fix(ecstore): preserve remote preparation corruption

test(ecstore): reject conflicting native preparation locators
fix(ecstore): publish native stream files atomically
2026-10-29 15:29:35 +00:00 · 2026-10-16 05:56:35 +08:00 · 2026-10-15 16:30:27 +08:00 · 2026-10-15 16:14:01 +08:00 · 2026-10-15 13:38:03 +08:00 · 2026-10-15 12:56:25 +08:00
427 changed files with 62320 additions and 4867 deletions
@@ -50,6 +50,7 @@ Update this file only when an advisory adds or changes a reusable lesson, affect

 - `GHSA-g8w9-qw9q-fghr` and `GHSA-xm99-m3gq-83g8`: a valid presigned `PutObject` accepted extra `x-amz-*` headers omitted from `SignedHeaders`; `x-amz-copy-source` could turn that upload capability into a cross-bucket read performed as the signer. Lesson: a presigned URL is a bounded capability; inspect all received security-sensitive headers and reject unsigned ones before selecting the S3 operation or reaching storage.
 - `GHSA-3ppv-fx5m-m749`: explicit `versionId` reads and copy sources authorized `s3:GetObject` instead of `s3:GetObjectVersion`. Lesson: version-specific object access must select version-specific actions for direct reads, `CopyObject`, and `UploadPartCopy`, with tests proving the backend is not reached on denial.
+- `GHSA-gjph-xcgw-4p7w`: `PutObject` and browser POST accepted a caller-supplied `versionId` and replaced an unlocked historical version with only `s3:PutObject`. Lesson: only replication-authorized writes may choose an existing version ID; ordinary uploads must not rewrite version history.
 - `GHSA-x298-9x87-fvjq`: anonymous `ListObjectVersions` fell back to `ListBucket` and returned before public-access-block gates. Lesson: compatibility fallbacks must converge on the same post-authorization checks as direct grants, especially `RestrictPublicBuckets` and anonymous data-plane denies.
 - `GHSA-mx42-j6wv-px98`: `UploadPartCopy` missed source authorization and allowed cross-bucket object exfiltration. Lesson: multipart copy must enforce the same source and destination contract as `CopyObject`.
 - `GHSA-wfxj-ph3v-7mjf`: `UploadPartCopy` checked source and destination independently but missed destination copy-source policy constraints. Lesson: source read and destination write checks are not sufficient when policy constrains allowed copy sources.
@@ -60,12 +61,14 @@ Update this file only when an advisory adds or changes a reusable lesson, affect
 - `GHSA-3g29-xff2-92vp`: FTP `RETR` and `SIZE`/`MDTM` read paths authenticated the user but skipped IAM before calling storage. Lesson: non-HTTP protocol frontends must enforce the same per-operation authorization as the S3 API before backend access.
 - `GHSA-g3vq-vv42-f647`: FTPS `MKD` called `create_bucket` without checking `s3:CreateBucket`. Lesson: protocol command handlers need action-specific checks even when sibling handlers already authorize correctly.
 - `GHSA-3p3x-734c-h5vx`: FTPS and WebDAV compared secret keys with early-return string equality, while FTPS also returned distinguishable invalid-user and invalid-password failures. Lesson: password-style protocol auth needs constant-time secret comparison, indistinguishable failures where practical, and rate limiting.
+- `GHSA-rhg6-jrf2-29c4`: WebDAV, FTP/FTPS, and SFTP used identity-policy checks with empty conditions and bypassed S3 bucket-policy explicit denies and condition-key denies. Lesson: alternate listeners must evaluate the same effective policy and trustworthy request context before storage access.

 ### Filesystem paths and object key traversal

 - `GHSA-pq29-69jg-9mxc`: RPC `read_file_stream` joined untrusted paths under a volume directory without canonical boundary checks. Lesson: `PathBuf::join` plus length checks are not path security.
 - `GHSA-8r6f-hmq2-28rg`: object keys containing traversal sequences bypassed bucket/object authorization when mapped to filesystem paths. Lesson: reject traversal at object-key parsing and verify final storage paths remain under the expected bucket/key root.
 - `GHSA-f4vq-9ffr-m8m3`: Snowball auto-extract accepted archive entries such as `../victim-bucket/object`, authorized the raw attacker-bucket path, then storage path cleaning crossed bucket boundaries. Lesson: archive entries become object keys and need traversal rejection plus consistent authz/storage normalization before writes.
+- `GHSA-g7fq-6xjm-mxf4`: versioned `DeleteObjects` accepted dot-segment keys, authorized the literal prefix, then wrote delete markers under a resolved sibling prefix. Lesson: validate each batch key before authorization and storage normalization, using the same object-name rules as single delete.

 ### Secrets, defaults, and cryptographic misuse

@@ -106,6 +109,7 @@ Update this file only when an advisory adds or changes a reusable lesson, affect
 ### Object Lock and retention invariants

 - `GHSA-j548-9grx-fh4f`: Object Lock enforcement treated unreadable, fabricated, or unparsable bucket metadata as absent configuration and allowed retained objects to be deleted or expired. Lesson: retention must fail closed unless Object Lock absence is authoritative, and every delete, lifecycle, scanner, force-delete, and default-retention path needs the same state distinction.
+- `GHSA-28w4-rp2p-636w` and `GHSA-c7hx-vr97-45rg`: an authorized `DeleteBucket` with `x-rustfs-force-delete: true` removed active COMPLIANCE-retained versions even though ordinary version deletion was denied. Lesson: bucket force-delete must enforce retention across every version, including against root/owner callers.

 ### Serde deserialization and input validation

@@ -135,16 +139,17 @@ rg -n "deny_unknown_fields|serde.default|as u32|as usize|as i32" rustfs crates
 ## Minimum Regression Test Expectations

 - Authz fixes: include unauthenticated, valid low-privilege, wrong-action, correct-action, owner, non-owner, and root/admin cases as applicable.
- Protocol frontend authz fixes: include denied `RETR`, `SIZE`/`MDTM`, `MKD`, bucket probe, and sibling allowed-operation cases, and assert denied paths do not reach the storage backend.
+- Protocol frontend authz fixes: include denied `RETR`, `SIZE`/`MDTM`, `MKD`, bucket probe, and sibling allowed-operation cases; test bucket-policy explicit denies and condition-key denies on each enabled listener against an S3 control, and assert denied paths do not reach storage.
 - IAM fixes: include import/update/list service-account cases with attacker-controlled parent, claims, access key, secret key, and policy.
 - Copy/upload fixes: include cross-bucket, cross-user, source-denied, destination-denied, copy-source-condition, and multipart completion cases.
 - Presigned upload fixes: include a valid presign with extra unsigned tagging, redirect, storage-class, and cross-bucket copy-source headers; require rejection before source or destination storage access, and verify explicitly signed equivalents still work.
 - Version-action fixes: include historical UUID, explicit current version, `null`, range, partNumber, presigned, STS/session, service-account, anonymous bucket-policy, copy source, and multipart-copy source cases.
+- Version-write fixes: a `PutObject`-only principal must not replace an existing version through `PutObject` or browser POST `?versionId=`; verify replication-authorized writes and versioned, suspended, and unversioned buckets separately.
 - Policy-condition fixes: include reserved-key header collisions, missing keys, partially overlapping multi-value sets, plugin mode, and built-in policy mode.
- Path fixes: include encoded traversal, absolute path, nested traversal, archive entries with `..`, valid object keys that resemble traversal text but should be rejected, and canonical bucket/prefix boundary checks.
+- Path fixes: include encoded traversal, absolute path, nested traversal, archive entries with `..`, valid object keys that resemble traversal text but should be rejected, and canonical bucket/prefix boundary checks; test per-key `DeleteObjects` errors and prefix-scoped denials on versioned buckets.
 - Logging fixes: assert redacted output for structs and response bodies that may contain credentials.
 - IAM export fixes: assert exported archives omit plaintext user and service-account secrets unless the format deliberately encrypts or seals them.
 - RPC auth fixes: include captured metadata replay across two concrete methods, stale timestamps, wrong path, wrong method surrogate, wrong secret, and valid same-method calls.
 - Browser/CORS fixes: assert no credentials on reflected/default origins, correct behavior for explicit allowlists, and no same-origin script execution for previewed object content.
 - SSE fixes: inspect stored bytes and verify API metadata, read-back behavior, and on-disk ciphertext together; for ETags, upload equal plaintext under different keys and assert list-only callers cannot derive or correlate the plaintext digest, including multipart objects.
- Object Lock fixes: include unreadable metadata, fabricated metadata defaults, unparsable config, confirmed absent config, COMPLIANCE/GOVERNANCE retention, lifecycle expiry, scanner sweeps, and force-delete paths.
+- Object Lock fixes: include unreadable metadata, fabricated metadata defaults, unparsable config, confirmed absent config, COMPLIANCE/GOVERNANCE retention, lifecycle expiry, scanner sweeps, and force-delete paths; an authorized force-delete must preserve active COMPLIANCE versions.
@@ -1,2 +1,2 @@
 sha256-darwin=98dff0961c6eb21d6128bfd2d4a6100a624ddf1fb7db05dba360a5d3d0a310c2
-sha256-linux=9ff0da7406de9b89c0e848ff29f2238b30ba9da89c7e9df97c246ce86631350f
+sha256-linux=890f5925b87e1bf3710c356253f3b2276d901e156df9ccc04ec9315374de6d22
@@ -1,2 +1,2 @@
-sha256-darwin=55534a97fbd376f64c8f6c341d319017d11ff77cad6da8629a1a7f6a874e0315
-sha256-linux=c06fb8c19aed6f388b9dc61cb8251b7a44f8561a9bf764ad2b9e635598f8dc17
+sha256-darwin=a2cb67d91eb0d093928371d6e84a84b2dd41b2ba975a28094e81879bd65838ad
+sha256-linux=82db0561cf9b111f11782d647d0859c9d2a44c5c68481b5e39c48505828bd392
@@ -61,6 +61,11 @@
      "suite": "rustfs-ecstore",
      "name": "set_disk::core::io_primitives::tests::rename_rollback_incomplete_preserves_overwrite_data_dirs_and_staging"
    },
+    {
+      "invariant": "rollback-closed-inspection-channel",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::core::io_primitives::tests::rename_rollback_closed_heal_receiver_preserves_recovery_material"
+    },
    {
      "invariant": "rollback-cancellation",
      "suite": "rustfs-ecstore",
@@ -86,6 +91,11 @@
      "suite": "rustfs-ecstore",
      "name": "set_disk::ops::object::put_object_tmp_cleanup_tests::tail_drained_put_waits_for_tail_and_allows_immediate_cas"
    },
+    {
+      "invariant": "seeded-put-tail-pressure",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::ops::object::put_object_tmp_cleanup_tests::seeded_put_completion_pressure_drains_every_real_tail"
+    },
    {
      "invariant": "borrowed-lock-tail",
      "suite": "rustfs-ecstore",
@@ -151,6 +161,11 @@
      "suite": "rustfs-ecstore",
      "name": "set_disk::core::metadata_quorum::tests::metadata_observation_all_four_slot_states_and_arrival_orders_preserve_reduction"
    },
+    {
+      "invariant": "metadata-observation-mixed-current-identity-arrivals-preserve-selection",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::core::metadata_quorum::tests::metadata_observation_mixed_current_identity_arrivals_preserve_selection"
+    },
    {
      "invariant": "metadata-observation-pending-newer-version-and-same-time-different-directory-force-full-wait",
      "suite": "rustfs-ecstore",
@@ -166,6 +181,21 @@
      "suite": "rustfs-ecstore",
      "name": "set_disk::core::metadata_quorum::tests::metadata_observation_null_versions_markers_and_invalid_success_keep_their_meaning"
    },
+    {
+      "invariant": "metadata-cache-real-get-and-range-preserve-body-across-read-modes",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::read::metadata_cache_tests::metadata_cache_real_get_and_range_preserve_body_across_read_modes"
+    },
+    {
+      "invariant": "late-metadata-refresh-rejects-changed-generation-and-mapped-shards",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::read::metadata_cache_tests::late_metadata_refresh_rejects_changed_generation_and_mapped_shards"
+    },
+    {
+      "invariant": "late-full-get-rejects-changed-object-generation",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::prepared_get_object_metadata_tests::non_inline_two_phase_get_rejects_generation_change_before_late_refresh"
+    },
    {
      "invariant": "list-objects-shared-corpus-static-namespace",
      "suite": "rustfs-ecstore",
@@ -260,6 +290,46 @@
      "invariant": "logical-prefix-limit",
      "suite": "rustfs-ecstore",
      "name": "store::list_objects::test::list_path_gather_results_counts_common_prefixes_before_page_limit"
+    },
+    {
+      "invariant": "quota-same-name-owner-and-disk-authority",
+      "suite": "rustfs-ecstore",
+      "name": "bucket::metadata_sys::tests::quota_snapshot_reuses_same_name_owner_and_rereads_disk_authority"
+    },
+    {
+      "invariant": "quota-snapshot-scope-and-owner",
+      "suite": "rustfs-ecstore",
+      "name": "bucket::metadata_sys::tests::quota_snapshot_rejects_wrong_scope_incarnation_and_transaction_guard"
+    },
+    {
+      "invariant": "quota-snapshot-lifecycle-and-incarnation",
+      "suite": "rustfs-ecstore",
+      "name": "bucket::metadata_sys::tests::quota_snapshot_rejects_lifecycle_loss_during_read_and_changed_disk_incarnation"
+    },
+    {
+      "invariant": "marker-purge-receipt-reopen",
+      "suite": "rustfs-ecstore",
+      "name": "set_disk::ops::object::marker_purge_receipt_tests::quorum_marker_purge_receipt_survives_reopen_with_stale_member"
+    },
+    {
+      "invariant": "external-minio-shards",
+      "suite": "rustfs-ecstore::pinned_erasure_fixtures_test",
+      "name": "pinned_minio_shards_decode_and_reconstruct_plaintext"
+    },
+    {
+      "invariant": "external-legacy-shards",
+      "suite": "rustfs-ecstore::pinned_erasure_fixtures_test",
+      "name": "pinned_legacy_shards_decode_and_reconstruct_plaintext"
+    },
+    {
+      "invariant": "external-shard-bitrot",
+      "suite": "rustfs-ecstore::pinned_erasure_fixtures_test",
+      "name": "pinned_erasure_shards_reject_bitrot_corruption"
+    },
+    {
+      "invariant": "external-shard-quorum-codec",
+      "suite": "rustfs-ecstore::pinned_erasure_fixtures_test",
+      "name": "pinned_erasure_shards_reject_insufficient_quorum_and_wrong_codec"
    }
  ],
  "fixtures": [
@@ -302,6 +372,21 @@
      "path": "crates/ecstore/tests/fixtures/list_namespace_keys.json",
      "sha256": "5afddb60565deba0938e617470255563f0e15a15fab699645c59bb7ccc84898e",
      "source": "Shared static LIST namespace corpus; ECStore differential tests and ODM list_through_static_namespace_boundary_matrix"
+    },
+    {
+      "path": "crates/ecstore/tests/fixtures/erasure-shards/minio/manifest.json",
+      "sha256": "c398f8dcb76903f6cf8eee07931c76592895ced201ff9d59ed4ff9b0ead2b13c",
+      "source": "MinIO RELEASE.2025-07-23T15-54-02Z; crates/ecstore/tests/fixtures/erasure-shards/README.md"
+    },
+    {
+      "path": "crates/ecstore/tests/fixtures/erasure-shards/legacy/manifest.json",
+      "sha256": "dce397d5b2f42748eb5091b4355bb9d133a50cbba003626fafac28095dbe540b",
+      "source": "RustFS 1.0.0-alpha.37; crates/ecstore/tests/fixtures/erasure-shards/README.md"
+    },
+    {
+      "path": "crates/ecstore/tests/fixtures/erasure-shards/plaintext.bin",
+      "sha256": "b5a83332327961a31b65a8731c64559fd04ff8ac9f6124022201dbbdba5ffa01",
+      "source": "Independent fixed SHA256 counter-stream oracle; crates/ecstore/tests/fixtures/erasure-shards/README.md"
    }
  ]
 }
@@ -92,6 +92,12 @@ setup = 'ecstore-large-stack'
 filter = 'package(rustfs-ecstore) | package(rustfs-s3select-api) | package(rustfs-scanner) | (package(rustfs) & test(/^(app::multipart_usecase::tests::concurrent_completions_share_durable_bucket_quota_reservations|app::object::delete::tests::compressed_delete_requests_update_observed_usage_without_releasing_quota_floor|app::object::internal_put::tests::internal_multipart_roundtrip_completes_and_abort_leaves_nothing|app::object::restore::tests::execute_restore_object_maps_failures_to_typed_s3_errors|storage::access::tests::(delete_object_access_captures_authorized_bucket_incarnation|copy_operations_reject_recreated_source_bucket_after_authorization|request_slot_keeps_bucket_policy_bound_to_its_store))$/))'
 setup = 'ecstore-base-stack'

+# These handler/store fixtures and the relocated-pool GET compose deep storage
+# futures in libtest threads; on Linux they overflow the default 2 MiB stack.
+[[profile.default.scripts]]
+filter = 'package(rustfs) & (test(/^admin::handlers::table_catalog::tests::/) | test(/^table_catalog::store::tests::/) | test(=app::object::get::tests::execute_get_object_resumes_from_relocated_pool_without_splicing_body))'
+setup = 'ecstore-base-stack'
+
 [[profile.default.scripts]]
 filter = 'binary(lifecycle_integration_test) | (package(rustfs) & test(/^app::lifecycle_transition_api_test::/))'
 setup = 'lifecycle-large-stack'
@@ -170,10 +176,11 @@ filter = 'package(rustfs) & (binary(/^embedded.*_test$/) | binary(admin_diagnost
 test-group = 'embedded-test-ports'

 # Registration, inventory delivery, and real drive/object probes include
-# durable filesystem IO in short deadlines. Reserve capacity so sibling test
+# filesystem or HTTP IO in short deadlines. Reserve capacity so sibling test
 # processes cannot exhaust those budgets. Keep deadlines and assertions intact.
+# The short HTTP fixtures are tracked in rustfs/backlog#2714.
 [[profile.default.overrides]]
-filter = 'package(rustfs) & (binary(/^connect_registration$/) | binary(connect_inventory) | binary(connect_perf_drive) | (binary(connect_perf_object) & test(=real_rustfs_endpoint_and_production_cli_support_bounded_get_and_put)))'
+filter = 'package(rustfs) & (binary(/^connect_registration$/) | binary(connect_inventory) | binary(connect_perf_drive) | (binary(connect_perf_client) & test(=configured_proxy_and_direct_endpoint_failures_are_distinct)) | (binary(connect_perf_object) & test(/^(real_rustfs_endpoint_and_production_cli_support_bounded_get_and_put|cleanup_uses_a_budget_independent_of_the_measurement_window|oversized_response_is_rejected_without_buffering_it)$/)))'
 threads-required = "num-test-threads"

 # This cancellation probe hashes the serving executable before acquiring its
@@ -336,6 +343,10 @@ setup = 'ecstore-large-stack'
 filter = 'package(rustfs-ecstore) | package(rustfs-s3select-api) | package(rustfs-scanner) | (package(rustfs) & test(/^(app::multipart_usecase::tests::concurrent_completions_share_durable_bucket_quota_reservations|app::object::delete::tests::compressed_delete_requests_update_observed_usage_without_releasing_quota_floor|app::object::internal_put::tests::internal_multipart_roundtrip_completes_and_abort_leaves_nothing|app::object::restore::tests::execute_restore_object_maps_failures_to_typed_s3_errors|storage::access::tests::(delete_object_access_captures_authorized_bucket_incarnation|copy_operations_reject_recreated_source_bucket_after_authorization|request_slot_keeps_bucket_policy_bound_to_its_store))$/))'
 setup = 'ecstore-base-stack'

+[[profile.ci.scripts]]
+filter = 'package(rustfs) & (test(/^admin::handlers::table_catalog::tests::/) | test(/^table_catalog::store::tests::/) | test(=app::object::get::tests::execute_get_object_resumes_from_relocated_pool_without_splicing_body))'
+setup = 'ecstore-base-stack'
+
 [[profile.ci.scripts]]
 filter = 'binary(lifecycle_integration_test) | (package(rustfs) & test(/^app::lifecycle_transition_api_test::/))'
 setup = 'lifecycle-large-stack'
@@ -428,7 +439,7 @@ filter = 'package(rustfs) & (binary(/^embedded.*_test$/) | binary(admin_diagnost
 test-group = 'embedded-test-ports'

 [[profile.ci.overrides]]
-filter = 'package(rustfs) & (binary(/^connect_registration$/) | binary(connect_inventory) | binary(connect_perf_drive) | (binary(connect_perf_object) & test(=real_rustfs_endpoint_and_production_cli_support_bounded_get_and_put)))'
+filter = 'package(rustfs) & (binary(/^connect_registration$/) | binary(connect_inventory) | binary(connect_perf_drive) | (binary(connect_perf_client) & test(=configured_proxy_and_direct_endpoint_failures_are_distinct)) | (binary(connect_perf_object) & test(/^(real_rustfs_endpoint_and_production_cli_support_bounded_get_and_put|cleanup_uses_a_budget_independent_of_the_measurement_window|oversized_response_is_rejected_without_buffering_it)$/)))'
 threads-required = "num-test-threads"

 # Preserve the cancellation probe's startup budget and assertions in CI too.
@@ -98,6 +98,10 @@ runs:
      shell: bash
      run: make script-tests

+    - name: Check object generation protocol model
+      shell: bash
+      run: python3 ./scripts/test_object_generation_protocol_model.py
+
    - name: Check test wiring
      shell: bash
      run: python3 ./scripts/check_test_wiring.py
@@ -88,7 +88,7 @@ jobs:
  # database lookup with a second ignore list to maintain.
  cargo-deny:
    name: Cargo Deny
-    if: github.event_name != 'pull_request' || github.event.action != 'closed'
+    if: github.repository == 'rustfs/rustfs' && (github.event_name != 'pull_request' || github.event.action != 'closed')
    runs-on: sm-standard-2
    timeout-minutes: 20
    steps:
@@ -131,7 +131,7 @@ jobs:

  workflow-pin-report:
    name: Workflow Pin Report
-    if: github.event_name != 'pull_request' || github.event.action != 'closed'
+    if: github.repository == 'rustfs/rustfs' && (github.event_name != 'pull_request' || github.event.action != 'closed')
    runs-on: sm-standard-2
    timeout-minutes: 5
    steps:
@@ -203,7 +203,7 @@ jobs:
    # workflow produces, and until now it was only visible to whoever happened to
    # open the Actions tab. Same ci-8 mechanism coverage.yml and
    # e2e-replication-nightly.yml already use.
-    if: always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure')
+    if: github.repository == 'rustfs/rustfs' && (always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure'))
    runs-on: ubuntu-latest
    timeout-minutes: 10
    permissions:
@@ -440,6 +440,17 @@ jobs:
            fi
          done

+      - name: Verify macOS binary has no dynamic liblzma dependency
+        if: matrix.platform == 'macos'
+        shell: bash
+        run: |
+          binary="target/${{ matrix.target }}/release/rustfs"
+          if otool -L "$binary" | grep -E 'liblzma.*\.dylib'; then
+            echo "macOS release binary must not depend on a dynamic liblzma library" >&2
+            exit 1
+          fi
+          "$binary" --version
+
      - name: Generate the CPU symbol catalog
        if: endsWith(matrix.target, '-linux-gnu') && matrix.features == 'pyroscope'
        shell: bash
@@ -36,7 +36,9 @@ on:
      - "flake.lock"
  pull_request:
    types: [ opened, synchronize, reopened, closed ]
-    branches: [ main ]
+    # gateway/integration carries main's ruleset (rustfs/backlog#2734), so its
+    # pull requests must produce the same required checks.
+    branches: [ main, gateway/integration ]
  merge_group:
    types: [ checks_requested ]
  schedule:
@@ -509,7 +511,7 @@ jobs:
    runs-on: sm-standard-4
    timeout-minutes: 90
    strategy:
-      # Preserve both independent results when one protocol fails, so diagnosis
+      # Preserve independent results when one protocol fails, so diagnosis
      # and a focused fix do not require rebuilding an interrupted sibling.
      fail-fast: false
      matrix:
@@ -518,6 +520,8 @@ jobs:
            flags: "--features swift"
          - name: sftp
            flags: "--features sftp"
+          - name: http3
+            flags: "--features http3,swift"
    env:
      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
    steps:
@@ -78,6 +78,7 @@ concurrency:

 jobs:
  distributed:
+    if: github.repository == 'rustfs/rustfs'
    name: Distributed 4-node 4-disk e2e
    # GitHub-hosted VM: loop and tmpfs mounts work here. sm-standard-4 is an
    # ARC pod and rejects both (`mount -o loop` ENOENT, tmpfs "read-only").
@@ -167,11 +168,13 @@ jobs:
          FILTER: ${{ inputs.filter }}
        run: |
          set -euo pipefail
+          rm -f target/nextest/e2e-distributed/junit.xml
          if [ -n "${FILTER}" ]; then
            python3 scripts/e2e_binary.py run -- cargo nextest run --profile e2e-distributed -p e2e_test -E "${FILTER}"
          else
            python3 scripts/e2e_binary.py run -- cargo nextest run --profile e2e-distributed -p e2e_test --no-tests=fail
          fi
+          test -s target/nextest/e2e-distributed/junit.xml

      - name: Upload distributed e2e diagnostics
        if: always()
@@ -200,7 +203,7 @@ jobs:
  alert-on-failure:
    name: Alert on scheduled failure
    needs: [distributed]
-    if: always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure')
+    if: github.repository == 'rustfs/rustfs' && (always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure'))
    runs-on: ubuntu-latest
    timeout-minutes: 10
    permissions:
@@ -53,6 +53,7 @@ concurrency:

 jobs:
  repl-nightly:
+    if: github.repository == 'rustfs/rustfs'
    name: Replication e2e (nightly)
    # dual-node tests spawn two full rustfs servers each; the extra headroom
    # over the per-PR sm-standard-2 e2e job keeps the parallel process fan-out
@@ -119,6 +120,7 @@ jobs:
          if-no-files-found: ignore

  cluster-nightly:
+    if: github.repository == 'rustfs/rustfs'
    name: Cluster fault e2e (nightly)
    runs-on: sm-standard-4
    timeout-minutes: 90
@@ -168,6 +170,7 @@ jobs:
          if-no-files-found: warn

  protocols-nightly:
+    if: github.repository == 'rustfs/rustfs'
    name: Protocol e2e (nightly)
    runs-on: sm-standard-4
    timeout-minutes: 90
@@ -229,7 +232,7 @@ jobs:
    # Only scheduled runs open/append the tracking issue (backlog#1149 ci-8);
    # manual workflow_dispatch runs stay quiet so a debugging run never files a
    # spurious alert.
-    if: always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure')
+    if: github.repository == 'rustfs/rustfs' && (always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure'))
    runs-on: ubuntu-latest
    timeout-minutes: 10
    permissions:
@@ -11,6 +11,7 @@ concurrency:
  cancel-in-progress: false
 jobs:
  collect:
+    if: github.repository == 'rustfs/rustfs'
    runs-on: smoke-testing
    timeout-minutes: 15
    steps:
@@ -72,9 +72,11 @@ jobs:
  fuzz-build:
    name: Build Fuzz Harness
    if: >
-      (github.event_name == 'pull_request' && github.event.action != 'closed') ||
-      github.event_name == 'schedule' ||
-      github.event_name == 'workflow_dispatch'
+      github.repository == 'rustfs/rustfs' && (
+        (github.event_name == 'pull_request' && github.event.action != 'closed') ||
+        github.event_name == 'schedule' ||
+        github.event_name == 'workflow_dispatch'
+      )
    runs-on: sm-standard-4
    timeout-minutes: 60
    env:
@@ -239,7 +241,7 @@ jobs:
    # `always()` is required: without it this job is skipped when a needed
    # job fails. Alerts only for scheduled (nightly) runs (backlog#1149
    # ci-8); PR and manual dispatch failures are already watched by a human.
-    if: always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure')
+    if: github.repository == 'rustfs/rustfs' && (always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure'))
    runs-on: ubuntu-latest
    timeout-minutes: 10
    permissions:
@@ -0,0 +1,197 @@
+# Copyright 2026 RustFS Team
+# SPDX-License-Identifier: Apache-2.0
+
+name: Gateway Integration Sync
+
+# Merges `main` into the long-lived `gateway/integration` branch every weekday
+# so the s3s -> gateway migration (rustfs/backlog#2734, task T0.1 in #2736)
+# never accumulates merge debt. A clean merge lands as a `--no-ff` merge commit
+# pushed straight to the branch. A conflicting merge is aborted and surfaced as
+# a pull request from `sync/<source>-<date>`, whose head is the un-merged
+# source commit, so the conflict is resolved the same day, in review, with no
+# other change riding along. The run itself stays green in that case: the pull
+# request is the signal. Any other failure (push rejected, pull request not
+# created) fails the run and, on the schedule, files the shared failure issue.
+#
+# Resolving a conflict pull request: merge `origin/gateway/integration` into
+# the sync branch, resolve, push, get the review. Land it with a fast-forward
+# push of the resolved merge commit to `gateway/integration`, not with the
+# squash button: a squash drops the merge ancestry and the same hunks conflict
+# again on the next run.
+#
+# Write scope: only `gateway/integration` and `sync/*` are ever pushed, and
+# both push targets below are literal refs. The pushes use a one-hour
+# installation token of the organisation's own GitHub App (contents: write)
+# because that app is the only non-admin bypass actor of the
+# `gateway/integration` ruleset; GITHUB_TOKEN cannot be a ruleset bypass actor,
+# so it cannot push through the required reviews and checks. The conflict pull
+# request is opened with FLAKE_UPDATE_TOKEN, the PAT that already opens the
+# weekly flake.lock pull requests, because this repository keeps "Allow GitHub
+# Actions to create and approve pull requests" switched off. GITHUB_TOKEN is
+# therefore read-only here.
+#
+# `schedule` fires only from the default branch, so the cron is live once this
+# file exists on `main`. `workflow_dispatch` runs from any ref that carries the
+# file and always acts on `gateway/integration`.
+
+on:
+  schedule:
+    - cron: "0 2 * * 1-5" # Weekdays 02:00 UTC
+  workflow_dispatch:
+    inputs:
+      source_ref:
+        description: >-
+          Branch to merge into gateway/integration. The schedule always merges
+          main; only the conflict acceptance test (rustfs/backlog#2736) sets
+          another branch.
+        required: false
+        default: main
+        type: string
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}
+  cancel-in-progress: false
+
+env:
+  TARGET_BRANCH: gateway/integration
+
+jobs:
+  sync:
+    name: Merge main into gateway/integration
+    if: github.repository == 'rustfs/rustfs'
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - name: Create the branch push token
+        id: push-token
+        uses: actions/create-github-app-token@bcd2ba49218906704ab6c1aa796996da409d3eb1 # v3
+        with:
+          app-id: ${{ vars.CLA_BOT_APP_ID }}
+          private-key: ${{ secrets.CLA_BOT_APP_PRIVATE_KEY }}
+          owner: ${{ github.repository_owner }}
+          repositories: rustfs
+          permission-contents: write
+
+      - name: Checkout gateway/integration
+        uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          # persist-credentials-exempt: this checkout's token IS the push
+          # credential — the merge commit and the sync branch are pushed with it
+          # below. It is a one-hour installation token scoped to contents.
+          ref: gateway/integration
+          fetch-depth: 0
+          token: ${{ steps.push-token.outputs.token }}
+
+      - name: Configure the bot identity
+        run: |
+          git config user.name "rustfs-cla-assistant[bot]"
+          git config user.email "273121464+rustfs-cla-assistant[bot]@users.noreply.github.com"
+
+      - name: Merge the source branch
+        id: merge
+        env:
+          SOURCE_REF: ${{ inputs.source_ref || 'main' }}
+        run: |
+          set -euo pipefail
+          if ! [[ "$SOURCE_REF" =~ ^[A-Za-z0-9][A-Za-z0-9._/-]*$ ]] \
+            || [[ "$SOURCE_REF" == *..* ]] \
+            || [[ "$SOURCE_REF" == "$TARGET_BRANCH" ]]; then
+            echo "::error::refusing source_ref '${SOURCE_REF}'"
+            exit 1
+          fi
+          git fetch origin "refs/heads/${SOURCE_REF}:refs/remotes/origin/${SOURCE_REF}"
+          source_sha=$(git rev-parse "origin/${SOURCE_REF}")
+          target_sha=$(git rev-parse HEAD)
+          {
+            echo "source_ref=${SOURCE_REF}"
+            echo "source_sha=${source_sha}"
+            echo "target_sha=${target_sha}"
+          } >> "$GITHUB_OUTPUT"
+          if git merge --no-ff --no-edit "origin/${SOURCE_REF}"; then
+            if [[ "$(git rev-parse HEAD)" == "$target_sha" ]]; then
+              echo "result=up-to-date" >> "$GITHUB_OUTPUT"
+              echo "\`${TARGET_BRANCH}\` already contains \`${SOURCE_REF}\` (${source_sha}); nothing to push." >> "$GITHUB_STEP_SUMMARY"
+            else
+              echo "result=merged" >> "$GITHUB_OUTPUT"
+            fi
+            exit 0
+          fi
+          git diff --name-only --diff-filter=U > "$RUNNER_TEMP/conflicts.txt"
+          git merge --abort
+          if [[ ! -s "$RUNNER_TEMP/conflicts.txt" ]]; then
+            echo "::error::git merge failed without reporting conflicting paths; this is not a conflict"
+            exit 1
+          fi
+          echo "result=conflict" >> "$GITHUB_OUTPUT"
+
+      - name: Push the merge commit
+        if: steps.merge.outputs.result == 'merged'
+        env:
+          SOURCE_REF: ${{ steps.merge.outputs.source_ref }}
+          SOURCE_SHA: ${{ steps.merge.outputs.source_sha }}
+        run: |
+          set -euo pipefail
+          git push origin HEAD:refs/heads/gateway/integration
+          echo "Merged \`${SOURCE_REF}\` (${SOURCE_SHA}) into \`${TARGET_BRANCH}\` as $(git rev-parse HEAD)." >> "$GITHUB_STEP_SUMMARY"
+
+      - name: Open the conflict pull request
+        if: steps.merge.outputs.result == 'conflict'
+        env:
+          GH_TOKEN: ${{ secrets.FLAKE_UPDATE_TOKEN }}
+          SOURCE_REF: ${{ steps.merge.outputs.source_ref }}
+          SOURCE_SHA: ${{ steps.merge.outputs.source_sha }}
+          TARGET_SHA: ${{ steps.merge.outputs.target_sha }}
+          RUN_URL: ${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }}
+        run: |
+          set -euo pipefail
+          date_utc=$(date -u +%Y-%m-%d)
+          sync_branch="sync/${SOURCE_REF//\//-}-${date_utc}"
+          git push origin "refs/remotes/origin/${SOURCE_REF}:refs/heads/${sync_branch}"
+          {
+            echo "The scheduled merge of \`${SOURCE_REF}\` into \`${TARGET_BRANCH}\` stopped on conflicts and was aborted; \`${TARGET_BRANCH}\` is unchanged."
+            echo
+            echo "This branch is the un-merged \`${SOURCE_REF}\` head. Merge \`origin/${TARGET_BRANCH}\` into it, resolve the files below, push, and get the review. Land the resolved merge commit with a fast-forward push to \`${TARGET_BRANCH}\`; a squash merge drops the merge ancestry and the same hunks conflict again on the next run. Do not include any change that is not conflict resolution."
+            echo
+            echo "- Run: ${RUN_URL}"
+            echo "- Source: \`${SOURCE_REF}\` at ${SOURCE_SHA}"
+            echo "- Target: \`${TARGET_BRANCH}\` at ${TARGET_SHA}"
+            echo
+            echo "## Conflicting files"
+            echo
+            echo "Output of \`git diff --name-only --diff-filter=U\` captured before \`git merge --abort\`:"
+            echo
+            sed 's/^/- /' "$RUNNER_TEMP/conflicts.txt"
+          } > "$RUNNER_TEMP/pr-body.md"
+          title="sync: merge ${SOURCE_REF} ${date_utc} (conflicts)"
+          existing=$(gh pr list --repo "$GITHUB_REPOSITORY" --base "$TARGET_BRANCH" --head "$sync_branch" --state open --json number --jq '.[0].number // empty')
+          if [[ -n "$existing" ]]; then
+            gh pr edit "$existing" --repo "$GITHUB_REPOSITORY" --title "$title" --body-file "$RUNNER_TEMP/pr-body.md"
+            pr_url=$(gh pr view "$existing" --repo "$GITHUB_REPOSITORY" --json url --jq .url)
+          else
+            pr_url=$(gh pr create --repo "$GITHUB_REPOSITORY" --base "$TARGET_BRANCH" --head "$sync_branch" --title "$title" --body-file "$RUNNER_TEMP/pr-body.md")
+          fi
+          echo "Conflicts in \`${SOURCE_REF}\` -> \`${TARGET_BRANCH}\`; pull request: ${pr_url}" >> "$GITHUB_STEP_SUMMARY"
+
+  alert-on-failure:
+    name: Alert on scheduled failure
+    needs: [sync]
+    # Same ci-8 mechanism audit.yml and coverage.yml use: a scheduled run that
+    # could not push or could not open the conflict pull request must not stay
+    # visible only in the Actions tab.
+    if: github.repository == 'rustfs/rustfs' && (always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure'))
+    runs-on: ubuntu-latest
+    timeout-minutes: 10
+    permissions:
+      contents: read
+      issues: write
+    steps:
+      - uses: actions/checkout@f548e57e544e1ff5a4c46bf1e1b8685f8e4a348a # v7
+        with:
+          persist-credentials: false
+      - name: Open or update failure-tracking issue
+        uses: ./.github/actions/schedule-failure-issue
+        with:
+          github-token: ${{ secrets.BACKLOG_ISSUE_TOKEN }}
@@ -43,6 +43,7 @@ env:

 jobs:
  resolve-source:
+    if: github.repository == 'rustfs/rustfs'
    name: Resolve nightly source
    runs-on: sm-standard-2
    timeout-minutes: 10
@@ -496,8 +497,10 @@ jobs:
    name: Alert on scheduled failure
    needs: [build, kms-vault-lane, kms-vault-ha-failover]
    if: >-
-      always() && github.event_name == 'schedule' &&
-      (contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled'))
+      github.repository == 'rustfs/rustfs' && (
+        always() && github.event_name == 'schedule' &&
+        (contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled'))
+      )
    runs-on: ubuntu-latest
    timeout-minutes: 10
    permissions:
@@ -32,6 +32,7 @@ concurrency:

 jobs:
  update-flake:
+    if: github.repository == 'rustfs/rustfs'
    name: Update flake.lock
    runs-on: sm-standard-2
    timeout-minutes: 90
@@ -49,6 +49,7 @@ env:

 jobs:
  warp-ab:
+    if: github.repository == 'rustfs/rustfs'
    name: Warp A/B budget gate
    runs-on: sm-standard-2
    # A normal nightly restores the last successful binary and builds only the
@@ -364,8 +365,10 @@ jobs:
    # composite action already reports cancelled/timed-out jobs in the issue
    # body.
    if: >-
-      always() && github.event_name == 'schedule' &&
-      (contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled'))
+      github.repository == 'rustfs/rustfs' && (
+        always() && github.event_name == 'schedule' &&
+        (contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled'))
+      )
    runs-on: ubuntu-latest
    timeout-minutes: 10
    permissions:
@@ -42,6 +42,7 @@ concurrency:

 jobs:
  check-ephemerality:
+    if: github.repository == 'rustfs/rustfs'
    name: Check runner ephemerality
    runs-on: sm-standard-2
    timeout-minutes: 15
@@ -65,7 +66,7 @@ jobs:
    # Same ci-8 mechanism as coverage.yml, audit.yml and the nightly lanes:
    # scheduled runs file a tracking issue, manual dispatch stays quiet so
    # debugging never produces a spurious alert.
-    if: always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure')
+    if: github.repository == 'rustfs/rustfs' && (always() && github.event_name == 'schedule' && contains(needs.*.result, 'failure'))
    runs-on: ubuntu-latest
    timeout-minutes: 10
    permissions:
@@ -266,7 +266,10 @@ jobs:
        # the previous 45m step cap killed four consecutive runs mid-rebalance
        # with steps 1-7 green. 240m covers the observed pace (rebalance 1/4
        # in ~21m) plus one decommission pass; override per dispatch.
-        timeout-minutes: ${{ inputs.pool_timeout_minutes || 240 }}
+        # fromJSON coerces the string dispatch input to the number that
+        # timeout-minutes requires; the bare || form yielded the string '240'
+        # and the runner rejected the template (Unexpected value '240').
+        timeout-minutes: ${{ fromJSON(inputs.pool_timeout_minutes || '240') }}
        # Case failures keep the run green: the report and the backlog
        # issue manager carry the product signal.
        continue-on-error: true
@@ -583,25 +586,6 @@ jobs:
            echo "pool test script reported a failure (step FAIL verdict or die() abort)" >&2
            failed=1
          fi
-          # A runner-level kill (step timeout) truncates the suite without a
-          # FAIL verdict or an [ERROR] line, so require every expected
-          # [POOL-STEP] marker to be present with PASS - missing markers mean
-          # the run was truncated, not green. This also closes the gap for
-          # standalone dispatches, where the chain driver's evidence
-          # validation does not run.
-          last_step=6
-          if [ "${{ inputs.pools || '3' }}" = "3" ]; then
-            last_step=8
-            if [ "${{ inputs.run_decommission != 'false' }}" = "true" ]; then
-              last_step=9
-            fi
-          fi
-          for n in $(seq 1 "${last_step}"); do
-            grep -q "^\[POOL-STEP\] ${n} .* PASS" "${POOL_ARTIFACT_DIR}/pool-test.log" 2>/dev/null || {
-              echo "pool test log is missing [POOL-STEP] ${n} PASS (expected through step ${last_step}); the run was truncated" >&2
-              failed=1
-            }
-          done
          [ "${failed}" -eq 0 ] || exit 1

      - name: Upload functional report to dashboard
@@ -28,6 +28,7 @@ concurrency:

 jobs:
  check-freshness:
+    if: github.repository == 'rustfs/rustfs'
    name: Check scheduled validation freshness
    # Stay on GitHub-hosted runners: this monitor must keep running when the
    # self-hosted sm-standard-* pool is down, which is exactly when the
@@ -39,6 +39,7 @@ jobs:
  alert-on-incomplete-run:
    name: Alert on incomplete scheduled run
    if: >-
+      github.repository == 'rustfs/rustfs' &&
      github.event.workflow_run.event == 'schedule' &&
      github.event.workflow_run.conclusion != 'success' &&
      github.event.workflow_run.conclusion != 'failure'
@@ -26,6 +26,7 @@ on:

 jobs:
  stale:
+    if: github.repository == 'rustfs/rustfs'
    runs-on: sm-standard-2
    timeout-minutes: 30
    steps:
@@ -14,6 +14,7 @@ concurrency:

 jobs:
  update:
+    if: github.repository == 'rustfs/rustfs'
    runs-on: sm-standard-2
    timeout-minutes: 10
    steps:
@@ -40,6 +40,7 @@ env:

 jobs:
  targets-live:
+    if: github.repository == 'rustfs/rustfs'
    name: PostgreSQL, MySQL, AMQP, and NATS
    runs-on: sm-standard-2
    timeout-minutes: 90
@@ -175,8 +176,10 @@ jobs:
    name: Alert on scheduled failure
    needs: [targets-live]
    if: >-
-      always() && github.event_name == 'schedule' &&
-      (contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled'))
+      github.repository == 'rustfs/rustfs' && (
+        always() && github.event_name == 'schedule' &&
+        (contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled'))
+      )
    runs-on: ubuntu-latest
    timeout-minutes: 10
    permissions:
@@ -18,6 +18,7 @@ members = [
    "crates/audit", # Audit target management system with multi-target fan-out
    "crates/checksums", # client checksums
    "crates/common", # Shared utilities and data structures
+    "crates/capture-authority", # Durable capture authority storage
    "crates/data-usage", # Shared data usage models and algorithms
    "crates/config", # Configuration management
    "crates/connect-symbol-catalog", # Release-only CPU profile symbol catalog generator
@@ -73,7 +74,7 @@ resolver = "3"
 edition = "2024"
 license = "Apache-2.0"
 repository = "https://github.com/rustfs/rustfs"
-rust-version = "1.98.1"
+rust-version = "1.99.0"
 version = "1.0.1"
 homepage = "https://rustfs.com"
 description = "RustFS is a high-performance distributed object storage software built using Rust, one of the most popular languages worldwide. "
@@ -103,6 +104,7 @@ rustfs-config = { path = "./crates/config", version = "1.0.1" }
 rustfs-concurrency = { path = "./crates/concurrency", version = "1.0.1" }
 rustfs-credentials = { path = "crates/credentials", version = "1.0.1" }
 rustfs-crypto = { path = "crates/crypto", version = "1.0.1" }
+rustfs-capture-authority = { path = "crates/capture-authority", version = "1.0.1" }
 rustfs-ecstore = { path = "crates/ecstore", version = "1.0.1" }
 rustfs-filemeta = { path = "crates/filemeta", version = "1.0.1" }
 rustfs-iam = { path = "crates/iam", version = "1.0.1" }
@@ -157,7 +159,7 @@ futures-util = "0.3.34"
 pollster = "1.0.1"
 pulsar = { default-features = false, version = "6.9.0" }
 lapin = { default-features = false, version = "4.12.0" }
-hyper = { version = "1.11.1" }
+hyper = { version = "1.12.0" }
 hyper-rustls = { default-features = false, version = "0.27.10" }
 hyper-util = { version = "0.1.21" }
 http = "1.5.0"
@@ -166,9 +168,9 @@ http-body-util = "0.1.5"
 memchr = "2.8.3"
 minlz = "1.2.3"
 reqwest = "0.13.5"
-rustfs-kafka-async = { version = "1.3.1" }
+rustfs-kafka-async = { version = "1.4.0" }
 socket2 = { version = "0.6.5" }
-tokio = { version = "1.53.1" }
+tokio = { version = "1.53.2" }
 tokio-rustls = { default-features = false, version = "0.26.6" }
 tokio-stream = { version = "0.1.19" }
 tokio-test = "0.4.6"
@@ -178,6 +180,13 @@ tonic-prost = { version = "0.14.6" }
 tonic-prost-build = { version = "0.14.6" }
 tower = { version = "0.5.3" }
 tower-http = { version = "0.7.1" }
+h3 = "0.0.8"
+h3-quinn = "0.0.10"
+quinn = { version = "0.11.11", default-features = false, features = [
+      "futures-io",
+      "runtime-tokio",
+      "rustls-aws-lc-rs",
+] }

 # Serialization and Data Formats
 apache-avro = { version = "0.22.0", features = ["snappy", "zstandard"] }
@@ -219,13 +228,13 @@ x509-parser = "0.18.1"
 sha1 = "0.11.0"
 sha2 = "0.11.0"
 subtle = "2.6"
-zeroize = { version = "1.9.0" }
+zeroize = { version = "1.9.1" }
 proptest = "1"

 # Time and Date
 chrono = { version = "0.4.45" }
 humantime = "2.4.0"
-jiff = { version = "0.2.37" }
+jiff = { version = "0.2.38" }
 time = { version = "0.3.55" }

 # Database
@@ -244,9 +253,9 @@ atoi = "3.1.0"
 atomic_enum = "0.3.0"
 aws-config = { version = "1.12.0" }
 aws-credential-types = { version = "1.3.0" }
-aws-sdk-kms = { default-features = false, version = "1.122.0" }
-aws-sdk-s3 = { default-features = false, version = "1.151.0" }
-aws-sdk-sts = { default-features = false, version = "1.118.0" }
+aws-sdk-kms = { default-features = false, version = "1.123.0" }
+aws-sdk-s3 = { default-features = false, version = "1.152.0" }
+aws-sdk-sts = { default-features = false, version = "1.119.0" }
 aws-smithy-async = { version = "1.3.0" }
 aws-smithy-http-client = { default-features = false, version = "1.5.0" }
 aws-smithy-runtime-api = { version = "1.19.0" }
@@ -279,7 +288,7 @@ highway = { version = "1.3.0" }
 hostname = "0.4.2"
 ipnetwork = { version = "0.21.1" }
 lazy_static = "1.5.1"
-libc = "0.2.189"
+libc = "0.2.190"
 libsystemd = "0.7.2"
 local-ip-address = "0.6.13"
 log = "0.4"
@@ -299,7 +308,7 @@ percent-encoding = "2.3.2"
 # Server-side QR rendering for TOTP enrollment, so neither the console nor the
 # CLI needs its own QR encoder. No default features: the image/render backends
 # pull in an image stack this only needs SVG and text output from.
-qrcode-rs = { version = "2.1.1", default-features = false, features = ["std", "svg"] }
+qrcode-rs = { version = "2.2.0", default-features = false, features = ["std", "svg"] }
 pin-project-lite = "0.2.17"
 pretty_assertions = "1.4.1"
 rand = { version = "0.10.3" }
@@ -307,6 +316,8 @@ ratelimit = "2.0.0"
 rayon = "1.12.0"
 rustfs-erasure-codec = { version = "9.0.0" }
 reed-solomon-simd = "3.1.0"
+openraft = { version = "=0.9.25", default-features = false, features = ["serde"] }
+redb = "=4.3.0"
 regex = { version = "1.13.1" }
 rumqttc = { package = "rumqttc-next", version = "0.34.0" }
 redis = { version = "1.7.1" }
@@ -314,15 +325,15 @@ rustify = { version = "0.7", default-features = false }
 rustix = { version = "1.1.5" }
 rust-embed = { version = "8.12.0" }
 rustc-hash = { version = "2.1.3" }
-s3s = { version = "0.17.0", features = ["minio"] }
-s3s-sigv4 = "0.17.0"
+s3s = { git = "https://github.com/s3s-project/s3s", rev = "5761ddfe4c509fcac87e3b99fd5bd9e176b442c0", features = ["minio"] }
+s3s-sigv4 = { git = "https://github.com/s3s-project/s3s", rev = "5761ddfe4c509fcac87e3b99fd5bd9e176b442c0" }
 serial_test = "4.0.1"
 shadow-rs = { default-features = false, version = "2.0.0" }
 siphasher = "1.0.4"
 smallvec = { version = "1.16.2" }
 compact_str = "0.10.0"
 snap = "1.1.2"
-starshard = { version = "2.3.0" }
+starshard = { version = "2.4.0" }
 strum = { version = "0.28.0" }
 sysinfo = "0.39.6"
 temp-env = "0.3.6"
@@ -338,7 +349,7 @@ tracing-subscriber = { version = "0.3.23" }
 transform-stream = "0.3.1"
 url = "2.5.8"
 urlencoding = "2.1.3"
-uuid = { version = "1.26.1" }
+uuid = { version = "1.27.0" }
 vaultrs = { version = "0.8.0" }
 tar = "0.4.46"
 walkdir = "2.5.0"
@@ -352,7 +363,7 @@ zstd = "0.14.0"
 # Observability and Metrics
 metrics = "0.24.6"
 metrics-util = "0.20"
-dial9 = { version = "0.5.2", default-features = false }
+dial9 = { version = "0.5.3", default-features = false }
 opentelemetry = { version = "0.33.0" }
 opentelemetry-appender-tracing = { version = "0.33.0" }
 opentelemetry-otlp = { version = "0.33.0" }
@@ -367,7 +378,7 @@ libunftp = { version = "0.23.1" }
 unftp-core = "0.1.0"
 suppaftp = { version = "12.1.1" }
 rcgen = { version = "0.14.10", default-features = false, features = ["aws_lc_rs", "crypto", "pem"] }
-russh = { version = "0.63.3" }
+russh = { version = "0.64.1" }
 russh-sftp = "3.0.1"

 # WebDAV
@@ -376,9 +387,9 @@ dav-server = "0.11.0"
 # Performance Analysis and Memory Profiling
 rustfs-mimalloc = { version = "0.5.6" }
 # Preserve Unicode focus filters until rustfs/backlog#2302 is resolved.
-hotpath = { version = "0.28.0", default-features = false }
+hotpath = { version = "0.28.5", default-features = false }
 # Snapshot testing for output format regression detection
-insta = { version = "1.48" }
+insta = { version = "1.49" }

