setup Rust environment
+        uses: ./.github/actions/setup
+        with:
+          rust-version: stable
+          cache-shared-key: ci-fips-provider
+          cache-save-if: 'false'
+          install-build-packaging-tools: 'false'
+
+      - name: Install FIPS build tools
+        if: matrix.mode == 'fips'
+        run: sudo apt-get install -y cmake golang clang libclang-dev
+
+      - name: Verify provider installation
+        shell: bash
+        run: |
+          args=()
+          if [[ '${{ matrix.mode }}' == 'fips' ]]; then
+            args+=(--features fips)
+          fi
+          cargo nextest run --locked -p rustfs-tls-runtime --lib "${args[@]}" --no-tests fail -E 'test(provider_policy_isolated)'
+          cargo nextest run --locked -p rustfs-s3-client --lib "${args[@]}" --no-tests fail -E 'test(overflowing_closed_range_is_rejected_before_network_io)'
+          cargo nextest run --locked -p rustfs-targets --lib "${args[@]}" --no-tests fail -E 'test(validate_pulsar)'
+          cargo nextest run --locked -p rustfs-targets --lib "${args[@]}" --no-tests fail -E 'test(webhook_provider_policy_isolated)'
+          if [[ '${{ matrix.mode }}' == 'fips' ]]; then
+            cargo nextest run --locked -p rustfs-protocols --lib --features ftps,webdav,fips --no-tests fail -E 'test(protocols_reject_incompatible_host_provider)'
+            cargo nextest run --locked -p rustfs-targets --lib --features fips --no-tests fail -E 'test(fips_postgres_rejects_incompatible_host_provider)'
+            cargo nextest run --locked -p rustfs-s3-client --lib --features fips --no-tests fail -E 'test(fips_transition_rejects_incompatible_host_provider)'
+          fi
+          tests=(provider_installation_isolated prepared_generation_shares_crypto_and_serves_sni_with_mtls)
+          if [[ '${{ matrix.mode }}' == 'fips' ]]; then
+            tests+=(fips_provider_builds_fips_tls_configs fips_provider_tls_handshake_accepts_and_rejects_algorithms)
+          fi
+          for test_name in "${tests[@]}"; do
+            cargo nextest run --locked -p rustfs --lib "${args[@]}" --no-tests fail -E "test($test_name)"
+          done
+          cargo clippy --locked -p rustfs --lib "${args[@]}" -- -D warnings
+
  test-and-lint-rio-v2:
    name: Test and Lint (rio-v2)
    # Specialized boundary, hardware and alternate-feature lanes run only on
@@ -1397,6 +1454,7 @@ jobs:
      - workspace-shard-evidence
      - core-contract
      - connect-feature-boundaries
+      - fips-provider
      - test-and-lint-rio-v2
      - test-and-lint-protocols
      - build-rustfs-debug-binary
@@ -1424,6 +1482,7 @@ jobs:
    needs:
      - classify-changes
      - connect-feature-boundaries
+      - fips-provider
      - required-checks
      - typos
      - quick-checks
@@ -17,8 +17,8 @@
 # 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.
+# `[profile.e2e-distributed]` in `.config/nextest.toml`. The nightly schedule
+# and manual dispatches 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
@@ -33,32 +33,6 @@
 name: e2e-distributed

 on:
-  pull_request:
-    paths:
-      - "Cargo.lock"
-      - "Cargo.toml"
-      - ".config/nextest.toml"
-      - ".config/e2e-distributed-selection.txt"
-      - ".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:
@@ -12,30 +12,19 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

+# Single-node upgrade and rollback lane.
+#
+# The weekly Monday schedule, version tags, and workflow_dispatch run this
+# suite. Distributed upgrade cases stay on the nightly e2e-distributed lane.
+
 name: Upgrade Compatibility

 on:
-  pull_request:
-    paths:
-      - ".github/workflows/e2e-upgrade.yml"
-      - "crates/e2e_test/src/common.rs"
-      - "crates/e2e_test/src/distributed/**"
-      - "crates/e2e_test/src/fake_s3_target/**"
-      - "crates/e2e_test/src/lib.rs"
-      - "crates/e2e_test/src/replication_extension_test.rs"
-      - "crates/e2e_test/src/rpc_fault_proxy.rs"
-      - "crates/e2e_test/src/upgrade_compatibility_test.rs"
-      - "crates/ecstore/**"
-      - "crates/filemeta/**"
-      - "crates/heal/**"
-      - "crates/kms/**"
-      - "crates/storage-api/**"
-      - "rustfs/**"
-      - "Cargo.lock"
  push:
    tags:
      - "[0-9]*.[0-9]*.[0-9]*"
  schedule:
+    # 03:17 UTC every Monday.
    - cron: "17 3 * * 1"
  workflow_dispatch:

@@ -44,7 +33,7 @@ permissions:

 concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: ${{ github.event_name == 'pull_request' }}
+  cancel-in-progress: false

 env:
  CARGO_TERM_COLOR: always
@@ -346,12 +346,12 @@ jobs:
          path: ${{ steps.publish.outputs.candidate_file }}
          if-no-files-found: error

-      # Publish the deb/rpm pair to the auto-testing repo's `assets` branch so
-      # engineers can download and install the nightly directly. The branch is
-      # a single-commit orphan rewritten on every build, which keeps the repo
-      # small while the latest files stay reachable at stable raw URLs.
      # Publish the deb/rpm pair as assets of the rolling `nightly` release on
      # rustfs/auto-testing (see scripts/release/publish_nightly_assets.sh).
+      # The dispatch below runs only after this upload and the immutable R2
+      # candidate both succeed. It does not replace rustfs-functional-chain.yml,
+      # which still starts from this workflow's workflow_run completion.
+      # kms-vault-lane and kms-vault-ha-failover stay parallel with this job.
      - name: Publish packages to auto-testing release assets
        env:
          ASSETS_TOKEN: ${{ secrets.PF_TESTING_GH_TOKEN }}
@@ -360,6 +360,90 @@ jobs:
          DEB_DATE: ${{ steps.deb.outputs.deb_date }}
          BUILD_REF: ${{ env.NIGHTLY_BUILD_REF }}
        run: bash scripts/release/publish_nightly_assets.sh
+
+      # Cross-repo repository_dispatch cannot use GITHUB_TOKEN. The secret is a
+      # PAT or GitHub App token for rustfs/auto-testing: classic `repo` scope,
+      # or fine-grained Contents write (the permission GitHub checks for
+      # POST /repos/{owner}/{repo}/dispatches). A dispatch failure fails this
+      # step, and therefore this job, so the scheduled failure alert fires.
+      # GitHub allows at most 10 top-level client_payload properties, so the
+      # schema 2 candidate object is nested under `candidate` and the release
+      # fields sit beside it.
+      - name: Dispatch nightly candidate to auto-testing
+        if: ${{ steps.publish.outputs.candidate_file != '' }}
+        env:
+          DISPATCH_TOKEN: ${{ secrets.AUTO_TESTING_DISPATCH_TOKEN }}
+          CANDIDATE_FILE: ${{ steps.publish.outputs.candidate_file }}
+          DEB_DATE: ${{ steps.deb.outputs.deb_date }}
+          SOURCE_REPO: ${{ github.repository }}
+        shell: bash
+        run: |
+          set -euo pipefail
+          : "${DISPATCH_TOKEN:?AUTO_TESTING_DISPATCH_TOKEN is required to dispatch rustfs/auto-testing}"
+          : "${CANDIDATE_FILE:?nightly candidate manifest is required}"
+          : "${DEB_DATE:?DEB_DATE is required}"
+          [[ "${DEB_DATE}" =~ ^[0-9]{4}-[0-9]{2}-[0-9]{2}$ ]]
+          [[ "${SOURCE_REPO}" == "rustfs/rustfs" ]]
+
+          payload="${RUNNER_TEMP}/nightly-candidate-dispatch-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}.json"
+          jq -n --slurpfile candidate "${CANDIDATE_FILE}" \
+            --arg event_type "nightly-candidate-ready" \
+            --arg source_repo "${SOURCE_REPO}" \
+            --arg latest_alias_url "https://dl.rustfs.com/artifacts/rustfs/packages/nightly/rustfs-nightly-latest.deb" \
+            --arg auto_testing_release "https://github.com/rustfs/auto-testing/releases/tag/nightly" \
+            --arg channel "nightly" \
+            --arg deb_date "${DEB_DATE}" \
+            --arg nightly_workflow ".github/workflows/nightly-gnu.yml" \
+            '{
+              event_type: $event_type,
+              source_repo: $source_repo,
+              latest_alias_url: $latest_alias_url,
+              auto_testing_release: $auto_testing_release,
+              channel: $channel,
+              deb_date: $deb_date,
+              nightly_workflow: $nightly_workflow,
+              candidate: $candidate[0]
+            }' > "${payload}"
+          if ! jq -e '
+            .event_type == "nightly-candidate-ready"
+            and .source_repo == "rustfs/rustfs"
+            and .channel == "nightly"
+            and .deb_date == env.DEB_DATE
+            and .nightly_workflow == ".github/workflows/nightly-gnu.yml"
+            and .latest_alias_url == "https://dl.rustfs.com/artifacts/rustfs/packages/nightly/rustfs-nightly-latest.deb"
+            and .auto_testing_release == "https://github.com/rustfs/auto-testing/releases/tag/nightly"
+            and .candidate.schema == 2
+            and ((.candidate | keys | sort) == ["build_run_attempt","build_run_id","package_sha256","package_url","schema","source_ref","source_sha","workflow_sha"])
+            and ((keys | sort) == ["auto_testing_release","candidate","channel","deb_date","event_type","latest_alias_url","nightly_workflow","source_repo"])
+          ' "${payload}" >/dev/null; then
+            echo "nightly dispatch payload does not match the nightly-candidate-ready contract" >&2
+            exit 1
+          fi
+
+          body="${RUNNER_TEMP}/nightly-candidate-dispatch-body-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}.json"
+          jq -n --slurpfile client_payload "${payload}" \
+            '{event_type: "nightly-candidate-ready", client_payload: $client_payload[0]}' > "${body}"
+          response="${RUNNER_TEMP}/nightly-candidate-dispatch-response-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}.txt"
+          status="$(curl -sS --retry 3 --connect-timeout 15 --max-time 60 \
+            -o "${response}" -w '%{http_code}' \
+            -X POST \
+            -H "Authorization: Bearer ${DISPATCH_TOKEN}" \
+            -H "Accept: application/vnd.github+json" \
+            -H "X-GitHub-Api-Version: 2022-11-28" \
+            -H "Content-Type: application/json" \
+            --data-binary @"${body}" \
+            https://api.github.com/repos/rustfs/auto-testing/dispatches)" || {
+            echo "repository_dispatch request to rustfs/auto-testing failed before an HTTP status" >&2
+            exit 1
+          }
+          if [[ "${status}" != "204" ]]; then
+            echo "repository_dispatch to rustfs/auto-testing failed: HTTP ${status}" >&2
+            if [[ -s "${response}" ]]; then
+              cat "${response}" >&2
+            fi
+            exit 1
+          fi
+          echo "dispatched nightly-candidate-ready to rustfs/auto-testing (deb_date=${DEB_DATE}, run=${GITHUB_RUN_ID}, attempt=${GITHUB_RUN_ATTEMPT})"
  # Live-Vault lane for the rustfs-kms suite (rustfs/backlog#1774).
  #
  # RUSTFS_KMS_VAULT_TOKEN is the single switch that adds the Vault KV2 and
@@ -92,6 +92,17 @@ jobs:
          CHAIN_OUTPUT: ${{ runner.temp }}/validation-sweep-candidate.json
        run: |
          set -euo pipefail
+          # Scheduled runs carry no inputs and the resolver only derives a
+          # build from a workflow_run payload (which a schedule cannot
+          # provide), so resolve the newest successful nightly build
+          # explicitly before invoking it.
+          if [ -z "${BUILD_RUN_ID:-}" ]; then
+            LATEST="$(gh api "repos/rustfs/rustfs/actions/workflows/nightly-gnu.yml/runs?branch=main&status=success&per_page=1" --jq '.workflow_runs[0]')"
+            BUILD_RUN_ID="$(printf '%s' "${LATEST}" | jq -r '.id')"
+            BUILD_RUN_ATTEMPT="$(printf '%s' "${LATEST}" | jq -r '.run_attempt')"
+            export BUILD_RUN_ID BUILD_RUN_ATTEMPT
+            echo "scheduled sweep pinned to nightly ${BUILD_RUN_ID} attempt ${BUILD_RUN_ATTEMPT}"
+          fi
          python3 scripts/resolve_functional_candidate.py
          python3 - <<'PY'
          import json
@@ -99,15 +110,15 @@ jobs:
          from pathlib import Path
          root = Path(os.environ["RUNNER_TEMP"]) / "validation-sweep"
          root.mkdir(parents=True, exist_ok=True)
-          candidate = json.loads((Path(os.environ["RUNNER_TEMP"]) / "validation-sweep-candidate.json").read_text())
-          manifest = candidate["manifest"]
+          chain = json.loads((Path(os.environ["RUNNER_TEMP"]) / "validation-sweep-candidate.json").read_text())
+          manifest = chain["candidate"]["manifest"]
          (root / "candidate.json").write_text(json.dumps({
              "package_url": manifest["package_url"],
              "package_sha256": manifest["package_sha256"],
              "source_sha": manifest["source_sha"],
          }, sort_keys=True) + "\n")
          with open(os.environ["GITHUB_OUTPUT"], "a") as output:
-              output.write("testing_sha=" + candidate["testing_sha"] + "\n")
+              output.write("testing_sha=" + chain["testing_sha"] + "\n")
              output.write("package_url=" + manifest["package_url"] + "\n")
              output.write("source_sha=" + manifest["source_sha"] + "\n")
          PY
@@ -121,6 +132,26 @@ jobs:
          path: auto-testing
          persist-credentials: false

+      - name: Prepare validation Python runtime
+        # The pool / ilm / replication validation runners require a managed
+        # Python, exactly like the chain's validation job provides.
+        uses: actions/setup-python@5fda3b95a4ea91299a34e894583c3862153e4b97 # v7.0.0
+        with:
+          python-version: "3.12"
+
+      - name: Install validation Python requirements
+        # Case scripts call python3 and import boto3 at startup. Install the
+        # functional chain's client into a per-run env so that interpreter
+        # sees it without changing the runner's shared Python.
+        run: |
+          set -euo pipefail
+          validation_python="${RUNNER_TEMP}/rustfs-validation-python-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}"
+          test ! -e "${validation_python}"
+          python3 -m venv "${validation_python}"
+          "${validation_python}/bin/python" -m pip install \
+            -r auto-testing/scripts/validation_ilm/requirements.txt
+          printf '%s\n' "${validation_python}/bin" >> "${GITHUB_PATH}"
+
      - name: Select sweep cases
        id: select
        env:
@@ -224,7 +255,8 @@ jobs:
            # topology, this purge guarantees the next case starts clean even
            # when the previous one died mid-run.
            for node in ${RUSTFS_NODES}; do
-              ssh -o BatchMode=yes -o ConnectTimeout=10 "${RUSTFS_SSH_USER}@${node}" '
+              # -n keeps ssh from consuming the case list this loop reads from stdin
+              ssh -n -o BatchMode=yes -o ConnectTimeout=10 "${RUSTFS_SSH_USER}@${node}" '
                SUDO=""; [ "$(id -u)" -ne 0 ] && SUDO="sudo -n"
                ${SUDO} systemctl stop rustfs 2>/dev/null || true
                if ${SUDO} dpkg -l rustfs 2>/dev/null | grep -q "^ii"; then
@@ -243,6 +275,7 @@ jobs:
          SWEEP_ROOT: ${{ runner.temp }}/validation-sweep
          PACKAGE_URL: ${{ steps.candidate.outputs.package_url }}
          SOURCE_SHA: ${{ steps.candidate.outputs.source_sha }}
+          AUTO_TESTING_SHA: ${{ steps.candidate.outputs.testing_sha }}
        run: |
          set -euo pipefail
          python3 - <<'PY'
@@ -351,7 +384,10 @@ jobs:
            ${{ runner.temp }}/validation-sweep/sweep-cases.txt
            ${{ runner.temp }}/validation-sweep/candidate.json
            ${{ runner.temp }}/validation-sweep/results/
-          if-no-files-found: error
+            ${{ runner.temp }}/validation-sweep/run/
+          # warn (not error): an early infra failure legitimately leaves no
+          # per-case evidence, and the failing step is already red.
+          if-no-files-found: warn
          retention-days: 30

      - name: Cleanup environment (after)
@@ -107,32 +107,6 @@ jobs:
        shell: pwsh
        run: cargo test -p rustfs-ecstore --lib multipart_owned_record_parent_barriers_ -- --nocapture

-      # Native multipart completion publishes its payload from a different
-      # parent than its metadata, and abort renames under a precondition; both
-      # need their own Windows handle discipline.
-      - name: Test native multipart complete and abort publication
-        shell: pwsh
-        run: >-
-          cargo test -p rustfs-ecstore --lib --
-          native_abort_missing_state_with_actual_owned_records_never_becomes_legacy_or_missing
-          native_abort_pending_intent_is_not_a_delete_certificate_and_raw_cleanup_preserves_exact_evidence
-          --nocapture
-
-      # Owned-undo reclaim removes a directory it pinned, copy heal persists
-      # fault bytes through a handle, and metadata reads meet a directory.
-      - name: Test native multipart reclaim, copy heal, and directory reads
-        shell: pwsh
-        run: >-
-          cargo test -p rustfs-ecstore --lib --
-          native_multipart_owned_undo
-          native_multipart_copy_heal
-          orphan_multipart_cleanup_rejects_metadata_directory_on_one_disk
-          --nocapture
-
-      - name: Test directory removal fallback
-        shell: pwsh
-        run: "cargo test -p rustfs-ecstore --lib disk::fs::tests::test_remove -- --nocapture"
-
      - name: Test Windows path conversion
        shell: pwsh
        run: "cargo test -p rustfs-utils --no-default-features --features path --lib dunce:: -- --nocapture"
@@ -299,7 +299,7 @@ dependencies = [
 "strum",
 "thiserror 2.0.21",
 "uuid",
- "zstd 0.13.3",
+ "zstd 0.13.4",
 ]

 [[package]]
@@ -370,17 +370,17 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3f04e0a89cc175eaf3740282537851e6fd9d7276f97730b9d0ea2859c11a85df"
 dependencies = [
 "arrow-arith",
- "arrow-array",
- "arrow-buffer",
+ "arrow-array 59.3.2",
+ "arrow-buffer 59.3.2",
 "arrow-cast",
 "arrow-csv",
- "arrow-data",
- "arrow-ipc",
+ "arrow-data 59.3.2",
+ "arrow-ipc 59.3.2",
 "arrow-json",
 "arrow-ord",
 "arrow-row",
- "arrow-schema",
- "arrow-select",
+ "arrow-schema 59.3.2",
+ "arrow-select 59.3.2",
 "arrow-string",
 ]

@@ -390,10 +390,10 @@ version = "59.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "157943c720919e8f5b6731407ce5cd4c9a5ef47a75bf10edaf24b7081b5763ef"
 dependencies = [
- "arrow-array",
- "arrow-buffer",
- "arrow-data",
- "arrow-schema",
+ "arrow-array 59.3.2",
+ "arrow-buffer 59.3.2",
+ "arrow-data 59.3.2",
+ "arrow-schema 59.3.2",
 "chrono",
 "num-traits",
 ]
@@ -405,9 +405,9 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "72b5b1682924f11ecf6929689c16cbc69efbb6dd141b06ad98ee1cb10d858c43"
 dependencies = [
 "ahash",
- "arrow-buffer",
- "arrow-data",
- "arrow-schema",
+ "arrow-buffer 59.3.2",
+ "arrow-data 59.3.2",
+ "arrow-schema 59.3.2",
 "chrono",
 "chrono-tz",
 "half",
@@ -417,6 +417,24 @@ dependencies = [
 "num-traits",
 ]

+[[package]]
+name = "arrow-array"
+version = "60.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0f3d22a41d3df78b71df0b685af219ae3a94188ab98d6fb29c958eb55c3247f4"
+dependencies = [
+ "ahash",
+ "arrow-buffer 60.1.0",
+ "arrow-data 60.1.0",
+ "arrow-schema 60.1.0",
+ "chrono",
+ "half",
+ "hashbrown 0.17.1",
+ "num-complex",
+ "num-integer",
+ "num-traits",
+]
+
 [[package]]
 name = "arrow-buffer"
 version = "59.3.2"
@@ -429,18 +447,30 @@ dependencies = [
 "num-traits",
 ]

+[[package]]
+name = "arrow-buffer"
+version = "60.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7bd46a530bc1a8e763102fc36c1dc0f51f3ffa150c78e7f5658ccb54b1458ade"
+dependencies = [
+ "bytes",
+ "half",
+ "num-bigint 0.5.1",
+ "num-traits",
+]
+
 [[package]]
 name = "arrow-cast"
 version = "59.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3f23c3379feede42135ef27ca60e67f6e643cd4343ae74e4c8b16c83f55afd85"
 dependencies = [
- "arrow-array",
- "arrow-buffer",
- "arrow-data",
+ "arrow-array 59.3.2",
+ "arrow-buffer 59.3.2",
+ "arrow-data 59.3.2",
 "arrow-ord",
- "arrow-schema",
- "arrow-select",
+ "arrow-schema 59.3.2",
+ "arrow-select 59.3.2",
 "atoi 2.0.0",
 "base64 0.23.1",
 "chrono",
@@ -451,15 +481,26 @@ dependencies = [
 "ryu",
 ]

+[[package]]
+name = "arrow-cmp"
+version = "60.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1913e57a64fbf2c4502fa6d54b547d6a1ae59e3146cbc15321d47468c937abd1"
+dependencies = [
+ "arrow-array 60.1.0",
+ "arrow-buffer 60.1.0",
+ "arrow-schema 60.1.0",
+]
+
 [[package]]
 name = "arrow-csv"
 version = "59.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "85b1984628b091b8319b3c305fce21d04f9e9b2578bb733b0d0108462ae1d698"
 dependencies = [
- "arrow-array",
+ "arrow-array 59.3.2",
 "arrow-cast",
- "arrow-schema",
+ "arrow-schema 59.3.2",
 "chrono",
 "csv",
 "csv-core",
@@ -472,8 +513,21 @@ version = "59.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bb0a031c7ff5e2bdb443117054fb466af1d4df5b7f0ecbe197713dc8b113b1ba"
 dependencies = [
- "arrow-buffer",
- "arrow-schema",
+ "arrow-buffer 59.3.2",
+ "arrow-schema 59.3.2",
+ "half",
+ "num-integer",
+ "num-traits",
+]
+
+[[package]]
+name = "arrow-data"
+version = "60.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "415b7fbc4cf56b080bb5677a4ae74211b5c79e571ed37a1d547633ec07f3b43b"
+dependencies = [
+ "arrow-buffer 60.1.0",
+ "arrow-schema 60.1.0",
 "half",
 "num-integer",
 "num-traits",
@@ -485,14 +539,29 @@ version = "59.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f5dd1afadda2eb821aaf851bb4c0adef60c30cc0f1a29bd02fb99dd07fb9f58d"
 dependencies = [
- "arrow-array",
- "arrow-buffer",
- "arrow-data",
- "arrow-schema",
- "arrow-select",
+ "arrow-array 59.3.2",
+ "arrow-buffer 59.3.2",
+ "arrow-data 59.3.2",
+ "arrow-schema 59.3.2",
+ "arrow-select 59.3.2",
 "flatbuffers",
 "lz4_flex",
- "zstd 0.13.3",
+ "zstd 0.13.4",
+]
+
+[[package]]
+name = "arrow-ipc"
+version = "60.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "45f3e766bccf5a0cb57a5de52a4cba450dea5171547c2cb5939b486d08887cba"
+dependencies = [
+ "arrow-array 60.1.0",
+ "arrow-buffer 60.1.0",
+ "arrow-data 60.1.0",
+ "arrow-schema 60.1.0",
+ "arrow-select 60.1.0",
+ "flatbuffers",
+ "zstd 0.14.3",
 ]

 [[package]]
@@ -501,12 +570,12 @@ version = "59.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "e2bc4d56a0dad84cc51c5812ccff48b642047a1f3be80349cd84318983e79edb"
 dependencies = [
- "arrow-array",
- "arrow-buffer",
+ "arrow-array 59.3.2",
+ "arrow-buffer 59.3.2",
 "arrow-cast",
 "arrow-ord",
- "arrow-schema",
- "arrow-select",
+ "arrow-schema 59.3.2",
+ "arrow-select 59.3.2",
 "chrono",
 "half",
 "indexmap 2.14.2",
@@ -526,11 +595,11 @@ version = "59.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "561b8209faa2cd489764b7fd07d7cb9e8d274d78048c38de59f77c41140cb834"
 dependencies = [
- "arrow-array",
- "arrow-buffer",
- "arrow-data",
- "arrow-schema",
- "arrow-select",
+ "arrow-array 59.3.2",
+ "arrow-buffer 59.3.2",
+ "arrow-data 59.3.2",
+ "arrow-schema 59.3.2",
+ "arrow-select 59.3.2",
 ]

 [[package]]
@@ -539,10 +608,10 @@ version = "59.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0d8c68b87a569f92b90c46cb0a09d86bd9609ddc9b1bf1fc5a503c8c6907c14e"
 dependencies = [
- "arrow-array",
- "arrow-buffer",
- "arrow-data",
- "arrow-schema",
+ "arrow-array 59.3.2",
+ "arrow-buffer 59.3.2",
+ "arrow-data 59.3.2",
+ "arrow-schema 59.3.2",
 "half",
 ]

@@ -556,6 +625,12 @@ dependencies = [
 "serde_json",
 ]

+[[package]]
+name = "arrow-schema"
+version = "60.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "de2c8e714b466404cdc1189e89f892426efea614d938e16c15518c340cd43ba3"
+
 [[package]]
 name = "arrow-select"
 version = "59.3.2"
@@ -563,10 +638,25 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c844141958d805b47d1208e0d12aac53fcf388ce8011639fe82aa492adde0bb4"
 dependencies = [
 "ahash",
- "arrow-array",
- "arrow-buffer",
- "arrow-data",
- "arrow-schema",
+ "arrow-array 59.3.2",
+ "arrow-buffer 59.3.2",
+ "arrow-data 59.3.2",
+ "arrow-schema 59.3.2",
+ "num-traits",
+]
+
+[[package]]
+name = "arrow-select"
+version = "60.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a5af3f6656c5f2c6944f93d9a357a7f204c93d5f349385a44cf46c4a26fbcf30"
+dependencies = [
+ "ahash",
+ "arrow-array 60.1.0",
+ "arrow-buffer 60.1.0",
+ "arrow-cmp",
+ "arrow-data 60.1.0",
+ "arrow-schema 60.1.0",
 "num-traits",
 ]

@@ -576,11 +666,11 @@ version = "59.3.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "bc130892b1a9552da77971924b390e06cf375d1f7e4344a7b7b40ca4426fa470"
 dependencies = [
- "arrow-array",
- "arrow-buffer",
- "arrow-data",
- "arrow-schema",
- "arrow-select",
+ "arrow-array 59.3.2",
+ "arrow-buffer 59.3.2",
+ "arrow-data 59.3.2",
+ "arrow-schema 59.3.2",
+ "arrow-select 59.3.2",
 "memchr",
 "num-traits",
 "regex",
@@ -669,9 +759,9 @@ dependencies = [

 [[package]]
 name = "async-compression"
-version = "0.4.51"
+version = "0.4.52"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9d5755640120b24961f8231e05892e4a4f7e0d04e8aeca161508c9c397946970"
+checksum = "6ad47a740fb906fffdd891e8b6d5fde2511600f7bfaebd68939fa08a14999459"
 dependencies = [
 "compression-codecs",
 "compression-core",
@@ -976,9 +1066,9 @@ dependencies = [

 [[package]]
 name = "aws-sdk-kms"
-version = "1.128.0"
+version = "1.129.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ef519fda62fd0fdc066948f5ef33ecb8ae1bf0187bf1a12e8a775c93c4c47e7e"
+checksum = "5c5cdb29ad25d71c9ee54c408a672c98d90a24a0ea6eb063025483e0e894b411"
 dependencies = [
 "arc-swap",
 "aws-credential-types",
@@ -1001,9 +1091,9 @@ dependencies = [

 [[package]]
 name = "aws-sdk-s3"
-version = "1.156.0"
+version = "1.157.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1072c9aa20159b9b417115342d38c4c53108fc98e5359fcb86b7d4575747eb59"
+checksum = "6ea9794c741520d8baec9462bcd0ad12164d4f5f78d008a8da9d6a0b7222caa8"
 dependencies = [
 "arc-swap",
 "aws-credential-types",
@@ -1037,9 +1127,9 @@ dependencies = [

 [[package]]
 name = "aws-sdk-sso"
-version = "1.118.0"
+version = "1.119.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b4a89786db8ee6dff2c9302a354ef081354fcb0f527e16ee3cb249a6a5cb6df7"
+checksum = "3c414f25ec4aa9dc18f5cc8f55e18848bb509fcda4ce7c77bee6072e95806bc9"
 dependencies = [
 "arc-swap",
 "aws-credential-types",
@@ -1062,9 +1152,9 @@ dependencies = [

 [[package]]
 name = "aws-sdk-ssooidc"
-version = "1.120.0"
+version = "1.121.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "521c6a11a25f1244efddf947523f277a51a63f0a7ec228f0f62d826e88106366"
+checksum = "bc638bb4355e6514334dcc3228cd4622f5ee54d704bd6a9010a5a91d7856d9d9"
 dependencies = [
 "arc-swap",
 "aws-credential-types",
@@ -1087,9 +1177,9 @@ dependencies = [

 [[package]]
 name = "aws-sdk-sts"
-version = "1.123.0"
+version = "1.124.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8754d38d29448bd1d4aa4a6312ed010dfa296b3436eb2c551cc6e42fc6a66a38"
+checksum = "3f8a0dd8fbde04ba58ad722113c7addb11bf94cf549b1ea0eb9ea427fb87198a"
 dependencies = [
 "arc-swap",
 "aws-credential-types",
@@ -1993,9 +2083,9 @@ dependencies = [

 [[package]]
 name = "cc"
-version = "1.10.0"
+version = "1.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "86e055fa6e3e40ad6187ba25b6855eeb635d5271f8e0f0e1e9a5826b8d75bea0"
+checksum = "375989decec1a0fa424442bed520fa12ede43a48475ce8c7417d2c69b32b5f6b"
 dependencies = [
 "find-msvc-tools",
 "jobserver",
@@ -2251,9 +2341,9 @@ dependencies = [

 [[package]]
 name = "compression-codecs"
-version = "0.4.46"
+version = "0.4.47"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "cae36683101b2899f48bd31f8db9839437356188e3b1ff23096e73dab09ba8fe"
+checksum = "84f840cff84d10085260e460023934d0af699e71212b09b2a41452e8bd9cda43"
 dependencies = [
 "brotli 9.1.1",
 "bzip2",
@@ -2499,9 +2589,9 @@ dependencies = [

 [[package]]
 name = "crc32fast"
-version = "1.5.3"
+version = "1.5.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0ba106e844045ddc02c20768ac4b0b7cb45f2cc2ecbd833ac9f5dacb183c2055"
+checksum = "5f97d3416d5337d15c4b0d78c8c91df922ca143f57ec20e2d1c188333047ef04"
 dependencies = [
 "cfg-if",
 ]
@@ -2891,7 +2981,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4cd5317db77ba6bca25c761a8091b9d294a9f4188e4f9dcd0243e6a57de862ac"
 dependencies = [
 "arrow",
- "arrow-schema",
+ "arrow-schema 59.3.2",
 "async-trait",
 "chrono",
 "datafusion-catalog",
@@ -2925,7 +3015,7 @@ dependencies = [
 "log",
 "object_store",
 "parking_lot 0.12.5",
- "parquet",
+ "parquet 59.3.2",
 "sqlparser",
 "tempfile",
 "tokio",
@@ -2989,8 +3079,8 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d26626e12b224b83351239465dcd00f72d47b02fb5e5b65a643e2dc8ed37f687"
 dependencies = [
 "arrow",
- "arrow-ipc",
- "arrow-schema",
+ "arrow-ipc 59.3.2",
+ "arrow-schema 59.3.2",
 "chrono",
 "foldhash 0.2.0",
 "half",
@@ -3001,7 +3091,7 @@ dependencies = [
 "log",
 "num-traits",
 "object_store",
- "parquet",
+ "parquet 59.3.2",
 "recursive",
 "sqlparser",
 "tokio",
@@ -3057,7 +3147,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "65b1b041688cad707fa103c3e7f4c6d89b3134ebf67da0129070dad3381851eb"
 dependencies = [
 "arrow",
- "arrow-ipc",
+ "arrow-ipc 59.3.2",
 "async-trait",
 "bytes",
 "datafusion-common",
@@ -3127,7 +3217,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7c521ae91b6d899d7190812881f45fe9463c222c11d90149f5332a2e56555904"
 dependencies = [
 "arrow",
- "arrow-schema",
+ "arrow-schema 59.3.2",
 "async-trait",
 "bytes",
 "datafusion-common",
@@ -3148,7 +3238,7 @@ dependencies = [
 "log",
 "object_store",
 "parking_lot 0.12.5",
- "parquet",
+ "parquet 59.3.2",
 "tokio",
 ]

@@ -3165,7 +3255,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "39cfdd5ac0769d1e74a885450e49fde8f63ca7ab41eb479a068564aaea80b837"
 dependencies = [
 "arrow",
- "arrow-buffer",
+ "arrow-buffer 59.3.2",
 "async-trait",
 "bytes",
 "dashmap",
@@ -3191,7 +3281,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "86560fd7df4dda924f5474b588c56e7ae8e96dcf158c7327c9cfe9daff448ec5"
 dependencies = [
 "arrow",
- "arrow-schema",
+ "arrow-schema 59.3.2",
 "async-trait",
 "chrono",
 "datafusion-common",
@@ -3227,7 +3317,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "939ae2771da8a3fdc23ab4de610ae8fef9515476626b49d96d0d5c782b105771"
 dependencies = [
 "arrow",
- "arrow-buffer",
+ "arrow-buffer 59.3.2",
 "base64 0.23.1",
 "chrono",
 "chrono-tz",
@@ -3387,7 +3477,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9f002cef849e10c0597a217ec7891c2796c0f8b42d739f5d24abc24496373406"
 dependencies = [
 "arrow",
- "arrow-schema",
+ "arrow-schema 59.3.2",
 "datafusion-common",
 "datafusion-expr",
 "datafusion-expr-common",
@@ -3462,10 +3552,10 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b1451d6817e5ac0dd21819b19949d1c420ab51d748f55dd9a96ecb5418a24c97"
 dependencies = [
 "arrow",
- "arrow-data",
- "arrow-ipc",
+ "arrow-data 59.3.2",
+ "arrow-ipc 59.3.2",
 "arrow-ord",
- "arrow-schema",
+ "arrow-schema 59.3.2",
 "async-trait",
 "bytes",
 "datafusion-common",
@@ -3513,7 +3603,7 @@ version = "55.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "62dc5ce37a5ce1ff5362949776829aded2529a11dbaa13d24b0ffa6a059e106a"
 dependencies = [
- "arrow-schema",
+ "arrow-schema 59.3.2",
 "async-trait",
 "datafusion-common",
 "datafusion-execution",
@@ -3960,6 +4050,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292"
 dependencies = [
 "block-buffer 0.10.4",
+ "const-oid 0.9.6",
 "crypto-common 0.1.7",
 "subtle",
 ]
@@ -5295,9 +5386,9 @@ dependencies = [

 [[package]]
 name = "hotpath"
-version = "0.30.2"
+version = "0.30.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b39e941010d86063ff233bf8325269bf68097fb24ba67000fbddc0a410e8056e"
+checksum = "d9632c1fd213e63c1b88d8f3e59543da4999b17f859cc1e1d67cf8f2207e8935"
 dependencies = [
 "arc-swap",
 "async-channel",
@@ -5330,15 +5421,15 @@ dependencies = [

 [[package]]
 name = "hotpath-drain"
-version = "0.30.2"
+version = "0.30.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d0faae1a9e1fd8ce5debaaaf452dcbe833fc8a8a4f3779d2ec8828b116c3c16a"
+checksum = "738324f75aec847dea984f6806402f4f1bdfab437b4700fc8c3ebcad2dff6424"

 [[package]]
 name = "hotpath-macros"
-version = "0.30.2"
+version = "0.30.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a24f5179ff680132003306354efd0ac8cde12e87b1a788719c836e91af35cb58"
+checksum = "5d92fdfec803932e5764ea93c201cf7c6c2d8cff1871da44f64f5094c9c3ef8a"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -5959,7 +6050,7 @@ dependencies = [
 "lz4",
 "snap",
 "uuid",
- "zstd 0.13.3",
+ "zstd 0.13.4",
 ]

 [[package]]
@@ -6052,6 +6143,9 @@ name = "lazy_static"
 version = "1.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "20870f649af7073d53e38067b2a84312175d56ea15217e1b15bc83506ec50afb"
+dependencies = [
+ "spin 0.9.9",
+]

 [[package]]
 name = "lexical-core"
@@ -6261,9 +6355,9 @@ dependencies = [

 [[package]]
 name = "log"
-version = "0.4.34"
+version = "0.4.35"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f9f8bd3e56ce4dfc153cf470fffbfa98c7620958b312ca5c3a4b8d5181fd13c6"
+checksum = "5d15f86d12da44cd8d4ec3c73984537eb7f4cadcb9776a575a0a8a05ca11005c"

 [[package]]
 name = "lru"
@@ -6627,9 +6721,9 @@ dependencies = [

 [[package]]
 name = "ml-kem"
-version = "0.3.3"
+version = "0.3.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a9ba01ebcf90f6aff8e9b9f55bb69b19e6250e4ade5787d0040eaa27dd226238"
+checksum = "84ca78b66f46050d49fa21693ae8adab2081d8101b65d8067316b6b70b64a0d1"
 dependencies = [
 "hybrid-array",
 "kem",
@@ -6642,9 +6736,9 @@ dependencies = [

 [[package]]
 name = "module-lattice"
-version = "0.2.3"
+version = "0.2.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0c61b87c9683ab7cb1c6871d261ad5479b6b10ceb52c4352aaca3b5d35a8febe"
+checksum = "c4a1e412c8ee83a1c678f427806219c49c96b2511b9f07497514f0e43fb26155"
 dependencies = [
 "ctutils",
 "hybrid-array",
@@ -6940,6 +7034,22 @@ dependencies = [
 "rand_core 0.10.1",
 ]

+[[package]]
+name = "num-bigint-dig"
+version = "0.8.6"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "e661dda6640fad38e827a6d4a310ff4763082116fe217f279885c97f511bb0b7"
+dependencies = [
+ "lazy_static",
+ "libm",
+ "num-integer",
+ "num-iter",
+ "num-traits",
+ "rand 0.8.8",
+ "smallvec",
+ "zeroize",
+]
+
 [[package]]
 name = "num-complex"
 version = "0.4.6"
@@ -6964,6 +7074,16 @@ dependencies = [
 "num-traits",
 ]

+[[package]]
+name = "num-iter"
+version = "0.1.46"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c92800bd69a1eac91786bcfe9da64a897eb72911b8dc3095decbd07429e8048b"
+dependencies = [
+ "num-integer",
+ "num-traits",
+]
+
 [[package]]
 name = "num-rational"
 version = "0.4.2"
@@ -7035,18 +7155,19 @@ dependencies = [

 [[package]]
 name = "oauth2"
-version = "5.1.0"
-source = "git+https://github.com/ramosbugs/oauth2-rs?rev=89585f4a280afd7b285eaf1c9be0068874e03d22#89585f4a280afd7b285eaf1c9be0068874e03d22"
+version = "5.0.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "51e219e79014df21a225b1860a479e2dcd7cbd9130f4defd4bd0e191ea31d67d"
 dependencies = [
 "base64 0.22.1",
 "chrono",
 "getrandom 0.2.17",
 "http 1.5.0",
- "rand 0.10.3",
+ "rand 0.8.8",
 "serde",
 "serde_json",
 "serde_path_to_error",
- "sha2 0.11.1",
+ "sha2 0.10.9",
 "thiserror 1.0.69",
 "url",
 ]
@@ -7208,29 +7329,31 @@ checksum = "d6790f58c7ff633d8771f42965289203411a5e5c68388703c06e14f24770b41e"

 [[package]]
 name = "openidconnect"
-version = "4.0.1"
-source = "git+https://github.com/ramosbugs/openidconnect-rs?rev=2f77ee73bc4554e1cb6279adb386005429fb4e41#2f77ee73bc4554e1cb6279adb386005429fb4e41"
+version = "4.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "df0d098a59353679624d406825940139e2b2d554a52b64c9fb1b1bc42a11fe5c"
 dependencies = [
 "base64 0.22.1",
 "chrono",
 "dyn-clone",
 "ed25519-dalek 3.0.0",
- "hmac 0.13.0",
+ "getrandom 0.4.3",
+ "hmac 0.12.1",
 "http 1.5.0",
 "itertools 0.14.0",
 "log",
 "oauth2",
 "p256",
 "p384",
- "rand 0.10.3",
- "rsa",
+ "rand 0.8.8",
+ "rsa 0.9.10",
 "serde",
 "serde-value",
 "serde_json",
 "serde_path_to_error",
 "serde_plain",
 "serde_with",
- "sha2 0.11.1",
+ "sha2 0.10.9",
 "subtle",
 "thiserror 1.0.69",
 "url",
@@ -7558,12 +7681,12 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8eb260d566f2fbdabe356d26aaa5d13403aac29b5b4e44463b8f9bd7b2b12680"
 dependencies = [
 "ahash",
- "arrow-array",
- "arrow-buffer",
- "arrow-data",
- "arrow-ipc",
- "arrow-schema",
- "arrow-select",
+ "arrow-array 59.3.2",
+ "arrow-buffer 59.3.2",
+ "arrow-data 59.3.2",
+ "arrow-ipc 59.3.2",
+ "arrow-schema 59.3.2",
+ "arrow-select 59.3.2",
 "base64 0.23.1",
 "brotli 8.0.4",
 "bytes",
@@ -7582,7 +7705,33 @@ dependencies = [
 "snap",
 "tokio",
 "twox-hash",
- "zstd 0.13.3",
+ "zstd 0.13.4",
+]
+
+[[package]]
+name = "parquet"
+version = "60.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "806eaa0f1c0436dcb43bee8e6a82bf2efff90521a60172f8281cb9ef577b43a4"
+dependencies = [
+ "ahash",
+ "arrow-array 60.1.0",
+ "arrow-buffer 60.1.0",
+ "arrow-data 60.1.0",
+ "arrow-ipc 60.1.0",
+ "arrow-schema 60.1.0",
+ "arrow-select 60.1.0",
+ "base64 0.23.1",
+ "bytes",
+ "chrono",
+ "half",
+ "hashbrown 0.17.1",
+ "num-bigint 0.5.1",
+ "num-integer",
+ "num-traits",
+ "seq-macro",
+ "twox-hash",
+ "zstd 0.14.3",
 ]

 [[package]]
@@ -7834,6 +7983,17 @@ dependencies = [
 "futures-io",
 ]

+[[package]]
+name = "pkcs1"
+version = "0.7.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "c8ffb9f10fa047879315e6625af03c164b16962a5368d724ed16323b68ace47f"
+dependencies = [
+ "der 0.7.10",
+ "pkcs8 0.10.2",
+ "spki 0.7.3",
+]
+
 [[package]]
 name = "pkcs1"
 version = "0.8.0-rc.4"
@@ -8932,9 +9092,9 @@ dependencies = [

 [[package]]
 name = "reflink-copy"
-version = "0.1.30"
+version = "0.1.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d9dd7ab4af0363d5ccfd2838d782a28196cf32a5cc2e4fe3c5dc83f2be588b8b"
+checksum = "1ad64f61357bcf16105a0f353f9783733409c8bd889bc326ad857959605ea927"
 dependencies = [
 "cfg-if",
 "libc",
@@ -9096,6 +9256,26 @@ dependencies = [
 "xmlparser",
 ]

+[[package]]
+name = "rsa"
+version = "0.9.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b8573f03f5883dcaebdfcf4725caa1ecb9c15b2ef50c43a07b816e06799bb12d"
+dependencies = [
+ "const-oid 0.9.6",
+ "digest 0.10.7",
+ "num-bigint-dig",
+ "num-integer",
+ "num-traits",
+ "pkcs1 0.7.5",
+ "pkcs8 0.10.2",
+ "rand_core 0.6.4",
+ "signature 2.2.0",
+ "spki 0.7.3",
+ "subtle",
+ "zeroize",
+]
+
 [[package]]
 name = "rsa"
 version = "0.10.0-rc.18"
@@ -9106,7 +9286,7 @@ dependencies = [
 "crypto-bigint",
 "crypto-primes",
 "digest 0.11.3",
- "pkcs1",
+ "pkcs1 0.8.0-rc.4",
 "pkcs8 0.11.0",
 "rand_core 0.10.1",
 "sha2 0.11.1",
@@ -9211,13 +9391,13 @@ dependencies = [
 "p521",
 "pageant",
 "pbkdf2 0.13.0",
- "pkcs1",
+ "pkcs1 0.8.0-rc.4",
 "pkcs5 0.8.1",
 "pkcs8 0.11.0",
 "polyval",
 "rand 0.10.3",
 "rand_core 0.10.1",
- "rsa",
+ "rsa 0.10.0-rc.18",
 "russh-cryptovec",
 "russh-util",
 "salsa20 0.11.0",
@@ -9409,7 +9589,7 @@ dependencies = [
 "regex",
 "reqwest",
 "rmp-serde",
- "rsa",
+ "rsa 0.10.0-rc.18",
 "rust-embed",
 "rustfs-audit",
 "rustfs-common",
@@ -9612,7 +9792,7 @@ dependencies = [
 "jsonwebtoken",
 "pbkdf2 0.13.0",
 "rand 0.10.3",
- "rsa",
+ "rsa 0.10.0-rc.18",
 "serde",
 "serde_json",
 "sha2 0.11.1",
@@ -10150,18 +10330,18 @@ dependencies = [

 [[package]]
 name = "rustfs-mimalloc"
-version = "0.7.2"
+version = "0.7.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "8301a578772fc9823d68e083e42c3b414613d588a0eac594c1ca6bc0bbe0653f"
+checksum = "216ecfc9341b5d6e2506d47ce57c9232ae1b8020bb10264ea05286fb234286e7"
 dependencies = [
 "rustfs-mimalloc-sys",
 ]

 [[package]]
 name = "rustfs-mimalloc-sys"
-version = "0.7.2"
+version = "0.7.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6eb4082853043cb29556b4b776b05f28c3146706f43c53869dbf1ef262e00e91"
+checksum = "fb16ad68f96883407a75cdb88252b43dc2ea592428f0fc54ff7885756de4b814"
 dependencies = [
 "cc",
 ]
@@ -10878,6 +11058,7 @@ dependencies = [
 name = "rustfs-utils"
 version = "1.0.2"
 dependencies = [
+ "aws-lc-rs",
 "base64-simd",
 "blake2",
 "brotli 9.1.1",
@@ -12006,7 +12187,7 @@ dependencies = [
 "p384",
 "p521",
 "rand_core 0.10.1",
- "rsa",
+ "rsa 0.10.0-rc.18",
 "sec1",
 "serde",
 "sha1 0.11.0",
@@ -12356,18 +12537,18 @@ dependencies = [

 [[package]]
 name = "test-case"
-version = "3.4.0"
+version = "3.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "124953e7f67cb0b2fcfb87e899e4ae5a64fb68e2e160767933cc67b646fb0042"
+checksum = "ae6e80608d2037173d7d59a09089d249e43352c182f6dde6b00d138bcb7cfdbe"
 dependencies = [
 "test-case-macros",
 ]

 [[package]]
 name = "test-case-core"
-version = "3.4.0"
+version = "3.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "bd097615b407247e102046ba34426bb594923ab3554cf0f3cdfed050f1a5a3e8"
+checksum = "29b8e1f65819001927b8cdcbb5560f2428212f2e5ab5684e3c6815a4d80e72db"
 dependencies = [
 "cfg-if",
 "proc-macro2",
@@ -12377,9 +12558,9 @@ dependencies = [

 [[package]]
 name = "test-case-macros"
-version = "3.4.0"
+version = "3.4.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f4ce35ece947bccac166e1ded639133827a953f69a48ba134d17f91108333082"
+checksum = "802d6165a14df1598b7290a8c546df402940d3cfbe82d15c0fd35de176771693"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -13114,9 +13295,9 @@ dependencies = [

 [[package]]
 name = "unicode-properties"
-version = "0.1.4"
+version = "0.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7df058c713841ad818f1dc5d3fd88063241cc61f49f5fbea4b951e8cf5a8d71d"
+checksum = "993ad287eb20e8ba755312801cfe1413e0f25d604075863acaddca10b04c1833"

 [[package]]
 name = "unicode-segmentation"
@@ -14018,9 +14199,9 @@ dependencies = [

 [[package]]
 name = "zip"
-version = "9.1.3"
+version = "9.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1156b39d142a34b999c2310a50455e0cb3760bf7efad0122f796d813bee26ccb"
+checksum = "e408c6cc4699dd017d53dc080cd51c2cea2b5354af57a1479bd265ed42911487"
 dependencies = [
 "aes 0.9.3",
 "bzip2",
@@ -14031,6 +14212,7 @@ dependencies = [
 "getrandom 0.4.3",
 "hmac 0.13.0",
 "indexmap 2.14.2",
+ "libc",
 "lzma-rust2",
 "memchr",
 "nix 0.31.3",
@@ -14071,15 +14253,15 @@ dependencies = [

