Merge branch 'main' into overtrue/backlog-1831-heal-dead-variants

perf(storage): converge Wave 2 hot-path optimizations (#6065 )
* perf(get): share inline shards and lock clients Co-Authored-By: heihutu <heihutu@gmail.com> * perf(ecstore): converge PUT encoding on contiguous blocks Co-Authored-By: heihutu <heihutu@gmail.com> * perf(get): cache codec streaming gate config Co-Authored-By: heihutu <heihutu@gmail.com> * fix(sse): redact projected customer headers Co-Authored-By: heihutu <heihutu@gmail.com> * perf(ecstore): collapse GET metadata snapshots Co-Authored-By: heihutu <heihutu@gmail.com> * perf(ecstore): reuse decode stripe scratch Co-Authored-By: heihutu <heihutu@gmail.com> * refactor(ecstore): trim decode scratch adapters Co-Authored-By: heihutu <heihutu@gmail.com> * test(ecstore): adapt transition checks to metadata snapshots Co-Authored-By: heihutu <heihutu@gmail.com> * perf(get): release metadata snapshots at ownership boundary Co-Authored-By: heihutu <heihutu@gmail.com> * refactor(ecstore): close cumulative fast-path findings Co-Authored-By: heihutu <heihutu@gmail.com> * fix(storage): preserve lock and header invariants Co-Authored-By: heihutu <heihutu@gmail.com> * test(ecstore): adapt cumulative paths after rebase Co-Authored-By: heihutu <heihutu@gmail.com> * fix(rio-v2): adapt generated metadata fixture Co-Authored-By: heihutu <heihutu@gmail.com> --------- Co-authored-by: heihutu <heihutu@gmail.com>
2026-08-13 16:46:55 +00:00 · 2026-08-13 17:06:47 +08:00 · 2026-08-13 16:34:28 +08:00 · 2026-08-13 15:28:29 +08:00 · 2026-08-13 15:00:44 +08:00 · 2026-08-13 06:12:26 +00:00
272 changed files with 38508 additions and 8983 deletions
@@ -34,7 +34,8 @@ e2e-vault = { max-threads = 1 }

 # Reliability / fault-injection e2e tests each spawn a single-node 4-disk RustFS
 # server and manipulate its disk directories at runtime (crates/e2e_test:
-# reliability_disk_fault_test, degraded_read_eof_regression_test / dist-13). They
+# reliability_disk_fault_test, degraded_read_eof_regression_test / dist-13, and
+# replacement_privileged_e2e_test when explicitly run as root on Linux). They
 # are correct in isolation but resource-heavy; serialize them under nextest's
 # process boundary (serial_test's #[serial] does not cross it) so several 4-disk
 # servers never run at once. ci-7's nightly picks these up via the e2e suite;
@@ -90,7 +91,7 @@ test-group = 'ecstore-serial-flaky'
 # e2e-reliability test-group note above). The matching ci-profile override is at
 # the end of the file, after [profile.ci] is declared.
 [[profile.default.overrides]]
-filter = 'package(e2e_test) & test(/^(reliability_disk_fault|degraded_read_eof_regression)_test::/)'
+filter = 'package(e2e_test) & test(/^(reliability_disk_fault|degraded_read_eof_regression|replacement_privileged_e2e)_test::/)'
 test-group = 'e2e-reliability'

 [[profile.default.overrides]]
@@ -155,7 +156,7 @@ retries = 2
 # quarantine: no retries, just single-threaded so several 4-disk servers never
 # run concurrently when ci-7's nightly runs the full e2e suite.
 [[profile.ci.overrides]]
-filter = 'package(e2e_test) & test(/^(reliability_disk_fault|degraded_read_eof_regression)_test::/)'
+filter = 'package(e2e_test) & test(/^(reliability_disk_fault|degraded_read_eof_regression|replacement_privileged_e2e)_test::/)'
 test-group = 'e2e-reliability'

 # Serialize the multipart crash-consistency scenarios under the ci profile too
@@ -383,7 +384,7 @@ path = "junit.xml"
 # quarantine: no retries, just single-threaded so several 4-disk servers never
 # run concurrently.
 [[profile.e2e-full.overrides]]
-filter = 'package(e2e_test) & test(/^(reliability_disk_fault|degraded_read_eof_regression)_test::/)'
+filter = 'package(e2e_test) & test(/^(reliability_disk_fault|degraded_read_eof_regression|replacement_privileged_e2e)_test::/)'
 test-group = 'e2e-reliability'

 [[profile.e2e-full.overrides]]
@@ -17,9 +17,11 @@
 # =============================================================================
 #
 # Metric source: the KMS operation-policy choke point in
-# crates/kms/src/policy.rs. All label values are bounded static strings
-# (operation, op_class, outcome, error_class, backend, scope); key identifiers,
-# key material, and tokens never appear in labels.
+# crates/kms/src/policy.rs, except KmsKeyRotationOverdue, which reads the
+# label-less key-lifecycle gauge published by the deletion worker's sweep
+# (crates/kms/src/deletion_worker.rs). All label values are bounded static
+# strings (operation, op_class, outcome, error_class, backend, scope); key
+# identifiers, key material, and tokens never appear in labels.
 #
 # Response procedures: docs/operations/kms-observability-runbook.md
 #
@@ -212,3 +214,38 @@ groups:
            circuit_open until the half-open probe succeeds or returns
            a non-retryable failure.
          runbook_url: "https://github.com/rustfs/rustfs/blob/main/docs/operations/kms-observability-runbook.md#kmsbackendcircuitopen"
+
+      # ------------------------------------------------------------------
+      # 7. KmsKeyRotationOverdue
+      #    The least recently rotated usable key has gone more than 400
+      #    days without a rotation (measured from creation for keys with
+      #    no recorded rotation). Direct gauge state published by the
+      #    deletion worker's sweep, so no traffic guard applies; the
+      #    one-hour hold only bridges scrape gaps. The worker runs only
+      #    on backends with the schedule_deletion capability, so on the
+      #    Static backend the series never exists and this alert cannot
+      #    fire — that backend cannot rotate either; see the rotation
+      #    driver matrix in docs/operations/kms-backend-security.md.
+      #    Threshold: 400 days — conservative default sitting above a
+      #    one-year rotation policy. Align it with the rotation period
+      #    your compliance policy requires, and with
+      #    RUSTFS_KMS_ROTATION_MAX_AGE_SECS so the per-key rotation_due
+      #    verdict and this aggregate alert agree.
+      # ------------------------------------------------------------------
+      - alert: KmsKeyRotationOverdue
+        expr: |
+          rustfs_kms_oldest_key_rotation_age_seconds > (400 * 86400)
+        for: 1h
+        labels:
+          severity: warning
+          component: kms
+        annotations:
+          summary: "Oldest KMS key unrotated for more than 400 days"
+          description: >-
+            The least recently rotated usable KMS key was last rotated
+            {{ $value | humanizeDuration }} ago (measured from creation
+            for keys with no recorded rotation). List keys through the
+            admin API and read rotation_due / rotation_due_reason for
+            the per-key verdict; an "unsupported" reason means the
+            backend cannot rotate at all.
+          runbook_url: "https://github.com/rustfs/rustfs/blob/main/docs/operations/kms-observability-runbook.md#kmskeyrotationoverdue"
@@ -85,7 +85,7 @@ runs:
        repo-token: ${{ github.token }}

    - name: Install flatc
-      uses: Nugine/setup-flatc@e7855e994773ce90094a3f1626d4afc9080c23ae # v1
+      uses: Nugine/setup-flatc@698800de72a96bfb22cf60431dc21a2ff9a7e07b # v1
      with:
        version: "25.12.19"

@@ -55,3 +55,142 @@ jobs:

      - name: Build RustFS
        run: cargo build --release --locked --target x86_64-unknown-linux-gnu -p rustfs --bins
+
+  # Live-Vault lane for the rustfs-kms suite (rustfs/backlog#1774).
+  #
+  # RUSTFS_KMS_VAULT_TOKEN is the single switch that adds the Vault KV2 and
+  # Vault Transit backends to every for_each_backend spec in
+  # crates/kms/tests/behavior_*.rs (see crates/kms/AGENTS.md). rotate and
+  # versioning are advertised only by the Vault backends, so without this lane
+  # no CI run ever asserts the working half of behavior_rotation.rs — a
+  # rotation that silently dropped historical key versions would stay green.
+  # The same lane runs the dev-Vault #[ignore] tests and the two self-hosting
+  # live scripts (AppRole login, three-node Raft leader failover).
+  #
+  # GitHub-hosted ubuntu-latest, deliberately not the self-hosted sm-standard
+  # fleet: the HA failover script needs a working Docker daemon, and the
+  # self-hosted fleet is heterogeneous — a docker-dependent workflow has been
+  # burned by it before (see the banner in e2e-s3tests.yml, rustfs/backlog#1149).
+  kms-vault-lane:
+    name: KMS live Vault lane
+    runs-on: ubuntu-latest
+    timeout-minutes: 90
+    env:
+      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
+      # Root token of the ephemeral loopback dev server. Not a secret: the
+      # server lives only for this job, listens on 127.0.0.1, and holds only
+      # keys the tests create. The literal value matters — the dev-Vault
+      # #[ignore] fixtures in crates/kms/src/backends/vault.rs hardcode it.
+      VAULT_LANE_TOKEN: dev-only-token
+      VAULT_LANE_ADDR: http://127.0.0.1:8200
+      # Keeps a runner-level proxy from swallowing the loopback dev-server
+      # traffic (see crates/kms/AGENTS.md). Actions env keys are
+      # case-insensitive, so only the uppercase form is set; reqwest reads
+      # either casing.
+      NO_PROXY: 127.0.0.1,localhost
+    steps:
+      - name: Checkout main branch
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+          ref: main
+
+      - name: Setup Rust environment
+        uses: ./.github/actions/setup
+        with:
+          # Dedicated key: rust-cache cannot tell runner images apart, so
+          # sharing a key with an sm-standard lane would let two different
+          # system images overwrite each other's artifacts (same reasoning as
+          # ci.yml's ci-uring lane). Saved from this nightly job itself so the
+          # next night starts warm.
+          cache-shared-key: kms-vault-lane
+          cache-save-if: 'true'
+          install-build-packaging-tools: 'false'
+          install-test-tools: 'false'
+
+      - name: Install Vault CLI
+        run: |
+          set -euo pipefail
+          wget -qO- https://apt.releases.hashicorp.com/gpg | sudo gpg --dearmor -o /usr/share/keyrings/hashicorp-archive-keyring.gpg
+          echo "deb [signed-by=/usr/share/keyrings/hashicorp-archive-keyring.gpg] https://apt.releases.hashicorp.com $(lsb_release -cs) main" | sudo tee /etc/apt/sources.list.d/hashicorp.list >/dev/null
+          sudo apt-get update -qq
+          sudo apt-get install -y -qq vault
+          vault version
+
+      - name: Start Vault dev server with KV2 and Transit engines
+        run: |
+          set -euo pipefail
+          nohup vault server -dev \
+            -dev-root-token-id="${VAULT_LANE_TOKEN}" \
+            -dev-listen-address=127.0.0.1:8200 >/tmp/vault-dev.log 2>&1 &
+          for _ in $(seq 1 60); do
+            if curl -fsS "${VAULT_LANE_ADDR}/v1/sys/health" >/dev/null 2>&1; then
+              break
+            fi
+            sleep 1
+          done
+          curl -fsS "${VAULT_LANE_ADDR}/v1/sys/health"
+          export VAULT_ADDR="${VAULT_LANE_ADDR}" VAULT_TOKEN="${VAULT_LANE_TOKEN}"
+          # Dev mode mounts KV v2 at secret/ by default; Transit is explicit.
+          # Prove both engines actually work rather than assuming the defaults.
+          vault secrets enable transit
+          vault kv put secret/rustfs-ci-lane-probe value=ok >/dev/null
+          vault kv get secret/rustfs-ci-lane-probe >/dev/null
+          vault write -f transit/keys/rustfs-ci-lane-probe >/dev/null
+
+      - name: Run rustfs-kms suite with the Vault lane on
+        env:
+          RUSTFS_KMS_VAULT_TOKEN: ${{ env.VAULT_LANE_TOKEN }}
+          RUSTFS_KMS_VAULT_ADDR: ${{ env.VAULT_LANE_ADDR }}
+        run: cargo test -p rustfs-kms --locked
+
+      - name: Run dev-Vault ignored tests
+        env:
+          RUSTFS_KMS_VAULT_TOKEN: ${{ env.VAULT_LANE_TOKEN }}
+          RUSTFS_KMS_VAULT_ADDR: ${{ env.VAULT_LANE_ADDR }}
+        # Filters select the dev-Vault-only #[ignore] tests. The AWS #[ignore]
+        # tests (backends::aws, service_manager) stay excluded — they need real
+        # AWS credentials and create billable keys. The AppRole and HA #[ignore]
+        # tests are excluded here because their own scripts below provision the
+        # Vault topology they need.
+        run: |
+          set -euo pipefail
+          cargo test -p rustfs-kms --locked --lib backends::contract_tests -- --ignored
+          cargo test -p rustfs-kms --locked --lib backends::vault -- --ignored
+          cargo test -p rustfs-kms --locked --test vault_fault_injection -- --ignored
+
+      - name: Run AppRole live checks (self-hosting ephemeral Vault)
+        run: bash scripts/test/vault_approle_kms_live.sh
+
+      - name: Show Vault dev server log on failure
+        if: failure()
+        run: tail -n 200 /tmp/vault-dev.log || true
+
+  # Three-node Raft leader failover (crates/kms/tests/vault_ha_failover_live.rs,
+  # first validated by rustfs/rustfs#5653). Its own job so an election-timing
+  # flake cannot mask the main lane's verdict, and vice versa. The script
+  # provisions and tears down its own Docker cluster.
+  kms-vault-ha-failover:
+    name: KMS Vault HA failover lane
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    env:
+      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
+      NO_PROXY: 127.0.0.1,localhost
+    steps:
+      - name: Checkout main branch
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+          ref: main
+
+      - name: Setup Rust environment
+        uses: ./.github/actions/setup
+        with:
+          cache-shared-key: kms-vault-lane
+          cache-save-if: 'false'
+          install-build-packaging-tools: 'false'
+          install-test-tools: 'false'
+
+      - name: Run HA leader failover live checks (three-node Raft cluster in Docker)
+        run: bash scripts/test/vault_ha_kms_live.sh
@@ -1,6 +1,6 @@
 # ARCHITECTURE.md

-> Last updated: 2026-07-02 · Revision: 2
+> Last updated: 2026-08-12 · Revision: 3
 >
 > This document describes the high-level architecture of RustFS.
 > If you want to familiarize yourself with the code base, you are in the right place!
@@ -119,19 +119,44 @@ module split is tracked under `docs/architecture/`.

 3. **Each type has exactly one definition.** Types shared across crates must be defined
   in one crate and re-exported or imported by others.
-   - ⚠️ VIOLATED: `ReplicationStats` (4 copies), `LastMinuteLatency` (3 copies),
-     `BackpressureConfig` (3 copies), `DataUsageInfo` (2 copies).
+   - ⚠️ VIOLATED: `ReplicationStats` names three unrelated types
+     (`crates/data-usage/src/data_usage.rs`,
+     `crates/obs/src/metrics/collectors/replication.rs`,
+     `crates/ecstore/src/bucket/replication/replication_state.rs`) — a naming
+     collision, not copies; renaming is tracked in rustfs/backlog#1847.
+   - `LastMinuteLatency` has two deliberately different implementations: the
+     per-second bucketed accumulator in `crates/common/src/last_minute.rs` and
+     the in-memory endpoint-health sample tracker in
+     `crates/ecstore/src/bucket/bucket_target_sys.rs` (its doc comment explains
+     why it stays local).
+   - ✅ RESOLVED: `BackpressureConfig` and `DataUsageInfo` each have exactly one
+     definition (`crates/io-core/src/backpressure.rs`,
+     `crates/data-usage/src/data_usage.rs`). The zero-consumer
+     `BackpressureSettings` copy that lingered in io-metrics was removed
+     (rustfs/backlog#1833).

 4. **ecstore does not know about HTTP or S3 protocol details.** It operates on
   storage-level abstractions (objects, buckets, disks, pools).
+   - ⚠️ VIOLATED: 58 files under `crates/ecstore/src` reference `s3s`
+     (`rg -l 's3s' crates/ecstore/src | wc -l`), `crates/ecstore/src/client/`
+     is a ~9.4K-line embedded S3 HTTP client, and `crates/ecstore/Cargo.toml`
+     depends on `s3s`, `http`, `hyper`/`hyper-util`/`hyper-rustls`, and
+     `reqwest`. Target state: the engine's need to act as an S3 client
+     (tiering, replication targets) is served by an extracted client crate,
+     and ecstore holds no wire or DTO types.

 5. **The `rustfs` binary crate is the only place that wires everything together.**
   Individual crates should be testable in isolation.

 6. **Error types use `thiserror` with descriptive names** (e.g., `StorageError`,
   not bare `Error`).
-   - ⚠️ VIOLATED: 6 crates use `pub enum Error`; 2 crates use `snafu`;
-      `heal` use `anyhow` in library code.
+   - ✅ RESOLVED (strategy): `snafu` is gone from source
+     (`rg -l snafu crates/ rustfs/` is empty) and library code no longer uses
+     `anyhow` (remaining hits are test code and the `e2e_test` crate; `heal`
+     uses `thiserror`).
+   - ⚠️ VIOLATED (naming): 6 crates still export a bare `pub enum Error`:
+     `crypto`, `filemeta`, `heal`, `iam`, `policy`, and `replication`
+     (`src/resync.rs`) — all `thiserror`-derived.

 ## Known Structural Issues

@@ -140,13 +165,25 @@ module split is tracked under `docs/architecture/`.

 ### Critical

- **common/scanner code duplication (~3K lines).** `scanner` depends on `common`
-  but maintains its own copies of `DataUsageInfo`, `LastMinuteLatency`, and related
-  types instead of importing them.
+- **scanner/data-usage duplicate `.usage-cache.bin` serialization types.** The
+  original finding ("common/scanner code duplication, ~3K lines") is resolved:
+  `scanner` imports the shared data-usage types from `rustfs-data-usage` (see
+  the `pub use rustfs_data_usage::…` re-exports at the top of
+  `crates/scanner/src/data_usage_define.rs`). What remains: `scanner` and
+  `data-usage` each hold their own serialization types for the scanner cache
+  file (`DataUsageCacheInfo`/`DataUsageEntryInfo` in
+  `crates/scanner/src/data_usage_define.rs` vs
+  `DataUsageCacheInfo`/`DataUsageEntry` in
+  `crates/data-usage/src/data_usage.rs`); convergence is tracked in
+  rustfs/backlog#1828.

- **ecstore is a monolith (87K lines, 163 files).** It contains disk management,
-  bucket management, erasure coding, replication, lifecycle, RPC, and configuration
-  — all in one crate. It should be decomposed along its existing subdirectories.
+- **ecstore is a monolith (265 files, ~288K lines — roughly half is inline
+  `#[cfg(test)]` code).** Measured with
+  `find crates/ecstore/src -name '*.rs' | xargs wc -l`. It contains disk
+  management, bucket management, erasure coding, replication, lifecycle, RPC,
+  and configuration — all in one crate. It should be decomposed along its
+  existing subdirectories; the split plan lives in
+  [docs/architecture/ecstore-module-split-plan.md](docs/architecture/ecstore-module-split-plan.md).

 ### High

@@ -154,19 +191,26 @@ module split is tracked under `docs/architecture/`.
  `common → filemeta/madmin` edges must stay removed so leaf/helper crates do
  not regain upward dependencies.

- **Three-layer BackpressureConfig/DeadlockConfig duplication** across io-core,
-  concurrency, and `rustfs/src/storage`. Storage policies now expose and consume
-  explicit projections into the concurrency/io-core policy shapes, and workload
+- **Three-layer backpressure/deadlock policy bridging** across io-core,
+  concurrency, and `rustfs/src/storage`. The config types are no longer
+  duplicated (`BackpressureConfig` and `DeadlockDetectorConfig` are each
+  defined once, in io-core). Storage policies expose and consume explicit
+  projections into the concurrency/io-core policy shapes, and workload
  admission snapshots are composed through provider registries; later work
  should use those bridges before deleting compatibility wrappers.

 ### Medium

- **Inconsistent error handling.** Three strategies (thiserror/snafu/anyhow) and
-  mixed naming (bare `Error` vs descriptive names).
+- **Bare `Error` naming.** Error-handling strategy has converged on `thiserror`
+  (no `snafu`, no `anyhow` in library code); the remaining inconsistency is the
+  bare `pub enum Error` naming in the 6 crates listed under Invariant 6.

- **Ambiguous common vs utils boundary.** Both described as "utilities and data
-  structures." Need clear ownership rules.
+- **`common` is mostly parked domain code, not shared utilities.** Of its
+  6,724 lines, ~83% is scanner/heal domain code stranded there to break
+  dependency cycles (`metrics.rs`, ~4,810 lines of scanner-domain metrics;
+  `heal_channel.rs`, ~776 lines of heal-domain channel types). The
+  "common vs utils" naming ambiguity is secondary to moving that code to its
+  domain owners.