 [workspace.metadata.cargo-shear]
 ignored = ["hotpath", "rustfs"]
@@ -0,0 +1,28 @@
+[package]
+name = "rustfs-capture-authority"
+version.workspace = true
+edition.workspace = true
+license.workspace = true
+rust-version.workspace = true
+repository.workspace = true
+description = "Durable local storage for capture authority"
+
+[dependencies]
+openraft = { workspace = true, features = ["storage-v2"] }
+redb = { workspace = true }
+sha2 = { workspace = true }
+serde = { workspace = true, features = ["derive"] }
+thiserror = { workspace = true }
+tokio = { workspace = true, features = ["net", "io-util", "time", "sync", "rt"] }
+url = { workspace = true }
+tempfile = { workspace = true }
+serde_json = { workspace = true }
+
+[target.'cfg(any(target_os = "linux", target_os = "macos"))'.dependencies]
+rustix = { workspace = true, features = ["process"] }
+
+[dev-dependencies]
+tokio = { workspace = true, features = ["macros", "rt-multi-thread"] }
+
+[lints]
+workspace = true
@@ -0,0 +1,531 @@
+use super::*;
+use crate::application::{APPLY_RESULTS, EVENTS, OPERATIONS, VERSIONS};
+use crate::tests::CrashBackend;
+use redb::ReadableTableMetadata;
+use std::{
+    io::{BufRead, BufReader, Write},
+    process::{Command, Stdio},
+    sync::mpsc,
+    time::Duration,
+};
+fn identity() -> NodeIdentity {
+    NodeIdentity {
+        cluster_id: [7; 16],
+        node_id: 0,
+    }
+}
+fn binding() -> CaptureBinding {
+    CaptureBinding {
+        revision: 9,
+        binding_id: [8; 16],
+        target_id: [9; 16],
+    }
+}
+fn command() -> DecideCreated {
+    DecideCreated {
+        operation_id: [1; 16],
+        object: ObjectIdentity {
+            bucket_incarnation: [2; 16],
+            key: vec![0, 255, 3],
+        },
+        expected_head: None,
+        prepared: PreparedIdentity {
+            version_id: [3; 16],
+            preparation_id: [4; 16],
+            content_digest: [5; 32],
+            content_length: 123,
+        },
+        binding_revision: 9,
+    }
+}
+fn id(index: u64) -> LogId {
+    LogId {
+        term: 8,
+        leader_node: 0,
+        index,
+    }
+}
+fn append(store: &mut Store, index: u64, c: &DecideCreated) {
+    store
+        .append(&[LogEntry {
+            id: id(index),
+            payload: c.encode().expect("encode command"),
+        }])
+        .expect("append command");
+}
+fn apply(store: &mut Store, index: u64, c: &DecideCreated) -> CreatedResult {
+    append(store, index, c);
+    store.save_committed(id(index)).expect("commit log");
+    store.apply_next_created(id(index)).expect("apply command")
+}
+fn success(c: &DecideCreated, index: u64) -> CreatedResult {
+    CreatedResult::Created {
+        version_id: c.prepared.version_id,
+        event_id: c.operation_id,
+        decision_log_id: id(index),
+    }
+}
+fn complete(store: &Store) {
+    let c = command();
+    assert_eq!(store.capture_binding().expect("binding"), Some(binding()));
+    assert_eq!(
+        store.created_head(&c.object).expect("head"),
+        Some(c.prepared.version_id),
+        "U11 acknowledged head lost"
+    );
+    assert_eq!(
+        store.created_version(&c.object, c.prepared.version_id).expect("version"),
+        Some(CreatedVersion {
+            prepared: c.prepared.clone(),
+            decision_log_id: id(0),
+            binding_revision: 9
+        })
+    );
+    assert_eq!(
+        store.created_event(c.operation_id).expect("event"),
+        Some(CreatedEvent {
+            object: c.object.clone(),
+            prepared: c.prepared.clone(),
+            decision_log_id: id(0),
+            binding: binding()
+        }),
+        "U11 complete event missing"
+    );
+    assert_eq!(
+        store.created_operation(c.operation_id).expect("operation"),
+        Some(CreatedOperation {
+            canonical_command: c.encode().expect("encode"),
+            result: success(&c, 0)
+        })
+    );
+    assert_eq!(
+        store.created_apply_result(0).expect("result"),
+        Some(CreatedApplyResult {
+            log_id: id(0),
+            result: success(&c, 0)
+        })
+    );
+    assert_eq!(store.last_applied().expect("last applied"), Some(id(0)));
+}
+fn fixture() -> (tempfile::TempDir, Store) {
+    let dir = tempfile::tempdir().expect("tempdir");
+    let mut store = Store::initialize(dir.path().join("db"), identity()).expect("initialize");
+    store.pin_capture_binding(binding()).expect("pin");
+    (dir, store)
+}
+fn reopen(path: &std::path::Path) -> Store {
+    let result = Store::open(path, identity());
+    assert!(result.is_ok(), "U11 complete reopen: {:?}", result.as_ref().err());
+    result.expect("asserted successful reopen")
+}
+#[test]
+fn created_codec_is_canonical_and_strict() {
+    let c = command();
+    let b = c.encode().expect("encode");
+    let mut golden = b"CRTD\x01".to_vec();
+    golden.extend([1; 16]);
+    golden.extend([2; 16]);
+    golden.extend(3u32.to_be_bytes());
+    golden.extend([0, 255, 3]);
+    golden.push(0);
+    golden.extend([3; 16]);
+    golden.extend([4; 16]);
+    golden.extend([5; 32]);
+    golden.extend(123u64.to_be_bytes());
+    golden.extend(9u64.to_be_bytes());
+    assert_eq!(b, golden);
+    assert_eq!(DecideCreated::decode(&b).expect("decode"), c);
+    let mut variants = Vec::new();
+    let mut v = c.clone();
+    v.operation_id = [6; 16];
+    variants.push(v);
+    let mut v = c.clone();
+    v.object.bucket_incarnation = [6; 16];
+    variants.push(v);
+    let mut v = c.clone();
+    v.object.key.push(4);
+    variants.push(v);
+    let mut v = c.clone();
+    v.expected_head = Some([6; 16]);
+    variants.push(v);
+    let mut v = c.clone();
+    v.prepared.version_id = [6; 16];
+    variants.push(v);
+    let mut v = c.clone();
+    v.prepared.preparation_id = [6; 16];
+    variants.push(v);
+    let mut v = c.clone();
+    v.prepared.content_digest = [6; 32];
+    variants.push(v);
+    let mut v = c.clone();
+    v.prepared.content_length += 1;
+    variants.push(v);
+    let mut v = c.clone();
+    v.binding_revision += 1;
+    variants.push(v);
+    for v in variants {
+        assert_ne!(v.encode().expect("variant"), b)
+    }
+    for n in 0..b.len() {
+        assert!(DecideCreated::decode(&b[..n]).is_err())
+    }
+    let mut bad = b.clone();
+    bad[4] = 2;
+    assert!(DecideCreated::decode(&bad).is_err());
+    let mut bad = b.clone();
+    bad[44] = 2;
+    assert!(DecideCreated::decode(&bad).is_err());
+    let mut bad = b.clone();
+    bad.push(0);
+    assert!(DecideCreated::decode(&bad).is_err());
+    let mut bad = b;
+    bad[37..41].copy_from_slice(&4097u32.to_be_bytes());
+    assert!(DecideCreated::decode(&bad).is_err());
+    let mut bad = c.clone();
+    bad.operation_id = [0; 16];
+    assert!(bad.encode().is_err());
+    bad = c;
+    bad.object.key = vec![0; MAX_OBJECT_KEY_LENGTH + 1];
+    assert!(bad.encode().is_err());
+}
+#[test]
+fn created_committed_order_and_identity_are_enforced() {
+    let (_dir, mut s) = fixture();
+    let c = command();
+    append(&mut s, 0, &c);
+    append(&mut s, 1, &c);
+    assert!(s.apply_next_created(id(0)).is_err());
+    assert_eq!(s.last_applied().expect("no apply"), None);
+    s.save_committed(id(1)).expect("committed");
+    assert!(matches!(s.apply_next_created(id(1)), Err(StoreError::InvalidApplySequence)));
+    let mut wrong = id(0);
+    wrong.term += 1;
+    assert!(s.apply_next_created(wrong).is_err());
+    wrong = id(0);
+    wrong.leader_node += 1;
+    assert!(s.apply_next_created(wrong).is_err());
+    assert_eq!(s.apply_next_created(id(0)).expect("first"), success(&c, 0));
+    complete(&s);
+    assert_eq!(s.apply_next_created(id(0)).expect("retry"), success(&c, 0));
+    complete(&s);
+    assert!(s.apply_next_created(wrong).is_err());
+    assert_eq!(s.apply_next_created(id(1)).expect("duplicate log"), success(&c, 0));
+    assert_eq!(s.last_applied().expect("last"), Some(id(1)));
+    assert_eq!(
+        s.created_apply_result(1).expect("result"),
+        Some(CreatedApplyResult {
+            log_id: id(1),
+            result: success(&c, 0)
+        })
+    );
+    let tx = s.database.begin_read().expect("read");
+    assert_eq!(tx.open_table(EVENTS).expect("events").len().expect("len"), 1);
+}
+#[test]
+fn created_head_version_event_and_dedup_reopen_together() {
+    let (dir, mut s) = fixture();
+    let c = command();
+    assert_eq!(apply(&mut s, 0, &c), success(&c, 0));
+    drop(s);
+    let mut s = reopen(&dir.path().join("db"));
+    complete(&s);
+    assert_eq!(apply(&mut s, 1, &c), success(&c, 0));
+    let mut changed = c.clone();
+    changed.prepared.content_digest = [6; 32];
+    assert_eq!(
+        apply(&mut s, 2, &changed),
+        CreatedResult::OperationConflict,
+        "U11 full canonical operation conflict"
+    );
+    assert_eq!(
+        s.created_operation(c.operation_id)
+            .expect("original")
+            .expect("operation")
+            .result,
+        success(&c, 0)
+    );
+    let mut stale = c.clone();
+    stale.operation_id = [6; 16];
+    stale.prepared.version_id = [6; 16];
+    assert_eq!(
+        apply(&mut s, 3, &stale),
+        CreatedResult::HeadMismatch {
+            actual_head: Some(c.prepared.version_id)
+        }
+    );
+    let mut reuse = c.clone();
+    reuse.operation_id = [7; 16];
+    reuse.expected_head = Some(c.prepared.version_id);
+    assert_eq!(apply(&mut s, 4, &reuse), CreatedResult::VersionConflict);
+    let mut replacement = stale.clone();
+    replacement.operation_id = [10; 16];
+    replacement.expected_head = Some(c.prepared.version_id);
+    assert_eq!(apply(&mut s, 5, &replacement), success(&replacement, 5));
+    assert_eq!(s.created_head(&c.object).expect("head"), Some(replacement.prepared.version_id));
+    assert_eq!(
+        s.created_version(&c.object, c.prepared.version_id)
+            .expect("old version")
+            .expect("version")
+            .decision_log_id,
+        id(0)
+    );
+    assert!(s.created_event(c.operation_id).expect("old event").is_some());
+    assert!(s.created_event(replacement.operation_id).expect("new event").is_some());
+    assert!(s.created_event(stale.operation_id).expect("rejection event").is_none());
+    assert_eq!(
+        apply(&mut s, 6, &stale),
+        CreatedResult::HeadMismatch {
+            actual_head: Some(c.prepared.version_id)
+        }
+    );
+    drop(s);
+    let s = reopen(&dir.path().join("db"));
+    assert_eq!(s.last_applied().expect("last"), Some(id(6)));
+    let tx = s.database.begin_read().expect("read");
+    assert_eq!(tx.open_table(EVENTS).expect("events").len().expect("len"), 2);
+}
+#[test]
+fn created_binding_is_pinned_and_not_caller_selected() {
+    let (dir, mut s) = fixture();
+    let mut bad = command();
+    bad.binding_revision = 10;
+    assert_eq!(apply(&mut s, 0, &bad), CreatedResult::BindingMismatch);
+    assert!(s.created_head(&bad.object).expect("head").is_none());
+    assert!(s.created_event(bad.operation_id).expect("event").is_none());
+    assert_eq!(
+        s.created_operation(bad.operation_id)
+            .expect("operation")
+            .expect("saved")
+            .result,
+        CreatedResult::BindingMismatch
+    );
+    s.pin_capture_binding(binding()).expect("idempotent pin");
+    let mut changed = binding();
+    changed.target_id = [10; 16];
+    assert!(matches!(s.pin_capture_binding(changed), Err(StoreError::BindingConflict)));
+    drop(s);
+    let s = reopen(&dir.path().join("db"));
+    assert_eq!(s.capture_binding().expect("binding"), Some(binding()));
+}
+#[test]
+fn created_schema_compatibility_and_corruption_are_checked() {
+    let dir = tempfile::tempdir().expect("dir");
+    let path = dir.path().join("db");
+    let s = Store::initialize(&path, identity()).expect("schema1");
+    assert_eq!(s.capture_binding().expect("unmigrated"), None);
+    drop(s);
+    let mut s = reopen(&path);
+    assert_eq!(s.last_applied().expect("no apply"), None);
+    append(&mut s, 0, &command());
+    s.save_committed(id(0)).expect("schema2");
+    drop(s);
+    let mut s = reopen(&path);
+    assert_eq!(s.capture_binding().expect("schema2 unchanged"), None);
+    s.pin_capture_binding(binding()).expect("explicit upgrade");
+    drop(s);
+    let mut s = reopen(&path);
+    assert_eq!(s.capture_binding().expect("schema3"), Some(binding()));
+    s.apply_next_created(id(0)).expect("apply");
+    drop(s);
+    let db = Database::open(&path).expect("raw db");
+    let tx = db.begin_write().expect("write corruption");
+    tx.open_table(EVENTS)
+        .expect("events")
+        .remove(command().operation_id.as_slice())
+        .expect("remove reference");
+    tx.commit().expect("commit corruption");
+    drop(db);
+    let result = Store::open(&path, identity());
+    assert!(
+        matches!(result, Err(StoreError::CorruptRecord)),
+        "U11 missing event rejected before expect"
+    );
+    let (dir, mut s) = fixture();
+    assert_eq!(s.read_committed().expect("empty committed"), None);
+    append(&mut s, 0, &command());
+    s.save_committed(id(0)).expect("committed schema3");
+    assert_eq!(s.capture_binding().expect("no downgrade"), Some(binding()));
+    drop(s);
+    let path = dir.path().join("db");
+    let db = Database::open(&path).expect("raw db");
+    let tx = db.begin_write().expect("write");
+    let mut marker = vec![1];
+    marker.extend(encode_id(id(0)));
+    tx.open_table(META)
+        .expect("meta")
+        .insert("last_applied", marker.as_slice())
+        .expect("corrupt applied");
+    tx.commit().expect("commit");
+    drop(db);
+    assert!(
+        matches!(Store::open(&path, identity()), Err(StoreError::CorruptRecord)),
+        "U11 last applied missing result rejected"
+    );
+    let (dir, s) = fixture();
+    drop(s);
+    let path = dir.path().join("db");
+    let db = Database::open(&path).expect("raw pinned db");
+    let tx = db.begin_write().expect("missing binding write");
+    tx.open_table(META)
+        .expect("meta")
+        .remove("capture_binding")
+        .expect("remove binding");
+    tx.commit().expect("binding corruption commit");
+    drop(db);
+    assert!(
+        matches!(Store::open(&path, identity()), Err(StoreError::CorruptRecord)),
+        "U11 missing pinned binding rejected"
+    );
+
+    let (dir, mut s) = fixture();
+    let c = command();
+    assert_eq!(apply(&mut s, 0, &c), success(&c, 0));
+    assert_eq!(apply(&mut s, 1, &c), success(&c, 0));
+    drop(s);
+    let path = dir.path().join("db");
+    let db = Database::open(&path).expect("raw duplicate db");
+    let tx = db.begin_write().expect("retarget decision write");
+    let mut key = created::object_bytes(&c.object).expect("object key");
+    key.extend(c.prepared.version_id);
+    let mut version = Vec::new();
+    created::put_prepared(&mut version, &c.prepared);
+    version.extend(encode_id(id(1)));
+    version.extend(binding().revision.to_be_bytes());
+    tx.open_table(VERSIONS)
+        .expect("versions")
+        .insert(key.as_slice(), version.as_slice())
+        .expect("retarget version");
+    let mut event = created::object_bytes(&c.object).expect("event object");
+    created::put_prepared(&mut event, &c.prepared);
+    event.extend(encode_id(id(1)));
+    event.extend(created::binding_bytes(&binding()).expect("binding bytes"));
+    tx.open_table(EVENTS)
+        .expect("events")
+        .insert(c.operation_id.as_slice(), event.as_slice())
+        .expect("retarget event");
+    let canonical = c.encode().expect("canonical");
+    let mut operation = u64::try_from(canonical.len()).expect("length").to_be_bytes().to_vec();
+    operation.extend(canonical);
+    operation.extend(created::result_bytes(&success(&c, 1)));
+    tx.open_table(OPERATIONS)
+        .expect("operations")
+        .insert(c.operation_id.as_slice(), operation.as_slice())
+        .expect("retarget original response");
+    for index in 0..=1 {
+        let mut result = encode_id(id(index));
+        result.extend(created::result_bytes(&success(&c, 1)));
+        tx.open_table(APPLY_RESULTS)
+            .expect("results")
+            .insert(index, result.as_slice())
+            .expect("retarget duplicate response");
+    }
+    tx.commit().expect("coherent retarget commit");
+    drop(db);
+    assert!(
+        matches!(Store::open(&path, identity()), Err(StoreError::CorruptRecord)),
+        "U11 original decision cannot retarget duplicate log"
+    );
+}
+#[test]
+fn created_process_child() {
+    if let Some(path) = std::env::var_os("U11_CHILD_DB") {
+        let mut s = Store::initialize(path, identity()).expect("child init");
+        s.pin_capture_binding(binding()).expect("child pin");
+        assert_eq!(apply(&mut s, 0, &command()), success(&command(), 0));
+        complete(&s);
+        println!("U11_ACK");
+        std::io::stdout().flush().expect("flush ACK");
+        loop {
+            std::thread::park()
+        }
+    }
+    let (_dir, mut s) = fixture();
+    assert_eq!(apply(&mut s, 0, &command()), success(&command(), 0));
+    complete(&s);
+}
+#[test]
+fn created_acknowledged_apply_survives_child_kill() {
+    let dir = tempfile::tempdir().expect("dir");
+    let path = dir.path().join("db");
+    let mut child = Command::new(std::env::current_exe().expect("test executable"))
+        .args(["--exact", "application_tests::created_process_child", "--nocapture"])
+        .env("U11_CHILD_DB", &path)
+        .stdout(Stdio::piped())
+        .stderr(Stdio::inherit())
+        .spawn()
+        .expect("spawn");
+    let stdout = child.stdout.take().expect("stdout");
+    let (tx, rx) = mpsc::channel();
+    let reader = std::thread::spawn(move || {
+        for line in BufReader::new(stdout).lines() {
+            let line = line.expect("child line");
+            if line.ends_with("U11_ACK") {
+                let _ = tx.send(());
+                break;
+            }
+        }
+    });
+    let acknowledged = rx.recv_timeout(Duration::from_secs(20));
+    child.kill().expect("kill child");
+    child.wait().expect("wait child");
+    reader.join().expect("reader cleanup");
+    assert!(acknowledged.is_ok(), "U11 bounded child ACK missing");
+    complete(&reopen(&path));
+}
+#[test]
+fn created_power_cut_before_and_after_apply_flush() {
+    let (dir, mut s) = fixture();
+    append(&mut s, 0, &command());
+    s.save_committed(id(0)).expect("committed");
+    drop(s);
+    let path = dir.path().join("db");
+    let backend = CrashBackend::new(&path);
+    let mut s = Store::from_database(backend.database(), identity()).expect("backend");
+    assert_eq!(s.apply_next_created(id(0)).expect("apply ACK"), success(&command(), 0));
+    backend.cut();
+    drop(s);
+    complete(&reopen(&path));
+    for sync in 1..=2 {
+        for after in [false, true] {
+            let (dir, mut s) = fixture();
+            append(&mut s, 0, &command());
+            s.save_committed(id(0)).expect("committed");
+            drop(s);
+            let path = dir.path().join("db");
+            let backend = CrashBackend::new(&path);
+            let mut s = Store::from_database(backend.database(), identity()).expect("backend");
+            let (entered, resume) = backend.pause_sync(sync, after);
+            let writer = std::thread::spawn(move || {
+                let result = s.apply_next_created(id(0));
+                (s, result)
+            });
+            entered.wait();
+            backend.cut();
+            resume.wait();
+            let (s, result) = writer.join().expect("writer");
+            drop(s);
+            let reopened = Store::open(&path, identity());
+            assert!(
+                reopened.is_ok(),
+                "U11 apply cut must reopen whole old/new state sync={sync} after={after}: {:?}",
+                reopened.as_ref().err()
+            );
+            let s = reopened.expect("asserted reopen");
+            if s.last_applied().expect("last").is_some() {
+                complete(&s)
+            } else {
+                assert!(result.is_err());
+                assert_eq!(s.created_head(&command().object).expect("old head"), None);
+                assert_eq!(
+                    s.created_version(&command().object, command().prepared.version_id)
+                        .expect("old version"),
+                    None
+                );
+                assert_eq!(s.created_event(command().operation_id).expect("old event"), None);
+                assert_eq!(s.created_operation(command().operation_id).expect("old op"), None);
+                assert_eq!(s.created_apply_result(0).expect("old result"), None)
+            }
+            eprintln!("U11 cut boundary sync={sync} after={after}");
+        }
+    }
+}
@@ -0,0 +1,579 @@
+//! Immutable checkpoint state and disk-backed validation of the retained suffix.
+use crate::{
+    CaptureMembership, CreatedApplyResult, CreatedEvent, CreatedResult, CreatedVersion, LogId, StoreError, application,
+    created::{DecideCreated, Decoder, object_bytes, result_bytes},
+    native_checkpoint::{CheckpointLimits, recovery_scratch},
+    snapshot,
+};
+use redb::{ReadTransaction, ReadableDatabase, ReadableTable, ReadableTableMetadata, TableDefinition};
+
+pub(crate) const META_KEY: &str = "checkpoint_baseline";
+pub(crate) const PURGED_KEY: &str = "checkpoint_purged";
+pub(crate) const HEADS: TableDefinition<&[u8], &[u8]> = TableDefinition::new("checkpoint_heads_v1");
+pub(crate) const VERSIONS: TableDefinition<&[u8], &[u8]> = TableDefinition::new("checkpoint_versions_v1");
+pub(crate) const EVENTS: TableDefinition<&[u8], &[u8]> = TableDefinition::new("checkpoint_events_v1");
+pub(crate) const OPERATIONS: TableDefinition<&[u8], &[u8]> = TableDefinition::new("checkpoint_operations_v1");
+const FIRST_DECISIONS: TableDefinition<u64, &[u8]> = TableDefinition::new("checkpoint_validation_decisions");
+const VERSION_OPERATIONS: TableDefinition<&[u8], &[u8]> = TableDefinition::new("checkpoint_validation_versions");
+const METADATA_BYTES: usize = 1024 * 1024;
+
+#[derive(Clone, Debug)]
+pub(crate) struct BaselineMetadata {
+    pub(crate) generation: u64,
+    pub(crate) header: snapshot::Header,
+    pub(crate) limits: CheckpointLimits,
+}
+
+struct EmptySink;
+impl snapshot::StagingSink for EmptySink {
+    fn row(&mut self, _: snapshot::Section, _: &[u8], _: &[u8]) -> std::io::Result<()> {
+        Err(std::io::Error::other("checkpoint metadata contains rows"))
+    }
+}
+fn metadata_limits() -> snapshot::Limits {
+    snapshot::Limits {
+        total_bytes: 1024 * 1024,
+        membership_bytes: 512 * 1024,
+        key_bytes: 0,
+        value_bytes: 0,
+        records: 0,
+    }
+}
+pub(crate) fn encode_metadata(metadata: &BaselineMetadata) -> Result<Vec<u8>, StoreError> {
+    metadata.limits.validate()?;
+    if metadata.generation == 0 {
+        return Err(StoreError::CorruptRecord);
+    }
+    let mut bytes = b"CBAS\x01".to_vec();
+    for value in [
+        metadata.generation,
+        metadata.limits.snapshot.total_bytes,
+        metadata.limits.snapshot.membership_bytes,
+        metadata.limits.snapshot.key_bytes,
+        metadata.limits.snapshot.value_bytes,
+        metadata.limits.snapshot.records,
+        metadata.limits.staging_bytes,
+    ] {
+        bytes.extend(value.to_be_bytes());
+    }
+    let frame = snapshot::Encoder::new(Vec::new(), &metadata.header, [0; 4], metadata_limits())
+        .and_then(snapshot::Encoder::finish)
+        .map_err(|_| StoreError::CorruptRecord)?;
+    bytes.extend(frame);
+    if bytes.len() > METADATA_BYTES {
+        return Err(StoreError::CorruptRecord);
+    }
+    Ok(bytes)
+}
+pub(crate) fn read_metadata(
+    meta: &impl ReadableTable<&'static str, &'static [u8]>,
+) -> Result<Option<BaselineMetadata>, StoreError> {
+    let schema = crate::number(required(meta.get("schema")?)?.value())?;
+    let stored = meta.get(META_KEY)?;
+    if schema != 7 {
+        if stored.is_some() || meta.get(PURGED_KEY)?.is_some() {
+            return Err(StoreError::CorruptRecord);
+        }
+        return Ok(None);
+    }
+    let value = required(stored)?;
+    let bytes = value.value();
+    if bytes.len() > METADATA_BYTES || bytes.get(..5) != Some(b"CBAS\x01") {
+        return Err(StoreError::CorruptRecord);
+    }
+    let mut numbers = [0u64; 7];
+    for (index, number) in numbers.iter_mut().enumerate() {
+        let start = 5 + index * 8;
+        *number = crate::number(bytes.get(start..start + 8).ok_or(StoreError::CorruptRecord)?)?;
+    }
+    let limits = CheckpointLimits {
+        snapshot: snapshot::Limits {
+            total_bytes: numbers[1],
+            membership_bytes: numbers[2],
+            key_bytes: numbers[3],
+            value_bytes: numbers[4],
+            records: numbers[5],
+        },
+        staging_bytes: numbers[6],
+    };
+    limits.validate()?;
+    let header = snapshot::decode(bytes.get(61..).ok_or(StoreError::CorruptRecord)?, &mut EmptySink, metadata_limits())
+        .map_err(|_| StoreError::CorruptRecord)?;
+    let metadata = BaselineMetadata {
+        generation: numbers[0],
+        header,
+        limits,
+    };
+    if encode_metadata(&metadata)?.as_slice() != bytes {
+        return Err(StoreError::CorruptRecord);
+    }
+    Ok(Some(metadata))
+}
+pub(crate) fn read_baseline_cut(meta: &impl ReadableTable<&'static str, &'static [u8]>) -> Result<Option<LogId>, StoreError> {
+    Ok(read_metadata(meta)?.and_then(|metadata| metadata.header.applied))
+}
+pub(crate) fn encode_purged(id: LogId) -> Vec<u8> {
+    let mut bytes = b"CPUR\x01".to_vec();
+    bytes.extend(crate::encode_id(id));
+    bytes
+}
+pub(crate) fn read_purged(meta: &impl ReadableTable<&'static str, &'static [u8]>) -> Result<Option<LogId>, StoreError> {
+    let cut = read_baseline_cut(meta)?;
+    let Some(value) = meta.get(PURGED_KEY)? else {
+        return Ok(None);
+    };
+    let bytes = value.value();
+    if bytes.len() != 29 || bytes.get(..5) != Some(b"CPUR\x01") {
+        return Err(StoreError::CorruptRecord);
+    }
+    let id = crate::decode_id(&bytes[5..])?;
+    at_or_before(id, required(cut)?)?;
+    Ok(Some(id))
+}
+fn at_or_before(id: LogId, cut: LogId) -> Result<(), StoreError> {
+    if id.index > cut.index || (id.index == cut.index && id != cut) {
+        return Err(StoreError::CorruptRecord);
+    }
+    Ok(())
+}
+fn required<T>(value: Option<T>) -> Result<T, StoreError> {
+    value.ok_or(StoreError::CorruptRecord)
+}
+struct Budget {
+    limits: snapshot::Limits,
+    bytes: u64,
+    records: u64,
+}
+impl Budget {
+    fn new(limits: snapshot::Limits) -> Self {
+        Self {
+            limits,
+            bytes: 0,
+            records: 0,
+        }
+    }
+    fn row(&mut self, key: &[u8], value: &[u8]) -> Result<(), StoreError> {
+        let key = u64::try_from(key.len()).map_err(|_| StoreError::CorruptRecord)?;
+        let value = u64::try_from(value.len()).map_err(|_| StoreError::CorruptRecord)?;
+        if key > 1024 * 1024 || value > 1024 * 1024 {
+            return Err(StoreError::CorruptRecord);
+        }
+        self.bytes = self
+            .bytes
+            .checked_add(key)
+            .and_then(|n| n.checked_add(value))
+            .ok_or(StoreError::ResourceLimit)?;
+        self.records = self.records.checked_add(1).ok_or(StoreError::ResourceLimit)?;
+        if key > self.limits.key_bytes
+            || value > self.limits.value_bytes
+            || self.bytes > self.limits.total_bytes
+            || self.records > self.limits.records
+        {
+            return Err(StoreError::ResourceLimit);
+        }
+        Ok(())
+    }
+}
+
+pub(crate) fn resource_error(error: StoreError) -> StoreError {
+    use std::error::Error;
+    let mut cause: Option<&(dyn Error + 'static)> = Some(&error);
+    while let Some(current) = cause {
+        if current
+            .downcast_ref::<std::io::Error>()
+            .is_some_and(|io| io.kind() == std::io::ErrorKind::StorageFull)
+        {
+            return StoreError::ResourceLimit;
+        }
+        cause = current.source();
+    }
+    error
+}
+
+/// Validate relationships, preserving negative history rather than re-evaluating it.
+/// The caller separately establishes local capture or authenticated installation authority.
+pub(crate) fn validate_baseline_with_budget(
+    tx: &ReadTransaction,
+    metadata: &BaselineMetadata,
+    index_max: u64,
+) -> Result<(), StoreError> {
+    validate_baseline_inner(tx, metadata, index_max).map_err(resource_error)
+}
+
+fn validate_baseline_inner(tx: &ReadTransaction, metadata: &BaselineMetadata, index_max: u64) -> Result<(), StoreError> {
+    encode_metadata(metadata)?;
+    let Some(cut) = metadata.header.applied else {
+        if metadata.header.membership != CaptureMembership::new(None, vec![], std::collections::BTreeMap::new())? {
+            return Err(StoreError::CorruptRecord);
+        }
+        for definition in [HEADS, VERSIONS, EVENTS, OPERATIONS] {
+            if tx.open_table(definition)?.len()? != 0 {
+                return Err(StoreError::CorruptRecord);
+            }
+        }
+        return Ok(());
+    };
+    let heads = tx.open_table(HEADS)?;
+    let versions = tx.open_table(VERSIONS)?;
+    let events = tx.open_table(EVENTS)?;
+    let operations = tx.open_table(OPERATIONS)?;
+    let (_directory, index) = recovery_scratch(index_max)?;
+    let mut index_write = index.begin_write()?;
+    index_write.open_table(FIRST_DECISIONS)?;
+    index_write.open_table(VERSION_OPERATIONS)?;
+    let mut budget = Budget::new(metadata.limits.snapshot);
+    let mut pending = 0;
+    for row in operations.iter()? {
+        let (key, value) = row?;
+        budget.row(key.value(), value.value())?;
+        let provenance = application::decode_provenance(value.value())?;
+        let command = DecideCreated::decode(&provenance.operation.canonical_command)?;
+        let canonical = command.encode()?;
+        if key.value() != command.operation_id || canonical != provenance.operation.canonical_command {
+            return Err(StoreError::CorruptRecord);
+        }
+        at_or_before(provenance.first_decision_log_id, cut)?;
+        let mut original = u64::try_from(canonical.len())
+            .map_err(|_| StoreError::CorruptRecord)?
+            .to_be_bytes()
+            .to_vec();
+        original.extend(&canonical);
+        original.extend(result_bytes(&provenance.operation.result));
+        if application::encode_provenance(provenance.first_decision_log_id, &original)?.as_slice() != value.value() {
+            return Err(StoreError::CorruptRecord);
+        }
+        if index_write
+            .open_table(FIRST_DECISIONS)?
+            .insert(provenance.first_decision_log_id.index, key.value())?
+            .is_some()
+        {
+            return Err(StoreError::CorruptRecord);
+        }
+        if provenance.operation.result == CreatedResult::OperationConflict {
+            return Err(StoreError::CorruptRecord);
+        }
+        if let CreatedResult::Created {
+            version_id,
+            event_id,
+            decision_log_id,
+        } = provenance.operation.result
+        {
+            if version_id != command.prepared.version_id
+                || event_id != command.operation_id
+                || decision_log_id != provenance.first_decision_log_id
+                || command.binding_revision != metadata.header.binding.revision
+            {
+                return Err(StoreError::CorruptRecord);
+            }
+            let event = application::decode_event(required(events.get(key.value())?)?.value())?;
+            if event
+                != (CreatedEvent {
+                    object: command.object.clone(),
+                    prepared: command.prepared.clone(),
+                    decision_log_id,
+                    binding: metadata.header.binding.clone(),
+                })
+            {
+                return Err(StoreError::CorruptRecord);
+            }
+            let version_key = application::version_key(&command.object, version_id)?;
+            let version = application::decode_version(required(versions.get(version_key.as_slice())?)?.value())?;
+            if version
+                != (CreatedVersion {
+                    prepared: command.prepared.clone(),
+                    decision_log_id,
+                    binding_revision: metadata.header.binding.revision,
+                })
+            {
+                return Err(StoreError::CorruptRecord);
+            }
+            if index_write
+                .open_table(VERSION_OPERATIONS)?
+                .insert(version_key.as_slice(), key.value())?
+                .is_some()
+            {
+                return Err(StoreError::CorruptRecord);
+            }
+        } else if events.get(key.value())?.is_some() {
+            return Err(StoreError::CorruptRecord);
+        }
+        pending += 1;
+        if pending == 128 {
+            index_write.commit().map_err(StoreError::StagingCommit)?;
+            index_write = index.begin_write()?;
+            pending = 0;
+        }
+    }
+    index_write.commit().map_err(StoreError::StagingCommit)?;
+    let index_read = index.begin_read()?;
+    let version_operations = index_read.open_table(VERSION_OPERATIONS)?;
+    if version_operations.len()? != versions.len()? {
+        return Err(StoreError::CorruptRecord);
+    }
+    for row in versions.iter()? {
+        let (key, value) = row?;
+        budget.row(key.value(), value.value())?;
+        required(version_operations.get(key.value())?)?;
+        let mut decoder = Decoder::new(key.value());
+        let object = decoder.object()?;
+        let id = decoder.uuid()?;
+        decoder.finish()?;
+        let version = application::decode_version(value.value())?;
+        let head_bytes = required(heads.get(object_bytes(&object)?.as_slice())?)?;
+        let mut decoder = Decoder::new(head_bytes.value());
+        let head = decoder.uuid()?;
+        decoder.finish()?;
+        let current =
+            application::decode_version(required(versions.get(application::version_key(&object, head)?.as_slice())?)?.value())?;
+        if id != version.prepared.version_id
+            || current.decision_log_id.index < version.decision_log_id.index
+            || (current.decision_log_id.index == version.decision_log_id.index
+                && current.decision_log_id != version.decision_log_id)
+        {
+            return Err(StoreError::CorruptRecord);
+        }
+    }
+    for row in events.iter()? {
+        let (key, value) = row?;
+        budget.row(key.value(), value.value())?;
+        let event = application::decode_event(value.value())?;
+        let operation = application::decode_provenance(required(operations.get(key.value())?)?.value())?;
+        if operation.operation.result
+            != (CreatedResult::Created {
+                version_id: event.prepared.version_id,
+                event_id: key.value().try_into().map_err(|_| StoreError::CorruptRecord)?,
+                decision_log_id: event.decision_log_id,
+            })
+        {
+            return Err(StoreError::CorruptRecord);
+        }
+    }
+    for row in heads.iter()? {
+        let (key, value) = row?;
+        budget.row(key.value(), value.value())?;
+        let mut decoder = Decoder::new(key.value());
+        let object = decoder.object()?;
+        decoder.finish()?;
+        let mut decoder = Decoder::new(value.value());
+        let id = decoder.uuid()?;
+        decoder.finish()?;
+        required(versions.get(application::version_key(&object, id)?.as_slice())?)?;
+    }
+    Ok(())
+}
+
+pub(crate) fn validate_recovery(tx: &ReadTransaction) -> Result<(), StoreError> {
+    validate_recovery_with_limits(tx, None)
+}
+
+pub(crate) fn validate_recovery_with_limits(
+    tx: &ReadTransaction,
+    override_limits: Option<CheckpointLimits>,
+) -> Result<(), StoreError> {
+    validate_recovery_inner(tx, override_limits).map_err(resource_error)
+}
+
+fn validate_recovery_inner(tx: &ReadTransaction, override_limits: Option<CheckpointLimits>) -> Result<(), StoreError> {
+    let meta = tx.open_table(crate::META)?;
+    let metadata = required(read_metadata(&meta)?)?;
+    crate::raft_state_machine::selected_meta(&meta, &metadata)?;
+    let cut = metadata.header.applied;
+    let recovery_limits = override_limits.unwrap_or(metadata.limits);
+    recovery_limits.validate()?;
+    let identity = crate::decode_identity(required(meta.get("identity")?)?.value())?;
+    let binding = application::read_binding(&meta)?;
+    if identity.cluster_id != metadata.header.cluster_id || binding != metadata.header.binding {
+        return Err(StoreError::CorruptRecord);
+    }
+    let purged = read_purged(&meta)?;
+    validate_baseline_with_budget(tx, &metadata, recovery_limits.staging_bytes)?;
+    let last = application::read_applied(&meta)?;
+    optional_at_or_before(cut, last)?;
+    let logs = tx.open_table(crate::LOGS)?;
+    let committed = crate::committed::validate_committed(&meta, &logs)?;
+    optional_at_or_before(last, committed)?;
+    let declared_tail = meta.get("tail")?.map(|v| crate::decode_id(v.value())).transpose()?;
+    optional_at_or_before(committed, declared_tail)?;
+    let (_directory, scratch) = recovery_scratch(recovery_limits.staging_bytes)?;
+    for (source, target) in [
+        (HEADS, application::HEADS),
+        (VERSIONS, application::VERSIONS),
+        (EVENTS, application::EVENTS),
+        (OPERATIONS, application::OPERATIONS),
+    ] {
+        let table = tx.open_table(source)?;
+        let mut write = scratch.begin_write()?;
+        write.open_table(target)?;
+        let mut pending = 0;
+        for row in table.iter()? {
+            let (key, value) = row?;
+            write.open_table(target)?.insert(key.value(), value.value())?;
+            pending += 1;
+            if pending == 128 {
+                write.commit().map_err(StoreError::StagingCommit)?;
+                write = scratch.begin_write()?;
+                pending = 0;
+            }
+        }
+        write.commit().map_err(StoreError::StagingCommit)?;
+    }
+    let mut membership: CaptureMembership = metadata.header.membership.clone();
+    if let Some(cut) = cut
+        && let Some(retained_cut) = logs.get(cut.index)?
+        && crate::decode_log(retained_cut.value())?.id != cut
+    {
+        return Err(StoreError::CorruptRecord);
+    }
+    let results = tx.open_table(application::APPLY_RESULTS)?;
+    // Covered entries may remain until purge. Check their exact typed responses
+    // without rebuilding the historical state or allocating a prefix-sized map.
+    let operations = tx.open_table(OPERATIONS)?;
+    let mut actual_tail = purged;
+    let mut prefix_next = purged.map_or(Some(0), |id| id.index.checked_add(1));
+    if let Some(cut) = cut {
+        for row in logs.range(..=cut.index)? {
+            let (key, bytes) = row?;
+            if purged.is_some_and(|purged| key.value() <= purged.index) {
+                return Err(StoreError::CorruptRecord);
+            }
+            if bytes.value().len() > 1024 * 1024 {
+                return Err(StoreError::CorruptRecord);
+            }
+            let entry = crate::decode_log(bytes.value())?;
+            if entry.id.index != key.value() || prefix_next != Some(key.value()) {
+                return Err(StoreError::CorruptRecord);
+            }
+            actual_tail = Some(entry.id);
+            prefix_next = key.value().checked_add(1);
+            let native = application::decode_native_entry(entry)?;
+            let bytes = required(results.get(key.value())?)?;
+            if bytes.value().len() > 1024 * 1024 {
+                return Err(StoreError::CorruptRecord);
+            }
+            let response = application::decode_response(bytes.value())?;
+            application::validate_response_kind(&native, &response)?;
+            if let openraft::EntryPayload::Normal(command) = native.payload {
+                let operation = application::decode_provenance(
+                    required(operations.get(command.created().operation_id.as_slice())?)?.value(),
+                )?;
+                let canonical = command.created().encode()?;
+                let expected = if operation.operation.canonical_command == canonical {
+                    operation.operation.result
+                } else {
+                    CreatedResult::OperationConflict
+                };
+                if response.created_result() != Some(&expected) {
+                    return Err(StoreError::CorruptRecord);
+                }
+                if operation.first_decision_log_id.index == key.value()
+                    && operation.first_decision_log_id != LogId::from(native.log_id)
+                {
+                    return Err(StoreError::CorruptRecord);
+                }
+            }
+        }
+        for row in results.range(..=cut.index)? {
+            let (key, _) = row?;
+            if purged.is_some_and(|purged| key.value() <= purged.index) || logs.get(key.value())?.is_none() {
+                return Err(StoreError::CorruptRecord);
+            }
+        }
+        if purged.is_none_or(|purged| purged.index < cut.index) && logs.get(cut.index)?.is_none() {
+            return Err(StoreError::CorruptRecord);
+        }
+    }
+    if actual_tail != cut {
+        return Err(StoreError::CorruptRecord);
+    }
+
+    let mut observed = cut;
+    let mut next = cut.map_or(Some(0), |cut| cut.index.checked_add(1));
+    let mut budget = Budget::new(recovery_limits.snapshot);
+    if let Some(start) = next {
+        for row in logs.range(start..)? {
+            let (key, bytes) = row?;
+            budget.row(&key.value().to_be_bytes(), bytes.value())?;
+            let entry = crate::decode_log(bytes.value())?;
+            if next != Some(key.value()) || entry.id.index != key.value() {
+                return Err(StoreError::CorruptRecord);
+            }
+            next = key.value().checked_add(1);
+            actual_tail = Some(entry.id);
+            let native = application::decode_native_entry(entry.clone())?;
+            if last.is_none_or(|last| key.value() > last.index) {
+                continue;
+            }
+            let response_bytes = required(results.get(key.value())?)?;
+            if response_bytes.value().len() > 1024 * 1024 {
+                return Err(StoreError::CorruptRecord);
+            }
+            if u64::try_from(response_bytes.value().len()).map_err(|_| StoreError::CorruptRecord)?
+                > recovery_limits.snapshot.value_bytes
+            {
+                return Err(StoreError::ResourceLimit);
+            }
+            let response = application::decode_response(response_bytes.value())?;
+            application::validate_response_kind(&native, &response)?;
+            let write = scratch.begin_write()?;
+            let expected = match native.payload {
+                openraft::EntryPayload::Blank => crate::CaptureResponse::blank(entry.id)?,
+                openraft::EntryPayload::Membership(value) => {
+                    membership = CaptureMembership::try_from(&openraft::StoredMembership::new(Some(entry.id.into()), value))?;
+                    crate::CaptureResponse::membership(entry.id)?
+                }
+                openraft::EntryPayload::Normal(command) => {
+                    let canonical = command.created().encode()?;
+                    let result = application::apply_created(&write, entry.id, command.created(), &canonical, &binding, 6)?;
+                    crate::CaptureResponse::created(CreatedApplyResult {
+                        log_id: entry.id,
+                        result,
+                    })?
+                }
+            };
+            if response != expected {
+                return Err(StoreError::CorruptRecord);
+            }
+            write.commit().map_err(StoreError::StagingCommit)?;
+            observed = Some(entry.id);
+        }
+        for row in results.range(start..)? {
+            let (key, _) = row?;
+            if last.is_none_or(|last| key.value() > last.index) || logs.get(key.value())?.is_none() {
+                return Err(StoreError::CorruptRecord);
+            }
+        }
+    }
+    if actual_tail != declared_tail || observed != last || application::native_application_state(tx)?.1 != membership {
+        return Err(StoreError::CorruptRecord);
+    }
+    let recovered = scratch.begin_read()?;
+    let mut live_budget = Budget::new(recovery_limits.snapshot);
+    for definition in [
+        application::HEADS,
+        application::VERSIONS,
+        application::EVENTS,
+        application::OPERATIONS,
+    ] {
+        let expected = recovered.open_table(definition)?;
+        let actual = tx.open_table(definition)?;
+        if expected.len()? != actual.len()? {
+            return Err(StoreError::CorruptRecord);
+        }
+        for row in actual.iter()? {
+            let (key, value) = row?;
+            live_budget.row(key.value(), value.value())?;
+            if required(expected.get(key.value())?)?.value() != value.value() {
+                return Err(StoreError::CorruptRecord);
+            }
+        }
+    }
+    Ok(())
+}
+
+fn optional_at_or_before(id: Option<LogId>, cut: Option<LogId>) -> Result<(), StoreError> {
+    match (id, cut) {
+        (None, _) => Ok(()),
+        (Some(id), Some(cut)) => at_or_before(id, cut),
+        _ => Err(StoreError::CorruptRecord),
+    }
+}
@@ -0,0 +1,64 @@
+use crate::{LOGS, LogId, META, Store, StoreError, decode_id, decode_log, encode_id, number};
+use redb::{ReadableDatabase, ReadableTable};
+
+pub(crate) fn validate_committed(
+    meta: &impl ReadableTable<&'static str, &'static [u8]>,
+    logs: &impl ReadableTable<u64, &'static [u8]>,
+) -> Result<Option<LogId>, StoreError> {
+    let schema = number(meta.get("schema")?.ok_or(StoreError::CorruptRecord)?.value())?;
+    let marker = meta.get("committed")?;
+    match (schema, marker) {
+        (1 | 3 | 4 | 5 | 6 | 7, None) => {
+            if matches!(schema, 3..=7) && meta.get("last_applied")?.is_none_or(|v| v.value() != [0]) {
+                return Err(StoreError::CorruptRecord);
+            }
+            Ok(None)
+        }
+        (2 | 3 | 4 | 5 | 6 | 7, Some(marker)) => {
+            let id = decode_id(marker.value())?;
+            match logs.get(id.index)? {
+                Some(record) if decode_log(record.value())?.id == id => {}
+                None if schema == 7 && crate::checkpoint_baseline::read_baseline_cut(meta)? == Some(id) => {}
+                _ => return Err(StoreError::CorruptRecord),
+            }
+            Ok(Some(id))
+        }
+        (1 | 2, _) => Err(StoreError::CorruptRecord),
+        _ => Err(StoreError::UnsupportedSchema),
+    }
+}
+
+impl Store {
+    /// Read the exact committed boundary, validating it against the stored log.
+    pub fn read_committed(&self) -> Result<Option<LogId>, StoreError> {
+        let tx = self.database.begin_read()?;
+        validate_committed(&tx.open_table(META)?, &tx.open_table(LOGS)?)
+    }
+
+    /// Persist a monotonic committed boundary supplied by the authority owner.
+    pub fn save_committed(&mut self, id: LogId) -> Result<(), StoreError> {
+        let tx = self.write_transaction()?;
+        {
+            let mut meta = tx.open_table(META)?;
+            let logs = tx.open_table(LOGS)?;
+            let previous = validate_committed(&meta, &logs)?;
+            if previous == Some(id) {
+                return Ok(());
+            }
+            let record = logs.get(id.index)?.ok_or(StoreError::InvalidCommittedLog)?;
+            if decode_log(record.value())?.id != id {
+                return Err(StoreError::InvalidCommittedLog);
+            }
+            if let Some(previous) = previous {
+                if id.index < previous.index {
+                    return Err(StoreError::CommittedRegression);
+                }
+            }
+            meta.insert("committed", encode_id(id).as_slice())?;
+            if previous.is_none() && !matches!(number(meta.get("schema")?.ok_or(StoreError::CorruptRecord)?.value())?, 3..=7) {
+                meta.insert("schema", 2u64.to_be_bytes().as_slice())?;
+            }
+        }
+        self.commit(tx)
+    }
+}
@@ -0,0 +1,278 @@
+//! Reuse the existing actual Raft owner fixture, never a caller-constructed result proof.
+use super::*;
+
+async fn decide(cluster: &Cluster, value: u8) -> (DecideCreated, LogId) {
+    let input = command(value);
+    let expected = input.created().clone();
+    let response = cluster
+        .node(0)
+        .owner()
+        .client_write(input)
+        .await
+        .expect("actual Raft Created decision");
+    let Some(CreatedResult::Created { decision_log_id, .. }) = response.data.created_result() else {
+        panic!("successful Created result required")
+    };
+    (expected, *decision_log_id)
+}
+
+#[tokio::test]
+async fn committed_created_authorization_exact() {
+    let cluster = Cluster::start(&[0], 100).await;
+    let (expected, first) = decide(&cluster, 1).await;
+    let authorization = cluster
+        .node(0)
+        .owner()
+        .authorize_created(expected.clone(), first)
+        .await
+        .expect("durable exact authorization");
+    assert_eq!(authorization.identity(), identity(0));
+    assert_eq!(authorization.operation_id(), expected.operation_id);
+    assert_eq!(authorization.object(), &expected.object);
+    assert_eq!(authorization.prepared(), &expected.prepared);
+    assert_eq!(authorization.binding(), &binding());
+    assert_eq!(authorization.decision_log_id(), first);
+    assert_eq!(cluster.node(0).page().await.events[0].captured.decision_log_id, first);
+    cluster.stop().await;
+}
+
+#[tokio::test]
+async fn committed_created_authorization_retry_first_decision() {
+    let cluster = Cluster::start(&[0], 100).await;
+    let (expected, first) = decide(&cluster, 2).await;
+    let before = cluster
+        .node(0)
+        .owner()
+        .authorize_created(expected.clone(), first)
+        .await
+        .expect("first authorization");
+    let response = cluster
+        .node(0)
+        .owner()
+        .client_write(CaptureCommand::try_created(expected.clone()).expect("retry command"))
+        .await
+        .expect("real retry");
+    assert_ne!(LogId::from(response.log_id), first);
+    assert!(
+        matches!(response.data.created_result(), Some(CreatedResult::Created{decision_log_id,..}) if *decision_log_id==first)
+    );
+    let after = cluster
+        .node(0)
+        .owner()
+        .authorize_created(expected.clone(), first)
+        .await
+        .expect("first decision retained");
+    assert_eq!(before, after);
+    assert!(
+        cluster
+            .node(0)
+            .owner()
+            .authorize_created(expected, response.log_id.into())
+            .await
+            .is_err()
+    );
+    assert_eq!(cluster.node(0).page().await.events.len(), 1);
+    cluster.stop().await;
+}
+
+#[tokio::test]
+async fn committed_created_authorization_rejects_mismatch() {
+    let cluster = Cluster::start(&[0], 100).await;
+    let (expected, first) = decide(&cluster, 3).await;
+    let before = cluster.node(0).page().await;
+    let durable_before = cluster
+        .node(0)
+        .writer
+        .write(|store| {
+            Ok((
+                store.read_vote()?,
+                store.read_committed()?,
+                store.read_entries(0, u64::MAX)?,
+                store.native_application_state()?,
+                store.created_events_page(None, 256)?,
+            ))
+        })
+        .await
+        .expect("durable state before rejected authorization");
+    let mut variants = Vec::new();
+    let mut altered = expected.clone();
+    altered.operation_id = [4; 16];
+    variants.push(altered);
+    let mut altered = expected.clone();
+    altered.object.key.push(9);
+    variants.push(altered);
+    let mut altered = expected.clone();
+    altered.object.bucket_incarnation = [9; 16];
+    variants.push(altered);
+    let mut altered = expected.clone();
+    altered.expected_head = Some([9; 16]);
+    variants.push(altered);
+    let mut altered = expected.clone();
+    altered.prepared.preparation_id = [9; 16];
+    variants.push(altered);
+    let mut altered = expected.clone();
+    altered.prepared.version_id = [9; 16];
+    variants.push(altered);
+    let mut altered = expected.clone();
+    altered.prepared.content_digest = [9; 32];
+    variants.push(altered);
+    let mut altered = expected.clone();
+    altered.prepared.content_length += 1;
+    variants.push(altered);
+    let mut altered = expected.clone();
+    altered.binding_revision += 1;
+    variants.push(altered);
+    for altered in variants {
+        assert!(
+            cluster.node(0).owner().authorize_created(altered, first).await.is_err(),
+            "AUTHORIZATION_REQUIRES_EXACT_PREPARED_IDENTITY_AND_COMMAND"
+        );
+    }
+    for wrong in [
+        LogId {
+            term: first.term + 1,
+            ..first
+        },
+        LogId {
+            leader_node: first.leader_node + 1,
+            ..first
+        },
+        LogId {
+            index: first.index + 1,
+            ..first
+        },
+    ] {
+        assert!(
+            cluster
+                .node(0)
+                .owner()
+                .authorize_created(expected.clone(), wrong)
+                .await
+                .is_err()
+        );
+    }
+    let durable_after = cluster
+        .node(0)
+        .writer
+        .write(|store| {
+            Ok((
+                store.read_vote()?,
+                store.read_committed()?,
+                store.read_entries(0, u64::MAX)?,
+                store.native_application_state()?,
+                store.created_events_page(None, 256)?,
+            ))
+        })
+        .await
+        .expect("durable state after rejected authorization");
+    assert_eq!(durable_before, durable_after, "REJECTED_AUTHORIZATION_DOES_NOT_WRITE");
+    let mut failed = expected.clone();
+    failed.operation_id = [5; 16];
+    failed.prepared.version_id = [5; 16];
+    let response = cluster
+        .node(0)
+        .owner()
+        .client_write(CaptureCommand::try_created(failed.clone()).expect("failed command"))
+        .await
+        .expect("real head mismatch decision");
+    assert!(matches!(response.data.created_result(), Some(CreatedResult::HeadMismatch { .. })));
+    assert!(
+        cluster
+            .node(0)
+            .owner()
+            .authorize_created(failed, response.log_id.into())
+            .await
+            .is_err()
+    );
+    assert_eq!(cluster.node(0).page().await, before);
+    cluster.stop().await;
+}
+
+#[tokio::test]
+async fn committed_created_authorization_restart() {
+    let routes = Arc::new(Routes::default());
+    let mut node = Node::start(0, routes.clone(), config(100)).await;
+    node.owner().bootstrap(members(&[0])).await.expect("bootstrap actual owner");
+    node.owner()
+        .raft()
+        .wait(Some(Duration::from_secs(15)))
+        .state(ServerState::Leader, "leader")
+        .await
+        .expect("leader ready");
+    let expected = command(6).created().clone();
+    let response = node.owner().client_write(command(6)).await.expect("real first write");
+    let Some(CreatedResult::Created {
+        decision_log_id: first, ..
+    }) = response.data.created_result()
+    else {
+        panic!("Created")
+    };
+    let first = *first;
+    let before = node
+        .owner()
+        .authorize_created(expected.clone(), first)
+        .await
+        .expect("before restart");
+    node.stop(&routes).await;
+    let Node {
+        directory,
+        writer,
+        owner: _,
+        id,
+    } = node;
+    drop(writer);
+    let writer = NativeLogWriter::new(Store::open(directory.path().join("db"), identity(id)).expect("retained redb"));
+    // A retained commit-before-apply crash fixture is not a client-provided proof.
+    // Only real owner initialization may replay it before authorization can succeed.
+    use openraft::storage::{RaftLogStorage, RaftLogStorageExt};
+    let pending = command(7).created().clone();
+    let mut logs = NativeRaftLogStore::new(writer.clone());
+    let last = logs.get_log_state().await.expect("retained tail").last_log_id.expect("tail");
+    let pending_id = LogId {
+        term: last.leader_id.term,
+        leader_node: last.leader_id.node_id,
+        index: last.index + 1,
+    };
+    logs.blocking_append(vec![CaptureRaftEntry {
+        log_id: pending_id.into(),
+        payload: openraft::EntryPayload::Normal(CaptureCommand::try_created(pending.clone()).expect("pending command")),
+    }])
+    .await
+    .expect("retained crash suffix");
+    logs.save_committed(Some(pending_id.into()))
+        .await
+        .expect("retained commit marker");
+    assert!(
+        writer.authorize_created(pending.clone(), pending_id).await.is_err(),
+        "UNAPPLIED_COMMIT_IS_NOT_AUTHORIZATION"
+    );
+    drop(logs);
+    let owner = NativeRaftOwner::start(
+        writer.clone(),
+        id,
+        config(100),
+        CheckpointLimits::default(),
+        DirectFactory {
+            source: id,
+            routes: routes.clone(),
+        },
+    )
+    .await
+    .expect("real recovery");
+    routes.register(id, owner.raft().clone());
+    let after = owner
+        .authorize_created(expected, first)
+        .await
+        .expect("durable authorization after restart");
+    assert_eq!(before, after);
+    let replayed = owner
+        .authorize_created(pending.clone(), pending_id)
+        .await
+        .expect("actual owner committed recovery authorizes");
+    assert_eq!(replayed.prepared(), &pending.prepared);
+    assert_eq!(replayed.decision_log_id(), pending_id);
+    owner.shutdown().await.expect("drain recovered owner");
+    routes.remove(id);
+    drop(owner);
+    drop(writer);
+}
@@ -0,0 +1,297 @@
+use crate::{LogId, StoreError, decode_id, encode_id};
+
+pub const MAX_OBJECT_KEY_LENGTH: usize = 4096;
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct ObjectIdentity {
+    pub bucket_incarnation: [u8; 16],
+    pub key: Vec<u8>,
+}
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct PreparedIdentity {
+    pub version_id: [u8; 16],
+    pub preparation_id: [u8; 16],
+    pub content_digest: [u8; 32],
+    pub content_length: u64,
+}
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct DecideCreated {
+    pub operation_id: [u8; 16],
+    pub object: ObjectIdentity,
+    pub expected_head: Option<[u8; 16]>,
+    pub prepared: PreparedIdentity,
+    pub binding_revision: u64,
+}
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct CaptureBinding {
+    pub revision: u64,
+    pub binding_id: [u8; 16],
+    pub target_id: [u8; 16],
+}
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub enum CreatedResult {
+    Created {
+        version_id: [u8; 16],
+        event_id: [u8; 16],
+        decision_log_id: LogId,
+    },
+    HeadMismatch {
+        actual_head: Option<[u8; 16]>,
+    },
+    OperationConflict,
+    BindingMismatch,
+    VersionConflict,
+}
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct CreatedVersion {
+    pub prepared: PreparedIdentity,
+    pub decision_log_id: LogId,
+    pub binding_revision: u64,
+}
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct CreatedEvent {
+    pub object: ObjectIdentity,
+    pub prepared: PreparedIdentity,
+    pub decision_log_id: LogId,
+    pub binding: CaptureBinding,
+}
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct CreatedOperation {
+    pub canonical_command: Vec<u8>,
+    pub result: CreatedResult,
+}
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct CreatedOperationProvenance {
+    pub operation: CreatedOperation,
+    pub first_decision_log_id: LogId,
+}
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct CreatedApplyResult {
+    pub log_id: LogId,
+    pub result: CreatedResult,
+}
+
+/// Local publication authority re-read from the actual owner's durable Created decision.
+/// This is not a serializable remote proof and does not attest physical preparation durability.
+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct CommittedCreatedAuthorization {
+    identity: crate::NodeIdentity,
+    command: DecideCreated,
+    binding: CaptureBinding,
+    decision_log_id: LogId,
+}
+impl CommittedCreatedAuthorization {
+    pub(crate) fn verified(
+        identity: crate::NodeIdentity,
+        command: DecideCreated,
+        binding: CaptureBinding,
+        decision_log_id: LogId,
+    ) -> Self {
+        Self {
+            identity,
+            command,
+            binding,
+            decision_log_id,
+        }
+    }
+    pub fn identity(&self) -> crate::NodeIdentity {
+        self.identity
+    }
+    pub fn operation_id(&self) -> [u8; 16] {
+        self.command.operation_id
+    }
+    pub fn object(&self) -> &ObjectIdentity {
+        &self.command.object
+    }
+    pub fn expected_head(&self) -> Option<[u8; 16]> {
+        self.command.expected_head
+    }
+    pub fn prepared(&self) -> &PreparedIdentity {
+        &self.command.prepared
+    }
+    pub fn binding(&self) -> &CaptureBinding {
+        &self.binding
+    }
+    pub fn decision_log_id(&self) -> LogId {
+        self.decision_log_id
+    }
+}
+
+pub(crate) struct Decoder<'a> {
+    bytes: &'a [u8],
+}
+impl<'a> Decoder<'a> {
+    pub(crate) fn new(bytes: &'a [u8]) -> Self {
+        Self { bytes }
+    }
+    pub(crate) fn take<const N: usize>(&mut self) -> Result<[u8; N], StoreError> {
+        let a = self
+            .bytes
+            .get(..N)
+            .ok_or(StoreError::CorruptRecord)?
+            .try_into()
+            .map_err(|_| StoreError::CorruptRecord)?;
+        self.bytes = &self.bytes[N..];
+        Ok(a)
+    }
+    pub(crate) fn u64(&mut self) -> Result<u64, StoreError> {
+        Ok(u64::from_be_bytes(self.take()?))
+    }
+    pub(crate) fn uuid(&mut self) -> Result<[u8; 16], StoreError> {
+        let id = self.take()?;
+        if id == [0; 16] {
+            return Err(StoreError::CorruptRecord);
+        }
+        Ok(id)
+    }
+    pub(crate) fn option(&mut self) -> Result<Option<[u8; 16]>, StoreError> {
+        match self.take::<1>()?[0] {
+            0 => Ok(None),
+            1 => Ok(Some(self.uuid()?)),
+            _ => Err(StoreError::CorruptRecord),
+        }
+    }
+    pub(crate) fn object(&mut self) -> Result<ObjectIdentity, StoreError> {
+        let bucket_incarnation = self.uuid()?;
+        let len = usize::try_from(u32::from_be_bytes(self.take()?)).map_err(|_| StoreError::CorruptRecord)?;
+        if len > MAX_OBJECT_KEY_LENGTH {
+            return Err(StoreError::CorruptRecord);
+        }
+        let key = self.bytes.get(..len).ok_or(StoreError::CorruptRecord)?.to_vec();
+        self.bytes = &self.bytes[len..];
+        Ok(ObjectIdentity { bucket_incarnation, key })
+    }
+    pub(crate) fn prepared(&mut self) -> Result<PreparedIdentity, StoreError> {
+        Ok(PreparedIdentity {
+            version_id: self.uuid()?,
+            preparation_id: self.uuid()?,
+            content_digest: self.take()?,
+            content_length: self.u64()?,
+        })
+    }
+    pub(crate) fn id(&mut self) -> Result<LogId, StoreError> {
+        decode_id(&self.take::<24>()?)
+    }
+    pub(crate) fn binding(&mut self) -> Result<CaptureBinding, StoreError> {
+        Ok(CaptureBinding {
+            revision: self.u64()?,
+            binding_id: self.uuid()?,
+            target_id: self.uuid()?,
+        })
+    }
+    pub(crate) fn finish(self) -> Result<(), StoreError> {
+        if self.bytes.is_empty() {
+            Ok(())
+        } else {
+            Err(StoreError::CorruptRecord)
+        }
+    }
+    pub(crate) fn result(&mut self) -> Result<CreatedResult, StoreError> {
+        match self.take::<1>()?[0] {
+            0 => Ok(CreatedResult::Created {
+                version_id: self.uuid()?,
+                event_id: self.uuid()?,
+                decision_log_id: self.id()?,
+            }),
+            1 => Ok(CreatedResult::HeadMismatch {
+                actual_head: self.option()?,
+            }),
+            2 => Ok(CreatedResult::OperationConflict),
+            3 => Ok(CreatedResult::BindingMismatch),
+            4 => Ok(CreatedResult::VersionConflict),
+            _ => Err(StoreError::CorruptRecord),
+        }
+    }
+}
+pub(crate) fn put_option(b: &mut Vec<u8>, id: Option<[u8; 16]>) {
+    b.push(u8::from(id.is_some()));
+    if let Some(id) = id {
+        b.extend(id)
+    }
+}
+pub(crate) fn object_bytes(o: &ObjectIdentity) -> Result<Vec<u8>, StoreError> {
+    if o.bucket_incarnation == [0; 16] || o.key.len() > MAX_OBJECT_KEY_LENGTH {
+        return Err(StoreError::CorruptRecord);
+    }
+    let capacity = 20usize.checked_add(o.key.len()).ok_or(StoreError::CorruptRecord)?;
+    let mut b = Vec::with_capacity(capacity);
+    b.extend(o.bucket_incarnation);
+    b.extend(
+        u32::try_from(o.key.len())
+            .map_err(|_| StoreError::CorruptRecord)?
+            .to_be_bytes(),
+    );
+    b.extend(&o.key);
+    Ok(b)
+}
+pub(crate) fn put_prepared(b: &mut Vec<u8>, p: &PreparedIdentity) {
+    b.extend(p.version_id);
+    b.extend(p.preparation_id);
+    b.extend(p.content_digest);
+    b.extend(p.content_length.to_be_bytes())
+}
+pub(crate) fn binding_bytes(v: &CaptureBinding) -> Result<Vec<u8>, StoreError> {
+    let mut b = v.revision.to_be_bytes().to_vec();
+    b.extend(v.binding_id);
+    b.extend(v.target_id);
+    let mut d = Decoder::new(&b);
+    d.binding()?;
+    d.finish()?;
+    Ok(b)
+}
+pub(crate) fn result_bytes(r: &CreatedResult) -> Vec<u8> {
+    let mut b = Vec::new();
+    match r {
+        CreatedResult::Created {
+            version_id,
+            event_id,
+            decision_log_id,
+        } => {
+            b.push(0);
+            b.extend(version_id);
+            b.extend(event_id);
+            b.extend(encode_id(*decision_log_id))
+        }
+        CreatedResult::HeadMismatch { actual_head } => {
+            b.push(1);
+            put_option(&mut b, *actual_head)
+        }
+        CreatedResult::OperationConflict => b.push(2),
+        CreatedResult::BindingMismatch => b.push(3),
+        CreatedResult::VersionConflict => b.push(4),
+    }
+    b
+}
+impl DecideCreated {
+    pub fn encode(&self) -> Result<Vec<u8>, StoreError> {
+        let object = object_bytes(&self.object)?;
+        let capacity = object.len().checked_add(134).ok_or(StoreError::CorruptRecord)?;
+        let mut b = Vec::with_capacity(capacity);
+        b.extend(b"CRTD\x01");
+        b.extend(self.operation_id);
+        b.extend(object);
+        put_option(&mut b, self.expected_head);
+        put_prepared(&mut b, &self.prepared);
+        b.extend(self.binding_revision.to_be_bytes());
+        Self::decode(&b)?;
+        Ok(b)
+    }
+    pub fn decode(bytes: &[u8]) -> Result<Self, StoreError> {
+        let mut d = Decoder::new(bytes);
+        if d.take::<5>()? != *b"CRTD\x01" {
+            return Err(StoreError::CorruptRecord);
+        }
+        let c = Self {
+            operation_id: d.uuid()?,
+            object: d.object()?,
+            expected_head: d.option()?,
+            prepared: d.prepared()?,
+            binding_revision: d.u64()?,
+        };
+        d.finish()?;
+        Ok(c)
+    }
+}
@@ -0,0 +1,588 @@
+//! Local checkpoints are captured from this writer's owned transaction only.
+use crate::{
+    LogId, NativeLogWriter, NativeSnapshotError, NativeSnapshotView, StoreError,
+    checkpoint_baseline::{self, BaselineMetadata},
+    snapshot,
+};
+use redb::{ReadableDatabase, ReadableTable, StorageBackend};
+use std::{
+    fs::File,
+    io::{self, Seek, Write},
+    ops::Bound,
+    sync::{
+        Arc,
+        atomic::{AtomicBool, Ordering},
+    },
+};
+
+#[derive(Clone, Copy, Debug)]
+pub struct CheckpointLimits {
+    pub snapshot: snapshot::Limits,
+    pub staging_bytes: u64,
+}
+
+impl CheckpointLimits {
+    pub(crate) fn validate(&self) -> Result<(), StoreError> {
+        let rows = self.snapshot;
+        if rows.total_bytes == 0
+            || rows.membership_bytes == 0
+            || rows.membership_bytes > 512 * 1024
+            || rows.key_bytes == 0
+            || rows.key_bytes > 1024 * 1024
+            || rows.value_bytes == 0
+            || rows.value_bytes > 1024 * 1024
+            || rows.records == 0
+            || self.staging_bytes < 4 * 1024 * 1024
+        {
+            return Err(StoreError::ResourceLimit);
+        }
+        Ok(())
+    }
+}
+
+impl Default for CheckpointLimits {
+    fn default() -> Self {
+        Self {
+            snapshot: snapshot::Limits {
+                total_bytes: 256 * 1024 * 1024,
+                membership_bytes: 512 * 1024,
+                key_bytes: 8192,
+                value_bytes: 8192,
+                records: 1_000_000,
+            },
+            staging_bytes: 1024 * 1024 * 1024,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct CheckpointReceipt {
+    generation: u64,
+    cut: LogId,
+}
+
+impl CheckpointReceipt {
+    pub fn generation(&self) -> u64 {
+        self.generation
+    }
+    pub fn cut(&self) -> LogId {
+        self.cut
+    }
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum CheckpointError {
+    #[error(transparent)]
+    Store(#[from] StoreError),
+    #[error(transparent)]
+    Capture(#[from] NativeSnapshotError),
+    #[error(transparent)]
+    Snapshot(#[from] snapshot::SnapshotError),
+    #[error("checkpoint filesystem failure")]
+    Io(#[from] io::Error),
+    #[error("checkpoint worker failed")]
+    Worker(#[source] tokio::task::JoinError),
+    #[error("checkpoint store lock poisoned")]
+    Poisoned,
+    #[error("checkpoint conflict: {0}")]
+    Conflict(&'static str),
+    #[error("checkpoint cancelled before publication")]
+    Cancelled,
+}
+
+/// Enforces the file-length budget before redb can allocate or write a page.
+#[derive(Debug)]
+pub(crate) struct BoundedBackend {
+    inner: redb::backends::FileBackend,
+    maximum: u64,
+}
+
+impl BoundedBackend {
+    fn new(file: File, maximum: u64) -> Result<Self, redb::DatabaseError> {
+        Ok(Self {
+            inner: redb::backends::FileBackend::new(file)?,
+            maximum,
+        })
+    }
+
+    fn check(&self, end: u64) -> io::Result<()> {
+        if end > self.maximum {
+            return Err(io::Error::new(io::ErrorKind::StorageFull, "checkpoint staging budget"));
+        }
+        Ok(())
+    }
+}
+
+impl StorageBackend for BoundedBackend {
+    fn len(&self) -> io::Result<u64> {
+        self.inner.len()
+    }
+    fn read(&self, offset: u64, out: &mut [u8]) -> io::Result<()> {
+        self.inner.read(offset, out)
+    }
+    fn set_len(&self, len: u64) -> io::Result<()> {
+        self.check(len)?;
+        self.inner.set_len(len)
+    }
+    fn sync_data(&self) -> io::Result<()> {
+        self.inner.sync_data()
+    }
+    fn write(&self, offset: u64, data: &[u8]) -> io::Result<()> {
+        let length = u64::try_from(data.len()).map_err(io::Error::other)?;
+        self.check(
+            offset
+                .checked_add(length)
+                .ok_or_else(|| io::Error::other("checkpoint write overflow"))?,
+        )?;
+        self.inner.write(offset, data)
+    }
+    fn close(&self) -> io::Result<()> {
+        self.inner.close()
+    }
+    fn try_lock_range(&self, start: Bound<u64>, end: Bound<u64>) -> Result<bool, redb::BackendError> {
+        self.inner.try_lock_range(start, end)
+    }
+    fn try_lock_shared_range(&self, start: Bound<u64>, end: Bound<u64>) -> Result<bool, redb::BackendError> {
+        self.inner.try_lock_shared_range(start, end)
+    }
+    fn lock_range(&self, start: Bound<u64>, end: Bound<u64>) -> Result<(), redb::BackendError> {
+        self.inner.lock_range(start, end)
+    }
+    fn lock_shared_range(&self, start: Bound<u64>, end: Bound<u64>) -> Result<(), redb::BackendError> {
+        self.inner.lock_shared_range(start, end)
+    }
+    fn unlock_range(&self, start: Bound<u64>, end: Bound<u64>) -> Result<(), redb::BackendError> {
+        self.inner.unlock_range(start, end)
+    }
+    fn query_lock_range(&self, start: Bound<u64>, end: Bound<u64>) -> Result<bool, redb::BackendError> {
+        self.inner.query_lock_range(start, end)
+    }
+}
+
+pub(crate) fn recovery_scratch(maximum: u64) -> Result<(tempfile::TempDir, redb::Database), StoreError> {
+    let directory = tempfile::tempdir()?;
+    let database = stage_database(directory.path().join("stage.redb"), maximum)?;
+    Ok((directory, database))
+}
+
+pub(crate) fn stage_database(path: std::path::PathBuf, maximum: u64) -> Result<redb::Database, StoreError> {
+    let file = File::options().read(true).write(true).create_new(true).open(path)?;
+    let backend = BoundedBackend::new(file, maximum)?;
+    Ok(redb::Database::builder()
+        .set_cache_size(8 * 1024 * 1024)
+        .create_with_backend(backend)?)
+}
+
+struct PayloadFile {
+    file: File,
+    maximum: u64,
+    written: u64,
+}
+
+impl Write for PayloadFile {
+    fn write(&mut self, bytes: &[u8]) -> io::Result<usize> {
+        let length = u64::try_from(bytes.len()).map_err(io::Error::other)?;
+        let end = self
+            .written
+            .checked_add(length)
+            .ok_or_else(|| io::Error::other("checkpoint payload overflow"))?;
+        if end > self.maximum {
+            return Err(io::Error::new(io::ErrorKind::StorageFull, "checkpoint payload budget"));
+        }
+        let count = self.file.write(bytes)?;
+        self.written = self
+            .written
+            .checked_add(u64::try_from(count).map_err(io::Error::other)?)
+            .ok_or_else(|| io::Error::other("checkpoint payload overflow"))?;
+        Ok(count)
+    }
+    fn flush(&mut self) -> io::Result<()> {
+        self.file.flush()
+    }
+}
+
+pub(crate) struct StageSink<'a>(pub(crate) &'a redb::WriteTransaction);
+
+impl snapshot::StagingSink for StageSink<'_> {
+    fn row(&mut self, section: snapshot::Section, key: &[u8], value: &[u8]) -> io::Result<()> {
+        let definition = match section {
+            snapshot::Section::Heads => checkpoint_baseline::HEADS,
+            snapshot::Section::Versions => checkpoint_baseline::VERSIONS,
+            snapshot::Section::Events => checkpoint_baseline::EVENTS,
+            snapshot::Section::Operations => checkpoint_baseline::OPERATIONS,
+        };
+        let mut table = self.0.open_table(definition).map_err(io::Error::other)?;
+        if table.insert(key, value).map_err(io::Error::other)?.is_some() {
+            return Err(io::Error::new(io::ErrorKind::InvalidData, "duplicate checkpoint key"));
+        }
+        Ok(())
+    }
+}
+
+// This type has no external constructors, deserialization, or path-based import.
+// Its source remains the owned local read transaction, not the decoded artifact.
+struct LocalStage {
+    source: NativeSnapshotView,
+    database: redb::Database,
+    _directory: tempfile::TempDir,
+    metadata: BaselineMetadata,
+}
+
+fn capture_stage(source: NativeSnapshotView, limits: CheckpointLimits) -> Result<LocalStage, CheckpointError> {
+    limits.validate()?;
+    source
+        .header()
+        .applied
+        .ok_or(CheckpointError::Conflict("empty applied cut"))?;
+    let directory = tempfile::tempdir()?;
+    let payload = File::options()
+        .read(true)
+        .write(true)
+        .create_new(true)
+        .open(directory.path().join("capture.csnp"))?;
+    let artifact = source.stream_local_capability(
+        PayloadFile {
+            file: payload,
+            maximum: limits.staging_bytes.min(limits.snapshot.total_bytes),
+            written: 0,
+        },
+        limits.snapshot,
+    )?;
+    let remaining = limits
+        .staging_bytes
+        .checked_sub(artifact.payload.written)
+        .ok_or(CheckpointError::Conflict("staging budget"))?;
+    let mut payload = artifact.payload.file;
+    payload.sync_all()?;
+    payload.rewind()?;
+    let database = stage_database(directory.path().join("stage.redb"), remaining)?;
+    let transaction = database.begin_write().map_err(StoreError::from)?;
+    for definition in [
+        checkpoint_baseline::HEADS,
+        checkpoint_baseline::VERSIONS,
+        checkpoint_baseline::EVENTS,
+        checkpoint_baseline::OPERATIONS,
+    ] {
+        transaction.open_table(definition).map_err(StoreError::from)?;
+    }
+    let header = snapshot::decode_capability_artifact(
+        snapshot::CapabilityArtifact {
+            manifest: artifact.manifest,
+            payload,
+        },
+        &mut StageSink(&transaction),
+        limits.snapshot,
+    )?;
+    std::fs::remove_file(directory.path().join("capture.csnp"))?;
+    if header != *source.header() {
+        return Err(CheckpointError::Conflict("capture header"));
+    }
+    let metadata = BaselineMetadata {
+        generation: 1,
+        header,
+        limits,
+    };
+    let encoded = checkpoint_baseline::encode_metadata(&metadata)?;
+    {
+        let mut meta = transaction.open_table(crate::META).map_err(StoreError::from)?;
+        meta.insert("schema", 7u64.to_be_bytes().as_slice())
+            .map_err(StoreError::from)?;
+        meta.insert("checkpoint_baseline", encoded.as_slice())
+            .map_err(StoreError::from)?;
+    }
+    transaction.commit().map_err(StoreError::StagingCommit)?;
+    let stage = LocalStage {
+        source,
+        database,
+        _directory: directory,
+        metadata,
+    };
+    Ok(stage)
+}
+
+fn verify_local_stage(stage: &LocalStage) -> Result<(), CheckpointError> {
+    let transaction = stage.database.begin_read().map_err(StoreError::from)?;
+    for (source_definition, stage_definition) in [
+        (crate::application::HEADS, checkpoint_baseline::HEADS),
+        (crate::application::VERSIONS, checkpoint_baseline::VERSIONS),
+        (crate::application::EVENTS, checkpoint_baseline::EVENTS),
+        (crate::application::OPERATIONS, checkpoint_baseline::OPERATIONS),
+    ] {
+        let source = stage
+            .source
+            .transaction()
+            .open_table(source_definition)
+            .map_err(StoreError::from)?;
+        let staged = transaction.open_table(stage_definition).map_err(StoreError::from)?;
+        let mut source_rows = source.iter().map_err(StoreError::from)?;
+        let mut staged_rows = staged.iter().map_err(StoreError::from)?;
+        loop {
+            match (source_rows.next(), staged_rows.next()) {
+                (None, None) => break,
+                (Some(source), Some(staged)) => {
+                    let (source_key, source_value) = source.map_err(StoreError::from)?;
+                    let (stage_key, stage_value) = staged.map_err(StoreError::from)?;
+                    if source_key.value() != stage_key.value() || source_value.value() != stage_value.value() {
+                        return Err(CheckpointError::Conflict("staging differs from owned local capture"));
+                    }
+                }
+                _ => return Err(CheckpointError::Conflict("staging row count")),
+            }
+        }
+    }
+    let staged_meta = transaction.open_table(crate::META).map_err(StoreError::from)?;
+    let staged_metadata =
+        checkpoint_baseline::read_metadata(&staged_meta)?.ok_or(CheckpointError::Conflict("missing staging metadata"))?;
+    if checkpoint_baseline::encode_metadata(&staged_metadata)? != checkpoint_baseline::encode_metadata(&stage.metadata)? {
+        return Err(CheckpointError::Conflict("staging metadata differs from local capture"));
+    }
+    let stage_bytes = std::fs::metadata(stage._directory.path().join("stage.redb"))?.len();
+    let index_budget = stage
+        .metadata
+        .limits
+        .staging_bytes
+        .checked_sub(stage_bytes)
+        .ok_or(CheckpointError::Conflict("staging disk budget"))?;
+    checkpoint_baseline::validate_baseline_with_budget(&transaction, &stage.metadata, index_budget)?;
+    Ok(())
+}
+
+fn resource_error(error: CheckpointError) -> CheckpointError {
+    use std::error::Error;
+    if matches!(&error, CheckpointError::Store(StoreError::CommitIndeterminate(_))) {
+        return error;
+    }
+    let mut source: Option<&(dyn Error + 'static)> = Some(&error);
+    while let Some(current) = source {
+        if current
+            .downcast_ref::<io::Error>()
+            .is_some_and(|error| error.kind() == io::ErrorKind::StorageFull)
+        {
+            return CheckpointError::Store(StoreError::ResourceLimit);
+        }
+        source = current.source();
+    }
+    error
+}
+
+struct CancelOnDrop(Arc<AtomicBool>);
+impl Drop for CancelOnDrop {
+    fn drop(&mut self) {
+        self.0.store(true, Ordering::Release);
+    }
+}
+
+#[cfg(test)]
+type CaptureHook = Box<dyn FnOnce() + Send>;
+#[cfg(test)]
+type StageHook = Box<dyn FnOnce(&redb::Database) + Send>;
+#[cfg(test)]
+static CAPTURE_HOOKS: std::sync::Mutex<std::collections::BTreeMap<usize, CaptureHook>> =
+    std::sync::Mutex::new(std::collections::BTreeMap::new());
+#[cfg(test)]
+static STAGE_HOOKS: std::sync::Mutex<std::collections::BTreeMap<usize, StageHook>> =
+    std::sync::Mutex::new(std::collections::BTreeMap::new());
+
+#[cfg(test)]
+pub(crate) struct CheckpointHook {
+    owner: usize,
+    stage: bool,
+}
+#[cfg(test)]
+impl Drop for CheckpointHook {
+    fn drop(&mut self) {
+        if self.stage {
+            STAGE_HOOKS.lock().expect("stage hooks").remove(&self.owner);
+        } else {
+            CAPTURE_HOOKS.lock().expect("capture hooks").remove(&self.owner);
+        }
+    }
+}
+
+#[cfg(test)]
+pub(crate) fn after_local_capture(writer: &NativeLogWriter, hook: impl FnOnce() + Send + 'static) -> CheckpointHook {
+    let owner = Arc::as_ptr(&writer.store).addr();
+    assert!(
+        CAPTURE_HOOKS
+            .lock()
+            .expect("capture hooks")
+            .insert(owner, Box::new(hook))
+            .is_none()
+    );
+    CheckpointHook { owner, stage: false }
+}
+
+#[cfg(test)]
+pub(crate) fn after_local_stage(writer: &NativeLogWriter, hook: impl FnOnce(&redb::Database) + Send + 'static) -> CheckpointHook {
+    let owner = Arc::as_ptr(&writer.store).addr();
+    assert!(
+        STAGE_HOOKS
+            .lock()
+            .expect("stage hooks")
+            .insert(owner, Box::new(hook))
+            .is_none()
+    );
+    CheckpointHook { owner, stage: true }
+}
+
+#[cfg(test)]
+fn run_capture_hook(owner: &Arc<std::sync::Mutex<crate::Store>>) {
+    let hook = CAPTURE_HOOKS
+        .lock()
+        .expect("capture hooks")
+        .remove(&Arc::as_ptr(owner).addr());
+    if let Some(hook) = hook {
+        hook();
+    }
+}
+
+#[cfg(test)]
+fn run_stage_hook(owner: &Arc<std::sync::Mutex<crate::Store>>, database: &redb::Database) {
+    let hook = STAGE_HOOKS.lock().expect("stage hooks").remove(&Arc::as_ptr(owner).addr());
+    if let Some(hook) = hook {
+        hook(database);
+    }
+}
+
+impl NativeLogWriter {
+    pub async fn checkpoint_local(&self, limits: CheckpointLimits) -> Result<CheckpointReceipt, CheckpointError> {
+        self.checkpoint_local_inner(limits).await.map_err(resource_error)
+    }
+
+    async fn checkpoint_local_inner(&self, limits: CheckpointLimits) -> Result<CheckpointReceipt, CheckpointError> {
+        limits.validate()?;
+        let admission = self.checkpoint_admission.clone().lock_owned().await;
+        let owner = self.store.clone();
+        let stage = tokio::task::spawn_blocking(move || {
+            let source = NativeSnapshotView::acquire(owner.clone())?;
+            #[cfg(test)]
+            run_capture_hook(&owner);
+            let stage = capture_stage(source, limits)?;
+            #[cfg(test)]
+            run_stage_hook(&owner, &stage.database);
+            verify_local_stage(&stage)?;
+            Ok::<_, CheckpointError>((stage, admission))
+        })
+        .await
+        .map_err(CheckpointError::Worker)??;
+        let ordering = self.ordering.clone().lock_owned().await;
+        let owner = self.store.clone();
+        let cancellation = CancelOnDrop(Arc::new(AtomicBool::new(false)));
+        let cancelled = cancellation.0.clone();
+        tokio::task::spawn_blocking(move || {
+            // The blocking owner retains both guards until publication finishes,
+            // even if the async caller is dropped while redb is committing.
+            let _ordering = ordering;
+            let (stage, _admission) = stage;
+            if cancelled.load(Ordering::Acquire) {
+                return Err(CheckpointError::Cancelled);
+            }
+            if !Arc::ptr_eq(&owner, stage.source.owner()) {
+                return Err(CheckpointError::Conflict("local capture owner"));
+            }
+            let mut store = owner.lock().map_err(|_| CheckpointError::Poisoned)?;
+            let transaction = store.write_transaction()?;
+            let mut meta = transaction.open_table(crate::META).map_err(StoreError::from)?;
+            let schema = crate::number(
+                meta.get("schema")
+                    .map_err(StoreError::from)?
+                    .ok_or(StoreError::CorruptRecord)?
+                    .value(),
+            )?;
+            if !matches!(schema, 6 | 7) {
+                return Err(StoreError::UnsupportedSchema.into());
+            }
+            let previous = checkpoint_baseline::read_metadata(&meta)?;
+            if crate::application::read_binding(&meta)? != stage.metadata.header.binding {
+                return Err(CheckpointError::Conflict("live capture binding changed"));
+            }
+            let identity = crate::decode_identity(
+                meta.get("identity")
+                    .map_err(StoreError::from)?
+                    .ok_or(StoreError::CorruptRecord)?
+                    .value(),
+            )?;
+            if identity.cluster_id != stage.metadata.header.cluster_id {
+                return Err(CheckpointError::Conflict("live capture cluster changed"));
+            }
+            let cut = stage
+                .metadata
+                .header
+                .applied
+                .ok_or(CheckpointError::Conflict("empty applied cut"))?;
+            let applied = crate::application::read_applied(&meta)?.ok_or(StoreError::CorruptRecord)?;
+            let logs = transaction.open_table(crate::LOGS).map_err(StoreError::from)?;
+            let committed = crate::committed::validate_committed(&meta, &logs)?.ok_or(StoreError::CorruptRecord)?;
+            if cut.index > applied.index
+                || (cut.index == applied.index && cut != applied)
+                || applied.index > committed.index
+                || (applied.index == committed.index && applied != committed)
+            {
+                return Err(CheckpointError::Conflict("live applied and committed bounds"));
+            }
+            drop(logs);
+            if let Some(old) = previous.as_ref().and_then(|previous| previous.header.applied)
+                && (old.index > cut.index || (old.index == cut.index && old != cut))
+            {
+                return Err(CheckpointError::Conflict("published checkpoint cut"));
+            }
+            let mut metadata = stage.metadata;
+            metadata.generation = previous.map_or(Ok(1), |old| {
+                old.generation
+                    .checked_add(1)
+                    .ok_or(CheckpointError::Conflict("checkpoint generation exhausted"))
+            })?;
+            let stage_transaction = stage.database.begin_read().map_err(StoreError::from)?;
+            for (source_definition, definition) in [
+                (crate::application::HEADS, checkpoint_baseline::HEADS),
+                (crate::application::VERSIONS, checkpoint_baseline::VERSIONS),
+                (crate::application::EVENTS, checkpoint_baseline::EVENTS),
+                (crate::application::OPERATIONS, checkpoint_baseline::OPERATIONS),
+            ] {
+                let captured = stage
+                    .source
+                    .transaction()
+                    .open_table(source_definition)
+                    .map_err(StoreError::from)?;
+                let staged = stage_transaction.open_table(definition).map_err(StoreError::from)?;
+                let mut source_rows = captured.iter().map_err(StoreError::from)?;
+                let mut stage_rows = staged.iter().map_err(StoreError::from)?;
+                let mut target = transaction.open_table(definition).map_err(StoreError::from)?;
+                target.retain(|_, _| false).map_err(StoreError::from)?;
+                // Recheck disk bytes while copying; the earlier relationship
+                // validation remains outside the writer ordering interval.
+                loop {
+                    match (source_rows.next(), stage_rows.next()) {
+                        (None, None) => break,
+                        (Some(source), Some(staged)) => {
+                            let (source_key, source_value) = source.map_err(StoreError::from)?;
+                            let (key, value) = staged.map_err(StoreError::from)?;
+                            if source_key.value() != key.value() || source_value.value() != value.value() {
+                                return Err(CheckpointError::Conflict("publication differs from local capture"));
+                            }
+                            target.insert(key.value(), value.value()).map_err(StoreError::from)?;
+                        }
+                        _ => return Err(CheckpointError::Conflict("publication row count")),
+                    }
+                }
+            }
+            let encoded = checkpoint_baseline::encode_metadata(&metadata)?;
+            meta.insert("checkpoint_baseline", encoded.as_slice())
+                .map_err(StoreError::from)?;
+            meta.insert("schema", 7u64.to_be_bytes().as_slice())
+                .map_err(StoreError::from)?;
+            crate::raft_state_machine::publish_local_meta(&mut meta, &metadata)?;
+            drop(meta);
+            if cancelled.load(Ordering::Acquire) {
+                return Err(CheckpointError::Cancelled);
+            }
+            store.commit(transaction)?;
+            Ok(CheckpointReceipt {
+                generation: metadata.generation,
+                cut,
+            })
+        })
+        .await
+        .map_err(CheckpointError::Worker)?
+    }
+}
@@ -0,0 +1,607 @@
+use super::*;
+use std::collections::{BTreeMap, BTreeSet};
+
+fn identity() -> NodeIdentity {
+    NodeIdentity {
+        cluster_id: [7; 16],
+        node_id: 0,
+    }
+}
+
+fn binding() -> CaptureBinding {
+    CaptureBinding {
+        revision: 9,
+        binding_id: [8; 16],
+        target_id: [9; 16],
+    }
+}
+
+fn id(index: u64) -> LogId {
+    LogId {
+        term: 8,
+        leader_node: 0,
+        index,
+    }
+}
+
+fn command(operation: u8) -> DecideCreated {
+    DecideCreated {
+        operation_id: [operation; 16],
+        object: ObjectIdentity {
+            bucket_incarnation: [2; 16],
+            key: b"checkpoint-object".to_vec(),
+        },
+        expected_head: None,
+        prepared: PreparedIdentity {
+            version_id: [3; 16],
+            preparation_id: [4; 16],
+            content_digest: [5; 32],
+            content_length: 4097,
+        },
+        binding_revision: 9,
+    }
+}
+
+fn apply(store: &mut Store, index: u64, command: DecideCreated) -> CaptureResponse {
+    let entry = CaptureRaftEntry {
+        log_id: id(index).into(),
+        payload: openraft::EntryPayload::Normal(CaptureCommand::try_created(command).expect("canonical command")),
+    };
+    store.append_native(&[entry]).expect("actual durable append");
+    store.save_committed(id(index)).expect("actual durable commit");
+    store.apply_next_native(id(index)).expect("actual native apply")
+}
+
+fn fixture() -> (tempfile::TempDir, Store, DecideCreated, CreatedOperationProvenance) {
+    let dir = tempfile::tempdir().expect("actual native database directory");
+    let mut store = Store::initialize(dir.path().join("db"), identity()).expect("initialize actual redb");
+    store.pin_capture_binding(binding()).expect("bind actual store");
+    store.enable_native_empty().expect("native schema4");
+    store.promote_created_provenance().expect("native schema6 provenance");
+    let membership = CaptureMembership::new(
+        Some(id(0)),
+        vec![BTreeSet::from([0])],
+        BTreeMap::from([(0, CaptureNode::new("http://127.0.0.1:7000".into()).expect("native member"))]),
+    )
+    .expect("full native membership");
+    let stored = openraft::StoredMembership::try_from(&membership).expect("OpenRaft membership");
+    store
+        .append_native(&[CaptureRaftEntry {
+            log_id: id(0).into(),
+            payload: openraft::EntryPayload::Membership(stored.membership().clone()),
+        }])
+        .expect("append actual membership");
+    store.save_committed(id(0)).expect("commit membership");
+    store.apply_next_native(id(0)).expect("apply membership");
+    let negative = command(11);
+    let mut original = negative.clone();
+    original.expected_head = Some([12; 16]);
+    apply(&mut store, 1, original.clone());
+    let first = store
+        .created_operation_provenance(original.operation_id)
+        .expect("read first negative")
+        .expect("canonical negative operation");
+    assert_eq!(first.first_decision_log_id, id(1), "CHECKPOINT_FIRST_NEGATIVE_BOUND");
+    assert!(
+        matches!(first.operation.result, CreatedResult::HeadMismatch { .. }),
+        "CHECKPOINT_FIRST_NEGATIVE_BOUND"
+    );
+    apply(&mut store, 2, command(1));
+    (dir, store, original, first)
+}
+
+fn runtime() -> tokio::runtime::Runtime {
+    tokio::runtime::Builder::new_current_thread()
+        .enable_all()
+        .build()
+        .expect("checkpoint runtime")
+}
+
+fn schema(store: &Store) -> u64 {
+    let tx = store.database.begin_read().expect("read actual schema");
+    let table = tx.open_table(META).expect("schema table");
+    number(table.get("schema").expect("schema read").expect("schema exists").value()).expect("schema number")
+}
+
+fn assert_negative(store: &Store, original: &DecideCreated, first: &CreatedOperationProvenance) {
+    assert_eq!(
+        store
+            .created_operation_provenance(original.operation_id)
+            .expect("canonical read"),
+        Some(first.clone()),
+        "CHECKPOINT_FIRST_NEGATIVE_BOUND: preserve original bytes/result/full first ID"
+    );
+}
+
+#[test]
+fn native_checkpoint_negative_retry_and_purge_reopen() {
+    let (dir, store, original, first) = fixture();
+    assert_eq!(schema(&store), 6);
+    let writer = NativeLogWriter::new(store);
+    let receipt = runtime()
+        .block_on(writer.checkpoint_local(CheckpointLimits::default()))
+        .expect("publish local checkpoint");
+    assert_eq!(receipt.cut(), id(2), "CHECKPOINT_BASELINE_BOUND");
+    assert_eq!(receipt.generation(), 1);
+    {
+        let mut store = writer.store.lock().expect("actual writer store");
+        assert_eq!(schema(&store), 7);
+        assert_negative(&store, &original, &first);
+        apply(&mut store, 3, original.clone());
+        assert_negative(&store, &original, &first);
+        let mut conflict = original.clone();
+        conflict.binding_revision += 1;
+        apply(&mut store, 4, conflict);
+        assert_eq!(
+            store
+                .created_apply_result(4)
+                .expect("conflicting response")
+                .expect("response")
+                .result,
+            CreatedResult::OperationConflict,
+            "CHECKPOINT_RETRY_BOUND"
+        );
+        assert_negative(&store, &original, &first);
+    }
+    runtime()
+        .block_on(writer.purge_checkpoint(&receipt))
+        .expect("purge exact durable baseline");
+    drop(writer);
+    let reopened = Store::open(dir.path().join("db"), identity()).expect("reopen purged baseline plus suffix");
+    assert_negative(&reopened, &original, &first);
+    assert_eq!(reopened.last_applied().expect("suffix applied"), Some(id(4)), "CHECKPOINT_SUFFIX_BOUND");
+    assert!(reopened.read_entry(2).expect("purged cut").is_none());
+    assert!(reopened.read_entry(3).expect("retained suffix").is_some());
+}
+
+#[test]
+fn native_checkpoint_capture_keeps_concurrent_suffix() {
+    let (dir, store, original, first) = fixture();
+    let writer = NativeLogWriter::new(store);
+    let (entered_tx, entered_rx) = std::sync::mpsc::channel();
+    let (release_tx, release_rx) = std::sync::mpsc::channel();
+    let _hook = native_checkpoint::after_local_capture(&writer, move || {
+        entered_tx.send(()).expect("capture reached");
+        release_rx
+            .recv_timeout(std::time::Duration::from_secs(30))
+            .expect("release owned capture");
+    });
+    let worker = writer.clone();
+    let thread = std::thread::spawn(move || runtime().block_on(worker.checkpoint_local(CheckpointLimits::default())));
+    entered_rx
+        .recv_timeout(std::time::Duration::from_secs(30))
+        .expect("actual capture boundary");
+    {
+        let mut store = writer.store.lock().expect("capture releases live write owner");
+        apply(&mut store, 3, original.clone());
+    }
+    release_tx.send(()).expect("release capture");
+    let receipt = thread.join().expect("checkpoint worker").expect("concurrent checkpoint");
+    assert_eq!(receipt.cut(), id(2), "CHECKPOINT_SAME_CUT_BOUND");
+    runtime()
+        .block_on(writer.purge_checkpoint(&receipt))
+        .expect("purge captured prefix");
+    drop(writer);
+    let reopened = Store::open(dir.path().join("db"), identity()).expect("reopen concurrent suffix");
+    assert_eq!(reopened.last_applied().expect("actual suffix"), Some(id(3)), "CHECKPOINT_SUFFIX_BOUND");
+    assert_negative(&reopened, &original, &first);
+}
+
+#[test]
+fn native_checkpoint_staging_corruption_preserves_live_generation() {
+    let (dir, store, original, first) = fixture();
+    let writer = NativeLogWriter::new(store);
+    let before = runtime()
+        .block_on(writer.checkpoint_local(CheckpointLimits::default()))
+        .expect("initial baseline");
+    let _hook = native_checkpoint::after_local_stage(&writer, |database| {
+        let tx = database.begin_write().expect("actual isolated stage mutation");
+        tx.open_table(checkpoint_baseline::OPERATIONS)
+            .expect("stage operations")
+            .retain(|_, _| false)
+            .expect("damage cross-record relationships");
+        tx.commit().expect("persist actual damaged stage");
+    });
+    assert!(
+        runtime()
+            .block_on(writer.checkpoint_local(CheckpointLimits::default()))
+            .is_err(),
+        "CHECKPOINT_FAILURE_ATOMIC"
+    );
+    let store = writer.store.lock().expect("unchanged live store");
+    let tx = store.database.begin_read().expect("published generation");
+    let meta = tx.open_table(META).expect("published meta");
+    assert_eq!(
+        checkpoint_baseline::read_metadata(&meta)
+            .expect("valid old metadata")
+            .expect("old baseline")
+            .generation,
+        before.generation(),
+        "CHECKPOINT_FAILURE_ATOMIC"
+    );
+    assert_negative(&store, &original, &first);
+    drop(meta);
+    drop(tx);
+    drop(store);
+    drop(writer);
+    let reopened = Store::open(dir.path().join("db"), identity()).expect("reopen after failed staging validation");
+    assert_negative(&reopened, &original, &first);
+}
+
+#[test]
+fn native_checkpoint_budget_failure_preserves_schema_and_history() {
+    for staging in [false, true] {
+        let (dir, mut store, original, first) = fixture();
+        if staging {
+            for index in 0u64..256 {
+                let mut large = command(31);
+                large.operation_id[..8].copy_from_slice(&index.to_be_bytes());
+                large.prepared.version_id[..8].copy_from_slice(&index.to_be_bytes());
+                large.object.key = vec![b'x'; MAX_OBJECT_KEY_LENGTH];
+                large.object.key[..8].copy_from_slice(&index.to_be_bytes());
+                apply(&mut store, index + 3, large);
+            }
+        }
+        let writer = NativeLogWriter::new(store);
+        let mut limits = CheckpointLimits::default();
+        if staging {
+            limits.staging_bytes = 4 * 1024 * 1024;
+        } else {
+            limits.snapshot.total_bytes = 64;
+        }
+        let result = runtime().block_on(writer.checkpoint_local(limits));
+        if staging {
+            assert!(
+                matches!(&result, Err(CheckpointError::Store(StoreError::ResourceLimit))),
+                "CHECKPOINT_RESOURCE_BOUND: actual 4MiB physical staging limit: {result:?}"
+            );
+        } else {
+            assert!(
+                result.is_err(),
+                "CHECKPOINT_RESOURCE_BOUND: actual bounded output must refuse: {result:?}"
+            );
+        }
+        {
+            let store = writer.store.lock().expect("live after exhausted budget");
+            assert_eq!(schema(&store), 6, "CHECKPOINT_FAILURE_ATOMIC");
+            assert_negative(&store, &original, &first);
+            assert!(store.read_entry(0).expect("prefix intact").is_some());
+        }
+        drop(writer);
+        assert_negative(
+            &Store::open(dir.path().join("db"), identity()).expect("reopen budget failure"),
+            &original,
+            &first,
+        );
+    }
+}
+
+#[test]
+fn native_checkpoint_cancelled_capture_cannot_publish_or_purge() {
+    let (dir, store, original, first) = fixture();
+    let writer = NativeLogWriter::new(store);
+    let (entered_tx, entered_rx) = std::sync::mpsc::channel();
+    let (release_tx, release_rx) = std::sync::mpsc::channel();
+    let _hook = native_checkpoint::after_local_capture(&writer, move || {
+        entered_tx.send(()).expect("capture cancellation boundary");
+        release_rx
+            .recv_timeout(std::time::Duration::from_secs(30))
+            .expect("release pending actual capture");
+    });
+    let rt = runtime();
+    let worker = writer.clone();
+    rt.block_on(async {
+        let task = tokio::spawn(async move { worker.checkpoint_local(CheckpointLimits::default()).await });
+        tokio::task::spawn_blocking(move || entered_rx.recv_timeout(std::time::Duration::from_secs(30)))
+            .await
+            .expect("capture observer")
+            .expect("actual capture owned transaction");
+        task.abort();
+        assert!(task.await.expect_err("cancelled waiter").is_cancelled(), "CHECKPOINT_CANCEL_BOUND");
+        release_tx.send(()).expect("release actual worker after waiter cancellation");
+        // A fresh checkpoint waits for the same production admission; completion
+        // establishes that the cancelled capture worker released its real resources.
+        let receipt = writer
+            .checkpoint_local(CheckpointLimits::default())
+            .await
+            .expect("capture after cancellation");
+        assert_eq!(receipt.generation(), 1, "CHECKPOINT_CANCEL_BOUND: abandoned capture never published");
+    });
+    drop(rt);
+    drop(writer);
+    let reopened = Store::open(dir.path().join("db"), identity()).expect("reopen after cancellation and successful next capture");
+    assert_negative(&reopened, &original, &first);
+    assert!(
+        reopened.read_entry(0).expect("cancellation cannot purge").is_some(),
+        "CHECKPOINT_CANCEL_BOUND"
+    );
+}
+
+#[test]
+fn native_checkpoint_stale_receipt_cannot_remove_history() {
+    let (_dir, store, original, first) = fixture();
+    let writer = NativeLogWriter::new(store);
+    let rt = runtime();
+    let old = rt
+        .block_on(writer.checkpoint_local(CheckpointLimits::default()))
+        .expect("first generation");
+    let new = rt
+        .block_on(writer.checkpoint_local(CheckpointLimits::default()))
+        .expect("next generation");
+    assert!(new.generation() > old.generation());
+    assert!(
+        matches!(
+            rt.block_on(writer.purge_checkpoint(&old)),
+            Err(NativeWriteError::Store(StoreError::StaleCheckpoint))
+        ),
+        "CHECKPOINT_PURGE_BOUND"
+    );
+    let store = writer.store.lock().expect("history after stale purge");
+    assert_negative(&store, &original, &first);
+    assert!(store.read_entry(1).expect("first decision retained").is_some());
+    assert!(store.created_apply_result(1).expect("typed result retained").is_some());
+}
+
+#[test]
+fn native_checkpoint_reopen_rejects_conflicting_cut_and_suffix_gap() {
+    for corruption in 0..4 {
+        let gap = corruption == 1;
+        let (dir, store, original, _) = fixture();
+        let writer = NativeLogWriter::new(store);
+        let rt = runtime();
+        let receipt = rt
+            .block_on(writer.checkpoint_local(CheckpointLimits::default()))
+            .expect("valid checkpoint");
+        if gap {
+            {
+                let mut store = writer.store.lock().expect("suffix owner");
+                apply(&mut store, 3, original.clone());
+                apply(&mut store, 4, original.clone());
+            }
+            rt.block_on(writer.purge_checkpoint(&receipt))
+                .expect("purge before gap corruption");
+            {
+                let store = writer.store.lock().expect("actual purged metadata owner");
+                let tx = store.database.begin_write().expect("actual CPUR mutation");
+                let mut meta = tx.open_table(META).expect("actual CPUR table");
+                let mut conflict = receipt.cut();
+                conflict.term += 1;
+                conflict.leader_node += 1;
+                let encoded = checkpoint_baseline::encode_purged(conflict);
+                meta.insert(checkpoint_baseline::PURGED_KEY, encoded.as_slice())
+                    .expect("persist conflicting full purge ID");
+                drop(meta);
+                tx.commit().expect("durable conflicting CPUR");
+                let tx = store.database.begin_read().expect("actual conflicting CPUR read");
+                let meta = tx.open_table(META).expect("actual persisted CPUR");
+                assert!(
+                    matches!(checkpoint_baseline::read_purged(&meta), Err(StoreError::CorruptRecord)),
+                    "CHECKPOINT_FULL_CUT_BOUND: same index with different term/leader is corruption"
+                );
+                drop(meta);
+                drop(tx);
+                let tx = store
+                    .database
+                    .begin_write()
+                    .expect("restore valid purge boundary for gap scenario");
+                let encoded = checkpoint_baseline::encode_purged(receipt.cut());
+                tx.open_table(META)
+                    .expect("actual purge meta")
+                    .insert(checkpoint_baseline::PURGED_KEY, encoded.as_slice())
+                    .expect("restore exact cut");
+                tx.commit().expect("persist exact purge boundary");
+            }
+        }
+        if corruption >= 2 {
+            let mut store = writer.store.lock().expect("actual suffix before tail corruption");
+            apply(&mut store, 3, original.clone());
+        }
+        {
+            let store = writer.store.lock().expect("actual corruption owner");
+            let tx = store.database.begin_write().expect("actual redb mutation");
+            let mut logs = tx.open_table(LOGS).expect("actual log table");
+            if gap {
+                logs.remove(3).expect("remove required suffix entry");
+            } else if corruption >= 2 {
+                let mut tail = id(3);
+                if corruption == 2 {
+                    tail.index += 1;
+                } else {
+                    tail.term += 1;
+                    tail.leader_node += 1;
+                }
+                let encoded = encode_id(tail);
+                tx.open_table(META)
+                    .expect("actual metadata")
+                    .insert("tail", encoded.as_slice())
+                    .expect("corrupt durable tail");
+            } else {
+                let mut entry = store.read_entry(2).expect("retained cut").expect("cut entry");
+                entry.id.term += 1;
+                entry.id.leader_node += 1;
+                let encoded = encode_log(&entry).expect("valid conflicting full ID encoding");
+                logs.insert(2, encoded.as_slice()).expect("replace exact cut identity");
+            }
+            drop(logs);
+            tx.commit().expect("persist actual inconsistent recovery input");
+        }
+        assert!(
+            rt.block_on(writer.purge_checkpoint(&receipt)).is_err(),
+            "CHECKPOINT_PURGE_BOUND: invalid recovery cannot authorize cleanup"
+        );
+        if !gap {
+            let store = writer.store.lock().expect("refused purge retains original evidence");
+            assert!(store.read_entry(1).expect("original log retained").is_some(), "CHECKPOINT_PURGE_BOUND");
+            assert!(
+                store.created_apply_result(1).expect("original response retained").is_some(),
+                "CHECKPOINT_PURGE_BOUND"
+            );
+        }
+        drop(writer);
+        drop(rt);
+        assert!(Store::open(dir.path().join("db"), identity()).is_err(), "CHECKPOINT_RECOVERY_BOUND");
+    }
+}
+
+#[test]
+fn native_checkpoint_publication_fault_selects_complete_generation() {
+    for (sync, after) in [(1, false), (1, true), (2, false), (2, true)] {
+        let (dir, store, original, first) = fixture();
+        let vote = VoteRecord {
+            term: 9,
+            voted_node: 0,
+            committed: true,
+        };
+        let mut store = store;
+        store.save_vote(vote).expect("local hard state before checkpoint");
+        drop(store);
+        let path = dir.path().join("db");
+        let backend = tests::CrashBackend::new(&path);
+        let store = Store::from_database(backend.database(), identity()).expect("existing simulated durability backend");
+        let writer = NativeLogWriter::new(store);
+        let (entered, release) = backend.pause_sync(sync, after);
+        let worker = writer.clone();
+        let task = std::thread::spawn(move || runtime().block_on(worker.checkpoint_local(CheckpointLimits::default())));
+        entered.wait();
+        backend.cut();
+        release.wait();
+        let _unknown = task.join().expect("interrupted publication worker");
+        drop(writer);
+        let reopened = Store::open(&path, identity()).expect("recover complete generation after simulated sync cut");
+        assert!(matches!(schema(&reopened), 6 | 7), "CHECKPOINT_CRASH_ATOMIC");
+        assert_negative(&reopened, &original, &first);
+        assert_eq!(
+            reopened.read_vote().expect("preserved local vote"),
+            Some(vote),
+            "CHECKPOINT_LOCAL_STATE_BOUND"
+        );
+        assert_eq!(reopened.last_applied().expect("complete applied position"), Some(id(2)));
+    }
+}
+
+#[test]
+fn native_checkpoint_recovery_budget_can_be_raised_without_changing_baseline() {
+    use redb::ReadableTableMetadata;
+    let (dir, store, original, first) = fixture();
+    let mut limits = CheckpointLimits::default();
+    let tx = store.database.begin_read().expect("measure real baseline rows");
+    limits.snapshot.records = [
+        application::HEADS,
+        application::VERSIONS,
+        application::EVENTS,
+        application::OPERATIONS,
+    ]
+    .into_iter()
+    .map(|definition| {
+        tx.open_table(definition)
+            .expect("actual baseline table")
+            .len()
+            .expect("actual row count")
+    })
+    .sum();
+    drop(tx);
+    let writer = NativeLogWriter::new(store);
+    let receipt = runtime()
+        .block_on(writer.checkpoint_local(limits))
+        .expect("capture exact bounded baseline");
+    {
+        let mut store = writer.store.lock().expect("actual suffix writer");
+        let mut suffix = command(21);
+        suffix.object.key = b"suffix-new-key".to_vec();
+        suffix.prepared.version_id = [22; 16];
+        apply(&mut store, 3, suffix);
+    }
+    drop(writer);
+    let path = dir.path().join("db");
+    assert!(
+        matches!(Store::open(&path, identity()), Err(StoreError::ResourceLimit)),
+        "CHECKPOINT_RESOURCE_BOUND"
+    );
+    let mut raised = limits;
+    raised.snapshot.records += 32;
+    let reopened =
+        Store::open_with_checkpoint_limits(&path, identity(), raised).expect("explicitly raise bounded recovery budget");
+    assert_negative(&reopened, &original, &first);
+    assert_eq!(reopened.last_applied().expect("retained suffix position"), Some(id(3)));
+    let tx = reopened.database.begin_read().expect("unchanged baseline metadata");
+    let meta = tx.open_table(META).expect("actual metadata");
+    let baseline = checkpoint_baseline::read_metadata(&meta)
+        .expect("baseline decode")
+        .expect("baseline selected");
+    assert_eq!(baseline.generation, receipt.generation());
+    assert_eq!(
+        baseline.limits.snapshot.records, limits.snapshot.records,
+        "CHECKPOINT_RESOURCE_BOUND: override never rewrites evidence"
+    );
+}
+
+#[test]
+fn native_checkpoint_interrupted_purge_reopens_and_retries_idempotently() {