 [[package]]
 name = "zstd"
-version = "0.13.3"
+version = "0.13.4"
 dependencies = [
 "apache-avro",
- "arrow-array",
- "arrow-ipc",
- "arrow-schema",
+ "arrow-array 60.1.0",
+ "arrow-ipc 60.1.0",
+ "arrow-schema 60.1.0",
 "bytes",
 "kafka-protocol",
- "parquet",
+ "parquet 60.1.0",
 "semver",
 "toml",
 "zstd 0.14.3",
@@ -147,7 +147,7 @@ rustfs-zip = { path = "./crates/zip", version = "1.0.2" }
 async-channel = "2.5.0"
 async_zip = { default-features = false, version = "0.0.19" }
 mysql_async = { default-features = false, version = "0.37.1" }
-async-compression = { version = "0.4.51" }
+async-compression = { version = "0.4.52" }
 async-trait = "0.1.92"
 async-nats = { version = "0.51.0", default-features = false }
 axum = "0.8.9"
@@ -212,12 +212,13 @@ serde_urlencoded = "0.7.1"
 # releases.
 aes-gcm = { version = "0.11.1" }
 argon2 = { version = "0.6.0" }
+aws-lc-rs = { version = "1.18.0", default-features = false }
 blake2 = "0.11.0"
 chacha20poly1305 = { version = "0.11.0" }
 crc-fast = "1.10.0"
 hmac = { version = "0.13.0" }
 jsonwebtoken = { version = "11.1.0" }
-openidconnect = { default-features = false, version = "4.0" }
+openidconnect = { default-features = false, version = "4.1.0" }
 pbkdf2 = "0.13.0"
 p256 = { version = "0.14.0", features = ["ecdsa", "pkcs8"] }
 rsa = { version = "=0.10.0-rc.18" }
@@ -253,9 +254,9 @@ atoi = "3.1.1"
 atomic_enum = "0.3.0"
 aws-config = { version = "1.13.0" }
 aws-credential-types = { version = "1.3.1" }
-aws-sdk-kms = { default-features = false, version = "1.128.0" }
-aws-sdk-s3 = { default-features = false, version = "1.156.0" }
-aws-sdk-sts = { default-features = false, version = "1.123.0" }
+aws-sdk-kms = { default-features = false, version = "1.129.0" }
+aws-sdk-s3 = { default-features = false, version = "1.157.0" }
+aws-sdk-sts = { default-features = false, version = "1.124.0" }
 aws-smithy-async = { version = "1.3.1" }
 aws-smithy-http-client = { default-features = false, version = "1.5.2" }
 aws-smithy-runtime-api = { version = "1.19.2" }
@@ -337,7 +338,7 @@ strum = { version = "0.28.0" }
 sysinfo = "0.40.1"
 temp-env = "0.3.6"
 tempfile = "3.27.0"
-test-case = "3.4.0"
+test-case = "3.4.1"
 thiserror = "2.0.21"
 tracing = { version = "0.1.44" }
 tracing-appender = "0.2.5"
@@ -356,7 +357,7 @@ winapi-util = "0.1.11"
 windows = { version = "0.62.2" }
 windows-sys = "0.61.2"
 xxhash-rust = { version = "0.8.19" }
-zip = "9.1.3"
+zip = "9.2.2"
 zstd = "0.14.3"

 # Observability and Metrics
@@ -384,9 +385,9 @@ russh-sftp = "3.0.1"
 dav-server = "0.11.0"

 # Performance Analysis and Memory Profiling
-rustfs-mimalloc = { version = "0.7.2" }
+rustfs-mimalloc = { version = "0.7.3" }
 # Preserve Unicode focus filters until rustfs/backlog#2302 is resolved.
-hotpath = { version = "0.30.2", default-features = false }
+hotpath = { version = "0.30.3", default-features = false }
 # Snapshot testing for output format regression detection
 insta = { version = "1.49" }

@@ -427,7 +428,3 @@ strip = "none"
 # Pin upstream fixes until compatible crates.io releases are available.
 [patch.crates-io]
 zstd = { path = "crates/zstd-compat" }
-# Preserve RustFS OIDC flows while migrating the old elliptic-curve dependency chain.
-# Upstream: openidconnect-rs#244 and oauth2-rs#352.
-openidconnect = { git = "https://github.com/ramosbugs/openidconnect-rs", rev = "2f77ee73bc4554e1cb6279adb386005429fb4e41" }
-oauth2 = { git = "https://github.com/ramosbugs/oauth2-rs", rev = "89585f4a280afd7b285eaf1c9be0068874e03d22" }
@@ -26,7 +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) |
+| **distributed 4×4** | [`src/distributed/`](src/distributed) | 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 |

@@ -194,9 +194,9 @@ the same profile for membership and execution with one nightly worker.
 | `s3s-e2e` black-box | `e2e-tests` + `e2e-tests-rio-v2` jobs | **Active** (external conformance tool) |
 | ILM / lifecycle (ignored) | `test-ilm-integration-serial` lane, `-j1` | **Active** (backlog#1148 ilm-1) |
 | 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** |
+| Direct and mixed-version rolling upgrades from pinned previous release | `e2e-upgrade.yml`, release tags + weekly schedule + manual dispatch | **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) |
+| Distributed 4-node 4-disk (`e2e-distributed` profile) | `.github/workflows/e2e-distributed.yml` | **Active** (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) |
@@ -12,11 +12,249 @@
 // 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 super::harness::{DistCluster, DistLayout, TestResult, payload_for, s3_probe_client, sha256_hex, unique_bucket};
 use crate::common::init_logging;
+use aws_sdk_s3::{
+    Client,
+    error::{ProvideErrorMetadata, SdkError},
+    primitives::ByteStream,
+};
+use std::any::TypeId;
 use std::collections::BTreeSet;
+use std::fmt::Debug;
+use std::future::Future;
 use std::sync::Arc;
+use std::time::Duration;
 use tokio::sync::Barrier;
+use tokio::time::Instant;
+
+const LOCK_ADMISSION_BUDGET: Duration = Duration::from_secs(45);
+
+async fn after_lock_contention<T, E, F, Fut>(operation: &str, mut call: F) -> TestResult<Result<T, SdkError<E>>>
+where
+    E: Debug + ProvideErrorMetadata + 'static,
+    F: FnMut() -> Fut,
+    Fut: Future<Output = Result<T, SdkError<E>>>,
+{
+    let started = Instant::now();
+    let mut attempts = 0;
+    // The workload uses unique unversioned keys. Repeat only explicitly
+    // advertised lock backpressure; every step still needs its real result.
+    let outcome = tokio::time::timeout(LOCK_ADMISSION_BUDGET, async {
+        loop {
+            attempts += 1;
+            let result = call().await;
+            let retry = result.as_ref().err().is_some_and(|error| {
+                let Some(response) = error.raw_response() else {
+                    return false;
+                };
+                let Some(service_error) = error.as_service_error() else {
+                    return false;
+                };
+                // HEAD has no XML error body, so its SDK service error can
+                // lack a code even when the server supplies retry headers.
+                let namespace_code = service_error.code() == Some("ServiceUnavailable")
+                    || (service_error.code().is_none()
+                        && TypeId::of::<E>() == TypeId::of::<aws_sdk_s3::operation::head_object::HeadObjectError>());
+                response.status().as_u16() == 503
+                    && namespace_code
+                    && response.headers().get("x-rustfs-error-cause") == Some("namespace_lock_timeout")
+                    && response.headers().get("retry-after") == Some("1")
+            });
+            if !retry {
+                return result;
+            }
+            tokio::time::sleep(Duration::from_secs(1)).await;
+        }
+    })
+    .await;
+    if attempts > 1 || !matches!(&outcome, Ok(Ok(_))) {
+        eprintln!(
+            "mixed workload request: operation={operation}, attempts={attempts}, elapsed={:?}",
+            started.elapsed()
+        );
+    }
+    outcome.map_err(|_| {
+        format!("{operation} exceeded its {LOCK_ADMISSION_BUDGET:?} request budget after {attempts} attempts").into()
+    })
+}
+
+async fn assert_workload_bytes(client: &Client, bucket: &str, key: &str, expected: &[u8]) -> TestResult {
+    let output = after_lock_contention(&format!("GET {key}"), || client.get_object().bucket(bucket).key(key).send()).await??;
+    let actual = output.body.collect().await?.into_bytes();
+    if actual.as_ref() != expected {
+        return Err(format!(
+            "object {key} bytes mismatch: expected {} bytes sha256={} got {} bytes sha256={}",
+            expected.len(),
+            sha256_hex(expected),
+            actual.len(),
+            sha256_hex(&actual)
+        )
+        .into());
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use aws_smithy_http_client::test_util::infallible_client_fn;
+    use aws_smithy_types::body::SdkBody;
+    use std::sync::atomic::{AtomicUsize, Ordering};
+
+    type Reply = (u16, &'static str, Option<&'static str>, Option<&'static str>, &'static str);
+
+    fn client(respond: impl Fn(usize) -> Reply + Send + Sync + 'static) -> (Client, Arc<AtomicUsize>) {
+        let requests = Arc::new(AtomicUsize::new(0));
+        let seen = Arc::clone(&requests);
+        let transport = infallible_client_fn(move |_| {
+            let (status, code, cause, retry_after, body) = respond(seen.fetch_add(1, Ordering::SeqCst));
+            let body = if code.is_empty() {
+                body.to_string()
+            } else {
+                format!("<Error><Code>{code}</Code></Error>")
+            };
+            let mut response = http::Response::builder()
+                .status(status)
+                .header("content-type", "application/xml")
+                .header("content-length", body.len());
+            if let Some(cause) = cause {
+                response = response.header("x-rustfs-error-cause", cause);
+            }
+            if let Some(retry_after) = retry_after {
+                response = response.header("retry-after", retry_after);
+            }
+            response.body(SdkBody::from(body)).expect("scripted S3 response")
+        });
+        let config = crate::common::build_test_s3_config(
+            "http://contention.example.invalid",
+            "synthetic-access",
+            "synthetic-secret",
+            None,
+            "contention-test",
+        );
+        let client = Client::from_conf(config.to_builder().http_client(transport).build());
+        (s3_probe_client(&client), requests)
+    }
+
+    fn contention() -> Reply {
+        (503, "ServiceUnavailable", Some("namespace_lock_timeout"), Some("1"), "")
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn advertised_contention_recovers_after_standard_sdk_attempts_are_exhausted() -> TestResult {
+        let (client, requests) = client(|attempt| {
+            if attempt < 3 {
+                contention()
+            } else {
+                (200, "", None, None, "")
+            }
+        });
+        let standard = Client::from_conf(
+            client
+                .config()
+                .to_builder()
+                .retry_config(aws_sdk_s3::config::retry::RetryConfig::standard().with_max_attempts(3))
+                .build(),
+        );
+        standard
+            .put_object()
+            .bucket("bucket")
+            .key("key")
+            .body(ByteStream::from_static(b"body"))
+            .send()
+            .await
+            .expect_err("three advertised lock responses exhaust the standard SDK policy");
+        assert_eq!(requests.load(Ordering::SeqCst), 3);
+        requests.store(0, Ordering::SeqCst);
+        let started = Instant::now();
+        after_lock_contention("PUT recovery", || {
+            client
+                .put_object()
+                .bucket("bucket")
+                .key("key")
+                .body(ByteStream::from_static(b"body"))
+                .send()
+        })
+        .await??;
+        assert_eq!(requests.load(Ordering::SeqCst), 4);
+        assert!(started.elapsed() >= Duration::from_secs(3), "honor each Retry-After response");
+        Ok(())
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn bodyless_head_contention_uses_explicit_retry_headers() -> TestResult {
+        let (client, requests) = client(|attempt| {
+            if attempt == 0 {
+                (503, "", Some("namespace_lock_timeout"), Some("1"), "")
+            } else {
+                (200, "", None, None, "")
+            }
+        });
+        after_lock_contention("HEAD recovery", || client.head_object().bucket("bucket").key("key").send()).await??;
+        assert_eq!(requests.load(Ordering::SeqCst), 2);
+        Ok(())
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn unrelated_or_malformed_responses_are_not_retried() -> TestResult {
+        for reply in [
+            (500, "InternalError", Some("namespace_lock_timeout"), Some("1"), ""),
+            (403, "AccessDenied", None, None, ""),
+            (404, "NoSuchKey", None, None, ""),
+            (503, "ServiceUnavailable", Some("storage_write_quorum"), None, ""),
+            (503, "ServiceUnavailable", Some("namespace_lock_unavailable"), Some("1"), ""),
+            (503, "InternalError", Some("namespace_lock_timeout"), Some("1"), ""),
+            (503, "", Some("namespace_lock_timeout"), Some("1"), ""),
+            (503, "ServiceUnavailable", None, Some("1"), ""),
+            (503, "ServiceUnavailable", Some("namespace_lock_timeout"), None, ""),
+            (503, "ServiceUnavailable", Some("namespace_lock_timeout"), Some("2"), ""),
+        ] {
+            let (client, requests) = client(move |_| reply);
+            let result = after_lock_contention("PUT fatal response", || {
+                client
+                    .put_object()
+                    .bucket("bucket")
+                    .key("key")
+                    .body(ByteStream::from_static(b"body"))
+                    .send()
+            })
+            .await?;
+            assert!(result.is_err(), "fatal response cannot become success: {reply:?}");
+            assert_eq!(requests.load(Ordering::SeqCst), 1, "fatal response must remain visible: {reply:?}");
+        }
+        Ok(())
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn persistent_lock_contention_exhausts_the_request_budget() {
+        let (client, requests) = client(|_| contention());
+        let started = Instant::now();
+        let result = after_lock_contention("PUT unavailable", || {
+            client
+                .put_object()
+                .bucket("bucket")
+                .key("key")
+                .body(ByteStream::from_static(b"body"))
+                .send()
+        })
+        .await;
+        assert!(result.is_err(), "persistent contention cannot acknowledge an operation");
+        assert!(started.elapsed() >= LOCK_ADMISSION_BUDGET);
+        assert!(requests.load(Ordering::SeqCst) > 1);
+        assert!(requests.load(Ordering::SeqCst) <= LOCK_ADMISSION_BUDGET.as_secs() as usize + 1);
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn acknowledged_wrong_bytes_fail_without_retry() {
+        let (client, requests) = client(|_| (200, "", None, None, "wrong bytes"));
+        let error = assert_workload_bytes(&client, "bucket", "key", b"expected bytes")
+            .await
+            .expect_err("a successful but incorrect read is a consistency failure");
+        assert!(error.to_string().contains("bytes mismatch"));
+        assert_eq!(requests.load(Ordering::SeqCst), 1);
+    }
+}

 #[tokio::test]
 async fn four_node_high_concurrency_mixed_workload_is_consistent_on_every_node() -> TestResult {
@@ -24,7 +262,9 @@ async fn four_node_high_concurrency_mixed_workload_is_consistent_on_every_node()
    let dist = DistCluster::start(DistLayout::FourByFour).await?;
    let bucket = unique_bucket("conc");
    dist.create_bucket(&bucket).await?;
-    let clients = Arc::new(dist.clients()?);
+    // This consistency probe owns its admission policy; unrelated responses
+    // must not be hidden by the SDK's general-purpose retries.
+    let clients = Arc::new(dist.clients()?.iter().map(s3_probe_client).collect::<Vec<_>>());
    let worker_count = 24;
    let rounds = 4;
    let barrier = Arc::new(Barrier::new(worker_count));
@@ -44,31 +284,46 @@ async fn four_node_high_concurrency_mixed_workload_is_consistent_on_every_node()
                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 context = format!("worker={idx} round={round}");
+                after_lock_contention(&format!("{context} PUT source"), || {
+                    writer
+                        .put_object()
+                        .bucket(&bucket)
+                        .key(&key)
+                        .body(ByteStream::from(body.clone()))
+                        .send()
+                })
+                .await??;

-                let head = reader.head_object().bucket(&bucket).key(&key).send().await?;
+                let head = after_lock_contention(&format!("{context} HEAD source"), || {
+                    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?;
+                assert_workload_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?;
+                after_lock_contention(&format!("{context} COPY retained"), || {
+                    copier
+                        .copy_object()
+                        .bucket(&bucket)
+                        .key(&copy_key)
+                        .copy_source(format!("{bucket}/{key}"))
+                        .send()
+                })
+                .await??;
+                assert_workload_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");
+                after_lock_contention(&format!("{context} DELETE source"), || {
+                    writer.delete_object().bucket(&bucket).key(&key).send()
+                })
+                .await??;
+                let missing = after_lock_contention(&format!("{context} HEAD deleted source"), || {
+                    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());
                }
@@ -82,14 +337,22 @@ async fn four_node_high_concurrency_mixed_workload_is_consistent_on_every_node()
    for handle in handles {
        inventory.extend(handle.await??);
    }
+    assert_eq!(
+        inventory.len(),
+        worker_count * rounds,
+        "every worker round must retain its acknowledged copy"
+    );

    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 = after_lock_contention(&format!("node={node_idx} LIST retained"), || {
+            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)
+            assert_workload_bytes(client, &bucket, key, body)
                .await
                .map_err(|error| format!("node {node_idx} failed to read {key}: {error}"))?;
        }
@@ -31,8 +31,8 @@ use aws_sdk_s3::error::ProvideErrorMetadata;
 use aws_sdk_s3::operation::list_object_versions::ListObjectVersionsOutput;
 use aws_sdk_s3::primitives::ByteStream;
 use aws_sdk_s3::types::{
-    BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, DeleteMarkerEntry, ObjectVersion, ServerSideEncryption,
-    VersioningConfiguration,
+    BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, Delete, DeleteMarkerEntry, ObjectIdentifier, ObjectVersion,
+    ServerSideEncryption, VersioningConfiguration,
 };
 use base64_simd::STANDARD as BASE64_STANDARD;
 use bytes::Bytes;
@@ -3729,6 +3729,8 @@ async fn test_bucket_replication_converges_delete_marker_and_version_purge() ->
    init_logging();

    let mut source_env = RustFSTestEnvironment::new().await?;
+    let source_log = format!("{}/source.log", source_env.temp_dir);
+    source_env.capture_log_path = Some(source_log.clone());
    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?;
@@ -3791,19 +3793,66 @@ async fn test_bucket_replication_converges_delete_marker_and_version_purge() ->
        "target did not preserve the source delete-marker version ID"
    );

-    source_client
-        .delete_object()
+    let wait_for_source_writeback = || async {
+        let metrics_url = format!("{}/{source_bucket}?replication-metrics", source_env.url);
+        timeout(Duration::from_secs(30), async {
+            loop {
+                let response = signed_request(
+                    http::Method::GET,
+                    &metrics_url,
+                    &source_env.access_key,
+                    &source_env.secret_key,
+                    None,
+                    None,
+                )
+                .await?;
+                let metrics: serde_json::Value = response.error_for_status()?.json().await?;
+                let queued = metrics["queued"]["curr"]["count"]
+                    .as_f64()
+                    .ok_or("missing replication queue count")?;
+                if metrics["provider_available"] == true && queued == 0.0 {
+                    return Ok::<(), Box<dyn Error + Send + Sync>>(());
+                }
+                sleep(Duration::from_millis(50)).await;
+            }
+        })
+        .await?
+    };
+    wait_for_source_writeback().await?;
+
+    let purged = source_client
+        .delete_objects()
        .bucket(source_bucket)
-        .key(object_key)
-        .version_id(&purged_version_id)
+        .delete(
+            Delete::builder()
+                .objects(
+                    ObjectIdentifier::builder()
+                        .key(object_key)
+                        .version_id(&purged_version_id)
+                        .build()?,
+                )
+                .objects(
+                    ObjectIdentifier::builder()
+                        .key(object_key)
+                        .version_id(delete_marker_version_id)
+                        .build()?,
+                )
+                .build()?,
+        )
        .send()
        .await?;
+    assert!(purged.errors().is_empty(), "batch version purge failed: {:?}", purged.errors());
+    assert_eq!(purged.deleted().len(), 2, "both the data version and marker must be purged");
    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.iter().all(|entry| entry.version_id != purged_version_id),
        "target retained the explicitly purged object version"
    );
+    assert!(
+        target_state.iter().all(|entry| entry.version_id != delete_marker_version_id),
+        "target retained the explicitly purged delete marker"
+    );
    assert!(
        target_state
            .iter()
@@ -3819,6 +3868,19 @@ async fn test_bucket_replication_converges_delete_marker_and_version_purge() ->
        .await?;
    assert_eq!(retained.body.collect().await?.into_bytes().as_ref(), b"versioned replication payload v2");

+    // Wait for the source workers to finish their final writeback, which runs
+    // after the remote purge has already become visible (rustfs#9121).
+    wait_for_source_writeback().await?;
+    let logs = fs::read_to_string(source_log).await?;
+    let writeback_failures: Vec<_> = logs
+        .lines()
+        .filter(|line| line.contains("apply_replication_delete_state"))
+        .collect();
+    assert!(
+        writeback_failures.is_empty(),
+        "a completed batch purge must not report a source writeback failure: {writeback_failures:?}"
+    );
+
    Ok(())
 }

@@ -31,6 +31,7 @@ workspace = true

 [features]
 default = []
+fips = ["rustfs-s3-client/fips", "rustfs-utils/fips"]
 diagnostic-read-version-caller = ["rustfs-io-metrics/diagnostic-read-version-caller"]
 # Sampled wall/CPU diagnostics only; does not change the read backend.
 diagnostic-local-batch = ["rustix/time"]
@@ -3421,6 +3421,7 @@ mod tests {
        allow_first_read: Notify,
        read_count: AtomicUsize,
        delete_storage_calls: AtomicUsize,
+        delete_error: StdMutex<EcstoreError>,
        reported_size: StdMutex<Option<i64>>,
        stream_read_bytes: Arc<AtomicUsize>,
        write_count: AtomicUsize,
@@ -3672,7 +3673,12 @@ mod tests {
            _opts: Self::ObjectOptions,
        ) -> Result<Self::ObjectInfo, Self::Error> {
            self.shared.delete_storage_calls.fetch_add(1, Ordering::SeqCst);
-            Err(EcstoreError::NotImplemented)
+            Err(self
+                .shared
+                .delete_error
+                .lock()
+                .expect("delete error lock should not be poisoned")
+                .clone())
        }

        async fn delete_objects(
@@ -4171,6 +4177,7 @@ mod tests {
            allow_first_read: Notify::new(),
            read_count: AtomicUsize::new(0),
            delete_storage_calls: AtomicUsize::new(0),
+            delete_error: StdMutex::new(EcstoreError::NotImplemented),
            reported_size: StdMutex::new(None),
            stream_read_bytes: Arc::new(AtomicUsize::new(0)),
            write_count: AtomicUsize::new(0),
@@ -4221,6 +4228,103 @@ mod tests {
        }
    }

+    #[tokio::test]
+    async fn delete_worker_already_purged_source_completes_without_retry() {
+        let arn = "arn:test";
+        for delete_marker_version in [false, true] {
+            for version_id in [Uuid::new_v4(), Uuid::nil()] {
+                for error in [
+                    EcstoreError::VersionNotFound("source".to_string(), "purged".to_string(), version_id.to_string()),
+                    EcstoreError::ObjectNotFound("source".to_string(), "purged".to_string()),
+                    EcstoreError::FileVersionNotFound,
+                    EcstoreError::FileNotFound,
+                ] {
+                    let shared = empty_resync_shared_state();
+                    *shared.delete_error.lock().expect("delete error lock should not be poisoned") = error;
+                    let storage = Arc::new(LoadResyncNodeStore::new("purged-source", shared.clone()));
+                    let delete = DeletedObjectReplicationInfo {
+                        bucket: "source".to_string(),
+                        delete_object: ReplicationDeletedObject {
+                            object_name: "purged".to_string(),
+                            version_id: (!delete_marker_version).then_some(version_id),
+                            delete_marker_version_id: delete_marker_version.then_some(version_id),
+                            replication_state: Some(ReplicationState {
+                                replicate_decision_str: test_replicate_decision(&[arn]).to_string(),
+                                purge_targets: HashMap::from([(arn.to_string(), VersionPurgeStatusType::Complete)]),
+                                ..Default::default()
+                            }),
+                            ..Default::default()
+                        },
+                        op_type: ReplicationType::Delete,
+                        ..Default::default()
+                    };
+
+                    assert!(
+                        replicate_delete_with_outcome(delete, storage).await,
+                        "a confirmed target purge must complete when the source is already absent (marker={delete_marker_version}, version={version_id})"
+                    );
+                    assert_eq!(
+                        shared.delete_storage_calls.load(Ordering::SeqCst),
+                        2 + usize::from(!delete_marker_version && !version_id.is_nil()),
+                        "the worker must lock and attempt writeback, reading a non-null data version first"
+                    );
+                }
+            }
+        }
+    }
+
+    #[tokio::test]
+    async fn delete_worker_missing_source_does_not_hide_unfinished_replication() {
+        let arn = "arn:test";
+        for (version_purge, target_status, error) in [
+            (
+                true,
+                VersionPurgeStatusType::Pending,
+                EcstoreError::ObjectNotFound("source".to_string(), "purged".to_string()),
+            ),
+            (true, VersionPurgeStatusType::Complete, EcstoreError::ErasureWriteQuorum),
+            (
+                false,
+                VersionPurgeStatusType::Empty,
+                EcstoreError::ObjectNotFound("source".to_string(), "purged".to_string()),
+            ),
+        ] {
+            let shared = empty_resync_shared_state();
+            *shared.delete_error.lock().expect("delete error lock should not be poisoned") = error;
+            let storage = Arc::new(LoadResyncNodeStore::new("unfinished-delete", shared));
+            let delete = DeletedObjectReplicationInfo {
+                bucket: "source".to_string(),
+                delete_object: ReplicationDeletedObject {
+                    object_name: "purged".to_string(),
+                    version_id: version_purge.then(Uuid::new_v4),
+                    delete_marker: !version_purge,
+                    replication_state: Some(ReplicationState {
+                        replicate_decision_str: test_replicate_decision(&[arn]).to_string(),
+                        targets: if version_purge {
+                            HashMap::new()
+                        } else {
+                            HashMap::from([(arn.to_string(), ReplicationStatusType::Completed)])
+                        },
+                        purge_targets: if version_purge {
+                            HashMap::from([(arn.to_string(), target_status)])
+                        } else {
+                            HashMap::new()
+                        },
+                        ..Default::default()
+                    }),
+                    ..Default::default()
+                },
+                op_type: ReplicationType::Delete,
+                ..Default::default()
+            };
+
+            assert!(
+                !replicate_delete_with_outcome(delete, storage).await,
+                "unfinished target work, source I/O failures, and marker creation must remain retryable"
+            );
+        }
+    }
+
    #[tokio::test]
    async fn delete_worker_preserves_persisted_admission_without_live_config() {
        let arn = "arn:test";
@@ -5304,6 +5408,7 @@ mod tests {
                allow_first_read: Notify::new(),
                read_count: AtomicUsize::new(0),
                delete_storage_calls: AtomicUsize::new(0),
+                delete_error: StdMutex::new(EcstoreError::NotImplemented),
                reported_size: StdMutex::new(None),
                stream_read_bytes: Arc::new(AtomicUsize::new(0)),
                write_count: AtomicUsize::new(0),
@@ -2809,17 +2809,22 @@ pub(crate) async fn replicate_delete_with_outcome<S: ReplicationStorage>(
            true
        }
        Err(e) => {
-            error!(
-                event = EVENT_RESYNC_TARGET_OPERATION_FAILED,
-                component = LOG_COMPONENT_ECSTORE,
-                subsystem = LOG_SUBSYSTEM_REPLICATION_RESYNC,
-                bucket = %bucket,
-                arn = %dobj.target_arn,
-                object = %dobj.delete_object.object_name,
-                operation = "apply_replication_delete_state",
-                error = %e,
-                "Replication target operation failed"
-            );
+            // A batch purge or another worker may have removed this version
+            // already. Remote failures still control the final outcome below.
+            let source_already_purged = is_version_purge && (is_err_object_not_found(&e) || is_err_version_not_found(&e));
+            if !source_already_purged {
+                error!(
+                    event = EVENT_RESYNC_TARGET_OPERATION_FAILED,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_REPLICATION_RESYNC,
+                    bucket = %bucket,
+                    arn = %dobj.target_arn,
+                    object = %dobj.delete_object.object_name,
+                    operation = "apply_replication_delete_state",
+                    error = %e,
+                    "Replication target operation failed"
+                );
+            }
            send_event(EventArgs {
                event_name,
                bucket_name: bucket.clone(),
@@ -2832,7 +2837,7 @@ pub(crate) async fn replicate_delete_with_outcome<S: ReplicationStorage>(
                },
                ..Default::default()
            });
-            false
+            source_already_purged
        }
    };

@@ -32,7 +32,6 @@ use crate::storage_api_contracts::internode::{
    PUT_FILE_AUTH_TRAILER_MAGIC, PUT_FILE_CAPABILITY_VERSION,
 };

-use hmac::{Hmac, KeyInit, Mac};
 use http::uri::Authority;
 use http::{HeaderMap, HeaderValue, Method, Uri};
 #[cfg(test)]
@@ -47,6 +46,7 @@ use rustfs_io_metrics::internode_metrics::{
    global_internode_metrics,
 };
 use rustfs_object_data_cache::{MemoryBasis, resolve_effective_memory};
+use rustfs_utils::crypto::{HmacSha256, KeyInit, Mac};
 use rustfs_utils::get_env_bool;
 use sha2::Digest as _;
 use sha2::Sha256;
@@ -58,8 +58,6 @@ use time::OffsetDateTime;
 use tracing::{error, info, warn};
 use uuid::Uuid;

-type HmacSha256 = Hmac<Sha256>;
-
 const SIGNATURE_HEADER: &str = "x-rustfs-signature";
 pub const RPC_BACKGROUND_ADMISSION_HEADER: &str = "x-rustfs-rpc-background-admission";
 const RPC_BACKGROUND_ADMISSION_V1: &str = "v1";
@@ -1637,6 +1637,7 @@ impl RemoteDisk {
        let initial_probe_ok = Self::perform_connectivity_check(&addr).await.is_ok();
        if initial_probe_ok {
            health.record_operation_success(&endpoint, "connectivity_probe_success");
+            rustfs_io_metrics::internode_metrics::record_peer_reachable_if_offline(&addr);
        } else if health.mark_failure(&endpoint, "connectivity_probe_failed") {
            warn!(
                event = EVENT_REMOTE_DISK_HEALTH,
@@ -1708,6 +1709,7 @@ impl RemoteDisk {
                    // Perform basic connectivity check
                    if Self::perform_connectivity_check(&addr).await.is_ok() {
                        health.record_operation_success(&endpoint, "connectivity_probe_success");
+                        rustfs_io_metrics::internode_metrics::record_peer_reachable_if_offline(&addr);
                    } else if health.mark_failure(&endpoint, "connectivity_probe_failed") {
                        warn!(
                            event = EVENT_REMOTE_DISK_HEALTH,
@@ -1778,6 +1780,7 @@ impl RemoteDisk {
                }
                _ = interval.tick() => {
                    if Self::perform_recovery_probe(&addr, &endpoint).await.is_ok() {
+                        rustfs_io_metrics::internode_metrics::record_peer_reachable(&addr);
                        let became_online = health.mark_recovery_success(&endpoint, "disk_info_probe_success");
                        debug!(
                            event = EVENT_REMOTE_DISK_HEALTH,
@@ -2009,6 +2012,7 @@ impl RemoteDisk {
            let operation_result = operation().await;
            if operation_result.is_ok() {
                self.health.record_operation_success(&self.endpoint, "operation_success");
+                rustfs_io_metrics::internode_metrics::record_peer_reachable_if_offline(&self.addr);
            }
            self.handle_network_like_error(op, timeout_duration, &operation_result, failure_health_action)
                .await;
@@ -2023,6 +2027,7 @@ impl RemoteDisk {
                // Log success; the waiting guard balances every exit path.
                if operation_result.is_ok() {
                    self.health.record_operation_success(&self.endpoint, "operation_success");
+                    rustfs_io_metrics::internode_metrics::record_peer_reachable_if_offline(&self.addr);
                }
                self.handle_network_like_error(op, timeout_duration, &operation_result, failure_health_action)
                    .await;
@@ -445,6 +445,10 @@ impl RemoteClient {
                }
            }
        });
+        // A successful lock RPC over a cached or fresh channel proves the peer is
+        // reachable; reconcile the internode peer-health gauge so it does not stay
+        // stale when no fresh dial fires (issue #9119).
+        rustfs_io_metrics::internode_metrics::record_peer_reachable_if_offline(&self.addr);
    }

    fn admit_late_release(&self) -> bool {
@@ -672,6 +676,7 @@ impl RemoteClient {
                    with_lock_peer_health(&addr, |health| {
                        health.last_success = Some(Instant::now());
                    });
+                    rustfs_io_metrics::internode_metrics::record_peer_reachable_if_offline(&addr);
                    rustfs_io_metrics::lock_metrics::record_remote_lock_rpc_late_completion(op, "success");
                    if let Some(late) = late {
                        late(response).await;
@@ -18,6 +18,7 @@ use crate::storage_api_contracts::namespace::NamespaceLocking as _;
 use codec::{fixture_inline_pool_meta_replicas as replicas, fixture_split_inline_pool_meta_replicas as split};
 use std::path::PathBuf;
 use tokio::fs;
+use uuid::Uuid;

 struct PhysicalFixture {
    _directories: Vec<tempfile::TempDir>,
@@ -628,3 +629,559 @@ async fn pool_meta_native_physical_three_plus_one_checkpoint_converges_all_shard
        );
    }
 }
+
+async fn owner_repair_no_backups_fixture() -> PhysicalFixture {
+    let fixture = fixture(0).await;
+    for pool in disks(&fixture.store).await {
+        for disk in pool {
+            let object = object_path(&disk);
+            let mut entries = fs::read_dir(&object).await.expect("fixture metadata directory");
+            while let Some(entry) = entries.next_entry().await.expect("fixture directory entry") {
+                if entry.file_type().await.expect("fixture entry type").is_dir() {
+                    fs::remove_dir_all(entry.path())
+                        .await
+                        .expect("remove task-owned fixture rollback directory");
+                }
+            }
+        }
+    }
+    fixture
+}
+
+async fn owner_repair_identity_heads(store: &ECStore) -> Vec<Vec<(Uuid, Bytes)>> {
+    let mut all = Vec::new();
+    for pool in &store.pools {
+        let set = pool.get_disks_by_key(POOL_META_IDENTITY_NAME);
+        let members = set
+            .disks
+            .read()
+            .await
+            .iter()
+            .cloned()
+            .map(|disk| disk.expect("identity fixture disk must be online"))
+            .collect::<Vec<_>>();
+        let mut heads = Vec::new();
+        for disk in members {
+            match disk
+                .pool_meta_native_recovery(Request::InspectIdentity {})
+                .await
+                .expect("physical identity inspection")
+            {
+                Response::IdentityInspected { disk_id, raw } => heads.push((disk_id, raw)),
+                _ => panic!("identity inspection must return its physical head"),
+            }
+        }
+        all.push(heads);
+    }
+    all
+}
+
+async fn legacy_native_owner_fixture(version: u16, stale: bool) -> (PhysicalFixture, Vec<u8>) {
+    let fixture = owner_repair_no_backups_fixture().await;
+    let mut meta = match decode_pool_meta_replica(fixture.canonical.clone()) {
+        PoolMetaReplica::Valid {
+            meta, committed: true, ..
+        } => meta,
+        _ => panic!("fixture must provide valid committed pool metadata"),
+    };
+    meta.version = version;
+    let raw = meta
+        .encode_config_data_for_v2_gate(version == POOL_META_VERSION)
+        .expect("legacy fixture should encode in its original version");
+    assert_eq!(u16::from_le_bytes([raw[2], raw[3]]), version);
+    let all = disks(&fixture.store).await;
+    let now = OffsetDateTime::now_utc();
+    for pool in &all {
+        let heads = replicas(&raw, now).await;
+        for (index, disk) in pool.iter().enumerate() {
+            fs::write(object_path(disk).join("xl.meta"), &heads[index])
+                .await
+                .expect("install exact legacy head");
+        }
+    }
+    if stale {
+        let old = replicas(&raw, now - Duration::seconds(10)).await;
+        fs::write(object_path(&all[0][0]).join("xl.meta"), &old[0])
+            .await
+            .expect("install stale same-raw legacy head");
+    }
+    (fixture, raw)
+}
+
+#[tokio::test]
+#[serial_test::serial(pool_meta_version_env)]
+async fn pool_meta_native_heal_v1_with_all_upgrade_gates_disabled_keeps_exact_raw() {
+    temp_env::async_with_vars(
+        [
+            (rustfs_config::ENV_POOL_META_V2_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V2_FLEET_CONFIRMED, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_FLEET_CONFIRMED, Some("false")),
+        ],
+        async {
+            let (fixture, raw) = legacy_native_owner_fixture(POOL_META_V1_VERSION, false).await;
+            let guard = owner(&fixture.store).await;
+            let mut state = fixture.store.pool_meta_save_gate.lock().await;
+            let result = tokio::time::timeout(
+                std::time::Duration::from_secs(30),
+                fixture.store.heal_pool_meta_replica_under_fence(0, guard, &mut state),
+            )
+            .await
+            .expect("legacy native heal must terminate");
+            assert!(result.is_ok(), "fully online legacy native heal should succeed: {result:?}");
+            drop(state);
+            let reads = read_pool_meta_replicas(fixture.store.pools.clone(), true).await;
+            for read in &reads {
+                assert!(
+                    matches!(&read.replica, PoolMetaReplica::Valid { raw: bytes, revision, .. }
+                    if bytes == &raw && revision.version == POOL_META_V1_VERSION),
+                    "native heal must not upgrade a V1 raw payload when both upgrade gates are disabled"
+                );
+            }
+        },
+    )
+    .await;
+}
+
+#[tokio::test]
+#[serial_test::serial(pool_meta_version_env)]
+async fn pool_meta_native_heal_rejects_legacy_source_with_trailing_bytes_without_writes() {
+    temp_env::async_with_vars(
+        [
+            (rustfs_config::ENV_POOL_META_V2_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V2_FLEET_CONFIRMED, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_FLEET_CONFIRMED, Some("false")),
+        ],
+        async {
+            for version in [POOL_META_V1_VERSION, POOL_META_VERSION] {
+                let (fixture, raw) = legacy_native_owner_fixture(version, false).await;
+                let mut source_raw = raw.clone();
+                // A second MessagePack value is readable by legacy decoding but
+                // must not be copied as part of an unverified repair payload.
+                source_raw.push(0xc0);
+                assert!(
+                    matches!(decode_pool_meta_replica(source_raw.clone()), PoolMetaReplica::Valid { raw, .. } if raw == source_raw),
+                    "ordinary legacy reads must remain compatible with trailing bytes"
+                );
+                let all_disks = disks(&fixture.store).await;
+                let source_heads = replicas(&source_raw, OffsetDateTime::now_utc() + Duration::seconds(1)).await;
+                for (index, disk) in all_disks[0].iter().enumerate() {
+                    fs::write(object_path(disk).join("xl.meta"), &source_heads[index])
+                        .await
+                        .expect("install readable legacy source with trailing bytes");
+                }
+                let selection = load_pool_meta_replicas(fixture.store.pools.clone(), true)
+                    .await
+                    .expect("legacy semantic selection must remain readable");
+                assert_eq!(selection.legacy_raw.as_deref(), Some(source_raw.as_slice()));
+                let before = snapshots(&fixture.store).await;
+                let identities = owner_repair_identity_heads(&fixture.store).await;
+                let guard = owner(&fixture.store).await;
+                let mut state = fixture.store.pool_meta_save_gate.lock().await;
+                let result = tokio::time::timeout(
+                    std::time::Duration::from_secs(30),
+                    fixture.store.heal_pool_meta_replica_under_fence(1, guard, &mut state),
+                )
+                .await
+                .expect("legacy native heal must terminate");
+                drop(state);
+                assert!(result.is_err(), "version {version}: unparsed trailing source bytes must not be republished");
+                assert_eq!(snapshots(&fixture.store).await, before, "version {version}: no pool.bin head or backup may change");
+                assert_eq!(
+                    owner_repair_identity_heads(&fixture.store).await,
+                    identities,
+                    "version {version}: no identity shard may change"
+                );
+                let reads = read_pool_meta_replicas(fixture.store.pools.clone(), true).await;
+                assert!(matches!(&reads[0].replica, PoolMetaReplica::Valid { raw, .. } if raw == &source_raw));
+                assert!(matches!(&reads[1].replica, PoolMetaReplica::Valid { raw: target, .. } if target == &raw));
+            }
+        },
+    )
+    .await;
+}
+
+#[tokio::test]
+#[serial_test::serial(pool_meta_version_env)]
+async fn pool_meta_native_heal_mixed_v1_v2_preserves_observed_format_floor() {
+    temp_env::async_with_vars(
+        [
+            (rustfs_config::ENV_POOL_META_V2_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V2_FLEET_CONFIRMED, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_FLEET_CONFIRMED, Some("false")),
+        ],
+        async {
+            for first_version in [POOL_META_V1_VERSION, POOL_META_VERSION] {
+                for target_pool in [0, 1] {
+                    let (fixture, raw_v1) = legacy_native_owner_fixture(POOL_META_V1_VERSION, false).await;
+                    let mut meta = match decode_pool_meta_replica(raw_v1.clone()) {
+                        PoolMetaReplica::Valid { meta, committed: true, .. } => meta,
+                        _ => panic!("initial V1 metadata must decode"),
+                    };
+                    meta.version = POOL_META_VERSION;
+                    let raw_v2 = meta.encode_config_data_for_v2_gate(true).expect("equivalent V2 payload must encode");
+                    assert_eq!(u16::from_le_bytes([raw_v1[2], raw_v1[3]]), POOL_META_V1_VERSION);
+                    assert_eq!(u16::from_le_bytes([raw_v2[2], raw_v2[3]]), POOL_META_VERSION);
+                    let semantic_v1 = match decode_pool_meta_replica(raw_v1.clone()) {
+                        PoolMetaReplica::Valid { canonical, .. } => canonical,
+                        _ => panic!("V1 payload must remain valid"),
+                    };
+                    let semantic_v2 = match decode_pool_meta_replica(raw_v2.clone()) {
+                        PoolMetaReplica::Valid { canonical, .. } => canonical,
+                        _ => panic!("V2 payload must remain valid"),
+                    };
+                    assert_eq!(semantic_v1, semantic_v2, "the two formats must carry equivalent metadata");
+                    let other_version = if first_version == POOL_META_V1_VERSION { POOL_META_VERSION } else { POOL_META_V1_VERSION };
+                    // Pool two follows pool zero; pools zero and one cover both
+                    // orderings and both target formats in separate runs.
+                    let versions = [first_version, other_version, first_version];
+                    let physical = disks(&fixture.store).await;
+                    let timestamp = OffsetDateTime::now_utc();
+                    for (pool_index, pool) in physical.iter().enumerate() {
+                        let raw = if versions[pool_index] == POOL_META_V1_VERSION { &raw_v1 } else { &raw_v2 };
+                        let heads = replicas(raw, timestamp).await;
+                        for (index, disk) in pool.iter().enumerate() {
+                            fs::write(object_path(disk).join("xl.meta"), &heads[index]).await.expect("install mixed legacy pool head");
+                        }
+                    }
+                    let before = snapshots(&fixture.store).await;
+                    let selection = load_pool_meta_replicas(fixture.store.pools.clone(), true)
+                        .await.expect("equivalent legacy pools should be selectable");
+                    assert_eq!(selection.revision.version, POOL_META_VERSION, "the observed V2 format floor must be retained");
+                    assert_eq!(selection.canonical.as_deref(), Some(semantic_v2.as_slice()));
+                    let guard = owner(&fixture.store).await;
+                    let mut state = fixture.store.pool_meta_save_gate.lock().await;
+                    tokio::time::timeout(std::time::Duration::from_secs(30),
+                        fixture.store.heal_pool_meta_replica_under_fence(target_pool, guard, &mut state))
+                        .await.expect("mixed legacy heal must terminate")
+                        .expect("equivalent legacy metadata should heal without downgrading the format floor");
+                    drop(state);
+                    let after = snapshots(&fixture.store).await;
+                    let reads = read_pool_meta_replicas(fixture.store.pools.clone(), true).await;
+                    for (pool_index, read) in reads.iter().enumerate() {
+                        let expected_version = if pool_index == target_pool { POOL_META_VERSION } else { versions[pool_index] };
+                        let expected_raw = if expected_version == POOL_META_V1_VERSION { &raw_v1 } else { &raw_v2 };
+                        assert!(matches!(&read.replica,
+                            PoolMetaReplica::Valid { raw, canonical, revision, committed: true, .. }
+                                if raw == expected_raw && canonical == &semantic_v2 && revision.version == expected_version
+                        ), "pool {pool_index}, first={first_version}, target={target_pool}: preserve equivalent payload and expected legacy version");
+                        for (index, snapshot) in after[pool_index].iter().enumerate() {
+                            codec::validate_inline_pool_meta_shard(snapshot.current.as_ref().expect("physical head"), index, expected_raw)
+                                .await.expect("every physical shard must match its selected legacy payload");
+                        }
+                        if pool_index != target_pool {
+                            assert_eq!(after[pool_index], before[pool_index], "healing one pool must not modify other pools");
+                        }
+                    }
+                }
+            }
+        },
+    ).await;
+}
+
+#[tokio::test]
+#[serial_test::serial(pool_meta_version_env)]
+async fn pool_meta_native_heal_rejects_v1_and_v3_prepare_in_either_pool_order() {
+    temp_env::async_with_vars(
+        [
+            (rustfs_config::ENV_POOL_META_V2_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V2_FLEET_CONFIRMED, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_FLEET_CONFIRMED, Some("false")),
+        ],
+        async {
+            for prepare_pool in [0, 1] {
+                let (fixture, raw_v1) = legacy_native_owner_fixture(POOL_META_V1_VERSION, false).await;
+                let (legacy_meta, legacy_canonical) = match decode_pool_meta_replica(raw_v1.clone()) {
+                    PoolMetaReplica::Valid {
+                        meta,
+                        canonical,
+                        revision,
+                        committed: true,
+                        ..
+                    } if revision.version == POOL_META_V1_VERSION => (meta, canonical),
+                    _ => panic!("baseline V1 metadata must be committed and readable"),
+                };
+                let previous = PoolMetaCommittedCandidate {
+                    canonical: legacy_canonical,
+                    meta: legacy_meta.clone(),
+                    revision: PoolMetaRevision::legacy(POOL_META_V1_VERSION),
+                };
+                let mut next = legacy_meta;
+                next.version = POOL_META_GENERATION_VERSION;
+                next.pools[0].last_update += Duration::seconds(1);
+                let revision = PoolMetaRevision {
+                    version: POOL_META_GENERATION_VERSION,
+                    cluster_id: fixture.revision.cluster_id,
+                    epoch: fixture.revision.epoch,
+                    generation: 1,
+                    transaction_id: Some(Uuid::new_v4()),
+                };
+                let pending = encode_pool_meta_v3_envelope(&next, revision, false, Some(&previous))
+                    .expect("V3 PREPARE with V1 predecessor must encode");
+                assert!(matches!(decode_pool_meta_replica(pending.clone()),
+                    PoolMetaReplica::Valid { committed: false, revision: actual, .. } if actual == revision));
+                let heads = replicas(&pending, OffsetDateTime::now_utc() + Duration::seconds(1)).await;
+                let physical = disks(&fixture.store).await;
+                for (index, disk) in physical[prepare_pool].iter().enumerate() {
+                    fs::write(object_path(disk).join("xl.meta"), &heads[index])
+                        .await
+                        .expect("install task-owned V3 PREPARE evidence");
+                }
+                let reads = read_pool_meta_replicas(fixture.store.pools.clone(), true).await;
+                assert!(matches!(&reads[prepare_pool].replica,
+                    PoolMetaReplica::Valid { raw, committed: false, revision: actual, .. }
+                        if raw == &pending && *actual == revision));
+                let other_pool = 1 - prepare_pool;
+                assert!(matches!(&reads[other_pool].replica,
+                    PoolMetaReplica::Valid { raw, committed: true, revision: actual, .. }
+                        if raw == &raw_v1 && actual.version == POOL_META_V1_VERSION));
+                let before = snapshots(&fixture.store).await;
+                let identities = owner_repair_identity_heads(&fixture.store).await;
+                let guard = owner(&fixture.store).await;
+                let mut state = fixture.store.pool_meta_save_gate.lock().await;
+                let result = tokio::time::timeout(
+                    std::time::Duration::from_secs(30),
+                    fixture
+                        .store
+                        .heal_pool_meta_replica_under_fence(prepare_pool, guard, &mut state),
+                )
+                .await
+                .expect("mixed V1/V3 PREPARE heal must terminate");
+                drop(state);
+                assert!(
+                    result.is_err(),
+                    "pool {prepare_pool}: native heal must not overwrite a V3 PREPARE with legacy data"
+                );
+                assert_eq!(
+                    snapshots(&fixture.store).await,
+                    before,
+                    "pool {prepare_pool}: rejection must retain all physical head and backup evidence"
+                );
+                assert_eq!(
+                    owner_repair_identity_heads(&fixture.store).await,
+                    identities,
+                    "pool {prepare_pool}: rejection must not mutate identity evidence"
+                );
+            }
+        },
+    )
+    .await;
+}
+
+#[tokio::test]
+#[serial_test::serial(pool_meta_version_env)]
+async fn pool_meta_native_heal_rejects_v1_legacy_dont_save_true_without_writes() {
+    temp_env::async_with_vars(
+        [
+            (rustfs_config::ENV_POOL_META_V2_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V2_FLEET_CONFIRMED, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_FLEET_CONFIRMED, Some("false")),
+        ],
+        async {
+            let (fixture, safe_v1) = legacy_native_owner_fixture(POOL_META_V1_VERSION, false).await;
+            let meta = match decode_pool_meta_replica(safe_v1.clone()) {
+                PoolMetaReplica::Valid {
+                    meta, committed: true, ..
+                } => meta,
+                _ => panic!("baseline V1 metadata must decode"),
+            };
+            let statuses = meta
+                .pools
+                .iter()
+                .map(|pool| {
+                    assert!(pool.decommission.is_none(), "this regression fixture does not inject decommission state");
+                    LegacyPoolStatus {
+                        id: pool.id,
+                        cmd_line: pool.cmd_line.clone(),
+                        last_update: pool.last_update,
+                        decommission: None,
+                    }
+                })
+                .collect();
+            let legacy = LegacyPoolMeta {
+                version: POOL_META_V1_VERSION,
+                pools: statuses,
+                dont_save: true,
+            };
+            let mut dangerous_raw = Vec::new();
+            dangerous_raw.extend_from_slice(&POOL_META_FORMAT.to_le_bytes());
+            dangerous_raw.extend_from_slice(&POOL_META_V1_VERSION.to_le_bytes());
+            dangerous_raw.extend_from_slice(&rmp_serde::to_vec(&legacy).expect("legacy three-field tuple must serialize"));
+            let old_reader: LegacyPoolMeta =
+                rmp_serde::from_slice(&dangerous_raw[4..]).expect("legacy reader must decode source");
+            assert!(old_reader.dont_save, "the old reader would treat this source as permanently do-not-save");
+            assert_eq!(rmp::decode::read_array_len(&mut &dangerous_raw[4..]).expect("tuple length"), 3);
+            let source = match decode_pool_meta_replica(dangerous_raw.clone()) {
+                PoolMetaReplica::Valid {
+                    canonical,
+                    revision,
+                    committed: true,
+                    ..
+                } => {
+                    assert_eq!(revision.version, POOL_META_V1_VERSION);
+                    canonical
+                }
+                _ => panic!("current reader must recognize the legacy source"),
+            };
+            assert!(matches!(decode_pool_meta_replica(safe_v1.clone()),
+                PoolMetaReplica::Valid { canonical, revision, committed: true, .. }
+                    if canonical == source && revision.version == POOL_META_V1_VERSION));
+            let physical = disks(&fixture.store).await;
+            let source_heads = replicas(&dangerous_raw, OffsetDateTime::now_utc() + Duration::seconds(1)).await;
+            for (index, disk) in physical[0].iter().enumerate() {
+                fs::write(object_path(disk).join("xl.meta"), &source_heads[index])
+                    .await
+                    .expect("install legacy source on pool zero");
+            }
+            let reads = read_pool_meta_replicas(fixture.store.pools.clone(), true).await;
+            assert!(matches!(&reads[0].replica, PoolMetaReplica::Valid { raw, committed: true, .. } if raw == &dangerous_raw));
+            assert!(matches!(&reads[1].replica, PoolMetaReplica::Valid { raw, committed: true, .. } if raw == &safe_v1));
+            let before = snapshots(&fixture.store).await;
+            let identities = owner_repair_identity_heads(&fixture.store).await;
+            let guard = owner(&fixture.store).await;
+            let mut state = fixture.store.pool_meta_save_gate.lock().await;
+            let result = tokio::time::timeout(
+                std::time::Duration::from_secs(30),
+                fixture.store.heal_pool_meta_replica_under_fence(1, guard, &mut state),
+            )
+            .await
+            .expect("unsafe legacy heal must terminate");
+            drop(state);
+            let error = result.expect_err("native heal must reject a source whose old-reader dont_save is true");
+            assert!(
+                error.to_string().contains("dont_save"),
+                "rejection must identify the unsafe legacy field: {error}"
+            );
+            assert_eq!(
+                snapshots(&fixture.store).await,
+                before,
+                "rejection must preserve every original head and backup"
+            );
+            assert_eq!(
+                owner_repair_identity_heads(&fixture.store).await,
+                identities,
+                "rejection must not write identity shards"
+            );
+            let after = read_pool_meta_replicas(fixture.store.pools.clone(), true).await;
+            assert!(matches!(&after[0].replica, PoolMetaReplica::Valid { raw, .. } if raw == &dangerous_raw));
+            assert!(
+                matches!(&after[1].replica, PoolMetaReplica::Valid { raw, .. } if raw == &safe_v1),
+                "the dangerous source bytes must not be copied into the target pool"
+            );
+        },
+    )
+    .await;
+}
+
+#[tokio::test]
+#[serial_test::serial(pool_meta_version_env)]
+async fn pool_meta_native_heal_rejects_target_v1_dont_save_true_without_writes() {
+    temp_env::async_with_vars(
+        [
+            (rustfs_config::ENV_POOL_META_V2_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V2_FLEET_CONFIRMED, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_WRITE, Some("false")),
+            (rustfs_config::ENV_POOL_META_V3_FLEET_CONFIRMED, Some("false")),
+        ],
+        async {
+            let (fixture, safe_v1) = legacy_native_owner_fixture(POOL_META_V1_VERSION, false).await;
+            let meta = match decode_pool_meta_replica(safe_v1.clone()) {
+                PoolMetaReplica::Valid {
+                    meta, committed: true, ..
+                } => meta,
+                _ => panic!("baseline V1 metadata must decode"),
+            };
+            let statuses = meta
+                .pools
+                .iter()
+                .map(|pool| {
+                    assert!(pool.decommission.is_none(), "this regression fixture does not inject decommission state");
+                    LegacyPoolStatus {
+                        id: pool.id,
+                        cmd_line: pool.cmd_line.clone(),
+                        last_update: pool.last_update,
+                        decommission: None,
+                    }
+                })
+                .collect();
+            let legacy = LegacyPoolMeta {
+                version: POOL_META_V1_VERSION,
+                pools: statuses,
+                dont_save: true,
+            };
+            let mut dangerous_raw = Vec::new();
+            dangerous_raw.extend_from_slice(&POOL_META_FORMAT.to_le_bytes());
+            dangerous_raw.extend_from_slice(&POOL_META_V1_VERSION.to_le_bytes());
+            dangerous_raw.extend_from_slice(&rmp_serde::to_vec(&legacy).expect("legacy three-field tuple must serialize"));
+            let old_reader: LegacyPoolMeta =
+                rmp_serde::from_slice(&dangerous_raw[4..]).expect("legacy reader must decode target");
+            assert!(old_reader.dont_save, "the old reader interprets the target as do-not-save");
+            assert_eq!(rmp::decode::read_array_len(&mut &dangerous_raw[4..]).expect("tuple length"), 3);
+            let source_canonical = match decode_pool_meta_replica(safe_v1.clone()) {
+                PoolMetaReplica::Valid {
+                    canonical,
+                    revision,
+                    committed: true,
+                    ..
+                } => {
+                    assert_eq!(revision.version, POOL_META_V1_VERSION);
+                    canonical
+                }
+                _ => panic!("safe pool-zero source must decode"),
+            };
+            assert!(
+                matches!(decode_pool_meta_replica(dangerous_raw.clone()),
+                PoolMetaReplica::Valid { canonical, revision, committed: true, .. }
+                    if canonical == source_canonical && revision.version == POOL_META_V1_VERSION),
+                "the dangerous target must remain semantically equivalent to the safe source"
+            );
+            let physical = disks(&fixture.store).await;
+            let target_heads = replicas(&dangerous_raw, OffsetDateTime::now_utc() + Duration::seconds(1)).await;
+            for (index, disk) in physical[1].iter().enumerate() {
+                fs::write(object_path(disk).join("xl.meta"), &target_heads[index])
+                    .await
+                    .expect("install legacy dangerous target head");
+            }
+            let reads = read_pool_meta_replicas(fixture.store.pools.clone(), true).await;
+            assert!(matches!(&reads[0].replica, PoolMetaReplica::Valid { raw, committed: true, .. } if raw == &safe_v1));
+            assert!(matches!(&reads[1].replica, PoolMetaReplica::Valid { raw, committed: true, .. } if raw == &dangerous_raw));
+            let before = snapshots(&fixture.store).await;
+            let identities = owner_repair_identity_heads(&fixture.store).await;
+            let guard = owner(&fixture.store).await;
+            let mut state = fixture.store.pool_meta_save_gate.lock().await;
+            let result = tokio::time::timeout(
+                std::time::Duration::from_secs(30),
+                fixture.store.heal_pool_meta_replica_under_fence(1, guard, &mut state),
+            )
+            .await
+            .expect("unsafe legacy target heal must terminate");
+            drop(state);
+            let error = result.expect_err("native heal must reject a target whose old-reader dont_save is true");
+            assert!(
+                error.to_string().contains("dont_save"),
+                "rejection must identify the unsafe legacy field: {error}"
+            );
+            assert_eq!(
+                snapshots(&fixture.store).await,
+                before,
+                "rejection must preserve all physical heads and backups"
+            );
+            assert_eq!(
+                owner_repair_identity_heads(&fixture.store).await,
+                identities,
+                "rejection must not write identity shards"
+            );
+            let after = read_pool_meta_replicas(fixture.store.pools.clone(), true).await;
+            assert!(matches!(&after[0].replica, PoolMetaReplica::Valid { raw, .. } if raw == &safe_v1));
+            assert!(
+                matches!(&after[1].replica, PoolMetaReplica::Valid { raw, .. } if raw == &dangerous_raw),
+                "unsafe target evidence must not be overwritten by the safe source"
+            );
+        },
+    )
+    .await;
+}
@@ -94,7 +94,6 @@ use serde::{Deserialize, Serialize};
 use sha2::{Digest, Sha256};
 use std::collections::{HashMap, HashSet};
 use std::fmt::Display;
-#[cfg(test)]
 use std::io::Cursor;
 use std::io::Write;
 use std::path::PathBuf;
@@ -6839,11 +6838,47 @@ impl Drop for PoolMetaTransactionArm {
    }
 }

