fix(quota): account compressed deletes by committed size

chore(ecstore): standardize 24 bare TODO/FIXME to tracked format (#6346 )
feat(connect): add registration and credential rotation (#6353 )
2026-08-22 04:16:38 +00:00 · 2026-08-22 11:03:02 +08:00 · 2026-08-22 07:09:32 +08:00 · 2026-08-22 07:07:29 +08:00 · 2026-08-21 15:04:08 +00:00 · 2026-08-21 13:16:18 +00:00
92 changed files with 7801 additions and 571 deletions
@@ -0,0 +1,2 @@
+sha256-darwin=b4ae71aa894e5c7795ae3eb8116f1777a7601d0f5db3898be2e48faf3329bd9b
+sha256-linux=433debd9d9defa832986269abdf0f1d131597b2d7a417ce930e17c1fd47d85ba
@@ -0,0 +1 @@
+sha256=9b9bc336b43b70d0e06e0adb5455bf035bb18945d85d60936eb6fe4d48e0e680
@@ -0,0 +1,2 @@
+sha256-darwin=55534a97fbd376f64c8f6c341d319017d11ff77cad6da8629a1a7f6a874e0315
+sha256-linux=c06fb8c19aed6f388b9dc61cb8251b7a44f8561a9bf764ad2b9e635598f8dc17
@@ -0,0 +1 @@
+sha256=655a3f3c1d042e694339d15caba7580518320322d1bac0f09450b37e6c09e2e7
@@ -0,0 +1 @@
+sha256=ec27cde6ce6400723c4b372bfbd2ac61709c744294e4810af765e8a808d8e31d
@@ -75,6 +75,11 @@ embedded-secrets-check: ## Check no private key material or credential literal i
 	@echo "🔑 Checking embedded secret material guard..."
 	./scripts/check_embedded_secrets.sh

+.PHONY: test-wiring-check
+test-wiring-check: ## Check tests stay registered and selected by their intended runners
+	@echo "🧪 Checking test wiring..."
+	python3 ./scripts/check_test_wiring.py
+
 .PHONY: log-analyzer-rules-check
 log-analyzer-rules-check: core-deps ## Check log-analyzer rule anchors still exist verbatim in source
 	@echo "🩺 Checking log-analyzer rule anchors..."
@@ -19,13 +19,13 @@ planning-docs-check: ## Check that no planning-type documents are committed
 	./scripts/check_no_planning_docs.sh

 .PHONY: pre-commit
-pre-commit: fmt-check unsafe-code-check architecture-migration-check logging-guardrails-check tokio-io-uring-check extension-schema-check body-cache-whitelist-check s3s-footprint-check fips-wording-check embedded-secrets-check doc-paths-check planning-docs-check quick-check ## Run fast pre-commit checks without clippy/full tests
+pre-commit: fmt-check unsafe-code-check architecture-migration-check logging-guardrails-check tokio-io-uring-check extension-schema-check body-cache-whitelist-check s3s-footprint-check fips-wording-check embedded-secrets-check test-wiring-check doc-paths-check planning-docs-check quick-check ## Run fast pre-commit checks without clippy/full tests
 	@echo "✅ All pre-commit checks passed!"

 .PHONY: pre-pr
-pre-pr: fmt-check unsafe-code-check architecture-migration-check logging-guardrails-check tokio-io-uring-check extension-schema-check body-cache-whitelist-check s3s-footprint-check fips-wording-check embedded-secrets-check doc-paths-check planning-docs-check log-analyzer-rules-check clippy-check test ## Run full pre-PR checks with clippy and tests
+pre-pr: fmt-check unsafe-code-check architecture-migration-check logging-guardrails-check tokio-io-uring-check extension-schema-check body-cache-whitelist-check s3s-footprint-check fips-wording-check embedded-secrets-check test-wiring-check doc-paths-check planning-docs-check log-analyzer-rules-check clippy-check test ## Run full pre-PR checks with clippy and tests
 	@echo "✅ All pre-PR checks passed!"

 .PHONY: dev-check
-dev-check: fmt-check unsafe-code-check architecture-migration-check logging-guardrails-check tokio-io-uring-check extension-schema-check body-cache-whitelist-check s3s-footprint-check fips-wording-check embedded-secrets-check doc-paths-check planning-docs-check quick-check ## Run fast local development checks
+dev-check: fmt-check unsafe-code-check architecture-migration-check logging-guardrails-check tokio-io-uring-check extension-schema-check body-cache-whitelist-check s3s-footprint-check fips-wording-check embedded-secrets-check test-wiring-check doc-paths-check planning-docs-check quick-check ## Run fast local development checks
 	@echo "✅ Fast development checks passed!"
@@ -35,6 +35,8 @@ script-tests: ## Run shell script tests
 	./scripts/test_pinned_paired_abba_bench.sh
 	./scripts/test_manual_transition_runbooks.sh
 	./scripts/check_embedded_secrets.sh --self-test
+	python3 ./scripts/check_test_wiring.py --self-test
+	python3 ./scripts/s3-tests/test_report_compat.py
 	bash -n ./scripts/validate_object_data_cache_cold_stampede.sh
 	python3 ./scripts/check_object_data_cache_follower_samples.py --self-test
 	./scripts/validate_object_data_cache_cold_stampede.sh --self-test
@@ -38,10 +38,11 @@ e2e-vault = { max-threads = 1 }
 # 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;
+# servers never run at once. The e2e-full merge/main lane picks these up;
 # they are deliberately NOT in the fast PR `e2e-smoke` filter.
 e2e-reliability = { max-threads = 1 }
 e2e-inline-boundaries = { max-threads = 1 }
+e2e-cluster-nightly = { max-threads = 1 }

 # --- default profile (local): serialize the flaky groups, never retry --------
 [[profile.default.overrides]]
@@ -161,7 +162,7 @@ retries = 2
 # Serialize the 4-disk reliability / degraded-read e2e tests under the ci
 # profile too (see the e2e-reliability test-group note near the top). Not a
 # quarantine: no retries, just single-threaded so several 4-disk servers never
-# run concurrently when ci-7's nightly runs the full e2e suite.
+# run concurrently when e2e-full runs the suite.
 [[profile.ci.overrides]]
 filter = 'package(e2e_test) & test(/^(reliability_disk_fault|degraded_read_eof_regression|replacement_privileged_e2e)_test::/)'
 test-group = 'e2e-reliability'
@@ -230,8 +231,8 @@ test-group = 'ecstore-serial-flaky'
 # the nightly profile derives its set as "the replication module MINUS this
 # allowlist", so any new replication test lands in nightly by default (never
 # silently unrun) until it is explicitly blessed as fast here. Keep the two
-# regexes byte-identical. Count invariant: 20 here + 49 nightly = 69 total
-# (authority: `cargo nextest list`; docs/testing/e2e-suite-inventory.md).
+# regexes byte-identical. The committed profile selection digests make changes
+# visible in CI; current counts live in docs/testing/e2e-suite-inventory.md.
 # HISTORY (2026-07-11): the 20 fast tests were briefly pulled out of this lane
 # (#4724) because they set a loopback (127.0.0.1) replication target that the
 # SSRF egress guard rejected on every PR after repl-1 (#4712). That is fixed —
@@ -327,9 +328,8 @@ slow-timeout = { period = "60s", terminate-after = 2, grace-period = "10s" }
 # the STS dual-node test actually exercises its path (it skips gracefully with
 # a visible log line when awscurl is absent), and routes scheduled failures
 # through .github/actions/schedule-failure-issue (ci-8). Explicit division of
-# labor with ci-5's future e2e-full merge gate: these tests run ONLY here, not
-# double-run there. TODO(ci-7): fold this interim repl-owned lane into the ci
-# domain's consolidated scheduled e2e workflow once it exists.
+# labor with e2e-full: these tests run only in the consolidated nightly
+# workflow, not in the merge/main lane.
 [profile.e2e-repl-nightly]
 default-filter = """
    package(e2e_test)
@@ -343,26 +343,60 @@ fail-fast = false
 # workflow as the failure-triage artifact.
 path = "junit.xml"

+# ---------------------------------------------------------------------------
+# e2e-nightly profile — destructive multi-process cluster fault domains
+# ---------------------------------------------------------------------------
+# These seven modules are deliberately outside e2e-full's merge budget. Each
+# starts a real multi-process or multi-disk topology and exercises node/disk
+# loss, quorum, cleanup, notification fan-in, or admin-timeout behavior. The
+# consolidated nightly workflow runs them serially to avoid resource
+# starvation; failures are never retried.
+[profile.e2e-nightly]
+default-filter = """
+    package(e2e_test)
+    & test(/^(admin_timeout_regression_test|cluster_concurrency_test|cluster_multidrive_pool_test|heal_erasure_disk_rebuild_test|namespace_lock_quorum_test|object_lambda_test|stale_multipart_cleanup_cluster_test)::/)
+"""
+fail-fast = false
+
+[profile.e2e-nightly.junit]
+path = "junit.xml"
+
+[[profile.e2e-nightly.overrides]]
+filter = 'package(e2e_test)'
+test-group = 'e2e-cluster-nightly'
+
+# ---------------------------------------------------------------------------
+# e2e-protocols profile — serial protocol lane
+# ---------------------------------------------------------------------------
+# The suite owns fixed ports, so the nightly workflow runs this exact profile
+# with one nextest worker.
+[profile.e2e-protocols]
+default-filter = 'package(e2e_test) & test(/^protocols::/)'
+fail-fast = false
+
+[profile.e2e-protocols.junit]
+path = "junit.xml"
+
 # ---------------------------------------------------------------------------
 # e2e-full profile — merge-gate full single-node e2e lane (backlog#1149 ci-5)
 # ---------------------------------------------------------------------------
 # The merge gate (ci.yml `e2e-full` job: push main + merge_group +
-# workflow_dispatch). Runs the never-automated user-visible suites — KMS (40),
-# object_lock (33), multipart_auth (109), quota, checksum, encryption,
+# workflow_dispatch). Runs the user-visible KMS, object-lock, multipart-auth,
+# quota, checksum, encryption,
 # security-boundary, ... — that the fast PR `e2e-smoke` subset deliberately
 # skips. Budget <= 45 min; authority for the suite count is `cargo nextest list
 # --profile e2e-full` (see docs/testing/e2e-suite-inventory.md).
 #
 # The filter is "the whole e2e_test crate MINUS the sets owned by other lanes":
-#   * protocols:: — FTPS/SFTP/WebDAV, still pinned to --test-threads=1 by fixed
-#     ports; they join a scheduled lane once ci-6 randomises the ports (ci-7).
+#   * protocols:: — FTPS/SFTP/WebDAV, run from the dedicated protocol profile
+#     with one worker because the suite owns fixed ports.
 #   * the 7 cluster suites that spin up a RustFSTestClusterEnvironment
 #     (cluster_concurrency, cluster_multidrive_pool, stale_multipart_cleanup_cluster,
 #     namespace_lock_quorum, heal_erasure_disk_rebuild, admin_timeout_regression,
-#     object_lambda) — too heavy for the merge budget; they run in ci-7's
-#     nightly 4-node lane.
+#     object_lambda) — too heavy for the merge budget; they run in the
+#     e2e-nightly serial cluster-fault lane.
 #   * replication_extension_test — repl-1 already splits it into the PR
-#     `e2e-smoke` (20 fast) and `e2e-repl-nightly` (49 slow) lanes and reserves
+#     `e2e-smoke` (20 fast) and `e2e-repl-nightly` (55 slow) lanes and reserves
 #     it for those, so e2e-full does not double-run it.
 #   * #[ignore]d tests — nextest skips them by default (no --run-ignored); the
 #     manual-localhost:9000 reliant/policy tests are ci-13's migration.
@@ -46,10 +46,9 @@ lists when upstream changes.
  the PR.
 - **Weekly + manual**: `.github/workflows/e2e-s3tests.yml` runs the full
  upstream suite (`TEST_SCOPE=all`) against a Docker deployment (single node
-  or a 4-node distributed cluster behind HAProxy). It fails only on
-  regressions in the implemented whitelist and publishes a classification
-  report (`compat-report.md`, also shown in the job summary) listing promotion
-  candidates and unclassified tests.
+  or a 4-node distributed cluster behind HAProxy). The canonical gate policy
+  and compatibility-report behavior are documented in
+  [`scripts/s3-tests/README.md`](../../scripts/s3-tests/README.md).

 ## Running Tests Locally

@@ -125,6 +125,9 @@ jobs:
      - name: Check no embedded secret material
        run: ./scripts/check_embedded_secrets.sh

+      - name: Check test wiring
+        run: python3 ./scripts/check_test_wiring.py
+
      - name: Check no planning docs committed
        run: ./scripts/check_no_planning_docs.sh

@@ -160,6 +160,9 @@ jobs:
      - name: Check no embedded secret material
        run: ./scripts/check_embedded_secrets.sh

+      - name: Check test wiring
+        run: python3 ./scripts/check_test_wiring.py
+
      - name: Check no planning docs committed
        run: ./scripts/check_no_planning_docs.sh

@@ -686,9 +689,9 @@ jobs:
      - name: Make binary executable
        run: chmod +x ./target/debug/rustfs

-      # Build the e2e test graph once. The archive is reused by the security
-      # count-floor check and the smoke run below, avoiding a second compile of
-      # the same e2e_test target on cold runners (backlog#1645).
+      # Build the e2e test graph once. The archive is reused by the smoke
+      # selection guard, security exact-count check, and run below, avoiding a
+      # second compile of the same e2e_test target on cold runners (backlog#1645).
      - name: Archive e2e smoke test binaries
        env:
          NEXTEST_ARCHIVE: ${{ runner.temp }}/rustfs-e2e-smoke.tar.zst
@@ -696,6 +699,7 @@ jobs:
        run: |
          cargo nextest archive --profile e2e-smoke -p e2e_test --archive-file "${NEXTEST_ARCHIVE}"
          cargo nextest list --profile e2e-smoke --archive-file "${NEXTEST_ARCHIVE}" --message-format json > "${NEXTEST_LISTING}"
+          python3 ./scripts/check_test_wiring.py --check-profile e2e-smoke "${NEXTEST_LISTING}"
          ./scripts/check_security_smoke_count.sh check "${NEXTEST_LISTING}"

      # PR smoke subset of the in-repo e2e suite (backlog#1149 ci-4). The
@@ -760,7 +764,7 @@ jobs:
    # suites — KMS, object_lock, multipart_auth, quota, checksum, encryption,
    # security-boundary, ... — via the e2e-full nextest profile. Too heavy for
    # every PR, so it is gated to main pushes, the merge queue, and manual
-    # dispatch. protocols / the 6 cluster suites / replication / #[ignore] are
+    # dispatch. protocols / the 7 cluster suites / replication / #[ignore] are
    # owned by other lanes (see .config/nextest.toml profile.e2e-full).
    if: >-
      github.event_name == 'workflow_dispatch' ||
@@ -820,6 +824,13 @@ jobs:
      - name: Make binary executable
        run: chmod +x ./target/debug/rustfs

+      - name: Verify e2e full membership
+        env:
+          NEXTEST_LISTING: ${{ runner.temp }}/rustfs-e2e-full-list.json
+        run: |
+          cargo nextest list --profile e2e-full -p e2e_test --message-format json > "${NEXTEST_LISTING}"
+          python3 ./scripts/check_test_wiring.py --check-profile e2e-full "${NEXTEST_LISTING}"
+
      # Full single-node e2e lane (backlog#1149 ci-5). The e2e-full
      # default-filter in .config/nextest.toml is the single wiring mechanism —
      # extend that filter, never add ad-hoc e2e jobs here. Reuses the downloaded
@@ -832,7 +843,9 @@ jobs:
        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
        with:
          name: e2e-full-junit-${{ github.run_number }}
-          path: target/nextest/e2e-full/junit.xml
+          path: |
+            target/nextest/e2e-full/junit.xml
+            ${{ runner.temp }}/rustfs-e2e-full-list.json
          retention-days: 7

  e2e-tests-rio-v2:
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-# Nightly full replication e2e lane (backlog#1147 repl-1, deps: ci-4).
+# Consolidated nightly e2e lane for replication, cluster faults, and protocols.
 #
 # The per-PR gate (ci.yml `e2e-tests` job, `--profile e2e-smoke`) runs the
 # FAST replication tests. This scheduled lane runs the remaining heavier
@@ -28,15 +28,12 @@
 # add ad-hoc cargo-test steps here; change the filterset instead. The
 # authoritative membership and count come from
 # `cargo nextest list -p e2e_test --profile e2e-repl-nightly`; the PR/nightly
-# count invariant is maintained next to the filtersets in .config/nextest.toml
-# (deliberately not duplicated here).
+# selection digest is committed under .config/.
 #
-# Explicit division of labor: the nightly subset runs ONLY here, never double-run
-# in ci-5's future e2e-full merge gate. TODO(ci-7): once the ci domain's
-# consolidated scheduled e2e workflow exists, fold this interim repl-owned lane
-# into it rather than growing a second scheduled entrypoint.
+# Explicit division of labor: these subsets run only here and never double-run
+# in the e2e-full merge gate.

-name: e2e-replication-nightly
+name: e2e-nightly

 on:
  workflow_dispatch:
@@ -50,6 +47,10 @@ on:
 permissions:
  contents: read

+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: false
+
 jobs:
  repl-nightly:
    name: Replication e2e (nightly)
@@ -97,9 +98,20 @@ jobs:
      # demand otherwise, but a single explicit build avoids several parallel
      # nextest test processes racing to build it at once.
      - name: Build rustfs binary
-        run: cargo build -p rustfs --bins
+        run: |
+          cargo build -p rustfs --bins
+          : > target/debug/rustfs.features
+
+      - name: Verify replication e2e membership
+        env:
+          NEXTEST_LISTING: ${{ runner.temp }}/rustfs-e2e-repl-nightly-list.json
+        run: |
+          cargo nextest list --profile e2e-repl-nightly -p e2e_test --message-format json > "${NEXTEST_LISTING}"
+          python3 ./scripts/check_test_wiring.py --check-profile e2e-repl-nightly "${NEXTEST_LISTING}"

      - name: Run replication e2e nightly suite
+        env:
+          RUSTFS_E2E_LOG_DIR: ${{ runner.temp }}/rustfs-e2e-repl-nightly-logs
        run: cargo nextest run --profile e2e-repl-nightly -p e2e_test

      - name: Upload nextest junit report
@@ -107,13 +119,112 @@ jobs:
        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
        with:
          name: e2e-replication-nightly-junit-${{ github.run_number }}
-          path: target/nextest/e2e-repl-nightly/junit.xml
+          path: |
+            target/nextest/e2e-repl-nightly/junit.xml
+            ${{ runner.temp }}/rustfs-e2e-repl-nightly-list.json
+            ${{ runner.temp }}/rustfs-e2e-repl-nightly-logs/
          retention-days: 7
          if-no-files-found: ignore

+  cluster-nightly:
+    name: Cluster fault e2e (nightly)
+    runs-on: sm-standard-4
+    timeout-minutes: 90
+    env:
+      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+
+      - name: Setup Rust environment
+        uses: ./.github/actions/setup
+        with:
+          rust-version: stable
+          cache-shared-key: ci-e2e-nightly
+          cache-save-if: 'false'
+          install-build-packaging-tools: 'false'
+
+      - name: Build rustfs binary
+        run: |
+          cargo build -p rustfs --bins --features e2e-test-hooks
+          : > target/debug/rustfs.features
+
+      - name: Verify cluster fault e2e membership
+        env:
+          NEXTEST_LISTING: ${{ runner.temp }}/rustfs-e2e-nightly-list.json
+        run: |
+          cargo nextest list --profile e2e-nightly -p e2e_test --message-format json > "${NEXTEST_LISTING}"
+          python3 ./scripts/check_test_wiring.py --check-profile e2e-nightly "${NEXTEST_LISTING}"
+
+      - name: Run cluster fault e2e nightly suite
+        env:
+          RUSTFS_E2E_LOG_DIR: ${{ runner.temp }}/rustfs-e2e-nightly-logs
+        run: cargo nextest run --profile e2e-nightly -p e2e_test
+
+      - name: Upload cluster fault diagnostics
+        if: always()
+        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
+        with:
+          name: e2e-cluster-nightly-${{ github.run_number }}
+          path: |
+            target/nextest/e2e-nightly/junit.xml
+            ${{ runner.temp }}/rustfs-e2e-nightly-list.json
+            ${{ runner.temp }}/rustfs-e2e-nightly-logs/
+          retention-days: 7
+          if-no-files-found: warn
+
+  protocols-nightly:
+    name: Protocol e2e (nightly)
+    runs-on: sm-standard-4
+    timeout-minutes: 90
+    env:
+      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
+      RUSTFS_BUILD_FEATURES: ftps,webdav,sftp
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
+        with:
+          persist-credentials: false
+
+      - name: Setup Rust environment
+        uses: ./.github/actions/setup
+        with:
+          rust-version: stable
+          cache-shared-key: ci-e2e-protocols
+          cache-save-if: 'false'
+          install-build-packaging-tools: 'false'
+
+      # The suite owns fixed protocol ports and serializes its internal cases.
+      - name: Verify protocol e2e membership
+        env:
+          NEXTEST_LISTING: ${{ runner.temp }}/rustfs-e2e-protocols-list.json
+        run: |
+          cargo nextest list --profile e2e-protocols -p e2e_test --message-format json > "${NEXTEST_LISTING}"
+          python3 ./scripts/check_test_wiring.py --check-profile e2e-protocols "${NEXTEST_LISTING}"
+
+      - name: Run protocol e2e nightly suite
+        env:
+          RUSTFS_E2E_LOG_DIR: ${{ runner.temp }}/rustfs-protocol-e2e-logs
+        run: >-
+          cargo nextest run -j 1 --profile e2e-protocols -p e2e_test --no-capture
+
+      - name: Upload protocol diagnostics
+        if: always()
+        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
+        with:
+          name: e2e-protocol-nightly-${{ github.run_number }}
+          path: |
+            target/nextest/e2e-protocols/junit.xml
+            ${{ runner.temp }}/rustfs-e2e-protocols-list.json
+            ${{ runner.temp }}/rustfs-protocol-e2e-logs/
+          retention-days: 7
+          if-no-files-found: warn
+
  alert-on-failure:
    name: Alert on scheduled failure
-    needs: [repl-nightly]
+    needs: [repl-nightly, cluster-nightly, protocols-nightly]
    # Only scheduled runs open/append the tracking issue (backlog#1149 ci-8);
    # manual workflow_dispatch runs stay quiet so a debugging run never files a
    # spurious alert.
@@ -18,10 +18,9 @@
 # runs only the implemented_tests.txt whitelist. This workflow complements it:
 #
 #   - Scheduled weekly full sweep (TEST_SCOPE=all): runs the ENTIRE upstream
-#     suite and reports promotion candidates (tests that newly pass) and
-#     unclassified tests. The job fails only on regressions in the implemented
-#     whitelist or on infrastructure errors — expected failures from
-#     not-yet-implemented features do not turn the run red.
+#     suite and reports promotion candidates. Regressions, unclassified tests,
+#     incomplete execution, and infrastructure errors fail the job; classified
+#     failures for not-yet-implemented features remain informational.
 #   - Manual runs (workflow_dispatch): same, with configurable mode/scope.
 #
 # All test execution is delegated to scripts/s3-tests/run.sh (single source of
@@ -45,13 +44,6 @@
 # The PR gate (ci.yml s3-implemented-tests) is unaffected: it avoids Docker
 # via DEPLOY_MODE=binary and defers all pip setup to run.sh's self-bootstrap.

-# DISABLED. This workflow is switched off in the repository's Actions settings
-# (state: disabled_manually) and does not run on any trigger, including its cron
-# and workflow_dispatch. That state lives in GitHub's UI and is invisible when
-# reading this file, which has already misled at least one audit — hence this
-# banner. Re-enabling is a UI action; anyone doing so should first check that the
-# workflow still matches the current CI layout. See rustfs/backlog#1603.
-#
 name: e2e-s3tests

 on:
@@ -81,6 +73,19 @@ on:
        description: "Stop after N failures. '0' to run everything."
        required: false
        default: "0"
+      shard-count:
+        description: "Exact-node-ID shard count for a targeted manual run"
+        required: false
+        default: "1"
+        type: choice
+        options:
+          - "1"
+          - "2"
+          - "4"
+      shard-index:
+        description: "Zero-based shard index for a targeted manual run"
+        required: false
+        default: "0"
      markexpr:
        description: "Optional pytest -m expression"
        required: false
@@ -111,6 +116,8 @@ env:
  XDIST: ${{ github.event.inputs.xdist || '4' }}
  MAXFAIL: ${{ github.event.inputs.maxfail || '0' }}
  MARKEXPR: ${{ github.event.inputs.markexpr || '' }}
+  S3_SHARD_COUNT: ${{ github.event_name == 'schedule' && '4' || github.event.inputs.shard-count || '1' }}
+  TEST_TIMEOUT: "300"

 concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.event.inputs['test-mode'] || 'single' }}
@@ -127,19 +134,22 @@ defaults:

 jobs:
  s3tests:
+    name: s3tests (${{ matrix.test-mode }}, shard ${{ matrix.shard-index }})
    # GitHub-hosted: reliably provides Docker + docker compose + python3/pip.
    # See the header note (ci-1) for why the self-hosted sm-standard-4 label
-    # was abandoned. TODO(ci-8): scheduled-failure alerting (auto-open issue)
-    # is added by the ci-8 composite action; do not implement it here.
+    # was abandoned. Scheduled failures are handled by alert-on-failure below.
    runs-on: ubuntu-latest
    timeout-minutes: 180
    strategy:
      fail-fast: false
+      max-parallel: 2
      matrix:
        # Scheduled sweeps cover both topologies; manual runs use the input.
        test-mode: ${{ github.event_name == 'schedule' && fromJSON('["single", "multi"]') || fromJSON(format('["{0}"]', github.event.inputs.test-mode || 'single')) }}
+        shard-index: ${{ github.event_name == 'schedule' && fromJSON('[0, 1, 2, 3]') || fromJSON(format('[{0}]', github.event.inputs.shard-index || '0')) }}
    env:
      TEST_MODE: ${{ matrix.test-mode }}
+      S3_SHARD_INDEX: ${{ matrix.shard-index }}
    steps:
      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
        with:
@@ -181,6 +191,7 @@ jobs:
      - name: Start single RustFS
        if: env.TEST_MODE == 'single'
        run: |
+          SSE_KEY="$(head -c 32 /dev/zero | base64 -w0)"
          docker network inspect rustfs-net >/dev/null 2>&1 || docker network create rustfs-net
          docker rm -f rustfs-single >/dev/null 2>&1 || true
          # The four disks share one physical device on the runner (a single
@@ -193,6 +204,7 @@ jobs:
            -e RUSTFS_ADDRESS=0.0.0.0:9000 \
            -e RUSTFS_ACCESS_KEY="${S3_ACCESS_KEY}" \
            -e RUSTFS_SECRET_KEY="${S3_SECRET_KEY}" \
+            -e RUSTFS_SSE_S3_MASTER_KEY="${SSE_KEY}" \
            -e RUSTFS_VOLUMES="/data/rustfs{0...3}" \
            -e RUSTFS_UNSAFE_BYPASS_DISK_CHECK=true \
            -v /tmp/rustfs-single:/data \
@@ -201,6 +213,7 @@ jobs:
      - name: Start 4-node distributed cluster
        if: env.TEST_MODE == 'multi'
        run: |
+          SSE_KEY="$(head -c 32 /dev/zero | base64 -w0)"
          # A real distributed deployment: every node lists all endpoints in
          # RUSTFS_VOLUMES so data is erasure-coded ACROSS nodes. Do not use
          # node-local volume paths here — that would create four independent
@@ -213,6 +226,7 @@ jobs:
              RUSTFS_ADDRESS: "0.0.0.0:9000"
              RUSTFS_ACCESS_KEY: ${S3_ACCESS_KEY}
              RUSTFS_SECRET_KEY: ${S3_SECRET_KEY}
+              RUSTFS_SSE_S3_MASTER_KEY: "${SSE_KEY}"
              RUSTFS_VOLUMES: "http://rustfs{1...4}:9000/data/rustfs{0...3}"
              # Each node's four disks share one physical device inside its
              # container, so bypass the local physical-disk-independence guard
@@ -294,7 +308,6 @@ jobs:

      - name: Run ceph s3-tests
        run: |
-          set +e
          DEPLOY_MODE=existing \
          TEST_MODE="${TEST_MODE}" \
          TEST_SCOPE="${TEST_SCOPE}" \
@@ -302,26 +315,6 @@ jobs:
          MAXFAIL="${MAXFAIL}" \
          MARKEXPR="${MARKEXPR}" \
          ./scripts/s3-tests/run.sh
-          RC=$?
-          set -e
-
-          if [ "${TEST_SCOPE}" = "implemented" ]; then
-            # Whitelist run: every failure is a regression.
-            exit "${RC}"
-          fi
-
-          # Full sweep: failures outside the implemented whitelist are
-          # inventory (promotion candidates / unimplemented features), not a
-          # gate. Fail only on whitelist regressions or infrastructure errors.
-          JUNIT="artifacts/s3tests-${TEST_MODE}/junit.xml"
-          if [ ! -f "${JUNIT}" ]; then
-            echo "No junit.xml produced — infrastructure failure (exit ${RC})" >&2
-            exit "${RC}"
-          fi
-          python3 scripts/s3-tests/report_compat.py \
-            --junit "${JUNIT}" \
-            --lists-dir scripts/s3-tests \
-            --fail-on-regression

      - name: Publish compatibility report
        if: always()
@@ -346,7 +339,7 @@ jobs:
        if: always() && env.ACT != 'true'
        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
        with:
-          name: s3tests-${{ env.TEST_MODE }}
+          name: s3tests-${{ env.TEST_MODE }}-shard-${{ matrix.shard-index }}
          path: artifacts/**

  alert-on-failure:
@@ -12,27 +12,22 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.

-# DISABLED. This workflow is switched off in the repository's Actions settings
-# (state: disabled_manually) and does not run on any trigger, including its cron
-# and workflow_dispatch. That state lives in GitHub's UI and is invisible when
-# reading this file, which has already misled at least one audit — hence this
-# banner. Re-enabling is a UI action; anyone doing so should first check that the
-# workflow still matches the current CI layout. See rustfs/backlog#1603.
-#
 name: Fuzz

 on:
  pull_request:
    types: [ opened, synchronize, reopened, closed ]
-    # PR trigger is intentionally narrow: only changes to the fuzz harness
-    # itself gate a PR. Broad crate paths (ecstore/filemeta/utils/policy/…)
-    # are covered by the nightly `schedule` run below, which fuzzes against
-    # whatever landed on main. Widening these paths previously queued a
-    # ~45min fuzz-build on nearly every PR and is why this workflow was
-    # disabled; do not re-add crate paths here.
+    # Run when the harness or any directly fuzzed production crate changes.
    paths:
      - "fuzz/**"
      - "scripts/fuzz/**"
+      - "crates/ecstore/**"
+      - "crates/filemeta/**"
+      - "crates/policy/**"
+      - "crates/security-governance/**"
+      - "crates/utils/**"
+      - "Cargo.toml"
+      - "Cargo.lock"
      - ".github/workflows/fuzz.yml"
  schedule:
    - cron: "0 2 * * *"
@@ -81,7 +76,7 @@ jobs:
      github.event_name == 'schedule' ||
      github.event_name == 'workflow_dispatch'
    runs-on: sm-standard-4
-    timeout-minutes: 45
+    timeout-minutes: 60
    env:
      FORCE_JAVASCRIPT_ACTIONS_TO_NODE24: "true"
    steps:
@@ -121,12 +116,7 @@ jobs:
        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
        with:
          name: fuzz-prebuilt-binaries-${{ github.run_number }}
-          path: |
-            fuzz/prebuilt/${{ env.CARGO_BUILD_TARGET }}/release/archive_extract
-            fuzz/prebuilt/${{ env.CARGO_BUILD_TARGET }}/release/bucket_validation
-            fuzz/prebuilt/${{ env.CARGO_BUILD_TARGET }}/release/local_metadata
-            fuzz/prebuilt/${{ env.CARGO_BUILD_TARGET }}/release/path_containment
-            fuzz/prebuilt/${{ env.CARGO_BUILD_TARGET }}/release/policy_ingress
+          path: fuzz/prebuilt/${{ env.CARGO_BUILD_TARGET }}/release/
          if-no-files-found: error
          retention-days: 1
          compression-level: 0
@@ -192,10 +182,7 @@ jobs:
  nightly-fuzz-corpus:
    name: "Nightly / ${{ matrix.target }}"
    needs: fuzz-build
-    # TODO(ci-8): when the schedule-failure-issue composite action lands,
-    # add a step here (or a dependent job) that opens/updates a GitHub issue
-    # on nightly failure. ci-8 is the single alerting mechanism for all
-    # scheduled workflows; do not self-roll alerting in this workflow.
+    # Scheduled failures are handled by alert-on-failure below.
    if: >
      github.event_name == 'schedule' ||
      (github.event_name == 'workflow_dispatch' &&
@@ -189,6 +189,7 @@ jobs:
    timeout-minutes: 30
    strategy:
      fail-fast: false
+      max-parallel: 1
      matrix:
        include:
          - arch: x86_64
@@ -510,15 +511,13 @@ jobs:

          CHECKSUM_DIR="$(mktemp -d)"
          gh release download "$TAG" -p 'SHA256SUMS' -p 'SHA512SUMS' \
-            -D "$CHECKSUM_DIR" --clobber 2>/dev/null || true
+            -D "$CHECKSUM_DIR" --clobber

          for spec in "SHA256SUMS:sha256sum" "SHA512SUMS:sha512sum"; do
            asset="${spec%%:*}"
            checksum_cmd="${spec##*:}"
            checksum_file="${CHECKSUM_DIR}/${asset}"

-            touch "$checksum_file"
-
            for f in "$DEB_FILE" "$RPM_FILE"; do
              if [[ -n "$f" && -f "$f" ]]; then
                base="$(basename "$f")"
@@ -531,7 +530,8 @@ jobs:
                grep -Fv -- "$base" "$checksum_file" > "${checksum_file}.tmp" || true
                grep -Fv -- "$github_base" "${checksum_file}.tmp" > "${checksum_file}.tmp2" || true
                mv "${checksum_file}.tmp2" "$checksum_file"
-                (cd "$(dirname "$f")" && "$checksum_cmd" -- "$github_base") >> "$checksum_file"
+                digest=$("$checksum_cmd" -- "$f" | awk '{print $1}')
+                printf '%s  %s\n' "$digest" "$github_base" >> "$checksum_file"
              fi
            done

@@ -9257,6 +9257,7 @@ dependencies = [
 "url",
 "urlencoding",
 "uuid",
+ "x509-parser",
 "zeroize",
 "zip",
 "zstd",
@@ -204,6 +204,7 @@ rsa = { version = "=0.10.0-rc.18" }
 rustls = { default-features = false, version = "0.23.43" }
 rustls-native-certs = "0.8"
 rustls-pki-types = "1.15.1"
+x509-parser = "0.18.1"
 sha1 = "0.11.0"
 sha2 = "0.11.0"
 subtle = "2.6"
@@ -48,16 +48,14 @@ cargo nextest run --profile e2e-smoke -p e2e_test
 cargo nextest run -j1 --run-ignored ignored-only -p rustfs-scanner -p rustfs \
  -E 'binary(lifecycle_integration_test) or (package(rustfs) and test(lifecycle_transition_api_test))'

-# Protocols suite — fixed ports, MUST be single-threaded, gated by build features
-RUSTFS_BUILD_FEATURES=ftps,webdav,sftp \
-  cargo test -p e2e_test test_protocol_core_suite -- --test-threads=1 --nocapture
 ```

 The protocols suite has its own contract (fixed bind ports 9022–9301,
-`--test-threads=1`, feature-gated scheduling) documented in
+single-worker execution, feature-gated scheduling) documented in
 [`src/protocols/README.md`](src/protocols/README.md). `RUSTFS_BUILD_FEATURES`
 selects which features the spawned binary is built with; leave it unset to run
-every protocol entry.
+every protocol entry. Use the exact profile command under
+[Troubleshooting](#troubleshooting) for CI-equivalent execution.

 ### `#[ignore]` semantics

@@ -159,27 +157,26 @@ construction (random port + isolated temp dir) and need no serialization.
 ## CI map

 `e2e_test` is **excluded** from the main `cargo nextest run --profile ci --all`
-pass ([`.github/workflows/ci.yml`](../../.github/workflows/ci.yml) line 158,
-`--exclude e2e_test`) — the whole crate is too slow to gate every PR. Subsets
-join CI through the nextest profile system only (never as ad-hoc jobs):
+pass (`--exclude e2e_test`) — the whole crate is too slow to gate every PR.
+Subsets join CI through nextest profiles; the fixed-port protocol suite uses
+the same profile for membership and execution with one nightly worker.

 | Suite | Runs where | Status |
 | --- | --- | --- |
 | Smoke subset (`e2e-smoke` profile) | `e2e-tests` job, every PR | **Active** (backlog#1149 ci-4) |
+| Full single-node suite (`e2e-full` profile) | `e2e-full` job, merge queue + main | **Active** (backlog#1149 ci-5) |
 | `s3s-e2e` black-box | `e2e-tests` + `e2e-tests-rio-v2` jobs | **Active** (external conformance tool) |
 | ILM / lifecycle (ignored) | `test-ilm-integration-serial` lane, `-j1` | **Active** (backlog#1148 ilm-1) |
-| KMS suite | — | Not in CI yet (backlog#1149 ci-5) |
-| Protocols (FTPS/WebDAV/SFTP) | — | Not in CI yet (backlog#1149 ci-7) |
+| KMS suite | `e2e-full` job, merge queue + main | **Active** |
+| Cluster faults (`e2e-nightly` profile) | consolidated nightly workflow | **Active** (backlog#1149 ci-7) |
+| Protocols (FTPS/WebDAV/SFTP) | consolidated nightly workflow, serial | **Active** (backlog#1149 ci-7) |
 | Replication (fast subset) | `e2e-smoke` profile, `e2e-tests` job, every PR | **Active** (backlog#1147 repl-1) |
-| Replication (slow + dual-node) | `e2e-repl-nightly` profile, scheduled workflow | **Active** (backlog#1147 repl-1) |
-| `reliant/*` (pre-started server) | — | Manual only |
+| Replication (slow + multi-node) | `e2e-repl-nightly` profile, consolidated nightly workflow | **Active** (backlog#1147 repl-1) |
+| `reliant/*` | 19 tests in PR smoke; remaining default tests in `e2e-full` | **Active** except `#[ignore]` |

-Links: [`ci.yml`](../../.github/workflows/ci.yml) `e2e-tests` (line 347),
-`test-ilm-integration-serial` (line 196). The `e2e-smoke` `default-filter` in
-[`.config/nextest.toml`](../../.config/nextest.toml) is the **single wiring
-mechanism** — extend that filter (or add a sibling profile) to admit more
-tests; do not add e2e jobs to `ci.yml`. repl-1 / ilm-3 are landing in parallel
-and may add lanes; keep the table above easy to extend.
+The profile filters in [`.config/nextest.toml`](../../.config/nextest.toml) are
+the wiring source of truth. Committed test-ID digests under
+`.config/e2e-*-selection.txt` make every membership change explicit.

 ## Troubleshooting

@@ -188,9 +185,15 @@ and may add lanes; keep the table above easy to extend.
 ```bash
 # Smoke (e2e-tests job) — includes the 20 fast replication tests
 cargo nextest run --profile e2e-smoke -p e2e_test
-# Replication nightly lane (16 slow + dual-node tests; install awscurl for the
-# STS dual-node test, else it skips gracefully)
+# Full single-node merge/main lane
+cargo nextest run --profile e2e-full -p e2e_test
+# Cluster fault nightly lane
+cargo nextest run --profile e2e-nightly -p e2e_test
+# Replication nightly lane; install awscurl so STS paths do not skip
 cargo nextest run --profile e2e-repl-nightly -p e2e_test
+# Fixed-port protocol nightly lane
+RUSTFS_BUILD_FEATURES=ftps,webdav,sftp \
+  cargo nextest run -j 1 --profile e2e-protocols -p e2e_test --no-capture
 # ILM serial lane
 cargo nextest run -j1 --run-ignored ignored-only -p rustfs-scanner -p rustfs \
  -E 'binary(lifecycle_integration_test) or (package(rustfs) and test(lifecycle_transition_api_test))'
@@ -273,4 +276,6 @@ current subset is.
 `docs/testing/e2e-suite-inventory.md` records the per-module test counts as
 listed by `cargo nextest list -p e2e_test`. Regenerate it when adding or
 moving e2e tests so acceptance numbers in the test-strategy issues
-(backlog#1147–#1155) stay auditable.
+(backlog#1147–#1155) stay auditable. When a profile membership change is
+intentional, review its JSON listing before updating the matching
+`.config/e2e-*-selection.txt` test-ID digest.
@@ -11,10 +11,17 @@ test process directly.

 ## Running Tests

+Use the canonical CI-equivalent protocol command in the parent
+[`e2e_test` README](../../README.md#troubleshooting).
+
+For targeted debugging of the core suite only:
+
 ```bash
 RUSTFS_BUILD_FEATURES=ftps,webdav,sftp cargo test --package e2e_test test_protocol_core_suite -- --test-threads=1 --nocapture
 ```

+This targeted command does not cover the full `e2e-protocols` profile.
+
 `RUSTFS_BUILD_FEATURES` controls which features the test rustfs binary is
 built with. When this variable is set, the protocol test runner schedules
 only entries whose protocol is present in the requested feature list. Leave
@@ -133,4 +140,3 @@ property without consulting any external doc.
 Bind ports 9023 (SFTP) and 9100 (S3). Spawns rustfs with
 `RUSTFS_SFTP_IDLE_TIMEOUT=5`, sleeps 10 s past the timeout, then issues an
 SFTP request and asserts the server has closed the session.
-
@@ -317,8 +317,6 @@ 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,
@@ -330,6 +328,8 @@ pub mod data_usage {
        remove_bucket_usage_from_backend, replace_bucket_usage_memory_from_info, store_compression_total_in_backend,
        store_data_usage_in_backend,
    };
+    #[cfg(feature = "test-util")]
+    pub use crate::data_usage::{get_bucket_usage_memory, seed_bucket_usage_memory_for_test};
 }

 pub mod disk {
@@ -866,7 +866,7 @@ impl BucketTargetSys {
            return Some(cli);
        }

-        // TODO: spawn a task to reload the target
+        // TODO(backlog): spawn an async task to proactively reload the replication target
        if self.is_reloading_target(bucket, arn).await {
            return None;
        }
@@ -41,6 +41,7 @@ const IAM_FORMAT_FILE_PATH: &str = "config/iam/format.json";
 const IAM_USERS_PREFIX: &str = "config/iam/users/";
 const IAM_SERVICE_ACCOUNTS_PREFIX: &str = "config/iam/service-accounts/";
 const IAM_STS_PREFIX: &str = "config/iam/sts/";
+const MINIO_GO_ZERO_TIME: OffsetDateTime = time::macros::datetime!(0001-01-01 00:00 UTC);
 const IAM_GROUPS_PREFIX: &str = "config/iam/groups/";
 const IAM_POLICIES_PREFIX: &str = "config/iam/policies/";
 const IAM_POLICY_DB_PREFIX: &str = "config/iam/policydb/";
@@ -120,6 +121,15 @@ fn normalize_iam_config_blob(path: &str, data: &[u8]) -> std::result::Result<Opt
    if is_identity_path(path) {
        let mut identity: UserIdentity =
            serde_json::from_slice(data).map_err(|err| format!("parse IAM identity failed: {err}"))?;
+        if (path.starts_with(IAM_USERS_PREFIX) || path.starts_with(IAM_SERVICE_ACCOUNTS_PREFIX))
+            && identity
+                .credentials
+                .expiration
+                .as_ref()
+                .is_some_and(|expiration| *expiration == MINIO_GO_ZERO_TIME || *expiration == OffsetDateTime::UNIX_EPOCH)
+        {
+            identity.credentials.expiration = None;
+        }
        if identity.update_at.is_none() {
            identity.update_at = Some(OffsetDateTime::now_utc());
        }
@@ -441,7 +451,10 @@ mod tests {
    use crate::bucket::replication::{
        BucketReplicationResyncStatus, ReplicationMigrationBridge, ResyncStatusType, TargetReplicationResyncStatus,
    };
+    use rustfs_policy::auth::UserIdentity;
    use std::collections::HashMap;
+    use time::OffsetDateTime;
+    use time::format_description::well_known::Rfc3339;

    #[test]
    fn test_normalize_policy_mapping_legacy_timestamp_and_fields() {
@@ -493,6 +506,54 @@ mod tests {
        assert!(v.get("updatedAt").is_some(), "normalize should backfill updatedAt");
    }

+    #[test]
+    fn test_normalize_minio_permanent_credential_expiration() {
+        let cases = [
+            ("config/iam/users/alice/identity.json", "0001-01-01T00:00:00Z", true),
+            ("config/iam/users/alice/identity.json", "1970-01-01T00:00:00Z", true),
+            ("config/iam/service-accounts/svc/identity.json", "0001-01-01T00:00:00Z", true),
+            ("config/iam/service-accounts/svc/identity.json", "1970-01-01T00:00:00Z", true),
+            ("config/iam/service-accounts/svc/identity.json", "1970-01-01T00:00:00.000000001Z", false),
+            ("config/iam/sts/temp/identity.json", "0001-01-01T00:00:00Z", false),
+            ("config/iam/sts/temp/identity.json", "1970-01-01T00:00:00Z", false),
+            ("config/iam/users/alice/identity.json", "1969-12-31T23:59:59Z", false),
+            ("config/iam/users/alice/identity.json", "1970-01-01T00:00:00.000000001Z", false),
+            ("config/iam/users/alice/identity.json", "0001-01-01T00:00:00.000000001Z", false),
+            ("config/iam/users/alice/identity.json", "2030-01-01T00:00:00Z", false),
+        ];
+
+        for (path, expiration, should_clear) in cases {
+            let input = serde_json::json!({
+                "version": 1,
+                "credentials": {
+                    "accessKey": "test-access",
+                    "secretKey": "test-secret",
+                    "sessionToken": "test-session-token",
+                    "parentUser": "test-parent",
+                    "expiration": expiration,
+                }
+            });
+            let output = normalize_iam_config_blob(path, &serde_json::to_vec(&input).expect("serialize identity fixture"))
+                .expect("normalize should succeed")
+                .expect("identity path should be supported");
+            let identity: UserIdentity = serde_json::from_slice(&output).expect("deserialize normalized identity");
+
+            assert_eq!(identity.credentials.access_key, "test-access");
+            assert_eq!(identity.credentials.secret_key, "test-secret");
+            assert_eq!(identity.credentials.session_token, "test-session-token");
+            assert_eq!(identity.credentials.parent_user, "test-parent");
+            if should_clear {
+                assert_eq!(identity.credentials.expiration, None, "path: {path}, expiration: {expiration}");
+            } else {
+                assert_eq!(
+                    identity.credentials.expiration,
+                    Some(OffsetDateTime::parse(expiration, &Rfc3339).expect("parse expected expiration")),
+                    "path: {path}, expiration: {expiration}"
+                );
+            }
+        }
+    }
+
    #[test]
    fn test_normalize_bucket_meta_blob_resync_reencode() {
        let path = ".buckets/test/.replication/resync.bin";
@@ -2855,7 +2855,7 @@ fn replicate_object_info_from_object_info(
        .map(|v| OffsetDateTime::parse(&v, &Rfc3339).unwrap_or(OffsetDateTime::UNIX_EPOCH));
    let mut rstate = oi.replication_state();
    rstate.replicate_decision_str = dsc.to_string();
-    let asz = oi.get_actual_size().unwrap_or_default();
+    let asz = oi.get_actual_size_or_physical();
    let ssec = replication_object_is_ssec_encrypted(&oi.user_defined);
    let checksum = if ssec { oi.checksum.clone() } else { None };

@@ -1412,7 +1412,7 @@ pub async fn get_heal_replicate_object_info(oi: &ObjectInfo, rcfg: &ReplicationC
    };
    let mut replication_state = oi.replication_state();
    replication_state.replicate_decision_str = dsc.to_string();
-    let actual_size = oi.get_actual_size().unwrap_or_default();
+    let actual_size = oi.get_actual_size_or_physical();

    Ok(ReplicateObjectInfo {
        name: oi.name.clone(),
@@ -389,7 +389,7 @@ fn replication_source_object(object_info: &ObjectInfo) -> ReplicationSourceObjec
            .map(|mod_time| OffsetDateTime::from_unix_timestamp(mod_time.unix_timestamp()).unwrap_or(mod_time)),
        version_id: object_info.version_id.map(|version_id| version_id.to_string()),
        etag: object_info.etag.as_deref(),
-        actual_size: object_info.get_actual_size().unwrap_or_default(),
+        actual_size: object_info.get_actual_size_or_physical(),
        delete_marker: object_info.delete_marker,
        content_type: object_info.content_type.as_deref(),
        content_encoding: object_info.content_encoding.as_deref(),
@@ -542,6 +542,20 @@ mod tests {
        assert!(replication_target_head_is_newer_null_version(&source, &target));
    }

+    #[test]
+    fn replication_source_uses_physical_size_for_unknown_compressed_object() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(&mut metadata, rustfs_utils::http::SUFFIX_COMPRESSION, "zstd".to_string());
+        let source = ObjectInfo {
+            size: 128,
+            actual_size: -1,
+            user_defined: Arc::new(metadata),
+            ..Default::default()
+        };
+
+        assert_eq!(replication_source_object(&source).actual_size, 128);
+    }
+
    #[test]
    fn replication_target_head_content_matches_compare_etag_only() {
        let source = ObjectInfo {
@@ -454,7 +454,7 @@ impl S3PeerSys {
                }
            }
            topology_complete &= bucket_map.values().all(|count| *count >= quorum);
-            // TODO: MRF
+            // TODO(backlog): integrate MRF backlog stats into scanner bucket listing
        }

        let mut buckets: Vec<BucketInfo> = result_map.into_values().collect();
@@ -2406,7 +2406,7 @@ impl DiskAPI for RemoteDisk {
            return errors;
        }

-        // TODO: use Error not string
+        // TODO(backlog): replace string errors with typed `StorageError` variants

        let result = self
            .execute_with_timeout(
@@ -94,6 +94,9 @@ const DECOMMISSION_BUCKET_CONCURRENCY_DEFAULT_CAP: usize = 4;
 const DECOMMISSION_TARGET_CAPACITY_OVERHEAD_PERCENT: usize = 30;
 const DECOMMISSION_LISTING_MAX_ATTEMPTS: usize = 3;
 const DECOMMISSION_LISTING_RETRY_DELAY: std::time::Duration = std::time::Duration::from_secs(5);
+/// Background decommission walks must tolerate slow object migrations; the
+/// stall timeout is the drive-health bound, not the total listing duration.
+const DECOMMISSION_BACKGROUND_WALKDIR_STALL_TIMEOUT: std::time::Duration = std::time::Duration::from_secs(60);

 pub const POOL_META_NAME: &str = "pool.bin";
 pub const POOL_META_FORMAT: u16 = 1;
@@ -5047,6 +5050,8 @@ impl SetDisks {
                path: bucket_info.prefix.clone(),
                recursive: true,
                min_disks: listing_quorum,
+                skip_walkdir_total_timeout: true,
+                walkdir_stall_timeout: Some(DECOMMISSION_BACKGROUND_WALKDIR_STALL_TIMEOUT),
                agreed: Some(Box::new(move |entry: MetaCacheEntry| Box::pin(cb1(entry)))),
                partial: Some(Box::new(move |entries: MetaCacheEntries, _: &[Option<DiskError>]| {
                    let resolver = resolver.clone();
@@ -21,7 +21,7 @@ use crate::storage_api_contracts::{
    bucket::{BucketInfo, BucketOperations, BucketOptions, DeleteBucketOptions, MakeBucketOptions},
    list::{StorageListObjectVersionsInfo, StorageListObjectsV2Info, StorageObjectInfoOrErr, StorageWalkOptions},
    multipart::{CompletePart, ListMultipartsInfo, ListPartsInfo, MultipartInfo, MultipartUploadResult, PartInfo},
-    object::{DeletedObject, ObjectIO as _, ObjectOperations as _, ObjectToDelete},
+    object::{DeleteAccounting, DeletedObject, ObjectIO as _, ObjectOperations as _, ObjectToDelete},
    range::HTTPRangeSpec,
 };
 use crate::{
@@ -249,7 +249,7 @@ impl Sets {

        self.connect_disks().await;

-        // TODO: config interval
+        // TODO(backlog): make monitor_and_connect interval configurable instead of hardcoded 15s
        let mut interval = tokio::time::interval(Duration::from_secs(15));
        loop {
            tokio::select! {
@@ -414,6 +414,66 @@ fn apply_delete_objects_results(
    }
 }

+fn apply_delete_accounting_results(
+    accounting: &mut [Option<DeleteAccounting>],
+    set_objects: &[DelObj],
+    set_accounting: &[Option<DeleteAccounting>],
+) {
+    for (obj, value) in set_objects.iter().zip(set_accounting.iter()) {
+        accounting[obj.orig_idx] = value.clone();
+    }
+}
+
+impl Sets {
+    pub(crate) async fn delete_objects_with_accounting(
+        &self,
+        bucket: &str,
+        objects: Vec<ObjectToDelete>,
+        opts: ObjectOptions,
+    ) -> (Vec<DeletedObject>, Vec<Option<Error>>, Vec<Option<DeleteAccounting>>) {
+        let mut del_objects = vec![DeletedObject::default(); objects.len()];
+        let mut del_errs = vec![None; objects.len()];
+        let mut accounting = vec![None; objects.len()];
+        let mut set_obj_map = HashMap::new();
+
+        for (i, obj) in objects.iter().enumerate() {
+            let idx = self.get_hashed_set_index(obj.object_name.as_str());
+            set_obj_map.entry(idx).or_insert_with(Vec::new).push(DelObj {
+                orig_idx: i,
+                obj: obj.clone(),
+            });
+        }
+
+        let max_concurrent = set_obj_map.len().min(num_cpus::get()).max(1);
+        let semaphore = Arc::new(tokio::sync::Semaphore::new(max_concurrent));
+        let mut futures = FuturesUnordered::new();
+        let bucket = bucket.to_owned();
+
+        for (set_index, set_objects) in set_obj_map {
+            let disks = self.get_disks(set_index);
+            let objects = set_objects.iter().map(|entry| entry.obj.clone()).collect::<Vec<_>>();
+            let bucket = bucket.clone();
+            let opts = opts.clone();
+            let semaphore = semaphore.clone();
+            futures.push(async move {
+                let _permit = semaphore
+                    .acquire_owned()
+                    .await
+                    .expect("delete_objects semaphore should remain open");
+                let (deleted, errors, accounting) = disks.delete_objects_with_accounting(&bucket, objects, opts).await;
+                (set_objects, deleted, errors, accounting)
+            });
+        }
+
+        while let Some((set_objects, deleted, errors, set_accounting)) = futures.next().await {
+            apply_delete_objects_results(&mut del_objects, &mut del_errs, &set_objects, &deleted, errors);
+            apply_delete_accounting_results(&mut accounting, &set_objects, &set_accounting);
+        }
+
+        (del_objects, del_errs, accounting)
+    }
+}
+
 #[async_trait::async_trait]
 impl crate::storage_api_contracts::object::ObjectIO for Sets {
    type Error = Error;
@@ -655,65 +715,8 @@ impl crate::storage_api_contracts::object::ObjectOperations for Sets {
        objects: Vec<ObjectToDelete>,
        opts: ObjectOptions,
    ) -> (Vec<DeletedObject>, Vec<Option<Error>>) {
-        // Default return value
-        let mut del_objects = vec![DeletedObject::default(); objects.len()];
-
-        let mut del_errs = Vec::with_capacity(objects.len());
-        for _ in 0..objects.len() {
-            del_errs.push(None)
-        }
-
-        let mut set_obj_map = HashMap::new();
-
-        // hash key
-        for (i, obj) in objects.iter().enumerate() {
-            let idx = self.get_hashed_set_index(obj.object_name.as_str());
-
-            if !set_obj_map.contains_key(&idx) {
-                set_obj_map.insert(
-                    idx,
-                    vec![DelObj {
-                        // set_idx: idx,
-                        orig_idx: i,
-                        obj: obj.clone(),
-                    }],
-                );
-            } else if let Some(val) = set_obj_map.get_mut(&idx) {
-                val.push(DelObj {
-                    // set_idx: idx,
-                    orig_idx: i,
-                    obj: obj.clone(),
-                });
-            }
-        }
-
-        let max_concurrent = set_obj_map.len().min(num_cpus::get()).max(1);
-        let semaphore = Arc::new(tokio::sync::Semaphore::new(max_concurrent));
-        let mut futures = FuturesUnordered::new();
-        let bucket = bucket.to_string();
-
-        for (k, v) in set_obj_map {
-            let disks = self.get_disks(k);
-            let objs: Vec<ObjectToDelete> = v.iter().map(|v| v.obj.clone()).collect();
-            let bucket = bucket.clone();
-            let opts = opts.clone();
-            let semaphore = semaphore.clone();
-
-            futures.push(async move {
-                let _permit = semaphore
-                    .acquire_owned()
-                    .await
-                    .expect("delete_objects semaphore should remain open");
-                let (dobjects, errs) = disks.delete_objects(&bucket, objs, opts).await;
-                (v, dobjects, errs)
-            });
-        }
-
-        while let Some((v, dobjects, errs)) = futures.next().await {
-            apply_delete_objects_results(&mut del_objects, &mut del_errs, &v, &dobjects, errs);
-        }
-
-        (del_objects, del_errs)
+        let (deleted, errors, _) = self.delete_objects_with_accounting(bucket, objects, opts).await;
+        (deleted, errors)
    }