+    for (sync, after) in [(1, false), (1, true), (2, false), (2, true)] {
+        let (dir, store, original, first) = fixture();
+        let writer = NativeLogWriter::new(store);
+        let receipt = runtime()
+            .block_on(writer.checkpoint_local(CheckpointLimits::default()))
+            .expect("durable baseline before purge");
+        drop(writer);
+        let path = dir.path().join("db");
+        let backend = tests::CrashBackend::new(&path);
+        let store = Store::from_database(backend.database(), identity()).expect("simulated purge durability backend");
+        let writer = NativeLogWriter::new(store);
+        let (entered, release) = backend.pause_sync(sync, after);
+        let worker = writer.clone();
+        let task = std::thread::spawn(move || runtime().block_on(worker.purge_checkpoint(&receipt)));
+        entered.wait();
+        backend.cut();
+        release.wait();
+        let _unknown = task.join().expect("interrupted purge worker");
+        drop(writer);
+        let reopened = Store::open(&path, identity()).expect("baseline survives simulated interrupted purge");
+        assert_negative(&reopened, &original, &first);
+        let retained = (0..=2)
+            .map(|index| reopened.read_entry(index).expect("atomic purge log state").is_some())
+            .collect::<Vec<_>>();
+        assert!(
+            retained.iter().all(|present| *present) || retained.iter().all(|present| !*present),
+            "CHECKPOINT_PURGE_ATOMIC: complete old prefix or complete new purge"
+        );
+        let writer = NativeLogWriter::new(reopened);
+        let rt = runtime();
+        rt.block_on(writer.purge_checkpoint(&receipt)).expect("resume exact purge");
+        rt.block_on(writer.purge_checkpoint(&receipt))
+            .expect("repeat exact purge idempotently");
+        drop(writer);
+        let reopened = Store::open(&path, identity()).expect("reopen resumed purge");
+        assert_negative(&reopened, &original, &first);
+        assert!(reopened.read_entry(2).expect("purged cut absent").is_none(), "CHECKPOINT_PURGE_ATOMIC");
+    }
+}
+
+#[test]
+fn native_checkpoint_ack_and_purge_ack_survive_simulated_power_cut() {
+    let (dir, store, original, first) = fixture();
+    drop(store);
+    let path = dir.path().join("db");
+    let backend = tests::CrashBackend::new(&path);
+    let writer =
+        NativeLogWriter::new(Store::from_database(backend.database(), identity()).expect("actual checkpoint fault backend"));
+    let receipt = runtime()
+        .block_on(writer.checkpoint_local(CheckpointLimits::default()))
+        .expect("successful checkpoint ACK");
+    backend.cut();
+    drop(writer);
+    let reopened = Store::open(&path, identity()).expect("ACK checkpoint survives simulated power cut");
+    assert_eq!(schema(&reopened), 7, "CHECKPOINT_ACK_DURABLE");
+    assert_negative(&reopened, &original, &first);
+    let backend = tests::CrashBackend::new(&path);
+    drop(reopened);
+    let writer = NativeLogWriter::new(Store::from_database(backend.database(), identity()).expect("actual purge fault backend"));
+    runtime()
+        .block_on(writer.purge_checkpoint(&receipt))
+        .expect("successful purge ACK");
+    backend.cut();
+    drop(writer);
+    let reopened = Store::open(&path, identity()).expect("ACK purge survives simulated power cut");
+    assert_negative(&reopened, &original, &first);
+    assert!(reopened.read_entry(2).expect("ACK purge cut absent").is_none(), "CHECKPOINT_ACK_DURABLE");
+}
@@ -0,0 +1,182 @@
+//! Owned consistent native application snapshot view.
+use crate::{Store, StoreError, application, snapshot};
+use redb::{ReadTransaction, ReadableDatabase, ReadableTable, ReadableTableMetadata};
+use std::{
+    io::Write,
+    sync::{Arc, Mutex},
+};
+
+#[cfg(test)]
+std::thread_local! {
+    static AFTER_CAPTURE: std::cell::RefCell<Option<Box<dyn FnOnce()>>> = const { std::cell::RefCell::new(None) };
+}
+
+#[cfg(test)]
+pub(crate) struct CaptureHook;
+#[cfg(test)]
+impl Drop for CaptureHook {
+    fn drop(&mut self) {
+        AFTER_CAPTURE.with(|hook| {
+            hook.borrow_mut().take();
+        });
+    }
+}
+#[cfg(test)]
+pub(crate) fn after_capture(hook: impl FnOnce() + 'static) -> CaptureHook {
+    AFTER_CAPTURE.with(|slot| {
+        assert!(slot.borrow().is_none(), "one capture hook per test thread");
+        *slot.borrow_mut() = Some(Box::new(hook));
+    });
+    CaptureHook
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum NativeSnapshotError {
+    #[error(transparent)]
+    Store(#[from] StoreError),
+    #[error(transparent)]
+    Snapshot(#[from] snapshot::SnapshotError),
+    #[error("native snapshot Store lock poisoned")]
+    Poisoned,
+    #[error("native snapshot worker failed")]
+    Worker(#[source] tokio::task::JoinError),
+}
+
+pub struct NativeSnapshotView {
+    // Drop the transaction before the owner so the Database stays open throughout the view.
+    transaction: ReadTransaction,
+    header: snapshot::Header,
+    counts: [u64; 4],
+    schema: u64,
+    _owner: Arc<Mutex<Store>>,
+}
+impl NativeSnapshotView {
+    pub(crate) fn acquire(owner: Arc<Mutex<Store>>) -> Result<Self, NativeSnapshotError> {
+        let transaction = {
+            let store = owner.lock().map_err(|_| NativeSnapshotError::Poisoned)?;
+            if !store.writable {
+                return Err(StoreError::WriteDisabled.into());
+            }
+            store.database.begin_read().map_err(StoreError::from)?
+        };
+        #[cfg(test)]
+        if let Some(hook) = AFTER_CAPTURE.with(|slot| slot.borrow_mut().take()) {
+            hook();
+        }
+        let (applied, membership) = application::native_application_state(&transaction)?;
+        let (header, schema) = {
+            let meta = transaction.open_table(crate::META).map_err(StoreError::from)?;
+            let logs = transaction.open_table(crate::LOGS).map_err(StoreError::from)?;
+            let committed = crate::committed::validate_committed(&meta, &logs)?;
+            if let Some(applied) = applied {
+                let committed = committed.ok_or(StoreError::CorruptRecord)?;
+                if applied.index > committed.index || (applied.index == committed.index && applied != committed) {
+                    return Err(StoreError::CorruptRecord.into());
+                }
+            }
+            let identity = meta
+                .get("identity")
+                .map_err(StoreError::from)?
+                .ok_or(StoreError::CorruptRecord)?;
+            let schema = crate::number(
+                meta.get("schema")
+                    .map_err(StoreError::from)?
+                    .ok_or(StoreError::CorruptRecord)?
+                    .value(),
+            )?;
+            (
+                snapshot::Header {
+                    cluster_id: crate::decode_identity(identity.value())?.cluster_id,
+                    applied,
+                    membership,
+                    binding: application::read_binding(&meta)?,
+                },
+                schema,
+            )
+        };
+        let mut counts = [0; 4];
+        for (index, definition) in [
+            application::HEADS,
+            application::VERSIONS,
+            application::EVENTS,
+            application::OPERATIONS,
+        ]
+        .into_iter()
+        .enumerate()
+        {
+            counts[index] = transaction
+                .open_table(definition)
+                .map_err(StoreError::from)?
+                .len()
+                .map_err(StoreError::from)?;
+        }
+        Ok(Self {
+            transaction,
+            header,
+            counts,
+            schema,
+            _owner: owner,
+        })
+    }
+
+    pub fn stream_capability_artifact<W: Write>(
+        self,
+        writer: W,
+        limits: snapshot::Limits,
+    ) -> Result<snapshot::CapabilityArtifact<W>, NativeSnapshotError> {
+        if !matches!(self.schema, 6 | 7) {
+            return Err(StoreError::UnsupportedSchema.into());
+        }
+        let output = snapshot::CapabilityOutput::new(writer, limits.total_bytes);
+        Ok(self.stream(output, limits)?.finish())
+    }
+
+    pub(crate) fn transaction(&self) -> &ReadTransaction {
+        &self.transaction
+    }
+
+    pub(crate) fn header(&self) -> &snapshot::Header {
+        &self.header
+    }
+
+    pub(crate) fn owner(&self) -> &Arc<Mutex<Store>> {
+        &self._owner
+    }
+
+    pub(crate) fn stream_local_capability<W: Write>(
+        &self,
+        writer: W,
+        limits: snapshot::Limits,
+    ) -> Result<snapshot::CapabilityArtifact<W>, NativeSnapshotError> {
+        if !matches!(self.schema, 6 | 7) {
+            return Err(StoreError::UnsupportedSchema.into());
+        }
+        let output = snapshot::CapabilityOutput::new(writer, limits.total_bytes);
+        Ok(self.stream_borrowed(output, limits)?.finish())
+    }
+
+    pub fn stream<W: Write>(self, writer: W, limits: snapshot::Limits) -> Result<W, NativeSnapshotError> {
+        self.stream_borrowed(writer, limits)
+    }
+
+    fn stream_borrowed<W: Write>(&self, writer: W, limits: snapshot::Limits) -> Result<W, NativeSnapshotError> {
+        let mut encoder = snapshot::Encoder::new(writer, &self.header, self.counts, limits)?;
+        for (definition, section) in [
+            (application::HEADS, snapshot::Section::Heads),
+            (application::VERSIONS, snapshot::Section::Versions),
+            (application::EVENTS, snapshot::Section::Events),
+            (application::OPERATIONS, snapshot::Section::Operations),
+        ] {
+            let table = self.transaction.open_table(definition).map_err(StoreError::from)?;
+            for row in table.iter().map_err(StoreError::from)? {
+                let (key, value) = row.map_err(StoreError::from)?;
+                encoder.row(snapshot::Row {
+                    section,
+                    key: key.value(),
+                    value: value.value(),
+                })?;
+            }
+        }
+        Ok(encoder.finish()?)
+    }
+}
@@ -0,0 +1,197 @@
+//! Bounded snapshot transport bytes; receive ownership follows the buffer into install.
+use std::{
+    io::{self, Read, Seek, SeekFrom, Write},
+    pin::Pin,
+    task::{Context, Poll},
+};
+use tokio::io::{AsyncRead, AsyncSeek, AsyncWrite, ReadBuf};
+
+pub struct NativeSnapshotData {
+    bytes: Vec<u8>,
+    position: u64,
+    limit: u64,
+    pub(crate) receive_guard: Option<crate::raft_state_machine::ReceiveGuard>,
+}
+
+fn boundary() -> io::Error {
+    io::Error::new(io::ErrorKind::InvalidInput, "native snapshot byte limit")
+}
+
+impl NativeSnapshotData {
+    pub fn new(max_bytes: u64) -> io::Result<Self> {
+        if max_bytes == 0 || usize::try_from(max_bytes).is_err() {
+            return Err(boundary());
+        }
+        Ok(Self {
+            bytes: Vec::new(),
+            position: 0,
+            limit: max_bytes,
+            receive_guard: None,
+        })
+    }
+
+    pub(crate) fn as_slice(&self) -> &[u8] {
+        &self.bytes
+    }
+
+    pub(crate) fn attach_receive_guard(&mut self, guard: crate::raft_state_machine::ReceiveGuard) {
+        self.receive_guard = Some(guard);
+    }
+}
+
+impl Read for NativeSnapshotData {
+    fn read(&mut self, output: &mut [u8]) -> io::Result<usize> {
+        let position = usize::try_from(self.position).map_err(|_| boundary())?;
+        let available = self.bytes.get(position..).unwrap_or_default();
+        let count = output.len().min(available.len());
+        output[..count].copy_from_slice(&available[..count]);
+        self.position = self
+            .position
+            .checked_add(u64::try_from(count).map_err(|_| boundary())?)
+            .ok_or_else(boundary)?;
+        Ok(count)
+    }
+}
+
+impl Write for NativeSnapshotData {
+    fn write(&mut self, input: &[u8]) -> io::Result<usize> {
+        if input.is_empty() {
+            return Ok(0);
+        }
+        let count = u64::try_from(input.len()).map_err(|_| boundary())?;
+        let end = self
+            .position
+            .checked_add(count)
+            .filter(|end| *end <= self.limit)
+            .ok_or_else(boundary)?;
+        let position = usize::try_from(self.position).map_err(|_| boundary())?;
+        let length = usize::try_from(end).map_err(|_| boundary())?;
+        if length > self.bytes.len() {
+            self.bytes
+                .try_reserve_exact(length - self.bytes.len())
+                .map_err(|_| io::Error::new(io::ErrorKind::OutOfMemory, "native snapshot allocation"))?;
+            self.bytes.resize(length, 0);
+        }
+        self.bytes[position..length].copy_from_slice(input);
+        self.position = end;
+        Ok(input.len())
+    }
+
+    fn flush(&mut self) -> io::Result<()> {
+        Ok(())
+    }
+}
+
+impl Seek for NativeSnapshotData {
+    fn seek(&mut self, from: SeekFrom) -> io::Result<u64> {
+        let position = match from {
+            SeekFrom::Start(position) => Some(position),
+            SeekFrom::Current(offset) => self.position.checked_add_signed(offset),
+            SeekFrom::End(offset) => u64::try_from(self.bytes.len())
+                .map_err(|_| boundary())?
+                .checked_add_signed(offset),
+        }
+        .filter(|position| *position <= self.limit)
+        .ok_or_else(boundary)?;
+        self.position = position;
+        Ok(position)
+    }
+}
+
+impl AsyncRead for NativeSnapshotData {
+    fn poll_read(self: Pin<&mut Self>, _cx: &mut Context<'_>, output: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+        let this = self.get_mut();
+        let position = match usize::try_from(this.position) {
+            Ok(position) => position,
+            Err(_) => return Poll::Ready(Err(boundary())),
+        };
+        let available = this.bytes.get(position..).unwrap_or_default();
+        let count = output.remaining().min(available.len());
+        let end = this
+            .position
+            .checked_add(u64::try_from(count).map_err(|_| boundary())?)
+            .ok_or_else(boundary)?;
+        output.put_slice(&available[..count]);
+        this.position = end;
+        Poll::Ready(Ok(()))
+    }
+}
+
+impl AsyncWrite for NativeSnapshotData {
+    fn poll_write(self: Pin<&mut Self>, _cx: &mut Context<'_>, input: &[u8]) -> Poll<io::Result<usize>> {
+        Poll::Ready(Write::write(self.get_mut(), input))
+    }
+
+    fn poll_flush(self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        Poll::Ready(Ok(()))
+    }
+
+    fn poll_shutdown(self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        Poll::Ready(Ok(()))
+    }
+}
+
+impl AsyncSeek for NativeSnapshotData {
+    fn start_seek(self: Pin<&mut Self>, from: SeekFrom) -> io::Result<()> {
+        Seek::seek(self.get_mut(), from).map(|_| ())
+    }
+
+    fn poll_complete(self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<io::Result<u64>> {
+        Poll::Ready(Ok(self.position))
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn native_snapshot_data_rejects_growth_before_mutation() {
+        let mut data = NativeSnapshotData::new(4).expect("bounded snapshot");
+        Write::write_all(&mut data, b"ab").expect("initial bytes");
+        assert!(Write::write_all(&mut data, b"cde").is_err());
+        assert_eq!(data.as_slice(), b"ab", "rejected write leaves bytes intact");
+        assert_eq!(Seek::stream_position(&mut data).expect("position"), 2);
+        Seek::seek(&mut data, SeekFrom::Start(4)).expect("seek exactly to limit");
+        Write::write(&mut data, b"").expect("empty write");
+        assert_eq!(data.as_slice(), b"ab", "empty write never allocates a gap");
+        assert!(Write::write(&mut data, b"x").is_err());
+        assert!(Seek::seek(&mut data, SeekFrom::Current(1)).is_err());
+        assert!(Seek::seek(&mut data, SeekFrom::Current(i64::MIN)).is_err());
+        assert_eq!(Seek::stream_position(&mut data).expect("unchanged position"), 4);
+        Seek::seek(&mut data, SeekFrom::Start(3)).expect("bounded sparse position");
+        Write::write_all(&mut data, b"z").expect("write ending exactly at limit");
+        assert_eq!(data.as_slice(), b"ab\0z");
+    }
+
+    #[test]
+    fn native_snapshot_data_checks_length_overflow() {
+        assert!(NativeSnapshotData::new(0).is_err());
+        if usize::BITS == 64 {
+            let mut data = NativeSnapshotData::new(u64::MAX).expect("representable byte limit");
+            Seek::seek(&mut data, SeekFrom::Start(u64::MAX)).expect("seek does not allocate");
+            assert!(Write::write(&mut data, b"x").is_err());
+            assert!(Seek::seek(&mut data, SeekFrom::Current(1)).is_err());
+            assert!(data.as_slice().is_empty(), "overflow never allocates");
+        }
+    }
+
+    #[tokio::test]
+    async fn native_snapshot_data_async_io_uses_the_same_bounds() {
+        use tokio::io::{AsyncReadExt, AsyncSeekExt, AsyncWriteExt};
+        let mut data = NativeSnapshotData::new(4).expect("bounded async snapshot");
+        AsyncWriteExt::write_all(&mut data, b"abcd").await.expect("write exact limit");
+        assert!(AsyncWriteExt::write_all(&mut data, b"e").await.is_err());
+        assert!(AsyncSeekExt::seek(&mut data, SeekFrom::Start(5)).await.is_err());
+        AsyncSeekExt::seek(&mut data, SeekFrom::End(-2))
+            .await
+            .expect("rewind within snapshot");
+        let mut output = Vec::new();
+        AsyncReadExt::read_to_end(&mut data, &mut output)
+            .await
+            .expect("read snapshot suffix");
+        assert_eq!(output, b"cd");
+        AsyncWriteExt::shutdown(&mut data).await.expect("transport shutdown");
+        assert_eq!(data.as_slice(), b"abcd");
+    }
+}
@@ -0,0 +1,217 @@
+//! Validate isolated received bytes before one durable shared-owner publication.
+use crate::{
+    CaptureMembership, CheckpointLimits, NativeLogWriter, NativeSnapshotData, StoreError, application, checkpoint_baseline,
+    native_checkpoint, snapshot,
+};
+use redb::{ReadableDatabase, ReadableTable};
+use std::sync::{
+    Arc,
+    atomic::{AtomicBool, Ordering},
+};
+
+struct Cancel(Arc<AtomicBool>);
+impl Drop for Cancel {
+    fn drop(&mut self) {
+        self.0.store(true, Ordering::Release);
+    }
+}
+struct Stage {
+    database: redb::Database,
+    _directory: tempfile::TempDir,
+    baseline: checkpoint_baseline::BaselineMetadata,
+}
+fn stage(data: &NativeSnapshotData, limits: CheckpointLimits) -> Result<Stage, StoreError> {
+    let bytes = data.as_slice();
+    if u64::try_from(bytes.len()).map_err(|_| StoreError::ResourceLimit)?
+        > limits.snapshot.total_bytes.checked_add(56).ok_or(StoreError::ResourceLimit)?
+    {
+        return Err(StoreError::ResourceLimit);
+    }
+    let manifest = snapshot::CapabilityManifest::decode(bytes.get(..56).ok_or(StoreError::CorruptRecord)?)
+        .map_err(|_| StoreError::CorruptRecord)?;
+    let directory = tempfile::tempdir()?;
+    let database = native_checkpoint::stage_database(directory.path().join("received.redb"), limits.staging_bytes)?;
+    let tx = database.begin_write()?;
+    for definition in [
+        checkpoint_baseline::HEADS,
+        checkpoint_baseline::VERSIONS,
+        checkpoint_baseline::EVENTS,
+        checkpoint_baseline::OPERATIONS,
+    ] {
+        tx.open_table(definition)?;
+    }
+    let header = snapshot::decode_capability_artifact(
+        snapshot::CapabilityArtifact {
+            manifest,
+            payload: &bytes[56..],
+        },
+        &mut native_checkpoint::StageSink(&tx),
+        limits.snapshot,
+    )
+    .map_err(|_| StoreError::CorruptRecord)?;
+    let baseline = checkpoint_baseline::BaselineMetadata {
+        generation: 1,
+        header,
+        limits,
+    };
+    let encoded = checkpoint_baseline::encode_metadata(&baseline)?;
+    {
+        let mut meta = tx.open_table(crate::META)?;
+        meta.insert("schema", 7u64.to_be_bytes().as_slice())?;
+        meta.insert(checkpoint_baseline::META_KEY, encoded.as_slice())?;
+    }
+    tx.commit().map_err(StoreError::StagingCommit)?;
+    checkpoint_baseline::validate_baseline_with_budget(&database.begin_read()?, &baseline, limits.staging_bytes)?;
+    Ok(Stage {
+        database,
+        _directory: directory,
+        baseline,
+    })
+}
+pub(crate) async fn install(
+    writer: NativeLogWriter,
+    limits: CheckpointLimits,
+    meta: openraft::SnapshotMeta<u64, crate::CaptureNode>,
+    data: Box<NativeSnapshotData>,
+) -> Result<(), StoreError> {
+    if data
+        .receive_guard
+        .as_ref()
+        .is_none_or(|guard| !guard.belongs_to(&writer.snapshot_receives))
+    {
+        return Err(StoreError::CorruptRecord);
+    }
+    let (mut stage, data) = tokio::task::spawn_blocking(move || Ok::<_, StoreError>((stage(&data, limits)?, data)))
+        .await
+        .map_err(|_| StoreError::CorruptRecord)??;
+    if meta.last_log_id.map(crate::LogId::from) != stage.baseline.header.applied {
+        return Err(StoreError::CorruptRecord);
+    }
+    if CaptureMembership::try_from(&meta.last_membership)? != stage.baseline.header.membership {
+        return Err(StoreError::CorruptRecord);
+    }
+    let ordering = writer.ordering.clone().lock_owned().await;
+    let cancel = Cancel(Arc::new(AtomicBool::new(false)));
+    let cancelled = cancel.0.clone();
+    tokio::task::spawn_blocking(move || {
+        let _ordering = ordering;
+        let _data = data;
+        if cancelled.load(Ordering::Acquire) {
+            return Err(StoreError::WriteDisabled);
+        }
+        let mut store = writer.store.lock().map_err(|_| StoreError::CorruptRecord)?;
+        let read = store.database.begin_read()?;
+        let old_meta = read.open_table(crate::META)?;
+        let identity = crate::decode_identity(old_meta.get("identity")?.ok_or(StoreError::CorruptRecord)?.value())?;
+        if identity.cluster_id != stage.baseline.header.cluster_id
+            || application::read_binding(&old_meta)? != stage.baseline.header.binding
+        {
+            return Err(StoreError::CorruptRecord);
+        }
+        let old_applied = application::read_applied(&old_meta)?;
+        let old_logs = read.open_table(crate::LOGS)?;
+        let committed = crate::committed::validate_committed(&old_meta, &old_logs)?;
+        let cut = stage.baseline.header.applied;
+        if old_applied.is_some_and(|id| cut.is_none_or(|cut| id.index > cut.index || (id.index == cut.index && id != cut))) {
+            return Err(StoreError::CommittedRegression);
+        }
+        let compatible = match cut {
+            None => true,
+            Some(cut) => {
+                let at_cut = old_logs
+                    .get(cut.index)?
+                    .map(|v| crate::decode_log(v.value()).map(|v| v.id))
+                    .transpose()?
+                    .or(checkpoint_baseline::read_baseline_cut(&old_meta)?.filter(|id| id.index == cut.index));
+                if committed
+                    .is_some_and(|id| id.index >= cut.index && (at_cut != Some(cut) || (id.index == cut.index && id != cut)))
+                {
+                    return Err(StoreError::CommittedRegression);
+                }
+                at_cut == Some(cut)
+            }
+        };
+        let previous = checkpoint_baseline::read_metadata(&old_meta)?;
+        stage.baseline.generation =
+            previous.map_or(Ok(1), |old| old.generation.checked_add(1).ok_or(StoreError::ResourceLimit))?;
+        drop(old_logs);
+        drop(old_meta);
+        drop(read);
+        let source = stage.database.begin_read()?;
+        let tx = store.write_transaction()?;
+        for (baseline, live) in [
+            (checkpoint_baseline::HEADS, application::HEADS),
+            (checkpoint_baseline::VERSIONS, application::VERSIONS),
+            (checkpoint_baseline::EVENTS, application::EVENTS),
+            (checkpoint_baseline::OPERATIONS, application::OPERATIONS),
+        ] {
+            let table = source.open_table(baseline)?;
+            for target in [baseline, live] {
+                let mut target = tx.open_table(target)?;
+                target.retain(|_, _| false)?;
+                for row in table.iter()? {
+                    let (key, value) = row?;
+                    target.insert(key.value(), value.value())?;
+                }
+            }
+        }
+        {
+            let mut index = tx.open_table(crate::source_page::EVENT_SEQUENCE)?;
+            index.retain(|_, _| false)?;
+        }
+        for row in source.open_table(checkpoint_baseline::EVENTS)?.iter()? {
+            let (key, value) = row?;
+            let event = application::decode_event(value.value())?;
+            crate::source_page::insert_event_index(
+                &tx,
+                event.decision_log_id,
+                key.value().try_into().map_err(|_| StoreError::CorruptRecord)?,
+            )?;
+        }
+        let tail = {
+            let mut logs = tx.open_table(crate::LOGS)?;
+            logs.retain(|index, _| compatible && cut.is_none_or(|cut| index > cut.index))?;
+            logs.last()?
+                .map(|(_, v)| crate::decode_log(v.value()).map(|v| v.id))
+                .transpose()?
+                .or(cut)
+        };
+        tx.open_table(application::APPLY_RESULTS)?.retain(|_, _| false)?;
+        {
+            let mut target = tx.open_table(crate::META)?;
+            target.insert("schema", 7u64.to_be_bytes().as_slice())?;
+            target.insert(
+                checkpoint_baseline::META_KEY,
+                checkpoint_baseline::encode_metadata(&stage.baseline)?.as_slice(),
+            )?;
+            let mut applied = vec![0];
+            if let Some(cut) = cut {
+                target.insert(checkpoint_baseline::PURGED_KEY, checkpoint_baseline::encode_purged(cut).as_slice())?;
+                applied = vec![1];
+                applied.extend(crate::encode_id(cut));
+            } else {
+                target.remove(checkpoint_baseline::PURGED_KEY)?;
+            }
+            target.insert("last_applied", applied.as_slice())?;
+            target.insert("native_membership", stage.baseline.header.membership.encode()?.as_slice())?;
+            if let Some(tail) = tail {
+                target.insert("tail", crate::encode_id(tail).as_slice())?;
+            } else {
+                target.remove("tail")?;
+            }
+            let committed = committed.filter(|id| cut.is_none_or(|cut| id.index > cut.index)).or(cut);
+            if let Some(committed) = committed {
+                target.insert("committed", crate::encode_id(committed).as_slice())?;
+            } else {
+                target.remove("committed")?;
+            }
+            crate::raft_state_machine::publish_meta(&mut target, &stage.baseline, meta)?;
+        }
+        if cancelled.load(Ordering::Acquire) {
+            return Err(StoreError::WriteDisabled);
+        }
+        store.commit(tx)
+    })
+    .await
+    .map_err(|_| StoreError::CorruptRecord)?
+}
@@ -0,0 +1,536 @@
+//! Atomic installation failures over real native stores and disk fault backend.
+use super::raft_state_machine_tests::{Fixture, entries, entry, id, install, receive, snapshot_bytes};
+use super::*;
+use openraft::{
+    RaftLogReader, RaftSnapshotBuilder,
+    storage::{RaftLogStorage, RaftLogStorageExt, RaftStateMachine},
+};
+
+async fn media_target() -> (Fixture, tests::CrashBackend) {
+    let mut f = Fixture::new(1);
+    f.apply(entries(1)).await;
+    f.logs
+        .save_vote(&openraft::Vote::new_committed(9, 1))
+        .await
+        .expect("preserved target vote");
+    let Fixture {
+        directory,
+        writer,
+        logs,
+        sm,
+        identity,
+    } = f;
+    drop(sm);
+    drop(logs);
+    drop(writer);
+    let backend = tests::CrashBackend::new(&directory.path().join("db"));
+    let writer = NativeLogWriter::new(Store::from_database(backend.database(), identity).expect("actual redb fault target"));
+    let logs = NativeRaftLogStore::new(writer.clone());
+    let sm = NativeRaftStateMachine::new(writer.clone(), CheckpointLimits::default()).expect("fault target state machine");
+    (
+        Fixture {
+            directory,
+            writer,
+            logs,
+            sm,
+            identity,
+        },
+        backend,
+    )
+}
+
+#[tokio::test]
+async fn native_snapshot_install_corruption_preserves_state() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let mut snapshot = source.build().await;
+    let bytes = snapshot_bytes(&mut snapshot).await;
+    let mut target = Fixture::new(1);
+    target.apply(entries(1)).await;
+    let before = target.page().await;
+    let state = target.sm.applied_state().await.expect("old applied state");
+    let mut corrupted = bytes.clone();
+    *corrupted.last_mut().expect("nonempty snapshot") ^= 1;
+    for payload in [corrupted, bytes[..bytes.len() - 1].to_vec()] {
+        let stable = std::fs::read(target.directory.path().join("db")).expect("old stable bytes");
+        let data = receive(&mut target.sm, &payload).await;
+        assert!(
+            target.sm.install_snapshot(&snapshot.meta, data).await.is_err(),
+            "SNAPSHOT_INSTALL_CORRUPT_FRAME_REJECTED"
+        );
+        assert_eq!(target.sm.applied_state().await.expect("unchanged applied state"), state);
+        assert_eq!(target.page().await, before, "SNAPSHOT_INSTALL_CORRUPTION_NO_PARTIAL_TABLES");
+        assert_eq!(std::fs::read(target.directory.path().join("db")).expect("unchanged stable bytes"), stable);
+        assert!(
+            target
+                .sm
+                .get_current_snapshot()
+                .await
+                .expect("unpublished corrupt snapshot")
+                .is_none()
+        );
+    }
+    install(&mut target.sm, &mut snapshot).await;
+    assert_eq!(target.page().await, source.page().await, "SNAPSHOT_INSTALL_RETRY_AFTER_CORRUPT_RECEIVE");
+}
+
+#[tokio::test]
+async fn native_snapshot_install_stale_cut_rejected() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(1)).await;
+    let mut snapshot = source.build().await;
+    let bytes = snapshot_bytes(&mut snapshot).await;
+    let mut target = Fixture::new(1);
+    target.apply(entries(3)).await;
+    let current = target.build().await.meta;
+    let events = target.page().await;
+    let data = receive(&mut target.sm, &bytes).await;
+    assert!(
+        target.sm.install_snapshot(&snapshot.meta, data).await.is_err(),
+        "SNAPSHOT_INSTALL_NEVER_ROLLS_BACK_APPLIED_STATE"
+    );
+    assert_eq!(target.sm.applied_state().await.expect("non-regressed applied").0, Some(id(3).into()));
+    assert_eq!(
+        target
+            .sm
+            .get_current_snapshot()
+            .await
+            .expect("old current snapshot")
+            .expect("present")
+            .meta,
+        current
+    );
+    assert_eq!(target.page().await, events);
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let mut snapshot = source.build().await;
+    let bytes = snapshot_bytes(&mut snapshot).await;
+    let mut target = Fixture::new(1);
+    let mut batch = entries(3);
+    for entry in &mut batch[2..] {
+        entry.log_id = LogId {
+            term: 9,
+            ..id(entry.log_id.index)
+        }
+        .into();
+    }
+    target
+        .logs
+        .blocking_append(batch)
+        .await
+        .expect("conflicting committed suffix logs");
+    target.sm.apply(entries(1)).await.expect("existing applied prefix");
+    target
+        .logs
+        .save_committed(Some(LogId { term: 9, ..id(3) }.into()))
+        .await
+        .expect("protect conflicting committed suffix");
+    let before = target
+        .logs
+        .get_log_reader()
+        .await
+        .try_get_log_entries(..)
+        .await
+        .expect("protected logs");
+    let data = receive(&mut target.sm, &bytes).await;
+    assert!(
+        target.sm.install_snapshot(&snapshot.meta, data).await.is_err(),
+        "SNAPSHOT_INSTALL_NEVER_DISCARDS_COMMITTED_CONFLICT"
+    );
+    assert_eq!(
+        target
+            .logs
+            .get_log_reader()
+            .await
+            .try_get_log_entries(..)
+            .await
+            .expect("unchanged protected logs"),
+        before
+    );
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn native_snapshot_install_cancelled_owner() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let mut snapshot = source.build().await;
+    let bytes = snapshot_bytes(&mut snapshot).await;
+    let (mut target, backend) = media_target().await;
+    let data = receive(&mut target.sm, &bytes).await;
+    let (entered, resume) = backend.pause_sync(1, false);
+    let mut installing = target.sm.clone();
+    let meta = snapshot.meta.clone();
+    let task = tokio::spawn(async move { installing.install_snapshot(&meta, data).await });
+    entered.wait();
+    assert!(!task.is_finished(), "SNAPSHOT_INSTALL_NOT_ACKED_BEFORE_SYNC");
+    task.abort();
+    assert!(task.await.expect_err("cancelled install waiter").is_cancelled());
+    assert!(
+        target.writer.ordering.try_lock().is_err(),
+        "SNAPSHOT_INSTALL_CANCELLED_WAITER_RETAINS_OWNER"
+    );
+    resume.wait();
+    // Acquiring the production owner waits for the abandoned blocking commit.
+    drop(target.writer.ordering.lock().await);
+    assert_eq!(
+        target.sm.applied_state().await.expect("completed background install").0,
+        Some(id(2).into()),
+        "SNAPSHOT_INSTALL_COMMIT_NOT_INTERRUPTED_BY_WAITER_CANCEL"
+    );
+    let vote = target.logs.read_vote().await.expect("preserved vote");
+    assert_eq!(vote, Some(openraft::Vote::new_committed(9, 1)));
+    backend.cut();
+    let Fixture {
+        directory,
+        writer,
+        logs,
+        sm,
+        identity,
+    } = target;
+    drop(sm);
+    drop(logs);
+    drop(writer);
+    let writer = NativeLogWriter::new(Store::open(directory.path().join("db"), identity).expect("reopen cancelled installation"));
+    let mut sm =
+        NativeRaftStateMachine::new(writer.clone(), CheckpointLimits::default()).expect("reopen installed state machine");
+    assert_eq!(
+        sm.get_current_snapshot()
+            .await
+            .expect("completed durable snapshot")
+            .expect("present")
+            .meta,
+        snapshot.meta
+    );
+    assert_eq!(
+        writer.created_events_page(None, 256).await.expect("installed events"),
+        source.page().await
+    );
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn native_snapshot_install_power_cut_atomic() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let mut snapshot = source.build().await;
+    let bytes = snapshot_bytes(&mut snapshot).await;
+    let new_events = source.page().await;
+    for after_sync in [false, true] {
+        let (mut target, backend) = media_target().await;
+        let old_events = target.page().await;
+        let data = receive(&mut target.sm, &bytes).await;
+        let (entered, resume) = backend.pause_sync(1, after_sync);
+        let meta = snapshot.meta.clone();
+        let mut installing = target.sm.clone();
+        let task = tokio::spawn(async move { installing.install_snapshot(&meta, data).await });
+        entered.wait();
+        backend.cut();
+        resume.wait();
+        assert!(
+            task.await.expect("faulted install task").is_err(),
+            "SNAPSHOT_INSTALL_POWER_CUT_NO_FALSE_ACK"
+        );
+        assert!(
+            target.sm.get_current_snapshot().await.is_err(),
+            "SNAPSHOT_INSTALL_FENCED_HANDLE_CANNOT_EXPORT"
+        );
+        let Fixture {
+            directory,
+            writer,
+            logs,
+            sm,
+            identity,
+        } = target;
+        drop(sm);
+        drop(logs);
+        drop(writer);
+        let writer =
+            NativeLogWriter::new(Store::open(directory.path().join("db"), identity).expect("atomic installation recovery"));
+        let mut logs = NativeRaftLogStore::new(writer.clone());
+        let mut sm = NativeRaftStateMachine::new(writer.clone(), CheckpointLimits::default()).expect("recovered SM");
+        let events = writer
+            .created_events_page(None, 256)
+            .await
+            .expect("complete recovered tables");
+        let cut = sm.applied_state().await.expect("recovered applied marker").0;
+        if events == old_events {
+            assert_eq!(cut, Some(id(1).into()), "SNAPSHOT_INSTALL_OLD_TABLES_OLD_MARKER");
+            assert!(sm.get_current_snapshot().await.expect("old snapshot selection").is_none());
+        } else {
+            assert_eq!(events, new_events, "SNAPSHOT_INSTALL_ALL_TABLES_OLD_OR_NEW");
+            assert_eq!(cut, Some(id(2).into()), "SNAPSHOT_INSTALL_NEW_TABLES_NEW_MARKER");
+            let mut current = sm
+                .get_current_snapshot()
+                .await
+                .expect("new snapshot selection")
+                .expect("new snapshot published");
+            assert_eq!(current.meta, snapshot.meta);
+            assert_eq!(snapshot_bytes(&mut current).await, bytes);
+        }
+        assert_eq!(writer.durable_identity().await.expect("recovered local identity"), identity);
+        assert_eq!(
+            logs.read_vote().await.expect("recovered local vote"),
+            Some(openraft::Vote::new_committed(9, 1))
+        );
+        let state = logs.get_log_state().await.expect("coherent recovered log state");
+        assert!(
+            state
+                .last_log_id
+                .is_some_and(|last| cut.is_some_and(|applied| last.index >= applied.index)),
+            "SNAPSHOT_INSTALL_LOG_STATE_COVERS_APPLIED_CUT"
+        );
+    }
+}
+
+#[tokio::test]
+async fn native_snapshot_install_ack_survives_power_cut() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let mut snapshot = source.build().await;
+    let bytes = snapshot_bytes(&mut snapshot).await;
+    let (mut target, backend) = media_target().await;
+    install(&mut target.sm, &mut snapshot).await;
+    backend.cut();
+    let Fixture {
+        directory,
+        writer,
+        logs,
+        sm,
+        identity,
+    } = target;
+    drop(sm);
+    drop(logs);
+    drop(writer);
+    let writer = NativeLogWriter::new(
+        Store::open(directory.path().join("db"), identity).expect("acknowledged installation survives power cut"),
+    );
+    let mut logs = NativeRaftLogStore::new(writer.clone());
+    let mut sm = NativeRaftStateMachine::new(writer.clone(), CheckpointLimits::default()).expect("reopened durable SM");
+    let mut current = sm
+        .get_current_snapshot()
+        .await
+        .expect("durable current snapshot")
+        .expect("present");
+    assert_eq!(current.meta, snapshot.meta, "SNAPSHOT_INSTALL_ACK_COMPLETE_META_DURABLE");
+    assert_eq!(snapshot_bytes(&mut current).await, bytes);
+    assert_eq!(
+        writer.created_events_page(None, 256).await.expect("durable installed events"),
+        source.page().await
+    );
+    logs.blocking_append(vec![entry(3)]).await.expect("append after ACK recovery");
+    sm.apply(vec![entry(3)]).await.expect("apply after ACK recovery");
+    assert_eq!(sm.applied_state().await.expect("continued state").0, Some(id(3).into()));
+}
+
+#[tokio::test]
+async fn native_snapshot_install_purge_without_pending_rejected() {
+    let mut target = Fixture::new(1);
+    target.apply(entries(1)).await;
+    let before = target.page().await;
+    let bytes = std::fs::read(target.directory.path().join("db")).expect("old unauthorized purge bytes");
+    let result = tokio::time::timeout(std::time::Duration::from_secs(2), target.logs.purge(id(3).into()))
+        .await
+        .expect("no pending receiver means purge cannot wait for hypothetical installation");
+    assert!(result.is_err(), "SNAPSHOT_PURGE_WITHOUT_PENDING_FAILS_IMMEDIATELY");
+    assert_eq!(target.page().await, before);
+    assert_eq!(
+        std::fs::read(target.directory.path().join("db")).expect("rejected purge preserves bytes"),
+        bytes
+    );
+}
+
+async fn wait_for_pending_purge(writer: &NativeLogWriter) {
+    tokio::time::timeout(std::time::Duration::from_secs(5), async {
+        while writer.snapshot_receives.waiters.load(std::sync::atomic::Ordering::Acquire) == 0 {
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .expect("actual trait purge reached pending-receive wait");
+}
+
+#[tokio::test]
+async fn native_snapshot_install_pending_purge_waits_for_install() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(3)).await;
+    let mut snapshot = source.build().await;
+    let bytes = snapshot_bytes(&mut snapshot).await;
+    let mut target = Fixture::new(1);
+    target.apply(entries(1)).await;
+    let data = receive(&mut target.sm, &bytes).await;
+    let mut purging = target.logs.clone();
+    let purge = tokio::spawn(async move { purging.purge(id(3).into()).await });
+    wait_for_pending_purge(&target.writer).await;
+    assert!(!purge.is_finished(), "SNAPSHOT_PURGE_WAITS_FOR_INCOMING_CUT_BEYOND_LOCAL_TAIL");
+    target
+        .sm
+        .install_snapshot(&snapshot.meta, data)
+        .await
+        .expect("actual install while OpenRaft-style purge waits");
+    tokio::time::timeout(std::time::Duration::from_secs(5), purge)
+        .await
+        .expect("pending purge wakes after install")
+        .expect("purge task")
+        .expect("installed checkpoint now authorizes exact purge");
+    assert_eq!(
+        target
+            .logs
+            .get_log_state()
+            .await
+            .expect("installed log state")
+            .last_purged_log_id,
+        Some(id(3).into()),
+        "SNAPSHOT_PENDING_PURGE_USES_DURABLE_INSTALLED_CUT"
+    );
+    assert_eq!(target.page().await, source.page().await);
+    assert_eq!(
+        target
+            .writer
+            .snapshot_receives
+            .waiters
+            .load(std::sync::atomic::Ordering::Acquire),
+        0,
+        "SNAPSHOT_PURGE_WAIT_OWNER_RELEASED"
+    );
+}
+
+#[tokio::test]
+async fn native_snapshot_install_cancelled_receive_wakes_purge() {
+    let mut target = Fixture::new(1);
+    target.apply(entries(1)).await;
+    let page = target.page().await;
+    let before = std::fs::read(target.directory.path().join("db")).expect("stable target before cancelled receive");
+    let data = target.sm.begin_receiving_snapshot().await.expect("actual pending receiver");
+    let mut purging = target.logs.clone();
+    let purge = tokio::spawn(async move { purging.purge(id(3).into()).await });
+    wait_for_pending_purge(&target.writer).await;
+    assert!(!purge.is_finished(), "SNAPSHOT_CANCELLED_RECEIVE_TEST_HAS_REAL_WAITING_PURGE");
+    drop(data);
+    assert!(
+        tokio::time::timeout(std::time::Duration::from_secs(5), purge)
+            .await
+            .expect("cancelled receive wakes pending purge")
+            .expect("purge task")
+            .is_err(),
+        "SNAPSHOT_CANCELLED_RECEIVE_CANNOT_AUTHORIZE_PURGE"
+    );
+    assert_eq!(target.page().await, page);
+    assert_eq!(
+        std::fs::read(target.directory.path().join("db")).expect("retained cancelled-receive state"),
+        before
+    );
+    assert_eq!(
+        target
+            .writer
+            .snapshot_receives
+            .waiters
+            .load(std::sync::atomic::Ordering::Acquire),
+        0
+    );
+}
+
+#[tokio::test]
+async fn native_snapshot_install_failed_receive_wakes_purge() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(3)).await;
+    let mut snapshot = source.build().await;
+    let bytes = snapshot_bytes(&mut snapshot).await;
+    let mut target = Fixture::new(1);
+    target.apply(entries(1)).await;
+    let page = target.page().await;
+    let before = std::fs::read(target.directory.path().join("db")).expect("stable target before failed install");
+    let data = receive(&mut target.sm, &bytes).await;
+    let mut purging = target.logs.clone();
+    let purge = tokio::spawn(async move { purging.purge(id(3).into()).await });
+    wait_for_pending_purge(&target.writer).await;
+    assert!(!purge.is_finished(), "SNAPSHOT_FAILED_INSTALL_TEST_HAS_REAL_WAITING_PURGE");
+    let mut wrong = snapshot.meta.clone();
+    wrong.last_log_id = Some(LogId { term: 9, ..id(3) }.into());
+    assert!(
+        target.sm.install_snapshot(&wrong, data).await.is_err(),
+        "SNAPSHOT_INVALID_META_INSTALL_REJECTED"
+    );
+    assert!(
+        tokio::time::timeout(std::time::Duration::from_secs(5), purge)
+            .await
+            .expect("failed install wakes pending purge")
+            .expect("purge task")
+            .is_err(),
+        "SNAPSHOT_FAILED_INSTALL_CANNOT_AUTHORIZE_PURGE"
+    );
+    assert_eq!(target.page().await, page);
+    assert_eq!(
+        std::fs::read(target.directory.path().join("db")).expect("retained failed-install state"),
+        before
+    );
+    assert_eq!(
+        target
+            .writer
+            .snapshot_receives
+            .waiters
+            .load(std::sync::atomic::Ordering::Acquire),
+        0
+    );
+}
+
+#[tokio::test]
+async fn native_snapshot_install_completed_between_purge_query_and_wait() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let mut snapshot = source.build().await;
+    let data_bytes = snapshot_bytes(&mut snapshot).await;
+    let mut target = Fixture::new(1);
+    let data = receive(&mut target.sm, &data_bytes).await;
+    let entered = std::sync::Arc::new(tokio::sync::Notify::new());
+    let resume = std::sync::Arc::new(tokio::sync::Notify::new());
+    *target.writer.snapshot_receives.purge_recheck.lock().expect("test pause") = Some((entered.clone(), resume.clone()));
+    let mut logs = target.logs.clone();
+    let purge = tokio::spawn(async move { logs.purge(id(2).into()).await });
+    tokio::time::timeout(std::time::Duration::from_secs(5), entered.notified())
+        .await
+        .expect("purge failed coverage under owner before continuation");
+    target
+        .sm
+        .install_snapshot(&snapshot.meta, data)
+        .await
+        .expect("install completes before purge continuation");
+    resume.notify_one();
+    tokio::time::timeout(std::time::Duration::from_secs(5), purge)
+        .await
+        .expect("completed install wake retained")
+        .expect("purge task")
+        .expect("old StaleCheckpoint rechecked after install");
+    assert_eq!(
+        target.page().await,
+        source.page().await,
+        "SNAPSHOT_COMPLETED_INSTALL_RECHECKS_STALE_PURGE"
+    );
+}
+
+#[tokio::test]
+async fn native_snapshot_install_empty_baseline_pending_purge() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let mut snapshot = source.build().await;
+    let bytes = snapshot_bytes(&mut snapshot).await;
+    let mut target = Fixture::new(1);
+    target.sm.build_snapshot().await.expect("real local empty baseline");
+    let data = receive(&mut target.sm, &bytes).await;
+    let mut logs = target.logs.clone();
+    let purge = tokio::spawn(async move { logs.purge(id(2).into()).await });
+    wait_for_pending_purge(&target.writer).await;
+    target
+        .sm
+        .install_snapshot(&snapshot.meta, data)
+        .await
+        .expect("install supersedes empty baseline");
+    tokio::time::timeout(std::time::Duration::from_secs(5), purge)
+        .await
+        .expect("empty coverage wake")
+        .expect("purge task")
+        .expect("new covered purge");
+    assert_eq!(
+        target.page().await,
+        source.page().await,
+        "SNAPSHOT_EMPTY_BASELINE_IS_UNCOVERED_NOT_CORRUPT"
+    );
+}
@@ -0,0 +1,371 @@
+//! Validated capture peer metadata and its independent CNOD v1 codec.
+use serde::{Deserialize, Deserializer, Serialize, Serializer, ser::SerializeStruct};
+
+#[derive(Clone, Debug, Default, PartialEq, Eq)]
+pub struct CaptureNode {
+    rpc_uri: String,
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum CaptureNodeError {
+    #[error("invalid RPC URI")]
+    InvalidRpcUri,
+    #[error("invalid RPC URI")]
+    InvalidRpcUriParse(#[source] url::ParseError),
+    #[error("RPC URI too long")]
+    RpcUriTooLong,
+    #[error("invalid node magic")]
+    InvalidNodeMagic,
+    #[error("unsupported node version")]
+    UnsupportedNodeVersion,
+    #[error("invalid node length")]
+    InvalidNodeLength,
+    #[error("truncated node record")]
+    TruncatedNodeRecord,
+    #[error("invalid node UTF-8")]
+    InvalidNodeUtf8(#[source] std::str::Utf8Error),
+    #[error("trailing node bytes")]
+    TrailingNodeBytes,
+}
+impl CaptureNodeError {
+    pub fn code(&self) -> &'static str {
+        match self {
+            Self::InvalidRpcUri | Self::InvalidRpcUriParse(_) => "invalid_rpc_uri",
+            Self::RpcUriTooLong => "rpc_uri_too_long",
+            Self::InvalidNodeMagic => "invalid_node_magic",
+            Self::UnsupportedNodeVersion => "unsupported_node_version",
+            Self::InvalidNodeLength => "invalid_node_length",
+            Self::TruncatedNodeRecord => "truncated_node_record",
+            Self::InvalidNodeUtf8(_) => "invalid_node_utf8",
+            Self::TrailingNodeBytes => "trailing_node_bytes",
+        }
+    }
+}
+impl CaptureNode {
+    pub fn new(rpc_uri: String) -> Result<Self, CaptureNodeError> {
+        let node = Self { rpc_uri };
+        node.validate()?;
+        Ok(node)
+    }
+
+    pub fn rpc_uri(&self) -> &str {
+        &self.rpc_uri
+    }
+
+    pub fn validate(&self) -> Result<(), CaptureNodeError> {
+        let uri = &self.rpc_uri;
+        if uri.len() > 4096 {
+            return Err(CaptureNodeError::RpcUriTooLong);
+        }
+        if uri.is_empty() || uri.chars().any(|c| c.is_whitespace() || c.is_control() || c == '\\') {
+            return Err(CaptureNodeError::InvalidRpcUri);
+        }
+        let rest = uri
+            .strip_prefix("http://")
+            .or_else(|| uri.strip_prefix("https://"))
+            .ok_or(CaptureNodeError::InvalidRpcUri)?;
+        if uri.contains('?') || uri.contains('#') {
+            return Err(CaptureNodeError::InvalidRpcUri);
+        }
+        let (authority, path) = rest.split_once('/').map_or((rest, ""), |(a, p)| (a, p));
+        // An empty suffix after the first slash is the only allowed path.
+        if !path.is_empty() || authority.contains('@') {
+            return Err(CaptureNodeError::InvalidRpcUri);
+        }
+        let (host, port) = if authority.starts_with('[') {
+            let end = authority.find(']').ok_or(CaptureNodeError::InvalidRpcUri)?;
+            let port = authority
+                .get(end + 1..)
+                .and_then(|suffix| suffix.strip_prefix(':'))
+                .ok_or(CaptureNodeError::InvalidRpcUri)?;
+            (&authority[..=end], port)
+        } else {
+            authority.split_once(':').ok_or(CaptureNodeError::InvalidRpcUri)?
+        };
+        if host.is_empty() || port.is_empty() || !port.bytes().all(|b| b.is_ascii_digit()) {
+            return Err(CaptureNodeError::InvalidRpcUri);
+        }
+        let port = port.parse::<u16>().map_err(|_| CaptureNodeError::InvalidRpcUri)?;
+        if port == 0 {
+            return Err(CaptureNodeError::InvalidRpcUri);
+        }
+        let parsed_url = url::Url::parse(uri).map_err(CaptureNodeError::InvalidRpcUriParse)?;
+        if parsed_url.host().is_none() {
+            return Err(CaptureNodeError::InvalidRpcUri);
+        }
+        Ok(())
+    }
+
+    pub fn encode(&self) -> Result<Vec<u8>, CaptureNodeError> {
+        self.validate()?;
+        let length = u32::try_from(self.rpc_uri.len()).map_err(|_| CaptureNodeError::InvalidNodeLength)?;
+        let mut bytes = b"CNOD\x01".to_vec();
+        bytes.extend_from_slice(&length.to_be_bytes());
+        bytes.extend_from_slice(self.rpc_uri.as_bytes());
+        Ok(bytes)
+    }
+
+    pub fn decode(bytes: &[u8]) -> Result<Self, CaptureNodeError> {
+        if bytes.len() < 9 {
+            return Err(CaptureNodeError::TruncatedNodeRecord);
+        }
+        if &bytes[..4] != b"CNOD" {
+            return Err(CaptureNodeError::InvalidNodeMagic);
+        }
+        if bytes[4] != 1 {
+            return Err(CaptureNodeError::UnsupportedNodeVersion);
+        }
+        let length = u32::from_be_bytes(bytes[5..9].try_into().map_err(|_| CaptureNodeError::InvalidNodeLength)?);
+        if !(1..=4096).contains(&length) {
+            return Err(CaptureNodeError::InvalidNodeLength);
+        }
+        let length = usize::try_from(length).map_err(|_| CaptureNodeError::InvalidNodeLength)?;
+        let end = 9_usize.checked_add(length).ok_or(CaptureNodeError::InvalidNodeLength)?;
+        if bytes.len() < end {
+            return Err(CaptureNodeError::TruncatedNodeRecord);
+        }
+        if bytes.len() != end {
+            return Err(CaptureNodeError::TrailingNodeBytes);
+        }
+        let uri = std::str::from_utf8(&bytes[9..end]).map_err(CaptureNodeError::InvalidNodeUtf8)?;
+        Self::new(uri.to_owned())
+    }
+}
+impl Serialize for CaptureNode {
+    fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
+        self.validate().map_err(serde::ser::Error::custom)?;
+        let mut state = serializer.serialize_struct("CaptureNode", 1)?;
+        state.serialize_field("rpc_uri", &self.rpc_uri)?;
+        state.end()
+    }
+}
+impl<'de> Deserialize<'de> for CaptureNode {
+    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
+        #[derive(Deserialize)]
+        #[serde(deny_unknown_fields)]
+        struct Fields {
+            rpc_uri: String,
+        }
+        let fields = Fields::deserialize(deserializer).map_err(|_| serde::de::Error::custom("invalid capture node fields"))?;
+        Self::new(fields.rpc_uri).map_err(serde::de::Error::custom)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use std::error::Error;
+
+    fn record(payload: &[u8]) -> Vec<u8> {
+        let mut bytes = b"CNOD\x01".to_vec();
+        bytes.extend_from_slice(&u32::try_from(payload.len()).expect("test payload length").to_be_bytes());
+        bytes.extend_from_slice(payload);
+        bytes
+    }
+    fn reject_uri(uri: &str) {
+        let result = CaptureNode::new(uri.to_owned());
+        assert!(result.is_err(), "URI should reject: {uri:?}");
+        assert_eq!(result.expect_err("rejected URI").code(), "invalid_rpc_uri");
+    }
+    fn reject_record(bytes: &[u8], code: &str) {
+        let result = CaptureNode::decode(bytes);
+        assert!(result.is_err(), "record should reject with {code}");
+        assert_eq!(result.expect_err("rejected record").code(), code);
+    }
+    #[test]
+    fn actual_openraft_node_bound() {
+        fn require_node<T: openraft::Node>() {}
+        require_node::<CaptureNode>();
+        assert_eq!(CaptureNode::default().rpc_uri(), "");
+    }
+    #[test]
+    fn default_and_empty_rejected_at_every_boundary() {
+        let node = CaptureNode::default();
+        assert_eq!(node.rpc_uri(), "");
+        reject_uri("");
+        assert_eq!(node.validate().expect_err("empty validation").code(), "invalid_rpc_uri");
+        assert!(serde_json::to_string(&node).is_err());
+        assert!(serde_json::from_str::<CaptureNode>(r#"{"rpc_uri":""}"#).is_err());
+        assert_eq!(node.encode().expect_err("empty encoding").code(), "invalid_rpc_uri");
+        reject_record(&record(b""), "invalid_node_length");
+    }
+    #[test]
+    fn accepted_uris_preserve_exact_bytes() {
+        for uri in [
+            "http://node:80",
+            "https://node:443",
+            "http://node:1",
+            "https://node:65535/",
+            "http://127.0.0.1:9000",
+            "https://[::1]:443/",
+            "http://[2001:db8::1]:1",
+            "http://NoDe:00080/",
+        ] {
+            let result = CaptureNode::new(uri.to_owned());
+            assert!(result.is_ok(), "valid URI construction: {uri}");
+            let node = result.expect("valid URI");
+            assert_eq!(node.rpc_uri(), uri, "preserve explicit default port and raw URI");
+            assert_eq!(node.encode().expect("encode URI"), record(uri.as_bytes()));
+            assert_eq!(CaptureNode::decode(&record(uri.as_bytes())).expect("decode URI"), node);
+        }
+    }
+    #[test]
+    fn golden_cnod_bytes() {
+        let node = CaptureNode {
+            rpc_uri: "http://node:80".to_owned(),
+        };
+        let golden = b"CNOD\x01\x00\x00\x00\x0ehttp://node:80";
+        assert_eq!(node.encode().ok().as_deref(), Some(golden.as_slice()), "golden CNOD bytes");
+    }
+    #[test]
+    fn valid_cnod_decode() {
+        let golden = b"CNOD\x01\x00\x00\x00\x0ehttp://node:80";
+        let decoded = CaptureNode::decode(golden);
+        assert!(decoded.is_ok(), "golden decode succeeds");
+        assert_eq!(decoded.expect("golden decoded node").rpc_uri(), "http://node:80");
+    }
+    #[test]
+    fn serde_shape_roundtrip_and_strict_fields() {
+        let node = CaptureNode::new("https://node:443/".to_owned()).expect("serde node");
+        let json = r#"{"rpc_uri":"https://node:443/"}"#;
+        assert_eq!(serde_json::to_string(&node).expect("serialize node"), json);
+        assert_eq!(serde_json::from_str::<CaptureNode>(json).expect("deserialize node"), node);
+        for json in [
+            r#"{}"#,
+            r#"{"rpc_uri":1}"#,
+            r#"{"rpc_uri":null}"#,
+            r#"{"rpc_uri":"http://node:1","extra":0}"#,
+            r#"{"rpc_uri":"http://node:1","rpc_uri":"http://node:2"}"#,
+        ] {
+            assert!(serde_json::from_str::<CaptureNode>(json).is_err(), "strict fields: {json}");
+        }
+    }
+    #[test]
+    fn credentials_and_raw_uri_shapes_reject() {
+        for uri in [
+            "http://user@node:1",
+            "http://user:password@node:1",
+            "http://@node:1",
+            "HTTP://node:1",
+            "Https://node:1",
+            "http://node",
+            "http://node:",
+            "http://node:0",
+            "http://node:-1",
+            "http://node:+1",
+            "http://node:65536",
+            "http://node:999999999999999999999",
+            "http://[::1:1",
+            "http://[::1]extra:1",
+            "http://[invalid]:1",
+            "http://::1:1",
+            "http://:1",
+            "/disk/path",
+            "file:///disk/path",
+            "http://node:1/./",
+            "http://node:1/../",
+            "http://node:1/%2e/",
+            "http://node:1/%2E%2e/",
+            "http://node:1/path",
+            "http://node:1\\",
+            " http://node:1",
+            "http://node:1\n",
+            "http://no\tde:1",
+            "http://node:1\u{00a0}",
+            "http://node:1\u{0085}",
+            "http://node:1\0",
+        ] {
+            reject_uri(uri);
+        }
+    }
+    #[test]
+    fn query_and_fragment_markers_reject() {
+        for uri in [
+            "http://node:1?",
+            "http://node:1?x=y",
+            "http://node:1/?",
+            "http://node:1#",
+            "http://node:1#fragment",
+            "http://node:1/#",
+        ] {
+            reject_uri(uri);
+        }
+    }
+    #[test]
+    fn utf8_byte_length_boundaries() {
+        for length in [4095, 4096] {
+            let prefix = "http://node:";
+            let uri = format!("{prefix}{}1", "0".repeat(length - prefix.len() - 1));
+            assert_eq!(uri.len(), length);
+            assert!(url::Url::parse(&uri).is_ok(), "boundary fixture accepted by actual URL parser");
+            let node = CaptureNode::new(uri.clone()).expect("boundary node");
+            assert_eq!(node.rpc_uri(), uri);
+            assert_eq!(
+                CaptureNode::decode(&node.encode().expect("boundary encoding")).expect("boundary decoding"),
+                node
+            );
+        }
+        let uri = format!("http://node:{}1", "0".repeat(4097 - "http://node:".len() - 1));
+        assert_eq!(CaptureNode::new(uri).expect_err("oversize URI").code(), "rpc_uri_too_long");
+        let unicode = format!("http://{}:1", "é".repeat(2044));
+        assert!(unicode.chars().count() < 4096);
+        assert!(unicode.len() > 4096);
+        assert_eq!(CaptureNode::new(unicode).expect_err("UTF-8 byte limit").code(), "rpc_uri_too_long");
+    }
+    #[test]
+    fn cnod_headers_and_lengths_reject_exact_codes() {
+        let valid = record(b"http://node:80");
+        for end in 0..9 {
+            reject_record(&valid[..end], "truncated_node_record");
+        }
+        let mut bad = valid.clone();
+        bad[0] = b'X';
+        reject_record(&bad, "invalid_node_magic");
+        bad = valid.clone();
+        bad[4] = 2;
+        reject_record(&bad, "unsupported_node_version");
+        for length in [0_u32, 4097, u32::MAX] {
+            bad = b"CNOD\x01".to_vec();
+            bad.extend_from_slice(&length.to_be_bytes());
+            reject_record(&bad, "invalid_node_length");
+        }
+        for end in 9..valid.len() {
+            reject_record(&valid[..end], "truncated_node_record");
+        }
+        reject_record(&record(&[0xff]), "invalid_node_utf8");
+        reject_record(&record(&[0xc3, 0x28]), "invalid_node_utf8");
+        reject_record(&record(b"invalid URI"), "invalid_rpc_uri");
+    }
+    #[test]
+    fn trailing_bytes_reject() {
+        for tail in [&b"x"[..], &b"xyz"[..]] {
+            let mut bytes = record(b"http://node:80");
+            bytes.extend_from_slice(tail);
+            reject_record(&bytes, "trailing_node_bytes");
+        }
+    }
+    #[test]
+    fn errors_and_sources_redact_input() {
+        let sentinel = "http://SECRET_USER:SECRET_PASSWORD@node:1";
+        let error = CaptureNode::new(sentinel.to_owned()).expect_err("credentials reject");
+        let mut current: &dyn Error = &error;
+        loop {
+            let rendered = format!("{current} {current:?}");
+            assert!(!rendered.contains("SECRET"));
+            match current.source() {
+                Some(source) => current = source,
+                None => break,
+            }
+        }
+        let json = format!(r#"{{"rpc_uri":"{sentinel}"}}"#);
+        let error = serde_json::from_str::<CaptureNode>(&json).expect_err("serde credentials reject");
+        assert!(!format!("{error} {error:?}").contains("SECRET"));
+        let json = format!(r#"{{"{sentinel}":0}}"#);
+        let error = serde_json::from_str::<CaptureNode>(&json).expect_err("unknown fields reject");
+        assert!(!format!("{error} {error:?}").contains("SECRET"));
+        let error = CaptureNode::new("http://[bad]:1".to_owned()).expect_err("parser rejection");
+        assert!(error.source().is_some(), "retain URL parser source");
+        let error = CaptureNode::decode(&record(&[0xff])).expect_err("UTF-8 rejection");
+        assert!(error.source().is_some(), "retain UTF-8 source");
+    }
+}
@@ -0,0 +1,175 @@
+//! OpenRaft log storage backed by the existing native writer and local durable baseline.
+use crate::{CaptureRaftEntry, CaptureTypeConfig, NativeLogReader, NativeLogWriter, NativeWriteError, StoreError};
+use openraft::{
+    LogId, OptionalSend, RaftLogReader, StorageError, StorageIOError, Vote,
+    storage::{LogFlushed, RaftLogStorage},
+};
+use redb::{ReadableDatabase, ReadableTable};
+
+/// Clones share one native Store, write ordering and checkpoint authority.
+#[derive(Clone)]
+pub struct NativeRaftLogStore {
+    writer: NativeLogWriter,
+    // Last: owner drain is signaled only after this clone releases its writer.
+    pub(crate) owner_lease: Option<std::sync::Arc<crate::raft_owner::OwnerLease>>,
+}
+impl NativeRaftLogStore {
+    pub fn new(writer: NativeLogWriter) -> Self {
+        Self {
+            writer,
+            owner_lease: None,
+        }
+    }
+}
+fn error(error: NativeWriteError, operation: fn(openraft::AnyError) -> StorageIOError<u64>) -> StorageError<u64> {
+    StorageError::IO {
+        source: operation(openraft::AnyError::new(&error)),
+    }
+}
+impl RaftLogReader<CaptureTypeConfig> for NativeRaftLogStore {
+    async fn try_get_log_entries<RB: std::ops::RangeBounds<u64> + Clone + std::fmt::Debug + OptionalSend>(
+        &mut self,
+        range: RB,
+    ) -> Result<Vec<CaptureRaftEntry>, StorageError<u64>> {
+        let mut reader = self.writer.reader();
+        reader.owner_lease = self.owner_lease.clone();
+        reader.try_get_log_entries(range).await
+    }
+}
+impl RaftLogStorage<CaptureTypeConfig> for NativeRaftLogStore {
+    type LogReader = NativeLogReader;
+    async fn get_log_state(&mut self) -> Result<openraft::LogState<CaptureTypeConfig>, StorageError<u64>> {
+        self.writer
+            .write(|store| {
+                if !store.writable {
+                    return Err(StoreError::WriteDisabled);
+                }
+                let read = store.database.begin_read()?;
+                let meta = read.open_table(crate::META)?;
+                let purged = crate::checkpoint_baseline::read_purged(&meta)?;
+                let logs = read.open_table(crate::LOGS)?;
+                let last = logs
+                    .last()?
+                    .map(|(index, value)| {
+                        let id = crate::decode_log(value.value())?.id;
+                        if id.index != index.value() {
+                            return Err(StoreError::CorruptRecord);
+                        }
+                        Ok(id)
+                    })
+                    .transpose()?
+                    .or(purged);
+                let tail = meta.get("tail")?.map(|value| crate::decode_id(value.value())).transpose()?;
+                if tail != last {
+                    return Err(StoreError::CorruptRecord);
+                }
+                Ok(openraft::LogState {
+                    last_purged_log_id: purged.map(Into::into),
+                    last_log_id: last.map(Into::into),
+                })
+            })
+            .await
+            .map_err(|e| error(e, StorageIOError::read_logs))
+    }
+    async fn get_log_reader(&mut self) -> Self::LogReader {
+        let mut reader = self.writer.reader();
+        reader.owner_lease = self.owner_lease.clone();
+        reader
+    }
+    async fn save_vote(&mut self, vote: &Vote<u64>) -> Result<(), StorageError<u64>> {
+        self.writer
+            .save_vote(*vote)
+            .await
+            .map_err(|e| error(e, StorageIOError::write_vote))
+    }
+    async fn read_vote(&mut self) -> Result<Option<Vote<u64>>, StorageError<u64>> {
+        self.writer.read_vote().await.map_err(|e| error(e, StorageIOError::read_vote))
+    }
+    async fn save_committed(&mut self, committed: Option<LogId<u64>>) -> Result<(), StorageError<u64>> {
+        self.writer
+            .write(move |store| match committed {
+                Some(id) => store.save_committed(id.into()),
+                None if store.read_committed()?.is_none() => Ok(()),
+                None => Err(StoreError::CommittedRegression),
+            })
+            .await
+            .map_err(|e| error(e, StorageIOError::write_logs))
+    }
+    async fn read_committed(&mut self) -> Result<Option<LogId<u64>>, StorageError<u64>> {
+        self.writer
+            .write(|store| store.read_committed().map(|id| id.map(Into::into)))
+            .await
+            .map_err(|e| error(e, StorageIOError::read_logs))
+    }
+    async fn append<I>(&mut self, entries: I, callback: LogFlushed<CaptureTypeConfig>) -> Result<(), StorageError<u64>>
+    where
+        I: IntoIterator<Item = CaptureRaftEntry> + OptionalSend,
+        I::IntoIter: OptionalSend,
+    {
+        let entries = entries.into_iter().collect::<Vec<_>>();
+        // The original blocking worker owns ordering and callback through durable commit,
+        // even if OpenRaft cancels its awaiting append future.
+        self.writer
+            .write(move |store| {
+                let result = store.append_native(&entries);
+                callback.log_io_completed(
+                    result
+                        .as_ref()
+                        .map(|_| ())
+                        .map_err(|error| std::io::Error::other(error.code())),
+                );
+                result
+            })
+            .await
+            .map_err(|e| error(e, StorageIOError::write_logs))
+    }
+    async fn truncate(&mut self, log_id: LogId<u64>) -> Result<(), StorageError<u64>> {
+        self.writer
+            .truncate_suffix(log_id.index)
+            .await
+            .map_err(|e| error(e, StorageIOError::write_logs))
+    }
+    async fn purge(&mut self, log_id: LogId<u64>) -> Result<(), StorageError<u64>> {
+        let deadline = tokio::time::Instant::now() + std::time::Duration::from_secs(30);
+        loop {
+            if tokio::time::Instant::now() >= deadline {
+                return Err(error(NativeWriteError::Store(StoreError::StaleCheckpoint), StorageIOError::write_logs));
+            }
+            let notified = self.writer.snapshot_receives.changed.notified();
+            tokio::pin!(notified);
+            notified.as_mut().enable();
+            let receives = self.writer.snapshot_receives.clone();
+            // Observe coverage and receive ownership while install is excluded by ordering.
+            let (result, pending) = self
+                .writer
+                .write(move |store| Ok((store.purge_native_prefix(log_id.into(), None), receives.pending())))
+                .await
+                .map_err(|e| error(e, StorageIOError::write_logs))?;
+            if matches!(&result, Err(StoreError::StaleCheckpoint)) {
+                #[cfg(test)]
+                {
+                    let hook = self
+                        .writer
+                        .snapshot_receives
+                        .purge_recheck
+                        .lock()
+                        .expect("test hook lock")
+                        .take();
+                    if let Some((entered, resume)) = hook {
+                        entered.notify_one();
+                        resume.notified().await;
+                    }
+                }
+                if pending {
+                    #[cfg(test)]
+                    let _waiter = crate::raft_state_machine::Waiter::new(self.writer.snapshot_receives.clone());
+                    if tokio::time::timeout_at(deadline, notified).await.is_err() {
+                        return Err(error(NativeWriteError::Store(StoreError::StaleCheckpoint), StorageIOError::write_logs));
+                    }
+                    continue;
+                }
+            }
+            return result.map_err(|e| error(e.into(), StorageIOError::write_logs));
+        }
+    }
+}
@@ -0,0 +1,439 @@
+//! Exercises actual OpenRaft storage traits, not a substitute consensus engine.
+use super::*;
+use openraft::{
+    RaftLogReader,
+    storage::{RaftLogStorage, RaftLogStorageExt},
+};
+use std::collections::{BTreeMap, BTreeSet};
+
+fn identity() -> NodeIdentity {
+    NodeIdentity {
+        cluster_id: [7; 16],
+        node_id: 0,
+    }
+}
+fn binding() -> CaptureBinding {
+    CaptureBinding {
+        revision: 9,
+        binding_id: [8; 16],
+        target_id: [9; 16],
+    }
+}
+fn id(index: u64) -> LogId {
+    LogId {
+        term: 8,
+        leader_node: 0,
+        index,
+    }
+}
+fn entry(index: u64) -> CaptureRaftEntry {
+    let payload = if index == 0 {
+        let membership = CaptureMembership::new(
+            Some(id(0)),
+            vec![BTreeSet::from([0])],
+            BTreeMap::from([(0, CaptureNode::new("http://127.0.0.1:7000".into()).expect("test member"))]),
+        )
+        .expect("native membership");
+        let membership = openraft::StoredMembership::try_from(&membership).expect("stored native membership");
+        openraft::EntryPayload::Membership(membership.membership().clone())
+    } else {
+        let byte = u8::try_from(index).expect("small test event index");
+        let command = DecideCreated {
+            operation_id: [byte; 16],
+            object: ObjectIdentity {
+                bucket_incarnation: [2; 16],
+                key: vec![byte],
+            },
+            expected_head: None,
+            prepared: PreparedIdentity {
+                version_id: [byte; 16],
+                preparation_id: [4; 16],
+                content_digest: [5; 32],
+                content_length: 123,
+            },
+            binding_revision: binding().revision,
+        };
+        openraft::EntryPayload::Normal(CaptureCommand::try_created(command).expect("native test command"))
+    };
+    CaptureRaftEntry {
+        log_id: id(index).into(),
+        payload,
+    }
+}
+fn entries(end: u64) -> Vec<CaptureRaftEntry> {
+    (0..=end).map(entry).collect()
+}
+fn fixture() -> (tempfile::TempDir, NativeLogWriter, NativeRaftLogStore) {
+    let directory = tempfile::tempdir().expect("actual native database directory");
+    let mut store = Store::initialize(directory.path().join("db"), identity()).expect("actual redb initialization");
+    store.pin_capture_binding(binding()).expect("capture binding");
+    store.enable_native_empty().expect("native mode");
+    store.promote_created_provenance().expect("native provenance");
+    let writer = NativeLogWriter::new(store);
+    let adapter = NativeRaftLogStore::new(writer.clone());
+    (directory, writer, adapter)
+}
+fn media_fixture() -> (tempfile::TempDir, tests::CrashBackend, NativeLogWriter, NativeRaftLogStore) {
+    let (directory, writer, adapter) = fixture();
+    drop(adapter);
+    drop(writer);
+    let backend = tests::CrashBackend::new(&directory.path().join("db"));
+    let store = Store::from_database(backend.database(), identity()).expect("existing real disk fault backend");
+    let writer = NativeLogWriter::new(store);
+    let adapter = NativeRaftLogStore::new(writer.clone());
+    (directory, backend, writer, adapter)
+}
+async fn logs(adapter: &mut NativeRaftLogStore) -> Vec<CaptureRaftEntry> {
+    adapter
+        .get_log_reader()
+        .await
+        .try_get_log_entries(..)
+        .await
+        .expect("actual trait reader")
+}
+async fn checkpoint(writer: &NativeLogWriter, adapter: &mut NativeRaftLogStore) -> CheckpointReceipt {
+    let batch = entries(2);
+    adapter
+        .blocking_append(batch.clone())
+        .await
+        .expect("trait append before checkpoint");
+    // Prepares genuine applied/checkpoint evidence; this does not claim Raft consensus.
+    writer
+        .apply_committed(batch)
+        .await
+        .expect("trusted test owner applies preparation entries");
+    writer
+        .checkpoint_local(CheckpointLimits::default())
+        .await
+        .expect("real durable baseline")
+}
+
+#[tokio::test]
+async fn raft_log_storage_flush_restart() {
+    let (directory, backend, writer, mut adapter) = media_fixture();
+    let batch = entries(2);
+    // blocking_append constructs the real library-private LogFlushed callback.
+    adapter
+        .blocking_append(batch.clone())
+        .await
+        .expect("real durable flush callback");
+    let state = adapter.get_log_state().await.expect("trait log state");
+    assert_eq!(state.last_log_id, Some(id(2).into()), "RAFT_LOG_TAIL_NOT_APPLIED_STATE");
+    assert_eq!(state.last_purged_log_id, None);
+    assert_eq!(writer.applied_state().await.expect("application state").0, None);
+    assert_eq!(logs(&mut adapter).await, batch);
+    backend.cut();
+    drop(adapter);
+    drop(writer);
+    let reopened = Store::open(directory.path().join("db"), identity()).expect("reopen after successful callback and power cut");
+    let mut adapter = NativeRaftLogStore::new(NativeLogWriter::new(reopened));
+    assert_eq!(logs(&mut adapter).await, batch, "RAFT_LOG_FLUSH_CALLBACK_DURABLE");
+    assert_eq!(
+        adapter.get_log_state().await.expect("reopened trait state").last_log_id,
+        Some(id(2).into())
+    );
+}
+
+#[tokio::test]
+async fn raft_log_storage_vote_restart() {
+    let (directory, backend, writer, mut adapter) = media_fixture();
+    let vote = openraft::Vote::new_committed(9, 0);
+    adapter.save_vote(&vote).await.expect("trait durable vote");
+    assert_eq!(adapter.read_vote().await.expect("trait vote read"), Some(vote));
+    adapter.blocking_append(entries(2)).await.expect("trait append");
+    adapter
+        .save_committed(Some(id(1).into()))
+        .await
+        .expect("trait durable commit marker");
+    assert!(adapter.save_committed(None).await.is_err(), "RAFT_LOG_COMMITTED_NONE_CANNOT_RESET");
+    assert_eq!(adapter.read_committed().await.expect("exact marker"), Some(id(1).into()));
+    assert!(
+        adapter.save_committed(Some(id(0).into())).await.is_err(),
+        "RAFT_LOG_COMMITTED_NO_REGRESSION"
+    );
+    let mut wrong = id(2);
+    wrong.term += 1;
+    assert!(
+        adapter.save_committed(Some(wrong.into())).await.is_err(),
+        "RAFT_LOG_COMMITTED_EXACT_IDENTITY"
+    );
+    backend.cut();
+    drop(adapter);
+    drop(writer);
+    let mut reopened = NativeRaftLogStore::new(NativeLogWriter::new(
+        Store::open(directory.path().join("db"), identity()).expect("durable vote reopen"),
+    ));
+    assert_eq!(
+        reopened.read_vote().await.expect("reopened vote"),
+        Some(vote),
+        "RAFT_LOG_VOTE_ACK_DURABLE"
+    );
+    assert_eq!(reopened.read_committed().await.expect("reopened committed marker"), Some(id(1).into()));
+}
+
+#[tokio::test]
+async fn raft_log_storage_partial_purge_restart() {
+    let (directory, writer, mut adapter) = fixture();
+    let receipt = checkpoint(&writer, &mut adapter).await;
+    adapter
+        .blocking_append(vec![entry(3), entry(4)])
+        .await
+        .expect("uncovered suffix append");
+    let events = writer
+        .created_events_page(None, 256)
+        .await
+        .expect("pre-purge authoritative events");
+    adapter
+        .purge(id(1).into())
+        .await
+        .expect("checkpoint authorizes strict prefix");
+    assert_eq!(
+        logs(&mut adapter).await,
+        vec![entry(2), entry(3), entry(4)],
+        "RAFT_LOG_PARTIAL_PURGE_RETAINS_SUFFIX"
+    );
+    let state = adapter.get_log_state().await.expect("partially purged state");
+    assert_eq!(state.last_purged_log_id, Some(id(1).into()));
+    assert_eq!(state.last_log_id, Some(id(4).into()));
+    adapter.purge(id(1).into()).await.expect("repeat exact purge is idempotent");
+    assert!(adapter.purge(id(0).into()).await.is_err(), "RAFT_LOG_PURGE_NO_REGRESSION");
+    {
+        let store = writer.store.lock().expect("inspect task-owned store");
+        let tx = store.database.begin_read().expect("real persisted prefix inspection");
+        let results = tx.open_table(application::APPLY_RESULTS).expect("apply receipts");
+        assert!(results.get(0).expect("membership receipt").is_none());
+        assert!(results.get(1).expect("first receipt").is_none());
+        assert!(
+            results.get(2).expect("retained receipt").is_some(),
+            "RAFT_LOG_PARTIAL_PURGE_RECEIPT_SUFFIX"
+        );
+    }
+    drop(adapter);
+    drop(writer);
+    let writer = NativeLogWriter::new(Store::open(directory.path().join("db"), identity()).expect("restart partial purge"));
+    let mut adapter = NativeRaftLogStore::new(writer.clone());
+    assert_eq!(
+        logs(&mut adapter).await,
+        vec![entry(2), entry(3), entry(4)],
+        "RAFT_LOG_PARTIAL_PURGE_REOPEN"
+    );
+    adapter
+        .blocking_append(vec![entry(5)])
+        .await
+        .expect("continue append after partial purge");
+    writer
+        .apply_committed(vec![entry(3), entry(4), entry(5)])
+        .await
+        .expect("continue application from true applied state");
+    assert_eq!(
+        writer
+            .created_events_page(None, 256)
+            .await
+            .expect("retained plus new events")
+            .events
+            .len(),
+        5,
+        "RAFT_LOG_PURGE_NEVER_DELETES_EVENTS"
+    );
+    writer
+        .purge_checkpoint(&receipt)
+        .await
+        .expect("existing exact receipt API remains compatible after partial purge");
+    assert_eq!(
+        adapter
+            .get_log_state()
+            .await
+            .expect("exact compatibility state")
+            .last_purged_log_id,
+        Some(id(2).into())
+    );
+    assert_eq!(logs(&mut adapter).await, vec![entry(3), entry(4), entry(5)]);
+    assert_eq!(
+        writer
+            .created_events_page(None, 2)
+            .await
+            .expect("original events retained")
+            .events,
+        events.events
+    );
+    drop(adapter);
+    drop(writer);
+    let mut reopened = NativeRaftLogStore::new(NativeLogWriter::new(
+        Store::open(directory.path().join("db"), identity()).expect("reopen continued apply and exact purge"),
+    ));
+    assert_eq!(logs(&mut reopened).await, vec![entry(3), entry(4), entry(5)]);
+}
+
+#[tokio::test]
+async fn raft_log_storage_purge_rejects_uncovered() {
+    let (directory, writer, mut adapter) = fixture();
+    adapter
+        .blocking_append(entries(2))
+        .await
+        .expect("trait append without snapshot authority");
+    writer
+        .apply_committed(entries(2))
+        .await
+        .expect("applied preparation without checkpoint");
+    let before = std::fs::read(directory.path().join("db")).expect("stable pre-rejection bytes");
+    assert!(
+        adapter.purge(id(1).into()).await.is_err(),
+        "PURGE_REQUIRES_CHECKPOINT_COVERAGE: application progress is not a durable checkpoint"
+    );
+    assert_eq!(
+        std::fs::read(directory.path().join("db")).expect("rejected purge bytes"),
+        before,
+        "RAFT_LOG_UNAUTHORIZED_PURGE_NO_MUTATION"
+    );
+    writer
+        .checkpoint_local(CheckpointLimits::default())
+        .await
+        .expect("checkpoint at applied cut 2");
+    adapter
+        .blocking_append(vec![entry(3)])
+        .await
+        .expect("log beyond checkpoint cut");
+    let before = std::fs::read(directory.path().join("db")).expect("post-checkpoint bytes");
+    assert!(
+        adapter.purge(id(3).into()).await.is_err(),
+        "PURGE_REQUIRES_CHECKPOINT_COVERAGE: retained target exceeds durable cut"
+    );
+    assert_eq!(
+        std::fs::read(directory.path().join("db")).expect("uncovered target rejection bytes"),
+        before,
+        "RAFT_LOG_UNCOVERED_PURGE_NO_MUTATION"
+    );
+    assert_eq!(logs(&mut adapter).await, entries(3));
+    assert_eq!(adapter.get_log_state().await.expect("unchanged marker").last_purged_log_id, None);
+}
+
+#[tokio::test]
+async fn raft_log_storage_purge_identity() {
+    let (directory, writer, mut adapter) = fixture();
+    checkpoint(&writer, &mut adapter).await;
+    for target in [
+        LogId { term: 9, ..id(1) },
+        LogId { leader_node: 1, ..id(1) },
+        LogId { term: 9, ..id(2) },
+    ] {
+        let before = std::fs::read(directory.path().join("db")).expect("before wrong complete LogId");
+        assert!(adapter.purge(target.into()).await.is_err(), "RAFT_LOG_PURGE_EXACT_LOG_ID");
+        assert_eq!(std::fs::read(directory.path().join("db")).expect("after identity rejection"), before);
+    }
+    adapter.purge(id(1).into()).await.expect("valid identity");
+    let mut wrong = id(1);
+    wrong.term += 1;
+    assert!(adapter.purge(wrong.into()).await.is_err(), "RAFT_LOG_PURGED_REPEAT_IDENTITY_BOUND");
+    assert_eq!(
+        adapter
+            .get_log_state()
+            .await
+            .expect("unchanged identity marker")
+            .last_purged_log_id,
+        Some(id(1).into())
+    );
+    adapter.purge(id(2).into()).await.expect("final exact covered target");
+    let state = adapter.get_log_state().await.expect("no remaining logs");
+    assert_eq!(state.last_purged_log_id, Some(id(2).into()));
+    assert_eq!(state.last_log_id, Some(id(2).into()), "RAFT_LOG_EMPTY_TAIL_USES_PURGED_LOG_ID");
+    assert!(logs(&mut adapter).await.is_empty());
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn raft_log_storage_cancelled_append() {
+    let (directory, backend, writer, adapter) = media_fixture();
+    let (entered, resume) = backend.pause_sync(1, false);
+    let mut first = adapter.clone();
+    let task = tokio::spawn(async move { first.blocking_append(entries(2)).await });
+    entered.wait(); // Actual redb commit has begun but its first sync is paused.
+    assert!(!task.is_finished(), "RAFT_LOG_FLUSH_NOT_BEFORE_DURABILITY");
+    task.abort();
+    assert!(task.await.expect_err("cancelled blocking append waiter").is_cancelled());
+    assert!(writer.ordering.try_lock().is_err(), "RAFT_LOG_CANCELLED_APPEND_RETAINS_ORDERING_OWNER");
+    let mut following = adapter.clone();
+    let next = tokio::spawn(async move { following.blocking_append(vec![entry(3)]).await });
+    assert!(writer.ordering.try_lock().is_err(), "RAFT_LOG_NEXT_APPEND_CANNOT_BYPASS_PENDING_SYNC");
+    resume.wait();
+    next.await.expect("next trait append task").expect("ordered follow-up append");
+    let mut adapter = adapter;
+    assert_eq!(
+        logs(&mut adapter).await,
+        entries(3),
+        "RAFT_LOG_CANCELLED_WAITER_BACKGROUND_COMMIT_COMPLETE"
+    );
+    backend.cut();
+    drop(adapter);
+    drop(writer);
+    let mut reopened = NativeRaftLogStore::new(NativeLogWriter::new(
+        Store::open(directory.path().join("db"), identity()).expect("reopen cancelled waiter background commit"),
+    ));
+    assert_eq!(logs(&mut reopened).await, entries(3));
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn raft_log_storage_failed_append() {
+    let (directory, backend, writer, adapter) = media_fixture();
+    let (entered, resume) = backend.pause_sync(1, false);
+    let mut appending = adapter.clone();
+    let task = tokio::spawn(async move { appending.blocking_append(entries(2)).await });
+    entered.wait();
+    assert!(!task.is_finished(), "RAFT_LOG_FAILED_FLUSH_NOT_EARLY_SUCCESS");
+    backend.cut();
+    resume.wait();
+    assert!(
+        task.await.expect("failed actual trait append task").is_err(),
+        "RAFT_LOG_FAILED_APPEND_NEVER_REPORTS_CALLBACK_SUCCESS"
+    );
+    let mut adapter = adapter;
+    assert!(
+        adapter.blocking_append(entries(2)).await.is_err(),
+        "RAFT_LOG_INDETERMINATE_STORE_REJECTS_LATER_APPEND"
+    );
+    assert!(
+        adapter.save_vote(&openraft::Vote::new_committed(9, 0)).await.is_err(),
+        "RAFT_LOG_INDETERMINATE_STORE_REJECTS_LATER_VOTE"
+    );
+    drop(adapter);
+    drop(writer);
+    let mut reopened = NativeRaftLogStore::new(NativeLogWriter::new(
+        Store::open(directory.path().join("db"), identity()).expect("reopen failed-before-first-sync append"),
+    ));
+    assert!(logs(&mut reopened).await.is_empty(), "RAFT_LOG_FAILURE_NO_PARTIAL_BATCH");
+}
+
+#[tokio::test]
+async fn raft_log_storage_truncate_uncommitted_suffix() {
+    let (directory, writer, mut adapter) = fixture();
+    adapter.blocking_append(entries(4)).await.expect("trait append full log");
+    adapter
+        .save_committed(Some(id(2).into()))
+        .await
+        .expect("protect committed prefix");
+    assert!(
+        adapter.truncate(id(2).into()).await.is_err(),
+        "RAFT_LOG_TRUNCATE_PROTECTS_COMMITTED_PREFIX"
+    );
+    adapter
+        .truncate(id(3).into())
+        .await
+        .expect("truncate includes uncommitted target");
+    assert_eq!(logs(&mut adapter).await, entries(2), "RAFT_LOG_TRUNCATE_INCLUSIVE");
+    let mut replacement = entry(3);
+    replacement.log_id = LogId { term: 9, ..id(3) }.into();
+    adapter
+        .blocking_append(vec![replacement.clone()])
+        .await
+        .expect("trait append replacement suffix");
+    assert_eq!(
+        adapter.get_log_state().await.expect("replacement tail").last_log_id,
+        Some(replacement.log_id)
+    );
+    drop(adapter);
+    drop(writer);
+    let mut reopened = NativeRaftLogStore::new(NativeLogWriter::new(
+        Store::open(directory.path().join("db"), identity()).expect("reopen truncate and replacement"),
+    ));
+    let mut expected = entries(2);
+    expected.push(replacement);
+    assert_eq!(logs(&mut reopened).await, expected);
+}
@@ -0,0 +1,251 @@
+//! Explicit lifecycle for real OpenRaft over one durable native capture owner.
+use crate::{
+    CaptureCommand, CaptureNode, CaptureNodeError, CaptureTypeConfig, CheckpointLimits, NativeLogWriter, NativeRaftLogStore,
+    NativeRaftStateMachine, NativeSnapshotError, NativeWriteError, StoreError,
+};
+use openraft::{
+    Config, ConfigError, Raft, RaftMetrics, RaftNetworkFactory,
+    error::{ClientWriteError, Fatal, InitializeError, RaftError},
+    raft::ClientWriteResponse,
+};
+use redb::ReadableDatabase;
+use std::{
+    collections::BTreeMap,
+    sync::{
+        Arc,
+        atomic::{AtomicBool, Ordering},
+    },
+};
+use tokio::sync::watch;
+
+#[derive(Debug, thiserror::Error)]
+pub enum NativeRaftOwnerError {
+    #[error(transparent)]
+    Config(#[from] Box<ConfigError>),
+    #[error(transparent)]
+    Store(#[from] NativeWriteError),
+    #[error(transparent)]
+    Snapshot(#[from] NativeSnapshotError),
+    #[error("runtime node ID {runtime} differs from durable node ID {durable}")]
+    IdentityMismatch { runtime: u64, durable: u64 },
+    #[error("bootstrap must contain the local node")]
+    MissingLocalMember,
+    #[error(transparent)]
+    Node(#[from] CaptureNodeError),
+    #[error("native Raft owner is stopping or stopped")]
+    Stopped,
+    #[error(transparent)]
+    Fatal(Box<Fatal<u64>>),
+    #[error(transparent)]
+    Initialize(Box<RaftError<u64, InitializeError<u64, CaptureNode>>>),
+    #[error(transparent)]
+    Write(Box<RaftError<u64, ClientWriteError<u64, CaptureNode>>>),
+    #[error("native Raft shutdown task failed")]
+    Shutdown(#[source] tokio::task::JoinError),
+    #[cfg(any(target_os = "linux", target_os = "macos"))]
+    #[error(transparent)]
+    Reader(#[from] crate::source_reader::ReaderError),
+}
+
+#[derive(Default)]
+struct DrainState {
+    complete: AtomicBool,
+    changed: tokio::sync::Notify,
+    #[cfg(test)]
+    waiters: Arc<std::sync::atomic::AtomicUsize>,
+}
+impl DrainState {
+    async fn wait(&self) {
+        #[cfg(test)]
+        let _waiter = DrainWaiter::new(self.waiters.clone());
+        loop {
+            let notified = self.changed.notified();
+            tokio::pin!(notified);
+            notified.as_mut().enable();
+            if self.complete.load(Ordering::Acquire) {
+                return;
+            }
+            notified.await;
+        }
+    }
+}
+#[cfg(test)]
+struct DrainWaiter(Arc<std::sync::atomic::AtomicUsize>);
+#[cfg(test)]
+impl DrainWaiter {
+    fn new(waiters: Arc<std::sync::atomic::AtomicUsize>) -> Self {
+        waiters.fetch_add(1, Ordering::AcqRel);
+        Self(waiters)
+    }
+}
+#[cfg(test)]
+impl Drop for DrainWaiter {
+    fn drop(&mut self) {
+        self.0.fetch_sub(1, Ordering::AcqRel);
+    }
+}
+
+// The real worker, snapshot builders and replication readers share this one lease.
+// Its final drop runs after their NativeLogWriter fields have been released.
+pub(crate) struct OwnerLease {
+    drain: Arc<DrainState>,
+}
+impl Drop for OwnerLease {
+    fn drop(&mut self) {
+        self.drain.complete.store(true, Ordering::Release);
+        self.drain.changed.notify_waiters();
+    }
+}
+
+/// A lifecycle owner, not a transport. All storage adapters share its original writer.
+pub struct NativeRaftOwner {
+    raft: Raft<CaptureTypeConfig>,
+    writer: NativeLogWriter,
+    identity: crate::NodeIdentity,
+    accepting: AtomicBool,
+    drain: Arc<DrainState>,
+    #[cfg(test)]
+    pub(crate) snapshot_build_pause: Arc<crate::raft_state_machine::TestPause>,
+}
+impl NativeRaftOwner {
+    pub async fn start<N>(
+        writer: NativeLogWriter,
+        node_id: u64,
+        config: Arc<Config>,
+        limits: CheckpointLimits,
+        network: N,
+    ) -> Result<Self, NativeRaftOwnerError>
+    where
+        N: RaftNetworkFactory<CaptureTypeConfig>,
+    {
+        // All operator-input checks precede Raft::new and its background workers.
+        let config = Arc::new(
+            (*config)
+                .clone()
+                .validate()
+                .map_err(|error| NativeRaftOwnerError::Config(Box::new(error)))?,
+        );
+        let mut state_machine = NativeRaftStateMachine::new(writer.clone(), limits)?;
+        let identity = writer
+            .write(|store| {
+                if !store.writable {
+                    return Err(StoreError::WriteDisabled);
+                }
+                let transaction = store.database.begin_read()?;
+                let meta = transaction.open_table(crate::META)?;
+                crate::application::native_application_state(&transaction)?;
+                crate::application::read_binding(&meta)?;
+                crate::decode_identity(meta.get("identity")?.ok_or(StoreError::CorruptRecord)?.value())
+            })
+            .await?;
+        if identity.node_id != node_id {
+            return Err(NativeRaftOwnerError::IdentityMismatch {
+                runtime: node_id,
+                durable: identity.node_id,
+            });
+        }
+        let drain = Arc::new(DrainState::default());
+        let lease = Arc::new(OwnerLease { drain: drain.clone() });
+        state_machine.owner_lease = Some(lease.clone());
+        let mut logs = NativeRaftLogStore::new(writer.clone());
+        logs.owner_lease = Some(lease);
+        #[cfg(test)]
+        let snapshot_build_pause = state_machine.build_before_ordering.clone();
+        let raft = Raft::new(node_id, config, network, logs, state_machine)
+            .await
+            .map_err(|error| NativeRaftOwnerError::Fatal(Box::new(error)))?;
+        Ok(Self {
+            raft,
+            writer,
+            identity,
+            accepting: AtomicBool::new(true),
+            drain,
+            #[cfg(test)]
+            snapshot_build_pause,
+        })
+    }
+
+    /// Identity read from durable META and checked before starting this owner.
+    pub fn identity(&self) -> crate::NodeIdentity {
+        self.identity
+    }
+
+    fn require_running(&self) -> Result<(), NativeRaftOwnerError> {
+        if self.accepting.load(Ordering::Acquire) {
+            Ok(())
+        } else {
+            Err(NativeRaftOwnerError::Stopped)
+        }
+    }
+
+    /// Only explicit first-cluster formation calls initialize; start never does.
+    pub async fn bootstrap(&self, members: BTreeMap<u64, CaptureNode>) -> Result<(), NativeRaftOwnerError> {
+        self.require_running()?;
+        if !members.contains_key(&self.identity.node_id) {
+            return Err(NativeRaftOwnerError::MissingLocalMember);
+        }
+        for node in members.values() {
+            node.validate()?;
+        }
+        self.raft
+            .initialize(members)
+            .await
+            .map_err(|error| NativeRaftOwnerError::Initialize(Box::new(error)))
+    }
+
+    /// The returned LogId and application response come from actual consensus apply.
+    pub async fn client_write(
+        &self,
+        command: CaptureCommand,
+    ) -> Result<ClientWriteResponse<CaptureTypeConfig>, NativeRaftOwnerError> {
+        self.require_running()?;
+        self.raft
+            .client_write(command)
+            .await
+            .map_err(|error| NativeRaftOwnerError::Write(Box::new(error)))
+    }
+
+    /// Re-read an exact successful first decision from this owner's durable view.
+    /// Use the Created result's decision LogId, not an idempotent retry entry's outer LogId.
+    pub async fn authorize_created(
+        &self,
+        expected: crate::DecideCreated,
+        expected_decision: crate::LogId,
+    ) -> Result<crate::CommittedCreatedAuthorization, NativeRaftOwnerError> {
+        self.require_running()?;
+        Ok(self.writer.authorize_created(expected, expected_decision).await?)
+    }
+
+    pub fn metrics(&self) -> watch::Receiver<RaftMetrics<u64, CaptureNode>> {
+        self.raft.metrics()
+    }
+
+    /// Transport registration only; raw clones remain trusted caller-owned handles.
+    pub fn raft(&self) -> &Raft<CaptureTypeConfig> {
+        &self.raft
+    }
+
+    /// The caller owns serve/shutdown and must drain it before releasing the Store.
+    #[cfg(any(target_os = "linux", target_os = "macos"))]
+    pub fn bind_reader(
+        &self,
+        path: impl AsRef<std::path::Path>,
+    ) -> Result<(crate::source_reader::SourceReaderServer, crate::source_reader::Shutdown), NativeRaftOwnerError> {
+        self.require_running()?;
+        Ok(crate::source_reader::SourceReaderServer::bind(path, self.writer.clone())?)
+    }
+
+    #[cfg(test)]
+    pub(crate) fn shutdown_waiters(&self) -> Arc<std::sync::atomic::AtomicUsize> {
+        self.drain.waiters.clone()
+    }
+
+    /// Stop admission before signaling Raft; await its real core/tick shutdown.
+    pub async fn shutdown(&self) -> Result<(), NativeRaftOwnerError> {
+        self.accepting.store(false, Ordering::Release);
+        self.raft.shutdown().await.map_err(NativeRaftOwnerError::Shutdown)?;
+        // OpenRaft joins core/tick but detaches the SM worker and snapshot builders.
+        self.drain.wait().await;
+        Ok(())
+    }
+}
@@ -0,0 +1,885 @@
+//! Real Raft engines and real redb; direct RPC routing exists only inside these tests.
+use super::*;
+use openraft::{
+    Raft, RaftNetwork, RaftNetworkFactory, ServerState,
+    error::{InstallSnapshotError, RPCError, RaftError, RemoteError, Unreachable},
+    network::{Backoff, RPCOption},
+    raft::{
+        AppendEntriesRequest, AppendEntriesResponse, InstallSnapshotRequest, InstallSnapshotResponse, VoteRequest, VoteResponse,
+    },
+};
+use std::{
+    collections::{BTreeMap, BTreeSet},
+    sync::{
+        Arc, Mutex,
+        atomic::{AtomicUsize, Ordering},
+    },
+    time::Duration,
+};
+
+type Engine = Raft<CaptureTypeConfig>;
+type RpcError = RPCError<u64, CaptureNode, RaftError<u64>>;
+type SnapshotRpcError = RPCError<u64, CaptureNode, RaftError<u64, InstallSnapshotError>>;
+type PausedLinks = Mutex<BTreeMap<(u64, u64), (Arc<tokio::sync::Notify>, Arc<tokio::sync::Notify>)>>;
+#[derive(Default)]
+struct Routes {
+    nodes: Mutex<BTreeMap<u64, Engine>>,
+    blocked: Mutex<BTreeSet<(u64, u64)>>,
+    snapshot_installs: AtomicUsize,
+    pause_append: PausedLinks,
+}
+impl Routes {
+    fn register(&self, id: u64, raft: Engine) {
+        self.nodes.lock().expect("routes").insert(id, raft);
+    }
+    fn remove(&self, id: u64) {
+        self.nodes.lock().expect("routes").remove(&id);
+    }
+    fn isolate(&self, id: u64, members: &[u64]) {
+        let mut blocked = self.blocked.lock().expect("partition routes");
+        for peer in members {
+            if *peer != id {
+                blocked.insert((id, *peer));
+                blocked.insert((*peer, id));
+            }
+        }
+    }
+    fn heal(&self) {
+        self.blocked.lock().expect("restore routes").clear();
+    }
+    fn endpoint(&self, source: u64, target: u64) -> Result<Engine, Unreachable> {
+        if self.blocked.lock().expect("partition check").contains(&(source, target)) {
+            return Err(Unreachable::new(&std::io::Error::other("test partition")));
+        }
+        self.nodes
+            .lock()
+            .expect("endpoint registry")
+            .get(&target)
+            .cloned()
+            .ok_or_else(|| Unreachable::new(&std::io::Error::other("test endpoint unavailable")))
+    }
+}
+#[derive(Clone)]
+struct DirectFactory {
+    source: u64,
+    routes: Arc<Routes>,
+}
+struct DirectClient {
+    source: u64,
+    target: u64,
+    node: CaptureNode,
+    routes: Arc<Routes>,
+}
+impl RaftNetworkFactory<CaptureTypeConfig> for DirectFactory {
+    type Network = DirectClient;
+    async fn new_client(&mut self, target: u64, node: &CaptureNode) -> DirectClient {
+        DirectClient {
+            source: self.source,
+            target,
+            node: node.clone(),
+            routes: self.routes.clone(),
+        }
+    }
+}
+impl RaftNetwork<CaptureTypeConfig> for DirectClient {
+    async fn append_entries(
+        &mut self,
+        request: AppendEntriesRequest<CaptureTypeConfig>,
+        _option: RPCOption,
+    ) -> Result<AppendEntriesResponse<u64>, RpcError> {
+        let pause = self
+            .routes
+            .pause_append
+            .lock()
+            .expect("actual RPC pause hook")
+            .remove(&(self.source, self.target));
+        if let Some((entered, resume)) = pause {
+            entered.notify_one();
+            resume.notified().await;
+        }
+        let raft = self.routes.endpoint(self.source, self.target)?;
+        raft.append_entries(request)
+            .await
+            .map_err(|error| RemoteError::new_with_node(self.target, self.node.clone(), error).into())
+    }
+    async fn vote(&mut self, request: VoteRequest<u64>, _option: RPCOption) -> Result<VoteResponse<u64>, RpcError> {
+        let raft = self.routes.endpoint(self.source, self.target)?;
+        raft.vote(request)
+            .await
+            .map_err(|error| RemoteError::new_with_node(self.target, self.node.clone(), error).into())
+    }
+    async fn install_snapshot(
+        &mut self,
+        request: InstallSnapshotRequest<CaptureTypeConfig>,
+        _option: RPCOption,
+    ) -> Result<InstallSnapshotResponse<u64>, SnapshotRpcError> {
+        let raft = self.routes.endpoint(self.source, self.target)?;
+        let done = request.done;
+        let result = raft
+            .install_snapshot(request)
+            .await
+            .map_err(|error| RemoteError::new_with_node(self.target, self.node.clone(), error).into());
+        if done && result.is_ok() {
+            self.routes.snapshot_installs.fetch_add(1, Ordering::AcqRel);
+        }
+        result
+    }
+    fn backoff(&self) -> Backoff {
+        Backoff::new(std::iter::repeat(Duration::from_millis(25)))
+    }
+}
+fn identity(node_id: u64) -> NodeIdentity {
+    NodeIdentity {
+        cluster_id: [7; 16],
+        node_id,
+    }
+}
+fn binding() -> CaptureBinding {
+    CaptureBinding {
+        revision: 9,
+        binding_id: [8; 16],
+        target_id: [9; 16],
+    }
+}
+fn command(value: u8) -> CaptureCommand {
+    CaptureCommand::try_created(DecideCreated {
+        operation_id: [value; 16],
+        object: ObjectIdentity {
+            bucket_incarnation: [2; 16],
+            key: vec![value],
+        },
+        expected_head: None,
+        prepared: PreparedIdentity {
+            version_id: [value; 16],
+            preparation_id: [4; 16],
+            content_digest: [5; 32],
+            content_length: 123,
+        },
+        binding_revision: 9,
+    })
+    .expect("native client command")
+}
+fn config(snapshot_every: u64) -> Arc<openraft::Config> {
+    Arc::new(
+        openraft::Config {
+            cluster_name: "native-owner-tests".into(),
+            heartbeat_interval: 100,
+            election_timeout_min: 500,
+            election_timeout_max: 1000,
+            snapshot_policy: openraft::SnapshotPolicy::LogsSinceLast(snapshot_every),
+            max_in_snapshot_log_to_keep: 0,
+            purge_batch_size: 1,
+            ..openraft::Config::default()
+        }
+        .validate()
+        .expect("valid real Raft test configuration"),
+    )
+}
+fn members(ids: &[u64]) -> BTreeMap<u64, CaptureNode> {
+    ids.iter()
+        .map(|id| {
+            (
+                *id,
+                CaptureNode::new(format!("http://127.0.0.1:{}", 7000 + id)).expect("bounded test endpoint"),
+            )
+        })
+        .collect()
+}
+struct Node {
+    directory: tempfile::TempDir,
+    writer: NativeLogWriter,
+    owner: Option<NativeRaftOwner>,
+    id: u64,
+}
+impl Node {
+    async fn start(id: u64, routes: Arc<Routes>, config: Arc<openraft::Config>) -> Self {
+        let directory = tempfile::tempdir().expect("real Raft owner directory");
+        let mut store = Store::initialize(directory.path().join("db"), identity(id)).expect("real owner redb");
+        store.pin_capture_binding(binding()).expect("captured binding");
+        store.enable_native_empty().expect("native storage mode");
+        store.promote_created_provenance().expect("native provenance");
+        let writer = NativeLogWriter::new(store);
+        let owner = NativeRaftOwner::start(
+            writer.clone(),
+            id,
+            config,
+            CheckpointLimits::default(),
+            DirectFactory {
+                source: id,
+                routes: routes.clone(),
+            },
+        )
+        .await
+        .expect("actual owner start");
+        routes.register(id, owner.raft().clone());
+        Self {
+            directory,
+            writer,
+            owner: Some(owner),
+            id,
+        }
+    }
+    fn owner(&self) -> &NativeRaftOwner {
+        self.owner.as_ref().expect("live owner")
+    }
+    async fn page(&self) -> EventPage {
+        self.writer
+            .created_events_page(None, 256)
+            .await
+            .expect("actual committed events")
+    }
+    async fn stop(&mut self, routes: &Routes) {
+        self.owner
+            .take()
+            .expect("running owner")
+            .shutdown()
+            .await
+            .expect("real Raft shutdown");
+        routes.remove(self.id);
+    }
+}
+struct Cluster {
+    nodes: Vec<Node>,
+    routes: Arc<Routes>,
+}
+impl Cluster {
+    async fn start(ids: &[u64], snapshot_every: u64) -> Self {
+        let routes = Arc::new(Routes::default());
+        let mut nodes = Vec::new();
+        for id in ids {
+            nodes.push(Node::start(*id, routes.clone(), config(snapshot_every)).await);
+        }
+        nodes[0]
+            .owner()
+            .bootstrap(members(ids))
+            .await
+            .expect("explicit initial cluster bootstrap");
+        let cluster = Self { nodes, routes };
+        cluster.leader(ids).await;
+        cluster
+    }
+    fn node(&self, id: u64) -> &Node {
+        self.nodes.iter().find(|node| node.id == id).expect("cluster node")
+    }
+    async fn leader(&self, eligible: &[u64]) -> u64 {
+        tokio::time::timeout(Duration::from_secs(15), async {
+            loop {
+                for id in eligible {
+                    let node = self.node(*id);
+                    let Some(owner) = &node.owner else {
+                        continue;
+                    };
+                    let metrics = owner.metrics().borrow().clone();
+                    assert!(metrics.running_state.is_ok(), "real Raft owner fatal state: {:?}", metrics.running_state);
+                    if metrics.state == ServerState::Leader && metrics.current_leader == Some(*id) {
+                        return *id;
+                    }
+                }
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("real leader election readiness")
+    }
+    async fn applied(&self, ids: &[u64], index: u64) {
+        for id in ids {
+            self.node(*id)
+                .owner()
+                .raft()
+                .wait(Some(Duration::from_secs(15)))
+                .applied_index_at_least(Some(index), "real replicated apply readiness")
+                .await
+                .expect("committed log applied on real node");
+        }
+    }
+    async fn stop(mut self) {
+        for node in &mut self.nodes {
+            if node.owner.is_some() {
+                node.stop(&self.routes).await;
+            }
+        }
+    }
+}
+
+#[tokio::test]
+async fn raft_owner_single_node_bootstrap() {
+    let cluster = Cluster::start(&[0], 100).await;
+    let response = cluster
+        .node(0)
+        .owner()
+        .client_write(command(1))
+        .await
+        .expect("real client_write");
+    assert!(
+        matches!(response.data.created_result(),Some(CreatedResult::Created { event_id,.. }) if *event_id==[1;16]),
+        "OWNER_TRUE_RAFT_CREATED_RESPONSE"
+    );
+    let page = cluster.node(0).page().await;
+    assert_eq!(page.events.len(), 1);
+    assert_eq!(
+        page.events[0].captured.decision_log_id,
+        LogId::from(response.log_id),
+        "OWNER_COMMITTED_EVENT_LOG_ID"
+    );
+    assert!(
+        cluster.node(0).owner().bootstrap(members(&[0])).await.is_err(),
+        "OWNER_BOOTSTRAP_NOT_REINITIALIZATION"
+    );
+    cluster.stop().await;
+}
+
+#[tokio::test]
+async fn raft_owner_single_node_restart() {
+    let routes = Arc::new(Routes::default());
+    let mut node = Node::start(0, routes.clone(), config(100)).await;
+    node.owner().bootstrap(members(&[0])).await.expect("first bootstrap only");
+    node.owner()
+        .raft()
+        .wait(Some(Duration::from_secs(15)))
+        .state(ServerState::Leader, "single leader")
+        .await
+        .expect("leader ready");
+    node.owner().client_write(command(1)).await.expect("first genuine write");
+    let before = node.page().await;
+    node.stop(&routes).await;
+    let Node {
+        directory,
+        writer,
+        owner: _,
+        id,
+    } = node;
+    let vote = writer.read_vote().await.expect("durable vote before reopen");
+    drop(writer);
+    let store = Store::open(directory.path().join("db"), identity(id)).expect("restart same durable identity");
+    assert_eq!(store.read_vote().expect("persisted vote"), vote.map(Into::into));
+    let writer = NativeLogWriter::new(store);
+    let owner = NativeRaftOwner::start(
+        writer.clone(),
+        id,
+        config(100),
+        CheckpointLimits::default(),
+        DirectFactory {
+            source: id,
+            routes: routes.clone(),
+        },
+    )
+    .await
+    .expect("restart existing Raft");
+    routes.register(id, owner.raft().clone());
+    assert!(
+        owner.raft().is_initialized().await.expect("restored cluster membership"),
+        "OWNER_RESTART_DOES_NOT_REQUIRE_BOOTSTRAP"
+    );
+    owner
+        .raft()
+        .wait(Some(Duration::from_secs(15)))
+        .state(ServerState::Leader, "restored leader")
+        .await
+        .expect("restarted leader ready");
+    owner
+        .client_write(command(2))
+        .await
+        .expect("continue after restart without initialize");
+    let after = writer.created_events_page(None, 256).await.expect("restart continued events");
+    assert_eq!(after.events.len(), 2);
+    assert_eq!(after.events[0], before.events[0], "OWNER_RESTART_PRESERVES_FIRST_DECISION");
+    owner.shutdown().await.expect("restart owner shutdown");
+    routes.remove(id);
+}
+
+#[tokio::test]
+async fn raft_owner_idempotent_retry() {
+    let cluster = Cluster::start(&[0], 100).await;
+    let owner = cluster.node(0).owner();
+    let first = owner.client_write(command(1)).await.expect("first write");
+    let replay = owner.client_write(command(1)).await.expect("same operation retry");
+    assert_eq!(
+        replay.data.created_result(),
+        first.data.created_result(),
+        "OWNER_RETRY_PRESERVES_FIRST_DECISION_RESULT"
+    );
+    assert!(replay.log_id.index > first.log_id.index, "real retry can append another Raft log");
+    assert_eq!(cluster.node(0).page().await.events.len(), 1, "OWNER_RETRY_NO_DUPLICATE_EVENT");
+    cluster.stop().await;
+}
+
+#[tokio::test]
+async fn raft_owner_identity_binding() {
+    let routes = Arc::new(Routes::default());
+    let mut node = Node::start(0, routes.clone(), config(100)).await;
+    node.owner().bootstrap(members(&[0])).await.expect("real durable membership");
+    node.owner()
+        .raft()
+        .wait(Some(Duration::from_secs(15)))
+        .state(ServerState::Leader, "single leader")
+        .await
+        .expect("leader");
+    node.owner().client_write(command(1)).await.expect("nonempty durable store");
+    node.stop(&routes).await;
+    let before = std::fs::read(node.directory.path().join("db")).expect("stable identity-bound store");
+    let result = NativeRaftOwner::start(
+        node.writer.clone(),
+        1,
+        config(100),
+        CheckpointLimits::default(),
+        DirectFactory {
+            source: 1,
+            routes: routes.clone(),
+        },
+    )
+    .await;
+    assert!(
+        matches!(result, Err(NativeRaftOwnerError::IdentityMismatch { runtime: 1, durable: 0 })),
+        "OWNER_REQUIRES_DURABLE_NODE_ID: runtime node cannot reinterpret stored membership/identity"
+    );
+    assert_eq!(
+        std::fs::read(node.directory.path().join("db")).expect("unchanged identity-bound store"),
+        before
+    );
+    let invalid_config = Arc::new(openraft::Config {
+        election_timeout_min: 1000,
+        election_timeout_max: 500,
+        ..openraft::Config::default()
+    });
+    let invalid = NativeRaftOwner::start(
+        node.writer.clone(),
+        0,
+        invalid_config,
+        CheckpointLimits::default(),
+        DirectFactory {
+            source: 0,
+            routes: routes.clone(),
+        },
+    )
+    .await;
+    assert!(
+        matches!(invalid, Err(NativeRaftOwnerError::Config(_))),
+        "OWNER_REJECTS_INVALID_CONFIG_BEFORE_START"
+    );
+    let mut limits = CheckpointLimits::default();
+    limits.snapshot.total_bytes = 0;
+    let invalid = NativeRaftOwner::start(node.writer.clone(), 0, config(100), limits, DirectFactory { source: 0, routes }).await;
+    assert!(
+        matches!(invalid, Err(NativeRaftOwnerError::Snapshot(_))),
+        "OWNER_REJECTS_INVALID_LIMITS_BEFORE_START"
+    );
+    assert_eq!(
+        std::fs::read(node.directory.path().join("db")).expect("invalid start leaves stable bytes"),