+fn ensure_legacy_pool_meta_republish_safe(data: &[u8]) -> Result<()> {
+    if data.len() >= 4
+        && LittleEndian::read_u16(&data[2..4]) == POOL_META_V1_VERSION
+        // The selection already validates the payload. The runtime-only
+        // flag can only appear in the legacy three-field sequence or map.
+        && rmp::decode::read_array_len(&mut &data[4..])
+            .or_else(|_| rmp::decode::read_map_len(&mut &data[4..]))
+            .ok()
+            == Some(3)
+        && let Ok(legacy) = rmp_serde::from_slice::<LegacyPoolMeta>(&data[4..])
+        && legacy.dont_save
+    {
+        return Err(Error::other("pool metadata legacy source contains dont_save=true"));
+    }
+    Ok(())
+}
+
+// Ordinary legacy reads retain their compatibility behavior, but bytes that
+// will be copied to another pool must not contain an unparsed extension.
+fn ensure_legacy_pool_meta_payload_complete(data: &[u8]) -> Result<()> {
+    let payload = data
+        .get(4..)
+        .ok_or_else(|| Error::other("pool metadata legacy source is too short"))?;
+    let mut decoder = rmp_serde::Deserializer::new(Cursor::new(payload));
+    serde::de::IgnoredAny::deserialize(&mut decoder).map_err(|_| Error::other("pool metadata legacy source is invalid"))?;
+    let payload_len = u64::try_from(payload.len()).map_err(|_| Error::other("pool metadata legacy source is too large"))?;
+    if decoder.get_ref().position() != payload_len {
+        return Err(Error::other("pool metadata legacy source has trailing bytes"));
+    }
+    Ok(())
+}
+
 #[derive(Debug)]
 struct PoolMetaSelection {
    meta: PoolMeta,
    revision: PoolMetaRevision,
    canonical: Option<Vec<u8>>,
+    // Legacy semantic comparison canonicalizes to V2; a heal must instead
+    // preserve verified raw bytes at the highest observed format floor.
+    legacy_raw: Option<Vec<u8>>,
+    legacy_has_unsafe_flag: bool,
    replica_state: PoolMetaReplicaState,
    cas_tokens: Vec<PoolMetaCasToken>,
    absent: bool,
@@ -7234,6 +7269,8 @@ fn select_pool_meta_replica_reads(reads: Vec<PoolMetaReplicaRead>) -> Result<Poo
            meta: PoolMeta::default(),
            revision: PoolMetaRevision::legacy(0),
            canonical: None,
+            legacy_raw: None,
+            legacy_has_unsafe_flag: false,
            replica_state: PoolMetaReplicaState {
                needs_repair: false,
                repair_write_safe: true,
@@ -7288,6 +7325,8 @@ fn select_pool_meta_replica_reads(reads: Vec<PoolMetaReplicaRead>) -> Result<Poo
            meta: selected.meta.clone(),
            revision: selected.revision,
            canonical: Some(selected.canonical.clone()),
+            legacy_raw: None,
+            legacy_has_unsafe_flag: false,
            replica_state: PoolMetaReplicaState {
                needs_repair,
                repair_write_safe,
@@ -7301,6 +7340,10 @@ fn select_pool_meta_replica_reads(reads: Vec<PoolMetaReplicaRead>) -> Result<Poo

    // Legacy V1/V2 has no durable generation. Pool zero remains the commit
    // record; divergent backups without it are ambiguous and fail closed.
+    let legacy_has_unsafe_flag = committed
+        .iter()
+        .any(|(_, raw, _)| ensure_legacy_pool_meta_republish_safe(raw).is_err());
+    let mut floor_raw = None;
    let mut selected: Option<(usize, Vec<u8>, PoolMetaCommittedCandidate)> = None;
    for (idx, raw, candidate) in committed {
        if let Some((selected_idx, selected_raw, selected_candidate)) = selected.as_ref() {
@@ -7314,19 +7357,36 @@ fn select_pool_meta_replica_reads(reads: Vec<PoolMetaReplicaRead>) -> Result<Poo
                }
            } else {
                needs_repair |= selected_raw != &raw;
+                if floor_raw.is_none()
+                    && candidate.revision.version == observed_version
+                    && raw
+                        .get(2..4)
+                        .is_some_and(|version| LittleEndian::read_u16(version) == observed_version)
+                {
+                    floor_raw = Some(raw);
+                }
            }
        } else {
            selected = Some((idx, raw, candidate));
        }
    }
-    let (_, _, mut selected) =
+    let (_, selected_raw, mut selected) =
        selected.ok_or_else(|| Error::other("pool metadata recovery required: no valid legacy replica is available"))?;
+    if selected.revision.version == observed_version
+        && selected_raw
+            .get(2..4)
+            .is_some_and(|version| LittleEndian::read_u16(version) == observed_version)
+    {
+        floor_raw = Some(selected_raw);
+    }
    selected.meta.version = observed_version;
    selected.revision.version = observed_version;
    Ok(PoolMetaSelection {
        meta: selected.meta,
        revision: selected.revision,
        canonical: Some(selected.canonical),
+        legacy_raw: floor_raw,
+        legacy_has_unsafe_flag,
        replica_state: PoolMetaReplicaState {
            needs_repair,
            repair_write_safe,
@@ -14546,6 +14606,23 @@ impl ECStore {
            .canonical
            .as_ref()
            .ok_or_else(|| Error::other("pool metadata heal has no committed source"))?;
+        if selection.generation_protocol_observed && !selection.revision.is_generation_protocol() {
+            return Err(Error::other("pool metadata heal requires recovery of the observed generation protocol"));
+        }
+        if selection.legacy_has_unsafe_flag {
+            return Err(Error::other("pool metadata legacy replica contains dont_save=true"));
+        }
+        let repair_data = if selection.revision.is_generation_protocol() {
+            canonical.clone()
+        } else {
+            let raw = selection
+                .legacy_raw
+                .clone()
+                .ok_or_else(|| Error::other("pool metadata heal cannot prove the legacy format floor"))?;
+            ensure_legacy_pool_meta_payload_complete(&raw)?;
+            raw
+        };
+        ensure_legacy_pool_meta_republish_safe(&repair_data)?;
        let pool = self
            .pools
            .get(pool_index)
@@ -14556,10 +14633,10 @@ impl ECStore {
            .get(pool_index)
            .ok_or_else(|| Error::other("pool metadata heal has no target revision"))?;
        let mut arm = state.arm_transaction();
-        save_pool_meta_object_cas(pool.clone(), POOL_META_NAME, canonical.clone(), token, &fence, "native_heal", &mut arm)
+        save_pool_meta_object_cas(pool.clone(), POOL_META_NAME, repair_data.clone(), token, &fence, "native_heal", &mut arm)
            .await?;
        let (confirmed, _) = read_config_no_lock_preserve_empty_with_metadata(pool, POOL_META_NAME).await?;
-        if confirmed != *canonical {
+        if confirmed != repair_data {
            return Err(Error::other("pool metadata native heal readback diverged"));
        }
        fence.ensure_held()?;
@@ -29482,6 +29559,53 @@ mod tests {
        );
    }

+    #[test]
+    fn legacy_republish_guard_checks_only_three_field_sequence_or_map() {
+        let frame = |payload: Vec<u8>| {
+            let mut data = Vec::new();
+            data.extend_from_slice(&POOL_META_FORMAT.to_le_bytes());
+            data.extend_from_slice(&POOL_META_V1_VERSION.to_le_bytes());
+            data.extend_from_slice(&payload);
+            data
+        };
+        let two_fields = PersistedPoolMetaV1 {
+            version: POOL_META_V1_VERSION,
+            pools: Vec::new(),
+        };
+        for payload in [rmp_serde::to_vec(&two_fields), rmp_serde::to_vec_named(&two_fields)] {
+            let data = frame(payload.expect("two-field V1 payload must serialize"));
+            assert!(matches!(decode_pool_meta_replica(data.clone()), PoolMetaReplica::Valid { .. }));
+            assert!(ensure_legacy_pool_meta_republish_safe(&data).is_ok());
+            assert!(ensure_legacy_pool_meta_payload_complete(&data).is_ok());
+            let mut trailing = data;
+            trailing.push(0xc0);
+            assert!(ensure_legacy_pool_meta_payload_complete(&trailing).is_err());
+        }
+        for dont_save in [false, true] {
+            let legacy = LegacyPoolMeta {
+                version: POOL_META_V1_VERSION,
+                pools: Vec::new(),
+                dont_save,
+            };
+            for payload in [rmp_serde::to_vec(&legacy), rmp_serde::to_vec_named(&legacy)] {
+                let data = frame(payload.expect("three-field V1 payload must serialize"));
+                assert!(matches!(decode_pool_meta_replica(data.clone()), PoolMetaReplica::Valid { .. }));
+                assert_eq!(ensure_legacy_pool_meta_republish_safe(&data).is_err(), dont_save);
+                assert!(ensure_legacy_pool_meta_payload_complete(&data).is_ok());
+                let mut trailing = data;
+                trailing.push(0xc0);
+                assert!(ensure_legacy_pool_meta_payload_complete(&trailing).is_err());
+            }
+        }
+        // This guard is only called for already-decoded replicas; malformed
+        // containers must fail selection rather than becoming repair sources.
+        let truncated = frame(vec![0x93, 0x01]);
+        assert!(matches!(
+            decode_pool_meta_replica(truncated),
+            PoolMetaReplica::Corrupt(_) | PoolMetaReplica::Incompatible(_)
+        ));
+    }
+
    #[test]
    fn pool_meta_decode_supports_legacy_payload() {
        let start_time = OffsetDateTime::now_utc();
@@ -853,6 +853,8 @@ pub(crate) struct DiskHealthMetricEpoch {
    fallback_operation_metrics: StdRwLock<HashMap<&'static str, Arc<DiskOperationMetrics>>>,
    #[cfg(test)]
    multipart_cleanup_observer: std::sync::Mutex<Option<std::sync::Weak<ScopedMultipartCleanupCompletions>>>,
+    #[cfg(test)]
+    operation_timeout_override: std::sync::Mutex<Option<Duration>>,
    /// Caller API operations currently executing through the disk health wrapper.
    api_waiting: AtomicU32,
    /// Operations rejected because the disk was unavailable for the caller.
@@ -880,6 +882,8 @@ impl Default for DiskHealthMetricEpoch {
            fallback_operation_metrics: StdRwLock::new(HashMap::new()),
            #[cfg(test)]
            multipart_cleanup_observer: std::sync::Mutex::new(None),
+            #[cfg(test)]
+            operation_timeout_override: std::sync::Mutex::new(None),
            api_waiting: AtomicU32::new(0),
            total_errors_availability: AtomicU64::new(0),
            total_errors_timeout: AtomicU64::new(0),
@@ -1573,6 +1577,20 @@ impl Default for DiskHealthTracker {
    }
 }

+#[cfg(test)]
+#[derive(Debug)]
+pub(crate) struct ScopedDiskOperationTimeout {
+    metrics: Arc<DiskHealthMetricEpoch>,
+    previous: Option<Duration>,
+}
+
+#[cfg(test)]
+impl Drop for ScopedDiskOperationTimeout {
+    fn drop(&mut self) {
+        *self.metrics.operation_timeout_override.lock().expect("test timeout lock") = self.previous;
+    }
+}
+
 /// LocalDiskWrapper wraps a DiskStore with health tracking capabilities.
 /// This is similar to Go's xlStorageDiskIDCheck.
 #[derive(Debug, Clone)]
@@ -1914,6 +1932,23 @@ impl LocalDiskWrapper {
        self.health_check
    }

+    #[cfg(test)]
+    pub(crate) fn override_operation_timeout_for_test(&self, timeout: Duration) -> ScopedDiskOperationTimeout {
+        assert!(!timeout.is_zero(), "fault injection needs a positive timeout");
+        // The metrics epoch follows this drive across clones and reconnects.
+        // Its scoped override cannot shorten unrelated healthy drives' I/O.
+        let previous = self
+            .metrics
+            .operation_timeout_override
+            .lock()
+            .expect("test timeout lock")
+            .replace(timeout);
+        ScopedDiskOperationTimeout {
+            metrics: Arc::clone(&self.metrics),
+            previous,
+        }
+    }
+
    /// Enable health monitoring after disk creation.
    /// Used to defer health checks until after startup format loading completes.
    pub fn enable_health_check(&self) {
@@ -2638,6 +2673,13 @@ impl LocalDiskWrapper {
            }
            return result;
        }
+        #[cfg(test)]
+        let timeout_duration = self
+            .metrics
+            .operation_timeout_override
+            .lock()
+            .expect("test timeout lock")
+            .map_or(timeout_duration, |injected| timeout_duration.min(injected));
        // Execute the operation with timeout
        let result = tokio::time::timeout(timeout_duration, operation).await;

@@ -6044,6 +6086,67 @@ mod tests {
        assert!(health.offline_duration().is_none());
    }

+    #[tokio::test]
+    async fn scoped_operation_timeout_is_per_disk_and_preserves_caller_policy() {
+        let directories = [
+            tempfile::tempdir().expect("faulted drive"),
+            tempfile::tempdir().expect("healthy drive"),
+        ];
+        let mut wrappers = Vec::new();
+        for directory in &directories {
+            let endpoint = Endpoint::try_from(directory.path().to_str().expect("UTF-8 drive path")).expect("drive endpoint");
+            let disk = Arc::new(LocalDisk::new(&endpoint, false).await.expect("local drive"));
+            wrappers.push(LocalDiskWrapper::new(disk, false));
+        }
+        let reconnected = LocalDiskWrapper::new_with_reconnect_state(
+            Arc::clone(&wrappers[0].disk),
+            false,
+            Some(wrappers[0].health_tracker_epoch_for_reconnect()),
+        );
+        let scope = wrappers[0].override_operation_timeout_for_test(Duration::from_secs(1));
+        tokio::time::pause();
+        let operation = || async {
+            tokio::time::sleep(Duration::from_secs(2)).await;
+            Ok(())
+        };
+        let (faulted, healthy) = tokio::join!(
+            reconnected.track_disk_health_with_op_and_timeout_action(
+                "read_metadata",
+                operation,
+                Duration::from_secs(3),
+                TimeoutHealthAction::IgnoreFailure,
+            ),
+            wrappers[1].track_disk_health_with_op_and_timeout_action(
+                "read_metadata",
+                operation,
+                Duration::from_secs(3),
+                TimeoutHealthAction::IgnoreFailure,
+            ),
+        );
+        assert_eq!(faulted.expect_err("only the selected drive must time out"), DiskError::Timeout);
+        healthy.expect("healthy drive keeps its caller's deadline");
+        let shorter_started = tokio::time::Instant::now();
+        let shorter = wrappers[0]
+            .track_disk_health_with_op_and_timeout_action(
+                "read_metadata",
+                operation,
+                Duration::from_millis(500),
+                TimeoutHealthAction::IgnoreFailure,
+            )
+            .await;
+        assert_eq!(shorter.expect_err("fault injection cannot extend a caller deadline"), DiskError::Timeout);
+        assert!(shorter_started.elapsed() < Duration::from_secs(1));
+        wrappers[0]
+            .track_disk_health_with_op("read_metadata", operation, Duration::ZERO)
+            .await
+            .expect("explicit unbounded physical ownership cannot acquire a test deadline");
+        drop(scope);
+        reconnected
+            .track_disk_health_with_op("read_metadata", operation, Duration::from_secs(3))
+            .await
+            .expect("dropping the scope restores reconnected callers' deadlines");
+    }
+
    #[tokio::test]
    async fn ignored_timeout_does_not_mark_drive_failure() {
        let dir = tempfile::tempdir().expect("temp dir should be created");
@@ -5,6 +5,7 @@ use super::{FdCache, FdCacheEntry, FdKey};
 use std::fs::File;
 use std::os::unix::fs::FileExt;
 use std::sync::Arc;
+use std::sync::atomic::AtomicBool;

 fn key(volume: &str, path: &str, direct: bool) -> FdKey {
    FdKey {
@@ -52,7 +53,11 @@ async fn uring_fd_cache_exact_invalidation_removes_both_modes_and_preserves_othe
 async fn uring_fd_cache_exact_invalidation_bumps_one_generation_and_rejects_old_epoch_inserts() {
    let (_directory, file) = fixture();
    let cache = FdCache::new();
-    let entry = Arc::new(FdCacheEntry { file, len: 3 });
+    let entry = Arc::new(FdCacheEntry {
+        file,
+        len: 3,
+        invalidated: AtomicBool::new(false),
+    });
    let old_generation = cache.generation();
    cache.invalidate_exact("v", "part").await;
    let current_generation = cache.generation();
@@ -125,7 +130,7 @@ async fn uring_fd_cache_volume_and_clear_invalidation_cover_both_modes() {
        cache.insert(key("v", "part", direct), Arc::clone(&file)).await;
        assert!(
            cache.get(&key("v", "part", direct)).await.is_some(),
-            "new entries must survive the earlier volume predicate"
+            "new entries must survive the earlier volume invalidation"
        );
    }
    cache.clear();
@@ -164,6 +169,7 @@ async fn uring_fd_cache_invalidation_keeps_borrowed_old_inode_alive_and_allows_r
    let replacement = Arc::new(FdCacheEntry {
        file: Arc::new(File::open(directory.path().join("part")).expect("open replacement inode")),
        len: 11,
+        invalidated: AtomicBool::new(false),
    });
    for direct in [false, true] {
        let key = key("v", "part", direct);
@@ -148,58 +148,47 @@ pub async fn make_dir_all(path: impl AsRef<Path>) -> io::Result<()> {
    native_io(move || create_dir_all_std(&path)).await
 }

-fn is_dir_error(path: &Path, e: &io::Error) -> bool {
+fn is_dir_error(e: &io::Error) -> bool {
    e.raw_os_error() == Some(libc::EISDIR)
        || e.kind() == io::ErrorKind::IsADirectory
        // macOS: remove_file on a directory returns EPERM
        || (cfg!(target_os = "macos") && e.raw_os_error() == Some(libc::EPERM))
-        || is_windows_dir_access_denied(path, e)
-}
-
-/// Windows reports `remove_file` on a directory (or directory symlink) as
-/// ERROR_ACCESS_DENIED, which is indistinguishable from a real permission
-/// failure without inspecting the entry itself.
-#[cfg(windows)]
-fn is_windows_dir_access_denied(path: &Path, e: &io::Error) -> bool {
-    use std::os::windows::fs::FileTypeExt;
-
-    e.kind() == io::ErrorKind::PermissionDenied
-        && std::fs::symlink_metadata(path).is_ok_and(|metadata| {
-            let file_type = metadata.file_type();
-            file_type.is_dir() || file_type.is_symlink_dir()
-        })
-}
-
-#[cfg(not(windows))]
-fn is_windows_dir_access_denied(_path: &Path, _e: &io::Error) -> bool {
-    false
 }

 #[tracing::instrument(level = "debug", skip_all)]
 pub async fn remove(path: impl AsRef<Path>) -> io::Result<()> {
-    let path = path.as_ref().to_path_buf();
-    native_io(move || remove_std(path)).await
+    // Try remove_file first; fall back to remove_dir if it's a directory
+    let file_path = path.as_ref().to_path_buf();
+    match native_io(move || std::fs::remove_file(file_path)).await {
+        Ok(()) => Ok(()),
+        Err(e) if is_dir_error(&e) => {
+            let path = path.as_ref().to_path_buf();
+            native_io(move || std::fs::remove_dir(path)).await
+        }
+        Err(e) => Err(e),
+    }
 }

 #[allow(dead_code, reason = "asserted by this file's tests (backlog#1823)")]
 pub async fn remove_all(path: impl AsRef<Path>) -> io::Result<()> {
    // Try remove_file first; fall back to remove_dir_all if it's a directory
-    let path = path.as_ref().to_path_buf();
-    native_io(move || match std::fs::remove_file(&path) {
+    let file_path = path.as_ref().to_path_buf();
+    match native_io(move || std::fs::remove_file(file_path)).await {
        Ok(()) => Ok(()),
-        Err(e) if is_dir_error(&path, &e) => std::fs::remove_dir_all(&path),
+        Err(e) if is_dir_error(&e) => {
+            let path = path.as_ref().to_path_buf();
+            native_io(move || std::fs::remove_dir_all(path)).await
+        }
        Err(e) => Err(e),
-    })
-    .await
+    }
 }

 #[tracing::instrument(level = "debug", skip_all)]
 pub fn remove_std(path: impl AsRef<Path>) -> io::Result<()> {
    // Try remove_file first; fall back to remove_dir if it's a directory
-    let path = path.as_ref();
-    match std::fs::remove_file(path) {
+    match std::fs::remove_file(path.as_ref()) {
        Ok(()) => Ok(()),
-        Err(e) if is_dir_error(path, &e) => std::fs::remove_dir(path),
+        Err(e) if is_dir_error(&e) => std::fs::remove_dir(path.as_ref()),
        Err(e) => Err(e),
    }
 }
@@ -3012,7 +3012,7 @@ fn should_fail_local_inline_rollback_hardlink(dst_path: &Path) -> bool {
        .remove(dst_path)
 }

-#[cfg(test)]
+#[cfg(all(test, target_os = "linux"))]
 pub(super) fn remove_dst_base_before_commit_std(
    dst_path: &str,
    guard: os::RenameCommitGuard,
@@ -3290,7 +3290,7 @@ fn should_fail_local_inline_rollback_hardlink(_dst_path: &Path) -> bool {
    false
 }

-#[cfg(not(test))]
+#[cfg(all(not(test), target_os = "linux"))]
 pub(super) fn remove_dst_base_before_commit_std(
    _dst_path: &str,
    guard: os::RenameCommitGuard,
@@ -5167,6 +5167,7 @@ impl LocalIoBackend for StdBackend {
                    Arc::new(FdCacheEntry {
                        file: Arc::new(file),
                        len: metadata_len,
+                        invalidated: AtomicBool::new(false),
                    })
                });
                #[cfg(not(target_os = "linux"))]
@@ -5607,6 +5608,9 @@ const FD_CACHE_CAPACITY: u64 = 512;
 #[cfg(any(target_os = "linux", all(unix, test)))]
 const FD_CACHE_TTL: std::time::Duration = std::time::Duration::from_secs(5);

+#[cfg(any(target_os = "linux", all(unix, test)))]
+const FD_CACHE_MAINTENANCE_INTERVAL: std::time::Duration = std::time::Duration::from_secs(1);
+
 #[cfg(target_os = "linux")]
 fn is_io_uring_fd_cache_enabled() -> bool {
    rustfs_utils::get_env_bool(ENV_RUSTFS_IO_URING_FD_CACHE, DEFAULT_RUSTFS_IO_URING_FD_CACHE)
@@ -5702,6 +5706,8 @@ struct FdCacheEntry {
    file: Arc<std::fs::File>,
    /// File length captured together with the descriptor.
    len: u64,
+    #[cfg(any(target_os = "linux", all(unix, test)))]
+    invalidated: AtomicBool,
 }

 /// Per-disk cache of open descriptors for io_uring reads (backlog#1145).
@@ -5735,6 +5741,13 @@ struct FdCacheEntry {
 #[cfg(any(target_os = "linux", all(unix, test)))]
 struct FdCache {
    cache: moka::future::Cache<FdKey, Arc<FdCacheEntry>>,
+    // Moka's entry count can lag inserts. Only a cache that has never accepted
+    // an insert can safely skip the foreground invalidation scan.
+    has_inserted_entries: AtomicBool,
+    // The only sender belongs to this cache. Dropping it lets the worker finish
+    // maintenance and release its cache clone without retaining the backend.
+    maintenance_tx: tokio::sync::watch::Sender<()>,
+    maintenance_task: std::sync::OnceLock<tokio::task::JoinHandle<()>>,
    /// Bumped by every invalidation. A miss-path open snapshots this before it
    /// opens and refuses to insert if it moved, so an fd opened before a
    /// heal/delete commit can never be resurrected into the cache after the
@@ -5752,20 +5765,78 @@ impl FdCache {
    /// `new` (`FD_CACHE_TTL`); tests inject a short TTL to exercise the backstop
    /// eviction (rustfs/backlog#1180) without a multi-second wait.
    fn with_ttl(ttl: std::time::Duration) -> Self {
+        let (maintenance_tx, _) = tokio::sync::watch::channel(());
        Self {
            cache: moka::future::Cache::builder()
                .max_capacity(FD_CACHE_CAPACITY)
                .time_to_live(ttl)
-                // Required for `invalidate_entries_if`; without it that call
-                // fails at runtime instead of dropping stale descriptors.
-                .support_invalidation_closures()
                .build(),
+            has_inserted_entries: AtomicBool::new(false),
+            maintenance_tx,
+            maintenance_task: std::sync::OnceLock::new(),
            generation: std::sync::atomic::AtomicU64::new(0),
        }
    }

+    fn schedule_maintenance(&self) {
+        self.maintenance_task.get_or_init(|| {
+            let cache = self.cache.clone();
+            let mut notifications = self.maintenance_tx.subscribe();
+            // Invalidation runs on the disk API's Tokio runtime. Starting lazily
+            // keeps construction runtime-independent and idle caches task-free.
+            tokio::spawn(async move {
+                let mut interval = tokio::time::interval_at(
+                    tokio::time::Instant::now() + FD_CACHE_MAINTENANCE_INTERVAL,
+                    FD_CACHE_MAINTENANCE_INTERVAL,
+                );
+                interval.set_missed_tick_behavior(tokio::time::MissedTickBehavior::Skip);
+                loop {
+                    let stopping = tokio::select! {
+                        changed = notifications.changed() => changed.is_err(),
+                        _ = interval.tick() => false,
+                    };
+                    if stopping {
+                        cache.invalidate_all();
+                    }
+                    cache.run_pending_tasks().await;
+                    if stopping {
+                        break;
+                    }
+                    if cache.entry_count() != 0 {
+                        for (key, entry) in cache.iter() {
+                            if entry.invalidated.load(Ordering::Acquire) {
+                                // Concurrent replacement may evict a fresh fd;
+                                // a later read safely reopens it.
+                                cache.invalidate(key.as_ref()).await;
+                            }
+                        }
+                        cache.run_pending_tasks().await;
+                    }
+                }
+            })
+        });
+        self.maintenance_tx.send_replace(());
+    }
+
+    fn invalidate_matching(&self, matches: impl Fn(&FdKey) -> bool) {
+        if !self.has_inserted_entries.load(Ordering::Acquire) {
+            return;
+        }
+        // Entries already present are visited unless concurrently removed. The
+        // generation fence rejects an older open inserted after this scan.
+        for (key, entry) in self.cache.iter() {
+            if matches(&key) {
+                entry.invalidated.store(true, Ordering::Release);
+            }
+        }
+        self.schedule_maintenance();
+    }
+
    async fn get(&self, key: &FdKey) -> Option<Arc<FdCacheEntry>> {
-        self.cache.get(key).await
+        self.cache
+            .get(key)
+            .await
+            .filter(|entry| !entry.invalidated.load(Ordering::Acquire))
    }

    /// Current invalidation generation, snapshotted by the miss path before it
@@ -5783,6 +5854,7 @@ impl FdCache {
        if self.generation.load(Ordering::Acquire) != gen_at_open {
            return;
        }
+        self.has_inserted_entries.store(true, Ordering::Release);
        self.cache.insert(key.clone(), entry).await;
        if self.generation.load(Ordering::Acquire) != gen_at_open {
            self.cache.invalidate(&key).await;
@@ -5809,38 +5881,24 @@ impl FdCache {
    /// Drop every descriptor for `volume` whose path is `prefix` or lies under it.
    /// A component-boundary check keeps `a/b` from invalidating `a/bc`.
    ///
-    /// moka applies the predicate to entries inserted at or before this call and
-    /// guarantees a later `get` never returns one of them, so the invalidation is
-    /// effective immediately even though the removal itself is deferred. Reserved
-    /// for paths whose exact keys are unknown (`delete`); every predicate is
-    /// re-evaluated by subsequent `get`s, so registering one per write would tax
-    /// the read path.
+    /// Mark matching entries immediately so reads reject them while maintenance
+    /// removes them. The scan borrows its paths and retains no invalidation
+    /// closures when the cache is empty or idle.
    fn invalidate_under(&self, volume: &str, prefix: &str) {
-        // Bump before registering the predicate so a concurrent miss-path insert
+        // Bump before scanning the cache so a concurrent miss-path insert
        // that snapshotted the old generation is refused (rustfs/backlog#1176).
        self.generation.fetch_add(1, Ordering::AcqRel);
-        let volume = volume.to_owned();
-        let prefix = prefix.trim_end_matches('/').to_owned();
-        let matches = move |k: &FdKey, _: &Arc<FdCacheEntry>| {
-            k.volume == volume && (k.path == prefix || k.path.strip_prefix(&prefix).is_some_and(|r| r.starts_with('/')))
-        };
-        if self.cache.invalidate_entries_if(matches).is_err() {
-            // Closure support is enabled at construction, so this is unreachable.
-            // If it ever changes, over-invalidate rather than serve a stale
-            // descriptor: correctness first, the cache refills on the next read.
-            self.cache.invalidate_all();
-        }
+        let prefix = prefix.trim_end_matches('/');
+        self.invalidate_matching(|k| {
+            k.volume == volume && (k.path == prefix || k.path.strip_prefix(prefix).is_some_and(|r| r.starts_with('/')))
+        });
    }

    /// Drop every descriptor for `volume` — the whole bucket is gone
    /// (rustfs/backlog#1177).
    fn invalidate_volume(&self, volume: &str) {
        self.generation.fetch_add(1, Ordering::AcqRel);
-        let volume = volume.to_owned();
-        let matches = move |k: &FdKey, _: &Arc<FdCacheEntry>| k.volume == volume;
-        if self.cache.invalidate_entries_if(matches).is_err() {
-            self.cache.invalidate_all();
-        }
+        self.invalidate_matching(|k| k.volume == volume);
    }

    /// Drop ALL descriptors — this disk instance is being retired
@@ -5848,6 +5906,7 @@ impl FdCache {
    fn clear(&self) {
        self.generation.fetch_add(1, Ordering::AcqRel);
        self.cache.invalidate_all();
+        self.schedule_maintenance();
    }

    /// Unconditional insert. Production reads go through `insert_if_fresh` (the
@@ -5856,7 +5915,17 @@ impl FdCache {
    #[cfg(test)]
    async fn insert(&self, key: FdKey, file: Arc<std::fs::File>) {
        let len = file.metadata().map(|metadata| metadata.len()).unwrap_or_default();
-        self.cache.insert(key, Arc::new(FdCacheEntry { file, len })).await;
+        self.has_inserted_entries.store(true, Ordering::Release);
+        self.cache
+            .insert(
+                key,
+                Arc::new(FdCacheEntry {
+                    file,
+                    len,
+                    invalidated: AtomicBool::new(false),
+                }),
+            )
+            .await;
    }

    #[cfg(test)]
@@ -6455,6 +6524,7 @@ impl UringBackend {
                let file = Arc::new(FdCacheEntry {
                    file: Arc::new(file),
                    len,
+                    invalidated: AtomicBool::new(false),
                });
                if let (Some((cache, key)), Some(gen_at_open)) = (cache_entry, gen_at_open) {
                    cache.insert_if_fresh(key, Arc::clone(&file), gen_at_open).await;
@@ -6627,6 +6697,7 @@ impl UringBackend {
                let entry = Arc::new(FdCacheEntry {
                    file: Arc::new(file),
                    len,
+                    invalidated: AtomicBool::new(false),
                });
                if let (Some(cache), Some(gen_at_open)) = (cache, gen_at_open) {
                    cache.insert_if_fresh(key.clone(), Arc::clone(&entry), gen_at_open).await;
@@ -34128,11 +34199,217 @@ mod test {
        });
    }

+    #[cfg(unix)]
+    #[tokio::test]
+    async fn fd_cache_invalidation_releases_descriptors_without_further_operations() {
+        let cache = FdCache::new();
+        for round in 0..4 {
+            let file = Arc::new(tempfile::tempfile().expect("create a descriptor to invalidate"));
+            let released = Arc::downgrade(&file);
+            let parts = if round == 1 || round == 2 { FD_CACHE_CAPACITY } else { 1 };
+            for part in 0..parts {
+                cache
+                    .insert(
+                        FdKey {
+                            volume: "bucket".into(),
+                            path: format!("object-{round}/part.{part}"),
+                            direct: false,
+                        },
+                        file.clone(),
+                    )
+                    .await;
+            }
+            drop(file);
+            if round % 2 == 0 {
+                cache.invalidate_under("bucket", &format!("object-{round}"));
+            } else {
+                cache.invalidate_volume("bucket");
+            }
+            tokio::time::timeout(std::time::Duration::from_secs(10), async {
+                while released.upgrade().is_some() {
+                    tokio::task::yield_now().await;
+                }
+            })
+            .await
+            .expect("invalidation must release descriptors without another cache operation");
+        }
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn fd_cache_invalidation_rejects_entries_before_maintenance() {
+        let cache = FdCache::new();
+        let key = FdKey {
+            volume: "bucket".into(),
+            path: "object/part.1".into(),
+            direct: false,
+        };
+        cache
+            .insert(key.clone(), Arc::new(tempfile::tempfile().expect("create a cached descriptor")))
+            .await;
+        assert!(cache.get(&key).await.is_some());
+        cache.invalidate_under("bucket", "object");
+
+        // No yield has allowed the maintenance task to remove the entry yet.
+        assert!(cache.cache.contains_key(&key));
+        assert!(cache.get(&key).await.is_none(), "a marked descriptor must never be served");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn fd_cache_idle_maintenance_releases_expired_descriptors() {
+        let cache = FdCache::with_ttl(std::time::Duration::from_millis(50));
+        let file = Arc::new(tempfile::tempfile().expect("create a descriptor with a short TTL"));
+        let released = Arc::downgrade(&file);
+        let key = FdKey {
+            volume: "retained".into(),
+            path: "part.1".into(),
+            direct: false,
+        };
+        cache.insert(key.clone(), file).await;
+        cache.invalidate_volume("other");
+        assert!(cache.get(&key).await.is_some(), "invalidation must preserve other volumes");
+
+        tokio::time::timeout(std::time::Duration::from_secs(10), async {
+            while released.upgrade().is_some() {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("periodic maintenance must release expired descriptors while idle");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn fd_cache_maintenance_exits_when_cache_is_dropped() {
+        let mut cache = FdCache::new();
+        let file = Arc::new(tempfile::tempfile().expect("create a descriptor to retire"));
+        let released = Arc::downgrade(&file);
+        cache
+            .insert(
+                FdKey {
+                    volume: "bucket".into(),
+                    path: "part.1".into(),
+                    direct: false,
+                },
+                file,
+            )
+            .await;
+        cache.invalidate_volume("bucket");
+        let maintenance = cache.maintenance_task.take().expect("invalidation starts maintenance");
+        drop(cache);
+        tokio::time::timeout(std::time::Duration::from_secs(10), maintenance)
+            .await
+            .expect("dropping the cache must stop its maintenance task")
+            .expect("maintenance must complete without a panic");
+        assert!(released.upgrade().is_none(), "shutdown must release cached descriptors");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn fd_cache_maintenance_survives_request_cancellation() {
+        let cache = FdCache::new();
+        let file = Arc::new(tempfile::tempfile().expect("create a descriptor to invalidate"));
+        let released = Arc::downgrade(&file);
+        cache
+            .insert(
+                FdKey {
+                    volume: "bucket".into(),
+                    path: "part.1".into(),
+                    direct: false,
+                },
+                file,
+            )
+            .await;
+        let mut request = Box::pin(async {
+            cache.invalidate_volume("bucket");
+            std::future::pending::<()>().await;
+        });
+        let mut context = std::task::Context::from_waker(std::task::Waker::noop());
+        assert!(std::future::Future::poll(request.as_mut(), &mut context).is_pending());
+        drop(request);
+
+        tokio::time::timeout(std::time::Duration::from_secs(10), async {
+            while released.upgrade().is_some() {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("maintenance must continue after the invalidating request is cancelled");
+    }
+
+    #[cfg(unix)]
+    #[test]
+    fn fd_cache_construction_does_not_require_a_runtime() {
+        let cache = FdCache::new();
+        assert!(cache.maintenance_task.get().is_none());
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn fd_cache_invalidation_fences_an_open_before_the_first_insert() {
+        for mode in 0..2 {
+            let cache = FdCache::new();
+            let key = FdKey {
+                volume: "bucket".into(),
+                path: "object/part.1".into(),
+                direct: false,
+            };
+            let opened = Arc::new(FdCacheEntry {
+                file: Arc::new(tempfile::tempfile().expect("open the descriptor before invalidation")),
+                len: 0,
+                invalidated: AtomicBool::new(false),
+            });
+            let generation = cache.generation();
+            match mode {
+                0 => cache.invalidate_under("bucket", "object"),
+                _ => cache.invalidate_volume("bucket"),
+            }
+            cache.insert_if_fresh(key.clone(), opened, generation).await;
+            assert!(cache.get(&key).await.is_none(), "mode {mode} must reject a pre-invalidation open");
+
+            let fresh = Arc::new(FdCacheEntry {
+                file: Arc::new(tempfile::tempfile().expect("open a fresh descriptor after invalidation")),
+                len: 0,
+                invalidated: AtomicBool::new(false),
+            });
+            cache.insert_if_fresh(key.clone(), fresh.clone(), cache.generation()).await;
+            let cached = cache.get(&key).await.expect("a post-invalidation open must remain cacheable");
+            assert!(Arc::ptr_eq(&cached, &fresh));
+        }
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn fd_cache_clear_releases_descriptors_without_further_reads() {
+        let cache = FdCache::new();
+        let file = Arc::new(tempfile::tempfile().expect("create a descriptor to retire"));
+        let released = Arc::downgrade(&file);
+        cache
+            .insert(
+                FdKey {
+                    volume: "bucket".into(),
+                    path: "part.1".into(),
+                    direct: false,
+                },
+                file,
+            )
+            .await;
+        cache.clear();
+        tokio::time::timeout(std::time::Duration::from_secs(10), async {
+            while released.upgrade().is_some() {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("clear must release retired descriptors without another cache operation");
+    }
+
    /// Prefix invalidation must respect component boundaries: dropping the
    /// descriptors for object `a/b` must not also drop those for `a/bc`.
-    /// moka applies the predicate lazily, so this also pins the guarantee we rely
-    /// on — a `get` after `invalidate_under` never returns a matched entry.
-    #[cfg(target_os = "linux")]
+    /// A `get` after `invalidate_under` must reject matched entries while
+    /// background maintenance removes them.
+    #[cfg(unix)]
    #[tokio::test(flavor = "multi_thread")]
    async fn fd_cache_prefix_invalidation_respects_component_boundary() {
        use std::fs::File;
@@ -34368,8 +34645,32 @@ mod test {
            }
        }

-        // Invalidating by the object prefix drops the cached descriptor.
+        let key = FdKey {
+            volume: volume.into(),
+            path: object.into(),
+            direct: false,
+        };
+        let released = {
+            let entry = cache
+                .get(&key)
+                .await
+                .expect("repeated reads must retain the cached descriptor");
+            Arc::downgrade(&entry.file)
+        };
+
+        // Reads reject invalidated entries immediately; maintenance releases the descriptor.
        backend.invalidate_cached_fds_under(volume, "obj/abc");
+        assert!(
+            cache.get(&key).await.is_none(),
+            "prefix invalidation must immediately reject the cached descriptor"
+        );
+        tokio::time::timeout(std::time::Duration::from_secs(10), async {
+            while released.upgrade().is_some() {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("prefix invalidation must release the descriptor without further cache operations");
        assert_eq!(cache.entry_count().await, 0, "prefix invalidation must drop the cached descriptor");
    }

