Merge branch 'main' into test/distributed-e2e-hardening

.config/e2e-full-selection.txt

@@ -1,2 +1,2 @@
 sha256-darwin=a881fd7d3f5cb94654221ca85b8b30cce1b95e608824a55a15339cbc294e6d34
-sha256-linux=e9a8d64e73f627c4d26c236dbbba690c9ee03a9e26d42a4244515b4439365535
+sha256-linux=a2933d83dfe74ffa03410a0959333a1c48288b8469ca9f17273d449d7510c24b

.config/ecstore-required-tests.json

@@ -0,0 +1,72 @@
+{
+  "lane": "ci/test-and-lint",
+  "tests": [
+    {
+      "invariant": "write-quorum",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::ops::object::inline_put_commit_path_tests::inline_put_direct_commit_accepts_exact_quorum_and_rejects_quorum_minus_one"
+    },
+    {
+      "invariant": "metadata-rollback",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::core::io_primitives::tests::write_unique_file_info_reverts_metadata_when_write_quorum_fails"
+    },
+    {
+      "invariant": "stale-writer",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::ops::object::put_object_tmp_cleanup_tests::put_object_no_lock_aborts_after_outer_namespace_lock_loss"
+    },
+    {
+      "invariant": "range-body",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::ops::object::transition_upload_integrity_tests::transitioned_compressed_object_range_get_returns_plaintext_slice"
+    },
+    {
+      "invariant": "multipart-cancellation",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::ops::multipart::tests::cancelled_complete_keeps_upload_lock_through_tail_cleanup"
+    },
+    {
+      "invariant": "list-uncommitted-version",
+      "suite": "rustfs-filemeta",
+      "name": "metacache::tests::resolve_with_write_quorum_slack_keeps_partial_latest_hidden_during_merge"
+    },
+    {
+      "invariant": "minio-object-fixture",
+      "suite": "rustfs-filemeta",
+      "name": "filemeta::test::parses_real_minio_object_xlmeta"
+    },
+    {
+      "invariant": "corrupt-part-arrays",
+      "suite": "rustfs-filemeta",
+      "name": "filemeta::test::crc_valid_but_part_arrays_corrupt_into_fileinfo_errors_not_panics"
+    }
+  ],
+  "fixtures": [
+    {
+      "path": "crates/filemeta/tests/fixtures/minio/object_large_bin.xlmeta.hex",
+      "sha256": "e8093767806d701e639b48d023190e858fbc4cde69bcfd83c22af8cba8452ce5",
+      "source": "MinIO RELEASE.2025-07-23T15-54-02Z; crates/ecstore/tests/fixtures/minio/README.md"
+    },
+    {
+      "path": "crates/filemeta/tests/fixtures/minio/object_small_txt.xlmeta.hex",
+      "sha256": "2a415ad3a3be5a9440035d4026ff880e0e8c1ec1701be9f4e077734e8dce03da",
+      "source": "MinIO RELEASE.2025-07-23T15-54-02Z; crates/ecstore/tests/fixtures/minio/README.md"
+    },
+    {
+      "path": "crates/filemeta/tests/fixtures/minio/object_versioned_txt.xlmeta.hex",
+      "sha256": "7f21f50c326dd8b0228deb6dbdb7052b3d0a3f8ee6c85d43486f0e6bb7a97261",
+      "source": "MinIO RELEASE.2025-07-23T15-54-02Z; crates/ecstore/tests/fixtures/minio/README.md"
+    },
+    {
+      "path": "crates/ecstore/tests/fixtures/minio/bucket_metadata.blob.hex",
+      "sha256": "f2b6e260aff106adf6039feb1c645686e84e75404ff725491fb18668be5db203",
+      "source": "MinIO RELEASE.2025-07-23T15-54-02Z; crates/ecstore/tests/fixtures/minio/README.md"
+    },
+    {
+      "path": "crates/ecstore/tests/fixtures/minio/bucket_metadata_full.xlmeta.hex",
+      "sha256": "3b6de589519c08a1614c8bd409bb8199c17d42043861b07bce513075e6fbfc12",
+      "source": "MinIO RELEASE.2025-07-23T15-54-02Z; crates/ecstore/tests/fixtures/minio/README.md"
+    }
+  ]
+}

.config/make/pre-commit.mak

@@ -3,9 +3,10 @@
 .NOTPARALLEL: pre-commit pre-pr dev-check
 
 .PHONY: setup-hooks
-setup-hooks: ## Set up git hooks
+setup-hooks: ## Install the configured pre-commit hooks
 	@echo "🔧 Setting up git hooks..."
-	chmod +x .git/hooks/pre-commit
+	pre-commit validate-config
+	pre-commit install
 	@echo "✅ Git hooks setup complete!"
 
 .PHONY: doc-paths-check

.config/make/tests.mak

@@ -40,6 +40,7 @@ script-tests: ## Run shell script tests
 	$(RUSTFS_PYTHON_BIN) ./scripts/check_test_wiring.py --self-test
 	$(RUSTFS_PYTHON_BIN) ./scripts/check_security_coverage.py --self-test
 	$(RUSTFS_PYTHON_BIN) ./scripts/check_scheduled_validation_freshness.py --self-test
+	$(RUSTFS_PYTHON_BIN) ./scripts/test_security_workflow.py
 	$(RUSTFS_PYTHON_BIN) ./scripts/s3-tests/test_report_compat.py
 	bash -n ./scripts/validate_object_data_cache_cold_stampede.sh
 	$(RUSTFS_PYTHON_BIN) ./scripts/check_object_data_cache_follower_samples.py --self-test

.github/actions/quick-checks/action.yml

@@ -0,0 +1,115 @@
+# Copyright 2024 RustFS Team
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: Quick Checks
+description: Run the shared compile-free RustFS quality checks.
+
+runs:
+  using: composite
+  steps:
+    - name: Install quality tools
+      uses: taiki-e/install-action@bffeee26d4db9be238a4ea78d8826604ebcb594d # v2
+      with:
+        tool: |
+          ripgrep@15.2.0
+          shellcheck@0.11.0
+
+    - name: Install actionlint
+      shell: bash
+      run: |
+        actionlint_dir="$(mktemp -d "${RUNNER_TEMP}/actionlint.XXXXXX")"
+        curl --fail --location --silent --show-error \
+          --output "$actionlint_dir/actionlint.tar.gz" \
+          https://github.com/rhysd/actionlint/releases/download/v1.7.12/actionlint_1.7.12_linux_amd64.tar.gz
+        echo "8aca8db96f1b94770f1b0d72b6dddcb1ebb8123cb3712530b08cc387b349a3d8  $actionlint_dir/actionlint.tar.gz" | sha256sum --check --status
+        tar -xzf "$actionlint_dir/actionlint.tar.gz" -C "$actionlint_dir" actionlint
+        rm "$actionlint_dir/actionlint.tar.gz"
+        echo "$actionlint_dir" >> "$GITHUB_PATH"
+
+    - name: Install Rust toolchain
+      uses: dtolnay/rust-toolchain@29eef336d9b2848a0b548edc03f92a220660cdb8 # stable
+      with:
+        components: rustfmt
+
+    - name: Check workflow syntax and shell scripts
+      shell: bash
+      run: shellcheck --version && actionlint
+
+    - name: Check code formatting
+      shell: bash
+      run: cargo fmt --all --check
+
+    - name: Check unsafe code allowances
+      shell: bash
+      run: ./scripts/check_unsafe_code_allowances.sh
+
+    - name: Check layered dependencies
+      shell: bash
+      run: ./scripts/check_layer_dependencies.sh
+
+    - name: Check architecture migration rules
+      shell: bash
+      run: ./scripts/check_architecture_migration_rules.sh
+
+    - name: Check logging guardrails
+      shell: bash
+      run: ./scripts/check_logging_guardrails.sh
+
+    - name: Check error other(format!) ratchet
+      shell: bash
+      run: ./scripts/check_error_other_format_ratchet.sh
+
+    - name: Check tokio io-uring feature guard
+      shell: bash
+      run: ./scripts/check_no_tokio_io_uring.sh
+
+    - name: Check extension schema boundaries
+      shell: bash
+      run: ./scripts/check_extension_schema_boundaries.sh
+
+    - name: Check body-cache whitelist guard
+      shell: bash
+      run: ./scripts/check_body_cache_whitelist.sh
+
+    - name: Check s3s footprint ratchet
+      shell: bash
+      run: ./scripts/check_s3s_footprint.sh
+
+    - name: Check cryptographic capability wording
+      shell: bash
+      run: ./scripts/check_fips_wording.sh
+
+    - name: Check no embedded secret material
+      shell: bash
+      run: ./scripts/check_embedded_secrets.sh
+
+    - name: Check test wiring
+      shell: bash
+      run: |
+        python3 ./scripts/check_test_wiring.py --self-test
+        python3 ./scripts/check_scheduled_validation_freshness.py --self-test
+        python3 ./scripts/test_security_workflow.py
+        python3 ./scripts/check_test_wiring.py
+
+    - name: Check no planning docs committed
+      shell: bash
+      run: ./scripts/check_no_planning_docs.sh
+
+    - name: Check CI paths stay in sync
+      shell: bash
+      run: ./scripts/check_ci_paths_sync.sh
+
+    - name: Check io_uring lane --lib precondition
+      shell: bash
+      run: ./scripts/check_uring_lane_lib_only.sh

.github/pull_request_template.md

@@ -10,16 +10,16 @@ Use N/A when there is no related issue.
 
 ## Summary of Changes
 <!--
-Briefly explain what changed and why reviewers should accept it.
+Describe the concrete problem and resulting behavior. For a behavior change, name the input or state that triggers it and the expected outcome. Explain any new dependency or abstraction that the change needs.
-Focus on behavior, compatibility, and review-relevant context.
 -->
 
 ## Verification
 <!--
-List the commands or checks you ran, for example:
+Give 1–3 concrete pieces of evidence for the changed behavior: the test or command, its observed result, and the regression it catches. For a bug fix, record a failing-before/passing-after check or explain why it was unavailable.
-- `make pre-commit`
 
-Use N/A only when verification is not applicable.
+Identify the tested commit and any local changes. When testing a prebuilt binary or external service, include its source/version and artifact identity; a successful run against a different build is not evidence for this change.
+
+List relevant checks not run and the remaining risk. Use the validation tier in AGENTS.md; do not run broader checks solely to fill this section. For documentation-only changes, list the applicable documentation checks. Use N/A only when verification is not applicable.
 -->
 
 ## Impact

.github/workflows/ci-docs-only.yml

@@ -12,24 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-# Companion to ci.yml for required status checks.
+# Reports the existing required checks for paths excluded by ci.yml.
-#
+# Mixed PRs can trigger both workflows; their Quick Checks jobs use one shared
-# ci.yml skips docs-only pull requests via paths-ignore, but the branch ruleset
+# action to keep validation coverage aligned. Keep this paths list in sync with
-# requires a check named "Test and Lint" — without this workflow a docs-only PR
+# ci.yml's pull_request.paths-ignore via scripts/check_ci_paths_sync.sh.
-# would wait on it forever. This workflow triggers on exactly the paths ci.yml
-# ignores and reports success under the same job name. Mixed PRs trigger both
-# workflows and the real check still gates: a required check with any failing
-# run blocks the merge.
-# https://docs.github.com/en/repositories/configuring-branches-and-merges-in-your-repository/defining-the-mergeability-of-pull-requests/troubleshooting-required-status-checks#handling-skipped-but-required-checks
-#
-# "Quick Checks" is mirrored here ahead of the ruleset change that will make it
-# required too (rustfs/backlog#1599). Until that change lands this job is
-# inert; mirroring it first is what lets the ruleset change happen without
-# stranding docs-only PRs on a check nobody reports.
-#
-# Keep the paths list below in sync with the pull_request paths-ignore list
-# in ci.yml, and keep the quick-checks steps below byte-identical to the
-# quick-checks job in ci.yml.
 
 name: Continuous Integration (docs only)
 
@@ -59,19 +45,6 @@ permissions:
   contents: read
 
 jobs:
-  # Deliberately NOT a bare `echo`. Once "Quick Checks" becomes a required
-  # check, ci.yml gates every expensive job behind it, so a mixed PR reports
-  # two check runs with this name: the real one (45-51s) and this companion.
-  # GitHub has no written contract for how it picks between same-named
-  # required check runs ("latest wins" vs "any failure blocks"), so instead of
-  # relying on ordering we make both runs execute the same commands against
-  # the same merge ref — their conclusions are then necessarily identical and
-  # the choice does not matter. Keep these steps byte-identical to the
-  # quick-checks job in ci.yml (a guard script that asserts this, and the paths
-  # sync below, is tracked in rustfs/backlog#1603).
-  #
-  # For a genuinely docs-only PR this adds no strictness (no code changed, so
-  # fmt and the guards always pass) and costs ~50s of ubuntu-latest.
   quick-checks:
     name: Quick Checks
     runs-on: ubuntu-latest
@@ -82,63 +55,8 @@ jobs:
         with:
           persist-credentials: false
 
-      - name: Install ripgrep
+      - name: Run shared quick checks
-        uses: taiki-e/install-action@bffeee26d4db9be238a4ea78d8826604ebcb594d # v2
+        uses: ./.github/actions/quick-checks
-        with:
-          tool: ripgrep@15.2.0
-
-      - name: Install Rust toolchain
-        uses: dtolnay/rust-toolchain@29eef336d9b2848a0b548edc03f92a220660cdb8 # stable
-        with:
-          components: rustfmt
-
-      - name: Check code formatting
-        run: cargo fmt --all --check
-
-      - name: Check unsafe code allowances
-        run: ./scripts/check_unsafe_code_allowances.sh
-
-      - name: Check layered dependencies
-        run: ./scripts/check_layer_dependencies.sh
-
-      - name: Check architecture migration rules
-        run: ./scripts/check_architecture_migration_rules.sh
-
-      - name: Check logging guardrails
-        run: ./scripts/check_logging_guardrails.sh
-
-      - name: Check tokio io-uring feature guard
-        run: ./scripts/check_no_tokio_io_uring.sh
-
-      - name: Check extension schema boundaries
-        run: ./scripts/check_extension_schema_boundaries.sh
-
-      - name: Check body-cache whitelist guard
-        run: ./scripts/check_body_cache_whitelist.sh
-
-      - name: Check s3s footprint ratchet
-        run: ./scripts/check_s3s_footprint.sh
-
-      - name: Check cryptographic capability wording
-        run: ./scripts/check_fips_wording.sh
-
-      - name: Check no embedded secret material
-        run: ./scripts/check_embedded_secrets.sh
-
-      - name: Check test wiring
-        run: |
-          python3 ./scripts/check_test_wiring.py --self-test
-          python3 ./scripts/check_scheduled_validation_freshness.py --self-test
-          python3 ./scripts/check_test_wiring.py
-
-      - name: Check no planning docs committed
-        run: ./scripts/check_no_planning_docs.sh
-
-      - name: Check CI paths stay in sync
-        run: ./scripts/check_ci_paths_sync.sh
-
-      - name: Check io_uring lane --lib precondition
-        run: ./scripts/check_uring_lane_lib_only.sh
 
   test-and-lint:
     name: Test and Lint

.github/workflows/ci.yml

@@ -100,12 +100,7 @@ jobs:
       - name: Typos check with custom config file
         uses: crate-ci/typos@37bb98842b0d8c4ffebdb75301a13db0267cef89 # master
 
-  # Fast, compile-free checks that fail early so contributors get feedback in
+  # Fail early with compile-free checks shared with docs-only CI.
-  # ~1 minute instead of waiting for the full test job.
-  #
-  # These steps are mirrored byte-for-byte in ci-docs-only.yml so that a mixed
-  # PR, which reports two check runs named "Quick Checks", cannot get one red
-  # and one green. Edit both jobs together.
   quick-checks:
     name: Quick Checks
     if: github.event_name != 'pull_request' || github.event.action != 'closed'
@@ -117,66 +112,8 @@ jobs:
         with:
           persist-credentials: false
 
-      - name: Install ripgrep
+      - name: Run shared quick checks
-        uses: taiki-e/install-action@bffeee26d4db9be238a4ea78d8826604ebcb594d # v2
+        uses: ./.github/actions/quick-checks
-        with:
-          tool: ripgrep@15.2.0
-
-      - name: Install Rust toolchain
-        uses: dtolnay/rust-toolchain@29eef336d9b2848a0b548edc03f92a220660cdb8 # stable
-        with:
-          components: rustfmt
-
-      - name: Check code formatting
-        run: cargo fmt --all --check
-
-      - name: Check unsafe code allowances
-        run: ./scripts/check_unsafe_code_allowances.sh
-
-      - name: Check layered dependencies
-        run: ./scripts/check_layer_dependencies.sh
-
-      - name: Check architecture migration rules
-        run: ./scripts/check_architecture_migration_rules.sh
-
-      - name: Check logging guardrails
-        run: ./scripts/check_logging_guardrails.sh
-
-      - name: Check error other(format!) ratchet
-        run: ./scripts/check_error_other_format_ratchet.sh
-
-      - name: Check tokio io-uring feature guard
-        run: ./scripts/check_no_tokio_io_uring.sh
-
-      - name: Check extension schema boundaries
-        run: ./scripts/check_extension_schema_boundaries.sh
-
-      - name: Check body-cache whitelist guard
-        run: ./scripts/check_body_cache_whitelist.sh
-
-      - name: Check s3s footprint ratchet
-        run: ./scripts/check_s3s_footprint.sh
-
-      - name: Check cryptographic capability wording
-        run: ./scripts/check_fips_wording.sh
-
-      - name: Check no embedded secret material
-        run: ./scripts/check_embedded_secrets.sh
-
-      - name: Check test wiring
-        run: |
-          python3 ./scripts/check_test_wiring.py --self-test
-          python3 ./scripts/check_scheduled_validation_freshness.py --self-test
-          python3 ./scripts/check_test_wiring.py
-
-      - name: Check no planning docs committed
-        run: ./scripts/check_no_planning_docs.sh
-
-      - name: Check CI paths stay in sync
-        run: ./scripts/check_ci_paths_sync.sh
-
-      - name: Check io_uring lane --lib precondition
-        run: ./scripts/check_uring_lane_lib_only.sh
 
   test-and-lint:
     name: Test and Lint
@@ -269,6 +206,7 @@ jobs:
           CARGO_BUILD_JOBS: ${{ (github.event_name == 'push' || github.event_name == 'workflow_dispatch') && '3' || '2' }}
         run: |
           mkdir -p artifacts/test-and-lint
+          rm -f target/nextest/ci/junit.xml
           ./scripts/ci/resource_sampler.sh start nextest
           trap './scripts/ci/resource_sampler.sh stop' EXIT
           set +e
@@ -277,6 +215,12 @@ jobs:
             --status-level all --final-status-level all \
             2>&1 | tee artifacts/test-and-lint/nextest.log
           status=${PIPESTATUS[0]}
+          if [[ "${status}" -eq 0 ]]; then
+            cargo nextest list --profile ci --all --exclude e2e_test --message-format json \
+              > artifacts/test-and-lint/core-test-listing.json \
+              && python3 scripts/check_test_wiring.py --check-core artifacts/test-and-lint/core-test-listing.json \
+              && test -s target/nextest/ci/junit.xml || status=$?
+          fi
           {
             echo "command=cargo nextest run --profile ci --all --exclude e2e_test"
             echo "exit_status=${status}"

.github/workflows/rustfs-heal-test.yml

@@ -54,9 +54,6 @@ env:
 jobs:
   heal-test:
     runs-on: smoke-testing
-    # Requirement: a failing suite must not fail the workflow; failures
-    # are filed to rustfs/backlog and the chain continues.
-    continue-on-error: true
     timeout-minutes: 480
     # Standalone manual run, or one link of the nightly functional chain
     # (storage -> heal -> pool). Pool expansion no longer re-runs heal.

.github/workflows/rustfs-kms-test.yml

@@ -49,7 +49,6 @@ env:
 jobs:
   kms-test:
     runs-on: smoke-testing
-    continue-on-error: true
     timeout-minutes: 420
     if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'repository_dispatch' }}
     steps:
@@ -109,7 +108,6 @@ jobs:
 
       - name: Run KMS suite
         id: test
-        continue-on-error: true
         env:
           LOG_FILE: /tmp/rustfs-kms.log
         run: |

.github/workflows/rustfs-performance-test.yml

@@ -84,9 +84,6 @@ env:
 jobs:
   performance-test:
     runs-on: pf-testing
-    # Requirement: a failing benchmark must not fail the workflow;
-    # failures are filed to rustfs/backlog.
-    continue-on-error: true
     timeout-minutes: 900
     # Run on manual dispatch, or when the nightly build completed successfully.
     # Skipped when nightly failed.

.github/workflows/rustfs-pool-expand-test.yml

@@ -76,9 +76,6 @@ jobs:
   pool-expansion-test:
     name: Pool expansion / decommission test
     runs-on: smoke-testing
-    # Requirement: a failing suite must not fail the workflow; failures
-    # are filed to rustfs/backlog and the chain continues.
-    continue-on-error: true
     timeout-minutes: 360
     if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'repository_dispatch' }}
     env:

.github/workflows/rustfs-replication-test.yml

@@ -62,9 +62,6 @@ env:
 jobs:
   replication-test:
     runs-on: smoke-testing
-    # A failed replication run must not break the chain or the workflow: the
-    # failure is reported to rustfs/backlog instead (see the issue step).
-    continue-on-error: true
     timeout-minutes: 360
     if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'repository_dispatch' }}
     steps:
@@ -116,7 +113,6 @@ jobs:
 
       - name: Run replication suite
         id: test
-        continue-on-error: true
         env:
           LOG_FILE: /tmp/rustfs-replication.log
         run: |

.github/workflows/rustfs-s3-compat-test.yml

@@ -37,7 +37,6 @@ env:
 jobs:
   s3-compat-test:
     runs-on: smoke-testing
-    continue-on-error: true
     timeout-minutes: 360
     if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'repository_dispatch' }}
     steps:
@@ -88,7 +87,6 @@ jobs:
 
       - name: Run S3 compatibility suite
         id: test
-        continue-on-error: true
         env:
           LOG_FILE: /tmp/rustfs-s3-compat.log
         run: |

.github/workflows/rustfs-security-test.yml

@@ -74,10 +74,23 @@ env:
 jobs:
   security-test:
     runs-on: smoke-testing
-    continue-on-error: true
     timeout-minutes: 360
     if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'repository_dispatch' }}
     steps:
+      - name: Checkout repository (for the OIDC live gate script)
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+
+      - name: Initialize security evidence
+        id: evidence
+        run: |
+          set -euo pipefail
+          umask 077
+          SECURITY_ARTIFACTS_DIR="${RUNNER_TEMP}/rustfs-security-${GITHUB_RUN_ID}-${GITHUB_RUN_ATTEMPT}"
+          mkdir -- "${SECURITY_ARTIFACTS_DIR}"
+          printf 'SECURITY_ARTIFACTS_DIR=%s\n' "${SECURITY_ARTIFACTS_DIR}" >> "${GITHUB_ENV}"
+
       # auto-testing is private: clone it with the dedicated PF token (not
       # GITHUB_TOKEN) and retry transient GitHub/network failures.
       - name: Checkout auto-testing scripts (with retry)
@@ -98,11 +111,6 @@ jobs:
           echo "ERROR: unable to clone rustfs/auto-testing after 5 attempts" >&2
           exit 1
 
-      - name: Checkout repository (for the OIDC live gate script)
-        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
-        with:
-          persist-credentials: false
-
       - name: Show environment
         run: |
           uname -a
@@ -135,7 +143,8 @@ jobs:
         id: test
         continue-on-error: true
         env:
-          REPORT_FILE: /tmp/rustfs-security-report.md
+          REPORT_FILE: ${{ env.SECURITY_ARTIFACTS_DIR }}/suite-report.md
+          TMPDIR: ${{ env.SECURITY_ARTIFACTS_DIR }}
           RUSTFS_SECURITY_OIDC_LIVE_SCRIPT: ${{ github.workspace }}/scripts/test/oidc_keycloak_live.sh
         run: |
           set -euo pipefail
@@ -159,29 +168,48 @@ jobs:
           else
             ARGS+=(--package-url "${RUSTFS_NIGHTLY_PACKAGE_URL}")
           fi
-          ./auto-testing/rustfs-security-test.sh "${ARGS[@]}"
+          GITHUB_STEP_SUMMARY=/dev/null ./auto-testing/rustfs-security-test.sh "${ARGS[@]}"
 
       - name: Generate report
-        if: always()
+        id: report
+        if: ${{ always() && steps.evidence.outcome == 'success' }}
+        env:
+          TEST_OUTCOME: ${{ steps.test.outcome }}
         run: |
           set -euo pipefail
-          if [ ! -f /tmp/rustfs-security-report.md ]; then
+          RESULT=failure
-            {
+          if [ "${TEST_OUTCOME}" = "success" ] && [ -s "${SECURITY_ARTIFACTS_DIR}/suite-report.md" ]; then
-              echo "# RustFS security test report"
+            RESULT=success
-              echo ""
-              echo "- Run: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}"
-              echo "- Trigger: ${{ github.event_name }}"
-              echo "- Test Step Outcome: failure (suite did not produce a report)"
-            } > /tmp/rustfs-security-report.md
           fi
-          cat /tmp/rustfs-security-report.md >> "${GITHUB_STEP_SUMMARY}"
+          {
+            echo "# RustFS security test report"
+            echo ""
+            echo "- Run: ${GITHUB_SERVER_URL}/${GITHUB_REPOSITORY}/actions/runs/${GITHUB_RUN_ID}"
+            echo "- Attempt: ${GITHUB_RUN_ATTEMPT}"
+            echo "- Workflow Commit: ${GITHUB_SHA}"
+            echo "- Trigger: ${GITHUB_EVENT_NAME}"
+            echo "- Test Step Outcome: ${RESULT}"
+            echo "- Suite Step Outcome: ${TEST_OUTCOME}"
+            echo ""
+            # The dashboard prioritizes case rows over the step outcome.
+            # Keep partial case results in the artifact when the suite fails.
+            if [ "${RESULT}" = "success" ]; then
+              cat "${SECURITY_ARTIFACTS_DIR}/suite-report.md"
+            elif [ -s "${SECURITY_ARTIFACTS_DIR}/suite-report.md" ]; then
+              echo "The suite did not complete successfully. See suite-report.md in this run's artifact for diagnostics."
+            else
+              echo "The suite did not produce a non-empty report."
+            fi
+          } > "${SECURITY_ARTIFACTS_DIR}/report.md"
+          cat "${SECURITY_ARTIFACTS_DIR}/report.md" >> "${GITHUB_STEP_SUMMARY}"
+          [ "${RESULT}" = "success" ]
 
       - name: Upload functional report to dashboard
-        if: always()
+        if: ${{ always() && steps.evidence.outcome == 'success' }}
         continue-on-error: true
         env:
           GH_TOKEN: ${{ env.PF_TESTING_GH_TOKEN }}
-          REPORT_FILE: /tmp/rustfs-security-report.md
+          REPORT_FILE: ${{ env.SECURITY_ARTIFACTS_DIR }}/report.md
           SUITE: security
         run: |
           set -euo pipefail
@@ -210,8 +238,9 @@ jobs:
           GH_TOKEN: ${{ secrets.PF_TESTING_GH_TOKEN }}
           SUITE: 'security'
           SUITE_LABEL: 'Security'
+          EVIDENCE_OUTCOME: ${{ steps.evidence.outcome }}
           RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
-          REPORT_FILE: '/tmp/rustfs-security-report.md'
+          REPORT_FILE: ${{ env.SECURITY_ARTIFACTS_DIR }}/report.md
           LOG_FILE: ''
         run: |
           set -euo pipefail
@@ -245,7 +274,7 @@ jobs:
             echo ""
             echo "## Report (errors and symptoms)"
             echo ""
-            if [ -s "${REPORT_FILE}" ]; then
+            if [ "${EVIDENCE_OUTCOME}" = "success" ] && [ -s "${REPORT_FILE}" ]; then
               redact < "${REPORT_FILE}"
             elif [ -s "${LOG_FILE:-}" ]; then
               echo "(report file missing; log tail below)"
@@ -263,14 +292,12 @@ jobs:
           echo "filed backlog issue for suite ${SUITE}"
 
       - name: Upload report and logs
-        if: always()
+        if: ${{ always() && steps.evidence.outcome == 'success' }}
         uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
         with:
-          name: rustfs-security-test-${{ github.run_id }}
+          name: rustfs-security-test-${{ github.run_id }}-${{ github.run_attempt }}
-          path: |
+          path: ${{ env.SECURITY_ARTIFACTS_DIR }}/
-            /tmp/rustfs-security-report.md
+          if-no-files-found: error
-            /tmp/rustfs-security.*/*
-          if-no-files-found: ignore
           retention-days: 3
 
       - name: Cleanup environment (after)

.github/workflows/rustfs-storage-test.yml

@@ -46,7 +46,6 @@ env:
 jobs:
   storage-test:
     runs-on: smoke-testing
-    continue-on-error: true
     timeout-minutes: 360
     if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'repository_dispatch' }}
     steps:
@@ -97,7 +96,6 @@ jobs:
 
       - name: Run storage engine suite
         id: test
-        continue-on-error: true
         env:
           LOG_FILE: /tmp/rustfs-storage.log
         run: |

.github/workflows/rustfs-tier-test.yml

@@ -61,9 +61,6 @@ env:
 jobs:
   tier-test:
     runs-on: smoke-testing
-    # Requirement: a failing suite must not fail the workflow; failures
-    # are filed to rustfs/backlog and the chain continues.
-    continue-on-error: true
     timeout-minutes: 420
     if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'repository_dispatch' }}
     steps:

.github/workflows/rustfs-upgrade-test.yml

@@ -18,7 +18,7 @@ on:
   workflow_dispatch:
     inputs:
       from_version:
-        description: 'OLD RustFS release tag (must ship a .deb asset, e.g. 1.0.0-rc.3)'
+        description: 'OLD RustFS release tag, e.g. 1.0.0-rc.3 (its release must ship a .deb asset). Leave empty for the default.'
         required: false
         default: '1.0.0-rc.3'
       from_url:
@@ -26,7 +26,7 @@ on:
         required: false
         type: string
       to_version:
-        description: 'NEW RustFS release tag (leave empty for latest nightly)'
+        description: 'NEW RustFS release tag, e.g. 1.0.0-rc.5 (any version with a .deb asset). Leave empty for latest nightly.'
         required: false
       to_url:
         description: 'NEW .deb URL. Overrides to_version / nightly default.'
@@ -79,7 +79,6 @@ env:
 jobs:
   upgrade-test:
     runs-on: smoke-testing
-    continue-on-error: true
     timeout-minutes: 420
     if: ${{ github.event_name == 'workflow_dispatch' || github.event_name == 'repository_dispatch' }}
     steps:
@@ -142,9 +141,9 @@ jobs:
 
       - name: Run upgrade compatibility suite
         id: test
-        continue-on-error: true
         env:
           LOG_FILE: /tmp/rustfs-upgrade.log
+          GH_TOKEN: ${{ secrets.PF_TESTING_GH_TOKEN }}
         run: |
           set -euo pipefail
           chmod +x auto-testing/rustfs-upgrade-test.sh
@@ -175,6 +174,29 @@ jobs:
           else
             ARGS+=(--to-url "${RUSTFS_NIGHTLY_PACKAGE_URL}")
           fi
+          # Fail fast with a clear message when a requested release tag has
+          # no .deb asset (e.g. 1.0.0-rc.4 ships only zips), instead of
+          # letting the suite die mid-run on a 404.
+          check_release_asset() {
+            local version="$1" tag asset url
+            [ -n "${version}" ] && [ "${version}" != "null" ] || return 0
+            tag="${version#v}"
+            asset="rustfs_${tag//-/.}_amd64.deb"
+            url="https://github.com/rustfs/rustfs/releases/download/${tag}/${asset}"
+            if ! gh api "repos/rustfs/rustfs/releases/tags/${tag}" --jq '.assets[].name' 2>/dev/null | grep -qxF "${asset}"; then
+              echo "ERROR: release ${tag} has no downloadable asset ${asset}:" >&2
+              echo "       ${url}" >&2
+              echo "Pick a tag whose release ships a .deb (check its release assets)." >&2
+              exit 1
+            fi
+            echo "resolved ${tag} -> ${url}"
+          }
+          if [ -z "${FROM_URL}" ]; then
+            check_release_asset "${FROM_VERSION}"
+          fi
+          if [ -z "${TO_URL}" ]; then
+            check_release_asset "${TO_VERSION}"
+          fi
           ./auto-testing/rustfs-upgrade-test.sh "${ARGS[@]}"
 
       - name: Generate report

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

@@ -42,6 +42,7 @@ jobs:
       - name: Check latest scheduled runs
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          RUSTFS_DEFAULT_BRANCH: ${{ github.event.repository.default_branch }}
         run: |
           set +e
           python3 scripts/check_scheduled_validation_freshness.py \

.gitignore

@@ -33,6 +33,7 @@ profile.json
 *.zst
 .secrets
 *.go
+!crates/zip/tests/fixtures/snowball/**/generate/*.go
 *.pb
 *.svg
 deploy/logs/*.log.*

.pre-commit-config.yaml

@@ -3,9 +3,9 @@
 repos:
   - repo: local
     hooks:
-      - id: rustfs-dev-check
+      - id: rustfs-fmt-check
-        name: rustfs dev-check
+        name: Rust formatting
-        entry: make dev-check
+        entry: cargo fmt --all --check
         language: system
         types: [rust]
         pass_filenames: false

CONTRIBUTING.md

@@ -109,24 +109,17 @@ affected boundaries and risks. CI still runs its configured repository gates.
 
 ### 🔒 Git Pre-commit Hooks (optional)
 
-Git hooks are **not** versioned in this repository, so a fresh clone has no
+The optional hook uses the checked-in `.pre-commit-config.yaml`. Install [pre-commit](https://pre-commit.com/#installation), then run this from the checkout or a linked worktree:
-active pre-commit hook. If you add your own `.git/hooks/pre-commit` (a good
-choice is a one-liner that runs `make pre-commit`), you can mark it executable
-with:
 
 ```bash
 make setup-hooks
 ```
 
-Or manually:
+The hook runs `cargo fmt --all --check` when staged files include Rust source. It does not compile the workspace or run tests. Fix formatting with `cargo fmt --all`, inspect and stage the result, then commit again.
 
-```bash
+`pre-commit install` resolves Git's hook directory for linked worktrees and preserves an existing hook in migration mode. If you use `core.hooksPath`, keep that hook manager and integrate `pre-commit run` there; the installer refuses to silently replace that configuration.
-chmod +x .git/hooks/pre-commit
-```
 
-With or without a hook, follow the verification tiers in `AGENTS.md`. Run the
+A local hook provides early formatting feedback. With or without it, follow the verification tiers in `AGENTS.md`, run relevant behavioral tests, and satisfy the CI merge gates. `make pre-commit` and `make dev-check` remain explicit broader commands.
-applicable scoped checks, and reserve `make pre-pr` for broad cross-module
-changes whose impact cannot be bounded by those checks.
 
 ### 📝 Formatting Configuration
 
@@ -138,31 +131,11 @@ fn_call_width = 90
 single_line_let_else_max_width = 100
 ```
 
-### 🚫 Commit Prevention
-
-If you set up a pre-commit hook and your code doesn't meet the formatting requirements, the hook will:
-
-1. **Block the commit** and show clear error messages
-2. **Provide exact commands** to fix the issues
-3. **Guide you through** the resolution process
-
-Example output when formatting fails:
-
-```
-❌ Code formatting check failed!
-💡 Please run 'cargo fmt --all' to format your code before committing.
-
-🔧 Quick fix:
-   cargo fmt --all
-   git add .
-   git commit
-```
-
 ### 🔄 Development Workflow
 
 1. **Make your changes**
 2. **Format your code**: `make fmt` or `cargo fmt --all`
-3. **Run the fast gate**: `make pre-commit` (no clippy, no tests)
+3. **Select relevant checks** using the validation tier in `AGENTS.md`; use `make pre-commit` when its broader fast gate adds useful coverage
 4. **Commit your changes**: `git commit -m "your message"`
 5. **Complete the applicable multi-role adversarial review** for non-exempt changes (see `AGENTS.md`)
 6. **Run applicable scoped checks before opening/updating a PR**; consider
@@ -206,11 +179,12 @@ Configure your IDE to:
 #### Pre-commit hook not running?
 