+        before
+    );
+}
+
+#[cfg(any(target_os = "linux", target_os = "macos"))]
+#[tokio::test]
+async fn raft_owner_shared_reader() {
+    use source_reader::SourceReaderClient;
+    use std::os::unix::fs::PermissionsExt;
+    let cluster = Cluster::start(&[0], 100).await;
+    let node = cluster.node(0);
+    std::fs::set_permissions(node.directory.path(), std::fs::Permissions::from_mode(0o700)).expect("private reader parent");
+    let socket = node.directory.path().join("reader.sock");
+    let (server, shutdown) = node
+        .owner()
+        .bind_reader(&socket)
+        .expect("owner shares live writer with reader");
+    let task = tokio::spawn(server.serve());
+    let client = SourceReaderClient::new(&socket).expect("live reader");
+    assert!(
+        client
+            .page(identity(0).cluster_id, None, 1)
+            .await
+            .expect("initial reader page")
+            .page
+            .events
+            .is_empty()
+    );
+    let response = node
+        .owner()
+        .client_write(command(1))
+        .await
+        .expect("real write while reader runs");
+    let page = client
+        .page(identity(0).cluster_id, None, 1)
+        .await
+        .expect("live committed event");
+    assert_eq!(page.identity, identity(0));
+    assert_eq!(
+        page.page.events[0].captured.decision_log_id,
+        LogId::from(response.log_id),
+        "OWNER_READER_USES_REAL_COMMITTED_WRITER"
+    );
+    shutdown.shutdown();
+    task.await.expect("reader serve task").expect("reader drains");
+    cluster.stop().await;
+}
+
+#[tokio::test]
+async fn raft_owner_shutdown() {
+    let cluster = Cluster::start(&[0], 100).await;
+    let owner = cluster.node(0).owner();
+    owner.client_write(command(1)).await.expect("accepted before shutdown");
+    owner.shutdown().await.expect("actual owner shutdown");
+    assert!(owner.client_write(command(2)).await.is_err(), "OWNER_SHUTDOWN_REJECTS_NEW_WRITES");
+    assert_eq!(cluster.node(0).page().await.events.len(), 1);
+    cluster.stop().await;
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn raft_owner_three_node_majority() {
+    let cluster = Cluster::start(&[0, 1, 2], 100).await;
+    let leader = cluster.leader(&[0, 1, 2]).await;
+    let first = cluster
+        .node(leader)
+        .owner()
+        .client_write(command(1))
+        .await
+        .expect("real quorum write");
+    cluster.applied(&[0, 1, 2], first.log_id.index).await;
+    let follower = [0, 1, 2].into_iter().find(|id| *id != leader).expect("nonleader");
+    let rejected = cluster.node(follower).owner().client_write(command(254)).await;
+    match rejected {
+        Err(NativeRaftOwnerError::Write(error)) => {
+            assert!(
+                matches!(error.api_error(), Some(openraft::error::ClientWriteError::ForwardToLeader(_))),
+                "OWNER_PRESERVES_TYPED_FORWARD_TO_LEADER"
+            );
+        }
+        other => panic!("OWNER_NONLEADER_WRITE_MUST_NOT_SUCCEED: {other:?}"),
+    }
+    cluster.routes.isolate(leader, &[0, 1, 2]);
+    let uncertain =
+        tokio::time::timeout(Duration::from_millis(1500), cluster.node(leader).owner().client_write(command(2))).await;
+    assert!(!matches!(uncertain, Ok(Ok(_))), "OWNER_MAJORITY_REQUIRED_FOR_SUCCESS");
+    let others: Vec<_> = [0, 1, 2].into_iter().filter(|id| *id != leader).collect();
+    let next = cluster.leader(&others).await;
+    let response = cluster
+        .node(next)
+        .owner()
+        .client_write(command(2))
+        .await
+        .expect("majority-owned retry of exact same operation");
+    cluster.routes.heal();
+    cluster.applied(&[0, 1, 2], response.log_id.index).await;
+    for node in &cluster.nodes {
+        assert_eq!(node.page().await.events.len(), 2, "OWNER_QUORUM_RETRY_NO_DUPLICATE_EVENT");
+    }
+    cluster.stop().await;
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn raft_owner_leader_failover() {
+    let mut cluster = Cluster::start(&[0, 1, 2], 100).await;
+    let leader = cluster.leader(&[0, 1, 2]).await;
+    let first = cluster
+        .node(leader)
+        .owner()
+        .client_write(command(1))
+        .await
+        .expect("first confirmed write");
+    cluster.applied(&[0, 1, 2], first.log_id.index).await;
+    let before = cluster.node(leader).page().await.events[0].clone();
+    cluster
+        .nodes
+        .iter_mut()
+        .find(|node| node.id == leader)
+        .expect("old leader")
+        .stop(&cluster.routes)
+        .await;
+    let survivors: Vec<_> = [0, 1, 2].into_iter().filter(|id| *id != leader).collect();
+    let next = cluster.leader(&survivors).await;
+    let response = cluster
+        .node(next)
+        .owner()
+        .client_write(command(2))
+        .await
+        .expect("new leader actual quorum write");
+    cluster.applied(&survivors, response.log_id.index).await;
+    for id in survivors {
+        let page = cluster.node(id).page().await;
+        assert_eq!(page.events.len(), 2);
+        assert_eq!(page.events[0], before, "OWNER_FAILOVER_PRESERVES_CONFIRMED_DECISION");
+    }
+    cluster.stop().await;
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn raft_owner_snapshot_catchup() {
+    let cluster = Cluster::start(&[0, 1, 2], 1000).await;
+    let leader = cluster.leader(&[0, 1, 2]).await;
+    let follower = [0, 1, 2].into_iter().find(|id| *id != leader).expect("lagging follower");
+    let first = cluster
+        .node(leader)
+        .owner()
+        .client_write(command(1))
+        .await
+        .expect("initial confirmed event");
+    cluster.applied(&[0, 1, 2], first.log_id.index).await;
+    // Keep the catch-up test focused on replication rather than an isolated follower's election.
+    cluster.node(follower).owner().raft().runtime_config().elect(false);
+    cluster.routes.isolate(follower, &[0, 1, 2]);
+    let mut last = first;
+    for value in 2..=12 {
+        last = cluster
+            .node(leader)
+            .owner()
+            .client_write(command(value))
+            .await
+            .expect("majority writes while follower offline");
+    }
+    cluster
+        .node(leader)
+        .owner()
+        .raft()
+        .trigger()
+        .snapshot()
+        .await
+        .expect("true Raft snapshot trigger");
+    cluster
+        .node(leader)
+        .owner()
+        .raft()
+        .wait(Some(Duration::from_secs(15)))
+        .metrics(
+            |metrics| metrics.purged.is_some_and(|cut| cut.index >= last.log_id.index),
+            "snapshot durably purged lagging log prefix",
+        )
+        .await
+        .expect("real snapshot and purge");
+    let before = cluster.routes.snapshot_installs.load(Ordering::Acquire);
+    cluster.routes.heal();
+    cluster.applied(&[0, 1, 2], last.log_id.index).await;
+    cluster
+        .node(follower)
+        .owner()
+        .raft()
+        .wait(Some(Duration::from_secs(15)))
+        .metrics(
+            |metrics| metrics.snapshot.is_some_and(|cut| cut.index >= last.log_id.index),
+            "follower installed actual transferred snapshot",
+        )
+        .await
+        .expect("real snapshot catch-up");
+    tokio::time::timeout(Duration::from_secs(15), async {
+        while cluster.routes.snapshot_installs.load(Ordering::Acquire) <= before {
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .expect("snapshot readiness includes actual successful done-RPC return");
+    assert!(
+        cluster.routes.snapshot_installs.load(Ordering::Acquire) > before,
+        "OWNER_CATCHUP_USED_ACTUAL_SNAPSHOT_RPC"
+    );
+    let expected = cluster.node(leader).page().await;
+    for node in &cluster.nodes {
+        assert_eq!(node.page().await, expected, "OWNER_SNAPSHOT_CATCHUP_FULL_SOURCE_EVENTS");
+    }
+    cluster.node(follower).owner().raft().runtime_config().elect(true);
+    let current = cluster.leader(&[0, 1, 2]).await;
+    let next = cluster
+        .node(current)
+        .owner()
+        .client_write(command(13))
+        .await
+        .expect("continue after real snapshot install");
+    cluster.applied(&[0, 1, 2], next.log_id.index).await;
+    for node in &cluster.nodes {
+        assert_eq!(node.page().await.events.len(), 13);
+    }
+    cluster.stop().await;
+}
+
+async fn wait_for_shutdown_drain(waiters: &AtomicUsize) {
+    tokio::time::timeout(Duration::from_secs(5), async {
+        while waiters.load(Ordering::Acquire) == 0 {
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .expect("OWNER_SHUTDOWN_DRAINS_INFLIGHT_WORK: owner reached real worker drain wait");
+}
+
+#[tokio::test]
+async fn raft_owner_shutdown_inflight_snapshot() {
+    let routes = Arc::new(Routes::default());
+    let mut node = Node::start(0, routes.clone(), config(1000)).await;
+    node.owner()
+        .bootstrap(members(&[0]))
+        .await
+        .expect("real single-node bootstrap");
+    node.owner()
+        .raft()
+        .wait(Some(Duration::from_secs(15)))
+        .state(ServerState::Leader, "leader before snapshot")
+        .await
+        .expect("leader readiness");
+    node.owner()
+        .client_write(command(1))
+        .await
+        .expect("real committed snapshot input");
+    let entered = Arc::new(tokio::sync::Notify::new());
+    let resume = Arc::new(tokio::sync::Notify::new());
+    *node.owner().snapshot_build_pause.lock().expect("real builder hook") = Some((entered.clone(), resume.clone()));
+    node.owner()
+        .raft()
+        .trigger()
+        .snapshot()
+        .await
+        .expect("real Raft snapshot request");
+    tokio::time::timeout(Duration::from_secs(5), entered.notified())
+        .await
+        .expect("real detached snapshot builder paused");
+    let owner = node.owner.take().expect("running owner");
+    let waiters = owner.shutdown_waiters();
+    let observer = owner.raft().clone();
+    let shutdown = tokio::spawn(async move {
+        let result = owner.shutdown().await;
+        (owner, result)
+    });
+    wait_for_shutdown_drain(&waiters).await;
+    assert!(
+        observer.metrics().borrow().running_state.is_err(),
+        "OWNER_CORE_STOPPED_WHILE_SNAPSHOT_DRAIN_PENDING"
+    );
+    assert!(!shutdown.is_finished(), "OWNER_SHUTDOWN_DRAINS_INFLIGHT_SNAPSHOT");
+    resume.notify_one();
+    let (owner, result) = tokio::time::timeout(Duration::from_secs(10), shutdown)
+        .await
+        .expect("released snapshot lets shutdown finish")
+        .expect("shutdown task");
+    result.expect("snapshot work drained");
+    routes.remove(0);
+    drop(observer);
+    drop(owner);
+    drop(waiters);
+    let Node {
+        directory,
+        writer,
+        owner: _,
+        id,
+    } = node;
+    drop(writer);
+    let store = Store::open(directory.path().join("db"), identity(id)).expect("OWNER_SHUTDOWN_SNAPSHOT_RELEASES_LIVE_DATABASE");
+    assert_eq!(
+        store
+            .created_events_page(None, 256)
+            .expect("retained committed source events")
+            .events
+            .len(),
+        1
+    );
+}
+
+#[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+async fn raft_owner_shutdown_inflight_replication() {
+    // OpenRaft times append_entries by heartbeat_interval. Keep the real deadline
+    // longer than readiness checks so only explicit RPC release ends this pause.
+    let slow = Arc::new(
+        openraft::Config {
+            heartbeat_interval: 10_000,
+            election_timeout_min: 20_000,
+            election_timeout_max: 30_000,
+            snapshot_policy: openraft::SnapshotPolicy::LogsSinceLast(1000),
+            ..openraft::Config::default()
+        }
+        .validate()
+        .expect("long real RPC deadline"),
+    );
+    let routes = Arc::new(Routes::default());
+    let mut nodes = Vec::new();
+    for id in [0, 1, 2] {
+        nodes.push(Node::start(id, routes.clone(), slow.clone()).await);
+    }
+    nodes[0]
+        .owner()
+        .bootstrap(members(&[0, 1, 2]))
+        .await
+        .expect("real three-node bootstrap");
+    nodes[0]
+        .owner()
+        .raft()
+        .trigger()
+        .elect()
+        .await
+        .expect("explicit test election readiness");
+    let mut cluster = Cluster { nodes, routes };
+    let leader = cluster.leader(&[0, 1, 2]).await;
+    let response = cluster
+        .node(leader)
+        .owner()
+        .client_write(command(1))
+        .await
+        .expect("real replicated write before pause");
+    cluster.applied(&[0, 1, 2], response.log_id.index).await;
+    let follower = [0, 1, 2].into_iter().find(|id| *id != leader).expect("replication target");
+    let entered = Arc::new(tokio::sync::Notify::new());
+    let resume = Arc::new(tokio::sync::Notify::new());
+    cluster
+        .routes
+        .pause_append
+        .lock()
+        .expect("actual request pause")
+        .insert((leader, follower), (entered.clone(), resume.clone()));
+    cluster
+        .node(leader)
+        .owner()
+        .raft()
+        .trigger()
+        .heartbeat()
+        .await
+        .expect("real replication RPC");
+    tokio::time::timeout(Duration::from_secs(5), entered.notified())
+        .await
+        .expect("actual AppendEntries in flight");
+    let node = cluster.nodes.iter_mut().find(|node| node.id == leader).expect("leader node");
+    let owner = node.owner.take().expect("leader owner");
+    let waiters = owner.shutdown_waiters();
+    let observer = owner.raft().clone();
+    let shutdown = tokio::spawn(async move {
+        let result = owner.shutdown().await;
+        (owner, result)
+    });
+    wait_for_shutdown_drain(&waiters).await;
+    assert!(
+        observer.metrics().borrow().running_state.is_err(),
+        "OWNER_CORE_STOPPED_WHILE_REPLICATION_DRAIN_PENDING"
+    );
+    assert!(!shutdown.is_finished(), "OWNER_SHUTDOWN_DRAINS_INFLIGHT_REPLICATION");
+    resume.notify_one();
+    let (owner, result) = tokio::time::timeout(Duration::from_secs(10), shutdown)
+        .await
+        .expect("released real RPC lets shutdown finish")
+        .expect("replication shutdown task");
+    result.expect("actual replication reader drained");
+    cluster.routes.remove(leader);
+    drop(observer);
+    drop(owner);
+    drop(waiters);
+    let position = cluster
+        .nodes
+        .iter()
+        .position(|node| node.id == leader)
+        .expect("stopped leader");
+    let Node {
+        directory,
+        writer,
+        owner: _,
+        id,
+    } = cluster.nodes.remove(position);
+    drop(writer);
+    let store =
+        Store::open(directory.path().join("db"), identity(id)).expect("OWNER_SHUTDOWN_REPLICATION_RELEASES_LIVE_DATABASE");
+    assert_eq!(
+        store
+            .created_events_page(None, 256)
+            .expect("real quorum event remains")
+            .events
+            .len(),
+        1
+    );
+    drop(store);
+    cluster.stop().await;
+}
+
+#[path = "committed_created_authorization_tests.rs"]
+mod authorization;
@@ -0,0 +1,327 @@
+//! Durable OpenRaft application state and selected baseline snapshots.
+use crate::{
+    CaptureRaftEntry, CaptureResponse, CaptureTypeConfig, CheckpointLimits, NativeLogWriter, NativeSnapshotData,
+    NativeSnapshotError, StoreError, checkpoint_baseline, snapshot,
+};
+use openraft::{
+    OptionalSend, RaftSnapshotBuilder, Snapshot, SnapshotMeta, StorageError, StorageIOError, StoredMembership,
+    storage::RaftStateMachine,
+};
+use redb::{ReadableDatabase, ReadableTable, ReadableTableMetadata};
+use std::{
+    io::{Seek, Write},
+    sync::{
+        Arc,
+        atomic::{AtomicUsize, Ordering},
+    },
+};
+
+pub(crate) const SNAPSHOT_META: &str = "native_snapshot_meta_v1";
+type Meta = SnapshotMeta<u64, crate::CaptureNode>;
+#[derive(serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+struct SelectedMeta {
+    generation: u64,
+    meta: Meta,
+}
+
+pub(crate) fn publish_meta(
+    meta: &mut redb::Table<&str, &[u8]>,
+    baseline: &checkpoint_baseline::BaselineMetadata,
+    value: Meta,
+) -> Result<(), StoreError> {
+    let bytes = serde_json::to_vec(&SelectedMeta {
+        generation: baseline.generation,
+        meta: value,
+    })
+    .map_err(|_| StoreError::CorruptRecord)?;
+    if bytes.len() > 1024 * 1024 {
+        return Err(StoreError::ResourceLimit);
+    }
+    meta.insert(SNAPSHOT_META, bytes.as_slice())?;
+    meta.insert("native_snapshot_meta_required", &[1u8][..])?;
+    Ok(())
+}
+pub(crate) fn local_meta(
+    baseline: &checkpoint_baseline::BaselineMetadata,
+    identity: crate::NodeIdentity,
+) -> Result<Meta, StoreError> {
+    Ok(Meta {
+        last_log_id: baseline.header.applied.map(Into::into),
+        last_membership: StoredMembership::try_from(&baseline.header.membership)?,
+        snapshot_id: format!("native-local-{:x?}-{}-{}", identity.cluster_id, identity.node_id, baseline.generation),
+    })
+}
+pub(crate) fn publish_local_meta(
+    meta: &mut redb::Table<&str, &[u8]>,
+    baseline: &checkpoint_baseline::BaselineMetadata,
+) -> Result<(), StoreError> {
+    let identity = crate::decode_identity(meta.get("identity")?.ok_or(StoreError::CorruptRecord)?.value())?;
+    publish_meta(meta, baseline, local_meta(baseline, identity)?)
+}
+pub(crate) fn selected_meta(
+    meta: &impl ReadableTable<&'static str, &'static [u8]>,
+    baseline: &checkpoint_baseline::BaselineMetadata,
+) -> Result<Meta, StoreError> {
+    let Some(bytes) = meta.get(SNAPSHOT_META)? else {
+        if meta.get("native_snapshot_meta_required")?.is_some() {
+            return Err(StoreError::CorruptRecord);
+        }
+        return local_meta(
+            baseline,
+            crate::decode_identity(meta.get("identity")?.ok_or(StoreError::CorruptRecord)?.value())?,
+        );
+    };
+    if bytes.value().len() > 1024 * 1024 {
+        return Err(StoreError::CorruptRecord);
+    }
+    let selected: SelectedMeta = serde_json::from_slice(bytes.value()).map_err(|_| StoreError::CorruptRecord)?;
+    if selected.generation != baseline.generation
+        || selected.meta.last_log_id.map(crate::LogId::from) != baseline.header.applied
+        || crate::CaptureMembership::try_from(&selected.meta.last_membership)? != baseline.header.membership
+    {
+        return Err(StoreError::CorruptRecord);
+    }
+    Ok(selected.meta)
+}
+
+#[cfg(test)]
+pub(crate) type TestPause = std::sync::Mutex<Option<(Arc<tokio::sync::Notify>, Arc<tokio::sync::Notify>)>>;
+
+#[derive(Default)]
+pub(crate) struct ReceiveState {
+    count: AtomicUsize,
+    #[cfg(test)]
+    pub(crate) purge_recheck: TestPause,
+    pub(crate) changed: tokio::sync::Notify,
+    #[cfg(test)]
+    pub(crate) waiters: AtomicUsize,
+}
+pub(crate) struct ReceiveGuard {
+    state: Arc<ReceiveState>,
+}
+impl ReceiveGuard {
+    pub(crate) fn belongs_to(&self, state: &Arc<ReceiveState>) -> bool {
+        Arc::ptr_eq(&self.state, state)
+    }
+    fn new(state: Arc<ReceiveState>) -> Self {
+        state.count.fetch_add(1, Ordering::AcqRel);
+        Self { state }
+    }
+}
+impl Drop for ReceiveGuard {
+    fn drop(&mut self) {
+        self.state.count.fetch_sub(1, Ordering::AcqRel);
+        self.state.changed.notify_waiters();
+    }
+}
+impl ReceiveState {
+    pub(crate) fn pending(&self) -> bool {
+        self.count.load(Ordering::Acquire) > 0
+    }
+}
+
+/// Both adapters must use clones of the same writer. Limits apply while receiving.
+#[derive(Clone)]
+pub struct NativeRaftStateMachine {
+    pub(crate) writer: NativeLogWriter,
+    pub(crate) limits: CheckpointLimits,
+    #[cfg(test)]
+    pub(crate) build_before_ordering: Arc<TestPause>,
+    // Declared last: every clone releases its writer before the final owner lease.
+    pub(crate) owner_lease: Option<Arc<crate::raft_owner::OwnerLease>>,
+}
+impl NativeRaftStateMachine {
+    pub fn new(writer: NativeLogWriter, limits: CheckpointLimits) -> Result<Self, NativeSnapshotError> {
+        limits.validate()?;
+        limits.snapshot.total_bytes.checked_add(56).ok_or(StoreError::ResourceLimit)?;
+        Ok(Self {
+            writer,
+            limits,
+            #[cfg(test)]
+            build_before_ordering: Arc::new(TestPause::default()),
+            owner_lease: None,
+        })
+    }
+    fn maximum(&self) -> u64 {
+        self.limits.snapshot.total_bytes + 56
+    }
+    async fn current(&self) -> Result<Option<Snapshot<CaptureTypeConfig>>, NativeSnapshotError> {
+        let owner = self.writer.store.clone();
+        let limits = self.limits;
+        tokio::task::spawn_blocking(move || {
+            let store = owner.lock().map_err(|_| StoreError::CorruptRecord)?;
+            if !store.writable {
+                return Err(StoreError::WriteDisabled);
+            }
+            let tx = store.database.begin_read()?;
+            drop(store);
+            let table = tx.open_table(crate::META)?;
+            let Some(baseline) = checkpoint_baseline::read_metadata(&table)? else {
+                return Ok(None);
+            };
+            let meta = selected_meta(&table, &baseline)?;
+            let definitions = [
+                checkpoint_baseline::HEADS,
+                checkpoint_baseline::VERSIONS,
+                checkpoint_baseline::EVENTS,
+                checkpoint_baseline::OPERATIONS,
+            ];
+            let mut counts = [0; 4];
+            for (i, definition) in definitions.iter().enumerate() {
+                counts[i] = tx.open_table(*definition)?.len()?;
+            }
+            let out = snapshot::CapabilityOutput::new(Vec::new(), limits.snapshot.total_bytes);
+            let mut encoder =
+                snapshot::Encoder::new(out, &baseline.header, counts, limits.snapshot).map_err(|_| StoreError::CorruptRecord)?;
+            for (definition, section) in definitions.into_iter().zip([
+                snapshot::Section::Heads,
+                snapshot::Section::Versions,
+                snapshot::Section::Events,
+                snapshot::Section::Operations,
+            ]) {
+                for row in tx.open_table(definition)?.iter()? {
+                    let (key, value) = row?;
+                    encoder
+                        .row(snapshot::Row {
+                            section,
+                            key: key.value(),
+                            value: value.value(),
+                        })
+                        .map_err(|_| StoreError::CorruptRecord)?;
+                }
+            }
+            let artifact = encoder.finish().map_err(|_| StoreError::CorruptRecord)?.finish();
+            let mut data = NativeSnapshotData::new(limits.snapshot.total_bytes + 56).map_err(|_| StoreError::ResourceLimit)?;
+            data.write_all(&artifact.manifest.bytes)
+                .map_err(|_| StoreError::ResourceLimit)?;
+            data.write_all(&artifact.payload).map_err(|_| StoreError::ResourceLimit)?;
+            data.rewind().map_err(|_| StoreError::CorruptRecord)?;
+            Ok::<_, StoreError>(Some(Snapshot {
+                meta,
+                snapshot: Box::new(data),
+            }))
+        })
+        .await
+        .map_err(NativeSnapshotError::Worker)?
+        .map_err(NativeSnapshotError::Store)
+    }
+}
+fn failure(error: impl std::error::Error + 'static) -> StorageError<u64> {
+    StorageIOError::read_state_machine(openraft::AnyError::new(&error)).into()
+}
+impl RaftStateMachine<CaptureTypeConfig> for NativeRaftStateMachine {
+    type SnapshotBuilder = Self;
+    async fn applied_state(
+        &mut self,
+    ) -> Result<(Option<openraft::LogId<u64>>, StoredMembership<u64, crate::CaptureNode>), StorageError<u64>> {
+        let (id, membership) = self.writer.applied_state().await.map_err(failure)?;
+        Ok((id.map(Into::into), StoredMembership::try_from(&membership).map_err(failure)?))
+    }
+    async fn apply<I>(&mut self, entries: I) -> Result<Vec<CaptureResponse>, StorageError<u64>>
+    where
+        I: IntoIterator<Item = CaptureRaftEntry> + OptionalSend,
+        I::IntoIter: OptionalSend,
+    {
+        self.writer
+            .apply_committed(entries.into_iter().collect())
+            .await
+            .map_err(failure)
+    }
+    async fn get_snapshot_builder(&mut self) -> Self {
+        self.clone()
+    }
+    async fn begin_receiving_snapshot(&mut self) -> Result<Box<NativeSnapshotData>, StorageError<u64>> {
+        let mut data = NativeSnapshotData::new(self.maximum()).map_err(failure)?;
+        data.attach_receive_guard(ReceiveGuard::new(self.writer.snapshot_receives.clone()));
+        Ok(Box::new(data))
+    }
+    async fn install_snapshot(&mut self, meta: &Meta, data: Box<NativeSnapshotData>) -> Result<(), StorageError<u64>> {
+        crate::native_snapshot_install::install(self.writer.clone(), self.limits, meta.clone(), data)
+            .await
+            .map_err(failure)
+    }
+    async fn get_current_snapshot(&mut self) -> Result<Option<Snapshot<CaptureTypeConfig>>, StorageError<u64>> {
+        self.current().await.map_err(failure)
+    }
+}
+impl RaftSnapshotBuilder<CaptureTypeConfig> for NativeRaftStateMachine {
+    async fn build_snapshot(&mut self) -> Result<Snapshot<CaptureTypeConfig>, StorageError<u64>> {
+        #[cfg(test)]
+        {
+            let hook = self.build_before_ordering.lock().expect("build test hook").take();
+            if let Some((entered, resume)) = hook {
+                entered.notify_one();
+                resume.notified().await;
+            }
+        }
+        let limits = self.limits;
+        let published_empty = self
+            .writer
+            .write(move |store| {
+                let tx = store.write_transaction()?;
+                let mut meta = tx.open_table(crate::META)?;
+                // Decide the empty/nonempty path under the original writer ordering.
+                if crate::application::read_applied(&meta)?.is_some() {
+                    return Ok(false);
+                }
+                let identity = crate::decode_identity(meta.get("identity")?.ok_or(StoreError::CorruptRecord)?.value())?;
+                let membership =
+                    crate::CaptureMembership::decode(meta.get("native_membership")?.ok_or(StoreError::CorruptRecord)?.value())?;
+                let previous = checkpoint_baseline::read_metadata(&meta)?;
+                let generation = previous.map_or(Ok(1), |v| v.generation.checked_add(1).ok_or(StoreError::ResourceLimit))?;
+                let baseline = checkpoint_baseline::BaselineMetadata {
+                    generation,
+                    header: snapshot::Header {
+                        cluster_id: identity.cluster_id,
+                        applied: None,
+                        membership,
+                        binding: crate::application::read_binding(&meta)?,
+                    },
+                    limits,
+                };
+                for (live, base) in [
+                    (crate::application::HEADS, checkpoint_baseline::HEADS),
+                    (crate::application::VERSIONS, checkpoint_baseline::VERSIONS),
+                    (crate::application::EVENTS, checkpoint_baseline::EVENTS),
+                    (crate::application::OPERATIONS, checkpoint_baseline::OPERATIONS),
+                ] {
+                    if tx.open_table(live)?.len()? != 0 {
+                        return Err(StoreError::CorruptRecord);
+                    }
+                    tx.open_table(base)?.retain(|_, _| false)?;
+                }
+                meta.insert("schema", 7u64.to_be_bytes().as_slice())?;
+                meta.insert(checkpoint_baseline::META_KEY, checkpoint_baseline::encode_metadata(&baseline)?.as_slice())?;
+                publish_local_meta(&mut meta, &baseline)?;
+                drop(meta);
+                store.commit(tx)?;
+                Ok(true)
+            })
+            .await
+            .map_err(failure)?;
+        if !published_empty {
+            self.writer.checkpoint_local(self.limits).await.map_err(failure)?;
+        }
+        self.current()
+            .await
+            .map_err(failure)?
+            .ok_or_else(|| failure(StoreError::CorruptRecord))
+    }
+}
+
+#[cfg(test)]
+pub(crate) struct Waiter(Arc<ReceiveState>);
+#[cfg(test)]
+impl Waiter {
+    pub(crate) fn new(state: Arc<ReceiveState>) -> Self {
+        state.waiters.fetch_add(1, Ordering::AcqRel);
+        Self(state)
+    }
+}
+#[cfg(test)]
+impl Drop for Waiter {
+    fn drop(&mut self) {
+        self.0.waiters.fetch_sub(1, Ordering::AcqRel);
+    }
+}
@@ -0,0 +1,583 @@
+//! Actual OpenRaft storage traits over real redb; no simulated state machine.
+use super::*;
+use openraft::{
+    RaftLogReader, RaftSnapshotBuilder,
+    storage::{RaftLogStorage, RaftLogStorageExt, RaftStateMachine},
+};
+use std::{
+    collections::{BTreeMap, BTreeSet},
+    io::SeekFrom,
+};
+use tokio::io::{AsyncReadExt, AsyncSeekExt, AsyncWriteExt};
+
+pub(super) type Snapshot = openraft::Snapshot<CaptureTypeConfig>;
+pub(super) fn identity(node_id: u64) -> NodeIdentity {
+    NodeIdentity {
+        cluster_id: [7; 16],
+        node_id,
+    }
+}
+pub(super) fn binding() -> CaptureBinding {
+    CaptureBinding {
+        revision: 9,
+        binding_id: [8; 16],
+        target_id: [9; 16],
+    }
+}
+pub(super) fn id(index: u64) -> LogId {
+    LogId {
+        term: 8,
+        leader_node: 0,
+        index,
+    }
+}
+pub(super) fn entry(index: u64) -> CaptureRaftEntry {
+    let payload = if index == 0 {
+        let membership = CaptureMembership::new(
+            Some(id(0)),
+            vec![BTreeSet::from([0])],
+            BTreeMap::from([(0, CaptureNode::new("http://127.0.0.1:7000".into()).expect("member"))]),
+        )
+        .expect("native membership");
+        let stored = openraft::StoredMembership::try_from(&membership).expect("stored membership");
+        openraft::EntryPayload::Membership(stored.membership().clone())
+    } else {
+        let byte = u8::try_from(index).expect("bounded test event");
+        openraft::EntryPayload::Normal(
+            CaptureCommand::try_created(DecideCreated {
+                operation_id: [byte; 16],
+                object: ObjectIdentity {
+                    bucket_incarnation: [2; 16],
+                    key: vec![byte],
+                },
+                expected_head: None,
+                prepared: PreparedIdentity {
+                    version_id: [byte; 16],
+                    preparation_id: [4; 16],
+                    content_digest: [5; 32],
+                    content_length: 123,
+                },
+                binding_revision: binding().revision,
+            })
+            .expect("native command"),
+        )
+    };
+    CaptureRaftEntry {
+        log_id: id(index).into(),
+        payload,
+    }
+}
+pub(super) fn entries(end: u64) -> Vec<CaptureRaftEntry> {
+    (0..=end).map(entry).collect()
+}
+pub(super) struct Fixture {
+    pub directory: tempfile::TempDir,
+    pub writer: NativeLogWriter,
+    pub logs: NativeRaftLogStore,
+    pub sm: NativeRaftStateMachine,
+    pub identity: NodeIdentity,
+}
+impl Fixture {
+    pub fn new(node_id: u64) -> Self {
+        Self::with_identity_binding(identity(node_id), binding())
+    }
+    pub fn with_identity_binding(identity: NodeIdentity, binding: CaptureBinding) -> Self {
+        let directory = tempfile::tempdir().expect("real state machine directory");
+        let mut store = Store::initialize(directory.path().join("db"), identity).expect("real redb initialization");
+        store.pin_capture_binding(binding).expect("native binding");
+        store.enable_native_empty().expect("native mode");
+        store.promote_created_provenance().expect("provenance");
+        let writer = NativeLogWriter::new(store);
+        let logs = NativeRaftLogStore::new(writer.clone());
+        let sm = NativeRaftStateMachine::new(writer.clone(), CheckpointLimits::default()).expect("real state machine adapter");
+        Self {
+            directory,
+            writer,
+            logs,
+            sm,
+            identity,
+        }
+    }
+    pub async fn apply(&mut self, batch: Vec<CaptureRaftEntry>) -> Vec<CaptureResponse> {
+        self.logs
+            .blocking_append(batch.clone())
+            .await
+            .expect("actual trait append and library flush callback");
+        self.sm.apply(batch).await.expect("actual state machine trait apply")
+    }
+    pub async fn build(&mut self) -> Snapshot {
+        self.sm
+            .get_snapshot_builder()
+            .await
+            .build_snapshot()
+            .await
+            .expect("actual durable snapshot builder")
+    }
+    pub async fn page(&self) -> EventPage {
+        self.writer
+            .created_events_page(None, 256)
+            .await
+            .expect("authoritative native events")
+    }
+}
+pub(super) async fn snapshot_bytes(snapshot: &mut Snapshot) -> Vec<u8> {
+    snapshot.snapshot.seek(SeekFrom::Start(0)).await.expect("snapshot rewind");
+    let mut bytes = Vec::new();
+    snapshot
+        .snapshot
+        .read_to_end(&mut bytes)
+        .await
+        .expect("actual snapshot bytes");
+    bytes
+}
+pub(super) async fn receive(
+    sm: &mut NativeRaftStateMachine,
+    bytes: &[u8],
+) -> Box<<CaptureTypeConfig as openraft::RaftTypeConfig>::SnapshotData> {
+    let mut data = sm.begin_receiving_snapshot().await.expect("actual bounded receive handle");
+    data.write_all(bytes).await.expect("receive real snapshot bytes");
+    data.seek(SeekFrom::Start(0)).await.expect("received snapshot rewind");
+    data
+}
+pub(super) async fn install(sm: &mut NativeRaftStateMachine, snapshot: &mut Snapshot) {
+    let bytes = snapshot_bytes(snapshot).await;
+    let data = receive(sm, &bytes).await;
+    sm.install_snapshot(&snapshot.meta, data)
+        .await
+        .expect("actual state machine trait install");
+}
+
+#[tokio::test]
+async fn raft_state_machine_apply_restart() {
+    let mut f = Fixture::new(0);
+    assert_eq!(f.sm.applied_state().await.expect("empty applied state").0, None);
+    assert!(f.sm.get_current_snapshot().await.expect("no published snapshot").is_none());
+    let batch = entries(2);
+    let responses = f.apply(batch.clone()).await;
+    assert_eq!(responses.len(), 3, "RAFT_SM_REAL_RESPONSE_COUNT");
+    assert_eq!(responses[0].kind(), CaptureResponseKind::Membership);
+    assert!(matches!(responses[1].created_result(), Some(CreatedResult::Created { .. })));
+    let state = f.sm.applied_state().await.expect("durable applied state");
+    assert_eq!(state.0, Some(id(2).into()));
+    assert_eq!(state.1.log_id(), &Some(id(0).into()));
+    let before = f.page().await;
+    let Fixture {
+        directory,
+        writer,
+        logs,
+        sm,
+        identity,
+    } = f;
+    drop(sm);
+    drop(logs);
+    drop(writer);
+    let writer = NativeLogWriter::new(Store::open(directory.path().join("db"), identity).expect("restart real store"));
+    let mut sm = NativeRaftStateMachine::new(writer.clone(), CheckpointLimits::default()).expect("reopened state machine");
+    assert_eq!(
+        sm.applied_state().await.expect("reopened applied state"),
+        state,
+        "RAFT_SM_APPLY_ACK_DURABLE"
+    );
+    assert_eq!(writer.created_events_page(None, 256).await.expect("reopened events"), before);
+    assert_eq!(
+        sm.apply(batch).await.expect("exact replay responses"),
+        responses,
+        "RAFT_SM_REPLAY_SAME_RESPONSES"
+    );
+    assert_eq!(
+        writer
+            .created_events_page(None, 256)
+            .await
+            .expect("no duplicate replay events"),
+        before
+    );
+}
+
+#[tokio::test]
+async fn raft_state_machine_snapshot_build_get_restart() {
+    let mut f = Fixture::new(0);
+    f.apply(entries(2)).await;
+    let mut built = f.build().await;
+    let meta = built.meta.clone();
+    let bytes = snapshot_bytes(&mut built).await;
+    assert_eq!(meta.last_log_id, Some(id(2).into()), "RAFT_SM_SNAPSHOT_META_CUT");
+    assert_eq!(meta.last_membership, f.sm.applied_state().await.expect("membership").1);
+    f.apply(vec![entry(3)]).await;
+    let mut current =
+        f.sm.get_current_snapshot()
+            .await
+            .expect("persisted snapshot")
+            .expect("snapshot exists");
+    assert_eq!(current.meta, meta, "RAFT_SM_LIVE_APPLY_DOES_NOT_REPIN_SNAPSHOT");
+    assert_eq!(snapshot_bytes(&mut current).await, bytes, "RAFT_SM_SNAPSHOT_FROZEN_BYTES");
+    let Fixture {
+        directory,
+        writer,
+        logs,
+        sm,
+        identity,
+    } = f;
+    drop(current);
+    drop(built);
+    drop(sm);
+    drop(logs);
+    drop(writer);
+    let writer = NativeLogWriter::new(Store::open(directory.path().join("db"), identity).expect("restart snapshot owner"));
+    let mut sm = NativeRaftStateMachine::new(writer, CheckpointLimits::default()).expect("reopen adapter");
+    let mut current = sm
+        .get_current_snapshot()
+        .await
+        .expect("restarted current snapshot")
+        .expect("durable snapshot exists");
+    assert_eq!(current.meta, meta, "RAFT_SM_COMPLETE_META_RECONSTRUCTED");
+    assert_eq!(snapshot_bytes(&mut current).await, bytes, "RAFT_SM_RESTART_SNAPSHOT_BYTES");
+}
+
+#[tokio::test]
+async fn raft_state_machine_receive_bounds() {
+    let f = Fixture::new(0);
+    let mut limits = CheckpointLimits::default();
+    limits.snapshot.total_bytes = 1024;
+    let mut sm = NativeRaftStateMachine::new(f.writer.clone(), limits).expect("bounded state machine");
+    let mut data = sm.begin_receiving_snapshot().await.expect("bounded receive");
+    data.write_all(&[1; 16]).await.expect("small write");
+    assert!(data.seek(SeekFrom::Start(u64::MAX)).await.is_err(), "RAFT_SM_RECEIVE_SEEK_BOUND");
+    data.seek(SeekFrom::Start(0)).await.expect("restore seek");
+    let mut bytes = [0; 16];
+    data.read_exact(&mut bytes).await.expect("small received bytes");
+    assert_eq!(bytes, [1; 16], "RAFT_SM_BAD_SEEK_PRESERVES_BYTES");
+    assert!(data.write_all(&[2; 2048]).await.is_err(), "RAFT_SM_RECEIVE_WRITE_BOUND_BEFORE_INSTALL");
+    assert_eq!(sm.applied_state().await.expect("receive is not install").0, None);
+    assert!(sm.get_current_snapshot().await.expect("no received publication").is_none());
+}
+
+#[tokio::test]
+async fn raft_state_machine_install_rejects_meta_identity() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let mut snapshot = source.build().await;
+    let bytes = snapshot_bytes(&mut snapshot).await;
+    let mut target = Fixture::new(1);
+    target.apply(entries(1)).await;
+    let old_page = target.page().await;
+    let stable = std::fs::read(target.directory.path().join("db")).expect("nonempty target bytes");
+    let mut wrong = snapshot.meta.clone();
+    wrong.last_log_id = Some(LogId { term: 9, ..id(2) }.into());
+    let data = receive(&mut target.sm, &bytes).await;
+    assert!(
+        target.sm.install_snapshot(&wrong, data).await.is_err(),
+        "SNAPSHOT_META_BOUND_TO_FRAME: full LogId term cannot differ from snapshot header"
+    );
+    assert_eq!(target.sm.applied_state().await.expect("rejection state").0, Some(id(1).into()));
+    assert_eq!(target.page().await, old_page, "RAFT_SM_META_REJECTION_RETAINS_EVENTS");
+    assert_eq!(
+        std::fs::read(target.directory.path().join("db")).expect("unchanged nonempty target"),
+        stable,
+        "RAFT_SM_META_REJECTION_NO_MUTATION"
+    );
+    assert!(target.sm.get_current_snapshot().await.expect("rejection snapshot").is_none());
+    let mut wrong = snapshot.meta.clone();
+    wrong.last_membership = openraft::StoredMembership::default();
+    let data = receive(&mut target.sm, &bytes).await;
+    assert!(
+        target.sm.install_snapshot(&wrong, data).await.is_err(),
+        "RAFT_SM_MEMBERSHIP_BOUND_TO_FRAME"
+    );
+    let mut foreign = Fixture::with_identity_binding(
+        NodeIdentity {
+            cluster_id: [6; 16],
+            ..identity(1)
+        },
+        binding(),
+    );
+    let data = receive(&mut foreign.sm, &bytes).await;
+    assert!(
+        foreign.sm.install_snapshot(&snapshot.meta, data).await.is_err(),
+        "RAFT_SM_CLUSTER_BOUND_TO_DURABLE_OWNER"
+    );
+    let mut foreign = Fixture::with_identity_binding(
+        identity(1),
+        CaptureBinding {
+            target_id: [5; 16],
+            ..binding()
+        },
+    );
+    let data = receive(&mut foreign.sm, &bytes).await;
+    assert!(
+        foreign.sm.install_snapshot(&snapshot.meta, data).await.is_err(),
+        "RAFT_SM_CAPTURE_BINDING_BOUND"
+    );
+}
+
+#[tokio::test]
+async fn raft_state_machine_install_preserves_local_identity_vote() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let mut snapshot = source.build().await;
+    snapshot.meta.snapshot_id = "foreign-snapshot-id-preserved-verbatim".into();
+    let mut target = Fixture::new(1);
+    let vote = openraft::Vote::new_committed(10, 1);
+    target.logs.save_vote(&vote).await.expect("local election vote");
+    install(&mut target.sm, &mut snapshot).await;
+    assert_eq!(
+        target.writer.durable_identity().await.expect("local identity"),
+        identity(1),
+        "RAFT_SM_INSTALL_PRESERVES_LOCAL_NODE"
+    );
+    assert_eq!(
+        target.logs.read_vote().await.expect("local vote"),
+        Some(vote),
+        "RAFT_SM_INSTALL_PRESERVES_LOCAL_VOTE"
+    );
+    let mut current = target
+        .sm
+        .get_current_snapshot()
+        .await
+        .expect("installed snapshot")
+        .expect("present");
+    assert_eq!(current.meta, snapshot.meta, "RAFT_SM_INSTALL_PRESERVES_COMPLETE_INCOMING_META");
+    assert_eq!(snapshot_bytes(&mut current).await, snapshot_bytes(&mut snapshot).await);
+    let Fixture {
+        directory,
+        writer,
+        logs,
+        sm,
+        identity,
+    } = target;
+    drop(current);
+    drop(sm);
+    drop(logs);
+    drop(writer);
+    let writer = NativeLogWriter::new(Store::open(directory.path().join("db"), identity).expect("restart installed target"));
+    let mut logs = NativeRaftLogStore::new(writer.clone());
+    let mut sm = NativeRaftStateMachine::new(writer.clone(), CheckpointLimits::default()).expect("reopened installed SM");
+    assert_eq!(writer.durable_identity().await.expect("restarted local identity"), identity);
+    assert_eq!(logs.read_vote().await.expect("restarted vote"), Some(vote));
+    assert_eq!(
+        sm.get_current_snapshot()
+            .await
+            .expect("restarted snapshot")
+            .expect("present")
+            .meta,
+        snapshot.meta,
+        "RAFT_SM_INCOMING_SNAPSHOT_ID_DURABLE"
+    );
+}
+
+#[tokio::test]
+async fn raft_state_machine_install_suffix() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let mut snapshot = source.build().await;
+    let mut target = Fixture::new(1);
+    target
+        .logs
+        .blocking_append(entries(4))
+        .await
+        .expect("retained local log suffix");
+    target.sm.apply(entries(1)).await.expect("old local applied prefix");
+    target
+        .logs
+        .save_committed(Some(id(4).into()))
+        .await
+        .expect("known committed compatible suffix");
+    install(&mut target.sm, &mut snapshot).await;
+    assert_eq!(
+        target
+            .logs
+            .get_log_reader()
+            .await
+            .try_get_log_entries(3..)
+            .await
+            .expect("compatible suffix"),
+        vec![entry(3), entry(4)],
+        "RAFT_SM_INSTALL_PRESERVES_COMPATIBLE_COMMITTED_SUFFIX"
+    );
+    assert_eq!(
+        target.logs.read_committed().await.expect("preserved committed marker"),
+        Some(id(4).into())
+    );
+    target
+        .sm
+        .apply(vec![entry(3), entry(4)])
+        .await
+        .expect("apply retained committed suffix");
+    assert_eq!(target.page().await.events.len(), 4);
+    let mut conflict = Fixture::new(2);
+    let mut batch = entries(4);
+    for entry in &mut batch[2..] {
+        entry.log_id = LogId {
+            term: 9,
+            ..id(entry.log_id.index)
+        }
+        .into();
+    }
+    conflict
+        .logs
+        .blocking_append(batch)
+        .await
+        .expect("conflicting uncommitted suffix");
+    conflict.sm.apply(entries(1)).await.expect("protected applied prefix");
+    install(&mut conflict.sm, &mut snapshot).await;
+    assert!(
+        conflict
+            .logs
+            .get_log_reader()
+            .await
+            .try_get_log_entries(3..)
+            .await
+            .expect("discarded conflict")
+            .is_empty(),
+        "RAFT_SM_INSTALL_DROPS_ONLY_UNCOMMITTED_CONFLICT"
+    );
+    conflict.apply(vec![entry(3)]).await;
+    assert_eq!(conflict.sm.applied_state().await.expect("new suffix applied").0, Some(id(3).into()));
+}
+
+#[tokio::test]
+async fn raft_state_machine_install_restart_continue() {
+    let mut source = Fixture::new(0);
+    source.apply(entries(2)).await;
+    let expected = source.page().await;
+    let mut snapshot = source.build().await;
+    let mut target = Fixture::new(1);
+    install(&mut target.sm, &mut snapshot).await;
+    assert_eq!(target.page().await, expected, "RAFT_SM_INSTALL_REBUILDS_SOURCE_EVENTS");
+    let Fixture {
+        directory,
+        writer,
+        logs,
+        sm,
+        identity,
+    } = target;
+    drop(sm);
+    drop(logs);
+    drop(writer);
+    let writer = NativeLogWriter::new(Store::open(directory.path().join("db"), identity).expect("restart installed redb"));
+    let mut logs = NativeRaftLogStore::new(writer.clone());
+    let mut sm = NativeRaftStateMachine::new(writer.clone(), CheckpointLimits::default()).expect("restart installed adapter");
+    assert_eq!(sm.applied_state().await.expect("snapshot applied state").0, Some(id(2).into()));
+    logs.blocking_append(vec![entry(3)])
+        .await
+        .expect("append after install restart");
+    sm.apply(vec![entry(3)]).await.expect("apply after install restart");
+    let all = writer.created_events_page(None, 256).await.expect("retained plus new event");
+    assert_eq!(all.events.len(), 3);
+    assert_eq!(all.events[..2], expected.events);
+    let next = writer
+        .created_events_page(expected.next, 1)
+        .await
+        .expect("snapshot cursor resumes");
+    assert_eq!(next.events[0].event_id, [3; 16], "RAFT_SM_INSTALL_CURSOR_CONTINUITY");
+}
+
+#[tokio::test]
+async fn raft_state_machine_empty_snapshot_restart_continue() {
+    let mut source = Fixture::new(0);
+    let mut target = Fixture::new(42);
+    let mut snapshot = source.sm.build_snapshot().await.expect("durable empty baseline");
+    assert_eq!(snapshot.meta.last_log_id, None);
+    let empty_meta = snapshot.meta.clone();
+    let received = receive(&mut target.sm, &snapshot_bytes(&mut snapshot).await).await;
+    target
+        .sm
+        .install_snapshot(&empty_meta, received)
+        .await
+        .expect("install empty frame");
+    let mut target = target.reopen();
+    let current = target
+        .sm
+        .get_current_snapshot()
+        .await
+        .expect("empty after reopen")
+        .expect("durable selection");
+    assert_eq!(current.meta, empty_meta);
+    target
+        .logs
+        .blocking_append(entries(1))
+        .await
+        .expect("real append after empty install");
+    target.sm.apply(entries(1)).await.expect("real apply after empty install");
+    let mut target = target.reopen();
+    assert_eq!(
+        target.sm.applied_state().await.expect("replay empty baseline + suffix").0,
+        Some(id(1).into())
+    );
+    assert_eq!(
+        target
+            .writer
+            .created_events_page(None, 16)
+            .await
+            .expect("event after empty baseline")
+            .events
+            .len(),
+        1
+    );
+    let refreshed = target
+        .sm
+        .build_snapshot()
+        .await
+        .expect("nonempty checkpoint supersedes empty");
+    assert_eq!(refreshed.meta.last_log_id, Some(id(1).into()));
+}
+
+impl Fixture {
+    fn reopen(self) -> Self {
+        let Self {
+            directory,
+            writer,
+            logs,
+            sm,
+            identity,
+        } = self;
+        drop(sm);
+        drop(logs);
+        drop(writer);
+        let writer =
+            NativeLogWriter::new(Store::open(directory.path().join("db"), identity).expect("real empty baseline restart"));
+        let logs = NativeRaftLogStore::new(writer.clone());
+        let sm =
+            NativeRaftStateMachine::new(writer.clone(), CheckpointLimits::default()).expect("reopened empty baseline adapter");
+        Self {
+            directory,
+            writer,
+            logs,
+            sm,
+            identity,
+        }
+    }
+}
+
+#[tokio::test]
+async fn raft_state_machine_build_competes_with_initial_apply() {
+    let mut target = Fixture::new(0);
+    target
+        .logs
+        .blocking_append(entries(1))
+        .await
+        .expect("real pending initial batch");
+    let entered = std::sync::Arc::new(tokio::sync::Notify::new());
+    let resume = std::sync::Arc::new(tokio::sync::Notify::new());
+    *target.sm.build_before_ordering.lock().expect("build pause") = Some((entered.clone(), resume.clone()));
+    let mut builder = target.sm.get_snapshot_builder().await;
+    let task = tokio::spawn(async move { builder.build_snapshot().await });
+    tokio::time::timeout(std::time::Duration::from_secs(5), entered.notified())
+        .await
+        .expect("builder began before initial apply");
+    target
+        .sm
+        .apply(entries(1))
+        .await
+        .expect("actual initial apply while builder pending");
+    resume.notify_one();
+    let snapshot = tokio::time::timeout(std::time::Duration::from_secs(5), task)
+        .await
+        .expect("build continues")
+        .expect("builder task")
+        .expect("initial apply does not make snapshot build a fatal storage error");
+    assert_eq!(
+        snapshot.meta.last_log_id,
+        Some(id(1).into()),
+        "SNAPSHOT_BUILD_INITIAL_APPLY_RACE_USES_NONEMPTY_CHECKPOINT"
+    );
+}
@@ -0,0 +1,559 @@
+//! Disposable event ordering index. EVENTS remains the sole event authority.
+use crate::{
+    CaptureBinding, CreatedEvent, CreatedResult, LogId, NativeLogWriter, NativeWriteError, Store, StoreError, application,
+    decode_id, encode_id,
+};
+use redb::{ReadableDatabase, ReadableTable, ReadableTableMetadata, TableDefinition, WriteTransaction};
+
+pub const MAX_EVENT_PAGE: usize = 256;
+pub(crate) const EVENT_SEQUENCE: TableDefinition<u64, &[u8]> = TableDefinition::new("event_sequence_cache_v1");
+
+/// Persist only after every event preceding this cursor has a durable consumer ACK.
+/// Deserialization does not grant authority: each use checks the stored event.
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct EventCursor {
+    cluster: [u8; 16],
+    term: u64,
+    leader: u64,
+    index: u64,
+    event: [u8; 16],
+    binding_id: [u8; 16],
+    target_id: [u8; 16],
+    revision: u64,
+}
+impl EventCursor {
+    pub(crate) fn new(cluster: [u8; 16], event: [u8; 16], value: &CreatedEvent) -> Self {
+        Self {
+            cluster,
+            term: value.decision_log_id.term,
+            leader: value.decision_log_id.leader_node,
+            index: value.decision_log_id.index,
+            event,
+            binding_id: value.binding.binding_id,
+            target_id: value.binding.target_id,
+            revision: value.binding.revision,
+        }
+    }
+    fn matches(&self, cluster: [u8; 16], event: [u8; 16], value: &CreatedEvent) -> bool {
+        self == &Self::new(cluster, event, value)
+    }
+}
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct SourceEvent {
+    pub event_id: [u8; 16],
+    pub captured: CreatedEvent,
+}
+#[derive(Debug, Clone, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct EventPage {
+    pub events: Vec<SourceEvent>,
+    pub next: Option<EventCursor>,
+}
+fn encoded_index(id: LogId, event: [u8; 16]) -> Vec<u8> {
+    let mut bytes = encode_id(id).to_vec();
+    bytes.extend(event);
+    bytes
+}
+fn decode_index(bytes: &[u8]) -> Result<(LogId, [u8; 16]), StoreError> {
+    if bytes.len() != 40 {
+        return Err(StoreError::CorruptRecord);
+    }
+    let id = decode_id(&bytes[..24])?;
+    let event: [u8; 16] = bytes[24..].try_into().map_err(|_| StoreError::CorruptRecord)?;
+    if event == [0; 16] {
+        return Err(StoreError::CorruptRecord);
+    }
+    Ok((id, event))
+}
+pub(crate) fn insert_event_index(tx: &WriteTransaction, id: LogId, event: [u8; 16]) -> Result<(), StoreError> {
+    let bytes = encoded_index(id, event);
+    let mut index = tx.open_table(EVENT_SEQUENCE)?;
+    if index.get(id.index)?.is_some() {
+        return Err(StoreError::CorruptRecord);
+    }
+    index.insert(id.index, bytes.as_slice())?;
+    Ok(())
+}
+
+/// Check the event semantics required to derive ordering from checkpoint authority.
+/// This deliberately does not grant snapshot installation or source authority.
+fn validated_event(
+    key: &[u8],
+    bytes: &[u8],
+    operations: &impl ReadableTable<&'static [u8], &'static [u8]>,
+    versions: &impl ReadableTable<&'static [u8], &'static [u8]>,
+    schema: u64,
+    binding: &CaptureBinding,
+    applied: Option<LogId>,
+) -> Result<([u8; 16], CreatedEvent), StoreError> {
+    let event_id: [u8; 16] = key.try_into().map_err(|_| StoreError::CorruptRecord)?;
+    let event = application::decode_event(bytes)?;
+    let cut = applied.ok_or(StoreError::CorruptRecord)?;
+    let operation = operations.get(key)?.ok_or(StoreError::CorruptRecord)?;
+    let first_decision = if matches!(schema, 5 | 6 | 7) {
+        Some(application::decode_provenance(operation.value())?.first_decision_log_id)
+    } else {
+        None
+    };
+    let operation = application::decode_stored_operation(operation.value(), schema)?;
+    let command = crate::DecideCreated::decode(&operation.canonical_command)?;
+    let version_key = application::version_key(&event.object, event.prepared.version_id)?;
+    let version = versions.get(version_key.as_slice())?.ok_or(StoreError::CorruptRecord)?;
+    let version = application::decode_version(version.value())?;
+    if first_decision.is_some_and(|first| first != event.decision_log_id)
+        || event_id == [0; 16]
+        || event.decision_log_id.index > cut.index
+        || (event.decision_log_id.index == cut.index && event.decision_log_id != cut)
+        || event.binding != *binding
+        || command.operation_id != event_id
+        || command.object != event.object
+        || command.prepared != event.prepared
+        || command.binding_revision != event.binding.revision
+        || operation.result
+            != (CreatedResult::Created {
+                version_id: event.prepared.version_id,
+                event_id,
+                decision_log_id: event.decision_log_id,
+            })
+        || version
+            != (crate::CreatedVersion {
+                prepared: event.prepared.clone(),
+                decision_log_id: event.decision_log_id,
+                binding_revision: event.binding.revision,
+            })
+    {
+        return Err(StoreError::CorruptRecord);
+    }
+    Ok((event_id, event))
+}
+
+/// Rebuild from the same authority rows emitted by the existing four-section checkpoint.
+/// Callers must independently establish authority and validate all other state before promotion.
+pub(crate) fn rebuild_from_authority(
+    tx: &WriteTransaction,
+    schema: u64,
+    binding: &CaptureBinding,
+    applied: Option<LogId>,
+) -> Result<(), StoreError> {
+    let events = tx.open_table(application::EVENTS)?;
+    let operations = tx.open_table(application::OPERATIONS)?;
+    let versions = tx.open_table(application::VERSIONS)?;
+    {
+        let mut index = tx.open_table(EVENT_SEQUENCE)?;
+        index.retain(|_, _| false)?;
+    }
+    for row in events.iter()? {
+        let (key, bytes) = row?;
+        let (event, value) = validated_event(key.value(), bytes.value(), &operations, &versions, schema, binding, applied)?;
+        insert_event_index(tx, value.decision_log_id, event)?;
+    }
+    Ok(())
+}
+impl Store {
+    pub(crate) fn ensure_event_index(&mut self) -> Result<(), StoreError> {
+        let read = self.database.begin_read()?;
+        let meta = read.open_table(crate::META)?;
+        let schema = crate::number(meta.get("schema")?.ok_or(StoreError::CorruptRecord)?.value())?;
+        if !matches!(schema, 3..=7) {
+            return Ok(());
+        }
+        let binding = application::read_binding(&meta)?;
+        let applied = application::read_applied(&meta)?;
+        let events = read.open_table(application::EVENTS)?;
+        let complete = match read.open_table(EVENT_SEQUENCE) {
+            Ok(index) => {
+                let mut complete = index.len()? == events.len()?;
+                if complete {
+                    for row in events.iter()? {
+                        let (key, bytes) = row?;
+                        let value = application::decode_event(bytes.value())?;
+                        let expected =
+                            encoded_index(value.decision_log_id, key.value().try_into().map_err(|_| StoreError::CorruptRecord)?);
+                        if index
+                            .get(value.decision_log_id.index)?
+                            .is_none_or(|stored| stored.value() != expected.as_slice())
+                        {
+                            complete = false;
+                            break;
+                        }
+                    }
+                }
+                complete
+            }
+            Err(redb::TableError::TableDoesNotExist(_)) => false,
+            Err(error) => return Err(error.into()),
+        };
+        drop(events);
+        drop(meta);
+        drop(read);
+        if !complete {
+            let tx = self.write_transaction()?;
+            rebuild_from_authority(&tx, schema, &binding, applied)?;
+            self.commit(tx)?;
+        }
+        Ok(())
+    }
+    /// Read a bounded page without changing any delivery ACK or cursor state.
+    pub fn created_events_page(&self, cursor: Option<&EventCursor>, limit: usize) -> Result<EventPage, StoreError> {
+        if !self.writable {
+            return Err(StoreError::WriteDisabled);
+        }
+        if limit == 0 || limit > MAX_EVENT_PAGE {
+            return Err(StoreError::InvalidEventPage);
+        }
+        let tx = self.database.begin_read()?;
+        let meta = tx.open_table(crate::META)?;
+        let schema = crate::number(meta.get("schema")?.ok_or(StoreError::CorruptRecord)?.value())?;
+        if !matches!(schema, 3..=7) {
+            return Err(StoreError::UnsupportedSchema);
+        }
+        let identity = crate::decode_identity(meta.get("identity")?.ok_or(StoreError::CorruptRecord)?.value())?;
+        let events = tx.open_table(application::EVENTS)?;
+        let index = tx.open_table(EVENT_SEQUENCE)?;
+        if index.len()? != events.len()? {
+            return Err(StoreError::CorruptRecord);
+        }
+        let read_event = |id: u64, bytes: &[u8]| -> Result<SourceEvent, StoreError> {
+            let (log, event_id) = decode_index(bytes)?;
+            let value = events.get(event_id.as_slice())?.ok_or(StoreError::CorruptRecord)?;
+            let captured = application::decode_event(value.value())?;
+            if log.index != id || captured.decision_log_id != log {
+                return Err(StoreError::CorruptRecord);
+            }
+            Ok(SourceEvent { event_id, captured })
+        };
+        if let Some(cursor) = cursor {
+            let stored = index.get(cursor.index)?.ok_or(StoreError::IdentityMismatch)?;
+            let event = read_event(cursor.index, stored.value())?;
+            if !cursor.matches(identity.cluster_id, event.event_id, &event.captured) {
+                return Err(StoreError::IdentityMismatch);
+            }
+        }
+        let start = cursor.map_or(0, |c| c.index);
+        let range = if cursor.is_some() {
+            (std::ops::Bound::Excluded(start), std::ops::Bound::Unbounded)
+        } else {
+            (std::ops::Bound::Included(start), std::ops::Bound::Unbounded)
+        };
+        let mut page = EventPage {
+            events: Vec::with_capacity(limit),
+            next: cursor.cloned(),
+        };
+        for row in index.range(range)?.take(limit) {
+            let (id, bytes) = row?;
+            let event = read_event(id.value(), bytes.value())?;
+            page.next = Some(EventCursor::new(identity.cluster_id, event.event_id, &event.captured));
+            page.events.push(event);
+        }
+        Ok(page)
+    }
+}
+impl NativeLogWriter {
+    pub async fn created_events_page(&self, cursor: Option<EventCursor>, limit: usize) -> Result<EventPage, NativeWriteError> {
+        let owner = self.store.clone();
+        tokio::task::spawn_blocking(move || {
+            owner
+                .lock()
+                .map_err(|_| NativeWriteError::Poisoned)?
+                .created_events_page(cursor.as_ref(), limit)
+                .map_err(NativeWriteError::Store)
+        })
+        .await
+        .map_err(NativeWriteError::Worker)?
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::{CaptureCommand, CaptureRaftEntry, DecideCreated, NodeIdentity, ObjectIdentity, PreparedIdentity, snapshot};
+    use std::io;
+    fn identity() -> NodeIdentity {
+        NodeIdentity {
+            cluster_id: [7; 16],
+            node_id: 0,
+        }
+    }
+    fn binding() -> CaptureBinding {
+        CaptureBinding {
+            revision: 9,
+            binding_id: [8; 16],
+            target_id: [9; 16],
+        }
+    }
+    fn id(index: u64) -> LogId {
+        LogId {
+            term: 8,
+            leader_node: 0,
+            index,
+        }
+    }
+    fn command(event: u8) -> DecideCreated {
+        DecideCreated {
+            operation_id: [event; 16],
+            object: ObjectIdentity {
+                bucket_incarnation: [2; 16],
+                key: vec![event],
+            },
+            expected_head: None,
+            prepared: PreparedIdentity {
+                version_id: [event; 16],
+                preparation_id: [4; 16],
+                content_digest: [5; 32],
+                content_length: 123,
+            },
+            binding_revision: 9,
+        }
+    }
+    fn apply(store: &mut Store, index: u64, command: DecideCreated) -> CreatedResult {
+        store
+            .append_native(&[CaptureRaftEntry {
+                log_id: id(index).into(),
+                payload: openraft::EntryPayload::Normal(CaptureCommand::try_created(command).expect("command")),
+            }])
+            .expect("append native");
+        store.save_committed(id(index)).expect("commit native");
+        store
+            .apply_next_native(id(index))
+            .expect("apply native")
+            .created_result()
+            .expect("created response")
+            .clone()
+    }
+    fn fixture() -> (tempfile::TempDir, Store) {
+        let dir = tempfile::tempdir().expect("directory");
+        let mut store = Store::initialize(dir.path().join("source.redb"), identity()).expect("initialize");
+        store.pin_capture_binding(binding()).expect("pin");
+        store.enable_native_empty().expect("native");
+        apply(&mut store, 0, command(9));
+        apply(&mut store, 1, command(7));
+        (dir, store)
+    }
+    #[test]
+    fn committed_events_page_in_decision_order() {
+        let (_dir, store) = fixture();
+        let first = store.created_events_page(None, 1).expect("first");
+        assert_eq!(
+            first.events.iter().map(|e| e.event_id).collect::<Vec<_>>(),
+            vec![[9; 16]],
+            "SOURCE_PAGE_ORDER_BOUND"
+        );
+        let next = store.created_events_page(first.next.as_ref(), 1).expect("second");
+        assert_eq!(next.events[0].event_id, [7; 16]);
+        assert_eq!(next.events[0].captured.binding, binding());
+        assert!(
+            store
+                .created_events_page(next.next.as_ref(), 1)
+                .expect("end")
+                .events
+                .is_empty()
+        );
+    }
+    #[test]
+    fn replay_and_rejections_do_not_duplicate_events() {
+        let (_dir, mut store) = fixture();
+        apply(&mut store, 2, command(9));
+        let mut rejected = command(3);
+        rejected.object = command(9).object;
+        assert!(matches!(apply(&mut store, 3, rejected), CreatedResult::HeadMismatch { .. }));
+        let page = store.created_events_page(None, 256).expect("page");
+        assert_eq!(page.events.len(), 2, "SOURCE_PAGE_REPLAY_BOUND");
+    }
+    #[test]
+    fn reopen_rebuilds_missing_or_stale_index() {
+        let (dir, mut store) = fixture();
+        apply(&mut store, 2, command(1));
+        let expected = store.created_events_page(None, 256).expect("before");
+        let tx = store.write_transaction().expect("remove cache");
+        tx.delete_table(EVENT_SEQUENCE).expect("delete cache");
+        store.commit(tx).expect("cache deletion");
+        drop(store);
+        let mut store = Store::open(dir.path().join("source.redb"), identity()).expect("rebuild missing");
+        assert_eq!(store.created_events_page(None, 256).expect("rebuilt"), expected);
+        let tx = store.write_transaction().expect("stale cache");
+        {
+            let mut index = tx.open_table(EVENT_SEQUENCE).expect("cache");
+            index.remove(1).expect("omit middle item");
+            index.insert(200, b"bad".as_slice()).expect("same-count stale item");
+        }
+        store.commit(tx).expect("cache damage");
+        drop(store);
+        let store = Store::open(dir.path().join("source.redb"), identity()).expect("rebuild stale");
+        assert_eq!(store.created_events_page(None, 256).expect("rebuilt stale"), expected);
+    }
+    #[test]
+    fn cursor_binds_cluster_term_leader_and_event() {
+        let (_dir, store) = fixture();
+        let cursor = store.created_events_page(None, 1).expect("page").next.expect("cursor");
+        let mut bad = Vec::new();
+        let mut c = cursor.clone();
+        c.cluster = [6; 16];
+        bad.push(c);
+        let mut c = cursor.clone();
+        c.term += 1;
+        bad.push(c);
+        let mut c = cursor.clone();
+        c.leader += 1;
+        bad.push(c);
+        let mut c = cursor.clone();
+        c.event = [7; 16];
+        bad.push(c);
+        let mut c = cursor.clone();
+        c.revision += 1;
+        bad.push(c);
+        let mut c = cursor.clone();
+        c.binding_id = [6; 16];
+        bad.push(c);
+        let mut c = cursor.clone();
+        c.target_id = [6; 16];
+        bad.push(c);
+        for cursor in bad {
+            assert_eq!(
+                store.created_events_page(Some(&cursor), 1).err().map(|e| e.code()),
+                Some("identity_mismatch"),
+                "SOURCE_PAGE_CURSOR_BOUND"
+            );
+        }
+        let encoded = serde_json::to_vec(&cursor).expect("cursor bytes");
+        let decoded: EventCursor = serde_json::from_slice(&encoded).expect("cursor decode");
+        assert_eq!(store.created_events_page(Some(&decoded), 1).expect("resume").events[0].event_id, [7; 16]);
+    }
+    #[test]
+    fn page_is_bounded_and_new_lower_uuid_is_not_lost() {
+        let (_dir, mut store) = fixture();
+        let page = store.created_events_page(None, 1).expect("one");
+        assert_eq!(page.events.len(), 1);
+        apply(&mut store, 2, command(1));
+        let next = store.created_events_page(page.next.as_ref(), 256).expect("later");
+        assert_eq!(next.events.iter().map(|e| e.event_id).collect::<Vec<_>>(), vec![[7; 16], [1; 16]]);
+        for limit in [0, MAX_EVENT_PAGE + 1, usize::MAX] {
+            assert_eq!(store.created_events_page(None, limit).err().map(|e| e.code()), Some("invalid_event_page"));
+        }
+        store.writable = false;
+        assert_eq!(store.created_events_page(None, 1).err().map(|e| e.code()), Some("write_disabled"));
+        store.writable = true;
+        let mut max = page.next.expect("cursor");
+        max.index = u64::MAX;
+        assert_eq!(
+            store.created_events_page(Some(&max), 1).err().map(|e| e.code()),
+            Some("identity_mismatch")
+        );
+    }
+    struct Staging<'a>(&'a WriteTransaction);
+    impl snapshot::StagingSink for Staging<'_> {
+        fn row(&mut self, section: snapshot::Section, key: &[u8], value: &[u8]) -> io::Result<()> {
+            let table = match section {
+                snapshot::Section::Heads => application::HEADS,
+                snapshot::Section::Versions => application::VERSIONS,
+                snapshot::Section::Events => application::EVENTS,
+                snapshot::Section::Operations => application::OPERATIONS,
+            };
+            self.0
+                .open_table(table)
+                .map_err(io::Error::other)?
+                .insert(key, value)
+                .map_err(io::Error::other)?;
+            Ok(())
+        }
+    }
+    fn limits() -> snapshot::Limits {
+        snapshot::Limits {
+            total_bytes: 1_048_576,
+            membership_bytes: 4096,
+            key_bytes: 4096,
+            value_bytes: 8192,
+            records: 1024,
+        }
+    }
+    async fn artifact(store: Store) -> (snapshot::CapabilityArtifact<Vec<u8>>, EventPage) {
+        let writer = NativeLogWriter::new(store);
+        let before = writer.created_events_page(None, 256).await.expect("async page");
+        let artifact = writer
+            .snapshot_view()
+            .await
+            .expect("snapshot view")
+            .stream_capability_artifact(Vec::new(), limits())
+            .expect("capability export");
+        (artifact, before)
+    }
+    fn stage(artifact: snapshot::CapabilityArtifact<Vec<u8>>, damage: bool) -> Result<(tempfile::TempDir, Store), StoreError> {
+        let dir = tempfile::tempdir().expect("stage directory");
+        let database = redb::Database::create(dir.path().join("stage.redb")).expect("isolated database");
+        let tx = database.begin_write()?;
+        let header = snapshot::decode_capability_artifact(
+            snapshot::CapabilityArtifact {
+                manifest: artifact.manifest,
+                payload: std::io::Cursor::new(artifact.payload),
+            },
+            &mut Staging(&tx),
+            limits(),
+        )
+        .expect("real capability decoder");
+        if damage {
+            let mut events = tx.open_table(application::EVENTS)?;
+            let value = events.get([9; 16].as_slice())?.expect("event").value().to_vec();
+            events.insert([6; 16].as_slice(), value.as_slice())?;
+            events.remove([9; 16].as_slice())?;
+        }
+        rebuild_from_authority(&tx, 6, &header.binding, header.applied)?;
+        {
+            let mut meta = tx.open_table(crate::META)?;
+            meta.insert("schema", 6u64.to_be_bytes().as_slice())?;
+            meta.insert(
+                "identity",
+                crate::encode_identity(NodeIdentity {
+                    cluster_id: header.cluster_id,
+                    ..identity()
+                })
+                .as_slice(),
+            )?;
+            meta.insert("capture_binding", crate::created::binding_bytes(&header.binding)?.as_slice())?;
+            let mut applied = vec![1];
+            applied.extend(encode_id(header.applied.expect("cut")));
+            meta.insert("last_applied", applied.as_slice())?;
+        }
+        tx.commit().expect("stage commit");
+        // A test-only staging handle, not a production installer or authority grant.
+        Ok((
+            dir,
+            Store {
+                database,
+                writable: true,
+            },
+        ))
+    }
+    #[tokio::test]
+    async fn snapshot_authority_rebuild_matches_source() {
+        let (_dir, mut store) = fixture();
+        apply(&mut store, 2, command(9));
+        let mut rejected = command(3);
+        rejected.object = command(9).object;
+        apply(&mut store, 3, rejected);
+        store.promote_created_provenance().expect("schema6");
+        let (artifact, before) = artifact(store).await;
+        let (_stage, restored) = stage(artifact, false).expect("staged authority index");
+        assert_eq!(
+            restored.created_events_page(None, 256).expect("restored page"),
+            before,
+            "SOURCE_PAGE_SNAPSHOT_BOUND"
+        );
+        let first = restored.created_events_page(None, 1).expect("first");
+        assert_eq!(
+            restored.created_events_page(first.next.as_ref(), 1).expect("resume").events[0],
+            before.events[1]
+        );
+    }
+    #[tokio::test]
+    async fn snapshot_event_operation_mismatch_rejected() {
+        let (_dir, mut store) = fixture();
+        store.promote_created_provenance().expect("schema6");
+        let (artifact, _) = artifact(store).await;
+        assert_eq!(
+            stage(artifact, true).err().map(|e| e.code()),
+            Some("corrupt_record"),
+            "SOURCE_PAGE_SNAPSHOT_BOUND"
+        );
+    }
+}
@@ -0,0 +1,451 @@
+//! Same-UID local reads from an owner's live native store. No ACK or cursor writes.
+use crate::{EventCursor, EventPage, NativeLogWriter, NativeWriteError, NodeIdentity, StoreError};
+use redb::ReadableDatabase;
+use serde::{Deserialize, Serialize};
+use std::{
+    fs, io,
+    os::unix::fs::{FileTypeExt, MetadataExt, PermissionsExt},
+    path::{Path, PathBuf},
+    time::Duration,
+};
+use tokio::{
+    io::{AsyncReadExt, AsyncWriteExt},
+    net::{UnixListener, UnixStream},
+    sync::{OwnedSemaphorePermit, watch},
+    task::JoinSet,
+};
+
+pub const WIRE_VERSION: u16 = 1;
+pub const MAX_REQUEST_BYTES: usize = 4096;
+pub const MAX_RESPONSE_BYTES: usize = 8 * 1024 * 1024;
+pub const MAX_CONNECTIONS: usize = 4;
+pub const IO_TIMEOUT: Duration = Duration::from_secs(5);
+
+#[derive(Debug, Clone, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct ReadPage {
+    pub version: u16,
+    pub expected_cluster: [u8; 16],
+    pub cursor: Option<EventCursor>,
+    pub limit: usize,
+}
+#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub struct ReadPageResponse {
+    pub version: u16,
+    pub identity: NodeIdentity,
+    pub page: EventPage,
+}
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+pub enum Rejection {
+    InvalidRequest,
+    WrongCluster,
+    InvalidCursor,
+    StoreUnavailable,
+    ResponseLimit,
+}
+#[derive(Debug, Serialize, Deserialize)]
+#[serde(tag = "status", deny_unknown_fields)]
+pub enum ReadPageReply {
+    Page { response: ReadPageResponse },
+    Rejected { version: u16, error: Rejection },
+}
+#[derive(Debug, thiserror::Error)]
+pub enum ReaderError {
+    #[error("SOURCE_READER_IO")]
+    Io(#[from] io::Error),
+    #[error("SOURCE_READER_PERMISSION")]
+    Permission,
+    #[error("SOURCE_READER_PATH_OCCUPIED")]
+    Occupied,
+    #[error("SOURCE_READER_PATH_CHANGED")]
+    PathChanged,
+    #[error("SOURCE_READER_TIMEOUT")]
+    Timeout,
+    #[error("SOURCE_READER_PROTOCOL")]
+    Protocol,
+    #[error("SOURCE_READER_REJECTED: {0:?}")]
+    Rejected(Rejection),
+    #[error("SOURCE_READER_WORKER")]
+    Worker(#[source] tokio::task::JoinError),
+}
+
+impl crate::Store {
+    fn durable_source_identity(&self) -> Result<NodeIdentity, StoreError> {
+        let tx = self.database.begin_read()?;
+        let meta = tx.open_table(crate::META)?;
+        let identity = meta.get("identity")?.ok_or(StoreError::CorruptRecord)?;
+        crate::decode_identity(identity.value())
+    }
+}
+impl NativeLogWriter {
+    /// Read the validated durable identity, never an owner-supplied cluster label.
+    pub async fn durable_identity(&self) -> Result<NodeIdentity, NativeWriteError> {
+        let owner = self.store.clone();
+        tokio::task::spawn_blocking(move || {
+            owner
+                .lock()
+                .map_err(|_| NativeWriteError::Poisoned)?
+                .durable_source_identity()
+                .map_err(NativeWriteError::Store)
+        })
+        .await
+        .map_err(NativeWriteError::Worker)?
+    }
+}
+
+fn uid() -> u32 {
+    rustix::process::geteuid().as_raw()
+}
+fn fingerprint(meta: &fs::Metadata) -> (u64, u64) {
+    (meta.dev(), meta.ino())
+}
+fn private_parent(path: &Path) -> Result<(PathBuf, (u64, u64)), ReaderError> {
+    if !path.is_absolute() || path.file_name().is_none() {
+        return Err(ReaderError::Permission);
+    }
+    let parent = path.parent().ok_or(ReaderError::Permission)?;
+    let meta = fs::symlink_metadata(parent)?;
+    if !meta.is_dir() || meta.uid() != uid() || meta.mode() & 0o077 != 0 {
+        return Err(ReaderError::Permission);
+    }
+    let resolved = fs::canonicalize(parent)?;
+    let resolved_meta = fs::symlink_metadata(&resolved)?;
+    if fingerprint(&meta) != fingerprint(&resolved_meta) {
+        return Err(ReaderError::PathChanged);
+    }
+    Ok((resolved.join(path.file_name().ok_or(ReaderError::Permission)?), fingerprint(&meta)))
+}
+fn owned_socket(path: &Path) -> Result<fs::Metadata, ReaderError> {
+    let meta = fs::symlink_metadata(path)?;
+    if !meta.file_type().is_socket() || meta.uid() != uid() {
+        return Err(ReaderError::Permission);
+    }
+    Ok(meta)
+}
+fn socket_metadata(path: &Path) -> Result<(u64, u64), ReaderError> {
+    let meta = owned_socket(path)?;
+    if meta.mode() & 0o077 != 0 {
+        return Err(ReaderError::Permission);
+    }
+    Ok(fingerprint(&meta))
+}
+fn peer(stream: &UnixStream) -> Result<(), ReaderError> {
+    if stream.peer_cred()?.uid() != uid() {
+        return Err(ReaderError::Permission);
+    }
+    Ok(())
+}
+struct SocketPath {
+    path: PathBuf,
+    inode: Option<(u64, u64)>,
+    parent: (u64, u64),
+}
+impl SocketPath {
+    fn remove(&mut self) -> Result<(), ReaderError> {
+        // One cleanup attempt only: never let Drop re-evaluate a later namespace.
+        let Some(inode) = self.inode.take() else {
+            return Ok(());
+        };
+        let (_, parent) = private_parent(&self.path)?;
+        if parent != self.parent || fingerprint(&owned_socket(&self.path)?) != inode {
+            return Err(ReaderError::PathChanged);
+        }
+        fs::remove_file(&self.path)?;
+        Ok(())
+    }
+}
+impl Drop for SocketPath {
+    fn drop(&mut self) {
+        // Explicit shutdown reports cleanup failure; cancellation preserves detected replacements.
+        let _ = self.remove();
+    }
+}
+
+/// Signal shutdown, then await the owner's `serve` task to drain workers and clean up.
+#[derive(Clone)]
+pub struct Shutdown(watch::Sender<bool>);
+impl Shutdown {
+    pub fn shutdown(&self) {
+        let _ = self.0.send(true);
+    }
+}
+pub struct SourceReaderServer {
+    listener: UnixListener,
+    writer: NativeLogWriter,
+    socket: SocketPath,
+    shutdown: watch::Receiver<bool>,
+}
+impl SourceReaderServer {
+    pub fn bind(path: impl AsRef<Path>, writer: NativeLogWriter) -> Result<(Self, Shutdown), ReaderError> {
+        let (path, parent) = private_parent(path.as_ref())?;
+        match fs::symlink_metadata(&path) {
+            Ok(_) => return Err(ReaderError::Occupied),
+            Err(error) if error.kind() == io::ErrorKind::NotFound => {}
+            Err(error) => return Err(error.into()),
+        }
+        // The private parent prevents a foreign UID from observing the initial socket mode.
+        let listener = UnixListener::bind(&path)?;
+        let inode = fingerprint(&fs::symlink_metadata(&path)?);
+        let socket = SocketPath {
+            path,
+            inode: Some(inode),
+            parent,
+        };
+        fs::set_permissions(&socket.path, fs::Permissions::from_mode(0o600))?;
+        if private_parent(&socket.path)?.1 != parent || socket_metadata(&socket.path)? != inode {
+            return Err(ReaderError::PathChanged);
+        }
+        let (sender, shutdown) = watch::channel(false);
+        Ok((
+            Self {
+                listener,
+                writer,
+                socket,
+                shutdown,
+            },
+            Shutdown(sender),
+        ))
+    }
+    pub async fn serve(mut self) -> Result<(), ReaderError> {
+        let mut tasks = JoinSet::new();
+        let outcome = async {
+            loop {
+                if *self.shutdown.borrow() {
+                    break;
+                }
+                // Admission is shared by every clone of this writer, including re-bound services.
+                let admission = self.writer.source_reader_admission.clone();
+                let permit = tokio::select! {
+                    result = admission.acquire_owned() => result.map_err(|_| ReaderError::Protocol)?,
+                    _ = self.shutdown.changed() => break,
+                };
+                let stream = tokio::select! {
+                    result = self.listener.accept() => result?.0,
+                    _ = self.shutdown.changed() => break,
+                };
+                if peer(&stream).is_err() {
+                    continue;
+                }
+                let writer = self.writer.clone();
+                tasks.spawn(async move {
+                    let _ = connection(stream, writer, permit).await;
+                });
+                while let Some(result) = tasks.try_join_next() {
+                    result.map_err(ReaderError::Worker)?;
+                }
+            }
+            Ok(())
+        }
+        .await;
+        // Ordinary shutdown drains connections and blocking workers, never aborts storage work.
+        while let Some(result) = tasks.join_next().await {
+            result.map_err(ReaderError::Worker)?;
+        }
+        drop(self.listener);
+        self.socket.remove()?;
+        outcome
+    }
+}
+
+fn reject(error: Rejection) -> ReadPageReply {
+    ReadPageReply::Rejected {
+        version: WIRE_VERSION,
+        error,
+    }
+}
+fn read_page(writer: &NativeLogWriter, request: ReadPage) -> ReadPageReply {
+    if request.version != WIRE_VERSION
+        || request.expected_cluster == [0; 16]
+        || request.limit == 0
+        || request.limit > crate::MAX_EVENT_PAGE
+    {
+        return reject(Rejection::InvalidRequest);
+    }
+    // Lock order: connection admission, then Store. Native reads take no writer-ordering lock.
+    let read = || -> Result<ReadPageReply, NativeWriteError> {
+        let store = writer.store.lock().map_err(|_| NativeWriteError::Poisoned)?;
+        let identity = store.durable_source_identity()?;
+        if identity.cluster_id != request.expected_cluster {
+            return Ok(reject(Rejection::WrongCluster));
+        }
+        let page = match store.created_events_page(request.cursor.as_ref(), request.limit) {
+            Ok(page) => page,
+            Err(StoreError::IdentityMismatch) => return Ok(reject(Rejection::InvalidCursor)),
+            Err(error) => return Err(error.into()),
+        };
+        Ok(ReadPageReply::Page {
+            response: ReadPageResponse {
+                version: WIRE_VERSION,
+                identity,
+                page,
+            },
+        })
+    };
+    read().unwrap_or_else(|_| reject(Rejection::StoreUnavailable))
+}
+
+struct BoundedBytes {
+    bytes: Vec<u8>,
+    limit: usize,
+}
+impl io::Write for BoundedBytes {
+    fn write(&mut self, bytes: &[u8]) -> io::Result<usize> {
+        if bytes.len() > self.limit.saturating_sub(self.bytes.len()) {
+            return Err(io::Error::other("source reader frame limit"));
+        }
+        self.bytes.extend_from_slice(bytes);
+        Ok(bytes.len())
+    }
+    fn flush(&mut self) -> io::Result<()> {
+        Ok(())
+    }
+}
+fn encode(value: &impl Serialize, limit: usize) -> Result<Vec<u8>, ReaderError> {
+    let mut buffer = BoundedBytes {
+        bytes: Vec::new(),
+        limit,
+    };
+    serde_json::to_writer(&mut buffer, value).map_err(|_| ReaderError::Protocol)?;
+    Ok(buffer.bytes)
+}
+async fn frame_read(stream: &mut UnixStream, limit: usize) -> Result<Vec<u8>, ReaderError> {
+    let len = usize::try_from(stream.read_u32().await?).map_err(|_| ReaderError::Protocol)?;
+    if len == 0 || len > limit {
+        return Err(ReaderError::Protocol);
+    }
+    let mut bytes = vec![0; len];
+    stream.read_exact(&mut bytes).await?;
+    Ok(bytes)
+}
+async fn frame_write(stream: &mut UnixStream, bytes: &[u8]) -> Result<(), ReaderError> {
+    stream
+        .write_u32(u32::try_from(bytes.len()).map_err(|_| ReaderError::Protocol)?)
+        .await?;
+    stream.write_all(bytes).await?;
+    Ok(())
+}
+async fn connection(mut stream: UnixStream, writer: NativeLogWriter, permit: OwnedSemaphorePermit) -> Result<(), ReaderError> {
+    let bytes = tokio::time::timeout(IO_TIMEOUT, frame_read(&mut stream, MAX_REQUEST_BYTES))
+        .await
+        .map_err(|_| ReaderError::Timeout)??;
+    let request: ReadPage = match serde_json::from_slice(&bytes) {
+        Ok(request) => request,
+        Err(_) => {
+            let bytes = encode(&reject(Rejection::InvalidRequest), MAX_RESPONSE_BYTES)?;
+            return tokio::time::timeout(IO_TIMEOUT, frame_write(&mut stream, &bytes))
+                .await
+                .map_err(|_| ReaderError::Timeout)?;
+        }
+    };
+    // The blocking worker owns admission until read + bounded serialization finish, even
+    // if a deadline, shutdown or a dropped serve task cancels its awaiting connection.
+    let (bytes, _permit) = tokio::task::spawn_blocking(move || {
+        let reply = read_page(&writer, request);
+        let bytes = encode(&reply, MAX_RESPONSE_BYTES).or_else(|_| encode(&reject(Rejection::ResponseLimit), MAX_RESPONSE_BYTES));
+        (bytes, permit)
+    })
+    .await
+    .map_err(ReaderError::Worker)?;
+    tokio::time::timeout(IO_TIMEOUT, frame_write(&mut stream, &bytes?))
+        .await
+        .map_err(|_| ReaderError::Timeout)?
+}
+
+/// Connect only to a private socket owned by the effective UID, verified at both ends.
+#[derive(Clone)]
+pub struct SourceReaderClient {
+    path: PathBuf,
+}
+impl SourceReaderClient {
+    pub fn new(path: impl AsRef<Path>) -> Result<Self, ReaderError> {
+        let (path, _) = private_parent(path.as_ref())?;
+        socket_metadata(&path)?;
+        Ok(Self { path })
+    }
+    pub async fn page(
+        &self,
+        expected_cluster: [u8; 16],
+        cursor: Option<EventCursor>,
+        limit: usize,
+    ) -> Result<ReadPageResponse, ReaderError> {
+        let request = ReadPage {
+            version: WIRE_VERSION,
+            expected_cluster,
+            cursor,
+            limit,
+        };
+        tokio::time::timeout(IO_TIMEOUT, self.page_inner(request))
+            .await
+            .map_err(|_| ReaderError::Timeout)?
+    }
+    async fn page_inner(&self, request: ReadPage) -> Result<ReadPageResponse, ReaderError> {
+        let (_, parent) = private_parent(&self.path)?;
+        let inode = socket_metadata(&self.path)?;
+        let mut stream = UnixStream::connect(&self.path).await?;
+        peer(&stream)?;
+        if private_parent(&self.path)?.1 != parent || socket_metadata(&self.path)? != inode {
+            return Err(ReaderError::PathChanged);
+        }
+        frame_write(&mut stream, &encode(&request, MAX_REQUEST_BYTES)?).await?;
+        let bytes = frame_read(&mut stream, MAX_RESPONSE_BYTES).await?;
+        match serde_json::from_slice::<ReadPageReply>(&bytes).map_err(|_| ReaderError::Protocol)? {
+            ReadPageReply::Page { response } => {
+                if response.version != WIRE_VERSION
+                    || response.identity.cluster_id != request.expected_cluster
+                    || response.identity.cluster_id == [0; 16]
+                    || response.page.events.len() > request.limit
+                    || request.limit == 0
+                    || request.limit > crate::MAX_EVENT_PAGE
+                {
+                    return Err(ReaderError::Protocol);
+                }
+                let next = response
+                    .page
+                    .events
+                    .last()
+                    .map(|event| EventCursor::new(response.identity.cluster_id, event.event_id, &event.captured))
+                    .or(request.cursor);
+                if response.page.next != next {
+                    return Err(ReaderError::Protocol);
+                }
+                Ok(response)
+            }
+            ReadPageReply::Rejected { version, error } if version == WIRE_VERSION => Err(ReaderError::Rejected(error)),
+            _ => Err(ReaderError::Protocol),
+        }
+    }
+}
+
+#[cfg(test)]
+mod cleanup_tests {
+    use super::*;
+    #[test]
+    fn source_reader_cleanup_is_disarmed_before_rebind() {
+        let directory = tempfile::tempdir().expect("private socket owner");
+        fs::set_permissions(directory.path(), fs::Permissions::from_mode(0o700)).expect("private directory mode");
+        let (path, parent) = private_parent(&directory.path().join("reader.sock")).expect("private path");
+        let listener = std::os::unix::net::UnixListener::bind(&path).expect("first real socket");
+        fs::set_permissions(&path, fs::Permissions::from_mode(0o600)).expect("first socket mode");
+        let inode = socket_metadata(&path).expect("first socket identity");
+        let mut guard = SocketPath {
+            path: path.clone(),
+            inode: Some(inode),
+            parent,
+        };
+        drop(listener);
+        guard.remove().expect("explicit owner cleanup");
+        assert!(guard.inode.is_none(), "SOURCE_READER_CLEANUP_DISARMED");
+        let listener = std::os::unix::net::UnixListener::bind(&path).expect("immediate new owner bind");
+        fs::set_permissions(&path, fs::Permissions::from_mode(0o600)).expect("new socket mode");
+        let replacement = socket_metadata(&path).expect("new owner socket");
+        drop(guard);
+        assert_eq!(
+            socket_metadata(&path).expect("new socket preserved"),
+            replacement,
+            "SOURCE_READER_OLD_GUARD_PRESERVES_REBIND"
+        );
+        let _client = std::os::unix::net::UnixStream::connect(&path).expect("new service reachable");
+        let _connection = listener.accept().expect("new owner accepts");
+    }
+}
@@ -0,0 +1,502 @@
+//! Real NativeLogWriter and socket owner is this test process, not a production S3 owner.
+use crate::{source_reader::*, *};
+use std::{
+    collections::{BTreeMap, BTreeSet},
+    os::unix::fs::{MetadataExt, PermissionsExt},
+    path::PathBuf,
+};
+use tokio::{
+    io::{AsyncReadExt, AsyncWriteExt},
+    net::UnixStream,
+};
+fn identity() -> NodeIdentity {
+    NodeIdentity {
+        cluster_id: [7; 16],
+        node_id: 0,
+    }
+}
+fn binding() -> CaptureBinding {
+    CaptureBinding {
+        revision: 9,
+        binding_id: [8; 16],
+        target_id: [9; 16],
+    }
+}
+fn id(index: u64) -> LogId {
+    LogId {
+        term: 8,
+        leader_node: 0,
+        index,
+    }
+}
+fn command(event: u8) -> DecideCreated {
+    DecideCreated {
+        operation_id: [event; 16],
+        object: ObjectIdentity {
+            bucket_incarnation: [2; 16],
+            key: vec![event],
+        },
+        expected_head: None,
+        prepared: PreparedIdentity {
+            version_id: [event; 16],
+            preparation_id: [4; 16],
+            content_digest: [5; 32],
+            content_length: 123,
+        },
+        binding_revision: 9,
+    }
+}
+fn fixture() -> (tempfile::TempDir, NativeLogWriter) {
+    let dir = tempfile::tempdir().expect("real source reader directory");
+    std::fs::set_permissions(dir.path(), std::fs::Permissions::from_mode(0o700)).expect("private owner directory");
+    let mut store = Store::initialize(dir.path().join("db"), identity()).expect("initialize real redb");
+    store.pin_capture_binding(binding()).expect("binding");
+    store.enable_native_empty().expect("native schema");
+    store.promote_created_provenance().expect("provenance");
+    let membership = CaptureMembership::new(
+        Some(id(0)),
+        vec![BTreeSet::from([0])],
+        BTreeMap::from([(0, CaptureNode::new("http://127.0.0.1:7000".into()).expect("member"))]),
+    )
+    .expect("membership");
+    let membership = openraft::StoredMembership::try_from(&membership).expect("stored membership");
+    store
+        .append_native(&[CaptureRaftEntry {
+            log_id: id(0).into(),
+            payload: openraft::EntryPayload::Membership(membership.membership().clone()),
+        }])
+        .expect("membership log");
+    store.save_committed(id(0)).expect("membership commit");
+    store.apply_next_native(id(0)).expect("membership apply");
+    (dir, NativeLogWriter::new(store))
+}
+async fn apply(writer: &NativeLogWriter, index: u64, event: u8) {
+    let entry = CaptureRaftEntry {
+        log_id: id(index).into(),
+        payload: openraft::EntryPayload::Normal(CaptureCommand::try_created(command(event)).expect("canonical native command")),
+    };
+    writer
+        .append_native(vec![entry.clone()])
+        .await
+        .expect("live durable log append");
+    let response = writer.apply_committed(vec![entry]).await.expect("live committed apply");
+    assert!(
+        matches!(response[0].created_result(), Some(CreatedResult::Created { .. })),
+        "SOURCE_READER_REAL_APPLY"
+    );
+}
+fn start(
+    dir: &tempfile::TempDir,
+    writer: &NativeLogWriter,
+) -> (PathBuf, SourceReaderClient, Shutdown, tokio::task::JoinHandle<Result<(), ReaderError>>) {
+    let path = dir.path().join("reader.sock");
+    let (server, shutdown) = SourceReaderServer::bind(&path, writer.clone()).expect("owner socket bind");
+    let client = SourceReaderClient::new(&path).expect("trusted local client");
+    let task = tokio::spawn(server.serve());
+    (path, client, shutdown, task)
+}
+async fn stop(path: PathBuf, shutdown: Shutdown, task: tokio::task::JoinHandle<Result<(), ReaderError>>) {
+    shutdown.shutdown();
+    task.await.expect("serve task").expect("clean shutdown");
+    assert!(!path.exists(), "SOURCE_READER_SOCKET_CLEANUP");
+}
+async fn raw(path: &std::path::Path, value: serde_json::Value) -> ReadPageReply {
+    let mut stream = UnixStream::connect(path).await.expect("real socket connect");
+    let bytes = serde_json::to_vec(&value).expect("wire request");
+    stream
+        .write_u32(bytes.len().try_into().expect("request length"))
+        .await
+        .expect("frame prefix");
+    stream.write_all(&bytes).await.expect("frame bytes");
+    let len = stream.read_u32().await.expect("reply prefix");
+    let mut bytes = vec![0; usize::try_from(len).expect("reply length")];
+    stream.read_exact(&mut bytes).await.expect("reply bytes");
+    serde_json::from_slice(&bytes).expect("strict native reply")
+}
+fn request() -> serde_json::Value {
+    serde_json::to_value(ReadPage {
+        version: WIRE_VERSION,
+        expected_cluster: identity().cluster_id,
+        cursor: None,
+        limit: 2,
+    })
+    .expect("native request")
+}
+#[tokio::test]
+async fn source_reader_live_apply() {
+    let (dir, writer) = fixture();
+    let (path, client, shutdown, task) = start(&dir, &writer);
+    assert_eq!(writer.durable_identity().await.expect("durable META identity"), identity());
+    assert!(
+        client
+            .page(identity().cluster_id, None, 2)
+            .await
+            .expect("empty live page")
+            .page
+            .events
+            .is_empty()
+    );
+    apply(&writer, 1, 9).await;
+    let response = client
+        .page(identity().cluster_id, None, 2)
+        .await
+        .expect("live committed page");
+    assert_eq!(response.identity, identity(), "SOURCE_READER_DURABLE_IDENTITY");
+    assert_eq!(response.page.events[0].event_id, [9; 16], "SOURCE_READER_LIVE_EVENT");
+    assert_eq!(response.page.events[0].captured.binding, binding());
+    {
+        let mut corrupt = writer.store.lock().expect("test owner");
+        let tx = corrupt.write_transaction().expect("corrupt META transaction");
+        tx.open_table(META)
+            .expect("META")
+            .insert("identity", &[0; 24][..])
+            .expect("corrupt identity");
+        corrupt.commit(tx).expect("persist corruption");
+    }
+    assert!(writer.durable_identity().await.is_err(), "SOURCE_READER_META_CORRUPTION_REJECTED");
+    assert!(
+        matches!(
+            client.page(identity().cluster_id, None, 2).await,
+            Err(ReaderError::Rejected(Rejection::StoreUnavailable))
+        ),
+        "SOURCE_READER_NO_CALLER_IDENTITY_SUBSTITUTION"
+    );
+    stop(path, shutdown, task).await;
+}
+#[tokio::test]
+async fn source_reader_restart_cursor() {
+    let (dir, writer) = fixture();
+    apply(&writer, 1, 9).await;
+    apply(&writer, 2, 7).await;
+    let (path, client, shutdown, task) = start(&dir, &writer);
+    let first = client.page(identity().cluster_id, None, 1).await.expect("first");
+    stop(path, shutdown, task).await;
+    drop(writer);
+    let writer = NativeLogWriter::new(Store::open(dir.path().join("db"), identity()).expect("restart same durable store"));
+    let (path, client, shutdown, task) = start(&dir, &writer);
+    let next = client
+        .page(identity().cluster_id, first.page.next, 1)
+        .await
+        .expect("resume durable cursor");
+    assert_eq!(next.page.events[0].event_id, [7; 16], "SOURCE_READER_RESTART_CONTINUES");
+    stop(path, shutdown, task).await;
+}
+#[tokio::test]
+async fn source_reader_page_boundaries() {
+    let (dir, writer) = fixture();
+    for (index, event) in [(1, 9), (2, 7), (3, 1)] {
+        apply(&writer, index, event).await;
+    }
+    let (path, client, shutdown, task) = start(&dir, &writer);
+    let first = client.page(identity().cluster_id, None, 2).await.expect("first two");
+    assert_eq!(
+        first.page.events.iter().map(|e| e.event_id).collect::<Vec<_>>(),
+        vec![[9; 16], [7; 16]],
+        "SOURCE_READER_DECISION_ORDER"
+    );
+    let next = client
+        .page(identity().cluster_id, first.page.next, 2)
+        .await
+        .expect("remaining");
+    assert_eq!(next.page.events.len(), 1);
+    let end = client
+        .page(identity().cluster_id, next.page.next.clone(), 2)
+        .await
+        .expect("end");
+    assert!(end.page.events.is_empty());
+    assert_eq!(end.page.next, next.page.next, "SOURCE_READER_CURSOR_RETAINED_AT_END");
+    for limit in [0, MAX_EVENT_PAGE + 1] {
+        assert!(
+            matches!(
+                client.page(identity().cluster_id, None, limit).await,
+                Err(ReaderError::Rejected(Rejection::InvalidRequest))
+            ),
+            "SOURCE_READER_LIMIT_REJECTED"
+        );
+    }
+    stop(path, shutdown, task).await;
+}
+#[tokio::test]
+async fn source_reader_wrong_identity() {
+    let (dir, writer) = fixture();
+    apply(&writer, 1, 9).await;
+    let (path, client, shutdown, task) = start(&dir, &writer);
+    let mut wrong = request();
+    wrong["expected_cluster"] = serde_json::to_value([6; 16]).expect("wrong cluster");
+    assert!(
+        matches!(
+            raw(&path, wrong).await,
+            ReadPageReply::Rejected {
+                error: Rejection::WrongCluster,
+                ..
+            }
+        ),
+        "SOURCE_READER_CLUSTER: reject before reading or returning foreign events"
+    );
+    let first = client.page(identity().cluster_id, None, 1).await.expect("real cursor");
+    let mut wrong = serde_json::to_value(first.page.next.expect("cursor")).expect("cursor wire");
+    wrong["event"] = serde_json::to_value([1; 16]).expect("wrong event");
+    let wrong: EventCursor = serde_json::from_value(wrong).expect("syntactically valid foreign cursor");
+    assert!(
+        matches!(
+            client.page(identity().cluster_id, Some(wrong), 1).await,
+            Err(ReaderError::Rejected(Rejection::InvalidCursor))
+        ),
+        "SOURCE_READER_CURSOR_BOUND"
+    );
+    for (field, value) in [
+        ("version", serde_json::json!(2)),
+        ("unknown", serde_json::json!(true)),
+        ("cursor", serde_json::json!({"unknown":true})),
+    ] {
+        let mut wrong = request();
+        wrong[field] = value;
+        assert!(
+            matches!(
+                raw(&path, wrong).await,
+                ReadPageReply::Rejected {
+                    error: Rejection::InvalidRequest,
+                    ..
+                }
+            ),
+            "SOURCE_READER_STRICT_REQUEST"
+        );
+    }
+    stop(path, shutdown, task).await;
+}
+#[tokio::test]
+async fn source_reader_socket_ownership() {
+    let (dir, writer) = fixture();
+    let path = dir.path().join("reader.sock");
+    std::fs::set_permissions(dir.path(), std::fs::Permissions::from_mode(0o750)).expect("insecure directory");
+    assert!(
+        matches!(SourceReaderServer::bind(&path, writer.clone()), Err(ReaderError::Permission)),
+        "SOURCE_READER_PRIVATE_DIRECTORY"
+    );
+    std::fs::set_permissions(dir.path(), std::fs::Permissions::from_mode(0o700)).expect("restore private directory");
+    std::fs::write(&path, b"preserve me").expect("occupied file");
+    assert!(
+        matches!(SourceReaderServer::bind(&path, writer.clone()), Err(ReaderError::Occupied)),
+        "SOURCE_READER_OCCUPIED_PATH"
+    );
+    assert_eq!(std::fs::read(&path).expect("unchanged file"), b"preserve me");
+    std::fs::remove_file(&path).expect("remove own fixture");
+    std::os::unix::fs::symlink(dir.path().join("missing"), &path).expect("unknown symlink");
+    assert!(matches!(SourceReaderServer::bind(&path, writer.clone()), Err(ReaderError::Occupied)));
+    assert!(SourceReaderClient::new(&path).is_err());
+    std::fs::remove_file(&path).expect("remove own symlink");
+    let long = dir.path().join("x".repeat(300));
+    assert!(SourceReaderServer::bind(&long, writer.clone()).is_err(), "SOURCE_READER_LONG_SOCKET_PATH");
+    let parent_link = dir.path().join("parent-link");
+    std::os::unix::fs::symlink(dir.path(), &parent_link).expect("parent symlink");
+    assert!(
+        matches!(
+            SourceReaderServer::bind(parent_link.join("socket"), writer.clone()),
+            Err(ReaderError::Permission)
+        ),
+        "SOURCE_READER_PARENT_SYMLINK_DENIED"
+    );
+    let (path, client, shutdown, task) = start(&dir, &writer);
+    let meta = std::fs::symlink_metadata(&path).expect("socket metadata");
+    assert_eq!(meta.mode() & 0o777, 0o600);
+    assert_eq!(meta.uid(), rustix::process::geteuid().as_raw());
+    std::fs::set_permissions(&path, std::fs::Permissions::from_mode(0o660)).expect("insecure socket");
+    assert!(
+        matches!(client.page(identity().cluster_id, None, 1).await, Err(ReaderError::Permission)),
+        "SOURCE_READER_CLIENT_VALIDATES_SOCKET"
+    );
+    std::fs::set_permissions(&path, std::fs::Permissions::from_mode(0o600)).expect("restore socket");
+    std::fs::remove_file(&path).expect("test owner replaces socket");
+    std::fs::write(&path, b"replacement").expect("replacement file");
+    shutdown.shutdown();
+    assert!(task.await.expect("serve exit").is_err(), "SOURCE_READER_REPLACEMENT_CLEANUP_REJECTED");
+    assert_eq!(std::fs::read(&path).expect("replacement preserved"), b"replacement");
+}
+#[tokio::test]
+async fn source_reader_transport_bounds() {
+    let (dir, writer) = fixture();
+    let (path, client, shutdown, task) = start(&dir, &writer);
+    let mut stream = UnixStream::connect(&path).await.expect("oversized input peer");
+    stream
+        .write_u32(u32::try_from(MAX_REQUEST_BYTES + 1).expect("length"))
+        .await
+        .expect("overlong prefix");
+    let mut byte = [0];
+    assert_eq!(
+        stream.read(&mut byte).await.expect("bounded rejection closes"),
+        0,
+        "SOURCE_READER_INPUT_BOUND"
+    );
+    let mut idle = UnixStream::connect(&path).await.expect("idle connection");
+    let start = std::time::Instant::now();
+    assert_eq!(
+        idle.read(&mut byte).await.expect("deadline closes connection"),
+        0,
+        "SOURCE_READER_READ_DEADLINE"
+    );
+    assert!(start.elapsed() >= IO_TIMEOUT, "SOURCE_READER_REAL_READ_DEADLINE");
+    assert!(
+        client.page(identity().cluster_id, None, 1).await.is_ok(),
+        "SOURCE_READER_TIMEOUT_RELEASES_CONNECTION"
+    );
+    stop(path.clone(), shutdown, task).await;
+    // A same-UID peer still cannot supply unbounded responses to the client.
+    let listener = tokio::net::UnixListener::bind(&path).expect("controlled response peer");
+    std::fs::set_permissions(&path, std::fs::Permissions::from_mode(0o600)).expect("private response socket");
+    let client = SourceReaderClient::new(&path).expect("response client");
+    let peer = tokio::spawn(async move {
+        let (mut stream, _) = listener.accept().await.expect("client connects");
+        stream
+            .write_u32(u32::try_from(MAX_RESPONSE_BYTES + 1).expect("oversized response"))
+            .await
+            .expect("response prefix");
+        tokio::time::sleep(std::time::Duration::from_millis(50)).await;
+    });
+    assert!(
+        matches!(client.page(identity().cluster_id, None, 1).await, Err(ReaderError::Protocol)),
+        "SOURCE_READER_RESPONSE_BOUND"
+    );
+    peer.await.expect("peer task");
+}
+fn hold_store(writer: &NativeLogWriter) -> (std::sync::mpsc::Sender<()>, std::thread::JoinHandle<()>) {
+    let owner = writer.clone();
+    let (locked_tx, locked_rx) = std::sync::mpsc::channel();
+    let (release_tx, release_rx) = std::sync::mpsc::channel();
+    let thread = std::thread::spawn(move || {
+        let _store = owner.store.lock().expect("actual Store held by test owner");
+        locked_tx.send(()).expect("lock visible");
+        let _ = release_rx.recv();
+    });
+    locked_rx.recv().expect("store locked");
+    (release_tx, thread)
+}
+#[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+async fn source_reader_shutdown() {
+    let (dir, writer) = fixture();
+    apply(&writer, 1, 9).await;
+    let (path, _client, shutdown, task) = start(&dir, &writer);
+    let (release_tx, blocker) = hold_store(&writer);
+    let mut streams = Vec::new();
+    for _ in 0..MAX_CONNECTIONS {
+        let mut stream = UnixStream::connect(&path).await.expect("active read socket");
+        let bytes = serde_json::to_vec(&request()).expect("request");
+        stream
+            .write_u32(bytes.len().try_into().expect("length"))
+            .await
+            .expect("prefix");
+        stream.write_all(&bytes).await.expect("request bytes");
+        streams.push(stream);
+    }
+    tokio::time::timeout(std::time::Duration::from_secs(2), async {
+        while writer.source_reader_admission.available_permits() != 0 {
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .expect("all admissions active");
+    drop(streams);
+    tokio::time::sleep(IO_TIMEOUT + std::time::Duration::from_millis(100)).await;
+    assert_eq!(
+        writer.source_reader_admission.available_permits(),
+        0,
+        "SOURCE_READER_CLIENT_TIMEOUT_KEEPS_WORKER_ADMISSION"
+    );
+    shutdown.shutdown();
+    tokio::time::sleep(std::time::Duration::from_millis(100)).await;
+    assert!(!task.is_finished(), "SOURCE_READER_SHUTDOWN_DRAINS_READERS");
+    assert_eq!(
+        writer.source_reader_admission.available_permits(),
+        0,
+        "SOURCE_READER_CANCELLED_CLIENTS_KEEP_ADMISSION"
+    );
+    release_tx.send(()).expect("release real blocking reads");
+    blocker.join().expect("test lock owner exits");
+    task.await.expect("serve completes").expect("shutdown clean");
+    assert!(!path.exists());
+    assert_eq!(
+        writer.source_reader_admission.available_permits(),
+        MAX_CONNECTIONS,
+        "SOURCE_READER_ADMISSION_RELEASED_AFTER_WORKERS"
+    );
+    let (release, blocker) = hold_store(&writer);
+    let (path, _, _shutdown, task) = start(&dir, &writer);
+    let bytes = serde_json::to_vec(&request()).expect("native request");
+    let mut sockets = Vec::new();
+    for _ in 0..MAX_CONNECTIONS {
+        let mut stream = UnixStream::connect(&path).await.expect("cancelled service peer");
+        stream
+            .write_u32(bytes.len().try_into().expect("length"))
+            .await
+            .expect("prefix");
+        stream.write_all(&bytes).await.expect("request");
+        sockets.push(stream);
+    }
+    tokio::time::timeout(std::time::Duration::from_secs(2), async {
+        while writer.source_reader_admission.available_permits() != 0 {
+            tokio::task::yield_now().await;
+        }
+    })
+    .await
+    .expect("cancelled service admits bounded workers");
+    tokio::time::sleep(std::time::Duration::from_millis(100)).await;
+    task.abort();
+    assert!(task.await.expect_err("serve cancellation").is_cancelled());
+    drop(sockets);
+    assert_eq!(
+        writer.source_reader_admission.available_permits(),
+        0,
+        "SOURCE_READER_CANCELLED_SERVE_RETAINS_WORKER_ADMISSION"
+    );
+    let (path, client, shutdown, task) = start(&dir, &writer);
+    let pending = tokio::spawn(async move { client.page(identity().cluster_id, None, 1).await });
+    tokio::time::sleep(std::time::Duration::from_millis(100)).await;
+    assert!(!pending.is_finished(), "SOURCE_READER_REBIND_CANNOT_BYPASS_ADMISSION");
+    assert_eq!(writer.source_reader_admission.available_permits(), 0);
+    release.send(()).expect("release cancelled worker store lock");
+    blocker.join().expect("lock owner completes");
+    assert_eq!(
+        pending
+            .await
+            .expect("new service client")
+            .expect("new service waits for old workers")
+            .page
+            .events
+            .len(),
+        1
+    );
+    stop(path, shutdown, task).await;
+}
+#[tokio::test]
+async fn source_reader_checkpoint_replay() {
+    let (dir, writer) = fixture();
+    apply(&writer, 1, 9).await;
+    apply(&writer, 2, 7).await;
+    let (path, client, shutdown, task) = start(&dir, &writer);
+    let first = client
+        .page(identity().cluster_id, None, 1)
+        .await
+        .expect("pre-checkpoint cursor");
+    let checkpoint = writer
+        .checkpoint_local(CheckpointLimits::default())
+        .await
+        .expect("real checkpoint");
+    assert_eq!(checkpoint.cut(), id(2));
+    let next = client
+        .page(identity().cluster_id, first.page.next.clone(), 1)
+        .await
+        .expect("checkpoint retained events");
+    assert_eq!(next.page.events[0].event_id, [7; 16], "SOURCE_READER_CHECKPOINT_REPLAY");
+    stop(path, shutdown, task).await;
+    drop(writer);
+    let writer = NativeLogWriter::new(Store::open(dir.path().join("db"), identity()).expect("reopen checkpointed store"));
+    let (path, client, shutdown, task) = start(&dir, &writer);
+    assert_eq!(
+        client
+            .page(identity().cluster_id, first.page.next, 1)
+            .await
+            .expect("replay after restart"),
+        next,
+        "SOURCE_READER_CHECKPOINT_RESTART"
+    );
+    stop(path, shutdown, task).await;
+}
@@ -0,0 +1,47 @@
+use crate::{LOGS, META, Store, StoreError, decode_id, decode_log, encode_id};
+use redb::ReadableTable;
+
+impl Store {
+    /// Atomically remove every log entry at or after `from`.
+    pub fn truncate_suffix(&mut self, from: u64) -> Result<(), StoreError> {
+        let tx = self.write_transaction()?;
+        {
+            let mut meta = tx.open_table(META)?;
+            let logs = tx.open_table(LOGS)?;
+            if crate::committed::validate_committed(&meta, &logs)?.is_some_and(|id| from <= id.index) {
+                return Err(StoreError::CommittedLogProtected);
+            }
+            drop(logs);
+            let tail = meta.get("tail")?.map(|v| decode_id(v.value())).transpose()?;
+            if tail.is_none_or(|id| from > id.index) {
+                return Ok(());
+            }
+            let mut logs = tx.open_table(LOGS)?;
+            let new_tail = if from == 0 {
+                None
+            } else {
+                let index = from - 1;
+                let id = match logs.get(index)? {
+                    Some(record) => decode_log(record.value())?.id,
+                    None => crate::checkpoint_baseline::read_purged(&meta)?
+                        .filter(|id| id.index == index)
+                        .ok_or(StoreError::InvalidLogSequence)?,
+                };
+                if id.index != index {
+                    return Err(StoreError::CorruptRecord);
+                }
+                Some(id)
+            };
+            logs.retain(|index, _| index < from)?;
+            match new_tail {
+                Some(id) => {
+                    meta.insert("tail", encode_id(id).as_slice())?;
+                }
+                None => {
+                    meta.remove("tail")?;
+                }
+            }
+        }
+        self.commit(tx)
+    }
+}
@@ -18,10 +18,8 @@ use const_str::concat;
 /// Default value: RustFS
 /// Environment variable: RUSTFS_APP_NAME
 pub const APP_NAME: &str = "RustFS";
-/// Application version
-/// Default value: 1.0.0
-/// Environment variable: RUSTFS_VERSION
-pub const VERSION: &str = "1.0.0";
+/// Cargo package version for this workspace member.
+pub const VERSION: &str = env!("CARGO_PKG_VERSION");