    #[tracing::instrument(skip(self))]
@@ -1391,7 +1391,37 @@ impl BucketUsageAccumulator {
 }

 pub fn quota_object_size(object: &ObjectInfo) -> Result<u64, Error> {
-    let logical_size = u64::try_from(object.get_actual_size().map_err(Error::other)?).map_err(|_| Error::PartMissingOrCorrupt)?;
+    // A compressed object may carry -1 while the transformed size is unknown
+    // (legacy streaming sentinel). In that case the persisted physical size
+    // is still a valid accounting floor; every other negative value is corrupt.
+    // An explicit negative `actual-size` metadata value is corrupt, however:
+    // the sentinel is only valid in the in-memory/object-part field written by
+    // the legacy streaming path, not as a persisted declared size.
+    let compressed = object.is_compressed();
+    if object.actual_size < -1 || (object.actual_size == -1 && !compressed) {
+        return Err(Error::PartMissingOrCorrupt);
+    }
+    if object
+        .parts
+        .iter()
+        .any(|part| part.actual_size < -1 || (part.actual_size < 0 && !compressed))
+    {
+        return Err(Error::PartMissingOrCorrupt);
+    }
+    let declared_actual_size = rustfs_utils::http::get_str(&object.user_defined, rustfs_utils::http::SUFFIX_ACTUAL_SIZE)
+        .filter(|value| !value.is_empty());
+    if declared_actual_size
+        .as_deref()
+        .and_then(|value| value.parse::<i64>().ok())
+        .is_some_and(|size| size < 0)
+    {
+        return Err(Error::PartMissingOrCorrupt);
+    }
+    let logical_size = match object.get_actual_size().map_err(Error::other)? {
+        size if size == -1 && compressed && declared_actual_size.is_none() => None,
+        size if size >= 0 => Some(u64::try_from(size).map_err(|_| Error::PartMissingOrCorrupt)?),
+        _ => return Err(Error::PartMissingOrCorrupt),
+    };
    let persisted_part_size = if object.parts.is_empty() {
        u64::try_from(object.size).map_err(|_| Error::PartMissingOrCorrupt)?
    } else {
@@ -1399,12 +1429,8 @@ pub fn quota_object_size(object: &ObjectInfo) -> Result<u64, Error> {
            // Compressed streaming objects persist -1 when the transformed
            // part size is unknown. The physical part size remains a valid
            // quota floor; reject only non-negative values that overflow.
-            let actual_size = if part.actual_size < 0 {
-                if object.is_compressed() {
-                    0
-                } else {
-                    return Err(Error::PartMissingOrCorrupt);
-                }
+            let actual_size = if part.actual_size == -1 {
+                0
            } else {
                u64::try_from(part.actual_size).map_err(|_| Error::PartMissingOrCorrupt)?
            };
@@ -1412,7 +1438,7 @@ pub fn quota_object_size(object: &ObjectInfo) -> Result<u64, Error> {
            total.checked_add(part_size).ok_or(Error::PartMissingOrCorrupt)
        })?
    };
-    Ok(logical_size.max(persisted_part_size))
+    Ok(logical_size.unwrap_or(0).max(persisted_part_size))
 }

 type UsageVersionPage = StorageListObjectVersionsInfo<ObjectInfo>;
@@ -3320,6 +3346,80 @@ mod tests {
        );
    }

+    #[test]
+    fn quota_object_size_accepts_compressed_unknown_actual_size_sentinel() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        let object = ObjectInfo {
+            size: 400,
+            actual_size: -1,
+            user_defined: Arc::new(metadata),
+            ..Default::default()
+        };
+
+        assert_eq!(quota_object_size(&object).expect("compressed sentinel is valid"), 400);
+    }
+
+    #[test]
+    fn quota_object_size_rejects_compressed_part_sum_overflow() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        let object = ObjectInfo {
+            size: 1,
+            user_defined: Arc::new(metadata),
+            parts: Arc::new(vec![
+                rustfs_filemeta::ObjectPartInfo {
+                    actual_size: i64::MAX,
+                    ..Default::default()
+                },
+                rustfs_filemeta::ObjectPartInfo {
+                    actual_size: 1,
+                    ..Default::default()
+                },
+            ]),
+            ..Default::default()
+        };
+
+        assert!(matches!(quota_object_size(&object), Err(Error::Io(_))));
+    }
+
+    #[test]
+    fn quota_object_size_rejects_negative_values_other_than_the_compressed_sentinel() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        let corrupt_object = ObjectInfo {
+            size: 400,
+            actual_size: -2,
+            user_defined: Arc::new(metadata.clone()),
+            ..Default::default()
+        };
+        assert!(matches!(quota_object_size(&corrupt_object), Err(Error::PartMissingOrCorrupt)));
+
+        let corrupt_part = ObjectInfo {
+            size: 400,
+            user_defined: Arc::new(metadata),
+            parts: Arc::new(vec![rustfs_filemeta::ObjectPartInfo {
+                size: 400,
+                actual_size: -2,
+                ..Default::default()
+            }]),
+            ..Default::default()
+        };
+        assert!(matches!(quota_object_size(&corrupt_part), Err(Error::PartMissingOrCorrupt)));
+    }
+
    #[tokio::test]
    #[serial]
    async fn live_bucket_usage_refreshes_are_coalesced_only_while_in_flight() {
@@ -5215,8 +5215,8 @@ impl LocalDisk {

        let cache = Cache::new(update_fn, Duration::from_secs(1), Opts::default());

-        // TODO: DIRECT support
-        // TODD: DiskInfo
+        // TODO(backlog): add O_DIRECT I/O support for performance-critical paths
+        // TODO(backlog): populate DiskInfo in constructor
        let mut disk = Self {
            root: root.clone(),
            publication_root,
@@ -5751,7 +5751,7 @@ impl LocalDisk {

        // return Ok(());

-        // TODO: async notifications for disk space checks and trash cleanup
+        // TODO(backlog): make disk space checks and trash cleanup event-driven instead of poll-based

        let trash_path = self.io_get_object_path(RUSTFS_META_TMP_DELETED_BUCKET, Uuid::new_v4().to_string().as_str())?;
        // if let Some(parent) = trash_path.parent() {
@@ -5997,7 +5997,7 @@ impl LocalDisk {

    #[hotpath::measure(impl_type = "LocalDisk")]
    async fn read_all_data(&self, volume: &str, volume_dir: impl AsRef<Path>, file_path: impl AsRef<Path>) -> Result<Vec<u8>> {
-        // TODO: timeout support
+        // TODO(backlog): add configurable timeout for read_all_data operations
        let (data, _) = self.read_all_data_with_dmtime(volume, volume_dir, file_path).await?;
        Ok(data)
    }
@@ -6674,7 +6674,7 @@ impl LocalDisk {
            return Ok(());
        }

-        // TODO: add lock
+        // TODO(backlog): add directory listing lock to prevent concurrent enumeration

        let stall = opts.stall_timeout_duration();

@@ -8796,7 +8796,7 @@ impl DiskAPI for LocalDisk {
        Ok(entries)
    }

-    // FIXME: TODO: io.writer TODO cancel
+    // TODO(backlog): support io.writer cancellation and early termination in walk_dir
    #[tracing::instrument(level = "trace", skip_all)]
    async fn walk_dir<W: AsyncWrite + Unpin + Send>(&self, opts: WalkDirOptions, wr: &mut W) -> Result<()> {
        self.wait_for_startup_cleanup().await;
@@ -9880,7 +9880,7 @@ impl DiskAPI for LocalDisk {
                );
                return Err(e);
            }
-            // TODO: health check
+            // TODO(backlog): add post-setup disk health verification
        }
        Ok(())
    }
@@ -23,7 +23,7 @@ use std::{
    io,
    path::{Component, Path, PathBuf},
    sync::{Arc, LazyLock, Weak},
-    time::Instant,
+    time::{Duration, Instant},
 };
 use tokio::fs;
 use tokio::sync::{
@@ -328,6 +328,9 @@ const ENV_DST_DIR_FSYNC_GROUP_COMMIT_ENABLE: &str = "RUSTFS_EXPERIMENTAL_DST_DIR
 const DEFAULT_DST_DIR_FSYNC_GROUP_COMMIT_ENABLE: bool = false;
 const ENV_FILE_FDATASYNC_GROUP_COMMIT_ENABLE: &str = "RUSTFS_EXPERIMENTAL_FILE_FDATASYNC_GROUP_COMMIT_ENABLE";
 const DEFAULT_FILE_FDATASYNC_GROUP_COMMIT_ENABLE: bool = false;
+const ENV_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS: &str = "RUSTFS_EXPERIMENTAL_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS";
+const DEFAULT_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS: u64 = 0;
+const MAX_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS: u64 = 1_000;
 #[cfg(not(test))]
 const MAX_DST_DIR_FSYNC_GROUPS: usize = 1024;
 #[cfg(test)]
@@ -354,6 +357,16 @@ static DST_DIR_FSYNC_GROUP_COMMIT_ENABLED: LazyLock<bool> = LazyLock::new(|| {
 static FILE_FDATASYNC_GROUP_COMMIT_ENABLED: LazyLock<bool> = LazyLock::new(|| {
    rustfs_utils::get_env_bool(ENV_FILE_FDATASYNC_GROUP_COMMIT_ENABLE, DEFAULT_FILE_FDATASYNC_GROUP_COMMIT_ENABLE)
 });
+fn file_fdatasync_group_commit_wait_duration(wait_micros: u64) -> Duration {
+    Duration::from_micros(wait_micros.min(MAX_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS))
+}
+
+static FILE_FDATASYNC_GROUP_COMMIT_WAIT: LazyLock<Duration> = LazyLock::new(|| {
+    file_fdatasync_group_commit_wait_duration(rustfs_utils::get_env_u64(
+        ENV_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS,
+        DEFAULT_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS,
+    ))
+});

 #[cfg(test)]
 mod dst_dir_fsync_group_commit_override {
@@ -402,6 +415,7 @@ mod file_fdatasync_group_commit_override {
    use std::sync::{Mutex, MutexGuard, PoisonError, RwLock};

    static OVERRIDE: RwLock<Option<bool>> = RwLock::new(None);
+    static WAIT_OVERRIDE_MICROS: RwLock<Option<u64>> = RwLock::new(None);
    static SERIAL: Mutex<()> = Mutex::new(());

    pub(crate) fn get() -> Option<bool> {
@@ -415,6 +429,7 @@ mod file_fdatasync_group_commit_override {
    impl Drop for OverrideGuard {
        fn drop(&mut self) {
            *OVERRIDE.write().unwrap_or_else(PoisonError::into_inner) = None;
+            *WAIT_OVERRIDE_MICROS.write().unwrap_or_else(PoisonError::into_inner) = None;
        }
    }

@@ -423,6 +438,14 @@ mod file_fdatasync_group_commit_override {
        *OVERRIDE.write().unwrap_or_else(PoisonError::into_inner) = Some(enabled);
        OverrideGuard { _serial: serial }
    }
+
+    pub(crate) fn set_wait_micros(wait_micros: u64) {
+        *WAIT_OVERRIDE_MICROS.write().unwrap_or_else(PoisonError::into_inner) = Some(wait_micros);
+    }
+
+    pub(crate) fn wait_micros() -> Option<u64> {
+        *WAIT_OVERRIDE_MICROS.read().unwrap_or_else(PoisonError::into_inner)
+    }
 }

 #[cfg(test)]
@@ -430,6 +453,11 @@ pub(crate) fn set_file_fdatasync_group_commit_for_test(enabled: bool) -> file_fd
    file_fdatasync_group_commit_override::set(enabled)
 }

+#[cfg(test)]
+fn set_file_fdatasync_group_commit_wait_for_test(wait_micros: u64) {
+    file_fdatasync_group_commit_override::set_wait_micros(wait_micros);
+}
+
 fn file_fdatasync_group_commit_enabled() -> bool {
    #[cfg(test)]
    if let Some(enabled) = file_fdatasync_group_commit_override::get() {
@@ -439,6 +467,15 @@ fn file_fdatasync_group_commit_enabled() -> bool {
    *FILE_FDATASYNC_GROUP_COMMIT_ENABLED
 }

+fn file_fdatasync_group_commit_wait() -> Duration {
+    #[cfg(test)]
+    if let Some(wait_micros) = file_fdatasync_group_commit_override::wait_micros() {
+        return file_fdatasync_group_commit_wait_duration(wait_micros);
+    }
+
+    *FILE_FDATASYNC_GROUP_COMMIT_WAIT
+}
+
 #[derive(Clone, Eq, Hash, PartialEq)]
 struct DstDirFsyncGroupKey {
    canonical_path: PathBuf,
@@ -934,6 +971,10 @@ async fn run_file_fdatasync_group_worker(group: Arc<FileFdatasyncGroup>) {
        #[cfg(test)]
        file_sync_probe::run_before_group_batch();
        tokio::task::yield_now().await;
+        let wait = file_fdatasync_group_commit_wait();
+        if !wait.is_zero() {
+            tokio::time::sleep(wait).await;
+        }
        let (batch, batch_file_count): (Vec<FileFdatasyncWaiter>, usize) = {
            let mut group_state = group.inner.lock();
            let batch_file_count = group_state.pending_files;
@@ -6075,6 +6116,7 @@ mod tests {
        use std::sync::mpsc;

        let _group_commit = set_file_fdatasync_group_commit_for_test(true);
+        set_file_fdatasync_group_commit_wait_for_test(0);
        clear_file_fdatasync_group_commit_for_test();
        let temp_dir = tempdir().expect("create temp dir");
        let first_dir = temp_dir.path().join("first");
@@ -6141,12 +6183,105 @@ mod tests {
        assert_eq!(file_fdatasync_group_commit_counts_for_test(), (0, 0, 0));
    }

+    #[test]
+    fn file_fdatasync_group_commit_wait_duration_uses_default_and_cap() {
+        assert_eq!(DEFAULT_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS, 0);
+        assert_eq!(
+            file_fdatasync_group_commit_wait_duration(DEFAULT_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS),
+            Duration::ZERO
+        );
+        assert_eq!(file_fdatasync_group_commit_wait_duration(250), Duration::from_micros(250));
+        assert_eq!(
+            file_fdatasync_group_commit_wait_duration(MAX_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS),
+            Duration::from_micros(MAX_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS)
+        );
+        assert_eq!(
+            file_fdatasync_group_commit_wait_duration(u64::MAX),
+            Duration::from_micros(MAX_FILE_FDATASYNC_GROUP_COMMIT_WAIT_MICROS)
+        );
+    }
+
+    #[tokio::test(flavor = "current_thread", start_paused = true)]
+    #[serial_test::serial(file_sync_probe)]
+    async fn file_fdatasync_group_commit_wait_budget_batches_late_follower() {
+        use std::sync::mpsc;
+
+        let _group_commit = set_file_fdatasync_group_commit_for_test(true);
+        let wait_budget_micros = 1_000;
+        let wait_budget = file_fdatasync_group_commit_wait_duration(wait_budget_micros);
+        set_file_fdatasync_group_commit_wait_for_test(wait_budget_micros);
+        clear_file_fdatasync_group_commit_for_test();
+        let temp_dir = tempdir().expect("create temp dir");
+        let first_dir = temp_dir.path().join("first");
+        let second_dir = temp_dir.path().join("second");
+        std::fs::create_dir(&first_dir).expect("create first dir");
+        std::fs::create_dir(&second_dir).expect("create second dir");
+        std::fs::write(first_dir.join("part.1"), b"first").expect("write first part");
+        std::fs::write(second_dir.join("part.1"), b"second").expect("write second part");
+        let _probe = file_sync_probe::set_blocking(temp_dir.path());
+        let (entered_tx, entered_rx) = mpsc::channel();
+        file_sync_probe::set_before_group_batch(move || {
+            entered_tx.send(()).expect("signal first file fdatasync group worker");
+        });
+
+        let limiter = file_sync_limiter();
+        let first_limiter = limiter.clone();
+        let first_path = first_dir.clone();
+        let first = tokio::spawn(async move { sync_dir_files_with_limiter(first_path, first_limiter).await });
+        tokio::task::spawn_blocking(move || entered_rx.recv_timeout(Duration::from_secs(30)))
+            .await
+            .expect("group worker hook waiter should run")
+            .expect("first file fdatasync group worker should start");
+
+        let second_limiter = limiter.clone();
+        let second_path = second_dir.clone();
+        let second = tokio::spawn(async move { sync_dir_files_with_limiter(second_path, second_limiter).await });
+        tokio::time::timeout(Duration::from_secs(30), async {
+            loop {
+                if file_fdatasync_group_commit_counts_for_test().1 == 2 {
+                    return;
+                }
+                tokio::task::yield_now().await;
+            }
+        })
+        .await
+        .expect("second waiter should enqueue during the configured wait budget");
+        tokio::time::advance(wait_budget).await;
+        tokio::task::yield_now().await;
+        file_sync_probe::wait_for_active(1).await;
+
+        assert_eq!(
+            file_sync_probe::group_batches(),
+            vec![2],
+            "configured wait budget should let a follower join the leader's batch"
+        );
+        file_sync_probe::release();
+        first
+            .await
+            .expect("join first wait-budget file sync")
+            .expect("first wait-budget file sync must succeed");
+        second
+            .await
+            .expect("join second wait-budget file sync")
+            .expect("second wait-budget file sync must succeed");
+        assert!(
+            fsync_dir_recorder::was_fsynced(&first_dir),
+            "first source directory must still be fsynced"
+        );
+        assert!(
+            fsync_dir_recorder::was_fsynced(&second_dir),
+            "second source directory must still be fsynced"
+        );
+        assert_eq!(file_fdatasync_group_commit_counts_for_test(), (0, 0, 0));
+    }
+
    #[tokio::test(flavor = "multi_thread", worker_threads = 2)]
    #[serial_test::serial(file_sync_probe)]
    async fn file_fdatasync_group_commit_failure_fails_all_waiters_before_dir_fsync() {
        use std::sync::mpsc;

        let _group_commit = set_file_fdatasync_group_commit_for_test(true);
+        set_file_fdatasync_group_commit_wait_for_test(0);
        clear_file_fdatasync_group_commit_for_test();
        let temp_dir = tempdir().expect("create temp dir");
        let first_dir = temp_dir.path().join("first");
@@ -249,7 +249,7 @@ impl PoolEndpointList {
            endpoint.set_set_index(0);
            endpoint.set_disk_index(0);

-            // TODO Check for cross device mounts if any.
+            // TODO(backlog): check for cross-device mounts in single-drive setup

            return Ok(Self {
                inner: vec![Endpoints::from(vec![endpoint])],
@@ -264,7 +264,7 @@ impl PoolEndpointList {
                // Convert args to endpoints
                let mut eps = Endpoints::try_from(set_layout.as_slice())?;

-                // TODO Check for cross device mounts if any.
+                // TODO(backlog): check for cross-device mounts in multi-pool setup

                for (disk_idx, ep) in eps.as_mut().iter_mut().enumerate() {
                    ep.set_pool_index(pool_idx);
@@ -689,6 +689,9 @@ impl ObjectInfo {
    }

    pub fn get_actual_size(&self) -> std::io::Result<i64> {
+        if self.actual_size < -1 || (self.actual_size == -1 && !self.is_compressed()) {
+            return Err(std::io::Error::other("invalid negative actual size"));
+        }
        if self.actual_size > 0 {
            return Ok(self.actual_size);
        }
@@ -700,10 +703,25 @@ impl ObjectInfo {
                let size = size_str.parse::<i64>().map_err(|e| std::io::Error::other(e.to_string()))?;
                return Ok(size);
            }
-            let mut actual_size = 0;
-            self.parts.iter().for_each(|part| {
-                actual_size += part.actual_size;
-            });
+            if self.actual_size == -1 && self.parts.is_empty() {
+                return Ok(-1);
+            }
+            let mut actual_size = 0_i64;
+            let mut unknown = false;
+            for part in self.parts.iter() {
+                match part.actual_size {
+                    -1 => unknown = true,
+                    size if size >= 0 => {
+                        actual_size = actual_size
+                            .checked_add(size)
+                            .ok_or_else(|| std::io::Error::other("compressed actual size overflow"))?;
+                    }
+                    _ => return Err(std::io::Error::other("invalid negative compressed part size")),
+                }
+            }
+            if unknown {
+                return Ok(-1);
+            }
            if actual_size == 0 && actual_size != self.size {
                return Err(std::io::Error::other(format!("invalid decompressed size {} {}", actual_size, self.size)));
            }
@@ -718,6 +736,18 @@ impl ObjectInfo {
        Ok(self.size)
    }

+    /// Returns a non-negative size for client and replication boundaries.
+    ///
+    /// Compressed legacy metadata can retain the internal `-1` unknown-size
+    /// sentinel. Those boundaries cannot emit a negative length, so they use
+    /// the persisted physical size while quota accounting keeps the sentinel
+    /// distinction in [`crate::data_usage::quota_object_size`].
+    pub fn get_actual_size_or_physical(&self) -> i64 {
+        self.get_actual_size()
+            .map(|size| if size >= 0 { size } else { self.size.max(0) })
+            .unwrap_or_else(|_| self.size.max(0))
+    }
+
    pub fn from_file_info(fi: &FileInfo, bucket: &str, object: &str, versioned: bool) -> ObjectInfo {
        let mut version_id = fi.version_id;

@@ -1091,7 +1121,7 @@ impl ObjectInfo {
                    }
                };

-                // TODO:VersionPurgeStatus
+                // TODO(backlog): handle VersionPurgeStatus in object listing
                let versioned = vcfg.clone().map(|v| v.0.versioned(&entry.name)).unwrap_or_default();
                objects.push(ObjectInfo::from_file_info(&fi, bucket, &entry.name, versioned));

@@ -97,7 +97,7 @@ use crate::storage_api_contracts::{
        CompletePart, ListMultipartsInfo, ListPartsInfo, MultipartInfo, MultipartOperations as _, MultipartUploadResult, PartInfo,
    },
    namespace::NamespaceLocking as _,
-    object::{DeletedObject, HTTPPreconditions, ObjectIO as _, ObjectOperations as _, ObjectToDelete},
+    object::{DeleteAccounting, DeletedObject, HTTPPreconditions, ObjectIO as _, ObjectOperations as _, ObjectToDelete},
    range::HTTPRangeSpec,
 };
 use crate::store::utils::is_reserved_or_invalid_bucket;
@@ -922,14 +922,10 @@ mod prepared_get_object_metadata_tests {
            .expect("test should find an object whose initial fanout covers both data shards")
    }

-    #[allow(
-        dead_code,
-        reason = "test fixture no assertion in this module uses today; the live namesake lives in io_primitives tests (backlog#1823)"
-    )]
-    fn bounded_spare_disk_index(bucket: &str, object: &str) -> usize {
+    fn bounded_initial_parity_disk_index(bucket: &str, object: &str) -> usize {
        *bounded_metadata_fanout_order(bucket, object, 4, 2)
-            .get(3)
-            .expect("4-disk test geometry should leave one bounded spare disk")
+            .get(2)
+            .expect("4-disk test geometry should schedule one parity disk initially")
    }

    #[tokio::test]
@@ -1087,7 +1083,7 @@ mod prepared_get_object_metadata_tests {
                        ("RUSTFS_GET_METADATA_EARLY_STOP_BOUNDED_FANOUT", None::<&str>),
                    ],
                    async {
-                        let slow_parity_disk = bounded_spare_disk_index(bucket, &object);
+                        let slow_parity_disk = bounded_initial_parity_disk_index(bucket, &object);
                        let barrier =
                            rename_fanout_barrier::arm(&object, slow_parity_disk, rename_fanout_barrier::PHASE_READ_VERSION);
                        let calls = disk_call_counters::observe(&object);
@@ -1575,7 +1575,7 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
        let parts_metadata = vec![fi.clone(); disks.len()];

        if !user_defined.contains_key("content-type") {
-            // TODO: get content-type
+            // TODO(backlog): detect content-type from part data when header is missing
        }

        if let Some(sc) = user_defined.get(AMZ_STORAGE_CLASS)
@@ -1971,7 +1971,7 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
                return Err(Error::InvalidPart(p.part_num, ext_part.etag.clone(), p.etag.clone().unwrap_or_default()));
            }

-            // TODO: crypto
+            // TODO(backlog): integrate encryption verification during complete multipart

            if (i < uploaded_parts.len() - 1)
                && !(opts.data_movement && ext_part.actual_size < 0)
@@ -45,6 +45,7 @@ use crate::bucket::replication::{
    DeleteReplicationConfigSnapshot, ReplicationLifecycleBridge, ReplicationStatusType, VersionPurgeStatusType,
    replication_state_to_filemeta, replication_status_from_filemeta, version_purge_status_to_filemeta,
 };
+use crate::data_usage::quota_object_size;
 use crate::diagnostics::get::GetObjectFailureReason;
 use crate::disk::{DataDirDeleteStatus, OldCurrentSize};
 use crate::error::is_err_invalid_upload_id;
@@ -5655,7 +5656,18 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
        objects: Vec<ObjectToDelete>,
        opts: ObjectOptions,
    ) -> (Vec<DeletedObject>, Vec<Option<Error>>) {
+        let (deleted, errors, _) = self.delete_objects_with_accounting(bucket, objects, opts).await;
+        (deleted, errors)
+    }
+
+    async fn delete_objects_with_accounting(
+        &self,
+        bucket: &str,
+        objects: Vec<ObjectToDelete>,
+        opts: ObjectOptions,
+    ) -> (Vec<DeletedObject>, Vec<Option<Error>>, Vec<Option<DeleteAccounting>>) {
        let mut del_objects = vec![DeletedObject::default(); objects.len()];
+        let mut accounting = vec![None; objects.len()];
        let delete_config_snapshot = opts
            .delete_replication_config_snapshot
            .clone()
@@ -5745,7 +5757,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                                *item = Some(Error::other(message.clone()));
                            }
                        }
-                        return (del_objects, del_errs);
+                        return (del_objects, del_errs, accounting);
                    }
                },
            }
@@ -5792,6 +5804,22 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
            let source_missing = gerr
                .as_ref()
                .is_some_and(|err| is_err_object_not_found(err) || is_err_version_not_found(err));
+            // Resolve accounting from the generation selected under this
+            // object's write lock. A request-layer pre-stat is only an
+            // optimization and cannot identify a concurrent overwrite.
+            let (accounting_size, accounting_version_id, removed_current_object) = if source_missing
+                || dobj.synthetic_version_id
+                || set_disk_delete_creates_delete_marker(&check_opts)
+                || goi.delete_marker
+            {
+                (None, None, false)
+            } else {
+                (
+                    quota_object_size(&goi).ok(),
+                    goi.version_id.filter(|version_id| !version_id.is_nil()),
+                    (dobj.version_id.is_none() || is_explicit_null_version(dobj.version_id)) && !dobj.synthetic_version_id,
+                )
+            };
            // Normalize both sides before comparing. `goi.version_id` is the
            // client-facing identity, where `from_file_info` synthesizes
            // `Some(Uuid::nil())` for a null version on a versioned or
@@ -5920,7 +5948,12 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                    },
                    replication_state: vr.replication_state_internal.clone(),
                    ..Default::default()
-                }
+                };
+                accounting[i] = Some(DeleteAccounting {
+                    size: accounting_size,
+                    version_id: accounting_version_id,
+                    removed_current_object,
+                });
            }

            // Only add to vers_map if we hold the lock
@@ -5966,7 +5999,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                    });
                }
            }
-            return (del_objects, del_errs);
+            return (del_objects, del_errs, accounting);
        }

        let mut persisted_journal_entries = Vec::with_capacity(journal_entries.len());
@@ -6161,7 +6194,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {

        join_all(rollback_futures).await;

-        // TODO: add_partial
+        // TODO(backlog): support partial object deletion for multi-part objects

        if let Some(api) = opts.tier_delete_journal_api.as_ref() {
            for (idx, je) in persisted_journal_entries {
@@ -6204,7 +6237,16 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
            }
        }

-        (del_objects, del_errs)
+        // An accounting identity is actionable only when the delete result is
+        // successful. Never let a failed commit (including a partial quorum
+        // failure) reach the request-layer fast delta path.
+        for (index, err) in del_errs.iter().enumerate() {
+            if err.is_some() {
+                accounting[index] = None;
+            }
+        }
+
+        (del_objects, del_errs, accounting)
    }

    #[tracing::instrument(skip(self))]
@@ -6371,7 +6413,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
            }
        }

-        // TODO: Lifecycle
+        // TODO(backlog): integrate lifecycle evaluation before object deletion

        let mut version_found = true;
        // delete_object_version below derives its own majority quorum from the
@@ -6465,7 +6507,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                mark_deleted: mark_delete,
                mod_time: Some(mod_time),
                replication_state_internal: opts.delete_replication.as_ref().map(replication_state_to_filemeta),
-                ..Default::default() // TODO: Transition
+                ..Default::default() // TODO(backlog): populate transition state on delete markers
            };

            fi.set_tier_free_version_id(&find_vid.to_string());
@@ -6533,6 +6575,12 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {

        let mut obj_info = ObjectInfo::from_file_info(&dfi, bucket, object, opts.versioned || opts.version_suspended);
        obj_info.size = goi.size;
+        // Keep the committed source metadata on the internal delete result so
+        // the request layer can derive canonical accounting for this exact
+        // generation. Delete responses do not expose these fields.
+        obj_info.actual_size = goi.actual_size;
+        obj_info.user_defined = Arc::clone(&goi.user_defined);
+        obj_info.parts = Arc::clone(&goi.parts);
        obj_info.user_tags = Arc::clone(&goi.user_tags);
        self.invalidate_get_object_metadata_cache(bucket, object).await;
        Ok(obj_info)
@@ -7824,6 +7872,113 @@ mod replication_quota_safety_tests {
        assert_eq!(stored.get_actual_size().expect("stored logical size should parse"), 1);
    }

+    #[tokio::test]
+    async fn delete_returns_canonical_compressed_accounting_size() {
+        let (_temp_dirs, disks, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "compressed-delete-accounting";
+        for disk in &disks {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+
+        let mut user_defined = HashMap::new();
+        insert_str(
+            &mut user_defined,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        insert_str(&mut user_defined, SUFFIX_ACTUAL_SIZE, "1000".to_string());
+        let mut reader = PutObjReader::new(
+            HashReader::from_stream(Cursor::new(vec![0x5a; 400]), 400, 1000, None, None, false)
+                .expect("compressed fixture reader should be valid"),
+        );
+        set_disks
+            .put_object(
+                bucket,
+                "object",
+                &mut reader,
+                &ObjectOptions {
+                    user_defined,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("compressed object should be written");
+
+        let (deleted, errors, accounting) = set_disks
+            .delete_objects_with_accounting(
+                bucket,
+                vec![ObjectToDelete {
+                    object_name: "object".to_string(),
+                    ..Default::default()
+                }],
+                ObjectOptions {
+                    object_lock_config_snapshot: Some(Arc::new(ObjectLockConfigSnapshot::new(
+                        ObjectLockConfigState::ConfirmedAbsent,
+                    ))),
+                    ..Default::default()
+                },
+            )
+            .await;
+
+        assert!(errors[0].is_none(), "compressed delete should succeed: {:?}", errors[0]);
+        assert!(deleted[0].found, "the committed object must be reported as found");
+        assert_eq!(accounting[0].as_ref().and_then(|value| value.size), Some(1000));
+        assert!(accounting[0].as_ref().is_some_and(|value| value.version_id.is_none()));
+        assert!(accounting[0].as_ref().is_some_and(|value| value.removed_current_object));
+    }
+
+    #[tokio::test]
+    async fn suspended_delete_marker_does_not_return_body_accounting() {
+        let (_temp_dirs, disks, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "suspended-delete-accounting";
+        for disk in &disks {
+            disk.make_volume(bucket).await.expect("bucket volume should be created");
+        }
+
+        let mut user_defined = HashMap::new();
+        insert_str(
+            &mut user_defined,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        insert_str(&mut user_defined, SUFFIX_ACTUAL_SIZE, "1000".to_string());
+        let mut reader = PutObjReader::new(
+            HashReader::from_stream(Cursor::new(vec![0x5a; 400]), 400, 1000, None, None, false)
+                .expect("compressed fixture reader should be valid"),
+        );
+        let suspended_opts = ObjectOptions {
+            version_suspended: true,
+            delete_replication_config_snapshot: Some(Arc::new(DeleteReplicationConfigSnapshot::from_configs_for_test(
+                s3s::dto::VersioningConfiguration {
+                    status: Some(s3s::dto::BucketVersioningStatus::from_static(s3s::dto::BucketVersioningStatus::SUSPENDED)),
+                    ..Default::default()
+                },
+                None,
+            ))),
+            user_defined,
+            object_lock_config_snapshot: Some(Arc::new(ObjectLockConfigSnapshot::new(ObjectLockConfigState::ConfirmedAbsent))),
+            ..Default::default()
+        };
+        set_disks
+            .put_object(bucket, "object", &mut reader, &suspended_opts)
+            .await
+            .expect("compressed object should be written");
+
+        let (deleted, errors, accounting) = set_disks
+            .delete_objects_with_accounting(
+                bucket,
+                vec![ObjectToDelete {
+                    object_name: "object".to_string(),
+                    ..Default::default()
+                }],
+                suspended_opts,
+            )
+            .await;
+        assert!(errors[0].is_none(), "suspended delete should create a marker: {:?}", errors[0]);
+        assert!(deleted[0].delete_marker);
+        assert!(accounting[0].is_none(), "a delete marker must not carry body accounting");
+    }
+
    #[tokio::test]
    async fn direct_put_cannot_persist_a_tiny_logical_size() {
        let (_temp_dirs, disks, set_disks) = hermetic_set_disks(4).await;
@@ -62,8 +62,8 @@ pub(crate) mod object {
    use super::{Debug, Error, FileInfo, GetObjectReader, ObjectInfo, ObjectOptions, PutObjReader};
    use crate::storage_api_contracts::range::HTTPRangeSpec;
    pub(crate) use rustfs_storage_api::{
-        DeletedObject, HTTPPreconditions, ObjectIO, ObjectLockDeleteOptions, ObjectLockRetentionOptions, ObjectOperations,
-        ObjectPreconditionError, ObjectPreconditionPart, ObjectPreconditionState, ObjectToDelete,
+        DeleteAccounting, DeletedObject, HTTPPreconditions, ObjectIO, ObjectLockDeleteOptions, ObjectLockRetentionOptions,
+        ObjectOperations, ObjectPreconditionError, ObjectPreconditionPart, ObjectPreconditionState, ObjectToDelete,
    };

    pub(crate) trait EcstoreObjectIO:
@@ -601,7 +601,7 @@ impl ECStore {

    #[instrument(skip(self))]
    pub(super) async fn handle_list_bucket(&self, opts: &BucketOptions) -> Result<Vec<BucketInfo>> {
-        // TODO: opts.cached
+        // TODO(backlog): support cached bucket listing via opts.cached

        let mut buckets = self.peer_sys.list_bucket(opts).await?;

@@ -4673,7 +4673,7 @@ async fn gather_results(
            entry.name = entry.name.replace("\\", "/");
        }

-        // TODO: rx.recv()
+        // TODO(backlog): integrate rx.recv() for incremental listing results

        if let Some(marker) = &opts.marker
            && ((!opts.include_marker && &entry.name <= marker) || (opts.include_marker && &entry.name < marker))
@@ -4703,7 +4703,7 @@ async fn gather_results(
            continue;
        }

-        // TODO: Lifecycle
+        // TODO(backlog): integrate lifecycle evaluation during object listing

        entries.push(Some(entry));
        candidate_entries += 1;
@@ -343,6 +343,23 @@ impl ECStore {
        let (decommission, rebalance) = tokio::join!(self.is_decommission_running(), self.is_rebalance_started());
        decommission || rebalance
    }
+
+    /// Returns whether scanner metadata may still be hidden by a local
+    /// data-movement state. Terminal failed/canceled decommission entries
+    /// remain suspended until an operator clears or retries them, so they are
+    /// a publication barrier even after the worker has stopped.
+    pub async fn scanner_data_usage_publication_blocked(&self) -> bool {
+        if self.scanner_data_movement_active().await {
+            return true;
+        }
+
+        let pool_meta = self.pool_meta.read().await;
+        pool_meta.pools.iter().any(|pool| {
+            pool.decommission
+                .as_ref()
+                .is_some_and(|info| !info.queued && (info.failed || info.canceled))
+        })
+    }
 }

 // impl Clone for ECStore {
@@ -875,6 +892,7 @@ impl crate::storage_api_contracts::admin::StorageAdminApi for ECStore {
 #[cfg(test)]
 mod tests {
    use super::*;
+    use crate::core::pools::{PoolDecommissionInfo, PoolStatus};
    use crate::layout::endpoints::{Endpoints, PoolEndpoints, SetupType};
    use crate::runtime::global::reset_local_disk_test_state;
    use crate::runtime::sources::{clear_local_disk_id_map_for_test, local_disk_path_by_id};
@@ -911,6 +929,72 @@ mod tests {
        })
    }

+    #[tokio::test]
+    async fn scanner_data_usage_publication_blocks_active_and_unqueued_terminal_decommission() {
+        let store = build_store_with_ctx(Arc::new(InstanceContext::new()));
+        let cases = [
+            (
+                "active",
+                PoolDecommissionInfo {
+                    start_time: Some(OffsetDateTime::now_utc()),
+                    ..Default::default()
+                },
+                true,
+            ),
+            (
+                "failed",
+                PoolDecommissionInfo {
+                    failed: true,
+                    ..Default::default()
+                },
+                true,
+            ),
+            (
+                "canceled",
+                PoolDecommissionInfo {
+                    canceled: true,
+                    ..Default::default()
+                },
+                true,
+            ),
+            (
+                "queued_failed",
+                PoolDecommissionInfo {
+                    failed: true,
+                    queued: true,
+                    ..Default::default()
+                },
+                false,
+            ),
+            (
+                "complete",
+                PoolDecommissionInfo {
+                    complete: true,
+                    ..Default::default()
+                },
+                false,
+            ),
+            ("idle", PoolDecommissionInfo::default(), false),
+        ];
+
+        for (name, decommission, expected) in cases {
+            *store.pool_meta.write().await = PoolMeta {
+                pools: vec![PoolStatus {
+                    id: 0,
+                    cmd_line: format!("scanner-publication-{name}"),
+                    last_update: OffsetDateTime::now_utc(),
+                    decommission: Some(decommission),
+                }],
+                ..Default::default()
+            };
+            assert_eq!(
+                store.scanner_data_usage_publication_blocked().await,
+                expected,
+                "unexpected scanner publication barrier state for {name}"
+            );
+        }
+    }
+
    // The object graph is the isolation carrier: two ECStore instances holding
    // distinct contexts report independent erasure state through their real
    // `&self` accessors — no cross-contamination.
@@ -332,7 +332,7 @@ impl ECStore {
        let expected_incarnation_id = opts.expected_bucket_incarnation_id;

        if request.prefix.is_empty() {
-            // TODO: return from cache
+            // TODO(backlog): return cached multipart listing when prefix is empty
        }

        if self.single_pool() {
@@ -610,7 +610,7 @@ impl ECStore {
        let (opts, _bucket_lifecycle_guard) = self.guard_multipart_bucket_incarnation(bucket, opts).await?;
        let opts = &opts;

-        // TODO: defer DeleteUploadID
+        // TODO(backlog): defer DeleteUploadID to background for faster abort response

        if self.single_pool() {
            return self.pools[0].abort_multipart_upload(bucket, object, upload_id, opts).await;
@@ -41,7 +41,7 @@ use crate::set_disk::{
 };
 use crate::storage_api_contracts::{
    namespace::NamespaceLocking as _,
-    object::{ObjectIO as _, ObjectOperations as _},
+    object::{DeleteAccounting, ObjectIO as _, ObjectOperations as _},
 };
 use parking_lot::Mutex as ParkingMutex;
 use rustfs_io_metrics::{
@@ -1216,6 +1216,14 @@ fn return_batch_delete_lock_error(objects: &[ObjectToDelete], err: Error) -> (Ve
    (del_objects, del_errs)
 }

+fn return_batch_delete_lock_error_with_accounting(
+    objects: &[ObjectToDelete],
+    err: Error,
+) -> (Vec<DeletedObject>, Vec<Option<Error>>, Vec<Option<DeleteAccounting>>) {
+    let (deleted, errors) = return_batch_delete_lock_error(objects, err);
+    (deleted, errors, vec![None; objects.len()])
+}
+
 fn sorted_unique_delete_object_names(objects: &[ObjectToDelete]) -> Vec<&str> {
    let mut object_names: Vec<&str> = objects.iter().map(|object| object.object_name.as_str()).collect();
    object_names.sort_unstable();
@@ -2312,6 +2320,22 @@ impl ECStore {
        result
    }

+    pub async fn delete_objects_with_tier_delete_journal_and_accounting(
+        self: &Arc<Self>,
+        bucket: &str,
+        objects: Vec<ObjectToDelete>,
+        opts: ObjectOptions,
+    ) -> (Vec<DeletedObject>, Vec<Option<Error>>, Vec<Option<DeleteAccounting>>) {
+        let result = self
+            .handle_delete_objects_with_journal_and_accounting(bucket, objects, opts, Some(Arc::clone(self)))
+            .await;
+        let success_count = result.1.iter().filter(|err| err.is_none()).count();
+        if success_count > 0 {
+            list_objects::observe_list_objects_mutations(self, bucket, success_count).await;
+        }
+        result
+    }
+
    #[instrument(skip(self))]
    pub(super) async fn handle_delete_object(&self, bucket: &str, object: &str, opts: ObjectOptions) -> Result<ObjectInfo> {
        self.handle_delete_object_with_journal(bucket, object, opts, None).await
@@ -2689,6 +2713,19 @@ impl ECStore {
        opts: ObjectOptions,
        tier_journal_api: Option<Arc<ECStore>>,
    ) -> (Vec<DeletedObject>, Vec<Option<Error>>) {
+        let (deleted, errors, _) = self
+            .handle_delete_objects_with_journal_and_accounting(bucket, objects, opts, tier_journal_api)
+            .await;
+        (deleted, errors)
+    }
+
+    pub(super) async fn handle_delete_objects_with_journal_and_accounting(
+        &self,
+        bucket: &str,
+        objects: Vec<ObjectToDelete>,
+        opts: ObjectOptions,
+        tier_journal_api: Option<Arc<ECStore>>,
+    ) -> (Vec<DeletedObject>, Vec<Option<Error>>, Vec<Option<DeleteAccounting>>) {
        // encode object name
        let objects: Vec<ObjectToDelete> = objects
            .iter()
@@ -2701,6 +2738,7 @@ impl ECStore {

        // Default return value
        let mut del_objects = vec![DeletedObject::default(); objects.len()];
+        let mut accounting = vec![None; objects.len()];

        let mut del_errs = Vec::with_capacity(objects.len());
        for _ in 0..objects.len() {
@@ -2714,7 +2752,7 @@ impl ECStore {
        } else {
            match self.acquire_bucket_lifecycle_read_lock(bucket).await {
                Ok(guard) => Some(guard),
-                Err(err) => return return_batch_delete_lock_error(objects.as_slice(), err),
+                Err(err) => return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err),
            }
        };
        if let Some(guard) = _bucket_lifecycle_guard.as_ref() {
@@ -2726,21 +2764,21 @@ impl ECStore {
                Err(err) => {
                    let message = err.to_string();
                    let errors = (0..objects.len()).map(|_| Some(Error::other(message.clone()))).collect();
-                    return (del_objects, errors);
+                    return (del_objects, errors, accounting);
                }
            }
        }
        if !is_meta_bucketname(bucket)
            && let Err(err) = get_cached_bucket_incarnation_id_in(&self.ctx, bucket).await
        {
-            return return_batch_delete_lock_error(objects.as_slice(), err);
+            return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err);
        }
        let _object_lock_metadata_guard = if is_meta_bucketname(bucket) {
            None
        } else {
            Some(match acquire_bucket_metadata_transaction_read_lock_in(&self.ctx, bucket).await {
                Ok(guard) => guard,
-                Err(err) => return return_batch_delete_lock_error(objects.as_slice(), err),
+                Err(err) => return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err),
            })
        };
        if let Some(guard) = _object_lock_metadata_guard.as_ref() {
@@ -2750,7 +2788,7 @@ impl ECStore {
            let (state, incarnation_id, config_revision) =
                match get_object_lock_config_and_incarnation_from_disk_in(&self.ctx, bucket).await {
                    Ok(snapshot) => snapshot,
-                    Err(err) => return return_batch_delete_lock_error(objects.as_slice(), err),
+                    Err(err) => return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err),
                };
            opts.object_lock_config_snapshot = Some(Arc::new(ObjectLockConfigSnapshot::for_store_bucket(
                self.id,
@@ -2766,7 +2804,10 @@ impl ECStore {
        if let (Some(expected), Some(current)) = (opts.expected_bucket_incarnation_id, current_bucket_incarnation_id)
            && expected != current
        {
-            return return_batch_delete_lock_error(objects.as_slice(), StorageError::BucketNotFound(bucket.to_string()));
+            return return_batch_delete_lock_error_with_accounting(
+                objects.as_slice(),
+                StorageError::BucketNotFound(bucket.to_string()),
+            );
        }
        #[cfg(test)]
        if current_bucket_incarnation_id.is_some() {
@@ -2774,7 +2815,7 @@ impl ECStore {
        }
        let _object_lock_guards = match self.acquire_delete_objects_write_locks(bucket, &objects, &mut opts).await {
            Ok(guards) => guards,
-            Err(err) => return return_batch_delete_lock_error(objects.as_slice(), err),
+            Err(err) => return return_batch_delete_lock_error_with_accounting(objects.as_slice(), err),
        };

        let mut futures = Vec::with_capacity(self.pools.len());
@@ -2783,22 +2824,24 @@ impl ECStore {
            if self.is_pool_rebalancing(pool.pool_idx).await {
                continue;
            }
-            futures.push(pool.delete_objects(bucket, objects.clone(), opts.clone()));
+            futures.push(pool.delete_objects_with_accounting(bucket, objects.clone(), opts.clone()));
        }

        let results = join_all(futures).await;

        for idx in 0..del_objects.len() {
-            for (dels, errs) in results.iter() {
+            for (dels, errs, pool_accounting) in results.iter() {
                if errs[idx].is_none() && dels[idx].found {
                    del_errs[idx] = None;
                    del_objects[idx] = dels[idx].clone();
+                    accounting[idx] = pool_accounting[idx].clone();
                    break;
                }

                if del_errs[idx].is_none() {
                    del_errs[idx] = errs[idx].clone();
                    del_objects[idx] = dels[idx].clone();
+                    accounting[idx] = pool_accounting[idx].clone();
                }
            }
        }
@@ -2807,7 +2850,7 @@ impl ECStore {
            v.object_name = decode_dir_object(&v.object_name);
        });

-        (del_objects, del_errs)
+        (del_objects, del_errs, accounting)

        // let mut futures = Vec::with_capacity(objects.len());

@@ -385,7 +385,7 @@ impl ECStore {
    }

    pub(super) async fn is_suspended(&self, idx: usize) -> bool {
-        // TODO: LOCK
+        // TODO(backlog): acquire pool metadata lock for consistent suspension check

        let pool_meta = self.pool_meta.read().await;

@@ -17,11 +17,11 @@
 //! All direct `rustfs_ecstore` facade imports used by tests in this crate
 //! must go through this module (architecture migration rule:
 //! `check_architecture_migration_rules.sh`). Keep the surface minimal —
-//! only what the tests actually need to build a temp-disk ECStore fixture
-//! and to flip the erasure setup type for lock-quorum fault injection.
+//! only what the tests actually need to run storage-backed IAM scenarios.

