fix(replication): restore MinIO tag filter behavior

rename_all_missing_source_still_warns and rename_all_real_failure_still_warns
assert that `warn_reliable_rename_failure` emitted its WARN, but that is a
single production callsite shared with tests that call rename_all *without*
installing a subscriber — rename_all_missing_source_returns_file_not_found,
two tests above, is one of them.

tracing caches each callsite's Interest process-globally and the first thread
to reach a callsite fixes that value; while at most one dispatcher is
registered, tracing-core derives it from the registering thread's own
subscriber, and registration is once-only. When the subscriber-less sibling
wins, the callsite is cached as Interest::never() and the WARN never fires,
so the assertion sees empty output:

    ordinary missing-source failures must keep the WARN, got:

Reproduced at 3/25 with `disk::os::tests::rename_all_missing_source` (both
tests), against 0/20 for the victim alone. Fixed by pinning callsite interest
inside warn_capture(), so every current and future user of that helper is
covered rather than just the two tests that happen to fail today.

pin_callsite_interest_for_test() moves from cluster::rpc::background_monitor
to a new crate-level test_tracing module: it is domain-neutral and now has
consumers in two unrelated subsystems, and disk::os should not have to reach
into a cluster::rpc test helper.

Verified: repro filter 0/30 (was 3/25); disk::os:: 0/12; cluster::rpc:: 0/12
and its poisoner pair 0/15, confirming the moved helper still holds.

Follow-up to #5438. Closes the last item in #5439.

.agents/skills/arch-checks/SKILL.md

@@ -10,16 +10,10 @@ never weaken a check to get green.
 
 ## `check_layer_dependencies.sh` — layer DAG in `rustfs/src`
 
-Enforces `composition (server, startup/init) → interface (admin,
-storage/ecfs, storage/s3_api) → app → infra`; no upward imports. Server source
-files are composition roots, while imports of their exported HTTP contracts
-are classified as interface dependencies. Known legacy violations live in
+Enforces `interface (admin, storage/ecfs, storage/s3_api) → app → infra`; no
+upward imports. Known legacy violations live in
 `scripts/layer-dependency-baseline.txt`.
 
-Dedicated `*_test.rs` and `tests/` modules are outside this production guard.
-Inline `#[cfg(test)]` imports remain checked under their source file's layer;
-move architecture-crossing test scaffolding into a dedicated test module.
-
 - **New violation**: restructure your change so the dependency points
   downward (move the shared type/function to the lower layer).
 - **You legitimately removed a baseline entry**: run

.github/workflows/ci.yml

@@ -156,16 +156,103 @@ jobs:
           github-token: ${{ secrets.GITHUB_TOKEN }}
           cache-save-if: ${{ github.ref == 'refs/heads/main' }}
 
+      - name: Prepare test evidence
+        run: |
+          mkdir -p artifacts/test-and-lint
+          {
+            echo "run_id=${GITHUB_RUN_ID}"
+            echo "job=${GITHUB_JOB}"
+            echo "runner=${RUNNER_NAME}"
+            echo "started_at=$(date --utc --iso-8601=seconds)"
+          } > artifacts/test-and-lint/run-metadata.txt
+
       # Clippy runs before the test pass: lint failures are the most common
       # CI-only breakage and should surface in minutes, not after 20+ minutes
       # of tests.
       - name: Run clippy lints
         run: cargo clippy --all-targets -- -D warnings
 
-      - name: Run tests
+      - name: Run nextest tests
+        env:
+          # Three concurrent workspace test links saturate the self-hosted
+          # runner's overlay I/O and can wedge Cargo until the 75m timeout.
+          CARGO_BUILD_JOBS: "2"
         run: |
-          cargo nextest run --profile ci --all --exclude e2e_test
-          cargo test --all --doc
+          mkdir -p artifacts/test-and-lint
+          # Evidence sampler for issue #5394: the post-mortem pgrep below runs
+          # only after `timeout` has already TERM'd the whole cargo process
+          # group, so it cannot name a wedged process. Sample system and
+          # process state every 60s instead; the last samples before the
+          # timeout show what was stuck (rustc, linker, build script, memory
+          # pressure, ...). The log rides along in the existing artifact.
+          (
+            while true; do
+              {
+                echo "=== $(date --utc --iso-8601=seconds)"
+                echo "--- load";  cat /proc/loadavg
+                echo "--- psi";   grep -H . /proc/pressure/* 2>/dev/null || true
+                echo "--- mem";   free -m
+                echo "--- disk";  df -h / /home/runner 2>/dev/null || df -h /
+                echo "--- top-rss"
+                ps -eo pid,ppid,stat,etime,rss,pcpu,args --sort=-rss | head -15
+                echo "--- build/test processes"
+                ps -eo pid,ppid,stat,etime,rss,pcpu,args | grep -E '[c]argo|[r]ustc|[n]extest|[c]ollect2|rust-ll[d]|[b]uild-script|deps[/]' || true
+                echo "--- d-state (uninterruptible IO)"
+                ps -eo pid,stat,etime,args | awk 'NR > 1 && $2 ~ /D/' || true
+                echo
+              } >> artifacts/test-and-lint/sampler.log 2>&1 || true
+              sleep 60
+            done
+          ) &
+          sampler_pid=$!
+          trap 'kill "${sampler_pid}" 2>/dev/null || true' EXIT
+          set +e
+          NEXTEST_HIDE_PROGRESS_BAR=1 timeout --verbose --signal=TERM --kill-after=30s 75m \
+            cargo nextest run --profile ci --all --exclude e2e_test \
+            --status-level all --final-status-level all \
+            2>&1 | tee artifacts/test-and-lint/nextest.log
+          status=${PIPESTATUS[0]}
+          {
+            echo "command=cargo nextest run --profile ci --all --exclude e2e_test"
+            echo "exit_status=${status}"
+            echo "finished_at=$(date --utc --iso-8601=seconds)"
+            echo
+            echo "Remaining test-related processes:"
+            pgrep -af 'cargo|nextest|target/.*/deps/' || true
+            echo
+            echo "Kernel OOM / kill events:"
+            dmesg -T 2>/dev/null | grep -iE 'oom|out of memory|killed process' | tail -20 || true
+          } > artifacts/test-and-lint/nextest-diagnostics.txt
+          exit "${status}"
+
+      - name: Run documentation tests
+        run: |
+          mkdir -p artifacts/test-and-lint
+          set +e
+          timeout --verbose --signal=TERM --kill-after=30s 15m \
+            cargo test --all --doc \
+            2>&1 | tee artifacts/test-and-lint/doctest.log
+          status=${PIPESTATUS[0]}
+          {
+            echo "command=cargo test --all --doc"
+            echo "exit_status=${status}"
+            echo "finished_at=$(date --utc --iso-8601=seconds)"
+            echo
+            echo "Remaining test-related processes:"
+            pgrep -af 'cargo|rustdoc|target/.*/deps/' || true
+          } > artifacts/test-and-lint/doctest-diagnostics.txt
+          exit "${status}"
+
+      - name: Upload test reports and diagnostics
+        if: always()
+        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
+        with:
+          name: junit-test-and-lint-${{ github.run_number }}
+          path: |
+            target/nextest/ci/junit.xml
+            artifacts/test-and-lint
+          retention-days: 3
+          if-no-files-found: error
 
       # rustfs/backlog#1289: fail if a seed rule's log anchor no longer exists
       # verbatim in the source tree (log message drifted without updating the
@@ -174,15 +261,6 @@ jobs:
       - name: Check log-analyzer rule anchors
         run: ./scripts/check_log_analyzer_rules.sh
 
-      - name: Upload test junit report
-        if: always()
-        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
-        with:
-          name: junit-test-and-lint-${{ github.run_number }}
-          path: target/nextest/ci/junit.xml
-          retention-days: 3
-          if-no-files-found: ignore
-
       # Explicit gate for migration-critical suites. These tests already ran in
       # the full nextest pass above; a single filtered nextest invocation keeps
       # the named gate without rebuilding or re-running them one package at a time.
@@ -328,7 +406,7 @@ jobs:
           github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Build debug binary
-        run: cargo build -p rustfs --bins
+        run: cargo build -p rustfs --bins --features e2e-test-hooks
 
       - name: Upload debug binary
         uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6
@@ -358,7 +436,7 @@ jobs:
           github-token: ${{ secrets.GITHUB_TOKEN }}
 
       - name: Build debug binary with rio-v2
-        run: cargo build -p rustfs --bins --features rio-v2
+        run: cargo build -p rustfs --bins --features rio-v2,e2e-test-hooks
 
       - name: Upload debug binary
         uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6

.github/workflows/docker.yml

@@ -66,7 +66,7 @@ env:
   CARGO_TERM_COLOR: always
   REGISTRY_DOCKERHUB: rustfs/rustfs
   REGISTRY_GHCR: ghcr.io/${{ github.repository }}
-  REGISTRY_QUAY: quay.io/${{ secrets.QUAY_USERNAME }}/rustfs
+  REGISTRY_QUAY: quay.io/rustfs/rustfs
   DOCKER_PLATFORMS: linux/amd64,linux/arm64
 
 jobs:

Cargo.lock

@@ -3671,12 +3671,13 @@ dependencies = [
  "flatbuffers",
  "flate2",
  "futures",
+ "hex",
  "http 1.4.2",
  "http-body-util",
  "hyper",
  "hyper-util",
  "local-ip-address",
- "md5",
+ "md-5 0.11.0",
  "opentelemetry-proto",
  "prost 0.14.4",
  "rand 0.10.2",
@@ -5644,9 +5645,9 @@ dependencies = [
 
 [[package]]
 name = "lazy-regex"
-version = "3.6.0"
+version = "3.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6bae91019476d3ec7147de9aa291cadb6d870abf2f3015d2da73a90325ac1496"
+checksum = "4994ba703f78b083e2f7946dac9251abd83fd43a0365f030e99b69be5b4b9ef9"
 dependencies = [
  "lazy-regex-proc_macros",
  "once_cell",
@@ -5655,9 +5656,9 @@ dependencies = [
 
 [[package]]
 name = "lazy-regex-proc_macros"
-version = "3.6.0"
+version = "3.6.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4de9c1e1439d8b7b3061b2d209809f447ca33241733d9a3c01eabf2dc8d94358"
+checksum = "fd97232314824e6dbef1918a871bb93f51070455e3715bf26e19a6d01aa977a0"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -7236,12 +7237,6 @@ dependencies = [
  "path-dedot",
 ]
 
-[[package]]
-name = "path-clean"
-version = "1.0.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "17359afc20d7ab31fdb42bb844c8b3bb1dabd7dcf7e68428492da7f16966fcef"
-
 [[package]]
 name = "path-dedot"
 version = "4.0.1"
@@ -8890,7 +8885,7 @@ dependencies = [
  "libmimalloc-sys",
  "libsystemd",
  "matchit 0.9.2",
- "md5",
+ "md-5 0.11.0",
  "metrics",
  "metrics-util",
  "mimalloc",
@@ -9095,7 +9090,6 @@ name = "rustfs-data-usage"
 version = "1.0.0-beta.11"
 dependencies = [
  "async-trait",
- "path-clean",
  "rmp-serde",
  "rustfs-filemeta",
  "serde",
@@ -9211,11 +9205,14 @@ dependencies = [
  "tonic",
  "tower",
  "tracing",
+ "tracing-core",
  "tracing-opentelemetry",
  "tracing-subscriber",
  "url",
  "urlencoding",
  "uuid",
+ "winapi-util",
+ "windows-sys 0.61.2",
  "xxhash-rust",
 ]
 
@@ -9449,9 +9446,10 @@ dependencies = [
  "async-trait",
  "base64 0.23.0",
  "chacha20poly1305",
+ "hex",
  "insta",
  "jiff",
- "md5",
+ "md-5 0.11.0",
  "moka",
  "rand 0.10.2",
  "reqwest",
@@ -9460,6 +9458,7 @@ dependencies = [
  "serde",
  "serde_json",
  "sha2 0.11.0",
+ "subtle",
  "temp-env",
  "tempfile",
  "thiserror 2.0.19",
@@ -9717,7 +9716,7 @@ dependencies = [
  "hyper-util",
  "ipnetwork",
  "libunftp",
- "md5",
+ "md-5 0.11.0",
  "percent-encoding",
  "proptest",
  "quick-xml",
@@ -10128,6 +10127,7 @@ dependencies = [
  "bytes",
  "convert_case 0.11.0",
  "crc-fast",
+ "criterion",
  "flate2",
  "futures",
  "hex-simd",
@@ -11829,6 +11829,7 @@ dependencies = [
  "futures-util",
  "libc",
  "pin-project-lite",
+ "slab",
  "tokio",
 ]
 

Cargo.toml

@@ -265,7 +265,6 @@ memmap2 = "0.9.11"
 lz4 = "1.28.1"
 matchit = "0.9.2"
 md-5 = "0.11.0"
-md5 = "0.8.1"
 mime_guess = "2.0.5"
 moka = { version = "0.12.15" }
 netif = "0.1.6"
@@ -304,6 +303,7 @@ test-case = "3.3.1"
 thiserror = "2.0.19"
 tracing = { version = "0.1.44" }
 tracing-appender = "0.2.5"
+tracing-core = "0.1.36"
 tracing-error = "0.2.1"
 tracing-opentelemetry = { version = "0.33" }
 tracing-subscriber = { version = "0.3.23" }
@@ -314,7 +314,9 @@ uuid = { version = "1.24.0" }
 vaultrs = { version = "0.8.0" }
 tar = "0.4.46"
 walkdir = "2.5.0"
+winapi-util = "0.1.11"
 windows = { version = "0.62.2" }
+windows-sys = "0.61.2"
 xxhash-rust = { version = "0.8.18" }
 zip = "8.6.0"
 zstd = "0.13.3"

crates/checksums/Cargo.toml

@@ -25,6 +25,9 @@ keywords = ["checksum-calculation", "verification", "integrity", "authenticity",
 categories = ["web-programming", "development-tools", "network-programming"]
 documentation = "https://docs.rs/rustfs-checksums/latest/rustfs_checksum/"
 
+[lints]
+workspace = true
+
 [dependencies]
 bytes = { workspace = true, features = ["serde"] }
 crc-fast = { workspace = true }

crates/common/src/metrics.rs

@@ -1114,6 +1114,8 @@ pub struct ScannerLastMinute {
 pub struct ScannerMetricsReport {
     pub collected_at: DateTime<Utc>,
     pub current_cycle: u64,
+    #[serde(default)]
+    pub current_cycle_active: bool,
     pub current_started: DateTime<Utc>,
     pub cycles_completed_at: Vec<DateTime<Utc>>,
     pub ongoing_buckets: usize,
@@ -2051,7 +2053,7 @@ impl Metrics {
     pub fn record_scanner_transition_failed(&self, count: u64) {
         self.scanner_transition_failed.fetch_add(count, Ordering::Relaxed);
         self.record_scanner_source_failed(ScannerWorkSource::Lifecycle, count);
-        if !self.current_scan_cycle_work_active.load(Ordering::Relaxed) {
+        if !self.current_scan_cycle_work_active.load(Ordering::Acquire) {
             self.record_last_cycle_scanner_source_work(ScannerWorkSource::Lifecycle, ScannerSourceWorkUpdate::failed(count));
         }
     }
@@ -2336,6 +2338,21 @@ impl Metrics {
         *self.cycle_info.write().await = cycle;
     }
 
+    /// Publish a scanner cycle and its work-accounting baseline as one state transition.
+    pub async fn start_scan_cycle_work_with_cycle(&self, cycle: CurrentCycle) -> ScanCycleWorkSnapshot {
+        let mut current_cycle = self.cycle_info.write().await;
+        let snapshot = self.start_scan_cycle_work();
+        *current_cycle = Some(cycle);
+        snapshot
+    }
+
+    /// Publish the completed work snapshot and idle cycle state as one state transition.
+    pub async fn finish_scan_cycle_work_with_cycle(&self, start: ScanCycleWorkSnapshot, cycle: CurrentCycle) {
+        let mut current_cycle = self.cycle_info.write().await;
+        self.finish_scan_cycle_work(start);
+        *current_cycle = Some(cycle);
+    }
+
     /// Read the current cycle record.
     pub async fn get_cycle(&self) -> Option<CurrentCycle> {
         self.cycle_info.read().await.clone()
@@ -2464,7 +2481,7 @@ impl Metrics {
             &self.current_scan_cycle_replication_repair_work_start,
             &replication_repair_snapshot,
         );
-        self.current_scan_cycle_work_active.store(true, Ordering::Relaxed);
+        self.current_scan_cycle_work_active.store(true, Ordering::Release);
         snapshot
     }
 
@@ -2476,11 +2493,11 @@ impl Metrics {
         self.record_scan_cycle_work(work);
         self.record_scan_cycle_source_work(&source_work);
         self.record_scan_cycle_replication_repair_work(&replication_repair_work);
-        self.current_scan_cycle_work_active.store(false, Ordering::Relaxed);
+        self.current_scan_cycle_work_active.store(false, Ordering::Release);
     }
 
     pub fn current_scan_cycle_has_unresolved_heal_work(&self) -> bool {
-        if !self.current_scan_cycle_work_active.load(Ordering::Relaxed) {
+        if !self.current_scan_cycle_work_active.load(Ordering::Acquire) {
             return false;
         }
 
@@ -2746,13 +2763,41 @@ impl Metrics {
     pub async fn report(&self) -> ScannerMetricsReport {
         let mut m = ScannerMetricsReport::default();
 
-        let has_cycle = if let Some(cycle) = self.get_cycle().await {
-            m.current_cycle = cycle.current;
-            m.cycles_completed_at = cycle.cycle_completed;
-            m.current_started = cycle.started;
-            true
-        } else {
-            false
+        let has_cycle = {
+            let cycle = self.cycle_info.read().await;
+            let has_cycle = if let Some(cycle) = cycle.as_ref() {
+                m.current_cycle = cycle.current;
+                m.cycles_completed_at = cycle.cycle_completed.clone();
+                m.current_started = cycle.started;
+                true
+            } else {
+                false
+            };
+            m.current_cycle_active = self.current_scan_cycle_work_active.load(Ordering::Acquire);
+            if m.current_cycle_active {
+                let current_work = self.scan_cycle_work_since(self.current_scan_cycle_work_start());
+                let current_source_work = self.scanner_source_work_since(&self.current_scan_cycle_source_work_start_values());
+                let current_replication_repair_work =
+                    self.scanner_replication_repair_work_since(&self.current_scan_cycle_replication_repair_work_start_values());
+                m.current_cycle_objects_scanned = current_work.objects_scanned;
+                m.current_cycle_directories_scanned = current_work.directories_scanned;
+                m.current_cycle_bucket_drive_scans = current_work.bucket_drive_scans;
+                m.current_cycle_bucket_drive_failures = current_work.bucket_drive_failures;
+                m.current_cycle_yield_events = current_work.yield_events;
+                m.current_cycle_yield_duration_seconds = current_work.yield_duration_millis as f64 / 1000.0;
+                m.current_cycle_throttle_sleep_events = current_work.throttle_sleep_events;
+                m.current_cycle_throttle_sleep_duration_seconds = current_work.throttle_sleep_duration_millis as f64 / 1000.0;
+                m.current_cycle_ilm_actions = current_work.ilm_actions;
+                m.current_cycle_lifecycle_expiry_actions = current_work.lifecycle_expiry_actions;
+                m.current_cycle_lifecycle_transition_actions = current_work.lifecycle_transition_actions;
+                m.current_cycle_heal_objects = current_work.heal_objects;
+                m.current_cycle_replication_checks = current_work.replication_checks;
+                m.current_cycle_usage_saves = current_work.usage_saves;
+                m.current_cycle_source_work = self.scanner_source_work_snapshots(&current_source_work);
+                m.current_cycle_replication_repair =
+                    self.scanner_replication_repair_work_snapshots(&current_replication_repair_work);
+            }
+            has_cycle
         };
 
         if !has_cycle && let Some(init_time) = crate::get_global_init_time().await {
@@ -2793,28 +2838,6 @@ impl Metrics {
         m.current_disk_scan_concurrency_limit = disk_scan_concurrency_limit;
         m.current_disk_bucket_scans_queued = disk_bucket_scans_queued;
         m.current_disk_bucket_scans_active = disk_bucket_scans_active;
-        if self.current_scan_cycle_work_active.load(Ordering::Relaxed) {
-            let current_work = self.scan_cycle_work_since(self.current_scan_cycle_work_start());
-            let current_source_work = self.scanner_source_work_since(&self.current_scan_cycle_source_work_start_values());
-            let current_replication_repair_work =
-                self.scanner_replication_repair_work_since(&self.current_scan_cycle_replication_repair_work_start_values());
-            m.current_cycle_objects_scanned = current_work.objects_scanned;
-            m.current_cycle_directories_scanned = current_work.directories_scanned;
-            m.current_cycle_bucket_drive_scans = current_work.bucket_drive_scans;
-            m.current_cycle_bucket_drive_failures = current_work.bucket_drive_failures;
-            m.current_cycle_yield_events = current_work.yield_events;
-            m.current_cycle_yield_duration_seconds = current_work.yield_duration_millis as f64 / 1000.0;
-            m.current_cycle_throttle_sleep_events = current_work.throttle_sleep_events;
-            m.current_cycle_throttle_sleep_duration_seconds = current_work.throttle_sleep_duration_millis as f64 / 1000.0;
-            m.current_cycle_ilm_actions = current_work.ilm_actions;
-            m.current_cycle_lifecycle_expiry_actions = current_work.lifecycle_expiry_actions;
-            m.current_cycle_lifecycle_transition_actions = current_work.lifecycle_transition_actions;
-            m.current_cycle_heal_objects = current_work.heal_objects;
-            m.current_cycle_replication_checks = current_work.replication_checks;
-            m.current_cycle_usage_saves = current_work.usage_saves;
-            m.current_cycle_source_work = self.scanner_source_work_snapshots(&current_source_work);
-            m.current_cycle_replication_repair = self.scanner_replication_repair_work_snapshots(&current_replication_repair_work);
-        }
         let last_cycle_result = self.last_scan_cycle_result.load(Ordering::Relaxed);
         m.last_cycle_result = scan_cycle_result_label(last_cycle_result).to_string();
         m.last_cycle_result_code = last_cycle_result as u64;
@@ -4142,6 +4165,8 @@ mod tests {
 
         let report = metrics.report().await;
 
+        assert!(report.current_cycle_active);
+        assert_eq!(report.current_cycle, 0);
         assert_eq!(report.current_cycle_objects_scanned, 7);
         assert_eq!(report.current_cycle_directories_scanned, 3);
         assert_eq!(report.current_cycle_bucket_drive_scans, 2);
@@ -4158,6 +4183,8 @@ mod tests {
         metrics.finish_scan_cycle_work(start);
         let report = metrics.report().await;
 
+        assert!(!report.current_cycle_active);
+        assert_eq!(report.current_cycle, 0);
         assert_eq!(report.current_cycle_objects_scanned, 0);
         assert_eq!(report.current_cycle_directories_scanned, 0);
         assert_eq!(report.current_cycle_bucket_drive_scans, 0);
@@ -4184,6 +4211,91 @@ mod tests {
         assert_eq!(report.last_cycle_usage_saves, 2);
     }
 
+    #[tokio::test]
+    async fn scan_cycle_activity_and_cycle_state_publish_together() {
+        let metrics = Metrics::new();
+        let cycle_started = Utc::now() - chrono::Duration::seconds(5);
+        let active_cycle = CurrentCycle {
+            current: 12,
+            next: 13,
+            started: cycle_started,
+            ..Default::default()
+        };
+
+        let cycle_state = metrics.cycle_info.read().await;
+        let mut start_transition = Box::pin(metrics.start_scan_cycle_work_with_cycle(active_cycle));
+        let waker = std::task::Waker::noop();
+        let mut context = std::task::Context::from_waker(waker);
+        assert!(start_transition.as_mut().poll(&mut context).is_pending());
+        assert!(!metrics.current_scan_cycle_work_active.load(Ordering::Acquire));
+        drop(cycle_state);
+
+        let start = start_transition.await;
+        let active = metrics.report().await;
+        assert!(active.current_cycle_active);
+        assert_eq!(active.current_cycle, 12);
+        assert_eq!(active.current_started, cycle_started);
+
+        let idle_cycle = CurrentCycle {
+            current: 0,
+            next: 13,
+            started: cycle_started,
+            ..Default::default()
+        };
+        let cycle_state = metrics.cycle_info.read().await;
+        let mut finish_transition = Box::pin(metrics.finish_scan_cycle_work_with_cycle(start, idle_cycle));
+        assert!(finish_transition.as_mut().poll(&mut context).is_pending());
+        assert!(metrics.current_scan_cycle_work_active.load(Ordering::Acquire));
+        drop(cycle_state);
+
+        finish_transition.await;
+        let idle = metrics.report().await;
+        assert!(!idle.current_cycle_active);
+        assert_eq!(idle.current_cycle, 0);
+    }
+
+    #[tokio::test]
+    async fn report_keeps_cycle_identity_and_work_in_one_snapshot() {
+        let metrics = Metrics::new();
+        let cycle_ten = CurrentCycle {
+            current: 10,
+            next: 11,
+            started: Utc::now() - chrono::Duration::seconds(10),
+            ..Default::default()
+        };
+        let cycle_ten_start = metrics.start_scan_cycle_work_with_cycle(cycle_ten.clone()).await;
+        metrics.operations[Metric::ScanObject as usize].store(1, Ordering::Relaxed);
+
+        let paths = metrics.current_paths.write().await;
+        let mut report = Box::pin(metrics.report());
+        let waker = std::task::Waker::noop();
+        let mut context = std::task::Context::from_waker(waker);
+        assert!(report.as_mut().poll(&mut context).is_pending());
+
+        metrics
+            .finish_scan_cycle_work_with_cycle(cycle_ten_start, CurrentCycle { current: 0, ..cycle_ten })
+            .await;
+        let cycle_eleven_start = metrics
+            .start_scan_cycle_work_with_cycle(CurrentCycle {
+                current: 11,
+                next: 12,
+                started: Utc::now(),
+                ..Default::default()
+            })
+            .await;
+        metrics.operations[Metric::ScanObject as usize].store(101, Ordering::Relaxed);
+
+        drop(paths);
+        let snapshot = report.await;
+
+        assert_eq!(snapshot.current_cycle, 10);
+        assert_eq!(snapshot.current_cycle_objects_scanned, 1);
+
+        metrics
+            .finish_scan_cycle_work_with_cycle(cycle_eleven_start, CurrentCycle::default())
+            .await;
+    }
+
     #[tokio::test]
     async fn scanner_cycle_ilm_actions_ignore_global_ilm_work() {
         let metrics = Metrics::new();

crates/concurrency/Cargo.toml

@@ -10,6 +10,9 @@ description = "Shared concurrency contract types for RustFS - workload admission
 keywords = ["rustfs", "concurrency", "admission", "backpressure", "workers"]
 categories = ["concurrency", "filesystem"]
 
+[lints]
+workspace = true
+
 [dependencies]
 # Internal crates
 rustfs-io-core = { workspace = true }

crates/config/src/constants/internode.rs

@@ -158,17 +158,33 @@ pub const DEFAULT_INTERNODE_RPC_BODY_DIGEST_STRICT: bool = false;
 // rolling upgrades until the fleet-wide body-digest fallback counter reads zero.
 const _: () = assert!(!DEFAULT_INTERNODE_RPC_BODY_DIGEST_STRICT);
 
-/// Capacity (distinct nonces) of the process-local internode RPC replay cache that enforces
-/// one-time consumption of body-bound v2 signatures.
+/// Require the replay-scoped internode RPC signature after the fleet has converged on it.
 ///
-/// The cache retains each nonce for the ~10-minute signature freshness envelope, so the steady
-/// state holds roughly `mutating RPS x 601s` entries; the default sustains ~1,700 body-bound
-/// mutating RPCs per second (about 120 MiB worst case, allocated only under sustained load).
-/// Overflow fails closed — legitimate signed traffic is the only thing that can fill the cache
-/// (replays are rejected before insertion, and an attacker cannot mint valid nonces without the
-/// shared secret) — and increments
+/// The default keeps v1/v2 peers available during a rolling upgrade. Operators may set this only
+/// after `rustfs_system_network_internode_replay_scope_fallback_total` remains zero for a full
+/// release window. The node still accepts a v2-authenticated `Ping` carrying an epoch challenge:
+/// that narrowly scoped bootstrap lets an upgraded client learn the receiving process epoch and
+/// immediately retry with the replay-scoped signature after a peer restart.
+pub const ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT: &str = "RUSTFS_INTERNODE_RPC_REPLAY_SCOPE_STRICT";
+pub const DEFAULT_INTERNODE_RPC_REPLAY_SCOPE_STRICT: bool = false;
+
+// Compile-time invariant: mixed-version clusters must remain available until operators make the
+// observed fallback counter an explicit strictness decision.
+const _: () = assert!(!DEFAULT_INTERNODE_RPC_REPLAY_SCOPE_STRICT);
+
+/// Capacity (distinct nonces) of the process-local internode RPC replay cache that enforces
+/// one-time consumption of authenticated RPC signatures.
+///
+/// The cache retains each nonce for the ~10-minute signature freshness envelope. Once peers use
+/// replay-scoped v3 authentication, every authenticated RPC consumes one entry, so the steady
+/// state holds roughly `authenticated RPC RPS x 601s` entries. The default sustains about 1,700
+/// authenticated RPCs per second (about 120 MiB worst case, allocated only under sustained load);
+/// operators must size it for the node's aggregate peak RPC rate before enabling strict replay
+/// scope. Overflow fails closed — legitimate signed traffic is the only thing that can fill the
+/// cache (replays are rejected before insertion, and an attacker cannot mint valid nonces without
+/// the shared secret) — and increments
 /// `rustfs_system_network_internode_replay_cache_overflow_total`, so a sustained non-zero overflow
-/// counter means this capacity is undersized for the node's peak mutation rate.
+/// counter means this capacity is undersized for the node's peak authenticated RPC rate.
 pub const ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY: &str = "RUSTFS_INTERNODE_RPC_REPLAY_CACHE_CAPACITY";
 pub const DEFAULT_INTERNODE_RPC_REPLAY_CACHE_CAPACITY: usize = 1_048_576;
 
@@ -354,6 +370,12 @@ mod tests {
         assert_eq!(ENV_INTERNODE_RPC_BODY_DIGEST_STRICT, "RUSTFS_INTERNODE_RPC_BODY_DIGEST_STRICT");
     }
 
+    #[test]
+    fn internode_replay_scope_strict_env_name_is_stable() {
+        // The fail-open default invariant is asserted at compile time next to the definition.
+        assert_eq!(ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT, "RUSTFS_INTERNODE_RPC_REPLAY_SCOPE_STRICT");
+    }
+
     #[test]
     fn internode_replay_cache_capacity_defaults_and_env_name() {
         assert_eq!(ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY, "RUSTFS_INTERNODE_RPC_REPLAY_CACHE_CAPACITY");

crates/config/src/constants/object.rs

@@ -116,6 +116,27 @@ pub const ENV_OBJECT_GET_SKIP_BITROT_VERIFY: &str = "RUSTFS_OBJECT_GET_SKIP_BITR
 /// Default: bitrot verification is enabled on GetObject reads (do not skip).
 pub const DEFAULT_OBJECT_GET_SKIP_BITROT_VERIFY: bool = false;
 
+/// Request writing the complete remote-tier version state into object metadata.
+///
+/// This remains ineffective until
+/// [`ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED`] is also enabled.
+pub const ENV_TIER_REMOTE_VERSION_STATE_WRITE: &str = "RUSTFS_TIER_REMOTE_VERSION_STATE_WRITE";
+pub const DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE: bool = false;
+
+/// Operator-attested fleet-wide confirmation for
+/// [`ENV_TIER_REMOTE_VERSION_STATE_WRITE`].
+///
+/// This flag is an operational contract, not automatic capability discovery.
+/// Operators may enable it only after every node that can write or read
+/// transitioned object metadata supports the remote version-state schema and
+/// semantics. Keeping the confirmation separate makes a single-node request or
+/// a writer whose local opt-in is removed fail closed.
+pub const ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED: &str = "RUSTFS_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED";
+pub const DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED: bool = false;
+
+const _: () = assert!(!DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE);
+const _: () = assert!(!DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED);
+
 // =============================================================================
 // Concurrent Request Fix - Timeout and Backpressure Configuration
 // =============================================================================
@@ -617,3 +638,15 @@ pub const ENV_OBJECT_IO_RANDOM_READAHEAD_DISABLE_CONCURRENCY: &str = "RUSTFS_OBJ
 
 /// Default read-ahead disable concurrency threshold: 4.
 pub const DEFAULT_OBJECT_IO_RANDOM_READAHEAD_DISABLE_CONCURRENCY: usize = 4;
+
+#[cfg(test)]
+mod remote_version_state_tests {
+    #[test]
+    fn remote_version_state_gate_uses_stable_environment_names() {
+        assert_eq!(super::ENV_TIER_REMOTE_VERSION_STATE_WRITE, "RUSTFS_TIER_REMOTE_VERSION_STATE_WRITE");
+        assert_eq!(
+            super::ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED,
+            "RUSTFS_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED"
+        );
+    }
+}

crates/data-usage/Cargo.toml

@@ -29,7 +29,6 @@ workspace = true
 
 [dependencies]
 serde = { workspace = true, features = ["derive"] }
-path-clean = { workspace = true }
 rmp-serde = { workspace = true }
 async-trait = { workspace = true }
 rustfs-filemeta = { workspace = true }

crates/data-usage/src/data_usage.rs

@@ -12,12 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use path_clean::PathClean;
 use serde::{Deserialize, Serialize};
 use std::{
     collections::{HashMap, HashSet},
     hash::{DefaultHasher, Hash, Hasher},
-    path::Path,
     time::{Duration, SystemTime},
 };
 
@@ -334,7 +332,7 @@ impl<'de> Deserialize<'de> for SizeHistogram {
 impl SizeHistogram {
     pub fn add(&mut self, size: u64) {
         let intervals = [
-            (0, 1024),                                  // LESS_THAN_1024_B
+            (0, 1024 - 1),                              // LESS_THAN_1024_B
             (1024, 64 * 1024 - 1),                      // BETWEEN_1024_B_AND_64_KB
             (64 * 1024, 256 * 1024 - 1),                // BETWEEN_64_KB_AND_256_KB
             (256 * 1024, 512 * 1024 - 1),               // BETWEEN_256_KB_AND_512_KB
@@ -362,7 +360,7 @@ impl SizeHistogram {
         // the sub-ranges in [1 KiB, 512 KiB).
         const ONE_MIB: u64 = 1024 * 1024;
         let intervals = [
-            (0, 1024),                          // LESS_THAN_1024_B
+            (0, 1024 - 1),                      // LESS_THAN_1024_B
             (1024, 64 * 1024 - 1),              // BETWEEN_1024_B_AND_64_KB
             (64 * 1024, 256 * 1024 - 1),        // BETWEEN_64_KB_AND_256_KB
             (256 * 1024, 512 * 1024 - 1),       // BETWEEN_256_KB_AND_512_KB
@@ -506,8 +504,35 @@ pub struct ReplicationStats {
 }
 
 impl ReplicationStats {
+    pub fn is_empty(&self) -> bool {
+        let Self {
+            pending_size,
+            replicated_size,
+            failed_size,
+            failed_count,
+            pending_count,
+            missed_threshold_size,
+            after_threshold_size,
+            missed_threshold_count,
+            after_threshold_count,
+            replicated_count,
+        } = self;
+
+        *pending_size == 0
+            && *replicated_size == 0
+            && *failed_size == 0
+            && *failed_count == 0
+            && *pending_count == 0
+            && *missed_threshold_size == 0
+            && *after_threshold_size == 0
+            && *missed_threshold_count == 0
+            && *after_threshold_count == 0
+            && *replicated_count == 0
+    }
+
+    #[deprecated(note = "use is_empty instead")]
     pub fn empty(&self) -> bool {
-        self.replicated_size == 0 && self.failed_size == 0 && self.failed_count == 0
+        self.is_empty()
     }
 }
 
@@ -520,16 +545,19 @@ pub struct ReplicationAllStats {
 }
 
 impl ReplicationAllStats {
+    pub fn is_empty(&self) -> bool {
+        let Self {
+            replica_size,
+            replica_count,
+            targets,
+        } = self;
+
+        *replica_size == 0 && *replica_count == 0 && targets.values().all(ReplicationStats::is_empty)
+    }
+
+    #[deprecated(note = "use is_empty instead")]
     pub fn empty(&self) -> bool {
-        if self.replica_size != 0 && self.replica_count != 0 {
-            return false;
-        }
-        for v in self.targets.values() {
-            if !v.empty() {
-                return false;
-            }
-        }
-        true
+        self.is_empty()
     }
 }
 
@@ -783,7 +811,7 @@ impl DataUsageCache {
                     return Some(root);
                 }
                 let mut flat = self.flatten(&root);
-                if flat.replication_stats.as_ref().is_some_and(|stats| stats.empty()) {
+                if flat.replication_stats.as_ref().is_some_and(ReplicationAllStats::is_empty) {
                     flat.replication_stats = None;
                 }
                 Some(flat)
@@ -1080,9 +1108,39 @@ fn mark(duc: &DataUsageCache, entry: &DataUsageEntry, found: &mut HashSet<String
     }
 }
 
-/// Hash a path for data usage caching
+fn clean_data_usage_path(data: &str) -> String {
+    let rooted = data.starts_with('/');
+    let mut parts = Vec::new();
+
+    for part in data.split('/') {
+        match part {
+            "" | "." => {}
+            ".." => {
+                if parts.last().is_some_and(|last| *last != "..") {
+                    parts.pop();
+                } else if !rooted {
+                    parts.push(part);
+                }
+            }
+            _ => parts.push(part),
+        }
+    }
+
+    let clean = parts.join("/");
+    match (rooted, clean.is_empty()) {
+        (true, true) => "/".to_string(),
+        (true, false) => format!("/{clean}"),
+        (false, true) => ".".to_string(),
+        (false, false) => clean,
+    }
+}
+
+/// Hash a slash-separated path for data usage caching.
+///
+/// Cache identifiers are persisted and exchanged across nodes, so their
+/// normalization must not depend on the host operating system.
 pub fn hash_path(data: &str) -> DataUsageHash {
-    DataUsageHash(Path::new(&data).clean().to_string_lossy().to_string())
+    DataUsageHash(clean_data_usage_path(data))
 }
 
 impl DataUsageInfo {
@@ -1467,6 +1525,23 @@ mod tests {
         buckets_count: u64,
     }
 
+    #[test]
+    fn hash_path_uses_portable_slash_semantics() {
+        for (input, expected) in [
+            ("", "."),
+            (".", "."),
+            ("/", "/"),
+            ("//bucket///prefix/", "/bucket/prefix"),
+            ("bucket/./prefix//object", "bucket/prefix/object"),
+            ("bucket/a/../b", "bucket/b"),
+            ("../bucket/..", ".."),
+            ("/../../bucket", "/bucket"),
+            ("bucket\\prefix/object", "bucket\\prefix/object"),
+        ] {
+            assert_eq!(hash_path(input).key(), expected, "unexpected portable cache key for {input:?}");
+        }
+    }
+
     #[test]
     fn completeness_marker_is_additive_for_legacy_named_readers() {
         let current = DataUsageInfo {
@@ -1571,6 +1646,49 @@ mod tests {
         assert_eq!(map["BETWEEN_512_KB_AND_1_MB"], 1);
     }
 
+    #[test]
+    fn test_size_histogram_classifies_adjacent_boundaries_once() {
+        let cases = [
+            (1023, 0),
+            (1024, 1),
+            (64 * 1024 - 1, 1),
+            (64 * 1024, 2),
+            (256 * 1024 - 1, 2),
+            (256 * 1024, 3),
+            (512 * 1024 - 1, 3),
+            (512 * 1024, 4),
+            (1024 * 1024 - 1, 4),
+            (1024 * 1024, 6),
+            (10 * 1024 * 1024 - 1, 6),
+            (10 * 1024 * 1024, 7),
+            (64 * 1024 * 1024 - 1, 7),
+            (64 * 1024 * 1024, 8),
+            (128 * 1024 * 1024 - 1, 8),
+            (128 * 1024 * 1024, 9),
+            (512 * 1024 * 1024 - 1, 9),
+            (512 * 1024 * 1024, 10),
+        ];
+
+        for (size, expected_bucket) in cases {
+            let mut hist = SizeHistogram::default();
+            hist.add(size);
+
+            assert_eq!(hist.0.iter().sum::<u64>(), 1, "size {size} must have exactly one physical bucket");
+            assert_eq!(hist.0[expected_bucket], 1, "size {size} must select the expected bucket");
+        }
+    }
+
+    #[test]
+    fn test_size_histogram_1024_bytes_contributes_to_compat_rollup() {
+        let mut hist = SizeHistogram::default();
+        hist.add(1024);
+
+        let map = hist.to_map();
+        assert_eq!(map["LESS_THAN_1024_B"], 0);
+        assert_eq!(map["BETWEEN_1024_B_AND_64_KB"], 1);
+        assert_eq!(map["BETWEEN_1024B_AND_1_MB"], 1);
+    }
+
     #[test]
     fn test_size_histogram_compat_rollup_saturates_on_corrupt_counts() {
         let mut hist = SizeHistogram::default();
@@ -1582,6 +1700,126 @@ mod tests {
         assert_eq!(map["BETWEEN_1024B_AND_1_MB"], u64::MAX);
     }
 
+    #[test]
+    fn replication_stats_empty_checks_every_field() {
+        type SetField = fn(&mut ReplicationStats);
+
+        let cases: [(&str, SetField); 10] = [
+            ("pending_size", |stats| stats.pending_size = 1),
+            ("replicated_size", |stats| stats.replicated_size = 1),
+            ("failed_size", |stats| stats.failed_size = 1),
+            ("failed_count", |stats| stats.failed_count = 1),
+            ("pending_count", |stats| stats.pending_count = 1),
+            ("missed_threshold_size", |stats| stats.missed_threshold_size = 1),
+            ("after_threshold_size", |stats| stats.after_threshold_size = 1),
+            ("missed_threshold_count", |stats| stats.missed_threshold_count = 1),
+            ("after_threshold_count", |stats| stats.after_threshold_count = 1),
+            ("replicated_count", |stats| stats.replicated_count = 1),
+        ];
+
+        assert!(ReplicationStats::default().is_empty());
+        for (field, set_nonzero) in cases {
+            let mut stats = ReplicationStats::default();
+            set_nonzero(&mut stats);
+            assert!(!stats.is_empty(), "{field} must make replication stats non-empty");
+        }
+    }
+
+    #[test]
+    fn replication_all_stats_empty_checks_aggregate_fields_independently() {
+        let cases = [
+            (
+                "replica_size",
+                ReplicationAllStats {
+                    replica_size: 1,
+                    ..Default::default()
+                },
+            ),
+            (
+                "replica_count",
+                ReplicationAllStats {
+                    replica_count: 1,
+                    ..Default::default()
+                },
+            ),
+        ];
+
+        assert!(ReplicationAllStats::default().is_empty());
+        for (field, stats) in cases {
+            assert!(!stats.is_empty(), "{field} must make aggregate replication stats non-empty");
+        }
+
+        let empty_targets = ReplicationAllStats {
+            targets: HashMap::from([("arn:test:empty".to_string(), ReplicationStats::default())]),
+            ..Default::default()
+        };
+        assert!(empty_targets.is_empty(), "all-empty targets must keep aggregate stats empty");
+
+        let stats = ReplicationAllStats {
+            targets: HashMap::from([
+                ("arn:test:empty".to_string(), ReplicationStats::default()),
+                (
+                    "arn:test:non-empty".to_string(),
+                    ReplicationStats {
+                        pending_count: 1,
+                        ..Default::default()
+                    },
+                ),
+            ]),
+            ..Default::default()
+        };
+        assert!(!stats.is_empty(), "a non-empty target must make aggregate replication stats non-empty");
+    }
+
+    #[test]
+    fn size_recursive_prunes_empty_and_preserves_pending_replication_stats() {
+        let root = hash_path("bucket");
+        let child = hash_path("bucket/child");
+        let mut cache = DataUsageCache::default();
+        cache.replace_hashed(&root, &None, &DataUsageEntry::default());
+        cache.replace_hashed(
+            &child,
+            &Some(root.clone()),
+            &DataUsageEntry {
+                replication_stats: Some(ReplicationAllStats::default()),
+                ..Default::default()
+            },
+        );
+
+        assert!(
+            cache
+                .size_recursive("bucket")
+                .expect("bucket usage should flatten")
+                .replication_stats
+                .is_none()
+        );
+
+        cache.replace_hashed(
+            &child,
+            &Some(root.clone()),
+            &DataUsageEntry {
+                replication_stats: Some(ReplicationAllStats {
+                    targets: HashMap::from([(
+                        "arn:test:pending".to_string(),
+                        ReplicationStats {
+                            pending_count: 1,
+                            ..Default::default()
+                        },
+                    )]),
+                    ..Default::default()
+                }),
+                ..Default::default()
+            },
+        );
+
+        let flattened = cache.size_recursive("bucket").expect("bucket usage should flatten");
+        let replication = flattened
+            .replication_stats
+            .expect("pending-only replication stats must survive pruning");
+
+        assert_eq!(replication.targets["arn:test:pending"].pending_count, 1);
+    }
+
     #[test]
     fn test_data_usage_cache_merge_adds_missing_child() {
         let mut base = DataUsageCache::default();

crates/e2e_test/Cargo.toml

@@ -70,7 +70,8 @@ walkdir.workspace = true
 base64 = { workspace = true }
 rand = { workspace = true, features = ["serde"] }
 chrono = { workspace = true, features = ["serde"] }
-md5 = { workspace = true }
+hex = { workspace = true }
+md-5 = { workspace = true }
 opentelemetry-proto = { workspace = true }
 prost.workspace = true
 sha2 = { workspace = true }

crates/e2e_test/src/checksum_upload_test.rs

@@ -24,9 +24,10 @@ mod tests {
     use aws_sdk_s3::types::{ChecksumAlgorithm, ChecksumMode, CompletedMultipartUpload, CompletedPart};
     use aws_smithy_http_client::Builder as SmithyHttpClientBuilder;
     use base64::Engine;
+    use md5::{Digest as Md5Digest, Md5};
     use rustfs_rio::{Checksum, ChecksumType as RioChecksumType};
     use serial_test::serial;
-    use sha2::{Digest, Sha256};
+    use sha2::Sha256;
     use tracing::info;
 
     fn create_s3_client(env: &RustFSTestEnvironment) -> Client {
@@ -70,7 +71,9 @@ mod tests {
     }
 
     fn content_md5_base64(body: &[u8]) -> String {
-        let digest = md5::compute(body);
+        let mut hasher = Md5::new();
+        hasher.update(body);
+        let digest = hasher.finalize();
         base64::engine::general_purpose::STANDARD.encode(digest.as_slice())
     }
 

crates/e2e_test/src/fake_s3_target/mod.rs

@@ -25,6 +25,7 @@ use hyper::body::Incoming;
 use hyper::server::conn::http1;
 use hyper::service::service_fn;
 use hyper_util::rt::{TokioIo, TokioTimer};
+use md5::{Digest as Md5Digest, Md5};
 use s3s::access::{S3Access, S3AccessContext};
 use s3s::auth::SimpleAuth;
 use s3s::dto::{
@@ -827,13 +828,25 @@ fn ensure_body_growth(current: usize, added: usize) -> S3Result {
 
 async fn md5_digest(body: Bytes, permit: OwnedSemaphorePermit) -> S3Result<([u8; 16], OwnedSemaphorePermit)> {
     if body.len() < 1024 * 1024 {
-        return Ok((md5::compute(body).0, permit));
+        return Ok((md5_bytes(body), permit));
     }
-    tokio::task::spawn_blocking(move || (md5::compute(body).0, permit))
+    tokio::task::spawn_blocking(move || (md5_bytes(body), permit))
         .await
         .map_err(|error| s3s::s3_error!(InternalError, "MD5 worker failed: {error}"))
 }
 