 /// Default configuration logger level
 /// Default value: error
@@ -39,12 +37,6 @@ pub const SAMPLE_RATIO: f64 = 1.0;
 /// Default value: 30
 pub const METER_INTERVAL: u64 = 30;

-/// Default configuration service version
-/// Default value: 1.0.0
-/// Environment variable: RUSTFS_OBS_SERVICE_VERSION
-/// Uses the same value as VERSION constant
-pub const SERVICE_VERSION: &str = "1.0.0";
-
 /// Default configuration environment
 /// Default value: production
 pub const ENVIRONMENT: &str = "production";
@@ -410,7 +402,6 @@ mod tests {
            APP_NAME,
            VERSION,
            DEFAULT_LOG_LEVEL,
-            SERVICE_VERSION,
            ENVIRONMENT,
            RUSTFS_TLS_KEY,
            RUSTFS_TLS_CERT,
@@ -446,8 +437,7 @@ mod tests {
    fn test_configuration_consistency() {
        // Test configuration consistency

-        // Version consistency
-        assert_eq!(VERSION, SERVICE_VERSION, "Application version should match service version");
+        assert_eq!(VERSION, env!("CARGO_PKG_VERSION"));

        // Port conflict check
        let ports = [DEFAULT_PORT, DEFAULT_CONSOLE_PORT];
@@ -207,3 +207,9 @@ pub const ENV_TLS_RELOAD_INTERVAL: &str = "RUSTFS_TLS_RELOAD_INTERVAL";

 /// Default interval for TLS certificate reload check
 pub const DEFAULT_TLS_RELOAD_INTERVAL: u64 = 30;
+
+/// Environment variable to enable experimental HTTP/3 support.
+pub const ENV_HTTP3_ENABLE: &str = "RUSTFS_HTTP3_ENABLE";
+
+/// HTTP/3 is disabled unless explicitly enabled.
+pub const DEFAULT_HTTP3_ENABLE: bool = false;
@@ -33,6 +33,10 @@
 //! tracked separately.

 use crate::common::{ClusterTopology, RustFSTestClusterEnvironment};
+use crate::fault_proxy::FaultMode;
+use aws_sdk_s3::Client;
+use std::time::Duration;
+use tokio::time::{Instant, timeout};

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

@@ -144,7 +148,99 @@ async fn cluster_volume_fault_proxy_pass_smoke() -> TestResult {
        cluster.start().await?;
        cluster.create_test_bucket(BUCKET).await?;
        let payload = vec![0x6Du8; 256 * 1024];
-        put_get_roundtrip(&cluster, "volume-proxy/object", &payload).await
+        put_get_roundtrip(&cluster, "volume-proxy/object", &payload).await?;
+
+        // Node 1 owns only two disks; a successful three-vote write must
+        // authenticate at least one disk mutation through node 0's proxy.
+        let client = cluster.create_s3_client(1)?;
+        let client = Client::from_conf(
+            client
+                .config()
+                .to_builder()
+                .retry_config(aws_sdk_s3::config::retry::RetryConfig::standard().with_max_attempts(1))
+                .build(),
+        );
+        let forwarded_before = proxy.forwarded_bytes();
+        client
+            .put_object()
+            .bucket(BUCKET)
+            .key("volume-proxy/authenticated")
+            .body(bytes::Bytes::copy_from_slice(&payload).into())
+            .send()
+            .await?;
+        assert!(
+            proxy.forwarded_bytes() > forwarded_before,
+            "successful peer-disk publication must traverse the proxy"
+        );
+
+        let dropped_before = proxy.dropped_bytes();
+        proxy.set_mode(FaultMode::Blackhole);
+        assert_eq!(proxy.mode(), FaultMode::Blackhole);
+        let error = timeout(
+            Duration::from_secs(30),
+            client
+                .put_object()
+                .bucket(BUCKET)
+                .key("volume-proxy/rejected")
+                .body(bytes::Bytes::copy_from_slice(&payload).into())
+                .send(),
+        )
+        .await?
+        .expect_err("two local disks cannot commit a write while the peer is blackholed");
+        assert_eq!(error.raw_response().map(|response| response.status().as_u16()), Some(503));
+        assert_eq!(error.as_service_error().and_then(|error| error.meta().code()), Some("ServiceUnavailable"));
+        assert!(
+            proxy.dropped_bytes() > dropped_before,
+            "the injected blackhole must discard actual peer traffic"
+        );
+
+        proxy.set_mode(FaultMode::Pass);
+        assert_eq!(proxy.mode(), FaultMode::Pass);
+        let http = reqwest::Client::builder()
+            .no_proxy()
+            .timeout(Duration::from_secs(3))
+            .build()?;
+        let deadline = Instant::now() + Duration::from_secs(30);
+        loop {
+            let response = http.get(format!("{}/health/ready", cluster.nodes[1].url)).send().await?;
+            let status = response.status();
+            let body: serde_json::Value = response.json().await?;
+            // Reachable peers can still have Returning disk handles that reject
+            // writes. Wait for this node's live writer inventory to be Online.
+            if status.as_u16() == 200
+                && body["details"]["storage"]["source"] == "local_runtime"
+                && body["details"]["storage"]["writeQuorum"] == true
+                && body["details"]["storage"]["unavailableDrives"]
+                    .as_array()
+                    .is_some_and(Vec::is_empty)
+                && body["details"]["lock"]["ready"] == true
+            {
+                break;
+            }
+            assert!(Instant::now() < deadline, "peer write quorum did not recover: HTTP {status}, {body}");
+            tokio::time::sleep(Duration::from_millis(200)).await;
+        }
+        client
+            .put_object()
+            .bucket(BUCKET)
+            .key("volume-proxy/recovered")
+            .body(bytes::Bytes::copy_from_slice(&payload).into())
+            .send()
+            .await?;
+        for key in ["volume-proxy/object", "volume-proxy/authenticated", "volume-proxy/recovered"] {
+            let got = client.get_object().bucket(BUCKET).key(key).send().await?;
+            let body = got.body.collect().await?.into_bytes();
+            assert_eq!(body.as_ref(), payload, "proxy recovery must preserve the full object body for {key}");
+        }
+        let rejected = client
+            .head_object()
+            .bucket(BUCKET)
+            .key("volume-proxy/rejected")
+            .send()
+            .await
+            .expect_err("a rejected proxied write must never become visible after recovery");
+        assert_eq!(rejected.raw_response().map(|response| response.status().as_u16()), Some(404));
+        Ok(())
    }
    .await;

@@ -35,7 +35,7 @@ use s3s::Body;
 use serde_json;
 use std::fs as stdfs;
 use std::io::ErrorKind;
-use std::net::SocketAddr;
+use std::net::{Ipv6Addr, SocketAddr};
 use std::path::{Path, PathBuf};
 use std::process::{Child, Command, Stdio};
 use std::sync::Once;
@@ -1442,7 +1442,11 @@ impl RustFSTestClusterEnvironment {
            return Err(format!("a volume proxy is already configured for node {node_idx}").into());
        }
        let target = self.nodes[node_idx].address.parse::<SocketAddr>()?;
-        let proxy = crate::fault_proxy::FaultProxy::start(target).await?;
+        // Endpoint locality and RPC audience both use the advertised authority.
+        // Keep the node's port while the second loopback IP separates the proxy
+        // listener from the direct IPv4 server and S3 client address.
+        let listen = SocketAddr::from((Ipv6Addr::LOCALHOST, target.port()));
+        let proxy = crate::fault_proxy::FaultProxy::start_on(listen, target).await?;
        self.volume_proxy_addresses[node_idx] = Some(proxy.local_addr());
        Ok(proxy)
    }
@@ -2467,6 +2471,20 @@ mod tests {

        assert!(volumes.contains(&proxied), "volumes must use the proxy address: {volumes}");
        assert!(!volumes.contains(&direct), "volumes must not retain the direct address: {volumes}");
+        assert_eq!(proxy.local_addr().port(), proxy.target_addr().port());
+        assert_ne!(proxy.local_addr().ip(), proxy.target_addr().ip());
+        assert_eq!(env.nodes[0].address, direct, "S3 clients must retain the direct address");
+        let mut endpoint =
+            crate::storage_api::Endpoint::try_from(volumes.as_str()).expect("proxied volume should be a valid endpoint");
+        endpoint
+            .update_is_local(proxy.target_addr().port())
+            .expect("proxied endpoint locality should resolve");
+        assert!(endpoint.is_local, "the proxied node must keep its local disk and first-disk authority");
+        assert_eq!(
+            endpoint.host_port(),
+            proxied,
+            "the local RPC audience must be the advertised proxy authority"
+        );

        proxy.shutdown().await;
    }
@@ -31,6 +31,7 @@ use super::harness::{
 use crate::common::{
    AdminTransport, admin_add_canned_policy_via, admin_attach_user_policy_via, admin_create_user_via, init_logging,
 };
+use crate::upgrade_compatibility_test::wait_for_upgrade_write_readiness;
 use aws_sdk_s3::Client;
 use aws_sdk_s3::error::ProvideErrorMetadata;
 use std::path::{Path, PathBuf};
@@ -400,6 +401,19 @@ async fn four_node_direct_upgrade_preserves_history_and_iam_credentials() -> Tes
    Ok(())
 }

+async fn replace_node_and_wait_for_write_readiness(dist: &mut DistCluster, seed: &UpgradeSeed, node_idx: usize) -> TestResult {
+    dist.replace_node_with_current_binary(node_idx).await?;
+    // Node readiness proves reads can succeed; a rolling writer must also
+    // recover before replacing the next peer or checking upgrade contracts.
+    wait_for_upgrade_write_readiness(
+        &dist.clients()?,
+        &seed.history_bucket,
+        &format!("replaced-node-{node_idx}"),
+        CREDENTIAL_TIMEOUT,
+    )
+    .await
+}
+
 #[tokio::test]
 async fn four_node_rolling_upgrade_preserves_history_and_iam_credentials() -> TestResult {
    init_logging();
@@ -410,15 +424,15 @@ async fn four_node_rolling_upgrade_preserves_history_and_iam_credentials() -> Te

    let seed = seed_history_and_iam(&dist).await?;

-    dist.replace_node_with_current_binary(0).await?;
+    replace_node_and_wait_for_write_readiness(&mut dist, &seed, 0).await?;
    assert_history_and_iam(&dist, &seed, "one-current-node").await?;

    for node_idx in [1, 2] {
-        dist.replace_node_with_current_binary(node_idx).await?;
+        replace_node_and_wait_for_write_readiness(&mut dist, &seed, node_idx).await?;
    }
    assert_history_and_iam(&dist, &seed, "one-previous-node").await?;

-    dist.replace_node_with_current_binary(3).await?;
+    replace_node_and_wait_for_write_readiness(&mut dist, &seed, 3).await?;
    assert_history_and_iam(&dist, &seed, "homogeneous-current").await?;
    Ok(())
 }
@@ -50,8 +50,8 @@ use std::collections::{BTreeMap, HashMap, VecDeque};
 use std::error::Error;
 use std::io;
 use std::net::SocketAddr;
-use std::sync::{Arc, Mutex, MutexGuard};
-use std::time::{Duration, SystemTime};
+use std::sync::{Arc, Mutex, MutexGuard, OnceLock};
+use std::time::{Duration, Instant, SystemTime, UNIX_EPOCH};
 use tokio::net::TcpListener;
 use tokio::sync::{OwnedSemaphorePermit, Semaphore, watch};
 use tokio::task::JoinHandle;
@@ -363,6 +363,14 @@ pub struct RequestRecord {
    pub upload_id: Option<String>,
    pub part_number: Option<i32>,
    pub content_length: Option<u64>,
+    pub source_version_id: Option<String>,
+    pub source_mtime: Option<String>,
+    pub source_etag: Option<String>,
+    pub source_replication_request: bool,
+    pub journaled_at_unix_millis: Option<u64>,
+    /// Computed by the target, before response stall/disconnect faults. This
+    /// does not prove that any response bytes reached the client.
+    pub prepared_response: Option<PreparedResponse>,
    pub consumed_bytes: Option<usize>,
    pub replication_timestamps: ReplicationTimestampHeaders,
    pub proxy_headers: ProxyHeaderSnapshot,
@@ -381,6 +389,20 @@ pub struct RequestRecord {
    pub fault: Option<FaultAction>,
 }

+#[derive(Debug, Clone, PartialEq, Eq)]
+pub struct PreparedResponse {
+    pub status: u16,
+    pub version_id: Option<String>,
+    pub etag: Option<String>,
+    pub last_modified: Option<String>,
+    pub content_length: Option<u64>,
+    pub prepared_at_unix_millis: Option<u64>,
+    pub elapsed: Duration,
+}
+
+#[derive(Clone, Default)]
+struct RequestSequence(Arc<OnceLock<u64>>);
+
 /// Integrity and framing headers of an upload. A remote target's acceptance
 /// rules key off exactly these (rustfs#6853: `aws-chunked` framing stored
 /// verbatim; rustfs#7082: Object Lock PUTs need `Content-MD5` or
@@ -430,6 +452,10 @@ impl TransportSnapshot {
 #[derive(Debug, Clone, Default, PartialEq, Eq)]
 struct JournaledHeaders {
    content_length: Option<u64>,
+    source_version_id: Option<String>,
+    source_mtime: Option<String>,
+    source_etag: Option<String>,
+    source_replication_request: bool,
    replication_timestamps: ReplicationTimestampHeaders,
    proxy_headers: ProxyHeaderSnapshot,
    transport: TransportSnapshot,
@@ -444,6 +470,10 @@ impl JournaledHeaders {
                .get(CONTENT_LENGTH)
                .and_then(|value| value.to_str().ok())
                .and_then(|value| value.parse().ok()),
+            source_version_id: header_value(headers, &SOURCE_VERSION_ID_HEADERS).map(bounded_journal_value),
+            source_mtime: header_value(headers, &SOURCE_MTIME_HEADERS).map(bounded_journal_value),
+            source_etag: header_value(headers, &SOURCE_ETAG_HEADERS).map(bounded_journal_value),
+            source_replication_request: header_value(headers, &SOURCE_REPLICATION_REQUEST_HEADERS).as_deref() == Some("true"),
            replication_timestamps: ReplicationTimestampHeaders::from_headers(headers),
            proxy_headers: ProxyHeaderSnapshot::from_headers(headers),
            transport: TransportSnapshot::from_headers(headers),
@@ -511,6 +541,10 @@ struct StoreState {
    /// version id the target never had answers 404 `NoSuchVersion` instead of
    /// the idempotent 204 RustFS/MinIO give.
    reject_unknown_version_deletes: bool,
+    /// Models MinIO and its derivatives (Storj): `X-Minio-Internal-` is a reserved
+    /// metadata prefix, so a PutObject or UploadPart carrying any such request
+    /// header is refused with `InvalidArgument` (rustfs#8362).
+    reject_reserved_internal_headers: bool,
    limits: StoreLimits,
    buckets: HashMap<String, BucketState>,
    uploads: HashMap<String, MultipartState>,
@@ -738,6 +772,7 @@ impl FakeS3Target {
        let connection_limit = Arc::new(Semaphore::new(MAX_CONNECTIONS));
        let (active_connections_tx, active_connections) = watch::channel(0usize);
        let task_connection_limit = Arc::clone(&connection_limit);
+        let task_control = Arc::clone(&control);

        let task = tokio::spawn(async move {
            let mut connections = tokio::task::JoinSet::new();
@@ -751,6 +786,7 @@ impl FakeS3Target {
                        };
                        active_connections_tx.send_modify(|count| *count += 1);
                        let service = service.clone();
+                        let control = Arc::clone(&task_control);
                        let active_connections = active_connections_tx.clone();
                        let mut connection_gate = connection_gate.clone();
                        connections.spawn(async move {
@@ -764,7 +800,7 @@ impl FakeS3Target {
                            {
                                return;
                            }
-                            let handler = service_fn(move |request| handle_request(request, service.clone()));
+                            let handler = service_fn(move |request| handle_request(request, service.clone(), Arc::clone(&control)));
                            let mut connection = http1::Builder::new();
                            connection
                                .keep_alive(false)
@@ -975,6 +1011,43 @@ impl FakeS3Target {
        lock(&self.backend.store).reject_unknown_version_deletes = enabled;
    }

+    /// Refuse every PutObject / UploadPart that carries an `x-minio-internal-*`
+    /// request header, the way MinIO and Storj do.
+    pub fn reject_reserved_internal_headers(&self, enabled: bool) {
+        lock(&self.backend.store).reject_reserved_internal_headers = enabled;
+    }
+
+    /// Keys of every stored, non-delete object in the bucket, sorted.
+    pub fn stored_keys(&self, bucket: &str) -> Vec<String> {
+        let state = lock(&self.backend.store);
+        let mut keys: Vec<String> = state
+            .buckets
+            .get(bucket)
+            .map(|bucket| {
+                bucket
+                    .objects
+                    .iter()
+                    .filter(|(_, versions)| versions.last().is_some_and(|version| !version.delete_marker))
+                    .map(|(key, _)| key.clone())
+                    .collect()
+            })
+            .unwrap_or_default();
+        keys.sort();
+        keys
+    }
+
+    /// Body and user metadata (names as stored, without the `x-amz-meta-` prefix)
+    /// of the latest non-delete version of an object.
+    pub fn stored_object(&self, bucket: &str, key: &str) -> Option<(Bytes, HashMap<String, String>)> {
+        lock(&self.backend.store)
+            .buckets
+            .get(bucket)
+            .and_then(|bucket| bucket.objects.get(key))
+            .and_then(|versions| versions.last())
+            .filter(|version| !version.delete_marker)
+            .map(|version| (version.body.clone(), version.metadata.clone().unwrap_or_default()))
+    }
+
    pub fn require_checksum_for_object_lock(&self, enabled: bool) {
        lock(&self.backend.store).require_checksum_for_object_lock = enabled;
    }
@@ -1161,7 +1234,10 @@ impl S3Access for FaultAccess {
        let operation = operation_from_s3_name(context.s3_op().name());
        let journaled = JournaledHeaders::from_headers(context.headers());
        let aws_chunked = journaled.transport.aws_chunked;
-        let fault = record_request(&self.control, operation, context.method().clone(), parsed, journaled);
+        let (sequence, fault) = record_request(&self.control, operation, context.method().clone(), parsed, journaled);
+        if let Some(slot) = context.extensions_mut().get::<RequestSequence>() {
+            let _ = slot.0.set(sequence);
+        }
        if let Some(status) = fault.as_ref().and_then(|fault| scripted_status(&fault.action)) {
            return Err(scripted_status_error(status));
        }
@@ -1174,6 +1250,15 @@ impl S3Access for FaultAccess {
                "this target does not decode aws-chunked request bodies; send a plain signed payload with an exact Content-Length"
            ));
        }
+        if matches!(operation, Operation::PutObject | Operation::UploadPart)
+            && lock(&self.store).reject_reserved_internal_headers
+            && context
+                .headers()
+                .keys()
+                .any(|name| name.as_str().starts_with("x-minio-internal-"))
+        {
+            return Err(s3s::s3_error!(InvalidArgument, "Your metadata headers are not supported."));
+        }
        let prebody_permit = if operation == Operation::CompleteMultipartUpload {
            Some(
                timeout(
@@ -1228,7 +1313,7 @@ fn record_request(
    method: Method,
    parsed: ParsedRequest,
    headers: JournaledHeaders,
-) -> Option<RequestFault> {
+) -> (u64, Option<RequestFault>) {
    let mut state = lock(control);
    let action = parsed
        .key
@@ -1251,6 +1336,12 @@ fn record_request(
        upload_id: parsed.upload_id.map(bounded_journal_value),
        part_number: parsed.part_number,
        content_length: headers.content_length,
+        source_version_id: headers.source_version_id,
+        source_mtime: headers.source_mtime,
+        source_etag: headers.source_etag,
+        source_replication_request: headers.source_replication_request,
+        journaled_at_unix_millis: unix_millis(),
+        prepared_response: None,
        consumed_bytes: None,
        replication_timestamps: headers.replication_timestamps,
        proxy_headers: headers.proxy_headers,
@@ -1261,7 +1352,14 @@ fn record_request(
        continuation_token: parsed.continuation_token.map(bounded_journal_value),
        fault: action.clone(),
    });
-    action.map(|action| RequestFault { sequence, action })
+    (sequence, action.map(|action| RequestFault { sequence, action }))
+}
+
+fn unix_millis() -> Option<u64> {
+    SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .ok()
+        .and_then(|elapsed| u64::try_from(elapsed.as_millis()).ok())
 }

 fn bounded_journal_value(mut value: String) -> String {
@@ -1269,14 +1367,31 @@ fn bounded_journal_value(mut value: String) -> String {
    while !value.is_char_boundary(end) {
        end -= 1;
    }
-    value.truncate(end);
-    value
+    if value.capacity() > MAX_RETAINED_IDENTIFIER_BYTES {
+        value[..end].to_owned()
+    } else {
+        value.truncate(end);
+        value
+    }
 }

-async fn handle_request(request: Request<Incoming>, service: S3Service) -> Result<Response<Body>, BoxError> {
-    let mut response = timeout(MAX_REQUEST_DURATION, call_s3(service, request.map(Body::from)))
-        .await
-        .map_err(|_| io::Error::new(io::ErrorKind::TimedOut, "fake target request exceeded 65 seconds"))??;
+async fn handle_request(
+    mut request: Request<Incoming>,
+    service: S3Service,
+    control: Arc<Mutex<ControlState>>,
+) -> Result<Response<Body>, BoxError> {
+    let started = Instant::now();
+    let sequence = RequestSequence::default();
+    request.extensions_mut().insert(sequence.clone());
+    let mut response = timeout(MAX_REQUEST_DURATION, async {
+        let result = call_s3(service, request.map(Body::from)).await;
+        if let (Ok(response), Some(sequence)) = (&result, sequence.0.get()) {
+            update_prepared_response(&control, *sequence, response, started.elapsed());
+        }
+        result
+    })
+    .await
+    .map_err(|_| io::Error::new(io::ErrorKind::TimedOut, "fake target request exceeded 65 seconds"))??;
    if response.headers_mut().remove(DISCONNECT_HEADER).is_some() {
        return Err(io::Error::new(io::ErrorKind::ConnectionAborted, "fake target scripted disconnect").into());
    }
@@ -1531,6 +1646,25 @@ fn update_consumed(control: &Mutex<ControlState>, sequence: u64, consumed: usize
    }
 }

+fn update_prepared_response(control: &Mutex<ControlState>, sequence: u64, response: &Response<Body>, elapsed: Duration) {
+    let prepared = PreparedResponse {
+        status: response.status().as_u16(),
+        version_id: header_value(response.headers(), &["x-amz-version-id"]).map(bounded_journal_value),
+        etag: header_value(response.headers(), &[ETAG.as_str()]).map(bounded_journal_value),
+        last_modified: header_value(response.headers(), &[LAST_MODIFIED.as_str()]).map(bounded_journal_value),
+        content_length: response
+            .headers()
+            .get(CONTENT_LENGTH)
+            .and_then(|value| value.to_str().ok())
+            .and_then(|value| value.parse().ok()),
+        prepared_at_unix_millis: unix_millis(),
+        elapsed,
+    };
+    if let Some(record) = lock(control).requests.iter_mut().find(|record| record.sequence == sequence) {
+        record.prepared_response = Some(prepared);
+    }
+}
+
 async fn apply_non_body_fault(fault: Option<&RequestFault>, control: &Mutex<ControlState>) -> S3Result<()> {
    match fault.map(|fault| &fault.action) {
        Some(action @ (FaultAction::Status(_) | FaultAction::ResponseStatus(_))) => {
@@ -3230,10 +3364,38 @@ mod tests {
        let retry = client.clone();
        requests.spawn(async move { get_bytes(&retry, bucket, "held", None).await });
        timeout(Duration::from_secs(2), gate.entered.wait_for(|count| *count == 2)).await??;
+        let held_retry = target
+            .requests()
+            .into_iter()
+            .rev()
+            .find(|record| record.operation == Operation::GetObject && record.key.as_deref() == Some("held"))
+            .expect("the held retry must be journaled");
+        assert_eq!(held_retry.prepared_response, None);
        assert_eq!(
            timeout(Duration::from_secs(2), get_bytes(&client, bucket, "unrelated", None)).await??,
            Bytes::from_static(b"payload")
        );
+        let unrelated = target
+            .requests()
+            .into_iter()
+            .find(|record| record.operation == Operation::GetObject && record.key.as_deref() == Some("unrelated"))
+            .expect("the unrelated GET must be journaled");
+        let unrelated_response = unrelated.prepared_response.expect("the unrelated GET prepared a response");
+        assert_eq!(unrelated_response.status, 200);
+        assert_eq!(
+            unrelated_response.version_id.as_deref(),
+            Some(target.stored_versions(bucket, "unrelated")[0].0.as_str())
+        );
+        assert!(held_retry.sequence < unrelated.sequence);
+        assert_eq!(
+            target
+                .requests()
+                .into_iter()
+                .find(|record| record.sequence == held_retry.sequence)
+                .expect("the held retry remains journaled")
+                .prepared_response,
+            None
+        );
        assert!(requests.try_join_next().is_none(), "the retry must remain behind the gate");
        drop(gate);
        assert_eq!(
@@ -3242,6 +3404,18 @@ mod tests {
                .expect("retried GET task")??,
            Bytes::from_static(b"payload")
        );
+        let held_response = target
+            .requests()
+            .into_iter()
+            .find(|record| record.sequence == held_retry.sequence)
+            .and_then(|record| record.prepared_response)
+            .expect("the released retry prepared its own response");
+        assert_eq!(held_response.status, 200);
+        assert_eq!(
+            held_response.version_id.as_deref(),
+            Some(target.stored_versions(bucket, "held")[0].0.as_str())
+        );
+        assert_ne!(held_response.version_id, unrelated_response.version_id);
        assert_eq!(get_bytes(&client, bucket, "held", None).await?, Bytes::from_static(b"payload"));
        assert_eq!(target.count_requests(Operation::GetObject, "held"), 3);
        Ok(())
@@ -3586,6 +3760,23 @@ mod tests {
            .await?;
        assert_eq!(first_head.content_length(), Some(5));
        assert_eq!(first_head.version_id(), first.version_id.as_deref());
+        let journaled_head = target
+            .requests()
+            .into_iter()
+            .rev()
+            .find(|record| record.operation == Operation::HeadObject)
+            .expect("the version HEAD must be journaled");
+        assert_eq!(journaled_head.version_id, first.version_id);
+        assert!(journaled_head.journaled_at_unix_millis.is_some());
+        let prepared_head = journaled_head
+            .prepared_response
+            .expect("the version HEAD prepared a response");
+        assert_eq!(prepared_head.status, 200);
+        assert_eq!(prepared_head.version_id.as_deref(), first_head.version_id());
+        assert_eq!(prepared_head.etag.as_deref(), first_head.e_tag());
+        assert_eq!(prepared_head.content_length, Some(5));
+        assert!(prepared_head.last_modified.is_some());
+        assert!(prepared_head.prepared_at_unix_millis.is_some());
        assert_eq!(
            get_bytes(&client, "target-bucket", "nested/key", None).await?,
            Bytes::from_static(b"second")
@@ -3611,6 +3802,20 @@ mod tests {
            .await
            .expect_err("deleted version must not remain readable");
        assert_sdk_error!(deleted, 404, "NoSuchVersion");
+        let deleted_record = target
+            .requests()
+            .into_iter()
+            .rev()
+            .find(|record| record.operation == Operation::GetObject)
+            .expect("the deleted version GET must be journaled");
+        assert_eq!(deleted_record.version_id, first.version_id);
+        assert_eq!(
+            deleted_record
+                .prepared_response
+                .expect("the error response was prepared")
+                .status,
+            404
+        );
        let marker = client
            .delete_object()
            .bucket("target-bucket")
@@ -4042,7 +4247,21 @@ mod tests {
            .await
            .expect_err("wrong part ETag must fail without consuming the upload");
        assert_sdk_error!(invalid_complete, 400, "InvalidPart");
-        client
+        let rejected_complete = target
+            .requests()
+            .into_iter()
+            .rev()
+            .find(|record| record.operation == Operation::CompleteMultipartUpload)
+            .expect("the rejected completion must be journaled");
+        assert_eq!(rejected_complete.upload_id.as_deref(), Some(upload_id.as_str()));
+        assert_eq!(
+            rejected_complete
+                .prepared_response
+                .expect("the rejection was prepared")
+                .status,
+            400
+        );
+        let successful_complete = client
            .complete_multipart_upload()
            .bucket("target-bucket")
            .key("multipart")
@@ -4050,6 +4269,18 @@ mod tests {
            .multipart_upload(CompletedMultipartUpload::builder().set_parts(Some(parts)).build())
            .send()
            .await?;
+        let completed_record = target
+            .requests()
+            .into_iter()
+            .rev()
+            .find(|record| record.operation == Operation::CompleteMultipartUpload)
+            .expect("the successful completion must be journaled");
+        assert_eq!(completed_record.upload_id.as_deref(), Some(upload_id.as_str()));
+        let prepared_complete = completed_record
+            .prepared_response
+            .expect("the completion response was prepared");
+        assert_eq!(prepared_complete.status, 200);
+        assert_eq!(prepared_complete.version_id.as_deref(), successful_complete.version_id());
        let completed = get_bytes(&client, "target-bucket", "multipart", None).await?;
        assert_eq!(completed.len(), MIN_MULTIPART_PART_BYTES + 5);
        assert_eq!(&completed[..MIN_MULTIPART_PART_BYTES], first_body.as_ref());
@@ -4430,6 +4661,32 @@ mod tests {
            .send()
            .await?;
        assert_eq!(completed.version_id(), Some(completed_version.as_str()));
+        let receipt_records = target.requests();
+        let receipt_create = receipt_records
+            .iter()
+            .find(|record| {
+                record.operation == Operation::CreateMultipartUpload && record.key.as_deref() == Some("multipart-receipt-time")
+            })
+            .expect("the source version must be retained on MPU creation");
+        assert_eq!(receipt_create.source_version_id.as_deref(), Some(completed_version.as_str()));
+        let receipt_complete = receipt_records
+            .iter()
+            .find(|record| {
+                record.operation == Operation::CompleteMultipartUpload
+                    && record.upload_id.as_deref() == Some(receipt_upload.as_str())
+            })
+            .expect("the source mtime must be retained on completion");
+        assert_eq!(receipt_complete.source_mtime.as_deref(), Some("1970-01-01T00:00:00Z"));
+        assert!(receipt_complete.source_replication_request);
+        assert_eq!(
+            receipt_complete
+                .prepared_response
+                .as_ref()
+                .expect("the receipt-time response was prepared")
+                .version_id
+                .as_deref(),
+            completed.version_id()
+        );
        assert_eq!(
            get_bytes(&client, "target-bucket", "multipart-receipt-time", None).await?,
            Bytes::from_static(b"completed"),
@@ -4648,6 +4905,23 @@ mod tests {
            .expect_err("disconnected request must not store an object");
        assert_sdk_error!(disconnected, 404, "NotFound");

+        target.inject_for_key(Operation::PutObject, "lost-response", FaultAction::DisconnectAfterResponse, 1);
+        client
+            .put_object()
+            .bucket("target-bucket")
+            .key("lost-response")
+            .body(ByteStream::from_static(b"stored"))
+            .send()
+            .await
+            .expect_err("a prepared response does not prove client delivery");
+        let lost_response_head = client
+            .head_object()
+            .bucket("target-bucket")
+            .key("lost-response")
+            .send()
+            .await?;
+        assert_eq!(lost_response_head.content_length(), Some(6));
+
        assert_eq!(
            parse_request(&Method::PUT, &"/target-bucket/key?uploadId=upload&partNumber=bad".parse::<Uri>()?).operation,
            Operation::Unknown
@@ -4661,6 +4935,28 @@ mod tests {
                .count(),
            4
        );
+        assert!(
+            requests
+                .iter()
+                .filter(|request| request.fault == Some(FaultAction::Status(StatusCode::SERVICE_UNAVAILABLE)))
+                .all(|request| {
+                    request
+                        .prepared_response
+                        .as_ref()
+                        .is_some_and(|response| response.status == 503)
+                })
+        );
+        let lost_response = requests
+            .iter()
+            .find(|request| request.operation == Operation::PutObject && request.key.as_deref() == Some("lost-response"))
+            .expect("the disconnected response must be journaled");
+        assert_eq!(lost_response.fault, Some(FaultAction::DisconnectAfterResponse));
+        let prepared_lost_response = lost_response
+            .prepared_response
+            .as_ref()
+            .expect("the target prepared the successful response");
+        assert_eq!(prepared_lost_response.status, 200);
+        assert_eq!(prepared_lost_response.version_id.as_deref(), lost_response_head.version_id());
        assert!(
            requests
                .iter()
@@ -4851,8 +5147,44 @@ mod tests {
            assert_eq!(records.len(), MAX_REQUEST_RECORDS);
            assert_eq!(records.front().map(|record| record.sequence), Some(2));
            assert_eq!(records.back().map(|record| record.sequence), Some((MAX_REQUEST_RECORDS + 1) as u64));
+            let mut oversized_allocation = String::with_capacity(MAX_RETAINED_IDENTIFIER_BYTES * 16);
+            oversized_allocation.push_str("short");
+            let short_value = bounded_journal_value(oversized_allocation);
+            assert_eq!(short_value, "short");
+            assert!(short_value.capacity() <= MAX_RETAINED_IDENTIFIER_BYTES);
+            let mut response = Response::new(Body::empty());
+            for name in ["x-amz-version-id", "etag", "last-modified"] {
+                response.headers_mut().insert(
+                    name,
+                    HeaderValue::from_str(&"r".repeat(MAX_RETAINED_IDENTIFIER_BYTES + 1)).expect("header"),
+                );
+            }
+            response.headers_mut().insert(CONTENT_LENGTH, HeaderValue::from_static("17"));
+            response
+                .headers_mut()
+                .insert("authorization", HeaderValue::from_static("unrecorded-response-authorization"));
+            response.headers_mut().insert(
+                "x-amz-server-side-encryption-customer-key",
+                HeaderValue::from_static("unrecorded-response-customer-key"),
+            );
+            update_prepared_response(&control, 1, &response, Duration::ZERO);
+            update_prepared_response(&control, u64::MAX, &response, Duration::ZERO);
+            assert_eq!(lock(&control).requests, records, "misses and evicted records must not be resurrected");
            let bounded_control = Mutex::new(ControlState::default());
            let utf8_boundary = format!("{}é", "a".repeat(MAX_RETAINED_IDENTIFIER_BYTES - 1));
+            let mut source_headers = HeaderMap::new();
+            for name in ["x-rustfs-source-version-id", "x-minio-source-mtime", "x-rustfs-source-etag"] {
+                source_headers.insert(
+                    name,
+                    HeaderValue::from_str(&"s".repeat(MAX_RETAINED_IDENTIFIER_BYTES + 1)).expect("header"),
+                );
+            }
+            source_headers.insert("x-minio-source-replication-request", HeaderValue::from_static("true"));
+            source_headers.insert("authorization", HeaderValue::from_static("unrecorded-request-authorization"));
+            source_headers.insert(
+                "x-amz-server-side-encryption-customer-key",
+                HeaderValue::from_static("unrecorded-request-customer-key"),
+            );
            record_request(
                &bounded_control,
                Operation::GetObject,
@@ -4865,8 +5197,9 @@ mod tests {
                    upload_id: Some("u".repeat(MAX_RETAINED_IDENTIFIER_BYTES + 1)),
                    ..Default::default()
                },
-                JournaledHeaders::default(),
+                JournaledHeaders::from_headers(&source_headers),
            );
+            update_prepared_response(&bounded_control, 1, &response, Duration::from_millis(7));
            {
                let bounded_records = lock(&bounded_control);
                let bounded = &bounded_records.requests[0];
@@ -4874,7 +5207,39 @@ mod tests {
                assert_eq!(bounded.key.as_ref().map(String::len), Some(MAX_RETAINED_IDENTIFIER_BYTES - 1));
                assert_eq!(bounded.version_id.as_ref().map(String::len), Some(MAX_RETAINED_IDENTIFIER_BYTES));
                assert_eq!(bounded.upload_id.as_ref().map(String::len), Some(MAX_RETAINED_IDENTIFIER_BYTES));
+                assert_eq!(bounded.source_version_id.as_ref().map(String::len), Some(MAX_RETAINED_IDENTIFIER_BYTES));
+                assert_eq!(bounded.source_mtime.as_ref().map(String::len), Some(MAX_RETAINED_IDENTIFIER_BYTES));
+                assert_eq!(bounded.source_etag.as_ref().map(String::len), Some(MAX_RETAINED_IDENTIFIER_BYTES));
+                for value in [&bounded.source_version_id, &bounded.source_mtime, &bounded.source_etag]
+                    .into_iter()
+                    .flatten()
+                {
+                    assert!(value.capacity() <= MAX_RETAINED_IDENTIFIER_BYTES);
+                }
+                assert!(bounded.source_replication_request);
+                let prepared = bounded
+                    .prepared_response
+                    .as_ref()
+                    .expect("the retained record receives its response");
+                assert_eq!(prepared.version_id.as_ref().map(String::len), Some(MAX_RETAINED_IDENTIFIER_BYTES));
+                assert_eq!(prepared.etag.as_ref().map(String::len), Some(MAX_RETAINED_IDENTIFIER_BYTES));
+                assert_eq!(prepared.last_modified.as_ref().map(String::len), Some(MAX_RETAINED_IDENTIFIER_BYTES));
+                for value in [&prepared.version_id, &prepared.etag, &prepared.last_modified]
+                    .into_iter()
+                    .flatten()
+                {
+                    assert!(value.capacity() <= MAX_RETAINED_IDENTIFIER_BYTES);
+                }
+                assert_eq!(prepared.content_length, Some(17));
+                assert_eq!(prepared.elapsed, Duration::from_millis(7));
+                assert!(
+                    !format!("{bounded:?}").contains("unrecorded-"),
+                    "authorization and customer keys must not enter the journal"
+                );
            }
+            lock(&bounded_control).requests.clear();
+            update_prepared_response(&bounded_control, 1, &response, Duration::ZERO);
+            assert!(lock(&bounded_control).requests.is_empty(), "a drained record must not be resurrected");
            let target = FakeS3Target::start().await.expect("target");
            target.create_bucket("b".repeat(63));
            let empty_bucket = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
@@ -61,6 +61,8 @@

 use std::io;
 use std::net::{Ipv4Addr, SocketAddr};
+use std::sync::Arc;
+use std::sync::atomic::{AtomicUsize, Ordering};
 use std::time::Duration;

 use tokio::io::{AsyncReadExt, AsyncWriteExt};
@@ -108,19 +110,33 @@ pub struct FaultProxy {
    mode_tx: watch::Sender<FaultMode>,
    shutdown_tx: watch::Sender<bool>,
    accept_task: JoinHandle<()>,
+    traffic: Arc<ProxyTraffic>,
+}
+
+#[derive(Default)]
+struct ProxyTraffic {
+    forwarded_bytes: AtomicUsize,
+    dropped_bytes: AtomicUsize,
 }

 impl FaultProxy {
    /// Bind a listener on `127.0.0.1:0` and start forwarding accepted
    /// connections to `target`. Starts in [`FaultMode::Pass`].
    pub async fn start(target: SocketAddr) -> io::Result<Self> {
-        let listener = TcpListener::bind((Ipv4Addr::LOCALHOST, 0)).await?;
+        Self::start_on(SocketAddr::from((Ipv4Addr::LOCALHOST, 0)), target).await
+    }
+
+    /// Bind a specific address so advertised volume endpoints can retain the
+    /// server's port and authenticated RPC identity on another loopback IP.
+    pub async fn start_on(listen: SocketAddr, target: SocketAddr) -> io::Result<Self> {
+        let listener = TcpListener::bind(listen).await?;
        let listen_addr = listener.local_addr()?;

        let (mode_tx, mode_rx) = watch::channel(FaultMode::Pass);
        let (shutdown_tx, shutdown_rx) = watch::channel(false);

-        let accept_task = tokio::spawn(accept_loop(listener, target, mode_rx, shutdown_rx));
+        let traffic = Arc::new(ProxyTraffic::default());
+        let accept_task = tokio::spawn(accept_loop(listener, target, mode_rx, shutdown_rx, traffic.clone()));

        Ok(Self {
            listen_addr,
@@ -128,9 +144,20 @@ impl FaultProxy {
            mode_tx,
            shutdown_tx,
            accept_task,
+            traffic,
        })
    }

+    /// Bytes actually forwarded in either direction, excluding dropped traffic.
+    pub fn forwarded_bytes(&self) -> usize {
+        self.traffic.forwarded_bytes.load(Ordering::Relaxed)
+    }
+
+    /// Bytes actually discarded by a blackhole or one-way partition.
+    pub fn dropped_bytes(&self) -> usize {
+        self.traffic.dropped_bytes.load(Ordering::Relaxed)
+    }
+
    /// Local address the proxy is listening on. Point clients here instead of
    /// the real target to route their traffic through the proxy.
    pub fn local_addr(&self) -> SocketAddr {
@@ -171,6 +198,7 @@ async fn accept_loop(
    target: SocketAddr,
    mode_rx: watch::Receiver<FaultMode>,
    mut shutdown_rx: watch::Receiver<bool>,
+    traffic: Arc<ProxyTraffic>,
 ) {
    loop {
        tokio::select! {
@@ -188,6 +216,7 @@ async fn accept_loop(
                            target,
                            mode_rx.clone(),
                            shutdown_rx.clone(),
+                            traffic.clone(),
                        ));
                    }
                    // A transient accept error should not tear the proxy down.
@@ -205,6 +234,7 @@ async fn handle_connection(
    target: SocketAddr,
    mode_rx: watch::Receiver<FaultMode>,
    shutdown_rx: watch::Receiver<bool>,
+    traffic: Arc<ProxyTraffic>,
 ) {
    let server = match TcpStream::connect(target).await {
        Ok(s) => s,
@@ -221,8 +251,9 @@ async fn handle_connection(
        server_wr,
        mode_rx.clone(),
        shutdown_rx.clone(),
+        traffic.clone(),
    ));
-    let down = tokio::spawn(pump(Direction::ServerToClient, server_rd, client_wr, mode_rx, shutdown_rx));
+    let down = tokio::spawn(pump(Direction::ServerToClient, server_rd, client_wr, mode_rx, shutdown_rx, traffic));

    let _ = up.await;
    let _ = down.await;
@@ -237,6 +268,7 @@ async fn pump<R, W>(
    mut to: W,
    mut mode_rx: watch::Receiver<FaultMode>,
    mut shutdown_rx: watch::Receiver<bool>,
+    traffic: Arc<ProxyTraffic>,
 ) where
    R: AsyncReadExt + Unpin,
    W: AsyncWriteExt + Unpin,
@@ -257,15 +289,22 @@ async fn pump<R, W>(
        let mode = *mode_rx.borrow_and_update();
        match mode {
            // Whole connection is a black hole: drain and drop.
-            FaultMode::Blackhole => continue,
+            FaultMode::Blackhole => {
+                traffic.dropped_bytes.fetch_add(n, Ordering::Relaxed);
+                continue;
+            }
            // This direction is partitioned off: drop its bytes; the peer keeps
            // reading nothing while the other direction may still flow.
-            FaultMode::Partition(blocked) if blocked == dir => continue,
+            FaultMode::Partition(blocked) if blocked == dir => {
+                traffic.dropped_bytes.fetch_add(n, Ordering::Relaxed);
+                continue;
+            }
            FaultMode::Latency(delay) => {
                tokio::time::sleep(delay).await;
                if to.write_all(&buf[..n]).await.is_err() {
                    break;
                }
+                traffic.forwarded_bytes.fetch_add(n, Ordering::Relaxed);
                if to.flush().await.is_err() {
                    break;
                }
@@ -274,6 +313,7 @@ async fn pump<R, W>(
                if to.write_all(&buf[..n]).await.is_err() {
                    break;
                }
+                traffic.forwarded_bytes.fetch_add(n, Ordering::Relaxed);
                if to.flush().await.is_err() {
                    break;
                }
@@ -285,7 +325,6 @@ async fn pump<R, W>(
 #[cfg(test)]
 mod tests {
    use super::*;
-    use std::sync::Arc;
    use std::time::Instant;
    use tokio::sync::Mutex;

@@ -375,6 +414,8 @@ mod tests {

        assert_eq!(got, payload, "pass mode must forward bytes unchanged");
        assert_eq!(echo.received().await, payload, "server must have received the bytes");
+        assert_eq!(proxy.forwarded_bytes(), payload.len() * 2);
+        assert_eq!(proxy.dropped_bytes(), 0);

        proxy.shutdown().await;
    }
@@ -414,6 +455,8 @@ mod tests {

        assert!(got.is_empty(), "blackhole mode must not return any bytes, got {got:?}");
        assert!(echo.received().await.is_empty(), "blackhole must not forward to the server");
+        assert_eq!(proxy.forwarded_bytes(), 0);
+        assert_eq!(proxy.dropped_bytes(), b"into-the-void".len());

        // Proxy is still alive and steerable after a blackholed connection.
        proxy.set_mode(FaultMode::Pass);
@@ -562,17 +562,23 @@ mod tests {
        create_bucket(&client, bucket).await.expect("Failed to create bucket");

        let object_count = 1002;
-        for i in 0..object_count {
-            let key = format!("object{:04}.txt", i);
-            client
-                .put_object()
-                .bucket(bucket)
-                .key(&key)
-                .body(ByteStream::from_static(b"test content"))
-                .send()
-                .await
-                .expect("Failed to put object");
-        }
+        // Finish every durable PUT before listing, with the same bounded fanout as the delimiter fixtures.
+        stream::iter(0..object_count)
+            .for_each_concurrent(16, |i| {
+                let client = &client;
+                async move {
+                    let key = format!("object{i:04}.txt");
+                    client
+                        .put_object()
+                        .bucket(bucket)
+                        .key(&key)
+                        .body(ByteStream::from_static(b"test content"))
+                        .send()
+                        .await
+                        .unwrap_or_else(|err| panic!("Failed to put fixture object {key}: {err}"));
+                }
+            })
+            .await;

        eprintln!("Seeded {object_count} objects in {bucket}; starting ListObjectsV2 pagination");
        let deadline = Instant::now() + PAGINATION_TIMEOUT;
@@ -300,17 +300,28 @@ async fn test_degraded_cluster_read_quorum_follows_erasure_layout() -> TestResul
        }

        cluster.stop_node(read_quorum - 1)?;
+        let lock_read_quorum = node_count - node_count / 2;
+        // EC:2 on six nodes still has three shared lock votes after losing
+        // its fourth data shard. Preserve the distinct data-quorum error.
+        let expected_code = if read_quorum - 1 < lock_read_quorum {
+            "ServiceUnavailable"
+        } else {
+            "SlowDownRead"
+        };
        for client in clients.iter().take(read_quorum - 1) {
            match client.get_object().bucket(&bucket).key("below-quorum").send().await {
                Ok(response) => assert!(
                    response.body.collect().await.is_err(),
                    "fewer than {read_quorum} valid fragments must not reconstruct an uncached object"
                ),
-                Err(error) => assert_eq!(
-                    error.as_service_error().and_then(|error| error.meta().code()),
-                    Some("ServiceUnavailable"),
-                    "a quorum loss must not be mistaken for a missing object"
-                ),
+                Err(error) => {
+                    assert_eq!(error.raw_response().map(|response| response.status().as_u16()), Some(503));
+                    assert_eq!(
+                        error.as_service_error().and_then(|error| error.meta().code()),
+                        Some(expected_code),
+                        "a quorum loss must not be mistaken for a missing object"
+                    );
+                }
            }
        }

@@ -425,6 +436,60 @@ async fn test_concurrent_cluster_overwrites_do_not_fail_namespace_lock_quorum()
    Ok(())
 }

+fn unavailable_readiness_drive_indices(payload: &serde_json::Value) -> Vec<usize> {
+    let mut indices = payload["details"]["storage"]["unavailableDrives"]
+        .as_array()
+        .expect("node readiness must report unavailable drives")
+        .iter()
+        .map(|drive| {
+            usize::try_from(drive["diskIndex"].as_u64().expect("drive index must be unsigned"))
+                .expect("drive index must fit usize")
+        })
+        .collect::<Vec<_>>();
+    indices.sort_unstable();
+    indices
+}
+
+fn assert_node_readiness_contract(payload: &serde_json::Value, survivors: usize) {
+    let read_ready = survivors >= 2;
+    let write_ready = survivors >= 3;
+    let storage = &payload["details"]["storage"];
+    assert_eq!(payload["ready"], read_ready);
+    assert_eq!(storage["ready"], read_ready);
+    assert_eq!(storage["readQuorum"], read_ready);
+    assert_eq!(storage["writeQuorum"], write_ready);
+    assert_eq!(storage["readinessScope"], "read_quorum");
+    assert_eq!(storage["source"], "local_runtime");
+    assert_eq!(storage["status"], if read_ready { "connected" } else { "disconnected" });
+    assert_eq!(payload["details"]["poolMetadata"]["ready"], true);
+    assert_eq!(payload["details"]["iam"]["ready"], true);
+    assert_eq!(payload["details"]["lock"]["ready"], read_ready);
+    assert_eq!(unavailable_readiness_drive_indices(payload), (survivors..4).collect::<Vec<_>>());
+    for drive in storage["unavailableDrives"].as_array().expect("unavailable drives") {
+        assert_eq!(drive["poolIndex"], 0);
+        assert_eq!(drive["setIndex"], 0);
+        let state = drive["runtimeState"].as_str().expect("runtime state must be reported");
+        let host_online = drive["hostOnline"].as_bool().expect("host reachability must be reported");
+        assert!(
+            matches!(state, "online" | "suspect" | "offline" | "returning"),
+            "unknown runtime state: {state}"
+        );
+        assert!(
+            state != "online" || !host_online,
+            "a reachable online drive cannot be excluded from quorum"
+        );
+    }
+    if !read_ready {
+        assert!(
+            payload["degradedReasons"]
+                .as_array()
+                .expect("degraded reasons")
+                .iter()
+                .any(|reason| reason == "storage_and_lock_unavailable")
+        );
+    }
+}
+
 async fn assert_node_readiness_tracks_quorum(cluster: &mut RustFSTestClusterEnvironment) -> TestResult {
    let clients: Vec<_> = cluster
        .create_all_clients()?
@@ -483,27 +548,14 @@ async fn assert_node_readiness_tracks_quorum(cluster: &mut RustFSTestClusterEnvi
                    && payload["details"]["poolMetadata"]["ready"] == true
                    && payload["details"]["iam"]["ready"] == true
                    && payload["details"]["lock"]["ready"] == read_quorum
+                    && unavailable_readiness_drive_indices(&payload) == (survivors..4).collect::<Vec<_>>()
                {
                    break payload;
                }
                assert!(Instant::now() < deadline, "node {idx}, survivors={survivors}: HTTP {status}, {payload}");
                tokio::time::sleep(Duration::from_millis(200)).await;
            };
-            assert_eq!(payload["details"]["storage"]["readinessScope"], "read_quorum");
-            assert_eq!(payload["details"]["storage"]["source"], "local_runtime");
-            assert_eq!(
-                payload["details"]["storage"]["status"],
-                if node_ready { "connected" } else { "disconnected" }
-            );
-            if !node_ready {
-                assert!(
-                    payload["degradedReasons"]
-                        .as_array()
-                        .expect("degraded reasons")
-                        .iter()
-                        .any(|reason| reason == "storage_and_lock_unavailable")
-                );
-            }
+            assert_node_readiness_contract(&payload, survivors);

            for path in ["/health/ready", "/minio/health/ready"] {
                let head = http.head(format!("{url}{path}")).send().await?;
@@ -512,7 +564,9 @@ async fn assert_node_readiness_tracks_quorum(cluster: &mut RustFSTestClusterEnvi
                let response = http.get(format!("{url}{path}")).send().await?;
                assert_eq!(response.status().as_u16(), node_status);
                let body: serde_json::Value = response.json().await?;
-                assert_eq!(body["details"]["storage"], payload["details"]["storage"]);
+                // Each response samples live diagnostics; suspect -> offline
+                // must not change the quorum contract or excluded topology.
+                assert_node_readiness_contract(&body, survivors);
                assert_eq!(body["details"]["poolMetadata"], payload["details"]["poolMetadata"]);
            }
            let live = http.get(format!("{url}/health/live")).send().await?;
@@ -15,7 +15,7 @@
 #[cfg(test)]
 pub(crate) use rustfs_ecstore::api::bucket::bucket_target_sys::BucketTargetSys;
 #[cfg(test)]
-pub(crate) use rustfs_ecstore::api::disk::{RUSTFS_META_BUCKET, VolumeInfo, WalkDirOptions};
+pub(crate) use rustfs_ecstore::api::disk::{Endpoint, RUSTFS_META_BUCKET, VolumeInfo, WalkDirOptions};
 pub(crate) use rustfs_ecstore::api::rpc::{AuthenticatedChannel, TonicInterceptor, node_service_time_out_client_no_auth};
 #[cfg(test)]
 pub(crate) use rustfs_ecstore::api::rpc::{
@@ -165,4 +165,210 @@ mod tests {
        info!("RT-13c PASS: scanner/config persists across restart");
        Ok(())
    }
+
+    /// RT-14: a transition PUT must not carry reserved `x-minio-internal-*` headers.
+    ///
+    /// Regression pattern: transition forwarded the object's internal metadata as raw
+    /// request headers, so MinIO-derived remotes (Storj) refused every transition with
+    /// `400 InvalidArgument` while `tier add` still succeeded (rustfs#8362). The fake
+    /// target refuses any PutObject carrying such a header, like MinIO and Storj do.
+    #[tokio::test]
+    async fn test_transition_put_omits_reserved_internal_headers() -> TestResult {
+        use crate::fake_s3_target::{BucketMode, FAKE_ACCESS_KEY, FAKE_SECRET_KEY, FakeS3Target, Operation};
+        use aws_sdk_s3::primitives::ByteStream;
+        use aws_sdk_s3::types::{
+            BucketLifecycleConfiguration, ExpirationStatus, LifecycleRule, LifecycleRuleFilter, Transition,
+            TransitionStorageClass,
+        };
+        use std::time::{Duration, Instant};
+
+        init_logging();
+        info!("RT-14: transition PUT omits reserved internal headers");
+
+        const TIER: &str = "RT14COLD";
+        const TARGET_BUCKET: &str = "rt14-cold";
+        const SOURCE_BUCKET: &str = "rt14-hot";
+
+        let target = FakeS3Target::start().await.expect("start fake target");
+        target.create_bucket_with_mode(TARGET_BUCKET, BucketMode::Unversioned);
+        target.reject_reserved_internal_headers(true);
+
+        let mut env = RustFSTestEnvironment::new().await.expect("create test environment");
+        env.start_rustfs_server_with_env(
+            vec![],
+            &[
+                ("RUSTFS_CONSOLE_ENABLE", "false"),
+                ("RUSTFS_TIER_RUSTFS_ALLOW_LOOPBACK_ENDPOINT", "true"),
+            ],
+        )
+        .await
+        .expect("start RustFS");
+
+        let tier_body = serde_json::json!({
+            "type": "rustfs",
+            "rustfs": {
+                "name": TIER,
+                "endpoint": target.endpoint(),
+                "accessKey": FAKE_ACCESS_KEY,
+                "secretKey": FAKE_SECRET_KEY,
+                "bucket": TARGET_BUCKET,
+                "prefix": "cold/",
+                "region": "us-east-1",
+                "storageClass": ""
+            }
+        })
+        .to_string();
+        let deadline = Instant::now() + Duration::from_secs(30);
+        loop {
+            match admin_ok(&env, http::Method::PUT, "/rustfs/admin/v3/tier", Some(tier_body.clone())).await {
+                Ok(_) => break,
+                Err(err) if Instant::now() < deadline => {
+                    info!("  add tier not ready yet: {err}");
+                    tokio::time::sleep(Duration::from_millis(500)).await;
+                }
+                Err(err) => panic!("RT-14 FAIL: add tier: {err}"),
+            }
+        }
+
+        let client = env.create_s3_client();
+        client
+            .create_bucket()
+            .bucket(SOURCE_BUCKET)
+            .send()
+            .await
+            .expect("create bucket");
+
+        let objects: Vec<(&str, Vec<u8>)> = vec![
+            ("small.bin", (0..1024u32).map(|i| (i % 251) as u8).collect()),
+            ("medium.bin", (0..200_000u32).map(|i| (i.wrapping_mul(31) % 253) as u8).collect()),
+        ];
+        for (key, body) in &objects {
+            client
+                .put_object()
+                .bucket(SOURCE_BUCKET)
+                .key(*key)
+                .body(ByteStream::from(body.clone()))
+                .send()
+                .await
+                .expect("put source object");
+        }
+
+        let rule = LifecycleRule::builder()
+            .id("to-cold")
+            .status(ExpirationStatus::Enabled)
+            .filter(LifecycleRuleFilter::builder().prefix("").build())
+            .transitions(
+                Transition::builder()
+                    .days(0)
+                    .storage_class(TransitionStorageClass::from(TIER))
+                    .build(),
+            )
+            .build()
+            .expect("build rule");
+        client
+            .put_bucket_lifecycle_configuration()
+            .bucket(SOURCE_BUCKET)
+            .lifecycle_configuration(
+                BucketLifecycleConfiguration::builder()
+                    .rules(rule)
+                    .build()
+                    .expect("build config"),
+            )
+            .send()
+            .await
+            .expect("put lifecycle");
+
+        admin_ok(
+            &env,
+            http::Method::POST,
+            &format!("/rustfs/admin/v3/ilm/transition/run?bucket={SOURCE_BUCKET}&prefix=&tier={TIER}&dryRun=false"),
+            None,
+        )
+        .await
+        .expect("run transition");
+
+        let deadline = Instant::now() + Duration::from_secs(60);
+        while target.stored_keys(TARGET_BUCKET).len() < objects.len() {
+            assert!(
+                Instant::now() < deadline,
+                "RT-14 FAIL: transition did not reach the target; stored keys: {:?}",
+                target.stored_keys(TARGET_BUCKET)
+            );
+            tokio::time::sleep(Duration::from_millis(500)).await;
+        }
+
+        let keys = target.stored_keys(TARGET_BUCKET);
+        assert_eq!(keys.len(), objects.len(), "RT-14 FAIL: unexpected remote objects: {keys:?}");
+
+        let mut stored_bodies = Vec::new();
+        for key in &keys {
+            let (body, metadata) = target.stored_object(TARGET_BUCKET, key).expect("remote object");
+            for name in [
+                "x-rustfs-internal-transition-transaction-id",
+                "x-minio-internal-transition-transaction-id",
+                "x-rustfs-internal-transition-tier-destination-id",
+                "x-minio-internal-transition-tier-destination-id",
+            ] {
+                assert!(
+                    metadata.contains_key(name),
+                    "RT-14 FAIL: remote object {key} lost transition identity {name}: {:?}",
+                    metadata.keys().collect::<Vec<_>>()
+                );
+            }
+            stored_bodies.push(body.to_vec());
+        }
+        let mut expected_bodies: Vec<Vec<u8>> = objects.iter().map(|(_, body)| body.clone()).collect();
+        expected_bodies.sort();
+        stored_bodies.sort();
+        assert_eq!(stored_bodies, expected_bodies, "RT-14 FAIL: remote bytes differ from the source objects");
+
+        let transition_puts = target
+            .requests()
+            .iter()
+            .filter(|record| {
+                record.operation == Operation::PutObject
+                    && record
+                        .key
+                        .as_deref()
+                        .is_some_and(|key| key.contains("transition-transactions"))
+            })
+            .count();
+        assert_eq!(
+            transition_puts,
+            objects.len(),
+            "RT-14 FAIL: transition PUTs were retried or refused by the target"
+        );
+
+        for (key, body) in &objects {
+            let head = client
+                .head_object()
+                .bucket(SOURCE_BUCKET)
+                .key(*key)
+                .send()
+                .await
+                .expect("head transitioned");
+            assert_eq!(
+                head.storage_class().map(|class| class.as_str()),
+                Some(TIER),
+                "RT-14 FAIL: {key} is not reported as transitioned"
+            );
+            let got = client
+                .get_object()
+                .bucket(SOURCE_BUCKET)
+                .key(*key)
+                .send()
+                .await
+                .expect("get transitioned")
+                .body
+                .collect()
+                .await
+                .expect("read body")
+                .into_bytes();
+            assert_eq!(got.as_ref(), body.as_slice(), "RT-14 FAIL: read-back of {key} differs");
+        }
+
+        target.shutdown().await;
+        info!("RT-14 PASS: transition PUT omits reserved internal headers");
+        Ok(())
+    }
 }
@@ -33,6 +33,7 @@ use aws_sdk_s3::types::{
 };
 use http::{Method, StatusCode};
 use serde_json::Value;
+use std::future::Future;
 use std::io::Write;
 use std::path::{Path, PathBuf};
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
@@ -57,6 +58,10 @@ const UPGRADE_READINESS_BODY: &[u8] = b"upgrade write readiness";
 const PREVIOUS_RELEASE_SEED_KEY: &str = ".upgrade-readiness/previous-seed/node-0";
 const MULTIPART_WORKERS: usize = 16;
 const MULTIPART_UPLOADS_PER_WORKER: usize = 16;
+// Bound each 64 KiB upload, including completion acknowledgement. A peer
+// that accepts the request but never responds must fail the compatibility
+// check with the writer and object, rather than consume the whole CI job.
+const MIXED_MULTIPART_UPLOAD_TIMEOUT: Duration = Duration::from_secs(30);
 // Peers keep a restarted node's drive in Suspect/Returning for roughly
 // probe_interval (2s) x success_threshold (3) after it comes back; 30s
 // comfortably covers that window plus CI scheduling jitter.
@@ -335,14 +340,21 @@ fn configure_cluster_logs(cluster: &mut RustFSTestClusterEnvironment) -> TestRes
 async fn write_multipart_load(clients: &[Client], phase: &str) -> Result<Vec<String>, Box<dyn std::error::Error + Send + Sync>> {
    let mut tasks = JoinSet::new();
    for worker in 0..MULTIPART_WORKERS {
-        let client = clients[worker % clients.len()].clone();
+        let node = worker % clients.len();
+        let client = clients[node].clone();
        let phase = phase.to_string();
        tasks.spawn(async move {
            let mut keys = Vec::with_capacity(MULTIPART_UPLOADS_PER_WORKER);
            for upload in 0..MULTIPART_UPLOADS_PER_WORKER {
                let key = format!("{phase}/multipart/{worker:02}/{upload:02}");
                let part = vec![u8::try_from(worker)?; 64 * 1024];
-                write_multipart(&client, MIXED_BUCKET, &key, &[part]).await?;
+                mixed_multipart_before(
+                    Instant::now() + MIXED_MULTIPART_UPLOAD_TIMEOUT,
+                    node,
+                    &key,
+                    write_multipart(&client, MIXED_BUCKET, &key, &[part]),
+                )
+                .await?;
                keys.push(key);
            }
            Ok::<_, Box<dyn std::error::Error + Send + Sync>>(keys)
@@ -356,6 +368,172 @@ async fn write_multipart_load(clients: &[Client], phase: &str) -> Result<Vec<Str
    Ok(keys)
 }

+async fn mixed_multipart_before(
+    deadline: Instant,
+    node: usize,
+    key: &str,
+    upload: impl Future<Output = TestResult>,
+) -> TestResult {
+    tokio::time::timeout_at(deadline, upload)
+        .await
+        .map_err(|_| MultipartAcknowledgementDeadline {
+            node,
+            key: key.to_owned(),
+        })?
+}
+
+#[derive(Debug)]
+struct MultipartAcknowledgementDeadline {
+    node: usize,
+    key: String,
+}
+
+impl std::fmt::Display for MultipartAcknowledgementDeadline {
+    fn fmt(&self, formatter: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(
+            formatter,
+            "node {}: multipart upload {MIXED_BUCKET}/{} acknowledgement deadline exceeded",
+            self.node, self.key
+        )
+    }
+}
+
+impl std::error::Error for MultipartAcknowledgementDeadline {}
+
+// Decode only the public lock snapshot fields. Unknown fields, including
+// credential/configuration payloads, never enter the failure artifact.
+#[derive(serde::Deserialize, serde::Serialize)]
+struct MixedLockSnapshot {
+    total: Option<u64>,
+    truncated: Option<bool>,
+    capability_note: Option<String>,
+    locks: Vec<MixedLockEntry>,
+}
+
+#[derive(serde::Deserialize, serde::Serialize)]
+struct MixedLockEntry {
+    resource: Option<String>,
+    bucket: Option<String>,
+    object: Option<String>,
+    version: Option<String>,
+    owner: Option<String>,
+    #[serde(rename = "type")]
+    lock_type: Option<String>,
+    priority: Option<String>,
+    since: Option<String>,
+    elapsed_secs: Option<u64>,
+    ttl_secs: Option<u64>,
+}
+
+async fn read_mixed_lock_snapshot(response: &mut reqwest::Response) -> Result<MixedLockSnapshot, &'static str> {
+    const MAX_BODY_BYTES: usize = 1024 * 1024;
+    let mut body = Vec::new();
+    while let Some(chunk) = response.chunk().await.map_err(|_| "body_read_error")? {
+        if chunk.len() > MAX_BODY_BYTES - body.len() {
+            return Err("body_limit_exceeded");
+        }
+        body.extend_from_slice(&chunk);
+    }
+    serde_json::from_slice(&body).map_err(|_| "invalid_lock_snapshot")
+}
+
+async fn probe_mixed_multipart_failure(
+    cluster: &RustFSTestClusterEnvironment,
+    node: usize,
+    path: &str,
+    head: bool,
+    deadline: Instant,
+) -> Value {
+    let started = Instant::now();
+    let method = if head { Method::HEAD } else { Method::GET };
+    let mut observation = serde_json::json!({
+        "node": node, "method": method.as_str(), "path": path,
+        "status": null, "observation": "unknown", "timeout": false,
+    });
+    // The shared absolute deadline bounds both headers and a slow/infinite
+    // response body; a send-only timeout would leave the cluster alive forever.
+    let result = tokio::time::timeout_at(deadline, async {
+        let mut response = signed_request(
+            method,
+            &format!("{}{path}", cluster.nodes[node].url.trim_end_matches('/')),
+            &cluster.access_key,
+            &cluster.secret_key,
+            None,
+            None,
+        )
+        .await
+        .map_err(|_| "request_error")?;
+        let status = response.status();
+        observation["status"] = serde_json::json!(status.as_u16());
+        if head {
+            observation["observation"] = serde_json::json!(match status {
+                StatusCode::OK => "visible",
+                StatusCode::NOT_FOUND => "not_visible",
+                _ => "unknown",
+            });
+            for (field, header) in [
+                ("content_length", "content-length"),
+                ("etag", "etag"),
+                ("version", "x-amz-version-id"),
+            ] {
+                observation[field] = serde_json::json!(response.headers().get(header).and_then(|value| value.to_str().ok()));
+            }
+        } else if status == StatusCode::OK {
+            observation["snapshot"] =
+                serde_json::to_value(read_mixed_lock_snapshot(&mut response).await?).map_err(|_| "snapshot_encode_error")?;
+            observation["observation"] = serde_json::json!("captured");
+        }
+        Ok::<_, &'static str>(())
+    })
+    .await;
+    match result {
+        Ok(Ok(())) => {}
+        Ok(Err(reason)) => observation["error"] = serde_json::json!(reason),
+        Err(_) => observation["timeout"] = serde_json::json!(true),
+    }
+    observation["elapsed_secs"] = serde_json::json!(started.elapsed().as_secs_f64());
+    observation
+}
+
+async fn observe_mixed_multipart_failure(
+    cluster: &RustFSTestClusterEnvironment,
+    phase: &str,
+    error: BoxError,
+    log_dir: Option<&Path>,
+) -> TestResult {
+    if let (Some(log_dir), Some(failure)) = (log_dir, error.downcast_ref::<MultipartAcknowledgementDeadline>()) {
+        let started = Instant::now();
+        let deadline = started + Duration::from_secs(5);
+        // Reserve the last second for serialization and bounded artifact I/O.
+        // All probes run directly in this future; no diagnostic task is spawned.
+        let network_deadline = started + Duration::from_secs(4);
+        let path = format!("/{MIXED_BUCKET}/{}", failure.key);
+        let _ = tokio::time::timeout_at(deadline, async {
+            let (heads, locks) = tokio::join!(
+                futures::future::join_all((0..cluster.nodes.len()).map(|node| probe_mixed_multipart_failure(
+                    cluster,
+                    node,
+                    &path,
+                    true,
+                    network_deadline
+                ))),
+                probe_mixed_multipart_failure(cluster, 0, "/rustfs/admin/v3/top/locks?count=1000", false, network_deadline),
+            );
+            let snapshot = serde_json::json!({
+                "phase": phase, "bucket": MIXED_BUCKET, "key": failure.key,
+                "writer_node": failure.node, "heads": heads, "top_locks": locks,
+            });
+            let body = serde_json::to_vec_pretty(&snapshot).map_err(|_| ())?;
+            tokio::fs::write(log_dir.join("mixed-multipart-failure.json"), body)
+                .await
+                .map_err(|_| ())
+        })
+        .await;
+    }
+    // Visibility and diagnostic failures cannot acknowledge the original MPU.
+    Err(error)
+}
+
 /// Assert that `client` eventually lists exactly `expected` objects under
 /// `{phase}/`, polling until [`LISTING_CONVERGENCE_TIMEOUT`].
 ///
@@ -400,7 +578,7 @@ async fn exercise_mixed_cluster(
    previous_node: usize,
 ) -> TestResult {
    let clients = cluster.create_all_clients()?;
-    wait_for_upgrade_write_readiness(&clients, phase, LISTING_CONVERGENCE_TIMEOUT).await?;
+    wait_for_upgrade_write_readiness(&clients, MIXED_BUCKET, phase, LISTING_CONVERGENCE_TIMEOUT).await?;
    let current_client = &clients[current_node];
    let previous_client = &clients[previous_node];

@@ -426,7 +604,13 @@ async fn exercise_mixed_cluster(
        .await?;
    assert_eq!(read_object(current_client, MIXED_BUCKET, &previous_key, None).await?.1, previous_body);

-    let multipart_keys = write_multipart_load(&clients, phase).await?;
+    let multipart_keys = match write_multipart_load(&clients, phase).await {
+        Ok(keys) => keys,
+        Err(error) => {
+            let log_dir = std::env::var_os("RUSTFS_E2E_LOG_DIR");
+            return observe_mixed_multipart_failure(cluster, phase, error, log_dir.as_deref().map(Path::new)).await;
+        }
+    };
    let expected_count = multipart_keys.len() + 2;
    for (label, client) in [("current", current_client), ("previous", previous_client)] {
        assert_eq!(
@@ -462,7 +646,12 @@ fn upgrade_probe_client(client: &Client) -> Client {
    )
 }

-async fn wait_for_upgrade_write_readiness(clients: &[Client], phase: &str, budget: Duration) -> TestResult {
+pub(crate) async fn wait_for_upgrade_write_readiness(
+    clients: &[Client],
+    bucket: &str,
+    phase: &str,
+    budget: Duration,
+) -> TestResult {
    // ListBuckets can succeed before peers recover a restarted disk. Cluster
    // health also accepts Returning disks whose write health is still FAULTY.
    // Probe every writer outside the asserted phase prefix; compatibility
@@ -482,7 +671,7 @@ async fn wait_for_upgrade_write_readiness(clients: &[Client], phase: &str, budge
                deadline,
                client
                    .put_object()
-                    .bucket(MIXED_BUCKET)
+                    .bucket(bucket)
                    .key(&key)
                    .body(ByteStream::from_static(UPGRADE_READINESS_BODY))
                    .send(),
@@ -517,19 +706,20 @@ async fn prepare_previous_release_baseline(cluster: &mut RustFSTestClusterEnviro
    // waits for pool metadata (#7473). First prove the elected writer can
    // persist data, then restart each old process once with three peers still
    // readable. This preparation ends before any current binary is started.
-    wait_for_upgrade_write_readiness(&clients[..1], "previous-seed", LISTING_CONVERGENCE_TIMEOUT).await?;
+    wait_for_upgrade_write_readiness(&clients[..1], MIXED_BUCKET, "previous-seed", LISTING_CONVERGENCE_TIMEOUT).await?;
    for node in [1, 2, 3, 0] {
        cluster.stop_node(node)?;
        cluster.start_node_from_binary(node, previous_binary).await?;
        wait_for_upgrade_write_readiness(
            std::slice::from_ref(&clients[node]),
+            MIXED_BUCKET,
            &format!("previous-restart-{node}"),
            LISTING_CONVERGENCE_TIMEOUT,
        )
        .await?;
    }

-    wait_for_upgrade_write_readiness(&clients, "previous-baseline", LISTING_CONVERGENCE_TIMEOUT).await?;
+    wait_for_upgrade_write_readiness(&clients, MIXED_BUCKET, "previous-baseline", LISTING_CONVERGENCE_TIMEOUT).await?;
    for (reader, client) in clients.iter().enumerate() {
        assert_eq!(
            read_object(client, MIXED_BUCKET, PREVIOUS_RELEASE_SEED_KEY, None).await?.1,
@@ -556,9 +746,10 @@ mod upgrade_write_readiness_tests {
    #[tokio::test]
    async fn waits_for_each_writer_after_metadata_is_ready() -> TestResult {
        let target = FakeS3Target::start().await?;
-        target.create_bucket(MIXED_BUCKET);
+        let bucket = "distributed-upgrade-history";
+        target.create_bucket(bucket);
        let client = fake_source_client(&target);
-        client.head_bucket().bucket(MIXED_BUCKET).send().await?;
+        client.head_bucket().bucket(bucket).send().await?;

        let phase = "one-previous-node";
        let first_key = format!(".upgrade-readiness/{phase}/node-0");
@@ -578,7 +769,7 @@ mod upgrade_write_readiness_tests {
        // Metadata readiness does not prove that a data write can succeed.
        let premature = client
            .put_object()
-            .bucket(MIXED_BUCKET)
+            .bucket(bucket)
            .key(&first_key)
            .body(ByteStream::from_static(b"upgrade write readiness"))
            .send()
@@ -586,15 +777,15 @@ mod upgrade_write_readiness_tests {
            .expect_err("metadata readiness does not prove write readiness");
        assert_eq!(premature.raw_response().map(|response| response.status().as_u16()), Some(503));

-        wait_for_upgrade_write_readiness(&[client.clone(), client.clone()], phase, Duration::from_secs(5)).await?;
+        wait_for_upgrade_write_readiness(&[client.clone(), client.clone()], bucket, phase, Duration::from_secs(5)).await?;
        assert_eq!(target.count_requests(FakeTargetOperation::PutObject, &first_key), 3);
        assert_eq!(target.count_requests(FakeTargetOperation::PutObject, &second_key), 2);
-        assert!(target.has_object(MIXED_BUCKET, &first_key));
-        assert!(target.has_object(MIXED_BUCKET, &second_key));
+        assert!(target.has_object(bucket, &first_key));
+        assert!(target.has_object(bucket, &second_key));
        assert!(
            client
                .list_objects_v2()
-                .bucket(MIXED_BUCKET)
+                .bucket(bucket)
                .prefix(format!("{phase}/"))
                .send()
                .await?
@@ -624,9 +815,10 @@ mod upgrade_write_readiness_tests {
            let phase = format!("permanent-{}", status.as_u16());
            let key = format!(".upgrade-readiness/{phase}/node-0");
            target.inject_for_key(FakeTargetOperation::PutObject, &key, FaultAction::Status(status), 1);
-            let error = wait_for_upgrade_write_readiness(std::slice::from_ref(&client), &phase, Duration::from_secs(5))
-                .await
-                .expect_err("a permanent error must not be retried into success");
+            let error =
+                wait_for_upgrade_write_readiness(std::slice::from_ref(&client), MIXED_BUCKET, &phase, Duration::from_secs(5))
+                    .await
+                    .expect_err("a permanent error must not be retried into success");
            assert!(error.to_string().contains("node 0 write readiness failed"), "{error}");
            assert_eq!(target.count_requests(FakeTargetOperation::PutObject, &key), 1);
            assert!(!target.has_object(MIXED_BUCKET, &key));
@@ -649,7 +841,7 @@ mod upgrade_write_readiness_tests {
            FaultAction::Status(StatusCode::SERVICE_UNAVAILABLE),
            10,
        );
-        let error = wait_for_upgrade_write_readiness(&[client], phase, Duration::from_secs(1))
+        let error = wait_for_upgrade_write_readiness(&[client], MIXED_BUCKET, phase, Duration::from_secs(1))
            .await
            .expect_err("persistent unavailability must exhaust the shared deadline");
        // Transport scheduling consumes the same budget; do not require a
@@ -668,10 +860,134 @@ mod upgrade_write_readiness_tests {
        let phase = "stalled";
        let key = format!(".upgrade-readiness/{phase}/node-0");
        target.inject_for_key(FakeTargetOperation::PutObject, &key, FaultAction::Stall(Duration::from_secs(30)), 1);
-        let error = wait_for_upgrade_write_readiness(&[client], phase, Duration::from_secs(1))
+        let error = wait_for_upgrade_write_readiness(std::slice::from_ref(&client), MIXED_BUCKET, phase, Duration::from_secs(1))
            .await
            .expect_err("a stalled request must not outlive the readiness deadline");
        assert!(error.to_string().contains("deadline exceeded during PutObject"), "{error}");
+
+        let part = vec![2_u8; 64 * 1024];
+        let successful_key = "stalled/multipart/02/00";
+        mixed_multipart_before(
+            Instant::now() + MIXED_MULTIPART_UPLOAD_TIMEOUT,
+            2,
+            successful_key,
+            write_multipart(&client, MIXED_BUCKET, successful_key, std::slice::from_ref(&part)),
+        )
+        .await?;
+        assert_eq!(read_object(&client, MIXED_BUCKET, successful_key, None).await?.1, part);
+
+        let stalled_key = "stalled/multipart/02/01";
+        target.inject_for_key(
+            FakeTargetOperation::CompleteMultipartUpload,
+            stalled_key,
+            FaultAction::Stall(Duration::from_secs(30)),
+            1,
+        );
+        let error = {
+            let upload = write_multipart(&client, MIXED_BUCKET, stalled_key, std::slice::from_ref(&part));
+            tokio::pin!(upload);
+            tokio::select! {
+                result = &mut upload => panic!("stalled completion unexpectedly returned: {result:?}"),
+                observed = tokio::time::timeout(Duration::from_secs(10), async {
+                    while target.count_requests(FakeTargetOperation::CompleteMultipartUpload, stalled_key) == 0
+                        || !target.has_object(MIXED_BUCKET, stalled_key)
+                    {
+                        sleep(Duration::from_millis(10)).await;
+                    }
+                }) => observed.expect("multipart completion must reach and commit on the fake peer"),
+            }
+            // An already committed object is not an acknowledged client success.
+            // Expire only after observing the real completion request, so host
+            // scheduling cannot turn this into a timeout before admission.
+            tokio::time::timeout(Duration::from_secs(1), mixed_multipart_before(Instant::now(), 2, stalled_key, upload))
+                .await
+                .expect("expired deadline must stop the in-flight completion")
+                .expect_err("unacknowledged completion must fail even if the peer committed")
+        };
+        assert!(error.to_string().contains("node 2"), "{error}");
+        assert!(error.to_string().contains(stalled_key), "{error}");
+        assert!(error.to_string().contains("acknowledgement deadline exceeded"), "{error}");
+        assert_eq!(target.count_requests(FakeTargetOperation::CompleteMultipartUpload, stalled_key), 1);
+
+        let failure = error
+            .downcast_ref::<MultipartAcknowledgementDeadline>()
+            .expect("the timeout must retain typed writer/object context");
+        assert_eq!(failure.node, 2);
+        assert_eq!(failure.key, stalled_key);
+        let original_message = error.to_string();
+        // These four probe addresses share an in-process peer. This verifies
+        // bounded diagnostic fan-out, not a four-node RustFS cluster.
+        let mut probes = RustFSTestClusterEnvironment::new(MIXED_NODE_COUNT).await?;
+        probes.access_key = FAKE_ACCESS_KEY.to_owned();
+        probes.secret_key = FAKE_SECRET_KEY.to_owned();
+        for node in &mut probes.nodes {
+            node.url = target.endpoint().to_owned();
+        }
+        let log_dir = tempfile::tempdir()?;
+        target.inject_for_key(
+            FakeTargetOperation::HeadObject,
+            stalled_key,
+            FaultAction::Stall(Duration::from_secs(30)),
+            1,
+        );
+        let error = tokio::time::timeout(
+            Duration::from_secs(7),
+            observe_mixed_multipart_failure(&probes, phase, error, Some(log_dir.path())),
+        )
+        .await
+        .expect("the shared diagnostic budget must include artifact writing")
+        .expect_err("observing committed data cannot acknowledge the stalled upload");
+        assert_eq!(error.to_string(), original_message);
+        assert!(error.downcast_ref::<MultipartAcknowledgementDeadline>().is_some());
+        let snapshot: Value =
+            serde_json::from_slice(&tokio::fs::read(log_dir.path().join("mixed-multipart-failure.json")).await?)?;
+        assert_eq!(snapshot["key"], stalled_key);
+        assert_eq!(snapshot["writer_node"], 2);
+        let heads = snapshot["heads"]
+            .as_array()
+            .expect("failure artifact must retain every probe");
+        assert_eq!(heads.len(), MIXED_NODE_COUNT);
+        assert_eq!(
+            heads
+                .iter()
+                .filter(|head| head["timeout"] == true && head["observation"] == "unknown")
+                .count(),
+            1
+        );
+        assert_eq!(
+            heads
+                .iter()
+                .filter(|head| head["status"] == 200 && head["observation"] == "visible")
+                .count(),
+            3
+        );
+        assert_eq!(target.count_requests(FakeTargetOperation::HeadObject, stalled_key), MIXED_NODE_COUNT);
+        assert_eq!(target.count_requests(FakeTargetOperation::CompleteMultipartUpload, stalled_key), 1);
+        assert_eq!(snapshot["top_locks"]["observation"], "unknown", "the fake peer has no admin lock API");
+
+        let locks: MixedLockSnapshot = serde_json::from_value(serde_json::json!({
+            "total": 1, "truncated": true, "capability_note": "partial snapshot",
+            "credentials": { "secret_key": "must not persist" },
+            "locks": [{
+                "resource": format!("{MIXED_BUCKET}/{stalled_key}"), "owner": "writer", "type": "WRITE",
+                "elapsed_secs": 30, "ttl_secs": 1, "merged_config": { "secret_key": "must not persist" },
+            }],
+        }))?;
+        let safe_locks = serde_json::to_value(locks)?;
+        assert_eq!(safe_locks["total"], 1);
+        assert_eq!(safe_locks["truncated"], true);
+        assert_eq!(safe_locks["locks"][0]["owner"], "writer");
+        assert_eq!(safe_locks["locks"][0]["elapsed_secs"], 30);
+        assert!(safe_locks.get("credentials").is_none());
+        assert!(safe_locks["locks"][0].get("merged_config").is_none());
+
+        let mut body = br#"{"locks":[]}"#.to_vec();
+        body.resize(1024 * 1024, b' ');
+        let mut at_limit = reqwest::Response::from(http::Response::new(body.clone()));
+        assert!(read_mixed_lock_snapshot(&mut at_limit).await?.locks.is_empty());
+        body.push(b' ');
+        let mut oversized = reqwest::Response::from(http::Response::new(body));
+        assert_eq!(read_mixed_lock_snapshot(&mut oversized).await.err(), Some("body_limit_exceeded"));
        target.shutdown().await;
        Ok(())
    }
@@ -2181,9 +2497,9 @@ struct LayoutTransport {
    uploaded_parts: Vec<i32>,
    single_puts: usize,
    completes: usize,
-    /// Raw per-key journal in target order: (sequence, operation, part number,
-    /// upload id), so duplicate drives can be told apart from retries.
-    journal: Vec<(u64, String, Option<i32>, Option<String>)>,
+    /// Allowlisted request identity and prepared response metadata. A prepared
+    /// response does not prove that the client received it.
+    journal: Vec<Value>,
 }

 /// Configure every layout bucket to replicate its existing objects to a fresh
@@ -2245,12 +2561,28 @@ async fn replicate_layouts(
            journal: key_requests
                .iter()
                .map(|record| {
-                    (
-                        record.sequence,
-                        format!("{:?}", record.operation),
-                        record.part_number,
-                        record.upload_id.as_ref().map(|id| id.chars().take(12).collect()),
-                    )
+                    serde_json::json!({
+                        "sequence": record.sequence,
+                        "operation": format!("{:?}", record.operation),
+                        "part_number": record.part_number,
+                        "upload_id": record.upload_id,
+                        "requested_version_id": record.version_id,
+                        "source_version_id": record.source_version_id,
+                        "source_mtime": record.source_mtime,
+                        "source_etag": record.source_etag,
+                        "source_replication_request": record.source_replication_request,
+                        "request_content_length": record.content_length,
+                        "journaled_at_unix_millis": record.journaled_at_unix_millis,
+                        "prepared_response": record.prepared_response.as_ref().map(|response| serde_json::json!({
+                            "status": response.status,
+                            "version_id": response.version_id,
+                            "etag": response.etag,
+                            "last_modified": response.last_modified,
+                            "content_length": response.content_length,
+                            "prepared_at_unix_millis": response.prepared_at_unix_millis,
+                            "elapsed_secs": response.elapsed.as_secs_f64(),
+                        })),
+                    })
                })
                .collect(),
        };
@@ -122,6 +122,8 @@ test-util = []
 e2e-test-hooks = []