 ## Cross-Cutting Concerns

@@ -232,7 +276,7 @@ The binary (`main.rs`) boots in this order:

 ```
                            ┌─────────┐
-                            │  rustfs │  (binary + lib, 75K lines)
+                            │  rustfs │  (binary + lib)
                            │  main   │
                            └────┬────┘
                                 │
@@ -255,7 +299,7 @@ The binary (`main.rs`) boots in this order:
              │                  │                  │
        ┌─────▼──────┐    ┌──────▼──────┐    ┌──────▼──────┐
        │  ecstore   │    │     rio     │    │   io-core   │
-        │ (87K,core) │    │  (readers)  │    │ (zero-copy) │
+        │   (core)   │    │  (readers)  │    │ (zero-copy) │
        └─────┬──────┘    └─────────────┘    └─────────────┘
              │
     ┌─────┬──┼──┬─────┬──────┐
@@ -104,6 +104,12 @@ dependencies = [
 "memchr",
 ]

+[[package]]
+name = "aliasable"
+version = "0.1.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "250f629c0161ad8107cf89319e990051fae62832fd343083bea452d93e2205fd"
+
 [[package]]
 name = "aligned-vec"
 version = "0.6.4"
@@ -266,24 +272,24 @@ checksum = "330a5ed07fa54e4702c9d6c4174f74427fc0ef6e214bbd677ae50a5099946470"

 [[package]]
 name = "apache-avro"
-version = "0.21.0"
+version = "0.22.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "36fa98bc79671c7981272d91a8753a928ff6a1cd8e4f20a44c45bd5d313840bf"
+checksum = "312c1ea69e5fe9966e0029fb95aca8790100b85aff4f0d3b00a9337c74069a9c"
 dependencies = [
 "bigdecimal",
 "bon",
- "digest 0.10.7",
+ "digest 0.11.3",
 "log",
- "miniz_oxide",
+ "miniz_oxide 0.9.1",
 "num-bigint 0.4.8",
+ "ouroboros",
 "quad-rand",
- "rand 0.9.5",
+ "rand 0.10.2",
 "regex-lite",
 "serde",
 "serde_bytes",
 "serde_json",
- "strum 0.27.2",
- "strum_macros 0.27.2",
+ "strum",
 "thiserror 2.0.20",
 "uuid",
 ]
@@ -1156,9 +1162,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-eventstream"
-version = "0.61.1"
+version = "0.61.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5a9381123ab62d20c13082b151f30f962a3b112b727345394536dfa39a482944"
+checksum = "6de526c7b567420a31bc283657a7921b45c4cafe0827fdf2490713dcc770c28f"
 dependencies = [
 "aws-smithy-types",
 "bytes",
@@ -1189,9 +1195,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-http-client"
-version = "1.2.0"
+version = "1.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "635d23afda0a6ab48d666c4d447c4873e8d1e83518a2be2093122397e50b838e"
+checksum = "3c1c8a04cb31ba74d0115af5a890bb8c0d48fba64b52812fa13929a6ef0cc83c"
 dependencies = [
 "aws-smithy-async",
 "aws-smithy-protocol-test",
@@ -1271,9 +1277,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-runtime"
-version = "1.12.1"
+version = "1.13.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "07505b34e8f4b3591a4fa69e9792b52289b95488dbbc68c3c0075b7bedb245e1"
+checksum = "483b858ff67522011c4786310c5cd8fd88d0be7ea3d5f1a48328446300c4269e"
 dependencies = [
 "aws-smithy-async",
 "aws-smithy-http",
@@ -1337,9 +1343,9 @@ dependencies = [

 [[package]]
 name = "aws-smithy-types"
-version = "1.6.1"
+version = "1.6.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d6dc683efb34b9e755675b37fedbe0103141e5b6df7bdc9eb6967756a8c167d8"
+checksum = "fce83ce9abbb198d25bc7131e468d0f9fe1257125e58c39f3f9fc9f5098c9647"
 dependencies = [
 "base64-simd",
 "bytes",
@@ -1458,7 +1464,7 @@ dependencies = [
 "addr2line",
 "cfg-if",
 "libc",
- "miniz_oxide",
+ "miniz_oxide 0.8.9",
 "object 0.37.3",
 "rustc-demangle",
 "windows-link",
@@ -1801,6 +1807,15 @@ version = "0.3.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "37b2a672a2cb129a2e41c10b1224bb368f9f37a2b16b612598138befd7b37eb5"

+[[package]]
+name = "castaway"
+version = "0.2.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "dec551ab6e7578819132c713a93c022a05d60159dc86e7a7050223577484c55a"
+dependencies = [
+ "rustversion",
+]
+
 [[package]]
 name = "cbc"
 version = "0.1.2"
@@ -1994,7 +2009,7 @@ version = "4.6.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "d012d2b9d65aca7f18f4d9878a045bc17899bba951561ba5ec3c2ba1eed9a061"
 dependencies = [
- "heck",
+ "heck 0.5.0",
 "proc-macro2",
 "quote",
 "syn 3.0.3",
@@ -2063,6 +2078,19 @@ dependencies = [
 "unicode-width 0.2.2",
 ]

+[[package]]
+name = "compact_str"
+version = "0.10.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "79fcda08c33bb58b97008b2cdada6622500e949e060f5913361763121abd2416"
+dependencies = [
+ "castaway",
+ "cfg-if",
+ "itoa",
+ "static_assertions",
+ "zmij",
+]
+
 [[package]]
 name = "compression-codecs"
 version = "0.4.38"
@@ -4162,7 +4190,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "843fba2746e448b37e26a819579957415c8cef339bf08564fe8b7ddbd959573c"
 dependencies = [
 "crc32fast",
- "miniz_oxide",
+ "miniz_oxide 0.8.9",
 "zlib-rs",
 ]

@@ -4226,9 +4254,9 @@ checksum = "42703706b716c37f96a77aea830392ad231f44c9e9a67872fa5548707e11b11c"

 [[package]]
 name = "futures"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a88cf1f829d945f548cf8fec32c61b1f202b6d93b45848602fc02af4b12ad218"
+checksum = "9a31d2a3fbaaeb2af2368bbdd904aa8e812d3c04a1ee10d3171f52d556e5d0a3"
 dependencies = [
 "futures-channel",
 "futures-core",
@@ -4241,9 +4269,9 @@ dependencies = [

 [[package]]
 name = "futures-channel"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "262590f4fe6afeb0bc83be1daa64e52657fe185690a958af7f3ad0e92085c5ae"
+checksum = "b1f9e3d69d39e4862ffed03ed071a76f9a13ba1d9109d355b0f0aa6b15e393c4"
 dependencies = [
 "futures-core",
 "futures-sink",
@@ -4251,15 +4279,15 @@ dependencies = [

 [[package]]
 name = "futures-core"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2cd50c473c80f6d7c3670a752354b8e569b1a7cbfdc0419ec88e5edad85e0dc7"
+checksum = "92d699e522242e69e3003b94ecc1f960f3a5e015aa7c5d7486e65ad01dd94f5e"

 [[package]]
 name = "futures-executor"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6754879cc9f2c66f88c6e5c35344bb0bdb0708b0352b1201815667c7eabc7458"
+checksum = "031b47cf1a3c6cc8bc2fc76cd437f521619387907d469316e7c0bc278f1f5432"
 dependencies = [
 "futures-core",
 "futures-task",
@@ -4268,9 +4296,9 @@ dependencies = [

 [[package]]
 name = "futures-io"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4577ecaa3c4f96589d473f679a71b596316f6641bc350038b962a5daf0085d7a"
+checksum = "53c0fa8157de1303bfffdaa1cc2a673bfffb60102f76b0ef4441659124373fed"

 [[package]]
 name = "futures-lite"
@@ -4287,13 +4315,13 @@ dependencies = [

 [[package]]
 name = "futures-macro"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2d6d3cde68c518367be28956066ddfef33813991b77a55005a69dae04bf3b10b"
+checksum = "9fb9654ba8355388abeb8dcb4fc62f511300867002afc858860463bdd9fe0c44"
 dependencies = [
 "proc-macro2",
 "quote",
- "syn 2.0.119",
+ "syn 3.0.3",
 ]

 [[package]]
@@ -4309,21 +4337,21 @@ dependencies = [

 [[package]]
 name = "futures-sink"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e34418ac499d6305c2fb5ad0ed2f6ac998c5f8ca209b4510f7f94242c647e307"
+checksum = "1944426bf7d03f1d14f708785e4b33efd750b36d48a157b836b3efc15ede8e1d"

 [[package]]
 name = "futures-task"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b231ed28831efb4a61a08580c4bc233ec56bc009f4cd8f52da2c3cb97df0c109"
+checksum = "cd417de3d1d015fc3bfd2b1ea46dfc7bab72ef86f1cc7cc9c78e728b34a6d1fd"

 [[package]]
 name = "futures-util"
-version = "0.3.33"
+version = "0.3.34"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a77a90a256fce34da66415271e30f94ee91c57b04b8a2c042d9cf3220179deaa"
+checksum = "0d50a92467f8ba5dd6e3ee5d4bd04d73ab2e4e1c44474a0674821dfce14b79bc"
 dependencies = [
 "futures-channel",
 "futures-core",
@@ -4832,6 +4860,12 @@ dependencies = [
 "stable_deref_trait",
 ]

+[[package]]
+name = "heck"
+version = "0.4.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "95505c38b4572b2d910cecb0281560f54b440a19336cbbcb27bf6ce6adc6f5a8"
+
 [[package]]
 name = "heck"
 version = "0.5.0"
@@ -4991,9 +5025,9 @@ dependencies = [

 [[package]]
 name = "hotpath"
-version = "0.23.1"
+version = "0.23.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "be80823867e0c9820c9237c38b21f9f4aa1ebb0db1f98ff25ac0b1d2c088a470"
+checksum = "62e810bedda5a467ef5c9b5c8a20763fefebc89b63ef36f7ee44a143085204a2"
 dependencies = [
 "arc-swap",
 "async-channel",
@@ -5025,9 +5059,9 @@ dependencies = [

 [[package]]
 name = "hotpath-macros"
-version = "0.23.1"
+version = "0.23.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "61d1fb3ee80ae7b4743d29487665766ce5a1442e959521790e86317f89dcd5a3"
+checksum = "01bdc59bfc1a9984bee2ff5da63b2f6fccbaa57cd9a4119d709524632bddf341"
 dependencies = [
 "proc-macro2",
 "quote",
@@ -5036,15 +5070,15 @@ dependencies = [

 [[package]]
 name = "hotpath-macros-meta"
-version = "0.23.1"
+version = "0.23.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "feede71fa226b0b5d523e58e7b0a1462935c0b8a00584a6669f45d564086209d"
+checksum = "d9216e8a01abe1e1671c376dc8736fb1bf772d7a889538d25f9e1200120ced38"

 [[package]]
 name = "hotpath-meta"
-version = "0.23.1"
+version = "0.23.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "424fe0a13105d3731f65237785f5b95c3e4b8bfae4a039d932f56192cd74afc0"
+checksum = "f22a9d20435fb79511b19dae37b3607224cd98f342a410702d84657cc38fc72f"
 dependencies = [
 "hotpath-macros-meta",
 ]
@@ -5099,9 +5133,9 @@ dependencies = [

 [[package]]
 name = "http-body-util"
-version = "0.1.4"
+version = "0.1.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e9f41fd6a08e4d4ec69df65976da761afd5ad5e58a9d4acb46bd1c953a9e3ff2"
+checksum = "23169fe34a5fbcdd3f3862e78fb9b6fccd5f02a6dc6f732547005d45631ce71c"
 dependencies = [
 "bytes",
 "futures-core",
@@ -5420,9 +5454,9 @@ dependencies = [

 [[package]]
 name = "io-uring"
-version = "0.7.13"
+version = "0.7.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9080b15e63775b9a2ac7dca720f7050a8b955e092ea0f6020a4a80f69998cdc0"
+checksum = "d64d8ca234d152948ceaede1f419b6a83983a5ecccaac05fb337a809c96d3aa6"
 dependencies = [
 "bitflags 2.13.1",
 "cfg-if",
@@ -5896,18 +5930,18 @@ dependencies = [

 [[package]]
 name = "liblzma"
-version = "0.4.7"
+version = "0.4.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "45aec2360b3933207e27908049d8e4df4e476b58180afb1e56b2a4fb72efe4ba"
+checksum = "2fe0a34ca854fd4f20c07f696fc8675aec78f87d88d29f5e10257a7490a1b2e1"
 dependencies = [
 "liblzma-sys",
 ]

 [[package]]
 name = "liblzma-sys"
-version = "0.4.7"
+version = "0.4.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a046c7f353ba30f810545151e04f63545833803f5b86ee3ddf1517247fe560a5"
+checksum = "a0dad045e4b1b7b170be4b60b54b780cafb4490165461bac7d1cf7b703f61d5f"
 dependencies = [
 "cc",
 "libc",
@@ -5923,7 +5957,7 @@ checksum = "b6d2cec3eae94f9f509c767b45932f1ada8350c4bdb85af2fcab4a3c14807981"
 [[package]]
 name = "libmimalloc-sys"
 version = "0.1.49"
-source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=ce6338661179c8be22e516b00af7483f151485a7#ce6338661179c8be22e516b00af7483f151485a7"
+source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11#6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11"
 dependencies = [
 "cc",
 "cty",
@@ -6225,9 +6259,9 @@ dependencies = [

 [[package]]
 name = "metrique"
-version = "0.1.29"
+version = "0.1.30"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d2e394c63e2d1a30aeb3b9392ecf3439d8475d2df810a8f4f6e66d6866754017"
+checksum = "dedbf06ffeef4c37990c73636fbd993aa34fb1948afd736e6114f239220993db"
 dependencies = [
 "itoa",
 "jiff",
@@ -6255,9 +6289,9 @@ dependencies = [

 [[package]]
 name = "metrique-macro"
-version = "0.1.20"
+version = "0.1.21"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "786df1fd0abebd0db685f7e9a353c78756d4b370fb98a52376c2015fa55f141f"
+checksum = "f4fb1f30185f53f7f6e4c9e46745c1a1350af8e77fda5a88aded44b0637a82e0"
 dependencies = [
 "Inflector",
 "darling 0.23.0",
@@ -6284,9 +6318,9 @@ checksum = "2faca4e4480069ff02b1763b3b79f5cec7e8628e24d9dc5b6073f53d2577a4d9"

 [[package]]
 name = "metrique-writer"
-version = "0.1.25"
+version = "0.1.26"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "82cdde44d241dab7fc8b7a32e0eb5dae6cd28f8de80b59f9a1e9f2f0b05e485e"
+checksum = "20bd17c1a3ca2719e31f19ce77a853948dc2102f35976b92276c42a64fdc5f3f"
 dependencies = [
 "ahash",
 "crossbeam-queue",
@@ -6305,9 +6339,9 @@ dependencies = [

 [[package]]
 name = "metrique-writer-core"
-version = "0.1.19"
+version = "0.1.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e57379b7ee2272efaeaaa6de062503563e57333b24aadc7f2255b3d602899e8b"
+checksum = "f1a55b6aae1d85c557c729564c4e2b32a26dc65ba2d90d9647ca01f2bd4854c4"
 dependencies = [
 "derive-where",
 "itertools 0.14.0",
@@ -6332,7 +6366,7 @@ dependencies = [
 [[package]]
 name = "mimalloc"
 version = "0.1.52"
-source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=ce6338661179c8be22e516b00af7483f151485a7#ce6338661179c8be22e516b00af7483f151485a7"
+source = "git+https://github.com/xonatius/mimalloc_rust.git?rev=6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11#6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11"
 dependencies = [
 "libmimalloc-sys",
 ]
@@ -6369,6 +6403,15 @@ dependencies = [
 "simd-adler32",
 ]

+[[package]]
+name = "miniz_oxide"
+version = "0.9.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "b63fbc4a50860e98e7b2aa7804ded1db5cbc3aff9193adaff57a6931bf7c4b4c"
+dependencies = [
+ "adler2",
+]
+
 [[package]]
 name = "minlz"
 version = "1.2.3"
@@ -6416,9 +6459,9 @@ dependencies = [

 [[package]]
 name = "moka"
-version = "0.12.15"
+version = "0.12.16"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "957228ad12042ee839f93c8f257b62b4c0ab5eaae1d4fa60de53b27c9d7c5046"
+checksum = "4293f18e7567a1caf3c584855554377025c65e0aa445344d04171f5ad63d19b9"
 dependencies = [
 "async-lock",
 "crossbeam-channel",
@@ -6463,7 +6506,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a4db8a44120571277accfaa3f3d91e7d3989d601d817c2fc01a9391b86135666"
 dependencies = [
 "darling 0.23.0",
- "heck",
+ "heck 0.5.0",
 "manyhow",
 "num-bigint 0.4.8",
 "proc-macro-crate",
@@ -6747,9 +6790,9 @@ checksum = "521739c6d2bac4aa25192232afe6841231376b2b26d4d9fae5ecf8ca5772e441"

 [[package]]
 name = "num-integer"
-version = "0.1.46"
+version = "0.1.47"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7969661fd2958a5cb096e56c8e1ad0444ac2bbcd0061bd28660485a44879858f"
+checksum = "7ce2d95d4b3734dc35aa2f45e1aa22cd416814592a4f9d9205e11affd5b8e10b"
 dependencies = [
 "num-traits",
 ]
@@ -6839,7 +6882,7 @@ version = "5.0.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "51e219e79014df21a225b1860a479e2dcd7cbd9130f4defd4bd0e191ea31d67d"
 dependencies = [
- "base64 0.22.1",
+ "base64 0.21.7",
 "chrono",
 "getrandom 0.2.17",
 "http 1.5.0",
@@ -7168,6 +7211,30 @@ dependencies = [
 "num-traits",
 ]

+[[package]]
+name = "ouroboros"
+version = "0.18.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "1e0f050db9c44b97a94723127e6be766ac5c340c48f2c4bb3ffa11713744be59"
+dependencies = [
+ "aliasable",
+ "ouroboros_macro",
+ "static_assertions",
+]
+
+[[package]]
+name = "ouroboros_macro"
+version = "0.18.5"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "3c7028bdd3d43083f6d8d4d5187680d0d3560d54df4cc9d752005268b41e64d0"
+dependencies = [
+ "heck 0.4.1",
+ "proc-macro2",
+ "proc-macro2-diagnostics",
+ "quote",
+ "syn 2.0.119",
+]
+
 [[package]]
 name = "outref"
 version = "0.5.2"
@@ -7720,9 +7787,9 @@ dependencies = [

 [[package]]
 name = "portable-atomic"
-version = "1.14.0"
+version = "1.15.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3d20d5497ef88037a52ff98267d066e7f11fcc5e99bbfbd58a42336193aacec3"
+checksum = "05c8b63e8d9609db387f0324918f81d68fe27748f084ef092fb35954d0539a85"

 [[package]]
 name = "portable-atomic-util"
@@ -7903,6 +7970,19 @@ dependencies = [
 "unicode-ident",
 ]

+[[package]]
+name = "proc-macro2-diagnostics"
+version = "0.10.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "af066a9c399a26e020ada66a034357a868728e72cd426f3adcd35f80d88d88c8"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn 2.0.119",
+ "version_check",
+ "yansi",
+]
+
 [[package]]
 name = "prometheus"
 version = "0.14.0"
@@ -7962,8 +8042,8 @@ version = "0.13.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "be769465445e8c1474e9c5dac2018218498557af32d9ed057325ec9a41ae81bf"
 dependencies = [
- "heck",
- "itertools 0.14.0",
+ "heck 0.5.0",
+ "itertools 0.10.5",
 "log",
 "multimap",
 "once_cell",
@@ -7982,8 +8062,8 @@ version = "0.14.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "03da047801ff44bb6a4d407d4860c05fd70bb81714e6b2f3812603d5b145b042"
 dependencies = [
- "heck",
- "itertools 0.14.0",
+ "heck 0.5.0",
+ "itertools 0.10.5",
 "log",
 "multimap",
 "petgraph 0.8.3",
@@ -8004,7 +8084,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8a56d757972c98b346a9b766e3f02746cde6dd1cd1d1d563472929fdd74bec4d"
 dependencies = [
 "anyhow",
- "itertools 0.14.0",
+ "itertools 0.10.5",
 "proc-macro2",
 "quote",
 "syn 2.0.119",
@@ -8017,7 +8097,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "b570b25f7617e43d59005d0990ccb79e950a423952cea19671b7a876da390adf"
 dependencies = [
 "anyhow",
- "itertools 0.14.0",
+ "itertools 0.10.5",
 "proc-macro2",
 "quote",
 "syn 2.0.119",
@@ -8074,9 +8154,9 @@ dependencies = [

 [[package]]
 name = "pulldown-cmark-to-cmark"
-version = "22.0.0"
+version = "22.0.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "50793def1b900256624a709439404384204a5dc3a6ec580281bfaac35e882e90"
+checksum = "ab1ad36992cead65f02aa399a373a42730922f1525d988172634fdefdecb8a60"
 dependencies = [
 "pulldown-cmark",
 ]
@@ -8430,9 +8510,9 @@ dependencies = [

 [[package]]
 name = "rcgen"
-version = "0.14.8"
+version = "0.14.9"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "57f6d249aad744e274e682777a50283a225a32705394ee6d5fcc01efa25e4055"
+checksum = "091e7a8e7d86e6feb87a27ce8e2cba29d49eff9507afeebefab7eeb2ca667fb4"
 dependencies = [
 "aws-lc-rs",
 "pem",
@@ -8837,9 +8917,9 @@ dependencies = [

 [[package]]
 name = "russh"
-version = "0.62.5"
+version = "0.62.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "da7c230e0ed9cbeb92fbad6c8848985d6df2a1464c0dc247a021abd666e9005e"
+checksum = "b41043523e0edcbd4e31d00903e26f12994f63b21bae9904f7405c1ed92752a5"
 dependencies = [
 "aes 0.9.2",
 "aws-lc-rs",
@@ -9121,7 +9201,6 @@ dependencies = [
 "sha2 0.11.0",
 "shadow-rs",
 "socket2",
- "starshard",
 "subtle",
 "sysinfo",
 "temp-env",
@@ -9265,7 +9344,6 @@ dependencies = [
 name = "rustfs-data-usage"
 version = "1.0.0-rc.1"
 dependencies = [
- "async-trait",
 "hotpath",
 "rmp-serde",
 "rustfs-filemeta",
@@ -9542,9 +9620,7 @@ dependencies = [
 "metrics",
 "metrics-util",
 "num_cpus",
- "rustfs-common",
 "rustfs-s3-ops",
- "rustfs-utils",
 "sysinfo",
 "thiserror 2.0.20",
 "tokio",
@@ -9658,6 +9734,7 @@ dependencies = [
 "rustfs-utils",
 "rustify",
 "serde",
+ "serde_ignored",
 "serde_json",
 "sha2 0.11.0",
 "subtle",
@@ -9702,6 +9779,7 @@ name = "rustfs-lock"
 version = "1.0.0-rc.1"
 dependencies = [
 "async-trait",
+ "compact_str",
 "crossbeam-queue",
 "futures",
 "hotpath",
@@ -9712,7 +9790,6 @@ dependencies = [
 "serde",
 "serde_json",
 "smallvec",
- "smartstring",
 "thiserror 2.0.20",
 "tokio",
 "tonic",
@@ -9902,7 +9979,7 @@ dependencies = [
 "rustfs-crypto",
 "serde",
 "serde_json",
- "strum 0.28.0",
+ "strum",
 "temp-env",
 "test-case",
 "thiserror 2.0.20",
@@ -10482,9 +10559,9 @@ dependencies = [

 [[package]]
 name = "rustls-connector"
-version = "0.23.7"
+version = "0.23.8"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "09a5abe04eec18f8b9fbe87885bcaee6426de80bbc579958c0bc064b728ee617"
+checksum = "1babecfcc65b139b812e74bcc7f9ec7b4e00db659fd42d99567b7e77f0c714c6"
 dependencies = [
 "futures-io",
 "futures-rustls",
@@ -10546,9 +10623,9 @@ checksum = "f87165f0995f63a9fbeea62b64d10b4d9d8e78ec6d7d51fb2125fda7bb36788f"

 [[package]]
 name = "rustls-webpki"
-version = "0.103.13"
+version = "0.103.14"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "61c429a8649f110dddef65e2a5ad240f747e85f7758a6bccc7e5777bd33f756e"
+checksum = "0527518605e68109d875e248ea259b6758801cf165e4b2c2733ae3b51f12535a"
 dependencies = [
 "aws-lc-rs",
 "ring",
@@ -10860,6 +10937,16 @@ dependencies = [
 "syn 3.0.3",
 ]

+[[package]]
+name = "serde_ignored"
+version = "0.1.14"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "115dffd5f3853e06e746965a20dcbae6ee747ae30b543d91b0e089668bb07798"
+dependencies = [
+ "serde",
+ "serde_core",
+]
+
 [[package]]
 name = "serde_json"
 version = "1.0.151"
@@ -10928,9 +11015,9 @@ dependencies = [

 [[package]]
 name = "serde_with"
-version = "3.21.0"
+version = "3.22.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "76a5c54c7310e7b8b9577c286d7e399ddd876c3e12b3ed917a8aabc4b96e9e8c"
+checksum = "ee78f1fbe43ac4a0e47aadb3dbd357b69eb0d3793e948624cd03dd2750ab1c0a"
 dependencies = [
 "base64 0.22.1",
 "bs58",
@@ -10938,6 +11025,7 @@ dependencies = [
 "hex",
 "indexmap 1.9.3",
 "indexmap 2.14.0",
+ "jiff",
 "schemars 0.9.0",
 "schemars 1.2.2",
 "serde_core",
@@ -10948,9 +11036,9 @@ dependencies = [

 [[package]]
 name = "serde_with_macros"
-version = "3.21.0"
+version = "3.22.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "84d57bc0c8b9a17920c178daa6bb924850d54a9c97ab45194bb8c17ad66bb660"
+checksum = "8705578779c2b6bd90d84d66eb2e206b708b1a4d7b9f17641b293545bf1c7e46"
 dependencies = [
 "darling 0.23.0",
 "proc-macro2",
@@ -11244,17 +11332,6 @@ dependencies = [
 "serde",
 ]

-[[package]]
-name = "smartstring"
-version = "1.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3fb72c633efbaa2dd666986505016c32c3044395ceaf881518399d2f4127ee29"
-dependencies = [
- "autocfg",
- "static_assertions",
- "version_check",
-]
-
 [[package]]
 name = "snafu"
 version = "0.6.10"
@@ -11501,31 +11578,13 @@ version = "0.11.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7da8b5736845d9f2fcb837ea5d9e2628564b3b043a70948a3f0b778838c5fb4f"

-[[package]]
-name = "strum"
-version = "0.27.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "af23d6f6c1a224baef9d3f61e287d2761385a5b88fdab4eb4c6f11aeb54c4bcf"
-
 [[package]]
 name = "strum"
 version = "0.28.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9628de9b8791db39ceda2b119bbe13134770b56c138ec1d3af810d045c04f9bd"
 dependencies = [
- "strum_macros 0.28.0",
-]
-
-[[package]]
-name = "strum_macros"
-version = "0.27.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7695ce3845ea4b33927c055a39dc438a45b059f7c1b3d91d38d10355fb8cbca7"
-dependencies = [
- "heck",
- "proc-macro2",
- "quote",
- "syn 2.0.119",
+ "strum_macros",
 ]

 [[package]]
@@ -11534,7 +11593,7 @@ version = "0.28.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "ab85eea0270ee17587ed4156089e10b9e6880ee688791d45a905f5b1ca36f664"
 dependencies = [
- "heck",
+ "heck 0.5.0",
 "proc-macro2",
 "quote",
 "syn 2.0.119",
@@ -11741,7 +11800,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
 dependencies = [
 "fastrand",
- "getrandom 0.4.3",
+ "getrandom 0.3.4",
 "once_cell",
 "rustix",
 "windows-sys 0.61.2",
@@ -12807,9 +12866,9 @@ dependencies = [

 [[package]]
 name = "whoami"
-version = "2.1.2"
+version = "2.1.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "998767ef88740d1f5b0682a9c53c24431453923962269c2db68ee43788c5a40d"
+checksum = "626c4bac6755d76ffc12cb01b2eac751db1996b9e0041de9aa02c8c211ddc82c"
 dependencies = [
 "libc",
 "libredox",
@@ -142,10 +142,10 @@ async-recursion = "1.1.1"
 async-trait = "0.1.92"
 async-nats = { version = "0.50.0", default-features = false }
 axum = "0.8.9"
-futures = "0.3.33"
-futures-core = "0.3.33"
+futures = "0.3.34"
+futures-core = "0.3.34"
 futures-lite = "2.6.1"
-futures-util = "0.3.33"
+futures-util = "0.3.34"
 pollster = "1.0.1"
 pulsar = { default-features = false, version = "6.8.0" }
 lapin = { default-features = false, version = "4.10.0" }
@@ -154,7 +154,7 @@ hyper-rustls = { default-features = false, version = "0.27.9" }
 hyper-util = { version = "0.1.20" }
 http = "1.5.0"
 http-body = "1.1.0"
-http-body-util = "0.1.4"
+http-body-util = "0.1.5"
 minlz = "1.2.3"
 reqwest = "0.13.4"
 rustfs-kafka-async = { version = "1.2.0" }
@@ -171,7 +171,7 @@ tower = { version = "0.5.3" }
 tower-http = { version = "0.7.0" }

 # Serialization and Data Formats
-apache-avro = "0.21.0"
+apache-avro = "0.22.0"
 bytes = { version = "1.12.1" }
 bytesize = "2.7.0"
 byteorder = "1.5.0"
@@ -182,6 +182,7 @@ quick-xml = "0.41.0"
 rmp = { version = "0.8.15" }
 rmp-serde = { version = "1.3.1" }
 serde = { version = "1.0.229" }
+serde_ignored = { version = "0.1" }
 serde_json = { version = "1.0.151" }
 serde_urlencoded = "0.7.1"

@@ -230,9 +231,9 @@ aws-credential-types = { version = "1.3.0" }
 aws-sdk-kms = { default-features = false, version = "1.114.0" }
 aws-sdk-s3 = { default-features = false, version = "1.141.0" }
 aws-sdk-sts = { default-features = false, version = "1.110.0" }
-aws-smithy-http-client = { default-features = false, version = "1.2.0" }
+aws-smithy-http-client = { default-features = false, version = "1.3.0" }
 aws-smithy-runtime-api = { version = "1.14.0" }
-aws-smithy-types = { version = "1.6.1" }
+aws-smithy-types = { version = "1.6.2" }
 base64 = "0.23.1"
 base64-simd = "0.8.0"
 brotli = "8.0.4"
@@ -268,7 +269,7 @@ lz4 = "1.28.1"
 matchit = "0.9.2"
 md-5 = "0.11.0"
 mime_guess = "2.0.5"
-moka = { version = "0.12.15" }
+moka = { version = "0.12.16" }
 netif = "0.1.6"
 num_cpus = { version = "1.17.0" }
 nvml-wrapper = "0.12.1"
@@ -294,7 +295,7 @@ serial_test = "4.0.1"
 shadow-rs = { default-features = false, version = "2.0.0" }
 siphasher = "1.0.3"
 smallvec = { version = "1.15.2" }
-smartstring = "1.0.1"
+compact_str = "0.10.0"
 snap = "1.1.2"
 starshard = { version = "2.2.2" }
 strum = { version = "0.28.0" }
@@ -339,17 +340,17 @@ pyroscope = { version = "2.1.1" }
 libunftp = { version = "0.23.0" }
 unftp-core = "0.1.0"
 suppaftp = { version = "10.0.1" }
-rcgen = { version = "0.14.8", default-features = false, features = ["aws_lc_rs", "crypto", "pem"] }
-russh = { version = "0.62.5" }
+rcgen = { version = "0.14.9", default-features = false, features = ["aws_lc_rs", "crypto", "pem"] }
+russh = { version = "0.62.6" }
 russh-sftp = "2.4.0"

 # WebDAV
 dav-server = "0.11.0"

 # Performance Analysis and Memory Profiling
-mimalloc = { version = "0.1.52", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "ce6338661179c8be22e516b00af7483f151485a7" }
-libmimalloc-sys = { version = "0.1.49", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "ce6338661179c8be22e516b00af7483f151485a7", features = ["extended"] }
-hotpath = { version = "0.23.1", default-features = false }
+mimalloc = { version = "0.1.52", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11" }
+libmimalloc-sys = { version = "0.1.49", git = "https://github.com/xonatius/mimalloc_rust.git", rev = "6d4c41bb10c6d9da1d1b6f07b38c4cc051667f11", features = ["extended"] }
+hotpath = { version = "0.23.2", default-features = false }
 # Snapshot testing for output format regression detection
 insta = { version = "1.48" }

@@ -21,6 +21,13 @@ use crate::{
    Xxhash3, Xxhash64, Xxhash128,
 };

+// DELIBERATE DUPLICATION of the x-amz-checksum-* names that also exist as
+// AMZ_CHECKSUM_* in rustfs-utils' headers module (crates/utils/src/http/
+// headers.rs): this crate is a zero-internal-dependency leaf, so it cannot
+// import them, and it additionally owns the RustFS extension names
+// (sha512/xxhash*) that utils does not carry. Values are pinned by the S3
+// wire protocol; do not merge without a maintainer decision on the leaf
+// boundary (backlog#1833).
 pub const CRC_32_HEADER_NAME: &str = "x-amz-checksum-crc32";
 pub const CRC_32_C_HEADER_NAME: &str = "x-amz-checksum-crc32c";
 pub const SHA_1_HEADER_NAME: &str = "x-amz-checksum-sha1";
@@ -41,6 +41,14 @@ pub const XXHASH_64_NAME: &str = "xxhash64";
 pub const XXHASH_128_NAME: &str = "xxhash128";
 pub const MD5_NAME: &str = "md5";

+/// One of three deliberately separate checksum registries (backlog#1833):
+/// this enum owns the **streaming-hash algorithm registry**, including the
+/// RustFS extensions (sha512, xxhash3/64/128). The on-disk xl.meta bitset
+/// lives in `rustfs_rio::ChecksumType` (crates/rio/src/checksum.rs, varint
+/// bits are append-only), and the MinIO-port client keeps its own
+/// `ChecksumMode` (crates/ecstore/src/client/checksum.rs). When adding an
+/// algorithm, extend all three (or record why not) — they do not derive from
+/// each other.
 #[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
 #[non_exhaustive]
 pub enum ChecksumAlgorithm {
@@ -1,87 +0,0 @@
-// Copyright 2024 RustFS Team
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-use crate::last_minute::{self};
-use std::collections::HashMap;
-
-pub struct ReplicationLatency {
-    // Delays for single and multipart PUT requests
-    upload_histogram: last_minute::LastMinuteHistogram,
-}
-
-impl ReplicationLatency {
-    // Merge two ReplicationLatency
-    pub fn merge(&mut self, other: &mut ReplicationLatency) -> &ReplicationLatency {
-        self.upload_histogram.merge(&other.upload_histogram);
-        self
-    }
-
-    // Get upload delay (categorized by object size interval)
-    pub fn get_upload_latency(&mut self) -> HashMap<String, u64> {
-        let mut ret = HashMap::new();
-        let avg = self.upload_histogram.get_avg_data();
-        for (i, v) in avg.iter().enumerate() {
-            let avg_duration = v.avg();
-            ret.insert(self.size_tag_to_string(i), avg_duration.as_millis() as u64);
-        }
-        ret
-    }
-    pub fn update(&mut self, size: i64, during: std::time::Duration) {
-        self.upload_histogram.add(size, during);
-    }
-
-    // Simulate the conversion from size tag to string
-    fn size_tag_to_string(&self, tag: usize) -> String {
-        match tag {
-            0 => String::from("Size < 1 KiB"),
-            1 => String::from("Size < 1 MiB"),
-            2 => String::from("Size < 10 MiB"),
-            3 => String::from("Size < 100 MiB"),
-            4 => String::from("Size < 1 GiB"),
-            _ => String::from("Size > 1 GiB"),
-        }
-    }
-}
-
-// #[derive(Debug, Clone, Default)]
-// pub struct ReplicationLastMinute {
-//     pub last_minute: LastMinuteLatency,
-// }
-
-// impl ReplicationLastMinute {
-//     pub fn merge(&mut self, other: ReplicationLastMinute) -> ReplicationLastMinute {
-//         let mut nl = ReplicationLastMinute::default();
-//         nl.last_minute = self.last_minute.merge(&mut other.last_minute);
-//         nl
-//     }
-
-//     pub fn add_size(&mut self, n: i64) {
-//         let t = SystemTime::now()
-//             .duration_since(UNIX_EPOCH)
-//             .expect("Time went backwards")
-//             .as_secs();
-//         self.last_minute.add_all(t - 1, &AccElem { total: t - 1, size: n as u64, n: 1 });
-//     }
-
-//     pub fn get_total(&self) -> AccElem {
-//         self.last_minute.get_total()
-//     }
-// }
-
-// impl fmt::Display for ReplicationLastMinute {
-//     fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
-//         let t = self.last_minute.get_total();
-//         write!(f, "ReplicationLastMinute sz= {}, n= {}, dur= {}", t.size, t.n, t.total)
-//     }
-// }
@@ -572,44 +572,3 @@ mod tests {
        assert_eq!(total.n, 6);
    }
 }
-
-const SIZE_LAST_ELEM_MARKER: usize = 10; // Assumed marker size is 10, modify according to actual situation
-
-#[allow(dead_code)]
-#[derive(Debug, Default)]
-pub struct LastMinuteHistogram {
-    histogram: Vec<LastMinuteLatency>,
-    size: u32,
-}
-
-impl LastMinuteHistogram {
-    pub fn merge(&mut self, other: &LastMinuteHistogram) {
-        for i in 0..self.histogram.len() {
-            self.histogram[i].merge(&other.histogram[i]);
-        }
-    }
-
-    pub fn add(&mut self, size: i64, t: Duration) {
-        let index = size_to_tag(size);
-        self.histogram[index].add(&t);
-    }
-
-    pub fn get_avg_data(&mut self) -> [AccElem; SIZE_LAST_ELEM_MARKER] {
-        let mut res = [AccElem::default(); SIZE_LAST_ELEM_MARKER];
-        for (i, elem) in self.histogram.iter_mut().enumerate() {
-            res[i] = elem.get_total();
-        }
-        res
-    }
-}
-
-fn size_to_tag(size: i64) -> usize {
-    match size {
-        _ if size < 1024 => 0,               // sizeLessThan1KiB
-        _ if size < 1024 * 1024 => 1,        // sizeLessThan1MiB
-        _ if size < 10 * 1024 * 1024 => 2,   // sizeLessThan10MiB
-        _ if size < 100 * 1024 * 1024 => 3,  // sizeLessThan100MiB
-        _ if size < 1024 * 1024 * 1024 => 4, // sizeLessThan1GiB
-        _ => 5,                              // sizeGreaterThan1GiB
-    }
-}
@@ -12,13 +12,13 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

-pub mod bucket_stats;
 // pub mod error;
 pub mod globals;
 pub mod heal_channel;
 pub mod last_minute;
 pub mod metrics;
 mod readiness;
+pub mod table_catalog;

 pub use globals::*;
 pub use readiness::{GlobalReadiness, SystemStage};
@@ -915,11 +915,13 @@ const SCAN_CYCLE_RESULT_SUCCESS: u8 = 1;
 const SCAN_CYCLE_RESULT_ERROR: u8 = 2;
 const SCAN_CYCLE_RESULT_PARTIAL: u8 = 3;
 const SCAN_CYCLE_RESULT_SUPERSEDED: u8 = 4;
+const SCAN_CYCLE_RESULT_DEFERRED: u8 = 5;
 const SCAN_CYCLE_RESULT_UNKNOWN_LABEL: &str = "unknown";
 const SCAN_CYCLE_RESULT_SUCCESS_LABEL: &str = "success";
 const SCAN_CYCLE_RESULT_ERROR_LABEL: &str = "error";
 const SCAN_CYCLE_RESULT_PARTIAL_LABEL: &str = "partial";
 const SCAN_CYCLE_RESULT_SUPERSEDED_LABEL: &str = "superseded";
+const SCAN_CYCLE_RESULT_DEFERRED_LABEL: &str = "deferred";

 #[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
 pub enum ScanCyclePartialReason {
@@ -1424,6 +1426,7 @@ fn scan_cycle_result_label(result: u8) -> &'static str {
        SCAN_CYCLE_RESULT_ERROR => SCAN_CYCLE_RESULT_ERROR_LABEL,
        SCAN_CYCLE_RESULT_PARTIAL => SCAN_CYCLE_RESULT_PARTIAL_LABEL,
        SCAN_CYCLE_RESULT_SUPERSEDED => SCAN_CYCLE_RESULT_SUPERSEDED_LABEL,
+        SCAN_CYCLE_RESULT_DEFERRED => SCAN_CYCLE_RESULT_DEFERRED_LABEL,
        _ => SCAN_CYCLE_RESULT_UNKNOWN_LABEL,
    }
 }
@@ -1752,6 +1755,11 @@ pub fn emit_scan_cycle_superseded(duration: Duration) {
    metrics::counter!(OTEL_SCANNER_CYCLES, "result" => SCAN_CYCLE_RESULT_SUPERSEDED_LABEL).increment(1);
 }

+pub fn emit_scan_cycle_deferred(duration: Duration) {
+    global_metrics().record_scan_cycle_deferred(duration);
+    metrics::counter!(OTEL_SCANNER_CYCLES, "result" => SCAN_CYCLE_RESULT_DEFERRED_LABEL).increment(1);
+}
+
 pub fn emit_scan_bucket_drive_complete(success: bool, bucket: &str, disk: &str, duration: Duration) {
    let result = if success { "success" } else { "error" };
    global_metrics().record_scanner_bucket_drive_result(bucket, disk, result);
@@ -2549,6 +2557,17 @@ impl Metrics {
            .store(duration_millis_saturated(duration), Ordering::Relaxed);
    }

+    pub fn record_scan_cycle_deferred(&self, duration: Duration) {
+        self.record_scanner_cycle_end_time();
+        self.last_scan_cycle_result
+            .store(SCAN_CYCLE_RESULT_DEFERRED, Ordering::Relaxed);
+        self.last_scan_cycle_partial_reason
+            .store(ScanCyclePartialReason::Unknown as u8, Ordering::Relaxed);
+        self.last_scan_cycle_partial_source.store(0, Ordering::Relaxed);
+        self.last_scan_cycle_duration_millis
+            .store(duration_millis_saturated(duration), Ordering::Relaxed);
+    }
+
    pub fn record_scan_cycle_partial(&self, duration: Duration, reason: ScanCyclePartialReason) {
        self.record_scan_cycle_partial_with_source(duration, reason, None);
    }
@@ -4264,6 +4283,21 @@ mod tests {
        assert_eq!(report.partial_cycles, 0);
    }

+    #[tokio::test]
+    async fn report_tracks_deferred_cycle_without_failed_increment() {
+        let metrics = Metrics::new();
+        metrics.record_scan_cycle_deferred(Duration::from_millis(250));
+
+        let report = metrics.report().await;
+
+        assert_eq!(report.last_cycle_result, SCAN_CYCLE_RESULT_DEFERRED_LABEL);
+        assert_eq!(report.last_cycle_result_code, u64::from(SCAN_CYCLE_RESULT_DEFERRED));
+        assert_eq!(report.last_cycle_duration_seconds, 0.25);
+        assert_eq!(report.failed_cycles, 0);
+        assert_eq!(report.superseded_cycles, 0);
+        assert_eq!(report.partial_cycles, 0);
+    }
+
    #[tokio::test]
    async fn report_tracks_successful_scan_cycle_without_failed_increment() {
        let metrics = Metrics::new();
@@ -0,0 +1,17 @@
+// 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.
+
+/// Cross-crate lock identity used to fence table-bucket publication against
+/// object mutations that bypass the S3 request authorization layer.
+pub const TABLE_BUCKET_PUBLICATION_LOCK_PATH: &str = ".rustfs-table/warehouses/default/publication.lock";
@@ -353,6 +353,11 @@ pub const DEFAULT_OBS_TRACES_EXPORT_ENABLED: bool = true;
 /// Environment variable: RUSTFS_OBS_METRICS_EXPORT_ENABLED
 pub const DEFAULT_OBS_METRICS_EXPORT_ENABLED: bool = true;

+/// Default detailed PUT stage metrics enabled
+/// Default value: false
+/// Environment variable: RUSTFS_OBS_PUT_STAGE_METRICS_ENABLED
+pub const DEFAULT_OBS_PUT_STAGE_METRICS_ENABLED: bool = false;
+
 /// Default logs export enabled
 /// It is used to enable or disable exporting logs
 /// Default value: true
@@ -137,6 +137,21 @@ pub const DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED: bool = false;
 const _: () = assert!(!DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE);
 const _: () = assert!(!DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED);

+/// Request preserving legacy per-part checksum metadata during data movement.
+///
+/// This remains ineffective until
+/// [`ENV_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED`] is also enabled.
+pub const ENV_DATA_MOVEMENT_PART_CHECKSUMS_WRITE: &str = "RUSTFS_DATA_MOVEMENT_PART_CHECKSUMS_WRITE";
+pub const DEFAULT_DATA_MOVEMENT_PART_CHECKSUMS_WRITE: bool = false;
+
+/// Operator-attested confirmation that every serving node understands the
+/// data-movement per-part checksum sidecar.
+pub const ENV_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED: &str = "RUSTFS_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED";
+pub const DEFAULT_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED: bool = false;
+
+const _: () = assert!(!DEFAULT_DATA_MOVEMENT_PART_CHECKSUMS_WRITE);
+const _: () = assert!(!DEFAULT_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED);
+
 // =============================================================================
 // Concurrent Request Fix - Timeout and Backpressure Configuration
 // =============================================================================
@@ -649,4 +664,13 @@ mod remote_version_state_tests {
            "RUSTFS_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED"
        );
    }
+
+    #[test]
+    fn data_movement_part_checksum_gate_uses_stable_environment_names() {
+        assert_eq!(super::ENV_DATA_MOVEMENT_PART_CHECKSUMS_WRITE, "RUSTFS_DATA_MOVEMENT_PART_CHECKSUMS_WRITE");
+        assert_eq!(
+            super::ENV_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED,
+            "RUSTFS_DATA_MOVEMENT_PART_CHECKSUMS_FLEET_CONFIRMED"
+        );
+    }
 }
@@ -81,6 +81,9 @@ pub const ENV_TEST_IAM_FAIL_INIT_ATTEMPTS: &str = "RUSTFS_TEST_IAM_FAIL_INIT_ATT
 pub const ENV_TEST_IAM_RETRY_INTERVAL_MS: &str = "RUSTFS_TEST_IAM_RETRY_INTERVAL_MS";
 /// Runtime env var controlling the transition worker count.
 pub const ENV_TRANSITION_WORKERS: &str = "RUSTFS_MAX_TRANSITION_WORKERS";
+/// Runtime env var controlling the ILM expiry worker count. A set, parsable,
+/// non-zero value wins; anything else falls back to `min(cpus, 16)`.
+pub const ENV_MAX_EXPIRY_WORKERS: &str = "RUSTFS_MAX_EXPIRY_WORKERS";
 /// Runtime env var controlling the absolute maximum transition workers.
 pub const ENV_TRANSITION_WORKERS_ABSOLUTE_MAX: &str = "RUSTFS_ABSOLUTE_MAX_WORKERS";
 /// Runtime env var controlling the transition queue capacity.
@@ -44,6 +44,10 @@ pub const ENV_OBS_METRICS_EXPORT_ENABLED: &str = "RUSTFS_OBS_METRICS_EXPORT_ENAB
 pub const ENV_OBS_LOGS_EXPORT_ENABLED: &str = "RUSTFS_OBS_LOGS_EXPORT_ENABLED";
 pub const ENV_OBS_PROFILING_EXPORT_ENABLED: &str = "RUSTFS_OBS_PROFILING_EXPORT_ENABLED";

+/// Enables detailed per-stage PUT metrics. Disabled by default because each
+/// PUT records multiple timers and histograms when attribution is active.
+pub const ENV_OBS_PUT_STAGE_METRICS_ENABLED: &str = "RUSTFS_OBS_PUT_STAGE_METRICS_ENABLED";
+
 pub const ENV_OBS_LOGGER_LEVEL: &str = "RUSTFS_OBS_LOGGER_LEVEL";
 pub const ENV_OBS_LOG_STDOUT_ENABLED: &str = "RUSTFS_OBS_LOG_STDOUT_ENABLED";
 pub const ENV_OBS_LOG_DIRECTORY: &str = "RUSTFS_OBS_LOG_DIRECTORY";
@@ -141,6 +145,7 @@ mod tests {
        assert_eq!(ENV_OBS_METRICS_EXPORT_ENABLED, "RUSTFS_OBS_METRICS_EXPORT_ENABLED");
        assert_eq!(ENV_OBS_LOGS_EXPORT_ENABLED, "RUSTFS_OBS_LOGS_EXPORT_ENABLED");
        assert_eq!(ENV_OBS_PROFILING_EXPORT_ENABLED, "RUSTFS_OBS_PROFILING_EXPORT_ENABLED");
+        assert_eq!(ENV_OBS_PUT_STAGE_METRICS_ENABLED, "RUSTFS_OBS_PUT_STAGE_METRICS_ENABLED");
        // Test log cleanup related env keys
        assert_eq!(ENV_OBS_LOG_MAX_TOTAL_SIZE_BYTES, "RUSTFS_OBS_LOG_MAX_TOTAL_SIZE_BYTES");
        assert_eq!(ENV_OBS_LOG_MAX_SINGLE_FILE_SIZE_BYTES, "RUSTFS_OBS_LOG_MAX_SINGLE_FILE_SIZE_BYTES");
@@ -37,7 +37,6 @@ hotpath-cpu = ["hotpath", "hotpath/hotpath-cpu", "rustfs-filemeta/hotpath-cpu"]
 hotpath.workspace = true
 serde = { workspace = true, features = ["derive"] }
 rmp-serde = { workspace = true }
-async-trait = { workspace = true }
 rustfs-filemeta = { workspace = true }

 [lib]
@@ -846,8 +846,15 @@ impl DataUsageEntry {
    }
 }

-/// Data usage cache info
-#[derive(Clone, Debug, Default, Serialize, Deserialize)]
+/// Read-only projection of the scanner's `.usage-cache.bin` info block.
+///
+/// The canonical wire format is written by the hand-written map-encoded
+/// `Serialize` on the scanner-side `DataUsageCacheInfo`
+/// (`crates/scanner/src/data_usage_define.rs`), which carries 16 fields.
+/// This type decodes only the shared subset and is deliberately not
+/// `Serialize`: a derived (array) encoding of this 6-field subset would
+/// corrupt the cache for scanner readers, so no write path may exist here.
+#[derive(Clone, Debug, Default, Deserialize)]
 pub struct DataUsageCacheInfo {
    pub name: String,
    pub next_cycle: u64,
@@ -863,8 +870,12 @@ pub struct DataUsageCacheInfo {
    pub snapshot_complete: bool,
 }

-/// Data usage cache
-#[derive(Clone, Debug, Default, Serialize, Deserialize)]
+/// Read-only projection of a scanner-written `.usage-cache.bin` file.
+///
+/// The scanner-side `DataUsageCache` (`crates/scanner/src/data_usage_define.rs`)
+/// owns the persisted format; this type only decodes it (see
+/// [`DataUsageCacheInfo`]) and must never grow a serialization path.
+#[derive(Clone, Debug, Default, Deserialize)]
 pub struct DataUsageCache {
    pub info: DataUsageCacheInfo,
    pub cache: HashMap<String, DataUsageEntry>,
@@ -1186,31 +1197,10 @@ impl DataUsageCache {
        }
    }

-    pub fn marshal_msg(&self) -> Result<Vec<u8>, Box<dyn std::error::Error + Send + Sync>> {
-        let mut buf = Vec::new();
-        self.serialize(&mut rmp_serde::Serializer::new(&mut buf))?;
-        Ok(buf)
-    }
-
    pub fn unmarshal(buf: &[u8]) -> Result<Self, Box<dyn std::error::Error + Send + Sync>> {
        let t: Self = rmp_serde::from_slice(buf)?;
        Ok(t)
    }
-
-    // Note: load and save methods are storage-specific and should be implemented
-    // in the ecstore crate where storage access is available
-}
-
-/// Trait for storage-specific operations on DataUsageCache
-#[async_trait::async_trait]
-pub trait DataUsageCacheStorage {
-    /// Load data usage cache from backend storage
-    async fn load(store: &dyn std::any::Any, name: &str) -> Result<Self, Box<dyn std::error::Error + Send + Sync>>
-    where
-        Self: Sized;
-
-    /// Save data usage cache to backend storage
-    async fn save(&self, name: &str) -> Result<(), Box<dyn std::error::Error + Send + Sync>>;
 }

 // Helper structs and functions for cache operations
@@ -1832,6 +1822,82 @@ mod tests {
        assert!(decoded.all_tier_stats.is_none());
    }

+    /// Scanner-written `.usage-cache.bin` bytes: a 2-element array of the
+    /// canonical 16-field map-encoded info block and one map-encoded entry.
+    /// Captured from the canonical writer's `marshal_msg` — see
+    /// `usage_cache_wire_format_is_pinned` in
+    /// `crates/scanner/src/data_usage_define.rs`, which pins these exact
+    /// bytes and documents regeneration. Hardcoded here because a
+    /// dev-dependency on rustfs-scanner would pull the whole ecstore tree
+    /// into this crate's test build, and a fixture generated at test runtime
+    /// could not detect writer drift anyway.
+    const SCANNER_USAGE_CACHE_WIRE_FIXTURE: &[u8] = &[
+        0x92, 0xde, 0x00, 0x10, 0xa4, 0x6e, 0x61, 0x6d, 0x65, 0xab, 0x77, 0x69, 0x72, 0x65, 0x2d, 0x62, 0x75, 0x63, 0x6b, 0x65,
+        0x74, 0xaa, 0x6e, 0x65, 0x78, 0x74, 0x5f, 0x63, 0x79, 0x63, 0x6c, 0x65, 0x07, 0xac, 0x6c, 0x65, 0x61, 0x64, 0x65, 0x72,
+        0x5f, 0x65, 0x70, 0x6f, 0x63, 0x68, 0x09, 0xab, 0x6c, 0x61, 0x73, 0x74, 0x5f, 0x75, 0x70, 0x64, 0x61, 0x74, 0x65, 0x92,
+        0xce, 0x65, 0x53, 0xf1, 0x00, 0x00, 0xac, 0x73, 0x6b, 0x69, 0x70, 0x5f, 0x68, 0x65, 0x61, 0x6c, 0x69, 0x6e, 0x67, 0xc3,
+        0xa9, 0x6c, 0x69, 0x66, 0x65, 0x63, 0x79, 0x63, 0x6c, 0x65, 0xc0, 0xab, 0x72, 0x65, 0x70, 0x6c, 0x69, 0x63, 0x61, 0x74,
+        0x69, 0x6f, 0x6e, 0xc0, 0xae, 0x66, 0x61, 0x69, 0x6c, 0x65, 0x64, 0x5f, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x73, 0x81,
+        0xb0, 0x77, 0x69, 0x72, 0x65, 0x2d, 0x62, 0x75, 0x63, 0x6b, 0x65, 0x74, 0x2f, 0x6c, 0x6f, 0x73, 0x74, 0x0b, 0xb1, 0x73,
+        0x63, 0x61, 0x6e, 0x5f, 0x72, 0x65, 0x73, 0x75, 0x6d, 0x65, 0x5f, 0x61, 0x66, 0x74, 0x65, 0x72, 0xb2, 0x77, 0x69, 0x72,
+        0x65, 0x2d, 0x62, 0x75, 0x63, 0x6b, 0x65, 0x74, 0x2f, 0x72, 0x65, 0x73, 0x75, 0x6d, 0x65, 0xaf, 0x73, 0x63, 0x61, 0x6e,
+        0x5f, 0x63, 0x68, 0x65, 0x63, 0x6b, 0x70, 0x6f, 0x69, 0x6e, 0x74, 0xc0, 0xad, 0x70, 0x65, 0x6e, 0x64, 0x69, 0x6e, 0x67,
+        0x5f, 0x68, 0x65, 0x61, 0x6c, 0x73, 0x91, 0x9a, 0xa6, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0xab, 0x77, 0x69, 0x72, 0x65,
+        0x2d, 0x62, 0x75, 0x63, 0x6b, 0x65, 0x74, 0xa6, 0x62, 0x72, 0x6f, 0x6b, 0x65, 0x6e, 0xc0, 0x01, 0x64, 0xcc, 0xc8, 0x03,
+        0xa8, 0x64, 0x65, 0x66, 0x65, 0x72, 0x72, 0x65, 0x64, 0xa6, 0x62, 0x75, 0x64, 0x67, 0x65, 0x74, 0xab, 0x6f, 0x62, 0x6a,
+        0x65, 0x63, 0x74, 0x5f, 0x6c, 0x6f, 0x63, 0x6b, 0xc0, 0xa6, 0x73, 0x6f, 0x75, 0x72, 0x63, 0x65, 0x92, 0x01, 0x02, 0xb1,
+        0x73, 0x6e, 0x61, 0x70, 0x73, 0x68, 0x6f, 0x74, 0x5f, 0x63, 0x6f, 0x6d, 0x70, 0x6c, 0x65, 0x74, 0x65, 0xc3, 0xb0, 0x73,
+        0x63, 0x61, 0x6e, 0x5f, 0x70, 0x6c, 0x61, 0x6e, 0x5f, 0x64, 0x69, 0x67, 0x65, 0x73, 0x74, 0xdc, 0x00, 0x20, 0x03, 0x03,
+        0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03,
+        0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0xb0, 0x63, 0x61, 0x63, 0x68, 0x65, 0x5f, 0x6b, 0x65, 0x79,
+        0x5f, 0x66, 0x6f, 0x72, 0x6d, 0x61, 0x74, 0x01, 0x81, 0xab, 0x77, 0x69, 0x72, 0x65, 0x2d, 0x62, 0x75, 0x63, 0x6b, 0x65,
+        0x74, 0x8b, 0xa8, 0x63, 0x68, 0x69, 0x6c, 0x64, 0x72, 0x65, 0x6e, 0x90, 0xa4, 0x73, 0x69, 0x7a, 0x65, 0xcd, 0x10, 0x00,
+        0xa7, 0x6f, 0x62, 0x6a, 0x65, 0x63, 0x74, 0x73, 0x03, 0xa8, 0x76, 0x65, 0x72, 0x73, 0x69, 0x6f, 0x6e, 0x73, 0x05, 0xae,
+        0x64, 0x65, 0x6c, 0x65, 0x74, 0x65, 0x5f, 0x6d, 0x61, 0x72, 0x6b, 0x65, 0x72, 0x73, 0x01, 0xa9, 0x6f, 0x62, 0x6a, 0x5f,
+        0x73, 0x69, 0x7a, 0x65, 0x73, 0x9b, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xac, 0x6f, 0x62,
+        0x6a, 0x5f, 0x76, 0x65, 0x72, 0x73, 0x69, 0x6f, 0x6e, 0x73, 0x97, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xb1, 0x72,
+        0x65, 0x70, 0x6c, 0x69, 0x63, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x5f, 0x73, 0x74, 0x61, 0x74, 0x73, 0xc0, 0xa9, 0x63, 0x6f,
+        0x6d, 0x70, 0x61, 0x63, 0x74, 0x65, 0x64, 0xc3, 0xae, 0x66, 0x61, 0x69, 0x6c, 0x65, 0x64, 0x5f, 0x6f, 0x62, 0x6a, 0x65,
+        0x63, 0x74, 0x73, 0x02, 0xae, 0x61, 0x6c, 0x6c, 0x5f, 0x74, 0x69, 0x65, 0x72, 0x5f, 0x73, 0x74, 0x61, 0x74, 0x73, 0x91,
+        0x81, 0xa4, 0x57, 0x41, 0x52, 0x4d, 0x93, 0xcd, 0x08, 0x00, 0x02, 0x01,
+    ];
+
+    #[test]
+    fn thin_usage_cache_decodes_scanner_wire_fixture() {
+        let decoded =
+            DataUsageCache::unmarshal(SCANNER_USAGE_CACHE_WIRE_FIXTURE).expect("thin projection decodes a scanner-written cache");
+
+        // The six fields shared with the scanner's 16-field info block; the
+        // remaining ten (lifecycle, replication, checkpoint, heals, ...) must
+        // be skipped, not error.
+        assert_eq!(decoded.info.name, "wire-bucket");
+        assert_eq!(decoded.info.next_cycle, 7);
+        assert_eq!(
+            decoded.info.last_update,
+            Some(SystemTime::UNIX_EPOCH + Duration::from_secs(1_700_000_000))
+        );
+        assert!(decoded.info.skip_healing);
+        assert_eq!(decoded.info.failed_objects.get("wire-bucket/lost"), Some(&11));
+        assert!(decoded.info.snapshot_complete);
+
+        // Entries use the shared canonical map-encoded type end to end.
+        let entry = decoded.cache.get("wire-bucket").expect("fixture entry decodes");
+        assert_eq!(entry.size, 4096);
+        assert_eq!(entry.objects, 3);
+        assert_eq!(entry.versions, 5);
+        assert_eq!(entry.delete_markers, 1);
+        assert!(entry.compacted);
+        assert_eq!(entry.failed_objects, 2);
+        assert_eq!(
+            entry.all_tier_stats.as_ref().and_then(|tiers| tiers.tiers.get("WARM")),
+            Some(&TierStats {
+                total_size: 2048,
+                num_versions: 2,
+                num_objects: 1,
+            })
+        );
+    }
+
    #[test]
    fn hash_path_uses_portable_slash_semantics() {
        for (input, expected) in [
@@ -40,7 +40,8 @@ use http::header::{CONTENT_TYPE, HOST};
 use rustfs_signer::constants::UNSIGNED_PAYLOAD;
 use rustfs_signer::sign_v4;
 use s3s::Body;
-use std::collections::BTreeSet;
+use sha2::{Digest, Sha256};
+use std::collections::{BTreeMap, BTreeSet};
 use std::error::Error;
 use std::path::{Path, PathBuf};
 use tracing::info;
@@ -59,13 +60,26 @@ pub(crate) struct VersionShardCensus {
    pub version_id: Option<String>,
    pub has_xl_meta: bool,
    pub data_dir: Option<String>,
+    pub erasure_index: Option<usize>,
    pub expected_part_numbers: BTreeSet<usize>,
-    pub present_part_numbers: BTreeSet<usize>,
+    pub present_part_fingerprints: BTreeMap<usize, PartShardFingerprint>,
+    pub inline_data_fingerprint: Option<PartShardFingerprint>,
+}
+
+#[derive(Clone, Debug, Eq, PartialEq)]
+pub(crate) struct PartShardFingerprint {
+    pub size: u64,
+    pub sha256: String,
 }

 impl VersionShardCensus {
    pub(crate) fn is_complete(&self) -> bool {
-        self.has_xl_meta && self.expected_part_numbers == self.present_part_numbers
+        self.has_xl_meta
+            && self.expected_part_numbers.len() == self.present_part_fingerprints.len()
+            && self
+                .expected_part_numbers
+                .iter()
+                .all(|part_number| self.present_part_fingerprints.contains_key(part_number))
    }

    pub(crate) fn matches_manifest(&self, manifest: &Self) -> bool {
@@ -73,10 +87,25 @@ impl VersionShardCensus {
            && self.is_complete()
            && manifest.is_complete()
            && self.data_dir == manifest.data_dir
+            && self.erasure_index == manifest.erasure_index
            && self.expected_part_numbers == manifest.expected_part_numbers
+            && self.present_part_fingerprints == manifest.present_part_fingerprints
+            && self.inline_data_fingerprint == manifest.inline_data_fingerprint
    }
 }

+fn sha256_hex(data: &[u8]) -> String {
+    let digest = Sha256::digest(data);
+    digest.iter().map(|byte| format!("{byte:02x}")).collect()
+}
+
+fn shard_fingerprint(data: &[u8]) -> ChaosResult<PartShardFingerprint> {
+    Ok(PartShardFingerprint {
+        size: u64::try_from(data.len())?,
+        sha256: sha256_hex(data),
+    })
+}
+
 /// Single-node RustFS server with `disk_count` local volume directories that
 /// can be faulted individually while the server is running.
 pub struct DiskFaultHarness {
@@ -283,8 +312,10 @@ pub(crate) fn census_object_version_on_disk(
            version_id,
            has_xl_meta: false,
            data_dir: None,
+            erasure_index: None,
            expected_part_numbers: BTreeSet::new(),
-            present_part_numbers: BTreeSet::new(),
+            present_part_fingerprints: BTreeMap::new(),
+            inline_data_fingerprint: None,
        });
    }

@@ -296,20 +327,31 @@ pub(crate) fn census_object_version_on_disk(
        file_info.parts.iter().map(|part| part.number).collect()
    };
    let data_dir = file_info.data_dir.map(|id| id.to_string());
+    let erasure_index = Some(file_info.erasure.index);
+    let inline_data_fingerprint = file_info.data.as_deref().map(shard_fingerprint).transpose()?;
    let part_dir = data_dir.as_ref().map_or_else(|| object_dir.clone(), |id| object_dir.join(id));
-    let present_part_numbers = match std::fs::read_dir(&part_dir) {
-        Ok(entries) => entries
-            .filter_map(Result::ok)
-            .filter_map(|entry| {
-                entry
-                    .file_type()
-                    .ok()
-                    .filter(|kind| kind.is_file())
-                    .and_then(|_| entry.file_name().to_str().map(str::to_owned))
-            })
-            .filter_map(|name| name.strip_prefix("part.").and_then(|number| number.parse::<usize>().ok()))
-            .collect(),
-        Err(error) if error.kind() == std::io::ErrorKind::NotFound => BTreeSet::new(),
+    let present_part_fingerprints = match std::fs::read_dir(&part_dir) {
+        Ok(entries) => {
+            let mut fingerprints = BTreeMap::new();
+            for entry in entries {
+                let entry = entry?;
+                if !entry.file_type()?.is_file() {
+                    continue;
+                }
+                let file_name = entry.file_name();
+                let Some(part_number) = file_name
+                    .to_str()
+                    .and_then(|name| name.strip_prefix("part."))
+                    .and_then(|number| number.parse::<usize>().ok())
+                else {
+                    continue;
+                };
+                let data = std::fs::read(entry.path())?;
+                fingerprints.insert(part_number, shard_fingerprint(&data)?);
+            }
+            fingerprints
+        }
+        Err(error) if error.kind() == std::io::ErrorKind::NotFound => BTreeMap::new(),
        Err(error) => return Err(error.into()),
    };

@@ -317,8 +359,10 @@ pub(crate) fn census_object_version_on_disk(
        version_id,
        has_xl_meta: true,
        data_dir,
+        erasure_index,
        expected_part_numbers,
-        present_part_numbers,
+        present_part_fingerprints,
+        inline_data_fingerprint,
    })
 }

@@ -358,3 +402,43 @@ pub async fn signed_admin_post(url: &str, body: Option<&str>, access_key: &str,

    Ok(body)
 }
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    fn complete_census() -> VersionShardCensus {
+        VersionShardCensus {
+            version_id: Some("version".to_string()),
+            has_xl_meta: true,
+            data_dir: Some("data-dir".to_string()),
+            erasure_index: Some(3),
+            expected_part_numbers: BTreeSet::from([1]),
+            present_part_fingerprints: BTreeMap::from([(1, shard_fingerprint(b"part").unwrap())]),
+            inline_data_fingerprint: None,
+        }
+    }
+
+    #[test]
+    fn shard_fingerprint_uses_physical_length_and_sha256() {
+        assert_eq!(
+            shard_fingerprint(b"abc").unwrap(),
+            PartShardFingerprint {
+                size: 3,
+                sha256: "ba7816bf8f01cfea414140de5dae2223b00361a396177a9cb410ff61f20015ad".to_string(),
+            }
+        );
+    }
+
+    #[test]
+    fn manifest_requires_matching_inline_payload() {
+        let mut expected = complete_census();
+        expected.expected_part_numbers.clear();
+        expected.present_part_fingerprints.clear();
+        expected.inline_data_fingerprint = Some(shard_fingerprint(b"expected").unwrap());
+        let mut changed = expected.clone();
+        changed.inline_data_fingerprint = Some(shard_fingerprint(b"changed").unwrap());
+        assert!(expected.matches_manifest(&expected));
+        assert!(!changed.matches_manifest(&expected));
+    }
+}
@@ -67,6 +67,16 @@ fn configured_capture_log_path(temp_dir: &str) -> Option<String> {
    capture_log_path(Path::new(&log_dir), temp_dir).map(|path| path.to_string_lossy().into_owned())
 }

+fn capture_command_logs(command: &mut Command, log_path: Option<&str>) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+    let Some(log_path) = log_path else {
+        return Ok(());
+    };
+    let file = stdfs::OpenOptions::new().create(true).append(true).open(log_path)?;
+    let stderr_file = file.try_clone()?;
+    command.stdout(Stdio::from(file)).stderr(Stdio::from(stderr_file));
+    Ok(())
+}
+
 pub(crate) fn build_test_s3_config(
    endpoint_url: &str,
    access_key: &str,
@@ -557,13 +567,7 @@ impl RustFSTestEnvironment {
        for (key, value) in extra_env {
            command.env(key, value);
        }
-        // Optionally capture the child's stdout+stderr to a file so the test can
-        // grep server logs (e.g. to confirm which GET reader path was taken).
-        if let Some(log_path) = &self.capture_log_path {
-            let file = stdfs::OpenOptions::new().create(true).append(true).open(log_path)?;
-            let stderr_file = file.try_clone()?;
-            command.stdout(Stdio::from(file)).stderr(Stdio::from(stderr_file));
-        }
+        capture_command_logs(&mut command, self.capture_log_path.as_deref())?;
        let process = command.args(&args).spawn()?;

        self.process = Some(process);
@@ -1051,6 +1055,7 @@ pub struct RustFSTestClusterEnvironment {
    pub secret_key: String,
    pub extra_env: Vec<(String, String)>,
    pub node_extra_env: Vec<Vec<(String, String)>>,
+    pub node_capture_log_paths: Vec<Option<String>>,
    pub topology: ClusterTopology,
 }

@@ -1150,6 +1155,7 @@ impl RustFSTestClusterEnvironment {
            secret_key: "rustfs-cluster-test-secret".to_string(),
            extra_env,
            node_extra_env: vec![Vec::new(); topology.node_count],
+            node_capture_log_paths: vec![None; topology.node_count],
            topology,
        })
    }
@@ -1179,6 +1185,20 @@ impl RustFSTestClusterEnvironment {
        Ok(())
    }

+    /// Capture stdout+stderr for a single cluster node process.
+    pub fn set_node_capture_log_path<P>(
+        &mut self,
+        node_idx: usize,
+        path: P,
+    ) -> Result<(), Box<dyn std::error::Error + Send + Sync>>
+    where
+        P: Into<String>,
+    {
+        self.ensure_node_index(node_idx)?;
+        self.node_capture_log_paths[node_idx] = Some(path.into());
+        Ok(())
+    }
+
    fn ensure_node_index(&self, node_idx: usize) -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
        if node_idx >= self.nodes.len() {
            return Err(format!("node_idx {node_idx} is invalid").into());
@@ -1268,6 +1288,7 @@ impl RustFSTestClusterEnvironment {
            for (key, value) in &self.node_extra_env[i] {
                command.env(key, value);
            }
+            capture_command_logs(&mut command, self.node_capture_log_paths[i].as_deref())?;

            let process = command.current_dir(&node.data_dir).spawn()?;

@@ -1294,6 +1315,7 @@ impl RustFSTestClusterEnvironment {

        let binary_path = rustfs_binary_path();
        let volumes_arg = self.build_volumes_arg();
+        let log_path = self.node_capture_log_paths[node_idx].clone();
        let node = &mut self.nodes[node_idx];
        info!("Starting cluster node {} on {}", node_idx, node.address);

@@ -1312,6 +1334,7 @@ impl RustFSTestClusterEnvironment {
        for (key, value) in &self.node_extra_env[node_idx] {
            command.env(key, value);
        }
+        capture_command_logs(&mut command, log_path.as_deref())?;

        let process = command.current_dir(&node.data_dir).spawn()?;
        node.process = Some(process);
@@ -1563,6 +1586,7 @@ mod tests {
            secret_key: DEFAULT_SECRET_KEY.to_string(),
            extra_env: Vec::new(),
            node_extra_env: vec![Vec::new(); topology.node_count],
+            node_capture_log_paths: vec![None; topology.node_count],
            topology,
        }
    }
@@ -1658,6 +1682,16 @@ mod tests {
        );
    }

+    #[test]
+    fn cluster_node_log_capture_supports_per_node_paths() {
+        let mut env = fake_cluster(ClusterTopology::single_pool(3));
+        env.set_node_capture_log_path(1, "/tmp/node1.log").unwrap();
+        assert_eq!(env.node_capture_log_paths[0], None);
+        assert_eq!(env.node_capture_log_paths[1], Some("/tmp/node1.log".to_string()));
+        assert_eq!(env.node_capture_log_paths[2], None);
+        assert!(env.set_node_capture_log_path(3, "/tmp/invalid.log").is_err());
+    }
+
    #[test]
    fn cluster_node_env_rejects_invalid_index() {
        let mut env = fake_cluster(ClusterTopology::single_pool(4));
@@ -189,8 +189,6 @@ mod tests {
            ("RUSTFS_GET_CODEC_STREAMING_ROLLOUT_PCT", "100"),
            ("RUSTFS_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED", "true"),
            ("RUSTFS_GET_CODEC_STREAMING_HEADER_COMPAT_CONFIRMED", "true"),
-            // Lower the min-size floor so every non-inline object below is eligible.
-            ("RUSTFS_GET_CODEC_STREAMING_MIN_SIZE", "4096"),
            // Route multipart objects through per-part codec streaming too.
            ("RUSTFS_GET_CODEC_STREAMING_MULTIPART_ENABLE", "true"),
            // Lock optimization is on by default, but pin it so the gate's
@@ -315,6 +313,13 @@ mod tests {
                },
                payload(64 * 1024, 2),
            ),
+            (
+                Shape {
+                    key: "small-non-inline-256kib-plus",
+                    expect_large: true,
+                },
+                payload(256 * 1024 + 1, 6),
+            ),
            (
                Shape {
                    key: "mid-1_5mib",
@@ -14,7 +14,7 @@

 //! E2E tests for group management (fixes #2028).

-use crate::common::{RustFSTestEnvironment, awscurl_delete, awscurl_get, awscurl_put, init_logging};
+use crate::common::{RustFSTestEnvironment, admin_request, awscurl_delete, awscurl_get, awscurl_put, init_logging};
 use aws_sdk_s3::config::{Credentials, Region};
 use aws_sdk_s3::{Client, Config};
 use serial_test::serial;
@@ -32,6 +32,56 @@ fn create_user_s3_client(env: &RustFSTestEnvironment, access_key: &str, secret_k
    Client::from_conf(config)
 }

+#[tokio::test(flavor = "multi_thread")]
+async fn update_group_members_rejects_invalid_new_group_names() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
+    init_logging();
+
+    let mut env = RustFSTestEnvironment::new().await?;
+    env.start_rustfs_server(vec![]).await?;
+
+    let invalid_groups = [
+        ("test group", "group name contains whitespace"),
+        ("test=group", "group name contains reserved characters =,"),
+        ("test,group", "group name contains reserved characters =,"),
+    ];
+
+    for (group, expected_message) in invalid_groups {
+        let body = serde_json::json!({
+            "group": group,
+            "members": [],
+            "isRemove": false,
+            "groupStatus": "enabled"
+        })
+        .to_string();
+        let (status, response_body) = admin_request(
+            &env.url,
+            http::Method::PUT,
+            "/rustfs/admin/v3/update-group-members",
+            Some(body),
+            &env.access_key,
+            &env.secret_key,
+        )
+        .await?;
+
+        assert_eq!(
+            status,
+            reqwest::StatusCode::BAD_REQUEST,
+            "invalid group {group:?} must return HTTP 400, body: {response_body}"
+        );
+        assert!(
+            response_body.contains("<Code>InvalidArgument</Code>"),
+            "invalid group {group:?} must return InvalidArgument, body: {response_body}"
+        );
+        assert!(
+            response_body.contains(&format!("<Message>{expected_message}</Message>")),
+            "invalid group {group:?} returned an unexpected message: {response_body}"
+        );
+    }
+
+    env.stop_server();
+    Ok(())
+}
+
 /// Test that deleting a group with members fails, and deleting an empty group succeeds.
 #[tokio::test(flavor = "multi_thread")]
 #[serial]
@@ -0,0 +1,612 @@
+// 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.
+
+//! ILM on SSE-KMS buckets while per-key SSE authorization is enforced (backlog#1582).
+//!
+//! Per-key KMS authorization (`RUSTFS_KMS_ENFORCE_SSE_KEY_POLICY=true`) scopes the
+//! SSE-KMS data path to the requesting principal's `kms:GenerateDataKey` /
+//! `kms:Decrypt` grants. Internal callers — the lifecycle scanner's expiry deletes
+//! and the tier transition worker's reads — carry no request principal, and
+//! `authorize_sse_kms_key` (rustfs/src/storage/sse.rs) exempts a `None` principal
+//! so background maintenance keeps working on encrypted buckets.
+//!
+//! These tests pin that exemption end to end. If enforcement ever starts applying
+//! to the scanner's internal operations, expiry stops happening on SSE-KMS buckets
+//! and [`ilm_expiration_on_sse_kms_bucket_under_enforcement`] times out; if it
+//! starts applying to the transition worker or the read-through path,
+//! [`ilm_transition_on_sse_kms_bucket_under_enforcement_reads_back`] fails at the
+//! transition wait or the plaintext round-trip.
+//!
+//! The replication half of the same acceptance item lives in
+//! `crates/e2e_test/src/replication_extension_test.rs`
+//! (`test_bucket_replication_sse_kms_failure_contract`); ILM had no coverage
+//! before this file.
+//!
+//! Deployment constraint pinned by the transition test's setup: the RustFS warm
+//! backend forwards the object's stored `x-amz-server-side-encryption*` metadata
+//! as raw headers on the tier data PUT (`build_transition_put_options` +
+//! `api_put_object.rs` header mapping), so a RustFS tier target must itself have
+//! KMS enabled and hold the named key or it rejects every transition upload with
+//! 400 InvalidRequest. That rejection is independent of the enforcement switch;
+//! the cold server here therefore runs its own Local KMS with the same key id.
+
+use super::common::{LocalKMSTestEnvironment, create_key_with_specific_id};
+use crate::common::{RustFSTestEnvironment, admin_request, init_logging};
+use aws_sdk_s3::Client;
+use aws_sdk_s3::primitives::ByteStream;
+use aws_sdk_s3::types::{
+    BucketLifecycleConfiguration, ExpirationStatus, LifecycleExpiration, LifecycleRule, LifecycleRuleFilter, RestoreRequest,
+    ServerSideEncryption, ServerSideEncryptionByDefault, ServerSideEncryptionConfiguration, ServerSideEncryptionRule, Transition,
+    TransitionStorageClass,
+};
+use serde::Deserialize;
+use serial_test::serial;
+use std::time::{Duration as StdDuration, Instant};
+use tracing::info;
+
+type TestResult = Result<(), Box<dyn std::error::Error + Send + Sync>>;
+
+const SSE_KEY: &str = "kms-ilm-sse-key";
+const PAYLOAD: &[u8] = b"kms ilm sse payload: survives enforcement, expires and transitions on schedule";
+
+const EXPIRY_BUCKET: &str = "kms-ilm-expiry";
+const EXPIRE_KEY: &str = "expire/object.bin";
+const SURVIVOR_KEY: &str = "keep/object.bin";
+
+const TIER_NAME: &str = "KMSCOLD";
+const TIER_BUCKET: &str = "kms-ilm-cold-tier";
+const TIER_PREFIX: &str = "tiered";
+const TRANSITION_BUCKET: &str = "kms-ilm-transition";
+const TRANSITION_KEY: &str = "tier/object.bin";
+
+/// Generous CI safety net; with a 1s scanner cycle and 2s lifecycle days the
+/// terminal state normally lands within a few seconds.
+const ILM_DEADLINE: StdDuration = StdDuration::from_secs(90);
+
+/// Start a Local-KMS server with per-key SSE authorization enforced and the
+/// lifecycle clock accelerated.
+///
+/// KMS wiring matches `kms_authorization_negative_matrix_test.rs` (local backend,
+/// `--kms-default-key-id`, insecure dev defaults). The lifecycle env matches
+/// `reliant/lifecycle.rs::fast_lifecycle_env` plus `RUSTFS_ILM_DEBUG_DAY_SECS=2`,
+/// so a `Days=1` rule is due about two seconds after the write.
+async fn start_enforcing_ilm_server(env: &mut LocalKMSTestEnvironment) -> TestResult {
+    create_key_with_specific_id(&env.kms_keys_dir, SSE_KEY).await?;
+
+    let key_dir = env.kms_keys_dir.clone();
+    let args = vec![
+        "--kms-enable",
+        "--kms-backend",
+        "local",
+        "--kms-key-dir",
+        key_dir.as_str(),
+        "--kms-default-key-id",
+        SSE_KEY,
+    ];
+
+    let envs = [
+        ("RUSTFS_KMS_ALLOW_INSECURE_DEV_DEFAULTS", "true"),
+        ("RUSTFS_KMS_ENFORCE_SSE_KEY_POLICY", "false"),
+        ("RUSTFS_SCANNER_CYCLE", "1"),
+        ("RUSTFS_ILM_PROCESS_TIME", "1"),
+        ("RUSTFS_ILM_DEBUG_DAY_SECS", "2"),
+    ];
+
+    env.base_env.start_rustfs_server_with_env(args, &envs).await?;
+    Ok(())
+}
+
+/// Set the bucket's default encryption to SSE-KMS under [`SSE_KEY`], so plain
+/// PUTs (and internal rewrites) are encrypted without per-request SSE headers.
+async fn set_bucket_default_sse_kms(client: &Client, bucket: &str) -> TestResult {
+    let encryption_config = ServerSideEncryptionConfiguration::builder()
+        .rules(
+            ServerSideEncryptionRule::builder()
+                .apply_server_side_encryption_by_default(
+                    ServerSideEncryptionByDefault::builder()
+                        .sse_algorithm(ServerSideEncryption::AwsKms)
+                        .kms_master_key_id(SSE_KEY)
+                        .build()?,
+                )
+                .build(),
+        )
+        .build()?;
+    client
+        .put_bucket_encryption()
+        .bucket(bucket)
+        .server_side_encryption_configuration(encryption_config)
+        .send()
+        .await?;
+    Ok(())
+}
+
+/// Assert via `HeadObject` that the stored object is SSE-KMS encrypted under
+/// [`SSE_KEY`]. Without this, a bucket-default misconfiguration would let the
+/// tests pass on an unencrypted object and prove nothing about KMS.
+async fn assert_head_sse_kms(client: &Client, bucket: &str, key: &str) -> TestResult {
+    let head = client.head_object().bucket(bucket).key(key).send().await?;
+    assert_eq!(
+        head.server_side_encryption(),
+        Some(&ServerSideEncryption::AwsKms),
+        "{bucket}/{key} must be SSE-KMS encrypted via the bucket default"
+    );
+    assert_eq!(
+        head.ssekms_key_id(),
+        Some(SSE_KEY),
+        "{bucket}/{key} must be wrapped under the configured KMS key"
+    );
+    Ok(())
+}
+
+/// Returns `true` once `GET bucket/key` fails with `NoSuchKey`, `false` while it
+/// still succeeds. Any other error is surfaced. (Copied from
+/// `reliant/lifecycle.rs`; that helper is private to the reliant module.)
+async fn object_is_gone(client: &Client, bucket: &str, key: &str) -> Result<bool, Box<dyn std::error::Error + Send + Sync>> {
+    match client.get_object().bucket(bucket).key(key).send().await {
+        Ok(output) => {
+            output.body.collect().await?;
+            Ok(false)
+        }
+        Err(e) => {
+            if let Some(service_error) = e.as_service_error() {
+                if service_error.is_no_such_key() {
+                    return Ok(true);
+                }
+                return Err(format!("expected NoSuchKey, got: {e:?}").into());
+            }
+            Err(format!("expected a service error, got: {e:?}").into())
+        }
+    }
+}
+
+/// Poll until `GET bucket/key` returns `NoSuchKey`, or fail after `deadline`.
+async fn wait_for_object_expired(client: &Client, bucket: &str, key: &str, deadline: StdDuration) -> TestResult {
+    let start = Instant::now();
+    loop {
+        if object_is_gone(client, bucket, key).await? {
+            return Ok(());
+        }
+        if start.elapsed() >= deadline {
+            return Err(format!(
+                "object {bucket}/{key} was not expired by the lifecycle scanner within {}s; \
+                 SSE key-policy enforcement may have started blocking the scanner's internal deletes",
+                deadline.as_secs()
+            )
+            .into());
+        }
+        tokio::time::sleep(StdDuration::from_millis(500)).await;
+    }
+}
+
+/// Install a prefix-scoped `Days`-based expiration rule.
+async fn put_expiration_rule(client: &Client, bucket: &str, id: &str, prefix: &str, days: i32) -> TestResult {
+    let rule = LifecycleRule::builder()
+        .id(id)
+        .filter(LifecycleRuleFilter::builder().prefix(prefix).build())
+        .expiration(LifecycleExpiration::builder().days(days).build())
+        .status(ExpirationStatus::Enabled)
+        .build()?;
+    let lifecycle = BucketLifecycleConfiguration::builder().rules(rule).build()?;
+    client
+        .put_bucket_lifecycle_configuration()
+        .bucket(bucket)
+        .lifecycle_configuration(lifecycle)
+        .send()
+        .await?;
+    Ok(())
+}
+
+/// Install a prefix-scoped `Days`-based transition rule targeting [`TIER_NAME`].
+async fn put_transition_rule(client: &Client, bucket: &str, id: &str, prefix: &str, days: i32) -> TestResult {
+    let rule = LifecycleRule::builder()
+        .id(id)
+        .filter(LifecycleRuleFilter::builder().prefix(prefix).build())
+        .transitions(
+            Transition::builder()
+                .days(days)
+                .storage_class(TransitionStorageClass::from(TIER_NAME))
+                .build(),
+        )
+        .status(ExpirationStatus::Enabled)
+        .build()?;
+    let lifecycle = BucketLifecycleConfiguration::builder().rules(rule).build()?;
+    client
+        .put_bucket_lifecycle_configuration()
+        .bucket(bucket)
+        .lifecycle_configuration(lifecycle)
+        .send()
+        .await?;
+    Ok(())
+}
+
+/// Start a plain Local-KMS server (no enforcement, no lifecycle acceleration)
+/// holding [`SSE_KEY`], to serve as the cold tier target.
+///
+/// The RustFS warm backend forwards the stored SSE-KMS headers on the tier data
+/// PUT, so the target re-applies managed SSE-KMS under the named key and must
+/// be able to resolve it; without KMS it answers 400 InvalidRequest and the
+/// transition can never complete. Enforcement stays off here: the tier writes
+/// arrive under `cold`'s root credentials, and one enforcing side is enough to
+/// pin the exemption.
+async fn start_cold_tier_kms_server(env: &mut LocalKMSTestEnvironment) -> TestResult {
+    create_key_with_specific_id(&env.kms_keys_dir, SSE_KEY).await?;
+
+    let key_dir = env.kms_keys_dir.clone();
+    let args = vec![
+        "--kms-enable",
+        "--kms-backend",
+        "local",
+        "--kms-key-dir",
+        key_dir.as_str(),
+        "--kms-default-key-id",
+        SSE_KEY,
+    ];
+
+    env.base_env
+        .start_rustfs_server_with_env(args, &[("RUSTFS_KMS_ALLOW_INSECURE_DEV_DEFAULTS", "true")])
+        .await?;
+    Ok(())
+}
+
+/// The subset of the manual transition run report these tests assert on.
+///
+/// Unknown fields are ignored, so this stays compatible with report growth; the
+/// full shape is pinned by `reliant/tiering.rs`.
+#[derive(Debug, Deserialize)]
+struct ManualTransitionRunReport {
+    #[serde(default)]
+    scanned: u64,
+    #[serde(default)]
+    enqueued: u64,
+    #[serde(default)]
+    skipped_already_in_flight: u64,
+    #[serde(default)]
+    skipped_tier: u64,
+}
+
+#[derive(Debug, Deserialize)]
+struct ManualTransitionRunResponse {
+    state: String,
+    report: ManualTransitionRunReport,
+}
+
+/// One synchronous (enqueue-only) manual transition run over `bucket/prefix`,
+/// via the same admin endpoint `reliant/tiering.rs` drives.
+async fn manual_transition_run(
+    hot: &RustFSTestEnvironment,
+    bucket: &str,
+    prefix: &str,
+) -> Result<ManualTransitionRunResponse, Box<dyn std::error::Error + Send + Sync>> {
+    let bucket = urlencoding::encode(bucket);
+    let prefix = urlencoding::encode(prefix);
+    let tier = urlencoding::encode(TIER_NAME);
+    let path =
+        format!("/rustfs/admin/v3/ilm/transition/run?bucket={bucket}&prefix={prefix}&tier={tier}&dryRun=false&maxObjects=10");
+    let (status, body) = admin_request(&hot.url, http::Method::POST, &path, None, &hot.access_key, &hot.secret_key).await?;
+    if !status.is_success() {
+        return Err(format!("manual transition run failed: status={status}, body={body}").into());
+    }
+    Ok(serde_json::from_str(&body)?)
+}
+
+/// Drive manual transition runs until one reports the object as processed.
+///
+/// The `Days=1` rule becomes due about two seconds after the write
+/// (`RUSTFS_ILM_DEBUG_DAY_SECS=2`), so early runs may legitimately report the
+/// object as not yet eligible; the loop keeps running the endpoint until it
+/// either enqueues the transition, sees it already in flight (the 1s scanner
+/// backstop got there first), or finds it already on the tier.
+async fn run_manual_transition_until_processed(
+    hot: &RustFSTestEnvironment,
+    bucket: &str,
+    prefix: &str,
+    deadline: StdDuration,
+) -> TestResult {
+    let start = Instant::now();
+    loop {
+        let run = manual_transition_run(hot, bucket, prefix).await?;
+        assert_eq!(run.report.scanned, 1, "manual transition run must scan the object: {run:#?}");
+        if run.report.enqueued + run.report.skipped_already_in_flight + run.report.skipped_tier >= 1 {
+            info!(state = %run.state, report = ?run.report, "manual transition run processed the SSE-KMS object");
+            return Ok(());
+        }
+        if start.elapsed() >= deadline {
+            return Err(format!(
+                "manual transition runs never processed {bucket}/{prefix} within {}s; last report: {run:#?}",
+                deadline.as_secs()
+            )
+            .into());
+        }
+        tokio::time::sleep(StdDuration::from_millis(500)).await;
+    }
+}
+
+/// Wire `hot` -> `cold` as a `TierType::RustFS` remote tier via `AddTier`.
+///
+/// No `force`, so the server runs the real connectivity probe against `cold`
+/// (the tier bucket must already exist there). Mirrors
+/// `reliant/tiering.rs::add_rustfs_tier`, which is private to that module.
+async fn add_rustfs_tier(hot: &RustFSTestEnvironment, cold: &RustFSTestEnvironment) -> TestResult {
+    let body = serde_json::json!({
+        "type": "rustfs",
+        "rustfs": {
+            "name": TIER_NAME,
+            "endpoint": cold.url.as_str(),
+            "accessKey": cold.access_key.as_str(),
+            "secretKey": cold.secret_key.as_str(),
+            "bucket": TIER_BUCKET,
+            "prefix": TIER_PREFIX,
+            "region": "us-east-1",
+            "storageClass": ""
+        }
+    })
+    .to_string();
+
+    let (status, resp) = admin_request(
+        &hot.url,
+        http::Method::PUT,
+        "/rustfs/admin/v3/tier",
+        Some(body),
+        &hot.access_key,
+        &hot.secret_key,
+    )
+    .await?;
+    if !status.is_success() {
+        return Err(format!("AddTier(RustFS) failed: status={status}, body={resp}").into());
+    }
+    Ok(())
+}
+
+/// Poll `HEAD` until the object's storage class is the tier name (transition
+/// complete), or fail after `deadline`. (From `reliant/tiering.rs`.)
+async fn wait_for_transition(client: &Client, bucket: &str, key: &str, deadline: StdDuration) -> TestResult {
+    let start = Instant::now();
+    loop {
+        let head = client.head_object().bucket(bucket).key(key).send().await?;
+        if head.storage_class().map(|sc| sc.as_str()) == Some(TIER_NAME) {
+            return Ok(());
+        }
+        if start.elapsed() >= deadline {
+            return Err(format!(
+                "object {bucket}/{key} was not transitioned to {TIER_NAME} within {}s (storage_class={:?}); \
+                 SSE key-policy enforcement may have started blocking the transition worker's internal reads",
+                deadline.as_secs(),
+                head.storage_class()
+            )
+            .into());
+        }
+        tokio::time::sleep(StdDuration::from_millis(500)).await;
+    }
+}
+
+/// Poll `HEAD` until `x-amz-restore` reports a finished restore
+/// (`ongoing-request="false"`), or fail after `deadline`.
+async fn wait_for_restore_complete(client: &Client, bucket: &str, key: &str, deadline: StdDuration) -> TestResult {
+    let start = Instant::now();
+    loop {
+        let head = client.head_object().bucket(bucket).key(key).send().await?;
+        if head.restore().is_some_and(|r| r.contains("ongoing-request=\"false\"")) {
+            return Ok(());
+        }
+        if start.elapsed() >= deadline {
+            return Err(format!(
+                "object {bucket}/{key} restore did not complete within {}s (restore={:?}); \
+                 SSE key-policy enforcement may have started blocking the restore copy-back's internal reads",
+                deadline.as_secs(),
+                head.restore()
+            )
+            .into());
+        }
+        tokio::time::sleep(StdDuration::from_millis(500)).await;
+    }
+}
+
+/// ILM expiration keeps working on an SSE-KMS bucket while per-key SSE
+/// authorization is enforced.
+///
+/// The lifecycle scanner deletes expired objects with an internal (no-principal)
+/// identity that holds no `kms` grant. If enforcement ever starts applying to
+/// those internal deletes (or to the scanner's metadata reads) on encrypted
+/// buckets, expiry stops happening and this test times out.
+///
+/// A survivor object under a non-matching prefix isolates the rule's prefix
+/// filter as the cause of the deletion and proves the encrypted bucket stays
+/// readable end to end after the scanner has run.
+#[tokio::test]
+#[serial]
+async fn ilm_expiration_on_sse_kms_bucket_under_enforcement() -> TestResult {
+    init_logging();
+
+    let mut env = LocalKMSTestEnvironment::new().await?;
+    start_enforcing_ilm_server(&mut env).await?;
+    env.base_env.create_test_bucket(EXPIRY_BUCKET).await?;
+
+    let client = env.base_env.create_s3_client();
+    set_bucket_default_sse_kms(&client, EXPIRY_BUCKET).await?;
+
+    for key in [EXPIRE_KEY, SURVIVOR_KEY] {
+        client
+            .put_object()
+            .bucket(EXPIRY_BUCKET)
+            .key(key)
+            .body(ByteStream::from_static(PAYLOAD))
+            .send()
+            .await?;
+        assert_head_sse_kms(&client, EXPIRY_BUCKET, key).await?;
+    }
+    info!("both objects stored SSE-KMS encrypted under enforcement");
+
+    put_expiration_rule(&client, EXPIRY_BUCKET, "kms-ilm-expire", "expire/", 1).await?;
+
+    // The regression this pins: the scanner's internal delete must stay exempt
+    // from per-key SSE authorization, so the encrypted object actually expires.
+    wait_for_object_expired(&client, EXPIRY_BUCKET, EXPIRE_KEY, ILM_DEADLINE).await?;
+    info!("SSE-KMS object expired by the lifecycle scanner under enforcement");
+
+    // Negative control: same bucket, same encryption, non-matching prefix. It
+    // must survive the scanner and still decrypt for the requesting principal.