 #[allow(unused_imports)]
 pub(crate) mod fixture {
+    pub(crate) use rustfs_ecstore::api::bucket::migration::try_migrate_iam_config;
    pub(crate) use rustfs_ecstore::api::layout::SetupType;

    // `update_erasure_type` is a write-side global facade entry. Its use is
@@ -0,0 +1,182 @@
+// 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.
+
+mod ecstore_test_compat;
+
+use ecstore_test_compat::fixture::try_migrate_iam_config;
+use rustfs_credentials::{get_global_action_cred, init_global_action_credentials};
+use rustfs_iam::store::object::{
+    IAM_CONFIG_POLICY_DB_SERVICE_ACCOUNTS_PREFIX, IAM_CONFIG_POLICY_DB_USERS_PREFIX, IAM_CONFIG_SERVICE_ACCOUNTS_PREFIX,
+    IAM_CONFIG_USERS_PREFIX, ObjectStore,
+};
+use rustfs_iam::store::{Store, UserType};
+use rustfs_iam::utils::generate_jwt;
+use rustfs_policy::auth::UserIdentity;
+use serde_json::{Value, json};
+use std::collections::HashMap;
+
+const LEGACY_META_BUCKET: &str = ".minio.sys";
+const REGULAR_USER: &str = "minio-user";
+const SERVICE_ACCOUNT: &str = "minio-service-account";
+
+async fn seed_legacy_iam_object(env: &rustfs_test_utils::TestECStoreEnv, path: &str, value: &Value) {
+    env.put_object_bytes(
+        LEGACY_META_BUCKET,
+        path,
+        serde_json::to_vec(value).expect("legacy IAM object must serialize"),
+    )
+    .await;
+}
+
+fn assert_identity_fields(actual: &UserIdentity, expected: &Value) {
+    assert_eq!(
+        serde_json::to_value(actual).expect("loaded identity must serialize"),
+        *expected,
+        "migration must preserve every credential field except expiration",
+    );
+}
+
+async fn assert_identity_survives(
+    store: &ObjectStore,
+    identity_path: &str,
+    name: &str,
+    user_type: UserType,
+    source: &Value,
+    expected_policy: &Value,
+) {
+    let mut expected = source.clone();
+    expected["credentials"]["expiration"] = Value::Null;
+
+    let persisted: UserIdentity = store
+        .load_iam_config(identity_path)
+        .await
+        .expect("migrated identity must be persisted");
+    assert_identity_fields(&persisted, &expected);
+
+    for _ in 0..2 {
+        let actual = store
+            .load_user_identity(name, user_type)
+            .await
+            .expect("migrated permanent identity must remain loadable");
+        assert_identity_fields(&actual, &expected);
+    }
+
+    let mut mappings = HashMap::new();
+    store
+        .load_mapped_policy(name, user_type, false, &mut mappings)
+        .await
+        .expect("loading the identity must not delete its policy mapping");
+    let actual_policy = mappings.get(name).expect("migrated policy mapping must exist");
+    assert_eq!(
+        serde_json::to_value(actual_policy).expect("loaded policy mapping must serialize"),
+        *expected_policy,
+    );
+}
+
+#[tokio::test(flavor = "multi_thread")]
+async fn minio_permanent_identities_survive_migration_and_repeated_iam_loads() {
+    if get_global_action_cred().is_none() {
+        init_global_action_credentials(Some("MINIOMIGRATIONROOT".to_string()), Some("minio-migration-root-secret".to_string()))
+            .expect("root credentials must initialize for JWT validation");
+    }
+
+    let temp_dir = tempfile::TempDir::with_prefix("rustfs_minio_iam_migration_").expect("temp directory must be created");
+    let env = rustfs_test_utils::TestECStoreEnv::builder()
+        .base_dir(temp_dir.path())
+        .init_bucket_metadata(false)
+        .build()
+        .await;
+    for disk_path in &env.disk_paths {
+        tokio::fs::create_dir_all(disk_path.join(LEGACY_META_BUCKET))
+            .await
+            .expect("legacy metadata volume must be created");
+    }
+
+    let regular_source = json!({
+        "version": 1,
+        "credentials": {
+            "accessKey": REGULAR_USER,
+            "secretKey": "regular-user-secret",
+            "sessionToken": "",
+            "expiration": "0001-01-01T00:00:00Z",
+            "status": "on",
+            "parentUser": "regular-parent",
+            "groups": ["engineering", "operations"],
+            "claims": {"tenant": "alpha"},
+            "name": "MinIO regular user",
+            "description": "migrated regular identity"
+        },
+        "updatedAt": "2025-03-07T12:00:00Z"
+    });
+    let service_claims = json!({"sa-policy": "inherited-policy", "tenant": "alpha"});
+    let service_secret = "service-account-secret";
+    let service_source = json!({
+        "version": 1,
+        "credentials": {
+            "accessKey": SERVICE_ACCOUNT,
+            "secretKey": service_secret,
+            "sessionToken": generate_jwt(&service_claims, service_secret).expect("service-account JWT must be generated"),
+            "expiration": "1970-01-01T00:00:00Z",
+            "status": "on",
+            "parentUser": REGULAR_USER,
+            "groups": ["service-accounts"],
+            "claims": service_claims,
+            "name": "MinIO service account",
+            "description": "migrated service identity"
+        },
+        "updatedAt": "2025-03-07T12:00:00Z"
+    });
+    let regular_policy_source = json!({"version": 1, "policy": "readwrite", "updatedAt": "2025-03-07T12:00:00Z"});
+    let service_policy_source = json!({"version": 1, "policy": "readonly", "updatedAt": "2025-03-07T12:00:00Z"});
+
+    let regular_identity_path = format!("{}{REGULAR_USER}/identity.json", IAM_CONFIG_USERS_PREFIX.as_str());
+    let service_identity_path = format!("{}{SERVICE_ACCOUNT}/identity.json", IAM_CONFIG_SERVICE_ACCOUNTS_PREFIX.as_str());
+
+    seed_legacy_iam_object(&env, &regular_identity_path, &regular_source).await;
+    seed_legacy_iam_object(&env, &service_identity_path, &service_source).await;
+    seed_legacy_iam_object(
+        &env,
+        &format!("{}{REGULAR_USER}.json", IAM_CONFIG_POLICY_DB_USERS_PREFIX.as_str()),
+        &regular_policy_source,
+    )
+    .await;
+    seed_legacy_iam_object(
+        &env,
+        &format!("{}{SERVICE_ACCOUNT}.json", IAM_CONFIG_POLICY_DB_SERVICE_ACCOUNTS_PREFIX.as_str()),
+        &service_policy_source,
+    )
+    .await;
+
+    try_migrate_iam_config(env.ecstore.clone(), None).await;
+
+    let store = ObjectStore::new(env.ecstore);
+    assert_identity_survives(
+        &store,
+        &regular_identity_path,
+        REGULAR_USER,
+        UserType::Reg,
+        &regular_source,
+        &regular_policy_source,
+    )
+    .await;
+    assert_identity_survives(
+        &store,
+        &service_identity_path,
+        SERVICE_ACCOUNT,
+        UserType::Svc,
+        &service_source,
+        &service_policy_source,
+    )
+    .await;
+}
@@ -1081,18 +1081,6 @@ async fn run_data_scanner_cycle(
        }
    };
    let (sender, receiver) = mpsc::channel::<DataUsageInfo>(1);
-    let storeapi_clone = storeapi.clone();
-    let ctx_clone = ctx.clone();
-    let mut usage_persist_task = AbortOnDropHandle::new(tokio::spawn(async move {
-        store_data_usage_in_backend_with_outcome_for_epoch_and_baseline(
-            ctx_clone,
-            storeapi_clone,
-            receiver,
-            Some(leader_epoch),
-            Some(usage_persist_baseline),
-        )
-        .await
-    }));

    let done_cycle = Metrics::time(Metric::ScanCycle);
    let cycle_budget = ScannerCycleBudget::new(ctx, cycle_budget_config);
@@ -1107,47 +1095,78 @@ async fn run_data_scanner_cycle(
            scan_mode,
        )
        .await;
+    let publication_defer_reason = match &scan_result {
+        Ok(result) => final_data_usage_publication_defer_reason(storeapi.as_ref(), result.status).await,
+        Err(_) => Some(ScannerCycleDeferReason::ActivityBaselineUnavailable),
+    };
    let budget_elapsed = cycle_budget.budget_elapsed() && !ctx.is_cancelled();
-    let usage_persist_outcome = match wait_for_data_usage_persist_task(ctx, &mut usage_persist_task, usage_persist_timeout).await
-    {
-        DataUsagePersistTaskResult::Completed(outcome) => outcome,
-        DataUsagePersistTaskResult::JoinFailed(err) => {
-            error!(
-                target: "rustfs::scanner",
-                event = EVENT_SCANNER_PERSIST_STATE,
-                component = LOG_COMPONENT_SCANNER,
-                subsystem = LOG_SUBSYSTEM_RUNTIME,
-                cycle = cycle_info.current,
-                state = "usage_persist_task_failed",
-                error = %err,
-                "Scanner data usage persistence task failed"
-            );
-            DataUsagePersistOutcome::Failed
+    let usage_persist_outcome = match publication_defer_reason {
+        Some(reason) => {
+            drop(receiver);
+            DataUsagePersistOutcome::Deferred(reason)
        }
-        DataUsagePersistTaskResult::Cancelled => {
-            debug!(
-                target: "rustfs::scanner",
-                event = EVENT_SCANNER_PERSIST_STATE,
-                component = LOG_COMPONENT_SCANNER,
-                subsystem = LOG_SUBSYSTEM_RUNTIME,
-                cycle = cycle_info.current,
-                state = "usage_persist_task_cancelled",
-                "Scanner data usage persistence task cancelled"
-            );
-            DataUsagePersistOutcome::Failed
-        }
-        DataUsagePersistTaskResult::TimedOut => {
-            error!(
-                target: "rustfs::scanner",
-                event = EVENT_SCANNER_PERSIST_STATE,
-                component = LOG_COMPONENT_SCANNER,
-                subsystem = LOG_SUBSYSTEM_RUNTIME,
-                cycle = cycle_info.current,
-                timeout = ?usage_persist_timeout,
-                state = "usage_persist_task_timed_out",
-                "Scanner data usage persistence task timed out"
-            );
-            DataUsagePersistOutcome::Failed
+        None => {
+            // ScannerIO emits its complete or observational update only after
+            // all set workers finish. Persist after the final activity fence;
+            // this also avoids blocking the scanner on a denied publication.
+            let storeapi_clone = storeapi.clone();
+            let ctx_clone = ctx.clone();
+            let route_probe_store = storeapi.clone();
+            let mut usage_persist_task = AbortOnDropHandle::new(tokio::spawn(async move {
+                store_data_usage_in_backend_with_outcome_for_epoch_and_baseline_and_route_probe(
+                    ctx_clone,
+                    storeapi_clone,
+                    receiver,
+                    Some(leader_epoch),
+                    Some(usage_persist_baseline),
+                    move || {
+                        let storeapi = route_probe_store.clone();
+                        async move { storeapi.scanner_data_usage_publication_blocked().await }
+                    },
+                )
+                .await
+            }));
+            match wait_for_data_usage_persist_task(ctx, &mut usage_persist_task, usage_persist_timeout).await {
+                DataUsagePersistTaskResult::Completed(outcome) => outcome,
+                DataUsagePersistTaskResult::JoinFailed(err) => {
+                    error!(
+                        target: "rustfs::scanner",
+                        event = EVENT_SCANNER_PERSIST_STATE,
+                        component = LOG_COMPONENT_SCANNER,
+                        subsystem = LOG_SUBSYSTEM_RUNTIME,
+                        cycle = cycle_info.current,
+                        state = "usage_persist_task_failed",
+                        error = %err,
+                        "Scanner data usage persistence task failed"
+                    );
+                    DataUsagePersistOutcome::Failed
+                }
+                DataUsagePersistTaskResult::Cancelled => {
+                    debug!(
+                        target: "rustfs::scanner",
+                        event = EVENT_SCANNER_PERSIST_STATE,
+                        component = LOG_COMPONENT_SCANNER,
+                        subsystem = LOG_SUBSYSTEM_RUNTIME,
+                        cycle = cycle_info.current,
+                        state = "usage_persist_task_cancelled",
+                        "Scanner data usage persistence task cancelled"
+                    );
+                    DataUsagePersistOutcome::Failed
+                }
+                DataUsagePersistTaskResult::TimedOut => {
+                    error!(
+                        target: "rustfs::scanner",
+                        event = EVENT_SCANNER_PERSIST_STATE,
+                        component = LOG_COMPONENT_SCANNER,
+                        subsystem = LOG_SUBSYSTEM_RUNTIME,
+                        cycle = cycle_info.current,
+                        timeout = ?usage_persist_timeout,
+                        state = "usage_persist_task_timed_out",
+                        "Scanner data usage persistence task timed out"
+                    );
+                    DataUsagePersistOutcome::Failed
+                }
+            }
        }
    };
    let unresolved_heal_work = global_metrics().current_scan_cycle_has_unresolved_heal_work();
@@ -1191,33 +1210,51 @@ async fn run_data_scanner_cycle(
        mark_scan_cycle_idle(cycle_info, &mut cycle_metrics_guard).await;
        return ScannerCycleOutcome::Failed;
    }
-    if let Some(required_cycle) = scan_cycle_result.required_cycle_floor() {
-        warn!(
-            target: "rustfs::scanner",
-            event = EVENT_SCANNER_CYCLE_STATE,
-            component = LOG_COMPONENT_SCANNER,
-            subsystem = LOG_SUBSYSTEM_RUNTIME,
-            cycle = cycle_info.current,
-            required_cycle,
-            state = "cache_cycle_ahead",
-            "Scanner cycle is recovering to a newer durable cache generation"
-        );
-        emit_scan_cycle_partial_with_source(cycle_start.elapsed(), ScanCyclePartialReason::Unknown, None);
-        return if persist_required_scanner_cycle_floor(
-            ctx,
-            storeapi.clone(),
-            cycle_info,
-            cycle_revision,
-            leader_epoch,
-            required_cycle,
-            &mut cycle_metrics_guard,
-        )
-        .await
-        {
-            ScannerCycleOutcome::Partial
-        } else {
-            ScannerCycleOutcome::Failed
-        };
+    match scanner_cycle_pre_commit_outcome(scan_cycle_result.required_cycle_floor(), &usage_persist_outcome) {
+        Some(ScannerCyclePreCommitOutcome::RecoverCacheCycle(required_cycle)) => {
+            warn!(
+                target: "rustfs::scanner",
+                event = EVENT_SCANNER_CYCLE_STATE,
+                component = LOG_COMPONENT_SCANNER,
+                subsystem = LOG_SUBSYSTEM_RUNTIME,
+                cycle = cycle_info.current,
+                required_cycle,
+                state = "cache_cycle_ahead",
+                "Scanner cycle is recovering to a newer durable cache generation"
+            );
+            emit_scan_cycle_partial_with_source(cycle_start.elapsed(), ScanCyclePartialReason::Unknown, None);
+            return if persist_required_scanner_cycle_floor(
+                ctx,
+                storeapi.clone(),
+                cycle_info,
+                cycle_revision,
+                leader_epoch,
+                required_cycle,
+                &mut cycle_metrics_guard,
+            )
+            .await
+            {
+                ScannerCycleOutcome::Partial
+            } else {
+                ScannerCycleOutcome::Failed
+            };
+        }
+        Some(ScannerCyclePreCommitOutcome::Deferred(reason)) => {
+            info!(
+                target: "rustfs::scanner",
+                event = EVENT_SCANNER_CYCLE_STATE,
+                component = LOG_COMPONENT_SCANNER,
+                subsystem = LOG_SUBSYSTEM_RUNTIME,
+                cycle = cycle_info.current,
+                reason = reason.as_str(),
+                state = "deferred",
+                "Scanner cycle deferred before data usage publication"
+            );
+            emit_scan_cycle_deferred(cycle_start.elapsed());
+            mark_scan_cycle_idle(cycle_info, &mut cycle_metrics_guard).await;
+            return ScannerCycleOutcome::Deferred(reason);
+        }
+        None => {}
    }
    if usage_persist_outcome == DataUsagePersistOutcome::Failed {
        error!(
@@ -2000,6 +2037,56 @@ impl Drop for ScannerScanModeGuard {
    }
 }

+async fn final_data_usage_publication_defer_reason(
+    storeapi: &ECStore,
+    status: ScannerCycleStatus,
+) -> Option<ScannerCycleDeferReason> {
+    match status {
+        ScannerCycleStatus::Complete | ScannerCycleStatus::Superseded => {
+            if storeapi.scanner_data_usage_publication_blocked().await {
+                return Some(ScannerCycleDeferReason::DataMovement);
+            }
+            if status == ScannerCycleStatus::Complete {
+                let distributed = storeapi.setup_is_dist_erasure().await;
+                match probe_scanner_activity(storeapi, distributed).await {
+                    Ok(snapshot) if scanner_activity_allows_usage_publication(&snapshot) => None,
+                    Ok(_) => Some(ScannerCycleDeferReason::DataMovement),
+                    Err(_) => Some(ScannerCycleDeferReason::ActivityBaselineUnavailable),
+                }
+            } else {
+                // A superseded cycle is explicitly observational and cannot
+                // replace the authoritative snapshot. It may still be
+                // persisted as a convergence baseline for the next cycle.
+                None
+            }
+        }
+        ScannerCycleStatus::Deferred(reason) => Some(reason),
+        // Incomplete cycles do not publish a usage snapshot. Keep the
+        // decision permissive so existing partial-cycle handling remains
+        // unchanged if a future scanner path emits a bookkeeping update.
+        ScannerCycleStatus::Incomplete => None,
+    }
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum ScannerCyclePreCommitOutcome {
+    RecoverCacheCycle(u64),
+    Deferred(ScannerCycleDeferReason),
+}
+
+fn scanner_cycle_pre_commit_outcome(
+    required_cycle_floor: Option<u64>,
+    usage_persist_outcome: &DataUsagePersistOutcome,
+) -> Option<ScannerCyclePreCommitOutcome> {
+    // Keep the publication barrier fail-closed: `.bloomcycle.bin` uses the
+    // same routed writer and its floor must remain pending while data movement
+    // hides the source pool.
+    match usage_persist_outcome {
+        DataUsagePersistOutcome::Deferred(reason) => Some(ScannerCyclePreCommitOutcome::Deferred(*reason)),
+        _ => required_cycle_floor.map(ScannerCyclePreCommitOutcome::RecoverCacheCycle),
+    }
+}
+
 fn scanner_cycle_completion_outcome(
    scan_status: ScannerCycleStatus,
    usage_persist_outcome: DataUsagePersistOutcome,
@@ -2007,6 +2094,7 @@ fn scanner_cycle_completion_outcome(
    has_failed_dirty_usage: bool,
 ) -> ScannerCycleOutcome {
    match (scan_status, usage_persist_outcome) {
+        (_, DataUsagePersistOutcome::Deferred(reason)) => ScannerCycleOutcome::Deferred(reason),
        (_, DataUsagePersistOutcome::Failed) => ScannerCycleOutcome::Failed,
        (ScannerCycleStatus::Deferred(reason), DataUsagePersistOutcome::NoUpdate)
            if !has_dirty_usage && !has_failed_dirty_usage =>
@@ -153,6 +153,7 @@ struct MemoryConfigStore {
    objects: Mutex<HashMap<String, Vec<u8>>>,
    revisions: Mutex<HashMap<String, u64>>,
    fail_put_number: Mutex<HashMap<String, usize>>,
+    object_not_found_put_number: Mutex<HashMap<String, usize>>,
    error_after_commit_put_number: Mutex<HashMap<String, usize>>,
    interleaving_puts: Mutex<HashMap<String, (usize, Vec<u8>)>>,
    cancel_after_interleaving_puts: Mutex<HashMap<String, CancellationToken>>,
@@ -224,6 +225,9 @@ impl crate::storage_api::scanner_io::ObjectIO for MemoryConfigStore {
        if self.fail_put_number.lock().await.get(&key) == Some(&put_count) {
            return Err(EcstoreError::other("injected put failure"));
        }
+        if self.object_not_found_put_number.lock().await.get(&key) == Some(&put_count) {
+            return Err(EcstoreError::ObjectNotFound(bucket.to_string(), object.to_string()));
+        }

        let interleaving_data = {
            let mut interleaving_puts = self.interleaving_puts.lock().await;
@@ -1431,6 +1435,170 @@ async fn test_store_data_usage_in_backend_preserves_newer_snapshot() {
    assert_eq!(outcome, DataUsagePersistOutcome::Current);
 }

+#[tokio::test]
+async fn test_usage_save_object_not_found_defers_only_with_a_fresh_route_barrier() {
+    for (route_blocked, expected) in [
+        (true, DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement)),
+        (false, DataUsagePersistOutcome::Failed),
+    ] {
+        let store = Arc::new(MemoryConfigStore::default());
+        let key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_OBJ_NAME_PATH.as_str());
+        let baseline = complete_usage_with_bucket_count(Some(std::time::SystemTime::UNIX_EPOCH + Duration::from_secs(10)), 1);
+        let baseline_data = serde_json::to_vec(&baseline).expect("baseline usage snapshot should encode");
+        store.objects.lock().await.insert(key.clone(), baseline_data.clone());
+        store.revisions.lock().await.insert(key.clone(), 1);
+        store.object_not_found_put_number.lock().await.insert(key.clone(), 1);
+
+        let (sender, receiver) = mpsc::channel(1);
+        sender
+            .send(complete_usage_with_bucket_count(
+                Some(std::time::SystemTime::UNIX_EPOCH + Duration::from_secs(20)),
+                2,
+            ))
+            .await
+            .expect("new usage snapshot should enqueue");
+        drop(sender);
+        let probe_calls = Arc::new(std::sync::atomic::AtomicUsize::new(0));
+        let route_probe_calls = probe_calls.clone();
+
+        let outcome = store_data_usage_in_backend_with_outcome_for_epoch_and_baseline_and_route_probe(
+            CancellationToken::new(),
+            store.clone(),
+            receiver,
+            None,
+            Some(DataUsagePersistBaseline {
+                data: Some(Bytes::from(baseline_data.clone())),
+                revision: DataUsageCacheRevision::Etag("memory-1".to_string()),
+            }),
+            move || {
+                let probe_calls = route_probe_calls.clone();
+                async move {
+                    let call = probe_calls.fetch_add(1, std::sync::atomic::Ordering::SeqCst);
+                    route_blocked && call > 1
+                }
+            },
+        )
+        .await;
+
+        assert_eq!(outcome, expected);
+        assert_eq!(
+            probe_calls.load(std::sync::atomic::Ordering::SeqCst),
+            3,
+            "ObjectNotFound must be followed by a fresh route-barrier probe"
+        );
+        assert_eq!(
+            store.objects.lock().await.get(&key),
+            Some(&baseline_data),
+            "a route failure must not replace the authoritative baseline"
+        );
+    }
+}
+
+#[tokio::test]
+async fn test_usage_save_route_barrier_prevents_missing_snapshot_creation() {
+    for observational in [false, true] {
+        let store = Arc::new(MemoryConfigStore::default());
+        let target_path = if observational {
+            DATA_USAGE_OBSERVED_OBJ_NAME_PATH.as_str()
+        } else {
+            DATA_USAGE_OBJ_NAME_PATH.as_str()
+        };
+        let target_key = memory_config_key(RUSTFS_META_BUCKET, target_path);
+        let mut incoming = complete_usage_with_bucket_count(Some(std::time::SystemTime::UNIX_EPOCH + Duration::from_secs(20)), 1);
+        incoming.usage_snapshot_converged = Some(!observational);
+        let (sender, receiver) = mpsc::channel(1);
+        sender.send(incoming).await.expect("usage snapshot should enqueue");
+        drop(sender);
+
+        let outcome = store_data_usage_in_backend_with_outcome_for_epoch_and_baseline_and_route_probe(
+            CancellationToken::new(),
+            store.clone(),
+            receiver,
+            None,
+            Some(DataUsagePersistBaseline {
+                data: None,
+                revision: DataUsageCacheRevision::Missing,
+            }),
+            || async { true },
+        )
+        .await;
+
+        assert_eq!(outcome, DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement));
+        assert!(!store.objects.lock().await.contains_key(&target_key));
+        assert_eq!(
+            store.put_counts.lock().await.get(&target_key),
+            None,
+            "the final pool-state fence must run before the first PUT"
+        );
+    }
+}
+
+#[tokio::test]
+#[serial]
+async fn test_usage_route_barrier_precedes_durable_reconciliation() {
+    let store = Arc::new(MemoryConfigStore::default());
+    let key = memory_config_key(RUSTFS_META_BUCKET, DATA_USAGE_OBJ_NAME_PATH.as_str());
+    let snapshot = complete_usage_with_bucket_count(Some(std::time::SystemTime::UNIX_EPOCH + Duration::from_secs(20)), 1);
+    let snapshot_data = serde_json::to_vec(&snapshot).expect("usage snapshot should encode");
+    let (sender, receiver) = mpsc::channel(1);
+    sender.send(snapshot).await.expect("usage snapshot should enqueue");
+    drop(sender);
+
+    let outcome = store_data_usage_in_backend_with_outcome_for_epoch_and_baseline_and_route_probe(
+        CancellationToken::new(),
+        store.clone(),
+        receiver,
+        None,
+        Some(DataUsagePersistBaseline {
+            data: Some(Bytes::from(snapshot_data)),
+            revision: DataUsageCacheRevision::Etag("memory-1".to_string()),
+        }),
+        || async { true },
+    )
+    .await;
+
+    assert_eq!(outcome, DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement));
+    assert_eq!(store.put_counts.lock().await.get(&key), None);
+}
+
+#[tokio::test]
+#[serial]
+async fn test_deferred_usage_save_keeps_last_real_save_metric() {
+    let metrics = global_metrics();
+    metrics.record_scanner_usage_save_result(ScannerUsageSaveResult::Success);
+    let before = metrics.report().await.usage_freshness;
+
+    let store = Arc::new(MemoryConfigStore::default());
+    let (sender, receiver) = mpsc::channel(1);
+    sender
+        .send(complete_usage_with_bucket_count(
+            Some(std::time::SystemTime::UNIX_EPOCH + Duration::from_secs(20)),
+            1,
+        ))
+        .await
+        .expect("usage snapshot should enqueue");
+    drop(sender);
+
+    let outcome = store_data_usage_in_backend_with_outcome_for_epoch_and_baseline_and_route_probe(
+        CancellationToken::new(),
+        store,
+        receiver,
+        None,
+        Some(DataUsagePersistBaseline {
+            data: None,
+            revision: DataUsageCacheRevision::Missing,
+        }),
+        || async { true },
+    )
+    .await;
+
+    assert_eq!(outcome, DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement));
+    let after = metrics.report().await.usage_freshness;
+    assert_eq!(after.last_usage_save_result, before.last_usage_save_result);
+    assert_eq!(after.last_usage_save_result_code, before.last_usage_save_result_code);
+    assert_eq!(after.last_usage_save_unix_secs, before.last_usage_save_unix_secs);
+}
+
 #[tokio::test]
 async fn test_store_data_usage_in_backend_fences_interleaving_newer_writer() {
    let store = Arc::new(MemoryConfigStore::default());
@@ -2325,6 +2493,15 @@ async fn test_store_data_usage_in_backend_reports_missing_snapshot() {

 #[test]
 fn test_scanner_cycle_completion_prioritizes_persist_failure() {
+    assert_eq!(
+        scanner_cycle_completion_outcome(
+            ScannerCycleStatus::Complete,
+            DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement),
+            true,
+            false,
+        ),
+        ScannerCycleOutcome::Deferred(ScannerCycleDeferReason::DataMovement)
+    );
    assert_eq!(
        scanner_cycle_completion_outcome(
            ScannerCycleStatus::Deferred(ScannerCycleDeferReason::ActivityBaselineUnavailable),
@@ -2421,6 +2598,33 @@ fn test_scanner_cycle_completion_prioritizes_persist_failure() {
    );
 }

+#[test]
+fn scanner_cycle_cache_floor_stays_pending_during_deferred_usage_publication() {
+    for reason in [
+        ScannerCycleDeferReason::DataMovement,
+        ScannerCycleDeferReason::ActivityBaselineUnavailable,
+    ] {
+        let deferred = DataUsagePersistOutcome::Deferred(reason);
+        assert_eq!(
+            scanner_cycle_pre_commit_outcome(Some(19), &deferred),
+            Some(ScannerCyclePreCommitOutcome::Deferred(reason)),
+            "a blocked publication must not persist the routed scanner cycle floor"
+        );
+        assert_eq!(
+            scanner_cycle_pre_commit_outcome(None, &deferred),
+            Some(ScannerCyclePreCommitOutcome::Deferred(reason))
+        );
+    }
+    assert_eq!(
+        scanner_cycle_pre_commit_outcome(Some(19), &DataUsagePersistOutcome::Saved),
+        Some(ScannerCyclePreCommitOutcome::RecoverCacheCycle(19))
+    );
+    assert_eq!(
+        scanner_cycle_pre_commit_outcome(Some(19), &DataUsagePersistOutcome::Failed),
+        Some(ScannerCyclePreCommitOutcome::RecoverCacheCycle(19))
+    );
+}
+
 #[test]
 #[serial]
 fn finalizing_a_saved_cycle_acknowledges_its_exact_dirty_snapshot() {
@@ -2448,6 +2652,23 @@ fn finalizing_a_saved_cycle_acknowledges_its_exact_dirty_snapshot() {
    assert!(!crate::scanner_io::dirty_usage_buckets_pending());
 }

+#[test]
+#[serial]
+fn finalizing_a_deferred_usage_save_keeps_dirty_work_pending() {
+    crate::scanner_io::clear_dirty_usage_bucket("photos");
+    crate::scanner_io::record_dirty_usage_bucket("photos");
+    let dirty_snapshot = crate::scanner_io::dirty_usage_buckets_for_tests();
+    let deferred = crate::scanner_io::ScannerCycleResult::new(ScannerCycleStatus::Complete, Some(dirty_snapshot));
+
+    let (outcome, _, acknowledgements) =
+        finalize_scanner_cycle_result(deferred, DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement));
+
+    assert_eq!(outcome, ScannerCycleOutcome::Deferred(ScannerCycleDeferReason::DataMovement));
+    assert!(acknowledgements.is_empty());
+    assert!(crate::scanner_io::dirty_usage_buckets_pending());
+    crate::scanner_io::clear_dirty_usage_bucket("photos");
+}
+
 #[tokio::test]
 async fn scanner_cycle_keeps_remote_pending_acknowledgement() {
    let pending = remote_dirty_usage_acknowledgement_pending(7, 1, std::future::ready(Ok::<bool, std::io::Error>(true))).await;
@@ -22,6 +22,10 @@ pub(super) enum DataUsagePersistOutcome {
    AlreadyDurable,
    PriorCycleDurable,
    Saved,
+    /// The metadata route is temporarily unavailable (for example while a
+    /// terminal decommission state keeps the source pool suspended).  The
+    /// caller must retry without acknowledging dirty usage.
+    Deferred(ScannerCycleDeferReason),
    Failed,
 }

@@ -92,10 +96,33 @@ pub(super) async fn store_data_usage_in_backend_with_outcome_for_epoch(
 pub(super) async fn store_data_usage_in_backend_with_outcome_for_epoch_and_baseline(
    ctx: CancellationToken,
    storeapi: Arc<impl ScannerObjectIO + ScannerConfigObjectDelete>,
-    mut receiver: mpsc::Receiver<DataUsageInfo>,
+    receiver: mpsc::Receiver<DataUsageInfo>,
    leader_epoch: Option<u64>,
    initial_baseline: Option<DataUsagePersistBaseline>,
 ) -> DataUsagePersistOutcome {
+    store_data_usage_in_backend_with_outcome_for_epoch_and_baseline_and_route_probe(
+        ctx,
+        storeapi,
+        receiver,
+        leader_epoch,
+        initial_baseline,
+        || async { false },
+    )
+    .await
+}
+
+pub(super) async fn store_data_usage_in_backend_with_outcome_for_epoch_and_baseline_and_route_probe<F, Fut>(
+    ctx: CancellationToken,
+    storeapi: Arc<impl ScannerObjectIO + ScannerConfigObjectDelete>,
+    mut receiver: mpsc::Receiver<DataUsageInfo>,
+    leader_epoch: Option<u64>,
+    initial_baseline: Option<DataUsagePersistBaseline>,
+    route_probe: F,
+) -> DataUsagePersistOutcome
+where
+    F: Fn() -> Fut + Send + Sync,
+    Fut: Future<Output = bool> + Send,
+{
    let mut outcome = DataUsagePersistOutcome::NoUpdate;
    let mut next_baseline = initial_baseline;

@@ -113,6 +140,19 @@ pub(super) async fn store_data_usage_in_backend_with_outcome_for_epoch_and_basel
        } else {
            DATA_USAGE_OBJ_NAME_PATH.as_str()
        };
+        if route_probe().await {
+            debug!(
+                target: "rustfs::scanner",
+                event = EVENT_SCANNER_PERSIST_STATE,
+                component = LOG_COMPONENT_SCANNER,
+                subsystem = LOG_SUBSYSTEM_RUNTIME,
+                path = %target_path,
+                state = "publication_blocked_before_reconcile",
+                "Scanner data usage publication deferred by the pool-state fence"
+            );
+            outcome = DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement);
+            break;
+        }

        if observational && data_usage_info.usage_snapshot_authoritative_baseline.is_none() {
            let authoritative_data = match next_baseline.as_ref() {
@@ -275,6 +315,18 @@ pub(super) async fn store_data_usage_in_backend_with_outcome_for_epoch_and_basel
            if ctx.is_cancelled() {
                break 'updates;
            }
+            if route_probe().await {
+                debug!(
+                    target: "rustfs::scanner",
+                    event = EVENT_SCANNER_PERSIST_STATE,
+                    component = LOG_COMPONENT_SCANNER,
+                    subsystem = LOG_SUBSYSTEM_RUNTIME,
+                    path = %target_path,
+                    state = "publication_blocked_before_save",
+                    "Scanner data usage publication deferred by the final pool-state fence"
+                );
+                break DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement);
+            }

            let done_save = Metrics::time(Metric::SaveUsage);
            let save_result = save_config_shared_with_preconditions(
@@ -313,6 +365,33 @@ pub(super) async fn store_data_usage_in_backend_with_outcome_for_epoch_and_basel
                        "Scanner data usage CAS conflict will be reconciled"
                    );
                }
+                Err(e @ EcstoreError::ObjectNotFound(_, _)) => {
+                    let route_blocked = route_probe().await;
+                    if route_blocked {
+                        warn!(
+                            target: "rustfs::scanner",
+                            event = EVENT_SCANNER_PERSIST_STATE,
+                            component = LOG_COMPONENT_SCANNER,
+                            subsystem = LOG_SUBSYSTEM_RUNTIME,
+                            path = %target_path,
+                            state = "publication_deferred",
+                            error = %e,
+                            "Scanner data usage route is blocked by data movement; retrying later"
+                        );
+                        break DataUsagePersistOutcome::Deferred(ScannerCycleDeferReason::DataMovement);
+                    }
+                    error!(
+                        target: "rustfs::scanner",
+                        event = EVENT_SCANNER_PERSIST_STATE,
+                        component = LOG_COMPONENT_SCANNER,
+                        subsystem = LOG_SUBSYSTEM_RUNTIME,
+                        path = %target_path,
+                        state = "save_failed",
+                        error = %e,
+                        "Scanner data usage save failed"
+                    );
+                    break DataUsagePersistOutcome::Failed;
+                }
                Err(e) => {
                    error!(
                        target: "rustfs::scanner",
@@ -370,6 +449,13 @@ pub(super) async fn store_data_usage_in_backend_with_outcome_for_epoch_and_basel
                outcome = DataUsagePersistOutcome::Failed;
                continue;
            }
+            DataUsagePersistOutcome::Deferred(reason) => {
+                // A deferred publication is an intentional retryable state, not a
+                // failed save. Keep the last real save result so admin freshness
+                // reporting does not turn a pool-recovery fence into a false error.
+                outcome = DataUsagePersistOutcome::Deferred(reason);
+                break 'updates;
+            }
            DataUsagePersistOutcome::Saved => {
                if observational {
                    invalidate_admin_data_usage_snapshot_cache().await;
@@ -49,6 +49,25 @@ impl ScannerIOCycle for ECStore {
    ) -> Result<ScannerCycleResult> {
        let child_token = ctx.child_token();

+        // Check the local pool metadata before listing buckets. A failed or
+        // canceled decommission remains suspended after its worker exits, so
+        // starting a scan in that state could build a snapshot that cannot be
+        // routed to the authoritative metadata object.
+        if self.scanner_data_usage_publication_blocked().await {
+            debug!(
+                target: "rustfs::scanner::io",
+                event = EVENT_SCANNER_SET_STATE,
+                component = LOG_COMPONENT_SCANNER,
+                subsystem = LOG_SUBSYSTEM_IO,
+                state = "cycle_data_usage_route_blocked",
+                "Scanner cycle deferred while data usage metadata remains hidden by data movement"
+            );
+            return Ok(ScannerCycleResult::new(
+                ScannerCycleStatus::Deferred(ScannerCycleDeferReason::DataMovement),
+                None,
+            ));
+        }
+
        let distributed = self.setup_is_dist_erasure().await;
        let activity_before = match scanner_activity_preflight(crate::scanner::probe_scanner_activity(self, distributed).await) {
            ScannerActivityPreflight::Ready(snapshot) => snapshot,
@@ -17,7 +17,9 @@ use super::io_disk::tier_stats_template;
 use super::*;
 use crate::scanner_budget::ScannerCycleBudgetConfig;
 use crate::scanner_folder::ScannerItem;
-use crate::storage_api::owner::{EcstoreRebalStatus, EcstoreRebalanceInfo, EcstoreRebalanceMeta, EcstoreRebalanceStats};
+use crate::storage_api::owner::{
+    EcstorePoolDecommissionInfo, EcstoreRebalStatus, EcstoreRebalanceInfo, EcstoreRebalanceMeta, EcstoreRebalanceStats,
+};
 use crate::storage_api::scan::{BucketOperations as _, DeleteBucketOptions, MakeBucketOptions, ObjectIO as _};
 use crate::{
    DiskOption, ECStore, Endpoint, EndpointServerPools, Endpoints, InstanceContext, PoolEndpoints, ScannerObjectOptions,
@@ -182,6 +184,39 @@ async fn scanner_cycle_is_deferred_while_rebalance_is_active() {
    assert!(receiver.recv().await.is_none(), "rebalance-deferred cycle must not publish usage");
 }

+#[tokio::test]
+#[serial]
+async fn scanner_cycle_is_deferred_while_terminal_decommission_is_blocked() {
+    let (_temp_dir, store) = setup_two_pool_scanner_store().await;
+    for decommission in [
+        EcstorePoolDecommissionInfo {
+            failed: true,
+            ..Default::default()
+        },
+        EcstorePoolDecommissionInfo {
+            canceled: true,
+            ..Default::default()
+        },
+    ] {
+        store.pool_meta.write().await.pools[0].decommission = Some(decommission);
+        assert!(store.scanner_data_usage_publication_blocked().await);
+
+        let ctx = CancellationToken::new();
+        let budget = ScannerCycleBudget::new(&ctx, ScannerCycleBudgetConfig::default());
+        let (updates, mut receiver) = mpsc::channel(1);
+        let result = tokio::time::timeout(
+            Duration::from_secs(30),
+            ScannerIOCycle::nsscanner_with_status(store.as_ref(), ctx, budget, updates, 1, 1, HealScanMode::Normal),
+        )
+        .await
+        .expect("terminal-decommission-deferred scanner cycle should finish")
+        .expect("terminal-decommission-deferred scanner cycle should succeed");
+
+        assert_eq!(result.status, ScannerCycleStatus::Deferred(ScannerCycleDeferReason::DataMovement));
+        assert!(receiver.recv().await.is_none(), "blocked cycle must not publish usage");
+    }
+}
+
 #[tokio::test]
 async fn data_usage_publish_fails_when_receiver_is_closed() {
    let (updates, receiver) = mpsc::channel(1);
@@ -236,6 +271,10 @@ async fn multi_pool_scanner_cycle_publishes_combined_usage() {
    assert_eq!(bucket_usage.size, 11);
    assert_eq!(usage.objects_total_count, 2);
    assert_eq!(usage.objects_total_size, 11);
+    assert!(
+        receiver.recv().await.is_none(),
+        "a scanner cycle must publish at most one terminal usage snapshot"
+    );
 }

 #[tokio::test]
@@ -47,6 +47,8 @@ pub(crate) use rustfs_ecstore::api::bucket::versioning_sys::BucketVersioningSys
 pub(crate) use rustfs_ecstore::api::cache::{
    ListPathRawOptions as EcstoreListPathRawOptions, list_path_raw as ecstore_list_path_raw,
 };
+#[cfg(test)]
+pub(crate) use rustfs_ecstore::api::capacity::PoolDecommissionInfo as EcstorePoolDecommissionInfo;
 pub(crate) use rustfs_ecstore::api::capacity::{
    is_reserved_or_invalid_bucket as ecstore_is_reserved_or_invalid_bucket, path2_bucket_object as ecstore_path2_bucket_object,
    path2_bucket_object_with_base_path as ecstore_path2_bucket_object_with_base_path,
@@ -127,9 +129,9 @@ pub(crate) mod owner {
    #[cfg(test)]
    pub(crate) use super::{
        EcstoreDiskOption, EcstoreDiskStore, EcstoreEndpoint, EcstoreEndpointServerPools, EcstoreEndpoints,
-        EcstoreInstanceContext, EcstorePoolEndpoints, EcstoreRebalStatus, EcstoreRebalanceInfo, EcstoreRebalanceMeta,
-        EcstoreRebalanceStats, ecstore_config_init, ecstore_init_bucket_metadata_sys, ecstore_init_local_disks_with_instance_ctx,
-        ecstore_new_disk,
+        EcstoreInstanceContext, EcstorePoolDecommissionInfo, EcstorePoolEndpoints, EcstoreRebalStatus, EcstoreRebalanceInfo,
+        EcstoreRebalanceMeta, EcstoreRebalanceStats, ecstore_config_init, ecstore_init_bucket_metadata_sys,
+        ecstore_init_local_disks_with_instance_ctx, ecstore_new_disk,
    };
 }

@@ -76,6 +76,7 @@ pub use bucket::{BucketInfo, BucketOperations, BucketOptions, DeleteBucketOption
 pub use capability::{CapabilitySnapshotError, CapabilityState, CapabilityStatus};
 pub use error::{StorageErrorCode, StorageResult};
 pub use multipart::{CompletePart, ListMultipartsInfo, ListPartsInfo, MultipartInfo, MultipartUploadResult, PartInfo};
+pub use object::DeleteAccounting;
 pub use object::ObjectLockDeleteOptions;
 pub use object::{DeletedObject, ObjectToDelete};
 pub use object::{ExpirationOptions, TransitionedObject};
@@ -218,6 +218,17 @@ pub struct DeletedObject {
    pub force_delete_generation: Option<i64>,
 }

+/// Accounting identity returned by the internal commit-time delete path.
+///
+/// This is carried separately from [`DeletedObject`] so adding quota details
+/// does not change the source shape of the public S3 delete result contract.
+#[derive(Debug, Default, Clone, PartialEq, Eq)]
+pub struct DeleteAccounting {
+    pub size: Option<u64>,
+    pub version_id: Option<Uuid>,
+    pub removed_current_object: bool,
+}
+
 impl DeletedObject {
    pub fn version_purge_status(&self) -> VersionPurgeStatusType {
        self.replication_state
@@ -341,6 +352,19 @@ pub trait ObjectOperations: Send + Sync + fmt::Debug {
        objects: Vec<Self::ObjectToDelete>,
        opts: Self::ObjectOptions,
    ) -> (Vec<Self::DeletedObject>, Vec<Option<Self::Error>>);
+    /// Delete objects and optionally return commit-time accounting identities.
+    /// The default preserves the ordinary delete contract for implementations
+    /// that do not expose storage-level accounting details.
+    async fn delete_objects_with_accounting(
+        &self,
+        bucket: &str,
+        objects: Vec<Self::ObjectToDelete>,
+        opts: Self::ObjectOptions,
+    ) -> (Vec<Self::DeletedObject>, Vec<Option<Self::Error>>, Vec<Option<DeleteAccounting>>) {
+        let object_count = objects.len();
+        let (deleted, errors) = self.delete_objects(bucket, objects, opts).await;
+        (deleted, errors, vec![None; object_count])
+    }
    async fn put_object_metadata(
        &self,
        bucket: &str,
@@ -622,7 +622,8 @@ mod test {
        let _resolver_lock = DNS_RESOLVER_TEST_LOCK.lock().unwrap();
        reset_dns_resolver_inner();

-        let err = resolve_domain("rustfs-resolver-provenance.invalid").unwrap_err();
+        // DNS labels are limited to 63 bytes, so the system resolver rejects this before lookup.
+        let err = resolve_domain("aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa.invalid").unwrap_err();

        assert_ne!(err.kind(), std::io::ErrorKind::Other, "system resolver error was wrapped: {err}");
    }
@@ -5,12 +5,10 @@
 > ```bash
 > cargo nextest list -p e2e_test --message-format json | jq -r '.["rust-suites"][]?.testcases | to_entries[] | select(.value.ignored == false) | .key | split("::")[0]' | sort | uniq -c
 > ```
-> Modules marked ✅ are in the PR smoke profile `e2e-smoke`
-> (`.config/nextest.toml`); admission criteria: `crates/e2e_test/README.md`.
-> 🌙 marks tests in the scheduled `e2e-repl-nightly` profile (backlog#1147
-> repl-1): `replication_extension_test` splits 20 fast tests into the PR smoke
-> lane and 28 slow / `_real_dual_node` / `_real_three_node` / `_real_single_node` tests into the
-> nightly lane (`.github/workflows/e2e-replication-nightly.yml`).
+> Modules marked ✅ are in the PR smoke profile `e2e-smoke`; 🌙 marks the
+> cluster, protocol, and replication subsets in the consolidated nightly
+> workflow. The `e2e-full` merge/main profile covers the remaining default
+> single-node tests. Committed test-ID digests are enforced before each run.
 > Note: counts exclude `#[ignore]`d tests (nextest lists them separately).
 > Managed-SSE (SSE-S3/SSE-KMS) replication contracts assert successful
 > re-encryption on the target (backlog#1783); SSE-C replication still pins a
@@ -19,19 +17,21 @@
 | module | tests | PR smoke |
 |---|---|---|
 | admin_auth_test | 4 | ✅ |
-| admin_iam_crud_test | 2 | ✅ |
+| admin_iam_crud_test | 3 | ✅ |
 | admin_pools_test | 1 | ✅ |
-| admin_timeout_regression_test | 1 |  |
+| admin_timeout_regression_test | 1 | 🌙 |
 | anonymous_access_test | 4 | ✅ |
 | api_rate_limit_test | 3 |  |
 | archive_download_integrity_test | 13 |  |
 | bucket_logging_test | 3 |  |
 | bucket_policy_check_test | 1 | ✅ |
+| bucket_stats_regression_test | 3 |  |
+| chaos | 2 |  |
 | checksum_upload_test | 7 |  |
-| cluster_concurrency_test | 2 |  |
-| cluster_multidrive_pool_test | 2 |  |
-| common | 12 |  |
-| compression_test | 1 |  |
+| cluster_concurrency_test | 2 | 🌙 |
+| cluster_multidrive_pool_test | 2 | 🌙 |
+| common | 14 |  |
+| compression_test | 6 | ✅ |
 | connection_cap_test | 2 |  |
 | console_smoke_test | 1 | ✅ |
 | content_encoding_test | 3 | ✅ |
@@ -45,48 +45,58 @@
 | delete_marker_migration_semantics_test | 2 | ✅ |
 | delete_object_no_content_length_test | 1 |  |
 | delete_objects_versioning_test | 2 | ✅ |
+| delete_regression_test | 5 |  |
+| distributed_startup_regression_test | 3 |  |
 | existing_object_tag_policy_test | 4 |  |
-| fake_s3_target | 4 | ✅ |
+| fake_s3_target | 6 | ✅ |
 | fault_proxy | 7 |  |
 | get_codec_streaming_compat_test | 1 |  |
+| get_stream_failure_observability_test | 1 |  |
+| group_delete_test | 1 |  |
 | head_object_consistency_test | 1 | ✅ |
 | head_object_range_test | 1 | ✅ |
-| heal_erasure_disk_rebuild_test | 3 |  |
-| inline_fast_path_cluster_test | 14 |  |
+| heal_erasure_disk_rebuild_test | 4 | 🌙 |
+| inline_fast_path_cluster_test | 16 |  |
 | internode_rpc_signature_e2e_test | 5 |  |
-| kms | 41 |  |
+| kms | 48 |  |
 | leading_slash_key_test | 2 | ✅ |
+| lifecycle_regression_test | 4 |  |
+| list_buckets_auth_test | 1 | ✅ |
 | list_buckets_double_slash_test | 3 | ✅ |
+| list_buckets_iam_filter_test | 1 | ✅ |
 | list_object_versions_metadata_extension_test | 1 |  |
 | list_object_versions_regression_test | 2 | ✅ |
 | list_objects_duplicates_test | 3 | ✅ |
 | list_objects_v2_metadata_extension_test | 1 |  |
 | list_objects_v2_pagination_test | 12 | ✅ |
+| listing_regression_test | 4 |  |
 | mc_mirror_small_bucket_test | 1 |  |
 | multipart_auth_test | 75 |  |
 | multipart_storage_class_test | 3 | ✅ |
-| namespace_lock_quorum_test | 2 |  |
+| namespace_lock_quorum_test | 2 | 🌙 |
 | negative_sigv4_test | 6 | ✅ |
+| notification_startup_regression_test | 2 |  |
 | notification_webhook_test | 3 | ✅ |
-| object_lambda_test | 16 |  |
-| object_lock | 33 |  |
+| object_lambda_test | 16 | 🌙 |
+| object_lock | 34 |  |
 | overwrite_cleanup_regression_test | 1 |  |
 | presigned_negative_test | 7 | ✅ |
-| protocols | 16 |  |
+| protocols | 16 | 🌙 |
 | quota_test | 14 |  |
-| reliability_disk_fault_test | 3 |  |
-| reliant | 24 | 18 ✅ |
-| replication_extension_test | 50 | 20 ✅ +30 🌙 |
+| reliability_disk_fault_test | 4 |  |
+| reliant | 25 | 19 ✅ |
+| replication_extension_test | 75 | 20 ✅ +55 🌙 |
 | security_boundary_test | 4 |  |
-| ssec_copy_test | 2 | ✅ |
 | server_startup_failfast_test | 1 |  |
 | snowball_auto_extract_test | 6 |  |
 | special_chars_test | 14 | ✅ |
-| stale_multipart_cleanup_cluster_test | 1 |  |
+| ssec_copy_test | 2 | ✅ |
+| stale_multipart_cleanup_cluster_test | 1 | 🌙 |
 | storage_class_capability_test | 4 | ✅ |
-| sts_query_compat_test | 3 | ✅ |
+| sts_query_compat_test | 6 | ✅ |
+| tier_transition_regression_test | 3 |  |
 | tls_gen | 3 |  |
 | tls_hot_reload_test | 1 | ✅ |
 | version_id_regression_test | 10 | ✅ |

-**Total listed: 530 tests across 70 modules · PR smoke subset: 148 tests / 33 modules** (31 full modules + 18 `reliant` tests + 20 of `replication_extension_test`) **· nightly `e2e-repl-nightly`: 30 tests** · updated 2026-08-09.
+**Total listed: 577 tests across 82 modules · PR smoke: 163 tests / 36 modules · merge/main full: 455 tests / 73 modules · nightly replication: 55 tests · nightly cluster faults: 28 tests / 7 modules · nightly protocols: 16 tests** · updated 2026-08-21.
@@ -48,11 +48,8 @@ fn materialize_case(path: String, prefix: Option<String>, flags: &[String]) -> (
        .fold((path, prefix), |(path, prefix), flag| apply_flag(path, prefix, flag))
 }

-fn has_dot_segments(path: &str) -> bool {
-    path.split(['/', '\\']).any(|segment| {
-        let trimmed = segment.trim();
-        trimmed == "." || trimmed == ".."
-    })
+fn has_parent_segments(path: &str) -> bool {
+    path.split(['/', '\\']).any(|segment| segment == "..")
 }