 [dependencies]
+rustfs-capture-authority = { workspace = true }
+openraft = { workspace = true, features = ["storage-v2"] }
 hotpath.workspace = true
 rustfs-filemeta.workspace = true
 rustfs-utils = { workspace = true, features = ["full"] }
@@ -176,7 +178,7 @@ regex = { workspace = true }
 path-absolutize = { workspace = true }
 rmp.workspace = true
 rmp-serde.workspace = true
-tokio-util = { workspace = true, features = ["io", "compat"] }
+tokio-util = { workspace = true, features = ["io", "compat", "rt"] }
 tokio-stream = { workspace = true, features = ["sync"] }
 hmac = { workspace = true }
 sha2 = { workspace = true }
@@ -240,7 +240,10 @@ pub mod bucket {
    }

    pub mod quota {
-        pub use crate::bucket::quota::{BucketQuota, QuotaCheckResult, QuotaError, QuotaOperation};
+        pub use crate::bucket::quota::{
+            BucketQuota, QUOTA_RESERVATION_PROTOCOL_V1, QUOTA_RESERVATION_PROTOCOL_V2, QuotaCheckResult, QuotaError,
+            QuotaOperation,
+        };

        pub mod checker {
            pub use crate::bucket::quota::checker::QuotaChecker;
@@ -362,11 +365,49 @@ pub mod config {
            ServerConfigCorruptError, ServerConfigSaveResult, ServerConfigSnapshot, delete_config, delete_config_no_lock,
            is_server_config_corrupt_error, lookup_configs, read_config, read_config_no_lock, read_config_with_metadata,
            read_config_without_migrate, read_config_without_migrate_no_lock, read_existing_server_config_no_lock,
-            read_server_config_snapshot, save_config, save_config_no_lock, save_config_with_opts, save_server_config,
-            save_server_config_no_lock, save_server_config_snapshot, save_server_config_snapshot_with_generation,
-            server_config_path, try_migrate_server_config, with_config_object_read_lock, with_config_object_write_lock,
-            with_server_config_read_lock, with_server_config_write_lock,
+            read_server_config_snapshot, save_config, save_config_no_lock, save_config_with_opts, save_config_with_opts_quiet,
+            save_server_config, save_server_config_no_lock, save_server_config_snapshot,
+            save_server_config_snapshot_with_generation, server_config_path, try_migrate_server_config,
+            with_config_object_read_lock, with_config_object_write_lock, with_server_config_read_lock,
+            with_server_config_write_lock,
        };
+
+        /// Observe persisted config bytes and their reader-associated metadata within a byte limit.
+        pub async fn read_config_observation(
+            api: std::sync::Arc<crate::store::ECStore>,
+            file: &str,
+            max_bytes: usize,
+        ) -> crate::error::Result<(Vec<u8>, crate::object_api::ObjectInfo)> {
+            crate::config::com::read_config_observation(api, file, max_bytes).await
+        }
+
+        /// Observe bounded config bytes while the caller retains the object's write owner.
+        pub async fn read_config_observation_no_lock(
+            api: std::sync::Arc<crate::store::ECStore>,
+            file: &str,
+            max_bytes: usize,
+            guard: &crate::object_api::WriteCommitGuard,
+        ) -> crate::error::Result<(Vec<u8>, crate::object_api::ObjectInfo)> {
+            crate::config::com::read_config_observation_no_lock(api, file, max_bytes, guard).await
+        }
+
+        /// Read config bytes and metadata without treating an empty payload as absent.
+        pub async fn read_config_preserve_empty_with_metadata_opts(
+            api: std::sync::Arc<crate::store::ECStore>,
+            file: &str,
+            opts: &crate::object_api::ObjectOptions,
+        ) -> crate::error::Result<(Vec<u8>, crate::object_api::ObjectInfo)> {
+            crate::config::com::read_config_preserve_empty_with_metadata_opts(api, file, opts).await
+        }
+
+        pub async fn read_config_limited_preserve_empty_with_metadata_opts(
+            api: std::sync::Arc<crate::store::ECStore>,
+            file: &str,
+            opts: &crate::object_api::ObjectOptions,
+            max_bytes: usize,
+        ) -> crate::error::Result<(Vec<u8>, crate::object_api::ObjectInfo)> {
+            crate::config::com::read_config_limited_preserve_empty_with_metadata_opts(api, file, opts, max_bytes).await
+        }
    }

    pub mod storageclass {
@@ -468,7 +509,7 @@ pub mod event {
 pub mod global {
    pub use crate::runtime::global::{
        set_global_endpoints, set_global_region, set_global_rustfs_port, set_object_store_resolver, shutdown_background_services,
-        update_erasure_type,
+        update_erasure_type, wait_for_detached_mutations,
    };
 }

@@ -556,6 +597,7 @@ pub mod rio {
 }

 pub mod rpc {
+    pub use crate::cluster::rpc::capture_raft::{CaptureRaftError, CaptureRaftLimits, CaptureRaftNetworkFactory, CaptureRaftRpc};
    pub use crate::cluster::rpc::{
        AuthenticatedChannel, KMS_SIGNAL_SUBSYSTEM, LocalPeerS3Client, MAX_NETWORK_PROBE_BYTES, MAX_NETWORK_PROBE_DURATION,
        NetworkPeerProbeClient, NetworkPeerProbeError, NetworkPeerProbeMeasurement, NetworkPeerTarget, PEER_RESTDRY_RUN,
@@ -602,7 +644,13 @@ pub mod store_list {

 pub mod storage {
    pub use crate::core::pools::{HealLifecycleExpiryContext, POOL_META_NAME};
+    pub use crate::disk::local::native_prepare::{
+        NativePreparationDescription, NativePreparationProbe, NativePreparationReceipt,
+    };
    pub use crate::store::HealWalkVersion;
+    pub use crate::store::native_prepare::{
+        NativePreparationFinalizeRequest, NativePreparationRecovery, NativePreparationTarget,
+    };
    pub use crate::store::{
        BootstrapLocalTarget, ECStore, SCANNER_PUBLICATION_LEASE_TTL_MS, ScannerDataMovementPauseStatus, all_local_disk,
        all_local_disk_path, find_local_disk_by_ref, init_local_disks, init_local_disks_with_instance_ctx, init_lock_clients,
@@ -87,6 +87,7 @@ use rustfs_filemeta::{
    FileInfo, FileInfoOpts, NULL_VERSION_ID, RestoreStatus, RestoreStatusOps, TRANSITION_COMPLETE, get_file_info,
    is_restored_object_on_disk,
 };
+use rustfs_lock::{NamespaceLockGuard, error::LockError};
 use rustfs_scanner_metrics::metrics::{
    IlmAction, Metrics, ScannerLifecycleExpiryStateUpdate, ScannerLifecycleTransitionStateUpdate, global_metrics,
 };
@@ -190,6 +191,8 @@ const TIER_FREE_VERSION_RECOVERY_BASE_INTERVAL: StdDuration = StdDuration::from_
 // before another full sweep can discover work persisted by another node.
 const TIER_FREE_VERSION_RECOVERY_MAX_IDLE_INTERVAL: StdDuration = StdDuration::from_secs(10 * 60);
 const TIER_FREE_VERSION_RECOVERY_JITTER_PERCENT: u64 = 10;
+const TIER_FREE_VERSION_RECOVERY_LEADER_LOCK: &str = "tier-free-version-recovery/leader.lock";
+const TIER_FREE_VERSION_RECOVERY_LOCK_TIMEOUT: StdDuration = StdDuration::from_secs(1);
 const DATE_EXPIRY_EXISTING_OBJECTS_GRACE_SECS: i64 = 5;
 const EXPIRY_WORKER_QUEUE_CAPACITY: usize = 1000;
 const DEFAULT_MANUAL_TRANSITION_JOB_RECOVERY_LIMIT: usize = 100;
@@ -1441,13 +1444,24 @@ pub(crate) async fn enqueue_committed_free_versions(api: &ECStore, free_versions
    queued
 }

-async fn enqueue_recovered_free_version_with_state(state: &Arc<RwLock<ExpiryState>>, oi: ObjectInfo) -> bool {
+async fn enqueue_recovered_free_version_with_state(
+    state: &Arc<RwLock<ExpiryState>>,
+    oi: ObjectInfo,
+    cancel: &CancellationToken,
+) -> bool {
    let task = FreeVersionTask(oi);
    let hash = task.op_hash();
    let (wrkr, stats) = {
-        let state = state.read().await;
+        let state = select! {
+            biased;
+            _ = cancel.cancelled() => return false,
+            state = state.read() => state,
+        };
        (state.get_worker_ch(hash), Arc::clone(&state.stats))
    };
+    if cancel.is_cancelled() {
+        return false;
+    }
    let Some(wrkr) = wrkr else {
        stats.increment_missed_freevers_tasks();
        stats.record_scanner_expiry_state();
@@ -1497,9 +1511,13 @@ fn set_recovered_free_version_enqueue_observer(
    RecoveredFreeVersionEnqueueObserverGuard
 }

-pub async fn enqueue_recovered_free_version(api: &ECStore, oi: ObjectInfo) -> bool {
+pub(super) async fn enqueue_recovered_free_version_with_cancel(
+    api: &ECStore,
+    oi: ObjectInfo,
+    cancel: &CancellationToken,
+) -> bool {
    let expiry_state = api.ctx.expiry_state();
-    let queued = enqueue_recovered_free_version_with_state(&expiry_state, oi).await;
+    let queued = enqueue_recovered_free_version_with_state(&expiry_state, oi, cancel).await;

    #[cfg(test)]
    if let Some(observer) = RECOVERED_FREE_VERSION_ENQUEUE_OBSERVER
@@ -2921,6 +2939,20 @@ fn resolve_tier_free_version_recovery_enabled(value: Result<String, env::VarErro
 }

 fn spawn_tier_free_version_recovery_once(api: Arc<ECStore>, started: &OnceLock<()>) -> Option<JoinHandle<()>> {
+    spawn_tier_free_version_recovery_once_with_options(
+        api,
+        started,
+        TIER_FREE_VERSION_RECOVERY_LEADER_LOCK.to_string(),
+        tier_free_version_recovery_leader_retry_delay,
+    )
+}
+
+fn spawn_tier_free_version_recovery_once_with_options(
+    api: Arc<ECStore>,
+    started: &OnceLock<()>,
+    leader_lock_name: String,
+    retry_delay: fn() -> StdDuration,
+) -> Option<JoinHandle<()>> {
    if !tier_free_version_recovery_enabled() {
        warn!(
            event = EVENT_LIFECYCLE_WORKER_STATE,
@@ -2937,20 +2969,43 @@ fn spawn_tier_free_version_recovery_once(api: Arc<ECStore>, started: &OnceLock<(

    Some(tokio::spawn(async move {
        let cancel_token = api.ctx.background_cancel_token().unwrap_or_default();
-        let expiry_state = api.ctx.expiry_state();
-        run_tier_free_version_recovery_loop(
+        let acquire_api = Arc::clone(&api);
+        let acquire_lock_name = Arc::<str>::from(leader_lock_name);
+        run_tier_free_version_recovery_coordinator(
            cancel_token,
-            expiry_state,
-            jitter_tier_free_version_recovery_delay,
-            move |bucket_marker, object_marker, recovery_cancel| {
+            retry_delay,
+            move || {
+                let api = Arc::clone(&acquire_api);
+                let lock_name = Arc::clone(&acquire_lock_name);
+                async move {
+                    let lock = api.new_ns_lock(RUSTFS_META_BUCKET, &lock_name).await?;
+                    match lock.get_write_lock_quiet(TIER_FREE_VERSION_RECOVERY_LOCK_TIMEOUT).await {
+                        Ok(guard) => Ok(Some(guard)),
+                        Err(LockError::Timeout { .. }) => Ok(None),
+                        Err(err) => Err(err.into()),
+                    }
+                }
+            },
+            move |leader_cancel| {
                let api = Arc::clone(&api);
                async move {
-                    recover_tier_free_versions_with_cancel(
-                        api,
-                        DEFAULT_FREE_VERSION_RECOVERY_LIMIT,
-                        bucket_marker,
-                        object_marker,
-                        recovery_cancel,
+                    run_tier_free_version_recovery_loop(
+                        leader_cancel,
+                        api.ctx.expiry_state(),
+                        jitter_tier_free_version_recovery_delay,
+                        move |bucket_marker, object_marker, recovery_cancel| {
+                            let api = Arc::clone(&api);
+                            async move {
+                                recover_tier_free_versions_with_cancel(
+                                    api,
+                                    DEFAULT_FREE_VERSION_RECOVERY_LIMIT,
+                                    bucket_marker,
+                                    object_marker,
+                                    recovery_cancel,
+                                )
+                                .await
+                            }
+                        },
                    )
                    .await
                }
@@ -2960,6 +3015,86 @@ fn spawn_tier_free_version_recovery_once(api: Arc<ECStore>, started: &OnceLock<(
    }))
 }

+fn tier_free_version_recovery_leader_retry_delay() -> StdDuration {
+    StdDuration::from_secs(rand::rng().random_range(15..=30))
+}
+
+async fn run_tier_free_version_recovery_coordinator<A, AF, R, RF>(
+    cancel: CancellationToken,
+    retry_delay: fn() -> StdDuration,
+    mut acquire: A,
+    mut recover: R,
+) where
+    A: FnMut() -> AF,
+    AF: Future<Output = crate::error::Result<Option<NamespaceLockGuard>>>,
+    R: FnMut(CancellationToken) -> RF,
+    RF: Future<Output = ()>,
+{
+    loop {
+        let acquired = select! {
+            biased;
+            _ = cancel.cancelled() => return,
+            result = acquire() => result,
+        };
+        match acquired {
+            Ok(Some(mut guard)) => {
+                let leader_cancel = cancel.child_token();
+                // Hold ownership across every page, cooldown and idle wait. A
+                // follower taking over starts a fresh sweep of persisted markers.
+                let recovery = recover(leader_cancel.clone());
+                tokio::pin!(recovery);
+                debug!(
+                    event = EVENT_LIFECYCLE_WORKER_STATE,
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_LIFECYCLE,
+                    state = "free_version_recovery_leader_acquired",
+                    lock_name = TIER_FREE_VERSION_RECOVERY_LEADER_LOCK,
+                    "Tier free-version recovery leader acquired"
+                );
+                select! {
+                    biased;
+                    _ = cancel.cancelled() => {
+                        leader_cancel.cancel();
+                        recovery.await;
+                    }
+                    _ = guard.lock_lost_notified() => {
+                        debug!(
+                            event = EVENT_LIFECYCLE_WORKER_STATE,
+                            component = LOG_COMPONENT_ECSTORE,
+                            subsystem = LOG_SUBSYSTEM_LIFECYCLE,
+                            state = "free_version_recovery_leader_lost",
+                            lock_name = TIER_FREE_VERSION_RECOVERY_LEADER_LOCK,
+                            "Tier free-version recovery leader lost"
+                        );
+                        // Dropping the recovery future would detach its active
+                        // walk. Cancel it and poll the same future through cleanup.
+                        leader_cancel.cancel();
+                        recovery.await;
+                    }
+                    _ = &mut recovery => {}
+                }
+                leader_cancel.cancel();
+                guard.release();
+            }
+            Ok(None) => {}
+            Err(err) => warn!(
+                event = EVENT_LIFECYCLE_WORKER_STATE,
+                component = LOG_COMPONENT_ECSTORE,
+                subsystem = LOG_SUBSYSTEM_LIFECYCLE,
+                state = "free_version_recovery_leader_acquire_failed",
+                lock_name = TIER_FREE_VERSION_RECOVERY_LEADER_LOCK,
+                error = %err,
+                "Tier free-version recovery leader acquisition failed"
+            ),
+        }
+        select! {
+            biased;
+            _ = cancel.cancelled() => return,
+            _ = tokio::time::sleep(retry_delay()) => {}
+        }
+    }
+}
+
 async fn run_tier_free_version_recovery_loop<F, Fut>(
    cancel_token: CancellationToken,
    expiry_state: Arc<RwLock<ExpiryState>>,
@@ -2969,7 +3104,11 @@ async fn run_tier_free_version_recovery_loop<F, Fut>(
    F: FnMut(Option<String>, Option<String>, CancellationToken) -> Fut,
    Fut: Future<Output = crate::error::Result<FreeVersionRecoveryStats>>,
 {
-    let recovery_notify = Arc::clone(&expiry_state.read().await.recovery_notify);
+    let recovery_notify = select! {
+        biased;
+        _ = cancel_token.cancelled() => return,
+        state = expiry_state.read() => Arc::clone(&state.recovery_notify),
+    };
    let mut schedule = TierFreeVersionRecoverySchedule::default();

    loop {
@@ -2992,7 +3131,11 @@ async fn run_tier_free_version_recovery_loop<F, Fut>(
                schedule.record_success(&stats);
                rustfs_io_metrics::record_stage_duration("lifecycle_free_version_recovery", elapsed.as_secs_f64() * 1000.0);
                let (pending_tasks, active_tasks) = {
-                    let state = expiry_state.read().await;
+                    let state = select! {
+                        biased;
+                        _ = cancel_token.cancelled() => return,
+                        state = expiry_state.read() => state,
+                    };
                    (state.pending_tasks(), state.stats.active_tasks())
                };
                debug!(
@@ -4748,7 +4891,7 @@ async fn expire_transitioned_object_with_lock_lost_signal(
    bucket_incarnation_id: Uuid,
    lock_lost_signal: Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
 ) -> Result<ObjectInfo, std::io::Error> {
-    let publication_guard = lifecycle_expiry_publication_guard(&api, oi, bucket_incarnation_id)
+    let publication_guard = lifecycle_expiry_publication_guard(&api, &oi.bucket, &oi.name, bucket_incarnation_id)
        .await
        .ok_or_else(|| std::io::Error::other("lifecycle expiry is not allowed for this bucket"))?;
    let snapshot = lifecycle_delete_config_snapshot(&api, oi)
@@ -4838,10 +4981,14 @@ pub fn gen_transition_objname(bucket: &str) -> Result<String, Error> {
 pub async fn transition_object(api: Arc<ECStore>, oi: &ObjectInfo, lae: LcAuditEvent) -> Result<(), Error> {
    let time_ilm = Metrics::time_ilm(lae.event.action);

+    let (publication_guard, bucket_incarnation_id) = lifecycle_transition_publication_guard(&api, oi)
+        .await
+        .ok_or_else(|| Error::other("lifecycle transition is not allowed for this bucket"))?;
+
    let etag = if let Some(etag) = &oi.etag { etag } else { "" };
    let etag = etag.to_string();

-    let opts = ObjectOptions {
+    let mut opts = ObjectOptions {
        transition: TransitionOptions {
            status: lifecycle::TRANSITION_PENDING.to_string(),
            tier: lae.event.storage_class,
@@ -4853,8 +5000,10 @@ pub async fn transition_object(api: Arc<ECStore>, oi: &ObjectInfo, lae: LcAuditE
        versioned: BucketVersioningSys::prefix_enabled(&oi.bucket, &oi.name).await,
        version_suspended: BucketVersioningSys::prefix_suspended(&oi.bucket, &oi.name).await,
        mod_time: oi.mod_time,
+        expected_bucket_incarnation_id: Some(bucket_incarnation_id),
        ..Default::default()
    };
+    opts.add_namespace_lock_guard(&publication_guard);
    let result = api.transition_object(&oi.bucket, &oi.name, &opts).await;
    time_ilm(1)();
    result
@@ -5458,20 +5607,26 @@ pub async fn apply_transition_rule(event: &lifecycle::Event, src: &LcEventSrc, o
        .await
 }

-async fn lifecycle_expiry_publication_guard(
+async fn acquire_lifecycle_publication_guard(api: &ECStore, bucket: &str) -> Result<rustfs_lock::NamespaceLockGuard, Error> {
+    let lock = api
+        .new_ns_lock(bucket, rustfs_common::table_catalog::TABLE_BUCKET_PUBLICATION_LOCK_PATH)
+        .await?;
+    let guard = lock.get_read_lock(get_lock_acquire_timeout()).await.map_err(Error::other)?;
+    if guard.is_lock_lost() {
+        return Err(Error::other("table-bucket publication lock was lost before lifecycle admission"));
+    }
+    Ok(guard)
+}
+
+pub(crate) async fn lifecycle_expiry_publication_guard(
    api: &ECStore,
-    oi: &ObjectInfo,
+    bucket: &str,
+    object: &str,
    bucket_incarnation_id: Uuid,
 ) -> Option<rustfs_lock::NamespaceLockGuard> {
-    let result = async {
-        let lock = api
-            .new_ns_lock(&oi.bucket, rustfs_common::table_catalog::TABLE_BUCKET_PUBLICATION_LOCK_PATH)
-            .await?;
-        let guard = lock.get_read_lock(get_lock_acquire_timeout()).await.map_err(Error::other)?;
-        if guard.is_lock_lost() {
-            return Err(Error::other("table-bucket publication lock was lost before lifecycle delete admission"));
-        }
-        if !metadata_boundary::lifecycle_expiry_allowed(api, &oi.bucket, bucket_incarnation_id).await? {
+    let result: std::result::Result<Option<rustfs_lock::NamespaceLockGuard>, Error> = async {
+        let guard = acquire_lifecycle_publication_guard(api, bucket).await?;
+        if !metadata_boundary::lifecycle_expiry_allowed(api, bucket, bucket_incarnation_id).await? {
            return Ok(None);
        }
        Ok(Some(guard))
@@ -5484,8 +5639,8 @@ async fn lifecycle_expiry_publication_guard(
                event = EVENT_LIFECYCLE_DELETE_FAILED,
                component = LOG_COMPONENT_ECSTORE,
                subsystem = LOG_SUBSYSTEM_LIFECYCLE,
-                bucket = %oi.bucket,
-                object = %oi.name,
+                bucket,
+                object,
                operation = "authorize_lifecycle_expiry",
                error = %err,
                "Lifecycle delete admission failed"
@@ -5495,6 +5650,34 @@ async fn lifecycle_expiry_publication_guard(
    }
 }

+async fn lifecycle_transition_publication_guard(
+    api: &ECStore,
+    oi: &ObjectInfo,
+) -> Option<(rustfs_lock::NamespaceLockGuard, Uuid)> {
+    let result: std::result::Result<(rustfs_lock::NamespaceLockGuard, Uuid), Error> = async {
+        let guard = acquire_lifecycle_publication_guard(api, &oi.bucket).await?;
+        let configs = metadata_boundary::get_expiry_configs(api, &oi.bucket).await?;
+        Ok((guard, configs.bucket_incarnation_id))
+    }
+    .await;
+    match result {
+        Ok(state) => Some(state),
+        Err(err) => {
+            warn!(
+                event = EVENT_LIFECYCLE_TIER_OPERATION_FAILED,
+                component = LOG_COMPONENT_ECSTORE,
+                subsystem = LOG_SUBSYSTEM_LIFECYCLE,
+                bucket = %oi.bucket,
+                object = %oi.name,
+                operation = "authorize_lifecycle_transition",
+                error = %err,
+                "Lifecycle transition admission failed"
+            );
+            None
+        }
+    }
+}
+
 pub async fn apply_expiry_on_transitioned_object(
    api: Arc<ECStore>,
    oi: &ObjectInfo,
@@ -5551,7 +5734,8 @@ async fn apply_expiry_on_non_transitioned_objects_with_lock_lost_signal(
    bucket_incarnation_id: Uuid,
    lock_lost_signal: Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>>,
 ) -> bool {
-    let Some(publication_guard) = lifecycle_expiry_publication_guard(&api, oi, bucket_incarnation_id).await else {
+    let Some(publication_guard) = lifecycle_expiry_publication_guard(&api, &oi.bucket, &oi.name, bucket_incarnation_id).await
+    else {
        return false;
    };
    let snapshot = match lifecycle_delete_config_snapshot(&api, oi).await {
@@ -5611,8 +5795,10 @@ async fn apply_expiry_on_non_transitioned_objects_with_lock_lost_signal(
    } else {
        None
    };
+    // The delete owns the publication guard from here on, so a cancelled
+    // expiry worker cannot let table-bucket publication overtake it.
    let mut dobj = match api
-        .delete_object_with_tier_delete_journal(&oi.bucket, &encode_dir_object(&oi.name), opts)
+        .delete_object_with_tier_delete_journal_and_guards(&oi.bucket, &encode_dir_object(&oi.name), opts, [publication_guard])
        .await
    {
        Ok(dobj) => dobj,
@@ -5844,6 +6030,7 @@ pub async fn apply_lifecycle_action(event: &lifecycle::Event, src: &LcEventSrc,
 #[cfg(test)]
 mod tests {
    use super::expiry_worker_count;
+    use super::lifecycle_transition_publication_guard;
    use super::{
        DATE_EXPIRY_EXISTING_OBJECTS_GRACE_SECS, DEFAULT_TRANSITION_QUEUE_CAPACITY, DEFAULT_TRANSITION_WORKERS_ABSOLUTE_MAX,
        DEFAULT_TRANSITION_WORKERS_CAP, EVENT_LIFECYCLE_EVALUATION_FAILED, EVENT_LIFECYCLE_EXPIRED_DETECTED,
@@ -5862,10 +6049,10 @@ mod tests {
        merge_stale_multipart_candidate, persist_manual_transition_job_progress_if_owned,
        persist_manual_transition_page_checkpoint, recover_manual_transition_job, recover_manual_transition_jobs,
        resolve_tier_free_version_recovery_enabled, resolve_transition_queue_capacity, resolve_transition_queue_send_timeout,
-        resolve_transition_worker_count, resolve_transition_workers_absolute_max, run_tier_free_version_recovery_loop,
-        select_restore_s3_location, set_lifecycle_observability_observer, set_recovered_free_version_enqueue_observer,
-        should_defer_date_expiry_for_recent_config_update, transitioned_cleanup_tuple, transitioned_object_delete_opts,
-        wait_for_tier_free_version_recovery,
+        resolve_transition_worker_count, resolve_transition_workers_absolute_max, run_tier_free_version_recovery_coordinator,
+        run_tier_free_version_recovery_loop, select_restore_s3_location, set_lifecycle_observability_observer,
+        set_recovered_free_version_enqueue_observer, should_defer_date_expiry_for_recent_config_update,
+        transitioned_cleanup_tuple, transitioned_object_delete_opts, wait_for_tier_free_version_recovery,
    };
    #[cfg(feature = "test-util")]
    use super::{delete_free_version_remote_object_then, encode_dir_object, get_transitioned_object_reader_with_tier_manager};
@@ -5959,6 +6146,231 @@ mod tests {
    use tokio_util::sync::CancellationToken;
    use uuid::Uuid;

+    fn recovery_test_lock() -> (Arc<rustfs_lock::NamespaceLock>, Vec<Arc<dyn rustfs_lock::LockClient>>) {
+        let clients: Vec<Arc<dyn rustfs_lock::LockClient>> = (0..3)
+            .map(|_| {
+                Arc::new(rustfs_lock::LocalClient::with_manager(Arc::new(rustfs_lock::GlobalLockManager::new())))
+                    as Arc<dyn rustfs_lock::LockClient>
+            })
+            .collect();
+        let lock =
+            rustfs_lock::NamespaceLock::with_clients_and_quorum(format!("recovery-test-{}", Uuid::new_v4()), clients.clone(), 2);
+        (Arc::new(lock), clients)
+    }
+
+    async fn acquire_recovery_test_lock(
+        lock: Arc<rustfs_lock::NamespaceLock>,
+        owner: String,
+    ) -> crate::error::Result<Option<rustfs_lock::NamespaceLockGuard>> {
+        // Keep this lease longer than the virtual-time cooldown/idle advances;
+        // production uses the namespace wrapper's renewable 30-second lease.
+        lock.lock_guard(
+            rustfs_lock::ObjectKey::new(crate::disk::RUSTFS_META_BUCKET, super::TIER_FREE_VERSION_RECOVERY_LEADER_LOCK),
+            &owner,
+            StdDuration::from_millis(5),
+            StdDuration::from_secs(3600),
+        )
+        .await
+        .map_err(Into::into)
+    }
+
+    fn recovery_test_retry_delay() -> StdDuration {
+        StdDuration::from_secs(15)
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial]
+    async fn tier_free_version_recovery_eight_coordinators_hold_one_leader_and_take_over() {
+        let (lock, _clients) = recovery_test_lock();
+        let (pages_tx, mut pages_rx) = tokio::sync::mpsc::unbounded_channel();
+        let mut tasks = Vec::new();
+        let mut cancels = Vec::new();
+        for node in 0..8 {
+            let cancel = CancellationToken::new();
+            cancels.push(cancel.clone());
+            let lock = Arc::clone(&lock);
+            let pages_tx = pages_tx.clone();
+            tasks.push(tokio::spawn(async move {
+                run_tier_free_version_recovery_coordinator(
+                    cancel,
+                    recovery_test_retry_delay,
+                    move || acquire_recovery_test_lock(Arc::clone(&lock), format!("node-{node}")),
+                    move |leader_cancel| {
+                        let pages_tx = pages_tx.clone();
+                        async move {
+                            let mut pages = 0;
+                            run_tier_free_version_recovery_loop(
+                                leader_cancel,
+                                ExpiryState::new(),
+                                std::convert::identity,
+                                move |bucket, object, _| {
+                                    pages += 1;
+                                    pages_tx.send((node, bucket, object)).expect("page observer should be open");
+                                    std::future::ready(Ok(free_version_recovery_stats(0, 0, pages == 1)))
+                                },
+                            )
+                            .await;
+                        }
+                    },
+                )
+                .await;
+            }));
+        }
+        let (leader, bucket, object) = pages_rx.recv().await.expect("one coordinator should start recovery");
+        assert!(bucket.is_none() && object.is_none());
+        // The same leader must survive both a truncated page's cooldown and
+        // complete idle sweeps. All other coordinators keep probing the lock.
+        let mut observed_pages = 1;
+        for _ in 0..27 {
+            tokio::time::advance(StdDuration::from_secs(60)).await;
+            for _ in 0..100 {
+                tokio::task::yield_now().await;
+            }
+            while let Ok((node, bucket, object)) = pages_rx.try_recv() {
+                assert_eq!(node, leader, "followers must not repeat a namespace sweep");
+                if observed_pages == 1 {
+                    assert_eq!(bucket.as_deref(), Some("bucket"));
+                    assert_eq!(object.as_deref(), Some("object"));
+                } else {
+                    assert!(bucket.is_none() && object.is_none(), "complete sweeps should reset their cursors");
+                }
+                observed_pages += 1;
+            }
+        }
+        assert!(observed_pages >= 7, "the leader should finish a full ten-minute idle wait");
+        cancels[leader].cancel();
+        let (successor, bucket, object) = tokio::time::timeout(StdDuration::from_secs(60), pages_rx.recv())
+            .await
+            .expect("a follower should take over after leader shutdown")
+            .expect("page observer should remain open");
+        assert_ne!(successor, leader);
+        assert!(bucket.is_none() && object.is_none(), "a new leader must restart discovery safely");
+        for cancel in cancels {
+            cancel.cancel();
+        }
+        for task in tasks {
+            task.await.expect("all coordinators should shut down cleanly");
+        }
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial]
+    async fn tier_free_version_recovery_lock_loss_waits_for_active_page_cleanup() {
+        let (lock, _clients) = recovery_test_lock();
+        let cancel = CancellationToken::new();
+        let loop_cancel = cancel.clone();
+        let (page_tx, mut page_rx) = tokio::sync::mpsc::unbounded_channel();
+        let cleanup_started = Arc::new(tokio::sync::Notify::new());
+        let finish_cleanup = Arc::new(tokio::sync::Notify::new());
+        let observed_cleanup = Arc::clone(&cleanup_started);
+        let release_cleanup = Arc::clone(&finish_cleanup);
+        let task = tokio::spawn(async move {
+            run_tier_free_version_recovery_coordinator(
+                loop_cancel,
+                recovery_test_retry_delay,
+                move || acquire_recovery_test_lock(Arc::clone(&lock), "old-leader".to_string()),
+                move |leader_cancel| {
+                    let page_tx = page_tx.clone();
+                    let cleanup_started = Arc::clone(&cleanup_started);
+                    let finish_cleanup = Arc::clone(&finish_cleanup);
+                    async move {
+                        run_tier_free_version_recovery_loop(
+                            leader_cancel,
+                            ExpiryState::new(),
+                            std::convert::identity,
+                            move |_, _, page_cancel| {
+                                let cleanup_started = Arc::clone(&cleanup_started);
+                                let finish_cleanup = Arc::clone(&finish_cleanup);
+                                page_tx.send(()).expect("page observer should remain open");
+                                async move {
+                                    page_cancel.cancelled().await;
+                                    cleanup_started.notify_one();
+                                    finish_cleanup.notified().await;
+                                    Err(std::io::Error::new(std::io::ErrorKind::Interrupted, "cancelled page").into())
+                                }
+                            },
+                        )
+                        .await;
+                    }
+                },
+            )
+            .await;
+        });
+        page_rx.recv().await.expect("leader should start an active page");
+        // Let authoritative backend leases expire before the next heartbeat
+        // can renew them. A failed refresh must cancel the active page.
+        tokio::time::advance(StdDuration::from_secs(3600)).await;
+        tokio::time::timeout(StdDuration::from_secs(1), observed_cleanup.notified())
+            .await
+            .expect("expired backend leases must trigger recovery cancellation");
+        assert!(!task.is_finished(), "lock loss must wait for the active page's cleanup");
+        assert!(page_rx.try_recv().is_err(), "the old leader must not start another page");
+        cancel.cancel();
+        release_cleanup.notify_one();
+        task.await.expect("coordinator should finish cleanup before exiting");
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn tier_free_version_recovery_coordinator_retries_without_scanning_after_lock_error() {
+        let cancel = CancellationToken::new();
+        let loop_cancel = cancel.clone();
+        let attempts = Arc::new(AtomicUsize::new(0));
+        let observed_attempts = Arc::clone(&attempts);
+        let task = tokio::spawn(async move {
+            run_tier_free_version_recovery_coordinator(
+                loop_cancel,
+                recovery_test_retry_delay,
+                move || {
+                    observed_attempts.fetch_add(1, Ordering::SeqCst);
+                    std::future::ready(Err(rustfs_lock::LockError::InsufficientNodes {
+                        required: 2,
+                        available: 1,
+                    }
+                    .into()))
+                },
+                |_| async { panic!("a node without ownership must never start discovery") },
+            )
+            .await;
+        });
+        tokio::task::yield_now().await;
+        assert_eq!(attempts.load(Ordering::SeqCst), 1);
+        tokio::time::advance(StdDuration::from_secs(14)).await;
+        assert_eq!(attempts.load(Ordering::SeqCst), 1, "lock errors must not cause a tight retry loop");
+        tokio::time::advance(StdDuration::from_secs(1)).await;
+        tokio::task::yield_now().await;
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        cancel.cancel();
+        task.await.expect("a follower should respond to shutdown during retry wait");
+    }
+
+    #[tokio::test]
+    async fn tier_free_version_recovery_enqueue_cancels_while_waiting_for_state_lock() {
+        let state = ExpiryState::new();
+        let (tx, mut rx) = tokio::sync::mpsc::channel(1);
+        let mut held = state.write().await;
+        held.tasks_tx.push(tx);
+        let cancel = CancellationToken::new();
+        let enqueue_cancel = cancel.clone();
+        let enqueue_state = Arc::clone(&state);
+        let enqueue = tokio::spawn(async move {
+            enqueue_recovered_free_version_with_state(&enqueue_state, ObjectInfo::default(), &enqueue_cancel).await
+        });
+        tokio::task::yield_now().await;
+        assert!(!enqueue.is_finished(), "enqueue should be blocked on the state lock");
+        cancel.cancel();
+        let queued = tokio::time::timeout(StdDuration::from_secs(1), enqueue)
+            .await
+            .expect("cancellation must not wait for the state write guard")
+            .expect("enqueue should return cleanly");
+        assert!(!queued);
+        assert_eq!(held.pending_tasks(), 0);
+        assert_eq!(held.stats.missed_free_vers_tasks(), 0);
+        assert!(rx.try_recv().is_err(), "a cancelled recovery must not enqueue a task");
+        drop(held);
+        assert!(!enqueue_recovered_free_version_with_state(&state, ObjectInfo::default(), &cancel).await);
+        assert!(rx.try_recv().is_err(), "cancellation must win even when the lock is available");
+    }
+
    fn free_version_recovery_stats(enqueued: usize, failed: usize, truncated: bool) -> FreeVersionRecoveryStats {
        FreeVersionRecoveryStats {
            scanned: enqueued.saturating_add(failed),
@@ -7638,7 +8050,7 @@ mod tests {
        assert_eq!(state.read().await.stats.pending_tasks(), 1);

        // The single-slot queue is full for both enqueue paths.
-        assert!(!enqueue_recovered_free_version_with_state(&state, oi.clone()).await);
+        assert!(!enqueue_recovered_free_version_with_state(&state, oi.clone(), &CancellationToken::new()).await);
        assert_eq!(state.read().await.stats.pending_tasks(), 1);
        assert_eq!(state.read().await.stats.missed_free_vers_tasks(), 1);

@@ -7670,7 +8082,7 @@ mod tests {
            ..Default::default()
        };

-        let queued = enqueue_recovered_free_version_with_state(&state, oi).await;
+        let queued = enqueue_recovered_free_version_with_state(&state, oi, &CancellationToken::new()).await;
        let state = state.read().await;

        assert!(!queued);
@@ -7799,8 +8211,8 @@ mod tests {
            ..Default::default()
        };

-        let first = enqueue_recovered_free_version_with_state(&state, oi.clone()).await;
-        let second = enqueue_recovered_free_version_with_state(&state, oi).await;
+        let first = enqueue_recovered_free_version_with_state(&state, oi.clone(), &CancellationToken::new()).await;
+        let second = enqueue_recovered_free_version_with_state(&state, oi, &CancellationToken::new()).await;
        let state = state.read().await;

        assert!(first);
@@ -7901,7 +8313,7 @@ mod tests {
        };

        assert!(
-            super::enqueue_recovered_free_version(&ecstore, oi).await,
+            super::enqueue_recovered_free_version_with_cancel(&ecstore, oi, &CancellationToken::new()).await,
            "the resized production worker queue should accept the task"
        );
        stop_tx.send(None).await.expect("worker stop signal should be delivered");
@@ -13028,6 +13440,171 @@ mod tests {
        }
    }

+    #[tokio::test]
+    #[serial]
+    async fn queued_noncurrent_batch_expiry_does_not_delete_from_table_bucket() {
+        let (_disk_paths, ecstore) = setup_test_env().await;
+        let bucket = format!("table-noncurrent-lifecycle-{}", Uuid::new_v4().simple());
+        let object = "tables/table-id/data/part-00001.parquet";
+        create_test_bucket(&ecstore, &bucket).await;
+        metadata_sys::update_in(
+            &ecstore.ctx,
+            &bucket,
+            BUCKET_VERSIONING_CONFIG,
+            b"<VersioningConfiguration><Status>Enabled</Status></VersioningConfiguration>".to_vec(),
+        )
+        .await
+        .expect("bucket versioning should be enabled");
+
+        let mut old_reader = PutObjReader::from_vec(b"old table data".to_vec());
+        let old = ecstore
+            .put_object(
+                &bucket,
+                object,
+                &mut old_reader,
+                &ObjectOptions {
+                    versioned: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("the historical table version should be created");
+        let old_version_id = old.version_id.expect("the historical version should have an identity");
+        let mut current_reader = PutObjReader::from_vec(b"current table data".to_vec());
+        ecstore
+            .put_object(
+                &bucket,
+                object,
+                &mut current_reader,
+                &ObjectOptions {
+                    versioned: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("the current table version should be created");
+
+        let incarnation = ecstore
+            .bucket_incarnation_id_from_disk(&bucket)
+            .await
+            .expect("bucket incarnation should be available");
+        let publication_lock = ecstore
+            .new_ns_lock(&bucket, rustfs_common::table_catalog::TABLE_BUCKET_PUBLICATION_LOCK_PATH)
+            .await
+            .expect("table-bucket publication lock should be created");
+        let enable_guard = publication_lock
+            .get_write_lock(get_lock_acquire_timeout())
+            .await
+            .expect("table-bucket enablement should acquire the publication lock");
+
+        let delete_store = ecstore.clone();
+        let delete_bucket = bucket.clone();
+        let delete_object = object.to_string();
+        let mut delete = tokio::spawn(async move {
+            let target = ObjectToDelete {
+                object_name: delete_object,
+                version_id: Some(old_version_id),
+                ..Default::default()
+            };
+            crate::bucket::lifecycle::object_handlers_common::delete_object_versions(
+                &delete_store,
+                &delete_bucket,
+                std::slice::from_ref(&target),
+                lifecycle::Event::default(),
+                incarnation,
+            )
+            .await
+        });
+        tokio::task::yield_now().await;
+        assert!(
+            tokio::time::timeout(StdDuration::from_millis(100), &mut delete)
+                .await
+                .is_err(),
+            "noncurrent lifecycle cleanup must wait for the table-bucket publication fence"
+        );
+
+        let sys = metadata_sys::bucket_metadata_sys_of(&ecstore.ctx).expect("metadata system should be initialized");
+        let sys = sys.read().await.clone();
+        let mut metadata = (*sys.get(&bucket).await.expect("bucket metadata should exist")).clone();
+        metadata.table_bucket_config_json = br#"{"enabled":true}"#.to_vec();
+        sys.persist_and_set(metadata)
+            .await
+            .expect("table bucket marker should be persisted");
+        sys.reload_from_store(&bucket)
+            .await
+            .expect("table bucket marker should become authoritative");
+        drop(enable_guard);
+
+        let failed = tokio::time::timeout(StdDuration::from_secs(2), delete)
+            .await
+            .expect("noncurrent lifecycle cleanup should resume after the fence is released")
+            .expect("noncurrent lifecycle cleanup task should join");
+        assert_eq!(failed, 1, "table-bucket noncurrent cleanup must be rejected");
+        assert!(
+            ecstore
+                .get_object_info(
+                    &bucket,
+                    object,
+                    &ObjectOptions {
+                        version_id: Some(old_version_id.to_string()),
+                        versioned: true,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .is_ok(),
+            "table data must remain after queued noncurrent cleanup is rejected"
+        );
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn lifecycle_transition_admission_holds_publication_reader() {
+        let (_disk_paths, ecstore) = setup_test_env().await;
+        let bucket = format!("lifecycle-transition-{}", Uuid::new_v4().simple());
+        let object = "tables/table-id/data/part-00001.parquet";
+        create_test_bucket(&ecstore, &bucket).await;
+        let object_info = ObjectInfo {
+            bucket: bucket.clone(),
+            name: object.to_string(),
+            ..Default::default()
+        };
+
+        let publication_lock = ecstore
+            .new_ns_lock(&bucket, rustfs_common::table_catalog::TABLE_BUCKET_PUBLICATION_LOCK_PATH)
+            .await
+            .expect("table-bucket publication lock should be created");
+        let enable_guard = publication_lock
+            .get_write_lock(get_lock_acquire_timeout())
+            .await
+            .expect("publication writer should be acquired");
+        let admission_store = ecstore.clone();
+        let admission_object = object_info.clone();
+        let mut admission =
+            tokio::spawn(async move { lifecycle_transition_publication_guard(&admission_store, &admission_object).await });
+        assert!(
+            tokio::time::timeout(StdDuration::from_millis(100), &mut admission)
+                .await
+                .is_err(),
+            "transition admission must wait for an in-progress table publication"
+        );
+        drop(enable_guard);
+
+        let (transition_guard, incarnation) = tokio::time::timeout(StdDuration::from_secs(2), admission)
+            .await
+            .expect("transition admission should resume after publication")
+            .expect("transition admission task should join")
+            .expect("transition admission should succeed");
+        assert_ne!(incarnation, Uuid::nil(), "transition admission must bind a bucket incarnation");
+        let blocked_writer = publication_lock.get_write_lock(StdDuration::from_millis(100)).await;
+        assert!(blocked_writer.is_err(), "transition admission must retain the publication read lock");
+        drop(transition_guard);
+        publication_lock
+            .get_write_lock(get_lock_acquire_timeout())
+            .await
+            .expect("publication writer should proceed after transition admission releases its guard");
+    }
+
    #[tokio::test]
    #[serial]
    async fn queued_delete_all_rechecks_a_same_id_rule_moved_into_the_future() {
@@ -13630,6 +14207,10 @@ mod tests {
        .await;
    }

+    fn fast_tier_free_version_recovery_leader_retry_delay() -> StdDuration {
+        StdDuration::from_millis(10)
+    }
+
    #[tokio::test]
    #[serial]
    async fn tier_free_version_recovery_production_entrypoint_enqueues_seeded_item() {
@@ -13645,8 +14226,13 @@ mod tests {
        seed_recoverable_free_version(&disk_paths, &bucket, object, None, None).await;

        let started = OnceLock::new();
-        let recovery = super::spawn_tier_free_version_recovery_once(Arc::clone(&ecstore), &started)
-            .expect("production recovery entrypoint should start once");
+        let recovery = super::spawn_tier_free_version_recovery_once_with_options(
+            Arc::clone(&ecstore),
+            &started,
+            format!("tier-free-version-recovery/test/{}", Uuid::new_v4()),
+            fast_tier_free_version_recovery_leader_retry_delay,
+        )
+        .expect("production recovery entrypoint should start once");
        let task = tokio::time::timeout(StdDuration::from_secs(30), async { recovery_rx.lock().await.recv().await })
            .await
            .expect("production recovery entrypoint should enqueue the seeded item")
@@ -13689,13 +14275,14 @@ mod tests {
        };
        let mut recovery_rx = recovery_rx.lock().await;
        assert!(
-            super::enqueue_recovered_free_version(
+            super::enqueue_recovered_free_version_with_cancel(
                &ecstore,
                ObjectInfo {
                    bucket: "prefill".to_string(),
                    name: "prefill".to_string(),
                    ..Default::default()
                },
+                &CancellationToken::new(),
            )
            .await,
            "the production recovery queue should accept its first task"
@@ -36,6 +36,10 @@ pub async fn delete_object_versions(
    _lc_event: lifecycle::Event,
    bucket_incarnation_id: Uuid,
 ) -> usize {
+    if to_del.is_empty() {
+        return 0;
+    }
+
    if to_del.iter().any(|target| {
        target.version_id.is_none()
            || (target.version_id.is_some_and(|version_id| version_id.is_nil()) && target.expected_identity.is_none())
@@ -52,6 +56,25 @@ pub async fn delete_object_versions(
        return to_del.len();
    }

+    let Some(publication_guard) = super::bucket_lifecycle_ops::lifecycle_expiry_publication_guard(
+        api,
+        bucket,
+        to_del[0].object_name.as_str(),
+        bucket_incarnation_id,
+    )
+    .await
+    else {
+        debug!(
+            event = EVENT_LIFECYCLE_CLEANUP_SKIPPED,
+            component = LOG_COMPONENT_ECSTORE,
+            subsystem = LOG_SUBSYSTEM_LIFECYCLE,
+            bucket,
+            target_count = to_del.len(),
+            reason = "publication_admission_rejected",
+            "Skipped lifecycle noncurrent version cleanup"
+        );
+        return to_del.len();
+    };
    let delete_config_snapshot = match ReplicationObjectBridge::delete_request_config(api, bucket).await {
        Ok(snapshot) => Arc::new(snapshot),
        Err(err) => {
@@ -78,17 +101,13 @@ pub async fn delete_object_versions(
            remaining = &[];
        }

-        let (mut deleted_objs, errors) = api
-            .delete_objects(
-                bucket,
-                to_del.to_vec(),
-                ObjectOptions {
-                    delete_replication_config_snapshot: Some(Arc::clone(&delete_config_snapshot)),
-                    expected_bucket_incarnation_id: Some(bucket_incarnation_id),
-                    ..Default::default()
-                },
-            )
-            .await;
+        let mut delete_options = ObjectOptions {
+            delete_replication_config_snapshot: Some(Arc::clone(&delete_config_snapshot)),
+            expected_bucket_incarnation_id: Some(bucket_incarnation_id),
+            ..Default::default()
+        };
+        delete_options.add_namespace_lock_guard(&publication_guard);
+        let (mut deleted_objs, errors) = api.delete_objects(bucket, to_del.to_vec(), delete_options).await;
        failed += errors.iter().filter(|err| err.is_some()).count();

        for (i, deleted_obj) in deleted_objs.iter_mut().enumerate() {
@@ -20,7 +20,7 @@ use std::time::Duration;
 use tokio::sync::mpsc;
 use tokio_util::sync::CancellationToken;

-use crate::bucket::lifecycle::bucket_lifecycle_ops::enqueue_recovered_free_version;
+use crate::bucket::lifecycle::bucket_lifecycle_ops::enqueue_recovered_free_version_with_cancel;
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::Result;
 use crate::object_api::ObjectInfo;
@@ -201,7 +201,11 @@ pub(super) async fn recover_tier_free_versions_with_cancel(
            return Err(tier_free_version_recovery_cancelled());
        }
        retry_cursor.visit(&oi);
-        if !record_recovered_free_version_enqueue(&mut stats, enqueue_recovered_free_version(&api, oi).await) {
+        let queued = enqueue_recovered_free_version_with_cancel(&api, oi, &cancel_token).await;
+        if cancel_token.is_cancelled() {
+            return Err(tier_free_version_recovery_cancelled());
+        }
+        if !record_recovered_free_version_enqueue(&mut stats, queued) {
            let (bucket_marker, object_marker) = retry_cursor.retry_markers();
            stats.truncated = true;
            stats.next_bucket_marker = bucket_marker;
@@ -1210,6 +1210,7 @@ impl BucketMetadata {
                "Failed to parse bucket metadata config"
            );
        }
+        self.notification_config = None;
        if !self.notification_config_xml.is_empty()
            && let Err(e) = deserialize::<NotificationConfiguration>(&self.notification_config_xml)
                .map(|c| self.notification_config = Some(c))
@@ -1224,6 +1225,7 @@ impl BucketMetadata {
                "Failed to parse bucket metadata config"
            );
        }
+        self.lifecycle_config = None;
        if !self.lifecycle_config_xml.is_empty()
            && let Err(e) =
                deserialize::<BucketLifecycleConfiguration>(&self.lifecycle_config_xml).map(|c| self.lifecycle_config = Some(c))
@@ -1238,6 +1240,7 @@ impl BucketMetadata {
                "Failed to parse bucket metadata config"
            );
        }
+        self.object_lock_config = None;
        if !self.object_lock_config_xml.is_empty()
            && let Err(e) =
                deserialize::<ObjectLockConfiguration>(&self.object_lock_config_xml).map(|c| self.object_lock_config = Some(c))
@@ -1267,6 +1270,7 @@ impl BucketMetadata {
                "Failed to parse bucket metadata config"
            );
        }
+        self.sse_config = None;
        if !self.encryption_config_xml.is_empty()
            && let Err(e) =
                deserialize::<ServerSideEncryptionConfiguration>(&self.encryption_config_xml).map(|c| self.sse_config = Some(c))
@@ -1281,6 +1285,7 @@ impl BucketMetadata {
                "Failed to parse bucket metadata config"
            );
        }
+        self.tagging_config = None;
        if !self.tagging_config_xml.is_empty()
            && let Err(e) = deserialize::<Tagging>(&self.tagging_config_xml).map(|c| self.tagging_config = Some(c))
        {
@@ -1346,6 +1351,7 @@ impl BucketMetadata {
        } else {
            self.bucket_target_config = Some(BucketTargets::default());
        }
+        self.cors_config = None;
        if !self.cors_config_xml.is_empty()
            && let Err(e) = deserialize::<CORSConfiguration>(&self.cors_config_xml).map(|c| self.cors_config = Some(c))
        {
@@ -1359,6 +1365,7 @@ impl BucketMetadata {
                "Failed to parse bucket metadata config"
            );
        }
+        self.logging_config = None;
        if !self.logging_config_xml.is_empty()
            && let Err(e) = deserialize::<BucketLoggingStatus>(&self.logging_config_xml).map(|c| self.logging_config = Some(c))
        {
@@ -1372,6 +1379,7 @@ impl BucketMetadata {
                "Failed to parse bucket metadata config"
            );
        }
+        self.website_config = None;
        if !self.website_config_xml.is_empty()
            && let Err(e) = deserialize::<WebsiteConfiguration>(&self.website_config_xml).map(|c| self.website_config = Some(c))
        {
@@ -1385,6 +1393,7 @@ impl BucketMetadata {
                "Failed to parse bucket metadata config"
            );
        }
+        self.accelerate_config = None;
        if !self.accelerate_config_xml.is_empty()
            && let Err(e) =
                deserialize::<AccelerateConfiguration>(&self.accelerate_config_xml).map(|c| self.accelerate_config = Some(c))
@@ -1399,6 +1408,7 @@ impl BucketMetadata {
                "Failed to parse bucket metadata config"
            );
        }
+        self.request_payment_config = None;
        if !self.request_payment_config_xml.is_empty()
            && let Err(e) = deserialize::<RequestPaymentConfiguration>(&self.request_payment_config_xml)
                .map(|c| self.request_payment_config = Some(c))
@@ -1413,6 +1423,7 @@ impl BucketMetadata {
                "Failed to parse bucket metadata config"
            );
        }
+        self.public_access_block_config = None;
        if !self.public_access_block_config_xml.is_empty()
            && let Err(e) = deserialize::<PublicAccessBlockConfiguration>(&self.public_access_block_config_xml)
                .map(|c| self.public_access_block_config = Some(c))
@@ -1973,6 +1984,156 @@ mod test {
        assert_eq!(bm.lifecycle_config_xml, malformed_xml, "raw bytes are retained");
    }

+    const REPARSE_XML_FIXTURES: [(&str, &[u8]); 13] = [
+        (BUCKET_NOTIFICATION_CONFIG, br#"<NotificationConfiguration/>"#),
+        (
+            BUCKET_LIFECYCLE_CONFIG,
+            br#"<LifecycleConfiguration><Rule><ID>keep</ID><Status>Enabled</Status><Expiration><Days>30</Days></Expiration></Rule></LifecycleConfiguration>"#,
+        ),
+        (
+            OBJECT_LOCK_CONFIG,
+            br#"<ObjectLockConfiguration><ObjectLockEnabled>Enabled</ObjectLockEnabled><Rule><DefaultRetention><Mode>GOVERNANCE</Mode><Days>7</Days></DefaultRetention></Rule></ObjectLockConfiguration>"#,
+        ),
+        (BUCKET_VERSIONING_CONFIG, br#"<VersioningConfiguration><Status>Enabled</Status></VersioningConfiguration>"#),
+        (
+            BUCKET_SSECONFIG,
+            br#"<ServerSideEncryptionConfiguration><Rule><ApplyServerSideEncryptionByDefault><SSEAlgorithm>AES256</SSEAlgorithm></ApplyServerSideEncryptionByDefault></Rule></ServerSideEncryptionConfiguration>"#,
+        ),
+        (BUCKET_TAGGING_CONFIG, br#"<Tagging><TagSet><Tag><Key>team</Key><Value>blue</Value></Tag></TagSet></Tagging>"#),
+        (
+            BUCKET_REPLICATION_CONFIG,
+            br#"<ReplicationConfiguration><Role>arn:aws:iam::123456789012:role/replication-role</Role><Rule><ID>copy</ID><Status>Enabled</Status><Prefix>documents/</Prefix><Destination><Bucket>arn:aws:s3:::destination-bucket</Bucket></Destination></Rule></ReplicationConfiguration>"#,
+        ),
+        (
+            BUCKET_CORS_CONFIG,
+            br#"<CORSConfiguration><CORSRule><AllowedOrigin>https://example.test</AllowedOrigin><AllowedMethod>GET</AllowedMethod></CORSRule></CORSConfiguration>"#,
+        ),
+        (BUCKET_LOGGING_CONFIG, br#"<BucketLoggingStatus/>"#),
+        (BUCKET_WEBSITE_CONFIG, br#"<WebsiteConfiguration><IndexDocument><Suffix>index.html</Suffix></IndexDocument></WebsiteConfiguration>"#),
+        (BUCKET_ACCELERATE_CONFIG, br#"<AccelerateConfiguration><Status>Enabled</Status></AccelerateConfiguration>"#),
+        (BUCKET_REQUEST_PAYMENT_CONFIG, br#"<RequestPaymentConfiguration><Payer>BucketOwner</Payer></RequestPaymentConfiguration>"#),
+        (
+            BUCKET_PUBLIC_ACCESS_BLOCK_CONFIG,
+            br#"<PublicAccessBlockConfiguration><BlockPublicAcls>true</BlockPublicAcls><IgnorePublicAcls>true</IgnorePublicAcls><BlockPublicPolicy>true</BlockPublicPolicy><RestrictPublicBuckets>true</RestrictPublicBuckets></PublicAccessBlockConfiguration>"#,
+        ),
+    ];
+
+    fn reparse_xml_state(bm: &BucketMetadata, config: &str) -> &'static str {
+        fn state<T>(raw: &[u8], parsed: &Option<T>) -> &'static str {
+            match ConfigState::of(raw, parsed) {
+                ConfigState::Absent => "absent",
+                ConfigState::Valid(_) => "valid",
+                ConfigState::Unreadable { .. } => "unreadable",
+            }
+        }
+        match config {
+            BUCKET_NOTIFICATION_CONFIG => state(&bm.notification_config_xml, &bm.notification_config),
+            BUCKET_LIFECYCLE_CONFIG => state(&bm.lifecycle_config_xml, &bm.lifecycle_config),
+            OBJECT_LOCK_CONFIG => state(&bm.object_lock_config_xml, &bm.object_lock_config),
+            BUCKET_VERSIONING_CONFIG => state(&bm.versioning_config_xml, &bm.versioning_config),
+            BUCKET_SSECONFIG => state(&bm.encryption_config_xml, &bm.sse_config),
+            BUCKET_TAGGING_CONFIG => state(&bm.tagging_config_xml, &bm.tagging_config),
+            BUCKET_REPLICATION_CONFIG => state(&bm.replication_config_xml, &bm.replication_config),
+            BUCKET_CORS_CONFIG => state(&bm.cors_config_xml, &bm.cors_config),
+            BUCKET_LOGGING_CONFIG => state(&bm.logging_config_xml, &bm.logging_config),
+            BUCKET_WEBSITE_CONFIG => state(&bm.website_config_xml, &bm.website_config),
+            BUCKET_ACCELERATE_CONFIG => state(&bm.accelerate_config_xml, &bm.accelerate_config),
+            BUCKET_REQUEST_PAYMENT_CONFIG => state(&bm.request_payment_config_xml, &bm.request_payment_config),
+            BUCKET_PUBLIC_ACCESS_BLOCK_CONFIG => state(&bm.public_access_block_config_xml, &bm.public_access_block_config),
+            _ => panic!("unknown XML config: {config}"),
+        }
+    }
+
+    // A persisted payload can change without going through the request decoder.
+    fn replace_xml_raw(bm: &mut BucketMetadata, config: &str, raw: Vec<u8>) {
+        match config {
+            BUCKET_NOTIFICATION_CONFIG => bm.notification_config_xml = raw,
+            BUCKET_LIFECYCLE_CONFIG => bm.lifecycle_config_xml = raw,
+            OBJECT_LOCK_CONFIG => bm.object_lock_config_xml = raw,
+            BUCKET_VERSIONING_CONFIG => bm.versioning_config_xml = raw,
+            BUCKET_SSECONFIG => bm.encryption_config_xml = raw,
+            BUCKET_TAGGING_CONFIG => bm.tagging_config_xml = raw,
+            BUCKET_REPLICATION_CONFIG => bm.replication_config_xml = raw,
+            BUCKET_CORS_CONFIG => bm.cors_config_xml = raw,
+            BUCKET_LOGGING_CONFIG => bm.logging_config_xml = raw,
+            BUCKET_WEBSITE_CONFIG => bm.website_config_xml = raw,
+            BUCKET_ACCELERATE_CONFIG => bm.accelerate_config_xml = raw,
+            BUCKET_REQUEST_PAYMENT_CONFIG => bm.request_payment_config_xml = raw,
+            BUCKET_PUBLIC_ACCESS_BLOCK_CONFIG => bm.public_access_block_config_xml = raw,
+            _ => panic!("unknown XML config: {config}"),
+        }
+    }
+
+    #[test]
+    fn reparse_xml_rejects_stale_views_after_corruption() {
+        let corrupt = b"<not-a-valid-document";
+        for (config, valid) in REPARSE_XML_FIXTURES {
+            let mut bm = BucketMetadata::new("reparse-corrupt");
+            bm.update_config(config, valid.to_vec()).expect("store valid XML");
+            bm.parse_all_configs().expect("parse valid XML");
+            assert_eq!(reparse_xml_state(&bm, config), "valid", "{config}: valid setup");
+
+            replace_xml_raw(&mut bm, config, corrupt.to_vec());
+            let before = bm.marshal_msg().expect("marshal corrupt raw metadata");
+            bm.parse_all_configs()
+                .expect("one corrupt config must not fail the metadata load");
+
+            assert_eq!(reparse_xml_state(&bm, config), "unreadable", "{config}: stale view accepted");
+            assert_eq!(bm.xml_config_unreadable_len(config), Some(corrupt.len()), "{config}: missing refusal");
+            assert_eq!(
+                bm.marshal_msg().expect("marshal reparsed metadata"),
+                before,
+                "{config}: persisted bytes changed"
+            );
+        }
+    }
+
+    #[test]
+    fn reparse_xml_clears_stale_views_after_raw_is_removed() {
+        for (config, valid) in REPARSE_XML_FIXTURES {
+            let mut bm = BucketMetadata::new("reparse-removed");
+            bm.update_config(config, valid.to_vec()).expect("store valid XML");
+            bm.parse_all_configs().expect("parse valid XML");
+            assert_eq!(reparse_xml_state(&bm, config), "valid", "{config}: valid setup");
+
+            replace_xml_raw(&mut bm, config, Vec::new());
+            let before = bm.marshal_msg().expect("marshal absent raw metadata");
+            bm.parse_all_configs().expect("parse removed XML");
+
+            assert_eq!(reparse_xml_state(&bm, config), "absent", "{config}: stale view survived removal");
+            assert!(!bm.xml_config_unreadable(config), "{config}: absence is not corruption");
+            assert_eq!(
+                bm.marshal_msg().expect("marshal reparsed metadata"),
+                before,
+                "{config}: persisted bytes changed"
+            );
+        }
+    }
+
+    #[test]
+    fn reparse_xml_recovers_after_corrupt_raw_is_repaired() {
+        for (config, valid) in REPARSE_XML_FIXTURES {
+            let mut bm = BucketMetadata::new("reparse-repaired");
+            bm.update_config(config, valid.to_vec()).expect("store valid XML");
+            bm.parse_all_configs().expect("parse valid XML");
+            replace_xml_raw(&mut bm, config, b"<not-a-valid-document".to_vec());
+            bm.parse_all_configs().expect("parse corrupt XML");
+            assert_eq!(reparse_xml_state(&bm, config), "unreadable", "{config}: corrupt setup");
+
+            replace_xml_raw(&mut bm, config, valid.to_vec());
+            let before = bm.marshal_msg().expect("marshal repaired raw metadata");
+            bm.parse_all_configs().expect("parse repaired XML");
+
+            assert_eq!(reparse_xml_state(&bm, config), "valid", "{config}: repaired config must become readable");
+            assert!(!bm.xml_config_unreadable(config), "{config}: repaired config is not corrupt");
+            assert_eq!(
+                bm.marshal_msg().expect("marshal reparsed metadata"),
+                before,
+                "{config}: persisted bytes changed"
+            );
+        }
+    }
+
    #[test]
    fn lifecycle_update_config_clears_parsed_config_on_delete() {
        let mut bm = BucketMetadata::new("test-bucket");
@@ -576,6 +576,27 @@ pub(crate) async fn update_in(
    Box::pin(update_with_sys(require_bucket_metadata_sys_in(ctx)?, bucket, config_file, data)).await
 }

+/// Validate a caller's fence after acquiring the lifecycle and metadata
+/// transaction locks, before changing the configuration.
+pub(crate) async fn update_validated_in<F>(
+    ctx: &crate::runtime::instance::InstanceContext,
+    bucket: &str,
+    config_file: &str,
+    data: Vec<u8>,
+    validate: F,
+) -> Result<OffsetDateTime>
+where
+    F: FnOnce() -> Result<()> + Send,
+{
+    Box::pin(async {
+        let sys = require_bucket_metadata_sys_in(ctx)?;
+        let guard = acquire_config_write_guard(sys.clone(), bucket).await?;
+        validate()?;
+        update_under_config_write_guard(sys, &guard, config_file, data, None).await
+    })
+    .await
+}
+
 pub async fn delete(bucket: &str, config_file: &str) -> Result<OffsetDateTime> {
    delete_with_sys(get_bucket_metadata_sys()?, bucket, config_file).await
 }
@@ -1132,7 +1153,7 @@ async fn acquire_transaction_lock_with_sys(
 /// agree on, so renaming it would leave a mixed-version cluster with two
 /// disjoint keys — and old and new nodes would stop excluding each other on
 /// the very writes that are serialized today.
-fn bucket_metadata_transaction_lock_key(bucket: &str) -> String {
+pub(crate) fn bucket_metadata_transaction_lock_key(bucket: &str) -> String {
    format!("bucket-targets/{bucket}/transaction.lock")
 }

@@ -1271,14 +1292,34 @@ pub(crate) async fn get_object_lock_config_state_in(
 pub(crate) async fn get_object_lock_config_and_incarnation_from_disk_in(
    ctx: &crate::runtime::instance::InstanceContext,
    bucket: &str,
+) -> Result<(ObjectLockConfigState, Uuid, OffsetDateTime)> {
+    get_object_lock_config_and_incarnation_from_disk_with_guard_in(ctx, bucket, None).await
+}
+
+/// Bucket deletion already owns the namespace/lifecycle fences. Reuse its
+/// live guard rather than recursively acquiring the bucket namespace lock.
+/// The caller must also hold the metadata transaction read lock.
+pub(crate) async fn get_object_lock_config_and_incarnation_from_disk_with_guard_in(
+    ctx: &crate::runtime::instance::InstanceContext,
+    bucket: &str,
+    guard: Option<&rustfs_lock::NamespaceLockGuard>,
 ) -> Result<(ObjectLockConfigState, Uuid, OffsetDateTime)> {
    let bucket_meta_sys_lock = bucket_metadata_sys_of(ctx)?;
    let bucket_meta_sys = bucket_meta_sys_lock.read().await.clone();

-    match bucket_meta_sys
-        .read_authoritative_metadata_from_disk_under_transaction_lock(bucket)
-        .await?
-    {
+    let authority = match guard {
+        Some(guard) => {
+            bucket_meta_sys
+                .read_authoritative_metadata_from_disk_under_guard(bucket, guard)
+                .await?
+        }
+        None => {
+            bucket_meta_sys
+                .read_authoritative_metadata_from_disk_under_transaction_lock(bucket)
+                .await?
+        }
+    };
+    match authority {
        BucketMetadataAuthority::Authoritative(metadata)
            if metadata.bucket_incarnation_sidecar && !metadata.bucket_incarnation_id.is_nil() =>
        {
@@ -1331,10 +1372,54 @@ pub(crate) async fn get_quota_config_and_incarnation_from_disk_in(
    let bucket_meta_sys_lock = bucket_metadata_sys_of(ctx)?;
    let bucket_meta_sys = bucket_meta_sys_lock.read().await.clone();

-    match bucket_meta_sys
-        .read_authoritative_metadata_from_disk_under_transaction_lock(bucket)
-        .await?
+    quota_config_and_incarnation_from_authority(
+        bucket,
+        bucket_meta_sys
+            .read_authoritative_metadata_from_disk_under_transaction_lock(bucket)
+            .await?,
+    )
+}
+
+pub(crate) async fn get_quota_config_and_incarnation_from_disk_for_options_in(
+    ctx: &crate::runtime::instance::InstanceContext,
+    bucket: &str,
+    opts: &ObjectOptions,
+    transaction_guard: &Arc<rustfs_lock::NamespaceLockGuard>,
+) -> Result<(Option<BucketQuota>, Uuid, OffsetDateTime)> {
+    let Some(snapshot) = opts.object_lock_config_snapshot.as_ref() else {
+        return get_quota_config_and_incarnation_from_disk_in(ctx, bucket).await;
+    };
+    let sys = bucket_metadata_sys_of(ctx)?.read().await.clone();
+    let store = sys.object_store();
+    let snapshot_guard = snapshot
+        .metadata_transaction_guard_for(store.id, bucket, opts.expected_bucket_incarnation_id)
+        .filter(|guard| Arc::ptr_eq(guard, transaction_guard))
+        .ok_or_else(|| Error::other("quota snapshot does not hold this bucket's metadata transaction fence"))?;
+    if snapshot_guard.is_released() {
+        return Err(Error::other("quota snapshot metadata transaction fence was released"));
+    }
+
+    // The snapshot carries the caller's lifecycle fence, excluding deletion.
+    // Its transaction guard excludes config writes, so a same-name object
+    // owner need not recursively acquire the bucket namespace here.
+    let authority = sys
+        .read_authoritative_metadata_from_disk_under_guard(bucket, transaction_guard)
+        .await?;
+    if snapshot_guard.is_lock_lost()
+        || snapshot_guard.is_released()
+        || matches!(&authority, BucketMetadataAuthority::Authoritative(metadata)
+            if !snapshot.is_valid_for_destructive_put(store.id, bucket, metadata.bucket_incarnation_id))
    {
+        return Err(Error::other("quota snapshot fence or bucket incarnation changed during the disk read"));
+    }
+    quota_config_and_incarnation_from_authority(bucket, authority)
+}
+
+fn quota_config_and_incarnation_from_authority(
+    bucket: &str,
+    authority: BucketMetadataAuthority,
+) -> Result<(Option<BucketQuota>, Uuid, OffsetDateTime)> {
+    match authority {
        BucketMetadataAuthority::Authoritative(metadata)
            if metadata.bucket_incarnation_sidecar && !metadata.bucket_incarnation_id.is_nil() =>
        {
@@ -2703,29 +2788,30 @@ impl BucketMetadataSys {
    async fn read_authoritative_metadata_from_disk_under_transaction_lock(
        &self,
        bucket: &str,
+    ) -> Result<BucketMetadataAuthority> {
+        let namespace_lock = self.object_store().new_ns_lock(bucket, bucket).await?;
+        let namespace_guard = namespace_lock
+            .get_read_lock(crate::set_disk::get_lock_acquire_timeout())
+            .await?;
+        self.read_authoritative_metadata_from_disk_under_guard(bucket, &namespace_guard)
+            .await
+    }
+
+    async fn read_authoritative_metadata_from_disk_under_guard(
+        &self,
+        bucket: &str,
+        guard: &rustfs_lock::NamespaceLockGuard,
    ) -> Result<BucketMetadataAuthority> {
        #[cfg(test)]
        if self.object_lock_disk_read_errors.write().await.remove(bucket) {
            return Err(Error::other(format!("injected Object Lock metadata disk read failure: {bucket}")));
        }

-        let namespace_lock = self.object_store().new_ns_lock(bucket, bucket).await?;
-        let namespace_guard = namespace_lock
-            .get_read_lock(crate::set_disk::get_lock_acquire_timeout())
-            .await?;
-        match await_bucket_namespace_operation(
-            Some(&namespace_guard),
-            bucket,
-            "bucket metadata snapshot existence check",
-            async {
-                self.object_store()
-                    .get_bucket_info_from_sets_at_read_quorum(
-                        bucket,
-                        &crate::storage_api_contracts::bucket::BucketOptions::default(),
-                    )
-                    .await
-            },
-        )
+        match await_bucket_namespace_operation(Some(guard), bucket, "bucket metadata snapshot existence check", async {
+            self.object_store()
+                .get_bucket_info_from_sets_at_read_quorum(bucket, &crate::storage_api_contracts::bucket::BucketOptions::default())
+                .await
+        })
        .await
        {
            Ok(_) => {}
@@ -2734,7 +2820,7 @@ impl BucketMetadataSys {
        }

        let (metadata, persisted) = await_bucket_namespace_operation(
-            Some(&namespace_guard),
+            Some(guard),
            bucket,
            "bucket metadata authoritative snapshot",
            load_bucket_metadata_parse_with_presence(self.object_store(), bucket, true),
@@ -4121,6 +4207,114 @@ mod tests {
        sys.persist_new_and_set(bm).await.expect("initial metadata should persist");
    }

+    #[tokio::test]
+    async fn reparse_xml_rmw_corruption_refuses_cached_and_reloaded_encryption() {
+        use crate::bucket::metadata::{BUCKET_SSECONFIG, is_unreadable_config_error};
+        use std::sync::atomic::{AtomicUsize, Ordering};
+
+        let (dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let sys = BucketMetadataSys::new(ecstore);
+        let bucket = "reparse-encryption-corrupt";
+        let valid = br#"<ServerSideEncryptionConfiguration><Rule><ApplyServerSideEncryptionByDefault><SSEAlgorithm>AES256</SSEAlgorithm></ApplyServerSideEncryptionByDefault></Rule></ServerSideEncryptionConfiguration>"#;
+        let corrupt = b"<ServerSideEncryptionConfiguration";
+        persist_bucket_with_raw_config(&sys, &dirs, bucket, BUCKET_SSECONFIG, valid).await;
+        sys.get_sse_config(bucket).await.expect("valid encryption must be readable");
+
+        let updated = sys
+            .update_config_with(bucket, BUCKET_SSECONFIG, |bm| {
+                assert!(bm.sse_config.is_some(), "rewrite must start with parsed encryption");
+                Ok(corrupt.to_vec())
+            })
+            .await
+            .expect("persist future or corrupt raw encryption without failing the whole bucket");
+
+        let err = sys
+            .get_sse_config(bucket)
+            .await
+            .expect_err("cached old encryption must not hide corrupt raw");
+        assert!(is_unreadable_config_error(&err), "cached refusal must retain its type: {err}");
+        let (cached, _) = sys.get_config(bucket).await.expect("bucket remains readable");
+        assert_eq!(cached.encryption_config_xml, corrupt);
+        assert_eq!(cached.encryption_config_updated_at, updated);
+
+        sys.metadata_map.write().await.clear();
+        let err = sys
+            .get_sse_config(bucket)
+            .await
+            .expect_err("reloaded encryption must still refuse corrupt raw");
+        assert!(is_unreadable_config_error(&err), "reloaded refusal must retain its type: {err}");
+
+        let calls = AtomicUsize::new(0);
+        let err = sys
+            .update_config_with(bucket, BUCKET_SSECONFIG, |_| {
+                calls.fetch_add(1, Ordering::SeqCst);
+                Ok(Vec::new())
+            })
+            .await
+            .expect_err("another rewrite must refuse before it consumes corrupt encryption");
+        assert!(is_unreadable_config_error(&err));
+        assert_eq!(calls.load(Ordering::SeqCst), 0, "refused rewrite must not erase the original raw");
+        let disk = sys
+            .get_config_from_disk(bucket)
+            .await
+            .expect("read unchanged metadata from disk");
+        assert_eq!(disk.encryption_config_xml, corrupt);
+        assert_eq!(disk.encryption_config_updated_at, updated);
+
+        sys.update(bucket, BUCKET_SSECONFIG, valid.to_vec())
+            .await
+            .expect("explicit repair must remain possible");
+        sys.get_sse_config(bucket)
+            .await
+            .expect("repaired cached encryption must be readable");
+        sys.metadata_map.write().await.clear();
+        sys.get_sse_config(bucket).await.expect("repair must survive a disk reload");
+    }
+
+    #[tokio::test]
+    async fn reparse_xml_rmw_removal_clears_cached_and_reloaded_encryption() {
+        use crate::bucket::metadata::BUCKET_SSECONFIG;
+
+        let (dirs, ecstore) = isolated_store_over_temp_disks().await;
+        let sys = BucketMetadataSys::new(ecstore);
+        let bucket = "reparse-encryption-removed";
+        let valid = br#"<ServerSideEncryptionConfiguration><Rule><ApplyServerSideEncryptionByDefault><SSEAlgorithm>AES256</SSEAlgorithm></ApplyServerSideEncryptionByDefault></Rule></ServerSideEncryptionConfiguration>"#;
+        persist_bucket_with_raw_config(&sys, &dirs, bucket, BUCKET_SSECONFIG, valid).await;
+        sys.get_sse_config(bucket).await.expect("valid encryption must be readable");
+
+        let updated = sys
+            .update_config_with(bucket, BUCKET_SSECONFIG, |bm| {
+                assert!(bm.sse_config.is_some(), "rewrite must start with parsed encryption");
+                Ok(Vec::new())
+            })
+            .await
+            .expect("removing encryption must persist");
+
+        assert_eq!(
+            sys.get_sse_config(bucket)
+                .await
+                .expect_err("old cached encryption must be gone"),
+            Error::ConfigNotFound
+        );
+        let (cached, _) = sys.get_config(bucket).await.expect("bucket remains readable");
+        assert!(cached.encryption_config_xml.is_empty());
+        assert_eq!(cached.encryption_config_updated_at, updated);
+
+        sys.metadata_map.write().await.clear();
+        assert_eq!(
+            sys.get_sse_config(bucket)
+                .await
+                .expect_err("removed encryption must stay absent after reload"),
+            Error::ConfigNotFound
+        );
+        let disk = sys
+            .get_config_from_disk(bucket)
+            .await
+            .expect("read removed encryption from disk");
+        assert!(disk.encryption_config_xml.is_empty());
+        assert_eq!(disk.encryption_config_updated_at, updated);
+    }
+
    /// rustfs/backlog#1734: a read-modify-write of a stored config that cannot
    /// be parsed must be refused before `mutate` runs. Otherwise `mutate` sees
    /// the unreadable config as absent, rebuilds it from nothing, and the
@@ -4736,6 +4930,231 @@ mod tests {
            .expect("DeleteBucket should succeed");
    }

+    #[tokio::test]
+    #[serial]
+    async fn quota_snapshot_reuses_same_name_owner_and_rereads_disk_authority() {
+        let (_dirs, store) = isolated_store_over_temp_disks().await;
+        init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let bucket = "quota-same-name-owner";
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("quota fixture bucket");
+        let snapshot = store
+            .object_lock_config_snapshot(bucket)
+            .await
+            .expect("guarded Object Lock snapshot");
+        let incarnation = store.bucket_incarnation_id(bucket).await.expect("fixture incarnation");
+        let transaction_guard = snapshot
+            .metadata_transaction_guard_for(store.id, bucket, Some(incarnation))
+            .expect("snapshot's actual transaction guard");
+        let expected = get_quota_config_and_incarnation_from_disk_for_options_in(
+            &store.ctx,
+            bucket,
+            &ObjectOptions::default(),
+            &transaction_guard,
+        )
+        .await
+        .expect("a reader without a snapshot must retain the ordinary namespace path");
+
+        let sys = bucket_metadata_sys_of(&store.ctx).expect("fixture metadata system");
+        let sys = sys.read().await.clone();
+        let mut stale = (*sys.get(bucket).await.expect("cached fixture metadata")).clone();
+        let metadata_path = stale.save_file_path();
+        stale.quota_config = Some(BucketQuota {
+            quota: Some(123),
+            ..Default::default()
+        });
+        sys.set(bucket.to_string(), Arc::new(stale)).await;
+        let namespace = store.new_ns_lock(bucket, bucket).await.expect("same-name object namespace");
+        let _owner = namespace
+            .get_write_lock(crate::set_disk::get_lock_acquire_timeout())
+            .await
+            .expect("same-name object write owner");
+        let opts = ObjectOptions {
+            object_lock_config_snapshot: Some(snapshot),
+            expected_bucket_incarnation_id: Some(incarnation),
+            ..Default::default()
+        };
+        assert_eq!(
+            get_quota_config_and_incarnation_from_disk_for_options_in(&store.ctx, bucket, &opts, &transaction_guard)
+                .await
+                .expect("quota disk read must reuse the held lifecycle and transaction fences"),
+            expected,
+            "the cached quota must not replace disk authority"
+        );
+
+        crate::config::com::save_config_with_opts(
+            store.clone(),
+            &metadata_path,
+            b"corrupt bucket metadata".to_vec(),
+            &ObjectOptions {
+                write_completion: crate::object_api::WriteCompletion::TailDrained,
+                ..Default::default()
+            },
+        )
+        .await
+        .expect("inject committed metadata corruption");
+        assert!(
+            get_quota_config_and_incarnation_from_disk_for_options_in(&store.ctx, bucket, &opts, &transaction_guard)
+                .await
+                .is_err(),
+            "a valid snapshot must not hide corrupt disk metadata"
+        );
+        crate::config::com::delete_config(store.clone(), &metadata_path)
+            .await
+            .expect("remove the persisted metadata while retaining its sidecar");
+        assert!(
+            get_quota_config_and_incarnation_from_disk_for_options_in(&store.ctx, bucket, &opts, &transaction_guard)
+                .await
+                .is_err(),
+            "a valid snapshot must not promote fabricated metadata to authority"
+        );
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn quota_snapshot_rejects_wrong_scope_incarnation_and_transaction_guard() {
+        let (_dirs, store) = isolated_store_over_temp_disks().await;
+        init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let bucket = "quota-snapshot-scope";
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("quota fixture bucket");
+        let incarnation = store.bucket_incarnation_id(bucket).await.expect("fixture incarnation");
+        let snapshot = store
+            .object_lock_config_snapshot(bucket)
+            .await
+            .expect("guarded Object Lock snapshot");
+        let transaction_guard = snapshot
+            .metadata_transaction_guard_for(store.id, bucket, Some(incarnation))
+            .expect("snapshot's actual transaction guard");
+        let mut opts = ObjectOptions {
+            object_lock_config_snapshot: Some(snapshot),
+            expected_bucket_incarnation_id: Some(incarnation),
+            ..Default::default()
+        };
+        assert!(
+            get_quota_config_and_incarnation_from_disk_for_options_in(&store.ctx, "another-bucket", &opts, &transaction_guard)
+                .await
+                .is_err()
+        );
+        opts.expected_bucket_incarnation_id = Some(Uuid::new_v4());
+        assert!(
+            get_quota_config_and_incarnation_from_disk_for_options_in(&store.ctx, bucket, &opts, &transaction_guard)
+                .await
+                .is_err()
+        );
+        opts.expected_bucket_incarnation_id = Some(incarnation);
+        let other_guard = Arc::new(
+            acquire_bucket_metadata_transaction_read_lock_in(&store.ctx, bucket)
+                .await
+                .expect("distinct real transaction read guard"),
+        );
+        assert!(
+            get_quota_config_and_incarnation_from_disk_for_options_in(&store.ctx, bucket, &opts, &other_guard)
+                .await
+                .is_err(),
+            "the same namespace key does not prove that this is the snapshot's actual guard"
+        );
+        let (_other_dirs, other_store) = isolated_store_over_temp_disks().await;
+        init_bucket_metadata_sys(other_store.clone(), Vec::new()).await;
+        assert!(
+            get_quota_config_and_incarnation_from_disk_for_options_in(&other_store.ctx, bucket, &opts, &transaction_guard)
+                .await
+                .is_err()
+        );
+        opts.object_lock_config_snapshot = Some(Arc::new(crate::object_api::ObjectLockConfigSnapshot::new(
+            ObjectLockConfigState::ConfirmedAbsent,
+        )));
+        assert!(
+            get_quota_config_and_incarnation_from_disk_for_options_in(&store.ctx, bucket, &opts, &transaction_guard)
+                .await
+                .is_err(),
+            "an unguarded snapshot must not authorize namespace reuse"
+        );
+    }
+
+    #[tokio::test]
+    #[serial]
+    async fn quota_snapshot_rejects_lifecycle_loss_during_read_and_changed_disk_incarnation() {
+        let (_dirs, store) = isolated_store_over_temp_disks().await;
+        init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+        let bucket = "quota-snapshot-loss";
+        store
+            .make_bucket(bucket, &MakeBucketOptions::default())
+            .await
+            .expect("quota fixture bucket");
+        let sys = bucket_metadata_sys_of(&store.ctx).expect("fixture metadata system");
+        let sys = sys.read().await.clone();
+        let mut metadata = (*sys.get(bucket).await.expect("fixture metadata")).clone();
+        let incarnation = metadata.bucket_incarnation_id;
+        let _lifecycle_guard = store
+            .acquire_bucket_lifecycle_read_lock(bucket)
+            .await
+            .expect("real lifecycle owner");
+        let transaction_guard = Arc::new(
+            acquire_bucket_metadata_transaction_read_lock_in(&store.ctx, bucket)
+                .await
+                .expect("real metadata transaction owner"),
+        );
+        let (lifecycle_fence, lost) = crate::object_api::NamespaceLockFence::loss_handle_for_test();
+        let snapshot = crate::object_api::ObjectLockConfigSnapshot::for_store_bucket_under_lifecycle_fence(
+            store.id,
+            bucket,
+            incarnation,
+            metadata.object_lock_config_updated_at,
+            ObjectLockConfigState::ConfirmedAbsent,
+            lifecycle_fence,
+            Arc::clone(&transaction_guard),
+        );
+        let opts = ObjectOptions {
+            object_lock_config_snapshot: Some(Arc::new(snapshot)),
+            expected_bucket_incarnation_id: Some(incarnation),
+            ..Default::default()
+        };
+        let mut read = Box::pin(get_quota_config_and_incarnation_from_disk_for_options_in(
+            &store.ctx,
+            bucket,
+            &opts,
+            &transaction_guard,
+        ));
+        assert!(
+            futures::poll!(read.as_mut()).is_pending(),
+            "the authoritative disk read must be in flight"
+        );
+        lost.store(true, std::sync::atomic::Ordering::Release);
+        assert!(read.await.is_err(), "a lifecycle fence lost during the disk read must reject its result");
+
+        let snapshot = store
+            .object_lock_config_snapshot(bucket)
+            .await
+            .expect("fresh guarded snapshot");
+        let transaction_guard = snapshot
+            .metadata_transaction_guard_for(store.id, bucket, Some(incarnation))
+            .expect("snapshot's actual transaction guard");
+        let opts = ObjectOptions {
+            object_lock_config_snapshot: Some(snapshot),
+            expected_bucket_incarnation_id: Some(incarnation),
+            ..Default::default()
+        };
+        metadata.bucket_incarnation_id = Uuid::new_v4();
+        save_bucket_incarnation(store.clone(), bucket, metadata.bucket_incarnation_id)
+            .await
+            .expect("inject a new disk incarnation");
+        metadata
+            .save_with_store_committed(store.clone())
+            .await
+            .expect("persist its matching metadata blob");
+        assert!(
+            get_quota_config_and_incarnation_from_disk_for_options_in(&store.ctx, bucket, &opts, &transaction_guard)
+                .await
+                .is_err(),
+            "a matching sidecar and blob from another incarnation must not reuse the old snapshot"
+        );
+    }
+
    /// Writers on different nodes updating *different* config files of one
    /// bucket must both survive. Each rewrites the whole metadata blob, so
    /// without a lock spanning the read-modify-write the later save carries
@@ -32,7 +32,10 @@ pub enum QuotaType {
    Hard,
 }

-pub(crate) const QUOTA_RESERVATION_PROTOCOL_V1: u32 = 1;
+pub const QUOTA_RESERVATION_PROTOCOL_V1: u32 = 1;
+/// Sharded durable reservations. Older nodes reject this protocol instead of
+/// silently falling back to the single-ledger protocol.
+pub const QUOTA_RESERVATION_PROTOCOL_V2: u32 = 2;

 /// Bucket quota configuration. quota_type defaults to Hard when omitted.
 #[derive(Debug, Default, Clone, PartialEq)]
@@ -89,7 +92,10 @@ impl<'de> Deserialize<'de> for BucketQuota {
        D: Deserializer<'de>,
    {
        let wire = BucketQuotaWire::deserialize(deserializer)?;
-        let quota = if wire.reservation_protocol == Some(QUOTA_RESERVATION_PROTOCOL_V1) {
+        let quota = if matches!(
+            wire.reservation_protocol,
+            Some(QUOTA_RESERVATION_PROTOCOL_V1) | Some(QUOTA_RESERVATION_PROTOCOL_V2)
+        ) {
            Some(
                wire.reservation_quota
                    .ok_or_else(|| D::Error::custom("reservation_quota is required for reservation protocol v1"))?,
@@ -129,17 +135,33 @@ impl BucketQuota {
        }
    }

+    /// Creates a quota using the sharded durable reservation protocol.
+    /// Deployment must complete the mixed-version rollout before enabling it;
+    /// older nodes fail closed when they see protocol v2.
+    pub fn new_sharded(quota: u64) -> Self {
+        let mut config = Self::new(Some(quota));
+        config.reservation_protocol = Some(QUOTA_RESERVATION_PROTOCOL_V2);
+        config
+    }
+
    pub fn get_quota_limit(&self) -> Option<u64> {
        self.quota
    }

    pub fn uses_durable_reservations(&self) -> bool {
-        self.reservation_protocol == Some(QUOTA_RESERVATION_PROTOCOL_V1)
+        matches!(
+            self.reservation_protocol,
+            Some(QUOTA_RESERVATION_PROTOCOL_V1) | Some(QUOTA_RESERVATION_PROTOCOL_V2)
+        )
+    }
+
+    pub fn uses_sharded_reservations(&self) -> bool {
+        self.reservation_protocol == Some(QUOTA_RESERVATION_PROTOCOL_V2)
    }

    pub fn has_unsupported_reservation_protocol(&self) -> bool {
        self.reservation_protocol
-            .is_some_and(|version| version != QUOTA_RESERVATION_PROTOCOL_V1)
+            .is_some_and(|version| !matches!(version, QUOTA_RESERVATION_PROTOCOL_V1 | QUOTA_RESERVATION_PROTOCOL_V2))
    }

    pub fn check_operation_allowed(&self, current_usage: u64, operation_size: u64) -> bool {
@@ -322,9 +344,21 @@ mod tests {
    }

    #[test]
-    fn unknown_reservation_protocol_does_not_activate_v1() {
+    fn sharded_quota_roundtrips_with_authoritative_reservation_quota() {
+        let quota = BucketQuota::new_sharded(4096);
+        let json = serde_json::to_value(&quota).expect("sharded quota should serialize");
+        assert_eq!(json["quota"], 0);
+        assert_eq!(json["reservation_protocol"], QUOTA_RESERVATION_PROTOCOL_V2);
+        assert_eq!(json["reservation_quota"], 4096);
+        let restored: BucketQuota = serde_json::from_value(json).expect("sharded quota should parse");
+        assert_eq!(restored.quota, Some(4096));
+        assert!(restored.uses_sharded_reservations());
+    }
+
+    #[test]
+    fn unknown_reservation_protocol_does_not_activate_any_known_protocol() {
        let quota: BucketQuota =
-            serde_json::from_str(r#"{"quota":0,"quota_type":"Hard","reservation_protocol":2,"reservation_quota":2048}"#)
+            serde_json::from_str(r#"{"quota":0,"quota_type":"Hard","reservation_protocol":3,"reservation_quota":2048}"#)
                .expect("future protocol should remain parseable");

        assert!(!quota.uses_durable_reservations());
@@ -12,9 +12,11 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

+#[cfg(test)]
+use crate::disk::DiskIoClass;
 use crate::disk::disk_store::{get_drive_walkdir_peek_timeout, get_drive_walkdir_stall_timeout};
 use crate::disk::error::DiskError;
-use crate::disk::{self, DiskAPI, DiskStore, WalkDirOptions};
+use crate::disk::{self, DiskAPI, DiskStore, WalkDirOptions, current_disk_io_context, with_optional_disk_io_context};
 use futures::future::join_all;
 use metrics::counter;
 use rustfs_filemeta::{MetaCacheEntries, MetaCacheEntry, MetacacheReader, is_io_eof};
@@ -245,6 +247,8 @@ pub struct ListPathRawOptions {
    pub(crate) test_fallback_reader_behaviors: Vec<TestReaderBehavior>,
    #[cfg(test)]
    pub(crate) peek_timeout: Option<Duration>,
+    #[cfg(test)]
+    pub(crate) test_disk_io_context_observed: Option<Arc<AtomicBool>>,
    // pub agreed: Option<Arc<dyn Fn(MetaCacheEntry) + Send + Sync>>,
    // pub partial: Option<Arc<dyn Fn(MetaCacheEntries, &[Option<Error>]) + Send + Sync>>,
    // pub finished: Option<Arc<dyn Fn(&[Option<Error>]) + Send + Sync>>,
@@ -276,6 +280,8 @@ impl Clone for ListPathRawOptions {
            test_fallback_reader_behaviors: self.test_fallback_reader_behaviors.clone(),
            #[cfg(test)]
            peek_timeout: self.peek_timeout,
+            #[cfg(test)]
+            test_disk_io_context_observed: self.test_disk_io_context_observed.clone(),
            ..Default::default()
        }
    }
@@ -344,6 +350,7 @@ async fn list_path_raw_inner(
    let log_path = opts.path.clone();

    let mut jobs = JoinSet::new();
+    let disk_io_context = current_disk_io_context();
    let mut readers = Vec::with_capacity(opts.disks.len());
    let fds = Arc::new(TokioMutex::new(opts.fallback_disks.iter().flatten().cloned().collect::<VecDeque<_>>()));
    #[cfg(test)]
@@ -364,11 +371,19 @@ async fn list_path_raw_inner(
        let test_fallbacks_clone = test_fallbacks.clone();
        let cancel_rx_clone = cancel_rx.clone();
        let producer_errs_clone = producer_errs.clone();
+        let disk_io_context_clone = disk_io_context.clone();
        let (rd, wr) = tokio::io::duplex(64);
        readers.push(MetacacheReader::new(rd));
-        jobs.spawn(async move {
+        jobs.spawn(with_optional_disk_io_context(disk_io_context_clone, async move {
            let mut wr = PublishedBytesWriter::new(wr);
            #[cfg(test)]
+            if let Some(observer) = &opts_clone.test_disk_io_context_observed {
+                observer.store(
+                    crate::disk::current_disk_io_context().is_some_and(|context| context.class == DiskIoClass::Background),
+                    std::sync::atomic::Ordering::Release,
+                );
+            }
+            #[cfg(test)]
            let test_primary_error = if let Some(behavior) = opts_clone.test_reader_behaviors.get(disk_idx).cloned() {
                match behavior {
                    TestReaderBehavior::Eof => return Ok(()),
@@ -652,7 +667,7 @@ async fn list_path_raw_inner(

            // warn!("list_path_raw: while need_fallback done");
            Ok(())
-        });
+        }));
    }

    let revjob_rx = rx.clone();
@@ -1168,6 +1183,68 @@ mod tests {
        assert_eq!(err, DiskError::ErasureReadQuorum);
    }