+    assert!(
+        !object_is_gone(&client, EXPIRY_BUCKET, SURVIVOR_KEY).await?,
+        "non-matching-prefix object must not be expired by a prefix-scoped rule"
+    );
+    let survivor = client.get_object().bucket(EXPIRY_BUCKET).key(SURVIVOR_KEY).send().await?;
+    assert_eq!(
+        survivor.body.collect().await?.into_bytes().as_ref(),
+        PAYLOAD,
+        "surviving SSE-KMS object must still decrypt after the scanner has run"
+    );
+
+    Ok(())
+}
+
+/// ILM transition to a remote tier keeps working on an SSE-KMS bucket while
+/// per-key SSE authorization is enforced, and the transitioned object reads
+/// back as plaintext.
+///
+/// The transition worker moves the stored (encrypted) bytes to the cold tier
+/// with an internal (no-principal) identity; the read-through `GET` then
+/// decrypts the envelope for the requesting principal. If enforcement ever
+/// starts applying to the worker's internal reads, the transition wait times
+/// out; if the stored envelope is mishandled across the tier round trip, the
+/// plaintext comparison fails.
+///
+/// The transition is driven through the manual transition-run admin endpoint
+/// (the mechanism `reliant/tiering.rs` established), so the test does not
+/// depend on scanner scheduling; the 1s scanner cycle stays on as a backstop.
+#[tokio::test]
+#[serial]
+#[ignore = "pins rustfs/rustfs#6025: GET on a transitioned managed-SSE object silently returns corrupt bytes (fails with enforcement on AND off, so it is not an authorization regression); un-ignore with the fix"]
+async fn ilm_transition_on_sse_kms_bucket_under_enforcement_reads_back() -> TestResult {
+    init_logging();
+
+    // Cold-tier server: independent credentials, its own Local KMS holding the
+    // same key id (see the module docs for why the tier target needs KMS).
+    // Started first; each server's startup cleanup only matches its own unique
+    // address and temp dir, so the two instances coexist.
+    let mut cold = LocalKMSTestEnvironment::new().await?;
+    cold.base_env.access_key = "kmscoldtieradmin".to_string();
+    cold.base_env.secret_key = "kmscoldtiersecret".to_string();
+    start_cold_tier_kms_server(&mut cold).await?;
+    let cold_client = cold.base_env.create_s3_client();
+    cold_client.create_bucket().bucket(TIER_BUCKET).send().await?;
+
+    // Hot server: Local KMS + enforcement + accelerated lifecycle clock.
+    let mut env = LocalKMSTestEnvironment::new().await?;
+    start_enforcing_ilm_server(&mut env).await?;
+    let hot_client = env.base_env.create_s3_client();
+
+    add_rustfs_tier(&env.base_env, &cold.base_env).await?;
+
+    env.base_env.create_test_bucket(TRANSITION_BUCKET).await?;
+    set_bucket_default_sse_kms(&hot_client, TRANSITION_BUCKET).await?;
+
+    hot_client
+        .put_object()
+        .bucket(TRANSITION_BUCKET)
+        .key(TRANSITION_KEY)
+        .body(ByteStream::from_static(PAYLOAD))
+        .send()
+        .await?;
+    assert_head_sse_kms(&hot_client, TRANSITION_BUCKET, TRANSITION_KEY).await?;
+    info!("object stored SSE-KMS encrypted under enforcement");
+
+    // Days=1 is due ~2s after the write with RUSTFS_ILM_DEBUG_DAY_SECS=2.
+    put_transition_rule(&hot_client, TRANSITION_BUCKET, "kms-ilm-transition", "tier/", 1).await?;
+
+    // Drive the transition deterministically via the manual run endpoint, then
+    // wait for HEAD to report the tier as the object's storage class.
+    run_manual_transition_until_processed(&env.base_env, TRANSITION_BUCKET, "tier/", ILM_DEADLINE).await?;
+    wait_for_transition(&hot_client, TRANSITION_BUCKET, TRANSITION_KEY, ILM_DEADLINE).await?;
+    info!("SSE-KMS object transitioned to the remote tier under enforcement");
+
+    let head = hot_client
+        .head_object()
+        .bucket(TRANSITION_BUCKET)
+        .key(TRANSITION_KEY)
+        .send()
+        .await?;
+    assert!(
+        head.restore().is_none(),
+        "a freshly transitioned object must not advertise x-amz-restore, got {:?}",
+        head.restore()
+    );
+
+    // The remote copy exists on the cold tier. The payload the tier holds is the
+    // hot server's stored ciphertext, wrapped once more under the cold server's
+    // own managed SSE-KMS layer (the forwarded headers re-request encryption).
+    let remote = cold_client.list_objects_v2().bucket(TIER_BUCKET).send().await?;
+    assert!(!remote.contents().is_empty(), "cold-tier bucket must hold the transitioned object's data");
+
+    // Read-through GET under enforcement must succeed (not AccessDenied) and
+    // keep advertising SSE-KMS. Its BODY is deliberately not compared here:
+    // the transitioned read path skips managed-SSE decryption — a product gap
+    // unrelated to enforcement — so a direct GET streams the stored ciphertext
+    // (`new_getobjectreader` in crates/ecstore/src/client/object_api_utils.rs
+    // hardcodes `is_encrypted = false` and never applies the
+    // `ReadTransform::Encrypted` wrapping the hot-read path builds in
+    // crates/ecstore/src/object_api/readers.rs). Plaintext recovery is pinned
+    // through restore semantics below; when the read-through gap is fixed, a
+    // byte assertion can be added here too.
+    let read_through = hot_client
+        .get_object()
+        .bucket(TRANSITION_BUCKET)
+        .key(TRANSITION_KEY)
+        .send()
+        .await?;
+    assert_eq!(
+        read_through.server_side_encryption(),
+        Some(&ServerSideEncryption::AwsKms),
+        "transitioned object must still report SSE-KMS on read-through"
+    );
+    let read_through_body = read_through.body.collect().await?.into_bytes();
+    assert_eq!(
+        read_through_body.len(),
+        PAYLOAD.len(),
+        "read-through GET must stream the object's full logical size under enforcement"
+    );
+
+    // RestoreObject copies the ciphertext back from the tier under the original
+    // envelope metadata; the restored copy is then served by the normal
+    // decrypting read path. The copy-back runs with an internal (no-principal)
+    // identity, so this also pins the exemption on the restore path. Days=300
+    // because RUSTFS_ILM_DEBUG_DAY_SECS=2 accelerates the restored copy's
+    // expiry as well (300 accelerated days == 600s of validity).
+    hot_client
+        .restore_object()
+        .bucket(TRANSITION_BUCKET)
+        .key(TRANSITION_KEY)
+        .restore_request(RestoreRequest::builder().days(300).build())
+        .send()
+        .await?;
+    wait_for_restore_complete(&hot_client, TRANSITION_BUCKET, TRANSITION_KEY, ILM_DEADLINE).await?;
+    info!("SSE-KMS object restored from the remote tier under enforcement");
+
+    // The KMS-relevant half: the restored envelope decrypts back to the exact
+    // plaintext for the requesting principal.
+    let restored = hot_client
+        .get_object()
+        .bucket(TRANSITION_BUCKET)
+        .key(TRANSITION_KEY)
+        .send()
+        .await?;
+    assert_eq!(
+        restored.server_side_encryption(),
+        Some(&ServerSideEncryption::AwsKms),
+        "restored object must still report SSE-KMS"
+    );
+    let body = restored.body.collect().await?.into_bytes();
+    assert_eq!(body.as_ref(), PAYLOAD, "restored SSE-KMS object must round-trip byte-identical plaintext");
+
+    Ok(())
+}
@@ -59,3 +59,6 @@ mod configured_roundtrip_test;

 #[cfg(test)]
 mod kms_authorization_negative_matrix_test;
+
+#[cfg(test)]
+mod kms_ilm_sse_kms_test;
@@ -39,6 +39,10 @@ pub mod fault_proxy;
 #[cfg(test)]
 mod reliability_disk_fault_test;

+// Privileged Linux-only 3x4 replacement rebuild proof for rustfs#5869/#1791.
+#[cfg(all(test, target_os = "linux"))]
+mod replacement_privileged_e2e_test;
+
 // dist-13 (backlog#1150/#1155): e2e regression net proving a large-object
 // degraded EC read never returns a silently truncated body (rustfs#4594/#4560/#4585).
 #[cfg(test)]
@@ -2854,7 +2854,7 @@ pub(crate) mod cmptst_30 {
        result
    }

-    #[ignore]
+    #[ignore = "timing-sensitive backend-pressure latency probe; run explicitly with --ignored"]
    #[tokio::test]
    async fn regression() -> Result<(), Box<dyn std::error::Error + Send + Sync>> {
        crate::common::init_logging();
@@ -252,6 +252,7 @@ impl QuotaTestEnv {
 #[cfg(test)]
 mod integration_tests {
    use super::*;
+    use aws_sdk_s3::error::ProvideErrorMetadata;

    #[tokio::test]
    #[serial]
@@ -963,9 +964,27 @@ mod integration_tests {
            .send()
            .await;

-        assert!(complete_result.is_err());
+        let complete_error = complete_result.expect_err("multipart completion above quota must be rejected");
+        assert_eq!(complete_error.as_service_error().and_then(|error| error.code()), Some("InvalidRequest"));
        assert!(!env.object_exists("over_quota.txt").await?);

+        let staged_parts = env
+            .client
+            .list_parts()
+            .bucket(&env.bucket_name)
+            .key("over_quota.txt")
+            .upload_id(upload_id2)
+            .send()
+            .await?;
+        assert_eq!(staged_parts.parts().len(), 2, "quota rejection must preserve the multipart upload");
+        env.client
+            .abort_multipart_upload()
+            .bucket(&env.bucket_name)
+            .key("over_quota.txt")
+            .upload_id(upload_id2)
+            .send()
+            .await?;
+
        env.cleanup_bucket().await?;

        Ok(())
@@ -349,11 +349,32 @@ mod tests {
            .send()
            .await?;

+        let first_inline = client
+            .put_object()
+            .bucket(bucket)
+            .key("versions/inline.bin")
+            .body(ByteStream::from(payload(8 * 1024, 40)))
+            .send()
+            .await?;
+        let first_inline_version = first_inline
+            .version_id()
+            .ok_or("first inline PUT did not return a version ID")?;
+        let second_inline = client
+            .put_object()
+            .bucket(bucket)
+            .key("versions/inline.bin")
+            .body(ByteStream::from(payload(8 * 1024, 41)))
+            .send()
+            .await?;
+        let second_inline_version = second_inline
+            .version_id()
+            .ok_or("second inline PUT did not return a version ID")?;
+
        let first = client
            .put_object()
            .bucket(bucket)
            .key(key)
-            .body(ByteStream::from(payload(256 * 1024, 41)))
+            .body(ByteStream::from(payload(128 * 1024, 41)))
            .send()
            .await?;
        let first_version = first.version_id().ok_or("first PUT did not return a version ID")?;
@@ -361,16 +382,36 @@ mod tests {
            .put_object()
            .bucket(bucket)
            .key(key)
-            .body(ByteStream::from(payload(256 * 1024, 42)))
+            .body(ByteStream::from(payload(3 * 1024 * 1024, 42)))
            .send()
            .await?;
        let second_version = second.version_id().ok_or("second PUT did not return a version ID")?;
        let delete = client.delete_object().bucket(bucket).key(key).send().await?;
        let delete_version = delete.version_id().ok_or("delete marker did not return a version ID")?;

+        let first_inline_census = harness.census_object_version(0, bucket, "versions/inline.bin", Some(first_inline_version))?;
+        let second_inline_census =
+            harness.census_object_version(0, bucket, "versions/inline.bin", Some(second_inline_version))?;
        let first_census = harness.census_object_version(0, bucket, key, Some(first_version))?;
+        let first_other_disk_census = harness.census_object_version(1, bucket, key, Some(first_version))?;
        let second_census = harness.census_object_version(0, bucket, key, Some(second_version))?;
        let delete_census = harness.census_object_version(0, bucket, key, Some(delete_version))?;
+        assert!(
+            first_inline_census.is_complete() && second_inline_census.is_complete(),
+            "inline version physical census is incomplete: first={first_inline_census:?} second={second_inline_census:?}"
+        );
+        assert!(
+            first_inline_census.present_part_fingerprints.is_empty() && second_inline_census.present_part_fingerprints.is_empty(),
+            "inline versions must not select external shard files: first={first_inline_census:?} second={second_inline_census:?}"
+        );
+        assert!(
+            first_inline_census.inline_data_fingerprint.is_some() && second_inline_census.inline_data_fingerprint.is_some(),
+            "inline versions must fingerprint payload bytes stored in xl.meta"
+        );
+        assert_ne!(
+            first_inline_census.inline_data_fingerprint, second_inline_census.inline_data_fingerprint,
+            "same-size inline versions with different payloads must retain distinct xl.meta fingerprints"
+        );
        assert!(
            first_census.is_complete(),
            "first version physical census is incomplete: {first_census:?}"
@@ -379,6 +420,14 @@ mod tests {
            second_census.is_complete(),
            "second version physical census is incomplete: {second_census:?}"
        );
+        assert!(
+            first_other_disk_census.is_complete(),
+            "first version physical census on the second disk is incomplete: {first_other_disk_census:?}"
+        );
+        assert_ne!(
+            first_census.erasure_index, first_other_disk_census.erasure_index,
+            "physical census must preserve each disk's erasure index"
+        );
        assert_ne!(
            first_census.data_dir, second_census.data_dir,
            "distinct object versions must select distinct physical data directories"
@@ -387,6 +436,24 @@ mod tests {
            first_census.expected_part_numbers, second_census.expected_part_numbers,
            "same single-part shape should expose the same part numbers"
        );
+        let first_part = first_census
+            .present_part_fingerprints
+            .values()
+            .next()
+            .ok_or("first version did not expose a physical part fingerprint")?;
+        let second_part = second_census
+            .present_part_fingerprints
+            .values()
+            .next()
+            .ok_or("second version did not expose a physical part fingerprint")?;
+        assert_ne!(
+            first_part.size, second_part.size,
+            "different shard lengths must retain their physical sizes"
+        );
+        assert_ne!(
+            first_part.sha256, second_part.sha256,
+            "different shard contents must retain their physical hashes"
+        );
        assert!(
            delete_census.is_complete(),
            "delete marker physical census is incomplete: {delete_census:?}"
@@ -396,7 +463,7 @@ mod tests {
            "delete marker must not declare object shards: {delete_census:?}"
        );
        assert!(
-            delete_census.present_part_numbers.is_empty(),
+            delete_census.present_part_fingerprints.is_empty(),
            "delete marker must not select stale object shards: {delete_census:?}"
        );
        Ok(())
@@ -2401,15 +2401,20 @@ async fn wait_for_site_replication_info<F>(
 where
    F: Fn(&SiteReplicationInfo) -> bool,
 {
-    for _ in 0..40 {
+    // 30s to match wait_for_replication_state: the three-node site tests run
+    // several full rustfs processes on one runner, so peer-state propagation
+    // can take well over 10s under CI load.
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(30);
+    loop {
        let info = site_replication_info(env).await?;
        if predicate(&info) {
            return Ok(info);
        }
+        if tokio::time::Instant::now() >= deadline {
+            return Err(format!("site replication info did not reach expected state on {}", env.address).into());
+        }
        sleep(Duration::from_millis(250)).await;
    }
-
-    Err(format!("site replication info did not reach expected state on {}", env.address).into())
 }

 async fn wait_for_site_replication_status<F>(
@@ -2420,15 +2425,19 @@ async fn wait_for_site_replication_status<F>(
 where
    F: Fn(&SRStatusInfo) -> bool,
 {
-    for _ in 0..40 {
+    // Same 30s ceiling as wait_for_site_replication_info: the status probes
+    // fan out to every peer, so they see the same multi-process CI load.
+    let deadline = tokio::time::Instant::now() + Duration::from_secs(30);
+    loop {
        let status = site_replication_status(env, query).await?;
        if predicate(&status) {
            return Ok(status);
        }
+        if tokio::time::Instant::now() >= deadline {
+            return Err(format!("site replication status did not reach expected state on {}", env.address).into());
+        }
        sleep(Duration::from_millis(250)).await;
    }
-
-    Err(format!("site replication status did not reach expected state on {}", env.address).into())
 }

 async fn wait_for_replication_reset_target<F>(
@@ -4235,37 +4244,49 @@ async fn test_bucket_replication_acceptance_matrix_local_dual_targets() -> TestR
        "tag rule with disabled delete-marker replication created a marker: {tagged_state:?}"
    );

-    set_bucket_versioning(&source_env, source_bucket, BucketVersioningStatus::Suspended).await?;
-    set_bucket_versioning(&target_env_a, target_bucket_a, BucketVersioningStatus::Suspended).await?;
-    let null_put = source_client
+    // AWS S3 and MinIO both reject suspending versioning on a bucket that
+    // carries a replication configuration (InvalidBucketState): suspension
+    // would mint null versions that versioned replication can never converge.
+    let suspend_err = source_client
+        .put_bucket_versioning()
+        .bucket(source_bucket)
+        .versioning_configuration(
+            VersioningConfiguration::builder()
+                .status(BucketVersioningStatus::Suspended)
+                .build(),
+        )
+        .send()
+        .await
+        .expect_err("suspending versioning on a replication source must be rejected");
+    assert_eq!(
+        suspend_err.as_service_error().and_then(|error| error.code()),
+        Some("InvalidBucketState"),
+        "suspension on a replication source must fail with InvalidBucketState: {suspend_err:?}"
+    );
+
+    // The rejected suspension must leave the versioning + replication state
+    // fully intact: a fresh matched PUT still replicates with a real version.
+    let post_reject_put = source_client
        .put_object()
        .bucket(source_bucket)
-        .key("prefix/null.txt")
-        .body(ByteStream::from_static(b"null version"))
+        .key("prefix/after-rejected-suspend.txt")
+        .body(ByteStream::from_static(b"still replicating"))
        .send()
        .await?;
-    assert!(null_put.version_id().is_none(), "suspended source PUT must create a null version");
-    wait_for_replication_state(&target_client_a, target_bucket_a, "null version did not replicate", |state| {
-        state
-            .iter()
-            .any(|entry| entry.key == "prefix/null.txt" && entry.version_id == "null" && !entry.delete_marker)
-    })
-    .await?;
-    let null_delete = source_client
-        .delete_object()
-        .bucket(source_bucket)
-        .key("prefix/null.txt")
-        .send()
-        .await?;
-    assert!(
-        null_delete.version_id().is_none(),
-        "suspended source DELETE must create a null delete marker"
-    );
-    wait_for_replication_state(&target_client_a, target_bucket_a, "null delete marker did not replicate", |state| {
-        state
-            .iter()
-            .any(|entry| entry.key == "prefix/null.txt" && entry.version_id == "null" && entry.delete_marker)
-    })
+    let post_reject_version_id = post_reject_put
+        .version_id()
+        .ok_or("PUT after rejected suspension omitted version ID")?
+        .to_string();
+    wait_for_replication_state(
+        &target_client_a,
+        target_bucket_a,
+        "replication stopped after rejected versioning suspension",
+        |state| {
+            state
+                .iter()
+                .any(|entry| entry.key == "prefix/after-rejected-suspend.txt" && entry.version_id == post_reject_version_id)
+        },
+    )
    .await?;

    Ok(())
@@ -32,6 +32,11 @@ workspace = true

 [features]
 default = []
+# Compiles the controlled list-objects namespace-journal chaos injector into a
+# production binary (it is always available to tests). Off by default so the
+# RUSTFS_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_* env vars cannot rewrite journal
+# state in a stock build (backlog#1832).
+list-chaos = []
 rio-v2 = ["dep:rustfs-rio-v2"]
 hotpath = [
    "hotpath/hotpath",
@@ -69,6 +69,7 @@ fn build_non_inline_writers(config: &BenchConfig) -> Vec<Option<BitrotWriterWrap
 fn bench_single_block_non_inline_fast_path(c: &mut Criterion) {
    let configs = vec![
        BenchConfig::new(4 * 1024, 4, 2, 128 * 1024),
+        BenchConfig::new(16 * 1024, 4, 2, 128 * 1024),
        BenchConfig::new(64 * 1024, 4, 2, 128 * 1024),
        BenchConfig::new(128 * 1024, 4, 2, 128 * 1024),
    ];
@@ -112,7 +113,12 @@ fn bench_single_block_non_inline_fast_path(c: &mut Criterion) {
                rt.block_on(async {
                    erasure
                        .clone()
-                        .encode_single_block_non_inline(reader, &mut writers, config.data_shards)
+                        .encode_single_block_non_inline_with_size_hint(
+                            reader,
+                            &mut writers,
+                            config.data_shards,
+                            config.payload_size,
+                        )
                        .await
                        .expect("single block candidate benchmark");
                });
@@ -61,9 +61,11 @@ pub mod bucket {
                delete_manual_transition_scope_admission_if_current, load_manual_transition_job_record,
                load_manual_transition_job_record_with_etag, load_manual_transition_scope_admission,
                manual_transition_job_lease_expired, manual_transition_scope_admission_lease_expired,
-                manual_transition_scope_key, persist_manual_transition_job_progress, renew_manual_transition_job_lease,
-                request_manual_transition_job_cancel, save_manual_transition_job_record,
-                save_manual_transition_job_record_if_current, save_manual_transition_scope_admission_if_absent,
+                manual_transition_scope_key, persist_manual_transition_job_progress,
+                persist_manual_transition_job_progress_if_owned, renew_manual_transition_job_lease,
+                renew_manual_transition_job_lease_if_owned, request_manual_transition_job_cancel,
+                save_manual_transition_job_record, save_manual_transition_job_record_if_current,
+                save_manual_transition_scope_admission_if_absent, update_manual_transition_job_record,
            };
        }

@@ -308,6 +310,8 @@ pub mod config {
 }

 pub mod data_usage {
+    #[cfg(feature = "test-util")]
+    pub use crate::data_usage::seed_bucket_usage_memory_for_test;
    pub use crate::data_usage::{
        DATA_USAGE_CACHE_NAME, apply_bucket_usage_memory_overlay, compute_bucket_usage,
        init_compression_total_memory_from_backend, invalidate_admin_data_usage_snapshot_cache,
@@ -342,7 +346,7 @@ pub mod disk {
    }

    pub mod error {
-        pub use crate::disk::error::{BitrotErrorType, DiskError, Error, FileAccessDeniedWithContext, Result};
+        pub use crate::disk::error::{DiskError, Error, FileAccessDeniedWithContext, Result};
    }

    pub mod error_reduce {
@@ -409,10 +413,10 @@ pub mod object {
    pub use crate::object_api::{
        BLOCK_SIZE_V2, ERASURE_ALGORITHM, EncryptionResolutionError, EncryptionResolutionErrorKind, GetObjectBodyCacheHook,
        GetObjectBodyCacheHookLookup, GetObjectBodySource, GetObjectReader, NamespaceLockFence, ObjectEncryptionResolver,
-        ObjectInfo, ObjectLockConfigSnapshot, ObjectMutationHook, ObjectOptions, PutObjReader, RangedDecompressReader,
-        ReadEncryptionMaterial, ReadEncryptionMode, ReadEncryptionRequest, StreamConsumer, get_object_body_cache_plaintext_len,
-        lookup_get_object_body_cache_hook, register_get_object_body_cache_hook, register_object_mutation_hook,
-        unregister_get_object_body_cache_hook, unregister_object_mutation_hook,
+        ObjectInfo, ObjectLockConfigSnapshot, ObjectMutationHook, ObjectOptions, PutObjReader, QuotaAdmission,
+        RangedDecompressReader, ReadEncryptionMaterial, ReadEncryptionMode, ReadEncryptionRequest, StreamConsumer,
+        get_object_body_cache_plaintext_len, lookup_get_object_body_cache_hook, register_get_object_body_cache_hook,
+        register_object_mutation_hook, unregister_get_object_body_cache_hook, unregister_object_mutation_hook,
    };
    pub use crate::store::{
        PrepareSelectObjectSnapshotError, PreparedGetObjectReader, SelectObjectSnapshot, SelectObjectSnapshotReadError,
@@ -64,10 +64,41 @@ impl BucketDurabilityConfig {
    }
 }

+/// Default durability tier seeded into a newly created bucket's metadata
+/// (rustfs/backlog#1811). `relaxed` aligns new buckets with MinIO's default
+/// posture: object data is still fdatasynced, while xl.meta and directory-entry
+/// fsyncs follow the relaxed durability gate.
+pub const ENV_NEW_BUCKET_DURABILITY_MODE: &str = "RUSTFS_NEW_BUCKET_DURABILITY_MODE";
+pub const DEFAULT_NEW_BUCKET_DURABILITY_MODE: &str = BUCKET_DURABILITY_MODE_RELAXED;
+
+/// The `durability.json` bytes to seed into a freshly created bucket's metadata.
+/// Empty means "no override" (the bucket then follows the global
+/// `RUSTFS_DURABILITY_MODE`); otherwise the serialized chosen tier. Operators
+/// can set `inherit` to disable the new-bucket override. Invalid values also
+/// fail closed to inherit the global mode instead of seeding a surprising tier.
+pub fn new_bucket_durability_config_json() -> Vec<u8> {
+    let raw = std::env::var(ENV_NEW_BUCKET_DURABILITY_MODE).unwrap_or_else(|_| DEFAULT_NEW_BUCKET_DURABILITY_MODE.to_string());
+    let mode = raw.trim();
+    if mode.eq_ignore_ascii_case("inherit") || mode.is_empty() || !BucketDurabilityConfig::is_valid_mode(mode) {
+        return Vec::new();
+    }
+    serde_json::to_vec(&BucketDurabilityConfig::new(mode)).expect("BucketDurabilityConfig serialization cannot fail")
+}
+
 #[cfg(test)]
 mod tests {
    use super::*;

+    fn new_bucket_seeded_mode() -> Option<String> {
+        let json = new_bucket_durability_config_json();
+        if json.is_empty() {
+            return None;
+        }
+        serde_json::from_slice::<BucketDurabilityConfig>(&json)
+            .expect("new-bucket durability config must serialize")
+            .normalized_mode()
+    }
+
    #[test]
    fn valid_modes_are_recognized() {
        assert!(BucketDurabilityConfig::is_valid_mode("strict"));
@@ -99,4 +130,33 @@ mod tests {
        let empty: BucketDurabilityConfig = serde_json::from_slice(b"{}").expect("deserialize empty");
        assert_eq!(empty.normalized_mode(), None);
    }
+
+    #[test]
+    fn new_bucket_default_seeds_relaxed_when_unset() {
+        temp_env::with_var_unset(ENV_NEW_BUCKET_DURABILITY_MODE, || {
+            assert_eq!(new_bucket_seeded_mode().as_deref(), Some(BUCKET_DURABILITY_MODE_RELAXED));
+        });
+    }
+
+    #[test]
+    fn new_bucket_default_honors_explicit_tiers() {
+        for mode in [
+            BUCKET_DURABILITY_MODE_STRICT,
+            BUCKET_DURABILITY_MODE_RELAXED,
+            BUCKET_DURABILITY_MODE_NONE,
+        ] {
+            temp_env::with_var(ENV_NEW_BUCKET_DURABILITY_MODE, Some(mode), || {
+                assert_eq!(new_bucket_seeded_mode().as_deref(), Some(mode));
+            });
+        }
+    }
+
+    #[test]
+    fn new_bucket_default_can_inherit_global_mode() {
+        for mode in ["inherit", "", "bogus"] {
+            temp_env::with_var(ENV_NEW_BUCKET_DURABILITY_MODE, Some(mode), || {
+                assert_eq!(new_bucket_seeded_mode(), None);
+            });
+        }
+    }
 }
@@ -86,6 +86,21 @@ where
    com::save_config_with_opts(api, file, data, opts).await
 }

+pub(crate) async fn save_config_with_opts_quiet<S>(api: Arc<S>, file: &str, data: Vec<u8>, opts: &ObjectOptions) -> Result<()>
+where
+    S: ObjectIO<
+            Error = Error,
+            RangeSpec = HTTPRangeSpec,
+            HeaderMap = HeaderMap,
+            ObjectOptions = ObjectOptions,
+            ObjectInfo = ObjectInfo,
+            GetObjectReader = GetObjectReader,
+            PutObjectReader = PutObjReader,
+        >,
+{
+    com::save_config_with_opts_quiet(api, file, data, opts).await
+}
+
 pub(crate) async fn delete_config<S>(api: Arc<S>, file: &str) -> Result<()>
 where
    S: ObjectOperations<
@@ -45,6 +45,104 @@ const MANUAL_TRANSITION_JOB_LEASE_SECONDS: i128 = 60;
 const MANUAL_TRANSITION_LEGACY_SCOPE_SCAN_LIMIT: i32 = 1000;
 const MANUAL_TRANSITION_TASK_SCAN_LIMIT: i32 = 1000;
 const MANUAL_TRANSITION_WORKER_RESULT_SCAN_LIMIT: i32 = 1000;
+const MANUAL_TRANSITION_JOB_CAS_RETRIES: usize = 4;
+
+#[cfg(test)]
+struct ManualTransitionJobCasBarrierState {
+    job_id: Uuid,
+    paused: std::sync::atomic::AtomicBool,
+    arrived: tokio::sync::Notify,
+    release: tokio::sync::Semaphore,
+}
+
+#[cfg(test)]
+pub(crate) struct ManualTransitionJobCasBarrier {
+    state: Arc<ManualTransitionJobCasBarrierState>,
+}
+
+#[cfg(test)]
+static MANUAL_TRANSITION_JOB_CAS_BARRIER: std::sync::OnceLock<std::sync::Mutex<Option<Arc<ManualTransitionJobCasBarrierState>>>> =
+    std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl ManualTransitionJobCasBarrier {
+    pub(crate) fn install(job_id: Uuid) -> Self {
+        let state = Arc::new(ManualTransitionJobCasBarrierState {
+            job_id,
+            paused: std::sync::atomic::AtomicBool::new(false),
+            arrived: tokio::sync::Notify::new(),
+            release: tokio::sync::Semaphore::new(0),
+        });
+        let mut slot = MANUAL_TRANSITION_JOB_CAS_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("manual transition progress CAS barrier mutex should not poison");
+        assert!(
+            slot.is_none(),
+            "manual transition job CAS barrier must be installed by one test at a time"
+        );
+        *slot = Some(Arc::clone(&state));
+        drop(slot);
+        Self { state }
+    }
+
+    pub(crate) async fn wait_until_paused(&self) {
+        tokio::time::timeout(std::time::Duration::from_secs(30), async {
+            loop {
+                let arrived = self.state.arrived.notified();
+                if self.state.paused.load(std::sync::atomic::Ordering::Acquire) {
+                    return;
+                }
+                arrived.await;
+            }
+        })
+        .await
+        .expect("manual transition job update should reach the deterministic CAS barrier");
+    }
+
+    pub(crate) fn release(&self) {
+        self.state.release.add_permits(1);
+    }
+}
+
+#[cfg(test)]
+impl Drop for ManualTransitionJobCasBarrier {
+    fn drop(&mut self) {
+        self.release();
+        let mut slot = MANUAL_TRANSITION_JOB_CAS_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("manual transition progress CAS barrier mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+async fn pause_manual_transition_job_before_first_cas(job_id: Uuid) {
+    let barrier = MANUAL_TRANSITION_JOB_CAS_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("manual transition progress CAS barrier mutex should not poison")
+        .as_ref()
+        .filter(|barrier| barrier.job_id == job_id)
+        .cloned();
+    if let Some(barrier) = barrier
+        && barrier
+            .paused
+            .compare_exchange(false, true, std::sync::atomic::Ordering::AcqRel, std::sync::atomic::Ordering::Acquire)
+            .is_ok()
+    {
+        barrier.arrived.notify_one();
+        barrier
+            .release
+            .acquire()
+            .await
+            .expect("manual transition job CAS barrier should remain open")
+            .forget();
+    }
+}

 fn is_false(value: &bool) -> bool {
    !*value
@@ -148,7 +246,6 @@ impl ManualTransitionJobRecord {

    pub fn fail(&mut self, error: impl Into<String>) {
        self.state = ManualTransitionJobState::Failed;
-        self.report.tier_failure = self.report.tier_failure.saturating_add(1);
        self.error = Some(error.into());
        self.mark_updated_terminal();
    }
@@ -1040,7 +1137,7 @@ pub async fn save_manual_transition_job_record_if_current(
    }
    let object = manual_transition_job_record_object_name(job.job_id).map_err(manual_transition_job_store_error)?;
    let data = job.encode().map_err(manual_transition_job_store_error)?;
-    config_boundary::save_config_with_opts(
+    config_boundary::save_config_with_opts_quiet(
        api,
        &object,
        data,
@@ -1056,6 +1153,54 @@ pub async fn save_manual_transition_job_record_if_current(
    .await
 }

+/// Applies a job-record mutation with optimistic concurrency control.
+///
+/// The mutation returns whether the record needs to be persisted. When a lease
+/// is supplied, ownership is checked again after every conflicting write.
+pub async fn update_manual_transition_job_record<F>(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Option<Uuid>,
+    update: F,
+) -> EcstoreResult<ManualTransitionJobRecord>
+where
+    F: FnMut(&mut ManualTransitionJobRecord) -> bool,
+{
+    update_manual_transition_job_record_from(api, job_id, expected_lease_id, None, update).await
+}
+
+async fn update_manual_transition_job_record_from<F>(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Option<Uuid>,
+    mut current: Option<(ManualTransitionJobRecord, String)>,
+    mut update: F,
+) -> EcstoreResult<ManualTransitionJobRecord>
+where
+    F: FnMut(&mut ManualTransitionJobRecord) -> bool,
+{
+    for _ in 0..MANUAL_TRANSITION_JOB_CAS_RETRIES {
+        let (mut record, etag) = match current.take() {
+            Some(current) => current,
+            None => load_manual_transition_job_record_with_etag(api.clone(), job_id).await?,
+        };
+        if expected_lease_id.is_some_and(|lease_id| record.lease_id != lease_id) {
+            return Err(Error::PreconditionFailed);
+        }
+        if !update(&mut record) {
+            return Ok(record);
+        }
+        #[cfg(test)]
+        pause_manual_transition_job_before_first_cas(job_id).await;
+        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
+            Ok(()) => return Ok(record),
+            Err(Error::PreconditionFailed) => continue,
+            Err(err) => return Err(err),
+        }
+    }
+    Err(Error::PreconditionFailed)
+}
+
 pub(crate) async fn save_manual_transition_worker_result_if_absent(
    api: Arc<ECStore>,
    record: &ManualTransitionWorkerResultRecord,
@@ -1314,99 +1459,113 @@ pub async fn reconcile_manual_transition_worker_results(
    api: Arc<ECStore>,
    job_id: Uuid,
    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    reconcile_manual_transition_worker_results_inner(api, job_id, None, queue_snapshot, false).await
+}
+
+pub(crate) async fn reconcile_manual_transition_worker_results_if_owned(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Uuid,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    reconcile_manual_transition_worker_results_inner(api, job_id, Some(expected_lease_id), queue_snapshot, false).await
+}
+
+async fn reconcile_manual_transition_worker_results_inner(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Option<Uuid>,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+    mark_missing_results_unknown: bool,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
    let task_stats = match scan_manual_transition_task_journal(api.clone(), job_id).await? {
        ManualTransitionTaskJournal::Stats(stats) => stats,
        ManualTransitionTaskJournal::Corrupt(error) => {
-            return mark_manual_transition_job_unknown_for_task_journal_error(api, job_id, error, queue_snapshot).await;
+            return mark_manual_transition_job_unknown_for_task_journal_error(
+                api,
+                job_id,
+                expected_lease_id,
+                error,
+                queue_snapshot,
+            )
+            .await;
        }
    };
    let stats = match scan_manual_transition_worker_result_journal(api.clone(), job_id).await? {
        ManualTransitionWorkerResultJournal::Stats(stats) => stats,
        ManualTransitionWorkerResultJournal::Corrupt(error) => {
-            return mark_manual_transition_job_unknown_for_worker_result_journal_error(api, job_id, error, queue_snapshot).await;
+            return mark_manual_transition_job_unknown_for_worker_result_journal_error(
+                api,
+                job_id,
+                expected_lease_id,
+                error,
+                queue_snapshot,
+            )
+            .await;
        }
    };
-    for _ in 0..4 {
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-        let changed = record.apply_worker_result_counts(
+    let mut changed = false;
+    let record = update_manual_transition_job_record(api.clone(), job_id, expected_lease_id, |record| {
+        let counts_changed = record.apply_worker_result_counts(
            stats.stats.completed,
            stats.stats.failed,
            &stats.stats.tier_failure_by_reason,
            task_stats.queued,
            queue_snapshot,
        );
-        if !changed {
-            return Ok(record);
-        }
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
-            Ok(()) => {
-                if record.is_terminal() {
-                    delete_manual_transition_scope_admission_if_current(
-                        api.clone(),
-                        &record.scope_key,
-                        record.job_id,
-                        record.lease_id,
-                    )
-                    .await?;
-                } else {
-                    renew_manual_transition_scope_admission_from_job(api, &record).await?;
-                }
-                return Ok(record);
-            }
-            Err(Error::PreconditionFailed) => continue,
-            Err(err) => return Err(err),
-        }
+        let became_unknown = mark_missing_results_unknown && record.mark_unknown_if_worker_results_lost(queue_snapshot);
+        changed = counts_changed || became_unknown;
+        changed
+    })
+    .await?;
+    if !changed {
+        return Ok(record);
    }
-    Err(Error::PreconditionFailed)
+    if record.is_terminal() {
+        delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id).await?;
+    } else {
+        renew_manual_transition_scope_admission_from_job(api, &record).await?;
+    }
+    Ok(record)
 }

 async fn mark_manual_transition_job_unknown_for_task_journal_error(
    api: Arc<ECStore>,
    job_id: Uuid,
+    expected_lease_id: Option<Uuid>,
    error: String,
    queue_snapshot: ManualTransitionQueueSnapshot,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
-    for _ in 0..4 {
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-        if !record.mark_unknown_for_task_journal_error(error.clone(), queue_snapshot) {
-            return Ok(record);
-        }
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
-            Ok(()) => {
-                delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id)
-                    .await?;
-                return Ok(record);
-            }
-            Err(Error::PreconditionFailed) => continue,
-            Err(err) => return Err(err),
-        }
+    let mut changed = false;
+    let record = update_manual_transition_job_record(api.clone(), job_id, expected_lease_id, |record| {
+        changed = record.mark_unknown_for_task_journal_error(error.clone(), queue_snapshot);
+        changed
+    })
+    .await?;
+    if changed && record.is_terminal() {
+        delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id).await?;
    }
-    Err(Error::PreconditionFailed)
+    Ok(record)
 }

 async fn mark_manual_transition_job_unknown_for_worker_result_journal_error(
    api: Arc<ECStore>,
    job_id: Uuid,
+    expected_lease_id: Option<Uuid>,
    error: String,
    queue_snapshot: ManualTransitionQueueSnapshot,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
-    for _ in 0..4 {
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-        if !record.mark_unknown_for_worker_result_journal_error(error.clone(), queue_snapshot) {
-            return Ok(record);
-        }
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
-            Ok(()) => {
-                delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id)
-                    .await?;
-                return Ok(record);
-            }
-            Err(Error::PreconditionFailed) => continue,
-            Err(err) => return Err(err),
-        }
+    let mut changed = false;
+    let record = update_manual_transition_job_record(api.clone(), job_id, expected_lease_id, |record| {
+        changed = record.mark_unknown_for_worker_result_journal_error(error.clone(), queue_snapshot);
+        changed
+    })
+    .await?;
+    if changed && record.is_terminal() {
+        delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id).await?;
    }
-    Err(Error::PreconditionFailed)
+    Ok(record)
 }

 pub async fn save_manual_transition_scope_admission_if_absent(
@@ -1603,19 +1762,14 @@ async fn find_active_legacy_manual_transition_scope_conflict(
 }

 pub async fn request_manual_transition_job_cancel(api: Arc<ECStore>, job_id: Uuid) -> EcstoreResult<ManualTransitionJobRecord> {
-    for _ in 0..4 {
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
+    update_manual_transition_job_record(api, job_id, None, |record| {
        if record.is_terminal() || record.cancel_requested {
-            return Ok(record);
+            return false;
        }
        record.mark_cancel_requested();
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
-            Ok(()) => return Ok(record),
-            Err(Error::PreconditionFailed) => continue,
-            Err(err) => return Err(err),
-        }
-    }
-    Err(Error::PreconditionFailed)
+        true
+    })
+    .await
 }

 pub async fn persist_manual_transition_job_progress(
@@ -1624,10 +1778,39 @@ pub async fn persist_manual_transition_job_progress(
    report: &ManualTransitionRunReport,
    queue_snapshot: ManualTransitionQueueSnapshot,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
-    let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-    record.update_running_progress(report.clone(), queue_snapshot);
-    save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await?;
-    renew_manual_transition_scope_admission_from_job(api, &record).await?;
+    let current = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
+    persist_manual_transition_job_progress_inner(api, job_id, current.0.lease_id, Some(current), report, queue_snapshot).await
+}
+
+pub async fn persist_manual_transition_job_progress_if_owned(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Uuid,
+    report: &ManualTransitionRunReport,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    persist_manual_transition_job_progress_inner(api, job_id, expected_lease_id, None, report, queue_snapshot).await
+}
+
+async fn persist_manual_transition_job_progress_inner(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Uuid,
+    current: Option<(ManualTransitionJobRecord, String)>,
+    report: &ManualTransitionRunReport,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    let record = update_manual_transition_job_record_from(api.clone(), job_id, Some(expected_lease_id), current, |record| {
+        if record.state != ManualTransitionJobState::Running {
+            return false;
+        }
+        record.update_running_progress(report.clone(), queue_snapshot);
+        true
+    })
+    .await?;
+    if record.state == ManualTransitionJobState::Running {
+        renew_manual_transition_scope_admission_from_job(api, &record).await?;
+    }
    Ok(record)
 }

@@ -1661,25 +1844,58 @@ pub async fn renew_manual_transition_job_lease(
    job_id: Uuid,
    queue_snapshot: ManualTransitionQueueSnapshot,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
-    let (mut record, mut etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-    if record.state == ManualTransitionJobState::Running {
-        if record.scan_completed && queue_snapshot.queued == 0 && queue_snapshot.active == 0 {
-            record = reconcile_manual_transition_worker_results(api.clone(), job_id, queue_snapshot).await?;
-            if record.is_terminal() || !record.report.worker_transition_pending() {
-                return Ok(record);
+    let current = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
+    renew_manual_transition_job_lease_inner(api, job_id, current.0.lease_id, Some(current), queue_snapshot).await
+}
+
+pub async fn renew_manual_transition_job_lease_if_owned(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Uuid,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    renew_manual_transition_job_lease_inner(api, job_id, expected_lease_id, None, queue_snapshot).await
+}
+
+async fn renew_manual_transition_job_lease_inner(
+    api: Arc<ECStore>,
+    job_id: Uuid,
+    expected_lease_id: Uuid,
+    current: Option<(ManualTransitionJobRecord, String)>,
+    queue_snapshot: ManualTransitionQueueSnapshot,
+) -> EcstoreResult<ManualTransitionJobRecord> {
+    let (current, current_etag) = match current {
+        Some(current) => current,
+        None => load_manual_transition_job_record_with_etag(api.clone(), job_id).await?,
+    };
+    if current.lease_id != expected_lease_id {
+        return Err(Error::PreconditionFailed);
+    }
+    if current.state != ManualTransitionJobState::Running {
+        return Ok(current);
+    }
+    if current.scan_completed && queue_snapshot.queued == 0 && queue_snapshot.active == 0 {
+        return reconcile_manual_transition_worker_results_inner(api, job_id, Some(expected_lease_id), queue_snapshot, true)
+            .await;
+    }
+    let record = update_manual_transition_job_record_from(
+        api.clone(),
+        job_id,
+        Some(expected_lease_id),
+        Some((current, current_etag)),
+        |record| {
+            if record.state != ManualTransitionJobState::Running {
+                return false;
            }
-            (record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-        }
-        let became_terminal = record.mark_unknown_if_worker_results_lost(queue_snapshot);
-        if !became_terminal {
            record.renew_lease(queue_snapshot);
-        }
-        save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await?;
-        if became_terminal {
-            delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id).await?;
-        } else {
-            renew_manual_transition_scope_admission_from_job(api, &record).await?;
-        }
+            true
+        },
+    )
+    .await?;
+    if record.is_terminal() {
+        delete_manual_transition_scope_admission_if_current(api, &record.scope_key, record.job_id, record.lease_id).await?;
+    } else if record.state == ManualTransitionJobState::Running {
+        renew_manual_transition_scope_admission_from_job(api, &record).await?;
    }
    Ok(record)
 }
@@ -1688,15 +1904,31 @@ async fn renew_manual_transition_scope_admission_from_job(
    api: Arc<ECStore>,
    record: &ManualTransitionJobRecord,
 ) -> EcstoreResult<()> {
-    if let Ok((admission, admission_etag)) =
-        load_manual_transition_scope_admission_with_etag(api.clone(), &record.scope_key).await
-        && admission.job_id == record.job_id
-        && admission.lease_id == record.lease_id
-    {
-        let renewed_admission = ManualTransitionScopeAdmission::from_job(record);
-        save_manual_transition_scope_admission_if_current(api, &renewed_admission, &admission_etag).await?;
+    for _ in 0..MANUAL_TRANSITION_JOB_CAS_RETRIES {
+        let (admission, admission_etag) =
+            match load_manual_transition_scope_admission_with_etag(api.clone(), &record.scope_key).await {
+                Ok(admission) => admission,
+                Err(Error::ConfigNotFound) => return Ok(()),
+                Err(err) => return Err(err),
+            };
+        if admission.job_id != record.job_id || admission.lease_id != record.lease_id {
+            return Err(Error::PreconditionFailed);
+        }
+        let mut renewed_admission = ManualTransitionScopeAdmission::from_job(record);
+        renewed_admission.lease_expires_at_unix_nanos = renewed_admission
+            .lease_expires_at_unix_nanos
+            .max(admission.lease_expires_at_unix_nanos);
+        renewed_admission.updated_at_unix_nanos = renewed_admission.updated_at_unix_nanos.max(admission.updated_at_unix_nanos);
+        if renewed_admission == admission {
+            return Ok(());
+        }
+        match save_manual_transition_scope_admission_if_current(api.clone(), &renewed_admission, &admission_etag).await {
+            Ok(()) => return Ok(()),
+            Err(Error::PreconditionFailed) => continue,
+            Err(err) => return Err(err),
+        }
    }
-    Ok(())
+    Err(Error::PreconditionFailed)
 }

 pub async fn delete_manual_transition_scope_admission_if_current(
@@ -2386,14 +2618,14 @@ mod tests {
    }

    #[test]
-    fn manual_transition_job_record_failure_counts_tier_failure() {
+    fn manual_transition_job_record_control_plane_failure_does_not_count_tier_failure() {
        let options = ManualTransitionRunOptions::default();
        let mut record = ManualTransitionJobRecord::new(Uuid::new_v4(), "bucket", &options, TEST_OWNER);

        record.fail("missing tier");

        assert_eq!(record.state, ManualTransitionJobState::Failed);
-        assert_eq!(record.report.tier_failure, 1);
+        assert_eq!(record.report.tier_failure, 0);
        assert_eq!(record.error.as_deref(), Some("missing tier"));
    }

@@ -18,6 +18,7 @@ use s3s::dto::{BucketLifecycleConfiguration, ObjectLockConfiguration};
 use time::OffsetDateTime;
 use uuid::Uuid;

+use crate::bucket::metadata::BucketMetadata;
 use crate::bucket::metadata_sys::{self, ObjectLockConfigState};
 use crate::error::{Error, Result};

@@ -26,16 +27,37 @@ pub(crate) struct LifecycleExpiryConfigs {
    pub(crate) lifecycle: Option<Arc<BucketLifecycleConfiguration>>,
    pub(crate) object_lock: Option<Arc<ObjectLockConfiguration>>,
    pub(crate) bucket_incarnation_id: Uuid,
+    pub(crate) table_bucket_enabled: bool,
 }

-pub(crate) async fn get_expiry_configs(api: &crate::store::ECStore, bucket: &str) -> Result<LifecycleExpiryConfigs> {
-    let bucket_incarnation_id = api.bucket_incarnation_id_from_disk(bucket).await?;
+async fn get_authoritative_metadata(
+    api: &crate::store::ECStore,
+    bucket: &str,
+    bucket_incarnation_id: Uuid,
+) -> Result<Arc<BucketMetadata>> {
    let sys = metadata_sys::bucket_metadata_sys_of(&api.ctx)?;
    let sys = sys.read().await.clone();
    let metadata = sys.get_authoritative_metadata(bucket).await?;
    if !metadata.bucket_incarnation_sidecar || metadata.bucket_incarnation_id != bucket_incarnation_id {
        return Err(Error::other(format!("bucket lifecycle metadata is not authoritative: {bucket}")));
    }
+    Ok(metadata)
+}
+
+pub(crate) async fn lifecycle_expiry_allowed(
+    api: &crate::store::ECStore,
+    bucket: &str,
+    bucket_incarnation_id: Uuid,
+) -> Result<bool> {
+    Ok(!get_authoritative_metadata(api, bucket, bucket_incarnation_id)
+        .await?
+        .table_bucket_enabled())
+}
+
+pub(crate) async fn get_expiry_configs(api: &crate::store::ECStore, bucket: &str) -> Result<LifecycleExpiryConfigs> {
+    let bucket_incarnation_id = api.bucket_incarnation_id_from_disk(bucket).await?;
+    let metadata = get_authoritative_metadata(api, bucket, bucket_incarnation_id).await?;
+    let table_bucket_enabled = metadata.table_bucket_enabled();

    let lifecycle = if metadata.lifecycle_config.is_none() && !metadata.lifecycle_config_xml.is_empty() {
        return Err(Error::other("persisted bucket lifecycle configuration is invalid"));
@@ -51,6 +73,7 @@ pub(crate) async fn get_expiry_configs(api: &crate::store::ECStore, bucket: &str
            lifecycle: None,
            object_lock: None,
            bucket_incarnation_id,
+            table_bucket_enabled,
        });
    }
    let object_lock = match metadata_sys::object_lock_config_state_from_authoritative_metadata(&metadata)? {
@@ -65,6 +88,7 @@ pub(crate) async fn get_expiry_configs(api: &crate::store::ECStore, bucket: &str
        lifecycle,
        object_lock,
        bucket_incarnation_id,
+        table_bucket_enabled,
    })
 }

@@ -125,6 +149,7 @@ mod tests {
        let lifecycle = lifecycle_config();
        metadata.lifecycle_config_xml = crate::bucket::utils::serialize(&lifecycle).unwrap();
        metadata.lifecycle_config = Some(lifecycle);
+        metadata.table_bucket_config_json = br#"{"enabled":true}"#.to_vec();
        metadata_sys::set_new_bucket_metadata_in(&store_a.ctx, metadata)
            .await
            .unwrap();
@@ -132,7 +157,14 @@ mod tests {
            .await
            .unwrap();

-        assert!(get_expiry_configs(&store_a, bucket).await.unwrap().lifecycle.is_some());
+        let configs = get_expiry_configs(&store_a, bucket).await.unwrap();
+        assert!(configs.lifecycle.is_some());
+        assert!(configs.table_bucket_enabled);
+        assert!(
+            !lifecycle_expiry_allowed(&store_a, bucket, configs.bucket_incarnation_id)
+                .await
+                .unwrap()
+        );
        assert!(get_expiry_configs(&store_b, bucket).await.unwrap().lifecycle.is_none());
    }
 }
@@ -425,6 +425,15 @@ impl BucketMetadata {
        }
    }

+    /// Metadata for a physically new user bucket. Existing or fabricated legacy
+    /// metadata must use [`Self::new`] so upgrades do not rewrite their
+    /// durability posture.
+    pub fn new_with_default_durability(name: &str) -> Self {
+        let mut metadata = Self::new(name);
+        metadata.durability_config_json = super::durability::new_bucket_durability_config_json();
+        metadata
+    }
+
    pub fn save_file_path(&self) -> String {
        format!("{}/{}/{}", BUCKET_META_PREFIX, self.name.as_str(), BUCKET_METADATA_FILE)
    }
@@ -1302,7 +1311,7 @@ mod test {
        assert!(bm.object_locking(), "object lock active via parsed config");
    }

-    /// backlog#580: KNOWN GAP (weisd 2026-03-06 "inline_data 前缀不同"). RustFS's
+    /// backlog#580: KNOWN GAP (flagged 2026-03-06: "inline_data 前缀不同"). RustFS's
    /// inline-data extraction does not yet recover the object body from a
    /// MinIO-written bucket-metadata object: `into_fileinfo(read_data=true).data`
    /// returns bytes that are not the `.metadata.bin` blob (no `format|version`
@@ -1310,7 +1319,7 @@ mod test {
    /// inline-data framing is handled on the read path.
    /// backlog#580: prove RustFS reads a MinIO-written **inlined** bucket-metadata
    /// object end-to-end. MinIO stores inline data as `[bitrot hash][object body]`
-    /// (the "`inline_data` 前缀不同" that weisd flagged on 2026-03-06 is that
+    /// (the "`inline_data` 前缀不同" gap flagged on 2026-03-06 is that
    /// bitrot prefix, not a format incompatibility). Running the raw inline shard
    /// through RustFS's `BitrotReader` with the default `HighwayHash256S` must
    /// verify the checksum and yield the exact `.metadata.bin` blob.
@@ -1378,6 +1387,43 @@ mod test {
        assert_ne!(old.bucket_incarnation_id, new.bucket_incarnation_id);
    }

+    #[test]
+    fn regular_bucket_metadata_constructor_does_not_seed_durability() {
+        temp_env::with_var_unset(crate::bucket::durability::ENV_NEW_BUCKET_DURABILITY_MODE, || {
+            let metadata = BucketMetadata::new("legacy-or-fabricated");
+            assert!(metadata.durability_config_json.is_empty());
+            assert!(metadata.durability_config().is_none());
+        });
+    }
+
+    #[test]
+    fn new_bucket_metadata_constructor_seeds_default_durability() {
+        temp_env::with_var_unset(crate::bucket::durability::ENV_NEW_BUCKET_DURABILITY_MODE, || {
+            let metadata = BucketMetadata::new_with_default_durability("new-user-bucket");
+            assert_eq!(
+                metadata.durability_config().and_then(|cfg| cfg.normalized_mode()).as_deref(),
+                Some(crate::bucket::durability::BUCKET_DURABILITY_MODE_RELAXED)
+            );
+
+            let encoded = metadata.marshal_msg().expect("marshal metadata");
+            let decoded = BucketMetadata::unmarshal(&encoded).expect("unmarshal metadata");
+            assert_eq!(decoded.durability_config_json, metadata.durability_config_json);
+            assert_eq!(
+                decoded.durability_config().and_then(|cfg| cfg.normalized_mode()).as_deref(),
+                Some(crate::bucket::durability::BUCKET_DURABILITY_MODE_RELAXED)
+            );
+        });
+    }
+
+    #[test]
+    fn new_bucket_metadata_constructor_can_inherit_global_durability() {
+        temp_env::with_var(crate::bucket::durability::ENV_NEW_BUCKET_DURABILITY_MODE, Some("inherit"), || {
+            let metadata = BucketMetadata::new_with_default_durability("strict-fleet-new-bucket");
+            assert!(metadata.durability_config_json.is_empty());
+            assert!(metadata.durability_config().is_none());
+        });
+    }
+
    #[test]
    fn site_replication_config_updates_cannot_replace_bucket_incarnation() {
        let mut metadata = BucketMetadata::new("site-replication-update");
@@ -288,6 +288,13 @@ pub(crate) fn bucket_metadata_sys_of(ctx: &crate::runtime::instance::InstanceCon
    get_bucket_metadata_sys()
 }

+pub(crate) fn require_bucket_metadata_sys_in(
+    ctx: &crate::runtime::instance::InstanceContext,
+) -> Result<Arc<RwLock<BucketMetadataSys>>> {
+    ctx.bucket_metadata_sys()
+        .ok_or_else(|| Error::other("bucket metadata sys not initialized for this instance"))
+}
+
 pub(crate) async fn object_store_in(ctx: &crate::runtime::instance::InstanceContext) -> Result<Arc<ECStore>> {
    let sys = bucket_metadata_sys_of(ctx)?;
    Ok(sys.read().await.api.clone())
@@ -376,6 +383,15 @@ pub async fn update(bucket: &str, config_file: &str, data: Vec<u8>) -> Result<Of
    Box::pin(update_with_sys(get_bucket_metadata_sys()?, bucket, config_file, data)).await
 }

+pub(crate) async fn update_in(
+    ctx: &crate::runtime::instance::InstanceContext,
+    bucket: &str,
+    config_file: &str,
+    data: Vec<u8>,
+) -> Result<OffsetDateTime> {
+    Box::pin(update_with_sys(require_bucket_metadata_sys_in(ctx)?, bucket, config_file, data)).await
+}
+
 pub async fn delete(bucket: &str, config_file: &str) -> Result<OffsetDateTime> {
    delete_with_sys(get_bucket_metadata_sys()?, bucket, config_file).await
 }
@@ -14,6 +14,7 @@

 use super::metadata_sys::get_bucket_metadata_sys;
 use crate::error::{Result, StorageError};
+use crate::store::ECStore;
 use rustfs_policy::policy::{BucketPolicy, BucketPolicyArgs};

 pub struct PolicySys {}
@@ -27,6 +28,10 @@ impl PolicySys {
        Self::is_allowed_with_policy(args, Self::get(args.bucket).await).await
    }

+    pub async fn try_is_allowed_for_store(store: &ECStore, args: &BucketPolicyArgs<'_>) -> Result<bool> {
+        Self::is_allowed_with_policy(args, store.get_bucket_policy(args.bucket).await.map(|(policy, _)| policy)).await
+    }
+
    async fn is_allowed_with_policy(args: &BucketPolicyArgs<'_>, policy: Result<BucketPolicy>) -> Result<bool> {
        match policy {
            Ok(policy) => Ok(policy.is_allowed(args).await),
@@ -1,171 +0,0 @@
-// 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.
-#![allow(unused_imports)]
-#![allow(unused_variables)]
-#![allow(unused_mut)]
-#![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
-
-use http::{HeaderMap, StatusCode};
-use http_body_util::BodyExt;
-use hyper::body::Body;
-use hyper::body::Bytes;
-use std::collections::HashMap;
-
-use crate::client::{
-    api_error_response::http_resp_to_error_response,
-    transition_api::{ReaderImpl, RequestMetadata, TransitionClient},
-};
-use rustfs_utils::hash::EMPTY_STRING_SHA256_HASH;
-
-impl TransitionClient {
-    pub async fn set_bucket_policy(&self, bucket_name: &str, policy: &str) -> Result<(), std::io::Error> {
-        if policy == "" {
-            return self.remove_bucket_policy(bucket_name).await;
-        }
-
-        self.put_bucket_policy(bucket_name, policy).await
-    }
-
-    pub async fn put_bucket_policy(&self, bucket_name: &str, policy: &str) -> Result<(), std::io::Error> {
-        let mut url_values = HashMap::new();
-        url_values.insert("policy".to_string(), "".to_string());
-
-        let mut req_metadata = RequestMetadata {
-            bucket_name: bucket_name.to_string(),
-            query_values: url_values,
-            content_body: ReaderImpl::Body(Bytes::from(policy.as_bytes().to_vec())),
-            content_length: policy.len() as i64,
-            object_name: "".to_string(),
-            custom_header: HeaderMap::new(),
-            content_md5_base64: "".to_string(),
-            content_sha256_hex: "".to_string(),
-            stream_sha256: false,
-            trailer: HeaderMap::new(),
-            pre_sign_url: Default::default(),
-            add_crc: Default::default(),
-            extra_pre_sign_header: Default::default(),
-            bucket_location: Default::default(),
-            expires: Default::default(),
-        };
-
-        let resp = self.execute_method(http::Method::PUT, &mut req_metadata).await?;
-        //defer closeResponse(resp)
-
-        let resp_status = resp.status();
-        let h = resp.headers().clone();
-
-        //if resp != nil {
-        if resp_status != StatusCode::NO_CONTENT && resp.status() != StatusCode::OK {
-            return Err(std::io::Error::other(http_resp_to_error_response(
-                resp_status,
-                &h,
-                vec![],
-                bucket_name,
-                "",
-            )));
-        }
-        //}
-        Ok(())
-    }
-
-    pub async fn remove_bucket_policy(&self, bucket_name: &str) -> Result<(), std::io::Error> {
-        let mut url_values = HashMap::new();
-        url_values.insert("policy".to_string(), "".to_string());
-
-        let resp = self
-            .execute_method(
-                http::Method::DELETE,
-                &mut RequestMetadata {
-                    bucket_name: bucket_name.to_string(),
-                    query_values: url_values,
-                    content_sha256_hex: EMPTY_STRING_SHA256_HASH.to_string(),
-                    object_name: "".to_string(),
-                    custom_header: HeaderMap::new(),
-                    content_body: ReaderImpl::Body(Bytes::new()),
-                    content_length: 0,
-                    content_md5_base64: "".to_string(),
-                    stream_sha256: false,
-                    trailer: HeaderMap::new(),
-                    pre_sign_url: Default::default(),
-                    add_crc: Default::default(),
-                    extra_pre_sign_header: Default::default(),
-                    bucket_location: Default::default(),
-                    expires: Default::default(),
-                },
-            )
-            .await?;
-        //defer closeResponse(resp)
-
-        let resp_status = resp.status();
-        let h = resp.headers().clone();
-
-        if resp_status != StatusCode::NO_CONTENT {
-            return Err(std::io::Error::other(http_resp_to_error_response(
-                resp_status,
-                &h,
-                vec![],
-                bucket_name,
-                "",
-            )));
-        }
-
-        Ok(())
-    }
-
-    pub async fn get_bucket_policy(&self, bucket_name: &str) -> Result<String, std::io::Error> {
-        let bucket_policy = self.get_bucket_policy_inner(bucket_name).await?;
-        Ok(bucket_policy)
-    }
-
-    pub async fn get_bucket_policy_inner(&self, bucket_name: &str) -> Result<String, std::io::Error> {
-        let mut url_values = HashMap::new();
-        url_values.insert("policy".to_string(), "".to_string());
-
-        let resp = self
-            .execute_method(
-                http::Method::GET,
-                &mut RequestMetadata {
-                    bucket_name: bucket_name.to_string(),
-                    query_values: url_values,
-                    content_sha256_hex: EMPTY_STRING_SHA256_HASH.to_string(),
-                    object_name: "".to_string(),
-                    custom_header: HeaderMap::new(),
-                    content_body: ReaderImpl::Body(Bytes::new()),
-                    content_length: 0,
-                    content_md5_base64: "".to_string(),
-                    stream_sha256: false,
-                    trailer: HeaderMap::new(),
-                    pre_sign_url: Default::default(),
-                    add_crc: Default::default(),
-                    extra_pre_sign_header: Default::default(),
-                    bucket_location: Default::default(),
-                    expires: Default::default(),
-                },
-            )
-            .await?;
-
-        let mut body_vec = Vec::new();
-        let mut body = resp.into_body();
-        while let Some(frame) = body.frame().await {
-            let frame = frame.map_err(|e| std::io::Error::new(std::io::ErrorKind::Other, e.to_string()))?;
-            if let Some(data) = frame.data_ref() {
-                body_vec.extend_from_slice(data);
-            }
-        }
-        let policy = String::from_utf8_lossy(&body_vec).to_string();
-        Ok(policy)
-    }
-}
@@ -1,199 +0,0 @@
-// 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.
-#![allow(unused_imports)]
-#![allow(unused_variables)]
-#![allow(unused_mut)]
-#![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
-
-use crate::client::{
-    api_error_response::http_resp_to_error_response,
-    api_get_options::GetObjectOptions,
-    transition_api::{ObjectInfo, ReaderImpl, RequestMetadata, TransitionClient},
-};
-use bytes::Bytes;
-use http::{HeaderMap, HeaderValue};
-use http_body_util::BodyExt;
-use rustfs_config::MAX_S3_CLIENT_RESPONSE_SIZE;
-use rustfs_utils::EMPTY_STRING_SHA256_HASH;
-use s3s::dto::Owner;
-use std::collections::HashMap;
-
-#[derive(Clone, Debug, Default, serde::Serialize, serde::Deserialize)]
-pub struct Grantee {
-    pub id: String,
-    pub display_name: String,
-    pub uri: String,
-}
-
-#[derive(Clone, Debug, Default, serde::Serialize, serde::Deserialize)]
-pub struct Grant {
-    pub grantee: Grantee,
-    pub permission: String,
-}
-
-#[derive(Debug, Default, serde::Serialize, serde::Deserialize)]
-pub struct AccessControlList {
-    pub grant: Vec<Grant>,
-    pub permission: String,
-}
-
-#[derive(Debug, Default, serde::Deserialize)]
-pub struct AccessControlPolicy {
-    #[serde(skip)]
-    owner: Owner,
-    pub access_control_list: AccessControlList,
-}
-
-impl TransitionClient {
-    pub async fn get_object_acl(&self, bucket_name: &str, object_name: &str) -> Result<ObjectInfo, std::io::Error> {
-        let mut url_values = HashMap::new();
-        url_values.insert("acl".to_string(), "".to_string());
-        let mut resp = self
-            .execute_method(
-                http::Method::GET,
-                &mut RequestMetadata {
-                    bucket_name: bucket_name.to_string(),
-                    object_name: object_name.to_string(),
-                    query_values: url_values,
-                    custom_header: HeaderMap::new(),
-                    content_sha256_hex: EMPTY_STRING_SHA256_HASH.to_string(),
-                    content_body: ReaderImpl::Body(Bytes::new()),
-                    content_length: 0,
-                    content_md5_base64: "".to_string(),
-                    stream_sha256: false,
-                    trailer: HeaderMap::new(),
-                    pre_sign_url: Default::default(),
-                    add_crc: Default::default(),
-                    extra_pre_sign_header: Default::default(),
-                    bucket_location: Default::default(),
-                    expires: Default::default(),
-                },
-            )
-            .await?;
-
-        let resp_status = resp.status();
-        let h = resp.headers().clone();
-
-        let mut body_vec = Vec::new();
-        let mut body = resp.into_body();
-        while let Some(frame) = body.frame().await {
-            let frame = frame.map_err(|e| std::io::Error::new(std::io::ErrorKind::Other, e.to_string()))?;
-            if let Some(data) = frame.data_ref() {
-                body_vec.extend_from_slice(data);
-            }
-        }
-
-        if resp_status != http::StatusCode::OK {
-            return Err(std::io::Error::other(http_resp_to_error_response(
-                resp_status,
-                &h,
-                body_vec,
-                bucket_name,
-                object_name,
-            )));
-        }
-
-        let mut res = match quick_xml::de::from_str::<AccessControlPolicy>(&String::from_utf8(body_vec).unwrap()) {
-            Ok(result) => result,
-            Err(err) => {
-                return Err(std::io::Error::other(err.to_string()));
-            }
-        };
-
-        let mut obj_info = self
-            .stat_object(bucket_name, object_name, &GetObjectOptions::default())
-            .await?;
-
-        obj_info.owner.display_name = res.owner.display_name.clone();
-        obj_info.owner.id = res.owner.id.clone();
-
-        //obj_info.grant.extend(res.access_control_list.grant);
-
-        let canned_acl = get_canned_acl(&res);
-        if canned_acl != "" {
-            obj_info
-                .metadata
-                .insert("X-Amz-Acl", HeaderValue::from_str(&canned_acl).unwrap());
-            return Ok(obj_info);
-        }
-
-        let grant_acl = get_amz_grant_acl(&res);
-        /*for (k, v) in grant_acl {
-            obj_info.metadata.insert(HeaderName::from_bytes(k.as_bytes()).unwrap(), HeaderValue::from_str(&v.to_string()).unwrap());
-        }*/
-
-        Ok(obj_info)
-    }
-}
-
-fn get_canned_acl(ac_policy: &AccessControlPolicy) -> String {
-    let grants = ac_policy.access_control_list.grant.clone();
-
-    if grants.len() == 1 {
-        if grants[0].grantee.uri == "" && grants[0].permission == "FULL_CONTROL" {
-            return "private".to_string();
-        }
-    } else if grants.len() == 2 {
-        for g in grants {
-            if g.grantee.uri == "http://acs.amazonaws.com/groups/global/AuthenticatedUsers" && &g.permission == "READ" {
-                return "authenticated-read".to_string();
-            }
-            if g.grantee.uri == "http://acs.amazonaws.com/groups/global/AllUsers" && &g.permission == "READ" {
-                return "public-read".to_string();
-            }
-            if g.permission == "READ" && g.grantee.id == ac_policy.owner.id.clone().unwrap() {
-                return "bucket-owner-read".to_string();
-            }
-        }
-    } else if grants.len() == 3 {
-        for g in grants {
-            if g.grantee.uri == "http://acs.amazonaws.com/groups/global/AllUsers" && g.permission == "WRITE" {
-                return "public-read-write".to_string();
-            }
-        }
-    }
-    "".to_string()
-}
-
-pub fn get_amz_grant_acl(ac_policy: &AccessControlPolicy) -> HashMap<String, Vec<String>> {
-    let grants = ac_policy.access_control_list.grant.clone();
-    let mut res = HashMap::<String, Vec<String>>::new();
-
-    for g in grants {
-        let mut id = "id=".to_string();
-        id.push_str(&g.grantee.id);
-        let permission: &str = &g.permission;
-        match permission {
-            "READ" => {
-                res.entry("X-Amz-Grant-Read".to_string()).or_insert(vec![]).push(id);
-            }
-            "WRITE" => {
-                res.entry("X-Amz-Grant-Write".to_string()).or_insert(vec![]).push(id);
-            }
-            "READ_ACP" => {
-                res.entry("X-Amz-Grant-Read-Acp".to_string()).or_insert(vec![]).push(id);
-            }
-            "WRITE_ACP" => {
-                res.entry("X-Amz-Grant-Write-Acp".to_string()).or_insert(vec![]).push(id);
-            }
-            "FULL_CONTROL" => {
-                res.entry("X-Amz-Grant-Full-Control".to_string()).or_insert(vec![]).push(id);
-            }
-            _ => (),
-        }
-    }
-    res
-}
@@ -1,266 +0,0 @@
-// 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.
-#![allow(unused_imports)]
-#![allow(unused_variables)]
-#![allow(unused_mut)]
-#![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
-
-use http::{HeaderMap, HeaderValue};
-use std::collections::HashMap;
-use time::OffsetDateTime;
-
-use crate::client::constants::{GET_OBJECT_ATTRIBUTES_MAX_PARTS, GET_OBJECT_ATTRIBUTES_TAGS, ISO8601_DATEFORMAT};
-use crate::client::{
-    api_get_object_acl::AccessControlPolicy,
-    transition_api::{ReaderImpl, RequestMetadata, TransitionClient},
-};
-use http_body_util::BodyExt;
-use hyper::body::Body;
-use hyper::body::Bytes;
-use hyper::body::Incoming;
-use rustfs_config::MAX_S3_CLIENT_RESPONSE_SIZE;
-use rustfs_utils::EMPTY_STRING_SHA256_HASH;
-use s3s::header::{X_AMZ_MAX_PARTS, X_AMZ_OBJECT_ATTRIBUTES, X_AMZ_PART_NUMBER_MARKER, X_AMZ_VERSION_ID};
-
-pub struct ObjectAttributesOptions {
-    pub max_parts: i64,
-    pub version_id: String,
-    pub part_number_marker: i64,
-    //server_side_encryption: encrypt::ServerSide,
-}
-
-pub struct ObjectAttributes {
-    pub version_id: String,
-    pub last_modified: OffsetDateTime,
-    pub object_attributes_response: ObjectAttributesResponse,
-}
-
-impl ObjectAttributes {
-    fn new() -> Self {
-        Self {
-            version_id: "".to_string(),
-            last_modified: OffsetDateTime::now_utc(),
-            object_attributes_response: ObjectAttributesResponse::new(),
-        }
-    }
-}
-
-#[derive(Debug, Default, serde::Deserialize)]
-pub struct Checksum {
-    checksum_crc32: String,
-    checksum_crc32c: String,
-    checksum_sha1: String,
-    checksum_sha256: String,
-}
-
-impl Checksum {
-    fn new() -> Self {
-        Self {
-            checksum_crc32: "".to_string(),
-            checksum_crc32c: "".to_string(),
-            checksum_sha1: "".to_string(),
-            checksum_sha256: "".to_string(),
-        }
-    }
-}
-
-#[derive(Debug, Default, serde::Deserialize)]
-pub struct ObjectParts {
-    pub parts_count: i64,
-    pub part_number_marker: i64,
-    pub next_part_number_marker: i64,
-    pub max_parts: i64,
-    is_truncated: bool,
-    parts: Vec<ObjectAttributePart>,
-}
-
-impl ObjectParts {
-    fn new() -> Self {
-        Self {
-            parts_count: 0,
-            part_number_marker: 0,
-            next_part_number_marker: 0,
-            max_parts: 0,
-            is_truncated: false,
-            parts: Vec::new(),
-        }
-    }
-}
-
-#[derive(Debug, Default, serde::Deserialize)]
-pub struct ObjectAttributesResponse {
-    pub etag: String,
-    pub storage_class: String,
-    pub object_size: i64,
-    pub checksum: Checksum,
-    pub object_parts: ObjectParts,
-}
-
-impl ObjectAttributesResponse {
-    fn new() -> Self {
-        Self {
-            etag: "".to_string(),
-            storage_class: "".to_string(),
-            object_size: 0,
-            checksum: Checksum::new(),
-            object_parts: ObjectParts::new(),
-        }
-    }
-}
-
-#[derive(Debug, Default, serde::Deserialize)]
-struct ObjectAttributePart {
-    checksum_crc32: String,
-    checksum_crc32c: String,
-    checksum_sha1: String,
-    checksum_sha256: String,
-    part_number: i64,
-    size: i64,
-}
-
-impl ObjectAttributes {
-    pub async fn parse_response(&mut self, h: &HeaderMap, body_vec: Vec<u8>) -> Result<(), std::io::Error> {
-        let last_modified = h
-            .get("Last-Modified")
-            .ok_or_else(|| std::io::Error::other("missing Last-Modified header"))?
-            .to_str()
-            .map_err(|e| std::io::Error::other(format!("invalid Last-Modified header: {e}")))?;
-        let mod_time = OffsetDateTime::parse(last_modified, ISO8601_DATEFORMAT)
-            .map_err(|e| std::io::Error::other(format!("invalid Last-Modified date: {e}")))?;
-        self.last_modified = mod_time;
-
-        let version_id = h
-            .get(X_AMZ_VERSION_ID)
-            .ok_or_else(|| std::io::Error::other("missing version ID header"))?
-            .to_str()
-            .map_err(|e| std::io::Error::other(format!("invalid version ID header: {e}")))?;
-        self.version_id = version_id.to_string();
-
-        let body_str = String::from_utf8(body_vec).map_err(|e| std::io::Error::other(format!("invalid UTF-8 body: {e}")))?;
-        let mut response = match quick_xml::de::from_str::<ObjectAttributesResponse>(&body_str) {
-            Ok(result) => result,
-            Err(err) => {
-                return Err(std::io::Error::other(err.to_string()));
-            }
-        };
-        self.object_attributes_response = response;
-
-        Ok(())
-    }
-}
-
-impl TransitionClient {
-    pub async fn get_object_attributes(
-        &self,
-        bucket_name: &str,
-        object_name: &str,
-        opts: ObjectAttributesOptions,
-    ) -> Result<ObjectAttributes, std::io::Error> {
-        let mut url_values = HashMap::new();
-        url_values.insert("attributes".to_string(), "".to_string());
-        if opts.version_id != "" {
-            url_values.insert("versionId".to_string(), opts.version_id);
-        }
-
-        let mut headers = HeaderMap::new();
-        headers.insert(
-            X_AMZ_OBJECT_ATTRIBUTES,
-            HeaderValue::from_str(GET_OBJECT_ATTRIBUTES_TAGS).expect("valid header value"),
-        );
-
-        if opts.part_number_marker > 0 {
-            headers.insert(
-                X_AMZ_PART_NUMBER_MARKER,
-                HeaderValue::from_str(&opts.part_number_marker.to_string()).expect("valid header value"),
-            );
-        }
-
-        if opts.max_parts > 0 {
-            headers.insert(
-                X_AMZ_MAX_PARTS,
-                HeaderValue::from_str(&opts.max_parts.to_string()).expect("valid header value"),
-            );
-        } else {
-            headers.insert(
-                X_AMZ_MAX_PARTS,
-                HeaderValue::from_str(&GET_OBJECT_ATTRIBUTES_MAX_PARTS.to_string()).expect("valid header value"),
-            );
-        }
-
-        /*if opts.server_side_encryption.is_some() {
-            opts.server_side_encryption.Marshal(headers);
-        }*/
-
-        let mut resp = self
-            .execute_method(
-                http::Method::HEAD,
-                &mut RequestMetadata {
-                    bucket_name: bucket_name.to_string(),
-                    object_name: object_name.to_string(),
-                    query_values: url_values,
-                    custom_header: headers,
-                    content_sha256_hex: EMPTY_STRING_SHA256_HASH.to_string(),
-                    content_md5_base64: "".to_string(),
-                    content_body: ReaderImpl::Body(Bytes::new()),
-                    content_length: 0,
-                    stream_sha256: false,
-                    trailer: HeaderMap::new(),
-                    pre_sign_url: Default::default(),
-                    add_crc: Default::default(),
-                    extra_pre_sign_header: Default::default(),
-                    bucket_location: Default::default(),
-                    expires: Default::default(),
-                },
-            )
-            .await?;
-
-        let resp_status = resp.status();
-        let h = resp.headers().clone();
-        let has_etag = h.get("ETag").and_then(|v| v.to_str().ok()).unwrap_or("");
-        if !has_etag.is_empty() {
-            return Err(std::io::Error::other(
-                "get_object_attributes is not supported by the current endpoint version",
-            ));
-        }
-
-        let mut body_vec = Vec::new();
-        let mut body = resp.into_body();
-        while let Some(frame) = body.frame().await {
-            let frame = frame.map_err(|e| std::io::Error::new(std::io::ErrorKind::Other, e.to_string()))?;
-            if let Some(data) = frame.data_ref() {
-                body_vec.extend_from_slice(data);
-            }
-        }
-
-        if resp_status != http::StatusCode::OK {
-            let err_body =
-                String::from_utf8(body_vec).map_err(|e| std::io::Error::other(format!("invalid UTF-8 error body: {e}")))?;
-            let mut er = match quick_xml::de::from_str::<AccessControlPolicy>(&err_body) {
-                Ok(result) => result,
-                Err(err) => {
-                    return Err(std::io::Error::other(err.to_string()));
-                }
-            };
-
-            return Err(std::io::Error::other(er.access_control_list.permission));
-        }
-
-        let mut oa = ObjectAttributes::new();
-        oa.parse_response(&h, body_vec).await?;
-
-        Ok(oa)
-    }
-}
@@ -1,159 +0,0 @@
-// Copyright 2024 RustFS Team
-//
-// Licensed under the Apache License, Version 2.0 (the "License");
-// you may not use this file except in compliance with the License.
-// You may obtain a copy of the License at
-//
-//     http://www.apache.org/licenses/LICENSE-2.0
-//
-// Unless required by applicable law or agreed to in writing, software
-// distributed under the License is distributed on an "AS IS" BASIS,
-// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-// See the License for the specific language governing permissions and
-// limitations under the License.
-
-use std::io;
-use std::path::{Path, PathBuf};
-
-#[cfg(not(windows))]
-use std::os::unix::fs::PermissionsExt;
-
-use tokio::fs::{self, OpenOptions};
-use tokio::io::{AsyncSeekExt, AsyncWriteExt, SeekFrom};
-
-use crate::client::{
-    api_error_response::err_invalid_argument, api_get_options::GetObjectOptions, transition_api::TransitionClient,
-};
-
-async fn prepare_download_target(file_path: &Path) -> io::Result<()> {
-    match fs::metadata(file_path).await {
-        Ok(metadata) if metadata.is_dir() => {
-            return Err(io::Error::other(err_invalid_argument("filename is a directory.")));
-        }
-        Ok(_) => {}
-        Err(err) if err.kind() == io::ErrorKind::NotFound => {}
-        Err(err) => return Err(err),
-    }
-
-    if let Some(parent) = file_path.parent()
-        && !parent.as_os_str().is_empty()
-    {
-        fs::create_dir_all(parent).await?;
-
-        #[cfg(not(windows))]
-        {
-            let mut permissions = fs::metadata(parent).await?.permissions();
-            permissions.set_mode(0o700);
-            fs::set_permissions(parent, permissions).await?;
-        }
-    }
-
-    Ok(())
-}
-
-fn build_part_path(file_path: &Path) -> PathBuf {
-    PathBuf::from(format!("{}.part.rustfs", file_path.display()))
-}
-
-async fn open_download_part_file(file_part_path: &Path) -> io::Result<tokio::fs::File> {
-    let mut options = OpenOptions::new();
-    options.create(true).truncate(false).read(true).write(true);
-
-    #[cfg(not(windows))]
-    options.mode(0o600);
-
-    options.open(file_part_path).await
-}
-
-async fn cleanup_part_file(file_part_path: &Path) {
-    let _ = fs::remove_file(file_part_path).await;
-}
-
-impl TransitionClient {