+fn md5_bytes(input: impl AsRef<[u8]>) -> [u8; 16] {
+    let mut hasher = Md5::new();
+    hasher.update(input.as_ref());
+    hasher.finalize().into()
+}
+
+fn md5_hex(input: impl AsRef<[u8]>) -> String {
+    let mut hasher = Md5::new();
+    hasher.update(input.as_ref());
+    hex::encode(hasher.finalize())
+}
+
 fn ensure_store_budget(state: &StoreState, removed_bytes: usize, added_bytes: usize, adds_version: bool) -> S3Result {
     let total_bytes = state
         .total_bytes
@@ -1005,7 +1018,7 @@ impl S3 for FakeBackend {
             Some(value) => value,
             None => {
                 let (digest, _body_permit) = md5_digest(body.clone(), _body_permit).await?;
-                format!("{:x}", md5::Digest(digest))
+                hex::encode(digest)
             }
         };
         let version = ObjectVersion {
@@ -1208,7 +1221,7 @@ impl S3 for FakeBackend {
         }
         let body = collect_stream(input.body, input.content_length, fault.as_ref(), &self.control).await?;
         let (digest, _body_permit) = md5_digest(body.clone(), _body_permit).await?;
-        let e_tag = format!("{:x}", md5::Digest(digest));
+        let e_tag = hex::encode(digest);
         let mut state = lock(&self.store);
         let existing_bytes = state
             .uploads
@@ -1336,7 +1349,7 @@ impl S3 for FakeBackend {
             .collect();
         let (body, digests, _body_permits) = assemble_multipart(assembly_parts, total_len, _body_permits).await?;
         let part_count = requested.len();
-        let e_tag = source_etag(&headers)?.unwrap_or_else(|| format!("{:x}-{part_count}", md5::compute(digests)));
+        let e_tag = source_etag(&headers)?.unwrap_or_else(|| format!("{}-{part_count}", md5_hex(digests)));
         let version = ObjectVersion {
             version_id: upload.version_id.clone(),
             body,

crates/e2e_test/src/internode_rpc_signature_e2e_test.rs

@@ -13,8 +13,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-//! Cross-process replay / tamper acceptance for the internode NodeService v2 RPC
-//! signature (<https://github.com/rustfs/backlog/issues/1327>).
+//! Cross-process replay / tamper acceptance for internode NodeService RPC
+//! signatures (<https://github.com/rustfs/backlog/issues/1327>,
+//! <https://github.com/rustfs/backlog/issues/1542>).
 //!
 //! # Why this exists on top of the in-process tests
 //!
@@ -78,6 +79,8 @@
 //! | mixed version: legacy-only still served, not blocked | [`legacy_only_signature_is_accepted_in_default_posture`] |
 //! | strict flip closes the signature downgrade | [`signature_strict_rejects_legacy_only_downgrade`] |
 //! | strict flip closes the body-digest downgrade, incl. v1 | [`body_digest_strict_rejects_digestless_mutation`] |
+//! | replay scope binds every RPC and rejects restart replay | [`replay_scope_rejects_replay_path_transplant_and_stale_epoch_e2e`] |
+//! | strict replay scope allows only Ping bootstrap before v3 | [`replay_scope_strict_requires_v3_after_ping_bootstrap_e2e`] |
 //!
 //! Two acceptance items are deliberately left to the in-process tests. A stale
 //! timestamp cannot be forged from outside — it is inside the HMAC — so
@@ -88,18 +91,20 @@
 
 use crate::common::{RustFSTestEnvironment, init_logging};
 use crate::storage_api::internode_rpc_signature::{
-    TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_signature_headers, node_service_time_out_client_no_auth,
+    TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
+    node_service_time_out_client_no_auth, verify_tonic_boot_epoch_response,
 };
 use http::{HeaderMap, Method};
 use rustfs_config::{
-    ENV_INTERNODE_RPC_BODY_DIGEST_STRICT, ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY, ENV_INTERNODE_RPC_SIGNATURE_STRICT,
+    ENV_INTERNODE_RPC_BODY_DIGEST_STRICT, ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY, ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT,
+    ENV_INTERNODE_RPC_SIGNATURE_STRICT,
 };
 use rustfs_protos::canonical_make_volume_request_body;
-use rustfs_protos::proto_gen::node_service::{MakeVolumeRequest, MakeVolumeResponse};
+use rustfs_protos::proto_gen::node_service::{MakeVolumeRequest, MakeVolumeResponse, PingRequest, PingResponse};
 use serial_test::serial;
 use sha2::{Digest, Sha256};
 use std::error::Error;
-use tonic::{Code, Request, Status};
+use tonic::{Code, Request, Response, Status};
 use uuid::Uuid;
 
 type TestResult = Result<(), Box<dyn Error + Send + Sync>>;
@@ -115,12 +120,14 @@ const TEST_RPC_SECRET: &str = "rustfs-internode-signature-e2e-secret";
 /// clears authentication stops harmlessly at `find_disk`.
 const ABSENT_DISK: &str = "/nonexistent/rustfs-signature-e2e-disk";
 
-/// Wire names of the two v2 headers these tests edit. They are `pub(crate)` in
-/// ecstore, so they are repeated here rather than imported — [`overwrite_header`]
-/// asserts the header it replaces was actually present, which turns a rename
-/// into a loud failure instead of silently reducing an attack to a no-op.
+/// Wire names of the v2 and replay-scope headers these black-box tests edit. They are
+/// `pub(crate)` in ecstore, so they are repeated here rather than imported.
+/// [`overwrite_header`] asserts the header it replaces was actually present, which turns a
+/// rename into a loud failure instead of silently reducing an attack to a no-op.
 const CONTENT_SHA256_HEADER: &str = "x-rustfs-content-sha256";
 const NONCE_HEADER: &str = "x-rustfs-rpc-nonce";
+const TIMESTAMP_HEADER: &str = "x-rustfs-timestamp";
+const BOOT_EPOCH_CHALLENGE_HEADER: &str = "x-rustfs-rpc-boot-epoch-challenge";
 
 /// gRPC service name carried in the signed scope, i.e. `TONIC_RPC_PREFIX`
 /// without its leading `/`.
@@ -155,19 +162,28 @@ fn align_rpc_secret_with_server() {
 ///
 /// Uses the no-cleanup spawn so a `pkill` pattern cannot reap servers belonging
 /// to other tests running in the same binary.
-async fn start_server(extra_env: &[(&str, &str)]) -> Result<RustFSTestEnvironment, Box<dyn Error + Send + Sync>> {
-    let mut env = RustFSTestEnvironment::new().await?;
+fn server_env(extra_env: &[(&'static str, &'static str)]) -> Vec<(&'static str, &'static str)> {
     let mut child_env = vec![
         ("RUSTFS_RPC_SECRET", TEST_RPC_SECRET),
         (ENV_INTERNODE_RPC_SIGNATURE_STRICT, "false"),
         (ENV_INTERNODE_RPC_BODY_DIGEST_STRICT, "false"),
+        (ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT, "false"),
         (ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY, "1048576"),
     ];
     child_env.extend_from_slice(extra_env);
-    env.start_rustfs_server_without_cleanup_with_env(&child_env).await?;
+    child_env
+}
+
+async fn start_server_with_env(child_env: &[(&str, &str)]) -> Result<RustFSTestEnvironment, Box<dyn Error + Send + Sync>> {
+    let mut env = RustFSTestEnvironment::new().await?;
+    env.start_rustfs_server_without_cleanup_with_env(child_env).await?;
     Ok(env)
 }
 
+async fn start_server(extra_env: &[(&'static str, &'static str)]) -> Result<RustFSTestEnvironment, Box<dyn Error + Send + Sync>> {
+    start_server_with_env(&server_env(extra_env)).await
+}
+
 /// Stop the child and drop the cached gRPC channel for its address.
 ///
 /// `node_service_time_out_client_no_auth` memoises channels in a process-global
@@ -238,12 +254,83 @@ fn overwrite_header(headers: &mut HeaderMap, name: &'static str, value: &str) {
 /// and nothing else — no interceptor adds or rewrites auth metadata, so the
 /// bytes on the wire are the ones the test chose.
 async fn call_make_volume(url: &str, request: MakeVolumeRequest, headers: HeaderMap) -> Result<MakeVolumeResponse, Status> {
+    call_make_volume_response(url, request, headers)
+        .await
+        .map(Response::into_inner)
+}
+
+async fn call_make_volume_response(
+    url: &str,
+    request: MakeVolumeRequest,
+    headers: HeaderMap,
+) -> Result<Response<MakeVolumeResponse>, Status> {
     let mut client = node_service_time_out_client_no_auth(&url.to_string())
         .await
         .map_err(|err| Status::unavailable(format!("cannot reach the node service: {err}")))?;
     let mut rpc_request = Request::new(request);
     rpc_request.metadata_mut().as_mut().extend(headers);
-    client.make_volume(rpc_request).await.map(|response| response.into_inner())
+    client.make_volume(rpc_request).await
+}
+
+async fn call_ping_response(url: &str, headers: HeaderMap) -> Result<Response<PingResponse>, Status> {
+    let mut client = node_service_time_out_client_no_auth(&url.to_string())
+        .await
+        .map_err(|err| Status::unavailable(format!("cannot reach the node service: {err}")))?;
+    let mut rpc_request = Request::new(PingRequest {
+        version: 1,
+        body: bytes::Bytes::new(),
+    });
+    rpc_request.metadata_mut().as_mut().extend(headers);
+    client.ping(rpc_request).await
+}
+
+fn attach_boot_epoch_challenge(headers: &mut HeaderMap) -> Uuid {
+    let challenge = Uuid::new_v4();
+    headers.insert(
+        BOOT_EPOCH_CHALLENGE_HEADER,
+        challenge.to_string().parse().expect("UUID must be a valid header value"),
+    );
+    challenge
+}
+
+fn mint_replay_scope_headers(audience: &str, path: &str, content_sha256: &str, boot_epoch: Uuid) -> HeaderMap {
+    let mut headers = mint_v2_headers(audience, "MakeVolume", Some(content_sha256));
+    let timestamp = headers
+        .get(TIMESTAMP_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .expect("v2 headers must carry a timestamp")
+        .to_string();
+    headers.extend(
+        gen_tonic_replay_scope_headers(audience, path, &timestamp, content_sha256, boot_epoch)
+            .expect("replay-scope headers must mint with the aligned RPC secret"),
+    );
+    headers
+}
+
+async fn learn_boot_epoch_from_make_volume(url: &str, audience: &str) -> Uuid {
+    let request = make_volume_request("signature-e2e-epoch-bootstrap");
+    let mut headers = mint_v2_headers(audience, "MakeVolume", Some(&canonical_digest(&request)));
+    let challenge = attach_boot_epoch_challenge(&mut headers);
+    let response = call_make_volume_response(url, request, headers)
+        .await
+        .expect("v2 request with epoch challenge must clear default authentication");
+    let boot_epoch = verify_tonic_boot_epoch_response(audience, challenge, response.metadata().as_ref())
+        .expect("server must HMAC-authenticate the advertised boot epoch");
+    assert_authenticated(
+        Ok(response.into_inner()),
+        "a v2 epoch-challenge request in the default replay-scope posture",
+    );
+    boot_epoch
+}
+
+async fn learn_boot_epoch_from_ping(url: &str, audience: &str) -> Uuid {
+    let mut headers = mint_v2_headers(audience, "Ping", None);
+    let challenge = attach_boot_epoch_challenge(&mut headers);
+    let response = call_ping_response(url, headers)
+        .await
+        .expect("v2 Ping with an epoch challenge must bootstrap strict replay scope");
+    verify_tonic_boot_epoch_response(audience, challenge, response.metadata().as_ref())
+        .expect("strict replay-scope Ping must return a valid boot epoch proof")
 }
 
 /// Assert a call cleared authentication.
@@ -332,6 +419,122 @@ async fn internode_rpc_signature_default_posture_e2e() -> TestResult {
     Ok(())
 }
 
+/// A replay-scoped signature is usable exactly once against the exact gRPC path and the server
+/// process epoch that minted it. This crosses the child-process boundary twice: the HMAC-protected
+/// epoch is learned from a real response, then the same server is restarted in place to prove its
+/// replacement epoch rejects the captured request even though the nonce cache is necessarily new.
+#[tokio::test]
+#[serial]
+async fn replay_scope_rejects_replay_path_transplant_and_stale_epoch_e2e() -> TestResult {
+    init_logging();
+    align_rpc_secret_with_server();
+    let child_env = server_env(&[]);
+    let mut env = start_server_with_env(&child_env).await?;
+    let url = env.url.clone();
+    let audience = audience_of(&env);
+    let boot_epoch = learn_boot_epoch_from_make_volume(&url, &audience).await;
+
+    let request = make_volume_request("replay-scope-e2e-once");
+    let captured = mint_replay_scope_headers(
+        &audience,
+        &format!("{TONIC_RPC_PREFIX}/MakeVolume"),
+        &canonical_digest(&request),
+        boot_epoch,
+    );
+    assert_authenticated(
+        call_make_volume(&url, request.clone(), captured.clone()).await,
+        "the first replay-scoped mutation delivery",
+    );
+    assert_rejected(
+        call_make_volume(&url, request.clone(), captured).await,
+        Code::Unauthenticated,
+        None,
+        "the same replay-scoped mutation delivered twice",
+    );
+
+    let transplanted =
+        mint_replay_scope_headers(&audience, &format!("{TONIC_RPC_PREFIX}/Ping"), &canonical_digest(&request), boot_epoch);
+    assert_rejected(
+        call_make_volume(&url, request.clone(), transplanted).await,
+        Code::Unauthenticated,
+        None,
+        "a replay-scoped Ping signature transplanted onto MakeVolume",
+    );
+
+    let stale_epoch = mint_replay_scope_headers(
+        &audience,
+        &format!("{TONIC_RPC_PREFIX}/MakeVolume"),
+        &canonical_digest(&request),
+        boot_epoch,
+    );
+    env.restart_server_preserving_data(Vec::new(), &child_env).await?;
+    rustfs_protos::evict_failed_connection(&url).await;
+    assert_rejected(
+        call_make_volume(&url, request.clone(), stale_epoch).await,
+        Code::Unauthenticated,
+        None,
+        "a replay-scoped signature captured before the receiving process restart",
+    );
+
+    let restarted_epoch = learn_boot_epoch_from_make_volume(&url, &audience).await;
+    assert_ne!(boot_epoch, restarted_epoch, "a restarted child process must advertise a new boot epoch");
+    let fresh_epoch = mint_replay_scope_headers(
+        &audience,
+        &format!("{TONIC_RPC_PREFIX}/MakeVolume"),
+        &canonical_digest(&request),
+        restarted_epoch,
+    );
+    assert_authenticated(
+        call_make_volume(&url, request, fresh_epoch).await,
+        "a replay-scoped mutation signed with the replacement process epoch",
+    );
+
+    stop_server(env, &url).await;
+    Ok(())
+}
+
+/// Strict replay scope leaves one authenticated v2 bootstrap: `Ping` carrying a fresh challenge.
+/// A mutating v2 request cannot use that lane; once the epoch proof is returned, the first v3
+/// mutation succeeds. This protects a server restart without reopening a general downgrade path.
+#[tokio::test]
+#[serial]
+async fn replay_scope_strict_requires_v3_after_ping_bootstrap_e2e() -> TestResult {
+    init_logging();
+    align_rpc_secret_with_server();
+    let env = start_server(&[(ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT, "true")]).await?;
+    let url = env.url.clone();
+    let audience = audience_of(&env);
+
+    let v2_request = make_volume_request("replay-scope-e2e-strict-v2");
+    assert_rejected(
+        call_make_volume(
+            &url,
+            v2_request.clone(),
+            mint_v2_headers(&audience, "MakeVolume", Some(&canonical_digest(&v2_request))),
+        )
+        .await,
+        Code::Unauthenticated,
+        None,
+        "a v2 mutation after replay-scope strictness is enabled",
+    );
+
+    let boot_epoch = learn_boot_epoch_from_ping(&url, &audience).await;
+    let request = make_volume_request("replay-scope-e2e-strict-v3");
+    let replay_scoped = mint_replay_scope_headers(
+        &audience,
+        &format!("{TONIC_RPC_PREFIX}/MakeVolume"),
+        &canonical_digest(&request),
+        boot_epoch,
+    );
+    assert_authenticated(
+        call_make_volume(&url, request, replay_scoped).await,
+        "a replay-scoped mutation after Ping bootstrap under strict replay scope",
+    );
+
+    stop_server(env, &url).await;
+    Ok(())
+}
+
 /// Baseline: correctly signed mutations are accepted, both with and without a
 /// body digest.
 ///

crates/e2e_test/src/kms/common.rs

@@ -30,6 +30,7 @@ use aws_sdk_s3::primitives::ByteStream;
 use aws_sdk_s3::types::ServerSideEncryption;
 use base64::{Engine, engine::general_purpose::STANDARD as BASE64};
 use http::header::{CONTENT_TYPE, HOST};
+use md5::{Digest as Md5Digest, Md5};
 use rustfs_signer::constants::UNSIGNED_PAYLOAD;
 use rustfs_signer::sign_v4;
 use s3s::Body;
@@ -68,7 +69,9 @@ pub fn skip_if_kms_admin_tool_unavailable(test_name: &str) -> bool {
 }
 
 pub fn sse_customer_key_md5_base64(key: &str) -> String {
-    BASE64.encode(md5::compute(key).0)
+    let mut hasher = Md5::new();
+    hasher.update(key.as_bytes());
+    BASE64.encode(hasher.finalize())
 }
 
 pub async fn kms_admin_request(

crates/e2e_test/src/kms/kms_edge_cases_test.rs

@@ -25,12 +25,18 @@ use super::common::{LocalKMSTestEnvironment, sse_customer_key_md5_base64};
 use crate::common::{TEST_BUCKET, init_logging};
 use aws_sdk_s3::types::ServerSideEncryption;
 use base64::Engine;
-use md5::compute;
+use md5::{Digest as Md5Digest, Md5};
 use serial_test::serial;
 use std::sync::Arc;
 use tokio::sync::Semaphore;
 use tracing::{info, warn};
 
+fn md5_hex(input: impl AsRef<[u8]>) -> String {
+    let mut hasher = Md5::new();
+    hasher.update(input.as_ref());
+    hex::encode(hasher.finalize())
+}
+
 /// Test encryption of zero-byte files (empty files)
 #[tokio::test]
 #[serial]
@@ -294,7 +300,7 @@ async fn test_kms_invalid_key_scenarios() -> Result<(), Box<dyn std::error::Erro
     info!("🔍 Testing invalid SSE-C key length");
     let invalid_short_key = "short"; // Too short
     let invalid_key_b64 = base64::engine::general_purpose::STANDARD.encode(invalid_short_key);
-    let invalid_key_md5 = format!("{:x}", compute(invalid_short_key));
+    let invalid_key_md5 = md5_hex(invalid_short_key);
 
     let invalid_key_result = s3_client
         .put_object()

crates/e2e_test/src/multipart_auth_test.rs

@@ -26,6 +26,7 @@ use chrono::{Duration as ChronoDuration, Utc};
 use flate2::{Compression, write::GzEncoder};
 use http::HeaderValue;
 use http::header::{CONTENT_TYPE, HOST};
+use md5::{Digest as Md5Digest, Md5};
 use rustfs_signer::constants::UNSIGNED_PAYLOAD;
 use rustfs_signer::sign_v4;
 use s3s::Body;
@@ -50,7 +51,15 @@ fn encode_post_policy(conditions: Vec<serde_json::Value>) -> String {
 }
 
 fn sse_customer_key_md5_base64(key: &str) -> String {
-    base64::engine::general_purpose::STANDARD.encode(md5::compute(key).0)
+    let mut hasher = Md5::new();
+    hasher.update(key.as_bytes());
+    base64::engine::general_purpose::STANDARD.encode(hasher.finalize())
+}
+
+fn md5_hex(input: impl AsRef<[u8]>) -> String {
+    let mut hasher = Md5::new();
+    hasher.update(input.as_ref());
+    hex::encode(hasher.finalize())
 }
 
 /// Env var consumed by the local SSE-S3 DEK provider when KMS is not configured.
@@ -5664,7 +5673,7 @@ async fn test_signed_put_object_extract_returns_archive_etag() -> Result<(), Box
     client.create_bucket().bucket(bucket).send().await?;
 
     let archive = make_tar(&[("alpha.txt", b"alpha-body")], &[]).await;
-    let expected_etag = format!("\"{:x}\"", md5::compute(&archive));
+    let expected_etag = format!("\"{}\"", md5_hex(&archive));
 
     let response = client
         .put_object()

crates/e2e_test/src/notification_webhook_test.rs

@@ -21,18 +21,22 @@
 //! function, never as an S3 event sink.
 //!
 //! Coverage:
-//!   * PUT / multipart-complete / DELETE each deliver one event with the correct
+//!   * PUT / multipart-complete / DeleteObject / DeleteObjects each deliver one event with the correct
 //!     eventName, bucket, key, versionId and eTag.
 //!   * prefix/suffix filters drop non-matching keys (rule-engine gate).
 //!   * an event queued while the target endpoint is unreachable is redelivered
 //!     from the on-disk store once the endpoint recovers (store-and-forward).
+//!   * responseElements and the S3 response use the canonical request ID while
+//!     requestParameters preserve a conflicting client-supplied value.
 
 use crate::common::{RustFSTestEnvironment, init_logging};
 use aws_sdk_s3::Client;
+use aws_sdk_s3::operation::RequestId;
 use aws_sdk_s3::primitives::ByteStream;
 use aws_sdk_s3::types::{
-    BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, Event, FilterRule, FilterRuleName,
-    NotificationConfiguration, NotificationConfigurationFilter, QueueConfiguration, S3KeyFilter, VersioningConfiguration,
+    BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, Delete, Event, FilterRule, FilterRuleName,
+    NotificationConfiguration, NotificationConfigurationFilter, ObjectIdentifier, QueueConfiguration, S3KeyFilter,
+    VersioningConfiguration,
 };
 use http::header::{CONTENT_TYPE, HOST};
 use local_ip_address::local_ip;
@@ -40,10 +44,12 @@ use reqwest::StatusCode;
 use rustfs_signer::constants::UNSIGNED_PAYLOAD;
 use rustfs_signer::sign_v4;
 use rustfs_utils::egress::ENV_OUTBOUND_ALLOW_ORIGINS;
+use rustfs_utils::http::headers::{AMZ_REQUEST_ID, REQUEST_ID_HEADER};
 use s3s::Body;
 use serde_json::Value;
 use serial_test::serial;
 use std::error::Error;
+use std::io::Cursor;
 use std::path::Path;
 use std::sync::{
     Arc, Once,
@@ -63,6 +69,8 @@ type BoxError = Box<dyn Error + Send + Sync>;
 /// for target lookup (see `process_queue_configurations`), so the region is
 /// nominal, but it must be present for `ARN::parse` to succeed.
 const NOTIFY_REGION: &str = "us-east-1";
+const CLIENT_REQUEST_ID: &str = "client-supplied-request-id";
+const CLIENT_AMZ_REQUEST_ID: &str = "client-supplied-amz-request-id";
 
 /// Webhook targets are registered as `TargetID { id: <name>, name: "webhook" }`,
 /// so the ARN a notification rule references is
@@ -579,6 +587,36 @@ fn trimmed_etag(value: Option<&str>) -> Option<String> {
     value.map(|e| e.trim_matches('"').to_string())
 }
 
+fn assert_conflicting_request_id_correlation(record: &Value, server_request_id: &str) {
+    assert_eq!(
+        record["requestParameters"][REQUEST_ID_HEADER].as_str(),
+        Some(CLIENT_REQUEST_ID),
+        "notification request parameters should retain the actual client header: {record}"
+    );
+    assert_eq!(
+        record["requestParameters"][AMZ_REQUEST_ID].as_str(),
+        Some(CLIENT_AMZ_REQUEST_ID),
+        "notification request parameters should retain the actual client header: {record}"
+    );
+    assert_eq!(
+        record["responseElements"][AMZ_REQUEST_ID].as_str(),
+        Some(server_request_id),
+        "notification response elements should use the canonical request ID: {record}"
+    );
+}
+
+fn assert_generated_request_id_correlation(record: &Value, request_id: &str) {
+    assert!(
+        record["requestParameters"][AMZ_REQUEST_ID].is_null(),
+        "notification request parameters must not invent a client request header: {record}"
+    );
+    assert_eq!(
+        record["responseElements"][AMZ_REQUEST_ID].as_str(),
+        Some(request_id),
+        "notification response elements should match the S3 response request ID: {record}"
+    );
+}
+
 // ---------------------------------------------------------------------------
 // Tests
 // ---------------------------------------------------------------------------
@@ -671,8 +709,17 @@ async fn test_webhook_event_delivery_and_filtering() -> TestResult {
         .bucket(bucket)
         .key(put_key)
         .body(ByteStream::from_static(b"peri-1 put body"))
+        .customize()
+        .mutate_request(|request| {
+            request.headers_mut().insert(REQUEST_ID_HEADER, CLIENT_REQUEST_ID);
+            request.headers_mut().insert(AMZ_REQUEST_ID, CLIENT_AMZ_REQUEST_ID);
+        })
         .send()
         .await?;
+    let put_request_id = put.request_id().ok_or("PUT response missing request ID")?.to_owned();
+    assert!(uuid::Uuid::parse_str(&put_request_id).is_ok());
+    assert_ne!(put_request_id, CLIENT_REQUEST_ID);
+    assert_ne!(put_request_id, CLIENT_AMZ_REQUEST_ID);
     let put_version = put
         .version_id()
         .ok_or("PUT response missing versionId (versioning not enabled?)")?;
@@ -692,6 +739,7 @@ async fn test_webhook_event_delivery_and_filtering() -> TestResult {
         "record eventName: {record}"
     );
     assert_eq!(record["s3"]["bucket"]["name"].as_str(), Some(bucket), "bucket in event: {record}");
+    assert_conflicting_request_id_correlation(record, &put_request_id);
     assert_eq!(object["versionId"].as_str(), Some(put_version), "versionId in event: {object}");
     assert_eq!(
         trimmed_etag(object["eTag"].as_str()),
@@ -744,6 +792,35 @@ async fn test_webhook_event_delivery_and_filtering() -> TestResult {
         "multipart eTag in event: {mp_record}"
     );
 
+    // --- Snowball extract: direct notification path keeps response correlation
+    let snowball_key = "uploads/snowball.dat";
+    let snowball_body = b"snowball notification body";
+    let mut archive_builder = tokio_tar::Builder::new(Cursor::new(Vec::new()));
+    let mut archive_header = tokio_tar::Header::new_gnu();
+    archive_header.set_size(u64::try_from(snowball_body.len()).expect("snowball fixture length should fit in u64"));
+    archive_header.set_mode(0o644);
+    archive_header.set_cksum();
+    archive_builder
+        .append_data(&mut archive_header, snowball_key, Cursor::new(snowball_body))
+        .await?;
+    let archive = archive_builder.into_inner().await?.into_inner();
+    let snowball = client
+        .put_object()
+        .bucket(bucket)
+        .key("snowball-fixture.tar")
+        .body(ByteStream::from(archive))
+        .customize()
+        .mutate_request(|request| {
+            request.headers_mut().insert("x-amz-meta-snowball-auto-extract", "true");
+        })
+        .send()
+        .await?;
+    let snowball_request_id = snowball.request_id().ok_or("Snowball response missing request ID")?;
+
+    let snowball_event = wait_for_event(&mut rx, snowball_key, "s3:ObjectCreated:", Duration::from_secs(20)).await?;
+    let snowball_record = &snowball_event["Records"][0];
+    assert_generated_request_id_correlation(snowball_record, snowball_request_id);
+
     // --- Filter: non-matching prefix and suffix must never be delivered ------
     let wrong_prefix = "logs/report.dat"; // right suffix, wrong prefix
     let wrong_suffix = "uploads/report.txt"; // right prefix, wrong suffix
@@ -772,7 +849,32 @@ async fn test_webhook_event_delivery_and_filtering() -> TestResult {
         "wrong-suffix key {wrong_suffix} bypassed the filter; saw {seen:?}"
     );
 
-    // --- DELETE on a versioned bucket: ObjectRemoved:* with delete-marker version
+    // --- DeleteObjects: direct notification path keeps response correlation --
+    let delete_many_key = "uploads/delete-many.dat";
+    client
+        .put_object()
+        .bucket(bucket)
+        .key(delete_many_key)
+        .body(ByteStream::from_static(b"delete objects notification body"))
+        .send()
+        .await?;
+    wait_for_event(&mut rx, delete_many_key, "s3:ObjectCreated:", Duration::from_secs(20)).await?;
+
+    let delete_many = client
+        .delete_objects()
+        .bucket(bucket)
+        .delete(
+            Delete::builder()
+                .objects(ObjectIdentifier::builder().key(delete_many_key).build()?)
+                .build()?,
+        )
+        .send()
+        .await?;
+    let delete_many_request_id = delete_many.request_id().ok_or("DeleteObjects response missing request ID")?;
+    let delete_many_event = wait_for_event(&mut rx, delete_many_key, "s3:ObjectRemoved:", Duration::from_secs(20)).await?;
+    assert_generated_request_id_correlation(&delete_many_event["Records"][0], delete_many_request_id);
+
+    // --- DeleteObject on a versioned bucket: delete-marker version ----------
     let delete = client.delete_object().bucket(bucket).key(put_key).send().await?;
     let removed = wait_for_event(&mut rx, put_key, "s3:ObjectRemoved:", Duration::from_secs(20)).await?;
     let removed_record = &removed["Records"][0];

crates/e2e_test/src/reliant/grpc_lock_client.rs

@@ -24,7 +24,7 @@ use rustfs_protos::proto_gen::node_service::{BatchGenerallyLockRequest, Generall
 use tonic::Request;
 use tracing::{info, warn};
 
-use crate::storage_api::grpc_lock::{TonicInterceptor, node_service_time_out_client_no_auth};
+use crate::storage_api::grpc_lock::{AuthenticatedChannel, TonicInterceptor, node_service_time_out_client_no_auth};
 
 /// gRPC lock client without authentication for testing
 /// Similar to RemoteClient but uses no_auth client
@@ -42,7 +42,7 @@ impl GrpcLockClient {
         &self,
     ) -> Result<
         rustfs_protos::proto_gen::node_service::node_service_client::NodeServiceClient<
-            tonic::service::interceptor::InterceptedService<tonic::transport::Channel, TonicInterceptor>,
+            tonic::service::interceptor::InterceptedService<AuthenticatedChannel, TonicInterceptor>,
         >,
     > {
         node_service_time_out_client_no_auth(&self.addr)

crates/e2e_test/src/reliant/grpc_lock_server.rs

@@ -23,7 +23,8 @@ use rustfs_protos::{
     proto_gen::node_service::{
         BatchGenerallyLockRequest, BatchGenerallyLockResponse, BatchReadVersionRequest, BatchReadVersionResponse,
         GenerallyLockRequest, GenerallyLockResponse, GenerallyLockResult, PingRequest, PingResponse,
-        node_service_server::NodeService,
+        SnapshotLeaseMutationResponse, SnapshotLeaseReleaseRequest, SnapshotLeaseRenewRequest, SnapshotLeaseRequest,
+        SnapshotLeaseResponse, node_service_server::NodeService,
     },
 };
 use std::pin::Pin;
@@ -104,6 +105,27 @@ impl NodeService for MinimalLockNodeService {
         Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
     }
 
+    async fn acquire_snapshot_lease(
+        &self,
+        _request: Request<SnapshotLeaseRequest>,
+    ) -> Result<Response<SnapshotLeaseResponse>, Status> {
+        Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
+    }
+
+    async fn renew_snapshot_lease(
+        &self,
+        _request: Request<SnapshotLeaseRenewRequest>,
+    ) -> Result<Response<SnapshotLeaseResponse>, Status> {
+        Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
+    }
+
+    async fn release_snapshot_lease(
+        &self,
+        _request: Request<SnapshotLeaseReleaseRequest>,
+    ) -> Result<Response<SnapshotLeaseMutationResponse>, Status> {
+        Err(Status::unimplemented("MinimalLockNodeService only supports lock RPCs"))
+    }
+
     async fn lock(&self, request: Request<GenerallyLockRequest>) -> Result<Response<GenerallyLockResponse>, Status> {
         let request = request.into_inner();
         let args: LockRequest = match serde_json::from_str(&request.args) {

crates/e2e_test/src/reliant/tiering.rs

@@ -95,6 +95,7 @@ const MANUAL_ASYNC_PARALLEL_OBJECTS: usize = 64;
 const MANUAL_ACTIVE_CANCEL_OBJECTS: usize = 512;
 const MANUAL_RESTART_CANCEL_OBJECTS: usize = 512;
 const MANUAL_ACTIVE_CANCEL_RUNNING_TIMEOUT: StdDuration = StdDuration::from_secs(15);
+const MANUAL_TRANSITION_CANCEL_BARRIER_ENV: &str = "RUSTFS_E2E_MANUAL_TRANSITION_CANCEL_BARRIER";
 const MANUAL_ASYNC_CONFLICT_TERMINAL_TIMEOUT: StdDuration = StdDuration::from_secs(90);
 const MANUAL_RESTART_RECOVERY_TIMEOUT: StdDuration = StdDuration::from_secs(80);
 const OBJECT_KEY: &str = "tier/鲁A12345/report.bin";
@@ -1684,8 +1685,15 @@ async fn test_manual_transition_async_active_cancel_reports_terminal_cancelled()
     cold_client.create_bucket().bucket(TIER_BUCKET).send().await?;
 
     let mut hot = RustFSTestEnvironment::new().await?;
-    hot.start_rustfs_server_with_env(vec![], &[("RUSTFS_SCANNER_ENABLED", "false"), ("RUSTFS_SCANNER_CYCLE", "3600")])
-        .await?;
+    hot.start_rustfs_server_with_env(
+        vec![],
+        &[
+            ("RUSTFS_SCANNER_ENABLED", "false"),
+            ("RUSTFS_SCANNER_CYCLE", "3600"),
+            (MANUAL_TRANSITION_CANCEL_BARRIER_ENV, "1"),
+        ],
+    )
+    .await?;
     let hot_client = hot.create_s3_client();
     add_rustfs_tier(&hot, &cold).await?;
 

crates/e2e_test/src/ssec_copy_test.rs

@@ -22,6 +22,7 @@ use aws_sdk_s3::primitives::ByteStream;
 use aws_sdk_s3::types::{BucketVersioningStatus, CompletedMultipartUpload, CompletedPart, VersioningConfiguration};
 use aws_smithy_http_client::Builder as SmithyHttpClientBuilder;
 use base64::Engine;
+use md5::{Digest as Md5Digest, Md5};
 use std::collections::HashMap;
 use std::sync::atomic::{AtomicUsize, Ordering};
 use std::sync::{Arc, Mutex};
@@ -102,10 +103,12 @@ impl Intercept for ResponseHeaderCapture {
 
 fn customer_key(byte: u8) -> CustomerKey {
     let raw = [byte; 32];
+    let mut hasher = Md5::new();
+    hasher.update(raw);
     CustomerKey {
         raw: String::from_utf8_lossy(&raw).into_owned(),
         encoded: base64::engine::general_purpose::STANDARD.encode(raw),
-        md5: base64::engine::general_purpose::STANDARD.encode(md5::compute(raw).0),
+        md5: base64::engine::general_purpose::STANDARD.encode(hasher.finalize()),
     }
 }
 

crates/e2e_test/src/storage_api.rs

@@ -16,9 +16,12 @@
 pub(crate) use rustfs_ecstore::api::bucket::bucket_target_sys::BucketTargetSys;
 #[cfg(test)]
 pub(crate) use rustfs_ecstore::api::disk::{VolumeInfo, WalkDirOptions};
+pub(crate) use rustfs_ecstore::api::rpc::{AuthenticatedChannel, TonicInterceptor, node_service_time_out_client_no_auth};
 #[cfg(test)]
-pub(crate) use rustfs_ecstore::api::rpc::{TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_signature_headers};
-pub(crate) use rustfs_ecstore::api::rpc::{TonicInterceptor, node_service_time_out_client_no_auth};
+pub(crate) use rustfs_ecstore::api::rpc::{
+    TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
+    verify_tonic_boot_epoch_response,
+};
 #[cfg(test)]
 pub(crate) use rustfs_ecstore::api::rpc::{gen_tonic_signature_interceptor, node_service_time_out_client};
 
@@ -30,7 +33,7 @@ pub(crate) mod node_interact {
 }
 
 pub(crate) mod grpc_lock {
-    pub(crate) use super::{TonicInterceptor, node_service_time_out_client_no_auth};
+    pub(crate) use super::{AuthenticatedChannel, TonicInterceptor, node_service_time_out_client_no_auth};
 }
 
 /// Signing/transport surface used by the cross-process internode RPC signature
@@ -40,7 +43,8 @@ pub(crate) mod grpc_lock {
 #[cfg(test)]
 pub(crate) mod internode_rpc_signature {
     pub(crate) use super::{
-        TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_signature_headers, node_service_time_out_client_no_auth,
+        TONIC_RPC_PREFIX, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
+        node_service_time_out_client_no_auth, verify_tonic_boot_epoch_response,
     };
 }
 

crates/ecstore/Cargo.toml

@@ -153,11 +153,18 @@ metrics = { workspace = true }
 [target.'cfg(target_os = "linux")'.dependencies]
 rustfs-uring = "0.2.1"
 
+[target.'cfg(windows)'.dependencies]
+winapi-util.workspace = true
+windows-sys = { workspace = true, features = ["Win32_Foundation", "Win32_Storage_FileSystem"] }
+
 [dev-dependencies]
 tokio = { workspace = true, features = ["rt-multi-thread", "macros", "test-util", "fs"] }
 criterion = { workspace = true, features = ["html_reports"] }
 temp-env = { workspace = true, features = ["async_closure"] }
 tracing-subscriber = { workspace = true, features = ["json", "env-filter", "time"] }
+# Only for `pin_callsite_interest_for_test`, which registers a `NoSubscriber`
+# dispatcher to keep tracing's process-global callsite-interest cache honest.
+tracing-core = { workspace = true }
 serial_test = { workspace = true }
 opentelemetry_sdk = { workspace = true, features = ["rt-tokio"] }
 proptest = "1"

crates/ecstore/src/api/mod.rs

@@ -67,6 +67,14 @@ pub mod bucket {
             };
         }
 
+        pub mod transition_transaction {
+            pub use crate::bucket::lifecycle::transition_transaction::{
+                TransitionOperatorDeleteResult, TransitionOperatorError, TransitionOperatorProbe, TransitionOperatorStatus,
+                delete_transition_candidate_for_operator, finalize_missing_transition_transaction_for_operator,
+                inspect_transition_transaction_for_operator,
+            };
+        }
+
         pub mod evaluator {
             pub use crate::bucket::lifecycle::evaluator::Evaluator;
         }
@@ -122,13 +130,13 @@ pub mod bucket {
 
     pub mod metadata_sys {
         pub use crate::bucket::metadata_sys::{
-            BucketMetadataSys, acquire_bucket_targets_transaction_lock, delete, get, get_accelerate_config, get_bucket_policy,
+            BucketMetadataSys, acquire_bucket_metadata_transaction_lock, delete, get, get_accelerate_config, get_bucket_policy,
             get_bucket_policy_raw, get_bucket_targets_config, get_config_from_disk, get_cors_config, get_durability_config,
             get_global_bucket_metadata_sys, get_lifecycle_config, get_logging_config, get_notification_config,
             get_object_lock_config, get_public_access_block_config, get_quota_config, get_replication_config,
             get_request_payment_config, get_sse_config, get_tagging_config, get_versioning_config, get_website_config,
-            init_bucket_metadata_sys, list_bucket_targets, remove_bucket_metadata, set_bucket_metadata, update,
-            update_bucket_targets_under_transaction_lock, update_config_with,
+            init_bucket_metadata_sys, list_bucket_targets, reload_bucket_metadata, remove_bucket_metadata, set_bucket_metadata,
+            update, update_bucket_targets_under_transaction_lock, update_config_with, update_under_transaction_lock,
         };
     }
 
@@ -176,8 +184,8 @@ pub mod bucket {
             get_global_replication_stats, init_background_replication, read_durable_mrf_backlog, replication_state_to_filemeta,
             replication_status_to_filemeta, replication_statuses_map, replication_target_arns, resync_start_conflict_id,
             should_remove_replication_target, should_schedule_delete_replication, should_use_existing_delete_replication_info,
-            should_use_existing_delete_replication_source, validate_replication_config_target_arns,
-            version_purge_status_to_filemeta,
+            should_use_existing_delete_replication_source, unsupported_replication_config_field,
+            validate_replication_config_target_arns, version_purge_status_to_filemeta,
         };
     }
 
@@ -305,7 +313,8 @@ pub mod disk {
         CheckPartsResp, DeleteOptions, Disk, DiskAPI, DiskInfo, DiskInfoOptions, DiskLocation, DiskOption, DiskStore,
         FileInfoVersions, FileReader, FileWriter, HEALING_MARKER_PATH, NsScannerOpenRequest, OldCurrentSize,
         PartTransactionAction, RUSTFS_META_BUCKET, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp,
-        STORAGE_FORMAT_FILE, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, new_disk, validate_batch_read_version_item_count,
+        STORAGE_FORMAT_FILE, SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, new_disk,
+        validate_batch_read_version_item_count,
     };
     pub use bytes::Bytes;
     pub use endpoint::Endpoint;
@@ -376,6 +385,7 @@ pub mod metrics {
 pub mod notification {
     pub use crate::services::notification_sys::{
         NotificationPeerErr, NotificationSys, get_global_notification_sys, new_global_notification_sys,
+        start_remote_version_state_fleet_probe,
     };
 }
 
@@ -406,13 +416,15 @@ pub mod rio {
 
 pub mod rpc {
     pub use crate::cluster::rpc::{
-        LocalPeerS3Client, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS, PeerRestClient, PeerS3Client, S3PeerSys,
-        SERVICE_SIGNAL_REFRESH_CONFIG, SERVICE_SIGNAL_RELOAD_DYNAMIC, ScannerBucketListing, ScannerPeerActivity,
-        TONIC_RPC_PREFIX, TonicInterceptor, gen_signature_headers, gen_tonic_signature_headers, gen_tonic_signature_interceptor,
-        node_service_time_out_client, node_service_time_out_client_no_auth, normalize_tonic_rpc_audience,
-        set_tonic_canonical_body_digest, sign_ns_scanner_capability, sign_tonic_rpc_response_proof, verify_rpc_signature,
-        verify_tonic_canonical_body_digest, verify_tonic_mutation_body_digest, verify_tonic_rpc_response_proof,
-        verify_tonic_rpc_signature,
+        AuthenticatedChannel, LocalPeerS3Client, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS, PeerRestClient,
+        PeerS3Client, S3PeerSys, SERVICE_SIGNAL_REFRESH_CONFIG, SERVICE_SIGNAL_RELOAD_DYNAMIC, ScannerBucketListing,
+        ScannerPeerActivity, TONIC_RPC_PREFIX, TonicInterceptor, gen_signature_headers, gen_tonic_replay_scope_headers,
+        gen_tonic_signature_headers, gen_tonic_signature_interceptor, node_service_time_out_client,
+        node_service_time_out_client_no_auth, normalize_tonic_rpc_audience, set_tonic_canonical_body_digest,
+        sign_ns_scanner_capability, sign_tonic_rpc_response_proof, tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers,
+        verify_rpc_signature, verify_tonic_boot_epoch_response, verify_tonic_canonical_body_digest,
+        verify_tonic_mutation_body_digest, verify_tonic_rpc_response_proof, verify_tonic_rpc_signature,
+        verify_tonic_rpc_signature_with_bootstrap,
     };
 }
 

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

@@ -554,6 +554,7 @@ async fn delete_free_version_remote_object(
     oi: &ObjectInfo,
     tier_config_mgr: &Arc<RwLock<TierConfigMgr>>,
 ) -> Result<(), std::io::Error> {
+    let version_id_exact = validate_transition_remote_version(oi)?;
     let identity = tier_destination_id_from_metadata(&oi.user_defined)?
         .ok_or_else(|| std::io::Error::other("tier free-version has no durable backend identity"))?;
     delete_object_from_remote_tier_idempotent_with_manager_and_identity(
@@ -562,7 +563,7 @@ async fn delete_free_version_remote_object(
         &oi.transitioned_object.tier,
         identity,
         tier_config_mgr,
-        false,
+        version_id_exact,
     )
     .await?;
     Ok(())
@@ -4201,6 +4202,23 @@ pub async fn get_transitioned_object_reader(
     get_transitioned_object_reader_with_tier_manager(bucket, object, rs, h, oi, opts, &tier_config_mgr).await
 }
 
+fn validate_transition_remote_version(oi: &ObjectInfo) -> Result<bool, std::io::Error> {
+    let version = oi.transitioned_object.version_id.as_str();
+    match oi.transition_version_state {
+        rustfs_filemeta::TransitionVersionState::Unknown => Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier object version state is unknown",
+        )),
+        rustfs_filemeta::TransitionVersionState::KnownDisabled if version.is_empty() => Ok(false),
+        rustfs_filemeta::TransitionVersionState::SuspendedNull if version == "null" => Ok(true),
+        rustfs_filemeta::TransitionVersionState::Exact if !version.is_empty() && version != "null" => Ok(true),
+        _ => Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier object version state conflicts with its version ID",
+        )),
+    }
+}
+
 pub(crate) async fn get_transitioned_object_reader_with_tier_manager(
     bucket: &str,
     object: &str,
@@ -4210,6 +4228,7 @@ pub(crate) async fn get_transitioned_object_reader_with_tier_manager(
     opts: &ObjectOptions,
     tier_config_mgr: &Arc<RwLock<TierConfigMgr>>,
 ) -> Result<GetObjectReader, std::io::Error> {
+    validate_transition_remote_version(oi)?;
     let expected_identity = tier_destination_id_from_metadata(&oi.user_defined)?;
     let lease = match expected_identity {
         Some(identity) => {
@@ -5506,6 +5525,7 @@ mod tests {
                 tier: tier.clone(),
                 ..Default::default()
             },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Exact,
             ..Default::default()
         };
 
@@ -5569,6 +5589,7 @@ mod tests {
                 tier,
                 ..Default::default()
             },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Exact,
             ..Default::default()
         };
 