+    #[tokio::test]
+    async fn list_path_raw_joinset_producers_inherit_scanner_disk_io_class() {
+        let observed = Arc::new(AtomicBool::new(false));
+        crate::disk::with_background_disk_io(
+            CancellationToken::new(),
+            list_path_raw(
+                CancellationToken::new(),
+                ListPathRawOptions {
+                    disks: vec![None],
+                    min_disks: 1,
+                    test_reader_behaviors: vec![TestReaderBehavior::Eof],
+                    test_disk_io_context_observed: Some(observed.clone()),
+                    ..Default::default()
+                },
+            ),
+        )
+        .await
+        .expect("empty test producer should complete the listing");
+
+        assert!(
+            observed.load(std::sync::atomic::Ordering::Acquire),
+            "the per-disk JoinSet producer must retain the scanner background class"
+        );
+    }
+
+    #[tokio::test]
+    async fn scanner_joinset_walk_uses_the_local_disk_background_gate() {
+        let root = tempfile::tempdir().expect("local disk root should be created");
+        let endpoint = crate::disk::endpoint::Endpoint::try_from(root.path().to_str().expect("disk path should be UTF-8"))
+            .expect("endpoint should parse");
+        let local = Arc::new(
+            crate::disk::local::LocalDisk::new(&endpoint, false)
+                .await
+                .expect("local disk should open"),
+        );
+        let wrapper = crate::disk::disk_store::LocalDiskWrapper::new(local, false);
+        wrapper.make_volume("bucket").await.expect("bucket volume should be created");
+        let disk = Arc::new(crate::disk::Disk::Local(Box::new(wrapper.clone())));
+        let cancellation = CancellationToken::new();
+
+        crate::disk::with_background_disk_io(
+            cancellation.clone(),
+            list_path_raw(
+                cancellation,
+                ListPathRawOptions {
+                    disks: vec![Some(disk)],
+                    bucket: "bucket".to_string(),
+                    min_disks: 1,
+                    recursive: true,
+                    ..Default::default()
+                },
+            ),
+        )
+        .await
+        .expect("scanner metacache walk should complete");
+
+        assert!(
+            wrapper.background_io_admissions_for_tests() > 0,
+            "the LocalDiskWrapper walk should acquire background admission inside its JoinSet producer"
+        );
+    }
+
    #[tokio::test]
    async fn list_path_raw_returns_typed_quorum_error_for_multiple_drive_failures() {
        let err = list_path_raw(
@@ -0,0 +1,806 @@
+//! Bounded authenticated Raft RPC, mountable without a second storage owner.
+use super::{client, http_auth};
+use openraft::{
+    DefensiveError, ErrorSubject, ErrorVerb, RaftNetwork, RaftNetworkFactory, StorageError, StorageIOError, Violation,
+    error::{Fatal, InstallSnapshotError, RPCError, RaftError, RemoteError, Unreachable},
+    network::RPCOption,
+    raft::{
+        AppendEntriesRequest, AppendEntriesResponse, InstallSnapshotRequest, InstallSnapshotResponse, VoteRequest, VoteResponse,
+    },
+};
+use rustfs_capture_authority::{CaptureMembership, CaptureNode, CaptureTypeConfig, NativeRaftOwner, NodeIdentity};
+use rustfs_protos::proto_gen::node_service::{
+    CaptureRaftRequest, CaptureRaftResponse,
+    capture_raft_service_server::{CaptureRaftService, CaptureRaftServiceServer},
+};
+use serde::{Deserialize, Serialize, de::DeserializeOwned};
+use serde_json::value::RawValue;
+use std::{collections::BTreeMap, io::Write, sync::Arc, time::Duration};
+use tonic::{Request, Response, Status};
+
+#[derive(Clone, Copy, Debug)]
+pub struct CaptureRaftLimits {
+    pub max_request_bytes: usize,
+    pub max_response_bytes: usize,
+    pub max_entries: usize,
+    pub max_snapshot_chunk_bytes: usize,
+    pub max_inflight: usize,
+    pub connect_timeout: Duration,
+    pub request_timeout: Duration,
+}
+impl Default for CaptureRaftLimits {
+    fn default() -> Self {
+        Self {
+            max_request_bytes: 16 * 1024 * 1024,
+            max_response_bytes: 1024 * 1024,
+            max_entries: 300,
+            max_snapshot_chunk_bytes: 3 * 1024 * 1024,
+            max_inflight: 16,
+            connect_timeout: Duration::from_secs(3),
+            request_timeout: Duration::from_secs(5),
+        }
+    }
+}
+impl CaptureRaftLimits {
+    fn validate(&self) -> Result<(), CaptureRaftError> {
+        let chunk = self
+            .max_snapshot_chunk_bytes
+            .checked_mul(4)
+            .and_then(|n| n.checked_add(16384))
+            .ok_or(CaptureRaftError::Limits)?;
+        if self.max_request_bytes < chunk
+            || self.max_request_bytes > 64 * 1024 * 1024
+            || self.max_response_bytes < 4096
+            || self.max_response_bytes > 4 * 1024 * 1024
+            || self.max_entries == 0
+            || self.max_entries > 1024
+            || self.max_snapshot_chunk_bytes == 0
+            || self.max_inflight == 0
+            || self.max_inflight > 128
+            || self.connect_timeout.is_zero()
+            || self.request_timeout.is_zero()
+            || self.request_timeout > Duration::from_secs(60)
+            || self.connect_timeout > Duration::from_secs(60)
+        {
+            return Err(CaptureRaftError::Limits);
+        }
+        Ok(())
+    }
+}
+#[derive(Debug, thiserror::Error)]
+pub enum CaptureRaftError {
+    #[error("capture Raft message limit exceeded")]
+    Limits,
+    #[error("capture Raft request requires a smaller append batch")]
+    RequestTooLarge,
+    #[error("invalid capture Raft message")]
+    Protocol,
+    #[error("capture Raft identity mismatch")]
+    Identity,
+    #[error("capture Raft authentication failed")]
+    Authentication,
+    #[error("capture Raft channel unavailable")]
+    ChannelUnavailable,
+    #[error("capture Raft deadline exceeded")]
+    Deadline,
+    #[error(transparent)]
+    Io(#[from] std::io::Error),
+    #[error(transparent)]
+    Json(#[from] serde_json::Error),
+    #[error(transparent)]
+    Node(#[from] rustfs_capture_authority::CaptureNodeError),
+    #[error(transparent)]
+    Status(Box<Status>),
+}
+#[derive(Clone, Copy, Debug, PartialEq, Eq, Serialize, Deserialize)]
+pub(crate) enum Method {
+    AppendEntries,
+    Vote,
+    InstallSnapshot,
+}
+impl Method {
+    pub(crate) fn path(self) -> &'static str {
+        match self {
+            Self::AppendEntries => "/node_service.CaptureRaftService/AppendEntries",
+            Self::Vote => "/node_service.CaptureRaftService/Vote",
+            Self::InstallSnapshot => "/node_service.CaptureRaftService/InstallSnapshot",
+        }
+    }
+}
+#[derive(Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+pub(crate) struct WireRequest {
+    pub version: u32,
+    pub cluster: [u8; 16],
+    pub source: u64,
+    pub target: u64,
+    pub method: Method,
+    pub nonce: [u8; 16],
+    pub payload: Box<RawValue>,
+}
+struct Bounded {
+    bytes: Vec<u8>,
+    max: usize,
+}
+impl Write for Bounded {
+    fn write(&mut self, bytes: &[u8]) -> std::io::Result<usize> {
+        let end = self
+            .bytes
+            .len()
+            .checked_add(bytes.len())
+            .ok_or_else(|| std::io::Error::other("capture RPC limit"))?;
+        if end > self.max {
+            return Err(std::io::Error::other("capture RPC limit"));
+        }
+        self.bytes.try_reserve(bytes.len()).map_err(std::io::Error::other)?;
+        self.bytes.extend_from_slice(bytes);
+        Ok(bytes.len())
+    }
+    fn flush(&mut self) -> std::io::Result<()> {
+        Ok(())
+    }
+}
+fn encoded<T: Serialize>(value: &T, max: usize) -> Result<Vec<u8>, CaptureRaftError> {
+    let mut out = Bounded { bytes: Vec::new(), max };
+    serde_json::to_writer(&mut out, value).map_err(|error| {
+        if error.is_io() {
+            CaptureRaftError::Limits
+        } else {
+            CaptureRaftError::Json(error)
+        }
+    })?;
+    Ok(out.bytes)
+}
+fn raw<T: Serialize>(value: &T, max: usize) -> Result<Box<RawValue>, CaptureRaftError> {
+    Ok(RawValue::from_string(
+        String::from_utf8(encoded(value, max)?).map_err(|_| CaptureRaftError::Protocol)?,
+    )?)
+}
+pub(crate) fn encode_request<P: Serialize>(
+    local: NodeIdentity,
+    target: u64,
+    method: Method,
+    payload: &P,
+    limits: CaptureRaftLimits,
+) -> Result<CaptureRaftRequest, CaptureRaftError> {
+    let wire = WireRequest {
+        version: 1,
+        cluster: local.cluster_id,
+        source: local.node_id,
+        target,
+        method,
+        nonce: *uuid::Uuid::new_v4().as_bytes(),
+        payload: raw(payload, limits.max_request_bytes)?,
+    };
+    Ok(CaptureRaftRequest {
+        body: encoded(&wire, limits.max_request_bytes)?.into(),
+    })
+}
+pub(crate) fn canonical_request(body: &[u8]) -> Vec<u8> {
+    let mut out = b"rustfs.capture-raft.request.v1\0".to_vec();
+    out.extend_from_slice(body);
+    out
+}
+fn decode_wire(body: &[u8], max: usize) -> Result<WireRequest, CaptureRaftError> {
+    if body.len() > max {
+        return Err(CaptureRaftError::Limits);
+    }
+    let wire: WireRequest = serde_json::from_slice(body)?;
+    if wire.version != 1 || wire.cluster == [0; 16] || wire.nonce == [0; 16] || encoded(&wire, max)? != body {
+        return Err(CaptureRaftError::Protocol);
+    }
+    Ok(wire)
+}
+fn decode_payload<P: Serialize + DeserializeOwned>(wire: &WireRequest, max: usize) -> Result<P, CaptureRaftError> {
+    let value: P = serde_json::from_str(wire.payload.get())?;
+    if encoded(&value, max)?.as_slice() != wire.payload.get().as_bytes() {
+        return Err(CaptureRaftError::Protocol);
+    }
+    Ok(value)
+}
+#[derive(Serialize, Deserialize)]
+#[serde(tag = "kind", deny_unknown_fields)]
+pub(crate) enum RemoteFailure {
+    Stopped,
+    Panicked,
+    StorageIo {
+        operation: Option<IoOperation>,
+    },
+    StorageDefensive {
+        subject: ErrorSubject<u64>,
+        violation: Violation<u64>,
+    },
+    SnapshotMismatch {
+        error: InstallSnapshotError,
+    },
+}
+#[derive(Serialize, Deserialize)]
+pub(crate) struct IoOperation {
+    subject: ErrorSubject<u64>,
+    verb: ErrorVerb,
+}
+fn fatal(error: Fatal<u64>) -> RemoteFailure {
+    match error {
+        Fatal::Stopped => RemoteFailure::Stopped,
+        Fatal::Panicked => RemoteFailure::Panicked,
+        Fatal::StorageError(StorageError::IO { source }) => RemoteFailure::StorageIo {
+            operation: encoded(&source, 16384)
+                .ok()
+                .and_then(|bytes| serde_json::from_slice(&bytes).ok()),
+        },
+        Fatal::StorageError(StorageError::Defensive { source }) => RemoteFailure::StorageDefensive {
+            subject: source.subject,
+            violation: source.violation,
+        },
+    }
+}
+#[derive(Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+enum WireReply {
+    Success { payload: Box<RawValue> },
+    Failure { error: RemoteFailure },
+}
+fn response_context(request: &CaptureRaftRequest, body: &[u8]) -> Result<Vec<u8>, CaptureRaftError> {
+    let mut out = b"rustfs.capture-raft.response.v1\0".to_vec();
+    out.try_reserve(
+        request
+            .body
+            .len()
+            .checked_add(body.len())
+            .and_then(|n| n.checked_add(16))
+            .ok_or(CaptureRaftError::Limits)?,
+    )
+    .map_err(std::io::Error::other)?;
+    out.extend(
+        u64::try_from(request.body.len())
+            .map_err(|_| CaptureRaftError::Limits)?
+            .to_be_bytes(),
+    );
+    out.extend_from_slice(&request.body);
+    out.extend(u64::try_from(body.len()).map_err(|_| CaptureRaftError::Limits)?.to_be_bytes());
+    out.extend_from_slice(body);
+    Ok(out)
+}
+fn reply<P: Serialize>(
+    request: &CaptureRaftRequest,
+    result: Result<P, RemoteFailure>,
+    limits: CaptureRaftLimits,
+) -> Result<CaptureRaftResponse, CaptureRaftError> {
+    let wire = match result {
+        Ok(value) => WireReply::Success {
+            payload: raw(&value, limits.max_response_bytes)?,
+        },
+        Err(error) => WireReply::Failure { error },
+    };
+    let body = encoded(&wire, limits.max_response_bytes)?;
+    let response_proof = http_auth::sign_tonic_rpc_response_proof(&response_context(request, &body)?)?;
+    Ok(CaptureRaftResponse {
+        body: body.into(),
+        response_proof: response_proof.into(),
+    })
+}
+pub(crate) fn verify_response<P: Serialize + DeserializeOwned>(
+    request: &CaptureRaftRequest,
+    response: CaptureRaftResponse,
+    limits: CaptureRaftLimits,
+) -> Result<Result<P, RemoteFailure>, CaptureRaftError> {
+    if request.body.len() > limits.max_request_bytes
+        || response.body.len() > limits.max_response_bytes
+        || response.response_proof.len() != 32
+    {
+        return Err(CaptureRaftError::Limits);
+    }
+    http_auth::verify_tonic_rpc_response_proof(&response_context(request, &response.body)?, &response.response_proof)?;
+    let wire: WireReply = serde_json::from_slice(&response.body)?;
+    if encoded(&wire, limits.max_response_bytes)?.as_slice() != response.body {
+        return Err(CaptureRaftError::Protocol);
+    }
+    match wire {
+        WireReply::Failure { error } => Ok(Err(error)),
+        WireReply::Success { payload } => {
+            let value: P = serde_json::from_str(payload.get())?;
+            if encoded(&value, limits.max_response_bytes)?.as_slice() != payload.get().as_bytes() {
+                return Err(CaptureRaftError::Protocol);
+            }
+            Ok(Ok(value))
+        }
+    }
+}
+fn restored_fatal(error: RemoteFailure) -> Result<Fatal<u64>, CaptureRaftError> {
+    Ok(match error {
+        RemoteFailure::Stopped => Fatal::Stopped,
+        RemoteFailure::Panicked => Fatal::Panicked,
+        RemoteFailure::StorageIo { operation } => {
+            let (subject, verb) = operation
+                .map(|op| (op.subject, op.verb))
+                .unwrap_or((ErrorSubject::None, ErrorVerb::Read));
+            Fatal::StorageError(
+                StorageIOError::new(
+                    subject,
+                    verb,
+                    openraft::AnyError::new(&std::io::Error::other("remote Raft storage I/O failed; detailed source redacted")),
+                )
+                .into(),
+            )
+        }
+        RemoteFailure::StorageDefensive { subject, violation } => Fatal::StorageError(StorageError::Defensive {
+            source: DefensiveError {
+                subject,
+                violation,
+                backtrace: None,
+            },
+        }),
+        RemoteFailure::SnapshotMismatch { .. } => return Err(CaptureRaftError::Protocol),
+    })
+}
+fn peers_valid(local: NodeIdentity, peers: &BTreeMap<u64, CaptureNode>) -> Result<(), CaptureRaftError> {
+    if local.cluster_id == [0; 16] || !peers.contains_key(&local.node_id) {
+        return Err(CaptureRaftError::Identity);
+    }
+    let mut addresses = std::collections::BTreeSet::new();
+    for node in peers.values() {
+        node.validate()?;
+        if !addresses.insert(node.rpc_uri()) {
+            return Err(CaptureRaftError::Identity);
+        }
+    }
+    Ok(())
+}
+
+#[cfg(test)]
+type WorkerPause = Arc<std::sync::Mutex<Option<(Arc<tokio::sync::Notify>, Arc<tokio::sync::Notify>)>>>;
+#[derive(Clone)]
+pub struct CaptureRaftRpc {
+    owner: Arc<NativeRaftOwner>,
+    peers: Arc<BTreeMap<u64, CaptureNode>>,
+    audience: String,
+    limits: CaptureRaftLimits,
+    admission: Arc<tokio::sync::Semaphore>,
+    accepting: Arc<std::sync::atomic::AtomicBool>,
+    #[cfg(test)]
+    pub(crate) worker_pause: WorkerPause,
+}
+impl CaptureRaftRpc {
+    pub fn new(
+        owner: Arc<NativeRaftOwner>,
+        peers: BTreeMap<u64, CaptureNode>,
+        limits: CaptureRaftLimits,
+    ) -> Result<Self, CaptureRaftError> {
+        limits.validate()?;
+        let identity = owner.identity();
+        peers_valid(identity, &peers)?;
+        let node = peers.get(&identity.node_id).ok_or(CaptureRaftError::Identity)?;
+        let uri: http::Uri = node.rpc_uri().parse().map_err(|_| CaptureRaftError::Identity)?;
+        let audience = http_auth::normalize_tonic_rpc_audience(uri.authority().ok_or(CaptureRaftError::Identity)?.as_str())?;
+        Ok(Self {
+            owner,
+            peers: Arc::new(peers),
+            audience,
+            limits,
+            #[cfg(test)]
+            worker_pause: Arc::new(std::sync::Mutex::new(None)),
+            accepting: Arc::new(std::sync::atomic::AtomicBool::new(true)),
+            admission: Arc::new(tokio::sync::Semaphore::new(limits.max_inflight)),
+        })
+    }
+    pub fn into_service(self) -> CaptureRaftServiceServer<Self> {
+        let request = self.limits.max_request_bytes + 64;
+        let response = self.limits.max_response_bytes + 128;
+        CaptureRaftServiceServer::new(self)
+            .max_decoding_message_size(request)
+            .max_encoding_message_size(response)
+    }
+    fn prepare<P: Serialize + DeserializeOwned>(
+        &self,
+        request: &Request<CaptureRaftRequest>,
+        method: Method,
+    ) -> Result<(WireRequest, P, tokio::sync::OwnedSemaphorePermit), Status> {
+        if !self.accepting.load(std::sync::atomic::Ordering::Acquire) {
+            return Err(Status::unavailable("capture Raft service stopping"));
+        }
+        let permit = self
+            .admission
+            .clone()
+            .try_acquire_owned()
+            .map_err(|_| Status::resource_exhausted("capture Raft admission full"))?;
+        if !self.accepting.load(std::sync::atomic::Ordering::Acquire) {
+            return Err(Status::unavailable("capture Raft service stopping"));
+        }
+        if request.get_ref().body.len() > self.limits.max_request_bytes {
+            return Err(Status::resource_exhausted("capture Raft request limit"));
+        }
+        if request
+            .metadata()
+            .get(http_auth::RPC_REPLAY_SCOPE_VERSION_HEADER)
+            .and_then(|v| v.to_str().ok())
+            != Some("3")
+        {
+            return Err(auth_rejected("missing_replay_scope"));
+        }
+        http_auth::verify_tonic_rpc_signature_with_bootstrap(&self.audience, method.path(), request.metadata().as_ref(), false)
+            .map_err(|_| auth_rejected("invalid_signature"))?;
+        http_auth::verify_tonic_canonical_body_digest(request, &canonical_request(&request.get_ref().body))
+            .map_err(|_| auth_rejected("body_digest_mismatch"))?;
+        let wire = decode_wire(&request.get_ref().body, self.limits.max_request_bytes)
+            .map_err(|_| Status::invalid_argument("capture Raft wire rejected"))?;
+        let local = self.owner.identity();
+        if wire.cluster != local.cluster_id
+            || wire.target != local.node_id
+            || wire.method != method
+            || !self.peers.contains_key(&wire.source)
+        {
+            return Err(Status::permission_denied("capture Raft identity rejected"));
+        }
+        let payload = decode_payload(&wire, self.limits.max_request_bytes)
+            .map_err(|_| Status::invalid_argument("capture Raft payload rejected"))?;
+        Ok((wire, payload, permit))
+    }
+    #[cfg(test)]
+    async fn paused(pause: WorkerPause) {
+        let gate = pause.lock().expect("worker pause lock").take();
+        if let Some((entered, resume)) = gate {
+            entered.notify_one();
+            resume.notified().await;
+        }
+    }
+    /// Stop admission and wait for all actual RPC tasks, including cancelled HTTP waiters.
+    pub async fn shutdown(&self) -> Result<(), CaptureRaftError> {
+        self.accepting.store(false, std::sync::atomic::Ordering::Release);
+        let count = u32::try_from(self.limits.max_inflight).map_err(|_| CaptureRaftError::Limits)?;
+        let _all = self
+            .admission
+            .clone()
+            .acquire_many_owned(count)
+            .await
+            .map_err(|_| CaptureRaftError::ChannelUnavailable)?;
+        Ok(())
+    }
+}
+fn signed_response<P: Serialize>(
+    raw: &CaptureRaftRequest,
+    result: Result<P, RemoteFailure>,
+    limits: CaptureRaftLimits,
+    audience: &str,
+    challenge: Option<uuid::Uuid>,
+) -> Result<Response<CaptureRaftResponse>, Status> {
+    let mut response =
+        Response::new(reply(raw, result, limits).map_err(|_| Status::internal("capture Raft response unavailable"))?);
+    if let Some(challenge) = challenge {
+        response.metadata_mut().as_mut().extend(
+            http_auth::tonic_boot_epoch_response_headers(audience, challenge)
+                .map_err(|_| Status::internal("capture Raft epoch proof unavailable"))?,
+        );
+    }
+    Ok(response)
+}
+#[tonic::async_trait]
+impl CaptureRaftService for CaptureRaftRpc {
+    async fn append_entries(&self, request: Request<CaptureRaftRequest>) -> Result<Response<CaptureRaftResponse>, Status> {
+        let (wire, payload, permit) = self.prepare::<AppendEntriesRequest<CaptureTypeConfig>>(&request, Method::AppendEntries)?;
+        validate_append(&payload, wire.source, self.limits)
+            .map_err(|_| Status::invalid_argument("capture Raft append rejected"))?;
+        let challenge = http_auth::tonic_boot_epoch_challenge(request.metadata().as_ref())
+            .map_err(|_| Status::unauthenticated("capture Raft challenge rejected"))?;
+        let raw = request.into_inner();
+        let owner = self.owner.clone();
+        let audience = self.audience.clone();
+        let limits = self.limits;
+        #[cfg(test)]
+        let pause = self.worker_pause.clone();
+        tokio::spawn(async move {
+            let _permit = permit;
+            #[cfg(test)]
+            Self::paused(pause).await;
+            let result = owner.raft().append_entries(payload).await.map_err(|error| match error {
+                RaftError::Fatal(error) => fatal(error),
+                RaftError::APIError(error) => match error {},
+            });
+            drop(owner);
+            signed_response(&raw, result, limits, &audience, challenge)
+        })
+        .await
+        .map_err(|_| Status::internal("capture Raft worker failed"))?
+    }
+    async fn vote(&self, request: Request<CaptureRaftRequest>) -> Result<Response<CaptureRaftResponse>, Status> {
+        let (wire, payload, permit) = self.prepare::<VoteRequest<u64>>(&request, Method::Vote)?;
+        if payload.vote.leader_id.node_id != wire.source {
+            return Err(Status::invalid_argument("capture Raft candidate rejected"));
+        }
+        let challenge = http_auth::tonic_boot_epoch_challenge(request.metadata().as_ref())
+            .map_err(|_| Status::unauthenticated("capture Raft challenge rejected"))?;
+        let raw = request.into_inner();
+        let owner = self.owner.clone();
+        let audience = self.audience.clone();
+        let limits = self.limits;
+        #[cfg(test)]
+        let pause = self.worker_pause.clone();
+        tokio::spawn(async move {
+            let _permit = permit;
+            #[cfg(test)]
+            Self::paused(pause).await;
+            let result = owner.raft().vote(payload).await.map_err(|error| match error {
+                RaftError::Fatal(error) => fatal(error),
+                RaftError::APIError(error) => match error {},
+            });
+            drop(owner);
+            signed_response(&raw, result, limits, &audience, challenge)
+        })
+        .await
+        .map_err(|_| Status::internal("capture Raft worker failed"))?
+    }
+    async fn install_snapshot(&self, request: Request<CaptureRaftRequest>) -> Result<Response<CaptureRaftResponse>, Status> {
+        let (wire, payload, permit) =
+            self.prepare::<InstallSnapshotRequest<CaptureTypeConfig>>(&request, Method::InstallSnapshot)?;
+        validate_snapshot(&payload, wire.source, self.limits)
+            .map_err(|_| Status::invalid_argument("capture Raft snapshot rejected"))?;
+        let challenge = http_auth::tonic_boot_epoch_challenge(request.metadata().as_ref())
+            .map_err(|_| Status::unauthenticated("capture Raft challenge rejected"))?;
+        let raw = request.into_inner();
+        let owner = self.owner.clone();
+        let audience = self.audience.clone();
+        let limits = self.limits;
+        #[cfg(test)]
+        let pause = self.worker_pause.clone();
+        tokio::spawn(async move {
+            let _permit = permit;
+            #[cfg(test)]
+            Self::paused(pause).await;
+            let result = owner.raft().install_snapshot(payload).await.map_err(|error| match error {
+                RaftError::Fatal(error) => fatal(error),
+                RaftError::APIError(error) => RemoteFailure::SnapshotMismatch { error },
+            });
+            drop(owner);
+            signed_response(&raw, result, limits, &audience, challenge)
+        })
+        .await
+        .map_err(|_| Status::internal("capture Raft worker failed"))?
+    }
+}
+fn validate_append(
+    request: &AppendEntriesRequest<CaptureTypeConfig>,
+    source: u64,
+    limits: CaptureRaftLimits,
+) -> Result<(), CaptureRaftError> {
+    if request.vote.leader_id.node_id != source {
+        return Err(CaptureRaftError::Identity);
+    }
+    if request.entries.len() > limits.max_entries {
+        return Err(CaptureRaftError::RequestTooLarge);
+    }
+    for entry in &request.entries {
+        if let openraft::EntryPayload::Membership(value) = &entry.payload {
+            CaptureMembership::try_from(&openraft::StoredMembership::new(Some(entry.log_id), value.clone()))
+                .map_err(|_| CaptureRaftError::Protocol)?;
+        }
+    }
+    Ok(())
+}
+fn validate_snapshot(
+    request: &InstallSnapshotRequest<CaptureTypeConfig>,
+    source: u64,
+    limits: CaptureRaftLimits,
+) -> Result<(), CaptureRaftError> {
+    if request.vote.leader_id.node_id != source
+        || request.data.len() > limits.max_snapshot_chunk_bytes
+        || request
+            .offset
+            .checked_add(u64::try_from(request.data.len()).map_err(|_| CaptureRaftError::Limits)?)
+            .is_none()
+        || request.meta.snapshot_id.is_empty()
+        || request.meta.snapshot_id.len() > 1024
+    {
+        return Err(CaptureRaftError::Limits);
+    }
+    CaptureMembership::try_from(&request.meta.last_membership).map_err(|_| CaptureRaftError::Protocol)?;
+    Ok(())
+}
+
+#[derive(Clone)]
+pub struct CaptureRaftNetworkFactory {
+    local: NodeIdentity,
+    peers: Arc<BTreeMap<u64, CaptureNode>>,
+    limits: CaptureRaftLimits,
+    admission: Arc<tokio::sync::Semaphore>,
+}
+impl CaptureRaftNetworkFactory {
+    pub fn new(
+        local: NodeIdentity,
+        peers: BTreeMap<u64, CaptureNode>,
+        config: Arc<openraft::Config>,
+        limits: CaptureRaftLimits,
+    ) -> Result<Self, CaptureRaftError> {
+        limits.validate()?;
+        peers_valid(local, &peers)?;
+        if usize::try_from(config.max_payload_entries).map_err(|_| CaptureRaftError::Limits)? > limits.max_entries {
+            return Err(CaptureRaftError::Limits);
+        }
+        if usize::try_from(config.snapshot_max_chunk_size).map_err(|_| CaptureRaftError::Limits)?
+            > limits.max_snapshot_chunk_bytes
+        {
+            return Err(CaptureRaftError::Limits);
+        }
+        Ok(Self {
+            local,
+            peers: Arc::new(peers),
+            limits,
+            admission: Arc::new(tokio::sync::Semaphore::new(limits.max_inflight)),
+        })
+    }
+}
+pub struct CaptureRaftPeer {
+    factory: CaptureRaftNetworkFactory,
+    target: u64,
+    node: CaptureNode,
+    valid: bool,
+}
+impl RaftNetworkFactory<CaptureTypeConfig> for CaptureRaftNetworkFactory {
+    type Network = CaptureRaftPeer;
+    async fn new_client(&mut self, target: u64, node: &CaptureNode) -> Self::Network {
+        CaptureRaftPeer {
+            factory: self.clone(),
+            target,
+            node: node.clone(),
+            valid: self.peers.get(&target) == Some(node),
+        }
+    }
+}
+impl CaptureRaftPeer {
+    async fn exchange<P: Serialize, R: Serialize + DeserializeOwned>(
+        &self,
+        method: Method,
+        payload: &P,
+        option: &RPCOption,
+    ) -> Result<Result<R, RemoteFailure>, CaptureRaftError> {
+        if !self.valid {
+            return Err(CaptureRaftError::Identity);
+        }
+        let ttl = self.factory.limits.request_timeout.min(option.soft_ttl());
+        tokio::time::timeout(ttl, async {
+            let _permit = self
+                .factory
+                .admission
+                .clone()
+                .acquire_owned()
+                .await
+                .map_err(|_| CaptureRaftError::ChannelUnavailable)?;
+            let raw =
+                encode_request(self.factory.local, self.target, method, payload, self.factory.limits).map_err(
+                    |error| match error {
+                        CaptureRaftError::Limits => CaptureRaftError::RequestTooLarge,
+                        error => error,
+                    },
+                )?;
+            let mut request = Request::new(raw.clone());
+            http_auth::set_tonic_canonical_body_digest(&mut request, &canonical_request(&raw.body))?;
+            request.set_timeout(ttl);
+            let mut client = client::capture_raft_time_out_client(
+                self.node.rpc_uri(),
+                self.factory.limits.connect_timeout.min(ttl),
+                self.factory.limits.max_request_bytes + 64,
+                self.factory.limits.max_response_bytes + 128,
+            )
+            .await?;
+            let response = match method {
+                Method::AppendEntries => client.append_entries(request).await,
+                Method::Vote => client.vote(request).await,
+                Method::InstallSnapshot => client.install_snapshot(request).await,
+            }
+            .map_err(|status| CaptureRaftError::Status(Box::new(status)))?;
+            verify_response(&raw, response.into_inner(), self.factory.limits)
+        })
+        .await
+        .map_err(|_| CaptureRaftError::Deadline)?
+    }
+    fn network_error<E: std::error::Error>(
+        &self,
+        error: CaptureRaftError,
+        method: Method,
+        option: &RPCOption,
+    ) -> RPCError<u64, CaptureNode, E> {
+        match error {
+            CaptureRaftError::Deadline => RPCError::Timeout(openraft::error::Timeout {
+                action: match method {
+                    Method::AppendEntries => openraft::RPCTypes::AppendEntries,
+                    Method::Vote => openraft::RPCTypes::Vote,
+                    Method::InstallSnapshot => openraft::RPCTypes::InstallSnapshot,
+                },
+                id: self.factory.local.node_id,
+                target: self.target,
+                timeout: self.factory.limits.request_timeout.min(option.soft_ttl()),
+            }),
+            error @ CaptureRaftError::ChannelUnavailable => RPCError::Unreachable(Unreachable::new(&error)),
+            error => RPCError::Network(openraft::error::NetworkError::new(&error)),
+        }
+    }
+    fn append_error<E: std::error::Error>(
+        &self,
+        error: CaptureRaftError,
+        count: usize,
+        option: &RPCOption,
+    ) -> RPCError<u64, CaptureNode, E> {
+        if matches!(error, CaptureRaftError::RequestTooLarge) && count > 1 {
+            let hint = u64::try_from((count / 2).max(1).min(self.factory.limits.max_entries)).unwrap_or(1);
+            return RPCError::PayloadTooLarge(openraft::error::PayloadTooLarge::new_entries_hint(hint));
+        }
+        self.network_error(error, Method::AppendEntries, option)
+    }
+    fn remote<E: std::error::Error>(&self, error: E) -> RPCError<u64, CaptureNode, E> {
+        RPCError::RemoteError(RemoteError::new_with_node(self.target, self.node.clone(), error))
+    }
+}
+impl RaftNetwork<CaptureTypeConfig> for CaptureRaftPeer {
+    async fn append_entries(
+        &mut self,
+        request: AppendEntriesRequest<CaptureTypeConfig>,
+        option: RPCOption,
+    ) -> Result<AppendEntriesResponse<u64>, RPCError<u64, CaptureNode, RaftError<u64>>> {
+        let count = request.entries.len();
+        validate_append(&request, self.factory.local.node_id, self.factory.limits)
+            .map_err(|error| self.append_error(error, count, &option))?;
+        match self
+            .exchange(Method::AppendEntries, &request, &option)
+            .await
+            .map_err(|error| self.append_error(error, count, &option))?
+        {
+            Ok(value) => Ok(value),
+            Err(error) => Err(self.remote(RaftError::Fatal(
+                restored_fatal(error).map_err(|error| self.network_error(error, Method::AppendEntries, &option))?,
+            ))),
+        }
+    }
+    async fn vote(
+        &mut self,
+        request: VoteRequest<u64>,
+        option: RPCOption,
+    ) -> Result<VoteResponse<u64>, RPCError<u64, CaptureNode, RaftError<u64>>> {
+        if request.vote.leader_id.node_id != self.factory.local.node_id {
+            return Err(self.network_error(CaptureRaftError::Identity, Method::Vote, &option));
+        }
+        match self
+            .exchange(Method::Vote, &request, &option)
+            .await
+            .map_err(|error| self.network_error(error, Method::Vote, &option))?
+        {
+            Ok(value) => Ok(value),
+            Err(error) => Err(self.remote(RaftError::Fatal(
+                restored_fatal(error).map_err(|error| self.network_error(error, Method::Vote, &option))?,
+            ))),
+        }
+    }
+    async fn install_snapshot(
+        &mut self,
+        request: InstallSnapshotRequest<CaptureTypeConfig>,
+        option: RPCOption,
+    ) -> Result<InstallSnapshotResponse<u64>, RPCError<u64, CaptureNode, RaftError<u64, InstallSnapshotError>>> {
+        validate_snapshot(&request, self.factory.local.node_id, self.factory.limits)
+            .map_err(|error| self.network_error(error, Method::InstallSnapshot, &option))?;
+        match self
+            .exchange(Method::InstallSnapshot, &request, &option)
+            .await
+            .map_err(|error| self.network_error(error, Method::InstallSnapshot, &option))?
+        {
+            Ok(value) => Ok(value),
+            Err(error) => {
+                let source = match error {
+                    RemoteFailure::SnapshotMismatch { error } => RaftError::APIError(error),
+                    error => RaftError::Fatal(
+                        restored_fatal(error).map_err(|error| self.network_error(error, Method::InstallSnapshot, &option))?,
+                    ),
+                };
+                Err(self.remote(source))
+            }
+        }
+    }
+}
+
+fn auth_rejected(reason: &'static str) -> Status {
+    use rustfs_io_metrics::internode_metrics::{
+        INTERNODE_OPERATION_GRPC_OTHER, INTERNODE_TRANSPORT_BACKEND_GRPC, global_internode_metrics,
+    };
+    global_internode_metrics().record_rpc_auth_failure_for_operation_and_backend(
+        INTERNODE_OPERATION_GRPC_OTHER,
+        INTERNODE_TRANSPORT_BACKEND_GRPC,
+        reason,
+    );
+    Status::unauthenticated("capture Raft authentication rejected")
+}
@@ -535,6 +535,62 @@ impl tonic::service::Interceptor for TonicInterceptor {
    }
 }

+pub(crate) type CaptureRaftClient = rustfs_protos::proto_gen::node_service::capture_raft_service_client::CaptureRaftServiceClient<
+    InterceptedService<AuthenticatedChannel, TonicInterceptor>,
+>;
+
+/// Capture RPCs always perform the real authenticated epoch handshake, never seed replay state.
+pub(crate) async fn capture_raft_time_out_client(
+    addr: &str,
+    connect_timeout: std::time::Duration,
+    request_max: usize,
+    response_max: usize,
+) -> Result<CaptureRaftClient, super::capture_raft::CaptureRaftError> {
+    use super::capture_raft::CaptureRaftError;
+    let interceptor = TonicInterceptor::Signature(gen_tonic_signature_interceptor()).with_rpc_audience(addr)?;
+    let channel = tokio::time::timeout(connect_timeout, async {
+        match runtime_sources::cached_node_channel(addr).await {
+            Some(channel) => Ok(channel),
+            None => create_new_channel(addr)
+                .await
+                .map_err(|_| CaptureRaftError::ChannelUnavailable),
+        }
+    })
+    .await
+    .map_err(|_| CaptureRaftError::Deadline)??;
+    let audience = interceptor
+        .replay_scope_audience()
+        .ok_or(CaptureRaftError::Authentication)?
+        .to_owned();
+    let authenticated = ReplayScopeChannel::new(channel, Some(audience.clone()));
+    let mut ping = NodeServiceClient::with_interceptor(authenticated.clone(), interceptor)
+        .max_encoding_message_size(1024)
+        .max_decoding_message_size(8192);
+    ping.ping(tonic::Request::new(rustfs_protos::proto_gen::node_service::PingRequest {
+        version: 1,
+        body: bytes::Bytes::new(),
+    }))
+    .await
+    .map_err(|status| CaptureRaftError::Status(Box::new(status)))?;
+    {
+        let peers = PEER_REPLAY_STATES.lock().map_err(|_| CaptureRaftError::Authentication)?;
+        let state = peers.get(&audience).ok_or(CaptureRaftError::Authentication)?;
+        if !matches!(state.cache_capability, Some(PeerReplayCapability::Capable { boot_epoch }) if Some(boot_epoch) == state.boot_epoch)
+        {
+            return Err(CaptureRaftError::Authentication);
+        }
+    }
+    let interceptor = TonicInterceptor::Signature(gen_tonic_signature_interceptor()).with_rpc_audience(addr)?;
+    Ok(
+        rustfs_protos::proto_gen::node_service::capture_raft_service_client::CaptureRaftServiceClient::with_interceptor(
+            authenticated,
+            interceptor,
+        )
+        .max_encoding_message_size(request_max)
+        .max_decoding_message_size(response_max),
+    )
+}
+
 #[cfg(test)]
 mod tests {
    use super::*;
@@ -13,6 +13,9 @@
 // limitations under the License.

 pub(crate) mod background_monitor;
+pub mod capture_raft;
+#[cfg(test)]
+mod capture_raft_tests;
 pub(crate) mod client;
 pub(crate) mod context_propagation;
 pub(crate) mod http_auth;
@@ -640,6 +640,7 @@ pub struct PeerLiveEventsBatch {
 pub struct ScannerPublicationLease {
    pub token: Uuid,
    pub movement_generation: u64,
+    pub observational_only: bool,
    /// Stable storage owner identity. This is distinct from the activity
    /// session and is bound into both acquire and release proofs.
    pub owner_id: String,
@@ -658,6 +659,7 @@ fn validate_scanner_publication_lease_response_fields(
    response: &ScannerPublicationLeaseResponse,
    expected_session_id: &str,
    expected_generation: u64,
+    expected_observational_only: bool,
 ) -> Result<(Uuid, String)> {
    if !response.success {
        return Err(Error::other(
@@ -674,6 +676,9 @@ fn validate_scanner_publication_lease_response_fields(
    if response.session_id != expected_session_id {
        return Err(Error::other("peer returned a different scanner publication lease session"));
    }
+    if response.observational_only != expected_observational_only {
+        return Err(Error::other("peer returned a different scanner publication lease purpose"));
+    }
    let owner_id = Uuid::parse_str(&response.owner_id)
        .ok()
        .filter(|owner_id| !owner_id.is_nil())
@@ -2418,6 +2423,7 @@ impl PeerRestClient {
        &self,
        expected_session_id: &str,
        expected_generation: u64,
+        observational_only: bool,
    ) -> Result<ScannerPublicationLease> {
        let request_started = std::time::Instant::now();
        self.finalize_result(
@@ -2434,6 +2440,7 @@ impl PeerRestClient {
                    ttl_ms: crate::store::SCANNER_PUBLICATION_LEASE_TTL_MS,
                    expected_session_id: expected_session_id.to_string(),
                    token: Bytes::new(),
+                    observational_only,
                });
                let canonical = rustfs_protos::canonical_scanner_publication_lease_request_body(request.get_ref())
                    .map_err(|_| Error::other("scanner publication lease request is too large to authenticate"))?;
@@ -2444,11 +2451,16 @@ impl PeerRestClient {
                        .map_err(|_| Error::other("scanner publication lease response is too large to authenticate"))?;
                verify_tonic_rpc_response_proof(&response_body, &response.response_proof)
                    .map_err(|_| Error::other("peer returned an invalid scanner publication lease proof"))?;
-                let (token, owner_id) =
-                    validate_scanner_publication_lease_response_fields(&response, expected_session_id, expected_generation)?;
+                let (token, owner_id) = validate_scanner_publication_lease_response_fields(
+                    &response,
+                    expected_session_id,
+                    expected_generation,
+                    observational_only,
+                )?;
                Ok(ScannerPublicationLease {
                    token,
                    movement_generation: response.movement_generation,
+                    observational_only,
                    owner_id,
                    session_id: response.session_id,
                    expires_at: scanner_publication_lease_deadline(
@@ -2483,6 +2495,7 @@ impl PeerRestClient {
                    ttl_ms: crate::store::SCANNER_PUBLICATION_LEASE_TTL_MS,
                    expected_session_id: lease.session_id.clone(),
                    token: lease.token.as_bytes().to_vec().into(),
+                    observational_only: lease.observational_only,
                });
                let canonical = rustfs_protos::canonical_scanner_publication_lease_request_body(request.get_ref())
                    .map_err(|_| Error::other("scanner publication lease validation request is too large to authenticate"))?;
@@ -2494,8 +2507,12 @@ impl PeerRestClient {
                    )?;
                verify_tonic_rpc_response_proof(&response_body, &response.response_proof)
                    .map_err(|_| Error::other("peer returned an invalid scanner publication lease validation proof"))?;
-                let (token, owner_id) =
-                    validate_scanner_publication_lease_response_fields(&response, &lease.session_id, lease.movement_generation)?;
+                let (token, owner_id) = validate_scanner_publication_lease_response_fields(
+                    &response,
+                    &lease.session_id,
+                    lease.movement_generation,
+                    lease.observational_only,
+                )?;
                if token != lease.token {
                    return Err(Error::other("peer returned a different scanner publication lease token"));
                }
@@ -3001,15 +3018,16 @@ mod tests {
            response_proof: Bytes::new(),
            owner_id: Uuid::new_v4().to_string(),
            session_id: "session-a".to_string(),
+            observational_only: false,
        };

-        assert!(validate_scanner_publication_lease_response_fields(&response, "session-a", 7).is_ok());
+        assert!(validate_scanner_publication_lease_response_fields(&response, "session-a", 7, false).is_ok());

        let stale_generation = ScannerPublicationLeaseResponse {
            movement_generation: 6,
            ..response.clone()
        };
-        let error = validate_scanner_publication_lease_response_fields(&stale_generation, "session-a", 7)
+        let error = validate_scanner_publication_lease_response_fields(&stale_generation, "session-a", 7, false)
            .expect_err("a response from an older movement generation must be rejected");
        assert!(error.to_string().contains("different scanner publication lease generation"));

@@ -3017,9 +3035,21 @@ mod tests {
            session_id: "session-b".to_string(),
            ..response
        };
-        let error = validate_scanner_publication_lease_response_fields(&stale_session, "session-a", 7)
+        let error = validate_scanner_publication_lease_response_fields(&stale_session, "session-a", 7, false)
            .expect_err("a response from an older scanner session must be rejected");
        assert!(error.to_string().contains("different scanner publication lease session"));
+
+        let observational = ScannerPublicationLeaseResponse {
+            observational_only: true,
+            ..stale_session
+        };
+        assert!(
+            validate_scanner_publication_lease_response_fields(&observational, "session-b", 7, true).is_ok(),
+            "the signed purpose echo should validate an observation lease"
+        );
+        let error = validate_scanner_publication_lease_response_fields(&observational, "session-b", 7, false)
+            .expect_err("a peer must not upgrade an observational lease to authoritative");
+        assert!(error.to_string().contains("different scanner publication lease purpose"));
    }

    #[test]
@@ -58,13 +58,15 @@ use rustfs_protos::proto_gen::node_service::{
    BatchReadVersionRequest, BatchReadVersionResponse, CheckPartsRequest, CompareAndUpdateFileOutcome,
    CompareAndUpdateFileRequest, DeletePathsRequest, DeleteRequest, DeleteVersionRequest, DeleteVersionsRequest,
    DeleteVersionsResponse, DeleteVolumeRequest, DiskInfoRequest, ListDirRequest, ListVolumesRequest, MakeVolumeRequest,
-    MakeVolumesRequest, PreparePartTransactionRequest, ReadAllRequest, ReadMetadataRequest, ReadMultipleRequest,
-    ReadMultipleResponse, ReadPartsRequest, ReadVersionRequest, ReadXlRequest, RenameDataRequest, RenameFileRequest,
-    SettlePartTransactionRequest, SnapshotLeaseReleaseRequest, SnapshotLeaseRenewRequest, SnapshotLeaseRequest,
-    SnapshotLeaseResponse, StatVolumeRequest, UpdateMetadataRequest, VerifyFileRequest, WriteAllRequest, WriteMetadataRequest,
-    node_service_client::NodeServiceClient,
+    MakeVolumesRequest, NativePreparationBeginRequest, NativePreparationFinalizeRequest, NativePreparationLocateRequest,
+    NativePreparationProbeRequest, NativePreparationRecoverRequest, PreparePartTransactionRequest, ReadAllRequest,
+    ReadMetadataRequest, ReadMultipleRequest, ReadMultipleResponse, ReadPartsRequest, ReadVersionRequest, ReadXlRequest,
+    RenameDataRequest, RenameFileRequest, SettlePartTransactionRequest, SnapshotLeaseReleaseRequest, SnapshotLeaseRenewRequest,
+    SnapshotLeaseRequest, SnapshotLeaseResponse, StatVolumeRequest, UpdateMetadataRequest, VerifyFileRequest, WriteAllRequest,
+    WriteMetadataRequest, node_service_client::NodeServiceClient,
 };
 use serde::{Serialize, de::DeserializeOwned};
+use sha2::{Digest, Sha256};
 use std::{
    future::Future,
    io::Cursor,
@@ -176,6 +178,172 @@ fn attach_mutation_body_digest<T>(
    crate::cluster::rpc::set_tonic_rolling_canonical_body_digest(request, &canonical_body).map_err(Error::other)
 }

+fn native_response_proof(body: std::result::Result<Vec<u8>, std::num::TryFromIntError>, proof: &[u8]) -> Result<()> {
+    let body = body.map_err(|_| Error::other("native preparation response is too large to authenticate"))?;
+    if proof.len() != 32 {
+        return Err(Error::other("native preparation response proof has invalid length"));
+    }
+    crate::cluster::rpc::verify_tonic_rpc_response_proof(&body, proof)
+        .map_err(|_| Error::other("native preparation response proof is invalid"))
+}
+
+fn native_response_error(error: Option<rustfs_protos::proto_gen::node_service::Error>) -> DiskError {
+    let error = DiskError::from(error.unwrap_or_default());
+    match error {
+        DiskError::FileCorrupt => DiskError::Io(std::io::Error::new(std::io::ErrorKind::InvalidData, error)),
+        error => error,
+    }
+}
+
+fn native_uuid_field(value: &[u8], field: &str) -> Result<Uuid> {
+    let id = Uuid::from_slice(value).map_err(|_| Error::other(format!("native preparation returned invalid {field}")))?;
+    if id.is_nil() {
+        return Err(Error::other(format!("native preparation returned nil {field}")));
+    }
+    Ok(id)
+}
+
+fn native_version_field(value: &[u8]) -> Result<Uuid> {
+    Uuid::from_slice(value).map_err(|_| Error::other("native preparation returned invalid version id"))
+}
+
+fn native_object_length_from_metadata(input: &crate::disk::local::native_prepare::NativePrepareInput) -> Result<u64> {
+    let raw = rustfs_filemeta::FileMeta::load(&input.metadata).map_err(Error::other)?;
+    if raw.versions.len() != 1 {
+        return Err(Error::other("native preparation metadata must have one version"));
+    }
+    let info = rustfs_filemeta::get_file_info(
+        &input.metadata,
+        &input.bucket,
+        &input.key,
+        &input.version.to_string(),
+        rustfs_filemeta::FileInfoOpts {
+            data: true,
+            include_free_versions: false,
+            include_part_checksums: true,
+        },
+    )
+    .map_err(Error::other)?;
+    if info.version_id != Some(input.version) || info.deleted || info.mark_deleted || info.size < 0 {
+        return Err(Error::other("native preparation metadata identity mismatch"));
+    }
+    u64::try_from(info.size).map_err(Error::other)
+}
+
+fn validate_native_metadata_for_version(
+    metadata: &[u8],
+    expected_digest: Option<&[u8; 32]>,
+    bucket: &str,
+    key: &str,
+    version: Uuid,
+) -> Result<()> {
+    validate_native_metadata(metadata, expected_digest)?;
+    let raw = rustfs_filemeta::FileMeta::load(metadata).map_err(Error::other)?;
+    if raw.versions.len() != 1 {
+        return Err(Error::other("native preparation metadata must have one version"));
+    }
+    let info = rustfs_filemeta::get_file_info(
+        metadata,
+        bucket,
+        key,
+        &version.to_string(),
+        rustfs_filemeta::FileInfoOpts {
+            data: true,
+            include_free_versions: false,
+            include_part_checksums: true,
+        },
+    )
+    .map_err(Error::other)?;
+    if info.version_id != Some(version) || info.deleted || info.mark_deleted || info.size < 0 {
+        return Err(Error::other("native preparation metadata identity mismatch"));
+    }
+    Ok(())
+}
+
+fn native_digest_field(value: &[u8], field: &str) -> Result<[u8; 32]> {
+    value
+        .try_into()
+        .map_err(|_| Error::other(format!("native preparation returned invalid {field}")))
+}
+
+fn native_receipt_from_wire(
+    receipt: rustfs_protos::proto_gen::node_service::NativePreparationReceipt,
+    expected_disk: Uuid,
+    expected_store: Uuid,
+    expected_preparation: Uuid,
+    expected_bucket: &str,
+    expected_key: &str,
+    expected_incarnation: Uuid,
+    expected_object: Option<(&[u8], u64)>,
+    expected_version: Option<Uuid>,
+) -> Result<crate::disk::local::native_prepare::NativePreparationReceipt> {
+    let disk = native_uuid_field(&receipt.disk_id, "disk id")?;
+    let store = native_uuid_field(&receipt.store_id, "store id")?;
+    let deployment = native_uuid_field(&receipt.deployment_id, "deployment id")?;
+    let preparation = native_uuid_field(&receipt.preparation_id, "preparation id")?;
+    let version = native_version_field(&receipt.version_id)?;
+    let incarnation = native_uuid_field(&receipt.bucket_incarnation_id, "bucket incarnation")?;
+    if disk != expected_disk
+        || store != expected_store
+        || deployment != expected_store
+        || preparation != expected_preparation
+        || incarnation != expected_incarnation
+        || receipt.bucket != expected_bucket
+        || receipt.key != expected_key
+        || expected_version.is_some_and(|expected| version != expected)
+    {
+        return Err(Error::other("native preparation receipt identity mismatch"));
+    }
+    let object_digest = native_digest_field(&receipt.object_digest, "object digest")?;
+    if expected_object.is_some_and(|(digest, length)| digest != object_digest.as_slice() || length != receipt.object_length) {
+        return Err(Error::other("native preparation receipt object mismatch"));
+    }
+    let set_digest = native_digest_field(&receipt.set_digest, "set digest")?;
+    let distribution_digest = native_digest_field(&receipt.distribution_digest, "distribution digest")?;
+    let metadata_digest = native_digest_field(&receipt.metadata_digest, "metadata digest")?;
+    let shard_digest = native_digest_field(&receipt.shard_digest, "shard digest")?;
+    let data_blocks = usize::try_from(receipt.data_blocks).map_err(Error::other)?;
+    let parity_blocks = usize::try_from(receipt.parity_blocks).map_err(Error::other)?;
+    let shard_index = usize::try_from(receipt.shard_index).map_err(Error::other)?;
+    if data_blocks == 0
+        || parity_blocks == 0
+        || data_blocks
+            .checked_add(parity_blocks)
+            .is_none_or(|total| shard_index == 0 || shard_index > total)
+    {
+        return Err(Error::other("native preparation receipt has invalid erasure geometry"));
+    }
+    Ok(crate::disk::local::native_prepare::NativePreparationReceipt {
+        preparation,
+        version,
+        disk,
+        set_digest,
+        distribution_digest,
+        metadata_digest,
+        shard_digest,
+        shard_length: receipt.shard_length,
+        bucket: receipt.bucket,
+        key: receipt.key,
+        deployment,
+        incarnation,
+        object_digest,
+        object_length: receipt.object_length,
+        data_blocks,
+        parity_blocks,
+        shard_index,
+    })
+}
+
+fn validate_native_metadata(metadata: &[u8], expected_digest: Option<&[u8; 32]>) -> Result<()> {
+    if metadata.len() > rustfs_protos::NATIVE_PREPARATION_METADATA_MAX_BYTES {
+        return Err(Error::other("native preparation metadata exceeds the protocol limit"));
+    }
+    if expected_digest.is_some_and(|expected| Sha256::digest(metadata).as_slice() != expected.as_slice()) {
+        return Err(Error::other("native preparation metadata digest mismatch"));
+    }
+    Ok(())
+}
+
 fn decode_volume_infos(volume_infos: Vec<String>) -> Result<Vec<VolumeInfo>> {
    volume_infos
        .into_iter()
@@ -896,6 +1064,339 @@ impl RemoteDisk {
        .await
    }

+    pub(crate) async fn native_begin_stream(
+        &self,
+        preparation: Uuid,
+        store_id: Uuid,
+        bucket: &str,
+        key: &str,
+        incarnation: Uuid,
+    ) -> Result<()> {
+        let disk_id = self
+            .get_disk_id()
+            .await?
+            .filter(|id| !id.is_nil())
+            .ok_or(DiskError::InvalidPath)?;
+        if preparation.is_nil() || store_id.is_nil() || incarnation.is_nil() || bucket.is_empty() || key.is_empty() {
+            return Err(DiskError::InvalidPath);
+        }
+        self.execute_with_timeout_for_op(
+            "native_preparation_begin",
+            || async {
+                let mut client = self.get_client().await?;
+                let mut request = Request::new(NativePreparationBeginRequest {
+                    disk: self.disk_ref().await,
+                    disk_id: disk_id.as_bytes().to_vec().into(),
+                    store_id: store_id.as_bytes().to_vec().into(),
+                    preparation_id: preparation.as_bytes().to_vec().into(),
+                    bucket: bucket.to_owned(),
+                    key: key.to_owned(),
+                    bucket_incarnation_id: incarnation.as_bytes().to_vec().into(),
+                });
+                let canonical = rustfs_protos::canonical_native_preparation_begin_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical, "native_preparation_begin")?;
+                let canonical_request = request.get_ref().clone();
+                let response = client.native_preparation_begin(request).await?.into_inner();
+                native_response_proof(
+                    rustfs_protos::canonical_native_preparation_begin_response_body(&canonical_request, &response),
+                    &response.response_proof,
+                )?;
+                if !response.success {
+                    return Err(native_response_error(response.error));
+                }
+                Ok(())
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    pub(crate) async fn native_finalize(
+        &self,
+        input: crate::disk::local::native_prepare::NativePrepareInput,
+    ) -> Result<crate::disk::local::native_prepare::NativePreparationReceipt> {
+        let disk_id = self
+            .get_disk_id()
+            .await?
+            .filter(|id| !id.is_nil())
+            .ok_or(DiskError::InvalidPath)?;
+        if input.preparation.is_nil() || input.store_id.is_nil() || input.bucket_incarnation.is_nil() {
+            return Err(DiskError::InvalidPath);
+        }
+        validate_native_metadata(&input.metadata, None)?;
+        let expected_object_digest = input.object_digest;
+        let expected_object_length = native_object_length_from_metadata(&input)?;
+        let expected_version = input.version;
+        self.execute_with_timeout_for_op(
+            "native_preparation_finalize",
+            || async move {
+                let mut client = self.get_client().await?;
+                let mut request = Request::new(NativePreparationFinalizeRequest {
+                    disk: self.disk_ref().await,
+                    disk_id: disk_id.as_bytes().to_vec().into(),
+                    store_id: input.store_id.as_bytes().to_vec().into(),
+                    preparation_id: input.preparation.as_bytes().to_vec().into(),
+                    bucket: input.bucket.clone(),
+                    key: input.key.clone(),
+                    bucket_incarnation_id: input.bucket_incarnation.as_bytes().to_vec().into(),
+                    object_digest: expected_object_digest.to_vec().into(),
+                    object_length: expected_object_length,
+                    version_id: expected_version.as_bytes().to_vec().into(),
+                    metadata: input.metadata.clone(),
+                });
+                let canonical = rustfs_protos::canonical_native_preparation_finalize_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical, "native_preparation_finalize")?;
+                let canonical_request = request.get_ref().clone();
+                let response = client.native_preparation_finalize(request).await?.into_inner();
+                native_response_proof(
+                    rustfs_protos::canonical_native_preparation_finalize_response_body(&canonical_request, &response),
+                    &response.response_proof,
+                )?;
+                if !response.success {
+                    return Err(native_response_error(response.error));
+                }
+                let receipt = response
+                    .receipt
+                    .ok_or_else(|| Error::other("native finalize response omitted its receipt"))?;
+                let receipt = native_receipt_from_wire(
+                    receipt,
+                    disk_id,
+                    input.store_id,
+                    input.preparation,
+                    &input.bucket,
+                    &input.key,
+                    input.bucket_incarnation,
+                    Some((&expected_object_digest, expected_object_length)),
+                    Some(expected_version),
+                )?;
+                validate_native_metadata_for_version(
+                    &input.metadata,
+                    Some(&receipt.metadata_digest),
+                    &input.bucket,
+                    &input.key,
+                    expected_version,
+                )?;
+                Ok(receipt)
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    pub(crate) async fn native_probe(
+        &self,
+        preparation: Uuid,
+        store_id: Uuid,
+        bucket: &str,
+        key: &str,
+        incarnation: Uuid,
+    ) -> Result<crate::disk::local::native_prepare::NativePreparationProbe> {
+        let disk_id = self
+            .get_disk_id()
+            .await?
+            .filter(|id| !id.is_nil())
+            .ok_or(DiskError::InvalidPath)?;
+        if preparation.is_nil() || store_id.is_nil() || incarnation.is_nil() || bucket.is_empty() || key.is_empty() {
+            return Err(DiskError::InvalidPath);
+        }
+        self.execute_with_timeout_for_op(
+            "native_preparation_probe",
+            || async {
+                let mut client = self.get_client().await?;
+                let mut request = Request::new(NativePreparationProbeRequest {
+                    disk: self.disk_ref().await,
+                    disk_id: disk_id.as_bytes().to_vec().into(),
+                    store_id: store_id.as_bytes().to_vec().into(),
+                    preparation_id: preparation.as_bytes().to_vec().into(),
+                    bucket: bucket.to_owned(),
+                    key: key.to_owned(),
+                    bucket_incarnation_id: incarnation.as_bytes().to_vec().into(),
+                });
+                let canonical = rustfs_protos::canonical_native_preparation_probe_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical, "native_preparation_probe")?;
+                let canonical_request = request.get_ref().clone();
+                let response = client.native_preparation_probe(request).await?.into_inner();
+                native_response_proof(
+                    rustfs_protos::canonical_native_preparation_probe_response_body(&canonical_request, &response),
+                    &response.response_proof,
+                )?;
+                if !response.success {
+                    return Err(native_response_error(response.error));
+                }
+                validate_native_metadata(&response.metadata, None)?;
+                if !response.exists {
+                    if response.prepared
+                        || !response.metadata.is_empty()
+                        || !response.version_id.is_empty()
+                        || !response.object_digest.is_empty()
+                        || response.object_length != 0
+                        || !response.set_digest.is_empty()
+                        || !response.distribution_digest.is_empty()
+                    {
+                        return Err(Error::other("native probe absent response carries object state"));
+                    }
+                    return Ok(crate::disk::local::native_prepare::NativePreparationProbe::Absent);
+                }
+                let version = native_version_field(&response.version_id)?;
+                let object_digest = native_digest_field(&response.object_digest, "object digest")?;
+                let set_digest = native_digest_field(&response.set_digest, "set digest")?;
+                let distribution_digest = native_digest_field(&response.distribution_digest, "distribution digest")?;
+                if !response.metadata.is_empty() {
+                    validate_native_metadata_for_version(&response.metadata, None, bucket, key, version)?;
+                }
+                Ok(crate::disk::local::native_prepare::NativePreparationProbe::Existing {
+                    version,
+                    prepared: response.prepared,
+                    object_digest,
+                    object_length: response.object_length,
+                    set_digest,
+                    distribution_digest,
+                    metadata: (!response.metadata.is_empty()).then_some(response.metadata),
+                })
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    pub(crate) async fn native_recover(
+        &self,
+        preparation: Uuid,
+        store_id: Uuid,
+        bucket: &str,
+        key: &str,
+        incarnation: Uuid,
+    ) -> Result<(crate::disk::local::native_prepare::NativePreparationReceipt, Bytes)> {
+        let disk_id = self
+            .get_disk_id()
+            .await?
+            .filter(|id| !id.is_nil())
+            .ok_or(DiskError::InvalidPath)?;
+        if preparation.is_nil() || store_id.is_nil() || incarnation.is_nil() || bucket.is_empty() || key.is_empty() {
+            return Err(DiskError::InvalidPath);
+        }
+        self.execute_with_timeout_for_op(
+            "native_preparation_recover",
+            || async {
+                let mut client = self.get_client().await?;
+                let mut request = Request::new(NativePreparationRecoverRequest {
+                    disk: self.disk_ref().await,
+                    disk_id: disk_id.as_bytes().to_vec().into(),
+                    store_id: store_id.as_bytes().to_vec().into(),
+                    preparation_id: preparation.as_bytes().to_vec().into(),
+                    bucket: bucket.to_owned(),
+                    key: key.to_owned(),
+                    bucket_incarnation_id: incarnation.as_bytes().to_vec().into(),
+                });
+                let canonical = rustfs_protos::canonical_native_preparation_recover_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical, "native_preparation_recover")?;
+                let canonical_request = request.get_ref().clone();
+                let response = client.native_preparation_recover(request).await?.into_inner();
+                native_response_proof(
+                    rustfs_protos::canonical_native_preparation_recover_response_body(&canonical_request, &response),
+                    &response.response_proof,
+                )?;
+                if !response.success {
+                    return Err(native_response_error(response.error));
+                }
+                let receipt = response
+                    .receipt
+                    .ok_or_else(|| Error::other("native recover response omitted its receipt"))?;
+                let receipt =
+                    native_receipt_from_wire(receipt, disk_id, store_id, preparation, bucket, key, incarnation, None, None)?;
+                validate_native_metadata_for_version(
+                    &response.metadata,
+                    Some(&receipt.metadata_digest),
+                    bucket,
+                    key,
+                    receipt.version,
+                )?;
+                Ok((receipt, response.metadata))
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    pub(crate) async fn native_locate(
+        &self,
+        store_id: Uuid,
+        preparation: Uuid,
+    ) -> Result<Option<crate::disk::local::native_prepare::NativePreparationDescription>> {
+        let disk_id = self
+            .get_disk_id()
+            .await?
+            .filter(|id| !id.is_nil())
+            .ok_or(DiskError::InvalidPath)?;
+        if preparation.is_nil() || store_id.is_nil() {
+            return Err(DiskError::InvalidPath);
+        }
+        self.execute_with_timeout_for_op(
+            "native_preparation_locate",
+            || async {
+                let mut client = self.get_client().await?;
+                let mut request = Request::new(NativePreparationLocateRequest {
+                    disk: self.disk_ref().await,
+                    disk_id: disk_id.as_bytes().to_vec().into(),
+                    store_id: store_id.as_bytes().to_vec().into(),
+                    preparation_id: preparation.as_bytes().to_vec().into(),
+                });
+                let canonical = rustfs_protos::canonical_native_preparation_locate_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical, "native_preparation_locate")?;
+                let canonical_request = request.get_ref().clone();
+                let response = client.native_preparation_locate(request).await?.into_inner();
+                native_response_proof(
+                    rustfs_protos::canonical_native_preparation_locate_response_body(&canonical_request, &response),
+                    &response.response_proof,
+                )?;
+                if !response.success {
+                    return Err(native_response_error(response.error));
+                }
+                if !response.found {
+                    if !response.bucket.is_empty()
+                        || !response.key.is_empty()
+                        || !response.bucket_incarnation_id.is_empty()
+                        || !response.metadata.is_empty()
+                        || response.prepared
+                    {
+                        return Err(Error::other("native locate absent response carries preparation state"));
+                    }
+                    return Ok(None);
+                }
+                if response.bucket.is_empty()
+                    || response.bucket.len() > 255
+                    || response.key.is_empty()
+                    || response.key.len() > 4096
+                    || preparation.is_nil()
+                {
+                    return Err(Error::other("native locate response has invalid description"));
+                }
+                let incarnation = native_uuid_field(&response.bucket_incarnation_id, "bucket incarnation")?;
+                validate_native_metadata(&response.metadata, None)?;
+                if response.metadata.is_empty() {
+                    return Err(Error::other("native locate response omitted metadata"));
+                }
+                let raw_metadata = rustfs_filemeta::FileMeta::load(&response.metadata)
+                    .map_err(|_| Error::other("native locate response metadata is invalid"))?;
+                if raw_metadata.versions.len() != 1 {
+                    return Err(Error::other("native locate metadata must have one version"));
+                }
+                let version = raw_metadata.versions[0].header.version_id.unwrap_or_else(Uuid::nil);
+                Ok(Some(crate::disk::local::native_prepare::NativePreparationDescription {
+                    preparation,
+                    bucket: response.bucket,
+                    key: response.key,
+                    incarnation,
+                    version,
+                    metadata: response.metadata,
+                    prepared: response.prepared,
+                }))
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
    pub(crate) async fn ns_scanner_server_epoch(&self) -> Result<Option<Uuid>> {
        if self.health.is_faulty() {
            return Err(DiskError::FaultyDisk);
@@ -449,6 +449,64 @@ where
    Ok(data)
 }

+pub(crate) async fn read_config_observation<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<(Vec<u8>, ObjectInfo)>
+where
+    S: EcstoreObjectIO,
+{
+    let opts = ObjectOptions {
+        suppress_read_repair: true,
+        no_lock: false,
+        skip_verify_bitrot: false,
+        ..Default::default()
+    };
+    crate::object_api::without_get_object_body_cache_hook(read_config_limited_preserve_empty_with_metadata_opts(
+        api, file, &opts, max_bytes,
+    ))
+    .await
+}
+
+pub(crate) async fn read_config_observation_no_lock<S>(
+    api: Arc<S>,
+    file: &str,
+    max_bytes: usize,
+    guard: &crate::object_api::WriteCommitGuard,
+) -> Result<(Vec<u8>, ObjectInfo)>
+where
+    S: EcstoreObjectIO,
+{
+    if !guard.matches(RUSTFS_META_BUCKET, file) {
+        return Err(Error::InvalidArgument(
+            RUSTFS_META_BUCKET.to_owned(),
+            file.to_owned(),
+            "locked observation requires a matching write owner".to_owned(),
+        ));
+    }
+    let lock_lost_error = || Error::NamespaceLockQuorumUnavailable {
+        mode: "locked_config_observation",
+        bucket: RUSTFS_META_BUCKET.to_owned(),
+        object: file.to_owned(),
+        required: 1,
+        achieved: 0,
+    };
+    if guard.is_lock_lost() {
+        return Err(lock_lost_error());
+    }
+    let opts = ObjectOptions {
+        suppress_read_repair: true,
+        no_lock: true,
+        skip_verify_bitrot: false,
+        ..Default::default()
+    };
+    let result = crate::object_api::without_get_object_body_cache_hook(read_config_limited_preserve_empty_with_metadata_opts(
+        api, file, &opts, max_bytes,
+    ))
+    .await;
+    if guard.is_lock_lost() {
+        return Err(lock_lost_error());
+    }
+    result
+}
+
 pub(crate) async fn read_config_limited_preserve_empty<S>(api: Arc<S>, file: &str, max_bytes: usize) -> Result<Vec<u8>>
 where
    S: EcstoreObjectIO,
@@ -495,10 +553,22 @@ pub(crate) async fn read_config_preserve_empty<S>(api: Arc<S>, file: &str) -> Re
 where
    S: EcstoreObjectIO,
 {
-    let (data, _obj) = read_config_with_metadata_inner(api, file, &ObjectOptions::default(), true, None).await?;
+    let (data, _obj) = read_config_preserve_empty_with_metadata_opts(api, file, &ObjectOptions::default()).await?;
    Ok(data)
 }

+/// Preserve an empty config payload, its metadata, and caller-supplied read options.
+pub(crate) async fn read_config_preserve_empty_with_metadata_opts<S>(
+    api: Arc<S>,
+    file: &str,
+    opts: &ObjectOptions,
+) -> Result<(Vec<u8>, ObjectInfo)>
+where
+    S: EcstoreObjectIO,
+{
+    read_config_with_metadata_inner(api, file, opts, true, None).await
+}
+
 pub async fn read_config_no_lock<S>(api: Arc<S>, file: &str) -> Result<Vec<u8>>
 where
    S: ObjectIO<
@@ -527,15 +597,13 @@ pub(crate) async fn read_config_no_lock_preserve_empty_with_metadata<S>(api: Arc
 where
    S: EcstoreObjectIO,
 {
-    read_config_with_metadata_inner(
+    read_config_preserve_empty_with_metadata_opts(
        api,
        file,
        &ObjectOptions {
            no_lock: true,
            ..Default::default()
        },
-        true,
-        None,
    )
    .await
 }
@@ -2815,9 +2883,10 @@ mod tests {
        config_task_join_error, configs_semantically_equal, decode_server_config_blob, encode_server_config_blob,
        heal_config_descriptor, is_standard_object_server_config, lookup_configs, map_system_metadata_write_error,
        new_and_save_server_config, read_config, read_config_no_lock_preserve_empty_with_metadata, read_config_preserve_empty,
-        read_config_with_metadata, read_config_without_migrate, read_server_config_snapshot, save_config_with_opts_inner,
-        save_server_config, save_server_config_snapshot, save_server_config_snapshot_with_generation,
-        server_config_transaction_lock_path, should_warn_ignored_scalar_section, storage_class_kvs_mut,
+        read_config_preserve_empty_with_metadata_opts, read_config_with_metadata, read_config_without_migrate,
+        read_server_config_snapshot, save_config_with_opts_inner, save_server_config, save_server_config_snapshot,
+        save_server_config_snapshot_with_generation, server_config_transaction_lock_path, should_warn_ignored_scalar_section,
+        storage_class_kvs_mut,
    };
    use crate::config::{audit, heal, notify, oidc, scanner};
    use crate::disk::{RUSTFS_META_BUCKET, endpoint::Endpoint};
@@ -5065,6 +5134,7 @@ mod tests {
        heal_replacement: Option<Vec<u8>>,
        heal_calls: AtomicUsize,
        write_calls: AtomicUsize,
+        last_get_no_lock: AtomicBool,
        last_put_no_lock: AtomicBool,
        last_put_preconditions: Mutex<Option<HTTPPreconditions>>,
        revision: AtomicUsize,
@@ -5080,6 +5150,7 @@ mod tests {
                heal_replacement,
                heal_calls: AtomicUsize::new(0),
                write_calls: AtomicUsize::new(0),
+                last_get_no_lock: AtomicBool::new(false),
                last_put_no_lock: AtomicBool::new(false),
                last_put_preconditions: Mutex::new(None),
                revision: AtomicUsize::new(1),
@@ -5135,8 +5206,9 @@ mod tests {
            _object: &str,
            _range: Option<HTTPRangeSpec>,
            _h: HeaderMap,
-            _opts: &ObjectOptions,
+            opts: &ObjectOptions,
        ) -> Result<GetObjectReader> {
+            self.last_get_no_lock.store(opts.no_lock, Ordering::SeqCst);
            let data = match &*self.state.lock().expect("state lock poisoned") {
                RecoveryReadState::Missing => return Err(Error::ConfigNotFound),
                RecoveryReadState::Blob(data) => data.clone(),
@@ -5411,6 +5483,329 @@ mod tests {
        assert_eq!(store.write_calls.load(Ordering::SeqCst), 0);
    }

+    struct ObservationStream {
+        inner: Cursor<Vec<u8>>,
+        consumed: Arc<AtomicUsize>,
+        fail: bool,
+        eof_gate: Option<(Option<tokio::sync::oneshot::Sender<()>>, tokio::sync::oneshot::Receiver<()>)>,
+    }
+
+    impl AsyncRead for ObservationStream {
+        fn poll_read(mut self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &mut ReadBuf<'_>) -> Poll<std::io::Result<()>> {
+            if self.fail {
+                return Poll::Ready(Err(std::io::Error::other("observation stream failed")));
+            }
+            if self.inner.position() == u64::try_from(self.inner.get_ref().len()).expect("fixture length")
+                && let Some((ready, release)) = self.eof_gate.as_mut()
+            {
+                if let Some(ready) = ready.take() {
+                    ready.send(()).expect("EOF observer remains active");
+                }
+                match std::future::Future::poll(Pin::new(release), cx) {
+                    Poll::Pending => return Poll::Pending,
+                    Poll::Ready(result) => result.expect("EOF gate released"),
+                }
+                self.eof_gate = None;
+            }
+            let before = buf.filled().len();
+            let result = Pin::new(&mut self.inner).poll_read(cx, buf);
+            self.consumed.fetch_add(buf.filled().len() - before, Ordering::SeqCst);
+            result
+        }
+    }
+
+    struct ObservationStore {
+        body: Vec<u8>,
+        declared_size: i64,
+        error: Option<Error>,
+        stream_error: bool,
+        gets: AtomicUsize,
+        eof_gate: std::sync::Mutex<Option<(Option<tokio::sync::oneshot::Sender<()>>, tokio::sync::oneshot::Receiver<()>)>>,
+        consumed: Arc<AtomicUsize>,
+        options: std::sync::Mutex<Option<(bool, bool, bool)>>,
+        writes: AtomicUsize,
+    }
+
+    impl Debug for ObservationStore {
+        fn fmt(&self, f: &mut Formatter<'_>) -> std::fmt::Result {
+            f.debug_struct("ObservationStore").finish()
+        }
+    }
+
+    impl ObservationStore {
+        fn new(body: Vec<u8>, declared_size: i64) -> Self {
+            Self {
+                body,
+                declared_size,
+                error: None,
+                stream_error: false,
+                gets: AtomicUsize::new(0),
+                eof_gate: std::sync::Mutex::new(None),
+                consumed: Arc::new(AtomicUsize::new(0)),
+                options: std::sync::Mutex::new(None),
+                writes: AtomicUsize::new(0),
+            }
+        }
+    }
+
+    #[async_trait::async_trait]
+    impl crate::storage_api_contracts::object::ObjectIO for ObservationStore {
+        type Error = Error;
+        type RangeSpec = HTTPRangeSpec;
+        type HeaderMap = HeaderMap;
+        type ObjectOptions = ObjectOptions;
+        type ObjectInfo = ObjectInfo;
+        type GetObjectReader = GetObjectReader;
+        type PutObjectReader = PutObjReader;
+
+        async fn get_object_reader(
+            &self,
+            bucket: &str,
+            object: &str,
+            range: Option<HTTPRangeSpec>,
+            _headers: HeaderMap,
+            opts: &ObjectOptions,
+        ) -> Result<GetObjectReader> {
+            assert_eq!(bucket, RUSTFS_META_BUCKET);
+            assert_eq!(object, "config/observation.bin");
+            self.gets.fetch_add(1, Ordering::SeqCst);
+            assert!(range.is_none());
+            *self.options.lock().expect("observation options") =
+                Some((opts.suppress_read_repair, opts.no_lock, opts.skip_verify_bitrot));
+            if let Some(error) = &self.error {
+                return Err(error.clone());
+            }
+            Ok(GetObjectReader {
+                stream: Box::new(ObservationStream {
+                    inner: Cursor::new(self.body.clone()),
+                    consumed: self.consumed.clone(),
+                    fail: self.stream_error,
+                    eof_gate: self.eof_gate.lock().expect("EOF gate").take(),
+                }),
+                object_info: ObjectInfo {
+                    size: self.declared_size,
+                    etag: Some("observation-etag".to_string()),
+                    ..Default::default()
+                },
+                buffered_body: None,
+                body_source: Default::default(),
+            })
+        }
+
+        async fn put_object(
+            &self,
+            _bucket: &str,
+            _object: &str,
+            _data: &mut PutObjReader,
+            _opts: &ObjectOptions,
+        ) -> Result<ObjectInfo> {
+            self.writes.fetch_add(1, Ordering::SeqCst);
+            Err(Error::other("observation must not write"))
+        }
+    }
+
+    #[tokio::test]
+    async fn locked_observation_preserves_options_and_typed_errors() {
+        let object = "config/observation.bin";
+        let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
+        let wrapper = rustfs_lock::NamespaceLockWrapper::new(
+            rustfs_lock::NamespaceLock::with_local_manager("locked-observation-options".to_owned(), manager.clone()),
+            rustfs_lock::ObjectKey::new(RUSTFS_META_BUCKET, object),
+            "locked-observation-options".to_owned(),
+        );
+        let guard = crate::object_api::WriteCommitGuard::acquire(&wrapper, std::time::Duration::from_secs(5))
+            .await
+            .expect("matching real owner");
+        let store = Arc::new(ObservationStore::new(b"raw".to_vec(), 3));
+        let result = super::read_config_observation_no_lock(store.clone(), object, 3, &guard).await;
+        assert_eq!(result.is_ok(), true, "I06L_READ_FLAGS: {result:?}");
+        assert_eq!(
+            *store.options.lock().expect("options") == Some((true, true, false)),
+            true,
+            "I06L_READ_FLAGS"
+        );
+        assert_eq!(result.expect("raw observation").0, b"raw");
+        assert_eq!(store.writes.load(Ordering::SeqCst), 0);
+        for (bucket, key) in [(RUSTFS_META_BUCKET, "config/other.bin"), ("wrong-bucket", object)] {
+            let wrapper = rustfs_lock::NamespaceLockWrapper::new(
+                rustfs_lock::NamespaceLock::with_local_manager("locked-observation-target".to_owned(), manager.clone()),
+                rustfs_lock::ObjectKey::new(bucket, key),
+                "locked-observation-target".to_owned(),
+            );
+            let wrong = crate::object_api::WriteCommitGuard::acquire(&wrapper, std::time::Duration::from_secs(5))
+                .await
+                .expect("wrong-target real owner");
+            let store = Arc::new(ObservationStore::new(Vec::new(), 0));
+            let result = super::read_config_observation_no_lock(store.clone(), object, 8, &wrong).await;
+            assert_eq!(matches!(result, Err(Error::InvalidArgument(..))), true, "I06L_TARGET_REJECTED");
+            assert_eq!(store.gets.load(Ordering::SeqCst) == 0, true, "I06L_TARGET_REJECTED");
+        }
+        for error in [
+            Error::FileNotFound,
+            Error::ObjectNotFound(RUSTFS_META_BUCKET.to_owned(), object.to_owned()),
+            Error::ErasureReadQuorum,
+        ] {
+            let expected = if error == Error::FileNotFound || matches!(error, Error::ObjectNotFound(..)) {
+                Error::ConfigNotFound
+            } else {
+                error.clone()
+            };
+            let mut fixture = ObservationStore::new(Vec::new(), 0);
+            fixture.error = Some(error);
+            let store = Arc::new(fixture);
+            assert_eq!(
+                super::read_config_observation_no_lock(store.clone(), object, 8, &guard)
+                    .await
+                    .expect_err("typed error"),
+                expected
+            );
+            assert_eq!(store.writes.load(Ordering::SeqCst), 0);
+        }
+        let mut fixture = ObservationStore::new(Vec::new(), 0);
+        fixture.stream_error = true;
+        let store = Arc::new(fixture);
+        let err = super::read_config_observation_no_lock(store.clone(), object, 8, &guard)
+            .await
+            .expect_err("stream failure");
+        assert_eq!(err.to_string().contains("observation stream failed"), true, "I06L_TYPED_STREAM_ERROR");
+        assert_eq!(store.writes.load(Ordering::SeqCst), 0);
+        let store = Arc::new(ObservationStore::new(vec![7; 64], 0));
+        let result = super::read_config_observation_no_lock(store.clone(), object, 8, &guard).await;
+        assert_eq!(matches!(result, Err(Error::CorruptedFormat)), true, "I06L_STREAM_BOUND");
+        assert_eq!(store.consumed.load(Ordering::SeqCst) == 9, true, "I06L_STREAM_BOUND");
+        assert_eq!(store.writes.load(Ordering::SeqCst), 0);
+    }
+
+    #[tokio::test]
+    async fn locked_observation_rejects_actual_owner_loss_before_and_after_read() {
+        let object = "config/observation.bin";
+        for after_read in [false, true] {
+            let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
+            let client = Arc::new(RefreshControlledClient::new(manager));
+            let wrapper = rustfs_lock::NamespaceLockWrapper::new(
+                rustfs_lock::NamespaceLock::new("locked-observation-loss".to_owned(), client.clone()),
+                rustfs_lock::ObjectKey::new(RUSTFS_META_BUCKET, object),
+                "locked-observation-loss".to_owned(),
+            );
+            let guard = crate::object_api::WriteCommitGuard::acquire(&wrapper, std::time::Duration::from_secs(5))
+                .await
+                .expect("real distributed owner");
+            let signal = guard.lock_lost_signal().expect("acquired distributed owner signal");
+            let store = Arc::new(ObservationStore::new(b"raw".to_vec(), 3));
+            if after_read {
+                let (ready_tx, ready_rx) = tokio::sync::oneshot::channel();
+                let (release_tx, release_rx) = tokio::sync::oneshot::channel();
+                *store.eof_gate.lock().expect("EOF gate") = Some((Some(ready_tx), release_rx));
+                let mut read = Box::pin(super::read_config_observation_no_lock(store.clone(), object, 3, &guard));
+                tokio::time::timeout(std::time::Duration::from_secs(60), async {
+                    tokio::select! {
+                        ready = ready_rx => ready.expect("read reached EOF"),
+                        result = &mut read => assert_eq!(false, true, "I06L_LOSS_AFTER_READ: returned before EOF gate: {result:?}"),
+                    }
+                }).await.expect("EOF watchdog");
+                client.reject_refreshes();
+                tokio::time::timeout(std::time::Duration::from_secs(60), signal.notified())
+                    .await
+                    .expect("actual owner loss watchdog");
+                assert_eq!(guard.is_lock_lost(), true, "I06L_LOSS_AFTER_READ");
+                release_tx.send(()).expect("release EOF");
+                let result = tokio::time::timeout(std::time::Duration::from_secs(60), read)
+                    .await
+                    .expect("read watchdog");
+                assert_eq!(
+                    matches!(
+                        result,
+                        Err(Error::NamespaceLockQuorumUnavailable {
+                            mode: "locked_config_observation",
+                            ..
+                        })
+                    ),
+                    true,
+                    "I06L_LOSS_AFTER_READ"
+                );
+                assert_eq!(store.gets.load(Ordering::SeqCst), 1);
+            } else {
+                client.reject_refreshes();
+                tokio::time::timeout(std::time::Duration::from_secs(60), signal.notified())
+                    .await
+                    .expect("actual owner loss watchdog");
+                assert_eq!(guard.is_lock_lost(), true, "I06L_LOSS_BEFORE_GET");
+                let result = super::read_config_observation_no_lock(store.clone(), object, 3, &guard).await;
+                assert_eq!(
+                    matches!(
+                        result,
+                        Err(Error::NamespaceLockQuorumUnavailable {
+                            mode: "locked_config_observation",
+                            ..
+                        })
+                    ),
+                    true,
+                    "I06L_LOSS_BEFORE_GET"
+                );
+                assert_eq!(store.gets.load(Ordering::SeqCst) == 0, true, "I06L_LOSS_BEFORE_GET");
+            }
+            assert_eq!(store.writes.load(Ordering::SeqCst), 0);
+        }
+    }
+
+    #[tokio::test]
+    async fn observation_preserves_options_and_errors() {
+        for body in [Vec::new(), b"raw config".to_vec()] {
+            let store = Arc::new(ObservationStore::new(body.clone(), i64::try_from(body.len()).expect("fixture size")));
+            let result = super::read_config_observation(store.clone(), "config/observation.bin", 32).await;
+            assert!(result.is_ok(), "observation must preserve empty and nonempty objects: {result:?}");
+            let (observed, info) = result.expect("successful observation");
+            assert_eq!(observed, body);
+            assert_eq!(info.etag.as_deref(), Some("observation-etag"));
+            assert_eq!(
+                *store.options.lock().expect("options"),
+                Some((true, false, false)),
+                "I05_REPAIR_SUPPRESSED"
+            );
+            assert_eq!(store.writes.load(Ordering::SeqCst), 0);
+        }
+        for (source, expected) in [
+            (Error::FileNotFound, Error::ConfigNotFound),
+            (
+                Error::ObjectNotFound(RUSTFS_META_BUCKET.to_string(), "config/observation.bin".to_string()),
+                Error::ConfigNotFound,
+            ),
+            (Error::ErasureReadQuorum, Error::ErasureReadQuorum),
+        ] {
+            let mut fixture = ObservationStore::new(Vec::new(), 0);
+            fixture.error = Some(source);
+            let store = Arc::new(fixture);
+            let result = super::read_config_observation(store.clone(), "config/observation.bin", 32).await;
+            assert_eq!(result.expect_err("read error must propagate"), expected);
+            assert_eq!(store.writes.load(Ordering::SeqCst), 0);
+        }
+        let mut fixture = ObservationStore::new(Vec::new(), 0);
+        fixture.stream_error = true;
+        let store = Arc::new(fixture);
+        let result = super::read_config_observation(store.clone(), "config/observation.bin", 32).await;
+        let error = result.expect_err("stream error must propagate");
+        assert!(error.to_string().contains("observation stream failed"));
+        assert_eq!(store.writes.load(Ordering::SeqCst), 0);
+    }
+
+    #[tokio::test]
+    async fn observation_enforces_byte_limit() {
+        let store = Arc::new(ObservationStore::new(vec![7; 64], 0));
+        let result = super::read_config_observation(store.clone(), "config/observation.bin", 8).await;
+        assert_eq!(result.is_err(), true, "I05_BOUND_ENFORCED");
+        assert_eq!(
+            result.expect_err("bounded observation error"),
+            Error::CorruptedFormat,
+            "I05_BOUND_ENFORCED"
+        );
+        assert_eq!(store.consumed.load(Ordering::SeqCst), 9, "I05_BOUND_ENFORCED: drain exactly limit + 1");
+        let store = Arc::new(ObservationStore::new(vec![7; 8], 8));
+        let result = super::read_config_observation(store.clone(), "config/observation.bin", 8).await;
+        assert!(result.is_ok(), "exact byte limit must succeed: {result:?}");
+        assert_eq!(result.expect("exact limit").0, vec![7; 8]);
+        assert_eq!(store.consumed.load(Ordering::SeqCst), 8);
+    }
+
    #[tokio::test]
    async fn read_config_preserve_empty_distinguishes_empty_object() {
        let store = Arc::new(RecoveryMockStore::new(RecoveryReadState::Blob(Vec::new()), None));
@@ -5431,6 +5826,60 @@ mod tests {
        assert!(data.is_empty());
    }

+    #[tokio::test]
+    async fn read_config_preserve_empty_metadata_keeps_empty_and_options() {
+        let store = Arc::new(RecoveryMockStore::new(RecoveryReadState::Blob(Vec::new()), None));
+        for no_lock in [false, true] {
+            let opts = ObjectOptions {
+                no_lock,
+                ..Default::default()
+            };
+            let (data, info) = read_config_preserve_empty_with_metadata_opts(store.clone(), "config/empty.json", &opts)
+                .await
+                .expect("an existing empty config must retain its payload and metadata");
+            assert!(data.is_empty());
+            assert_eq!(info.size, 0);
+            assert_eq!(info.actual_size, 0);
+            assert_eq!(info.etag.as_deref(), Some("config-1"));
+            assert_eq!(info.data_dir, Some(uuid::Uuid::from_u128(1)));
+            assert_eq!(store.last_get_no_lock.load(Ordering::SeqCst), no_lock);
+        }
+        let err = read_config(store, "config/empty.json")
+            .await
+            .expect_err("the existing reader must still treat empty configs as missing");
+        assert!(matches!(err, Error::ConfigNotFound));
+    }
+
+    #[tokio::test]
+    async fn read_config_preserve_empty_metadata_keeps_payload_and_metadata() {
+        let payload = b"original config bytes".to_vec();
+        let store = Arc::new(RecoveryMockStore::new(RecoveryReadState::Blob(payload.clone()), None));
+        let (data, info) =
+            read_config_preserve_empty_with_metadata_opts(store, "config/nonempty.json", &ObjectOptions::default())
+                .await
+                .expect("a nonempty config must retain its payload and metadata");
+        assert_eq!(data, payload);
+        let size = i64::try_from(payload.len()).expect("test payload size fits i64");
+        assert_eq!(info.size, size);
+        assert_eq!(info.actual_size, size);
+        assert_eq!(info.etag.as_deref(), Some("config-1"));
+        assert_eq!(info.data_dir, Some(uuid::Uuid::from_u128(1)));
+    }
+
+    #[tokio::test]
+    async fn read_config_preserve_empty_metadata_preserves_read_errors() {
+        let missing = Arc::new(RecoveryMockStore::new(RecoveryReadState::Missing, None));
+        let err = read_config_preserve_empty_with_metadata_opts(missing, "config/missing.json", &ObjectOptions::default())
+            .await
+            .expect_err("physical absence must remain ConfigNotFound");
+        assert!(matches!(err, Error::ConfigNotFound));
+        let quorum = Arc::new(RecoveryMockStore::new(RecoveryReadState::QuorumError, None));
+        let err = read_config_preserve_empty_with_metadata_opts(quorum, "config/quorum.json", &ObjectOptions::default())
+            .await
+            .expect_err("a quorum failure must retain its typed read error");
+        assert!(matches!(err, Error::ErasureReadQuorum));
+    }
+
    #[async_trait::async_trait]
    impl StorageAdminApi for RecoveryMockStore {
        type BackendInfo = rustfs_madmin::BackendInfo;
@@ -738,7 +738,12 @@ where
        .filter(|etag| !etag.is_empty())
        .map(str::to_owned)
        .ok_or_else(|| Error::other("data usage snapshot has no ETag"))?;
-    let mut data_usage_info = parse_usage_snapshot(&reader.read_all().await?)?;
+    let mut data_usage_info = parse_usage_snapshot(
+        &reader
+            .read_all()
+            .await
+            .map_err(|err| map_data_usage_metadata_read_error(err, object))?,
+    )?;
    populate_backward_compatible_usage_maps(&mut data_usage_info);
    validate_complete_usage_snapshot(&mut data_usage_info);
    Ok(Some((data_usage_info, revision)))
@@ -2997,6 +3002,61 @@ mod tests {
        }
    }

+    #[derive(Debug)]
+    struct UsageSnapshotReadFailureStore;
+
+    struct UsageSnapshotFailureReader;
+
+    impl tokio::io::AsyncRead for UsageSnapshotFailureReader {
+        fn poll_read(
+            self: std::pin::Pin<&mut Self>,
+            _cx: &mut std::task::Context<'_>,
+            _buf: &mut tokio::io::ReadBuf<'_>,
+        ) -> std::task::Poll<std::io::Result<()>> {
+            std::task::Poll::Ready(Err(std::io::Error::other(Error::VolumeNotFound)))
+        }
+    }
+
+    #[async_trait::async_trait]
+    impl crate::storage_api_contracts::object::ObjectIO for UsageSnapshotReadFailureStore {
+        type Error = Error;
+        type RangeSpec = crate::storage_api_contracts::range::HTTPRangeSpec;
+        type HeaderMap = http::HeaderMap;
+        type ObjectOptions = ObjectOptions;
+        type ObjectInfo = ObjectInfo;
+        type GetObjectReader = crate::object_api::GetObjectReader;
+        type PutObjectReader = PutObjReader;
+
+        async fn get_object_reader(
+            &self,
+            _bucket: &str,
+            _object: &str,
+            _range: Option<Self::RangeSpec>,
+            _headers: Self::HeaderMap,
+            _opts: &Self::ObjectOptions,
+        ) -> Result<Self::GetObjectReader, Self::Error> {
+            Ok(crate::object_api::GetObjectReader {
+                stream: Box::new(UsageSnapshotFailureReader),
+                object_info: ObjectInfo {
+                    etag: Some("usage-revision".to_string()),
+                    ..Default::default()
+                },
+                buffered_body: None,
+                body_source: Default::default(),
+            })
+        }
+
+        async fn put_object(
+            &self,
+            _bucket: &str,
+            _object: &str,
+            _data: &mut Self::PutObjectReader,
+            _opts: &Self::ObjectOptions,
+        ) -> Result<Self::ObjectInfo, Self::Error> {
+            unimplemented!("the read-failure fixture does not write")
+        }
+    }
+
    /// Minimal ObjectIO backing `load_data_usage_cache` tests: records the keys
    /// read and fails the first N reads with a transient (non-absence) error.
    #[derive(Debug, Default)]
@@ -3161,6 +3221,20 @@ mod tests {
        }
    }

+    #[tokio::test]
+    async fn data_usage_removal_maps_missing_system_volume_during_snapshot_stream_read() {
+        let store = UsageSnapshotReadFailureStore;
+
+        let error = load_data_usage_for_bucket_removal(&store, "bucket-metadata/.usage.json")
+            .await
+            .expect_err("a system volume disappearing during body read must remain a retryable storage error");
+
+        assert_eq!(
+            error,
+            Error::InsufficientReadQuorum(RUSTFS_META_BUCKET.to_string(), "bucket-metadata/.usage.json".to_string())
+        );
+    }
+
    #[tokio::test]
    async fn load_data_usage_cache_retries_a_transient_failure() {
        let name = "usage-cache";
@@ -12,6 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

+use rustfs_io_metrics::record_object_lock_acquire_failure;
 use rustfs_lock::{LockError, NamespaceLockGuard, NamespaceLockWrapper};
 use std::time::{Duration, Instant};
 use tracing::Span;
@@ -21,6 +22,26 @@ const EVENT_OBJECT_NAMESPACE_LOCK: &str = "object_namespace_lock";
 const LOG_COMPONENT_ECSTORE: &str = "ecstore";
 const LOG_SUBSYSTEM_OBJECT_LOCK: &str = "object_lock";

+pub(crate) fn record_lock_acquire_failure(op: &'static str, bucket: &str, object: &str, mode: &'static str, error: &LockError) {
+    let resource_class = if bucket == crate::disk::RUSTFS_META_MULTIPART_BUCKET {
+        "multipart_metadata"
+    } else if bucket == crate::disk::RUSTFS_META_BUCKET || bucket == ".minio.sys" {
+        "system_metadata"
+    } else if object.is_empty() || object == bucket {
+        "bucket_namespace"
+    } else {
+        "object_namespace"
+    };
+    let reason = match error {
+        LockError::Timeout { .. } => "timeout",
+        LockError::AlreadyLocked { .. } => "contention",
+        LockError::QuorumNotReached { .. } | LockError::InsufficientNodes { .. } => "quorum",
+        LockError::Network { .. } | LockError::Internal { .. } => "transport",
+        _ => "other",
+    };
+    record_object_lock_acquire_failure(op, mode, resource_class, reason);
+}
+
 /// Diagnostic lifetime only; it never owns or changes a namespace lock.
 /// Keep it alongside the real guard when ownership moves to a commit tail.
 #[derive(Default)]
@@ -12,6 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