 ```bash
-# Check if hook is executable
+pre-commit validate-config
-ls -la .git/hooks/pre-commit
+pre-commit run --all-files
-
+# Inspect any configured hook manager; do not overwrite it.
-# Make it executable if needed
+git config --get core.hooksPath
-chmod +x .git/hooks/pre-commit
+# Install if no separate hook manager is configured.
+make setup-hooks
 ```
 
 #### Formatting issues?

Cargo.toml

@@ -168,7 +168,7 @@ reqwest = "0.13.4"
 rustfs-kafka-async = { version = "1.3.1" }
 socket2 = { version = "0.6.5" }
 tokio = { version = "1.53.1" }
-tokio-rustls = { default-features = false, version = "0.26.4" }
+tokio-rustls = { default-features = false, version = "0.26.5" }
 tokio-stream = { version = "0.1.19" }
 tokio-test = "0.4.5"
 tokio-util = { version = "0.7.19" }
@@ -199,10 +199,10 @@ serde_urlencoded = "0.7.1"
 # matching stable releases are not available yet, while previous stable lines
 # have incompatible APIs. Keep them exact-pinned and monitor upstream for stable
 # releases.
-aes-gcm = { version = "=0.11.1" }
+aes-gcm = { version = "0.11.1" }
-argon2 = { version = "=0.6.0" }
+argon2 = { version = "0.6.0" }
-blake2 = "=0.11.0"
+blake2 = "0.11.0"
-chacha20poly1305 = { version = "=0.11.0" }
+chacha20poly1305 = { version = "0.11.0" }
 crc-fast = "1.10.0"
 hmac = { version = "0.13.0" }
 jsonwebtoken = { version = "11.0.0" }
@@ -234,15 +234,19 @@ tokio-postgres-rustls = "0.14.0"
 # Utilities and Tools
 anyhow = "1.0.104"
 arc-swap = "1.9.2"
-# RUSTFS_COMPAT_TODO(tokio-tar-extension-limits): keep the fork pin until every parser hardening used by Snowball is released upstream. Remove after astral-sh/tokio-tar#118 is merged and a published release includes extension, physical-entry, and sparse limits, cancellation-safe sparse parsing, and error-fused entry streams.
+# RUSTFS_COMPAT_TODO(tokio-tar-extension-limits): keep the fork pin while Snowball and Swift still depend on it. Remove after Snowball uses a released tar-codec/tar-framing API that exposes precedence-resolved MinIO vendor records, RustFS preserves cancellation-safe ownership of large streamed members, footerless minio-go input is accepted only at an authenticated complete request boundary, the existing resource-limit, cancellation, and error-fuse regressions pass, and Swift no longer needs this fork.
 astral-tokio-tar = { git = "https://github.com/cxymds/tokio-tar.git", rev = "603756478b7668436e464519c77ccac22a99ba96" }
+# Candidate Snowball parser versions exercised by rustfs-zip compatibility fixtures.
+tar-codec = "0.0.14"
+tar-framing = "0.0.14"
 atoi = "3.1.0"
 atomic_enum = "0.3.0"
-aws-config = { version = "1.11.0" }
+aws-config = { version = "1.12.0" }
 aws-credential-types = { version = "1.3.0" }
-aws-sdk-kms = { default-features = false, version = "1.117.0" }
+aws-sdk-kms = { default-features = false, version = "1.118.0" }
-aws-sdk-s3 = { default-features = false, version = "1.144.0" }
+aws-sdk-s3 = { default-features = false, version = "1.145.0" }
-aws-sdk-sts = { default-features = false, version = "1.113.0" }
+aws-sdk-sts = { default-features = false, version = "1.114.0" }
+aws-smithy-async = { version = "1.3.0" }
 aws-smithy-http-client = { default-features = false, version = "1.4.0" }
 aws-smithy-runtime-api = { version = "1.16.0" }
 aws-smithy-types = { version = "1.6.3" }
@@ -339,7 +343,7 @@ windows = { version = "0.62.2" }
 windows-sys = "0.61.2"
 xxhash-rust = { version = "0.8.18" }
 zip = "8.6.0"
-zstd = "0.13.3"
+zstd = "0.14.0"
 
 # Observability and Metrics
 metrics = "0.24.6"

crates/config/README.md

@@ -130,6 +130,21 @@ Scanner cycle budget controls:
   - timeout returns S3 `SlowDown`, so clients should use normal SDK retry handling.
   - this is not a fdatasync or group-commit switch. Track fdatasync batching separately with `rustfs_s3_put_object_rename_fdatasync_batch_files`.
 
+## Remote tier timeout environment variables
+
+- `RUSTFS_TIER_REMOTE_CONNECT_TIMEOUT_SECS`
+  - remote tier TCP connect timeout.
+  - default is `10`.
+  - must be positive; zero fails tier client initialization, while an invalid integer is logged and falls back to the default.
+- `RUSTFS_TIER_REMOTE_REQUEST_TIMEOUT_SECS`
+  - remote tier request timeout through response headers.
+  - default is `86400` so large transition uploads keep a production-safe budget.
+  - must be positive; zero fails tier client initialization, while an invalid integer is logged and falls back to the default. Very large values are accepted and act as a correspondingly long budget.
+- `RUSTFS_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS`
+  - maximum idle time between remote tier response-body chunks.
+  - default is `60`; the timer resets only when non-empty body data keeps progressing.
+  - must be positive; zero fails tier client initialization, while an invalid integer is logged and falls back to the default.
+
 ## Drive timeout environment variables
 
 - `RUSTFS_DRIVE_METADATA_TIMEOUT_SECS`

crates/config/src/constants/object.rs

@@ -137,6 +137,28 @@ pub const DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED: bool = false;
 const _: () = assert!(!DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE);
 const _: () = assert!(!DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED);
 
+/// Environment variable for remote tier TCP connect timeout in seconds.
+pub const ENV_TIER_REMOTE_CONNECT_TIMEOUT_SECS: &str = "RUSTFS_TIER_REMOTE_CONNECT_TIMEOUT_SECS";
+/// Default remote tier TCP connect timeout in seconds.
+pub const DEFAULT_TIER_REMOTE_CONNECT_TIMEOUT_SECS: u64 = 10;
+
+/// Environment variable for the remote tier request timeout in seconds.
+///
+/// This bounds upload/download request progress through response headers. The
+/// default is intentionally large so multi-TiB transition uploads keep their
+/// previous production budget while black-hole remotes no longer wait forever.
+pub const ENV_TIER_REMOTE_REQUEST_TIMEOUT_SECS: &str = "RUSTFS_TIER_REMOTE_REQUEST_TIMEOUT_SECS";
+/// Default remote tier request timeout in seconds.
+pub const DEFAULT_TIER_REMOTE_REQUEST_TIMEOUT_SECS: u64 = 24 * 60 * 60;
+
+/// Environment variable for remote tier response-body idle timeout in seconds.
+///
+/// The timer is re-armed on every non-empty response-body chunk, so slow but
+/// progressing remotes can continue while silent response bodies are cancelled.
+pub const ENV_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS: &str = "RUSTFS_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS";
+/// Default remote tier response-body idle timeout in seconds.
+pub const DEFAULT_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS: u64 = 60;
+
 /// Request the object-transaction fencing contract used by storage-owned
 /// cleanup receipts and lock-window optimizations.
 ///
@@ -812,6 +834,16 @@ mod remote_version_state_tests {
         );
     }
 
+    #[test]
+    fn remote_tier_timeout_env_names_are_stable() {
+        assert_eq!(super::ENV_TIER_REMOTE_CONNECT_TIMEOUT_SECS, "RUSTFS_TIER_REMOTE_CONNECT_TIMEOUT_SECS");
+        assert_eq!(super::ENV_TIER_REMOTE_REQUEST_TIMEOUT_SECS, "RUSTFS_TIER_REMOTE_REQUEST_TIMEOUT_SECS");
+        assert_eq!(
+            super::ENV_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS,
+            "RUSTFS_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS"
+        );
+    }
+
     #[test]
     fn data_movement_part_checksum_gate_uses_stable_environment_names() {
         assert_eq!(super::ENV_DATA_MOVEMENT_PART_CHECKSUMS_WRITE, "RUSTFS_DATA_MOVEMENT_PART_CHECKSUMS_WRITE");

crates/e2e_test/src/replication_extension_test.rs

@@ -6743,6 +6743,99 @@ async fn test_site_replication_replicates_object_with_bucket_versioning_real_dua
     Ok(())
 }
 
+#[tokio::test]
+async fn test_site_replication_replays_bucket_created_during_peer_outage_real_dual_node() -> TestResult {
+    init_logging();
+
+    // Keep compilation outside the scenario timeout. Recovery itself waits
+    // for the production 30-second lightweight retry tick.
+    let _rustfs_binary = rustfs_binary_path();
+
+    match timeout(Duration::from_secs(150), async {
+        let mut site_env = replication_fast_env();
+        site_env.extend_from_slice(LOOPBACK_REPLICATION_TARGET_ENV);
+
+        let mut site_a_env = RustFSTestEnvironment::new().await?;
+        site_a_env.start_rustfs_server_with_env(vec![], &site_env).await?;
+
+        let mut site_b_env = RustFSTestEnvironment::new().await?;
+        site_b_env.start_rustfs_server_without_cleanup_with_env(&site_env).await?;
+
+        let site_a_client = site_a_env.create_s3_client();
+        let site_b_client = site_b_env.create_s3_client();
+        let bucket = "site-repl-peer-outage";
+        let key = "after-recovery.txt";
+        let payload = b"site replication recovered the missed bucket".to_vec();
+
+        let add_status = site_replication_add(
+            &site_a_env,
+            &[
+                PeerSite {
+                    name: "outage-site-a".to_string(),
+                    endpoint: site_a_env.url.clone(),
+                    access_key: site_a_env.access_key.clone(),
+                    secret_key: site_a_env.secret_key.clone(),
+                    ..Default::default()
+                },
+                PeerSite {
+                    name: "outage-site-b".to_string(),
+                    endpoint: site_b_env.url.clone(),
+                    access_key: site_b_env.access_key.clone(),
+                    secret_key: site_b_env.secret_key.clone(),
+                    ..Default::default()
+                },
+            ],
+        )
+        .await?;
+        assert!(add_status.success, "unexpected site add result: {add_status:?}");
+        wait_for_site_replication_enabled(&site_a_env, 2).await?;
+        wait_for_site_replication_enabled(&site_b_env, 2).await?;
+
+        site_b_env.stop_server();
+        site_a_client.create_bucket().bucket(bucket).send().await?;
+        site_a_client.head_bucket().bucket(bucket).send().await?;
+
+        let queued = site_replication_info(&site_a_env)
+            .await?
+            .retry_stats
+            .ok_or("peer outage did not persist a site replication retry event")?;
+        assert!(queued.pending + queued.failed > 0, "peer outage retry queue was unexpectedly empty");
+
+        site_b_env.restart_server_preserving_data(vec![], &site_env).await?;
+        let recovery_deadline = tokio::time::Instant::now() + Duration::from_secs(75);
+        loop {
+            let bucket_recovered = site_b_client.head_bucket().bucket(bucket).send().await.is_ok();
+            let queue_empty = site_replication_info(&site_a_env).await?.retry_stats.is_none();
+            if bucket_recovered && queue_empty {
+                break;
+            }
+            if tokio::time::Instant::now() >= recovery_deadline {
+                return Err(format!(
+                    "site replication retry did not settle after peer recovery; bucket_recovered={bucket_recovered}, queue_empty={queue_empty}"
+                )
+                .into());
+            }
+            sleep(Duration::from_millis(250)).await;
+        }
+
+        site_a_client
+            .put_object()
+            .bucket(bucket)
+            .key(key)
+            .body(ByteStream::from(payload.clone()))
+            .send()
+            .await?;
+        assert_eq!(wait_for_object_on_target(&site_b_client, bucket, key).await?, payload);
+
+        Ok(())
+    })
+    .await
+    {
+        Ok(result) => result,
+        Err(_) => Err("site replication peer-outage recovery timed out after 150 seconds".into()),
+    }
+}
+
 /// Re-applying a site's own replication config must not disable the peer's reverse direction.
 ///
 /// `PutBucketReplication` broadcasts the config to every peer — the console's replication

crates/ecstore/Cargo.toml

@@ -215,6 +215,7 @@ serde_urlencoded.workspace = true
 google-cloud-storage = { workspace = true }
 google-cloud-auth = { workspace = true }
 faster-hex = { workspace = true }
+quick-xml = { workspace = true }
 ratelimit = { workspace = true }
 aws-smithy-http-client = { workspace = true, default-features = false, features = ["rustls-aws-lc"] }
 
@@ -244,6 +245,7 @@ windows-sys = { workspace = true, features = [
 windows-sys = { workspace = true, features = ["Win32_System_Ioctl"] }
 
 [dev-dependencies]
+aws-smithy-async.workspace = true
 tokio = { workspace = true, features = ["rt-multi-thread", "macros", "test-util", "fs"] }
 criterion = { workspace = true, features = ["html_reports"] }
 temp-env = { workspace = true, features = ["async_closure"] }

crates/ecstore/src/api/mod.rs

@@ -153,12 +153,13 @@ pub mod bucket {
             LastSourceError, LatencyBucketSnapshot, NEGATIVE_CACHE_MAX_ENTRIES, NegativeCache, OdmBucketSnapshot, OdmLookup,
             OdmOp, OdmOutcome, OdmStateError, OdmStats, OdmStatsSnapshot, OnDemandMigrationSys, PullError, PullFailureReason,
             PullFollower, PullLeader, PullOutcome, PullPath, PullResult, PullSlot, SOURCE_LATENCY_BUCKET_BOUNDS_MS,
-            SourceLatencySnapshot, source_client_spec,
+            SourceLatencySnapshot, source_backend_spec, source_client_spec,
         };
         pub use crate::bucket::on_demand_migration::{
-            ConfigPublishHook, FilterConfig, HeadPolicy, ON_DEMAND_MIGRATION_CONFIG_HOOK, ON_DEMAND_MIGRATION_CONFIG_VERSION,
+            AzureSourceConfig, ConfigPublishHook, FilterConfig, GcsSourceConfig, HeadPolicy, ON_DEMAND_MIGRATION_CONFIG_HOOK,
-            OnDemandMigrationConfig, OnDemandMigrationConfigError, PathStyle, PolicyConfig, Provider, RangeGetPolicy,
+            ON_DEMAND_MIGRATION_CONFIG_VERSION, OnDemandMigrationConfig, OnDemandMigrationConfigError, PathStyle, PolicyConfig,
-            SourceConfig, SourceCredentials, SourceErrorPolicy, SourceTimeout, TlsConfig, ValidationContext,
+            Provider, RangeGetPolicy, SourceConfig, SourceCredentials, SourceErrorPolicy, SourceTimeout, TlsConfig,
+            ValidationContext,
         };
         pub use crate::bucket::on_demand_migration::{
             EnqueueOutcome, LocalObject, MAX_MULTIPART_PARTS, OdmWriteBack, PULL_MAX_RETRIES, PULL_RETRY_BASE_DELAYS,
@@ -167,9 +168,10 @@ pub mod bucket {
             idle_guarded_body,
         };
         pub use crate::bucket::on_demand_migration::{
-            FetchRequest, LIST_THROUGH_TOKEN_VERSION, ListEntryKey, ListThroughCursor, ListThroughMerger, ListThroughToken,
+            FetchRequest, LIST_THROUGH_TOKEN_VERSION, ListEntryKey, ListPageError, ListThroughCursor, ListThroughMerger,
-            ListThroughTokenError, MAX_LIST_FETCHES_PER_SIDE, MergeOutcome, MergePick, MergeSide, SOURCE_LIST_MAX_RATE_WAIT,
+            ListThroughToken, ListThroughTokenError, MAX_LIST_FETCHES_PER_SIDE, MAX_LIST_NO_PROGRESS_PAGES, MergeOutcome,
-            SOURCE_LIST_RATE_PER_SEC, SourceListPlan, SourceListRateLimiter, decode_continuation_token, source_list_plan,
+            MergePick, MergeSide, SOURCE_LIST_MAX_RATE_WAIT, SOURCE_LIST_RATE_PER_SEC, SourceListPlan, SourceListRateLimiter,
+            decode_continuation_token, source_list_plan,
         };
         pub mod backfill {
             pub use crate::bucket::on_demand_migration::backfill::{
@@ -184,9 +186,9 @@ pub mod bucket {
         }
         pub mod source_client {
             pub use crate::bucket::on_demand_migration::source_client::{
-                SourceClient, SourceClientSpec, SourceError, SourceGet, SourceHead, SourceListRequest, SourceObject, SourcePage,
+                AzureAuth, AzureSourceSpec, GcsSourceSpec, SourceBackendSpec, SourceClient, SourceClientSpec, SourceError,
-                SourceProbe, SourceProvider, SourceSse, SourceTimeouts, USER_AGENT_SUFFIX, is_multipart_etag, range_header_value,
+                SourceGet, SourceHead, SourceListRequest, SourceObject, SourcePage, SourceProbe, SourceProvider, SourceSse,
-                resolve_path_style,
+                SourceTimeouts, USER_AGENT_SUFFIX, is_multipart_etag, range_header_value, resolve_path_style,
             };
         }
     }
@@ -196,15 +198,16 @@ pub mod bucket {
         pub use crate::bucket::metadata_sys::ConfigWriteLockProbe;
         pub use crate::bucket::metadata_sys::{
             BucketMetadataMutationGuard, BucketMetadataSys, ObjectLockConfigState, acquire_bucket_metadata_transaction_lock,
-            acquire_bucket_metadata_transaction_lock_for_incarnation, capture_bucket_metadata_incarnation, delete,
+            acquire_bucket_metadata_transaction_lock_for_incarnation, acquire_scanner_bucket_incarnation_fence,
-            delete_if_incarnation, delete_under_transaction_lock, get, get_accelerate_config, get_bucket_policy,
+            capture_bucket_metadata_incarnation, delete, delete_if_incarnation, delete_under_transaction_lock, get,
-            get_bucket_policy_raw, get_bucket_targets_config, get_config_from_disk, get_cors_config, get_durability_config,
+            get_accelerate_config, get_bucket_policy, get_bucket_policy_raw, get_bucket_targets_config, get_config_from_disk,
-            get_global_bucket_metadata_sys, get_lifecycle_config, get_logging_config, get_notification_config,
+            get_cors_config, get_durability_config, get_global_bucket_metadata_sys, get_lifecycle_config, get_logging_config,
-            get_object_lock_config, get_object_lock_config_state, get_on_demand_migration_config, get_public_access_block_config,
+            get_notification_config, get_object_lock_config, get_object_lock_config_state, get_on_demand_migration_config,
-            get_quota_config, get_replication_config, get_request_payment_config, get_sse_config, get_tagging_config,
+            get_public_access_block_config, get_quota_config, get_replication_config, get_request_payment_config, get_sse_config,
-            get_versioning_config, get_website_config, init_bucket_metadata_sys, list_bucket_targets, reload_bucket_metadata,
+            get_tagging_config, get_versioning_config, get_website_config, init_bucket_metadata_sys, list_bucket_targets,
-            remove_bucket_metadata, set_bucket_metadata, update, update_bucket_targets_under_transaction_lock,
+            reload_bucket_metadata, remove_bucket_metadata, set_bucket_metadata, update,
-            update_config_with, update_if_incarnation, update_quota_if_incarnation, update_under_transaction_lock,
+            update_bucket_targets_under_transaction_lock, update_config_with, update_if_incarnation, update_quota_if_incarnation,
+            update_under_transaction_lock,
         };
     }
 
@@ -560,6 +563,12 @@ pub mod set_disk {
     pub mod test_util {
         pub use crate::bucket::quota::reservation::fail_next_quota_ledger_save_for_test;
         pub use crate::set_disk::{MultipartCommitBarrier, MultipartCommitPause, PutObjectCommitBarrier, PutObjectCommitPause};
+
+        /// Keep a namespace commit pending until the returned owner is dropped.
+        #[must_use]
+        pub fn hold_namespace_commit(store: &crate::store::ECStore) -> impl Send + Sync {
+            store.ctx.begin_namespace_commit()
+        }
     }
 }
 

crates/ecstore/src/bucket/bucket_target_sys.rs

@@ -59,7 +59,7 @@ use rustfs_utils::http::{
     insert_header,
 };
 use serde::{Deserialize, Serialize};
-use std::collections::HashMap;
+use std::collections::{HashMap, HashSet};
 use std::error::Error;
 use std::fmt;
 use std::str::FromStr as _;
@@ -376,6 +376,11 @@ pub struct BucketTargetSys {
     /// [`SsecPassthroughCapability`]; reset alongside `arn_remotes_map`.
     ssec_passthrough_map: Arc<RwLock<HashMap<String, SsecPassthroughRecord>>>,
     pub targets_map: Arc<RwLock<HashMap<String, Vec<BucketTarget>>>>,
+    /// Buckets whose persisted `bucket-targets.json` exists but cannot be
+    /// decoded (rustfs/backlog#2282). Written under the bucket's update mutex
+    /// alongside `targets_map`, and read before it so an unreadable
+    /// configuration surfaces as a typed error instead of an empty target set.
+    unreadable_targets: Arc<RwLock<HashSet<String>>>,
     pub h_mutex: Arc<RwLock<HashMap<String, EpHealth>>>,
     target_h_mutex: Arc<RwLock<HashMap<String, EpHealth>>>,
     pub hc_client: Arc<HttpClient>,
@@ -419,6 +424,7 @@ impl BucketTargetSys {
             arn_remotes_map: Arc::new(RwLock::new(HashMap::new())),
             ssec_passthrough_map: Arc::new(RwLock::new(HashMap::new())),
             targets_map: Arc::new(RwLock::new(HashMap::new())),
+            unreadable_targets: Arc::new(RwLock::new(HashSet::new())),
             h_mutex: Arc::new(RwLock::new(HashMap::new())),
             target_h_mutex: Arc::new(RwLock::new(HashMap::new())),
             hc_client: Arc::new(build_health_check_client()),
@@ -628,30 +634,40 @@ impl BucketTargetSys {
         health_map.clone()
     }
 
-    pub async fn list_targets(&self, bucket: &str, arn_type: &str) -> Vec<BucketTarget> {
+    /// Targets of one bucket, or of every bucket when `bucket` is empty.
+    ///
+    /// A bucket that simply has no targets yields an empty list; a bucket
+    /// whose persisted configuration cannot be decoded is an error, so an
+    /// admin listing reports the fault instead of an empty list that reads as
+    /// "replication is not configured" (rustfs/backlog#2282).
+    pub async fn list_targets(&self, bucket: &str, arn_type: &str) -> Result<Vec<BucketTarget>, BucketTargetError> {
         let health_stats = self.target_health_stats().await;
         let mut targets = Vec::new();
 
         if !bucket.is_empty() {
-            if let Ok(bucket_targets) = self.list_bucket_targets(bucket).await {
+            match self.list_bucket_targets(bucket).await {
-                for mut target in bucket_targets.targets {
+                Ok(bucket_targets) => {
-                    if arn_type.is_empty() || target.target_type.to_string() == arn_type {
+                    for mut target in bucket_targets.targets {
-                        if let Some(health) = health_stats.get(&target.arn) {
+                        if arn_type.is_empty() || target.target_type.to_string() == arn_type {
-                            target.total_downtime = health.offline_duration;
+                            if let Some(health) = health_stats.get(&target.arn) {
-                            target.online = health.online;
+                                target.total_downtime = health.offline_duration;
-                            target.last_online = health.last_online;
+                                target.online = health.online;
-                            target.latency = target::LatencyStat {
+                                target.last_online = health.last_online;
-                                curr: health.latency.curr,
+                                target.latency = target::LatencyStat {
-                                avg: health.latency.avg,
+                                    curr: health.latency.curr,
-                                max: health.latency.peak,
+                                    avg: health.latency.avg,
-                            };
+                                    max: health.latency.peak,
-                            target.offline_count = health.offline_count;
+                                };
+                                target.offline_count = health.offline_count;
+                            }
+                            targets.push(target);
                         }
-                        targets.push(target);
                     }
                 }
+                Err(BucketTargetError::BucketRemoteTargetNotFound { .. }) => {}
+                Err(err) => return Err(err),
             }
-            return targets;
+            return Ok(targets);
         }
 
         let targets_map = self.targets_map.read().await;
@@ -674,10 +690,16 @@ impl BucketTargetSys {
             }
         }
 
-        targets
+        Ok(targets)
     }
 
     pub async fn list_bucket_targets(&self, bucket: &str) -> Result<BucketTargets, BucketTargetError> {
+        if self.unreadable_targets.read().await.contains(bucket) {
+            return Err(BucketTargetError::BucketRemoteTargetsUnreadable {
+                bucket: bucket.to_string(),
+            });
+        }
+
         let targets_map = self.targets_map.read().await;
         if let Some(targets) = targets_map.get(bucket) {
             Ok(BucketTargets {
@@ -690,13 +712,30 @@ impl BucketTargetSys {
         }
     }
 
+    /// Record that this bucket's persisted targets configuration exists but
+    /// cannot be decoded (rustfs/backlog#2282).
+    ///
+    /// Any snapshot published from an earlier readable load is deliberately
+    /// left in place: withdrawing it would produce exactly the silent "no
+    /// targets configured" state this marker exists to prevent. The marker is
+    /// cleared by the next successful publish, which is what makes a repaired
+    /// configuration take effect without a restart.
+    pub async fn mark_targets_unreadable(&self, bucket: &str) {
+        let update_mutex = self.target_update_mutex(bucket).await;
+        let _update_guard = update_mutex.lock().await;
+
+        self.unreadable_targets.write().await.insert(bucket.to_string());
+    }
+
     pub async fn delete(&self, bucket: &str) {
         let update_mutex = self.target_update_mutex(bucket).await;
         let _update_guard = update_mutex.lock().await;
 
-        // Lock order: targets_map, then arn_remotes_map, then target_h_mutex,
+        // Lock order: unreadable_targets, then targets_map, then
-        // then ssec_passthrough_map (always last; also taken standalone by the
+        // arn_remotes_map, then target_h_mutex, then ssec_passthrough_map
-        // capability accessors).
+        // (always last; also taken standalone by the capability accessors).
+        self.unreadable_targets.write().await.remove(bucket);
+
         let mut targets_map = self.targets_map.write().await;
         let mut arn_remotes_map = self.arn_remotes_map.write().await;
         let mut health_map = self.target_h_mutex.write().await;
@@ -1093,6 +1132,11 @@ impl BucketTargetSys {
     /// Keeping persisted-config reads under the same mutex prevents a stale
     /// reload from overwriting a concurrent credential rotation.
     async fn update_all_targets_locked(&self, bucket: &str, targets: Option<&BucketTargets>) {
+        // Reaching here means the persisted configuration decoded, so the
+        // unreadable marker (if any) is stale. Cleared before the maps below
+        // so `unreadable_targets` stays the outermost of this module's locks.
+        self.unreadable_targets.write().await.remove(bucket);
+
         let mut clients = Vec::new();
         if let Some(new_targets) = targets {
             for target in &new_targets.targets {
@@ -1100,9 +1144,9 @@ impl BucketTargetSys {
             }
         }
 
-        // Lock order: targets_map, then arn_remotes_map, then target_h_mutex,
+        // Lock order: unreadable_targets (above), then targets_map, then
-        // then ssec_passthrough_map (always last; also taken standalone by the
+        // arn_remotes_map, then target_h_mutex, then ssec_passthrough_map
-        // capability accessors).
+        // (always last; also taken standalone by the capability accessors).
         let mut targets_map = self.targets_map.write().await;
         let mut arn_remotes_map = self.arn_remotes_map.write().await;
         let mut health_map = self.target_h_mutex.write().await;
@@ -1161,6 +1205,11 @@ impl BucketTargetSys {
     }
 
     pub async fn set(&self, bucket: &str, meta: &BucketMetadata) {
+        if meta.bucket_targets_unreadable() {
+            self.mark_targets_unreadable(bucket).await;
+            return;
+        }
+
         let Some(config) = &meta.bucket_target_config else {
             return;
         };
@@ -2276,6 +2325,13 @@ pub enum BucketTargetError {
     BucketRemoteTargetNotFound {
         bucket: String,
     },
+    /// The bucket's persisted targets configuration exists but cannot be
+    /// decoded. Distinct from `BucketRemoteTargetNotFound`, which means the
+    /// bucket genuinely has no targets: callers must not degrade this one to
+    /// an empty target set (rustfs/backlog#2282).
+    BucketRemoteTargetsUnreadable {
+        bucket: String,
+    },
     BucketRemoteArnTypeInvalid {
         bucket: String,
     },
@@ -2309,6 +2365,9 @@ impl fmt::Display for BucketTargetError {
             BucketTargetError::BucketRemoteTargetNotFound { bucket } => {
                 write!(f, "Remote target not found for bucket: {bucket}")
             }
+            BucketTargetError::BucketRemoteTargetsUnreadable { bucket } => {
+                write!(f, "Persisted replication target configuration is unreadable for bucket: {bucket}")
+            }
             BucketTargetError::BucketRemoteArnTypeInvalid { bucket } => {
                 write!(f, "Invalid ARN type for bucket: {bucket}")
             }
@@ -3256,7 +3315,7 @@ mod tests {
             }],
         );
 
-        let targets = sys.list_targets("", "").await;
+        let targets = sys.list_targets("", "").await.expect("listing every bucket's targets");
 
         assert_eq!(targets.len(), 1);
         assert!(!targets[0].online);

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

@@ -584,33 +584,173 @@ impl ExpiryOp for FreeVersionTask {
     }
 }
 
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+enum TransitionDeleteVersionPlan {
+    Direct { version_id_exact: bool },
+    ProbeLegacyUnknown,
+}
+
+fn legacy_transition_version_state_missing(oi: &ObjectInfo) -> Result<bool, std::io::Error> {
+    use rustfs_utils::http::metadata_compat::{
+        SUFFIX_TRANSITIONED_VERSION_ID, SUFFIX_TRANSITIONED_VERSION_STATE, contains_key_str, get_consistent_str,
+    };
+
+    if !contains_key_str(&oi.user_defined, SUFFIX_TRANSITIONED_VERSION_STATE) {
+        let version_key_present = contains_key_str(&oi.user_defined, SUFFIX_TRANSITIONED_VERSION_ID);
+        if version_key_present {
+            if oi.transitioned_object.version_id.is_empty() {
+                let has_non_empty_version = oi.user_defined.iter().any(|(key, value)| {
+                    rustfs_utils::http::metadata_compat::strip_internal_prefix_preserving_case(key)
+                        .is_some_and(|suffix| suffix.eq_ignore_ascii_case(SUFFIX_TRANSITIONED_VERSION_ID))
+                        && !value.is_empty()
+                });
+                if !has_non_empty_version {
+                    // MinIO writes the transitioned-versionID key with an empty value
+                    // for unversioned tier objects. The backend probe remains the proof.
+                    return Ok(true);
+                }
+            } else if get_consistent_str(&oi.user_defined, SUFFIX_TRANSITIONED_VERSION_ID)
+                == Some(oi.transitioned_object.version_id.as_str())
+            {
+                return Ok(true);
+            }
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidData,
+                "legacy remote tier version metadata is conflicting or malformed",
+            ));
+        }
+        if !oi.transitioned_object.version_id.is_empty() {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidData,
+                "legacy remote tier version metadata is missing or inconsistent",
+            ));
+        }
+        return Ok(true);
+    }
+    let persisted = get_consistent_str(&oi.user_defined, SUFFIX_TRANSITIONED_VERSION_STATE).ok_or_else(|| {
+        std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier object has conflicting transition version state metadata",
+        )
+    })?;
+    if persisted != oi.transition_version_state.as_str() {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier object transition version state metadata changed during decoding",
+        ));
+    }
+    Ok(false)
+}
+
+fn transition_remote_version_delete_plan(oi: &ObjectInfo) -> Result<TransitionDeleteVersionPlan, std::io::Error> {
+    match oi.transition_version_state {
+        rustfs_filemeta::TransitionVersionState::Unknown => {
+            if legacy_transition_version_state_missing(oi)? {
+                Ok(TransitionDeleteVersionPlan::ProbeLegacyUnknown)
+            } else {
+                validate_transition_remote_version(oi)
+                    .map(|version_id_exact| TransitionDeleteVersionPlan::Direct { version_id_exact })
+            }
+        }
+        _ => validate_transition_remote_version(oi)
+            .map(|version_id_exact| TransitionDeleteVersionPlan::Direct { version_id_exact }),
+    }
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+struct ResolvedTransitionDeleteVersion {
+    version_id_exact: bool,
+    remote_already_missing: bool,
+}
+
 async fn acquire_free_version_tier_lease(
     oi: &ObjectInfo,
     tier_config_mgr: &Arc<RwLock<TierConfigMgr>>,
-) -> Result<(TierOperationLease, bool), std::io::Error> {
+) -> Result<(TierOperationLease, TransitionDeleteVersionPlan), std::io::Error> {
-    let version_id_exact = validate_transition_remote_version(oi)?;
+    let delete_plan = transition_remote_version_delete_plan(oi)?;
     let identity = tier_destination_id_from_metadata(&oi.user_defined)?
         .ok_or_else(|| std::io::Error::other("tier free-version has no durable backend identity"))?;
     let lease =
         TierConfigMgr::acquire_operation_lease_for_backend_identity(tier_config_mgr, &oi.transitioned_object.tier, identity)
             .await
             .map_err(std::io::Error::other)?;
-    Ok((lease, version_id_exact))
+    Ok((lease, delete_plan))
+}
+
+async fn resolve_transition_delete_version_plan(
+    oi: &ObjectInfo,
+    lease: &TierOperationLease,
+    delete_plan: TransitionDeleteVersionPlan,
+) -> Result<ResolvedTransitionDeleteVersion, std::io::Error> {
+    match delete_plan {
+        TransitionDeleteVersionPlan::Direct { version_id_exact } => Ok(ResolvedTransitionDeleteVersion {
+            version_id_exact,
+            remote_already_missing: false,
+        }),
+        TransitionDeleteVersionPlan::ProbeLegacyUnknown => {
+            let expected_version = oi.transitioned_object.version_id.as_str();
+            if expected_version.is_empty() {
+                return Err(std::io::Error::new(
+                    std::io::ErrorKind::WouldBlock,
+                    "remote tier cannot safely delete a legacy object without an exact version ID",
+                ));
+            }
+            let probe = lease
+                .probe_transition_version(&oi.transitioned_object.name, expected_version)
+                .await?;
+            match (expected_version, probe) {
+                (expected, crate::services::tier::warm_backend::TransitionCandidateProbe::VersionedPresent(actual))
+                    if expected == actual =>
+                {
+                    lease.validate_remote_version_id(expected)?;
+                    Ok(ResolvedTransitionDeleteVersion {
+                        version_id_exact: true,
+                        remote_already_missing: false,
+                    })
+                }
+                (_, crate::services::tier::warm_backend::TransitionCandidateProbe::Missing) => {
+                    Ok(ResolvedTransitionDeleteVersion {
+                        version_id_exact: false,
+                        remote_already_missing: true,
+                    })
+                }
+                (_, crate::services::tier::warm_backend::TransitionCandidateProbe::Unsupported) => Err(std::io::Error::new(
+                    std::io::ErrorKind::Unsupported,
+                    "remote tier cannot prove legacy transition delete state",
+                )),
+                _ => Err(std::io::Error::new(
+                    std::io::ErrorKind::WouldBlock,
+                    "remote tier object version state is unknown",
+                )),
+            }
+        }
+    }
+}
+
+async fn execute_resolved_transition_delete(
+    oi: &ObjectInfo,
+    lease: &TierOperationLease,
+    resolved: ResolvedTransitionDeleteVersion,
+) -> Result<(), std::io::Error> {
+    if !resolved.remote_already_missing {
+        delete_object_from_remote_tier_with_lease_idempotent(
+            &oi.transitioned_object.name,
+            &oi.transitioned_object.version_id,
+            lease,
+            resolved.version_id_exact,
+        )
+        .await?;
+    }
+    Ok(())
 }
 
 async fn delete_free_version_remote_object_with_lease(
     oi: &ObjectInfo,
     lease: &TierOperationLease,
-    version_id_exact: bool,
+    delete_plan: TransitionDeleteVersionPlan,
 ) -> Result<(), std::io::Error> {
-    delete_object_from_remote_tier_with_lease_idempotent(
+    let resolved = resolve_transition_delete_version_plan(oi, lease, delete_plan).await?;
-        &oi.transitioned_object.name,
+    execute_resolved_transition_delete(oi, lease, resolved).await
-        &oi.transitioned_object.version_id,
-        lease,
-        version_id_exact,
-    )
-    .await?;
-    Ok(())
 }
 
 fn free_version_physical_topology_generation(api: &ECStore) -> String {
@@ -641,6 +781,16 @@ fn free_version_remote_tuple_matches(candidate: &ObjectInfo, expected: &ObjectIn
     if candidate.transition_version_state == rustfs_filemeta::TransitionVersionState::Unknown
         || expected.transition_version_state == rustfs_filemeta::TransitionVersionState::Unknown
     {
+        let candidate_legacy_missing = legacy_transition_version_state_missing(candidate)?;
+        let expected_legacy_missing = legacy_transition_version_state_missing(expected)?;
+        if candidate.transition_version_state == rustfs_filemeta::TransitionVersionState::Unknown
+            && expected.transition_version_state == rustfs_filemeta::TransitionVersionState::Unknown
+            && candidate_legacy_missing
+            && expected_legacy_missing
+            && candidate.transitioned_object.version_id == expected.transitioned_object.version_id
+        {
+            return Ok(true);
+        }
         return Err(std::io::Error::new(
             std::io::ErrorKind::WouldBlock,
             "tier free-version remote version state is unknown",
@@ -716,7 +866,7 @@ async fn cleanup_free_version_exact(api: Arc<ECStore>, oi: &ObjectInfo, cancel:
         .acquire_bucket_lifecycle_read_lock(&oi.bucket)
         .await
         .map_err(std::io::Error::other)?;
-    let (lease, version_id_exact) = acquire_free_version_tier_lease(oi, &api.tier_config_mgr()).await?;
+    let (lease, delete_plan) = acquire_free_version_tier_lease(oi, &api.tier_config_mgr()).await?;
     let local_object = encode_dir_object(&oi.name);
     let object_guards = api
         .acquire_all_physical_object_write_locks("tier_free_version_cleanup", &oi.bucket, &local_object)
@@ -734,16 +884,30 @@ async fn cleanup_free_version_exact(api: Arc<ECStore>, oi: &ObjectInfo, cancel:
             "tier free-version cleanup fence is invalid before remote delete",
         ));
     }
+    let resolved = tokio::select! {
+        _ = cancel.cancelled() => {
+            return Err(std::io::Error::new(std::io::ErrorKind::Interrupted, "tier free-version cleanup was cancelled"));
+        }
+        result = tokio::time::timeout_at(deadline, resolve_transition_delete_version_plan(oi, &lease, delete_plan)) => {
+            result.map_err(|_| {
+                std::io::Error::new(std::io::ErrorKind::TimedOut, "tier free-version remote probe timed out")
+            })??
+        }
+    };
+    if !free_version_cleanup_fences_current(&topology_generation, &api, &bucket_guard, &object_guards, &lease, cancel, deadline) {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::WouldBlock,
+            "tier free-version cleanup fence changed after remote probe",
+        ));
+    }
     tokio::select! {
         _ = cancel.cancelled() => {
             return Err(std::io::Error::new(std::io::ErrorKind::Interrupted, "tier free-version cleanup was cancelled"));
         }
-        result = tokio::time::timeout_at(
+        result = tokio::time::timeout_at(deadline, execute_resolved_transition_delete(oi, &lease, resolved)) => {
-            deadline,
+            result.map_err(|_| {
-            delete_free_version_remote_object_with_lease(oi, &lease, version_id_exact),
+                std::io::Error::new(std::io::ErrorKind::TimedOut, "tier free-version remote delete timed out")
-        ) => {
+            })??;
-            result
-                .map_err(|_| std::io::Error::new(std::io::ErrorKind::TimedOut, "tier free-version remote delete timed out"))??;
         }
     }
     if !free_version_cleanup_fences_current(&topology_generation, &api, &bucket_guard, &object_guards, &lease, cancel, deadline) {
@@ -791,8 +955,8 @@ async fn delete_free_version_remote_object(
     oi: &ObjectInfo,
     tier_config_mgr: &Arc<RwLock<TierConfigMgr>>,
 ) -> Result<(), std::io::Error> {
-    let (lease, version_id_exact) = acquire_free_version_tier_lease(oi, tier_config_mgr).await?;
+    let (lease, delete_plan) = acquire_free_version_tier_lease(oi, tier_config_mgr).await?;
-    delete_free_version_remote_object_with_lease(oi, &lease, version_id_exact).await
+    delete_free_version_remote_object_with_lease(oi, &lease, delete_plan).await
 }
 
 #[allow(
@@ -808,8 +972,8 @@ where
     F: FnOnce() -> Fut,
     Fut: std::future::Future<Output = T>,
 {
-    let (lease, version_id_exact) = acquire_free_version_tier_lease(oi, tier_config_mgr).await?;
+    let (lease, delete_plan) = acquire_free_version_tier_lease(oi, tier_config_mgr).await?;
-    delete_free_version_remote_object_with_lease(oi, &lease, version_id_exact).await?;
+    delete_free_version_remote_object_with_lease(oi, &lease, delete_plan).await?;
     let result = delete_local().await;
     drop(lease);
     Ok(result)
@@ -4688,6 +4852,39 @@ fn validate_transition_remote_version(oi: &ObjectInfo) -> Result<bool, std::io::
     }
 }
 
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+enum TransitionReadVersionPlan {
+    Direct,
+    ProbeLegacyUnversioned,
+}
+
+const LEGACY_TRANSITION_READ_PROBE_TIMEOUT: StdDuration = StdDuration::from_secs(30);
+
+fn transition_remote_version_read_plan(oi: &ObjectInfo) -> Result<TransitionReadVersionPlan, std::io::Error> {
+    let version = oi.transitioned_object.version_id.as_str();
+    match oi.transition_version_state {
+        rustfs_filemeta::TransitionVersionState::Unknown => {
+            if !legacy_transition_version_state_missing(oi)? {
+                return validate_transition_remote_version(oi).map(|_| TransitionReadVersionPlan::Direct);
+            }
+            if version.is_empty() {
+                Ok(TransitionReadVersionPlan::ProbeLegacyUnversioned)
+            } else {
+                Ok(TransitionReadVersionPlan::Direct)
+            }
+        }
+        rustfs_filemeta::TransitionVersionState::KnownDisabled if version.is_empty() => Ok(TransitionReadVersionPlan::Direct),
+        rustfs_filemeta::TransitionVersionState::SuspendedNull if version == "null" => Ok(TransitionReadVersionPlan::Direct),
+        rustfs_filemeta::TransitionVersionState::Exact if !version.is_empty() && version != "null" => {
+            Ok(TransitionReadVersionPlan::Direct)
+        }
+        _ => Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier object version state conflicts with its version ID",
+        )),
+    }
+}
+
 // The resolver joins the tier manager as the second injected port this read
 // needs; grouping the request half into a struct would churn every call site of
 // a bug fix.
@@ -4702,7 +4899,12 @@ pub(crate) async fn get_transitioned_object_reader_with_tier_manager(
     tier_config_mgr: &Arc<RwLock<TierConfigMgr>>,
     resolver: Option<&dyn ObjectEncryptionResolver>,
 ) -> Result<GetObjectReader, std::io::Error> {
-    validate_transition_remote_version(oi)?;
+    let read_plan = transition_remote_version_read_plan(oi)?;
+    // Reject invalid ranges and encryption requests before a compatibility
+    // probe can amplify them into remote listing work.
+    let plan = ReadPlan::build_for_request(rs.clone(), oi, opts, h, resolver)
+        .await
+        .map_err(|err| std::io::Error::other(format!("building the read plan for {bucket}/{object} failed: {err}")))?;
     let expected_identity = tier_destination_id_from_metadata(&oi.user_defined)?;
     let lease = match expected_identity {
         Some(identity) => {
@@ -4716,7 +4918,36 @@ pub(crate) async fn get_transitioned_object_reader_with_tier_manager(
         Err(err) => return Err(std::io::Error::other(err)),
     };
 
-    tgt_client.validate_remote_version_id(&oi.transitioned_object.version_id)?;
+    match read_plan {
+        TransitionReadVersionPlan::Direct => {
+            tgt_client.validate_remote_version_id(&oi.transitioned_object.version_id)?;
+        }
+        TransitionReadVersionPlan::ProbeLegacyUnversioned => {
+            // RUSTFS_COMPAT_TODO(backlog#2203): remove operation-time probing
+            // after an admin reconcile can persist every proven legacy state.
+            let probe = tokio::time::timeout(
+                LEGACY_TRANSITION_READ_PROBE_TIMEOUT,
+                tgt_client.probe_transition_candidate(&oi.transitioned_object.name),
+            )
+            .await
+            .map_err(|_| std::io::Error::new(std::io::ErrorKind::TimedOut, "legacy remote tier version probe timed out"))??;
+            match probe {
+                crate::services::tier::warm_backend::TransitionCandidateProbe::UnversionedPresent => {}
+                crate::services::tier::warm_backend::TransitionCandidateProbe::Unsupported => {
+                    return Err(std::io::Error::new(
+                        std::io::ErrorKind::Unsupported,
+                        "remote tier cannot prove legacy unversioned transition state",
+                    ));
+                }
+                _ => {
+                    return Err(std::io::Error::new(
+                        std::io::ErrorKind::InvalidData,
+                        "remote tier object version state is unknown",
+                    ));
+                }
+            }
+        }
+    }
 
     // The same read plan the local path uses, so the tier fetch is positioned in
     // the object's *stored* coordinate system and the stream is handed the same
@@ -4724,9 +4955,6 @@ pub(crate) async fn get_transitioned_object_reader_with_tier_manager(
     // through a plaintext-coordinate range and skipping the transform is how a
     // transitioned SSE object used to come back as silently corrupt bytes of the
     // right length (rustfs/rustfs#6025).
-    let plan = ReadPlan::build_for_request(rs.clone(), oi, opts, h, resolver)
-        .await
-        .map_err(|err| std::io::Error::other(format!("building the read plan for {bucket}/{object} failed: {err}")))?;
     let (off, length) = (plan.storage_offset() as i64, plan.storage_length());
     let mut gopts = WarmBackendGetOpts::default();
 
@@ -5599,11 +5827,13 @@ mod tests {
     use crate::layout::endpoints::{EndpointServerPools, Endpoints, PoolEndpoints};
     use crate::object_api::{ObjectInfo, ObjectOptions, PutObjReader};
     #[cfg(feature = "test-util")]
+    use crate::services::tier::test_util::MockWarmOp;
+    #[cfg(feature = "test-util")]
     use crate::services::tier::test_util::register_mock_tier;
     #[cfg(feature = "test-util")]
     use crate::services::tier::tier::TierConfigMgr;
     #[cfg(feature = "test-util")]
-    use crate::services::tier::warm_backend::WarmBackend as _;
+    use crate::services::tier::warm_backend::{TransitionCandidateProbe, WarmBackend as _};
     use crate::set_disk::{MultipartCommitBarrier, MultipartCommitPause};
     use crate::set_disk::{RUSTFS_MULTIPART_BUCKET_KEY, RUSTFS_MULTIPART_OBJECT_KEY};
     use crate::storage_api_contracts::namespace::NamespaceLocking as _;
@@ -6299,7 +6529,75 @@ mod tests {
 
     #[cfg(feature = "test-util")]
     #[tokio::test]
-    async fn transitioned_get_rejects_unknown_version_state_before_backend_io() {
+    async fn transitioned_get_allows_legacy_unknown_exact_version_for_non_destructive_read() {
+        let manager = TierConfigMgr::new();
+        let tier = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let backend = register_mock_tier(&manager, &tier).await;
+        let remote_object = format!("remote/{}", Uuid::new_v4());
+        let body = Bytes::from_static(b"legacy transitioned object body");
+        let remote_version = backend
+            .put(
+                &remote_object,
+                ReaderImpl::Body(body.clone()),
+                i64::try_from(body.len()).expect("body length should fit"),
+            )
+            .await
+            .expect("mock remote object should be stored");
+        let mut user_defined = HashMap::new();
+        insert_legacy_transition_version_id(&mut user_defined, &remote_version);
+        let object_info = ObjectInfo {
+            bucket: "bucket".to_string(),
+            name: "object".to_string(),
+            size: i64::try_from(body.len()).expect("body length should fit"),
+            transitioned_object: TransitionedObject {
+                name: remote_object,
+                version_id: remote_version,
+                status: crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE.to_string(),
+                tier: tier.clone(),
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            user_defined: user_defined.into(),
+            ..Default::default()
+        };
+
+        let range = Some(crate::storage_api_contracts::range::HTTPRangeSpec {
+            is_suffix_length: false,
+            start: 7,
+            end: 18,
+        });
+        let mut reader = get_transitioned_object_reader_with_tier_manager(
+            &object_info.bucket,
+            &object_info.name,
+            &range,
+            &HeaderMap::new(),
+            &object_info,
+            &ObjectOptions::default(),
+            &manager,
+            None,
+        )
+        .await
+        .expect("legacy unknown state should still allow a non-destructive read");
+        let mut got = Vec::new();
+        reader
+            .stream
+            .read_to_end(&mut got)
+            .await
+            .expect("transitioned reader should drain");
+
+        assert_eq!(got, &body.as_ref()[7..=18]);
+        assert_eq!(backend.get_count().await, 1);
+        assert_eq!(backend.remove_count().await, 0);
+        assert_eq!(
+            TierConfigMgr::active_operation_lease_count(&manager, &tier).await,
+            0,
+            "tier generation lease should release after EOF"
+        );
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn transitioned_get_rejects_explicit_unknown_version_state_before_backend_io() {
         let manager = TierConfigMgr::new();
         let tier = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
         let backend = register_mock_tier(&manager, &tier).await;
@@ -6315,6 +6613,181 @@ mod tests {
                 ..Default::default()
             },
             transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            user_defined: user_defined_with_transition_version_state(rustfs_filemeta::TransitionVersionState::Unknown).into(),
+            ..Default::default()
+        };
+
+        let err = match get_transitioned_object_reader_with_tier_manager(
+            &object_info.bucket,
+            &object_info.name,
+            &None,
+            &HeaderMap::new(),
+            &object_info,
+            &ObjectOptions::default(),
+            &manager,
+            None,
+        )
+        .await
+        {
+            Ok(_) => panic!("explicit unknown remote version state must fail before backend IO"),
+            Err(err) => err,
+        };
+
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.op_log().await, Vec::<MockWarmOp>::new());
+        assert_eq!(backend.get_count().await, 0);
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn transitioned_get_rejects_present_but_invalid_legacy_version_metadata() {
+        let manager = TierConfigMgr::new();
+        let tier = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let backend = register_mock_tier(&manager, &tier).await;
+
+        for persisted_version in [
+            Uuid::nil().to_string(),
+            "\u{fffd}".to_string(),
+            "bad\u{0001}version".to_string(),
+        ] {
+            let mut user_defined = HashMap::new();
+            insert_legacy_transition_version_id(&mut user_defined, &persisted_version);
+            let object_info = ObjectInfo {
+                bucket: "bucket".to_string(),
+                name: "object".to_string(),
+                size: 1,
+                transitioned_object: TransitionedObject {
+                    name: "remote/object".to_string(),
+                    version_id: String::new(),
+                    status: crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE.to_string(),
+                    tier: tier.clone(),
+                    ..Default::default()
+                },
+                transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+                user_defined: user_defined.into(),
+                ..Default::default()
+            };
+
+            let err = match get_transitioned_object_reader_with_tier_manager(
+                &object_info.bucket,
+                &object_info.name,
+                &None,
+                &HeaderMap::new(),
+                &object_info,
+                &ObjectOptions::default(),
+                &manager,
+                None,
+            )
+            .await
+            {
+                Ok(_) => panic!("present but invalid legacy version metadata must fail before backend IO"),
+                Err(err) => err,
+            };
+
+            assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        }
+
+        assert_eq!(backend.op_log().await, Vec::<MockWarmOp>::new());
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn transitioned_get_probes_legacy_empty_unknown_state_before_unversioned_read() {
+        let manager = TierConfigMgr::new();
+        let tier = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let backend = register_mock_tier(&manager, &tier).await;
+        backend.set_put_remote_version(Some(String::new())).await;
+        let remote_object = format!("remote/{}", Uuid::new_v4());
+        let body = Bytes::from_static(b"legacy unversioned transitioned object body");
+        let remote_version = backend
+            .put(
+                &remote_object,
+                ReaderImpl::Body(body.clone()),
+                i64::try_from(body.len()).expect("body length should fit"),
+            )
+            .await
+            .expect("mock remote object should be stored");
+        assert!(remote_version.is_empty());
+        let object_info = ObjectInfo {
+            bucket: "bucket".to_string(),
+            name: "object".to_string(),
+            size: i64::try_from(body.len()).expect("body length should fit"),
+            transitioned_object: TransitionedObject {
+                name: remote_object.clone(),
+                version_id: String::new(),
+                status: crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE.to_string(),
+                tier: tier.clone(),
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            user_defined: HashMap::from([("x-minio-internal-transitioned-versionID".to_string(), String::new())]).into(),
+            ..Default::default()
+        };
+
+        let mut reader = get_transitioned_object_reader_with_tier_manager(
+            &object_info.bucket,
+            &object_info.name,
+            &None,
+            &HeaderMap::new(),
+            &object_info,
+            &ObjectOptions::default(),
+            &manager,
+            None,
+        )
+        .await
+        .expect("probe-proven legacy unversioned state should allow a non-destructive read");
+        let mut got = Vec::new();
+        reader
+            .stream
+            .read_to_end(&mut got)
+            .await
+            .expect("transitioned reader should drain");
+
+        assert_eq!(got, body.as_ref());
+        assert_eq!(backend.remove_count().await, 0);
+        assert_eq!(
+            backend.op_log().await,
+            vec![
+                MockWarmOp::Put {
+                    object: remote_object.clone()
+                },
+                MockWarmOp::Probe {
+                    object: remote_object.clone()
+                },
+                MockWarmOp::Get { object: remote_object },
+            ]
+        );
+        assert_eq!(
+            TierConfigMgr::active_operation_lease_count(&manager, &tier).await,
+            0,
+            "tier generation lease should release after EOF"
+        );
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn transitioned_get_rejects_ambiguous_empty_unknown_state_without_backend_get() {
+        let manager = TierConfigMgr::new();
+        let tier = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let backend = register_mock_tier(&manager, &tier).await;
+        let remote_object = format!("remote/{}", Uuid::new_v4());
+        backend
+            .set_transition_candidate_probe_override(Some(TransitionCandidateProbe::VersionedPresent(
+                "versioned-candidate".to_string(),
+            )))
+            .await;
+        let object_info = ObjectInfo {
+            bucket: "bucket".to_string(),
+            name: "object".to_string(),
+            size: 1,
+            transitioned_object: TransitionedObject {
+                name: remote_object.clone(),
+                version_id: String::new(),
+                status: crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE.to_string(),
+                tier,
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
             ..Default::default()
         };
 
@@ -6330,19 +6803,28 @@ mod tests {
         )
         .await
         {
-            Ok(_) => panic!("unknown remote version state must fail before backend IO"),
+            Ok(_) => panic!("versioned legacy unknown state without stored version must fail before backend GET"),
             Err(err) => err,
         };
 
         assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.op_log().await, vec![MockWarmOp::Probe { object: remote_object }]);
         assert_eq!(backend.get_count().await, 0);
+        assert_eq!(backend.remove_count().await, 0);
     }
 
     #[cfg(feature = "test-util")]
     #[tokio::test]
-    async fn free_version_delete_rejects_unknown_version_state_before_backend_io() {
+    async fn free_version_delete_rejects_explicit_unknown_before_backend_io() {
         let manager = TierConfigMgr::new();
         let backend = register_mock_tier(&manager, "WARM").await;
+        let identity = test_tier_destination_identity(&manager, "WARM").await;
+        let mut user_defined = user_defined_with_tier_destination_identity(identity);
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut user_defined,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITIONED_VERSION_STATE,
+            rustfs_filemeta::TransitionVersionState::Unknown.as_str().to_string(),
+        );
         let object_info = ObjectInfo {
             transitioned_object: TransitionedObject {
                 name: "remote/object".to_string(),
@@ -6351,17 +6833,251 @@ mod tests {
                 ..Default::default()
             },
             transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            user_defined: user_defined.into(),
             ..Default::default()
         };
 
         let err = super::delete_free_version_remote_object(&object_info, &manager)
             .await
-            .expect_err("unknown remote version state must fail before backend IO");
+            .expect_err("explicit unknown cleanup must fail before backend IO");
 
         assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert!(err.to_string().contains("version state is unknown"));
+        assert_eq!(backend.op_log().await, Vec::<MockWarmOp>::new());
         assert_eq!(backend.remove_count().await, 0);
     }
 
+    #[cfg(feature = "test-util")]
+    async fn test_tier_destination_identity(
+        manager: &Arc<tokio::sync::RwLock<TierConfigMgr>>,
+        tier: &str,
+    ) -> crate::services::tier::tier::TierDestinationId {
+        TierConfigMgr::acquire_operation_lease(manager, tier)
+            .await
+            .expect("test tier lease should be available")
+            .backend_identity()
+    }
+
+    #[cfg(feature = "test-util")]
+    fn user_defined_with_tier_destination_identity(
+        identity: crate::services::tier::tier::TierDestinationId,
+    ) -> HashMap<String, String> {
+        let mut user_defined = HashMap::new();
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut user_defined,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
+            rustfs_utils::crypto::hex(identity),
+        );
+        user_defined
+    }
+
+    #[cfg(feature = "test-util")]
+    fn user_defined_with_transition_version_state(state: rustfs_filemeta::TransitionVersionState) -> HashMap<String, String> {
+        let mut user_defined = HashMap::new();
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut user_defined,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITIONED_VERSION_STATE,
+            state.as_str().to_string(),
+        );
+        user_defined
+    }
+
+    #[cfg(feature = "test-util")]
+    fn insert_legacy_transition_version_id(user_defined: &mut HashMap<String, String>, version_id: &str) {
+        rustfs_utils::http::metadata_compat::insert_str(
+            user_defined,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITIONED_VERSION_ID,
+            version_id.to_string(),
+        );
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn free_version_tuple_rejects_mixed_legacy_missing_and_explicit_unknown() {
+        let manager = TierConfigMgr::new();
+        register_mock_tier(&manager, "WARM").await;
+        let identity = test_tier_destination_identity(&manager, "WARM").await;
+        let mut legacy_metadata = user_defined_with_tier_destination_identity(identity);
+        insert_legacy_transition_version_id(&mut legacy_metadata, "legacy-version");
+        let mut explicit_metadata = legacy_metadata.clone();
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut explicit_metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITIONED_VERSION_STATE,
+            rustfs_filemeta::TransitionVersionState::Unknown.as_str().to_string(),
+        );
+        let make_info = |user_defined: HashMap<String, String>| ObjectInfo {
+            transitioned_object: TransitionedObject {
+                name: "remote/object".to_string(),
+                version_id: "legacy-version".to_string(),
+                tier: "WARM".to_string(),
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            user_defined: user_defined.into(),
+            ..Default::default()
+        };
+
+        let err = super::free_version_remote_tuple_matches(&make_info(legacy_metadata), &make_info(explicit_metadata))
+            .expect_err("mixed legacy-missing and explicit unknown provenance must fail closed");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::WouldBlock);
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn free_version_delete_probes_exact_version_hidden_by_current_delete_marker() {
+        let manager = TierConfigMgr::new();
+        let tier = "WARM";
+        let backend = register_mock_tier(&manager, tier).await;
+        let identity = test_tier_destination_identity(&manager, tier).await;
+        let remote_object = format!("remote/{}", Uuid::new_v4());
+        let body = Bytes::from_static(b"legacy exact cleanup body");
+        let remote_version = backend
+            .put(
+                &remote_object,
+                ReaderImpl::Body(body),
+                i64::try_from(b"legacy exact cleanup body".len()).expect("body length should fit"),
+            )
+            .await
+            .expect("mock remote object should be stored");
+        let mut user_defined = user_defined_with_tier_destination_identity(identity);
+        insert_legacy_transition_version_id(&mut user_defined, &remote_version);
+        backend
+            .set_transition_candidate_probe_override(Some(TransitionCandidateProbe::Missing))
+            .await;
+        assert_eq!(
+            backend
+                .probe_transition_candidate_state(&remote_object)
+                .await
+                .expect("current remote view should be readable"),
+            TransitionCandidateProbe::Missing,
+            "a current delete marker must hide the historical data version from an unversioned probe"
+        );
+        backend.clear_op_log().await;
+        let object_info = ObjectInfo {
+            transitioned_object: TransitionedObject {
+                name: remote_object.clone(),
+                version_id: remote_version,
+                tier: tier.to_string(),
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            user_defined: user_defined.into(),
+            ..Default::default()
+        };
+
+        super::delete_free_version_remote_object(&object_info, &manager)
+            .await
+            .expect("probe-proven legacy exact cleanup should delete the remote version");
+        super::delete_free_version_remote_object(&object_info, &manager)
+            .await
+            .expect("a retry after the exact remote version is already missing should be idempotent");
+
+        assert_eq!(
+            backend.op_log().await,
+            vec![
+                MockWarmOp::Get {
+                    object: remote_object.clone()
+                },
+                MockWarmOp::Remove {
+                    object: remote_object.clone()
+                },
+                MockWarmOp::Get {
+                    object: remote_object.clone()
+                },
+            ]
+        );
+        assert_eq!(
+            backend.remove_versions().await,
+            vec![(remote_object, object_info.transitioned_object.version_id)]
+        );
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn free_version_delete_retains_legacy_unknown_unversioned_object() {
+        let manager = TierConfigMgr::new();
+        let tier = "WARM";
+        let backend = register_mock_tier(&manager, tier).await;
+        backend.set_put_remote_version(Some(String::new())).await;
+        let identity = test_tier_destination_identity(&manager, tier).await;
+        let remote_object = format!("remote/{}", Uuid::new_v4());
+        let body = Bytes::from_static(b"legacy unversioned cleanup body");
+        let remote_version = backend
+            .put(
+                &remote_object,
+                ReaderImpl::Body(body),
+                i64::try_from(b"legacy unversioned cleanup body".len()).expect("body length should fit"),
+            )
+            .await
+            .expect("mock remote object should be stored");
+        assert!(remote_version.is_empty());
+        backend.clear_op_log().await;
+        let mut user_defined = user_defined_with_tier_destination_identity(identity);
+        user_defined.insert("x-minio-internal-transitioned-versionID".to_string(), String::new());
+        let object_info = ObjectInfo {
+            transitioned_object: TransitionedObject {
+                name: remote_object.clone(),
+                version_id: String::new(),
+                tier: tier.to_string(),
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            user_defined: user_defined.into(),
+            ..Default::default()
+        };
+
+        let err = super::delete_free_version_remote_object(&object_info, &manager)
+            .await
+            .expect_err("legacy unversioned cleanup cannot exclude a versioning-state race");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::WouldBlock);
+        assert!(backend.op_log().await.is_empty());
+        assert_eq!(backend.remove_count().await, 0);
+        assert!(backend.remove_versions().await.is_empty());
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn free_version_delete_does_not_remove_a_different_remote_version() {
+        let manager = TierConfigMgr::new();
+        let tier = "WARM";
+        let backend = register_mock_tier(&manager, tier).await;
+        let identity = test_tier_destination_identity(&manager, tier).await;
+        let remote_object = format!("remote/{}", Uuid::new_v4());
+        backend.set_put_remote_version(Some("different-version".to_string())).await;
+        backend
+            .put(
+                &remote_object,
+                ReaderImpl::Body(Bytes::from_static(b"different remote version")),
+                i64::try_from(b"different remote version".len()).expect("body length should fit"),
+            )
+            .await
+            .expect("different remote version should be stored");
+        backend.clear_op_log().await;
+        let mut user_defined = user_defined_with_tier_destination_identity(identity);
+        insert_legacy_transition_version_id(&mut user_defined, "legacy-version");
+        let object_info = ObjectInfo {
+            transitioned_object: TransitionedObject {
+                name: remote_object.clone(),
+                version_id: "legacy-version".to_string(),
+                tier: tier.to_string(),
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            user_defined: user_defined.into(),
+            ..Default::default()
+        };
+
+        super::delete_free_version_remote_object(&object_info, &manager)
+            .await
+            .expect("a missing exact legacy version should be an idempotent cleanup success");
+
+        assert_eq!(backend.op_log().await, vec![MockWarmOp::Get { object: remote_object }]);
+        assert_eq!(backend.remove_count().await, 0);
+        assert!(backend.remove_versions().await.is_empty());
+    }
+
     #[cfg(feature = "test-util")]
     #[tokio::test]
     async fn free_version_remote_delete_requires_persisted_destination_identity() {

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

@@ -1170,6 +1170,7 @@ pub async fn save_manual_transition_job_record_if_current(
         data.clone(),
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_match: Some(current_etag.to_string()),
                 ..Default::default()
@@ -1242,6 +1243,7 @@ pub(crate) async fn save_manual_transition_worker_result_if_absent(
         data,
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_none_match: Some("*".to_string()),
                 ..Default::default()
@@ -1270,6 +1272,7 @@ pub(crate) async fn save_manual_transition_task_if_absent(
         data,
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_none_match: Some("*".to_string()),
                 ..Default::default()
@@ -1621,6 +1624,7 @@ pub async fn save_manual_transition_scope_admission_if_absent(
         data.clone(),
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_none_match: Some("*".to_string()),
                 ..Default::default()
@@ -1672,6 +1676,7 @@ pub async fn save_manual_transition_scope_admission_if_current(
         data.clone(),
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_match: Some(current_etag.to_string()),
                 ..Default::default()

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

@@ -1733,6 +1733,7 @@ async fn save_config_if_none_fenced(
         data.clone(),
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_none_match: Some("*".to_string()),
                 ..Default::default()
@@ -1832,6 +1833,7 @@ async fn save_decommission_manifest_checkpoint_if_match(
 
         let mut opts = ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             no_lock: true,
             http_preconditions: Some(HTTPPreconditions {
                 if_match: Some(observed_etag),
@@ -1960,6 +1962,7 @@ async fn save_config_if_match_fenced(
         data.clone(),
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_match: Some(etag.to_string()),
                 ..Default::default()
@@ -3780,6 +3783,7 @@ where
                 data.clone(),
                 &ObjectOptions {
                     max_parity: true,
+                    write_completion: crate::object_api::WriteCompletion::TailDrained,
                     http_preconditions: Some(HTTPPreconditions {
                         if_none_match: Some("*".to_string()),
                         ..Default::default()
@@ -3869,6 +3873,7 @@ where
                 data.clone(),
                 &ObjectOptions {
                     max_parity: true,
+                    write_completion: crate::object_api::WriteCompletion::TailDrained,
                     http_preconditions: Some(HTTPPreconditions {
                         if_match: Some(etag),
                         ..Default::default()
@@ -3893,6 +3898,7 @@ where
                 data.clone(),
                 &ObjectOptions {
                     max_parity: true,
+                    write_completion: crate::object_api::WriteCompletion::TailDrained,
                     http_preconditions: Some(HTTPPreconditions {
                         if_none_match: Some("*".to_string()),
                         ..Default::default()

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

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use super::runtime_boundary as runtime_sources;
 use crate::bucket::lifecycle::bucket_lifecycle_ops::ExpiryOp;
@@ -72,9 +70,11 @@ static REMOTE_DELETE_BREAKER: LazyLock<Mutex<RemoteDeleteBreaker>> = LazyLock::n
 });
 
 #[cfg(test)]
-static REMOTE_TIER_DELETE_TEST_HOOK: std::sync::LazyLock<
+type RemoteTierDeleteTestHook = Box<dyn Fn(&str, &str, &str) -> std::io::Result<()> + Send + Sync>;
-    std::sync::Mutex<Option<Box<dyn Fn(&str, &str, &str) -> std::io::Result<()> + Send + Sync>>>,
+
-> = std::sync::LazyLock::new(|| std::sync::Mutex::new(None));
+#[cfg(test)]
+static REMOTE_TIER_DELETE_TEST_HOOK: std::sync::LazyLock<std::sync::Mutex<Option<RemoteTierDeleteTestHook>>> =
+    std::sync::LazyLock::new(|| std::sync::Mutex::new(None));
 
 #[derive(Debug)]
 struct RemoteDeleteBreaker {
@@ -107,7 +107,7 @@ impl RemoteDeleteBreaker {
     fn prune(&mut self, now: Instant) {
         while let Some(ts) = self.failures.front().copied() {
             if now.duration_since(ts) > self.window {
-                self.failures.pop_front();
+                let _ = self.failures.pop_front();
             } else {
                 break;
             }
@@ -137,10 +137,10 @@ fn is_signer_header_error(err: &std::io::Error) -> bool {
         return false;
     }
 
-    if let Some(source) = err.get_ref() {
+    if let Some(source) = err.get_ref()
-        if error_chain_contains_signer_header_marker(source) {
+        && error_chain_contains_signer_header_marker(source)
-            return true;
+    {
-        }
+        return true;
     }
 
     let message = err.to_string().to_ascii_lowercase();
@@ -205,7 +205,7 @@ impl ObjSweeper {
 
     #[allow(dead_code, reason = "MinIO-parity surface with no caller in this port (backlog#1823)")]
     pub fn with_version(&mut self, vid: Option<Uuid>) -> &Self {
-        self.version_id = vid.clone();
+        self.version_id = vid;
         self
     }
 
@@ -219,7 +219,7 @@ impl ObjSweeper {
     #[allow(dead_code, reason = "MinIO-parity surface with no caller in this port (backlog#1823)")]
     pub fn get_opts(&self) -> lifecycle::ObjectOpts {
         let mut opts = ObjectOpts {
-            version_id: self.version_id.clone(),
+            version_id: self.version_id,
             versioned: self.versioned,
             version_suspended: self.suspended,
             ..Default::default()
@@ -388,8 +388,8 @@ impl Jentry {
 impl ExpiryOp for Jentry {
     fn op_hash(&self) -> u64 {
         let mut hasher = Sha256::new();
-        hasher.update(format!("{}", self.tier_name).as_bytes());
+        hasher.update(self.tier_name.as_bytes());
-        hasher.update(format!("{}", self.obj_name).as_bytes());
+        hasher.update(self.obj_name.as_bytes());
         xxh64::xxh64(hasher.finalize().as_slice(), XXHASH_SEED)
     }
 
@@ -436,7 +436,7 @@ async fn delete_object_from_remote_tier_raw_with_manager(
     tier_name: &str,
     tier_config_mgr: &Arc<tokio::sync::RwLock<TierConfigMgr>>,
 ) -> Result<(), std::io::Error> {
-    let lease = TierConfigMgr::acquire_operation_lease(&tier_config_mgr, tier_name)
+    let lease = TierConfigMgr::acquire_operation_lease(tier_config_mgr, tier_name)
         .await
         .map_err(std::io::Error::other)?;
     delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, &lease, false, true).await

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

@@ -612,6 +612,7 @@ pub(crate) async fn save_transition_transaction_record(
         data.clone(),
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_none_match: Some("*".to_string()),
                 ..Default::default()
@@ -658,6 +659,7 @@ pub(crate) async fn save_transition_transaction_record_if_current(
         data.clone(),
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_match: Some(etag),
                 ..Default::default()

crates/ecstore/src/bucket/metadata.rs

@@ -477,6 +477,18 @@ impl BucketMetadata {
         !self.table_bucket_config_json.is_empty()
     }
 
+    /// `bucket-targets.json` is stored for this bucket but this build cannot
+    /// decode it.
+    ///
+    /// Keeps "no replication targets configured" and "the target
+    /// configuration cannot be read" apart, the same distinction the
+    /// `fabricated` marker draws for the bucket metadata as a whole. Only
+    /// meaningful after [`Self::parse_all_configs`] has run; readers must fail
+    /// closed on `true` instead of serving an empty target set.
+    pub fn bucket_targets_unreadable(&self) -> bool {
+        !self.bucket_targets_config_json.is_empty() && self.bucket_target_config.is_none()
+    }
+
     /// Parsed per-bucket durability override, if a valid one is stored.
     ///
     /// Absent/empty/unparsable payloads all mean "no override" (the bucket
@@ -964,7 +976,32 @@ impl BucketMetadata {
         Ok(())
     }
 
-    fn parse_all_configs(&mut self) -> Result<()> {
+    /// Decode every stored sub-configuration into its typed field.
+    ///
+    /// A decode failure never fails the whole load: this runs on every bucket
+    /// metadata read, including startup and peer reload, so one bucket's
+    /// corrupt sub-configuration must not make the bucket — or the node —
+    /// unloadable. Instead the failure is *retained*: the raw bytes stay
+    /// untouched and the typed field stays `None`, so `!raw.is_empty() &&
+    /// typed.is_none()` is the durable "exists but cannot be read" signal that
+    /// each accessor keys off. Which accessors must fail closed on it:
+    ///
+    /// | Config | Verdict |
+    /// |---|---|
+    /// | policy | Fails closed: `get_bucket_policy` re-parses the raw JSON and propagates the error; `get_bucket_policy_raw` returns the stored bytes. |
+    /// | object lock | Fails closed in `object_lock_config_state_from_authoritative_metadata`; a retention decision may never be taken on a guess. |
+    /// | versioning | Fails closed in `get_versioning_config`; guessing Unversioned would make delete markers and version ids diverge from what is on disk. |
+    /// | replication | Fails closed in `get_replication_config`. |
+    /// | bucket targets | Fails closed in `get_bucket_targets_config`, and `sync_bucket_target_sys` marks the bucket unreadable in `BucketTargetSys` instead of publishing an empty target set (rustfs/backlog#2282). |
+    /// | encryption | Fails closed in `get_sse_config`: degrading to "no default encryption" stores plaintext objects the operator required to be encrypted. |
+    /// | public access block | Fails closed in `get_public_access_block_config`: degrading grants the anonymous access the operator asked to block. |
+    /// | quota | Fails closed in `get_quota_config`; the enforcement path in `quota::checker` already re-parses the raw JSON and refuses on error. |
+    /// | lifecycle | Safe to degrade: no rules means no expiration and no transition, so nothing is deleted or moved on the strength of an unreadable rule set. The bucket keeps serving reads and writes. |
+    /// | notification | Safe to degrade: events are an outbound side channel; no consumer draws a durability or authorization conclusion from their absence. |
+    /// | tagging | Safe to degrade: bucket tags are cost-allocation labels here; object-level tag conditions come from object metadata, not this blob. |
+    /// | CORS | Safe to degrade: an absent CORS configuration rejects cross-origin browser requests, which is already the restrictive direction. |
+    /// | logging, website, accelerate, request payment, bucket ACL | Safe to degrade: each only shapes an optional response or an optional side channel, and none of them authorizes an action or decides whether data is retained. |
+    pub(super) fn parse_all_configs(&mut self) -> Result<()> {
         if let Err(e) = self.parse_policy_config() {
             tracing::warn!(
                 event = "bucket_metadata_parse_failed",
@@ -1088,20 +1125,26 @@ impl BucketMetadata {
                 "Failed to parse bucket metadata config"
             );
         }
+        // A stored targets blob that cannot be decoded must not collapse into
+        // the empty target set: that is indistinguishable from "no replication
+        // configured", so replication stops and no caller ever sees an error
+        // (rustfs/backlog#2282). Leaving the typed field `None` while the raw
+        // bytes stay non-empty is the retained parse failure every targets
+        // reader keys off; the bytes are preserved so the configuration is
+        // still recoverable.
+        self.bucket_target_config = None;
         if !self.bucket_targets_config_json.is_empty() {
-            if let Err(e) = serde_json::from_slice::<BucketTargets>(&self.bucket_targets_config_json)
+            match serde_json::from_slice::<BucketTargets>(&self.bucket_targets_config_json) {
-                .map(|t| self.bucket_target_config = Some(t))
+                Ok(targets) => self.bucket_target_config = Some(targets),
-            {
+                Err(e) => tracing::error!(
-                tracing::warn!(
                     event = "bucket_metadata_parse_failed",
                     component = "ecstore",
                     subsystem = "bucket_metadata",