@@ -5591,6 +5612,70 @@ mod tests {
         assert_eq!(backend.get_count().await, 0);
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn transitioned_get_rejects_unknown_version_state_before_backend_io() {
+        let manager = TierConfigMgr::new();
+        let tier = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let backend = register_mock_tier(&manager, &tier).await;
+        let object_info = ObjectInfo {
+            bucket: "bucket".to_string(),
+            name: "object".to_string(),
+            size: 1,
+            transitioned_object: TransitionedObject {
+                name: "remote/object".to_string(),
+                version_id: String::new(),
+                status: crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE.to_string(),
+                tier,
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            ..Default::default()
+        };
+
+        let err = match get_transitioned_object_reader_with_tier_manager(
+            &object_info.bucket,
+            &object_info.name,
+            &None,
+            &HeaderMap::new(),
+            &object_info,
+            &ObjectOptions::default(),
+            &manager,
+        )
+        .await
+        {
+            Ok(_) => panic!("unknown remote version state must fail before backend IO"),
+            Err(err) => err,
+        };
+
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.get_count().await, 0);
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn free_version_delete_rejects_unknown_version_state_before_backend_io() {
+        let manager = TierConfigMgr::new();
+        let backend = register_mock_tier(&manager, "WARM").await;
+        let object_info = ObjectInfo {
+            transitioned_object: TransitionedObject {
+                name: "remote/object".to_string(),
+                version_id: "legacy-version".to_string(),
+                tier: "WARM".to_string(),
+                ..Default::default()
+            },
+            transition_version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+            ..Default::default()
+        };
+
+        let err = super::delete_free_version_remote_object(&object_info, &manager)
+            .await
+            .expect_err("unknown remote version state must fail before backend IO");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.remove_count().await, 0);
+    }
+
     #[cfg(feature = "test-util")]
     #[tokio::test]
     async fn free_version_remote_delete_requires_persisted_destination_identity() {
@@ -5640,6 +5725,7 @@ mod tests {
         oi.transitioned_object.tier = "WARM".to_string();
         oi.transitioned_object.name = "remote/object".to_string();
         oi.transitioned_object.version_id = "remote-version".to_string();
+        oi.transition_version_state = rustfs_filemeta::TransitionVersionState::Exact;
         let local_delete_calls = Arc::new(std::sync::atomic::AtomicUsize::new(0));
 
         let legacy_err = delete_free_version_remote_object_then(&oi, &manager, {
@@ -5650,7 +5736,8 @@ mod tests {
         })
         .await
         .expect_err("legacy free-version without identity must be retained");
-        assert!(legacy_err.to_string().contains("no durable backend identity"));
+        assert_eq!(legacy_err.kind(), std::io::ErrorKind::Other);
+        assert_eq!(old_backend.remove_count().await, 0);
         assert_eq!(local_delete_calls.load(Ordering::Relaxed), 0);
 
         let mut invalid_metadata = HashMap::new();
@@ -5781,6 +5868,7 @@ mod tests {
         oi.transitioned_object.tier = "WARM".to_string();
         oi.transitioned_object.name = "remote/object".to_string();
         oi.transitioned_object.version_id = "remote-version".to_string();
+        oi.transition_version_state = rustfs_filemeta::TransitionVersionState::Exact;
 
         let err = match get_transitioned_object_reader_with_tier_manager(
             "bucket",
@@ -5796,7 +5884,12 @@ mod tests {
             Ok(_) => panic!("identity-bound GET must reject a same-name tier rebind"),
             Err(err) => err,
         };
-        assert!(err.to_string().contains("identity no longer matches"));
+        assert_eq!(err.kind(), std::io::ErrorKind::Other);
+        let admin_err = err
+            .get_ref()
+            .and_then(|source| source.downcast_ref::<crate::client::admin_handler_utils::AdminError>())
+            .expect("identity mismatch should retain the typed tier error");
+        assert_eq!(admin_err.code, crate::services::tier::tier::ERR_TIER_INVALID_CONFIG.code);
         assert_eq!(new_backend.get_count().await, 0);
 
         oi.user_defined = Arc::new(HashMap::new());
@@ -5902,7 +5995,8 @@ mod tests {
             version_id: "remote-version".to_string(),
             tier_name: "WARM".to_string(),
             backend_identity: Some([1; 32]),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         };
 
         let err = state
@@ -6013,7 +6107,8 @@ mod tests {
             version_id: "remote-version".to_string(),
             tier_name: "WARM".to_string(),
             backend_identity: Some([1; 32]),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         };
 
         state
@@ -8879,7 +8974,7 @@ mod tests {
         .await
         .expect("first worker result should persist");
         assert_eq!(first.state, ManualTransitionJobState::Running);
-        assert_eq!(first.report.transition_completed, 1);
+        assert_eq!(first.report.transition_completed, 0);
         assert_eq!(first.report.transition_failed, 0);
 
         let duplicate = record_manual_transition_worker_result(
@@ -8892,11 +8987,11 @@ mod tests {
         .await
         .expect("duplicate worker result should be idempotent");
         assert_eq!(duplicate.state, ManualTransitionJobState::Running);
-        assert_eq!(duplicate.report.transition_completed, 1);
+        assert_eq!(duplicate.report.transition_completed, 0);
         assert_eq!(duplicate.report.transition_failed, 0);
 
         let second_key = manual_transition_worker_result_task_key(&bucket, "logs/b", None);
-        let final_record = record_manual_transition_worker_result(
+        let pending_record = record_manual_transition_worker_result(
             ecstore.clone(),
             job_id,
             &second_key,
@@ -8905,7 +9000,14 @@ mod tests {
         )
         .await
         .expect("second distinct worker result should persist");
+        assert_eq!(pending_record.state, ManualTransitionJobState::Running);
+        assert_eq!(pending_record.report.transition_completed, 0);
+        assert_eq!(pending_record.report.transition_failed, 0);
 
+        let final_record =
+            reconcile_manual_transition_worker_results(ecstore.clone(), job_id, ManualTransitionQueueSnapshot::default())
+                .await
+                .expect("worker result journal should reconcile");
         assert_eq!(final_record.state, ManualTransitionJobState::Partial);
         assert_eq!(final_record.report.transition_completed, 1);
         assert_eq!(final_record.report.transition_failed, 1);
@@ -8940,7 +9042,7 @@ mod tests {
             .expect("worker result job record should save");
 
         let task_key = manual_transition_worker_result_task_key(&bucket, "logs/fail", None);
-        let final_record = record_manual_transition_worker_result_with_reason(
+        let pending_record = record_manual_transition_worker_result_with_reason(
             ecstore.clone(),
             job_id,
             &task_key,
@@ -8950,7 +9052,12 @@ mod tests {
         )
         .await
         .expect("worker result with failure reason should persist");
+        assert!(pending_record.report.tier_failure_by_reason.is_empty());
+        assert_eq!(pending_record.report.transition_failed, 0);
 
+        let final_record = reconcile_manual_transition_worker_results(ecstore, job_id, ManualTransitionQueueSnapshot::default())
+            .await
+            .expect("worker failure reason should reconcile");
         assert_eq!(
             final_record
                 .report
@@ -10081,6 +10188,87 @@ mod tests {
         (backend, identity_hex)
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn journal_replay_rejects_unknown_version_state_before_backend_io() {
+        let (_disk_paths, ecstore) = setup_test_env().await;
+        let (backend, _) = register_recovery_mock_tier(&ecstore).await;
+        let identity = TierConfigMgr::acquire_operation_lease(&ecstore.tier_config_mgr(), "WARM")
+            .await
+            .expect("mock tier lease should be available")
+            .backend_identity();
+        let je = Jentry {
+            obj_name: "remote/object".to_string(),
+            version_id: "legacy-version".to_string(),
+            tier_name: "WARM".to_string(),
+            backend_identity: Some(identity),
+            version_id_exact: false,
+            version_state: rustfs_filemeta::TransitionVersionState::Unknown,
+        };
+
+        let err = crate::bucket::lifecycle::tier_delete_journal::process_tier_delete_journal_entry(ecstore, &je)
+            .await
+            .expect_err("unknown journal state must fail before backend IO");
+
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+        assert_eq!(backend.remove_count().await, 0);
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    async fn journal_replay_deletes_confirmed_exact_provider_token() {
+        let (_disk_paths, ecstore) = setup_test_env().await;
+        let (backend, _) = register_recovery_mock_tier(&ecstore).await;
+        let lease = TierConfigMgr::acquire_operation_lease(&ecstore.tier_config_mgr(), "WARM")
+            .await
+            .expect("mock tier lease should be available");
+        let identity = lease.backend_identity();
+        backend
+            .set_put_remote_version(Some("provider-version-token".to_string()))
+            .await;
+        lease
+            .put(
+                "remote/object",
+                crate::client::transition_api::ReaderImpl::Body(bytes::Bytes::from_static(b"candidate")),
+                9,
+            )
+            .await
+            .expect("confirmed remote candidate should be seeded");
+        backend.set_remove_failure(true);
+        backend.set_reject_non_empty_remote_versions(true);
+        let je = Jentry {
+            obj_name: "remote/object".to_string(),
+            version_id: "provider-version-token".to_string(),
+            tier_name: "WARM".to_string(),
+            backend_identity: Some(identity),
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
+        };
+
+        crate::set_disk::cleanup_rejected_transition_upload_durably(
+            &lease,
+            &je.obj_name,
+            &je.version_id,
+            true,
+            Some(ecstore.clone()),
+        )
+        .await
+        .expect("failed immediate cleanup should remain durable in the journal");
+        assert!(backend.contains(&je.obj_name).await);
+
+        backend.set_remove_failure(false);
+        crate::bucket::lifecycle::tier_delete_journal::process_tier_delete_journal_entry(ecstore, &je)
+            .await
+            .expect("identity-bound exact journal must retry confirmed candidate cleanup");
+
+        assert!(!backend.contains(&je.obj_name).await);
+        assert_eq!(backend.exact_remove_count(), 2);
+        assert_eq!(
+            backend.remove_versions().await,
+            vec![("remote/object".to_string(), "provider-version-token".to_string())]
+        );
+    }
+
     async fn seed_recoverable_free_version(
         disk_paths: &[PathBuf],
         bucket: &str,
@@ -10100,6 +10288,7 @@ mod tests {
                 identity,
             );
         }
+        let transition_version_id = Uuid::new_v4();
         let mut metadata = FileMeta::new();
         metadata
             .add_version(FileInfo {
@@ -10108,7 +10297,9 @@ mod tests {
                 version_id: Some(object_version_id),
                 transition_status: crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE.to_string(),
                 transitioned_objname: format!("remote/{bucket}/{object}"),
-                transition_version_id: Some(Uuid::new_v4()),
+                transition_version_id: Some(transition_version_id),
+                transition_version: Some(transition_version_id.to_string()),
+                transition_version_state: rustfs_filemeta::TransitionVersionState::Exact,
                 transition_tier: "WARM".to_string(),
                 mod_time: Some(OffsetDateTime::now_utc()),
                 metadata: transitioned_metadata,
@@ -11127,23 +11318,25 @@ mod tests {
 
         // Distinct payloads with distinct sizes: a mixed-generation reassembly
         // would produce bytes matching none of them (or fail the read outright).
-        let candidates: Vec<Vec<u8>> = (0..3)
+        let candidates: Vec<Vec<u8>> = (0..2)
             .map(|g| {
                 let len = 4096 + g * 512;
                 vec![b'a' + g as u8; len]
             })
             .collect();
 
-        let commit_barrier = MultipartCommitBarrier::install(&bucket, object, MultipartCommitPause::PutPartBeforeLockLost);
-        let start = Arc::new(tokio::sync::Barrier::new(candidates.len() + 1));
+        let commit_barrier = MultipartCommitBarrier::install_for_arrivals(
+            &bucket,
+            object,
+            MultipartCommitPause::PutPartBeforeLockAcquire,
+            candidates.len(),
+        );
         let mut tasks = tokio::task::JoinSet::new();
         for payload in candidates.iter().cloned() {
             let store = ecstore.clone();
             let bucket = bucket.clone();
             let upload_id = upload.upload_id.clone();
-            let start = Arc::clone(&start);
             tasks.spawn(async move {
-                start.wait().await;
                 let mut data = PutObjReader::from_vec(payload.clone());
                 store
                     .put_object_part(&bucket, object, &upload_id, 1, &mut data, &ObjectOptions::default())
@@ -11151,11 +11344,10 @@ mod tests {
                     .map(|info| (info, payload))
             });
         }
-        start.wait().await;
 
-        // The first writer holds the uploadId commit lock while the other
-        // resends reach the same critical section. Releasing it proves the
-        // handoff without depending on saturated CI disk latency.
+        // Both writers finish streaming before racing for the uploadId commit
+        // lock. Two generations are sufficient to exercise the mixed-shard
+        // hazard, while each waiter sits behind at most one cross-disk rename.
         commit_barrier.wait_until_paused().await;
         commit_barrier.release();
 

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

@@ -1646,40 +1646,14 @@ pub async fn record_manual_transition_worker_result_with_reason(
     job_id: Uuid,
     task_key: &str,
     result: ManualTransitionWorkerResult,
-    queue_snapshot: ManualTransitionQueueSnapshot,
+    _queue_snapshot: ManualTransitionQueueSnapshot,
     failure_reason: Option<ManualTransitionWorkerFailureReason>,
 ) -> EcstoreResult<ManualTransitionJobRecord> {
     let result_record = ManualTransitionWorkerResultRecord::new_with_reason(job_id, task_key, result, failure_reason);
     if !save_manual_transition_worker_result_if_absent(api.clone(), &result_record).await? {
         return load_manual_transition_job_record(api, job_id).await;
     }
-
-    for _ in 0..4 {
-        let (mut record, etag) = load_manual_transition_job_record_with_etag(api.clone(), job_id).await?;
-        if record.is_terminal() {
-            return Ok(record);
-        }
-        record.record_worker_result_with_reason(result, queue_snapshot, failure_reason);
-        match save_manual_transition_job_record_if_current(api.clone(), &record, &etag).await {
-            Ok(()) => {
-                if record.is_terminal() {
-                    delete_manual_transition_scope_admission_if_current(
-                        api.clone(),
-                        &record.scope_key,
-                        record.job_id,
-                        record.lease_id,
-                    )
-                    .await?;
-                } else {
-                    renew_manual_transition_scope_admission_from_job(api, &record).await?;
-                }
-                return Ok(record);
-            }
-            Err(Error::PreconditionFailed) => continue,
-            Err(err) => return Err(err),
-        }
-    }
-    Err(Error::PreconditionFailed)
+    load_manual_transition_job_record(api, job_id).await
 }
 
 pub async fn renew_manual_transition_job_lease(

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

@@ -20,7 +20,10 @@ use tokio_util::sync::CancellationToken;
 use tracing::{debug, warn};
 
 use crate::bucket::lifecycle::config_boundary;
-use crate::bucket::lifecycle::tier_sweeper::{Jentry, delete_object_from_remote_tier_idempotent_with_manager_and_identity};
+use crate::bucket::lifecycle::tier_sweeper::{
+    Jentry, delete_confirmed_transition_candidate_exact_with_manager_and_identity,
+    delete_object_from_remote_tier_idempotent_with_manager_and_identity,
+};
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::{Error, Result};
 use crate::object_api::{GetObjectReader, ObjectInfo, ObjectOptions, PutObjReader};
@@ -42,6 +45,7 @@ const TIER_DELETE_JOURNAL_RECOVERY_INTERVAL: Duration = Duration::from_secs(60);
 const TIER_DELETE_JOURNAL_RECOVERY_TIMEOUT: Duration = Duration::from_secs(300);
 const TIER_DELETE_JOURNAL_VERSION: u8 = 2;
 const TIER_DELETE_JOURNAL_EXACT_VERSION: u8 = 3;
+const TIER_DELETE_JOURNAL_STATE_VERSION: u8 = 4;
 pub(crate) const TIER_DELETE_JOURNAL_PREFIX: &str = "ilm/tier-delete-journal/";
 
 #[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
@@ -55,24 +59,35 @@ struct PersistedTierDeleteJournalEntry {
     backend_identity: Option<[u8; 32]>,
     #[serde(default, skip_serializing_if = "Option::is_none")]
     version_id_exact: Option<bool>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    version_state: Option<rustfs_filemeta::TransitionVersionState>,
 }
 
 impl PersistedTierDeleteJournalEntry {
-    fn from_jentry(je: &Jentry) -> Self {
-        Self {
-            version: if je.version_id_exact {
-                TIER_DELETE_JOURNAL_EXACT_VERSION
-            } else if je.backend_identity.is_some() {
+    fn from_jentry(je: &Jentry) -> Result<Self> {
+        validate_version_state(je.version_state, &je.version_id, je.version_id_exact)?;
+        let legacy_unknown = je.version_state == rustfs_filemeta::TransitionVersionState::Unknown;
+        let version = if legacy_unknown {
+            if je.backend_identity.is_some() {
                 TIER_DELETE_JOURNAL_VERSION
             } else {
                 1
-            },
+            }
+        } else {
+            if je.backend_identity.is_none() {
+                return Err(Error::other("new tier delete journal entry is missing its backend identity"));
+            }
+            TIER_DELETE_JOURNAL_STATE_VERSION
+        };
+        Ok(Self {
+            version,
             obj_name: je.obj_name.clone(),
             version_id: je.version_id.clone(),
             tier_name: je.tier_name.clone(),
             backend_identity: je.backend_identity,
             version_id_exact: je.version_id_exact.then_some(true),
-        }
+            version_state: (!legacy_unknown).then_some(je.version_state),
+        })
     }
 
     fn into_jentry(self) -> Result<Jentry> {
@@ -84,19 +99,23 @@ impl PersistedTierDeleteJournalEntry {
         if self.obj_name.is_empty() || self.tier_name.is_empty() {
             return Err(Error::other("tier delete journal entry is incomplete"));
         }
-        if self.version != TIER_DELETE_JOURNAL_EXACT_VERSION && self.version_id_exact.unwrap_or(false) {
+        if self.version != TIER_DELETE_JOURNAL_EXACT_VERSION
+            && self.version != TIER_DELETE_JOURNAL_STATE_VERSION
+            && self.version_id_exact.unwrap_or(false)
+        {
             return Err(Error::other(
                 "legacy tier delete journal entry has an unsupported exact version constraint",
             ));
         }
-        let (backend_identity, version_id_exact) = match self.version {
-            1 => (None, false),
+        let (backend_identity, version_id_exact, version_state) = match self.version {
+            1 => (None, false, rustfs_filemeta::TransitionVersionState::Unknown),
             TIER_DELETE_JOURNAL_VERSION => (
                 Some(
                     self.backend_identity
                         .ok_or_else(|| Error::other("tier delete journal v2 entry is missing its backend identity"))?,
                 ),
                 false,
+                rustfs_filemeta::TransitionVersionState::Unknown,
             ),
             TIER_DELETE_JOURNAL_EXACT_VERSION => {
                 if self.version_id.is_empty() || self.version_id_exact != Some(true) {
@@ -108,6 +127,22 @@ impl PersistedTierDeleteJournalEntry {
                             .ok_or_else(|| Error::other("tier delete journal v3 entry is missing its backend identity"))?,
                     ),
                     true,
+                    rustfs_filemeta::TransitionVersionState::Exact,
+                )
+            }
+            TIER_DELETE_JOURNAL_STATE_VERSION => {
+                let state = self
+                    .version_state
+                    .ok_or_else(|| Error::other("tier delete journal v4 entry is missing its version state"))?;
+                let exact = self.version_id_exact.unwrap_or(false);
+                validate_version_state(state, &self.version_id, exact)?;
+                (
+                    Some(
+                        self.backend_identity
+                            .ok_or_else(|| Error::other("tier delete journal v4 entry is missing its backend identity"))?,
+                    ),
+                    exact,
+                    state,
                 )
             }
             version => return Err(Error::other(format!("unsupported tier delete journal version {version}"))),
@@ -118,10 +153,30 @@ impl PersistedTierDeleteJournalEntry {
             tier_name: self.tier_name,
             backend_identity,
             version_id_exact,
+            version_state,
         })
     }
 }
 
+fn validate_version_state(
+    state: rustfs_filemeta::TransitionVersionState,
+    version_id: &str,
+    version_id_exact: bool,
+) -> Result<()> {
+    use rustfs_filemeta::TransitionVersionState::{Exact, KnownDisabled, SuspendedNull, Unknown};
+
+    let valid = match state {
+        Unknown => !version_id_exact,
+        KnownDisabled => version_id.is_empty() && !version_id_exact,
+        SuspendedNull => version_id == "null" && version_id_exact,
+        Exact => !version_id.is_empty() && version_id != "null" && version_id_exact,
+    };
+    if !valid {
+        return Err(Error::other("tier delete journal version state conflicts with its version id"));
+    }
+    Ok(())
+}
+
 #[derive(Debug, Clone, PartialEq, Eq)]
 pub struct TierDeleteJournalRecoveryStats {
     pub scanned: usize,
@@ -159,7 +214,7 @@ pub(crate) fn decode_tier_delete_journal_entry(data: &[u8]) -> Result<Jentry> {
 }
 
 pub(crate) fn encode_tier_delete_journal_entry(je: &Jentry) -> Result<Vec<u8>> {
-    serde_json::to_vec(&PersistedTierDeleteJournalEntry::from_jentry(je))
+    serde_json::to_vec(&PersistedTierDeleteJournalEntry::from_jentry(je)?)
         .map_err(|err| Error::other(format!("encode tier delete journal failed: {err}")))
 }
 
@@ -209,18 +264,35 @@ where
 }
 
 pub async fn process_tier_delete_journal_entry(api: Arc<ECStore>, je: &Jentry) -> std::io::Result<()> {
+    if je.version_state == rustfs_filemeta::TransitionVersionState::Unknown {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "tier delete journal remote version state is unknown",
+        ));
+    }
     let backend_identity = je
         .backend_identity
         .ok_or_else(|| std::io::Error::other("legacy tier delete journal has no durable backend identity"))?;
-    delete_object_from_remote_tier_idempotent_with_manager_and_identity(
-        &je.obj_name,
-        &je.version_id,
-        &je.tier_name,
-        backend_identity,
-        &api.tier_config_mgr(),
-        je.version_id_exact,
-    )
-    .await?;
+    if je.version_id_exact {
+        delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+            &je.obj_name,
+            &je.version_id,
+            &je.tier_name,
+            backend_identity,
+            &api.tier_config_mgr(),
+        )
+        .await?;
+    } else {
+        delete_object_from_remote_tier_idempotent_with_manager_and_identity(
+            &je.obj_name,
+            &je.version_id,
+            &je.tier_name,
+            backend_identity,
+            &api.tier_config_mgr(),
+            false,
+        )
+        .await?;
+    }
     remove_tier_delete_journal_entry(api, je).await
 }
 
@@ -406,8 +478,9 @@ where
 #[cfg(test)]
 mod tests {
     use super::{
-        TIER_DELETE_JOURNAL_EXACT_VERSION, await_tier_delete_journal_recovery, decode_tier_delete_journal_entry,
-        encode_tier_delete_journal_entry, record_tier_delete_journal_backend_identity, tier_delete_journal_object_name,
+        TIER_DELETE_JOURNAL_EXACT_VERSION, TIER_DELETE_JOURNAL_STATE_VERSION, await_tier_delete_journal_recovery,
+        decode_tier_delete_journal_entry, encode_tier_delete_journal_entry, record_tier_delete_journal_backend_identity,
+        tier_delete_journal_object_name,
     };
     use crate::bucket::lifecycle::tier_sweeper::Jentry;
     use crate::error::Result;
@@ -420,7 +493,8 @@ mod tests {
             version_id: "remote-version".to_string(),
             tier_name: "WARM".to_string(),
             backend_identity: Some([7; 32]),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         }
     }
 
@@ -436,6 +510,7 @@ mod tests {
         assert_eq!(decoded.tier_name, je.tier_name);
         assert_eq!(decoded.backend_identity, je.backend_identity);
         assert_eq!(decoded.version_id_exact, je.version_id_exact);
+        assert_eq!(decoded.version_state, je.version_state);
     }
 
     #[test]
@@ -450,7 +525,7 @@ mod tests {
         let persisted: serde_json::Value = serde_json::from_slice(&encoded).expect("exact journal JSON should decode");
         let decoded = decode_tier_delete_journal_entry(&encoded).expect("exact journal entry should decode");
 
-        assert_eq!(persisted["version"], TIER_DELETE_JOURNAL_EXACT_VERSION);
+        assert_eq!(persisted["version"], TIER_DELETE_JOURNAL_STATE_VERSION);
         assert_eq!(persisted["version_id_exact"], true);
         assert!(decoded.version_id_exact);
         assert_ne!(tier_delete_journal_object_name(&exact), tier_delete_journal_object_name(&normalized));
@@ -513,6 +588,46 @@ mod tests {
         }
     }
 
+    #[test]
+    fn tier_delete_journal_rejects_conflicting_v4_version_states() {
+        let identity = vec![7_u8; 32];
+        let invalid = [
+            ("known-disabled", "unexpected", false),
+            ("suspended-null", "", true),
+            ("suspended-null", "null", false),
+            ("exact", "", true),
+            ("exact", "null", true),
+            ("exact", "version", false),
+            ("unknown", "version", true),
+        ];
+
+        for (state, version_id, exact) in invalid {
+            let persisted = serde_json::json!({
+                "version": TIER_DELETE_JOURNAL_STATE_VERSION,
+                "obj_name": "remote/object",
+                "version_id": version_id,
+                "tier_name": "WARM",
+                "backend_identity": identity,
+                "version_id_exact": exact.then_some(true),
+                "version_state": state,
+            });
+            let encoded = serde_json::to_vec(&persisted).expect("invalid journal fixture should encode");
+            decode_tier_delete_journal_entry(&encoded).expect_err("conflicting v4 version state must fail closed");
+        }
+    }
+
+    #[test]
+    fn legacy_journals_decode_with_unknown_version_state() {
+        let v1 = br#"{"version":1,"obj_name":"remote/object","version_id":"opaque","tier_name":"WARM"}"#;
+        let v2 = br#"{"version":2,"obj_name":"remote/object","version_id":"opaque","tier_name":"WARM","backend_identity":[7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7,7]}"#;
+
+        for payload in [v1.as_slice(), v2.as_slice()] {
+            let decoded = decode_tier_delete_journal_entry(payload).expect("legacy journal should decode");
+            assert_eq!(decoded.version_state, rustfs_filemeta::TransitionVersionState::Unknown);
+            assert!(!decoded.version_id_exact);
+        }
+    }
+
     #[test]
     fn tier_delete_journal_path_is_stable_and_sanitized() {
         let je = journal_entry();
@@ -530,6 +645,8 @@ mod tests {
     fn tier_delete_journal_paths_separate_legacy_and_backend_identities() {
         let mut legacy = journal_entry();
         legacy.backend_identity = None;
+        legacy.version_id_exact = false;
+        legacy.version_state = rustfs_filemeta::TransitionVersionState::Unknown;
         let mut backend_a = journal_entry();
         backend_a.backend_identity = Some([1; 32]);
         let mut backend_b = journal_entry();
@@ -575,6 +692,8 @@ mod tests {
     fn tier_delete_journal_without_transition_identity_stays_legacy() {
         let mut je = journal_entry();
         je.backend_identity = None;
+        je.version_id_exact = false;
+        je.version_state = rustfs_filemeta::TransitionVersionState::Unknown;
 
         let encoded = encode_tier_delete_journal_entry(&je).expect("legacy journal should remain encodable");
         let persisted: serde_json::Value = serde_json::from_slice(&encoded).expect("journal JSON should decode");

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

@@ -185,6 +185,7 @@ struct ObjSweeper {
     transition_status: String,
     transition_tier: String,
     transition_version_id: String,
+    transition_version_state: rustfs_filemeta::TransitionVersionState,
     remote_object: String,
 }
 
@@ -231,7 +232,9 @@ impl ObjSweeper {
     }
 
     pub fn should_remove_remote_object(&self) -> Option<Jentry> {
-        if self.transition_status != lifecycle::TRANSITION_COMPLETE {
+        if self.transition_status != lifecycle::TRANSITION_COMPLETE
+            || self.transition_version_state == rustfs_filemeta::TransitionVersionState::Unknown
+        {
             return None;
         }
 
@@ -249,7 +252,11 @@ impl ObjSweeper {
                 version_id: self.transition_version_id.clone(),
                 tier_name: self.transition_tier.clone(),
                 backend_identity: None,
-                version_id_exact: false,
+                version_id_exact: matches!(
+                    self.transition_version_state,
+                    rustfs_filemeta::TransitionVersionState::SuspendedNull | rustfs_filemeta::TransitionVersionState::Exact
+                ),
+                version_state: self.transition_version_state,
             });
         }
         None
@@ -286,6 +293,7 @@ pub struct Jentry {
     pub(crate) tier_name: String,
     pub(crate) backend_identity: Option<TierDestinationId>,
     pub(crate) version_id_exact: bool,
+    pub(crate) version_state: rustfs_filemeta::TransitionVersionState,
 }
 
 impl ExpiryOp for Jentry {
@@ -330,7 +338,7 @@ async fn delete_object_from_remote_tier_raw_with_manager(
     let lease = TierConfigMgr::acquire_operation_lease(&tier_config_mgr, tier_name)
         .await
         .map_err(std::io::Error::other)?;
-    delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, &lease, false).await
+    delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, &lease, false, true).await
 }
 
 async fn delete_object_from_remote_tier_raw_with_lease(
@@ -338,8 +346,11 @@ async fn delete_object_from_remote_tier_raw_with_lease(
     rv_id: &str,
     lease: &TierOperationLease,
     version_id_exact: bool,
+    validate_remote_version_id: bool,
 ) -> Result<(), std::io::Error> {
-    lease.validate_remote_version_id(rv_id)?;
+    if validate_remote_version_id {
+        lease.validate_remote_version_id(rv_id)?;
+    }
 
     if remote_delete_breaker_is_open(Instant::now()).await {
         metrics::counter!(METRIC_DELETE_REMOTE_BREAKER_TOTAL).increment(1);
@@ -435,7 +446,53 @@ pub(crate) async fn delete_object_from_remote_tier_with_lease_idempotent(
     lease: &TierOperationLease,
     version_id_exact: bool,
 ) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
-    match delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, lease, version_id_exact).await {
+    delete_object_from_remote_tier_with_lease_idempotent_inner(obj_name, rv_id, lease, version_id_exact, true).await
+}
+
+pub(crate) async fn delete_confirmed_transition_candidate_exact_with_lease_idempotent(
+    obj_name: &str,
+    rv_id: &str,
+    lease: &TierOperationLease,
+) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
+    if rv_id.is_empty() {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidInput,
+            "confirmed versioned transition candidate requires a non-empty remote version",
+        ));
+    }
+    #[cfg(test)]
+    if obj_name == "remote/empty-guard-probe" {
+        CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES.fetch_add(1, std::sync::atomic::Ordering::Relaxed);
+    }
+    delete_object_from_remote_tier_with_lease_idempotent_inner(obj_name, rv_id, lease, true, false).await
+}
+
+#[cfg(test)]
+static CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES: std::sync::atomic::AtomicUsize = std::sync::atomic::AtomicUsize::new(0);
+
+pub(crate) async fn delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+    obj_name: &str,
+    rv_id: &str,
+    tier_name: &str,
+    backend_identity: TierDestinationId,
+    tier_config_mgr: &Arc<tokio::sync::RwLock<TierConfigMgr>>,
+) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
+    let lease = TierConfigMgr::acquire_operation_lease_for_backend_identity(tier_config_mgr, tier_name, backend_identity)
+        .await
+        .map_err(std::io::Error::other)?;
+    delete_confirmed_transition_candidate_exact_with_lease_idempotent(obj_name, rv_id, &lease).await
+}
+
+async fn delete_object_from_remote_tier_with_lease_idempotent_inner(
+    obj_name: &str,
+    rv_id: &str,
+    lease: &TierOperationLease,
+    version_id_exact: bool,
+    validate_remote_version_id: bool,
+) -> Result<RemoteTierDeleteOutcome, std::io::Error> {
+    match delete_object_from_remote_tier_raw_with_lease(obj_name, rv_id, lease, version_id_exact, validate_remote_version_id)
+        .await
+    {
         Ok(()) => Ok(RemoteTierDeleteOutcome::Deleted),
         Err(err) if is_remote_tier_not_found_error(&err) => Ok(RemoteTierDeleteOutcome::AlreadyRemoved),
         Err(err) => {
@@ -460,6 +517,7 @@ pub fn transitioned_delete_journal_entry(
     versioned: bool,
     suspended: bool,
     transitioned: &TransitionedObject,
+    transition_version_state: rustfs_filemeta::TransitionVersionState,
 ) -> Option<Jentry> {
     let sweeper = ObjSweeper {
         version_id,
@@ -468,6 +526,7 @@ pub fn transitioned_delete_journal_entry(
         transition_status: transitioned.status.clone(),
         transition_tier: transitioned.tier.clone(),
         transition_version_id: transitioned.version_id.clone(),
+        transition_version_state,
         remote_object: transitioned.name.clone(),
         ..Default::default()
     };
@@ -475,8 +534,13 @@ pub fn transitioned_delete_journal_entry(
     sweeper.should_remove_remote_object()
 }
 
-pub fn transitioned_force_delete_journal_entry(transitioned: &TransitionedObject) -> Option<Jentry> {
-    if transitioned.status != lifecycle::TRANSITION_COMPLETE {
+pub fn transitioned_force_delete_journal_entry(
+    transitioned: &TransitionedObject,
+    transition_version_state: rustfs_filemeta::TransitionVersionState,
+) -> Option<Jentry> {
+    if transitioned.status != lifecycle::TRANSITION_COMPLETE
+        || transition_version_state == rustfs_filemeta::TransitionVersionState::Unknown
+    {
         return None;
     }
 
@@ -485,7 +549,11 @@ pub fn transitioned_force_delete_journal_entry(transitioned: &TransitionedObject
         version_id: transitioned.version_id.clone(),
         tier_name: transitioned.tier.clone(),
         backend_identity: None,
-        version_id_exact: false,
+        version_id_exact: matches!(
+            transition_version_state,
+            rustfs_filemeta::TransitionVersionState::SuspendedNull | rustfs_filemeta::TransitionVersionState::Exact
+        ),
+        version_state: transition_version_state,
     })
 }
 
@@ -494,11 +562,14 @@ mod test {
     use crate::client::signer_error::invalid_utf8_header_error;
 
     use super::{
-        ERR_REMOTE_DELETE_BREAKER_OPEN, ERR_REMOTE_DELETE_LIMITER_CLOSED, RemoteDeleteBreaker, RemoteTierDeleteOutcome,
+        CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES, ERR_REMOTE_DELETE_BREAKER_OPEN, ERR_REMOTE_DELETE_LIMITER_CLOSED,
+        RemoteDeleteBreaker, RemoteTierDeleteOutcome, delete_confirmed_transition_candidate_exact_with_manager_and_identity,
         delete_object_from_remote_tier_idempotent, delete_object_from_remote_tier_idempotent_with_manager_and_identity,
-        is_remote_tier_not_found_error, is_signer_header_error, set_remote_tier_delete_test_hook,
-        should_record_remote_delete_failure,
+        is_remote_tier_not_found_error, is_signer_header_error, lifecycle, set_remote_tier_delete_test_hook,
+        should_record_remote_delete_failure, transitioned_delete_journal_entry, transitioned_force_delete_journal_entry,
     };
+    use crate::storage_api_contracts::lifecycle::TransitionedObject;
+    use rustfs_filemeta::TransitionVersionState;
     use std::io::{Error, ErrorKind};
     use std::time::{Duration, Instant};
 
@@ -542,6 +613,43 @@ mod test {
         assert!(should_record_remote_delete_failure(&Error::other("NoSuchVersion")));
     }
 
+    #[test]
+    fn transitioned_delete_journal_preserves_remote_version_state() {
+        let cases = [
+            (TransitionVersionState::Unknown, "legacy-version", None),
+            (TransitionVersionState::KnownDisabled, "", Some(false)),
+            (TransitionVersionState::SuspendedNull, "null", Some(true)),
+            (TransitionVersionState::Exact, "opaque-version", Some(true)),
+        ];
+
+        for (state, version_id, expected_exact) in cases {
+            let transitioned = TransitionedObject {
+                name: "remote/object".to_string(),
+                version_id: version_id.to_string(),
+                tier: "WARM".to_string(),
+                status: lifecycle::TRANSITION_COMPLETE.to_string(),
+                ..Default::default()
+            };
+            let regular = transitioned_delete_journal_entry(None, false, false, &transitioned, state);
+            let forced = transitioned_force_delete_journal_entry(&transitioned, state);
+
+            match expected_exact {
+                Some(expected_exact) => {
+                    let regular = regular.expect("known version state should produce a regular delete journal entry");
+                    assert_eq!(regular.version_state, state);
+                    assert_eq!(regular.version_id_exact, expected_exact);
+                    let forced = forced.expect("known version state should produce a forced delete journal entry");
+                    assert_eq!(forced.version_state, state);
+                    assert_eq!(forced.version_id_exact, expected_exact);
+                }
+                None => {
+                    assert!(regular.is_none());
+                    assert!(forced.is_none());
+                }
+            }
+        }
+    }
+
     #[tokio::test]
     #[serial_test::serial]
     async fn idempotent_remote_delete_treats_hooked_nosuchversion_as_already_removed() {
@@ -664,6 +772,55 @@ mod test {
         assert_eq!(backend.remove_versions().await, vec![("remote/object".to_string(), String::new())]);
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn confirmed_transition_cleanup_deletes_exact_provider_token() {
+        CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES.store(0, std::sync::atomic::Ordering::Relaxed);
+        let manager = crate::services::tier::tier::TierConfigMgr::new();
+        let backend = crate::services::tier::test_util::register_mock_tier(&manager, "WARM").await;
+        let lease = crate::services::tier::tier::TierConfigMgr::acquire_operation_lease(&manager, "WARM")
+            .await
+            .expect("test tier lease should be available");
+        let identity = lease.backend_identity();
+        drop(lease);
+        backend.set_reject_non_empty_remote_versions(true);
+
+        let outcome = delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+            "remote/object",
+            "provider-version-token",
+            "WARM",
+            identity,
+            &manager,
+        )
+        .await
+        .expect("confirmed upload compensation should delete the exact provider token");
+
+        assert_eq!(outcome, RemoteTierDeleteOutcome::Deleted);
+        assert_eq!(backend.exact_remove_count(), 1);
+        assert_eq!(
+            backend.remove_versions().await,
+            vec![("remote/object".to_string(), "provider-version-token".to_string())]
+        );
+
+        let err = delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+            "remote/empty-guard-probe",
+            "",
+            "WARM",
+            identity,
+            &manager,
+        )
+        .await
+        .expect_err("confirmed versioned cleanup must reject an empty token");
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidInput);
+        assert_eq!(backend.remove_count().await, 1);
+        assert_eq!(
+            CONFIRMED_TRANSITION_EMPTY_GUARD_DISPATCHES.load(std::sync::atomic::Ordering::Relaxed),
+            0,
+            "empty remote versions must be rejected before exact cleanup dispatch"
+        );
+    }
+
     #[test]
     fn breaker_opens_at_threshold_and_recovers_after_window() {
         let mut breaker = RemoteDeleteBreaker::new(3, Duration::from_secs(30));

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

@@ -22,7 +22,11 @@ use uuid::Uuid;
 
 use crate::bucket::lifecycle::config_boundary;
 use crate::bucket::lifecycle::lifecycle::TRANSITION_COMPLETE;
-use crate::bucket::lifecycle::tier_sweeper::delete_object_from_remote_tier_idempotent_with_manager_and_identity;
+use crate::bucket::lifecycle::tier_sweeper::{
+    delete_confirmed_transition_candidate_exact_with_lease_idempotent,
+    delete_confirmed_transition_candidate_exact_with_manager_and_identity,
+    delete_object_from_remote_tier_idempotent_with_manager_and_identity,
+};
 use crate::disk::RUSTFS_META_BUCKET;
 use crate::error::{Error, Result as EcstoreResult};
 use crate::object_api::ObjectOptions;
@@ -613,6 +617,199 @@ pub enum TransitionTransactionRecoveryOutcome {
     Retained,
 }
 
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+#[serde(rename_all = "snake_case")]
+pub enum TransitionOperatorProbe {
+    Missing,
+    UnversionedPresent,
+    VersionedPresent(String),
+    Ambiguous,
+    Unsupported,
+}
+
+impl From<TransitionCandidateProbe> for TransitionOperatorProbe {
+    fn from(value: TransitionCandidateProbe) -> Self {
+        match value {
+            TransitionCandidateProbe::Missing => Self::Missing,
+            TransitionCandidateProbe::UnversionedPresent => Self::UnversionedPresent,
+            TransitionCandidateProbe::VersionedPresent(version_id) => Self::VersionedPresent(version_id),
+            TransitionCandidateProbe::Ambiguous => Self::Ambiguous,
+            TransitionCandidateProbe::Unsupported => Self::Unsupported,
+        }
+    }
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+pub struct TransitionOperatorStatus {
+    pub transaction_id: Uuid,
+    pub state: TransitionTransactionState,
+    pub tier_name: String,
+    pub remote_object: String,
+    pub not_after_unix_nanos: i64,
+    pub probe: TransitionOperatorProbe,
+}
+
+#[derive(Debug, Clone, PartialEq, Eq, Serialize)]
+pub struct TransitionOperatorDeleteResult {
+    pub status: TransitionOperatorStatus,
+    pub journal_observed_after_delete: bool,
+}
+
+#[derive(Debug, thiserror::Error)]
+pub enum TransitionOperatorError {
+    #[error("transition transaction was not found")]
+    NotFound,
+    #[error("transition transaction is still inside its active ownership window")]
+    NotExpired,
+    #[error("transition transaction state is not eligible for operator reconciliation: {0:?}")]
+    InvalidState(TransitionTransactionState),
+    #[error("an exact non-empty remote version is required")]
+    RemoteVersionRequired,
+    #[error("remote candidate is not proven missing: {0:?}")]
+    CandidateNotMissing(TransitionOperatorProbe),
+    #[error("remote candidate version does not match requested exact version: expected {expected}, observed {actual:?}")]
+    CandidateVersionMismatch {
+        expected: String,
+        actual: TransitionOperatorProbe,
+    },
+    #[error("transition transaction store failed: {0}")]
+    Store(#[source] Error),
+    #[error("remote tier reconciliation failed: {0}")]
+    Remote(#[source] std::io::Error),
+}
+
+type TransitionOperatorResult<T> = std::result::Result<T, TransitionOperatorError>;
+
+fn validate_operator_reconcile_transaction(
+    transaction: &TransitionTransaction,
+    now_unix_nanos: i128,
+) -> TransitionOperatorResult<()> {
+    transaction
+        .validate()
+        .map_err(|err| TransitionOperatorError::Store(Error::other(err)))?;
+    if transaction.state != TransitionTransactionState::UploadOutcomeUnknown {
+        return Err(TransitionOperatorError::InvalidState(transaction.state));
+    }
+    if now_unix_nanos < i128::from(transaction.not_after_unix_nanos) {
+        return Err(TransitionOperatorError::NotExpired);
+    }
+    Ok(())
+}
+
+async fn load_operator_reconcile_transaction(
+    api: Arc<ECStore>,
+    transaction_id: Uuid,
+) -> TransitionOperatorResult<TransitionTransaction> {
+    match load_transition_transaction_record(api, transaction_id).await {
+        Ok(transaction) => Ok(transaction),
+        Err(Error::ConfigNotFound) => Err(TransitionOperatorError::NotFound),
+        Err(err) => Err(TransitionOperatorError::Store(err)),
+    }
+}
+
+async fn operator_probe_transition_candidate(
+    api: Arc<ECStore>,
+    transaction: &TransitionTransaction,
+) -> TransitionOperatorResult<TransitionOperatorProbe> {
+    let lease = TierConfigMgr::acquire_operation_lease_for_backend_identity(
+        &api.tier_config_mgr(),
+        &transaction.tier_name,
+        transaction.backend_fingerprint,
+    )
+    .await
+    .map_err(|err| TransitionOperatorError::Remote(std::io::Error::other(err)))?;
+    lease
+        .probe_transition_candidate_for(&transaction.remote_object, transaction.transaction_id)
+        .await
+        .map(TransitionOperatorProbe::from)
+        .map_err(TransitionOperatorError::Remote)
+}
+
+pub async fn inspect_transition_transaction_for_operator(
+    api: Arc<ECStore>,
+    transaction_id: Uuid,
+) -> TransitionOperatorResult<TransitionOperatorStatus> {
+    let transaction = load_operator_reconcile_transaction(api.clone(), transaction_id).await?;
+    validate_operator_reconcile_transaction(&transaction, time::OffsetDateTime::now_utc().unix_timestamp_nanos())?;
+    let probe = operator_probe_transition_candidate(api, &transaction).await?;
+    Ok(TransitionOperatorStatus {
+        transaction_id,
+        state: transaction.state,
+        tier_name: transaction.tier_name,
+        remote_object: transaction.remote_object,
+        not_after_unix_nanos: transaction.not_after_unix_nanos,
+        probe,
+    })
+}
+
+pub async fn delete_transition_candidate_for_operator(
+    api: Arc<ECStore>,
+    transaction_id: Uuid,
+    remote_version_id: &str,
+) -> TransitionOperatorResult<TransitionOperatorDeleteResult> {
+    if remote_version_id.is_empty() {
+        return Err(TransitionOperatorError::RemoteVersionRequired);
+    }
+    let transaction = load_operator_reconcile_transaction(api.clone(), transaction_id).await?;
+    validate_operator_reconcile_transaction(&transaction, time::OffsetDateTime::now_utc().unix_timestamp_nanos())?;
+    let lease = TierConfigMgr::acquire_operation_lease_for_backend_identity(
+        &api.tier_config_mgr(),
+        &transaction.tier_name,
+        transaction.backend_fingerprint,
+    )
+    .await
+    .map_err(|err| TransitionOperatorError::Remote(std::io::Error::other(err)))?;
+    lease
+        .validate_remote_version_id(remote_version_id)
+        .map_err(TransitionOperatorError::Remote)?;
+    let before_delete_probe = lease
+        .probe_transition_candidate_for(&transaction.remote_object, transaction.transaction_id)
+        .await
+        .map(TransitionOperatorProbe::from)
+        .map_err(TransitionOperatorError::Remote)?;
+    if !matches!(&before_delete_probe, TransitionOperatorProbe::VersionedPresent(version_id) if version_id == remote_version_id) {
+        return Err(TransitionOperatorError::CandidateVersionMismatch {
+            expected: remote_version_id.to_string(),
+            actual: before_delete_probe,
+        });
+    }
+    delete_confirmed_transition_candidate_exact_with_lease_idempotent(&transaction.remote_object, remote_version_id, &lease)
+        .await
+        .map_err(TransitionOperatorError::Remote)?;
+    let probe = operator_probe_transition_candidate(api.clone(), &transaction).await?;
+    let journal_observed_after_delete = match load_transition_transaction_record(api, transaction_id).await {
+        Ok(_) => true,
+        Err(Error::ConfigNotFound) => false,
+        Err(err) => return Err(TransitionOperatorError::Store(err)),
+    };
+    Ok(TransitionOperatorDeleteResult {
+        status: TransitionOperatorStatus {
+            transaction_id,
+            state: transaction.state,
+            tier_name: transaction.tier_name,
+            remote_object: transaction.remote_object,
+            not_after_unix_nanos: transaction.not_after_unix_nanos,
+            probe,
+        },
+        journal_observed_after_delete,
+    })
+}
+
+pub async fn finalize_missing_transition_transaction_for_operator(
+    api: Arc<ECStore>,
+    transaction_id: Uuid,
+) -> TransitionOperatorResult<()> {
+    let transaction = load_operator_reconcile_transaction(api.clone(), transaction_id).await?;
+    validate_operator_reconcile_transaction(&transaction, time::OffsetDateTime::now_utc().unix_timestamp_nanos())?;
+    let probe = operator_probe_transition_candidate(api.clone(), &transaction).await?;
+    if probe != TransitionOperatorProbe::Missing {
+        return Err(TransitionOperatorError::CandidateNotMissing(probe));
+    }
+    delete_transition_transaction_record(api, transaction_id)
+        .await
+        .map_err(TransitionOperatorError::Store)
+}
+
 pub(crate) fn decode_transition_transaction_record(object: &str, data: &[u8]) -> Result<TransitionTransaction> {
     let transaction_id = transition_transaction_id_from_record_object_name(object)?;
     TransitionTransaction::decode(transaction_id, data)
@@ -697,7 +894,7 @@ async fn recover_unknown_upload_outcome(
     .map_err(Error::other)?;
 
     match lease
-        .probe_transition_candidate(&transaction.remote_object)
+        .probe_transition_candidate_for(&transaction.remote_object, transaction.transaction_id)
         .await
         .map_err(Error::other)?
     {
@@ -708,6 +905,21 @@ async fn recover_unknown_upload_outcome(
         TransitionCandidateProbe::UnversionedPresent => {
             cleanup_recovered_unknown_upload_candidate(api, transaction, TransitionRemoteVersion::unversioned()).await
         }
+        TransitionCandidateProbe::VersionedPresent(version_id)
+            if Uuid::parse_str(&version_id).is_ok_and(|version_id| version_id.is_nil()) =>
+        {
+            delete_confirmed_transition_candidate_exact_with_manager_and_identity(
+                &transaction.remote_object,
+                &version_id,
+                &transaction.tier_name,
+                transaction.backend_fingerprint,
+                &api.tier_config_mgr(),
+            )
+            .await
+            .map_err(Error::other)?;
+            delete_transition_transaction_record(api, transaction.transaction_id).await?;
+            Ok(TransitionTransactionRecoveryOutcome::RemoteCandidateDeleted)
+        }
         TransitionCandidateProbe::VersionedPresent(version_id) => {
             cleanup_recovered_unknown_upload_candidate(api, transaction, TransitionRemoteVersion::versioned(version_id)).await
         }
@@ -1079,6 +1291,27 @@ mod tests {
             .expect("upload state change should succeed")
     }
 