-    pub async fn fget_object(
-        &self,
-        bucket_name: &str,
-        object_name: &str,
-        file_path: &str,
-        mut opts: GetObjectOptions,
-    ) -> Result<(), io::Error> {
-        let file_path = Path::new(file_path);
-        prepare_download_target(file_path).await?;
-
-        let file_part_path = build_part_path(file_path);
-        let mut file_part = open_download_part_file(&file_part_path).await?;
-        let existing_len = file_part.metadata().await?.len();
-        if existing_len > 0 {
-            opts.set_range(existing_len as i64, 0)?;
-            file_part.seek(SeekFrom::Start(existing_len)).await?;
-        }
-
-        let (_object_info, _headers, mut object_reader) = self.get_object_inner(bucket_name, object_name, &opts).await?;
-        if let Err(err) = tokio::io::copy(&mut object_reader, &mut file_part).await {
-            cleanup_part_file(&file_part_path).await;
-            return Err(err);
-        }
-
-        if let Err(err) = file_part.flush().await {
-            cleanup_part_file(&file_part_path).await;
-            return Err(err);
-        }
-        drop(file_part);
-
-        if let Err(err) = fs::rename(&file_part_path, file_path).await {
-            cleanup_part_file(&file_part_path).await;
-            return Err(err);
-        }
-
-        Ok(())
-    }
-}
-
-#[cfg(test)]
-mod tests {
-    use super::*;
-    use tempfile::tempdir;
-
-    #[tokio::test]
-    async fn prepare_download_target_allows_missing_file_and_creates_parent_dirs() {
-        let dir = tempdir().expect("temp dir");
-        let target = dir.path().join("nested").join("object.bin");
-
-        prepare_download_target(&target)
-            .await
-            .expect("missing target should be accepted");
-
-        assert!(target.parent().expect("parent").exists(), "parent directory should be created");
-        assert!(
-            fs::metadata(&target).await.is_err(),
-            "preparing the target should not create the final file eagerly"
-        );
-    }
-
-    #[tokio::test]
-    async fn prepare_download_target_rejects_directory_paths() {
-        let dir = tempdir().expect("temp dir");
-        let target_dir = dir.path().join("download-dir");
-        fs::create_dir_all(&target_dir).await.expect("target dir");
-
-        let err = prepare_download_target(&target_dir)
-            .await
-            .expect_err("directory targets must be rejected");
-
-        assert!(err.to_string().contains("directory"), "unexpected error for directory target: {err}");
-    }
-
-    #[tokio::test]
-    async fn open_download_part_file_creates_part_file() {
-        let dir = tempdir().expect("temp dir");
-        let target = dir.path().join("object.bin");
-        let part_path = build_part_path(&target);
-
-        let file = open_download_part_file(&part_path)
-            .await
-            .expect("part file should be created");
-        drop(file);
-
-        assert!(part_path.exists(), "part file should exist after creation");
-    }
-}
@@ -1,134 +0,0 @@
-// 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.
-#![allow(unused_imports)]
-#![allow(unused_variables)]
-#![allow(unused_mut)]
-#![allow(unused_assignments)]
-#![allow(unused_must_use)]
-#![allow(clippy::all)]
-
-use crate::client::{
-    api_error_response::{err_invalid_argument, http_resp_to_error_response},
-    api_get_object_acl::AccessControlList,
-    api_get_options::GetObjectOptions,
-    transition_api::{ObjectInfo, ReadCloser, ReaderImpl, RequestMetadata, TransitionClient, to_object_info},
-};
-use http::HeaderMap;
-use http_body_util::BodyExt;
-use hyper::body::Body;
-use hyper::body::Bytes;
-use s3s::dto::RestoreRequest;
-use std::collections::HashMap;
-use std::io::Cursor;
-use tokio::io::BufReader;
-
-const TIER_STANDARD: &str = "Standard";
-const TIER_BULK: &str = "Bulk";
-const TIER_EXPEDITED: &str = "Expedited";
-
-#[derive(Debug, Default, serde::Serialize, serde::Deserialize)]
-pub struct Encryption {
-    pub encryption_type: String,
-    pub kms_context: String,
-    pub kms_key_id: String,
-}
-
-#[derive(Debug, Default, serde::Serialize, serde::Deserialize)]
-pub struct MetadataEntry {
-    pub name: String,
-    pub value: String,
-}
-
-#[derive(Debug, Default, serde::Serialize)]
-pub struct S3 {
-    pub access_control_list: AccessControlList,
-    pub bucket_name: String,
-    pub prefix: String,
-    pub canned_acl: String,
-    pub encryption: Encryption,
-    pub storage_class: String,
-    //tagging:            Tags,
-    pub user_metadata: MetadataEntry,
-}
-
-impl TransitionClient {
-    pub async fn restore_object(
-        &self,
-        bucket_name: &str,
-        object_name: &str,
-        version_id: &str,
-        restore_req: &RestoreRequest,
-    ) -> Result<(), std::io::Error> {
-        /*let restore_request = match quick_xml::se::to_string(restore_req) {
-            Ok(buf) => buf,
-            Err(e) => {
-                return Err(std::io::Error::other(e));
-            }
-        };*/
-        let restore_request = "".to_string();
-        let restore_request_bytes = restore_request.as_bytes().to_vec();
-
-        let mut url_values = HashMap::new();
-        url_values.insert("restore".to_string(), "".to_string());
-        if version_id != "" {
-            url_values.insert("versionId".to_string(), version_id.to_string());
-        }
-
-        let restore_request_buffer = Bytes::from(restore_request_bytes.clone());
-        let resp = self
-            .execute_method(
-                http::Method::HEAD,
-                &mut RequestMetadata {
-                    bucket_name: bucket_name.to_string(),
-                    object_name: object_name.to_string(),
-                    query_values: url_values,
-                    custom_header: HeaderMap::new(),
-                    content_sha256_hex: "".to_string(), //sum_sha256_hex(&restore_request_bytes),
-                    content_md5_base64: "".to_string(), //sum_md5_base64(&restore_request_bytes),
-                    content_body: ReaderImpl::Body(restore_request_buffer),
-                    content_length: restore_request_bytes.len() as i64,
-                    stream_sha256: false,
-                    trailer: HeaderMap::new(),
-                    pre_sign_url: Default::default(),
-                    add_crc: Default::default(),
-                    extra_pre_sign_header: Default::default(),
-                    bucket_location: Default::default(),
-                    expires: Default::default(),
-                },
-            )
-            .await?;
-
-        let resp_status = resp.status();
-        let h = resp.headers().clone();
-
-        let mut body_vec = Vec::new();
-        let mut body = resp.into_body();
-        while let Some(frame) = body.frame().await {
-            let frame = frame.map_err(|e| std::io::Error::new(std::io::ErrorKind::Other, e.to_string()))?;
-            if let Some(data) = frame.data_ref() {
-                body_vec.extend_from_slice(data);
-            }
-        }
-        if resp_status != http::StatusCode::ACCEPTED && resp_status != http::StatusCode::OK {
-            return Err(std::io::Error::other(http_resp_to_error_response(
-                resp_status,
-                &h,
-                body_vec,
-                bucket_name,
-                "",
-            )));
-        }
-        Ok(())
-    }
-}
@@ -27,12 +27,24 @@ use crate::client::utils::base64_decode;
 use crate::client::utils::base64_encode;
 use crate::client::{api_put_object::PutObjectOptions, api_s3_datatypes::ObjectPart};
 use crate::{disk::DiskAPI, object_api::GetObjectReader};
+// s3s::header has no CRC64NVME constant yet; the canonical RustFS copy lives
+// in rustfs-utils' headers module.
+use rustfs_utils::http::headers::AMZ_CHECKSUM_CRC64NVME;
 use s3s::header::{
    X_AMZ_CHECKSUM_ALGORITHM, X_AMZ_CHECKSUM_CRC32, X_AMZ_CHECKSUM_CRC32C, X_AMZ_CHECKSUM_SHA1, X_AMZ_CHECKSUM_SHA256,
 };

 use enumset::{EnumSet, EnumSetType, enum_set};

+/// One of three deliberately separate checksum registries (backlog#1833):
+/// this enum is the MinIO-port client's wire vocabulary and stops at the
+/// standard S3 set (CRC64NVME is its newest member; the RustFS extensions do
+/// not exist on this client path). The streaming-hash registry lives in
+/// `rustfs_checksums::ChecksumAlgorithm` (crates/checksums/src/lib.rs) and
+/// the on-disk xl.meta bitset in `rustfs_rio::ChecksumType`
+/// (crates/rio/src/checksum.rs, varint bits are append-only). When adding an
+/// algorithm, extend all three (or record why not) — they do not derive from
+/// each other.
 #[derive(Debug, EnumSetType, Default)]
 #[enumset(repr = "u8")]
 pub enum ChecksumMode {
@@ -57,8 +69,6 @@ lazy_static! {
    static ref C_ChecksumFullObjectCRC32C: EnumSet<ChecksumMode> =
        enum_set!(ChecksumMode::ChecksumCRC32C | ChecksumMode::ChecksumFullObject);
 }
-const AMZ_CHECKSUM_CRC64NVME: &str = "x-amz-checksum-crc64nvme";
-
 impl ChecksumMode {
    //pub const CRC64_NVME_POLYNOMIAL: i64 = 0xad93d23594c93659;

@@ -37,6 +37,3 @@ pub const TOTAL_WORKERS: i64 = 4;
 pub const SIGN_V4_ALGORITHM: &str = "AWS4-HMAC-SHA256";
 pub const ISO8601_DATEFORMAT: &[FormatItem<'_>] =
    format_description!("[year]-[month]-[day]T[hour]:[minute]:[second].[subsecond]Z");
-
-pub const GET_OBJECT_ATTRIBUTES_TAGS: &str = "ETag,Checksum,StorageClass,ObjectSize,ObjectParts";
-pub const GET_OBJECT_ATTRIBUTES_MAX_PARTS: i64 = 1000;
@@ -16,12 +16,8 @@
 #![allow(dead_code)]

 pub mod admin_handler_utils;
-pub mod api_bucket_policy;
 pub mod api_error_response;
 pub mod api_get_object;
-pub mod api_get_object_acl;
-pub mod api_get_object_attributes;
-pub mod api_get_object_file;
 pub mod api_get_options;
 pub mod api_list;
 pub mod api_put_object;
@@ -29,7 +25,6 @@ pub mod api_put_object_common;
 pub mod api_put_object_multipart;
 pub mod api_put_object_streaming;
 pub mod api_remove;
-pub mod api_restore;
 pub mod api_s3_datatypes;
 pub mod api_stat;
 pub mod bucket_cache;
@@ -1006,16 +1006,6 @@ impl TransitionCore {
        client.abort_multipart_upload(bucket_name, object, upload_id).await
    }

-    pub async fn get_bucket_policy(&self, bucket_name: &str) -> Result<String, std::io::Error> {
-        let client = self.0.clone();
-        client.get_bucket_policy(bucket_name).await
-    }
-
-    pub async fn put_bucket_policy(&self, bucket_name: &str, bucket_policy: &str) -> Result<(), std::io::Error> {
-        let client = self.0.clone();
-        client.put_bucket_policy(bucket_name, bucket_policy).await
-    }
-
    pub async fn get_object(
        &self,
        bucket_name: &str,
@@ -31,7 +31,7 @@ use rustfs_config::{
    DEFAULT_INTERNODE_DATA_TRANSPORT, ENV_RUSTFS_INTERNODE_DATA_TRANSPORT, INTERNODE_DATA_TRANSPORT_TCP,
    KNOWN_INTERNODE_DATA_TRANSPORT_BACKENDS,
 };
-use rustfs_rio::{HttpReader, HttpWriter};
+use rustfs_rio::{ChunkReaderBox, HttpChunkReader, HttpReader, HttpWriter};
 use sha2::{Digest, Sha256};
 use std::collections::HashMap;
 use std::future::Future;
@@ -221,6 +221,11 @@ pub struct NsScannerCapabilityRequest {
 #[async_trait]
 pub trait InternodeDataTransport: Send + Sync + std::fmt::Debug {
    async fn open_read(&self, request: ReadStreamRequest) -> Result<FileReader>;
+    /// Opens an owned-chunk stream when this transport can retain receive-buffer
+    /// ownership. `None` preserves the established `open_read` fallback.
+    async fn open_read_chunks(&self, _request: ReadStreamRequest) -> Result<Option<ChunkReaderBox>> {
+        Ok(None)
+    }
    async fn open_write(&self, request: WriteStreamRequest) -> Result<FileWriter>;
    async fn open_walk_dir(&self, request: WalkDirStreamRequest) -> Result<FileReader>;
    async fn open_ns_scanner(&self, _request: NsScannerStreamRequest) -> Result<FileReader> {
@@ -247,6 +252,15 @@ impl InternodeDataTransport for TcpHttpInternodeDataTransport {
        ))
    }

+    async fn open_read_chunks(&self, request: ReadStreamRequest) -> Result<Option<ChunkReaderBox>> {
+        let url = build_read_file_stream_url(&request);
+        let mut headers = json_headers();
+        build_auth_headers(&url, &Method::GET, &mut headers)?;
+        Ok(Some(Box::new(
+            HttpChunkReader::new_with_stall_timeout(url, Method::GET, headers, None, request.stall_timeout).await?,
+        )))
+    }
+
    async fn open_write(&self, request: WriteStreamRequest) -> Result<FileWriter> {
        let server_epoch = self.put_file_auth_capability(&request.endpoint).await?;
        let nonce = server_epoch.map(|_| Uuid::new_v4());
@@ -522,6 +522,33 @@ impl RemoteDisk {
        }
    }

+    async fn open_read_chunks_with_retry(&self, request: ReadStreamRequest) -> Result<Option<rustfs_rio::ChunkReaderBox>> {
+        let mut attempt = 1;
+        let mut last_retry_classification = None;
+        loop {
+            match self.data_transport.open_read_chunks(request.clone()).await {
+                Ok(reader) => {
+                    if attempt > 1
+                        && let Some(classification) = last_retry_classification
+                    {
+                        crate::cluster::rpc::runtime_sources::record_remote_disk_open_read_retry_success(classification);
+                    }
+                    return Ok(reader);
+                }
+                Err(err) if attempt < REMOTE_DISK_OPEN_READ_MAX_ATTEMPTS && Self::is_retryable_open_read_error(&err) => {
+                    if let Some(classification) = err.internode_http_error_kind() {
+                        let classification = classification.metric_label();
+                        crate::cluster::rpc::runtime_sources::record_remote_disk_open_read_retry(classification);
+                        last_retry_classification = Some(classification);
+                    }
+                    tokio::time::sleep(REMOTE_DISK_OPEN_READ_RETRY_BACKOFF).await;
+                    attempt += 1;
+                }
+                Err(err) => return Err(err),
+            }
+        }
+    }
+
    pub fn record_capacity_probe(&self, total: u64, used: u64, free: u64) {
        self.health.record_capacity_probe(total, used, free);
    }
@@ -846,31 +873,49 @@ impl RemoteDisk {
    /// default to 1 (see [`internode_idempotent_read_retries`]). MUST NOT be used for write/lock
    /// RPCs — those must never auto-retry (quorum/idempotency safety). The `operation` closure is
    /// re-invoked per attempt, so it must be `Fn` (rebuild the request from borrowed inputs, do not
-    /// move captured state out).
+    /// move captured state out). Attempts and backoff share one total timeout budget.
    async fn execute_read_with_retry<T, F, Fut>(&self, op: &'static str, operation: F, timeout_duration: Duration) -> Result<T>
    where
        F: Fn() -> Fut,
        Fut: std::future::Future<Output = Result<T>>,
    {
+        let deadline = (!timeout_duration.is_zero()).then(|| {
+            time::Instant::now()
+                .checked_add(timeout_duration)
+                .unwrap_or_else(|| time::sleep(timeout_duration).deadline())
+        });
        let max_retries = internode_idempotent_read_retries();
        let mut attempt = 0usize;
        loop {
-            // Only the final attempt marks the disk faulty / evicts the channel. Earlier retries
-            // ignore the failure, so a transient error cannot flip the disk into a faulty
-            // short-circuit (which would defeat the retry) or over-count failures.
+            let attempt_timeout = deadline
+                .map(|deadline| deadline.saturating_duration_since(time::Instant::now()))
+                .unwrap_or(Duration::ZERO);
+            if deadline.is_some() && attempt_timeout.is_zero() {
+                self.record_timeout(op, timeout_duration);
+                return Err(DiskError::Timeout);
+            }
+
            let health_action = if attempt >= max_retries {
                FailureHealthAction::MarkFailure
            } else {
                FailureHealthAction::IgnoreFailure
            };
            match self
-                .execute_with_timeout_for_op_and_health_action(op, &operation, timeout_duration, health_action)
+                .execute_with_timeout_for_op_and_health_action(op, &operation, attempt_timeout, health_action)
                .await
            {
                Err(err) if attempt < max_retries && is_network_like_disk_error(&err) => {
+                    if matches!(err, DiskError::Timeout) && deadline.is_some_and(|deadline| time::Instant::now() >= deadline) {
+                        self.mark_faulty("read_operation_deadline");
+                        return Err(err);
+                    }
                    attempt += 1;
                    let backoff = REMOTE_DISK_READ_RETRY_BASE_BACKOFF
                        .saturating_mul(1u32 << u32::try_from(attempt - 1).unwrap_or(4).min(4));
+                    if deadline.is_some_and(|deadline| deadline.saturating_duration_since(time::Instant::now()) <= backoff) {
+                        attempt = max_retries;
+                        continue;
+                    }
                    debug!(
                        endpoint = %self.endpoint,
                        addr = %self.addr,
@@ -878,7 +923,17 @@ impl RemoteDisk {
                        attempt,
                        "retrying idempotent read-only RPC after transient network error"
                    );
-                    tokio::time::sleep(backoff).await;
+                    if let Some(deadline) = deadline {
+                        if time::timeout_at(deadline, time::sleep(backoff)).await.is_err() {
+                            self.record_timeout(op, timeout_duration);
+                            return Err(DiskError::Timeout);
+                        }
+                    } else {
+                        time::sleep(backoff).await;
+                    }
+                    if self.health.is_faulty() {
+                        return Err(DiskError::FaultyDisk);
+                    }
                }
                other => return other,
            }
@@ -957,32 +1012,35 @@ impl RemoteDisk {
                operation_result
            }
            Err(_) => {
-                // Timeout occurred, mark disk as potentially faulty
-                counter!(
-                    "rustfs_drive_op_timeout_total",
-                    "endpoint" => self.endpoint.to_string(),
-                    "op" => op.to_string()
-                )
-                .increment(1);
+                self.record_timeout(op, timeout_duration);
                if failure_health_action == FailureHealthAction::MarkFailure {
                    self.mark_faulty_and_evict("operation_timeout").await;
                }
-                warn!(
-                    event = EVENT_REMOTE_DISK_RPC,
-                    component = LOG_COMPONENT_ECSTORE,
-                    subsystem = LOG_SUBSYSTEM_REMOTE_DISK,
-                    endpoint = %self.endpoint,
-                    addr = %self.addr,
-                    op,
-                    timeout_ms = timeout_duration.as_millis(),
-                    state = "timeout",
-                    "Remote disk operation timed out"
-                );
                Err(DiskError::Timeout)
            }
        }
    }

+    fn record_timeout(&self, op: &'static str, timeout_duration: Duration) {
+        counter!(
+            "rustfs_drive_op_timeout_total",
+            "endpoint" => self.endpoint.to_string(),
+            "op" => op.to_string()
+        )
+        .increment(1);
+        warn!(
+            event = EVENT_REMOTE_DISK_RPC,
+            component = LOG_COMPONENT_ECSTORE,
+            subsystem = LOG_SUBSYSTEM_REMOTE_DISK,
+            endpoint = %self.endpoint,
+            addr = %self.addr,
+            op,
+            timeout_ms = timeout_duration.as_millis(),
+            state = "timeout",
+            "Remote disk operation timed out"
+        );
+    }
+
    async fn handle_network_like_error<T>(
        &self,
        op: &'static str,
@@ -1016,7 +1074,7 @@ impl RemoteDisk {
        }
    }

-    async fn mark_faulty_and_evict(&self, reason: &'static str) {
+    fn mark_faulty(&self, reason: &'static str) -> bool {
        let previous_state = self.runtime_state();
        let transitioned_to_offline = self.mark_suspect_or_offline(reason);
        let state = self.runtime_state();
@@ -1053,6 +1111,12 @@ impl RemoteDisk {
                    "Remote disk marked suspect"
                );
            }
+        }
+        state != previous_state
+    }
+
+    async fn mark_faulty_and_evict(&self, reason: &'static str) {
+        if self.mark_faulty(reason) {
            counter!(
                "rustfs_drive_connection_evict_total",
                "endpoint" => self.endpoint.to_string(),
@@ -2068,7 +2132,7 @@ impl DiskAPI for RemoteDisk {

                Ok(file_info)
            },
-            get_max_timeout_duration(),
+            get_drive_metadata_timeout(),
        )
        .await
    }
@@ -2417,6 +2481,30 @@ impl DiskAPI for RemoteDisk {
        .await
    }

+    async fn read_file_stream_chunks(
+        &self,
+        volume: &str,
+        path: &str,
+        offset: usize,
+        length: usize,
+    ) -> Result<Option<rustfs_rio::ChunkReaderBox>> {
+        if self.health.is_faulty() {
+            return Err(DiskError::FaultyDisk);
+        }
+        let disk = self.disk_ref().await;
+        let stall_timeout = get_object_disk_read_timeout();
+        self.open_read_chunks_with_retry(ReadStreamRequest {
+            endpoint: self.endpoint.grid_host(),
+            disk,
+            volume: volume.to_string(),
+            path: path.to_string(),
+            offset,
+            length,
+            stall_timeout: (!stall_timeout.is_zero()).then_some(stall_timeout),
+        })
+        .await
+    }
+
    /// Buffered read for remote disks.
    /// The transport stream is collected into owned Bytes for caller sharing.
    #[tracing::instrument(level = "trace", skip_all)]
@@ -5094,6 +5182,452 @@ mod tests {
        );
    }

+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_reset_during_backoff_preserves_recovery() {
+        let remote_disk = Arc::new(new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await);
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let first_attempt = Arc::new(tokio::sync::Notify::new());
+        let started = time::Instant::now();
+
+        let task_disk = Arc::clone(&remote_disk);
+        let task_attempts = Arc::clone(&attempts);
+        let task_first_attempt = Arc::clone(&first_attempt);
+        let task = tokio::spawn(async move {
+            task_disk
+                .execute_read_with_retry(
+                    "read_version",
+                    move || {
+                        let attempt = task_attempts.fetch_add(1, Ordering::SeqCst);
+                        let first_attempt = Arc::clone(&task_first_attempt);
+                        async move {
+                            if attempt == 0 {
+                                time::sleep(Duration::from_millis(20)).await;
+                                first_attempt.notify_one();
+                                return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                    std_io::ErrorKind::ConnectionRefused,
+                                    "connection refused",
+                                )));
+                            }
+                            Ok(())
+                        }
+                    },
+                    Duration::from_millis(100),
+                )
+                .await
+        });
+
+        first_attempt.notified().await;
+        tokio::task::yield_now().await;
+        remote_disk.health.reset_for_store_init_retry(&remote_disk.endpoint);
+        let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+            .expect("remote disk address should parse")
+            .connect_lazy();
+        runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+        task.await
+            .expect("retry task should finish")
+            .expect("the retry should succeed after the health reset");
+
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(started.elapsed(), Duration::from_millis(70));
+        assert_eq!(
+            remote_disk.health.waiting_count(),
+            0,
+            "health reset must not underflow the waiting counter"
+        );
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Online);
+        assert!(
+            runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await,
+            "a recovered channel must survive the retry backoff"
+        );
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_still_retries_within_shared_deadline() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+            .expect("remote disk address should parse")
+            .connect_lazy();
+        runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+
+        remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    let attempt = attempts.fetch_add(1, Ordering::SeqCst);
+                    async move {
+                        if attempt == 0 {
+                            return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionReset,
+                                "connection reset",
+                            )));
+                        }
+                        Ok(())
+                    }
+                },
+                Duration::from_millis(100),
+            )
+            .await
+            .expect("a retry that fits the shared deadline should succeed");
+
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Online);
+        assert!(runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_uses_remaining_budget_for_final_attempt() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let started = time::Instant::now();
+        let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+            .expect("remote disk address should parse")
+            .connect_lazy();
+        runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+
+        let err = remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    let attempt = attempts.fetch_add(1, Ordering::SeqCst);
+                    async move {
+                        if attempt == 0 {
+                            time::sleep(Duration::from_millis(20)).await;
+                            return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionRefused,
+                                "connection refused",
+                            )));
+                        }
+                        std::future::pending::<Result<()>>().await
+                    }
+                },
+                Duration::from_millis(100),
+            )
+            .await
+            .expect_err("the final retry should consume only the remaining total budget");
+
+        assert_eq!(err, DiskError::Timeout);
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(started.elapsed(), Duration::from_millis(100));
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+        assert!(!runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_uses_final_attempt_at_exact_backoff_boundary() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let started = time::Instant::now();
+
+        let err = remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    let attempt = attempts.fetch_add(1, Ordering::SeqCst);
+                    async move {
+                        if attempt == 0 {
+                            time::sleep(Duration::from_millis(50)).await;
+                            return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionRefused,
+                                "connection refused",
+                            )));
+                        }
+                        std::future::pending::<Result<()>>().await
+                    }
+                },
+                Duration::from_millis(100),
+            )
+            .await
+            .expect_err("the exact backoff boundary should be reserved for a final attempt");
+
+        assert_eq!(err, DiskError::Timeout);
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(started.elapsed(), Duration::from_millis(100));
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_uses_final_attempt_below_backoff_budget() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let started = time::Instant::now();
+
+        let err = remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    let attempt = attempts.fetch_add(1, Ordering::SeqCst);
+                    async move {
+                        if attempt == 0 {
+                            time::sleep(Duration::from_millis(80)).await;
+                            return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionRefused,
+                                "connection refused",
+                            )));
+                        }
+                        std::future::pending::<Result<()>>().await
+                    }
+                },
+                Duration::from_millis(100),
+            )
+            .await
+            .expect_err("remaining budget below backoff should be reserved for a final attempt");
+
+        assert_eq!(err, DiskError::Timeout);
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(started.elapsed(), Duration::from_millis(100));
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_zero_timeout_disables_the_deadline() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let started = time::Instant::now();
+
+        remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    let attempt = attempts.fetch_add(1, Ordering::SeqCst);
+                    async move {
+                        if attempt == 0 {
+                            return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionReset,
+                                "connection reset",
+                            )));
+                        }
+                        Ok(())
+                    }
+                },
+                Duration::ZERO,
+            )
+            .await
+            .expect("zero timeout should allow a retry without a deadline");
+
+        assert_eq!(attempts.load(Ordering::SeqCst), 2);
+        assert_eq!(started.elapsed(), REMOTE_DISK_READ_RETRY_BASE_BACKOFF);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_accepts_max_metadata_timeout() {
+        temp_env::async_with_vars([(rustfs_config::ENV_DRIVE_METADATA_TIMEOUT_SECS, Some(u64::MAX.to_string()))], async {
+            let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+
+            remote_disk
+                .execute_read_with_retry("read_version", || async { Ok::<(), Error>(()) }, get_drive_metadata_timeout())
+                .await
+                .expect("the maximum configured metadata timeout must not panic");
+
+            remote_disk.cancel_token.cancel();
+        })
+        .await;
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_zero_retries_runs_once() {
+        temp_env::async_with_vars([(rustfs_config::ENV_INTERNODE_IDEMPOTENT_READ_RETRIES, Some("0"))], async {
+            let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+            let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+            let started = time::Instant::now();
+            let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+                .expect("remote disk address should parse")
+                .connect_lazy();
+            runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+
+            let err = remote_disk
+                .execute_read_with_retry(
+                    "read_version",
+                    || {
+                        attempts.fetch_add(1, Ordering::SeqCst);
+                        async {
+                            Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionReset,
+                                "connection reset",
+                            )))
+                        }
+                    },
+                    Duration::from_secs(1),
+                )
+                .await
+                .expect_err("zero retries should return the first network error");
+
+            assert!(matches!(err, DiskError::Io(ref io_err) if io_err.kind() == std_io::ErrorKind::ConnectionReset));
+            assert_eq!(attempts.load(Ordering::SeqCst), 1);
+            assert_eq!(started.elapsed(), Duration::ZERO);
+            assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+            assert!(!runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await);
+            remote_disk.cancel_token.cancel();
+        })
+        .await;
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_attempt_timeout_marks_health_without_evicting() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let recorder = crate::test_metrics::CapturingRecorder::default();
+        let _recorder_guard = metrics::set_default_local_recorder(&recorder);
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+            .expect("remote disk address should parse")
+            .connect_lazy();
+        runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+
+        let err = remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    attempts.fetch_add(1, Ordering::SeqCst);
+                    std::future::pending::<Result<()>>()
+                },
+                Duration::from_millis(100),
+            )
+            .await
+            .expect_err("an in-flight attempt that consumes the deadline should time out");
+
+        assert_eq!(err, DiskError::Timeout);
+        assert_eq!(attempts.load(Ordering::SeqCst), 1);
+        assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+        assert!(runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await);
+        assert_eq!(
+            recorder.counter_value(
+                "rustfs_drive_op_timeout_total",
+                &[
+                    ("endpoint", remote_disk.endpoint.to_string().as_str()),
+                    ("op", "read_version")
+                ]
+            ),
+            1
+        );
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_does_not_retry_business_errors() {
+        let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+
+        let err = remote_disk
+            .execute_read_with_retry(
+                "read_version",
+                || {
+                    attempts.fetch_add(1, Ordering::SeqCst);
+                    async { Err::<(), Error>(DiskError::FileNotFound) }
+                },
+                Duration::from_secs(1),
+            )
+            .await
+            .expect_err("business errors should be returned directly");
+
+        assert_eq!(err, DiskError::FileNotFound);
+        assert_eq!(attempts.load(Ordering::SeqCst), 1);
+        remote_disk.cancel_token.cancel();
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_honors_configured_retry_count() {
+        temp_env::async_with_vars([(rustfs_config::ENV_INTERNODE_IDEMPOTENT_READ_RETRIES, Some("2"))], async {
+            let remote_disk = new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await;
+            let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+            let started = time::Instant::now();
+            let channel = TonicEndpoint::from_shared(remote_disk.addr.clone())
+                .expect("remote disk address should parse")
+                .connect_lazy();
+            runtime_sources::cache_test_node_channel(remote_disk.addr.clone(), channel).await;
+
+            let err = remote_disk
+                .execute_read_with_retry(
+                    "read_version",
+                    || {
+                        attempts.fetch_add(1, Ordering::SeqCst);
+                        async {
+                            Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                std_io::ErrorKind::ConnectionReset,
+                                "connection reset",
+                            )))
+                        }
+                    },
+                    Duration::from_secs(1),
+                )
+                .await
+                .expect_err("exhausted retries should return the last network error");
+
+            assert!(matches!(err, DiskError::Io(ref io_err) if io_err.kind() == std_io::ErrorKind::ConnectionReset));
+            assert_eq!(attempts.load(Ordering::SeqCst), 3);
+            assert_eq!(started.elapsed(), Duration::from_millis(150));
+            assert_eq!(remote_disk.runtime_state(), RuntimeDriveHealthState::Suspect);
+            assert!(!runtime_sources::test_node_channel_is_cached(&remote_disk.addr).await);
+            remote_disk.cancel_token.cancel();
+        })
+        .await;
+    }
+
+    #[tokio::test(start_paused = true)]
+    #[serial(remote_disk_read_retry)]
+    async fn execute_read_with_retry_stops_when_disk_turns_offline_during_backoff() {
+        let remote_disk = Arc::new(new_remote_disk_with_transport(Arc::new(RecordingInternodeDataTransport::default())).await);
+        let attempts = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let first_attempt = Arc::new(tokio::sync::Notify::new());
+        let task_disk = Arc::clone(&remote_disk);
+        let task_attempts = Arc::clone(&attempts);
+        let task_first_attempt = Arc::clone(&first_attempt);
+
+        let task = tokio::spawn(async move {
+            task_disk
+                .execute_read_with_retry(
+                    "read_version",
+                    move || {
+                        let attempt = task_attempts.fetch_add(1, Ordering::SeqCst);
+                        let first_attempt = Arc::clone(&task_first_attempt);
+                        async move {
+                            if attempt == 0 {
+                                first_attempt.notify_one();
+                                return Err::<(), Error>(DiskError::Io(std_io::Error::new(
+                                    std_io::ErrorKind::ConnectionReset,
+                                    "connection reset",
+                                )));
+                            }
+                            Ok(())
+                        }
+                    },
+                    Duration::from_secs(1),
+                )
+                .await
+        });
+
+        first_attempt.notified().await;
+        tokio::task::yield_now().await;
+        remote_disk
+            .health
+            .force_runtime_state_for_test(RuntimeDriveHealthState::Offline);
+        time::advance(REMOTE_DISK_READ_RETRY_BASE_BACKOFF).await;
+        let err = task
+            .await
+            .expect("retry task should finish")
+            .expect_err("an offline disk must stop before the next attempt");
+
+        assert_eq!(err, DiskError::FaultyDisk);
+        assert_eq!(attempts.load(Ordering::SeqCst), 1);
+        remote_disk.cancel_token.cancel();
+    }
+
    #[tokio::test]
    async fn test_execute_with_timeout_evicts_cached_connection() {
        let addr = "http://127.0.0.1:59991".to_string();
@@ -5603,6 +6137,40 @@ mod tests {
        accept_task.abort();
    }

+    #[tokio::test]
+    async fn read_version_uses_the_metadata_timeout_on_a_stalled_peer() {
+        runtime_sources::ensure_test_rpc_secret();
+        let Some((base_addr, accept_task)) = spawn_stalled_grpc_peer().await else {
+            return;
+        };
+        let remote_disk = remote_disk_for_addr(&base_addr).await;
+
+        temp_env::async_with_vars(
+            [
+                (rustfs_config::ENV_DRIVE_METADATA_TIMEOUT_SECS, Some("1")),
+                (rustfs_config::ENV_DRIVE_MAX_TIMEOUT_DURATION, Some("10")),
+            ],
+            async {
+                let started = time::Instant::now();
+                let err = tokio::time::timeout(
+                    Duration::from_secs(5),
+                    remote_disk.read_version("bucket", "bucket", "object", "", &ReadOptions::default()),
+                )
+                .await
+                .expect("read_version must use the shorter metadata deadline")
+                .expect_err("a stalled peer must fail read_version");
+
+                assert!(matches!(err, DiskError::Timeout), "expected the metadata deadline to fire, got {err:?}");
+                assert!(started.elapsed() >= Duration::from_millis(900));
+                assert!(started.elapsed() < Duration::from_secs(2));
+            },
+        )
+        .await;
+
+        remote_disk.cancel_token.cancel();
+        accept_task.abort();
+    }
+
    #[tokio::test]
    async fn delete_volume_bounds_the_wait_on_a_stalled_peer() {
        runtime_sources::ensure_test_rpc_secret();
@@ -2266,15 +2266,19 @@ fn decommission_delete_marker_opts(
    version: &rustfs_filemeta::FileInfo,
    version_id: Option<String>,
    src_pool_idx: usize,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
 ) -> ObjectOptions {
+    let version_suspended = version.version_id.is_none() && version_id.is_none();
    ObjectOptions {
-        versioned: true,
-        version_id,
+        versioned: !version_suspended,
+        version_suspended,
+        version_id: version_id.or_else(|| version_suspended.then(|| uuid::Uuid::nil().to_string())),
        mod_time: version.mod_time,
        src_pool_idx,
        data_movement: true,
        delete_marker: true,
        skip_decommissioned: true,
+        expected_bucket_incarnation_id,
        delete_replication: version
            .replication_state_internal
            .as_ref()
@@ -2299,6 +2303,7 @@ fn decommission_remote_tiered_opts(
    version: &rustfs_filemeta::FileInfo,
    version_id: Option<String>,
    src_pool_idx: usize,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
 ) -> ObjectOptions {
    ObjectOptions {
        versioned: version_id.is_some(),
@@ -2307,6 +2312,9 @@ fn decommission_remote_tiered_opts(
        user_defined: version.metadata.clone(),
        src_pool_idx,
        data_movement: true,
+        include_part_checksums: true,
+        http_preconditions: Some(crate::data_movement::data_movement_target_precondition()),
+        expected_bucket_incarnation_id,
        ..Default::default()
    }
 }
@@ -2805,6 +2813,7 @@ impl ECStore {
        lifecycle_config: Option<BucketLifecycleConfiguration>,
        object_lock_config: Option<ObjectLockConfiguration>,
        replication_config: Option<(ReplicationConfiguration, OffsetDateTime)>,
+        expected_bucket_incarnation_id: Option<uuid::Uuid>,
    ) -> Result<()> {
        debug!(
            event = EVENT_DECOMMISSION_ENTRY,
@@ -2834,6 +2843,11 @@ impl ECStore {
        }
        decommission_cancel_signal_result(rx.is_cancelled())?;

+        let bucket_incarnation_fence = match expected_bucket_incarnation_id {
+            Some(expected) => Some(self.acquire_bucket_incarnation_fence(&bucket, expected).await?),
+            None => None,
+        };
+
        let mut fivs = load_decommission_entry_exact_versions(&set, &entry, &bucket, "file_info_versions").await?;

        fivs.versions
@@ -2894,7 +2908,7 @@ impl ECStore {
                    .delete_object(
                        bucket.as_str(),
                        &version.name,
-                        decommission_delete_marker_opts(version, version_id.clone(), idx),
+                        decommission_delete_marker_opts(version, version_id.clone(), idx, expected_bucket_incarnation_id),
                    )
                    .await
                {
@@ -2984,7 +2998,7 @@ impl ECStore {
                            bucket.as_str(),
                            &version.name,
                            version,
-                            &decommission_remote_tiered_opts(version, version_id.clone(), idx),
+                            &decommission_remote_tiered_opts(version, version_id.clone(), idx, expected_bucket_incarnation_id),
                        )
                        .await
                    {
@@ -3056,7 +3070,11 @@ impl ECStore {
                )
                .await?;

-                if let Err(err) = self.clone().decommission_object(idx, bucket, rd).await {
+                if let Err(err) = self
+                    .clone()
+                    .decommission_object(idx, bucket, rd, expected_bucket_incarnation_id)
+                    .await
+                {
                    if is_decommission_copy_cleanup_safe_error(&err) {
                        ignore = true;
                        cleanup_ignored = true;
@@ -3133,6 +3151,9 @@ impl ECStore {
        }

        if should_cleanup_decommission_source_entry(decommissioned, fivs.versions.len(), expired) {
+            if bucket_incarnation_fence.as_ref().is_some_and(|guard| guard.is_lock_lost()) {
+                return Err(Error::other("decommission bucket incarnation fence was lost before source cleanup"));
+            }
            decommission_cancel_signal_result(rx.is_cancelled())?;

            self.save_decommission_entry_progress_stage(
@@ -3157,6 +3178,12 @@ impl ECStore {
                entry.name.as_str(),
                &fivs,
                &cleanup_preflight_allowed_missing,
+                data_movement::SourceCleanupBucketFence {
+                    expected_incarnation_id: expected_bucket_incarnation_id,
+                    lifecycle_guard: bucket_incarnation_fence
+                        .as_ref()
+                        .and_then(|guard| guard.namespace_lock_guard()),
+                },
                "decommission",
            )
            .await
@@ -3268,6 +3295,11 @@ impl ECStore {
        let mut lifecycle_config = None;
        let mut object_lock_config = None;
        let mut replication_config = None;
+        let expected_bucket_incarnation_id = if bi.name == RUSTFS_META_BUCKET {
+            None
+        } else {
+            Some(self.bucket_incarnation_id_from_disk(&bi.name).await?)
+        };

        if bi.name != RUSTFS_META_BUCKET {
            let _ = resolve_decommission_optional_bucket_config_result(
@@ -3321,6 +3353,7 @@ impl ECStore {
                    let lifecycle_config = lifecycle_config.clone();
                    let object_lock_config = object_lock_config.clone();
                    let replication_config = replication_config.clone();
+                    let expected_bucket_incarnation_id = expected_bucket_incarnation_id;
                    let entry_error = entry_error.clone();
                    let callback_rx = callback_rx.clone();

@@ -3383,6 +3416,7 @@ impl ECStore {
                                lifecycle_config,
                                object_lock_config,
                                replication_config,
+                                expected_bucket_incarnation_id,
                            )
                            .await
                        {
@@ -4168,10 +4202,24 @@ impl ECStore {
    }

    #[tracing::instrument(skip(self, rd))]
-    async fn decommission_object(self: Arc<Self>, pool_idx: usize, bucket: String, rd: GetObjectReader) -> Result<()> {
+    async fn decommission_object(
+        self: Arc<Self>,
+        pool_idx: usize,
+        bucket: String,
+        rd: GetObjectReader,
+        expected_bucket_incarnation_id: Option<uuid::Uuid>,
+    ) -> Result<()> {
        warn!("decommission_object: start {} {}", &bucket, &rd.object_info.name);
        let object_name = rd.object_info.name.clone();
-        let result = data_movement::migrate_object(self, pool_idx, bucket.clone(), rd, "decommission_object").await;
+        let result = data_movement::migrate_object(
+            self,
+            pool_idx,
+            bucket.clone(),
+            rd,
+            expected_bucket_incarnation_id,
+            "decommission_object",
+        )
+        .await;
        if result.is_ok() {
            warn!("decommission_object: migrated {} {}", &bucket, &object_name);
        }
@@ -4347,7 +4395,8 @@ mod tests {
            ..Default::default()
        };

-        let opts = decommission_delete_marker_opts(&version, Some("version-id".to_string()), 7);
+        let incarnation = uuid::Uuid::new_v4();
+        let opts = decommission_delete_marker_opts(&version, Some("version-id".to_string()), 7, Some(incarnation));
        let replication = opts.delete_replication.expect("replication state should be preserved");

        assert!(opts.versioned);
@@ -4357,11 +4406,25 @@ mod tests {
        assert_eq!(opts.src_pool_idx, 7);
        assert_eq!(opts.version_id.as_deref(), Some("version-id"));
        assert_eq!(opts.mod_time, Some(mod_time));
+        assert_eq!(opts.expected_bucket_incarnation_id, Some(incarnation));
        assert_eq!(replication.replica_status, ReplicationStatusType::Replica);
        assert!(replication.delete_marker);
        assert_eq!(replication.replicate_decision_str, "existing");
    }

+    #[test]
+    fn decommission_delete_marker_opts_preserves_suspended_null_version() {
+        let version = rustfs_filemeta::FileInfo {
+            deleted: true,
+            ..Default::default()
+        };
+        let opts = decommission_delete_marker_opts(&version, None, 7, None);
+
+        assert!(!opts.versioned);
+        assert!(opts.version_suspended);
+        assert_eq!(opts.version_id.as_deref(), Some(uuid::Uuid::nil().to_string().as_str()));
+    }
+
    #[test]
    fn test_decommission_object_migration_read_opts_are_raw_data_movement() {
        let opts = decommission_object_migration_read_opts(Some("vid-1".to_string()));
@@ -4383,7 +4446,8 @@ mod tests {
            ..Default::default()
        };

-        let opts = decommission_remote_tiered_opts(&version, Some("version-id".to_string()), 9);
+        let incarnation = uuid::Uuid::new_v4();
+        let opts = decommission_remote_tiered_opts(&version, Some("version-id".to_string()), 9, Some(incarnation));

        assert!(opts.versioned);
        assert!(opts.data_movement);
@@ -4391,6 +4455,9 @@ mod tests {
        assert_eq!(opts.version_id.as_deref(), Some("version-id"));
        assert_eq!(opts.mod_time, Some(mod_time));
        assert_eq!(opts.user_defined.get("x-amz-meta-key").map(String::as_str), Some("value"));
+        assert!(opts.include_part_checksums);
+        assert!(opts.http_preconditions.is_some());
+        assert_eq!(opts.expected_bucket_incarnation_id, Some(incarnation));
    }

    #[test]
@@ -1638,7 +1638,7 @@ fn preserve_unknown_dirty_usage(
    Some(preserved)
 }

-#[cfg(test)]
+#[cfg(any(test, feature = "test-util"))]
 async fn replace_bucket_usage_memory_from_authoritative(bucket: &str, usage: BucketUsageInfo, refresh_started_at: SystemTime) {
    let mut cache = memory_cache().write().await;
    if let Some(existing) = cache.get(bucket)
@@ -1650,6 +1650,19 @@ async fn replace_bucket_usage_memory_from_authoritative(bucket: &str, usage: Buc
    cache.insert(bucket.to_string(), cached_bucket_usage_from_backend(usage, refresh_started_at, true));
 }

+#[cfg(feature = "test-util")]
+pub async fn seed_bucket_usage_memory_for_test(bucket: &str, size: u64) {
+    replace_bucket_usage_memory_from_authoritative(
+        bucket,
+        BucketUsageInfo {
+            size,
+            ..Default::default()
+        },
+        SystemTime::now(),
+    )
+    .await;
+}
+
 /// Fast in-memory update for immediate quota and admin usage consistency.
 pub async fn record_bucket_object_write_memory(bucket: &str, previous_current_size: Option<u64>, new_size: u64) {
    record_bucket_object_write_memory_inner(bucket, previous_current_size, new_size, false).await;
@@ -2137,30 +2150,6 @@ pub async fn load_data_usage_cache(store: &crate::set_disk::SetDisks, name: &str
    Ok(d)
 }

-#[instrument(skip(cache))]
-pub async fn save_data_usage_cache(cache: &DataUsageCache, name: &str) -> crate::error::Result<()> {
-    use crate::config::com::save_config;
-    use crate::disk::BUCKET_META_PREFIX;
-    use std::path::Path;
-
-    let Some(store) = runtime_sources::object_store_handle() else {
-        return Err(Error::other("errServerNotInitialized"));
-    };
-    let buf = cache.marshal_msg().map_err(Error::other)?;
-    let buf_clone = buf.clone();
-
-    let store_clone = store.clone();
-
-    let name = Path::new(BUCKET_META_PREFIX).join(name).to_string_lossy().to_string();
-
-    let name_clone = name.clone();
-    tokio::spawn(async move {
-        let _ = save_config(store_clone, &format!("{}{}", name_clone, ".bkp"), buf_clone).await;
-    });
-    save_config(store, &name, buf).await?;
-    Ok(())
-}
-
 /// Persist the current in-memory compression total to the backend.
 /// Resets the debounce counter so the next auto-persist won't fire
 /// immediately after this manual flush (intended for shutdown paths).
@@ -12,7 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.

-use crate::cluster::rpc::{TonicInterceptor, gen_tonic_signature_interceptor, node_service_time_out_client};
+use crate::cluster::rpc::{
+    ScannerBucketListing, TonicInterceptor, gen_tonic_signature_interceptor, node_service_time_out_client,
+};
 use crate::data_usage::{DATA_USAGE_CACHE_NAME, DATA_USAGE_ROOT, load_data_usage_from_backend_cached};
 use crate::error::{Error, Result};
 use crate::{
@@ -23,6 +25,7 @@ use crate::{

 use crate::data_usage::load_data_usage_cache;
 use crate::storage_api_contracts::admin::StorageAdminApi;
+use crate::storage_api_contracts::bucket::BucketOptions;
 use rustfs_common::heal_channel::DriveState;
 use rustfs_madmin::{
    BackendDisks, Disk, ErasureSetInfo, ITEM_INITIALIZING, ITEM_OFFLINE, ITEM_ONLINE, ITEM_UNKNOWN, InfoMessage, MemStats,
@@ -74,6 +77,19 @@ fn apply_data_usage_result(
    }
 }

+fn apply_bucket_namespace_count(result: Result<ScannerBucketListing>, buckets: &mut rustfs_madmin::Buckets) {
+    if let Ok(listing) = result
+        && listing.topology_complete
+    {
+        let count = listing.buckets.iter().filter(|bucket| !bucket.name.starts_with('.')).count();
+        let Ok(count) = u64::try_from(count) else {
+            return;
+        };
+        buckets.count = count;
+        buckets.error = None;
+    }
+}
+
 // pub const ITEM_OFFLINE: &str = "offline";
 // pub const ITEM_INITIALIZING: &str = "initializing";
 // pub const ITEM_ONLINE: &str = "online";
@@ -285,6 +301,18 @@ pub async fn get_server_info(get_pools: bool) -> InfoMessage {
            &mut delete_markers,
            &mut usage,
        );
+        if buckets.error.is_some() {
+            apply_bucket_namespace_count(
+                store
+                    .list_bucket_for_scanner(&BucketOptions {
+                        cached: true,
+                        no_metadata: true,
+                        ..Default::default()
+                    })
+                    .await,
+                &mut buckets,
+            );
+        }

        let after3 = OffsetDateTime::now_utc();

@@ -705,12 +733,13 @@ mod tests {
        endpoints::{EndpointServerPools, Endpoints, PoolEndpoints},
    };
    use crate::runtime::sources as runtime_sources;
+    use crate::storage_api_contracts::bucket::BucketInfo;
    use rustfs_madmin::{Disk, ITEM_OFFLINE, ITEM_ONLINE, ITEM_UNKNOWN, ServerProperties};

    use super::{
-        DATA_USAGE_ROOT, DATA_USAGE_UNAVAILABLE_ERROR, apply_data_usage_result, apply_erasure_set_usage,
-        get_local_server_property, get_online_offline_disks_stats, get_server_info, reconcile_servers_with_endpoint_topology,
-        server_topology_completeness_report,
+        DATA_USAGE_ROOT, DATA_USAGE_UNAVAILABLE_ERROR, apply_bucket_namespace_count, apply_data_usage_result,
+        apply_erasure_set_usage, get_local_server_property, get_online_offline_disks_stats, get_server_info,
+        reconcile_servers_with_endpoint_topology, server_topology_completeness_report,
    };

    fn disk_with_state(endpoint: &str, state: &str) -> Disk {
@@ -960,6 +989,75 @@ mod tests {
        assert_eq!(usage.error.as_deref(), Some(DATA_USAGE_UNAVAILABLE_ERROR));
    }

+    #[test]
+    fn live_bucket_namespace_count_survives_unavailable_data_usage() {
+        let mut buckets = rustfs_madmin::Buckets {
+            count: 0,
+            error: Some(DATA_USAGE_UNAVAILABLE_ERROR.to_string()),
+        };
+
+        apply_bucket_namespace_count(
+            Ok(crate::cluster::rpc::ScannerBucketListing {
+                buckets: vec![
+                    BucketInfo {
+                        name: "bucket-a".to_string(),
+                        ..Default::default()
+                    },
+                    BucketInfo {
+                        name: ".rustfs.sys".to_string(),
+                        ..Default::default()
+                    },
+                    BucketInfo {
+                        name: "bucket-b".to_string(),
+                        ..Default::default()
+                    },
+                ],
+                set_buckets: Vec::new(),
+                topology_complete: true,
+            }),
+            &mut buckets,
+        );
+
+        assert_eq!(buckets.count, 2);
+        assert_eq!(buckets.error, None);
+    }
+
+    #[test]
+    fn incomplete_bucket_namespace_lookup_preserves_usage_state() {
+        let mut buckets = rustfs_madmin::Buckets {
+            count: 7,
+            error: Some(DATA_USAGE_UNAVAILABLE_ERROR.to_string()),
+        };
+
+        apply_bucket_namespace_count(
+            Ok(crate::cluster::rpc::ScannerBucketListing {
+                buckets: vec![BucketInfo {
+                    name: "bucket-a".to_string(),
+                    ..Default::default()
+                }],
+                set_buckets: Vec::new(),
+                topology_complete: false,
+            }),
+            &mut buckets,
+        );
+
+        assert_eq!(buckets.count, 7);
+        assert_eq!(buckets.error.as_deref(), Some(DATA_USAGE_UNAVAILABLE_ERROR));
+    }
+
+    #[test]
+    fn failed_bucket_namespace_lookup_preserves_usage_state() {
+        let mut buckets = rustfs_madmin::Buckets {
+            count: 7,
+            error: Some(DATA_USAGE_UNAVAILABLE_ERROR.to_string()),
+        };
+
+        apply_bucket_namespace_count(Err(crate::error::Error::DiskNotFound), &mut buckets);
+
+        assert_eq!(buckets.count, 7);
+        assert_eq!(buckets.error.as_deref(), Some(DATA_USAGE_UNAVAILABLE_ERROR));
+    }
+
    #[test]
    fn incomplete_erasure_set_cache_is_not_reported_as_zero() {
        let mut cache = rustfs_data_usage::DataUsageCache::default();
@@ -137,6 +137,7 @@ pub(crate) const GET_METADATA_CACHE_REASON_NO_LOCK: &str = "no_lock";
 pub(crate) const GET_METADATA_CACHE_REASON_NOT_FOUND_OR_EXPIRED: &str = "not_found_or_expired";
 pub(crate) const GET_METADATA_CACHE_REASON_NOT_READ_DATA: &str = "not_read_data";
 pub(crate) const GET_METADATA_CACHE_REASON_PART_NUMBER: &str = "part_number";
+pub(crate) const GET_METADATA_CACHE_REASON_PART_CHECKSUMS: &str = "part_checksums";
 pub(crate) const GET_METADATA_CACHE_REASON_RAW_DATA_MOVEMENT_READ: &str = "raw_data_movement_read";
 pub(crate) const GET_METADATA_CACHE_REASON_STALE_PUBLICATION: &str = "stale_publication";
 pub(crate) const GET_METADATA_CACHE_REASON_USABLE: &str = "usable";
@@ -480,6 +481,7 @@ mod tests {
        assert_eq!(GET_METADATA_CACHE_REASON_NO_LOCK, "no_lock");
        assert_eq!(GET_METADATA_CACHE_REASON_NOT_FOUND_OR_EXPIRED, "not_found_or_expired");
        assert_eq!(GET_METADATA_CACHE_REASON_NOT_READ_DATA, "not_read_data");
+        assert_eq!(GET_METADATA_CACHE_REASON_PART_CHECKSUMS, "part_checksums");
        assert_eq!(GET_METADATA_CACHE_REASON_PART_NUMBER, "part_number");
        assert_eq!(GET_METADATA_CACHE_REASON_RAW_DATA_MOVEMENT_READ, "raw_data_movement_read");
        assert_eq!(GET_METADATA_CACHE_REASON_STALE_PUBLICATION, "stale_publication");
@@ -2022,6 +2022,21 @@ impl DiskAPI for LocalDiskWrapper {
        .await
    }

+    async fn read_file_stream_chunks(
+        &self,
+        volume: &str,
+        path: &str,
+        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
+    }
+
    async fn read_file_mmap_copy(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<bytes::Bytes> {
        self.track_disk_health_with_op(
            "read_file_mmap_copy",
@@ -113,6 +113,9 @@ pub enum DiskError {
    #[error("bit-rot hash algorithm is invalid")]
    BitrotHashAlgoInvalid,

+    /// Never constructed locally by RustFS (only reachable through wire
+    /// decoding, and no current node sends it). The wire code is kept for
+    /// cross-version compatibility — do not renumber or remove (backlog#1831).
    #[error("Rename across devices not allowed, please fix your backend configuration")]
    CrossDeviceLink,

@@ -143,6 +146,9 @@ pub enum DiskError {
    #[error("io error {0}")]
    Io(#[source] io::Error),

+    /// Never constructed locally by RustFS (only reachable through wire
+    /// decoding, and no current node sends it). The wire code is kept for
+    /// cross-version compatibility — do not renumber or remove (backlog#1831).
    #[error("source stalled")]
    SourceStalled,

@@ -331,7 +337,14 @@ impl From<std::io::Error> for DiskError {
        }
        match e.downcast::<DiskError>() {
            Ok(disk_error) => disk_error,
-            Err(io_error) => DiskError::Io(io_error),
+            // Mirror `From<io::Error> for StorageError`: a StorageError boxed
+            // through `From<StorageError> for io::Error` must recover its typed
+            // classification instead of degrading to `DiskError::Io`, which
+            // quorum aggregation (`reduce_errs`) would count as a distinct error.
+            Err(io_error) => match io_error.downcast::<crate::error::StorageError>() {
+                Ok(storage_error) => storage_error.into(),
+                Err(io_error) => DiskError::Io(io_error),
+            },
        }
    }
 }
@@ -635,19 +648,6 @@ impl Hash for DiskError {
 // is currently commented out to avoid complexity. These can be re-enabled
 // when needed for specific disk quorum checking and error aggregation logic.

-/// Bitrot errors
-#[derive(Debug, thiserror::Error)]
-pub enum BitrotErrorType {
-    #[error("bitrot checksum verification failed")]
-    BitrotChecksumMismatch { expected: String, got: String },
-}
-
-impl From<BitrotErrorType> for DiskError {
-    fn from(e: BitrotErrorType) -> Self {
-        DiskError::other(e)
-    }
-}
-
 /// Context wrapper for file access errors
 #[derive(Debug, thiserror::Error)]
 pub struct FileAccessDeniedWithContext {
@@ -862,19 +862,6 @@ mod tests {
        let _disk_error: DiskError = json_error.into();
    }

-    #[test]
-    fn test_bitrot_error_type() {
-        let bitrot_error = BitrotErrorType::BitrotChecksumMismatch {
-            expected: "abc123".to_string(),
-            got: "def456".to_string(),
-        };
-
-        assert!(bitrot_error.to_string().contains("bitrot checksum verification failed"));
-
-        let disk_error: DiskError = bitrot_error.into();
-        assert!(matches!(disk_error, DiskError::Io(_)));
-    }
-
    #[test]
    fn test_file_access_denied_with_context() {
        let path = PathBuf::from("/test/path");
@@ -953,6 +940,27 @@ mod tests {
        assert_eq!(original_disk_error, recovered_disk_error);
    }

+    #[test]
+    fn test_io_error_with_storage_error_inside() {
+        use crate::error::StorageError;
+
+        // An io::Error boxing a disk-representable StorageError (as produced by
+        // `From<StorageError> for io::Error`) must recover the typed DiskError
+        // variant instead of degrading to an opaque DiskError::Io.
+        let io_with_storage_error: std::io::Error = StorageError::FaultyRemoteDisk.into();
+        let recovered: DiskError = io_with_storage_error.into();
+        assert_eq!(recovered, DiskError::FaultyRemoteDisk);
+
+        let io_with_storage_error: std::io::Error = StorageError::FileAccessDenied.into();
+        let recovered: DiskError = io_with_storage_error.into();
+        assert_eq!(recovered, DiskError::FileAccessDenied);
+
+        // A StorageError with no DiskError analog stays an opaque Io error.
+        let io_with_bucket_error: std::io::Error = StorageError::BucketNotFound("bucket".to_string()).into();
+        let recovered: DiskError = io_with_bucket_error.into();
+        assert!(matches!(recovered, DiskError::Io(_)));
+    }
+
    #[test]
    fn test_io_error_different_kinds() {
        use std::io::ErrorKind;
@@ -2027,6 +2027,9 @@ type InlinePreparationHook = Box<dyn FnOnce() + Send>;
 static INLINE_PREPARATION_BEFORE_BACKUP: std::sync::LazyLock<std::sync::Mutex<HashMap<String, InlinePreparationHook>>> =
    std::sync::LazyLock::new(|| std::sync::Mutex::new(HashMap::new()));
 #[cfg(test)]
+static INLINE_BEFORE_FILE_SYNC_ADMISSION: std::sync::LazyLock<std::sync::Mutex<HashMap<String, InlinePreparationHook>>> =
+    std::sync::LazyLock::new(|| std::sync::Mutex::new(HashMap::new()));
+#[cfg(test)]
 static RENAME_DATA_AFTER_FIRST_PUBLICATION: std::sync::LazyLock<std::sync::Mutex<HashMap<String, InlinePreparationHook>>> =
    std::sync::LazyLock::new(|| std::sync::Mutex::new(HashMap::new()));
 #[cfg(test)]
@@ -2091,6 +2094,14 @@ fn set_inline_preparation_before_backup(dst_path: &str, hook: impl FnOnce() + Se
        .insert(dst_path.to_string(), Box::new(hook));
 }

+#[cfg(test)]
+fn set_inline_before_file_sync_admission(dst_path: &str, hook: impl FnOnce() + Send + 'static) {
+    INLINE_BEFORE_FILE_SYNC_ADMISSION
+        .lock()
+        .expect("test admission hook lock should not be poisoned")
+        .insert(dst_path.to_string(), Box::new(hook));
+}
+
 #[cfg(test)]
 fn set_rename_data_after_first_publication(dst_path: &str, hook: impl FnOnce() + Send + 'static) {
    RENAME_DATA_AFTER_FIRST_PUBLICATION
@@ -2236,6 +2247,17 @@ fn run_inline_preparation_before_backup(dst_path: &str) {
    }
 }

+#[cfg(test)]
+fn run_inline_before_file_sync_admission(dst_path: &str) {
+    let hook = INLINE_BEFORE_FILE_SYNC_ADMISSION
+        .lock()
+        .expect("test admission hook lock should not be poisoned")
+        .remove(dst_path);
+    if let Some(hook) = hook {
+        hook();
+    }
+}
+
 #[cfg(test)]
 fn run_rename_data_after_first_publication(dst_path: &str) {
    let hook = RENAME_DATA_AFTER_FIRST_PUBLICATION
@@ -2907,23 +2929,82 @@ pub(crate) trait LocalIoBackend: Send + Sync + Debug + 'static {
 /// Default [`LocalIoBackend`]: tokio blocking-pool file I/O plus the
 /// mmap-copy / direct-read-copy positioned read, moved verbatim from the
 /// former `DiskAPI` method bodies on `LocalDisk`.
-#[derive(Debug)]
 pub(crate) struct StdBackend {
    root: PathBuf,
    #[cfg(target_os = "linux")]
    direct_io: Arc<DirectIoReadState>,
    #[cfg(target_os = "linux")]
    direct_io_write: Arc<DirectIoWriteState>,
+    /// Per-disk descriptor cache for buffered reads (rustfs/backlog#1801).
+    /// `None` when disabled by env, blocked by a low `RLIMIT_NOFILE`, or on
+    /// non-Linux (where the cache type is unavailable). Like the io_uring
+    /// cache, only the buffered read path populates it; O_DIRECT reads keep
+    /// opening their own aligned descriptors.
+    #[cfg(target_os = "linux")]
+    fd_cache: Option<FdCache>,
+}
+
+// Manual `Debug` mirrors `UringBackend`: the fd cache (and the Linux-only
+// direct-IO state) hold types that do not implement `Debug`, so a derive would
+// force `FdCache: Debug`. `finish_non_exhaustive` skips them.
+impl std::fmt::Debug for StdBackend {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("StdBackend").field("root", &self.root).finish_non_exhaustive()
+    }
 }

 impl StdBackend {
    pub(crate) fn new(root: PathBuf) -> Self {
+        Self::build(root, true)
+    }
+
+    /// Construct without the descriptor cache.
+    ///
+    /// `UringBackend` wraps a `StdBackend` and runs its own `FdCache` over the
+    /// same positioned reads. If the inner `StdBackend` also built a cache, a
+    /// fallback read (`UringBackend::pread_bytes` delegates to the inner backend
+    /// on latch-off / O_DIRECT / buffered errors) would populate a *second*
+    /// cache that `UringBackend`'s invalidation never touches — re-opening the
+    /// stale-inode hazard `FdCache` exists to close (rustfs/backlog#1176/#1801).
+    /// The wrapper therefore owns the only cache for the disk; the inner backend
+    /// opens per read. This also avoids double-counting `FD_CACHE_CAPACITY`
+    /// against `RLIMIT_NOFILE` (rustfs/backlog#1178).
+    #[cfg(target_os = "linux")]
+    pub(crate) fn new_without_fd_cache(root: PathBuf) -> Self {
+        Self::build(root, false)
+    }
+
+    fn build(root: PathBuf, build_fd_cache: bool) -> Self {
+        // Gate the fd cache on RLIMIT_NOFILE headroom (rustfs/backlog#1178):
+        // 512 fds/disk with a low soft limit and several disks would hit EMFILE.
+        // Fall back to open-per-read when the limit is too small.
+        #[cfg(target_os = "linux")]
+        let fd_cache = if build_fd_cache && is_local_fd_cache_enabled() {
+            if rlimit_allows_fd_cache() {
+                Some(FdCache::new())
+            } else {
+                warn!(
+                    component = LOG_COMPONENT_ECSTORE,
+                    subsystem = LOG_SUBSYSTEM_DISK_LOCAL,
+                    "std fd cache disabled: RLIMIT_NOFILE soft limit too low for 512 fds/disk; using open-per-read"
+                );
+                None
+            }
+        } else {
+            None
+        };
+        // `build_fd_cache` is only consulted on Linux (for the fd cache); on
+        // other platforms it has no effect and would trip the unused-variable lint.
+        #[cfg(not(target_os = "linux"))]
+        let _ = build_fd_cache;
        Self {
            root,
            #[cfg(target_os = "linux")]
            direct_io: Arc::new(DirectIoReadState::new()),
            #[cfg(target_os = "linux")]
            direct_io_write: Arc::new(DirectIoWriteState::new()),
+            #[cfg(target_os = "linux")]
+            fd_cache,
        }
    }

@@ -3001,6 +3082,9 @@ impl LocalIoBackend for StdBackend {
                direct_read_copy_fault_delta: MmapPageFaultDelta,
                blocking_task_duration: StdDuration,
                used_direct_io: bool,
+                /// The descriptor opened by THIS call (None on a cache hit), handed
+                /// back so the async caller can index it in the fd cache.
+                opened_fd: Option<Arc<std::fs::File>>,
            }

            enum MmapCopyReadError {
@@ -3028,28 +3112,72 @@ impl LocalIoBackend for StdBackend {
            let direct_io_state = self.direct_io.clone();
            let offset_u64 = u64::try_from(offset).map_err(|_| DiskError::FileCorrupt)?;
            let end_offset_u64 = u64::try_from(end_offset).map_err(|_| DiskError::FileCorrupt)?;
+
+            // Descriptor cache (rustfs/backlog#1801): on a hit the read reuses an
+            // already-open descriptor (via dup below) and skips `access` +
+            // `File::open`. Linux-only — on other Unix `cached_fd` is None and the
+            // read opens per call exactly as before. `fd_lookup` snapshots the
+            // invalidation generation BEFORE the open so a heal/delete that lands
+            // while the blocking open is in flight prevents the now-stale descriptor
+            // from being inserted (rustfs/backlog#1176).
+            #[cfg(target_os = "linux")]
+            let fd_lookup = self.fd_cache.as_ref().map(|cache| {
+                let key = FdKey {
+                    volume: volume.to_owned(),
+                    path: path.to_owned(),
+                    direct: false,
+                };
+                let gen_at_open = cache.generation();
+                (cache, key, gen_at_open)
+            });
+            #[cfg(target_os = "linux")]
+            let cached_fd: Option<Arc<std::fs::File>> = match &fd_lookup {
+                Some((cache, key, _)) => cache.get(key).await,
+                None => None,
+            };
+            #[cfg(not(target_os = "linux"))]
+            let cached_fd: Option<Arc<std::fs::File>> = None;
+
            let blocking_wait_start = metrics_enabled.then(std::time::Instant::now);
            let read_result = tokio::task::spawn_blocking(move || {
                let blocking_task_start = metrics_enabled.then(StdInstant::now);

-                let access_check_start = metrics_enabled.then(StdInstant::now);
-                let volume_dir = local_disk_bucket_path(&root, &volume_owned)?;
-                if !skip_access_checks(&volume_owned) {
-                    crate::disk::fs::access_std(&volume_dir)
-                        .map_err(|e| DiskError::from(to_access_error(e, DiskError::VolumeAccessDenied)))?;
-                }
-                let access_check_duration = access_check_start.map_or(StdDuration::ZERO, |started_at| started_at.elapsed());
-
+                // Resolve the part path unconditionally: the O_DIRECT branch (large
+                // reads) opens its own aligned descriptor by path even on a cache hit.
                let path_resolve_start = metrics_enabled.then(StdInstant::now);
                let file_path = local_disk_object_path(&root, &volume_owned, &path_owned)?;
                check_path_length(file_path.to_string_lossy().as_ref())?;
                let path_resolve_duration = path_resolve_start.map_or(StdDuration::ZERO, |started_at| started_at.elapsed());

                let file_open_start = metrics_enabled.then(StdInstant::now);
-                let mut file = std::fs::File::open(&file_path).map_err(DiskError::from)?;
+                // Acquire the read handle (rustfs/backlog#1801). On a descriptor-cache
+                // hit this reuses the cached descriptor via `dup` (one syscall, no path
+                // resolution or permission re-check) and skips the volume access probe;
+                // on a miss it resolves the volume, access-checks, and opens the file.
+                // `File::try_clone` shares the cached descriptor's open-file offset, so
+                // the read below is positioned (mmap offset argument / `read_exact_at`)
+                // and never depends on the descriptor's current offset. `cached_fd` being
+                // None also marks this call as a miss for the cache-insert side-channel.
+                let (file, access_check_duration) = if let Some(cached) = cached_fd.as_ref() {
+                    (cached.as_ref().try_clone().map_err(DiskError::from)?, StdDuration::ZERO)
+                } else {
+                    // Measure the volume access probe only — the part-path resolution
+                    // above is accounted in `path_resolve_duration` (rustfs/backlog#1801).
+                    let access_check_start = metrics_enabled.then(StdInstant::now);
+                    let volume_dir = local_disk_bucket_path(&root, &volume_owned)?;
+                    if !skip_access_checks(&volume_owned) {
+                        crate::disk::fs::access_std(&volume_dir)
+                            .map_err(|e| DiskError::from(to_access_error(e, DiskError::VolumeAccessDenied)))?;
+                    }
+                    let access_check_duration = access_check_start.map_or(StdDuration::ZERO, |started_at| started_at.elapsed());
+                    (std::fs::File::open(&file_path).map_err(DiskError::from)?, access_check_duration)
+                };
                let file_open_duration = file_open_start.map_or(StdDuration::ZERO, |started_at| started_at.elapsed());

                let metadata_lookup_start = metrics_enabled.then(StdInstant::now);
+                // On a cache hit this fstats the cached descriptor — the inode it was
+                // opened against, which invalidation keeps current for live entries. EC
+                // shards are fixed-length, so a still-cached pre-heal length is benign.
                let meta = file.metadata().map_err(DiskError::from)?;
                let metadata_lookup_duration = metadata_lookup_start.map_or(StdDuration::ZERO, |started_at| started_at.elapsed());

@@ -3160,13 +3288,15 @@ impl LocalIoBackend for StdBackend {
                            bytes
                        }
                        LocalReadCopyMethod::DirectReadCopy => {
-                            use std::io::{Read as _, Seek as _};
+                            use std::os::unix::fs::FileExt;

                            let direct_read_copy_start = metrics_enabled.then(StdInstant::now);
                            let direct_read_copy_faults_before = read_mmap_page_fault_counts(metrics_enabled);
-                            file.seek(SeekFrom::Start(offset_u64)).map_err(DiskError::from)?;
                            let mut buffer = vec![0; length];
-                            file.read_exact(&mut buffer).map_err(DiskError::from)?;
+                            // Positioned read: a cache hit reads through a `dup`'d handle
+                            // that shares the cached descriptor's offset, so this must not
+                            // touch the descriptor offset (rustfs/backlog#1801).
+                            file.read_exact_at(&mut buffer, offset_u64).map_err(DiskError::from)?;
                            let direct_read_copy_faults_after = read_mmap_page_fault_counts(metrics_enabled);
                            direct_read_copy_duration =
                                direct_read_copy_start.map_or(StdDuration::ZERO, |started_at| started_at.elapsed());
@@ -3193,6 +3323,16 @@ impl LocalIoBackend for StdBackend {

                let blocking_task_duration = blocking_task_start.map_or(StdDuration::ZERO, |started_at| started_at.elapsed());

+                // Hand the freshly opened descriptor back so the async caller can index
+                // the cache — None on a hit (the cache already holds it). mmap/reclaim
+                // above only borrowed `file`, so it is still owned here and moves into the
+                // Arc; `cached_fd.is_none()` is true exactly when this call did the open.
+                // Non-Linux has no fd cache, so skip the Arc allocation there.
+                #[cfg(target_os = "linux")]
+                let opened_fd: Option<Arc<std::fs::File>> = cached_fd.is_none().then(|| Arc::new(file));
+                #[cfg(not(target_os = "linux"))]
+                let opened_fd: Option<Arc<std::fs::File>> = None;
+
                Ok::<MmapCopyReadResult, MmapCopyReadError>(MmapCopyReadResult {
                    bytes,
                    access_check_duration,
@@ -3208,6 +3348,7 @@ impl LocalIoBackend for StdBackend {
                    direct_read_copy_fault_delta,
                    blocking_task_duration,
                    used_direct_io,
+                    opened_fd,
                })
            })
            .await
@@ -3313,6 +3454,16 @@ impl LocalIoBackend for StdBackend {
                    }
                }
            }
+            // Index the freshly opened descriptor for future cache hits
+            // (rustfs/backlog#1801). `insert_if_fresh` refuses to cache if an
+            // invalidation (heal/delete/rename) bumped the generation between the
+            // open snapshot and now, so a stale pre-mutation inode is never served
+            // (rustfs/backlog#1176). On a cache hit `opened_fd` is None; on non-Linux
+            // there is no fd cache, so this is gated out entirely.
+            #[cfg(target_os = "linux")]
+            if let (Some((cache, key, gen_at_open)), Some(opened)) = (fd_lookup, read_result.opened_fd) {
+                cache.insert_if_fresh(key, opened, gen_at_open).await;
+            }
            let bytes = read_result.bytes;

            // Log successful mmap read metrics
@@ -3516,6 +3667,39 @@ impl LocalIoBackend for StdBackend {
            }
        }
    }
+
+    // Descriptor-cache invalidation for StdBackend (rustfs/backlog#1801). On
+    // non-Linux `fd_cache` does not exist, so these overrides are absent and the
+    // trait's default no-op impls apply. On Linux they mirror UringBackend so
+    // the existing LocalDisk mutation hooks (rename_data/rename_file/delete/
+    // delete_volume/close) drop stale descriptors on every inode swap.
+    #[cfg(target_os = "linux")]
+    async fn invalidate_cached_fd(&self, volume: &str, path: &str) {
+        if let Some(cache) = self.fd_cache.as_ref() {
+            cache.invalidate_exact(volume, path).await;
+        }
+    }
+
+    #[cfg(target_os = "linux")]
+    fn invalidate_cached_fds_under(&self, volume: &str, path: &str) {
+        if let Some(cache) = self.fd_cache.as_ref() {
+            cache.invalidate_under(volume, path);
+        }
+    }
+
+    #[cfg(target_os = "linux")]
+    fn invalidate_cached_fds_for_volume(&self, volume: &str) {
+        if let Some(cache) = self.fd_cache.as_ref() {
+            cache.invalidate_volume(volume);
+        }
+    }
+
+    #[cfg(target_os = "linux")]
+    async fn clear_cached_fds(&self) {
+        if let Some(cache) = self.fd_cache.as_ref() {
+            cache.clear();
+        }
+    }
 }

 /// Enable the per-disk descriptor cache for io_uring reads (backlog#1145).
@@ -3541,6 +3725,20 @@ fn is_io_uring_fd_cache_enabled() -> bool {
    rustfs_utils::get_env_bool(ENV_RUSTFS_IO_URING_FD_CACHE, DEFAULT_RUSTFS_IO_URING_FD_CACHE)
 }

+/// Enable the per-disk descriptor cache for the default `StdBackend` reads
+/// (rustfs/backlog#1801). Independent of the io_uring switch so each backend is
+/// separately controllable; both share the same `rlimit_allows_fd_cache` guard
+/// because each may hold up to `FD_CACHE_CAPACITY` (512) descriptors per disk.
+#[cfg(target_os = "linux")]
+const ENV_RUSTFS_LOCAL_FD_CACHE: &str = "RUSTFS_LOCAL_FD_CACHE";
+#[cfg(target_os = "linux")]
+const DEFAULT_RUSTFS_LOCAL_FD_CACHE: bool = true;
+
+#[cfg(target_os = "linux")]
+fn is_local_fd_cache_enabled() -> bool {
+    rustfs_utils::get_env_bool(ENV_RUSTFS_LOCAL_FD_CACHE, DEFAULT_RUSTFS_LOCAL_FD_CACHE)
+}
+
 /// Whether the soft `RLIMIT_NOFILE` has enough headroom to run the fd cache
 /// safely (rustfs/backlog#1178). The cache holds up to `FD_CACHE_CAPACITY` (512)
 /// descriptors PER DISK and `try_new` cannot know the disk count, so a low limit
@@ -3946,7 +4144,7 @@ impl UringBackend {
                // struct (rustfs/backlog#1185).
                let root_label = root.display().to_string();
                Some(Self {
-                    inner: StdBackend::new(root.clone()),
+                    inner: StdBackend::new_without_fd_cache(root.clone()),
                    root,
                    root_label,
                    driver: std::mem::ManuallyDrop::new(driver),
@@ -8976,7 +9174,7 @@ impl DiskAPI for LocalDisk {

            if let Some(src_file_path_parent) = src_file_path.parent() {
                if src_volume != super::RUSTFS_META_MULTIPART_BUCKET {
-                    let _ = remove_std(src_file_path_parent);
+                    let _ = std::fs::remove_dir(src_file_path_parent);
                } else {
                    let _ = self
                        .delete_file(&dst_volume_dir, &src_file_path_parent.to_path_buf(), true, false)
@@ -9021,7 +9219,20 @@ impl DiskAPI for LocalDisk {
            #[cfg(windows)]
            let source_parent = src_file_parent.to_path_buf();
            let rename_commit_guard_for_preparation = rename_commit_guard.clone();
-            let inline_preparation = os::run_blocking_namespace_operation(mutation_lease.clone(), move || {
+            let sync = durability.syncs_commit_metadata();
+            #[cfg(test)]
+            run_inline_before_file_sync_admission(dst_path);
+            let mut file_sync_admission = if sync {
+                Some(