 fuzz_target!(|data: &[u8]| {
@@ -66,8 +63,8 @@ fuzz_target!(|data: &[u8]| {

    if let Ok(key) = normalize_extract_entry_key(&path, prefix.as_deref(), is_dir) {
        assert!(
-            !has_dot_segments(&key),
-            "accepted archive entry retained dot segments: path={:?} prefix={:?} key={:?}",
+            !has_parent_segments(&key),
+            "accepted archive entry retained parent segments: path={:?} prefix={:?} key={:?}",
            path,
            prefix,
            key
@@ -69,7 +69,11 @@ fuzz_target!(|data: &[u8]| {
        && let Some(object) = value.as_object()
    {
        let mut legacy_doc = Map::new();
-        if let Some(policy) = object.get("Policy").or_else(|| object.get("policy")) {
+        if let Some(policy) = object
+            .get("Policy")
+            .or_else(|| object.get("policy"))
+            .filter(|policy| serde_json::from_value::<Policy>((*policy).clone()).is_ok())
+        {
            legacy_doc.insert("version".to_string(), json!(1));
            legacy_doc.insert("policy".to_string(), policy.clone());
            legacy_doc.insert("create_date".to_string(), json!("2025-03-07T12:00:00Z"));
@@ -1,5 +1,6 @@
 3d602080f7ca4c32ba9e37ad1a32665c78560726b30aeee08fd9e95eb2f36194  accept-vectors.json
 d3c19946288717088145592e0e8d6f2fa684443ba2f73d4c7bc49c415d6dd051  certificate-profile.json
-060485263c51003274c056a0e04bec1b7d76157cf599ba79eebe040bc7cee71b  error-codes.json
+299a2ae34a8ca74bcf31deeb53a08f9eff279efa09d3f5358a0cf10866fe1a5d  error-codes.json
 43fe297ffb512b1b9f4af62f1832f3aa3905157893bfdc3dcc6d56f5a98aaef6  reject-vectors.json
-b946175b094f4a8d75091b652fbe3d4327c9c795f28e02c96e1ab90a429e418d  surface-separation.json
+0cf26a7332fa6e3f57390e081f2cceead3236f6ca57f7038e3b55c2582af1733  rotation-proof.json
+7c07100460fa23fca482466df9ca22f91c7f26b36087a7207a001f7d987f527e  surface-separation.json
@@ -2,7 +2,7 @@
  "protocolVersion": "v1",
  "fixtureSet": "auth",
  "fixture": "error-codes",
-  "description": "Frozen ErrorInfo reasons for agent authentication and negotiation. Clients branch on status and reason, never on message.",
+  "description": "Frozen ErrorInfo reasons for agent authentication, negotiation, and credential-rotation authorization. Clients branch on status and reason, never on message.",
  "domain": "rustfs.connect",
  "detailType": "type.googleapis.com/google.rpc.ErrorInfo",
  "disclosureRules": [
@@ -81,6 +81,24 @@
      "httpStatus": 401,
      "status": "UNAUTHENTICATED",
      "meaning": "A browser session cookie was presented to an authenticated agent operation. The agent surface never accepts it."
+    },
+    {
+      "reason": "ROTATION_CREDENTIAL_NOT_CURRENT",
+      "httpStatus": 409,
+      "status": "ABORTED",
+      "meaning": "The authenticated certificate is valid for ordinary agent operations but is not the device's current ACTIVE credential and therefore cannot authorize another rotation."
+    },
+    {
+      "reason": "ROTATION_REQUEST_CONFLICT",
+      "httpStatus": 409,
+      "status": "ABORTED",
+      "meaning": "The requestId already belongs to a rotation with different transcript inputs and cannot be reused."
+    },
+    {
+      "reason": "ROTATION_PROOF_INVALID",
+      "httpStatus": 401,
+      "status": "UNAUTHENTICATED",
+      "meaning": "The credential proof does not verify over the canonical rotation transcript under the presented certificate key."
    }
  ]
 }
@@ -0,0 +1,306 @@
+{
+  "protocolVersion": "v1",
+  "fixtureSet": "auth",
+  "fixture": "rotation-proof",
+  "description": "Frozen proof-of-possession transcript for rotating an online device credential. The current certificate key authorizes one new certificate request for one device and one idempotent request.",
+  "operation": {
+    "method": "POST",
+    "path": "/agent/clusterDevices/{device}:rotateCredential",
+    "operationId": "rotateClusterDeviceCredential",
+    "authenticatedSurface": "/agent/*",
+    "currentCredentialRequiredForNewRotation": true,
+    "outgoingOverlapCredentialMayAuthenticateOrdinaryOperations": true,
+    "outgoingOverlapCredentialMayRotate": false,
+    "outgoingOverlapCredentialMayReplayItsCompletedRotation": true
+  },
+  "replayPolicy": {
+    "recordState": "COMPLETED",
+    "bindingFields": [
+      "currentCertificateFingerprint",
+      "clusterDeviceName",
+      "requestId",
+      "certificateRequestSha256"
+    ],
+    "exactMatch": "returnStoredResult",
+    "mismatchReason": "ROTATION_REQUEST_CONFLICT",
+    "retentionLowerBound": "outgoingCredential.validUntil",
+    "sideEffects": "No issuer call, credential write, or overlap extension."
+  },
+  "completedReplayRecord": {
+    "state": "COMPLETED",
+    "currentCertificateFingerprint": "1bc816e2d285c52ee3a0aa06dc7aec17e788626043e795259a67f2436f970d09",
+    "clusterDeviceName": "organizations/0198f4b0-1a00-7c10-8d21-2e3f4a5b6c70/clusters/0198f4b0-2b00-7d20-9e31-3f4a5b6c7d81/clusterDevices/0198f4b0-3c00-7e30-8f41-4a5b6c7d8e92",
+    "requestId": "7c5d6e2a-8b91-4f03-a4d5-6e7f8091a2b3",
+    "certificateRequestSha256": "kVS-bXYxD6F22cZNy4Vnpgb5gFZbr4EGP5GvIz7uOaw",
+    "resultReference": "credential-rotation-result-01"
+  },
+  "transcript": {
+    "domain": "RUSTFS-CONNECT-CREDENTIAL-ROTATION-V1",
+    "domainTerminator": "0x0a",
+    "fieldSeparator": "0x3a",
+    "fieldTerminator": "0x0a",
+    "fieldCount": 4,
+    "encoding": "US-ASCII",
+    "normalisationPermitted": false,
+    "fieldOrder": [
+      "currentCertificateFingerprint",
+      "clusterDeviceName",
+      "requestId",
+      "certificateRequestSha256"
+    ],
+    "fields": [
+      {
+        "name": "currentCertificateFingerprint",
+        "position": 1,
+        "source": "the exact certificate accepted by trusted ingress and resolved by Connect",
+        "pattern": "^[0-9a-f]{64}$",
+        "binds": "the one currently active credential authorizing the rotation",
+        "absenceWouldAllow": "A proof captured from one current certificate to authorize another credential after the device rotated."
+      },
+      {
+        "name": "clusterDeviceName",
+        "position": 2,
+        "source": "the authenticated identity after it is matched to the target resource",
+        "pattern": "^organizations/[0-9a-f-]{36}/clusters/[0-9a-f-]{36}/clusterDevices/[0-9a-f-]{36}$",
+        "binds": "the organization, cluster, and device being rotated",
+        "absenceWouldAllow": "A proof produced by one device to be presented against another device resource."
+      },
+      {
+        "name": "requestId",
+        "position": 3,
+        "source": "the request body",
+        "pattern": "^[0-9a-f]{8}-[0-9a-f]{4}-4[0-9a-f]{3}-[89ab][0-9a-f]{3}-[0-9a-f]{12}$",
+        "binds": "the single idempotent rotation attempt",
+        "absenceWouldAllow": "A captured proof to be replayed as a fresh attempt instead of the same idempotent request."
+      },
+      {
+        "name": "certificateRequestSha256",
+        "position": 4,
+        "source": "recomputed over the exact PKCS#10 DER octets in certificateRequest",
+        "pattern": "^[A-Za-z0-9_-]{43}$",
+        "binds": "the next device public key and certificate request",
+        "absenceWouldAllow": "A valid proof to be combined with an attacker certificate request."
+      }
+    ]
+  },
+  "signature": {
+    "algorithmField": "proof.algorithm",
+    "algorithmEnumeration": ["ES256"],
+    "curve": "P-256",
+    "signatureEncoding": "fixed-width-r-s",
+    "signatureLengthBytes": 64,
+    "signatureTransferEncoding": "base64url-unpadded",
+    "signatureValuePattern": "^[A-Za-z0-9_-]{86}$",
+    "lowSRequired": true,
+    "verifyingKeySource": "the SubjectPublicKeyInfo of the exact presented certificate accepted by trusted ingress",
+    "newKeyPossession": "the PKCS#10 self-signature is verified separately under the public key inside certificateRequest",
+    "sharedEncodingContract": "protocol/agent/v1/registration-proof.md#the-signature"
+  },
+  "verificationOrder": [
+    {
+      "stage": "protocolVersion",
+      "rule": "protocolVersion must be the supported major version.",
+      "reason": "UNSUPPORTED_PROTOCOL"
+    },
+    {
+      "stage": "encoding",
+      "rule": "proof.algorithm and proof.value obey the shared ES256 fixed-width low-S contract.",
+      "reasons": ["UNSUPPORTED_ALGORITHM", "SIGNATURE_MALFORMED", "SIGNATURE_NOT_CANONICAL"]
+    },
+    {
+      "stage": "certificateRequest",
+      "rule": "certificateRequest is one self-signed PKCS#10 request whose public key is P-256.",
+      "reasons": ["CERTIFICATE_REQUEST_MALFORMED", "DEVICE_KEY_UNSUPPORTED"]
+    },
+    {
+      "stage": "proof",
+      "rule": "Connect rebuilds the transcript and verifies proof.value under the presented certificate public key.",
+      "reason": "ROTATION_PROOF_INVALID"
+    },
+    {
+      "stage": "replay",
+      "rule": "A completed record bound to the presented credential, device, requestId, and CSR digest returns its stored result without issuing again. A requestId bound to different transcript inputs is refused.",
+      "reason": "ROTATION_REQUEST_CONFLICT"
+    },
+    {
+      "stage": "credential",
+      "rule": "After replay lookup misses, the authenticated certificate must be the device current ACTIVE credential, not a credential in the outgoing overlap.",
+      "reason": "ROTATION_CREDENTIAL_NOT_CURRENT"
+    }
+  ],
+  "reasonSources": {
+    "auth": ["UNSUPPORTED_PROTOCOL", "ROTATION_CREDENTIAL_NOT_CURRENT", "ROTATION_REQUEST_CONFLICT", "ROTATION_PROOF_INVALID"],
+    "sharedFromRegistration": [
+      "UNSUPPORTED_ALGORITHM",
+      "SIGNATURE_MALFORMED",
+      "SIGNATURE_NOT_CANONICAL",
+      "CERTIFICATE_REQUEST_MALFORMED",
+      "DEVICE_KEY_UNSUPPORTED"
+    ]
+  },
+  "example": {
+    "inputs": {
+      "currentCertificateFingerprint": "1bc816e2d285c52ee3a0aa06dc7aec17e788626043e795259a67f2436f970d09",
+      "clusterDeviceName": "organizations/0198f4b0-1a00-7c10-8d21-2e3f4a5b6c70/clusters/0198f4b0-2b00-7d20-9e31-3f4a5b6c7d81/clusterDevices/0198f4b0-3c00-7e30-8f41-4a5b6c7d8e92",
+      "requestId": "7c5d6e2a-8b91-4f03-a4d5-6e7f8091a2b3",
+      "certificateRequestSha256": "kVS-bXYxD6F22cZNy4Vnpgb5gFZbr4EGP5GvIz7uOaw"
+    },
+    "artifacts": {
+      "currentPublicKeySpki": "MFkwEwYHKoZIzj0CAQYIKoZIzj0DAQcDQgAENEpjuFZTqp0Hxh/OWV3TGkluNjCo15dk+4CozuR6aT9Vaxhkb2M9nhaVGfk8+aSiSIiKFSCsYonKl8jh743Qow==",
+      "currentPublicKeyFingerprint": "28608e223c75ed89e72041f12afae4fc1cd3f1d3bcca38cbb62d1a41687610c2",
+      "certificateRequest": "MIIBHjCBxAIBADBiMRswGQYDVQQDDBJpZ25vcmVkLWJ5LWNvbm5lY3QxCzAJBgNVBAYTAkFVMRMwEQYDVQQIDApTb21lLVN0YXRlMSEwHwYDVQQKDBhJbnRlcm5ldCBXaWRnaXRzIFB0eSBMdGQwWTATBgcqhkjOPQIBBggqhkjOPQMBBwNCAAQWPosvIKr5V3CpTUdLfMQxDx31B3SglLKyqRg/oH3J+PhUnqf7pDWW1sTkP2aRIDjAwRn0DpLrz405CcvGHvYEoAAwCgYIKoZIzj0EAwIDSQAwRgIhAMJzXo/CK4E9BfjOxP35he9LLlqENhK7HTzZQTuIgLX2AiEAwOZHibk5HEijTWcJ/UT117nssfJesWZVOWwz/KTIpi8=",
+      "otherCertificateRequest": "MIIBHTCBxAIBADBiMRswGQYDVQQDDBJpZ25vcmVkLWJ5LWNvbm5lY3QxCzAJBgNVBAYTAkFVMRMwEQYDVQQIDApTb21lLVN0YXRlMSEwHwYDVQQKDBhJbnRlcm5ldCBXaWRnaXRzIFB0eSBMdGQwWTATBgcqhkjOPQIBBggqhkjOPQMBBwNCAATi0KiZGcbPJixDiQN/aIVky7xqaYhjmwAn1aozrC2eanaDENYQ3D5zw5qoZNIp8YUk/CETZrba3C5KYR5ocdlLoAAwCgYIKoZIzj0EAwIDSAAwRQIhAL0ZtvWyTSu18RF5J4ZVIuOGjJpJwSdP+87CVxNKJJduAiAahIruc3FLtO3RI7B8Ome8IsVDUpSAQilThjdOeDMsYg=="
+    },
+    "canonicalTranscript": "RUSTFS-CONNECT-CREDENTIAL-ROTATION-V1\n64:1bc816e2d285c52ee3a0aa06dc7aec17e788626043e795259a67f2436f970d09\n148:organizations/0198f4b0-1a00-7c10-8d21-2e3f4a5b6c70/clusters/0198f4b0-2b00-7d20-9e31-3f4a5b6c7d81/clusterDevices/0198f4b0-3c00-7e30-8f41-4a5b6c7d8e92\n36:7c5d6e2a-8b91-4f03-a4d5-6e7f8091a2b3\n43:kVS-bXYxD6F22cZNy4Vnpgb5gFZbr4EGP5GvIz7uOaw\n",
+    "canonicalTranscriptLengthBytes": 346,
+    "canonicalTranscriptSha256": "e5f0e9cd0d5d420bc7e51217512de76e1cf4dceb1d67804b6518c4fb5d9fe434"
+  },
+  "acceptVectors": [
+    {
+      "name": "current credential authorizes one new certificate request",
+      "requestId": "7c5d6e2a-8b91-4f03-a4d5-6e7f8091a2b3",
+      "proof": "vgz7Hm-xZ5JQ7B3oqUEKXB4tQxEzWZ988iPxAklaIZIsd4nSTUv5YriqES5yKGHvKOgWPyWjvrgI0j0T-fYEsQ",
+      "expected": {"accepted": true, "reason": null, "stage": "credential", "replayed": false}
+    },
+    {
+      "name": "the same certificate request under a new requestId needs its own proof",
+      "requestId": "82e1f3a4-9b05-4c67-8d90-1e2f3a4b5c6d",
+      "proof": "s-rkF9w41sqgdQNkOlhjcLYlrNHN_HL5xh4C3k_1a9ZQu7Uq0LoET_x7i4Rw3xQ75ec6B6tfqZ3XsHTS0FfN7Q",
+      "expected": {"accepted": true, "reason": null, "stage": "credential", "replayed": false}
+    },
+    {
+      "name": "outgoing credential replays its completed byte-equivalent rotation",
+      "requestId": "7c5d6e2a-8b91-4f03-a4d5-6e7f8091a2b3",
+      "proof": "vgz7Hm-xZ5JQ7B3oqUEKXB4tQxEzWZ988iPxAklaIZIsd4nSTUv5YriqES5yKGHvKOgWPyWjvrgI0j0T-fYEsQ",
+      "replayRecord": "completedExample",
+      "mutation": {"currentCredential": false},
+      "expected": {"accepted": true, "reason": null, "stage": "replay", "replayedResultReference": "credential-rotation-result-01"}
+    }
+  ],
+  "rejectVectors": [
+    {
+      "name": "outgoing overlap credential attempts another rotation",
+      "stage": "credential",
+      "requestId": "82e1f3a4-9b05-4c67-8d90-1e2f3a4b5c6d",
+      "proof": "s-rkF9w41sqgdQNkOlhjcLYlrNHN_HL5xh4C3k_1a9ZQu7Uq0LoET_x7i4Rw3xQ75ec6B6tfqZ3XsHTS0FfN7Q",
+      "mutation": {"currentCredential": false},
+      "expected": {"accepted": false, "reason": "ROTATION_CREDENTIAL_NOT_CURRENT"}
+    },
+    {
+      "name": "rotation declares protocol v2",
+      "stage": "protocolVersion",
+      "mutation": {"protocolVersion": "v2"},
+      "expected": {"accepted": false, "reason": "UNSUPPORTED_PROTOCOL"}
+    },
+    {
+      "name": "proof declares ES384",
+      "stage": "encoding",
+      "mutation": {"proofAlgorithm": "ES384"},
+      "expected": {"accepted": false, "reason": "UNSUPPORTED_ALGORITHM"}
+    },
+    {
+      "name": "DER encoded proof",
+      "stage": "encoding",
+      "mutation": {"proofValue": "MEUCIQC-DPseb7FnklDsHeipQQpcHi1DETNZn3zyI_ECSVohkgIgLHeJ0k1L-WK4qhEucihh7yjoFj8lo764CNI9E_n2BLE"},
+      "expected": {"accepted": false, "reason": "SIGNATURE_MALFORMED"}
+    },
+    {
+      "name": "padded base64url proof",
+      "stage": "encoding",
+      "mutation": {"proofValue": "vgz7Hm-xZ5JQ7B3oqUEKXB4tQxEzWZ988iPxAklaIZIsd4nSTUv5YriqES5yKGHvKOgWPyWjvrgI0j0T-fYEsQ=="},
+      "expected": {"accepted": false, "reason": "SIGNATURE_MALFORMED"}
+    },
+    {
+      "name": "protocol negotiation precedes proof encoding and stale credential state",
+      "stage": "protocolVersion",
+      "mutation": {"currentCredential": false, "protocolVersion": "v2", "proofValue": "not-a-signature"},
+      "expected": {"accepted": false, "reason": "UNSUPPORTED_PROTOCOL"}
+    },
+    {
+      "name": "declared algorithm precedes signature bytes",
+      "stage": "encoding",
+      "mutation": {"proofAlgorithm": "ES384", "proofValue": "not-a-signature"},
+      "expected": {"accepted": false, "reason": "UNSUPPORTED_ALGORITHM"}
+    },
+    {
+      "name": "malleated high-S proof",
+      "stage": "encoding",
+      "mutation": {"proofValue": "vgz7Hm-xZ5JQ7B3oqUEKXB4tQxEzWZ988iPxAklaIZLTiHYssrQGnkdV7tGN154Qk_7kboFz38zq542vAm0goA"},
+      "expected": {"accepted": false, "reason": "SIGNATURE_NOT_CANONICAL"}
+    },
+    {
+      "name": "proof over the rotation fields uses the registration domain",
+      "stage": "proof",
+      "mutation": {"proofValue": "as27GKXEKXtym-BU8NUl0BYhJkEUooZWPadrxwOWa40pAaN6p4VxKCCFggZl4ZQsl5CUtaMoMaxM_HzT1c1qUA"},
+      "expected": {"accepted": false, "reason": "ROTATION_PROOF_INVALID"}
+    },
+    {
+      "name": "proof verification precedes stale credential state",
+      "stage": "proof",
+      "mutation": {"currentCredential": false, "proofValue": "as27GKXEKXtym-BU8NUl0BYhJkEUooZWPadrxwOWa40pAaN6p4VxKCCFggZl4ZQsl5CUtaMoMaxM_HzT1c1qUA"},
+      "expected": {"accepted": false, "reason": "ROTATION_PROOF_INVALID"}
+    },
+    {
+      "name": "new certificate request key signs instead of the current credential key",
+      "stage": "proof",
+      "mutation": {"proofValue": "QEN8wMQQk5HjriDw2aeAxIfsRvrWgIjGR3KlpK81Y34kaTAacO-6_bOMAjQ0hMMIk-YYN6MDCw0Via_Ri0X0BA"},
+      "expected": {"accepted": false, "reason": "ROTATION_PROOF_INVALID"}
+    },
+    {
+      "name": "proof is moved to another current certificate",
+      "stage": "proof",
+      "mutation": {"currentCertificateFingerprint": "2bc816e2d285c52ee3a0aa06dc7aec17e788626043e795259a67f2436f970d09"},
+      "expected": {"accepted": false, "reason": "ROTATION_PROOF_INVALID"}
+    },
+    {
+      "name": "proof is moved to another device",
+      "stage": "proof",
+      "mutation": {"clusterDeviceName": "organizations/0198f4b0-1a00-7c10-8d21-2e3f4a5b6c70/clusters/0198f4b0-2b00-7d20-9e31-3f4a5b6c7d81/clusterDevices/0198f4b0-3c00-7e30-8f41-4a5b6c7d8e93"},
+      "expected": {"accepted": false, "reason": "ROTATION_PROOF_INVALID"}
+    },
+    {
+      "name": "proof is replayed under another requestId",
+      "stage": "proof",
+      "mutation": {"requestId": "82e1f3a4-9b05-4c67-8d90-1e2f3a4b5c6d"},
+      "expected": {"accepted": false, "reason": "ROTATION_PROOF_INVALID"}
+    },
+    {
+      "name": "proof is combined with another certificate request",
+      "stage": "proof",
+      "mutation": {"certificateRequest": "otherCertificateRequest"},
+      "expected": {"accepted": false, "reason": "ROTATION_PROOF_INVALID"}
+    },
+    {
+      "name": "completed replay belongs to another certificate fingerprint",
+      "stage": "replay",
+      "replayRecord": "completedExample",
+      "replayRecordMutation": {"currentCertificateFingerprint": "2bc816e2d285c52ee3a0aa06dc7aec17e788626043e795259a67f2436f970d09"},
+      "mutation": {"currentCredential": false},
+      "expected": {"accepted": false, "reason": "ROTATION_REQUEST_CONFLICT"}
+    },
+    {
+      "name": "completed replay belongs to another device",
+      "stage": "replay",
+      "replayRecord": "completedExample",
+      "replayRecordMutation": {"clusterDeviceName": "organizations/0198f4b0-1a00-7c10-8d21-2e3f4a5b6c70/clusters/0198f4b0-2b00-7d20-9e31-3f4a5b6c7d81/clusterDevices/0198f4b0-3c00-7e30-8f41-4a5b6c7d8e93"},
+      "mutation": {"currentCredential": false},
+      "expected": {"accepted": false, "reason": "ROTATION_REQUEST_CONFLICT"}
+    },
+    {
+      "name": "completed replay belongs to another requestId",
+      "stage": "replay",
+      "replayRecord": "completedExample",
+      "replayRecordMutation": {"requestId": "82e1f3a4-9b05-4c67-8d90-1e2f3a4b5c6d"},
+      "mutation": {"currentCredential": false},
+      "expected": {"accepted": false, "reason": "ROTATION_REQUEST_CONFLICT"}
+    },
+    {
+      "name": "completed replay belongs to another CSR digest",
+      "stage": "replay",
+      "replayRecord": "completedExample",
+      "replayRecordMutation": {"certificateRequestSha256": "PXxfNGWtIbrqVocag6lOLDDR41AjpHXJ7UcAWkbTSAs"},
+      "mutation": {"currentCredential": false},
+      "expected": {"accepted": false, "reason": "ROTATION_REQUEST_CONFLICT"}
+    }
+  ]
+}
@@ -49,7 +49,7 @@
      "schemeType": "mutualTLS",
      "forbiddenSecuritySchemes": ["sessionCookie"],
      "defaultSecurity": ["agentMutualTls"],
-      "publicOperations": ["getProtocolStatus"]
+      "publicOperations": ["getProtocolStatus", "exchangeRegistrationToken"]
    },
    {
      "document": "openapi/control.json",
@@ -11,7 +11,7 @@
    {
      "name": "auth",
      "status": "populated",
-      "purpose": "Client certificate profile, RFC 9440 header profile, authentication accept and reject vectors, surface separation, and the frozen error reason registry."
+      "purpose": "Client certificate profile, RFC 9440 header profile, authentication and credential-rotation proof vectors, surface separation, and the frozen error reason registry."
    },
    {
      "name": "version",
@@ -288,6 +288,7 @@ zstd.workspace = true
 # Cryptography and Security
 rustls = { workspace = true, default-features = false, features = ["aws-lc-rs", "logging", "tls12", "prefer-post-quantum", "std"] }
 rustls-pki-types = { workspace = true }
+x509-parser = { workspace = true }
 subtle = { workspace = true }
 jiff = { workspace = true, features = ["serde"] }
 time = { workspace = true, features = ["parsing", "formatting", "serde", "macros"] }
@@ -1021,7 +1021,7 @@ fn build_list_objects_v2_metadata_output(
                object: Object {
                    key: Some(encode_list_objects_v2_value(&object.name, encoding_type)),
                    last_modified: object.mod_time.map(Timestamp::from),
-                    size: Some(object.get_actual_size().unwrap_or_default()),
+                    size: Some(object.get_actual_size_or_physical()),
                    e_tag: object.etag.clone().map(|etag| to_s3s_etag(&etag)),
                    storage_class: Some(ObjectStorageClass::from(
                        object
@@ -3969,6 +3969,63 @@ fn delete_creates_delete_marker(opts: &ObjectOptions) -> bool {
    opts.version_id.is_none() && opts.versioned && !opts.version_suspended
 }

+fn delete_removes_current_object(opts: &ObjectOptions) -> bool {
+    delete_request_targets_current(
+        opts.version_id
+            .as_deref()
+            .and_then(|version_id| Uuid::parse_str(version_id).ok()),
+    )
+}
+
+fn delete_request_targets_current(version_id: Option<Uuid>) -> bool {
+    version_id.is_none() || version_id.is_some_and(|version_id| version_id.is_nil())
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum DeleteMemoryUpdate {
+    DeleteMarker,
+    Object { size: u64, removed_current_object: bool },
+}
+
+fn delete_memory_update(
+    creates_delete_marker: bool,
+    committed_delete_marker: bool,
+    requested_current: bool,
+    accounting_size: Option<u64>,
+    removed_current_object: bool,
+) -> Option<DeleteMemoryUpdate> {
+    if creates_delete_marker || (committed_delete_marker && requested_current) {
+        return Some(DeleteMemoryUpdate::DeleteMarker);
+    }
+
+    (!committed_delete_marker)
+        .then_some(accounting_size)
+        .flatten()
+        .map(|size| DeleteMemoryUpdate::Object {
+            size,
+            removed_current_object,
+        })
+}
+
+async fn apply_delete_memory_update(bucket: &str, update: Option<DeleteMemoryUpdate>) {
+    match update {
+        Some(DeleteMemoryUpdate::DeleteMarker) => record_bucket_delete_marker_memory(bucket).await,
+        Some(DeleteMemoryUpdate::Object {
+            size,
+            removed_current_object,
+        }) => record_bucket_object_delete_memory(bucket, size, removed_current_object).await,
+        None => {}
+    }
+}
+
+/// `DeleteObjects` is idempotent. A raw filesystem `NotFound` can cross the
+/// distributed delete path instead of its usual typed missing-object error.
+fn is_delete_objects_not_found(error: &EcstoreError) -> bool {
+    is_err_object_not_found(error)
+        || is_err_version_not_found(error)
+        || matches!(error, StorageError::Io(source) if source.kind() == std::io::ErrorKind::NotFound)
+}
+
 /// Bounded concurrency for the per-object pre-delete stat fanout in
 /// `execute_delete_objects` (backlog#929 / HP-8). Keeps the metadata reads for
 /// a 1000-key batch from serializing while capping the disk fanout pressure.
@@ -4030,6 +4087,27 @@ fn delete_response_version_id(version_id: Option<Uuid>, synthetic_version_id: bo
    }
 }

+fn reduce_delete_objects_result<'a>(
+    object: &ObjectToDelete,
+    deleted: &'a StorageDeletedObject,
+    error: Option<&EcstoreError>,
+    synthetic_version_id: bool,
+) -> Result<&'a StorageDeletedObject, s3s::dto::Error> {
+    match error {
+        None => Ok(deleted),
+        Some(error) if is_delete_objects_not_found(error) => Ok(deleted),
+        Some(error) => {
+            let api_error = ApiError::from(error.clone());
+            Err(s3s::dto::Error {
+                code: Some(api_error.code.as_str().to_string()),
+                key: Some(object.object_name.clone()),
+                message: Some(api_error.message),
+                version_id: delete_response_version_id(object.version_id, synthetic_version_id),
+            })
+        }
+    }
+}
+
 fn resolve_put_object_extract_options(headers: &HeaderMap) -> S3Result<PutObjectExtractOptions> {
    let prefix = snowball_meta_value(headers, SNOWBALL_PREFIX_HEADER_KEYS, SNOWBALL_PREFIX_SUFFIX_LOWER)
        .map(|value| normalize_snowball_prefix(&value))
@@ -8380,8 +8458,6 @@ impl DefaultObjectUsecase {
            object: ObjectToDelete,
            versioned: bool,
            version_suspended: bool,
-            size: i64,
-            existing: Option<ObjectInfo>,
        }

        // Phase 2 (bounded concurrency, backlog#929 / HP-8): collect the
@@ -8399,32 +8475,23 @@ impl DefaultObjectUsecase {
                    skip_stat,
                } = prepared;
                let synthetic_version_id = object.version_id.is_none() && is_dir_object(&object.object_name);
-                let (goi, source_missing) = if skip_stat {
-                    (ObjectInfo::default(), false)
-                } else {
+                if !skip_stat {
                    match store_ref.get_object_info(bucket_ref, &object.object_name, &opts).await {
-                        Ok(res) => (res, false),
-                        Err(err) if is_err_object_not_found(&err) || is_err_version_not_found(&err) => {
-                            (ObjectInfo::default(), true)
-                        }
+                        Ok(_) => {}
+                        Err(err) if is_err_object_not_found(&err) || is_err_version_not_found(&err) => {}
                        Err(err) => return Err(ApiError::from(err)),
                    }
-                };
-
-                let size = goi.size;
+                }

                if synthetic_version_id {
                    object.version_id = Some(Uuid::nil());
                }

-                let existing = (!skip_stat && !source_missing).then_some(goi);
                Ok::<_, ApiError>(AdmittedDelete {
                    idx,
                    object,
                    versioned: opts.versioned,
                    version_suspended: opts.version_suspended,
-                    size,
-                    existing,
                })
            }))
            .buffered(DELETE_OBJECTS_PRE_STAT_CONCURRENCY)
@@ -8435,15 +8502,11 @@ impl DefaultObjectUsecase {
        // per-key success/failure reporting is unchanged.
        let mut object_to_delete = Vec::new();
        let mut object_to_delete_idx = Vec::new();
-        let mut object_sizes = Vec::new();
-        let mut existing_object_infos = Vec::new();
        let mut object_versioning = Vec::new();
        for admitted in admitted_deletes {
-            object_sizes.push(admitted.size);
            object_to_delete_idx.push(admitted.idx);
            object_versioning.push((admitted.versioned, admitted.version_suspended));
            object_to_delete.push(admitted.object);
-            existing_object_infos.push(admitted.existing);
        }
        let cache_adapter = self.object_data_cache();
        let cache_keys_before_delete = object_to_delete
@@ -8460,8 +8523,8 @@ impl DefaultObjectUsecase {
            ..Default::default()
        };
        apply_bucket_generation_guard(&req, &bucket, &mut storage_delete_opts)?;
-        let (dobjs, errs) = store
-            .delete_objects_with_tier_delete_journal(&bucket, object_to_delete.clone(), storage_delete_opts)
+        let (dobjs, errs, accounting) = store
+            .delete_objects_with_tier_delete_journal_and_accounting(&bucket, object_to_delete.clone(), storage_delete_opts)
            .await;

        let _manager = get_concurrency_manager();
@@ -8476,39 +8539,30 @@ impl DefaultObjectUsecase {
        for (i, err) in errs.iter().enumerate() {
            let didx = object_to_delete_idx[i];

-            if err.is_none()
-                || err
-                    .clone()
-                    .is_some_and(|v| is_err_object_not_found(&v) || is_err_version_not_found(&v))
-            {
-                delete_results[didx].delete_object = Some(dobjs[i].clone());
-                let (versioned, version_suspended) = object_versioning[i];
-                let creates_delete_marker = object_to_delete[i].version_id.is_none() && versioned && !version_suspended;
-                if creates_delete_marker {
-                    record_bucket_delete_marker_memory(&bucket).await;
-                } else {
-                    let size = object_sizes[i].max(0) as u64;
-                    record_bucket_object_delete_memory(
-                        &bucket,
-                        size,
-                        existing_object_infos[i].is_some() && object_to_delete[i].version_id.is_none(),
-                    )
-                    .await;
+            match reduce_delete_objects_result(
+                &object_to_delete[i],
+                &dobjs[i],
+                err.as_ref(),
+                delete_results[didx].synthetic_version_id,
+            ) {
+                Ok(deleted_object) => {
+                    delete_results[didx].delete_object = Some(deleted_object.clone());
+                    let (versioned, version_suspended) = object_versioning[i];
+                    let creates_delete_marker = object_to_delete[i].version_id.is_none() && versioned && !version_suspended;
+                    let committed_delete_marker = dobjs[i].delete_marker;
+                    let delete_accounting = accounting.get(i).and_then(Option::as_ref);
+                    let update = delete_memory_update(
+                        creates_delete_marker,
+                        committed_delete_marker,
+                        delete_request_targets_current(object_to_delete[i].version_id),
+                        delete_accounting.and_then(|value| value.size),
+                        delete_accounting.is_some_and(|value| value.removed_current_object),
+                    );
+                    apply_delete_memory_update(&bucket, update).await;
+                }
+                Err(error) => {
+                    delete_results[didx].error = Some(error);
                }
-                continue;
-            }
-
-            if let Some(err) = err.clone() {
-                let api_error = ApiError::from(err);
-                delete_results[didx].error = Some(s3s::dto::Error {
-                    code: Some(api_error.code.as_str().to_string()),
-                    key: Some(object_to_delete[i].object_name.clone()),
-                    message: Some(api_error.message),
-                    version_id: delete_response_version_id(
-                        object_to_delete[i].version_id,
-                        delete_results[didx].synthetic_version_id,
-                    ),
-                });
            }
        }

@@ -8782,12 +8836,24 @@ impl DefaultObjectUsecase {
            let _ = invalidate_object_data_cache_after_delete_success(&cache_adapter, &bucket, &key).await;
        }

-        // Fast in-memory update for immediate quota and admin usage consistency
-        if delete_creates_delete_marker(&opts) {
-            record_bucket_delete_marker_memory(&bucket).await;
+        // Fast in-memory update for immediate quota and admin usage consistency.
+        // Prefix/force deletes and synthetic directory entries do not carry one
+        // committed object identity; leave their cache delta to reconciliation.
+        let update = if force_delete || obj_info.name.is_empty() || synthetic_version_id {
+            None
        } else {
-            record_bucket_object_delete_memory(&bucket, obj_info.size.max(0) as u64, opts.version_id.is_none()).await;
-        }
+            // The storage commit returns this object's metadata while its
+            // generation lock is held. Never fall back to a pre-delete stat:
+            // an overwrite can commit between that stat and this delete.
+            delete_memory_update(
+                delete_creates_delete_marker(&opts),
+                obj_info.delete_marker,
+                opts.version_id.is_none(),
+                quota_object_size(&obj_info).ok(),
+                delete_removes_current_object(&opts),
+            )
+        };
+        apply_delete_memory_update(&bucket, update).await;

        if obj_info.name.is_empty() {
            if let Some((operation_id, target_arns, generation)) = force_delete_intent {
@@ -17692,6 +17758,35 @@ mod tests {
        assert_eq!(internal_version_id, None);
    }

+    #[test]
+    fn delete_objects_treats_raw_io_not_found_as_idempotent() {
+        assert!(is_delete_objects_not_found(&StorageError::FileNotFound));
+        assert!(is_delete_objects_not_found(&StorageError::Io(std::io::Error::from(
+            std::io::ErrorKind::NotFound,
+        ))));
+        assert!(!is_delete_objects_not_found(&StorageError::Io(std::io::Error::from(
+            std::io::ErrorKind::PermissionDenied,
+        ))));
+        assert!(!is_delete_objects_not_found(&StorageError::DiskNotFound));
+    }
+
+    #[test]
+    fn delete_objects_result_reducer_reports_raw_not_found_as_deleted() {
+        let object = ObjectToDelete {
+            object_name: "missing-key".to_string(),
+            ..Default::default()
+        };
+        let deleted = StorageDeletedObject {
+            object_name: object.object_name.clone(),
+            ..Default::default()
+        };
+        let error = StorageError::Io(std::io::Error::from(std::io::ErrorKind::NotFound));
+
+        let deleted = reduce_delete_objects_result(&object, &deleted, Some(&error), false)
+            .expect("raw not-found must produce a deleted result");
+        assert_eq!(deleted.object_name, "missing-key");
+    }
+
    #[test]
    fn recursive_force_delete_requires_administrative_or_replica_context() {
        let mut headers = HeaderMap::new();
@@ -17811,6 +17906,158 @@ mod tests {
        assert!(!can_skip_delete_objects_pre_stat(false, &delete_marker_creating_opts(), false));
    }

+    #[test]
+    fn delete_accounting_recognizes_explicit_null_as_current_object() {
+        let opts = ObjectOptions {
+            version_id: Some(Uuid::nil().to_string()),
+            version_suspended: true,
+            ..Default::default()
+        };
+        assert!(delete_removes_current_object(&opts));
+        assert!(delete_request_targets_current(Some(Uuid::nil())));
+        assert!(!delete_request_targets_current(Some(Uuid::new_v4())));
+        assert!(!delete_removes_current_object(&ObjectOptions {
+            version_id: Some(Uuid::new_v4().to_string()),
+            ..Default::default()
+        }));
+    }
+
+    #[test]
+    fn compressed_object_delete_restores_usage_baseline() {
+        let mut metadata = HashMap::new();
+        insert_str(&mut metadata, SUFFIX_COMPRESSION, "klauspost/compress/s2".to_string());
+        let object = ObjectInfo {
+            size: 400,
+            actual_size: 1000,
+            user_defined: Arc::new(metadata),
+            ..Default::default()
+        };
+        let accounting_size = quota_object_size(&object).expect("logical compressed size should be canonical");
+
+        assert_eq!(
+            delete_memory_update(false, false, true, Some(accounting_size), true),
+            Some(DeleteMemoryUpdate::Object {
+                size: 1000,
+                removed_current_object: true,
+            })
+        );
+    }
+
+    #[test]
+    fn invalid_accounting_metadata_is_reconciled_without_overflow() {
+        assert_eq!(delete_memory_update(false, false, true, None, true), None);
+        assert_eq!(
+            delete_memory_update(false, true, true, None, true),
+            Some(DeleteMemoryUpdate::DeleteMarker)
+        );
+    }
+
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn compressed_delete_requests_restore_usage_baseline() {
+        use crate::app::storage_api::test::contract::bucket::{BucketOperations as _, DeleteBucketOptions, MakeBucketOptions};
+
+        let store = crate::app::gating_test_env::shared_gating_ecstore().await;
+        if current_app_context().is_none() {
+            crate::app::runtime_sources::install_test_app_context(Arc::clone(&store)).await;
+        }
+        let bucket = format!("compressed-delete-request-{}", Uuid::new_v4().simple());
+        store
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("create compressed delete request bucket");
+
+        // Seed the process-local usage with the canonical logical bytes. The
+        // direct storage PUT below intentionally does not apply an app-layer
+        // usage delta; the two real DELETE requests must remove exactly this
+        // amount through their request-layer wiring.
+        crate::app::storage_api::test::data_usage::seed_bucket_usage_memory_for_test(&bucket, 2_000).await;
+
+        for object in ["single", "batch"] {
+            let mut metadata = HashMap::new();
+            insert_str(&mut metadata, SUFFIX_COMPRESSION, "klauspost/compress/s2".to_string());
+            insert_str(&mut metadata, SUFFIX_ACTUAL_SIZE, "1000".to_string());
+            let reader = HashReader::from_stream(std::io::Cursor::new(vec![0x5a; 400]), 400, 1000, None, None, false)
+                .expect("compressed fixture reader should be valid");
+            let mut reader = PutObjReader::new(reader);
+            store
+                .put_object(
+                    &bucket,
+                    object,
+                    &mut reader,
+                    &ObjectOptions {
+                        user_defined: metadata,
+                        ..Default::default()
+                    },
+                )
+                .await
+                .expect("compressed fixture object should be written");
+        }
+
+        let mut single_req = build_request(
+            DeleteObjectInput::builder()
+                .bucket(bucket.clone())
+                .key("single".to_string())
+                .build()
+                .expect("single delete input should build"),
+            Method::DELETE,
+        );
+        single_req.extensions.insert(crate::storage::access::ReqInfo {
+            cred: Some(rustfs_credentials::Credentials::default()),
+            is_owner: true,
+            ..Default::default()
+        });
+        DefaultObjectUsecase::from_global()
+            .execute_delete_object(single_req)
+            .await
+            .expect("single compressed delete should succeed");
+        assert_eq!(
+            crate::app::storage_api::test::data_usage::get_bucket_usage_memory(&bucket).await,
+            Some(1_000),
+            "single delete must subtract the logical accounting size"
+        );
+
+        let mut batch_req = build_request(
+            DeleteObjectsInput::builder()
+                .bucket(bucket.clone())
+                .delete(Delete {
+                    objects: vec![ObjectIdentifier {
+                        key: "batch".to_string(),
+                        ..Default::default()
+                    }],
+                    quiet: None,
+                })
+                .build()
+                .expect("batch delete input should build"),
+            Method::POST,
+        );
+        batch_req.extensions.insert(crate::storage::access::ReqInfo {
+            cred: Some(rustfs_credentials::Credentials::default()),
+            is_owner: true,
+            ..Default::default()
+        });
+        DefaultObjectUsecase::from_global()
+            .execute_delete_objects(batch_req)
+            .await
+            .expect("batch compressed delete should succeed");
+        assert_eq!(
+            crate::app::storage_api::test::data_usage::get_bucket_usage_memory(&bucket).await,
+            Some(0),
+            "batch delete must subtract the committed logical accounting size"
+        );
+
+        store
+            .delete_bucket(
+                &bucket,
+                &DeleteBucketOptions {
+                    force: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("clean up compressed delete request bucket");
+    }
+
    #[tokio::test]
    async fn execute_get_object_attributes_returns_internal_error_when_store_uninitialized() {
        let input = GetObjectAttributesInput::builder()
@@ -72,6 +72,11 @@ pub(crate) mod data_usage {
        compute_bucket_usage, live_bucket_usage_computations, seed_bucket_usage_memory_for_test, store_data_usage_in_backend,
    };

+    #[cfg(test)]
+    pub(crate) async fn get_bucket_usage_memory(bucket: &str) -> Option<u64> {
+        crate::storage::storage_api::ecstore_data_usage::get_bucket_usage_memory(bucket).await
+    }
+
    pub(crate) async fn record_bucket_object_delete_memory(bucket: &str, deleted_size: u64, removed_current_object: bool) {
        crate::storage::storage_api::ecstore_data_usage::record_bucket_object_delete_memory(
            bucket,
@@ -1233,7 +1238,10 @@ pub(crate) mod test {

    pub(crate) use super::access::ReqInfo;
    pub(crate) use super::options::VERSIONING_CONFIG_LOOKUPS;
-    pub(crate) use super::{bucket, data_usage, ecfs, object_utils, runtime};
+    pub(crate) use super::{bucket, ecfs, object_utils, runtime};
+    pub(crate) mod data_usage {
+        pub(crate) use super::super::data_usage::*;
+    }
    pub(crate) use crate::storage::storage_api::test_consumer::{get_global_bucket_metadata_sys, set_bucket_metadata};
    pub(crate) use crate::storage::storage_api::{
        ECStore, Endpoint, Endpoints, PoolEndpoints, StorageObjectInfo, StorageObjectOptions, StoragePutObjReader,
@@ -17,6 +17,220 @@
 //! This binary shares RustFS's existing subcommand dispatcher and provides the
 //! documented entry point for offline tooling such as `inspect bucket-meta`.