@@ -1594,13 +1594,12 @@ impl LocalDisk {
                );
            }
            if let Some((src_data_path, dst_data_path)) = has_data_dir_path.as_ref() {
-                #[cfg(any(target_os = "linux", windows))]
+                #[cfg(target_os = "linux")]
                let payload_guard = if multipart_publication.is_some() {
                    Some(
                        os::multipart_payload_commit_guard(
                            &commit.directory_guard,
                            src_data_path.parent().ok_or(DiskError::InvalidPath)?,
-                            commit.publication_root,
                            commit.mutation_lease.clone(),
                        )
                        .await
@@ -1609,9 +1608,9 @@ impl LocalDisk {
                } else {
                    None
                };
-                #[cfg(any(target_os = "linux", windows))]
+                #[cfg(target_os = "linux")]
                let directory_guard = payload_guard.as_ref().unwrap_or(&commit.directory_guard);
-                #[cfg(not(any(target_os = "linux", windows)))]
+                #[cfg(not(target_os = "linux"))]
                let directory_guard = &commit.directory_guard;
                if let Err(err) = os::rename_all_with_commit_guard(
                    src_data_path,
@@ -1527,7 +1527,7 @@ impl LocalDisk {
                    }
                    Err(error) => return Err(error),
                };
-                let directory_pin = os::pin_multipart_source_directory(&directory, base, root)?;
+                let _directory = os::pin_multipart_source_directory(&directory, base, root)?;
                if !parts.is_empty() {
                    let anchor_path = directory.join(RETIREMENT_ANCHOR);
                    let bytes = os::read_multipart_source_metadata(&anchor_path, base, root)?;
@@ -1560,13 +1560,7 @@ impl LocalDisk {
                }
                // An anchor-less empty DD is only rmdir'ed. A new child makes
                // this fail closed; no recursive removal can consume its bytes.
-                // A Windows pin refuses DELETE sharing, so the leaf pin is
-                // released first; the object pin still holds the parent.
-                #[cfg(windows)]
-                drop(directory_pin);
                std::fs::remove_dir(&directory)?;
-                #[cfg(not(windows))]
-                drop(directory_pin);
                if sync {
                    os::fsync_dir_std(&object_path)?;
                }
@@ -1942,6 +1936,7 @@ impl NativeMultipartPublication {
        *self.validated_prepared.lock() = Some(validated);
        Ok(())
    }
+    #[cfg(target_os = "linux")]
    pub(super) fn prepare_and_install_head_on_thread(
        &self,
        new_metadata: &[u8],
@@ -1677,8 +1677,230 @@ const MAX_GLOBAL_FILE_SYNCS: usize = 512;
 #[cfg(test)]
 const TEST_GLOBAL_FILE_SYNCS: usize = 64;

+const PATH_LOCK_SHARDS: usize = 64;
+const PATH_LOCK_INITIAL_GC_THRESHOLD: usize = 64;
+
+#[cfg(unix)]
+#[derive(Clone, Copy, Default)]
+struct Xxh3SliceHasher(u64);
+
+#[cfg(unix)]
+impl std::hash::Hasher for Xxh3SliceHasher {
+    #[inline]
+    fn write(&mut self, bytes: &[u8]) {
+        self.0 = xxhash_rust::xxh3::xxh3_64(bytes);
+    }
+
+    #[inline]
+    fn write_usize(&mut self, _i: usize) {}
+
+    #[inline]
+    fn finish(&self) -> u64 {
+        self.0
+    }
+}
+
+#[cfg(unix)]
+type Xxh3SliceBuildHasher = std::hash::BuildHasherDefault<Xxh3SliceHasher>;
+
+#[cfg(unix)]
+#[inline]
+fn is_canonical_unix_key_bytes(bytes: &[u8]) -> bool {
+    if bytes.starts_with(b"./") {
+        return !bytes[1..].windows(2).any(|w| w == b"//") && !bytes[1..].windows(3).any(|w| w == b"/./");
+    }
+    const SLASH_MASK: u64 = 0x2f2f_2f2f_2f2f_2f2f;
+    const LOW_BITS: u64 = 0x0101_0101_0101_0101;
+    const HIGH_BITS: u64 = 0x8080_8080_8080_8080;
+
+    let mut i = 0usize;
+    let len = bytes.len();
+    let mut prev_s = 0u64;
+    while i + 8 <= len {
+        let mut buf = [0u8; 8];
+        buf.copy_from_slice(&bytes[i..i + 8]);
+        let w = u64::from_le_bytes(buf);
+        let x = w ^ SLASH_MASK;
+        let s = (x.wrapping_sub(LOW_BITS)) & !x & HIGH_BITS;
+        let s_prev1 = (s << 8) | (prev_s >> 56);
+        let s_prev2 = (s << 16) | (prev_s >> 48);
+        if (s & (s_prev1 | s_prev2)) != 0 {
+            let start = i.saturating_sub(2);
+            let window = &bytes[start..i + 8];
+            if window.windows(2).any(|pair| pair == b"//") || window.windows(3).any(|triple| triple == b"/./") {
+                return false;
+            }
+        }
+        prev_s = s;
+        i += 8;
+    }
+    let tail_start = i.saturating_sub(2);
+    let tail = &bytes[tail_start..];
+    !tail.windows(2).any(|pair| pair == b"//") && !tail.windows(3).any(|triple| triple == b"/./")
+}
+
+#[cfg(unix)]
+#[cold]
+fn canonical_unix_path_key_slow(path: &Path) -> std::borrow::Cow<'static, [u8]> {
+    use std::os::unix::ffi::OsStrExt;
+    let mut out = Vec::with_capacity(path.as_os_str().len());
+    for comp in path.components() {
+        match comp {
+            Component::RootDir => {
+                out.clear();
+                out.push(b'/');
+            }
+            Component::CurDir => {
+                if out.is_empty() {
+                    out.push(b'.');
+                }
+            }
+            Component::ParentDir => {
+                if out.len() > 1 || (out.len() == 1 && out[0] != b'/') {
+                    out.push(b'/');
+                }
+                out.extend_from_slice(b"..");
+            }
+            Component::Normal(part) => {
+                if out.len() > 1 || (out.len() == 1 && out[0] != b'/') {
+                    out.push(b'/');
+                }
+                out.extend_from_slice(part.as_bytes());
+            }
+            Component::Prefix(_) => {}
+        }
+    }
+    std::borrow::Cow::Owned(out)
+}
+
+#[cfg(unix)]
+#[inline]
+fn canonical_unix_path_key(path: &Path) -> std::borrow::Cow<'_, [u8]> {
+    use std::os::unix::ffi::OsStrExt;
+    let mut bytes = path.as_os_str().as_bytes();
+    while bytes.len() > 1 {
+        if bytes[bytes.len() - 1] == b'/' {
+            bytes = &bytes[..bytes.len() - 1];
+        } else if bytes.len() >= 2 && bytes[bytes.len() - 2] == b'/' && bytes[bytes.len() - 1] == b'.' {
+            bytes = if bytes.len() == 2 { b"/" } else { &bytes[..bytes.len() - 2] };
+        } else {
+            break;
+        }
+    }
+    if is_canonical_unix_key_bytes(bytes) {
+        return std::borrow::Cow::Borrowed(bytes);
+    }
+    canonical_unix_path_key_slow(path)
+}
+
+#[cfg(unix)]
+#[inline]
+fn unix_canonical_key_starts_with(path_key: &[u8], root_key: &[u8]) -> bool {
+    if root_key.is_empty() {
+        return true;
+    }
+    if root_key == b"/" {
+        return path_key.starts_with(b"/");
+    }
+    path_key.starts_with(root_key) && (path_key.len() == root_key.len() || path_key.get(root_key.len()) == Some(&b'/'))
+}
+
+struct PathLockShard<L> {
+    #[cfg(unix)]
+    entries: HashMap<Box<[u8]>, Weak<L>, Xxh3SliceBuildHasher>,
+    #[cfg(not(unix))]
+    entries: HashMap<PathBuf, Weak<L>>,
+    next_gc_threshold: usize,
+}
+
+impl<L> Default for PathLockShard<L> {
+    fn default() -> Self {
+        Self {
+            entries: HashMap::default(),
+            next_gc_threshold: PATH_LOCK_INITIAL_GC_THRESHOLD,
+        }
+    }
+}
+
+struct ShardedPathLockRegistry<L> {
+    shards: [Mutex<PathLockShard<L>>; PATH_LOCK_SHARDS],
+}
+
+impl<L> ShardedPathLockRegistry<L> {
+    fn new() -> Self {
+        Self {
+            shards: std::array::from_fn(|_| Mutex::new(PathLockShard::default())),
+        }
+    }
+
+    #[cfg(unix)]
+    #[inline]
+    fn get_or_create_by_key(&self, key: &[u8], init: impl FnOnce() -> L) -> Arc<L> {
+        let hash = xxhash_rust::xxh3::xxh3_64(key);
+        let shard_idx = ((hash >> 32) as usize) & (PATH_LOCK_SHARDS - 1);
+        let mut shard = self.shards[shard_idx].lock();
+        if let Some(existing) = shard.entries.get(key).and_then(Weak::upgrade) {
+            return existing;
+        }
+        let created = Arc::new(init());
+        let weak = Arc::downgrade(&created);
+        if let Some(slot) = shard.entries.get_mut(key) {
+            *slot = weak;
+            return created;
+        }
+        if shard.entries.len() >= shard.next_gc_threshold {
+            shard.entries.retain(|_, entry| entry.strong_count() > 0);
+            shard.next_gc_threshold = shard
+                .entries
+                .len()
+                .saturating_add(1)
+                .saturating_mul(2)
+                .max(PATH_LOCK_INITIAL_GC_THRESHOLD);
+        }
+        shard.entries.insert(key.into(), weak);
+        created
+    }
+
+    #[inline]
+    fn get_or_create(&self, path: &Path, init: impl FnOnce() -> L) -> Arc<L> {
+        #[cfg(unix)]
+        {
+            let key = canonical_unix_path_key(path);
+            self.get_or_create_by_key(key.as_ref(), init)
+        }
+        #[cfg(not(unix))]
+        {
+            use std::hash::{Hash, Hasher};
+            let mut hasher = std::collections::hash_map::DefaultHasher::new();
+            path.hash(&mut hasher);
+            let shard_idx = ((hasher.finish() >> 32) as usize) & (PATH_LOCK_SHARDS - 1);
+            let mut shard = self.shards[shard_idx].lock();
+            if let Some(existing) = shard.entries.get(path).and_then(Weak::upgrade) {
+                return existing;
+            }
+            let created = Arc::new(init());
+            let weak = Arc::downgrade(&created);
+            if let Some(slot) = shard.entries.get_mut(path) {
+                *slot = weak;
+                return created;
+            }
+            if shard.entries.len() >= shard.next_gc_threshold {
+                shard.entries.retain(|_, entry| entry.strong_count() > 0);
+                shard.next_gc_threshold = shard
+                    .entries
+                    .len()
+                    .saturating_add(1)
+                    .saturating_mul(2)
+                    .max(PATH_LOCK_INITIAL_GC_THRESHOLD);
+            }
+            shard.entries.insert(path.to_path_buf(), weak);
+            created
+        }
+    }
+}
+
 static FILE_SYNC_PERMITS: LazyLock<Semaphore> = LazyLock::new(|| Semaphore::new(global_file_sync_limit()));
-static DISK_FILE_SYNC_LIMITERS: LazyLock<Mutex<HashMap<PathBuf, Weak<Semaphore>>>> = LazyLock::new(|| Mutex::new(HashMap::new()));
+static DISK_FILE_SYNC_LIMITERS: LazyLock<ShardedPathLockRegistry<Semaphore>> = LazyLock::new(ShardedPathLockRegistry::new);

 /// Dedicated tokio runtime for fsync/fdatasync blocking operations. When
 /// configured with >1 threads, isolates device-bound fsync from the main
@@ -1742,12 +1964,10 @@ fn fsync_spawn_blocking<T: Send + 'static>(f: impl FnOnce() -> T + Send + 'stati
        f()
    })
 }
-static DISK_VOLUME_MUTATION_LOCKS: LazyLock<Mutex<HashMap<PathBuf, Weak<RwLock<()>>>>> =
-    LazyLock::new(|| Mutex::new(HashMap::new()));
+static DISK_VOLUME_MUTATION_LOCKS: LazyLock<ShardedPathLockRegistry<RwLock<()>>> = LazyLock::new(ShardedPathLockRegistry::new);
 type NamespaceMutationLock = AsyncMutex<()>;
-type NamespaceMutationLockRegistry = HashMap<PathBuf, Weak<NamespaceMutationLock>>;
-static DISK_NAMESPACE_MUTATION_LOCKS: LazyLock<Mutex<NamespaceMutationLockRegistry>> =
-    LazyLock::new(|| Mutex::new(HashMap::new()));
+static DISK_NAMESPACE_MUTATION_LOCKS: LazyLock<ShardedPathLockRegistry<NamespaceMutationLock>> =
+    LazyLock::new(ShardedPathLockRegistry::new);

 fn default_global_file_sync_limit(cpu_count: usize, max_blocking_threads: usize) -> usize {
    let cpu_scaled = cpu_count
@@ -1771,15 +1991,7 @@ fn global_file_sync_limit() -> usize {

 /// Reuse a disk's limiter across reconnects while detached sync calls still hold it.
 pub(crate) fn disk_file_sync_limiter(root: &Path) -> Arc<Semaphore> {
-    let mut limiters = DISK_FILE_SYNC_LIMITERS.lock();
-    limiters.retain(|_, limiter| limiter.strong_count() > 0);
-    if let Some(limiter) = limiters.get(root).and_then(Weak::upgrade) {
-        return limiter;
-    }
-
-    let limiter = Arc::new(Semaphore::new(MAX_PARALLEL_FILE_SYNCS));
-    limiters.insert(root.to_path_buf(), Arc::downgrade(&limiter));
-    limiter
+    DISK_FILE_SYNC_LIMITERS.get_or_create(root, || Semaphore::new(MAX_PARALLEL_FILE_SYNCS))
 }

 // Keep warm draining workers per disk/lane and spin briefly in user space
@@ -2147,45 +2359,131 @@ pub(crate) fn enter_blocking_test_hook() -> BlockingTestHookGuard {
    BlockingTestHookGuard { queue }
 }

-static DISK_BLOCKING_QUEUES: LazyLock<Mutex<HashMap<PathBuf, Weak<DiskBlockingQueue>>>> =
-    LazyLock::new(|| Mutex::new(HashMap::new()));
+struct RegisteredDiskBlockingQueue {
+    #[cfg(not(unix))]
+    root: PathBuf,
+    #[cfg(unix)]
+    canonical_root: Box<[u8]>,
+    queue: Weak<DiskBlockingQueue>,
+}
+
+static DISK_BLOCKING_QUEUES: LazyLock<parking_lot::RwLock<Vec<RegisteredDiskBlockingQueue>>> =
+    LazyLock::new(|| parking_lot::RwLock::new(Vec::new()));
 static FALLBACK_BLOCKING_QUEUES: LazyLock<[Arc<DiskBlockingQueue>; FALLBACK_NAMESPACE_QUEUE_SHARDS]> =
    LazyLock::new(|| std::array::from_fn(|_| DiskBlockingQueue::new()));

 pub(crate) fn register_disk_blocking_queue(root: &Path, io_root: &Path) -> Arc<DiskBlockingQueue> {
-    let mut queues = DISK_BLOCKING_QUEUES.lock();
-    queues.retain(|_, q| q.strong_count() > 0);
-    let queue = queues.get(root).and_then(Weak::upgrade).unwrap_or_else(|| {
-        let queue = DiskBlockingQueue::new();
-        queues.insert(root.to_path_buf(), Arc::downgrade(&queue));
-        queue
-    });
-    if io_root != root {
-        queues.insert(io_root.to_path_buf(), Arc::downgrade(&queue));
+    #[cfg(unix)]
+    let root_key = canonical_unix_path_key(root);
+    #[cfg(unix)]
+    let io_root_key = canonical_unix_path_key(io_root);
+
+    let mut queues = DISK_BLOCKING_QUEUES.write();
+    queues.retain(|entry| entry.queue.strong_count() > 0);
+    let queue = queues
+        .iter()
+        .find_map(|entry| {
+            #[cfg(unix)]
+            let matches = entry.canonical_root.as_ref() == root_key.as_ref();
+            #[cfg(not(unix))]
+            let matches = entry.root == root;
+            matches.then(|| entry.queue.upgrade()).flatten()
+        })
+        .unwrap_or_else(|| {
+            let queue = DiskBlockingQueue::new();
+            queues.push(RegisteredDiskBlockingQueue {
+                #[cfg(not(unix))]
+                root: root.to_path_buf(),
+                #[cfg(unix)]
+                canonical_root: root_key.as_ref().into(),
+                queue: Arc::downgrade(&queue),
+            });
+            queue
+        });
+    #[cfg(unix)]
+    let io_matches_root = io_root_key.as_ref() == root_key.as_ref();
+    #[cfg(not(unix))]
+    let io_matches_root = io_root == root;
+    if !io_matches_root {
+        let weak = Arc::downgrade(&queue);
+        if let Some(existing) = queues.iter_mut().find(|entry| {
+            #[cfg(unix)]
+            {
+                entry.canonical_root.as_ref() == io_root_key.as_ref()
+            }
+            #[cfg(not(unix))]
+            {
+                entry.root == io_root
+            }
+        }) {
+            existing.queue = weak;
+        } else {
+            queues.push(RegisteredDiskBlockingQueue {
+                #[cfg(not(unix))]
+                root: io_root.to_path_buf(),
+                #[cfg(unix)]
+                canonical_root: io_root_key.as_ref().into(),
+                queue: weak,
+            });
+        }
    }
    queue
 }

 fn disk_blocking_queue_for_root(root: &Path) -> Arc<DiskBlockingQueue> {
-    register_disk_blocking_queue(root, root)
-}
-
-fn disk_blocking_queue_for_path(path: &Path) -> Arc<DiskBlockingQueue> {
+    #[cfg(unix)]
+    let root_key = canonical_unix_path_key(root);
    {
-        let queues = DISK_BLOCKING_QUEUES.lock();
-        for (root, weak) in queues.iter() {
-            if path.starts_with(root)
-                && let Some(queue) = weak.upgrade()
-            {
+        let queues = DISK_BLOCKING_QUEUES.read();
+        for entry in queues.iter() {
+            #[cfg(unix)]
+            let matches = entry.canonical_root.as_ref() == root_key.as_ref();
+            #[cfg(not(unix))]
+            let matches = entry.root == root;
+            if matches && let Some(queue) = entry.queue.upgrade() {
                return queue;
            }
        }
    }
-    use std::hash::{Hash, Hasher};
-    let mut hasher = std::collections::hash_map::DefaultHasher::new();
-    path.hash(&mut hasher);
-    let idx = usize::try_from(hasher.finish() % (FALLBACK_NAMESPACE_QUEUE_SHARDS as u64)).unwrap_or(0);
-    Arc::clone(&FALLBACK_BLOCKING_QUEUES[idx])
+    register_disk_blocking_queue(root, root)
+}
+
+fn disk_blocking_queue_for_path(path: &Path) -> Arc<DiskBlockingQueue> {
+    #[cfg(unix)]
+    {
+        let path_key = canonical_unix_path_key(path);
+        {
+            let queues = DISK_BLOCKING_QUEUES.read();
+            for entry in queues.iter() {
+                if unix_canonical_key_starts_with(path_key.as_ref(), entry.canonical_root.as_ref())
+                    && let Some(queue) = entry.queue.upgrade()
+                {
+                    return queue;
+                }
+            }
+        }
+        let hash = xxhash_rust::xxh3::xxh3_64(path_key.as_ref());
+        let idx = usize::try_from(hash % (FALLBACK_NAMESPACE_QUEUE_SHARDS as u64)).unwrap_or(0);
+        Arc::clone(&FALLBACK_BLOCKING_QUEUES[idx])
+    }
+    #[cfg(not(unix))]
+    {
+        {
+            let queues = DISK_BLOCKING_QUEUES.read();
+            for entry in queues.iter() {
+                if path.starts_with(&entry.root)
+                    && let Some(queue) = entry.queue.upgrade()
+                {
+                    return queue;
+                }
+            }
+        }
+        use std::hash::{Hash, Hasher};
+        let mut hasher = std::collections::hash_map::DefaultHasher::new();
+        path.hash(&mut hasher);
+        let idx = usize::try_from(hasher.finish() % (FALLBACK_NAMESPACE_QUEUE_SHARDS as u64)).unwrap_or(0);
+        Arc::clone(&FALLBACK_BLOCKING_QUEUES[idx])
+    }
 }

 /// Serialize a bucket's local metadata commits with physical bucket removal.
@@ -2194,32 +2492,46 @@ fn disk_blocking_queue_for_path(path: &Path) -> Arc<DiskBlockingQueue> {
 /// [`LocalDisk`](super::local::LocalDisk) instances share the same lock while
 /// disconnected disks do not keep the registry alive.
 pub(crate) fn disk_volume_mutation_lock(root: &Path, volume: &str) -> Arc<RwLock<()>> {
-    let key = root.join(volume);
-    let mut locks = DISK_VOLUME_MUTATION_LOCKS.lock();
-    locks.retain(|_, lock| lock.strong_count() > 0);
-    if let Some(lock) = locks.get(&key).and_then(Weak::upgrade) {
-        return lock;
+    #[cfg(unix)]
+    {
+        let vol_bytes = volume.as_bytes();
+        if !vol_bytes.is_empty()
+            && vol_bytes[0] != b'/'
+            && vol_bytes != b"."
+            && !vol_bytes.starts_with(b"./")
+            && !vol_bytes.ends_with(b"/")
+            && !vol_bytes.ends_with(b"/.")
+            && is_canonical_unix_key_bytes(vol_bytes)
+            && let std::borrow::Cow::Borrowed(root_bytes) = canonical_unix_path_key(root)
+            && !root_bytes.is_empty()
+            && root_bytes != b"/"
+        {
+            let total_len = root_bytes.len() + 1 + vol_bytes.len();
+            if total_len <= 512 {
+                let mut stack_key = [0u8; 512];
+                stack_key[..root_bytes.len()].copy_from_slice(root_bytes);
+                stack_key[root_bytes.len()] = b'/';
+                stack_key[root_bytes.len() + 1..total_len].copy_from_slice(vol_bytes);
+                return DISK_VOLUME_MUTATION_LOCKS.get_or_create_by_key(&stack_key[..total_len], || RwLock::new(()));
+            }
+        }
    }
-
-    let lock = Arc::new(RwLock::new(()));
-    locks.insert(key, Arc::downgrade(&lock));
-    lock
+    let key = root.join(volume);
+    DISK_VOLUME_MUTATION_LOCKS.get_or_create(&key, || RwLock::new(()))
 }

 fn disk_namespace_mutation_lock(path: &Path) -> Arc<NamespaceMutationLock> {
-    let mut locks = DISK_NAMESPACE_MUTATION_LOCKS.lock();
-    locks.retain(|_, lock| lock.strong_count() > 0);
-    if let Some(lock) = locks.get(path).and_then(Weak::upgrade) {
-        return lock;
-    }
-
-    let lock = Arc::new(AsyncMutex::new(()));
-    locks.insert(path.to_path_buf(), Arc::downgrade(&lock));
-    lock
+    DISK_NAMESPACE_MUTATION_LOCKS.get_or_create(path, || AsyncMutex::new(()))
 }

-static DISK_METADATA_MUTATION_LOCKS: LazyLock<Mutex<NamespaceMutationLockRegistry>> =
-    LazyLock::new(|| Mutex::new(HashMap::new()));
+#[cfg(unix)]
+#[inline]
+fn disk_namespace_mutation_lock_by_key(key: &[u8]) -> Arc<NamespaceMutationLock> {
+    DISK_NAMESPACE_MUTATION_LOCKS.get_or_create_by_key(key, || AsyncMutex::new(()))
+}
+
+static DISK_METADATA_MUTATION_LOCKS: LazyLock<ShardedPathLockRegistry<NamespaceMutationLock>> =
+    LazyLock::new(ShardedPathLockRegistry::new);

 /// Serializes the complete xl.meta read/modify/commit transaction. This domain
 /// precedes namespace/volume publication locks, whose narrower syscall leases
@@ -2233,17 +2545,7 @@ pub(crate) async fn acquire_metadata_mutation_lease(
    object: &Path,
    owner: Option<Arc<dyn Send + Sync>>,
 ) -> Arc<MetadataMutationLease> {
-    let lock = {
-        let mut locks = DISK_METADATA_MUTATION_LOCKS.lock();
-        locks.retain(|_, lock| lock.strong_count() > 0);
-        if let Some(lock) = locks.get(object).and_then(Weak::upgrade) {
-            lock
-        } else {
-            let lock = Arc::new(AsyncMutex::new(()));
-            locks.insert(object.to_path_buf(), Arc::downgrade(&lock));
-            lock
-        }
-    };
+    let lock = DISK_METADATA_MUTATION_LOCKS.get_or_create(object, || AsyncMutex::new(()));
    Arc::new(MetadataMutationLease {
        _guard: lock.lock_owned().await,
        _owner: owner,
@@ -2306,17 +2608,45 @@ pub(crate) async fn acquire_rename_data_mutation_lease_with_owner(
 ) -> Arc<NamespaceMutationLease> {
    // Metadata ownership precedes namespace paths in lexical order, then volume
    // ownership. Deduplicate identical keys so TMP-to-TMP publication cannot self-lock.
-    let destination_lock = disk_namespace_mutation_lock(destination_object);
-    let (namespace_guard, staging_guard) = match staging_prefix {
-        Some(source) if source < destination_object => {
-            let staging = disk_namespace_mutation_lock(source).lock_owned().await;
-            (destination_lock.lock_owned().await, Some(staging))
+    #[cfg(unix)]
+    let (namespace_guard, staging_guard) = {
+        let dst_key = canonical_unix_path_key(destination_object);
+        let destination_lock = disk_namespace_mutation_lock_by_key(dst_key.as_ref());
+        match staging_prefix {
+            Some(source) => {
+                let src_key = canonical_unix_path_key(source);
+                match src_key.as_ref().cmp(dst_key.as_ref()) {
+                    std::cmp::Ordering::Less => {
+                        let staging = disk_namespace_mutation_lock_by_key(src_key.as_ref()).lock_owned().await;
+                        (destination_lock.lock_owned().await, Some(staging))
+                    }
+                    std::cmp::Ordering::Greater => {
+                        let destination = destination_lock.lock_owned().await;
+                        (
+                            destination,
+                            Some(disk_namespace_mutation_lock_by_key(src_key.as_ref()).lock_owned().await),
+                        )
+                    }
+                    std::cmp::Ordering::Equal => (destination_lock.lock_owned().await, None),
+                }
+            }
+            None => (destination_lock.lock_owned().await, None),
        }
-        Some(source) if source > destination_object => {
-            let destination = destination_lock.lock_owned().await;
-            (destination, Some(disk_namespace_mutation_lock(source).lock_owned().await))
+    };
+    #[cfg(not(unix))]
+    let (namespace_guard, staging_guard) = {
+        let destination_lock = disk_namespace_mutation_lock(destination_object);
+        match staging_prefix {
+            Some(source) if source < destination_object => {
+                let staging = disk_namespace_mutation_lock(source).lock_owned().await;
+                (destination_lock.lock_owned().await, Some(staging))
+            }
+            Some(source) if source > destination_object => {
+                let destination = destination_lock.lock_owned().await;
+                (destination, Some(disk_namespace_mutation_lock(source).lock_owned().await))
+            }
+            _ => (destination_lock.lock_owned().await, None),
        }
-        _ => (destination_lock.lock_owned().await, None),
    };
    let volume_guard = disk_volume_mutation_lock(root, volume).read_owned().await;
    let blocking_queue = disk_blocking_queue_for_root(root);
@@ -4903,9 +5233,7 @@ impl PublicationRoot {
    pub(crate) fn open_readonly(&self, path: &Path) -> io::Result<std::fs::File> {
        #[cfg(target_os = "linux")]
        return self.open_relative(path, rustix::fs::OFlags::RDONLY, rustix::fs::Mode::empty());
-        #[cfg(windows)]
-        return open_windows_readonly(path);
-        #[cfg(not(any(target_os = "linux", windows)))]
+        #[cfg(not(target_os = "linux"))]
        std::fs::File::open(path)
    }

@@ -5481,47 +5809,18 @@ pub(crate) fn multipart_payload_commit_guard_std(
    })
 }

-#[cfg(windows)]
-fn multipart_payload_commit_guard_std(
-    metadata_guard: &RenameCommitGuard,
-    source_parent: &Path,
-    publication_root: &PublicationRoot,
-) -> io::Result<RenameCommitGuard> {
-    // Pin the payload staging tree like a distinct-tree metadata source; the
-    // target stays the exact parent handle the metadata transaction holds.
-    let source_parent_guard = lock_windows_directory_tree(source_parent, Some(source_parent), publication_root)?;
-    Ok(RenameCommitGuard {
-        source_parent: source_parent.to_path_buf(),
-        destination_parent: metadata_guard.destination_parent.clone(),
-        source_parent_guard,
-        destination_parent_guard: metadata_guard.destination_parent_guard.clone(),
-    })
-}
-
 /// Native MPU metadata and payload have distinct source parents. Reuse the
 /// exact target identity already held by the metadata transaction, while
 /// independently pinning the private payload source under the same root.
-#[cfg(any(target_os = "linux", windows))]
+#[cfg(target_os = "linux")]
 pub(crate) async fn multipart_payload_commit_guard(
    metadata_guard: &RenameCommitGuard,
    source_parent: &Path,
-    publication_root: &PublicationRoot,
    lease: Arc<NamespaceMutationLease>,
 ) -> io::Result<RenameCommitGuard> {
    let metadata_guard = metadata_guard.clone();
    let source_parent = source_parent.to_path_buf();
-    #[cfg(target_os = "linux")]
-    let operation = {
-        // The Linux guard already carries its publication root.
-        let _ = publication_root;
-        move || multipart_payload_commit_guard_std(&metadata_guard, &source_parent)
-    };
-    #[cfg(windows)]
-    let operation = {
-        let publication_root = publication_root.clone();
-        move || multipart_payload_commit_guard_std(&metadata_guard, &source_parent, &publication_root)
-    };
-    run_blocking_namespace_operation(lease, operation).await
+    run_blocking_namespace_operation(lease, move || multipart_payload_commit_guard_std(&metadata_guard, &source_parent)).await
 }

 pub(crate) fn prepare_rename_commit_guard(
@@ -5596,28 +5895,6 @@ pub(crate) fn prepare_rename_commit_guard(
    }
 }

-/// Unix opens a directory read-only and lets the reader classify it, while
-/// Windows refuses the plain open with ACCESS_DENIED. Reopen only an actual
-/// directory with backup semantics so the same `is_dir` checks apply.
-#[cfg(windows)]
-fn open_windows_readonly(path: &Path) -> io::Result<std::fs::File> {
-    use std::os::windows::fs::OpenOptionsExt;
-    use windows_sys::Win32::Storage::FileSystem::FILE_FLAG_BACKUP_SEMANTICS;
-
-    match std::fs::File::open(path) {
-        Err(error)
-            if error.kind() == io::ErrorKind::PermissionDenied
-                && std::fs::symlink_metadata(path).is_ok_and(|metadata| metadata.is_dir()) =>
-        {
-            std::fs::OpenOptions::new()
-                .read(true)
-                .custom_flags(FILE_FLAG_BACKUP_SEMANTICS)
-                .open(path)
-        }
-        result => result,
-    }
-}
-
 #[cfg(windows)]
 fn open_windows_publication_root(path: &Path) -> io::Result<(PathBuf, WindowsDirectoryHandle)> {
    use std::os::windows::fs::OpenOptionsExt;
@@ -6654,31 +6931,12 @@ pub(crate) fn rename_multipart_source_owned_std(
    {
        let _ = &guard;
        let source_guard = pin_multipart_source_directory(source_parent, base, root)?;
-        // The prepared Windows source holds DELETE access and shares only
-        // read, so the precondition's own opens of that entry (or, for a
-        // directory, of its guarded tree) would fail with a sharing violation.
-        // Check it first under a non-exclusive identity anchor, then require
-        // the exclusive handle to name the same file.
-        #[cfg(windows)]
-        let (_source_identity_anchor, expected_source_identity) = {
-            let anchor = open_windows_rename_source_identity(source, &source_guard)?;
-            precondition()?;
-            anchor
-        };
        let preparation = prepare_rename_with_retry(source, destination, base, root)?;
-        #[cfg(windows)]
-        if windows_file_identity(&preparation.source)? != expected_source_identity {
-            return Err(io::Error::new(
-                io::ErrorKind::InvalidData,
-                "rename source identity changed after its precondition",
-            ));
-        }
        #[cfg(all(any(test, feature = "test-util"), not(windows)))]
        {
            prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Rename, source);
            prepared_publication_test_hooks::run(prepared_publication_test_hooks::Stage::Rename, destination);
        }
-        #[cfg(not(windows))]
        precondition()?;
        #[cfg(unix)]
        validate_prepared_rename_source(&pinned_source, source)?;
@@ -7265,7 +7523,7 @@ mod multipart_source_preservation_tests {
        let metadata_guard =
            prepare_rename_commit_guard(&canonical, &target, &bucket, &publication).expect("canonical metadata guard");
        let lease = acquire_namespace_mutation_lease(&target).await;
-        let payload_guard = multipart_payload_commit_guard(&metadata_guard, &temporary, &publication, lease.clone())
+        let payload_guard = multipart_payload_commit_guard(&metadata_guard, &temporary, lease.clone())
            .await
            .expect("distinct pinned payload source");
        assert!(
@@ -7300,61 +7558,6 @@ mod multipart_source_preservation_tests {
        );
    }

-    #[cfg(windows)]
-    #[tokio::test]
-    async fn windows_multipart_payload_guard_publishes_distinct_source_into_metadata_target() {
-        // Build paths from the configured root: the resolved handle path may
-        // be verbatim, where `/` in a joined key is not a separator.
-        let (root, publication, _source, _destination) = fixture();
-        let root = root.path();
-        let canonical = root.join("multipart/upload");
-        let temporary = root.join("tmp/attempt");
-        let bucket = root.join("bucket");
-        let target = bucket.join("object");
-        for directory in [&canonical, &temporary, &target] {
-            std::fs::create_dir_all(directory).expect("two-source guard directories");
-        }
-        let payload = temporary.join("payload");
-        std::fs::create_dir(&payload).expect("private payload DD");
-        std::fs::write(payload.join("part.1"), b"native selected shard").expect("private payload shard");
-        let head = canonical.join("xl.meta");
-        let destination = target.join("xl.meta");
-        let metadata_guard =
-            prepare_rename_commit_guard(&canonical, &target, &bucket, &publication).expect("canonical metadata guard");
-        let lease = acquire_namespace_mutation_lease(&target).await;
-        assert!(
-            rename_all_with_commit_guard(&payload, target.join("payload"), &bucket, &publication, &metadata_guard, lease.clone())
-                .await
-                .is_err(),
-            "the metadata guard does not pin the private payload parent"
-        );
-        assert!(
-            payload.join("part.1").exists(),
-            "a rejected payload rename must leave its source in place"
-        );
-        let payload_guard = multipart_payload_commit_guard(&metadata_guard, &temporary, &publication, lease.clone())
-            .await
-            .expect("distinct pinned payload source");
-        let mut prepared = create_prepared_rename_source_with_commit_guard(&head, &destination, &metadata_guard)
-            .expect("canonical prepared metadata");
-        prepared
-            .write_all(b"prepared native metadata", true)
-            .expect("durable canonical metadata");
-        rename_all_with_commit_guard(&payload, target.join("payload"), &bucket, &publication, &payload_guard, lease.clone())
-            .await
-            .expect("payload rename uses its own pinned TMP parent");
-        assert!(head.exists(), "payload publication must not consume the canonical head");
-        rename_all_with_prepared_source(prepared, &head, &destination, &bucket, &publication, &metadata_guard, lease)
-            .await
-            .expect("metadata rename keeps its canonical parent");
-        assert!(!head.exists());
-        assert_eq!(std::fs::read(&destination).expect("published metadata"), b"prepared native metadata");
-        assert_eq!(
-            std::fs::read(target.join("payload/part.1")).expect("published selected shard"),
-            b"native selected shard"
-        );
-    }
-
    #[cfg(target_os = "linux")]
    #[tokio::test]
    async fn multipart_payload_guard_rejects_replaced_private_source_or_shared_destination() {
@@ -7376,7 +7579,7 @@ mod multipart_source_preservation_tests {
            let metadata_guard =
                prepare_rename_commit_guard(&canonical, &target, &bucket, &publication).expect("canonical metadata guard");
            let lease = acquire_namespace_mutation_lease(&target).await;
-            let payload_guard = multipart_payload_commit_guard(&metadata_guard, &temporary, &publication, lease.clone())
+            let payload_guard = multipart_payload_commit_guard(&metadata_guard, &temporary, lease.clone())
                .await
                .expect("pin payload and shared target");
            let replaced = if replace_source { &temporary } else { &target };
@@ -12119,6 +12322,121 @@ mod tests {
        );
        assert!(preserved_final.exists(), "final file must remain when intermediate removal fails");
    }
+
+    #[cfg(unix)]
+    #[test]
+    fn canonical_unix_path_key_preserves_std_path_equality_and_prefix_semantics() {
+        let samples = [
+            "",
+            "/",
+            "///",
+            "/.",
+            "/./.",
+            ".",
+            "./",
+            "./.",
+            "./a",
+            "././a/",
+            ".//a/.",
+            "a",
+            "a/",
+            "a/.",
+            "a/b/c",
+            "a/b/c/",
+            "a//b/./c/.",
+            "/tmp/d0/bucket/obj",
+            "/tmp/d0/bucket/obj/",
+            "/tmp/d0//bucket/./obj/.",
+            "/tmp/d0/.rustfs.sys/multipart/sha/upload",
+            "/proc/self/fd/12/.",
+            "/proc/self/fd/12",
+            "/proc/self/fd/12/./bucket/obj",
+            "/proc/self/fd/12/bucket/obj",
+            "/tmp/d1",
+            "/tmp/d10",
+            "a/../b",
+            "a/../b/",
+            "/utf8/\u{00af}/segment/",
+        ];
+        for s in samples {
+            let p = Path::new(s);
+            let fast = canonical_unix_path_key(p);
+            let slow = canonical_unix_path_key_slow(p);
+            assert_eq!(fast.as_ref(), slow.as_ref(), "fast and slow key mismatch for {s:?}");
+        }
+        for a in samples {
+            for b in samples {
+                let pa = Path::new(a);
+                let pb = Path::new(b);
+                let ka = canonical_unix_path_key(pa);
+                let kb = canonical_unix_path_key(pb);
+                assert_eq!(
+                    pa == pb,
+                    ka.as_ref() == kb.as_ref(),
+                    "Path::eq vs canonical key mismatch for {a:?} and {b:?}"
+                );
+                assert_eq!(
+                    pa.starts_with(pb),
+                    unix_canonical_key_starts_with(ka.as_ref(), kb.as_ref()),
+                    "Path::starts_with vs canonical key prefix mismatch for {a:?} and {b:?}"
+                );
+            }
+        }
+
+        let dir = tempdir().expect("tempdir");
+        let p_clean = dir.path().join("bucket/object");
+        let p_trailing = PathBuf::from(format!("{}/", p_clean.display()));
+        let p_redundant = PathBuf::from(format!("{}//bucket/./object/.", dir.path().display()));
+        let lock1 = disk_namespace_mutation_lock(&p_clean);
+        let lock2 = disk_namespace_mutation_lock(&p_trailing);
+        let lock3 = disk_namespace_mutation_lock(&p_redundant);
+        assert!(Arc::ptr_eq(&lock1, &lock2));
+        assert!(Arc::ptr_eq(&lock1, &lock3));
+
+        let vlock1 = disk_volume_mutation_lock(dir.path(), "bucket");
+        let vlock2 = disk_volume_mutation_lock(&PathBuf::from(format!("{}/.", dir.path().display())), "bucket/");
+        assert!(Arc::ptr_eq(&vlock1, &vlock2));
+
+        let io_root = dir.path().join("io_root/.");
+        let q_reg = register_disk_blocking_queue(dir.path(), &io_root);
+        let q_root = disk_blocking_queue_for_root(dir.path());
+        let q_path = disk_blocking_queue_for_path(&dir.path().join("io_root/bucket/key"));
+        assert!(Arc::ptr_eq(&q_reg, &q_root));
+        assert!(Arc::ptr_eq(&q_reg, &q_path));
+    }
+
+    #[test]
+    fn sharded_path_lock_registry_reuses_slots_and_bounds_dead_weak_entries() {
+        let registry: ShardedPathLockRegistry<AsyncMutex<()>> = ShardedPathLockRegistry::new();
+        let target = Path::new("/tmp/sharded-lock-test/persistent-key");
+        let held = registry.get_or_create(target, || AsyncMutex::new(()));
+        let held_again = registry.get_or_create(target, || AsyncMutex::new(()));
+        assert!(Arc::ptr_eq(&held, &held_again), "live lock must be shared");
+
+        // Re-acquiring a key after drop must reuse the existing slot in-place.
+        let transient = Path::new("/tmp/sharded-lock-test/reused-slot");
+        drop(registry.get_or_create(transient, || AsyncMutex::new(())));
+        let reused = registry.get_or_create(transient, || AsyncMutex::new(()));
+        drop(reused);
+
+        // Flood transient keys across shards and verify `held` is never evicted
+        // and total dead entries remain bounded by the amortized GC threshold.
+        for i in 0..(PATH_LOCK_SHARDS * PATH_LOCK_INITIAL_GC_THRESHOLD * 3) {
+            let p = PathBuf::from(format!("/tmp/sharded-lock-test/ephemeral-{i}"));
+            drop(registry.get_or_create(&p, || AsyncMutex::new(())));
+        }
+        let still_held = registry.get_or_create(target, || AsyncMutex::new(()));
+        assert!(Arc::ptr_eq(&held, &still_held), "GC must never evict a live lock");
+        for shard in &registry.shards {
+            let guard = shard.lock();
+            assert!(
+                guard.entries.len() <= guard.next_gc_threshold,
+                "shard entries ({}) exceeded GC threshold ({})",
+                guard.entries.len(),
+                guard.next_gc_threshold
+            );
+        }
+    }
 }

 #[cfg(all(test, target_os = "linux"))]
@@ -318,6 +318,21 @@ fn is_rebalance_transient_lock_error(err: &rustfs_lock::LockError) -> bool {
    }
 }

+fn is_rebalance_metadata_admission_retryable(err: &Error) -> bool {
+    match rebalance_error_source(err) {
+        Error::Lock(err) => is_rebalance_transient_lock_error(err),
+        // No metadata was written when lock admission failed. Each attempt
+        // must acquire its quorum again before loading or saving this run.
+        Error::NamespaceLockQuorumUnavailable {
+            mode: "write",
+            bucket,
+            object,
+            ..
+        } => bucket == crate::disk::RUSTFS_META_BUCKET && object == REBAL_META_NAME,
+        _ => false,
+    }
+}
+
 fn is_rebalance_transient_io_error(err: &std::io::Error) -> bool {
    if err.kind() == std::io::ErrorKind::TimedOut {
        return true;
@@ -385,9 +400,7 @@ where
        match result {
            Ok(value) => return Ok(value),
            Err(err) => {
-                if attempt.saturating_add(1) >= max_attempts.max(1)
-                    || !matches!(rebalance_error_source(&err), Error::Lock(lock_err) if is_rebalance_transient_lock_error(lock_err))
-                {
+                if attempt.saturating_add(1) >= max_attempts.max(1) || !is_rebalance_metadata_admission_retryable(&err) {
                    return Err(err);
                }
                #[cfg(test)]
@@ -719,6 +732,82 @@ mod error_source_tests {
        }
    }

+    #[tokio::test(start_paused = true)]
+    async fn rebalance_metadata_retry_recovers_after_write_quorum_failure() {
+        let mut attempts = 0;
+        retry_rebalance_metadata_access(None, 3, || {
+            attempts += 1;
+            std::future::ready(if attempts == 1 {
+                resolve_rebalance_meta_save_result(
+                    Err(rebalance_meta_lock_error(
+                        rustfs_lock::LockError::QuorumNotReached {
+                            required: 1,
+                            achieved: 0,
+                        },
+                        "write",
+                    )),
+                    "stop_rebalance",
+                )
+            } else {
+                Ok(())
+            })
+        })
+        .await
+        .expect("a later quorum-protected admission must be allowed to succeed");
+        assert_eq!(attempts, 2);
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn rebalance_metadata_retry_retains_persistent_write_quorum_failure() {
+        for max_attempts in [0, 1, 3] {
+            let mut attempts = 0;
+            let result = retry_rebalance_metadata_access(None, max_attempts, || {
+                attempts += 1;
+                std::future::ready(Err::<(), _>(rebalance_meta_lock_error(
+                    rustfs_lock::LockError::QuorumNotReached {
+                        required: 1,
+                        achieved: 0,
+                    },
+                    "write",
+                )))
+            })
+            .await;
+            let error = result.expect_err("exhausted admission attempts must never acknowledge a write");
+            assert_eq!(attempts, max_attempts.max(1));
+            assert!(matches!(
+                rebalance_error_source(&error),
+                Error::NamespaceLockQuorumUnavailable {
+                    mode: "write",
+                    required: 1,
+                    achieved: 0,
+                    ..
+                }
+            ));
+        }
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn rebalance_metadata_retry_stops_after_admission_is_cancelled() {
+        let mut attempts = 0;
+        let result = retry_rebalance_metadata_access(None, 3, || {
+            attempts += 1;
+            std::future::ready(Err::<(), _>(if attempts == 1 {
+                rebalance_meta_lock_error(
+                    rustfs_lock::LockError::QuorumNotReached {
+                        required: 1,
+                        achieved: 0,
+                    },
+                    "write",
+                )
+            } else {
+                Error::OperationCanceled
+            }))
+        })
+        .await;
+        assert_eq!(attempts, 2, "a stale or cancelled run cannot enter another admission attempt");
+        assert!(matches!(result, Err(Error::OperationCanceled)));
+    }
+
    #[tokio::test]
    async fn rebalance_metadata_retry_does_not_retry_permanent_or_untyped_errors() {
        for err in [
@@ -731,6 +820,20 @@ mod error_source_tests {
                required: 3,
                achieved: 2,
            },
+            Error::NamespaceLockQuorumUnavailable {
+                mode: "write",
+                bucket: "data-bucket".to_string(),
+                object: REBAL_META_NAME.to_string(),
+                required: 1,
+                achieved: 0,
+            },
+            Error::NamespaceLockQuorumUnavailable {
+                mode: "write",
+                bucket: crate::disk::RUSTFS_META_BUCKET.to_string(),
+                object: "other-metadata.bin".to_string(),
+                required: 1,
+                achieved: 0,
+            },
            Error::other("stale rebalance run rejected: lock acquisition timed out"),
            Error::other("rebalance distributed run fence lost"),
            Error::Io(std::io::Error::from(std::io::ErrorKind::TimedOut)),
@@ -4696,6 +4696,17 @@ pub(crate) fn same_distributed_lock_domain(left: &[Arc<dyn LockClient>], right:
            .all(|right_client| left.iter().any(|left_client| Arc::ptr_eq(left_client, right_client)))
 }

+// Reusing a held writer additionally requires the same quorum, so repeated
+// clients must not give different weight to an otherwise identical domain.
+fn same_distributed_lock_quorum(left: &[Arc<dyn LockClient>], right: &[Arc<dyn LockClient>]) -> bool {
+    !left.is_empty()
+        && left.len() == right.len()
+        && left.iter().all(|client| {
+            left.iter().filter(|other| Arc::ptr_eq(client, other)).count() == 1
+                && right.iter().filter(|other| Arc::ptr_eq(client, other)).count() == 1
+        })
+}
+
 const ERASURE_CACHE_MAX_ENTRIES: usize = 32;

 #[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
@@ -5645,9 +5656,9 @@ impl SetDisks {
    pub(crate) async fn shares_namespace_lock_domain(&self, other: &Self) -> bool {
        match (self.ctx.is_dist_erasure().await, other.ctx.is_dist_erasure().await) {
            (false, false) => Arc::ptr_eq(&self.local_lock_manager, &other.local_lock_manager),
-            (true, true) => {
-                self.set_lock_namespace == other.set_lock_namespace && same_distributed_lock_domain(&self.lockers, &other.lockers)
-            }
+            // Distributed lock clients receive the raw ObjectKey; the pool
+            // namespace is not part of the key stored by their backends.
+            (true, true) => same_distributed_lock_quorum(&self.lockers, &other.lockers),
            _ => false,
        }
    }
@@ -9882,7 +9893,7 @@ mod tests {

    #[tokio::test(flavor = "multi_thread")]
    #[serial]
-    async fn namespace_lock_domain_includes_pool_namespace() {
+    async fn distributed_namespace_lock_domain_follows_backend_not_pool_name() {
        let _setup_type_guard = SetupTypeGuard::switch_to(SetupType::DistErasure).await;

        let first: Arc<dyn LockClient> = Arc::new(LocalClient::with_manager(Arc::new(rustfs_lock::GlobalLockManager::new())));
@@ -9921,11 +9932,25 @@ mod tests {
            "sets in the same pool share the object namespace lock domain"
        );
        assert!(
-            !same_pool_first_set.shares_namespace_lock_domain(&other_pool_set).await,
-            "different pool namespaces must not be deduplicated solely by identical clients"
+            same_pool_first_set.shares_namespace_lock_domain(&other_pool_set).await,
+            "different pools using identical backends lock the same raw object key"
        );
    }