+    #[test]
+    fn operator_reconcile_requires_expired_unknown_upload_outcome() {
+        let mut transaction = new_transaction();
+        let active_deadline = transaction.not_after_unix_nanos;
+
+        assert!(matches!(
+            validate_operator_reconcile_transaction(&transaction, i128::from(active_deadline) + 1),
+            Err(TransitionOperatorError::InvalidState(TransitionTransactionState::UploadStarted))
+        ));
+
+        transaction
+            .advance(transaction.fence(), TransitionTransactionState::UploadOutcomeUnknown, None)
+            .expect("unknown upload outcome should be recorded");
+        assert!(matches!(
+            validate_operator_reconcile_transaction(&transaction, i128::from(active_deadline) - 1),
+            Err(TransitionOperatorError::NotExpired)
+        ));
+        validate_operator_reconcile_transaction(&transaction, i128::from(active_deadline))
+            .expect("expired unknown upload outcome should be eligible");
+    }
+
     fn cleanup_proof(transaction: &TransitionTransaction, decision: TransitionCleanupDecision) -> TransitionCleanupProof {
         TransitionCleanupProof {
             transaction_id: transaction.transaction_id,

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

@@ -46,7 +46,7 @@ mod runtime_boundary;
 pub use datatypes::ResyncStatusType;
 pub use replication_config_boundary::{
     ObjectOpts, ReplicationConfigurationExt, ReplicationTargetValidationError, replication_target_arns,
-    should_remove_replication_target, validate_replication_config_target_arns,
+    should_remove_replication_target, unsupported_replication_config_field, validate_replication_config_target_arns,
 };
 #[cfg(test)]
 pub(crate) use replication_filemeta_boundary::ReplicateTargetDecision;

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

@@ -14,5 +14,5 @@
 
 pub use rustfs_replication::{
     ObjectOpts, ReplicationConfigurationExt, ReplicationTargetValidationError, replication_target_arns,
-    should_remove_replication_target, validate_replication_config_target_arns,
+    should_remove_replication_target, unsupported_replication_config_field, validate_replication_config_target_arns,
 };

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

@@ -14,8 +14,11 @@
 
 use std::{collections::HashMap, sync::Arc};
 
+use super::replication_error_boundary::Result;
 use super::replication_filemeta_boundary::{ReplicateDecision, ReplicationStatusType, ReplicationType};
-use super::replication_object_config::{check_replicate_delete, get_must_replicate_options, must_replicate};
+use super::replication_object_config::{
+    check_replicate_delete, check_replicate_delete_strict, get_must_replicate_options, must_replicate,
+};
 use super::replication_object_decision_boundary::MustReplicateOptions;
 use super::replication_pool::{schedule_replication, schedule_replication_delete};
 use super::replication_queue_boundary::DeletedObjectReplicationInfo;
@@ -50,6 +53,16 @@ impl ReplicationObjectBridge {
         check_replicate_delete(bucket, object, source, opts, get_error).await
     }
 
+    pub async fn check_delete_strict(
+        bucket: &str,
+        object: &ObjectToDelete,
+        source: &ObjectInfo,
+        opts: &ObjectOptions,
+        get_error: Option<String>,
+    ) -> Result<ReplicateDecision> {
+        check_replicate_delete_strict(bucket, object, source, opts, get_error).await
+    }
+
     pub async fn schedule_object<S: ReplicationStorage>(
         object: ObjectInfo,
         storage: Arc<S>,

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

@@ -169,9 +169,8 @@ pub(crate) async fn check_replicate_delete(
     del_opts: &ObjectOptions,
     gerr: Option<String>,
 ) -> ReplicateDecision {
-    let rcfg = match get_replication_config(bucket).await {
-        Ok(Some(config)) => config,
-        Ok(None) => return ReplicateDecision::default(),
+    match check_replicate_delete_strict(bucket, dobj, oi, del_opts, gerr).await {
+        Ok(decision) => decision,
         Err(err) => {
             error!(
                 event = EVENT_RESYNC_CONFIG_LOOKUP_SKIPPED,
@@ -182,16 +181,30 @@ pub(crate) async fn check_replicate_delete(
                 error = %err,
                 "Failed to look up replication config for delete replication"
             );
-            return ReplicateDecision::default();
+            ReplicateDecision::default()
         }
+    }
+}
+
+pub(crate) async fn check_replicate_delete_strict(
+    bucket: &str,
+    dobj: &ObjectToDelete,
+    oi: &ObjectInfo,
+    del_opts: &ObjectOptions,
+    gerr: Option<String>,
+) -> Result<ReplicateDecision> {
+    let rcfg = match get_replication_config(bucket).await {
+        Ok(Some(config)) => config,
+        Ok(None) => return Ok(ReplicateDecision::default()),
+        Err(err) => return Err(err),
     };
 
     if del_opts.replication_request {
-        return ReplicateDecision::default();
+        return Ok(ReplicateDecision::default());
     }
 
-    if !del_opts.versioned {
-        return ReplicateDecision::default();
+    if !del_opts.versioned && !del_opts.version_suspended {
+        return Ok(ReplicateDecision::default());
     }
 
     let replication_delete = object_to_delete_for_replication(dobj);
@@ -209,7 +222,7 @@ pub(crate) async fn check_replicate_delete(
     let mut dsc = ReplicateDecision::new();
 
     if tgt_arns.is_empty() {
-        return dsc;
+        return Ok(dsc);
     }
 
     for tgt_arn in tgt_arns {
@@ -239,7 +252,7 @@ pub(crate) async fn check_replicate_delete(
         dsc.set(tgt_dsc);
     }
 
-    dsc
+    Ok(dsc)
 }
 
 pub(crate) async fn must_replicate(bucket: &str, object: &str, mopts: MustReplicateOptions) -> ReplicateDecision {

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

@@ -1151,60 +1151,6 @@ pub async fn replicate_delete<S: ReplicationStorage>(dobj: DeletedObjectReplicat
         dobj.delete_object.version_id
     };
 
-    let _rcfg = match get_replication_config(&bucket).await {
-        Ok(Some(config)) => config,
-        Ok(None) => {
-            debug!(
-                event = EVENT_REPLICATION_DELETE_SKIPPED,
-                component = LOG_COMPONENT_ECSTORE,
-                subsystem = LOG_SUBSYSTEM_REPLICATION_RESYNC,
-                bucket = %bucket,
-                reason = "replication_config_missing",
-                "Skipping replication delete because replication config is missing"
-            );
-            send_local_event(EventArgs {
-                event_name: EventName::ObjectReplicationNotTracked.to_string(),
-                bucket_name: bucket.clone(),
-                object: ObjectInfo {
-                    bucket: bucket.clone(),
-                    name: dobj.delete_object.object_name.clone(),
-                    version_id,
-                    delete_marker: dobj.delete_object.delete_marker,
-                    ..Default::default()
-                },
-                user_agent: "Internal: [Replication]".to_string(),
-                ..Default::default()
-            });
-
-            return;
-        }
-        Err(err) => {
-            debug!(
-                event = EVENT_REPLICATION_DELETE_SKIPPED,
-                component = LOG_COMPONENT_ECSTORE,
-                subsystem = LOG_SUBSYSTEM_REPLICATION_RESYNC,
-                bucket = %bucket,
-                error = %err,
-                reason = "replication_config_lookup_failed",
-                "Skipping replication delete because replication config lookup failed"
-            );
-            send_local_event(EventArgs {
-                event_name: EventName::ObjectReplicationNotTracked.to_string(),
-                bucket_name: bucket.clone(),
-                object: ObjectInfo {
-                    bucket: bucket.clone(),
-                    name: dobj.delete_object.object_name.clone(),
-                    version_id,
-                    delete_marker: dobj.delete_object.delete_marker,
-                    ..Default::default()
-                },
-                user_agent: "Internal: [Replication]".to_string(),
-                ..Default::default()
-            });
-            return;
-        }
-    };
-
     if dobj.delete_object.delete_marker
         && let Some(delete_marker_version_id) = dobj.delete_object.delete_marker_version_id
     {

crates/ecstore/src/client/provider_versions.rs

@@ -172,7 +172,7 @@ impl ProviderVersionCapabilities {
     }
 }
 
-fn validate_remote_version_id(version_id: &str) -> Result<(), Error> {
+pub(crate) fn validate_remote_version_id(version_id: &str) -> Result<(), Error> {
     if version_id.is_empty() {
         return Err(Error::new(
             ErrorKind::InvalidData,

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

@@ -12,11 +12,20 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use crate::cluster::rpc::http_auth::RPC_CONTENT_SHA256_HEADER;
-use crate::cluster::rpc::{gen_tonic_signature_headers, normalize_tonic_rpc_audience};
-use crate::disk::error::{DiskError, Error as DiskErrorType};
+#[cfg(test)]
+use crate::cluster::rpc::http_auth::RPC_REPLAY_SCOPE_VERSION_HEADER;
+use crate::cluster::rpc::http_auth::{
+    RPC_AUTH_VERSION_HEADER, RPC_AUTH_VERSION_V2, RPC_BOOT_EPOCH_CHALLENGE_HEADER, RPC_BOOT_EPOCH_HEADER,
+    RPC_BOOT_EPOCH_PROOF_HEADER, RPC_CONTENT_SHA256_HEADER, TIMESTAMP_HEADER,
+};
+use crate::cluster::rpc::{
+    gen_tonic_replay_scope_headers, gen_tonic_signature_headers, normalize_tonic_rpc_audience, verify_tonic_boot_epoch_response,
+};
+#[cfg(test)]
+use crate::cluster::rpc::{tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers};
+use crate::disk::error::{DiskError, Error as DiskErrorType, RpcStatusError};
 use crate::runtime::sources as runtime_sources;
-use http::Uri;
+use http::{Request as HttpRequest, Response as HttpResponse, Uri};
 use rustfs_protos::{
     ChannelClass, create_new_channel, get_channel_for_class,
     proto_gen::node_service::{
@@ -24,9 +33,19 @@ use rustfs_protos::{
         tier_mutation_control_service_client::TierMutationControlServiceClient,
     },
 };
-use std::{error::Error, io::ErrorKind};
+use std::{
+    collections::HashMap,
+    error::Error,
+    future::Future,
+    io::ErrorKind,
+    pin::Pin,
+    sync::{LazyLock, Mutex},
+    task::{Context, Poll},
+};
 use tonic::{service::interceptor::InterceptedService, transport::Channel};
+use tower::Service;
 use tracing::debug;
+use uuid::Uuid;
 
 use super::context_propagation::{inject_request_id_into_metadata, inject_trace_context_into_metadata};
 
@@ -35,7 +54,7 @@ use super::context_propagation::{inject_request_id_into_metadata, inject_trace_c
 pub async fn node_service_time_out_client(
     addr: &String,
     interceptor: TonicInterceptor,
-) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>, Box<dyn Error>> {
+) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>, Box<dyn Error>> {
     // Default to the latency-sensitive control channel; bulk `bytes` RPCs opt in via the
     // `_for_class` variant below (grpc-optimization P1).
     node_service_time_out_client_for_class(addr, interceptor, ChannelClass::Control).await
@@ -44,13 +63,14 @@ pub async fn node_service_time_out_client(
 pub async fn heal_control_time_out_client(
     addr: &str,
     interceptor: TonicInterceptor,
-) -> Result<HealControlServiceClient<InterceptedService<Channel, TonicInterceptor>>, Box<dyn Error>> {
+) -> Result<HealControlServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>, Box<dyn Error>> {
     let interceptor = interceptor.with_rpc_audience(addr)?;
     let channel = match runtime_sources::cached_node_channel(addr).await {
         Some(channel) => channel,
         None => create_new_channel(addr).await?,
     };
     let max_message_size = rustfs_protos::HEAL_CONTROL_RPC_MAX_MESSAGE_SIZE;
+    let channel = ReplayScopeChannel::new(channel, interceptor.replay_scope_audience());
     Ok(HealControlServiceClient::with_interceptor(channel, interceptor)
         .max_decoding_message_size(max_message_size)
         .max_encoding_message_size(max_message_size))
@@ -59,13 +79,14 @@ pub async fn heal_control_time_out_client(
 pub async fn tier_mutation_control_time_out_client(
     addr: &str,
     interceptor: TonicInterceptor,
-) -> Result<TierMutationControlServiceClient<InterceptedService<Channel, TonicInterceptor>>, Box<dyn Error>> {
+) -> Result<TierMutationControlServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>, Box<dyn Error>> {
     let interceptor = interceptor.with_rpc_audience(addr)?;
     let channel = match runtime_sources::cached_node_channel(addr).await {
         Some(channel) => channel,
         None => create_new_channel(addr).await?,
     };
     let max_message_size = rustfs_protos::TIER_MUTATION_RPC_MAX_MESSAGE_SIZE;
+    let channel = ReplayScopeChannel::new(channel, interceptor.replay_scope_audience());
     Ok(TierMutationControlServiceClient::with_interceptor(channel, interceptor)
         .max_decoding_message_size(max_message_size)
         .max_encoding_message_size(max_message_size))
@@ -81,7 +102,7 @@ pub async fn node_service_time_out_client_for_class(
     addr: &String,
     interceptor: TonicInterceptor,
     class: ChannelClass,
-) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>, Box<dyn Error>> {
+) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>, Box<dyn Error>> {
     let interceptor = interceptor.with_rpc_audience(addr)?;
     let channel = match class {
         ChannelClass::Control => match runtime_sources::cached_node_channel(addr).await {
@@ -96,6 +117,7 @@ pub async fn node_service_time_out_client_for_class(
     };
 
     let max_message_size = rustfs_protos::internode_rpc_max_message_size();
+    let channel = ReplayScopeChannel::new(channel, interceptor.replay_scope_audience());
     Ok(NodeServiceClient::with_interceptor(channel, interceptor)
         .max_decoding_message_size(max_message_size)
         .max_encoding_message_size(max_message_size))
@@ -103,14 +125,83 @@ pub async fn node_service_time_out_client_for_class(
 
 pub async fn node_service_time_out_client_no_auth(
     addr: &String,
-) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>, Box<dyn Error>> {
+) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>, Box<dyn Error>> {
     node_service_time_out_client(addr, TonicInterceptor::NoOp(NoOpInterceptor)).await
 }
 
+/// The typed `tonic::Status` an internode RPC failure was converted from, if
+/// this error carries one.
+pub(crate) fn embedded_tonic_status(io_err: &std::io::Error) -> Option<&tonic::Status> {
+    io_err.get_ref()?.downcast_ref::<RpcStatusError>().map(RpcStatusError::status)
+}
+
+/// Decide whether a gRPC status reports a peer we cannot currently reach,
+/// rather than an application outcome from a live peer.
+///
+/// `Unavailable` is the one code that means "no service behind this channel":
+/// the client transport raises it when the connection is broken, and the
+/// server's own not-ready gates use it deliberately.
+///
+/// `Unknown` is the client transport's escape hatch for a cause it could not
+/// map to a code — tower's "Service was not ready: <cause>", an h2 error with
+/// no gRPC mapping. Our handlers never return it, so there its message is the
+/// only evidence available and the anchored needles decide.
+///
+/// Every other code is an answer from a live peer and is never a transport
+/// failure, whatever its message says. That distinction is the point of
+/// classifying by code: a peer relaying its own downstream trouble as
+/// `Internal("connection refused ...")`, or a handler interpolating a local
+/// `io::Error` into `Status::internal`, answered us perfectly well. Marking it
+/// offline over that text is the bug this classification replaces. Likewise a
+/// `Cancelled` "Timeout expired" from the per-RPC channel deadline means the
+/// peer is slow, not gone; gating it would turn load into a partition.
+pub(crate) fn is_network_like_status(status: &tonic::Status) -> bool {
+    match status.code() {
+        tonic::Code::Unavailable => true,
+        tonic::Code::Unknown => message_has_network_needle(&status.to_string()),
+        _ => false,
+    }
+}
+
+/// Substring fallback for failures that only exist as text: dial errors
+/// wrapped by `get_client`, remote `error_info` payloads, and statuses
+/// flattened through `format!`. Needles must stay anchored to transport
+/// context — a bare word like "unavailable" also matches application text
+/// (e.g. a bucket named "unavailable-logs") and would take a healthy peer
+/// offline.
+pub(crate) fn message_has_network_needle(message: &str) -> bool {
+    let message = message.to_ascii_lowercase();
+    [
+        "temporarily offline",
+        "transport error",
+        // tonic >= 0.14 renders Code::Unavailable as
+        // `code: 'The service is currently unavailable'`.
+        "code: 'the service is currently unavailable'",
+        // RUSTFS_COMPAT_TODO(tonic-013-status-render): releases up to 1.0.0-alpha.38 shipped tonic 0.13, which rendered the same status as `status: Unavailable`, and peers relay that text in error_info. Remove after the minimum supported RustFS peer version ships tonic >= 0.14.
+        "status: unavailable",
+        "error trying to connect",
+        "connection refused",
+        "connection reset",
+        "broken pipe",
+        "not connected",
+        "unexpected eof",
+        "timed out",
+        "deadline has elapsed",
+        "connection closed",
+        "connection aborted",
+        "tcp connect error",
+    ]
+    .iter()
+    .any(|needle| message.contains(needle))
+}
+
 pub(crate) fn is_network_like_disk_error(err: &DiskErrorType) -> bool {
     match err {
         DiskError::Timeout => true,
         DiskError::Io(io_err) => {
+            if let Some(status) = embedded_tonic_status(io_err) {
+                return is_network_like_status(status);
+            }
             if matches!(
                 io_err.kind(),
                 ErrorKind::TimedOut
@@ -124,29 +215,110 @@ pub(crate) fn is_network_like_disk_error(err: &DiskErrorType) -> bool {
                 return true;
             }
 
-            let message = io_err.to_string().to_ascii_lowercase();
-            [
-                "transport error",
-                "unavailable",
-                "error trying to connect",
-                "connection refused",
-                "connection reset",
-                "broken pipe",
-                "not connected",
-                "unexpected eof",
-                "timed out",
-                "deadline has elapsed",
-                "connection closed",
-                "connection aborted",
-                "tcp connect error",
-            ]
-            .iter()
-            .any(|needle| message.contains(needle))
+            message_has_network_needle(&io_err.to_string())
         }
         _ => false,
     }
 }
 
+/// The transport service that learns an authenticated peer boot epoch and adds the replay-scoped
+/// signature only after one has been observed. The v1/v2 interceptor stays inside this wrapper so
+/// old servers continue receiving precisely the metadata they understand.
+#[derive(Clone, Debug)]
+pub struct ReplayScopeChannel<S> {
+    inner: S,
+    audience: Option<String>,
+}
+
+/// The channel type used by internode clients after v2 authentication and replay-scope handling.
+pub type AuthenticatedChannel = ReplayScopeChannel<Channel>;
+
+static PEER_BOOT_EPOCHS: LazyLock<Mutex<HashMap<String, Uuid>>> = LazyLock::new(|| Mutex::new(HashMap::new()));
+
+impl<S> ReplayScopeChannel<S> {
+    fn new(inner: S, audience: Option<String>) -> Self {
+        Self { inner, audience }
+    }
+}
+
+fn cached_peer_boot_epoch(audience: &str) -> Option<Uuid> {
+    PEER_BOOT_EPOCHS.lock().ok().and_then(|epochs| epochs.get(audience).copied())
+}
+
+fn remember_peer_boot_epoch(audience: String, epoch: Uuid) {
+    if let Ok(mut epochs) = PEER_BOOT_EPOCHS.lock() {
+        epochs.insert(audience, epoch);
+    }
+}
+
+impl<S, ReqBody, ResBody> Service<HttpRequest<ReqBody>> for ReplayScopeChannel<S>
+where
+    S: Service<HttpRequest<ReqBody>, Response = HttpResponse<ResBody>>,
+    S::Error: Send + 'static,
+    S::Future: Send + 'static,
+    ReqBody: Send + 'static,
+    ResBody: Send + 'static,
+{
+    type Response = HttpResponse<ResBody>;
+    type Error = S::Error;
+    type Future = Pin<Box<dyn Future<Output = Result<Self::Response, Self::Error>> + Send>>;
+
+    fn poll_ready(&mut self, cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
+        self.inner.poll_ready(cx)
+    }
+
+    fn call(&mut self, mut request: HttpRequest<ReqBody>) -> Self::Future {
+        let authenticated = self.audience.as_ref().is_some_and(|_| {
+            request
+                .headers()
+                .get(RPC_AUTH_VERSION_HEADER)
+                .and_then(|value| value.to_str().ok())
+                == Some(RPC_AUTH_VERSION_V2)
+        });
+        let challenge = authenticated.then(Uuid::new_v4);
+        if let (Some(audience), Some(challenge)) = (self.audience.as_deref(), challenge) {
+            // The challenge is independently HMAC-authenticated by the response proof. It is not
+            // part of v2 so old peers ignore it, while a new peer can safely advertise its epoch.
+            request.headers_mut().insert(
+                RPC_BOOT_EPOCH_CHALLENGE_HEADER,
+                challenge.to_string().parse().expect("UUID must be a valid header value"),
+            );
+            if let (Some(boot_epoch), Some(timestamp), Some(content_sha256)) = (
+                cached_peer_boot_epoch(audience),
+                request.headers().get(TIMESTAMP_HEADER).and_then(|value| value.to_str().ok()),
+                request
+                    .headers()
+                    .get(RPC_CONTENT_SHA256_HEADER)
+                    .and_then(|value| value.to_str().ok()),
+            ) {
+                match gen_tonic_replay_scope_headers(audience, request.uri().path(), timestamp, content_sha256, boot_epoch) {
+                    Ok(headers) => request.headers_mut().extend(headers),
+                    Err(error) => debug!(error = %error, "could not attach replay-scoped RPC signature"),
+                }
+            }
+        }
+
+        let audience = self.audience.clone();
+        let future = self.inner.call(request);
+        Box::pin(async move {
+            let response = future.await?;
+            if let (Some(audience), Some(challenge)) = (audience, challenge) {
+                match verify_tonic_boot_epoch_response(&audience, challenge, response.headers()) {
+                    Ok(epoch) => remember_peer_boot_epoch(audience, epoch),
+                    Err(error)
+                        if response.headers().contains_key(RPC_BOOT_EPOCH_HEADER)
+                            || response.headers().contains_key(RPC_BOOT_EPOCH_PROOF_HEADER) =>
+                    {
+                        debug!(error = %error, "peer boot epoch response proof was rejected")
+                    }
+                    Err(_) => {}
+                }
+            }
+            Ok(response)
+        })
+    }
+}
+
 pub struct TonicSignatureInterceptor {
     audience: Option<String>,
 }
@@ -205,6 +377,13 @@ impl TonicInterceptor {
         }
         Ok(self)
     }
+
+    fn replay_scope_audience(&self) -> Option<String> {
+        match self {
+            Self::Signature(interceptor) => interceptor.audience.clone(),
+            Self::NoOp(_) => None,
+        }
+    }
 }
 
 impl tonic::service::Interceptor for TonicInterceptor {
@@ -227,6 +406,38 @@ mod tests {
     use tracing_opentelemetry::OpenTelemetrySpanExt;
     use tracing_subscriber::{Registry, layer::SubscriberExt};
 
+    #[derive(Clone)]
+    struct EpochProofService {
+        audience: String,
+        seen_headers: std::sync::Arc<Mutex<Vec<http::HeaderMap>>>,
+    }
+
+    impl Service<HttpRequest<()>> for EpochProofService {
+        type Response = HttpResponse<()>;
+        type Error = std::convert::Infallible;
+        type Future = std::future::Ready<Result<Self::Response, Self::Error>>;
+
+        fn poll_ready(&mut self, _cx: &mut Context<'_>) -> Poll<Result<(), Self::Error>> {
+            Poll::Ready(Ok(()))
+        }
+
+        fn call(&mut self, request: HttpRequest<()>) -> Self::Future {
+            self.seen_headers
+                .lock()
+                .expect("test header capture lock must not be poisoned")
+                .push(request.headers().clone());
+            let challenge = tonic_boot_epoch_challenge(request.headers())
+                .expect("client challenge must be syntactically valid")
+                .expect("authenticated client request must carry a boot epoch challenge");
+            let mut response = HttpResponse::new(());
+            response.headers_mut().extend(
+                tonic_boot_epoch_response_headers(&self.audience, challenge)
+                    .expect("test server must be able to sign an epoch proof"),
+            );
+            std::future::ready(Ok(response))
+        }
+    }
+
     fn ensure_test_rpc_secret() {
         runtime_sources::ensure_test_rpc_secret();
     }
@@ -269,6 +480,70 @@ mod tests {
         let _ = provider.shutdown();
     }
 
+    #[test]
+    fn network_like_disk_error_uses_typed_status_code() {
+        // Transport-level Unavailable statuses justify retry/eviction.
+        assert!(is_network_like_disk_error(&DiskError::from(tonic::Status::unavailable(
+            "storage layer is not initialized"
+        ))));
+        // Application statuses from a live peer must not look network-like,
+        // even when their message contains transport-sounding words.
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::internal(
+            "failed to heal bucket \"unavailable-logs\""
+        ))));
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::unauthenticated(
+            "No valid auth token"
+        ))));
+        // A slow peer that blew the per-RPC deadline is still answering.
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::cancelled("Timeout expired"))));
+    }
+
+    #[test]
+    fn embedded_tonic_status_is_recovered_across_error_conversions() {
+        // DiskError and StorageError share one wrapper, so a status keeps its
+        // typed classification whichever error it was converted into first.
+        let from_storage: DiskErrorType = crate::error::Error::from(tonic::Status::unavailable("peer gone")).into();
+        let DiskError::Io(io_err) = &from_storage else {
+            panic!("status-derived disk error should stay an Io error");
+        };
+        assert_eq!(embedded_tonic_status(io_err).map(|status| status.code()), Some(tonic::Code::Unavailable));
+
+        let from_disk = crate::error::Error::from(DiskError::from(tonic::Status::unavailable("peer gone")));
+        let crate::error::Error::Io(io_err) = &from_disk else {
+            panic!("status-derived storage error should stay an Io error");
+        };
+        assert_eq!(embedded_tonic_status(io_err).map(|status| status.code()), Some(tonic::Code::Unavailable));
+    }
+
+    #[test]
+    fn network_like_disk_error_ignores_transport_words_in_application_statuses() {
+        // Same contract as the peer client: a status the peer answered with
+        // is not a transport failure, so it must not drive a reconnect even
+        // when its message describes one.
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::internal(
+            "connection refused while dialing downstream backend"
+        ))));
+        assert!(!is_network_like_disk_error(&DiskError::from(tonic::Status::unauthenticated(
+            "connection reset while validating token"
+        ))));
+    }
+
+    #[test]
+    fn network_like_disk_error_requires_anchored_unavailable_needle() {
+        // Regression: a bare "unavailable" needle used to match application
+        // text such as a bucket name.
+        assert!(!is_network_like_disk_error(&DiskError::other("bucket \"unavailable-logs\" not found")));
+        // Anchored renderings of a flattened Unavailable status still match.
+        assert!(is_network_like_disk_error(&DiskError::other(
+            "code: 'The service is currently unavailable', message: \"peer gone\""
+        )));
+        assert!(is_network_like_disk_error(&DiskError::other(
+            "status: Unavailable, message: \"peer gone\""
+        )));
+        assert!(is_network_like_disk_error(&DiskError::other("connection refused")));
+        assert!(!is_network_like_disk_error(&DiskError::FileNotFound));
+    }
+
     #[test]
     fn test_signature_interceptor_keeps_auth_headers() {
         ensure_test_rpc_secret();
@@ -304,6 +579,52 @@ mod tests {
         assert_eq!(interceptor.audience.as_deref(), Some("node-a:9000"));
     }
 
+    #[test]
+    fn replay_scope_channel_uses_epoch_proof_before_sending_v3() {
+        ensure_test_rpc_secret();
+        let audience = "replay-scope-client-test:9000";
+        PEER_BOOT_EPOCHS
+            .lock()
+            .expect("peer epoch cache lock must not be poisoned")
+            .remove(audience);
+        let seen_headers = std::sync::Arc::new(Mutex::new(Vec::new()));
+        let service = EpochProofService {
+            audience: audience.to_string(),
+            seen_headers: seen_headers.clone(),
+        };
+        let mut channel = ReplayScopeChannel::new(service, Some(audience.to_string()));
+        let make_request = || {
+            let mut request = HttpRequest::builder()
+                .uri("/node_service.NodeService/Ping")
+                .body(())
+                .expect("test RPC request must build");
+            request.headers_mut().extend(
+                gen_tonic_signature_headers(audience, "node_service.NodeService", "Ping", None)
+                    .expect("v2 test headers must mint"),
+            );
+            request
+        };
+
+        futures::executor::block_on(channel.call(make_request())).expect("first request must complete");
+        futures::executor::block_on(channel.call(make_request())).expect("second request must complete");
+
+        let headers = seen_headers.lock().expect("test header capture lock must not be poisoned");
+        assert_eq!(headers.len(), 2);
+        assert!(headers[0].contains_key(RPC_BOOT_EPOCH_CHALLENGE_HEADER));
+        assert!(
+            !headers[0].contains_key(RPC_REPLAY_SCOPE_VERSION_HEADER),
+            "the first request must remain v2-compatible until the peer proves its epoch"
+        );
+        assert!(
+            headers[1].contains_key(RPC_REPLAY_SCOPE_VERSION_HEADER),
+            "the second request must carry the replay-scoped v3 signature"
+        );
+        PEER_BOOT_EPOCHS
+            .lock()
+            .expect("peer epoch cache lock must not be poisoned")
+            .remove(audience);
+    }
+
     #[test]
     fn test_signature_interceptor_requires_generated_method_metadata() {
         ensure_test_rpc_secret();

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

@@ -20,8 +20,8 @@
 //! rustfs/rustfs#4402) is anchored by the `ghsa_r5qv_*` tests in the module
 //! below, plus the broader negative-signature suite. The advisory class is: a
 //! node must never accept an RPC whose auth is missing, malformed, or signed
-//! with the default/empty shared secret. Body-bound v2 requests additionally
-//! receive process-local replay protection. See
+//! with the default/empty shared secret. Body-bound v2 requests and all replay-scoped v3
+//! requests additionally receive process-local replay protection. See
 //! `docs/testing/security-regressions.md` for the full advisory -> test map.
 //!
 //! Advisory: <https://github.com/rustfs/rustfs/security/advisories/GHSA-r5qv-rc46-hv8q>
@@ -50,13 +50,22 @@ use uuid::Uuid;
 type HmacSha256 = Hmac<Sha256>;
 
 const SIGNATURE_HEADER: &str = "x-rustfs-signature";
-const TIMESTAMP_HEADER: &str = "x-rustfs-timestamp";
-const RPC_AUTH_VERSION_HEADER: &str = "x-rustfs-rpc-auth-version";
+pub(crate) const TIMESTAMP_HEADER: &str = "x-rustfs-timestamp";
+pub(crate) const RPC_AUTH_VERSION_HEADER: &str = "x-rustfs-rpc-auth-version";
 const RPC_SIGNATURE_V2_HEADER: &str = "x-rustfs-rpc-signature-v2";
 const RPC_NONCE_HEADER: &str = "x-rustfs-rpc-nonce";
 pub(crate) const RPC_CONTENT_SHA256_HEADER: &str = "x-rustfs-content-sha256";
-const RPC_AUTH_VERSION_V2: &str = "2";
+pub(crate) const RPC_AUTH_VERSION_V2: &str = "2";
+pub const RPC_REPLAY_SCOPE_VERSION_HEADER: &str = "x-rustfs-rpc-replay-scope-version";
+pub const RPC_REPLAY_SCOPE_SIGNATURE_HEADER: &str = "x-rustfs-rpc-signature-v3";
+pub const RPC_REPLAY_SCOPE_NONCE_HEADER: &str = "x-rustfs-rpc-replay-nonce";
+pub const RPC_BOOT_EPOCH_HEADER: &str = "x-rustfs-rpc-boot-epoch";
+pub const RPC_BOOT_EPOCH_CHALLENGE_HEADER: &str = "x-rustfs-rpc-boot-epoch-challenge";
+pub const RPC_BOOT_EPOCH_PROOF_HEADER: &str = "x-rustfs-rpc-boot-epoch-proof";
+const RPC_REPLAY_SCOPE_VERSION_V3: &str = "3";
 const RPC_RESPONSE_PROOF_DOMAIN: &[u8] = b"rustfs-rpc-response-proof-v1\0";
+const RPC_REPLAY_SCOPE_DOMAIN: &[u8] = b"rustfs-rpc-replay-scope-v3\0";
+const RPC_BOOT_EPOCH_PROOF_DOMAIN: &[u8] = b"rustfs-rpc-boot-epoch-proof-v1\0";
 const UNSIGNED_PAYLOAD: &str = "UNSIGNED-PAYLOAD";
 const UNSIGNED_PAYLOAD_NONCE: &str = "unsigned";
 const SIGNATURE_VALID_DURATION: i64 = 300; // 5 minutes
@@ -75,9 +84,15 @@ static INTERNODE_RPC_BODY_DIGEST_STRICT: LazyLock<bool> = LazyLock::new(|| {
         rustfs_config::DEFAULT_INTERNODE_RPC_BODY_DIGEST_STRICT,
     )
 });
-// Sized for peak legitimate body-bound mutation RPS x the retention window; overflow fails closed
-// and increments the replay-cache overflow counter. Clamped to at least 1 so a misconfigured zero
-// cannot disable replay protection by rejecting every body-bound request.
+static INTERNODE_RPC_REPLAY_SCOPE_STRICT: LazyLock<bool> = LazyLock::new(|| {
+    get_env_bool(
+        rustfs_config::ENV_INTERNODE_RPC_REPLAY_SCOPE_STRICT,
+        rustfs_config::DEFAULT_INTERNODE_RPC_REPLAY_SCOPE_STRICT,
+    )
+});
+// Sized for peak legitimate authenticated RPC RPS x the retention window once replay scope is
+// active; overflow fails closed and increments the replay-cache overflow counter. Clamped to at
+// least 1 so a misconfigured zero cannot disable replay protection by rejecting every request.
 static REPLAY_CACHE_CAPACITY: LazyLock<usize> = LazyLock::new(|| {
     rustfs_utils::get_env_usize(
         rustfs_config::ENV_INTERNODE_RPC_REPLAY_CACHE_CAPACITY,
@@ -86,6 +101,7 @@ static REPLAY_CACHE_CAPACITY: LazyLock<usize> = LazyLock::new(|| {
     .max(1)
 });
 static RPC_SECRET_RESOLUTION_LOG_ONCE: Once = Once::new();
+static RPC_BOOT_EPOCH: LazyLock<Uuid> = LazyLock::new(Uuid::new_v4);
 
 #[derive(Default)]
 struct RpcNonceCache {
@@ -313,6 +329,189 @@ fn verify_signature_v2(secret: &str, scope: SignatureV2Scope<'_>, signature: &st
     mac.verify_slice(&signature).is_ok()
 }
 
+#[derive(Clone, Copy)]
+struct ReplayScope<'a> {
+    audience: &'a str,
+    path: &'a str,
+    timestamp: &'a str,
+    nonce: Uuid,
+    content_sha256: &'a str,
+    boot_epoch: Uuid,
+}
+
+fn update_replay_scope(mac: &mut HmacSha256, scope: ReplayScope<'_>) {
+    mac.update(RPC_REPLAY_SCOPE_DOMAIN);
+    for part in [
+        scope.audience.as_bytes(),
+        b"|",
+        scope.path.as_bytes(),
+        b"|POST|",
+        scope.timestamp.as_bytes(),
+        b"|",
+        scope.nonce.as_bytes(),
+        b"|",
+        scope.content_sha256.as_bytes(),
+        b"|",
+        scope.boot_epoch.as_bytes(),
+    ] {
+        mac.update(part);
+    }
+}
+
+fn generate_replay_scope_signature(secret: &str, scope: ReplayScope<'_>) -> std::io::Result<String> {
+    let mut mac =
+        <HmacSha256 as KeyInit>::new_from_slice(secret.as_bytes()).map_err(|_| std::io::Error::other("Invalid RPC HMAC key"))?;
+    update_replay_scope(&mut mac, scope);
+    Ok(general_purpose::STANDARD.encode(mac.finalize().into_bytes()))
+}
+
+fn verify_replay_scope_signature(secret: &str, scope: ReplayScope<'_>, signature: &str) -> bool {
+    let Ok(signature) = general_purpose::STANDARD.decode(signature) else {
+        return false;
+    };
+    let Ok(mut mac) = <HmacSha256 as KeyInit>::new_from_slice(secret.as_bytes()) else {
+        return false;
+    };
+    update_replay_scope(&mut mac, scope);
+    mac.verify_slice(&signature).is_ok()
+}
+
+fn update_boot_epoch_proof(mac: &mut HmacSha256, audience: &str, challenge: Uuid, boot_epoch: Uuid) {
+    mac.update(RPC_BOOT_EPOCH_PROOF_DOMAIN);
+    mac.update(audience.as_bytes());
+    mac.update(b"|");
+    mac.update(challenge.as_bytes());
+    mac.update(boot_epoch.as_bytes());
+}
+
+fn generate_boot_epoch_proof(secret: &str, audience: &str, challenge: Uuid, boot_epoch: Uuid) -> std::io::Result<String> {
+    if audience.is_empty() || challenge.is_nil() || boot_epoch.is_nil() {
+        return Err(std::io::Error::other("Invalid RPC boot epoch proof scope"));
+    }
+    let mut mac =
+        <HmacSha256 as KeyInit>::new_from_slice(secret.as_bytes()).map_err(|_| std::io::Error::other("Invalid RPC HMAC key"))?;
+    update_boot_epoch_proof(&mut mac, audience, challenge, boot_epoch);
+    Ok(general_purpose::STANDARD.encode(mac.finalize().into_bytes()))
+}
+
+fn verify_boot_epoch_proof(secret: &str, audience: &str, challenge: Uuid, boot_epoch: Uuid, proof: &str) -> std::io::Result<()> {
+    if audience.is_empty() || challenge.is_nil() || boot_epoch.is_nil() {
+        return Err(std::io::Error::other("Invalid RPC boot epoch proof scope"));
+    }
+    let proof = general_purpose::STANDARD
+        .decode(proof)
+        .map_err(|_| std::io::Error::other("Invalid RPC boot epoch proof"))?;
+    let mut mac =
+        <HmacSha256 as KeyInit>::new_from_slice(secret.as_bytes()).map_err(|_| std::io::Error::other("Invalid RPC HMAC key"))?;
+    update_boot_epoch_proof(&mut mac, audience, challenge, boot_epoch);
+    mac.verify_slice(&proof)
+        .map_err(|_| std::io::Error::new(std::io::ErrorKind::PermissionDenied, "Invalid RPC boot epoch proof"))
+}
+
+fn non_nil_uuid(value: &str, name: &str) -> std::io::Result<Uuid> {
+    let value = Uuid::parse_str(value).map_err(|_| std::io::Error::other(format!("Invalid {name}")))?;
+    (!value.is_nil())
+        .then_some(value)
+        .ok_or_else(|| std::io::Error::other(format!("Invalid {name}")))
+}
+
+fn parse_tonic_rpc_path(path: &str) -> std::io::Result<(&str, &str)> {
+    path.strip_prefix('/')
+        .and_then(|path| path.split_once('/'))
+        .filter(|(service, rpc_method)| !service.is_empty() && !rpc_method.is_empty() && !rpc_method.contains('/'))
+        .ok_or_else(|| std::io::Error::other("Invalid RPC request path"))
+}
+
+/// The process-unique epoch included in every replay-scoped server verification.
+///
+/// A fresh process gets a fresh value, so a signature captured before a server restart cannot be
+/// admitted even though the bounded in-memory nonce cache necessarily starts empty again.
+pub fn tonic_rpc_boot_epoch() -> Uuid {
+    *RPC_BOOT_EPOCH
+}
+
+/// Build the additive replay-scope headers for a request that already carries rolling-upgrade-safe
+/// v1/v2 metadata. `timestamp` and `content_sha256` are deliberately reused from the v2 scope so
+/// old servers can continue validating the same request unchanged.
+pub fn gen_tonic_replay_scope_headers(
+    audience: &str,
+    path: &str,
+    timestamp: &str,
+    content_sha256: &str,
+    boot_epoch: Uuid,
+) -> std::io::Result<HeaderMap> {
+    if audience.is_empty() || !path.starts_with('/') || !valid_content_sha256(content_sha256) || boot_epoch.is_nil() {
+        return Err(std::io::Error::other("Invalid replay-scoped RPC signing scope"));
+    }
+    parse_tonic_rpc_path(path)?;
+    timestamp
+        .parse::<i64>()
+        .map_err(|_| std::io::Error::other("Invalid timestamp format"))?;
+
+    let nonce = Uuid::new_v4();
+    let signature = generate_replay_scope_signature(
+        &get_shared_secret()?,
+        ReplayScope {
+            audience,
+            path,
+            timestamp,
+            nonce,
+            content_sha256,
+            boot_epoch,
+        },
+    )?;
+    let mut headers = HeaderMap::new();
+    headers.insert(RPC_REPLAY_SCOPE_VERSION_HEADER, HeaderValue::from_static(RPC_REPLAY_SCOPE_VERSION_V3));
+    headers.insert(
+        RPC_REPLAY_SCOPE_SIGNATURE_HEADER,
+        header_value(&signature, RPC_REPLAY_SCOPE_SIGNATURE_HEADER)?,
+    );
+    headers.insert(
+        RPC_REPLAY_SCOPE_NONCE_HEADER,
+        header_value(&nonce.to_string(), RPC_REPLAY_SCOPE_NONCE_HEADER)?,
+    );
+    headers.insert(RPC_BOOT_EPOCH_HEADER, header_value(&boot_epoch.to_string(), RPC_BOOT_EPOCH_HEADER)?);
+    Ok(headers)
+}
+
+/// Parse the optional client challenge used to authenticate a server boot-epoch advertisement.
+pub fn tonic_boot_epoch_challenge(headers: &HeaderMap) -> std::io::Result<Option<Uuid>> {
+    headers
+        .get(RPC_BOOT_EPOCH_CHALLENGE_HEADER)
+        .map(|value| {
+            value
+                .to_str()
+                .map_err(|_| std::io::Error::other("Invalid RPC boot epoch challenge"))
+                .and_then(|value| non_nil_uuid(value, "RPC boot epoch challenge"))
+        })
+        .transpose()
+}
+
+/// Build the authenticated response headers for a client boot-epoch challenge.
+pub fn tonic_boot_epoch_response_headers(audience: &str, challenge: Uuid) -> std::io::Result<HeaderMap> {
+    let boot_epoch = tonic_rpc_boot_epoch();
+    let proof = generate_boot_epoch_proof(&get_shared_secret()?, audience, challenge, boot_epoch)?;
+    let mut headers = HeaderMap::new();
+    headers.insert(RPC_BOOT_EPOCH_HEADER, header_value(&boot_epoch.to_string(), RPC_BOOT_EPOCH_HEADER)?);
+    headers.insert(RPC_BOOT_EPOCH_PROOF_HEADER, header_value(&proof, RPC_BOOT_EPOCH_PROOF_HEADER)?);
+    Ok(headers)
+}
+
+/// Verify the server boot-epoch response for a challenge generated by this client.
+pub fn verify_tonic_boot_epoch_response(audience: &str, challenge: Uuid, headers: &HeaderMap) -> std::io::Result<Uuid> {
+    let boot_epoch = headers
+        .get(RPC_BOOT_EPOCH_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC boot epoch"))
+        .and_then(|value| non_nil_uuid(value, "RPC boot epoch"))?;
+    let proof = headers
+        .get(RPC_BOOT_EPOCH_PROOF_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC boot epoch proof"))?;
+    verify_boot_epoch_proof(&get_shared_secret()?, audience, challenge, boot_epoch, proof)?;
+    Ok(boot_epoch)
+}
+
 fn valid_content_sha256(value: &str) -> bool {
     value == UNSIGNED_PAYLOAD
         || (value.len() == 64
@@ -443,6 +642,16 @@ pub fn set_tonic_canonical_body_digest<T>(request: &mut tonic::Request<T>, canon
     Ok(())
 }
 
+pub fn set_tonic_mutation_body_digest<T: rustfs_protos::CanonicalMutationBody>(
+    request: &mut tonic::Request<T>,
+) -> std::io::Result<()> {
+    let canonical_body = request
+        .get_ref()
+        .canonical_body()
+        .map_err(|_| std::io::Error::other("RPC mutation body length cannot be represented"))?;
+    set_tonic_canonical_body_digest(request, &canonical_body)
+}
+
 pub fn verify_tonic_canonical_body_digest<T>(request: &tonic::Request<T>, canonical_body: &[u8]) -> std::io::Result<()> {
     let version = request
         .metadata()
@@ -466,7 +675,7 @@ pub fn verify_tonic_canonical_body_digest<T>(request: &tonic::Request<T>, canoni
     Ok(())
 }
 
-/// Verify a mutating disk RPC's canonical body digest with a rolling-upgrade fallback.
+/// Verify a mutating RPC's canonical body digest with a rolling-upgrade fallback.
 ///
 /// When the request carries a real (non-`UNSIGNED-PAYLOAD`) content SHA-256 it is verified exactly
 /// like [`verify_tonic_canonical_body_digest`]. The digest value is a member of the signed v2
@@ -497,7 +706,7 @@ fn verify_tonic_mutation_body_digest_with_strictness<T>(
         Some(digest) if digest != UNSIGNED_PAYLOAD => verify_tonic_canonical_body_digest(request, canonical_body),
         _ => {
             // RUSTFS_COMPAT_TODO(disk-mutation-body-digest): accept digestless peers during rolling upgrades. Remove after the
-            // minimum supported RustFS peer version body-binds every mutating disk RPC.
+            // minimum supported RustFS peer version body-binds every mutating RPC.
             if strict {
                 return Err(std::io::Error::other("RPC mutation requires a body-bound v2 signature"));
             }
@@ -521,6 +730,17 @@ fn has_v2_auth_headers(headers: &HeaderMap) -> bool {
     .any(|name| headers.contains_key(*name))
 }
 