-fn main() {
+use std::fs;
+use std::io::{Read as _, Write as _};
+use std::path::{Path, PathBuf};
+use std::process::ExitCode;
+use std::time::{SystemTime, UNIX_EPOCH};
+
+use rustfs::connect::offline::{OfflineEnrollment, OfflineKeyStore};
+
+/// Owner read/write only. The response names the key being enrolled and the
+/// challenge it answers; neither belongs to anyone else on the machine.
+#[cfg(unix)]
+const RESPONSE_MODE: u32 = 0o600;
+
+const USAGE: &str = "\
+Usage: rustfs-cli connect offline enroll --challenge <path|-> --output <path> [--key-dir <path>]
+
+Answers a Connect offline enrolment challenge without a network. Reads the
+challenge from a file or from stdin when the path is `-`, verifies it against the
+enrolment root compiled into this binary, mints the key being enrolled on first
+use, and writes the signed response.
+
+No secret is ever accepted on the command line.
+";
+
+fn main() -> ExitCode {
+    let arguments: Vec<String> = std::env::args().skip(1).collect();
+
+    // Offline enrolment is handled before the server dispatcher is reached, and
+    // the reason is the surface's whole point: `run_process` builds a Tokio
+    // runtime and enters the server's async main. An air-gapped enrolment must
+    // not start a runtime, a task, or anything that could open a socket, so the
+    // two paths cannot share an entry.
+    if matches!(
+        arguments.first().map(String::as_str),
+        Some("connect") if matches!(arguments.get(1).map(String::as_str), Some("offline"))
+    ) {
+        return match run_offline(&arguments[2..]) {
+            Ok(()) => ExitCode::SUCCESS,
+            Err(message) => {
+                eprintln!("rustfs-cli: {message}");
+                ExitCode::FAILURE
+            }
+        };
+    }
+
    rustfs::startup_entrypoint::run_process();
+
+    ExitCode::SUCCESS
+}
+
+fn run_offline(arguments: &[String]) -> Result<(), String> {
+    match arguments.first().map(String::as_str) {
+        Some("enroll") => enroll(&arguments[1..]),
+        Some(other) => Err(format!("unknown offline subcommand `{other}`\n\n{USAGE}")),
+        None => Err(format!("missing offline subcommand\n\n{USAGE}")),
+    }
+}
+
+fn enroll(arguments: &[String]) -> Result<(), String> {
+    let mut challenge_path: Option<String> = None;
+    let mut output_path: Option<String> = None;
+    let mut key_directory: Option<String> = None;
+
+    let mut index = 0;
+    while index < arguments.len() {
+        let flag = arguments[index].as_str();
+        let take_value = |name: &str| -> Result<String, String> {
+            arguments
+                .get(index + 1)
+                .cloned()
+                .ok_or_else(|| format!("`{name}` needs a value\n\n{USAGE}"))
+        };
+
+        match flag {
+            "--challenge" => challenge_path = Some(take_value("--challenge")?),
+            "--output" => output_path = Some(take_value("--output")?),
+            "--key-dir" => key_directory = Some(take_value("--key-dir")?),
+            "-h" | "--help" => {
+                println!("{USAGE}");
+                return Ok(());
+            }
+            other => return Err(format!("unknown option `{other}`\n\n{USAGE}")),
+        }
+
+        index += 2;
+    }
+
+    let challenge_path = challenge_path.ok_or_else(|| format!("`--challenge` is required\n\n{USAGE}"))?;
+    let output_path = output_path.ok_or_else(|| format!("`--output` is required\n\n{USAGE}"))?;
+    let key_directory = key_directory.unwrap_or_else(|| ".".to_string());
+
+    let challenge = read_challenge(&challenge_path)?;
+
+    let now = SystemTime::now()
+        .duration_since(UNIX_EPOCH)
+        .map_err(|_| "the system clock is before the Unix epoch".to_string())?
+        .as_secs() as i64;
+
+    let verified = OfflineEnrollment::verify_challenge(&challenge, now).map_err(|error| error.to_string())?;
+
+    // First use mints the key; a retry answers with the one already enrolled,
+    // because the operator may already be carrying a response naming it.
+    let key = OfflineKeyStore::new(&key_directory)
+        .load_or_create()
+        .map_err(|error| error.to_string())?;
+
+    let mut device_nonce = [0u8; 32];
+    getrandom(&mut device_nonce)?;
+
+    let response = OfflineEnrollment::build_response(&verified, &key, &device_nonce, now).map_err(|error| error.to_string())?;
+
+    write_response(Path::new(&output_path), &response)?;
+
+    Ok(())
+}
+
+/// Reads the challenge from a file, or from stdin when the path is `-`.
+///
+/// A challenge is not a secret — it is signed, public, and carried in by hand —
+/// so accepting a path is safe. The response's key never arrives this way.
+fn read_challenge(path: &str) -> Result<Vec<u8>, String> {
+    if path == "-" {
+        let mut buffer = Vec::new();
+        std::io::stdin()
+            .read_to_end(&mut buffer)
+            .map_err(|error| format!("cannot read the challenge from stdin: {error}"))?;
+        return Ok(buffer);
+    }
+
+    fs::read(path).map_err(|error| format!("cannot read the challenge at {path}: {error}"))
+}
+
+/// Writes the response durably and atomically at mode 0600.
+///
+/// Not the no-clobber publish `IdentityStore` performs for a key: an operator
+/// who reruns an enrolment expects the response file to be replaced, whereas a
+/// second key would strand the first. Same durability, deliberately different
+/// publication rule.
+fn write_response(path: &Path, response: &[u8]) -> Result<(), String> {
+    let parent = path.parent().filter(|parent| !parent.as_os_str().is_empty());
+    let temporary: PathBuf = match parent {
+        Some(parent) => parent.join(format!(".{}.tmp", file_name(path))),
+        None => PathBuf::from(format!(".{}.tmp", file_name(path))),
+    };
+
+    let mut options = fs::OpenOptions::new();
+    options.write(true).create(true).truncate(true);
+    #[cfg(unix)]
+    {
+        use std::os::unix::fs::OpenOptionsExt as _;
+        options.mode(RESPONSE_MODE);
+    }
+
+    let write = (|| -> std::io::Result<()> {
+        let mut file = options.open(&temporary)?;
+        file.write_all(response)?;
+
+        // The umask can only narrow the creation mode, so set the exact mode
+        // before the bytes become durable.
+        #[cfg(unix)]
+        {
+            use std::os::unix::fs::PermissionsExt as _;
+            file.set_permissions(fs::Permissions::from_mode(RESPONSE_MODE))?;
+        }
+
+        file.sync_all()
+    })();
+
+    if let Err(error) = write {
+        let _ = fs::remove_file(&temporary);
+        return Err(format!("cannot write the response to {}: {error}", path.display()));
+    }
+
+    fs::rename(&temporary, path).map_err(|error| {
+        let _ = fs::remove_file(&temporary);
+        format!("cannot publish the response at {}: {error}", path.display())
+    })?;
+
+    if let Some(parent) = parent {
+        sync_directory(parent);
+    }
+
+    Ok(())
+}
+
+fn file_name(path: &Path) -> String {
+    path.file_name()
+        .map(|name| name.to_string_lossy().into_owned())
+        .unwrap_or_else(|| "response".to_string())
+}
+
+/// Fsync the directory so the renamed entry survives power loss. Directories
+/// cannot be opened for syncing on Windows, where this is a no-op.
+fn sync_directory(directory: &Path) {
+    #[cfg(unix)]
+    {
+        if let Ok(handle) = fs::File::open(directory) {
+            let _ = handle.sync_all();
+        }
+    }
+    #[cfg(not(unix))]
+    let _ = directory;
+}
+
+/// Fills `buffer` with operating-system randomness.
+///
+/// The device nonce must be unpredictable: it is what stops a captured response
+/// being replayed as a fresh one. Sourced through p256's pinned rand_core 0.6
+/// rather than the workspace `rand` 0.10, matching `identity.rs`; the two are
+/// different crate versions and only the pinned one is on p256's own path.
+fn getrandom(buffer: &mut [u8]) -> Result<(), String> {
+    use p256::elliptic_curve::rand_core::{OsRng, RngCore as _};
+
+    OsRng
+        .try_fill_bytes(buffer)
+        .map_err(|error| format!("the operating system random source failed: {error}"))
 }
@@ -0,0 +1,566 @@
+// 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::time::Duration;
+
+use base64::Engine as _;
+use reqwest::{Client, StatusCode, Url};
+use rustls::RootCertStore;
+use rustls::pki_types::{CertificateDer, pem::PemObject as _};
+use serde::Deserialize;
+use uuid::Uuid;
+use zeroize::Zeroizing;
+
+use super::credential_store::{
+    CompletedRegistration, CredentialStore, CredentialStoreError, DeviceCredential, PendingRegistration, PendingRotation,
+};
+use super::identity::{IdentityError, RegistrationTranscript};
+use super::identity_store::{IdentityStore, StoreError};
+use super::registration::{
+    CredentialResponse, CredentialValidationError, ExpectedDevice, RegistrationRequest, RegistrationToken, RotationRequest,
+    certificate_fingerprint, certificate_request_matches, public_key_fingerprint, validate_credential,
+    validate_stored_credential,
+};
+
+const MAX_ATTEMPTS: usize = 3;
+const MAX_RESPONSE_BYTES: usize = 1024 * 1024;
+const ROTATION_THRESHOLD_SECONDS: i64 = 8 * 60 * 60;
+
+pub struct ConnectConfig<'a> {
+    pub endpoint: &'a str,
+    pub root_ca_pem: &'a [u8],
+    pub timeout: Duration,
+}
+
+pub struct ConnectClient {
+    endpoint: Url,
+    roots: RootCertStore,
+    root_certificates: Vec<CertificateDer<'static>>,
+    client: Client,
+    timeout: Duration,
+}
+
+impl ConnectClient {
+    pub fn from_optional_config(config: Option<ConnectConfig<'_>>) -> Result<Option<Self>, ClientError> {
+        config.map(Self::new).transpose()
+    }
+
+    pub fn new(config: ConnectConfig<'_>) -> Result<Self, ClientError> {
+        let mut endpoint = Url::parse(config.endpoint).map_err(|_| ClientError::Endpoint)?;
+        if endpoint.scheme() != "https"
+            || endpoint.cannot_be_a_base()
+            || !endpoint.username().is_empty()
+            || endpoint.password().is_some()
+            || endpoint.query().is_some()
+            || endpoint.fragment().is_some()
+        {
+            return Err(ClientError::Endpoint);
+        }
+        if !endpoint.path().ends_with('/') {
+            let path = format!("{}/", endpoint.path());
+            endpoint.set_path(&path);
+        }
+
+        let root_certificates = CertificateDer::pem_slice_iter(config.root_ca_pem)
+            .collect::<Result<Vec<_>, _>>()
+            .map_err(|_| ClientError::RootCertificate)?;
+        if root_certificates.is_empty() {
+            return Err(ClientError::RootCertificate);
+        }
+        let mut roots = RootCertStore::empty();
+        let (accepted, rejected) = roots.add_parsable_certificates(root_certificates.clone());
+        if accepted != root_certificates.len() || rejected != 0 {
+            return Err(ClientError::RootCertificate);
+        }
+
+        let client = build_client(&root_certificates, config.timeout, None)?;
+        Ok(Self {
+            endpoint,
+            roots,
+            root_certificates,
+            client,
+            timeout: config.timeout,
+        })
+    }
+
+    pub async fn register(
+        &self,
+        identity_store: &IdentityStore,
+        credential_store: &CredentialStore,
+        token: &RegistrationToken,
+    ) -> Result<DeviceCredential, ClientError> {
+        let _lock = credential_store.lock().await?;
+        if let Some((credential, _)) = self.load_valid_credential(identity_store, credential_store)? {
+            ensure_credential_time(&credential, unix_now())?;
+            return Ok(credential);
+        }
+
+        let identity = identity_store.load_or_create()?;
+        let candidate = PendingRegistration {
+            token_uid: token.registration_token_uid.clone(),
+            request_id: Uuid::new_v4().to_string(),
+            certificate_request: identity.certificate_request_base64()?,
+            previous_credential_fingerprint: None,
+            next_public_key_sha256: None,
+        };
+        let pending = credential_store.claim_pending_registration(&candidate)?;
+        if pending.token_uid != token.registration_token_uid
+            || pending.previous_credential_fingerprint.is_some()
+            || pending.next_public_key_sha256.is_some()
+            || !is_request_id(&pending.request_id)
+            || !certificate_request_matches(&pending.certificate_request, &identity)?
+        {
+            return Err(ClientError::PendingRegistration);
+        }
+
+        let credential = match self.exchange_registration(token, &pending, &identity).await {
+            Ok(credential) => credential,
+            Err(error @ (ClientError::AccessRevoked { .. } | ClientError::Rejected { .. })) => {
+                credential_store.clear_pending_registration()?;
+                return Err(error);
+            }
+            Err(error) => return Err(error),
+        };
+        credential_store.save(&credential)?;
+        credential_store.clear_pending_registration()?;
+        Ok(credential)
+    }
+
+    pub async fn reenroll(
+        &self,
+        identity_store: &IdentityStore,
+        credential_store: &CredentialStore,
+        token: &RegistrationToken,
+    ) -> Result<DeviceCredential, ClientError> {
+        let _lock = credential_store.lock().await?;
+        let (credential, _) = self
+            .load_valid_credential(identity_store, credential_store)?
+            .ok_or(ClientError::NotRegistered)?;
+        let fingerprint = certificate_fingerprint(&credential.certificate)?;
+        if credential_store.load_completed_registration()?.is_some_and(|completed| {
+            completed.token_uid == token.registration_token_uid && completed.credential_fingerprint == fingerprint
+        }) {
+            return Ok(credential);
+        }
+        credential_store.clear_pending_rotation()?;
+        let next = identity_store.load_or_create_next()?;
+        let next_fingerprint = public_key_fingerprint(&next);
+        let candidate = PendingRegistration {
+            token_uid: token.registration_token_uid.clone(),
+            request_id: Uuid::new_v4().to_string(),
+            certificate_request: next.certificate_request_base64()?,
+            previous_credential_fingerprint: Some(fingerprint.clone()),
+            next_public_key_sha256: Some(next_fingerprint.clone()),
+        };
+        let pending = credential_store.claim_pending_registration(&candidate)?;
+        if pending.token_uid != token.registration_token_uid
+            || pending.previous_credential_fingerprint.as_deref() != Some(&fingerprint)
+            || pending.next_public_key_sha256.as_deref() != Some(&next_fingerprint)
+            || !is_request_id(&pending.request_id)
+            || !certificate_request_matches(&pending.certificate_request, &next)?
+        {
+            return Err(ClientError::PendingRegistration);
+        }
+        let enrolled = match self.exchange_registration(token, &pending, &next).await {
+            Ok(credential) => credential,
+            Err(error @ (ClientError::AccessRevoked { .. } | ClientError::Rejected { .. })) => {
+                credential_store.clear_pending_registration()?;
+                identity_store.clear_next()?;
+                return Err(error);
+            }
+            Err(error) => return Err(error),
+        };
+        credential_store.save(&enrolled)?;
+        identity_store.commit_next(&next)?;
+        credential_store.save_completed_registration(&CompletedRegistration {
+            token_uid: token.registration_token_uid.clone(),
+            credential_fingerprint: certificate_fingerprint(&enrolled.certificate)?,
+        })?;
+        credential_store.clear_pending_registration()?;
+        Ok(enrolled)
+    }
+
+    async fn exchange_registration(
+        &self,
+        token: &RegistrationToken,
+        pending: &PendingRegistration,
+        identity: &super::identity::DeviceIdentity,
+    ) -> Result<DeviceCredential, ClientError> {
+        let csr_der = base64::engine::general_purpose::STANDARD
+            .decode(&pending.certificate_request)
+            .map_err(|_| ClientError::PendingRegistration)?;
+        let transcript = RegistrationTranscript::build(
+            &token.registration_token_uid,
+            &token.organization_uid,
+            &token.cluster_uid,
+            &pending.request_id,
+            &token.challenge_nonce,
+            token.expires_unix,
+            &csr_der,
+        )?;
+        let proof = identity.sign_registration(&transcript);
+        let body = RegistrationRequest::new(token, &pending.request_id, &pending.certificate_request, &proof);
+        let url = self.url("./registrationTokens:exchange")?;
+        let response = self
+            .send(StatusCode::CREATED, || self.client.post(url.clone()).json(&body))
+            .await?;
+        let cluster = format!("organizations/{}/clusters/{}", token.organization_uid, token.cluster_uid);
+        let credential = validate_credential(
+            response,
+            identity,
+            &self.roots,
+            &self.root_certificates,
+            ExpectedDevice::Registration { cluster: &cluster },
+        )?;
+        Ok(credential)
+    }
+
+    pub async fn rotate_if_due(
+        &self,
+        identity_store: &IdentityStore,
+        credential_store: &CredentialStore,
+        now_unix: i64,
+    ) -> Result<Option<DeviceCredential>, ClientError> {
+        let _lock = credential_store.lock().await?;
+        let (credential, identity) = self
+            .load_valid_credential(identity_store, credential_store)?
+            .ok_or(ClientError::NotRegistered)?;
+        if credential_store.load_pending_registration()?.is_some() {
+            return Err(ClientError::PendingRegistration);
+        }
+        ensure_credential_time(&credential, now_unix)?;
+        if credential.not_after_unix - now_unix > ROTATION_THRESHOLD_SECONDS {
+            return Ok(None);
+        }
+        let fingerprint = certificate_fingerprint(&credential.certificate)?;
+        let next = identity_store.load_or_create_next()?;
+        let candidate = PendingRotation {
+            credential_fingerprint: fingerprint.clone(),
+            device_name: credential.name.clone(),
+            request_id: Uuid::new_v4().to_string(),
+            certificate_request: next.certificate_request_base64()?,
+            next_public_key_sha256: public_key_fingerprint(&next),
+        };
+        let pending = credential_store.claim_pending_rotation(&candidate)?;
+        if pending.credential_fingerprint != fingerprint
+            || pending.device_name != credential.name
+            || !is_request_id(&pending.request_id)
+            || pending.next_public_key_sha256 != public_key_fingerprint(&next)
+            || !certificate_request_matches(&pending.certificate_request, &next)?
+        {
+            return Err(ClientError::PendingRotation);
+        }
+        let body = RotationRequest::new(
+            &identity,
+            &fingerprint,
+            &credential.name,
+            &pending.request_id,
+            &pending.certificate_request,
+        )?;
+        let private_key = identity.to_pkcs8_pem()?;
+        let mut identity_pem = Zeroizing::new(Vec::with_capacity(credential.certificate_chain.len() + private_key.len() + 1));
+        identity_pem.extend_from_slice(credential.certificate_chain.as_bytes());
+        identity_pem.push(b'\n');
+        identity_pem.extend_from_slice(private_key.as_bytes());
+        let tls_identity = reqwest::Identity::from_pem(&identity_pem).map_err(|_| ClientError::IdentityCertificate)?;
+        let client = build_client(&self.root_certificates, self.timeout, Some(tls_identity))?;
+        let path = format!("clusterDevices/{}:rotateCredential", credential.uid);
+        let url = self.url(&path)?;
+        let response = self.send(StatusCode::OK, || client.post(url.clone()).json(&body)).await?;
+        let rotated = validate_credential(
+            response,
+            &next,
+            &self.roots,
+            &self.root_certificates,
+            ExpectedDevice::Rotation { name: &credential.name },
+        )?;
+        if rotated.name != credential.name || rotated.uid != credential.uid {
+            return Err(ClientError::Credential(CredentialValidationError::Identity));
+        }
+        credential_store.save(&rotated)?;
+        identity_store.commit_next(&next)?;
+        credential_store.clear_pending_rotation()?;
+        Ok(Some(rotated))
+    }
+
+    fn load_valid_credential(
+        &self,
+        identity_store: &IdentityStore,
+        credential_store: &CredentialStore,
+    ) -> Result<Option<(DeviceCredential, super::identity::DeviceIdentity)>, ClientError> {
+        let Some(credential) = credential_store.load()? else {
+            return Ok(None);
+        };
+        let current = identity_store.load()?.ok_or(ClientError::IdentityMissing)?;
+        if let Some(pending) = credential_store.load_pending_registration()? {
+            let Some(previous) = pending.previous_credential_fingerprint.as_deref() else {
+                if pending.next_public_key_sha256.is_some()
+                    || !is_request_id(&pending.request_id)
+                    || !certificate_request_matches(&pending.certificate_request, &current)?
+                {
+                    return Err(ClientError::PendingRegistration);
+                }
+                validate_stored_credential(&credential, &current, &self.roots, &self.root_certificates)?;
+                credential_store.clear_pending_registration()?;
+                return Ok(Some((credential, current)));
+            };
+            let next_fingerprint = pending
+                .next_public_key_sha256
+                .as_deref()
+                .ok_or(ClientError::PendingRegistration)?;
+            if !is_request_id(&pending.request_id) {
+                return Err(ClientError::PendingRegistration);
+            }
+            let fingerprint = certificate_fingerprint(&credential.certificate)?;
+            if fingerprint == previous {
+                validate_stored_credential(&credential, &current, &self.roots, &self.root_certificates)?;
+                let next = identity_store.load_next()?.ok_or(ClientError::PendingRegistration)?;
+                if public_key_fingerprint(&next) != next_fingerprint
+                    || !certificate_request_matches(&pending.certificate_request, &next)?
+                {
+                    return Err(ClientError::PendingRegistration);
+                }
+                return Ok(Some((credential, current)));
+            }
+            if public_key_fingerprint(&current) == next_fingerprint {
+                if !certificate_request_matches(&pending.certificate_request, &current)? {
+                    return Err(ClientError::PendingRegistration);
+                }
+                validate_stored_credential(&credential, &current, &self.roots, &self.root_certificates)?;
+            } else {
+                let next = identity_store.load_next()?.ok_or(ClientError::PendingRegistration)?;
+                if public_key_fingerprint(&next) != next_fingerprint
+                    || !certificate_request_matches(&pending.certificate_request, &next)?
+                {
+                    return Err(ClientError::PendingRegistration);
+                }
+                validate_stored_credential(&credential, &next, &self.roots, &self.root_certificates)?;
+                identity_store.commit_next(&next)?;
+            }
+            credential_store.save_completed_registration(&CompletedRegistration {
+                token_uid: pending.token_uid,
+                credential_fingerprint: fingerprint,
+            })?;
+            credential_store.clear_pending_registration()?;
+            let current = identity_store.load()?.ok_or(ClientError::IdentityMissing)?;
+            return Ok(Some((credential, current)));
+        }
+        let Some(pending) = credential_store.load_pending_rotation()? else {
+            validate_stored_credential(&credential, &current, &self.roots, &self.root_certificates)?;
+            return Ok(Some((credential, current)));
+        };
+        let fingerprint = certificate_fingerprint(&credential.certificate)?;
+        if pending.device_name != credential.name || !is_request_id(&pending.request_id) {
+            return Err(ClientError::PendingRotation);
+        }
+        if fingerprint == pending.credential_fingerprint {
+            validate_stored_credential(&credential, &current, &self.roots, &self.root_certificates)?;
+            let next = identity_store.load_next()?.ok_or(ClientError::PendingRotation)?;
+            if pending.next_public_key_sha256 != public_key_fingerprint(&next)
+                || !certificate_request_matches(&pending.certificate_request, &next)?
+            {
+                return Err(ClientError::PendingRotation);
+            }
+            return Ok(Some((credential, current)));
+        }
+
+        if public_key_fingerprint(&current) == pending.next_public_key_sha256 {
+            if !certificate_request_matches(&pending.certificate_request, &current)? {
+                return Err(ClientError::PendingRotation);
+            }
+            validate_stored_credential(&credential, &current, &self.roots, &self.root_certificates)?;
+        } else {
+            let next = identity_store.load_next()?.ok_or(ClientError::PendingRotation)?;
+            if public_key_fingerprint(&next) != pending.next_public_key_sha256
+                || !certificate_request_matches(&pending.certificate_request, &next)?
+            {
+                return Err(ClientError::PendingRotation);
+            }
+            validate_stored_credential(&credential, &next, &self.roots, &self.root_certificates)?;
+            identity_store.commit_next(&next)?;
+        }
+        credential_store.clear_pending_rotation()?;
+        let current = identity_store.load()?.ok_or(ClientError::IdentityMissing)?;
+        Ok(Some((credential, current)))
+    }
+
+    async fn send<F>(&self, success: StatusCode, mut request: F) -> Result<CredentialResponse, ClientError>
+    where
+        F: FnMut() -> reqwest::RequestBuilder,
+    {
+        let mut last_status = None;
+        for attempt in 0..MAX_ATTEMPTS {
+            match request().send().await {
+                Ok(response) if response.status() == success => return decode_response(response).await,
+                Ok(response) if matches!(response.status(), StatusCode::UNAUTHORIZED | StatusCode::FORBIDDEN) => {
+                    let status = response.status();
+                    let reason = decode_reason(response).await;
+                    return Err(ClientError::AccessRevoked { status, reason });
+                }
+                Ok(response) if matches!(response.status(), StatusCode::REQUEST_TIMEOUT | StatusCode::TOO_MANY_REQUESTS) => {
+                    last_status = Some(response.status());
+                }
+                Ok(response) if response.status().is_client_error() => {
+                    let status = response.status();
+                    let reason = decode_reason(response).await;
+                    return Err(ClientError::Rejected { status, reason });
+                }
+                Ok(response) if response.status().is_server_error() => {
+                    last_status = Some(response.status());
+                }
+                Ok(response) => {
+                    let status = response.status();
+                    let reason = decode_reason(response).await;
+                    return Err(ClientError::Rejected { status, reason });
+                }
+                Err(error) if !error.is_timeout() && !error.is_connect() => return Err(ClientError::Transport(error)),
+                Err(_) => {}
+            }
+
+            if attempt + 1 < MAX_ATTEMPTS {
+                tokio::time::sleep(Duration::from_millis(50 * (attempt as u64 + 1))).await;
+            }
+        }
+        Err(ClientError::Unavailable { status: last_status })
+    }
+
+    fn url(&self, path: &str) -> Result<Url, ClientError> {
+        self.endpoint.join(path).map_err(|_| ClientError::Endpoint)
+    }
+}
+
+fn is_request_id(value: &str) -> bool {
+    Uuid::parse_str(value).is_ok_and(|uuid| uuid.get_version() == Some(uuid::Version::Random) && uuid.to_string() == value)
+}
+
+fn unix_now() -> i64 {
+    std::time::SystemTime::now()
+        .duration_since(std::time::UNIX_EPOCH)
+        .map_or(0, |duration| duration.as_secs() as i64)
+}
+
+fn ensure_credential_time(credential: &DeviceCredential, now_unix: i64) -> Result<(), ClientError> {
+    if credential.not_before_unix > now_unix {
+        return Err(ClientError::CredentialNotYetValid);
+    }
+    if credential.not_after_unix <= now_unix {
+        return Err(ClientError::CredentialExpired);
+    }
+    Ok(())
+}
+
+fn build_client(
+    roots: &[CertificateDer<'static>],
+    timeout: Duration,
+    identity: Option<reqwest::Identity>,
+) -> Result<Client, ClientError> {
+    let certificates = roots
+        .iter()
+        .map(|root| reqwest::Certificate::from_der(root.as_ref()))
+        .collect::<Result<Vec<_>, _>>()?;
+    let mut builder = Client::builder()
+        .https_only(true)
+        .redirect(reqwest::redirect::Policy::none())
+        .timeout(timeout)
+        .tls_certs_only(certificates);
+    if let Some(identity) = identity {
+        builder = builder.identity(identity);
+    }
+    builder.build().map_err(ClientError::Transport)
+}
+
+async fn decode_response(mut response: reqwest::Response) -> Result<CredentialResponse, ClientError> {
+    let body = read_body(&mut response).await?;
+    serde_json::from_slice(&body).map_err(|_| ClientError::Response)
+}
+
+async fn read_body(response: &mut reqwest::Response) -> Result<Vec<u8>, ClientError> {
+    let mut body = Vec::new();
+    while let Some(chunk) = response.chunk().await.map_err(ClientError::Transport)? {
+        if body.len() + chunk.len() > MAX_RESPONSE_BYTES {
+            return Err(ClientError::ResponseTooLarge);
+        }
+        body.extend_from_slice(&chunk);
+    }
+    Ok(body)
+}
+
+#[derive(Deserialize)]
+struct ErrorEnvelope {
+    #[serde(default)]
+    details: Vec<ErrorDetail>,
+}
+
+#[derive(Deserialize)]
+struct ErrorDetail {
+    #[serde(default)]
+    reason: String,
+}
+
+async fn decode_reason(mut response: reqwest::Response) -> Option<String> {
+    let body = read_body(&mut response).await.ok()?;
+    serde_json::from_slice::<ErrorEnvelope>(&body)
+        .ok()?
+        .details
+        .into_iter()
+        .find_map(|detail| (!detail.reason.is_empty()).then_some(detail.reason))
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum ClientError {
+    #[error("Connect endpoint must be an HTTPS base URL without credentials, query, or fragment")]
+    Endpoint,
+    #[error("Connect root CA configuration is invalid")]
+    RootCertificate,
+    #[error(
+        "Connect registration has a pending attempt for a different token; restore the original protected token configuration"
+    )]
+    PendingRegistration,
+    #[error(
+        "Connect credential rotation has an unfinished attempt for a different current certificate; inspect the local credential store"
+    )]
+    PendingRotation,
+    #[error("RustFS is not registered with Connect")]
+    NotRegistered,
+    #[error("the Connect device private key is missing; restore device.key before using the stored certificate")]
+    IdentityMissing,
+    #[error("the Connect device certificate has expired; call ConnectClient::reenroll with a fresh registration token")]
+    CredentialExpired,
+    #[error("the Connect device certificate is not yet valid; fix local clock skew or call ConnectClient::reenroll")]
+    CredentialNotYetValid,
+    #[error("the stored Connect certificate and device private key cannot form a TLS identity")]
+    IdentityCertificate,
+    #[error(
+        "Connect rejected the device credential with HTTP {status}; reason={reason:?}; call ConnectClient::reenroll with a fresh registration token if revoked"
+    )]
+    AccessRevoked { status: StatusCode, reason: Option<String> },
+    #[error("Connect rejected the request with HTTP {status}; reason={reason:?}")]
+    Rejected { status: StatusCode, reason: Option<String> },
+    #[error("Connect remained unavailable after bounded retries; last_status={status:?}")]
+    Unavailable { status: Option<StatusCode> },
+    #[error("Connect response exceeded the 1 MiB credential-response limit")]
+    ResponseTooLarge,
+    #[error("Connect returned an invalid credential response")]
+    Response,
+    #[error(transparent)]
+    Transport(#[from] reqwest::Error),
+    #[error(transparent)]
+    Identity(#[from] IdentityError),
+    #[error(transparent)]
+    IdentityStore(#[from] StoreError),
+    #[error(transparent)]
+    CredentialStore(#[from] CredentialStoreError),
+    #[error(transparent)]
+    Credential(#[from] CredentialValidationError),
+}
@@ -0,0 +1,370 @@
+// 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::fs;
+use std::io::{self, Write as _};
+use std::path::{Path, PathBuf};
+use std::sync::atomic::{AtomicU64, Ordering};
+
+use serde::{Deserialize, Serialize};
+
+const CREDENTIAL_FILE: &str = "device.crt.json";
+const REGISTRATION_COMPLETED_FILE: &str = "registration.completed.json";
+const REGISTRATION_PENDING_FILE: &str = "registration.pending.json";
+const ROTATION_PENDING_FILE: &str = "rotation.pending.json";
+const LOCK_FILE: &str = ".state.lock";
+
+#[cfg(unix)]
+const FILE_MODE: u32 = 0o600;
+
+static STAGING_SEQUENCE: AtomicU64 = AtomicU64::new(0);
+
+#[derive(Clone, Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub struct DeviceCredential {
+    pub name: String,
+    pub uid: String,
+    pub protocol_version: String,
+    pub key_id: String,
+    pub certificate_serial: String,
+    pub certificate: String,
+    pub certificate_chain: String,
+    pub not_before_unix: i64,
+    pub not_after_unix: i64,
+}
+
+impl std::fmt::Debug for DeviceCredential {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("DeviceCredential")
+            .field("name", &self.name)
+            .field("key_id", &self.key_id)
+            .field("certificate_serial", &self.certificate_serial)
+            .field("not_after_unix", &self.not_after_unix)
+            .finish_non_exhaustive()
+    }
+}
+
+#[derive(Clone, Debug, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct PendingRegistration {
+    pub token_uid: String,
+    pub request_id: String,
+    pub certificate_request: String,
+    #[serde(default)]
+    pub previous_credential_fingerprint: Option<String>,
+    #[serde(default)]
+    pub next_public_key_sha256: Option<String>,
+}
+
+#[derive(Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct CompletedRegistration {
+    pub token_uid: String,
+    pub credential_fingerprint: String,
+}
+
+#[derive(Clone, Debug, Serialize, Deserialize)]
+#[serde(deny_unknown_fields)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct PendingRotation {
+    pub credential_fingerprint: String,
+    pub device_name: String,
+    pub request_id: String,
+    pub certificate_request: String,
+    pub next_public_key_sha256: String,
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum CredentialStoreError {
+    #[error("connect credential store I/O failed at {path}: {source}")]
+    Io {
+        path: PathBuf,
+        #[source]
+        source: io::Error,
+    },
+
+    #[error("connect credential data at {path} is invalid: {source}")]
+    Invalid {
+        path: PathBuf,
+        #[source]
+        source: serde_json::Error,
+    },
+
+    #[cfg(unix)]
+    #[error("connect credential file at {path} has mode {mode:o}, expected {expected:o}")]
+    Permissions { path: PathBuf, mode: u32, expected: u32 },
+}
+
+#[derive(Clone, Debug)]
+pub struct CredentialStore {
+    directory: PathBuf,
+}
+
+pub(crate) struct CredentialLock {
+    _file: fs::File,
+}
+
+impl CredentialStore {
+    pub fn new(directory: impl Into<PathBuf>) -> Self {
+        Self {
+            directory: directory.into(),
+        }
+    }
+
+    pub(crate) fn load(&self) -> Result<Option<DeviceCredential>, CredentialStoreError> {
+        self.read(CREDENTIAL_FILE)
+    }
+
+    pub(crate) async fn lock(&self) -> Result<CredentialLock, CredentialStoreError> {
+        let directory = self.directory.clone();
+        tokio::task::spawn_blocking(move || {
+            fs::create_dir_all(&directory).map_err(|source| CredentialStoreError::Io {
+                path: directory.clone(),
+                source,
+            })?;
+            let path = directory.join(LOCK_FILE);
+            let mut options = fs::OpenOptions::new();
+            options.create(true).truncate(false).read(true).write(true);
+            #[cfg(unix)]
+            {
+                use std::os::unix::fs::OpenOptionsExt as _;
+                options.mode(FILE_MODE);
+            }
+            let file = options.open(&path).map_err(|source| CredentialStoreError::Io {
+                path: path.clone(),
+                source,
+            })?;
+            check_mode(&path)?;
+            file.lock().map_err(|source| CredentialStoreError::Io { path, source })?;
+            Ok(CredentialLock { _file: file })
+        })
+        .await
+        .map_err(|source| CredentialStoreError::Io {
+            path: self.directory.join(LOCK_FILE),
+            source: io::Error::other(source),
+        })?
+    }
+
+    pub(crate) fn save(&self, credential: &DeviceCredential) -> Result<(), CredentialStoreError> {
+        self.write(CREDENTIAL_FILE, credential)
+    }
+
+    pub(crate) fn claim_pending_registration(
+        &self,
+        pending: &PendingRegistration,
+    ) -> Result<PendingRegistration, CredentialStoreError> {
+        self.claim(REGISTRATION_PENDING_FILE, pending)
+    }
+
+    pub(crate) fn load_pending_registration(&self) -> Result<Option<PendingRegistration>, CredentialStoreError> {
+        self.read(REGISTRATION_PENDING_FILE)
+    }
+
+    pub(crate) fn clear_pending_registration(&self) -> Result<(), CredentialStoreError> {
+        self.remove(REGISTRATION_PENDING_FILE)
+    }
+
+    pub(crate) fn load_completed_registration(&self) -> Result<Option<CompletedRegistration>, CredentialStoreError> {
+        self.read(REGISTRATION_COMPLETED_FILE)
+    }
+
+    pub(crate) fn save_completed_registration(&self, completed: &CompletedRegistration) -> Result<(), CredentialStoreError> {
+        self.write(REGISTRATION_COMPLETED_FILE, completed)
+    }
+
+    pub(crate) fn load_pending_rotation(&self) -> Result<Option<PendingRotation>, CredentialStoreError> {
+        self.read(ROTATION_PENDING_FILE)
+    }
+
+    pub(crate) fn claim_pending_rotation(&self, pending: &PendingRotation) -> Result<PendingRotation, CredentialStoreError> {
+        self.claim(ROTATION_PENDING_FILE, pending)
+    }
+
+    pub(crate) fn clear_pending_rotation(&self) -> Result<(), CredentialStoreError> {
+        self.remove(ROTATION_PENDING_FILE)
+    }
+
+    fn read<T: for<'de> Deserialize<'de>>(&self, file: &str) -> Result<Option<T>, CredentialStoreError> {
+        let path = self.directory.join(file);
+        let bytes = match fs::read(&path) {
+            Ok(bytes) => bytes,
+            Err(source) if source.kind() == io::ErrorKind::NotFound => return Ok(None),
+            Err(source) => return Err(CredentialStoreError::Io { path, source }),
+        };
+
+        check_mode(&path)?;
+        serde_json::from_slice(&bytes)
+            .map(Some)
+            .map_err(|source| CredentialStoreError::Invalid { path, source })
+    }
+
+    fn write<T: Serialize>(&self, file: &str, value: &T) -> Result<(), CredentialStoreError> {
+        let bytes = serde_json::to_vec(value).map_err(|source| CredentialStoreError::Invalid {
+            path: self.directory.join(file),
+            source,
+        })?;
+
+        fs::create_dir_all(&self.directory).map_err(|source| CredentialStoreError::Io {
+            path: self.directory.clone(),
+            source,
+        })?;
+
+        let final_path = self.directory.join(file);
+        let temp_path = self.stage(file, &bytes)?;
+        let result = fs::rename(&temp_path, &final_path)
+            .map_err(|source| CredentialStoreError::Io {
+                path: final_path,
+                source,
+            })
+            .and_then(|()| {
+                fsync_dir(&self.directory).map_err(|source| CredentialStoreError::Io {
+                    path: self.directory.clone(),
+                    source,
+                })
+            });
+
+        if result.is_err() {
+            let _ = fs::remove_file(&temp_path);
+        }
+        result
+    }
+
+    fn claim<T>(&self, file: &str, value: &T) -> Result<T, CredentialStoreError>
+    where
+        T: Clone + Serialize + for<'de> Deserialize<'de>,
+    {
+        if let Some(existing) = self.read(file)? {
+            return Ok(existing);
+        }
+        let bytes = serde_json::to_vec(value).map_err(|source| CredentialStoreError::Invalid {
+            path: self.directory.join(file),
+            source,
+        })?;
+        fs::create_dir_all(&self.directory).map_err(|source| CredentialStoreError::Io {
+            path: self.directory.clone(),
+            source,
+        })?;
+        let final_path = self.directory.join(file);
+        let temp_path = self.stage(file, &bytes)?;
+        let published = fs::hard_link(&temp_path, &final_path);
+        let _ = fs::remove_file(&temp_path);
+        match published {
+            Ok(()) => {
+                fsync_dir(&self.directory).map_err(|source| CredentialStoreError::Io {
+                    path: self.directory.clone(),
+                    source,
+                })?;
+                Ok(value.clone())
+            }
+            Err(source) if source.kind() == io::ErrorKind::AlreadyExists => {
+                fsync_dir(&self.directory).map_err(|source| CredentialStoreError::Io {
+                    path: self.directory.clone(),
+                    source,
+                })?;
+                self.read(file)?.ok_or_else(|| CredentialStoreError::Io {
+                    path: final_path,
+                    source: io::Error::new(io::ErrorKind::NotFound, "pending state vanished after publication"),
+                })
+            }
+            Err(source) => Err(CredentialStoreError::Io {
+                path: final_path,
+                source,
+            }),
+        }
+    }
+
+    fn stage(&self, file: &str, bytes: &[u8]) -> Result<PathBuf, CredentialStoreError> {
+        loop {
+            let path = self.directory.join(format!(
+                ".{file}.{}.{}.tmp",
+                std::process::id(),
+                STAGING_SEQUENCE.fetch_add(1, Ordering::Relaxed)
+            ));
+            let mut options = fs::OpenOptions::new();
+            options.write(true).create_new(true);
+            #[cfg(unix)]
+            {
+                use std::os::unix::fs::OpenOptionsExt as _;
+                options.mode(FILE_MODE);
+            }
+            let mut staging = match options.open(&path) {
+                Ok(staging) => staging,
+                Err(source) if source.kind() == io::ErrorKind::AlreadyExists => continue,
+                Err(source) => return Err(CredentialStoreError::Io { path, source }),
+            };
+            let result = staging
+                .write_all(bytes)
+                .and_then(|()| staging.sync_all())
+                .map_err(|source| CredentialStoreError::Io {
+                    path: path.clone(),
+                    source,
+                })
+                .and_then(|()| check_mode(&path));
+            if let Err(error) = result {
+                let _ = fs::remove_file(&path);
+                return Err(error);
+            }
+            return Ok(path);
+        }
+    }
+
+    fn remove(&self, file: &str) -> Result<(), CredentialStoreError> {
+        let path = self.directory.join(file);
+        match fs::remove_file(&path) {
+            Ok(()) => fsync_dir(&self.directory).map_err(|source| CredentialStoreError::Io {
+                path: self.directory.clone(),
+                source,
+            }),
+            Err(source) if source.kind() == io::ErrorKind::NotFound => Ok(()),
+            Err(source) => Err(CredentialStoreError::Io { path, source }),
+        }
+    }
+}
+
+#[cfg(unix)]
+fn check_mode(path: &Path) -> Result<(), CredentialStoreError> {
+    use std::os::unix::fs::PermissionsExt as _;
+
+    let mode = fs::metadata(path)
+        .map_err(|source| CredentialStoreError::Io {
+            path: path.to_path_buf(),
+            source,
+        })?
+        .permissions()
+        .mode()
+        & 0o7777;
+    if mode != FILE_MODE {
+        return Err(CredentialStoreError::Permissions {
+            path: path.to_path_buf(),
+            mode,
+            expected: FILE_MODE,
+        });
+    }
+    Ok(())
+}
+
+#[cfg(not(unix))]
+fn check_mode(_path: &Path) -> Result<(), CredentialStoreError> {
+    Ok(())
+}
+
+fn fsync_dir(directory: &Path) -> io::Result<()> {
+    #[cfg(unix)]
+    fs::File::open(directory)?.sync_all()?;
+    #[cfg(not(unix))]
+    let _ = directory;
+    Ok(())
+}
@@ -24,7 +24,7 @@ use base64::Engine as _;
 use base64::engine::general_purpose::{STANDARD as BASE64_STANDARD, URL_SAFE_NO_PAD as BASE64_URL_NO_PAD};
 use p256::ecdsa::signature::Signer as _;
 use p256::ecdsa::{Signature, SigningKey};
-use p256::pkcs8::{DecodePrivateKey as _, EncodePrivateKey as _};
+use p256::pkcs8::{DecodePrivateKey as _, EncodePrivateKey as _, LineEnding};
 use sha2::{Digest as _, Sha256};
 use zeroize::Zeroizing;

@@ -213,6 +213,12 @@ impl DeviceIdentity {
            .map_err(|error| IdentityError::MalformedKey(error.to_string()))
    }

+    pub(crate) fn to_pkcs8_pem(&self) -> Result<Zeroizing<String>, IdentityError> {
+        self.signing_key
+            .to_pkcs8_pem(LineEnding::LF)
+            .map_err(|error| IdentityError::MalformedKey(error.to_string()))
+    }
+
    /// Build the PKCS#10 certificate request Connect consumes.
    ///
    /// Connect reads the request for its SubjectPublicKeyInfo and its
@@ -32,6 +32,7 @@ use super::identity::{DeviceIdentity, IdentityError};

 /// Name of the key file inside the store directory.
 const KEY_FILE: &str = "device.key";
+const NEXT_KEY_FILE: &str = "device.key.next";

 /// Owner read/write only. The key is the device's whole identity.
 #[cfg(unix)]
@@ -94,7 +95,15 @@ impl IdentityStore {
    /// been enrolled. Reading never creates anything, so an unconfigured
    /// server can ask without acquiring an identity as a side effect.
    pub fn load(&self) -> Result<Option<DeviceIdentity>, StoreError> {
-        let path = self.key_path();
+        self.load_file(KEY_FILE)
+    }
+
+    pub(crate) fn load_next(&self) -> Result<Option<DeviceIdentity>, StoreError> {
+        self.load_file(NEXT_KEY_FILE)
+    }
+
+    fn load_file(&self, file: &str) -> Result<Option<DeviceIdentity>, StoreError> {
+        let path = self.directory.join(file);

        let der = match fs::read(&path) {
            Ok(der) => Zeroizing::new(der),
@@ -142,11 +151,15 @@ impl IdentityStore {
        let candidate = DeviceIdentity::generate();
        let der = candidate.to_pkcs8_der()?;

-        match self.publish(&der) {
+        match self.publish(KEY_FILE, &der) {
            Ok(()) => Ok(candidate),
            // Another process published first. Its key is the identity; ours
            // was never written anywhere and simply goes out of scope.
            Err(StoreError::Io { source, .. }) if source.kind() == io::ErrorKind::AlreadyExists => {
+                fsync_dir(&self.directory).map_err(|source| StoreError::Io {
+                    path: self.directory.clone(),
+                    source,
+                })?;
                self.load()?.ok_or_else(|| StoreError::Io {
                    path: self.key_path(),
                    source: io::Error::new(
@@ -159,35 +172,106 @@ impl IdentityStore {
        }
    }

+    pub(crate) fn load_or_create_next(&self) -> Result<DeviceIdentity, StoreError> {
+        if let Some(identity) = self.load_next()? {
+            return Ok(identity);
+        }
+        fs::create_dir_all(&self.directory).map_err(|source| StoreError::Io {
+            path: self.directory.clone(),
+            source,
+        })?;
+        let candidate = DeviceIdentity::generate();
+        let der = candidate.to_pkcs8_der()?;
+        match self.publish(NEXT_KEY_FILE, &der) {
+            Ok(()) => Ok(candidate),
+            Err(StoreError::Io { source, .. }) if source.kind() == io::ErrorKind::AlreadyExists => {
+                fsync_dir(&self.directory).map_err(|source| StoreError::Io {
+                    path: self.directory.clone(),
+                    source,
+                })?;
+                self.load_next()?.ok_or_else(|| StoreError::Io {
+                    path: self.directory.join(NEXT_KEY_FILE),
+                    source: io::Error::new(io::ErrorKind::NotFound, "next device key vanished after publication"),
+                })
+            }
+            Err(error) => Err(error),
+        }
+    }
+
+    pub(crate) fn commit_next(&self, expected: &DeviceIdentity) -> Result<(), StoreError> {
+        let next_path = self.directory.join(NEXT_KEY_FILE);
+        match fs::rename(&next_path, self.key_path()) {
+            Ok(()) => fsync_dir(&self.directory).map_err(|source| StoreError::Io {
+                path: self.directory.clone(),
+                source,
+            }),
+            Err(source) if source.kind() == io::ErrorKind::NotFound => {
+                let current = self.load()?.ok_or_else(|| StoreError::Io {
+                    path: self.key_path(),
+                    source,
+                })?;
+                if current.public_key_der() == expected.public_key_der() {
+                    fsync_dir(&self.directory).map_err(|source| StoreError::Io {
+                        path: self.directory.clone(),
+                        source,
+                    })
+                } else {
+                    Err(StoreError::Io {
+                        path: next_path,
+                        source: io::Error::new(io::ErrorKind::NotFound, "next device key is missing"),
+                    })
+                }
+            }
+            Err(source) => Err(StoreError::Io { path: next_path, source }),
+        }
+    }
+
+    pub(crate) fn clear_next(&self) -> Result<(), StoreError> {
+        let path = self.directory.join(NEXT_KEY_FILE);
+        match fs::remove_file(&path) {
+            Ok(()) => fsync_dir(&self.directory).map_err(|source| StoreError::Io {
+                path: self.directory.clone(),
+                source,
+            }),
+            Err(source) if source.kind() == io::ErrorKind::NotFound => Ok(()),
+            Err(source) => Err(StoreError::Io { path, source }),
+        }
+    }
+
    /// Write, seal, fsync, then link into place and fsync the directory. The
    /// key is durable before it is reachable, and it is reachable only once.
-    fn publish(&self, der: &[u8]) -> Result<(), StoreError> {
+    fn publish(&self, file: &str, der: &[u8]) -> Result<(), StoreError> {
        use std::io::Write as _;

-        let final_path = self.key_path();
-        let temp_path = self.directory.join(format!(
-            "{KEY_FILE}.{}.{}.tmp",
-            std::process::id(),
-            STAGING_SEQUENCE.fetch_add(1, Ordering::Relaxed)
-        ));
+        let final_path = self.directory.join(file);

        let io_at = |path: &Path| {
            let path = path.to_path_buf();
            move |source| StoreError::Io { path, source }
        };

-        let mut options = fs::OpenOptions::new();
-        options.write(true).create(true).truncate(true);
-        #[cfg(unix)]
-        {
-            use std::os::unix::fs::OpenOptionsExt as _;
-            options.mode(KEY_MODE);
-        }
-
-        let mut file = options.open(&temp_path).map_err(io_at(&temp_path))?;
+        let (temp_path, mut staging) = loop {
+            let temp_path = self.directory.join(format!(
+                ".{file}.{}.{}.tmp",
+                std::process::id(),
+                STAGING_SEQUENCE.fetch_add(1, Ordering::Relaxed)
+            ));
+            let mut options = fs::OpenOptions::new();
+            options.write(true).create_new(true);
+            #[cfg(unix)]
+            {
+                use std::os::unix::fs::OpenOptionsExt as _;
+                options.mode(KEY_MODE);
+            }
+            match options.open(&temp_path) {
+                Ok(staging) => break (temp_path, staging),
+                Err(source) if source.kind() == io::ErrorKind::AlreadyExists => continue,
+                Err(source) => return Err(io_at(&temp_path)(source)),
+            }
+        };

        let result = (|| -> Result<(), StoreError> {
-            file.write_all(der).map_err(io_at(&temp_path))?;
+            staging.write_all(der).map_err(io_at(&temp_path))?;

            // The umask can only narrow the creation mode, so set and verify
            // the exact mode before the bytes become durable.
@@ -195,9 +279,10 @@ impl IdentityStore {
            {
                use std::os::unix::fs::PermissionsExt as _;

-                file.set_permissions(fs::Permissions::from_mode(KEY_MODE))
+                staging
+                    .set_permissions(fs::Permissions::from_mode(KEY_MODE))
                    .map_err(io_at(&temp_path))?;
-                let mode = file.metadata().map_err(io_at(&temp_path))?.permissions().mode() & 0o7777;
+                let mode = staging.metadata().map_err(io_at(&temp_path))?.permissions().mode() & 0o7777;
                if mode != KEY_MODE {
                    return Err(StoreError::Permissions {
                        path: temp_path.clone(),
@@ -207,11 +292,11 @@ impl IdentityStore {
                }
            }

-            file.sync_all().map_err(io_at(&temp_path))?;
+            staging.sync_all().map_err(io_at(&temp_path))?;
            Ok(())
        })();

-        drop(file);
+        drop(staging);

        if let Err(error) = result {
            let _ = fs::remove_file(&temp_path);
@@ -25,8 +25,16 @@
 //! not been enrolled into a Connect control plane never calls into it, so an
 //! unconfigured server generates no key and holds no identity.

+pub mod client;
+pub mod credential_store;
 pub mod identity;
 pub mod identity_store;
+pub mod offline;
+pub mod registration;