                     bucket = %self.name,
                     config = "bucket_targets",
                     error = %e,
-                    "Failed to parse bucket metadata config"
+                    "Bucket replication targets are unreadable; replication for this bucket fails closed"
-                );
+                ),
-                self.bucket_target_config = Some(BucketTargets::default());
             }
         } else {
             self.bucket_target_config = Some(BucketTargets::default());
@@ -1535,6 +1578,117 @@ mod test {
         assert_eq!(bucket_targets.targets[0].target_bucket, "target-bucket");
     }
 
+    /// rustfs/backlog#2282: a stored targets blob this build cannot decode
+    /// must not become the empty target set, and must stay distinguishable
+    /// from a bucket that never configured a target.
+    #[test]
+    fn unreadable_bucket_targets_never_degrade_to_an_empty_target_set() {
+        let truncated = br#"{"targets":[{"endpoint":"s3.example.com","#.to_vec();
+        let mut corrupt = BucketMetadata::new("corrupt-targets");
+        corrupt.bucket_targets_config_json = truncated.clone();
+
+        corrupt
+            .parse_all_configs()
+            .expect("one unreadable sub-config must not fail the whole metadata load");
+
+        assert!(
+            corrupt.bucket_target_config.is_none(),
+            "an undecodable targets blob must not produce a target set at all"
+        );
+        assert!(corrupt.bucket_targets_unreadable());
+        assert_eq!(
+            corrupt.bucket_targets_config_json, truncated,
+            "the raw bytes must survive so the configuration stays recoverable"
+        );
+
+        // The genuinely-absent case is unchanged, and the two now diverge.
+        let mut absent = BucketMetadata::new("no-targets");
+        absent.parse_all_configs().expect("absent targets parse");
+        assert!(
+            absent.bucket_target_config.as_ref().is_some_and(BucketTargets::is_empty),
+            "a bucket that configured no target still reads as an empty target set"
+        );
+        assert!(!absent.bucket_targets_unreadable());
+    }
+
+    /// `Credentials` carries no struct-level `serde(default)`, so one target
+    /// missing `secretKey` is a hard parse error for the whole document. That
+    /// must surface as "unreadable", never as "no targets configured".
+    #[test]
+    fn bucket_targets_missing_secret_key_are_unreadable_not_empty() {
+        let mut bm = BucketMetadata::new("missing-secret-key");
+        bm.bucket_targets_config_json = br#"{"targets":[{"endpoint":"s3.example.com","targetbucket":"remote","arn":"arn:rustfs:replication:us-east-1:src:1","credentials":{"accessKey":"AKIAEXAMPLE"}}]}"#.to_vec();
+
+        bm.parse_all_configs()
+            .expect("a rejected targets document must not fail the whole metadata load");
+
+        assert!(
+            bm.bucket_targets_unreadable(),
+            "a targets document rejected for a missing secretKey is unreadable, not empty"
+        );
+        assert!(bm.bucket_target_config.is_none());
+    }
+
+    /// The invariant every branch of `parse_all_configs` shares: a stored but
+    /// undecodable payload keeps its raw bytes and leaves the typed field
+    /// `None`, so no branch fabricates a value. What a reader may then do with
+    /// that state is decided per config; see the table on `parse_all_configs`.
+    #[test]
+    fn every_config_branch_retains_its_parse_failure_instead_of_defaulting() {
+        let malformed_xml = b"<not-a-valid-document".to_vec();
+        let malformed_json = b"{not-json".to_vec();
+
+        let mut bm = BucketMetadata::new("all-configs-malformed");
+        bm.policy_config_json = malformed_json.clone();
+        bm.quota_config_json = malformed_json.clone();
+        bm.bucket_targets_config_json = malformed_json.clone();
+        bm.notification_config_xml = malformed_xml.clone();
+        bm.lifecycle_config_xml = malformed_xml.clone();
+        bm.object_lock_config_xml = malformed_xml.clone();
+        bm.versioning_config_xml = malformed_xml.clone();
+        bm.encryption_config_xml = malformed_xml.clone();
+        bm.tagging_config_xml = malformed_xml.clone();
+        bm.replication_config_xml = malformed_xml.clone();
+        bm.cors_config_xml = malformed_xml.clone();
+        bm.logging_config_xml = malformed_xml.clone();
+        bm.website_config_xml = malformed_xml.clone();
+        bm.accelerate_config_xml = malformed_xml.clone();
+        bm.request_payment_config_xml = malformed_xml.clone();
+        bm.public_access_block_config_xml = malformed_xml.clone();
+        // `bucket_acl_config_json` is only checked for UTF-8, so only invalid
+        // UTF-8 exercises its failure branch.
+        bm.bucket_acl_config_json = vec![0xff, 0xfe];
+
+        bm.parse_all_configs()
+            .expect("a bucket whose every config is corrupt must still load its metadata");
+
+        let cleared: [(&str, bool); 17] = [
+            ("policy", bm.policy_config.is_none()),
+            ("quota", bm.quota_config.is_none()),
+            ("bucket_targets", bm.bucket_target_config.is_none()),
+            ("notification", bm.notification_config.is_none()),
+            ("lifecycle", bm.lifecycle_config.is_none()),
+            ("object_lock", bm.object_lock_config.is_none()),
+            ("versioning", bm.versioning_config.is_none()),
+            ("encryption", bm.sse_config.is_none()),
+            ("tagging", bm.tagging_config.is_none()),
+            ("replication", bm.replication_config.is_none()),
+            ("cors", bm.cors_config.is_none()),
+            ("logging", bm.logging_config.is_none()),
+            ("website", bm.website_config.is_none()),
+            ("accelerate", bm.accelerate_config.is_none()),
+            ("request_payment", bm.request_payment_config.is_none()),
+            ("public_access_block", bm.public_access_block_config.is_none()),
+            ("bucket_acl", bm.bucket_acl_config.is_none()),
+        ];
+        for (config, is_cleared) in cleared {
+            assert!(is_cleared, "{config}: a corrupt payload must not be replaced by a default");
+        }
+
+        assert_eq!(bm.bucket_targets_config_json, malformed_json, "raw bytes are retained");
+        assert_eq!(bm.lifecycle_config_xml, malformed_xml, "raw bytes are retained");
+    }
+
     #[test]
     fn lifecycle_update_config_clears_parsed_config_on_delete() {
         let mut bm = BucketMetadata::new("test-bucket");

crates/ecstore/src/bucket/metadata_sys.rs

@@ -360,6 +360,16 @@ async fn refresh_buckets_metadata_once(sys: Arc<RwLock<BucketMetadataSys>>) {
 }
 
 async fn sync_bucket_target_sys(bucket: &str, bm: &BucketMetadata) {
+    if bm.bucket_targets_unreadable() {
+        // "The configuration cannot be read" is not "no targets configured".
+        // Publishing an empty snapshot here is what silently stopped
+        // replication (rustfs/backlog#2282): mark the bucket instead, so every
+        // targets reader gets a typed error, and leave any snapshot from an
+        // earlier readable load in place rather than withdrawing it.
+        BucketTargetSys::get().mark_targets_unreadable(bucket).await;
+        return;
+    }
+
     BucketTargetSys::get()
         .update_all_targets(bucket, bm.bucket_target_config.as_ref())
         .await;
@@ -645,6 +655,12 @@ pub struct BucketMetadataMutationGuard {
 }
 
 impl BucketMetadataMutationGuard {
+    /// Returns the storage-verified identity while both incarnation fences remain valid.
+    pub fn checked_bucket_incarnation(&self) -> Result<(&str, Uuid)> {
+        self.ensure_valid(&self.bucket)?;
+        Ok((&self.bucket, self.incarnation_id))
+    }
+
     fn ensure_valid(&self, bucket: &str) -> Result<()> {
         if self.bucket != bucket {
             return Err(Error::other("bucket metadata mutation guard does not match bucket"));
@@ -664,6 +680,29 @@ async fn acquire_config_write_guard_for_incarnation(
     sys: Arc<RwLock<BucketMetadataSys>>,
     bucket: &str,
     expected_incarnation_id: Option<Uuid>,
+) -> Result<BucketMetadataMutationGuard> {
+    acquire_config_write_guard_with_migration(sys, bucket, expected_incarnation_id, true).await
+}
+
+/// Scanner probes must not create an incarnation to make a capability available.
+pub async fn acquire_scanner_bucket_incarnation_fence(
+    bucket: &str,
+    expected_incarnation_id: Uuid,
+    expected_owner_id: Uuid,
+) -> Result<BucketMetadataMutationGuard> {
+    super::utils::check_valid_bucket_name(bucket)?;
+    let sys = get_bucket_metadata_sys()?;
+    if expected_owner_id.is_nil() || sys.read().await.api.id != expected_owner_id || expected_incarnation_id.is_nil() {
+        return Err(Error::other("scanner bucket incarnation owner does not match"));
+    }
+    acquire_config_write_guard_with_migration(sys, bucket, Some(expected_incarnation_id), false).await
+}
+
+async fn acquire_config_write_guard_with_migration(
+    sys: Arc<RwLock<BucketMetadataSys>>,
+    bucket: &str,
+    expected_incarnation_id: Option<Uuid>,
+    migrate: bool,
 ) -> Result<BucketMetadataMutationGuard> {
     let metadata_sys = sys.read().await.clone();
     let lifecycle_guard = metadata_sys.api.acquire_bucket_lifecycle_read_lock(bucket).await?;
@@ -671,13 +710,15 @@ async fn acquire_config_write_guard_for_incarnation(
     // Legacy buckets are migrated while the lifecycle fence prevents a
     // same-name replacement. The second read under the write transaction is
     // the CAS source of truth for the actual rewrite.
-    await_bucket_namespace_operation(
+    if migrate {
-        Some(&lifecycle_guard),
+        await_bucket_namespace_operation(
-        bucket,
+            Some(&lifecycle_guard),
-        "bucket config incarnation migration",
+            bucket,
-        metadata_sys.get_bucket_incarnation_id(bucket),
+            "bucket config incarnation migration",
-    )
+            metadata_sys.get_bucket_incarnation_id(bucket),
-    .await?;
+        )
+        .await?;
+    }
     let transaction_guard = await_bucket_namespace_operation(
         Some(&lifecycle_guard),
         bucket,
@@ -2118,7 +2159,9 @@ impl BucketMetadataSys {
     pub async fn get_public_access_block_config(&self, bucket: &str) -> Result<(PublicAccessBlockConfiguration, OffsetDateTime)> {
         let (bm, _) = self.get_config(bucket).await?;
 
-        if let Some(config) = &bm.public_access_block_config {
+        if !bm.public_access_block_config_xml.is_empty() && bm.public_access_block_config.is_none() {
+            Err(Error::other("persisted bucket public access block configuration is invalid"))
+        } else if let Some(config) = &bm.public_access_block_config {
             Ok((config.clone(), bm.public_access_block_config_updated_at))
         } else {
             Err(Error::ConfigNotFound)
@@ -2429,7 +2472,9 @@ impl BucketMetadataSys {
     pub async fn get_sse_config(&self, bucket: &str) -> Result<(ServerSideEncryptionConfiguration, OffsetDateTime)> {
         let (bm, _) = self.get_config(bucket).await?;
 
-        if let Some(config) = &bm.sse_config {
+        if !bm.encryption_config_xml.is_empty() && bm.sse_config.is_none() {
+            Err(Error::other("persisted bucket encryption configuration is invalid"))
+        } else if let Some(config) = &bm.sse_config {
             Ok((config.clone(), bm.encryption_config_updated_at))
         } else {
             Err(Error::ConfigNotFound)
@@ -2500,7 +2545,9 @@ impl BucketMetadataSys {
     pub async fn get_quota_config(&self, bucket: &str) -> Result<(BucketQuota, OffsetDateTime)> {
         let (bm, _) = self.get_config(bucket).await?;
 
-        if let Some(config) = &bm.quota_config {
+        if !bm.quota_config_json.is_empty() && bm.quota_config.is_none() {
+            Err(Error::other("persisted bucket quota configuration is invalid"))
+        } else if let Some(config) = &bm.quota_config {
             Ok((config.clone(), bm.quota_config_updated_at))
         } else {
             Err(Error::ConfigNotFound)
@@ -2522,7 +2569,9 @@ impl BucketMetadataSys {
     pub async fn get_bucket_targets_config(&self, bucket: &str) -> Result<BucketTargets> {
         let (bm, _) = self.get_config(bucket).await?;
 
-        if let Some(config) = &bm.bucket_target_config {
+        if bm.bucket_targets_unreadable() {
+            Err(Error::other("persisted bucket replication target configuration is invalid"))
+        } else if let Some(config) = &bm.bucket_target_config {
             Ok(config.clone())
         } else {
             Err(Error::ConfigNotFound)
@@ -2593,6 +2642,7 @@ pub(crate) mod test_support {
 mod tests {
     use super::test_support::isolated_store_over_temp_disks;
     use super::*;
+    use crate::bucket::bucket_target_sys::BucketTargetError;
     use crate::bucket::metadata::{
         BUCKET_ACCELERATE_CONFIG, BUCKET_CORS_CONFIG, BUCKET_LIFECYCLE_CONFIG, BUCKET_LOGGING_CONFIG, BUCKET_NOTIFICATION_CONFIG,
         BUCKET_POLICY_CONFIG, BUCKET_PUBLIC_ACCESS_BLOCK_CONFIG, BUCKET_REPLICATION_CONFIG, BUCKET_REQUEST_PAYMENT_CONFIG,
@@ -2788,6 +2838,36 @@ mod tests {
         );
     }
 
+    /// The `parse_all_configs` audit (rustfs/backlog#2282): every accessor
+    /// whose configuration grants something — plaintext storage, anonymous
+    /// access, capacity, replication targets — reports a corrupt payload as
+    /// invalid rather than as absent, because "absent" is what grants it.
+    #[tokio::test]
+    async fn malformed_permissive_configs_are_not_reported_as_absent() {
+        let (_dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let sys = BucketMetadataSys::new(ecstore);
+        let bucket = "malformed-permissive-config";
+        let mut metadata = BucketMetadata::new(bucket);
+        metadata.encryption_config_xml = b"<ServerSideEncryptionConfiguration".to_vec();
+        metadata.public_access_block_config_xml = b"<PublicAccessBlockConfiguration".to_vec();
+        metadata.quota_config_json = b"{not-json".to_vec();
+        metadata.bucket_targets_config_json = b"{not-json".to_vec();
+        metadata
+            .parse_all_configs()
+            .expect("a corrupt sub-config must not fail the load");
+        sys.set(bucket.to_string(), Arc::new(metadata)).await;
+
+        for (config, result) in [
+            ("encryption", sys.get_sse_config(bucket).await.err()),
+            ("public access block", sys.get_public_access_block_config(bucket).await.err()),
+            ("quota", sys.get_quota_config(bucket).await.err()),
+            ("bucket targets", sys.get_bucket_targets_config(bucket).await.err()),
+        ] {
+            let err = result.unwrap_or_else(|| panic!("malformed {config} metadata must not read as a value"));
+            assert_ne!(err, Error::ConfigNotFound, "malformed {config} metadata must not be reported as absent");
+        }
+    }
+
     #[tokio::test]
     async fn config_states_distinguish_authoritative_absence_from_fabricated_metadata() {
         use std::sync::atomic::Ordering;
@@ -3127,6 +3207,82 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn scoped_dirty_usage_incarnation_probe_does_not_migrate_legacy_metadata() {
+        let (dirs, store) = isolated_store_over_temp_disks().await;
+        let sys = Arc::new(RwLock::new(BucketMetadataSys::new(store.clone())));
+        let bucket = "scoped-ack-legacy";
+        for dir in &dirs {
+            std::fs::create_dir_all(dir.path().join(bucket)).expect("create legacy bucket");
+        }
+        let mut metadata = BucketMetadata::new(bucket);
+        metadata.bucket_incarnation_id = Uuid::nil();
+        sys.read()
+            .await
+            .persist_and_set(metadata)
+            .await
+            .expect("persist legacy metadata");
+        assert!(
+            acquire_config_write_guard_with_migration(sys.clone(), bucket, Some(Uuid::new_v4()), false)
+                .await
+                .is_err()
+        );
+        assert!(load_bucket_incarnation(store, bucket).await.expect("read sidecar").is_none());
+        assert!(
+            sys.read()
+                .await
+                .get_config_from_disk(bucket)
+                .await
+                .expect("read metadata")
+                .bucket_incarnation_id
+                .is_nil()
+        );
+    }
+
+    #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+    #[serial]
+    async fn scoped_dirty_usage_incarnation_rejects_deleted_and_recreated_bucket() {
+        let (_dirs, store) = isolated_store_over_temp_disks().await;
+        init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let sys = bucket_metadata_sys_of(&store.ctx).expect("metadata owner");
+        let bucket = "scoped-ack-recreated";
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create bucket");
+        let old = store.bucket_incarnation_id_from_disk(bucket).await.expect("old incarnation");
+        let guard = acquire_config_write_guard_with_migration(sys.clone(), bucket, Some(old), false)
+            .await
+            .expect("trusted incarnation fence");
+        assert_eq!(guard.checked_bucket_incarnation().expect("valid fences"), (bucket, old));
+        drop(guard);
+        store
+            .delete_bucket(bucket, &DeleteBucketOptions::default())
+            .await
+            .expect("delete bucket");
+        assert!(
+            acquire_config_write_guard_with_migration(sys.clone(), bucket, Some(old), false)
+                .await
+                .is_err()
+        );
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("recreate bucket");
+        let new = store.bucket_incarnation_id_from_disk(bucket).await.expect("new incarnation");
+        assert_ne!(old, new);
+        assert!(
+            acquire_config_write_guard_with_migration(sys.clone(), bucket, Some(old), false)
+                .await
+                .is_err()
+        );
+        assert!(
+            acquire_config_write_guard_with_migration(sys, bucket, Some(new), false)
+                .await
+                .is_ok()
+        );
+    }
+
     #[tokio::test]
     async fn old_node_metadata_rewrite_cannot_replace_bucket_incarnation_sidecar() {
         let (dirs, ecstore) = isolated_store_over_temp_disks().await;
@@ -4066,6 +4222,114 @@ mod tests {
         target_sys.delete(bucket).await;
     }
 
+    /// rustfs/backlog#2282: an unreadable `bucket-targets.json` reaches every
+    /// targets reader as a typed error; it neither withdraws a snapshot a
+    /// previous readable load published, nor collapses into the "no targets
+    /// configured" state that a bucket with an absent configuration reports.
+    #[tokio::test]
+    #[serial]
+    async fn unreadable_bucket_targets_fail_closed_and_stay_distinct_from_absent() {
+        let (_dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let sys = BucketMetadataSys::new(ecstore);
+        let target_sys = BucketTargetSys::get();
+        let unreadable = "targets-unreadable";
+        let absent = "targets-absent";
+        target_sys.delete(unreadable).await;
+        target_sys.delete(absent).await;
+
+        // A readable load publishes this bucket's targets.
+        let mut readable = BucketMetadata::new(unreadable);
+        readable.bucket_target_config = Some(BucketTargets {
+            targets: vec![target(unreadable, "live")],
+        });
+        sync_bucket_target_sys(unreadable, &readable).await;
+        assert_eq!(
+            target_sys
+                .list_bucket_targets(unreadable)
+                .await
+                .expect("readable targets publish")
+                .targets
+                .len(),
+            1
+        );
+
+        // The same bucket reloaded with a blob that cannot be decoded.
+        let mut corrupt = BucketMetadata::new(unreadable);
+        corrupt.bucket_targets_config_json = br#"{"targets":[{"endpoint":"#.to_vec();
+        corrupt
+            .parse_all_configs()
+            .expect("an unreadable targets blob must not fail the metadata load");
+        sys.set(unreadable.to_string(), Arc::new(corrupt)).await;
+
+        assert!(
+            matches!(
+                target_sys.list_bucket_targets(unreadable).await,
+                Err(BucketTargetError::BucketRemoteTargetsUnreadable { .. })
+            ),
+            "an unreadable configuration must not read as an empty or a missing target set"
+        );
+        assert!(
+            target_sys.list_targets(unreadable, "").await.is_err(),
+            "the admin listing must surface the fault instead of an empty list"
+        );
+        let err = sys
+            .get_bucket_targets_config(unreadable)
+            .await
+            .expect_err("an unreadable targets configuration must not read as a value");
+        assert_ne!(err, Error::ConfigNotFound, "unreadable must not be reported as absent");
+
+        // A bucket that never configured a target keeps its previous behavior.
+        let mut no_targets = BucketMetadata::new(absent);
+        no_targets.parse_all_configs().expect("absent targets parse");
+        sys.set(absent.to_string(), Arc::new(no_targets)).await;
+        assert!(
+            matches!(
+                target_sys.list_bucket_targets(absent).await,
+                Err(BucketTargetError::BucketRemoteTargetNotFound { .. })
+            ),
+            "an absent configuration must still report as a missing target set"
+        );
+        assert!(
+            target_sys
+                .list_targets(absent, "")
+                .await
+                .expect("an absent configuration lists no targets")
+                .is_empty()
+        );
+        assert!(
+            sys.get_bucket_targets_config(absent)
+                .await
+                .expect("an absent targets configuration still reads as an empty set")
+                .is_empty(),
+            "the absent path must keep returning an empty target set, exactly as before"
+        );
+
+        // One bucket's unreadable configuration does not reach another bucket.
+        assert!(!matches!(
+            target_sys.list_bucket_targets(absent).await,
+            Err(BucketTargetError::BucketRemoteTargetsUnreadable { .. })
+        ));
+
+        // A repaired configuration takes effect on the next load, no restart.
+        let mut repaired = BucketMetadata::new(unreadable);
+        repaired.bucket_target_config = Some(BucketTargets {
+            targets: vec![target(unreadable, "repaired")],
+        });
+        sync_bucket_target_sys(unreadable, &repaired).await;
+        assert_eq!(
+            target_sys
+                .list_bucket_targets(unreadable)
+                .await
+                .expect("a repaired configuration clears the unreadable marker")
+                .targets
+                .len(),
+            1
+        );
+
+        target_sys.delete(unreadable).await;
+        target_sys.delete(absent).await;
+    }
+
     #[tokio::test]
     #[serial]
     async fn metadata_reload_clears_stale_bucket_targets_when_config_is_removed() {

crates/ecstore/src/bucket/on_demand_migration/backend_contract.rs

@@ -0,0 +1,172 @@
+// 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.
+
+//! One contract every [`SourceBackend`] implementation must satisfy.
+//!
+//! The migration pipeline talks to a source only through the trait, so a new
+//! provider is correct exactly when it answers the same questions the same way:
+//! the same head fields, the same range semantics, the same page shape, the
+//! same error classes. Each backend supplies a fixture that answers this fixed
+//! corpus in its own dialect and then runs [`assert_backend_contract`], so a
+//! provider-specific mapping bug shows up as a contract failure rather than as
+//! a surprise in the pull pipeline.
+//!
+//! Backends differ in two documented ways, declared through
+//! [`BackendCapabilities`]: whether the provider's ETag is a content digest,
+//! and whether the provider can resume a listing from a key.
+
+use super::source_client::{SourceBackend, SourceError, SourceListRequest};
+use crate::storage_api_contracts::range::HTTPRangeSpec;
+use std::collections::HashMap;
+
+/// The single object every fixture serves.
+pub(super) const OBJECT_KEY: &str = "dir/a.txt";
+pub(super) const OBJECT_BODY: &[u8] = b"hello";
+/// MD5 of [`OBJECT_BODY`]; the ETag of the object on a digest provider.
+pub(super) const OBJECT_MD5: &str = "5d41402abc4b2a76b9719d911017c592";
+/// The second key the fixture's listing returns, on its second page.
+pub(super) const SECOND_KEY: &str = "dir/b.txt";
+pub(super) const COMMON_PREFIX: &str = "dir/sub/";
+pub(super) const LIST_CURSOR: &str = "cursor-1";
+/// A key the fixture answers with the provider's "no such object".
+pub(super) const MISSING_KEY: &str = "missing";
+/// A key the fixture answers with the provider's "not authorized".
+pub(super) const FORBIDDEN_KEY: &str = "secret";
+
+/// Where backends are allowed to differ.
+#[derive(Clone, Copy, Debug)]
+pub(super) struct BackendCapabilities {
+    /// The provider's ETag is an opaque token, not a digest of the bytes.
+    pub(super) etag_is_opaque: bool,
+    /// The provider can resume a listing from a key rather than only from an
+    /// opaque cursor.
+    pub(super) supports_start_after: bool,
+    /// The provider has an object-tagging concept at all. GCS does not, and
+    /// answers with an empty map instead of failing a pull.
+    pub(super) supports_tagging: bool,
+}
+
+/// Drives `backend` through the shared corpus. Fixtures are scripted in
+/// request order, so the call order here is part of the contract.
+pub(super) async fn assert_backend_contract(backend: &dyn SourceBackend, caps: BackendCapabilities) {
+    // 1. HEAD maps the object's shared fields.
+    let head = backend.head(OBJECT_KEY).await.expect("HEAD of the fixture object");
+    assert_eq!(head.size, OBJECT_BODY.len() as u64, "HEAD reports the object size");
+    assert_eq!(head.content_type.as_deref(), Some("text/plain"));
+    assert_eq!(
+        head.user_metadata,
+        HashMap::from([("owner".to_string(), "alice".to_string())]),
+        "user metadata is keyed without the provider prefix"
+    );
+    assert!(head.storage_class.is_some(), "the provider's tier is recorded");
+    assert!(head.last_modified.is_some(), "the provider's timestamp is parsed");
+    assert!(head.sse.is_none(), "the fixture object is not server-side encrypted");
+    assert!(!head.is_multipart_etag);
+    assert_eq!(head.etag_is_opaque, caps.etag_is_opaque);
+    match caps.etag_is_opaque {
+        false => assert_eq!(head.etag.as_deref(), Some(OBJECT_MD5), "a digest ETag is mapped verbatim"),
+        true => assert!(head.etag.is_some(), "an opaque ETag is still recorded"),
+    }
+
+    // 2. An unranged GET streams the whole object and reports no range.
+    let got = backend.get(OBJECT_KEY, None).await.expect("unranged GET");
+    assert_eq!(got.head.size, OBJECT_BODY.len() as u64);
+    assert!(got.content_range.is_none(), "an unranged GET has no content-range");
+    assert_eq!(got.head.etag_is_opaque, caps.etag_is_opaque, "GET and HEAD agree about the ETag");
+    let body = got.body.collect().await.expect("body streams").into_bytes();
+    assert_eq!(body.as_ref(), OBJECT_BODY);
+
+    // 3. A ranged GET returns exactly the requested interval, and `size` is
+    //    the length of the returned bytes rather than of the object.
+    let range = HTTPRangeSpec {
+        is_suffix_length: false,
+        start: 1,
+        end: 3,
+    };
+    let got = backend.get(OBJECT_KEY, Some(&range)).await.expect("ranged GET");
+    assert_eq!(got.head.size, 3, "a ranged GET reports the range length");
+    assert_eq!(got.content_range.as_deref(), Some("bytes 1-3/5"));
+    let body = got.body.collect().await.expect("body streams").into_bytes();
+    assert_eq!(body.as_ref(), &OBJECT_BODY[1..=3]);
+
+    // 4. A delimiter listing rolls prefixes up and hands back a cursor.
+    let page = backend
+        .list(&SourceListRequest {
+            prefix: Some("dir/"),
+            delimiter: Some("/"),
+            max_keys: 2,
+            ..Default::default()
+        })
+        .await
+        .expect("first listing page");
+    assert_eq!(page.objects.len(), 1, "the first page holds one object");
+    assert_eq!(page.objects[0].key, OBJECT_KEY, "listing keys are in the source namespace");
+    assert_eq!(page.objects[0].size, OBJECT_BODY.len() as u64);
+    assert!(page.objects[0].last_modified.is_some());
+    assert_eq!(page.common_prefixes, vec![COMMON_PREFIX.to_string()]);
+    assert!(page.is_truncated);
+    assert_eq!(page.next_continuation_token.as_deref(), Some(LIST_CURSOR));
+
+    // 5. The cursor is passed back verbatim and the last page ends the walk.
+    let page = backend
+        .list(&SourceListRequest {
+            prefix: Some("dir/"),
+            delimiter: Some("/"),
+            continuation_token: Some(LIST_CURSOR),
+            max_keys: 2,
+            ..Default::default()
+        })
+        .await
+        .expect("second listing page");
+    assert_eq!(page.objects.len(), 1);
+    assert_eq!(page.objects[0].key, SECOND_KEY);
+    assert!(!page.is_truncated);
+    assert!(page.next_continuation_token.is_none(), "a complete listing carries no cursor");
+
+    // 6. Tags come back as a flat map, empty on a provider without tags.
+    let tags = backend.tagging(OBJECT_KEY).await.expect("object tags");
+    match caps.supports_tagging {
+        true => assert_eq!(tags, HashMap::from([("env".to_string(), "prod".to_string())])),
+        false => assert!(tags.is_empty(), "a provider without tags reports none: {tags:?}"),
+    }
+
+    // 7. The probe confirms the bucket or container answers.
+    backend.probe().await.expect("probe of the fixture bucket");
+
+    // 8. A missing object is `NotFound`, and never retried.
+    let err = backend.head(MISSING_KEY).await.expect_err("a missing object must fail");
+    assert!(matches!(err, SourceError::NotFound), "{err:?}");
+    assert_eq!(err.class_label(), "not_found");
+    assert!(!err.is_retryable());
+
+    // 9. A denied object is `AccessDenied`, and never retried.
+    let err = backend.head(FORBIDDEN_KEY).await.expect_err("a denied object must fail");
+    assert!(matches!(err, SourceError::AccessDenied), "{err:?}");
+    assert_eq!(err.class_label(), "access_denied");
+    assert!(!err.is_retryable());
+
+    // 10. A provider without a key cursor must refuse one instead of listing
+    //     from the wrong position. This issues no request either way.
+    if !caps.supports_start_after {
+        let err = backend
+            .list(&SourceListRequest {
+                start_after: Some(OBJECT_KEY),
+                max_keys: 1,
+                ..Default::default()
+            })
+            .await
+            .expect_err("a backend without a key cursor must refuse start_after");
+        assert!(matches!(err, SourceError::Unsupported(_)), "{err:?}");
+    }
+}

crates/ecstore/src/bucket/on_demand_migration/backfill.rs

@@ -25,8 +25,8 @@
 //!   [`BACKFILL_SAVE_INTERVAL`], and at every page end, with an `If-Match`
 //!   compare-and-set so a concurrent cancel or takeover is never overwritten.
 //! - The `continuation_token` only advances once every pull queued from the
-//!   page before it has reported back, so a crash re-lists at most one page
+//!   page before it has succeeded. After a failure it stays at that page,
-//!   (already-present keys are then skipped, never re-pulled).
+//!   so crash recovery cannot skip failed pulls (existing keys are skipped).
 //! - The owner holds a lease of [`BACKFILL_LEASE`] renewed by every save. The
 //!   recovery loop ([`run_backfill_recovery_loop`]) scans the buckets this
 //!   node has an ODM state for every [`BACKFILL_RECOVERY_INTERVAL`] and takes
@@ -367,9 +367,8 @@ pub struct LocalBackfillObject {
     pub source_etag: Option<String>,
 }
 
-/// Receiver of one queued pull's report; `None` when the pull was coalesced
+/// Shared report of a new or coalesced pull; absent only when not admitted.
-/// into one already running.
+pub type PullReport = Option<super::pull::QueuedPullReport>;
-pub type PullReport = Option<oneshot::Receiver<QueuedPullOutcome>>;
 
 /// Everything the job needs from its bucket, so the loop can run against a
 /// mock in unit tests. Production: [`BucketBackfillContext`].
@@ -684,6 +683,7 @@ async fn write_checkpoint(
     };
     let opts = ObjectOptions {
         max_parity: true,
+        write_completion: crate::object_api::WriteCompletion::TailDrained,
         http_preconditions: Some(preconditions),
         ..Default::default()
     };
@@ -1190,9 +1190,11 @@ impl Job {
     }
 
     async fn main_loop(&mut self) -> Result<(), Stop> {
+        let mut cursor = self.checkpoint.continuation_token.clone();
+        let failed_at_resume = self.checkpoint.failed;
         loop {
             self.check_cancel()?;
-            let page = self.list_page().await?;
+            let page = self.list_page(cursor.as_deref()).await?;
             for object in &page.objects {
                 self.check_cancel()?;
                 self.checkpoint.listed += 1;
@@ -1204,10 +1206,13 @@ impl Job {
                 self.drain_ready();
                 self.tick(false).await?;
             }
-            // Only advance the cursor once every pull of this page reported
+            // A persisted cursor certifies successful work, not just listing
-            // back, so a takeover re-lists at most this page.
+            // progress. Keep it at the first failed page for crash recovery.
             self.drain_all().await?;
-            self.checkpoint.continuation_token = page.next_continuation_token.clone();
+            cursor = page.next_continuation_token;
+            if self.checkpoint.failed == failed_at_resume {
+                self.checkpoint.continuation_token = cursor.clone();
+            }
             self.tick(true).await?;
             if !page.is_truncated {
                 return Ok(());
@@ -1222,7 +1227,7 @@ impl Job {
         }
     }
 
-    async fn list_page(&mut self) -> Result<SourcePage, Stop> {
+    async fn list_page(&mut self, cursor: Option<&str>) -> Result<SourcePage, Stop> {
         let mut attempt = 0;
         loop {
             while !self.context.source_available() {
@@ -1230,7 +1235,7 @@ impl Job {
                 self.tick(false).await?;
             }
             let prefix = self.checkpoint.prefix.clone();
-            let token = self.checkpoint.continuation_token.clone();
+            let token = cursor.map(str::to_string);
             match self
                 .context
                 .list_page(prefix.as_deref(), token.as_deref(), BACKFILL_LIST_PAGE_SIZE)
@@ -1304,9 +1309,10 @@ impl Job {
         }
         loop {
             match self.context.enqueue(key) {
-                (EnqueueOutcome::Enqueued, report) => {
+                (EnqueueOutcome::Enqueued | EnqueueOutcome::Coalesced, report) => {
                     self.checkpoint.enqueued += 1;
-                    if let Some(rx) = report {
+                    let rx = report.ok_or(Stop::Unavailable)?;
+                    {
                         let key = key.to_string();
                         self.outstanding.push(Box::pin(async move { (key, rx.await) }));
                     }
@@ -1321,11 +1327,6 @@ impl Job {
                     );
                     return Ok(());
                 }
-                (EnqueueOutcome::Coalesced, _) => {
-                    // Someone else pulls it; its result is not ours to count.
-                    self.checkpoint.enqueued += 1;
-                    return Ok(());
-                }
                 (EnqueueOutcome::QueueFull, _) => {
                     // Wait, never drop: one completion frees a slot.
                     if self.outstanding.is_empty() {
@@ -1639,6 +1640,7 @@ mod tests {
         queue_capacity: usize,
         pending: Mutex<Vec<(String, oneshot::Sender<QueuedPullOutcome>)>>,
         fail_keys: HashSet<String>,
+        coalesced: bool,
         auto_complete: AtomicBool,
         cancel: CancellationToken,
         config_updated_at: Mutex<Option<OffsetDateTime>>,
@@ -1666,6 +1668,7 @@ mod tests {
                 queue_capacity: usize::MAX,
                 pending: Mutex::new(Vec::new()),
                 fail_keys: HashSet::new(),
+                coalesced: false,
                 auto_complete: AtomicBool::new(true),
                 cancel: CancellationToken::new(),
                 config_updated_at: Mutex::new(Some(ts(1_700_000_000))),
@@ -1745,7 +1748,12 @@ mod tests {
             } else {
                 self.pending.lock().push((key.to_string(), tx));
             }
-            (EnqueueOutcome::Enqueued, Some(rx))
+            let outcome = if self.coalesced {
+                EnqueueOutcome::Coalesced
+            } else {
+                EnqueueOutcome::Enqueued
+            };
+            (outcome, Some(futures::FutureExt::shared(rx)))
         }
 
         fn cancel_token(&self) -> CancellationToken {
@@ -1911,7 +1919,7 @@ mod tests {
     #[tokio::test]
     async fn failed_pulls_are_counted_hashed_and_finish_with_failures() {
         let bucket = "backfill-failed";
-        let mut context = MockContext::new(5, 1000);
+        let mut context = MockContext::new(5, 2);
         Arc::get_mut(&mut context)
             .expect("unshared")
             .fail_keys
@@ -1926,12 +1934,52 @@ mod tests {
             .checkpoint;
         assert_eq!(cp.state, BackfillState::CompletedWithFailures);
         assert_eq!((cp.pulled, cp.failed), (4, 1));
+        assert_eq!(cp.continuation_token.as_deref(), Some("2"), "retain the first failed page for recovery");
         assert_eq!(cp.failed_keys, vec![key_hash("k/00002")]);
         let last = cp.last_error.expect("last error");
         assert_eq!(last.class, "local_write");
         assert_eq!(last.key_hash.as_deref(), Some(key_hash("k/00002").as_str()));
     }
 
+    #[tokio::test]
+    async fn coalesced_pulls_block_the_checkpoint_and_report_failures() {
+        let bucket = "backfill-coalesced";
+        let mut context = MockContext::new(1, 1);
+        {
+            let ctx = Arc::get_mut(&mut context).expect("unshared");
+            ctx.coalesced = true;
+            ctx.auto_complete = AtomicBool::new(false);
+            ctx.fail_keys.insert("k/00000".to_string());
+        }
+        let (_dirs, store, runner) = runner_with("node-a", bucket, Arc::clone(&context)).await;
+        runner.start(bucket, BackfillRequest::default()).await.expect("start");
+        tokio::time::timeout(Duration::from_secs(10), async {
+            while context.pending.lock().is_empty() {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("job enqueued");
+        assert!(runner.is_running_locally(bucket), "coalescing is not completion");
+        let cp = read_checkpoint(&store, bucket)
+            .await
+            .expect("read")
+            .expect("checkpoint")
+            .checkpoint;
+        assert!(cp.state.is_active());
+        assert!(cp.continuation_token.is_none());
+        context.complete_pending();
+        runner.wait_until_idle(bucket).await;
+        let cp = read_checkpoint(&store, bucket)
+            .await
+            .expect("read")
+            .expect("checkpoint")
+            .checkpoint;
+        assert_eq!(cp.state, BackfillState::CompletedWithFailures);
+        assert_eq!((cp.enqueued, cp.pulled, cp.failed), (1, 0, 1));
+        assert_eq!(cp.failed_keys, vec![key_hash("k/00000")]);
+    }
+
     #[tokio::test]
     async fn listing_failure_marks_the_job_failed_with_the_error_class() {
         let bucket = "backfill-list-error";
@@ -2144,6 +2192,68 @@ mod tests {
         assert_eq!(runner.recover_once().await.taken_over, 0, "a finished job is not recovered");
     }
 
+    #[tokio::test]
+    async fn recovery_advances_past_historical_failures_but_pins_new_failures() {
+        let bucket = "backfill-takeover-failed";
+        let mut context = MockContext::new(8, 2);
+        {
+            let ctx = Arc::get_mut(&mut context).expect("unshared");
+            ctx.auto_complete = AtomicBool::new(false);
+            ctx.fail_keys.insert("k/00004".to_string());
+        }
+        let (_dirs, store, runner) = runner_with("node-b", bucket, Arc::clone(&context)).await;
+        let crashed_at = OffsetDateTime::now_utc() - Duration::from_secs(300);
+        let mut crashed = BackfillCheckpoint::new(&BackfillRequest::default(), ts(1_700_000_000), "node-a", crashed_at);
+        crashed.continuation_token = Some("2".to_string());
+        crashed.failed = 1;
+        crashed.record_failure("local_write", Some("k/00002"), crashed_at);
+        write_checkpoint(&store, bucket, &crashed, None)
+            .await
+            .expect("seed failed page with an expired lease");
+
+        assert_eq!(runner.recover_once().await.taken_over, 1);
+        for (page_start, durable_token, failures) in [(2, "2", 1), (4, "4", 1), (6, "4", 2)] {
+            tokio::time::timeout(Duration::from_secs(10), async {
+                loop {
+                    if context.pending.lock().len() == 2 {
+                        break;
+                    }
+                    tokio::task::yield_now().await;
+                }
+            })
+            .await
+            .expect("resumed page enqueued before its reports complete");
+            assert_eq!(
+                context.pending.lock().iter().map(|(key, _)| key.clone()).collect::<Vec<_>>(),
+                vec![format!("k/{page_start:05}"), format!("k/{:05}", page_start + 1)]
+            );
+            let cp = read_checkpoint(&store, bucket)
+                .await
+                .expect("read persisted page boundary")
+                .expect("checkpoint")
+                .checkpoint;
+            assert_eq!(cp.job_id, crashed.job_id);
+            assert_eq!(cp.owner.as_ref().map(|owner| owner.node.as_str()), Some("node-b"));
+            assert_eq!(cp.continuation_token.as_deref(), Some(durable_token));
+            assert_eq!(cp.failed, failures);
+            context.complete_pending();
+        }
+        runner.wait_until_idle(bucket).await;
+        let cp = read_checkpoint(&store, bucket)
+            .await
+            .expect("read completed checkpoint")
+            .expect("checkpoint")
+            .checkpoint;
+        assert_eq!(cp.state, BackfillState::CompletedWithFailures);
+        assert_eq!((cp.pulled, cp.failed), (5, 2));
+        assert_eq!(cp.continuation_token.as_deref(), Some("4"));
+        assert_eq!(cp.failed_keys, vec![key_hash("k/00002"), key_hash("k/00004")]);
+        assert_eq!(
+            context.list_requests.lock().as_slice(),
+            &[Some("2".to_string()), Some("4".to_string()), Some("6".to_string())]
+        );
+    }
+
     #[tokio::test]
     async fn recovery_cancels_a_job_whose_config_changed_and_reclaims_own_node_jobs() {
         let bucket = "backfill-recovery-config";

crates/ecstore/src/bucket/on_demand_migration/breaker.rs

@@ -86,7 +86,12 @@ impl BreakerVerdict {
             Some(SourceError::Throttled | SourceError::Timeout | SourceError::Connect(_) | SourceError::ServerError(_)) => {
                 BreakerVerdict::Failure
             }
-            Some(SourceError::AccessDenied | SourceError::Unsupported(_) | SourceError::Other(_)) => BreakerVerdict::Neutral,
+            Some(
+                SourceError::AccessDenied
+                | SourceError::Unsupported(_)
+                | SourceError::InvalidPagination(_)
+                | SourceError::Other(_),
+            ) => BreakerVerdict::Neutral,
         }
     }
 }

crates/ecstore/src/bucket/on_demand_migration/config.rs

@@ -30,6 +30,10 @@ pub const ON_DEMAND_MIGRATION_CONFIG_VERSION: u32 = 1;
 const REDACTED: &str = "REDACTED";
 const AUTO_REGION: &str = "auto";
 const AUTO_REGION_FALLBACK: &str = "us-east-1";
+/// Public Azure Blob host suffix; the account name is the first label.
+pub const AZURE_BLOB_SUFFIX: &str = "blob.core.windows.net";
+/// Public Google Cloud Storage endpoint for the native provider.
+pub const GCS_DEFAULT_ENDPOINT: &str = "https://storage.googleapis.com";
 
 const KIB: u64 = 1024;
 const MIB: u64 = 1024 * KIB;
@@ -75,14 +79,25 @@ pub struct SourceConfig {
     pub bucket: String,
     #[serde(default)]
     pub path_style: PathStyle,
-    /// `None` means anonymous access to a public source bucket.
+    /// `None` means anonymous access to a public source bucket. Only the
+    /// SigV4 providers read it; `azure` and `gcs_native` carry their own
+    /// credentials in `azure` / `gcs`.
     #[serde(default)]
     pub credentials: Option<SourceCredentials>,
     #[serde(default)]
     pub tls: TlsConfig,
+    /// Required for [`Provider::Azure`] and rejected for every other
+    /// provider.
+    #[serde(default)]
+    pub azure: Option<AzureSourceConfig>,
+    /// Required for [`Provider::GcsNative`] and rejected for every other
+    /// provider. [`Provider::Gcs`] keeps using `credentials` because it
+    /// speaks the S3 interoperability API.
+    #[serde(default)]
+    pub gcs: Option<GcsSourceConfig>,
 }
 
-/// Source vendor family. `azure` is deliberately absent from this version.
+/// Source vendor family.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
 #[serde(rename_all = "lowercase")]
 pub enum Provider {
@@ -94,6 +109,12 @@ pub enum Provider {
     R2,
     /// GCS XML interoperability API with HMAC keys.
     Gcs,
+    /// Native Azure Blob service; parameters in `source.azure`.
+    Azure,
+    /// Native GCS JSON API with a service-account key; parameters in
+    /// `source.gcs`.
+    #[serde(rename = "gcs_native")]
+    GcsNative,
 }
 
 impl Provider {
@@ -105,13 +126,22 @@ impl Provider {
             Provider::Rustfs => "rustfs",
             Provider::R2 => "r2",
             Provider::Gcs => "gcs",
+            Provider::Azure => "azure",
+            Provider::GcsNative => "gcs_native",
         }
     }
 
+    /// Providers that do not speak S3 and therefore ignore `region`,
+    /// `path_style` and `credentials`.
+    pub fn is_native(&self) -> bool {
+        matches!(self, Provider::Azure | Provider::GcsNative)
+    }
+
     /// Providers whose SDKs accept `region = "auto"`; RustFS maps it to
-    /// `us-east-1` for signing.
+    /// `us-east-1` for signing. The native providers never sign with a
+    /// region, so they accept it as well.
     fn accepts_auto_region(&self) -> bool {
-        matches!(self, Provider::R2 | Provider::Minio | Provider::Rustfs)
+        matches!(self, Provider::R2 | Provider::Minio | Provider::Rustfs) || self.is_native()
     }
 }
 
@@ -164,6 +194,73 @@ impl fmt::Debug for SourceCredentials {
     }
 }
 
+/// Native Azure Blob source parameters. The container is `source.bucket`,
+/// so a config never carries two names for the same container. Exactly one
+/// of `account_key` and `sas_token` must be set: the account key signs with
+/// Shared Key, the SAS token is appended to every request URL.
+#[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct AzureSourceConfig {
+    /// Storage account name; also derives the default `blob.core.windows.net`
+    /// endpoint when `source.endpoint` is absent.
+    pub account: String,
+    /// Base64 shared key of the storage account.
+    #[serde(default)]
+    pub account_key: Option<String>,
+    /// SAS query string without the leading `?`.
+    #[serde(default)]
+    pub sas_token: Option<String>,
+}
+
+impl AzureSourceConfig {
+    /// A copy safe to return to admin clients or log: both secrets are
+    /// replaced by `REDACTED`, and whether each is set stays visible.
+    pub fn redacted(&self) -> Self {
+        Self {
+            account: self.account.clone(),
+            account_key: self.account_key.as_ref().map(|_| REDACTED.to_string()),
+            sas_token: self.sas_token.as_ref().map(|_| REDACTED.to_string()),
+        }
+    }
+}
+
+impl fmt::Debug for AzureSourceConfig {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("AzureSourceConfig")
+            .field("account", &self.account)
+            .field("account_key", &self.account_key.as_ref().map(|_| REDACTED))
+            .field("sas_token", &self.sas_token.as_ref().map(|_| REDACTED))
+            .finish()
+    }
+}
+
+/// Native Google Cloud Storage source parameters. The bucket is
+/// `source.bucket`; only the service-account key lives here.
+#[derive(Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct GcsSourceConfig {
+    /// Service-account key JSON, verbatim as downloaded from Google Cloud.
+    pub service_account_json: String,
+}
+
+impl GcsSourceConfig {
+    /// A copy safe to return to admin clients or log: the whole key JSON is
+    /// a secret (it embeds the private key), so it is replaced wholesale.
+    pub fn redacted(&self) -> Self {
+        Self {
+            service_account_json: REDACTED.to_string(),
+        }
+    }
+}
+
+impl fmt::Debug for GcsSourceConfig {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("GcsSourceConfig")
+            .field("service_account_json", &REDACTED)
+            .finish()
+    }
+}
+
 #[derive(Debug, Clone, PartialEq, Eq, Default, Serialize, Deserialize)]
 #[serde(deny_unknown_fields)]
 pub struct TlsConfig {
@@ -354,6 +451,14 @@ pub enum OnDemandMigrationConfigError {
     InvalidBucket(&'static str),
     #[error("source credentials field {0} must not be empty")]
     EmptyCredential(&'static str),
+    #[error("source.{0} is required for provider {1}")]
+    MissingProviderBlock(&'static str, Provider),
+    #[error("source.{0} is not valid for provider {1}")]
+    UnexpectedProviderBlock(&'static str, Provider),
+    /// Carries only the reason: the block holds account keys, SAS tokens and
+    /// service-account JSON, so no value of it is ever echoed.
+    #[error("source.{0} is invalid: {1}")]
+    InvalidProviderBlock(&'static str, &'static str),
     #[error("source tls.ca_cert_pem is not a PEM certificate")]
     InvalidCaCert,
     #[error("filter.{0} must be null or a non-empty string")]
@@ -388,6 +493,8 @@ impl OnDemandMigrationConfig {
     pub fn redacted(&self) -> Self {
         let mut copy = self.clone();
         copy.source.credentials = self.source.credentials.as_ref().map(SourceCredentials::redacted);
+        copy.source.azure = self.source.azure.as_ref().map(AzureSourceConfig::redacted);
+        copy.source.gcs = self.source.gcs.as_ref().map(GcsSourceConfig::redacted);
         copy
     }
 
@@ -433,6 +540,12 @@ impl SourceConfig {
         match (&self.endpoint, self.provider) {
             (Some(endpoint), _) => endpoint.clone(),
             (None, Provider::Aws) => format!("https://s3.{}.amazonaws.com", self.region),
+            (None, Provider::Azure) => self
+                .azure
+                .as_ref()
+                .map(|azure| format!("https://{}.{AZURE_BLOB_SUFFIX}", azure.account))
+                .unwrap_or_default(),
+            (None, Provider::GcsNative) => GCS_DEFAULT_ENDPOINT.to_string(),
             (None, _) => String::new(),
         }
     }
@@ -448,6 +561,8 @@ impl SourceConfig {
     }
 
     fn validate(&self) -> Result<(), OnDemandMigrationConfigError> {
+        self.validate_provider_block()?;
+
         if self.region.is_empty() {
             return Err(OnDemandMigrationConfigError::EmptyRegion);
         }
@@ -466,6 +581,9 @@ impl SourceConfig {
                     ));
                 }
             }
+            // Both native providers derive a fixed endpoint; Azure's is built
+            // from the account name, already checked by `validate_provider_block`.
+            None if self.provider.is_native() => {}
             None => return Err(OnDemandMigrationConfigError::MissingEndpoint(self.provider)),
         }
 
@@ -496,6 +614,84 @@ impl SourceConfig {
 
         Ok(())
     }
+
+    /// The provider-specific block must be present for exactly its own
+    /// provider: a stray `azure` block on an `s3` source would otherwise be
+    /// accepted, stored, and silently ignored by the client builder.
+    fn validate_provider_block(&self) -> Result<(), OnDemandMigrationConfigError> {
+        let missing = OnDemandMigrationConfigError::MissingProviderBlock;
+        let unexpected = OnDemandMigrationConfigError::UnexpectedProviderBlock;
+        let invalid = OnDemandMigrationConfigError::InvalidProviderBlock;
+
+        if self.provider != Provider::Azure && self.azure.is_some() {
+            return Err(unexpected("azure", self.provider));
+        }
+        if self.provider != Provider::GcsNative && self.gcs.is_some() {
+            return Err(unexpected("gcs", self.provider));
+        }
+
+        match self.provider {
+            Provider::Azure => {
+                let azure = self.azure.as_ref().ok_or(missing("azure", self.provider))?;
+                if azure.account.is_empty() {
+                    return Err(invalid("azure", "account must not be empty"));
+                }
+                // The account feeds a hostname when the endpoint is derived:
+                // keep it to label characters so it cannot rewrite the host.
+                if !azure.account.bytes().all(|b| b.is_ascii_alphanumeric() || b == b'-') {
+                    return Err(invalid("azure", "account contains characters outside [A-Za-z0-9-]"));
+                }
+                match (azure.account_key.as_deref(), azure.sas_token.as_deref()) {
+                    (Some(_), Some(_)) => return Err(invalid("azure", "account_key and sas_token are mutually exclusive")),
+                    (None, None) => return Err(invalid("azure", "one of account_key and sas_token is required")),
+                    (Some(key), None) => {
+                        if key.is_empty() {
+                            return Err(invalid("azure", "account_key must not be empty"));
+                        }
+                        // Decoded here so a mistyped key fails at the admin
+                        // boundary instead of on the first source request.
+                        if base64_simd::STANDARD.decode_to_vec(key.as_bytes()).is_err() {
+                            return Err(invalid("azure", "account_key is not base64"));
+                        }
+                    }
+                    (None, Some(sas)) => {
+                        if sas.is_empty() {
+                            return Err(invalid("azure", "sas_token must not be empty"));
+                        }
+                        if sas.starts_with('?') {
+                            return Err(invalid("azure", "sas_token must not start with '?'"));
+                        }
+                        if sas.chars().any(char::is_whitespace) {
+                            return Err(invalid("azure", "sas_token must not contain whitespace"));
+                        }
+                    }
+                }
+            }
+            Provider::GcsNative => {
+                let gcs = self.gcs.as_ref().ok_or(missing("gcs", self.provider))?;
+                let key: serde_json::Value = serde_json::from_str(&gcs.service_account_json)
+                    .map_err(|_| invalid("gcs", "service_account_json is not valid JSON"))?;
+                let Some(object) = key.as_object() else {
+                    return Err(invalid("gcs", "service_account_json is not a JSON object"));
+                };
+                if object.get("type").and_then(serde_json::Value::as_str) != Some("service_account") {
+                    return Err(invalid("gcs", "service_account_json is not a service_account key"));
+                }
+                for field in ["client_email", "private_key"] {
+                    if object
+                        .get(field)
+                        .and_then(serde_json::Value::as_str)
+                        .is_none_or(str::is_empty)
+                    {
+                        return Err(invalid("gcs", "service_account_json is missing client_email or private_key"));
+                    }
+                }
+            }
+            Provider::S3 | Provider::Aws | Provider::Minio | Provider::Rustfs | Provider::R2 | Provider::Gcs => {}
+        }
+
+        Ok(())
+    }
 }
 
 fn validate_endpoint(endpoint: &str) -> Result<(), OnDemandMigrationConfigError> {
@@ -699,7 +895,15 @@ mod tests {
             ),
             (
                 "provider enum",
-                r#"{"source":{"provider":"azure","endpoint":"https://h","region":"r","bucket":"b"}}"#,
+                r#"{"source":{"provider":"swift","endpoint":"https://h","region":"r","bucket":"b"}}"#,
+            ),
+            (
+                "azure block",
+                r#"{"source":{"provider":"azure","region":"auto","bucket":"b","azure":{"account":"acct","account_key":"a2V5","extra":1}}}"#,
+            ),
+            (
+                "gcs block",
+                r#"{"source":{"provider":"gcs_native","region":"auto","bucket":"b","gcs":{"service_account_json":"{}","extra":1}}}"#,
             ),
         ] {
             let err = OnDemandMigrationConfig::from_json(json.as_bytes()).expect_err(label);
@@ -820,9 +1024,201 @@ mod tests {
                 "{provider}"
             );
         }
+        // The native providers never sign with a region, so "auto" is the
+        // honest value to write for them.
+        for cfg in [azure_cfg(), gcs_native_cfg()] {
+            assert_eq!(cfg.source.region, "auto");
+            cfg.validate(empty_ctx())
+                .unwrap_or_else(|err| panic!("{}: {err}", cfg.source.provider));
+        }
         assert_eq!(sample().source.effective_region(), "us-west-1");
     }
 
+    const SERVICE_ACCOUNT_JSON: &str = r#"{"type":"service_account","project_id":"p","client_email":"a@b.iam.gserviceaccount.com","private_key":"-----BEGIN PRIVATE KEY-----\nsecret\n-----END PRIVATE KEY-----"}"#;
+
+    fn azure_cfg() -> OnDemandMigrationConfig {
+        let mut cfg = sample();
+        cfg.source.provider = Provider::Azure;
+        cfg.source.endpoint = None;
+        cfg.source.region = "auto".to_string();
+        cfg.source.credentials = None;
+        cfg.source.azure = Some(AzureSourceConfig {
+            account: "legacyaccount".to_string(),
+            account_key: Some("c2VjcmV0LWtleQ==".to_string()),
+            sas_token: None,
+        });
+        cfg
+    }
+
+    fn gcs_native_cfg() -> OnDemandMigrationConfig {
+        let mut cfg = sample();
+        cfg.source.provider = Provider::GcsNative;
+        cfg.source.endpoint = None;
+        cfg.source.region = "auto".to_string();
+        cfg.source.credentials = None;
+        cfg.source.gcs = Some(GcsSourceConfig {
+            service_account_json: SERVICE_ACCOUNT_JSON.to_string(),
+        });
+        cfg
+    }
+
+    #[test]
+    fn native_providers_derive_their_endpoint_and_round_trip_on_the_wire() {
+        let azure = azure_cfg();
+        assert_eq!(azure.source.effective_endpoint(), "https://legacyaccount.blob.core.windows.net");
+        let gcs = gcs_native_cfg();
+        assert_eq!(gcs.source.effective_endpoint(), "https://storage.googleapis.com");
+
+        for cfg in [azure_cfg(), gcs_native_cfg()] {
+            let json = cfg.to_json().expect("config must serialize");
+            assert_eq!(OnDemandMigrationConfig::from_json(&json).expect("config must parse"), cfg);
+        }
+        // The wire labels are part of the admin contract.
+        assert!(
+            String::from_utf8(azure_cfg().to_json().expect("json"))
+                .expect("utf8")
+                .contains(r#""provider":"azure""#)
+        );
+        assert!(
+            String::from_utf8(gcs_native_cfg().to_json().expect("json"))
+                .expect("utf8")
+                .contains(r#""provider":"gcs_native""#)
+        );
+    }
+
+    #[test]
+    fn an_explicit_endpoint_overrides_the_derived_native_one() {
+        // Azurite and fake-gcs-server are addressed this way.
+        let mut cfg = azure_cfg();
+        cfg.source.endpoint = Some("http://azurite.example.com:10000".to_string());
+        cfg.validate(empty_ctx()).expect("an explicit native endpoint is allowed");
+        assert_eq!(cfg.source.effective_endpoint(), "http://azurite.example.com:10000");
+
+        cfg.source.endpoint = Some("http://azurite.example.com:10000/devstoreaccount1".to_string());
+        assert!(
+            matches!(cfg.validate(empty_ctx()), Err(OnDemandMigrationConfigError::InvalidEndpoint(_))),
+            "a native endpoint is still an origin"
+        );
+    }
+
+    #[test]
+    fn a_provider_block_belongs_to_exactly_its_own_provider() {
+        let mut cfg = sample();
+        cfg.source.azure = azure_cfg().source.azure;
+        assert_eq!(
+            cfg.validate(empty_ctx()),
+            Err(OnDemandMigrationConfigError::UnexpectedProviderBlock("azure", Provider::S3))
+        );
+
+        let mut cfg = sample();
+        cfg.source.gcs = gcs_native_cfg().source.gcs;
+        assert_eq!(
+            cfg.validate(empty_ctx()),
+            Err(OnDemandMigrationConfigError::UnexpectedProviderBlock("gcs", Provider::S3))
+        );
+
+        let mut cfg = azure_cfg();
+        cfg.source.azure = None;
+        assert_eq!(
+            cfg.validate(empty_ctx()),
+            Err(OnDemandMigrationConfigError::MissingProviderBlock("azure", Provider::Azure))
+        );
+
+        let mut cfg = gcs_native_cfg();
+        cfg.source.gcs = None;
+        assert_eq!(
+            cfg.validate(empty_ctx()),
+            Err(OnDemandMigrationConfigError::MissingProviderBlock("gcs", Provider::GcsNative))
+        );
+    }
+
+    #[test]
+    fn azure_block_rules() {
+        let with = |account: &str, key: Option<&str>, sas: Option<&str>| {
+            let mut cfg = azure_cfg();
+            cfg.source.azure = Some(AzureSourceConfig {
+                account: account.to_string(),
+                account_key: key.map(str::to_string),
+                sas_token: sas.map(str::to_string),
+            });
+            cfg.validate(empty_ctx())
+        };
+
+        with("legacyaccount", None, Some("sv=2021-08-06&sig=abc%3D")).expect("a SAS token is a complete credential");
+        with("legacyaccount", Some("c2VjcmV0LWtleQ=="), None).expect("an account key is a complete credential");
+
+        for (label, result) in [
+            ("empty account", with("", Some("c2VjcmV0LWtleQ=="), None)),
+            // The account becomes the first label of the derived hostname.
+            ("account with a dot", with("legacy.account", Some("c2VjcmV0LWtleQ=="), None)),
+            ("account with a slash", with("legacy/account", Some("c2VjcmV0LWtleQ=="), None)),
+            ("no credential", with("legacyaccount", None, None)),
+            ("both credentials", with("legacyaccount", Some("c2VjcmV0LWtleQ=="), Some("sv=1"))),
+            ("empty key", with("legacyaccount", Some(""), None)),
+            ("key that is not base64", with("legacyaccount", Some("not base64!"), None)),
+            ("empty sas", with("legacyaccount", None, Some(""))),
+            ("sas with a leading question mark", with("legacyaccount", None, Some("?sv=1"))),
+            ("sas with whitespace", with("legacyaccount", None, Some("sv=1 &sig=a"))),
+        ] {
+            assert!(
+                matches!(result, Err(OnDemandMigrationConfigError::InvalidProviderBlock("azure", _))),
+                "{label}: {result:?}"
+            );
+        }
+    }
+
+    #[test]
+    fn gcs_native_block_requires_a_usable_service_account_key() {
+        let with = |json: &str| {
+            let mut cfg = gcs_native_cfg();
+            cfg.source.gcs = Some(GcsSourceConfig {
+                service_account_json: json.to_string(),
+            });
+            cfg.validate(empty_ctx())
+        };
+
+        with(SERVICE_ACCOUNT_JSON).expect("a service-account key is accepted");
+        for (label, json) in [
+            ("empty", ""),
+            ("not json", "not json"),
+            ("not an object", "[]"),
+            ("wrong type", r#"{"type":"authorized_user","client_email":"a@b","private_key":"k"}"#),
+            ("no private key", r#"{"type":"service_account","client_email":"a@b"}"#),
+            ("empty client email", r#"{"type":"service_account","client_email":"","private_key":"k"}"#),
+        ] {
+            let result = with(json);
+            assert!(
+                matches!(result, Err(OnDemandMigrationConfigError::InvalidProviderBlock("gcs", _))),
+                "{label}: {result:?}"
+            );
+        }
+    }
+
+    #[test]
+    fn native_secrets_never_survive_redaction_or_debug() {
+        let mut azure = azure_cfg();
+        azure.source.azure.as_mut().expect("block").sas_token = Some("sv=2021-08-06&sig=top-secret".to_string());
+        azure.source.azure.as_mut().expect("block").account_key = None;
+        let gcs = gcs_native_cfg();
+
+        for rendered in [
+            format!("{:?}", azure.redacted()),
+            format!("{azure:?}"),
+            String::from_utf8(azure.redacted().to_json().expect("json")).expect("utf8"),
+        ] {
+            assert!(!rendered.contains("top-secret"), "{rendered}");
+            assert!(rendered.contains("legacyaccount"), "the account name is not a secret: {rendered}");
+        }
+        for rendered in [
+            format!("{:?}", gcs.redacted()),
+            format!("{gcs:?}"),
+            String::from_utf8(gcs.redacted().to_json().expect("json")).expect("utf8"),
+        ] {
+            assert!(!rendered.contains("PRIVATE KEY-----"), "{rendered}");
+            assert!(!rendered.contains("gserviceaccount"), "{rendered}");
+        }
+    }
+
     #[test]
     fn bucket_rules() {
         let mut cfg = sample();

crates/ecstore/src/bucket/on_demand_migration/gcs.rs

@@ -0,0 +1,506 @@
+// 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.
+
+//! Native Google Cloud Storage source backend.
+//!
+//! The `gcs` provider already reaches GCS through its S3 interoperability API,
+//! which needs an HMAC key pair. This backend is the other half: it authorizes
+//! with a service-account key, the credential most GCS projects actually issue,
+//! by minting OAuth tokens through the shared `google-cloud-auth` credential
+//! machinery the tier layer already uses.
+//!
+//! Two GCS surfaces are involved, each for the half it describes best. The read
+//! path uses the XML API (`/{bucket}/{object}`), whose responses carry
+//! `x-goog-meta-*` user metadata and the `x-goog-hash` digest in one round trip.
+//! Listing uses the JSON API (`objects.list`), whose `pageToken` maps directly
+//! onto the shared page cursor and whose `prefixes` are the delimiter roll-up.
+//! Both accept the same bearer token.
+//!
+//! Every call this backend makes needs only `storage.objects.get` and
+//! `storage.objects.list`, the two permissions of the `objectViewer` role, so a
+//! key scoped to exactly the migration's needs works.
+//!
+//! `x-goog-hash` carries a base64 MD5 for every non-composite object; it is
+//! converted to hex and becomes the head's ETag, so a pulled object is checked
+//! against the digest GCS itself computed. A composite object has no MD5, and
+//! its ETag is then marked opaque rather than checked.
+
+use super::native_http::{
+    NativeHeadFields, NativeHttp, base64_md5_to_hex, header, native_source_head, parse_http_timestamp, read_text, response_body,
+};
+use super::source_client::{
+    GcsSourceSpec, SourceBackend, SourceError, SourceGet, SourceHead, SourceListRequest, SourceObject, SourcePage,
+    SourceTimeouts, range_header_value,
+};
+use crate::bucket::remote_s3_client::RemoteS3ClientError;
+use crate::storage_api_contracts::range::HTTPRangeSpec;
+use google_cloud_auth::credentials::service_account::{AccessSpecifier, Builder as ServiceAccountBuilder};
+use google_cloud_auth::credentials::{CacheableResource, Credentials};
+use http::{HeaderMap, HeaderValue, Method};
+use serde::Deserialize;
+use std::collections::HashMap;
+use url::Url;
+
+/// Read-only object scope: this backend never writes to the source.
+const READ_ONLY_SCOPE: &str = "https://www.googleapis.com/auth/devstorage.read_only";
+const METADATA_PREFIX: &str = "x-goog-meta-";
+/// GCS reports its error code in the response body, not a header; the shared
+/// transport takes a header name, so it is given one that never matches and
+/// classification falls back to the status.
+const NO_ERROR_CODE_HEADER: &str = "x-goog-unused-error-code";
+/// One `objects.list` page is small; refuse an unbounded document.
+const MAX_JSON_BYTES: usize = 8 * 1024 * 1024;
+
+pub struct GcsNativeSourceBackend {
+    http: NativeHttp,
+    bucket: String,
+    credentials: Credentials,
+}
+
+impl GcsNativeSourceBackend {
+    pub fn new(
+        endpoint: &str,
+        bucket: &str,
+        spec: &GcsSourceSpec,
+        timeouts: SourceTimeouts,
+        skip_tls_verify: bool,
+        ca_cert_pem: Option<&str>,
+    ) -> Result<Self, RemoteS3ClientError> {
+        let key: serde_json::Value = serde_json::from_str(&spec.service_account_json)
+            .map_err(|_| RemoteS3ClientError::Credentials("gcs service account key is not valid JSON"))?;
+        let credentials = ServiceAccountBuilder::new(key)
+            .with_access_specifier(AccessSpecifier::from_scopes([READ_ONLY_SCOPE]))
+            .build()
+            .map_err(|_| RemoteS3ClientError::Credentials("gcs service account key is not usable"))?;
+        Ok(Self {
+            http: NativeHttp::new(endpoint, timeouts, skip_tls_verify, ca_cert_pem)?,
+            bucket: bucket.to_string(),
+            credentials,
+        })
+    }
+
+    /// Authorization headers for one request. A credential failure is reported
+    /// as `AccessDenied` with no message: the renderer of a credential error
+    /// has the key material in scope, and the class is what callers act on.
+    async fn auth_headers(&self) -> Result<HeaderMap, SourceError> {
+        match self.credentials.headers(http::Extensions::new()).await {
+            Ok(CacheableResource::New { data, .. }) => Ok(data),
+            // Only returned when the caller passes an entity tag, which this
+            // backend never does; an empty set is still the honest answer.
+            Ok(CacheableResource::NotModified) => Ok(HeaderMap::new()),
+            Err(_) => Err(SourceError::AccessDenied),
+        }
+    }
+
+    /// XML API URL of one object; `/` in the key stay path separators.
+    fn object_url(&self, key: &str) -> Result<Url, SourceError> {
+        self.http.url(std::iter::once(self.bucket.as_str()).chain(key.split('/')))
+    }
+
+    /// JSON API URL of the bucket's object collection.
+    fn objects_url(&self) -> Result<Url, SourceError> {
+        self.http.url(["storage", "v1", "b", self.bucket.as_str(), "o"])
+    }
+
+    async fn request(&self, method: Method, url: Url, mut headers: HeaderMap) -> Result<reqwest::Request, SourceError> {
+        for (name, value) in self.auth_headers().await? {
+            if let Some(name) = name {
+                headers.insert(name, value);
+            }
+        }
+        let mut request = reqwest::Request::new(method, url);
+        *request.headers_mut() = headers;
+        Ok(request)
+    }
+
+    /// Shared mapping for the XML API's HEAD and GET responses.
+    fn head_from_response(headers: &HeaderMap) -> Result<SourceHead, SourceError> {
+        if header(headers, "x-goog-encryption-key-sha256").is_some() {
+            return Err(SourceError::Unsupported(
+                "source object uses a customer-supplied encryption key; customer-key sources are not supported".to_string(),
+            ));
+        }
+        // `x-goog-hash` lists digests as `name=base64`, comma separated, and may
+        // repeat across header lines. Only the MD5 describes the whole object.
+        let md5 = headers
+            .get_all("x-goog-hash")
+            .iter()
+            .filter_map(|value| value.to_str().ok())
+            .flat_map(|value| value.split(','))
+            .filter_map(|digest| digest.trim().strip_prefix("md5="))
+            .find_map(base64_md5_to_hex);
+
+        let (etag, etag_is_opaque) = match md5 {
+            Some(md5) => (Some(md5), false),
+            // A composite object has no MD5; its ETag describes the composition
+            // rather than the bytes, so it is provenance only.
+            None => (header(headers, "etag").map(str::to_string), true),
+        };
+        native_source_head(
+            headers,
+            METADATA_PREFIX,
+            NativeHeadFields {
+                etag,
+                etag_is_opaque,
+                version_id: header(headers, "x-goog-generation").map(str::to_string),
+                storage_class: header(headers, "x-goog-storage-class").map(str::to_string),
+            },
+        )
+    }
+}
+
+#[async_trait::async_trait]
+impl SourceBackend for GcsNativeSourceBackend {
+    async fn head(&self, key: &str) -> Result<SourceHead, SourceError> {
+        let request = self.request(Method::HEAD, self.object_url(key)?, HeaderMap::new()).await?;
+        let response = self.http.send(request, NO_ERROR_CODE_HEADER).await?;
+        Self::head_from_response(response.headers())
+    }
+
+    async fn get(&self, key: &str, range: Option<&HTTPRangeSpec>) -> Result<SourceGet, SourceError> {
+        let mut headers = HeaderMap::new();
+        if let Some(range) = range.map(range_header_value).transpose()? {
+            headers.insert(
+                http::header::RANGE,
+                HeaderValue::from_str(&range).map_err(|_| SourceError::Other("invalid range header".to_string()))?,
+            );
+        }
+        let request = self.request(Method::GET, self.object_url(key)?, headers).await?;
+        let response = self.http.send(request, NO_ERROR_CODE_HEADER).await?;
+        let head = Self::head_from_response(response.headers())?;
+        let content_range = header(response.headers(), "content-range").map(str::to_string);
+        Ok(SourceGet {
+            head,
+            body: response_body(response),
+            content_range,
+        })
+    }
+
+    async fn list(&self, request: &SourceListRequest<'_>) -> Result<SourcePage, SourceError> {
+        // `objects.list` offers `startOffset`, which is inclusive, so it cannot
+        // express "resume after this key" without silently repeating it.
+        if request.start_after.is_some() {
+            return Err(SourceError::Unsupported(
+                "gcs sources cannot resume a listing from a key; use the continuation token".to_string(),
+            ));
+        }
+        let mut url = self.objects_url()?;
+        {
+            let mut query = url.query_pairs_mut();
+            if let Some(prefix) = request.prefix.filter(|prefix| !prefix.is_empty()) {
+                query.append_pair("prefix", prefix);
+            }
+            if let Some(delimiter) = request.delimiter.filter(|delimiter| !delimiter.is_empty()) {
+                query.append_pair("delimiter", delimiter);
+            }
+            if let Some(token) = request.continuation_token.filter(|token| !token.is_empty()) {
+                query.append_pair("pageToken", token);
+            }
+            if request.max_keys > 0 {
+                query.append_pair("maxResults", &request.max_keys.to_string());
+            }
+        }
+
+        let request = self.request(Method::GET, url, HeaderMap::new()).await?;
+        let response = self.http.send(request, NO_ERROR_CODE_HEADER).await?;
+        let body = read_text(response, MAX_JSON_BYTES).await?;
+        parse_objects_list(&body)
+    }
+
+    /// GCS has no object tagging API; user metadata is already carried by the
+    /// head mapping. An empty map keeps `policy.copy_tags` from failing a pull
+    /// over a concept the provider does not have.
+    async fn tagging(&self, _key: &str) -> Result<HashMap<String, String>, SourceError> {
+        Ok(HashMap::new())
+    }
+
+    /// A one-object listing, not `buckets.get`: the migration pipeline only
+    /// ever needs `storage.objects.list` and `storage.objects.get`, and a key
+    /// scoped to exactly those (the `objectViewer` role) cannot read the bucket
+    /// resource. Probing with `buckets.get` would reject a correct key.
+    async fn probe(&self) -> Result<(), SourceError> {
+        let mut url = self.objects_url()?;
+        url.query_pairs_mut().append_pair("maxResults", "1");
+        let request = self.request(Method::GET, url, HeaderMap::new()).await?;
+        let response = self.http.send(request, NO_ERROR_CODE_HEADER).await?;
+        read_text(response, MAX_JSON_BYTES)
+            .await
+            .and_then(|body| parse_objects_list(&body))?;
+        Ok(())
+    }
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct ObjectsList {
+    #[serde(default)]
+    items: Vec<ListedObject>,
+    #[serde(default)]
+    prefixes: Vec<String>,
+    #[serde(default)]
+    next_page_token: Option<String>,
+}
+
+#[derive(Debug, Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct ListedObject {
+    name: String,
+    /// GCS renders the size as a decimal string, not a JSON number.
+    #[serde(default)]
+    size: Option<String>,
+    #[serde(default)]
+    updated: Option<String>,
+    #[serde(default)]
+    md5_hash: Option<String>,
+    #[serde(default)]
+    etag: Option<String>,
+    #[serde(default)]
+    storage_class: Option<String>,
+}
+
+fn parse_objects_list(body: &str) -> Result<SourcePage, SourceError> {
+    let listing: ObjectsList =
+        serde_json::from_str(body).map_err(|err| SourceError::Other(format!("source listing is not valid JSON: {err}")))?;
+    let next_continuation_token = listing.next_page_token.filter(|token| !token.is_empty());
+    let objects = listing
+        .items
+        .into_iter()
+        .map(|item| {
+            let etag = item
+                .md5_hash
+                .as_deref()
+                .and_then(base64_md5_to_hex)
+                .or_else(|| item.etag.map(|etag| etag.trim_matches('"').to_string()))
+                .filter(|etag| !etag.is_empty());
+            SourceObject {
+                key: item.name,
+                etag,
+                size: item.size.and_then(|size| size.parse().ok()).unwrap_or(0),
+                last_modified: item.updated.as_deref().and_then(parse_http_timestamp),
+                storage_class: item.storage_class,
+                // GCS never encodes a part count in a digest or an ETag.
+                is_multipart_etag: false,
+            }
+        })
+        .collect();
+
+    Ok(SourcePage {
+        objects,
+        common_prefixes: listing.prefixes,
+        is_truncated: next_continuation_token.is_some(),
+        next_continuation_token,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::bucket::on_demand_migration::backend_contract::{BackendCapabilities, assert_backend_contract};
+    use crate::bucket::on_demand_migration::test_http_fixture::{ScriptedResponse, scripted_server};
+    use google_cloud_auth::credentials::anonymous::Builder as AnonymousBuilder;
+
+    const LIST_PAGE_ONE: &str = r#"{
+      "kind": "storage#objects",
+      "nextPageToken": "cursor-1",
+      "prefixes": ["dir/sub/"],
+      "items": [
+        {
+          "name": "dir/a.txt",
+          "size": "5",
+          "updated": "2015-10-21T07:28:00.000Z",
+          "md5Hash": "XUFAKrxLKna5cZ2REBfFkg==",
+          "etag": "CJizy9Wq0McCEAE=",
+          "storageClass": "STANDARD"
+        }
+      ]
+    }"#;
+
+    const LIST_PAGE_TWO: &str = r#"{
+      "kind": "storage#objects",
+      "items": [
+        {
+          "name": "dir/b.txt",
+          "size": "7",
+          "updated": "2015-10-21T07:28:00.000Z",
+          "etag": "\"CJizy9Wq0McCEAI=\""
+        }
+      ]
+    }"#;
+
+    fn backend(endpoint: &Url) -> GcsNativeSourceBackend {
+        GcsNativeSourceBackend {
+            http: NativeHttp::for_test(endpoint.clone()),
+            bucket: "legacy".to_string(),
+            // Anonymous credentials add no headers, so the fixture sees exactly
+            // the request this backend builds.
+            credentials: AnonymousBuilder::new().build(),
+        }
+    }
+
+    fn object_headers() -> Vec<(&'static str, String)> {
+        vec![
+            ("Content-Type", "text/plain".to_string()),
+            ("Last-Modified", "Wed, 21 Oct 2015 07:28:00 GMT".to_string()),
+            ("ETag", "\"CJizy9Wq0McCEAE=\"".to_string()),
+            ("x-goog-hash", "crc32c=AAAAAA==,md5=XUFAKrxLKna5cZ2REBfFkg==".to_string()),
+            ("x-goog-meta-owner", "alice".to_string()),
+            ("x-goog-storage-class", "STANDARD".to_string()),
+            ("x-goog-generation", "1445412480000000".to_string()),
+        ]
+    }
+
+    #[test]
+    fn objects_list_maps_items_prefixes_and_the_page_token() {
+        let page = parse_objects_list(LIST_PAGE_ONE).expect("page should parse");
+        assert_eq!(page.common_prefixes, vec!["dir/sub/"]);
+        assert!(page.is_truncated);
+        assert_eq!(page.next_continuation_token.as_deref(), Some("cursor-1"));
+        assert_eq!(page.objects.len(), 1);
+        assert_eq!(page.objects[0].key, "dir/a.txt");
+        assert_eq!(page.objects[0].size, 5, "the string size is parsed");
+        assert_eq!(
+            page.objects[0].etag.as_deref(),
+            Some("5d41402abc4b2a76b9719d911017c592"),
+            "the base64 md5Hash becomes a hex ETag"
+        );
+        assert_eq!(page.objects[0].storage_class.as_deref(), Some("STANDARD"));
+        assert!(page.objects[0].last_modified.is_some(), "RFC 3339 `updated` is parsed");
+
+        let page = parse_objects_list(LIST_PAGE_TWO).expect("page should parse");
+        assert!(!page.is_truncated);
+        assert!(page.next_continuation_token.is_none());
+        assert_eq!(
+            page.objects[0].etag.as_deref(),
+            Some("CJizy9Wq0McCEAI="),
+            "without md5Hash the raw etag is carried"
+        );
+
+        assert!(parse_objects_list("not json").is_err());
+    }
+
+    #[tokio::test]
+    async fn head_prefers_the_goog_hash_md5_over_the_etag() {
+        let (endpoint, recorded) = scripted_server(vec![ScriptedResponse::new(200, object_headers(), String::new())]).await;
+        let head = backend(&endpoint).head("dir/a b.txt").await.expect("HEAD should map");
+