+    #[test]
+    fn distributed_lock_domain_rejects_partial_or_duplicate_backends() {
+        let a: Arc<dyn LockClient> = Arc::new(LocalClient::with_manager(Arc::new(rustfs_lock::GlobalLockManager::new())));
+        let b: Arc<dyn LockClient> = Arc::new(LocalClient::with_manager(Arc::new(rustfs_lock::GlobalLockManager::new())));
+        let c: Arc<dyn LockClient> = Arc::new(LocalClient::with_manager(Arc::new(rustfs_lock::GlobalLockManager::new())));
+        assert!(same_distributed_lock_quorum(&[a.clone(), b.clone()], &[b.clone(), a.clone()]));
+        assert!(!same_distributed_lock_quorum(&[a.clone(), b.clone()], &[a.clone(), c]));
+        assert!(!same_distributed_lock_quorum(
+            &[a.clone(), a.clone(), b.clone()],
+            &[a.clone(), b.clone(), b]
+        ));
+        assert!(!same_distributed_lock_quorum(&[], &[]));
+    }
+
    #[tokio::test(flavor = "multi_thread")]
    #[serial]
    async fn streaming_reader_holds_read_lock_until_eof() {
@@ -10060,7 +10060,7 @@ mod tests {

        // Real disk callbacks and health probes must share wall-clock deadlines.
        // Advancing Tokio time can expire unrelated healthy-drive probes.
-        temp_env::async_with_vars([(rustfs_config::ENV_DRIVE_MAX_TIMEOUT_DURATION, Some("2"))], async {
+        temp_env::async_with_vars([(rustfs_config::ENV_DRIVE_MAX_TIMEOUT_DURATION, None::<&str>)], async {
            let (dirs, store) = crate::bucket::metadata_sys::test_support::isolated_store_over_temp_disks().await;
            crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
            let bucket = "mpu-quorum-success-physical";
@@ -10087,6 +10087,8 @@ mod tests {
                local.observe_multipart_cleanup_for_test(&upload_path)
            }).collect::<Vec<_>>();
            let disk::Disk::Local(paused_local) = disks[3].as_ref() else { panic!("actual local physical tail") };
+            // Fault only the held leg; healthy disks keep their normal I/O budgets.
+            let _paused_timeout = paused_local.override_operation_timeout_for_test(Duration::from_secs(2));
            let io_target = paused_local.get_disk().get_object_path_for_io(bucket, object).expect("real IO target");
            let mutation_key = paused_local.get_disk().get_object_path(bucket, object).expect("logical native mutation key");
            let (entered_tx, entered_rx) = tokio::sync::oneshot::channel();
@@ -10243,7 +10245,15 @@ mod tests {
                let size = copy_info.get_actual_size().expect("actual source size");
                copy_info.put_object_reader = Some(PutObjReader::new(HashReader::from_stream(reader.stream, size, size, None, None, false).expect("actual Copy body stream")));
                tokio::time::timeout(Duration::from_secs(5), store.copy_object(bucket, object, bucket, copy_object, &mut copy_info, &ObjectOptions::default(), &ObjectOptions::default()))
-                    .await.expect("cross-key Copy cannot wait on the old successful logical owner").expect("actual CopyObject");
+                    .await.expect("cross-key Copy cannot wait on the old successful logical owner")
+                    .unwrap_or_else(|error| {
+                        let health = disks.iter().enumerate().map(|(slot, disk)| {
+                            let disk::Disk::Local(local) = disk.as_ref() else { panic!("actual local Copy backend") };
+                            let metrics = local.metrics_snapshot();
+                            (slot, disk.runtime_state(), metrics.total_errors_timeout, metrics.total_errors_availability, metrics.total_waiting, metrics.api_calls)
+                        }).collect::<Vec<_>>();
+                        panic!("actual CopyObject: {error:?}; drive health (slot, state, timeouts, unavailable, waiting, calls): {health:?}");
+                    });
                drop(copy_info);
                let mut copied = store.get_object_reader(bucket, copy_object, None, HeaderMap::new(), &ObjectOptions::default()).await.expect("actual copied body");
                let mut actual = Vec::new();
@@ -3692,6 +3692,177 @@ mod tests {
        );
    }

+    // The RPC lock service uses the raw ObjectKey and one backend per node,
+    // even when the requesting sets carry different pool namespace strings.
+    async fn pool_meta_heal_store_with_lock_domains(source: &Arc<ECStore>, shared: bool) -> Arc<ECStore> {
+        use rustfs_lock::client::{LockClient, local::LocalClient};
+
+        let clients = || {
+            (0..2)
+                .map(|_| {
+                    Arc::new(LocalClient::with_manager(Arc::new(rustfs_lock::GlobalLockManager::new()))) as Arc<dyn LockClient>
+                })
+                .collect::<Vec<_>>()
+        };
+        let first = clients();
+        let second = if shared { first.clone() } else { clients() };
+        let mut pools = source.pools.clone();
+        for (pool, lockers) in pools.iter_mut().zip([first, second]) {
+            for set in &mut Arc::make_mut(pool).disk_set {
+                Arc::make_mut(set).lockers = lockers.clone();
+            }
+        }
+        let endpoints: EndpointServerPools = pools.iter().map(|pool| pool.endpoints.clone()).collect::<Vec<_>>().into();
+        let state = source.pool_meta_save_gate.lock().await.independent_clone_for_test();
+        Arc::new(ECStore {
+            id: Uuid::new_v4(),
+            disk_map: Default::default(),
+            pools,
+            peer_sys: crate::cluster::rpc::S3PeerSys::new_with_instance_ctx(&endpoints, Arc::clone(&source.ctx)),
+            pool_meta: tokio::sync::RwLock::new(source.pool_meta.read().await.clone()),
+            rebalance_meta: tokio::sync::RwLock::new(None),
+            decommission_cancelers: tokio::sync::RwLock::new(vec![None; 2]),
+            start_gate: tokio::sync::Mutex::new(()),
+            pool_meta_save_gate: tokio::sync::Mutex::new(state),
+            ctx: Arc::clone(&source.ctx),
+            bucket_fence_registry: Arc::default(),
+        })
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn pool_meta_same_distributed_backend_has_one_raw_lock_across_pools() {
+        let (_dirs, source, _peer) = test_two_pool_stores(None).await;
+        let store = pool_meta_heal_store_with_lock_domains(&source, true).await;
+        let first = store.pools[0].get_disks_by_key(POOL_META_NAME);
+        let second = store.pools[1].get_disks_by_key(POOL_META_NAME);
+        let first_lock = first
+            .new_ns_lock(RUSTFS_META_BUCKET, POOL_META_NAME)
+            .await
+            .expect("first pool lock");
+        let second_lock = second
+            .new_ns_lock(RUSTFS_META_BUCKET, POOL_META_NAME)
+            .await
+            .expect("second pool lock");
+        let guard = first_lock
+            .get_write_lock(std::time::Duration::from_secs(2))
+            .await
+            .expect("first writer");
+        assert!(
+            second_lock
+                .get_write_lock(std::time::Duration::from_millis(150))
+                .await
+                .is_err(),
+            "different pool namespaces still conflict on the same backend and object key"
+        );
+        drop(guard);
+        second_lock
+            .get_write_lock(std::time::Duration::from_secs(3))
+            .await
+            .expect("second writer should acquire after the first releases");
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn pool_meta_heal_reuses_same_distributed_backend_across_pools() {
+        let (_dirs, source, _peer) = test_two_pool_stores(None).await;
+        let store = pool_meta_heal_store_with_lock_domains(&source, true).await;
+        let (item, error) = tokio::time::timeout(
+            std::time::Duration::from_secs(3),
+            store.handle_heal_object(
+                RUSTFS_META_BUCKET,
+                POOL_META_NAME,
+                "",
+                &HealOpts {
+                    pool: Some(1),
+                    dry_run: true,
+                    ..Default::default()
+                },
+            ),
+        )
+        .await
+        .expect("inspection must reuse the fixed pool's writer when the lock backend is shared")
+        .expect("pool metadata heal inspection should return a result");
+        assert!(error.is_none(), "inspection should succeed: {error:?}");
+        assert_eq!(item.object, POOL_META_NAME);
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn pool_meta_heal_repairs_other_pool_with_shared_distributed_backend() {
+        let (_dirs, source, _peer) = test_two_pool_stores(None).await;
+        let store = pool_meta_heal_store_with_lock_domains(&source, true).await;
+        let missing = remove_pool_meta_shard(&store, 1).await;
+        let (item, error) = tokio::time::timeout(
+            std::time::Duration::from_secs(10),
+            store.handle_heal_object(
+                RUSTFS_META_BUCKET,
+                POOL_META_NAME,
+                "",
+                &HealOpts {
+                    pool: Some(1),
+                    ..Default::default()
+                },
+            ),
+        )
+        .await
+        .expect("cross-pool repair must not reenter the shared writer")
+        .expect("pool metadata repair should return a result");
+        assert!(error.is_none(), "cross-pool repair should complete: {error:?}");
+        assert_eq!(item.object, POOL_META_NAME);
+        assert!(
+            missing.read_xl(RUSTFS_META_BUCKET, POOL_META_NAME, false).await.is_ok(),
+            "the selected missing shard must be restored"
+        );
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn pool_meta_heal_keeps_target_lock_for_independent_distributed_backend() {
+        let (_dirs, source, _peer) = test_two_pool_stores(None).await;
+        let store = pool_meta_heal_store_with_lock_domains(&source, false).await;
+        let target = store.pools[1].get_disks_by_key(POOL_META_NAME);
+        let lock = target
+            .new_ns_lock(RUSTFS_META_BUCKET, POOL_META_NAME)
+            .await
+            .expect("target lock");
+        let guard = lock
+            .get_write_lock(std::time::Duration::from_secs(2))
+            .await
+            .expect("hold target lock");
+        let barrier = DecommissionCapacityLockOrderBarrier::install(store.id, store.id);
+        let heal_store = Arc::clone(&store);
+        let mut heal = tokio::spawn(async move {
+            heal_store
+                .handle_heal_object(
+                    RUSTFS_META_BUCKET,
+                    POOL_META_NAME,
+                    "",
+                    &HealOpts {
+                        pool: Some(1),
+                        dry_run: true,
+                        ..Default::default()
+                    },
+                )
+                .await
+        });
+        tokio::time::timeout(
+            std::time::Duration::from_secs(3),
+            barrier.wait_until_external_heal_target_lock_attempted(),
+        )
+        .await
+        .expect("inspection must attempt the independent target lock");
+        assert!(!heal.is_finished(), "inspection must wait while the target writer is held");
+        drop(guard);
+        let (item, error) = tokio::time::timeout(std::time::Duration::from_secs(3), &mut heal)
+            .await
+            .expect("inspection should progress after target lock release")
+            .expect("heal task should not panic")
+            .expect("pool metadata heal inspection should return a result");
+        assert!(error.is_none(), "inspection should succeed: {error:?}");
+        assert_eq!(item.object, POOL_META_NAME);
+    }
+
    #[tokio::test]
    #[serial_test::serial]
    async fn pool_meta_heal_cas_repairs_only_the_requested_pool() {
@@ -318,10 +318,7 @@ mod multipart_copy_heal_tests {
            let mut corrupt = original.clone();
            corrupt[0] ^= 0x80;
            std::fs::write(&shard, &corrupt).expect("actual bitrot corruption in shared physical shard");
-            // Windows only flushes a handle opened for writing.
-            std::fs::OpenOptions::new()
-                .write(true)
-                .open(&shard)
+            std::fs::File::open(&shard)
                .expect("open corrupted actual shard")
                .sync_all()
                .expect("persist actual fault bytes");
@@ -925,6 +925,13 @@ impl MetaCacheEntriesSorted {

 const METACACHE_STREAM_VERSION: u8 = 2;
 const METACACHE_READ_CHUNK_SIZE: usize = 64 << 10;
+/// Upper bound on one incremental reservation while accumulating a single
+/// metacache element. Read rounds double from [`METACACHE_READ_CHUNK_SIZE`]
+/// up to this bound so a legitimately large xl.meta does not pay one small
+/// read per chunk, while a corrupt length prefix still cannot force a
+/// reservation larger than the payload that has already arrived (see
+/// rustfs/rustfs#2715 and rustfs/rustfs#9120).
+const METACACHE_READ_MAX_CHUNK_SIZE: usize = 4 << 20;
 /// msgpack `false`: the entry-stream terminator, and the only byte a legacy
 /// reader accepts after it (see `MetacacheWriter::close_with_limit_reached`).
 const METACACHE_TERMINATOR: u8 = 0xc2;
@@ -1157,13 +1164,14 @@ impl<R: AsyncRead + Unpin> MetacacheReader<R> {
        // A whole xl.meta can legitimately exceed the limit on its individual
        // msgpack values. Grow only as payload arrives, so a corrupt length
        // prefix cannot cause a speculative allocation of the advertised size.
+        let mut chunk = METACACHE_READ_CHUNK_SIZE;
        while self.buf.len() < read_size {
            let start = self.buf.len();
            let extra = if lookahead {
                let min_chunk = METACACHE_READ_CHUNK_SIZE.saturating_sub(start);
-                (read_size - start).clamp(min_chunk, METACACHE_READ_CHUNK_SIZE)
+                (read_size - start).clamp(min_chunk, chunk)
            } else {
-                (read_size - start).min(METACACHE_READ_CHUNK_SIZE)
+                (read_size - start).min(chunk)
            };
            if let Err(e) = self.buf.try_reserve(extra) {
                let err = Error::other(format!("metacache stream: buffer allocation of {extra} bytes failed: {e}"));
@@ -1194,6 +1202,7 @@ impl<R: AsyncRead + Unpin> MetacacheReader<R> {
                    return Err(err);
                }
            }
+            chunk = chunk.saturating_mul(2).min(METACACHE_READ_MAX_CHUNK_SIZE);
        }

        self.pos = read_size;
@@ -2155,6 +2164,76 @@ mod tests {
        }
    }

+    /// rustfs/rustfs#9120: a listing stream made of many ordinary entries must
+    /// decode completely even when the aggregate stream exceeds the per-value
+    /// `MAX_MSGP_ELEMENT_SIZE` bound that release 1.0.4 applied to every
+    /// element.
+    #[tokio::test]
+    async fn test_reader_wide_listing_aggregate_exceeds_per_value_limit() {
+        // ~2.9 KiB of metadata per entry matches the per-entry cost observed in
+        // the #9120 production report; 6,000 entries push the stream past 16 MiB.
+        const ENTRIES: usize = 6_000;
+        let metadata = vec![0xab; 2_880];
+        let entries: Vec<MetaCacheEntry> = (0..ENTRIES)
+            .map(|i| MetaCacheEntry {
+                name: format!("bigdir/object-{i:05}"),
+                metadata: metadata.clone(),
+                ..Default::default()
+            })
+            .collect();
+        let mut output = Cursor::new(Vec::new());
+        let mut writer = MetacacheWriter::new(&mut output);
+        writer.write(&entries).await.expect("wide listing must encode");
+        writer.close().await.expect("stream must close");
+        let bytes = output.into_inner();
+        assert!(
+            bytes.len() > 16 << 20,
+            "fixture must exceed the per-value limit, got {} bytes",
+            bytes.len()
+        );
+        let mut reader = MetacacheReader::new(Cursor::new(bytes));
+        let decoded = reader.read_all().await.expect("wide listing must decode");
+        assert_eq!(decoded, entries);
+    }
+
+    /// The #9120 failure shape seen in production: one entry whose metadata is
+    /// an xl.meta that grew past 16 MiB with its version history, listed among
+    /// ordinary neighbors and skipped past without losing the framing.
+    #[tokio::test]
+    async fn test_reader_oversized_entry_among_ordinary_neighbors_stays_listable() {
+        let small = |i: u8| MetaCacheEntry {
+            name: format!("bigdir/object-{i}"),
+            metadata: vec![i; 64],
+            ..Default::default()
+        };
+        let oversized = MetaCacheEntry {
+            name: "bigdir/huge-versions".to_string(),
+            metadata: vec![0xcd; (16 << 20) + 12_345],
+            ..Default::default()
+        };
+        let entries = vec![small(0), small(1), oversized.clone(), small(2), small(3)];
+        let mut output = Cursor::new(Vec::new());
+        let mut writer = MetacacheWriter::new(&mut output);
+        writer.write(&entries).await.expect("mixed listing must encode");
+        writer.close().await.expect("stream must close");
+
+        let mut reader = MetacacheReader::new(Cursor::new(output.into_inner()));
+        assert_eq!(
+            reader.peek().await.expect("first entry").expect("present"),
+            entries[0],
+            "an ordinary entry before the oversized one must decode"
+        );
+        reader.skip(1).await.expect("skip the first entry");
+        assert_eq!(reader.peek().await.expect("second entry").expect("present"), entries[1]);
+        assert_eq!(
+            reader.peek().await.expect("oversized entry").expect("present"),
+            oversized,
+            "an entry past the per-value limit must round-trip whole"
+        );
+        reader.skip(1).await.expect("skipping the oversized entry must stay aligned");
+        assert_eq!(reader.read_all().await.expect("tail must decode"), entries[3..]);
+    }
+
    #[test]
    fn file_info_versions_with_free_versions_includes_persisted_tier_cleanup_records() {
        let version_id = Uuid::new_v4();
@@ -810,7 +810,14 @@ fn publish_heal_queue_length(queue: &PriorityHealQueue) {
 fn active_heal_for_dedup_key(active_heals: &HashMap<String, Arc<HealTask>>, key: &str) -> Option<(String, HealType)> {
    active_heals
        .iter()
-        .find(|(_, task)| PriorityHealQueue::make_dedup_key_for_scope(&task.heal_type, &task.options) == key)
+        .find(|(_, task)| {
+            PriorityHealQueue::make_dedup_key_for_owner(
+                &task.heal_type,
+                &task.options,
+                task.source,
+                task.expected_mrf_bucket_incarnation_id,
+            ) == key
+        })
        .map(|(task_id, task)| (task_id.clone(), task.heal_type.clone()))
 }

@@ -611,7 +611,104 @@ impl PriorityHealQueue {

    /// Create a deduplication key from a heal request
    pub(super) fn make_dedup_key(request: &HealRequest) -> String {
-        Self::make_dedup_key_for_scope(&request.heal_type, &request.options)
+        Self::make_dedup_key_for_owner(
+            &request.heal_type,
+            &request.options,
+            request.source,
+            request.expected_mrf_bucket_incarnation_id,
+        )
+    }
+
+    /// Bound MRF work can only merge with work for the same bucket lifetime.
+    /// Length-prefix user-controlled components so a key cannot forge another
+    /// version, scope, or incarnation through delimiter characters.
+    pub(super) fn make_dedup_key_for_owner(
+        heal_type: &HealType,
+        options: &HealOptions,
+        source: HealRequestSource,
+        incarnation: Option<Uuid>,
+    ) -> String {
+        if source == HealRequestSource::Mrf
+            && let Some(incarnation) = incarnation
+        {
+            use std::fmt::Write as _;
+
+            let (pool, set) = heal_scope_indices(heal_type, options);
+            let mut key = format!("mrf-bound-v1:{incarnation}:{pool:?}:{set:?}:");
+            let mut part = |value: &str| {
+                // Formatting into a String is infallible.
+                let _ = write!(key, "{}:{value}", value.len());
+            };
+            match heal_type {
+                HealType::Object {
+                    bucket,
+                    object,
+                    version_id,
+                } => {
+                    part("object");
+                    part(bucket);
+                    part(object);
+                    match version_id {
+                        None => part("none"),
+                        Some(version_id) => {
+                            part("some");
+                            part(version_id);
+                        }
+                    }
+                }
+                HealType::Metadata { bucket, object } => {
+                    part("metadata");
+                    part(bucket);
+                    part(object);
+                }
+                HealType::ECDecode {
+                    bucket,
+                    object,
+                    version_id,
+                } => {
+                    part("ecdecode");
+                    part(bucket);
+                    part(object);
+                    match version_id {
+                        None => part("none"),
+                        Some(version_id) => {
+                            part("some");
+                            part(version_id);
+                        }
+                    }
+                }
+                HealType::DeleteMarkerPurge {
+                    bucket,
+                    object,
+                    version_id,
+                    purge,
+                } => {
+                    part("delete-marker-purge");
+                    part(bucket);
+                    part(object);
+                    part(version_id);
+                    part(&purge.bucket_incarnation_id.to_string());
+                    part(&purge.marker_incarnation_id.to_string());
+                    part(&base64_simd::URL_SAFE_NO_PAD.encode_to_string(purge.marker_identity));
+                }
+                HealType::Bucket { bucket } => {
+                    part("bucket");
+                    part(bucket);
+                }
+                HealType::Prefix { bucket, prefix } => {
+                    part("prefix");
+                    part(bucket);
+                    part(prefix);
+                }
+                // Cluster and erasure-set work are not bucket-bound; preserve
+                // their existing keys, including contains_erasure_set lookups.
+                HealType::Cluster | HealType::ErasureSet { .. } => {
+                    return Self::make_dedup_key_for_scope(heal_type, options);
+                }
+            }
+            return key;
+        }
+        Self::make_dedup_key_for_scope(heal_type, options)
    }

    pub(super) fn make_dedup_key_for_scope(heal_type: &HealType, options: &HealOptions) -> String {
@@ -117,6 +117,453 @@ async fn process_manager_queue_once(manager: &HealManager) {

 struct MockStorage;

+fn incarnation_bound_mrf_fallback(incarnation: Uuid) -> HealRequest {
+    let mut request = HealRequest::object("mrf-recreated-bucket".to_owned(), "object".to_owned(), None);
+    request.source = HealRequestSource::Mrf;
+    request.expected_mrf_bucket_incarnation_id = Some(incarnation);
+    request.options.pool_index = Some(0);
+    request.options.set_index = Some(0);
+    request
+}
+
+async fn assert_recreated_bucket_fallback_has_its_own_owner(existing_state: &str) {
+    let storage: Arc<dyn HealStorageAPI> = Arc::new(MockStorage);
+    let manager = HealManager::new_without_root_recovery_for_test(Arc::clone(&storage), None);
+    let old_incarnation = Uuid::new_v4();
+    let new_incarnation = Uuid::new_v4();
+    let old_request = incarnation_bound_mrf_fallback(old_incarnation);
+    let old_id = old_request.id.clone();
+    match existing_state {
+        "queued" => {
+            let receipt = manager
+                .submit_heal_request_with_receipt(old_request)
+                .await
+                .expect("old fallback must be admitted");
+            assert_eq!(receipt.result, HealAdmissionResult::Accepted);
+        }
+        "active" => {
+            manager
+                .active_heals
+                .lock()
+                .await
+                .insert(old_id.clone(), Arc::new(HealTask::from_request(old_request, Arc::clone(&storage))));
+        }
+        "retrying" => {
+            manager.retrying_heals.lock().await.insert(
+                old_id.clone(),
+                RetryingHeal {
+                    request: old_request,
+                    error: "retry pending".to_owned(),
+                    cancel_token: CancellationToken::new(),
+                },
+            );
+        }
+        other => panic!("unknown fixture state: {other}"),
+    }
+
+    let new_request = incarnation_bound_mrf_fallback(new_incarnation);
+    let new_id = new_request.id.clone();
+    assert_ne!(new_id, old_id, "the fallback must use a new request ID");
+    let receipt = manager
+        .submit_heal_request_with_receipt(new_request)
+        .await
+        .expect("new incarnation admission must be decided");
+    assert_eq!(
+        receipt.result,
+        HealAdmissionResult::Accepted,
+        "a {existing_state} fallback for an older bucket incarnation cannot own the new repair"
+    );
+    assert_eq!(receipt.task_id, new_id, "the new incarnation must retain its own task identity");
+    let queue = manager.heal_queue.lock().await;
+    assert!(
+        queue
+            .requests()
+            .any(|request| { request.id == new_id && request.expected_mrf_bucket_incarnation_id == Some(new_incarnation) })
+    );
+    if existing_state == "queued" {
+        assert!(
+            queue
+                .requests()
+                .any(|request| { request.id == old_id && request.expected_mrf_bucket_incarnation_id == Some(old_incarnation) })
+        );
+    }
+}
+
+#[tokio::test]
+async fn mrf_fallback_recreated_bucket_does_not_merge_into_queued_old_incarnation() {
+    assert_recreated_bucket_fallback_has_its_own_owner("queued").await;
+}
+
+#[tokio::test]
+async fn mrf_fallback_recreated_bucket_does_not_merge_into_active_old_incarnation() {
+    assert_recreated_bucket_fallback_has_its_own_owner("active").await;
+}
+
+#[tokio::test]
+async fn mrf_fallback_recreated_bucket_does_not_merge_into_retrying_old_incarnation() {
+    assert_recreated_bucket_fallback_has_its_own_owner("retrying").await;
+}
+
+fn incarnation_bound_mrf_fallback_for_version(incarnation: Uuid, version_id: Option<&str>) -> HealRequest {
+    let mut request = incarnation_bound_mrf_fallback(incarnation);
+    if let HealType::Object { version_id: id, .. } = &mut request.heal_type {
+        *id = version_id.map(str::to_owned);
+    } else {
+        unreachable!("fallback fixture must be an object request");
+    }
+    request
+}
+
+#[tokio::test]
+async fn versioned_mrf_fallbacks_merge_into_queued_owner() {
+    for version in [Uuid::nil(), Uuid::new_v4()] {
+        let manager = HealManager::new_without_root_recovery_for_test(Arc::new(MockStorage), None);
+        let incarnation = Uuid::new_v4();
+        let version_id = version.to_string();
+        let original = incarnation_bound_mrf_fallback_for_version(incarnation, Some(&version_id));
+        let original_id = original.id.clone();
+        let accepted = manager
+            .submit_heal_request_with_receipt(original)
+            .await
+            .expect("first queued fallback");
+        assert_eq!(accepted.result, HealAdmissionResult::Accepted);
+        let merged = manager
+            .submit_heal_request_with_receipt(incarnation_bound_mrf_fallback_for_version(incarnation, Some(&version_id)))
+            .await
+            .expect("duplicate queued fallback");
+        assert_eq!(merged.result, HealAdmissionResult::Merged);
+        assert_eq!(merged.task_id, original_id);
+        assert_eq!(manager.heal_queue.lock().await.len(), 1);
+    }
+}
+
+#[tokio::test]
+async fn mrf_fallback_same_incarnation_merges_but_version_and_scope_remain_distinct() {
+    let manager = HealManager::new_without_root_recovery_for_test(Arc::new(MockStorage), None);
+    let incarnation = Uuid::new_v4();
+    let original = incarnation_bound_mrf_fallback(incarnation);
+    let original_id = original.id.clone();
+    assert_eq!(
+        manager
+            .submit_heal_request_with_receipt(original)
+            .await
+            .expect("first fallback")
+            .result,
+        HealAdmissionResult::Accepted
+    );
+    let same = manager
+        .submit_heal_request_with_receipt(incarnation_bound_mrf_fallback(incarnation))
+        .await
+        .expect("same-incarnation fallback");
+    assert_eq!(same.result, HealAdmissionResult::Merged);
+    assert_eq!(same.task_id, original_id);
+
+    let mut versioned = incarnation_bound_mrf_fallback(incarnation);
+    versioned.heal_type = HealType::Object {
+        bucket: "mrf-recreated-bucket".to_owned(),
+        object: "object".to_owned(),
+        version_id: Some(Uuid::new_v4().to_string()),
+    };
+    assert_eq!(
+        manager
+            .submit_heal_request_with_receipt(versioned)
+            .await
+            .expect("different version")
+            .result,
+        HealAdmissionResult::Accepted
+    );
+    let mut other_scope = incarnation_bound_mrf_fallback(incarnation);
+    other_scope.options.pool_index = Some(1);
+    assert_eq!(
+        manager
+            .submit_heal_request_with_receipt(other_scope)
+            .await
+            .expect("different pool")
+            .result,
+        HealAdmissionResult::Accepted
+    );
+    let mut same_id_other_incarnation = incarnation_bound_mrf_fallback(Uuid::new_v4());
+    same_id_other_incarnation.id = original_id;
+    assert_eq!(
+        manager
+            .submit_heal_request_with_receipt(same_id_other_incarnation)
+            .await
+            .expect("same-ID conflict must be decided")
+            .result,
+        HealAdmissionResult::Dropped(HealAdmissionDropReason::AlreadyRunning)
+    );
+}
+
+#[test]
+fn mrf_bound_dedup_keys_separate_incarnation_source_and_untrusted_components() {
+    let incarnation = Uuid::new_v4();
+    let bound = incarnation_bound_mrf_fallback(incarnation);
+    let mut unbound = bound.clone();
+    unbound.expected_mrf_bucket_incarnation_id = None;
+    let mut non_mrf = bound.clone();
+    non_mrf.source = HealRequestSource::ReadRepair;
+    assert_ne!(PriorityHealQueue::make_dedup_key(&bound), PriorityHealQueue::make_dedup_key(&unbound));
+    assert_eq!(
+        PriorityHealQueue::make_dedup_key(&non_mrf),
+        PriorityHealQueue::make_dedup_key_for_scope(&non_mrf.heal_type, &non_mrf.options),
+        "an unrelated source retains its previous scope key even if the MRF-only field is set"
+    );
+    assert_eq!(PriorityHealQueue::make_dedup_key(&non_mrf), PriorityHealQueue::make_dedup_key(&unbound));
+
+    let mut left = bound.clone();
+    left.heal_type = HealType::Object {
+        bucket: "mrf-recreated-bucket".to_owned(),
+        object: "part:a".to_owned(),
+        version_id: Some("b".to_owned()),
+    };
+    let mut right = bound.clone();
+    right.heal_type = HealType::Object {
+        bucket: "mrf-recreated-bucket".to_owned(),
+        object: "part".to_owned(),
+        version_id: Some("a:b".to_owned()),
+    };
+    assert_eq!(
+        PriorityHealQueue::make_dedup_key_for_scope(&left.heal_type, &left.options),
+        PriorityHealQueue::make_dedup_key_for_scope(&right.heal_type, &right.options),
+        "the legacy delimiter-only key demonstrates the collision"
+    );
+    assert_ne!(
+        PriorityHealQueue::make_dedup_key(&left),
+        PriorityHealQueue::make_dedup_key(&right),
+        "length-prefixed object and version components must not alias"
+    );
+    let mut forged = unbound;
+    forged.heal_type = HealType::Object {
+        bucket: "mrf-recreated-bucket".to_owned(),
+        object: PriorityHealQueue::make_dedup_key(&bound),
+        version_id: None,
+    };
+    assert_ne!(PriorityHealQueue::make_dedup_key(&bound), PriorityHealQueue::make_dedup_key(&forged));
+}
+
+#[test]
+fn mrf_bound_optional_versions_do_not_alias_empty_or_forged_components() {
+    let incarnation = Uuid::new_v4();
+    for decode in [false, true] {
+        let make_request = |object: &str, version_id: Option<&str>| {
+            let mut request = incarnation_bound_mrf_fallback(incarnation);
+            let bucket = "mrf-recreated-bucket".to_owned();
+            let object = object.to_owned();
+            let version_id = version_id.map(str::to_owned);
+            request.heal_type = if decode {
+                HealType::ECDecode {
+                    bucket,
+                    object,
+                    version_id,
+                }
+            } else {
+                HealType::Object {
+                    bucket,
+                    object,
+                    version_id,
+                }
+            };
+            request
+        };
+        let absent = make_request("object", None);
+        let empty = make_request("object", Some(""));
+        let forged_object = make_request("object4:some0:", None);
+        let forged_version = make_request("object", Some("4:none"));
+        let keys = [&absent, &empty, &forged_object, &forged_version].map(PriorityHealQueue::make_dedup_key);
+        for (index, key) in keys.iter().enumerate() {
+            assert!(keys.iter().skip(index + 1).all(|other| key != other));
+        }
+
+        let mut queue = PriorityHealQueue::new();
+        assert_eq!(queue.push(absent.clone()), QueuePushOutcome::Accepted);
+        assert_eq!(queue.push(empty.clone()), QueuePushOutcome::Accepted);
+        assert_eq!(queue.push(make_request("object", None)), QueuePushOutcome::Merged);
+        assert_eq!(queue.push(make_request("object", Some(""))), QueuePushOutcome::Merged);
+        assert_eq!(queue.len(), 2);
+
+        for source in [HealRequestSource::Mrf, HealRequestSource::ReadRepair] {
+            let mut legacy_absent = absent.clone();
+            let mut legacy_empty = empty.clone();
+            legacy_absent.source = source;
+            legacy_empty.source = source;
+            if source == HealRequestSource::Mrf {
+                legacy_absent.expected_mrf_bucket_incarnation_id = None;
+                legacy_empty.expected_mrf_bucket_incarnation_id = None;
+            }
+            assert_eq!(
+                PriorityHealQueue::make_dedup_key(&legacy_absent),
+                PriorityHealQueue::make_dedup_key(&legacy_empty),
+                "unbound and non-MRF work retain the legacy empty-version key"
+            );
+        }
+    }
+}
+
+#[test]
+fn mrf_bound_queue_releases_each_incarnation_representative_independently() {
+    let mut queue = PriorityHealQueue::new();
+    let old = incarnation_bound_mrf_fallback(Uuid::new_v4());
+    let old_id = old.id.clone();
+    let old_key = PriorityHealQueue::make_dedup_key(&old);
+    let new = incarnation_bound_mrf_fallback(Uuid::new_v4());
+    let new_id = new.id.clone();
+    let new_key = PriorityHealQueue::make_dedup_key(&new);
+    let mut forced_old = old.clone();
+    forced_old.id = Uuid::new_v4().to_string();
+    forced_old.force_start = true;
+    let forced_id = forced_old.id.clone();
+
+    assert_eq!(queue.push(old), QueuePushOutcome::Accepted);
+    assert_eq!(queue.push(new), QueuePushOutcome::Accepted);
+    assert_eq!(queue.push(forced_old), QueuePushOutcome::Accepted);
+    assert_eq!(queue.queued_request_id_for_dedup_key(&old_key), Some(old_id.as_str()));
+    assert_eq!(queue.queued_request_id_for_dedup_key(&new_key), Some(new_id.as_str()));
+    queue
+        .remove_request_id(&old_id)
+        .expect("remove the old incarnation representative");
+    assert_eq!(queue.queued_request_id_for_dedup_key(&old_key), Some(forced_id.as_str()));
+    assert_eq!(queue.queued_request_id_for_dedup_key(&new_key), Some(new_id.as_str()));
+    assert_eq!(queue.pop_next().expect("new incarnation stays queued").id, new_id);
+    assert!(queue.queued_request_id_for_dedup_key(&new_key).is_none());
+    assert_eq!(queue.pop_next().expect("forced old owner stays queued").id, forced_id);
+    assert!(queue.queued_request_id_for_dedup_key(&old_key).is_none());
+}
+
+#[tokio::test]
+async fn durable_mrf_recreated_bucket_is_not_merged_into_queued_old_fallback() {
+    use rustfs_common::mrf_channel::{MrfIngressResult, MrfIntent, MrfKind, MrfScope, try_rearm_mrf_replay_intent};
+
+    let manager = HealManager::new_without_root_recovery_for_test(Arc::new(MockStorage), None);
+    let old = incarnation_bound_mrf_fallback(Uuid::new_v4());
+    assert_eq!(
+        manager
+            .submit_heal_request_with_receipt(old)
+            .await
+            .expect("old fallback")
+            .result,
+        HealAdmissionResult::Accepted
+    );
+    let incarnation = Uuid::new_v4();
+    let request = incarnation_bound_mrf_fallback(incarnation);
+    let request_id = request.id.clone();
+    let mut intent = MrfIntent {
+        bucket: Arc::from("mrf-recreated-bucket"),
+        object: Arc::from("object"),
+        version_id: None,
+        kind: MrfKind::PartialWrite,
+        delete_marker_purge: None,
+        scope: Some(MrfScope {
+            pool_index: 0,
+            set_index: 0,
+        }),
+        lease: None,
+        enqueued_at_ms: 1,
+        attempts: 0,
+    };
+    assert_eq!(try_rearm_mrf_replay_intent(&mut intent), MrfIngressResult::Enqueued);
+    let anchor = rustfs_common::mrf_channel::MrfDurableRepairAnchor::from_intent(&intent, incarnation).expect("durable identity");
+    let target = MrfRepairNoticeTarget {
+        bucket: Arc::clone(&intent.bucket),
+        object: Arc::clone(&intent.object),
+        version_id: intent.version_id,
+        kind: intent.kind,
+        scope: intent.scope,
+        delete_marker_purge: None,
+        lease: intent.lease,
+        durable_anchor: Some(anchor.clone()),
+    };
+    let receipt = manager
+        .submit_mrf_heal_request_with_receipt_and_identity(request, target)
+        .await
+        .expect("durable admission must be decided");
+    assert_eq!(
+        receipt.result,
+        HealAdmissionResult::Accepted,
+        "a durable new incarnation needs its own owner"
+    );
+    assert_eq!(receipt.task_id, request_id);
+    assert!(
+        manager
+            .heal_queue
+            .lock()
+            .await
+            .requests()
+            .any(|queued| { queued.id == request_id && queued.expected_mrf_bucket_incarnation_id == Some(incarnation) })
+    );
+    assert!(
+        manager
+            .try_in_flight_durable_mrf_anchors()
+            .expect("stable owner snapshot")
+            .contains(&anchor)
+    );
+}
+
+#[tokio::test]
+async fn durable_mrf_same_incarnation_merge_retains_the_durable_anchor() {
+    use rustfs_common::mrf_channel::{MrfIngressResult, MrfIntent, MrfKind, MrfScope, try_rearm_mrf_replay_intent};
+
+    let manager = HealManager::new_without_root_recovery_for_test(Arc::new(MockStorage), None);
+    let incarnation = Uuid::new_v4();
+    let object = format!("same-generation-{}", Uuid::new_v4());
+    let make_request = || {
+        let mut request = incarnation_bound_mrf_fallback(incarnation);
+        request.heal_type = HealType::Object {
+            bucket: "mrf-recreated-bucket".to_owned(),
+            object: object.clone(),
+            version_id: None,
+        };
+        request
+    };
+    let old = make_request();
+    let old_id = old.id.clone();
+    assert_eq!(
+        manager
+            .submit_heal_request_with_receipt(old)
+            .await
+            .expect("old fallback")
+            .result,
+        HealAdmissionResult::Accepted
+    );
+    let mut intent = MrfIntent {
+        bucket: Arc::from("mrf-recreated-bucket"),
+        object: Arc::from(object.as_str()),
+        version_id: None,
+        kind: MrfKind::PartialWrite,
+        delete_marker_purge: None,
+        scope: Some(MrfScope {
+            pool_index: 0,
+            set_index: 0,
+        }),
+        lease: None,
+        enqueued_at_ms: 1,
+        attempts: 0,
+    };
+    assert_eq!(try_rearm_mrf_replay_intent(&mut intent), MrfIngressResult::Enqueued);
+    let anchor = rustfs_common::mrf_channel::MrfDurableRepairAnchor::from_intent(&intent, incarnation).expect("durable identity");
+    let target = MrfRepairNoticeTarget {
+        bucket: Arc::clone(&intent.bucket),
+        object: Arc::clone(&intent.object),
+        version_id: intent.version_id,
+        kind: intent.kind,
+        scope: intent.scope,
+        delete_marker_purge: None,
+        lease: intent.lease,
+        durable_anchor: Some(anchor.clone()),
+    };
+    let merged = manager
+        .submit_mrf_heal_request_with_receipt_and_identity(make_request(), target)
+        .await
+        .expect("durable same-incarnation admission");
+    assert_eq!(merged.result, HealAdmissionResult::Merged);
+    assert_eq!(merged.task_id, old_id);
+    assert!(
+        manager
+            .try_in_flight_durable_mrf_anchors()
+            .expect("stable owner snapshot")
+            .contains(&anchor)
+    );
+}
+
 fn completed_retention_fixture(completed_at: SystemTime) -> CompletedHealStatus {
    CompletedHealStatus {
        task_id: "fixture-token".to_string(),
@@ -480,8 +480,9 @@ impl HealTask {
                message: format!("Missing source bucket incarnation for durable MRF repair {bucket}/{object}"),
            })?;
        if current_bucket_incarnation_id != bucket_incarnation_id {
-            return Err(Error::TaskExecutionFailed {
-                message: format!("Bucket incarnation changed before durable MRF repair {bucket}/{object}"),
+            return Err(Error::StaleBucketIncarnation {
+                bucket: bucket.to_owned(),
+                expected: Some(bucket_incarnation_id),
            });
        }
        expected.bucket_incarnation_id = Some(bucket_incarnation_id);
@@ -3375,10 +3375,67 @@ async fn durable_mrf_heal_rejects_a_recreated_bucket_before_storage_heal() {

    let result = task.heal_object("bucket", "object", None).await;

-    assert!(matches!(result, Err(Error::TaskExecutionFailed { .. })));
+    assert!(matches!(
+        result,
+        Err(Error::StaleBucketIncarnation { expected: Some(id), .. }) if id == original_incarnation
+    ));
    assert!(storage.heal_object_calls.lock().unwrap().is_empty());
 }

+#[tokio::test]
+async fn durable_mrf_stale_incarnation_object_name_cannot_change_retry_classification() {
+    for object in ["timed out", "network error", "ordinary"] {
+        let original_incarnation = Uuid::new_v4();
+        let recreated_incarnation = Uuid::new_v4();
+        let storage = Arc::new(MockStorage {
+            bucket_incarnation_id: Mutex::new(Some(recreated_incarnation)),
+            ..Default::default()
+        });
+        let mut request = HealRequest::object("bucket".to_owned(), object.to_owned(), None);
+        request.source = HealRequestSource::Mrf;
+        request.expected_mrf_bucket_incarnation_id = Some(original_incarnation);
+        request.options.pool_index = Some(0);
+        request.options.set_index = Some(1);
+        let task = HealTask::from_request(request, storage.clone());
+
+        let error = task
+            .execute()
+            .await
+            .expect_err("a recreated bucket must reject the old MRF responsibility");
+        assert!(
+            !error.is_recoverable_heal(),
+            "user-controlled object names must not make a stale incarnation retryable: {object}"
+        );
+        assert!(
+            matches!(
+                &error,
+                Error::StaleBucketIncarnation { bucket, expected }
+                    if bucket == "bucket" && *expected == Some(original_incarnation)
+            ),
+            "the failure must retain the original bucket incarnation: {error:?}"
+        );
+        assert!(
+            matches!(crate::error::classify_heal_error(&error), crate::error::HealErrorClass::Aborted),
+            "stale responsibility must be classified independently of the object name"
+        );
+        assert!(
+            storage.heal_object_calls.lock().expect("object heal calls").is_empty(),
+            "an old responsibility must not reach storage repair in the recreated bucket"
+        );
+        let outcome = task.get_outcome().await;
+        assert_eq!(outcome.counters.healed, 0, "a stale task cannot claim a repair");
+        assert!(
+            outcome.objects.iter().all(|item| !matches!(
+                &item.disposition,
+                HealObjectDisposition::Repaired
+                    | HealObjectDisposition::VerifiedHealthy
+                    | HealObjectDisposition::AuthoritativelyAbsent
+            )),
+            "the stale task must not produce a positive repair proof"
+        );
+    }
+}
+
 #[tokio::test(start_paused = true)]
 async fn read_repair_object_heal_is_not_failed_by_flat_task_timeout() {
    let storage = Arc::new(MockStorage {
@@ -16,9 +16,10 @@ use super::{
    transport::{DiscoveryRequest, OidcHttpError, ReqwestHttpClient, oidc_extra_root_certs, oidc_http_error_diagnostics},
 };
 use crate::{EVENT_OIDC_DIAGNOSTICS, LOG_COMPONENT_IAM, LOG_SUBSYSTEM_OIDC};
-use openidconnect::core::{CoreIdTokenVerifier, CoreJsonWebKeySet, CoreJwsSigningAlgorithm};
+use openidconnect::core::{CoreIdToken, CoreIdTokenClaims, CoreIdTokenVerifier, CoreJsonWebKeySet, CoreJwsSigningAlgorithm};
 use openidconnect::{
-    AsyncHttpClient, ClientId, ClientSecret, DiscoveryError, IssuerUrl, JsonWebKeySetUrl, ProviderMetadataWithLogout, TokenUrl,
+    AsyncHttpClient, ClaimsVerificationError, ClientId, ClientSecret, DiscoveryError, IssuerUrl, JsonWebKeySetUrl, NonceVerifier,
+    ProviderMetadataWithLogout, SignatureVerificationError, TokenUrl,
 };
 use serde::Deserialize;
 use std::collections::HashMap;
@@ -212,6 +213,35 @@ impl DiscoveredProviderMetadata {
    }
 }

+pub(super) fn verify_id_token_claims<'a, N: NonceVerifier>(
+    token: &'a CoreIdToken,
+    verifier: &CoreIdTokenVerifier<'_>,
+    nonce_verifier: N,
+) -> Result<&'a CoreIdTokenClaims, ClaimsVerificationError> {
+    let expected_length = match token.signing_alg().map_err(ClaimsVerificationError::SignatureVerification)? {
+        CoreJwsSigningAlgorithm::HmacSha256 => 32,
+        CoreJwsSigningAlgorithm::HmacSha384 => 48,
+        CoreJwsSigningAlgorithm::HmacSha512 => 64,
+        _ => return token.claims(verifier, nonce_verifier),
+    };
+    // RUSTFS_COMPAT_TODO(openidconnect-hmac-signature-width): Reject widths that upstream converts to fixed-size arrays. Remove after upstream returns verification errors for arbitrary HMAC signature lengths.
+    // openidconnect 4.1.0 converts HMAC signatures to fixed-size arrays before verifying them.
+    let invalid_length = || {
+        ClaimsVerificationError::SignatureVerification(SignatureVerificationError::CryptoError(
+            "invalid HMAC signature length".to_owned(),
+        ))
+    };
+    let compact = token.to_string();
+    let (_, signature) = compact.rsplit_once('.').ok_or_else(invalid_length)?;
+    let signature = base64_simd::URL_SAFE_NO_PAD
+        .decode_to_vec(signature)
+        .map_err(|_| invalid_length())?;
+    if signature.len() != expected_length {
+        return Err(invalid_length());
+    }
+    token.claims(verifier, nonce_verifier)
+}
+
 // This adapter is used only for discovery/JWKS fetches, never token exchange.
 struct JwksAcceptClient<'a> {
    inner: &'a ReqwestHttpClient,
@@ -242,7 +242,7 @@ fn configured_verifier_rejects_malformed_hmac_signature_lengths() {
    let now = time::OffsetDateTime::now_utc().unix_timestamp();
    let payload = serde_json::json!({
        "iss": config.config_url, "sub": "synthetic-subject", "aud": config.client_id,
-        "iat": now, "exp": now + 300,
+        "iat": now, "exp": now + 300, "nonce": "synthetic-nonce",
    });

    for (algorithm, name, signature_length) in [
@@ -275,22 +275,28 @@ fn configured_verifier_rejects_malformed_hmac_signature_lengths() {
                .expect("parse configured provider metadata");
            let verifier = metadata.verifier(&config);
            assert!(
-                valid.claims(&verifier, |_: Option<&Nonce>| Ok(())).is_ok(),
+                verify_id_token_claims(&valid, &verifier, |_: Option<&Nonce>| Ok(())).is_ok(),
                "valid {name} token must pass for workload={workload}"
            );
+            assert!(verify_id_token_claims(&valid, &verifier, &Nonce::new("synthetic-nonce".into())).is_ok());
+            assert!(verify_id_token_claims(&valid, &verifier, &Nonce::new("wrong-nonce".into())).is_err());

            for length in [1, 0, signature_length - 1, signature_length, signature_length + 1, 128] {
                let signature = base64_simd::URL_SAFE_NO_PAD.encode_to_string(vec![0_u8; length]);
                let malformed: CoreIdToken = format!("{signing_input}.{signature}")
                    .parse()
-                    .expect("invalid signature length must still reach cryptographic verification");
+                    .expect("invalid signature length must remain parseable before verification");
                assert!(
-                    malformed.claims(&verifier, |_: Option<&Nonce>| Ok(())).is_err(),
+                    verify_id_token_claims(&malformed, &verifier, |_: Option<&Nonce>| Ok(())).is_err(),
                    "{name} signature length {length} must return an error for workload={workload}"
                );
+                assert!(
+                    verify_id_token_claims(&malformed, &verifier, &Nonce::new("synthetic-nonce".into())).is_err(),
+                    "{name} signature length {length} must return an error with a nonce for workload={workload}"
+                );
            }
            assert!(
-                valid.claims(&verifier, |_: Option<&Nonce>| Ok(())).is_ok(),
+                verify_id_token_claims(&valid, &verifier, |_: Option<&Nonce>| Ok(())).is_ok(),
                "invalid inputs must not prevent later valid {name} verification"
            );
        }
@@ -16,7 +16,7 @@ use super::{
        OidcConfigQuery, OidcConfigSnapshot, OidcProviderConfig, OidcProviderSummary, OidcSiteReplicationSnapshot,
        SourcedOidcProviderConfig, load_oidc_config_snapshot,
    },
-    provider::{ProviderRuntime, discover_provider_for_recovery},
+    provider::{ProviderRuntime, discover_provider_for_recovery, verify_id_token_claims},
    state::{OidcAuthSession, OidcLogoutSession, OidcStateStore},
    transport::{OidcExtraRootCaProvider, ReqwestHttpClient, inspect_token_response_body, oidc_http_error_diagnostics},
 };
@@ -535,7 +535,7 @@ impl OidcSys {
        let verifier = client
            .id_token_verifier()
            .set_other_audience_verifier_fn(|aud| trusted_aud(&config.other_audiences, aud));
-        let verified = id_token.claims(&verifier, &Nonce::new(session.nonce.clone()));
+        let verified = verify_id_token_claims(id_token, &verifier, &Nonce::new(session.nonce.clone()));
        if let Err(e) = verified {
            let verification_error = e.to_string();
            warn!(
@@ -583,8 +583,7 @@ impl OidcSys {
            let verifier = client
                .id_token_verifier()
                .set_other_audience_verifier_fn(|aud| trusted_aud(&config.other_audiences, aud));
-            id_token
-                .claims(&verifier, &Nonce::new(session.nonce.clone()))
+            verify_id_token_claims(id_token, &verifier, &Nonce::new(session.nonce.clone()))
                .map_err(|retry_err| {
                    error!(
                        event = EVENT_OIDC_DIAGNOSTICS,
@@ -767,7 +766,7 @@ impl OidcSys {
            .metadata
            .verifier(&config)
            .set_other_audience_verifier_fn(|aud| trusted_aud(&config.other_audiences, aud));
-        if let Err(e) = id_token.claims(&verifier, |_: Option<&Nonce>| Ok(())) {
+        if let Err(e) = verify_id_token_claims(&id_token, &verifier, |_: Option<&Nonce>| Ok(())) {
            state = self
                .provider_runtime
                .refresh_provider_state(&provider_id, &config)
@@ -780,8 +779,7 @@ impl OidcSys {
                .metadata
                .verifier(&config)
                .set_other_audience_verifier_fn(|aud| trusted_aud(&config.other_audiences, aud));
-            id_token
-                .claims(&verifier, |_: Option<&Nonce>| Ok(()))
+            verify_id_token_claims(&id_token, &verifier, |_: Option<&Nonce>| Ok(()))
                .map_err(|retry_err| format!("ID token verification failed after JWKS refresh: {retry_err}"))?;
        }