+fn has_replay_scope_headers(headers: &HeaderMap) -> bool {
+    [
+        RPC_REPLAY_SCOPE_VERSION_HEADER,
+        RPC_REPLAY_SCOPE_SIGNATURE_HEADER,
+        RPC_REPLAY_SCOPE_NONCE_HEADER,
+        RPC_BOOT_EPOCH_HEADER,
+    ]
+    .iter()
+    .any(|name| headers.contains_key(*name))
+}
+
 /// Whether the server requires target-bound v2 authentication on every internode gRPC request,
 /// rejecting the legacy constant-target fallback instead of accepting it. Default-off rollout
 /// lever gated on the v1-fallback counter reading zero fleet-wide; see
@@ -530,9 +750,127 @@ fn internode_rpc_signature_strict() -> bool {
     *INTERNODE_RPC_SIGNATURE_STRICT
 }
 
+fn internode_rpc_replay_scope_strict() -> bool {
+    *INTERNODE_RPC_REPLAY_SCOPE_STRICT
+}
+
+fn verify_tonic_replay_scope_signature(audience: &str, path: &str, headers: &HeaderMap) -> std::io::Result<()> {
+    if audience.is_empty() {
+        return Err(std::io::Error::other("Missing RPC audience"));
+    }
+    parse_tonic_rpc_path(path)?;
+
+    let version = headers
+        .get(RPC_REPLAY_SCOPE_VERSION_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC replay scope version"))?;
+    if version != RPC_REPLAY_SCOPE_VERSION_V3 {
+        return Err(std::io::Error::other("Unsupported RPC replay scope version"));
+    }
+    let signature = headers
+        .get(RPC_REPLAY_SCOPE_SIGNATURE_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC replay scope signature"))?;
+    let timestamp = headers
+        .get(TIMESTAMP_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing timestamp header"))?;
+    let signed_at = timestamp
+        .parse::<i64>()
+        .map_err(|_| std::io::Error::other("Invalid timestamp format"))?;
+    check_timestamp(signed_at)?;
+    let nonce = headers
+        .get(RPC_REPLAY_SCOPE_NONCE_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC replay scope nonce"))
+        .and_then(|value| non_nil_uuid(value, "RPC replay scope nonce"))?;
+    let content_sha256 = headers
+        .get(RPC_CONTENT_SHA256_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC content SHA-256"))?;
+    if !valid_content_sha256(content_sha256) {
+        return Err(std::io::Error::other("Invalid RPC content SHA-256"));
+    }
+    let boot_epoch = headers
+        .get(RPC_BOOT_EPOCH_HEADER)
+        .and_then(|value| value.to_str().ok())
+        .ok_or_else(|| std::io::Error::other("Missing RPC boot epoch"))
+        .and_then(|value| non_nil_uuid(value, "RPC boot epoch"))?;
+    let secret = get_shared_secret()?;
+    if !verify_replay_scope_signature(
+        &secret,
+        ReplayScope {
+            audience,
+            path,
+            timestamp,
+            nonce,
+            content_sha256,
+            boot_epoch,
+        },
+        signature,
+    ) {
+        return Err(std::io::Error::other("Invalid RPC replay scope signature"));
+    }
+    if boot_epoch != tonic_rpc_boot_epoch() {
+        return Err(std::io::Error::other("RPC boot epoch is stale"));
+    }
+    check_and_record_nonce(nonce, signed_at)
+}
+
 /// Verify gRPC authentication, preferring v2 without downgrade on malformed v2 metadata.
 pub fn verify_tonic_rpc_signature(audience: &str, path: &str, headers: &HeaderMap) -> std::io::Result<()> {
-    verify_tonic_rpc_signature_with_strictness(audience, path, headers, internode_rpc_signature_strict())
+    verify_tonic_rpc_signature_with_policy(
+        audience,
+        path,
+        headers,
+        internode_rpc_signature_strict(),
+        internode_rpc_replay_scope_strict(),
+        false,
+    )
+}
+
+/// Verify gRPC authentication while allowing the narrowly scoped v2 `Ping` bootstrap used to
+/// obtain an authenticated server boot epoch when replay-scope strictness is enabled.
+pub fn verify_tonic_rpc_signature_with_bootstrap(
+    audience: &str,
+    path: &str,
+    headers: &HeaderMap,
+    allow_replay_scope_bootstrap: bool,
+) -> std::io::Result<()> {
+    verify_tonic_rpc_signature_with_policy(
+        audience,
+        path,
+        headers,
+        internode_rpc_signature_strict(),
+        internode_rpc_replay_scope_strict(),
+        allow_replay_scope_bootstrap,
+    )
+}
+
+fn verify_tonic_rpc_signature_with_policy(
+    audience: &str,
+    path: &str,
+    headers: &HeaderMap,
+    signature_strict: bool,
+    replay_scope_strict: bool,
+    allow_replay_scope_bootstrap: bool,
+) -> std::io::Result<()> {
+    if has_replay_scope_headers(headers) {
+        return verify_tonic_replay_scope_signature(audience, path, headers);
+    }
+
+    // Only a method-bound v2 Ping with a syntactically valid challenge may bootstrap a strict
+    // client after its peer restarts. Legacy metadata never gets this exception.
+    let bootstrap = allow_replay_scope_bootstrap
+        && has_v2_auth_headers(headers)
+        && tonic_boot_epoch_challenge(headers).is_ok_and(|challenge| challenge.is_some());
+    if replay_scope_strict && !bootstrap {
+        return Err(std::io::Error::other("RPC replay-scoped authentication required"));
+    }
+
+    verify_tonic_rpc_signature_with_strictness(audience, path, headers, signature_strict)?;
+    global_internode_metrics().record_replay_scope_fallback();
+    Ok(())
 }
 
 /// [`verify_tonic_rpc_signature`] with the strict gate injected as a parameter, so both rollout
@@ -677,11 +1015,28 @@ mod tests {
     use crate::cluster::rpc::context_propagation::REQUEST_ID_HEADER;
     use crate::runtime::sources as runtime_sources;
     use http::{HeaderMap, Method};
+    use rustfs_protos::{
+        CanonicalMutationBody as _, PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS,
+        proto_gen::node_service::{Mss, SignalServiceRequest},
+    };
+    use std::collections::HashMap;
     use std::io::{self, Write};
     use std::sync::{Arc, Mutex};
     use time::OffsetDateTime;
     use tracing_subscriber::fmt::MakeWriter;
 
+    fn signal_service_request(signal: &str, sub_system: &str, dry_run: &str) -> SignalServiceRequest {
+        SignalServiceRequest {
+            vars: Some(Mss {
+                value: HashMap::from([
+                    (PEER_RESTSIGNAL.to_string(), signal.to_string()),
+                    (PEER_RESTSUB_SYS.to_string(), sub_system.to_string()),
+                    (PEER_RESTDRY_RUN.to_string(), dry_run.to_string()),
+                ]),
+            }),
+        }
+    }
+
     #[derive(Clone, Default)]
     struct CapturedLogs {
         buffer: Arc<Mutex<Vec<u8>>>,
@@ -1246,6 +1601,104 @@ mod tests {
         assert_eq!(error.to_string(), "Invalid RPC v2 signature");
     }
 
+    #[test]
+    fn replay_scope_binds_path_epoch_and_random_nonce() {
+        ensure_test_rpc_secret();
+        let path = "/node_service.NodeService/Ping";
+        let mut headers = gen_tonic_signature_headers("node-a:9000", "node_service.NodeService", "Ping", None)
+            .expect("v2 compatibility headers should build");
+        let timestamp = headers
+            .get(TIMESTAMP_HEADER)
+            .and_then(|value| value.to_str().ok())
+            .expect("v2 timestamp")
+            .to_string();
+        let content_sha256 = headers
+            .get(RPC_CONTENT_SHA256_HEADER)
+            .and_then(|value| value.to_str().ok())
+            .expect("v2 content digest")
+            .to_string();
+        headers.extend(
+            gen_tonic_replay_scope_headers("node-a:9000", path, &timestamp, &content_sha256, tonic_rpc_boot_epoch())
+                .expect("replay-scope headers should build"),
+        );
+
+        assert!(
+            verify_tonic_rpc_signature_with_policy("node-a:9000", path, &headers, false, false, false).is_ok(),
+            "the first replay-scoped request must be accepted"
+        );
+        let replay = verify_tonic_rpc_signature_with_policy("node-a:9000", path, &headers, false, false, false)
+            .expect_err("the random replay-scope nonce must be single-use");
+        assert_eq!(replay.to_string(), "RPC request replay detected");
+
+        let path_error = verify_tonic_replay_scope_signature("node-a:9000", "/node_service.NodeService/SignalService", &headers)
+            .expect_err("a replay-scoped signature must not move to another method");
+        assert_eq!(path_error.to_string(), "Invalid RPC replay scope signature");
+    }
+
+    #[test]
+    fn replay_scope_rejects_partial_metadata_and_stale_epoch_without_fallback() {
+        ensure_test_rpc_secret();
+        let path = "/node_service.NodeService/Ping";
+        let mut partial = gen_tonic_signature_headers("node-a:9000", "node_service.NodeService", "Ping", None)
+            .expect("v2 compatibility headers should build");
+        partial.insert(RPC_REPLAY_SCOPE_VERSION_HEADER, HeaderValue::from_static(RPC_REPLAY_SCOPE_VERSION_V3));
+        let error = verify_tonic_rpc_signature_with_policy("node-a:9000", path, &partial, false, false, false)
+            .expect_err("partial replay-scope metadata must never downgrade to v2");
+        assert_eq!(error.to_string(), "Missing RPC replay scope signature");
+
+        let timestamp = partial
+            .get(TIMESTAMP_HEADER)
+            .and_then(|value| value.to_str().ok())
+            .expect("v2 timestamp")
+            .to_string();
+        let content_sha256 = partial
+            .get(RPC_CONTENT_SHA256_HEADER)
+            .and_then(|value| value.to_str().ok())
+            .expect("v2 content digest")
+            .to_string();
+        let stale_epoch = Uuid::new_v4();
+        partial.extend(
+            gen_tonic_replay_scope_headers("node-a:9000", path, &timestamp, &content_sha256, stale_epoch)
+                .expect("replay-scope headers should build"),
+        );
+        let stale = verify_tonic_rpc_signature_with_policy("node-a:9000", path, &partial, false, false, false)
+            .expect_err("a signature from a prior server boot epoch must be rejected");
+        assert_eq!(stale.to_string(), "RPC boot epoch is stale");
+    }
+
+    #[test]
+    fn replay_scope_strictness_allows_only_authenticated_ping_bootstrap() {
+        ensure_test_rpc_secret();
+        let mut headers = gen_tonic_signature_headers("node-a:9000", "node_service.NodeService", "Ping", None)
+            .expect("v2 compatibility headers should build");
+        let rejected =
+            verify_tonic_rpc_signature_with_policy("node-a:9000", "/node_service.NodeService/Ping", &headers, false, true, false)
+                .expect_err("strict replay scope must reject stripped new metadata");
+        assert_eq!(rejected.to_string(), "RPC replay-scoped authentication required");
+
+        headers.insert(
+            RPC_BOOT_EPOCH_CHALLENGE_HEADER,
+            HeaderValue::from_str(&Uuid::new_v4().to_string()).expect("UUID header"),
+        );
+        assert!(
+            verify_tonic_rpc_signature_with_policy("node-a:9000", "/node_service.NodeService/Ping", &headers, false, true, true,)
+                .is_ok(),
+            "only the signed Ping bootstrap may obtain a new server epoch in strict mode"
+        );
+    }
+
+    #[test]
+    fn boot_epoch_response_proof_binds_audience_challenge_and_epoch() {
+        ensure_test_rpc_secret();
+        let challenge = Uuid::new_v4();
+        let headers = tonic_boot_epoch_response_headers("node-a:9000", challenge).expect("proof headers should build");
+        let epoch =
+            verify_tonic_boot_epoch_response("node-a:9000", challenge, &headers).expect("matching proof headers should verify");
+        assert_eq!(epoch, tonic_rpc_boot_epoch());
+        assert!(verify_tonic_boot_epoch_response("node-b:9000", challenge, &headers).is_err());
+        assert!(verify_tonic_boot_epoch_response("node-a:9000", Uuid::new_v4(), &headers).is_err());
+    }
+
     #[test]
     fn malformed_v2_auth_does_not_downgrade_to_valid_legacy_signature() {
         ensure_test_rpc_secret();
@@ -1596,6 +2049,72 @@ mod tests {
         assert_eq!(stripped.to_string(), "RPC content SHA-256 mismatch");
     }
 
+    #[test]
+    fn signal_service_mutation_contract_rejects_tampering_and_replay() {
+        ensure_test_rpc_secret();
+        let body = signal_service_request("2", "scanner", "false")
+            .canonical_body()
+            .expect("small signal request should encode");
+        let mut request = tonic::Request::new(());
+        set_tonic_canonical_body_digest(&mut request, &body).expect("canonical body digest should be attached");
+        let content_sha256 = request
+            .metadata()
+            .get(RPC_CONTENT_SHA256_HEADER)
+            .and_then(|value| value.to_str().ok());
+        let headers = gen_tonic_signature_headers("node-a:9000", "node_service.NodeService", "SignalService", content_sha256)
+            .expect("body-bound auth headers should build");
+        request.metadata_mut().as_mut().extend(headers.clone());
+
+        assert!(
+            verify_tonic_rpc_signature("node-a:9000", "/node_service.NodeService/SignalService", &headers).is_ok(),
+            "the first body-bound signal request must authenticate"
+        );
+        assert!(verify_tonic_mutation_body_digest(&request, &body).is_ok());
+
+        let tampered = signal_service_request("1", "scanner", "false")
+            .canonical_body()
+            .expect("small signal request should encode");
+        let error = verify_tonic_mutation_body_digest(&request, &tampered)
+            .expect_err("changing the signal must invalidate the signed digest");
+        assert_eq!(error.to_string(), "RPC content SHA-256 mismatch");
+
+        let replay = verify_tonic_rpc_signature("node-a:9000", "/node_service.NodeService/SignalService", &headers)
+            .expect_err("reusing the signal nonce must fail");
+        assert_eq!(replay.to_string(), "RPC request replay detected");
+    }
+
+    #[test]
+    #[serial_test::serial(rpc_body_digest_fallback_counter)]
+    fn signal_service_mutation_contract_preserves_rollout_fallback_and_strictness() {
+        let body = signal_service_request("2", "scanner", "false")
+            .canonical_body()
+            .expect("small signal request should encode");
+        let before = global_internode_metrics().snapshot().body_digest_fallback_total;
+        let digestless = tonic::Request::new(());
+
+        assert!(
+            verify_tonic_mutation_body_digest_with_strictness(&digestless, &body, false).is_ok(),
+            "old peers must remain compatible while the rollout gate is open"
+        );
+        assert_eq!(
+            global_internode_metrics().snapshot().body_digest_fallback_total,
+            before + 1,
+            "accepted digestless signal requests must be visible in the fallback metric"
+        );
+
+        let error = verify_tonic_mutation_body_digest_with_strictness(&digestless, &body, true)
+            .expect_err("strict mode must reject a digestless signal request");
+        assert_eq!(error.to_string(), "RPC mutation requires a body-bound v2 signature");
+
+        let mut bound = tonic::Request::new(());
+        set_tonic_canonical_body_digest(&mut bound, &body).expect("canonical body digest should be attached");
+        bound
+            .metadata_mut()
+            .as_mut()
+            .insert(RPC_AUTH_VERSION_HEADER, HeaderValue::from_static(RPC_AUTH_VERSION_V2));
+        assert!(verify_tonic_mutation_body_digest_with_strictness(&bound, &body, true).is_ok());
+    }
+
     #[test]
     fn nonce_cache_rejects_replay_after_wall_clock_regression() {
         let now = Instant::now();

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

@@ -26,13 +26,18 @@ pub(crate) mod runtime_sources;
 pub use background_monitor::shutdown_background_monitors;
 pub(crate) use background_monitor::spawn_background_monitor;
 pub use client::{
-    TonicInterceptor, gen_tonic_signature_interceptor, node_service_time_out_client, node_service_time_out_client_no_auth,
+    AuthenticatedChannel, TonicInterceptor, gen_tonic_signature_interceptor, node_service_time_out_client,
+    node_service_time_out_client_no_auth,
 };
+// Re-exported through `api::rpc`; not every item is consumed inside this crate.
+#[allow(unused_imports)]
 pub use http_auth::{
-    TONIC_RPC_PREFIX, build_auth_headers, gen_signature_headers, gen_tonic_signature_headers, normalize_tonic_rpc_audience,
-    set_tonic_canonical_body_digest, sign_ns_scanner_capability, sign_tonic_rpc_response_proof, verify_ns_scanner_capability,
-    verify_rpc_signature, verify_tonic_canonical_body_digest, verify_tonic_mutation_body_digest, verify_tonic_rpc_response_proof,
-    verify_tonic_rpc_signature,
+    TONIC_RPC_PREFIX, build_auth_headers, gen_signature_headers, gen_tonic_replay_scope_headers, gen_tonic_signature_headers,
+    normalize_tonic_rpc_audience, set_tonic_canonical_body_digest, set_tonic_mutation_body_digest, sign_ns_scanner_capability,
+    sign_tonic_rpc_response_proof, tonic_boot_epoch_challenge, tonic_boot_epoch_response_headers, verify_ns_scanner_capability,
+    verify_rpc_signature, verify_tonic_boot_epoch_response, verify_tonic_canonical_body_digest,
+    verify_tonic_mutation_body_digest, verify_tonic_rpc_response_proof, verify_tonic_rpc_signature,
+    verify_tonic_rpc_signature_with_bootstrap,
 };
 #[cfg(test)]
 pub(crate) use internode_data_transport::TcpHttpInternodeDataTransport;

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

@@ -13,10 +13,10 @@
 // limitations under the License.
 
 use crate::cluster::rpc::client::{
-    TonicInterceptor, gen_tonic_signature_interceptor, heal_control_time_out_client, node_service_time_out_client,
-    tier_mutation_control_time_out_client,
+    AuthenticatedChannel, TonicInterceptor, embedded_tonic_status, gen_tonic_signature_interceptor, heal_control_time_out_client,
+    is_network_like_status, message_has_network_needle, node_service_time_out_client, tier_mutation_control_time_out_client,
 };
-use crate::cluster::rpc::{set_tonic_canonical_body_digest, verify_tonic_rpc_response_proof};
+use crate::cluster::rpc::{set_tonic_canonical_body_digest, set_tonic_mutation_body_digest, verify_tonic_rpc_response_proof};
 use crate::error::{Error, Result};
 use crate::storage_api_contracts::internode::{
     SCANNER_ACTIVITY_LEGACY_PROTOCOL_VERSION, SCANNER_ACTIVITY_PREVIOUS_PROTOCOL_VERSION, SCANNER_ACTIVITY_PROTOCOL_VERSION,
@@ -50,6 +50,7 @@ use rustfs_protos::proto_gen::node_service::{
     TierMutationPeerState, TierMutationPrepareRequest, node_service_client::NodeServiceClient,
     tier_mutation_control_service_client::TierMutationControlServiceClient,
 };
+pub use rustfs_protos::{PEER_RESTDRY_RUN, PEER_RESTSIGNAL, PEER_RESTSUB_SYS};
 use rustfs_protos::{TierMutationRpcPhase, evict_failed_connection};
 use rustfs_utils::XHost;
 use serde::{Deserialize, Serialize as _};
@@ -65,13 +66,9 @@ use std::{
 use tokio::{net::TcpStream, time::Duration};
 use tonic::Request;
 use tonic::service::interceptor::InterceptedService;
-use tonic::transport::Channel;
 use tracing::{debug, info, warn};
 use uuid::Uuid;
 
-pub const PEER_RESTSIGNAL: &str = "signal";
-pub const PEER_RESTSUB_SYS: &str = "sub-sys";
-pub const PEER_RESTDRY_RUN: &str = "dry-run";
 pub const SERVICE_SIGNAL_REFRESH_CONFIG: u64 = 1;
 pub const SERVICE_SIGNAL_RELOAD_DYNAMIC: u64 = 2;
 const BACKGROUND_HEAL_STATUS_MAX_MESSAGE_SIZE: usize = 64 * 1024;
@@ -224,6 +221,21 @@ fn validate_heal_control_response_proof(canonical_response: &[u8], proof: &[u8])
         .map_err(|_| Error::other("peer returned an invalid heal control response proof"))
 }
 
+fn decode_remote_version_state_capability(expected_member: &str, result: &[u8]) -> Result<Uuid> {
+    let (topology_member, process_epoch) = rustfs_protos::decode_remote_version_state_capability(result).map_err(Error::other)?;
+    if topology_member != expected_member {
+        return Err(Error::other(
+            "peer returned a remote version state capability for a different topology member",
+        ));
+    }
+    let server_epoch =
+        Uuid::from_slice(process_epoch).map_err(|_| Error::other("peer returned an invalid remote version state epoch"))?;
+    if server_epoch.is_nil() {
+        return Err(Error::other("peer returned a nil remote version state epoch"));
+    }
+    Ok(server_epoch)
+}
+
 #[derive(Clone, Debug)]
 pub struct PeerLiveEventsBatch {
     pub events: Vec<u8>,
@@ -235,6 +247,7 @@ pub struct PeerLiveEventsBatch {
 pub struct PeerRestClient {
     pub host: XHost,
     pub grid_host: String,
+    topology_member: String,
     offline: Arc<AtomicBool>,
     recovery_running: Arc<AtomicBool>,
 }
@@ -327,9 +340,11 @@ impl PeerRestClient {
     }
 
     pub fn new(host: XHost, grid_host: String) -> Self {
+        let topology_member = host.to_string();
         Self {
             host,
             grid_host,
+            topology_member,
             offline: Arc::new(AtomicBool::new(false)),
             recovery_running: Arc::new(AtomicBool::new(false)),
         }
@@ -349,7 +364,11 @@ impl PeerRestClient {
 
             let client = match grid_host {
                 Some(grid_host) => match XHost::try_from(peer_host_port.clone()) {
-                    Ok(host) => Some(PeerRestClient::new(host, grid_host)),
+                    Ok(host) => {
+                        let mut client = PeerRestClient::new(host, grid_host);
+                        client.topology_member = peer_host_port.clone();
+                        Some(client)
+                    }
                     Err(err) => {
                         warn!(peer = %peer_host_port, "Xhost parse failed while constructing peer client: {err:?}");
                         None
@@ -392,7 +411,7 @@ impl PeerRestClient {
         (remote, all, remote_topology_hosts)
     }
 
-    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
         if self.offline.load(Ordering::Acquire) {
             self.mark_offline_and_spawn_recovery();
             return Err(Error::other(format!("peer {} is temporarily offline", self.grid_host)));
@@ -413,7 +432,7 @@ impl PeerRestClient {
         &self,
     ) -> Result<
         rustfs_protos::proto_gen::node_service::heal_control_service_client::HealControlServiceClient<
-            InterceptedService<Channel, TonicInterceptor>,
+            InterceptedService<AuthenticatedChannel, TonicInterceptor>,
         >,
     > {
         if self.offline.load(Ordering::Acquire) {
@@ -434,7 +453,7 @@ impl PeerRestClient {
 
     async fn get_tier_mutation_control_client(
         &self,
-    ) -> Result<TierMutationControlServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    ) -> Result<TierMutationControlServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
         if self.offline.load(Ordering::Acquire) {
             self.mark_offline_and_spawn_recovery();
             return Err(Error::other(format!("peer {} is temporarily offline", self.grid_host)));
@@ -471,26 +490,21 @@ impl PeerRestClient {
         self.offline.store(false, Ordering::Release);
     }
 
+    /// Whether this failure means the peer is unreachable, so it should be
+    /// gated offline and its connection evicted.
+    ///
+    /// RPC failures are classified by their typed gRPC code first
+    /// (`is_network_like_status`); an application error from a live peer must
+    /// never take it offline no matter what its message says. The substring
+    /// fallback only covers failures that exist purely as text, such as the
+    /// dial errors `get_client` wraps.
     fn is_network_like_error(err: &Error) -> bool {
-        let message = err.to_string().to_ascii_lowercase();
-        [
-            "temporarily offline",
-            "transport error",
-            "unavailable",
-            "error trying to connect",
-            "connection refused",
-            "connection reset",
-            "broken pipe",
-            "not connected",
-            "unexpected eof",
-            "timed out",
-            "deadline has elapsed",
-            "connection closed",
-            "connection aborted",
-            "tcp connect error",
-        ]
-        .iter()
-        .any(|needle| message.contains(needle))
+        if let Error::Io(io_err) = err
+            && let Some(status) = embedded_tonic_status(io_err)
+        {
+            return is_network_like_status(status);
+        }
+        message_has_network_needle(&err.to_string())
     }
 
     fn mark_offline_and_spawn_recovery(&self) {
@@ -1161,14 +1175,24 @@ impl PeerRestClient {
         validate_heal_control_capability_proof(&canonical_ack, &proof)
     }
 
+    pub async fn probe_remote_version_state(&self, topology_fingerprint: String) -> Result<(String, Uuid)> {
+        let probe = rustfs_protos::remote_version_state_capability_probe(Uuid::new_v4().as_bytes());
+        let result = self
+            .heal_control(rustfs_protos::HEAL_CONTROL_PROTOCOL_VERSION, topology_fingerprint, probe)
+            .await?;
+        let epoch = decode_remote_version_state_capability(&self.topology_member, &result)?;
+        Ok((self.topology_member.clone(), epoch))
+    }
+
     pub async fn load_bucket_metadata(&self, bucket: &str, scanner_maintenance_change: bool) -> Result<()> {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(LoadBucketMetadataRequest {
+                let mut request = Request::new(LoadBucketMetadataRequest {
                     bucket: bucket.to_string(),
                     scanner_maintenance_change,
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.load_bucket_metadata(request).await?.into_inner();
                 if !response.success {
@@ -1188,9 +1212,10 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(DeleteBucketMetadataRequest {
+                let mut request = Request::new(DeleteBucketMetadataRequest {
                     bucket: bucket.to_string(),
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.delete_bucket_metadata(request).await?.into_inner();
                 if !response.success {
@@ -1210,9 +1235,10 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(DeletePolicyRequest {
+                let mut request = Request::new(DeletePolicyRequest {
                     policy_name: policy.to_string(),
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.delete_policy(request).await?.into_inner();
                 if !response.success {
@@ -1232,9 +1258,10 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(LoadPolicyRequest {
+                let mut request = Request::new(LoadPolicyRequest {
                     policy_name: policy.to_string(),
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.load_policy(request).await?.into_inner();
                 if !response.success {
@@ -1254,11 +1281,12 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(LoadPolicyMappingRequest {
+                let mut request = Request::new(LoadPolicyMappingRequest {
                     user_or_group: user_or_group.to_string(),
                     user_type,
                     is_group,
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.load_policy_mapping(request).await?.into_inner();
                 if !response.success {
@@ -1278,9 +1306,10 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(DeleteUserRequest {
+                let mut request = Request::new(DeleteUserRequest {
                     access_key: access_key.to_string(),
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.delete_user(request).await?.into_inner();
                 if !response.success {
@@ -1300,9 +1329,10 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(DeleteServiceAccountRequest {
+                let mut request = Request::new(DeleteServiceAccountRequest {
                     access_key: access_key.to_string(),
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.delete_service_account(request).await?.into_inner();
                 if !response.success {
@@ -1322,10 +1352,11 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(LoadUserRequest {
+                let mut request = Request::new(LoadUserRequest {
                     access_key: access_key.to_string(),
                     temp,
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.load_user(request).await?.into_inner();
                 if !response.success {
@@ -1345,9 +1376,10 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(LoadServiceAccountRequest {
+                let mut request = Request::new(LoadServiceAccountRequest {
                     access_key: access_key.to_string(),
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.load_service_account(request).await?.into_inner();
                 if !response.success {
@@ -1367,9 +1399,10 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(LoadGroupRequest {
+                let mut request = Request::new(LoadGroupRequest {
                     group: group.to_string(),
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.load_group(request).await?.into_inner();
                 if !response.success {
@@ -1389,7 +1422,8 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(ReloadSiteReplicationConfigRequest {});
+                let mut request = Request::new(ReloadSiteReplicationConfigRequest {});
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.reload_site_replication_config(request).await?.into_inner();
                 if !response.success {
@@ -1413,9 +1447,10 @@ impl PeerRestClient {
                 vars.insert(PEER_RESTSIGNAL.to_string(), sig.to_string());
                 vars.insert(PEER_RESTSUB_SYS.to_string(), sub_sys.to_string());
                 vars.insert(PEER_RESTDRY_RUN.to_string(), dry_run.to_string());
-                let request = Request::new(SignalServiceRequest {
+                let mut request = Request::new(SignalServiceRequest {
                     vars: Some(Mss { value: vars }),
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.signal_service(request).await?.into_inner();
                 if !response.success {
@@ -1484,7 +1519,8 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(ReloadPoolMetaRequest {});
+                let mut request = Request::new(ReloadPoolMetaRequest {});
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.reload_pool_meta(request).await?.into_inner();
                 if !response.success {
@@ -1505,9 +1541,10 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(StopRebalanceRequest {
+                let mut request = Request::new(StopRebalanceRequest {
                     expected_rebalance_id: expected_rebalance_id.unwrap_or_default().to_string(),
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.stop_rebalance(request).await?.into_inner();
                 if !response.success {
@@ -1528,7 +1565,8 @@ impl PeerRestClient {
         self.finalize_result(
             async {
                 let mut client = self.get_client().await?;
-                let request = Request::new(LoadRebalanceMetaRequest { start_rebalance });
+                let mut request = Request::new(LoadRebalanceMetaRequest { start_rebalance });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.load_rebalance_meta(request).await?.into_inner();
 
@@ -1567,7 +1605,8 @@ impl PeerRestClient {
                     })
                     .collect::<Result<Vec<_>>>()?;
                 let mut client = self.get_client().await?;
-                let request = Request::new(StartDecommissionRequest { pool_indices });
+                let mut request = Request::new(StartDecommissionRequest { pool_indices });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.start_decommission(request).await?.into_inner();
                 if !response.success {
@@ -1590,7 +1629,8 @@ impl PeerRestClient {
                 let pool_index = u32::try_from(pool_index)
                     .map_err(|_| Error::other(format!("decommission pool index {pool_index} exceeds RPC range")))?;
                 let mut client = self.get_client().await?;
-                let request = Request::new(CancelDecommissionRequest { pool_index });
+                let mut request = Request::new(CancelDecommissionRequest { pool_index });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.cancel_decommission(request).await?.into_inner();
                 if !response.success {
@@ -1613,7 +1653,8 @@ impl PeerRestClient {
                 let pool_index = u32::try_from(pool_index)
                     .map_err(|_| Error::other(format!("decommission pool index {pool_index} exceeds RPC range")))?;
                 let mut client = self.get_client().await?;
-                let request = Request::new(ClearDecommissionRequest { pool_index });
+                let mut request = Request::new(ClearDecommissionRequest { pool_index });
+                set_tonic_mutation_body_digest(&mut request)?;
 
                 let response = client.clear_decommission(request).await?.into_inner();
                 if !response.success {
@@ -1633,10 +1674,14 @@ impl PeerRestClient {
     pub async fn load_transition_tier_config(&self) -> Result<()> {
         match self.load_transition_tier_config_outcome().await {
             TierConfigReloadOutcome::Success => Ok(()),
-            TierConfigReloadOutcome::TransientReconnect(err) | TierConfigReloadOutcome::TransientRetrySameChannel(err) => {
-                self.finalize_result(Err(err)).await
-            }
-            TierConfigReloadOutcome::Terminal(err) => Err(err),
+            // Only a reconnect-class failure says anything about the channel.
+            // `finalize_result` marks the peer offline and evicts its connection
+            // whenever the message looks network-like, and a peer that answered
+            // and rejected the apply can easily report one ("release RPC failed:
+            // transport error"). Routing those through here would gate a healthy,
+            // responding peer out of every unrelated RPC.
+            TierConfigReloadOutcome::TransientReconnect(err) => self.finalize_result(Err(err)).await,
+            TierConfigReloadOutcome::TransientRetrySameChannel(err) | TierConfigReloadOutcome::Terminal(err) => Err(err),
         }
     }
 
@@ -1662,6 +1707,9 @@ impl PeerRestClient {
             Err(err) => return tier_config_reload_connection_outcome(err),
         };
         let mut request = Request::new(LoadTransitionTierConfigRequest {});
+        if let Err(err) = set_tonic_mutation_body_digest(&mut request) {
+            return TierConfigReloadOutcome::Terminal(Error::other(err));
+        }
         request.set_timeout(rustfs_protos::heal_control_execution_timeout());
 
         let response = match client.load_transition_tier_config(request).await {
@@ -1702,39 +1750,37 @@ fn tier_config_reload_connection_outcome(err: Error) -> TierConfigReloadOutcome
 }
 
 fn is_tier_config_reload_connection_failure(err: &Error) -> bool {
-    let message = err.to_string().to_ascii_lowercase();
+    let message = err.to_string();
+    // A bare "unavailable" is only trusted inside the local dial-failure
+    // wrapper from `get_client`, never in application text.
     if message
+        .to_ascii_lowercase()
         .split_once("can not get client, err:")
         .is_some_and(|(_, local_error)| local_error.contains("unavailable"))
     {
         return true;
     }
-    [
-        "temporarily offline",
-        "transport error",
-        "error trying to connect",
-        "connection refused",
-        "connection reset",
-        "connection closed",
-        "connection aborted",
-        "broken pipe",
-        "not connected",
-        "unexpected eof",
-        "timed out",
-        "deadline has elapsed",
-        "tcp connect error",
-    ]
-    .iter()
-    .any(|needle| message.contains(needle))
+    message_has_network_needle(&message)
 }
 
+/// Classifies a reload the peer answered but refused to apply.
+///
+/// The peer replied, so the channel is healthy and only the remote apply
+/// failed. Those failures are transient by nature: the reload reads the tier
+/// mutation intents and takes the distributed tier-config lock, both of which
+/// fail while any other node is restarting or while the lock quorum is briefly
+/// disturbed. Retiring the worker on the first such rejection leaves that peer
+/// pinned to the old configuration with nothing left to heal it, so it answers
+/// `TierNotFound` for a tier the rest of the cluster already committed until a
+/// second admin mutation happens to spawn a fresh worker.
+///
+/// Convergence is the whole point of this path, so a rejection is retried on
+/// the same channel. The worker's exponential backoff caps the cost at one
+/// reload every `TIER_CONFIG_RELOAD_RETRY_CAP`, and `Terminal` stays reachable
+/// for transport and gRPC status failures, which is where a genuinely
+/// unrecoverable peer surfaces.
 fn tier_config_reload_remote_failure(error_info: Option<String>) -> TierConfigReloadOutcome {
-    let error_info = error_info.unwrap_or_default();
-    if matches!(error_info.as_str(), "errServerNotInitialized" | "ServerNotInitialized") {
-        TierConfigReloadOutcome::TransientRetrySameChannel(Error::other(error_info))
-    } else {
-        TierConfigReloadOutcome::Terminal(Error::other(error_info))
-    }
+    TierConfigReloadOutcome::TransientRetrySameChannel(Error::other(error_info.unwrap_or_default()))
 }
 
 fn tier_config_reload_status_outcome(status: tonic::Status) -> TierConfigReloadOutcome {
@@ -1744,6 +1790,14 @@ fn tier_config_reload_status_outcome(status: tonic::Status) -> TierConfigReloadO
         TierConfigReloadOutcome::TransientReconnect(status.into())
     } else if status.code() == Code::Unknown && status.message().starts_with("Service was not ready:") {
         TierConfigReloadOutcome::TransientRetrySameChannel(status.into())
+    } else if status.code() == Code::Unknown
+        && is_tier_config_reload_connection_failure(&Error::other(status.message().to_string()))
+    {
+        // tonic reports a connection dropped mid-call as `Unknown` carrying the
+        // transport error text rather than as `Unavailable`, which is what a peer
+        // restarting under an active mutation produces. Reconnect and retry, so
+        // the restart does not permanently retire this peer's reload worker.
+        TierConfigReloadOutcome::TransientReconnect(status.into())
     } else {
         TierConfigReloadOutcome::Terminal(status.into())
     }
@@ -2151,6 +2205,126 @@ mod tests {
         assert!(!PeerRestClient::is_network_like_error(&Error::NotImplemented));
     }
 
+    #[test]
+    fn peer_rest_client_network_classifier_uses_typed_status_code() {
+        // The one code that means "nothing is answering on this channel".
+        assert!(PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unavailable(
+            "storage layer is not initialized"
+        ))));
+        // Application statuses from a live peer must not mark it offline,
+        // even when their message contains transport-sounding words.
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::internal(
+            "failed to reload metadata for bucket \"unavailable-logs\""
+        ))));
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unauthenticated(
+            "No valid auth token"
+        ))));
+        // A request-budget expiry answered by a live peer is an application
+        // outcome, not a transport failure.
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::deadline_exceeded(
+            "heal control request expired"
+        ))));
+        // Unknown is the transport's escape hatch for a cause it could not
+        // map, and our handlers never return it, so there the text decides.
+        assert!(PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unknown(
+            "Service was not ready: transport error"
+        ))));
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(tonic::Status::unknown(
+            "peer response unknown"
+        ))));
+    }
+
+    #[test]
+    fn peer_rest_client_network_classifier_ignores_transport_words_in_application_statuses() {
+        // The reason classification reads the code rather than the text: a
+        // peer that answers is reachable, even when what it says describes a
+        // connection failure of its own. A handler interpolating a local
+        // io::Error into Status::internal, or relaying trouble with its own
+        // downstream, must not cost us the channel to a healthy peer.
+        for status in [
+            tonic::Status::internal("connection refused while dialing downstream backend"),
+            tonic::Status::internal("write failed: broken pipe"),
+            tonic::Status::unauthenticated("connection reset while validating token"),
+            tonic::Status::failed_precondition("scanner lease timed out"),
+            tonic::Status::deadline_exceeded("heal control request timed out"),
+        ] {
+            let rendered = status.to_string();
+            assert!(
+                !PeerRestClient::is_network_like_error(&Error::from(status)),
+                "an answered application status must not mark the peer offline: {rendered}"
+            );
+        }
+    }
+
+    #[test]
+    fn peer_rest_client_network_classifier_keeps_slow_peers_online() {
+        // The per-RPC channel deadline (RUSTFS_INTERNODE_RPC_TIMEOUT, 30s)
+        // surfaces as Cancelled "Timeout expired" carrying the transport
+        // cause as its source. A peer that is merely slow must stay online:
+        // gating it would spend a full recovery cycle fast-failing every RPC
+        // to a host that is still answering, turning load into a partition.
+        let timeout_status = tonic::Status::cancelled("Timeout expired");
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(timeout_status)));
+
+        let sourced = tonic::Status::from_error(Box::new(std::io::Error::other("Timeout expired")));
+        assert!(
+            std::error::Error::source(&sourced).is_some(),
+            "the transport builds this status through Status::from_error, which attaches the cause"
+        );
+        assert!(!PeerRestClient::is_network_like_error(&Error::from(sourced)));
+    }
+
+    #[test]
+    fn rpc_status_errors_keep_their_rendering_and_hide_peer_metadata() {
+        let err = Error::from(tonic::Status::unavailable("peer gone"));
+        assert_eq!(
+            err.to_string(),
+            "Io error: code: 'The service is currently unavailable', message: \"peer gone\""
+        );
+
+        // tonic's own Debug prints the MetadataMap, i.e. every response header
+        // the peer sent; those must not reach a log through this error.
+        let mut status = tonic::Status::unavailable("peer gone");
+        status
+            .metadata_mut()
+            .insert("authorization", "Bearer secret".parse().expect("valid header value"));
+        let rendered = format!("{:?}", Error::from(status));
+        assert!(!rendered.contains("Bearer secret"), "{rendered}");
+        assert!(!rendered.contains("MetadataMap"), "{rendered}");
+    }
+
+    #[test]
+    fn peer_rest_client_network_classifier_ignores_application_text_containing_unavailable() {
+        // Regression: a bare "unavailable" needle used to match application
+        // strings like these and take a healthy peer offline.
+        assert!(!PeerRestClient::is_network_like_error(&Error::other(
+            "peer replication statistics provider is unavailable"
+        )));
+        assert!(!PeerRestClient::is_network_like_error(&Error::other(
+            "bucket \"unavailable-logs\" not found"
+        )));
+        // Anchored renderings of a flattened Unavailable status still match:
+        // tonic >= 0.14 form ...
+        assert!(PeerRestClient::is_network_like_error(&Error::other(
+            "peer tier mutation commit RPC failed: code: 'The service is currently unavailable', message: \"peer gone\""
+        )));
+        // ... which is only anchored as long as tonic renders Unavailable this
+        // way. A tonic bump that reworded it leaves the typed path correct but
+        // this needle stale, so pin the coupling rather than discover it in a
+        // partition.
+        assert!(
+            tonic::Status::unavailable("peer gone")
+                .to_string()
+                .to_ascii_lowercase()
+                .contains("code: 'the service is currently unavailable'"),
+            "tonic reworded Code::Unavailable; update the anchored needle"
+        );
+        // ... and the tonic <= 0.13 form peers may relay in error_info.
+        assert!(PeerRestClient::is_network_like_error(&Error::other(
+            "peer tier mutation commit RPC failed: status: Unavailable, message: \"peer gone\""
+        )));
+    }
+
     #[test]
     fn tier_config_reload_outcome_keeps_tonic_and_remote_errors_typed() {
         assert!(matches!(
@@ -2177,9 +2351,12 @@ mod tests {
             tier_config_reload_status_outcome(tonic::Status::cancelled("request cancelled")),
             TierConfigReloadOutcome::Terminal(_)
         ));
+        // A peer that answered and then refused the apply is retried rather than
+        // retired: the channel is healthy, so the rejection reflects remote state
+        // that the next attempt can find healed.
         assert!(matches!(
             tier_config_reload_remote_failure(Some("backend unavailable".to_string())),
-            TierConfigReloadOutcome::Terminal(_)
+            TierConfigReloadOutcome::TransientRetrySameChannel(_)
         ));
         assert!(matches!(
             tier_config_reload_remote_failure(Some("errServerNotInitialized".to_string())),
@@ -2193,6 +2370,58 @@ mod tests {
             tier_config_reload_connection_outcome(Error::other("can not get client, err: connection unavailable")),
             TierConfigReloadOutcome::TransientReconnect(_)
         ));
+        // The bare word is trusted only to the right of the dial-failure
+        // prefix, not anywhere in the message.
+        assert!(matches!(
+            tier_config_reload_connection_outcome(Error::other(
+                "bucket unavailable-logs rejected it, then: can not get client, err: some other reason"
+            )),
+            TierConfigReloadOutcome::Terminal(_)
+        ));
+    }
+
+    /// A tier mutation issued while another node restarts must still converge on
+    /// the nodes that stayed up. Those peers answer the reload RPC and reject the
+    /// apply, because reloading reads the tier mutation intents and takes the
+    /// distributed tier-config lock while the lock quorum is still disturbed.
+    /// Classifying those rejections as terminal retired the reload worker on its
+    /// first attempt and pinned the peer to the previous configuration, so it
+    /// served `TierNotFound` for an already-committed tier until an unrelated
+    /// second admin mutation spawned a new worker.
+    #[test]
+    fn tier_config_reload_retries_peers_that_reject_the_apply_mid_restart() {
+        for error_info in [
+            "Lock acquisition timeout for resource '.rustfs.sys/config/tier-config.bin.lock' after 5s",
+            "Resource '.rustfs.sys/config/tier-config.bin.lock' is already locked by node-3",
+            "Internal error: release RPC failed: transport error",
+            "save_config_with_opts: err: PreconditionFailed",
+            "erasure read quorum",
+        ] {
+            assert!(
+                matches!(
+                    tier_config_reload_remote_failure(Some(error_info.to_string())),
+                    TierConfigReloadOutcome::TransientRetrySameChannel(_)
+                ),
+                "a peer that rejected the apply must stay retryable so it converges: {error_info}"
+            );
+        }
+
+        // An absent error message is still a rejection, not a reason to stop.
+        assert!(matches!(
+            tier_config_reload_remote_failure(None),
+            TierConfigReloadOutcome::TransientRetrySameChannel(_)
+        ));
+
+        // tonic surfaces a connection dropped mid-call as `Unknown`, not `Unavailable`.
+        assert!(matches!(
+            tier_config_reload_status_outcome(tonic::Status::unknown("transport error")),
+            TierConfigReloadOutcome::TransientReconnect(_)
+        ));
+        // An `Unknown` that is not transport-shaped stays terminal.
+        assert!(matches!(
+            tier_config_reload_status_outcome(tonic::Status::unknown("peer response unknown")),
+            TierConfigReloadOutcome::Terminal(_)
+        ));
     }
 
     #[tokio::test]
@@ -2271,6 +2500,22 @@ mod tests {
         }
     }
 
+    #[test]
+    fn remote_version_state_capability_decoder_fails_closed() {
+        let epoch = Uuid::new_v4();
+        let result = rustfs_protos::encode_remote_version_state_capability("node-a:9000", epoch.as_bytes())
+            .expect("small capability response should encode");
+        assert_eq!(
+            decode_remote_version_state_capability("node-a:9000", &result).expect("valid epoch should decode"),
+            epoch
+        );
+        assert!(decode_remote_version_state_capability("node-b:9000", &result).is_err());
+        assert!(decode_remote_version_state_capability("node-a:9000", &result[..result.len() - 1]).is_err());
+        let nil = rustfs_protos::encode_remote_version_state_capability("node-a:9000", Uuid::nil().as_bytes())
+            .expect("small capability response should encode");
+        assert!(decode_remote_version_state_capability("node-a:9000", &nil).is_err());
+    }
+
     struct TierMutationResponseFixture<'a> {
         version: u32,
         phase: TierMutationRpcPhase,
@@ -2495,6 +2740,39 @@ mod tests {
         assert!(!client.offline.load(Ordering::Acquire));
     }
 