+                    os::acquire_file_sync_admission(self.file_sync_permits.clone())
+                        .await
+                        .map_err(to_file_error)
+                        .map_err(DiskError::from)?,
+                )
+            } else {
+                None
+            };
+            let prepare_inline_metadata = move || {
                let mut prepared_metadata_source =
                    os::create_prepared_rename_source_with_commit_guard(&src, &dst, &rename_commit_guard_for_preparation)?;
                #[cfg(windows)]
@@ -9058,7 +9269,6 @@ impl DiskAPI for LocalDisk {
                        None
                    }
                });
-                let sync = durability.syncs_commit_metadata();
                let mut staged_rollback_path = None;
                if let Some(d) = old_data_dir.as_ref() {
                    let _ = xlmeta.data.remove_two(version_id, *d);
@@ -9103,8 +9313,12 @@ impl DiskAPI for LocalDisk {
                    has_dst_buf.is_none(),
                    prepared_metadata_source,
                ))
-            })
-            .await
+            };
+            let inline_preparation = if let Some(admission) = file_sync_admission.as_ref() {
+                os::run_blocking_namespace_file_sync_operation(mutation_lease.clone(), admission, prepare_inline_metadata).await
+            } else {
+                os::run_blocking_namespace_operation(mutation_lease.clone(), prepare_inline_metadata).await
+            }
            .map_err(to_file_error)
            .map_err(DiskError::from);

@@ -9154,14 +9368,26 @@ impl DiskAPI for LocalDisk {
                let backup_path = dst_parent
                    .join(rollback_data_dir.to_string())
                    .join(STORAGE_FORMAT_FILE_BACKUP);
+                // rename_all acquires the backup path's namespace lease. Do not
+                // hold a disk admission while acquiring another namespace lock.
+                drop(file_sync_admission.take());
                if let Err(err) = rename_all(staged_backup, &backup_path, &dst_volume_dir, &self.publication_root).await {
                    let _ = remove_file_if_exists(staged_backup);
                    return Err(err);
                }
                run_rename_data_after_first_publication(dst_path);
-                if durability.syncs_commit_metadata()
+                if sync {
+                    file_sync_admission = Some(
+                        os::acquire_file_sync_admission(self.file_sync_permits.clone())
+                            .await
+                            .map_err(to_file_error)
+                            .map_err(DiskError::from)?,
+                    );
+                }
+                if let Some(admission) = file_sync_admission.as_ref()
                    && let Some(backup_parent) = backup_path.parent()
-                    && let Err(err) = os::fsync_dir(backup_parent).await
+                    && let Err(err) =
+                        os::fsync_dir_with_namespace_file_sync_limit(backup_parent, mutation_lease.clone(), admission).await
                {
                    return Err(DiskError::from(to_file_error(err)));
                }
@@ -9199,9 +9425,10 @@ impl DiskAPI for LocalDisk {
                }

                // Persist the commit rename's directory entry across power loss.
-                if durability.syncs_commit_metadata()
+                if let Some(admission) = file_sync_admission.as_ref()
                    && let Some(dst_parent) = dst_file_path.parent()
-                    && let Err(err) = os::fsync_dir(dst_parent).await
+                    && let Err(err) =
+                        os::fsync_dir_with_namespace_file_sync_limit(dst_parent, mutation_lease.clone(), admission).await
                {
                    rollback_inline_metadata_commit_std(&dst_file_path, rollback_data_dir, local_rollback_path.as_deref())?;
                    return Err(err);
@@ -9214,13 +9441,17 @@ impl DiskAPI for LocalDisk {
                // not its own entry, so for a new inline object fsync the ancestor
                // chain up to and including the bucket. Overwrites already have a
                // durable object dir; the starts_with guard bounds the walk.
-                if durability.syncs_commit_metadata() && destination_was_absent {
+                if let Some(admission) = file_sync_admission.as_ref()
+                    && destination_was_absent
+                {
                    let mut ancestor = dst_file_path.parent().and_then(|object_dir| object_dir.parent());
                    while let Some(ancestor_dir) = ancestor {
                        if !ancestor_dir.starts_with(&dst_volume_dir) {
                            break;
                        }
-                        if let Err(err) = os::fsync_dir(ancestor_dir).await {
+                        if let Err(err) =
+                            os::fsync_dir_with_namespace_file_sync_limit(ancestor_dir, mutation_lease.clone(), admission).await
+                        {
                            rollback_inline_metadata_commit_std(
                                &dst_file_path,
                                rollback_data_dir,
@@ -9239,6 +9470,10 @@ impl DiskAPI for LocalDisk {
            }
            .await;

+            // The disk admission protects the durability chain, not staging
+            // cleanup or cache invalidation after that chain has completed.
+            drop(file_sync_admission.take());
+
            // A post-commit rollback (for example, a commit-metadata fsync
            // failure under strict durability) restores the old metadata; drop any
            // descriptors cached during the committed window before propagating the
@@ -9264,7 +9499,7 @@ impl DiskAPI for LocalDisk {
            if let Some(ref cleanup) = cleanup_path {
                let _ = self.delete_file(&dst_volume_dir, cleanup, true, false).await;
            } else if let Some(parent) = src_file_path.parent() {
-                let _ = remove_std(parent);
+                let _ = std::fs::remove_dir(parent);
            }

            // Heal reuses a version's `data_dir` and lands the rebuilt shard on
@@ -9640,6 +9875,7 @@ impl DiskAPI for LocalDisk {
            FileInfoOpts {
                data: read_data,
                include_free_versions: opts.incl_free_versions,
+                include_part_checksums: false,
            },
        )?;

@@ -12204,6 +12440,10 @@ mod test {
            .join(RUSTFS_META_TMP_BUCKET)
            .join(tmp_object)
            .join(new_data_dir.to_string());
+        let tmp_parent = tmp_data_dir
+            .parent()
+            .expect("tmp data dir should have a parent")
+            .to_path_buf();
        fs::create_dir_all(&tmp_data_dir)
            .await
            .expect("new tmp data dir should be created");
@@ -12215,6 +12455,10 @@ mod test {
        disk.rename_data(RUSTFS_META_TMP_BUCKET, tmp_object, new_fi, bucket, object)
            .await
            .expect("rename_data should commit");
+        assert!(
+            !tmp_parent.exists(),
+            "successful non-inline commit should remove the empty staging parent"
+        );

        // The tmp xl.meta write point uses SyncMode::FileOnly: its parent dir
        // ({tmp}/{tmp_object}) must not be fsynced.
@@ -12405,7 +12649,7 @@ mod test {
    }

    #[tokio::test]
-    async fn test_rename_data_new_inline_object_fsyncs_new_ancestor_dirs() {
+    async fn windows_and_unix_rename_data_new_inline_object_fsyncs_new_ancestor_dirs() {
        // The inline commit path (fi.data present) has the same mkdir gap as the
        // non-inline path: a first PUT under a new prefix must fsync the newly
        // created prefix and bucket dirs.
@@ -12419,6 +12663,9 @@ mod test {
        let tmp_object = "tmp-new-inline";
        ensure_test_volume(&disk, bucket).await;
        ensure_test_volume(&disk, RUSTFS_META_TMP_BUCKET).await;
+        let tmp_parent = disk
+            .get_object_path(RUSTFS_META_TMP_BUCKET, tmp_object)
+            .expect("tmp parent should resolve");

        let _mode = durability_mode_override::set(DurabilityMode::Strict);
        let version_id = Uuid::parse_str("99999999-9999-9999-9999-999999999999").expect("version id should parse");
@@ -12427,6 +12674,7 @@ mod test {
        disk.rename_data(RUSTFS_META_TMP_BUCKET, tmp_object, new_fi, bucket, object)
            .await
            .expect("inline rename_data should commit the new object");
+        assert!(!tmp_parent.exists(), "successful inline commit should remove the empty staging parent");

        let bucket_dir = disk.get_bucket_path(bucket).expect("bucket path should resolve");
        let prefix_dir = disk.get_object_path(bucket, "prefix").expect("prefix path should resolve");
@@ -12434,10 +12682,150 @@ mod test {
            os::fsync_dir_recorder::was_fsynced(&prefix_dir),
            "the newly created prefix dir must be fsynced on an inline first PUT"
        );
+        assert_eq!(
+            os::fsync_dir_recorder::was_limited(&prefix_dir),
+            cfg!(unix),
+            "only Unix inline prefix fsyncs should use the disk file-sync limit"
+        );
        assert!(
            os::fsync_dir_recorder::was_fsynced(&bucket_dir),
            "the bucket dir must be fsynced on an inline first PUT"
        );
+        assert_eq!(
+            os::fsync_dir_recorder::was_limited(&bucket_dir),
+            cfg!(unix),
+            "only Unix inline bucket fsyncs should use the disk file-sync limit"
+        );
+    }
+
+    #[tokio::test]
+    async fn rename_data_inline_preserves_non_empty_staging_parent() {
+        use tempfile::tempdir;
+
+        let dir = tempdir().expect("temp dir should be created");
+        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
+        let disk = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+        let bucket = "inline-staging-sentinel-bucket";
+        let object = "inline-object";
+        let tmp_object = "inline-stage-with-sentinel";
+        ensure_test_volume(&disk, bucket).await;
+        ensure_test_volume(&disk, RUSTFS_META_TMP_BUCKET).await;
+
+        let tmp_parent = disk
+            .get_object_path(RUSTFS_META_TMP_BUCKET, tmp_object)
+            .expect("tmp parent should resolve");
+        fs::create_dir_all(&tmp_parent).await.expect("tmp parent should be created");
+        let sentinel = tmp_parent.join("sentinel");
+        fs::write(&sentinel, b"keep").await.expect("sentinel should be written");
+
+        let fi = test_file_info(object, Uuid::new_v4(), None, Some(Bytes::from_static(b"inline-payload")));
+        disk.rename_data(RUSTFS_META_TMP_BUCKET, tmp_object, fi, bucket, object)
+            .await
+            .expect("non-empty staging cleanup must not negate the committed object");
+
+        assert_eq!(fs::read(&sentinel).await.expect("sentinel should remain"), b"keep");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    #[allow(clippy::await_holding_lock)]
+    async fn strict_inline_rename_retains_admission_until_commit_fsync() {
+        use std::sync::atomic::{AtomicU8, Ordering};
+        use std::sync::mpsc;
+        use tempfile::tempdir;
+        use tokio::sync::oneshot;
+
+        const FIRST_BARRIER: u8 = 1;
+        const SECOND_PREPARATION: u8 = 2;
+
+        let _mode = durability_mode_override::set(DurabilityMode::Strict);
+        let dir = tempdir().expect("temp dir should be created");
+        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
+        let mut disk = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+        disk.file_sync_permits = Arc::new(Semaphore::new(1));
+        let disk = Arc::new(disk);
+        let bucket = "inline-admission-order";
+        let first_object = "first-object";
+        let second_object = "second-object";
+        ensure_test_volume(&disk, bucket).await;
+        ensure_test_volume(&disk, RUSTFS_META_TMP_BUCKET).await;
+
+        let (first_prepared_tx, first_prepared_rx) = mpsc::channel();
+        let (release_first_tx, release_first_rx) = mpsc::channel();
+        set_inline_preparation_before_backup(first_object, move || {
+            first_prepared_tx.send(()).expect("signal first preparation");
+            release_first_rx.recv().expect("wait for queued rename");
+        });
+        let first_disk = disk.clone();
+        let first = tokio::spawn(async move {
+            first_disk
+                .rename_data(
+                    RUSTFS_META_TMP_BUCKET,
+                    "first-stage",
+                    test_file_info(first_object, Uuid::new_v4(), None, Some(Bytes::from_static(b"first"))),
+                    bucket,
+                    first_object,
+                )
+                .await
+        });
+        tokio::task::spawn_blocking(move || first_prepared_rx.recv_timeout(Duration::from_secs(30)))
+            .await
+            .expect("first preparation waiter should run")
+            .expect("first rename should hold the only disk admission");
+
+        let first_event = Arc::new(AtomicU8::new(0));
+        let first_barrier_event = first_event.clone();
+        let first_object_dir = disk
+            .get_object_path_for_io(bucket, first_object)
+            .expect("first object path should resolve");
+        os::fsync_dir_recorder::set_before_limited(&first_object_dir, move || {
+            let _ = first_barrier_event.compare_exchange(0, FIRST_BARRIER, Ordering::SeqCst, Ordering::SeqCst);
+        });
+
+        let second_preparation_event = first_event.clone();
+        set_inline_preparation_before_backup(second_object, move || {
+            second_preparation_event.fetch_or(SECOND_PREPARATION, Ordering::SeqCst);
+        });
+        let (second_admission_tx, second_admission_rx) = oneshot::channel();
+        set_inline_before_file_sync_admission(second_object, move || {
+            second_admission_tx.send(()).expect("signal second admission attempt");
+        });
+        let second_disk = disk.clone();
+        let mut second = Box::pin(async move {
+            second_disk
+                .rename_data(
+                    RUSTFS_META_TMP_BUCKET,
+                    "second-stage",
+                    test_file_info(second_object, Uuid::new_v4(), None, Some(Bytes::from_static(b"second"))),
+                    bucket,
+                    second_object,
+                )
+                .await
+        });
+        let mut second_admission_rx = Box::pin(second_admission_rx);
+        tokio::time::timeout(Duration::from_secs(30), async {
+            tokio::select! {
+                _ = &mut second => panic!("second rename must wait for disk admission"),
+                signal = &mut second_admission_rx => signal.expect("second admission hook should run"),
+            }
+        })
+        .await
+        .expect("second rename should reach the admission queue");
+
+        release_first_tx.send(()).expect("release first preparation");
+        let (first_result, second_result) = tokio::time::timeout(Duration::from_secs(30), async { tokio::join!(first, second) })
+            .await
+            .expect("both inline renames should complete");
+        first_result
+            .expect("first rename task should join")
+            .expect("first inline rename should commit");
+        second_result.expect("second inline rename should commit");
+
+        assert_eq!(
+            first_event.load(Ordering::SeqCst),
+            FIRST_BARRIER | SECOND_PREPARATION,
+            "the queued rename must not overtake the admitted rename before its commit fsync"
+        );
    }

    #[cfg(windows)]
@@ -12512,7 +12900,10 @@ mod test {
            .expect("non-inline rename_data should commit");

        assert!(!replacement_dir.exists(), "the destination object directory must not be replaced");
-        assert!(staging_parent.exists(), "the guarded staging parent must retain its identity");
+        assert!(
+            !staging_parent.exists(),
+            "successful commit should remove the empty staging parent after releasing its guard"
+        );
        assert!(
            !replacement_staging_parent.exists(),
            "the staging parent must not be replaced between data and metadata publication"
@@ -13684,14 +14075,16 @@ mod test {
        );
    }

-    #[tokio::test]
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    #[allow(clippy::await_holding_lock)]
    async fn test_rename_data_writes_old_metadata_backup_for_inline_overwrite() {
+        use std::sync::mpsc;
        use tempfile::tempdir;

        let _mode = durability_mode_override::set(DurabilityMode::Strict);
        let dir = tempdir().expect("temp dir should be created");
        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
-        let disk = LocalDisk::new(&endpoint, false).await.expect("local disk should be created");
+        let disk = Arc::new(LocalDisk::new(&endpoint, false).await.expect("local disk should be created"));

        let bucket = "bucket";
        let object = "inline-object";
@@ -13717,10 +14110,32 @@ mod test {
            .await
            .expect("tmp object dir should be created");

+        let (published_tx, published_rx) = mpsc::channel();
+        let (release_tx, release_rx) = mpsc::channel();
+        set_rename_data_after_first_publication(object, move || {
+            published_tx.send(()).expect("signal backup publication");
+            release_rx.recv().expect("wait for lock-order assertion");
+        });
        let new_fi = test_file_info(object, version_id, None, Some(Bytes::from_static(b"inline-new")));
-        let resp = disk
-            .rename_data(RUSTFS_META_TMP_BUCKET, tmp_object, new_fi, bucket, object)
+        let rename_disk = disk.clone();
+        let rename = tokio::spawn(async move {
+            rename_disk
+                .rename_data(RUSTFS_META_TMP_BUCKET, tmp_object, new_fi, bucket, object)
+                .await
+        });
+        tokio::task::spawn_blocking(move || published_rx.recv_timeout(Duration::from_secs(10)))
            .await
+            .expect("publication waiter should run")
+            .expect("rollback backup must be published");
+        assert_eq!(
+            disk.file_sync_permits.available_permits(),
+            os::MAX_PARALLEL_FILE_SYNCS,
+            "backup publication must not acquire namespace while holding disk admission"
+        );
+        release_tx.send(()).expect("release backup publication");
+        let resp = rename
+            .await
+            .expect("inline rename task should join")
            .expect("inline rename_data should commit");

        assert_eq!(resp.old_data_dir, Some(old_data_dir));
@@ -13731,10 +14146,20 @@ mod test {
            os::fsync_dir_recorder::was_fsynced(backup_path.parent().expect("backup must have a parent")),
            "strict inline overwrite must persist the rollback backup directory entry"
        );
+        assert_eq!(
+            os::fsync_dir_recorder::was_limited(backup_path.parent().expect("backup must have a parent")),
+            cfg!(unix),
+            "only Unix rollback backup fsyncs should use the disk file-sync limit"
+        );
        assert!(
            os::fsync_dir_recorder::was_fsynced(&dst_object_dir),
            "strict inline overwrite must persist the committed xl.meta directory entry"
        );
+        assert_eq!(
+            os::fsync_dir_recorder::was_limited(&dst_object_dir),
+            cfg!(unix),
+            "only Unix inline commit fsyncs should use the disk file-sync limit"
+        );
        // The rollback backup must contain the previous metadata bytes verbatim so
        // that undo_write can restore the prior committed object; guards the inline
        // backup write against truncation/corruption regressions.
@@ -14331,10 +14756,12 @@ mod test {
    }

    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    #[allow(clippy::await_holding_lock)]
    async fn windows_and_unix_cancelled_inline_preparation_serializes_newer_commit() {
        use std::sync::mpsc;
        use tempfile::tempdir;

+        let _mode = durability_mode_override::set(DurabilityMode::Strict);
        let dir = tempdir().expect("temp dir should be created");
        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
        let disk = Arc::new(LocalDisk::new(&endpoint, false).await.expect("local disk should be created"));
@@ -14373,6 +14800,11 @@ mod test {
            .await
            .expect("preparation waiter should run")
            .expect("preparation must reach the backup hook");
+        assert_eq!(
+            disk.file_sync_permits.available_permits(),
+            os::MAX_PARALLEL_FILE_SYNCS - 1,
+            "strict inline preparation must hold one disk file-sync permit"
+        );
        cancelled.abort();
        assert!(cancelled.await.expect_err("operation should be cancelled").is_cancelled());

@@ -14446,6 +14878,56 @@ mod test {
        );
    }

+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    #[allow(clippy::await_holding_lock)]
+    async fn relaxed_inline_preparation_does_not_use_file_sync_limit() {
+        use std::sync::mpsc;
+        use tempfile::tempdir;
+
+        let _mode = durability_mode_override::set(DurabilityMode::Relaxed);
+        let dir = tempdir().expect("temp dir should be created");
+        let endpoint = Endpoint::try_from(dir.path().to_str().expect("temp dir should be utf8")).expect("endpoint should parse");
+        let disk = Arc::new(LocalDisk::new(&endpoint, false).await.expect("local disk should be created"));
+        let bucket = "relaxed-inline-preparation";
+        let object = "inline-object";
+        ensure_test_volume(&disk, bucket).await;
+        ensure_test_volume(&disk, RUSTFS_META_TMP_BUCKET).await;
+
+        let (entered_tx, entered_rx) = mpsc::channel();
+        let (release_tx, release_rx) = mpsc::channel();
+        set_inline_preparation_before_backup(object, move || {
+            entered_tx.send(()).expect("signal blocked preparation");
+            release_rx.recv().expect("wait for permit assertion");
+        });
+        let rename_disk = Arc::clone(&disk);
+        let rename = tokio::spawn(async move {
+            rename_disk
+                .rename_data(
+                    RUSTFS_META_TMP_BUCKET,
+                    "relaxed-inline-stage",
+                    test_file_info(object, Uuid::new_v4(), None, Some(Bytes::from_static(b"payload"))),
+                    bucket,
+                    object,
+                )
+                .await
+        });
+        tokio::task::spawn_blocking(move || entered_rx.recv_timeout(Duration::from_secs(10)))
+            .await
+            .expect("preparation waiter should run")
+            .expect("preparation must reach the hook");
+        assert_eq!(
+            disk.file_sync_permits.available_permits(),
+            os::MAX_PARALLEL_FILE_SYNCS,
+            "relaxed inline preparation must not consume strict sync capacity"
+        );
+
+        release_tx.send(()).expect("release inline preparation");
+        rename
+            .await
+            .expect("rename task should join")
+            .expect("relaxed inline rename should commit");
+    }
+
    #[tokio::test]
    async fn rename_purge_pending_payload_stays_object_and_cleans_local_backup() {
        use tempfile::tempdir;
@@ -19394,6 +19876,133 @@ mod test {
        );
    }

+    /// Same heal hazard as the io_uring test, but exercised through the default
+    /// `StdBackend` read path (rustfs/backlog#1801): a cached descriptor keeps
+    /// serving the pre-heal inode until `invalidate_cached_fds_under` drops it.
+    /// `StdBackend` reads via mmap/`try_clone`, so this proves the dup-based hit
+    /// path also defers to invalidation rather than masking a healed shard.
+    #[cfg(target_os = "linux")]
+    #[tokio::test(flavor = "multi_thread")]
+    async fn std_fd_cache_hides_a_healed_shard_until_invalidated() {
+        use tempfile::tempdir;
+
+        let root_dir = tempdir().expect("operation should succeed");
+        let root = root_dir.path().to_path_buf();
+        let backend = temp_env::with_vars([(ENV_RUSTFS_LOCAL_FD_CACHE, Some("true"))], || StdBackend::new(root.clone()));
+        if backend.fd_cache.is_none() {
+            // RLIMIT_NOFILE too low for 512 fds/disk (rustfs/backlog#1178): the
+            // cache is off, so there is nothing to exercise. Do not vacuously pass.
+            eprintln!(
+                "std_fd_cache_hides_a_healed_shard_until_invalidated: skipped \
+                 (RLIMIT_NOFILE too low for the std fd cache)"
+            );
+            return;
+        }
+
+        let volume = "bucket";
+        let object = "obj/0d1e2f/part.1";
+        let dir = root.join(volume).join("obj/0d1e2f");
+        std::fs::create_dir_all(&dir).expect("operation should succeed");
+        let part = root.join(volume).join(object);
+        std::fs::write(&part, b"corrupt-shard").expect("operation should succeed");
+
+        let before = backend
+            .pread_bytes(volume, object, 0, b"corrupt-shard".len(), None)
+            .await
+            .expect("operation should succeed");
+        assert_eq!(before, Bytes::from_static(b"corrupt-shard"));
+
+        // Heal: rename rebuilt content onto the same part path — inode swap, path
+        // unchanged. A cached descriptor would keep reading the old inode.
+        let rebuilt = dir.join("part.1.rebuilt");
+        std::fs::write(&rebuilt, b"healed--shard").expect("operation should succeed");
+        std::fs::rename(&rebuilt, &part).expect("operation should succeed");
+
+        let stale = backend
+            .pread_bytes(volume, object, 0, b"healed--shard".len(), None)
+            .await
+            .expect("operation should succeed");
+        assert_eq!(
+            stale,
+            Bytes::from_static(b"corrupt-shard"),
+            "a cached descriptor is expected to still see the pre-heal inode — this is the \
+             hazard invalidate_cached_fds exists to close, and the assertion proves the cache is live"
+        );
+
+        backend.invalidate_cached_fds_under(volume, "obj/0d1e2f");
+        let healed = backend
+            .pread_bytes(volume, object, 0, b"healed--shard".len(), None)
+            .await
+            .expect("operation should succeed");
+        assert_eq!(healed, Bytes::from_static(b"healed--shard"));
+    }
+
+    /// A repeated read of the same shard must (a) return correct bytes both times
+    /// and (b) actually populate the descriptor cache, so the second read can skip
+    /// `File::open` (rustfs/backlog#1801).
+    #[cfg(target_os = "linux")]
+    #[tokio::test(flavor = "multi_thread")]
+    async fn std_fd_cache_serves_repeated_reads_and_caches_descriptor() {
+        use tempfile::tempdir;
+
+        let root_dir = tempdir().expect("operation should succeed");
+        let root = root_dir.path().to_path_buf();
+        let backend = temp_env::with_vars([(ENV_RUSTFS_LOCAL_FD_CACHE, Some("true"))], || StdBackend::new(root.clone()));
+        let cache = match backend.fd_cache.as_ref() {
+            Some(c) => c,
+            None => {
+                eprintln!(
+                    "std_fd_cache_serves_repeated_reads_and_caches_descriptor: skipped \
+                     (RLIMIT_NOFILE too low for the std fd cache)"
+                );
+                return;
+            }
+        };
+
+        let volume = "bucket";
+        let object = "obj/abc/part.1";
+        std::fs::create_dir_all(root.join(volume).join("obj/abc")).expect("operation should succeed");
+        let payload = b"hello-small-shard-payload";
+        std::fs::write(root.join(volume).join(object), payload).expect("operation should succeed");
+
+        let first = backend
+            .pread_bytes(volume, object, 0, payload.len(), None)
+            .await
+            .expect("operation should succeed");
+        assert_eq!(first, Bytes::from_static(payload));
+
+        // After the first miss the freshly opened descriptor is indexed; a second
+        // read of the same path is a cache hit.
+        assert_eq!(cache.entry_count().await, 1, "the first read should have cached exactly one descriptor");
+
+        let second = backend
+            .pread_bytes(volume, object, 0, payload.len(), None)
+            .await
+            .expect("operation should succeed");
+        assert_eq!(second, Bytes::from_static(payload));
+
+        // Invalidating by the object prefix drops the cached descriptor.
+        backend.invalidate_cached_fds_under(volume, "obj/abc");
+        assert_eq!(cache.entry_count().await, 0, "prefix invalidation must drop the cached descriptor");
+    }
+
+    /// `StdBackend::new_without_fd_cache` must not build a descriptor cache.
+    /// `UringBackend` wraps a `StdBackend` and owns the only cache for the disk,
+    /// so an inner cache would be populated by fallback reads
+    /// (`UringBackend::pread_bytes` delegates inward) yet never invalidated —
+    /// the stale-inode hazard `FdCache` exists to close (backlog#1176/#1801).
+    /// This pins the contract so a future constructor change cannot regress it.
+    #[cfg(target_os = "linux")]
+    #[test]
+    fn new_without_fd_cache_builds_no_descriptor_cache() {
+        let root_dir = tempfile::tempdir().expect("operation should succeed");
+        let backend = StdBackend::new_without_fd_cache(root_dir.path().to_path_buf());
+        assert!(
+            backend.fd_cache.is_none(),
+            "new_without_fd_cache must not build a descriptor cache — UringBackend owns the only cache for the disk"
+        );
+    }
+
    /// The mutation paths on `LocalDisk` must actually call
    /// `invalidate_cached_fds`, not merely have it available (backlog#1145).
    /// `rename_file` replaces the inode at a path a reader has already cached;
@@ -65,6 +65,7 @@ use error::{Error, Result};
 use local::LocalDisk;
 use rustfs_filemeta::{FileInfo, ObjectPartInfo, RawFileInfo};
 use rustfs_madmin::info_commands::DiskMetrics;
+use rustfs_rio::ChunkReaderBox;
 use serde::{Deserialize, Serialize};
 use std::{fmt::Debug, path::PathBuf, sync::Arc, time::Duration};
 use time::OffsetDateTime;
@@ -427,6 +428,19 @@ impl DiskAPI for Disk {
        }
    }

+    async fn read_file_stream_chunks(
+        &self,
+        volume: &str,
+        path: &str,
+        offset: usize,
+        length: usize,
+    ) -> Result<Option<ChunkReaderBox>> {
+        match self {
+            Disk::Local(_) => Ok(None),
+            Disk::Remote(remote_disk) => remote_disk.read_file_stream_chunks(volume, path, offset, length).await,
+        }
+    }
+
    #[tracing::instrument(level = "trace", skip_all)]
    async fn read_file_mmap_copy(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<Bytes> {
        match self {
@@ -865,6 +879,18 @@ pub trait DiskAPI: Debug + Send + Sync + 'static {
    async fn read_file(&self, volume: &str, path: &str) -> Result<FileReader>;
    async fn read_file_stream(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<FileReader>;

+    /// Returns an owned-chunk stream when the backing transport can preserve
+    /// receive-buffer ownership. `None` retains the ordinary reader path.
+    async fn read_file_stream_chunks(
+        &self,
+        _volume: &str,
+        _path: &str,
+        _offset: usize,
+        _length: usize,
+    ) -> Result<Option<ChunkReaderBox>> {
+        Ok(None)
+    }
+
    /// File read using mmap-then-copy on Unix or an efficient read on non-Unix.
    async fn read_file_mmap_copy(&self, volume: &str, path: &str, offset: usize, length: usize) -> Result<Bytes>;

@@ -1251,7 +1277,7 @@ pub struct VolumeInfo {
    pub created: Option<OffsetDateTime>,
 }

-#[derive(Deserialize, Serialize, Debug, Default, Clone)]
+#[derive(Deserialize, Serialize, Debug, Default, Clone, Copy)]
 pub struct ReadOptions {
    pub incl_free_versions: bool,
    pub read_data: bool,
@@ -78,28 +78,59 @@ pub fn check_path_length(path_name: &str) -> Result<()> {
 /// their own unique tempdir to stay robust against parallel test execution.
 #[cfg(test)]
 pub(crate) mod fsync_dir_recorder {
+    use std::collections::HashMap;
    use std::path::{Path, PathBuf};
    use std::sync::Mutex;

-    static RECORDED: Mutex<Vec<PathBuf>> = Mutex::new(Vec::new());
+    type Hook = Box<dyn FnOnce() + Send>;

-    pub(crate) fn record(dir: &Path) {
-        let mut recorded = RECORDED.lock().expect("fsync dir recorder poisoned");
-        recorded.push(dir.to_path_buf());
-        if let Ok(canonical) = dir.canonicalize()
-            && canonical != dir
+    static RECORDED: Mutex<Vec<PathBuf>> = Mutex::new(Vec::new());
+    static LIMITED: Mutex<Vec<PathBuf>> = Mutex::new(Vec::new());
+    static BEFORE_LIMITED: std::sync::LazyLock<Mutex<HashMap<PathBuf, Hook>>> =
+        std::sync::LazyLock::new(|| Mutex::new(HashMap::new()));
+
+    fn record_path(paths: &Mutex<Vec<PathBuf>>, path: &Path, description: &str) {
+        let mut paths = paths.lock().expect(description);
+        paths.push(path.to_path_buf());
+        if let Ok(canonical) = path.canonicalize()
+            && canonical != path
        {
-            recorded.push(canonical);
+            paths.push(canonical);
        }
    }

-    pub(crate) fn was_fsynced(dir: &Path) -> bool {
-        let canonical = dir.canonicalize().ok();
-        RECORDED
-            .lock()
-            .expect("fsync dir recorder poisoned")
+    fn contains_path(paths: &[PathBuf], path: &Path) -> bool {
+        let canonical = path.canonicalize().ok();
+        paths
            .iter()
-            .any(|p| p == dir || canonical.as_ref().is_some_and(|canonical| p == canonical))
+            .any(|recorded| recorded == path || canonical.as_ref().is_some_and(|canonical| recorded == canonical))
+    }
+
+    pub(crate) fn record(dir: &Path) {
+        record_path(&RECORDED, dir, "fsync dir recorder");
+    }
+
+    pub(crate) fn was_fsynced(dir: &Path) -> bool {
+        contains_path(&RECORDED.lock().expect("fsync dir recorder poisoned"), dir)
+    }
+
+    pub(crate) fn record_limited(dir: &Path) {
+        record_path(&LIMITED, dir, "limited fsync dir recorder");
+        let hook = BEFORE_LIMITED.lock().expect("limited fsync hook poisoned").remove(dir);
+        if let Some(hook) = hook {
+            hook();
+        }
+    }
+
+    pub(crate) fn was_limited(dir: &Path) -> bool {
+        contains_path(&LIMITED.lock().expect("limited fsync dir recorder poisoned"), dir)
+    }
+
+    pub(crate) fn set_before_limited(dir: &Path, hook: impl FnOnce() + Send + 'static) {
+        BEFORE_LIMITED
+            .lock()
+            .expect("limited fsync hook poisoned")
+            .insert(dir.to_path_buf(), Box::new(hook));
    }
 }

@@ -1120,6 +1151,79 @@ pub(crate) async fn run_blocking_namespace_operation<T: Send + 'static>(
    .map_err(|err| io::Error::other(format!("blocking namespace operation failed: {err}")))?
 }

+/// Admit one strict inline commit under the disk sync limit. The caller already
+/// owns the namespace lease, establishing namespace -> disk ordering. Holding
+/// admission across adjacent durability barriers prevents one transaction from
+/// repeatedly joining the disk semaphore tail.
+pub(crate) struct FileSyncAdmission {
+    disk_permit: Arc<OwnedSemaphorePermit>,
+}
+
+pub(crate) async fn acquire_file_sync_admission(disk_permits: Arc<Semaphore>) -> io::Result<FileSyncAdmission> {
+    let disk_permit = disk_permits
+        .acquire_owned()
+        .await
+        .map_err(|_| io::Error::other("disk file sync concurrency limiter closed"))?;
+    Ok(FileSyncAdmission {
+        disk_permit: Arc::new(disk_permit),
+    })
+}
+
+/// Keep the disk admission and namespace lease with the blocking syscall if
+/// the async waiter is cancelled. The process-wide admission remains with the
+/// waiter so cancellation cannot starve healthy disks.
+pub(crate) async fn run_blocking_namespace_file_sync_operation<T: Send + 'static>(
+    lease: Arc<NamespaceMutationLease>,
+    admission: &FileSyncAdmission,
+    operation: impl FnOnce() -> io::Result<T> + Send + 'static,
+) -> io::Result<T> {
+    run_blocking_namespace_file_sync_operation_with_global(lease, admission, &FILE_SYNC_PERMITS, operation).await
+}
+
+async fn run_blocking_namespace_file_sync_operation_with_global<T: Send + 'static>(
+    lease: Arc<NamespaceMutationLease>,
+    admission: &FileSyncAdmission,
+    global_permits: &Semaphore,
+    operation: impl FnOnce() -> io::Result<T> + Send + 'static,
+) -> io::Result<T> {
+    let global_permit = global_permits
+        .acquire()
+        .await
+        .map_err(|_| io::Error::other("global file sync concurrency limiter closed"))?;
+    let disk_permit = admission.disk_permit.clone();
+    let result = tokio::task::spawn_blocking(move || {
+        let _lease = lease;
+        let _disk_permit = disk_permit;
+        operation()
+    })
+    .await;
+    drop(global_permit);
+    result.map_err(|err| io::Error::other(format!("blocking namespace file sync operation failed: {err}")))?
+}
+
+pub(crate) async fn fsync_dir_with_namespace_file_sync_limit(
+    dir: impl AsRef<Path>,
+    lease: Arc<NamespaceMutationLease>,
+    admission: &FileSyncAdmission,
+) -> io::Result<()> {
+    #[cfg(unix)]
+    {
+        let dir = dir.as_ref().to_path_buf();
+        run_blocking_namespace_file_sync_operation(lease, admission, move || {
+            #[cfg(test)]
+            fsync_dir_recorder::record_limited(&dir);
+            fsync_dir_std(dir)
+        })
+        .await
+    }
+
+    #[cfg(not(unix))]
+    {
+        let _ = (lease, admission);
+        fsync_dir_std(dir)
+    }
+}
+
 struct RenamePreparation {
    parent_guard: Option<ExistingBaseDirectoryGuard>,
    #[cfg(windows)]
@@ -2784,6 +2888,7 @@ pub fn is_dir_not_empty_error(err: &io::Error) -> bool {
 mod tests {
    use super::*;
    use std::sync::Mutex;
+    use std::time::Duration;
    use tempfile::tempdir;
    use tracing_subscriber::fmt::MakeWriter;

@@ -4553,6 +4658,79 @@ mod tests {
        fsync_dir(temp_dir.path()).await.expect("fsync dir must succeed");
    }

+    #[tokio::test]
+    async fn file_sync_admission_is_reused_across_commit_barriers() {
+        let temp_dir = tempdir().expect("create temp dir");
+        let limiter = Arc::new(Semaphore::new(1));
+        let lease = acquire_namespace_mutation_lease(temp_dir.path()).await;
+        let admission = acquire_file_sync_admission(limiter.clone())
+            .await
+            .expect("first commit should acquire admission");
+
+        run_blocking_namespace_file_sync_operation(lease.clone(), &admission, || Ok(()))
+            .await
+            .expect("first barrier should complete under the admission");
+        let mut waiting = Box::pin(acquire_file_sync_admission(limiter));
+        assert!(
+            futures::poll!(&mut waiting).is_pending(),
+            "another commit must remain queued between durability barriers"
+        );
+        run_blocking_namespace_file_sync_operation(lease, &admission, || Ok(()))
+            .await
+            .expect("later barrier should reuse admission without requeuing");
+
+        drop(admission);
+        tokio::time::timeout(Duration::from_secs(30), waiting)
+            .await
+            .expect("queued commit should acquire admission after release")
+            .expect("queued commit should acquire admission");
+    }
+
+    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
+    async fn cancelled_file_sync_waiter_keeps_disk_admission_until_blocking_work_finishes() {
+        use std::sync::mpsc;
+
+        let temp_dir = tempdir().expect("create temp dir");
+        let limiter = Arc::new(Semaphore::new(1));
+        let global_permits = Arc::new(Semaphore::new(1));
+        let lease = acquire_namespace_mutation_lease(temp_dir.path()).await;
+        let admission = acquire_file_sync_admission(limiter.clone())
+            .await
+            .expect("file sync admission should be acquired");
+        let (entered_tx, entered_rx) = mpsc::channel();
+        let (release_tx, release_rx) = mpsc::channel();
+        let waiter_global_permits = global_permits.clone();
+        let waiter = tokio::spawn(async move {
+            run_blocking_namespace_file_sync_operation_with_global(lease, &admission, waiter_global_permits.as_ref(), move || {
+                entered_tx.send(()).expect("signal blocking work");
+                release_rx.recv().expect("wait for blocking work release");
+                Ok(())
+            })
+            .await
+        });
+
+        tokio::task::spawn_blocking(move || entered_rx.recv_timeout(Duration::from_secs(30)))
+            .await
+            .expect("blocking work waiter should run")
+            .expect("blocking work should start");
+        waiter.abort();
+        assert!(waiter.await.expect_err("waiter should be cancelled").is_cancelled());
+        let returned_global_permit = global_permits
+            .try_acquire()
+            .expect("cancelled waiter must return global capacity for healthy disks");
+        assert!(
+            limiter.clone().try_acquire_owned().is_err(),
+            "cancelled waiter must not return disk capacity while blocking work is active"
+        );
+
+        release_tx.send(()).expect("release blocking work");
+        let _returned_permit = tokio::time::timeout(Duration::from_secs(30), limiter.acquire_owned())
+            .await
+            .expect("disk capacity should return after blocking work finishes")
+            .expect("disk limiter should remain open");
+        drop(returned_global_permit);
+    }
+
    #[tokio::test]
    #[serial_test::serial(file_sync_probe)]
    async fn sync_dir_files_syncs_regular_files_and_dir() {
@@ -76,6 +76,13 @@ impl ShardBufferPool {
        self.buffers[index] = Some(buf);
    }

+    #[cfg(test)]
+    pub(crate) fn stored_allocation(&self, index: usize) -> Option<(*const u8, usize)> {
+        self.buffers
+            .get(index)
+            .and_then(|buf| buf.as_ref().map(|buf| (buf.as_ptr(), buf.capacity())))
+    }
+
    #[cfg(test)]
    fn stored_capacity(&self, index: usize) -> Option<usize> {
        self.buffers.get(index).and_then(|buf| buf.as_ref().map(Vec::capacity))
@@ -14,11 +14,30 @@

 use pin_project_lite::pin_project;
 use rustfs_utils::HashAlgorithm;
+use std::future::poll_fn;
 use std::io::IoSlice;
+use std::pin::Pin;
+use std::task::{Context, Poll};
 use std::time::Duration;
 use tokio::io::{AsyncRead, AsyncReadExt, AsyncWrite, AsyncWriteExt};
 use tracing::error;
-use uuid::Uuid;
+
+const LOG_COMPONENT_ECSTORE: &str = "ecstore";
+const LOG_SUBSYSTEM_ERASURE: &str = "erasure";
+const EVENT_BITROT_SHORT_SHARD_READ: &str = "bitrot_short_shard_read";
+const EVENT_BITROT_HASH_MISMATCH: &str = "bitrot_hash_mismatch";
+const MAX_RETAINED_CHUNKS_PER_BLOCK: usize = 64;
+const MAX_CHUNK_POLLS_PER_YIELD: usize = MAX_RETAINED_CHUNKS_PER_BLOCK + 1;
+
+/// Result of polling an optional owned-chunk handoff.
+pub enum ShardChunkRead {
+    /// The source does not support owned-chunk handoff and remains untouched.
+    Unsupported,
+    /// The source reached EOF.
+    Eof,
+    /// A non-empty chunk containing at most the requested number of bytes.
+    Chunk(bytes::Bytes),
+}

 /// A shard source that may already hold its bytes in memory.
 ///
@@ -38,6 +57,12 @@ pub trait ShardSource: AsyncRead + Send + Sync + Unpin {
    fn try_take_block(&mut self, _n: usize) -> Option<bytes::Bytes> {
        None
    }
+
+    /// Polls one owned chunk when the source supports chunk handoff.
+    /// `Unsupported` must leave the source untouched.
+    fn poll_read_chunk(self: Pin<&mut Self>, _cx: &mut Context<'_>, _max: usize) -> Poll<std::io::Result<ShardChunkRead>> {
+        Poll::Ready(Ok(ShardChunkRead::Unsupported))
+    }
 }

 /// Borrowed and owned byte slices are ordinary streaming sources: they carry no
@@ -71,9 +96,11 @@ pin_project! {
        // contiguous on-disk `[hash][data]` block so both are pulled in a single
        // pass; grown lazily and never shrunk.
        buf: Vec<u8>,
+        // Reused owned chunk vector for the remote HTTP fast path. Keeping the
+        // allocation with the reader avoids allocating once per bitrot block.
+        chunks: Vec<bytes::Bytes>,
        skip_verify: bool,
        last_verify_duration: Duration,
-        id: Uuid,
    }
 }

@@ -88,9 +115,9 @@ where
            hash_algo: algo,
            shard_size,
            buf: Vec::new(),
+            chunks: Vec::new(),
            skip_verify,
            last_verify_duration: Duration::ZERO,
-            id: Uuid::new_v4(),
        }
    }

@@ -98,6 +125,11 @@ where
        self.last_verify_duration
    }

+    #[cfg(test)]
+    pub(crate) fn inner_ref(&self) -> &R {
+        &self.inner
+    }
+
    /// Read a single (hash+data) block, verify hash, and copy `out.len()` bytes
    /// into `out`. Returns an error if the shard is short, the hash mismatches,
    /// or `out` is larger than one shard. On error `out`'s contents are
@@ -118,7 +150,7 @@ where

        let need = self.hash_algo.size() + want;
        self.read_scratch_block(need, want).await?;
-        let (data, verify) = split_and_verify(&self.hash_algo, self.skip_verify, &self.buf[..need], &self.id)?;
+        let (data, verify) = split_and_verify(&self.hash_algo, self.skip_verify, &self.buf[..need])?;
        out.copy_from_slice(data);
        self.last_verify_duration = verify;
        Ok(want)
@@ -157,7 +189,7 @@ where
        }
        let filled = fill(&mut self.inner, &mut self.buf[..need]).await?;
        if filled < need {
-            return Err(short_shard_read(&self.id, filled.saturating_sub(self.hash_algo.size()), want));
+            return Err(short_shard_read(filled.saturating_sub(self.hash_algo.size()), want));
        }
        Ok(())
    }
@@ -166,15 +198,23 @@ where
    /// buffer returns its length, a short read is UnexpectedEof (backlog#799 B2).
    fn finish_len(&self, data_len: usize, want: usize) -> std::io::Result<usize> {
        if data_len < want {
-            return Err(short_shard_read(&self.id, data_len, want));
+            return Err(short_shard_read(data_len, want));
        }
        Ok(data_len)
    }
 }

 /// A truncated shard is `UnexpectedEof`, not a short success (backlog#799 B2).
-fn short_shard_read(id: &Uuid, got: usize, want: usize) -> std::io::Error {
-    error!("bitrot reader short shard read: id={id} got {got} of {want} bytes");
+fn short_shard_read(got: usize, want: usize) -> std::io::Error {
+    error!(
+        event = EVENT_BITROT_SHORT_SHARD_READ,
+        component = LOG_COMPONENT_ECSTORE,
+        subsystem = LOG_SUBSYSTEM_ERASURE,
+        state = "failed",
+        got,
+        want,
+        "short shard read: got {got} of {want} bytes"
+    );
    std::io::Error::new(std::io::ErrorKind::UnexpectedEof, format!("short shard read: got {got} of {want} bytes"))
 }

@@ -184,12 +224,7 @@ fn short_shard_read(id: &Uuid, got: usize, want: usize) -> std::io::Error {
 /// hash never reaches the caller's buffer. The verify duration is returned
 /// rather than stored so this stays a free function usable while `self` is
 /// borrowed for the block.
-fn split_and_verify<'a>(
-    hash_algo: &HashAlgorithm,
-    skip_verify: bool,
-    block: &'a [u8],
-    id: &Uuid,
-) -> std::io::Result<(&'a [u8], Duration)> {
+fn split_and_verify<'a>(hash_algo: &HashAlgorithm, skip_verify: bool, block: &'a [u8]) -> std::io::Result<(&'a [u8], Duration)> {
    let (hash, data) = block.split_at(hash_algo.size());
    if skip_verify {
        return Ok((data, Duration::ZERO));
@@ -198,7 +233,14 @@ fn split_and_verify<'a>(
    let actual_hash = hash_algo.hash_encode(data);
    let verify = verify_start.elapsed();
    if actual_hash.as_ref() != hash {
-        error!("bitrot reader hash mismatch, id={id} data_len={}", data.len());
+        error!(
+            event = EVENT_BITROT_HASH_MISMATCH,
+            component = LOG_COMPONENT_ECSTORE,
+            subsystem = LOG_SUBSYSTEM_ERASURE,
+            state = "failed",
+            data_len = data.len(),
+            "bitrot hash mismatch"
+        );
        return Err(std::io::Error::new(std::io::ErrorKind::InvalidData, "bitrot hash mismatch"));
    }
    Ok((data, verify))
@@ -248,23 +290,138 @@ where

        let need = hash_size + want;

-        // In-memory fast path: the block is already resident, so slice it instead
-        // of copying it into the scratch buffer first (rustfs/backlog#1159). One
-        // copy (`extend_from_slice`) instead of two. A source that cannot serve
-        // `need` bytes returns `None` and falls through to the scratch path,
-        // keeping the short-read contract.
        if let Some(block) = self.inner.try_take_block(need) {
-            let (data, verify) = split_and_verify(&self.hash_algo, self.skip_verify, &block, &self.id)?;
+            let (data, verify) = split_and_verify(&self.hash_algo, self.skip_verify, &block)?;
            out.extend_from_slice(data);
            self.last_verify_duration = verify;
            return Ok(want);
        }

+        self.chunks.clear();
+        let handed_off = {
+            let inner = &mut self.inner;
+            let chunks = &mut self.chunks;
+            let tail_buf = &mut self.buf;
+            let mut received = 0usize;
+            poll_fn(|cx| {
+                for _ in 0..MAX_CHUNK_POLLS_PER_YIELD {
+                    let next = match Pin::new(&mut *inner).poll_read_chunk(cx, need - received) {
+                        Poll::Ready(Ok(next)) => next,
+                        Poll::Ready(Err(err)) => return Poll::Ready(Err(err)),
+                        Poll::Pending => return Poll::Pending,
+                    };
+                    let chunk = match next {
+                        ShardChunkRead::Unsupported if received == 0 => return Poll::Ready(Ok(false)),
+                        ShardChunkRead::Unsupported => {
+                            return Poll::Ready(Err(std::io::Error::new(
+                                std::io::ErrorKind::InvalidData,
+                                "chunk handoff became unavailable after transferring data",
+                            )));
+                        }
+                        ShardChunkRead::Eof => {
+                            return Poll::Ready(Err(short_shard_read(received.saturating_sub(hash_size), want)));
+                        }
+                        ShardChunkRead::Chunk(chunk) => chunk,
+                    };
+
+                    if received == 0 {
+                        tail_buf.clear();
+                    }
+                    if chunk.is_empty() {
+                        return Poll::Ready(Err(std::io::Error::new(
+                            std::io::ErrorKind::InvalidData,
+                            "chunk handoff returned an empty chunk",
+                        )));
+                    }
+                    let remaining = need - received;
+                    if chunk.len() > remaining {
+                        return Poll::Ready(Err(std::io::Error::new(
+                            std::io::ErrorKind::InvalidData,
+                            "chunk handoff exceeded its requested boundary",
+                        )));
+                    }
+                    received += chunk.len();
+
+                    if chunks.len() == MAX_RETAINED_CHUNKS_PER_BLOCK {
+                        if tail_buf.is_empty() {
+                            tail_buf.reserve_exact(need - (received - chunk.len()));
+                        }
+                        tail_buf.extend_from_slice(&chunk);
+                    } else {
+                        chunks.push(chunk);
+                    }
+
+                    if received == need {
+                        return Poll::Ready(Ok(true));
+                    }
+                }
+                cx.waker().wake_by_ref();
+                Poll::Pending
+            })
+            .await?
+        };
+        if handed_off {
+            if self.chunks.len() == 1 && self.buf.is_empty() {
+                let block = &self.chunks[0];
+                let (data, verify) = split_and_verify(&self.hash_algo, self.skip_verify, block)?;
+                out.extend_from_slice(data);
+                self.last_verify_duration = verify;
+                return Ok(want);
+            }
+
+            let block_chunks = || {
+                self.chunks
+                    .iter()
+                    .map(|chunk| chunk.as_ref())
+                    .chain((!self.buf.is_empty()).then_some(self.buf.as_slice()))
+            };
+            if !self.skip_verify {
+                let verify_start = std::time::Instant::now();
+                let actual_hash = self
+                    .hash_algo
+                    .hash_encode_slices(block_chunks().scan(hash_size, |skip, chunk| {
+                        let start = (*skip).min(chunk.len());
+                        *skip -= start;
+                        Some(&chunk[start..])
+                    }));
+                let verify = verify_start.elapsed();
+                let mut hash_offset = 0;
+                let mut remaining = hash_size;
+                for chunk in block_chunks() {
+                    let take = remaining.min(chunk.len());
+                    if actual_hash.as_ref()[hash_offset..hash_offset + take] != chunk[..take] {
+                        error!(
+                            event = EVENT_BITROT_HASH_MISMATCH,
+                            component = LOG_COMPONENT_ECSTORE,
+                            subsystem = LOG_SUBSYSTEM_ERASURE,
+                            state = "failed",
+                            data_len = want,
+                            "bitrot hash mismatch"
+                        );
+                        return Err(std::io::Error::new(std::io::ErrorKind::InvalidData, "bitrot hash mismatch"));
+                    }
+                    hash_offset += take;
+                    remaining -= take;
+                    if remaining == 0 {
+                        break;
+                    }
+                }
+                self.last_verify_duration = verify;
+            }
+            let mut skip = hash_size;
+            for chunk in block_chunks() {
+                let start = skip.min(chunk.len());
+                skip -= start;
+                out.extend_from_slice(&chunk[start..]);
+            }
+            return Ok(want);
+        }
+
        // Streaming path: same single pass and same verification as `read`; only
        // the sink differs (`extend_from_slice` into `out` instead of
        // `copy_from_slice` into a pre-zeroed buffer).
        self.read_scratch_block(need, want).await?;
-        let (data, verify) = split_and_verify(&self.hash_algo, self.skip_verify, &self.buf[..need], &self.id)?;
+        let (data, verify) = split_and_verify(&self.hash_algo, self.skip_verify, &self.buf[..need])?;
        out.extend_from_slice(data);
        self.last_verify_duration = verify;
        Ok(want)
@@ -665,18 +822,167 @@ impl BitrotWriterWrapper {

 #[cfg(test)]
 mod tests {
-    use super::ShardSource;
    use super::{
        BitrotReader, BitrotWriter, BitrotWriterWrapper, CustomWriter, bitrot_shard_file_size, bitrot_verify, write_all_vectored,
    };
+    use super::{MAX_RETAINED_CHUNKS_PER_BLOCK, ShardChunkRead, ShardSource};
+    use bytes::Bytes;
    use rustfs_utils::HashAlgorithm;
-    use std::io::{Cursor, IoSlice};
+    use std::collections::VecDeque;
+    use std::io::{self, Cursor, IoSlice};
+    use std::pin::Pin;
    use std::sync::{
        Arc,
        atomic::{AtomicUsize, Ordering},
    };
    use std::task::{Context, Poll};
-    use tokio::io::{AsyncWrite, AsyncWriteExt};
+    use std::time::Duration;
+    use tokio::io::{AsyncRead, AsyncWrite, AsyncWriteExt, ReadBuf};
+
+    struct FragmentedSource {
+        chunks: VecDeque<Bytes>,
+    }
+
+    impl FragmentedSource {
+        fn new(bytes: Vec<u8>, fragment_sizes: &[usize]) -> Self {
+            let mut chunks = VecDeque::new();
+            let mut offset = 0;
+            for &size in fragment_sizes {
+                let end = (offset + size).min(bytes.len());
+                if offset < end {
+                    chunks.push_back(Bytes::copy_from_slice(&bytes[offset..end]));
+                }
+                offset = end;
+            }
+            if offset < bytes.len() {
+                chunks.push_back(Bytes::copy_from_slice(&bytes[offset..]));
+            }
+            Self { chunks }
+        }
+    }
+
+    impl AsyncRead for FragmentedSource {
+        fn poll_read(self: Pin<&mut Self>, _cx: &mut Context<'_>, _buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+            Poll::Ready(Err(io::Error::other("fragmented source must use chunk handoff")))
+        }
+    }
+
+    impl ShardSource for FragmentedSource {
+        fn poll_read_chunk(mut self: Pin<&mut Self>, _cx: &mut Context<'_>, max: usize) -> Poll<io::Result<ShardChunkRead>> {
+            let Some(mut chunk) = self.chunks.pop_front() else {
+                return Poll::Ready(Ok(ShardChunkRead::Eof));
+            };
+            if chunk.len() > max {
+                self.chunks.push_front(chunk.split_off(max));
+                chunk.truncate(max);
+            }
+            Poll::Ready(Ok(ShardChunkRead::Chunk(chunk)))
+        }
+    }
+
+    struct GeneratedChunkSource {
+        bytes: Bytes,
+        offset: usize,
+        fragment_size: usize,
+        fail_at: Option<usize>,
+    }
+
+    impl GeneratedChunkSource {
+        fn new(bytes: Vec<u8>, fragment_size: usize) -> Self {
+            assert!(fragment_size > 0);
+            Self {
+                bytes: Bytes::from(bytes),
+                offset: 0,
+                fragment_size,
+                fail_at: None,
+            }
+        }
+
+        fn failing(bytes: Vec<u8>, fragment_size: usize, fail_at: usize) -> Self {
+            Self {
+                fail_at: Some(fail_at),
+                ..Self::new(bytes, fragment_size)
+            }
+        }
+    }
+
+    impl AsyncRead for GeneratedChunkSource {
+        fn poll_read(self: Pin<&mut Self>, _cx: &mut Context<'_>, _buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+            Poll::Ready(Err(io::Error::other("generated source must use chunk handoff")))
+        }
+    }
+
+    impl ShardSource for GeneratedChunkSource {
+        fn poll_read_chunk(mut self: Pin<&mut Self>, _cx: &mut Context<'_>, max: usize) -> Poll<io::Result<ShardChunkRead>> {
+            if self.fail_at == Some(self.offset) {
+                return Poll::Ready(Err(rustfs_rio::new_test_internode_http_io_error(
+                    rustfs_rio::InternodeHttpErrorKind::BodyStreamAborted,
+                )));
+            }
+            if self.offset == self.bytes.len() {
+                return Poll::Ready(Ok(ShardChunkRead::Eof));
+            }
+            let error_limit = self.fail_at.unwrap_or(self.bytes.len());
+            let take = self
+                .fragment_size
+                .min(max)
+                .min(error_limit - self.offset)
+                .min(self.bytes.len() - self.offset);
+            let start = self.offset;
+            self.offset += take;
+            Poll::Ready(Ok(ShardChunkRead::Chunk(self.bytes.slice(start..start + take))))
+        }
+    }
+
+    struct InvalidChunkSource {
+        mode: InvalidChunkMode,
+    }
+
+    #[derive(Clone, Copy)]
+    enum InvalidChunkMode {
+        Empty,
+        Oversized,
+        UnsupportedAfterChunk,
+        Unsupported,
+    }
+
+    impl AsyncRead for InvalidChunkSource {
+        fn poll_read(self: Pin<&mut Self>, _cx: &mut Context<'_>, _buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+            Poll::Ready(Err(io::Error::other("invalid source must use chunk handoff")))
+        }
+    }
+
+    impl ShardSource for InvalidChunkSource {
+        fn poll_read_chunk(mut self: Pin<&mut Self>, _cx: &mut Context<'_>, max: usize) -> Poll<io::Result<ShardChunkRead>> {
+            match self.mode {
+                InvalidChunkMode::Empty => Poll::Ready(Ok(ShardChunkRead::Chunk(Bytes::new()))),
+                InvalidChunkMode::Oversized => Poll::Ready(Ok(ShardChunkRead::Chunk(Bytes::from(vec![0; max + 1])))),
+                InvalidChunkMode::UnsupportedAfterChunk => {
+                    self.mode = InvalidChunkMode::Unsupported;
+                    Poll::Ready(Ok(ShardChunkRead::Chunk(Bytes::from_static(b"x"))))
+                }
+                InvalidChunkMode::Unsupported => Poll::Ready(Ok(ShardChunkRead::Unsupported)),
+            }
+        }
+    }
+
+    struct ScratchReuseSource {
+        block: Option<Bytes>,
+        saw_reused_scratch: bool,
+    }
+
+    impl AsyncRead for ScratchReuseSource {
+        fn poll_read(mut self: Pin<&mut Self>, _cx: &mut Context<'_>, buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+            let Some(block) = self.block.take() else {
+                return Poll::Ready(Ok(()));
+            };
+            self.saw_reused_scratch = buf.initialize_unfilled()[..block.len()].iter().all(|byte| *byte == 0xa5);
+            buf.put_slice(&block);
+            Poll::Ready(Ok(()))
+        }
+    }
+
+    impl ShardSource for ScratchReuseSource {}

    #[derive(Default)]
    struct VectoredCountingWriter {
@@ -1434,6 +1740,70 @@ mod tests {
        assert!(out.is_empty(), "corrupt bytes must never reach the caller's buffer");
    }

+    #[tokio::test]
+    async fn chunked_handoff_verifies_data_split_across_hash_boundaries() {
+        const SHARD: usize = 4096;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data: Vec<u8> = (0..SHARD).map(|index| (index % 251) as u8).collect();
+        let mut encoded = Vec::new();
+        BitrotWriter::new(&mut encoded, SHARD, algo.clone())
+            .write(&data)
+            .await
+            .expect("write shard");
+
+        let mut output = Vec::with_capacity(SHARD);
+        BitrotReader::new(FragmentedSource::new(encoded, &[3, 11, 19, 37, 128]), SHARD, algo, false)
+            .read_appending(&mut output, SHARD)
+            .await
+            .expect("fragmented shard must verify");
+
+        assert_eq!(output, data);
+    }
+
+    #[tokio::test]
+    async fn chunked_handoff_never_appends_a_corrupt_shard() {
+        const SHARD: usize = 4096;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let mut encoded = Vec::new();
+        BitrotWriter::new(&mut encoded, SHARD, algo.clone())
+            .write(&vec![9u8; SHARD])
+            .await
+            .expect("write shard");
+        let last = encoded.len() - 1;
+        encoded[last] ^= 0xff;
+
+        let mut output = Vec::with_capacity(SHARD);
+        let err = BitrotReader::new(FragmentedSource::new(encoded, &[7, 17, 31]), SHARD, algo, false)
+            .read_appending(&mut output, SHARD)
+            .await
+            .expect_err("corrupt fragmented shard must fail");
+
+        assert_eq!(err.kind(), io::ErrorKind::InvalidData);
+        assert!(output.is_empty());
+    }
+
+    #[tokio::test]
+    async fn chunked_handoff_does_not_hash_when_verification_is_skipped() {
+        const SHARD: usize = 4096;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let mut encoded = Vec::new();
+        BitrotWriter::new(&mut encoded, SHARD, algo.clone())
+            .write(&vec![9u8; SHARD])
+            .await
+            .expect("write shard");
+        encoded[0] ^= 0xff;
+
+        let mut output = Vec::with_capacity(SHARD);
+        let mut reader = BitrotReader::new(FragmentedSource::new(encoded, &[7, 17, 31]), SHARD, algo, true);
+        reader
+            .read_appending(&mut output, SHARD)
+            .await
+            .expect("skipped verification must accept fragmented shard bytes");
+
+        assert_eq!(reader.last_verify_duration(), Duration::ZERO);
+        assert_eq!(output, vec![9u8; SHARD]);
+    }
+
    #[tokio::test]
    async fn read_appending_rejects_a_want_larger_than_the_shard() {
        let algo = HashAlgorithm::HighwayHash256;
@@ -1485,10 +1855,21 @@ mod tests {

        // Equivalence: same bytes out of both paths.
        let mut via_mem: Vec<u8> = Vec::with_capacity(SHARD);
-        BitrotReader::new(Cursor::new(Bytes::from(encoded.clone())), SHARD, algo.clone(), false)
+        let mut memory_reader = BitrotReader::new(Cursor::new(Bytes::from(encoded.clone())), SHARD, algo.clone(), false);
+        memory_reader
            .read_appending(&mut via_mem, SHARD)
            .await
            .expect("in-memory read");
+        assert_eq!(
+            memory_reader.chunks.capacity(),
+            0,
+            "the synchronous fast path must not allocate chunk storage"
+        );
+        assert_eq!(
+            memory_reader.buf.capacity(),
+            0,
+            "the synchronous fast path must not allocate scratch storage"
+        );

        let mut via_stream: Vec<u8> = Vec::with_capacity(SHARD);
        BitrotReader::new(Cursor::new(encoded), SHARD, algo, false)
@@ -1525,4 +1906,152 @@ mod tests {
        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
        assert!(out.is_empty(), "corrupt bytes must never reach the caller's buffer");
    }
+
+    #[tokio::test]
+    async fn streaming_fallback_reuses_initialized_scratch() {
+        const SHARD: usize = 4096;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data = vec![7u8; SHARD];
+        let encoded = encode_one_block(&data, SHARD, algo.clone()).await;
+        let source = ScratchReuseSource {
+            block: Some(Bytes::copy_from_slice(&encoded)),
+            saw_reused_scratch: false,
+        };
+        let mut reader = BitrotReader::new(source, SHARD, algo, false);
+        reader.buf = vec![0xa5; encoded.len()];
+        let mut output = Vec::new();
+
+        reader
+            .read_appending(&mut output, SHARD)
+            .await
+            .expect("streaming fallback should verify");
+
+        assert!(reader.inner.saw_reused_scratch, "capability probing must not clear reusable scratch");
+        assert_eq!(output, data);
+    }
+
+    #[tokio::test]
+    async fn chunked_handoff_bounds_production_sized_one_byte_fragments() {
+        const SHARD: usize = 1024 * 1024 / 4;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data: Vec<u8> = (0..SHARD).map(|index| (index % 251) as u8).collect();
+        let encoded = encode_one_block(&data, SHARD, algo.clone()).await;
+        let encoded_len = encoded.len();
+        let mut reader = BitrotReader::new(GeneratedChunkSource::new(encoded, 1), SHARD, algo, false);
+        let mut output = Vec::with_capacity(SHARD);
+
+        reader
+            .read_appending(&mut output, SHARD)
+            .await
+            .expect("one-byte fragments should verify with bounded retained state");
+
+        assert_eq!(output, data);
+        assert_eq!(reader.chunks.len(), MAX_RETAINED_CHUNKS_PER_BLOCK);
+        assert!(reader.chunks.capacity() <= MAX_RETAINED_CHUNKS_PER_BLOCK);
+        assert_eq!(reader.buf.len(), encoded_len - MAX_RETAINED_CHUNKS_PER_BLOCK);
+    }
+
+    #[tokio::test]
+    async fn chunked_handoff_keeps_sixty_four_frames_zero_copy_and_respects_poll_budget() {
+        const SHARD: usize = 1024 * 1024;
+        const FRAME: usize = 16 * 1024;
+        let algo = HashAlgorithm::HighwayHash256S;
+
+        let small_data = vec![3u8; 4096];
+        let small_encoded = encode_one_block(&small_data, 4096, algo.clone()).await;
+        let mut exact_reader =
+            BitrotReader::new(FragmentedSource::new(small_encoded.clone(), &[1; 63]), 4096, algo.clone(), false);
+        let mut exact_output = Vec::new();
+        exact_reader
+            .read_appending(&mut exact_output, 4096)
+            .await
+            .expect("exactly sixty-four frames should verify");
+        assert_eq!(exact_output, small_data);
+        assert_eq!(exact_reader.chunks.len(), MAX_RETAINED_CHUNKS_PER_BLOCK);
+        assert!(exact_reader.buf.is_empty(), "the threshold itself must remain zero-copy");
+
+        let mut yielded_reader = BitrotReader::new(FragmentedSource::new(small_encoded, &[1; 65]), 4096, algo.clone(), false);
+        let mut yielded_output = Vec::new();
+        let mut yielded_read = Box::pin(yielded_reader.read_appending(&mut yielded_output, 4096));
+        let mut cx = Context::from_waker(std::task::Waker::noop());
+        assert!(std::future::Future::poll(yielded_read.as_mut(), &mut cx).is_pending());
+        assert!(matches!(std::future::Future::poll(yielded_read.as_mut(), &mut cx), Poll::Ready(Ok(4096))));
+        drop(yielded_read);
+        assert_eq!(yielded_output, small_data);
+
+        let data = vec![7u8; SHARD];
+        let encoded = encode_one_block(&data, SHARD, algo.clone()).await;
+        let mut reader = BitrotReader::new(FragmentedSource::new(encoded, &[FRAME; 64]), SHARD, algo, false);
+        let mut output = Vec::with_capacity(SHARD);
+        let mut read = Box::pin(reader.read_appending(&mut output, SHARD));
+        assert!(
+            matches!(std::future::Future::poll(read.as_mut(), &mut cx), Poll::Ready(Ok(SHARD))),
+            "sixty-five normal HTTP frames should complete without a cooperative yield"
+        );
+        drop(read);
+
+        assert_eq!(output, data);
+        assert_eq!(reader.chunks.len(), MAX_RETAINED_CHUNKS_PER_BLOCK);
+        assert_eq!(reader.buf.len(), HashAlgorithm::HighwayHash256S.size());
+    }
+
+    #[tokio::test]
+    async fn chunked_tail_failures_preserve_errors_and_output() {
+        const SHARD: usize = 4096;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data = vec![7u8; SHARD];
+        let encoded = encode_one_block(&data, SHARD, algo.clone()).await;
+        let sentinel = vec![1u8, 2, 3];
+
+        let mut short_output = sentinel.clone();
+        let short_err = BitrotReader::new(GeneratedChunkSource::new(encoded[..100].to_vec(), 1), SHARD, algo.clone(), false)
+            .read_appending(&mut short_output, SHARD)
+            .await
+            .expect_err("EOF after the retention threshold must stay a short read");
+        assert_eq!(short_err.kind(), io::ErrorKind::UnexpectedEof);
+        assert_eq!(short_output, sentinel);
+
+        let mut corrupt = encoded.clone();
+        let last = corrupt.len() - 1;
+        corrupt[last] ^= 0xff;
+        let mut corrupt_output = sentinel.clone();
+        let corrupt_err = BitrotReader::new(GeneratedChunkSource::new(corrupt, 1), SHARD, algo.clone(), false)
+            .read_appending(&mut corrupt_output, SHARD)
+            .await
+            .expect_err("corrupt coalesced tail must fail verification");
+        assert_eq!(corrupt_err.kind(), io::ErrorKind::InvalidData);
+        assert_eq!(corrupt_output, sentinel);
+
+        let mut failed_output = sentinel.clone();
+        let body_err = BitrotReader::new(GeneratedChunkSource::failing(encoded, 1, 65), SHARD, algo, false)
+            .read_appending(&mut failed_output, SHARD)
+            .await
+            .expect_err("a terminal body error must not become EOF");
+        let source = body_err
+            .get_ref()
+            .and_then(|source| source.downcast_ref::<rustfs_rio::InternodeHttpError>())
+            .expect("body error should retain internode classification");
+        assert_eq!(source.kind(), rustfs_rio::InternodeHttpErrorKind::BodyStreamAborted);
+        assert_eq!(failed_output, sentinel);
+    }
+
+    #[tokio::test]
+    async fn chunked_handoff_rejects_invalid_source_contracts() {
+        const SHARD: usize = 64;
+        for mode in [
+            InvalidChunkMode::Empty,
+            InvalidChunkMode::Oversized,
+            InvalidChunkMode::UnsupportedAfterChunk,
+        ] {
+            let source = InvalidChunkSource { mode };
+            let mut output = vec![9u8];
+            let err = BitrotReader::new(source, SHARD, HashAlgorithm::HighwayHash256S, false)
+                .read_appending(&mut output, SHARD)
+                .await
+                .expect_err("invalid chunk contracts must fail closed");
+
+            assert_eq!(err.kind(), io::ErrorKind::InvalidData);
+            assert_eq!(output, vec![9u8]);
+        }
+    }
 }
@@ -25,10 +25,13 @@ use crate::disk::error_reduce::reduce_errs;
 use crate::erasure::codec::workspace::ShardBufferPool;
 use crate::erasure::coding::{BitrotReader, Erasure};
 use crate::io_support::bitrot::DeferredReaderStripeHandle;
-use crate::set_disk::shard_source::{ShardReadCost, ShardStripeSource, StripeReadState};
+use crate::set_disk::shard_source::{
+    INLINE_SHARD_SLOTS, ShardBuffers, ShardErrors, ShardReadCost, ShardStripeSource, StripeReadState,
+};
 use futures::FutureExt;
 use futures::stream::{FuturesUnordered, StreamExt};
 use pin_project_lite::pin_project;
+use smallvec::{SmallVec, smallvec};
 use std::future::Future;
 use std::io;
 use std::io::ErrorKind;
@@ -40,9 +43,12 @@ use tracing::{debug, error, warn};

 type ShardReadFuture<'a> = Pin<Box<dyn Future<Output = (usize, ShardReadCost, Result<Vec<u8>, Error>, bool)> + Send + 'a>>;

+type ShardIndexes = SmallVec<[usize; INLINE_SHARD_SLOTS]>;
+type ActiveReaders = SmallVec<[bool; INLINE_SHARD_SLOTS]>;
+
 /// One stripe's worth of shard buffers plus the per-shard read errors, as
 /// returned by `ParallelReader::read` / `read_stripe_timed`.
-type StripeReadOutput = (Vec<Option<Vec<u8>>>, Vec<Option<Error>>);
+type StripeReadOutput = (ShardBuffers, ShardErrors);

 const ENV_RUSTFS_SHARD_LOCALITY_SCHEDULING: &str = "RUSTFS_SHARD_LOCALITY_SCHEDULING";
 const ENV_RUSTFS_GET_SHARD_LOCALITY_PREFERENCE_ENABLE: &str = "RUSTFS_GET_SHARD_LOCALITY_PREFERENCE_ENABLE";
@@ -385,12 +391,13 @@ pub(crate) struct ParallelReader<R> {
    // Request-scoped shard buffers keyed by shard index. Keeping ownership in
    // `ParallelReader` avoids dropping unused parity/backup slot buffers between stripes.
    buffers: ShardBufferPool,
+    stripe_state: Option<Box<StripeReadState>>,
    // Lockstep-path state (verify_reconstruction == true). `engaged[i]` marks
    // readers that participate in each stripe read: all data slots from the
    // start, parity slots only once a data shard is missing/dead. Unengaged
    // parity stays an unopened deferred reader; `deferred_handles[i]` realigns
    // it to the current stripe when it is engaged mid-object (backlog#923).
-    engaged: Vec<bool>,
+    engaged: SmallVec<[bool; INLINE_SHARD_SLOTS]>,
    deferred_handles: Vec<Option<DeferredReaderStripeHandle>>,
    stripe_index: usize,
 }
@@ -573,7 +580,7 @@ where
        // behavior. With the gate on, only data slots start engaged; parity is
        // engaged on demand, stripe-aligned through its deferred handle.
        let data_shards_only = get_lockstep_data_shards_only_enabled();
-        let engaged = (0..readers.len())
+        let engaged: SmallVec<_> = (0..readers.len())
            .map(|index| !data_shards_only || index < e.data_shards)
            .collect();
        ParallelReader {
@@ -589,6 +596,7 @@ where
            verify_reconstruction,
            locality_preference_enabled: get_shard_locality_preference_enabled(),
            buffers: ShardBufferPool::new(e.data_shards + e.parity_shards),
+            stripe_state: None,
            engaged,
            deferred_handles: Vec::new(),
            stripe_index: 0,
@@ -612,7 +620,7 @@ where
 fn record_shard_read_result(
    shards: &mut [Option<Vec<u8>>],
    errs: &mut [Option<Error>],
-    retire_readers: &mut Vec<usize>,
+    retire_readers: &mut ShardIndexes,
    success: &mut usize,
    successful_costs: &mut ShardReadCostCounts,
    i: usize,
@@ -637,7 +645,7 @@ fn record_shard_read_result(
    }
 }

-fn retire_abandoned_readers(errs: &mut [Option<Error>], retire_readers: &mut Vec<usize>, active_readers: &[bool]) {
+fn retire_abandoned_readers(errs: &mut [Option<Error>], retire_readers: &mut ShardIndexes, active_readers: &[bool]) {
    for (i, active) in active_readers.iter().enumerate() {
        if !*active {
            continue;
@@ -692,7 +700,13 @@ where
    R: crate::erasure::coding::ShardSource,
 {
    #[hotpath::measure(impl_type = "ParallelReader")]
-    pub async fn read(&mut self) -> (Vec<Option<Vec<u8>>>, Vec<Option<Error>>) {
+    pub async fn read(&mut self) -> StripeReadOutput {
+        let mut state = StripeReadState::with_slot_count(self.readers.len(), self.data_shards);
+        self.read_into_state(&mut state).await;
+        state.into_parts()
+    }
+
+    async fn read_into_state(&mut self, state: &mut StripeReadState) {
        // On the reconstruction-verifying GET path, read every live shard reader
        // in lockstep so all readers advance one block per stripe and stay
        // mutually aligned. The adaptive data-first path below only reads
@@ -702,12 +716,14 @@ where
        // than the data shards, producing "inconsistent read source shards" and
        // truncating large-object GETs under concurrency (backlog#832).
        if self.verify_reconstruction {
-            return self.read_lockstep().await;
+            self.read_lockstep(state).await;
+            return;
        }
        // if self.readers.len() != self.total_shards {
        //     return Err(io::Error::new(ErrorKind::InvalidInput, "Invalid number of readers"));
        // }
        let num_readers = self.readers.len();
+        state.reset(num_readers, self.data_shards);

        let shard_size = if self.offset + self.shard_size > self.shard_file_size {
            self.shard_file_size - self.offset
@@ -716,7 +732,7 @@ where
        };

        if shard_size == 0 {
-            return (vec![None; num_readers], vec![None; num_readers]);
+            return;
        }

        // Advance to the next stripe so the following read() computes the correct
@@ -727,8 +743,7 @@ where
        // is only read above to derive `shard_size`, so advancing here is safe.
        self.offset += shard_size;

-        let mut shards: Vec<Option<Vec<u8>>> = vec![None; num_readers];
-        let mut errs = vec![None; num_readers];
+        let (shards, errs) = state.parts_mut();
        let read_costs = self.read_costs.as_slice();
        let locality_preference_enabled = self.locality_preference_enabled;
        let low_cost_available = self
@@ -759,11 +774,11 @@ where

        self.buffers.ensure_slots(num_readers);

-        let mut retire_readers = Vec::new();
+        let mut retire_readers = ShardIndexes::new();
        if num_readers >= self.data_shards {
            let mut reader_iter = ReaderLaunchIter::new(&mut self.readers, read_costs, locality_preference_enabled);
            let mut sets = FuturesUnordered::new();
-            let mut active_readers = vec![false; num_readers];
+            let mut active_readers: ActiveReaders = smallvec![false; num_readers];
            let stripe_read_start = self.metrics_path.map(|_| Instant::now());
            let mut scheduled = 0usize;
            for _ in 0..self.data_shards {
@@ -875,8 +890,8 @@ where
                }

                let result_is_err = record_shard_read_result(
-                    &mut shards,
-                    &mut errs,
+                    shards,
+                    errs,
                    &mut retire_readers,
                    &mut success,
                    &mut successful_costs,
@@ -937,8 +952,8 @@ where
                    active_readers[i] = false;
                    completed += 1;
                    if record_shard_read_result(
-                        &mut shards,
-                        &mut errs,
+                        shards,
+                        errs,
                        &mut retire_readers,
                        &mut success,
                        &mut successful_costs,
@@ -950,7 +965,7 @@ where
                        failed += 1;
                    }
                }
-                retire_abandoned_readers(&mut errs, &mut retire_readers, &active_readers);
+                retire_abandoned_readers(errs, &mut retire_readers, &active_readers);
            }

            if let Some(path) = self.metrics_path {
@@ -994,8 +1009,6 @@ where
        for i in retire_readers {
            self.readers[i] = None;
        }
-
-        (shards, errs)
    }

    /// Lockstep stripe read for the reconstruction-verifying GET path.
@@ -1023,18 +1036,18 @@ where
    /// stripe would reintroduce the desync. A parity reader that cannot be
    /// realigned (no pending deferred handle) is likewise retired instead of
    /// being read out of position.
-    async fn read_lockstep(&mut self) -> (Vec<Option<Vec<u8>>>, Vec<Option<Error>>) {
+    async fn read_lockstep(&mut self, state: &mut StripeReadState) {
        let num_readers = self.readers.len();
+        state.reset(num_readers, self.data_shards);
        let shard_size = if self.offset + self.shard_size > self.shard_file_size {
            self.shard_file_size - self.offset
        } else {
            self.shard_size
        };

-        let mut shards: Vec<Option<Vec<u8>>> = vec![None; num_readers];
-        let mut errs: Vec<Option<Error>> = vec![None; num_readers];
+        let (shards, errs) = state.parts_mut();
        if shard_size == 0 {
-            return (shards, errs);
+            return;
        }

        // Advance to the next stripe (see the matching note in `read`); the
@@ -1071,7 +1084,7 @@ where
        // Pre-claim per-slot buffers so the `self.readers` borrow below stays
        // disjoint from `self.buffers`; `Some(buffer)` also records which slots
        // participate, avoiding a per-stripe sidecar allocation.
-        let mut bufs: Vec<Option<Vec<u8>>> = Vec::with_capacity(num_readers);
+        let mut bufs: ShardBuffers = SmallVec::with_capacity(num_readers);
        for i in 0..num_readers {
            bufs.push(if self.engaged[i] && self.readers[i].is_some() {
                Some(self.buffers.take(i, shard_size))
@@ -1086,7 +1099,7 @@ where
        let locality_preference_enabled = self.locality_preference_enabled;
        let stripe_read_start = metrics_path.map(|_| Instant::now());

-        let mut retire_readers = Vec::new();
+        let mut retire_readers = ShardIndexes::new();
        let mut scheduled = 0usize;
        let mut success = 0usize;
        let mut completed = 0usize;
@@ -1272,8 +1285,6 @@ where
        for i in retire_readers {
            self.readers[i] = None;
        }
-
-        (shards, errs)
    }

    /// Attempt to bring an as-yet-unread parity reader into the lockstep read
@@ -1330,10 +1341,20 @@ impl<R> ShardStripeSource for ParallelReader<R>
 where
    R: crate::erasure::coding::ShardSource,
 {
-    async fn read_next_stripe(&mut self) -> StripeReadState {
-        let read_quorum = self.data_shards;
-        let (shards, errors) = ParallelReader::read(self).await;
-        StripeReadState::from_parts_with_read_costs(shards, errors, &self.read_costs, read_quorum)
+    async fn read_next_stripe(&mut self) -> Box<StripeReadState> {
+        let mut state = self
+            .stripe_state
+            .take()
+            .unwrap_or_else(|| Box::new(StripeReadState::with_slot_count(self.readers.len(), self.data_shards)));
+        self.read_into_state(&mut state).await;
+        state
+    }
+
+    fn recycle_stripe(&mut self, mut state: Box<StripeReadState>) {
+        self.recycle_shards(state.shards_mut());