+pub use client::{ClientError, ConnectClient, ConnectConfig};
+pub use credential_store::{CredentialStore, DeviceCredential};
 pub use identity::{DeviceIdentity, IdentityError, RegistrationProof, RegistrationTranscript};
 pub use identity_store::{IdentityStore, StoreError};
+pub use offline::{EnrollmentError, OfflineEnrollment, OfflineKeyStore, VerifiedChallenge};
+pub use registration::{RegistrationToken, TokenError};
@@ -0,0 +1,684 @@
+// 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.
+
+//! Challenge verification and response production for offline enrolment.
+//!
+//! Two invariants carry the security of this surface and both are easy to break
+//! by accident:
+//!
+//! - Every signature is checked over the octets that arrived, never over a
+//!   re-serialised document. Parsing happens only to route the verification, and
+//!   nothing a parse yields is believed until the signature over those same
+//!   octets has verified.
+//! - The enrolment root is the constant in this file. It is never taken from a
+//!   challenge, a configuration file, or an operator prompt, so there is no
+//!   trust-on-first-use path an operator could be talked into.
+//!
+//! The order of the checks in [`OfflineEnrollment::verify_challenge`] is frozen
+//! by `verificationOrder.enrollmentChallenge` in
+//! `protocol/agent/v1/fixtures/offline-enrollment/trust-model.json`, and the
+//! signature encoding, the domain separation tags, and every rejection reason
+//! are frozen beside it. Reordering the checks changes which reason a given
+//! artifact produces, which is itself part of the contract.
+
+use base64::Engine as _;
+use base64::engine::general_purpose::{STANDARD as BASE64_STANDARD, URL_SAFE_NO_PAD as BASE64_URL_NO_PAD};
+use p256::ecdsa::signature::{Signer as _, Verifier as _};
+use p256::ecdsa::{Signature, SigningKey, VerifyingKey};
+use p256::pkcs8::DecodePrivateKey as _;
+use serde::{Deserialize, Serialize};
+use sha2::{Digest as _, Sha256};
+use time::{Date, Month, OffsetDateTime, PrimitiveDateTime, Time};
+
+use crate::connect::identity::DeviceIdentity;
+
+/// The hosted enrolment root, compiled in. Both halves are pinned: the
+/// fingerprint identifies the root, and the point is what actually verifies the
+/// first link, so a build cannot be pointed at a different key by supplying one.
+const PINNED_ROOT_KEY_ID: &str = "df22e2806112debbe953672aafa186d699af0e97dd3fd2b09fa8359005fe348f";
+const PINNED_ROOT_PUBLIC_KEY: &str = "BFfx-K-FfEA5nK_Rz3IHacvRCkJyQ7JOd1geLyU6HKRZDgNezmVuKhvJ22VhemyjV__Gshk8JGGqOBzYPMD0p6s";
+
+/// Domain separation tags. A document that verifies under one of these must not
+/// be accepted for another artifact type, so the tag is part of the signature
+/// input rather than a property of the caller.
+const TAG_TRUST_LINK: &[u8] = b"rustfs-offline-trust-link-v1";
+const TAG_CHALLENGE: &[u8] = b"rustfs-offline-enrollment-challenge-v1";
+const TAG_RESPONSE: &[u8] = b"rustfs-offline-enrollment-response-v1";
+
+/// The single octet between the tag and the signed document.
+const DOMAIN_SEPARATOR: u8 = 0x00;
+
+const SIGNATURE_ALGORITHM: &str = "ES256";
+const PROTOCOL_VERSION: &str = "v1";
+const FORMAT_TRUST_LINK: &str = "rustfs.connect.offline.trustLink/1";
+const FORMAT_CHALLENGE: &str = "rustfs.connect.offline.enrollmentChallenge/1";
+const FORMAT_RESPONSE: &str = "rustfs.connect.offline.enrollmentResponse/1";
+
+/// DER SubjectPublicKeyInfo header for an uncompressed P-256 point. A keyId is
+/// the SHA-256 of this prefix followed by the 65 octet point, so the prefix is
+/// also how a device public key is recovered from its own DER encoding.
+const SPKI_PREFIX: [u8; 26] = [
+    0x30, 0x59, 0x30, 0x13, 0x06, 0x07, 0x2a, 0x86, 0x48, 0xce, 0x3d, 0x02, 0x01, 0x06, 0x08, 0x2a, 0x86, 0x48, 0xce, 0x3d, 0x03,
+    0x01, 0x07, 0x03, 0x42, 0x00,
+];
+
+/// Order of the P-256 group, and half of it. `r` and `s` must lie in `[1, n)`,
+/// and `s` additionally in `[1, n/2]`: ECDSA admits both `s` and `n - s`, and a
+/// signature with two spellings cannot serve as an artifact identity. Every
+/// ECDSA library accepts the malleated form, so the encoding layer rejects it.
+const GROUP_ORDER: [u8; 32] = [
+    0xff, 0xff, 0xff, 0xff, 0x00, 0x00, 0x00, 0x00, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xbc, 0xe6, 0xfa, 0xad, 0xa7,
+    0x17, 0x9e, 0x84, 0xf3, 0xb9, 0xca, 0xc2, 0xfc, 0x63, 0x25, 0x51,
+];
+const MAX_S: [u8; 32] = [
+    0x7f, 0xff, 0xff, 0xff, 0x80, 0x00, 0x00, 0x00, 0x7f, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xde, 0x73, 0x7d, 0x56, 0xd3,
+    0x8b, 0xcf, 0x42, 0x79, 0xdc, 0xe5, 0x61, 0x7e, 0x31, 0x92, 0xa8,
+];
+
+const SCALAR_OCTETS: usize = 32;
+const SIGNATURE_OCTETS: usize = 64;
+/// 64 octets as unpadded base64url. The length is checked before decoding so
+/// that `=` padding, the standard alphabet, DER, and a truncated value are all
+/// refused rather than repaired.
+const SIGNATURE_VALUE_CHARS: usize = 86;
+
+const PUBLIC_KEY_OCTETS: usize = 65;
+const PUBLIC_KEY_CHARS: usize = 87;
+/// SEC1 tag of an uncompressed point. Compressed and hybrid forms are refused.
+const UNCOMPRESSED_POINT: u8 = 0x04;
+
+const TIMESTAMP_CHARS: usize = 20;
+
+/// The chain is exactly two links: a pinned root issues the intermediate, and
+/// the intermediate issues the signing key. Roles are positional and the
+/// enumeration is closed.
+const CHAIN_LINK_COUNT: usize = 2;
+const CHAIN_ROLES: [&str; CHAIN_LINK_COUNT] = ["intermediate", "signing"];
+
+/// Skew allowed on the challenge window. A device may have no synchronised
+/// clock at all, so its own reading of "now" is advisory.
+const CLOCK_SKEW_TOLERANCE: i64 = 300;
+
+/// Longest life a challenge may claim. The issuer sets both ends of its own
+/// window, so the protocol bound is applied on top of the declared expiry
+/// rather than trusted from it.
+const MAX_CHALLENGE_LIFETIME: i64 = 604_800;
+
+/// A challenge that verified, with the fields the response has to echo.
+///
+/// Construction is the proof: a value of this type only exists after the chain
+/// closed on the pinned root and the challenge signature verified over the
+/// received octets.
+#[derive(Clone, Debug, PartialEq, Eq)]
+pub struct VerifiedChallenge {
+    pub challenge_id: String,
+    pub organization_name: String,
+    pub cluster_name: String,
+    pub nonce: String,
+    pub issued_at: String,
+    pub expires_at: String,
+    pub connect_key_id: String,
+    /// The signature value of the challenge, verbatim. It binds a response to
+    /// the one challenge it answers, so it is carried rather than recomputed.
+    pub challenge_proof: String,
+}
+
+/// Why an offline enrolment artifact was refused.
+///
+/// The variants are the frozen `reason` vocabulary of
+/// `fixtures/offline-enrollment/error-codes.json`, which spans both halves of
+/// the exchange. The device half implemented here produces the encoding, chain,
+/// version, and freshness reasons; the reasons that describe a response being
+/// evaluated against stored state — [`Self::ChallengeUnknown`],
+/// [`Self::ChallengeProofInvalid`], [`Self::DeviceProofInvalid`],
+/// [`Self::EnrollmentReplayed`], [`Self::OrganizationMismatch`], and
+/// [`Self::ClusterMismatch`] — are Connect's to raise and are named here so the
+/// two sides share one vocabulary.
+///
+/// No variant carries a payload: a rejection must never disclose key material,
+/// signature octets, nonces, or document bytes.
+#[derive(Debug, PartialEq, Eq, thiserror::Error)]
+pub enum EnrollmentError {
+    #[error("protocolVersion is missing, malformed, or names an unsupported major version")]
+    UnsupportedProtocol,
+
+    #[error("formatVersion is not a supported offline enrollment format")]
+    UnsupportedFormat,
+
+    #[error("the signature is not 64 octets of fixed-width r||s in unpadded base64url")]
+    SignatureMalformed,
+
+    #[error("the signature is not in its canonical low-S form")]
+    SignatureNotCanonical,
+
+    #[error("the signature does not verify over the received octets")]
+    SignatureInvalid,
+
+    #[error("the trust chain is not issued by a root pinned in this build")]
+    EnrollmentRootUnknown,
+
+    #[error("a trust link is invalid, misordered, or outside its validity at the challenge issuedAt")]
+    TrustChainInvalid,
+
+    #[error("connectKeyId is not the subject of the last trust link")]
+    ConnectKeyUnchained,
+
+    #[error("no issued challenge matches this challengeId")]
+    ChallengeUnknown,
+
+    #[error("the challenge is not yet valid at the evaluation time")]
+    ChallengeNotYetValid,
+
+    #[error("the challenge has expired at the evaluation time")]
+    ChallengeExpired,
+
+    #[error("the response nonce or challengeProof is not the one issued for this challenge")]
+    ChallengeProofInvalid,
+
+    #[error("the response does not prove possession of the device key it presents")]
+    DeviceProofInvalid,
+
+    #[error("the challenge was already consumed")]
+    EnrollmentReplayed,
+
+    #[error("the response names a different organization than the challenge it answers")]
+    OrganizationMismatch,
+
+    #[error("the response names a different cluster than the challenge it answers")]
+    ClusterMismatch,
+
+    /// The artifact could not be read as a signed enrolment document at all: the
+    /// envelope, the base64 of the signed octets, or a field the frozen order
+    /// reads before the signature verifies did not parse. The frozen reason set
+    /// has no code for a structurally unreadable document, so this variant maps
+    /// to none of them.
+    #[error("the offline enrollment document is not well formed")]
+    MalformedDocument,
+
+    /// A fault on this side of the exchange rather than in the artifact: the
+    /// device key did not round-trip through its own PKCS#8 encoding, or the
+    /// caller named an instant outside the representable calendar. Fails closed
+    /// because a half-produced response must never reach removable media.
+    #[error("the enrollment response could not be produced on this device")]
+    ResponseNotProduced,
+}
+
+impl EnrollmentError {
+    /// The frozen `reason` an operator and Connect both branch on.
+    ///
+    /// The `Display` message is prose and may be reworded; this is the stable
+    /// identifier, so nothing should parse the message instead. The two
+    /// variants with no frozen counterpart deliberately return codes outside
+    /// the frozen set rather than borrowing the nearest one, so a document that
+    /// simply failed to parse can never be reported as a signature or freshness
+    /// failure.
+    pub fn reason(&self) -> &'static str {
+        match self {
+            Self::UnsupportedProtocol => "UNSUPPORTED_PROTOCOL",
+            Self::UnsupportedFormat => "UNSUPPORTED_FORMAT",
+            Self::SignatureMalformed => "SIGNATURE_MALFORMED",
+            Self::SignatureNotCanonical => "SIGNATURE_NOT_CANONICAL",
+            Self::SignatureInvalid => "SIGNATURE_INVALID",
+            Self::EnrollmentRootUnknown => "ENROLLMENT_ROOT_UNKNOWN",
+            Self::TrustChainInvalid => "TRUST_CHAIN_INVALID",
+            Self::ConnectKeyUnchained => "CONNECT_KEY_UNCHAINED",
+            Self::ChallengeUnknown => "CHALLENGE_UNKNOWN",
+            Self::ChallengeNotYetValid => "CHALLENGE_NOT_YET_VALID",
+            Self::ChallengeExpired => "CHALLENGE_EXPIRED",
+            Self::ChallengeProofInvalid => "CHALLENGE_PROOF_INVALID",
+            Self::DeviceProofInvalid => "DEVICE_PROOF_INVALID",
+            Self::EnrollmentReplayed => "ENROLLMENT_REPLAYED",
+            Self::OrganizationMismatch => "ORGANIZATION_MISMATCH",
+            Self::ClusterMismatch => "CLUSTER_MISMATCH",
+            Self::MalformedDocument => "MALFORMED_DOCUMENT",
+            Self::ResponseNotProduced => "RESPONSE_NOT_PRODUCED",
+        }
+    }
+}
+
+/// A signed document, in the shape both directions carry it. `bytes` is
+/// standard padded base64 of the exact octets that were signed; nothing else is
+/// ever used as the signature input.
+#[derive(Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct SignedDocument {
+    bytes: String,
+    signature: DocumentSignature,
+}
+
+#[derive(Serialize, Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct DocumentSignature {
+    algorithm: String,
+    key_id: String,
+    value: String,
+}
+
+/// The three fields the frozen order permits reading before anything verifies.
+/// They route the verification and are not facts until it has.
+#[derive(Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct ChallengeRouting {
+    connect_key_id: String,
+    issued_at: String,
+    trust_chain: Vec<SignedDocument>,
+}
+
+#[derive(Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct ChallengeDocument {
+    format_version: String,
+    protocol_version: String,
+    challenge_id: String,
+    organization_name: String,
+    cluster_name: String,
+    nonce: String,
+    issued_at: String,
+    expires_at: String,
+    connect_key_id: String,
+}
+
+#[derive(Deserialize)]
+#[serde(rename_all = "camelCase")]
+struct TrustLink {
+    format_version: String,
+    protocol_version: String,
+    role: String,
+    issuer_key_id: String,
+    subject_key_id: String,
+    subject_public_key: String,
+    not_before: String,
+    not_after: String,
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+struct ResponseDocument<'a> {
+    format_version: &'a str,
+    protocol_version: &'a str,
+    challenge_id: &'a str,
+    organization_name: &'a str,
+    cluster_name: &'a str,
+    challenge_nonce: &'a str,
+    challenge_proof: &'a str,
+    device_key_id: String,
+    device_public_key: String,
+    device_nonce: String,
+    produced_at: String,
+}
+
+/// The device half of the offline enrolment exchange: bytes in, bytes out.
+pub struct OfflineEnrollment;
+
+impl OfflineEnrollment {
+    /// Verify an enrolment challenge and return what a response must echo.
+    ///
+    /// `now_unix` is the device's reading of the current time, which the clock
+    /// skew tolerance treats as advisory.
+    pub fn verify_challenge(document: &[u8], now_unix: i64) -> Result<VerifiedChallenge, EnrollmentError> {
+        let envelope: SignedDocument = serde_json::from_slice(document).map_err(|_| EnrollmentError::MalformedDocument)?;
+
+        // Step 1: the encoding is checked before anything is decoded from it, so
+        // a DER, padded, truncated, out-of-range, or high-S signature is refused
+        // on its spelling rather than handed to a library that would accept it.
+        let signature = decode_signature(&envelope.signature)?;
+
+        // The octets that were transmitted. They are never re-serialised: every
+        // later step signs and parses this same buffer.
+        let bytes = BASE64_STANDARD
+            .decode(envelope.bytes.as_bytes())
+            .map_err(|_| EnrollmentError::MalformedDocument)?;
+
+        // Step 2: routing only.
+        let routing: ChallengeRouting = serde_json::from_slice(&bytes).map_err(|_| EnrollmentError::MalformedDocument)?;
+        let issued_at = parse_timestamp(&routing.issued_at)?;
+
+        // Steps 3 to 5.
+        let connect_key = verify_trust_chain(&routing.trust_chain, &routing.connect_key_id, issued_at)?;
+
+        // Step 6. The verification key comes from the chain, so `signature.keyId`
+        // is a label rather than an input: a value naming some other key simply
+        // fails to verify here.
+        if !verifies(&connect_key, TAG_CHALLENGE, &bytes, &signature) {
+            return Err(EnrollmentError::SignatureInvalid);
+        }
+
+        // Step 7: only now is the document read as a fact.
+        let challenge: ChallengeDocument = serde_json::from_slice(&bytes).map_err(|_| EnrollmentError::MalformedDocument)?;
+        if challenge.protocol_version != PROTOCOL_VERSION {
+            return Err(EnrollmentError::UnsupportedProtocol);
+        }
+        if challenge.format_version != FORMAT_CHALLENGE {
+            return Err(EnrollmentError::UnsupportedFormat);
+        }
+
+        // Step 8.
+        let expires_at = parse_timestamp(&challenge.expires_at)?;
+        check_challenge_window(issued_at, expires_at, now_unix)?;
+
+        Ok(VerifiedChallenge {
+            challenge_id: challenge.challenge_id,
+            organization_name: challenge.organization_name,
+            cluster_name: challenge.cluster_name,
+            nonce: challenge.nonce,
+            issued_at: challenge.issued_at,
+            expires_at: challenge.expires_at,
+            connect_key_id: challenge.connect_key_id,
+            challenge_proof: envelope.signature.value,
+        })
+    }
+
+    /// Build the signed response an operator carries back to Connect.
+    ///
+    /// `device_nonce` is the response's own replay value and must come from a
+    /// cryptographic source. The private key never appears in the result: only
+    /// the public point, its fingerprint, and a signature over the document
+    /// that presents them, which is what makes presenting the key safe.
+    pub fn build_response(
+        challenge: &VerifiedChallenge,
+        key: &DeviceIdentity,
+        device_nonce: &[u8; 32],
+        produced_at_unix: i64,
+    ) -> Result<Vec<u8>, EnrollmentError> {
+        let issued_at = parse_timestamp(&challenge.issued_at)?;
+        let expires_at = parse_timestamp(&challenge.expires_at)?;
+        // Connect re-checks producedAt against the same window, so a response
+        // outside it is refused here rather than written to media and rejected
+        // after the operator has carried it out.
+        check_challenge_window(issued_at, expires_at, produced_at_unix)?;
+
+        let point = device_public_point(key)?;
+        let produced_at = format_timestamp(produced_at_unix)?;
+
+        let document = ResponseDocument {
+            format_version: FORMAT_RESPONSE,
+            protocol_version: PROTOCOL_VERSION,
+            challenge_id: &challenge.challenge_id,
+            organization_name: &challenge.organization_name,
+            cluster_name: &challenge.cluster_name,
+            challenge_nonce: &challenge.nonce,
+            challenge_proof: &challenge.challenge_proof,
+            device_key_id: key_id(&point),
+            device_public_key: BASE64_URL_NO_PAD.encode(point),
+            device_nonce: BASE64_URL_NO_PAD.encode(device_nonce),
+            produced_at,
+        };
+
+        // Serialised once. These octets are what is signed and what is carried,
+        // so no second serialisation can disagree with the signature.
+        let bytes = serde_json::to_vec(&document).map_err(|_| EnrollmentError::ResponseNotProduced)?;
+        let signature = sign(key, TAG_RESPONSE, &bytes)?;
+
+        let envelope = SignedDocument {
+            bytes: BASE64_STANDARD.encode(&bytes),
+            signature: DocumentSignature {
+                algorithm: SIGNATURE_ALGORITHM.to_owned(),
+                key_id: document.device_key_id,
+                value: signature,
+            },
+        };
+
+        serde_json::to_vec(&envelope).map_err(|_| EnrollmentError::ResponseNotProduced)
+    }
+}
+
+/// Walk the chain from the pinned root to the signing key, returning the key
+/// `connect_key_id` names once the chain vouches for it.
+fn verify_trust_chain(
+    chain: &[SignedDocument],
+    connect_key_id: &str,
+    challenge_issued_at: i64,
+) -> Result<VerifyingKey, EnrollmentError> {
+    // The pinned root gate runs before the chain's shape is examined, so a
+    // chain that is internally consistent under a foreign root — exactly what
+    // trust on first use would have accepted — is refused for its root rather
+    // than for its length.
+    let first = chain.first().ok_or(EnrollmentError::EnrollmentRootUnknown)?;
+    let root = decode_trust_link(first)?;
+    if root.0.issuer_key_id != PINNED_ROOT_KEY_ID {
+        return Err(EnrollmentError::EnrollmentRootUnknown);
+    }
+
+    let [_, second] = chain else {
+        return Err(EnrollmentError::TrustChainInvalid);
+    };
+    let links = [root, decode_trust_link(second)?];
+
+    let mut issuer_key_id = PINNED_ROOT_KEY_ID.to_owned();
+    let (mut issuer_key, _) = decode_public_key(PINNED_ROOT_PUBLIC_KEY).ok_or(EnrollmentError::EnrollmentRootUnknown)?;
+
+    for (index, ((link, link_bytes), entry)) in links.iter().zip(chain).enumerate() {
+        if link.format_version != FORMAT_TRUST_LINK
+            || link.protocol_version != PROTOCOL_VERSION
+            || link.role != CHAIN_ROLES[index]
+            || link.issuer_key_id != issuer_key_id
+            // A link that names itself as its own issuer would let a stolen
+            // intermediate mint its own root.
+            || link.subject_key_id == link.issuer_key_id
+        {
+            return Err(EnrollmentError::TrustChainInvalid);
+        }
+
+        let (subject_key, subject_point) =
+            decode_public_key(&link.subject_public_key).ok_or(EnrollmentError::TrustChainInvalid)?;
+        if key_id(&subject_point) != link.subject_key_id {
+            return Err(EnrollmentError::TrustChainInvalid);
+        }
+
+        let signature = decode_signature(&entry.signature)?;
+        if !verifies(&issuer_key, TAG_TRUST_LINK, link_bytes, &signature) {
+            return Err(EnrollmentError::TrustChainInvalid);
+        }
+
+        // The issuer controls both ends of a link's window, so it is evaluated
+        // with no skew tolerance, and against the challenge's issuedAt rather
+        // than against the device clock: a challenge carries the chain that was
+        // valid when it was issued.
+        let not_before = parse_timestamp(&link.not_before)?;
+        let not_after = parse_timestamp(&link.not_after)?;
+        if challenge_issued_at < not_before || challenge_issued_at > not_after {
+            return Err(EnrollmentError::TrustChainInvalid);
+        }
+
+        issuer_key_id = link.subject_key_id.clone();
+        issuer_key = subject_key;
+    }
+
+    if issuer_key_id != connect_key_id {
+        return Err(EnrollmentError::ConnectKeyUnchained);
+    }
+
+    Ok(issuer_key)
+}
+
+/// Decode a link and keep the octets it was signed over: the signature is
+/// checked against these, never against a re-encoding of the parsed link.
+fn decode_trust_link(entry: &SignedDocument) -> Result<(TrustLink, Vec<u8>), EnrollmentError> {
+    let bytes = BASE64_STANDARD
+        .decode(entry.bytes.as_bytes())
+        .map_err(|_| EnrollmentError::MalformedDocument)?;
+    let link = serde_json::from_slice(&bytes).map_err(|_| EnrollmentError::TrustChainInvalid)?;
+    Ok((link, bytes))
+}
+
+/// Check a signature's spelling and range, then admit it.
+///
+/// `r` and `s` are compared against the group order here rather than left to
+/// the ECDSA library, because a library that accepts high-S — every library
+/// does — would let a malleated copy of an artifact pass as a second artifact.
+fn decode_signature(signature: &DocumentSignature) -> Result<Signature, EnrollmentError> {
+    if signature.algorithm != SIGNATURE_ALGORITHM {
+        return Err(EnrollmentError::SignatureMalformed);
+    }
+
+    let value = signature.value.as_bytes();
+    if value.len() != SIGNATURE_VALUE_CHARS || !value.iter().all(|byte| is_base64url(*byte)) {
+        return Err(EnrollmentError::SignatureMalformed);
+    }
+
+    let decoded = BASE64_URL_NO_PAD
+        .decode(value)
+        .map_err(|_| EnrollmentError::SignatureMalformed)?;
+    let octets: [u8; SIGNATURE_OCTETS] = decoded
+        .as_slice()
+        .try_into()
+        .map_err(|_| EnrollmentError::SignatureMalformed)?;
+
+    // Big-endian octets of equal length order lexicographically exactly as the
+    // integers they spell, so a slice comparison is the range check.
+    let (r, s) = octets.split_at(SCALAR_OCTETS);
+    let out_of_range = |scalar: &[u8]| scalar.iter().all(|byte| *byte == 0) || scalar >= &GROUP_ORDER[..];
+    if out_of_range(r) || out_of_range(s) {
+        return Err(EnrollmentError::SignatureMalformed);
+    }
+    if s > &MAX_S[..] {
+        return Err(EnrollmentError::SignatureNotCanonical);
+    }
+
+    Signature::from_slice(&octets).map_err(|_| EnrollmentError::SignatureMalformed)
+}
+
+fn verifies(key: &VerifyingKey, tag: &[u8], bytes: &[u8], signature: &Signature) -> bool {
+    key.verify(&signature_input(tag, bytes), signature).is_ok()
+}
+
+fn signature_input(tag: &[u8], bytes: &[u8]) -> Vec<u8> {
+    let mut input = Vec::with_capacity(tag.len() + 1 + bytes.len());
+    input.extend_from_slice(tag);
+    input.push(DOMAIN_SEPARATOR);
+    input.extend_from_slice(bytes);
+    input
+}
+
+fn sign(key: &DeviceIdentity, tag: &[u8], bytes: &[u8]) -> Result<String, EnrollmentError> {
+    // `DeviceIdentity` publishes no general signing operation, so the key is
+    // rebuilt from its own PKCS#8 encoding; the encoding is wiped when the
+    // wrapper drops.
+    let pkcs8 = key.to_pkcs8_der().map_err(|_| EnrollmentError::ResponseNotProduced)?;
+    let signing_key = SigningKey::from_pkcs8_der(pkcs8.as_slice()).map_err(|_| EnrollmentError::ResponseNotProduced)?;
+
+    let signature: Signature = signing_key.sign(&signature_input(tag, bytes));
+    let canonical = signature.normalize_s().unwrap_or(signature);
+
+    Ok(BASE64_URL_NO_PAD.encode(canonical.to_bytes()))
+}
+
+/// The device's public point, recovered from the DER encoding the identity
+/// publishes so that one prefix constant governs both the fingerprint and the
+/// wire form.
+fn device_public_point(key: &DeviceIdentity) -> Result<[u8; PUBLIC_KEY_OCTETS], EnrollmentError> {
+    key.public_key_der()
+        .strip_prefix(&SPKI_PREFIX)
+        .and_then(|point| <[u8; PUBLIC_KEY_OCTETS]>::try_from(point).ok())
+        .ok_or(EnrollmentError::ResponseNotProduced)
+}
+
+/// Decode an uncompressed SEC1 point and check that it is on the curve.
+///
+/// The length and alphabet are checked before decoding so that a padded or
+/// standard-alphabet spelling is refused, and the point tag is checked so that
+/// the compressed and hybrid forms — which no keyId would match — cannot be
+/// spelled at all.
+fn decode_public_key(value: &str) -> Option<(VerifyingKey, [u8; PUBLIC_KEY_OCTETS])> {
+    let value = value.as_bytes();
+    if value.len() != PUBLIC_KEY_CHARS || !value.iter().all(|byte| is_base64url(*byte)) {
+        return None;
+    }
+
+    let point: [u8; PUBLIC_KEY_OCTETS] = BASE64_URL_NO_PAD.decode(value).ok()?.try_into().ok()?;
+    if point[0] != UNCOMPRESSED_POINT {
+        return None;
+    }
+
+    VerifyingKey::from_sec1_bytes(&point).ok().map(|key| (key, point))
+}
+
+/// Lowercase SHA-256 hex of the DER SubjectPublicKeyInfo built from a 65 octet
+/// uncompressed point.
+fn key_id(point: &[u8]) -> String {
+    let mut digest = Sha256::new();
+    digest.update(SPKI_PREFIX);
+    digest.update(point);
+    hex_simd::encode_to_string(digest.finalize(), hex_simd::AsciiCase::Lower)
+}
+
+fn is_base64url(byte: u8) -> bool {
+    byte.is_ascii_alphanumeric() || byte == b'-' || byte == b'_'
+}
+
+/// Parse `YYYY-MM-DDTHH:MM:SSZ` into a Unix instant.
+///
+/// The shape is checked before the fields are read: offsets other than `Z` and
+/// fractional seconds are refused rather than normalised, so two producers
+/// cannot spell the same instant two ways.
+fn parse_timestamp(value: &str) -> Result<i64, EnrollmentError> {
+    let octets = value.as_bytes();
+    if octets.len() != TIMESTAMP_CHARS
+        || octets[4] != b'-'
+        || octets[7] != b'-'
+        || octets[10] != b'T'
+        || octets[13] != b':'
+        || octets[16] != b':'
+        || octets[19] != b'Z'
+    {
+        return Err(EnrollmentError::MalformedDocument);
+    }
+
+    let field = |range: std::ops::Range<usize>| -> Result<u32, EnrollmentError> {
+        let text = &value[range];
+        if !text.bytes().all(|byte| byte.is_ascii_digit()) {
+            return Err(EnrollmentError::MalformedDocument);
+        }
+        text.parse().map_err(|_| EnrollmentError::MalformedDocument)
+    };
+
+    let month = Month::try_from(field(5..7)? as u8).map_err(|_| EnrollmentError::MalformedDocument)?;
+    let date = Date::from_calendar_date(field(0..4)? as i32, month, field(8..10)? as u8)
+        .map_err(|_| EnrollmentError::MalformedDocument)?;
+    let clock = Time::from_hms(field(11..13)? as u8, field(14..16)? as u8, field(17..19)? as u8)
+        .map_err(|_| EnrollmentError::MalformedDocument)?;
+
+    Ok(PrimitiveDateTime::new(date, clock).assume_utc().unix_timestamp())
+}
+
+fn format_timestamp(unix: i64) -> Result<String, EnrollmentError> {
+    let moment = OffsetDateTime::from_unix_timestamp(unix).map_err(|_| EnrollmentError::ResponseNotProduced)?;
+    Ok(format!(
+        "{:04}-{:02}-{:02}T{:02}:{:02}:{:02}Z",
+        moment.year(),
+        u8::from(moment.month()),
+        moment.day(),
+        moment.hour(),
+        moment.minute(),
+        moment.second()
+    ))
+}
+
+/// `at` must fall within `[issuedAt - 300, expiresAt + 300]`.
+///
+/// The declared expiry is capped at the protocol's maximum challenge lifetime
+/// because the issuer sets both ends of its own window; a challenge claiming a
+/// longer life expires at the bound.
+fn check_challenge_window(issued_at: i64, expires_at: i64, at: i64) -> Result<(), EnrollmentError> {
+    if at < issued_at.saturating_sub(CLOCK_SKEW_TOLERANCE) {
+        return Err(EnrollmentError::ChallengeNotYetValid);
+    }
+
+    let effective_expiry = expires_at.min(issued_at.saturating_add(MAX_CHALLENGE_LIFETIME));
+    if at > effective_expiry.saturating_add(CLOCK_SKEW_TOLERANCE) {
+        return Err(EnrollmentError::ChallengeExpired);
+    }
+
+    Ok(())
+}
@@ -0,0 +1,72 @@
+// 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.
+
+//! On-disk home of the offline enrollment key.
+//!
+//! An air-gapped device enrols with a key that is not its online device
+//! identity: the online key is minted during a registration exchange this
+//! device cannot perform, and an operator who carries an enrolment response out
+//! on removable media is enrolling exactly one key that Connect will pin. Losing
+//! it means asking for a fresh challenge, so it is written durably and published
+//! exactly once.
+//!
+//! The durability protocol is not reimplemented here. [`IdentityStore`] already
+//! seals a P-256 key at mode 0600, fsyncs it, and publishes it through a
+//! no-clobber link so a retry or a concurrent start converges on one key; it is
+//! pointed at a directory of this key's own rather than generalised into a
+//! key-store abstraction that would have to describe both lifecycles.
+
+use std::path::{Path, PathBuf};
+
+use super::super::identity::DeviceIdentity;
+use super::super::identity_store::{IdentityStore, StoreError};
+
+/// Subdirectory holding the offline enrolment key, kept apart from the online
+/// device identity so neither can be read in place of the other.
+const OFFLINE_DIRECTORY: &str = "offline";
+
+/// The offline enrolment key of one deployment.
+#[derive(Clone, Debug)]
+pub struct OfflineKeyStore {
+    inner: IdentityStore,
+}
+
+impl OfflineKeyStore {
+    pub fn new(directory: impl AsRef<Path>) -> Self {
+        Self {
+            inner: IdentityStore::new(directory.as_ref().join(OFFLINE_DIRECTORY)),
+        }
+    }
+
+    pub fn key_path(&self) -> PathBuf {
+        self.inner.key_path()
+    }
+
+    /// Return the stored key, or `None` when this deployment has never enrolled
+    /// offline. Reading never creates one, so a deployment that only ever
+    /// registers online holds no offline key.
+    pub fn load(&self) -> Result<Option<DeviceIdentity>, StoreError> {
+        self.inner.load()
+    }
+
+    /// Return the stored key, generating and publishing one the first time.
+    ///
+    /// A second enrolment attempt returns the original key rather than minting a
+    /// replacement: the operator may already be carrying a response for it, and
+    /// two keys would mean the response and the device disagree about which one
+    /// Connect pinned.
+    pub fn load_or_create(&self) -> Result<DeviceIdentity, StoreError> {
+        self.inner.load_or_create()
+    }
+}
@@ -0,0 +1,35 @@
+// 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.
+
+//! Offline enrolment: joining a Connect tenant without a network.
+//!
+//! An air-gapped cluster cannot perform the registration exchange, so an
+//! operator carries a signed challenge in and a signed response out. The device
+//! half of that exchange lives here: verifying the challenge against a root
+//! whose fingerprint is compiled into this binary, minting the key being
+//! enrolled, and signing the response.
+//!
+//! Nothing here opens a socket. That is the point of the surface, and it is
+//! asserted rather than assumed: the enrolment path takes bytes and returns
+//! bytes.
+//!
+//! The trust model, the signing convention, and every rejection reason are
+//! frozen by `protocol/agent/v1/fixtures/offline-enrollment/` and by
+//! `docs/adr/0009-offline-signing.md` on the Connect side.
+
+pub mod enrollment;
+pub mod key_store;
+
+pub use enrollment::{EnrollmentError, OfflineEnrollment, VerifiedChallenge};
+pub use key_store::OfflineKeyStore;
@@ -0,0 +1,527 @@
+// 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::Read;
+use std::sync::Arc;
+
+use base64::Engine as _;
+use base64::engine::general_purpose::URL_SAFE_NO_PAD as BASE64_URL_NO_PAD;
+use p256::ecdsa::signature::Signer as _;
+use p256::ecdsa::{Signature, SigningKey};
+use p256::pkcs8::DecodePrivateKey as _;
+use rustls::RootCertStore;
+use rustls::pki_types::{CertificateDer, UnixTime, pem::PemObject as _};
+use rustls::server::WebPkiClientVerifier;
+use serde::{Deserialize, Serialize};
+use sha2::{Digest as _, Sha256};
+use time::OffsetDateTime;
+use time::format_description::well_known::Rfc3339;
+use uuid::{Uuid, Version};
+use x509_parser::extensions::GeneralName;
+use x509_parser::oid_registry::OID_SIG_ECDSA_WITH_SHA256;
+use x509_parser::prelude::{FromDer as _, X509Certificate, X509CertificationRequest};
+use zeroize::Zeroizing;
+
+use super::credential_store::DeviceCredential;
+use super::identity::{DeviceIdentity, RegistrationProof};
+
+pub const PROTOCOL_VERSION: &str = "v1";
+
+const CERTIFICATE_LIFETIME_SECONDS: i64 = 86_400;
+const ROTATION_DOMAIN: &[u8] = b"RUSTFS-CONNECT-CREDENTIAL-ROTATION-V1";
+const MAX_TOKEN_BYTES: u64 = 16 * 1024;
+
+#[derive(Deserialize)]
+#[serde(deny_unknown_fields, rename_all = "camelCase")]
+struct RegistrationTokenDocument {
+    registration_token_uid: String,
+    registration_token_secret: String,
+    organization_uid: String,
+    cluster_uid: String,
+    challenge_nonce: String,
+    expires_unix: i64,
+}
+
+pub struct RegistrationToken {
+    pub registration_token_uid: String,
+    registration_token_secret: Zeroizing<String>,
+    pub organization_uid: String,
+    pub cluster_uid: String,
+    pub challenge_nonce: String,
+    pub expires_unix: i64,
+}
+
+impl std::fmt::Debug for RegistrationToken {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("RegistrationToken")
+            .field("registration_token_uid", &self.registration_token_uid)
+            .field("expires_unix", &self.expires_unix)
+            .finish_non_exhaustive()
+    }
+}
+
+impl RegistrationToken {
+    pub fn from_reader(reader: impl Read) -> Result<Self, TokenError> {
+        let mut bytes = Zeroizing::new(Vec::new());
+        reader
+            .take(MAX_TOKEN_BYTES + 1)
+            .read_to_end(&mut bytes)
+            .map_err(TokenError::Read)?;
+        if bytes.len() as u64 > MAX_TOKEN_BYTES {
+            return Err(TokenError::TooLarge);
+        }
+        let document: RegistrationTokenDocument = serde_json::from_slice(&bytes).map_err(TokenError::Invalid)?;
+        let decoded = BASE64_URL_NO_PAD
+            .decode(&document.registration_token_secret)
+            .map(Zeroizing::new)
+            .map_err(|_| TokenError::SecretShape)?;
+        if decoded.len() != 32 || BASE64_URL_NO_PAD.encode(&decoded) != document.registration_token_secret {
+            return Err(TokenError::SecretShape);
+        }
+        if !is_uuid_v7(&document.registration_token_uid)
+            || !is_uuid_v7(&document.organization_uid)
+            || !is_uuid_v7(&document.cluster_uid)
+            || document.expires_unix < 0
+            || document.challenge_nonce.len() != 64
+            || !document
+                .challenge_nonce
+                .bytes()
+                .all(|byte| byte.is_ascii_digit() || (b'a'..=b'f').contains(&byte))
+        {
+            return Err(TokenError::Shape);
+        }
+
+        Ok(Self {
+            registration_token_uid: document.registration_token_uid,
+            registration_token_secret: Zeroizing::new(document.registration_token_secret),
+            organization_uid: document.organization_uid,
+            cluster_uid: document.cluster_uid,
+            challenge_nonce: document.challenge_nonce,
+            expires_unix: document.expires_unix,
+        })
+    }
+
+    pub(crate) fn secret(&self) -> &str {
+        &self.registration_token_secret
+    }
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum TokenError {
+    #[error("failed to read the Connect registration token")]
+    Read(#[source] std::io::Error),
+    #[error("Connect registration token configuration is invalid")]
+    Invalid(#[source] serde_json::Error),
+    #[error("Connect registration token secret must be 32-byte unpadded base64url")]
+    SecretShape,
+    #[error("Connect registration token configuration exceeds 16 KiB")]
+    TooLarge,
+    #[error("Connect registration token fields do not match the protocol schema")]
+    Shape,
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct RegistrationRequest<'a> {
+    protocol_version: &'static str,
+    request_id: &'a str,
+    registration_token_uid: &'a str,
+    registration_token_secret: &'a str,
+    certificate_request: &'a str,
+    proof: ProofRef<'a>,
+}
+
+#[derive(Serialize)]
+struct ProofRef<'a> {
+    algorithm: &'a str,
+    value: &'a str,
+}
+
+impl<'a> RegistrationRequest<'a> {
+    pub(crate) fn new(
+        token: &'a RegistrationToken,
+        request_id: &'a str,
+        certificate_request: &'a str,
+        proof: &'a RegistrationProof,
+    ) -> Self {
+        Self {
+            protocol_version: PROTOCOL_VERSION,
+            request_id,
+            registration_token_uid: &token.registration_token_uid,
+            registration_token_secret: token.secret(),
+            certificate_request,
+            proof: ProofRef {
+                algorithm: &proof.algorithm,
+                value: &proof.value,
+            },
+        }
+    }
+}
+
+#[derive(Serialize)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct RotationRequest<'a> {
+    protocol_version: &'static str,
+    request_id: &'a str,
+    certificate_request: &'a str,
+    proof: ProofOwned,
+}
+
+#[derive(Serialize)]
+struct ProofOwned {
+    algorithm: String,
+    value: String,
+}
+
+impl<'a> RotationRequest<'a> {
+    pub(crate) fn new(
+        identity: &DeviceIdentity,
+        credential_fingerprint: &str,
+        device_name: &str,
+        request_id: &'a str,
+        certificate_request: &'a str,
+    ) -> Result<Self, CredentialValidationError> {
+        let csr_der = base64::engine::general_purpose::STANDARD
+            .decode(certificate_request)
+            .map_err(|_| CredentialValidationError::CertificateRequest)?;
+        let csr_digest = BASE64_URL_NO_PAD.encode(Sha256::digest(&csr_der));
+        let transcript = rotation_transcript(credential_fingerprint, device_name, request_id, &csr_digest)?;
+        let key = identity
+            .to_pkcs8_der()
+            .map_err(|_| CredentialValidationError::CertificateRequest)?;
+        let signing_key = SigningKey::from_pkcs8_der(&key).map_err(|_| CredentialValidationError::CertificateRequest)?;
+        let signature: Signature = signing_key.sign(&transcript);
+        let canonical = signature.normalize_s().unwrap_or(signature);
+
+        Ok(Self {
+            protocol_version: PROTOCOL_VERSION,
+            request_id,
+            certificate_request,
+            proof: ProofOwned {
+                algorithm: "ES256".to_string(),
+                value: BASE64_URL_NO_PAD.encode(canonical.to_bytes()),
+            },
+        })
+    }
+}
+
+fn rotation_transcript(
+    credential_fingerprint: &str,
+    device_name: &str,
+    request_id: &str,
+    csr_digest: &str,
+) -> Result<Vec<u8>, CredentialValidationError> {
+    let fields = [credential_fingerprint, device_name, request_id, csr_digest];
+    if fields
+        .iter()
+        .any(|field| !field.is_ascii() || field.as_bytes().contains(&b'\n'))
+    {
+        return Err(CredentialValidationError::RotationTranscript);
+    }
+
+    let mut transcript = Vec::with_capacity(346);
+    transcript.extend_from_slice(ROTATION_DOMAIN);
+    transcript.push(b'\n');
+    for field in fields {
+        transcript.extend_from_slice(field.len().to_string().as_bytes());
+        transcript.push(b':');
+        transcript.extend_from_slice(field.as_bytes());
+        transcript.push(b'\n');
+    }
+    Ok(transcript)
+}
+
+#[derive(Deserialize)]
+#[serde(rename_all = "camelCase")]
+pub(crate) struct CredentialResponse {
+    pub name: String,
+    #[serde(default)]
+    pub uid: String,
+    #[serde(default)]
+    pub cluster: String,
+    pub protocol_version: String,
+    pub key_id: String,
+    pub certificate_serial: String,
+    pub certificate: String,
+    pub certificate_chain: String,
+    pub not_before: String,
+    pub not_after: String,
+}
+
+pub(crate) enum ExpectedDevice<'a> {
+    Registration { cluster: &'a str },
+    Rotation { name: &'a str },
+    Stored,
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum CredentialValidationError {
+    #[error("Connect returned malformed certificate material")]
+    Certificate,
+    #[error("Connect returned a certificate chain that is not trusted")]
+    Chain,
+    #[error("Connect returned a certificate for the wrong device identity")]
+    Identity,
+    #[error("Connect returned a certificate for a different device key")]
+    Key,
+    #[error("Connect returned an invalid certificate validity window")]
+    Validity,
+    #[error("the device certificate request could not be prepared")]
+    CertificateRequest,
+    #[error("the credential rotation transcript contains an invalid field")]
+    RotationTranscript,
+}
+
+pub(crate) fn validate_credential(
+    response: CredentialResponse,
+    identity: &DeviceIdentity,
+    roots: &RootCertStore,
+    root_certificates: &[CertificateDer<'static>],
+    expected: ExpectedDevice<'_>,
+) -> Result<DeviceCredential, CredentialValidationError> {
+    validate_credential_at(response, identity, roots, root_certificates, expected, true)
+}
+
+pub(crate) fn validate_stored_credential(
+    credential: &DeviceCredential,
+    identity: &DeviceIdentity,
+    roots: &RootCertStore,
+    root_certificates: &[CertificateDer<'static>],
+) -> Result<(), CredentialValidationError> {
+    let not_before = OffsetDateTime::from_unix_timestamp(credential.not_before_unix)
+        .map_err(|_| CredentialValidationError::Validity)?
+        .format(&Rfc3339)
+        .map_err(|_| CredentialValidationError::Validity)?;
+    let not_after = OffsetDateTime::from_unix_timestamp(credential.not_after_unix)
+        .map_err(|_| CredentialValidationError::Validity)?
+        .format(&Rfc3339)
+        .map_err(|_| CredentialValidationError::Validity)?;
+    let response = CredentialResponse {
+        name: credential.name.clone(),
+        uid: credential.uid.clone(),
+        cluster: String::new(),
+        protocol_version: credential.protocol_version.clone(),
+        key_id: credential.key_id.clone(),
+        certificate_serial: credential.certificate_serial.clone(),
+        certificate: credential.certificate.clone(),
+        certificate_chain: credential.certificate_chain.clone(),
+        not_before,
+        not_after,
+    };
+    validate_credential_at(response, identity, roots, root_certificates, ExpectedDevice::Stored, false).map(|_| ())
+}
+
+fn validate_credential_at(
+    response: CredentialResponse,
+    identity: &DeviceIdentity,
+    roots: &RootCertStore,
+    root_certificates: &[CertificateDer<'static>],
+    expected: ExpectedDevice<'_>,
+    verify_now: bool,
+) -> Result<DeviceCredential, CredentialValidationError> {
+    if response.protocol_version != PROTOCOL_VERSION {
+        return Err(CredentialValidationError::Identity);
+    }
+
+    let leaves = CertificateDer::pem_slice_iter(response.certificate.as_bytes())
+        .collect::<Result<Vec<_>, _>>()
+        .map_err(|_| CredentialValidationError::Certificate)?;
+    if leaves.len() != 1 {
+        return Err(CredentialValidationError::Certificate);
+    }
+    let chain = CertificateDer::pem_slice_iter(response.certificate_chain.as_bytes())
+        .collect::<Result<Vec<_>, _>>()
+        .map_err(|_| CredentialValidationError::Certificate)?;
+    if chain.is_empty()
+        || chain[0].as_ref() != leaves[0].as_ref()
+        || chain
+            .iter()
+            .skip(1)
+            .any(|certificate| root_certificates.iter().any(|root| root.as_ref() == certificate.as_ref()))
+    {
+        return Err(CredentialValidationError::Chain);
+    }
+
+    let (remaining, certificate) =
+        X509Certificate::from_der(leaves[0].as_ref()).map_err(|_| CredentialValidationError::Certificate)?;
+    if !remaining.is_empty() {
+        return Err(CredentialValidationError::Certificate);
+    }
+
+    let uid = match expected {
+        ExpectedDevice::Registration { cluster } => {
+            if response.uid.is_empty()
+                || response.cluster != cluster
+                || response.name != format!("{cluster}/clusterDevices/{}", response.uid)
+            {
+                return Err(CredentialValidationError::Identity);
+            }
+            response.uid.clone()
+        }
+        ExpectedDevice::Rotation { name } => {
+            if response.name != name || !response.uid.is_empty() || !response.cluster.is_empty() {
+                return Err(CredentialValidationError::Identity);
+            }
+            name.rsplit_once("/clusterDevices/")
+                .map(|(_, uid)| uid.to_string())
+                .ok_or(CredentialValidationError::Identity)?
+        }
+        ExpectedDevice::Stored => {
+            let (cluster, name_uid) = response
+                .name
+                .rsplit_once("/clusterDevices/")
+                .ok_or(CredentialValidationError::Identity)?;
+            if response.uid != name_uid
+                || !response.cluster.is_empty()
+                || !valid_cluster_name(cluster)
+                || response.name.matches("/clusterDevices/").count() != 1
+            {
+                return Err(CredentialValidationError::Identity);
+            }
+            response.uid.clone()
+        }
+    };
+    let parsed_uid = Uuid::parse_str(&uid).map_err(|_| CredentialValidationError::Identity)?;
+    if parsed_uid.get_version() != Some(Version::SortRand) || parsed_uid.to_string() != uid {
+        return Err(CredentialValidationError::Identity);
+    }
+    let expected_uri = format!("urn:rustfs:connect:device:{uid}");
+    let common_names = certificate
+        .subject()
+        .iter_common_name()
+        .map(|name| name.as_str())
+        .collect::<Result<Vec<_>, _>>()
+        .map_err(|_| CredentialValidationError::Identity)?;
+    let san = certificate
+        .subject_alternative_name()
+        .map_err(|_| CredentialValidationError::Identity)?
+        .ok_or(CredentialValidationError::Identity)?;
+    let san_matches = matches!(san.value.general_names.as_slice(), [GeneralName::URI(uri)] if *uri == expected_uri);
+
+    if common_names.as_slice() != [uid.as_str()]
+        || !san_matches
+        || certificate.subject().iter().count() != 1
+        || certificate.subject().iter_attributes().count() != 1
+    {
+        return Err(CredentialValidationError::Identity);
+    }
+    if certificate.public_key().raw != identity.public_key_der() {
+        return Err(CredentialValidationError::Key);
+    }
+
+    let not_before = OffsetDateTime::parse(&response.not_before, &Rfc3339)
+        .map_err(|_| CredentialValidationError::Validity)?
+        .unix_timestamp();
+    let not_after = OffsetDateTime::parse(&response.not_after, &Rfc3339)
+        .map_err(|_| CredentialValidationError::Validity)?
+        .unix_timestamp();
+    let verify_unix = if verify_now {
+        UnixTime::now()
+    } else {
+        let midpoint = certificate.validity().not_before.timestamp()
+            + (certificate.validity().not_after.timestamp() - certificate.validity().not_before.timestamp()) / 2;
+        UnixTime::since_unix_epoch(std::time::Duration::from_secs(
+            midpoint.try_into().map_err(|_| CredentialValidationError::Validity)?,
+        ))
+    };
+    let verifier = WebPkiClientVerifier::builder(Arc::new(roots.clone()))
+        .build()
+        .map_err(|_| CredentialValidationError::Chain)?;
+    verifier
+        .verify_client_cert(&leaves[0], &chain[1..], verify_unix)
+        .map_err(|_| CredentialValidationError::Chain)?;
+    if not_before != certificate.validity().not_before.timestamp()
+        || not_after != certificate.validity().not_after.timestamp()
+        || not_after - not_before != CERTIFICATE_LIFETIME_SECONDS
+        || certificate.signature_algorithm.algorithm != OID_SIG_ECDSA_WITH_SHA256
+        || response.certificate_serial != canonical_serial(certificate.raw_serial())?
+        || response.key_id != format!("x509-{}", response.certificate_serial)
+    {
+        return Err(CredentialValidationError::Validity);
+    }
+
+    Ok(DeviceCredential {
+        name: response.name,
+        uid,
+        protocol_version: response.protocol_version,
+        key_id: response.key_id,
+        certificate_serial: response.certificate_serial,
+        certificate: response.certificate,
+        certificate_chain: response.certificate_chain,
+        not_before_unix: not_before,
+        not_after_unix: not_after,
+    })
+}
+
+pub(crate) fn public_key_fingerprint(identity: &DeviceIdentity) -> String {
+    hex_lower(&Sha256::digest(identity.public_key_der()))
+}
+
+pub(crate) fn certificate_request_matches(encoded: &str, identity: &DeviceIdentity) -> Result<bool, CredentialValidationError> {
+    let der = base64::engine::general_purpose::STANDARD
+        .decode(encoded)
+        .map_err(|_| CredentialValidationError::CertificateRequest)?;
+    let (remaining, request) =
+        X509CertificationRequest::from_der(&der).map_err(|_| CredentialValidationError::CertificateRequest)?;
+    Ok(remaining.is_empty() && request.certification_request_info.subject_pki.raw == identity.public_key_der())
+}
+
+fn canonical_serial(raw: &[u8]) -> Result<String, CredentialValidationError> {
+    let magnitude = match raw {
+        [0, first, rest @ ..] if first & 0x80 != 0 => {
+            if rest.len() + 1 > 16 {
+                return Err(CredentialValidationError::Validity);
+            }
+            &raw[1..]
+        }
+        [0] => raw,
+        [0, ..] | [] => return Err(CredentialValidationError::Validity),
+        [first, ..] if first & 0x80 != 0 => return Err(CredentialValidationError::Validity),
+        _ if raw.len() > 16 => return Err(CredentialValidationError::Validity),
+        _ => raw,
+    };
+    let mut padded = [0u8; 16];
+    padded[16 - magnitude.len()..].copy_from_slice(magnitude);
+    Ok(hex_lower(&padded))
+}
+
+fn is_uuid_v7(value: &str) -> bool {
+    Uuid::parse_str(value).is_ok_and(|uuid| uuid.get_version() == Some(Version::SortRand) && uuid.to_string() == value)
+}
+
+fn valid_cluster_name(name: &str) -> bool {
+    let Some((organization, cluster)) = name
+        .strip_prefix("organizations/")
+        .and_then(|rest| rest.split_once("/clusters/"))
+    else {
+        return false;
+    };
+    !cluster.contains('/') && is_uuid_v7(organization) && is_uuid_v7(cluster)
+}
+
+pub(crate) fn certificate_fingerprint(certificate_pem: &str) -> Result<String, CredentialValidationError> {
+    let certificate = CertificateDer::pem_slice_iter(certificate_pem.as_bytes())
+        .next()
+        .ok_or(CredentialValidationError::Certificate)?
+        .map_err(|_| CredentialValidationError::Certificate)?;
+    Ok(hex_lower(&Sha256::digest(certificate.as_ref())))
+}
+
+fn hex_lower(bytes: &[u8]) -> String {
+    bytes.iter().fold(String::with_capacity(bytes.len() * 2), |mut output, byte| {
+        use std::fmt::Write as _;
+        let _ = write!(output, "{byte:02x}");
+        output
+    })
+}
@@ -1537,7 +1537,7 @@ fn process_connection(
                None
            }
        };
-        // ── Canonical Middleware Stack Order (outermost → innermost) ──
+        // ── Canonical External Middleware Stack Order (outermost → innermost) ──
        // This order MUST be preserved across refactorings.
        // Only AddExtensionLayer (layers 1-2) are per-connection; most remaining layers are stateless.
        //
@@ -1565,6 +1565,8 @@ fn process_connection(
        // 22. PublicHealthEndpointLayer              — handles public health before s3s host parsing
        // 23. VirtualHostStyleHintLayer              — actionable error for unroutable virtual-hosted-style (conditional)
        // 24. DoubleSlashListBucketsCompatLayer      — rewrites `GET //` to `GET /` for ListBuckets (MinIO browser compat)
+        // The internode lane below intentionally keeps only the shared
+        // transport/auth/observability subset needed by `/rustfs/rpc/...`.
        // ─────────────────────────────────────────────────────────────
        let build_external_stack = |service| {
            ServiceBuilder::new()
@@ -1747,16 +1749,9 @@ fn process_connection(
                .layer(PropagateRequestIdLayer::x_request_id())
                .layer(CompressionLayer::new().compress_when(PathAwareHttpCompressionPredicate::new(compression_config.clone())))
                .option_layer(compression_config.enabled.then_some(PathCategoryInjectionLayer))
-                .layer(S3ErrorMessageCompatLayer)
-                .layer(IcebergRestErrorCompatLayer)
-                .layer(ObjectAttributesEtagFixLayer)
-                .layer(ConditionalCorsLayer::new())
-                .option_layer(if is_console { Some(RedirectLayer) } else { None })
-                .layer(BodylessStatusFixLayer)
-                .layer(HeadRequestBodyFixLayer)
-                .layer(PublicHealthEndpointLayer::new(Arc::clone(&server_ctx)))
-                .option_layer((!server_domains_configured && !is_console).then_some(VirtualHostStyleHintLayer))
-                .layer(DoubleSlashListBucketsCompatLayer)
+                // The internode lane only serves `/rustfs/rpc/...` gRPC requests.
+                // Keep safety/observability layers above, but leave S3/REST
+                // compatibility rewrites on the external lane.
                .service(service)
        };
        let external_stack_service = build_external_stack(external_service);
@@ -296,7 +296,10 @@ pub(crate) fn build_list_objects_v2_output(
            let mut obj = Object {
                key: Some(key),
                last_modified: v.mod_time.map(Timestamp::from),
-                size: Some(v.get_actual_size().unwrap_or_default()),
+                // Compressed legacy objects may retain an unknown (-1)
+                // logical-size sentinel; never expose that internal value in
+                // an S3 response.
+                size: Some(v.get_actual_size_or_physical()),
                e_tag: v.etag.clone().map(|etag| to_s3s_etag(&etag)),
                storage_class: v.storage_class.clone().map(ObjectStorageClass::from),
                ..Default::default()
@@ -656,6 +659,45 @@ mod tests {
        assert_eq!(output.common_prefixes.as_ref().map(std::vec::Vec::len), Some(2));
    }

+    #[test]
+    fn list_objects_never_exposes_compressed_unknown_size_sentinel() {
+        let mut metadata = std::collections::HashMap::new();
+        rustfs_utils::http::insert_str(
+            &mut metadata,
+            rustfs_utils::http::SUFFIX_COMPRESSION,
+            "klauspost/compress/s2".to_string(),
+        );
+        let output = build_list_objects_v2_output(
+            ListObjectsV2Info {
+                objects: vec![ObjectInfo {
+                    name: "legacy-compressed".to_string(),
+                    size: 128,
+                    actual_size: -1,
+                    user_defined: std::sync::Arc::new(metadata),
+                    ..Default::default()
+                }],
+                ..Default::default()
+            },
+            false,
+            1000,
+            "bucket".to_string(),
+            String::new(),
+            None,
+            None,
+            None,
+            None,
+        );
+
+        assert_eq!(
+            output
+                .contents
+                .as_ref()
+                .and_then(|objects| objects.first())
+                .and_then(|object| object.size),
+            Some(128)
+        );
+    }
+
    #[test]
    fn list_responses_report_standard_for_legacy_label_only_file_metadata() {
        let version_id = Uuid::parse_str("11111111-2222-3333-4444-555555555555").expect("fixture version ID should be valid");
@@ -429,6 +429,8 @@ pub(crate) mod ecstore_config {
 }