+        let recorded = recorded.lock().expect("recorder lock").clone();
+        assert_eq!(recorded[0].method, "HEAD");
+        assert_eq!(recorded[0].target, "/legacy/dir/a%20b.txt", "the XML API addresses the object by path");
+        assert_eq!(
+            head.etag.as_deref(),
+            Some("5d41402abc4b2a76b9719d911017c592"),
+            "the x-goog-hash md5 is the content digest"
+        );
+        assert!(!head.etag_is_opaque, "a GCS md5 may be checked against the pulled bytes");
+        assert_eq!(head.user_metadata, HashMap::from([("owner".to_string(), "alice".to_string())]));
+        assert_eq!(head.version_id.as_deref(), Some("1445412480000000"));
+        assert_eq!(head.storage_class.as_deref(), Some("STANDARD"));
+    }
+
+    #[tokio::test]
+    async fn a_composite_object_without_an_md5_keeps_an_opaque_etag() {
+        let headers = object_headers()
+            .into_iter()
+            .map(|(name, value)| {
+                if name == "x-goog-hash" {
+                    (name, "crc32c=AAAAAA==".to_string())
+                } else {
+                    (name, value)
+                }
+            })
+            .collect();
+        let (endpoint, _) = scripted_server(vec![ScriptedResponse::new(200, headers, String::new())]).await;
+        let head = backend(&endpoint).head("composed").await.expect("HEAD should map");
+        assert_eq!(head.etag.as_deref(), Some("CJizy9Wq0McCEAE="));
+        assert!(head.etag_is_opaque, "a composite ETag describes the composition, not the bytes");
+    }
+
+    #[tokio::test]
+    async fn customer_supplied_key_objects_are_refused() {
+        let mut headers = object_headers();
+        headers.push(("x-goog-encryption-key-sha256", "abc".to_string()));
+        let (endpoint, _) = scripted_server(vec![ScriptedResponse::new(200, headers, String::new())]).await;
+        let err = backend(&endpoint)
+            .head("a.txt")
+            .await
+            .expect_err("CSEK objects are unsupported");
+        assert!(matches!(err, SourceError::Unsupported(_)), "{err:?}");
+    }
+
+    #[tokio::test]
+    async fn list_and_probe_address_the_json_api() {
+        let (endpoint, recorded) = scripted_server(vec![
+            ScriptedResponse::new(200, Vec::new(), LIST_PAGE_ONE.to_string()),
+            ScriptedResponse::new(200, Vec::new(), "{}".to_string()),
+        ])
+        .await;
+        let backend = backend(&endpoint);
+
+        backend
+            .list(&SourceListRequest {
+                prefix: Some("dir/"),
+                delimiter: Some("/"),
+                continuation_token: Some("cursor-0"),
+                max_keys: 2,
+                ..Default::default()
+            })
+            .await
+            .expect("listing should succeed");
+        backend.probe().await.expect("probe should succeed");
+
+        let recorded = recorded.lock().expect("recorder lock").clone();
+        assert!(recorded[0].target.starts_with("/storage/v1/b/legacy/o?"), "{}", recorded[0].target);
+        assert!(recorded[0].target.contains("prefix=dir%2F"), "{}", recorded[0].target);
+        assert!(recorded[0].target.contains("delimiter=%2F"), "{}", recorded[0].target);
+        assert!(recorded[0].target.contains("pageToken=cursor-0"), "{}", recorded[0].target);
+        assert!(recorded[0].target.contains("maxResults=2"), "{}", recorded[0].target);
+        assert_eq!(
+            recorded[1].target, "/storage/v1/b/legacy/o?maxResults=1",
+            "the probe uses the listing permission the pipeline already needs"
+        );
+    }
+
+    #[tokio::test]
+    async fn gcs_native_backend_satisfies_the_shared_backend_contract() {
+        let mut ranged = object_headers();
+        ranged.push(("Content-Range", "bytes 1-3/5".to_string()));
+        // A HEAD reports the object size with no body, exactly as GCS does.
+        let mut head_only = object_headers();
+        head_only.push(("Content-Length", "5".to_string()));
+        let (endpoint, _) = scripted_server(vec![
+            ScriptedResponse::new(200, head_only, String::new()),
+            ScriptedResponse::new(200, object_headers(), "hello".to_string()),
+            ScriptedResponse::new(206, ranged, "ell".to_string()),
+            ScriptedResponse::new(200, Vec::new(), LIST_PAGE_ONE.to_string()),
+            ScriptedResponse::new(200, Vec::new(), LIST_PAGE_TWO.to_string()),
+            // GCS has no tagging call, so the contract's tag step issues no
+            // request; the probe is the next one on the wire.
+            ScriptedResponse::new(200, Vec::new(), "{}".to_string()),
+            ScriptedResponse::new(404, Vec::new(), String::new()),
+            ScriptedResponse::new(403, Vec::new(), String::new()),
+        ])
+        .await;
+
+        assert_backend_contract(
+            &backend(&endpoint),
+            BackendCapabilities {
+                etag_is_opaque: false,
+                supports_start_after: false,
+                // GCS objects have no tags; the contract's tag step is skipped.
+                supports_tagging: false,
+            },
+        )
+        .await;
+    }
+}

crates/ecstore/src/bucket/on_demand_migration/list_through.rs

@@ -25,13 +25,21 @@ use parking_lot::Mutex;
 use serde::{Deserialize, Serialize};
 use std::time::{Duration, Instant};
 
-/// The only continuation-token envelope version this build reads and writes.
+/// The continuation-token version used by ordinary progressing pages.
 pub const LIST_THROUGH_TOKEN_VERSION: u32 = 1;
+const LIST_THROUGH_PROGRESS_TOKEN_VERSION: u32 = 2;
+
+/// The sixteenth consecutive merged page without a key or new EOF fails.
+/// This also bounds legitimate sparse listings; it is not a cycle detector.
+pub const MAX_LIST_NO_PROGRESS_PAGES: u8 = 16;
 
 /// Envelope marker. A bucket that is *not* merging hands out the local
 /// listing's own marker, so the decoder needs a positive signal before it
 /// treats an opaque token as a merged one.
 const LIST_THROUGH_TOKEN_TAG: &str = "odm-list";
+// Object keys cannot contain NUL (bucket::utils::is_valid_object_prefix),
+// so this framing cannot collide with a local key used as an opaque marker.
+const LIST_THROUGH_TOKEN_PREFIX: &str = "\0odm-list:";
 
 /// Pages fetched per side per request: the first page, plus at most one refill
 /// when the first one was mostly consumed by the previous page. Two pages of
@@ -86,8 +94,7 @@ pub struct MergePick {
 }
 
 /// The continuation-token envelope. Opaque to clients: it is serialized as
-/// JSON and then base64-encoded by the same helper that encodes a plain local
+/// framed JSON and then base64-encoded by the same helper as a local marker.
-/// marker, so the wire shape is `base64(json)`.
 ///
 /// A `null` cursor with `done = false` means "list that side from the start";
 /// `done = true` means the side is finished and must not be listed again.
@@ -111,6 +118,10 @@ pub struct ListThroughToken {
     /// common prefix compares as itself, never as its members.
     #[serde(default)]
     pub last_key: Option<String>,
+    /// Consecutive empty truncated merged pages, present only in v2 tokens.
+    /// Ordinary v1 tokens retain their original serialized shape.
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub no_progress: Option<u8>,
 }
 
 impl ListThroughToken {
@@ -123,13 +134,14 @@ impl ListThroughToken {
             source: source.token,
             source_done: source.done,
             last_key,
+            no_progress: None,
         }
     }
 
     pub fn encode(&self) -> String {
         // The envelope is built here from owned strings, so serialization
         // cannot fail; the fallback keeps the signature infallible.
-        serde_json::to_string(self).unwrap_or_default()
+        format!("{LIST_THROUGH_TOKEN_PREFIX}{}", serde_json::to_string(self).unwrap_or_default())
     }
 }
 
@@ -153,24 +165,35 @@ pub enum ListThroughTokenError {
 
 /// Classifies an already base64-decoded continuation token.
 ///
-/// Only a JSON object carrying the envelope marker is read as a merged token;
+/// Only a framed JSON object is read as a merged token;
 /// anything else is a local marker, so a bucket that turns `list_through` off
 /// keeps paginating with the tokens it handed out. A token that *is* an
 /// envelope but was tampered with (unknown version, unknown field, truncated
 /// JSON) is an error, never a silent fallback.
 pub fn decode_continuation_token(decoded: &str) -> Result<ListThroughCursor, ListThroughTokenError> {
-    if !decoded.starts_with('{') {
+    let Some(payload) = decoded.strip_prefix(LIST_THROUGH_TOKEN_PREFIX) else {
-        return Ok(ListThroughCursor::Local(decoded.to_string()));
-    }
-    let Ok(value) = serde_json::from_str::<serde_json::Value>(decoded) else {
-        // Not JSON at all: an object key may legitimately start with '{'.
         return Ok(ListThroughCursor::Local(decoded.to_string()));
     };
+    let value = serde_json::from_str::<serde_json::Value>(payload).map_err(|_| ListThroughTokenError::Malformed)?;
     if value.get("t").and_then(serde_json::Value::as_str) != Some(LIST_THROUGH_TOKEN_TAG) {
-        return Ok(ListThroughCursor::Local(decoded.to_string()));
+        return Err(ListThroughTokenError::Malformed);
     }
     match value.get("v").and_then(serde_json::Value::as_u64) {
-        Some(version) if version == u64::from(LIST_THROUGH_TOKEN_VERSION) => {}
+        Some(version) if version == u64::from(LIST_THROUGH_TOKEN_VERSION) => {
+            // v1 readers reject this field even when it is null or zero.
+            if value.get("no_progress").is_some() {
+                return Err(ListThroughTokenError::Malformed);
+            }
+        }
+        Some(version) if version == u64::from(LIST_THROUGH_PROGRESS_TOKEN_VERSION) => {
+            if !value
+                .get("no_progress")
+                .and_then(serde_json::Value::as_u64)
+                .is_some_and(|count| (1..u64::from(MAX_LIST_NO_PROGRESS_PAGES)).contains(&count))
+            {
+                return Err(ListThroughTokenError::Malformed);
+            }
+        }
         Some(version) => return Err(ListThroughTokenError::UnsupportedVersion(version.min(u64::from(u32::MAX)) as u32)),
         None => return Err(ListThroughTokenError::Malformed),
     }
@@ -189,8 +212,8 @@ pub enum SourceListPlan {
     /// delimiter — the source's own roll-up boundary matches the request's.
     Page { prefix: String },
     /// `filter.prefix` reaches past a delimiter, so every key the source could
-    /// contribute rolls into this one common prefix. One bounded probe listing
+    /// contribute rolls into this one common prefix. Bounded probes follow
-    /// decides whether it exists; there is nothing to paginate.
+    /// empty progressing pages until a key proves existence or the source ends.
     Folded { probe_prefix: String, common_prefix: String },
 }
 
@@ -279,6 +302,31 @@ pub struct FetchRequest {
     pub token: Option<String>,
 }
 
+/// Invalid pagination metadata. Opaque cursor values are never included in errors.
+#[derive(Clone, Copy, Debug, PartialEq, Eq, thiserror::Error)]
+pub enum ListPageError {
+    #[error("truncated listing has no continuation token")]
+    Missing,
+    #[error("truncated listing has an empty continuation token")]
+    Empty,
+    #[error("truncated listing repeats a continuation token")]
+    Repeated,
+    #[error("listing exhausted its consecutive no-progress page budget")]
+    NoProgress(MergeSide),
+}
+
+pub(crate) fn validate_list_page(is_truncated: bool, token: Option<&str>, next_token: Option<&str>) -> Result<(), ListPageError> {
+    if is_truncated {
+        match next_token {
+            None => return Err(ListPageError::Missing),
+            Some("") => return Err(ListPageError::Empty),
+            Some(next) if Some(next) == token => return Err(ListPageError::Repeated),
+            Some(_) => {}
+        }
+    }
+    Ok(())
+}
+
 #[derive(Debug, Default)]
 struct SideState {
     start: SideCursor,
@@ -329,6 +377,7 @@ pub struct MergeOutcome {
 #[derive(Debug)]
 pub struct ListThroughMerger {
     max_keys: usize,
+    no_progress: Option<u8>,
     last_key: Option<String>,
     local: SideState,
     source: SideState,
@@ -348,6 +397,7 @@ impl ListThroughMerger {
         };
         Self {
             max_keys,
+            no_progress: token.and_then(|token| token.no_progress),
             last_key,
             local,
             source,
@@ -364,6 +414,11 @@ impl ListThroughMerger {
     /// or `filter.prefix` excludes it.
     pub fn disable_source(&mut self) {
         self.source.disabled = true;
+        // A refill can fail after a valid first page. A local-only response
+        // must discard both that source payload and its ordering horizon.
+        self.source.entries.clear();
+        self.source.pages.clear();
+        self.source.more = false;
     }
 
     pub fn next_fetch(&self) -> Option<FetchRequest> {
@@ -378,7 +433,13 @@ impl ListThroughMerger {
     /// Records one fetched page. `entries` must be sorted by `name` and already
     /// filtered with [`Self::accepts`]; the caller keeps the matching payloads
     /// in the same order.
-    pub fn push_page(&mut self, side: MergeSide, entries: Vec<ListEntryKey>, is_truncated: bool, next_token: Option<String>) {
+    pub fn push_page(
+        &mut self,
+        side: MergeSide,
+        entries: Vec<ListEntryKey>,
+        is_truncated: bool,
+        next_token: Option<String>,
+    ) -> Result<(), ListPageError> {
         let state = match side {
             MergeSide::Local => &mut self.local,
             MergeSide::Source => &mut self.source,
@@ -387,24 +448,33 @@ impl ListThroughMerger {
             Some(last) => last.next_token.clone(),
             None => state.start.token.clone(),
         };
-        // A truncated page without a cursor cannot be continued; treating the
+        validate_list_page(is_truncated, token.as_deref(), next_token.as_deref())?;
-        // side as finished is the only alternative to looping on it forever.
+        // Also reject a cycle through an earlier page in this bounded fetch.
-        state.more = is_truncated && next_token.is_some();
+        if is_truncated && state.pages.iter().any(|page| page.token == next_token) {
+            return Err(ListPageError::Repeated);
+        }
+        state.more = is_truncated;
         state.pages.push(FetchedPage {
             token,
             count: entries.len(),
             next_token: is_truncated.then_some(next_token).flatten(),
         });
         state.entries.extend(entries);
+        Ok(())
     }
 
-    pub fn finish(self) -> MergeOutcome {
+    /// `issue_progress_tokens` allows a v1 chain to start carrying a budget.
+    /// An existing v2 budget is always enforced, including on reader-only nodes.
+    /// Borrowing lets a source failure re-merge the fetched local buffers.
+    pub fn finish(&self, issue_progress_tokens: bool) -> Result<MergeOutcome, ListPageError> {
         let Self {
             max_keys,
+            no_progress,
             last_key,
             local,
             source,
         } = self;
+        let max_keys = *max_keys;
 
         // A side with more pages behind it can only be trusted up to the last
         // key it handed over: past that horizon the other side's entries could
@@ -470,12 +540,44 @@ impl ListThroughMerger {
         let source_left = !source.disabled && (!source_cursor.done || consumed_source < source.entries.len());
         let is_truncated = local_left || source_left;
 
-        let last_key = consumed_key.or(last_key);
+        let reached_eof = (!local.start.done && local_cursor.done) || (!source.start.done && source_cursor.done);
-        MergeOutcome {
+        let next_no_progress = if !is_truncated || !picks.is_empty() || reached_eof {
+            None
+        } else if max_keys == 0 {
+            // A zero-sized request cannot consume entries. Preserve an existing
+            // budget without spending it or starting a new one.
+            *no_progress
+        } else if issue_progress_tokens || no_progress.is_some() {
+            let count = no_progress.unwrap_or(0).saturating_add(1);
+            if count >= MAX_LIST_NO_PROGRESS_PAGES {
+                // An empty truncated side closes the merge horizon. Local
+                // failure takes precedence; disabling the source cannot fix it.
+                let side = if local.more && local.entries.is_empty() {
+                    MergeSide::Local
+                } else if !source.disabled && source.more && source.entries.is_empty() {
+                    MergeSide::Source
+                } else {
+                    MergeSide::Local
+                };
+                return Err(ListPageError::NoProgress(side));
+            }
+            Some(count)
+        } else {
+            None
+        };
+        let last_key = consumed_key.or_else(|| last_key.clone());
+        Ok(MergeOutcome {
             picks,
             is_truncated,
-            next_token: is_truncated.then(|| ListThroughToken::new(local_cursor, source_cursor, last_key)),
+            next_token: is_truncated.then(|| {
-        }
+                let mut token = ListThroughToken::new(local_cursor, source_cursor, last_key);
+                if let Some(count) = next_no_progress {
+                    token.v = LIST_THROUGH_PROGRESS_TOKEN_VERSION;
+                    token.no_progress = Some(count);
+                }
+                token
+            }),
+        })
     }
 }
 
@@ -599,9 +701,15 @@ mod tests {
                 let (entries, truncated, next) = reference_page(keys, prefix, delimiter, fetch.token.as_deref(), max_keys);
                 let kept: Vec<ListEntryKey> = entries.into_iter().filter(|entry| merger.accepts(&entry.name)).collect();
                 buffers[usize::from(fetch.side == MergeSide::Source)].extend(kept.iter().cloned());
-                merger.push_page(fetch.side, kept, truncated, next);
+                merger
+                    .push_page(fetch.side, kept, truncated, next)
+                    .expect("reference provider pages must advance");
+            }
+            let outcome = merger.finish(false).expect("valid merge outcome");
+            assert_eq!(outcome.is_truncated, outcome.next_token.is_some());
+            if outcome.is_truncated {
+                assert_ne!(outcome.next_token, token, "every truncated merged page must make progress");
             }
-            let outcome = merger.finish();
             page_sizes.push(outcome.picks.len());
             for pick in &outcome.picks {
                 let entry = buffers[usize::from(pick.side == MergeSide::Source)][pick.index].clone();
@@ -616,11 +724,25 @@ mod tests {
     }
 
     fn expected(local: &[String], source: &[String], prefix: &str, delimiter: Option<&str>) -> Vec<ListEntryKey> {
-        let mut all: Vec<String> = local.iter().chain(source.iter()).cloned().collect();
+        // This oracle builds the complete namespace independently of the
-        all.sort();
+        // provider's page/marker helper and the production merger.
-        all.dedup();
+        let mut namespace = std::collections::BTreeMap::new();
-        let (entries, _, _) = reference_page(&all, prefix, delimiter, None, usize::MAX);
+        for key in local.iter().chain(source) {
-        entries
+            let Some(suffix) = key.strip_prefix(prefix) else {
+                continue;
+            };
+            if let Some(delimiter) = delimiter.filter(|delimiter| !delimiter.is_empty())
+                && let Some((directory, _)) = suffix.split_once(delimiter)
+            {
+                namespace.insert(format!("{prefix}{directory}{delimiter}"), true);
+                continue;
+            }
+            namespace.insert(key.clone(), false);
+        }
+        namespace
+            .into_iter()
+            .map(|(name, is_prefix)| ListEntryKey { name, is_prefix })
+            .collect()
     }
 
     #[test]
@@ -662,9 +784,11 @@ mod tests {
                 token: None
             })
         );
-        merger.push_page(MergeSide::Local, vec![ListEntryKey::object("a")], false, None);
+        merger
+            .push_page(MergeSide::Local, vec![ListEntryKey::object("a")], false, None)
+            .expect("local EOF is valid");
         assert_eq!(merger.next_fetch(), None);
-        let outcome = merger.finish();
+        let outcome = merger.finish(false).expect("valid merge outcome");
         assert_eq!(outcome.picks.len(), 1);
         assert!(!outcome.is_truncated);
         assert!(outcome.next_token.is_none());
@@ -680,16 +804,19 @@ mod tests {
             source: Some("source-1".to_string()),
             source_done: false,
             last_key: Some("a".to_string()),
+            no_progress: None,
         };
         let mut merger = ListThroughMerger::new(1, Some(&resume));
         merger.disable_source();
-        merger.push_page(
+        merger
-            MergeSide::Local,
+            .push_page(
-            vec![ListEntryKey::object("b"), ListEntryKey::object("c")],
+                MergeSide::Local,
-            true,
+                vec![ListEntryKey::object("b"), ListEntryKey::object("c")],
-            Some("local-2".to_string()),
+                true,
-        );
+                Some("local-2".to_string()),
-        let outcome = merger.finish();
+            )
+            .expect("local cursor advances");
+        let outcome = merger.finish(false).expect("valid merge outcome");
         assert!(outcome.is_truncated);
         let token = outcome.next_token.expect("truncated page carries a token");
         assert_eq!(token.source.as_deref(), Some("source-1"), "the source cursor must not move");
@@ -698,6 +825,212 @@ mod tests {
         assert_eq!(token.local.as_deref(), Some("local-1"), "a partly read page is re-listed");
     }
 
+    #[test]
+    fn truncated_pages_require_a_nonempty_advancing_cursor() {
+        for side in [MergeSide::Local, MergeSide::Source] {
+            for entries in [vec![], vec![ListEntryKey::object("a")]] {
+                for (next, expected) in [
+                    (None, Err(ListPageError::Missing)),
+                    (Some(""), Err(ListPageError::Empty)),
+                    (Some("stuck"), Err(ListPageError::Repeated)),
+                    (Some("advances"), Ok(())),
+                ] {
+                    let resume = ListThroughToken::new(
+                        SideCursor {
+                            token: Some("stuck".into()),
+                            done: false,
+                        },
+                        SideCursor {
+                            token: Some("stuck".into()),
+                            done: false,
+                        },
+                        None,
+                    );
+                    let mut merger = ListThroughMerger::new(2, Some(&resume));
+                    let result = merger.push_page(side, entries.clone(), true, next.map(str::to_string));
+                    assert_eq!(result, expected, "{side:?}, {entries:?}, {next:?}");
+                    let state = if side == MergeSide::Local {
+                        &merger.local
+                    } else {
+                        &merger.source
+                    };
+                    assert_eq!(state.pages.len(), usize::from(result.is_ok()), "invalid page must not be accepted");
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn repeated_empty_cursor_is_rejected_before_an_identical_page_can_escape() {
+        let resume = ListThroughToken::new(
+            SideCursor { token: None, done: true },
+            SideCursor {
+                token: Some("stuck".into()),
+                done: false,
+            },
+            None,
+        );
+        let mut merger = ListThroughMerger::new(2, Some(&resume));
+        assert_eq!(
+            merger.next_fetch(),
+            Some(FetchRequest {
+                side: MergeSide::Source,
+                token: Some("stuck".into())
+            })
+        );
+        assert_eq!(
+            merger.push_page(MergeSide::Source, vec![], true, Some("stuck".into())),
+            Err(ListPageError::Repeated)
+        );
+    }
+
+    #[test]
+    fn empty_pages_may_advance_within_the_fetch_budget_until_eof() {
+        let mut merger = ListThroughMerger::new(2, None);
+        merger.push_page(MergeSide::Local, vec![], false, None).expect("local EOF");
+        for next in ["opaque-z", "opaque-a"] {
+            assert_eq!(merger.next_fetch().expect("bounded source fetch").side, MergeSide::Source);
+            merger
+                .push_page(MergeSide::Source, vec![], true, Some(next.into()))
+                .expect("opaque cursor advances regardless of sort order");
+        }
+        assert!(merger.next_fetch().is_none(), "two source fetches exhaust the request budget");
+        let outcome = merger.finish(false).expect("valid merge outcome");
+        assert!(outcome.picks.is_empty());
+        assert!(outcome.is_truncated);
+        let token = outcome.next_token.expect("empty progressing page has a cursor");
+        assert_eq!(token.source.as_deref(), Some("opaque-a"));
+        let mut merger = ListThroughMerger::new(2, Some(&token));
+        assert_eq!(merger.next_fetch().expect("source resumes").token.as_deref(), Some("opaque-a"));
+        merger
+            .push_page(MergeSide::Source, vec![ListEntryKey::object("result")], false, None)
+            .expect("source EOF");
+        let outcome = merger.finish(false).expect("valid merge outcome");
+        assert_eq!(
+            outcome.picks,
+            vec![MergePick {
+                side: MergeSide::Source,
+                index: 0
+            }]
+        );
+        assert!(!outcome.is_truncated);
+        assert!(outcome.next_token.is_none());
+    }
+
+    #[test]
+    fn a_cursor_cycle_inside_the_fetch_budget_is_rejected() {
+        let resume = ListThroughToken::new(
+            SideCursor { token: None, done: true },
+            SideCursor {
+                token: Some("first".into()),
+                done: false,
+            },
+            None,
+        );
+        let mut merger = ListThroughMerger::new(2, Some(&resume));
+        merger
+            .push_page(MergeSide::Source, vec![], true, Some("second".into()))
+            .expect("first page advances");
+        assert_eq!(
+            merger.push_page(MergeSide::Source, vec![], true, Some("first".into())),
+            Err(ListPageError::Repeated)
+        );
+    }
+
+    #[test]
+    fn source_refill_failure_discards_buffered_source_entries_and_horizon() {
+        let mut merger = ListThroughMerger::new(2, None);
+        merger
+            .push_page(MergeSide::Local, vec![ListEntryKey::object("z")], false, None)
+            .expect("local EOF");
+        merger
+            .push_page(MergeSide::Source, vec![ListEntryKey::object("a")], true, Some("stuck".into()))
+            .expect("first source page advances");
+        assert_eq!(merger.next_fetch().expect("source refill is required").token.as_deref(), Some("stuck"));
+        assert_eq!(
+            merger.push_page(MergeSide::Source, vec![], true, Some("stuck".into())),
+            Err(ListPageError::Repeated)
+        );
+        merger.disable_source();
+        let outcome = merger.finish(false).expect("valid merge outcome");
+        assert_eq!(
+            outcome.picks,
+            vec![MergePick {
+                side: MergeSide::Local,
+                index: 0
+            }]
+        );
+        assert!(!outcome.is_truncated);
+        assert!(outcome.next_token.is_none());
+    }
+
+    #[test]
+    fn list_through_static_namespace_boundary_matrix() {
+        let corpus = [
+            "a",
+            "a/",
+            "a/b",
+            "a/b/child",
+            "a0",
+            "b",
+            "b/leaf",
+            "quote\"&<",
+            "space key",
+            "z",
+            "é",
+            "中/文",
+        ];
+        for count in [0, 1, 3, 4, corpus.len()] {
+            let keys: Vec<String> = corpus[..count].iter().map(|key| (*key).to_string()).collect();
+            for placement in 0..3 {
+                let (local, source): (Vec<_>, Vec<_>) =
+                    keys.iter()
+                        .enumerate()
+                        .fold((vec![], vec![]), |(mut local, mut source), (index, key)| {
+                            if placement != 1 || index % 2 == 0 {
+                                local.push(key.clone());
+                            }
+                            if placement != 0 || index % 2 == 0 {
+                                source.push(key.clone());
+                            }
+                            (local, source)
+                        });
+                for prefix in ["", "a", "a/", "中/"] {
+                    for delimiter in [None, Some("/")] {
+                        for max_keys in [1, 3, 4] {
+                            let oracle = expected(&local, &source, prefix, delimiter);
+                            let (emitted, sizes) = walk(&local, &source, prefix, delimiter, max_keys);
+                            assert_eq!(
+                                emitted.iter().map(|(entry, _)| entry.clone()).collect::<Vec<_>>(),
+                                oracle,
+                                "count={count}, placement={placement}, prefix={prefix}, delimiter={delimiter:?}, max={max_keys}"
+                            );
+                            let expected_sizes: Vec<_> = if oracle.is_empty() {
+                                vec![0]
+                            } else {
+                                oracle.chunks(max_keys).map(<[ListEntryKey]>::len).collect()
+                            };
+                            assert_eq!(sizes, expected_sizes, "exact max and max+1 boundaries must agree");
+                        }
+                    }
+                }
+            }
+        }
+    }
+
+    #[test]
+    fn list_through_large_overlap_walk_keeps_all_5300_keys() {
+        let source: Vec<_> = (0..5000).map(|index| format!("k{index:05}")).collect();
+        let local: Vec<_> = (4800..5300).map(|index| format!("k{index:05}")).collect();
+        let (emitted, sizes) = walk(&local, &source, "", None, 333);
+        assert_eq!(emitted.len(), 5300);
+        for (index, (entry, side)) in emitted.iter().enumerate() {
+            assert_eq!(entry.name, format!("k{index:05}"));
+            assert_eq!(*side, if index >= 4800 { MergeSide::Local } else { MergeSide::Source });
+        }
+        assert_eq!(sizes, [vec![333; 15], vec![305]].concat());
+    }
+
     #[test]
     fn token_round_trips_and_rejects_tampering() {
         let token = ListThroughToken::new(
@@ -711,21 +1044,279 @@ mod tests {
         let encoded = token.encode();
         assert_eq!(decode_continuation_token(&encoded), Ok(ListThroughCursor::Merged(Box::new(token))));
 
-        let bumped = encoded.replace("\"v\":1", "\"v\":2");
+        let bumped = encoded.replace("\"v\":1", "\"v\":3");
-        assert_eq!(decode_continuation_token(&bumped), Err(ListThroughTokenError::UnsupportedVersion(2)));
+        assert_eq!(decode_continuation_token(&bumped), Err(ListThroughTokenError::UnsupportedVersion(3)));
 
         let extra = encoded.replace("{", "{\"x\":1,");
         assert_eq!(decode_continuation_token(&extra), Err(ListThroughTokenError::Malformed));
 
         let truncated = &encoded[..encoded.len() - 3];
-        assert_eq!(decode_continuation_token(truncated), Ok(ListThroughCursor::Local(truncated.to_string())));
+        assert_eq!(decode_continuation_token(truncated), Err(ListThroughTokenError::Malformed));
 
-        let no_version = "{\"t\":\"odm-list\"}";
+        let no_version = "\0odm-list:{\"t\":\"odm-list\"}";
         assert_eq!(decode_continuation_token(no_version), Err(ListThroughTokenError::Malformed));
     }
 
+    fn progress_token(count: Option<u8>, local_done: bool, source_done: bool) -> ListThroughToken {
+        let mut token = ListThroughToken::new(
+            SideCursor {
+                token: None,
+                done: local_done,
+            },
+            SideCursor {
+                token: Some("A".into()),
+                done: source_done,
+            },
+            Some("last-key".into()),
+        );
+        if let Some(count) = count {
+            token.v = LIST_THROUGH_PROGRESS_TOKEN_VERSION;
+            token.no_progress = Some(count);
+        }
+        token
+    }
+
+    fn push_empty_pages(merger: &mut ListThroughMerger, side: MergeSide) {
+        for _ in 0..MAX_LIST_FETCHES_PER_SIDE {
+            let fetch = merger.next_fetch().expect("empty truncated side must be fetched");
+            assert_eq!(fetch.side, side);
+            let next = format!("{}:next", fetch.token.unwrap_or_default());
+            merger
+                .push_page(side, vec![], true, Some(next))
+                .expect("opaque cursor advances");
+        }
+    }
+
+    #[test]
+    fn progress_tokens_preserve_v1_bytes_and_validate_v2_counts() {
+        let token = progress_token(None, true, false);
+        assert_eq!(
+            token.encode(),
+            r#"{"t":"odm-list","v":1,"local":null,"local_done":true,"source":"A","source_done":false,"last_key":"last-key"}"#
+        );
+        for count in 1..MAX_LIST_NO_PROGRESS_PAGES {
+            let token = progress_token(Some(count), true, false);
+            assert_eq!(decode_continuation_token(&token.encode()), Ok(ListThroughCursor::Merged(Box::new(token))));
+        }
+        for version in [1, 2] {
+            for value in ["null", "0", "16", "-1", "1.5", "256", "18446744073709551616", "\"1\""] {
+                let encoded = format!(r#"{{"t":"odm-list","v":{version},"no_progress":{value}}}"#);
+                assert_eq!(decode_continuation_token(&encoded), Err(ListThroughTokenError::Malformed), "{encoded}");
+            }
+        }
+        for encoded in [
+            r#"{"t":"odm-list","v":1,"no_progress":1}"#,
+            r#"{"t":"odm-list","v":2}"#,
+            r#"{"t":"odm-list","v":2,"no_progress":1,"extra":true}"#,
+        ] {
+            assert_eq!(decode_continuation_token(encoded), Err(ListThroughTokenError::Malformed), "{encoded}");
+        }
+    }
+
+    #[test]
+    fn reader_only_nodes_do_not_start_a_budget_but_mixed_readers_preserve_one() {
+        let mut token = progress_token(None, true, false);
+        for _ in 0..MAX_LIST_NO_PROGRESS_PAGES {
+            let mut merger = ListThroughMerger::new(2, Some(&token));
+            push_empty_pages(&mut merger, MergeSide::Source);
+            token = merger
+                .finish(false)
+                .expect("reader-only v1 behavior")
+                .next_token
+                .expect("truncated cursor");
+            assert_eq!(token.v, 1);
+            assert_eq!(token.no_progress, None);
+        }
+        for count in 1..=MAX_LIST_NO_PROGRESS_PAGES {
+            let mut merger = ListThroughMerger::new(2, Some(&token));
+            push_empty_pages(&mut merger, MergeSide::Source);
+            assert!(merger.next_fetch().is_none(), "the per-request two-fetch limit stays intact");
+            let outcome = merger.finish(count % 2 == 1);
+            if count == MAX_LIST_NO_PROGRESS_PAGES {
+                assert_eq!(outcome, Err(ListPageError::NoProgress(MergeSide::Source)));
+                break;
+            }
+            token = outcome.expect("budget not exhausted").next_token.expect("truncated cursor");
+            assert_eq!(token.no_progress, Some(count));
+            let ListThroughCursor::Merged(decoded) = decode_continuation_token(&token.encode()).expect("round-trip v2") else {
+                panic!("merged cursor expected");
+            };
+            token = *decoded;
+        }
+    }
+
+    #[test]
+    fn objects_and_common_prefixes_reset_a_budget_at_the_boundary() {
+        for entry in [ListEntryKey::object("result"), ListEntryKey::prefix("result/")] {
+            for issue_tokens in [false, true] {
+                let resume = progress_token(Some(MAX_LIST_NO_PROGRESS_PAGES - 1), true, false);
+                let mut merger = ListThroughMerger::new(2, Some(&resume));
+                merger
+                    .push_page(MergeSide::Source, vec![], true, Some("B".into()))
+                    .expect("empty advancing page");
+                merger
+                    .push_page(MergeSide::Source, vec![entry.clone()], true, Some("C".into()))
+                    .expect("real progress");
+                let outcome = merger
+                    .finish(issue_tokens)
+                    .expect("real progress does not exhaust the budget");
+                assert_eq!(
+                    outcome.picks,
+                    vec![MergePick {
+                        side: MergeSide::Source,
+                        index: 0
+                    }]
+                );
+                let next = outcome.next_token.expect("source remains truncated");
+                assert_eq!(next.last_key.as_deref(), Some(entry.name.as_str()));
+                assert_eq!(next.v, 1);
+                assert_eq!(next.no_progress, None);
+                assert!(!next.encode().contains("no_progress"));
+            }
+        }
+    }
+
+    #[test]
+    fn only_a_new_eof_transition_resets_the_empty_page_budget() {
+        for finished_side in [MergeSide::Local, MergeSide::Source] {
+            let resume = progress_token(Some(MAX_LIST_NO_PROGRESS_PAGES - 1), false, false);
+            let mut merger = ListThroughMerger::new(2, Some(&resume));
+            if finished_side == MergeSide::Local {
+                merger
+                    .push_page(MergeSide::Local, vec![], false, None)
+                    .expect("new local EOF");
+                push_empty_pages(&mut merger, MergeSide::Source);
+            } else {
+                push_empty_pages(&mut merger, MergeSide::Local);
+                merger
+                    .push_page(MergeSide::Source, vec![], false, None)
+                    .expect("new source EOF");
+            }
+            let next = merger
+                .finish(false)
+                .expect("new EOF is progress")
+                .next_token
+                .expect("other side truncated");
+            assert_eq!(next.no_progress, None);
+            assert_eq!(next.v, 1);
+            assert_eq!(next.local_done, finished_side == MergeSide::Local);
+            assert_eq!(next.source_done, finished_side == MergeSide::Source);
+            let mut merger = ListThroughMerger::new(2, Some(&next));
+            let remaining = if finished_side == MergeSide::Local {
+                MergeSide::Source
+            } else {
+                MergeSide::Local
+            };
+            push_empty_pages(&mut merger, remaining);
+            let next = merger
+                .finish(true)
+                .expect("a new budget starts")
+                .next_token
+                .expect("truncated");
+            assert_eq!(next.no_progress, Some(1), "an already-done side cannot reset every page");
+        }
+        let resume = progress_token(Some(MAX_LIST_NO_PROGRESS_PAGES - 1), true, false);
+        let mut merger = ListThroughMerger::new(2, Some(&resume));
+        merger.push_page(MergeSide::Source, vec![], false, None).expect("final EOF");
+        let outcome = merger.finish(false).expect("EOF succeeds at the budget boundary");
+        assert!(!outcome.is_truncated);
+        assert!(outcome.next_token.is_none());
+    }
+
+    #[test]
+    fn filtered_duplicates_cannot_reset_the_no_progress_budget() {
+        let resume = progress_token(Some(MAX_LIST_NO_PROGRESS_PAGES - 1), true, false);
+        let mut merger = ListThroughMerger::new(2, Some(&resume));
+        for next in ["B", "C"] {
+            let entries = [ListEntryKey::object("last-key"), ListEntryKey::object("earlier")]
+                .into_iter()
+                .filter(|entry| merger.accepts(&entry.name))
+                .collect::<Vec<_>>();
+            assert!(entries.is_empty(), "both provider entries were already consumed");
+            merger
+                .push_page(MergeSide::Source, entries, true, Some(next.into()))
+                .expect("advancing cursor");
+        }
+        assert_eq!(merger.finish(false), Err(ListPageError::NoProgress(MergeSide::Source)));
+    }
+
+    #[test]
+    fn no_progress_is_attributed_to_local_when_source_cannot_unblock_it() {
+        for source_mode in ["disabled", "done", "empty", "data"] {
+            let resume = progress_token(Some(MAX_LIST_NO_PROGRESS_PAGES - 1), false, source_mode == "done");
+            let mut merger = ListThroughMerger::new(2, Some(&resume));
+            if source_mode == "disabled" {
+                merger.disable_source();
+            }
+            push_empty_pages(&mut merger, MergeSide::Local);
+            match source_mode {
+                "empty" => push_empty_pages(&mut merger, MergeSide::Source),
+                "data" => merger
+                    .push_page(MergeSide::Source, vec![ListEntryKey::object("source")], false, None)
+                    .expect("source data"),
+                _ => {}
+            }
+            assert_eq!(merger.finish(false), Err(ListPageError::NoProgress(MergeSide::Local)), "{source_mode}");
+        }
+    }
+
+    #[test]
+    fn source_budget_failure_remerges_local_objects_and_prefixes_without_refetching() {
+        let resume = progress_token(Some(MAX_LIST_NO_PROGRESS_PAGES - 1), false, false);
+        let mut merger = ListThroughMerger::new(2, Some(&resume));
+        merger
+            .push_page(MergeSide::Local, vec![ListEntryKey::object("local")], true, Some("L1".into()))
+            .expect("local object");
+        merger
+            .push_page(MergeSide::Local, vec![ListEntryKey::prefix("prefix/")], true, Some("L2".into()))
+            .expect("local prefix");
+        push_empty_pages(&mut merger, MergeSide::Source);
+        assert_eq!(merger.finish(false), Err(ListPageError::NoProgress(MergeSide::Source)));
+        merger.disable_source();
+        assert!(merger.next_fetch().is_none(), "fallback does not perform another fetch");
+        let outcome = merger.finish(false).expect("local data makes progress");
+        assert_eq!(
+            outcome.picks,
+            vec![
+                MergePick {
+                    side: MergeSide::Local,
+                    index: 0
+                },
+                MergePick {
+                    side: MergeSide::Local,
+                    index: 1
+                }
+            ]
+        );
+        let token = outcome.next_token.expect("remaining local page");
+        assert_eq!(token.local.as_deref(), Some("L2"));
+        assert_eq!(token.source.as_deref(), Some("A"));
+        assert_eq!(token.last_key.as_deref(), Some("prefix/"));
+        assert_eq!(token.no_progress, None);
+        assert_eq!(token.v, 1);
+    }
+
+    #[test]
+    fn a_zero_sized_merge_preserves_an_existing_budget() {
+        let resume = progress_token(Some(MAX_LIST_NO_PROGRESS_PAGES - 1), true, false);
+        let mut merger = ListThroughMerger::new(0, Some(&resume));
+        merger
+            .push_page(MergeSide::Source, vec![ListEntryKey::object("result")], true, Some("B".into()))
+            .expect("source page");
+        let outcome = merger.finish(false).expect("a zero-sized request cannot consume entries");
+        assert!(outcome.picks.is_empty());
+        assert_eq!(outcome.next_token.expect("unconsumed source").no_progress, resume.no_progress);
+    }
+
     #[test]
     fn a_plain_local_marker_stays_local() {
+        for marker in [
+            r#"{"t":"odm-list","v":1}"#,
+            r#"{"t":"odm-list","v":2,"local_done":true}"#,
+            r#"{"t":"odm-list"}"#,
+        ] {
+            assert_eq!(decode_continuation_token(marker), Ok(ListThroughCursor::Local(marker.to_string())));
+        }
         assert_eq!(
             decode_continuation_token("photos/2024/01.jpg"),
             Ok(ListThroughCursor::Local("photos/2024/01.jpg".to_string()))
@@ -796,7 +1387,10 @@ mod tests {
     }
 
     proptest! {
-        #![proptest_config(ProptestConfig::with_cases(256))]
+        #![proptest_config(ProptestConfig {
+            rng_seed: proptest::test_runner::RngSeed::Fixed(0xec5706),
+            ..ProptestConfig::with_cases(256)
+        })]
 
         /// Full pagination of a merged listing equals the sorted, deduplicated
         /// union of both sides, with every shared key served by local, and no

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

@@ -19,30 +19,42 @@
 //! client, and the per-node runtime (`sys`) that turns configs into live
 //! clients guarded by a breaker, a negative cache, singleflight and a pull
 //! concurrency limit (rustfs/backlog#2147).
+//!
+//! A source is reached through one `SourceBackend`: the S3 dialect for every
+//! S3-compatible provider, and a native backend for the providers that have no
+//! S3 API (`azure`, `gcs_native`).
 
+pub mod azure;
+#[cfg(test)]
+mod backend_contract;
 pub mod backfill;
 pub mod breaker;
 pub mod config;
+pub mod gcs;
 pub mod list_through;
+mod native_http;
 pub mod negative_cache;
 pub mod pull;
 pub mod source_client;
 pub mod stats;
 pub mod sys;
+#[cfg(test)]
+mod test_http_fixture;
 
 pub use breaker::{
     BREAKER_FAILURE_THRESHOLD, BREAKER_FAILURE_WINDOW, BREAKER_HALF_OPEN_MAX_PROBES, BREAKER_OPEN_DURATION, Breaker,
     BreakerState, BreakerTransition, BreakerVerdict,
 };
 pub use config::{
-    ConfigPublishHook, FilterConfig, HeadPolicy, ON_DEMAND_MIGRATION_CONFIG_HOOK, ON_DEMAND_MIGRATION_CONFIG_VERSION,
+    AzureSourceConfig, ConfigPublishHook, FilterConfig, GcsSourceConfig, HeadPolicy, ON_DEMAND_MIGRATION_CONFIG_HOOK,
-    OnDemandMigrationConfig, OnDemandMigrationConfigError, PathStyle, PolicyConfig, Provider, RangeGetPolicy, SourceConfig,
+    ON_DEMAND_MIGRATION_CONFIG_VERSION, OnDemandMigrationConfig, OnDemandMigrationConfigError, PathStyle, PolicyConfig, Provider,
-    SourceCredentials, SourceErrorPolicy, SourceTimeout, TlsConfig, ValidationContext,
+    RangeGetPolicy, SourceConfig, SourceCredentials, SourceErrorPolicy, SourceTimeout, TlsConfig, ValidationContext,
 };
 pub use list_through::{
-    FetchRequest, LIST_THROUGH_TOKEN_VERSION, ListEntryKey, ListThroughCursor, ListThroughMerger, ListThroughToken,
+    FetchRequest, LIST_THROUGH_TOKEN_VERSION, ListEntryKey, ListPageError, ListThroughCursor, ListThroughMerger,
-    ListThroughTokenError, MAX_LIST_FETCHES_PER_SIDE, MergeOutcome, MergePick, MergeSide, SOURCE_LIST_MAX_RATE_WAIT,
+    ListThroughToken, ListThroughTokenError, MAX_LIST_FETCHES_PER_SIDE, MAX_LIST_NO_PROGRESS_PAGES, MergeOutcome, MergePick,
-    SOURCE_LIST_RATE_PER_SEC, SourceListPlan, SourceListRateLimiter, decode_continuation_token, source_list_plan,
+    MergeSide, SOURCE_LIST_MAX_RATE_WAIT, SOURCE_LIST_RATE_PER_SEC, SourceListPlan, SourceListRateLimiter,
+    decode_continuation_token, source_list_plan,
 };
 pub use negative_cache::{NEGATIVE_CACHE_MAX_ENTRIES, NegativeCache};
 pub use pull::{
@@ -56,5 +68,6 @@ pub use stats::{
 };
 pub use sys::{
     ApplyOutcome, BucketOdmState, GLOBAL_ON_DEMAND_MIGRATION_SYS, OdmBucketSnapshot, OdmLookup, OdmStateError,
-    OnDemandMigrationSys, PullError, PullFollower, PullLeader, PullOutcome, PullResult, PullSlot, source_client_spec,
+    OnDemandMigrationSys, PullError, PullFollower, PullLeader, PullOutcome, PullResult, PullSlot, source_backend_spec,
+    source_client_spec,
 };

crates/ecstore/src/bucket/on_demand_migration/native_http.rs

@@ -0,0 +1,415 @@
+// 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.
+
+//! Shared HTTP transport for the on-demand migration source backends that do
+//! not speak S3 (Azure Blob, native GCS).
+//!
+//! The S3 backend rides the AWS SDK; these providers have no SigV4 dialect, so
+//! they talk plain HTTP through one `reqwest` client that carries the same
+//! connect/read timeouts and TLS policy the operator configured for the source.
+//! Redirects are refused: the endpoint passed the outbound policy gate once, and
+//! following a source-chosen `Location` would leave that gate behind.
+//!
+//! Errors never render the request URL. A SAS token lives in the query string,
+//! so a `reqwest` error rendered with its URL would print the credential into
+//! the log line and the admin response.
+
+use super::source_client::{SourceError, SourceHead, SourceTimeouts, USER_AGENT_SUFFIX, classify_status, is_multipart_etag};
+use crate::bucket::remote_s3_client::{RemoteS3ClientError, validate_remote_endpoint, validate_target_ca_pem};
+use aws_sdk_s3::primitives::ByteStream;
+use aws_smithy_types::body::SdkBody;
+use futures::StreamExt;
+use http::HeaderMap;
+use std::collections::HashMap;
+use std::time::SystemTime;
+use time::OffsetDateTime;
+use time::format_description::well_known::{Rfc2822, Rfc3339};
+use url::Url;
+
+/// Origin the native backends are allowed to address, plus the HTTP client
+/// that reaches it.
+pub(super) struct NativeHttp {
+    client: reqwest::Client,
+    endpoint: Url,
+}
+
+impl NativeHttp {
+    /// `endpoint` must be a bare `scheme://host[:port]` origin; it is checked
+    /// against the outbound policy exactly like an S3 source endpoint.
+    pub(super) fn new(
+        endpoint: &str,
+        timeouts: SourceTimeouts,
+        skip_tls_verify: bool,
+        ca_cert_pem: Option<&str>,
+    ) -> Result<Self, RemoteS3ClientError> {
+        let endpoint = Url::parse(endpoint.trim()).map_err(|err| RemoteS3ClientError::InvalidEndpoint(err.to_string()))?;
+        if !matches!(endpoint.scheme(), "http" | "https") {
+            return Err(RemoteS3ClientError::InvalidEndpoint(format!(
+                "unsupported scheme {}; expected http or https",
+                endpoint.scheme()
+            )));
+        }
+        if endpoint.host_str().is_none_or(str::is_empty) {
+            return Err(RemoteS3ClientError::InvalidEndpoint("endpoint has no host".to_string()));
+        }
+        if !endpoint.username().is_empty() || endpoint.password().is_some() {
+            return Err(RemoteS3ClientError::InvalidEndpoint("endpoint must not carry userinfo".to_string()));
+        }
+        if !matches!(endpoint.path(), "" | "/") || endpoint.query().is_some() || endpoint.fragment().is_some() {
+            return Err(RemoteS3ClientError::InvalidEndpoint(
+                "endpoint must be an origin without path, query or fragment".to_string(),
+            ));
+        }
+        validate_remote_endpoint(&endpoint).map_err(RemoteS3ClientError::EndpointNotAllowed)?;
+
+        let mut builder = reqwest::Client::builder()
+            .connect_timeout(timeouts.connect)
+            .read_timeout(timeouts.read)
+            .redirect(reqwest::redirect::Policy::none())
+            .user_agent(USER_AGENT_SUFFIX);
+        if skip_tls_verify {
+            builder = builder.danger_accept_invalid_certs(true);
+        } else if let Some(pem) = ca_cert_pem.map(str::trim).filter(|pem| !pem.is_empty()) {
+            // Reject a malformed bundle the same way the S3 path does, so the
+            // operator sees "invalid CA PEM" instead of a TLS handshake failure.
+            validate_target_ca_pem(pem)?;
+            let certificate = reqwest::Certificate::from_pem(pem.as_bytes())
+                .map_err(|err| RemoteS3ClientError::InvalidCaPem(err.to_string()))?;
+            builder = builder.add_root_certificate(certificate);
+        }
+
+        let client = builder
+            .build()
+            .map_err(|err| RemoteS3ClientError::InvalidEndpoint(format!("http client cannot be built: {err}")))?;
+        Ok(Self { client, endpoint })
+    }
+
+    #[cfg(test)]
+    pub(super) fn for_test(endpoint: Url) -> Self {
+        Self {
+            client: reqwest::Client::builder()
+                .redirect(reqwest::redirect::Policy::none())
+                .build()
+                .expect("test http client should build"),
+            endpoint,
+        }
+    }
+
+    /// A URL under the endpoint origin. `segments` are percent-encoded as
+    /// path segments, so a key containing `?`, `#` or a space cannot rewrite
+    /// the request target.
+    pub(super) fn url<'a>(&self, segments: impl IntoIterator<Item = &'a str>) -> Result<Url, SourceError> {
+        let mut url = self.endpoint.clone();
+        {
+            let mut path = url
+                .path_segments_mut()
+                .map_err(|_| SourceError::Other("source endpoint cannot carry a path".to_string()))?;
+            path.clear();
+            path.extend(segments);
+        }
+        Ok(url)
+    }
+
+    /// Sends the request and returns the response only for a 2xx status.
+    /// Non-2xx statuses are classified from the status and the provider's own
+    /// error-code header; response bodies are not read, so no provider message
+    /// can smuggle credentials or markup into a log line.
+    pub(super) async fn send(
+        &self,
+        request: reqwest::Request,
+        error_code_header: &str,
+    ) -> Result<reqwest::Response, SourceError> {
+        let response = self.client.execute(request).await.map_err(classify_transport_error)?;
+        let status = response.status();
+        if status.is_success() {
+            return Ok(response);
+        }
+        let code = response
+            .headers()
+            .get(error_code_header)
+            .and_then(|value| value.to_str().ok())
+            .map(str::to_string);
+        Err(classify_status(
+            status.as_u16(),
+            None,
+            match &code {
+                Some(code) => format!("source returned HTTP {status} ({code})"),
+                None => format!("source returned HTTP {status}"),
+            },
+        ))
+    }
+}
+
+/// Renders a transport failure without the request URL: a SAS token or a
+/// signed query would otherwise reach logs and admin responses.
+pub(super) fn classify_transport_error(err: reqwest::Error) -> SourceError {
+    let is_timeout = err.is_timeout();
+    let is_connect = err.is_connect();
+    let message = err.without_url().to_string();
+    if is_timeout {
+        SourceError::Timeout
+    } else if is_connect {
+        SourceError::Connect(message)
+    } else {
+        SourceError::Other(message)
+    }
+}
+
+/// Streams the response body without buffering it.
+pub(super) fn response_body(response: reqwest::Response) -> ByteStream {
+    let stream = response.bytes_stream().map(|chunk| {
+        chunk
+            .map(http_body::Frame::data)
+            .map_err(|err| std::io::Error::other(err.without_url().to_string()))
+    });
+    ByteStream::new(SdkBody::from_body_1_x(http_body_util::StreamBody::new(stream)))
+}
+
+/// Reads a bounded response body as UTF-8, for the XML and JSON listings.
+pub(super) async fn read_text(response: reqwest::Response, max_bytes: usize) -> Result<String, SourceError> {
+    let mut body = Vec::new();
+    let mut stream = response.bytes_stream();
+    while let Some(chunk) = stream.next().await {
+        let chunk = chunk.map_err(classify_transport_error)?;
+        if body.len().saturating_add(chunk.len()) > max_bytes {
+            return Err(SourceError::Other("source listing response exceeded the size limit".to_string()));
+        }
+        body.extend_from_slice(&chunk);
+    }
+    String::from_utf8(body).map_err(|_| SourceError::Other("source listing response is not valid UTF-8".to_string()))
+}
+
+/// Base64 digest (`Content-MD5`, `md5Hash`, `x-goog-hash`) as lowercase hex.
+/// `None` when the value is not a 16-byte digest, so a CRC32C never passes as
+/// an MD5.
+pub(super) fn base64_md5_to_hex(value: &str) -> Option<String> {
+    let raw = base64_simd::STANDARD.decode_to_vec(value.trim().as_bytes()).ok()?;
+    (raw.len() == 16).then(|| faster_hex::hex_string(&raw))
+}
+
+pub(super) fn header<'a>(headers: &'a HeaderMap, name: &str) -> Option<&'a str> {
+    headers.get(name).and_then(|value| value.to_str().ok()).map(str::trim)
+}
+
+fn header_string(headers: &HeaderMap, name: &str) -> Option<String> {
+    header(headers, name).filter(|value| !value.is_empty()).map(str::to_string)
+}
+
+/// `Last-Modified` and friends arrive as an HTTP date; the JSON dialects use
+/// RFC 3339 for the same field, so both are accepted.
+pub(super) fn parse_http_timestamp(value: &str) -> Option<SystemTime> {
+    OffsetDateTime::parse(value, &Rfc2822)
+        .or_else(|_| OffsetDateTime::parse(value, &Rfc3339))
+        .ok()
+        .map(SystemTime::from)
+}
+
+/// Provider-specific fields the shared header mapping cannot infer.
+pub(super) struct NativeHeadFields {
+    pub(super) etag: Option<String>,
+    /// The ETag is an opaque token rather than a digest of the bytes.
+    pub(super) etag_is_opaque: bool,
+    pub(super) version_id: Option<String>,
+    pub(super) storage_class: Option<String>,
+}
+
+/// Maps a HEAD or GET response onto [`SourceHead`]. `metadata_prefix` is the
+/// provider's user-metadata header prefix (`x-ms-meta-`, `x-goog-meta-`); the
+/// stored shape drops it, matching the `x-amz-meta-` handling of the S3 path.
+pub(super) fn native_source_head(
+    headers: &HeaderMap,
+    metadata_prefix: &str,
+    fields: NativeHeadFields,
+) -> Result<SourceHead, SourceError> {
+    let size = header(headers, "content-length")
+        .and_then(|value| value.parse::<u64>().ok())
+        .ok_or_else(|| SourceError::Other("source response has no valid content-length".to_string()))?;
+
+    let mut user_metadata = HashMap::new();
+    for (name, value) in headers {
+        let name = name.as_str();
+        if let Some(key) = name.strip_prefix(metadata_prefix)
+            && !key.is_empty()
+            && let Ok(value) = value.to_str()
+        {
+            user_metadata.insert(key.to_string(), value.to_string());
+        }
+    }
+
+    let etag = fields
+        .etag
+        .map(|etag| etag.trim().trim_matches('"').to_string())
+        .filter(|etag| !etag.is_empty());
+    // An opaque ETag never encodes a part count, so the multipart flag stays
+    // false for it however the provider happens to spell the token.
+    let is_multipart_etag = !fields.etag_is_opaque && etag.as_deref().is_some_and(is_multipart_etag);
+
+    Ok(SourceHead {
+        etag,
+        size,
+        last_modified: header(headers, "last-modified").and_then(parse_http_timestamp),
+        content_type: header_string(headers, "content-type"),
+        content_encoding: header_string(headers, "content-encoding"),
+        content_disposition: header_string(headers, "content-disposition"),
+        content_language: header_string(headers, "content-language"),
+        cache_control: header_string(headers, "cache-control"),
+        expires: header_string(headers, "expires"),
+        user_metadata,
+        version_id: fields.version_id,
+        storage_class: fields.storage_class,
+        // Neither native provider hands back ciphertext: a customer-key object
+        // is refused by the backend before it reaches this mapping, and the
+        // service-managed encryption is transparent to the reader.
+        sse: None,
+        is_multipart_etag,
+        etag_is_opaque: fields.etag_is_opaque,
+    })
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use http::HeaderValue;
+
+    fn headers(pairs: &[(&str, &str)]) -> HeaderMap {
+        let mut headers = HeaderMap::new();
+        for (name, value) in pairs {
+            headers.insert(
+                http::HeaderName::from_bytes(name.as_bytes()).expect("test header name"),
+                HeaderValue::from_str(value).expect("test header value"),
+            );
+        }
+        headers
+    }
+
+    fn fields() -> NativeHeadFields {
+        NativeHeadFields {
+            etag: None,
+            etag_is_opaque: false,
+            version_id: None,
+            storage_class: None,
+        }
+    }
+
+    #[test]
+    fn native_source_head_maps_content_headers_and_prefixed_metadata() {
+        let headers = headers(&[
+            ("content-length", "1234"),
+            ("content-type", "text/plain"),
+            ("content-encoding", "gzip"),
+            ("content-language", "en"),
+            ("content-disposition", "attachment"),
+            ("cache-control", "max-age=60"),
+            ("expires", "Thu, 01 Jan 2026 00:00:00 GMT"),
+            ("last-modified", "Wed, 21 Oct 2015 07:28:00 GMT"),
+            ("x-ms-meta-owner", "alice"),
+            ("x-goog-meta-owner", "not-mine"),
+        ]);
+        let head = native_source_head(
+            &headers,
+            "x-ms-meta-",
+            NativeHeadFields {
+                etag: Some("\"0x8DCE1D2\"".to_string()),
+                etag_is_opaque: true,
+                version_id: Some("2026-01-01T00:00:00.0000000Z".to_string()),
+                storage_class: Some("Hot".to_string()),
+            },
+        )
+        .expect("head should map");
+
+        assert_eq!(head.size, 1234);
+        assert_eq!(head.content_type.as_deref(), Some("text/plain"));
+        assert_eq!(head.content_encoding.as_deref(), Some("gzip"));
+        assert_eq!(head.content_language.as_deref(), Some("en"));
+        assert_eq!(head.content_disposition.as_deref(), Some("attachment"));
+        assert_eq!(head.cache_control.as_deref(), Some("max-age=60"));
+        assert_eq!(head.expires.as_deref(), Some("Thu, 01 Jan 2026 00:00:00 GMT"));
+        assert_eq!(
+            head.last_modified,
+            Some(SystemTime::UNIX_EPOCH + std::time::Duration::from_secs(1_445_412_480)),
+            "HTTP-date Last-Modified must parse"
+        );
+        assert_eq!(
+            head.user_metadata,
+            HashMap::from([("owner".to_string(), "alice".to_string())]),
+            "only the provider's own metadata prefix is read"
+        );
+        assert_eq!(head.etag.as_deref(), Some("0x8DCE1D2"), "quotes are stripped, the token is kept");
+        assert!(head.etag_is_opaque);
+        assert!(!head.is_multipart_etag);
+        assert_eq!(head.storage_class.as_deref(), Some("Hot"));
+        assert!(head.sse.is_none());
+    }
+
+    #[test]
+    fn native_source_head_requires_a_content_length() {
+        let err = native_source_head(&headers(&[("content-type", "text/plain")]), "x-ms-meta-", fields())
+            .expect_err("a response without content-length is unusable");
+        assert!(matches!(err, SourceError::Other(_)), "{err:?}");
+    }
+
+    #[test]
+    fn opaque_etag_never_reads_as_a_multipart_etag() {
+        // A digest-shaped ETag keeps the S3 reading; the same string marked
+        // opaque must not be split into "digest-partcount".
+        for (opaque, expected) in [(false, true), (true, false)] {
+            let head = native_source_head(
+                &headers(&[("content-length", "1")]),
+                "x-ms-meta-",
+                NativeHeadFields {
+                    etag: Some("d41d8cd98f00b204e9800998ecf8427e-3".to_string()),
+                    etag_is_opaque: opaque,
+                    ..fields()
+                },
+            )
+            .expect("head should map");
+            assert_eq!(head.is_multipart_etag, expected, "opaque = {opaque}");
+        }
+    }
+
+    #[test]
+    fn base64_md5_converts_only_sixteen_byte_digests() {
+        assert_eq!(
+            base64_md5_to_hex("1B2M2Y8AsgTpgAmY7PhCfg==").as_deref(),
+            Some("d41d8cd98f00b204e9800998ecf8427e")
+        );
+        assert_eq!(base64_md5_to_hex("not base64!").as_deref(), None);
+        // A CRC32C digest is four bytes: it must not pass as an MD5.
+        assert_eq!(base64_md5_to_hex("AAAAAA==").as_deref(), None);
+    }
+
+    #[test]
+    fn native_http_rejects_endpoints_that_are_not_bare_origins() {
+        for bad in [
+            "ftp://source.example.com",
+            "https://user:pw@source.example.com",
+            "https://source.example.com/container",
+            "https://source.example.com/?x=1",
+            "not a url",
+        ] {
+            assert!(
+                NativeHttp::new(bad, SourceTimeouts::default(), false, None).is_err(),
+                "{bad} must be rejected"
+            );
+        }
+    }
+
+    #[test]
+    fn native_http_percent_encodes_every_path_segment() {
+        let http = NativeHttp::for_test(Url::parse("https://acct.blob.core.windows.net").expect("origin"));
+        let url = http.url(["container", "dir", "a b?c#d.txt"]).expect("url should build");
+        assert_eq!(url.as_str(), "https://acct.blob.core.windows.net/container/dir/a%20b%3Fc%23d.txt");
+        assert_eq!(url.query(), None, "a key with '?' must not become a query");
+    }
+}

crates/ecstore/src/bucket/on_demand_migration/pull.rs

@@ -46,10 +46,10 @@ use super::stats::{PullFailureReason, PullPath};
 use super::sys::{BucketOdmState, OnDemandMigrationSys, PullError, PullOutcome, PullSlot};
 use async_trait::async_trait;
 use bytes::Bytes;
-use futures::{Stream, StreamExt};
+use futures::{FutureExt, Stream, StreamExt, future::Shared};
 use parking_lot::Mutex;
 use rand::RngExt;
-use std::collections::{HashMap, HashSet};
+use std::collections::HashMap;
 use std::fmt;
 use std::io;
 use std::pin::Pin;
@@ -133,6 +133,8 @@ pub enum QueuedPullOutcome {
     Failed(PullError),
 }
 
+pub type QueuedPullReport = Shared<oneshot::Receiver<QueuedPullOutcome>>;
+
 /// Result of [`PullQueue::enqueue`].
 #[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
 pub enum EnqueueOutcome {
@@ -251,6 +253,7 @@ pub struct WriteBackRequest {
     pub preserve_etag: bool,
     /// `policy.emit_events`.
     pub emit_events: bool,
+    pub respect_delete_marker: bool,
     /// Source tags to copy (`policy.copy_tags`), `None` to skip.
     pub tags: Option<HashMap<String, String>>,
 }
@@ -266,6 +269,7 @@ impl WriteBackRequest {
             pulled_at: OffsetDateTime::now_utc(),
             preserve_etag: config.policy.preserve_etag,
             emit_events: config.policy.emit_events,
+            respect_delete_marker: config.policy.respect_local_delete_marker,
             tags,
         }
     }
@@ -830,7 +834,7 @@ pub struct PullQueue {
     bucket: String,
     tx: mpsc::Sender<PullJob>,
     /// Keys queued or running; the job removes its key when it ends.
-    pending: Mutex<HashSet<String>>,
+    pending: Mutex<HashMap<String, QueuedPullReport>>,
     capacity: usize,
     cancel: CancellationToken,
     stats: Arc<super::stats::OdmStats>,
@@ -869,7 +873,7 @@ impl PullQueue {
         let queue = Arc::new(Self {
             bucket: state.bucket().to_string(),
             tx,
-            pending: Mutex::new(HashSet::new()),
+            pending: Mutex::new(HashMap::new()),
             capacity,
             cancel: state.cancel_token(),
             stats: Arc::clone(state.stats()),
@@ -903,29 +907,24 @@ impl PullQueue {
         self.enqueue_with_report(key, reason).0
     }
 
-    /// [`Self::enqueue`] that also hands back the job's report channel when
+    /// [`Self::enqueue`] with a shared report, including for coalesced pulls.
-    /// a new job was queued (`Coalesced` pulls report to their first
+    pub fn enqueue_with_report(&self, key: &str, reason: PullReason) -> (EnqueueOutcome, Option<QueuedPullReport>) {
-    /// requester only).
-    pub fn enqueue_with_report(
-        &self,
-        key: &str,
-        reason: PullReason,
-    ) -> (EnqueueOutcome, Option<oneshot::Receiver<QueuedPullOutcome>>) {
         if self.cancel.is_cancelled() {
             return (EnqueueOutcome::Unavailable, None);
         }
         let mut pending = self.pending.lock();
-        if pending.contains(key) {
+        if let Some(report) = pending.get(key) {
-            return (EnqueueOutcome::Coalesced, None);
+            return (EnqueueOutcome::Coalesced, Some(report.clone()));
         }
         let (report_tx, report_rx) = oneshot::channel();
+        let report_rx = report_rx.shared();
         match self.tx.try_send(PullJob {
             key: key.to_string(),
             reason,
             report: Some(report_tx),
         }) {
             Ok(()) => {
-                pending.insert(key.to_string());
+                pending.insert(key.to_string(), report_rx.clone());
                 (EnqueueOutcome::Enqueued, Some(report_rx))
             }
             Err(TrySendError::Full(_)) => {
@@ -1072,7 +1071,7 @@ impl BucketOdmState {
         self: &Arc<Self>,
         key: &str,
         reason: PullReason,
-    ) -> (EnqueueOutcome, Option<oneshot::Receiver<QueuedPullOutcome>>) {
+    ) -> (EnqueueOutcome, Option<QueuedPullReport>) {
         match self.pull_queue() {
             Some(queue) => queue.enqueue_with_report(key, reason),
             None => (EnqueueOutcome::Unavailable, None),
@@ -1119,6 +1118,8 @@ mod tests {
                     session_token: None,
                 }),
                 tls: TlsConfig::default(),
+                azure: None,
+                gcs: None,
             },
             filter: FilterConfig::default(),
             policy: PolicyConfig::default(),
@@ -1399,13 +1400,21 @@ mod tests {
         assert_eq!(queue.capacity(), 1024);
 
         let mut outcomes = HashMap::new();
+        let mut shared_report = None;
         for _ in 0..100 {
-            *outcomes.entry(queue.enqueue("a", PullReason::RangeGet)).or_insert(0) += 1;
+            let (outcome, report) = queue.enqueue_with_report("a", PullReason::RangeGet);
+            *outcomes.entry(outcome).or_insert(0) += 1;
+            shared_report = report;
         }
         assert_eq!(outcomes.get(&EnqueueOutcome::Enqueued), Some(&1));
         assert_eq!(outcomes.get(&EnqueueOutcome::Coalesced), Some(&99));
         assert_eq!(queue.pending_keys(), 1);
 
+        assert_eq!(
+            shared_report.expect("coalesced report").await,
+            Ok(QueuedPullOutcome::Stored { size: 1000 })
+        );
+
         wait_until("first pull to finish", || queue.pending_keys() == 0).await;
         assert_eq!(source.head_calls.load(Ordering::SeqCst), 1);
         assert_eq!(source.get_calls.load(Ordering::SeqCst), 1);
@@ -1438,6 +1447,23 @@ mod tests {
         assert_eq!(queue.enqueue("a", PullReason::RangeGet), EnqueueOutcome::Unavailable);
     }
 
+    #[tokio::test]
+    async fn coalesced_enqueues_share_failure_reports() {
+        let sys = OnDemandMigrationSys::new();
+        let state = enabled_state(&sys, &config()).await;
+        let source = MockSource::with_object("missing", 1000, BodyKind::Bytes(body_bytes(1000)));
+        let queue = PullQueue::start(Arc::clone(&state), source, Arc::new(MockWriteBack::default()));
+        let (first, first_report) = queue.enqueue_with_report("absent", PullReason::RangeGet);
+        let (second, second_report) = queue.enqueue_with_report("absent", PullReason::Backfill);
+        assert_eq!(first, EnqueueOutcome::Enqueued);
+        assert_eq!(second, EnqueueOutcome::Coalesced);
+        let (first, second) = tokio::join!(first_report.expect("leader report"), second_report.expect("coalesced report"));
+        assert_eq!(first, second);
+        assert!(matches!(first, Ok(QueuedPullOutcome::Failed(_))));
+        sys.remove(BUCKET);
+        queue.wait_until_stopped().await;
+    }
+
     #[tokio::test]
     async fn queue_full_is_reported_and_cancel_drains_without_leaking_tasks() {
         let sys = OnDemandMigrationSys::new();
@@ -1467,7 +1493,8 @@ mod tests {
         wait_until("dispatcher to wait for a slot", || state.stats().queue_depth() == 1).await;
         assert_eq!(queue.enqueue("c", PullReason::LargeObject), EnqueueOutcome::Enqueued);
         assert_eq!(queue.enqueue("d", PullReason::LargeObject), EnqueueOutcome::QueueFull);
-        assert_eq!(queue.enqueue("c", PullReason::LargeObject), EnqueueOutcome::Coalesced);
+        let (coalesced, canceled_report) = queue.enqueue_with_report("c", PullReason::LargeObject);
+        assert_eq!(coalesced, EnqueueOutcome::Coalesced);
         assert_eq!(queue.pending_keys(), 3);
         assert_eq!(failures(&state).get("queue_full"), Some(&1));
         assert!(!queue.is_stopped());
@@ -1477,6 +1504,12 @@ mod tests {
             .await
             .expect("dispatcher and in-flight job must exit after cancel");
         assert!(queue.is_stopped());
+        assert!(
+            tokio::time::timeout(Duration::from_secs(5), canceled_report.expect("coalesced cancellation report"))
+                .await
+                .expect("cancellation closes the report")
+                .is_err()
+        );
         assert_eq!(queue.pending_keys(), 0);
         assert_eq!(state.inflight_keys(), 0);
         assert_eq!(state.stats().inflight_pulls(), 0);

crates/ecstore/src/bucket/on_demand_migration/source_client.rs

@@ -25,6 +25,9 @@
 //! Client-supplied `If-*`, `Authorization`, `Host` and SSE-C headers are never
 //! forwarded: v1 rejects SSE-C source objects outright.
 
+use super::azure::AzureSourceBackend;
+use super::gcs::GcsNativeSourceBackend;
+use super::list_through::{ListPageError, validate_list_page};
 use crate::bucket::remote_s3_client::{
     PathStyle, RemoteCredentials, RemoteS3ClientError, RemoteS3EndpointSpec, RemoteS3RetryPolicy, build_remote_s3_config,
 };
@@ -64,6 +67,10 @@ pub enum SourceProvider {
     /// Generic S3-compatible service.
     #[default]
     S3,
+    /// Native Azure Blob service; not an S3 dialect.
+    Azure,
+    /// Native GCS JSON API with a service-account key; not an S3 dialect.
+    GcsNative,
 }
 
 impl SourceProvider {
@@ -75,6 +82,8 @@ impl SourceProvider {
             "minio" => Some(Self::Minio),
             "rustfs" => Some(Self::Rustfs),
             "s3" => Some(Self::S3),
+            "azure" => Some(Self::Azure),
+            "gcs_native" => Some(Self::GcsNative),
             _ => None,
         }
     }
@@ -87,6 +96,8 @@ impl SourceProvider {
             Self::Minio => "minio",
             Self::Rustfs => "rustfs",
             Self::S3 => "s3",
+            Self::Azure => "azure",
+            Self::GcsNative => "gcs_native",
         }
     }
 
@@ -152,12 +163,75 @@ pub struct SourceClientSpec {
     /// Wire requests one logical source call may cost. The pull pipeline and
     /// the backfill job own the retry budget (`pull.rs` `PULL_MAX_RETRIES`,
     /// `backfill.rs` `LIST_MAX_RETRIES`) and the breaker counts logical calls,
-    /// so ODM declares [`RemoteS3RetryPolicy::Disabled`] and keeps one counted
+    /// so ODM declares [`RemoteS3RetryPolicy::Disabled`]. An ambiguous HEAD
-    /// failure equal to one request against a struggling source.
+    /// 404 additionally probes the bucket before declaring a key absent.
     pub retry: RemoteS3RetryPolicy,
     /// Bytes per second the pull pipeline may consume from this source;
     /// `None` means unlimited. Enforced by the consumer, not by this client.
     pub bandwidth_limit: Option<NonZeroU64>,
+    /// Which [`SourceBackend`] to build. The S3 variant reads `region`,
+    /// `path_style` and `credentials`; the native variants ignore all three
+    /// and carry their own credentials.
+    pub backend: SourceBackendSpec,
+}
+
+/// Provider-specific half of [`SourceClientSpec`].
+#[derive(Clone, Debug, Default, PartialEq, Eq)]
+pub enum SourceBackendSpec {
+    #[default]
+    S3,
+    Azure(AzureSourceSpec),
+    Gcs(GcsSourceSpec),
+}
+
+/// Native Azure Blob parameters. The container is [`SourceClientSpec::bucket`].
+#[derive(Clone, PartialEq, Eq)]
+pub struct AzureSourceSpec {
+    pub account: String,
+    pub auth: AzureAuth,
+}
+
+impl fmt::Debug for AzureSourceSpec {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("AzureSourceSpec")
+            .field("account", &self.account)
+            .field("auth", &self.auth)
+            .finish()
+    }
+}
+
+/// How Azure requests are authorized.
+#[derive(Clone, PartialEq, Eq)]
+pub enum AzureAuth {
+    /// Base64 storage-account key, signed per request with Shared Key.
+    SharedKey(String),
+    /// SAS query string without the leading `?`, appended to every URL.
+    Sas(String),
+}
+
+impl fmt::Debug for AzureAuth {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        // Both variants are secrets; only the scheme may be rendered.
+        f.write_str(match self {
+            Self::SharedKey(_) => "SharedKey(REDACTED)",
+            Self::Sas(_) => "Sas(REDACTED)",
+        })
+    }
+}
+
+/// Native GCS parameters. The bucket is [`SourceClientSpec::bucket`].
+#[derive(Clone, PartialEq, Eq)]
+pub struct GcsSourceSpec {
+    /// Service-account key JSON.
+    pub service_account_json: String,
+}
+
+impl fmt::Debug for GcsSourceSpec {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_struct("GcsSourceSpec")
+            .field("service_account_json", &"REDACTED")
+            .finish()
+    }
 }
 
 impl SourceClientSpec {
@@ -223,6 +297,8 @@ pub enum SourceError {
     ServerError(u16),
     #[error("unsupported source object: {0}")]
     Unsupported(String),
+    #[error("invalid source listing: {0}")]
+    InvalidPagination(#[from] ListPageError),
     #[error("source request failed: {0}")]
     Other(String),
 }
@@ -245,6 +321,7 @@ impl SourceError {
             SourceError::Connect(_) => "connect",
             SourceError::ServerError(_) => "server_error",
             SourceError::Unsupported(_) => "unsupported",
+            SourceError::InvalidPagination(_) => "invalid_pagination",
             SourceError::Other(_) => "other",
         }
     }
@@ -258,7 +335,7 @@ const THROTTLE_CODES: &[&str] = &[
     "TooManyRequests",
     "RequestThrottled",
 ];
-const NOT_FOUND_CODES: &[&str] = &["NoSuchKey", "NotFound", "NoSuchBucket", "NoSuchVersion"];
+const NOT_FOUND_CODES: &[&str] = &["NoSuchKey"];
 const ACCESS_DENIED_CODES: &[&str] = &[
     "AccessDenied",
     "InvalidAccessKeyId",
@@ -268,7 +345,7 @@ const ACCESS_DENIED_CODES: &[&str] = &[
     "InvalidToken",
 ];
 
-fn classify_status(status: u16, code: Option<&str>, message: String) -> SourceError {
+pub(super) fn classify_status(status: u16, code: Option<&str>, message: String) -> SourceError {
     if let Some(code) = code {
         if THROTTLE_CODES.contains(&code) {
             return SourceError::Throttled;
@@ -281,7 +358,6 @@ fn classify_status(status: u16, code: Option<&str>, message: String) -> SourceEr
         }
     }
     match status {
-        404 => SourceError::NotFound,
         401 | 403 => SourceError::AccessDenied,
         429 | 503 => SourceError::Throttled,
         500..=599 => SourceError::ServerError(status),
@@ -341,6 +417,11 @@ pub struct SourceHead {
     pub storage_class: Option<String>,
     pub sse: Option<SourceSse>,
     pub is_multipart_etag: bool,
+    /// The provider's ETag is not derived from the object bytes (Azure
+    /// stamps an opaque concurrency token). Such an ETag is recorded for
+    /// provenance but must never be read as a content digest, so the
+    /// write-back path refuses to use it as the expected MD5.
+    pub etag_is_opaque: bool,
 }
 
 /// Per-operation fields shared by HEAD and GET outputs.
@@ -362,7 +443,7 @@ struct HeadParts {
     sse_customer_algorithm: Option<String>,
 }
 
-fn normalize_etag(etag: Option<String>) -> Option<String> {
+pub(super) fn normalize_etag(etag: Option<String>) -> Option<String> {
     etag.map(|etag| etag.trim().trim_matches('"').to_string())
         .filter(|etag| !etag.is_empty())
 }
@@ -411,6 +492,7 @@ fn source_head(parts: HeadParts) -> Result<SourceHead, SourceError> {
         storage_class: parts.storage_class,
         sse,
         is_multipart_etag,
+        etag_is_opaque: false,
     })
 }
 
@@ -621,14 +703,52 @@ impl fmt::Debug for SourceClient {
 
 impl SourceClient {
     pub async fn new(spec: &SourceClientSpec) -> Result<Self, RemoteS3ClientError> {
-        let endpoint = spec.endpoint_spec()?;
+        match &spec.backend {
-        let config = build_remote_s3_config(&endpoint).await?;
+            SourceBackendSpec::S3 => {
-        Ok(Self::from_config_builder(config, endpoint.endpoint_url(), spec))
+                let endpoint = spec.endpoint_spec()?;
+                let config = build_remote_s3_config(&endpoint).await?;
+                Ok(Self::from_config_builder(config, endpoint.endpoint_url(), spec))
+            }
+            SourceBackendSpec::Azure(azure) => {
+                let backend = AzureSourceBackend::new(
+                    &spec.endpoint,
+                    &spec.bucket,
+                    azure,
+                    spec.timeouts,
+                    spec.skip_tls_verify,
+                    spec.ca_cert_pem.as_deref(),
+                )?;
+                Ok(Self::from_backend(Box::new(backend), spec))
+            }
+            SourceBackendSpec::Gcs(gcs) => {
+                let backend = GcsNativeSourceBackend::new(
+                    &spec.endpoint,
+                    &spec.bucket,
+                    gcs,
+                    spec.timeouts,
+                    spec.skip_tls_verify,
+                    spec.ca_cert_pem.as_deref(),
+                )?;
+                Ok(Self::from_backend(Box::new(backend), spec))
+            }
+        }
+    }
+
+    /// Wraps a ready backend in the prefix-mapping client. The endpoint is
+    /// kept only for `Debug` and admin status.
+    fn from_backend(backend: Box<dyn SourceBackend>, spec: &SourceClientSpec) -> Self {
+        Self {
+            backend,
+            endpoint: spec.endpoint.clone(),
+            bucket: spec.bucket.clone(),
+            source_prefix: spec.source_prefix.clone().filter(|prefix| !prefix.is_empty()),
+            timeouts: spec.timeouts,
+            bandwidth_limit: spec.bandwidth_limit,
+        }
     }
 
     /// `config` must come from [`SourceClientSpec::endpoint_spec`], which is
-    /// where the retry policy that keeps one logical call equal to one wire
+    /// where the policy disabling SDK-level retries is declared.
-    /// request is declared.
     fn from_config_builder(config: aws_sdk_s3::config::Builder, endpoint: String, spec: &SourceClientSpec) -> Self {
         let client = S3Client::from_conf(config.interceptor(SourceProxyMarkerInterceptor::new()).build());
         Self {
@@ -714,6 +834,7 @@ impl SourceClient {
                 ..*request