+use crate::disk::io_admission::{
+    AdmissionReader, BoxedChunkReader, DiskIoAdmission, DiskIoClass, DiskIoContext, current_disk_io_context,
+    with_disk_io_context, with_disk_io_permit,
+};
 use crate::disk::{
    CheckPartsResp, DataDirDeleteStatus, DeleteOptions, DiskAPI, DiskError, DiskInfo, DiskInfoOptions, DiskLocation, Endpoint,
    Error, FileInfoVersions, MmapCopyStageMetrics, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp, Result,
@@ -35,13 +39,15 @@ use std::sync::atomic::AtomicBool;
 use std::{
    collections::HashMap,
    path::PathBuf,
+    pin::Pin,
    sync::{
        Arc, LazyLock, RwLock as StdRwLock,
        atomic::{AtomicI64, AtomicU32, AtomicU64, Ordering},
    },
+    task::{Context, Poll},
    time::{Duration, Instant, SystemTime, UNIX_EPOCH},
 };
-use tokio::{sync::RwLock, time};
+use tokio::{io::AsyncWrite, sync::RwLock, time};
 use tokio_util::sync::CancellationToken;
 use tracing::{info, warn};
 use uuid::Uuid;
@@ -54,6 +60,121 @@ const LOG_SUBSYSTEM_DISK: &str = "disk";
 const EVENT_DISK_HEALTH_CHECK_FAILED: &str = "disk_health_check_failed";
 const EVENT_DISK_RECOVERY_PROBE_STATE: &str = "disk_recovery_probe_state";
 const EVENT_DISK_TIMEOUT_POLICY_FALLBACK: &str = "disk_timeout_policy_fallback";
+const EVENT_DISK_OPERATION_SLOW: &str = "disk_operation_slow";
+const SLOW_DISK_OPERATION_THRESHOLD: Duration = Duration::from_secs(5);
+
+fn warn_slow_disk_operation(endpoint: &Endpoint, op: &'static str, elapsed: Duration) {
+    warn!(
+        event = EVENT_DISK_OPERATION_SLOW,
+        component = LOG_COMPONENT_ECSTORE,
+        subsystem = LOG_SUBSYSTEM_DISK,
+        state = "pending",
+        endpoint = %endpoint,
+        op,
+        elapsed_ms = elapsed.as_millis(),
+        threshold_ms = SLOW_DISK_OPERATION_THRESHOLD.as_millis(),
+        "Disk operation is still pending"
+    );
+}
+
+async fn observe_disk_operation<F: std::future::Future>(
+    endpoint: impl FnOnce() -> Endpoint,
+    op: &'static str,
+    operation: F,
+) -> F::Output {
+    if !tracing::enabled!(tracing::Level::WARN) {
+        return operation.await;
+    }
+    let started = time::Instant::now();
+    tokio::pin!(operation);
+    tokio::select! {
+        biased;
+        result = &mut operation => result,
+        _ = time::sleep(SLOW_DISK_OPERATION_THRESHOLD) => {
+            warn_slow_disk_operation(&endpoint(), op, started.elapsed());
+            operation.await
+        }
+    }
+}
+
+struct SlowDiskWriter {
+    inner: crate::disk::FileWriter,
+    endpoint: Endpoint,
+    pending: Option<(&'static str, time::Instant)>,
+    timer: Option<Pin<Box<time::Sleep>>>,
+    warned: bool,
+}
+
+impl SlowDiskWriter {
+    fn new(inner: crate::disk::FileWriter, endpoint: Endpoint) -> Self {
+        Self {
+            inner,
+            endpoint,
+            pending: None,
+            timer: None,
+            warned: false,
+        }
+    }
+
+    fn observe<T>(&mut self, cx: &mut Context<'_>, op: &'static str, result: Poll<T>) -> Poll<T> {
+        if result.is_ready() {
+            self.pending = None;
+            self.warned = false;
+        } else if tracing::enabled!(tracing::Level::WARN) {
+            if self.pending.is_none_or(|(pending_op, _)| pending_op != op) {
+                let started = time::Instant::now();
+                self.pending = Some((op, started));
+                self.warned = false;
+                let deadline = started + SLOW_DISK_OPERATION_THRESHOLD;
+                // Reuse one timer per stream; ready writes need no timer allocation.
+                if let Some(timer) = &mut self.timer {
+                    timer.as_mut().reset(deadline);
+                } else {
+                    self.timer = Some(Box::pin(time::sleep_until(deadline)));
+                }
+            }
+            if !self.warned
+                && let Some(timer) = &mut self.timer
+                && std::future::Future::poll(timer.as_mut(), cx).is_ready()
+                && let Some((_, started)) = self.pending
+            {
+                warn_slow_disk_operation(&self.endpoint, op, started.elapsed());
+                self.warned = true;
+            }
+        }
+        result
+    }
+}
+
+impl AsyncWrite for SlowDiskWriter {
+    fn poll_write(mut self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &[u8]) -> Poll<std::io::Result<usize>> {
+        let result = Pin::new(&mut self.inner).poll_write(cx, buf);
+        self.observe(cx, "write", result)
+    }
+
+    fn poll_write_vectored(
+        mut self: Pin<&mut Self>,
+        cx: &mut Context<'_>,
+        bufs: &[std::io::IoSlice<'_>],
+    ) -> Poll<std::io::Result<usize>> {
+        let result = Pin::new(&mut self.inner).poll_write_vectored(cx, bufs);
+        self.observe(cx, "write_vectored", result)
+    }
+
+    fn is_write_vectored(&self) -> bool {
+        self.inner.is_write_vectored()
+    }
+
+    fn poll_flush(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<std::io::Result<()>> {
+        let result = Pin::new(&mut self.inner).poll_flush(cx);
+        self.observe(cx, "flush", result)
+    }
+
+    fn poll_shutdown(mut self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<std::io::Result<()>> {
+        let result = Pin::new(&mut self.inner).poll_shutdown(cx);
+        self.observe(cx, "shutdown", result)
+    }
+}

 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 enum TimeoutHealthAction {
@@ -813,6 +934,7 @@ impl Drop for DiskMetricWaitingGuard<'_> {
 pub(crate) struct ReconnectDiskHealthState {
    pub(crate) health: Arc<DiskHealthTracker>,
    pub(crate) metrics: Arc<DiskHealthMetricEpoch>,
+    pub(crate) io_admission: Arc<DiskIoAdmission>,
 }

 #[derive(Debug)]
@@ -1329,6 +1451,8 @@ pub struct LocalDiskWrapper {
    health: Arc<DiskHealthTracker>,
    /// Internal metrics epoch preserved across local disk reconnects.
    metrics: Arc<DiskHealthMetricEpoch>,
+    /// Per-drive foreground/background admission shared across wrapper reconnects.
+    io_admission: Arc<DiskIoAdmission>,
    /// Whether health checking is enabled
    health_check: bool,
    /// Cancellation token for monitoring tasks
@@ -1342,11 +1466,13 @@ pub struct LocalDiskWrapper {
 impl LocalDiskWrapper {
    /// Create a new LocalDiskWrapper
    pub fn new(disk: Arc<LocalDisk>, health_check: bool) -> Self {
+        let admission = DiskIoAdmission::new(disk.endpoint().to_string());
        Self::new_with_health_and_metrics(
            disk,
            health_check,
            Arc::new(DiskHealthTracker::new()),
            Arc::new(DiskHealthMetricEpoch::default()),
+            admission,
        )
    }

@@ -1386,11 +1512,13 @@ impl LocalDiskWrapper {
        health_check: bool,
        reconnect: Option<ReconnectDiskHealthState>,
    ) -> Self {
+        let endpoint = disk.endpoint().to_string();
        let reconnect = reconnect.unwrap_or_else(|| ReconnectDiskHealthState {
            health: Arc::new(DiskHealthTracker::new()),
            metrics: Arc::new(DiskHealthMetricEpoch::default()),
+            io_admission: DiskIoAdmission::new(endpoint),
        });
-        Self::new_with_health_and_metrics(disk, health_check, reconnect.health, reconnect.metrics)
+        Self::new_with_health_and_metrics(disk, health_check, reconnect.health, reconnect.metrics, reconnect.io_admission)
    }

    fn new_with_health_and_metrics(
@@ -1398,6 +1526,7 @@ impl LocalDiskWrapper {
        health_check: bool,
        health: Arc<DiskHealthTracker>,
        metrics: Arc<DiskHealthMetricEpoch>,
+        io_admission: Arc<DiskIoAdmission>,
    ) -> Self {
        // Check environment variable for health check override.
        // Only enable if both param and env are true.
@@ -1408,6 +1537,7 @@ impl LocalDiskWrapper {
            disk,
            health,
            metrics,
+            io_admission,
            health_check: health_check && env_health_check,
            cancel_token: CancellationToken::new(),
            disk_id: Arc::new(RwLock::new(None)),
@@ -1421,6 +1551,7 @@ impl LocalDiskWrapper {
        ReconnectDiskHealthState {
            health: Arc::new(self.health.metric_epoch_for_reconnect()),
            metrics: self.metrics.clone(),
+            io_admission: self.io_admission.clone(),
        }
    }

@@ -1428,6 +1559,58 @@ impl LocalDiskWrapper {
        self.disk.clone()
    }

+    #[cfg(test)]
+    pub(crate) fn background_io_admissions_for_tests(&self) -> usize {
+        self.io_admission.background_admissions_for_tests()
+    }
+
+    #[cfg(test)]
+    pub(crate) fn background_io_active_for_tests(&self) -> bool {
+        self.io_admission.background_active_for_tests()
+    }
+
+    fn disk_io_context(&self) -> DiskIoContext {
+        current_disk_io_context().unwrap_or_else(|| DiskIoContext {
+            class: DiskIoClass::Foreground,
+            cancellation: CancellationToken::new(),
+            admission: None,
+            permit: None,
+        })
+    }
+
+    fn disk_walk_io_context(&self) -> DiskIoContext {
+        let mut context = self.disk_io_context();
+        context.admission = Some(self.io_admission.clone());
+        context.permit = None;
+        context
+    }
+
+    async fn acquire_disk_io_permit(&self) -> Result<Option<crate::disk::io_admission::DiskIoPermit>> {
+        if let Some(context) = current_disk_io_context() {
+            if let Some(permit) = context.permit
+                && permit.is_for(&self.io_admission)
+            {
+                return Ok(Some(permit));
+            }
+            if context
+                .admission
+                .as_ref()
+                .is_some_and(|admission| Arc::ptr_eq(admission, &self.io_admission))
+            {
+                return Ok(None);
+            }
+            if context.class == DiskIoClass::Background {
+                return self
+                    .io_admission
+                    .background(&context.cancellation)
+                    .await
+                    .map(Some)
+                    .map_err(DiskError::Io);
+            }
+        }
+        Ok(Some(self.io_admission.foreground()))
+    }
+
    pub fn get_object_path_if_local(&self, volume: &str, path: &str) -> crate::disk::error::Result<std::path::PathBuf> {
        self.disk.get_object_path(volume, path)
    }
@@ -1899,6 +2082,31 @@ impl LocalDiskWrapper {
    where
        F: FnOnce() -> Fut,
        Fut: std::future::Future<Output = Result<T>>,
+    {
+        self.track_disk_health_with_op_timeout_action_and_mutation_with_io_permit(
+            op,
+            move |io_permit| async move {
+                let _io_permit = io_permit;
+                operation().await
+            },
+            timeout_duration,
+            timeout_health_action,
+            mutation,
+        )
+        .await
+    }
+
+    async fn track_disk_health_with_op_timeout_action_and_mutation_with_io_permit<T, F, Fut>(
+        &self,
+        op: &'static str,
+        operation: F,
+        timeout_duration: Duration,
+        timeout_health_action: TimeoutHealthAction,
+        mutation: DiskMetricMutation,
+    ) -> Result<T>
+    where
+        F: FnOnce(Option<crate::disk::io_admission::DiskIoPermit>) -> Fut,
+        Fut: std::future::Future<Output = Result<T>>,
    {
        self.metrics.record_operation_call(op);
        // Check if disk is faulty
@@ -1921,14 +2129,17 @@ impl LocalDiskWrapper {
            return Err(err);
        }

+        let io_permit = self.acquire_disk_io_permit().await?;
+
        // Record operation start
        self.health.last_started.store(current_unix_nanos(), Ordering::Relaxed);
        let _waiting_guard = self.health.waiting_guard();
        let _metric_waiting_guard = self.metrics.waiting_guard();
        let started = Instant::now();
+        let operation = observe_disk_operation(|| self.endpoint(), op, operation(io_permit));

        if timeout_duration == Duration::ZERO {
-            let result = operation().await;
+            let result = operation.await;
            self.metrics.record_operation_latency(op, started.elapsed());
            self.record_result_error_metrics(&result);
            if result.is_ok() {
@@ -1938,7 +2149,7 @@ impl LocalDiskWrapper {
            return result;
        }
        // Execute the operation with timeout
-        let result = tokio::time::timeout(timeout_duration, operation()).await;
+        let result = tokio::time::timeout(timeout_duration, operation).await;

        match result {
            Ok(operation_result) => {
@@ -1989,11 +2200,12 @@ impl DiskAPI for LocalDiskWrapper {
    }

    async fn read_metadata(&self, volume: &str, path: &str) -> Result<Bytes> {
-        self.track_disk_health_with_op_and_timeout_action(
+        self.track_disk_health_with_op_timeout_action_and_mutation_with_io_permit(
            "read_metadata",
-            || async { self.disk.read_metadata(volume, path).await },
+            |io_permit| async { self.disk.read_metadata_with_permit(volume, path, io_permit).await },
            get_drive_metadata_timeout(),
            self.scanner_timeout_health_action(),
+            DiskMetricMutation::None,
        )
        .await
    }
@@ -2165,12 +2377,15 @@ impl DiskAPI for LocalDiskWrapper {
            opts.timeout_duration().unwrap_or_else(get_drive_walkdir_timeout)
        };

-        self.track_disk_health_with_op_and_timeout_action(
-            "walk_dir",
-            || async { self.disk.walk_dir(opts, wr).await },
-            timeout_duration,
-            // Listing/scanner backpressure should fail only the current walk, not poison drive health.
-            TimeoutHealthAction::IgnoreFailure,
+        with_disk_io_context(
+            self.disk_walk_io_context(),
+            self.track_disk_health_with_op_and_timeout_action(
+                "walk_dir",
+                || async { self.disk.walk_dir(opts, wr).await },
+                timeout_duration,
+                // Listing/scanner backpressure should fail only the current walk, not poison drive health.
+                TimeoutHealthAction::IgnoreFailure,
+            ),
        )
        .await
    }
@@ -2193,6 +2408,7 @@ impl DiskAPI for LocalDiskWrapper {
    }

    async fn delete_versions(&self, volume: &str, versions: Vec<FileInfoVersions>, opts: DeleteOptions) -> Vec<Option<Error>> {
+        let version_count = versions.len();
        self.metrics.record_operation_call("delete_versions");
        // Check if disk is faulty before proceeding
        if self.health.is_faulty() {
@@ -2206,9 +2422,14 @@ impl DiskAPI for LocalDiskWrapper {
            return vec![Some(e); versions.len()];
        }

+        let _io_permit = match self.acquire_disk_io_permit().await {
+            Ok(permit) => permit,
+            Err(error) => return vec![Some(error); version_count],
+        };
+
        // Record operation start
        self.health.last_started.store(current_unix_nanos(), Ordering::Relaxed);
-        self.health.increment_waiting();
+        let waiting_guard = self.health.waiting_guard();
        let metric_waiting_guard = self.metrics.waiting_guard();
        let started = Instant::now();

@@ -2217,7 +2438,7 @@ impl DiskAPI for LocalDiskWrapper {
        self.metrics.record_operation_latency("delete_versions", started.elapsed());
        self.record_batch_delete_error_metrics(&result);

-        self.health.decrement_waiting();
+        drop(waiting_guard);
        drop(metric_waiting_guard);
        let has_err = result.iter().any(|e| e.is_some());
        if !has_err {
@@ -2322,19 +2543,25 @@ impl DiskAPI for LocalDiskWrapper {
        version_id: &str,
        opts: &ReadOptions,
    ) -> Result<FileInfo> {
-        self.track_disk_health_with_op(
+        self.track_disk_health_with_op_timeout_action_and_mutation_with_io_permit(
            "read_version",
-            || async { self.disk.read_version(org_volume, volume, path, version_id, opts).await },
+            |io_permit| async move {
+                with_disk_io_permit(io_permit, self.disk.read_version(org_volume, volume, path, version_id, opts)).await
+            },
            get_max_timeout_duration(),
+            TimeoutHealthAction::MarkFailure,
+            DiskMetricMutation::None,
        )
        .await
    }

    async fn read_xl(&self, volume: &str, path: &str, read_data: bool) -> Result<RawFileInfo> {
-        self.track_disk_health_with_op(
+        self.track_disk_health_with_op_timeout_action_and_mutation_with_io_permit(
            "read_xl",
-            || async { self.disk.read_xl(volume, path, read_data).await },
+            |io_permit| async move { with_disk_io_permit(io_permit, self.disk.read_xl(volume, path, read_data)).await },
            get_max_timeout_duration(),
+            TimeoutHealthAction::MarkFailure,
+            DiskMetricMutation::None,
        )
        .await
    }
@@ -2362,21 +2589,25 @@ impl DiskAPI for LocalDiskWrapper {
    }

    async fn read_file(&self, volume: &str, path: &str) -> Result<crate::disk::FileReader> {
-        self.track_disk_health_with_op(
-            "read_file",
-            || async { self.disk.read_file(volume, path).await },
-            get_max_timeout_duration(),
-        )
-        .await
+        let reader = self
+            .track_disk_health_with_op(
+                "read_file",
+                || async { self.disk.read_file(volume, path).await },
+                get_max_timeout_duration(),
+            )
+            .await?;
+        Ok(Box::new(AdmissionReader::new(reader, self.io_admission.clone(), self.disk_io_context())))
    }

    async fn read_file_stream(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<crate::disk::FileReader> {
-        self.track_disk_health_with_op(
-            "read_file_stream",
-            || async { self.disk.read_file_stream(volume, path, offset, length).await },
-            get_max_timeout_duration(),
-        )
-        .await
+        let reader = self
+            .track_disk_health_with_op(
+                "read_file_stream",
+                || async { self.disk.read_file_stream(volume, path, offset, length).await },
+                get_max_timeout_duration(),
+            )
+            .await?;
+        Ok(Box::new(AdmissionReader::new(reader, self.io_admission.clone(), self.disk_io_context())))
    }

    async fn read_file_stream_chunks(
@@ -2386,12 +2617,20 @@ impl DiskAPI for LocalDiskWrapper {
        offset: usize,
        length: usize,
    ) -> Result<Option<rustfs_rio::ChunkReaderBox>> {
-        self.track_disk_health_with_op(
-            "read_file_stream_chunks",
-            || async { self.disk.read_file_stream_chunks(volume, path, offset, length).await },
-            get_max_timeout_duration(),
-        )
-        .await
+        let reader = self
+            .track_disk_health_with_op(
+                "read_file_stream_chunks",
+                || async { self.disk.read_file_stream_chunks(volume, path, offset, length).await },
+                get_max_timeout_duration(),
+            )
+            .await?;
+        Ok(reader.map(|reader| {
+            Box::new(AdmissionReader::new(
+                BoxedChunkReader(reader),
+                self.io_admission.clone(),
+                self.disk_io_context(),
+            )) as rustfs_rio::ChunkReaderBox
+        }))
    }

    async fn read_file_mmap_copy(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<bytes::Bytes> {
@@ -2424,18 +2663,22 @@ impl DiskAPI for LocalDiskWrapper {
    }

    async fn append_file(&self, volume: &str, path: &str) -> Result<crate::disk::FileWriter> {
-        self.track_disk_health_with_op("append_file", || async { self.disk.append_file(volume, path).await }, Duration::ZERO)
-            .await
+        let writer = self
+            .track_disk_health_with_op("append_file", || async { self.disk.append_file(volume, path).await }, Duration::ZERO)
+            .await?;
+        Ok(Box::new(SlowDiskWriter::new(writer, self.endpoint())))
    }

    async fn create_file(&self, origvolume: &str, volume: &str, path: &str, file_size: i64) -> Result<crate::disk::FileWriter> {
-        self.track_disk_health_mutation(
-            "create_file",
-            DiskMetricMutation::Write,
-            || async { self.disk.create_file(origvolume, volume, path, file_size).await },
-            Duration::ZERO,
-        )
-        .await
+        let writer = self
+            .track_disk_health_mutation(
+                "create_file",
+                DiskMetricMutation::Write,
+                || async { self.disk.create_file(origvolume, volume, path, file_size).await },
+                Duration::ZERO,
+            )
+            .await?;
+        Ok(Box::new(SlowDiskWriter::new(writer, self.endpoint())))
    }

    async fn rename_file(&self, src_volume: &str, src_path: &str, dst_volume: &str, dst_path: &str) -> Result<()> {
@@ -2524,8 +2767,14 @@ impl DiskAPI for LocalDiskWrapper {
    }

    async fn read_parts(&self, bucket: &str, paths: &[String]) -> Result<Vec<ObjectPartInfo>> {
-        self.track_disk_health_with_op("read_parts", || async { self.disk.read_parts(bucket, paths).await }, Duration::ZERO)
-            .await
+        self.track_disk_health_with_op_timeout_action_and_mutation_with_io_permit(
+            "read_parts",
+            |io_permit| async move { with_disk_io_permit(io_permit, self.disk.read_parts(bucket, paths)).await },
+            Duration::ZERO,
+            TimeoutHealthAction::MarkFailure,
+            DiskMetricMutation::None,
+        )
+        .await
    }

    async fn read_multiple(&self, req: ReadMultipleReq) -> Result<Vec<ReadMultipleResp>> {
@@ -2582,7 +2831,268 @@ mod tests {
        pin::Pin,
        task::{Context, Poll},
    };
-    use tokio::io::AsyncWrite;
+    use tokio::io::AsyncWriteExt;
+    use tracing_subscriber::fmt::MakeWriter;
+
+    #[derive(Clone, Default)]
+    struct SlowOperationLogs(Arc<std::sync::Mutex<Vec<u8>>>);
+
+    impl SlowOperationLogs {
+        fn contents(&self) -> String {
+            String::from_utf8(self.0.lock().expect("log buffer should not be poisoned").clone()).expect("logs should be UTF-8")
+        }
+    }
+
+    impl io::Write for SlowOperationLogs {
+        fn write(&mut self, buf: &[u8]) -> io::Result<usize> {
+            self.0
+                .lock()
+                .expect("log buffer should not be poisoned")
+                .extend_from_slice(buf);
+            Ok(buf.len())
+        }
+
+        fn flush(&mut self) -> io::Result<()> {
+            Ok(())
+        }
+    }
+
+    impl<'a> MakeWriter<'a> for SlowOperationLogs {
+        type Writer = Self;
+
+        fn make_writer(&'a self) -> Self::Writer {
+            self.clone()
+        }
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn slow_disk_operation_warns_once_while_pending_and_preserves_result() {
+        let dir = tempfile::tempdir().expect("temp dir should be created");
+        let endpoint = Endpoint::try_from(dir.path().to_str().expect("path should be UTF-8")).expect("endpoint should parse");
+        let disk = Arc::new(LocalDisk::new(&endpoint, false).await.expect("local disk should be created"));
+        let wrapper = LocalDiskWrapper::new(disk, false);
+        let logs = SlowOperationLogs::default();
+        let subscriber = tracing_subscriber::fmt()
+            .with_writer(logs.clone())
+            .with_ansi(false)
+            .without_time()
+            .finish();
+        let _subscriber = tracing::subscriber::set_default(subscriber);
+        let _interest = crate::test_tracing::pin_callsite_interest_for_test();
+        let (release, wait) = tokio::sync::oneshot::channel();
+        let operation = wrapper.track_disk_health_with_op(
+            "rename_data",
+            || async { wait.await.expect("operation should be released") },
+            Duration::ZERO,
+        );
+        tokio::pin!(operation);
+        assert!(futures::poll!(&mut operation).is_pending());
+        time::advance(SLOW_DISK_OPERATION_THRESHOLD - Duration::from_millis(1)).await;
+        assert!(futures::poll!(&mut operation).is_pending());
+        assert!(logs.contents().is_empty(), "normal disk latency should not warn");
+        time::advance(Duration::from_millis(1)).await;
+        assert!(futures::poll!(&mut operation).is_pending());
+        let warning = logs.contents();
+        assert!(warning.contains("WARN"));
+        assert!(warning.contains("event=\"disk_operation_slow\""));
+        assert!(warning.contains("state=\"pending\""));
+        assert!(warning.contains(&format!("endpoint={endpoint}")));
+        assert!(warning.contains("op=\"rename_data\""));
+        assert!(warning.contains("elapsed_ms=5000"));
+        time::advance(Duration::from_secs(120)).await;
+        assert!(futures::poll!(&mut operation).is_pending(), "warning must not cancel the disk operation");
+        assert_eq!(logs.contents(), warning, "a pending operation should warn only once");
+        release
+            .send(Err::<(), _>(DiskError::FileAccessDenied))
+            .expect("operation should still be alive");
+        assert_eq!(operation.await, Err(DiskError::FileAccessDenied));
+        assert_eq!(wrapper.runtime_state(), RuntimeDriveHealthState::Online);
+        assert_eq!(wrapper.metrics_snapshot().total_waiting, 0);
+
+        let timed = wrapper.track_disk_health_with_op(
+            "write_all",
+            std::future::pending::<Result<()>>,
+            SLOW_DISK_OPERATION_THRESHOLD + Duration::from_secs(1),
+        );
+        tokio::pin!(timed);
+        assert!(futures::poll!(&mut timed).is_pending());
+        time::advance(SLOW_DISK_OPERATION_THRESHOLD).await;
+        assert!(futures::poll!(&mut timed).is_pending());
+        assert_eq!(logs.contents().matches(EVENT_DISK_OPERATION_SLOW).count(), 2);
+        assert_eq!(wrapper.runtime_state(), RuntimeDriveHealthState::Online);
+        time::advance(Duration::from_secs(1)).await;
+        assert_eq!(
+            timed.await,
+            Err(DiskError::Timeout),
+            "existing timeout must still apply after the warning"
+        );
+        assert_eq!(wrapper.metrics_snapshot().total_errors_timeout, 1);
+        assert_eq!(wrapper.metrics_snapshot().total_waiting, 0);
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn slow_disk_operation_fast_path_is_silent_and_does_not_resolve_endpoint() {
+        let logs = SlowOperationLogs::default();
+        let subscriber = tracing_subscriber::fmt()
+            .with_writer(logs.clone())
+            .with_ansi(false)
+            .without_time()
+            .finish();
+        let _subscriber = tracing::subscriber::set_default(subscriber);
+        let _interest = crate::test_tracing::pin_callsite_interest_for_test();
+        for delay in [Duration::ZERO, SLOW_DISK_OPERATION_THRESHOLD] {
+            let result =
+                observe_disk_operation(|| panic!("completed operation should not clone its endpoint"), "write_all", async {
+                    if !delay.is_zero() {
+                        time::sleep(delay).await;
+                    }
+                    Ok::<_, DiskError>(17)
+                })
+                .await;
+            assert_eq!(result, Ok(17));
+        }
+        assert!(logs.contents().is_empty());
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn slow_disk_diagnostics_are_bypassed_when_warnings_are_disabled() {
+        let _subscriber = tracing::subscriber::set_default(tracing::subscriber::NoSubscriber::default());
+        let endpoint = Endpoint::try_from("/slow-disk").expect("endpoint should parse");
+        let mut writer = SlowDiskWriter::new(Box::new(PendingWriter), endpoint);
+        {
+            let write = writer.write(b"bytes");
+            tokio::pin!(write);
+            assert!(futures::poll!(&mut write).is_pending());
+        }
+        assert!(writer.timer.is_none(), "disabled warnings should not allocate a stream timer");
+        let result = observe_disk_operation(|| panic!("disabled warnings should not resolve endpoints"), "write_all", async {
+            time::sleep(SLOW_DISK_OPERATION_THRESHOLD * 2).await;
+            17
+        })
+        .await;
+        assert_eq!(result, 17);
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn slow_disk_writer_preserves_vectored_bytes_and_warns_once_while_blocked() {
+        let endpoint = Endpoint::try_from("/slow-disk").expect("endpoint should parse");
+        let logs = SlowOperationLogs::default();
+        let subscriber = tracing_subscriber::fmt()
+            .with_writer(logs.clone())
+            .with_ansi(false)
+            .without_time()
+            .finish();
+        let _subscriber = tracing::subscriber::set_default(subscriber);
+        let _interest = crate::test_tracing::pin_callsite_interest_for_test();
+        let (stream, mut receiver) = tokio::io::duplex(1);
+        let mut writer = SlowDiskWriter::new(Box::new(stream), endpoint.clone());
+        writer.write_all(b"a").await.expect("first byte should fit in the buffer");
+        assert!(writer.timer.is_none(), "ready writes should not allocate a timer");
+        assert!(logs.contents().is_empty());
+        let mut received = [0; 1];
+        {
+            let bufs = [io::IoSlice::new(b"b"), io::IoSlice::new(b"c")];
+            let write = writer.write_vectored(&bufs);
+            tokio::pin!(write);
+            assert!(futures::poll!(&mut write).is_pending());
+            time::advance(SLOW_DISK_OPERATION_THRESHOLD).await;
+            assert!(futures::poll!(&mut write).is_pending());
+            assert!(logs.contents().contains("op=\"write_vectored\""));
+            time::advance(SLOW_DISK_OPERATION_THRESHOLD).await;
+            assert!(futures::poll!(&mut write).is_pending());
+            assert_eq!(logs.contents().matches(EVENT_DISK_OPERATION_SLOW).count(), 1);
+            tokio::io::AsyncReadExt::read_exact(&mut receiver, &mut received)
+                .await
+                .expect("buffered byte should be readable");
+            assert_eq!(received, *b"a");
+            assert_eq!(write.await.expect("blocked write should finish when capacity returns"), 1);
+        }
+        tokio::io::AsyncReadExt::read_exact(&mut receiver, &mut received)
+            .await
+            .expect("written byte should be readable");
+        assert_eq!(received, *b"b");
+
+        drop(receiver);
+        let err = writer.write_all(b"d").await.expect_err("closed receiver should fail writes");
+        assert_eq!(err.kind(), io::ErrorKind::BrokenPipe);
+    }
+
+    #[tokio::test]
+    async fn slow_disk_writer_preserves_full_vectored_write_and_capability() {
+        let endpoint = Endpoint::try_from("/slow-disk").expect("endpoint should parse");
+        let (stream, mut receiver) = tokio::io::duplex(64);
+        let vectored = stream.is_write_vectored();
+        let mut writer = SlowDiskWriter::new(Box::new(stream), endpoint);
+        assert_eq!(writer.is_write_vectored(), vectored);
+        let bufs = [io::IoSlice::new(b"hash"), io::IoSlice::new(b"data")];
+        assert_eq!(writer.write_vectored(&bufs).await.expect("vectored write should succeed"), 8);
+        writer.shutdown().await.expect("shutdown should succeed");
+        let mut received = Vec::new();
+        tokio::io::AsyncReadExt::read_to_end(&mut receiver, &mut received)
+            .await
+            .expect("all bytes should be readable");
+        assert_eq!(received, b"hashdata");
+        assert!(writer.timer.is_none());
+    }
+
+    struct StalledWriter(Arc<AtomicBool>);
+
+    impl AsyncWrite for StalledWriter {
+        fn poll_write(self: Pin<&mut Self>, _cx: &mut Context<'_>, buf: &[u8]) -> Poll<io::Result<usize>> {
+            if self.0.load(Ordering::Relaxed) {
+                Poll::Ready(Ok(buf.len()))
+            } else {
+                Poll::Pending
+            }
+        }
+
+        fn poll_flush(self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+            if self.0.load(Ordering::Relaxed) {
+                Poll::Ready(Ok(()))
+            } else {
+                Poll::Pending
+            }
+        }
+
+        fn poll_shutdown(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+            self.poll_flush(cx)
+        }
+    }
+
+    #[tokio::test(start_paused = true)]
+    async fn slow_disk_writer_warns_for_write_flush_and_shutdown_and_rearms_after_progress() {
+        let endpoint = Endpoint::try_from("/slow-disk").expect("endpoint should parse");
+        let logs = SlowOperationLogs::default();
+        let subscriber = tracing_subscriber::fmt()
+            .with_writer(logs.clone())
+            .with_ansi(false)
+            .without_time()
+            .finish();
+        let _subscriber = tracing::subscriber::set_default(subscriber);
+        let _interest = crate::test_tracing::pin_callsite_interest_for_test();
+        let ready = Arc::new(AtomicBool::new(false));
+        let mut writer = SlowDiskWriter::new(Box::new(StalledWriter(Arc::clone(&ready))), endpoint);
+        for (index, op) in ["write", "flush", "shutdown"].into_iter().enumerate() {
+            ready.store(false, Ordering::Relaxed);
+            let operation = async {
+                match op {
+                    "write" => writer.write_all(b"bytes").await,
+                    "flush" => writer.flush().await,
+                    "shutdown" => writer.shutdown().await,
+                    _ => unreachable!("only write operations are tested"),
+                }
+            };
+            tokio::pin!(operation);
+            assert!(futures::poll!(&mut operation).is_pending());
+            time::advance(SLOW_DISK_OPERATION_THRESHOLD).await;
+            assert!(futures::poll!(&mut operation).is_pending());
+            assert!(logs.contents().contains(&format!("op=\"{op}\"")));
+            assert_eq!(logs.contents().matches(EVENT_DISK_OPERATION_SLOW).count(), index + 1);
+            ready.store(true, Ordering::Relaxed);
+            operation.await.expect("warning should not prevent operation completion");
+        }
+        assert!(writer.pending.is_none());
+    }

    struct DropProbe(Arc<std::sync::atomic::AtomicUsize>);

@@ -2592,6 +3102,84 @@ mod tests {
        }
    }

+    struct StalledMetacacheWriter {
+        entered: Arc<tokio::sync::Notify>,
+    }
+
+    impl AsyncWrite for StalledMetacacheWriter {
+        fn poll_write(self: Pin<&mut Self>, _cx: &mut Context<'_>, _buf: &[u8]) -> Poll<io::Result<usize>> {
+            self.entered.notify_one();
+            Poll::Pending
+        }
+
+        fn poll_flush(self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+            Poll::Ready(Ok(()))
+        }
+
+        fn poll_shutdown(self: Pin<&mut Self>, _cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+            Poll::Ready(Ok(()))
+        }
+    }
+
+    #[tokio::test]
+    async fn scanner_walk_releases_disk_admission_while_output_is_backpressured() {
+        let root = tempfile::tempdir().expect("disk root should be created");
+        let endpoint =
+            Endpoint::try_from(root.path().to_str().expect("disk root should be UTF-8")).expect("endpoint should parse");
+        let disk = Arc::new(LocalDisk::new(&endpoint, false).await.expect("local disk should open"));
+        let wrapper = Arc::new(LocalDiskWrapper::new(disk, false));
+        wrapper.make_volume("bucket").await.expect("bucket should be created");
+        let object_path = root.path().join("bucket/object");
+        tokio::fs::create_dir_all(&object_path)
+            .await
+            .expect("object directory should be created");
+        tokio::fs::write(object_path.join("xl.meta"), b"scanner metadata")
+            .await
+            .expect("object metadata should be created");
+
+        let entered = Arc::new(tokio::sync::Notify::new());
+        let writer = StalledMetacacheWriter {
+            entered: entered.clone(),
+        };
+        let walk_wrapper = wrapper.clone();
+        let walk = tokio::spawn(async move {
+            crate::disk::with_background_disk_io(CancellationToken::new(), async move {
+                let mut writer = writer;
+                walk_wrapper
+                    .walk_dir(
+                        WalkDirOptions {
+                            bucket: "bucket".to_string(),
+                            recursive: true,
+                            skip_total_timeout: true,
+                            ..Default::default()
+                        },
+                        &mut writer,
+                    )
+                    .await
+            })
+            .await
+        });
+
+        tokio::time::timeout(Duration::from_secs(2), entered.notified())
+            .await
+            .expect("walk should reach the blocked output writer");
+        assert!(
+            !wrapper.background_io_active_for_tests(),
+            "filesystem read permits must be released before metacache output backpressure"
+        );
+
+        let checkpoint_permit = tokio::time::timeout(
+            Duration::from_secs(1),
+            crate::disk::with_background_disk_io(CancellationToken::new(), wrapper.acquire_disk_io_permit()),
+        )
+        .await
+        .expect("checkpoint admission should progress while the scanner consumer is paused")
+        .expect("checkpoint should acquire the disk gate");
+        drop(checkpoint_permit);
+        walk.abort();
+        let _ = walk.await;
+    }
+
    #[tokio::test]
    async fn owned_mutation_keeps_publication_guard_after_waiter_cancellation() {
        let drops = Arc::new(std::sync::atomic::AtomicUsize::new(0));
@@ -2742,6 +3330,7 @@ mod tests {
        let reconnect = ReconnectDiskHealthState {
            health: Arc::new(health.metric_epoch_for_reconnect()),
            metrics: metrics.clone(),
+            io_admission: DiskIoAdmission::new("test-disk"),
        };
        metrics.record_operation_call("read_all");
        reconnect.metrics.record_timeout_error();
@@ -3023,6 +3612,40 @@ mod tests {
        assert_eq!(snapshot.total_deletes, 1);
    }

+    // A batch delete dropped while the disk call is in flight (for example by a
+    // cancelled caller) must not leave the drive looking busy forever.
+    #[tokio::test]
+    async fn delete_versions_releases_waiting_slot_when_dropped() {
+        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"));
+        let wrapper = LocalDiskWrapper::new(Arc::clone(&disk), false);
+        wrapper.make_volume("bucket").await.expect("volume should be created");
+        let object_path = disk.get_object_path("bucket", "object").expect("object path");
+        // Holding the object's metadata lease parks the disk call before it mutates anything.
+        let metadata_lease = crate::disk::os::acquire_metadata_mutation_lease(&object_path, None).await;
+
+        let mut delete = Box::pin(wrapper.delete_versions(
+            "bucket",
+            vec![FileInfoVersions {
+                name: "object".to_string(),
+                versions: vec![FileInfo::default()],
+                ..Default::default()
+            }],
+            DeleteOptions::default(),
+        ));
+        assert!(
+            futures::poll!(delete.as_mut()).is_pending(),
+            "the disk call should wait for the metadata lease"
+        );
+        assert_eq!(wrapper.health.waiting_count(), 1);
+
+        drop(delete);
+        assert_eq!(wrapper.health.waiting_count(), 0, "a dropped delete must release its waiting slot");
+        drop(metadata_lease);
+    }
+
    #[tokio::test]
    async fn delete_versions_counts_faulty_drive_availability_rejection() {
        let dir = tempfile::tempdir().expect("temp dir should be created");
@@ -0,0 +1,706 @@
+// 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.
+
+//! Per-local-disk foreground-priority admission for scanner I/O.
+
+use std::collections::VecDeque;
+use std::future::Future;
+use std::io;
+use std::pin::Pin;
+use std::sync::atomic::{AtomicBool, AtomicUsize, Ordering};
+use std::sync::{Arc, Mutex};
+use std::task::{Context, Poll};
+use std::time::Duration;
+
+use metrics::{counter, gauge, histogram};
+use rustfs_rio::ChunkReader;
+use tokio::io::{AsyncRead, ReadBuf};
+use tokio::sync::Notify;
+use tokio::time::Instant;
+use tokio_util::sync::CancellationToken;
+
+const BACKGROUND_FOREGROUND_BURST: u32 = 8;
+const BACKGROUND_MAX_WAIT: Duration = Duration::from_millis(100);
+
+tokio::task_local! {
+    static DISK_IO_CONTEXT: DiskIoContext;
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub(crate) enum DiskIoClass {
+    Foreground,
+    Background,
+}
+
+#[derive(Clone, Debug)]
+pub(crate) struct DiskIoContext {
+    pub(crate) class: DiskIoClass,
+    pub(crate) cancellation: CancellationToken,
+    pub(crate) admission: Option<Arc<DiskIoAdmission>>,
+    pub(crate) permit: Option<DiskIoPermit>,
+}
+
+/// Mark disk I/O issued by `future` as background work for local-disk admission.
+#[doc(hidden)]
+pub async fn with_background_disk_io<F: Future>(cancellation: CancellationToken, future: F) -> F::Output {
+    with_disk_io_context(
+        DiskIoContext {
+            class: DiskIoClass::Background,
+            cancellation,
+            admission: None,
+            permit: None,
+        },
+        future,
+    )
+    .await
+}
+
+pub(crate) async fn with_disk_io_context<F: Future>(context: DiskIoContext, future: F) -> F::Output {
+    DISK_IO_CONTEXT.scope(context, future).await
+}
+
+pub(crate) async fn with_disk_io_permit<F: Future>(permit: Option<DiskIoPermit>, future: F) -> F::Output {
+    let Some(permit) = permit else {
+        return future.await;
+    };
+    let mut context = current_disk_io_context().unwrap_or_else(|| DiskIoContext {
+        class: DiskIoClass::Foreground,
+        cancellation: CancellationToken::new(),
+        admission: None,
+        permit: None,
+    });
+    context.admission = Some(permit.admission());
+    with_disk_io_context(
+        DiskIoContext {
+            permit: Some(permit),
+            ..context
+        },
+        future,
+    )
+    .await
+}
+
+pub(crate) async fn with_optional_disk_io_context<F: Future>(context: Option<DiskIoContext>, future: F) -> F::Output {
+    match context {
+        Some(context) => with_disk_io_context(context, future).await,
+        None => future.await,
+    }
+}
+
+pub(crate) fn current_disk_io_context() -> Option<DiskIoContext> {
+    DISK_IO_CONTEXT.try_with(Clone::clone).ok()
+}
+
+/// Acquire one low-level permit only when the current walker explicitly owns a gate.
+/// Ordinary wrapper calls use their outer operation permit instead.
+pub(crate) async fn acquire_context_disk_io_permit() -> io::Result<Option<DiskIoPermit>> {
+    let Some(context) = current_disk_io_context() else {
+        return Ok(None);
+    };
+    let Some(admission) = context.admission else {
+        return Ok(None);
+    };
+    if let Some(permit) = context.permit
+        && permit.is_for(&admission)
+    {
+        return Ok(Some(permit));
+    }
+    match context.class {
+        DiskIoClass::Foreground => Ok(Some(admission.foreground())),
+        DiskIoClass::Background => Ok(Some(admission.background(&context.cancellation).await?)),
+    }
+}
+
+#[derive(Debug, Default)]
+struct AdmissionState {
+    foreground_completions: u32,
+    next_background_ticket: u64,
+    background_queue: VecDeque<u64>,
+    background_front_since: Option<Instant>,
+}
+
+/// One gate is shared by all handles to a local disk, including reconnects.
+/// Foreground I/O never waits on this gate. Background I/O runs exclusively
+/// while the disk is idle, or after a bounded foreground burst/wait.
+#[derive(Debug)]
+pub(crate) struct DiskIoAdmission {
+    state: Mutex<AdmissionState>,
+    changed: Notify,
+    disk: String,
+    foreground_active: AtomicUsize,
+    background_active: AtomicBool,
+    background_waiters: AtomicUsize,
+    #[cfg(test)]
+    background_admissions: AtomicUsize,
+}
+
+impl DiskIoAdmission {
+    pub(crate) fn new(disk: impl Into<String>) -> Arc<Self> {
+        Arc::new(Self {
+            state: Mutex::new(AdmissionState::default()),
+            changed: Notify::new(),
+            disk: disk.into(),
+            foreground_active: AtomicUsize::new(0),
+            background_active: AtomicBool::new(false),
+            background_waiters: AtomicUsize::new(0),
+            #[cfg(test)]
+            background_admissions: AtomicUsize::new(0),
+        })
+    }
+
+    #[cfg(test)]
+    pub(crate) fn background_admissions_for_tests(&self) -> usize {
+        self.background_admissions.load(Ordering::Relaxed)
+    }
+
+    #[cfg(test)]
+    pub(crate) fn background_active_for_tests(&self) -> bool {
+        self.background_active.load(Ordering::Acquire)
+    }
+
+    #[cfg(test)]
+    fn foreground_active_for_tests(&self) -> usize {
+        self.foreground_active.load(Ordering::Acquire)
+    }
+
+    pub(crate) fn foreground(self: &Arc<Self>) -> DiskIoPermit {
+        self.foreground_active.fetch_add(1, Ordering::AcqRel);
+        if self.background_waiters.load(Ordering::Acquire) > 0 {
+            self.record_state();
+        }
+        DiskIoPermit {
+            lease: Arc::new(DiskIoPermitLease {
+                admission: self.clone(),
+                class: DiskIoClass::Foreground,
+            }),
+        }
+    }
+
+    pub(crate) async fn background(self: &Arc<Self>, cancellation: &CancellationToken) -> std::io::Result<DiskIoPermit> {
+        let started_at = Instant::now();
+        let ticket = {
+            let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
+            let ticket = state.next_background_ticket;
+            state.next_background_ticket = state.next_background_ticket.wrapping_add(1);
+            if state.background_queue.is_empty() && !self.background_active.load(Ordering::Acquire) {
+                state.background_front_since = Some(Instant::now());
+            }
+            state.background_queue.push_back(ticket);
+            self.background_waiters.fetch_add(1, Ordering::Release);
+            ticket
+        };
+        self.record_state();
+
+        let mut waiter = BackgroundWaiter {
+            admission: self.clone(),
+            ticket,
+            waiting: true,
+        };
+        loop {
+            let notified = self.changed.notified();
+            tokio::pin!(notified);
+            notified.as_mut().enable();
+
+            let granted = {
+                let mut state = self.state.lock().unwrap_or_else(|error| error.into_inner());
+                let is_front = state.background_queue.front() == Some(&ticket);
+                let bounded_wait_elapsed = is_front
+                    && state
+                        .background_front_since
+                        .is_some_and(|front_since| front_since.elapsed() >= BACKGROUND_MAX_WAIT);
+                if is_front
+                    && !self.background_active.load(Ordering::Acquire)
+                    && (self.foreground_active.load(Ordering::Acquire) == 0
+                        || state.foreground_completions >= BACKGROUND_FOREGROUND_BURST
+                        || bounded_wait_elapsed)
+                {
+                    self.background_active.store(true, Ordering::Release);
+                    state.background_queue.pop_front();
+                    state.background_front_since = None;
+                    self.background_waiters.fetch_sub(1, Ordering::AcqRel);
+                    state.foreground_completions = 0;
+                    true
+                } else {
+                    false
+                }
+            };
+
+            if granted {
+                waiter.waiting = false;
+                #[cfg(test)]
+                self.background_admissions.fetch_add(1, Ordering::Relaxed);
+                histogram!("rustfs_disk_io_admission_wait_seconds", "disk" => self.disk.clone(), "class" => "background")
+                    .record(started_at.elapsed().as_secs_f64());
+                counter!("rustfs_disk_io_admission_total", "disk" => self.disk.clone(), "class" => "background").increment(1);
+                self.record_state();
+                return Ok(DiskIoPermit {
+                    lease: Arc::new(DiskIoPermitLease {
+                        admission: self.clone(),
+                        class: DiskIoClass::Background,
+                    }),
+                });
+            }
+
+            let remaining = {
+                let state = self.state.lock().unwrap_or_else(|error| error.into_inner());
+                if state.background_queue.front() == Some(&ticket) {
+                    state
+                        .background_front_since
+                        .map(|front_since| BACKGROUND_MAX_WAIT.saturating_sub(front_since.elapsed()))
+                } else {
+                    None
+                }
+            };
+            if let Some(remaining) = remaining.filter(|remaining| !remaining.is_zero()) {
+                tokio::select! {
+                    _ = cancellation.cancelled() => {
+                        return Err(std::io::Error::new(std::io::ErrorKind::Interrupted, "background disk I/O admission cancelled"));
+                    }
+                    _ = tokio::time::sleep(remaining) => {
+                        self.changed.notify_waiters();
+                    }
+                    _ = &mut notified => {}
+                }
+            } else {
+                tokio::select! {
+                    _ = cancellation.cancelled() => {
+                        return Err(std::io::Error::new(std::io::ErrorKind::Interrupted, "background disk I/O admission cancelled"));
+                    }
+                    _ = &mut notified => {}
+                }
+            }
+        }
+    }
+
+    fn record_state(&self) {
+        let (foreground_active, background_waiters) = {
+            let state = self.state.lock().unwrap_or_else(|error| error.into_inner());
+            (
+                if state.background_queue.is_empty() {
+                    0
+                } else {
+                    self.foreground_active.load(Ordering::Acquire)
+                },
+                state.background_queue.len(),
+            )
+        };
+        let foreground_active = u32::try_from(foreground_active).unwrap_or(u32::MAX);
+        let background_waiters = u32::try_from(background_waiters).unwrap_or(u32::MAX);
+        gauge!("rustfs_disk_io_admission_active", "disk" => self.disk.clone(), "class" => "foreground")
+            .set(f64::from(foreground_active));
+        gauge!("rustfs_disk_io_admission_active", "disk" => self.disk.clone(), "class" => "background").set(
+            if self.background_active.load(Ordering::Acquire) {
+                1.0
+            } else {
+                0.0
+            },
+        );
+        gauge!("rustfs_disk_io_admission_waiters", "disk" => self.disk.clone(), "class" => "background")
+            .set(f64::from(background_waiters));
+    }
+}
+
+struct BackgroundWaiter {
+    admission: Arc<DiskIoAdmission>,
+    ticket: u64,
+    waiting: bool,
+}
+
+impl Drop for BackgroundWaiter {
+    fn drop(&mut self) {
+        if !self.waiting {
+            return;
+        }
+        let mut state = self.admission.state.lock().unwrap_or_else(|error| error.into_inner());
+        if let Some(index) = state.background_queue.iter().position(|ticket| *ticket == self.ticket) {
+            state.background_queue.remove(index);
+            if index == 0 {
+                state.background_front_since =
+                    if state.background_queue.is_empty() || self.admission.background_active.load(Ordering::Acquire) {
+                        None
+                    } else {
+                        Some(Instant::now())
+                    };
+            }
+            self.admission.background_waiters.fetch_sub(1, Ordering::AcqRel);
+        }
+        drop(state);
+        self.admission.record_state();
+        self.admission.changed.notify_waiters();
+    }
+}
+
+#[derive(Debug, Clone)]
+pub(crate) struct DiskIoPermit {
+    lease: Arc<DiskIoPermitLease>,
+}
+
+impl DiskIoPermit {
+    pub(crate) fn is_for(&self, admission: &Arc<DiskIoAdmission>) -> bool {
+        Arc::ptr_eq(&self.lease.admission, admission)
+    }
+
+    fn admission(&self) -> Arc<DiskIoAdmission> {
+        self.lease.admission.clone()
+    }
+}
+
+#[derive(Debug)]
+struct DiskIoPermitLease {
+    admission: Arc<DiskIoAdmission>,
+    class: DiskIoClass,
+}
+
+pub(crate) fn spawn_blocking_with_disk_io_permit<T, F>(
+    io_permit: Option<DiskIoPermit>,
+    operation: F,
+) -> tokio::task::JoinHandle<T>
+where
+    T: Send + 'static,
+    F: FnOnce() -> T + Send + 'static,
+{
+    tokio::task::spawn_blocking(move || {
+        let _io_permit = io_permit;
+        operation()
+    })
+}
+
+type AdmissionWait = Pin<Box<dyn Future<Output = io::Result<DiskIoPermit>> + Send>>;
+
+impl Drop for DiskIoPermitLease {
+    fn drop(&mut self) {
+        match self.class {
+            DiskIoClass::Foreground => {
+                self.admission.foreground_active.fetch_sub(1, Ordering::AcqRel);
+                let record_state = self.admission.background_waiters.load(Ordering::Acquire) > 0;
+                if record_state {
+                    let mut state = self.admission.state.lock().unwrap_or_else(|error| error.into_inner());
+                    if !state.background_queue.is_empty() {
+                        state.foreground_completions = state.foreground_completions.saturating_add(1);
+                    }
+                }
+                if record_state {
+                    self.admission.record_state();
+                    self.admission.changed.notify_waiters();
+                }
+            }
+            DiskIoClass::Background => {
+                let mut state = self.admission.state.lock().unwrap_or_else(|error| error.into_inner());
+                self.admission.background_active.store(false, Ordering::Release);
+                state.background_front_since = if state.background_queue.is_empty() {
+                    None
+                } else {
+                    Some(Instant::now())
+                };
+                drop(state);
+                self.admission.record_state();
+                self.admission.changed.notify_waiters();
+            }
+        }
+    }
+}
+
+pub(crate) struct AdmissionReader<R> {
+    inner: R,
+    admission: Arc<DiskIoAdmission>,
+    context: DiskIoContext,
+    wait: Mutex<Option<AdmissionWait>>,
+    permit: Option<DiskIoPermit>,
+}
+
+impl<R> AdmissionReader<R> {
+    pub(crate) fn new(inner: R, admission: Arc<DiskIoAdmission>, context: DiskIoContext) -> Self {
+        Self {
+            inner,
+            admission,
+            context,
+            wait: Mutex::new(None),
+            permit: None,
+        }
+    }
+
+    fn poll_permit(&mut self, cx: &mut Context<'_>) -> Poll<io::Result<()>> {
+        if self.permit.is_some() {
+            return Poll::Ready(Ok(()));
+        }
+
+        if self.context.class == DiskIoClass::Foreground {
+            self.permit = Some(self.admission.foreground());
+            return Poll::Ready(Ok(()));
+        }
+
+        let mut wait = self.wait.lock().unwrap_or_else(|error| error.into_inner());
+        if wait.is_none() {
+            let admission = self.admission.clone();
+            let cancellation = self.context.cancellation.clone();
+            *wait = Some(Box::pin(async move { admission.background(&cancellation).await }));
+        }
+        let Some(wait_future) = wait.as_mut() else {
+            return Poll::Ready(Err(io::Error::other("background disk I/O admission waiter is missing")));
+        };
+        let result = wait_future.as_mut().poll(cx);
+        match result {
+            Poll::Ready(Ok(permit)) => {
+                wait.take();
+                drop(wait);
+                self.permit = Some(permit);
+                Poll::Ready(Ok(()))
+            }
+            Poll::Ready(Err(error)) => {
+                wait.take();
+                Poll::Ready(Err(error))
+            }
+            Poll::Pending => Poll::Pending,
+        }
+    }
+
+    fn finish_poll<T>(&mut self, result: Poll<io::Result<T>>) -> Poll<io::Result<T>> {
+        if result.is_ready() {
+            self.permit = None;
+        }
+        result
+    }
+}
+
+impl<R: AsyncRead + Unpin> AsyncRead for AdmissionReader<R> {
+    fn poll_read(mut self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+        match self.poll_permit(cx) {
+            Poll::Ready(Ok(())) => {}
+            Poll::Ready(Err(error)) => return Poll::Ready(Err(error)),
+            Poll::Pending => return Poll::Pending,
+        }
+        let result = Pin::new(&mut self.inner).poll_read(cx, buf);
+        self.finish_poll(result)
+    }
+}
+
+impl<R: ChunkReader + Unpin> ChunkReader for AdmissionReader<R> {
+    fn poll_read_chunk(mut self: Pin<&mut Self>, cx: &mut Context<'_>, max: usize) -> Poll<io::Result<Option<bytes::Bytes>>> {
+        match self.poll_permit(cx) {
+            Poll::Ready(Ok(())) => {}
+            Poll::Ready(Err(error)) => return Poll::Ready(Err(error)),
+            Poll::Pending => return Poll::Pending,
+        }
+        let result = Pin::new(&mut self.inner).poll_read_chunk(cx, max);
+        self.finish_poll(result)
+    }
+}
+
+pub(crate) struct BoxedChunkReader(pub(crate) rustfs_rio::ChunkReaderBox);
+
+impl AsyncRead for BoxedChunkReader {
+    fn poll_read(mut self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+        Pin::new(&mut self.0).poll_read(cx, buf)
+    }
+}
+
+impl ChunkReader for BoxedChunkReader {
+    fn poll_read_chunk(self: Pin<&mut Self>, cx: &mut Context<'_>, max: usize) -> Poll<io::Result<Option<bytes::Bytes>>> {
+        Pin::new(self.get_mut().0.as_mut()).poll_read_chunk(cx, max)
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use tokio::io::{AsyncReadExt, AsyncWriteExt};
+
+    struct BlockingWorker(tokio::task::JoinHandle<()>);
+
+    async fn wait_for_background_waiters(admission: &DiskIoAdmission, expected: usize) {
+        tokio::time::timeout(Duration::from_secs(1), async {
+            loop {
+                let waiters = admission
+                    .state
+                    .lock()
+                    .unwrap_or_else(|error| error.into_inner())
+                    .background_queue
+                    .len();
+                if waiters == expected {
+                    return;
+                }
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("background admission waiter should reach expected count");
+    }
+
+    async fn wait_for_foreground_active(admission: &DiskIoAdmission, expected: usize) {
+        tokio::time::timeout(Duration::from_secs(1), async {
+            loop {
+                let active = admission.foreground_active.load(Ordering::Acquire);
+                if active == expected {
+                    return;
+                }
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("foreground disk read should retain its permit while pending");
+    }
+
+    #[tokio::test]
+    async fn foreground_bypasses_background_and_background_progress_is_bounded() {
+        let admission = DiskIoAdmission::new("disk-a");
+        let foreground = admission.foreground();
+        let background_admission = admission.clone();
+        let cancellation = CancellationToken::new();
+        let background = tokio::spawn(async move { background_admission.background(&cancellation).await });
+
+        wait_for_background_waiters(&admission, 1).await;
+        for _ in 0..BACKGROUND_FOREGROUND_BURST {
+            drop(admission.foreground());
+        }
+
+        let background_permit = tokio::time::timeout(Duration::from_secs(1), background)
+            .await
+            .expect("background admission should progress after the foreground burst")
+            .expect("background waiter task should complete")
+            .expect("background admission should not be cancelled");
+        assert_eq!(
+            admission.foreground_active.load(Ordering::Acquire),
+            1,
+            "foreground operations remain admitted while one background operation runs"
+        );
+        drop(background_permit);
+        drop(foreground);
+    }
+
+    #[tokio::test]
+    async fn cancellation_removes_a_queued_background_waiter() {
+        let admission = DiskIoAdmission::new("disk-a");
+        let background = admission
+            .background(&CancellationToken::new())
+            .await
+            .expect("first background operation should be admitted");
+        let cancellation = CancellationToken::new();
+        let waiter_admission = admission.clone();
+        let waiter_cancellation = cancellation.clone();
+        let waiter = tokio::spawn(async move { waiter_admission.background(&waiter_cancellation).await });
+        wait_for_background_waiters(&admission, 1).await;
+
+        cancellation.cancel();
+        let error = tokio::time::timeout(Duration::from_secs(1), waiter)
+            .await
+            .expect("cancelled background waiter should stop")
+            .expect("background waiter task should complete")
+            .expect_err("cancelled admission must not start disk I/O");
+        assert_eq!(error.kind(), io::ErrorKind::Interrupted);
+        assert_eq!(
+            admission
+                .state
+                .lock()
+                .unwrap_or_else(|error| error.into_inner())
+                .background_queue
+                .len(),
+            0
+        );
+        drop(background);
+    }
+
+    #[tokio::test]
+    async fn background_admission_is_isolated_per_disk() {
+        let disk_a = DiskIoAdmission::new("disk-a");
+        let disk_b = DiskIoAdmission::new("disk-b");
+        let _disk_a_permit = disk_a
+            .background(&CancellationToken::new())
+            .await
+            .expect("disk A background operation should be admitted");
+        let _disk_b_permit = tokio::time::timeout(Duration::from_secs(1), disk_b.background(&CancellationToken::new()))
+            .await
+            .expect("disk A must not block disk B")
+            .expect("disk B background operation should be admitted");
+    }
+
+    #[tokio::test]
+    async fn streamed_read_holds_admission_until_disk_read_resolves() {
+        let admission = DiskIoAdmission::new("disk-a");
+        let (mut writer, reader) = tokio::io::duplex(8);
+        let mut reader = AdmissionReader::new(
+            reader,
+            admission.clone(),
+            DiskIoContext {
+                class: DiskIoClass::Foreground,
+                cancellation: CancellationToken::new(),
+                admission: None,
+                permit: None,
+            },
+        );
+        let read = tokio::spawn(async move {
+            let mut bytes = [0; 4];
+            reader
+                .read_exact(&mut bytes)
+                .await
+                .expect("streamed disk read should complete");
+            bytes
+        });
+
+        wait_for_foreground_active(&admission, 1).await;
+        writer.write_all(b"rust").await.expect("test bytes should be written");
+        assert_eq!(read.await.expect("reader task should finish"), *b"rust");
+        wait_for_foreground_active(&admission, 0).await;
+    }
+
+    #[tokio::test]
+    async fn blocking_worker_keeps_permit_after_caller_timeout() {
+        let admission = DiskIoAdmission::new("disk-a");
+        let (started_tx, started_rx) = tokio::sync::oneshot::channel();
+        let worker = spawn_blocking_with_disk_io_permit(Some(admission.foreground()), move || {
+            started_tx.send(()).expect("worker start should be observed");
+            std::thread::sleep(Duration::from_millis(100));
+        });
+
+        started_rx.await.expect("blocking worker should start");
+        assert!(tokio::time::timeout(Duration::from_millis(10), worker).await.is_err());
+        assert_eq!(admission.foreground_active_for_tests(), 1);
+        wait_for_foreground_active(&admission, 0).await;
+    }
+
+    #[tokio::test]
+    async fn cloned_permit_remains_active_until_all_workers_finish() {
+        let admission = DiskIoAdmission::new("disk-a");
+        let permit = admission.foreground();
+        let worker_permit = permit.clone();
+        drop(permit);
+        assert_eq!(admission.foreground_active_for_tests(), 1);
+
+        drop(worker_permit);
+        assert_eq!(admission.foreground_active_for_tests(), 0);
+    }
+
+    #[tokio::test]
+    async fn scoped_foreground_permit_is_inherited_by_blocking_worker() {
+        let admission = DiskIoAdmission::new("disk-a");
+        let permit = admission.foreground();
+        let worker = with_disk_io_permit(Some(permit), async {
+            let worker_permit = acquire_context_disk_io_permit()
+                .await
+                .expect("context permit should be available")
+                .expect("scoped permit should be inherited");
+            let (started_tx, started_rx) = tokio::sync::oneshot::channel();
+            let worker = spawn_blocking_with_disk_io_permit(Some(worker_permit), move || {
+                started_tx.send(()).expect("worker start should be observed");
+                std::thread::sleep(Duration::from_millis(100));
+            });
+            started_rx.await.expect("blocking worker should start");
+            BlockingWorker(worker)
+        })
+        .await;
+
+        assert_eq!(admission.foreground_active_for_tests(), 1);
+        assert!(tokio::time::timeout(Duration::from_millis(10), worker.0).await.is_err());
+        wait_for_foreground_active(&admission, 0).await;
+    }
+}
@@ -22,9 +22,16 @@ pub mod error_reduce;
 pub mod format;
 pub mod fs;
 pub mod health_state;
+mod io_admission;
 pub mod local;
+pub(crate) mod native_prepare_rpc;
 pub mod os;

+#[cfg(test)]
+pub(crate) use io_admission::DiskIoClass;
+pub use io_admission::with_background_disk_io;
+pub(crate) use io_admission::{acquire_context_disk_io_permit, current_disk_io_context, with_optional_disk_io_context};
+
 pub const RUSTFS_META_BUCKET: &str = ".rustfs.sys";
 pub const MIGRATING_META_BUCKET: &str = ".minio.sys";
 pub const RUSTFS_META_MULTIPART_BUCKET: &str = ".rustfs.sys/multipart";
@@ -224,6 +231,152 @@ pub enum Disk {
 }

 impl Disk {
+    #[cfg(unix)]
+    pub(crate) async fn native_begin_stream(
+        &self,
+        preparation: Uuid,
+        store_id: Uuid,
+        bucket: &str,
+        key: &str,
+        incarnation: Uuid,
+        fence: Arc<crate::object_api::ObjectLockConfigSnapshot>,
+    ) -> Result<()> {
+        match self {
+            Disk::Local(local) => {
+                local
+                    .get_disk()
+                    .begin_native_stream(preparation, store_id, bucket, incarnation, fence)
+                    .await
+            }
+            Disk::Remote(remote) => {
+                remote
+                    .native_begin_stream(preparation, store_id, bucket, key, incarnation)
+                    .await
+            }
+        }
+    }
+
+    #[cfg(not(unix))]
+    pub(crate) async fn native_begin_stream(
+        &self,
+        _preparation: Uuid,
+        _store_id: Uuid,
+        _bucket: &str,
+        _key: &str,
+        _incarnation: Uuid,
+        _fence: Arc<crate::object_api::ObjectLockConfigSnapshot>,
+    ) -> Result<()> {
+        Err(DiskError::UnsupportedDisk)
+    }
+
+    #[cfg(unix)]
+    pub(crate) async fn native_finalize(
+        &self,
+        input: local::native_prepare::NativePrepareInput,
+    ) -> Result<local::native_prepare::NativePreparationReceipt> {
+        match self {
+            Disk::Local(local) => local.get_disk().prepare_native(input).await,
+            Disk::Remote(remote) => remote.native_finalize(input).await,
+        }
+    }
+
+    #[cfg(not(unix))]
+    pub(crate) async fn native_finalize(
+        &self,
+        _input: local::native_prepare::NativePrepareInput,
+    ) -> Result<local::native_prepare::NativePreparationReceipt> {
+        Err(DiskError::UnsupportedDisk)
+    }
+
+    #[cfg(unix)]
+    pub(crate) async fn native_recover(
+        &self,
+        preparation: Uuid,
+        store_id: Uuid,
+        bucket: &str,
+        key: &str,
+        incarnation: Uuid,
+        fence: Arc<crate::object_api::ObjectLockConfigSnapshot>,
+    ) -> Result<(local::native_prepare::NativePreparationReceipt, Bytes)> {
+        match self {
+            Disk::Local(local) => {
+                local
+                    .get_disk()
+                    .recover_native_with_metadata(preparation, store_id, bucket, incarnation, fence)
+                    .await
+            }
+            Disk::Remote(remote) => remote.native_recover(preparation, store_id, bucket, key, incarnation).await,
+        }
+    }
+
+    #[cfg(not(unix))]
+    pub(crate) async fn native_recover(
+        &self,
+        _preparation: Uuid,
+        _store_id: Uuid,
+        _bucket: &str,
+        _key: &str,
+        _incarnation: Uuid,
+        _fence: Arc<crate::object_api::ObjectLockConfigSnapshot>,
+    ) -> Result<(local::native_prepare::NativePreparationReceipt, Bytes)> {
+        Err(DiskError::UnsupportedDisk)
+    }
+
+    #[cfg(unix)]
+    pub(crate) async fn native_probe(
+        &self,
+        preparation: Uuid,
+        store_id: Uuid,
+        bucket: &str,
+        key: &str,
+        incarnation: Uuid,
+        fence: Arc<crate::object_api::ObjectLockConfigSnapshot>,
+    ) -> Result<local::native_prepare::NativePreparationProbe> {
+        match self {
+            Disk::Local(local) => {
+                local
+                    .get_disk()
+                    .probe_native(preparation, store_id, bucket, key, incarnation, fence)
+                    .await
+            }
+            Disk::Remote(remote) => remote.native_probe(preparation, store_id, bucket, key, incarnation).await,
+        }
+    }
+
+    #[cfg(not(unix))]
+    pub(crate) async fn native_probe(
+        &self,
+        _preparation: Uuid,
+        _store_id: Uuid,
+        _bucket: &str,
+        _key: &str,
+        _incarnation: Uuid,
+        _fence: Arc<crate::object_api::ObjectLockConfigSnapshot>,
+    ) -> Result<local::native_prepare::NativePreparationProbe> {
+        Err(DiskError::UnsupportedDisk)
+    }
+
+    #[cfg(unix)]
+    pub(crate) async fn native_locate(
+        &self,
+        store_id: Uuid,
+        preparation: Uuid,
+    ) -> Result<Option<local::native_prepare::NativePreparationDescription>> {
+        match self {
+            Disk::Local(local) => local.get_disk().locate_native_preparation(store_id, preparation).await,
+            Disk::Remote(remote) => remote.native_locate(store_id, preparation).await,
+        }
+    }
+
+    #[cfg(not(unix))]
+    pub(crate) async fn native_locate(
+        &self,
+        _store_id: Uuid,
+        _preparation: Uuid,
+    ) -> Result<Option<local::native_prepare::NativePreparationDescription>> {
+        Err(DiskError::UnsupportedDisk)
+    }
+
    pub(crate) async fn set_disk_id_state(&self, id: Option<Uuid>) -> Result<()> {
        match self {
            Disk::Local(local_disk) => {
@@ -0,0 +1,145 @@
+// Copyright 2026 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use crate::disk::{
+    RUSTFS_META_BUCKET,
+    error::{DiskError, Result},
+};
+use rustfs_utils::path::path_join_buf;
+use uuid::Uuid;
+
+const NATIVE_STREAM_PREFIX: &str = "native-preparations/";
+const NATIVE_STREAM_ROOT: &str = "native-preparations";
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) enum NativeStreamFile {
+    Part(usize),
+    Staging(Uuid),
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+pub(crate) struct NativeStreamPath {
+    pub(crate) preparation: Uuid,
+    pub(crate) data_dir: Uuid,
+    pub(crate) file: NativeStreamFile,
+}
+
+fn targets_native_stream_namespace(volume: &str, path: &str) -> bool {
+    if !volume.contains(RUSTFS_META_BUCKET) && !path.contains(NATIVE_STREAM_ROOT) {
+        return false;
+    }
+    let normalized = path_join_buf(&[volume, path]);
+    let prefix = format!("{RUSTFS_META_BUCKET}/{NATIVE_STREAM_PREFIX}");
+    normalized == format!("{RUSTFS_META_BUCKET}/{NATIVE_STREAM_ROOT}") || normalized.starts_with(&prefix)
+}
+
+/// Recognize only the reserved per-preparation shard path used by the native
+/// stream writer. Malformed paths under that namespace must not fall through to
+/// the general disk writer.
+pub(crate) fn parse_native_stream_path(volume: &str, path: &str) -> Result<Option<NativeStreamPath>> {
+    if !targets_native_stream_namespace(volume, path) {
+        return Ok(None);
+    }
+    if volume != RUSTFS_META_BUCKET {
+        return Err(DiskError::InvalidPath);
+    }
+    if !path.starts_with(NATIVE_STREAM_PREFIX) {
+        return Err(DiskError::InvalidPath);
+    }
+
+    let mut components = path.split('/');
+    if components.next() != Some("native-preparations") {
+        return Err(DiskError::InvalidPath);
+    }
+    let preparation = components
+        .next()
+        .and_then(|value| Uuid::parse_str(value).ok())
+        .filter(|value| !value.is_nil())
+        .ok_or(DiskError::InvalidPath)?;
+    let data_dir = components
+        .next()
+        .and_then(|value| Uuid::parse_str(value).ok())
+        .filter(|value| !value.is_nil())
+        .ok_or(DiskError::InvalidPath)?;
+    let file_name = components.next().ok_or(DiskError::InvalidPath)?;
+    let file = file_name
+        .strip_prefix("part.")
+        .and_then(|value| value.parse::<usize>().ok())
+        .filter(|value| *value > 0)
+        .filter(|part| file_name == format!("part.{part}"))
+        .map(NativeStreamFile::Part)
+        .or_else(|| {
+            file_name
+                .strip_prefix(".rustfs-put-")
+                .and_then(|value| Uuid::parse_str(value).ok())
+                .filter(|value| !value.is_nil())
+                .filter(|id| file_name == format!(".rustfs-put-{id}"))
+                .map(NativeStreamFile::Staging)
+        });
+
+    if components.next().is_some()
+        || file.is_none()
+        || path != format!("native-preparations/{preparation}/{data_dir}/{file_name}")
+    {
+        return Err(DiskError::InvalidPath);
+    }
+    Ok(Some(NativeStreamPath {
+        preparation,
+        data_dir,
+        file: file.expect("validated native stream filename"),
+    }))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn native_stream_path_requires_canonical_identity_and_part() {
+        let preparation = Uuid::from_u128(1);
+        let data_dir = Uuid::from_u128(2);
+        let valid = format!("native-preparations/{preparation}/{data_dir}/part.1");
+        assert!(
+            parse_native_stream_path(RUSTFS_META_BUCKET, &valid)
+                .expect("valid path")
+                .is_some()
+        );
+        let staging = format!("native-preparations/{preparation}/{data_dir}/.rustfs-put-{data_dir}");
+        assert!(
+            parse_native_stream_path(RUSTFS_META_BUCKET, &staging)
+                .expect("valid staging path")
+                .is_some()
+        );
+        assert!(parse_native_stream_path("bucket", &valid).expect("other volume").is_none());
+        assert!(
+            parse_native_stream_path(RUSTFS_META_BUCKET, "ordinary/object")
+                .expect("ordinary path")
+                .is_none()
+        );
+        for invalid in [
+            "native-preparations/00000000-0000-0000-0000-000000000000/00000000-0000-0000-0000-000000000002/part.1".to_owned(),
+            format!("native-preparations/{preparation}/{data_dir}/part.0"),
+            format!("native-preparations/{preparation}/{data_dir}/part.01"),
+            format!("native-preparations/{preparation}/{data_dir}/.rustfs-put-{}", Uuid::nil()),
+            format!("native-preparations/{preparation}/{data_dir}/.rustfs-put-not-a-uuid"),
+            format!("native-preparations/{preparation}/{data_dir}/part.1/extra"),
+            format!("native-preparations/{preparation}/../part.1"),
+        ] {
+            assert!(matches!(
+                parse_native_stream_path(RUSTFS_META_BUCKET, &invalid),
+                Err(DiskError::InvalidPath)
+            ));
+        }
+    }
+}
@@ -497,7 +497,7 @@ where

    /// Write a (hash+data) block. Returns the number of data bytes written.
    /// Returns an error if called after a short write or if data exceeds shard_size.
-    #[hotpath::measure(label = "BitrotWriter::write", impl_type = "BitrotWriter")]
+    #[hotpath::measure(impl_type = "BitrotWriter", future = true)]
    pub async fn write(&mut self, buf: &[u8]) -> std::io::Result<usize> {
        if buf.is_empty() {
            return Ok(0);
@@ -791,7 +791,7 @@ impl Erasure {
    }

    #[tracing::instrument(level = "debug", skip_all, fields(data_len=data.len()))]
-    #[hotpath::measure(label = "Erasure::encode_data", impl_type = "Erasure")]
+    #[hotpath::measure(impl_type = "Erasure")]
    pub(crate) fn encode_data_block(&self, data: &[u8]) -> io::Result<EncodedBlock> {
        self.encode_data_block_inner(data)
    }
@@ -810,7 +810,7 @@ impl Erasure {
            .map(|block| block.into_shards(self.total_shard_count()))
    }

-    #[hotpath::measure(label = "Erasure::encode_data_owned", impl_type = "Erasure")]
+    #[hotpath::measure(impl_type = "Erasure")]
    pub(crate) fn encode_data_owned_block(&self, data: Vec<u8>) -> io::Result<EncodedBlock> {
        self.encode_data_owned_block_inner(data)
    }
@@ -844,7 +844,7 @@ impl Erasure {
            .map(|block| block.into_shards(self.total_shard_count()))
    }

-    #[hotpath::measure(label = "Erasure::encode_data_bytes_mut", impl_type = "Erasure")]
+    #[hotpath::measure(impl_type = "Erasure")]
    pub(crate) fn encode_data_bytes_mut_block(&self, data_buffer: BytesMut, data_len: usize) -> io::Result<EncodedBlock> {
        self.encode_buffer(data_buffer, data_len)
    }
@@ -442,12 +442,19 @@ impl WritePlan {
        }

        if let Some(encryption) = self.encryption {
+            let etag_nonce = match encryption.mode {
+                WriteEncryptionMode::Singlepart { base_nonce } | WriteEncryptionMode::MultipartLegacy { base_nonce, .. } => {
+                    base_nonce
+                }
+                WriteEncryptionMode::SinglepartObjectKey | WriteEncryptionMode::MultipartObjectKey { .. } => [0u8; 12],
+            };
+            let protected_reader = rustfs_rio::EtagReader::new_encrypted(reader, encryption.key_bytes, etag_nonce)?;
            reader = match encryption.mode {
                WriteEncryptionMode::SinglepartObjectKey => HashReader::from_reader(
                    #[cfg(feature = "rio-v2")]
-                    EncryptReader::new_with_object_key(reader, encryption.key_bytes),
+                    EncryptReader::new_with_object_key(protected_reader, encryption.key_bytes),
                    #[cfg(not(feature = "rio-v2"))]
-                    EncryptReader::new(reader, encryption.key_bytes, [0u8; 12]),
+                    EncryptReader::new(protected_reader, encryption.key_bytes, [0u8; 12]),
                    HashReader::SIZE_PRESERVE_LAYER,
                    actual_size,
                    None,
@@ -457,12 +464,12 @@ impl WritePlan {
                WriteEncryptionMode::Singlepart { base_nonce } => {
                    #[cfg(not(feature = "rio-v2"))]
                    let encrypt_reader = if encryption_frame_v2_enabled() {
-                        EncryptReader::new_v2(reader, encryption.key_bytes, base_nonce)
+                        EncryptReader::new_v2(protected_reader, encryption.key_bytes, base_nonce)
                    } else {
-                        EncryptReader::new(reader, encryption.key_bytes, base_nonce)
+                        EncryptReader::new(protected_reader, encryption.key_bytes, base_nonce)
                    };
                    #[cfg(feature = "rio-v2")]
-                    let encrypt_reader = EncryptReader::new(reader, encryption.key_bytes, base_nonce);
+                    let encrypt_reader = EncryptReader::new(protected_reader, encryption.key_bytes, base_nonce);
                    HashReader::from_reader(encrypt_reader, HashReader::SIZE_PRESERVE_LAYER, actual_size, None, None, false)?
                }
                WriteEncryptionMode::MultipartLegacy {
@@ -471,20 +478,25 @@ impl WritePlan {
                } => {
                    #[cfg(not(feature = "rio-v2"))]
                    let encrypt_reader = if encryption_frame_v2_enabled() {
-                        EncryptReader::new_multipart_v2(reader, encryption.key_bytes, base_nonce, multipart_part_number)
+                        EncryptReader::new_multipart_v2(protected_reader, encryption.key_bytes, base_nonce, multipart_part_number)
                    } else {
-                        EncryptReader::new_multipart(reader, encryption.key_bytes, base_nonce, multipart_part_number)
+                        EncryptReader::new_multipart(protected_reader, encryption.key_bytes, base_nonce, multipart_part_number)
                    };
                    #[cfg(feature = "rio-v2")]
                    let encrypt_reader =
-                        EncryptReader::new_multipart(reader, encryption.key_bytes, base_nonce, multipart_part_number);
+                        EncryptReader::new_multipart(protected_reader, encryption.key_bytes, base_nonce, multipart_part_number);
                    HashReader::from_reader(encrypt_reader, HashReader::SIZE_PRESERVE_LAYER, actual_size, None, None, false)?
                }
                WriteEncryptionMode::MultipartObjectKey { multipart_part_number } => HashReader::from_reader(
                    #[cfg(feature = "rio-v2")]
-                    EncryptReader::new_multipart_with_object_key(reader, encryption.key_bytes, multipart_part_number),
+                    EncryptReader::new_multipart_with_object_key(protected_reader, encryption.key_bytes, multipart_part_number),
                    #[cfg(not(feature = "rio-v2"))]
-                    EncryptReader::new_multipart(reader, encryption.key_bytes, [0u8; 12], multipart_part_number as usize),
+                    EncryptReader::new_multipart(
+                        protected_reader,
+                        encryption.key_bytes,
+                        [0u8; 12],
+                        multipart_part_number as usize,
+                    ),
                    HashReader::SIZE_PRESERVE_LAYER,
                    actual_size,
                    None,
@@ -497,6 +509,7 @@ impl WritePlan {
        // `ignore_value` deliberately avoids a second hasher over compressed or
        // encrypted bytes. The inner reader still validates the plaintext request
        // checksum while this outer reader exposes the request checksum context.
+        reader.encrypted_etag = encrypted;
        reader.add_non_trailing_checksum(checksum, true)?;
        reader.set_trailer(trailer);

@@ -513,6 +526,36 @@ mod tests {
    use std::io::Cursor;
    use tokio::io::AsyncReadExt;

+    #[tokio::test]
+    async fn write_plan_encryption_persists_protected_etags_for_every_mode() {
+        let data = b"Salary 2026: 85000 EUR";
+        let size = i64::try_from(data.len()).expect("test data size");
+        let mut control =
+            HashReader::from_stream(Cursor::new(data), size, size, None, None, false).expect("create control reader");
+        control.read_to_end(&mut Vec::new()).await.expect("read control plaintext");
+        let md5 = control.try_resolve_etag().expect("control MD5");
+        for encryption in [
+            WriteEncryption::singlepart([1; 32], [1; 12]),
+            WriteEncryption::singlepart_object_key([1; 32]),
+            WriteEncryption::multipart([1; 32], [1; 12], 1),
+            WriteEncryption::multipart_object_key([1; 32], 1),
+        ] {
+            let plaintext = HashReader::from_stream(Cursor::new(data), size, size, Some(md5.clone()), None, false)
+                .expect("create checked reader");
+            let mut reader = WritePlan::new()
+                .with_encryption(encryption)
+                .apply(plaintext, size)
+                .expect("apply encryption");
+            let mut ciphertext = Vec::new();
+            reader.read_to_end(&mut ciphertext).await.expect("stream ciphertext");
+            assert_ne!(ciphertext, data);
+            let tag = reader.try_resolve_etag().expect("resolve persisted encrypted ETag");
+            assert_ne!(tag, md5, "stored ETag must not reveal the plaintext MD5");
+            assert_eq!(tag.len(), 32, "multipart ETags require a 16-byte hexadecimal part tag");
+            assert_eq!(reader.try_resolve_etag().as_ref(), Some(&tag));
+        }
+    }
+
    async fn assert_non_trailing_checksum_survives(plan: WritePlan) {
        let plaintext = b"checksum-context-through-write-plan".repeat(256);
        let actual_size = plaintext.len() as i64;
@@ -55,6 +55,7 @@ mod set_disk;
 mod storage_api_contracts;
 mod store;

+pub use disk::with_background_disk_io;
 pub use store::PoolMetaWriteGateStatus;

 // pub mod checksum;
@@ -1237,7 +1237,8 @@ impl ObjectOptions {
        self.version_id.is_some() || !self.versioned || self.version_suspended
    }

-    pub(crate) fn add_namespace_lock_lost_signal(&mut self, signal: Arc<rustfs_lock::distributed_lock::LockLostSignal>) {
+    #[doc(hidden)]
+    pub fn add_namespace_lock_lost_signal(&mut self, signal: Arc<rustfs_lock::distributed_lock::LockLostSignal>) {
        #[cfg(test)]
        let test_fence = NAMESPACE_LOCK_SIGNAL_TEST_FENCES
            .get_or_init(|| std::sync::Mutex::new(Vec::new()))
@@ -58,11 +58,17 @@ impl WriteCommitGuard {
        }
    }

+    /// Return the loss signal of the namespace owner retained by acquisition.
+    #[doc(hidden)]
+    pub fn lock_lost_signal(&self) -> Option<Arc<rustfs_lock::distributed_lock::LockLostSignal>> {
+        self.guards.first().and_then(|guard| guard.lock_lost_signal())
+    }
+
    pub fn is_lock_lost(&self) -> bool {
        self.guards.iter().any(|guard| guard.is_lock_lost() || guard.is_released())
    }

-    fn matches(&self, bucket: &str, object: &str) -> bool {
+    pub(crate) fn matches(&self, bucket: &str, object: &str) -> bool {
        self.bucket.as_ref() == bucket && self.object.as_ref() == object
    }
 }
@@ -307,6 +307,13 @@ pub fn shutdown_background_services() {
    }
 }

+/// Wait up to `timeout` for object mutations that outlived their callers
+/// (see `InstanceContext::wait_for_detached_mutations`). Returns `false` when
+/// some are still running at the deadline.
+pub async fn wait_for_detached_mutations(timeout: std::time::Duration) -> bool {
+    current_ctx().wait_for_detached_mutations(timeout).await
+}
+
 /// Set the global lock client (first LocalClient created)
 ///
 /// # Arguments
@@ -59,6 +59,8 @@ use std::sync::{
 use tokio::sync::{Mutex, Notify, OnceCell, OwnedRwLockReadGuard, RwLock};
 use tokio::time::Instant;
 use tokio_util::sync::CancellationToken;
+use tokio_util::task::TaskTracker;
+use tracing::Instrument as _;
 use uuid::Uuid;

 const SCANNER_PUBLICATION_STATE_UNKNOWN: u8 = 0;
@@ -72,10 +74,19 @@ pub(crate) const SCANNER_PUBLICATION_LEASE_MAX_ENTRIES: usize = 256;
 /// have started movement after the lease was abandoned.
 pub(crate) const SCANNER_PUBLICATION_LEASE_TTL: std::time::Duration = std::time::Duration::from_secs(60);

+/// Observational leases hold data-movement fencing but may coexist with
+/// ordinary namespace commits; authoritative leases require a stable namespace.
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub(crate) enum ScannerPublicationLeasePurpose {
+    Authoritative,
+    Observational,
+}
+
 pub(crate) struct ScannerPublicationLeaseEntry {
    pub(crate) expires_at: Instant,
    pub(crate) movement_generation: u64,
-    pub(crate) namespace_generation: u64,
+    pub(crate) namespace_generation: Option<u64>,
+    pub(crate) purpose: ScannerPublicationLeasePurpose,
    pub(crate) _operation_guard: OwnedRwLockReadGuard<()>,
 }

@@ -234,6 +245,10 @@ pub struct InstanceContext {
    suppress_tier_delete_journal_recovery: bool,
    transition_transaction_recovery_stores: std::sync::Mutex<HashSet<Uuid>>,
    tier_delete_journal_recovery_wakeup: tokio::sync::Notify,
+    /// Object mutations that run on their own task so that dropping the
+    /// caller (for example, an S3 client disconnect) cannot stop them between
+    /// their on-disk steps. Shutdown waits for them after the HTTP drain.
+    detached_mutations: TaskTracker,
 }

 impl InstanceContext {
@@ -284,6 +299,7 @@ impl InstanceContext {
            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(),
+            detached_mutations: TaskTracker::new(),
        }
    }

@@ -297,6 +313,31 @@ impl InstanceContext {
        self.lock_manager.clone()
    }

+    /// Run a multi-step mutation on its own task and wait for its output.
+    ///
+    /// Dropping the returned future leaves the task running to completion, so
+    /// the mutation, together with the locks and guards it owns, never stops
+    /// between two of its on-disk steps because its caller went away.
+    pub(crate) async fn run_detached_mutation<F>(&self, mutation: F) -> Result<F::Output, tokio::task::JoinError>
+    where
+        F: std::future::Future + Send + 'static,
+        F::Output: Send + 'static,
+    {
+        self.detached_mutations.spawn(mutation.in_current_span()).await
+    }
+
+    /// Wait up to `timeout` for every detached mutation to finish. Returns
+    /// `false` when some are still running at the deadline.
+    pub async fn wait_for_detached_mutations(&self, timeout: std::time::Duration) -> bool {
+        self.detached_mutations.close();
+        tokio::time::timeout(timeout, self.detached_mutations.wait()).await.is_ok()
+    }
+
+    #[cfg(test)]
+    pub(crate) fn detached_mutation_count(&self) -> usize {
+        self.detached_mutations.len()
+    }
+
    pub(crate) fn data_movement_operation_gate(&self) -> Arc<RwLock<()>> {
        Arc::clone(&self.data_movement_operation_gate)
    }
@@ -306,7 +347,8 @@ impl InstanceContext {
        token: Uuid,
        expires_at: Instant,
        movement_generation: u64,
-        namespace_generation: u64,
+        namespace_generation: Option<u64>,
+        purpose: ScannerPublicationLeasePurpose,
        operation_guard: OwnedRwLockReadGuard<()>,
    ) -> bool {
        let mut leases = self.scanner_publication_leases.lock().await;
@@ -319,6 +361,7 @@ impl InstanceContext {
                expires_at,
                movement_generation,
                namespace_generation,
+                purpose,
                _operation_guard: operation_guard,
            },
        );
@@ -329,21 +372,24 @@ impl InstanceContext {
        self.scanner_publication_leases.lock().await.remove(&token).is_some()
    }

-    /// Return both generations from the same live lease while the caller holds
-    /// the movement read guard. Namespace commits do not take that guard, so
-    /// the caller must compare the saved namespace generation after this await.
+    /// Return the generations and purpose from the same live lease while the
+    /// caller holds the movement read guard. Namespace commits do not take that
+    /// guard, so strict leases compare their saved generation after this await.
    /// The process-owned table rejects tokens from a prior instance or expiry.
-    pub(crate) async fn scanner_publication_lease_generations(&self, token: Uuid) -> Option<(u64, u64)> {
+    pub(crate) async fn scanner_publication_lease_generations(
+        &self,
+        token: Uuid,
+    ) -> Option<(u64, Option<u64>, ScannerPublicationLeasePurpose)> {
        let mut leases = self.scanner_publication_leases.lock().await;
        let now = Instant::now();
-        let (expires_at, movement_generation, namespace_generation) = leases
+        let (expires_at, movement_generation, namespace_generation, purpose) = leases
            .get(&token)
-            .map(|entry| (entry.expires_at, entry.movement_generation, entry.namespace_generation))?;
+            .map(|entry| (entry.expires_at, entry.movement_generation, entry.namespace_generation, entry.purpose))?;
        if expires_at <= now {
            leases.remove(&token);
            return None;
        }
-        Some((movement_generation, namespace_generation))
+        Some((movement_generation, namespace_generation, purpose))
    }

    pub(crate) async fn expire_scanner_publication_lease(&self, token: Uuid, expires_at: Instant) {
@@ -859,12 +905,22 @@ mod tests {
        let gate = ctx.data_movement_operation_gate();
        let permit = gate.clone().read_owned().await;
        assert!(
-            ctx.install_scanner_publication_lease(token, Instant::now() + SCANNER_PUBLICATION_LEASE_TTL, 7, 11, permit)
-                .await
+            ctx.install_scanner_publication_lease(
+                token,
+                Instant::now() + SCANNER_PUBLICATION_LEASE_TTL,
+                7,
+                Some(11),
+                ScannerPublicationLeasePurpose::Authoritative,
+                permit,
+            )
+            .await
        );
        drop(ctx.begin_namespace_commit());
        assert_eq!(ctx.namespace_commit_generation(), 2);
-        assert_eq!(ctx.scanner_publication_lease_generations(token).await, Some((7, 11)));
+        assert_eq!(
+            ctx.scanner_publication_lease_generations(token).await,
+            Some((7, Some(11), ScannerPublicationLeasePurpose::Authoritative))
+        );
        assert!(gate.clone().try_write_owned().is_err(), "lookup must retain the stored permit");
        tokio::time::advance(SCANNER_PUBLICATION_LEASE_TTL).await;
        assert_eq!(ctx.scanner_publication_lease_generations(token).await, None);
@@ -896,6 +952,33 @@ mod tests {
        }
    }

+    // A detached mutation keeps running after its caller is dropped, and
+    // shutdown waits for it instead of tearing it down mid-way.
+    #[tokio::test]
+    async fn detached_mutation_outlives_dropped_caller_and_is_awaited() {
+        let ctx = InstanceContext::new();
+        let finished = Arc::new(AtomicBool::new(false));
+        let (release_tx, release_rx) = tokio::sync::oneshot::channel::<()>();
+        let mut caller = Box::pin(ctx.run_detached_mutation({
+            let finished = Arc::clone(&finished);
+            async move {
+                let _ = release_rx.await;
+                finished.store(true, Ordering::Release);
+            }
+        }));
+        assert!(futures::poll!(caller.as_mut()).is_pending());
+        drop(caller);
+        assert_eq!(ctx.detached_mutation_count(), 1, "dropping the caller must not cancel the mutation");
+
+        assert!(
+            !ctx.wait_for_detached_mutations(std::time::Duration::from_millis(50)).await,
+            "the wait must report a mutation that is still running"
+        );
+        release_tx.send(()).expect("the detached mutation should still be waiting");
+        assert!(ctx.wait_for_detached_mutations(std::time::Duration::from_secs(5)).await);
+        assert!(finished.load(Ordering::Acquire));
+    }
+
    // A fresh context (before any update) reflects the initial all-false state.
    #[tokio::test]
    async fn fresh_context_is_all_false() {
@@ -2888,8 +2888,23 @@ impl NotificationSys {
    /// leases are released before returning so a partial acquisition cannot
    /// pin movement on one peer.
    pub async fn acquire_scanner_publication_leases(
+        &self,
+        targets: Vec<(String, String, u64)>,
+    ) -> Result<Vec<ScannerPublicationLeaseGrant>> {
+        self.acquire_scanner_publication_leases_with_purpose(targets, false).await
+    }
+
+    pub async fn acquire_observational_scanner_publication_leases(
+        &self,
+        targets: Vec<(String, String, u64)>,
+    ) -> Result<Vec<ScannerPublicationLeaseGrant>> {
+        self.acquire_scanner_publication_leases_with_purpose(targets, true).await
+    }
+
+    async fn acquire_scanner_publication_leases_with_purpose(
        &self,
        mut targets: Vec<(String, String, u64)>,
+        observational_only: bool,
    ) -> Result<Vec<ScannerPublicationLeaseGrant>> {
        targets.sort_by(|left, right| left.0.cmp(&right.0));
        for pair in targets.windows(2) {
@@ -2910,7 +2925,10 @@ impl NotificationSys {
                let _ = self.release_scanner_publication_leases(grants).await;
                return Err(Error::other(format!("scanner publication lease peer {host} is unavailable")));
            };
-            match client.acquire_scanner_publication_lease(&session_id, generation).await {
+            match client
+                .acquire_scanner_publication_lease(&session_id, generation, observational_only)
+                .await
+            {
                Ok(lease) => grants.push(ScannerPublicationLeaseGrant { host, lease }),
                Err(err) => {
                    let _ = self.release_scanner_publication_leases(grants).await;
@@ -4949,6 +4967,7 @@ mod tests {
                lease: ScannerPublicationLease {
                    token: Uuid::new_v4(),
                    movement_generation: 3,
+                    observational_only: false,
                    owner_id: Uuid::new_v4().to_string(),
                    session_id: "session-a".to_string(),
                    expires_at: Instant::now() + Duration::from_secs(30),
@@ -298,6 +298,11 @@ pub fn build_transition_put_options(storage_class: String, mut metadata: HashMap
        }
    }

+    // x-minio-internal-* is reserved by MinIO and its derivatives; they reject
+    // any request carrying it with 400 InvalidArgument. The transition ids above
+    // are the only internal keys the remote needs, and they travel as x-amz-meta-*.
+    metadata.retain(|key, _| !rustfs_utils::http::is_internal_key(key));
+
    opts.user_metadata = metadata;
    opts
 }
@@ -2244,6 +2249,38 @@ mod tests {
        }
    }

+    #[test]
+    fn build_transition_put_options_drops_reserved_internal_headers() {
+        let transaction_suffix = rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID;
+        let mut metadata = HashMap::new();
+        for suffix in [
+            rustfs_utils::http::SUFFIX_INLINE_DATA,
+            rustfs_utils::http::SUFFIX_ACTUAL_SIZE,
+            rustfs_utils::http::SUFFIX_PART_CHECKSUMS,
+            transaction_suffix,
+        ] {
+            rustfs_utils::http::metadata_compat::insert_str(&mut metadata, suffix, "x".to_string());
+        }
+        metadata.insert("X-Minio-Internal-Compression".to_string(), "klauspost/compress/s2".to_string());
+        metadata.insert("x-amz-meta-foo".to_string(), "bar".to_string());
+        metadata.insert("name".to_string(), "object".to_string());
+
+        let opts = build_transition_put_options("COLD".to_string(), metadata);
+
+        let leaked: Vec<_> = opts
+            .user_metadata
+            .keys()
+            .filter(|key| rustfs_utils::http::is_internal_key(key))
+            .collect();
+        assert!(leaked.is_empty(), "reserved internal keys must not reach the remote: {leaked:?}");
+        assert_eq!(opts.user_metadata.get("x-amz-meta-foo").map(String::as_str), Some("bar"));
+        assert_eq!(opts.user_metadata.get("name").map(String::as_str), Some("object"));
+        assert!(opts.user_metadata.contains_key(&format!(
+            "x-amz-meta-{}",
+            rustfs_utils::http::metadata_compat::internal_key_rustfs(transaction_suffix)
+        )));
+    }
+
    #[test]
    fn build_transition_put_options_requests_no_checksum_and_content_md5() {
        // Regression for rustfs/rustfs#4811: transition uploads must leave the
--- a/Show More
+++ b/Show More