 pub(crate) mod ecstore_data_usage {
+    #[cfg(test)]
+    pub(crate) use rustfs_ecstore::api::data_usage::get_bucket_usage_memory;
    pub(crate) use rustfs_ecstore::api::data_usage::{
        apply_bucket_usage_memory_overlay, init_compression_total_memory_from_backend, load_admin_data_usage_from_backend_cached,
        load_data_usage_from_backend, quota_object_size, record_bucket_delete_marker_memory, record_bucket_object_delete_memory,
@@ -0,0 +1,951 @@
+// 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.
+
+//! Offline enrollment conformance against the frozen Connect fixtures.
+//!
+//! The device half of the air-gapped exchange verifies a challenge Connect
+//! signed and produces a response Connect will verify. Neither side can talk to
+//! the other while it does so, which means every disagreement about encoding,
+//! trust, or clock windows surfaces as a failed enrollment in the field rather
+//! than as an error at development time. The fixtures under
+//! `protocol/agent/v1/fixtures/offline-enrollment/` are the shared statement of
+//! what both sides must do, so this suite replays them rather than restating
+//! them: accept vectors must be accepted with the fields the document carries,
+//! reject vectors must fail with the single reason `error-codes.json` freezes,
+//! and the signature encoding rules in `trust-model.json` must hold even where
+//! the underlying ECDSA library is happy.
+
+use std::fs;
+use std::path::PathBuf;
+
+use base64::Engine as _;
+use base64::engine::general_purpose::STANDARD as BASE64_STANDARD;
+use base64::engine::general_purpose::URL_SAFE_NO_PAD as BASE64_URL_NO_PAD;
+use rustfs::connect::identity::DeviceIdentity;
+use rustfs::connect::offline::{EnrollmentError, OfflineEnrollment, VerifiedChallenge};
+use serde_json::Value;
+use sha2::{Digest as _, Sha256};
+
+/// DER prefix of a P-256 `SubjectPublicKeyInfo`, frozen by
+/// `trust-model.json` as `signature.subjectPublicKeyInfoDerPrefix`. The 65
+/// octet uncompressed point follows it, so a SEC1 point published in a fixture
+/// becomes a decodable public key by concatenation.
+const SPKI_PREFIX_HEX: &str = "3059301306072a8648ce3d020106082a8648ce3d030107034200";
+
+/// `clockSkew.toleranceSeconds` in `trust-model.json`.
+const SKEW_TOLERANCE_SECONDS: i64 = 300;
+
+// ---------------------------------------------------------------------------
+// Fixture access
+// ---------------------------------------------------------------------------
+
+fn fixture_dir() -> PathBuf {
+    PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("../protocol/agent/v1/fixtures/offline-enrollment")
+}
+
+fn sha256_hex(bytes: &[u8]) -> String {
+    Sha256::digest(bytes).iter().map(|byte| format!("{byte:02x}")).collect()
+}
+
+/// Read one fixture file and refuse it unless its bytes match the digest
+/// `MANIFEST.sha256` freezes.
+///
+/// Every vector in this suite arrives through here. A fixture edited on this
+/// side therefore fails the tests that depend on it instead of quietly
+/// redefining what conformance means, which is the failure mode a
+/// fixture-driven suite is otherwise blind to.
+fn read_fixture(name: &str) -> Vec<u8> {
+    let dir = fixture_dir();
+    let manifest = fs::read_to_string(dir.join("MANIFEST.sha256")).expect("read MANIFEST.sha256");
+
+    let expected = manifest
+        .lines()
+        .filter(|line| !line.trim().is_empty())
+        .find_map(|line| {
+            let (digest, file) = line
+                .split_once("  ")
+                .unwrap_or_else(|| panic!("malformed manifest line: {line}"));
+            (file == name).then(|| digest.to_string())
+        })
+        .unwrap_or_else(|| panic!("{name} is not listed in MANIFEST.sha256"));
+
+    let bytes = fs::read(dir.join(name)).unwrap_or_else(|error| panic!("read {name}: {error}"));
+    assert_eq!(sha256_hex(&bytes), expected, "{name} does not match the digest MANIFEST.sha256 freezes");
+    bytes
+}
+
+fn fixture_json(name: &str) -> Value {
+    serde_json::from_slice(&read_fixture(name)).unwrap_or_else(|error| panic!("{name} parses: {error}"))
+}
+
+fn accept_vectors() -> Value {
+    fixture_json("accept-vectors.json")
+}
+
+fn reject_vectors() -> Value {
+    fixture_json("reject-vectors.json")
+}
+
+fn trust_model() -> Value {
+    fixture_json("trust-model.json")
+}
+
+fn vector_list(fixture: &Value) -> Vec<Value> {
+    fixture["vectors"].as_array().expect("fixture carries a vector list").clone()
+}
+
+fn field<'a>(value: &'a Value, key: &str) -> &'a str {
+    value[key]
+        .as_str()
+        .unwrap_or_else(|| panic!("expected a string at '{key}' in {value}"))
+}
+
+/// The octets an operator carries in on removable media.
+///
+/// The fixture's `document` object *is* the transmitted artifact: a padded
+/// base64 `bytes` field holding the raw signed octets, plus the detached
+/// signature over them. Only `bytes` is covered by the signature, so
+/// re-serialising the surrounding envelope here cannot change what a verifier
+/// checks.
+fn envelope(document: &Value) -> Vec<u8> {
+    serde_json::to_vec(document).expect("envelope serialises")
+}
+
+/// The raw octets the signature covers, exactly as transmitted.
+fn signed_octets(document: &Value) -> Vec<u8> {
+    BASE64_STANDARD
+        .decode(field(document, "bytes"))
+        .expect("document bytes are padded base64")
+}
+
+/// The parsed signed document. Parsing is a convenience for the assertions
+/// below; the implementation under test is required to verify before it parses.
+fn signed_document(document: &Value) -> Value {
+    serde_json::from_slice(&signed_octets(document)).expect("signed document parses")
+}
+
+fn unix(rfc3339: &str) -> i64 {
+    chrono::DateTime::parse_from_rfc3339(rfc3339)
+        .unwrap_or_else(|error| panic!("'{rfc3339}' is not RFC 3339: {error}"))
+        .timestamp()
+}
+
+fn hex_to_bytes(hex: &str) -> Vec<u8> {
+    (0..hex.len())
+        .step_by(2)
+        .map(|i| u8::from_str_radix(&hex[i..i + 2], 16).expect("valid hex"))
+        .collect()
+}
+
+/// Turn a fixture's unpadded-base64url SEC1 point into a usable verifying key.
+fn verifying_key(sec1_base64url: &str) -> p256::ecdsa::VerifyingKey {
+    let point = BASE64_URL_NO_PAD.decode(sec1_base64url).expect("public key is base64url");
+    assert_eq!(point.len(), 65, "the protocol freezes a 65 octet uncompressed SEC1 point");
+
+    let mut der = hex_to_bytes(SPKI_PREFIX_HEX);
+    der.extend_from_slice(&point);
+    <p256::ecdsa::VerifyingKey as p256::pkcs8::DecodePublicKey>::from_public_key_der(&der).expect("public key decodes")
+}
+
+fn published_key(role_or_name: &str) -> Value {
+    fixture_json("trust-chain.json")["keys"]
+        .as_array()
+        .expect("trust chain publishes keys")
+        .iter()
+        .find(|key| field(key, "name") == role_or_name)
+        .unwrap_or_else(|| panic!("trust-chain.json publishes no key named '{role_or_name}'"))
+        .clone()
+}
+
+/// `signatureInput = domainSeparationTag || 0x00 || the received octets`, the
+/// rule `trust-model.json` freezes under `domainSeparation`.
+fn signing_input(artifact_tag: &str, received: &[u8]) -> Vec<u8> {
+    let mut input = artifact_tag.as_bytes().to_vec();
+    input.push(0x00);
+    input.extend_from_slice(received);
+    input
+}
+
+fn domain_tag(artifact: &str) -> String {
+    let model = trust_model();
+    assert_eq!(
+        field(&model["domainSeparation"], "separatorByte"),
+        "0x00",
+        "the separator byte this suite encodes is the one the trust model freezes"
+    );
+    field(&model["domainSeparation"]["tags"], artifact).to_string()
+}
+
+/// Locate an accept vector by the name other vectors reference it by.
+fn accept_vector_named(name: &str) -> Value {
+    vector_list(&accept_vectors())
+        .into_iter()
+        .find(|vector| field(vector, "name") == name)
+        .unwrap_or_else(|| panic!("accept-vectors.json carries no vector named '{name}'"))
+}
+
+/// Verify the challenge a response vector answers, at that challenge's own
+/// evaluation time.
+fn answered_challenge(response_vector: &Value) -> (Value, VerifiedChallenge) {
+    let challenge_vector = accept_vector_named(field(response_vector, "answersChallenge"));
+    let now = unix(field(&challenge_vector, "evaluationTime"));
+    let verified = OfflineEnrollment::verify_challenge(&envelope(&challenge_vector["document"]), now)
+        .expect("the answered challenge is an accept vector and must verify");
+    (challenge_vector, verified)
+}
+
+fn device_nonce_of(document: &Value) -> [u8; 32] {
+    let raw = BASE64_URL_NO_PAD
+        .decode(field(&signed_document(document), "deviceNonce"))
+        .expect("deviceNonce is base64url");
+    raw.try_into().expect("replay.nonceLengthBytes freezes a 32 octet nonce")
+}
+
+// ---------------------------------------------------------------------------
+// Accept vectors
+// ---------------------------------------------------------------------------
+
+/// Every challenge accept vector must verify at its own evaluation time and
+/// expose exactly what the signed document says.
+///
+/// Two of these vectors sit on the skew boundary — 120 seconds before
+/// `issuedAt` and 300 seconds after `expiresAt` — so a verifier that compares
+/// against the raw window instead of the tolerated one fails here rather than
+/// in an air-gapped data centre. `challenge_proof` is pinned to the challenge's
+/// own detached signature value because that is what the response has to echo;
+/// deriving it from anything else would silently break the binding.
+#[test]
+fn every_challenge_accept_vector_verifies_and_exposes_the_signed_fields() {
+    let mut verified_count = 0usize;
+
+    for vector in vector_list(&accept_vectors()) {
+        if field(&vector, "artifact") != "challenge" {
+            continue;
+        }
+
+        let name = field(&vector, "name");
+        let document = &vector["document"];
+        let now = unix(field(&vector, "evaluationTime"));
+
+        let verified = OfflineEnrollment::verify_challenge(&envelope(document), now)
+            .unwrap_or_else(|error| panic!("accept vector '{name}' must verify: {}", error.reason()));
+
+        let signed = signed_document(document);
+        assert_eq!(verified.challenge_id, field(&signed, "challengeId"), "vector '{name}' challengeId");
+        assert_eq!(
+            verified.organization_name,
+            field(&signed, "organizationName"),
+            "vector '{name}' organizationName"
+        );
+        assert_eq!(verified.cluster_name, field(&signed, "clusterName"), "vector '{name}' clusterName");
+        assert_eq!(verified.nonce, field(&signed, "nonce"), "vector '{name}' nonce");
+        assert_eq!(verified.issued_at, field(&signed, "issuedAt"), "vector '{name}' issuedAt");
+        assert_eq!(verified.expires_at, field(&signed, "expiresAt"), "vector '{name}' expiresAt");
+        assert_eq!(verified.connect_key_id, field(&signed, "connectKeyId"), "vector '{name}' connectKeyId");
+        assert_eq!(
+            verified.challenge_proof,
+            field(&document["signature"], "value"),
+            "vector '{name}' must carry the challenge's own signature as the proof a response echoes"
+        );
+
+        verified_count += 1;
+    }
+
+    assert_eq!(
+        verified_count, 3,
+        "accept-vectors.json publishes three challenge vectors; a fourth is a protocol change"
+    );
+}
+
+/// Connect's own producer wrote the response accept vectors. Rebuilding them
+/// from the challenge they answer, with the device nonce and production time
+/// they used, must reproduce every field that does not depend on which device
+/// key signed — including the discarded-unknown-field vector, whose extra
+/// `telemetryHint` must not survive into anything this side produces.
+#[test]
+fn response_accept_vectors_are_reproduced_field_for_field_by_build_response() {
+    let key = DeviceIdentity::generate();
+    let mut reproduced = 0usize;
+
+    for vector in vector_list(&accept_vectors()) {
+        if field(&vector, "artifact") != "response" {
+            continue;
+        }
+
+        let name = field(&vector, "name");
+        let published = signed_document(&vector["document"]);
+        let (_, challenge) = answered_challenge(&vector);
+
+        let produced_at = unix(field(&published, "producedAt"));
+        let nonce = device_nonce_of(&vector["document"]);
+
+        let built_envelope: Value = serde_json::from_slice(
+            &OfflineEnrollment::build_response(&challenge, &key, &nonce, produced_at)
+                .unwrap_or_else(|error| panic!("vector '{name}' must be reproducible: {}", error.reason())),
+        )
+        .expect("the built response is JSON");
+        let built = signed_document(&built_envelope);
+
+        for shared in [
+            "formatVersion",
+            "protocolVersion",
+            "challengeId",
+            "organizationName",
+            "clusterName",
+            "challengeNonce",
+            "challengeProof",
+            "deviceNonce",
+        ] {
+            assert_eq!(
+                field(&built, shared),
+                field(&published, shared),
+                "vector '{name}' field {shared} must match the response Connect published"
+            );
+        }
+        assert_eq!(
+            unix(field(&built, "producedAt")),
+            produced_at,
+            "vector '{name}' producedAt must be the instant it was given"
+        );
+
+        // `versioning.additive` says an unknown optional field is discarded and
+        // never echoed back; a producer that copied the challenge or a previous
+        // response wholesale would carry it forward.
+        assert!(
+            built.get("telemetryHint").is_none(),
+            "vector '{name}' must not echo an unknown optional field"
+        );
+
+        reproduced += 1;
+    }
+
+    assert_eq!(
+        reproduced, 2,
+        "accept-vectors.json publishes two response vectors; a third is a protocol change"
+    );
+}
+
+// ---------------------------------------------------------------------------
+// Reject vectors
+// ---------------------------------------------------------------------------
+
+/// Every challenge reject vector must fail, and fail for the one reason
+/// `error-codes.json` freezes.
+///
+/// Asserting only that verification failed would pass for an implementation
+/// that rejects everything, and would let a tampered document be reported as an
+/// expiry — a rejection reason is what an operator acts on, so it is part of the
+/// contract rather than a diagnostic detail.
+#[test]
+fn every_challenge_reject_vector_fails_with_its_frozen_reason() {
+    let known_reasons: Vec<String> = fixture_json("error-codes.json")["reasons"]
+        .as_array()
+        .expect("error-codes.json carries reasons")
+        .iter()
+        .map(|entry| field(entry, "reason").to_string())
+        .collect();
+
+    let mut rejected = 0usize;
+
+    for vector in vector_list(&reject_vectors()) {
+        if field(&vector, "artifact") != "challenge" {
+            continue;
+        }
+
+        let name = field(&vector, "name");
+        let expected = field(&vector["expected"], "reason");
+        assert!(
+            known_reasons.iter().any(|reason| reason == expected),
+            "vector '{name}' names reason {expected}, which error-codes.json does not freeze"
+        );
+
+        let now = unix(field(&vector, "evaluationTime"));
+        let error = OfflineEnrollment::verify_challenge(&envelope(&vector["document"]), now)
+            .expect_err(&format!("reject vector '{name}' must not verify"));
+
+        assert_eq!(error.reason(), expected, "vector '{name}' must fail as {expected}");
+        rejected += 1;
+    }
+
+    assert_eq!(
+        rejected, 8,
+        "reject-vectors.json publishes eight challenge vectors; losing one silently narrows the suite"
+    );
+}
+
+/// The response reject vectors are artifacts Connect refuses. This side never
+/// verifies a response, so the device-side statement is the stronger one: given
+/// the challenge each vector answers, `build_response` must not be capable of
+/// emitting that artifact in the first place.
+///
+/// Each arm pins the specific field a compromised or careless producer would
+/// have to get wrong, so an implementation that copied values out of the wrong
+/// place — the response's own document, an operator-supplied argument, a
+/// previous exchange — fails here.
+#[test]
+fn response_reject_vectors_are_artifacts_build_response_cannot_emit() {
+    let key = DeviceIdentity::generate();
+    let mut covered = 0usize;
+
+    for vector in vector_list(&reject_vectors()) {
+        if field(&vector, "artifact") != "response" {
+            continue;
+        }
+
+        let name = field(&vector, "name");
+        let refused = signed_document(&vector["document"]);
+        let (_, challenge) = answered_challenge(&vector);
+        let produced_at = unix(field(&refused, "producedAt"));
+        let nonce = device_nonce_of(&vector["document"]);
+
+        let outcome = OfflineEnrollment::build_response(&challenge, &key, &nonce, produced_at);
+
+        match field(&vector["expected"], "reason") {
+            // `responseWindow` in trust-model.json: a device that emits a
+            // response outside the tolerated challenge window has produced an
+            // artifact Connect will refuse, so the refusal belongs here rather
+            // than at the far end of a courier run.
+            "CHALLENGE_EXPIRED" => {
+                let error = outcome.expect_err(&format!("vector '{name}': producing this response must be refused"));
+                assert_eq!(error.reason(), "CHALLENGE_EXPIRED", "vector '{name}' must refuse as CHALLENGE_EXPIRED");
+                covered += 1;
+                continue;
+            }
+            reason => {
+                let built_envelope: Value = serde_json::from_slice(
+                    &outcome.unwrap_or_else(|error| panic!("vector '{name}' baseline must build: {}", error.reason())),
+                )
+                .expect("the built response is JSON");
+                let built = signed_document(&built_envelope);
+
+                match reason {
+                    "ORGANIZATION_MISMATCH" => {
+                        assert_ne!(
+                            field(&refused, "organizationName"),
+                            challenge.organization_name,
+                            "vector '{name}' is only a mismatch if it names another organization"
+                        );
+                        assert_eq!(
+                            field(&built, "organizationName"),
+                            challenge.organization_name,
+                            "vector '{name}': the organization must come from the challenge, never from elsewhere"
+                        );
+                    }
+                    "CLUSTER_MISMATCH" => {
+                        assert_ne!(
+                            field(&refused, "clusterName"),
+                            challenge.cluster_name,
+                            "vector '{name}' is only a mismatch if it names another cluster"
+                        );
+                        assert_eq!(
+                            field(&built, "clusterName"),
+                            challenge.cluster_name,
+                            "vector '{name}': the cluster must come from the challenge, never from elsewhere"
+                        );
+                    }
+                    "CHALLENGE_PROOF_INVALID" => {
+                        // Two distinct vectors land here: a nonce the challenge
+                        // never carried, and a proof lifted from another
+                        // challenge. Both must be impossible to produce.
+                        assert_eq!(
+                            field(&built, "challengeNonce"),
+                            challenge.nonce,
+                            "vector '{name}': the echoed nonce must be the challenge's own"
+                        );
+                        assert_eq!(
+                            field(&built, "challengeProof"),
+                            challenge.challenge_proof,
+                            "vector '{name}': the proof must be the answered challenge's signature"
+                        );
+                        assert!(
+                            field(&refused, "challengeNonce") != challenge.nonce
+                                || field(&refused, "challengeProof") != challenge.challenge_proof,
+                            "vector '{name}' must differ from the challenge in nonce or proof to be rejectable"
+                        );
+                    }
+                    "DEVICE_PROOF_INVALID" => {
+                        // The refused vector presents one key and is signed by
+                        // another; hold the fixture to that claim, then require
+                        // the built response to be the opposite. Proof of
+                        // possession is the only thing that makes presenting a
+                        // key in an unauthenticated document safe.
+                        use p256::ecdsa::signature::Verifier as _;
+
+                        let presented = verifying_key(field(&refused, "devicePublicKey"));
+                        let raw = BASE64_URL_NO_PAD
+                            .decode(field(&vector["document"]["signature"], "value"))
+                            .expect("signature is base64url");
+                        let signature = p256::ecdsa::Signature::from_slice(&raw).expect("signature parses");
+                        assert!(
+                            presented
+                                .verify(
+                                    &signing_input(&domain_tag("enrollmentResponse"), &signed_octets(&vector["document"])),
+                                    &signature
+                                )
+                                .is_err(),
+                            "vector '{name}' is only a possession failure if it does not verify under the key it presents"
+                        );
+
+                        assert_response_proves_possession(&built_envelope, name);
+                    }
+                    "UNSUPPORTED_FORMAT" => {
+                        assert_ne!(
+                            field(&refused, "formatVersion"),
+                            field(&built, "formatVersion"),
+                            "vector '{name}' is only unsupported if it names another format version"
+                        );
+                        assert_eq!(
+                            field(&built, "formatVersion"),
+                            "rustfs.connect.offline.enrollmentResponse/1",
+                            "vector '{name}': the format version is frozen"
+                        );
+                    }
+                    "UNSUPPORTED_PROTOCOL" => {
+                        assert_ne!(
+                            field(&refused, "protocolVersion"),
+                            field(&built, "protocolVersion"),
+                            "vector '{name}' is only unsupported if it names another protocol major"
+                        );
+                        assert_eq!(field(&built, "protocolVersion"), "v1", "vector '{name}': the protocol major is frozen");
+                    }
+                    "ENROLLMENT_REPLAYED" => {
+                        // The vector claims to be a byte-identical replay of an
+                        // accepted response; hold it to that, because a replay
+                        // vector that is not byte identical proves nothing about
+                        // single use.
+                        let accepted = accept_vector_named("response binding the device public key and the challenge proof");
+                        assert_eq!(
+                            signed_octets(&vector["document"]),
+                            signed_octets(&accepted["document"]),
+                            "vector '{name}' must be the accepted response octet for octet"
+                        );
+                        assert_eq!(
+                            field(&vector["document"]["signature"], "value"),
+                            field(&accepted["document"]["signature"], "value"),
+                            "vector '{name}' must carry the accepted response's signature"
+                        );
+
+                        // A fresh device nonce is a different artifact, so a
+                        // second enrollment is never mistaken for a replay of
+                        // the first.
+                        let other = OfflineEnrollment::build_response(&challenge, &key, &[0x5a; 32], produced_at)
+                            .expect("a second response builds");
+                        assert_ne!(
+                            signed_octets(&built_envelope),
+                            signed_octets(&serde_json::from_slice::<Value>(&other).expect("JSON")),
+                            "vector '{name}': a different device nonce must yield a different artifact"
+                        );
+                    }
+                    other => panic!("vector '{name}' names an unhandled reason {other}; extend this test"),
+                }
+            }
+        }
+
+        covered += 1;
+    }
+
+    assert_eq!(
+        covered, 9,
+        "reject-vectors.json publishes nine response vectors; losing one silently narrows the suite"
+    );
+}
+
+// ---------------------------------------------------------------------------
+// Signature encoding
+// ---------------------------------------------------------------------------
+
+/// The high-S malleation is the rejection the whole encoding rule exists for.
+///
+/// `(r, n - s)` is a second valid signature over the same document under the
+/// same key. Every mainstream ECDSA library verifies it, so an implementation
+/// that hands the decoded octets straight to `p256` accepts a forged-looking
+/// duplicate of a genuine challenge — and because the 64 octets differ, that
+/// duplicate is a distinct artifact identity that slips past any deduplication
+/// keyed on the signature. This test proves the rejection came from the
+/// encoding rule and not from a failed verification: it first shows the
+/// malleated signature verifying mathematically, then requires
+/// `verify_challenge` to refuse it as SIGNATURE_NOT_CANONICAL.
+#[test]
+fn malleated_high_s_signature_is_refused_although_it_verifies_mathematically() {
+    use p256::ecdsa::signature::Verifier as _;
+
+    let model = trust_model();
+    let malleated = model["rejectedSignatureEncodings"]
+        .as_array()
+        .expect("trust-model.json publishes rejected encodings")
+        .iter()
+        .find(|entry| field(entry, "reason") == "SIGNATURE_NOT_CANONICAL")
+        .expect("trust-model.json publishes the high-S malleation")
+        .clone();
+    assert!(
+        malleated["acceptedByALenientVerifier"].as_bool() == Some(true),
+        "this vector is only interesting because a lenient verifier accepts it"
+    );
+
+    let vector = accept_vector_named("challenge signed by a chained signing key under the pinned root");
+    let genuine_value = field(&vector["document"]["signature"], "value").to_string();
+    let malleated_value = field(&malleated, "value").to_string();
+    assert_ne!(genuine_value, malleated_value, "the malleation must be a different encoding");
+
+    let genuine = BASE64_URL_NO_PAD.decode(&genuine_value).expect("signature is base64url");
+    let raw = BASE64_URL_NO_PAD.decode(&malleated_value).expect("signature is base64url");
+    assert_eq!(raw.len(), 64, "the malleation is well formed at 64 octets");
+    assert_eq!(raw[..32], genuine[..32], "the malleation shares r with the genuine signature");
+    assert_ne!(raw[32..], genuine[32..], "the malleation replaces s with n - s");
+
+    // Step one: the malleated pair really does verify under the signing key, so
+    // a verifier cannot be excused for accepting it on mathematical grounds.
+    let signature = p256::ecdsa::Signature::from_slice(&raw).expect("the malleated signature parses");
+    assert!(signature.normalize_s().is_some(), "the malleated signature must be the high-S form");
+    let key = verifying_key(field(&published_key("signing"), "publicKey"));
+    let input = signing_input(&domain_tag("enrollmentChallenge"), &signed_octets(&vector["document"]));
+    key.verify(&input, &signature)
+        .expect("the malleated signature must verify mathematically, or this test proves nothing");
+
+    // Step two: the implementation must refuse it anyway, and say why.
+    let mut tampered = vector["document"].clone();
+    tampered["signature"]["value"] = Value::String(malleated_value);
+
+    let now = unix(field(&vector, "evaluationTime"));
+    let error = OfflineEnrollment::verify_challenge(&envelope(&tampered), now)
+        .expect_err("a high-S signature must be refused even though it verifies");
+    assert_eq!(
+        error.reason(),
+        "SIGNATURE_NOT_CANONICAL",
+        "a malleated signature is a canonicality failure, not a verification failure"
+    );
+}
+
+/// Every encoding `trust-model.json` names as rejected must fail with the
+/// reason it names — DER, padded base64url, truncation, and out-of-range
+/// scalars alongside the malleation. Three of the five are accepted by a
+/// lenient verifier, so a single blanket "signature did not verify" answer would
+/// be both wrong and undiagnosable.
+#[test]
+fn every_rejected_signature_encoding_fails_with_its_frozen_reason() {
+    let vector = accept_vector_named("challenge signed by a chained signing key under the pinned root");
+    let now = unix(field(&vector, "evaluationTime"));
+    let model = trust_model();
+    let encodings = model["rejectedSignatureEncodings"]
+        .as_array()
+        .expect("trust-model.json publishes rejected encodings");
+
+    for entry in encodings {
+        let name = field(entry, "name");
+        let mut tampered = vector["document"].clone();
+        tampered["signature"]["value"] = Value::String(field(entry, "value").to_string());
+
+        let error: EnrollmentError = OfflineEnrollment::verify_challenge(&envelope(&tampered), now)
+            .err()
+            .unwrap_or_else(|| panic!("rejected encoding '{name}' must not verify"));
+
+        assert_eq!(error.reason(), field(entry, "reason"), "rejected encoding '{name}'");
+    }
+
+    assert_eq!(encodings.len(), 5, "trust-model.json freezes five rejected encodings");
+}
+
+// ---------------------------------------------------------------------------
+// Clock window
+// ---------------------------------------------------------------------------
+
+/// The tolerated window is `[issuedAt - 300, expiresAt + 300]`, inclusive at
+/// both ends. An air-gapped device has no synchronised clock, so an
+/// off-by-one here either strands a legitimate enrollment or widens the window
+/// a stolen challenge stays usable in. Both ends are checked at the exact bound
+/// and one second past it, and the reason distinguishes the two directions.
+#[test]
+fn challenge_is_accepted_at_the_exact_skew_bound_and_refused_one_second_past_it() {
+    let vector = accept_vector_named("challenge signed by a chained signing key under the pinned root");
+    let document = envelope(&vector["document"]);
+    let signed = signed_document(&vector["document"]);
+
+    let issued_at = unix(field(&signed, "issuedAt"));
+    let expires_at = unix(field(&signed, "expiresAt"));
+
+    let earliest = issued_at - SKEW_TOLERANCE_SECONDS;
+    OfflineEnrollment::verify_challenge(&document, earliest).expect("the earliest tolerated instant is inside the window");
+    let error =
+        OfflineEnrollment::verify_challenge(&document, earliest - 1).expect_err("one second earlier is outside the window");
+    assert_eq!(error.reason(), "CHALLENGE_NOT_YET_VALID");
+
+    let latest = expires_at + SKEW_TOLERANCE_SECONDS;
+    OfflineEnrollment::verify_challenge(&document, latest).expect("the latest tolerated instant is inside the window");
+    let error = OfflineEnrollment::verify_challenge(&document, latest + 1).expect_err("one second later is outside the window");
+    assert_eq!(error.reason(), "CHALLENGE_EXPIRED");
+}
+
+// ---------------------------------------------------------------------------
+// Response production
+// ---------------------------------------------------------------------------
+
+/// Assert a built response proves possession of the key it presents: the
+/// fingerprint matches the presented key, and the detached signature is a
+/// canonical low-S ES256 signature that verifies under that key over the exact
+/// octets transmitted.
+fn assert_response_proves_possession(built_envelope: &Value, label: &str) {
+    use p256::ecdsa::signature::Verifier as _;
+
+    let raw = signed_octets(built_envelope);
+    let built = signed_document(built_envelope);
+    let signature_block = &built_envelope["signature"];
+
+    assert_eq!(field(signature_block, "algorithm"), "ES256", "{label}: the algorithm is frozen");
+
+    let value = field(signature_block, "value");
+    assert_eq!(value.len(), 86, "{label}: the transfer encoding is 86 unpadded base64url characters");
+    assert!(
+        value.bytes().all(|b| b.is_ascii_alphanumeric() || b == b'-' || b == b'_'),
+        "{label}: the signature must use the base64url alphabet with no padding"
+    );
+
+    let bytes = BASE64_URL_NO_PAD.decode(value).expect("signature is base64url");
+    assert_eq!(bytes.len(), 64, "{label}: the signature is a fixed-width r || s");
+    let signature = p256::ecdsa::Signature::from_slice(&bytes).expect("signature parses");
+    assert!(
+        signature.normalize_s().is_none(),
+        "{label}: this side must never emit the malleated high-S form it refuses to accept"
+    );
+
+    let presented = field(&built, "devicePublicKey");
+    let key = verifying_key(presented);
+    key.verify(&signing_input(&domain_tag("enrollmentResponse"), &raw), &signature)
+        .unwrap_or_else(|error| panic!("{label}: the response must verify under the key it presents: {error}"));
+
+    // `signature.keyIdAlgorithm`: the lowercase SHA-256 of the DER
+    // SubjectPublicKeyInfo, not of the bare point and not of the transfer
+    // encoding.
+    let mut spki = hex_to_bytes(SPKI_PREFIX_HEX);
+    spki.extend_from_slice(&BASE64_URL_NO_PAD.decode(presented).expect("public key is base64url"));
+    let fingerprint = sha256_hex(&spki);
+    assert_eq!(
+        field(&built, "deviceKeyId"),
+        fingerprint,
+        "{label}: deviceKeyId must be the fingerprint of the key the document presents"
+    );
+    assert_eq!(
+        field(signature_block, "keyId"),
+        fingerprint,
+        "{label}: the detached signature must name the same key"
+    );
+}
+
+/// A response is the only thing Connect will ever see from this device, so it
+/// has to carry the whole binding on its own: the challenge it answers, the
+/// proof that challenge was genuine, the key being enrolled, and possession of
+/// that key.
+#[test]
+fn built_response_binds_the_challenge_proof_and_proves_possession_of_the_device_key() {
+    let vector = accept_vector_named("response binding the device public key and the challenge proof");
+    let (challenge_vector, challenge) = answered_challenge(&vector);
+    let key = DeviceIdentity::generate();
+    let produced_at = unix(field(&signed_document(&vector["document"]), "producedAt"));
+
+    let bytes = OfflineEnrollment::build_response(&challenge, &key, &[0x11; 32], produced_at).expect("the response builds");
+    let built_envelope: Value = serde_json::from_slice(&bytes).expect("the response is JSON");
+    let built = signed_document(&built_envelope);
+
+    assert_response_proves_possession(&built_envelope, "built response");
+
+    // The proof is the challenge's own detached signature. A producer that
+    // echoed the nonce alone, or hashed something, would let a response be
+    // built from an unverified challenge.
+    assert_eq!(
+        field(&built, "challengeProof"),
+        field(&challenge_vector["document"]["signature"], "value"),
+        "the proof must be the signature of the challenge being answered"
+    );
+    assert_eq!(field(&built, "challengeNonce"), challenge.nonce);
+    assert_eq!(field(&built, "challengeId"), challenge.challenge_id);
+
+    assert_eq!(
+        field(&built, "devicePublicKey"),
+        BASE64_URL_NO_PAD.encode(&key.public_key_der()[hex_to_bytes(SPKI_PREFIX_HEX).len()..]),
+        "the presented key must be the key that was passed in"
+    );
+    assert_eq!(
+        field(&built, "deviceNonce"),
+        BASE64_URL_NO_PAD.encode([0x11; 32]),
+        "the device nonce must be the one that was passed in"
+    );
+    assert!(field(&built, "producedAt").ends_with('Z'), "producedAt is a UTC RFC 3339 instant");
+}
+
+/// The response leaves the air gap on removable media and is read by anyone who
+/// handles it. A producer that serialised the key pair instead of the public
+/// key, or logged a debug rendering into the document, would put the enrolled
+/// private key on that medium — and the enrollment would still succeed, so
+/// nothing else in this suite would notice.
+#[test]
+fn built_response_carries_no_private_key_material() {
+    let vector = accept_vector_named("response binding the device public key and the challenge proof");
+    let (_, challenge) = answered_challenge(&vector);
+    let key = DeviceIdentity::generate();
+    let produced_at = unix(field(&signed_document(&vector["document"]), "producedAt"));
+
+    let response = OfflineEnrollment::build_response(&challenge, &key, &[0x22; 32], produced_at).expect("the response builds");
+
+    // The envelope carries the signed document base64-encoded, so a needle
+    // present in the document is not present in the envelope octets. Both
+    // layers are searched: an operator handling the medium can read either.
+    let envelope_value: Value = serde_json::from_slice(&response).expect("the response is JSON");
+    let mut haystack = response;
+    haystack.extend_from_slice(&signed_octets(&envelope_value));
+
+    let pkcs8 = key.to_pkcs8_der().expect("serialise the key");
+    let secret = <p256::SecretKey as p256::pkcs8::DecodePrivateKey>::from_pkcs8_der(&pkcs8).expect("the key parses");
+    let scalar = secret.to_bytes();
+
+    // Every spelling the scalar could plausibly reach a document in: raw, and
+    // the three encodings this protocol already uses elsewhere.
+    let scalar_hex: String = scalar.iter().map(|byte| format!("{byte:02x}")).collect();
+    for (description, needle) in [
+        ("the PKCS#8 encoding", pkcs8.to_vec()),
+        ("the raw private scalar", scalar.to_vec()),
+        ("the scalar in base64url", BASE64_URL_NO_PAD.encode(scalar).into_bytes()),
+        ("the scalar in standard base64", BASE64_STANDARD.encode(scalar).into_bytes()),
+        ("the scalar in hex", scalar_hex.into_bytes()),
+    ] {
+        assert!(
+            !haystack.windows(needle.len()).any(|window| window == needle.as_slice()),
+            "the response must not contain {description}"
+        );
+    }
+
+    // The public half must be there, so the absence above is a statement about
+    // what was excluded rather than about a haystack that would not have found
+    // the private half either.
+    let point = BASE64_URL_NO_PAD.encode(&key.public_key_der()[hex_to_bytes(SPKI_PREFIX_HEX).len()..]);
+    assert!(
+        haystack.windows(point.len()).any(|window| window == point.as_bytes()),
+        "the response must still present the public key"
+    );
+}
+
+// ---------------------------------------------------------------------------
+// The offline invariant
+// ---------------------------------------------------------------------------
+
+/// The whole surface exists because there is no network. This asserts that
+/// three different ways, because no single one of them is conclusive on its own.
+///
+/// 1. The process opens no descriptor across a full verify-and-respond cycle. A
+///    socket, a DNS resolver, a pooled HTTP client, or a revocation-list fetch
+///    all show up here — including one that is opened and cached rather than
+///    opened and closed, which is what a lazily built client does.
+/// 2. The cycle is a pure byte transform: the same inputs produce the same
+///    verified fields, and the evaluation instant is an argument rather than an
+///    ambient read, so nothing about the outcome can depend on reachability.
+/// 3. Repeating the cycle changes nothing observable, so a first call cannot be
+///    quietly initialising shared state that a later one reuses.
+#[cfg(unix)]
+#[test]
+fn enrollment_opens_no_descriptor_and_is_a_pure_byte_transform() {
+    let vector = accept_vector_named("challenge signed by a chained signing key under the pinned root");
+    let document = envelope(&vector["document"]);
+    let now = unix(field(&vector, "evaluationTime"));
+    let key = DeviceIdentity::generate();
+
+    // Warm anything the test harness itself lazily opens before the baseline.
+    let _ = open_descriptors();
+    let baseline = open_descriptors();
+    assert!(
+        !baseline.is_empty(),
+        "the descriptor table must be readable for this test to mean anything"
+    );
+
+    let mut fields = Vec::new();
+    for _ in 0..2 {
+        let challenge = OfflineEnrollment::verify_challenge(&document, now).expect("the challenge verifies");
+        let response = OfflineEnrollment::build_response(&challenge, &key, &[0x33; 32], now).expect("the response builds");
+        fields.push((
+            challenge.challenge_id.clone(),
+            challenge.nonce.clone(),
+            challenge.challenge_proof.clone(),
+            signed_octets(&serde_json::from_slice::<Value>(&response).expect("JSON")),
+        ));
+    }
+
+    assert_eq!(
+        open_descriptors(),
+        baseline,
+        "the enrollment path must not open a descriptor: no socket, no resolver, no cached client"
+    );
+
+    let (first, second) = (&fields[0], &fields[1]);
+    assert_eq!(first.0, second.0, "verification must be deterministic");
+    assert_eq!(first.1, second.1, "verification must be deterministic");
+    assert_eq!(first.2, second.2, "verification must be deterministic");
+    assert_eq!(
+        first.3, second.3,
+        "the signed response octets are a function of the challenge, the key, the nonce, and the instant"
+    );
+}
+
+#[cfg(unix)]
+fn open_descriptors() -> Vec<String> {
+    // Linux publishes the table at /proc/self/fd; the BSDs and macOS at /dev/fd.
+    let path = if PathBuf::from("/proc/self/fd").is_dir() {
+        "/proc/self/fd"
+    } else {
+        "/dev/fd"
+    };
+
+    let mut entries: Vec<String> = fs::read_dir(path)
+        .unwrap_or_else(|error| panic!("read {path}: {error}"))
+        .map(|entry| entry.expect("read dir entry").file_name().to_string_lossy().into_owned())
+        .collect();
+    entries.sort();
+    entries
+}
+
+/// A descriptor count taken around a call cannot see a socket that was opened
+/// and closed inside it, so the invariant is also asserted where it can be
+/// stated absolutely: the implementation names no network API at all.
+///
+/// This is the shape the regression actually takes — someone adds a
+/// revocation-list fetch, a time-server check, or a "just confirm the challenge
+/// with Connect" call — and it is caught at the source rather than by observing
+/// its effects.
+#[test]
+fn enrollment_implementation_names_no_network_api() {
+    let path = PathBuf::from(env!("CARGO_MANIFEST_DIR")).join("src/connect/offline/enrollment.rs");
+    let source = fs::read_to_string(&path).unwrap_or_else(|error| panic!("read {}: {error}", path.display()));
+
+    // Prose is allowed to discuss the invariant it is documenting, so only code
+    // is scanned.
+    let code: String = source
+        .lines()
+        .filter(|line| !line.trim_start().starts_with("//"))
+        .collect::<Vec<_>>()
+        .join("\n");
+
+    for forbidden in [
+        "std::net",
+        "tokio::net",
+        "TcpStream",
+        "TcpListener",
+        "UdpSocket",
+        "UnixStream",
+        "ToSocketAddrs",
+        "reqwest",
+        "hyper",
+        "tonic",
+    ] {
+        assert!(
+            !code.contains(forbidden),
+            "offline enrollment must not reach the network, but the implementation names {forbidden}"
+        );
+    }
+}
@@ -0,0 +1,911 @@
+// Copyright 2024 RustFS Team
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+use std::collections::VecDeque;
+use std::fs;
+use std::sync::{Arc, Mutex};
+use std::time::Duration;
+
+use base64::Engine as _;
+use base64::engine::general_purpose::{STANDARD as BASE64_STANDARD, URL_SAFE_NO_PAD as BASE64_URL_NO_PAD};
+use bytes::Bytes;
+use http_body_util::{BodyExt as _, Full};
+use hyper::service::service_fn;
+use hyper::{Request, Response, StatusCode};
+use hyper_util::rt::TokioIo;
+use p256::ecdsa::signature::Verifier as _;
+use p256::ecdsa::{Signature, VerifyingKey};
+use p256::pkcs8::DecodePublicKey as _;
+use rcgen::{
+    BasicConstraints, CertificateParams, DistinguishedName, DnType, ExtendedKeyUsagePurpose, IsCa, Issuer, KeyPair,
+    KeyUsagePurpose, SanType, SerialNumber,
+};
+use rustfs::connect::{ClientError, ConnectClient, ConnectConfig, CredentialStore, IdentityStore, RegistrationToken, TokenError};
+use rustls::RootCertStore;
+use rustls::pki_types::{CertificateDer, PrivateKeyDer, PrivatePkcs8KeyDer, pem::PemObject as _};
+use rustls::server::WebPkiClientVerifier;
+use serde_json::{Value, json};
+use sha2::{Digest as _, Sha256};
+use time::OffsetDateTime;
+use time::format_description::well_known::Rfc3339;
+use tokio::net::TcpListener;
+use tokio_rustls::TlsAcceptor;
+
+const ORGANIZATION_UID: &str = "0198f4b0-1a00-7c10-8d21-2e3f4a5b6c70";
+const CLUSTER_UID: &str = "0198f4b0-2b00-7d20-9e31-3f4a5b6c7d81";
+const DEVICE_UID: &str = "0198f4b0-3c00-7e30-8f41-4a5b6c7d8e92";
+const TOKEN_UID: &str = "0198f4b0-6f00-7b60-9271-7d8e9fa0b1c5";
+const FRESH_TOKEN_UID: &str = "0198f4b0-7f00-7c70-a381-8e9fa0b1c2d6";
+const SECOND_TOKEN_UID: &str = "0198f4b0-8f00-7d80-b491-9fa0b1c2d3e7";
+
+struct TestPki {
+    root_params: CertificateParams,
+    root_key: KeyPair,
+    root_der: CertificateDer<'static>,
+    root_pem: String,
+    server_der: CertificateDer<'static>,
+    server_key: PrivatePkcs8KeyDer<'static>,
+}
+
+impl TestPki {
+    fn new() -> Self {
+        let now = OffsetDateTime::now_utc();
+        let root_key = KeyPair::generate().expect("generate root key");
+        let mut root_params = CertificateParams::default();
+        root_params.is_ca = IsCa::Ca(BasicConstraints::Unconstrained);
+        root_params.not_before = now - time::Duration::days(1);
+        root_params.not_after = now + time::Duration::days(30);
+        root_params.key_usages = vec![KeyUsagePurpose::KeyCertSign, KeyUsagePurpose::DigitalSignature];
+        root_params.distinguished_name.push(DnType::CommonName, "Connect test root");
+        let root = root_params.self_signed(&root_key).expect("sign root");
+
+        let server_key = KeyPair::generate().expect("generate server key");
+        let mut server_params = CertificateParams::default();
+        server_params.not_before = now - time::Duration::hours(1);
+        server_params.not_after = now + time::Duration::days(2);
+        server_params
+            .subject_alt_names
+            .push(SanType::DnsName("localhost".try_into().expect("valid DNS name")));
+        server_params.extended_key_usages = vec![ExtendedKeyUsagePurpose::ServerAuth];
+        let issuer = Issuer::from_params(&root_params, &root_key);
+        let server = server_params
+            .signed_by(&server_key, &issuer)
+            .expect("sign server certificate");
+
+        Self {
+            root_params,
+            root_key,
+            root_der: root.der().clone(),
+            root_pem: root.pem(),
+            server_der: server.der().clone(),
+            server_key: PrivatePkcs8KeyDer::from(server_key.serialize_der()),
+        }
+    }
+
+    fn credential(&self, identity: &rustfs::connect::DeviceIdentity, uri: &str, serial_byte: u8) -> Value {
+        let now = OffsetDateTime::now_utc().replace_nanosecond(0).expect("whole second");
+        self.credential_window(identity, uri, serial_byte, now, now + time::Duration::days(1))
+    }
+
+    fn credential_window(
+        &self,
+        identity: &rustfs::connect::DeviceIdentity,
+        uri: &str,
+        serial_byte: u8,
+        not_before: OffsetDateTime,
+        not_after: OffsetDateTime,
+    ) -> Value {
+        let mut params = CertificateParams::default();
+        params.not_before = not_before;
+        params.not_after = not_after;
+        params.serial_number = Some(SerialNumber::from(vec![serial_byte; 16]));
+        params.key_usages = vec![KeyUsagePurpose::DigitalSignature];
+        params.extended_key_usages = vec![ExtendedKeyUsagePurpose::ClientAuth];
+        params.distinguished_name = DistinguishedName::new();
+        params.distinguished_name.push(DnType::CommonName, DEVICE_UID);
+        params
+            .subject_alt_names
+            .push(SanType::URI(uri.try_into().expect("valid URI SAN")));
+        let private_key = identity.to_pkcs8_der().expect("serialize device key");
+        let private_key = PrivatePkcs8KeyDer::from(private_key.to_vec());
+        let device_key =
+            KeyPair::from_pkcs8_der_and_sign_algo(&private_key, &rcgen::PKCS_ECDSA_P256_SHA256).expect("parse device key");
+        let issuer = Issuer::from_params(&self.root_params, &self.root_key);
+        let certificate = params.signed_by(&device_key, &issuer).expect("sign device certificate");
+        let serial = format!("{serial_byte:02x}").repeat(16);
+        let cluster = format!("organizations/{ORGANIZATION_UID}/clusters/{CLUSTER_UID}");
+
+        json!({
+            "name": format!("{cluster}/clusterDevices/{DEVICE_UID}"),
+            "uid": DEVICE_UID,
+            "cluster": cluster,
+            "protocolVersion": "v1",
+            "keyId": format!("x509-{serial}"),
+            "certificateSerial": serial,
+            "certificate": certificate.pem(),
+            "certificateChain": certificate.pem(),
+            "notBefore": not_before.format(&Rfc3339).expect("format notBefore"),
+            "notAfter": not_after.format(&Rfc3339).expect("format notAfter"),
+        })
+    }
+
+    fn server_config(&self, require_client: bool) -> rustls::ServerConfig {
+        let mut roots = RootCertStore::empty();
+        roots.add(self.root_der.clone()).expect("add client root");
+        let verifier = WebPkiClientVerifier::builder(Arc::new(roots));
+        let verifier = if require_client {
+            verifier.build()
+        } else {
+            verifier.allow_unauthenticated().build()
+        }
+        .expect("build client verifier");
+        rustls::ServerConfig::builder()
+            .with_client_cert_verifier(verifier)
+            .with_single_cert(vec![self.server_der.clone()], PrivateKeyDer::Pkcs8(self.server_key.clone_key()))
+            .expect("build server TLS")
+    }
+}
+
+#[derive(Clone)]
+enum Reply {
+    Json(StatusCode, Value),
+    DelayedClose(Duration),
+    VerifiedRotation {
+        response: Value,
+        current_public_key: Vec<u8>,
+        current_certificate_fingerprint: String,
+        device_name: String,
+    },
+}
+
+struct TestServer {
+    endpoint: String,
+    seen: Arc<Mutex<Vec<Value>>>,
+    task: tokio::task::JoinHandle<()>,
+}
+
+impl Drop for TestServer {
+    fn drop(&mut self) {
+        self.task.abort();
+    }
+}
+
+async fn server(pki: &TestPki, replies: Vec<Reply>) -> TestServer {
+    server_with_client_auth(pki, replies, false).await
+}
+
+async fn server_with_client_auth(pki: &TestPki, replies: Vec<Reply>, require_client: bool) -> TestServer {
+    let listener = TcpListener::bind("127.0.0.1:0").await.expect("bind test server");
+    let address = listener.local_addr().expect("server address");
+    let acceptor = TlsAcceptor::from(Arc::new(pki.server_config(require_client)));
+    let replies = Arc::new(Mutex::new(VecDeque::from(replies)));
+    let seen = Arc::new(Mutex::new(Vec::new()));
+    let captured = seen.clone();
+    let task = tokio::spawn(async move {
+        loop {
+            let Ok((stream, _)) = listener.accept().await else {
+                return;
+            };
+            let acceptor = acceptor.clone();
+            let replies = replies.clone();
+            let seen = captured.clone();
+            tokio::spawn(async move {
+                let Ok(stream) = acceptor.accept(stream).await else {
+                    return;
+                };
+                let service = service_fn(move |request: Request<hyper::body::Incoming>| {
+                    let replies = replies.clone();
+                    let seen = seen.clone();
+                    async move {
+                        let body = request.into_body().collect().await.expect("read request body").to_bytes();
+                        let value: Value = serde_json::from_slice(&body).expect("request JSON");
+                        seen.lock().expect("seen lock").push(value.clone());
+                        let reply = replies.lock().expect("reply lock").pop_front().expect("planned reply");
+                        match reply {
+                            Reply::Json(status, value) => Ok::<_, hyper::Error>(
+                                Response::builder()
+                                    .status(status)
+                                    .header("content-type", "application/json")
+                                    .body(Full::new(Bytes::from(serde_json::to_vec(&value).expect("reply JSON"))))
+                                    .expect("response"),
+                            ),
+                            Reply::DelayedClose(delay) => {
+                                tokio::time::sleep(delay).await;
+                                Ok(Response::builder()
+                                    .status(StatusCode::SERVICE_UNAVAILABLE)
+                                    .body(Full::new(Bytes::new()))
+                                    .expect("response"))
+                            }
+                            Reply::VerifiedRotation {
+                                response,
+                                current_public_key,
+                                current_certificate_fingerprint,
+                                device_name,
+                            } => {
+                                verify_rotation_request(
+                                    &value,
+                                    &current_public_key,
+                                    &current_certificate_fingerprint,
+                                    &device_name,
+                                );
+                                Ok(Response::builder()
+                                    .status(StatusCode::OK)
+                                    .header("content-type", "application/json")
+                                    .body(Full::new(Bytes::from(serde_json::to_vec(&response).expect("reply JSON"))))
+                                    .expect("response"))
+                            }
+                        }
+                    }
+                });
+                let _ = hyper::server::conn::http1::Builder::new()
+                    .serve_connection(TokioIo::new(stream), service)
+                    .await;
+            });
+        }
+    });
+    TestServer {
+        endpoint: format!("https://localhost:{}/agent/", address.port()),
+        seen,
+        task,
+    }
+}
+
+fn verify_rotation_request(request: &Value, current_public_key: &[u8], fingerprint: &str, device_name: &str) {
+    assert_eq!(request["protocolVersion"], "v1");
+    assert_eq!(request["proof"]["algorithm"], "ES256");
+    let csr = BASE64_STANDARD
+        .decode(request["certificateRequest"].as_str().expect("certificateRequest"))
+        .expect("CSR base64");
+    let csr_digest = BASE64_URL_NO_PAD.encode(Sha256::digest(&csr));
+    let request_id = request["requestId"].as_str().expect("requestId");
+    let transcript = rebuilt_rotation_transcript(
+        b"RUSTFS-CONNECT-CREDENTIAL-ROTATION-V1",
+        [fingerprint, device_name, request_id, &csr_digest],
+    );
+    let encoded = request["proof"]["value"].as_str().expect("proof value");
+    assert_eq!(encoded.len(), 86);
+    let raw = BASE64_URL_NO_PAD.decode(encoded).expect("proof base64url");
+    let signature = Signature::from_slice(&raw).expect("fixed-width signature");
+    assert!(signature.normalize_s().is_none(), "rotation proof must be low-S");
+    let verifying = VerifyingKey::from_public_key_der(current_public_key).expect("current public key");
+    verifying.verify(&transcript, &signature).expect("rotation proof verifies");
+
+    let wrong_domain = rebuilt_rotation_transcript(
+        b"RUSTFS-CONNECT-CREDENTIAL-ROTATION-V2",
+        [fingerprint, device_name, request_id, &csr_digest],
+    );
+    assert!(verifying.verify(&wrong_domain, &signature).is_err());
+    let wrong_order = rebuilt_rotation_transcript(
+        b"RUSTFS-CONNECT-CREDENTIAL-ROTATION-V1",
+        [device_name, fingerprint, request_id, &csr_digest],
+    );
+    assert!(verifying.verify(&wrong_order, &signature).is_err());
+}
+
+fn rebuilt_rotation_transcript(domain: &[u8], fields: [&str; 4]) -> Vec<u8> {
+    let mut transcript = Vec::new();
+    transcript.extend_from_slice(domain);
+    transcript.push(b'\n');
+    for field in fields {
+        transcript.extend_from_slice(field.len().to_string().as_bytes());
+        transcript.push(b':');
+        transcript.extend_from_slice(field.as_bytes());
+        transcript.push(b'\n');
+    }
+    transcript
+}
+
+fn certificate_fingerprint(pem: &str) -> String {
+    let certificate = CertificateDer::pem_slice_iter(pem.as_bytes())
+        .next()
+        .expect("leaf certificate")
+        .expect("certificate PEM");
+    Sha256::digest(certificate.as_ref())
+        .iter()
+        .map(|byte| format!("{byte:02x}"))
+        .collect()
+}
+
+fn token_document() -> Value {
+    json!({
+        "registrationTokenUid": TOKEN_UID,
+        "registrationTokenSecret": "AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA",
+        "organizationUid": ORGANIZATION_UID,
+        "clusterUid": CLUSTER_UID,
+        "challengeNonce": "a3f1c07d9b2e4856af0c1d3b5e7f9012c4a6b8d0e2f4061738495a6b7c8d9e0f",
+        "expiresUnix": OffsetDateTime::now_utc().unix_timestamp() + 3600,
+    })
+}
+
+fn token() -> RegistrationToken {
+    token_with_uid(TOKEN_UID)
+}
+
+fn token_with_uid(uid: &str) -> RegistrationToken {
+    let document = token_document();
+    let mut document = document;
+    document["registrationTokenUid"] = json!(uid);
+    RegistrationToken::from_reader(serde_json::to_vec(&document).expect("token JSON").as_slice()).expect("token parses")
+}
+
+fn stores(temp: &tempfile::TempDir) -> (IdentityStore, CredentialStore) {
+    (
+        IdentityStore::new(temp.path().join("identity")),
+        CredentialStore::new(temp.path().join("credential")),
+    )
+}
+
+fn client(server: &TestServer, pki: &TestPki, timeout: Duration) -> ConnectClient {
+    ConnectClient::new(ConnectConfig {
+        endpoint: &server.endpoint,
+        root_ca_pem: pki.root_pem.as_bytes(),
+        timeout,
+    })
+    .expect("build Connect client")
+}
+
+fn rotation_response(pki: &TestPki, identity: &rustfs::connect::DeviceIdentity, serial: u8) -> (Value, Value) {
+    let stored = pki.credential(identity, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), serial);
+    let mut wire = stored.clone();
+    wire.as_object_mut().expect("response object").remove("uid");
+    wire.as_object_mut().expect("response object").remove("cluster");
+    (wire, stored)
+}
+
+fn write_stored_credential(path: &std::path::Path, response: &Value) {
+    let not_before = OffsetDateTime::parse(response["notBefore"].as_str().expect("notBefore"), &Rfc3339)
+        .expect("parse notBefore")
+        .unix_timestamp();
+    let not_after = OffsetDateTime::parse(response["notAfter"].as_str().expect("notAfter"), &Rfc3339)
+        .expect("parse notAfter")
+        .unix_timestamp();
+    let stored = json!({
+        "name": response["name"],
+        "uid": DEVICE_UID,
+        "protocolVersion": response["protocolVersion"],
+        "keyId": response["keyId"],
+        "certificateSerial": response["certificateSerial"],
+        "certificate": response["certificate"],
+        "certificateChain": response["certificateChain"],
+        "notBeforeUnix": not_before,
+        "notAfterUnix": not_after,
+    });
+    fs::write(path, serde_json::to_vec(&stored).expect("stored credential JSON")).expect("write credential");
+    set_owner_only(path);
+}
+
+#[cfg(unix)]
+fn set_owner_only(path: &std::path::Path) {
+    use std::os::unix::fs::PermissionsExt as _;
+    fs::set_permissions(path, fs::Permissions::from_mode(0o600)).expect("set owner-only mode");
+}
+
+#[cfg(not(unix))]
+fn set_owner_only(_path: &std::path::Path) {}
+
+#[tokio::test]
+async fn registration_reuses_request_and_csr_after_timeout_and_restart() {
+    let temp = tempfile::tempdir().expect("temp dir");
+    let (identity_store, credential_store) = stores(&temp);
+    let identity = identity_store.load_or_create().expect("create identity");
+    let pki = TestPki::new();
+    let response = pki.credential(&identity, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 0x80);
+    let server = server(
+        &pki,
+        vec![
+            Reply::DelayedClose(Duration::from_millis(200)),
+            Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})),
+            Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})),
+            Reply::Json(StatusCode::CREATED, response),
+        ],
+    )
+    .await;
+    let first = client(&server, &pki, Duration::from_millis(80));
+    assert!(matches!(
+        first.register(&identity_store, &credential_store, &token()).await,
+        Err(ClientError::Unavailable { .. })
+    ));
+
+    let restarted = client(&server, &pki, Duration::from_secs(2));
+    let credential = restarted
+        .register(&identity_store, &credential_store, &token())
+        .await
+        .expect("restart replays completed exchange");
+    assert_eq!(credential.uid, DEVICE_UID);
+    assert_eq!(credential.certificate_serial, "80".repeat(16));
+
+    let seen = server.seen.lock().expect("seen lock");
+    assert_eq!(seen.len(), 4);
+    for request in &seen[1..] {
+        assert_eq!(request["requestId"], seen[0]["requestId"]);
+        assert_eq!(request["certificateRequest"], seen[0]["certificateRequest"]);
+    }
+}
+
+#[tokio::test]
+async fn registration_rejects_untrusted_or_misbound_credentials() {
+    for case in ["san", "chain", "key", "key_id", "cluster", "name"] {
+        let temp = tempfile::tempdir().expect("temp dir");
+        let (identity_store, credential_store) = stores(&temp);
+        let identity = identity_store.load_or_create().expect("create identity");
+        let pki = TestPki::new();
+        let mut response = match case {
+            "san" => pki.credential(&identity, "urn:rustfs:connect:device:0198f4b0-3c00-7e30-8f41-4a5b6c7d8e93", 2),
+            "chain" => TestPki::new().credential(&identity, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 2),
+            "key" => pki.credential(
+                &rustfs::connect::DeviceIdentity::generate(),
+                &format!("urn:rustfs:connect:device:{DEVICE_UID}"),
+                2,
+            ),
+            _ => pki.credential(&identity, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 2),
+        };
+        match case {
+            "key_id" => response["keyId"] = json!("x509-deadbeef"),
+            "cluster" => response["cluster"] = json!(format!("organizations/{ORGANIZATION_UID}/clusters/other")),
+            "name" => {
+                response["name"] = json!(format!("organizations/{ORGANIZATION_UID}/clusters/other/clusterDevices/{DEVICE_UID}"))
+            }
+            _ => {}
+        }
+        let server = server(&pki, vec![Reply::Json(StatusCode::CREATED, response)]).await;
+        let error = client(&server, &pki, Duration::from_secs(2))
+            .register(&identity_store, &credential_store, &token())
+            .await
+            .expect_err("invalid returned identity must fail closed");
+        assert!(matches!(error, ClientError::Credential(_)));
+        assert!(!temp.path().join("credential/device.crt.json").exists());
+    }
+}
+
+#[tokio::test]
+async fn stored_credential_is_revalidated_before_reuse() {
+    let temp = tempfile::tempdir().expect("temp dir");
+    let (identity_store, credential_store) = stores(&temp);
+    let identity = identity_store.load_or_create().expect("create identity");
+    let pki = TestPki::new();
+    let issued = pki.credential(&identity, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 4);
+    let server = server(&pki, vec![Reply::Json(StatusCode::CREATED, issued)]).await;
+    let client = client(&server, &pki, Duration::from_secs(2));
+    client
+        .register(&identity_store, &credential_store, &token())
+        .await
+        .expect("register");
+
+    let path = temp.path().join("credential/device.crt.json");
+    let mut stored: Value = serde_json::from_slice(&fs::read(&path).expect("read credential")).expect("credential JSON");
+    stored["certificateSerial"] = json!("00".repeat(16));
+    fs::write(&path, serde_json::to_vec(&stored).expect("credential JSON")).expect("tamper credential");
+    let error = client
+        .register(&identity_store, &credential_store, &token())
+        .await
+        .expect_err("tampered stored credential must fail closed");
+    assert!(matches!(error, ClientError::Credential(_)));
+}
+
+#[tokio::test]
+async fn register_rejects_expired_and_not_yet_valid_stored_credentials() {
+    let temp = tempfile::tempdir().expect("temp dir");
+    let (identity_store, credential_store) = stores(&temp);
+    let identity = identity_store.load_or_create().expect("create identity");
+    let pki = TestPki::new();
+    let issued = pki.credential(&identity, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 9);
+    let registration = server(&pki, vec![Reply::Json(StatusCode::CREATED, issued)]).await;
+    let client = client(&registration, &pki, Duration::from_secs(2));
+    client
+        .register(&identity_store, &credential_store, &token())
+        .await
+        .expect("register");
+
+    let now = OffsetDateTime::now_utc().replace_nanosecond(0).expect("whole second");
+    let path = temp.path().join("credential/device.crt.json");
+    let expired = pki.credential_window(
+        &identity,
+        &format!("urn:rustfs:connect:device:{DEVICE_UID}"),
+        10,
+        now - time::Duration::days(2),
+        now - time::Duration::days(1),
+    );
+    write_stored_credential(&path, &expired);
+    assert!(matches!(
+        client.register(&identity_store, &credential_store, &token()).await,
+        Err(ClientError::CredentialExpired)
+    ));
+
+    let future = pki.credential_window(
+        &identity,
+        &format!("urn:rustfs:connect:device:{DEVICE_UID}"),
+        11,
+        now + time::Duration::hours(1),
+        now + time::Duration::hours(25),
+    );
+    write_stored_credential(&path, &future);
+    assert!(matches!(
+        client.register(&identity_store, &credential_store, &token()).await,
+        Err(ClientError::CredentialNotYetValid)
+    ));
+}
+
+#[tokio::test]
+async fn concurrent_rotation_retries_converge_and_promote_the_next_key() {
+    let temp = tempfile::tempdir().expect("temp dir");
+    let (identity_store, credential_store) = stores(&temp);
+    let current = identity_store.load_or_create().expect("create identity");
+    let pki = TestPki::new();
+    let mut issued = pki.credential(&current, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 5);
+    issued["certificateChain"] = json!(
+        issued["certificateChain"]
+            .as_str()
+            .expect("certificate chain")
+            .trim_end_matches('\n')
+    );
+    let registration = server(&pki, vec![Reply::Json(StatusCode::CREATED, issued)]).await;
+    let registered = client(&registration, &pki, Duration::from_secs(2))
+        .register(&identity_store, &credential_store, &token())
+        .await
+        .expect("register");
+
+    let retries = server(
+        &pki,
+        vec![
+            Reply::DelayedClose(Duration::from_millis(200)),
+            Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})),
+            Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})),
+            Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})),
+            Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})),
+            Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})),
+        ],
+    )
+    .await;
+    let retry_client = client(&retries, &pki, Duration::from_millis(80));
+    let due = registered.not_after_unix - 8 * 60 * 60;
+    let (first, second) = tokio::join!(
+        retry_client.rotate_if_due(&identity_store, &credential_store, due),
+        retry_client.rotate_if_due(&identity_store, &credential_store, due)
+    );
+    assert!(matches!(first, Err(ClientError::Unavailable { .. })));
+    assert!(matches!(second, Err(ClientError::Unavailable { .. })));
+    let (request_id, certificate_request) = {
+        let seen = retries.seen.lock().expect("seen lock");
+        assert!(seen.len() >= 3, "bounded retries must reach the server");
+        for request in &seen[1..] {
+            assert_eq!(request["requestId"], seen[0]["requestId"]);
+            assert_eq!(request["certificateRequest"], seen[0]["certificateRequest"]);
+        }
+        (seen[0]["requestId"].clone(), seen[0]["certificateRequest"].clone())
+    };
+
+    let next_der = fs::read(temp.path().join("identity/device.key.next")).expect("read staged next key");
+    let next = rustfs::connect::DeviceIdentity::from_pkcs8_der(&next_der).expect("parse next key");
+    assert_ne!(current.public_key_der(), next.public_key_der());
+    assert_eq!(
+        identity_store
+            .load()
+            .expect("load current key")
+            .expect("current key")
+            .public_key_der(),
+        current.public_key_der()
+    );
+    #[cfg(unix)]
+    {
+        use std::os::unix::fs::PermissionsExt as _;
+        let mode = fs::metadata(temp.path().join("identity/device.key.next"))
+            .expect("next key metadata")
+            .permissions()
+            .mode()
+            & 0o7777;
+        assert_eq!(mode, 0o600);
+    }
+    let (rotated, _) = rotation_response(&pki, &next, 6);
+    let success = server_with_client_auth(
+        &pki,
+        vec![Reply::VerifiedRotation {
+            response: rotated,
+            current_public_key: current.public_key_der(),
+            current_certificate_fingerprint: certificate_fingerprint(&registered.certificate),
+            device_name: registered.name.clone(),
+        }],
+        true,
+    )
+    .await;
+    let success_client = client(&success, &pki, Duration::from_secs(2));
+    let (due_result, current_result) = tokio::join!(
+        success_client.rotate_if_due(&identity_store, &credential_store, due),
+        success_client.rotate_if_due(&identity_store, &credential_store, OffsetDateTime::now_utc().unix_timestamp())
+    );
+    let credential = due_result
+        .expect("retry rotation")
+        .or(current_result.expect("concurrent current-state check"))
+        .expect("exactly one rotation is due");
+    assert_eq!(credential.certificate_serial, "06".repeat(16));
+    let success_seen = success.seen.lock().expect("seen lock");
+    assert_eq!(success_seen.len(), 1, "the post-commit actor must not publish stale state");
+    assert_eq!(success_seen[0]["requestId"], request_id);
+    assert_eq!(success_seen[0]["certificateRequest"], certificate_request);
+    drop(success_seen);
+    assert_eq!(
+        identity_store
+            .load()
+            .expect("load key")
+            .expect("current key")
+            .public_key_der(),
+        next.public_key_der()
+    );
+    assert!(!temp.path().join("identity/device.key.next").exists());
+    assert!(!temp.path().join("credential/rotation.pending.json").exists());
+}
+
+#[tokio::test]
+async fn rotation_commit_recovers_after_each_durable_step() {
+    let temp = tempfile::tempdir().expect("temp dir");
+    let (identity_store, credential_store) = stores(&temp);
+    let current = identity_store.load_or_create().expect("create identity");
+    let pki = TestPki::new();
+    let issued = pki.credential(&current, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 7);
+    let registration = server(&pki, vec![Reply::Json(StatusCode::CREATED, issued)]).await;
+    let registered = client(&registration, &pki, Duration::from_secs(2))
+        .register(&identity_store, &credential_store, &token())
+        .await
+        .expect("register");
+    let failed = server(&pki, vec![Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})); 3]).await;
+    let due = registered.not_after_unix - 8 * 60 * 60;
+    assert!(matches!(
+        client(&failed, &pki, Duration::from_secs(2))
+            .rotate_if_due(&identity_store, &credential_store, due)
+            .await,
+        Err(ClientError::Unavailable { .. })
+    ));
+
+    let pending_path = temp.path().join("credential/rotation.pending.json");
+    let pending = fs::read(&pending_path).expect("read pending state");
+    let next_der = fs::read(temp.path().join("identity/device.key.next")).expect("read next key");
+    let next = rustfs::connect::DeviceIdentity::from_pkcs8_der(&next_der).expect("parse next key");
+    let (_, stored) = rotation_response(&pki, &next, 8);
+    write_stored_credential(&temp.path().join("credential/device.crt.json"), &stored);
+
+    let idle = server(&pki, vec![]).await;
+    assert!(
+        client(&idle, &pki, Duration::from_secs(2))
+            .rotate_if_due(&identity_store, &credential_store, OffsetDateTime::now_utc().unix_timestamp())
+            .await
+            .expect("recover after credential save")
+            .is_none()
+    );
+    assert_eq!(
+        identity_store
+            .load()
+            .expect("load key")
+            .expect("current key")
+            .public_key_der(),
+        next.public_key_der()
+    );
+
+    fs::write(&pending_path, pending).expect("restore pending after key commit");
+    set_owner_only(&pending_path);
+    assert!(
+        client(&idle, &pki, Duration::from_secs(2))
+            .rotate_if_due(&identity_store, &credential_store, OffsetDateTime::now_utc().unix_timestamp())
+            .await
+            .expect("recover after key commit")
+            .is_none()
+    );
+    assert!(!pending_path.exists());
+}
+
+#[tokio::test]
+async fn pending_reenrollment_blocks_rotation_and_resumes_original_exchange() {
+    let temp = tempfile::tempdir().expect("temp dir");
+    let (identity_store, credential_store) = stores(&temp);
+    let current = identity_store.load_or_create().expect("create identity");
+    let pki = TestPki::new();
+    let issued = pki.credential(&current, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 13);
+    let registration = server(&pki, vec![Reply::Json(StatusCode::CREATED, issued)]).await;
+    let registered = client(&registration, &pki, Duration::from_secs(2))
+        .register(&identity_store, &credential_store, &token())
+        .await
+        .expect("register");
+
+    let failed = server(&pki, vec![Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})); 3]).await;
+    assert!(matches!(
+        client(&failed, &pki, Duration::from_secs(2))
+            .reenroll(&identity_store, &credential_store, &token_with_uid(FRESH_TOKEN_UID))
+            .await,
+        Err(ClientError::Unavailable { .. })
+    ));
+
+    let pending_path = temp.path().join("credential/registration.pending.json");
+    let pending = fs::read(&pending_path).expect("read pending reenrollment");
+    let pending_document: Value = serde_json::from_slice(&pending).expect("pending JSON");
+    let next_der = fs::read(temp.path().join("identity/device.key.next")).expect("read next key");
+    let next = rustfs::connect::DeviceIdentity::from_pkcs8_der(&next_der).expect("parse next key");
+    let enrolled = pki.credential(&next, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 14);
+
+    let rotation = server(&pki, vec![]).await;
+    let error = client(&rotation, &pki, Duration::from_secs(2))
+        .rotate_if_due(&identity_store, &credential_store, registered.not_after_unix - 8 * 60 * 60)
+        .await
+        .expect_err("pending reenrollment blocks rotation");
+    assert!(matches!(error, ClientError::PendingRegistration));
+    assert!(rotation.seen.lock().expect("seen lock").is_empty());
+
+    let resumed = server(&pki, vec![Reply::Json(StatusCode::CREATED, enrolled)]).await;
+    let credential = client(&resumed, &pki, Duration::from_secs(2))
+        .reenroll(&identity_store, &credential_store, &token_with_uid(FRESH_TOKEN_UID))
+        .await
+        .expect("resume reenrollment");
+    assert_eq!(credential.certificate_serial, "0e".repeat(16));
+    let resumed_seen = resumed.seen.lock().expect("seen lock");
+    assert_eq!(resumed_seen.len(), 1);
+    assert_eq!(resumed_seen[0]["requestId"], pending_document["requestId"]);
+    assert_eq!(resumed_seen[0]["certificateRequest"], pending_document["certificateRequest"]);
+}
+
+#[tokio::test]
+async fn reenrollment_commit_recovers_after_each_durable_step() {
+    let temp = tempfile::tempdir().expect("temp dir");
+    let (identity_store, credential_store) = stores(&temp);
+    let current = identity_store.load_or_create().expect("create identity");
+    let pki = TestPki::new();
+    let issued = pki.credential(&current, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 13);
+    let registration = server(&pki, vec![Reply::Json(StatusCode::CREATED, issued)]).await;
+    client(&registration, &pki, Duration::from_secs(2))
+        .register(&identity_store, &credential_store, &token())
+        .await
+        .expect("register");
+
+    let failed = server(&pki, vec![Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})); 3]).await;
+    assert!(matches!(
+        client(&failed, &pki, Duration::from_secs(2))
+            .reenroll(&identity_store, &credential_store, &token_with_uid(FRESH_TOKEN_UID))
+            .await,
+        Err(ClientError::Unavailable { .. })
+    ));
+
+    let pending_path = temp.path().join("credential/registration.pending.json");
+    let pending = fs::read(&pending_path).expect("read pending reenrollment");
+    let next_der = fs::read(temp.path().join("identity/device.key.next")).expect("read next key");
+    let next = rustfs::connect::DeviceIdentity::from_pkcs8_der(&next_der).expect("parse next key");
+    let enrolled = pki.credential(&next, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 14);
+    write_stored_credential(&temp.path().join("credential/device.crt.json"), &enrolled);
+
+    let idle = server(&pki, vec![]).await;
+    let idle_client = client(&idle, &pki, Duration::from_secs(2));
+    let recovered = idle_client
+        .reenroll(&identity_store, &credential_store, &token_with_uid(FRESH_TOKEN_UID))
+        .await
+        .expect("recover after reenrollment credential save");
+    assert_eq!(recovered.certificate_serial, "0e".repeat(16));
+    assert_eq!(
+        identity_store
+            .load()
+            .expect("load key")
+            .expect("current key")
+            .public_key_der(),
+        next.public_key_der()
+    );
+
+    fs::remove_file(temp.path().join("credential/registration.completed.json")).expect("remove completed receipt");
+    fs::write(&pending_path, pending).expect("restore pending after key commit");
+    set_owner_only(&pending_path);
+    let recovered = idle_client
+        .reenroll(&identity_store, &credential_store, &token_with_uid(FRESH_TOKEN_UID))
+        .await
+        .expect("recover after reenrollment key commit");
+    assert_eq!(recovered.certificate_serial, "0e".repeat(16));
+    assert!(!pending_path.exists());
+    let recovered = idle_client
+        .reenroll(&identity_store, &credential_store, &token_with_uid(FRESH_TOKEN_UID))
+        .await
+        .expect("completed reenrollment is idempotent after pending cleanup");
+    assert_eq!(recovered.certificate_serial, "0e".repeat(16));
+    assert!(idle.seen.lock().expect("seen lock").is_empty());
+
+    let different = server(&pki, vec![Reply::Json(StatusCode::SERVICE_UNAVAILABLE, json!({})); 3]).await;
+    assert!(matches!(
+        client(&different, &pki, Duration::from_secs(2))
+            .reenroll(&identity_store, &credential_store, &token_with_uid(SECOND_TOKEN_UID))
+            .await,
+        Err(ClientError::Unavailable { .. })
+    ));
+    assert_eq!(different.seen.lock().expect("seen lock").len(), 3);
+}
+
+#[test]
+fn registration_token_schema_is_strict_and_bounded() {
+    let mut document = serde_json::to_value(token_document()).expect("token document");
+    document["unexpected"] = json!(true);
+    assert!(matches!(
+        RegistrationToken::from_reader(serde_json::to_vec(&document).expect("token JSON").as_slice()),
+        Err(TokenError::Invalid(_))
+    ));
+    assert!(matches!(
+        RegistrationToken::from_reader(vec![b' '; 16 * 1024 + 1].as_slice()),
+        Err(TokenError::TooLarge)
+    ));
+    let mut malformed = token_document();
+    malformed["challengeNonce"] = json!("A".repeat(64));
+    assert!(matches!(
+        RegistrationToken::from_reader(serde_json::to_vec(&malformed).expect("token JSON").as_slice()),
+        Err(TokenError::Shape)
+    ));
+}
+
+#[tokio::test]
+async fn rotation_waits_for_threshold_and_stops_on_revocation() {
+    let temp = tempfile::tempdir().expect("temp dir");
+    let (identity_store, credential_store) = stores(&temp);
+    let identity = identity_store.load_or_create().expect("create identity");
+    let pki = TestPki::new();
+    let issued = pki.credential(&identity, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 3);
+    let rotation_server = server(
+        &pki,
+        vec![
+            Reply::Json(StatusCode::CREATED, issued),
+            Reply::Json(StatusCode::UNAUTHORIZED, json!({"details": [{"reason": "DEVICE_REVOKED"}]})),
+        ],
+    )
+    .await;
+    let connect = client(&rotation_server, &pki, Duration::from_secs(2));
+    let current = connect
+        .register(&identity_store, &credential_store, &token())
+        .await
+        .expect("register");
+    assert!(
+        connect
+            .rotate_if_due(&identity_store, &credential_store, current.not_before_unix)
+            .await
+            .expect("not due")
+            .is_none()
+    );
+    let error = connect
+        .rotate_if_due(&identity_store, &credential_store, current.not_after_unix - 8 * 60 * 60)
+        .await
+        .expect_err("revocation must stop rotation");
+    assert!(matches!(error, ClientError::AccessRevoked { .. }));
+    assert!(error.to_string().contains("ConnectClient::reenroll"));
+    assert_eq!(rotation_server.seen.lock().expect("seen lock").len(), 2);
+    let stored: Value =
+        serde_json::from_slice(&fs::read(temp.path().join("credential/device.crt.json")).expect("read stored credential"))
+            .expect("stored credential JSON");
+    assert_eq!(stored["certificateSerial"], current.certificate_serial);
+
+    let next_der = fs::read(temp.path().join("identity/device.key.next")).expect("read staged next key");
+    let next = rustfs::connect::DeviceIdentity::from_pkcs8_der(&next_der).expect("parse staged next key");
+    let enrolled = pki.credential(&next, &format!("urn:rustfs:connect:device:{DEVICE_UID}"), 12);
+    let reenrollment = server(&pki, vec![Reply::Json(StatusCode::CREATED, enrolled)]).await;
+    let fresh = client(&reenrollment, &pki, Duration::from_secs(2))
+        .reenroll(&identity_store, &credential_store, &token_with_uid(FRESH_TOKEN_UID))
+        .await
+        .expect("fresh token reenrolls revoked credential");
+    assert_eq!(fresh.certificate_serial, "0c".repeat(16));
+    assert_eq!(
+        identity_store
+            .load()
+            .expect("load identity")
+            .expect("identity")
+            .public_key_der(),
+        next.public_key_der()
+    );
+}
+
+#[test]
+fn unconfigured_connect_has_no_side_effects() {
+    let temp = tempfile::tempdir().expect("temp dir");
+    let directory = temp.path().join("connect");
+    assert!(
+        ConnectClient::from_optional_config(None)
+            .expect("unconfigured is valid")
+            .is_none()
+    );
+    assert!(!directory.exists());
+}
@@ -0,0 +1,419 @@
+#!/usr/bin/env python3
+"""Fail when committed tests silently fall out of their execution wiring."""
+
+from __future__ import annotations
+
+import hashlib
+import json
+import re
+import sys
+import tempfile
+import tomllib
+import unittest
+from unittest import mock
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+
+
+def words(value: str) -> set[str]:
+    return {item.strip() for item in value.split(",") if item.strip()}
+
+
+def rust_code_only(source: str) -> str:
+    """Blank Rust comments and literals while preserving byte positions."""
+    def quoted_end(quote_index: int, delimiter: str) -> int:
+        end = quote_index + 1
+        while end < len(source):
+            if source[end] == "\\":
+                end += 2
+            elif source[end] == delimiter:
+                return end + 1
+            else:
+                end += 1
+        return end
+
+    code = list(source)
+    index = 0
+    while index < len(source):
+        if source.startswith("//", index):
+            end = source.find("\n", index)
+            end = len(source) if end < 0 else end
+        elif source.startswith("/*", index):
+            depth = 1
+            end = index + 2
+            while end < len(source) and depth:
+                if source.startswith("/*", end):
+                    depth += 1
+                    end += 2
+                elif source.startswith("*/", end):
+                    depth -= 1
+                    end += 2
+                else:
+                    end += 1
+        else:
+            raw = re.match(r'(?:br|cr|r)(?P<hashes>#{0,})"', source[index:])
+            if raw:
+                marker = '"' + raw.group("hashes")
+                end = source.find(marker, index + raw.end())
+                end = len(source) if end < 0 else end + len(marker)
+            elif source[index] == '"' or source.startswith(('b"', 'c"'), index):
+                start_quote = index if source[index] == '"' else index + 1
+                end = quoted_end(start_quote, '"')
+            elif source[index] == "'" and index + 2 < len(source) and (
+                source[index + 1] == "\\" or source[index + 2] == "'"
+            ):
+                end = quoted_end(index, "'")
+            elif source.startswith("b'", index):
+                end = quoted_end(index + 1, "'")
+            else:
+                index += 1
+                continue
+
+        for offset in range(index, end):
+            if code[offset] != "\n":
+                code[offset] = " "
+        index = end
+    return "".join(code)
+
+
+def declared(parent: Path, module: str) -> bool:
+    source = parent.read_text()
+    code = rust_code_only(source)
+    pattern = re.compile(rf"^\s*(?:pub(?:\([^)]*\))?\s+)?mod\s+{re.escape(module)}\s*;", re.MULTILINE)
+    allowed_preambles = {
+        "#[cfg(test)]": "#[cfg(test)]",
+        "#[cfg(all(test,target_os=))]": '#[cfg(all(test,target_os="linux"))]',
+    }
+    for match in pattern.finditer(code):
+        prefix = code[: match.start()]
+        depths = {"(": 0, "[": 0, "{": 0}
+        pairs = {")": "(", "]": "[", "}": "{"}
+        for char in prefix:
+            if char in depths:
+                depths[char] += 1
+            elif char in pairs:
+                depths[pairs[char]] -= 1
+        if any(depths.values()):
+            continue
+
+        boundary = max(prefix.rfind(";"), prefix.rfind("{"), prefix.rfind("}"))
+        code_preamble = re.sub(r"\s+", "", prefix[boundary + 1 :])
+        if not code_preamble:
+            return True
+        if code_preamble in allowed_preambles:
+            attr_start = prefix.rfind("#[cfg", boundary + 1)
+            if attr_start >= 0 and re.sub(r"\s+", "", source[attr_start : match.start()]) == allowed_preambles[code_preamble]:
+                return True
+    return False
+
+
+def module_source(src: Path, directory: Path) -> Path | None:
+    if not directory.parts:
+        return src / "lib.rs"
+
+    mod_file = src / directory / "mod.rs"
+    if mod_file.is_file():
+        return mod_file
+
+    sibling = src.joinpath(*directory.parts[:-1], f"{directory.name}.rs")
+    return sibling if sibling.is_file() else None
+
+
+def check_e2e_modules(root: Path) -> list[str]:
+    src = root / "crates/e2e_test/src"
+    errors: list[str] = []
+    for test_file in sorted(src.rglob("*_test.rs")):
+        relative = test_file.relative_to(root).as_posix()
+        directory = test_file.relative_to(src).parent
+        parent = module_source(src, directory)
+        if parent is None:
+            errors.append(f"{relative}: no canonical parent module")
+            continue
+
+        if not declared(parent, test_file.stem):
+            errors.append(f"{relative}: not declared by {parent.relative_to(root).as_posix()}")
+        while directory.parts:
+            module = directory.name
+            directory = directory.parent
+            parent = module_source(src, directory)
+            if parent is None:
+                errors.append(f"{relative}: module {module} has no canonical parent")
+                break
+            if not declared(parent, module):
+                errors.append(f"{relative}: module {module} not declared by {parent.relative_to(root).as_posix()}")
+    return errors
+
+
+def check_fuzz_targets(root: Path) -> list[str]:
+    manifest = tomllib.loads((root / "fuzz/Cargo.toml").read_text())
+    expected = {item["name"] for item in manifest.get("bin", []) if "name" in item}
+    errors: list[str] = []
+    if not expected:
+        return ["fuzz/Cargo.toml: no [[bin]] fuzz targets found"]
+
+    runner = (root / "scripts/fuzz/run.sh").read_text()
+    match = re.search(r'^targets="([^"]+)"', runner, re.MULTILINE)
+    runner_targets = set(match.group(1).split()) if match else set()
+    if runner_targets != expected:
+        errors.append(f"scripts/fuzz/run.sh targets {sorted(runner_targets)} != manifest {sorted(expected)}")
+
+    workflow = (root / ".github/workflows/fuzz.yml").read_text()
+    matrices = [words(value) for value in re.findall(r"^\s*target:\s*\[([^]]+)]", workflow, re.MULTILINE)]
+    if len(matrices) != 2:
+        errors.append(f".github/workflows/fuzz.yml: expected smoke and nightly target matrices, found {len(matrices)}")
+    for index, matrix in enumerate(matrices, start=1):
+        if matrix != expected:
+            errors.append(f".github/workflows/fuzz.yml matrix {index} {sorted(matrix)} != manifest {sorted(expected)}")
+
+    runtime_targets = re.findall(r"^\s*FUZZ_TARGET:\s*(\S.*?)\s*$", workflow, re.MULTILINE)
+    if runtime_targets != ["${{ matrix.target }}", "${{ matrix.target }}"]:
+        errors.append(".github/workflows/fuzz.yml: smoke and nightly jobs must pass matrix.target to FUZZ_TARGET")
+
+    dependency_paths = {
+        f"{path.removeprefix('../')}/**"
+        for dependency in manifest.get("dependencies", {}).values()
+        if isinstance(dependency, dict)
+        and isinstance(path := dependency.get("path"), str)
+        and path.startswith("../crates/")
+    }
+    missing_paths = sorted(path for path in dependency_paths if f'"{path}"' not in workflow)
+    if missing_paths:
+        errors.append(f".github/workflows/fuzz.yml missing direct dependency paths: {', '.join(missing_paths)}")
+
+    staged_matches = re.findall(r"^\s*for target in ([^;]+); do", workflow, re.MULTILINE)
+    staged = set(staged_matches[0].split()) if staged_matches else set()
+    if staged != expected:
+        errors.append(f".github/workflows/fuzz.yml staged binaries {sorted(staged)} != manifest {sorted(expected)}")
+
+    return errors
+
+
+def check_runner_selection(root: Path) -> list[str]:
+    runner = (root / "scripts/run_e2e_tests.sh").read_text()
+    errors: list[str] = []
+    if "--include-ignored" not in runner:
+        errors.append("scripts/run_e2e_tests.sh: runner must include default and ignored tests")
+    if "--test-threads=1" not in runner:
+        errors.append("scripts/run_e2e_tests.sh: runner must serialize fixed-port protocol tests")
+    if re.search(r"(?<!include-)--ignored\b", runner):
+        errors.append("scripts/run_e2e_tests.sh: bare --ignored silently excludes default tests")
+    if "--exact" in runner:
+        errors.append("scripts/run_e2e_tests.sh: --test is documented as a pattern and must not force exact matching")
+    if 'eval "$test_cmd"' in runner:
+        errors.append("scripts/run_e2e_tests.sh: command construction must not use eval")
+    return errors
+
+
+def profile_selection(root: Path, profile: str) -> str:
+    if not re.fullmatch(r"e2e-[a-z0-9-]+", profile):
+        raise ValueError(f"invalid e2e profile name: {profile}")
+    path = root / f".config/{profile}-selection.txt"
+    lines = [line for line in path.read_text().splitlines() if line.strip()]
+    values = dict(line.split("=", 1) for line in lines if "=" in line)
+    if len(values) != len(lines) or any(not re.fullmatch(r"sha256(?:-[a-z0-9]+)?", key) for key in values):
+        raise ValueError(f"{path.relative_to(root).as_posix()}: invalid sha256 entry")
+    key = f"sha256-{sys.platform}"
+    digest = values.get(key, values.get("sha256", ""))
+    if not re.fullmatch(r"[0-9a-f]{64}", digest):
+        raise ValueError(f"{path.relative_to(root).as_posix()}: missing sha256 for {sys.platform}")
+    return digest
+
+
+def check_profile_definitions(root: Path) -> list[str]:
+    config = tomllib.loads((root / ".config/nextest.toml").read_text())
+    profiles = {
+        profile
+        for profile in config.get("profile", {})
+        if profile.startswith("e2e-")
+    }
+    selection_profiles = {
+        path.name.removesuffix("-selection.txt") for path in (root / ".config").glob("e2e-*-selection.txt")
+    }
+    errors: list[str] = []
+    for profile in sorted(profiles | selection_profiles):
+        if profile not in profiles:
+            errors.append(f".config/nextest.toml: missing profile.{profile}")
+        if profile not in selection_profiles:
+            errors.append(f".config/{profile}-selection.txt: missing expected profile selection")
+            continue
+        try:
+            profile_selection(root, profile)
+        except (FileNotFoundError, ValueError) as error:
+            errors.append(str(error))
+    return errors
+
+
+def check_profile_listing(root: Path, profile: str, listing: Path) -> list[str]:
+    try:
+        expected_digest = profile_selection(root, profile)
+        data = json.loads(listing.read_text())
+        selected = sorted(
+            f"{suite_id}::{test_name}"
+            for suite_id, suite in data["rust-suites"].items()
+            for test_name, testcase in suite["testcases"].items()
+            if testcase.get("filter-match", {}).get("status") == "matches"
+        )
+        digest = hashlib.sha256(("\n".join(selected) + "\n").encode()).hexdigest()
+    except (FileNotFoundError, KeyError, TypeError, ValueError, json.JSONDecodeError) as error:
+        return [f"cannot read {profile} nextest listing: {error}"]
+    if digest != expected_digest:
+        return [
+            f"{profile} selection changed: count={len(selected)} sha256={digest}; "
+            f"expected sha256={expected_digest}"
+        ]
+    print(f"{profile} selection OK: {len(selected)} tests, sha256={digest}")
+    return []
+
+
+def validate(root: Path) -> list[str]:
+    errors: list[str] = []
+    errors.extend(check_e2e_modules(root))
+    errors.extend(check_fuzz_targets(root))
+    errors.extend(check_runner_selection(root))
+    errors.extend(check_profile_definitions(root))
+    return errors
+
+
+class SelfTests(unittest.TestCase):
+    def test_e2e_requires_registration(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            root = Path(tmp)
+            src = root / "crates/e2e_test/src"
+            src.mkdir(parents=True)
+            (src / "lib.rs").write_text("")
+            test_file = src / "boundary_test.rs"
+            test_file.write_text("#[test]\nfn boundary() {}\n")
+            self.assertEqual(len(check_e2e_modules(root)), 1)
+            (src / "lib.rs").write_text("mod boundary_test;\n")
+            self.assertEqual(check_e2e_modules(root), [])
+
+            nested = src / "protocols"
+            nested.mkdir()
+            (nested / "mod.rs").write_text("mod fixed_port_test;\n")
+            (nested / "fixed_port_test.rs").write_text("#[test]\nfn fixed_port() {}\n")
+            self.assertEqual(len(check_e2e_modules(root)), 1)
+            (src / "lib.rs").write_text("mod boundary_test;\nmod protocols;\n")
+            self.assertEqual(check_e2e_modules(root), [])
+
+            (src / "lib.rs").write_text("#[cfg(any())]\nmod boundary_test;\nmod protocols;\n")
+            self.assertEqual(len(check_e2e_modules(root)), 1)
+
+            (src / "lib.rs").write_text(
+                "#[cfg(any())]\n/// hidden module\nmod boundary_test;\n#[cfg_attr(test, cfg(any()))]\nmod protocols;\n"
+            )
+            self.assertEqual(len(check_e2e_modules(root)), 2)
+
+            (src / "lib.rs").write_text(
+                'const PHANTOM: &str = r#"{\nmod boundary_test;\n"#;\ndiscard! { mod protocols; }\n'
+            )
+            self.assertEqual(len(check_e2e_modules(root)), 2)
+
+            (src / "lib.rs").write_text(
+                '#[cfg(all(test, target_os = r"windows" /* target_os = "linux" */))]\n'
+                "mod boundary_test;\nmod protocols;\n"
+            )
+            self.assertEqual(len(check_e2e_modules(root)), 1)
+
+            (src / "lib.rs").write_text(
+                '#[cfg(all(test, target_os = r"windows"))] // #[cfg(all(test, target_os = "linux"))]\n'
+                "mod boundary_test;\nmod protocols;\n"
+            )
+            self.assertEqual(len(check_e2e_modules(root)), 1)
+
+    def test_fuzz_runtime_uses_matrix_target(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            root = Path(tmp)
+            (root / "fuzz").mkdir()
+            (root / "scripts/fuzz").mkdir(parents=True)
+            (root / ".github/workflows").mkdir(parents=True)
+            (root / "fuzz/Cargo.toml").write_text(
+                'dep = { path = "../crates/dep" }\n[[bin]]\nname = "one"\n'
+            )
+            (root / "scripts/fuzz/run.sh").write_text('targets="one"\n')
+            (root / ".github/workflows/fuzz.yml").write_text(
+                'paths:\n  - "crates/dep/**"\n'
+                "target: [one]\nFUZZ_TARGET: fixed\n"
+                "target: [one]\nFUZZ_TARGET: ${{ matrix.target }}\n"
+                "for target in one; do\n"
+                "  fuzz/prebuilt/${{ env.CARGO_BUILD_TARGET }}/release/one\n"
+            )
+            self.assertEqual(len(check_fuzz_targets(root)), 1)
+
+    def test_profile_listing_enforces_selection(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            root = Path(tmp)
+            (root / ".config").mkdir()
+            digest = hashlib.sha256(b"suite::two\n").hexdigest()
+            (root / ".config/e2e-smoke-selection.txt").write_text(f"sha256={digest}\n")
+            listing = root / "listing.json"
+            listing.write_text(
+                json.dumps(
+                    {
+                        "rust-suites": {
+                            "suite": {
+                                "testcases": {
+                                    "one": {"filter-match": {"status": "matches"}},
+                                    "two": {"filter-match": {"status": "mismatch"}},
+                                }
+                            }
+                        }
+                    }
+                )
+            )
+            self.assertEqual(len(check_profile_listing(root, "e2e-smoke", listing)), 1)
+
+    def test_profile_listing_binds_platform_digest(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            root = Path(tmp)
+            (root / ".config").mkdir()
+            darwin_digest = hashlib.sha256(b"suite::darwin\n").hexdigest()
+            linux_digest = hashlib.sha256(b"suite::linux\n").hexdigest()
+            (root / ".config/e2e-full-selection.txt").write_text(
+                f"sha256-darwin={darwin_digest}\nsha256-linux={linux_digest}\n"
+            )
+            listing = root / "listing.json"
+            listing.write_text(
+                json.dumps(
+                    {
+                        "rust-suites": {
+                            "suite": {
+                                "testcases": {"darwin": {"filter-match": {"status": "matches"}}}
+                            }
+                        }
+                    }
+                )
+            )
+            with mock.patch.object(sys, "platform", "linux"):
+                self.assertEqual(len(check_profile_listing(root, "e2e-full", listing)), 1)
+
+def main() -> int:
+    if sys.argv[1:] == ["--self-test"]:
+        suite = unittest.defaultTestLoader.loadTestsFromTestCase(SelfTests)
+        return 0 if unittest.TextTestRunner(verbosity=2).run(suite).wasSuccessful() else 1
+    if len(sys.argv) == 4 and sys.argv[1] == "--check-profile":
+        errors = check_profile_listing(ROOT, sys.argv[2], Path(sys.argv[3]))
+        if errors:
+            for error in errors:
+                print(f"ERROR: {error}", file=sys.stderr)
+            return 1
+        return 0
+    if sys.argv[1:]:
+        print(
+            "usage: check_test_wiring.py [--self-test | --check-profile PROFILE LISTING]",
+            file=sys.stderr,
+        )
+        return 2
+
+    errors = validate(ROOT)
+    if errors:
+        for error in errors:
+            print(f"ERROR: {error}", file=sys.stderr)
+        return 1
+    print("OK: e2e modules, runner selection, fuzz matrices, and profile guards are wired")
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
@@ -20,6 +20,8 @@ DATA_DIR="$TARGET_DIR/rustfs_test_data"
 RUSTFS_PID=""
 TEST_FILTER=""
 TEST_TYPE="all"
+RUSTFS_BUILD_FEATURES="${RUSTFS_BUILD_FEATURES:-ftps,webdav,sftp}"
+export RUSTFS_BUILD_FEATURES