+        state.reset(0, self.data_shards);
+        debug_assert!(self.stripe_state.is_none(), "a stripe cannot be recycled twice");
+        self.stripe_state = Some(state);
    }
 }

@@ -1351,10 +1372,7 @@ fn get_data_block_len(shards: &[Option<Vec<u8>>], data_blocks: usize) -> usize {
 /// stripe-read stage timer. Factored out so the depth-1 prefetch loop and the
 /// serial loop time reads identically. A free `async fn` (rather than a closure)
 /// so the returned future's borrow of `reader` is correctly tied to the call.
-async fn read_stripe_timed<R>(
-    reader: &mut ParallelReader<R>,
-    stage_metrics_enabled: bool,
-) -> (Vec<Option<Vec<u8>>>, Vec<Option<Error>>)
+async fn read_stripe_timed<R>(reader: &mut ParallelReader<R>, stage_metrics_enabled: bool) -> StripeReadOutput
 where
    R: crate::erasure::coding::ShardSource,
 {
@@ -1967,6 +1985,93 @@ mod tests {

    type BoxedShardReader = crate::io_support::bitrot::ShardReader;

+    #[test]
+    fn parallel_reader_keeps_stripe_scratch_out_of_line() {
+        eprintln!(
+            "parallel_reader={} stripe_state={} cached_state={}",
+            std::mem::size_of::<ParallelReader<Cursor<Vec<u8>>>>(),
+            std::mem::size_of::<StripeReadState>(),
+            std::mem::size_of::<Option<Box<StripeReadState>>>()
+        );
+        assert_eq!(
+            std::mem::size_of::<Option<Box<StripeReadState>>>(),
+            std::mem::size_of::<usize>(),
+            "the request-scoped cache must remain pointer-sized",
+        );
+    }
+
+    #[tokio::test]
+    async fn parallel_reader_preserves_slot_count_above_inline_capacity() {
+        const DATA_SHARDS: usize = INLINE_SHARD_SLOTS;
+        const TOTAL_SHARDS: usize = INLINE_SHARD_SLOTS + 1;
+        let readers = std::iter::repeat_with(|| None).take(TOTAL_SHARDS).collect();
+        let erasure = Erasure::new(DATA_SHARDS, 1, DATA_SHARDS);
+        let mut reader: ParallelReader<Cursor<Vec<u8>>> = ParallelReader::new(readers, erasure, 0, DATA_SHARDS);
+
+        let (shards, errors) = reader.read().await;
+
+        assert!(shards.spilled());
+        assert!(errors.spilled());
+        assert_eq!(shards.len(), TOTAL_SHARDS);
+        assert_eq!(errors.len(), TOTAL_SHARDS);
+    }
+
+    #[tokio::test]
+    async fn codec_reader_reuses_inline_and_spilled_stripe_scratch_between_reads() {
+        for total_shards in [INLINE_SHARD_SLOTS, INLINE_SHARD_SLOTS + 1] {
+            let data_shards = total_shards - 1;
+            let readers = std::iter::repeat_with(|| None).take(total_shards).collect();
+            let erasure = Erasure::new(data_shards, 1, data_shards * 2);
+            let mut reader: ParallelReader<Cursor<Vec<u8>>> = ParallelReader::new(readers, erasure, 0, data_shards * 2);
+
+            let first = ShardStripeSource::read_next_stripe(&mut reader).await;
+            let first_state = (&*first) as *const StripeReadState;
+            let first_storage = first.scratch_storage();
+            assert_eq!(first_storage.2, total_shards > INLINE_SHARD_SLOTS);
+            assert_eq!(first_storage.3, total_shards > INLINE_SHARD_SLOTS);
+            ShardStripeSource::recycle_stripe(&mut reader, first);
+
+            let second = ShardStripeSource::read_next_stripe(&mut reader).await;
+            let second_storage = second.scratch_storage();
+
+            assert_eq!(
+                (&*second) as *const StripeReadState,
+                first_state,
+                "the request-scoped state must be reused"
+            );
+            assert_eq!(second_storage.0, first_storage.0, "shard slots must reuse their allocation");
+            assert_eq!(second_storage.1, first_storage.1, "error slots must reuse their allocation");
+            assert_eq!(second.into_parts().0.len(), total_shards);
+        }
+    }
+
+    #[tokio::test]
+    async fn codec_reader_returns_shard_allocations_to_the_request_pool() {
+        const SHARD_SIZE: usize = 16;
+        let hash_algo = HashAlgorithm::None;
+        let readers = vec![Some(create_reader(SHARD_SIZE, 2, 0x5a, &hash_algo, false).await)];
+        let erasure = Erasure::new(1, 0, SHARD_SIZE);
+        let mut reader = ParallelReader::new(readers, erasure, 0, SHARD_SIZE * 2);
+
+        let first = ShardStripeSource::read_next_stripe(&mut reader).await;
+        let first_allocation = first
+            .shard_allocation(0)
+            .expect("the first stripe should own its shard allocation");
+        ShardStripeSource::recycle_stripe(&mut reader, first);
+        assert_eq!(
+            reader.buffers.stored_allocation(0),
+            Some(first_allocation),
+            "recycling a stripe must return its shard allocation to the request pool"
+        );
+
+        let second = ShardStripeSource::read_next_stripe(&mut reader).await;
+        assert_eq!(
+            second.shard_allocation(0),
+            Some(first_allocation),
+            "the next stripe must reuse the pooled shard allocation"
+        );
+    }
+
    /// Counts the raw bytes pulled from a shard stream, to prove which shards
    /// a decode path actually touches (backlog#923 call-count evidence).
    struct CountingShardReader {
@@ -2343,6 +2448,19 @@ mod tests {
        assert_eq!(err.expect("range beyond total length should fail").kind(), ErrorKind::InvalidInput);
    }

+    #[tokio::test]
+    async fn test_erasure_decode_zero_length_does_not_read_or_emit() {
+        let erasure = Erasure::new(2, 1, 64);
+        let readers: Vec<Option<BitrotReader<Cursor<Vec<u8>>>>> = vec![None, None, None];
+        let mut output = Vec::new();
+
+        let (written, err) = erasure.decode(&mut output, readers, 0, 0, 0).await;
+
+        assert_eq!(written, 0);
+        assert!(err.is_none());
+        assert!(output.is_empty());
+    }
+
    #[tokio::test]
    async fn test_erasure_decode_with_read_costs_restores_missing_data_shard_range() {
        const DATA_SHARDS: usize = 2;
@@ -65,7 +65,7 @@ enum FillPolicy {
 }

 impl FillPolicy {
-    fn from_env() -> Self {
+    fn load() -> Self {
        match rustfs_utils::get_env_usize(
            ENV_RUSTFS_GET_CODEC_STREAMING_MAX_INFLIGHT,
            DEFAULT_RUSTFS_GET_CODEC_STREAMING_MAX_INFLIGHT,
@@ -75,6 +75,22 @@ impl FillPolicy {
        }
    }

+    fn from_env() -> Self {
+        #[cfg(test)]
+        {
+            Self::load()
+        }
+        #[cfg(not(test))]
+        {
+            Self::cached_core(Self::load)
+        }
+    }
+
+    fn cached_core(load: impl FnOnce() -> Self) -> Self {
+        static CACHED: std::sync::OnceLock<FillPolicy> = std::sync::OnceLock::new();
+        *CACHED.get_or_init(load)
+    }
+
    const fn max_inflight(self) -> usize {
        match self {
            Self::SingleInFlight => 1,
@@ -479,22 +495,30 @@ where
    let mut deferred_error = None;
    let fill_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
    let stripe_read_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
-    let state = source.read_next_stripe().await;
+    let mut state = source.read_next_stripe().await;
    record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_STRIPE_READ, stripe_read_stage_start);
    let decode_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
    let mut output_buf = reusable_buffers.pop().unwrap_or_default();
-    let result =
-        match decode_stripe_into(metrics_path, stage_metrics_enabled, engine, workspace, state, remaining, &mut output_buf) {
-            Ok(true) => Ok(Some(output_buf)),
-            Ok(false) => {
-                reusable_buffers.push(output_buf);
-                Ok(None)
-            }
-            Err(err) => {
-                reusable_buffers.push(output_buf);
-                Err(err)
-            }
-        };
+    let result = match decode_stripe_into(
+        metrics_path,
+        stage_metrics_enabled,
+        engine,
+        workspace,
+        &mut state,
+        remaining,
+        &mut output_buf,
+    ) {
+        Ok(true) => Ok(Some(output_buf)),
+        Ok(false) => {
+            reusable_buffers.push(output_buf);
+            Ok(None)
+        }
+        Err(err) => {
+            reusable_buffers.push(output_buf);
+            Err(err)
+        }
+    };
+    source.recycle_stripe(state);
    record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_DECODE, decode_stage_start);
    if let Ok(Some(first_buf)) = result.as_ref() {
        let mut remaining_after_first = remaining.saturating_sub(first_buf.len());
@@ -503,7 +527,7 @@ where
                break;
            }
            let stripe_read_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
-            let state = source.read_next_stripe().await;
+            let mut state = source.read_next_stripe().await;
            record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_STRIPE_READ, stripe_read_stage_start);
            let decode_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
            let mut queued_buf = reusable_buffers.pop().unwrap_or_default();
@@ -512,10 +536,11 @@ where
                stage_metrics_enabled,
                engine,
                workspace,
-                state,
+                &mut state,
                remaining_after_first,
                &mut queued_buf,
            );
+            source.recycle_stripe(state);
            record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_DECODE, decode_stage_start);
            match queued_result {
                Ok(true) => {
@@ -717,7 +742,7 @@ fn decode_stripe_into<E>(
    stage_metrics_enabled: bool,
    engine: &E,
    workspace: &mut E::Workspace,
-    state: StripeReadState,
+    state: &mut StripeReadState,
    remaining: usize,
    output: &mut Vec<u8>,
 ) -> io::Result<bool>
@@ -725,7 +750,7 @@ where
    E: ErasureDecodeEngine,
 {
    output.clear();
-    if state.slots().is_empty() {
+    if state.is_empty() {
        return Ok(false);
    }
    if !state.can_decode() {
@@ -741,13 +766,12 @@ where
        );
        record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_RECONSTRUCT, reconstruct_stage_start);
        let emit_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
-        emit_data_shards_into(&state, engine.data_shards(), engine.block_size(), remaining, output)?;
+        emit_data_shards_into(state, engine.data_shards(), engine.block_size(), remaining, output)?;
        record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_EMIT, emit_stage_start);
        return Ok(true);
    }

-    let (mut shards, _errs) = state.into_parts();
-    let reconstruct_outcome = match engine.reconstruct_into(&mut shards, workspace) {
+    let reconstruct_outcome = match engine.reconstruct_into(state.shards_mut(), workspace) {
        Ok(outcome) => outcome,
        Err(err) => {
            record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_RECONSTRUCT, reconstruct_stage_start);
@@ -757,7 +781,7 @@ where
    rustfs_io_metrics::record_get_object_reconstruct_outcome(metrics_path, engine.engine_name(), reconstruct_outcome);
    record_get_stage_duration_if_enabled(metrics_path, GET_STAGE_RECONSTRUCT, reconstruct_stage_start);

-    if shards.len() < engine.data_shards() {
+    if state.shards_mut().len() < engine.data_shards() {
        return Err(io::Error::new(
            ErrorKind::UnexpectedEof,
            "decoded stripe has fewer shards than data shard count",
@@ -766,7 +790,7 @@ where

    let emit_stage_start = get_stage_timer_if_enabled(stage_metrics_enabled);
    reserve_output_capacity(output, engine.block_size().min(remaining));
-    for shard in shards.iter().take(engine.data_shards()) {
+    for shard in state.shards_mut().iter().take(engine.data_shards()) {
        if output.len() >= remaining {
            break;
        }
@@ -806,10 +830,7 @@ fn emit_data_shards_into(
        if output.len() >= remaining {
            break;
        }
-        let Some(slot) = state.slot_by_index(index) else {
-            return Err(io::Error::new(ErrorKind::UnexpectedEof, "decoded stripe is missing a data shard"));
-        };
-        let Some(shard) = slot.data_bytes() else {
+        let Some(shard) = state.data_bytes(index) else {
            return Err(io::Error::new(ErrorKind::UnexpectedEof, "decoded stripe is missing a data shard"));
        };
        let copy_len = shard.len().min(remaining - output.len());
@@ -826,7 +847,7 @@ mod tests {
    };
    use crate::erasure::coding::decode::ParallelReader;
    use crate::erasure::coding::{BitrotReader, BitrotWriter, Erasure};
-    use crate::set_disk::shard_source::{ShardSlot, StripeReadState};
+    use crate::set_disk::shard_source::StripeReadState;
    use rustfs_utils::HashAlgorithm;
    use std::collections::VecDeque;
    use std::future::{pending, poll_fn};
@@ -845,6 +866,13 @@ mod tests {
        read_count: Option<Arc<AtomicUsize>>,
    }

+    struct RecordingStripeSource {
+        stripes: VecDeque<StripeReadState>,
+        read_quorum: usize,
+        reads: usize,
+        recycles: usize,
+    }
+
    struct BlockingSource {
        started: Arc<Notify>,
        dropped: Arc<AtomicUsize>,
@@ -899,25 +927,43 @@ mod tests {

    #[async_trait::async_trait]
    impl ShardStripeSource for VecStripeSource {
-        async fn read_next_stripe(&mut self) -> StripeReadState {
+        async fn read_next_stripe(&mut self) -> Box<StripeReadState> {
            if let Some(read_count) = &self.read_count {
                read_count.fetch_add(1, Ordering::SeqCst);
            }
-            self.stripes
-                .pop_front()
-                .unwrap_or_else(|| StripeReadState::new(Vec::new(), self.read_quorum))
+            Box::new(
+                self.stripes
+                    .pop_front()
+                    .unwrap_or_else(|| StripeReadState::from_parts(Vec::new(), Vec::new(), self.read_quorum)),
+            )
+        }
+    }
+
+    #[async_trait::async_trait]
+    impl ShardStripeSource for RecordingStripeSource {
+        async fn read_next_stripe(&mut self) -> Box<StripeReadState> {
+            self.reads += 1;
+            Box::new(
+                self.stripes
+                    .pop_front()
+                    .unwrap_or_else(|| StripeReadState::from_parts(Vec::new(), Vec::new(), self.read_quorum)),
+            )
+        }
+
+        fn recycle_stripe(&mut self, _state: Box<StripeReadState>) {
+            self.recycles += 1;
        }
    }

    #[async_trait::async_trait]
    impl ShardStripeSource for BlockingSource {
-        async fn read_next_stripe(&mut self) -> StripeReadState {
+        async fn read_next_stripe(&mut self) -> Box<StripeReadState> {
            let _guard = BlockingSourceDropGuard {
                dropped: Arc::clone(&self.dropped),
            };
            self.started.notify_one();
            pending::<()>().await;
-            StripeReadState::new(Vec::new(), self.read_quorum)
+            Box::new(StripeReadState::from_parts(Vec::new(), Vec::new(), self.read_quorum))
        }
    }

@@ -1090,6 +1136,23 @@ mod tests {
        });
    }

+    #[test]
+    fn fill_policy_production_cache_loads_once() {
+        use std::cell::Cell;
+
+        let loads = Cell::new(0);
+        for _ in 0..3 {
+            assert_eq!(
+                FillPolicy::cached_core(|| {
+                    loads.set(loads.get() + 1);
+                    FillPolicy::DualInFlight
+                }),
+                FillPolicy::DualInFlight
+            );
+        }
+        assert_eq!(loads.get(), 1, "the production fill policy must not re-read the environment per reader");
+    }
+
    #[test]
    fn erasure_decode_reader_rejects_invalid_engine_shape() {
        let source = VecStripeSource {
@@ -1689,7 +1752,10 @@ mod tests {
            .pop_front()
            .expect("first stripe should exist");
        let mut source = VecStripeSource {
-            stripes: VecDeque::from([first_state, StripeReadState::new(Vec::new(), erasure.data_shards)]),
+            stripes: VecDeque::from([
+                first_state,
+                StripeReadState::from_parts(Vec::new(), Vec::new(), erasure.data_shards),
+            ]),
            read_quorum: erasure.data_shards,
            read_count: None,
        };
@@ -1724,13 +1790,14 @@ mod tests {
            .stripes
            .pop_front()
            .expect("first stripe should exist");
-        let mut source = VecStripeSource {
+        let mut source = RecordingStripeSource {
            stripes: VecDeque::from([
                first_state,
-                StripeReadState::new(vec![ShardSlot::data(0, vec![1])], erasure.data_shards),
+                StripeReadState::from_parts(vec![Some(vec![1])], Vec::new(), erasure.data_shards),
            ]),
            read_quorum: erasure.data_shards,
-            read_count: None,
+            reads: 0,
+            recycles: 0,
        };
        let engine = LegacyEcDecodeEngine::new(erasure);
        let mut workspace = engine.prepare_workspace(4).expect("workspace should be prepared");
@@ -1756,6 +1823,8 @@ mod tests {
                .kind(),
            ErrorKind::Other
        );
+        assert_eq!(source.reads, 2, "the fill must read the primary and queued stripe");
+        assert_eq!(source.recycles, source.reads, "every completed stripe read must be recycled");
    }

    #[tokio::test]
@@ -1768,7 +1837,7 @@ mod tests {
                    .stripes
                    .pop_front()
                    .expect("first stripe should exist"),
-                StripeReadState::new(Vec::new(), erasure.data_shards),
+                StripeReadState::from_parts(Vec::new(), Vec::new(), erasure.data_shards),
            ]),
            read_quorum: erasure.data_shards,
            read_count: None,
@@ -2028,17 +2097,11 @@ mod tests {
    }

    #[test]
-    fn emit_data_shards_preserves_output_order_for_out_of_order_slots() {
-        let state = StripeReadState::new(
-            vec![
-                ShardSlot::data(1, b"cd".to_vec()),
-                ShardSlot::data(0, b"ab".to_vec()),
-                ShardSlot::data(2, b"ef".to_vec()),
-            ],
-            2,
-        );
+    fn emit_data_shards_preserves_output_order() {
+        let state =
+            StripeReadState::from_parts(vec![Some(b"ab".to_vec()), Some(b"cd".to_vec()), Some(b"ef".to_vec())], Vec::new(), 2);

-        let output = emit_data_shards(&state, 3, 6, 5).expect("out-of-order data slots should emit by shard index");
+        let output = emit_data_shards(&state, 3, 6, 5).expect("data slots should emit by shard index");

        assert_eq!(output, b"abcde");
    }
@@ -2051,27 +2114,27 @@ mod tests {
        };
        let mut workspace = engine.prepare_workspace(4).expect("workspace should be prepared");
        let mut output = Vec::with_capacity(1);
-        let short_state = StripeReadState::new(vec![ShardSlot::data(0, vec![1, 2, 3, 4])], 1);
+        let mut short_state = StripeReadState::from_parts(vec![Some(vec![1, 2, 3, 4])], Vec::new(), 1);

        let err = decode_stripe_into(
            GET_OBJECT_PATH_CODEC_STREAMING,
            false,
            &engine,
            &mut workspace,
-            short_state,
+            &mut short_state,
            8,
            &mut output,
        )
        .expect_err("decoded stripe shorter than data shard count must fail");
        assert_eq!(err.kind(), ErrorKind::UnexpectedEof);

-        let missing_state = StripeReadState::from_parts(vec![None, Some(vec![5, 6, 7, 8])], Vec::new(), 1);
+        let mut missing_state = StripeReadState::from_parts(vec![None, Some(vec![5, 6, 7, 8])], Vec::new(), 1);
        let err = decode_stripe_into(
            GET_OBJECT_PATH_CODEC_STREAMING,
            false,
            &engine,
            &mut workspace,
-            missing_state,
+            &mut missing_state,
            8,
            &mut output,
        )
@@ -2082,6 +2145,35 @@ mod tests {
        assert!(output.capacity() >= 32);
    }

+    #[test]
+    fn decode_stripe_reconstructs_in_place_without_replacing_slot_storage() {
+        let erasure = Erasure::new(2, 1, 8);
+        let engine = LegacyEcDecodeEngine::new(erasure.clone());
+        let mut workspace = engine.prepare_workspace(4).expect("workspace should be prepared");
+        let encoded = erasure.encode_data(b"abcdefgh").expect("test stripe should encode");
+        let mut shards = encoded.into_iter().map(|shard| Some(shard.to_vec())).collect::<Vec<_>>();
+        shards[0] = None;
+        let mut state = StripeReadState::from_parts(shards, vec![Some(DiskError::FileCorrupt)], 2);
+        let before = state.scratch_storage();
+        let mut output = Vec::new();
+
+        let decoded = decode_stripe_into(
+            GET_OBJECT_PATH_CODEC_STREAMING,
+            false,
+            &engine,
+            &mut workspace,
+            &mut state,
+            8,
+            &mut output,
+        )
+        .expect("degraded stripe should reconstruct");
+
+        assert!(decoded);
+        assert_eq!(output, b"abcdefgh");
+        assert_eq!(state.scratch_storage().0, before.0, "reconstruction must retain shard slot storage");
+        assert_eq!(state.scratch_storage().1, before.1, "unused error storage must not be rebuilt");
+    }
+
    #[tokio::test]
    async fn erasure_decode_reader_reports_short_source() {
        let erasure = Erasure::new(4, 2, 32);
@@ -18,10 +18,12 @@ use crate::disk::error_reduce::{
 };
 use crate::erasure::coding::BitrotWriterWrapper;
 use crate::erasure::coding::Erasure;
+use crate::erasure::coding::erasure::EncodedBlock;
 use crate::runtime::sources as runtime_sources;
 use bytes::{Bytes, BytesMut};
 use futures::StreamExt;
 use futures::stream::FuturesUnordered;
+use rustfs_utils::HashAlgorithm;
 use std::sync::Arc;
 use std::time::Instant;
 use std::vec;
@@ -91,6 +93,11 @@ fn use_bytesmut_ingest() -> bool {
    })
 }

+fn small_ingest_capacity(erasure: &Erasure, size_hint: usize) -> usize {
+    let data_len = size_hint.min(erasure.block_size);
+    erasure.encoded_capacity_for_data_len(data_len).min(erasure.block_size)
+}
+
 /// Keeps the encoder producer scoped to its parent future. Tokio detaches a
 /// task when its `JoinHandle` is dropped, so the producer must be aborted when
 /// an upload is cancelled before the encode pipeline finishes.
@@ -218,8 +225,8 @@ async fn send_queued<T>(
    sender.send(InflightEntry::new(entry, bytes)).await
 }

-fn queued_batch_bytes(batch: &[Vec<Bytes>]) -> usize {
-    batch.iter().map(|block| queued_block_bytes(block)).sum()
+fn queued_batch_bytes(batch: &[EncodedBlock]) -> usize {
+    batch.iter().map(EncodedBlock::queued_bytes).sum()
 }

 fn dominant_error_summary_label(summary: &WriteQuorumFailureSummary) -> &'static str {
@@ -331,7 +338,7 @@ impl<'a> MultiWriter<'a> {
        }
    }

-    async fn write_shard(writer_opt: &mut Option<BitrotWriterWrapper>, err: &mut Option<Error>, shard: &Bytes) {
+    async fn write_shard(writer_opt: &mut Option<BitrotWriterWrapper>, err: &mut Option<Error>, shard: &[u8]) {
        match writer_opt {
            Some(writer) => {
                match writer.write(shard).await {
@@ -356,12 +363,20 @@ impl<'a> MultiWriter<'a> {
    }

    pub async fn write(&mut self, data: Vec<Bytes>) -> std::io::Result<()> {
-        assert_eq!(data.len(), self.writers.len());
+        self.write_shards(data.iter().map(Bytes::as_ref)).await
+    }
+
+    async fn write_block(&mut self, block: &EncodedBlock) -> std::io::Result<()> {
+        self.write_shards(block.shards()).await
+    }
+
+    async fn write_shards<'b>(&mut self, shards: impl ExactSizeIterator<Item = &'b [u8]>) -> std::io::Result<()> {
+        assert_eq!(shards.len(), self.writers.len());

        let budget = self.next_progress_budget();
        {
            let mut futures = FuturesUnordered::new();
-            for ((writer_opt, err), shard) in self.writers.iter_mut().zip(self.errs.iter_mut()).zip(data.iter()) {
+            for ((writer_opt, err), shard) in self.writers.iter_mut().zip(self.errs.iter_mut()).zip(shards) {
                if err.is_some() {
                    continue; // Skip if we already have an error for this writer
                }
@@ -485,10 +500,10 @@ impl<'a> MultiWriter<'a> {
 }

 impl Erasure {
-    async fn encode_block(self: Arc<Self>, encode_buf: Vec<u8>, len: usize) -> std::io::Result<(Vec<Bytes>, Vec<u8>)> {
+    async fn encode_block(self: Arc<Self>, encode_buf: Vec<u8>, len: usize) -> std::io::Result<(EncodedBlock, Vec<u8>)> {
        let encode_stage_start = stage_timer_if_enabled();
        let encode_once = move || {
-            let res = self.encode_data(&encode_buf[..len]);
+            let res = self.encode_data_block(&encode_buf[..len]);
            (res, encode_buf)
        };

@@ -513,9 +528,9 @@ impl Erasure {
        Ok((res?, returned_buf))
    }

-    async fn encode_block_bytes_mut(self: Arc<Self>, encode_buf: BytesMut, len: usize) -> std::io::Result<Vec<Bytes>> {
+    async fn encode_block_bytes_mut(self: Arc<Self>, encode_buf: BytesMut, len: usize) -> std::io::Result<EncodedBlock> {
        let encode_stage_start = stage_timer_if_enabled();
-        let encode_once = move || self.encode_data_bytes_mut(encode_buf, len);
+        let encode_once = move || self.encode_data_bytes_mut_block(encode_buf, len);

        let res = match tokio::runtime::Handle::current().runtime_flavor() {
            // Same rationale as encode_block: inline the short EC burst on the
@@ -540,13 +555,14 @@ impl Erasure {
        writers: &mut [Option<BitrotWriterWrapper>],
        quorum: usize,
        require_single_block: bool,
+        size_hint: usize,
    ) -> std::io::Result<(R, usize)>
    where
        R: AsyncRead + Send + Sync + Unpin,
    {
        use tokio::io::AsyncReadExt;

-        let mut buf = Vec::with_capacity(self.block_size);
+        let mut buf = Vec::with_capacity(small_ingest_capacity(&self, size_hint));
        let total = if require_single_block {
            let read_limit = self
                .block_size
@@ -570,13 +586,46 @@ impl Erasure {
            ));
        }

-        let shards = self.encode_data_owned(buf)?;
+        let block = self.encode_data_owned_block(buf)?;
        let mut mw = MultiWriter::new(writers, quorum);
-        mw.write(shards).await?;
+        mw.write_block(&block).await?;
        mw.shutdown().await?;
        Ok((reader, total))
    }

+    /// Encode a small inline object directly into its per-disk bitrot payloads.
+    /// The returned bytes are the same `[hash][shard]` representation produced
+    /// by `BitrotWriter`, ready to be embedded in each disk's staged `xl.meta`.
+    #[hotpath::measure(impl_type = "Erasure")]
+    pub(crate) async fn encode_inline_shards_with_size_hint<R>(
+        self: Arc<Self>,
+        mut reader: R,
+        size_hint: usize,
+    ) -> std::io::Result<(R, usize, Vec<Bytes>)>
+    where
+        R: AsyncRead + Send + Sync + Unpin,
+    {
+        use tokio::io::AsyncReadExt;
+
+        let mut buf = Vec::with_capacity(small_ingest_capacity(&self, size_hint));
+        let total = reader.read_to_end(&mut buf).await?;
+        if total == 0 {
+            return Ok((reader, 0, Vec::new()));
+        }
+
+        let block = self.encode_data_owned_block(buf)?;
+        let mut inline_shards = Vec::with_capacity(block.shards().len());
+        for shard in block.shards() {
+            let hash = HashAlgorithm::HighwayHash256S.hash_encode(shard);
+            let mut encoded = BytesMut::with_capacity(hash.as_ref().len() + shard.len());
+            encoded.extend_from_slice(hash.as_ref());
+            encoded.extend_from_slice(shard);
+            inline_shards.push(encoded.freeze());
+        }
+
+        Ok((reader, total, inline_shards))
+    }
+
    #[hotpath::measure(impl_type = "Erasure")]
    pub async fn encode<R>(
        self: Arc<Self>,
@@ -618,7 +667,7 @@ impl Erasure {
        let expanded_block_bytes = self.shard_size().saturating_mul(self.total_shard_count());
        let max_inflight_bytes = erasure_encode_max_inflight_bytes();
        let inflight_blocks = encode_channel_capacity(expanded_block_bytes, max_inflight_bytes);
-        let (tx, mut rx) = mpsc::channel::<InflightEntry<Vec<Bytes>>>(inflight_blocks);
+        let (tx, mut rx) = mpsc::channel::<InflightEntry<EncodedBlock>>(inflight_blocks);

        let mut task = AbortOnDropTask::new(tokio::spawn(async move {
            let block_size = self.block_size;
@@ -640,7 +689,7 @@ impl Erasure {
                            let encode_buf = buf;
                            let res = self.clone().encode_block_bytes_mut(encode_buf, n).await?;
                            buf = BytesMut::with_capacity(ingest_capacity);
-                            let queued_bytes = queued_block_bytes(&res);
+                            let queued_bytes = res.queued_bytes();
                            let _producer_stage = rustfs_io_metrics::track_ec_encode_producer_bytes(queued_bytes);
                            let send_wait_stage_start = stage_timer_if_enabled();
                            if let Err(err) = send_queued(&tx, res, queued_bytes).await {
@@ -670,7 +719,7 @@ impl Erasure {
                            let encode_buf = std::mem::take(&mut buf);
                            let (res, returned_buf) = self.clone().encode_block(encode_buf, n).await?;
                            buf = returned_buf;
-                            let queued_bytes = queued_block_bytes(&res);
+                            let queued_bytes = res.queued_bytes();
                            let _producer_stage = rustfs_io_metrics::track_ec_encode_producer_bytes(queued_bytes);
                            let send_wait_stage_start = stage_timer_if_enabled();
                            if let Err(err) = send_queued(&tx, res, queued_bytes).await {
@@ -714,9 +763,9 @@ impl Erasure {
            if block.is_empty() {
                break;
            }
-            let _writer_stage = rustfs_io_metrics::track_ec_encode_writer_bytes(queued_block_bytes(&block));
+            let _writer_stage = rustfs_io_metrics::track_ec_encode_writer_bytes(block.queued_bytes());
            let write_stage_start = stage_timer_if_enabled();
-            if let Err(err) = writers.write(block).await {
+            if let Err(err) = writers.write_block(&block).await {
                write_err = Some(err);
                break;
            }
@@ -763,7 +812,7 @@ impl Erasure {
        let inflight_blocks = encode_channel_capacity(expanded_block_bytes, max_inflight_bytes);
        let batch_blocks = encode_batch_block_count().min(inflight_blocks);
        let channel_capacity = inflight_blocks.div_ceil(batch_blocks).max(1);
-        let (tx, mut rx) = mpsc::channel::<InflightEntry<Vec<Vec<Bytes>>>>(channel_capacity);
+        let (tx, mut rx) = mpsc::channel::<InflightEntry<Vec<EncodedBlock>>>(channel_capacity);

        let mut task = AbortOnDropTask::new(tokio::spawn(async move {
            let block_size = self.block_size;
@@ -780,7 +829,7 @@ impl Erasure {
                        let encode_buf = std::mem::take(&mut buf);
                        let (res, returned_buf) = self.clone().encode_block(encode_buf, n).await?;
                        buf = returned_buf;
-                        let queued_bytes = queued_block_bytes(&res);
+                        let queued_bytes = res.queued_bytes();
                        pending_batch_bytes = pending_batch_bytes.saturating_add(queued_bytes);
                        pending_batch.push(res);
                        drop(pending_batch_stage.take());
@@ -839,7 +888,7 @@ impl Erasure {
            let _writer_stage = rustfs_io_metrics::track_ec_encode_writer_bytes(queued_batch_bytes(&batch));
            let write_stage_start = stage_timer_if_enabled();
            for block in batch {
-                if let Err(err) = writers.write(block).await {
+                if let Err(err) = writers.write_block(&block).await {
                    write_err = Some(err);
                    break;
                }
@@ -880,7 +929,24 @@ impl Erasure {
    where
        R: AsyncRead + Send + Sync + Unpin,
    {
-        self.encode_small_direct(reader, writers, quorum, false).await
+        let size_hint = self.block_size;
+        self.encode_small_direct(reader, writers, quorum, false, size_hint).await
+    }
+
+    /// Size-aware inline fast path. `size_hint` only controls the bounded initial
+    /// allocation; reads remain authoritative.
+    #[hotpath::measure(impl_type = "Erasure")]
+    pub async fn encode_inline_small_with_size_hint<R>(
+        self: Arc<Self>,
+        reader: R,
+        writers: &mut [Option<BitrotWriterWrapper>],
+        quorum: usize,
+        size_hint: usize,
+    ) -> std::io::Result<(R, usize)>
+    where
+        R: AsyncRead + Send + Sync + Unpin,
+    {
+        self.encode_small_direct(reader, writers, quorum, false, size_hint).await
    }

    /// Fast path for single-block non-inline objects: avoids the producer/consumer
@@ -895,7 +961,24 @@ impl Erasure {
    where
        R: AsyncRead + Send + Sync + Unpin,
    {
-        self.encode_small_direct(reader, writers, quorum, true).await
+        let size_hint = self.block_size;
+        self.encode_small_direct(reader, writers, quorum, true, size_hint).await
+    }
+
+    /// Size-aware single-block fast path. `size_hint` only controls the bounded
+    /// initial allocation; reads remain authoritative.
+    #[hotpath::measure(impl_type = "Erasure")]
+    pub async fn encode_single_block_non_inline_with_size_hint<R>(
+        self: Arc<Self>,
+        reader: R,
+        writers: &mut [Option<BitrotWriterWrapper>],
+        quorum: usize,
+        size_hint: usize,
+    ) -> std::io::Result<(R, usize)>
+    where
+        R: AsyncRead + Send + Sync + Unpin,
+    {
+        self.encode_small_direct(reader, writers, quorum, true, size_hint).await
    }
 }

@@ -1855,7 +1938,11 @@ mod tests {
        let baseline = rustfs_io_metrics::current_ec_encode_inflight_bytes();
        let (tx, rx) = mpsc::channel(2);
        let mut rx = rx;
-        let batch = vec![vec![Bytes::from_static(b"queued")], vec![Bytes::from_static(b"batch")]];
+        let erasure = Erasure::new(1, 0, 16);
+        let batch = vec![
+            erasure.encode_data_block(b"queued").expect("first block should encode"),
+            erasure.encode_data_block(b"batch").expect("second block should encode"),
+        ];
        let batch_bytes = queued_batch_bytes(&batch);

        send_queued(&tx, batch, batch_bytes).await.expect("batch should be queued");
@@ -2077,6 +2164,39 @@ mod tests {
        );
    }

+    #[tokio::test]
+    async fn cancelling_inline_small_drops_stalled_write() {
+        const BLOCK_SIZE: usize = 16;
+
+        let (writer_entered_tx, writer_entered) = oneshot::channel();
+        let writes = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let mut writers = vec![Some(bitrot_writer_plain(
+            StallOnWriteWithSignal {
+                entered: Some(writer_entered_tx),
+                writes: writes.clone(),
+            },
+            BLOCK_SIZE,
+        ))];
+        let erasure = Arc::new(Erasure::new(1, 0, BLOCK_SIZE));
+        let reader = tokio::io::BufReader::new(Cursor::new(vec![0xA5; BLOCK_SIZE - 1]));
+        let encode = tokio::spawn(async move { erasure.encode_inline_small(reader, &mut writers, 1).await });
+
+        tokio::time::timeout(Duration::from_secs(1), writer_entered)
+            .await
+            .expect("inline writer should enter before cancellation")
+            .expect("stalling writer should signal entry");
+        encode.abort();
+        assert!(
+            matches!(encode.await, Err(err) if err.is_cancelled()),
+            "inline encode task should be cancelled"
+        );
+        assert_eq!(
+            writes.load(std::sync::atomic::Ordering::SeqCst),
+            1,
+            "cancellation must drop the stalled write instead of polling it again"
+        );
+    }
+
    #[tokio::test]
    async fn encode_returns_unexpected_eof_for_truncated_limited_reader() {
        let committed = Arc::new(Mutex::new(Vec::new()));
@@ -2196,11 +2316,11 @@ mod tests {
            .expect("bytesmut encode should succeed on current-thread runtime");

        let expected_shard_size = payload.len().div_ceil(erasure.data_shards);
-        assert_eq!(shards.len(), erasure.total_shard_count());
-        assert!(shards.iter().all(|shard| shard.len() == expected_shard_size));
+        assert_eq!(shards.shards().len(), erasure.total_shard_count());
+        assert!(shards.shards().all(|shard| shard.len() == expected_shard_size));

        let mut restored = Vec::new();
-        for shard in shards.iter().take(erasure.data_shards) {
+        for shard in shards.shards().take(erasure.data_shards) {
            restored.extend_from_slice(shard);
        }
        restored.truncate(payload.len());
@@ -2293,13 +2413,51 @@ mod tests {

        let erasure = Arc::new(Erasure::new(1, 0, 16));
        let reader = tokio::io::BufReader::new(Cursor::new(Vec::<u8>::new()));
-        let (_reader, total) = erasure.encode_inline_small(reader, &mut writers, 1).await.unwrap();
+        let (_reader, total) = erasure
+            .encode_inline_small_with_size_hint(reader, &mut writers, 1, 0)
+            .await
+            .unwrap();

        assert_eq!(total, 0);
        // No shutdown was called, so nothing should be committed
        assert!(committed.lock().unwrap().is_empty());
    }

+    #[tokio::test]
+    async fn encode_inline_shards_matches_writer_bitrot_layout() {
+        const DATA_SHARDS: usize = 2;
+        const PARITY_SHARDS: usize = 2;
+        const BLOCK_SIZE: usize = 64;
+        let checksum_algo = HashAlgorithm::HighwayHash256S;
+        for uses_legacy in [false, true] {
+            let erasure = Arc::new(Erasure::new_with_options(DATA_SHARDS, PARITY_SHARDS, BLOCK_SIZE, uses_legacy));
+            for payload in [Vec::new(), vec![0xA5], vec![0x5A; BLOCK_SIZE - 1], vec![0xC3; BLOCK_SIZE]] {
+                let reader = tokio::io::BufReader::new(Cursor::new(payload.clone()));
+                let (_reader, total, inline_shards) = erasure
+                    .clone()
+                    .encode_inline_shards_with_size_hint(reader, payload.len())
+                    .await
+                    .expect("inline shards should encode");
+
+                assert_eq!(total, payload.len());
+                if payload.is_empty() {
+                    assert!(inline_shards.is_empty());
+                    continue;
+                }
+
+                let raw_shards = erasure.encode_data(&payload).expect("reference shards should encode");
+                assert_eq!(inline_shards.len(), DATA_SHARDS + PARITY_SHARDS);
+                for (inline, raw) in inline_shards.iter().zip(raw_shards) {
+                    let mut writer =
+                        BitrotWriterWrapper::new(CustomWriter::new_inline_buffer(), raw.len(), checksum_algo.clone());
+                    writer.write(&raw).await.expect("reference writer should accept shard");
+                    writer.shutdown().await.expect("reference writer should shutdown");
+                    assert_eq!(inline.as_ref(), writer.into_inline_data().expect("reference writer should retain bytes"));
+                }
+            }
+        }
+    }
+
    /// encode_inline_small: small payload is encoded into the correct number of shards
    /// and each writer receives data after shutdown.
    #[tokio::test]
@@ -2325,7 +2483,10 @@ mod tests {
        let payload = b"hello inline small";
        let erasure = Arc::new(Erasure::new(DATA_SHARDS, PARITY_SHARDS, BLOCK_SIZE));
        let reader = tokio::io::BufReader::new(Cursor::new(payload.to_vec()));
-        let (_reader, total) = erasure.encode_inline_small(reader, &mut writers, DATA_SHARDS).await.unwrap();
+        let (_reader, total) = erasure
+            .encode_inline_small_with_size_hint(reader, &mut writers, DATA_SHARDS, 1)
+            .await
+            .unwrap();

        assert_eq!(total, payload.len());
        // All shards must have received data (shutdown flushed the bitrot header + shard bytes)
@@ -2392,7 +2553,7 @@ mod tests {
        let erasure = Arc::new(Erasure::new(DATA_SHARDS, PARITY_SHARDS, BLOCK_SIZE));
        let reader = tokio::io::BufReader::new(Cursor::new(payload));
        let err = erasure
-            .encode_single_block_non_inline(reader, &mut writers, DATA_SHARDS)
+            .encode_single_block_non_inline_with_size_hint(reader, &mut writers, DATA_SHARDS, BLOCK_SIZE)
            .await
            .expect_err("single-block fast path must reject oversized readers");

@@ -2403,6 +2564,21 @@ mod tests {
        }
    }

+    #[test]
+    fn small_ingest_capacity_uses_bounded_size_hint() {
+        let erasure = Erasure::new(4, 2, 1024 * 1024);
+        assert_eq!(small_ingest_capacity(&erasure, 0), 0);
+        assert_eq!(small_ingest_capacity(&erasure, 4 * 1024), 6 * 1024);
+        assert_eq!(small_ingest_capacity(&erasure, 16 * 1024), 24 * 1024);
+        assert_eq!(small_ingest_capacity(&erasure, usize::MAX), 1024 * 1024);
+
+        let legacy = Erasure::new_with_options(4, 2, 1024 * 1024, true);
+        assert_eq!(small_ingest_capacity(&legacy, 4 * 1024), 6 * 1024);
+
+        let high_parity = Erasure::new(4, 12, 1024 * 1024);
+        assert_eq!(small_ingest_capacity(&high_parity, usize::MAX), 1024 * 1024);
+    }
+
    #[tokio::test]
    async fn read_full_buf_or_eof_returns_none_on_empty_reader() {
        let mut reader = Cursor::new(Vec::<u8>::new());
@@ -2445,7 +2621,7 @@ mod tests {
        assert_eq!(&next[..], &data[16..]);
    }

-    async fn committed_shards_for_ingest_mode(use_bytesmut_ingest: bool, uses_legacy: bool, payload: &[u8]) -> Vec<Vec<u8>> {
+    async fn committed_shards_for_pipeline(pipeline: EncodePipeline, uses_legacy: bool, payload: &[u8]) -> Vec<Vec<u8>> {
        const DATA_SHARDS: usize = 2;
        const PARITY_SHARDS: usize = 2;
        const TOTAL_SHARDS: usize = DATA_SHARDS + PARITY_SHARDS;
@@ -2459,10 +2635,16 @@ mod tests {

        let erasure = Arc::new(Erasure::new_with_options(DATA_SHARDS, PARITY_SHARDS, BLOCK_SIZE, uses_legacy));
        let reader = tokio::io::BufReader::new(Cursor::new(payload.to_vec()));
-        let (_reader, total) = erasure
-            .encode_with_ingest_mode(reader, &mut writers, DATA_SHARDS, use_bytesmut_ingest)
-            .await
-            .expect("encode should succeed");
+        let (_reader, total) = match pipeline {
+            EncodePipeline::Vec => {
+                erasure
+                    .encode_with_ingest_mode(reader, &mut writers, DATA_SHARDS, false)
+                    .await
+            }
+            EncodePipeline::BytesMut => erasure.encode_with_ingest_mode(reader, &mut writers, DATA_SHARDS, true).await,
+            EncodePipeline::Batched => erasure.encode_batched(reader, &mut writers, DATA_SHARDS).await,
+        }
+        .expect("encode should succeed");
        assert_eq!(total, payload.len());

        committed
@@ -2471,31 +2653,64 @@ mod tests {
            .collect()
    }

-    /// HP-10 (rustfs/backlog#931) merge gate: the BytesMut ingest path must produce
-    /// byte-for-byte identical shard streams to the default Vec ingest path, for both
-    /// legacy-aware shard-size formulas, across empty, sub-block, exactly-full-block,
-    /// and multi-block-with-partial-tail payloads.
+    async fn expected_committed_shards(uses_legacy: bool, payload: &[u8]) -> Vec<Vec<u8>> {
+        const DATA_SHARDS: usize = 2;
+        const PARITY_SHARDS: usize = 2;
+        const TOTAL_SHARDS: usize = DATA_SHARDS + PARITY_SHARDS;
+        const BLOCK_SIZE: usize = 64;
+
+        let committed: Vec<Arc<Mutex<Vec<u8>>>> = (0..TOTAL_SHARDS).map(|_| Arc::new(Mutex::new(Vec::new()))).collect();
+        let mut writers: Vec<BitrotWriterWrapper> = committed
+            .iter()
+            .map(|c| bitrot_writer(DeferredCommitWriter::new(c.clone()), BLOCK_SIZE / DATA_SHARDS))
+            .collect();
+        let erasure = Erasure::new_with_options(DATA_SHARDS, PARITY_SHARDS, BLOCK_SIZE, uses_legacy);
+
+        for block in payload.chunks(BLOCK_SIZE) {
+            let shards = erasure.encode_data(block).expect("reference block should encode");
+            for (writer, shard) in writers.iter_mut().zip(shards) {
+                let written = writer.write(&shard).await.expect("reference shard should write");
+                assert_eq!(written, shard.len());
+            }
+        }
+        for writer in &mut writers {
+            writer.shutdown().await.expect("reference writer should commit");
+        }
+
+        committed
+            .iter()
+            .map(|c| c.lock().expect("committed buffer should be lockable").clone())
+            .collect()
+    }
+
+    /// The streaming and batched paths must produce the same bitrot-wrapped shard
+    /// bytes as the public block encoder for both shard-size formulas and all block
+    /// boundary shapes.
    #[tokio::test]
    async fn bytesmut_ingest_matches_vec_ingest_byte_for_byte() {
        const BLOCK_SIZE: usize = 64;
        let payloads: Vec<Vec<u8>> = vec![
            Vec::new(),
-            b"tiny".to_vec(),
+            vec![1],
+            vec![2; BLOCK_SIZE - 1],
            (0..BLOCK_SIZE as u32).map(|i| i as u8).collect(), // exactly one full block
-            vec![3u8; BLOCK_SIZE * 4],                         // whole number of blocks
+            vec![4; BLOCK_SIZE + 1],
+            vec![3u8; BLOCK_SIZE * 4],                                           // whole number of blocks
            (0..(BLOCK_SIZE * 3 + 7) as u32).map(|i| (i % 251) as u8).collect(), // partial tail
        ];

        for uses_legacy in [false, true] {
            for payload in &payloads {
-                let vec_path = committed_shards_for_ingest_mode(false, uses_legacy, payload).await;
-                let bytesmut_path = committed_shards_for_ingest_mode(true, uses_legacy, payload).await;
-                assert_eq!(
-                    vec_path,
-                    bytesmut_path,
-                    "ingest paths must be byte-identical (legacy={uses_legacy}, payload_len={})",
-                    payload.len()
-                );
+                let expected = expected_committed_shards(uses_legacy, payload).await;
+                for pipeline in [EncodePipeline::Vec, EncodePipeline::BytesMut, EncodePipeline::Batched] {
+                    let actual = committed_shards_for_pipeline(pipeline, uses_legacy, payload).await;
+                    assert_eq!(
+                        actual,
+                        expected,
+                        "streaming shards must match the public block encoder (legacy={uses_legacy}, payload_len={})",
+                        payload.len()
+                    );
+                }
            }
        }
    }
@@ -29,6 +29,46 @@ use tokio::io::AsyncRead;
 use tracing::warn;
 use uuid::Uuid;

+pub(crate) struct EncodedBlock {
+    data: Bytes,
+    shard_size: usize,
+}
+
+impl EncodedBlock {
+    fn empty() -> Self {
+        Self {
+            data: Bytes::new(),
+            shard_size: 0,
+        }
+    }
+
+    pub(crate) fn is_empty(&self) -> bool {
+        self.data.is_empty()
+    }
+
+    pub(crate) fn queued_bytes(&self) -> usize {
+        self.data.len()
+    }
+
+    pub(crate) fn shards(&self) -> impl ExactSizeIterator<Item = &[u8]> {
+        debug_assert!(self.shard_size > 0, "only non-empty encoded blocks reach shard writers");
+        debug_assert_eq!(self.data.len() % self.shard_size, 0);
+        self.data.chunks_exact(self.shard_size)
+    }
+
+    fn into_shards(mut self, shard_count: usize) -> Vec<Bytes> {
+        if self.shard_size == 0 {
+            return vec![Bytes::new(); shard_count];
+        }
+
+        let mut shards = Vec::with_capacity(shard_count);
+        for _ in 0..shard_count {
+            shards.push(self.data.split_to(self.shard_size));
+        }
+        shards
+    }
+}
+
 const MODERN_MAX_TOTAL_SHARDS: usize = <reed_solomon_erasure::galois_8::Field as reed_solomon_erasure::Field>::ORDER;
 const MODERN_REED_SOLOMON_CACHE_MAX_ENTRIES: usize = 64;

@@ -675,106 +715,48 @@ impl Erasure {
    #[tracing::instrument(level = "debug", skip_all, fields(data_len=data.len()))]
    #[hotpath::measure(impl_type = "Erasure")]
    pub fn encode_data(&self, data: &[u8]) -> io::Result<Vec<Bytes>> {
-        let shard_size_fn = if self.uses_legacy {
-            calc_shard_size_legacy
-        } else {
-            calc_shard_size
-        };
-        let per_shard_size = shard_size_fn(data.len(), self.data_shards);
-        if per_shard_size == 0 {
-            return Ok(vec![Bytes::new(); self.total_shard_count()]);
-        }
-        let need_total_size = per_shard_size * self.total_shard_count();
+        self.encode_data_block_inner(data)
+            .map(|block| block.into_shards(self.total_shard_count()))
+    }

-        let mut data_buffer = BytesMut::with_capacity(need_total_size);
+    #[tracing::instrument(level = "debug", skip_all, fields(data_len=data.len()))]
+    #[hotpath::measure(label = "Erasure::encode_data", impl_type = "Erasure")]
+    pub(crate) fn encode_data_block(&self, data: &[u8]) -> io::Result<EncodedBlock> {
+        self.encode_data_block_inner(data)
+    }
+
+    fn encode_data_block_inner(&self, data: &[u8]) -> io::Result<EncodedBlock> {
+        let mut data_buffer = BytesMut::with_capacity(self.encoded_capacity_for_data_len(data.len()));
        data_buffer.extend_from_slice(data);
-        data_buffer.resize(need_total_size, 0u8);
-
-        {
-            let data_slices: SmallVec<[&mut [u8]; 16]> = data_buffer.chunks_exact_mut(per_shard_size).collect();
-
-            if self.parity_shards > 0 {
-                if self.uses_legacy {
-                    if let Some(encoder) = self.legacy_encoder.as_ref() {
-                        encoder.encode(data_slices)?;
-                    } else {
-                        warn!("parity_shards > 0, uses_legacy but legacy_encoder is None");
-                    }
-                } else if let Some(encoder) = self.encoder.as_ref() {
-                    encoder.encode(data_slices)?;
-                } else {
-                    warn!("parity_shards > 0, but encoder is None");
-                }
-            }
-        }
-
-        // Zero-copy split, all shards reference data_buffer
-        let mut data_buffer = data_buffer.freeze();
-        let mut shards = Vec::with_capacity(self.total_shard_count());
-        for _ in 0..self.total_shard_count() {
-            let shard = data_buffer.split_to(per_shard_size);
-            shards.push(shard);
-        }
-
-        Ok(shards)
+        self.encode_buffer(data_buffer, data.len())
    }

    /// Encode owned data, avoiding a copy when the caller already has a heap buffer.
    /// Falls back to copying into a new buffer if zero-copy conversion fails.
    #[hotpath::measure(impl_type = "Erasure")]
    pub fn encode_data_owned(&self, data: Vec<u8>) -> io::Result<Vec<Bytes>> {
-        let shard_size_fn = if self.uses_legacy {
-            calc_shard_size_legacy
-        } else {
-            calc_shard_size
-        };
-        let per_shard_size = shard_size_fn(data.len(), self.data_shards);
-        if per_shard_size == 0 {
-            return Ok(vec![Bytes::new(); self.total_shard_count()]);
-        }
-        let need_total_size = per_shard_size * self.total_shard_count();
+        self.encode_data_owned_block_inner(data)
+            .map(|block| block.into_shards(self.total_shard_count()))
+    }

+    #[hotpath::measure(label = "Erasure::encode_data_owned", impl_type = "Erasure")]
+    pub(crate) fn encode_data_owned_block(&self, data: Vec<u8>) -> io::Result<EncodedBlock> {
+        self.encode_data_owned_block_inner(data)
+    }
+
+    fn encode_data_owned_block_inner(&self, data: Vec<u8>) -> io::Result<EncodedBlock> {
+        let data_len = data.len();
        // Try zero-copy: Vec<u8> -> Bytes -> BytesMut (succeeds when refcount == 1)
-        let mut data_buffer = match Bytes::from(data).try_into_mut() {
-            Ok(mut bm) => {
-                bm.resize(need_total_size, 0u8);
-                bm
-            }
+        let data_buffer = match Bytes::from(data).try_into_mut() {
+            Ok(data_buffer) => data_buffer,
            Err(b) => {
                // Rare path: refcount != 1, fall back to copy
-                let mut bm = BytesMut::with_capacity(need_total_size);
-                bm.extend_from_slice(&b);
-                bm.resize(need_total_size, 0u8);
-                bm
+                let mut data_buffer = BytesMut::with_capacity(self.encoded_capacity_for_data_len(data_len));
+                data_buffer.extend_from_slice(&b);
+                data_buffer
            }
        };
-
-        {
-            let data_slices: SmallVec<[&mut [u8]; 16]> = data_buffer.chunks_exact_mut(per_shard_size).collect();
-
-            if self.parity_shards > 0 {
-                if self.uses_legacy {
-                    if let Some(encoder) = self.legacy_encoder.as_ref() {
-                        encoder.encode(data_slices)?;
-                    } else {
-                        warn!("parity_shards > 0, uses_legacy but legacy_encoder is None");
-                    }
-                } else if let Some(encoder) = self.encoder.as_ref() {
-                    encoder.encode(data_slices)?;
-                } else {
-                    warn!("parity_shards > 0, but encoder is None");
-                }
-            }
-        }
-
-        let mut data_buffer = data_buffer.freeze();
-        let mut shards = Vec::with_capacity(self.total_shard_count());
-        for _ in 0..self.total_shard_count() {
-            let shard = data_buffer.split_to(per_shard_size);
-            shards.push(shard);
-        }
-
-        Ok(shards)
+        self.encode_buffer(data_buffer, data_len)
    }

    /// Encode data from an owned `BytesMut` buffer, avoiding the initial copy
@@ -786,7 +768,17 @@ impl Erasure {
    /// `data_len <= block_size` — both shard-size formulas are monotone in
    /// `data_len` — so this function never reallocates the buffer.
    #[hotpath::measure(impl_type = "Erasure")]
-    pub fn encode_data_bytes_mut(&self, mut data_buffer: BytesMut, data_len: usize) -> io::Result<Vec<Bytes>> {
+    pub fn encode_data_bytes_mut(&self, data_buffer: BytesMut, data_len: usize) -> io::Result<Vec<Bytes>> {
+        self.encode_buffer(data_buffer, data_len)
+            .map(|block| block.into_shards(self.total_shard_count()))
+    }
+
+    #[hotpath::measure(label = "Erasure::encode_data_bytes_mut", 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)
+    }
+
+    fn encode_buffer(&self, mut data_buffer: BytesMut, data_len: usize) -> io::Result<EncodedBlock> {
        let shard_size_fn = if self.uses_legacy {
            calc_shard_size_legacy
        } else {
@@ -794,7 +786,7 @@ impl Erasure {
        };
        let per_shard_size = shard_size_fn(data_len, self.data_shards);
        if per_shard_size == 0 {
-            return Ok(vec![Bytes::new(); self.total_shard_count()]);
+            return Ok(EncodedBlock::empty());
        }
        let need_total_size = per_shard_size * self.total_shard_count();

@@ -821,14 +813,10 @@ impl Erasure {
            }
        }

-        let mut data_buffer = data_buffer.freeze();
-        let mut shards = Vec::with_capacity(self.total_shard_count());
-        for _ in 0..self.total_shard_count() {
-            let shard = data_buffer.split_to(per_shard_size);
-            shards.push(shard);
-        }
-
-        Ok(shards)
+        Ok(EncodedBlock {
+            data: data_buffer.freeze(),
+            shard_size: per_shard_size,
+        })
    }

    /// Decode and reconstruct missing data shards in-place.
@@ -968,6 +956,15 @@ impl Erasure {
        self.data_shards + self.parity_shards
    }

+    pub(crate) fn encoded_capacity_for_data_len(&self, data_len: usize) -> usize {
+        let shard_size_fn = if self.uses_legacy {
+            calc_shard_size_legacy
+        } else {
+            calc_shard_size
+        };
+        shard_size_fn(data_len, self.data_shards).saturating_mul(self.total_shard_count())
+    }
+
    /// Whether the erasure dimensions are safe for the shard/offset arithmetic.
    ///
    /// `block_size` and `data_shards` come straight from on-disk metadata; a
@@ -1489,10 +1486,16 @@ mod tests {
    fn encode_data_owned_matches_borrowed_path() {
        for uses_legacy in [false, true] {
            let erasure = Erasure::new_with_options(4, 2, 64, uses_legacy);
-
-            assert_owned_encode_matches_borrowed(&erasure, Vec::new());
-            assert_owned_encode_matches_borrowed(&erasure, b"small payload".to_vec());
-            assert_owned_encode_matches_borrowed(&erasure, (0_u8..37).collect());
+            for data in [
+                Vec::new(),
+                vec![0xA5; 1],
+                b"small payload".to_vec(),
+                (0_u8..37).collect(),
+                vec![0xA5; erasure.block_size - 1],
+                vec![0x5A; erasure.block_size],
+            ] {
+                assert_owned_encode_matches_borrowed(&erasure, data);
+            }
        }
    }

@@ -1538,6 +1541,52 @@ mod tests {
        }
    }

+    #[test]
+    fn streaming_encoded_block_uses_one_contiguous_backing_buffer() {
+        for uses_legacy in [false, true] {
+            let erasure = Erasure::new_with_options(8, 8, 64, uses_legacy);
+
+            for data_len in [0, 1, 63, 64] {
+                let data = (0..data_len).map(|i| i as u8).collect::<Vec<_>>();
+                let expected = erasure.encode_data(&data).expect("public encode should succeed");
+                let borrowed = erasure
+                    .encode_data_block(&data)
+                    .expect("borrowed streaming encode should succeed");
+                let owned = erasure
+                    .encode_data_owned_block(data.clone())
+                    .expect("owned streaming encode should succeed");
+                let bytes_mut = erasure
+                    .encode_data_bytes_mut_block(BytesMut::from(&data[..]), data.len())
+                    .expect("BytesMut streaming encode should succeed");
+
+                assert_eq!(borrowed.queued_bytes(), owned.queued_bytes());
+                assert_eq!(borrowed.queued_bytes(), bytes_mut.queued_bytes());
+
+                if data_len == 0 {
+                    assert!(expected.iter().all(Bytes::is_empty));
+                    assert!(borrowed.is_empty());
+                    assert!(owned.is_empty());
+                    assert!(bytes_mut.is_empty());
+                    continue;
+                }
+
+                assert!(borrowed.shards().eq(expected.iter().map(Bytes::as_ref)));
+                assert!(owned.shards().eq(expected.iter().map(Bytes::as_ref)));
+                assert!(bytes_mut.shards().eq(expected.iter().map(Bytes::as_ref)));
+                assert_eq!(borrowed.shards().len(), 16);
+                let first = borrowed.shards().next().expect("encoded block should have shards").as_ptr();
+                for (index, shard) in borrowed.shards().enumerate() {
+                    assert_eq!(shard.as_ptr(), first.wrapping_add(index * shard.len()));
+                }
+            }
+        }
+        assert_eq!(
+            std::mem::size_of::<EncodedBlock>(),
+            std::mem::size_of::<Bytes>() + std::mem::size_of::<usize>(),
+            "queue entries must contain one backing buffer handle, not per-shard handles"
+        );
+    }
+
    /// HP-10 capacity invariant: both shard-size formulas are monotone in `data_len`,
    /// so pre-reserving `shard_size(block_size) * total_shard_count` covers the
    /// `need_total_size` of every block-or-smaller payload and the ingest buffer
@@ -204,6 +204,8 @@ pub enum StorageError {
        required: usize,
        achieved: usize,
    },
+    #[error("Bucket quota exceeded. Current usage: {current} bytes, limit: {limit} bytes")]
+    QuotaExceeded { current: u64, limit: u64 },

    // ── Generic ──────────────────────────────────────────────────────
    #[error("Unexpected error")]
@@ -356,6 +358,13 @@ impl From<StorageError> for DiskError {
            StorageError::VolumeNotFound => DiskError::VolumeNotFound,
            StorageError::VolumeExists => DiskError::VolumeExists,
            StorageError::FileNameTooLong => DiskError::FileNameTooLong,
+            StorageError::FaultyRemoteDisk => DiskError::FaultyRemoteDisk,
+            StorageError::DiskAccessDenied => DiskError::DiskAccessDenied,
+            StorageError::DriveIsRoot => DiskError::DriveIsRoot,
+            StorageError::IsNotRegular => DiskError::IsNotRegular,
+            StorageError::VolumeNotEmpty => DiskError::VolumeNotEmpty,
+            StorageError::VolumeAccessDenied => DiskError::VolumeAccessDenied,
+            StorageError::FileAccessDenied => DiskError::FileAccessDenied,
            _ => DiskError::other(val),
        }
    }
@@ -540,6 +549,10 @@ impl Clone for StorageError {
                required: *required,
                achieved: *achieved,
            },
+            StorageError::QuotaExceeded { current, limit } => StorageError::QuotaExceeded {
+                current: *current,
+                limit: *limit,
+            },
        }
    }
 }
@@ -627,6 +640,7 @@ impl StorageError {
            StorageError::NotModified => StorageErrorCode::NotModified,
            StorageError::InvalidPartNumber(_) => StorageErrorCode::InvalidPartNumber,
            StorageError::NamespaceLockQuorumUnavailable { .. } => StorageErrorCode::NamespaceLockQuorumUnavailable,
+            StorageError::QuotaExceeded { .. } => StorageErrorCode::QuotaExceeded,
        }
    }

@@ -752,6 +766,10 @@ impl StorageError {
                required: Default::default(),
                achieved: Default::default(),
            }),
+            StorageErrorCode::QuotaExceeded => Some(StorageError::QuotaExceeded {
+                current: Default::default(),
+                limit: Default::default(),
+            }),
        }
    }
 }
@@ -1301,6 +1319,7 @@ mod tests {
            .to_u32(),
            0x42
        );
+        assert_eq!(StorageError::QuotaExceeded { current: 1, limit: 2 }.to_u32(), 0x53);
    }

    #[test]
@@ -1319,6 +1338,10 @@ mod tests {
            StorageError::from_u32(0x42),
            Some(StorageError::NamespaceLockQuorumUnavailable { .. })
        ));
+        assert!(matches!(
+            StorageError::from_u32(0x53),
+            Some(StorageError::QuotaExceeded { current: 0, limit: 0 })
+        ));

        // Test invalid code returns None
        assert!(StorageError::from_u32(0xFF).is_none());
@@ -1476,6 +1499,49 @@ mod tests {
        }
    }

+    // Every DiskError variant must survive DiskError -> StorageError -> DiskError
+    // unchanged. A variant that degrades to `DiskError::Io` on the way back loses
+    // its identity for quorum aggregation (`reduce_errs` classifies by variant
+    // equality), so ignore-list entries such as FaultyRemoteDisk and
+    // DiskAccessDenied would silently stop matching.
+    #[test]
+    fn test_disk_error_storage_error_round_trip_identity_all_variants() {
+        // DiskError codes are contiguous from 0x01, so enumerating via from_u32
+        // covers every variant and picks up newly appended ones automatically.
+        let all_variants: Vec<DiskError> = (1u32..).map_while(DiskError::from_u32).collect();
+        assert!(
+            all_variants.len() >= 42,
+            "DiskError variant enumeration shrank: got {}, expected at least 42",
+            all_variants.len()
+        );
+
+        for original in all_variants {
+            let storage_error: StorageError = original.clone().into();
+            let round_tripped: DiskError = storage_error.into();
+
+            assert_eq!(
+                std::mem::discriminant(&original),
+                std::mem::discriminant(&round_tripped),
+                "round trip changed variant: {original:?} -> {round_tripped:?}"
+            );
+            assert_eq!(original, round_tripped, "round trip not identical for {original:?}");
+        }
+
+        // Io is the only payload-carrying variant: a representative kind and
+        // message must both survive the round trip.
+        let io_original = DiskError::Io(IoError::new(ErrorKind::PermissionDenied, "denied"));
+        let storage_error: StorageError = io_original.clone().into();
+        let io_round_tripped: DiskError = storage_error.into();
+        assert_eq!(io_original, io_round_tripped);
+        match io_round_tripped {
+            DiskError::Io(inner) => {
+                assert_eq!(inner.kind(), ErrorKind::PermissionDenied);
+                assert_eq!(inner.to_string(), "denied");
+            }
+            other => panic!("expected DiskError::Io, got {other:?}"),
+        }
+    }
+
    #[test]
    fn test_storage_error_from_io_error() {
        // Test direct IO error conversion
@@ -1549,6 +1615,7 @@ mod tests {
            StorageError::DecommissionAlreadyRunning,
            StorageError::RebalanceAlreadyRunning,
            StorageError::OperationCanceled,
+            StorageError::QuotaExceeded { current: 1, limit: 2 },
        ];

        for original_error in test_errors {
@@ -22,12 +22,13 @@ use crate::diagnostics::get::{
 #[cfg(feature = "hotpath")]
 use crate::disk::FileWriter;
 use crate::disk::{self, DiskAPI as _, DiskStore, FileReader, MmapCopyStageMetrics, error::DiskError};
-use crate::erasure::coding::{BitrotReader, BitrotWriterWrapper, CustomWriter};
+use crate::erasure::coding::{BitrotReader, BitrotWriterWrapper, CustomWriter, ShardChunkRead};
 use bytes::Bytes;
 use rustfs_config::{
    DEFAULT_OBJECT_MMAP_READ_ENABLE, DEFAULT_OBJECT_MMAP_READ_MAX_LENGTH, ENV_OBJECT_MMAP_READ_ENABLE,
    ENV_OBJECT_MMAP_READ_MAX_LENGTH, ENV_OBJECT_ZERO_COPY_ENABLE,
 };
+use rustfs_rio::ChunkReaderBox;
 use rustfs_utils::HashAlgorithm;
 use std::future::Future;
 use std::io::{self, Cursor};
@@ -51,13 +52,25 @@ tokio::task_local! {
 /// (rustfs/backlog#1159). Everything else is a stream and keeps the old path.
 pub enum ShardReader {
    InMemory(Cursor<Bytes>),
+    Chunked(ChunkReaderBox),
    Stream(Box<dyn AsyncRead + Send + Sync + Unpin>),
 }

+#[cfg(test)]
+impl ShardReader {
+    pub(crate) fn inline_bytes(&self) -> Option<&Bytes> {
+        match self {
+            Self::InMemory(cursor) => Some(cursor.get_ref()),
+            Self::Chunked(_) | Self::Stream(_) => None,
+        }
+    }
+}
+
 impl AsyncRead for ShardReader {
    fn poll_read(self: Pin<&mut Self>, cx: &mut Context<'_>, buf: &mut tokio::io::ReadBuf<'_>) -> Poll<std::io::Result<()>> {
        match self.get_mut() {
            Self::InMemory(cursor) => Pin::new(cursor).poll_read(cx, buf),
+            Self::Chunked(reader) => Pin::new(&mut **reader).poll_read(cx, buf),
            Self::Stream(reader) => Pin::new(reader).poll_read(cx, buf),
        }
    }
@@ -67,7 +80,19 @@ impl crate::erasure::coding::ShardSource for ShardReader {
    fn try_take_block(&mut self, n: usize) -> Option<Bytes> {
        match self {
            Self::InMemory(cursor) => cursor.try_take_block(n),
-            Self::Stream(_) => None,
+            Self::Chunked(_) | Self::Stream(_) => None,
+        }
+    }
+
+    fn poll_read_chunk(self: Pin<&mut Self>, cx: &mut Context<'_>, max: usize) -> Poll<io::Result<ShardChunkRead>> {
+        let Self::Chunked(reader) = self.get_mut() else {
+            return Poll::Ready(Ok(ShardChunkRead::Unsupported));
+        };
+        match Pin::new(&mut **reader).poll_read_chunk(cx, max) {
+            Poll::Ready(Ok(Some(chunk))) => Poll::Ready(Ok(ShardChunkRead::Chunk(chunk))),
+            Poll::Ready(Ok(None)) => Poll::Ready(Ok(ShardChunkRead::Eof)),
+            Poll::Ready(Err(err)) => Poll::Ready(Err(err)),
+            Poll::Pending => Poll::Pending,
        }
    }
 }
@@ -120,26 +145,41 @@ struct BitrotReaderSource {

 impl BitrotReaderSource {
    async fn open(self) -> disk::error::Result<Option<BoxedObjectReader>> {
-        if let Some(data) = self.inline_data {
-            let mut rd = Cursor::new(data);
-            let offset = u64::try_from(self.offset).map_err(|_| DiskError::FileCorrupt)?;
-            rd.set_position(offset);
-            Ok(Some(ShardReader::InMemory(rd)))
-        } else if let Some(disk) = self.disk {
-            open_disk_reader(
-                &disk,
-                &self.bucket,
-                &self.path,
-                self.offset,
-                self.length,
-                self.use_mmap_read,
-                self.stage_metrics.map(|metrics| metrics.path),
-            )
+        open_reader_source(
+            self.inline_data,
+            self.disk.as_ref(),
+            &self.bucket,
+            &self.path,
+            self.offset,
+            self.length,
+            self.use_mmap_read,
+            self.stage_metrics.map(|metrics| metrics.path),
+        )
+        .await
+    }
+}
+
+#[allow(clippy::too_many_arguments)]
+async fn open_reader_source(
+    inline_data: Option<Bytes>,
+    disk: Option<&DiskStore>,
+    bucket: &str,
+    path: &str,
+    offset: usize,
+    length: usize,
+    use_mmap_read: bool,
+    metrics_path: Option<&'static str>,
+) -> disk::error::Result<Option<BoxedObjectReader>> {
+    if let Some(data) = inline_data {
+        let mut reader = Cursor::new(data);
+        reader.set_position(u64::try_from(offset).map_err(|_| DiskError::FileCorrupt)?);
+        Ok(Some(ShardReader::InMemory(reader)))
+    } else if let Some(disk) = disk {
+        open_disk_reader(disk, bucket, path, offset, length, use_mmap_read, metrics_path)
            .await
            .map(Some)
-        } else {
-            Ok(None)
-        }
+    } else {
+        Ok(None)
    }
 }

@@ -330,6 +370,17 @@ async fn open_disk_reader(
    let metrics_path = metrics_path.filter(|_| rustfs_io_metrics::get_stage_metrics_enabled());
    let stage_metrics_enabled = metrics_path.is_some();

+    // Preserve HTTP body ownership only on healthy remote reads. Instrumented
+    // and local paths retain their existing AsyncRead wrappers.
+    if use_mmap_read
+        && !disk.is_local()
+        && !stage_metrics_enabled
+        && !cfg!(feature = "hotpath")
+        && let Some(reader) = disk.read_file_stream_chunks(bucket, path, offset, length).await?
+    {
+        return Ok(ShardReader::Chunked(reader));
+    }
+
    // Mmap-copy materializes the whole `offset..offset+length` range as one
    // owned allocation before any byte is served, and GET/heal shard reads
    // request the entire part span in one call. Over-cap reads (e.g. a huge
@@ -605,7 +656,7 @@ pub async fn create_bitrot_reader_from_bytes(
 }

 #[allow(clippy::too_many_arguments)]
-async fn create_bitrot_reader_from_bytes_with_stage_metrics(
+pub(crate) async fn create_bitrot_reader_from_bytes_with_stage_metrics(
    inline_data: Option<Bytes>,
    disk: Option<&DiskStore>,
    bucket: &str,
@@ -623,22 +674,22 @@ async fn create_bitrot_reader_from_bytes_with_stage_metrics(

    let reader_construction_start = stage_metrics_enabled.then(Instant::now);
    let (offset, length) = bitrot_encoded_range(offset, length, shard_size, checksum_algo.clone());
-    let source = BitrotReaderSource {
-        inline_data,
-        disk: disk.cloned(),
-        bucket: bucket.to_string(),
-        path: path.to_string(),
-        offset,
-        length,
-        use_mmap_read,
-        stage_metrics,
-    };
    if let Some(metrics) = stage_metrics {
        record_get_stage_duration_if_enabled(metrics.path, metrics.reader_construction_stage, reader_construction_start);
    }

    let file_open_start = stage_metrics_enabled.then(Instant::now);
-    let reader = source.open().await?;
+    let reader = open_reader_source(
+        inline_data,
+        disk,
+        bucket,
+        path,
+        offset,
+        length,
+        use_mmap_read,
+        stage_metrics.map(|metrics| metrics.path),
+    )
+    .await?;
    if let Some(metrics) = stage_metrics {
        record_get_stage_duration_if_enabled(metrics.path, metrics.file_open_stage, file_open_start);
    }
@@ -698,11 +749,12 @@ pub(crate) fn create_deferred_bitrot_reader_with_stripe_handle(
 ) -> (BitrotReader<ShardReader>, DeferredReaderStripeHandle) {
    let stripe_stride = shard_size + checksum_algo.size();
    let (offset, length) = bitrot_encoded_range(offset, length, shard_size, checksum_algo.clone());
+    let inline_source = inline_data.is_some();
    let source = BitrotReaderSource {
        inline_data,
        disk,
-        bucket: bucket.to_string(),
-        path: path.to_string(),
+        bucket: if inline_source { String::new() } else { bucket.to_string() },
+        path: if inline_source { String::new() } else { path.to_string() },
        offset,
        length,
        use_mmap_read,
@@ -764,6 +816,50 @@ pub async fn create_bitrot_writer(
 #[cfg(test)]
 mod tests {
    use super::*;
+    use rustfs_rio::ChunkReader;
+    use std::collections::VecDeque;
+
+    struct TestChunkReader {
+        chunks: VecDeque<Bytes>,
+    }
+
+    impl TestChunkReader {
+        fn new(bytes: Bytes, fragment_sizes: &[usize]) -> Self {
+            let mut chunks = VecDeque::new();
+            let mut offset = 0;
+            for &size in fragment_sizes {
+                let end = (offset + size).min(bytes.len());
+                if offset < end {
+                    chunks.push_back(bytes.slice(offset..end));
+                }
+                offset = end;
+            }
+            if offset < bytes.len() {
+                chunks.push_back(bytes.slice(offset..));
+            }
+            Self { chunks }
+        }
+    }
+
+    impl AsyncRead for TestChunkReader {
+        fn poll_read(self: Pin<&mut Self>, _cx: &mut Context<'_>, _buf: &mut ReadBuf<'_>) -> Poll<io::Result<()>> {
+            Poll::Ready(Err(io::Error::other("test chunk reader must use chunk handoff")))
+        }
+    }
+
+    impl ChunkReader for TestChunkReader {
+        fn poll_read_chunk(mut self: Pin<&mut Self>, _cx: &mut Context<'_>, max: usize) -> Poll<io::Result<Option<Bytes>>> {
+            let Some(mut chunk) = self.chunks.pop_front() else {
+                return Poll::Ready(Ok(None));
+            };
+            let take = chunk.len().min(max);
+            if take < chunk.len() {
+                self.chunks.push_front(chunk.split_off(take));
+            }
+            chunk.truncate(take);
+            Poll::Ready(Ok(Some(chunk)))
+        }
+    }

    #[cfg(feature = "hotpath")]
    use crate::cluster::rpc::RemoteDisk;
@@ -1653,4 +1749,49 @@ mod tests {
        println!("error: {error:?}");
        assert_eq!(error, DiskError::DiskNotFound);
    }
+
+    #[tokio::test]
+    async fn shard_reader_chunked_path_verifies_fragmented_remote_block() {
+        const SHARD_SIZE: usize = 1024;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data = vec![42u8; SHARD_SIZE];
+        let mut encoded = Vec::new();
+        crate::erasure::coding::BitrotWriter::new(&mut encoded, SHARD_SIZE, algo.clone())
+            .write(&data)
+            .await
+            .expect("test shard should encode");
+
+        let source = TestChunkReader::new(Bytes::from(encoded), &[3, 7, 17, 31]);
+        let mut reader = BitrotReader::new(ShardReader::Chunked(Box::new(source)), SHARD_SIZE, algo, false);
+        let mut output = Vec::with_capacity(SHARD_SIZE);
+        reader
+            .read_appending(&mut output, SHARD_SIZE)
+            .await
+            .expect("fragmented remote shard should verify");
+
+        assert_eq!(output, data);
+    }
+
+    #[tokio::test]
+    async fn shard_reader_chunked_path_handles_more_than_one_poll_budget() {
+        const SHARD_SIZE: usize = 1024;
+        let algo = HashAlgorithm::HighwayHash256S;
+        let data = vec![42u8; SHARD_SIZE];
+        let mut encoded = Vec::new();
+        crate::erasure::coding::BitrotWriter::new(&mut encoded, SHARD_SIZE, algo.clone())
+            .write(&data)
+            .await
+            .expect("test shard should encode");
+
+        let fragment_sizes = vec![1; encoded.len()];
+        let source = TestChunkReader::new(Bytes::from(encoded), &fragment_sizes);
+        let mut reader = BitrotReader::new(ShardReader::Chunked(Box::new(source)), SHARD_SIZE, algo, false);
+        let mut output = Vec::with_capacity(SHARD_SIZE);
+        reader
+            .read_appending(&mut output, SHARD_SIZE)
+            .await
+            .expect("fragmented remote shard should verify after multiple polls");
+
+        assert_eq!(output, data);
+    }
 }
@@ -21,7 +21,8 @@ use tracing::debug;

 /// Supported set sizes this is used to find the optimal
 /// single set size.
-const SET_SIZES: [usize; 15] = [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16];
+pub(crate) const MAX_ERASURE_SET_DRIVE_COUNT: usize = 16;
+const SET_SIZES: [usize; 15] = [2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, MAX_ERASURE_SET_DRIVE_COUNT];
 const ENV_RUSTFS_ERASURE_SET_DRIVE_COUNT: &str = "RUSTFS_ERASURE_SET_DRIVE_COUNT";

 #[derive(Deserialize, Debug, Default)]
@@ -327,7 +328,7 @@ fn possible_set_counts(set_size: usize) -> Vec<usize> {

 /// checks whether given count is a valid set size for erasure coding.
 fn is_valid_set_size(count: usize) -> bool {
-    count >= SET_SIZES[0] && count <= SET_SIZES[SET_SIZES.len() - 1]
+    count >= SET_SIZES[0] && count <= MAX_ERASURE_SET_DRIVE_COUNT
 }

 /// Final set size with all the symmetry accounted for.
@@ -234,11 +234,17 @@ mod test {

    #[test]
    fn test_format_v1() {
+        // A freshly created format must survive a serialize -> parse roundtrip
+        // unchanged (identity on every on-disk field).
        let format = FormatV3::new(1, 4);
+        let serialized = serde_json::to_string(&format).expect("FormatV3 must serialize to JSON");
+        let reparsed = FormatV3::try_from(serialized.as_str()).expect("serialized FormatV3 must parse back");
+        assert_eq!(reparsed, format);

-        let str = serde_json::to_string(&format);
-        println!("{str:?}");
-
+        // minio-file-format-compat: this literal pins the on-disk format.json
+        // shape (erasure version "1", distributionAlgo "CRCMOD"). `this` always
+        // carries the disk's own UUID in real format.json files; a JSON null
+        // there was never parseable and never written by MinIO or RustFS.
        let data = r#"
        {
            "version": "1",
@@ -246,7 +252,7 @@ mod test {
            "id": "321b3874-987d-4c15-8fa5-757c956b1243",
            "xl": {
                "version": "1",
-                "this": null,
+                "this": "8ab9a908-f869-4f1f-8e42-eb067ffa7eb5",
                "sets": [
                    [
                        "8ab9a908-f869-4f1f-8e42-eb067ffa7eb5",
@@ -259,9 +265,23 @@ mod test {
            }
        }"#;

-        let p = FormatV3::try_from(data);
+        let parsed = FormatV3::try_from(data).expect("pinned v1 format.json literal must keep parsing");

-        println!("{p:?}");
+        assert_eq!(parsed.version, FormatMetaVersion::V1);
+        assert_eq!(parsed.format, FormatBackend::Erasure);
+        assert_eq!(
+            parsed.id,
+            Uuid::parse_str("321b3874-987d-4c15-8fa5-757c956b1243").expect("literal id is a valid UUID")
+        );
+        assert_eq!(parsed.erasure.version, FormatErasureVersion::V1);
+        assert_eq!(
+            parsed.erasure.this,
+            Uuid::parse_str("8ab9a908-f869-4f1f-8e42-eb067ffa7eb5").expect("literal this is a valid UUID")
+        );
+        assert_eq!(parsed.erasure.sets.len(), 1);
+        assert_eq!(parsed.erasure.sets[0].len(), 4);
+        assert_eq!(parsed.erasure.sets[0][0], parsed.erasure.this);
+        assert_eq!(parsed.erasure.distribution_algo, DistributionAlgoVersion::V1);
    }

    #[test]
@@ -211,6 +211,26 @@ impl ObjectLockConfigSnapshot {
    }
 }

+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+pub struct QuotaAdmission {
+    current_usage: u64,
+    quota_limit: u64,
+}
+
+impl QuotaAdmission {
+    pub(crate) fn current_usage(self) -> u64 {
+        self.current_usage
+    }
+
+    pub(crate) fn quota_limit(self) -> u64 {
+        self.quota_limit
+    }
+
+    pub(crate) fn remaining(self) -> u64 {
+        self.quota_limit - self.current_usage
+    }
+}
+
 #[derive(Debug, Default, Clone)]
 pub struct ObjectOptions {
    // Use the maximum parity (N/2), used when saving server configuration files
@@ -240,6 +260,9 @@ pub struct ObjectOptions {

    pub data_movement: bool,
    pub raw_data_movement_read: bool,
+    /// Materialize the data-movement per-part checksum sidecar for APIs that
+    /// return part checksums. Ordinary object reads leave it encoded.
+    pub include_part_checksums: bool,
    pub src_pool_idx: usize,
    pub user_defined: HashMap<String, String>,
    pub preserve_etag: Option<String>,
@@ -275,12 +298,22 @@ pub struct ObjectOptions {
    pub want_checksum: Option<Checksum>,
    pub skip_verify_bitrot: bool,
    pub capacity_scope_token: Option<Uuid>,
+    /// Server-derived bucket-quota snapshot for commit-boundary admission.
+    pub quota_admission: Option<QuotaAdmission>,
    /// Storage-owned journal writer used by the atomic delete path. This is
    /// populated only by the `ECStore` wrapper that holds the namespace locks.
    pub tier_delete_journal_api: Option<Arc<crate::store::ECStore>>,
 }

 impl ObjectOptions {
+    pub fn set_quota_admission(&mut self, current_usage: u64, quota_limit: u64) -> bool {
+        self.quota_admission = (current_usage <= quota_limit).then_some(QuotaAdmission {
+            current_usage,
+            quota_limit,
+        });
+        self.quota_admission.is_some()
+    }
+
    pub(crate) fn overwrites_existing_version(&self) -> bool {
        self.version_id.is_some() || !self.versioned || self.version_suspended
    }
@@ -164,6 +164,9 @@ pub(crate) async fn local_node_name() -> String {
 }

 pub(crate) async fn set_local_node_name(node_name: String) {
+    // Also stamp the internode-metrics server label: io-metrics is a leaf
+    // crate and no longer resolves node identity itself (backlog#1834).
+    rustfs_io_metrics::internode_metrics::set_internode_server_label(node_name.as_str());
    rustfs_common::set_global_local_node_name(&node_name).await;
 }

@@ -16,7 +16,7 @@ use super::meta::{
    clone_arc_by_index, ensure_valid_rebalance_pool_index, invalid_rebalance_pool_index_error,
    rebalance_metadata_not_initialized_error, should_ignore_rebalance_data_usage_cache,
 };
-use super::migration::migrate_entry_version;
+use super::migration::{RebalanceMigrationBackend, migrate_entry_version};
 use super::worker::{
    RebalanceEntryCleanupResult, RebalanceEntryTask, load_rebalance_bucket_configs, rebalance_max_attempts,
    resolve_rebalance_bucket_error, resolve_rebalance_entry_cleanup_delete_result, resolve_rebalance_file_info_versions_result,
@@ -144,6 +144,11 @@ impl ECStore {
            return Ok(RebalanceEntryOutcome::Completed);
        }

+        let bucket_incarnation_fence = match bucket_configs.bucket_incarnation_id {
+            Some(expected) => Some(self.acquire_bucket_incarnation_fence(&bucket, expected).await?),
+            None => None,