             })
             .await?;
+        validate_list_page(page.is_truncated, request.continuation_token, page.next_continuation_token.as_deref())?;
         page.objects = page
             .objects
             .into_iter()
@@ -749,15 +870,16 @@ impl SourceClient {
 #[async_trait::async_trait]
 impl SourceBackend for S3SourceBackend {
     async fn head(&self, key: &str) -> Result<SourceHead, SourceError> {
-        let output = self
+        match self.client.head_object().bucket(&self.bucket).key(key).send().await {
-            .client
+            Ok(output) => source_head_from_head_output(output),
-            .head_object()
+            Err(err) if err.raw_response().is_some_and(|response| response.status().as_u16() == 404) => {
-            .bucket(&self.bucket)
+                // HEAD has no error body: a missing bucket must not poison
-            .key(key)
+                // the per-key negative cache as though only the key was absent.
-            .send()
+                self.probe().await?;
-            .await
+                Err(SourceError::NotFound)
-            .map_err(classify_sdk_error)?;
+            }
-        source_head_from_head_output(output)
+            Err(err) => Err(classify_sdk_error(err)),
+        }
     }
 
     /// Streams the object; `range` is passed through as an HTTP `Range`
@@ -800,17 +922,12 @@ impl SourceBackend for S3SourceBackend {
 
         let is_truncated = output.is_truncated.unwrap_or(false);
         let next_continuation_token = output.next_continuation_token;
-        if is_truncated && next_continuation_token.is_none() {
-            return Err(SourceError::Other(
-                "source reported a truncated listing without a continuation token".to_string(),
-            ));
-        }
         let objects = output
             .contents
             .unwrap_or_default()
             .into_iter()
-            .filter_map(s3_source_object)
+            .map(s3_source_object)
-            .collect();
+            .collect::<Result<Vec<_>, _>>()?;
         let common_prefixes = output
             .common_prefixes
             .unwrap_or_default()
@@ -849,14 +966,20 @@ impl SourceBackend for S3SourceBackend {
     }
 }
 
-fn s3_source_object(object: SdkObject) -> Option<SourceObject> {
+fn s3_source_object(object: SdkObject) -> Result<SourceObject, SourceError> {
-    let key = object.key?;
+    let key = object
+        .key
+        .ok_or_else(|| SourceError::Other("source listing object has no key".to_string()))?;
+    let size = object
+        .size
+        .and_then(|size| u64::try_from(size).ok())
+        .ok_or_else(|| SourceError::Other("source listing object has no valid size".to_string()))?;
     let etag = normalize_etag(object.e_tag);
     let is_multipart_etag = etag.as_deref().is_some_and(is_multipart_etag);
-    Some(SourceObject {
+    Ok(SourceObject {
         key,
         etag,
-        size: object.size.and_then(|size| u64::try_from(size).ok()).unwrap_or(0),
+        size,
         last_modified: system_time(object.last_modified),
         storage_class: object.storage_class.map(|class| class.as_str().to_string()),
         is_multipart_etag,
@@ -866,6 +989,7 @@ fn s3_source_object(object: SdkObject) -> Option<SourceObject> {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use crate::bucket::on_demand_migration::backend_contract::{BackendCapabilities, OBJECT_MD5, assert_backend_contract};
     use aws_smithy_runtime_api::client::http::{HttpConnector, HttpConnectorFuture, SharedHttpConnector, http_client_fn};
     use aws_smithy_runtime_api::client::orchestrator::HttpRequest;
     use aws_smithy_runtime_api::client::result::ConnectorError;
@@ -983,6 +1107,7 @@ mod tests {
             retry: RemoteS3RetryPolicy::Disabled,
             timeouts: SourceTimeouts::default(),
             bandwidth_limit: NonZeroU64::new(1_000_000),
+            backend: SourceBackendSpec::S3,
         }
     }
 
@@ -1274,7 +1399,9 @@ mod tests {
 <CommonPrefixes><Prefix>data/photos/</Prefix></CommonPrefixes>
 <CommonPrefixes><Prefix>outside/</Prefix></CommonPrefixes>
 </ListBucketResult>"#;
-        let (client, requests) = scripted_client(&spec(Some("data/")), vec![ok(Vec::new(), body), ok(Vec::new(), body)]).await;
+        let next_body = body.replace("data/opaque", "data/next");
+        let (client, requests) =
+            scripted_client(&spec(Some("data/")), vec![ok(Vec::new(), body), ok(Vec::new(), &next_body)]).await;
         let first = client
             .list_page(&SourceListRequest {
                 prefix: Some("photos/"),
@@ -1336,7 +1463,104 @@ mod tests {
             .list_objects_v2(None, None, 10)
             .await
             .expect_err("truncated page without token is corrupt");
-        assert!(matches!(err, SourceError::Other(_)), "{err:?}");
+        assert!(matches!(err, SourceError::InvalidPagination(ListPageError::Missing)), "{err:?}");
+    }
+
+    #[tokio::test]
+    async fn list_page_validates_s3_cursor_progress_before_mapping_entries() {
+        for contents in ["", "<Contents><Key>data/a</Key><Size>1</Size></Contents>"] {
+            for (truncated, next, expected) in [
+                (true, None, Some(ListPageError::Missing)),
+                (true, Some(""), Some(ListPageError::Empty)),
+                (true, Some("stuck"), Some(ListPageError::Repeated)),
+                (true, Some("opaque-next"), None),
+                (false, None, None),
+                (false, Some("stuck"), None),
+            ] {
+                let next_xml = next
+                    .map(|next| format!("<NextContinuationToken>{next}</NextContinuationToken>"))
+                    .unwrap_or_default();
+                let body = format!(
+                    "<ListBucketResult xmlns=\"http://s3.amazonaws.com/doc/2006-03-01/\"><IsTruncated>{truncated}</IsTruncated>{next_xml}{contents}</ListBucketResult>"
+                );
+                let (client, requests) = scripted_client(&spec(Some("data/")), vec![ok(Vec::new(), &body)]).await;
+                let result = client
+                    .list_page(&SourceListRequest {
+                        continuation_token: Some("stuck"),
+                        max_keys: 2,
+                        ..Default::default()
+                    })
+                    .await;
+                match expected {
+                    Some(expected) => {
+                        let error = result.expect_err("malformed pagination must fail at the provider boundary");
+                        assert!(
+                            matches!(&error, SourceError::InvalidPagination(actual) if *actual == expected),
+                            "{error:?}"
+                        );
+                        assert_eq!(error.class_label(), "invalid_pagination");
+                        assert!(!error.is_retryable());
+                        assert!(!error.to_string().contains("stuck"), "errors must not echo opaque tokens");
+                    }
+                    None => {
+                        let page = result.expect("progressing empty/nonempty pages and EOF are valid");
+                        assert_eq!(page.is_truncated, truncated);
+                        assert_eq!(page.next_continuation_token.as_deref(), next);
+                        assert_eq!(page.objects.len(), usize::from(!contents.is_empty()));
+                        if let Some(object) = page.objects.first() {
+                            assert_eq!(object.key, "a");
+                        }
+                    }
+                }
+                let requests = recorded(&requests);
+                assert_eq!(requests.len(), 1, "invalid pagination must not be retried");
+                assert!(requests[0].uri.contains("continuation-token=stuck"));
+            }
+        }
+    }
+
+    struct ListOnlyBackend(SourcePage);
+
+    #[async_trait::async_trait]
+    impl SourceBackend for ListOnlyBackend {
+        async fn list(&self, request: &SourceListRequest<'_>) -> Result<SourcePage, SourceError> {
+            assert_eq!(request.continuation_token, Some("stuck"), "opaque cursors reach every provider unchanged");
+            Ok(self.0.clone())
+        }
+
+        async fn head(&self, _key: &str) -> Result<SourceHead, SourceError> {
+            panic!("unexpected HEAD in list test")
+        }
+        async fn get(&self, _key: &str, _range: Option<&HTTPRangeSpec>) -> Result<SourceGet, SourceError> {
+            panic!("unexpected GET in list test")
+        }
+        async fn tagging(&self, _key: &str) -> Result<HashMap<String, String>, SourceError> {
+            panic!("unexpected tagging in list test")
+        }
+        async fn probe(&self) -> Result<(), SourceError> {
+            panic!("unexpected probe in list test")
+        }
+    }
+
+    #[tokio::test]
+    async fn list_page_validates_non_s3_provider_cursors_at_the_common_boundary() {
+        for (next, expected) in [
+            (None, ListPageError::Missing),
+            (Some(""), ListPageError::Empty),
+            (Some("stuck"), ListPageError::Repeated),
+        ] {
+            let mut client = prefix_client(Some("data/".into()));
+            client.backend = Box::new(ListOnlyBackend(SourcePage {
+                is_truncated: true,
+                next_continuation_token: next.map(str::to_string),
+                ..Default::default()
+            }));
+            let error = client
+                .list_objects_v2(None, Some("stuck"), 2)
+                .await
+                .expect_err("all providers must advance pagination");
+            assert!(matches!(error, SourceError::InvalidPagination(actual) if actual == expected));
+        }
     }
 
     const TAGGING_BODY: &str = r#"<?xml version="1.0" encoding="UTF-8"?>
@@ -1390,7 +1614,10 @@ mod tests {
     #[tokio::test]
     async fn source_error_classification_covers_every_class() {
         let cases: Vec<(Scripted, &str, bool)> = vec![
-            (status(404, ""), "not_found", false),
+            (status(404, ""), "other", false),
+            (status(404, "<Error><Code>NoSuchKey</Code></Error>"), "not_found", false),
+            (status(404, "<Error><Code>NoSuchBucket</Code></Error>"), "other", false),
+            (status(404, "<Error><Code>NoSuchVersion</Code></Error>"), "other", false),
             (status(403, ACCESS_DENIED_BODY), "access_denied", false),
             (status(401, ""), "access_denied", false),
             (status(429, ""), "throttled", true),
@@ -1413,14 +1640,35 @@ mod tests {
             }
         }
 
-        // HEAD carries no error body, so the classification must work from the
+        let (client, requests) = scripted_client(&spec(None), vec![status(404, ""), status(200, "")]).await;
-        // status alone as well.
-        let (client, _) = scripted_client(&spec(None), vec![status(404, "")]).await;
         assert!(matches!(client.head_object("missing").await, Err(SourceError::NotFound)));
+        assert_eq!(recorded(&requests).len(), 2, "ambiguous HEAD 404 must check the bucket");
+        let (client, _) = scripted_client(&spec(None), vec![status(404, ""), status(404, "")]).await;
+        assert!(matches!(client.head_object("missing").await, Err(SourceError::Other(_))));
+        let (client, _) = scripted_client(&spec(None), vec![status(404, ""), status(403, "")]).await;
+        assert!(matches!(client.head_object("missing").await, Err(SourceError::AccessDenied)));
         let (client, _) = scripted_client(&spec(None), vec![status(403, "")]).await;
         assert!(matches!(client.head_object("secret").await, Err(SourceError::AccessDenied)));
     }
 
+    #[test]
+    fn source_listing_rejects_missing_and_negative_sizes() {
+        for size in [None, Some(-1)] {
+            let object = SdkObject::builder().key("key").set_size(size).build();
+            assert!(matches!(s3_source_object(object), Err(SourceError::Other(_))));
+        }
+        assert!(matches!(
+            s3_source_object(SdkObject::builder().size(0).build()),
+            Err(SourceError::Other(_))
+        ));
+        assert_eq!(
+            s3_source_object(SdkObject::builder().key("empty").size(0).build())
+                .expect("empty object")
+                .size,
+            0
+        );
+    }
+
     #[tokio::test]
     async fn source_client_debug_redacts_credentials() {
         let (client, _) = scripted_client(&spec(Some("data/")), Vec::new()).await;
@@ -1487,7 +1735,101 @@ mod tests {
         assert_eq!(resolve_path_style(PathStyle::VirtualHost, Minio, "10.0.0.1"), PathStyle::VirtualHost);
         assert_eq!(resolve_path_style(PathStyle::Path, Aws, "s3.amazonaws.com"), PathStyle::Path);
         assert_eq!(SourceProvider::from_label(" AWS "), Some(Aws));
-        assert_eq!(SourceProvider::from_label("azure"), None);
+        assert_eq!(SourceProvider::from_label(" Azure "), Some(Azure));
+        assert_eq!(SourceProvider::from_label("gcs_native"), Some(GcsNative));
+        assert_eq!(SourceProvider::from_label("swift"), None);
+    }
+
+    const CONTRACT_LIST_PAGE_ONE: &str = r#"<?xml version="1.0" encoding="UTF-8"?>
+<ListBucketResult xmlns="http://s3.amazonaws.com/doc/2006-03-01/">
+  <Name>source-bucket</Name>
+  <IsTruncated>true</IsTruncated>
+  <NextContinuationToken>cursor-1</NextContinuationToken>
+  <Contents>
+    <Key>dir/a.txt</Key>
+    <LastModified>2015-10-21T07:28:00.000Z</LastModified>
+    <ETag>&quot;5d41402abc4b2a76b9719d911017c592&quot;</ETag>
+    <Size>5</Size>
+    <StorageClass>STANDARD</StorageClass>
+  </Contents>
+  <CommonPrefixes><Prefix>dir/sub/</Prefix></CommonPrefixes>
+</ListBucketResult>"#;
+
+    const CONTRACT_LIST_PAGE_TWO: &str = r#"<?xml version="1.0" encoding="UTF-8"?>
+<ListBucketResult xmlns="http://s3.amazonaws.com/doc/2006-03-01/">
+  <Name>source-bucket</Name>
+  <IsTruncated>false</IsTruncated>
+  <Contents>
+    <Key>dir/b.txt</Key>
+    <LastModified>2015-10-21T07:28:00.000Z</LastModified>
+    <ETag>&quot;7d41402abc4b2a76b9719d911017c592&quot;</ETag>
+    <Size>7</Size>
+  </Contents>
+</ListBucketResult>"#;
+
+    const CONTRACT_TAGGING: &str = r#"<?xml version="1.0" encoding="UTF-8"?>
+<Tagging xmlns="http://s3.amazonaws.com/doc/2006-03-01/"><TagSet>
+  <Tag><Key>env</Key><Value>prod</Value></Tag>
+</TagSet></Tagging>"#;
+
+    fn contract_object_headers(content_length: u64) -> Vec<(&'static str, String)> {
+        vec![
+            ("etag", format!("\"{OBJECT_MD5}\"")),
+            ("content-length", content_length.to_string()),
+            ("content-type", "text/plain".to_string()),
+            ("last-modified", "Wed, 21 Oct 2015 07:28:00 GMT".to_string()),
+            ("x-amz-meta-owner", "alice".to_string()),
+            ("x-amz-storage-class", "STANDARD".to_string()),
+        ]
+    }
+
+    /// The S3 backend behind the scripted connector, without the prefix-mapping
+    /// client on top: the contract is a property of the backend itself.
+    async fn scripted_s3_backend(responses: Vec<Scripted>) -> S3SourceBackend {
+        let spec = spec(None);
+        let connector = SharedHttpConnector::new(ScriptedConnector {
+            requests: Arc::new(Mutex::new(Vec::new())),
+            responses: Arc::new(Mutex::new(responses.into_iter().collect())),
+        });
+        let http_client = http_client_fn(move |_settings, _components| connector.clone());
+        let endpoint = spec.endpoint_spec().expect("test spec endpoint should parse");
+        let config = build_remote_s3_config(&endpoint)
+            .await
+            .expect("test spec should build")
+            .http_client(http_client)
+            .interceptor(SourceProxyMarkerInterceptor::new());
+        S3SourceBackend {
+            client: S3Client::from_conf(config.build()),
+            bucket: spec.bucket.clone(),
+        }
+    }
+
+    #[tokio::test]
+    async fn s3_backend_satisfies_the_shared_backend_contract() {
+        let mut ranged = contract_object_headers(3);
+        ranged.push(("content-range", "bytes 1-3/5".to_string()));
+        let backend = scripted_s3_backend(vec![
+            ok(contract_object_headers(5), ""),
+            ok(contract_object_headers(5), "hello"),
+            ok(ranged, "ell"),
+            ok(Vec::new(), CONTRACT_LIST_PAGE_ONE),
+            ok(Vec::new(), CONTRACT_LIST_PAGE_TWO),
+            ok(Vec::new(), CONTRACT_TAGGING),
+            ok(Vec::new(), ""),
+            status(404, ""),
+            status(403, ACCESS_DENIED_BODY),
+        ])
+        .await;
+
+        assert_backend_contract(
+            &backend,
+            BackendCapabilities {
+                etag_is_opaque: false,
+                supports_start_after: true,
+                supports_tagging: true,
+            },
+        )
+        .await;
     }
 
     fn prefix_client(prefix: Option<String>) -> SourceClient {

crates/ecstore/src/bucket/on_demand_migration/stats.rs

@@ -177,7 +177,7 @@ impl From<&SourceError> for PullFailureReason {
             SourceError::Connect(_) => PullFailureReason::SourceConnect,
             SourceError::ServerError(_) => PullFailureReason::SourceServerError,
             SourceError::Unsupported(_) => PullFailureReason::SourceUnsupported,
-            SourceError::Other(_) => PullFailureReason::SourceOther,
+            SourceError::InvalidPagination(_) | SourceError::Other(_) => PullFailureReason::SourceOther,
         }
     }
 }

crates/ecstore/src/bucket/on_demand_migration/sys.rs

@@ -47,7 +47,10 @@ use super::config::{
 use super::list_through::{SOURCE_LIST_RATE_PER_SEC, SourceListRateLimiter};
 use super::negative_cache::NegativeCache;
 use super::pull::{OdmWriteBack, PullQueue};
-use super::source_client::{SourceClient, SourceClientSpec, SourceError, SourceProvider, SourceTimeouts};
+use super::source_client::{
+    AzureAuth, AzureSourceSpec, GcsSourceSpec, SourceBackendSpec, SourceClient, SourceClientSpec, SourceError, SourceProvider,
+    SourceTimeouts,
+};
 use super::stats::{GaugeGuard, OdmStats, OdmStatsSnapshot, PullFailureReason};
 use crate::bucket::remote_s3_client::{
     PathStyle as ClientPathStyle, RemoteCredentials, RemoteS3ClientError, RemoteS3RetryPolicy,
@@ -619,6 +622,7 @@ pub fn source_client_spec(config: &OnDemandMigrationConfig) -> SourceClientSpec
         // load on a source that is already failing.
         retry: RemoteS3RetryPolicy::Disabled,
         bandwidth_limit: policy.bandwidth_limit_bytes_per_sec.and_then(NonZeroU64::new),
+        backend: source_backend_spec(source),
     }
 }
 
@@ -630,6 +634,31 @@ fn source_provider(provider: Provider) -> SourceProvider {
         Provider::Rustfs => SourceProvider::Rustfs,
         Provider::R2 => SourceProvider::R2,
         Provider::Gcs => SourceProvider::Gcs,
+        Provider::Azure => SourceProvider::Azure,
+        Provider::GcsNative => SourceProvider::GcsNative,
+    }
+}
+
+/// Which backend the client builds. A native provider whose block is missing
+/// falls back to the S3 spec, where the builder reports the missing
+/// credentials: the config layer already refuses to store that shape, so this
+/// only covers a config written by an older or hand-edited build.
+pub fn source_backend_spec(source: &SourceConfig) -> SourceBackendSpec {
+    match (source.provider, source.azure.as_ref(), source.gcs.as_ref()) {
+        (Provider::Azure, Some(azure), _) => SourceBackendSpec::Azure(AzureSourceSpec {
+            account: azure.account.clone(),
+            auth: match (&azure.account_key, &azure.sas_token) {
+                (Some(key), _) => AzureAuth::SharedKey(key.clone()),
+                (None, Some(sas)) => AzureAuth::Sas(sas.clone()),
+                // Refused by `SourceConfig::validate`; an empty shared key
+                // fails closed at the builder rather than signing with none.
+                (None, None) => AzureAuth::SharedKey(String::new()),
+            },
+        }),
+        (Provider::GcsNative, _, Some(gcs)) => SourceBackendSpec::Gcs(GcsSourceSpec {
+            service_account_json: gcs.service_account_json.clone(),
+        }),
+        _ => SourceBackendSpec::S3,
     }
 }
 
@@ -929,6 +958,8 @@ mod tests {
                     session_token: None,
                 }),
                 tls: TlsConfig::default(),
+                azure: None,
+                gcs: None,
             },
             filter: FilterConfig {
                 prefix: prefix.map(str::to_string),

crates/ecstore/src/bucket/on_demand_migration/test_http_fixture.rs

@@ -0,0 +1,120 @@
+// 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.
+
+//! Scripted HTTP server for the native source backends' tests.
+//!
+//! The S3 backend can be driven through the SDK's own connector; the native
+//! backends talk to a real socket, so their tests need a server that answers a
+//! fixed script and records what it was asked. Every response closes its
+//! connection, which keeps one request on one socket and makes the script order
+//! exactly the request order.
+
+use std::sync::{Arc, Mutex};
+use tokio::io::{AsyncReadExt, AsyncWriteExt};
+use tokio::net::TcpListener;
+use url::Url;
+
+pub(super) struct ScriptedResponse {
+    status: u16,
+    headers: Vec<(&'static str, String)>,
+    body: String,
+}
+
+impl ScriptedResponse {
+    pub(super) fn new(status: u16, headers: Vec<(&'static str, String)>, body: String) -> Self {
+        Self { status, headers, body }
+    }
+}
+
+#[derive(Clone, Debug)]
+pub(super) struct RecordedRequest {
+    pub(super) method: String,
+    /// Request target as it appeared on the wire: path plus query.
+    pub(super) target: String,
+    pub(super) headers: Vec<(String, String)>,
+}
+
+impl RecordedRequest {
+    pub(super) fn header(&self, name: &str) -> Option<&str> {
+        self.headers
+            .iter()
+            .find(|(key, _)| key.eq_ignore_ascii_case(name))
+            .map(|(_, value)| value.as_str())
+    }
+}
+
+pub(super) type Recorder = Arc<Mutex<Vec<RecordedRequest>>>;
+
+/// Binds a loopback listener that answers `responses` in order and returns its
+/// origin plus the recorder. The task ends once the script is exhausted.
+pub(super) async fn scripted_server(responses: Vec<ScriptedResponse>) -> (Url, Recorder) {
+    let listener = TcpListener::bind("127.0.0.1:0").await.expect("fixture listener should bind");
+    let port = listener.local_addr().expect("fixture address").port();
+    let recorder: Recorder = Arc::new(Mutex::new(Vec::new()));
+    let sink = Arc::clone(&recorder);
+
+    tokio::spawn(async move {
+        for response in responses {
+            let Ok((mut stream, _)) = listener.accept().await else {
+                return;
+            };
+            let mut request = Vec::new();
+            let mut buffer = [0_u8; 2048];
+            while !request.windows(4).any(|window| window == b"\r\n\r\n") {
+                match stream.read(&mut buffer).await {
+                    Ok(0) | Err(_) => break,
+                    Ok(read) => request.extend_from_slice(&buffer[..read]),
+                }
+            }
+            let text = String::from_utf8_lossy(&request).into_owned();
+            let mut lines = text.lines();
+            let start = lines.next().unwrap_or_default().to_string();
+            let mut parts = start.split_whitespace();
+            sink.lock().expect("recorder lock").push(RecordedRequest {
+                method: parts.next().unwrap_or_default().to_string(),
+                target: parts.next().unwrap_or_default().to_string(),
+                headers: lines
+                    .take_while(|line| !line.is_empty())
+                    .filter_map(|line| line.split_once(':'))
+                    .map(|(name, value)| (name.trim().to_string(), value.trim().to_string()))
+                    .collect(),
+            });
+
+            // A scripted HEAD declares the object size in its own headers while
+            // carrying no body, so an explicit `Content-Length` wins over the
+            // body length.
+            let declares_length = response
+                .headers
+                .iter()
+                .any(|(name, _)| name.eq_ignore_ascii_case("content-length"));
+            let mut rendered = match declares_length {
+                true => format!("HTTP/1.1 {} Scripted\r\nConnection: close\r\n", response.status),
+                false => format!(
+                    "HTTP/1.1 {} Scripted\r\nContent-Length: {}\r\nConnection: close\r\n",
+                    response.status,
+                    response.body.len()
+                ),
+            };
+            for (name, value) in response.headers {
+                rendered.push_str(&format!("{name}: {value}\r\n"));
+            }
+            rendered.push_str("\r\n");
+            rendered.push_str(&response.body);
+            let _ = stream.write_all(rendered.as_bytes()).await;
+            let _ = stream.flush().await;
+        }
+    });
+
+    (Url::parse(&format!("http://127.0.0.1:{port}")).expect("fixture endpoint"), recorder)
+}

crates/ecstore/src/bucket/remote_s3_client.rs

@@ -652,9 +652,10 @@ async fn build_aws_s3_http_client_from_tls_path() -> Option<SharedHttpClient> {
 #[cfg(test)]
 mod tests {
     use super::*;
+    use aws_smithy_async::time::TimeSource;
     use aws_smithy_runtime_api::http::StatusCode as SmithyStatusCode;
     use std::sync::Mutex;
-    use std::sync::atomic::{AtomicUsize, Ordering};
+    use std::sync::atomic::{AtomicU64, AtomicUsize, Ordering};
 
     fn spec(endpoint: &str, secure: bool) -> RemoteS3EndpointSpec {
         RemoteS3EndpointSpec {
@@ -824,6 +825,174 @@ mod tests {
         );
     }
 
+    #[derive(Clone, Debug)]
+    struct ClockSkewTimeSource(Arc<AtomicU64>);
+
+    impl TimeSource for ClockSkewTimeSource {
+        fn now(&self) -> SystemTime {
+            SystemTime::UNIX_EPOCH + Duration::from_secs(self.0.load(Ordering::SeqCst))
+        }
+    }
+
+    #[derive(Clone, Debug)]
+    struct ClockSkewConnector {
+        request_headers: RecordedHeaders,
+        error_code: &'static str,
+        skew_seconds: i64,
+        clock: ClockSkewTimeSource,
+    }
+
+    fn recorded_header<'a>(headers: &'a [(String, String)], name: &str) -> &'a str {
+        headers
+            .iter()
+            .find(|(key, _)| key.eq_ignore_ascii_case(name))
+            .map(|(_, value)| value.as_str())
+            .unwrap_or_else(|| panic!("signed request must contain {name}"))
+    }
+
+    fn signing_time(headers: &[(String, String)]) -> chrono::NaiveDateTime {
+        chrono::NaiveDateTime::parse_from_str(recorded_header(headers, "x-amz-date"), "%Y%m%dT%H%M%SZ")
+            .expect("SDK signing timestamp must use the SigV4 format")
+    }
+
+    impl SmithyHttpConnector for ClockSkewConnector {
+        fn call(&self, request: HttpRequest) -> HttpConnectorFuture {
+            let mut headers = self.request_headers.lock().expect("clock skew request capture lock");
+            assert!(headers.len() < 3, "clock skew fixture must not exceed two GET attempts and one HEAD");
+            headers.push(
+                request
+                    .headers()
+                    .iter()
+                    .map(|(key, value)| (key.to_string(), value.to_string()))
+                    .collect(),
+            );
+            let server_time = chrono::DateTime::<chrono::Utc>::from(self.clock.now()).naive_utc()
+                + chrono::Duration::seconds(self.skew_seconds);
+            let (status, body) = if headers.len() == 1 {
+                (
+                    403,
+                    format!("<Error><Code>{}</Code><Message>Clock skew fixture</Message></Error>", self.error_code),
+                )
+            } else {
+                (200, String::new())
+            };
+            let response = http::Response::builder()
+                .status(status)
+                .header("date", server_time.format("%a, %d %b %Y %H:%M:%S GMT").to_string())
+                .header("content-type", "application/xml")
+                .header("content-length", body.len())
+                .body(SdkBody::from(body))
+                .expect("clock skew fixture response");
+            HttpConnectorFuture::ready(Ok(HttpResponse::try_from(response).expect("Smithy fixture response")))
+        }
+    }
+
+    async fn clock_skew_client(
+        error_code: &'static str,
+        skew_seconds: i64,
+        retry: RemoteS3RetryPolicy,
+    ) -> (S3Client, RecordedHeaders, ClockSkewTimeSource) {
+        let headers: RecordedHeaders = Arc::new(Mutex::new(Vec::new()));
+        let clock = ClockSkewTimeSource(Arc::new(AtomicU64::new(1_700_000_000)));
+        let connector = SharedHttpConnector::new(ClockSkewConnector {
+            request_headers: Arc::clone(&headers),
+            error_code,
+            skew_seconds,
+            clock: clock.clone(),
+        });
+        let mut spec = spec("s3.example.com", true);
+        spec.retry = retry;
+        let config = build_remote_s3_config(&spec)
+            .await
+            .expect("clock skew fixture uses the production outbound configuration")
+            .http_client(http_client_fn(move |_settings, _components| connector.clone()))
+            .time_source(clock.clone())
+            .build();
+        (S3Client::from_conf(config), headers, clock)
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn remote_s3_clock_skew_retries_resign_and_seed_next_operation() {
+        for error_code in ["RequestTimeTooSkewed", "SignatureDoesNotMatch"] {
+            for skew_seconds in [-600, 600] {
+                let (client, headers, clock) = clock_skew_client(error_code, skew_seconds, REPLICATION_TARGET_RETRY_POLICY).await;
+                let initial = chrono::DateTime::<chrono::Utc>::from(clock.now()).naive_utc();
+                client
+                    .get_object()
+                    .bucket("bucket")
+                    .key("object")
+                    .send()
+                    .await
+                    .expect("clock skew GET must retry successfully");
+                assert_eq!(
+                    headers.lock().expect("captured requests").len(),
+                    2,
+                    "{error_code}: GET needs exactly one retry"
+                );
+                clock.0.fetch_add(17, Ordering::SeqCst);
+                // SDK signing time is independent of Tokio's retry/scheduler clock.
+                tokio::time::advance(Duration::from_secs(61)).await;
+                client
+                    .head_bucket()
+                    .bucket("bucket")
+                    .send()
+                    .await
+                    .expect("subsequent HEAD must use the client's cached skew");
+                let headers = headers.lock().expect("captured signed requests");
+                assert_eq!(headers.len(), 3, "subsequent operation must succeed on its first attempt");
+                assert_eq!(signing_time(&headers[0]), initial, "the first attempt must use the injected clock");
+                assert_eq!(
+                    signing_time(&headers[1]),
+                    initial + chrono::Duration::seconds(skew_seconds),
+                    "{error_code}: retry must apply the measured offset exactly"
+                );
+                assert_eq!(
+                    signing_time(&headers[2]),
+                    initial + chrono::Duration::seconds(skew_seconds + 17),
+                    "{error_code}: the next operation must apply cached skew to the advanced signing clock"
+                );
+                let signature = |index: usize| {
+                    recorded_header(&headers[index], "authorization")
+                        .rsplit_once("Signature=")
+                        .expect("SigV4 authorization contains a signature")
+                        .1
+                };
+                assert_ne!(
+                    signature(0),
+                    signature(1),
+                    "{error_code}: retry must be signed again after adjusting its date"
+                );
+            }
+        }
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn remote_s3_clock_skew_respects_one_attempt_policy() {
+        use aws_smithy_types::error::metadata::ProvideErrorMetadata;
+
+        for error_code in ["RequestTimeTooSkewed", "SignatureDoesNotMatch"] {
+            for retry in [
+                RemoteS3RetryPolicy::Disabled,
+                RemoteS3RetryPolicy::Standard { max_attempts: 1 },
+            ] {
+                let (client, headers, _clock) = clock_skew_client(error_code, 600, retry).await;
+                let error = client
+                    .get_object()
+                    .bucket("bucket")
+                    .key("object")
+                    .send()
+                    .await
+                    .expect_err("clock skew must not override the caller's one-attempt budget");
+                assert_eq!(error.as_service_error().and_then(ProvideErrorMetadata::code), Some(error_code));
+                assert_eq!(
+                    headers.lock().expect("captured requests").len(),
+                    1,
+                    "{error_code}: {retry:?} must send exactly one request"
+                );
+            }
+        }
+    }
+
     #[test]
     fn path_style_auto_and_path_force_path_style() {
         assert!(PathStyle::Auto.force_path_style());

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

@@ -46,7 +46,7 @@ use super::replication_storage_boundary::{
     HTTPPreconditions, ObjectInfo, ObjectOptions, ObjectToDelete, ReplicationDeletedObject, ReplicationObjectIO,
     ReplicationStorage,
 };
-use super::replication_target_boundary::{ReplicationTargetStore, replication_object_is_ssec_encrypted};
+use super::replication_target_boundary::{BucketTargetError, ReplicationTargetStore, replication_object_is_ssec_encrypted};
 use super::replication_versioning_boundary::ReplicationVersioningStore;
 use super::runtime_boundary as runtime_sources;
 use futures_util::stream::{self, StreamExt};
@@ -3084,6 +3084,23 @@ pub async fn queue_replication_heal(bucket: &str, oi: ObjectInfo, retry_count: u
 
     let tgts = match ReplicationTargetStore::list_bucket_targets(bucket).await {
         Ok(targets) => Some(targets),
+        // A bucket whose persisted target configuration cannot be decoded has
+        // an unknown target set, not an empty one: scheduling against `None`
+        // here would drop every heal for it without a trace
+        // (rustfs/backlog#2282). Report it missed so the object is retried
+        // once the configuration is readable again.
+        Err(BucketTargetError::BucketRemoteTargetsUnreadable { .. }) => {
+            warn!(
+                event = EVENT_REPLICATION_CONFIG_LOOKUP_SKIPPED,
+                component = LOG_COMPONENT_ECSTORE,
+                subsystem = LOG_SUBSYSTEM_REPLICATION,
+                bucket,
+                reason = "target_config_unreadable",
+                "Bucket replication targets are unreadable; replication heal queue fails closed"
+            );
+
+            return ReplicationQueueAdmission::Missed;
+        }
         Err(err) => {
             debug!(
                 event = EVENT_REPLICATION_CONFIG_LOOKUP_SKIPPED,

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

@@ -15,7 +15,8 @@
 use std::collections::HashMap;
 use std::sync::Arc;
 
-use crate::bucket::bucket_target_sys::{BucketTargetError, BucketTargetSys};
+pub(crate) use crate::bucket::bucket_target_sys::BucketTargetError;
+use crate::bucket::bucket_target_sys::BucketTargetSys;
 use aws_sdk_s3::operation::head_object::HeadObjectOutput;
 use aws_sdk_s3::types::{ObjectLockLegalHoldStatus, ObjectLockRetentionMode};
 use http::HeaderMap;

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

@@ -30,6 +30,7 @@ use rustfs_protos::{
     ChannelClass, create_new_channel, get_channel_for_class,
     proto_gen::node_service::{
         heal_control_service_client::HealControlServiceClient, node_service_client::NodeServiceClient,
+        scanner_control_service_client::ScannerControlServiceClient,
         tier_mutation_control_service_client::TierMutationControlServiceClient,
     },
 };
@@ -60,6 +61,24 @@ pub async fn node_service_time_out_client(
     node_service_time_out_client_for_class(addr, interceptor, ChannelClass::Control).await
 }
 
+pub(crate) async fn scanner_control_time_out_client(
+    addr: &str,
+    interceptor: TonicInterceptor,
+) -> crate::error::Result<ScannerControlServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
+    let interceptor = interceptor.with_rpc_audience(addr)?;
+    let channel = match runtime_sources::cached_node_channel(addr).await {
+        Some(channel) => channel,
+        None => create_new_channel(addr)
+            .await
+            .map_err(|err| crate::error::Error::other(err.to_string()))?,
+    };
+    let channel = ReplayScopeChannel::new(channel, interceptor.replay_scope_audience());
+    let limit = rustfs_protos::scoped_dirty_usage::SCOPED_DIRTY_USAGE_MAX_REQUEST_BYTES as usize;
+    Ok(ScannerControlServiceClient::with_interceptor(channel, interceptor)
+        .max_decoding_message_size(limit)
+        .max_encoding_message_size(limit))
+}
+
 pub async fn heal_control_time_out_client(
     addr: &str,
     interceptor: TonicInterceptor,

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

@@ -2050,6 +2050,53 @@ impl PeerRestClient {
         .await
     }
 
+    /// Probe only: scoped ACK production requires a durable per-bucket proof.
+    pub async fn scanner_scoped_dirty_usage_capability(
+        &self,
+        owner_id: String,
+        instance_id: String,
+        entries: Vec<rustfs_protos::proto_gen::node_service::ScannerScopedDirtyUsageEntry>,
+    ) -> Result<bool> {
+        use rustfs_protos::scoped_dirty_usage::*;
+        let payload = rustfs_protos::proto_gen::node_service::ScannerScopedDirtyUsageAckRequest {
+            challenge: Uuid::new_v4().as_bytes().to_vec().into(),
+            protocol_version: SCOPED_DIRTY_USAGE_PROTOCOL_VERSION,
+            owner_id,
+            instance_id,
+            scope: SCOPED_DIRTY_USAGE_BUCKET_SCOPE,
+            probe_only: true,
+            entries,
+        };
+        let canonical = canonical_scoped_dirty_usage_request(&payload).map_err(|err| Error::other(err.to_string()))?;
+        self.finalize_result(
+            async {
+                let mut client = super::client::scanner_control_time_out_client(
+                    &self.grid_host,
+                    TonicInterceptor::Signature(gen_tonic_signature_interceptor()),
+                )
+                .await?;
+                let mut request = Request::new(payload.clone());
+                set_tonic_canonical_body_digest(&mut request, &canonical)?;
+                let response = client.scanner_scoped_dirty_usage_ack(request).await?.into_inner();
+                let body = canonical_scoped_dirty_usage_response(&canonical, &response)
+                    .map_err(|_| Error::other("scoped dirty usage capability response is too large"))?;
+                verify_tonic_rpc_response_proof(&body, response.response_proof.as_ref())?;
+                if response.protocol_version != SCOPED_DIRTY_USAGE_PROTOCOL_VERSION
+                    || response.owner_id != payload.owner_id
+                    || response.instance_id != payload.instance_id
+                    || response.max_entries != SCOPED_DIRTY_USAGE_MAX_ENTRIES
+                    || response.max_request_bytes != SCOPED_DIRTY_USAGE_MAX_REQUEST_BYTES
+                    || response.cleared != 0
+                {
+                    return Err(Error::other("scoped dirty usage capability response does not match request"));
+                }
+                Ok(response.supported)
+            }
+            .await,
+        )
+        .await
+    }
+
     pub async fn acknowledge_scanner_dirty_usage(&self, instance_id: String, generation: u64) -> Result<ScannerPeerActivity> {
         let result = self
             .scanner_activity_request_with_protocol(instance_id.clone(), generation, SCANNER_ACTIVITY_PROTOCOL_VERSION)

crates/ecstore/src/core/pools.rs

@@ -5493,6 +5493,7 @@ where
     fence.ensure_held()?;
     let mut opts = ObjectOptions {
         max_parity: true,
+        write_completion: crate::object_api::WriteCompletion::TailDrained,
         no_lock: true,
         http_preconditions: Some(pool_meta_cas_preconditions(token, object)?),
         ..Default::default()
@@ -14412,6 +14413,7 @@ impl ECStore {
                 encoded.clone(),
                 &ObjectOptions {
                     max_parity: true,
+                    write_completion: crate::object_api::WriteCompletion::TailDrained,
                     http_preconditions: Some(HTTPPreconditions {
                         if_none_match: Some("*".to_string()),
                         ..Default::default()
@@ -14566,6 +14568,7 @@ impl ECStore {
                 encoded,
                 &ObjectOptions {
                     max_parity: true,
+                    write_completion: crate::object_api::WriteCompletion::TailDrained,
                     http_preconditions: Some(http_preconditions),
                     ..Default::default()
                 },
@@ -14957,6 +14960,7 @@ impl ECStore {
                 encoded,
                 &ObjectOptions {
                     max_parity: true,
+                    write_completion: crate::object_api::WriteCompletion::TailDrained,
                     http_preconditions: Some(HTTPPreconditions {
                         if_match: Some(etag),
                         ..Default::default()

crates/ecstore/src/disk/disk_store.rs

@@ -317,6 +317,22 @@ impl DiskStoreRenameDataExt for LocalDiskWrapper {
         dst_path: &str,
         external_guard: Option<Arc<dyn Send + Sync>>,
     ) -> Result<RenameDataResp> {
+        self.rename_data_observed(src_volume, src_path, fi, dst_volume, dst_path, external_guard)
+            .await
+            .result
+    }
+}
+
+impl LocalDiskWrapper {
+    pub(in crate::disk) async fn rename_data_observed(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        fi: &FileInfo,
+        dst_volume: &str,
+        dst_path: &str,
+        external_guard: Option<Arc<dyn Send + Sync>>,
+    ) -> super::RenameDataObservation {
         let operation = self.clone();
         let src_volume = src_volume.to_owned();
         let src_path = src_path.to_owned();
@@ -333,22 +349,35 @@ impl DiskStoreRenameDataExt for LocalDiskWrapper {
         } else {
             get_max_timeout_duration()
         };
-        run_owned_mutation(external_guard, move || async move {
+        let observed = run_owned_mutation(external_guard, move || async move {
-            operation
+            let mut preflight_rejection = None;
+            let result = operation
                 .track_disk_health_mutation(
                     "rename_data",
                     DiskMetricMutation::Write,
                     || async {
-                        operation
+                        // Preserve the former DiskAPI future's single boxing boundary.
-                            .disk
+                        let observed =
-                            .rename_data_borrowed(&src_volume, &src_path, &fi, &dst_volume, &dst_path)
+                            Box::pin(
-                            .await
+                                operation
+                                    .disk
+                                    .rename_data_observed(&src_volume, &src_path, &fi, &dst_volume, &dst_path),
+                            )
+                            .await;
+                        preflight_rejection = observed.preflight_rejection;
+                        observed.result
                     },
                     timeout_duration,
                 )
-                .await
+                .await;
+            // Health tracking must observe the real disk error, not an Ok tuple.
+            Ok(super::RenameDataObservation {
+                result,
+                preflight_rejection,
+            })
         })
-        .await
+        .await;
+        observed.unwrap_or_else(|error| super::RenameDataObservation::unknown(Err(error)))
     }
 }
 
@@ -2588,6 +2617,46 @@ mod tests {
         assert_eq!(wrapper.metrics_snapshot().api_calls.get("unknown"), Some(&1));
     }
 
+    #[tokio::test]
+    async fn rename_preflight_evidence_preserves_health_errors_and_owned_reply() {
+        for source_exists in [false, true] {
+            for guarded in [false, true] {
+                let dir = tempfile::tempdir().expect("temp dir should be created");
+                let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be valid UTF-8"))
+                    .expect("endpoint should parse");
+                let disk = Arc::new(LocalDisk::new(&endpoint, false).await.expect("local disk should be created"));
+                if source_exists {
+                    disk.make_volume("source").await.expect("source volume should exist");
+                }
+                let wrapper = LocalDiskWrapper::new(disk, false);
+                let drops = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+                let external_guard = guarded.then(|| Arc::new(DropProbe(Arc::clone(&drops))) as Arc<dyn Send + Sync>);
+                let mut file_info = FileInfo::new("object", 1, 0);
+                file_info.mod_time = Some(::time::OffsetDateTime::now_utc());
+                file_info.erasure.index = 1;
+                let observed = wrapper
+                    .rename_data_observed("source", "object", &file_info, "missing-destination", "object", external_guard)
+                    .await;
+                assert!(observed.rejected_before_publication(), "normal access rejection must carry proof");
+                assert!(matches!(observed.result, Err(DiskError::VolumeNotFound)));
+                let snapshot = wrapper.metrics_snapshot();
+                assert_eq!(snapshot.api_calls.get("rename_data"), Some(&1));
+                assert_eq!(snapshot.total_writes, 0, "health tracking must not observe the rejection as Ok");
+                assert_eq!(drops.load(Ordering::SeqCst), usize::from(guarded));
+
+                wrapper.health.force_runtime_state_for_test(RuntimeDriveHealthState::Offline);
+                let observed = wrapper
+                    .rename_data_observed("source", "object", &file_info, "missing-destination", "object", None)
+                    .await;
+                assert!(!observed.rejected_before_publication(), "wrapper errors carry no local preflight proof");
+                assert!(matches!(observed.result, Err(DiskError::FaultyDisk)));
+                let snapshot = wrapper.metrics_snapshot();
+                assert_eq!(snapshot.total_errors_availability, 1);
+                assert_eq!(snapshot.total_writes, 0);
+            }
+        }
+    }
+
     #[tokio::test]
     async fn local_disk_health_wrapper_counts_returned_availability_errors() {
         let dir = tempfile::tempdir().expect("temp dir should be created");

crates/ecstore/src/disk/mod.rs

@@ -75,6 +75,25 @@ use time::OffsetDateTime;
 use tokio::io::{AsyncRead, AsyncWrite};
 use uuid::Uuid;
 
+/// Local preflight evidence stays outside DiskAPI and the RPC response format.
+pub(crate) struct RenameDataObservation {
+    pub(crate) result: Result<RenameDataResp>,
+    preflight_rejection: Option<local::LocalRenamePreflightRejection>,
+}
+
+impl RenameDataObservation {
+    fn unknown(result: Result<RenameDataResp>) -> Self {
+        Self {
+            result,
+            preflight_rejection: None,
+        }
+    }
+
+    pub(crate) fn rejected_before_publication(&self) -> bool {
+        self.result.is_err() && self.preflight_rejection.is_some()
+    }
+}
+
 const QUOTA_MUTATION_FENCE_PREFIX: &str = "tmp/quota-mutation-fences/";
 pub(crate) const QUOTA_MUTATION_FENCE_METADATA_SUFFIX: &str = "quota-mutation-fence-token";
 
@@ -711,6 +730,36 @@ impl Disk {
             .await
     }
 
+    pub(crate) async fn rename_data_borrowed_with_fence_observed(
+        &self,
+        src_volume: &str,
+        src_path: &str,
+        fi: &FileInfo,
+        dst_volume: &str,
+        dst_path: &str,
+        scanner_publication_lease_token: Option<Uuid>,
+    ) -> RenameDataObservation {
+        match self {
+            Disk::Local(local_disk) => {
+                local_disk
+                    .rename_data_observed(src_volume, src_path, fi, dst_volume, dst_path, None)
+                    .await
+            }
+            Disk::Remote(remote_disk) => RenameDataObservation::unknown(
+                remote_disk
+                    .rename_data_borrowed_with_fence(
+                        src_volume,
+                        src_path,
+                        fi,
+                        dst_volume,
+                        dst_path,
+                        scanner_publication_lease_token,
+                    )
+                    .await,
+            ),
+        }
+    }
+
     pub(crate) async fn rename_data_borrowed_with_fence(
         &self,
         src_volume: &str,

crates/ecstore/src/object_api/types.rs

@@ -870,6 +870,18 @@ impl TierFreeVersionReceiptSink {
     }
 }
 
+/// Internal PUT completion boundary; this does not change fsync or write quorum.
+#[doc(hidden)]
+#[derive(Debug, Default, Clone, Copy, PartialEq, Eq)]
+pub enum WriteCompletion {
+    /// Return at write quorum when the commit owner can retain its guards.
+    #[default]
+    Quorum,
+    /// Drain the rename fan-out before returning. Minority failures still heal
+    /// after a successful quorum commit; this does not require every disk to succeed.
+    TailDrained,
+}
+
 #[derive(Default, Clone)]
 pub struct ObjectOptions {
     // Use the maximum parity (N/2), used when saving server configuration files
@@ -896,6 +908,10 @@ pub struct ObjectOptions {
     /// Persisted bucket incarnation observed before authorization.
     pub expected_bucket_incarnation_id: Option<Uuid>,
     pub no_lock: bool,
+    /// Control-plane writers that immediately read or CAS the same namespace
+    /// key use TailDrained without changing namespace lock ownership.
+    #[doc(hidden)]
+    pub write_completion: WriteCompletion,
     /// True when an upper layer already holds the object read lock before
     /// forwarding a no_lock read to the set layer.
     pub metadata_cache_safe: bool,
@@ -940,6 +956,9 @@ pub struct ObjectOptions {
     pub preserve_etag: Option<String>,
     pub metadata_chg: bool,
     pub http_preconditions: Option<HTTPPreconditions>,
+    /// Internal create-only writes may also preserve an acknowledged deletion.
+    /// Evaluated with `http_preconditions` under the namespace commit lock.
+    pub preserve_delete_marker: bool,
 
     pub delete_replication: Option<ReplicationState>,
     pub delete_replication_config_snapshot: Option<Arc<DeleteReplicationConfigSnapshot>>,

crates/ecstore/src/runtime/instance.rs

@@ -78,6 +78,21 @@ pub(crate) struct ScannerPublicationLeaseEntry {
     pub(crate) _operation_guard: OwnedRwLockReadGuard<()>,
 }
 
+pub(crate) struct NamespaceCommitGuard {
+    ctx: Arc<InstanceContext>,
+    counted: bool,
+}
+
+impl Drop for NamespaceCommitGuard {
+    fn drop(&mut self) {
+        if self.counted {
+            // Publish the new generation before a zero-pending publication probe.
+            self.ctx.advance_namespace_commit_generation();
+            self.ctx.namespace_commits.fetch_sub(1, Ordering::AcqRel);
+        }
+    }
+}
+
 /// Runtime state owned by a single `ECStore` instance.
 ///
 /// This is intentionally minimal in the first migration slice; subsequent
@@ -209,9 +224,13 @@ pub struct InstanceContext {
     /// Last storage-owned movement snapshot observed under the operation
     /// gate. SetDisks cache writers fail closed until ECStore refreshes it.
     scanner_publication_state: AtomicU8,
+    namespace_commits: AtomicU64,
+    namespace_commit_generation: AtomicU64,
     /// Resolves object-encryption material at the application boundary.
     object_encryption_resolver: OnceLock<Arc<dyn ObjectEncryptionResolver>>,
     tier_delete_journal_recovery_stores: std::sync::Mutex<HashSet<Uuid>>,
+    #[cfg(test)]
+    suppress_tier_delete_journal_recovery: bool,
     transition_transaction_recovery_stores: std::sync::Mutex<HashSet<Uuid>>,
     tier_delete_journal_recovery_wakeup: tokio::sync::Notify,
 }
@@ -256,8 +275,12 @@ impl InstanceContext {
             data_movement_generation_exhausted: AtomicBool::new(false),
             data_movement_generation_notify: Arc::new(Notify::new()),
             scanner_publication_state: AtomicU8::new(SCANNER_PUBLICATION_STATE_UNKNOWN),
+            namespace_commits: AtomicU64::new(0),
+            namespace_commit_generation: AtomicU64::new(0),
             object_encryption_resolver: OnceLock::new(),
             tier_delete_journal_recovery_stores: std::sync::Mutex::new(HashSet::new()),
+            #[cfg(test)]
+            suppress_tier_delete_journal_recovery: false,
             transition_transaction_recovery_stores: std::sync::Mutex::new(HashSet::new()),
             tier_delete_journal_recovery_wakeup: tokio::sync::Notify::new(),
         }
@@ -385,6 +408,36 @@ impl InstanceContext {
             && self.scanner_publication_state.load(Ordering::Acquire) == SCANNER_PUBLICATION_STATE_ALLOWED
     }
 
+    pub(crate) fn begin_namespace_commit(self: &Arc<Self>) -> Arc<NamespaceCommitGuard> {
+        let counted = self
+            .namespace_commits
+            .fetch_update(Ordering::AcqRel, Ordering::Acquire, |count| count.checked_add(1))
+            .is_ok();
+        if counted {
+            self.advance_namespace_commit_generation();
+        } else {
+            self.namespace_commit_generation.store(u64::MAX, Ordering::Release);
+        }
+        Arc::new(NamespaceCommitGuard {
+            ctx: Arc::clone(self),
+            counted,
+        })
+    }
+
+    fn advance_namespace_commit_generation(&self) {
+        let _ = self
+            .namespace_commit_generation
+            .fetch_update(Ordering::AcqRel, Ordering::Acquire, |generation| Some(generation.saturating_add(1)));
+    }
+
+    pub(crate) fn namespace_commit_generation(&self) -> u64 {
+        self.namespace_commit_generation.load(Ordering::Acquire)
+    }
+
+    pub(crate) fn namespace_commits_pending(&self) -> bool {
+        self.namespace_commits.load(Ordering::Acquire) != 0 || self.namespace_commit_generation() == u64::MAX
+    }
+
     pub(crate) fn set_scanner_publication_state(&self, blocked: bool) {
         self.scanner_publication_state.store(
             if blocked {
@@ -640,12 +693,21 @@ impl InstanceContext {
     }
 
     pub(crate) fn mark_tier_delete_journal_recovery_started(&self, store_id: Uuid) -> bool {
+        #[cfg(test)]
+        if self.suppress_tier_delete_journal_recovery {
+            return false;
+        }
         self.tier_delete_journal_recovery_stores
             .lock()
             .unwrap_or_else(std::sync::PoisonError::into_inner)
             .insert(store_id)
     }
 
+    #[cfg(test)]
+    pub(crate) fn suppress_tier_delete_journal_recovery_for_test(&mut self) {
+        self.suppress_tier_delete_journal_recovery = true;
+    }
+
     pub(crate) fn mark_transition_transaction_recovery_started(&self, store_id: Uuid) -> bool {
         self.transition_transaction_recovery_stores
             .lock()
@@ -756,6 +818,50 @@ pub fn bootstrap_ctx() -> Arc<InstanceContext> {
 mod tests {
     use super::*;
 
+    #[test]
+    fn namespace_commit_guards_are_instance_local_and_count_until_last_owner() {
+        let first = Arc::new(InstanceContext::new());
+        let other = Arc::new(InstanceContext::new());
+        first.set_scanner_publication_state(false);
+        other.set_scanner_publication_state(false);
+        assert!(first.scanner_publication_state_allowed());
+        let one = first.begin_namespace_commit();
+        let shared_owner = Arc::clone(&one);
+        let two = first.begin_namespace_commit();
+        assert!(first.namespace_commits_pending());
+        assert!(first.scanner_publication_state_allowed(), "pending writes must not block scan admission");
+        assert_eq!(first.namespace_commit_generation(), 2);
+        assert!(!other.namespace_commits_pending());
+        assert_eq!(other.namespace_commit_generation(), 0);
+        assert!(other.scanner_publication_state_allowed());
+        drop(one);
+        assert_eq!(first.namespace_commit_generation(), 2);
+        drop(shared_owner);
+        assert!(first.namespace_commits_pending());
+        assert_eq!(first.namespace_commit_generation(), 3);
+        drop(two);
+        assert!(!first.namespace_commits_pending());
+        assert_eq!(first.namespace_commit_generation(), 4);
+        assert!(first.scanner_publication_state_allowed());
+    }
+
+    #[test]
+    fn namespace_commit_counter_exhaustion_keeps_publication_blocked() {
+        for (count, generation) in [(0, u64::MAX - 1), (u64::MAX, 0)] {
+            let ctx = Arc::new(InstanceContext::new());
+            ctx.set_scanner_publication_state(false);
+            ctx.namespace_commits.store(count, Ordering::Release);
+            ctx.namespace_commit_generation.store(generation, Ordering::Release);
+            let guard = ctx.begin_namespace_commit();
+            assert!(ctx.namespace_commits_pending());
+            assert_eq!(ctx.namespace_commit_generation(), u64::MAX);
+            drop(guard);
+            assert!(ctx.namespace_commits_pending());
+            assert_eq!(ctx.namespace_commit_generation(), u64::MAX);
+            assert_eq!(ctx.namespace_commits.load(Ordering::Acquire), count);
+        }
+    }
+
     // The SetupType inputs must derive the exact (is_erasure,
     // is_dist_erasure, is_erasure_sd) triples that the original three
     // process-global erasure bools produced via update_erasure_type().
@@ -1073,6 +1179,12 @@ mod tests {
         assert!(!ctx_a.mark_tier_delete_journal_recovery_started(store_a));
         assert!(ctx_a.mark_tier_delete_journal_recovery_started(store_b));
         assert!(ctx_b.mark_tier_delete_journal_recovery_started(store_a));
+
+        let mut manual_ctx = InstanceContext::new();
+        manual_ctx.suppress_tier_delete_journal_recovery_for_test();
+        assert!(!manual_ctx.mark_tier_delete_journal_recovery_started(store_a));
+        assert!(!manual_ctx.mark_tier_delete_journal_recovery_started(store_b));
+        assert!(ctx_b.mark_tier_delete_journal_recovery_started(store_b));
     }
 
     #[test]

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

@@ -701,7 +701,7 @@ impl WarmBackend for MockWarmBackend {
         Ok(version)
     }
 
-    async fn get(&self, object: &str, _rv: &str, opts: WarmBackendGetOpts) -> Result<ReadCloser, std::io::Error> {
+    async fn get(&self, object: &str, rv: &str, opts: WarmBackendGetOpts) -> Result<ReadCloser, std::io::Error> {
         self.precondition().await?;
         let barrier = self.inner.get_barrier.lock().await.take();
         if let Some(barrier) = barrier {
@@ -719,6 +719,9 @@ impl WarmBackend for MockWarmBackend {
         let Some(stored) = objects.get(object) else {
             return Err(std::io::Error::new(std::io::ErrorKind::NotFound, "mock object not found"));
         };
+        if !rv.is_empty() && stored.remote_version_id != rv {
+            return Err(std::io::Error::new(std::io::ErrorKind::NotFound, "NoSuchVersion"));
+        }
         let bytes = &stored.bytes;
 
         let start = opts.start_offset.max(0) as usize;

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

@@ -2346,6 +2346,10 @@ impl WarmBackend for SharedWarmBackendProxy {
         self.0.probe_transition_candidate(object).await
     }
 
+    async fn probe_transition_version(&self, object: &str, remote_version_id: &str) -> io::Result<TransitionCandidateProbe> {
+        self.0.probe_transition_version(object, remote_version_id).await
+    }
+
     async fn in_use(&self) -> io::Result<bool> {
         self.0.in_use().await
     }
@@ -2458,6 +2462,15 @@ impl TierOperationLease {
         Ok(())
     }
 
+    pub(crate) async fn probe_transition_version(
+        &self,
+        object: &str,
+        remote_version_id: &str,
+    ) -> io::Result<TransitionCandidateProbe> {
+        self.validate_remote_version_id(remote_version_id)?;
+        self.inner.driver.probe_transition_version(object, remote_version_id).await
+    }
+
     pub(crate) fn is_current_generation(&self) -> bool {
         lock_unpoisoned(&self.runtime)
             .generations
@@ -3528,7 +3541,7 @@ impl TierConfigMgr {
         // Get tier configuration and create new driver
         let tier_config = self.tiers.get(tier_name).ok_or_else(|| ERR_TIER_NOT_FOUND.clone())?;
 
-        let driver = new_warm_backend(tier_config, false).await?;
+        let driver = construct_warm_backend(tier_config).await?;
 
         self.replace_driver(tier_name, driver)?;
         Ok(self
@@ -4473,6 +4486,11 @@ impl TierConfigMgr {
                     let committed_coordinator_intent =
                         committed_tier_mutation_intent(coordinator_intent.as_ref(), &committed_config_etag)
                             .map_err(TierConfigUpdateError::Save)?;
+                    // Persist Committed before notifying refresh; a Prepared disk record
+                    // would restore the prepared block and invalidate our publish allowance.
+                    let coordinator_commit =
+                        commit_coordinator_tier_mutation_intent(api.clone(), coordinator_intent.as_ref(), &committed_config_etag)
+                            .await;
                     if let Some(intent) = committed_coordinator_intent.as_ref() {
                         TierConfigMgr::apply_committed_mutation_intent_block(&handle, intent)
                             .await
@@ -4483,9 +4501,9 @@ impl TierConfigMgr {
                                 .map_err(TierConfigUpdateError::Publish)?,
                         );
                     }
-                    commit_coordinator_tier_mutation_intent(api.clone(), coordinator_intent.as_ref(), &committed_config_etag)
+                    // Config is already saved: retain the committed fence and wake recovery
-                        .await
+                    // even when the coordinator commit failed or its outcome is unknown.
-                        .map_err(TierConfigUpdateError::Save)?;
+                    coordinator_commit.map_err(TierConfigUpdateError::Save)?;
                     if coordinated_config_update {
                         drop(update.take());
                         drop(config_lock.take());
@@ -10590,6 +10608,11 @@ mod tests {
             .expect_err("coordinator committed-state CAS failure must be observable");
         assert!(matches!(err, TierConfigUpdateError::Save(_)));
         assert!(manager.read().await.tiers.contains_key("COLD-A"));
+        assert!(TierConfigMgr::has_committed_mutation_block(&manager).await);
+        let refresh = TierConfigMgr::mutation_refresh_notifier(&manager).await;
+        tokio::time::timeout(Duration::from_secs(1), refresh.notified())
+            .await
+            .expect("failed coordinator commit must notify recovery after saving config");
         let blocked = match TierConfigMgr::acquire_operation_lease(&manager, "COLD-A").await {
             Ok(_) => panic!("failed coordinator commit CAS must retain the local committed fence"),
             Err(err) => err,
@@ -14316,6 +14339,12 @@ mod tests {
         after_commit: bool,
     }
 
+    #[derive(Debug, Default)]
+    struct CasCoordinatorCommitBarrier {
+        arrived: tokio::sync::Notify,
+        release: tokio::sync::Notify,
+    }
+
     #[derive(Debug)]
     struct CasConfigStore {
         objects: tokio::sync::Mutex<HashMap<String, (Vec<u8>, String)>>,
@@ -14328,6 +14357,7 @@ mod tests {
         fail_delete_prefix: tokio::sync::Mutex<Option<(String, usize)>>,
         delete_log: tokio::sync::Mutex<Vec<String>>,
         list_barrier: tokio::sync::Mutex<Option<Arc<CasListBarrier>>>,
+        coordinator_commit_barrier: tokio::sync::Mutex<Option<Arc<CasCoordinatorCommitBarrier>>>,
         intent_list_calls: AtomicUsize,
         fail_reference_walk: AtomicBool,
         reference_walk_send_count: AtomicUsize,
@@ -14350,6 +14380,7 @@ mod tests {
                 fail_delete_prefix: tokio::sync::Mutex::new(None),
                 delete_log: tokio::sync::Mutex::new(Vec::new()),
                 list_barrier: tokio::sync::Mutex::new(None),
+                coordinator_commit_barrier: tokio::sync::Mutex::new(None),
                 intent_list_calls: AtomicUsize::new(0),
                 fail_reference_walk: AtomicBool::new(false),
                 reference_walk_send_count: AtomicUsize::new(0),
@@ -14541,6 +14572,19 @@ mod tests {
             }
             let mut payload = Vec::new();
             tokio::io::AsyncReadExt::read_to_end(&mut data.stream, &mut payload).await?;
+            if object.starts_with(crate::services::tier::tier_mutation_intent::TIER_COORDINATOR_MUTATION_INTENT_RECORD_PREFIX)
+                && opts
+                    .http_preconditions
+                    .as_ref()
+                    .and_then(HTTPPreconditions::if_match_value)
+                    .is_some()
+            {
+                let barrier = self.coordinator_commit_barrier.lock().await.take();
+                if let Some(barrier) = barrier {
+                    barrier.arrived.notify_one();
+                    barrier.release.notified().await;
+                }
+            }
             let race_rewrite = if opts
                 .http_preconditions
                 .as_ref()
@@ -15638,14 +15682,7 @@ mod tests {
         );
     }
 
-    #[tokio::test]
+    async fn assert_lifecycle_only_reference_obeys_force(clear: bool, force: bool) {
-    async fn force_remove_and_save_bypasses_lifecycle_only_reference() {
-        // rustfs/rustfs#6832: reproduces the admin RemoveTier path (not just the lower-level
-        // reference-proof function) for a tier with zero transitioned objects but a lifecycle
-        // rule still pointing at it — the exact shape of
-        // `test_manual_transition_async_tier_failure_reports_terminal_partial` in e2e_test,
-        // which force-removes a tier a lifecycle rule still references to simulate a
-        // decommissioned backend.
         let store = Arc::new(CasConfigStore::default());
         let tier = build_rustfs_tier("COLD-A");
         let mut persisted = empty_mgr();
@@ -15686,22 +15723,55 @@ mod tests {
 
         let manager = TierConfigMgr::new();
         manager.write().await.tiers.insert("COLD-A".to_string(), tier);
-        TierConfigMgr::remove_and_save_with(&manager, store.clone(), "COLD-A", true)
+        let mutation = if clear {
-            .await
+            TierCandidateMutation::Clear(force)
-            .expect("force remove must bypass a lifecycle-config-only reference");
+        } else {
+            TierCandidateMutation::Remove("COLD-A".to_string(), force)
+        };
+        let result = TIER_DRIVER_TEST_FACTORY
+            .scope(
+                healthy_driver_factory(),
+                TierConfigMgr::update_candidate_with_config_lock(&manager, store.clone(), mutation),
+            )
+            .await;
+        if force {
+            result.expect("force mutation must bypass a lifecycle-config-only reference");
+        } else {
+            let err = result.expect_err("non-force mutation must reject a lifecycle-only reference");
+            let TierConfigUpdateError::Publish(err) = err else {
+                panic!("non-force mutation must fail during reference proof: {err:?}");
+            };
+            assert_eq!(err.code, ERR_TIER_BACKEND_IN_USE.code);
+            assert!(err.message.contains("move-current"), "{err}");
+        }
 
-        assert!(!manager.read().await.tiers.contains_key("COLD-A"));
+        assert_eq!(manager.read().await.tiers.contains_key("COLD-A"), !force);
-        assert!(
+        assert_eq!(
-            !load_tier_config_for_update(store)
+            load_tier_config_for_update(store)
                 .await
                 .expect("config should still reload")
                 .0
                 .tiers
                 .contains_key("COLD-A"),
-            "force removal must persist the empty candidate"
+            !force,
+            "persisted state must match the force mutation result"
         );
     }
 
+    #[tokio::test]
+    async fn remove_with_config_lock_obeys_force_for_lifecycle_only_reference() {
+        for force in [false, true] {
+            assert_lifecycle_only_reference_obeys_force(false, force).await;
+        }
+    }
+
+    #[tokio::test]
+    async fn clear_with_config_lock_obeys_force_for_lifecycle_only_reference() {
+        for force in [false, true] {
+            assert_lifecycle_only_reference_obeys_force(true, force).await;
+        }
+    }
+
     #[tokio::test]
     async fn zero_reference_proof_blocks_clear_before_config_save() {
         let store = Arc::new(CasConfigStore::default());
@@ -17242,6 +17312,98 @@ mod tests {
         assert_ne!(manager_a.read().await.empty(), manager_b.read().await.empty());
     }
 
+    async fn assert_coordinator_commit_refresh_succeeds(mutation: TierCandidateMutation) {
+        let adding = matches!(mutation, TierCandidateMutation::Add(..));
+        let manager = TierConfigMgr::new();
+        let store = Arc::new(CasConfigStore::default());
+        if !adding {
+            let mut persisted = empty_mgr();
+            persisted.tiers.insert("COLD-A".to_string(), build_rustfs_tier("COLD-A"));
+            persisted
+                .save_tiering_config_if_current(store.clone(), None)
+                .await
+                .expect("existing tier fixture should persist");
+            let mut guard = manager.write().await;
+            install_lease_backend(&mut guard, "COLD-A", LeaseTestBackend::ready("old"));
+        }
+        let barrier = Arc::new(CasCoordinatorCommitBarrier::default());
+        *store.coordinator_commit_barrier.lock().await = Some(barrier.clone());
+        let update_manager = manager.clone();
+        let update_store = store.clone();
+        let update = tokio::spawn(async move {
+            TIER_DRIVER_TEST_FACTORY
+                .scope(
+                    healthy_driver_factory(),
+                    TIER_MUTATION_TEST_PEERS.scope(
+                        Vec::new(),
+                        TierConfigMgr::update_candidate_with_config_lock(&update_manager, update_store, mutation),
+                    ),
+                )
+                .await
+        });
+        tokio::time::timeout(Duration::from_secs(5), barrier.arrived.notified())
+            .await
+            .expect("mutation should reach coordinator commit after saving config");
+        assert_eq!(
+            load_tier_config_for_update(store.clone())
+                .await
+                .expect("saved config should be readable before coordinator commit")
+                .0
+                .tiers
+                .contains_key("COLD-A"),
+            adding
+        );
+        assert_eq!(
+            TierConfigMgr::load_coordinator_mutation_intents(store.clone())
+                .await
+                .expect("coordinator intent should remain readable")[0]
+                .state,
+            TierMutationIntentState::Prepared
+        );
+
+        let lock_requests = lock_unpoisoned(&store.lock_requests).len();
+        // Also exercise an independently scheduled refresh while the durable
+        // coordinator record is still Prepared, before its commit notification.
+        TierConfigMgr::request_committed_mutation_refresh(&manager).await;
+        TIER_MUTATION_TEST_PEERS
+            .scope(Vec::new(), async {
+                let worker = TierConfigMgr::refresh_tier_config_handle_with(manager.clone(), store.clone());
+                tokio::pin!(worker);
+                tokio::time::timeout(Duration::from_secs(5), async {
+                    while lock_unpoisoned(&store.lock_requests).len() == lock_requests {
+                        tokio::select! {
+                            _ = &mut worker => panic!("refresh worker must remain available"),
+                            _ = tokio::task::yield_now() => {}
+                        }
+                    }
+                })
+                .await
+                .expect("refresh should reconcile the Prepared record before waiting for the config lock");
+                barrier.release.notify_one();
+                let result = tokio::time::timeout(Duration::from_secs(5), async {
+                    tokio::select! {
+                        _ = &mut worker => panic!("refresh worker must remain available"),
+                        result = update => result.expect("tier mutation task should join"),
+                    }
+                })
+                .await
+                .expect("tier mutation should finish with refresh running");
+                result.expect("saved tier mutation must publish successfully on the first attempt");
+            })
+            .await;
+        assert_eq!(manager.read().await.tiers.contains_key("COLD-A"), adding);
+    }
+
+    #[tokio::test]
+    async fn tier_add_succeeds_with_refresh_during_coordinator_commit() {
+        assert_coordinator_commit_refresh_succeeds(TierCandidateMutation::Add(build_rustfs_tier("COLD-A"), true)).await;
+    }
+
+    #[tokio::test]
+    async fn tier_remove_succeeds_with_refresh_during_coordinator_commit() {
+        assert_coordinator_commit_refresh_succeeds(TierCandidateMutation::Remove("COLD-A".to_string(), true)).await;
+    }
+
     async fn committed_refresh_fixture(fail_cleanup: bool) -> (Arc<RwLock<TierConfigMgr>>, Arc<CasConfigStore>, uuid::Uuid) {
         let manager = TierConfigMgr::new();
         {

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

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use serde::{Deserialize, Deserializer, Serialize, Serializer, de};
 
@@ -145,7 +143,7 @@ mod tests {
 
         assert_eq!(creds.access_key, "access");
         assert_eq!(creds.secret_key, "secret");
-        assert_eq!(creds.creds_json.as_slice(), &service_account[..]);
+        assert_eq!(creds.creds_json.as_slice(), service_account);
 
         let wire = serde_json::to_value(&creds).expect("madmin tier credentials should encode");
         assert_eq!(wire["access"], "access");
@@ -162,7 +160,7 @@ mod tests {
         .expect("the former RustFS field names and byte-array encoding should remain readable");
         assert_eq!(legacy.access_key, "legacy-access");
         assert_eq!(legacy.secret_key, "legacy-secret");
-        assert_eq!(legacy.creds_json.as_slice(), &service_account[..]);
+        assert_eq!(legacy.creds_json.as_slice(), service_account);
     }
 
     #[test]

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

@@ -460,6 +460,7 @@ where
         data,
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_none_match: Some("*".to_string()),
                 ..Default::default()
@@ -556,6 +557,7 @@ where
         data,
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_match: Some(current_etag.to_string()),
                 ..Default::default()

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

@@ -494,6 +494,7 @@ where
         data,
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_none_match: Some("*".to_string()),
                 ..Default::default()
@@ -549,6 +550,7 @@ where
         data,
         &ObjectOptions {
             max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
             http_preconditions: Some(HTTPPreconditions {
                 if_match: Some(current.record_etag.clone()),
                 ..Default::default()

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

@@ -37,9 +37,10 @@ use crate::services::tier::{
 use bytes::Bytes;
 use http::StatusCode;
 use rustfs_s3_client::credentials::{Credentials, SignatureType, Static, Value};
-use rustfs_s3_client::transition_api::{BucketLookupType, Options, TransitionClient, TransitionCore};
+use rustfs_s3_client::transition_api::{BucketLookupType, Options, TransitionClient, TransitionClientTimeouts, TransitionCore};
 use rustfs_s3_client::{
     admin_handler_utils::AdminError,
+    api_error_response::to_error_response,
     api_put_object::{AdvancedPutOptions, PutObjectOptions},
     transition_api::{ReadCloser, ReaderImpl},
 };
@@ -48,11 +49,14 @@ use rustfs_utils::egress::validate_outbound_url;
 use rustfs_utils::http::headers::{
     CACHE_CONTROL, CONTENT_DISPOSITION, CONTENT_ENCODING, CONTENT_LANGUAGE, CONTENT_TYPE, EXPIRES, HeaderExt as _,
 };
-use s3s::dto::{ObjectLockLegalHoldStatus, ObjectLockRetentionMode, ReplicationStatus};
 use s3s::header::{
     X_AMZ_OBJECT_LOCK_LEGAL_HOLD, X_AMZ_OBJECT_LOCK_MODE, X_AMZ_OBJECT_LOCK_RETAIN_UNTIL_DATE, X_AMZ_REPLICATION_STATUS,
     X_AMZ_STORAGE_CLASS,
 };
+use s3s::{
+    S3ErrorCode,
+    dto::{ObjectLockLegalHoldStatus, ObjectLockRetentionMode, ReplicationStatus},
+};
 use std::collections::HashMap;
 use std::sync::Arc;
 use std::time::Duration;
@@ -141,6 +145,42 @@ pub trait WarmBackend {
     async fn probe_transition_candidate(&self, _object: &str) -> Result<TransitionCandidateProbe, std::io::Error> {
         Ok(TransitionCandidateProbe::Unsupported)
     }
+    async fn probe_transition_version(
+        &self,
+        object: &str,
+        remote_version_id: &str,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        if remote_version_id.is_empty() {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidInput,
+                "an exact tier probe requires a remote version ID",
+            ));
+        }
+        self.validate_remote_version_id(remote_version_id)?;
+        match self
+            .get(
+                object,
+                remote_version_id,
+                WarmBackendGetOpts {
+                    start_offset: 0,
+                    length: 1,
+                },
+            )
+            .await
+        {
+            Ok(_) => Ok(TransitionCandidateProbe::VersionedPresent(remote_version_id.to_string())),
+            Err(err) if matches!(to_error_response(&err).code, S3ErrorCode::InvalidRange) => {
+                Ok(TransitionCandidateProbe::VersionedPresent(remote_version_id.to_string()))
+            }
+            Err(err)
+                if err.kind() == std::io::ErrorKind::NotFound
+                    || matches!(to_error_response(&err).code, S3ErrorCode::NoSuchKey | S3ErrorCode::NoSuchVersion) =>
+            {
+                Ok(TransitionCandidateProbe::Missing)
+            }
+            Err(err) => Err(err),
+        }
+    }
     async fn in_use(&self) -> Result<bool, std::io::Error>;
 }
 
@@ -280,6 +320,27 @@ pub(crate) fn endpoint_authority(url: &url::Url) -> Result<String, std::io::Erro
     }
 }
 
+fn transition_timeout_from_env(env_key: &str, default_secs: u64) -> Duration {
+    Duration::from_secs(rustfs_utils::get_env_u64(env_key, default_secs))
+}
+
+pub(crate) fn transition_client_timeouts_from_env() -> TransitionClientTimeouts {
+    TransitionClientTimeouts::new(
+        transition_timeout_from_env(
+            rustfs_config::ENV_TIER_REMOTE_CONNECT_TIMEOUT_SECS,
+            rustfs_config::DEFAULT_TIER_REMOTE_CONNECT_TIMEOUT_SECS,
+        ),
+        transition_timeout_from_env(
+            rustfs_config::ENV_TIER_REMOTE_REQUEST_TIMEOUT_SECS,
+            rustfs_config::DEFAULT_TIER_REMOTE_REQUEST_TIMEOUT_SECS,
+        ),
+        transition_timeout_from_env(
+            rustfs_config::ENV_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS,
+            rustfs_config::DEFAULT_TIER_REMOTE_RESPONSE_BODY_IDLE_TIMEOUT_SECS,
+        ),
+    )
+}
+
 /// Build the [`WarmBackendS3`] shared by the S3-compatible warm backend providers.
 ///
 /// Credential, bucket, and endpoint validation run in this order because the
@@ -310,6 +371,7 @@ pub(crate) async fn new_s3_compatible_warm_backend(
         signer_type: SignatureType::SignatureV4,
         ..Default::default()
     }));
+    let timeouts = transition_client_timeouts_from_env();
     let opts = Options {
         creds,
         secure: u.scheme() == "https",
@@ -322,7 +384,7 @@ pub(crate) async fn new_s3_compatible_warm_backend(
     // Run the SSRF guard after the host-presence check so a host-less endpoint
     // keeps this constructor's stable error text.
     (params.validate_endpoint)(&u).map_err(|err| std::io::Error::other(format!("tier endpoint is not allowed: {err}")))?;
-    let client = TransitionClient::new(&endpoint, opts, params.provider_tag).await?;
+    let client = TransitionClient::new_with_timeouts(&endpoint, opts, params.provider_tag, timeouts).await?;
 
     let client = Arc::new(client);
     let core = TransitionCore(Arc::clone(&client));
@@ -437,6 +499,17 @@ impl WarmBackend for MeteredWarmBackend {
         Self::record(TierRequestOperation::Probe, result)
     }
 
+    async fn probe_transition_version(
+        &self,
+        object: &str,
+        remote_version_id: &str,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        Self::record(
+            TierRequestOperation::Probe,
+            self.inner.probe_transition_version(object, remote_version_id).await,
+        )
+    }
+
     async fn in_use(&self) -> Result<bool, std::io::Error> {
         Self::record(TierRequestOperation::InUse, self.inner.in_use().await)
     }

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

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use std::collections::HashMap;
 

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

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use std::collections::HashMap;
 

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

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use std::collections::{HashMap, HashSet};
 use std::future::Future;
@@ -146,11 +144,11 @@ pub struct WarmBackendGCS {
 
 impl WarmBackendGCS {
     pub async fn new(conf: &TierGCS, tier: &str) -> Result<Self, std::io::Error> {
-        if conf.creds == "" {
+        if conf.creds.is_empty() {
             return Err(std::io::Error::other("both access and secret keys are required"));
         }
 
-        if conf.bucket == "" {