+    #[tokio::test]
+    async fn peer_rest_client_finalize_result_keeps_online_for_app_errors_mentioning_unavailable() {
+        // Regression: application error text containing "unavailable" (a
+        // remote error_info payload, or a bucket named "unavailable-logs" in
+        // a typed application status) must not take a healthy peer offline.
+        let client = test_peer_client();
+        let err = client
+            .finalize_result::<()>(Err(Error::other("peer replication statistics provider is unavailable")))
+            .await
+            .expect_err("application error should still be returned");
+        assert!(err.to_string().contains("provider is unavailable"));
+        assert!(!client.offline.load(Ordering::Acquire));
+
+        let err = client
+            .finalize_result::<()>(Err(Error::from(tonic::Status::internal(
+                "failed to reload metadata for bucket \"unavailable-logs\"",
+            ))))
+            .await
+            .expect_err("application status should still be returned");
+        assert!(err.to_string().contains("unavailable-logs"));
+        assert!(!client.offline.load(Ordering::Acquire));
+    }
+
+    #[tokio::test]
+    async fn peer_rest_client_finalize_result_marks_offline_for_typed_unavailable_status() {
+        let client = test_peer_client();
+        client
+            .finalize_result::<()>(Err(Error::from(tonic::Status::unavailable("storage layer is not initialized"))))
+            .await
+            .expect_err("network error should still be returned");
+        assert!(client.offline.load(Ordering::Acquire));
+    }
+
     #[tokio::test(flavor = "current_thread")]
     async fn peer_rest_recovery_probe_logs_keep_request_id_span_context() {
         let logs = CapturedLogs::default();
@@ -2509,6 +2787,11 @@ mod tests {
                 .with_span_list(true),
         );
         let _guard = tracing::subscriber::set_default(subscriber);
+        // The `recovery-monitor` callsite is shared with the production
+        // `mark_offline_and_spawn_recovery` path that sibling tests exercise from
+        // subscriber-less threads; without this the span can be cached as
+        // `Interest::never()` and silently degrade to `Span::none()`.
+        let _callsite_pin = crate::test_tracing::pin_callsite_interest_for_test();
 
         let client = test_peer_client();
         let span = tracing::info_span!("request-span", request_id = "req-peer-rest");

crates/ecstore/src/cluster/rpc/peer_s3_client.rs

@@ -14,8 +14,10 @@
 
 use crate::bucket::metadata_sys;
 use crate::cluster::rpc::client::{
-    TonicInterceptor, gen_tonic_signature_interceptor, is_network_like_disk_error, node_service_time_out_client,
+    AuthenticatedChannel, TonicInterceptor, gen_tonic_signature_interceptor, is_network_like_disk_error,
+    node_service_time_out_client,
 };
+use crate::cluster::rpc::set_tonic_mutation_body_digest;
 use crate::disk::error::DiskError;
 use crate::disk::error::{Error, Result};
 use crate::disk::error_reduce::{BUCKET_OP_IGNORED_ERRS, is_all_buckets_not_found, reduce_write_quorum_errs};
@@ -39,16 +41,84 @@ use rustfs_protos::proto_gen::node_service::node_service_client::NodeServiceClie
 use rustfs_protos::proto_gen::node_service::{
     DeleteBucketRequest, GetBucketInfoRequest, HealBucketRequest, ListBucketRequest, MakeBucketRequest,
 };
+#[cfg(test)]
+use std::sync::{
+    Mutex as StdMutex,
+    atomic::{AtomicBool, Ordering},
+};
 use std::{collections::HashMap, fmt::Debug, sync::Arc, time::Duration};
+#[cfg(test)]
+use tokio::sync::Notify;
 use tokio::{net::TcpStream, sync::RwLock, time};
 use tokio_util::sync::CancellationToken;
 use tonic::Request;
 use tonic::service::interceptor::InterceptedService;
-use tonic::transport::Channel;
 use tracing::{debug, info, warn};
 
 type Client = Arc<Box<dyn PeerS3Client>>;
 
+#[cfg(test)]
+#[derive(Default)]
+pub(crate) struct DeleteBucketEmptyScanBarrier {
+    arrived: AtomicBool,
+    arrived_notify: Notify,
+    released: AtomicBool,
+    release_notify: Notify,
+}
+
+#[cfg(test)]
+impl DeleteBucketEmptyScanBarrier {
+    pub(crate) async fn wait_until_paused(&self) {
+        loop {
+            let notified = self.arrived_notify.notified();
+            if self.arrived.load(Ordering::Acquire) {
+                return;
+            }
+            notified.await;
+        }
+    }
+
+    pub(crate) fn release(&self) {
+        self.released.store(true, Ordering::Release);
+        self.release_notify.notify_waiters();
+    }
+
+    async fn pause(&self) {
+        self.arrived.store(true, Ordering::Release);
+        self.arrived_notify.notify_waiters();
+        loop {
+            let notified = self.release_notify.notified();
+            if self.released.load(Ordering::Acquire) {
+                return;
+            }
+            notified.await;
+        }
+    }
+}
+
+#[cfg(test)]
+static DELETE_BUCKET_EMPTY_SCAN_BARRIER: StdMutex<Option<Arc<DeleteBucketEmptyScanBarrier>>> = StdMutex::new(None);
+
+#[cfg(test)]
+pub(crate) fn install_delete_bucket_empty_scan_barrier() -> Arc<DeleteBucketEmptyScanBarrier> {
+    let barrier = Arc::new(DeleteBucketEmptyScanBarrier::default());
+    *DELETE_BUCKET_EMPTY_SCAN_BARRIER
+        .lock()
+        .expect("empty scan barrier lock should not be poisoned") = Some(barrier.clone());
+    barrier
+}
+
+#[cfg(test)]
+async fn pause_after_delete_bucket_empty_scan() {
+    let barrier = DELETE_BUCKET_EMPTY_SCAN_BARRIER
+        .lock()
+        .expect("empty scan barrier lock should not be poisoned")
+        .take();
+    if let Some(barrier) = barrier {
+        barrier.pause().await;
+    }
+}
+
 #[derive(Clone, Debug)]
 pub struct ScannerBucketListing {
     pub buckets: Vec<BucketInfo>,
@@ -89,6 +159,22 @@ fn reduce_pool_write_quorum_errs(per_pool_errs: &[Option<Error>]) -> Option<Erro
     reduce_write_quorum_errs(per_pool_errs, BUCKET_OP_IGNORED_ERRS, pool_write_quorum(per_pool_errs.len()))
 }
 
+fn resolve_heal_bucket_mode(opts: &mut HealOpts, pool_errs: &[Option<Error>]) -> Result<()> {
+    if opts.recreate {
+        return Ok(());
+    }
+    if let Some(err) = pool_errs
+        .iter()
+        .flatten()
+        .find(|err| **err != Error::DiskNotFound && **err != Error::VolumeNotFound)
+    {
+        return Err(err.clone());
+    }
+    opts.remove = is_all_buckets_not_found(pool_errs);
+    opts.recreate = !opts.remove;
+    Ok(())
+}
+
 #[async_trait]
 pub trait PeerS3Client: Debug + Sync + Send + 'static {
     async fn heal_bucket(&self, bucket: &str, opts: &HealOpts) -> Result<HealResultItem>;
@@ -159,10 +245,7 @@ impl S3PeerSys {
             pool_errs.push(reduce_pool_write_quorum_errs(&per_pool_errs));
         }
 
-        if !opts.recreate {
-            opts.remove = is_all_buckets_not_found(&pool_errs);
-            opts.recreate = !opts.remove;
-        }
+        resolve_heal_bucket_mode(&mut opts, &pool_errs)?;
 
         let mut futures = Vec::new();
         let heal_bucket_results = Arc::new(RwLock::new(vec![HealResultItem::default(); self.clients.len()]));
@@ -636,24 +719,22 @@ impl PeerS3Client for LocalPeerS3Client {
             return Err(Error::ErasureWriteQuorum);
         }
 
-        let force = if opts.force_if_empty && !opts.force {
+        if opts.force_if_empty && !opts.force {
             for disk in local_disks.iter() {
                 if has_xlmeta_files(&disk.path().join(bucket)).await.map_err(Error::Io)? {
                     return Err(Error::VolumeNotEmpty);
                 }
             }
-            true
-        } else {
-            opts.force
-        };
+            #[cfg(test)]
+            pause_after_delete_bucket_empty_scan().await;
+        }
 
         let mut futures = Vec::with_capacity(local_disks.len());
 
         for disk in local_disks.iter() {
-            // Non-force delete refuses a non-empty bucket (VolumeNotEmpty), which
-            // the recreate loop below turns into BucketNotEmpty; only an explicit
-            // force delete removes recursively (backlog#799 B1).
-            futures.push(disk.delete_volume(bucket, force));
+            // `force_if_empty` is validation-only. Passing it as force would let
+            // a PutObject committed after the scan be removed recursively.
+            futures.push(disk.delete_volume(bucket, opts.force));
         }
 
         let results = join_all(futures).await;
@@ -716,6 +797,15 @@ pub struct RemotePeerS3Client {
 }
 
 impl RemotePeerS3Client {
+    fn encode_delete_bucket_options(opts: &DeleteBucketOptions) -> Result<String> {
+        let mut remote_opts = opts.clone();
+        // Older peers promote `force_if_empty` to recursive force after their
+        // metadata scan. Keep this coordinator-only hint off the wire so a
+        // mixed-version delete fails closed on non-empty directory remnants.
+        remote_opts.force_if_empty = false;
+        serde_json::to_string(&remote_opts).map_err(Into::into)
+    }
+
     fn recovery_monitor_span(addr: &str) -> tracing::Span {
         tracing::info_span!(
             "recovery-monitor",
@@ -742,7 +832,7 @@ impl RemotePeerS3Client {
         client
     }
 
-    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
         node_service_time_out_client(&self.addr, TonicInterceptor::Signature(gen_tonic_signature_interceptor()))
             .await
             .map_err(|err| Error::other(format!("can not get client, err: {err}")))
@@ -917,10 +1007,11 @@ impl PeerS3Client for RemotePeerS3Client {
             || async {
                 let options: String = serde_json::to_string(opts)?;
                 let mut client = self.get_client().await?;
-                let request = Request::new(HealBucketRequest {
+                let mut request = Request::new(HealBucketRequest {
                     bucket: bucket.to_string(),
                     options,
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
                 let response = client.heal_bucket(request).await?.into_inner();
                 if !response.success {
                     return if let Some(err) = response.error {
@@ -973,10 +1064,11 @@ impl PeerS3Client for RemotePeerS3Client {
             || async {
                 let options = serde_json::to_string(opts)?;
                 let mut client = self.get_client().await?;
-                let request = Request::new(MakeBucketRequest {
+                let mut request = Request::new(MakeBucketRequest {
                     name: bucket.to_string(),
                     options,
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
                 let response = client.make_bucket(request).await?.into_inner();
 
                 if !response.success {
@@ -1024,13 +1116,14 @@ impl PeerS3Client for RemotePeerS3Client {
     async fn delete_bucket(&self, bucket: &str, opts: &DeleteBucketOptions) -> Result<()> {
         self.execute_with_timeout(
             || async {
-                let options = serde_json::to_string(opts)?;
+                let options = Self::encode_delete_bucket_options(opts)?;
                 let mut client = self.get_client().await?;
 
-                let request = Request::new(DeleteBucketRequest {
+                let mut request = Request::new(DeleteBucketRequest {
                     bucket: bucket.to_string(),
                     options,
                 });
+                set_tonic_mutation_body_digest(&mut request)?;
                 let response = client.delete_bucket(request).await?.into_inner();
                 if !response.success {
                     return if let Some(err) = response.error {
@@ -1397,6 +1490,49 @@ mod tests {
         }
     }
 
+    #[test]
+    fn remote_delete_bucket_options_fail_closed_for_legacy_peers() {
+        let encoded = RemotePeerS3Client::encode_delete_bucket_options(&DeleteBucketOptions {
+            no_lock: true,
+            no_recreate: true,
+            force_if_empty: true,
+            ..Default::default()
+        })
+        .expect("remote delete options should serialize");
+        let legacy_opts: DeleteBucketOptions =
+            serde_json::from_str(&encoded).expect("legacy peer should decode remote delete options");
+
+        assert!(legacy_opts.no_lock);
+        assert!(legacy_opts.no_recreate);
+        assert!(!legacy_opts.force);
+        assert!(!legacy_opts.force_if_empty);
+
+        let legacy_recursive_force = if legacy_opts.force_if_empty && !legacy_opts.force {
+            true
+        } else {
+            legacy_opts.force
+        };
+        assert!(
+            !legacy_recursive_force,
+            "legacy peer must not upgrade empty-only delete to recursive force"
+        );
+    }
+
+    #[test]
+    fn remote_delete_bucket_options_preserve_explicit_force() {
+        let encoded = RemotePeerS3Client::encode_delete_bucket_options(&DeleteBucketOptions {
+            force: true,
+            force_if_empty: true,
+            ..Default::default()
+        })
+        .expect("remote force-delete options should serialize");
+        let remote_opts: DeleteBucketOptions =
+            serde_json::from_str(&encoded).expect("remote peer should decode force-delete options");
+
+        assert!(remote_opts.force);
+        assert!(!remote_opts.force_if_empty);
+    }
+
     #[tokio::test]
     async fn test_execute_with_timeout_marks_remote_peer_faulty_on_network_like_error() {
         let client = test_remote_peer("http://peer-network-error:9000");
@@ -1554,6 +1690,54 @@ mod tests {
         reset_local_disk_test_state().await;
     }
 
+    #[tokio::test]
+    #[serial]
+    async fn local_peer_force_if_empty_preserves_unclassified_file_in_selected_pool() {
+        reset_local_disk_test_state().await;
+
+        let temp_dir = TempDir::new().expect("create temp dir for empty-only delete regression");
+        let disks = init_test_local_disks_for_pools(
+            &temp_dir,
+            &[(0, 1), (1, 1)],
+            "local-peer-force-if-empty-preserves-unclassified-file",
+        )
+        .await;
+        let bucket = "empty-only-delete-bucket";
+        let marker = "object/commit-marker";
+        let data = bytes::Bytes::from_static(b"committed object data");
+
+        disks[1]
+            .make_volume(bucket)
+            .await
+            .expect("bucket should be created in the selected pool");
+        disks[1]
+            .write_all(bucket, marker, data.clone())
+            .await
+            .expect("unclassified committed file should be written");
+
+        let err = LocalPeerS3Client::new_with_local_disks(None, Some(vec![1]), disks.clone())
+            .delete_bucket(
+                bucket,
+                &DeleteBucketOptions {
+                    force_if_empty: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect_err("empty-only delete must not recursively remove an unclassified file");
+
+        assert_eq!(err, Error::VolumeNotEmpty);
+        assert_eq!(
+            disks[1]
+                .read_all(bucket, marker)
+                .await
+                .expect("unclassified committed file should be preserved"),
+            data
+        );
+
+        reset_local_disk_test_state().await;
+    }
+
     #[tokio::test]
     #[serial]
     async fn heal_bucket_local_recreates_missing_bucket_volumes() {
@@ -1618,6 +1802,30 @@ mod tests {
         assert_eq!(err, Error::VolumeExists);
     }
 
+    #[test]
+    fn heal_bucket_mode_fails_closed_on_incomplete_topology() {
+        let mut opts = HealOpts::default();
+        assert_eq!(
+            resolve_heal_bucket_mode(&mut opts, &[Some(Error::ErasureWriteQuorum)]),
+            Err(Error::ErasureWriteQuorum)
+        );
+        assert!(!opts.recreate);
+        assert!(!opts.remove);
+    }
+
+    #[test]
+    fn heal_bucket_mode_distinguishes_deleted_and_partial_buckets() {
+        let mut deleted = HealOpts::default();
+        resolve_heal_bucket_mode(&mut deleted, &[Some(Error::VolumeNotFound)]).unwrap();
+        assert!(deleted.remove);
+        assert!(!deleted.recreate);
+
+        let mut partial = HealOpts::default();
+        resolve_heal_bucket_mode(&mut partial, &[None, Some(Error::VolumeNotFound)]).unwrap();
+        assert!(!partial.remove);
+        assert!(partial.recreate);
+    }
+
     #[tokio::test]
     async fn test_make_bucket_reduces_quorum_by_pool_participants() {
         let peer_sys = S3PeerSys {

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

@@ -13,8 +13,8 @@
 // limitations under the License.
 
 use crate::cluster::rpc::client::{
-    TonicInterceptor, gen_tonic_signature_interceptor, is_network_like_disk_error, node_service_time_out_client,
-    node_service_time_out_client_for_class, node_service_time_out_client_no_auth,
+    AuthenticatedChannel, TonicInterceptor, gen_tonic_signature_interceptor, is_network_like_disk_error,
+    node_service_time_out_client, node_service_time_out_client_for_class, node_service_time_out_client_no_auth,
 };
 use crate::cluster::rpc::http_auth::set_tonic_canonical_body_digest;
 use crate::cluster::rpc::internode_data_transport::{
@@ -25,7 +25,7 @@ use crate::disk::error::{Error, Result};
 use crate::disk::{
     BatchReadVersionReq, BatchReadVersionResp, CheckPartsResp, DeleteOptions, DiskAPI, DiskInfo, DiskInfoOptions, DiskLocation,
     DiskOption, FileInfoVersions, FileReader, FileWriter, PartTransactionAction, ReadMultipleReq, ReadMultipleResp, ReadOptions,
-    RenameDataResp, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, batch_read_version_one_by_one,
+    RenameDataResp, SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions, batch_read_version_one_by_one,
     disk_store::{
         DEFAULT_RUSTFS_DRIVE_ACTIVE_MONITORING, ENV_RUSTFS_DRIVE_ACTIVE_MONITORING, SKIP_IF_SUCCESS_BEFORE,
         get_drive_active_check_interval, get_drive_active_check_timeout, get_drive_disk_info_timeout, get_drive_list_dir_timeout,
@@ -50,8 +50,9 @@ use rustfs_protos::proto_gen::node_service::{
     DeleteVersionRequest, DeleteVersionsRequest, DeleteVolumeRequest, DiskInfoRequest, ListDirRequest, ListVolumesRequest,
     MakeVolumeRequest, MakeVolumesRequest, PreparePartTransactionRequest, ReadAllRequest, ReadMetadataRequest,
     ReadMultipleRequest, ReadMultipleResponse, ReadPartsRequest, ReadVersionRequest, ReadXlRequest, RenameDataRequest,
-    RenameFileRequest, SettlePartTransactionRequest, StatVolumeRequest, UpdateMetadataRequest, VerifyFileRequest,
-    WriteAllRequest, WriteMetadataRequest, node_service_client::NodeServiceClient,
+    RenameFileRequest, SettlePartTransactionRequest, SnapshotLeaseReleaseRequest, SnapshotLeaseRenewRequest,
+    SnapshotLeaseRequest, SnapshotLeaseResponse, StatVolumeRequest, UpdateMetadataRequest, VerifyFileRequest, WriteAllRequest,
+    WriteMetadataRequest, node_service_client::NodeServiceClient,
 };
 use serde::{Serialize, de::DeserializeOwned};
 use std::{
@@ -70,7 +71,7 @@ use tokio::{
     time::timeout,
 };
 use tokio_util::sync::CancellationToken;
-use tonic::{Code, Request, service::interceptor::InterceptedService, transport::Channel};
+use tonic::{Code, Request, service::interceptor::InterceptedService};
 use tracing::{debug, trace, warn};
 use uuid::Uuid;
 
@@ -100,6 +101,18 @@ const LOG_COMPONENT_ECSTORE: &str = "ecstore";
 const LOG_SUBSYSTEM_REMOTE_DISK: &str = "remote_disk";
 const EVENT_REMOTE_DISK_HEALTH: &str = "remote_disk_health";
 const EVENT_REMOTE_DISK_RPC: &str = "remote_disk_rpc";
+const SNAPSHOT_LEASE_PROTOCOL_VERSION: u32 = 1;
+pub const REMOTE_SNAPSHOT_LEASE_TTL: Duration = Duration::from_secs(60);
+
+fn snapshot_lease_token_from_response(response: SnapshotLeaseResponse) -> Result<SnapshotLeaseToken> {
+    if !response.success {
+        return Err(response.error.unwrap_or_default().into());
+    }
+    if response.protocol_version != SNAPSHOT_LEASE_PROTOCOL_VERSION {
+        return Err(Error::other("remote snapshot lease protocol is incompatible"));
+    }
+    SnapshotLeaseToken::from_slice(&response.token)
+}
 
 /// Bind a mutating disk RPC to its canonical body: the digest lands in the request metadata, and
 /// the signing interceptor folds it (plus a replay-protected nonce) into the v2 signature scope
@@ -1070,7 +1083,7 @@ impl RemoteDisk {
         internode_offline_bypass_reason(&self.addr).map(Error::other)
     }
 
-    async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
         if let Some(err) = self.offline_bypass_error() {
             return Err(err);
         }
@@ -1083,7 +1096,7 @@ impl RemoteDisk {
     /// Routes onto the isolated bulk channel pool so large transfers cannot head-of-line block
     /// lock/health RPCs (grpc-optimization P1). Falls back to the control channel when isolation
     /// is disabled.
-    async fn get_bulk_client(&self) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    async fn get_bulk_client(&self) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
         if let Some(err) = self.offline_bypass_error() {
             return Err(err);
         }
@@ -1784,6 +1797,81 @@ impl DiskAPI for RemoteDisk {
         .await
     }
 
+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(SnapshotLeaseRequest {
+                    disk: self.endpoint.to_string(),
+                    volume: volume.to_string(),
+                    path: path.to_string(),
+                    ttl_ms: u64::try_from(REMOTE_SNAPSHOT_LEASE_TTL.as_millis())
+                        .map_err(|_| Error::other("snapshot lease TTL cannot be represented"))?,
+                });
+                let canonical_body = rustfs_protos::canonical_snapshot_lease_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "acquire_snapshot_lease")?;
+                let response = client.acquire_snapshot_lease(request).await?.into_inner();
+                snapshot_lease_token_from_response(response)
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(SnapshotLeaseRenewRequest {
+                    disk: self.endpoint.to_string(),
+                    volume: volume.to_string(),
+                    path: path.to_string(),
+                    token: token.as_bytes().to_vec().into(),
+                    ttl_ms: u64::try_from(REMOTE_SNAPSHOT_LEASE_TTL.as_millis())
+                        .map_err(|_| Error::other("snapshot lease TTL cannot be represented"))?,
+                });
+                let canonical_body = rustfs_protos::canonical_snapshot_lease_renew_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "renew_snapshot_lease")?;
+                let response = client.renew_snapshot_lease(request).await?.into_inner();
+                snapshot_lease_token_from_response(response)
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<()> {
+        self.execute_with_timeout(
+            || async {
+                let mut client = self
+                    .get_client()
+                    .await
+                    .map_err(|err| Error::other(format!("can not get client, err: {err}")))?;
+                let mut request = Request::new(SnapshotLeaseReleaseRequest {
+                    disk: self.endpoint.to_string(),
+                    volume: volume.to_string(),
+                    path: path.to_string(),
+                    token: token.as_bytes().to_vec().into(),
+                });
+                let canonical_body = rustfs_protos::canonical_snapshot_lease_release_request_body(request.get_ref());
+                attach_mutation_body_digest(&mut request, canonical_body, "release_snapshot_lease")?;
+                let response = client.release_snapshot_lease(request).await?.into_inner();
+                if !response.success {
+                    return Err(response.error.unwrap_or_default().into());
+                }
+                Ok(())
+            },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
     #[tracing::instrument(level = "trace", skip_all)]
     async fn write_metadata(&self, _org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()> {
         trace!(
@@ -2917,6 +3005,7 @@ mod tests {
     use crate::cluster::rpc::internode_data_transport::{InternodeDataTransportCapabilities, TcpHttpInternodeDataTransport};
     use crate::runtime::sources as runtime_sources;
     use serde_json::Value;
+    use serial_test::serial;
     use std::io::{self as std_io, Write};
     use std::pin::Pin;
     use std::sync::{Arc, Mutex, Mutex as StdMutex, Once};
@@ -2930,6 +3019,48 @@ mod tests {
 
     static INIT: Once = Once::new();
 
+    // `#[serial(internode_metrics)]` marks every test that observes
+    // `global_internode_metrics()`. Those counters are a process-wide singleton:
+    // some of these tests snapshot a counter, run one decode, and assert on the
+    // delta, while others deliberately record decode errors or call
+    // `reset_internode_metrics_for_test()`. Run concurrently in one process they
+    // corrupt each other's deltas — a sibling's error bumps the "no decode error"
+    // assertion off zero, and a sibling's reset can drive an `after > before`
+    // assertion backwards.
+    //
+    // The marker only takes effect under the `cargo test` fallback; nextest
+    // already isolates each test in its own process, so every test there gets its
+    // own copy of the counters (see `docs/testing/README.md`). Any new test that
+    // reads or mutates the global internode metrics belongs in this group.
+
+    #[test]
+    fn snapshot_lease_response_requires_current_protocol_and_valid_token() {
+        let token = SnapshotLeaseToken::new();
+        let response = SnapshotLeaseResponse {
+            success: true,
+            token: token.as_bytes().to_vec().into(),
+            protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+            error: None,
+        };
+        assert_eq!(snapshot_lease_token_from_response(response).unwrap(), token);
+
+        let incompatible = SnapshotLeaseResponse {
+            success: true,
+            token: token.as_bytes().to_vec().into(),
+            protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION + 1,
+            error: None,
+        };
+        assert!(snapshot_lease_token_from_response(incompatible).is_err());
+
+        let malformed = SnapshotLeaseResponse {
+            success: true,
+            token: Bytes::from_static(b"not-a-uuid"),
+            protocol_version: SNAPSHOT_LEASE_PROTOCOL_VERSION,
+            error: None,
+        };
+        assert!(snapshot_lease_token_from_response(malformed).is_err());
+    }
+
     #[test]
     fn list_volumes_decode_rejects_a_malformed_entry() {
         let valid = serde_json::to_string(&VolumeInfo {
@@ -3190,6 +3321,7 @@ mod tests {
     }
 
     #[test]
+    #[serial(internode_metrics)]
     fn read_multiple_response_decode_prefers_msgpack_payloads() {
         let endpoint = sample_remote_endpoint();
         let msgpack_resp = sample_read_multiple_resp("msgpack", b"binary");
@@ -3212,6 +3344,7 @@ mod tests {
     }
 
     #[test]
+    #[serial(internode_metrics)]
     fn read_multiple_response_decode_falls_back_to_json_payloads() {
         let endpoint = sample_remote_endpoint();
         let json_resp = sample_read_multiple_resp("json", b"fallback");
@@ -3233,6 +3366,7 @@ mod tests {
     }
 
     #[test]
+    #[serial(internode_metrics)]
     fn rename_data_response_accepts_legacy_json_without_decode_error() {
         crate::cluster::rpc::runtime_sources::reset_internode_metrics_for_test();
         let response = RenameDataResp {
@@ -3384,6 +3518,7 @@ mod tests {
     }
 
     #[test]
+    #[serial(internode_metrics)]
     fn read_multiple_response_decode_reports_corrupt_msgpack_item() {
         let endpoint = sample_remote_endpoint();
         let response = ReadMultipleResponse {
@@ -3409,6 +3544,7 @@ mod tests {
     }
 
     #[test]
+    #[serial(internode_metrics)]
     fn read_multiple_response_decode_reports_corrupt_json_item() {
         let endpoint = sample_remote_endpoint();
         let response = ReadMultipleResponse {
@@ -3449,6 +3585,7 @@ mod tests {
     }
 
     #[test]
+    #[serial(internode_metrics)]
     fn batch_read_version_response_decode_prefers_msgpack_payloads() {
         let endpoint = sample_remote_endpoint();
         let msgpack_resp = sample_batch_read_version_resp(7, "msgpack-object", true);
@@ -3469,6 +3606,7 @@ mod tests {
     }
 
     #[test]
+    #[serial(internode_metrics)]
     fn batch_read_version_response_rejects_invalid_success_metadata() {
         let endpoint = sample_remote_endpoint();
         let mut response_item = sample_batch_read_version_resp(0, "invalid-object", true);
@@ -3488,6 +3626,7 @@ mod tests {
     }
 
     #[test]
+    #[serial(internode_metrics)]
     fn batch_read_version_response_decode_reports_corrupt_msgpack_item() {
         let endpoint = sample_remote_endpoint();
         let response = BatchReadVersionResponse {
@@ -3514,6 +3653,7 @@ mod tests {
     }
 
     #[test]
+    #[serial(internode_metrics)]
     fn batch_read_version_response_decode_reports_corrupt_json_item() {
         let endpoint = sample_remote_endpoint();
         let response = BatchReadVersionResponse {
@@ -4303,6 +4443,7 @@ mod tests {
     }
 
     #[tokio::test]
+    #[serial(internode_metrics)]
     async fn test_remote_disk_create_file_retries_once_on_retryable_open_write_error() {
         let transport = RetryingOpenWriteInternodeDataTransport::with_steps(vec![
             OpenWriteTestStep::Error(DiskError::from(rustfs_rio::new_test_internode_http_io_error(
@@ -4341,6 +4482,7 @@ mod tests {
     }
 
     #[tokio::test]
+    #[serial(internode_metrics)]
     async fn test_remote_disk_read_file_stream_retries_once_on_retryable_open_read_error() {
         // A transient reset-by-peer on a shard read during the read-after-write window must be
         // absorbed by one re-dial rather than eroding read quorum (issue #2761).
@@ -5190,6 +5332,11 @@ mod tests {
                 .with_span_list(true),
         );
         let _guard = tracing::subscriber::set_default(subscriber);
+        // The `recovery-monitor` span and the monitor's own log events are
+        // production callsites that sibling tests exercise from subscriber-less
+        // threads; without this they can be cached as `Interest::never()` and go
+        // silently missing here.
+        let _callsite_pin = crate::test_tracing::pin_callsite_interest_for_test();
 
         let endpoint = Endpoint {
             url: url::Url::parse("http://127.0.0.1:59996/data").expect("endpoint URL should parse"),
@@ -5244,6 +5391,11 @@ mod tests {
                 .with_span_list(true),
         );
         let _guard = tracing::subscriber::set_default(subscriber);
+        // The `recovery-monitor` span and the monitor's own log events are
+        // production callsites that sibling tests exercise from subscriber-less
+        // threads; without this they can be cached as `Interest::never()` and go
+        // silently missing here.
+        let _callsite_pin = crate::test_tracing::pin_callsite_interest_for_test();
 
         let addr = "http://127.0.0.1:59997".to_string();
         let endpoint = Endpoint {

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

@@ -12,7 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-use crate::cluster::rpc::client::{TonicInterceptor, gen_tonic_signature_interceptor, node_service_time_out_client};
+use crate::cluster::rpc::client::{
+    AuthenticatedChannel, TonicInterceptor, gen_tonic_signature_interceptor, node_service_time_out_client,
+};
+use crate::cluster::rpc::set_tonic_mutation_body_digest;
 use async_trait::async_trait;
 use bytes::Bytes;
 use rustfs_lock::{
@@ -28,7 +31,6 @@ use std::time::Duration;
 use tokio::time::timeout;
 use tonic::Request;
 use tonic::service::interceptor::InterceptedService;
-use tonic::transport::Channel;
 use tracing::{debug, info, warn};
 
 /// Remote lock client implementation
@@ -77,7 +79,7 @@ impl RemoteClient {
         }
     }
 
-    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<Channel, TonicInterceptor>>> {
+    pub async fn get_client(&self) -> Result<NodeServiceClient<InterceptedService<AuthenticatedChannel, TonicInterceptor>>> {
         // P3-2 offline bypass (now covering the lock path too): fast-fail a peer already marked
         // offline instead of paying the connect timeout, so dsync reaches quorum sooner. Does not
         // change quorum; the self-healing re-probe keeps the peer recoverable.
@@ -313,10 +315,11 @@ impl LockClient for RemoteClient {
         info!("remote acquire_exclusive for {}", request.resource);
         let mut client = self.get_client().await?;
         let resource_summary = request.resource.to_string();
-        let req = Request::new(GenerallyLockRequest {
+        let mut req = Request::new(GenerallyLockRequest {
             args: serde_json::to_string(&request)
                 .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?,
         });
+        set_tonic_mutation_body_digest(&mut req)?;
 
         let resp = match self.execute_rpc("lock", &resource_summary, client.lock(req)).await {
             Ok(resp) => resp.into_inner(),
@@ -347,7 +350,7 @@ impl LockClient for RemoteClient {
 
         let mut client = self.get_client().await?;
         let resource_summary = Self::summarize_resources(requests);
-        let req = Request::new(BatchGenerallyLockRequest {
+        let mut req = Request::new(BatchGenerallyLockRequest {
             args: requests
                 .iter()
                 .map(|request| {
@@ -355,6 +358,7 @@ impl LockClient for RemoteClient {
                 })
                 .collect::<Result<Vec<_>>>()?,
         });
+        set_tonic_mutation_body_digest(&mut req)?;
 
         let resp = match self
             .execute_rpc("lock_batch", &resource_summary, client.lock_batch(req))
@@ -395,7 +399,8 @@ impl LockClient for RemoteClient {
             .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?;
         let mut client = self.get_client().await?;
         let resource_summary = unlock_request.resource.to_string();
-        let req = Request::new(GenerallyLockRequest { args: request_string });
+        let mut req = Request::new(GenerallyLockRequest { args: request_string });
+        set_tonic_mutation_body_digest(&mut req)?;
         let resp = self
             .execute_rpc("release", &resource_summary, client.un_lock(req))
             .await?
@@ -414,7 +419,7 @@ impl LockClient for RemoteClient {
         let unlock_requests = lock_ids.iter().map(Self::create_unlock_request).collect::<Vec<_>>();
         let mut client = self.get_client().await?;
         let resource_summary = Self::summarize_resources(&unlock_requests);
-        let req = Request::new(BatchGenerallyLockRequest {
+        let mut req = Request::new(BatchGenerallyLockRequest {
             args: unlock_requests
                 .iter()
                 .map(|request| {
@@ -422,6 +427,7 @@ impl LockClient for RemoteClient {
                 })
                 .collect::<Result<Vec<_>>>()?,
         });
+        set_tonic_mutation_body_digest(&mut req)?;
 
         let resp = self
             .execute_rpc("release_batch", &resource_summary, client.un_lock_batch(req))
@@ -440,10 +446,11 @@ impl LockClient for RemoteClient {
         let refresh_request = Self::create_unlock_request(lock_id);
         let mut client = self.get_client().await?;
         let resource_summary = refresh_request.resource.to_string();
-        let req = Request::new(GenerallyLockRequest {
+        let mut req = Request::new(GenerallyLockRequest {
             args: serde_json::to_string(&refresh_request)
                 .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?,
         });
+        set_tonic_mutation_body_digest(&mut req)?;
         let resp = self
             .execute_rpc("refresh", &resource_summary, client.refresh(req))
             .await?
@@ -459,10 +466,11 @@ impl LockClient for RemoteClient {
         let force_request = Self::create_unlock_request(lock_id);
         let mut client = self.get_client().await?;
         let resource_summary = force_request.resource.to_string();
-        let req = Request::new(GenerallyLockRequest {
+        let mut req = Request::new(GenerallyLockRequest {
             args: serde_json::to_string(&force_request)
                 .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?,
         });
+        set_tonic_mutation_body_digest(&mut req)?;
         let resp = self
             .execute_rpc("force_release", &resource_summary, client.force_un_lock(req))
             .await?
@@ -483,10 +491,11 @@ impl LockClient for RemoteClient {
         let mut client = self.get_client().await?;
 
         // Try to acquire a very short-lived lock to test availability
-        let req = Request::new(GenerallyLockRequest {
+        let mut req = Request::new(GenerallyLockRequest {
             args: serde_json::to_string(&status_request)
                 .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?,
         });
+        set_tonic_mutation_body_digest(&mut req)?;
 
         // Try exclusive lock first with very short timeout
         let resp = match self.execute_rpc("check_status", &resource_summary, client.lock(req)).await {
@@ -497,10 +506,11 @@ impl LockClient for RemoteClient {
         if resp.success {
             // If we successfully acquired the lock, the resource was free.
             // Immediately release it on a best-effort basis.
-            let release_req = Request::new(GenerallyLockRequest {
+            let mut release_req = Request::new(GenerallyLockRequest {
                 args: serde_json::to_string(&status_request)
                     .map_err(|e| LockError::internal(format!("Failed to serialize request: {e}")))?,
             });
+            set_tonic_mutation_body_digest(&mut release_req)?;
             let _ = self
                 .execute_rpc("check_status_release", &resource_summary, client.un_lock(release_req))
                 .await;

crates/ecstore/src/config/com.rs

@@ -2543,6 +2543,7 @@ mod tests {
                 data_dir: None,
                 delete_marker: false,
                 transitioned_object: Default::default(),
+                transition_version_state: Default::default(),
                 restore_ongoing: false,
                 restore_expires: None,
                 user_tags: Arc::new(String::new()),

crates/ecstore/src/disk/disk_store.rs

@@ -13,9 +13,9 @@
 // limitations under the License.
 
 use crate::disk::{
-    CheckPartsResp, DeleteOptions, DiskAPI, DiskError, DiskInfo, DiskInfoOptions, DiskLocation, Endpoint, Error,
-    FileInfoVersions, MmapCopyStageMetrics, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp, Result,
-    UpdateMetadataOpts, VolumeInfo, WalkDirOptions,
+    CheckPartsResp, DataDirDeleteStatus, DeleteOptions, DiskAPI, DiskError, DiskInfo, DiskInfoOptions, DiskLocation, Endpoint,
+    Error, FileInfoVersions, MmapCopyStageMetrics, ReadMultipleReq, ReadMultipleResp, ReadOptions, RenameDataResp, Result,
+    SnapshotLeaseToken, UpdateMetadataOpts, VolumeInfo, WalkDirOptions,
     health_state::{
         RuntimeDriveHealthState, classify_drive_recovery, get_drive_returning_probe_interval,
         get_drive_returning_success_threshold, get_drive_suspect_failure_threshold, record_drive_offline_duration,
@@ -1349,6 +1349,38 @@ impl DiskAPI for LocalDiskWrapper {
             .await
     }
 
+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
+        self.track_disk_health(
+            || async { self.disk.acquire_snapshot_lease(volume, path).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<()> {
+        self.track_disk_health(
+            || async { self.disk.release_snapshot_lease(volume, path, token).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        self.track_disk_health(
+            || async { self.disk.renew_snapshot_lease(volume, path, token).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
+    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
+        self.track_disk_health(
+            || async { self.disk.delete_data_dir(volume, path, opts).await },
+            get_max_timeout_duration(),
+        )
+        .await
+    }
+
     async fn write_metadata(&self, org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()> {
         self.track_disk_health(
             || async { self.disk.write_metadata(org_volume, volume, path, fi).await },

crates/ecstore/src/disk/error.rs

@@ -337,9 +337,54 @@ impl From<DiskError> for std::io::Error {
     }
 }
 
+/// The single in-band representation of a failed internode RPC: it carries the
+/// typed `tonic::Status` so failure classifiers can read the gRPC code instead
+/// of substring-matching the rendered message (see `is_network_like_status`).
+/// Both `DiskError` and `StorageError` wrap statuses in this type, so one
+/// downcast recovers the status regardless of which error the status was
+/// converted into first.
+pub(crate) struct RpcStatusError(tonic::Status);
+
+impl RpcStatusError {
+    pub(crate) fn status(&self) -> &tonic::Status {
+        &self.0
+    }
+}
+
+impl From<tonic::Status> for RpcStatusError {
+    fn from(status: tonic::Status) -> Self {
+        Self(status)
+    }
+}
+
+impl std::fmt::Display for RpcStatusError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        std::fmt::Display::fmt(&self.0, f)
+    }
+}
+
+/// `tonic::Status`'s own `Debug` prints its `MetadataMap`, i.e. every response
+/// header and trailer the peer sent. Those are remote-controlled and can carry
+/// credentials injected by a proxy in front of the peer, so keep them out of
+/// anything that reaches a log.
+impl std::fmt::Debug for RpcStatusError {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("RpcStatusError")
+            .field("code", &self.0.code())
+            .field("message", &self.0.message())
+            .finish()
+    }
+}
+
+impl StdError for RpcStatusError {
+    fn source(&self) -> Option<&(dyn StdError + 'static)> {
+        Some(&self.0)
+    }
+}
+
 impl From<tonic::Status> for DiskError {
     fn from(e: tonic::Status) -> Self {
-        DiskError::other(e.message().to_string())
+        DiskError::Io(io::Error::other(RpcStatusError(e)))
     }
 }
 

crates/ecstore/src/disk/mod.rs

@@ -72,6 +72,39 @@ pub type DiskStore = Arc<Disk>;
 pub type FileReader = Box<dyn AsyncRead + Send + Sync + Unpin>;
 pub type FileWriter = Box<dyn AsyncWrite + Send + Sync + Unpin>;
 
+#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq, Serialize, Deserialize)]
+pub struct SnapshotLeaseToken(Uuid);
+
+impl SnapshotLeaseToken {
+    pub fn new() -> Self {
+        Self(Uuid::new_v4())
+    }
+
+    pub fn from_slice(bytes: &[u8]) -> Result<Self> {
+        let uuid = Uuid::from_slice(bytes).map_err(|_| Error::other("invalid snapshot lease token"))?;
+        if uuid.is_nil() {
+            return Err(Error::other("invalid snapshot lease token"));
+        }
+        Ok(Self(uuid))
+    }
+
+    pub fn as_bytes(&self) -> &[u8; 16] {
+        self.0.as_bytes()
+    }
+}
+
+impl Default for SnapshotLeaseToken {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[derive(Clone, Copy, Debug, Eq, PartialEq)]
+pub enum DataDirDeleteStatus {
+    Deleted,
+    Deferred,
+}
+
 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 pub enum PartTransactionAction {
     Commit,
@@ -249,6 +282,34 @@ impl DiskAPI for Disk {
         }
     }
 
+    async fn acquire_snapshot_lease(&self, volume: &str, path: &str) -> Result<SnapshotLeaseToken> {
+        match self {
+            Disk::Local(local_disk) => local_disk.acquire_snapshot_lease(volume, path).await,
+            Disk::Remote(remote_disk) => remote_disk.acquire_snapshot_lease(volume, path).await,
+        }
+    }
+
+    async fn release_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<()> {
+        match self {
+            Disk::Local(local_disk) => local_disk.release_snapshot_lease(volume, path, token).await,
+            Disk::Remote(remote_disk) => remote_disk.release_snapshot_lease(volume, path, token).await,
+        }
+    }
+
+    async fn renew_snapshot_lease(&self, volume: &str, path: &str, token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        match self {
+            Disk::Local(local_disk) => local_disk.renew_snapshot_lease(volume, path, token).await,
+            Disk::Remote(remote_disk) => remote_disk.renew_snapshot_lease(volume, path, token).await,
+        }
+    }
+
+    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
+        match self {
+            Disk::Local(local_disk) => local_disk.delete_data_dir(volume, path, opts).await,
+            Disk::Remote(remote_disk) => remote_disk.delete_data_dir(volume, path, opts).await,
+        }
+    }
+
     #[tracing::instrument(level = "trace", skip_all)]
     async fn write_metadata(&self, _org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()> {
         match self {
@@ -646,6 +707,19 @@ pub trait DiskAPI: Debug + Send + Sync + 'static {
     ) -> Result<()>;
     async fn delete_versions(&self, volume: &str, versions: Vec<FileInfoVersions>, opts: DeleteOptions) -> Vec<Option<Error>>;
     async fn delete_paths(&self, volume: &str, paths: &[String]) -> Result<()>;
+    async fn acquire_snapshot_lease(&self, _volume: &str, _path: &str) -> Result<SnapshotLeaseToken> {
+        Err(Error::other("snapshot leases are not supported by this disk"))
+    }
+    async fn release_snapshot_lease(&self, _volume: &str, _path: &str, _token: SnapshotLeaseToken) -> Result<()> {
+        Err(Error::other("snapshot leases are not supported by this disk"))
+    }
+    async fn renew_snapshot_lease(&self, _volume: &str, _path: &str, _token: SnapshotLeaseToken) -> Result<SnapshotLeaseToken> {
+        Err(Error::other("snapshot leases are not supported by this disk"))
+    }
+    async fn delete_data_dir(&self, volume: &str, path: &str, opts: DeleteOptions) -> Result<DataDirDeleteStatus> {
+        self.delete(volume, path, opts).await?;
+        Ok(DataDirDeleteStatus::Deleted)
+    }
     async fn write_metadata(&self, org_volume: &str, volume: &str, path: &str, fi: FileInfo) -> Result<()>;
     async fn update_metadata(&self, volume: &str, path: &str, fi: FileInfo, opts: &UpdateMetadataOpts) -> Result<()>;
     async fn read_version(

crates/ecstore/src/disk/os.rs

@@ -25,7 +25,7 @@ use std::{
     sync::{Arc, LazyLock, Weak},
 };
 use tokio::fs;
-use tokio::sync::{OwnedSemaphorePermit, Semaphore, SemaphorePermit};
+use tokio::sync::{OwnedSemaphorePermit, RwLock, Semaphore, SemaphorePermit};
 use tracing::warn;
 
 /// Check path length according to OS limits.
@@ -123,6 +123,8 @@ const TEST_GLOBAL_FILE_SYNCS: usize = 64;
 
 static FILE_SYNC_PERMITS: LazyLock<Semaphore> = LazyLock::new(|| Semaphore::new(global_file_sync_limit()));
 static DISK_FILE_SYNC_LIMITERS: LazyLock<Mutex<HashMap<PathBuf, Weak<Semaphore>>>> = LazyLock::new(|| Mutex::new(HashMap::new()));
+static DISK_VOLUME_MUTATION_LOCKS: LazyLock<Mutex<HashMap<PathBuf, Weak<RwLock<()>>>>> =
+    LazyLock::new(|| Mutex::new(HashMap::new()));
 
 fn default_global_file_sync_limit(cpu_count: usize, max_blocking_threads: usize) -> usize {
     let cpu_scaled = cpu_count
@@ -157,6 +159,24 @@ pub(crate) fn disk_file_sync_limiter(root: &Path) -> Arc<Semaphore> {
     limiter
 }
 
+/// Serialize a bucket's local metadata commits with physical bucket removal.
+///
+/// The key includes the canonical disk root, so independently reconnected
+/// [`LocalDisk`](super::local::LocalDisk) instances share the same lock while
+/// disconnected disks do not keep the registry alive.
+pub(crate) fn disk_volume_mutation_lock(root: &Path, volume: &str) -> Arc<RwLock<()>> {
+    let key = root.join(volume);
+    let mut locks = DISK_VOLUME_MUTATION_LOCKS.lock();
+    locks.retain(|_, lock| lock.strong_count() > 0);
+    if let Some(lock) = locks.get(&key).and_then(Weak::upgrade) {
+        return lock;
+    }
+
+    let lock = Arc::new(RwLock::new(()));
+    locks.insert(key, Arc::downgrade(&lock));
+    lock
+}
+
 /// Always acquire the per-disk permit before the process-wide permit. Keeping
 /// this order uniform prevents one slow disk from reserving global capacity
 /// while it waits for its own concurrency slot.
@@ -572,15 +592,14 @@ async fn reliable_rename_inner(
     base_dir: impl AsRef<Path>,
     warn_on_missing_source: bool,
 ) -> io::Result<()> {
-    if let Some(parent) = dst_file_path.as_ref().parent()
-        && !file_exists(parent)
-    {
-        reliable_mkdir_all(parent, base_dir.as_ref()).await?;
-    }
+    let parent_guard = match dst_file_path.as_ref().parent() {
+        Some(parent) => Some(mkdir_all_below_existing_base(parent, base_dir.as_ref()).await?),
+        None => None,
+    };
 
     let mut i = 0;
     loop {
-        if let Err(e) = super::fs::rename_std(src_file_path.as_ref(), dst_file_path.as_ref()) {
+        if let Err(e) = rename_into_existing_parent(src_file_path.as_ref(), dst_file_path.as_ref(), parent_guard.as_ref()) {
             if should_retry_rename(&e, i) {
                 i += 1;
                 continue;
@@ -597,6 +616,159 @@ async fn reliable_rename_inner(
     Ok(())
 }
 