@@ -1411,6 +1411,28 @@ pub fn record_peer_reachable(addr: &str) {
    publish_offline_gauge(&peers);
 }

+/// Record a successful interaction with a cluster peer, but only when the peer is currently
+/// marked offline. This is the fast recovery path for cached-channel RPCs, recovery probes,
+/// and lock-path successes that do not go through a fresh dial.
+///
+/// Fast path (peer unknown or already online): a shared read lock, no write, no gauge
+/// republish — the overwhelmingly common case on a healthy cluster. Slow path (offline →
+/// online transition): upgrades to the write lock and delegates to [`record_peer_reachable`].
+pub fn record_peer_reachable_if_offline(addr: &str) {
+    let key = normalize_peer_key(addr);
+    // Fast path: read lock to check whether the peer is currently offline.
+    {
+        let peers = CLUSTER_PEER_HEALTH.read().unwrap_or_else(|poisoned| poisoned.into_inner());
+        match peers.get(key) {
+            None => return,
+            Some(entry) if entry.online => return,
+            Some(_) => {} // offline: fall through to the write path
+        }
+    }
+    // Slow path: the peer is offline, so flip it back online and republish the gauge.
+    record_peer_reachable(addr);
+}
+
 /// Record a failed interaction with a cluster peer (dial failure or RPC-triggered eviction). After
 /// `failure_threshold` (>= 1) consecutive failures the peer flips offline.
 pub fn record_peer_unreachable(addr: &str, failure_threshold: u32) {
@@ -2188,6 +2210,34 @@ mod tests {
        assert_eq!(CLUSTER_SERVERS_OFFLINE_TOTAL, "rustfs_cluster_servers_offline_total");
    }

+    #[test]
+    fn record_peer_reachable_if_offline_recovers_stale_offline_state() {
+        let addr = "http://cluster-peer-recover-if-offline-test:9000";
+
+        // Mark the peer offline.
+        record_peer_unreachable(addr, 1);
+        assert!(cluster_peer_is_offline(addr));
+
+        // A successful cached-channel RPC (no fresh dial) clears the offline state.
+        record_peer_reachable_if_offline(addr);
+        assert!(!cluster_peer_is_offline(addr));
+        assert_eq!(cluster_peer_observed_online_status(addr), Some(true));
+    }
+
+    #[test]
+    fn record_peer_reachable_if_offline_no_op_for_online_peer() {
+        let addr = "http://cluster-peer-recover-noop-test:9000";
+
+        record_peer_reachable(addr);
+        // Should not panic or change state for an already-online peer.
+        record_peer_reachable_if_offline(addr);
+        assert_eq!(cluster_peer_observed_online_status(addr), Some(true));
+
+        // Should also be a no-op for unknown peers.
+        record_peer_reachable_if_offline("http://cluster-peer-recover-unknown-test:9000");
+        assert_eq!(cluster_peer_observed_online_status("http://cluster-peer-recover-unknown-test:9000"), None);
+    }
+
    #[test]
    fn cluster_peer_should_bypass_is_self_healing() {
        let addr = "http://cluster-peer-bypass-selfheal-test:9000";
@@ -30,6 +30,7 @@ workspace = true

 [features]
 default = []
+fips = ["rustfs-tls-runtime?/fips"]
 hotpath = [
    "hotpath/hotpath",
    "hotpath/tokio",
@@ -110,6 +110,9 @@ where
            ftps_required = self.config.ftps_required,
            "FTPS server starting"
        );
+        if self.config.tls_enabled && self.config.cert_dir.is_some() {
+            rustfs_tls_runtime::ensure_crypto_provider().map_err(|err| FtpsInitError::InvalidConfig(err.to_string()))?;
+        }
        let (reload_shutdown_tx, reload_shutdown_rx) = watch::channel(false);

        let storage_clone = self.storage.clone();
@@ -191,8 +194,6 @@ where
                );

                // Build ServerConfig with SNI support
-                let _ = rustls::crypto::aws_lc_rs::default_provider().install_default();
-
                let server_config = rustls::ServerConfig::builder()
                    .with_no_client_auth()
                    .with_cert_resolver(resolver);
@@ -43,3 +43,70 @@ pub use webdav::{config::WebDavConfig, server::WebDavServer};

 #[cfg(feature = "sftp")]
 pub use sftp::{SftpConfig, SftpInitError, SftpServer};
+
+#[cfg(all(test, feature = "fips", feature = "ftps", feature = "webdav"))]
+mod fips_tests {
+    use crate::common::dummy_storage::DummyBackend;
+    use crate::{FtpsConfig, FtpsServer, WebDavConfig, WebDavServer};
+
+    #[test]
+    fn protocols_reject_incompatible_host_provider() {
+        const CHILD: &str = "RUSTFS_TEST_PROTOCOL_FIPS_PROVIDER";
+        if std::env::var_os(CHILD).is_some() {
+            let mut provider = rustfs_tls_runtime::default_crypto_provider();
+            provider.kx_groups = vec![rustls::crypto::aws_lc_rs::kx_group::X25519];
+            provider.install_default().expect("install incompatible host provider");
+            let cert_dir = tempfile::tempdir().expect("certificate directory");
+            let cert_dir = cert_dir.path().to_string_lossy().into_owned();
+            let runtime = tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+                .expect("test runtime");
+            runtime.block_on(async {
+                let ftps_config = FtpsConfig {
+                    cert_dir: Some(cert_dir.clone()),
+                    ..Default::default()
+                };
+                let ftps = FtpsServer::new(ftps_config, DummyBackend::default())
+                    .await
+                    .expect("FTPS config");
+                let (_sender, receiver) = tokio::sync::broadcast::channel(1);
+                let error = ftps
+                    .start(receiver)
+                    .await
+                    .expect_err("FTPS must reject the incompatible provider");
+                assert!(error.to_string().contains("process-wide rustls crypto provider"));
+
+                let webdav_config = WebDavConfig {
+                    bind_addr: ([127, 0, 0, 1], 0).into(),
+                    cert_dir: Some(cert_dir),
+                    ..Default::default()
+                };
+                let webdav = WebDavServer::new(webdav_config, DummyBackend::default())
+                    .await
+                    .expect("WebDAV config");
+                let (_sender, receiver) = tokio::sync::broadcast::channel(1);
+                let error = webdav
+                    .start(receiver)
+                    .await
+                    .expect_err("WebDAV must reject the incompatible provider");
+                assert!(error.to_string().contains("process-wide rustls crypto provider"));
+            });
+            return;
+        }
+        let output = std::process::Command::new(std::env::current_exe().expect("test executable"))
+            .arg("protocols_reject_incompatible_host_provider")
+            .arg("--nocapture")
+            .env(CHILD, "1")
+            .env("NO_PROXY", "localhost,127.0.0.1,::1")
+            .env("no_proxy", "localhost,127.0.0.1,::1")
+            .output()
+            .expect("start isolated protocol TLS test");
+        assert!(
+            output.status.success(),
+            "stdout={} stderr={}",
+            String::from_utf8_lossy(&output.stdout),
+            String::from_utf8_lossy(&output.stderr)
+        );
+    }
+}
@@ -118,6 +118,10 @@ where
            "WebDAV server starting"
        );

+        if self.config.tls_enabled && self.config.cert_dir.is_some() {
+            rustfs_tls_runtime::ensure_crypto_provider().map_err(|err| WebDavInitError::Tls(err.to_string()))?;
+        }
+
        let listener = TcpListener::bind(self.config.bind_addr).await?;
        self.serve(listener, shutdown_rx).await
    }
@@ -155,8 +159,6 @@ where
                    reload_shutdown_rx.clone(),
                );

-                let _ = rustls::crypto::aws_lc_rs::default_provider().install_default();
-
                let server_config = ServerConfig::builder().with_no_client_auth().with_cert_resolver(resolver);

                Some(TlsAcceptor::from(Arc::new(server_config)))
@@ -25,6 +25,10 @@ keywords = ["s3", "client", "tiering", "rustfs", "Minio"]
 categories = ["web-programming", "development-tools", "network-programming"]
 documentation = "https://docs.rs/rustfs-s3-client/latest/rustfs_s3_client/"

+[features]
+default = []
+fips = ["rustfs-tls-runtime/fips"]
+
 [dependencies]
 rustfs-checksums.workspace = true
 rustfs-config.workspace = true
@@ -81,7 +81,7 @@ use time::Duration;
 use time::OffsetDateTime;
 use tokio::io::BufReader;
 use tokio::io::{AsyncRead, AsyncReadExt};
-use tracing::{debug, error, trace, warn};
+use tracing::{error, trace, warn};
 use url::{Url, form_urlencoded};
 use uuid::Uuid;

@@ -448,16 +448,7 @@ impl TransitionClient {
        tier_type: &str,
        timeouts: TransitionClientTimeouts,
    ) -> Result<TransitionClient, std::io::Error> {
-        if rustls::crypto::CryptoProvider::get_default().is_none() {
-            // No default provider is set yet; try to install aws-lc-rs.
-            // `install_default` can only fail if another thread races us and installs a provider
-            // between our check and this call, which is still safe to ignore.
-            if rustls::crypto::aws_lc_rs::default_provider().install_default().is_err() {
-                debug!("rustls crypto provider was installed concurrently, skipping aws-lc-rs install");
-            }
-        } else {
-            debug!("rustls crypto provider already installed, skipping aws-lc-rs install");
-        }
+        rustfs_tls_runtime::ensure_crypto_provider().map_err(std::io::Error::other)?;

        let endpoint_url = get_endpoint_url(endpoint, opts.secure)?;
        let timeouts = timeouts.validate()?;
@@ -1862,21 +1853,36 @@ mod tests {
        assert!(outcome.is_ok(), "TLS config creation should not panic");
    }

-    /// Installing the rustls crypto provider when one is already set must not panic or return
-    /// an error that surfaces to callers (the race-safe `get_default` check guards the install).
+    #[cfg(feature = "fips")]
    #[test]
-    fn provider_install_is_idempotent() {
-        // Install once (may already be set by another test in this binary — that's fine).
-        let _ = rustls::crypto::aws_lc_rs::default_provider().install_default();
-
-        // A second install attempt on an already-set provider must not panic.
-        let outcome = std::panic::catch_unwind(|| {
-            if rustls::crypto::CryptoProvider::get_default().is_none() {
-                let _ = rustls::crypto::aws_lc_rs::default_provider().install_default();
-            }
-            // If a default is already present, the branch above is simply skipped.
-        });
-        assert!(outcome.is_ok(), "provider install guard must not panic when a provider is already set");
+    fn fips_transition_rejects_incompatible_host_provider() {
+        const CHILD: &str = "RUSTFS_TEST_TRANSITION_FIPS_PROVIDER";
+        if std::env::var_os(CHILD).is_some() {
+            let mut provider = rustfs_tls_runtime::default_crypto_provider();
+            provider.kx_groups = vec![rustls::crypto::aws_lc_rs::kx_group::X25519];
+            provider.install_default().expect("install incompatible host provider");
+            let runtime = tokio::runtime::Builder::new_current_thread()
+                .enable_all()
+                .build()
+                .expect("test runtime");
+            let error = runtime
+                .block_on(super::TransitionClient::new("127.0.0.1:0", super::Options::default(), ""))
+                .expect_err("transition client must reject the incompatible provider");
+            assert!(error.to_string().contains("process-wide rustls crypto provider"));
+            return;
+        }
+        let output = std::process::Command::new(std::env::current_exe().expect("test executable"))
+            .arg("fips_transition_rejects_incompatible_host_provider")
+            .arg("--nocapture")
+            .env(CHILD, "1")
+            .output()
+            .expect("start isolated transition TLS test");
+        assert!(
+            output.status.success(),
+            "stdout={} stderr={}",
+            String::from_utf8_lossy(&output.stdout),
+            String::from_utf8_lossy(&output.stderr)
+        );
    }

    #[test]
@@ -640,12 +640,12 @@ fn http_range_spec_from_get_range(range: &GetRange) -> HTTPRangeSpec {
        GetRange::Bounded(range) => http_range_spec_from_range(range.clone()),
        GetRange::Offset(start) => HTTPRangeSpec {
            is_suffix_length: false,
-            start: *start as i64,
+            start: i64::try_from(*start).unwrap_or(i64::MAX),
            end: -1,
        },
        GetRange::Suffix(length) => HTTPRangeSpec {
            is_suffix_length: true,
-            start: *length as i64,
+            start: i64::try_from(*length).unwrap_or(i64::MAX),
            end: -1,
        },
    }
@@ -654,15 +654,15 @@ fn http_range_spec_from_get_range(range: &GetRange) -> HTTPRangeSpec {
 fn http_range_spec_from_range(range: Range<u64>) -> HTTPRangeSpec {
    HTTPRangeSpec {
        is_suffix_length: false,
-        start: range.start as i64,
-        end: range.end.saturating_sub(1) as i64,
+        start: i64::try_from(range.start).unwrap_or(i64::MAX),
+        end: i64::try_from(range.end.saturating_sub(1)).unwrap_or(i64::MAX),
    }
 }

 fn http_range_spec_from_start(start: u64) -> HTTPRangeSpec {
    HTTPRangeSpec {
        is_suffix_length: false,
-        start: start as i64,
+        start: i64::try_from(start).unwrap_or(i64::MAX),
        end: -1,
    }
 }
@@ -2476,6 +2476,48 @@ mod test {
        assert_eq!(range.end, 19);
    }

+    #[test]
+    fn test_get_range_conversion_for_suffix_ranges() {
+        // Normal suffix: last 100 bytes.
+        let range = http_range_spec_from_get_range(&GetRange::Suffix(100));
+        assert!(range.is_suffix_length);
+        assert_eq!(range.start, 100);
+        assert_eq!(range.end, -1);
+
+        // Overflowing suffix: `u64::MAX` previously wrapped to `-1` via `as i64`
+        // and was misread as "last 1 byte". It now clamps to `i64::MAX`, which
+        // `get_length` treats as "the whole object".
+        let range = http_range_spec_from_get_range(&GetRange::Suffix(u64::MAX));
+        assert!(range.is_suffix_length);
+        assert_eq!(range.start, i64::MAX);
+        assert_eq!(range.end, -1);
+    }
+
+    #[test]
+    fn test_get_range_conversion_for_offset_overflow() {
+        // Normal offset.
+        let range = http_range_spec_from_get_range(&GetRange::Offset(42));
+        assert!(!range.is_suffix_length);
+        assert_eq!(range.start, 42);
+        assert_eq!(range.end, -1);
+
+        // Overflowing offset clamps to `i64::MAX`, which is then rejected as
+        // out-of-range (416) by `get_offset_length`.
+        let range = http_range_spec_from_get_range(&GetRange::Offset(u64::MAX));
+        assert!(!range.is_suffix_length);
+        assert_eq!(range.start, i64::MAX);
+        assert_eq!(range.end, -1);
+    }
+
+    #[test]
+    fn test_get_range_conversion_for_bounded_overflow() {
+        // Overflowing bounded range clamps start/end to `i64::MAX`.
+        let range = http_range_spec_from_get_range(&GetRange::Bounded(u64::MAX - 10..u64::MAX));
+        assert!(!range.is_suffix_length);
+        assert_eq!(range.start, i64::MAX);
+        assert_eq!(range.end, i64::MAX);
+    }
+
    #[tokio::test]
    async fn test_scan_range_output_can_convert_field_delimiter() {
        let chunks = stream::iter(vec![Ok::<_, std::io::Error>(Bytes::from_static(b"a&&1\nb&&2\n"))]);
@@ -13,6 +13,7 @@ documentation = "https://docs.rs/rustfs-targets/latest/rustfs_targets/"

 [features]
 default = []
+fips = ["rustfs-tls-runtime/fips"]
 # Exposes the builder-style `testkit::MockTarget` to downstream test suites. Never enable this
 # from a production `[dependencies]` entry; in-crate unit tests get the module via cfg(test).
 test-support = []
@@ -162,7 +162,7 @@ pub async fn check_mysql_server_available(args: &crate::target::mysql::MySqlArgs
        .db_name(Some(dsn.database.clone()));

    if dsn.tls {
-        ensure_rustls_provider_installed();
+        ensure_rustls_provider_installed()?;
        let mut ssl_opts = SslOpts::default();
        if !args.tls_ca.is_empty() {
            ssl_opts = ssl_opts.with_root_certs(vec![PathBuf::from(args.tls_ca.clone()).into()]);
@@ -293,6 +293,10 @@ fn map_lapin_error(err: lapin::Error, context: &str) -> TargetError {

 pub async fn connect_amqp(args: &AMQPArgs) -> Result<AMQPConnection, TargetError> {
    args.validate()?;
+    #[cfg(feature = "fips")]
+    if args.url.scheme() == "amqps" {
+        super::ensure_rustls_provider_installed()?;
+    }
    tokio::time::timeout(Duration::from_secs(5), async {
        let url = connection_url(args)?;
        // Reconnect explicitly so every new channel enables publisher confirms below.
@@ -27,7 +27,7 @@ use std::sync::Arc;
 use std::sync::atomic::{AtomicBool, AtomicU64, Ordering};
 use std::thread_local;
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
-use tracing::{debug, warn};
+use tracing::warn;

 pub mod amqp;
 pub mod kafka;
@@ -947,16 +947,12 @@ pub(crate) fn encode_failed_entry(

 /// Ensures a rustls crypto provider is installed before any TLS operation.
 ///
-/// Multiple target modules (MySQL, Redis, Postgres, MQTT) need this because
-/// each may be the first to perform a TLS handshake. Idempotent: if a
-/// provider is already registered, returns immediately.
-pub(crate) fn ensure_rustls_provider_installed() {
-    if rustls::crypto::CryptoProvider::get_default().is_some() {
-        return;
-    }
-    if let Err(err) = rustls::crypto::aws_lc_rs::default_provider().install_default() {
-        debug!("rustls provider already installed or unavailable: {err:?}");
-    }
+/// Target modules may be the first to construct a TLS client. In a FIPS build,
+/// an incompatible provider installed by the host is rejected.
+pub(crate) fn ensure_rustls_provider_installed() -> Result<(), TargetError> {
+    rustfs_tls_runtime::ensure_crypto_provider()
+        .map(|_| ())
+        .map_err(|err| TargetError::Configuration(err.to_string()))
 }

 #[cfg(test)]
@@ -241,7 +241,7 @@ fn build_root_store(ca_path: &str, trust_leaf_as_ca: bool) -> Result<rustls::Roo
 }

 fn build_mqtt_tls_transport(broker: &Url, tls: &MQTTTlsConfig) -> Result<Transport, TargetError> {
-    super::ensure_rustls_provider_installed();
+    super::ensure_rustls_provider_installed()?;

    let client_config = match tls
        .policy
@@ -881,7 +881,7 @@ async fn build_mysql_pool_from_args(args: &MySqlArgs) -> Result<(Pool, bool), Ta
        .db_name(Some(dsn.database.clone()));

    if dsn.tls {
-        super::ensure_rustls_provider_installed();
+        super::ensure_rustls_provider_installed()?;
        let mut ssl_opts = SslOpts::default();
        if !args.tls_ca.is_empty() {
            let _ =
@@ -204,6 +204,8 @@ fn nats_sends_credentials_without_tls(args: &NATSArgs) -> bool {

 pub async fn connect_nats(args: &NATSArgs) -> Result<async_nats::Client, TargetError> {
    args.validate()?;
+    #[cfg(feature = "fips")]
+    super::ensure_rustls_provider_installed()?;

    let mut options = async_nats::ConnectOptions::new().require_tls(args.tls_required);

@@ -462,7 +462,7 @@ pub fn table_probe_sql(schema: &str, table: &str) -> String {
 /// When `tls_client_cert` and `tls_client_key` are both set the connection
 /// uses mTLS authentication; otherwise no client cert is sent.
 pub fn build_tls_config(args: &PostgresArgs) -> Result<rustls::ClientConfig, TargetError> {
-    super::ensure_rustls_provider_installed();
+    super::ensure_rustls_provider_installed()?;

    let mut root_store = rustls::RootCertStore::empty();

@@ -939,6 +939,32 @@ mod tests {
        }
    }

+    #[cfg(feature = "fips")]
+    #[test]
+    fn fips_postgres_rejects_incompatible_host_provider() {
+        const CHILD: &str = "RUSTFS_TEST_POSTGRES_FIPS_PROVIDER";
+        if std::env::var_os(CHILD).is_some() {
+            let mut provider = rustfs_tls_runtime::default_crypto_provider();
+            provider.kx_groups = vec![rustls::crypto::aws_lc_rs::kx_group::X25519];
+            provider.install_default().expect("install incompatible host provider");
+            let error = build_tls_config(&base_args()).expect_err("Postgres must reject the incompatible provider");
+            assert!(error.to_string().contains("process-wide rustls crypto provider"));
+            return;
+        }
+        let output = std::process::Command::new(std::env::current_exe().expect("test executable"))
+            .arg("fips_postgres_rejects_incompatible_host_provider")
+            .arg("--nocapture")
+            .env(CHILD, "1")
+            .output()
+            .expect("start isolated Postgres TLS test");
+        assert!(
+            output.status.success(),
+            "stdout={} stderr={}",
+            String::from_utf8_lossy(&output.stdout),
+            String::from_utf8_lossy(&output.stderr)
+        );
+    }
+
    #[test]
    fn validate_disabled_skips_all_checks() {
        let args = PostgresArgs {
@@ -90,7 +90,7 @@ impl PulsarArgs {
            return Ok(());
        }

-        validate_pulsar_broker(&self.broker)?;
+        let parsed = validate_pulsar_broker(&self.broker)?;

        if self.topic.trim().is_empty() {
            return Err(TargetError::Configuration("Pulsar topic cannot be empty".to_string()));
@@ -116,8 +116,6 @@ impl PulsarArgs {
            return Err(TargetError::Configuration("Pulsar queue directory must be an absolute path".to_string()));
        }

-        let parsed = Url::parse(&self.broker)
-            .map_err(|e| TargetError::Configuration(format!("Invalid Pulsar broker URL: {e} (value: '{}')", self.broker)))?;
        let tls_enabled = parsed.scheme() == "pulsar+ssl";
        // A CA bundle is TLS trust material a plaintext `pulsar://` broker
        // silently ignores, so treat it as a genuine misconfiguration. The
@@ -173,6 +171,10 @@ fn pulsar_producer_name(target_id: &TargetID, target_type: TargetType, nonce: Uu

 pub async fn connect_pulsar(args: &PulsarArgs) -> Result<Pulsar<TokioExecutor>, TargetError> {
    args.validate()?;
+    #[cfg(feature = "fips")]
+    if Url::parse(&args.broker).is_ok_and(|url| url.scheme() == "pulsar+ssl") {
+        super::ensure_rustls_provider_installed()?;
+    }

    let mut builder = Pulsar::builder(args.broker.clone(), TokioExecutor);

@@ -766,7 +766,7 @@ pub(crate) fn build_redis_client(args: &RedisArgs) -> Result<Client, TargetError

    let secure_scheme = matches!(args.url.scheme(), "rediss" | "valkeys");
    if secure_scheme {
-        super::ensure_rustls_provider_installed();
+        super::ensure_rustls_provider_installed()?;
        let tls_certs = TlsCertificates {
            client_tls: read_client_tls(&args.tls)?,
            root_cert: read_root_cert(&args.tls)?,
@@ -305,6 +305,8 @@ where
        args: &WebhookArgs,
        resolver: impl reqwest::dns::Resolve + 'static,
    ) -> Result<Client, TargetError> {
+        #[cfg(feature = "fips")]
+        super::ensure_rustls_provider_installed()?;
        let mut client_builder = Client::builder()
            .no_proxy()
            .dns_resolver(resolver)
@@ -848,6 +850,51 @@ mod tests {
        }
    }

+    #[test]
+    fn webhook_provider_policy_isolated() {
+        const CHILD: &str = "RUSTFS_TEST_WEBHOOK_PROVIDER_POLICY";
+        if std::env::var_os(CHILD).is_some() {
+            if cfg!(feature = "fips") {
+                let mut provider = rustfs_tls_runtime::default_crypto_provider();
+                provider.kx_groups = vec![rustls::crypto::aws_lc_rs::kx_group::X25519];
+                provider.install_default().expect("install incompatible host provider");
+            } else {
+                assert!(rustls::crypto::CryptoProvider::get_default().is_none());
+            }
+            let args = WebhookArgs {
+                skip_tls_verify: true,
+                ..base_args()
+            };
+            let result = WebhookTarget::<serde_json::Value>::build_http_client_with_resolver(
+                &args,
+                StaticResolver("127.0.0.1".parse().expect("loopback IP")),
+            );
+            if cfg!(feature = "fips") {
+                let error = result.expect_err("FIPS webhook must reject the incompatible provider");
+                assert!(error.to_string().contains("process-wide rustls crypto provider"));
+            } else {
+                result.expect("default webhook client should build");
+                assert!(
+                    rustls::crypto::CryptoProvider::get_default().is_none(),
+                    "building a default webhook must not install a process-wide provider"
+                );
+            }
+            return;
+        }
+        let output = std::process::Command::new(std::env::current_exe().expect("test executable"))
+            .arg("webhook_provider_policy_isolated")
+            .arg("--nocapture")
+            .env(CHILD, "1")
+            .output()
+            .expect("start isolated webhook provider test");
+        assert!(
+            output.status.success(),
+            "stdout={} stderr={}",
+            String::from_utf8_lossy(&output.stdout),
+            String::from_utf8_lossy(&output.stderr)
+        );
+    }
+
    async fn http_status_url(status: u16) -> Url {
        let listener = tokio::net::TcpListener::bind("127.0.0.1:0").await.expect("bind test server");
        let address = listener.local_addr().expect("test server address");
@@ -610,10 +610,12 @@ pub fn validate_certified_key(
            }
        }
    }
-    let certified = CertifiedKey::from_der(certs.to_vec(), key.clone_key(), &rustls::crypto::aws_lc_rs::default_provider())
-        .map_err(|source| TlsRuntimeError::InvalidCertifiedKey {
-            name: name.to_string(),
-            source,
+    let certified =
+        CertifiedKey::from_der(certs.to_vec(), key.clone_key(), &crate::default_crypto_provider()).map_err(|source| {
+            TlsRuntimeError::InvalidCertifiedKey {
+                name: name.to_string(),
+                source,
+            }
        })?;
    certified
        .keys_match()
@@ -21,6 +21,7 @@ pub mod fingerprint;
 pub mod material;
 pub mod metrics;
 pub mod outbound;
+pub mod provider;
 pub mod server;
 pub mod source;
 pub mod state;
@@ -45,6 +46,7 @@ pub use outbound::{
    GlobalOutboundTlsStateSummary, GlobalPublishedOutboundTlsState, load_global_outbound_tls_generation,
    load_global_outbound_tls_state, publish_global_outbound_tls_state, summarize_global_outbound_tls_state,
 };
+pub use provider::{CryptoProviderError, CryptoProviderStatus, default_crypto_provider, ensure_crypto_provider};
 pub use server::{ReloadableServerCertResolver, spawn_server_cert_reload_loop};
 pub use source::{TlsFileLayout, TlsSource, TlsSourceKind, TlsSourceSnapshot};
 pub use state::OutboundOnlySnapshotArgs;
@@ -0,0 +1,108 @@
+// 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 rustls::crypto::CryptoProvider;
+
+/// The outcome of ensuring the process-wide rustls provider.
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub enum CryptoProviderStatus {
+    Installed,
+    AlreadyInstalled,
+}
+
+/// An incompatible process-wide rustls provider prevents FIPS startup.
+#[derive(Debug, thiserror::Error)]
+#[error(
+    "FIPS TLS requested, but the process-wide rustls crypto provider is not FIPS compatible; configure the host to install a FIPS provider before starting RustFS"
+)]
+pub struct CryptoProviderError {
+    _private: (),
+}
+
+/// Construct the provider selected by this crate's feature configuration.
+pub fn default_crypto_provider() -> CryptoProvider {
+    #[cfg(feature = "fips")]
+    {
+        rustls::crypto::default_fips_provider()
+    }
+    #[cfg(not(feature = "fips"))]
+    {
+        rustls::crypto::aws_lc_rs::default_provider()
+    }
+}
+
+/// Install or reuse the process-wide rustls provider.
+///
+/// # Errors
+/// Returns an error in a FIPS build if the provider actually installed in the
+/// process is absent or does not satisfy rustls's FIPS checks.
+pub fn ensure_crypto_provider() -> Result<CryptoProviderStatus, CryptoProviderError> {
+    // Rustls's process-wide provider is write-once; check again after a race.
+    let status = if CryptoProvider::get_default().is_some() || default_crypto_provider().install_default().is_err() {
+        CryptoProviderStatus::AlreadyInstalled
+    } else {
+        CryptoProviderStatus::Installed
+    };
+    #[cfg(feature = "fips")]
+    if !CryptoProvider::get_default().is_some_and(|provider| provider.fips()) {
+        return Err(CryptoProviderError { _private: () });
+    }
+    Ok(status)
+}
+
+#[cfg(test)]
+mod tests {
+    use super::{CryptoProviderStatus, default_crypto_provider, ensure_crypto_provider};
+
+    #[test]
+    fn provider_policy_isolated() {
+        const SCENARIO: &str = "RUSTFS_TLS_TEST_PROVIDER_SCENARIO";
+        if let Ok(scenario) = std::env::var(SCENARIO) {
+            match scenario.as_str() {
+                "fresh" => {
+                    assert_eq!(ensure_crypto_provider().expect("initial install"), CryptoProviderStatus::Installed);
+                    assert_eq!(ensure_crypto_provider().expect("provider reuse"), CryptoProviderStatus::AlreadyInstalled);
+                }
+                "incompatible" => {
+                    let mut provider = default_crypto_provider();
+                    provider.kx_groups = vec![rustls::crypto::aws_lc_rs::kx_group::X25519];
+                    provider.install_default().expect("install incompatible provider");
+                    let result = ensure_crypto_provider();
+                    if cfg!(feature = "fips") {
+                        assert!(result.is_err(), "FIPS mode must reject the installed provider");
+                    } else {
+                        assert_eq!(result.expect("reuse existing provider"), CryptoProviderStatus::AlreadyInstalled);
+                    }
+                }
+                _ => panic!("unknown provider scenario"),
+            }
+            return;
+        }
+
+        for scenario in ["fresh", "incompatible"] {
+            let output = std::process::Command::new(std::env::current_exe().expect("test executable"))
+                .arg("provider_policy_isolated")
+                .arg("--nocapture")
+                .env(SCENARIO, scenario)
+                .output()
+                .expect("start isolated provider test");
+            assert!(
+                output.status.success(),
+                "{scenario}: stdout={} stderr={}",
+                String::from_utf8_lossy(&output.stdout),
+                String::from_utf8_lossy(&output.stderr)
+            );
+        }
+    }
+}
@@ -26,6 +26,7 @@ categories = ["web-programming", "development-tools", "cryptography"]

 [dependencies]
 hotpath.workspace = true
+aws-lc-rs = { workspace = true, optional = true }
 base64-simd = { workspace = true, optional = true }
 blake2 = { workspace = true, optional = true }
 brotli = { workspace = true, optional = true }
@@ -89,6 +90,7 @@ path = []  # path manipulation features
 compress = ["dep:flate2", "dep:brotli", "dep:snap", "dep:lz4", "dep:zstd"]
 string = ["dep:regex"]
 crypto = ["dep:base64-simd", "dep:hex-simd", "dep:hmac", "dep:hyper", "dep:sha1", "dep:sha2"]
+fips = ["crypto", "dep:aws-lc-rs", "aws-lc-rs/fips"]
 hash = ["dep:highway", "dep:md-5", "dep:sha2", "dep:blake2", "dep:serde", "dep:siphasher", "dep:hex-simd", "dep:crc-fast"]
 os = ["dep:rustix", "dep:tempfile", "dep:windows"]  # operating system utilities
 integration = []  # integration test features
@@ -13,12 +13,77 @@
 // limitations under the License.

 use hex_simd::{AsOut, AsciiCase};
-use hmac::{Hmac, KeyInit, Mac};
+use hmac::Hmac;
+#[cfg(not(feature = "fips"))]
+pub use hmac::{KeyInit, Mac};
 use hyper::body::Bytes;
+#[cfg(not(feature = "fips"))]
 use sha1::Sha1;
 use sha2::{Digest, Sha256};
 use std::mem::MaybeUninit;

+#[cfg(not(feature = "fips"))]
+pub type HmacSha256 = Hmac<Sha256>;
+
+#[cfg(feature = "fips")]
+pub struct HmacSha256Output([u8; 32]);
+
+#[cfg(feature = "fips")]
+impl HmacSha256Output {
+    #[must_use]
+    pub fn into_bytes(self) -> hmac::digest::Output<Hmac<Sha256>> {
+        self.0.into()
+    }
+}
+
+#[cfg(feature = "fips")]
+pub trait KeyInit: Sized {
+    fn new_from_slice(key: &[u8]) -> Result<Self, hmac::digest::InvalidLength>;
+}
+
+#[cfg(feature = "fips")]
+pub trait Mac: Sized {
+    fn update(&mut self, data: &[u8]);
+    fn finalize(self) -> HmacSha256Output;
+    fn verify_slice(self, tag: &[u8]) -> Result<(), hmac::digest::MacError>;
+}
+
+#[cfg(feature = "fips")]
+#[derive(Clone)]
+pub struct HmacSha256(aws_lc_rs::hmac::Context);
+
+#[cfg(feature = "fips")]
+impl HmacSha256 {
+    pub fn new_from_slice(key: &[u8]) -> Result<Self, hmac::digest::InvalidLength> {
+        let signing_key = aws_lc_rs::hmac::Key::new(aws_lc_rs::hmac::HMAC_SHA256, key);
+        Ok(Self(aws_lc_rs::hmac::Context::with_key(&signing_key)))
+    }
+}
+
+#[cfg(feature = "fips")]
+impl KeyInit for HmacSha256 {
+    fn new_from_slice(key: &[u8]) -> Result<Self, hmac::digest::InvalidLength> {
+        Self::new_from_slice(key)
+    }
+}
+
+#[cfg(feature = "fips")]
+impl Mac for HmacSha256 {
+    fn update(&mut self, data: &[u8]) {
+        self.0.update(data);
+    }
+
+    fn finalize(self) -> HmacSha256Output {
+        let tag = self.0.sign();
+        HmacSha256Output(tag.as_ref().try_into().expect("HMAC-SHA256 tag is 32 bytes"))
+    }
+
+    fn verify_slice(self, tag: &[u8]) -> Result<(), hmac::digest::MacError> {
+        let computed = self.0.sign();
+        aws_lc_rs::constant_time::verify_slices_are_equal(computed.as_ref(), tag).map_err(|_| hmac::digest::MacError)
+    }
+}
+
 /// Base64 URL safe encoding without padding
 /// `base64_encode_url_safe_no_pad(input)`
 ///
@@ -84,9 +149,18 @@ pub fn is_sha256_checksum(s: &str) -> bool {
 /// A 20-byte array containing the HMAC-SHA1 hash of the input data using the provided key
 ///
 pub fn hmac_sha1(key: impl AsRef<[u8]>, data: impl AsRef<[u8]>) -> [u8; 20] {
-    let mut m = <Hmac<Sha1>>::new_from_slice(key.as_ref()).expect("operation should succeed");
-    m.update(data.as_ref());
-    m.finalize().into_bytes().into()
+    #[cfg(feature = "fips")]
+    {
+        let signing_key = aws_lc_rs::hmac::Key::new(aws_lc_rs::hmac::HMAC_SHA1_FOR_LEGACY_USE_ONLY, key.as_ref());
+        let tag = aws_lc_rs::hmac::sign(&signing_key, data.as_ref());
+        tag.as_ref().try_into().expect("HMAC-SHA1 tag is 20 bytes")
+    }
+    #[cfg(not(feature = "fips"))]
+    {
+        let mut m = <Hmac<Sha1>>::new_from_slice(key.as_ref()).expect("operation should succeed");
+        m.update(data.as_ref());
+        m.finalize().into_bytes().into()
+    }
 }

 /// HMAC-SHA256 hashing
@@ -100,7 +174,7 @@ pub fn hmac_sha1(key: impl AsRef<[u8]>, data: impl AsRef<[u8]>) -> [u8; 20] {
 /// A 32-byte array containing the HMAC-SHA256 hash of the input data using the provided key
 ///
 pub fn hmac_sha256(key: impl AsRef<[u8]>, data: impl AsRef<[u8]>) -> [u8; 32] {
-    let mut m = Hmac::<Sha256>::new_from_slice(key.as_ref()).expect("operation should succeed");
+    let mut m = HmacSha256::new_from_slice(key.as_ref()).expect("operation should succeed");
    m.update(data.as_ref());
    m.finalize().into_bytes().into()
 }
@@ -167,3 +241,20 @@ fn sha256_checksum_rejects_wrong_length_or_non_lowercase_hex() {
    assert!(!is_sha256_checksum("0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdeg"));
    assert!(!is_sha256_checksum("0123456789ABCDEF0123456789abcdef0123456789abcdef0123456789abcdef"));
 }
+
+#[test]
+fn test_hmac_sha256_rfc4231() {
+    #[cfg(feature = "fips")]
+    aws_lc_rs::try_fips_mode().expect("aws-lc-rs FIPS mode must be active when `rustfs-utils/fips` is enabled");
+
+    let key = [0x0b_u8; 20];
+    let data = b"Hi There";
+    assert_eq!(hex(hmac_sha1(key, data)), "b617318655057264e28bc0b6fb378c8ef146be00");
+    let mac = hmac_sha256(key, data);
+    assert_eq!(hex(mac), "b0344c61d8db38535ca8afceaf0bf12b881dc200c9833da726e9376c2e32cff7");
+
+    let mut verifier = <HmacSha256 as KeyInit>::new_from_slice(&key).expect("valid key");
+    verifier.update(data);
+    assert!(verifier.clone().verify_slice(&mac).is_ok());
+    assert!(verifier.verify_slice(&[0_u8; 32]).is_err());
+}
@@ -1,6 +1,6 @@
 [package]
 name = "zstd"
-version = "0.13.3"
+version = "0.13.4"
 publish = false
 edition.workspace = true
 license.workspace = true
@@ -33,10 +33,10 @@ zstd-safe = { version = "8.2.0", features = ["seekable"] }
 toml = "1.1.8"
 semver = "1.0.28"
 apache-avro.workspace = true
-arrow-array = "59.3.2"
-arrow-ipc = { version = "59.3.2", features = ["zstd"] }
-arrow-schema = "59.3.2"
-parquet = { version = "59.3.2", default-features = false, features = ["arrow", "zstd"] }
+arrow-array = "60.1.0"
+arrow-ipc = { version = "60.1.0", features = ["zstd"] }
+arrow-schema = "60.1.0"
+parquet = { version = "60.1.0", default-features = false, features = ["arrow", "zstd"] }
 kafka-protocol = { version = "0.18.0", default-features = false, features = ["zstd"] }
 bytes.workspace = true

@@ -6,6 +6,6 @@ This private package satisfies the version constraints of Avro, Arrow IPC, Parqu

 The fixed implementation closes RUSTSEC-2026-0391, RUSTSEC-2026-0398, and RUSTSEC-2026-0403. A successful advisory scan alone is insufficient evidence: cargo-deny skips local path packages. The dependency-graph test therefore requires this adapter's sole runtime dependency to be the exact fixed registry implementation and rejects vulnerable registry codec versions. Security tests exercise abandoned stream contexts, malformed legacy frames, and seekable offsets past EOF. Consumer tests retain zstandard support in all four upstream formats.

-Run `cargo nextest run --locked -p zstd@0.13.3` for the focused tests. The seekable feature is enabled by a test-only dependency to cover the optional vulnerable API without enabling it in production.
+Run `cargo nextest run --locked -p zstd@0.13.4` for the focused tests. The seekable feature is enabled by a test-only dependency to cover the optional vulnerable API without enabling it in production.

 Remove this package and the root Cargo patch when all consumers accept the fixed zstd release line directly. The cleanup marker is registered in the architecture compatibility register.
@@ -57,7 +57,7 @@ fn legacy_constraints_resolve_only_to_the_fixed_registry_implementation() {
        let version = Version::parse(package["version"].as_str().expect("package version")).expect("valid package semver");
        if name == "zstd" && package.get("source").is_none() {
            adapters += 1;
-            assert_eq!(version, Version::new(0, 13, 3));
+            assert_eq!(version, Version::new(0, 13, 4));
            let runtime = package["dependencies"].as_array().expect("adapter dependency edges");
            assert!(runtime.iter().any(|value| value.as_str() == Some("zstd 0.14.3")));
        } else {
@@ -36,10 +36,6 @@ allow-git = [
    # Official s3s repository pinned for Iceberg SigV4 path encoding support.
    # owner: GatewayJ  review: 2026-10
    "https://github.com/s3s-project/s3s",
-    # Upstream OIDC/OAuth2 crypto migrations, pinned to reviewed PR commits.
-    # owner: rustfs-maintainers  review: 2026-10
-    "https://github.com/ramosbugs/openidconnect-rs",
-    "https://github.com/ramosbugs/oauth2-rs",
    # Completion ownership patch fixed to rustfs/uring#20.
    # owner: overtrue  review: 2026-10
    "https://github.com/rustfs/uring.git",
@@ -11,6 +11,8 @@

 ## Open Items

+- `openidconnect-hmac-signature-width` ID token verification guard: openidconnect 4.1.0 converts decoded HMAC signatures to fixed-size arrays, which panic on unexpected lengths. Check HS256, HS384, and HS512 signature widths before delegating all cryptographic, issuer, audience, expiry, and nonce checks to the upstream verifier. Remove the guard after a supported upstream release returns verification errors for arbitrary HMAC signature lengths.
+
 - `zstd-013-fixed-backend` dependency constraint bridge: Avro, Arrow IPC, Parquet, and Kafka still require zstd 0.13. The local adapter re-exports the fixed registry zstd 0.14.3 implementation and its zstd-safe types. Remove the adapter and Cargo patch after every dependency accepts the fixed zstd line without the 0.13 bridge.

 - `8565` historical rebalance replacement recovery: older schema-seven intents can exhaust their retry budget before format repair when retained rebalance metadata describes removed pools. After validating blank targets, terminal historical topology, and ownership, grant one durable handoff and one rebuild attempt. Remove after supported direct upgrades no longer admit snapshots lacking replacement_preformat_handoff_retry_done and all retained affected intents have been resolved.
@@ -64,7 +64,7 @@ RustFS S3 Tables is an Iceberg REST Catalog and table-bucket implementation on t
 | Namespaces | Supported | Create, list, load, exists, drop on both prefixes; `pageSize`/`pageToken` pagination with context-bound tokens; identifiers limited to 512 ASCII characters. |
 | Tables | Supported | Direct and staged create, register, list, load, exists, rename, commit, metadata-location get/update, and drop on both prefixes. Staged create is stateless until an `assert-create` commit atomically publishes the initial metadata and table entry. Rename uses a bucket-scoped persistent fence, recoverable intent, and conditional publication; the source name is reusable only via an ETag-conditional tombstone replacement. Commit identifiers must match the URL; unknown requirements/updates fail as bad requests; register overwrite, purge-on-drop, and v3-only encryption-key updates return an explicit unsupported-operation response. |
 | Commit CAS | Supported | Single-table commits validate base metadata, version token, referenced object existence, warehouse scope, and Iceberg requirements before advancing the pointer; external metadata transitions preserve monotonic assignment watermarks and immutable retained definitions. `idempotency-key-lifetime` is not advertised; mutation-wide `Idempotency-Key` semantics are unsupported. |
-| Commit recovery | Supported | Commit log, idempotency lookup, diagnostics, and recovery routes expose and repair finalization gaps without moving the pointer. |
+| Commit recovery | Supported | Commit log, idempotency lookup, diagnostics, and recovery routes expose and repair finalization gaps without moving the pointer. Object-backed recovery reports a staged request as SUPERSEDED only when a reachable finalized same-base winner proves it lost CAS and its idempotency binding is intact. The original staged receipt is retained; its old request still conflicts. The winning history must be retained with the losing receipt; missing or contradictory history requires manual review. |
 | Snapshot refs | Supported | List, create/replace, delete via commits; `main` is protected; refs with explicit retention need forced delete. |
 | Iceberg views | Supported | Create, list, load, replace, exists, drop with view-scoped authorization; only view format version 1. |
 | LoadTable and table credentials endpoint | Supported | Vending only on negotiated `vended-credentials`; one temporary session scoped to the warehouse prefix and current metadata location; missing credential permission falls back to metadata-only with an explicit reason; responses carry `Cache-Control: no-store, private`. |
@@ -1,29 +1,30 @@
 # Cryptographic compliance positioning

 **Use this when:** writing README, CHANGELOG, release notes, marketing, RFP, or security-questionnaire text that touches FIPS, or reasoning about the `rustfs` / `rustfs-tls-runtime` / `rustfs-crypto` `fips` features and algorithm deprecation.
-**Source of truth:** `scripts/check_fips_wording.sh` (the enforced guard); `rustfs/Cargo.toml` and `crates/tls-runtime/Cargo.toml` (`fips` feature); `crates/crypto/Cargo.toml` (`fips` feature); `crates/crypto/src/encdec/id.rs` (`ID` algorithm bytes); `rustfs/src/startup_runtime_hooks.rs` (`install_default_crypto_provider`).
+**Source of truth:** `scripts/check_fips_wording.sh` (the enforced guard); the `fips` features in `rustfs/Cargo.toml` and the affected library manifests; `crates/tls-runtime/src/provider.rs` (provider policy); `crates/crypto/Cargo.toml` (`fips` feature); `crates/crypto/src/encdec/id.rs` (`ID` algorithm bytes).

 This document records where RustFS stands on cryptographic module validation and what may and may not be said about it, so the question is answered once, from the code. For where master key material lives per backend, see [KMS backend security properties](kms-backend-security.md).

 ## Status: not FIPS 140-3 validated

-**RustFS is not FIPS 140-3 (or 140-2) validated as a product, and in the default build no component it links runs as a validated cryptographic module.** There is no CMVP certificate covering RustFS itself, or the libraries it uses in the default shipped configuration. An opt-in `--features fips` build flag on `rustfs` switches the `rustls` / `aws-lc-rs` provider to `aws-lc-fips-sys` for TLS, X.509, and JWT operations, while object data-path and local KMS operations remain on unvalidated RustCrypto software implementations. Validation is a property of a specific module build, its documented boundary, and a certificate — none of which RustFS as a whole has or pursues.
+**RustFS is not FIPS 140-3 (or 140-2) validated as a product, and in the default build no component it links runs as a validated cryptographic module.** There is no CMVP certificate covering RustFS itself, or the libraries it uses in the default shipped configuration. An opt-in `--features fips` build flag on `rustfs` enables the `aws-lc-fips-sys` backend in the applicable dependency graph; it does not establish that every TLS, X.509, or JWT operation uses an approved service. Object data-path and local KMS operations remain on unvalidated RustCrypto software implementations. Validation is a property of a specific module build, its documented boundary, and a certificate — none of which RustFS as a whole has or pursues.