+        if conf.bucket.is_empty() {
             return Err(std::io::Error::other("no bucket name was provided"));
         }
 
@@ -195,11 +193,11 @@ impl WarmBackendGCS {
     }
 
     pub fn get_dest(&self, object: &str) -> String {
-        let mut dest_obj = object.to_string();
+        if self.prefix.is_empty() {
-        if self.prefix != "" {
+            object.to_string()
-            dest_obj = format!("{}/{}", &self.prefix, object);
+        } else {
+            format!("{}/{}", self.prefix, object)
         }
-        return dest_obj;
     }
 }
 
@@ -223,7 +221,7 @@ impl WarmBackend for WarmBackendGCS {
         let bucket = gcs_bucket_resource_name(&self.bucket);
         let Ok(res) = Box::pin(
             self.client
-                .write_object(&bucket, &self.get_dest(object), Bytes::from(d))
+                .write_object(&bucket, self.get_dest(object), Bytes::from(d))
                 .send_buffered(),
         )
         .await
@@ -240,7 +238,7 @@ impl WarmBackend for WarmBackendGCS {
 
     async fn get(&self, object: &str, rv: &str, opts: WarmBackendGetOpts) -> Result<ReadCloser, std::io::Error> {
         let bucket = gcs_bucket_resource_name(&self.bucket);
-        let mut req = self.client.read_object(&bucket, &self.get_dest(object));
+        let mut req = self.client.read_object(&bucket, self.get_dest(object));
         let mut max_response_bytes = None;
         if let Some(generation) = parse_generation(rv)? {
             req = req.set_generation(generation);

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

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use std::collections::HashMap;
 

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

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use std::collections::HashMap;
 

crates/ecstore/src/services/tier/warm_backend_r2.rs

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use std::collections::HashMap;
 

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

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use std::collections::HashMap;
 

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

@@ -26,7 +26,7 @@ use crate::services::tier::{
     tier_config::TierS3,
     warm_backend::{
         TransitionCandidateIdentity, TransitionCandidateProbe, TransitionCandidateReconciler, WarmBackend, WarmBackendGetOpts,
-        build_transition_put_options, endpoint_authority,
+        build_transition_put_options, endpoint_authority, transition_client_timeouts_from_env,
     },
 };
 use http::HeaderMap;
@@ -139,6 +139,7 @@ impl WarmBackendS3 {
         } else {
             return Err(std::io::Error::other("insufficient parameters for S3 backend authentication"));
         }
+        let timeouts = transition_client_timeouts_from_env();
         let opts = Options {
             creds,
             secure: u.scheme() == "https",
@@ -147,7 +148,7 @@ impl WarmBackendS3 {
             ..Default::default()
         };
         let endpoint = endpoint_authority(&u)?;
-        let client = TransitionClient::new(&endpoint, opts, tier_type).await?;
+        let client = TransitionClient::new_with_timeouts(&endpoint, opts, tier_type, timeouts).await?;
 
         let client = Arc::new(client);
         let core = TransitionCore(Arc::clone(&client));
@@ -529,6 +530,10 @@ mod tests {
                 "HTTP/1.1 404 Not Found\r\nContent-Type: application/xml\r\nContent-Length: 63\r\nConnection: close\r\n\r\n<Error><Code>NoSuchKey</Code><Message>missing</Message></Error>",
                 "HTTP/1.1 404 Not Found\r\nContent-Type: application/xml\r\nContent-Length: 66\r\nConnection: close\r\n\r\n<Error><Code>NoSuchObject</Code><Message>missing</Message></Error>",
                 "HTTP/1.1 403 Forbidden\r\nContent-Type: application/xml\r\nContent-Length: 65\r\nConnection: close\r\n\r\n<Error><Code>AccessDenied</Code><Message>denied</Message></Error>",
+                "HTTP/1.1 404 Not Found\r\nContent-Type: application/xml\r\nContent-Length: 63\r\nConnection: close\r\n\r\n<Error><Code>NoSuchKey</Code><Message>missing</Message></Error>",
+                "HTTP/1.1 416 Range Not Satisfiable\r\nContent-Type: application/xml\r\nContent-Length: 72\r\nConnection: close\r\n\r\n<Error><Code>InvalidRange</Code><Message>empty version</Message></Error>",
+                "HTTP/1.1 404 Not Found\r\nContent-Type: application/xml\r\nContent-Length: 67\r\nConnection: close\r\n\r\n<Error><Code>NoSuchVersion</Code><Message>missing</Message></Error>",
+                "HTTP/1.1 404 Not Found\r\nContent-Type: application/xml\r\nContent-Length: 63\r\nConnection: close\r\n\r\n<Error><Code>NoSuchKey</Code><Message>missing</Message></Error>",
             ];
             let mut requests = Vec::new();
             for response in responses {
@@ -622,15 +627,52 @@ mod tests {
             .await
             .expect_err("an authorization failure must not be mistaken for a missing key");
         assert_eq!(to_error_response(&err).code, S3ErrorCode::AccessDenied);
+        assert_eq!(
+            backend
+                .probe_transition_candidate("delete-marker-hidden")
+                .await
+                .expect("a current delete marker should hide the data version"),
+            TransitionCandidateProbe::Missing
+        );
+        assert_eq!(
+            backend
+                .probe_transition_version("delete-marker-hidden", "historical-version")
+                .await
+                .expect("the stored historical version should be probed exactly"),
+            TransitionCandidateProbe::VersionedPresent("historical-version".to_string())
+        );
+        assert_eq!(
+            backend
+                .probe_transition_version("delete-marker-hidden", "missing-version")
+                .await
+                .expect("a missing exact version should be classified"),
+            TransitionCandidateProbe::Missing
+        );
+        assert_eq!(
+            backend
+                .probe_transition_version("missing-object", "historical-version")
+                .await
+                .expect("a missing key for an exact version probe should be classified"),
+            TransitionCandidateProbe::Missing
+        );
 
         let requests = fixture.await.expect("candidate fixture should join");
-        for request in requests {
+        for request in &requests[..6] {
             let request = request.to_ascii_lowercase();
             assert!(request.starts_with("get /bucket/"), "candidate discovery must use object GET");
             assert!(request.contains("\r\nrange: bytes=0-0\r\n"));
             assert!(!request.contains("?versioning"));
             assert!(!request.contains("?versions"));
         }
+        for request in &requests[6..] {
+            let request = request.to_ascii_lowercase();
+            assert!(request.starts_with("get /bucket/"), "exact discovery must use object GET");
+            assert!(request.contains("\r\nrange: bytes=0-0\r\n"));
+        }
+        assert!(!requests[5].to_ascii_lowercase().contains("versionid="));
+        assert!(requests[6].to_ascii_lowercase().contains("?versionid=historical-version"));
+        assert!(requests[7].to_ascii_lowercase().contains("?versionid=missing-version"));
+        assert!(requests[8].to_ascii_lowercase().contains("?versionid=historical-version"));
     }
 
     fn list_versions(versions: &[(&str, &str)], delete_markers: &[(&str, &str)], is_truncated: bool) -> ListVersionsResult {

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

@@ -15,8 +15,6 @@
 #![allow(unused_variables)]
 #![allow(unused_mut)]
 #![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
 
 use std::collections::HashMap;
 

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

@@ -0,0 +1,385 @@
+// 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.
+
+//! Pure metadata quorum and early-stop decisions for `SetDisks` reads.
+//!
+//! Disk scheduling, coalescing, cancellation, and late shard materialization
+//! remain with their existing owners; this module only classifies observations.
+
+use crate::diagnostics::get::{
+    GET_METADATA_EARLY_STOP_REASON_CONFLICTING_METADATA, GET_METADATA_EARLY_STOP_REASON_DELETE_MARKER,
+    GET_METADATA_EARLY_STOP_REASON_ERROR, GET_METADATA_EARLY_STOP_REASON_INSUFFICIENT_QUORUM,
+    GET_METADATA_EARLY_STOP_REASON_NOT_FOUND, GET_METADATA_EARLY_STOP_REASON_UNSAFE_REQUEST,
+    GET_METADATA_EARLY_STOP_REASON_VALID_QUORUM, GET_METADATA_EARLY_STOP_REASON_VERSION_MATCH_QUORUM,
+    GET_METADATA_EARLY_STOP_REASON_VERSION_NOT_FOUND,
+};
+use crate::disk::error::DiskError;
+use crate::disk::error_reduce::OBJECT_OP_IGNORED_ERRS;
+use crate::set_disk::file_info_is_valid_for_metadata;
+use rustfs_filemeta::FileInfo;
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub(in crate::set_disk) struct MetadataEarlyStopDecision {
+    pub(in crate::set_disk) reason: &'static str,
+}
+
+#[derive(Clone, Debug)]
+pub(in crate::set_disk) struct MetadataQuorumAccumulator {
+    pub(in crate::set_disk) total_disks: usize,
+    pub(in crate::set_disk) default_parity_count: usize,
+    pub(in crate::set_disk) allow_early_stop: bool,
+    pub(in crate::set_disk) valid_responses: usize,
+    pub(in crate::set_disk) not_found_responses: usize,
+    pub(in crate::set_disk) version_not_found_responses: usize,
+    pub(in crate::set_disk) ignored_errors: usize,
+    pub(in crate::set_disk) hard_errors: usize,
+    pub(in crate::set_disk) candidate: Option<FileInfo>,
+    pub(in crate::set_disk) candidate_votes: usize,
+    // Bitset of shard indexes whose metadata matches the candidate. Erasure
+    // layouts are capped at 16 shards, so this stays allocation-free on the
+    // GET metadata hot path.
+    candidate_shard_mask: u16,
+    pub(in crate::set_disk) conflicting_metadata: bool,
+    pub(in crate::set_disk) delete_marker_seen: bool,
+    pub(in crate::set_disk) delete_marker_candidates: Vec<(FileInfo, usize)>,
+    pub(in crate::set_disk) delete_marker_votes: usize,
+    pub(in crate::set_disk) requested_version_id: String,
+    pub(in crate::set_disk) matching_version_votes: usize,
+}
+
+impl MetadataQuorumAccumulator {
+    pub(in crate::set_disk) fn new(total_disks: usize, default_parity_count: usize, allow_early_stop: bool) -> Self {
+        Self {
+            total_disks,
+            default_parity_count,
+            allow_early_stop,
+            valid_responses: 0,
+            not_found_responses: 0,
+            version_not_found_responses: 0,
+            ignored_errors: 0,
+            hard_errors: 0,
+            candidate: None,
+            candidate_votes: 0,
+            candidate_shard_mask: 0,
+            conflicting_metadata: false,
+            delete_marker_seen: false,
+            delete_marker_candidates: Vec::new(),
+            delete_marker_votes: 0,
+            requested_version_id: String::new(),
+            matching_version_votes: 0,
+        }
+    }
+
+    pub(in crate::set_disk) fn with_requested_version_id(mut self, version_id: &str) -> Self {
+        self.requested_version_id = version_id.to_string();
+        self
+    }
+
+    pub(in crate::set_disk) fn observe_file_info(&mut self, file_info: &FileInfo) {
+        self.observe_file_info_with_index(None, file_info);
+    }
+
+    pub(in crate::set_disk) fn observe_file_info_at(&mut self, disk_index: usize, file_info: &FileInfo) {
+        self.observe_file_info_with_index(Some(disk_index), file_info);
+    }
+
+    fn observe_file_info_with_index(&mut self, disk_index: Option<usize>, file_info: &FileInfo) {
+        if !file_info_is_valid_for_metadata(file_info) {
+            self.hard_errors = self.hard_errors.saturating_add(1);
+            return;
+        }
+
+        self.valid_responses = self.valid_responses.saturating_add(1);
+
+        // Track version match for versioned requests
+        if !self.requested_version_id.is_empty()
+            && let Some(ref vid) = file_info.version_id
+            && vid.to_string() == self.requested_version_id
+        {
+            self.matching_version_votes = self.matching_version_votes.saturating_add(1);
+        }
+
+        if file_info.is_canonical_delete_marker() {
+            self.delete_marker_seen = true;
+            if let Some((_, votes)) = self
+                .delete_marker_candidates
+                .iter_mut()
+                .find(|(candidate, _)| metadata_early_stop_candidate_matches(candidate, file_info))
+            {
+                *votes = votes.saturating_add(1);
+            } else {
+                self.delete_marker_candidates.push((file_info.clone(), 1));
+            }
+            self.delete_marker_votes = self
+                .delete_marker_candidates
+                .iter()
+                .map(|(_, votes)| *votes)
+                .max()
+                .unwrap_or_default();
+            self.conflicting_metadata |= self.delete_marker_candidates.len() > 1;
+            return;
+        }
+
+        match &self.candidate {
+            Some(candidate) if metadata_early_stop_candidate_matches(candidate, file_info) => {
+                self.candidate_votes = self.candidate_votes.saturating_add(1);
+                if let Some(disk_index) = disk_index
+                    && let Some(bit) = Self::candidate_shard_bit(candidate, file_info, disk_index)
+                {
+                    self.candidate_shard_mask |= bit;
+                }
+            }
+            Some(_) => {
+                self.conflicting_metadata = true;
+            }
+            None => {
+                self.candidate = Some(file_info.clone());
+                self.candidate_votes = 1;
+                if let Some(disk_index) = disk_index
+                    && let Some(bit) = Self::candidate_shard_bit(file_info, file_info, disk_index)
+                {
+                    self.candidate_shard_mask |= bit;
+                }
+            }
+        }
+    }
+
+    fn candidate_shard_bit(candidate: &FileInfo, file_info: &FileInfo, disk_index: usize) -> Option<u16> {
+        let &erasure_index = candidate.erasure.distribution.get(disk_index)?;
+        if erasure_index == 0 || erasure_index > u16::BITS as usize || file_info.erasure.index != erasure_index {
+            return None;
+        }
+        Some(1u16 << (erasure_index - 1))
+    }
+
+    pub(in crate::set_disk) fn candidate_has_read_reserve(&self) -> bool {
+        self.candidate_read_reserve_target()
+            .is_some_and(|required| self.candidate_shard_mask.count_ones() as usize >= required)
+    }
+
+    pub(in crate::set_disk) fn candidate_read_reserve_target(&self) -> Option<usize> {
+        let candidate = self.candidate.as_ref()?;
+        Some(
+            candidate
+                .erasure
+                .data_blocks
+                .saturating_add(usize::from(candidate.erasure.parity_blocks > 0)),
+        )
+    }
+
+    pub(in crate::set_disk) fn observe_error(&mut self, err: &DiskError) {
+        match err {
+            DiskError::FileNotFound | DiskError::VolumeNotFound => {
+                self.not_found_responses = self.not_found_responses.saturating_add(1);
+            }
+            DiskError::FileVersionNotFound => {
+                self.version_not_found_responses = self.version_not_found_responses.saturating_add(1);
+            }
+            _ if is_metadata_fanout_ignored_error(err) => {
+                self.ignored_errors = self.ignored_errors.saturating_add(1);
+            }
+            _ => {
+                self.hard_errors = self.hard_errors.saturating_add(1);
+            }
+        }
+    }
+
+    pub(in crate::set_disk) fn early_stop_decision(&self) -> Option<MetadataEarlyStopDecision> {
+        if !self.allow_early_stop {
+            return None;
+        }
+        if self.delete_marker_votes >= self.default_write_quorum() {
+            return Some(MetadataEarlyStopDecision {
+                reason: GET_METADATA_EARLY_STOP_REASON_DELETE_MARKER,
+            });
+        }
+        if self.conflicting_metadata
+            || self.delete_marker_seen
+            || self.not_found_responses > 0
+            || self.version_not_found_responses > 0
+            || self.hard_errors > 0
+        {
+            return None;
+        }
+        if self
+            .candidate
+            .as_ref()
+            .and_then(|candidate| self.candidate_latest_quorum(candidate))
+            .is_some_and(|latest_quorum| self.candidate_votes >= latest_quorum)
+        {
+            return Some(MetadataEarlyStopDecision {
+                reason: GET_METADATA_EARLY_STOP_REASON_VALID_QUORUM,
+            });
+        }
+        None
+    }
+
+    /// Check if a versioned request can early-stop because the requested
+    /// version_id has reached quorum across disks.
+    pub(in crate::set_disk) fn version_early_stop_decision(&self) -> Option<MetadataEarlyStopDecision> {
+        if !self.allow_early_stop {
+            return None;
+        }
+        if self.requested_version_id.is_empty() {
+            return None;
+        }
+        if self.conflicting_metadata
+            || self.delete_marker_seen
+            || self.not_found_responses > 0
+            || self.version_not_found_responses > 0
+            || self.hard_errors > 0
+        {
+            return None;
+        }
+        if self.matching_version_votes >= self.read_quorum_for_version() {
+            return Some(MetadataEarlyStopDecision {
+                reason: GET_METADATA_EARLY_STOP_REASON_VERSION_MATCH_QUORUM,
+            });
+        }
+        None
+    }
+
+    pub(in crate::set_disk) fn can_still_reach_early_stop_with_pending(&self, pending: usize) -> bool {
+        if !self.allow_early_stop {
+            return false;
+        }
+        if self.delete_marker_votes.saturating_add(pending) >= self.default_write_quorum() {
+            return true;
+        }
+        if self.conflicting_metadata
+            || self.delete_marker_seen
+            || self.not_found_responses > 0
+            || self.version_not_found_responses > 0
+            || self.hard_errors > 0
+        {
+            return false;
+        }
+        if !self.requested_version_id.is_empty()
+            && self.matching_version_votes.saturating_add(pending) >= self.read_quorum_for_version()
+        {
+            return true;
+        }
+        match &self.candidate {
+            Some(candidate) => self
+                .candidate_latest_quorum(candidate)
+                .is_some_and(|latest_quorum| self.candidate_votes.saturating_add(pending) >= latest_quorum),
+            None => pending >= self.default_write_quorum(),
+        }
+    }
+
+    /// Compute the read quorum threshold for version-aware early-stop.
+    /// Uses `total_disks / 2` (like `missing_response_quorum`) when
+    /// `default_parity_count` is set, otherwise requires all disks.
+    pub(in crate::set_disk) fn read_quorum_for_version(&self) -> usize {
+        self.missing_response_quorum()
+    }
+
+    pub(in crate::set_disk) fn final_miss_reason(&self) -> &'static str {
+        if !self.allow_early_stop {
+            return GET_METADATA_EARLY_STOP_REASON_UNSAFE_REQUEST;
+        }
+        if self.conflicting_metadata {
+            return GET_METADATA_EARLY_STOP_REASON_CONFLICTING_METADATA;
+        }
+        if self.delete_marker_seen {
+            return GET_METADATA_EARLY_STOP_REASON_DELETE_MARKER;
+        }
+        let missing_response_quorum = self.missing_response_quorum();
+        if self.version_not_found_responses >= missing_response_quorum {
+            return GET_METADATA_EARLY_STOP_REASON_VERSION_NOT_FOUND;
+        }
+        if self.not_found_responses >= missing_response_quorum {
+            return GET_METADATA_EARLY_STOP_REASON_NOT_FOUND;
+        }
+        if self.hard_errors > 0 {
+            return GET_METADATA_EARLY_STOP_REASON_ERROR;
+        }
+        if self.ignored_errors > 0 {
+            return GET_METADATA_EARLY_STOP_REASON_INSUFFICIENT_QUORUM;
+        }
+        GET_METADATA_EARLY_STOP_REASON_INSUFFICIENT_QUORUM
+    }
+
+    pub(in crate::set_disk) fn candidate_latest_quorum(&self, candidate: &FileInfo) -> Option<usize> {
+        if self.default_parity_count == 0 {
+            return Some(self.total_disks);
+        }
+        if candidate.is_canonical_delete_marker() || candidate.size == 0 || candidate.erasure.parity_blocks >= self.total_disks {
+            return None;
+        }
+        let data_blocks = candidate.erasure.data_blocks;
+        Some(if data_blocks == candidate.erasure.parity_blocks {
+            data_blocks.saturating_add(1)
+        } else {
+            data_blocks
+        })
+    }
+
+    pub(crate) fn default_write_quorum(&self) -> usize {
+        if self.default_parity_count == 0 || self.default_parity_count >= self.total_disks {
+            return self.total_disks;
+        }
+        let data_blocks = self.total_disks.saturating_sub(self.default_parity_count);
+        if data_blocks == self.default_parity_count {
+            data_blocks.saturating_add(1)
+        } else {
+            data_blocks
+        }
+    }
+
+    pub(in crate::set_disk) fn missing_response_quorum(&self) -> usize {
+        if self.default_parity_count == 0 || self.default_parity_count >= self.total_disks {
+            self.total_disks
+        } else {
+            self.total_disks / 2
+        }
+    }
+}
+
+pub(in crate::set_disk) fn metadata_early_stop_candidate_matches(left: &FileInfo, right: &FileInfo) -> bool {
+    left.volume == right.volume
+        && left.name == right.name
+        && left.version_id == right.version_id
+        && left.is_latest == right.is_latest
+        && left.deleted == right.deleted
+        && left.mark_deleted == right.mark_deleted
+        && left.transition_status == right.transition_status
+        && left.transitioned_objname == right.transitioned_objname
+        && left.transition_tier == right.transition_tier
+        && left.transition_version_id == right.transition_version_id
+        && left.transition_version == right.transition_version
+        && left.transition_version_state == right.transition_version_state
+        && left.expire_restored == right.expire_restored
+        && left.size == right.size
+        && left.mod_time == right.mod_time
+        && left.mode == right.mode
+        && left.written_by_version == right.written_by_version
+        && left.metadata == right.metadata
+        && left.replication_state_internal == right.replication_state_internal
+        && left.parts == right.parts
+        && left.checksum == right.checksum
+        && left.versioned == right.versioned
+        && left.num_versions == right.num_versions
+        && left.successor_mod_time == right.successor_mod_time
+        && left.data_dir == right.data_dir
+        && left.erasure.algorithm == right.erasure.algorithm
+        && left.erasure.data_blocks == right.erasure.data_blocks
+        && left.erasure.parity_blocks == right.erasure.parity_blocks
+        && left.erasure.block_size == right.erasure.block_size
+        && left.erasure.distribution == right.erasure.distribution
+}
+
+pub(in crate::set_disk) fn is_metadata_fanout_ignored_error(err: &DiskError) -> bool {
+    OBJECT_OP_IGNORED_ERRS.iter().any(|ignored| ignored == err)
+}

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

@@ -18,3 +18,4 @@
 //! duplicating read/write/erasure logic.
 
 pub(crate) mod io_primitives;
+mod metadata_quorum;

crates/ecstore/src/set_disk/mod.rs

@@ -876,7 +876,7 @@ pub use ops::multipart::{MultipartCommitBarrier, MultipartCommitPause};
 pub(crate) use ops::object::DeleteObjectCommitBarrier;
 #[cfg(any(test, feature = "test-util"))]
 pub(crate) use ops::object::TransitionCleanupStoreBarrier as SetDiskTransitionCleanupStoreBarrier;
-#[cfg(test)]
+#[cfg(all(test, feature = "test-util"))]
 pub(crate) use ops::object::TransitionUploadedCommitBarrier as SetDiskTransitionUploadedCommitBarrier;
 pub(crate) use ops::object::body_cache_plaintext_len;
 #[cfg(all(test, feature = "test-util"))]

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

@@ -2490,9 +2490,9 @@ impl crate::storage_api_contracts::heal::HealOperations for SetDisks {
             return Ok((result, err.map(|e| e.into())));
         }
 
-        let disks = self.disks.read().await;
+        // The inner heal and missing-object report read the registry again;
-
+        // release this snapshot guard before a topology writer can queue between reads.
-        let disks = disks.clone();
+        let disks = self.get_disks_internal().await;
         let (_, errs) = Self::read_all_fileinfo(&disks, "", bucket, object, version_id, false, false, false)
             .await
             .map_err(|e| to_object_err(e.into(), vec![bucket, object]))?;
@@ -3419,6 +3419,366 @@ mod heal_result_report_tests {
         assert_eq!(unformatted, DiskError::UnformattedDisk);
     }
 
+    #[derive(Clone, Copy)]
+    enum InventoryWriterHealCase {
+        Existing,
+        Missing,
+        MissingVersion,
+    }
+
+    async fn assert_heal_object_inventory_writer(case: InventoryWriterHealCase) {
+        use crate::set_disk::core::io_primitives::disk_call_counters;
+        use std::time::Duration;
+        use tokio::io::AsyncReadExt;
+
+        let (_temp_dirs, disks, set) = hermetic_set_disks_isolated(4).await;
+        let bucket = "heal-inventory-writer-bucket";
+        let object = match case {
+            InventoryWriterHealCase::Existing => "heal-inventory-writer-existing",
+            InventoryWriterHealCase::Missing => "heal-inventory-writer-missing",
+            InventoryWriterHealCase::MissingVersion => "heal-inventory-writer-missing-version",
+        };
+        set.make_bucket(
+            bucket,
+            &MakeBucketOptions {
+                versioning_enabled: true,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("heal fixture bucket should be created");
+        let body = vec![0x67; 64 * 1024];
+        let stored_version = Uuid::new_v4();
+        let stored_version_string = stored_version.to_string();
+        let published = if matches!(case, InventoryWriterHealCase::Missing) {
+            None
+        } else {
+            let mut reader = PutObjReader::from_vec(body.clone());
+            let info = set
+                .put_object(
+                    bucket,
+                    object,
+                    &mut reader,
+                    &ObjectOptions {
+                        no_lock: true,
+                        versioned: true,
+                        version_id: Some(stored_version_string.clone()),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("full-fanout PUT should seed the heal fixture");
+            for disk in &disks {
+                let metadata = disk
+                    .read_version("", bucket, object, &stored_version_string, &ReadOptions::default())
+                    .await
+                    .expect("the seeded version must be present on every disk");
+                assert_eq!(metadata.version_id, Some(stored_version));
+                assert_eq!(metadata.size, i64::try_from(body.len()).expect("fixture size should fit i64"));
+            }
+            Some(info)
+        };
+        let requested_version = match case {
+            InventoryWriterHealCase::Existing => stored_version_string.clone(),
+            InventoryWriterHealCase::Missing => String::new(),
+            InventoryWriterHealCase::MissingVersion => Uuid::new_v4().to_string(),
+        };
+        let opts = HealOpts {
+            no_lock: true,
+            ..Default::default()
+        };
+        let calls = disk_call_counters::observe(object);
+        let read_gate = set.disks.read().await;
+        // UFCS selects the trait's outer precheck, not the same-named inherent heal.
+        let heal = <SetDisks as crate::storage_api_contracts::heal::HealOperations>::heal_object(
+            set.as_ref(),
+            bucket,
+            object,
+            &requested_version,
+            &opts,
+        );
+        tokio::pin!(heal);
+        assert!(matches!(
+            futures::poll!(tokio::task::unconstrained(heal.as_mut())),
+            std::task::Poll::Pending
+        ));
+        // These tests use the current-thread runtime: full-wait metadata tasks
+        // have been spawned, but cannot run during the single unconstrained poll.
+        assert_eq!(calls.total(disk_call_counters::KIND_READ_VERSION), 0);
+        let writer = set.disks.write();
+        tokio::pin!(writer);
+        assert!(matches!(
+            futures::poll!(tokio::task::unconstrained(writer.as_mut())),
+            std::task::Poll::Pending
+        ));
+        assert!(set.disks.try_read().is_err(), "the writer must already block new inventory readers");
+        tokio::time::timeout(Duration::from_secs(5), async {
+            while calls.total(disk_call_counters::KIND_READ_VERSION) < 4 {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("the suspended trait heal must have started the real metadata fanout");
+        for disk_index in 0..4 {
+            assert_eq!(calls.for_disk(disk_call_counters::KIND_READ_VERSION, disk_index), 1);
+        }
+        drop(read_gate);
+
+        let (_, outcome) =
+            tokio::time::timeout(Duration::from_secs(5), async { tokio::join!(async { drop(writer.await) }, heal) })
+                .await
+                .expect("trait heal must not deadlock its nested inventory read with the queued writer");
+        let (result, error) = outcome.expect("heal should report the object's outcome");
+        match case {
+            InventoryWriterHealCase::Existing => assert!(error.is_none(), "existing object heal failed: {error:?}"),
+            InventoryWriterHealCase::Missing => assert!(matches!(error, Some(Error::FileNotFound))),
+            InventoryWriterHealCase::MissingVersion => assert!(matches!(error, Some(Error::FileVersionNotFound))),
+        }
+        assert_eq!(result.bucket, bucket);
+        assert_eq!(result.object, object);
+        assert_eq!(result.version_id, requested_version);
+        assert_eq!(result.disk_count, 4);
+        assert_eq!(result.before.drives.len(), 4);
+        assert_eq!(result.after.drives.len(), 4);
+        for disk_index in 0..4 {
+            let endpoint = set.set_endpoints[disk_index].to_string();
+            assert_eq!(result.before.drives[disk_index].endpoint, endpoint);
+            assert_eq!(result.after.drives[disk_index].endpoint, endpoint);
+        }
+        if let Some(published) = published {
+            tokio::time::timeout(Duration::from_secs(10), async {
+                let mut reader = set
+                    .get_object_reader(
+                        bucket,
+                        object,
+                        None,
+                        Default::default(),
+                        &ObjectOptions {
+                            versioned: true,
+                            version_id: Some(stored_version_string),
+                            ..Default::default()
+                        },
+                    )
+                    .await
+                    .expect("the stored version must remain readable after heal");
+                assert_eq!(reader.object_info.etag, published.etag);
+                assert_eq!(reader.object_info.version_id, Some(stored_version));
+                let mut observed_body = Vec::new();
+                reader
+                    .stream
+                    .read_to_end(&mut observed_body)
+                    .await
+                    .expect("stored body should stream");
+                assert_eq!(observed_body, body);
+            })
+            .await
+            .expect("GET must finish after the inventory writer and heal");
+        }
+    }
+
+    #[tokio::test]
+    async fn heal_object_inventory_writer_existing() {
+        assert_heal_object_inventory_writer(InventoryWriterHealCase::Existing).await;
+    }
+
+    #[tokio::test]
+    async fn heal_object_inventory_writer_missing() {
+        assert_heal_object_inventory_writer(InventoryWriterHealCase::Missing).await;
+    }
+
+    #[tokio::test]
+    async fn heal_object_inventory_writer_missing_version() {
+        assert_heal_object_inventory_writer(InventoryWriterHealCase::MissingVersion).await;
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn heal_object_with_queued_disk_renewal() {
+        use crate::layout::endpoints::SetupType;
+        use crate::runtime::instance::InstanceContext;
+        use crate::set_disk::core::io_primitives::disk_call_counters;
+        use std::collections::HashMap;
+        use std::future::Future;
+        use std::task::Poll;
+        use std::time::Duration;
+        use tokio::io::AsyncReadExt;
+
+        // renew_disk still registers local disks on the ambient context. Match
+        // the default serial group used by its other setup/registry fixtures,
+        // and restore only this temporary endpoint, including on a failed join.
+        struct RenewDiskTestState {
+            ctx: Arc<InstanceContext>,
+            was_dist_erasure: bool,
+            map: Arc<RwLock<HashMap<String, Option<DiskStore>>>>,
+            endpoint: String,
+            previous_disk: Option<Option<DiskStore>>,
+        }
+
+        impl Drop for RenewDiskTestState {
+            fn drop(&mut self) {
+                let ctx = self.ctx.clone();
+                let was_dist_erasure = self.was_dist_erasure;
+                let map = self.map.clone();
+                let endpoint = self.endpoint.clone();
+                let previous_disk = self.previous_disk.take();
+                let handle = tokio::runtime::Handle::current();
+                std::thread::spawn(move || {
+                    handle.block_on(async move {
+                        let mut map = map.write().await;
+                        match previous_disk {
+                            Some(disk) => {
+                                map.insert(endpoint, disk);
+                            }
+                            None => {
+                                map.remove(&endpoint);
+                            }
+                        }
+                        drop(map);
+                        if was_dist_erasure {
+                            ctx.update_erasure_type(SetupType::DistErasure).await;
+                        }
+                    });
+                })
+                .join()
+                .expect("renew fixture state restoration should finish");
+            }
+        }
+
+        let (_temp_dirs, disks, set) = hermetic_set_disks_isolated(4).await;
+        let endpoint = set.set_endpoints[0].clone();
+        let ctx = crate::runtime::global::current_ctx();
+        let map = ctx.local_disk_map();
+        let restore = RenewDiskTestState {
+            ctx: ctx.clone(),
+            was_dist_erasure: ctx.is_dist_erasure().await,
+            map: map.clone(),
+            endpoint: endpoint.to_string(),
+            previous_disk: map.read().await.get(&endpoint.to_string()).cloned(),
+        };
+        // Only distributed erasure needs an override to avoid the ambient slot array.
+        if restore.was_dist_erasure {
+            ctx.update_erasure_type(SetupType::Erasure).await;
+        }
+
+        let bucket = "heal-disk-renewal-bucket";
+        let object = "heal-disk-renewal-object";
+        set.make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("renew fixture bucket should be created");
+        let body = vec![0x73; 64 * 1024];
+        let mut reader = PutObjReader::from_vec(body.clone());
+        let published = set
+            .put_object(
+                bucket,
+                object,
+                &mut reader,
+                &ObjectOptions {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("full-fanout PUT should seed the renewal fixture");
+        for disk in &disks {
+            let metadata = disk
+                .read_version("", bucket, object, "", &ReadOptions::default())
+                .await
+                .expect("the seeded object must be present on every disk");
+            assert_eq!(metadata.size, i64::try_from(body.len()).expect("fixture size should fit i64"));
+        }
+
+        let opts = HealOpts {
+            no_lock: true,
+            ..Default::default()
+        };
+        let calls = disk_call_counters::observe(object);
+        let read_gate = set.disks.read().await;
+        let heal = <SetDisks as crate::storage_api_contracts::heal::HealOperations>::heal_object(
+            set.as_ref(),
+            bucket,
+            object,
+            "",
+            &opts,
+        );
+        tokio::pin!(heal);
+        assert!(matches!(futures::poll!(tokio::task::unconstrained(heal.as_mut())), Poll::Pending));
+        assert_eq!(calls.total(disk_call_counters::KIND_READ_VERSION), 0);
+
+        let renew = set.renew_disk(&endpoint);
+        tokio::pin!(renew);
+        tokio::time::timeout(
+            Duration::from_secs(5),
+            futures::future::poll_fn(|cx| {
+                assert!(
+                    std::pin::pin!(tokio::task::unconstrained(renew.as_mut()))
+                        .poll(cx)
+                        .is_pending(),
+                    "renewal must reach its inventory write before returning"
+                );
+                if set.disks.try_read().is_err() {
+                    Poll::Ready(())
+                } else {
+                    Poll::Pending
+                }
+            }),
+        )
+        .await
+        .expect("real renewal must queue its topology writer behind the read gate");
+        let registered = map
+            .read()
+            .await
+            .get(&endpoint.to_string())
+            .cloned()
+            .flatten()
+            .expect("renewal must register the connected disk before its inventory write");
+        assert!(!Arc::ptr_eq(&registered, &disks[0]), "renewal must construct a new disk handle");
+        tokio::time::timeout(Duration::from_secs(5), async {
+            while calls.total(disk_call_counters::KIND_READ_VERSION) < 4 {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("the suspended trait heal must have started the real metadata fanout");
+        for disk_index in 0..4 {
+            assert_eq!(calls.for_disk(disk_call_counters::KIND_READ_VERSION, disk_index), 1);
+        }
+        drop(read_gate);
+
+        let (_, outcome) = tokio::time::timeout(Duration::from_secs(5), async { tokio::join!(renew, heal) })
+            .await
+            .expect("trait heal and real disk renewal must finish without a nested inventory read deadlock");
+        let (report, error) = outcome.expect("heal should report the existing object");
+        assert!(error.is_none(), "existing object heal failed after renewal: {error:?}");
+        assert_eq!(report.bucket, bucket);
+        assert_eq!(report.object, object);
+        assert_eq!(report.disk_count, 4);
+        let renewed = set.get_disks_internal().await[0]
+            .clone()
+            .expect("the renewed slot must remain online");
+        assert!(Arc::ptr_eq(&renewed, &registered), "the set must publish the newly connected handle");
+        assert_eq!(renewed.endpoint(), endpoint);
+        let format = load_format_erasure(&renewed, false)
+            .await
+            .expect("renewed disk format should remain readable");
+        assert_eq!(format.erasure.this, set.format.erasure.sets[0][0]);
+        tokio::time::timeout(Duration::from_secs(10), async {
+            let mut reader = set
+                .get_object_reader(bucket, object, None, Default::default(), &ObjectOptions::default())
+                .await
+                .expect("the object must remain readable after renewal and heal");
+            assert_eq!(reader.object_info.etag, published.etag);
+            let mut observed_body = Vec::new();
+            reader
+                .stream
+                .read_to_end(&mut observed_body)
+                .await
+                .expect("stored body should stream");
+            assert_eq!(observed_body, body);
+        })
+        .await
+        .expect("GET must finish after renewal and heal");
+    }
+
     // Regression for #955: an offline disk must contribute exactly one drive
     // record. Before the fix the offline branch fell through and pushed a second
     // (Corrupt) record for the same disk, so `before/after.drives` grew to

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

@@ -2452,10 +2452,9 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
         let write_quorum = fi.write_quorum(self.default_write_quorum());
         let read_quorum = fi.read_quorum(self.default_read_quorum());
 
-        let disks = self.disks.read().await;
+        // Release the registry guard before recovery and cleanup read it again:
-
+        // a queued topology writer would otherwise deadlock those nested reads.
-        let disks = disks.clone();
+        let disks = self.get_disks_internal().await;
-        // let disks = Self::shuffle_disks(&disks, &fi.erasure.distribution);
 
         let part_path = format!("{}/{}/", upload_id_path, fi.data_dir.unwrap_or(Uuid::nil()));
         self.recover_part_transactions(&part_path, read_quorum, write_quorum)
@@ -4051,6 +4050,7 @@ mod tests {
             let _ = drain_global_dirty_scopes();
 
             let rename_barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier::PHASE_RENAME);
+            let rename_tasks = rename_fanout_barrier::observe_tasks(object);
             let complete_store = Arc::clone(&set_disks);
             let mut complete = tokio::spawn(async move {
                 let mut opts = ObjectOptions::default();
@@ -4062,16 +4062,6 @@ mod tests {
             tokio::time::timeout(Duration::from_secs(30), rename_barrier.wait_until_paused())
                 .await
                 .expect("multipart completion should pause one tail disk during rename");
-            assert!(
-                tokio::time::timeout(Duration::from_millis(100), &mut complete).await.is_err(),
-                "multipart completion must not publish success while a tail rename is still paused"
-            );
-
-            let initial = drain_global_dirty_scopes().into_iter().collect::<HashSet<_>>();
-            assert!(
-                initial.is_empty(),
-                "capacity must not be marked as committed before the full multipart rename finishes"
-            );
 
             let abort_store = Arc::clone(&set_disks);
             let abort = tokio::spawn(async move {
@@ -4080,21 +4070,46 @@ mod tests {
                     .await
             });
             signaling.wait_for_attempts(2).await;
-            assert!(!abort.is_finished(), "the in-flight completion must retain the multipart upload guard");
 
-            let retained_staging = futures::future::join_all(
+            // A paused rename does not establish that the other disks reached quorum.
-                disk_stores
+            let retained_staging = tokio::time::timeout(Duration::from_secs(30), async {
-                    .iter()
+                loop {
-                    .map(|disk| disk.read_all(RUSTFS_META_MULTIPART_BUCKET, &staged_part)),
+                    let mut retained = 0;
-            )
+                    for result in futures::future::join_all(
+                        disk_stores
+                            .iter()
+                            .map(|disk| disk.read_all(RUSTFS_META_MULTIPART_BUCKET, &staged_part)),
+                    )
+                    .await
+                    {
+                        match result {
+                            Ok(_) => retained += 1,
+                            Err(DiskError::FileNotFound) => {}
+                            Err(error) => panic!("staged rename source lookup failed: {error}"),
+                        }
+                    }
+                    if retained <= 1 && rename_tasks.running() == 1 {
+                        break retained;
+                    }
+                    tokio::time::sleep(Duration::from_millis(10)).await;
+                }
+            })
             .await
-            .into_iter()
+            .expect("unpaused multipart renames should finish before the tail is released");
-            .filter(|result| result.is_ok())
-            .count();
             assert_eq!(
                 retained_staging, 1,
                 "only the paused tail disk should still retain the multipart rename source"
             );
+            assert!(
+                tokio::time::timeout(Duration::from_millis(100), &mut complete).await.is_err(),
+                "multipart completion must not publish success while a tail rename is still paused"
+            );
+            let initial = drain_global_dirty_scopes().into_iter().collect::<HashSet<_>>();
+            assert!(
+                initial.is_empty(),
+                "capacity must not be marked as committed before the full multipart rename finishes"
+            );
+            assert!(!abort.is_finished(), "the in-flight completion must retain the multipart upload guard");
 
             signaling.set_target(rustfs_lock::ObjectKey::new(bucket, object));
             let object_attempt = signaling.attempts.load(Ordering::Acquire) + 1;
@@ -6743,6 +6758,87 @@ mod tests {
         .await;
     }
 
+    #[tokio::test(flavor = "multi_thread")]
+    #[serial]
+    async fn complete_multipart_releases_disk_snapshot_before_cleanup() {
+        let (temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "multipart-topology-lock-bucket";
+        let object = "object";
+        let body = vec![0x65; 4096];
+        make_bucket_on_all(&disk_stores, bucket).await;
+        let (upload_id, parts) =
+            stage_upload_with_create_opts(&set_disks, bucket, object, &body, &ObjectOptions::default()).await;
+        let upload_id_path = SetDisks::get_upload_id_dir(bucket, object, &upload_id);
+        for dir in &temp_dirs {
+            assert!(
+                dir.path().join(RUSTFS_META_MULTIPART_BUCKET).join(&upload_id_path).exists(),
+                "the test must create real upload staging on every disk"
+            );
+        }
+        let barrier = MultipartCommitBarrier::install(bucket, object, MultipartCommitPause::AfterObjectPublication);
+        let complete_store = set_disks.clone();
+        let complete_upload_id = upload_id.clone();
+        let complete = tokio::spawn(async move {
+            complete_store
+                .complete_multipart_upload(bucket, object, &complete_upload_id, parts, &ObjectOptions::default())
+                .await
+        });
+        barrier.wait_until_paused().await;
+
+        // Hold a separate read gate so the real writer queues even when completion
+        // correctly releases its snapshot guard. Polling Pending proves admission
+        // to Tokio's write-preferring queue before the cleanup attempts another read.
+        let read_gate = set_disks.disks.read().await;
+        let writer = set_disks.disks.write();
+        tokio::pin!(writer);
+        assert!(matches!(
+            futures::poll!(tokio::task::unconstrained(writer.as_mut())),
+            std::task::Poll::Pending
+        ));
+        assert!(
+            set_disks.disks.try_read().is_err(),
+            "the pending writer must already block new readers before the cleanup resumes"
+        );
+        drop(read_gate);
+        barrier.release();
+
+        let writer_guard = tokio::time::timeout(Duration::from_secs(5), writer)
+            .await
+            .expect("a queued topology writer must not deadlock with multipart cleanup's disk snapshot");
+        // A reconnect can publish the same handles; this test isolates admission
+        // order without changing the disks that contain the committed object.
+        drop(writer_guard);
+        tokio::time::timeout(Duration::from_secs(10), complete)
+            .await
+            .expect("multipart cleanup must finish after the topology writer releases")
+            .expect("completion task should not panic")
+            .expect("completion should preserve the successful object commit");
+
+        let mut reader = tokio::time::timeout(
+            Duration::from_secs(10),
+            set_disks.get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default()),
+        )
+        .await
+        .expect("GET should finish after completion")
+        .expect("the completed object should remain readable");
+        let mut observed_body = Vec::new();
+        tokio::time::timeout(Duration::from_secs(10), reader.stream.read_to_end(&mut observed_body))
+            .await
+            .expect("the completed object body should finish streaming")
+            .expect("the completed object body should be readable");
+        assert_eq!(observed_body, body);
+        assert!(matches!(
+            set_disks.check_upload_id_exists(bucket, object, &upload_id, false).await,
+            Err(StorageError::InvalidUploadID(..))
+        ));
+        for dir in &temp_dirs {
+            assert!(
+                !dir.path().join(RUSTFS_META_MULTIPART_BUCKET).join(&upload_id_path).exists(),
+                "successful completion must remove its upload staging from every disk"
+            );
+        }
+    }
+
     #[tokio::test(flavor = "multi_thread")]
     #[serial]
     async fn complete_releases_object_lock_before_cleanup_and_keeps_upload_lock() {

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

@@ -299,11 +299,11 @@ use crate::error::is_err_invalid_upload_id;
 use crate::object_api::{GetObjectBodySource, get_object_body_cache_hook_suppressed};
 use crate::object_api::{
     NamespaceLockFence, ReplicationStatusWritebackCondition, ReplicationStatusWritebackMode,
-    SCANNER_PUBLICATION_LEASE_FENCE_METADATA_KEY,
+    SCANNER_PUBLICATION_LEASE_FENCE_METADATA_KEY, WriteCompletion,
 };
 use crate::services::notification_sys::RemoteVersionStateFleetProofToken;
 use crate::services::tier::tier::{TierConfigMgr, TierDestinationId, TierOperationLease, tier_destination_id_from_metadata};
-use crate::set_disk::core::io_primitives::{RenameTailCleanup, finish_rename_tail_heal};
+use crate::set_disk::core::io_primitives::{RenameRollbackReceipt, RenameTailCleanup, finish_rename_tail_heal};
 #[cfg(test)]
 use crate::storage_api_contracts::namespace::NamespaceLocking;
 #[cfg(test)]
@@ -3548,6 +3548,7 @@ impl SetDisks {
                 (None, None, None)
             };
         let mut tmp_cleanup_owned = false;
+        let rollback_receipt = RenameRollbackReceipt::default();
         let operation = async {
             let erasure = Arc::new(erasure_from_file_info(&fi, false)?);
 
@@ -4256,6 +4257,7 @@ impl SetDisks {
             let commit_bucket = bucket.to_owned();
             let commit_object = object.to_owned();
             let commit_tmp_dir = tmp_dir.clone();
+            let commit_rollback_receipt = rollback_receipt.clone();
             let commit_object_lock_guard = object_lock_guard.take();
             let commit_decommission_object_lock_guard = decommission_object_lock_guard.take();
             let commit_publication_guard = publication_commit_guard.take();
@@ -4266,13 +4268,17 @@ impl SetDisks {
             // complete rename fan-out drains. Keep this path synchronous so
             // its terminal state is known before the coordinator releases
             // remote leases.
-            let commit_allows_early_ack = !(opts.data_movement && opts.has_decommission_capacity_reservation())
+            let commit_owns_namespace_guard = commit_object_lock_guard.is_some()
-                && (commit_object_lock_guard.is_some()
+                || commit_decommission_object_lock_guard.is_some()
-                    || commit_decommission_object_lock_guard.is_some()
+                || commit_publication_guard.is_some();
-                    || commit_publication_guard.is_some())
+            let commit_allows_early_ack = opts.write_completion == WriteCompletion::Quorum
+                && !(opts.data_movement && opts.has_decommission_capacity_reservation())
+                && commit_owns_namespace_guard
                 && commit_scanner_publication_scope.is_none();
+            // Full-tail callers also transfer owned guards to the coordinator:
+            // cancelling their ACK waiter must not cancel an in-flight rename.
             let detach_commit_owner = commit_scanner_publication_scope.is_some()
-                || commit_allows_early_ack
+                || commit_owns_namespace_guard
                 || commit_bucket_lifecycle_guard.is_some()
                 || quota_mutation_fence;
             let commit_write_path_label = write_path.metric_label();
@@ -4452,7 +4458,11 @@ impl SetDisks {
                         write_quorum,
                         commit_scanner_publication_lease_tokens.as_ref(),
                     )
-                    .with_publication_scope(commit_scanner_publication_scope.clone()),
+                    .with_publication_scope(commit_scanner_publication_scope.clone())
+                    .with_rollback_receipt(commit_rollback_receipt.clone())
+                    .with_namespace_commit_guard(
+                        (!is_meta_bucketname(&commit_bucket)).then(|| commit_set.ctx.begin_namespace_commit()),
+                    ),
                 )
                 .await;
                 if let Some(scope) = commit_scanner_publication_scope.as_ref() {
@@ -4585,6 +4595,11 @@ impl SetDisks {
                 let rename_commit = match rename_result {
                     Ok(commit) => commit,
                     Err(err) => {
+                        if commit_rollback_receipt.is_incomplete() {
+                            // Incomplete undo retains the staging source and
+                            // rollback backup for recovery; cleanup is unsafe.
+                            return Err(err.into());
+                        }
                         if let Err(cleanup_err) = commit_set.delete_all(RUSTFS_META_TMP_BUCKET, &commit_tmp_dir).await {
                             warn!(tmp_dir = %commit_tmp_dir, error = ?cleanup_err, "failed to cleanup put_object temporary data");
                         } else if issue3031_diag_enabled() {
@@ -4617,9 +4632,8 @@ impl SetDisks {
                     request.object_version_id = committed_version_id
                         .or_else(|| commit_version_suspended.then(Uuid::nil))
                         .map(|version_id| version_id.to_string());
-                    tokio::spawn(async move {
+                    let heal_set = commit_set.clone();
-                        let _ = rustfs_heal_contracts::heal_channel::send_heal_request(request).await;
+                    tokio::spawn(async move { heal_set.submit_rename_tail_heal(request).await });
-                    });
                 }
 
                 let rename_stage_elapsed = rename_stage_start.elapsed();
@@ -4885,7 +4899,7 @@ impl SetDisks {
                     );
                 }
             });
-        } else {
+        } else if !rollback_receipt.is_incomplete() {
             // Failure path (quorum loss / rollback): keep the cleanup inline so
             // a failed PUT never returns while its tmp shards are still on disk
             // (state-residue hardening tracked by backlog#864 / backlog#898).
@@ -17494,27 +17508,69 @@ mod put_object_tmp_cleanup_tests {
     }
 
     #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
     async fn put_object_failure_cleans_tmp_workspace_inline() {
-        let (temp_dirs, _disk_stores, set_disks) = hermetic_set_disks(4).await;
+        temp_env::async_with_vars([(ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true"))], async {
+            for write_completion in [WriteCompletion::Quorum, WriteCompletion::TailDrained] {
+                let (temp_dirs, _disk_stores, set_disks) = hermetic_set_disks(4).await;
+                let bucket = "tmp-clean-missing-bucket";
+                let object = "orphan-object";
+                let barrier = PutObjectCommitBarrier::install(bucket, object, PutObjectCommitPause::BeforeNamespace);
+                let writer = Arc::clone(&set_disks);
+                let put = tokio::spawn(async move {
+                    let mut reader = PutObjReader::from_vec(vec![9u8; TEST_OBJECT_SIZE]);
+                    writer
+                        .put_object(
+                            bucket,
+                            object,
+                            &mut reader,
+                            &ObjectOptions {
+                                write_completion,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                });
+                tokio::time::timeout(Duration::from_secs(30), barrier.wait_until_paused())
+                    .await
+                    .expect("missing-bucket PUT must stage before rename");
+                let staged = non_trash_tmp_entries(&temp_dirs).await;
+                assert_eq!(staged.len(), 4, "every disk must have a staged workspace before rejection");
+                for workspace in staged {
+                    let mut entries = tokio::fs::read_dir(&workspace)
+                        .await
+                        .expect("staged workspace should be readable");
+                    let mut shards = 0;
+                    while let Some(entry) = entries.next_entry().await.expect("staged data directory should be readable") {
+                        if entry.file_type().await.expect("staged entry type").is_dir() {
+                            let part = tokio::fs::metadata(entry.path().join("part.1"))
+                                .await
+                                .expect("staging must contain an actual erasure shard");
+                            assert!(part.len() > 0, "the shard must be written before the missing-bucket failure");
+                            shards += 1;
+                        }
+                    }
+                    assert_eq!(shards, 1);
+                }
+                assert!(temp_dirs.iter().all(|dir| !dir.path().join(bucket).exists()));
+                barrier.release();
+                let err = tokio::time::timeout(Duration::from_secs(30), put)
+                    .await
+                    .expect("missing-bucket PUT must finish")
+                    .expect("PUT task should join")
+                    .expect_err("put_object into a missing bucket volume must fail");
+                assert!(matches!(err, StorageError::VolumeNotFound), "original disk error expected: {err}");
 
-        // The bucket volume is never created, so the shards are written into
+                // No polling: known pre-publication rejection must clean staging
-        // the tmp workspace and the commit fails at rename_data with a quorum
+                // inline, before PUT returns (backlog#864 / backlog#898).
-        // error — exercising the failure-path cleanup.
+                let leftovers = non_trash_tmp_entries(&temp_dirs).await;
-        let mut reader = PutObjReader::from_vec(vec![9u8; TEST_OBJECT_SIZE]);
+                assert!(
-        let err = set_disks
+                    leftovers.is_empty(),
-            .put_object("tmp-clean-missing-bucket", "orphan-object", &mut reader, &ObjectOptions::default())
+                    "failed PUT must not leave tmp shards behind, leftovers: {leftovers:?}, err: {err}"
-            .await
+                );
-            .expect_err("put_object into a missing bucket volume must fail");
+            }
-
+        })
-        // No polling: the failure path must clean the tmp workspace inline,
+        .await;
-        // before put_object returns (backlog#864 / backlog#898 hardening).
-        let leftovers = non_trash_tmp_entries(&temp_dirs).await;
-        assert!(
-            leftovers.is_empty(),
-            "failed PUT must not leave tmp shards behind, leftovers: {leftovers:?}, err: {err}"
-        );
-
-        drop(temp_dirs);
     }
 
     #[tokio::test]
@@ -18157,6 +18213,354 @@ mod put_object_tmp_cleanup_tests {
         .await;
     }
 
+    async fn make_completion_test_bucket(disks: &[DiskStore], bucket: &str) {
+        for disk in disks {
+            disk.make_volume(bucket)
+                .await
+                .expect("completion test bucket should be created");
+        }
+    }
+
+    /// Observe the actual metadata quorum while the remaining rename is parked.
+    /// A completed task count alone can race tasks that have not started yet.
+    async fn wait_for_paused_tail_metadata_quorum(disks: &[DiskStore], bucket: &str, object: &str) {
+        tokio::time::timeout(Duration::from_secs(30), async {
+            loop {
+                let mut committed = 0;
+                for disk in disks {
+                    match disk.read_version("", bucket, object, "", &ReadOptions::default()).await {
+                        Ok(_) => committed += 1,
+                        Err(DiskError::FileNotFound | DiskError::FileVersionNotFound) => {}
+                        Err(err) => panic!("unexpected metadata error while observing {bucket}/{object}: {err}"),
+                    }
+                }
+                if committed == 3 {
+                    break;
+                }
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("three disks must publish metadata while the fourth rename remains paused");
+    }
+
+    #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
+    async fn tail_drained_put_waits_for_tail_and_allows_immediate_cas() {
+        temp_env::async_with_vars([(ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true"))], async {
+            for size in [4096, 1024 * 1024] {
+                let (_dirs, disks, set) = hermetic_set_disks(4).await;
+                let bucket = "put-full-tail-cas";
+                let object = "full-tail-cas-object";
+                make_completion_test_bucket(&disks, bucket).await;
+                let tasks = rename_fanout_barrier::observe_tasks(object);
+                let barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier::PHASE_RENAME);
+                let writer = Arc::clone(&set);
+                let put = tokio::spawn(async move {
+                    let mut reader = PutObjReader::from_vec(vec![b'1'; size]);
+                    writer
+                        .put_object(
+                            bucket,
+                            object,
+                            &mut reader,
+                            &ObjectOptions {
+                                write_completion: WriteCompletion::TailDrained,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                });
+                tokio::time::timeout(Duration::from_secs(30), barrier.wait_until_paused())
+                    .await
+                    .expect("full-tail PUT must reach the rename barrier");
+                wait_for_paused_tail_metadata_quorum(&disks, bucket, object).await;
+                assert!(!put.is_finished(), "full-tail PUT must remain pending after metadata quorum");
+                let mut lock_probe = Box::pin(set.acquire_write_lock_diag("full_tail_probe", bucket, object));
+                assert!(
+                    futures::poll!(lock_probe.as_mut()).is_pending(),
+                    "the owned namespace guard must remain held"
+                );
+                barrier.release();
+                let written = tokio::time::timeout(Duration::from_secs(30), put)
+                    .await
+                    .expect("full-tail PUT should finish after release")
+                    .expect("full-tail PUT task should join")
+                    .expect("full-tail PUT must commit");
+                assert_eq!(tasks.running(), 0, "full-tail response must follow every rename task");
+                drop(
+                    tokio::time::timeout(Duration::from_secs(5), lock_probe)
+                        .await
+                        .expect("same-key lock should be available on return")
+                        .expect("same-key lock probe should succeed"),
+                );
+                for disk in &disks {
+                    disk.read_version("", bucket, object, "", &ReadOptions::default())
+                        .await
+                        .expect("successful full-tail PUT must publish on every healthy disk");
+                }
+                drop(barrier);
+                let mut replacement = PutObjReader::from_vec(b"cas successor".to_vec());
+                set.put_object(
+                    bucket,
+                    object,
+                    &mut replacement,
+                    &ObjectOptions {
+                        write_completion: WriteCompletion::TailDrained,
+                        http_preconditions: Some(HTTPPreconditions {
+                            if_match: written.etag,
+                            ..Default::default()
+                        }),
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("immediate same-key CAS must acquire the namespace guard");
+                let mut read = set
+                    .get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default())
+                    .await
+                    .expect("CAS successor must be immediately readable");
+                let mut body = Vec::new();
+                read.stream.read_to_end(&mut body).await.expect("successor body must drain");
+                assert_eq!(body, b"cas successor");
+            }
+        })
+        .await;
+    }
+
+    #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
+    async fn tail_drained_put_preserves_quorum_success_and_heals_failed_tail() {
+        let (_dirs, disks, set) = hermetic_set_disks(4).await;
+        let bucket = "put-full-tail-heal";
+        let object = "full-tail-heal-object";
+        make_completion_test_bucket(&disks, bucket).await;
+        let mut heals = set.capture_test_rename_tail_heals();
+        let tasks = rename_fanout_barrier::observe_tasks(object);
+        let barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier::PHASE_RENAME);
+        let _fault = rename_fault_injection::fail_rename_on(object, &[0]);
+        let writer = Arc::clone(&set);
+        let put = tokio::spawn(async move {
+            let mut reader = PutObjReader::from_vec(vec![b'1'; TEST_OBJECT_SIZE]);
+            writer
+                .put_object(
+                    bucket,
+                    object,
+                    &mut reader,
+                    &ObjectOptions {
+                        write_completion: WriteCompletion::TailDrained,
+                        ..Default::default()
+                    },
+                )
+                .await
+        });
+        tokio::time::timeout(Duration::from_secs(30), barrier.wait_until_paused())
+            .await
+            .expect("failed tail must first reach the rename barrier");
+        wait_for_paused_tail_metadata_quorum(&disks, bucket, object).await;
+        assert!(!put.is_finished(), "committed quorum must still wait for the failing tail");
+        barrier.release();
+        tokio::time::timeout(Duration::from_secs(30), put)
+            .await
+            .expect("failed tail should drain")
+            .expect("PUT task should join")
+            .expect("a minority tail error must not negate committed quorum");
+        assert_eq!(tasks.running(), 0);
+        let heal = tokio::time::timeout(Duration::from_secs(30), heals.recv())
+            .await
+            .expect("failed tail must schedule heal")
+            .expect("heal capture must remain connected");
+        assert_eq!(heal.bucket, bucket);
+        assert_eq!(heal.object_prefix.as_deref(), Some(object));
+        let info = set
+            .get_object_info(bucket, object, &ObjectOptions::default())
+            .await
+            .expect("committed object must remain readable despite the failed tail");
+        assert_eq!(info.size, TEST_OBJECT_SIZE as i64);
+    }
+
+    #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
+    async fn tail_drained_put_rejects_quorum_minus_one() {
+        let (_dirs, disks, set) = hermetic_set_disks(4).await;
+        let bucket = "put-full-tail-no-quorum";
+        let object = "full-tail-no-quorum-object";
+        make_completion_test_bucket(&disks, bucket).await;
+        let _fault = rename_fault_injection::fail_rename_on(object, &[0, 1]);
+        let tasks = rename_fanout_barrier::observe_tasks(object);
+        let mut reader = PutObjReader::from_vec(vec![b'1'; TEST_OBJECT_SIZE]);
+        let err = set
+            .put_object(
+                bucket,
+                object,
+                &mut reader,
+                &ObjectOptions {
+                    write_completion: WriteCompletion::TailDrained,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect_err("draining two successful disks cannot satisfy write quorum three");
+        assert!(
+            matches!(err, Error::ErasureWriteQuorum | Error::InsufficientWriteQuorum(_, _)),
+            "original quorum error expected: {err}"
+        );
+        assert_eq!(tasks.running(), 0, "failed fan-out and rollback must complete before return");
+        assert!(
+            set.get_object_info(bucket, object, &ObjectOptions::default()).await.is_err(),
+            "failed fresh write must not become visible"
+        );
+    }
+
+    #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
+    async fn put_incomplete_rollback_preserves_staging_and_old_version_backup() {
+        use crate::set_disk::core::io_primitives::rollback_fault_injection;
+
+        temp_env::async_with_vars([(ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true"))], async {
+            for write_completion in [WriteCompletion::Quorum, WriteCompletion::TailDrained] {
+                for fault in [
+                    rollback_fault_injection::Fault::Io,
+                    rollback_fault_injection::Fault::VolumeNotFoundAfterRename,
+                ] {
+                    let (dirs, disks, set) = hermetic_set_disks(4).await;
+                    let bucket = "put-incomplete-undo";
+                    let object = "incomplete-undo-object";
+                    make_completion_test_bucket(&disks, bucket).await;
+                    let mut old_reader = PutObjReader::from_vec(vec![b'0'; TEST_OBJECT_SIZE]);
+                    set.put_object(
+                        bucket,
+                        object,
+                        &mut old_reader,
+                        &ObjectOptions {
+                            write_completion: WriteCompletion::TailDrained,
+                            ..Default::default()
+                        },
+                    )
+                    .await
+                    .expect("old generation should be completely committed");
+                    wait_for_tmp_workspace_to_drain(&dirs, "old PUT must leave no unrelated staging").await;
+                    let old = disks[0]
+                        .read_version("", bucket, object, "", &ReadOptions::default())
+                        .await
+                        .expect("old metadata must be readable");
+                    let old_data_dir = old.data_dir.expect("non-inline old version needs a data directory");
+                    let tasks = rename_fanout_barrier::observe_tasks(object);
+                    let barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier::PHASE_RENAME);
+                    let _rename_fault = rename_fault_injection::fail_rename_on(object, &[2, 3]);
+                    let _undo_fault = rollback_fault_injection::arm(object, 0, fault);
+                    let writer = Arc::clone(&set);
+                    let put = tokio::spawn(async move {
+                        let mut reader = PutObjReader::from_vec(vec![b'1'; TEST_OBJECT_SIZE]);
+                        writer
+                            .put_object(
+                                bucket,
+                                object,
+                                &mut reader,
+                                &ObjectOptions {
+                                    write_completion,
+                                    ..Default::default()
+                                },
+                            )
+                            .await
+                    });
+                    tokio::time::timeout(Duration::from_secs(30), barrier.wait_until_paused())
+                        .await
+                        .expect("overwrite must enter the actual rename fan-out before failure injection");
+                    barrier.release();
+                    let err = tokio::time::timeout(Duration::from_secs(30), put)
+                        .await
+                        .expect("incomplete undo must return without hanging")
+                        .expect("PUT task should join")
+                        .expect_err("two renamed disks cannot satisfy write quorum three");
+                    assert!(
+                        matches!(err, Error::ErasureWriteQuorum | Error::InsufficientWriteQuorum(_, _)),
+                        "original quorum error expected: {err}"
+                    );
+                    assert_eq!(tasks.running(), 0, "every rename and undo task must be reaped before return");
+                    let leftovers = non_trash_tmp_entries(&dirs).await;
+                    assert!(!leftovers.is_empty(), "incomplete undo must retain the new staging source for recovery");
+                    let backups = dirs
+                        .iter()
+                        .filter(|dir| {
+                            dir.path()
+                                .join(bucket)
+                                .join(object)
+                                .join(old_data_dir.to_string())
+                                .join(crate::disk::STORAGE_FORMAT_FILE_BACKUP)
+                                .exists()
+                        })
+                        .count();
+                    assert_eq!(backups, 1, "exactly the failed undo disk must retain its old-version backup");
+                    // The remaining three disks still serve the old generation;
+                    // the failed minority must never become an acknowledged write.
+                    let mut read = set
+                        .get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default())
+                        .await
+                        .expect("old generation must remain readable after incomplete rollback");
+                    let mut body = Vec::new();
+                    read.stream
+                        .read_to_end(&mut body)
+                        .await
+                        .expect("old generation should stream");
+                    assert_eq!(body, vec![b'0'; TEST_OBJECT_SIZE]);
+                }
+            }
+        })
+        .await;
+    }
+
+    #[tokio::test]
+    #[serial_test::serial(capacity_dirty_scope)]
+    async fn tail_drained_put_owned_commit_survives_waiter_cancellation() {
+        let (dirs, disks, set) = hermetic_set_disks(4).await;
+        let bucket = RUSTFS_META_BUCKET;
+        let object = "full-tail-cancelled-receipt";
+        // Internal config writes do not own a bucket lifecycle guard. The object
+        // guard alone must keep the full-tail coordinator alive after cancellation.
+        let tasks = rename_fanout_barrier::observe_tasks(object);
+        let barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier::PHASE_RENAME);
+        let writer = Arc::clone(&set);
+        let put = tokio::spawn(async move {
+            let mut reader = PutObjReader::from_vec(vec![b'1'; TEST_OBJECT_SIZE]);
+            writer
+                .put_object(
+                    bucket,
+                    object,
+                    &mut reader,
+                    &ObjectOptions {
+                        write_completion: WriteCompletion::TailDrained,
+                        ..Default::default()
+                    },
+                )
+                .await
+        });
+        tokio::time::timeout(Duration::from_secs(30), barrier.wait_until_paused())
+            .await
+            .expect("cancelled receipt must first reach the rename barrier");
+        wait_for_paused_tail_metadata_quorum(&disks, bucket, object).await;
+        put.abort();
+        assert!(put.await.expect_err("ACK waiter should cancel").is_cancelled());
+        let mut lock_probe = Box::pin(set.acquire_write_lock_diag("cancelled_full_tail_probe", bucket, object));
+        assert!(
+            futures::poll!(lock_probe.as_mut()).is_pending(),
+            "owned coordinator must retain the namespace guard after waiter cancellation"
+        );
+        barrier.release();
+        drop(
+            tokio::time::timeout(Duration::from_secs(30), lock_probe)
+                .await
+                .expect("cancelled coordinator must eventually release its guard")
+                .expect("post-commit lock probe should succeed"),
+        );
+        assert_eq!(tasks.running(), 0, "cancelled coordinator must reap every rename task");
+        for disk in &disks {
+            disk.read_version("", bucket, object, "", &ReadOptions::default())
+                .await
+                .expect("caller cancellation must not interrupt committed receipt materialization");
+        }
+        wait_for_tmp_workspace_to_drain(&dirs, "cancelled full-tail commit should release staging ownership").await;
+    }
+
     #[tokio::test]
     #[serial_test::serial(capacity_dirty_scope)]
     async fn no_lock_put_waits_for_rename_tail_under_outer_guard() {
@@ -18184,6 +18588,7 @@ mod put_object_tmp_cleanup_tests {
                         &mut reader,
                         &ObjectOptions {
                             no_lock: true,
+                            write_completion: WriteCompletion::TailDrained,
                             ..Default::default()
                         },
                     )
@@ -18209,7 +18614,18 @@ mod put_object_tmp_cleanup_tests {
             put.await
                 .expect("no-lock PUT task should join")
                 .expect("no-lock PUT should commit after the rename tail releases");
+            let mut lock_probe = Box::pin(set_disks.acquire_write_lock_diag("borrowed_full_tail_probe", bucket, object));
+            assert!(
+                futures::poll!(lock_probe.as_mut()).is_pending(),
+                "full-tail PUT must not release the caller's outer guard"
+            );
             drop(outer_guard);
+            drop(
+                tokio::time::timeout(Duration::from_secs(5), lock_probe)
+                    .await
+                    .expect("outer owner releasing its guard should unblock the probe")
+                    .expect("post-outer-guard probe should succeed"),
+            );
         })
         .await;
     }

crates/ecstore/src/set_disk/transition_matrix_tests.rs

@@ -18,6 +18,7 @@ use super::{
 };
 use crate::bucket::lifecycle::lifecycle::{TRANSITION_COMPLETE, TRANSITION_PENDING, TransitionOptions, expected_expiry_time};
 use crate::ecstore_validation_blackbox::make_local_set_disks;
+use crate::object_api::WriteCompletion;
 use crate::services::tier::test_util::register_mock_tier;
 use crate::storage_api_contracts::bucket::BucketOperations;
 use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations as _};
@@ -72,7 +73,7 @@ async fn transition_and_restore_reclaim_prior_metadata_generations() {
             object,
             &mut reader,
             &ObjectOptions {
-                no_lock: true,
+                write_completion: WriteCompletion::TailDrained,
                 ..Default::default()
             },
         )
@@ -185,7 +186,7 @@ async fn prepared_snapshot_transition_duplicate_and_late_get_use_committed_remot
             object,
             &mut reader,
             &ObjectOptions {
-                no_lock: true,
+                write_completion: WriteCompletion::TailDrained,
                 ..Default::default()
             },
         )

crates/ecstore/src/store/bucket.rs

@@ -1059,6 +1059,7 @@ mod tests {
     use crate::storage_api_contracts::{
         bucket::{BucketOperations as _, BucketOptions, DeleteBucketOptions, MakeBucketOptions, SRBucketDeleteOp},
         list::ListOperations as _,
+        namespace::NamespaceLocking as _,
         object::{ObjectIO as _, ObjectOperations as _},
     };
     use crate::store::{ECStore, init_local_disks_with_instance_ctx};
@@ -1486,10 +1487,19 @@ mod tests {
             .put_object(bucket, object, &mut reader, &ObjectOptions::default())
             .await
             .expect("object should be written");
+        let lock = ecstore.pools[0].disk_set[0]
+            .new_ns_lock(bucket, object)
+            .await
+            .expect("fixture namespace lock should be created");
+        drop(
+            lock.get_write_lock(Duration::from_secs(30))
+                .await
+                .expect("fixture rename tail should finish before checking its generation"),
+        );
         assert_eq!(
             ecstore.scanner_namespace_mutation_generation(),
-            generation_before_put.saturating_add(1),
+            generation_before_put.saturating_add(3),
-            "successful object creation should advance scanner namespace activity"
+            "successful object creation must observe the logical mutation and both fanout boundaries"
         );
         ecstore
             .get_object_info(bucket, object, &ObjectOptions::default())

crates/ecstore/src/store/init.rs

@@ -787,6 +787,12 @@ impl ECStore {
     pub fn single_pool(&self) -> bool {
         self.pools.len() == 1
     }
+
+    /// The set-local create-only check is atomic only when every object
+    /// mutation uses that same, enabled namespace lock domain.
+    pub fn supports_atomic_create_only_write_back(&self) -> bool {
+        !self.ctx.lock_manager().is_disabled() && self.pools.len() == 1 && self.pools[0].disk_set.len() == 1
+    }
 }
 
 #[cfg(test)]
@@ -2127,7 +2133,7 @@ mod tests {
             .iter()
             .map(|&drives_per_set| (1, drives_per_set))
             .collect::<Vec<_>>();
-        build_isolated_test_store_with_layout(temp_dir, cmd_line, &pool_layouts, shutdown).await
+        build_isolated_test_store_with_layout(temp_dir, cmd_line, &pool_layouts, shutdown, None).await
     }
 
     async fn build_isolated_test_store_with_layout(