+#[cfg(unix)]
+fn rename_into_existing_parent(
+    src_file_path: &Path,
+    dst_file_path: &Path,
+    parent_guard: Option<&ExistingBaseDirectoryGuard>,
+) -> io::Result<()> {
+    use rustix::fs::{Mode, OFlags, open, renameat};
+
+    let Some(parent_guard) = parent_guard else {
+        return super::fs::rename_std(src_file_path, dst_file_path);
+    };
+    let src_parent = src_file_path
+        .parent()
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "rename source must have a parent directory"))?;
+    let src_name = src_file_path
+        .file_name()
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "rename source must have a file name"))?;
+    let dst_name = dst_file_path
+        .file_name()
+        .ok_or_else(|| io::Error::new(io::ErrorKind::InvalidInput, "rename destination must have a file name"))?;
+    let src_parent = open(
+        src_parent,
+        OFlags::RDONLY | OFlags::DIRECTORY | OFlags::NOFOLLOW | OFlags::CLOEXEC,
+        Mode::empty(),
+    )
+    .map_err(io::Error::from)?;
+    let dst_parent = parent_guard
+        .last()
+        .ok_or_else(|| io::Error::other("rename destination parent guard is empty"))?;
+
+    renameat(&src_parent, src_name, dst_parent, dst_name).map_err(io::Error::from)
+}
+
+#[cfg(not(unix))]
+fn rename_into_existing_parent(
+    src_file_path: &Path,
+    dst_file_path: &Path,
+    _parent_guard: Option<&ExistingBaseDirectoryGuard>,
+) -> io::Result<()> {
+    super::fs::rename_std(src_file_path, dst_file_path)
+}
+
+async fn mkdir_all_below_existing_base(dir_path: &Path, base_dir: &Path) -> io::Result<ExistingBaseDirectoryGuard> {
+    let dir_path = dir_path.to_path_buf();
+    let base_dir = base_dir.to_path_buf();
+
+    tokio::task::spawn_blocking(move || mkdir_all_below_existing_base_std(&dir_path, &base_dir)).await?
+}
+
+#[cfg(windows)]
+pub(crate) type ExistingBaseDirectoryGuard = Vec<winapi_util::Handle>;
+
+#[cfg(unix)]
+pub(crate) type ExistingBaseDirectoryGuard = Vec<std::os::fd::OwnedFd>;
+
+#[cfg(all(not(unix), not(windows)))]
+pub(crate) type ExistingBaseDirectoryGuard = ();
+
+#[cfg(windows)]
+fn lock_windows_directory(path: &Path) -> io::Result<winapi_util::Handle> {
+    use std::os::windows::fs::OpenOptionsExt;
+
+    const FILE_ATTRIBUTE_DIRECTORY: u64 = 0x10;
+    const FILE_ATTRIBUTE_REPARSE_POINT: u32 = 0x400;
+    const FILE_FLAG_BACKUP_SEMANTICS: u32 = 0x0200_0000;
+    const FILE_FLAG_OPEN_REPARSE_POINT: u32 = 0x0020_0000;
+    const FILE_SHARE_READ: u32 = 0x1;
+
+    let file = std::fs::OpenOptions::new()
+        .read(true)
+        .share_mode(FILE_SHARE_READ)
+        .custom_flags(FILE_FLAG_BACKUP_SEMANTICS | FILE_FLAG_OPEN_REPARSE_POINT)
+        .open(path)?;
+    let handle = winapi_util::Handle::from_file(file);
+    let info = winapi_util::file::information(&handle)?;
+    if info.file_attributes() & FILE_ATTRIBUTE_DIRECTORY == 0
+        || info.file_attributes() & u64::from(FILE_ATTRIBUTE_REPARSE_POINT) != 0
+    {
+        return Err(io::Error::from(io::ErrorKind::NotADirectory));
+    }
+    Ok(handle)
+}
+
+pub(crate) fn mkdir_all_below_existing_base_std(dir_path: &Path, base_dir: &Path) -> io::Result<ExistingBaseDirectoryGuard> {
+    let relative = dir_path
+        .strip_prefix(base_dir)
+        .map_err(|_| io::Error::new(io::ErrorKind::InvalidInput, "rename destination must remain below its base directory"))?;
+    for component in relative.components() {
+        if !matches!(component, Component::Normal(_) | Component::CurDir) {
+            return Err(io::Error::new(
+                io::ErrorKind::InvalidInput,
+                "rename destination contains an invalid path component",
+            ));
+        }
+    }
+
+    #[cfg(unix)]
+    {
+        use rustix::fs::{Mode, OFlags, mkdirat, open, openat};
+        use rustix::io::Errno;
+
+        let flags = OFlags::RDONLY | OFlags::DIRECTORY | OFlags::NOFOLLOW | OFlags::CLOEXEC;
+        let mode = Mode::RWXU | Mode::RWXG | Mode::RWXO;
+        let mut parents = vec![open(base_dir, flags, Mode::empty()).map_err(io::Error::from)?];
+
+        for component in relative.components() {
+            let Component::Normal(component) = component else {
+                continue;
+            };
+            let parent = parents
+                .last()
+                .expect("base directory guard should contain the base directory");
+            match mkdirat(parent, component, mode) {
+                Ok(()) => {}
+                Err(Errno::EXIST) => {}
+                Err(err) => return Err(err.into()),
+            }
+            parents.push(openat(parent, component, flags, Mode::empty()).map_err(io::Error::from)?);
+        }
+
+        Ok(parents)
+    }
+
+    #[cfg(windows)]
+    {
+        let mut handles = vec![lock_windows_directory(base_dir)?];
+        let mut current = base_dir.to_path_buf();
+        for component in relative.components() {
+            let Component::Normal(component) = component else {
+                continue;
+            };
+            current.push(component);
+            match std::fs::create_dir(&current) {
+                Ok(()) => {}
+                Err(err) if err.kind() == io::ErrorKind::AlreadyExists => {}
+                Err(err) => return Err(err),
+            }
+            handles.push(lock_windows_directory(&current)?);
+        }
+
+        Ok(handles)
+    }
+
+    #[cfg(all(not(unix), not(windows)))]
+    {
+        let _ = relative;
+        Err(io::Error::new(
+            io::ErrorKind::Unsupported,
+            "safe recursive directory creation is unavailable on this platform",
+        ))
+    }
+}
+
 fn warn_reliable_rename_failure(src_file_path: &Path, dst_file_path: &Path, base_dir: &Path, err: &io::Error) {
     warn!(
         "reliable_rename failed. src_file_path: {:?}, dst_file_path: {:?}, base_dir: {:?}, err: {:?}",
@@ -727,6 +899,20 @@ mod tests {
         Arc::new(Semaphore::new(MAX_PARALLEL_FILE_SYNCS))
     }
 
+    #[tokio::test]
+    async fn disk_volume_mutation_lock_is_shared_per_root_and_volume() {
+        let temp_dir = tempdir().expect("create temp dir");
+        let first = disk_volume_mutation_lock(temp_dir.path(), "bucket");
+        let second = disk_volume_mutation_lock(temp_dir.path(), "bucket");
+        let other = disk_volume_mutation_lock(temp_dir.path(), "other-bucket");
+
+        assert!(Arc::ptr_eq(&first, &second), "reconnected disks must share a bucket mutation lock");
+        assert!(!Arc::ptr_eq(&first, &other), "different buckets must not serialize each other");
+
+        let _write_guard = first.write().await;
+        assert!(second.try_read().is_err(), "a bucket delete lock must exclude local commits");
+    }
+
     #[derive(Clone, Default)]
     struct CapturedLogs {
         buffer: Arc<Mutex<Vec<u8>>>,
@@ -771,9 +957,26 @@ mod tests {
         }
     }
 
+    /// Holds a `warn_capture()` capture alive: the thread-local subscriber, plus
+    /// the pin that keeps tracing's process-global callsite-interest cache from
+    /// being decided by some other test's thread.
+    struct WarnCaptureGuard {
+        _subscriber: tracing::subscriber::DefaultGuard,
+        _callsite_pin: tracing::Dispatch,
+    }
+
     /// Capture WARN-level output on the current thread; tokio tests here run on
     /// the current-thread runtime, so the guard covers the whole test body.
-    fn warn_capture() -> (CapturedLogs, tracing::subscriber::DefaultGuard) {
+    ///
+    /// The callsite pin matters because `warn_reliable_rename_failure` is a
+    /// single production callsite shared with tests that call `rename_all`
+    /// *without* installing a subscriber — `rename_all_missing_source_returns_file_not_found`
+    /// is one. Whichever thread reaches it first fixes its `Interest`
+    /// process-wide, so without the pin that sibling can cache
+    /// `Interest::never()` and the WARN never fires here at all, leaving the
+    /// "must keep the WARN" assertions staring at empty output. See
+    /// [`crate::test_tracing::pin_callsite_interest_for_test`].
+    fn warn_capture() -> (CapturedLogs, WarnCaptureGuard) {
         let logs = CapturedLogs::default();
         let subscriber = tracing_subscriber::fmt()
             .with_max_level(tracing::Level::WARN)
@@ -781,7 +984,10 @@ mod tests {
             .with_ansi(false)
             .without_time()
             .finish();
-        let guard = tracing::subscriber::set_default(subscriber);
+        let guard = WarnCaptureGuard {
+            _subscriber: tracing::subscriber::set_default(subscriber),
+            _callsite_pin: crate::test_tracing::pin_callsite_interest_for_test(),
+        };
         (logs, guard)
     }
 
@@ -968,6 +1174,112 @@ mod tests {
         assert_eq!(std::fs::read(dst.join("nested").join("part.1")).expect("read moved part"), b"payload");
     }
 
+    #[tokio::test]
+    async fn rename_all_does_not_recreate_missing_base_directory() {
+        let temp_dir = tempdir().expect("create temp dir");
+        let base = temp_dir.path().join("bucket");
+        std::fs::create_dir(&base).expect("create destination base");
+        let src = temp_dir.path().join("staged-object");
+        std::fs::write(&src, b"payload").expect("write staged object");
+        let dst = base.join("object").join("xl.meta");
+        std::fs::remove_dir(&base).expect("delete destination base before commit");
+
+        let err = rename_all(&src, &dst, &base)
+            .await
+            .expect_err("rename must not recreate a deleted destination base");
+
+        assert!(matches!(err, DiskError::FileNotFound));
+        assert!(src.exists(), "failed commit must preserve the staged source");
+        assert!(!base.exists(), "failed commit must not recreate the deleted bucket");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn rename_all_rejects_a_replaced_base_with_an_existing_parent() {
+        use std::os::unix::fs::symlink;
+
+        let temp_dir = tempdir().expect("create temp dir");
+        let base = temp_dir.path().join("bucket");
+        let outside = temp_dir.path().join("outside");
+        std::fs::create_dir(&base).expect("create destination base");
+        std::fs::create_dir_all(outside.join("object")).expect("create outside destination parent");
+        let src = temp_dir.path().join("staged-object");
+        std::fs::write(&src, b"payload").expect("write staged object");
+        let dst = base.join("object").join("xl.meta");
+
+        std::fs::remove_dir(&base).expect("remove destination base before replacement");
+        symlink(&outside, &base).expect("replace destination base with a symlink");
+
+        rename_all(&src, &dst, &base)
+            .await
+            .expect_err("rename must reject an existing destination parent below a replaced base");
+
+        assert!(src.exists(), "rejected rename must preserve the staged source");
+        assert!(
+            !outside.join("object/xl.meta").exists(),
+            "rename must not publish through the replacement symlink"
+        );
+    }
+
+    #[cfg(windows)]
+    #[test]
+    fn windows_parent_guard_blocks_base_and_intermediate_replacement() {
+        let temp_dir = tempdir().expect("create temp dir");
+        let base = temp_dir.path().join("bucket");
+        std::fs::create_dir(&base).expect("create destination base");
+        let parent = base.join("object").join("nested");
+        let guard = mkdir_all_below_existing_base_std(&parent, &base).expect("create and lock destination parents");
+
+        std::fs::rename(&base, temp_dir.path().join("replacement-base"))
+            .expect_err("the locked base must not be replaceable before commit");
+        std::fs::rename(base.join("object"), base.join("replacement-object"))
+            .expect_err("a locked intermediate directory must not be replaceable before commit");
+
+        drop(guard);
+        std::fs::rename(base.join("object"), base.join("replacement-object"))
+            .expect("replacement should succeed after the commit guard is released");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn rename_parent_creation_rejects_symlinked_base() {
+        use std::os::unix::fs::symlink;
+
+        let temp_dir = tempdir().expect("create temp dir");
+        let outside = temp_dir.path().join("outside");
+        std::fs::create_dir(&outside).expect("create outside directory");
+        let base = temp_dir.path().join("bucket");
+        symlink(&outside, &base).expect("create symlinked base");
+
+        mkdir_all_below_existing_base(&base.join("object"), &base)
+            .await
+            .expect_err("symlinked base must be rejected");
+
+        assert!(!outside.join("object").exists(), "parent creation must remain confined to the base");
+    }
+
+    #[cfg(unix)]
+    #[tokio::test]
+    async fn rename_parent_creation_rejects_symlink_below_base() {
+        use std::os::unix::fs::symlink;
+
+        let temp_dir = tempdir().expect("create temp dir");
+        let base = temp_dir.path().join("bucket");
+        let outside = temp_dir.path().join("outside");
+        std::fs::create_dir(&base).expect("create destination base");
+        std::fs::create_dir(&outside).expect("create outside directory");
+        symlink(&outside, base.join("linked")).expect("create symlink below base");
+
+        mkdir_all_below_existing_base(&base.join("linked/object"), &base)
+            .await
+            .expect_err("symlink below base must be rejected");
+
+        assert!(
+            !outside.join("object").exists(),
+            "parent creation must not follow a symlink outside the base"
+        );
+    }
+
     #[tokio::test]
     async fn fsync_dir_succeeds_on_directory() {
         let temp_dir = tempdir().expect("create temp dir");

crates/ecstore/src/error/mod.rs

@@ -818,7 +818,11 @@ impl From<s3s::xml::DeError> for Error {
 
 impl From<tonic::Status> for Error {
     fn from(e: tonic::Status) -> Self {
-        Error::other(e.to_string())
+        // Keep the typed status as the io::Error payload instead of a
+        // flattened string so RPC failure classifiers can read the gRPC code
+        // via downcast (see `PeerRestClient::is_network_like_error`).
+        // `RpcStatusError` renders exactly as `e.to_string()` did.
+        Error::other(crate::disk::error::RpcStatusError::from(e))
     }
 }
 

crates/ecstore/src/lib.rs

@@ -93,3 +93,6 @@ pub(crate) mod ecstore_validation_blackbox;
 
 #[cfg(test)]
 pub(crate) mod test_metrics;
+
+#[cfg(test)]
+pub(crate) mod test_tracing;

crates/ecstore/src/object_api/types.rs

@@ -180,6 +180,7 @@ pub struct ObjectInfo {
     pub data_dir: Option<Uuid>,
     pub delete_marker: bool,
     pub transitioned_object: TransitionedObject,
+    pub transition_version_state: rustfs_filemeta::TransitionVersionState,
     pub restore_ongoing: bool,
     pub restore_expires: Option<OffsetDateTime>,
     pub user_tags: Arc<String>,
@@ -220,6 +221,7 @@ impl Clone for ObjectInfo {
             data_dir: self.data_dir,
             delete_marker: self.delete_marker,
             transitioned_object: self.transitioned_object.clone(),
+            transition_version_state: self.transition_version_state,
             restore_ongoing: self.restore_ongoing,
             restore_expires: self.restore_expires,
             user_tags: self.user_tags.clone(),
@@ -464,11 +466,11 @@ impl ObjectInfo {
 
         let transitioned_object = TransitionedObject {
             name: fi.transitioned_objname.clone(),
-            version_id: if let Some(transition_version_id) = fi.transition_version_id {
-                transition_version_id.to_string()
-            } else {
-                "".to_string()
-            },
+            version_id: fi
+                .transition_version
+                .clone()
+                .or_else(|| fi.transition_version_id.map(|version_id| version_id.to_string()))
+                .unwrap_or_default(),
             status: fi.transition_status.clone(),
             free_version: fi.tier_free_version(),
             tier: fi.transition_tier.clone(),
@@ -537,6 +539,7 @@ impl ObjectInfo {
             inlined,
             user_defined: Arc::new(metadata),
             transitioned_object,
+            transition_version_state: fi.transition_version_state,
             checksum: fi.checksum.clone(),
             storage_class,
             restore_ongoing,

crates/ecstore/src/runtime/sources.rs

@@ -207,10 +207,6 @@ pub(crate) async fn ensure_boot_time() {
     GLOBAL_BOOT_TIME.get_or_init(|| async { SystemTime::now() }).await;
 }
 
-pub(crate) async fn scanner_init_time() -> Option<chrono::DateTime<chrono::Utc>> {
-    rustfs_common::get_global_init_time().await
-}
-
 pub(crate) async fn root_disk_threshold_for_erasure_disk() -> Option<u64> {
     if is_erasure_sd().await {
         None

crates/ecstore/src/services/metrics_realtime.rs

@@ -71,6 +71,7 @@ fn to_madmin_scanner_metrics(metrics: rustfs_common::metrics::ScannerMetricsRepo
     MadminScannerMetrics {
         collected_at: metrics.collected_at,
         current_cycle: metrics.current_cycle,
+        current_cycle_active: Some(metrics.current_cycle_active),
         current_started: metrics.current_started,
         cycles_completed_at: metrics.cycles_completed_at,
         ongoing_buckets: metrics.ongoing_buckets,
@@ -398,10 +399,7 @@ pub async fn collect_local_metrics(types: MetricType, opts: &CollectMetricsOpts)
 
     if types.contains(&MetricType::SCANNER) {
         debug!("start get scanner metrics");
-        let mut metrics = global_metrics().report().await;
-        if let Some(init_time) = runtime_sources::scanner_init_time().await {
-            metrics.current_started = init_time;
-        }
+        let metrics = global_metrics().report().await;
         real_time_metrics.aggregated.scanner = Some(to_madmin_scanner_metrics(metrics));
     }
 
@@ -540,7 +538,9 @@ async fn collect_local_disks_metrics(disks: &HashSet<String>) -> HashMap<String,
 #[cfg(test)]
 mod test {
     use super::*;
+    use rustfs_common::metrics::CurrentCycle;
     use rustfs_io_metrics::internode_metrics::global_internode_metrics;
+    use serial_test::serial;
     use std::time::Duration;
 
     #[test]
@@ -588,7 +588,10 @@ mod test {
 
     #[test]
     fn scanner_metrics_mapping_preserves_partial_source_status() {
+        let current_started = Utc::now() - chrono::Duration::seconds(5);
         let scanner = to_madmin_scanner_metrics(rustfs_common::metrics::ScannerMetricsReport {
+            current_cycle_active: true,
+            current_started,
             last_cycle_partial_source: "usage".to_string(),
             last_cycle_partial_source_code: 1,
             partial_cycles_by_source: vec![rustfs_common::metrics::ScannerSourceCycleSnapshot {
@@ -598,6 +601,8 @@ mod test {
             ..Default::default()
         });
 
+        assert_eq!(scanner.current_cycle_active, Some(true));
+        assert_eq!(scanner.current_started, current_started);
         assert_eq!(scanner.last_cycle_partial_source, "usage");
         assert_eq!(scanner.last_cycle_partial_source_code, 1);
         let usage = scanner
@@ -608,6 +613,39 @@ mod test {
         assert_eq!(usage.cycles, 2);
     }
 
+    #[tokio::test]
+    #[serial]
+    async fn collect_local_metrics_preserves_scanner_cycle_started_time() {
+        let previous_init_time = *rustfs_common::globals::GLOBAL_INIT_TIME.read().await;
+        let previous_cycle = global_metrics().get_cycle().await;
+        let init_time = Utc::now() - chrono::Duration::hours(1);
+        let cycle_started = Utc::now() - chrono::Duration::seconds(5);
+        *rustfs_common::globals::GLOBAL_INIT_TIME.write().await = Some(init_time);
+        let cycle = CurrentCycle {
+            current: 0,
+            next: 1,
+            started: cycle_started,
+            ..Default::default()
+        };
+        let cycle_start = global_metrics().start_scan_cycle_work_with_cycle(cycle).await;
+
+        let realtime = collect_local_metrics(MetricType::SCANNER, &CollectMetricsOpts::default()).await;
+
+        global_metrics()
+            .finish_scan_cycle_work_with_cycle(cycle_start, previous_cycle.clone().unwrap_or_default())
+            .await;
+        global_metrics().set_cycle(previous_cycle).await;
+        *rustfs_common::globals::GLOBAL_INIT_TIME.write().await = previous_init_time;
+
+        let encoded = rmp_serde::to_vec_named(&realtime).expect("realtime metrics should encode");
+        let decoded: RealtimeMetrics = rmp_serde::from_slice(&encoded).expect("realtime metrics should decode");
+        let mut aggregated = RealtimeMetrics::default();
+        aggregated.merge(decoded);
+        let scanner = aggregated.aggregated.scanner.expect("scanner metrics");
+        assert_eq!(scanner.current_cycle_active, Some(true));
+        assert_eq!(scanner.current_started, cycle_started);
+    }
+
     #[test]
     fn scanner_metrics_mapping_preserves_pacing_pressure() {
         let scanner = to_madmin_scanner_metrics(rustfs_common::metrics::ScannerMetricsReport {

crates/ecstore/src/services/notification_sys.rs

@@ -29,11 +29,11 @@ use rustfs_madmin::metrics::RealtimeMetrics;
 use rustfs_madmin::net::NetInfo;
 use rustfs_madmin::{ItemState, ServerProperties, StorageInfo};
 use rustfs_utils::XHost;
-use std::collections::{HashMap, hash_map::DefaultHasher};
+use std::collections::{BTreeMap, HashMap, hash_map::DefaultHasher};
 use std::future::Future;
 use std::hash::{Hash, Hasher};
 use std::sync::{Arc, Mutex, OnceLock};
-use std::time::{Duration, SystemTime};
+use std::time::{Duration, Instant, SystemTime};
 use tokio::time::{sleep, timeout};
 use tokio_util::sync::CancellationToken;
 use tracing::{debug, error, info, warn};
@@ -47,6 +47,9 @@ const EVENT_NOTIFICATION_PEER_PROPAGATION: &str = "notification_peer_propagation
 const SCANNER_ACTIVITY_PROBE_TIMEOUT: Duration = Duration::from_secs(5);
 const TIER_CONFIG_RELOAD_RETRY_BASE: Duration = Duration::from_millis(100);
 const TIER_CONFIG_RELOAD_RETRY_CAP: Duration = Duration::from_secs(5);
+const REMOTE_VERSION_STATE_PROBE_INTERVAL: Duration = Duration::from_secs(10);
+const REMOTE_VERSION_STATE_PROBE_TIMEOUT: Duration = Duration::from_secs(5);
+const REMOTE_VERSION_STATE_PROOF_TTL: Duration = Duration::from_secs(30);
 
 /// Cached result from the last successful admin call to a peer.
 struct PeerAdminCache {
@@ -91,6 +94,180 @@ lazy_static! {
     pub static ref GLOBAL_NOTIFICATION_SYS: OnceLock<Arc<NotificationSys>> = OnceLock::new();
 }
 
+#[derive(Clone)]
+struct RemoteVersionStateFleetProof {
+    topology_fingerprint: String,
+    peer_epochs: Arc<BTreeMap<String, Uuid>>,
+    expires_at: Instant,
+}
+
+impl RemoteVersionStateFleetProof {
+    fn token(&self) -> RemoteVersionStateFleetProofToken {
+        RemoteVersionStateFleetProofToken {
+            topology_fingerprint: self.topology_fingerprint.clone(),
+            peer_epochs: self.peer_epochs.clone(),
+        }
+    }
+}
+
+#[derive(Clone, PartialEq, Eq)]
+pub(crate) struct RemoteVersionStateFleetProofToken {
+    topology_fingerprint: String,
+    peer_epochs: Arc<BTreeMap<String, Uuid>>,
+}
+
+#[derive(Default)]
+struct RemoteVersionStateFleetProofState {
+    proof: Option<RemoteVersionStateFleetProof>,
+    topology_conflict: bool,
+}
+
+static REMOTE_VERSION_STATE_FLEET_PROOF: OnceLock<std::sync::RwLock<RemoteVersionStateFleetProofState>> = OnceLock::new();
+static REMOTE_VERSION_STATE_PROBE_TOPOLOGY: OnceLock<String> = OnceLock::new();
+
+fn remote_version_state_fleet_proof_slot() -> &'static std::sync::RwLock<RemoteVersionStateFleetProofState> {
+    REMOTE_VERSION_STATE_FLEET_PROOF.get_or_init(|| std::sync::RwLock::new(RemoteVersionStateFleetProofState::default()))
+}
+
+fn replace_remote_version_state_fleet_proof(proof: Option<RemoteVersionStateFleetProof>) {
+    replace_remote_version_state_fleet_proof_in(remote_version_state_fleet_proof_slot(), proof);
+}
+
+fn replace_remote_version_state_fleet_proof_in(
+    slot: &std::sync::RwLock<RemoteVersionStateFleetProofState>,
+    proof: Option<RemoteVersionStateFleetProof>,
+) {
+    slot.write().unwrap_or_else(std::sync::PoisonError::into_inner).proof = proof;
+}
+
+fn publish_remote_version_state_probe_result(
+    slot: &std::sync::RwLock<RemoteVersionStateFleetProofState>,
+    topology_fingerprint: &str,
+    result: Result<BTreeMap<String, Uuid>>,
+    observed_at: Instant,
+) -> Option<Error> {
+    match result {
+        Ok(peer_epochs) => {
+            let mut state = slot.write().unwrap_or_else(std::sync::PoisonError::into_inner);
+            let peer_epochs = state
+                .proof
+                .as_ref()
+                .filter(|proof| proof.topology_fingerprint == topology_fingerprint && proof.peer_epochs.as_ref() == &peer_epochs)
+                .map(|proof| Arc::clone(&proof.peer_epochs))
+                .unwrap_or_else(|| Arc::new(peer_epochs));
+            state.proof = Some(RemoteVersionStateFleetProof {
+                topology_fingerprint: topology_fingerprint.to_string(),
+                peer_epochs,
+                expires_at: observed_at + REMOTE_VERSION_STATE_PROOF_TTL,
+            });
+            None
+        }
+        Err(err) => {
+            replace_remote_version_state_fleet_proof_in(slot, None);
+            Some(err)
+        }
+    }
+}
+
+pub(crate) fn acquire_remote_version_state_fleet_proof() -> Option<RemoteVersionStateFleetProofToken> {
+    let expected_topology = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get()?;
+    let state = remote_version_state_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    acquire_remote_version_state_fleet_proof_from(&state, expected_topology, Instant::now())
+}
+
+fn acquire_remote_version_state_fleet_proof_from(
+    state: &RemoteVersionStateFleetProofState,
+    expected_topology: &str,
+    now: Instant,
+) -> Option<RemoteVersionStateFleetProofToken> {
+    if state.topology_conflict || !remote_version_state_fleet_proof_valid_at(state.proof.as_ref(), expected_topology, now) {
+        return None;
+    }
+    state.proof.as_ref().map(RemoteVersionStateFleetProof::token)
+}
+
+pub(crate) fn remote_version_state_fleet_proof_matches(proof: &RemoteVersionStateFleetProofToken) -> bool {
+    let Some(expected_topology) = REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() else {
+        return false;
+    };
+    let state = remote_version_state_fleet_proof_slot()
+        .read()
+        .unwrap_or_else(std::sync::PoisonError::into_inner);
+    if state.topology_conflict {
+        return false;
+    }
+    state.proof.as_ref().is_some_and(|current| {
+        current.topology_fingerprint == *expected_topology
+            && current.topology_fingerprint == proof.topology_fingerprint
+            && Arc::ptr_eq(&current.peer_epochs, &proof.peer_epochs)
+            && Instant::now() < current.expires_at
+    })
+}
+
+fn remote_version_state_fleet_proof_valid_at(
+    proof: Option<&RemoteVersionStateFleetProof>,
+    expected_topology: &str,
+    now: Instant,
+) -> bool {
+    proof.is_some_and(|proof| proof.topology_fingerprint == expected_topology && now < proof.expires_at)
+}
+
+fn insert_remote_version_state_peer(peer_epochs: &mut BTreeMap<String, Uuid>, peer: String, epoch: Uuid) -> Result<()> {
+    if epoch.is_nil() || peer_epochs.values().any(|existing| *existing == epoch) || peer_epochs.insert(peer, epoch).is_some() {
+        return Err(Error::other("remote version state capability peer identity is invalid"));
+    }
+    Ok(())
+}
+
+pub fn start_remote_version_state_fleet_probe(topology_fingerprint: String) {
+    if REMOTE_VERSION_STATE_PROBE_TOPOLOGY.set(topology_fingerprint.clone()).is_err() {
+        if REMOTE_VERSION_STATE_PROBE_TOPOLOGY.get() != Some(&topology_fingerprint) {
+            let mut state = remote_version_state_fleet_proof_slot()
+                .write()
+                .unwrap_or_else(std::sync::PoisonError::into_inner);
+            state.topology_conflict = true;
+            state.proof = None;
+        }
+        return;
+    }
+
+    tokio::spawn(async move {
+        loop {
+            let result = match get_global_notification_sys() {
+                Some(notification_sys) => {
+                    match timeout(
+                        REMOTE_VERSION_STATE_PROBE_TIMEOUT,
+                        notification_sys.probe_remote_version_state_fleet(&topology_fingerprint),
+                    )
+                    .await
+                    {
+                        Ok(result) => result,
+                        Err(_) => Err(Error::other("remote version state fleet capability probe timed out")),
+                    }
+                }
+                None => Err(Error::other("remote version state fleet capability notification system is unavailable")),
+            };
+            let topology_conflict = remote_version_state_fleet_proof_slot()
+                .read()
+                .unwrap_or_else(std::sync::PoisonError::into_inner)
+                .topology_conflict;
+            if topology_conflict {
+                replace_remote_version_state_fleet_proof(None);
+            } else if let Some(err) = publish_remote_version_state_probe_result(
+                remote_version_state_fleet_proof_slot(),
+                &topology_fingerprint,
+                result,
+                Instant::now(),
+            ) {
+                debug!(error = %err, "remote version state fleet capability probe failed closed");
+            }
+            sleep(REMOTE_VERSION_STATE_PROBE_INTERVAL).await;
+        }
+    });
+}
+
 pub async fn new_global_notification_sys(eps: EndpointServerPools) -> Result<()> {
     let _ = GLOBAL_NOTIFICATION_SYS
         .set(Arc::new(NotificationSys::new(eps).await))
@@ -115,7 +292,17 @@ pub struct NotificationSys {
 
 impl NotificationSys {
     pub async fn new(eps: EndpointServerPools) -> Self {
+        let expected_remote_hosts = eps
+            .peer_grid_host_slots_sorted()
+            .into_iter()
+            .filter_map(|(peer, _, is_local)| (!is_local).then_some(peer))
+            .collect::<Vec<_>>();
         let (peer_clients, all_peer_clients, peer_topology_hosts) = PeerRestClient::new_clients_with_topology(eps).await;
+        let peer_topology_hosts = if peer_topology_hosts.is_empty() {
+            expected_remote_hosts
+        } else {
+            peer_topology_hosts
+        };
         let peer_admin_caches = (0..peer_clients.len()).map(|_| Mutex::new(PeerAdminCache::new())).collect();
         Self {
             peer_clients,
@@ -125,6 +312,24 @@ impl NotificationSys {
             tier_config_reload_workers: Default::default(),
         }
     }
+
+    async fn probe_remote_version_state_fleet(&self, topology_fingerprint: &str) -> Result<BTreeMap<String, Uuid>> {
+        if self.peer_clients.len() != self.peer_topology_hosts.len() {
+            return Err(Error::other("remote version state capability fleet membership is incomplete"));
+        }
+        let probes = self.peer_clients.iter().map(|client| async {
+            let client = client
+                .as_ref()
+                .ok_or_else(|| Error::other("remote version state capability peer is unreachable"))?;
+            client.probe_remote_version_state(topology_fingerprint.to_string()).await
+        });
+        let mut peer_epochs = BTreeMap::new();
+        for result in join_all(probes).await {
+            let (peer, epoch) = result?;
+            insert_remote_version_state_peer(&mut peer_epochs, peer, epoch)?;
+        }
+        Ok(peer_epochs)
+    }
 }
 
 pub struct NotificationPeerErr {
@@ -1344,8 +1549,11 @@ async fn run_tier_config_reload_worker<F, Fut>(
                 }
                 TierConfigReloadFinish::Pending => retry_attempt = 0,
             },
-            TierConfigReloadOutcome::Terminal(_) => match sys.finish_tier_config_reload_worker(&host) {
+            TierConfigReloadOutcome::Terminal(err) => match sys.finish_tier_config_reload_worker(&host) {
                 TierConfigReloadFinish::Completed => {
+                    // This peer keeps the previous tier configuration for good, so record
+                    // why. Dropping the error here hides the only evidence of a divergent
+                    // node behind an outcome label that cannot be acted on.
                     warn!(
                         event = EVENT_NOTIFICATION_PEER_PROPAGATION,
                         component = LOG_COMPONENT_ECSTORE,
@@ -1353,6 +1561,7 @@ async fn run_tier_config_reload_worker<F, Fut>(
                         action = "reload_transition_tier_config",
                         host,
                         outcome = "terminal",
+                        error = ?err,
                         "tier configuration reload stopped after a terminal outcome"
                     );
                     return;
@@ -1886,6 +2095,183 @@ fn aggregate_scanner_dirty_usage_acknowledgement_results(
 mod tests {
     use super::*;
 
+    #[test]
+    fn remote_version_state_fleet_proof_rejects_stale_or_mismatched_membership() {
+        let now = Instant::now();
+        let mut peer_epochs = BTreeMap::new();
+        peer_epochs.insert("peer-a".to_string(), Uuid::new_v4());
+        let proof = RemoteVersionStateFleetProof {
+            topology_fingerprint: "topology-a".to_string(),
+            peer_epochs: Arc::new(peer_epochs),
+            expires_at: now + Duration::from_secs(1),
+        };
+
+        assert!(remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-a", now));
+        assert!(!remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-b", now));
+        assert!(!remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-a", proof.expires_at));
+        assert!(!remote_version_state_fleet_proof_valid_at(None, "topology-a", now));
+    }
+
+    #[test]
+    fn remote_version_state_fleet_proof_rejects_nil_process_epoch() {
+        let mut peer_epochs = BTreeMap::new();
+
+        assert!(insert_remote_version_state_peer(&mut peer_epochs, "peer-a".to_string(), Uuid::nil()).is_err());
+        assert!(peer_epochs.is_empty());
+    }
+
+    #[test]
+    fn remote_version_state_fleet_proof_accepts_single_node_membership() {
+        let now = Instant::now();
+        let proof = RemoteVersionStateFleetProof {
+            topology_fingerprint: "topology-a".to_string(),
+            peer_epochs: Arc::new(BTreeMap::new()),
+            expires_at: now + Duration::from_secs(1),
+        };
+
+        assert!(remote_version_state_fleet_proof_valid_at(Some(&proof), "topology-a", now));
+    }
+
+    #[test]
+    fn remote_version_state_fleet_proof_token_changes_with_process_epoch() {
+        let now = Instant::now();
+        let proof = RemoteVersionStateFleetProof {
+            topology_fingerprint: "topology-a".to_string(),
+            peer_epochs: Arc::new(BTreeMap::from([("peer-a".to_string(), Uuid::new_v4())])),
+            expires_at: now + Duration::from_secs(1),
+        };
+        let captured = proof.token();
+        let restarted = RemoteVersionStateFleetProof {
+            topology_fingerprint: proof.topology_fingerprint.clone(),
+            peer_epochs: Arc::new(BTreeMap::from([("peer-a".to_string(), Uuid::new_v4())])),
+            expires_at: proof.expires_at,
+        };
+
+        assert!(captured != restarted.token());
+    }
+
+    #[test]
+    fn remote_version_state_fleet_proof_renewal_preserves_only_same_epoch_token() {
+        let slot = std::sync::RwLock::new(RemoteVersionStateFleetProofState::default());
+        let now = Instant::now();
+        let epoch = Uuid::new_v4();
+        let peers = BTreeMap::from([("peer-a".to_string(), epoch)]);
+        assert!(publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peers.clone()), now).is_none());
+        let original = slot
+            .read()
+            .expect("proof slot should not poison")
+            .proof
+            .as_ref()
+            .expect("successful probe should publish proof")
+            .token();
+
+        assert!(
+            publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peers), now + Duration::from_millis(1)).is_none()
+        );
+        let renewed = slot
+            .read()
+            .expect("proof slot should not poison")
+            .proof
+            .as_ref()
+            .expect("renewal should retain proof")
+            .token();
+        assert!(Arc::ptr_eq(&original.peer_epochs, &renewed.peer_epochs));
+
+        let restarted = BTreeMap::from([("peer-a".to_string(), Uuid::new_v4())]);
+        assert!(
+            publish_remote_version_state_probe_result(&slot, "topology-a", Ok(restarted), now + Duration::from_millis(2))
+                .is_none()
+        );
+        let replaced = slot
+            .read()
+            .expect("proof slot should not poison")
+            .proof
+            .as_ref()
+            .expect("restarted peer should publish a new proof")
+            .token();
+        assert!(!Arc::ptr_eq(&original.peer_epochs, &replaced.peer_epochs));
+    }
+
+    #[test]
+    fn remote_version_state_fleet_proof_conflict_revokes_atomic_snapshot() {
+        let now = Instant::now();
+        let mut state = RemoteVersionStateFleetProofState {
+            proof: Some(RemoteVersionStateFleetProof {
+                topology_fingerprint: "topology-a".to_string(),
+                peer_epochs: Arc::new(BTreeMap::new()),
+                expires_at: now + Duration::from_secs(1),
+            }),
+            topology_conflict: false,
+        };
+        assert!(acquire_remote_version_state_fleet_proof_from(&state, "topology-a", now).is_some());
+
+        state.topology_conflict = true;
+        assert!(acquire_remote_version_state_fleet_proof_from(&state, "topology-a", now).is_none());
+    }
+
+    #[test]
+    fn remote_version_state_fleet_probe_rejects_duplicate_member_or_process_epoch() {
+        let epoch = Uuid::new_v4();
+        let mut peer_epochs = BTreeMap::new();
+        insert_remote_version_state_peer(&mut peer_epochs, "node-a:9000".to_string(), epoch)
+            .expect("first member should be admitted");
+        assert!(insert_remote_version_state_peer(&mut peer_epochs, "node-b:9000".to_string(), epoch).is_err());
+        assert!(insert_remote_version_state_peer(&mut peer_epochs, "node-a:9000".to_string(), Uuid::new_v4()).is_err());
+        assert!(insert_remote_version_state_peer(&mut peer_epochs, "node-c:9000".to_string(), Uuid::nil()).is_err());
+    }
+
+    #[test]
+    fn remote_version_state_fleet_probe_failure_revokes_previous_proof() {
+        let slot = std::sync::RwLock::new(RemoteVersionStateFleetProofState::default());
+        let now = Instant::now();
+        let peer_epochs = BTreeMap::from([("node-a:9000".to_string(), Uuid::new_v4())]);
+        assert!(publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peer_epochs), now).is_none());
+        assert!(slot.read().expect("proof slot should not poison").proof.is_some());
+
+        assert!(
+            publish_remote_version_state_probe_result(&slot, "topology-a", Err(Error::other("peer unavailable")), now,).is_some()
+        );
+        assert!(slot.read().expect("proof slot should not poison").proof.is_none());
+
+        let peer_epochs = BTreeMap::from([("node-a:9000".to_string(), Uuid::new_v4())]);
+        assert!(publish_remote_version_state_probe_result(&slot, "topology-a", Ok(peer_epochs), now).is_none());
+        assert!(slot.read().expect("proof slot should not poison").proof.is_some());
+    }
+
+    #[tokio::test]
+    async fn remote_version_state_fleet_probe_rejects_unreachable_member() {
+        let notification_sys = NotificationSys {
+            peer_clients: vec![None],
+            all_peer_clients: vec![None, None],
+            peer_topology_hosts: vec!["peer-a".to_string()],
+            peer_admin_caches: vec![Mutex::new(PeerAdminCache::new())],
+            tier_config_reload_workers: Default::default(),
+        };
+
+        let err = notification_sys
+            .probe_remote_version_state_fleet("topology-a")
+            .await
+            .expect_err("an unreachable configured member must fail the fleet proof");
+        assert!(err.to_string().contains("unreachable"));
+    }
+
+    #[tokio::test]
+    async fn remote_version_state_fleet_probe_rejects_missing_member_slot() {
+        let notification_sys = NotificationSys {
+            peer_clients: Vec::new(),
+            all_peer_clients: vec![None],
+            peer_topology_hosts: vec!["peer-a".to_string()],
+            peer_admin_caches: Vec::new(),
+            tier_config_reload_workers: Default::default(),
+        };
+
+        let err = notification_sys
+            .probe_remote_version_state_fleet("topology-a")
+            .await
+            .expect_err("a missing configured member slot must fail the fleet proof");
+        assert!(err.to_string().contains("incomplete"));
+    }
+
     fn build_props(endpoint: &str) -> ServerProperties {
         ServerProperties {
             endpoint: endpoint.to_string(),

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

@@ -931,7 +931,10 @@ pub async fn read_transition_meta(disk_path: &Path, bucket: &str, object: &str)
         status: fi.transition_status.clone(),
         tier: fi.transition_tier.clone(),
         remote_object: fi.transitioned_objname.clone(),
-        remote_version_id: fi.transition_version_id.map(|id| id.to_string()),
+        remote_version_id: fi
+            .transition_version
+            .clone()
+            .or_else(|| fi.transition_version_id.map(|id| id.to_string())),
         free_version_count,
     })
 }

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

@@ -236,6 +236,7 @@ struct TierPublishTransition {
 
 struct PreparedTierDriver {
     tier_name: String,
+    tier_config: TierConfig,
     config_fingerprint: TierDriverFingerprint,
     backend_identity: TierDestinationId,
     exact_get_delete: bool,
@@ -1342,6 +1343,7 @@ fn tier_exact_get_delete(config: &TierConfig) -> bool {
 
 struct TierDriverGeneration {
     tier_name: Arc<str>,
+    tier_config: TierConfig,
     generation: DriverRevision,
     // Process-local only: this may reflect credential changes and must never be persisted or logged.
     config_fingerprint: TierDriverFingerprint,
@@ -1349,6 +1351,9 @@ struct TierDriverGeneration {
     backend_identity: TierDestinationId,
     exact_get_delete: bool,
     driver: SharedWarmBackend,
+    reconciler: tokio::sync::OnceCell<
+        Option<Arc<dyn crate::services::tier::warm_backend::TransitionCandidateReconciler + Send + Sync + 'static>>,
+    >,
     accepting: AtomicBool,
     active_leases: AtomicUsize,
     drained: tokio::sync::Notify,
@@ -1473,6 +1478,35 @@ impl TierOperationLease {
         self.inner.backend_identity
     }
 
+    pub(crate) async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        transaction_id: uuid::Uuid,
+    ) -> io::Result<TransitionCandidateProbe> {
+        let Some(reconciler) = self
+            .inner
+            .reconciler
+            .get_or_try_init(|| async {
+                crate::services::tier::warm_backend::new_transition_candidate_reconciler(&self.inner.tier_config)
+                    .await
+                    .map(|reconciler| reconciler.map(Arc::from))
+            })
+            .await
+            .map_err(|err| io::Error::other(err.message))?
+        else {
+            return self.inner.driver.probe_transition_candidate(object).await;
+        };
+        reconciler
+            .probe_transition_candidate_for(
+                object,
+                crate::services::tier::warm_backend::TransitionCandidateIdentity {
+                    transaction_id,
+                    destination_id: self.backend_identity(),
+                },
+            )
+            .await
+    }
+
     pub(crate) fn validate_remote_version_id(&self, remote_version_id: &str) -> io::Result<()> {
         self.inner.driver.validate_remote_version_id(remote_version_id)?;
         if !remote_version_id.is_empty() && !self.inner.exact_get_delete {
@@ -2833,6 +2867,7 @@ impl TierConfigMgr {
                     let exact_get_delete = tier_exact_get_delete(config);
                     Some(PreparedTierDriver {
                         tier_name: tier_name.to_string(),
+                        tier_config: config.clone(),
                         config_fingerprint,
                         backend_identity,
                         exact_get_delete,
@@ -2861,11 +2896,13 @@ impl TierConfigMgr {
             })?;
             let entry = Arc::new(TierDriverGeneration {
                 tier_name: Arc::from(prepared.tier_name.as_str()),
+                tier_config: prepared.tier_config.clone(),
                 generation,
                 config_fingerprint: prepared.config_fingerprint,
                 backend_identity: prepared.backend_identity,
                 exact_get_delete: prepared.exact_get_delete,
                 driver: prepared.driver.clone(),
+                reconciler: tokio::sync::OnceCell::new(),
                 accepting: AtomicBool::new(true),
                 active_leases: AtomicUsize::new(0),
                 drained: tokio::sync::Notify::new(),
@@ -3609,11 +3646,13 @@ impl TierConfigMgr {
         let driver: SharedWarmBackend = Arc::from(driver);
         let entry = Arc::new(TierDriverGeneration {
             tier_name: Arc::from(tier_name),
+            tier_config: config.clone(),
             generation,
             config_fingerprint,
             backend_identity,
             exact_get_delete,
             driver: driver.clone(),
+            reconciler: tokio::sync::OnceCell::new(),
             accepting: AtomicBool::new(true),
             active_leases: AtomicUsize::new(0),
             drained: tokio::sync::Notify::new(),
@@ -9274,7 +9313,8 @@ mod tests {
             version_id: "v1".to_string(),
             tier_name: "COLD-A".to_string(),
             backend_identity: Some(current_identity),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         };
         journal_store
             .insert_config_object(

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

@@ -73,6 +73,21 @@ pub enum TransitionCandidateProbe {
     Unsupported,
 }
 
+#[derive(Clone, Copy)]
+pub(crate) struct TransitionCandidateIdentity {
+    pub transaction_id: uuid::Uuid,
+    pub destination_id: [u8; 32],
+}
+
+#[async_trait::async_trait]
+pub(crate) trait TransitionCandidateReconciler {
+    async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        identity: TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error>;
+}
+
 #[async_trait::async_trait]
 pub trait WarmBackend {
     async fn validate(&self) -> Result<(), std::io::Error> {
@@ -189,6 +204,20 @@ pub fn build_transition_put_options(storage_class: String, mut metadata: HashMap
         metadata.remove(key);
     }
 
+    for suffix in [
+        rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+        rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
+    ] {
+        for key in [
+            rustfs_utils::http::metadata_compat::internal_key_rustfs(suffix),
+            format!("{}{}", rustfs_utils::http::metadata_compat::MINIO_INTERNAL_PREFIX, suffix),
+        ] {
+            if let Some(value) = metadata.remove(&key) {
+                metadata.insert(format!("x-amz-meta-{key}"), value);
+            }
+        }
+    }
+
     opts.user_metadata = metadata;
     opts
 }
@@ -446,6 +475,51 @@ pub async fn new_warm_backend(tier: &TierConfig, probe: bool) -> Result<WarmBack
     Ok(d)
 }
 