 ### What the process actually links

 | Layer | Where | Implementation | Primitives | Validated module |
 | --- | --- | --- | --- | --- |
-| TLS (S3 server, internode, outbound clients) | Process-wide default provider installed by `install_default_crypto_provider` in `rustfs/src/startup_runtime_hooks.rs` | `rustls` with the `aws-lc-rs` provider (`default_provider` by default; `default_fips_provider` when built with `--features fips`) | TLS 1.2/1.3 suites, `prefer-post-quantum` hybrid key exchange (restricted to FIPS-approved suites and groups when `--features fips` is enabled) | No in the default build (`aws-lc-sys`); opt-in `--features fips` links `aws-lc-fips-sys` and asserts `provider.fips()` at startup |
+| TLS using the process-wide rustls provider | Provider installed or checked by `install_default_crypto_provider` in `rustfs/src/startup_runtime_hooks.rs` | `rustls` with the process-wide provider (AWS-LC is installed when none exists) | TLS 1.2/1.3 suites, `prefer-post-quantum` hybrid key exchange (restricted to FIPS-approved suites and groups when `--features fips` is enabled) | No in the default build (`aws-lc-sys`); opt-in `--features fips` enables `aws-lc-fips-sys` and checks the installed provider at startup; module validation depends on the deployment |
+| Internode RPC HMAC authentication | `crates/ecstore/src/cluster/rpc/http_auth.rs` via `rustfs_utils::crypto::HmacSha256` | RustCrypto `hmac` + `sha2` in the default build; `aws-lc-rs` when `fips` is enabled | HMAC-SHA256 (`x-rustfs-signature`, `x-rustfs-rpc-signature-v2`, `x-rustfs-rpc-signature-v3`) | No in the default build; opt-in `--features fips` enables `rustfs-utils/fips` (`aws-lc-fips-sys`) |
 | Object data path AEAD (SSE) | `crates/kms/src/encryption/ciphers.rs`, `crates/rio/src/encrypt_reader.rs`, `crates/rio-v2/src/encrypt_reader.rs` | RustCrypto `aes-gcm`, `chacha20poly1305` | AES-256-GCM, ChaCha20-Poly1305 | No |
 | DEK wrapping | `crates/kms/src/encryption/dek.rs` | RustCrypto `aes-gcm` | AES-256-GCM | No |
 | Local KMS backend master key | `crates/kms/src/backends/local.rs` | RustCrypto `argon2`, `aes-gcm` | Argon2id KDF, AES-256-GCM | No |
 | Config and IAM blobs at rest | `crates/crypto/src/encdec/` (`rustfs-crypto`) | RustCrypto `pbkdf2`/`argon2`, `aes-gcm`, `chacha20poly1305`, `sha2` | see [the `rustfs-crypto` `fips` feature](#the-rustfs-crypto-fips-feature-what-it-actually-does) | No |
-| JWT signing and verification | `jsonwebtoken` with the `aws_lc_rs` feature (`crates/crypto`, `crates/iam`, `crates/policy`) | AWS-LC through `aws-lc-rs` | Non-FIPS `aws-lc-sys` build by default; unified with `aws-lc-fips-sys` when `rustfs` is built with `--features fips` | No in the default build; links `aws-lc-fips-sys` only when built with `-p rustfs --features fips` |
+| JWT signing and verification | `jsonwebtoken` with the `aws_lc_rs` feature (`crates/crypto`, `crates/iam`, `crates/policy`) | AWS-LC through `aws-lc-rs` | Non-FIPS `aws-lc-sys` in the standalone default build; the FIPS backend can be enabled by Cargo feature unification | No in the standalone default build; linking the FIPS backend does not establish validation of each operation |

 Two consequences are commonly assumed the other way:

- **AWS-LC being present does not imply FIPS in the default build.** `aws-lc-rs` has a FIPS variant (`aws-lc-fips-sys`); default builds do not enable it. Building `rustfs` with `--features fips` enables `rustls/fips`, `tokio-rustls/fips`, and `rustfs-tls-runtime/fips`, which links `aws-lc-fips-sys` and installs `rustls::crypto::default_fips_provider()` (see [below](#the-rustfs-and-rustfs-tls-runtime-fips-feature)).
- **The object data path never touches AWS-LC.** Every byte of object plaintext is encrypted by RustCrypto software implementations, even when `rustfs` is built with `--features fips`. Enabling the TLS FIPS provider does not change the object encryption data path; see route 1 below.
+- **AWS-LC being present does not imply FIPS in the default build.** `aws-lc-rs` has a FIPS variant (`aws-lc-fips-sys`); default builds do not enable it. Building `rustfs` with `--features fips` enables `rustls/fips`, `tokio-rustls/fips`, and `rustfs-tls-runtime/fips`, which enables `aws-lc-fips-sys` and installs or checks the process-wide rustls provider (see [below](#the-rustfs-and-rustfs-tls-runtime-fips-feature)).
+- **The object encryption data path uses RustCrypto.** When RustFS encrypts object data, it uses RustCrypto software implementations, even when `rustfs` is built with `--features fips`. Enabling the TLS FIPS provider does not change the object encryption data path; see route 1 below.

 ## Terminology red lines for external material

@@ -44,7 +45,7 @@ These rules apply to the README, CHANGELOG, release notes, marketing pages, sale

 Boilerplate when the topic cannot be avoided:

-> RustFS encrypts object data with AES-256-GCM and supports ChaCha20-Poly1305. These are FIPS-approved algorithms, but the implementations are not FIPS 140-3 validated cryptographic modules and RustFS makes no FIPS validation claim.
+> RustFS supports AES-256-GCM and ChaCha20-Poly1305 for object encryption. AES-256-GCM can be a FIPS-approved algorithm when used under the applicable conditions; ChaCha20-Poly1305 is not FIPS-approved. These implementations are not FIPS 140-3 validated cryptographic modules and RustFS makes no FIPS validation claim.

 ### Guard

@@ -52,13 +53,25 @@ Boilerplate when the topic cannot be avoided:

 ## The `rustfs` and `rustfs-tls-runtime` `fips` feature

-`rustfs/Cargo.toml` (`fips = ["rustls/fips", "tokio-rustls/fips", "rustfs-tls-runtime/fips"]`) and `crates/tls-runtime/Cargo.toml` (`fips = ["rustls/fips"]`) define an opt-in Cargo feature (disabled by default) for builds that require the `aws-lc-fips-sys` cryptographic module on the `aws-lc-rs` / `rustls` path:
+`rustfs/Cargo.toml` and the TLS runtime, protocols, targets, S3 client, and ECStore library manifests define opt-in `fips` features (disabled by default) for the `aws-lc-fips-sys` / `rustls` path. The RustFS feature propagates the policy to these libraries; when enabling a protocol library independently, also enable the required protocol feature.

- **What it enables:** Activates `rustls/fips` and `tokio-rustls/fips`, which pull in `aws-lc-rs/fips` (`aws-lc-fips-sys`) and `rustls-webpki/aws-lc-rs-fips`. Because Cargo unifies features across the dependency graph, other `aws-lc-rs` consumers in the binary (`jsonwebtoken`, `rcgen`, `google-cloud-auth`, `async-nats`) also link against `aws-lc-fips-sys`.
- **Runtime provider installation:** In `rustfs/src/startup_runtime_hooks.rs`, `build_default_crypto_provider()` selects `rustls::crypto::default_fips_provider()` (instead of `rustls::crypto::aws_lc_rs::default_provider()`) and asserts `provider.fips()` on startup, restricting TLS negotiation to FIPS-approved cipher suites (e.g., `TLS_AES_256_GCM_SHA384`, `TLS_AES_128_GCM_SHA256`, `TLS_ECDHE_*_AES_*_GCM_*`) and key exchange groups (excluding `TLS_CHACHA20_POLY1305_SHA256` and `X25519`).
- **Build requirements:** Compiling `aws-lc-fips-sys` requires CMake, Go, and Clang/`libclang` in the build environment (unlike the default `aws-lc-sys` build).
+Cargo's `--all-features` also enables `fips`, including for these individual libraries. Such builds inherit the FIPS backend's target and native-tool requirements; select supported feature combinations deliberately rather than assuming every all-features build is supported.
+
+```bash
+cargo build --release --locked -p rustfs --bin rustfs --features fips
+```
+
+- **What it enables:** Activates `rustls/fips` and `tokio-rustls/fips`, which pull in `aws-lc-rs/fips` (`aws-lc-fips-sys`) and `rustls-webpki/aws-lc-rs-fips`. Cargo unifies features for the same dependency version in a build graph; consumers of that `aws-lc-rs` instance use its FIPS backend, but this does not establish approval for every operation or cover independent dependency versions.
+- **Runtime provider installation:** `rustfs-tls-runtime` installs or reuses the process-wide provider and checks the actual provider in FIPS builds. RustFS performs this check before observability initialization in standalone and embedded startup. FTPS, WebDAV, transition clients, and relevant notification TLS clients also invoke the shared policy before using the process-wide provider. An incompatible host provider causes a configuration error. The check does not inspect clients that use their own provider or a different TLS implementation.
+- **Build requirements:** The FIPS backend requires a C/C++ toolchain, CMake, and Go; targets without suitable pre-generated bindings also require bindgen and Clang/`libclang`. Check the locked dependency version and target platform for their exact requirements.
 - **What it does not change:** It does not alter the object data-path AEAD (`crates/kms`, `crates/rio`, `crates/rio-v2`) or `rustfs-crypto` config/IAM blob encryption, which continue to use pure-Rust `RustCrypto` crates.

+The RustFS server TLS configuration uses the process-wide rustls provider. The Kafka notification client and remote replication SDK explicitly select AWS-LC, so their providers are independent of the process default; for the locked rustls version, the unified `rustls/fips` feature restricts that provider's suites and groups. ECStore's insecure replication-client path still uses the process default: an application using ECStore independently must directly depend on `rustfs-tls-runtime` with its `fips` feature and call `rustfs_tls_runtime::ensure_crypto_provider()` before constructing that client when it requires the FIPS policy. Skipping TLS certificate verification does not authenticate the remote target, regardless of provider selection. Feature unification applies only to the same dependency version in a build graph; independent builds and third-party TLS stacks require separate verification. No complete outbound TLS or platform validation matrix is asserted here.
+
+NATS can upgrade a connection to TLS based on the server's response, so the targets library checks the process provider before connecting in a FIPS build, including for a plaintext-looking NATS URL. Webhook targets construct a TLS-capable HTTP client even for an HTTP endpoint and also check the provider when built with `fips`. An incompatible host provider therefore prevents these initializations even if a particular connection would not use TLS.
+
+A provider check or linked FIPS backend does not establish CMVP validation for a deployment or approval of every operation. Verify the specific module version, certificate status, security policy, build, operating environment, and approved services against the [CMVP validated modules list](https://csrc.nist.gov/projects/cryptographic-module-validation-program/validated-modules). The feature is opt-in; before enabling it, test client algorithms, certificates, TLS versions, rolling upgrades, and rollback in the target environment.
+
 ## The `rustfs-crypto` `fips` feature: what it actually does

 `crates/crypto/Cargo.toml` declares `default = ["crypto", "fips"]`, so the feature is on in every normal build. Its entire effect is **which algorithm the write path selects**; the implementation is RustCrypto either way.
@@ -80,8 +93,8 @@ Selection sites are `crates/crypto/src/encdec/encrypt.rs` and `crates/crypto/src

 | Route | Scope | What it buys | Verdict |
 | --- | --- | --- | --- |
-| 1. Adopt the `aws-lc-rs` FIPS variant (`aws-lc-fips-sys`) across the data path | The TLS/X.509/JWT provider swap is wired via the opt-in `-p rustfs --features fips` build flag. Extending `aws-lc-fips-sys` to the data path would require re-implementing every AEAD/KDF call in `crates/kms/src/encryption/`, `crates/rio*/src/encrypt_reader.rs`, `crates/kms/src/backends/local.rs`, and `crates/crypto/src/encdec/` against `aws-lc-rs` primitives; Argon2id (no FIPS status) and ChaCha20-Poly1305 would need read-only compatibility stories. Build needs CMake, a C toolchain, Clang/`libclang`, and Go, on a narrower target set — the static musl release already hits AWS-LC's `getentropy` abort on kernels older than 3.17, and the FIPS variant constrains the matrix strictly harder | The accurate claim becomes "cryptographic operations are performed by a FIPS 140-3 validated module", not "RustFS is FIPS validated"; a product-level claim additionally needs a documented boundary, approved-mode enforcement, self-tests, zeroization, and entropy documentation | Data-path rewrite remains heavy and re-opens platform-support constraints; TLS/X.509/JWT is opt-in via `--features fips` |
-| 2. Let an externally validated KMS carry key operations | Mostly built: the `VaultTransit` backend never lets key-encryption key material leave Vault, so master key generation, wrapping, unwrapping, and rotation happen inside whatever module Vault's seal/HSM is validated against. Remaining work is configuration guidance and a supported-deployment statement | A defensible partial answer to "where do keys live and who validated that". The object data path stays RustCrypto and TLS stays non-FIPS AWS-LC: "key management operations are performed by an externally validated module; the object data path is not validated" | Nearest partial step, no code rewrite. Point customers here when the requirement is key custody rather than a certificate covering the storage layer |
+| 1. Adopt the `aws-lc-rs` FIPS variant (`aws-lc-fips-sys`) across the data path | The opt-in `-p rustfs --features fips` flag enables the FIPS backend for the relevant dependency graph; individual TLS, X.509, and JWT services still require verification. Extending `aws-lc-fips-sys` to the data path would require re-implementing every AEAD/KDF call in `crates/kms/src/encryption/`, `crates/rio*/src/encrypt_reader.rs`, `crates/kms/src/backends/local.rs`, and `crates/crypto/src/encdec/` against `aws-lc-rs` primitives; Argon2id (no FIPS status) and ChaCha20-Poly1305 would need read-only compatibility stories. Build needs a C toolchain, CMake, Go, and target-dependent bindings tooling, on a narrower target set — the static musl release already hits AWS-LC's `getentropy` abort on kernels older than 3.17, and the FIPS variant constrains the matrix strictly harder | A claim that particular operations use a validated module requires verified module and deployment boundaries, certificate status, approved-mode operation, and service coverage; it does not establish product-level validation | Data-path rewrite remains heavy and re-opens platform-support constraints; the FIPS backend is opt-in via `--features fips` |
+| 2. Let an externally validated KMS carry key operations | Mostly built: the `VaultTransit` backend never lets key-encryption key material leave Vault, so master key generation, wrapping, unwrapping, and rotation happen inside whatever module Vault's seal/HSM is validated against. Remaining work is configuration guidance and a supported-deployment statement | A defensible partial answer to "where do keys live and who validated that". The object data path stays RustCrypto; TLS backend selection remains a separate decision: "key management operations are performed by an externally validated module; the object data path is not validated" | Nearest partial step, no code rewrite. Point customers here when the requirement is key custody rather than a certificate covering the storage layer |
 | 3. Make no validation claim (current default) | This document plus the grep guard | Nothing further | The current decision. FIPS 140-3 validation is not a roadmap target; adjacent items (PKCS#11, KMIP, BYOK, signing keys) are deferred for lack of demand and because HSM-dependent paths cannot be exercised in CI |

 ## Algorithm disablement and migration policy
@@ -14,7 +14,7 @@ Pick the lowest layer that can prove the change; add a higher-layer test only wh
 | Outbound target matrix | Replication of every object shape (empty, plain, retention, legal hold, multipart, checksummed, SSE-S3) against every remote-target failure mode the fake target models; an explicit expectation table pins known-red cells to an open issue | `python3 scripts/e2e_binary.py build`, then `python3 scripts/e2e_binary.py run -- cargo nextest run -p e2e_test -E 'test(/^replication_target_matrix_test::/)'` | With `e2e-repl-nightly`; required locally for any change to outbound client defaults (SOP: [`docs/postmortems/2026-09-03-replication-checksum-default-regression.md`](../postmortems/2026-09-03-replication-checksum-default-regression.md)) |
 | s3s-e2e conformance | External S3 conformance tool against a live server | `./scripts/e2e-run.sh ./target/debug/rustfs <data-dir>` | Non-documentation PRs, required through `Test and Lint` (second half of the `End-to-End Tests` job) |
 | S3 compatibility | `ceph/s3-tests` (boto3; allow-list `scripts/s3-tests/implemented_tests.txt`) and MinIO `mint` | `scripts/s3-tests/run.sh`; mint via `.github/workflows/mint.yml` | Non-documentation PRs: implemented and lifecycle S3 lanes required through `Test and Lint`; the separate s3-tests workflow provides additional PR reporting and a weekly full sweep; mint: weekly, report-only |
-| Chaos / fault-injection | Single-node disk fault injection (`crates/e2e_test/src/chaos.rs`, `crates/e2e_test/src/fault_proxy.rs`) plus the 4-node kill/fresh-drive/blackhole cases in `crates/e2e_test/src/distributed/chaos_test.rs` | Part of the e2e crate (`e2e-reliability` and `e2e-distributed`) | Reliability cases with `e2e-full`; 4-node chaos on storage-sensitive PRs and nightly via `e2e-distributed` |
+| Chaos / fault-injection | Single-node disk fault injection (`crates/e2e_test/src/chaos.rs`, `crates/e2e_test/src/fault_proxy.rs`) plus the 4-node kill/fresh-drive/blackhole cases in `crates/e2e_test/src/distributed/chaos_test.rs` | Part of the e2e crate (`e2e-reliability` and `e2e-distributed`) | Reliability cases with `e2e-full`; 4-node chaos nightly via `e2e-distributed` |
 | Fuzz | `cargo-fuzz` targets over untrusted parsing surfaces; isolated sub-workspace under `fuzz/` | `./scripts/fuzz/run.sh` (see [`fuzz/README.md`](../../fuzz/README.md)) | PR smoke on the paths listed in `.github/workflows/fuzz.yml`, plus nightly corpus |
 | Benchmarks | Criterion benches under each crate's `benches/` | `cargo bench -p <crate>` | On demand; never a gate |
 | Conditional reads | Paired ordinary/hit/miss GET latency, throughput and read-lock measurements | `scripts/run_conditional_read_abba.py` | Local conditional-read acceptance; [guide and scope](conditional-read-benchmark.md) |
@@ -88,7 +88,7 @@ All profiles are defined in `.config/nextest.toml`; its block comments hold the
 | `e2e-full` | Merge-queue / main-push single-node e2e lane |
 | `e2e-repl-nightly` | Nightly slow / cross-process replication lane |
 | `e2e-nightly` | Nightly serial multi-process cluster fault lane |
-| `e2e-distributed` | Storage-sensitive PR and nightly 4-node 4-disk S3 / lock / versioning / replication / quota / expand / decommission / rebalance / site-replication / chaos / upgrade (history + IAM AK/SK) lane |
+| `e2e-distributed` | Nightly 4-node 4-disk S3 / lock / versioning / replication / quota / expand / decommission / rebalance / site-replication / chaos / upgrade (history + IAM AK/SK) lane |
 | `e2e-protocols` | Nightly fixed-port FTPS/SFTP/WebDAV lane, run with `-j 1` |