+        };
+
        let mut fivs =
            resolve_rebalance_file_info_versions_result(entry.file_info_versions(&bucket), bucket.as_str(), entry.name.as_str())?;

@@ -203,9 +208,14 @@ impl ECStore {
            }

            let version_id = version.version_id.map(|v| v.to_string());
+            let expected_bucket_incarnation_id = bucket_configs.bucket_incarnation_id;
            let mut transfer = |src_pool_idx: usize, bucket: String, rd: GetObjectReader| {
                let store = self.clone();
-                async move { store.rebalance_object(src_pool_idx, bucket, rd).await }
+                async move {
+                    store
+                        .rebalance_object(src_pool_idx, bucket, rd, expected_bucket_incarnation_id)
+                        .await
+                }
            };
            // Route delete-marker migration through the store layer so it lands on the
            // cross-pool target (excluding the source pool), not back onto the source set.
@@ -214,11 +224,12 @@ impl ECStore {
                async move { store.delete_object(&bucket, &object, opts).await }
            };
            let result = migrate_entry_version(
-                set.as_ref(),
+                &RebalanceMigrationBackend::new(set.as_ref(), self.as_ref()),
                bucket.clone(),
                pool_index,
                version,
                version_id.clone(),
+                expected_bucket_incarnation_id,
                rebalance_max_attempts(),
                should_ignore_rebalance_data_usage_cache(bucket.as_str()),
                &mut transfer,
@@ -303,6 +314,9 @@ impl ECStore {
        }

        if should_cleanup_rebalance_source_entry(rebalanced, fivs.versions.len(), expired) {
+            if bucket_incarnation_fence.as_ref().is_some_and(|guard| guard.is_lock_lost()) {
+                return Err(Error::other("rebalance bucket incarnation fence was lost before source cleanup"));
+            }
            let cleanup_result = self
                .finish_rebalance_entry_after_cleanup(
                    pool_index,
@@ -315,6 +329,12 @@ impl ECStore {
                        entry.name.as_str(),
                        &fivs,
                        &cleanup_preflight_allowed_missing,
+                        data_movement::SourceCleanupBucketFence {
+                            expected_incarnation_id: bucket_configs.bucket_incarnation_id,
+                            lifecycle_guard: bucket_incarnation_fence
+                                .as_ref()
+                                .and_then(|guard| guard.namespace_lock_guard()),
+                        },
                        "rebalance",
                    ),
                )
@@ -389,8 +409,14 @@ impl ECStore {
    }

    #[tracing::instrument(skip(self, rd))]
-    async fn rebalance_object(self: Arc<Self>, pool_idx: usize, bucket: String, rd: GetObjectReader) -> Result<()> {
-        data_movement::migrate_object(self, pool_idx, bucket, rd, "rebalance_object").await
+    async fn rebalance_object(
+        self: Arc<Self>,
+        pool_idx: usize,
+        bucket: String,
+        rd: GetObjectReader,
+        expected_bucket_incarnation_id: Option<uuid::Uuid>,
+    ) -> Result<()> {
+        data_movement::migrate_object(self, pool_idx, bucket, rd, expected_bucket_incarnation_id, "rebalance_object").await
    }

    async fn update_rebalance_last_error(&self, pool_idx: usize, message: String) -> Result<()> {
@@ -5,6 +5,7 @@ use crate::error::{Error, Result, is_err_object_not_found, is_err_version_not_fo
 use crate::object_api::{GetObjectReader, ObjectInfo, ObjectOptions};
 use crate::set_disk::SetDisks;
 use crate::storage_api_contracts::{object::ObjectIO, range::HTTPRangeSpec};
+use crate::store::ECStore;
 use http::HeaderMap;
 use rustfs_filemeta::FileInfo;
 use rustfs_utils::path::encode_dir_object;
@@ -21,15 +22,23 @@ pub(crate) struct MigrationVersionResult {
    pub error: Option<Error>,
 }

-pub(super) fn rebalance_delete_marker_opts(version: &FileInfo, version_id: Option<String>, src_pool_idx: usize) -> ObjectOptions {
+pub(super) fn rebalance_delete_marker_opts(
+    version: &FileInfo,
+    version_id: Option<String>,
+    src_pool_idx: usize,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
+) -> ObjectOptions {
+    let version_suspended = version.version_id.is_none() && version_id.is_none();
    ObjectOptions {
-        versioned: true,
-        version_id,
+        versioned: !version_suspended,
+        version_suspended,
+        version_id: version_id.or_else(|| version_suspended.then(|| uuid::Uuid::nil().to_string())),
        mod_time: version.mod_time,
        src_pool_idx,
        data_movement: true,
        delete_marker: true,
        skip_decommissioned: true,
+        expected_bucket_incarnation_id,
        delete_replication: version
            .replication_state_internal
            .as_ref()
@@ -38,7 +47,12 @@ pub(super) fn rebalance_delete_marker_opts(version: &FileInfo, version_id: Optio
    }
 }

-fn rebalance_remote_tiered_opts(version: &FileInfo, version_id: Option<String>, src_pool_idx: usize) -> ObjectOptions {
+fn rebalance_remote_tiered_opts(
+    version: &FileInfo,
+    version_id: Option<String>,
+    src_pool_idx: usize,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
+) -> ObjectOptions {
    ObjectOptions {
        versioned: version_id.is_some(),
        version_id,
@@ -46,6 +60,21 @@ fn rebalance_remote_tiered_opts(version: &FileInfo, version_id: Option<String>,
        user_defined: version.metadata.clone(),
        src_pool_idx,
        data_movement: true,
+        include_part_checksums: true,
+        http_preconditions: Some(crate::data_movement::data_movement_target_precondition()),
+        expected_bucket_incarnation_id,
+        ..Default::default()
+    }
+}
+
+pub(super) fn rebalance_object_migration_read_opts(version_id: Option<String>) -> ObjectOptions {
+    ObjectOptions {
+        version_id,
+        no_lock: true,
+        data_movement: true,
+        raw_data_movement_read: true,
+        skip_decommissioned: true,
+        skip_rebalancing: true,
        ..Default::default()
    }
 }
@@ -70,8 +99,19 @@ pub(crate) trait MigrationBackend: Send + Sync {
    ) -> Result<()>;
 }

+pub(crate) struct RebalanceMigrationBackend<'a> {
+    source: &'a SetDisks,
+    store: &'a ECStore,
+}
+
+impl<'a> RebalanceMigrationBackend<'a> {
+    pub(crate) fn new(source: &'a SetDisks, store: &'a ECStore) -> Self {
+        Self { source, store }
+    }
+}
+
 #[async_trait::async_trait]
-impl MigrationBackend for SetDisks {
+impl MigrationBackend for RebalanceMigrationBackend<'_> {
    async fn get_object_reader_for_migration(
        &self,
        bucket: &str,
@@ -80,7 +120,7 @@ impl MigrationBackend for SetDisks {
        h: HeaderMap,
        opts: &ObjectOptions,
    ) -> Result<GetObjectReader> {
-        self.get_object_reader(bucket, object, range, h, opts).await
+        self.source.get_object_reader(bucket, object, range, h, opts).await
    }

    async fn move_remote_version_for_migration(
@@ -90,7 +130,7 @@ impl MigrationBackend for SetDisks {
        fi: &FileInfo,
        opts: &ObjectOptions,
    ) -> Result<()> {
-        self.decommission_tiered_object(bucket, object, fi, opts).await
+        self.store.decommission_tiered_object(bucket, object, fi, opts).await
    }
 }

@@ -101,6 +141,7 @@ pub(crate) async fn migrate_entry_version<Backend, F, Fut, D, DFut>(
    pool_index: usize,
    version: &FileInfo,
    version_id: Option<String>,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
    max_attempts: usize,
    ignore_data_usage_cache: bool,
    transfer: F,
@@ -113,12 +154,13 @@ where
    D: FnMut(String, String, ObjectOptions) -> DFut + Send,
    DFut: Future<Output = Result<ObjectInfo>> + Send,
 {
-    migrate_entry_version_with_retry_wait(
+    migrate_entry_version_with_retry_wait_and_incarnation(
        set,
        bucket,
        pool_index,
        version,
        version_id,
+        expected_bucket_incarnation_id,
        max_attempts,
        ignore_data_usage_cache,
        transfer,
@@ -137,6 +179,45 @@ pub(super) async fn migrate_entry_version_with_retry_wait<Backend, F, Fut, D, DF
    version_id: Option<String>,
    max_attempts: usize,
    ignore_data_usage_cache: bool,
+    transfer: F,
+    delete_marker: D,
+    wait_retry: W,
+) -> MigrationVersionResult
+where
+    Backend: MigrationBackend + ?Sized,
+    F: FnMut(usize, String, GetObjectReader) -> Fut + Send,
+    Fut: Future<Output = Result<()>> + Send,
+    D: FnMut(String, String, ObjectOptions) -> DFut + Send,
+    DFut: Future<Output = Result<ObjectInfo>> + Send,
+    W: FnMut(Duration) -> WFut + Send,
+    WFut: Future<Output = ()> + Send,
+{
+    migrate_entry_version_with_retry_wait_and_incarnation(
+        set,
+        bucket,
+        pool_index,
+        version,
+        version_id,
+        None,
+        max_attempts,
+        ignore_data_usage_cache,
+        transfer,
+        delete_marker,
+        wait_retry,
+    )
+    .await
+}
+
+#[allow(clippy::too_many_arguments)]
+async fn migrate_entry_version_with_retry_wait_and_incarnation<Backend, F, Fut, D, DFut, W, WFut>(
+    set: &Backend,
+    bucket: String,
+    pool_index: usize,
+    version: &FileInfo,
+    version_id: Option<String>,
+    expected_bucket_incarnation_id: Option<uuid::Uuid>,
+    max_attempts: usize,
+    ignore_data_usage_cache: bool,
    mut transfer: F,
    mut delete_marker: D,
    mut wait_retry: W,
@@ -169,7 +250,7 @@ where
                &bucket,
                &version.name,
                version,
-                &rebalance_remote_tiered_opts(version, version_id, pool_index),
+                &rebalance_remote_tiered_opts(version, version_id, pool_index, expected_bucket_incarnation_id),
            )
            .await
        {
@@ -212,7 +293,7 @@ where
        if let Err(err) = delete_marker(
            bucket.clone(),
            version.name.clone(),
-            rebalance_delete_marker_opts(version, version_id, pool_index),
+            rebalance_delete_marker_opts(version, version_id, pool_index, expected_bucket_incarnation_id),
        )
        .await
        {
@@ -255,11 +336,7 @@ where
                &encode_dir_object(&version.name),
                None,
                HeaderMap::new(),
-                &ObjectOptions {
-                    version_id: version_id.clone(),
-                    no_lock: true,
-                    ..Default::default()
-                },
+                &rebalance_object_migration_read_opts(version_id.clone()),
            )
            .await
        {
@@ -113,6 +113,8 @@ struct LegacyRebalanceMeta {
 struct MigrationBackendSpy {
    get_object_reader: Mutex<Option<core::result::Result<GetObjectReader, Error>>>,
    move_remote: Mutex<Option<core::result::Result<(), Error>>>,
+    get_opts: Mutex<Vec<ObjectOptions>>,
+    move_remote_opts: Mutex<Vec<ObjectOptions>>,
    get_calls: AtomicUsize,
    move_remote_calls: AtomicUsize,
 }
@@ -125,6 +127,8 @@ impl MigrationBackendSpy {
        Self {
            get_object_reader: Mutex::new(get_object_reader),
            move_remote: Mutex::new(move_remote),
+            get_opts: Mutex::new(Vec::new()),
+            move_remote_opts: Mutex::new(Vec::new()),
            get_calls: AtomicUsize::new(0),
            move_remote_calls: AtomicUsize::new(0),
        }
@@ -138,6 +142,24 @@ impl MigrationBackendSpy {
        self.move_remote_calls.load(Ordering::SeqCst)
    }

+    fn last_get_opts(&self) -> ObjectOptions {
+        self.get_opts
+            .lock()
+            .unwrap()
+            .last()
+            .cloned()
+            .expect("reader opts should be captured")
+    }
+
+    fn last_move_remote_opts(&self) -> ObjectOptions {
+        self.move_remote_opts
+            .lock()
+            .unwrap()
+            .last()
+            .cloned()
+            .expect("remote opts should be captured")
+    }
+
    fn make_reader() -> GetObjectReader {
        GetObjectReader {
            stream: Box::new(Cursor::new(vec![0_u8; 3])),
@@ -156,9 +178,10 @@ impl MigrationBackend for MigrationBackendSpy {
        _object: &str,
        _range: Option<HTTPRangeSpec>,
        _h: http::HeaderMap,
-        _opts: &ObjectOptions,
+        opts: &ObjectOptions,
    ) -> Result<GetObjectReader> {
        self.get_calls.fetch_add(1, Ordering::SeqCst);
+        self.get_opts.lock().unwrap().push(opts.clone());
        if let Some(result) = self.get_object_reader.lock().unwrap().take() {
            return result;
        }
@@ -171,9 +194,10 @@ impl MigrationBackend for MigrationBackendSpy {
        _bucket: &str,
        _object: &str,
        _fi: &FileInfo,
-        _opts: &ObjectOptions,
+        opts: &ObjectOptions,
    ) -> Result<()> {
        self.move_remote_calls.fetch_add(1, Ordering::SeqCst);
+        self.move_remote_opts.lock().unwrap().push(opts.clone());
        if let Some(result) = self.move_remote.lock().unwrap().take() {
            return result;
        }
@@ -217,7 +241,8 @@ fn test_rebalance_delete_marker_opts_preserves_replication_state() {
        ..version_deleted()
    };

-    let opts = rebalance_delete_marker_opts(&version, Some("version-id".to_string()), 7);
+    let incarnation = uuid::Uuid::new_v4();
+    let opts = rebalance_delete_marker_opts(&version, Some("version-id".to_string()), 7, Some(incarnation));
    let replication = opts.delete_replication.expect("replication state should be preserved");

    assert!(opts.versioned);
@@ -227,11 +252,22 @@ fn test_rebalance_delete_marker_opts_preserves_replication_state() {
    assert_eq!(opts.src_pool_idx, 7);
    assert_eq!(opts.version_id.as_deref(), Some("version-id"));
    assert_eq!(opts.mod_time, Some(mod_time));
+    assert_eq!(opts.expected_bucket_incarnation_id, Some(incarnation));
    assert_eq!(replication.replica_status, ReplicationStatusType::Replica);
    assert!(replication.delete_marker);
    assert_eq!(replication.replicate_decision_str, "existing");
 }

+#[test]
+fn test_rebalance_delete_marker_opts_preserves_suspended_null_version() {
+    let version = version_deleted();
+    let opts = rebalance_delete_marker_opts(&version, None, 7, None);
+
+    assert!(!opts.versioned);
+    assert!(opts.version_suspended);
+    assert_eq!(opts.version_id.as_deref(), Some(uuid::Uuid::nil().to_string().as_str()));
+}
+
 #[tokio::test]
 async fn test_migrate_entry_version_remote_version_is_moved_without_transfer() {
    let backend = MigrationBackendSpy::new(None, Some(Ok(())));
@@ -248,12 +284,14 @@ async fn test_migrate_entry_version_remote_version_is_moved_without_transfer() {
        }
    };

+    let incarnation = uuid::Uuid::new_v4();
    let result = migrate_entry_version(
        &backend,
        "bucket".to_string(),
        0,
        &version,
        version.version_id.map(|v| v.to_string()),
+        Some(incarnation),
        3,
        false,
        &mut transfer,
@@ -269,6 +307,10 @@ async fn test_migrate_entry_version_remote_version_is_moved_without_transfer() {
    assert_eq!(transfer_count.load(Ordering::SeqCst), 0);
    assert_eq!(backend.move_remote_calls(), 1);
    assert_eq!(backend.get_calls(), 0);
+    let remote_opts = backend.last_move_remote_opts();
+    assert!(remote_opts.include_part_checksums);
+    assert!(remote_opts.http_preconditions.is_some());
+    assert_eq!(remote_opts.expected_bucket_incarnation_id, Some(incarnation));
 }

 #[tokio::test]
@@ -294,6 +336,7 @@ async fn test_migrate_entry_version_remote_not_found_is_cleanup_ignored() {
        0,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        3,
        false,
        &mut transfer,
@@ -330,6 +373,7 @@ async fn test_migrate_entry_version_remote_overwrite_is_not_ignored() {
        0,
        &version,
        Some("vid-1".to_string()),
+        None,
        3,
        false,
        &mut transfer,
@@ -368,6 +412,7 @@ async fn test_migrate_entry_version_remote_failure_is_reported() {
        0,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        3,
        false,
        &mut transfer,
@@ -410,6 +455,7 @@ async fn test_migrate_entry_version_deleted_version_routes_delete_through_store_
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        3,
        false,
        &mut transfer,
@@ -449,6 +495,7 @@ async fn test_migrate_entry_version_deleted_version_not_found_is_ignored() {
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        3,
        false,
        &mut transfer,
@@ -491,6 +538,7 @@ async fn test_migrate_entry_version_deleted_version_overwrite_is_not_ignored() {
        1,
        &version,
        Some("vid-1".to_string()),
+        None,
        3,
        false,
        &mut transfer,
@@ -520,6 +568,7 @@ async fn test_migrate_entry_version_reader_not_found_is_ignored() {
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        3,
        false,
        &mut transfer,
@@ -647,6 +696,7 @@ async fn test_migrate_entry_version_reader_fails_after_retries() {
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        3,
        false,
        &mut transfer,
@@ -685,6 +735,7 @@ async fn test_migrate_entry_version_zero_max_attempts_still_attempts_once() {
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        0,
        false,
        &mut transfer,
@@ -750,6 +801,13 @@ async fn test_migrate_entry_version_transfer_retries_before_success() {
    assert_eq!(backend.get_calls(), 2);
    assert_eq!(transfer_count.load(Ordering::SeqCst), 2);
    assert_eq!(wait_count.load(Ordering::SeqCst), 1);
+    let read_opts = backend.last_get_opts();
+    assert_eq!(read_opts.version_id.as_deref(), version.version_id.map(|id| id.to_string()).as_deref());
+    assert!(read_opts.no_lock);
+    assert!(read_opts.data_movement);
+    assert!(read_opts.raw_data_movement_read);
+    assert!(read_opts.skip_decommissioned);
+    assert!(read_opts.skip_rebalancing);
 }

 #[tokio::test]
@@ -822,6 +880,7 @@ async fn test_migrate_entry_version_transfer_fails_after_retries() {
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        2,
        false,
        &mut transfer,
@@ -860,6 +919,7 @@ async fn test_migrate_entry_version_transfer_not_found_is_ignored() {
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        3,
        false,
        &mut transfer,
@@ -901,6 +961,7 @@ async fn test_migrate_entry_version_transfer_overwrite_is_not_ignored() {
        1,
        &version,
        Some("vid-1".to_string()),
+        None,
        3,
        false,
        &mut transfer,
@@ -943,6 +1004,7 @@ async fn test_migrate_entry_version_ignores_data_usage_cache_when_enabled() {
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        2,
        true,
        &mut transfer,
@@ -985,6 +1047,7 @@ async fn test_migrate_entry_version_data_usage_cache_moves_when_ignore_disabled(
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        2,
        false,
        &mut transfer,
@@ -2026,6 +2089,7 @@ async fn test_migrate_entry_version_transfer_failure_reports_write_target_stage(
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        1,
        false,
        &mut transfer,
@@ -2050,6 +2114,7 @@ async fn test_migrate_entry_version_reader_failure_reports_read_source_stage() {
        1,
        &version,
        version.version_id.map(|v| v.to_string()),
+        None,
        1,
        false,
        &mut transfer,
@@ -36,6 +36,7 @@ pub type RStats = Vec<Arc<RebalanceStats>>;

 #[derive(Debug, Default)]
 pub(super) struct RebalanceBucketConfigs {
+    pub(super) bucket_incarnation_id: Option<uuid::Uuid>,
    pub(super) lifecycle_config: Option<s3s::dto::BucketLifecycleConfiguration>,
    pub(super) object_lock_config: Option<s3s::dto::ObjectLockConfiguration>,
    pub(super) replication_config: Option<(s3s::dto::ReplicationConfiguration, OffsetDateTime)>,
@@ -406,6 +406,7 @@ pub(super) async fn load_rebalance_bucket_configs(api: &ECStore, bucket: &str) -

    let expiry_configs = crate::bucket::lifecycle::get_expiry_configs(api, bucket).await?;
    Ok(RebalanceBucketConfigs {
+        bucket_incarnation_id: Some(api.bucket_incarnation_id_from_disk(bucket).await?),
        lifecycle_config: expiry_configs.lifecycle.map(|config| (*config).clone()),
        object_lock_config: expiry_configs.object_lock.map(|config| (*config).clone()),
        replication_config: resolve_rebalance_optional_bucket_config_result(
@@ -54,10 +54,12 @@ use crate::disk::{
 use crate::erasure::coding::BitrotReader;
 use crate::io_support::bitrot::ShardReader;
 use crate::io_support::bitrot::{
-    BitrotReaderStageMetrics, DeferredReaderStripeHandle, adjust_shard_read_params, create_bitrot_reader_with_stage_metrics,
-    create_deferred_bitrot_reader_with_stripe_handle, object_mmap_read_enabled, object_mmap_read_max_length,
+    BitrotReaderStageMetrics, DeferredReaderStripeHandle, adjust_shard_read_params,
+    create_bitrot_reader_from_bytes_with_stage_metrics, create_deferred_bitrot_reader_with_stripe_handle,
+    object_mmap_read_enabled, object_mmap_read_max_length,
 };
 use crate::set_disk::shard_source::ShardReadCost;
+use futures::FutureExt as _;
 use futures::stream::{FuturesUnordered, StreamExt};
 use metrics::counter;
 use std::{
@@ -221,7 +223,7 @@ impl MetadataFanoutDiagnostics {
        self.observations.iter().filter(|observation| observation.ignored).count()
    }

-    pub(in crate::set_disk) fn error_responses(&self) -> usize {
+    pub(in crate::set_disk) fn non_valid_responses(&self) -> usize {
        self.total_responses().saturating_sub(self.valid_responses())
    }

@@ -272,7 +274,7 @@ impl MetadataFanoutDiagnostics {
            self.total_responses(),
            self.valid_responses(),
            self.ignored_responses(),
-            self.error_responses(),
+            self.non_valid_responses(),
        );
        for observation in &self.observations {
            rustfs_io_metrics::record_get_object_metadata_response(path, observation.outcome);
@@ -1261,13 +1263,13 @@ pub(in crate::set_disk) fn schedule_bitrot_reader_task<'a>(
        return;
    }

-    let inline_data = files[idx].data.as_deref();
+    let inline_data = files[idx].data.clone();
    let data_dir = files[idx].data_dir.unwrap_or_default();
    let disk = disks[idx].as_ref();
    let path = format!("{object}/{data_dir}/part.{part_number}");

    reader_tasks.push(Box::pin(async move {
-        let result = create_bitrot_reader_with_stage_metrics(
+        let result = create_bitrot_reader_from_bytes_with_stage_metrics(
            inline_data,
            disk,
            bucket,
@@ -1559,14 +1561,14 @@ pub(in crate::set_disk) async fn create_bitrot_readers_until_quorum_all_shards(
    let schedule_stage_start = stage_metrics.map(|_| Instant::now());
    for (idx, disk_op) in disks.iter().enumerate() {
        setup.mark_scheduled(idx);
-        let inline_data = files[idx].data.as_deref();
+        let inline_data = files[idx].data.clone();
        let data_dir = files[idx].data_dir.unwrap_or_default();
        let disk = disk_op.as_ref();
        let path = format!("{object}/{data_dir}/part.{part_number}");
        let checksum_algo = checksum_algo.clone();

        reader_tasks.push(async move {
-            let result = create_bitrot_reader_with_stage_metrics(
+            let result = create_bitrot_reader_from_bytes_with_stage_metrics(
                inline_data,
                disk,
                bucket,
@@ -2222,18 +2224,18 @@ impl SetDisks {
        let mut ress = Vec::with_capacity(disks.len());
        let mut errors = Vec::with_capacity(disks.len());
        let mut observations = observe.then(|| Vec::with_capacity(disks.len()));
-        let opts = Arc::new(ReadOptions {
+        let opts = ReadOptions {
            incl_free_versions,
            read_data,
            healing,
-        });
-        let org_bucket = Arc::new(org_bucket.to_string());
-        let bucket = Arc::new(bucket.to_string());
-        let object = Arc::new(object.to_string());
-        let version_id = Arc::new(version_id.to_string());
+        };
+        let org_bucket: Arc<str> = Arc::from(org_bucket);
+        let bucket: Arc<str> = Arc::from(bucket);
+        let object: Arc<str> = Arc::from(object);
+        let version_id: Arc<str> = Arc::from(version_id);
        let futures = disks.iter().enumerate().map(|(disk_index, disk)| {
            let disk = disk.clone();
-            let opts = opts.clone();
+            let task_opts = opts;
            let org_bucket = org_bucket.clone();
            let bucket = bucket.clone();
            let object = object.clone();
@@ -2242,7 +2244,8 @@ impl SetDisks {
                let response_start = observe.then(Instant::now);
                let result = if let Some(disk) = disk {
                    Self::record_read_version_call(&object, disk_index);
-                    disk.read_version(&org_bucket, &bucket, &object, &version_id, &opts).await
+                    disk.read_version(&org_bucket, &bucket, &object, &version_id, &task_opts)
+                        .await
                } else {
                    Err(DiskError::DiskNotFound)
                };
@@ -2307,21 +2310,21 @@ impl SetDisks {
        let mut observations = Vec::with_capacity(disks.len());
        let mut accumulator =
            MetadataQuorumAccumulator::new(disks.len(), default_parity_count, true).with_requested_version_id(version_id);
-        let opts = Arc::new(ReadOptions {
+        let opts = ReadOptions {
            incl_free_versions,
            read_data,
            healing,
-        });
-        let org_bucket = Arc::new(org_bucket.to_string());
-        let bucket = Arc::new(bucket.to_string());
-        let object = Arc::new(object.to_string());
-        let version_id = Arc::new(version_id.to_string());
+        };
+        let org_bucket: Arc<str> = Arc::from(org_bucket);
+        let bucket: Arc<str> = Arc::from(bucket);
+        let object: Arc<str> = Arc::from(object);
+        let version_id: Arc<str> = Arc::from(version_id);
        let mut join_set = JoinSet::new();
        let bounded_fanout = is_get_metadata_early_stop_bounded_fanout_enabled();
        let mut next_disk_index = 0usize;
        let spawn_read_version =
            |join_set: &mut JoinSet<(usize, disk::error::Result<FileInfo>, Duration)>, index: usize, disk: Option<DiskStore>| {
-                let opts = opts.clone();
+                let task_opts = opts;
                let org_bucket = org_bucket.clone();
                let bucket = bucket.clone();
                let object = object.clone();
@@ -2332,7 +2335,8 @@ impl SetDisks {
                        Self::record_read_version_call(&object, index);
                        #[cfg(test)]
                        Self::read_version_fanout_barrier(&object, index).await;
-                        disk.read_version(&org_bucket, &bucket, &object, &version_id, &opts).await
+                        disk.read_version(&org_bucket, &bucket, &object, &version_id, &task_opts)
+                            .await
                    } else {
                        Err(DiskError::DiskNotFound)
                    };
@@ -2440,13 +2444,14 @@ impl SetDisks {
        bucket: &str,
        object: &str,
    ) -> Result<Option<rustfs_filemeta::FileInfoVersions>> {
+        let disk_object = rustfs_utils::path::encode_dir_object(object);
        let disks = self.get_disks_internal().await;
        if disks.is_empty() {
            return Err(to_object_err(StorageError::ErasureReadQuorum, vec![bucket, object]));
        }

        let read_quorum = disks.len().div_ceil(2).max(1);
-        let (raw_fileinfos, errs) = Self::read_all_raw_file_info(&disks, bucket, object, false).await;
+        let (raw_fileinfos, errs) = Self::read_all_raw_file_info(&disks, bucket, disk_object.as_str(), false).await;

        if let Some(err) = reduce_read_quorum_errs(&errs, OBJECT_OP_IGNORED_ERRS, read_quorum) {
            let object_err = to_object_err(err.into(), vec![bucket, object]);
@@ -2602,7 +2607,7 @@ impl SetDisks {
        //
        // `into_fileinfo` with an empty version_id selects the first non-free version
        // (see FileMeta::into_fileinfo); replicate that selection from the header here.
-        let vid = match meta.into_fileinfo(bucket, object, "", true, incl_free_vers, true) {
+        let vid = match meta.into_fileinfo_without_part_checksums(bucket, object, "", true, incl_free_vers) {
            Ok(finfo) if file_info_is_valid_for_metadata(&finfo) => finfo.version_id.unwrap_or(Uuid::nil()),
            _ => match meta
                .versions
@@ -2625,7 +2630,13 @@ impl SetDisks {

        for (idx, meta_op) in metadata_array.iter().enumerate() {
            if let Some(meta) = meta_op {
-                match meta.into_fileinfo(bucket, object, vid.to_string().as_str(), read_data, incl_free_vers, true) {
+                match meta.into_fileinfo_without_part_checksums(
+                    bucket,
+                    object,
+                    vid.to_string().as_str(),
+                    read_data,
+                    incl_free_vers,
+                ) {
                    Ok(res) => match res.validate_for_metadata_read() {
                        Ok(_) => meta_file_infos[idx] = res,
                        Err(err) => errs[idx] = Some(err.into()),
@@ -2854,8 +2865,6 @@ impl SetDisks {
                file_info.validate_for_erasure_write()?;
            }
        }
-        let mut futures = Vec::with_capacity(disks.len());
-
        let mut errs = Vec::with_capacity(disks.len());

        let src_bucket = Arc::new(src_bucket.to_string());
@@ -2863,48 +2872,65 @@ impl SetDisks {
        let dst_bucket = Arc::new(dst_bucket.to_string());
        let dst_object = Arc::new(dst_object.to_string());

-        for (i, (disk, file_info)) in disks.iter().zip(file_infos.iter()).enumerate() {
-            let mut file_info = file_info.clone();
-            let disk = disk.clone();
-            let src_bucket = src_bucket.clone();
-            let src_object = src_object.clone();
-            let dst_object = dst_object.clone();
-            let dst_bucket = dst_bucket.clone();
+        let disk_count = disks.len();
+        let fanout_disks = disks.to_vec();
+        let fanout_file_infos = file_infos.to_vec();
+        let fanout_src_bucket = src_bucket.clone();
+        let fanout_src_object = src_object.clone();
+        let fanout_dst_bucket = dst_bucket.clone();
+        let fanout_dst_object = dst_object.clone();
+        // Keep one coordinator task so a cancelled caller cannot drop partially
+        // completed disk mutations. Per-disk futures stay ordered in `join_all`,
+        // preserving slot-indexed quorum and convergence accounting without a
+        // scheduler task for every disk.
+        let fanout = tokio::spawn(async move {
+            let futures = fanout_disks
+                .into_iter()
+                .zip(fanout_file_infos)
+                .enumerate()
+                .map(|(i, (disk, mut file_info))| {
+                    let src_bucket = fanout_src_bucket.clone();
+                    let src_object = fanout_src_object.clone();
+                    let dst_object = fanout_dst_object.clone();
+                    let dst_bucket = fanout_dst_bucket.clone();

-            futures.push(tokio::spawn(async move {
-                // Test-only introspection guard: counts this task as in-flight for
-                // the whole body. Compiles to `()` in production (no behavior).
-                #[allow(clippy::let_unit_value)]
-                let _fanout_task_guard = Self::rename_fanout_task_guard(&dst_object);
+                    std::panic::AssertUnwindSafe(async move {
+                        // Test-only introspection guard: counts this operation as
+                        // in-flight for the whole body. Compiles to `()` in production.
+                        #[allow(clippy::let_unit_value)]
+                        let _fanout_task_guard = Self::rename_fanout_task_guard(&dst_object);

-                let Some(disk) = disk else {
-                    return Err(DiskError::DiskNotFound);
-                };
+                        let Some(disk) = disk else {
+                            return Err(DiskError::DiskNotFound);
+                        };

-                let is_delete_marker = file_info.is_canonical_delete_marker();
-                if file_info.erasure.index == 0 {
-                    file_info.erasure.index = i + 1;
-                }
+                        let is_delete_marker = file_info.is_canonical_delete_marker();
+                        if file_info.erasure.index == 0 {
+                            file_info.erasure.index = i + 1;
+                        }

-                if !is_delete_marker && !file_info.has_valid_erasure_geometry() {
-                    return Err(DiskError::FileCorrupt);
-                }
+                        if !is_delete_marker && !file_info.has_valid_erasure_geometry() {
+                            return Err(DiskError::FileCorrupt);
+                        }

-                // Test-only awaitable pause point right before the disk rename.
-                // A no-op immediately-ready future in production.
-                Self::rename_fanout_barrier(&dst_object, i, rename_fanout_barrier_phase::RENAME).await;
+                        // Test-only awaitable pause point right before the disk rename.
+                        // A no-op immediately-ready future in production.
+                        Self::rename_fanout_barrier(&dst_object, i, rename_fanout_barrier_phase::RENAME).await;

-                disk.rename_data(&src_bucket, &src_object, file_info, &dst_bucket, &dst_object)
-                    .await
-            }));
-        }
+                        disk.rename_data(&src_bucket, &src_object, file_info, &dst_bucket, &dst_object)
+                            .await
+                    })
+                    .catch_unwind()
+                });
+            join_all(futures).await
+        });

-        let mut disk_versions = vec![None; disks.len()];
-        let mut data_dirs = vec![None; disks.len()];
-        let mut cleanup_data_dirs = vec![None; disks.len()];
-        let mut old_current_sizes = vec![None; disks.len()];
+        let mut disk_versions = vec![None; disk_count];
+        let mut data_dirs = vec![None; disk_count];
+        let mut cleanup_data_dirs = vec![None; disk_count];
+        let mut old_current_sizes = vec![None; disk_count];

-        let results = join_all(futures).await;
+        let results = fanout.await.map_err(|_| DiskError::Unexpected)?;

        for (idx, result) in results.iter().enumerate() {
            match result.as_ref().map_err(|_| DiskError::Unexpected)? {
@@ -4515,26 +4541,29 @@ impl SetDisks {
        object: &str,
        opts: &ObjectOptions,
    ) -> Option<StorageError> {
-        let mut opts = opts.clone();
+        let mut lookup_opts = opts.clone();

-        let http_preconditions = opts.http_preconditions?;
-        opts.http_preconditions = None;
+        let http_preconditions = lookup_opts.http_preconditions?;
+        lookup_opts.http_preconditions = None;

        // Never claim a lock here, to avoid deadlock
        // - If no_lock is false, we must have obtained the lock out side of this function
        // - If no_lock is true, we should not obtain locks
-        opts.no_lock = true;
-        let oi = self.get_object_info(bucket, object, &opts).await;
+        lookup_opts.no_lock = true;
+        let oi = self.get_object_info(bucket, object, &lookup_opts).await;

        match oi {
            Ok(oi) => {
-                // If top level is a delete marker proceed to upload.
+                // Ordinary writes may proceed past a top-level delete marker;
+                // data movement must not replace an acknowledged deletion.
                if oi.delete_marker {
-                    return None;
+                    return opts.data_movement.then_some(StorageError::PreconditionFailed);
                }
                let if_none_match = http_preconditions.if_none_match_value().map(str::to_owned);
                let if_match = http_preconditions.if_match_value().map(str::to_owned);
-                if should_prevent_write(&oi, if_none_match, if_match) {
+                if should_prevent_write(&oi, if_none_match, if_match)
+                    && !crate::data_movement::can_replace_stale_data_movement_target(&oi, opts)
+                {
                    return Some(StorageError::PreconditionFailed);
                }
            }
@@ -5865,6 +5894,51 @@ mod tests {
        drop(dirs);
    }

+    #[tokio::test]
+    async fn rename_fanout_drains_after_caller_cancellation() {
+        const DISKS: usize = 4;
+        let bucket = "rename-cancel-bucket";
+        let object = "rename-cancel-object";
+        let (dirs, disks) = call_counter_local_disks(bucket, DISKS).await;
+        let marker = metadata_test_delete_marker(object, Uuid::new_v4(), OffsetDateTime::now_utc());
+        let file_infos = vec![marker; DISKS];
+        let tracker = rename_fanout_barrier::observe_tasks(object);
+        let barrier = rename_fanout_barrier::arm(object, 0, rename_fanout_barrier::PHASE_RENAME);
+
+        let rename =
+            tokio::spawn(
+                async move { SetDisks::rename_data(&disks, bucket, object, &file_infos, bucket, object, DISKS - 1).await },
+            );
+        tokio::time::timeout(BARRIER_PAUSE_GUARD, barrier.wait_until_paused())
+            .await
+            .expect("rename fan-out must reach the armed barrier");
+        rename.abort();
+        assert!(
+            rename
+                .await
+                .expect_err("aborted caller should report cancellation")
+                .is_cancelled(),
+            "caller task should be cancelled, not panic"
+        );
+        assert!(tracker.running() >= 1, "the coordinator must retain in-flight disk mutations");
+
+        barrier.release();
+        tokio::time::timeout(BARRIER_PAUSE_GUARD, async {
+            while tracker.running() != 0 {
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("cancelled caller's disk mutations must drain");
+
+        for (idx, dir) in dirs.iter().enumerate() {
+            assert!(
+                dir.path().join(bucket).join(object).join(STORAGE_FORMAT_FILE).exists(),
+                "disk {idx} must finish the rename after caller cancellation"
+            );
+        }
+    }
+
    /// Demo / regression guard for the barrier on the commit (old-data-dir)
    /// cleanup fan-out. Serves the same #1312/#1319 "no background disk write
    /// after release" shape, on the reclamation path that runs *after* a write is
@@ -6035,7 +6109,7 @@ mod tests {
        assert_eq!(diagnostics.total_responses(), 3);
        assert_eq!(diagnostics.valid_responses(), 1);
        assert_eq!(diagnostics.ignored_responses(), 1);
-        assert_eq!(diagnostics.error_responses(), 2);
+        assert_eq!(diagnostics.non_valid_responses(), 2);
        assert_eq!(diagnostics.first_response_latency(), Some(Duration::from_millis(10)));
        assert_eq!(diagnostics.first_valid_response_latency(), Some(Duration::from_millis(30)));
        assert_eq!(diagnostics.slowest_response_latency(), Some(Duration::from_millis(30)));
@@ -6386,6 +6460,32 @@ mod tests {
        assert_eq!(versions.versions[0].name, object);
    }

+    #[tokio::test]
+    async fn load_file_info_versions_exact_encodes_directory_key_but_returns_logical_name() {
+        let bucket = "exact-directory-versions-bucket";
+        let object = "prefix/directory/";
+        let disk_object = rustfs_utils::path::encode_dir_object(object);
+        let (_dir, disk) = read_multiple_test_disk(bucket, &[]).await;
+        let mut fi = metadata_test_fileinfo(object);
+        fi.version_id = Some(Uuid::new_v4());
+        fi.mod_time = Some(OffsetDateTime::now_utc());
+        disk.write_metadata(bucket, bucket, disk_object.as_str(), fi.clone())
+            .await
+            .expect("directory metadata should be written under the encoded key");
+        let set = io_primitives_test_set(vec![Some(disk)], 0).await;
+
+        let versions = set
+            .load_file_info_versions_exact(bucket, object)
+            .await
+            .expect("exact directory version load should succeed")
+            .expect("exact directory version load should find metadata");
+
+        assert_eq!(versions.name, object);
+        assert_eq!(versions.versions.len(), 1);
+        assert_eq!(versions.versions[0].name, object);
+        assert_eq!(versions.versions[0].version_id, fi.version_id);
+    }
+
    #[tokio::test]
    async fn load_file_info_versions_exact_rejects_transitioned_duplicate_parts() {
        let bucket = "exact-versions-bucket";
@@ -85,6 +85,15 @@ impl SetDisks {
        format!("{}/{}", Self::get_multipart_sha_dir(bucket, object), upload_uuid)
    }

+    pub(super) fn get_multipart_upload_dir(bucket: &str, object: &str, upload_id: &str, data_movement: bool) -> String {
+        let upload_dir = Self::get_upload_id_dir(bucket, object, upload_id);
+        if data_movement {
+            format!("{DATA_MOVEMENT_MULTIPART_PREFIX}/{upload_dir}")
+        } else {
+            upload_dir
+        }
+    }
+
    pub(super) fn get_multipart_sha_dir(bucket: &str, object: &str) -> String {
        let path = format!("{bucket}/{object}");
        let mut hasher = Sha256::new();
@@ -466,6 +475,28 @@ impl SetDisks {
        Self::find_file_info_in_quorum(metas, &mod_time, &etag, quorum)
    }

+    pub(crate) fn hydrate_selected_fileinfo_part_checksums(fi: &mut FileInfo) -> disk::error::Result<()> {
+        fi.hydrate_data_movement_part_checksums().map_err(DiskError::from)?;
+        for part in &fi.parts {
+            let Some(checksums) = part.checksums.as_ref() else {
+                continue;
+            };
+            let mut algorithms = HashSet::with_capacity(checksums.len());
+            for (name, value) in checksums {
+                let Some(checksum) = rustfs_rio::Checksum::new_from_string(name, value) else {
+                    return Err(DiskError::FileCorrupt);
+                };
+                if checksum.checksum_type.is(rustfs_rio::ChecksumType::MULTIPART) {
+                    return Err(DiskError::FileCorrupt);
+                }
+                if !algorithms.insert(checksum.checksum_type.base().0) {
+                    return Err(DiskError::FileCorrupt);
+                }
+            }
+        }
+        Ok(())
+    }
+
    fn update_hash_bytes(hasher: &mut Sha256, value: &[u8]) {
        hasher.update(value.len().to_le_bytes());
        hasher.update(value);
@@ -1079,6 +1110,25 @@ impl SetDisks {
        shuffled_disks
    }

+    pub(super) fn shuffle_disks_owned(mut disks: Vec<Option<DiskStore>>, distribution: &[usize]) -> Vec<Option<DiskStore>> {
+        if distribution.is_empty() {
+            return disks;
+        }
+
+        let mut shuffled_disks = vec![None; disks.len()];
+        for (index, disk) in disks.iter_mut().enumerate() {
+            let Some(slot) = distribution
+                .get(index)
+                .and_then(|block_index| block_index.checked_sub(1))
+                .filter(|slot| *slot < shuffled_disks.len())
+            else {
+                continue;
+            };
+            shuffled_disks[slot] = disk.take();
+        }
+        shuffled_disks
+    }
+
    pub(super) fn shuffle_check_parts(parts_errs: &[usize], distribution: &[usize]) -> Vec<usize> {
        if distribution.is_empty() {
            return parts_errs.to_vec();
@@ -1390,6 +1440,23 @@ mod tests {
        assert_eq!(owned_slots, expected_slots, "fallback disk slots must match the borrowing variant");
    }

+    #[tokio::test]
+    async fn owned_shuffle_preserves_fresh_put_metadata() {
+        let tempdir = tempfile::tempdir().expect("tempdir should be created");
+        let fi = FileInfo::new("bucket/object", 2, 1);
+        let parts = vec![fi.clone(); fi.erasure.distribution.len()];
+        let disks = shuffle_test_disks(&tempdir, parts.len()).await;
+
+        let (owned_disks, owned_parts) = SetDisks::shuffle_disks_and_parts_metadata_by_index_owned(disks, parts, &fi);
+
+        assert!(owned_disks.iter().all(Option::is_some), "fresh PUT must retain every online disk");
+        assert_eq!(
+            owned_parts,
+            vec![fi; owned_disks.len()],
+            "fresh PUT metadata with pending shard indexes must survive init fallback"
+        );
+    }
+
    // backlog#949: corrupt/adversarial distribution values (0 or > N) must not
    // trigger a `usize` underflow / out-of-bounds panic in the shuffle helpers.
    #[test]
@@ -1419,6 +1486,22 @@ mod tests {
        assert_eq!(result.len(), disks.len(), "output length must be preserved");
    }

+    #[tokio::test]
+    async fn owned_disk_shuffle_matches_borrowing_variant() {
+        let tempdir = tempfile::tempdir().expect("tempdir should be created");
+        let mut disks = shuffle_test_disks(&tempdir, 4).await;
+        disks[1] = None;
+        disks[3] = None;
+        let distribution = [3, 1, 4, 2];
+
+        let expected = SetDisks::shuffle_disks(&disks, &distribution);
+        let actual = SetDisks::shuffle_disks_owned(disks, &distribution);
+
+        let expected_slots = expected.iter().map(Option::is_some).collect::<Vec<_>>();
+        let actual_slots = actual.iter().map(Option::is_some).collect::<Vec<_>>();
+        assert_eq!(actual_slots, expected_slots, "owned shuffle must preserve disk placement");
+    }
+
    #[tokio::test]
    async fn shuffle_disks_and_parts_metadata_survives_corrupt_distribution() {
        let tempdir = tempfile::tempdir().expect("tempdir should be created");
@@ -362,9 +362,9 @@ impl SetDisks {
            healing: true,
        };
        let checks = target_disks.into_iter().map(|disk| {
-            let read_options = read_options.clone();
+            let task_read_options = read_options;
            async move {
-                let file_info = match disk.read_version("", bucket, object, version_id, &read_options).await {
+                let file_info = match disk.read_version("", bucket, object, version_id, &task_read_options).await {
                    Ok(file_info) => file_info,
                    Err(
                        DiskError::DiskNotFound
@@ -542,7 +542,8 @@ impl SetDisks {

                let filter_by_etag = quorum_etag.is_some();
                match Self::pick_valid_fileinfo(&parts_metadata, quorum_mod_time, quorum_etag.clone(), read_quorum as usize) {
-                    Ok(latest_meta) => {
+                    Ok(mut latest_meta) => {
+                        Self::hydrate_selected_fileinfo_part_checksums(&mut latest_meta)?;
                        trace!(
                            event = EVENT_SET_DISK_HEAL,
                            component = LOG_COMPONENT_ECSTORE,
@@ -3170,10 +3171,11 @@ mod heal_result_report_tests {
            .await
            .expect("object should be written");

-        let (fi, _, _) = set
+        let snapshot = set
            .get_object_fileinfo(bucket, object, &opts, true, false)
            .await
            .expect("object metadata should resolve");
+        let fi = snapshot.fi();
        assert_eq!(fi.erasure.parity_blocks, 0);
        let data_dir = fi.data_dir.expect("non-inline object should have a data directory");
        let part_path = dir.path().join(bucket).join(object).join(data_dir.to_string()).join("part.1");
@@ -36,19 +36,23 @@ impl crate::storage_api_contracts::namespace::NamespaceLocking for SetDisks {
        // test's transient DistErasure window) would push this set's namespace
        // locking onto its own — possibly empty — dist locker list.
        let set_lock = if self.ctx.is_dist_erasure().await {
-            // Calculate quorum based on lockers count (majority)
-            let lockers_count = self.lockers.len();
+            let lockers = if self.lockers.len() == self.shared_lockers.len()
+                && self
+                    .lockers
+                    .iter()
+                    .zip(self.shared_lockers.iter())
+                    .all(|(current, shared)| Arc::ptr_eq(current, shared))
+            {
+                self.shared_lockers.clone()
+            } else {
+                Arc::from(self.lockers.clone())
+            };
+            // Calculate quorum from the exact client domain used by this lock.
+            let lockers_count = lockers.len();
            let write_quorum = if lockers_count > 1 { (lockers_count / 2) + 1 } else { 1 };
-            NamespaceLock::with_clients_and_quorum(
-                format!("set-{}-{}", self.pool_index, self.set_index),
-                self.lockers.clone(),
-                write_quorum,
-            )
+            NamespaceLock::with_clients_and_quorum_shared(self.set_lock_namespace.clone(), lockers, write_quorum)
        } else {
-            NamespaceLock::Local(LocalLock::new(
-                format!("set-{}-{}", self.pool_index, self.set_index),
-                self.local_lock_manager.clone(),
-            ))
+            NamespaceLock::with_local_manager_shared(self.set_lock_namespace.clone(), self.local_lock_manager.clone())
        };

        let resource = ObjectKey {
@@ -19,7 +19,7 @@ use crate::diagnostics::get::{
    GET_METADATA_CACHE_REASON_DATA_MOVEMENT, GET_METADATA_CACHE_REASON_DELETE_MARKER, GET_METADATA_CACHE_REASON_DIST_ERASURE,
    GET_METADATA_CACHE_REASON_INCL_FREE_VERSIONS, GET_METADATA_CACHE_REASON_INSUFFICIENT_CACHED_QUORUM,
    GET_METADATA_CACHE_REASON_META_BUCKET, GET_METADATA_CACHE_REASON_NO_LOCK, GET_METADATA_CACHE_REASON_NOT_FOUND_OR_EXPIRED,
-    GET_METADATA_CACHE_REASON_NOT_READ_DATA, GET_METADATA_CACHE_REASON_PART_NUMBER,
+    GET_METADATA_CACHE_REASON_NOT_READ_DATA, GET_METADATA_CACHE_REASON_PART_CHECKSUMS, GET_METADATA_CACHE_REASON_PART_NUMBER,
    GET_METADATA_CACHE_REASON_RAW_DATA_MOVEMENT_READ, GET_METADATA_CACHE_REASON_STALE_PUBLICATION,
    GET_METADATA_CACHE_REASON_USABLE, GET_METADATA_CACHE_REASON_VERSION_ID, GET_METADATA_CACHE_REASON_VERSION_SUSPENDED,
    GET_METADATA_CACHE_REASON_VERSIONED, GET_METADATA_EARLY_STOP_REASON_CONFLICTING_METADATA,
@@ -30,12 +30,12 @@ use crate::diagnostics::get::{
    GET_METADATA_RESPONSE_CORRUPT, GET_METADATA_RESPONSE_DISK_NOT_FOUND, GET_METADATA_RESPONSE_ERROR,
    GET_METADATA_RESPONSE_IGNORED, GET_METADATA_RESPONSE_NOT_FOUND, GET_METADATA_RESPONSE_TIMEOUT, GET_METADATA_RESPONSE_VALID,
    GET_METADATA_RESPONSE_VERSION_NOT_FOUND, GET_OBJECT_PATH_CODEC_STREAMING, GET_OBJECT_PATH_DIRECT_MEMORY,
-    GET_OBJECT_PATH_LEGACY_DUPLEX, GET_OBJECT_PATH_SET_DISK, GET_STAGE_DECODE, GET_STAGE_METADATA_CACHE_LOOKUP,
-    GET_STAGE_METADATA_RESOLVE, GET_STAGE_RANGE, GET_STAGE_READER_SETUP, GET_STAGE_READER_SETUP_DROP_PENDING,
-    GET_STAGE_READER_SETUP_SCHEDULE, GET_STAGE_READER_SETUP_WAIT_QUORUM, GET_STAGE_READER_TASK_BITROT_READER_INIT,
-    GET_STAGE_READER_TASK_FILE_OPEN, GET_STAGE_READER_TASK_READER_CONSTRUCTION, GetObjectFailureReason, classify_disk_error,
-    get_stage_timer_if_enabled, mark_get_object_downstream_closed, record_get_object_pipeline_failure,
-    record_get_object_pipeline_failure_for_path, record_get_stage_duration_if_enabled,
+    GET_OBJECT_PATH_INTERNAL_META, GET_OBJECT_PATH_LEGACY_DUPLEX, GET_OBJECT_PATH_SET_DISK, GET_STAGE_DECODE,
+    GET_STAGE_METADATA_CACHE_LOOKUP, GET_STAGE_METADATA_RESOLVE, GET_STAGE_RANGE, GET_STAGE_READER_SETUP,
+    GET_STAGE_READER_SETUP_DROP_PENDING, GET_STAGE_READER_SETUP_SCHEDULE, GET_STAGE_READER_SETUP_WAIT_QUORUM,
+    GET_STAGE_READER_TASK_BITROT_READER_INIT, GET_STAGE_READER_TASK_FILE_OPEN, GET_STAGE_READER_TASK_READER_CONSTRUCTION,
+    GetObjectFailureReason, classify_disk_error, get_stage_timer_if_enabled, mark_get_object_downstream_closed,
+    record_get_object_pipeline_failure, record_get_object_pipeline_failure_for_path, record_get_stage_duration_if_enabled,
 };
 use crate::erasure::coding::BitrotReader;
 use crate::io_support::bitrot::{
@@ -116,9 +116,9 @@ impl SetDisks {
            .then_some(GET_METADATA_CACHE_REASON_DIST_ERASURE)
    }

-    async fn cached_get_object_fileinfo(&self, bucket: &str, object: &str) -> Option<GetObjectMetadataCacheEntry> {
+    async fn cached_get_object_fileinfo(&self, bucket: &str, object: &str) -> Option<Arc<GetObjectMetadataCacheEntry>> {
        match self.lookup_cached_get_object_fileinfo(bucket, object).await {
-            MetadataCacheLookup::Hit(entry) => Some((*entry).clone()),
+            MetadataCacheLookup::Hit(entry) => Some(entry),
            MetadataCacheLookup::Miss | MetadataCacheLookup::RejectedInsufficientQuorum => None,
        }
    }
@@ -221,10 +221,10 @@ impl SetDisks {
        let disks = self.disks.read().await.clone();
        let required_reads = self.default_read_quorum();

-        let bucket = bucket.to_string();
-        let object = object.to_string();
-        let version_id = version_id.to_string();
-        let opts = opts.clone();
+        let bucket: Arc<str> = Arc::from(bucket);
+        let object: Arc<str> = Arc::from(object);
+        let version_id: Arc<str> = Arc::from(version_id);
+        let opts = *opts;

        let processor = runtime_sources::batch_processors().read_processor();
        let tasks: Vec<_> = disks
@@ -235,9 +235,9 @@ impl SetDisks {
                    let bucket = bucket.clone();
                    let object = object.clone();
                    let version_id = version_id.clone();
-                    let opts = opts.clone();
+                    let task_opts = opts;

-                    async move { disk.read_version(&bucket, &bucket, &object, &version_id, &opts).await }
+                    async move { disk.read_version(&bucket, &bucket, &object, &version_id, &task_opts).await }
                })
            })
            .collect();
@@ -257,7 +257,7 @@ impl SetDisks {
        opts: &ObjectOptions,
        read_data: bool,
        caller_allows_early_stop: bool,
-    ) -> Result<(FileInfo, Vec<FileInfo>, Vec<Option<DiskStore>>)> {
+    ) -> Result<GetObjectFileInfo> {
        self.get_object_fileinfo_gated(bucket, object, opts, read_data, caller_allows_early_stop)
            .await
    }
@@ -274,7 +274,7 @@ impl SetDisks {
        opts: &ObjectOptions,
        read_data: bool,
        allow_early_stop: bool,
-    ) -> Result<(FileInfo, Vec<FileInfo>, Vec<Option<DiskStore>>)> {
+    ) -> Result<GetObjectFileInfo> {
        let vid = opts.version_id.clone().unwrap_or_default();
        let stage_metrics_enabled = rustfs_io_metrics::get_stage_metrics_enabled();

@@ -300,7 +300,7 @@ impl SetDisks {
                        GET_STAGE_METADATA_CACHE_LOOKUP,
                        metadata_cache_lookup_start,
                    );
-                    return Ok((cached.fi.clone(), cached.parts_metadata.clone(), cached.online_disks.clone()));
+                    return Ok(GetObjectFileInfo::shared(cached));
                }
                MetadataCacheLookup::Miss => {
                    rustfs_io_metrics::record_get_object_metadata_cache_decision(
@@ -336,7 +336,7 @@ impl SetDisks {
        // read_all_fileinfo_observed (see read_all_fileinfo_early_stop in
        // core/io_primitives.rs); unsafe requests and callers that opt out
        // (allow_early_stop=false) fall back to full-wait.
-        let (parts_metadata, errs, metadata_fanout_diagnostics) = Self::read_all_fileinfo_observed(
+        let (mut parts_metadata, errs, metadata_fanout_diagnostics) = Self::read_all_fileinfo_observed(
            &disks,
            "",
            bucket,
@@ -349,7 +349,12 @@ impl SetDisks {
            self.default_parity_count,
        )
        .await?;
-        metadata_fanout_diagnostics.record(GET_OBJECT_PATH_LEGACY_DUPLEX);
+        let metadata_metrics_path = if crate::bucket::utils::is_meta_bucketname(bucket) {
+            GET_OBJECT_PATH_INTERNAL_META
+        } else {
+            GET_OBJECT_PATH_LEGACY_DUPLEX
+        };
+        metadata_fanout_diagnostics.record(metadata_metrics_path);
        let metadata_fanout_complete = metadata_fanout_diagnostics.total_responses() >= disks.len();
        // warn!("get_object_fileinfo parts_metadata {:?}", &parts_metadata);
        // warn!("get_object_fileinfo {}/{} errs {:?}", bucket, object, &errs);
@@ -385,9 +390,18 @@ impl SetDisks {
            return Err(to_object_err(err.into(), vec![bucket, object]));
        }

-        let (op_online_disks, fi, fileinfo_selection_quorum) =
+        let (op_online_disks, mut fi, fileinfo_selection_quorum) =
            Self::select_valid_fileinfo(&disks, &parts_metadata, &errs, vid.as_str(), read_quorum, write_quorum)?;
-        metadata_fanout_diagnostics.record_quorum_candidate_latency(GET_OBJECT_PATH_LEGACY_DUPLEX, fileinfo_selection_quorum);
+        let include_part_checksums =
+            opts.include_part_checksums || opts.part_number.is_some() || opts.data_movement || opts.raw_data_movement_read;
+        if include_part_checksums {
+            Self::hydrate_selected_fileinfo_part_checksums(&mut fi)?;
+        } else {
+            for metadata in std::iter::once(&mut fi).chain(parts_metadata.iter_mut()) {
+                rustfs_utils::http::remove_str(&mut metadata.metadata, rustfs_utils::http::SUFFIX_PART_CHECKSUMS);
+            }
+        }
+        metadata_fanout_diagnostics.record_quorum_candidate_latency(metadata_metrics_path, fileinfo_selection_quorum);
        if errs.iter().any(|err| err.is_some()) {
            let version_id = resolved_read_repair_version_id(&fi, opts.version_id.as_deref());
            submit_read_repair_heal(
@@ -418,7 +432,7 @@ impl SetDisks {

        // let online_disks: Vec<Option<DiskStore>> = op_online_disks.iter().filter(|v| v.is_some()).cloned().collect();

-        Ok((fi, parts_metadata, op_online_disks))
+        Ok(GetObjectFileInfo::owned(fi, parts_metadata, op_online_disks))
    }

    #[hotpath::measure(impl_type = "SetDisks")]
@@ -428,14 +442,15 @@ impl SetDisks {
        object: &str,
        opts: &ObjectOptions,
    ) -> (ObjectInfo, usize, Option<StorageError>) {
-        let fi = match self.get_object_fileinfo(bucket, object, opts, false, false).await {
-            Ok((fi, _, _)) => fi,
+        let snapshot = match self.get_object_fileinfo(bucket, object, opts, false, false).await {
+            Ok(snapshot) => snapshot,
            Err(e) => return (ObjectInfo::default(), 0, Some(e)),
        };
+        let fi = snapshot.fi();

        let write_quorum = fi.write_quorum(self.default_write_quorum());

-        let oi = ObjectInfo::from_file_info(&fi, bucket, object, opts.versioned || opts.version_suspended);
+        let oi = ObjectInfo::from_file_info(fi, bucket, object, opts.versioned || opts.version_suspended);

        if !fi.version_purge_status().is_empty() && opts.version_id.is_some() {
            return (
@@ -1813,6 +1828,9 @@ fn get_object_metadata_cache_request_bypass_reason(bucket: &str, opts: &ObjectOp
    if opts.part_number.is_some() {
        return Some(GET_METADATA_CACHE_REASON_PART_NUMBER);
    }
+    if opts.include_part_checksums {
+        return Some(GET_METADATA_CACHE_REASON_PART_CHECKSUMS);
+    }
    if opts.data_movement {
        return Some(GET_METADATA_CACHE_REASON_DATA_MOVEMENT);
    }
@@ -2484,6 +2502,16 @@ mod metadata_cache_tests {
            Some(GET_METADATA_CACHE_REASON_PART_NUMBER)
        );

+        opts = ObjectOptions {
+            include_part_checksums: true,
+            ..Default::default()
+        };
+        assert!(!is_get_object_metadata_cache_request_eligible("bucket", &opts, true));
+        assert_eq!(
+            get_object_metadata_cache_request_bypass_reason("bucket", &opts, true),
+            Some(GET_METADATA_CACHE_REASON_PART_CHECKSUMS)
+        );
+
        opts = ObjectOptions {
            data_movement: true,
            ..Default::default()
@@ -2674,6 +2702,31 @@ mod metadata_cache_tests {
        assert_eq!(cached.read_quorum, 0);
    }

+    #[tokio::test]
+    async fn get_object_fileinfo_cache_hit_shares_cached_metadata() {
+        let set = new_metadata_cache_test_set().await;
+        let fi = valid_test_fileinfo("object");
+        let parts_metadata = vec![fi.clone()];
+        let online_disks = Vec::new();
+        let generation = set.get_object_metadata_cache_generation("bucket", "object");
+        set.cache_get_object_fileinfo(("bucket", "object"), generation, &fi, &parts_metadata, &online_disks, 0)
+            .await;
+        let cached = set
+            .cached_get_object_fileinfo("bucket", "object")
+            .await
+            .expect("fresh cache entry should be returned");
+
+        let returned = set
+            .get_object_fileinfo("bucket", "object", &ObjectOptions::default(), true, false)
+            .await
+            .expect("cache-backed metadata lookup should succeed");
+
+        assert!(
+            returned.shared_entry().is_some_and(|value| Arc::ptr_eq(value, &cached)),
+            "cache hits must share the complete metadata snapshot"
+        );
+    }
+
    #[tokio::test]
    async fn get_object_metadata_cache_rejects_deleted_and_invalid_fileinfo() {
        let set = new_metadata_cache_test_set().await;
@@ -2729,9 +2782,6 @@ mod metadata_cache_tests {

    #[tokio::test]
    async fn get_object_metadata_cache_rejects_stale_entries() {
-        // moka handles TTL expiry automatically via time_to_live(250ms).
-        // This test verifies that entries inserted with the cache API are retrievable
-        // while fresh, and that the cache API works correctly.
        let set = new_metadata_cache_test_set().await;
        let fi = valid_test_fileinfo("object");

@@ -2743,6 +2793,14 @@ mod metadata_cache_tests {
            set.cached_get_object_fileinfo("bucket", "object").await.is_some(),
            "freshly inserted entry should be returned"
        );
+
+        tokio::time::timeout(GET_OBJECT_METADATA_CACHE_TTL + Duration::from_secs(1), async {
+            while set.cached_get_object_fileinfo("bucket", "object").await.is_some() {
+                tokio::time::sleep(Duration::from_millis(10)).await;
+            }
+        })
+        .await
+        .expect("metadata cache entry should expire after its TTL");
    }

    #[tokio::test]
@@ -2804,9 +2862,12 @@ mod metadata_cache_tests {
        barrier.wait_until_paused().await;
        set.invalidate_get_object_metadata_cache(bucket, object).await;
        barrier.release();
-        read.await
+        let snapshot = read
+            .await
            .expect("metadata read task should not panic")
            .expect("metadata fanout should still return its selected FileInfo");
+        assert!(snapshot.owned.is_some());
+        assert!(snapshot.has_valid_representation());

        assert!(
            set.get_object_metadata_cache
@@ -3415,7 +3476,7 @@ mod tests {
        );
        assert_eq!(diagnostics.total_responses(), 9);
        assert_eq!(diagnostics.valid_responses(), 1);
-        assert_eq!(diagnostics.error_responses(), 8);
+        assert_eq!(diagnostics.non_valid_responses(), 8);
    }

    #[test]
@@ -3430,7 +3491,7 @@ mod tests {
        );

        assert_eq!(diagnostics.ignored_responses(), 2);
-        assert_eq!(diagnostics.error_responses(), 3);
+        assert_eq!(diagnostics.non_valid_responses(), 3);
        assert_eq!(diagnostics.observations[0].outcome, GET_METADATA_RESPONSE_DISK_NOT_FOUND);
        assert_eq!(diagnostics.observations[1].outcome, GET_METADATA_RESPONSE_IGNORED);
        assert_eq!(diagnostics.observations[2].outcome, GET_METADATA_RESPONSE_NOT_FOUND);
@@ -3498,7 +3559,7 @@ mod tests {

        assert_eq!(diagnostics.total_responses(), 3);
        assert_eq!(diagnostics.valid_responses(), 3);
-        assert_eq!(diagnostics.error_responses(), 0);
+        assert_eq!(diagnostics.non_valid_responses(), 0);
        assert!(
            diagnostics
                .observations
@@ -4028,8 +4089,8 @@ mod tests {
        get_codec_streaming_reader_gate(
            CODEC_STREAMING_TEST_BUCKET,
            CODEC_STREAMING_TEST_OBJECT,
-            range,
            None,
+            classify_get_codec_streaming_object_class(range, object_info, fi),
            object_info,
            fi,
            lock_optimization_enabled,
@@ -4046,8 +4107,8 @@ mod tests {
        get_codec_streaming_reader_gate(
            CODEC_STREAMING_TEST_BUCKET,
            CODEC_STREAMING_TEST_OBJECT,
-            range,
            part_number,
+            classify_get_codec_streaming_object_class(range, object_info, fi),
            object_info,
            fi,
            lock_optimization_enabled,
@@ -4773,6 +4834,114 @@ mod tests {
        .await
    }

+    async fn encoded_inline_blocks(blocks: &[&[u8]], shard_size: usize, hash_algo: HashAlgorithm) -> Bytes {
+        let mut writer = BitrotWriter::new(Cursor::new(Vec::new()), shard_size, hash_algo);
+        for block in blocks {
+            writer.write(block).await.expect("test block should be encoded");
+        }
+        Bytes::from(writer.into_inner().into_inner())
+    }
+
+    fn assert_reader_shares_inline_allocation(reader: &ObjectBitrotReader, source: &Bytes) {
+        let reader_bytes = reader
+            .inner_ref()
+            .inline_bytes()
+            .expect("inline scheduler should retain an in-memory Bytes source");
+        assert_eq!(
+            reader_bytes.as_ptr(),
+            source.as_ptr(),
+            "the scheduler must clone Bytes ownership instead of copying the inline shard payload"
+        );
+    }
+
+    #[tokio::test]
+    async fn inline_range_scheduler_shares_bytes_and_rejects_bitrot_mismatch() {
+        const SHARD_SIZE: usize = 16;
+        let hash_algo = HashAlgorithm::HighwayHash256S;
+        let first = [b'a'; SHARD_SIZE];
+        let second = [b'b'; SHARD_SIZE];
+        let mut source = encoded_inline_blocks(&[&first, &second], SHARD_SIZE, hash_algo.clone()).await;
+        let second_payload = hash_algo.size() * 2 + SHARD_SIZE;
+        source = {
+            let mut corrupt = source.to_vec();
+            corrupt[second_payload] ^= 0xff;
+            Bytes::from(corrupt)
+        };
+        let files = vec![encoded_reader_setup_fileinfo(Some(source.to_vec()))];
+        let source = files[0].data.clone().expect("inline shard should exist");
+        let disks = vec![None];
+
+        let mut setup = create_bitrot_readers_until_quorum_with_preference(
+            &files,
+            &disks,
+            "bucket",
+            "object",
+            1,
+            SHARD_SIZE,
+            SHARD_SIZE,
+            SHARD_SIZE,
+            hash_algo,
+            false,
+            false,
+            1,
+            0,
+            BitrotReaderSetupMode::ReadQuorum,
+            true,
+            None,
+            None,
+        )
+        .await;
+        let mut reader = setup.readers[0].take().expect("range reader should be ready");
+        assert_reader_shares_inline_allocation(&reader, &source);
+
+        let err = reader
+            .read(&mut [0; SHARD_SIZE])
+            .await
+            .expect_err("corrupt ranged inline block must fail bitrot verification");
+        assert_eq!(err.kind(), ErrorKind::InvalidData);
+    }
+
+    #[tokio::test]
+    async fn inline_part_scheduler_shares_bytes_and_rejects_bitrot_mismatch() {
+        const SHARD_SIZE: usize = 16;
+        let hash_algo = HashAlgorithm::HighwayHash256S;
+        let block = [b'p'; SHARD_SIZE];
+        let encoded = encoded_inline_blocks(&[&block], SHARD_SIZE, hash_algo.clone()).await;
+        let mut corrupt = encoded.to_vec();
+        corrupt[hash_algo.size()] ^= 0xff;
+        let files = vec![encoded_reader_setup_fileinfo(Some(corrupt))];
+        let source = files[0].data.clone().expect("inline shard should exist");
+        let disks = vec![None];
+
+        let mut setup = create_bitrot_readers_until_quorum_all_shards(
+            &files,
+            &disks,
+            "bucket",
+            "object",
+            7,
+            0,
+            SHARD_SIZE,
+            SHARD_SIZE,
+            hash_algo,
+            false,
+            false,
+            1,
+            0,
+            BitrotReaderSetupMode::VerifyReconstruction,
+            None,
+            None,
+        )
+        .await;
+        let mut reader = setup.readers[0].take().expect("part reader should be ready");
+        assert_reader_shares_inline_allocation(&reader, &source);
+
+        let err = reader
+            .read(&mut [0; SHARD_SIZE])
+            .await
+            .expect_err("corrupt inline part must fail bitrot verification");
+        assert_eq!(err.kind(), ErrorKind::InvalidData);
+    }
+
    async fn decode_codec_data_blocks_first_setup(
        erasure: coding::Erasure,
        data: &[u8],
@@ -5474,35 +5643,95 @@ mod tests {
    }

    #[test]
-    fn rustfs_codec_streaming_uses_conservative_default_min_size() {
+    fn codec_streaming_config_cache_loads_once() {
+        use std::cell::Cell;
+
+        let loads = Cell::new(0);
+        let expected = GetCodecStreamingConfig {
+            enabled: true,
+            rollout: GetCodecStreamingRollout::Off,
+            rollout_pct: 100,
+            body_compat_confirmed: true,
+            header_compat_confirmed: true,
+            engine: GetCodecStreamingEngine::Legacy,
+            min_size: DEFAULT_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE,
+        };
+
+        for _ in 0..3 {
+            assert_eq!(
+                get_codec_streaming_config_cached_core(|| {
+                    loads.set(loads.get() + 1);
+                    expected
+                }),
+                expected
+            );
+        }
+        assert_eq!(loads.get(), 1, "production config cache must not reload env per GET");
+    }
+
+    #[test]
+    fn codec_streaming_config_loader_preserves_all_gate_env_overrides() {
        temp_env::with_vars(
            [
-                (ENV_RUSTFS_GET_CODEC_STREAMING_ENABLE, Some("true")),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_ENABLE, Some("false")),
                (ENV_RUSTFS_GET_CODEC_STREAMING_ENGINE, Some(GET_CODEC_STREAMING_ENGINE_RUSTFS)),
-                (ENV_RUSTFS_GET_CODEC_STREAMING_ROLLOUT, Some("benchmark")),
-                (ENV_RUSTFS_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED, Some("true")),
-                (ENV_RUSTFS_GET_CODEC_STREAMING_HEADER_COMPAT_CONFIRMED, Some("true")),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_ROLLOUT, Some("production")),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_ROLLOUT_PCT, Some("37")),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED, Some("false")),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_HEADER_COMPAT_CONFIRMED, Some("false")),
                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, None::<&str>),
-                (ENV_RUSTFS_GET_CODEC_STREAMING_RUSTFS_MIN_SIZE, None::<&str>),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_RUSTFS_MIN_SIZE, Some("262144")),
            ],
            || {
-                let below_threshold_fi = codec_streaming_test_fileinfo(512 * 1024, 1);
-                let below_threshold_object_info = codec_streaming_test_object_info(&below_threshold_fi);
                assert_eq!(
-                    codec_streaming_reader_gate_for_test(&None, &below_threshold_object_info, &below_threshold_fi, true).decision,
-                    GetCodecStreamingDecision::Fallback(GetCodecStreamingFallbackReason::BelowMinSize)
-                );
-
-                let threshold_fi = codec_streaming_test_fileinfo(1_048_576, 1);
-                let threshold_object_info = codec_streaming_test_object_info(&threshold_fi);
-                assert_eq!(
-                    codec_streaming_reader_gate_for_test(&None, &threshold_object_info, &threshold_fi, true).decision,
-                    GetCodecStreamingDecision::Use
+                    load_get_codec_streaming_config(),
+                    GetCodecStreamingConfig {
+                        enabled: false,
+                        rollout: GetCodecStreamingRollout::On,
+                        rollout_pct: 37,
+                        body_compat_confirmed: false,
+                        header_compat_confirmed: false,
+                        engine: GetCodecStreamingEngine::Rustfs,
+                        min_size: 262144,
+                    }
                );
            },
        );
    }

+    #[test]
+    fn codec_streaming_default_min_size_meets_direct_memory_ceiling() {
+        for engine in [None, Some(GET_CODEC_STREAMING_ENGINE_RUSTFS)] {
+            temp_env::with_vars(
+                [
+                    (ENV_RUSTFS_GET_CODEC_STREAMING_ENABLE, Some("true")),
+                    (ENV_RUSTFS_GET_CODEC_STREAMING_ENGINE, engine),
+                    (ENV_RUSTFS_GET_CODEC_STREAMING_ROLLOUT, Some("benchmark")),
+                    (ENV_RUSTFS_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED, Some("true")),
+                    (ENV_RUSTFS_GET_CODEC_STREAMING_HEADER_COMPAT_CONFIRMED, Some("true")),
+                    (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, None::<&str>),
+                    (ENV_RUSTFS_GET_CODEC_STREAMING_RUSTFS_MIN_SIZE, None::<&str>),
+                ],
+                || {
+                    let below_threshold_fi = codec_streaming_test_fileinfo(128 * 1024 - 1, 1);
+                    let below_threshold_object_info = codec_streaming_test_object_info(&below_threshold_fi);
+                    assert_eq!(
+                        codec_streaming_reader_gate_for_test(&None, &below_threshold_object_info, &below_threshold_fi, true)
+                            .decision,
+                        GetCodecStreamingDecision::Fallback(GetCodecStreamingFallbackReason::BelowMinSize)
+                    );
+
+                    let threshold_fi = codec_streaming_test_fileinfo(128 * 1024, 1);
+                    let threshold_object_info = codec_streaming_test_object_info(&threshold_fi);
+                    assert_eq!(
+                        codec_streaming_reader_gate_for_test(&None, &threshold_object_info, &threshold_fi, true).decision,
+                        GetCodecStreamingDecision::Use
+                    );
+                },
+            );
+        }
+    }
+
    #[test]
    fn rustfs_codec_streaming_min_size_override_can_lower_threshold() {
        temp_env::with_vars(
@@ -5798,10 +6027,10 @@ mod tests {
                (ENV_RUSTFS_GET_CODEC_STREAMING_ROLLOUT, None::<&str>),
                (ENV_RUSTFS_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED, None::<&str>),
                (ENV_RUSTFS_GET_CODEC_STREAMING_HEADER_COMPAT_CONFIRMED, None::<&str>),
-                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, Some("1")),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, None::<&str>),
            ],
            || {
-                let fi = codec_streaming_test_fileinfo(1024, 1);
+                let fi = codec_streaming_test_fileinfo(128 * 1024, 1);
                let object_info = codec_streaming_test_object_info(&fi);

                assert_eq!(
@@ -5822,10 +6051,10 @@ mod tests {
                (ENV_RUSTFS_GET_CODEC_STREAMING_ROLLOUT, Some("on")),
                (ENV_RUSTFS_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED, None::<&str>),
                (ENV_RUSTFS_GET_CODEC_STREAMING_HEADER_COMPAT_CONFIRMED, None::<&str>),
-                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, Some("1")),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, None::<&str>),
            ],
            || {
-                let fi = codec_streaming_test_fileinfo(1024, 1);
+                let fi = codec_streaming_test_fileinfo(128 * 1024, 1);
                let object_info = codec_streaming_test_object_info(&fi);

                assert_eq!(
@@ -5843,10 +6072,10 @@ mod tests {
            [
                (ENV_RUSTFS_GET_CODEC_STREAMING_ENABLE, Some("false")),
                (ENV_RUSTFS_GET_CODEC_STREAMING_ROLLOUT, Some("on")),
-                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, Some("1")),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, None::<&str>),
            ],
            || {
-                let fi = codec_streaming_test_fileinfo(1024, 1);
+                let fi = codec_streaming_test_fileinfo(128 * 1024, 1);
                let object_info = codec_streaming_test_object_info(&fi);

                assert_eq!(
@@ -5926,10 +6155,10 @@ mod tests {
                (ENV_RUSTFS_GET_CODEC_STREAMING_ROLLOUT_PCT, Some("0")),
                (ENV_RUSTFS_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED, Some("true")),
                (ENV_RUSTFS_GET_CODEC_STREAMING_HEADER_COMPAT_CONFIRMED, Some("true")),
-                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, Some("1")),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, None::<&str>),
            ],
            || {
-                let fi = codec_streaming_test_fileinfo(1024, 1);
+                let fi = codec_streaming_test_fileinfo(128 * 1024, 1);
                let object_info = codec_streaming_test_object_info(&fi);

                assert_eq!(
@@ -5946,10 +6175,10 @@ mod tests {
                (ENV_RUSTFS_GET_CODEC_STREAMING_ROLLOUT_PCT, Some("100")),
                (ENV_RUSTFS_GET_CODEC_STREAMING_BODY_COMPAT_CONFIRMED, Some("true")),
                (ENV_RUSTFS_GET_CODEC_STREAMING_HEADER_COMPAT_CONFIRMED, Some("true")),
-                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, Some("1")),
+                (ENV_RUSTFS_GET_CODEC_STREAMING_MIN_SIZE, None::<&str>),
            ],
            || {
-                let fi = codec_streaming_test_fileinfo(1024, 1);
+                let fi = codec_streaming_test_fileinfo(128 * 1024, 1);
                let object_info = codec_streaming_test_object_info(&fi);

                assert_eq!(
@@ -75,11 +75,13 @@ impl SetDisks {
            version_id,
            versioned: opts.versioned,
            version_suspended: opts.version_suspended,
+            include_part_checksums: true,
            ..Default::default()
        };
        let (mut fi, _, disks) = self
            .get_object_fileinfo_gated(bucket, object, &read_opts, false, false)
-            .await?;
+            .await?
+            .into_owned();
        if let Some(expected_operation_id) = expected_operation_id {
            require_restore_operation_id(&fi.metadata, expected_operation_id)?;
        }
@@ -101,7 +103,7 @@ impl SetDisks {
            bucket,
            object,
            fi.clone(),
-            disks.as_slice(),
+            &disks,
            &UpdateMetadataOpts {
                replace_user_metadata: true,
                ..Default::default()
@@ -141,11 +143,13 @@ impl SetDisks {
            version_id,
            versioned: opts.versioned,
            version_suspended: opts.version_suspended,
+            include_part_checksums: true,
            ..Default::default()
        };
        let (mut fi, _, disks) = self
            .get_object_fileinfo_gated(bucket, object, &read_opts, false, false)
-            .await?;
+            .await?
+            .into_owned();
        if let Some(expected_operation_id) = expected_operation_id {
            match restore_operation_id_from_metadata(&fi.metadata)? {
                Some(actual_operation_id) if actual_operation_id == expected_operation_id => {}
@@ -170,7 +174,7 @@ impl SetDisks {
            bucket,
            object,
            fi,
-            disks.as_slice(),
+            &disks,
            &UpdateMetadataOpts {
                replace_user_metadata: true,
                ..Default::default()
@@ -16,6 +16,14 @@ use crate::diagnostics::get::{
    GET_SHARD_READ_COST_LOCAL, GET_SHARD_READ_COST_REMOTE, GET_SHARD_READ_COST_SAME_NODE, GET_SHARD_READ_COST_UNKNOWN,
 };
 use crate::disk::error::Error;
+use crate::layout::disks_layout::MAX_ERASURE_SET_DRIVE_COUNT;
+use smallvec::SmallVec;
+
+/// Generic codec callers may exceed the production set limit; `SmallVec` then
+/// spills without changing slot semantics.
+pub(crate) const INLINE_SHARD_SLOTS: usize = MAX_ERASURE_SET_DRIVE_COUNT;
+pub(crate) type ShardBuffers = SmallVec<[Option<Vec<u8>>; INLINE_SHARD_SLOTS]>;
+pub(crate) type ShardErrors = SmallVec<[Option<Error>; INLINE_SHARD_SLOTS]>;

 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub(crate) enum ShardReadCost {
@@ -43,201 +51,137 @@ impl ShardReadCost {
    }
 }

-#[derive(Debug, Clone, PartialEq, Eq)]
-pub(crate) struct ShardSlot {
-    index: usize,
-    read_cost: ShardReadCost,
-    data: Option<Vec<u8>>,
-    error: Option<Error>,
-}
-
-impl ShardSlot {
-    pub(crate) fn new(index: usize, data: Option<Vec<u8>>, error: Option<Error>) -> Self {
-        Self::with_read_cost(index, ShardReadCost::Unknown, data, error)
-    }
-
-    pub(crate) fn with_read_cost(index: usize, read_cost: ShardReadCost, data: Option<Vec<u8>>, error: Option<Error>) -> Self {
-        Self {
-            index,
-            read_cost,
-            data,
-            error,
-        }
-    }
-
-    pub(crate) fn data(index: usize, data: Vec<u8>) -> Self {
-        Self::new(index, Some(data), None)
-    }
-
-    pub(crate) fn data_with_read_cost(index: usize, read_cost: ShardReadCost, data: Vec<u8>) -> Self {
-        Self::with_read_cost(index, read_cost, Some(data), None)
-    }
-
-    pub(crate) fn missing(index: usize, error: Error) -> Self {
-        Self::new(index, None, Some(error))
-    }
-
-    pub(crate) fn missing_with_read_cost(index: usize, read_cost: ShardReadCost, error: Error) -> Self {
-        Self::with_read_cost(index, read_cost, None, Some(error))
-    }
-
-    pub(crate) fn index(&self) -> usize {
-        self.index
-    }
-
-    pub(crate) fn read_cost(&self) -> ShardReadCost {
-        self.read_cost
-    }
-
-    pub(crate) fn has_data(&self) -> bool {
-        self.data.is_some()
-    }
-
-    pub(crate) fn data_bytes(&self) -> Option<&[u8]> {
-        self.data.as_deref()
-    }
-
-    pub(crate) fn error(&self) -> Option<&Error> {
-        self.error.as_ref()
-    }
-}
-
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub(crate) struct StripeReadState {
-    slots: Vec<ShardSlot>,
+    shards: ShardBuffers,
+    errors: ShardErrors,
    read_quorum: usize,
 }

 impl StripeReadState {
-    pub(crate) fn new(slots: Vec<ShardSlot>, read_quorum: usize) -> Self {
-        Self { slots, read_quorum }
-    }
-
+    #[cfg(test)]
    pub(crate) fn from_parts(shards: Vec<Option<Vec<u8>>>, errors: Vec<Option<Error>>, read_quorum: usize) -> Self {
-        Self::from_parts_with_read_costs(shards, errors, &[], read_quorum)
+        let mut shards = SmallVec::from_vec(shards);
+        let mut errors = SmallVec::from_vec(errors);
+        let slot_count = shards.len().max(errors.len());
+        shards.resize_with(slot_count, || None);
+        errors.resize_with(slot_count, || None);
+        Self {
+            shards,
+            errors,
+            read_quorum,
+        }
    }

-    pub(crate) fn from_parts_with_read_costs(
-        shards: Vec<Option<Vec<u8>>>,
-        errors: Vec<Option<Error>>,
-        read_costs: &[ShardReadCost],
-        read_quorum: usize,
-    ) -> Self {
-        let slot_count = shards.len().max(errors.len());