@@ -2135,6 +2141,7 @@ mod tests {
         cmd_line: &str,
         pool_layouts: &[(usize, usize)],
         shutdown: CancellationToken,
+        instance_ctx: Option<Arc<crate::runtime::instance::InstanceContext>>,
     ) -> (
         Arc<crate::runtime::instance::InstanceContext>,
         Arc<crate::store::ECStore>,
@@ -2167,7 +2174,7 @@ mod tests {
         let endpoint_pools = EndpointServerPools(pools);
         crate::services::notification_sys::install_cross_pool_fence_fleet_proof_for_test();
 
-        let instance_ctx = Arc::new(crate::runtime::instance::InstanceContext::new());
+        let instance_ctx = instance_ctx.unwrap_or_else(|| Arc::new(crate::runtime::instance::InstanceContext::new()));
         crate::store::init_local_disks_with_instance_ctx(&instance_ctx, endpoint_pools.clone())
             .await
             .expect("register local disks into the fresh context");
@@ -2535,6 +2542,348 @@ mod tests {
         shutdown.cancel();
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    #[serial_test::serial(storage_class_env)]
+    async fn early_ack_put_tails_block_scanner_publication_until_all_renames_finish() {
+        use crate::storage_api_contracts::namespace::NamespaceLocking as _;
+
+        let temp_dir = tempfile::tempdir().expect("create scanner PUT tail store dir");
+        let (ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "scanner-put-tails", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(Arc::clone(&store), Vec::new()).await;
+        let bucket = format!("scanner-put-tails-{}", Uuid::new_v4());
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create scanner PUT tail bucket");
+        let set = &store.pools[0].disk_set[0];
+        let objects = [("scanner-tail-a", vec![0xA1; 273]), ("scanner-tail-b", vec![0xB2; 379])];
+
+        temp_env::async_with_vars([(crate::set_disk::ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true"))], async {
+            let (active, blocked, movement_generation) = store.scanner_data_movement_activity().await;
+            assert!(!active && !blocked);
+            assert!(ctx.scanner_publication_state_allowed(), "the set admission cache should start allowed");
+            let (old_lease, _) = store
+                .acquire_scanner_publication_lease(movement_generation, crate::runtime::instance::SCANNER_PUBLICATION_LEASE_TTL)
+                .await
+                .expect("publication lease should be admitted before either PUT starts");
+
+            let barriers: Vec<_> = objects
+                .iter()
+                .map(|(object, _)| {
+                    crate::set_disk::rename_fanout_barrier::arm(object, 0, crate::set_disk::rename_fanout_barrier::PHASE_RENAME)
+                })
+                .collect();
+            let trackers: Vec<_> = objects
+                .iter()
+                .map(|(object, _)| crate::set_disk::rename_fanout_barrier::observe_tasks(object))
+                .collect();
+            let puts: Vec<_> = objects
+                .iter()
+                .map(|(object, body)| {
+                    let put_store = Arc::clone(&store);
+                    let put_bucket = bucket.clone();
+                    let object = *object;
+                    let body = body.clone();
+                    tokio::spawn(async move {
+                        let mut reader = PutObjReader::from_vec(body);
+                        put_store
+                            .put_object(&put_bucket, object, &mut reader, &ObjectOptions::default())
+                            .await
+                    })
+                })
+                .collect();
+            let committed = tokio::time::timeout(Duration::from_secs(30), async {
+                for barrier in &barriers {
+                    barrier.wait_until_paused().await;
+                }
+                let mut committed = Vec::with_capacity(puts.len());
+                for put in puts {
+                    committed.push(
+                        put.await
+                            .expect("early-ACK PUT task should join while its tail is paused")
+                            .expect("root PUT should return after quorum without waiting for its tail"),
+                    );
+                }
+                committed
+            })
+            .await
+            .expect("both root PUTs must quorum-ACK while their tail disks remain paused");
+
+            assert!(trackers.iter().all(|tracker| tracker.running() >= 1));
+            assert!(ctx.namespace_commits_pending());
+            assert!(
+                ctx.scanner_publication_state_allowed(),
+                "pending PUT tails must not disable scanner namespace walks"
+            );
+            let (active, blocked, observed_movement_generation) = store.scanner_data_movement_activity().await;
+            assert!(!active, "ordinary PUT tails are not decommission or rebalance work");
+            assert!(!blocked, "ordinary PUT tails must not block the movement-only scan baseline");
+            assert_eq!(observed_movement_generation, movement_generation);
+            assert!(store.scanner_data_usage_publication_blocked().await);
+            assert!(store.scanner_data_usage_publication_admission_guard().await.is_some());
+            assert!(set.scanner_data_usage_publication_admission_guard().await.is_some());
+            for error in [
+                store
+                    .acquire_scanner_publication_lease(
+                        movement_generation,
+                        crate::runtime::instance::SCANNER_PUBLICATION_LEASE_TTL,
+                    )
+                    .await
+                    .expect_err("a new remote publication lease must reject pending PUT tails"),
+                store
+                    .validate_scanner_publication_lease(old_lease, movement_generation)
+                    .await
+                    .expect_err("an existing remote lease must not bypass pending PUT tails"),
+                store
+                    .acquire_scanner_publication_lease_guard(old_lease)
+                    .await
+                    .expect_err("target-side publication admission must reject pending PUT tails"),
+            ] {
+                assert!(
+                    error.to_string().contains("blocked"),
+                    "publication must fail because of active tails: {error}"
+                );
+            }
+            store.release_scanner_publication_lease(old_lease).await;
+
+            for (index, barrier) in barriers.iter().enumerate() {
+                let commit_generation = ctx.namespace_commit_generation();
+                let namespace_generation = store.scanner_namespace_mutation_generation();
+                barrier.release();
+                tokio::time::timeout(Duration::from_secs(30), async {
+                    while trackers[index].running() != 0 || ctx.namespace_commit_generation() <= commit_generation {
+                        tokio::task::yield_now().await;
+                    }
+                    if index + 1 == barriers.len() {
+                        while ctx.namespace_commits_pending() {
+                            tokio::task::yield_now().await;
+                        }
+                    }
+                })
+                .await
+                .expect("released tail must drain and publish its terminal namespace generation");
+                assert!(store.scanner_namespace_mutation_generation() > namespace_generation);
+                let pending = index + 1 < barriers.len();
+                assert_eq!(ctx.namespace_commits_pending(), pending);
+                assert_eq!(store.scanner_data_usage_publication_blocked().await, pending);
+                assert!(!store.scanner_data_movement_activity().await.1);
+                assert!(store.scanner_data_usage_publication_admission_guard().await.is_some());
+                assert!(set.scanner_data_usage_publication_admission_guard().await.is_some());
+            }
+
+            let (lease, generation) = store
+                .acquire_scanner_publication_lease(movement_generation, crate::runtime::instance::SCANNER_PUBLICATION_LEASE_TTL)
+                .await
+                .expect("remote publication lease should resume after both tails drain");
+            store
+                .validate_scanner_publication_lease(lease, generation)
+                .await
+                .expect("a resumed remote publication lease should validate");
+            drop(
+                store
+                    .acquire_scanner_publication_lease_guard(lease)
+                    .await
+                    .expect("target-side publication admission should resume after both tails drain"),
+            );
+            assert!(store.release_scanner_publication_lease(lease).await);
+
+            let disks = set.disk_inventory().await;
+            assert_eq!(disks.len(), 4);
+            for ((object, body), committed) in objects.iter().zip(&committed) {
+                let logical_size = i64::try_from(body.len()).expect("fixture payload size should fit i64");
+                let etag = committed.etag.as_ref().expect("root PUT should return a committed ETag");
+                for (disk_index, disk) in disks.iter().enumerate() {
+                    let file_info = disk
+                        .as_ref()
+                        .expect("every fixture disk should remain online")
+                        .read_version(
+                            "",
+                            &bucket,
+                            object,
+                            "",
+                            &crate::disk::ReadOptions {
+                                read_data: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                        .unwrap_or_else(|err| panic!("disk {disk_index} should publish {object} after its tail finishes: {err}"));
+                    assert_eq!(file_info.size, logical_size);
+                    assert_eq!(file_info.metadata.get(http::header::ETAG.as_str()), Some(etag));
+                    assert!(
+                        file_info.inline_data(),
+                        "small fixture payloads should have an inline shard on every disk"
+                    );
+                    let inline_data = file_info.data.as_ref().expect("every disk should retain its inline shard");
+                    let erasure = crate::erasure::coding::Erasure::try_new_with_options(
+                        file_info.erasure.data_blocks,
+                        file_info.erasure.parity_blocks,
+                        file_info.erasure.block_size,
+                        file_info.uses_legacy_checksum,
+                    )
+                    .expect("persisted erasure geometry should be valid");
+                    let shard_size =
+                        usize::try_from(erasure.shard_file_size(logical_size)).expect("fixture shard size should fit usize");
+                    crate::erasure::coding::bitrot_verify(
+                        Cursor::new(inline_data.clone()),
+                        inline_data.len(),
+                        shard_size,
+                        rustfs_utils::HashAlgorithm::HighwayHash256S,
+                        erasure.shard_size(),
+                    )
+                    .await
+                    .unwrap_or_else(|err| panic!("disk {disk_index} should retain a complete valid shard for {object}: {err}"));
+                }
+                let mut reader = store
+                    .get_object_reader(&bucket, object, None, HeaderMap::new(), &ObjectOptions::default())
+                    .await
+                    .expect("fully drained PUT should be readable");
+                let mut actual = Vec::new();
+                reader.stream.read_to_end(&mut actual).await.expect("PUT body should drain");
+                assert_eq!(&actual, body);
+            }
+
+            let generation_before_internal_put = ctx.namespace_commit_generation();
+            let internal_object = "scanner-tail-regression/internal-metadata";
+            let internal_body = b"scanner metadata must not invalidate its own publication";
+            let mut internal_reader = PutObjReader::from_vec(internal_body.to_vec());
+            store
+                .put_object(RUSTFS_META_BUCKET, internal_object, &mut internal_reader, &ObjectOptions::default())
+                .await
+                .expect("internal metadata PUT should commit without scanner self-invalidation");
+            let internal_lock = set
+                .new_ns_lock(RUSTFS_META_BUCKET, internal_object)
+                .await
+                .expect("internal metadata tail lock should be available");
+            drop(
+                internal_lock
+                    .get_write_lock(Duration::from_secs(30))
+                    .await
+                    .expect("internal metadata tail should drain"),
+            );
+            assert_eq!(ctx.namespace_commit_generation(), generation_before_internal_put);
+            assert!(!ctx.namespace_commits_pending());
+            assert!(store.scanner_data_usage_publication_admission_guard().await.is_some());
+            assert!(set.scanner_data_usage_publication_admission_guard().await.is_some());
+            let mut internal_reader = store
+                .get_object_reader(RUSTFS_META_BUCKET, internal_object, None, HeaderMap::new(), &ObjectOptions::default())
+                .await
+                .expect("internal metadata should remain readable");
+            let mut actual = Vec::new();
+            internal_reader
+                .stream
+                .read_to_end(&mut actual)
+                .await
+                .expect("internal metadata body should drain");
+            assert_eq!(actual, internal_body);
+        })
+        .await;
+        shutdown.cancel();
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    #[serial_test::serial(storage_class_env)]
+    async fn cancelled_early_ack_put_keeps_scanner_publication_blocked_until_tail_finishes() {
+        let temp_dir = tempfile::tempdir().expect("create cancelled scanner PUT tail store dir");
+        let (ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "scanner-cancelled-put-tail", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(Arc::clone(&store), Vec::new()).await;
+        let bucket = format!("scanner-cancelled-put-tail-{}", Uuid::new_v4());
+        let object = "scanner-cancelled-tail";
+        let body = vec![0xC3; 273];
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create cancelled scanner PUT tail bucket");
+
+        temp_env::async_with_vars([(crate::set_disk::ENV_RUSTFS_PUT_RENAME_EARLY_ACK_ENABLE, Some("true"))], async {
+            let tracker = crate::set_disk::rename_fanout_barrier::observe_tasks(object);
+            let tail =
+                crate::set_disk::rename_fanout_barrier::arm(object, 0, crate::set_disk::rename_fanout_barrier::PHASE_RENAME);
+            let quorum = crate::set_disk::PutObjectCommitBarrier::install(
+                &bucket,
+                object,
+                crate::set_disk::PutObjectCommitPause::AfterRenameQuorum,
+            );
+            let handoff = crate::set_disk::PutObjectCommitBarrier::install(
+                &bucket,
+                object,
+                crate::set_disk::PutObjectCommitPause::AfterRenameHandoff,
+            );
+            let put_store = Arc::clone(&store);
+            let put_bucket = bucket.clone();
+            let put_body = body.clone();
+            let put = tokio::spawn(async move {
+                let mut reader = PutObjReader::from_vec(put_body);
+                put_store
+                    .put_object(&put_bucket, object, &mut reader, &ObjectOptions::default())
+                    .await
+            });
+            tokio::time::timeout(Duration::from_secs(30), tail.wait_until_paused())
+                .await
+                .expect("cancelled PUT should pause one disk before rename");
+            quorum.wait_until_paused().await;
+            put.abort();
+            assert!(
+                put.await
+                    .expect_err("caller should be cancelled after rename quorum")
+                    .is_cancelled()
+            );
+            quorum.release();
+            handoff.wait_until_paused().await;
+            assert!(tracker.running() >= 1);
+            assert!(ctx.namespace_commits_pending());
+            assert!(!store.scanner_data_movement_activity().await.1);
+            assert!(store.scanner_data_usage_publication_blocked().await);
+            assert!(store.scanner_data_usage_publication_admission_guard().await.is_some());
+            assert!(
+                store.pools[0].disk_set[0]
+                    .scanner_data_usage_publication_admission_guard()
+                    .await
+                    .is_some()
+            );
+            let generation = store.scanner_namespace_mutation_generation();
+
+            handoff.release();
+            tail.release();
+            tokio::time::timeout(Duration::from_secs(30), async {
+                while tracker.running() != 0 || ctx.namespace_commits_pending() {
+                    tokio::task::yield_now().await;
+                }
+            })
+            .await
+            .expect("cancelled request's detached fanout must release scanner admission after finishing");
+            assert!(store.scanner_namespace_mutation_generation() > generation);
+            assert!(!store.scanner_data_usage_publication_blocked().await);
+            assert!(store.scanner_data_usage_publication_admission_guard().await.is_some());
+            for (disk_index, disk) in store.pools[0].disk_set[0].disk_inventory().await.iter().enumerate() {
+                let file_info = disk
+                    .as_ref()
+                    .expect("cancelled PUT fixture disk should remain online")
+                    .read_version("", &bucket, object, "", &crate::disk::ReadOptions::default())
+                    .await
+                    .unwrap_or_else(|err| panic!("cancelled PUT must still publish on disk {disk_index}: {err}"));
+                assert_eq!(file_info.size, i64::try_from(body.len()).expect("fixture body size should fit i64"));
+            }
+            let mut reader = store
+                .get_object_reader(&bucket, object, None, HeaderMap::new(), &ObjectOptions::default())
+                .await
+                .expect("a cancelled caller must not discard its quorum-committed object");
+            let mut actual = Vec::new();
+            reader
+                .stream
+                .read_to_end(&mut actual)
+                .await
+                .expect("cancelled PUT body should drain");
+            assert_eq!(actual, body);
+        })
+        .await;
+        shutdown.cancel();
+    }
+
     #[cfg(feature = "test-util")]
     #[test]
     #[serial_test::serial(storage_class_env)]
@@ -2979,6 +3328,35 @@ mod tests {
     #[cfg(feature = "test-util")]
     const DECOMMISSION_TEST_FAULT_STAGE_TIERED: &str = "decommission_tiered_object";
 
+    fn decommission_retry_fault_hook(
+        bucket: &str,
+        object: &str,
+        faults: Arc<AtomicUsize>,
+    ) -> crate::core::pools::DecommissionTestFaultDecision {
+        let target_bucket = bucket.to_string();
+        let target_object = object.to_string();
+        Arc::new(move |stage, bucket, object, attempt, succeeded| {
+            if !succeeded
+                || attempt >= crate::core::pools::DECOMMISSION_VERSION_COPY_ATTEMPTS
+                || stage != DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT
+                || bucket != target_bucket
+                || object != target_object
+            {
+                return false;
+            }
+
+            // Entry retries reset the local attempt; real copy errors can skip
+            // successful attempts. Only injected faults spend this global budget.
+            // A real failure may consume an attempt, so preserve the final chance.
+            faults
+                .fetch_update(Ordering::SeqCst, Ordering::SeqCst, |faults| {
+                    (faults < crate::core::pools::DECOMMISSION_VERSION_COPY_ATTEMPTS.saturating_sub(1))
+                        .then_some(faults.saturating_add(1))
+                })
+                .is_ok()
+        })
+    }
+
     async fn seed_decommission_source(
         store: &Arc<crate::store::ECStore>,
         bucket: &str,
@@ -4991,6 +5369,7 @@ mod tests {
             "decommission-delete-fence",
             &[(2, 4), (1, 4)],
             CancellationToken::new(),
+            None,
         ))
         .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
@@ -5120,6 +5499,43 @@ mod tests {
         shutdown.cancel();
     }
 
+    #[test]
+    fn decommission_retry_fault_budget_counts_successes_across_attempt_changes() {
+        let cases: &[&[(usize, bool, bool)]] = &[
+            &[(1, true, true), (2, true, true), (3, true, false)],
+            &[(1, true, true), (1, true, true), (2, true, false)],
+            &[(1, true, true), (3, true, false), (3, true, false)],
+            &[(1, true, true), (2, false, false), (1, true, true), (2, true, false)],
+            &[(1, true, true), (2, false, false), (3, true, false)],
+            &[(3, true, false), (4, true, false)],
+        ];
+        for case in cases {
+            let faults = Arc::new(AtomicUsize::new(0));
+            let hook = decommission_retry_fault_hook("bucket", "object", Arc::clone(&faults));
+
+            for (stage, bucket, object, succeeded) in [
+                ("other-stage", "bucket", "object", true),
+                (DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT, "other-bucket", "object", true),
+                (DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT, "bucket", "other-object", true),
+                (DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT, "bucket", "object", false),
+            ] {
+                assert!(!hook(stage, bucket, object, 1, succeeded));
+            }
+            assert_eq!(faults.load(Ordering::SeqCst), 0, "unrelated or failed copies must not consume faults");
+
+            let mut expected_faults = 0;
+            for &(attempt, succeeded, expected) in *case {
+                assert_eq!(
+                    hook(DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT, "bucket", "object", attempt, succeeded),
+                    expected,
+                    "fault plan {case:?} at attempt {attempt}"
+                );
+                expected_faults += usize::from(expected);
+                assert_eq!(faults.load(Ordering::SeqCst), expected_faults);
+            }
+        }
+    }
+
     #[test]
     #[serial_test::serial(storage_class_env)]
     fn decommission_entry_retries_source_changed_without_canceling_other_bucket() {
@@ -5214,31 +5630,8 @@ mod tests {
                     ));
 
                     let ordinary_faults = Arc::new(AtomicUsize::new(0));
-                    let ordinary_faults_for_hook = Arc::clone(&ordinary_faults);
+                    let fault_hook = decommission_retry_fault_hook(&other_bucket, other_object, Arc::clone(&ordinary_faults));
-                    let fault_bucket = other_bucket.clone();
+                    let _fault_guard = crate::core::pools::DecommissionTestFaultGuard::install(fault_hook);
-                    let _fault_guard = crate::core::pools::DecommissionTestFaultGuard::install(Arc::new(
-                        move |stage, bucket, object, attempt, succeeded| {
-                            let candidate = succeeded
-                                && stage == DECOMMISSION_TEST_FAULT_STAGE_MIGRATE_OBJECT
-                                && bucket == fault_bucket.as_str()
-                                && object == other_object;
-                            if !candidate {
-                                return false;
-                            }
-
-                            // Keep the fault budget global across any
-                            // entry-level re-list; its inner attempt counter
-                            // restarts after SourceChanged.
-                            ordinary_faults_for_hook
-                                .fetch_update(Ordering::SeqCst, Ordering::SeqCst, |faults| {
-                                    let next_fault = faults.saturating_add(1);
-                                    (faults < crate::core::pools::DECOMMISSION_VERSION_COPY_ATTEMPTS.saturating_sub(1)
-                                        && attempt == next_fault)
-                                        .then_some(next_fault)
-                                })
-                                .is_ok()
-                        },
-                    ));
 
                     let rx = CancellationToken::new();
                     let source_changed_exhaustions = Arc::new(AtomicUsize::new(0));
@@ -5275,6 +5668,15 @@ mod tests {
                     changed_result.expect("SourceChanged entry retry must converge");
                     other_result.expect("other bucket entry must continue through ordinary copy retries");
 
+                    assert_eq!(
+                        store.pool_meta.read().await.pools[0]
+                            .decommission
+                            .as_ref()
+                            .expect("decommission progress should remain available")
+                            .items_decommission_failed,
+                        0,
+                        "entry completion must not hide an exhausted copy failure"
+                    );
                     assert!(!rx.is_cancelled(), "entry-level SourceChanged must not cancel the shared worker token");
                     assert_eq!(mutation_calls.load(Ordering::SeqCst), 2, "entry must be re-listed after SourceChanged");
                     assert_eq!(ordinary_faults.load(Ordering::SeqCst), 2, "ordinary copy must consume the retry budget");
@@ -5903,6 +6305,7 @@ mod tests {
             "reverse-decommission-fixed-target",
             &[(1, 4), (1, 4)],
             CancellationToken::new(),
+            None,
         ))
         .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
@@ -6324,6 +6727,7 @@ mod tests {
             "multi-set-decommission-source-cleanup",
             &[(2, 4)],
             CancellationToken::new(),
+            None,
         ))
         .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
@@ -8045,10 +8449,15 @@ mod tests {
         );
         assert!(com::read_config(store.pools[0].clone(), &second_page_path).await.is_ok());
 
-        com::save_config(store.pools[target_pool_idx].clone(), &second_page_path, receipt_bytes.clone())
+        let full_tail = ObjectOptions {
+            max_parity: true,
+            write_completion: crate::object_api::WriteCompletion::TailDrained,
+            ..Default::default()
+        };
+        com::save_config_with_opts(store.pools[target_pool_idx].clone(), &second_page_path, receipt_bytes.clone(), &full_tail)
             .await
             .expect("second page receipt should restore");
-        com::save_config(store.pools[target_pool_idx].clone(), &second_page_path, b"{corrupt".to_vec())
+        com::save_config_with_opts(store.pools[target_pool_idx].clone(), &second_page_path, b"{corrupt".to_vec(), &full_tail)
             .await
             .expect("second page receipt should corrupt deterministically");
         let corrupt = store
@@ -8834,18 +9243,17 @@ mod tests {
         const MANIFEST_COUNT: usize = 10;
 
         let temp_dir = tempfile::tempdir().expect("create fast manifest pass recovery store dir");
-        let (ctx, store, _shutdown) =
+        let mut instance_ctx = crate::runtime::instance::InstanceContext::new();
-            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "tier-delete-fast-manifest-pass", &[4])).await;
+        instance_ctx.suppress_tier_delete_journal_recovery_for_test();
+        let (ctx, store, shutdown) = without_storage_class_env(build_isolated_test_store_with_layout(
+            temp_dir.path(),
+            "tier-delete-fast-manifest-pass",
+            &[(1, 4)],
+            CancellationToken::new(),
+            Some(Arc::new(instance_ctx)),
+        ))
+        .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
-        let bucket = "tier-delete-fast-manifest-pass-bucket";
-        store
-            .make_bucket(bucket, &MakeBucketOptions::default())
-            .await
-            .expect("fast manifest pass bucket should be created");
-        let incarnation = store
-            .bucket_incarnation_id(bucket)
-            .await
-            .expect("fast manifest pass bucket incarnation should resolve");
         let tier_name = "FAST-MANIFEST-PASS";
         let backend = register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
         let backend_identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
@@ -8853,9 +9261,19 @@ mod tests {
             .expect("fast manifest pass tier lease should resolve")
             .backend_identity();
         for index in 0..MANIFEST_COUNT {
+            // Pagination must not depend on same-bucket lock wait deadlines.
+            let bucket = format!("tier-delete-fast-manifest-pass-{index}");
+            store
+                .make_bucket(&bucket, &MakeBucketOptions::default())
+                .await
+                .expect("fast manifest pass bucket should be created");
+            let incarnation = store
+                .bucket_incarnation_id(&bucket)
+                .await
+                .expect("fast manifest pass bucket incarnation should resolve");
             install_aborting_dispatch_fixture(
                 store.clone(),
-                bucket,
+                &bucket,
                 incarnation,
                 &format!("manifest-page-{index:06}/"),
                 tier_name,
@@ -8886,12 +9304,78 @@ mod tests {
             "one production pass must cross the default eight-manifest page limit"
         );
         assert_eq!(stats.manifests.scanned, MANIFEST_COUNT);
-        assert_eq!(stats.manifests.deleted, MANIFEST_COUNT);
+        assert_eq!(stats.manifests.deleted, MANIFEST_COUNT, "full recovery result: {stats:?}");
-        assert_eq!(stats.manifests.failed, 0);
+        assert_eq!(stats.manifests.failed, 0, "full recovery result: {stats:?}");
         assert_eq!(manifest_marker, None);
         assert_eq!(tier_delete_dispatch_manifest_count(store.clone()).await, 0);
         assert_eq!(tier_delete_journal_count(store).await, 0);
         assert_eq!(backend.remove_count().await, 0, "rollback recovery must not call the remote tier");
+        shutdown.cancel();
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn tier_delete_manual_pass_retains_manifest_owned_by_startup_recovery() {
+        let temp_dir = tempfile::tempdir().expect("create automatic recovery ownership store dir");
+        let (ctx, store, shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "tier-delete-auto-owner", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let bucket = "tier-delete-auto-owner-bucket";
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("automatic recovery bucket should be created");
+        let incarnation = store.bucket_incarnation_id(bucket).await.expect("bucket incarnation");
+        let tier_name = "AUTO-OWNER";
+        let backend = register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
+            .await
+            .expect("automatic recovery tier lease")
+            .backend_identity();
+
+        // The automatic worker must not observe a partially installed fixture.
+        let lifecycle_guard = store
+            .acquire_bucket_lifecycle_write_lock(bucket)
+            .await
+            .expect("fixture lifecycle lock");
+        let (manifest_name, entries) =
+            install_aborting_dispatch_fixture(store.clone(), bucket, incarnation, "auto-owner/", tier_name, identity, 1).await;
+        let journal_name = tier_delete_journal_object_name(&entries[0]);
+        let hook = TierDeleteDispatchRollbackTestHook::install_slow_delete(&journal_name, &journal_name);
+        drop(lifecycle_guard);
+        ctx.wake_tier_delete_journal_recovery();
+        tokio::time::timeout(Duration::from_secs(30), hook.wait_until_delete_paused())
+            .await
+            .expect("startup recovery should own the manifest before a manual pass");
+        assert!(tier_delete_dispatch_manifest_recovery_inflight_for_test(&store, &manifest_name));
+
+        let stats = recover_tier_delete_dispatch_manifests(store.clone(), 8, None)
+            .await
+            .expect("manual recovery scan");
+        assert_eq!(stats.scanned, 1, "{stats:?}");
+        assert_eq!(stats.retained, 1, "{stats:?}");
+        assert_eq!(stats.deleted, 0, "{stats:?}");
+        assert_eq!(stats.failed, 0, "{stats:?}");
+        assert_eq!(tier_delete_dispatch_manifest_count(store.clone()).await, 1);
+        assert_eq!(tier_delete_journal_count(store.clone()).await, 1);
+
+        hook.release_delete();
+        tokio::time::timeout(Duration::from_secs(30), async {
+            loop {
+                let manifest_gone = matches!(com::read_config(store.clone(), &manifest_name).await, Err(Error::ConfigNotFound));
+                if manifest_gone && !tier_delete_dispatch_manifest_recovery_inflight_for_test(&store, &manifest_name) {
+                    break;
+                }
+                tokio::time::sleep(Duration::from_millis(10)).await;
+            }
+        })
+        .await
+        .expect("automatic recovery should converge without a manual retry");
+        assert_eq!(tier_delete_dispatch_manifest_count(store.clone()).await, 0);
+        assert_eq!(tier_delete_journal_count(store).await, 0);
+        assert_eq!(backend.remove_count().await, 0, "rollback must not delete from the remote tier");
+        shutdown.cancel();
     }
 
     #[cfg(feature = "test-util")]
@@ -10266,8 +10750,17 @@ mod tests {
         const JOURNAL_COUNT: usize = 40;
 
         let temp_dir = tempfile::tempdir().expect("create rollback retry store dir");
-        let (ctx, store, _shutdown) =
+        // Manual retries must own progress between fault removal and the next attempt.
-            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "dispatch-rollback-retry", &[4])).await;
+        let mut instance_ctx = crate::runtime::instance::InstanceContext::new();
+        instance_ctx.suppress_tier_delete_journal_recovery_for_test();
+        let (ctx, store, shutdown) = without_storage_class_env(build_isolated_test_store_with_layout(
+            temp_dir.path(),
+            "dispatch-rollback-retry",
+            &[(1, 4)],
+            CancellationToken::new(),
+            Some(Arc::new(instance_ctx)),
+        ))
+        .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
         let bucket = "dispatch-rollback-retry-bucket";
         store
@@ -10341,6 +10834,7 @@ mod tests {
 
         assert_eq!(tier_delete_dispatch_manifest_count(store.clone()).await, 0);
         assert_eq!(backend.remove_count().await, 0, "rollback retries must never call the remote tier");
+        shutdown.cancel();
     }
 
     #[cfg(feature = "test-util")]
@@ -11575,6 +12069,7 @@ mod tests {
         pool_index: usize,
         bucket: &str,
         object: &str,
+        minio_unversioned: bool,
     ) {
         for disk_index in 0..4 {
             let metadata_path =
@@ -11608,6 +12103,11 @@ mod tests {
                 ] {
                     rustfs_utils::http::metadata_compat::remove_bytes(&mut object_meta.meta_sys, suffix);
                 }
+                if minio_unversioned {
+                    object_meta
+                        .meta_sys
+                        .insert("x-minio-internal-transitioned-versionID".to_string(), Vec::new());
+                }
                 *shallow = rustfs_filemeta::FileMetaShallowVersion::try_from(version)
                     .expect("legacy transitioned version should re-encode");
             }
@@ -11618,6 +12118,152 @@ mod tests {
         }
     }
 
+    #[cfg(feature = "test-util")]
+    async fn read_store_body(
+        store: &Arc<crate::store::ECStore>,
+        bucket: &str,
+        object: &str,
+        range: Option<HTTPRangeSpec>,
+        opts: &ObjectOptions,
+    ) -> Vec<u8> {
+        let mut reader = store
+            .get_object_reader(bucket, object, range, HeaderMap::new(), opts)
+            .await
+            .expect("object reader should open");
+        let mut body = Vec::new();
+        reader.stream.read_to_end(&mut body).await.expect("object body should drain");
+        body
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn legacy_unknown_unversioned_transition_supports_head_get_and_range_without_backfill() {
+        let temp_dir = tempfile::tempdir().expect("create legacy unknown unversioned store dir");
+        let (ctx, store, _shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "legacy-unknown-unversioned-read", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let tier_name = "LEGACY-UNKNOWN-UNVERSIONED-READ";
+        let backend = register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
+        backend.set_put_remote_version(Some(String::new())).await;
+        let bucket = "legacy-unknown-unversioned-read-bucket";
+        let object = "object.bin";
+        let payload = b"legacy unversioned remote tier object remains readable".repeat(1024);
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("legacy source bucket should be created");
+        let mut reader = PutObjReader::from_vec(payload.clone());
+        let source = store
+            .put_object(bucket, object, &mut reader, &ObjectOptions::default())
+            .await
+            .expect("legacy source should be written");
+        store
+            .transition_object(
+                bucket,
+                object,
+                &ObjectOptions {
+                    transition: TransitionOptions {
+                        status: TRANSITION_PENDING.to_string(),
+                        tier: tier_name.to_string(),
+                        etag: source.etag.clone().expect("legacy source should have an etag"),
+                        ..Default::default()
+                    },
+                    mod_time: source.mod_time,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("legacy source should transition");
+        rewrite_transitioned_xlmeta_as_legacy_unknown(temp_dir.path(), 0, bucket, object, true).await;
+        backend.clear_op_log().await;
+
+        let opts = ObjectOptions {
+            metadata_cache_safe: false,
+            ..Default::default()
+        };
+        let head = store
+            .get_object_info(bucket, object, &opts)
+            .await
+            .expect("legacy transitioned HEAD should use local metadata");
+        assert_eq!(head.transition_version_state, rustfs_filemeta::TransitionVersionState::Unknown);
+        assert!(head.transitioned_object.version_id.is_empty());
+        assert_eq!(
+            head.user_defined
+                .get("x-minio-internal-transitioned-versionID")
+                .map(String::as_str),
+            Some(""),
+            "the MinIO empty version-key provenance must survive xl.meta decoding"
+        );
+        assert!(
+            !rustfs_utils::http::metadata_compat::contains_key_str(
+                &head.user_defined,
+                rustfs_utils::http::metadata_compat::SUFFIX_TRANSITIONED_VERSION_STATE,
+            ),
+            "the compatibility read must not synthesize version-state metadata"
+        );
+
+        let full_body = read_store_body(&store, bucket, object, None, &opts).await;
+        assert_eq!(full_body, payload);
+
+        let range = HTTPRangeSpec {
+            is_suffix_length: false,
+            start: 7,
+            end: 38,
+        };
+        let ranged_body = read_store_body(&store, bucket, object, Some(range), &opts).await;
+        assert_eq!(ranged_body, &payload[7..=38]);
+
+        let after_read = store.pools[0]
+            .get_disks_by_key(object)
+            .load_file_info_versions_exact(bucket, object)
+            .await
+            .expect("legacy metadata should remain readable after GET")
+            .expect("legacy object metadata should remain on disk")
+            .versions
+            .into_iter()
+            .find(|version| version.transition_status == rustfs_filemeta::TRANSITION_COMPLETE)
+            .expect("legacy transitioned source should remain visible after GET");
+        assert_eq!(after_read.transition_version_state, rustfs_filemeta::TransitionVersionState::Unknown);
+        assert!(after_read.transition_version.is_none());
+        assert!(after_read.transition_version_id.is_none());
+        assert_eq!(
+            after_read
+                .metadata
+                .get("x-minio-internal-transitioned-versionID")
+                .map(String::as_str),
+            Some(""),
+            "the MinIO empty version-key provenance must remain after GET and Range GET"
+        );
+        assert!(
+            !rustfs_utils::http::metadata_compat::contains_key_str(
+                &after_read.metadata,
+                rustfs_utils::http::metadata_compat::SUFFIX_TRANSITIONED_VERSION_STATE,
+            ),
+            "the compatibility read must remain side-effect free"
+        );
+
+        assert_eq!(
+            backend.op_log().await,
+            vec![
+                MockWarmOp::Probe {
+                    object: after_read.transitioned_objname.clone(),
+                },
+                MockWarmOp::Get {
+                    object: after_read.transitioned_objname.clone(),
+                },
+                MockWarmOp::Probe {
+                    object: after_read.transitioned_objname.clone(),
+                },
+                MockWarmOp::Get {
+                    object: after_read.transitioned_objname,
+                },
+            ],
+            "legacy reads should probe before each unversioned GET and never mutate local metadata"
+        );
+        assert_eq!(backend.remove_count().await, 0);
+    }
+
     #[cfg(feature = "test-util")]
     #[tokio::test]
     #[serial_test::serial(storage_class_env)]
@@ -11658,7 +12304,7 @@ mod tests {
                 )
                 .await
                 .expect("legacy source should transition");
-            rewrite_transitioned_xlmeta_as_legacy_unknown(temp_dir.path(), 0, bucket, object).await;
+            rewrite_transitioned_xlmeta_as_legacy_unknown(temp_dir.path(), 0, bucket, object, false).await;
             let legacy = store.pools[0]
                 .get_disks_by_key(object)
                 .load_file_info_versions_exact(bucket, object)
@@ -12799,7 +13445,7 @@ mod tests {
                 .expect("merge-loser source should transition");
             copy_test_xlmeta_between_pools(temp_dir.path(), 0, 1, bucket, object).await;
         }
-        rewrite_transitioned_xlmeta_as_legacy_unknown(temp_dir.path(), 1, bucket, "legacy/item.bin").await;
+        rewrite_transitioned_xlmeta_as_legacy_unknown(temp_dir.path(), 1, bucket, "legacy/item.bin", false).await;
         backend.set_remove_failure(true);
         store.pools[1]
             .delete_object(bucket, "hidden/item.bin", ObjectOptions::default())
@@ -13016,6 +13662,7 @@ mod tests {
             "partial-set-prefix-delete",
             &[(2, 4)],
             CancellationToken::new(),
+            None,
         ))
         .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
@@ -16388,6 +17035,7 @@ mod tests {
             "prepared-directory-recovery",
             &[(2, 4)],
             shutdown,
+            None,
         ))
         .await;
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
@@ -16866,6 +17514,10 @@ mod tests {
                             .find(|version| version.version_id == history.version_id)
                             .expect("transitioned history should exist");
                         transitioned.transition_version_state = rustfs_filemeta::TransitionVersionState::Unknown;
+                        rustfs_utils::http::metadata_compat::remove_str(
+                            &mut transitioned.metadata,
+                            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITIONED_VERSION_STATE,
+                        );
                         metadata
                             .add_version(transitioned)
                             .expect("unknown state should replace the transitioned version");
@@ -17036,6 +17688,38 @@ mod tests {
             .expect("test thread should complete");
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn odm_write_back_requires_one_set_and_enabled_namespace_locking() {
+        for (layout, locking, supported) in [
+            (&[(1, 4)][..], true, true),
+            (&[(1, 4), (1, 4)][..], true, false),
+            (&[(2, 4)][..], true, false),
+            (&[(1, 4)][..], false, false),
+        ] {
+            temp_env::async_with_vars([("RUSTFS_LOCK_ENABLED", Some(if locking { "true" } else { "false" }))], async {
+                let dir = tempfile::tempdir().expect("isolated topology");
+                let shutdown = CancellationToken::new();
+                let (_ctx, store, _) = without_storage_class_env(build_isolated_test_store_with_layout(
+                    dir.path(),
+                    "odm-topology",
+                    layout,
+                    shutdown.clone(),
+                    None,
+                ))
+                .await;
+                assert_eq!(
+                    store.supports_atomic_create_only_write_back(),
+                    supported,
+                    "layout={layout:?}, locking={locking}"
+                );
+                shutdown.cancel();
+            })
+            .await;
+        }
+    }
+
     #[cfg(feature = "test-util")]
     #[tokio::test]
     #[serial_test::serial(storage_class_env)]

crates/ecstore/src/store/mod.rs

@@ -425,7 +425,7 @@ pub(crate) mod init_format;
 pub(crate) mod list_objects;
 mod multipart;
 mod object;
-#[cfg(any(test, feature = "test-util"))]
+#[cfg(feature = "test-util")]
 pub use object::DeleteAfterObjectLockSnapshotBarrier;
 pub(crate) use object::{
     DecommissionFixedReadAnchor, ObjectLockDiagGuard, RemoteTuplePublicationCommitGuard, RemoteTuplePublicationFence,
@@ -848,7 +848,7 @@ impl ECStore {
     }
 
     pub fn scanner_namespace_mutation_generation(&self) -> u64 {
-        list_objects::scanner_namespace_mutation_generation()
+        list_objects::scanner_namespace_mutation_generation().saturating_add(self.ctx.namespace_commit_generation())
     }
 
     pub async fn scanner_data_movement_active(&self) -> bool {
@@ -857,7 +857,7 @@ impl ECStore {
     }
 
     /// Return the storage-owned movement state and generation as one
-    /// authenticated activity snapshot.  The read lock is acquired before
+    /// authenticated activity snapshot. The read lock is acquired before
     /// the state locks (cancelers, pool metadata, then rebalance metadata),
     /// matching the transition writer order and preventing a terminal state
     /// from being reported with the preceding generation.
@@ -886,11 +886,12 @@ impl ECStore {
     /// Returns whether scanner metadata may still be hidden by a local
     /// data-movement state. Terminal failed/canceled decommission entries
     /// remain suspended until an operator clears or retries them, so they are
-    /// a publication barrier even after the worker has stopped.
+    /// a publication barrier even after the worker has stopped. Active PUT
+    /// rename fanouts also defer publication, including post-ACK tails.
     pub async fn scanner_data_usage_publication_blocked(&self) -> bool {
         let operation_gate = self.ctx.data_movement_operation_gate();
         let _operation_guard = operation_gate.read_owned().await;
-        self.scanner_data_usage_publication_snapshot_blocked().await
+        self.scanner_data_usage_publication_snapshot_blocked().await || self.ctx.namespace_commits_pending()
     }
 
     pub async fn scanner_data_movement_pause_status(&self) -> ScannerDataMovementPauseStatus {
@@ -1070,7 +1071,7 @@ impl ECStore {
         {
             return Err(Error::other("scanner publication lease generation is stale"));
         }
-        if self.scanner_data_movement_snapshot_locked().await.1 {
+        if self.scanner_data_movement_snapshot_locked().await.1 || self.ctx.namespace_commits_pending() {
             return Err(Error::other("scanner publication lease is blocked by data movement"));
         }
 
@@ -1109,7 +1110,7 @@ impl ECStore {
         {
             return Err(Error::other("scanner publication lease generation is stale"));
         }
-        if self.scanner_data_movement_snapshot_locked().await.1 {
+        if self.scanner_data_movement_snapshot_locked().await.1 || self.ctx.namespace_commits_pending() {
             return Err(Error::other("scanner publication lease is blocked by data movement"));
         }
         if !self.ctx.scanner_publication_lease_is_active(token).await {
@@ -1129,7 +1130,7 @@ impl ECStore {
         if self.ctx.data_movement_generation_exhausted() || self.ctx.data_movement_operation_epoch_exhausted() {
             return Err(Error::other("scanner publication lease generation is exhausted"));
         }
-        if self.scanner_data_movement_snapshot_locked().await.1 {
+        if self.scanner_data_movement_snapshot_locked().await.1 || self.ctx.namespace_commits_pending() {
             return Err(Error::other("scanner publication lease is blocked by data movement"));
         }
         let Some(lease_generation) = self.ctx.scanner_publication_lease_generation(token).await else {

crates/filemeta/src/filemeta/version.rs

@@ -297,6 +297,20 @@ fn transitioned_version_from_bytes(value: Option<&[u8]>, state: TransitionVersio
     }
 }
 
+fn transition_version_metadata_value(raw: &[u8], decoded: Option<&str>) -> String {
+    decoded.map(str::to_owned).unwrap_or_else(|| {
+        if raw.is_empty() {
+            String::new()
+        } else {
+            String::from_utf8_lossy(raw).into_owned()
+        }
+    })
+}
+
+fn is_transition_version_metadata_key(key: &str) -> bool {
+    strip_internal_prefix_preserving_case(key).is_some_and(|suffix| suffix.eq_ignore_ascii_case(SUFFIX_TRANSITIONED_VERSION_ID))
+}
+
 fn validate_transition_version_state(state: TransitionVersionState, version: Option<&str>) -> Result<()> {
     let valid = match state {
         TransitionVersionState::Unknown | TransitionVersionState::KnownDisabled => version.is_none(),
@@ -366,14 +380,26 @@ impl<'a> DerivedInternalMetadata<'a> {
             }
             *slot = Some(value.as_slice());
         }
+        fn merge_consistent<'a>(canonical: Option<&'a [u8]>, legacy: Option<&'a [u8]>) -> Result<Option<&'a [u8]>> {
+            if let (Some(canonical), Some(legacy)) = (canonical, legacy)
+                && canonical != legacy
+            {
+                return Err(Error::FileCorrupt);
+            }
+            Ok(canonical.or(legacy))
+        }
+
         Ok(Self {
             checksum: canonical.checksum.or(legacy.checksum),
             part_checksums: canonical.part_checksums.or(legacy.part_checksums),
-            transition_status: canonical.transition_status.or(legacy.transition_status),
+            transition_status: merge_consistent(canonical.transition_status, legacy.transition_status)?,
-            transitioned_object: canonical.transitioned_object.or(legacy.transitioned_object),
+            transitioned_object: merge_consistent(canonical.transitioned_object, legacy.transitioned_object)?,
-            transitioned_version: canonical.transitioned_version.or(legacy.transitioned_version),
+            transitioned_version: merge_consistent(canonical.transitioned_version, legacy.transitioned_version)?,
-            transitioned_version_state: canonical.transitioned_version_state.or(legacy.transitioned_version_state),
+            transitioned_version_state: merge_consistent(
-            transition_tier: canonical.transition_tier.or(legacy.transition_tier),
+                canonical.transitioned_version_state,
+                legacy.transitioned_version_state,
+            )?,
+            transition_tier: merge_consistent(canonical.transition_tier, legacy.transition_tier)?,
         })
     }
 }
@@ -438,8 +464,14 @@ impl FileInfo {
     }
 }
 
-fn set_transition_version_state(meta_sys: &mut HashMap<String, Vec<u8>>, state: TransitionVersionState) {
+fn set_transition_version_state(
-    if state == TransitionVersionState::Unknown {
+    meta_sys: &mut HashMap<String, Vec<u8>>,
+    state: TransitionVersionState,
+    source_metadata: &HashMap<String, String>,
+) {
+    if state == TransitionVersionState::Unknown
+        && !rustfs_utils::http::metadata_compat::contains_key_str(source_metadata, SUFFIX_TRANSITIONED_VERSION_STATE)
+    {
         remove_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE);
     } else {
         insert_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE, state.as_str().as_bytes().to_vec());
@@ -2643,6 +2675,11 @@ impl MetaObject {
         if derived_metadata.transitioned_version_state.is_some() {
             validate_transition_version_state(transition_version_state, transition_version.as_deref())?;
         }
+        for (key, value) in &self.meta_sys {
+            if is_transition_version_metadata_key(key) {
+                metadata.insert(key.to_owned(), transition_version_metadata_value(value, transition_version.as_deref()));
+            }
+        }
         let transition_version_id = transition_version.as_deref().and_then(|value| Uuid::parse_str(value).ok());
         let transition_tier = derived_metadata
             .transition_tier
@@ -2689,7 +2726,7 @@ impl MetaObject {
         } else {
             remove_bytes(&mut self.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID);
         }
-        set_transition_version_state(&mut self.meta_sys, fi.transition_version_state);
+        set_transition_version_state(&mut self.meta_sys, fi.transition_version_state, &fi.metadata);
         insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITION_TIER, fi.transition_tier.as_bytes().to_vec());
         if let Some(destination_id) = get_str(&fi.metadata, SUFFIX_TRANSITION_TIER_DESTINATION_ID) {
             insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITION_TIER_DESTINATION_ID, destination_id.into_bytes());
@@ -2830,7 +2867,7 @@ impl From<FileInfo> for MetaObject {
             insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
         }
         if !value.transition_status.is_empty() {
-            set_transition_version_state(&mut meta_sys, value.transition_version_state);
+            set_transition_version_state(&mut meta_sys, value.transition_version_state, &value.metadata);
         }
 
         if !value.transition_tier.is_empty() {
@@ -2985,6 +3022,12 @@ impl MetaDeleteMarker {
             fi.transition_version_state = transition_version_state_from_bytes(derived_metadata.transitioned_version_state)?;
             fi.transition_version =
                 transitioned_version_from_bytes(derived_metadata.transitioned_version, fi.transition_version_state);
+            for (key, value) in &self.meta_sys {
+                if is_transition_version_metadata_key(key) {
+                    fi.metadata
+                        .insert(key.to_owned(), transition_version_metadata_value(value, fi.transition_version.as_deref()));
+                }
+            }
             fi.transition_version_id = fi.transition_version.as_deref().and_then(|value| Uuid::parse_str(value).ok());
             if derived_metadata.transitioned_version_state.is_some() {
                 validate_transition_version_state(fi.transition_version_state, fi.transition_version.as_deref())?;
@@ -3152,7 +3195,7 @@ impl From<FileInfo> for MetaDeleteMarker {
             insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
         }
         if !value.transition_status.is_empty() || value.tier_free_version() {
-            set_transition_version_state(&mut meta_sys, value.transition_version_state);
+            set_transition_version_state(&mut meta_sys, value.transition_version_state, &value.metadata);
         }
         if !value.transition_tier.is_empty() {
             insert_bytes(&mut meta_sys, SUFFIX_TRANSITION_TIER, value.transition_tier.as_bytes().to_vec());
@@ -4574,6 +4617,7 @@ mod tests {
             .into_fileinfo("b", "k", false)
             .expect("into_fileinfo");
         assert_eq!(fi.transition_version_id, None);
+        assert_eq!(get_str(&fi.metadata, SUFFIX_TRANSITIONED_VERSION_ID), Some(String::new()));
     }
 
     #[test]
@@ -4585,6 +4629,10 @@ mod tests {
             .into_fileinfo("b", "k", false)
             .expect("into_fileinfo");
         assert_eq!(fi.transition_version_id, None);
+        assert!(
+            get_str(&fi.metadata, SUFFIX_TRANSITIONED_VERSION_ID).is_some_and(|value| !value.is_empty()),
+            "nil UUID bytes must remain distinguishable from an empty MinIO version"
+        );
     }
 
     #[test]
@@ -4598,6 +4646,7 @@ mod tests {
         assert_eq!(fi.transition_version_id, Some(id));
         assert_eq!(fi.transition_version, Some(id.to_string()));
         assert_eq!(fi.transition_version_state, TransitionVersionState::Unknown);
+        assert_eq!(get_str(&fi.metadata, SUFFIX_TRANSITIONED_VERSION_ID), Some(id.to_string()));
     }
 
     #[test]
@@ -4637,6 +4686,36 @@ mod tests {
         assert_eq!(fi.transition_version_state, TransitionVersionState::Unknown);
     }
 
+    #[test]
+    fn meta_object_transition_version_state_explicit_unknown_is_not_legacy_missing() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut metadata,
+            SUFFIX_TRANSITIONED_VERSION_STATE,
+            TransitionVersionState::Unknown.as_str().to_string(),
+        );
+        let fi = FileInfo {
+            transition_status: "complete".to_string(),
+            transition_version_state: TransitionVersionState::Unknown,
+            metadata,
+            ..Default::default()
+        };
+
+        let object = MetaObject::from(fi);
+        assert_eq!(
+            get_consistent_bytes(&object.meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE),
+            Some(b"unknown".as_slice())
+        );
+        let decoded = object
+            .into_fileinfo("b", "k", false)
+            .expect("explicit unknown state should decode");
+        assert_eq!(decoded.transition_version_state, TransitionVersionState::Unknown);
+        assert_eq!(
+            rustfs_utils::http::metadata_compat::get_consistent_str(&decoded.metadata, SUFFIX_TRANSITIONED_VERSION_STATE,),
+            Some("unknown")
+        );
+    }
+
     #[test]
     fn meta_object_transition_version_state_exact_round_trips_dual_keys() {
         let id = sample_version_id();
@@ -4753,6 +4832,10 @@ mod tests {
             .expect("invalid transition version bytes must not fail the object read");
         assert_eq!(fi.transition_version_id, None);
         assert_eq!(fi.transition_version, None);
+        assert!(
+            get_str(&fi.metadata, SUFFIX_TRANSITIONED_VERSION_ID).is_some_and(|value| !value.is_empty()),
+            "invalid raw bytes must remain distinguishable from an empty MinIO version"
+        );
     }
 
     #[test]
@@ -4795,6 +4878,10 @@ mod tests {
         .into_fileinfo("b", "k", false)
         .expect("nil tier version should remain an absent remote version");
         assert_eq!(fi.transition_version_id, None);
+        assert!(
+            get_str(&fi.metadata, SUFFIX_TRANSITIONED_VERSION_ID).is_some_and(|value| !value.is_empty()),
+            "nil UUID bytes must remain distinguishable from an empty MinIO version"
+        );
     }
 
     #[test]
@@ -4812,6 +4899,7 @@ mod tests {
         .expect("legacy binary UUID tier version should decode");
         assert_eq!(fi.transition_version_id, Some(id));
         assert_eq!(fi.transition_version, Some(id.to_string()));
+        assert_eq!(get_str(&fi.metadata, SUFFIX_TRANSITIONED_VERSION_ID), Some(id.to_string()));
     }
 
     #[test]
@@ -4910,6 +4998,23 @@ mod tests {
         assert_eq!(err, Error::FileCorrupt);
     }
 
+    #[test]
+    fn meta_object_transition_version_state_mixed_case_alias_conflict_fails_closed() {
+        let sys = HashMap::from([
+            (
+                format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_STATE}"),
+                b"unknown".to_vec(),
+            ),
+            ("X-Minio-Internal-transitioned-version-state".to_string(), b"exact".to_vec()),
+        ]);
+
+        let err = make_meta_object_with_sys(sys)
+            .into_fileinfo("b", "k", false)
+            .expect_err("mixed-case transition state aliases must agree");
+
+        assert_eq!(err, Error::FileCorrupt);
+    }
+
     #[test]
     fn version_header_sorts_before_prefers_object_over_delete_marker_on_equal_mod_time() {
         let object = FileMetaVersionHeader {

crates/heal/src/heal/manager.rs

@@ -45,6 +45,11 @@ use tracing::{debug, error, info, warn};
 use super::{DiskError, Endpoint, HealDiskExt as _, local_disk_map_read};
 
 const KEEP_HEAL_TASK_STATUS_DURATION: Duration = Duration::from_secs(10 * 60);
+// Each cache includes alias tokens in its count and byte budget. Eviction
+// removes every token sharing a snapshot; neither cache retains repair state.
+const MAX_COMPLETED_HEAL_TOKENS: usize = 1024;
+const MAX_COMPLETED_HEAL_BYTES: usize = 64 * 1024 * 1024;
+const MAX_COMPLETED_HEAL_RESULT_BYTES: usize = 1024 * 1024;
 const DISPLACED_HEAL_REASON: &str = "reason=displaced; retry_hint=submit_again";
 const LOG_COMPONENT_HEAL: &str = "heal";
 const LOG_SUBSYSTEM_DISK_SCANNER: &str = "disk_scanner";
@@ -180,6 +185,8 @@ fn record_displaced_terminal(
     request: &HealRequest,
 ) -> Arc<CompletedHealStatus> {
     let terminal = Arc::new(CompletedHealStatus {
+        progress: None,
+        retained_bytes: std::sync::OnceLock::new(),
         heal_type: request.heal_type.clone(),
         status: HealTaskStatus::Failed {
             error: format!("heal task displaced by a higher-priority request ({DISPLACED_HEAL_REASON})"),
@@ -193,6 +200,7 @@ fn record_displaced_terminal(
     let mut terminals = lock_displaced_terminals(registry);
     prune_completed_heal_statuses(&mut terminals);
     terminals.insert(request.id.clone(), Arc::clone(&terminal));
+    prune_completed_heal_statuses(&mut terminals);
     terminal
 }
 
@@ -209,9 +217,15 @@ async fn remove_displaced_task_aliases(
         .collect::<Vec<_>>();
     let mut displaced_terminals = lock_displaced_terminals(terminals);
     prune_completed_heal_statuses(&mut displaced_terminals);
-    for alias_id in alias_ids {
+    if displaced_terminals
-        displaced_terminals.insert(alias_id, Arc::clone(terminal));
+        .get(task_id)
+        .is_some_and(|current| Arc::ptr_eq(current, terminal))
+    {
+        for alias_id in alias_ids {
+            displaced_terminals.insert(alias_id, Arc::clone(terminal));
+        }
     }
+    prune_completed_heal_statuses(&mut displaced_terminals);
     aliases.retain(|alias_id, alias| alias_id != task_id && alias.task_id != task_id);
 }
 
@@ -222,6 +236,36 @@ async fn remove_task_aliases_for_task(registry: &Arc<Mutex<HashMap<String, HealT
         .retain(|alias_id, alias| alias_id != task_id && alias.task_id != task_id);
 }
 
+// Callers hold active ownership until publication. Lock order is active ->
+// retrying (when needed) -> aliases -> completed; queries release aliases
+// before looking up active state. Publishing aliases before removing their
+// mapping keeps both an already-resolved token and a new lookup valid.
+async fn publish_completed_heal(
+    completed_heals: &Mutex<HashMap<String, Arc<CompletedHealStatus>>>,
+    task_aliases: &Mutex<HashMap<String, HealTaskAlias>>,
+    task_id: &str,
+    completed: CompletedHealStatus,
+    terminal: bool,
+) {
+    let completed = Arc::new(completed);
+    completed.retained_bytes();
+    let mut aliases = task_aliases.lock().await;
+    let mut retained = completed_heals.lock().await;
+    if let Some(previous) = retained.get(task_id).cloned() {
+        for entry in retained.values_mut().filter(|entry| Arc::ptr_eq(entry, &previous)) {
+            *entry = Arc::clone(&completed);
+        }
+    }
+    retained.insert(task_id.to_owned(), Arc::clone(&completed));
+    if terminal {
+        for (alias_id, _) in aliases.iter().filter(|(_, alias)| alias.task_id == task_id) {
+            retained.insert(alias_id.clone(), Arc::clone(&completed));
+        }
+        aliases.retain(|alias_id, alias| alias_id != task_id && alias.task_id != task_id);
+    }
+    prune_completed_heal_statuses(&mut retained);
+}
+
 #[derive(Debug, Clone)]
 pub struct HealTaskReport {
     pub status: HealTaskStatus,
@@ -268,7 +312,7 @@ fn completed_task_report(completed: &CompletedHealStatus, since: Option<u64>) ->
     let result_items = match since {
         None => completed.seqed_items.iter().map(|(_, item)| item.clone()).collect(),
         Some(cursor) => {
-            if cursor + 1 < completed.min_seq {
+            if cursor.saturating_add(1) < completed.min_seq {
                 lagged = true;
             }
             completed
@@ -283,7 +327,7 @@ fn completed_task_report(completed: &CompletedHealStatus, since: Option<u64>) ->
         status: completed.status.clone(),
         result_items,
         result_items_truncated: completed.result_items_truncated || lagged,
-        progress: None,
+        progress: completed.progress.clone(),
         next_seq: completed.next_seq,
         min_seq: completed.min_seq,
     }
@@ -1847,14 +1891,14 @@ impl HealManager {
 
     pub async fn get_task_progress(&self, task_id: &str) -> Result<HealProgress> {
         let canonical_task_id = self.canonical_task_id(task_id).await;
-        let active_heals = self.active_heals.lock().await;
+        let progress = match self.lookup_task_state(&canonical_task_id, None).await {
-        if let Some(task) = active_heals.get(&canonical_task_id) {
+            TaskStateLookup::Active(task) => Some(task.get_progress().await),
-            Ok(task.get_progress().await)
+            TaskStateLookup::Completed(completed) => completed.progress.clone(),
-        } else {
+            _ => None,
-            Err(Error::TaskNotFound {
+        };
-                task_id: task_id.to_string(),
+        progress.ok_or_else(|| Error::TaskNotFound {
-            })
+            task_id: task_id.to_string(),
-        }
+        })
     }
 
     /// Cancel task
@@ -1864,6 +1908,8 @@ impl HealManager {
             let mut active_heals = self.active_heals.lock().await;
             if let Some(task) = active_heals.get(&canonical_task_id) {
                 task.cancel().await?;
+                let completed = CompletedHealStatus::snapshot(task, HealTaskStatus::Cancelled).await;
+                publish_completed_heal(&self.completed_heals, &self.task_aliases, &canonical_task_id, completed, true).await;
                 active_heals.remove(&canonical_task_id);
                 publish_active_heal_count(&active_heals);
                 info!(
@@ -1940,6 +1986,8 @@ impl HealManager {
             for task_id in &task_ids {
                 if let Some(task) = active_heals.get(task_id) {
                     task.cancel().await?;
+                    let completed = CompletedHealStatus::snapshot(task, HealTaskStatus::Cancelled).await;
+                    publish_completed_heal(&self.completed_heals, &self.task_aliases, task_id, completed, true).await;
                 }
                 active_heals.remove(task_id);
                 cancelled += 1;

crates/heal/src/heal/manager/queue.rs

@@ -82,6 +82,8 @@ pub(super) enum QueuePushOutcome {
 pub(super) struct CompletedHealStatus {
     pub(super) heal_type: HealType,
     pub(super) status: HealTaskStatus,
+    pub(super) progress: Option<HealProgress>,
+    pub(super) retained_bytes: std::sync::OnceLock<usize>,
     pub(super) result_items_truncated: bool,
     pub(super) completed_at: SystemTime,
     /// Sequence-stamped retained window, archived with the completion so
@@ -92,6 +94,133 @@ pub(super) struct CompletedHealStatus {
     pub(super) min_seq: u64,
 }
 
+impl CompletedHealStatus {
+    // Account for owned capacities, including nested drive arrays. Aliases
+    // conservatively charge the shared allocation again, keeping both token
+    // count and retained payload bounded without a second ownership index.
+    pub(super) fn retained_bytes(&self) -> usize {
+        *self.retained_bytes.get_or_init(|| self.measure_retained_bytes())
+    }
+
+    fn measure_retained_bytes(&self) -> usize {
+        let mut bytes = size_of::<Self>();
+        let mut add = |amount: usize| bytes = bytes.saturating_add(amount);
+        match &self.heal_type {
+            HealType::Cluster => {}
+            HealType::Bucket { bucket } => add(bucket.capacity()),
+            HealType::Object {
+                bucket,
+                object,
+                version_id,
+            }
+            | HealType::ECDecode {
+                bucket,
+                object,
+                version_id,
+            } => {
+                add(bucket.capacity());
+                add(object.capacity());
+                add(version_id.as_ref().map_or(0, String::capacity));
+            }
+            HealType::Prefix { bucket, prefix } => {
+                add(bucket.capacity());
+                add(prefix.capacity());
+            }
+            HealType::Metadata { bucket, object } => {
+                add(bucket.capacity());
+                add(object.capacity());
+            }
+            HealType::ErasureSet { buckets, set_disk_id } => {
+                add(buckets.capacity().saturating_mul(size_of::<String>()));
+                for bucket in buckets {
+                    add(bucket.capacity());
+                }
+                add(set_disk_id.capacity());
+            }
+        }
+        if let HealTaskStatus::Failed { error } | HealTaskStatus::Retrying { error, .. } = &self.status {
+            add(error.capacity());
+        }
+        add(self
+            .progress
+            .as_ref()
+            .and_then(|progress| progress.current_object.as_ref())
+            .map_or(0, String::capacity));
+        add(self.seqed_items.capacity().saturating_mul(size_of::<(u64, HealResultItem)>()));
+        for (_, item) in &self.seqed_items {
+            add(Self::result_item_heap_bytes(item));
+        }
+        bytes
+    }
+
+    fn result_item_heap_bytes(item: &HealResultItem) -> usize {
+        let mut bytes = 0usize;
+        let mut add = |amount: usize| bytes = bytes.saturating_add(amount);
+        for value in [
+            &item.heal_item_type,
+            &item.bucket,
+            &item.object,
+            &item.version_id,
+            &item.detail,
+        ] {
+            add(value.capacity());
+        }
+        for infos in [&item.before, &item.after] {
+            add(infos
+                .drives
+                .capacity()
+                .saturating_mul(size_of::<rustfs_madmin::heal_commands::HealDriveInfo>()));
+            for drive in &infos.drives {
+                add(drive.uuid.capacity());
+                add(drive.endpoint.capacity());
+                add(drive.state.capacity());
+            }
+        }
+        bytes
+    }
+
+    pub(super) fn bound_result_window(&mut self) {
+        let mut bytes = 0usize;
+        let retained = self
+            .seqed_items
+            .iter()
+            .rev()
+            .take_while(|(_, item)| {
+                bytes = bytes
+                    .saturating_add(size_of::<(u64, HealResultItem)>())
+                    .saturating_add(Self::result_item_heap_bytes(item));
+                bytes <= MAX_COMPLETED_HEAL_RESULT_BYTES
+            })
+            .count();
+        let truncated = retained < self.seqed_items.len();
+        if truncated {
+            self.seqed_items.drain(..self.seqed_items.len() - retained);
+            self.seqed_items.shrink_to_fit();
+            self.min_seq = self.seqed_items.first().map_or(self.next_seq, |(seq, _)| *seq);
+            self.result_items_truncated = true;
+            self.retained_bytes.take();
+        }
+    }
+
+    pub(super) async fn snapshot(task: &HealTask, status: HealTaskStatus) -> Self {
+        let seqed_items = task.get_seqed_result_items().await;
+        let (next_seq, min_seq) = task.result_seq_cursors();
+        let mut snapshot = Self {
+            heal_type: task.heal_type.clone(),
+            status,
+            progress: Some(task.get_progress().await),
+            retained_bytes: std::sync::OnceLock::new(),
+            result_items_truncated: task.result_items_truncated(),
+            completed_at: SystemTime::now(),
+            seqed_items,
+            next_seq,
+            min_seq,
+        };
+        snapshot.bound_result_window();
+        snapshot
+    }
+}
+
 #[derive(Debug, Clone)]
 pub(super) struct HealTaskAlias {
     pub(super) task_id: String,

crates/heal/src/heal/manager/scheduler.rs

@@ -264,7 +264,7 @@ impl HealManager {
                         error: error.clone(),
                         retry_attempt: request.retry_attempts,
                     });
-                    let retry_request_for_queue = retry_request;
+                    let mut retry_request_for_queue = retry_request;
                     let retry_cancel_token = retry_request_for_queue.as_ref().map(|_| CancellationToken::new());
                     if retry_request_for_queue.is_none() {
                         replacement_recovery_anchors_clone
@@ -272,7 +272,35 @@ impl HealManager {
                             .unwrap_or_else(|poisoned| poisoned.into_inner())
                             .remove(&task_id);
                     }
+                    let mut completed_status = match retry_request_for_status {
+                        Some(status) => status,
+                        None => task.get_status().await,
+                    };
+                    let mut completed_status_entry = CompletedHealStatus::snapshot(&task, completed_status.clone()).await;
+                    let completed_progress = task.get_progress().await;
+                    #[cfg(test)]
+                    tests::pause_completed_retention_before_publish(&task_id, &completed_status).await;
                     let mut active_heals_guard = active_heals_clone.lock().await;
+                    let owns_completion = active_heals_guard.contains_key(&task_id);
+                    let cancelled_completion = if owns_completion {
+                        false
+                    } else {
+                        // Cancellation can win while a finished worker waits
+                        // for active ownership. It must not resurrect a retry
+                        // or replace an acknowledged cancellation with success.
+                        retry_request_for_queue = None;
+                        completed_heals_clone
+                            .lock()
+                            .await
+                            .get(&task_id)
+                            .is_some_and(|completed| completed.status == HealTaskStatus::Cancelled)
+                    };
+                    if cancelled_completion {
+                        completed_status = HealTaskStatus::Cancelled;
+                        completed_status_entry.status = HealTaskStatus::Cancelled;
+                    }
+                    let terminal_completion = !matches!(completed_status, HealTaskStatus::Retrying { .. });
+                    let successful_completion = matches!(completed_status, HealTaskStatus::Completed);
                     // Keep retry ownership continuous: status snapshots acquire
                     // these locks in the same active -> retrying order.
                     let mut retrying_heals_guard = if let (Some((request, _, error)), Some(cancel_token)) =
@@ -295,6 +323,16 @@ impl HealManager {
                     } else {
                         None
                     };
+                    if owns_completion || cancelled_completion {
+                        publish_completed_heal(
+                            &completed_heals_clone,
+                            &task_aliases_clone,
+                            &task_id,
+                            completed_status_entry,
+                            terminal_completion,
+                        )
+                        .await;
+                    }
                     let completed_task = active_heals_guard.remove(&task_id);
                     if let Some(completed_task) = completed_task.as_ref() {
                         publish_active_heal_count(&active_heals_guard);
@@ -304,33 +342,10 @@ impl HealManager {
                     drop(retrying_heals_guard.take());
                     drop(active_heals_guard);
 
-                    if let Some(completed_task) = completed_task {
+                    #[cfg(test)]
-                        let completed_status = if let Some(status) = retry_request_for_status {
+                    tests::pause_completed_retention_handoff(&task_id).await;
-                            status
+
-                        } else {
+                    if completed_task.is_some() {
-                            completed_task.get_status().await
-                        };
-                        let terminal_completion = !matches!(completed_status, HealTaskStatus::Retrying { .. });
-                        let successful_completion = matches!(completed_status, HealTaskStatus::Completed);
-                        let completed_progress = completed_task.get_progress().await;
-                        // Single snapshot of the retained window: the task is
-                        // finished and already off the active map, so there is
-                        // no concurrent writer to race with.
-                        let seqed_items = completed_task.get_seqed_result_items().await;
-                        let (next_seq, min_seq) = completed_task.result_seq_cursors();
-                        let completed_status_entry = CompletedHealStatus {
-                            heal_type: completed_task.heal_type.clone(),
-                            status: completed_status.clone(),
-                            result_items_truncated: completed_task.result_items_truncated(),
-                            completed_at: SystemTime::now(),
-                            seqed_items,
-                            next_seq,
-                            min_seq,
-                        };
-                        let mut completed_heals_guard = completed_heals_clone.lock().await;
-                        prune_completed_heal_statuses(&mut completed_heals_guard);
-                        completed_heals_guard.insert(task_id.clone(), Arc::new(completed_status_entry));
-                        drop(completed_heals_guard);
                         // update statistics
                         let mut stats = statistics_clone.write().await;
                         match completed_status {
@@ -352,10 +367,6 @@ impl HealManager {
                             } else {
                                 release_mrf_repair_notice_targets(notice_targets);
                             }
-                            task_aliases_clone
-                                .lock()
-                                .await
-                                .retain(|alias_id, alias| alias_id != &task_id && alias.task_id != task_id);
                         }
                     }
 
@@ -718,17 +729,42 @@ pub(super) fn heal_request_set_key_for_task(task: &HealTask) -> Option<String> {
 }
 
 pub(super) fn prune_completed_heal_statuses(completed_heals: &mut HashMap<String, Arc<CompletedHealStatus>>) {
-    let Ok(now) = SystemTime::now().duration_since(SystemTime::UNIX_EPOCH) else {
+    prune_completed_heal_statuses_at(completed_heals, SystemTime::now());
-        return;
+}
-    };
 
+pub(super) fn prune_completed_heal_statuses_at(completed_heals: &mut HashMap<String, Arc<CompletedHealStatus>>, now: SystemTime) {
     completed_heals.retain(|_, completed| {
-        completed
+        now.duration_since(completed.completed_at)
-            .completed_at
+            .map(|age| age <= KEEP_HEAL_TASK_STATUS_DURATION)
-            .duration_since(SystemTime::UNIX_EPOCH)
-            .map(|completed_at| now.saturating_sub(completed_at) <= KEEP_HEAL_TASK_STATUS_DURATION)
             .unwrap_or(false)
     });
+    let entry_bytes = |key: &String, value: &Arc<CompletedHealStatus>| {
+        key.capacity()
+            .saturating_add(size_of::<(String, Arc<CompletedHealStatus>)>())
+            .saturating_add(value.retained_bytes())
+    };
+    let mut bytes = completed_heals
+        .iter()
+        .fold(0usize, |total, (key, value)| total.saturating_add(entry_bytes(key, value)));
+    while completed_heals.len() > MAX_COMPLETED_HEAL_TOKENS || bytes > MAX_COMPLETED_HEAL_BYTES {
+        let Some(oldest) = completed_heals
+            .iter()
+            .min_by(|(left_id, left), (right_id, right)| {
+                left.completed_at.cmp(&right.completed_at).then_with(|| left_id.cmp(right_id))
+            })
+            .map(|(_, value)| Arc::clone(value))
+        else {
+            break;
+        };
+        completed_heals.retain(|key, value| {
+            if Arc::ptr_eq(value, &oldest) {
+                bytes = bytes.saturating_sub(entry_bytes(key, value));
+                false
+            } else {
+                true
+            }
+        });
+    }
 }
 
 pub(super) fn can_schedule_request(

crates/heal/src/heal/manager/tests.rs

@@ -101,6 +101,326 @@ async fn process_manager_queue_once(manager: &HealManager) {
 
 struct MockStorage;
 
+fn completed_retention_fixture(completed_at: SystemTime) -> CompletedHealStatus {
+    CompletedHealStatus {
+        heal_type: HealType::Cluster,
+        status: HealTaskStatus::Completed,
+        progress: Some(HealProgress {
+            objects_scanned: 9,
+            objects_healed: 8,
+            objects_failed: 1,