 # Function to print colored output
 print_info() {
@@ -92,7 +94,7 @@ build_rustfs() {
    print_info "Building RustFS..."
    cd "$PROJECT_ROOT"
    
-    if ! cargo build --bin rustfs; then
+    if ! cargo build --bin rustfs --features "$RUSTFS_BUILD_FEATURES"; then
        print_error "Failed to build RustFS"
        exit 1
    fi
@@ -219,27 +221,28 @@ start_rustfs() {
 run_tests() {
    print_info "Running e2e tests..."
    cd "$PROJECT_ROOT"
-    
-    local test_cmd="cargo test --package e2e_test --lib"
-    
+
+    local test_cmd=(cargo test --package e2e_test --lib)
+
    case "$TEST_TYPE" in
        "specific")
-            test_cmd="$test_cmd -- $TEST_FILTER --exact --show-output --ignored"
+            test_cmd+=(-- "$TEST_FILTER")
            print_info "Running specific test: $TEST_FILTER"
            ;;
        "file")
-            test_cmd="$test_cmd -- $TEST_FILTER --show-output --ignored"
+            test_cmd+=(-- "$TEST_FILTER")
            print_info "Running tests in file/module: $TEST_FILTER"
            ;;
        "all")
-            test_cmd="$test_cmd -- --show-output --ignored"
+            test_cmd+=(--)
            print_info "Running all e2e tests"
            ;;
    esac
-    
-    print_info "Test command: $test_cmd"
-    
-    if eval "$test_cmd"; then
+    test_cmd+=(--show-output --include-ignored --test-threads=1)
+
+    print_info "Test command: ${test_cmd[*]}"
+
+    if "${test_cmd[@]}"; then
        print_success "All tests passed!"
        return 0
    else
@@ -253,6 +253,9 @@ Test results are saved in the `artifacts/s3tests-${TEST_MODE}/` directory (defau

 - `junit.xml`: Test results in JUnit format (compatible with CI/CD systems)
 - `pytest.log`: Detailed pytest logs with full test output
+- `all-collected-nodeids.txt`: Exact node IDs in the pinned upstream suite
+- `selected-nodeids.txt`: Exact node IDs expected in this run
+- `unsharded-selected-nodeids.txt`: Exact node IDs before deterministic sharding
 - `compat-report.md`: Classification report generated by `report_compat.py` —
  regressions against `implemented_tests.txt`, promotion candidates (tests
  that pass but are still listed as unimplemented/excluded), and tests missing
@@ -449,9 +452,11 @@ RustFS. Two GitHub Actions workflows delegate to it:
 - **Full sweep** (`.github/workflows/e2e-s3tests.yml`): weekly scheduled (and
  manually dispatchable) run of the ENTIRE upstream suite (`TEST_SCOPE=all`)
  against a Docker deployment — single node or a real 4-node distributed
-  cluster behind HAProxy. The sweep fails only on regressions in the
-  implemented whitelist; everything else is reported by `report_compat.py`
-  as promotion candidates or unclassified tests.
+  cluster behind HAProxy. Regressions, unclassified tests, incomplete
+  execution, and infrastructure errors fail the sweep; classified unsupported
+  behavior remains informational. Scheduled topology runs are split into four
+  deterministic exact-node-ID shards, and every case has a five-minute timeout,
+  so one stalled case cannot erase the entire sweep's evidence.

 Keeping both workflows on this script means local runs, the PR gate, and the
 scheduled sweep always execute tests the same way (same pinned s3-tests
@@ -466,8 +471,9 @@ pass/fail table in the job summary.
 ## Companion Tools

 - `report_compat.py` — diffs a junit.xml result against the classification
-  lists; run automatically at the end of `run.sh`, and used by the weekly
-  sweep to gate on whitelist regressions only (`--fail-on-regression`).
+  lists and the exact pytest collection; run before execution to reject stale
+  or missing classifications, then after execution to detect regressions and
+  incomplete parameterized cases.
 - `api_coverage.py` — quantifies S3 API surface coverage by comparing the
  s3s `S3` trait (at the revision pinned in Cargo.toml) against the methods
  RustFS overrides in `impl S3 for FS`:
@@ -307,3 +307,10 @@ test_object_acl_write
 test_object_acl_writeacp
 test_put_bucket_acl_grant_group_read
 test_object_raw_get_bucket_acl
+
+# Require upstream cloud-storage or IAM account services
+test_bucket_logging_requester_assumed_role
+test_lifecycle_cloud_transition_target_by_bucket
+test_lifecycle_cloud_transition_target_by_bucket_multiple_buckets
+test_list_object_versions_restore_status
+test_list_objects_restore_status
@@ -521,9 +521,13 @@ test_atomic_dual_conditional_write_1mb
 test_atomic_write_bucket_gone
 test_bucket_acl_canned_private_to_private
 test_bucket_concurrent_set_canned_acl
+test_bucket_create_delete
+test_bucket_policy
 test_bucket_policy_acl
 test_bucket_policy_put_obj_acl
 test_bucketv2_policy_acl
+test_copy_enc
+test_copy_part_enc
 test_copy_object_ifmatch_failed
 test_copy_object_ifnonematch_good
 test_cors_presigned_put_object_tenant_with_acl
@@ -21,14 +21,17 @@ Classifies every executed test into:
  - unclassified passes:    passed but not present in any list (new upstream tests)
  - unclassified failures:  failed and not present in any list (new upstream tests)

-Writes a markdown report and prints a summary to stdout. Exit code is 0 unless
--fail-on-regression is given and at least one regression was found.
+Writes a markdown report and prints a summary to stdout. Optional gates reject
+regressions, unclassified tests, stale classifications, and incomplete node-ID
+execution.
 """

 from __future__ import annotations

 import argparse
+from collections import Counter
 import pathlib
+import re
 import sys
 import xml.etree.ElementTree as ET

@@ -45,33 +48,47 @@ LIST_FILES = {
 }


-def load_list(path: pathlib.Path) -> set[str]:
-    names: set[str] = set()
+def load_entries(path: pathlib.Path) -> list[str]:
+    names: list[str] = []
    if not path.is_file():
        return names
    for line in path.read_text(encoding="utf-8").splitlines():
        line = line.strip()
        if line and not line.startswith("#"):
-            names.add(line)
+            names.append(line)
    return names


+def classification_errors(entries: dict[str, list[str]]) -> list[str]:
+    errors: list[str] = []
+    lists = {key: set(names) for key, names in entries.items()}
+    for key, names in entries.items():
+        duplicates = sorted(name for name, count in Counter(names).items() if count > 1)
+        if duplicates:
+            errors.append(f"{LIST_FILES[key]} has duplicates: {', '.join(duplicates)}")
+    keys = tuple(lists)
+    for index, left in enumerate(keys):
+        for right in keys[index + 1 :]:
+            overlap = sorted(lists[left] & lists[right])
+            if overlap:
+                errors.append(f"{left}/{right} classifications overlap: {', '.join(overlap)}")
+    return errors
+
+
 def base_name(testcase_name: str) -> str:
    """Strip pytest parametrization (test_foo[param]) to match list entries."""
    return testcase_name.split("[", 1)[0]


-def parse_junit(path: pathlib.Path) -> dict[str, str]:
-    """Return {test name: status} with status in passed/failed/error/skipped.
-
-    Parametrized cases collapse onto their base name; any failing variant marks
-    the whole test failed.
-    """
+def parse_junit(path: pathlib.Path) -> tuple[dict[str, str], list[str], list[tuple[str, str, str, str]]]:
+    """Return exact statuses, pytest-timeout cases, and failure summaries."""
    results: dict[str, str] = {}
+    timed_out: list[str] = []
+    failures: list[tuple[str, str, str, str]] = []
    severity = {"skipped": 0, "passed": 1, "failed": 2, "error": 2}
    root = ET.parse(path).getroot()
    for case in root.iter("testcase"):
-        name = base_name(case.get("name", ""))
+        name = case.get("name", "")
        if not name:
            continue
        if case.find("failure") is not None:
@@ -85,7 +102,35 @@ def parse_junit(path: pathlib.Path) -> dict[str, str]:
        prev = results.get(name)
        if prev is None or severity[status] > severity[prev]:
            results[name] = status
-    return results
+        node = case.find("failure") if status == "failed" else case.find("error")
+        if node is not None:
+            details = " ".join(filter(None, [node.get("message", ""), node.text or ""]))
+            message = node.get("message") or next(iter((node.text or "").strip().splitlines()), "")
+            failures.append((case.get("classname", ""), name, case.get("time", "0"), message))
+            if re.search(r"\bTimeout\s*(?:>|\()", details, re.IGNORECASE):
+                timed_out.append(name)
+    return results, timed_out, failures
+
+
+def collapse_results(results: dict[str, str]) -> dict[str, str]:
+    """Collapse parametrized cases for classification-level reporting."""
+    collapsed: dict[str, str] = {}
+    severity = {"skipped": 0, "passed": 1, "failed": 2, "error": 2}
+    for exact_name, status in results.items():
+        name = base_name(exact_name)
+        previous = collapsed.get(name)
+        if previous is None or severity[status] > severity[previous]:
+            collapsed[name] = status
+    return collapsed
+
+
+def load_collected_nodeids(path: pathlib.Path) -> set[str]:
+    names: set[str] = set()
+    for line in path.read_text(encoding="utf-8").splitlines():
+        nodeid = line.strip()
+        if nodeid:
+            names.add(nodeid.rsplit("::", 1)[-1])
+    return names


 def render_section(title: str, rows: list[str], hint: str = "") -> list[str]:
@@ -102,7 +147,7 @@ def render_section(title: str, rows: list[str], hint: str = "") -> list[str]:

 def main() -> int:
    parser = argparse.ArgumentParser(description=__doc__)
-    parser.add_argument("--junit", required=True, type=pathlib.Path, help="junit.xml produced by pytest")
+    parser.add_argument("--junit", type=pathlib.Path, help="junit.xml produced by pytest")
    parser.add_argument(
        "--lists-dir",
        type=pathlib.Path,
@@ -115,14 +160,60 @@ def main() -> int:
        action="store_true",
        help="exit non-zero when a test from implemented_tests.txt failed",
    )
+    parser.add_argument(
+        "--fail-on-unclassified",
+        action="store_true",
+        help="exit non-zero when an executed test is absent from every classification",
+    )
+    parser.add_argument(
+        "--collected-nodeids",
+        type=pathlib.Path,
+        help="exact pytest node IDs from the pinned suite's collect-only pass",
+    )
+    parser.add_argument(
+        "--check-classifications-only",
+        action="store_true",
+        help="validate classification names against collected node IDs without reading JUnit",
+    )
    args = parser.parse_args()

-    if not args.junit.is_file():
+    entries = {key: load_entries(args.lists_dir / fname) for key, fname in LIST_FILES.items()}
+    lists = {key: set(names) for key, names in entries.items()}
+    invalid_classifications = classification_errors(entries)
+    collected: set[str] = set()
+    if args.collected_nodeids:
+        collected = load_collected_nodeids(args.collected_nodeids)
+        collected_base = {base_name(name) for name in collected}
+        classified = set().union(*lists.values())
+        missing_classifications = sorted(collected_base - classified)
+        stale_classifications = sorted(classified - collected_base)
+    else:
+        missing_classifications = []
+        stale_classifications = []
+
+    if args.check_classifications_only:
+        if not args.collected_nodeids:
+            parser.error("--check-classifications-only requires --collected-nodeids")
+        for error in invalid_classifications:
+            print(f"[INVALID] {error}")
+        for name in missing_classifications:
+            print(f"[UNCLASSIFIED] {name}")
+        for name in stale_classifications:
+            print(f"[STALE] {name}")
+        return 1 if invalid_classifications or missing_classifications or stale_classifications else 0
+
+    if invalid_classifications:
+        for error in invalid_classifications:
+            print(f"[ERROR] {error}", file=sys.stderr)
+        return 2
+
+    if not args.junit or not args.junit.is_file():
        print(f"[ERROR] junit file not found: {args.junit}", file=sys.stderr)
        return 2

-    lists = {key: load_list(args.lists_dir / fname) for key, fname in LIST_FILES.items()}
-    results = parse_junit(args.junit)
+    exact_results, timed_out, failures = parse_junit(args.junit)
+    results = collapse_results(exact_results)
+    missing_results = sorted(collected - exact_results.keys()) if collected else []

    regressions: list[str] = []
    promotions: dict[str, list[str]] = {"unimplemented": [], "excluded": []}
@@ -155,7 +246,9 @@ def main() -> int:
    lines = [
        "# S3 compatibility report",
        "",
-        f"Executed: {len(results)} tests — "
+        f"Executed: {len(exact_results)} exact cases across {len(results)} classified tests.",
+        "",
+        "Classification status — "
        f"{counts['passed']} passed, {counts['failed']} failed, "
        f"{counts['error']} errored, {counts['skipped']} skipped.",
        "",
@@ -191,6 +284,16 @@ def main() -> int:
        unclassified_failed,
        "Failing and absent from every list — triage into `unimplemented_tests.txt` or `excluded_tests.txt`.",
    )
+    lines += render_section(
+        "Missing results",
+        missing_results,
+        "Present in the pinned upstream suite but absent from JUnit — the sweep was incomplete.",
+    )
+    lines += render_section(
+        "Timed out",
+        timed_out,
+        "Per-test timeout is an infrastructure failure regardless of compatibility classification.",
+    )

    report = "\n".join(lines)
    if args.output:
@@ -201,13 +304,28 @@ def main() -> int:
    print(
        f"[INFO] {len(regressions)} regression(s), "
        f"{len(promotions['unimplemented']) + len(promotions['excluded']) + len(unclassified_passed)} promotion candidate(s), "
-        f"{len(unclassified_failed)} unclassified failure(s)"
+        f"{len(unclassified_failed)} unclassified failure(s), "
+        f"{len(missing_results)} missing result(s), "
+        f"{len(timed_out)} timeout(s)"
    )
    for name in sorted(regressions):
        print(f"[REGRESSION] {name}")
+    if failures:
+        print("[ERROR] s3-tests failed testcase summary:")
+        for classname, name, duration, message in failures[:20]:
+            nodeid = f"{classname}::{name}" if classname else name
+            print(f"[ERROR] - {nodeid} ({duration}s): {message}")
+        if len(failures) > 20:
+            print(f"[ERROR] - ... {len(failures) - 20} additional failed testcases omitted")

    if args.fail_on_regression and regressions:
        return 1
+    if args.fail_on_unclassified and (unclassified_passed or unclassified_failed):
+        return 1
+    if args.collected_nodeids and missing_results:
+        return 1
+    if timed_out:
+        return 1
    return 0


@@ -58,6 +58,19 @@ if [[ "${TEST_SCOPE}" != "implemented" && "${TEST_SCOPE}" != "all" ]]; then
    echo "[ERROR] Invalid TEST_SCOPE: ${TEST_SCOPE} (must be \"implemented\" or \"all\")" >&2
    exit 1
 fi
+S3_SHARD_COUNT="${S3_SHARD_COUNT:-1}"
+S3_SHARD_INDEX="${S3_SHARD_INDEX:-0}"
+TEST_TIMEOUT="${TEST_TIMEOUT:-300}"
+if [[ ! "${S3_SHARD_COUNT}" =~ ^[1-9][0-9]*$ ]] \
+    || [[ ! "${S3_SHARD_INDEX}" =~ ^[0-9]+$ ]] \
+    || (( S3_SHARD_INDEX >= S3_SHARD_COUNT )); then
+    echo "[ERROR] Invalid S3 shard ${S3_SHARD_INDEX}/${S3_SHARD_COUNT}" >&2
+    exit 1
+fi
+if [[ ! "${TEST_TIMEOUT}" =~ ^[1-9][0-9]*$ ]]; then
+    echo "[ERROR] Invalid TEST_TIMEOUT: ${TEST_TIMEOUT}" >&2
+    exit 1
+fi

 # Upstream ceph/s3-tests suite, pinned for reproducible runs.
 # Bump S3TESTS_REV deliberately: upstream changes can rename tests or change
@@ -96,55 +109,6 @@ log_error() {
    echo -e "${RED}[ERROR]${NC} $*"
 }

-summarize_junit_failures() {
-    local junit_path="$1"
-
-    if [ ! -f "${junit_path}" ]; then
-        log_warn "JUnit report not found: ${junit_path}"
-        return 0
-    fi
-
-    python3 - "${junit_path}" <<'PY'
-import sys
-import xml.etree.ElementTree as ET
-
-junit_path = sys.argv[1]
-try:
-    root = ET.parse(junit_path).getroot()
-except Exception as exc:
-    print(f"[WARN] Failed to parse JUnit report {junit_path}: {exc}")
-    raise SystemExit(0)
-
-failures = []
-for case in root.iter("testcase"):
-    failure = case.find("failure")
-    error = case.find("error")
-    node = failure if failure is not None else error
-    if node is None:
-        continue
-
-    classname = case.attrib.get("classname", "")
-    name = case.attrib.get("name", "")
-    duration = case.attrib.get("time", "0")
-    message = node.attrib.get("message") or (node.text or "").strip().splitlines()[0:1]
-    if isinstance(message, list):
-        message = message[0] if message else ""
-    failures.append((classname, name, duration, message))
-
-if not failures:
-    print("[INFO] No failed testcases found in JUnit report")
-    raise SystemExit(0)
-
-print("[ERROR] s3-tests failed testcase summary:")
-for classname, name, duration, message in failures[:20]:
-    nodeid = f"{classname}::{name}" if classname else name
-    print(f"[ERROR] - {nodeid} ({duration}s): {message}")
-
-if len(failures) > 20:
-    print(f"[ERROR] - ... {len(failures) - 20} additional failed testcases omitted")
-PY
-}
-
 # =============================================================================
 # Test Classification Files
 # =============================================================================
@@ -322,6 +286,9 @@ Environment Variables:
  MAXFAIL                - Stop after N failures, 0 = never stop (default: 1)
  XDIST                  - Enable parallel execution with N workers (default: 0)
  TEST_SCOPE             - "implemented" (whitelist, default) or "all" (entire upstream suite)
+  S3_SHARD_COUNT         - Number of deterministic exact-node-ID shards (default: 1)
+  S3_SHARD_INDEX         - Zero-based shard index (default: 0)
+  TEST_TIMEOUT           - Per-test timeout in seconds (default: 300)
  S3TESTS_REPO           - s3-tests repository URL (default: https://github.com/ceph/s3-tests.git)
  S3TESTS_REV            - Pinned s3-tests commit; bump deliberately and reclassify test lists
  MARKEXPR               - pytest marker expression (default: no marker filtering)
@@ -982,9 +949,10 @@ mkdir -p "${ARTIFACTS_DIR}"
 XDIST_ARGS=""
 if [ "${XDIST}" != "0" ]; then
    # Add pytest-xdist to requirements.txt so tox installs it inside its virtualenv
-    echo "pytest-xdist" >> requirements.txt
+    grep -qxF "pytest-xdist" requirements.txt || echo "pytest-xdist" >> requirements.txt
    XDIST_ARGS="-n ${XDIST} --dist=loadgroup"
 fi
+grep -qxF "pytest-timeout" requirements.txt || echo "pytest-timeout" >> requirements.txt

 # Resolve config path (absolute path for tox)
 if [[ "${S3TESTS_CONF}" = /* ]]; then
@@ -1003,12 +971,69 @@ else
    PYTEST_SELECTION_ARGS=("${S3_TEST_FILE}")
 fi

+collect_nodeids() {
+    local output_path="$1"
+    shift
+    local collect_log="${output_path%.txt}.log"
+    local collect_rc=0
+    local node_prefix="${S3_TEST_FILE//./\\.}::"
+
+    set +e
+    S3TEST_CONF="${CONF_OUTPUT_PATH}" tox -- -q --collect-only "$@" 2>&1 | tee "${collect_log}"
+    collect_rc=${PIPESTATUS[0]}
+    set -e
+    if [ "${collect_rc}" -ne 0 ]; then
+        log_error "pytest collection failed with exit code ${collect_rc}"
+        return "${collect_rc}"
+    fi
+    grep -E "^${node_prefix}" "${collect_log}" > "${output_path}" || true
+    if [ ! -s "${output_path}" ]; then
+        log_error "pytest collection produced no S3 test node IDs"
+        return 1
+    fi
+}
+
+ALL_COLLECTED_NODEIDS="${ARTIFACTS_DIR}/all-collected-nodeids.txt"
+UNSHARDED_SELECTED_NODEIDS="${ARTIFACTS_DIR}/unsharded-selected-nodeids.txt"
+SELECTED_NODEIDS="${ARTIFACTS_DIR}/selected-nodeids.txt"
+collect_nodeids "${ALL_COLLECTED_NODEIDS}" "${S3_TEST_FILE}" -m "not rustfs_never_marker"
+python3 "${SCRIPT_DIR}/report_compat.py" \
+    --lists-dir "${SCRIPT_DIR}" \
+    --collected-nodeids "${ALL_COLLECTED_NODEIDS}" \
+    --check-classifications-only || {
+    log_error "S3 test classifications do not match pinned revision ${S3TESTS_REV}"
+    exit 1
+}
+if [[ "${TEST_SCOPE}" == "all" && -z "${TESTEXPR}" && "${MARKEXPR}" == "not rustfs_never_marker" ]]; then
+    cp "${ALL_COLLECTED_NODEIDS}" "${UNSHARDED_SELECTED_NODEIDS}"
+else
+    collect_nodeids "${UNSHARDED_SELECTED_NODEIDS}" "${PYTEST_SELECTION_ARGS[@]}" -m "${MARKEXPR}"
+fi
+
+if (( S3_SHARD_COUNT > 1 )); then
+    awk -v count="${S3_SHARD_COUNT}" -v shard_index="${S3_SHARD_INDEX}" \
+        '((NR - 1) % count) == shard_index' \
+        "${UNSHARDED_SELECTED_NODEIDS}" > "${SELECTED_NODEIDS}"
+    if [[ ! -s "${SELECTED_NODEIDS}" ]]; then
+        log_error "Shard ${S3_SHARD_INDEX}/${S3_SHARD_COUNT} selected no tests"
+        exit 1
+    fi
+    PYTEST_SELECTION_ARGS=()
+    while IFS= read -r nodeid; do
+        PYTEST_SELECTION_ARGS+=("${nodeid}")
+    done < "${SELECTED_NODEIDS}"
+    log_info "Selected shard ${S3_SHARD_INDEX}/${S3_SHARD_COUNT}: ${#PYTEST_SELECTION_ARGS[@]} exact cases"
+else
+    cp "${UNSHARDED_SELECTED_NODEIDS}" "${SELECTED_NODEIDS}"
+fi
+
 # Run tests from s3tests/functional
 set +e
 S3TEST_CONF="${CONF_OUTPUT_PATH}" \
    tox -- \
    -vv -ra --showlocals --tb=long \
    --maxfail="${MAXFAIL}" \
+    --timeout="${TEST_TIMEOUT}" \
    --junitxml="${ARTIFACTS_DIR}/junit.xml" \
    ${XDIST_ARGS} \
    "${PYTEST_SELECTION_ARGS[@]}" \
@@ -1033,19 +1058,22 @@ elif [ "${DEPLOY_MODE}" = "existing" ]; then
    echo "{\"host\": \"${S3_HOST}\", \"port\": ${S3_PORT}, \"mode\": \"existing\"}" > "${ARTIFACTS_DIR}/rustfs-${TEST_MODE}/inspect.json" || true
 fi

-# Step 11: Classification report (informational, never fails the run)
+# Step 11: Classification report and gate
 REPORT_SCRIPT="${SCRIPT_DIR}/report_compat.py"
-if [ -f "${REPORT_SCRIPT}" ] && [ -f "${ARTIFACTS_DIR}/junit.xml" ]; then
-    python3 "${REPORT_SCRIPT}" \
-        --junit "${ARTIFACTS_DIR}/junit.xml" \
-        --lists-dir "${SCRIPT_DIR}" \
-        --output "${ARTIFACTS_DIR}/compat-report.md" \
-        || log_warn "Compatibility report generation failed"
-fi
-
-if [ ${TEST_EXIT_CODE} -ne 0 ]; then
-    summarize_junit_failures "${ARTIFACTS_DIR}/junit.xml"
+REPORT_ARGS=(
+    --junit "${ARTIFACTS_DIR}/junit.xml"
+    --lists-dir "${SCRIPT_DIR}"
+    --collected-nodeids "${SELECTED_NODEIDS}"
+    --output "${ARTIFACTS_DIR}/compat-report.md"
+    --fail-on-regression
+)
+if [[ "${TEST_SCOPE}" == "all" ]]; then
+    REPORT_ARGS+=(--fail-on-unclassified)
 fi
+set +e
+python3 "${REPORT_SCRIPT}" "${REPORT_ARGS[@]}"
+REPORT_EXIT_CODE=$?
+set -e

 # Summary
 if [ ${TEST_EXIT_CODE} -eq 0 ]; then
@@ -1059,4 +1087,10 @@ else
    log_info "Check RustFS logs: ${ARTIFACTS_DIR}/rustfs-${TEST_MODE}/rustfs.log"
 fi

-exit ${TEST_EXIT_CODE}
+if [[ "${TEST_EXIT_CODE}" -ne 0 && "${TEST_EXIT_CODE}" -ne 1 ]]; then
+    exit "${TEST_EXIT_CODE}"
+fi
+if [[ "${TEST_SCOPE}" == "implemented" && "${TEST_EXIT_CODE}" -ne 0 ]]; then
+    exit "${TEST_EXIT_CODE}"
+fi
+exit "${REPORT_EXIT_CODE}"
@@ -0,0 +1,119 @@
+#!/usr/bin/env python3
+"""Regression tests for the S3 compatibility report."""
+
+from __future__ import annotations
+
+import importlib.util
+import subprocess
+import sys
+import tempfile
+import unittest
+from pathlib import Path
+
+
+REPORT_PATH = Path(__file__).with_name("report_compat.py")
+SPEC = importlib.util.spec_from_file_location("report_compat", REPORT_PATH)
+assert SPEC and SPEC.loader
+REPORT = importlib.util.module_from_spec(SPEC)
+SPEC.loader.exec_module(REPORT)
+
+
+class ReportCompatTests(unittest.TestCase):
+    def test_upstream_names_expose_incomplete_junit(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            directory = Path(tmp)
+            collected = directory / "collected.txt"
+            collected.write_text("s3tests/functional/test_s3.py::test_one[a]\ns3tests/functional/test_s3.py::test_one[b]\n")
+            junit = directory / "junit.xml"
+            junit.write_text('<testsuite><testcase name="test_one[a]" /></testsuite>')
+
+            expected = REPORT.load_collected_nodeids(collected)
+            results, _, _ = REPORT.parse_junit(junit)
+
+            self.assertEqual(expected - results.keys(), {"test_one[b]"})
+
+    def test_cli_fails_an_incomplete_sweep(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            directory = Path(tmp)
+            collected = directory / "collected.txt"
+            collected.write_text("s3tests/functional/test_s3.py::test_one\ns3tests/functional/test_s3.py::test_two\n")
+            junit = directory / "junit.xml"
+            junit.write_text('<testsuite><testcase name="test_one" /></testsuite>')
+            for filename in REPORT.LIST_FILES.values():
+                (directory / filename).write_text("")
+            (directory / "implemented_tests.txt").write_text("test_one\n")
+
+            result = subprocess.run(
+                [
+                    sys.executable,
+                    str(REPORT_PATH),
+                    "--junit",
+                    str(junit),
+                    "--lists-dir",
+                    str(directory),
+                    "--collected-nodeids",
+                    str(collected),
+                    "--fail-on-regression",
+                    "--fail-on-unclassified",
+                ],
+                check=False,
+                capture_output=True,
+                text=True,
+            )
+
+            self.assertEqual(result.returncode, 1)
+            self.assertIn("1 missing result(s)", result.stdout)
+
+    def test_preflight_rejects_missing_and_stale_classifications(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            directory = Path(tmp)
+            collected = directory / "collected.txt"
+            collected.write_text("s3tests/functional/test_s3.py::test_known[a]\ntest_new\n")
+            for filename in REPORT.LIST_FILES.values():
+                (directory / filename).write_text("")
+            (directory / "implemented_tests.txt").write_text("test_known\ntest_stale\ntest_stale\n")
+
+            result = subprocess.run(
+                [
+                    sys.executable,
+                    str(REPORT_PATH),
+                    "--lists-dir",
+                    str(directory),
+                    "--collected-nodeids",
+                    str(collected),
+                    "--check-classifications-only",
+                ],
+                check=False,
+                capture_output=True,
+                text=True,
+            )
+
+            self.assertEqual(result.returncode, 1)
+            self.assertIn("[UNCLASSIFIED] test_new", result.stdout)
+            self.assertIn("[STALE] test_stale", result.stdout)
+            self.assertIn("[INVALID] implemented_tests.txt has duplicates: test_stale", result.stdout)
+
+    def test_timeout_fails_even_when_test_is_excluded(self) -> None:
+        with tempfile.TemporaryDirectory() as tmp:
+            directory = Path(tmp)
+            junit = directory / "junit.xml"
+            junit.write_text(
+                '<testsuite><testcase name="test_slow"><failure message="Failed: Timeout (&gt;300.0s)" /></testcase></testsuite>'
+            )
+            for filename in REPORT.LIST_FILES.values():
+                (directory / filename).write_text("")
+            (directory / "excluded_tests.txt").write_text("test_slow\n")
+
+            result = subprocess.run(
+                [sys.executable, str(REPORT_PATH), "--junit", str(junit), "--lists-dir", str(directory)],
+                check=False,
+                capture_output=True,
+                text=True,
+            )
+
+            self.assertEqual(result.returncode, 1)
+            self.assertIn("1 timeout(s)", result.stdout)
+
+
+if __name__ == "__main__":
+    unittest.main()
@@ -11,8 +11,12 @@

 # Failed tests
 test_bucket_create_delete_bucket_ownership
+test_bucket_logging_request_id
 test_create_bucket_no_ownership_controls
 test_bucket_logging_owner
+test_head_object_404_with_policy_prefix
+test_multipart_reupload_checksum_and_etag
+test_multipart_upload_complete_without_create
 test_object_copy_not_owned_bucket
 test_bucket_policy_multipart
 test_post_object_upload_checksum
Author	SHA1	Message	Date
马登山	7beaa075b3	fix(quota): account compressed deletes by committed size	2026-08-22 11:03:02 +08:00
Zhengchao An	61b5edf16e	chore(ecstore): standardize 24 bare TODO/FIXME to tracked format (#6346 )	2026-08-22 07:09:32 +08:00
Zhengchao An	ab5d433fe1	feat(connect): add registration and credential rotation (#6353 )	2026-08-22 07:07:29 +08:00
Zhengchao An	bc07cfd115	ci: harden test selection and nightly coverage (#6341 )	2026-08-21 15:04:08 +00:00
Zhengchao An	bce5922aef	feat(connect): answer offline enrolment challenges without a network (#6335 )	2026-08-21 13:16:18 +00:00
cxymds	adb90fc6e1	fix(scanner): defer usage publication during pool recovery (#6333 ) * fix(scanner): defer usage publication during pool recovery * fix(scanner): preserve metrics when publication is deferred * fix(scanner): route test types through storage boundary * fix(scanner): keep cache floor deferred during movement	2026-08-21 17:32:59 +08:00
cxymds	cdfac5d7e3	fix(ecstore): avoid decommission walk deadline on backpressure (#6332 ) fix(ecstore): bound decommission background walks	2026-08-21 17:32:12 +08:00
houseme	ca4adea0c9	perf(server): trim internode REST compat stack (#6330 ) Co-authored-by: heihutu <heihutu@gmail.com>	2026-08-21 08:52:03 +00:00
GatewayJ	23a0f6324c	fix(iam): preserve MinIO permanent credentials in migration (#6328 ) * fix(iam): preserve MinIO permanent credentials in migration * test(iam): cover MinIO credential migration end to end	2026-08-21 15:34:25 +08:00
Zhengchao An	cdd9ab1124	fix(ci): update package checksums safely (#6329 )	2026-08-21 15:28:42 +08:00
houseme	122a69df65	feat(ecstore): tune fdatasync group wait budget (#6327 ) * feat(ecstore): tune fdatasync group wait budget Co-Authored-By: heihutu <heihutu@gmail.com> * test(ecstore): cover fdatasync wait budget contract Co-Authored-By: heihutu <heihutu@gmail.com> --------- Co-authored-by: heihutu <heihutu@gmail.com>	2026-08-21 14:08:21 +08:00
cxymds	dfeb732ac8	fix: make DeleteObjects idempotent for raw not-found errors (#6323 ) * fix: make DeleteObjects idempotent for raw not-found errors * fix: cover DeleteObjects raw not-found result dispatch	2026-08-21 03:12:37 +00:00
				`@@ -0,0 +1 @@`
				`sha256=9b9bc336b43b70d0e06e0adb5455bf035bb18945d85d60936eb6fe4d48e0e680`
				`@@ -0,0 +1 @@`
				`sha256=655a3f3c1d042e694339d15caba7580518320322d1bac0f09450b37e6c09e2e7`
				`@@ -0,0 +1 @@`
				`sha256=ec27cde6ce6400723c4b372bfbd2ac61709c744294e4810af765e8a808d8e31d`