-        let mut slots = Vec::with_capacity(slot_count);
-        let mut shards = shards.into_iter();
-        let mut errors = errors.into_iter();
-        for index in 0..slot_count {
-            let read_cost = read_costs.get(index).copied().unwrap_or(ShardReadCost::Unknown);
-            slots.push(ShardSlot::with_read_cost(
-                index,
-                read_cost,
-                shards.next().flatten(),
-                errors.next().flatten(),
-            ));
-        }
-        Self::new(slots, read_quorum)
+    pub(crate) fn with_slot_count(slot_count: usize, read_quorum: usize) -> Self {
+        let mut state = Self {
+            shards: SmallVec::new(),
+            errors: SmallVec::new(),
+            read_quorum,
+        };
+        state.reset(slot_count, read_quorum);
+        state
+    }
+
+    pub(crate) fn reset(&mut self, slot_count: usize, read_quorum: usize) {
+        self.shards.clear();
+        self.shards.resize_with(slot_count, || None);
+        self.errors.clear();
+        self.errors.resize_with(slot_count, || None);
+        self.read_quorum = read_quorum;
    }

    pub(crate) fn available_shards(&self) -> usize {
-        self.slots.iter().filter(|slot| slot.has_data()).count()
+        self.shards.iter().filter(|shard| shard.is_some()).count()
    }

    pub(crate) fn can_decode(&self) -> bool {
        self.available_shards() >= self.read_quorum
    }

-    pub(crate) fn slots(&self) -> &[ShardSlot] {
-        &self.slots
+    pub(crate) fn is_empty(&self) -> bool {
+        self.shards.is_empty()
    }

-    pub(crate) fn slot_by_index(&self, index: usize) -> Option<&ShardSlot> {
-        if let Some(slot) = self.slots.get(index)
-            && slot.index == index
-        {
-            return Some(slot);
-        }
-        self.slots.iter().find(|slot| slot.index == index)
+    pub(crate) fn data_bytes(&self, index: usize) -> Option<&[u8]> {
+        self.shards.get(index).and_then(Option::as_deref)
+    }
+
+    #[cfg(test)]
+    pub(crate) fn error(&self, index: usize) -> Option<&Error> {
+        self.errors.get(index).and_then(Option::as_ref)
    }

    pub(crate) fn data_shards_complete(&self, data_shards: usize) -> bool {
-        (0..data_shards).all(|index| self.slot_by_index(index).is_some_and(ShardSlot::has_data))
+        self.shards.len() >= data_shards && self.shards.iter().take(data_shards).all(Option::is_some)
    }

-    pub(crate) fn into_parts(self) -> (Vec<Option<Vec<u8>>>, Vec<Option<Error>>) {
-        let part_count = self.slots.iter().map(|slot| slot.index).max().map_or(0, |index| index + 1);
-        let mut shards = Vec::with_capacity(part_count);
-        shards.resize_with(part_count, || None);
-        let mut errors = Vec::with_capacity(part_count);
-        errors.resize_with(part_count, || None);
-        for slot in self.slots {
-            shards[slot.index] = slot.data;
-            errors[slot.index] = slot.error;
-        }
-        (shards, errors)
+    pub(crate) fn parts_mut(&mut self) -> (&mut ShardBuffers, &mut ShardErrors) {
+        (&mut self.shards, &mut self.errors)
+    }
+
+    pub(crate) fn shards_mut(&mut self) -> &mut ShardBuffers {
+        &mut self.shards
+    }
+
+    pub(crate) fn into_parts(self) -> (ShardBuffers, ShardErrors) {
+        (self.shards, self.errors)
+    }
+
+    #[cfg(test)]
+    pub(crate) fn scratch_storage(&self) -> (*const Option<Vec<u8>>, *const Option<Error>, bool, bool) {
+        (self.shards.as_ptr(), self.errors.as_ptr(), self.shards.spilled(), self.errors.spilled())
+    }
+
+    #[cfg(test)]
+    pub(crate) fn shard_allocation(&self, index: usize) -> Option<(*const u8, usize)> {
+        self.shards
+            .get(index)
+            .and_then(|shard| shard.as_ref().map(|shard| (shard.as_ptr(), shard.capacity())))
    }
 }

 #[async_trait::async_trait]
 pub(crate) trait ShardStripeSource: Send {
-    async fn read_next_stripe(&mut self) -> StripeReadState;
+    async fn read_next_stripe(&mut self) -> Box<StripeReadState>;
+
+    fn recycle_stripe(&mut self, _state: Box<StripeReadState>) {}
 }

 #[cfg(test)]
 mod tests {
    use super::*;
+    use std::mem::size_of;

    #[test]
-    fn stripe_read_state_tracks_decode_quorum() {
-        let state = StripeReadState::new(
-            vec![
-                ShardSlot::data_with_read_cost(0, ShardReadCost::Local, vec![1]),
-                ShardSlot::missing_with_read_cost(1, ShardReadCost::Remote, Error::FileNotFound),
-                ShardSlot::data_with_read_cost(2, ShardReadCost::SameNode, vec![2]),
-            ],
-            2,
-        );
+    fn stripe_scratch_capacity_matches_the_production_set_limit() {
+        type OversizedShardBuffers = SmallVec<[Option<Vec<u8>>; 32]>;
+        type OversizedShardErrors = SmallVec<[Option<Error>; 32]>;
+
+        assert_eq!(INLINE_SHARD_SLOTS, MAX_ERASURE_SET_DRIVE_COUNT);
+        assert!(size_of::<ShardBuffers>() < size_of::<OversizedShardBuffers>());
+        assert!(size_of::<ShardErrors>() < size_of::<OversizedShardErrors>());
+    }
+
+    #[test]
+    fn stripe_read_state_tracks_decode_quorum_and_slot_access() {
+        let state =
+            StripeReadState::from_parts(vec![Some(vec![1]), None, Some(vec![2])], vec![None, Some(Error::FileNotFound), None], 2);

        assert_eq!(state.available_shards(), 2);
        assert!(state.can_decode());
-        assert_eq!(state.slots()[1].index(), 1);
-        assert_eq!(state.slots()[0].read_cost(), ShardReadCost::Local);
-        assert!(state.slots()[2].read_cost().is_low_cost());
+        assert_eq!(state.data_bytes(0), Some(&[1][..]));
+        assert_eq!(state.error(1), Some(&Error::FileNotFound));
    }

    #[test]
    fn stripe_read_state_preserves_shards_and_errors() {
-        let state = StripeReadState::new(vec![ShardSlot::missing(1, Error::FileCorrupt), ShardSlot::data(0, vec![1, 2, 3])], 2);
+        let state = StripeReadState::from_parts(vec![Some(vec![1, 2, 3]), None], vec![None, Some(Error::FileCorrupt)], 2);

        assert!(!state.can_decode());
        let (shards, errors) = state.into_parts();
-        assert_eq!(shards, vec![Some(vec![1, 2, 3]), None]);
-        assert_eq!(errors, vec![None, Some(Error::FileCorrupt)]);
-    }
-
-    #[test]
-    fn stripe_read_state_builds_slots_from_parallel_reader_parts() {
-        let state =
-            StripeReadState::from_parts(vec![Some(vec![1]), None, Some(vec![3])], vec![None, Some(Error::FileNotFound)], 2);
-
-        assert!(state.can_decode());
-        assert_eq!(state.slots()[1].index(), 1);
-        assert_eq!(state.slots()[1].error(), Some(&Error::FileNotFound));
-    }
-
-    #[test]
-    fn stripe_read_state_preserves_read_cost_hints() {
-        let state = StripeReadState::from_parts_with_read_costs(
-            vec![Some(vec![1]), None, Some(vec![3])],
-            vec![None, Some(Error::FileNotFound)],
-            &[ShardReadCost::Local, ShardReadCost::Remote, ShardReadCost::Unknown],
-            2,
-        );
-
-        assert_eq!(state.slots()[0].read_cost(), ShardReadCost::Local);
-        assert_eq!(state.slots()[1].read_cost(), ShardReadCost::Remote);
-        assert_eq!(state.slots()[2].read_cost(), ShardReadCost::Unknown);
-        assert_eq!(ShardReadCost::SameNode.as_str(), GET_SHARD_READ_COST_SAME_NODE);
+        assert_eq!(shards.as_slice(), &[Some(vec![1, 2, 3]), None]);
+        assert_eq!(errors.as_slice(), &[None, Some(Error::FileCorrupt)]);
    }

    #[test]
@@ -257,8 +201,8 @@ mod tests {
        let state = StripeReadState::from_parts(vec![Some(vec![1]), Some(vec![2]), None], Vec::new(), 2);

        assert!(state.data_shards_complete(2));
-        assert_eq!(state.slots()[0].data_bytes(), Some(&[1][..]));
-        assert_eq!(state.slot_by_index(1).and_then(ShardSlot::data_bytes), Some(&[2][..]));
+        assert_eq!(state.data_bytes(0), Some(&[1][..]));
+        assert_eq!(state.data_bytes(1), Some(&[2][..]));
    }

    #[test]
@@ -267,12 +211,4 @@ mod tests {

        assert!(!state.data_shards_complete(2));
    }
-
-    #[test]
-    fn stripe_read_state_finds_out_of_order_slots_by_index() {
-        let state = StripeReadState::new(vec![ShardSlot::data(2, vec![3]), ShardSlot::data(0, vec![1])], 2);
-
-        assert_eq!(state.slot_by_index(0).and_then(ShardSlot::data_bytes), Some(&[1][..]));
-        assert!(state.slot_by_index(1).is_none());
-    }
 }
@@ -23,6 +23,7 @@ use crate::set_disk::get_lock_acquire_timeout;
 use crate::storage_api_contracts::bucket::{BUCKET_LIFECYCLE_LOCK_OBJECT, SRBucketDeleteOp};
 use crate::storage_api_contracts::namespace::NamespaceLocking as _;
 use futures::stream::{self, StreamExt};
+use rustfs_policy::policy::BucketPolicy;
 use std::collections::BTreeMap;
 use std::future::Future;

@@ -153,6 +154,31 @@ where
 }

 impl ECStore {
+    pub async fn get_bucket_metadata(&self, bucket: &str) -> Result<Arc<BucketMetadata>> {
+        let sys = metadata_sys::require_bucket_metadata_sys_in(&self.ctx)?;
+        sys.read().await.get(bucket).await
+    }
+
+    pub async fn get_bucket_policy(&self, bucket: &str) -> Result<(BucketPolicy, OffsetDateTime)> {
+        let sys = metadata_sys::require_bucket_metadata_sys_in(&self.ctx)?;
+        sys.read().await.get_bucket_policy(bucket).await
+    }
+
+    pub async fn get_bucket_policy_raw(&self, bucket: &str) -> Result<(String, OffsetDateTime)> {
+        let sys = metadata_sys::require_bucket_metadata_sys_in(&self.ctx)?;
+        sys.read().await.get_bucket_policy_raw(bucket).await
+    }
+
+    pub async fn restricts_public_bucket_access(&self, bucket: &str) -> Result<bool> {
+        let sys = metadata_sys::require_bucket_metadata_sys_in(&self.ctx)?;
+        let (config, _) = sys.read().await.get_public_access_block_config(bucket).await?;
+        Ok(config.restrict_public_buckets.unwrap_or(false))
+    }
+
+    pub async fn update_bucket_metadata_config(&self, bucket: &str, config_file: &str, data: Vec<u8>) -> Result<OffsetDateTime> {
+        metadata_sys::update_in(&self.ctx, bucket, config_file, data).await
+    }
+
    pub async fn bucket_incarnation_id(&self, bucket: &str) -> Result<Uuid> {
        metadata_sys::get_cached_bucket_incarnation_id_in(&self.ctx, bucket).await
    }
@@ -431,7 +457,13 @@ impl ECStore {
            None
        };

-        let mut meta = existing_metadata.unwrap_or_else(|| BucketMetadata::new(bucket));
+        let mut meta = existing_metadata.unwrap_or_else(|| {
+            if confirmed_missing && !is_meta_bucketname(bucket) {
+                BucketMetadata::new_with_default_durability(bucket)
+            } else {
+                BucketMetadata::new(bucket)
+            }
+        });
        let existing_incarnation_is_authoritative = meta.bucket_incarnation_sidecar;
        if confirmed_missing || is_meta_bucketname(bucket) {
            meta.set_created(opts.created_at);
@@ -1077,6 +1109,26 @@ mod tests {
        (temp_dir, ecstore)
    }

+    #[tokio::test]
+    async fn request_metadata_methods_fail_closed_before_instance_initialization() {
+        let (_temp_dir, store) = setup_multi_pool_scanner_listing_test_env().await;
+
+        let expected = "bucket metadata sys not initialized for this instance";
+        let errors = [
+            store.get_bucket_metadata("bucket").await.unwrap_err(),
+            store.get_bucket_policy("bucket").await.unwrap_err(),
+            store.get_bucket_policy_raw("bucket").await.unwrap_err(),
+            store.restricts_public_bucket_access("bucket").await.unwrap_err(),
+            store
+                .update_bucket_metadata_config("bucket", crate::bucket::metadata::BUCKET_POLICY_CONFIG, Vec::new())
+                .await
+                .unwrap_err(),
+        ];
+        for error in errors {
+            assert_eq!(error.to_string(), format!("Io error: {expected}"));
+        }
+    }
+
    async fn create_bucket_with_object(ecstore: &Arc<ECStore>, bucket: &str, object: &str) {
        let generation_before_make = ecstore.scanner_namespace_mutation_generation();
        ecstore
@@ -1511,6 +1563,76 @@ mod tests {
        );
    }

+    #[tokio::test(flavor = "multi_thread")]
+    #[serial]
+    async fn make_bucket_seeds_new_bucket_durability_override() {
+        temp_env::async_with_vars([(crate::bucket::durability::ENV_NEW_BUCKET_DURABILITY_MODE, None::<&str>)], async {
+            let (_disk_paths, ecstore) = setup_bucket_delete_test_env().await;
+            let bucket = format!("bucket-default-durability-{}", Uuid::new_v4().simple());
+
+            ecstore
+                .make_bucket(&bucket, &MakeBucketOptions::default())
+                .await
+                .expect("new bucket should be created");
+
+            let metadata = metadata_sys::get_in(&ecstore.ctx, &bucket)
+                .await
+                .expect("metadata should load for the new bucket");
+            assert_eq!(
+                metadata.durability_config().and_then(|cfg| cfg.normalized_mode()).as_deref(),
+                Some(crate::bucket::durability::BUCKET_DURABILITY_MODE_RELAXED)
+            );
+        })
+        .await;
+    }
+
+    #[tokio::test(flavor = "multi_thread")]
+    #[serial]
+    async fn force_create_existing_bucket_keeps_durability_override() {
+        let (_disk_paths, ecstore) = setup_bucket_delete_test_env().await;
+        let bucket = format!("bucket-force-durability-{}", Uuid::new_v4().simple());
+
+        temp_env::async_with_vars([(crate::bucket::durability::ENV_NEW_BUCKET_DURABILITY_MODE, Some("inherit"))], async {
+            ecstore
+                .make_bucket(&bucket, &MakeBucketOptions::default())
+                .await
+                .expect("plain bucket should be created without a durability override");
+        })
+        .await;
+        assert!(
+            metadata_sys::get_in(&ecstore.ctx, &bucket)
+                .await
+                .expect("metadata should load after initial create")
+                .durability_config()
+                .is_none(),
+            "test setup: the existing bucket must start without an override"
+        );
+
+        temp_env::async_with_vars([(crate::bucket::durability::ENV_NEW_BUCKET_DURABILITY_MODE, None::<&str>)], async {
+            ecstore
+                .make_bucket(
+                    &bucket,
+                    &MakeBucketOptions {
+                        force_create: true,
+                        lock_enabled: true,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("force create should update existing bucket metadata");
+        })
+        .await;
+
+        let metadata = metadata_sys::get_in(&ecstore.ctx, &bucket)
+            .await
+            .expect("metadata should load after force create");
+        assert!(metadata.lock_enabled, "force create sanity check: Object Lock should be enabled");
+        assert!(
+            metadata.durability_config().is_none(),
+            "force create must not apply the new-bucket default to existing bucket metadata"
+        );
+    }
+
    /// `DeleteBucket`'s emptiness check is a raw disk scan (`has_xlmeta_files`),
    /// not an S3-level listing, so "the client drained the bucket" and "the
    /// bucket is deletable" are two different contracts. Nothing pinned the
@@ -161,6 +161,10 @@ impl BucketIncarnationFenceGuard {
    pub(crate) fn is_lock_lost(&self) -> bool {
        self.inner.as_ref().is_some_and(NamespaceLockGuard::is_lock_lost)
    }
+
+    pub(crate) fn namespace_lock_guard(&self) -> Option<&NamespaceLockGuard> {
+        self.inner.as_ref()
+    }
 }

 impl Drop for BucketIncarnationFenceGuard {
@@ -309,9 +309,17 @@ const ENV_API_LIST_OBJECTS_INDEX_PROVIDER: &str = "RUSTFS_LIST_OBJECTS_INDEX_PRO
 const ENV_API_LIST_OBJECTS_INDEX_PROVIDER_PATH: &str = "RUSTFS_LIST_OBJECTS_INDEX_PROVIDER_PATH";
 const ENV_API_LIST_OBJECTS_INDEX_PROVIDER_GENERATION: &str = "RUSTFS_LIST_OBJECTS_INDEX_PROVIDER_GENERATION";
 const ENV_API_LIST_OBJECTS_NAMESPACE_JOURNAL_PATH: &str = "RUSTFS_LIST_OBJECTS_NAMESPACE_JOURNAL_PATH";
+// The chaos machinery below is compiled only for tests and the opt-in
+// `list-chaos` feature (backlog#1832): a production binary without the
+// feature carries no chaos symbols, so the two env vars cannot silently
+// rewrite a bucket's namespace-journal state.
+#[cfg(any(test, feature = "list-chaos"))]
 const ENV_API_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_ENABLED: &str = "RUSTFS_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_ENABLED";
+#[cfg(any(test, feature = "list-chaos"))]
 const ENV_API_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_BUCKET: &str = "RUSTFS_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_BUCKET";
+#[cfg(any(test, feature = "list-chaos"))]
 const ENV_API_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_SEQUENCE: &str = "RUSTFS_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_SEQUENCE";
+#[cfg(any(test, feature = "list-chaos"))]
 const ENV_API_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_STATUS: &str = "RUSTFS_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_STATUS";
 const ENV_API_LIST_OBJECTS_METADATA_FAST_ENABLED: &str = "RUSTFS_LIST_OBJECTS_METADATA_FAST_ENABLED";
 const ENV_API_LIST_OBJECTS_METADATA_FAST_STALENESS_MS: &str = "RUSTFS_LIST_OBJECTS_METADATA_FAST_STALENESS_MS";
@@ -552,7 +560,9 @@ static LIST_OBJECTS_MUTATION_SEQUENCE: AtomicU64 = AtomicU64::new(0);
 static SCANNER_NAMESPACE_MUTATION_GENERATION: AtomicU64 = AtomicU64::new(0);
 static LIST_OBJECTS_BUCKET_MUTATION_SEQUENCE: OnceCell<RwLock<HashMap<String, u64>>> = OnceCell::const_new();
 static LIST_OBJECTS_NAMESPACE_JOURNAL_DEGRADED_BUCKETS: OnceCell<RwLock<HashSet<String>>> = OnceCell::const_new();
+#[cfg(any(test, feature = "list-chaos"))]
 static LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_CONFIG: OnceCell<Option<NamespaceMutationJournalChaosConfig>> = OnceCell::const_new();
+#[cfg(any(test, feature = "list-chaos"))]
 static LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_APPLIED: OnceCell<RwLock<HashSet<String>>> = OnceCell::const_new();

 async fn persistent_key_only_index_cache() -> &'static RwLock<Option<PersistentKeyOnlyIndexCache>> {
@@ -579,6 +589,7 @@ async fn list_objects_namespace_journal_degraded_buckets() -> &'static RwLock<Ha
        .await
 }

+#[cfg(any(test, feature = "list-chaos"))]
 async fn list_objects_namespace_journal_chaos_config() -> Option<&'static NamespaceMutationJournalChaosConfig> {
    LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_CONFIG
        .get_or_init(|| async { namespace_mutation_journal_chaos_config_from_env() })
@@ -586,6 +597,7 @@ async fn list_objects_namespace_journal_chaos_config() -> Option<&'static Namesp
        .as_ref()
 }

+#[cfg(any(test, feature = "list-chaos"))]
 async fn list_objects_namespace_journal_chaos_applied() -> &'static RwLock<HashSet<String>> {
    LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_APPLIED
        .get_or_init(|| async { RwLock::new(HashSet::new()) })
@@ -681,6 +693,7 @@ enum NamespaceMutationJournalStatus {
 }

 impl NamespaceMutationJournalStatus {
+    #[cfg(any(test, feature = "list-chaos"))]
    fn from_env_value(value: &str) -> Option<Self> {
        if value.eq_ignore_ascii_case(LIST_OBJECTS_NAMESPACE_JOURNAL_STATUS_HEALTHY) {
            Some(Self::Healthy)
@@ -691,6 +704,7 @@ impl NamespaceMutationJournalStatus {
        }
    }

+    #[cfg(any(test, feature = "list-chaos"))]
    fn env_value(self) -> &'static str {
        match self {
            Self::Healthy => LIST_OBJECTS_NAMESPACE_JOURNAL_STATUS_HEALTHY,
@@ -712,6 +726,7 @@ struct NamespaceMutationJournalSnapshot {
    degraded: bool,
 }

+#[cfg(any(test, feature = "list-chaos"))]
 #[derive(Debug, Clone, PartialEq, Eq)]
 struct NamespaceMutationJournalChaosConfig {
    bucket: String,
@@ -795,30 +810,35 @@ fn list_objects_namespace_journal_root_from_env() -> Option<PathBuf> {
        .filter(|path| !path.as_os_str().is_empty())
 }

+#[cfg(any(test, feature = "list-chaos"))]
 fn namespace_mutation_journal_chaos_enabled_from_env() -> bool {
    std::env::var(ENV_API_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_ENABLED)
        .ok()
        .is_some_and(|value| value == "1" || value.eq_ignore_ascii_case("on") || value.eq_ignore_ascii_case("true"))
 }

+#[cfg(any(test, feature = "list-chaos"))]
 fn namespace_mutation_journal_chaos_bucket_from_env() -> Option<String> {
    std::env::var(ENV_API_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_BUCKET)
        .ok()
        .filter(|bucket| !bucket.is_empty())
 }

+#[cfg(any(test, feature = "list-chaos"))]
 fn namespace_mutation_journal_chaos_sequence_from_env() -> Option<u64> {
    std::env::var(ENV_API_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_SEQUENCE)
        .ok()
        .and_then(|value| value.parse::<u64>().ok())
 }

+#[cfg(any(test, feature = "list-chaos"))]
 fn namespace_mutation_journal_chaos_status_from_env() -> Option<NamespaceMutationJournalStatus> {
    std::env::var(ENV_API_LIST_OBJECTS_NAMESPACE_JOURNAL_CHAOS_STATUS)
        .ok()
        .and_then(|value| NamespaceMutationJournalStatus::from_env_value(&value))
 }

+#[cfg(any(test, feature = "list-chaos"))]
 fn namespace_mutation_journal_chaos_config_from_env() -> Option<NamespaceMutationJournalChaosConfig> {
    if !namespace_mutation_journal_chaos_enabled_from_env() {
        return None;
@@ -846,6 +866,7 @@ fn namespace_mutation_journal_chaos_config_from_env() -> Option<NamespaceMutatio
    })
 }

+#[cfg(any(test, feature = "list-chaos"))]
 fn namespace_mutation_journal_chaos_applied_key(bucket: &str, status: NamespaceMutationJournalStatus) -> String {
    let mut key = String::with_capacity(bucket.len() + 1 + status.env_value().len());
    key.push_str(bucket);
@@ -854,6 +875,13 @@ fn namespace_mutation_journal_chaos_applied_key(bucket: &str, status: NamespaceM
    key
 }

+/// Production no-op twin of the chaos injector: without `list-chaos` the
+/// injection point compiles to nothing (backlog#1832).
+#[cfg(not(any(test, feature = "list-chaos")))]
+#[inline]
+async fn maybe_apply_system_namespace_mutation_journal_chaos(_store: &ECStore, _bucket: &str, _default_sequence: u64) {}
+
+#[cfg(any(test, feature = "list-chaos"))]
 async fn maybe_apply_system_namespace_mutation_journal_chaos(store: &ECStore, bucket: &str, default_sequence: u64) {
    let Some(config) = list_objects_namespace_journal_chaos_config().await else {
        return;
@@ -9529,294 +9557,6 @@ mod test {
            .expect("a partial outage with a healthy set must not fail the walk");
    }

-    // use std::sync::Arc;
-
-    // use crate::cache_value::metacache_set::list_path_raw;
-    // use crate::cache_value::metacache_set::ListPathRawOptions;
-    // use crate::disk::endpoint::Endpoint;
-    // use crate::disk::error::is_err_eof;
-    // use crate::disk::format::FormatV3;
-    // use crate::disk::new_disk;
-    // use crate::disk::DiskAPI;
-    // use crate::disk::DiskOption;
-    // use crate::disk::MetaCacheEntries;
-    // use crate::disk::MetaCacheEntry;
-    // use crate::disk::WalkDirOptions;
-    // use crate::layout::endpoints::EndpointServerPools;
-    // use crate::error::Error;
-    // use crate::metacache::writer::MetacacheReader;
-    // use crate::set_disk::SetDisks;
-    // use crate::store::list_objects::ListPathOptions;
-    // use crate::store::list_objects::WalkOptions;
-    // use crate::store::list_objects::WalkVersionsSortOrder;
-    // use futures::future::join_all;
-    // use rustfs_lock::namespace_lock::NsLockMap;
-    // use tokio::sync::broadcast;
-    // use tokio::sync::mpsc;
-    // use tokio::sync::RwLock;
-    // use uuid::Uuid;
-
-    // #[tokio::test]
-    // async fn test_walk_dir() {
-    //     let mut ep = Endpoint::try_from("/Users/weisd/project/weisd/s3-rustfs/target/volume/test").unwrap();
-    //     ep.pool_idx = 0;
-    //     ep.set_idx = 0;
-    //     ep.disk_idx = 0;
-    //     ep.is_local = true;
-
-    //     let disk = new_disk(&ep, &DiskOption::default()).await.expect("init disk fail");
-
-    //     // let disk = match LocalDisk::new(&ep, false).await {
-    //     //     Ok(res) => res,
-    //     //     Err(err) => {
-    //     //         println!("LocalDisk::new err {:?}", err);
-    //     //         return;
-    //     //     }
-    //     // };
-
-    //     let (rd, mut wr) = tokio::io::duplex(64);
-
-    //     let job = tokio::spawn(async move {
-    //         let opts = WalkDirOptions {
-    //             bucket: "dada".to_owned(),
-    //             base_dir: "".to_owned(),
-    //             recursive: true,
-    //             ..Default::default()
-    //         };
-
-    //         println!("walk opts {:?}", opts);
-    //         if let Err(err) = disk.walk_dir(opts, &mut wr).await {
-    //             println!("walk_dir err {:?}", err);
-    //         }
-    //     });
-
-    //     let job2 = tokio::spawn(async move {
-    //         let mut mrd = MetacacheReader::new(rd);
-
-    //         loop {
-    //             match mrd.peek().await {
-    //                 Ok(res) => {
-    //                     if let Some(info) = res {
-    //                         println!("info {:?}", info.name)
-    //                     } else {
-    //                         break;
-    //                     }
-    //                 }
-    //                 Err(err) => {
-    //                     if is_err_eof(&err) {
-    //                         break;
-    //                     }
-
-    //                     println!("get err {:?}", err);
-    //                     break;
-    //                 }
-    //             }
-    //         }
-    //     });
-    //     join_all(vec![job, job2]).await;
-    // }
-
-    // #[tokio::test]
-    // async fn test_list_path_raw() {
-    //     let mut ep = Endpoint::try_from("/Users/weisd/project/weisd/s3-rustfs/target/volume/test").unwrap();
-    //     ep.pool_idx = 0;
-    //     ep.set_idx = 0;
-    //     ep.disk_idx = 0;
-    //     ep.is_local = true;
-
-    //     let disk = new_disk(&ep, &DiskOption::default()).await.expect("init disk fail");
-
-    //     // let disk = match LocalDisk::new(&ep, false).await {
-    //     //     Ok(res) => res,
-    //     //     Err(err) => {
-    //     //         println!("LocalDisk::new err {:?}", err);
-    //     //         return;
-    //     //     }
-    //     // };
-
-    //     let (_, rx) = broadcast::channel(1);
-    //     let bucket = "dada".to_owned();
-    //     let forward_to = None;
-    //     let disks = vec![Some(disk)];
-    //     let fallback_disks = Vec::new();
-
-    //     list_path_raw(
-    //         rx,
-    //         ListPathRawOptions {
-    //             disks,
-    //             fallback_disks,
-    //             bucket,
-    //             path: "".to_owned(),
-    //             recursice: true,
-    //             forward_to,
-    //             min_disks: 1,
-    //             report_not_found: false,
-    //             agreed: Some(Box::new(move |entry: MetaCacheEntry| {
-    //                 Box::pin(async move { println!("get entry: {}", entry.name) })
-    //             })),
-    //             partial: Some(Box::new(move |entries: MetaCacheEntries, _: &[Option<Error>]| {
-    //                 Box::pin(async move { println!("get entries: {:?}", entries) })
-    //             })),
-    //             finished: None,
-    //             ..Default::default()
-    //         },
-    //     )
-    //     .await
-    //     .unwrap();
-    // }
-
-    // #[tokio::test]
-    // async fn test_set_list_path() {
-    //     let mut ep = Endpoint::try_from("/Users/weisd/project/weisd/s3-rustfs/target/volume/test").unwrap();
-    //     ep.pool_idx = 0;
-    //     ep.set_idx = 0;
-    //     ep.disk_idx = 0;
-    //     ep.is_local = true;
-
-    //     let disk = new_disk(&ep, &DiskOption::default()).await.expect("init disk fail");
-    //     let _ = disk.set_disk_id(Some(Uuid::new_v4())).await;
-
-    //     let set = SetDisks {
-    //         lockers: Vec::new(),
-    //         locker_owner: String::new(),
-    //         ns_mutex: Arc::new(RwLock::new(NsLockMap::new(false))),
-    //         disks: RwLock::new(vec![Some(disk)]),
-    //         set_endpoints: Vec::new(),
-    //         set_drive_count: 1,
-    //         default_parity_count: 0,
-    //         set_index: 0,
-    //         pool_index: 0,
-    //         format: FormatV3::new(1, 1),
-    //     };
-
-    //     let (_tx, rx) = broadcast::channel(1);
-
-    //     let bucket = "dada".to_owned();
-
-    //     let opts = ListPathOptions {
-    //         bucket,
-    //         recursive: true,
-    //         ..Default::default()
-    //     };
-
-    //     let (sender, mut recv) = mpsc::channel(10);
-
-    //     set.list_path(rx, opts, sender).await.unwrap();
-
-    //     while let Some(entry) = recv.recv().await {
-    //         println!("get entry {:?}", entry.name)
-    //     }
-    // }
-
-    // #[tokio::test]
-    //walk() {
-    //     let server_address = "localhost:9000";
-
-    //     let (endpoint_pools, _setup_type) = EndpointServerPools::from_volumes(
-    //         server_address,
-    //         vec!["/Users/weisd/project/weisd/s3-rustfs/target/volume/test".to_string()],
-    //     )
-    //     .unwrap();
-
-    //     let store = ECStore::new(server_address.to_string(), endpoint_pools.clone())
-    //         .await
-    //         .unwrap();
-
-    //     let (_tx, rx) = broadcast::channel(1);
-
-    //     let bucket = "dada".to_owned();
-    //     let opts = ListPathOptions {
-    //         bucket,
-    //         recursive: true,
-    //         ..Default::default()
-    //     };
-
-    //     let (sender, mut recv) = mpsc::channel(10);
-
-    //     store.list_merged(rx, opts, sender).await.unwrap();
-
-    //     while let Some(entry) = recv.recv().await {
-    //         println!("get entry {:?}", entry.name)
-    //     }
-    // }
-
-    // #[tokio::test]
-    // async fn test_list_path() {
-    //     let server_address = "localhost:9000";
-
-    //     let (endpoint_pools, _setup_type) = EndpointServerPools::from_volumes(
-    //         server_address,
-    //         vec!["/Users/weisd/project/weisd/s3-rustfs/target/volume/test".to_string()],
-    //     )
-    //     .unwrap();
-
-    //     let store = ECStore::new(server_address.to_string(), endpoint_pools.clone())
-    //         .await
-    //         .unwrap();
-
-    //     let bucket = "dada".to_owned();
-    //     let opts = ListPathOptions {
-    //         bucket,
-    //         recursive: true,
-    //         limit: 100,
-
-    //         ..Default::default()
-    //     };
-
-    //     let ret = store.list_path(&opts).await.unwrap();
-    //     println!("ret {:?}", ret);
-    // }
-
-    // #[tokio::test]
-    // async fn test_list_objects_v2() {
-    //     let server_address = "localhost:9000";
-
-    //     let (endpoint_pools, _setup_type) = EndpointServerPools::from_volumes(
-    //         server_address,
-    //         vec!["/Users/weisd/project/weisd/s3-rustfs/target/volume/test".to_string()],
-    //     )
-    //     .unwrap();
-
-    //     let store = ECStore::new(server_address.to_string(), endpoint_pools.clone())
-    //         .await
-    //         .unwrap();
-
-    //     let ret = store.list_objects_v2("data", "", "", "", 100, false, "").await.unwrap();
-    //     println!("ret {:?}", ret);
-    // }
-
-    // #[tokio::test]
-    // async fn test_walk() {
-    //     let server_address = "localhost:9000";
-
-    //     let (endpoint_pools, _setup_type) = EndpointServerPools::from_volumes(
-    //         server_address,
-    //         vec!["/Users/weisd/project/weisd/s3-rustfs/target/volume/test".to_string()],
-    //     )
-    //     .unwrap();
-
-    //     let store = ECStore::new(server_address.to_string(), endpoint_pools.clone())
-    //         .await
-    //         .unwrap();
-
-    //     ECStore::init(store.clone()).await.unwrap();
-
-    //     let (_tx, rx) = broadcast::channel(1);
-
-    //     let bucket = ".rustfs.sys";
-    //     let prefix = "config/iam/sts/";
-
-    //     let (sender, mut recv) = mpsc::channel(10);
-
-    //     let opts = WalkOptions::default();
-
-    //     store.walk(rx, bucket, prefix, sender, opts).await.unwrap();
-
-    //     while let Some(entry) = recv.recv().await {
-    //         println!("get entry {:?}", entry)
-    //     }
-    // }
-
    #[tokio::test]
    async fn merge_entry_channels_produces_sorted_unique_output_from_two_channels() {
        let (tx_a, rx_a) = mpsc::channel(4);
@@ -389,7 +389,7 @@ impl crate::storage_api_contracts::object::ObjectIO for ECStore {
    type GetObjectReader = GetObjectReader;
    type PutObjectReader = PutObjReader;

-    #[instrument(level = "debug", skip(self))]
+    #[instrument(level = "debug", skip(self, h))]
    async fn get_object_reader(
        &self,
        bucket: &str,
@@ -66,6 +66,76 @@ fn ensure_multipart_bucket_lifecycle_guard_held(
    Ok(())
 }

+#[cfg(test)]
+struct DataMovementMultipartCompletionBarrierState {
+    bucket: String,
+    arrived: tokio::sync::Notify,
+    release: tokio::sync::Notify,
+}
+
+#[cfg(test)]
+pub(crate) struct DataMovementMultipartCompletionBarrier {
+    state: Arc<DataMovementMultipartCompletionBarrierState>,
+}
+
+#[cfg(test)]
+static DATA_MOVEMENT_MULTIPART_COMPLETION_BARRIER: std::sync::OnceLock<
+    std::sync::Mutex<Option<Arc<DataMovementMultipartCompletionBarrierState>>>,
+> = std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl DataMovementMultipartCompletionBarrier {
+    pub(crate) fn install(bucket: &str) -> Self {
+        let state = Arc::new(DataMovementMultipartCompletionBarrierState {
+            bucket: bucket.to_string(),
+            arrived: tokio::sync::Notify::new(),
+            release: tokio::sync::Notify::new(),
+        });
+        let mut slot = DATA_MOVEMENT_MULTIPART_COMPLETION_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("data movement multipart completion barrier mutex should not poison");
+        assert!(slot.is_none(), "data movement multipart completion barrier must be unique");
+        *slot = Some(Arc::clone(&state));
+        Self { state }
+    }
+
+    pub(crate) async fn wait_until_paused(&self) {
+        tokio::time::timeout(std::time::Duration::from_secs(30), self.state.arrived.notified())
+            .await
+            .expect("data movement multipart operation should reach selected completion");
+    }
+}
+
+#[cfg(test)]
+impl Drop for DataMovementMultipartCompletionBarrier {
+    fn drop(&mut self) {
+        self.state.release.notify_one();
+        let mut slot = DATA_MOVEMENT_MULTIPART_COMPLETION_BARRIER
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("data movement multipart completion barrier mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+async fn pause_data_movement_multipart_before_selected_completion(bucket: &str) {
+    let barrier = DATA_MOVEMENT_MULTIPART_COMPLETION_BARRIER
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("data movement multipart completion barrier mutex should not poison")
+        .as_ref()
+        .filter(|barrier| barrier.bucket == bucket)
+        .cloned();
+    if let Some(barrier) = barrier {
+        barrier.arrived.notify_one();
+        barrier.release.notified().await;
+    }
+}
+
 async fn list_pool_multipart_uploads_for_incarnation(
    pool: &crate::core::sets::Sets,
    bucket: &str,
@@ -332,7 +402,7 @@ impl ECStore {
    ) -> Result<MultipartUploadResult> {
        self.handle_new_multipart_upload_with_pool_idx(bucket, object, opts)
            .await
-            .map(|(res, _)| res)
+            .map(|(res, _, _)| res)
    }

    pub(crate) async fn handle_new_multipart_upload_with_pool_idx(
@@ -340,7 +410,7 @@ impl ECStore {
        bucket: &str,
        object: &str,
        opts: &ObjectOptions,
-    ) -> Result<(MultipartUploadResult, usize)> {
+    ) -> Result<(MultipartUploadResult, usize, Option<Uuid>)> {
        check_new_multipart_args(bucket, object)?;
        let (opts, _bucket_lifecycle_guard) = self.guard_multipart_bucket_incarnation(bucket, opts).await?;
        let opts = &opts;
@@ -349,7 +419,20 @@ impl ECStore {
            return self.pools[0]
                .new_multipart_upload(bucket, object, opts)
                .await
-                .map(|res| (res, 0));
+                .map(|res| (res, 0, opts.expected_bucket_incarnation_id));
+        }
+
+        if opts.data_movement && opts.version_id.is_some() {
+            let idx = self.select_data_movement_pool_idx(bucket, object, -1, opts, false).await?;
+            if idx == opts.src_pool_idx {
+                return Err(StorageError::DataMovementOverwriteErr(
+                    bucket.to_owned(),
+                    object.to_owned(),
+                    opts.version_id.clone().unwrap_or_default(),
+                ));
+            }
+            let res = self.pools[idx].new_multipart_upload(bucket, object, opts).await?;
+            return Ok((res, idx, opts.expected_bucket_incarnation_id));
        }

        for (idx, pool) in self.pools.iter().enumerate() {
@@ -372,7 +455,7 @@ impl ECStore {

            if !res.uploads.is_empty() {
                let res = self.pools[idx].new_multipart_upload(bucket, object, opts).await?;
-                return Ok((res, idx));
+                return Ok((res, idx, opts.expected_bucket_incarnation_id));
            }
        }
        let idx = self.get_pool_idx(bucket, object, -1).await?;
@@ -385,7 +468,7 @@ impl ECStore {
        }

        let res = self.pools[idx].new_multipart_upload(bucket, object, opts).await?;
-        Ok((res, idx))
+        Ok((res, idx, opts.expected_bucket_incarnation_id))
    }

    #[instrument(skip(self))]
@@ -456,6 +539,30 @@ impl ECStore {
        Err(StorageError::InvalidUploadID(bucket.to_owned(), object.to_owned(), upload_id.to_owned()))
    }

+    pub(crate) async fn put_object_part_for_data_movement(
+        &self,
+        target_pool_idx: usize,
+        bucket: &str,
+        object: &str,
+        upload_id: &str,
+        data: &mut PutObjReader,
+        opts: &ObjectOptions,
+    ) -> Result<PartInfo> {
+        let part_id = opts
+            .part_number
+            .ok_or_else(|| Error::other("targeted multipart upload requires a part number"))?;
+        check_put_object_part_args(bucket, object, upload_id)?;
+        if !opts.data_movement {
+            return Err(Error::other("targeted multipart upload requires data_movement options"));
+        }
+        let (opts, _bucket_lifecycle_guard) = self.guard_multipart_bucket_incarnation(bucket, opts).await?;
+        let pool = self
+            .pools
+            .get(target_pool_idx)
+            .ok_or_else(|| Error::other(format!("data movement target pool {target_pool_idx} is out of range")))?;
+        pool.put_object_part(bucket, object, upload_id, part_id, data, &opts).await
+    }
+
    #[instrument(skip(self))]
    pub(super) async fn handle_get_multipart_info(
        &self,
@@ -530,6 +637,26 @@ impl ECStore {
        Err(StorageError::InvalidUploadID(bucket.to_owned(), object.to_owned(), upload_id.to_owned()))
    }

+    pub(crate) async fn abort_multipart_upload_for_data_movement(
+        &self,
+        target_pool_idx: usize,
+        bucket: &str,
+        object: &str,
+        upload_id: &str,
+        opts: &ObjectOptions,
+    ) -> Result<()> {
+        check_abort_multipart_args(bucket, object, upload_id)?;
+        if !opts.data_movement {
+            return Err(Error::other("targeted multipart abort requires data_movement options"));
+        }
+        let (opts, _bucket_lifecycle_guard) = self.guard_multipart_bucket_incarnation(bucket, opts).await?;
+        let pool = self
+            .pools
+            .get(target_pool_idx)
+            .ok_or_else(|| Error::other(format!("data movement target pool {target_pool_idx} is out of range")))?;
+        pool.abort_multipart_upload(bucket, object, upload_id, &opts).await
+    }
+
    #[instrument(skip(self))]
    pub(super) async fn handle_complete_multipart_upload(
        self: Arc<Self>,
@@ -574,6 +701,62 @@ impl ECStore {

        Err(StorageError::InvalidUploadID(bucket.to_owned(), object.to_owned(), upload_id.to_owned()))
    }
+
+    pub(crate) async fn complete_multipart_upload_for_data_movement(
+        self: Arc<Self>,
+        target_pool_idx: usize,
+        bucket: &str,
+        object: &str,
+        upload_id: &str,
+        uploaded_parts: Vec<CompletePart>,
+        opts: &ObjectOptions,
+    ) -> Result<ObjectInfo> {
+        check_complete_multipart_args(bucket, object, upload_id)?;
+        if !opts.data_movement {
+            return Err(Error::other("targeted multipart completion requires data_movement options"));
+        }
+        let (mut opts, _bucket_lifecycle_guard) = self.guard_multipart_bucket_incarnation(bucket, opts).await?;
+        if opts.overwrites_existing_version() && !is_meta_bucketname(bucket) {
+            let expected_incarnation_id = opts
+                .expected_bucket_incarnation_id
+                .ok_or_else(|| Error::other("data movement completion is missing its bucket incarnation"))?;
+            let lifecycle_fence = opts
+                .bucket_lifecycle_lock_fence
+                .as_ref()
+                .ok_or_else(|| Error::other("data movement completion is missing its bucket lifecycle fence"))?;
+            let snapshot = match opts.object_lock_config_snapshot.as_ref() {
+                Some(snapshot) => Arc::clone(snapshot),
+                None => {
+                    self.object_lock_config_snapshot_under_lifecycle_fence(bucket, lifecycle_fence)
+                        .await?
+                }
+            };
+            if !snapshot.is_valid_for_destructive_put(self.id, bucket, expected_incarnation_id) {
+                return Err(Error::other(
+                    "data movement Object Lock snapshot does not match the target bucket generation",
+                ));
+            }
+            snapshot.add_lock_fences(&mut opts);
+            opts.object_lock_config_snapshot = Some(snapshot);
+        }
+        #[cfg(test)]
+        pause_data_movement_multipart_before_selected_completion(bucket).await;
+        let pool = self
+            .pools
+            .get(target_pool_idx)
+            .ok_or_else(|| Error::other(format!("data movement target pool {target_pool_idx} is out of range")))?
+            .clone();
+        let result = enqueue_transition_after_write(
+            pool.complete_multipart_upload(bucket, object, upload_id, uploaded_parts, &opts)
+                .await,
+            LcEventSrc::S3CompleteMultipartUpload,
+        )
+        .await;
+        if result.is_ok() {
+            list_objects::observe_list_objects_mutation(self.as_ref(), bucket).await;
+        }
+        result
+    }
 }

 /// Merges per-pool `ListMultipartUploads` pages into a single globally paginated
@@ -667,18 +667,6 @@ impl SelectObjectSnapshotLockLossWake {
    }
 }

-fn select_object_ssec_headers(headers: &HeaderMap) -> HeaderMap {
-    use rustfs_utils::http::headers::{SSEC_ALGORITHM_HEADER, SSEC_KEY_HEADER, SSEC_KEY_MD5_HEADER};
-
-    let mut selected = HeaderMap::new();
-    for name in [SSEC_ALGORITHM_HEADER, SSEC_KEY_HEADER, SSEC_KEY_MD5_HEADER] {
-        if let Some(value) = headers.get(name) {
-            selected.insert(name, value.clone());
-        }
-    }
-    selected
-}
-
 // LockRegistry clones its canonical client Arc for each endpoint host, so an
 // exact Arc set identifies one distributed namespace-lock quorum domain.
 fn same_distributed_lock_domain(left: &[Arc<dyn rustfs_lock::LockClient>], right: &[Arc<dyn rustfs_lock::LockClient>]) -> bool {
@@ -918,7 +906,7 @@ fn is_equivalent_data_movement_delete_marker(source: &ObjectInfo, target: &Objec
        && is_data_movement_delete_marker(target)
        && source.version_id == target.version_id
        && source.mod_time == target.mod_time
-        && source.user_defined == target.user_defined
+        && is_equivalent_data_movement_delete_marker_metadata(&source.user_defined, &target.user_defined)
        && source.user_tags == target.user_tags
        && source.replication_status_internal == target.replication_status_internal
        && source.replication_status == target.replication_status
@@ -926,24 +914,185 @@ fn is_equivalent_data_movement_delete_marker(source: &ObjectInfo, target: &Objec
        && source.version_purge_status == target.version_purge_status
 }

+fn is_equivalent_data_movement_delete_marker_metadata(
+    source: &HashMap<String, String>,
+    target: &HashMap<String, String>,
+) -> bool {
+    matches!(
+        (
+            data_movement_delete_marker_metadata_identity(source),
+            data_movement_delete_marker_metadata_identity(target)
+        ),
+        (Some(source), Some(target)) if source == target
+    )
+}
+
+fn data_movement_delete_marker_metadata_identity(metadata: &HashMap<String, String>) -> Option<HashMap<String, String>> {
+    let mut identity = HashMap::with_capacity(metadata.len());
+    let mut local_tier_free_version_id = None;
+    for (key, value) in metadata {
+        let Some(suffix) = rustfs_utils::http::strip_internal_prefix_preserving_case(key) else {
+            identity.insert(key.clone(), value.clone());
+            continue;
+        };
+
+        if suffix.eq_ignore_ascii_case(rustfs_utils::http::SUFFIX_TIER_FV_ID) {
+            let version_id = Uuid::parse_str(value).ok().filter(|version_id| !version_id.is_nil())?;
+            if local_tier_free_version_id.is_some_and(|expected| expected != version_id) {
+                return None;
+            }
+            local_tier_free_version_id = Some(version_id);
+            continue;
+        }
+
+        let canonical_suffix = [
+            rustfs_utils::http::SUFFIX_REPLICA_TIMESTAMP,
+            rustfs_utils::http::SUFFIX_REPLICA_STATUS,
+            rustfs_utils::http::SUFFIX_REPLICATION_TIMESTAMP,
+            rustfs_utils::http::SUFFIX_REPLICATION_STATUS,
+            rustfs_utils::http::SUFFIX_PURGESTATUS,
+        ]
+        .into_iter()
+        .find(|candidate| suffix.eq_ignore_ascii_case(candidate))
+        .map(str::to_string)
+        .or_else(|| {
+            [
+                rustfs_utils::http::SUFFIX_REPLICATION_RESET_ARN_PREFIX,
+                rustfs_utils::http::SUFFIX_REPLICATION_DELETE_MARKER_VERSION_ARN_PREFIX,
+            ]
+            .into_iter()
+            .find_map(|prefix| {
+                suffix
+                    .get(..prefix.len())
+                    .is_some_and(|candidate| candidate.eq_ignore_ascii_case(prefix))
+                    .then(|| format!("{prefix}{}", &suffix[prefix.len()..]))
+            })
+        })
+        .unwrap_or_else(|| suffix.to_string());
+        let canonical_value = if canonical_suffix.eq_ignore_ascii_case(rustfs_utils::http::SUFFIX_REPLICA_TIMESTAMP)
+            || canonical_suffix.eq_ignore_ascii_case(rustfs_utils::http::SUFFIX_REPLICATION_TIMESTAMP)
+        {
+            rustfs_filemeta::parse_replication_timestamp(value)?
+                .unix_timestamp_nanos()
+                .to_string()
+        } else {
+            value.clone()
+        };
+        let canonical_key = format!("{}{canonical_suffix}", rustfs_utils::http::RUSTFS_INTERNAL_PREFIX);
+        if identity
+            .insert(canonical_key, canonical_value.clone())
+            .is_some_and(|existing| existing != canonical_value)
+        {
+            return None;
+        }
+    }
+    for (status_suffix, timestamp_suffix) in [
+        (rustfs_utils::http::SUFFIX_REPLICA_STATUS, rustfs_utils::http::SUFFIX_REPLICA_TIMESTAMP),
+        (
+            rustfs_utils::http::SUFFIX_REPLICATION_STATUS,
+            rustfs_utils::http::SUFFIX_REPLICATION_TIMESTAMP,
+        ),
+    ] {
+        let status_key = format!("{}{status_suffix}", rustfs_utils::http::RUSTFS_INTERNAL_PREFIX);
+        let timestamp_key = format!("{}{timestamp_suffix}", rustfs_utils::http::RUSTFS_INTERNAL_PREFIX);
+        match (identity.contains_key(&status_key), identity.contains_key(&timestamp_key)) {
+            (true, false) => {
+                identity.insert(timestamp_key, OffsetDateTime::UNIX_EPOCH.unix_timestamp_nanos().to_string());
+            }
+            (false, true) => return None,
+            _ => {}
+        }
+    }
+    Some(identity)
+}
+
 fn is_data_movement_delete_marker(info: &ObjectInfo) -> bool {
    info.delete_marker
 }

+fn is_expected_data_movement_delete_marker_source(source: &ObjectInfo, expected_mod_time: Option<OffsetDateTime>) -> bool {
+    is_data_movement_delete_marker(source)
+        && source.mod_time.is_some()
+        && source.mod_time == expected_mod_time
+        && data_movement_delete_marker_metadata_identity(&source.user_defined).is_some()
+}
+
+fn current_data_movement_delete_marker_opts(source: &ObjectInfo, opts: &ObjectOptions) -> Option<ObjectOptions> {
+    let replica_status = rustfs_utils::http::get_str(&source.user_defined, rustfs_utils::http::SUFFIX_REPLICA_STATUS);
+    let replica_timestamp = rustfs_utils::http::get_str(&source.user_defined, rustfs_utils::http::SUFFIX_REPLICA_TIMESTAMP);
+    let (replica_status, replica_timestamp) = match (replica_status, replica_timestamp) {
+        (None, None) => Default::default(),
+        (Some(status), timestamp) => {
+            let status = crate::bucket::replication::ReplicationStatusType::from(status.as_str());
+            if status.is_empty() {
+                return None;
+            }
+            let timestamp = match timestamp {
+                Some(timestamp) => rustfs_filemeta::parse_replication_timestamp(&timestamp)?,
+                None => OffsetDateTime::UNIX_EPOCH,
+            };
+            (status, Some(timestamp))
+        }
+        (None, Some(_)) => return None,
+    };
+    let replication_status = rustfs_utils::http::get_str(&source.user_defined, rustfs_utils::http::SUFFIX_REPLICATION_STATUS);
+    let replication_timestamp =
+        rustfs_utils::http::get_str(&source.user_defined, rustfs_utils::http::SUFFIX_REPLICATION_TIMESTAMP);
+    let (replication_status, replication_timestamp, replication_targets) = match (replication_status, replication_timestamp) {
+        (None, None) => Default::default(),
+        (Some(status), timestamp) => {
+            let direct_status = crate::bucket::replication::ReplicationStatusType::from(status.as_str());
+            let targets = crate::bucket::replication::replication_statuses_map(status.as_str());
+            if direct_status.is_empty() && targets.is_empty() {
+                return None;
+            }
+            let timestamp = match timestamp {
+                Some(timestamp) => rustfs_filemeta::parse_replication_timestamp(&timestamp)?,
+                None => OffsetDateTime::UNIX_EPOCH,
+            };
+            (Some(status), Some(timestamp), targets)
+        }
+        (None, Some(_)) => return None,
+    };
+    let mut state = source.replication_state();
+    if state.target_delete_marker_version_ids_corrupt {
+        return None;
+    }
+    state.replica_status = replica_status;
+    state.replica_timestamp = replica_timestamp;
+    state.replication_status_internal = replication_status;
+    state.replication_timestamp = replication_timestamp;
+    state.targets = replication_targets;
+    state.replicate_decision_str = source.replication_decision.clone();
+    state.delete_marker = true;
+
+    let mut target_opts = opts.clone();
+    target_opts.mod_time = source.mod_time;
+    target_opts.delete_replication = Some(state);
+    Some(target_opts)
+}
+
 fn expected_data_movement_tiered_object(source: &rustfs_filemeta::FileInfo) -> ObjectInfo {
    ObjectInfo::from_file_info(source, "", &source.name, source.version_id.is_some())
 }

 fn is_equivalent_data_movement_tiered_object(source: &rustfs_filemeta::FileInfo, target: &ObjectInfo) -> bool {
    let expected = expected_data_movement_tiered_object(source);
+    let Some(source_actual_size) = effective_object_actual_size(&expected) else {
+        return false;
+    };
+    let Some(target_actual_size) = effective_object_actual_size(target) else {
+        return false;
+    };

    source.version_id == target.version_id
        && !target.delete_marker
        && source.size == target.size
        && source.get_etag() == target.etag
        && source.checksum == target.checksum
+        && crate::data_movement::are_equivalent_data_movement_parts(&source.parts, &target.parts)
        && source.mod_time == target.mod_time
-        && expected.user_defined == target.user_defined
+        && crate::data_movement::is_equivalent_data_movement_metadata(&expected, target, source_actual_size, target_actual_size)
        && expected.user_tags == target.user_tags
        && expected.expires == target.expires
        && expected.storage_class == target.storage_class
@@ -952,11 +1101,12 @@ fn is_equivalent_data_movement_tiered_object(source: &rustfs_filemeta::FileInfo,
        && expected.version_purge_status_internal == target.version_purge_status_internal
        && expected.version_purge_status == target.version_purge_status
        && expected.transitioned_object.status == target.transitioned_object.status
+        && expected.transition_version_state == target.transition_version_state
        && expected.transitioned_object.name == target.transitioned_object.name
        && expected.transitioned_object.tier == target.transitioned_object.tier
        && expected.transitioned_object.version_id == target.transitioned_object.version_id
        && expected.transitioned_object.free_version == target.transitioned_object.free_version
-        && effective_object_actual_size(target) == Some(source.size)
+        && source_actual_size == target_actual_size
 }

 fn should_check_data_movement_resume_target(src_pool_idx: usize, target_pool_idx: usize) -> bool {
@@ -1054,7 +1204,7 @@ impl ECStore {
        )))
    }

-    async fn object_lock_config_snapshot_under_lifecycle_fence(
+    pub(super) async fn object_lock_config_snapshot_under_lifecycle_fence(
        &self,
        bucket: &str,
        lifecycle_fence: &NamespaceLockFence,
@@ -1134,7 +1284,7 @@ impl ECStore {
            pool,
            bucket: bucket.to_owned(),
            object,
-            headers: select_object_ssec_headers(headers),
+            headers: rustfs_utils::http::project_ssec_transport_headers(headers),
            opts,
            object_info,
            logical_size,
@@ -1454,7 +1604,8 @@ impl ECStore {
        target_pool_idx: usize,
        opts: &ObjectOptions,
    ) -> Result<Option<ObjectInfo>> {
-        let lookup_opts = version_aware_lookup_opts(opts, true);
+        let mut lookup_opts = version_aware_lookup_opts(opts, true);
+        lookup_opts.include_part_checksums = true;

        let Some(pool) = self.pools.get(target_pool_idx) else {
            return Err(Error::other(format!(
@@ -1521,6 +1672,25 @@ impl ECStore {
    ) -> Result<()> {
        check_put_object_args(bucket, object)?;

+        let mut opts = opts.clone();
+        let bucket_incarnation_fence = if is_meta_bucketname(bucket) {
+            None
+        } else {
+            let expected = opts
+                .expected_bucket_incarnation_id
+                .ok_or_else(|| Error::other("tiered data movement is missing its bucket incarnation snapshot"))?;
+            let guard = self.acquire_bucket_incarnation_fence(bucket, expected).await?;
+            if let Some(namespace_guard) = guard.namespace_lock_guard() {
+                opts.add_bucket_lifecycle_lock_guard(namespace_guard);
+            }
+            Some(guard)
+        };
+
+        let mut fi = fi.clone();
+        if opts.data_movement {
+            crate::data_movement::prepare_tiered_data_movement_file_info(&mut fi)?;
+        }
+
        let object = encode_dir_object(object);

        if self.single_pool() {
@@ -1533,7 +1703,8 @@ impl ECStore {

        let idx = if opts.data_movement && opts.version_id.is_some() {
            Self::resolve_decommission_target_pool_idx_result(
-                self.select_data_movement_pool_idx(bucket, &object, fi.size, opts, true).await,
+                self.select_data_movement_pool_idx(bucket, &object, fi.size, &opts, true)
+                    .await,
                bucket,
                &object,
            )?
@@ -1550,7 +1721,7 @@ impl ECStore {
                .await;
            let target_pool_idx = resolve_data_movement_resume_target_pool(idx, resume_target_pool_idx, opts.src_pool_idx);
            if self
-                .has_equivalent_data_movement_tiered_object(bucket, &object, fi, opts, target_pool_idx)
+                .has_equivalent_data_movement_tiered_object(bucket, &object, &fi, &opts, target_pool_idx)
                .await?
            {
                return Ok(());
@@ -1563,17 +1734,30 @@ impl ECStore {
            ));
        }

-        Self::resolve_decommission_tiered_object_result(
-            self.pools[idx]
-                .get_disks_by_key(&object)
-                .decommission_tiered_object(bucket, &object, fi, opts)
-                .await,
-            bucket,
-            &object,
-        )
+        let result = self.pools[idx]
+            .get_disks_by_key(&object)
+            .decommission_tiered_object(bucket, &object, &fi, &opts)
+            .await;
+        if matches!(result, Err(Error::PreconditionFailed)) {
+            if self
+                .has_equivalent_data_movement_tiered_object(bucket, &object, &fi, &opts, idx)
+                .await?
+            {
+                return Ok(());
+            }
+            return Err(StorageError::DataMovementOverwriteErr(
+                bucket.to_owned(),
+                object,
+                opts.version_id.clone().unwrap_or_default(),
+            ));
+        }
+        if bucket_incarnation_fence.as_ref().is_some_and(|guard| guard.is_lock_lost()) {
+            return Err(Error::other("tiered data movement bucket incarnation fence was lost during target write"));
+        }
+        Self::resolve_decommission_tiered_object_result(result, bucket, &object)
    }

-    #[instrument(level = "debug", skip(self))]
+    #[instrument(level = "debug", skip(self, h))]
    #[hotpath::measure(impl_type = "ECStore")]
    pub(super) async fn handle_get_object_reader(
        &self,
@@ -1585,7 +1769,7 @@ impl ECStore {
    ) -> Result<GetObjectReader> {
        check_get_obj_args(bucket, object)?;

-        let object = encode_dir_object(object);
+        let object = rustfs_utils::path::encode_dir_object_ref(object);
        let mut opts = opts.clone();
        let read_lock_guard = self
            .acquire_object_read_lock_if_needed("get_object", bucket, &object, &mut opts)
@@ -1593,29 +1777,21 @@ impl ECStore {

        let reader = if self.single_pool() {
            self.pools[0]
-                .get_object_reader(bucket, object.as_str(), range, h, &opts)
+                .get_object_reader(bucket, object.as_ref(), range, h, &opts)
                .await?
        } else {
            let (_, idx) = self
                .get_latest_accessible_object_info_with_idx(bucket, &object, &opts)
                .await?;
            self.pools[idx]
-                .get_object_reader(bucket, object.as_str(), range, h, &opts)
+                .get_object_reader(bucket, object.as_ref(), range, h, &opts)
                .await?
        };

        Ok(Self::attach_read_lock_guard(reader, read_lock_guard))
    }

-    #[instrument(level = "debug", skip(self, data))]
-    #[hotpath::measure(impl_type = "ECStore")]
-    pub(super) async fn handle_put_object(
-        &self,
-        bucket: &str,
-        object: &str,
-        data: &mut PutObjReader,
-        opts: &ObjectOptions,
-    ) -> Result<(ObjectInfo, Option<OldCurrentSize>)> {
+    async fn prepare_put_object(&self, bucket: &str, object: &str, opts: &ObjectOptions) -> Result<(String, ObjectOptions)> {
        check_put_object_args(bucket, object)?;

        let object = encode_dir_object(object);
@@ -1642,22 +1818,20 @@ impl ECStore {
            };
            snapshot.add_lock_fences(&mut opts);
        }
+        Ok((object, opts))
+    }

-        // Keep PUT atomic-read friendly: SetDisks takes the object write lock only
-        // around precondition checks and the final rename/commit.
+    async fn select_put_object_pool_idx(&self, bucket: &str, object: &str, size: i64, opts: &ObjectOptions) -> Result<usize> {
        if self.single_pool() {
-            return self.pools[0]
-                .put_object_with_old_current_size(bucket, object.as_str(), data, &opts)
-                .await;
+            return Ok(0);
        }

        let idx = if opts.data_movement && opts.version_id.is_some() {
-            self.select_data_movement_pool_idx(bucket, &object, data.size(), &opts, false)
-                .await?
+            self.select_data_movement_pool_idx(bucket, object, size, opts, false).await?
        } else if opts.no_lock {
-            self.get_pool_idx_no_lock(bucket, &object, data.size()).await?
+            self.get_pool_idx_no_lock(bucket, object, size).await?
        } else {
-            self.get_pool_idx(bucket, &object, data.size()).await?
+            self.get_pool_idx(bucket, object, size).await?
        };

        if opts.data_movement && idx == opts.src_pool_idx {
@@ -1667,7 +1841,50 @@ impl ECStore {
                opts.version_id.clone().unwrap_or_default(),
            ));
        }
+        Ok(idx)
+    }

+    pub(crate) async fn put_object_for_data_movement(
+        &self,
+        bucket: &str,
+        object: &str,
+        data: &mut PutObjReader,
+        opts: &ObjectOptions,
+    ) -> Result<(usize, Result<ObjectInfo>)> {
+        if !opts.data_movement {
+            return Err(Error::other("data movement PUT requires data_movement options"));
+        }
+        let (object, opts) = self.prepare_put_object(bucket, object, opts).await?;
+        let idx = self
+            .select_put_object_pool_idx(bucket, object.as_str(), data.size(), &opts)
+            .await?;
+        let result = self.pools[idx]
+            .put_object_with_old_current_size(bucket, &object, data, &opts)
+            .await
+            .map(|(object_info, _)| object_info);
+        let result = enqueue_transition_after_write(result, LcEventSrc::S3PutObject).await;
+        if result.is_ok() {
+            list_objects::observe_list_objects_mutation(self, bucket).await;
+        }
+        Ok((idx, result))
+    }
+
+    #[instrument(level = "debug", skip(self, data))]
+    #[hotpath::measure(impl_type = "ECStore")]
+    pub(super) async fn handle_put_object(
+        &self,
+        bucket: &str,
+        object: &str,
+        data: &mut PutObjReader,
+        opts: &ObjectOptions,
+    ) -> Result<(ObjectInfo, Option<OldCurrentSize>)> {
+        let (object, opts) = self.prepare_put_object(bucket, object, opts).await?;
+        let idx = self
+            .select_put_object_pool_idx(bucket, object.as_str(), data.size(), &opts)
+            .await?;
+
+        // Keep PUT atomic-read friendly: SetDisks takes the object write lock only
+        // around precondition checks and the final rename/commit.
        self.pools[idx]
            .put_object_with_old_current_size(bucket, &object, data, &opts)
            .await
@@ -2110,6 +2327,50 @@ impl ECStore {
            };
            let target_pool_idx =
                resolve_data_movement_resume_target_pool(selected_target_pool_idx, resume_target_pool_idx, opts.src_pool_idx);
+            let mut delete_marker_target_opts = None;
+
+            if opts.delete_marker && should_check_data_movement_resume_target(opts.src_pool_idx, target_pool_idx) {
+                let source = self
+                    .find_data_movement_target_info(bucket, object, opts.src_pool_idx, &opts)
+                    .await?;
+                let Some(source) = source else {
+                    return Err(StorageError::DataMovementOverwriteErr(
+                        bucket.to_owned(),
+                        object.to_owned(),
+                        opts.version_id.unwrap_or_default(),
+                    ));
+                };
+                if !is_expected_data_movement_delete_marker_source(&source, opts.mod_time) {
+                    return Err(StorageError::DataMovementOverwriteErr(
+                        bucket.to_owned(),
+                        object.to_owned(),
+                        opts.version_id.unwrap_or_default(),
+                    ));
+                }
+                let Some(target_opts) = current_data_movement_delete_marker_opts(&source, &opts) else {
+                    return Err(StorageError::DataMovementOverwriteErr(
+                        bucket.to_owned(),
+                        object.to_owned(),
+                        opts.version_id.unwrap_or_default(),
+                    ));
+                };
+                let target = self
+                    .find_data_movement_target_info(bucket, object, target_pool_idx, &target_opts)
+                    .await?;
+                if let Some(target) = target {
+                    if is_equivalent_data_movement_delete_marker(&source, &target) {
+                        let mut target = target;
+                        target.name = decode_dir_object(object);
+                        return Ok(target);
+                    }
+                    return Err(StorageError::DataMovementOverwriteErr(
+                        bucket.to_owned(),
+                        object.to_owned(),
+                        opts.version_id.unwrap_or_default(),
+                    ));
+                }
+                delete_marker_target_opts = Some(target_opts);
+            }

            if !should_check_data_movement_resume_target(opts.src_pool_idx, target_pool_idx) {
                if let Ok((source_pool_info, _)) = existing_pool_info
@@ -2137,7 +2398,8 @@ impl ECStore {
                ));
            }

-            let mut obj = self.pools[target_pool_idx].delete_object(bucket, object, opts).await?;
+            let target_opts = delete_marker_target_opts.unwrap_or(opts);
+            let mut obj = self.pools[target_pool_idx].delete_object(bucket, object, target_opts).await?;
            obj.name = decode_dir_object(obj.name.as_str());
            return Ok(obj);
        }
@@ -3086,25 +3348,6 @@ mod tests {
        assert!(second_signal.is_lost());
    }

-    #[test]
-    fn select_snapshot_retains_only_ssec_headers() {
-        use rustfs_utils::http::headers::{SSEC_ALGORITHM_HEADER, SSEC_KEY_HEADER, SSEC_KEY_MD5_HEADER};
-
-        let mut headers = HeaderMap::new();
-        headers.insert(SSEC_ALGORITHM_HEADER, "AES256".parse().expect("valid SSE-C algorithm header"));
-        headers.insert(SSEC_KEY_HEADER, "secret-key".parse().expect("valid SSE-C key header"));
-        headers.insert(SSEC_KEY_MD5_HEADER, "key-md5".parse().expect("valid SSE-C key digest header"));
-        headers.insert("authorization", "credential".parse().expect("valid authorization header"));
-
-        let selected = select_object_ssec_headers(&headers);
-
-        assert_eq!(selected.len(), 3);
-        assert_eq!(selected.get(SSEC_ALGORITHM_HEADER), headers.get(SSEC_ALGORITHM_HEADER));
-        assert_eq!(selected.get(SSEC_KEY_HEADER), headers.get(SSEC_KEY_HEADER));
-        assert_eq!(selected.get(SSEC_KEY_MD5_HEADER), headers.get(SSEC_KEY_MD5_HEADER));
-        assert!(selected.get("authorization").is_none());
-    }
-
    #[test]
    fn tier_delete_entry_is_prepared_and_bound_to_source_generation() {
        let identity = [9_u8; 32];
@@ -3200,6 +3443,247 @@ mod tests {
        assert!(!is_equivalent_data_movement_delete_marker(&source, &mismatched));
    }

+    #[test]
+    fn equivalent_data_movement_delete_marker_accepts_distinct_local_free_version_ids() {
+        let mut source = ObjectInfo {
+            version_id: Some(Uuid::from_u128(1)),
+            delete_marker: true,
+            mod_time: Some(OffsetDateTime::UNIX_EPOCH),
+            ..Default::default()
+        };
+        rustfs_utils::http::insert_str(
+            Arc::make_mut(&mut source.user_defined),
+            rustfs_utils::http::SUFFIX_TIER_FV_ID,
+            Uuid::from_u128(2).to_string(),
+        );
+        let mut target = source.clone();
+        rustfs_utils::http::insert_str(
+            Arc::make_mut(&mut target.user_defined),
+            rustfs_utils::http::SUFFIX_TIER_FV_ID,
+            Uuid::from_u128(3).to_string(),
+        );
+
+        assert!(is_equivalent_data_movement_delete_marker(&source, &target));
+
+        Arc::make_mut(&mut target.user_defined).insert(
+            format!("{}{}", rustfs_utils::http::MINIO_INTERNAL_PREFIX, rustfs_utils::http::SUFFIX_TIER_FV_ID),
+            Uuid::from_u128(4).to_string(),
+        );
+        assert!(!is_equivalent_data_movement_delete_marker(&source, &target));
+    }
+
+    #[test]
+    fn equivalent_data_movement_delete_marker_accepts_replication_alias_expansion() {
+        let key = format!(
+            "{}{}",
+            rustfs_utils::http::MINIO_INTERNAL_PREFIX,
+            rustfs_utils::http::SUFFIX_REPLICATION_STATUS
+        );
+        let timestamp_key = format!(
+            "{}{}",
+            rustfs_utils::http::MINIO_INTERNAL_PREFIX,
+            rustfs_utils::http::SUFFIX_REPLICATION_TIMESTAMP
+        );
+        let source = ObjectInfo {
+            version_id: Some(Uuid::from_u128(1)),
+            delete_marker: true,
+            mod_time: Some(OffsetDateTime::UNIX_EPOCH),
+            user_defined: Arc::new(HashMap::from([
+                (key.clone(), "arn=COMPLETED;".to_string()),
+                (timestamp_key, "1970-01-01T00:00:01Z".to_string()),
+            ])),
+            ..Default::default()
+        };
+        let mut target = source.clone();
+        rustfs_utils::http::insert_str(
+            Arc::make_mut(&mut target.user_defined),
+            rustfs_utils::http::SUFFIX_REPLICATION_STATUS,
+            "arn=COMPLETED;".to_string(),
+        );
+        rustfs_utils::http::insert_str(
+            Arc::make_mut(&mut target.user_defined),
+            rustfs_utils::http::SUFFIX_REPLICATION_TIMESTAMP,
+            (OffsetDateTime::UNIX_EPOCH + time::Duration::SECOND).to_string(),
+        );
+        assert!(is_equivalent_data_movement_delete_marker(&source, &target));
+
+        Arc::make_mut(&mut target.user_defined).insert(key, "arn=FAILED;".to_string());
+        assert!(!is_equivalent_data_movement_delete_marker(&source, &target));
+    }
+
+    #[test]
+    fn data_movement_delete_marker_source_requires_persisted_mod_time() {
+        let source = ObjectInfo {
+            delete_marker: true,
+            ..Default::default()
+        };
+        assert!(!is_expected_data_movement_delete_marker_source(&source, None));
+
+        let source = ObjectInfo {
+            mod_time: Some(OffsetDateTime::UNIX_EPOCH),
+            ..source
+        };
+        assert!(is_expected_data_movement_delete_marker_source(&source, Some(OffsetDateTime::UNIX_EPOCH)));
+        assert!(!is_expected_data_movement_delete_marker_source(&source, None));
+    }
+
+    #[test]
+    fn data_movement_delete_marker_uses_current_source_replication_state() {
+        let expected_timestamp = OffsetDateTime::UNIX_EPOCH + time::Duration::SECOND;
+        let timestamp = expected_timestamp.to_string();
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_REPLICA_STATUS,
+            ReplicationStatusType::Replica.to_string(),
+        );
+        rustfs_utils::http::insert_str(&mut metadata, rustfs_utils::http::SUFFIX_REPLICA_TIMESTAMP, timestamp.clone());
+        rustfs_utils::http::insert_str(&mut metadata, rustfs_utils::http::SUFFIX_REPLICATION_TIMESTAMP, timestamp);
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_REPLICATION_STATUS,
+            "arn=COMPLETED;".to_string(),
+        );
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            &format!(
+                "{}{}",
+                rustfs_utils::http::SUFFIX_REPLICATION_RESET_ARN_PREFIX,
+                "arn:minio:replication::TenantA:bucket"
+            ),
+            "reset-id".to_string(),
+        );
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            &format!(
+                "{}{}",
+                rustfs_utils::http::SUFFIX_REPLICATION_DELETE_MARKER_VERSION_ARN_PREFIX,
+                "arn:minio:replication::TenantA:bucket"
+            ),
+            "target-version".to_string(),
+        );
+        let source = ObjectInfo {
+            delete_marker: true,
+            mod_time: Some(OffsetDateTime::UNIX_EPOCH),
+            replication_status_internal: Some("arn=COMPLETED;".to_string()),
+            replication_decision: "arn=replicate;".to_string(),
+            user_defined: Arc::new(metadata),
+            ..Default::default()
+        };
+        let opts = ObjectOptions {
+            mod_time: source.mod_time,
+            delete_replication: Some(ReplicationState {
+                replication_status_internal: Some("arn=PENDING;".to_string()),
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+
+        let target_opts = current_data_movement_delete_marker_opts(&source, &opts).expect("valid current source state");
+        let state = target_opts.delete_replication.as_ref().expect("current replication state");
+        assert_eq!(state.replication_status_internal.as_deref(), Some("arn=COMPLETED;"));
+        assert_eq!(state.replica_status, crate::bucket::replication::ReplicationStatusType::Replica);
+        assert_eq!(state.replica_timestamp, Some(expected_timestamp));
+        assert_eq!(state.replication_timestamp, state.replica_timestamp);
+        assert_eq!(state.replicate_decision_str, "arn=replicate;");
+        assert_eq!(
+            state
+                .reset_statuses_map
+                .get("arn:minio:replication::TenantA:bucket")
+                .map(String::as_str),
+            Some("reset-id")
+        );
+        assert_eq!(
+            state
+                .target_delete_marker_version_ids
+                .get("arn:minio:replication::TenantA:bucket")
+                .map(String::as_str),
+            Some("target-version")
+        );
+    }
+
+    #[test]
+    fn data_movement_delete_marker_rejects_corrupt_target_version_maps() {
+        let suffix = format!("{}not-an-arn", rustfs_utils::http::SUFFIX_REPLICATION_DELETE_MARKER_VERSION_ARN_PREFIX);
+        let mut malformed = HashMap::new();
+        rustfs_utils::http::insert_str(&mut malformed, &suffix, "target-version".to_string());
+        let malformed_source = ObjectInfo {
+            delete_marker: true,
+            mod_time: Some(OffsetDateTime::UNIX_EPOCH),
+            user_defined: Arc::new(malformed),
+            ..Default::default()
+        };
+        assert!(current_data_movement_delete_marker_opts(&malformed_source, &ObjectOptions::default()).is_none());
+
+        let mut conflicted = HashMap::new();
+        let suffix = format!(
+            "{}arn:minio:replication::target:bucket",
+            rustfs_utils::http::SUFFIX_REPLICATION_DELETE_MARKER_VERSION_ARN_PREFIX
+        );
+        rustfs_utils::http::insert_str(&mut conflicted, &suffix, "target-version-a".to_string());
+        conflicted.insert(
+            format!("{}{suffix}", rustfs_utils::http::MINIO_INTERNAL_PREFIX),
+            "target-version-b".to_string(),
+        );
+        let conflicted_source = ObjectInfo {
+            user_defined: Arc::new(conflicted),
+            ..malformed_source.clone()
+        };
+        assert!(current_data_movement_delete_marker_opts(&conflicted_source, &ObjectOptions::default()).is_none());
+
+        let mut over_cap = HashMap::new();
+        for index in 0..=1_000 {
+            let suffix = format!(
+                "{}arn:minio:replication::target:bucket-{index}",
+                rustfs_utils::http::SUFFIX_REPLICATION_DELETE_MARKER_VERSION_ARN_PREFIX
+            );
+            rustfs_utils::http::insert_str(&mut over_cap, &suffix, format!("target-version-{index}"));
+        }
+        let over_cap_source = ObjectInfo {
+            user_defined: Arc::new(over_cap),
+            ..malformed_source
+        };
+        assert!(current_data_movement_delete_marker_opts(&over_cap_source, &ObjectOptions::default()).is_none());
+    }
+
+    #[test]
+    fn data_movement_delete_marker_normalizes_legacy_missing_replication_timestamps() {
+        let mut source_metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut source_metadata,
+            rustfs_utils::http::SUFFIX_REPLICA_STATUS,
+            ReplicationStatusType::Replica.to_string(),
+        );
+        rustfs_utils::http::insert_str(
+            &mut source_metadata,
+            rustfs_utils::http::SUFFIX_REPLICATION_STATUS,
+            "arn=COMPLETED;".to_string(),
+        );
+        let source = ObjectInfo {
+            delete_marker: true,
+            mod_time: Some(OffsetDateTime::UNIX_EPOCH),
+            replication_status_internal: Some("arn=COMPLETED;".to_string()),
+            user_defined: Arc::new(source_metadata),
+            ..Default::default()
+        };
+
+        let target_opts = current_data_movement_delete_marker_opts(&source, &ObjectOptions::default())
+            .expect("legacy status-only metadata should remain migratable");
+        let state = target_opts
+            .delete_replication
+            .expect("replication state should be reconstructed");
+        assert_eq!(state.replica_timestamp, Some(OffsetDateTime::UNIX_EPOCH));
+        assert_eq!(state.replication_timestamp, Some(OffsetDateTime::UNIX_EPOCH));
+
+        let mut target_metadata = (*source.user_defined).clone();
+        let epoch = OffsetDateTime::UNIX_EPOCH
+            .format(&time::format_description::well_known::Rfc3339)
+            .unwrap();
+        rustfs_utils::http::insert_str(&mut target_metadata, rustfs_utils::http::SUFFIX_REPLICA_TIMESTAMP, epoch.clone());
+        rustfs_utils::http::insert_str(&mut target_metadata, rustfs_utils::http::SUFFIX_REPLICATION_TIMESTAMP, epoch);
+        assert!(is_equivalent_data_movement_delete_marker_metadata(&source.user_defined, &target_metadata));
+    }
+
    #[test]
    fn equivalent_data_movement_delete_marker_rejects_metadata_and_replication_mismatch() {
        let version_id = Uuid::nil();
@@ -3344,6 +3828,67 @@ mod tests {
        assert!(is_equivalent_data_movement_tiered_object(&source, &target));
    }

+    #[test]
+    fn equivalent_data_movement_tiered_object_uses_logical_compressed_and_encrypted_sizes() {
+        let mut compressed = tiered_equivalence_source();
+        compressed.size = 600;
+        rustfs_utils::http::insert_str(&mut compressed.metadata, rustfs_utils::http::SUFFIX_COMPRESSION, "S2".to_string());
+        rustfs_utils::http::insert_str(&mut compressed.metadata, rustfs_utils::http::SUFFIX_ACTUAL_SIZE, "1024".to_string());
+        let compressed_target = tiered_equivalence_target(&compressed);
+        assert!(is_equivalent_data_movement_tiered_object(&compressed, &compressed_target));
+
+        let mut encrypted = tiered_equivalence_source();
+        encrypted.size = 640;
+        encrypted.metadata.insert(
+            rustfs_utils::http::object_encryption_keys::INTERNAL_ENCRYPTION_KEY_ID_HEADER.to_string(),
+            "key-id".to_string(),
+        );
+        encrypted.metadata.insert(
+            rustfs_utils::http::object_encryption_keys::INTERNAL_ENCRYPTION_ORIGINAL_SIZE_HEADER.to_string(),
+            "1024".to_string(),
+        );
+        let encrypted_target = tiered_equivalence_target(&encrypted);
+        assert!(is_equivalent_data_movement_tiered_object(&encrypted, &encrypted_target));
+    }
+
+    #[test]
+    fn equivalent_data_movement_tiered_object_accepts_transition_alias_expansion() {
+        let mut source = tiered_equivalence_source();
+        let suffix = rustfs_utils::http::SUFFIX_TRANSITION_TIER;
+        source.metadata.insert(
+            format!("{}{suffix}", rustfs_utils::http::MINIO_INTERNAL_PREFIX),
+            source.transition_tier.clone(),
+        );
+        let mut target = tiered_equivalence_target(&source);
+        Arc::make_mut(&mut target.user_defined)
+            .insert(rustfs_utils::http::internal_key_rustfs(suffix), source.transition_tier.clone());
+
+        assert!(is_equivalent_data_movement_tiered_object(&source, &target));
+    }
+
+    #[test]
+    fn equivalent_data_movement_tiered_object_requires_hydrated_part_checksums() {
+        let mut source = tiered_equivalence_source();
+        source.parts = vec![rustfs_filemeta::ObjectPartInfo {
+            number: 1,
+            mod_time: Some(OffsetDateTime::UNIX_EPOCH + time::Duration::SECOND),
+            checksums: Some(HashMap::from([("CRC32C".to_string(), "AAAAAA==".to_string())])),
+            ..Default::default()
+        }];
+        rustfs_utils::http::insert_str(
+            &mut source.metadata,
+            rustfs_utils::http::SUFFIX_PART_CHECKSUMS,
+            r#"[[1,[["CRC32C","AAAAAA=="]]]]"#.to_string(),
+        );
+        let mut target = tiered_equivalence_target(&source);
+        Arc::make_mut(&mut target.parts)[0].mod_time = None;
+        assert!(is_equivalent_data_movement_tiered_object(&source, &target));
+
+        let mut missing = target;
+        Arc::make_mut(&mut missing.parts)[0].checksums = None;
+        assert!(!is_equivalent_data_movement_tiered_object(&source, &missing));
+    }
+
    #[test]
    fn equivalent_data_movement_tiered_object_rejects_transition_mismatch() {
        let source = tiered_equivalence_source();
@@ -3353,6 +3898,16 @@ mod tests {
        assert!(!is_equivalent_data_movement_tiered_object(&source, &target));
    }

+    #[test]
+    fn equivalent_data_movement_tiered_object_rejects_transition_version_state_mismatch() {
+        let mut source = tiered_equivalence_source();
+        source.transition_version_state = rustfs_filemeta::TransitionVersionState::Exact;
+        let mut target = tiered_equivalence_target(&source);
+        target.transition_version_state = rustfs_filemeta::TransitionVersionState::Unknown;
+
+        assert!(!is_equivalent_data_movement_tiered_object(&source, &target));
+    }
+
    #[test]
    fn equivalent_data_movement_tiered_object_rejects_user_metadata_mismatch() {
        let source = tiered_equivalence_source();
@@ -4022,6 +4577,8 @@ mod tests {
            assert_eq!(snapshot.headers.get(SSEC_KEY_HEADER), request_headers.get(SSEC_KEY_HEADER));
            assert_eq!(snapshot.headers.get(SSEC_KEY_MD5_HEADER), request_headers.get(SSEC_KEY_MD5_HEADER));
            assert!(snapshot.headers.get("authorization").is_none());
+            assert!(snapshot.headers.values().all(http::HeaderValue::is_sensitive));
+            assert!(!format!("{:?}", snapshot.headers).contains("secret-key"));
            assert_eq!(
                snapshot.logical_size(),
                u64::try_from(payload.len()).expect("test payload length should fit in u64")
@@ -109,6 +109,7 @@ async fn run_legacy_bitrot_test_for_object(root: &std::path::Path, disk_name: &s
        FileInfoOpts {
            data: true, // need inline data for inline objects
            include_free_versions: false,
+            include_part_checksums: true,
        },
    ) {
        Ok(f) => f,
--- a/Show More
+++ b/Show More