+            ..Default::default()
+        }),
+        retained_bytes: std::sync::OnceLock::new(),
+        result_items_truncated: false,
+        completed_at,
+        seqed_items: vec![(3, HealResultItem::default()), (4, HealResultItem::default())],
+        next_seq: 5,
+        min_seq: 3,
+    }
+}
+
+#[test]
+fn completed_retention_cursor_boundaries_preserve_progress() {
+    let completed = completed_retention_fixture(SystemTime::now());
+    for (cursor, count, lagged) in [
+        (0, 2, true),
+        (1, 2, true),
+        (2, 2, false),
+        (3, 1, false),
+        (4, 0, false),
+        (5, 0, false),
+        (u64::MAX, 0, false),
+    ] {
+        let report = completed_task_report(&completed, Some(cursor));
+        assert_eq!(report.result_items.len(), count, "cursor={cursor}");
+        assert_eq!(report.result_items_truncated, lagged, "cursor={cursor}");
+        assert_eq!(report.progress, completed.progress);
+        assert_eq!((report.next_seq, report.min_seq), (5, 3));
+    }
+    assert_eq!(completed_task_report(&completed, None).result_items.len(), 2);
+}
+
+#[tokio::test]
+async fn completed_retention_displaced_alias_does_not_resurrect_evicted_snapshot() {
+    let manager = HealManager::new(Arc::new(MockStorage), None);
+    let request = HealRequest::bucket("bucket".to_string());
+    manager.insert_task_alias("alias", &request.id).await;
+    let terminal = record_displaced_terminal(&manager.displaced_terminals, &request);
+    lock_displaced_terminals(&manager.displaced_terminals).remove(&request.id);
+    remove_displaced_task_aliases(&manager.task_aliases, &manager.displaced_terminals, &request.id, &terminal).await;
+    for token in [&request.id, &"alias".to_string()] {
+        assert!(matches!(manager.get_task_report(token).await, Err(Error::TaskNotFound { .. })));
+    }
+    assert!(manager.task_aliases.lock().await.is_empty());
+    assert!(lock_displaced_terminals(&manager.displaced_terminals).is_empty());
+}
+
+#[test]
+fn completed_retention_count_ttl_and_alias_eviction_are_bounded() {
+    let now = SystemTime::now();
+    let mut entries = HashMap::new();
+    let oldest = Arc::new(completed_retention_fixture(now - KEEP_HEAL_TASK_STATUS_DURATION));
+    entries.insert("oldest".to_string(), Arc::clone(&oldest));
+    entries.insert("oldest-alias".to_string(), Arc::clone(&oldest));
+    for index in 2..MAX_COMPLETED_HEAL_TOKENS {
+        entries.insert(format!("task-{index}"), Arc::new(completed_retention_fixture(now)));
+    }
+    prune_completed_heal_statuses_at(&mut entries, now);
+    assert_eq!(entries.len(), MAX_COMPLETED_HEAL_TOKENS);
+    entries.insert("cap-plus-one".to_string(), Arc::new(completed_retention_fixture(now)));
+    prune_completed_heal_statuses_at(&mut entries, now);
+    assert_eq!(entries.len(), MAX_COMPLETED_HEAL_TOKENS - 1);
+    assert!(!entries.contains_key("oldest"));
+    assert!(!entries.contains_key("oldest-alias"));
+    entries.clear();
+    entries.insert("ttl-boundary".to_string(), oldest);
+    entries.insert(
+        "expired".to_string(),
+        Arc::new(completed_retention_fixture(
+            now - KEEP_HEAL_TASK_STATUS_DURATION - Duration::from_nanos(1),
+        )),
+    );
+    entries.insert("future".to_string(), Arc::new(completed_retention_fixture(now + Duration::from_nanos(1))));
+    prune_completed_heal_statuses_at(&mut entries, now);
+    assert_eq!(entries.len(), 1);
+    assert!(entries.contains_key("ttl-boundary"));
+    prune_completed_heal_statuses_at(&mut entries, now + Duration::from_nanos(1));
+    assert!(entries.is_empty());
+}
+
+#[test]
+fn completed_retention_total_byte_cap_and_cap_plus_one() {
+    let now = SystemTime::now();
+    let key = "large".to_string();
+    let mut entry = completed_retention_fixture(now);
+    let base_bytes = entry.retained_bytes() + key.capacity() + size_of::<(String, Arc<CompletedHealStatus>)>();
+    entry.retained_bytes.take();
+    entry.status = HealTaskStatus::Failed {
+        error: "x".repeat(MAX_COMPLETED_HEAL_BYTES - base_bytes),
+    };
+    assert_eq!(
+        entry.retained_bytes() + key.capacity() + size_of::<(String, Arc<CompletedHealStatus>)>(),
+        MAX_COMPLETED_HEAL_BYTES
+    );
+    let mut entries = HashMap::from([(key, Arc::new(entry))]);
+    prune_completed_heal_statuses_at(&mut entries, now);
+    assert_eq!(entries.len(), 1, "exact byte cap remains retained");
+    let mut over = Arc::try_unwrap(entries.remove("large").expect("entry retained")).expect("entry not shared");
+    over.retained_bytes.take();
+    if let HealTaskStatus::Failed { error } = &mut over.status {
+        *error = "x".repeat(error.len() + 1);
+    }
+    entries.insert("large".to_string(), Arc::new(over));
+    prune_completed_heal_statuses_at(&mut entries, now);
+    assert!(entries.is_empty(), "oversized metadata cannot escape total byte bound");
+}
+
+#[tokio::test]
+async fn completed_retention_large_window_keeps_cursors_and_progress() {
+    let task = HealTask::from_request(HealRequest::bucket("bucket".to_string()), Arc::new(MockStorage));
+    let mut snapshot = completed_retention_fixture(SystemTime::now());
+    snapshot.seqed_items[0].1.detail = "x".repeat(MAX_COMPLETED_HEAL_RESULT_BYTES);
+    snapshot.bound_result_window();
+    assert_eq!(snapshot.seqed_items.len(), 1);
+    assert_eq!((snapshot.min_seq, snapshot.next_seq), (4, 5));
+    assert!(snapshot.result_items_truncated);
+    assert!(snapshot.retained_bytes() < MAX_COMPLETED_HEAL_RESULT_BYTES);
+    let report = completed_task_report(&snapshot, Some(0));
+    assert_eq!(report.progress.expect("progress retained").objects_scanned, 9);
+    assert!(report.result_items_truncated);
+    let active_max = task.get_result_items_since(Some(u64::MAX)).await;
+    assert!(active_max.items.is_empty());
+    assert!(!active_max.lagged);
+}
+
+#[test]
+fn completed_retention_result_byte_cap_and_cap_plus_one() {
+    for extra in [0, 1] {
+        let mut snapshot = completed_retention_fixture(SystemTime::now());
+        snapshot.seqed_items = vec![(
+            4,
+            HealResultItem {
+                detail: "x".repeat(MAX_COMPLETED_HEAL_RESULT_BYTES - size_of::<(u64, HealResultItem)>() + extra),
+                ..Default::default()
+            },
+        )];
+        snapshot.min_seq = 4;
+        snapshot.bound_result_window();
+        assert_eq!(snapshot.seqed_items.len(), 1 - extra);
+        assert_eq!(snapshot.result_items_truncated, extra == 1);
+        assert_eq!(snapshot.min_seq, if extra == 0 { 4 } else { 5 });
+        assert_eq!(snapshot.next_seq, 5);
+        assert_eq!(snapshot.progress.as_ref().expect("progress retained").objects_scanned, 9);
+    }
+}
+
+#[derive(Default)]
+struct CompletedRetentionHook {
+    started: Notify,
+    execute: Notify,
+    handoff: Notify,
+    finish: Notify,
+    pause_before_publish: bool,
+    before_publish: Notify,
+    publish: Notify,
+    prepared_status: Mutex<Option<HealTaskStatus>>,
+}
+
+static COMPLETED_RETENTION_HOOKS: LazyLock<Mutex<HashMap<String, Arc<CompletedRetentionHook>>>> =
+    LazyLock::new(|| Mutex::new(HashMap::new()));
+
+pub(super) async fn pause_completed_retention_handoff(task_id: &str) {
+    let hook = COMPLETED_RETENTION_HOOKS.lock().await.get(task_id).cloned();
+    if let Some(hook) = hook {
+        hook.handoff.notify_one();
+        hook.finish.notified().await;
+    }
+}
+
+pub(super) async fn pause_completed_retention_before_publish(task_id: &str, status: &HealTaskStatus) {
+    let hook = COMPLETED_RETENTION_HOOKS.lock().await.get(task_id).cloned();
+    if let Some(hook) = hook.filter(|hook| hook.pause_before_publish) {
+        *hook.prepared_status.lock().await = Some(status.clone());
+        hook.before_publish.notify_one();
+        hook.publish.notified().await;
+    }
+}
+
+#[tokio::test]
+async fn completed_retention_cancel_wins_over_a_prepared_retry_snapshot() {
+    let bucket = "completed-retention-retry-cancel";
+    let manager = HealManager::new(Arc::new(MockStorage), None);
+    let request = HealRequest::object(bucket.to_string(), "object".to_string(), None);
+    let task_id = request.id.clone();
+    let duplicate = HealRequest::object(bucket.to_string(), "object".to_string(), None);
+    let alias = duplicate.id.clone();
+    let hook = Arc::new(CompletedRetentionHook {
+        pause_before_publish: true,
+        ..Default::default()
+    });
+    {
+        let mut hooks = COMPLETED_RETENTION_HOOKS.lock().await;
+        hooks.insert(bucket.to_string(), Arc::clone(&hook));
+        hooks.insert(task_id.clone(), Arc::clone(&hook));
+    }
+    manager.submit_heal_request(request).await.expect("admit original");
+    manager.submit_heal_request(duplicate).await.expect("admit alias");
+    process_manager_queue_once(&manager).await;
+    tokio::time::timeout(Duration::from_secs(5), hook.started.notified())
+        .await
+        .expect("scheduler starts");
+    let task = manager.active_heals.lock().await.get(&task_id).cloned().expect("active task");
+    task.progress.write().await.update_object_progress(1, 1, 0, 0, 4096);
+    hook.execute.notify_one();
+    tokio::time::timeout(Duration::from_secs(5), hook.before_publish.notified())
+        .await
+        .expect("retry snapshot prepared");
+    manager.cancel_task(&alias).await.expect("cancel wins active ownership");
+    assert!(matches!(*hook.prepared_status.lock().await, Some(HealTaskStatus::Retrying { .. })));
+    hook.publish.notify_one();
+    tokio::time::timeout(Duration::from_secs(5), hook.handoff.notified())
+        .await
+        .expect("scheduler finishes handoff");
+    for token in [&task_id, &alias] {
+        let report = manager.get_task_report(token).await.expect("cancelled token retained");
+        assert_eq!(report.status, HealTaskStatus::Cancelled);
+        assert_eq!(report.progress.expect("frozen progress").objects_scanned, 1);
+    }
+    assert!(!manager.retrying_heals.lock().await.contains_key(&task_id));
+    assert!(!manager.heal_queue.lock().await.contains_request_id(&task_id));
+    hook.finish.notify_one();
+    COMPLETED_RETENTION_HOOKS
+        .lock()
+        .await
+        .retain(|key, _| key != bucket && key != &task_id);
+}
+
+#[tokio::test]
+async fn completed_retention_scheduler_preserves_progress_aliases_and_atomic_handoff() {
+    for outcome in ["success", "failed", "cancelled"] {
+        let bucket = format!("completed-retention-{outcome}");
+        let hook = Arc::new(CompletedRetentionHook::default());
+        let manager = Arc::new(HealManager::new(Arc::new(MockStorage), None));
+        let request = HealRequest::object(bucket.clone(), "object".to_string(), None);
+        let task_id = request.id.clone();
+        let duplicate = HealRequest::object(bucket.clone(), "object".to_string(), None);
+        let alias = duplicate.id.clone();
+        {
+            let mut hooks = COMPLETED_RETENTION_HOOKS.lock().await;
+            hooks.insert(bucket.clone(), Arc::clone(&hook));
+            hooks.insert(task_id.clone(), Arc::clone(&hook));
+        }
+        manager.submit_heal_request(request).await.expect("admit original");
+        manager.submit_heal_request(duplicate).await.expect("admit alias");
+        process_manager_queue_once(&manager).await;
+        tokio::time::timeout(Duration::from_secs(5), hook.started.notified())
+            .await
+            .expect("scheduler reaches storage");
+        let task = manager
+            .active_heals
+            .lock()
+            .await
+            .get(&task_id)
+            .cloned()
+            .expect("task is active");
+        task.progress.write().await.update_object_progress(1, 1, 0, 0, 4096);
+        let before = manager.get_task_report(&alias).await.expect("alias resolves active progress");
+        assert_eq!(before.progress.as_ref().expect("active progress").objects_scanned, 1);
+        let poll_manager = Arc::clone(&manager);
+        let poll_alias = alias.clone();
+        let stop = CancellationToken::new();
+        let poll_stop = stop.clone();
+        let polling = tokio::spawn(async move {
+            while !poll_stop.is_cancelled() {
+                let report = poll_manager
+                    .get_task_report(&poll_alias)
+                    .await
+                    .expect("handoff must never return NotFound");
+                assert!(report.progress.expect("progress never disappears").objects_scanned >= 1);
+                tokio::task::yield_now().await;
+            }
+        });
+        if outcome == "cancelled" {
+            manager.cancel_task(&alias).await.expect("cancel active task by alias");
+        } else {
+            hook.execute.notify_one();
+        }
+        tokio::time::timeout(Duration::from_secs(5), hook.handoff.notified())
+            .await
+            .expect("scheduler archives terminal");
+        assert!(!manager.active_heals.lock().await.contains_key(&task_id));
+        let expected = task.get_progress().await;
+        for token in [&task_id, &alias] {
+            assert_eq!(manager.get_task_progress(token).await.expect("terminal progress query"), expected);
+            let report = manager
+                .get_task_report_for_path_since(&format!("{bucket}/object"), token, Some(u64::MAX))
+                .await
+                .expect("terminal token remains queryable at handoff");
+            assert_eq!(report.progress.as_ref(), Some(&expected));
+            assert!(report.result_items.is_empty());
+            match outcome {
+                "success" => assert_eq!(report.status, HealTaskStatus::Completed),
+                "failed" => assert!(matches!(report.status, HealTaskStatus::Failed { .. })),
+                _ => assert_eq!(report.status, HealTaskStatus::Cancelled),
+            }
+        }
+        let retained = manager.completed_heals.lock().await;
+        assert!(Arc::ptr_eq(&retained[&task_id], &retained[&alias]));
+        drop(retained);
+        stop.cancel();
+        polling.await.expect("concurrent polling succeeds");
+        // Archived progress must not alias a mutable live progress object.
+        task.progress.write().await.objects_scanned = 999;
+        assert_eq!(manager.get_task_report(&alias).await.expect("frozen report").progress, Some(expected));
+        hook.finish.notify_one();
+        COMPLETED_RETENTION_HOOKS
+            .lock()
+            .await
+            .retain(|key, _| key != &bucket && key != &task_id);
+    }
+}
+
 #[async_trait::async_trait]
 impl HealStorageAPI for MockStorage {
     async fn get_object_meta(&self, _bucket: &str, _object: &str) -> Result<Option<HealObjectInfo>> {
@@ -123,6 +443,12 @@ impl HealStorageAPI for MockStorage {
     }
 
     async fn object_exists(&self, bucket: &str, _object: &str) -> Result<bool> {
+        let hook = COMPLETED_RETENTION_HOOKS.lock().await.get(bucket).cloned();
+        if let Some(hook) = hook {
+            hook.started.notify_one();
+            hook.execute.notified().await;
+            return Ok(true);
+        }
         Ok(bucket == "retry-transition")
     }
 
@@ -133,13 +459,18 @@ impl HealStorageAPI for MockStorage {
         _version_id: Option<&str>,
         _opts: &HealOpts,
     ) -> Result<(HealResultItem, Option<Error>)> {
+        if bucket == "completed-retention-failed" {
+            return Err(Error::TaskExecutionFailed {
+                message: "retention fixture failure".to_string(),
+            });
+        }
         if let Some(hook) = manager_recovery_test_hook() {
             *hook
                 .heal_object_calls
                 .lock()
                 .expect("manager recovery object call lock should not poison") += 1;
         }
-        if bucket == "retry-transition" {
+        if matches!(bucket, "retry-transition" | "completed-retention-retry-cancel") {
             return Ok((
                 HealResultItem::default(),
                 Some(Error::Storage(EcstoreError::InsufficientReadQuorum(
@@ -1145,7 +1476,13 @@ async fn test_active_duplicate_token_can_query_and_cancel_original_task() {
         .expect("duplicate token should cancel merged active task");
 
     assert!(manager.active_heals.lock().await.get(&active_task_id).is_none());
-    assert!(matches!(manager.get_task_status(&active_task_id).await, Err(Error::TaskNotFound { .. })));
+    assert_eq!(
+        manager
+            .get_task_status(&active_task_id)
+            .await
+            .expect("cancelled task remains queryable"),
+        HealTaskStatus::Cancelled
+    );
 }
 
 #[tokio::test]
@@ -1638,6 +1975,8 @@ async fn insert_retrying_request(manager: &HealManager, request: HealRequest) ->
     manager.completed_heals.lock().await.insert(
         task_id,
         Arc::new(CompletedHealStatus {
+            progress: None,
+            retained_bytes: std::sync::OnceLock::new(),
             heal_type: request.heal_type,
             status: HealTaskStatus::Retrying {
                 error: "Lock acquisition timeout".to_string(),
@@ -2053,7 +2392,7 @@ async fn admin_force_start_cancels_overlapping_active_task_first() {
         "the overlapping admin task must be cancelled (removed from the active table) before the new one starts"
     );
     assert!(
-        matches!(manager.get_task_status(&old_id).await, Err(Error::TaskNotFound { .. })),
+        matches!(manager.get_task_status(&old_id).await, Ok(HealTaskStatus::Cancelled)),
         "a cancelled task must no longer resolve as an active heal"
     );
 }
@@ -2360,6 +2699,8 @@ async fn test_retrying_completion_outranks_the_queue_for_the_same_id() {
     manager.completed_heals.lock().await.insert(
         task_id.clone(),
         Arc::new(CompletedHealStatus {
+            progress: None,
+            retained_bytes: std::sync::OnceLock::new(),
             heal_type: request.heal_type.clone(),
             status: HealTaskStatus::Retrying {
                 error: "transient disk failure".to_string(),
@@ -2395,6 +2736,8 @@ async fn test_get_task_status_reads_recent_completed_status() {
     manager.completed_heals.lock().await.insert(
         "completed-token".to_string(),
         Arc::new(CompletedHealStatus {
+            progress: None,
+            retained_bytes: std::sync::OnceLock::new(),
             heal_type: HealType::Bucket {
                 bucket: "bucket".to_string(),
             },
@@ -2424,6 +2767,8 @@ async fn test_get_task_report_for_path_reads_completed_items() {
     manager.completed_heals.lock().await.insert(
         "completed-token".to_string(),
         Arc::new(CompletedHealStatus {
+            progress: None,
+            retained_bytes: std::sync::OnceLock::new(),
             heal_type: HealType::Object {
                 bucket: "bucket".to_string(),
                 object: "object".to_string(),

crates/heal/src/heal/mrf_queue.rs

@@ -45,6 +45,10 @@ use uuid::Uuid;
 
 use crate::heal::task::{HealOptions, HealPriority, HealRequest, HealType};
 
+/// Read-only inspection of committed MRF checkpoints. The legacy consumer
+/// remains unchanged until ownership-aware replay is deployed.
+pub mod snapshot;
+
 /// Journal location inside the metadata bucket, following the resume-state
 /// layout.
 pub(crate) const MRF_JOURNAL_PATH: &str = "buckets/.heal/mrf/journal.bin";

crates/heal/src/heal/mrf_queue/snapshot.rs

@@ -0,0 +1,681 @@
+// 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.
+
+//! Reader-first support for owner-local MRF checkpoints.
+//!
+//! Each of two slots has a payload and a commit manifest. The manifest binds
+//! the writer identity, persistent sequence, length and whole-payload digest.
+//! Replacing the inactive slot must leave the previous committed slot intact.
+//! Production publication and reclamation are deliberately not enabled here.
+//! An unreadable commit path cannot prove that only legacy data exists. This
+//! explicit inspection API fails closed and never mutates recovery anchors.
+//! It is not wired into the legacy consumer: that transition requires the
+//! ownership-aware replay and producer handoff before writer activation.
+//! One surviving committed replica supports process restart recovery only;
+//! this reader does not establish a replication quorum or a power-loss policy.
+
+use super::{MRF_JOURNAL_PATH, MRF_SCOPED_JOURNAL_PATH, decode_journal};
+use crate::heal::RUSTFS_META_BUCKET;
+use crate::heal::storage_api::owner::{EcstoreDiskAPI, EcstoreDiskError, EcstoreDiskStore};
+use sha2::{Digest, Sha256};
+use std::collections::HashMap;
+use tokio::io::AsyncReadExt;
+use uuid::Uuid;
+
+// Root-level control files avoid requiring a new directory before the first
+// atomic commit. They remain inside the storage owner's metadata volume.
+const PAYLOAD_PATHS: [&str; 2] = [".heal-mrf-snapshot.0.bin", ".heal-mrf-snapshot.1.bin"];
+const MANIFEST_PATHS: [&str; 2] = [".heal-mrf-commit.0.bin", ".heal-mrf-commit.1.bin"];
+const MAGIC: &[u8; 8] = b"RFMRFC01";
+const MANIFEST_LEN: usize = 8 + 1 + 16 + 8 + 8 + 32 + 32;
+const VERSION: u8 = 1;
+
+#[derive(Debug, thiserror::Error)]
+pub enum SnapshotError {
+    #[error("MRF checkpoint has an invalid or incomplete commit record")]
+    Corrupt,
+    #[error("MRF checkpoint format is unsupported")]
+    Unsupported,
+    #[error("MRF checkpoint exceeds the configured byte limit")]
+    TooLarge,
+    #[error("MRF checkpoint replicas disagree at the same sequence")]
+    Conflict,
+    #[error("MRF checkpoint storage is unavailable")]
+    Disk(#[source] EcstoreDiskError),
+    #[error("MRF checkpoint body could not be read")]
+    Read(#[source] std::io::Error),
+}
+
+#[derive(Debug, PartialEq, Eq)]
+struct Manifest {
+    owner: Uuid,
+    sequence: u64,
+    payload_len: usize,
+    payload_digest: [u8; 32],
+}
+
+impl Manifest {
+    fn decode(bytes: &[u8], limit: usize) -> Result<Self, SnapshotError> {
+        if bytes.len() != MANIFEST_LEN || &bytes[..8] != MAGIC {
+            return Err(SnapshotError::Corrupt);
+        }
+        if bytes[8] != VERSION {
+            return Err(SnapshotError::Unsupported);
+        }
+        let signed = MANIFEST_LEN - 32;
+        let checksum: [u8; 32] = Sha256::digest(&bytes[..signed]).into();
+        if checksum != bytes[signed..] {
+            return Err(SnapshotError::Corrupt);
+        }
+        let owner = Uuid::from_slice(&bytes[9..25]).map_err(|_| SnapshotError::Corrupt)?;
+        let sequence = u64::from_le_bytes(bytes[25..33].try_into().map_err(|_| SnapshotError::Corrupt)?);
+        let payload_len = u64::from_le_bytes(bytes[33..41].try_into().map_err(|_| SnapshotError::Corrupt)?);
+        let payload_len = usize::try_from(payload_len).map_err(|_| SnapshotError::TooLarge)?;
+        if owner.is_nil() || sequence == 0 || sequence == u64::MAX {
+            return Err(SnapshotError::Corrupt);
+        }
+        if payload_len > limit {
+            return Err(SnapshotError::TooLarge);
+        }
+        Ok(Self {
+            owner,
+            sequence,
+            payload_len,
+            payload_digest: bytes[41..73].try_into().map_err(|_| SnapshotError::Corrupt)?,
+        })
+    }
+}
+
+#[derive(Debug)]
+pub struct CommittedSnapshot {
+    manifest: Manifest,
+    payload: Vec<u8>,
+}
+
+impl CommittedSnapshot {
+    /// Persistent single-writer sequence, not a process UUID ordering.
+    pub fn sequence(&self) -> u64 {
+        self.manifest.sequence
+    }
+
+    /// Identity recorded by the committed checkpoint's writer.
+    pub fn owner(&self) -> Uuid {
+        self.manifest.owner
+    }
+
+    /// Complete, checksum-validated record bytes. Inspection does not consume
+    /// these records or acknowledge completion to any producer.
+    pub fn payload(&self) -> &[u8] {
+        &self.payload
+    }
+
+    fn decode(manifest: &[u8], payload: Vec<u8>, limit: usize) -> Result<Self, SnapshotError> {
+        let manifest = Manifest::decode(manifest, limit)?;
+        let checksum: [u8; 32] = Sha256::digest(&payload).into();
+        if payload.len() != manifest.payload_len || checksum != manifest.payload_digest {
+            return Err(SnapshotError::Corrupt);
+        }
+        if decode_journal(&payload).1 != 0 {
+            return Err(SnapshotError::Corrupt);
+        }
+        Ok(Self { manifest, payload })
+    }
+}
+
+#[derive(Debug)]
+pub enum RecoverySnapshot {
+    /// An intact legacy snapshot, without a comparable commit sequence.
+    Legacy(Vec<u8>),
+    /// A committed checkpoint requiring ownership-aware replay before use.
+    Committed(CommittedSnapshot),
+}
+
+async fn read_bounded(disk: &EcstoreDiskStore, path: &str, limit: usize) -> Result<Option<Vec<u8>>, SnapshotError> {
+    let reader = match EcstoreDiskAPI::read_file(disk.as_ref(), RUSTFS_META_BUCKET, path).await {
+        Ok(reader) => reader,
+        Err(EcstoreDiskError::FileNotFound | EcstoreDiskError::VolumeNotFound) => return Ok(None),
+        Err(error) => return Err(SnapshotError::Disk(error)),
+    };
+    let maximum = limit.checked_add(1).ok_or(SnapshotError::TooLarge)?;
+    let maximum = u64::try_from(maximum).map_err(|_| SnapshotError::TooLarge)?;
+    let mut bytes = Vec::new();
+    reader
+        .take(maximum)
+        .read_to_end(&mut bytes)
+        .await
+        .map_err(SnapshotError::Read)?;
+    if bytes.len() > limit {
+        return Err(SnapshotError::TooLarge);
+    }
+    Ok(Some(bytes))
+}
+
+fn select_snapshot(selected: &mut Option<CommittedSnapshot>, candidate: CommittedSnapshot) -> Result<(), SnapshotError> {
+    if let Some(current) = selected {
+        if current.manifest.sequence == candidate.manifest.sequence
+            && (current.manifest != candidate.manifest || current.payload != candidate.payload)
+        {
+            return Err(SnapshotError::Conflict);
+        }
+        if current.manifest.sequence >= candidate.manifest.sequence {
+            return Ok(());
+        }
+    }
+    *selected = Some(candidate);
+    Ok(())
+}
+
+async fn read_committed(disks: &[EcstoreDiskStore], limit: usize) -> Result<Option<CommittedSnapshot>, SnapshotError> {
+    let mut selected = None;
+    let mut damaged = None;
+    let mut identities = HashMap::new();
+    for disk in disks {
+        for (manifest_path, payload_path) in MANIFEST_PATHS.into_iter().zip(PAYLOAD_PATHS) {
+            let candidate = async {
+                let Some(manifest) = read_bounded(disk, manifest_path, MANIFEST_LEN).await? else {
+                    return Ok(None);
+                };
+                let header = Manifest::decode(&manifest, limit)?;
+                let payload = read_bounded(disk, payload_path, header.payload_len)
+                    .await?
+                    .ok_or(SnapshotError::Corrupt)?;
+                CommittedSnapshot::decode(&manifest, payload, limit).map(Some)
+            }
+            .await;
+            match candidate {
+                Ok(Some(candidate)) => {
+                    let identity = (
+                        candidate.manifest.owner,
+                        candidate.manifest.payload_len,
+                        candidate.manifest.payload_digest,
+                    );
+                    if identities
+                        .insert(candidate.manifest.sequence, identity)
+                        .is_some_and(|previous| previous != identity)
+                    {
+                        return Err(SnapshotError::Conflict);
+                    }
+                    select_snapshot(&mut selected, candidate)?;
+                }
+                Ok(None) => {}
+                // A future committed format may supersede all readable slots.
+                Err(SnapshotError::Unsupported) => return Err(SnapshotError::Unsupported),
+                Err(error) => damaged = Some(error),
+            }
+        }
+    }
+    match (selected, damaged) {
+        (Some(snapshot), _) => Ok(Some(snapshot)),
+        (None, Some(error)) => Err(error),
+        (None, None) => Ok(None),
+    }
+}
+
+async fn read_legacy(disks: &[EcstoreDiskStore], path: &str, limit: usize) -> Result<Option<Vec<u8>>, SnapshotError> {
+    let mut selected = None;
+    let mut incomplete: Option<Vec<u8>> = None;
+    for disk in disks {
+        match read_bounded(disk, path, limit).await {
+            Ok(Some(payload)) if decode_journal(&payload).1 == 0 => {
+                if selected.as_ref().is_some_and(|current| *current != payload) {
+                    // Legacy snapshots have no sequence. There is no evidence
+                    // that the first, longest or nonempty replica is newest.
+                    return Err(SnapshotError::Conflict);
+                }
+                selected = Some(payload);
+            }
+            Ok(Some(payload)) => {
+                if let Some(previous) = &incomplete {
+                    if previous.starts_with(&payload) {
+                        continue;
+                    }
+                    if !payload.starts_with(previous) {
+                        return Err(SnapshotError::Corrupt);
+                    }
+                }
+                incomplete = Some(payload);
+            }
+            Ok(None) => {}
+            Err(error) => return Err(error),
+        }
+    }
+    if let Some(prefix) = incomplete
+        && !selected.as_ref().is_some_and(|payload| payload.starts_with(&prefix))
+    {
+        // In particular, an empty O_TRUNC replica cannot supersede another
+        // replica containing intact records followed by a torn tail.
+        return Err(SnapshotError::Corrupt);
+    }
+    Ok(selected)
+}
+
+/// Inspect local MRF checkpoints without replaying, acknowledging or deleting.
+///
+/// `max_bytes` bounds each payload read. Every local replica is examined and
+/// ambiguous identities, unavailable proof or unsupported formats return a
+/// typed error. This API must not authorize a writer without the separate
+/// ownership and mixed-version activation checks.
+pub async fn inspect_local_recovery_snapshot(max_bytes: usize) -> Result<Option<RecoverySnapshot>, SnapshotError> {
+    read_recovery_snapshot(&super::journal_disks().await, max_bytes).await
+}
+
+async fn read_recovery_snapshot(disks: &[EcstoreDiskStore], limit: usize) -> Result<Option<RecoverySnapshot>, SnapshotError> {
+    if let Some(snapshot) = read_committed(disks, limit).await? {
+        return Ok(Some(RecoverySnapshot::Committed(snapshot)));
+    }
+    // RUSTFS_COMPAT_TODO(backlog-2263): inspect retained legacy MRF journals. Remove after all supported upgrade and rollback readers understand committed snapshots and retained journals have migrated.
+    if let Some(payload) = read_legacy(disks, MRF_SCOPED_JOURNAL_PATH, limit).await? {
+        return Ok(Some(RecoverySnapshot::Legacy(payload)));
+    }
+    Ok(read_legacy(disks, MRF_JOURNAL_PATH, limit)
+        .await?
+        .map(RecoverySnapshot::Legacy))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::heal::mrf_queue::encode_intent;
+    use crate::heal::storage_api::owner::{EcstoreConditionalFileUpdate, EcstoreDiskBytes};
+    use crate::heal::{DiskOption, Endpoint, new_disk};
+    use rustfs_common::mrf_channel::{MrfIntent, MrfKind, MrfScope};
+    use std::sync::Arc;
+    use tempfile::TempDir;
+
+    fn payload(object: &str) -> Vec<u8> {
+        let intent = MrfIntent {
+            bucket: Arc::from("bucket"),
+            object: Arc::from(object),
+            version_id: None,
+            kind: MrfKind::PartialWrite,
+            scope: None,
+            lease: None,
+            enqueued_at_ms: 1234,
+            attempts: 0,
+        };
+        let mut bytes = Vec::new();
+        assert!(encode_intent(&intent, &mut bytes), "fixture must encode a full record");
+        bytes
+    }
+
+    fn manifest(owner: Uuid, sequence: u64, payload: &[u8]) -> Vec<u8> {
+        let mut bytes = Vec::with_capacity(MANIFEST_LEN);
+        bytes.extend_from_slice(MAGIC);
+        bytes.push(VERSION);
+        bytes.extend_from_slice(owner.as_bytes());
+        bytes.extend_from_slice(&sequence.to_le_bytes());
+        bytes.extend_from_slice(&u64::try_from(payload.len()).expect("fixture length fits").to_le_bytes());
+        bytes.extend_from_slice(&Sha256::digest(payload));
+        bytes.extend_from_slice(&Sha256::digest(&bytes));
+        bytes
+    }
+
+    async fn disk(root: &TempDir, name: &str) -> EcstoreDiskStore {
+        let path = root.path().join(name);
+        std::fs::create_dir_all(&path).expect("create disk directory");
+        let endpoint = Endpoint::try_from(path.to_string_lossy().as_ref()).expect("valid disk endpoint");
+        let disk = new_disk(
+            &endpoint,
+            &DiskOption {
+                cleanup: false,
+                health_check: false,
+            },
+        )
+        .await
+        .expect("open disk");
+        let result = EcstoreDiskAPI::make_volume(disk.as_ref(), RUSTFS_META_BUCKET).await;
+        assert!(
+            matches!(result, Ok(()) | Err(EcstoreDiskError::VolumeExists)),
+            "metadata volume: {result:?}"
+        );
+        disk
+    }
+
+    // Exercise the existing storage owner's atomic CAS primitive. No production
+    // caller publishes this format until ownership-aware replay is available.
+    async fn install(disk: &EcstoreDiskStore, path: &str, bytes: &[u8]) {
+        let expected = EcstoreDiskAPI::read_all(disk.as_ref(), RUSTFS_META_BUCKET, path).await.ok();
+        let result = EcstoreDiskAPI::compare_and_update_file(
+            disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            path,
+            expected,
+            Some(EcstoreDiskBytes::copy_from_slice(bytes)),
+        )
+        .await
+        .expect("atomic snapshot slot write");
+        assert_eq!(result, EcstoreConditionalFileUpdate::Updated);
+    }
+
+    async fn commit(disk: &EcstoreDiskStore, slot: usize, owner: Uuid, sequence: u64, bytes: &[u8]) {
+        install(disk, PAYLOAD_PATHS[slot], bytes).await;
+        install(disk, MANIFEST_PATHS[slot], &manifest(owner, sequence, bytes)).await;
+    }
+
+    #[test]
+    fn manifest_validates_identity_sequence_length_and_digest() {
+        let bytes = payload("object");
+        let owner = Uuid::new_v4();
+        assert!(CommittedSnapshot::decode(&manifest(owner, 1, &bytes), bytes.clone(), bytes.len()).is_ok());
+        for (owner, sequence) in [(Uuid::nil(), 1), (owner, 0), (owner, u64::MAX)] {
+            assert!(matches!(
+                Manifest::decode(&manifest(owner, sequence, &bytes), bytes.len()),
+                Err(SnapshotError::Corrupt)
+            ));
+        }
+        assert!(matches!(
+            Manifest::decode(&manifest(owner, 1, &bytes), bytes.len() - 1),
+            Err(SnapshotError::TooLarge)
+        ));
+        let mut corrupt = manifest(owner, 1, &bytes);
+        corrupt[25] ^= 1;
+        assert!(matches!(Manifest::decode(&corrupt, bytes.len()), Err(SnapshotError::Corrupt)));
+        let mut unsupported = manifest(owner, 1, &bytes);
+        unsupported[8] = 2;
+        assert!(matches!(Manifest::decode(&unsupported, bytes.len()), Err(SnapshotError::Unsupported)));
+    }
+
+    #[test]
+    fn whole_payload_integrity_is_required_even_with_a_valid_manifest() {
+        let bytes = payload("object");
+        let owner = Uuid::new_v4();
+        let header = manifest(owner, 1, &bytes);
+        assert!(matches!(
+            CommittedSnapshot::decode(&header, bytes[..bytes.len() - 1].to_vec(), bytes.len()),
+            Err(SnapshotError::Corrupt)
+        ));
+        let invalid = b"not an MRF record".to_vec();
+        assert!(matches!(
+            CommittedSnapshot::decode(&manifest(owner, 2, &invalid), invalid, bytes.len()),
+            Err(SnapshotError::Corrupt)
+        ));
+    }
+
+    #[tokio::test]
+    async fn newest_complete_replica_wins_in_both_disk_orders() {
+        let root = TempDir::new().expect("test directory");
+        let first = disk(&root, "first").await;
+        let second = disk(&root, "second").await;
+        let owner = Uuid::new_v4();
+        commit(&first, 0, owner, 1, &payload("old")).await;
+        commit(&second, 1, owner, 2, &payload("new")).await;
+        for disks in [vec![first.clone(), second.clone()], vec![second.clone(), first.clone()]] {
+            let recovered = read_committed(&disks, 4096)
+                .await
+                .expect("read replicas")
+                .expect("committed snapshot");
+            assert_eq!(recovered.manifest.sequence, 2);
+            assert_eq!(recovered.payload, payload("new"));
+        }
+    }
+
+    #[tokio::test]
+    async fn divergent_commits_at_same_sequence_fail_closed() {
+        let root = TempDir::new().expect("test directory");
+        let first = disk(&root, "first").await;
+        let second = disk(&root, "second").await;
+        let owner = Uuid::new_v4();
+        commit(&first, 0, owner, 7, &payload("a")).await;
+        commit(&second, 1, owner, 7, &payload("b")).await;
+        assert!(matches!(read_committed(&[first, second], 4096).await, Err(SnapshotError::Conflict)));
+    }
+
+    #[tokio::test]
+    async fn newer_slot_does_not_hide_a_conflicting_commit_history() {
+        let root = TempDir::new().expect("test directory");
+        let first = disk(&root, "first").await;
+        let second = disk(&root, "second").await;
+        let owner = Uuid::new_v4();
+        commit(&first, 0, owner, 8, &payload("newest")).await;
+        commit(&first, 1, owner, 7, &payload("a")).await;
+        commit(&second, 1, owner, 7, &payload("b")).await;
+        assert!(matches!(read_committed(&[first, second], 4096).await, Err(SnapshotError::Conflict)));
+    }
+
+    #[tokio::test]
+    async fn uncommitted_or_torn_successor_preserves_previous_slot() {
+        let root = TempDir::new().expect("test directory");
+        let disk = disk(&root, "disk").await;
+        let owner = Uuid::new_v4();
+        let old = payload("old");
+        let next = payload("next");
+        commit(&disk, 0, owner, 1, &old).await;
+        install(&disk, PAYLOAD_PATHS[1], &next).await;
+        let recovered = read_committed(std::slice::from_ref(&disk), 4096)
+            .await
+            .expect("staged payload is not a commit")
+            .expect("old snapshot");
+        assert_eq!(recovered.payload, old);
+        install(&disk, MANIFEST_PATHS[1], &manifest(owner, 2, &next)[..20]).await;
+        let recovered = read_committed(std::slice::from_ref(&disk), 4096)
+            .await
+            .expect("torn manifest preserves old slot")
+            .expect("old snapshot");
+        assert_eq!(recovered.manifest.sequence, 1);
+        install(&disk, MANIFEST_PATHS[1], &manifest(owner, 2, &next)).await;
+        install(&disk, PAYLOAD_PATHS[1], b"torn").await;
+        let recovered = read_committed(&[disk], 4096)
+            .await
+            .expect("torn payload preserves old slot")
+            .expect("old snapshot");
+        assert_eq!(recovered.manifest.sequence, 1);
+    }
+
+    #[tokio::test]
+    async fn stale_manifest_cas_cannot_replace_committed_anchor() {
+        let root = TempDir::new().expect("test directory");
+        let disk = disk(&root, "disk").await;
+        let owner = Uuid::new_v4();
+        let bytes = payload("object");
+        commit(&disk, 0, owner, 1, &bytes).await;
+        let result = EcstoreDiskAPI::compare_and_update_file(
+            disk.as_ref(),
+            RUSTFS_META_BUCKET,
+            MANIFEST_PATHS[0],
+            None,
+            Some(manifest(owner, 2, &bytes).into()),
+        )
+        .await
+        .expect("CAS call");
+        assert_eq!(result, EcstoreConditionalFileUpdate::Mismatch);
+        let recovered = read_committed(&[disk], 4096)
+            .await
+            .expect("read old anchor")
+            .expect("snapshot");
+        assert_eq!(recovered.manifest.sequence, 1);
+    }
+
+    #[tokio::test]
+    async fn legacy_import_requires_complete_consistent_replicas() {
+        let root = TempDir::new().expect("test directory");
+        let first = disk(&root, "first").await;
+        let second = disk(&root, "second").await;
+        let bytes = payload("object");
+        for (disk, data) in [(&first, &bytes[..bytes.len() - 1]), (&second, bytes.as_slice())] {
+            EcstoreDiskAPI::write_all(
+                disk.as_ref(),
+                RUSTFS_META_BUCKET,
+                MRF_SCOPED_JOURNAL_PATH,
+                EcstoreDiskBytes::copy_from_slice(data),
+            )
+            .await
+            .expect("legacy fixture");
+        }
+        let disks = [first.clone(), second];
+        assert!(
+            matches!(read_recovery_snapshot(&disks, 4096).await.expect("intact legacy replica"), Some(RecoverySnapshot::Legacy(data)) if data == bytes)
+        );
+        EcstoreDiskAPI::write_all(first.as_ref(), RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH, payload("different").into())
+            .await
+            .expect("divergent fixture");
+        assert!(matches!(read_recovery_snapshot(&disks, 4096).await, Err(SnapshotError::Conflict)));
+    }
+
+    #[tokio::test]
+    async fn committed_inspection_leaves_payload_and_manifest_unchanged() {
+        let root = TempDir::new().expect("test directory");
+        let disk = disk(&root, "disk").await;
+        let owner = Uuid::new_v4();
+        let bytes = payload("object");
+        commit(&disk, 0, owner, 3, &bytes).await;
+        assert!(matches!(
+            read_recovery_snapshot(std::slice::from_ref(&disk), 4096)
+                .await
+                .expect("new snapshot"),
+            Some(RecoverySnapshot::Committed(_))
+        ));
+        assert_eq!(
+            EcstoreDiskAPI::read_all(disk.as_ref(), RUSTFS_META_BUCKET, MANIFEST_PATHS[0])
+                .await
+                .expect("manifest retained")
+                .as_ref(),
+            manifest(owner, 3, &bytes)
+        );
+        assert_eq!(
+            EcstoreDiskAPI::read_all(disk.as_ref(), RUSTFS_META_BUCKET, PAYLOAD_PATHS[0])
+                .await
+                .expect("payload retained")
+                .as_ref(),
+            bytes
+        );
+    }
+
+    #[tokio::test]
+    async fn legacy_inspection_rejects_complete_subsets_and_scope_ambiguity() {
+        let scoped = |set_index| {
+            let intent = MrfIntent {
+                bucket: Arc::from("bucket"),
+                object: Arc::from("a"),
+                version_id: None,
+                kind: MrfKind::PartialWrite,
+                scope: Some(MrfScope {
+                    pool_index: 0,
+                    set_index,
+                }),
+                lease: None,
+                enqueued_at_ms: 1234,
+                attempts: 0,
+            };
+            let mut bytes = Vec::new();
+            assert!(encode_intent(&intent, &mut bytes), "scoped fixture must encode");
+            bytes
+        };
+        let mut superset = payload("a");
+        superset.extend_from_slice(&payload("b"));
+        for (case, first_bytes, second_bytes) in [
+            ("complete-subset", payload("a"), superset),
+            ("different-set", scoped(1), scoped(2)),
+            ("unknown-scope", payload("a"), scoped(1)),
+        ] {
+            let root = TempDir::new().expect("test directory");
+            let first = disk(&root, "first").await;
+            let second = disk(&root, "second").await;
+            for (disk, bytes) in [(&first, &first_bytes), (&second, &second_bytes)] {
+                assert_eq!(decode_journal(bytes).1, 0, "{case}: complete fixture");
+                EcstoreDiskAPI::write_all(
+                    disk.as_ref(),
+                    RUSTFS_META_BUCKET,
+                    MRF_SCOPED_JOURNAL_PATH,
+                    EcstoreDiskBytes::copy_from_slice(bytes),
+                )
+                .await
+                .expect("write legacy replica");
+            }
+            for disks in [vec![first.clone(), second.clone()], vec![second.clone(), first.clone()]] {
+                assert!(
+                    matches!(read_recovery_snapshot(&disks, 4096).await, Err(SnapshotError::Conflict)),
+                    "{case}: neither replica order proves a latest snapshot"
+                );
+            }
+            for (disk, bytes) in [(&first, &first_bytes), (&second, &second_bytes)] {
+                assert_eq!(
+                    EcstoreDiskAPI::read_all(disk.as_ref(), RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH)
+                        .await
+                        .expect("legacy evidence retained")
+                        .as_ref(),
+                    bytes.as_slice(),
+                    "{case}: inspection must preserve both source replicas"
+                );
+            }
+        }
+    }
+
+    #[tokio::test]
+    async fn oversized_or_corrupt_scoped_snapshot_never_falls_back_to_legacy() {
+        let root = TempDir::new().expect("test directory");
+        let disk = disk(&root, "disk").await;
+        EcstoreDiskAPI::write_all(disk.as_ref(), RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH, vec![0; 1025].into())
+            .await
+            .expect("oversized fixture");
+        EcstoreDiskAPI::write_all(disk.as_ref(), RUSTFS_META_BUCKET, MRF_JOURNAL_PATH, payload("old").into())
+            .await
+            .expect("legacy fixture");
+        assert!(matches!(
+            read_recovery_snapshot(std::slice::from_ref(&disk), 1024).await,
+            Err(SnapshotError::TooLarge)
+        ));
+        assert!(matches!(read_recovery_snapshot(&[disk], 2048).await, Err(SnapshotError::Corrupt)));
+    }
+
+    #[tokio::test]
+    async fn empty_legacy_replica_cannot_erase_records_in_a_torn_replica() {
+        let root = TempDir::new().expect("test directory");
+        let first = disk(&root, "first").await;
+        let second = disk(&root, "second").await;
+        let mut incomplete = payload("durable-object");
+        incomplete.extend_from_slice(b"torn");
+        EcstoreDiskAPI::write_all(first.as_ref(), RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH, Vec::new().into())
+            .await
+            .expect("empty truncated replica");
+        EcstoreDiskAPI::write_all(second.as_ref(), RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH, incomplete.clone().into())
+            .await
+            .expect("records and torn tail");
+        for disks in [vec![first.clone(), second.clone()], vec![second.clone(), first.clone()]] {
+            assert!(matches!(read_recovery_snapshot(&disks, 4096).await, Err(SnapshotError::Corrupt)));
+        }
+        assert_eq!(
+            EcstoreDiskAPI::read_all(second.as_ref(), RUSTFS_META_BUCKET, MRF_SCOPED_JOURNAL_PATH)
+                .await
+                .expect("recovery anchor preserved")
+                .as_ref(),
+            incomplete
+        );
+    }
+
+    #[tokio::test]
+    async fn unreadable_commit_record_never_implies_legacy_only() {
+        let root = TempDir::new().expect("test directory");
+        let disk = disk(&root, "disk").await;
+        let legacy = payload("old");
+        EcstoreDiskAPI::write_all(disk.as_ref(), RUSTFS_META_BUCKET, MRF_JOURNAL_PATH, legacy.clone().into())
+            .await
+            .expect("legacy fixture");
+        // Opening a directory as a record either fails at open or at read,
+        // depending on the platform. Neither outcome proves absence.
+        std::fs::create_dir(root.path().join("disk").join(RUSTFS_META_BUCKET).join(MANIFEST_PATHS[0]))
+            .expect("unreadable manifest fixture");
+        let recovered = read_recovery_snapshot(std::slice::from_ref(&disk), 4096).await;
+        assert!(
+            matches!(recovered, Err(SnapshotError::Disk(_) | SnapshotError::Read(_))),
+            "must preserve unavailable proof: {recovered:?}"
+        );
+        assert_eq!(
+            EcstoreDiskAPI::read_all(disk.as_ref(), RUSTFS_META_BUCKET, MRF_JOURNAL_PATH)
+                .await
+                .expect("legacy remains")
+                .as_ref(),
+            legacy
+        );
+    }
+}

crates/heal/src/heal/task.rs

@@ -999,7 +999,7 @@ impl HealTask {
         let items = match since {
             None => result_items.iter().map(|(_, item)| item.clone()).collect::<Vec<_>>(),
             Some(cursor) => {
-                if cursor + 1 < min_seq {
+                if cursor.saturating_add(1) < min_seq {
                     lagged = true;
                 }
                 result_items

crates/heal/tests/heal_b5_versioned_regression_test.rs

@@ -44,7 +44,9 @@ use walkdir::WalkDir;
 
 mod storage_api;
 
-use storage_api::integration::{BucketOperations, ECStore, MakeBucketOptions, ObjectIO as _, ObjectOperations as _};
+use storage_api::integration::{
+    BucketOperations, ECStore, MakeBucketOptions, NamespaceLocking as _, ObjectIO as _, ObjectOperations as _,
+};
 
 /// 256 KiB + change: large enough to be stored as non-inline erasure shards
 /// (so each data version materializes as an on-disk `part.*` file we can assert
@@ -106,6 +108,7 @@ async fn put_versioned(ecstore: &Arc<ECStore>, bucket: &str, object: &str, data:
         .put_object(bucket, object, &mut reader, &opts)
         .await
         .expect("versioned put_object failed");
+    wait_for_put_tail(ecstore, bucket, object).await;
     info.version_id
         .map(|u| u.to_string())
         .expect("versioned put must return a version id")
@@ -117,6 +120,7 @@ async fn put_unversioned(ecstore: &Arc<ECStore>, bucket: &str, object: &str, dat
         .put_object(bucket, object, &mut reader, &ObjectOptions::default())
         .await
         .expect("unversioned put_object failed");
+    wait_for_put_tail(ecstore, bucket, object).await;
 }
 
 /// Create a delete-marker as the latest version (versioned:true, no version_id)
@@ -160,20 +164,16 @@ fn xl_meta_path(obj_dir: &Path) -> PathBuf {
     obj_dir.join("xl.meta")
 }
 
-async fn wait_for_two_version_copies(disks: &[PathBuf], bucket: &str, object: &str) {
+async fn wait_for_put_tail(ecstore: &Arc<ECStore>, bucket: &str, object: &str) {
-    tokio::time::timeout(Duration::from_secs(5), async {
+    // Shards and xl.meta can exist before the detached PUT owner finishes.
-        loop {
+    let lock = ecstore
-            if disks.iter().all(|disk| {
+        .new_ns_lock(bucket, object)
-                let object_dir = object_dir(disk, bucket, object);
+        .await
-                xl_meta_path(&object_dir).exists() && count_part_files(&object_dir) >= 2
+        .expect("fixture namespace lock should be created");
-            }) {
+    let _settled = lock
-                break;
+        .get_write_lock(Duration::from_secs(30))
-            }
+        .await
-            tokio::time::sleep(Duration::from_millis(10)).await;
+        .expect("PUT rename tail must finish before inspecting or wiping the fixture");
-        }
-    })
-    .await
-    .expect("PUT rename tails must converge before wiping the versioned fixture");
 }
 
 fn recreate_heal_opts() -> HealOpts {
@@ -305,7 +305,13 @@ mod serial_tests {
         let data_v2 = versioned_test_data(20);
         let v1 = put_versioned(&ecstore, bucket, object, &data_v1).await; // OLD, non-latest
         let v2 = put_versioned(&ecstore, bucket, object, &data_v2).await; // latest
-        wait_for_two_version_copies(&disk_paths, bucket, object).await;
+        assert!(
+            disk_paths.iter().all(|disk| {
+                let dir = object_dir(disk, bucket, object);
+                xl_meta_path(&dir).exists() && count_part_files(&dir) >= 2
+            }),
+            "both versions must exist on every disk before wiping the fixture"
+        );
 
         // ── Pre-wipe: prove the fixture actually has 2 versions on disk[0] ──
         let obj_dir0 = object_dir(&disk_paths[0], bucket, object);

crates/heal/tests/storage_api.rs

@@ -23,6 +23,7 @@ pub(crate) mod integration {
     pub(crate) use rustfs_ecstore::api::storage::ECStore;
     pub(crate) use rustfs_storage_api::BucketOperations;
     pub(crate) use rustfs_storage_api::MakeBucketOptions;
+    pub(crate) use rustfs_storage_api::NamespaceLocking;
     pub(crate) use rustfs_storage_api::ObjectIO;
     pub(crate) use rustfs_storage_api::ObjectOperations;
 }

crates/lifecycle/src/core.rs

@@ -43,6 +43,10 @@ const ERR_LIFECYCLE_BUCKET_LOCKED: &str =
     "ExpiredObjectAllVersions element and DelMarkerExpiration action cannot be used on an object locked bucket";
 const ERR_LIFECYCLE_TOO_MANY_RULES: &str = "Lifecycle configuration should have at most 1000 rules";
 const ERR_LIFECYCLE_INVALID_EXPIRATION_DAYS: &str = "'Days' for Expiration action must be a positive integer";
+const ERR_LIFECYCLE_EXPIRATION_DAYS_DATE_CONFLICT: &str = "Expiration cannot specify both Days and Date";
+const ERR_LIFECYCLE_MULTIPLE_TRANSITIONS: &str = "Only one Transition action per lifecycle rule is supported";
+const ERR_LIFECYCLE_MULTIPLE_NONCURRENT_TRANSITIONS: &str =
+    "Only one NoncurrentVersionTransition action per lifecycle rule is supported";
 const ERR_LIFECYCLE_INVALID_NONCURRENT_EXPIRATION_DAYS: &str =
     "'NoncurrentDays' for NoncurrentVersionExpiration action must be a positive integer";
 const ERR_LIFECYCLE_INVALID_ABORT_INCOMPLETE_MPU_DAYS: &str =
@@ -361,6 +365,12 @@ impl Lifecycle for BucketLifecycleConfiguration {
                 {
                     return Err(std::io::Error::other(ERR_LIFECYCLE_INVALID_EXPIRED_OBJECT_ALL_VERSIONS));
                 }
+                if expiration.days.is_some() && expiration.date.is_some() {
+                    return Err(std::io::Error::new(
+                        std::io::ErrorKind::InvalidInput,
+                        ERR_LIFECYCLE_EXPIRATION_DAYS_DATE_CONFLICT,
+                    ));
+                }
                 if let Some(expiration_date) = &expiration.date {
                     let date = OffsetDateTime::from(expiration_date.clone());
                     if date.hour() != 0 || date.minute() != 0 || date.second() != 0 || date.nanosecond() != 0 {
@@ -394,11 +404,20 @@ impl Lifecycle for BucketLifecycleConfiguration {
                 }
             }
             if let Some(transitions) = &r.transitions {
+                if transitions.len() > 1 {
+                    return Err(std::io::Error::new(std::io::ErrorKind::InvalidInput, ERR_LIFECYCLE_MULTIPLE_TRANSITIONS));
+                }
                 for transition in transitions {
                     TransitionOps::validate(transition)?;
                 }
             }
             if let Some(noncurrent_transitions) = &r.noncurrent_version_transitions {
+                if noncurrent_transitions.len() > 1 {
+                    return Err(std::io::Error::new(
+                        std::io::ErrorKind::InvalidInput,
+                        ERR_LIFECYCLE_MULTIPLE_NONCURRENT_TRANSITIONS,
+                    ));
+                }
                 for transition in noncurrent_transitions {
                     NoncurrentVersionTransitionOps::validate(transition)?;
                 }
@@ -473,6 +492,8 @@ impl Lifecycle for BucketLifecycleConfiguration {
     }
 
     async fn eval(&self, obj: &ObjectOpts) -> Event {
+        // A single-object lookup cannot prove how many newer historical versions
+        // survive. Count-dependent actions wait for the complete-group evaluator.
         self.eval_inner(obj, OffsetDateTime::now_utc(), 0).await
     }
 
@@ -536,23 +557,8 @@ impl Lifecycle for BucketLifecycleConfiguration {
             return Event::default();
         };
 
-        if let Some(restore_expires) = obj.restore_expires
+        if let Some(event) = obj.restored_copy_expiry(now) {
-            && restore_expires.unix_timestamp() != 0
+            events.push(event);
-            && now.unix_timestamp() > restore_expires.unix_timestamp()
-        {
-            let mut action = IlmAction::DeleteRestoredAction;
-            if !obj.is_latest {
-                action = IlmAction::DeleteRestoredVersionAction;
-            }
-
-            events.push(Event {
-                action,
-                due: Some(now),
-                rule_id: "".into(),
-                noncurrent_days: 0,
-                newer_noncurrent_versions: 0,
-                storage_class: "".into(),
-            });
         }
 
         if let Some(ref lc_rules) = self.filter_rules(obj).await {
@@ -611,17 +617,12 @@ impl Lifecycle for BucketLifecycleConfiguration {
                     continue;
                 }
 
-                if !obj.is_latest
-                    && let Some(ref noncurrent_version_expiration) = rule.noncurrent_version_expiration
-                    && let Some(retain_newer_noncurrent_versions) = noncurrent_version_expiration.newer_noncurrent_versions
-                    && newer_noncurrent_versions < usize::try_from(retain_newer_noncurrent_versions).unwrap_or(usize::MAX)
-                {
-                    continue;
-                }
-
                 if !obj.is_latest
                     && let Some(ref noncurrent_version_expiration) = rule.noncurrent_version_expiration
                     && let Some(noncurrent_days) = noncurrent_version_expiration.noncurrent_days
+                    && noncurrent_version_expiration
+                        .newer_noncurrent_versions
+                        .is_none_or(|retain| usize::try_from(retain).is_ok_and(|retain| newer_noncurrent_versions >= retain))
                 {
                     if let Some(successor_mod_time) = obj.successor_mod_time {
                         let expected_expiry = expected_expiry_time(successor_mod_time, noncurrent_days);
@@ -651,7 +652,11 @@ impl Lifecycle for BucketLifecycleConfiguration {
                     && let Some(noncurrent_version_transition) = rule
                         .noncurrent_version_transitions
                         .as_ref()
+                        .filter(|transitions| transitions.len() == 1)
                         .and_then(|transitions| transitions.first())
+                    && noncurrent_version_transition
+                        .newer_noncurrent_versions
+                        .is_none_or(|retain| usize::try_from(retain).is_ok_and(|retain| newer_noncurrent_versions >= retain))
                     && let Some(storage_class) = noncurrent_version_transition.storage_class.as_ref()
                     && !storage_class.as_str().is_empty()
                     && !obj.delete_marker
@@ -735,7 +740,11 @@ impl Lifecycle for BucketLifecycleConfiguration {
                     }
 
                     if obj.transition_status != TRANSITION_COMPLETE
-                        && let Some(transition) = rule.transitions.as_ref().and_then(|transitions| transitions.first())
+                        && let Some(transition) = rule
+                            .transitions
+                            .as_ref()
+                            .filter(|transitions| transitions.len() == 1)
+                            .and_then(|transitions| transitions.first())
                         && let Some(storage_class) = transition.storage_class.as_ref()
                         && !storage_class.as_str().is_empty()
                     {
@@ -758,18 +767,15 @@ impl Lifecycle for BucketLifecycleConfiguration {
         }
 
         if !events.is_empty() {
-            // Select the winning event using a strict total order (MinIO semantics):
+            // Eligible expiration takes precedence over transition, even when a
-            // the earliest `due` wins, and ties break toward delete-type actions. A
+            // failed transition has an earlier deadline. Within each action class,
-            // missing `due` is treated as UNIX_EPOCH. This replaces a hand-written
+            // prefer the earliest deadline using a deterministic total order.
-            // `sort_by` comparator that was not a strict weak ordering (it could return
-            // `Ordering::Less` for both `(a, b)` and `(b, a)`), which panics on the
-            // repository toolchain and did not deterministically pick the earliest event.
             let event = events
                 .iter()
                 .min_by_key(|event| {
                     (
-                        event.due.unwrap_or(OffsetDateTime::UNIX_EPOCH).unix_timestamp(),
                         ilm_action_priority_rank(&event.action),
+                        event.due.unwrap_or(OffsetDateTime::UNIX_EPOCH).unix_timestamp(),
                     )
                 })
                 .cloned()
@@ -1042,6 +1048,27 @@ impl ObjectOpts {
     pub fn expired_object_deletemarker(&self) -> bool {
         self.delete_marker && self.is_latest && self.num_versions == 1
     }
+
+    pub(crate) fn restored_copy_expiry(&self, now: OffsetDateTime) -> Option<Event> {
+        let restore_expires = self.restore_expires?;
+        // Restore metadata alone does not prove that a durable remote copy exists.
+        if self.transition_status != TRANSITION_COMPLETE
+            || restore_expires.unix_timestamp() == 0
+            || now.unix_timestamp() <= restore_expires.unix_timestamp()
+        {
+            return None;
+        }
+        let action = if self.is_latest {
+            IlmAction::DeleteRestoredAction
+        } else {
+            IlmAction::DeleteRestoredVersionAction
+        };
+        expiration_action_has_valid_target(action, self.version_id, self.is_latest, self.delete_marker).then(|| Event {
+            action,
+            due: Some(now),
+            ..Default::default()
+        })
+    }
 }
 
 /// Returns whether an expiry action has enough identity to target the object
@@ -1064,11 +1091,8 @@ pub fn expiration_action_has_valid_target(
     }
 }
 
-/// Total-order rank for lifecycle actions used to break `due` ties.
+/// Eligible logical expiration takes precedence over transition and restore-copy
-///
+/// cleanup. Deadlines break ties within an action class.
-/// Delete-type actions rank before every other action so that, when two events
-/// share the same `due`, a delete wins (MinIO semantics). The concrete numeric
-/// values only matter relative to each other.
 fn ilm_action_priority_rank(action: &IlmAction) -> u8 {
     match action {
         IlmAction::DeleteAllVersionsAction
@@ -4159,6 +4183,392 @@ mod tests {
         assert_eq!(event.action, IlmAction::NoneAction);
     }
 
+    mod adversarial_regressions {
+        use super::*;
+        use s3s::dto::NoncurrentVersionExpiration;
+
+        fn run(test: impl std::future::Future<Output = ()>) {
+            with_default_ilm_process_time(|| {
+                tokio::runtime::Builder::new_current_thread()
+                    .build()
+                    .expect("lifecycle regression runtime should build")
+                    .block_on(test);
+            });
+        }
+
+        fn noncurrent_object() -> ObjectOpts {
+            ObjectOpts {
+                name: "logs/object".to_string(),
+                mod_time: Some(datetime!(2020-01-01 00:00:00 UTC)),
+                successor_mod_time: Some(datetime!(2020-01-02 00:00:00 UTC)),
+                version_id: Some(Uuid::from_u128(1)),
+                size: 1024 * 1024,
+                ..Default::default()
+            }
+        }
+
+        #[test]
+        #[serial]
+        fn noncurrent_transition_retains_the_requested_newer_versions() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("retain-two-hot-versions"));
+                rule.filter = Some(LifecycleRuleFilter::default());
+                rule.noncurrent_version_transitions = Some(vec![NoncurrentVersionTransition {
+                    noncurrent_days: Some(1),
+                    newer_noncurrent_versions: Some(2),
+                    storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                }]);
+                let lc = Arc::new(BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                });
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("valid noncurrent transition policy");
+                let objects = (0..4)
+                    .map(|index| ObjectOpts {
+                        mod_time: Some(datetime!(2020-01-05 00:00:00 UTC) - Duration::days(index)),
+                        successor_mod_time: (index > 0).then_some(datetime!(2020-01-06 00:00:00 UTC) - Duration::days(index)),
+                        version_id: Some(Uuid::from_u128(u128::try_from(index + 1).expect("small version index"))),
+                        is_latest: index == 0,
+                        num_versions: 4,
+                        ..noncurrent_object()
+                    })
+                    .collect::<Vec<_>>();
+                let actions = crate::Evaluator::new(lc)
+                    .eval(&objects)
+                    .await
+                    .expect("complete version chain should evaluate")
+                    .into_iter()
+                    .map(|event| event.action)
+                    .collect::<Vec<_>>();
+                assert_eq!(
+                    actions,
+                    [
+                        IlmAction::NoneAction,
+                        IlmAction::NoneAction,
+                        IlmAction::NoneAction,
+                        IlmAction::TransitionVersionAction
+                    ],
+                    "the two newest noncurrent versions must remain in their current storage class"
+                );
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn noncurrent_transition_checks_count_age_and_single_object_context() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("retain-two"));
+                rule.filter = Some(LifecycleRuleFilter::default());
+                rule.noncurrent_version_transitions = Some(vec![NoncurrentVersionTransition {
+                    noncurrent_days: Some(3),
+                    newer_noncurrent_versions: Some(2),
+                    storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                }]);
+                let mut lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("valid counted transition");
+                let object = noncurrent_object();
+                let now = datetime!(2020-01-10 00:00:00 UTC);
+                for (newer, expected) in [
+                    (0, IlmAction::NoneAction),
+                    (1, IlmAction::NoneAction),
+                    (2, IlmAction::TransitionVersionAction),
+                    (3, IlmAction::TransitionVersionAction),
+                ] {
+                    assert_eq!(lc.eval_inner(&object, now, newer).await.action, expected, "newer count: {newer}");
+                }
+                assert_eq!(
+                    lc.eval_inner(&object, datetime!(2020-01-04 00:00:00 UTC), 2).await.action,
+                    IlmAction::NoneAction,
+                    "the retention count does not replace the age condition"
+                );
+                assert_eq!(
+                    lc.eval(&object).await.action,
+                    IlmAction::NoneAction,
+                    "a single-object lookup must not assume a complete version history"
+                );
+                for retain in [None, Some(0), Some(-1), Some(i32::MAX)] {
+                    lc.rules[0]
+                        .noncurrent_version_transitions
+                        .as_mut()
+                        .expect("transition exists")[0]
+                        .newer_noncurrent_versions = retain;
+                    let expected = if matches!(retain, None | Some(0)) {
+                        IlmAction::TransitionVersionAction
+                    } else {
+                        IlmAction::NoneAction
+                    };
+                    assert_eq!(lc.eval_inner(&object, now, 2).await.action, expected, "retention: {retain:?}");
+                }
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn noncurrent_expiration_and_transition_have_independent_retention_counts() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("independent-counts"));
+                rule.filter = Some(LifecycleRuleFilter::default());
+                rule.noncurrent_version_expiration = Some(NoncurrentVersionExpiration {
+                    noncurrent_days: Some(90),
+                    newer_noncurrent_versions: Some(4),
+                });
+                rule.noncurrent_version_transitions = Some(vec![NoncurrentVersionTransition {
+                    noncurrent_days: Some(30),
+                    newer_noncurrent_versions: Some(2),
+                    storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                }]);
+                let lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("valid independent retention limits");
+                let object = noncurrent_object();
+                let now = datetime!(2020-05-01 00:00:00 UTC);
+                for (newer, expected) in [
+                    (1, IlmAction::NoneAction),
+                    (2, IlmAction::TransitionVersionAction),
+                    (3, IlmAction::TransitionVersionAction),
+                    (4, IlmAction::DeleteVersionAction),
+                ] {
+                    assert_eq!(lc.eval_inner(&object, now, newer).await.action, expected, "newer count: {newer}");
+                }
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn expiration_retention_does_not_skip_an_independent_transition() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("transition-then-expire"));
+                rule.filter = Some(LifecycleRuleFilter::default());
+                rule.noncurrent_version_transitions = Some(vec![NoncurrentVersionTransition {
+                    noncurrent_days: Some(1),
+                    newer_noncurrent_versions: None,
+                    storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                }]);
+                let mut lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                let object = noncurrent_object();
+                let now = datetime!(2020-01-10 00:00:00 UTC);
+                let transition_only = lc.eval_inner(&object, now, 0).await;
+                assert_eq!(transition_only.action, IlmAction::TransitionVersionAction);
+
+                lc.rules[0].noncurrent_version_expiration = Some(NoncurrentVersionExpiration {
+                    noncurrent_days: Some(90),
+                    newer_noncurrent_versions: Some(2),
+                });
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("valid combined policy");
+                let combined = lc.eval_inner(&object, now, 0).await;
+                assert_eq!(combined.action, transition_only.action, "retention limits expiration, not transition");
+                assert_eq!(combined.storage_class, transition_only.storage_class);
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn current_transition_rejects_multiple_stages_in_any_order() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("two-current-transitions"));
+                rule.transitions = Some(vec![
+                    Transition {
+                        date: Some(datetime!(2020-03-01 00:00:00 UTC).into()),
+                        days: None,
+                        storage_class: Some(TransitionStorageClass::from_static("COLD")),
+                    },
+                    Transition {
+                        date: Some(datetime!(2020-01-03 00:00:00 UTC).into()),
+                        days: None,
+                        storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                    },
+                ]);
+                let mut lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                let object = ObjectOpts {
+                    is_latest: true,
+                    ..noncurrent_object()
+                };
+                let now = datetime!(2020-01-10 00:00:00 UTC);
+                for status in [ExpirationStatus::ENABLED, ExpirationStatus::DISABLED] {
+                    lc.rules[0].status = ExpirationStatus::from_static(status);
+                    for _ in 0..2 {
+                        let err = lc
+                            .validate(&ObjectLockConfiguration::default())
+                            .await
+                            .expect_err("multiple transition stages must be rejected");
+                        assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+                        assert_eq!(err.to_string(), ERR_LIFECYCLE_MULTIPLE_TRANSITIONS);
+                        assert_eq!(
+                            lc.eval_inner(&object, now, 0).await.action,
+                            IlmAction::NoneAction,
+                            "legacy multi-stage configurations must not silently execute their first stage"
+                        );
+                        lc.rules[0]
+                            .transitions
+                            .as_mut()
+                            .expect("transition array is present")
+                            .reverse();
+                    }
+                }
+                lc.rules[0]
+                    .transitions
+                    .as_mut()
+                    .expect("transition array is present")
+                    .remove(0);
+                lc.rules[0].status = ExpirationStatus::from_static(ExpirationStatus::ENABLED);
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("one stage is supported");
+                let event = lc.eval_inner(&object, now, 0).await;
+                assert_eq!(event.action, IlmAction::TransitionAction);
+                assert_eq!(event.storage_class, "WARM");
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn noncurrent_transition_rejects_multiple_stages_in_any_order() {
+            run(async {
+                let mut rule = enabled_rule(None, None, Some("two-noncurrent-transitions"));
+                rule.noncurrent_version_transitions = Some(vec![
+                    NoncurrentVersionTransition {
+                        noncurrent_days: Some(30),
+                        newer_noncurrent_versions: None,
+                        storage_class: Some(TransitionStorageClass::from_static("COLD")),
+                    },
+                    NoncurrentVersionTransition {
+                        noncurrent_days: Some(1),
+                        newer_noncurrent_versions: None,
+                        storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                    },
+                ]);
+                let mut lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                let object = noncurrent_object();
+                let now = datetime!(2020-01-10 00:00:00 UTC);
+                for status in [ExpirationStatus::ENABLED, ExpirationStatus::DISABLED] {
+                    lc.rules[0].status = ExpirationStatus::from_static(status);
+                    for _ in 0..2 {
+                        let err = lc
+                            .validate(&ObjectLockConfiguration::default())
+                            .await
+                            .expect_err("multiple noncurrent transition stages must be rejected");
+                        assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+                        assert_eq!(err.to_string(), ERR_LIFECYCLE_MULTIPLE_NONCURRENT_TRANSITIONS);
+                        assert_eq!(
+                            lc.eval_inner(&object, now, 0).await.action,
+                            IlmAction::NoneAction,
+                            "legacy multi-stage configurations must not silently execute their first stage"
+                        );
+                        lc.rules[0]
+                            .noncurrent_version_transitions
+                            .as_mut()
+                            .expect("transition array is present")
+                            .reverse();
+                    }
+                }
+                lc.rules[0]
+                    .noncurrent_version_transitions
+                    .as_mut()
+                    .expect("transition array is present")
+                    .remove(0);
+                lc.rules[0].status = ExpirationStatus::from_static(ExpirationStatus::ENABLED);
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("one stage is supported");
+                let event = lc.eval_inner(&object, now, 0).await;
+                assert_eq!(event.action, IlmAction::TransitionVersionAction);
+                assert_eq!(event.storage_class, "WARM");
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn expiration_rejects_simultaneous_days_and_date() {
+            run(async {
+                let mut lc = BucketLifecycleConfiguration {
+                    rules: vec![enabled_rule(
+                        Some(LifecycleExpiration {
+                            days: Some(1),
+                            ..Default::default()
+                        }),
+                        None,
+                        Some("ambiguous-expiry"),
+                    )],
+                    expiry_updated_at: None,
+                };
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("a single Days expiration is valid");
+                lc.rules[0].expiration.as_mut().expect("expiration is present").date =
+                    Some(datetime!(2099-01-01 00:00:00 UTC).into());
+                let err = lc
+                    .validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect_err("Days and Date are mutually exclusive; accepting both silently overrides Days");
+                assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+                assert_eq!(err.to_string(), ERR_LIFECYCLE_EXPIRATION_DAYS_DATE_CONFLICT);
+            });
+        }
+
+        #[test]
+        #[serial]
+        fn overdue_transition_does_not_starve_permanent_expiration() {
+            run(async {
+                let mut rule = enabled_rule(
+                    Some(LifecycleExpiration {
+                        days: Some(90),
+                        ..Default::default()
+                    }),
+                    None,
+                    Some("archive-then-delete"),
+                );
+                rule.transitions = Some(vec![Transition {
+                    days: Some(30),
+                    date: None,
+                    storage_class: Some(TransitionStorageClass::from_static("WARM")),
+                }]);
+                let lc = BucketLifecycleConfiguration {
+                    rules: vec![rule],
+                    expiry_updated_at: None,
+                };
+                lc.validate(&ObjectLockConfiguration::default())
+                    .await
+                    .expect("valid transition and expiration policy");
+                let object = ObjectOpts {
+                    is_latest: true,
+                    version_id: None,
+                    transition_status: TRANSITION_PENDING.to_string(),
+                    ..noncurrent_object()
+                };
+                let before_expiration = lc.eval_inner(&object, datetime!(2020-02-15 00:00:00 UTC), 0).await;
+                assert_eq!(before_expiration.action, IlmAction::TransitionAction);
+                let overdue = lc.eval_inner(&object, datetime!(2020-05-01 00:00:00 UTC), 0).await;
+                assert_eq!(
+                    overdue.action,
+                    IlmAction::DeleteAction,
+                    "an unavailable tier must not prevent permanent expiration indefinitely"
+                );
+            });
+        }
+    }
+
     /// Property-based tests for the rule evaluator (backlog#1148 ilm-14,
     /// follow-up to backlog#1030 / rustfs#4455).
     ///
@@ -4169,7 +4579,7 @@ mod tests {
     ///
     /// * `eval_inner` never panics and is deterministic for a fixed input;
     /// * the winning event matches an independently recomputed candidate set:
-    ///   earliest `due` wins, ties break toward delete-class actions (the
+    ///   eligible expiration wins over transition, then earliest `due` wins (the
     ///   `min_by_key` selection that replaced the rustfs#4455 comparator);
     /// * `expected_expiry_time` is monotonically non-decreasing in `days` and
     ///   always lands on the processing boundary, both at production defaults
@@ -4458,8 +4868,8 @@ mod tests {
         /// consider for a live current version under `selection`-shaped rules
         /// (expiration and first-transition only, no filters): expiration
         /// fires when `now >= due`, transition when `now > due` and the object
-        /// has not already transitioned. Selection semantics under test:
+        /// has not already transitioned. Eligible expiration wins over transition;
-        /// earliest due wins, ties prefer delete-class.
+        /// the earliest deadline wins within the selected action class.
         fn oracle_candidates(lc: &BucketLifecycleConfiguration, obj: &ObjectOpts, now: OffsetDateTime) -> Vec<Candidate> {
             let mod_time = obj.mod_time.expect("selection strategy always sets mod_time");
             let mut candidates = Vec::new();
@@ -4548,8 +4958,8 @@ mod tests {
             /// Differential test of winner selection (the rustfs#4455 fix):
             /// for a live current version under randomized expiration and
             /// transition rules, `eval_inner`'s winner must carry the
-            /// minimum `(due, rank)` of the independently recomputed
+            /// earliest expiration from the independently recomputed candidate
-            /// candidate set — earliest due wins, ties prefer delete-class —
+            /// set, or the earliest transition when no expiration is eligible,
             /// and must be `NoneAction` exactly when that set is empty.
             #[test]
             #[serial]
@@ -4578,7 +4988,13 @@ mod tests {
 
                 // Oracle and evaluator must observe the same (pinned) time env.
                 let (event, expected) = with_production_time_env(|| {
-                    let expected = oracle_candidates(&lc, &obj, now).into_iter().min();
+                    let candidates = oracle_candidates(&lc, &obj, now);
+                    let expected = candidates
+                        .iter()
+                        .filter(|(_, rank)| *rank == 0)
+                        .min()
+                        .copied()
+                        .or_else(|| candidates.into_iter().min());
                     let rt = tokio::runtime::Builder::new_current_thread()
                         .enable_all()
                         .build()