 Membership of each e2e profile is pinned by a digest in `.config/e2e-<profile>-selection.txt` and checked by `scripts/check_test_wiring.py --check-profile <profile>` before the lane runs. To list what a profile selects on your platform (the result is platform-dependent because some modules are linux-only):
@@ -35,7 +35,7 @@ gh api repos/rustfs/rustfs/rulesets/6436880 \
  --jq '.rules[] | select(.type == "required_status_checks") | .parameters'
 ```

-The aggregate requires the validation lanes already selected by `ci.yml`; this closes the gap where a failing critical lane left the required workspace check green. Independent workflows remain report-only unless separately required. Before adding a new expensive lane or moving existing PR coverage to a schedule, collect representative execution and regression evidence, establish ownership and a working scheduled replacement, and update this reference with the resulting policy.
+The aggregate requires the validation lanes already selected by `ci.yml`; this closes the gap where a failing critical lane left the required workspace check green. Independent workflows remain report-only unless separately required. Before adding a new expensive lane or moving existing PR coverage to a schedule, collect representative execution and regression evidence, establish ownership and a working scheduled replacement, and update this reference with the resulting policy. `e2e-distributed.yml` and `e2e-upgrade.yml` are that kind of replacement: the 4-node lane runs on its nightly schedule, and the single-node upgrade lane runs on its weekly Monday schedule, on version tags, and on manual dispatch. Neither is part of `Test and Lint`. Dispatch the workflow against a candidate branch when that signal is needed before merge.

 Superseded PR attempts are cancelled. A running main validation finishes, with only the newest pending main run retained, so frequent merges cannot continuously cancel the full baseline. Having a `merge_group` trigger does not itself require use of the merge queue.

@@ -66,7 +66,6 @@ The four site-replication state-writer concurrency proofs reserve the available
 | PR touching `paths` in `fuzz.yml` | `Build Fuzz Harness`, `Smoke / <target>` | `fuzz.yml` `fuzz-build`, `pr-fuzz-smoke` | Report-only | `MAX_TOTAL_TIME=60 ./scripts/fuzz/run.sh` |
 | PR touching `paths` in `windows-filesystem.yml` | `Rename Safety`, `Startup Smoke` | `windows-filesystem.yml` `rename-safety`, `startup-smoke` | Report-only | the `cargo test -p rustfs-ecstore --lib <filter>` commands in the job, on Windows; `cargo build --locked -p rustfs --bin rustfs`, then `rustfs.exe --version` and a single-volume server polled on `/health/ready` |
 | PR touching `paths` in `coverage.yml` | `Workspace line coverage` | `coverage.yml` `coverage` | Report-only | `make coverage`; `python3 scripts/check_security_coverage.py target/llvm-cov/coverage.json` |
-| PR touching `paths` in `e2e-upgrade.yml` | `Upgrade compatibility` | `e2e-upgrade.yml` `upgrade` | Report-only | `python3 scripts/e2e_binary.py build`, then the job's `python3 scripts/e2e_binary.py run -- cargo test --locked -p e2e_test` commands with `RUSTFS_UPGRADE_SOURCE_BINARY` pointing at the pinned previous release (`UPGRADE_SOURCE_VERSION`) |
 | PR touching `paths` in `oidc-keycloak.yml` | `OIDC Keycloak live gate` | `oidc-keycloak.yml` `oidc-keycloak-live` | Report-only | `cargo build --locked -p rustfs --bin rustfs`, then `bash scripts/test/oidc_keycloak_live.sh ./target/debug/rustfs` |
 | PR touching `paths` in `targets-integration.yml` | `PostgreSQL, MySQL, AMQP, and NATS` | `targets-integration.yml` `targets-live` | Report-only | start the containers as in the job, export the `RUSTFS_TEST_*` DSNs, then the job's `cargo test --locked -p rustfs-targets --test <name> -- --ignored --test-threads=1` commands |
 | PR, documentation-only selection | `Quick Checks`, `Typos`, `Test and Lint` | `ci.yml` `quick-checks`, `typos`, `required-checks` | Required directly or via aggregate | Quick Checks commands; `python3 scripts/ci_gate.py --self-test` |
@@ -90,7 +89,7 @@ Scheduled lanes never block a PR. Their workflow-local gate fails the run. The v
 | `cache-warm.yml` (every six hours, plus main pushes that change `Cargo.lock`, `rust-toolchain.toml` or the setup action) | `warm-ci-dev`, `warm-ci-feat-rio`, `warm-ci-feat-proto`, `warm-ci-uring` | no verdict; sole writer of the `rust-cache` keys that `ci.yml` restores | yes | dispatch `cache-warm.yml` |
 | `build.yml` (nightly) | `build-rustfs` over the six-target platform matrix in `prepare-platform-matrix` (four Linux, macOS aarch64, Windows x86_64) | build/package integrity | yes | dispatch `build.yml` with an exact platform set |
 | `e2e-replication-nightly.yml` (nightly) | `repl-nightly`, `cluster-nightly`, `protocols-nightly` | three independent gates; JUnit, membership listing, server logs | yes | `python3 scripts/e2e_binary.py build --bins`, then `python3 scripts/e2e_binary.py run -- cargo nextest run --profile e2e-repl-nightly -p e2e_test`; `--features e2e-test-hooks` on both steps with `--profile e2e-nightly`; `--features ftps,webdav,sftp` on both steps with `-j 1 --profile e2e-protocols` |
-| `e2e-distributed.yml` (storage-sensitive PRs + nightly) | `distributed` | fail-closed 4-node 4-disk S3, durability, replication, movement, fault, and direct/rolling upgrade gate; JUnit, membership listing, per-node server logs | yes, with `never_ran_grace_until` | download the pinned previous release as in the workflow, export `RUSTFS_UPGRADE_SOURCE_BINARY`, run `python3 scripts/e2e_binary.py build --bins --features e2e-test-hooks`, then `python3 scripts/e2e_binary.py run --features e2e-test-hooks -- cargo nextest run --profile e2e-distributed -p e2e_test` |
+| `e2e-distributed.yml` (nightly) | `distributed` | fail-closed 4-node 4-disk S3, durability, replication, movement, fault, and direct/rolling upgrade gate; JUnit, membership listing, per-node server logs | yes | download the pinned previous release as in the workflow, export `RUSTFS_UPGRADE_SOURCE_BINARY`, run `python3 scripts/e2e_binary.py build --bins --features e2e-test-hooks`, then `python3 scripts/e2e_binary.py run --features e2e-test-hooks -- cargo nextest run --profile e2e-distributed -p e2e_test` |
 | `e2e-s3tests.yml` (weekly) | `s3tests` (single and distributed, four shards each), `upstream-head-canary` | compatibility gate; report, JUnit, node IDs, server logs | yes | `scripts/s3-tests/run.sh` against an existing single or distributed target |
 | `fuzz.yml` (nightly) | `nightly-fuzz-corpus` per target | gate; corpus and crash artifacts | yes (disabled) | `MAX_TOTAL_TIME=<seconds> ./scripts/fuzz/run.sh` |
 | `minio-interop.yml` (nightly) | `minio-interop` | EC + SSE read-parity gate | yes, with `never_ran_grace_until` | pinned Docker fixture steps in the workflow |
@@ -100,7 +99,7 @@ Scheduled lanes never block a PR. Their workflow-local gate fails the run. The v
 | `mint.yml` (weekly) | `mint` | report-only by design; per-suite PASS/FAIL/NA and raw `log.json` | yes | pinned Docker sequence in the workflow |
 | `coverage.yml` (weekly) | `coverage` | report-only trend; lcov and JSON artifact | yes (disabled) | `make coverage` |
 | `runner-hygiene.yml` (monthly) | `check-ephemerality` | runner ephemerality | yes | dispatch |
-| `e2e-upgrade.yml` (weekly) | `upgrade` (serial cases sharing one harness build) | upgrade and rollback gate; server logs | yes | see the PR row |
+| `e2e-upgrade.yml` (weekly, and on version tags) | `upgrade` (serial cases sharing one harness build) | upgrade and rollback gate; server logs | yes | `python3 scripts/e2e_binary.py build`, then the job's `python3 scripts/e2e_binary.py run -- cargo test --locked -p e2e_test` commands with `RUSTFS_UPGRADE_SOURCE_BINARY` pointing at the pinned previous release (`UPGRADE_SOURCE_VERSION`) |
 | `oidc-keycloak.yml` (weekly) | `oidc-keycloak-live` | live OIDC gate | yes | see the PR row |
 | `targets-integration.yml` (nightly) | `targets-live` | live target gate; container logs | yes (disabled) | see the PR row |
 | `gateway-integration-sync.yml` (weekdays) | `sync` | merges `main` into `gateway/integration` as a `--no-ff` merge commit; a conflicting merge is aborted and opens a `sync/<source>-<date>` pull request listing the conflicting files, which keeps the run green; a rejected push or missing pull request fails it | yes | dispatch, optionally with `source_ref` |
@@ -273,9 +273,9 @@ dependencies = [

 [[package]]
 name = "aws-sdk-s3"
-version = "1.156.0"
+version = "1.157.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1072c9aa20159b9b417115342d38c4c53108fc98e5359fcb86b7d4575747eb59"
+checksum = "6ea9794c741520d8baec9462bcd0ad12164d4f5f78d008a8da9d6a0b7222caa8"
 dependencies = [
 "arc-swap",
 "aws-credential-types",
@@ -1837,18 +1837,18 @@ dependencies = [

 [[package]]
 name = "hotpath"
-version = "0.30.2"
+version = "0.30.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b39e941010d86063ff233bf8325269bf68097fb24ba67000fbddc0a410e8056e"
+checksum = "d9632c1fd213e63c1b88d8f3e59543da4999b17f859cc1e1d67cf8f2207e8935"
 dependencies = [
 "hotpath-macros",
 ]

 [[package]]
 name = "hotpath-macros"
-version = "0.30.2"
+version = "0.30.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a24f5179ff680132003306354efd0ac8cde12e87b1a788719c836e91af35cb58"
+checksum = "5d92fdfec803932e5764ea93c201cf7c6c2d8cff1871da44f64f5094c9c3ef8a"

 [[package]]
 name = "http"
@@ -2418,9 +2418,9 @@ dependencies = [

 [[package]]
 name = "log"
-version = "0.4.34"
+version = "0.4.35"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f9f8bd3e56ce4dfc153cf470fffbfa98c7620958b312ca5c3a4b8d5181fd13c6"
+checksum = "5d15f86d12da44cd8d4ec3c73984537eb7f4cadcb9776a575a0a8a05ca11005c"

 [[package]]
 name = "lru"
@@ -4134,8 +4134,7 @@ dependencies = [
 [[package]]
 name = "rustfs-uring"
 version = "0.3.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f7deda770b7981ee3bfaa202ecd18625fc794c962b8d58cd4f436952b7e66dcf"
+source = "git+https://github.com/rustfs/uring.git?rev=eb8fc160e71d804427de1e905f5c226d580f71bf#eb8fc160e71d804427de1e905f5c226d580f71bf"
 dependencies = [
 "io-uring",
 "libc",
@@ -58,7 +58,7 @@ required-features = ["swift"]

 [features]
 default = ["ftps", "webdav", "gcs"]
-fips = ["rustls/fips", "tokio-rustls/fips", "rustfs-tls-runtime/fips"]
+fips = ["rustls/fips", "tokio-rustls/fips", "rustfs-tls-runtime/fips", "rustfs-ecstore/fips", "rustfs-protocols/fips", "rustfs-s3-client/fips", "rustfs-targets/fips", "rustfs-utils/fips"]
 gcs = ["rustfs-ecstore/gcs", "dep:google-cloud-auth"]
 metrics-gpu = ["rustfs-obs/gpu"]
 ftps = ["rustfs-protocols/ftps"]
@@ -63,6 +63,9 @@ pub(crate) async fn init_startup_server_preflight(
    config: &Config,
    env_compat_report: &ExternalEnvCompatReport,
 ) -> std::result::Result<(), StartupServerPreflightError> {
+    #[cfg(feature = "fips")]
+    crate::startup_runtime_hooks::install_default_crypto_provider().map_err(StartupServerPreflightError::Other)?;
+
    crate::config::init_config_snapshot(config);
    init_license(config.license.clone());
    init_startup_observability(config.obs_endpoint.clone()).await?;
@@ -25,7 +25,7 @@ pub(crate) async fn init_startup_runtime_foundation(config: &Config) -> Result<(
    log_startup_runtime_diagnostics();
    init_profiling_runtime().await;
    rustfs_trusted_proxies::init();
-    install_default_crypto_provider();
+    install_default_crypto_provider()?;
    init_object_max_versions_config()?;
    init_outbound_tls_material(config).await
 }
@@ -14,10 +14,7 @@

 use crate::license::license_status;
 use crate::startup_runtime_sources;
-#[cfg(not(feature = "fips"))]
-use rustls::crypto::aws_lc_rs::default_provider;
-#[cfg(feature = "fips")]
-use rustls::crypto::default_fips_provider as default_provider;
+use rustfs_tls_runtime::{CryptoProviderStatus, ensure_crypto_provider};
 use std::future::Future;
 use std::io::Result;
 use tracing::{debug, info, warn};
@@ -108,31 +105,24 @@ where
    shutdown();
 }

-fn build_default_crypto_provider() -> rustls::crypto::CryptoProvider {
-    let provider = default_provider();
-    #[cfg(feature = "fips")]
-    assert!(
-        provider.fips(),
-        "FIPS crypto provider requested via `fips` feature, but AWS-LC FIPS mode is not active"
-    );
-    provider
-}
-
-pub(crate) fn install_default_crypto_provider() {
-    if build_default_crypto_provider().install_default().is_err() {
+pub(crate) fn install_default_crypto_provider() -> Result<()> {
+    if ensure_crypto_provider().map_err(std::io::Error::other)? == CryptoProviderStatus::AlreadyInstalled {
        debug!(
            target: "rustfs::main",
            event = EVENT_CRYPTO_PROVIDER_STATE,
            component = LOG_COMPONENT_MAIN,
            subsystem = LOG_SUBSYSTEM_STARTUP,
-            provider = "aws_lc_rs",
+            provider = "process_default",
            state = "already_installed",
            "Rustls crypto provider state checked"
        );
    }
+    Ok(())
 }

 pub(crate) async fn init_embedded_runtime_hooks(obs_endpoint: String) -> Result<()> {
+    install_embedded_default_crypto_provider()?;
+
    // Observability is a process-global tracing subscriber and can only be
    // installed once. A second embedded server in the same process reuses the
    // first server's subscriber — try to publish anyway (first server wins),
@@ -161,37 +151,228 @@ pub(crate) async fn init_embedded_runtime_hooks(obs_endpoint: String) -> Result<
        }
    }

-    install_embedded_default_crypto_provider();
    rustfs_trusted_proxies::init();

    Ok(())
 }

-fn install_embedded_default_crypto_provider() {
-    if let Err(err) = build_default_crypto_provider().install_default() {
+fn install_embedded_default_crypto_provider() -> Result<()> {
+    if ensure_crypto_provider().map_err(std::io::Error::other)? == CryptoProviderStatus::AlreadyInstalled {
        debug!(
+            event = EVENT_CRYPTO_PROVIDER_STATE,
            component = LOG_COMPONENT_EMBEDDED,
            subsystem = LOG_SUBSYSTEM_EMBEDDED,
-            event = EVENT_CRYPTO_PROVIDER_STATE,
            state = "already_installed",
-            error = ?err,
            "Embedded crypto provider state changed"
        );
    }
+    Ok(())
 }

 #[cfg(test)]
 mod tests {
-    use super::{build_default_crypto_provider, init_profiling_runtime_with, shutdown_profiling_runtime_with};
+    use super::{
+        init_profiling_runtime_with, install_default_crypto_provider, install_embedded_default_crypto_provider,
+        shutdown_profiling_runtime_with,
+    };
+    use rustfs_tls_runtime::default_crypto_provider as build_default_crypto_provider;
    use std::sync::{
        Arc,
        atomic::{AtomicBool, Ordering},
    };

+    #[cfg(feature = "fips")]
    #[test]
    fn default_crypto_provider_matches_fips_feature() {
        let provider = build_default_crypto_provider();
-        assert_eq!(provider.fips(), cfg!(feature = "fips"));
+        assert!(provider.fips());
+    }
+
+    #[cfg(feature = "fips")]
+    #[test]
+    fn fips_provider_builds_fips_tls_configs() {
+        let provider = Arc::new(build_default_crypto_provider());
+        assert!(provider.cipher_suites.iter().all(|suite| suite.fips()));
+        assert!(provider.kx_groups.iter().all(|group| group.fips()));
+        assert!(
+            !provider
+                .cipher_suites
+                .iter()
+                .any(|suite| { suite.suite() == rustls::CipherSuite::TLS13_CHACHA20_POLY1305_SHA256 })
+        );
+        assert!(
+            !provider
+                .kx_groups
+                .iter()
+                .any(|group| group.name() == rustls::NamedGroup::X25519)
+        );
+
+        let client = rustls::ClientConfig::builder_with_provider(provider.clone())
+            .with_safe_default_protocol_versions()
+            .expect("supported protocol versions")
+            .with_root_certificates(rustls::RootCertStore::empty())
+            .with_no_client_auth();
+        let server = rustls::ServerConfig::builder_with_provider(provider)
+            .with_safe_default_protocol_versions()
+            .expect("supported protocol versions")
+            .with_no_client_auth()
+            .with_cert_resolver(Arc::new(rustls::server::ResolvesServerCertUsingSni::new()));
+        assert!(client.fips());
+        assert!(server.fips());
+    }
+
+    #[cfg(feature = "fips")]
+    fn tls_handshake_with_provider(
+        client_provider: rustls::crypto::CryptoProvider,
+        version: &'static rustls::SupportedProtocolVersion,
+    ) -> std::result::Result<(), rustls::Error> {
+        let cert = rcgen::generate_simple_self_signed(vec!["localhost".to_owned()]).expect("generate certificate");
+        let cert_der = cert.cert.der().clone();
+        let private_key = rustls::pki_types::PrivatePkcs8KeyDer::from(cert.signing_key.serialize_der());
+        let mut roots = rustls::RootCertStore::empty();
+        roots.add(cert_der.clone()).expect("trust certificate");
+
+        let server_config = rustls::ServerConfig::builder_with_provider(Arc::new(build_default_crypto_provider()))
+            .with_protocol_versions(&[version])
+            .expect("server protocol version")
+            .with_no_client_auth()
+            .with_single_cert(vec![cert_der], private_key.into())
+            .expect("server identity");
+        assert!(server_config.fips());
+        let client_config = rustls::ClientConfig::builder_with_provider(Arc::new(client_provider))
+            .with_protocol_versions(&[version])
+            .expect("client protocol version")
+            .with_root_certificates(roots)
+            .with_no_client_auth();
+
+        let mut server = rustls::ServerConnection::new(Arc::new(server_config)).expect("server connection");
+        let mut client = rustls::ClientConnection::new(
+            Arc::new(client_config),
+            rustls::pki_types::ServerName::try_from("localhost").expect("server name"),
+        )
+        .expect("client connection");
+        for _ in 0..8 {
+            let mut client_bytes = Vec::new();
+            client.write_tls(&mut client_bytes).expect("write client TLS");
+            if !client_bytes.is_empty() {
+                server
+                    .read_tls(&mut std::io::Cursor::new(client_bytes))
+                    .expect("read server TLS");
+                server.process_new_packets()?;
+            }
+
+            let mut server_bytes = Vec::new();
+            server.write_tls(&mut server_bytes).expect("write server TLS");
+            if !server_bytes.is_empty() {
+                client
+                    .read_tls(&mut std::io::Cursor::new(server_bytes))
+                    .expect("read client TLS");
+                client.process_new_packets()?;
+            }
+            if !client.is_handshaking() && !server.is_handshaking() {
+                return Ok(());
+            }
+        }
+        panic!("TLS handshake did not complete within eight exchanges");
+    }
+
+    #[cfg(feature = "fips")]
+    #[test]
+    fn fips_provider_tls_handshake_accepts_and_rejects_algorithms() {
+        for version in [&rustls::version::TLS12, &rustls::version::TLS13] {
+            tls_handshake_with_provider(build_default_crypto_provider(), version).expect("FIPS TLS handshake");
+            let mut client_provider = build_default_crypto_provider();
+            client_provider.kx_groups = vec![rustls::crypto::aws_lc_rs::kx_group::X25519];
+            assert!(tls_handshake_with_provider(client_provider, version).is_err());
+        }
+        let mut client_provider = build_default_crypto_provider();
+        client_provider.cipher_suites = vec![rustls::crypto::aws_lc_rs::cipher_suite::TLS13_CHACHA20_POLY1305_SHA256];
+        assert!(tls_handshake_with_provider(client_provider, &rustls::version::TLS13).is_err());
+    }
+
+    #[test]
+    fn provider_installation_isolated() {
+        const SCENARIO: &str = "RUSTFS_TEST_PROVIDER_SCENARIO";
+        if let Ok(scenario) = std::env::var(SCENARIO) {
+            let incompatible = || {
+                let mut provider = build_default_crypto_provider();
+                provider.kx_groups = vec![rustls::crypto::aws_lc_rs::kx_group::X25519];
+                assert!(!provider.fips());
+                provider
+            };
+            match scenario.as_str() {
+                "fresh" => {}
+                "compatible" => {
+                    build_default_crypto_provider().install_default().expect("install provider");
+                }
+                "incompatible" => {
+                    incompatible().install_default().expect("install provider");
+                }
+                "race" => {
+                    let barrier = Arc::new(std::sync::Barrier::new(2));
+                    let other_barrier = barrier.clone();
+                    let other = std::thread::spawn(move || {
+                        other_barrier.wait();
+                        let _ = incompatible().install_default();
+                    });
+                    barrier.wait();
+                    let result = install_default_crypto_provider();
+                    other.join().expect("provider installer thread");
+                    let actual_fips = rustls::crypto::CryptoProvider::get_default()
+                        .expect("provider installed")
+                        .fips();
+                    assert_eq!(result.is_ok(), !cfg!(feature = "fips") || actual_fips);
+                    assert_eq!(install_embedded_default_crypto_provider().is_ok(), !cfg!(feature = "fips") || actual_fips);
+                    return;
+                }
+                _ => panic!("unknown provider scenario"),
+            }
+            let result = install_default_crypto_provider();
+            let actual_fips = rustls::crypto::CryptoProvider::get_default()
+                .expect("provider installed")
+                .fips();
+            assert_eq!(result.is_ok(), !cfg!(feature = "fips") || actual_fips);
+            if cfg!(feature = "fips") && scenario == "incompatible" {
+                assert!(
+                    result
+                        .unwrap_err()
+                        .to_string()
+                        .contains("process-wide rustls crypto provider")
+                );
+                let runtime = tokio::runtime::Builder::new_current_thread()
+                    .enable_all()
+                    .build()
+                    .expect("embedded runtime");
+                let error = runtime
+                    .block_on(super::init_embedded_runtime_hooks(String::new()))
+                    .unwrap_err();
+                assert!(error.to_string().contains("process-wide rustls crypto provider"));
+                let config = crate::config::Config::new("127.0.0.1:9000", Vec::new());
+                let report = rustfs_utils::ExternalEnvCompatReport::default();
+                let error = runtime
+                    .block_on(crate::startup_preflight::init_startup_server_preflight(&config, &report))
+                    .unwrap_err();
+                assert!(error.to_string().contains("process-wide rustls crypto provider"));
+                assert!(rustfs_obs::get_global_guard().is_err(), "observability must not initialize first");
+            }
+            assert_eq!(install_embedded_default_crypto_provider().is_ok(), !cfg!(feature = "fips") || actual_fips);
+            return;
+        }
+
+        for scenario in ["fresh", "compatible", "incompatible", "race"] {
+            let output = std::process::Command::new(std::env::current_exe().expect("test executable"))
+                .arg("provider_installation_isolated")
+                .arg("--nocapture")
+                .env(SCENARIO, scenario)
+                .output()
+                .expect("start isolated test");
+            assert!(
+                output.status.success(),
+                "{scenario}: stdout={} stderr={}",
+                String::from_utf8_lossy(&output.stdout),
+                String::from_utf8_lossy(&output.stderr)
+            );
+        }
    }

    #[tokio::test]
@@ -44,45 +44,58 @@ pub(crate) fn table_matches_staged_base(table: &TableEntry, commit_log: &CommitL
 pub(crate) struct TableCommitHistoryIndex<'a> {
    table_id: &'a str,
    reachable_states: BTreeSet<(&'a str, &'a str)>,
+    reachable_version_tokens: BTreeSet<&'a str>,
+    committed_successors: BTreeMap<(&'a str, &'a str), &'a CommitLogEntry>,
    ambiguous_states: BTreeSet<(&'a str, &'a str)>,
    cycle_detected: bool,
+    version_token_reused: bool,
 }

 impl<'a> TableCommitHistoryIndex<'a> {
    pub(crate) fn new(table: &'a TableEntry, commits: impl IntoIterator<Item = &'a CommitLogEntry>) -> Self {
-        let mut by_new_state = BTreeMap::<(&str, &str), Option<(&str, &str)>>::new();
+        let mut by_new_state = BTreeMap::<(&str, &str), Option<&CommitLogEntry>>::new();
        let mut ambiguous_states = BTreeSet::new();
        for commit in commits
            .into_iter()
            .filter(|commit| commit.table_id == table.table_id && !matches!(commit.status, CommitLogStatus::Failed))
        {
            let key = (commit.new_metadata_location.as_str(), commit.new_version_token.as_str());
-            let previous = (commit.previous_metadata_location.as_str(), commit.expected_version_token.as_str());
            by_new_state
                .entry(key)
                .and_modify(|candidate| {
                    *candidate = None;
                    ambiguous_states.insert(key);
                })
-                .or_insert(Some(previous));
+                .or_insert(Some(commit));
        }

        let mut reachable_states = BTreeSet::new();
+        let mut reachable_version_tokens = BTreeSet::new();
+        let mut committed_successors = BTreeMap::new();
+        let mut version_token_reused = false;
        let mut state = (table.metadata_location.as_str(), table.version_token.as_str());
        let cycle_detected = loop {
            if !reachable_states.insert(state) {
                break true;
            }
-            let Some(Some(previous)) = by_new_state.get(&state) else {
+            version_token_reused |= !reachable_version_tokens.insert(state.1);
+            let Some(commit) = by_new_state.get(&state).copied().flatten() else {
                break false;
            };
-            state = *previous;
+            let previous = (commit.previous_metadata_location.as_str(), commit.expected_version_token.as_str());
+            if matches!(commit.status, CommitLogStatus::Committed) {
+                committed_successors.insert(previous, commit);
+            }
+            state = previous;
        };
        Self {
            table_id: &table.table_id,
            reachable_states,
+            reachable_version_tokens,
+            committed_successors,
            ambiguous_states,
            cycle_detected,
+            version_token_reused,
        }
    }

@@ -97,6 +110,49 @@ impl<'a> TableCommitHistoryIndex<'a> {
                .reachable_states
                .contains(&(target.new_metadata_location.as_str(), target.new_version_token.as_str()))
    }
+
+    fn proven_superseding_commit(&self, target: &CommitLogEntry) -> Option<&CommitLogEntry> {
+        if self.cycle_detected
+            || self.version_token_reused
+            || !self.ambiguous_states.is_empty()
+            || self.table_id != target.table_id
+            || !matches!(target.status, CommitLogStatus::Staged)
+            || self.reachable_version_tokens.contains(target.new_version_token.as_str())
+        {
+            return None;
+        }
+        let base = (target.previous_metadata_location.as_str(), target.expected_version_token.as_str());
+        self.committed_successors
+            .get(&base)
+            .copied()
+            .filter(|winner| winner.commit_id != target.commit_id)
+    }
+}
+
+/// Classifies an object-backed staged loser without rewriting its receipt.
+///
+/// A reachable finalized successor from the same pointer/token base proves that
+/// the old request lost CAS. Its STAGED record and idempotency binding remain
+/// available, and replay of that old request must still conflict. Missing or
+/// contradictory proof remains a manual-review state.
+pub(crate) fn table_commit_recovery_entry_from_history(
+    table: &TableEntry,
+    commit_log: &CommitLogEntry,
+    idempotency_commit: Option<&CommitLogEntry>,
+    history: &TableCommitHistoryIndex<'_>,
+) -> TableCommitRecoveryEntry {
+    let mut entry = table_commit_recovery_entry(table, commit_log, idempotency_commit, history.proves_committed(commit_log));
+    if entry.recovery_state == TableCommitRecoveryState::ManualReview
+        && matches!(
+            entry.idempotency_index_status,
+            TableCommitIdempotencyIndexStatus::NotRequired | TableCommitIdempotencyIndexStatus::Matches
+        )
+        && let Some(winner) = history.proven_superseding_commit(commit_log)
+    {
+        entry.recovery_state = TableCommitRecoveryState::Superseded;
+        entry.reason = format!("staged request lost CAS to reachable committed successor {}", winner.commit_id);
+    }
+    entry
 }

 pub(crate) fn table_catalog_recovery_summary(
@@ -1415,6 +1415,7 @@ pub(crate) struct TableCatalogDiagnosticsReport {
 #[serde(rename_all = "SCREAMING_SNAKE_CASE")]
 pub(crate) enum TableCommitRecoveryState {
    Committed,
+    Superseded,
    StagedBeforeTableUpdate,
    FinalizationRequired,
    IdempotencyIndexRepairRequired,
@@ -3065,11 +3065,11 @@ where
                }
                None => None,
            };
-            commits.push(table_commit_recovery_entry(
+            commits.push(table_commit_recovery_entry_from_history(
                entry,
                commit_log,
                idempotency_commit.as_ref(),
-                history.proves_committed(commit_log),
+                &history,
            ));
        }
        commits.sort_by(|left, right| left.commit_id.cmp(&right.commit_id));
@@ -3168,12 +3168,8 @@ where
                }
                None => None,
            };
-            let recovery_entry = table_commit_recovery_entry(
-                &entry,
-                commit_log,
-                idempotency_commit.as_ref(),
-                history.proves_committed(commit_log),
-            );
+            let recovery_entry =
+                table_commit_recovery_entry_from_history(&entry, commit_log, idempotency_commit.as_ref(), &history);
            if matches!(
                recovery_entry.recovery_state,
                TableCommitRecoveryState::FinalizationRequired | TableCommitRecoveryState::IdempotencyIndexRepairRequired
@@ -20,6 +20,7 @@ use std::sync::Arc;

 mod strong_maintenance_planning;
 mod strong_maintenance_planning_paths;
+mod superseded_commit_recovery;

 const TABLE_CATALOG_TEST_TIMEOUT: StdDuration = StdDuration::from_secs(30);

@@ -0,0 +1,487 @@
+// 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::*;
+
+struct StagedLoserFixture {
+    backend: TestCatalogObjectBackend,
+    store: ObjectTableCatalogStore<TestCatalogObjectBackend>,
+    loser: TableCommitRequest,
+    winner: TableCommitRequest,
+    committed: TableCommitResult,
+}
+
+async fn staged_loser_fixture(interrupt_winner_finalization: bool) -> StagedLoserFixture {
+    let backend = TestCatalogObjectBackend::default();
+    let store = ObjectTableCatalogStore::new(backend.clone());
+    let bucket = "analytics";
+    let namespace = Namespace::parse("sales").unwrap();
+    let table = IdentifierSegment::parse("orders").unwrap();
+    let initial = test_table_entry(
+        bucket,
+        &namespace,
+        &table,
+        default_table_metadata_file_path(&namespace, &table, "00001.metadata.json"),
+    );
+    store.put_table_bucket(test_bucket_entry(bucket)).await.unwrap();
+    store
+        .create_namespace(test_namespace_entry(bucket, &namespace))
+        .await
+        .unwrap();
+    store.create_table(initial.clone()).await.unwrap();
+    for name in ["00001", "loser", "winner", "reloaded"] {
+        let location = default_table_metadata_file_path(&namespace, &table, &format!("{name}.metadata.json"));
+        let mut metadata = super::super::test_support::table_metadata_json(&initial.table_uuid, &initial.warehouse_location);
+        metadata["properties"] = serde_json::json!({"writer": name});
+        backend
+            .seed_object(bucket, &location, serde_json::to_vec(&metadata).unwrap())
+            .await;
+    }
+    let paths = TableCatalogObjectPaths::default();
+    let table_path = paths.table_entry_path(bucket, &namespace, &table);
+    backend.fail_next_put(RUSTFS_META_BUCKET, &table_path).await;
+    let loser = TableCommitRequest {
+        table_bucket: bucket.to_string(),
+        namespace: namespace.public_name(),
+        table: table.as_str().to_string(),
+        commit_id: "paused-writer-a0".to_string(),
+        idempotency_key: Some("a0-request".to_string()),
+        operation: "set-properties".to_string(),
+        expected_version_token: initial.version_token.clone(),
+        expected_metadata_location: initial.metadata_location.clone(),
+        new_metadata_location: default_table_metadata_file_path(&namespace, &table, "loser.metadata.json"),
+        requirements: Vec::new(),
+        writer: None,
+    };
+    assert_matches!(store.commit_table(loser.clone()).await, Err(TableCatalogStoreError::Internal(_)));
+    let winner = TableCommitRequest {
+        commit_id: "winner-c1".to_string(),
+        idempotency_key: Some("c1-request".to_string()),
+        new_metadata_location: default_table_metadata_file_path(&namespace, &table, "winner.metadata.json"),
+        ..loser.clone()
+    };
+    if interrupt_winner_finalization {
+        let winner_log = paths.commit_log_entry_path(bucket, &initial.table_id, &winner.commit_id);
+        backend.fail_put_attempt(RUSTFS_META_BUCKET, &winner_log, 2).await;
+    }
+    let committed = store.commit_table(winner.clone()).await.unwrap();
+    StagedLoserFixture {
+        backend,
+        store,
+        loser,
+        winner,
+        committed,
+    }
+}
+
+#[tokio::test]
+async fn superseded_staged_commit_recovers_without_rewriting_receipts_or_pointer() {
+    let StagedLoserFixture {
+        backend,
+        store,
+        loser,
+        winner,
+        committed,
+    } = staged_loser_fixture(false).await;
+    let bucket = loser.table_bucket.as_str();
+    let before = backend.state.lock().await.objects.clone();
+
+    let recovered = store.recover_table_commits(bucket, "sales", "orders").await.unwrap();
+    assert_eq!(recovered.manual_review_count, 0, "a durable same-base winner proves the staged loser");
+    assert_eq!(recovered.finalized_count, 0, "a losing writer must never be finalized as committed");
+    let lost = recovered
+        .commits
+        .iter()
+        .find(|entry| entry.commit_id == loser.commit_id)
+        .unwrap();
+    assert_eq!(lost.status, CommitLogStatus::Staged);
+    assert_eq!(serde_json::to_value(&lost.recovery_state).unwrap(), "SUPERSEDED");
+    assert_eq!(lost.idempotency_index_status, TableCommitIdempotencyIndexStatus::Matches);
+    let after = backend.state.lock().await.objects.clone();
+    assert_eq!(after.len(), before.len());
+    for (path, original) in &before {
+        let actual = after.get(path).unwrap();
+        assert_eq!(actual.data, original.data, "recovery changed bytes at {path:?}");
+        assert_eq!(actual.etag, original.etag, "recovery rewrote {path:?}");
+    }
+
+    let restarted = ObjectTableCatalogStore::new(backend.clone());
+    let diagnostics = restarted.diagnose_table_catalog(bucket, "sales", "orders", 0).await.unwrap();
+    assert_eq!(diagnostics.recovery_status, TableCatalogRecoveryStatus::Healthy);
+    assert_matches!(restarted.commit_table(loser.clone()).await, Err(TableCatalogStoreError::Conflict(_)));
+    let mut stale_request = loser.clone();
+    stale_request.commit_id = "stale-new-request".to_string();
+    stale_request.idempotency_key = Some("stale-new-key".to_string());
+    assert_matches!(restarted.commit_table(stale_request).await, Err(TableCatalogStoreError::Conflict(_)));
+    let mut different_payload = loser.clone();
+    different_payload.operation = "replace".to_string();
+    assert_matches!(
+        restarted.commit_table(different_payload.clone()).await,
+        Err(TableCatalogStoreError::Conflict(_))
+    );
+    assert_eq!(
+        restarted.load_table(bucket, "sales", "orders").await.unwrap(),
+        Some(committed.table.clone())
+    );
+    assert_eq!(restarted.commit_table(winner).await.unwrap().table, committed.table);
+
+    let reloaded = TableCommitRequest {
+        commit_id: "reloaded-writer-a1".to_string(),
+        idempotency_key: Some("a1-request".to_string()),
+        expected_version_token: committed.table.version_token.clone(),
+        expected_metadata_location: committed.table.metadata_location.clone(),
+        new_metadata_location: default_table_metadata_file_path(
+            &Namespace::parse("sales").unwrap(),
+            &IdentifierSegment::parse("orders").unwrap(),
+            "reloaded.metadata.json",
+        ),
+        ..different_payload
+    };
+    let next = restarted.commit_table(reloaded.clone()).await.unwrap();
+    assert_eq!(next.table.generation, committed.table.generation + 1);
+    assert_eq!(restarted.commit_table(reloaded).await.unwrap().table, next.table);
+    let restarted = ObjectTableCatalogStore::new(backend);
+    assert_eq!(
+        restarted
+            .diagnose_table_catalog(bucket, "sales", "orders", 0)
+            .await
+            .unwrap()
+            .recovery_status,
+        TableCatalogRecoveryStatus::Healthy
+    );
+}
+
+#[tokio::test]
+async fn superseded_staged_commit_waits_for_durable_winner_finalization() {
+    let StagedLoserFixture {
+        backend,
+        store,
+        loser,
+        committed,
+        ..
+    } = staged_loser_fixture(true).await;
+    let paths = TableCatalogObjectPaths::default();
+    let log_path = paths.commit_log_entry_path(&loser.table_bucket, &committed.table.table_id, &loser.commit_id);
+    let before = backend.read_object(RUSTFS_META_BUCKET, &log_path).await.unwrap().unwrap();
+    let planned = store
+        .plan_table_commit_recovery("analytics", "sales", "orders")
+        .await
+        .unwrap();
+    assert_eq!(planned.manual_review_count, 1, "a merely staged winner is insufficient proof");
+    assert_eq!(planned.finalization_required_count, 1);
+
+    let recovered = store.recover_table_commits("analytics", "sales", "orders").await.unwrap();
+    assert_eq!(recovered.finalized_count, 1);
+    assert_eq!(recovered.manual_review_count, 0);
+    let loser_state = recovered
+        .commits
+        .iter()
+        .find(|entry| entry.commit_id == loser.commit_id)
+        .unwrap();
+    assert_eq!(loser_state.recovery_state, TableCommitRecoveryState::Superseded);
+    assert_eq!(loser_state.status, CommitLogStatus::Staged);
+    let after = backend.read_object(RUSTFS_META_BUCKET, &log_path).await.unwrap().unwrap();
+    assert_eq!(after.data, before.data);
+    assert_eq!(after.etag, before.etag);
+    assert_eq!(store.load_table("analytics", "sales", "orders").await.unwrap(), Some(committed.table));
+    assert_eq!(
+        store
+            .diagnose_table_catalog("analytics", "sales", "orders", 0)
+            .await
+            .unwrap()
+            .recovery_status,
+        TableCatalogRecoveryStatus::Healthy
+    );
+}
+
+#[tokio::test]
+async fn superseded_staged_commit_requires_retained_winner_proof_after_restart() {
+    let StagedLoserFixture {
+        backend,
+        store,
+        loser,
+        winner,
+        committed,
+    } = staged_loser_fixture(false).await;
+    let healthy = store.recover_table_commits("analytics", "sales", "orders").await.unwrap();
+    assert_eq!(healthy.manual_review_count, 0);
+    let winner_path =
+        TableCatalogObjectPaths::default().commit_log_entry_path("analytics", &committed.table.table_id, &winner.commit_id);
+    assert!(
+        backend
+            .state
+            .lock()
+            .await
+            .objects
+            .remove(&(RUSTFS_META_BUCKET.to_string(), winner_path))
+            .is_some()
+    );
+    let restarted = ObjectTableCatalogStore::new(backend);
+    let before = restarted.load_table("analytics", "sales", "orders").await.unwrap();
+    let report = restarted.recover_table_commits("analytics", "sales", "orders").await.unwrap();
+    assert_eq!(report.manual_review_count, 1);
+    assert_eq!(report.finalized_count, 0);
+    assert_eq!(report.commits[0].recovery_state, TableCommitRecoveryState::ManualReview);
+    assert_eq!(report.commits[0].status, CommitLogStatus::Staged);
+    assert_matches!(restarted.commit_table(loser).await, Err(TableCatalogStoreError::Conflict(_)));
+    assert_eq!(restarted.load_table("analytics", "sales", "orders").await.unwrap(), before);
+    let diagnostics = restarted
+        .diagnose_table_catalog("analytics", "sales", "orders", 0)
+        .await
+        .unwrap();
+    assert_eq!(diagnostics.recovery_status, TableCatalogRecoveryStatus::ManualReviewRequired);
+    assert!(
+        diagnostics
+            .recommended_actions
+            .contains(&TableCatalogRecoveryAction::ReviewCommitLog)
+    );
+}
+
+#[tokio::test]
+async fn superseded_staged_commit_does_not_repair_a_conflicting_index() {
+    let StagedLoserFixture {
+        backend,
+        store,
+        loser,
+        committed,
+        ..
+    } = staged_loser_fixture(false).await;
+    let mut conflicting = store
+        .get_commit_by_id("analytics", &committed.table.table_id, &loser.commit_id)
+        .await
+        .unwrap()
+        .unwrap();
+    conflicting.new_version_token = "contradictory-token".to_string();
+    let path = TableCatalogObjectPaths::default().commit_idempotency_entry_path(
+        "analytics",
+        &committed.table.table_id,
+        loser.idempotency_key.as_deref().unwrap(),
+    );
+    backend
+        .seed_object(RUSTFS_META_BUCKET, &path, serde_json::to_vec(&conflicting).unwrap())
+        .await;
+    let before = backend.state.lock().await.objects.clone();
+    let report = store.recover_table_commits("analytics", "sales", "orders").await.unwrap();
+    assert_eq!(report.manual_review_count, 1);
+    assert_eq!(report.finalized_count, 0);
+    let loser_state = report
+        .commits
+        .iter()
+        .find(|entry| entry.commit_id == loser.commit_id)
+        .unwrap();
+    assert_eq!(loser_state.recovery_state, TableCommitRecoveryState::ManualReview);
+    assert_eq!(loser_state.idempotency_index_status, TableCommitIdempotencyIndexStatus::Conflicting);
+    let after = backend.state.lock().await.objects.clone();
+    assert_eq!(before.len(), after.len());
+    for (key, original) in &before {
+        let actual = after.get(key).unwrap();
+        assert_eq!(actual.data, original.data);
+        assert_eq!(actual.etag, original.etag);
+    }
+    assert_matches!(store.commit_table(loser).await, Err(TableCatalogStoreError::Conflict(_)));
+}
+
+fn history_fixture() -> (TableEntry, CommitLogEntry, CommitLogEntry) {
+    let namespace = Namespace::parse("sales").unwrap();
+    let table_name = IdentifierSegment::parse("orders").unwrap();
+    let metadata = |name| default_table_metadata_file_path(&namespace, &table_name, name);
+    let mut table = test_table_entry("analytics", &namespace, &table_name, metadata("winner.metadata.json"));
+    table.version_token = "winner-token".to_string();
+    table.generation = 2;
+    let loser = CommitLogEntry {
+        version: TABLE_CATALOG_ENTRY_VERSION,
+        commit_id: "loser".to_string(),
+        idempotency_key: None,
+        table_id: table.table_id.clone(),
+        operation: "set-properties".to_string(),
+        expected_version_token: "base-token".to_string(),
+        new_version_token: "loser-token".to_string(),
+        previous_metadata_location: metadata("base.metadata.json"),
+        new_metadata_location: metadata("loser.metadata.json"),
+        requirements: Vec::new(),
+        status: CommitLogStatus::Staged,
+        writer: None,
+        created_at: None,
+        updated_at: None,
+    };
+    let winner = CommitLogEntry {
+        commit_id: "winner".to_string(),
+        new_version_token: table.version_token.clone(),
+        new_metadata_location: table.metadata_location.clone(),
+        status: CommitLogStatus::Committed,
+        ..loser.clone()
+    };
+    (table, loser, winner)
+}
+
+fn classify_history(
+    table: &TableEntry,
+    target: &CommitLogEntry,
+    commits: &[CommitLogEntry],
+    indexed: Option<&CommitLogEntry>,
+) -> TableCommitRecoveryEntry {
+    let history = TableCommitHistoryIndex::new(table, commits.iter());
+    table_commit_recovery_entry_from_history(table, target, indexed, &history)
+}
+
+#[test]
+fn superseded_history_requires_a_reachable_finalized_same_base_winner() {
+    for case in [
+        "staged-winner",
+        "failed-winner",
+        "wrong-base-token",
+        "wrong-base-location",
+        "foreign-table",
+        "disconnected-winner",
+        "same-commit-id",
+    ] {
+        let (mut table, loser, mut winner) = history_fixture();
+        match case {
+            "staged-winner" => winner.status = CommitLogStatus::Staged,
+            "failed-winner" => winner.status = CommitLogStatus::Failed,
+            "wrong-base-token" => winner.expected_version_token = "other-base-token".to_string(),
+            "wrong-base-location" => winner.previous_metadata_location = "other/base.metadata.json".to_string(),
+            "foreign-table" => winner.table_id = "other-table-id".to_string(),
+            "disconnected-winner" => {
+                table.metadata_location = "other/current.metadata.json".to_string();
+                table.version_token = "other-current-token".to_string();
+            }
+            "same-commit-id" => winner.commit_id.clone_from(&loser.commit_id),
+            _ => unreachable!(),
+        }
+        let recovery = classify_history(&table, &loser, &[loser.clone(), winner], None);
+        assert_eq!(recovery.recovery_state, TableCommitRecoveryState::ManualReview, "{case}");
+    }
+}
+
+#[test]
+fn superseded_history_is_stable_after_later_commits() {
+    let (mut table, loser, winner) = history_fixture();
+    let immediate = classify_history(&table, &loser, &[loser.clone(), winner.clone()], None);
+    assert_eq!(immediate.recovery_state, TableCommitRecoveryState::Superseded);
+    assert!(immediate.reason.contains(&winner.commit_id));
+    let later = CommitLogEntry {
+        commit_id: "later".to_string(),
+        previous_metadata_location: table.metadata_location.clone(),
+        expected_version_token: table.version_token.clone(),
+        new_metadata_location: "later/current.metadata.json".to_string(),
+        new_version_token: "later-token".to_string(),
+        ..winner.clone()
+    };
+    table.metadata_location.clone_from(&later.new_metadata_location);
+    table.version_token.clone_from(&later.new_version_token);
+    table.generation += 1;
+    let after = classify_history(&table, &loser, &[loser.clone(), winner, later], None);
+    assert_eq!(after.recovery_state, TableCommitRecoveryState::Superseded);
+    assert_eq!(after.reason, immediate.reason);
+}
+
+#[test]
+fn superseded_history_rejects_failed_or_committed_losers() {
+    for status in [CommitLogStatus::Failed, CommitLogStatus::Committed] {
+        let (table, mut loser, winner) = history_fixture();
+        loser.status = status;
+        let recovery = classify_history(&table, &loser, &[loser.clone(), winner], None);
+        assert_eq!(recovery.recovery_state, TableCommitRecoveryState::ManualReview);
+    }
+}
+
+#[test]
+fn superseded_history_rejects_ambiguity_cycles_and_reused_tokens() {
+    for case in [
+        "ambiguous-current",
+        "cycle",
+        "reused-history-token",
+        "loser-token-in-history",
+        "off-history-ambiguity",
+    ] {
+        let (mut table, mut loser, winner) = history_fixture();
+        let mut extra = CommitLogEntry {
+            commit_id: "extra".to_string(),
+            ..winner.clone()
+        };
+        let mut commits = vec![loser.clone(), winner.clone()];
+        match case {
+            "ambiguous-current" => commits.push(extra),
+            "cycle" => {
+                extra.previous_metadata_location.clone_from(&winner.new_metadata_location);
+                extra.expected_version_token.clone_from(&winner.new_version_token);
+                extra.new_metadata_location.clone_from(&winner.previous_metadata_location);
+                extra.new_version_token.clone_from(&winner.expected_version_token);
+                commits.push(extra);
+            }
+            "reused-history-token" => {
+                extra.previous_metadata_location.clone_from(&winner.new_metadata_location);
+                extra.expected_version_token.clone_from(&winner.new_version_token);
+                extra.new_metadata_location = "later/current.metadata.json".to_string();
+                extra.new_version_token.clone_from(&winner.expected_version_token);
+                table.metadata_location.clone_from(&extra.new_metadata_location);
+                table.version_token.clone_from(&extra.new_version_token);
+                commits.push(extra);
+            }
+            "loser-token-in-history" => {
+                loser.new_version_token.clone_from(&winner.expected_version_token);
+                commits[0] = loser.clone();
+            }
+            "off-history-ambiguity" => {
+                extra.new_metadata_location = "unknown/metadata.json".to_string();
+                extra.new_version_token = "unknown-token".to_string();
+                commits.push(extra.clone());
+                extra.commit_id = "another-extra".to_string();
+                commits.push(extra);
+            }
+            _ => unreachable!(),
+        }
+        assert_eq!(
+            classify_history(&table, &loser, &commits, None).recovery_state,
+            TableCommitRecoveryState::ManualReview,
+            "{case}"
+        );
+    }
+}
+
+#[test]
+fn superseded_history_requires_an_intact_idempotency_binding() {
+    let (table, mut loser, winner) = history_fixture();
+    loser.idempotency_key = Some("loser-request".to_string());
+    let commits = [loser.clone(), winner];
+    assert_eq!(
+        classify_history(&table, &loser, &commits, Some(&loser)).recovery_state,
+        TableCommitRecoveryState::Superseded
+    );
+    let missing = classify_history(&table, &loser, &commits, None);
+    assert_eq!(missing.recovery_state, TableCommitRecoveryState::ManualReview);
+    assert_eq!(missing.idempotency_index_status, TableCommitIdempotencyIndexStatus::Missing);
+    let mut stale = loser.clone();
+    stale.status = CommitLogStatus::Committed;
+    let state = classify_history(&table, &loser, &commits, Some(&stale));
+    assert_eq!(state.recovery_state, TableCommitRecoveryState::ManualReview);
+    assert_eq!(state.idempotency_index_status, TableCommitIdempotencyIndexStatus::Stale);
+    let mut conflicting = loser.clone();
+    conflicting.commit_id = "another-request".to_string();
+    let state = classify_history(&table, &loser, &commits, Some(&conflicting));
+    assert_eq!(state.recovery_state, TableCommitRecoveryState::ManualReview);
+    assert_eq!(state.idempotency_index_status, TableCommitIdempotencyIndexStatus::Conflicting);
+}
+
+#[test]
+fn superseded_history_never_relabels_a_committed_staged_ancestor() {
+    let (table, mut ancestor, winner) = history_fixture();
+    ancestor.new_metadata_location.clone_from(&winner.previous_metadata_location);
+    ancestor.new_version_token.clone_from(&winner.expected_version_token);
+    ancestor.previous_metadata_location = "older/base.metadata.json".to_string();
+    ancestor.expected_version_token = "older-base-token".to_string();
+    let recovery = classify_history(&table, &ancestor, &[ancestor.clone(), winner], None);
+    assert_eq!(recovery.recovery_state, TableCommitRecoveryState::FinalizationRequired);
+}
@@ -540,7 +540,7 @@ struct CleanupObservation {
    elapsed: Duration,
 }

-async fn observe_cleanup(replies: [CleanupReply; 2]) -> CleanupObservation {
+async fn observe_cleanup(replies: [CleanupReply; 2], measurement_delay: Duration) -> CleanupObservation {
    use tokio::io::{AsyncReadExt as _, AsyncWriteExt as _};

    const PRIVATE_BODY: &str = "private-cleanup-response-body";
@@ -583,6 +583,12 @@ async fn observe_cleanup(replies: [CleanupReply; 2]) -> CleanupObservation {
                .to_owned();
            let reply = if index < 2 {
                assert!(request_line.starts_with("PUT "));
+                if index == 1 && !measurement_delay.is_zero() {
+                    // Model descheduling during the measured PUT without a wall-clock sleep.
+                    tokio::time::pause();
+                    tokio::time::advance(measurement_delay).await;
+                    tokio::time::resume();
+                }
                CleanupReply::Status(200)
            } else {
                assert!(request_line.starts_with("DELETE "));
@@ -616,6 +622,8 @@ async fn observe_cleanup(replies: [CleanupReply; 2]) -> CleanupObservation {
    )
    .expect("cleanup probe");
    let mut request = request(ObjectOperation::PutObject);
+    // Status fixtures must reach cleanup independently of short measurement deadlines.
+    request.duration = REAL_SERVER_MEASUREMENT_BUDGET;
    request.traffic_bytes = 4;
    let log = CleanupLog::default();
    let writer = log.clone();
@@ -707,7 +715,7 @@ async fn cleanup_diagnostics_identify_each_failed_http_stage_without_private_con
        ([CleanupReply::Status(204), CleanupReply::Status(403)], "bucket_delete", 403),
        ([CleanupReply::Status(204), CleanupReply::Status(404)], "bucket_delete", 404),
    ] {
-        let observed = observe_cleanup(replies).await;
+        let observed = observe_cleanup(replies, Duration::ZERO).await;
        assert_cleanup_failure(&observed, stage, "CleanupFailed", Some(status));
        assert_eq!(observed.deletes.len(), if stage == "object_delete" { 1 } else { 2 });
    }
@@ -720,7 +728,7 @@ async fn cleanup_diagnostics_identify_each_disconnected_transport_stage() {
        ([CleanupReply::Disconnect, CleanupReply::Status(204)], "object_delete"),
        ([CleanupReply::Status(204), CleanupReply::Disconnect], "bucket_delete"),
    ] {
-        let observed = observe_cleanup(replies).await;
+        let observed = observe_cleanup(replies, Duration::ZERO).await;
        assert_cleanup_failure(&observed, stage, "ProtocolFailure", None);
    }
 }
@@ -728,13 +736,21 @@ async fn cleanup_diagnostics_identify_each_disconnected_transport_stage() {
 #[tokio::test]
 async fn cleanup_accepts_missing_object_and_still_deletes_the_bucket() {
    let _guard = TEST_LOCK.lock().await;
-    let observed = observe_cleanup([CleanupReply::Status(404), CleanupReply::Status(204)]).await;
-    assert_eq!(observed.measurement.result.outcome(), ObjectOutcome::Succeeded);
-    assert_eq!(observed.measurement.target.reason_code, ObjectTargetReasonCode::Complete);
-    assert_eq!(observed.deletes.len(), 2);
-    assert!(observed.deletes[0].ends_with("/synthetic-object HTTP/1.1"));
-    assert!(observed.deletes[1].ends_with("/ HTTP/1.1"));
-    assert!(!observed.log.contains("connect_object_cleanup_failed"));
+    for measurement_delay in [Duration::ZERO, Duration::from_secs(2)] {
+        let observed = observe_cleanup([CleanupReply::Status(404), CleanupReply::Status(204)], measurement_delay).await;
+        assert_eq!(
+            observed.measurement.result.outcome(),
+            ObjectOutcome::Succeeded,
+            "target result: {:?}; cleanup diagnostics: {}",
+            observed.measurement.target,
+            observed.log
+        );
+        assert_eq!(observed.measurement.target.reason_code, ObjectTargetReasonCode::Complete);
+        assert_eq!(observed.deletes.len(), 2);
+        assert!(observed.deletes[0].ends_with("/synthetic-object HTTP/1.1"));
+        assert!(observed.deletes[1].ends_with("/ HTTP/1.1"));
+        assert!(!observed.log.contains("connect_object_cleanup_failed"));
+    }
 }

 #[tokio::test]
@@ -750,7 +766,7 @@ async fn cleanup_diagnostics_identify_timeouts_without_resetting_the_shared_dead
            "bucket_delete",
        ),
    ] {
-        let observed = observe_cleanup(replies).await;
+        let observed = observe_cleanup(replies, Duration::ZERO).await;
        assert_cleanup_failure(&observed, stage, "TimedOut", None);
        assert!(observed.elapsed >= Duration::from_secs(5));
        assert!(observed.log.contains("remaining_ms=0"), "{}", observed.log);
@@ -16,7 +16,7 @@ import unittest
 ROOT = Path(__file__).resolve().parent.parent
 ALWAYS_JOBS = ("classify-changes", "typos", "quick-checks")
 CODE_JOBS = (
-    "workspace-lint-and-build", "workspace-tests", "core-contract", "connect-feature-boundaries", "test-and-lint-rio-v2",
+    "workspace-lint-and-build", "workspace-tests", "core-contract", "connect-feature-boundaries", "fips-provider", "test-and-lint-rio-v2",
    "test-and-lint-protocols", "build-rustfs-debug-binary", "uring-integration",
    "e2e-tests", "s3-implemented-tests", "s3-lifecycle-behavior-tests",
 )
@@ -28,12 +28,12 @@ CODE_JOBS = (
 # wall). Workspace lint and tests, the short core contract, the debug build,
 # E2E smoke and the S3 lanes stay on the pull-request path because they cover
 # the product's core correctness and S3 compatibility surface.
-NIGHTLY_JOBS = ("connect-feature-boundaries", "uring-integration", "test-and-lint-rio-v2", "test-and-lint-protocols")
+NIGHTLY_JOBS = ("connect-feature-boundaries", "fips-provider", "uring-integration", "test-and-lint-rio-v2", "test-and-lint-protocols")
 # Bump when a plan produced by an older `select` can no longer be checked
 # against this revision's expected job contract. The workflow refuses a plan
 # that does not carry the current marker and falls back to the full matrix, so
 # a base planner predating the contract cannot authorize a narrower selection.
-PLAN_CONTRACT = "2"
+PLAN_CONTRACT = "3"
 # These jobs still exercise the server and its black-box harness when only E2E
 # Rust sources change. Production code, manifests and shared test configuration
 # always select the full matrix. No production package depends on e2e_test.
@@ -210,6 +210,17 @@ def configurations(jobs: dict[str, str]) -> list[dict]:
        ("connect-offline-tests", "connect-feature-boundaries", "test", ["offline-enrollment-e2e-root"]),
        ("connect-release", "connect-feature-boundaries", "release", []),
        ("connect-release-rejection", "connect-feature-boundaries", "release", ["connect-e2e-short-credentials"]),
+        ("fips-provider-default-lint", "fips-provider", "dev", []),
+        ("fips-provider-fips-lint", "fips-provider", "dev", ["fips"]),
+        ("fips-tls-runtime-default", "fips-provider", "test", []),
+        ("fips-tls-runtime-fips", "fips-provider", "test", ["fips"]),
+        ("fips-protocols", "fips-provider", "test", ["ftps", "webdav", "fips"]),
+        ("fips-targets", "fips-provider", "test", ["fips"]),
+        ("fips-s3-client", "fips-provider", "test", ["fips"]),
+        ("fips-targets-default", "fips-provider", "test", []),
+        ("fips-s3-client-default", "fips-provider", "test", []),
+        ("fips-provider-default", "fips-provider", "test", []),
+        ("fips-provider-fips", "fips-provider", "test", ["fips"]),
        ("rio-v2-lint", "test-and-lint-rio-v2", "dev", ["rio-v2"]),
        ("rio-v2-tests", "test-and-lint-rio-v2", "test", ["rio-v2"]),
        ("protocol-swift-lint", "test-and-lint-protocols", "dev", ["swift"]),
@@ -234,7 +234,7 @@ class PlannerTests(unittest.TestCase):
        self.assertEqual(result.returncode, 0, result.stderr)
        values = dict(line.split('=', 1) for line in output.read_text().splitlines())
        self.assertEqual(set(values), {'mode', 'plan', 'connect', 'packages', 'full_validation', 'rehearse_shards', 'plan_contract'})
-        self.assertEqual(values['plan_contract'], '2')
+        self.assertEqual(values['plan_contract'], gate.PLAN_CONTRACT)
        plan = json.loads(values['plan'])
        self.assertEqual(plan['policy']['revision'], self.repo.base)
        self.assertEqual(plan['policy']['rollout'], 'shadow')
@@ -242,6 +242,15 @@ class PlannerTests(unittest.TestCase):


 class GateTests(unittest.TestCase):
+    def test_fips_provider_is_required_at_nightly_and_skipped_on_pr(self):
+        self.assertEqual(gate.expected_results('full', 'schedule', 'refs/heads/main')['fips-provider'], 'success')
+        self.assertEqual(gate.expected_results('full', 'pull_request', 'refs/heads/main')['fips-provider'], 'skipped')
+        plan = self.plan('schedule')
+        good = self.needs(plan)
+        self.assertEqual(gate.verify_results(good, 'schedule', plan['ref'], plan['head'], plan['tree']), [])
+        good['fips-provider']['result'] = 'skipped'
+        self.assertTrue(gate.verify_results(good, 'schedule', plan['ref']))
+
    def plan(self, event='pull_request', mode='full', connect=True):
        jobs = gate.expected_results(mode, event, 'refs/heads/main', connect)
        return {'schema_version': 1, 'event': event, 'ref': 'refs/heads/main', 'head': 'a' * 40, 'tree': 'b' * 40,
@@ -220,6 +220,126 @@ SH
        self.assertNotIn("        continue-on-error: true", self.publish)
        self.assertNotIn("          overwrite: true", upload)

+    def dispatch_shell(self):
+        step = self.steps["Dispatch nightly candidate to auto-testing"]
+        start = step.index("        run: |") + 1
+        return "\n".join(line[10:] for line in step[start:] if not line.strip() or line.startswith("          "))
+
+    def candidate_manifest(self, **overrides):
+        digest = overrides.pop("package_sha256", self.digest)
+        workflow_sha = overrides.pop("workflow_sha", self.sha)
+        manifest = {
+            "schema": 2,
+            "workflow_sha": workflow_sha,
+            "source_ref": "main",
+            "source_sha": self.sha,
+            "build_run_id": 12345,
+            "build_run_attempt": 1,
+            "package_sha256": digest,
+            "package_url": f"https://dl.rustfs.com/artifacts/rustfs/packages/nightly/runs/12345/1/{digest}/rustfs.deb",
+        }
+        manifest.update(overrides)
+        return manifest
+
+    def run_dispatch(self, manifest=None, **overrides):
+        manifest = self.candidate_manifest() if manifest is None else manifest
+        candidate = self.root / "candidate.json"
+        candidate.write_text(json.dumps(manifest))
+        shim = self.root / "dispatch-curl.sh"
+        shim.write_text(r'''curl() {
+  printf '%s\0' "$@" > "$FAKE_CURL_ARGS"
+  local outfile="" data=""
+  local args=("$@")
+  local i=0
+  while [[ $i -lt ${#args[@]} ]]; do
+    case "${args[$i]}" in
+      -o) outfile="${args[$((i+1))]}"; i=$((i+2));;
+      --data-binary)
+        data="${args[$((i+1))]}"
+        cp "${data#@}" "$FAKE_DISPATCH_BODY"
+        i=$((i+2));;
+      *) i=$((i+1));;
+    esac
+  done
+  [[ "$FAKE_DISPATCH_STATUS" != "curl-failure" ]] || return 28
+  if [[ -n "$outfile" ]]; then printf '%s' "$FAKE_DISPATCH_RESPONSE" > "$outfile"; fi
+  printf '%s' "$FAKE_DISPATCH_STATUS"
+}
+''')
+        env = dict(os.environ, BASH_ENV=str(shim), CANDIDATE_FILE=str(candidate), DEB_DATE="2026-09-06",
+                   SOURCE_REPO="rustfs/rustfs", DISPATCH_TOKEN="test-dispatch-token", RUNNER_TEMP=str(self.root),
+                   GITHUB_RUN_ID="12345", GITHUB_RUN_ATTEMPT="1", GITHUB_SHA=manifest["workflow_sha"],
+                   FAKE_CURL_ARGS=str(self.root / "curl-args"), FAKE_DISPATCH_BODY=str(self.root / "dispatch-body.json"),
+                   FAKE_DISPATCH_STATUS="204", FAKE_DISPATCH_RESPONSE="")
+        env.update(overrides)
+        return subprocess.run(["bash", "--noprofile", "--norc", "-e", "-o", "pipefail", "-c", self.dispatch_shell()],
+                              cwd=self.root, env=env, capture_output=True, text=True)
+
+    def test_dispatch_follows_release_publish_without_blocking_vault_lanes(self):
+        names = list(self.steps)
+        self.assertLess(names.index("Publish packages to auto-testing release assets"), names.index("Dispatch nightly candidate to auto-testing"))
+        step = "\n".join(self.steps["Dispatch nightly candidate to auto-testing"])
+        self.assertIn("        if: ${{ steps.publish.outputs.candidate_file != '' }}", step)
+        self.assertIn("          DISPATCH_TOKEN: ${{ secrets.AUTO_TESTING_DISPATCH_TOKEN }}", step)
+        self.assertNotIn("secrets.GITHUB_TOKEN", step)
+        self.assertNotIn("github.token", step)
+        self.assertNotIn("continue-on-error", step)
+        self.assertIn("https://api.github.com/repos/rustfs/auto-testing/dispatches", step)
+        self.assertIn("nightly-candidate-ready", step)
+        lines = WORKFLOW.read_text().splitlines()
+        for lane in ("kms-vault-lane", "kms-vault-ha-failover"):
+            job = yaml_block(lines, lane, 2)
+            needs = next(line for line in job if line.strip().startswith("needs:"))
+            self.assertEqual(needs.strip(), "needs: resolve-source")
+            self.assertNotIn("Dispatch nightly candidate", "\n".join(job))
+        alert = "\n".join(yaml_block(lines, "alert-on-failure", 2))
+        self.assertIn("needs: [build, kms-vault-lane, kms-vault-ha-failover]", alert)
+        chain = (ROOT / ".github/workflows/rustfs-functional-chain.yml").read_text()
+        self.assertIn('workflows: ["Nightly GNU Build"]', chain)
+        self.assertIn("types: [completed]", chain)
+
+    def test_dispatch_sends_schema2_candidate_without_the_token(self):
+        result = self.run_dispatch()
+        self.assertEqual(result.returncode, 0, result.stderr)
+        body = json.loads((self.root / "dispatch-body.json").read_text())
+        self.assertEqual(body["event_type"], "nightly-candidate-ready")
+        payload = body["client_payload"]
+        self.assertLessEqual(len(payload), 10)
+        self.assertEqual(set(payload), {"event_type", "source_repo", "latest_alias_url", "auto_testing_release", "channel", "deb_date", "nightly_workflow", "candidate"})
+        self.assertEqual(payload["candidate"], self.candidate_manifest())
+        self.assertEqual(payload["latest_alias_url"], "https://dl.rustfs.com/artifacts/rustfs/packages/nightly/rustfs-nightly-latest.deb")
+        self.assertEqual(payload["auto_testing_release"], "https://github.com/rustfs/auto-testing/releases/tag/nightly")
+        self.assertEqual(payload["channel"], "nightly")
+        self.assertEqual(payload["deb_date"], "2026-09-06")
+        self.assertEqual(payload["nightly_workflow"], ".github/workflows/nightly-gnu.yml")
+        self.assertEqual(payload["source_repo"], "rustfs/rustfs")
+        validate_manifest(payload["candidate"], {"id": 12345, "run_attempt": 1, "head_sha": self.sha})
+        args = (self.root / "curl-args").read_bytes().split(b"\0")
+        self.assertIn(b"https://api.github.com/repos/rustfs/auto-testing/dispatches", args)
+        self.assertIn(b"Authorization: Bearer test-dispatch-token", args)
+        self.assertNotIn("test-dispatch-token", json.dumps(body))
+        self.assertIn("dispatched nightly-candidate-ready", result.stdout)
+
+    def test_dispatch_failure_stops_before_treating_the_candidate_as_notified(self):
+        for status, response in (("422", '{"message":"No more than 10 properties are allowed"}'), ("500", "unavailable"), ("curl-failure", "")):
+            with self.subTest(status=status):
+                result = self.run_dispatch(FAKE_DISPATCH_STATUS=status, FAKE_DISPATCH_RESPONSE=response)
+                self.assertNotEqual(result.returncode, 0, result.stdout)
+                self.assertIn("repository_dispatch", result.stderr)
+                if response:
+                    self.assertIn(response, result.stderr)
+                self.assertNotIn("dispatched nightly-candidate-ready", result.stdout)
+        for marker in ("curl-args", "dispatch-body.json"):
+            (self.root / marker).unlink(missing_ok=True)
+        missing = self.run_dispatch(DISPATCH_TOKEN="")
+        self.assertNotEqual(missing.returncode, 0)
+        self.assertIn("AUTO_TESTING_DISPATCH_TOKEN is required", missing.stderr)
+        self.assertFalse((self.root / "curl-args").exists())
+        rejected = self.run_dispatch(manifest=self.candidate_manifest(extra="nope"))
+        self.assertNotEqual(rejected.returncode, 0, rejected.stdout)
+        self.assertIn("nightly-candidate-ready contract", rejected.stderr)
+        self.assertFalse((self.root / "dispatch-body.json").exists())
+

 class ValidationPackageTests(unittest.TestCase):
    def setUp(self):
Author	SHA1	Message	Date
Boyi (Elliot) Li	cc08e0d11b	feat(fips): route internode RPC HMAC through aws-lc-rs FIPS (#9137 ) * feat(fips): route internode RPC HMAC through aws-lc-rs FIPS * Revise KMS cryptographic compliance details Updated the KMS cryptographic compliance documentation to reflect changes in HMAC authentication and added details about the internode RPC HMAC authentication. Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com> Signed-off-by: Hauser Zhi <housemecn@gmail.com> --------- Signed-off-by: Hauser Zhi <housemecn@gmail.com> Co-authored-by: Hauser Zhi <housemecn@gmail.com> Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>	2026-11-03 06:56:56 +00:00
Hauser Zhi	69294523d5	fix(heal): preserve lock, format, and repair identity (#9132 ) * fix(heal): preserve lock, format, and repair identity Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * fix(ecstore): keep pool meta legacy error message stable for ratchet Co-authored-by: houseme <4829346+houseme@users.noreply.github.com> * fix(heal): drop redundant clone in MRF dedup test Co-authored-by: houseme <4829346+houseme@users.noreply.github.com> --------- Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com> Co-authored-by: Chris An <anzhengchao@gmail.com> Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com> Co-authored-by: houseme <4829346+houseme@users.noreply.github.com>	2026-11-03 04:59:06 +00:00
Chris An	cd4d86f28a	fix(replication): finish writeback for already purged versions (#9135 )	2026-11-03 02:08:32 +00:00
Hauser Zhi	2f55cf69e1	perf(filemeta): grow metacache read chunks geometrically (#9133 ) Large metacache elements (an xl.meta grown past 16 MiB with its version history) were accumulated in fixed 64 KiB reads, paying hundreds of read rounds per entry on the listing path. Double the per-round chunk up to a 4 MiB bound so each reservation stays incremental and bounded (rustfs/rustfs#2715 corrupt-length-prefix safety is unchanged) while wide listings and oversized entries decode in far fewer rounds. Add regression coverage for rustfs/rustfs#9120: a listing whose aggregate stream exceeds the per-value 16 MiB bound, and an oversized entry listed among ordinary neighbors with peek/skip interleaving.	2026-11-03 06:07:41 +08:00
Hauser Zhi	fff6e94378	fix(metrics): reconcile peer-health gauge on successful RPCs and probes (#9134 ) The rustfs_cluster_servers_offline_total gauge was event-driven only: record_peer_reachable fired solely after a fresh connector.connect() succeeded. When post-recovery traffic was served by a cached channel (tonic auto-reconnect, addr-form cache-key split, or delete-recovery channel), no fresh dial occurred and the gauge stayed stale until the process restarted. Add record_peer_reachable_if_offline (fast-path read / slow-path write) and wire it into every successful peer interaction: - remote-disk recovery probe success (gRPC DiskInfo) - periodic TCP connectivity check success (bounded-interval reconcile) - data-path RPC success (execute_with_timeout_for_op_and_health_action) - lock-path RPC success (record_rpc_success and detached completion) Fixes rustfs/rustfs#9119 Tracking: rustfs/backlog#3138	2026-11-02 19:07:37 +00:00
Hauser Zhi	41b13aeb8b	fix(tls): unify FIPS provider policy across clients (#9125 ) * fix(tls): enforce active FIPS provider at startup Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * ci: provision FIPS validation runner Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * fix(tls): validate FIPS before observability startup Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * perf(tls): reuse the installed crypto provider Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * fix(ecstore): gate Linux-only commit helpers Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * refactor(tls): centralize FIPS provider policy Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> --------- Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com> Co-authored-by: Chris An <anzhengchao@gmail.com>	2026-11-02 16:58:42 +00:00
weile321	88587a8f94	fix(s3select-api): clamp overflowing u64 GetRange values to i64::MAX (#9118 )	2026-11-02 14:05:07 +00:00
hector	8c8b5da63f	ci(nightly): dispatch nightly-candidate-ready after asset publish (#9131 ) After the nightly packages are on the auto-testing nightly release and the schema 2 R2 candidate exists, post repository_dispatch to rustfs/auto-testing. Vault lanes stay parallel, and the workflow_run functional chain is left in place. Co-authored-by: hector <majinghe@users.noreply.github.com>	2026-11-02 23:05:07 +08:00
Henry Guo	00a238f838	fix(table-catalog): resolve superseded staged commits (#9124 ) fix(table-catalog): diagnose superseded staged commits Co-authored-by: Chris An <anzhengchao@gmail.com>	2026-11-02 23:04:22 +08:00
hector	965d805a4f	ci: run distributed and upgrade e2e on schedule only (#9126 ) Remove pull_request from the distributed and upgrade workflows. Distributed stays on the nightly cron, and upgrade stays on the Monday cron plus version tags. Both remain manual-dispatch lanes and stay outside the required merge checks. Co-authored-by: majinghe <majinghe@users.noreply.github.com>	2026-11-02 23:03:36 +08:00
Hauser Zhi	9d66e04ae3	test(e2e): honor namespace backpressure in mixed workload (#9122 ) Give each mixed-workload request a bounded policy for the server's explicit namespace lock timeout response and Retry-After advice. Keep unrelated SDK failures and data mismatches immediately visible. Preserve 24 workers, four rounds, cross-node checks and all retained objects. Cover exhausted SDK retries, bodyless HEAD, fatal responses, persistent contention and incorrect acknowledged bytes, and include the new cases in the protected distributed profile inventory. Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-11-02 23:02:22 +08:00
dependabot[bot]	12c58d25b5	chore(deps): bump the dependencies group across 1 directory with 5 updates (#9113 ) * chore(deps): bump the dependencies group across 1 directory with 5 updates Bumps the dependencies group with 5 updates in the / directory: \| Package \| From \| To \| \| --- \| --- \| --- \| \| [pulsar](https://github.com/streamnative/pulsar-rs) \| `6.9.0` \| `6.9.1` \| \| [arrow-array](https://github.com/apache/arrow-rs) \| `59.3.2` \| `60.1.0` \| \| [arrow-ipc](https://github.com/apache/arrow-rs) \| `59.3.2` \| `60.1.0` \| \| [arrow-schema](https://github.com/apache/arrow-rs) \| `59.3.2` \| `60.1.0` \| \| [parquet](https://github.com/apache/arrow-rs) \| `59.3.2` \| `60.1.0` \| Updates `pulsar` from 6.9.0 to 6.9.1 - [Release notes](https://github.com/streamnative/pulsar-rs/releases) - [Commits](https://github.com/streamnative/pulsar-rs/compare/v6.9.0...v6.9.1) Updates `arrow-array` from 59.3.2 to 60.1.0 - [Release notes](https://github.com/apache/arrow-rs/releases) - [Changelog](https://github.com/apache/arrow-rs/blob/main/CHANGELOG.md) - [Commits](https://github.com/apache/arrow-rs/compare/59.3.2...60.1.0) Updates `arrow-ipc` from 59.3.2 to 60.1.0 - [Release notes](https://github.com/apache/arrow-rs/releases) - [Changelog](https://github.com/apache/arrow-rs/blob/main/CHANGELOG.md) - [Commits](https://github.com/apache/arrow-rs/compare/59.3.2...60.1.0) Updates `arrow-schema` from 59.3.2 to 60.1.0 - [Release notes](https://github.com/apache/arrow-rs/releases) - [Changelog](https://github.com/apache/arrow-rs/blob/main/CHANGELOG.md) - [Commits](https://github.com/apache/arrow-rs/compare/59.3.2...60.1.0) Updates `parquet` from 59.3.2 to 60.1.0 - [Release notes](https://github.com/apache/arrow-rs/releases) - [Changelog](https://github.com/apache/arrow-rs/blob/main/CHANGELOG.md) - [Commits](https://github.com/apache/arrow-rs/compare/59.3.2...60.1.0) --- updated-dependencies: - dependency-name: arrow-array dependency-version: 60.1.0 dependency-type: direct:production update-type: version-update:semver-major dependency-group: dependencies - dependency-name: arrow-ipc dependency-version: 60.1.0 dependency-type: direct:production update-type: version-update:semver-major dependency-group: dependencies - dependency-name: arrow-schema dependency-version: 60.1.0 dependency-type: direct:production update-type: version-update:semver-major dependency-group: dependencies - dependency-name: parquet dependency-version: 60.1.0 dependency-type: direct:production update-type: version-update:semver-major dependency-group: dependencies - dependency-name: pulsar dependency-version: 6.9.1 dependency-type: direct:production update-type: version-update:semver-patch dependency-group: dependencies ... Signed-off-by: dependabot[bot] <support@github.com> * chore(deps): update dependencies and pin pulsar 6.9.0 Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * chore(deps): ignore pulsar 6.9 updates Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Hauser <housemecn@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-11-02 23:01:55 +08:00
Ramakrishna Chilaka	1ac1d883f4	perf(ecstore): shard disk path lock registries with SWAR byte keys (#9127 )	2026-11-02 23:00:48 +08:00
GatewayJ	45ba33291a	fix(ecstore): release retained file descriptor cache invalidations (#9112 ) * fix(ecstore): release retained descriptor cache invalidations * test(ecstore): await descriptor cleanup after invalidation	2026-11-02 10:56:08 +00:00
hector	636eded277	ci(validation): resolve the newest nightly build on scheduled sweeps (#9114 ) * ci(validation): resolve the newest nightly build on scheduled sweeps The first scheduled sweep failed at candidate resolution: a schedule carries no workflow inputs and resolve_functional_candidate.py only derives a build from a workflow_run payload, so the empty BUILD_RUN_ID crashed int(). Look up the newest successful nightly-gnu run explicitly when the inputs are empty and feed it to the resolver. Also downgrade the evidence artifact to if-no-files-found: warn so an early infra failure does not cascade into a second red step. * ci(validation): read the resolver chain envelope correctly CHAIN_OUTPUT is the chain envelope {schema, run_id, attempt, workflow_sha, testing_sha, candidate}; the manifest lives under candidate.manifest and the testing sha at the top level. * ci(validation): sweep harness fixes from the first live run - ssh -n in the inter-case cleanup loop: plain ssh consumed the case list from the loop's stdin, so the first dispatch executed only one of the twelve selected cases. - Provide the managed Python 3.12 runtime for pool/ilm/replication cases, mirroring the chain validation job (pool.status reported INFRA_ERROR after its assertions passed; the missing runtime is the prime suspect). - Record the auto-testing sha in the report step environment. - Retain the last case's evidence tree in the sweep artifact so gate verdicts like INFRA_ERROR can be diagnosed from the run itself. * ci(validation): install sweep Python client dependencies Scheduled sweeps failed pool.status before any assertion because pool_behavior.py exits when boto3 is missing. Install the auto-testing validation requirements into a per-run virtualenv and put it on PATH, matching the functional chain, so case scripts can import the client. Co-authored-by: hector <majinghe@users.noreply.github.com> --------- Co-authored-by: hector <majinghe@users.noreply.github.com>	2026-11-02 08:34:13 +00:00
Hauser Zhi	ed4099ebe1	chore(deps): update crates and fix storage CI regressions (#9111 ) * chore(deps): refresh crates and adopt openidconnect 4.1 Use published OIDC/OAuth2 releases and reject malformed HMAC signature lengths before upstream verification. Refresh workspace and fuzz locks, and advance the local zstd compatibility patch to 0.13.4 while retaining the fixed 0.14.3 codec backend. Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * chore(deps): refresh compatible lockfile resolutions Resolve Prost build and derive dependencies with itertools 0.14 and prefer windows-sys 0.61.2 within the existing supported version ranges. Retain all package versions, sources and checksums. Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * test(connect): separate cleanup checks from short probe windows Use the functional measurement budget for the cleanup HTTP fixture so setup and measured I/O can finish before checking delete status handling. Exercise both normal scheduling and an in-flight PUT clock advance, and include target classification in unexpected-failure diagnostics. Retain the cleanup deadline, outer timeout and delete-stage assertions. Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * test(ecstore): isolate stalled multipart disk timeouts Apply the short injected timeout only to the deliberately stalled disk. Healthy disks retain their normal operation budgets, while logical lock, physical owner, immediate CopyObject and final shard assertions remain. Keep overrides scoped across cloned and reconnected handles, preserve shorter caller deadlines and explicit zero timeouts, and report drive health when the real CopyObject assertion fails. Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> * fix(rebalance): retry metadata write lock quorum admission Recognize typed write-lock quorum failures for rebalance.bin within the existing bounded metadata admission retry policy. Each attempt must acquire quorum before loading or saving metadata; data writes are not replayed by this retry. Cover recovered and persistent quorum failures, cancellation after a failed attempt, and rejection of unrelated resource or read-lock errors. Co-Authored-By: heihutu <heihutu@gmail.com> Co-Authored-By: zhi22915 <qiuzgang@gmail.com> --------- Co-authored-by: heihutu <heihutu@gmail.com> Co-authored-by: zhi22915 <qiuzgang@gmail.com>	2026-11-02 06:49:24 +00:00