+pub(crate) async fn new_transition_candidate_reconciler(
+    tier: &TierConfig,
+) -> Result<Option<Box<dyn TransitionCandidateReconciler + Send + Sync + 'static>>, AdminError> {
+    let reconciler: Box<dyn TransitionCandidateReconciler + Send + Sync + 'static> = match tier.tier_type {
+        TierType::S3 => Box::new(
+            WarmBackendS3::new(tier.s3.as_ref().ok_or_else(|| ERR_TIER_INVALID_CONFIG.clone())?, &tier.name)
+                .await
+                .map_err(|err| {
+                    let mut admin_err = ERR_TIER_INVALID_CONFIG.clone();
+                    admin_err.message = err.to_string();
+                    admin_err
+                })?,
+        ),
+        TierType::MinIO => Box::new(
+            WarmBackendMinIO::new(tier.minio.as_ref().ok_or_else(|| ERR_TIER_INVALID_CONFIG.clone())?, &tier.name)
+                .await
+                .map_err(|err| {
+                    let mut admin_err = ERR_TIER_INVALID_CONFIG.clone();
+                    admin_err.message = err.to_string();
+                    admin_err
+                })?,
+        ),
+        TierType::RustFS => Box::new(
+            WarmBackendRustFS::new(tier.rustfs.as_ref().ok_or_else(|| ERR_TIER_INVALID_CONFIG.clone())?, &tier.name)
+                .await
+                .map_err(|err| {
+                    let mut admin_err = ERR_TIER_INVALID_CONFIG.clone();
+                    admin_err.message = err.to_string();
+                    admin_err
+                })?,
+        ),
+        TierType::R2 => Box::new(
+            WarmBackendR2::new(tier.r2.as_ref().ok_or_else(|| ERR_TIER_INVALID_CONFIG.clone())?, &tier.name)
+                .await
+                .map_err(|err| {
+                    let mut admin_err = ERR_TIER_INVALID_CONFIG.clone();
+                    admin_err.message = err.to_string();
+                    admin_err
+                })?,
+        ),
+        _ => return Ok(None),
+    };
+    Ok(Some(reconciler))
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
@@ -775,6 +849,37 @@ mod tests {
         assert!(!opts.user_metadata.contains_key(X_AMZ_REPLICATION_STATUS.as_str()));
     }
 
+    #[test]
+    fn build_transition_put_options_persists_both_candidate_identity_keys_as_s3_metadata() {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            "aaaaaaaa-aaaa-4aaa-8aaa-aaaaaaaaaaaa".to_string(),
+        );
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
+            "5a".repeat(32),
+        );
+
+        let opts = build_transition_put_options("COLD".to_string(), metadata);
+
+        for suffix in [
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
+        ] {
+            assert!(opts.user_metadata.contains_key(&format!(
+                "x-amz-meta-{}",
+                rustfs_utils::http::metadata_compat::internal_key_rustfs(suffix)
+            )));
+            assert!(opts.user_metadata.contains_key(&format!(
+                "x-amz-meta-{}{suffix}",
+                rustfs_utils::http::metadata_compat::MINIO_INTERNAL_PREFIX
+            )));
+        }
+    }
+
     #[test]
     fn build_transition_put_options_requests_no_checksum_and_content_md5() {
         // Regression for rustfs/rustfs#4811: transition uploads must leave the

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

@@ -19,6 +19,7 @@
 #![allow(clippy::all)]
 
 use std::collections::HashMap;
+use std::io::{Error, ErrorKind};
 use std::sync::Arc;
 
 use bytes::Bytes;
@@ -45,6 +46,19 @@ const MAX_PARTS_COUNT: i64 = 10000;
 const _MAX_PART_SIZE: i64 = 1024 * 1024 * 1024 * 5;
 const MIN_PART_SIZE: i64 = 1024 * 1024 * 128;
 
+fn parse_generation(remote_version: &str) -> Result<Option<i64>, Error> {
+    if remote_version.is_empty() {
+        return Ok(None);
+    }
+    let generation = remote_version
+        .parse::<i64>()
+        .map_err(|_| Error::new(ErrorKind::InvalidData, "GCS remote version is not a valid generation"))?;
+    if generation <= 0 {
+        return Err(Error::new(ErrorKind::InvalidData, "GCS remote version generation must be positive"));
+    }
+    Ok(Some(generation))
+}
+
 pub struct WarmBackendGCS {
     pub client: Arc<Storage>,
     pub control: Arc<StorageControl>,
@@ -105,6 +119,10 @@ impl WarmBackendGCS {
 
 #[async_trait::async_trait]
 impl WarmBackend for WarmBackendGCS {
+    fn validate_remote_version_id(&self, remote_version_id: &str) -> Result<(), std::io::Error> {
+        parse_generation(remote_version_id).map(|_| ())
+    }
+
     async fn put_with_meta(
         &self,
         object: &str,
@@ -135,6 +153,9 @@ impl WarmBackend for WarmBackendGCS {
 
     async fn get(&self, object: &str, rv: &str, opts: WarmBackendGetOpts) -> Result<ReadCloser, std::io::Error> {
         let mut req = self.client.read_object(&self.bucket, &self.get_dest(object));
+        if let Some(generation) = parse_generation(rv)? {
+            req = req.set_generation(generation);
+        }
 
         // Honor the requested byte range so Range GETs on tiered objects return the exact
         // interval instead of the whole object (matches the s3/s3sdk/rustfs warm backends).
@@ -164,13 +185,15 @@ impl WarmBackend for WarmBackendGCS {
         // gRPC v2 DeleteObject requires the bucket in resource-name form. Without this the
         // deleted tiered object was never removed from GCS (empty impl returned Ok), leaking
         // remote data forever.
-        self.control
+        let mut req = self
+            .control
             .delete_object()
             .set_bucket(format!("projects/_/buckets/{}", self.bucket))
-            .set_object(self.get_dest(object))
-            .send()
-            .await
-            .map_err(|e| std::io::Error::other(e.to_string()))?;
+            .set_object(self.get_dest(object));
+        if let Some(generation) = parse_generation(rv)? {
+            req = req.set_generation(generation);
+        }
+        req.send().await.map_err(|e| std::io::Error::other(e.to_string()))?;
         Ok(())
     }
 
@@ -191,6 +214,30 @@ impl WarmBackend for WarmBackendGCS {
     }
 }
 
+#[cfg(test)]
+mod tests {
+    use super::parse_generation;
+    use std::io::ErrorKind;
+
+    #[test]
+    fn generation_parser_preserves_exact_numeric_versions() {
+        assert_eq!(parse_generation("").expect("empty generation means no version condition"), None);
+        assert_eq!(parse_generation("1").expect("minimum generation should parse"), Some(1));
+        assert_eq!(
+            parse_generation(&i64::MAX.to_string()).expect("maximum generation should parse"),
+            Some(i64::MAX)
+        );
+    }
+
+    #[test]
+    fn generation_parser_rejects_unknown_or_non_positive_versions() {
+        for value in ["unknown", "1.0", "-1", "0", "9223372036854775808"] {
+            let err = parse_generation(value).expect_err("unknown generation must fail closed");
+            assert_eq!(err.kind(), ErrorKind::InvalidData, "{value}");
+        }
+    }
+}
+
 /*fn gcs_to_object_error(err: Error, params: Vec<String>) -> Option<Error> {
   if err == nil {
     return nil

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

@@ -135,6 +135,20 @@ impl WarmBackend for WarmBackendMinIO {
     }
 }
 
+#[async_trait::async_trait]
+impl crate::services::tier::warm_backend::TransitionCandidateReconciler for WarmBackendMinIO {
+    async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        identity: crate::services::tier::warm_backend::TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        crate::services::tier::warm_backend::TransitionCandidateReconciler::probe_transition_candidate_for(
+            &self.0, object, identity,
+        )
+        .await
+    }
+}
+
 fn optimal_part_size(object_size: i64) -> Result<i64, std::io::Error> {
     let mut object_size = object_size;
     if object_size == -1 {

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

@@ -135,6 +135,20 @@ impl WarmBackend for WarmBackendR2 {
     }
 }
 
+#[async_trait::async_trait]
+impl crate::services::tier::warm_backend::TransitionCandidateReconciler for WarmBackendR2 {
+    async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        identity: crate::services::tier::warm_backend::TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        crate::services::tier::warm_backend::TransitionCandidateReconciler::probe_transition_candidate_for(
+            &self.0, object, identity,
+        )
+        .await
+    }
+}
+
 fn optimal_part_size(object_size: i64) -> Result<i64, std::io::Error> {
     let mut object_size = object_size;
     if object_size == -1 {

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

@@ -132,6 +132,20 @@ impl WarmBackend for WarmBackendRustFS {
     }
 }
 
+#[async_trait::async_trait]
+impl crate::services::tier::warm_backend::TransitionCandidateReconciler for WarmBackendRustFS {
+    async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        identity: crate::services::tier::warm_backend::TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        crate::services::tier::warm_backend::TransitionCandidateReconciler::probe_transition_candidate_for(
+            &self.0, object, identity,
+        )
+        .await
+    }
+}
+
 fn optimal_part_size(object_size: i64) -> Result<i64, std::io::Error> {
     let mut object_size = object_size;
     if object_size == -1 {

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

@@ -29,6 +29,7 @@ use crate::client::{
     api_remove::{RemoveObjectOptions, RemoveObjectResult},
     api_s3_datatypes::ListVersionsResult,
     credentials::{Credentials, SignatureType, Static, Value},
+    provider_versions::validate_remote_version_id,
     transition_api::{BucketLookupType, Options, TransitionClient, TransitionCore},
     transition_api::{ReadCloser, ReaderImpl},
 };
@@ -36,7 +37,10 @@ use crate::error::ErrorResponse;
 use crate::error::error_resp_to_object_err;
 use crate::services::tier::{
     tier_config::TierS3,
-    warm_backend::{TransitionCandidateProbe, WarmBackend, WarmBackendGetOpts, build_transition_put_options},
+    warm_backend::{
+        TransitionCandidateIdentity, TransitionCandidateProbe, TransitionCandidateReconciler, WarmBackend, WarmBackendGetOpts,
+        build_transition_put_options,
+    },
 };
 use http::HeaderMap;
 use rustfs_utils::egress::validate_outbound_url;
@@ -189,44 +193,188 @@ impl WarmBackendS3 {
         remote_bucket_versioning_from_status(config.status.as_ref().map(|status| status.as_str()))
     }
 
-    async fn list_transition_candidate_versions(&self, object: &str) -> Result<ListVersionsResult, std::io::Error> {
+    async fn probe_transition_candidate_versions(
+        &self,
+        object: &str,
+        bucket_versioning: RemoteBucketVersioning,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        let remote_object = self.get_dest(object);
         let mut opts = ListObjectsOptions::default();
-        opts.set("prefix", &self.get_dest(object));
-        opts.set("max-keys", "2");
-        self.client.list_object_versions_query(&self.bucket, &opts, "", "", "").await
+        opts.set("prefix", &remote_object);
+        opts.set("max-keys", "1000");
+
+        let mut key_marker = String::new();
+        let mut version_id_marker = String::new();
+        let mut candidates = TransitionCandidateVersions::default();
+        loop {
+            let versions = self
+                .client
+                .list_object_versions_query(&self.bucket, &opts, &key_marker, &version_id_marker, "")
+                .await?;
+            candidates.extend(&remote_object, &versions);
+            if candidates.is_ambiguous() {
+                return Ok(TransitionCandidateProbe::Ambiguous);
+            }
+            if !versions.is_truncated {
+                return classify_transition_candidates(candidates, bucket_versioning);
+            }
+
+            advance_version_markers(&mut key_marker, &mut version_id_marker, &versions)?;
+        }
+    }
+
+    async fn probe_transition_candidate_identity(
+        &self,
+        object: &str,
+        identity: TransitionCandidateIdentity,
+        bucket_versioning: RemoteBucketVersioning,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        let remote_object = self.get_dest(object);
+        let mut opts = ListObjectsOptions::default();
+        opts.set("prefix", &remote_object);
+        opts.set("max-keys", "1000");
+        let mut key_marker = String::new();
+        let mut version_id_marker = String::new();
+        let mut matched_version = None;
+        let mut saw_unproven_candidate = false;
+
+        loop {
+            let versions = self
+                .client
+                .list_object_versions_query(&self.bucket, &opts, &key_marker, &version_id_marker, "")
+                .await?;
+            for version in versions.versions.iter().filter(|version| version.key == remote_object) {
+                let mut stat_opts = GetObjectOptions::default();
+                stat_opts.version_id.clone_from(&version.version_id);
+                let info = self.client.stat_object(&self.bucket, &remote_object, &stat_opts).await?;
+                let mut metadata = info.user_metadata;
+                for (name, value) in &info.metadata {
+                    if (name
+                        .as_str()
+                        .starts_with(rustfs_utils::http::metadata_compat::RUSTFS_INTERNAL_PREFIX)
+                        || name
+                            .as_str()
+                            .starts_with(rustfs_utils::http::metadata_compat::MINIO_INTERNAL_PREFIX))
+                        && let Ok(value) = value.to_str()
+                    {
+                        metadata.insert(name.as_str().to_string(), value.to_string());
+                    }
+                }
+                if transition_candidate_metadata_matches(&metadata, identity)? {
+                    if matched_version.is_some() {
+                        return Ok(TransitionCandidateProbe::Ambiguous);
+                    }
+                    matched_version = Some(version.version_id.clone());
+                } else {
+                    saw_unproven_candidate = true;
+                }
+            }
+            if !versions.is_truncated {
+                if matched_version.is_none() && saw_unproven_candidate {
+                    return Ok(TransitionCandidateProbe::Unsupported);
+                }
+                let candidates = TransitionCandidateVersions {
+                    version_id: matched_version,
+                    ambiguous: false,
+                };
+                return classify_transition_candidates(candidates, bucket_versioning);
+            }
+            advance_version_markers(&mut key_marker, &mut version_id_marker, &versions)?;
+        }
     }
 }
 
-fn classify_transition_candidate_versions(
-    remote_object: &str,
-    bucket_versioning: RemoteBucketVersioning,
-    versions: &ListVersionsResult,
-) -> TransitionCandidateProbe {
-    if versions.is_truncated {
-        return TransitionCandidateProbe::Ambiguous;
-    }
-
-    if versions.delete_markers.iter().any(|marker| marker.key == remote_object) {
-        return TransitionCandidateProbe::Ambiguous;
-    }
-
-    let mut exact_versions = versions.versions.iter().filter(|version| version.key == remote_object);
-    let Some(version) = exact_versions.next() else {
-        return TransitionCandidateProbe::Missing;
+fn transition_candidate_metadata_matches(
+    metadata: &HashMap<String, String>,
+    identity: TransitionCandidateIdentity,
+) -> Result<bool, std::io::Error> {
+    use rustfs_utils::http::metadata_compat::{
+        SUFFIX_TRANSITION_TIER_DESTINATION_ID, SUFFIX_TRANSITION_TRANSACTION_ID, contains_key_str, get_consistent_str,
     };
-    if exact_versions.next().is_some() {
-        return TransitionCandidateProbe::Ambiguous;
+
+    let transaction_id = get_consistent_str(metadata, SUFFIX_TRANSITION_TRANSACTION_ID);
+    let destination_id = get_consistent_str(metadata, SUFFIX_TRANSITION_TIER_DESTINATION_ID);
+    if transaction_id.is_none() || destination_id.is_none() {
+        if contains_key_str(metadata, SUFFIX_TRANSITION_TRANSACTION_ID)
+            || contains_key_str(metadata, SUFFIX_TRANSITION_TIER_DESTINATION_ID)
+        {
+            return Err(std::io::Error::new(
+                std::io::ErrorKind::InvalidData,
+                "transition candidate identity metadata is empty or conflicting",
+            ));
+        }
+        return Ok(false);
+    }
+    let expected_transaction_id = identity.transaction_id.to_string();
+    let expected_destination_id = rustfs_utils::crypto::hex(identity.destination_id);
+    Ok(transaction_id == Some(expected_transaction_id.as_str()) && destination_id == Some(expected_destination_id.as_str()))
+}
+
+fn classify_transition_candidates(
+    candidates: TransitionCandidateVersions,
+    bucket_versioning: RemoteBucketVersioning,
+) -> Result<TransitionCandidateProbe, std::io::Error> {
+    let probe = candidates.classify(bucket_versioning);
+    if let TransitionCandidateProbe::VersionedPresent(version_id) = &probe {
+        validate_remote_version_id(version_id)?;
+    }
+    Ok(probe)
+}
+
+fn advance_version_markers(
+    key_marker: &mut String,
+    version_id_marker: &mut String,
+    versions: &ListVersionsResult,
+) -> Result<(), std::io::Error> {
+    let next_markers = (&versions.next_key_marker, &versions.next_version_id_marker);
+    if next_markers == (&*key_marker, &*version_id_marker) {
+        return Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "ListObjectVersions pagination markers did not advance",
+        ));
+    }
+    key_marker.clone_from(&versions.next_key_marker);
+    version_id_marker.clone_from(&versions.next_version_id_marker);
+    Ok(())
+}
+
+#[derive(Default)]
+struct TransitionCandidateVersions {
+    version_id: Option<String>,
+    ambiguous: bool,
+}
+
+impl TransitionCandidateVersions {
+    fn extend(&mut self, remote_object: &str, versions: &ListVersionsResult) {
+        for version in versions.versions.iter().filter(|version| version.key == remote_object) {
+            if self.version_id.is_some() {
+                self.ambiguous = true;
+                return;
+            }
+            self.version_id = Some(version.version_id.clone());
+        }
     }
 
-    match bucket_versioning {
-        RemoteBucketVersioning::Disabled => TransitionCandidateProbe::UnversionedPresent,
-        RemoteBucketVersioning::Suspended if version.version_id == "null" => {
-            TransitionCandidateProbe::VersionedPresent(version.version_id.clone())
+    fn is_ambiguous(&self) -> bool {
+        self.ambiguous
+    }
+
+    fn classify(self, bucket_versioning: RemoteBucketVersioning) -> TransitionCandidateProbe {
+        if self.ambiguous {
+            return TransitionCandidateProbe::Ambiguous;
         }
-        RemoteBucketVersioning::Suspended | RemoteBucketVersioning::Enabled if !version.version_id.is_empty() => {
-            TransitionCandidateProbe::VersionedPresent(version.version_id.clone())
+        let Some(version_id) = self.version_id else {
+            return TransitionCandidateProbe::Missing;
+        };
+
+        match bucket_versioning {
+            RemoteBucketVersioning::Disabled => TransitionCandidateProbe::UnversionedPresent,
+            RemoteBucketVersioning::Suspended if version_id == "null" => TransitionCandidateProbe::VersionedPresent(version_id),
+            RemoteBucketVersioning::Suspended | RemoteBucketVersioning::Enabled if !version_id.is_empty() => {
+                TransitionCandidateProbe::VersionedPresent(version_id)
+            }
+            RemoteBucketVersioning::Suspended | RemoteBucketVersioning::Enabled => TransitionCandidateProbe::Ambiguous,
         }
-        RemoteBucketVersioning::Suspended | RemoteBucketVersioning::Enabled => TransitionCandidateProbe::Ambiguous,
     }
 }
 
@@ -275,74 +423,163 @@ mod tests {
         }
     }
 
+    fn classify_pages(bucket_versioning: RemoteBucketVersioning, pages: &[ListVersionsResult]) -> TransitionCandidateProbe {
+        let mut candidates = TransitionCandidateVersions::default();
+        for page in pages {
+            candidates.extend("archive/object", page);
+        }
+        candidates.classify(bucket_versioning)
+    }
+
+    fn candidate_identity() -> TransitionCandidateIdentity {
+        TransitionCandidateIdentity {
+            transaction_id: uuid::Uuid::parse_str("aaaaaaaa-aaaa-4aaa-8aaa-aaaaaaaaaaaa").unwrap(),
+            destination_id: [0x5a; 32],
+        }
+    }
+
+    fn candidate_metadata(identity: TransitionCandidateIdentity) -> HashMap<String, String> {
+        let mut metadata = HashMap::new();
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            identity.transaction_id.to_string(),
+        );
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut metadata,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
+            rustfs_utils::crypto::hex(identity.destination_id),
+        );
+        metadata
+    }
+
+    #[test]
+    fn transition_candidate_identity_requires_exact_compatible_metadata() {
+        let identity = candidate_identity();
+        let metadata = candidate_metadata(identity);
+        assert!(transition_candidate_metadata_matches(&metadata, identity).unwrap());
+
+        let mut adjacent = metadata;
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut adjacent,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            uuid::Uuid::parse_str("bbbbbbbb-bbbb-4bbb-8bbb-bbbbbbbbbbbb")
+                .unwrap()
+                .to_string(),
+        );
+        assert!(!transition_candidate_metadata_matches(&adjacent, identity).unwrap());
+
+        assert!(!transition_candidate_metadata_matches(&HashMap::new(), identity).unwrap());
+    }
+
+    #[test]
+    fn transition_candidate_identity_rejects_conflicting_compatibility_keys() {
+        let identity = candidate_identity();
+        let mut metadata = candidate_metadata(identity);
+        metadata.insert(
+            rustfs_utils::http::metadata_compat::internal_key_rustfs(
+                rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            ),
+            uuid::Uuid::new_v4().to_string(),
+        );
+        assert!(transition_candidate_metadata_matches(&metadata, identity).is_err());
+    }
+
     #[test]
     fn transition_candidate_probe_classifier_is_fail_closed() {
         assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
-                RemoteBucketVersioning::Disabled,
-                &list_versions(&[], &[], false),
-            ),
+            classify_pages(RemoteBucketVersioning::Disabled, &[list_versions(&[], &[], false)],),
             TransitionCandidateProbe::Missing
         );
         assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
-                RemoteBucketVersioning::Disabled,
-                &list_versions(&[("archive/object", "")], &[], false),
-            ),
+            classify_pages(RemoteBucketVersioning::Disabled, &[list_versions(&[("archive/object", "")], &[], false)],),
             TransitionCandidateProbe::UnversionedPresent
         );
         assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
+            classify_pages(
                 RemoteBucketVersioning::Enabled,
-                &list_versions(&[("archive/object", "version-a")], &[], false),
+                &[list_versions(&[("archive/object", "version-a")], &[], false)],
             ),
             TransitionCandidateProbe::VersionedPresent("version-a".to_string())
         );
         assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
+            classify_pages(
                 RemoteBucketVersioning::Suspended,
-                &list_versions(&[("archive/object", "null")], &[], false),
+                &[list_versions(&[("archive/object", "null")], &[], false)],
             ),
             TransitionCandidateProbe::VersionedPresent("null".to_string())
         );
         assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
-                RemoteBucketVersioning::Enabled,
-                &list_versions(&[("archive/object", "")], &[], false),
-            ),
+            classify_pages(RemoteBucketVersioning::Enabled, &[list_versions(&[("archive/object", "")], &[], false)],),
             TransitionCandidateProbe::Ambiguous
         );
         assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
+            classify_pages(
                 RemoteBucketVersioning::Enabled,
-                &list_versions(&[("archive/object", "version-a"), ("archive/object", "version-b")], &[], false),
+                &[list_versions(
+                    &[("archive/object", "version-a"), ("archive/object", "version-b")],
+                    &[],
+                    false,
+                )],
             ),
             TransitionCandidateProbe::Ambiguous
         );
+    }
+
+    #[test]
+    fn transition_candidate_probe_reconciles_all_pages_and_ignores_delete_markers() {
+        assert_eq!(
+            classify_pages(
+                RemoteBucketVersioning::Enabled,
+                &[
+                    list_versions(&[], &[("archive/object", "marker-a")], true),
+                    list_versions(&[("archive/object", "version-a"), ("archive/object-adjacent", "unrelated"),], &[], false,),
+                ],
+            ),
+            TransitionCandidateProbe::VersionedPresent("version-a".to_string())
+        );
         assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
+            classify_pages(
                 RemoteBucketVersioning::Enabled,
-                &list_versions(&[("archive/object", "version-a")], &[("archive/object", "marker-a")], false),
-            ),
-            TransitionCandidateProbe::Ambiguous
-        );
-        assert_eq!(
-            classify_transition_candidate_versions(
-                "archive/object",
-                RemoteBucketVersioning::Enabled,
-                &list_versions(&[("archive/object", "version-a")], &[], true),
+                &[
+                    list_versions(&[("archive/object", "version-a")], &[], true),
+                    list_versions(&[("archive/object", "version-b")], &[], false),
+                ],
             ),
             TransitionCandidateProbe::Ambiguous
         );
     }
 
+    #[test]
+    fn transition_candidate_pagination_advances_both_markers() {
+        let mut key_marker = "old-key".to_string();
+        let mut version_id_marker = "old-version".to_string();
+        let page = ListVersionsResult {
+            next_key_marker: "next-key".to_string(),
+            next_version_id_marker: "next-version".to_string(),
+            ..Default::default()
+        };
+
+        advance_version_markers(&mut key_marker, &mut version_id_marker, &page)
+            .expect("new ListObjectVersions markers should advance pagination");
+        assert_eq!(key_marker, "next-key");
+        assert_eq!(version_id_marker, "next-version");
+
+        let err = advance_version_markers(&mut key_marker, &mut version_id_marker, &page)
+            .expect_err("repeated ListObjectVersions markers must fail closed");
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+    }
+
+    #[test]
+    fn transition_candidate_probe_rejects_untrusted_version_ids() {
+        let mut candidates = TransitionCandidateVersions::default();
+        candidates.extend("archive/object", &list_versions(&[("archive/object", "version\ninjection")], &[], false));
+
+        let err = classify_transition_candidates(candidates, RemoteBucketVersioning::Enabled)
+            .expect_err("control characters in listed version IDs must fail closed");
+        assert_eq!(err.kind(), std::io::ErrorKind::InvalidData);
+    }
+
     #[test]
     fn remote_bucket_versioning_status_parser_fails_closed() {
         assert_eq!(
@@ -397,12 +634,7 @@ impl WarmBackend for WarmBackendS3 {
 
     async fn probe_transition_candidate(&self, object: &str) -> Result<TransitionCandidateProbe, std::io::Error> {
         let bucket_versioning = self.remote_bucket_versioning().await?;
-        let versions = self.list_transition_candidate_versions(object).await?;
-        Ok(classify_transition_candidate_versions(
-            &self.get_dest(object),
-            bucket_versioning,
-            &versions,
-        ))
+        self.probe_transition_candidate_versions(object, bucket_versioning).await
     }
 
     async fn in_use(&self) -> Result<bool, std::io::Error> {
@@ -414,3 +646,16 @@ impl WarmBackend for WarmBackendS3 {
         Ok(result.common_prefixes.len() > 0 || result.contents.len() > 0)
     }
 }
+
+#[async_trait::async_trait]
+impl TransitionCandidateReconciler for WarmBackendS3 {
+    async fn probe_transition_candidate_for(
+        &self,
+        object: &str,
+        identity: TransitionCandidateIdentity,
+    ) -> Result<TransitionCandidateProbe, std::io::Error> {
+        let bucket_versioning = self.remote_bucket_versioning().await?;
+        self.probe_transition_candidate_identity(object, identity, bucket_versioning)
+            .await
+    }
+}

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

@@ -46,9 +46,10 @@ use crate::diagnostics::get::{
     GetObjectFailureReason, classify_disk_error, get_stage_timer_if_enabled, record_get_object_pipeline_failure,
     record_get_object_pipeline_failure_for_path, record_get_stage_duration_if_enabled,
 };
+use crate::disk::local::DELETE_DATA_DIR_MARKER_PREFIX;
 use crate::disk::{
-    OldCurrentSize, PART_TRANSACTION_NEW_META, PART_TRANSACTION_OLD_META, PART_TRANSACTION_ROLLBACK, PartTransactionAction,
-    part_transaction_path,
+    DataDirDeleteStatus, OldCurrentSize, PART_TRANSACTION_NEW_META, PART_TRANSACTION_OLD_META, PART_TRANSACTION_ROLLBACK,
+    PartTransactionAction, part_transaction_path,
 };
 use crate::erasure::coding::BitrotReader;
 use crate::io_support::bitrot::ShardReader;
@@ -547,6 +548,8 @@ pub(in crate::set_disk) fn metadata_early_stop_candidate_matches(left: &FileInfo
         && left.transitioned_objname == right.transitioned_objname
         && left.transition_tier == right.transition_tier
         && left.transition_version_id == right.transition_version_id
+        && left.transition_version == right.transition_version
+        && left.transition_version_state == right.transition_version_state
         && left.expire_restored == right.expire_restored
         && left.size == right.size
         && left.mod_time == right.mod_time
@@ -2985,29 +2988,48 @@ impl SetDisks {
                 Self::rename_fanout_barrier(&object_for_fault, idx, rename_fanout_barrier_phase::CLEANUP).await;
 
                 if let Some(err) = Self::cleanup_injected_error(&object_for_fault, idx) {
-                    return Some(err);
+                    return (false, Some(err));
                 }
                 if let Some(disk) = disk {
-                    disk.delete(
-                        &bucket,
-                        &file_path,
-                        DeleteOptions {
-                            recursive: true,
-                            ..Default::default()
-                        },
-                    )
-                    .await
-                    .err()
+                    match disk
+                        .delete_data_dir(
+                            &bucket,
+                            &file_path,
+                            DeleteOptions {
+                                recursive: true,
+                                ..Default::default()
+                            },
+                        )
+                        .await
+                    {
+                        Ok(DataDirDeleteStatus::Deleted) => (false, None),
+                        Ok(DataDirDeleteStatus::Deferred) => (true, None),
+                        Err(err) => (false, Some(err)),
+                    }
                 } else {
                     // `None` slot: ignored placeholder. It is not `attempted`, so
                     // classification excludes it from residue regardless.
-                    Some(DiskError::DiskNotFound)
+                    (false, Some(DiskError::DiskNotFound))
                 }
             })
         });
-        let errs: Vec<Option<DiskError>> = join_all(futures).await.into_iter().map(map_cleanup_join_result).collect();
+        let mut deferred = 0usize;
+        let errs: Vec<Option<DiskError>> = join_all(futures)
+            .await
+            .into_iter()
+            .map(|result| match result {
+                Ok((was_deferred, err)) => {
+                    deferred += usize::from(was_deferred);
+                    err
+                }
+                Err(join_err) => Some(DiskError::other(format!("old data dir cleanup task failed: {join_err}"))),
+            })
+            .collect();
 
-        classify_old_data_dir_cleanup(&errs, &attempted, write_quorum)
+        let mut cleanup = classify_old_data_dir_cleanup(&errs, &attempted, write_quorum);
+        cleanup.deferred = deferred;
+        cleanup.reclaimed = cleanup.reclaimed.saturating_sub(deferred);
+        cleanup
     }
 
     /// Test-only fault-injection seam for the old-data-dir cleanup path
@@ -3098,6 +3120,20 @@ impl SetDisks {
 
         rustfs_io_metrics::record_old_data_dir_cleanup(c.attempted, c.reclaimed, c.unreclaimed_disks.len(), c.below_quorum);
 
+        if c.deferred > 0 {
+            debug!(
+                event = EVENT_SET_DISK_WRITE,
+                component = LOG_COMPONENT_ECSTORE,
+                subsystem = LOG_SUBSYSTEM_SET_DISK,
+                bucket = %bucket,
+                object = %object,
+                old_data_dir = %old_dir,
+                deferred = c.deferred,
+                state = "old_data_cleanup_deferred",
+                "Old data directory cleanup deferred for active snapshot leases"
+            );
+        }
+
         if actions.warn {
             warn!(
                 component = LOG_COMPONENT_ECSTORE,
@@ -3917,9 +3953,9 @@ impl SetDisks {
     /// * The set of referenced data dirs is the UNION of `get_data_dirs()` across
     ///   every online disk's `xl.meta`, so a dir named by *any* replica is kept.
     /// * If a disk holds the object directory but its `xl.meta` is missing or
-    ///   unparsable, the object is treated as degraded and NOTHING is removed —
-    ///   the unreadable copy could be the only one naming a live data dir, and a
-    ///   heal must run first.
+    ///   unparsable, the object is treated as degraded and unmarked data dirs are
+    ///   never removed. A data dir carrying a committed delete-transaction marker
+    ///   remains reclaimable after a downgrade/re-upgrade cleanup interruption.
     /// * Only subdirectories whose names parse as a UUID are ever considered;
     ///   removal is non-recursive-safe via a recursive delete of the full stray
     ///   data-dir path only.
@@ -3934,7 +3970,7 @@ impl SetDisks {
         // physical UUID subdirectories present on each disk. Abort on any degraded
         // copy so a healable object is never stripped of a referenced data dir.
         let mut referenced: HashSet<Uuid> = HashSet::new();
-        let mut per_disk_dirs: Vec<(usize, Vec<Uuid>)> = Vec::new();
+        let mut per_disk_dirs: Vec<(usize, Vec<(Uuid, bool)>)> = Vec::new();
         let mut healthy_metas = 0usize;
 
         for (i, disk) in disks.iter().enumerate() {
@@ -3970,6 +4006,22 @@ impl SetDisks {
                         // to the orphan-dir / dangling-object heal paths.
                         continue;
                     }
+                    let mut committed = Vec::with_capacity(physical.len());
+                    for dir in physical {
+                        let data_dir = format!("{object}/{dir}");
+                        let committed_delete = disk.list_dir("", bucket, &data_dir, 0).await.is_ok_and(|entries| {
+                            entries.iter().any(|entry| {
+                                entry
+                                    .strip_prefix(DELETE_DATA_DIR_MARKER_PREFIX)
+                                    .is_some_and(|transaction| Uuid::parse_str(transaction).is_ok())
+                            })
+                        });
+                        committed.push((dir, committed_delete));
+                    }
+                    if committed.iter().all(|(_, committed_delete)| *committed_delete) {
+                        per_disk_dirs.push((i, committed));
+                        continue;
+                    }
                     warn!(
                         target: "rustfs_ecstore::set_disk",
                         bucket, object,
@@ -4006,22 +4058,16 @@ impl SetDisks {
 
             healthy_metas += 1;
             if !physical.is_empty() {
-                per_disk_dirs.push((i, physical));
+                per_disk_dirs.push((i, physical.into_iter().map(|dir| (dir, false)).collect()));
             }
         }
 
-        // No healthy metadata anywhere: this is not a live object, so surplus dirs
-        // (if any) belong to the dangling-object heal path, not here.
-        if healthy_metas == 0 {
-            return Ok(0);
-        }
-
         // Phase 2: delete every physical data dir not referenced by the union.
         let mut removed = 0usize;
         for (i, physical) in per_disk_dirs {
             let Some(disk) = disks[i].as_ref() else { continue };
-            for dir in physical {
-                if referenced.contains(&dir) {
+            for (dir, committed_delete) in physical {
+                if referenced.contains(&dir) || (healthy_metas == 0 && !committed_delete) {
                     continue;
                 }
                 let stray = format!("{object}/{dir}");
@@ -4129,6 +4175,9 @@ pub(in crate::set_disk) struct OldDataDirCleanup {
     /// Number of attempted disks that returned `Ok` or a not-found variant
     /// (a missing dir == already reclaimed).
     pub reclaimed: usize,
+    /// Number of attempted disks that retained the directory for an active
+    /// snapshot lease and registered it for deletion after the final release.
+    pub deferred: usize,
     /// Indices of attempted disks that failed with a non-ignored, non-not-found
     /// error (including task panic/cancel). This is the residue that actually
     /// leaks and drives the leak metric + heal enqueue.
@@ -4191,6 +4240,7 @@ fn classify_old_data_dir_cleanup(errs: &[Option<DiskError>], attempted: &[bool],
     OldDataDirCleanup {
         attempted: attempted_count,
         reclaimed,
+        deferred: 0,
         unreclaimed_disks,
         below_quorum,
     }
@@ -5131,6 +5181,40 @@ mod tests {
         drop((disk1, disk2));
     }
 
+    #[tokio::test]
+    async fn commit_cleanup_reports_and_releases_deferred_snapshot_data_dirs() {
+        let bucket = "cleanup-lease-bucket";
+        let object = "cleanup-lease-object";
+        let old_data_dir = "11111111-1111-1111-1111-111111111111";
+        let committed_data_dir = "22222222-2222-2222-2222-222222222222";
+        let data_dir_path = format!("{object}/{old_data_dir}");
+        let shard_path = format!("{data_dir_path}/part.1");
+        let (_dir1, disk1) = read_multiple_test_disk(bucket, &[(&shard_path, b"one".as_slice())]).await;
+        let set = io_primitives_test_set(vec![Some(disk1.clone())], 0).await;
+        let lease = disk1
+            .acquire_snapshot_lease(bucket, &data_dir_path)
+            .await
+            .expect("snapshot lease should be acquired before cleanup");
+
+        let cleanup = set
+            .commit_rename_data_dir(&[Some(disk1.clone())], bucket, object, old_data_dir, committed_data_dir, 1)
+            .await;
+        assert_eq!(cleanup.attempted, 1);
+        assert_eq!(cleanup.reclaimed, 0);
+        assert_eq!(cleanup.deferred, 1);
+        assert!(cleanup.unreclaimed_disks.is_empty());
+        disk1
+            .read_all(bucket, &shard_path)
+            .await
+            .expect("deferred cleanup must leave later shard opens available");
+
+        disk1
+            .release_snapshot_lease(bucket, &data_dir_path, lease)
+            .await
+            .expect("final lease release should reclaim the old data directory");
+        assert!(matches!(disk1.read_all(bucket, &shard_path).await, Err(DiskError::FileNotFound)));
+    }
+
     /// Isolation guard: an armed barrier / observed object only affects its own
     /// object. A fan-out for a different (unobserved, unarmed) object must not be
     /// paused and must not accrue any tracked task count — so concurrent tests

crates/ecstore/src/set_disk/metadata.rs

@@ -555,7 +555,7 @@ impl SetDisks {
         }
     }
 
-    fn file_info_quorum_hash(meta: &FileInfo) -> [u8; 32] {
+    pub(super) fn file_info_quorum_hash(meta: &FileInfo) -> [u8; 32] {
         let mut hasher = Sha256::new();
         Self::update_file_info_quorum_hash(&mut hasher, meta);
         let digest = hasher.finalize();
@@ -578,6 +578,13 @@ impl SetDisks {
         Self::update_hash_str(hasher, &meta.transition_tier);
         Self::update_hash_str(hasher, &meta.transitioned_objname);
         Self::update_hash_optional_uuid(hasher, meta.transition_version_id);
+        Self::update_hash_optional_str(hasher, meta.transition_version.as_deref());
+        hasher.update([match meta.transition_version_state {
+            rustfs_filemeta::TransitionVersionState::Unknown => 0,
+            rustfs_filemeta::TransitionVersionState::KnownDisabled => 1,
+            rustfs_filemeta::TransitionVersionState::SuspendedNull => 2,
+            rustfs_filemeta::TransitionVersionState::Exact => 3,
+        }]);
         Self::update_hash_optional_u32(hasher, meta.mode);
         Self::update_hash_optional_u64(hasher, meta.written_by_version);
 

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

@@ -92,6 +92,69 @@ struct PartFailureSummary {
     bitrot_failure: bool,
 }
 
+#[derive(Clone)]
+struct RecoverableMetaCandidate {
+    identity: [u8; 32],
+    file_info: FileInfo,
+    data_count: usize,
+    local_payload: bool,
+}
+
+#[derive(Clone, Copy, Debug, PartialEq, Eq)]
+enum DanglingDeleteSafety {
+    UnsafeToDelete,
+    NoRecoverableCandidate,
+}
+
+#[cfg(test)]
+struct DanglingCheckPartsFailure {
+    key: DanglingCheckPartsFailureKey,
+}
+
+#[cfg(test)]
+type DanglingCheckPartsFailureKey = (String, String, usize);
+
+#[cfg(test)]
+type DanglingCheckPartsFailures = HashMap<DanglingCheckPartsFailureKey, DiskError>;
+
+#[cfg(test)]
+fn dangling_check_parts_failures() -> &'static std::sync::Mutex<DanglingCheckPartsFailures> {
+    static FAILURES: std::sync::OnceLock<std::sync::Mutex<DanglingCheckPartsFailures>> = std::sync::OnceLock::new();
+    FAILURES.get_or_init(|| std::sync::Mutex::new(HashMap::new()))
+}
+
+#[cfg(test)]
+impl DanglingCheckPartsFailure {
+    fn install(bucket: &str, object: &str, disk_index: usize, error: DiskError) -> Self {
+        let key = (bucket.to_string(), object.to_string(), disk_index);
+        let previous = dangling_check_parts_failures()
+            .lock()
+            .expect("dangling check-parts failure registry should not poison")
+            .insert(key.clone(), error);
+        assert!(previous.is_none(), "dangling check-parts failure already installed");
+        Self { key }
+    }
+}
+
+#[cfg(test)]
+impl Drop for DanglingCheckPartsFailure {
+    fn drop(&mut self) {
+        dangling_check_parts_failures()
+            .lock()
+            .expect("dangling check-parts failure registry should not poison")
+            .remove(&self.key);
+    }
+}
+
+#[cfg(test)]
+fn injected_dangling_check_parts_error(bucket: &str, object: &str, disk_index: usize) -> Option<DiskError> {
+    dangling_check_parts_failures()
+        .lock()
+        .expect("dangling check-parts failure registry should not poison")
+        .get(&(bucket.to_string(), object.to_string(), disk_index))
+        .cloned()
+}
+
 fn first_unhealthy_part_summary(
     data_errs_by_part: &HashMap<usize, Vec<usize>>,
     parts: &[ObjectPartInfo],
@@ -125,39 +188,17 @@ impl SetDisks {
         version_id: &str,
         opts: &HealOpts,
     ) -> disk::error::Result<(HealResultItem, Option<DiskError>)> {
-        // `allow_meta_regen` is true on the first pass: a version whose data shards
-        // physically survive (>= data_blocks) but whose xl.meta fell below
-        // read-quorum is RESCUED (missing xl.meta regenerated) rather than
-        // dangling-deleted. The re-drive after a rescue sets it false so the
-        // regeneration can happen at most once (no unbounded recursion).
-        Box::pin(self.heal_object_with_regen(bucket, object, version_id, opts, true)).await
-    }
-
-    /// Best-effort orphan-data-dir reclaim for an object that is healthy on this
-    /// set. Wraps [`Self::reclaim_orphan_data_dirs`] with the shared logging so
-    /// both `heal_object` exits — the already-healthy early return and the
-    /// post-heal tail — reclaim identically. Never fails the heal: delete errors
-    /// are logged and swallowed. Callers must gate this on `!opts.dry_run`.
-    async fn reclaim_orphan_data_dirs_best_effort(&self, bucket: &str, object: &str) {
-        match self.reclaim_orphan_data_dirs(bucket, object).await {
-            Ok(removed) if removed > 0 => {
-                info!(bucket, object, removed, "heal_object: reclaimed orphaned data directories");
-            }
-            Ok(_) => {}
-            Err(e) => {
-                warn!(bucket, object, error = %e, "heal_object: orphan data-dir reclaim failed");
-            }
-        }
+        Box::pin(self.heal_object_with_explicit_version_regen(bucket, object, version_id, opts, true)).await
     }
 
     #[allow(clippy::too_many_lines)]
-    async fn heal_object_with_regen(
+    async fn heal_object_with_explicit_version_regen(
         &self,
         bucket: &str,
         object: &str,
         version_id: &str,
         opts: &HealOpts,
-        allow_meta_regen: bool,
+        allow_explicit_version_regen: bool,
     ) -> disk::error::Result<(HealResultItem, Option<DiskError>)> {
         info!(?opts, "Starting heal_object");
 
@@ -354,22 +395,6 @@ impl SetDisks {
                             }
                         }
 
-                        // DATA-SAFETY GUARD (backlog#920, decision 1): before any
-                        // dangling delete, if the version's DATA shards physically
-                        // survive on >= data_blocks disks it is RECONSTRUCTABLE.
-                        // Regenerate the missing xl.meta from a surviving valid
-                        // FileInfo and re-drive the heal instead of destroying a
-                        // recoverable version. Torn writes (< data_blocks data
-                        // shards) fall through to the existing dangling behavior.
-                        if cannot_heal
-                            && allow_meta_regen
-                            && self
-                                .try_regenerate_recoverable_meta(bucket, object, &parts_metadata, &errs, &disks)
-                                .await?
-                        {
-                            return Box::pin(self.heal_object_with_regen(bucket, object, version_id, opts, false)).await;
-                        }
-
                         if cannot_heal {
                             let total_disks = parts_metadata.len();
                             let healthy_count = total_disks.saturating_sub(disks_to_heal_count);
@@ -422,6 +447,20 @@ impl SetDisks {
                                 );
                             }
 
+                            // `disks_with_all_parts` normalizes conflicting entries
+                            // in `parts_metadata` to defaults. Re-read only before
+                            // destructive cleanup so the guard sees every original
+                            // identity.
+                            let (delete_guard_metadata, delete_guard_errs) =
+                                Self::read_all_fileinfo(&disks, "", bucket, object, version_id, true, true, false).await?;
+                            if self
+                                .dangling_delete_safety(bucket, object, &delete_guard_metadata, &delete_guard_errs, &disks)
+                                .await?
+                                == DanglingDeleteSafety::UnsafeToDelete
+                            {
+                                return Ok((result, Some(cannot_heal_err)));
+                            }
+
                             // Allow for dangling deletes, on versions that have DataDir missing etc.
                             // this would end up restoring the correct readable versions.
                             return match self
@@ -837,17 +876,25 @@ impl SetDisks {
                 }
             }
             Err(err) => {
-                // DATA-SAFETY GUARD (backlog#920, decision 1): meta quorum failed,
-                // but the version's DATA may still physically survive on enough
-                // disks (xl.meta lost on > parity disks while part files remain).
-                // Rescue it by regenerating the missing xl.meta and re-driving heal
-                // instead of dangling-deleting a reconstructable version.
-                if allow_meta_regen
+                if allow_explicit_version_regen
+                    && !version_id.is_empty()
                     && self
-                        .try_regenerate_recoverable_meta(bucket, object, &parts_metadata, &errs, &disks)
+                        .try_regenerate_explicit_version_meta(bucket, object, version_id, &parts_metadata, &errs, &disks)
                         .await?
                 {
-                    return Box::pin(self.heal_object_with_regen(bucket, object, version_id, opts, false)).await;
+                    return Box::pin(self.heal_object_with_explicit_version_regen(bucket, object, version_id, opts, false)).await;
+                }
+
+                if self
+                    .dangling_delete_safety(bucket, object, &parts_metadata, &errs, &disks)
+                    .await?
+                    == DanglingDeleteSafety::UnsafeToDelete
+                {
+                    return Ok((
+                        self.default_heal_result(FileInfo::default(), &errs, bucket, object, version_id)
+                            .await,
+                        Some(err),
+                    ));
                 }
 
                 let data_errs_by_part = HashMap::new();
@@ -883,129 +930,226 @@ impl SetDisks {
         }
     }
 
-    /// backlog#920 (decision 1): rescue a version that meta-quorum logic would
-    /// otherwise dangling-DELETE, when its DATA is still reconstructable.
-    ///
-    /// Returns `Ok(true)` if the version was rescued (missing xl.meta regenerated
-    /// on at least one disk, so a re-driven heal can reconstruct it), `Ok(false)`
-    /// to fall through to the existing dangling-delete behavior.
-    ///
-    /// Recoverability is computed by physically probing part files across ALL
-    /// disks in the set with `check_parts` — including disks whose xl.meta is
-    /// absent (a lost xl.meta does not lose the sibling `part.*` data). If at
-    /// least `data_blocks` disks hold every part of a surviving valid FileInfo,
-    /// the object is EC-reconstructable, so we regenerate that FileInfo's xl.meta
-    /// on every disk whose metadata is absent (via `write_metadata`, which merges
-    /// into any existing xl.meta). Delete markers, remote/transitioned versions,
-    /// and genuine torn writes (< `data_blocks` surviving data shards) are NOT
-    /// rescued — they keep the current dangling-delete-after-grace behavior, so no
-    /// regression on those paths.
-    async fn try_regenerate_recoverable_meta(
+    async fn try_regenerate_explicit_version_meta(
+        &self,
+        bucket: &str,
+        object: &str,
+        version_id: &str,
+        parts_metadata: &[FileInfo],
+        errs: &[Option<DiskError>],
+        disks: &[Option<DiskStore>],
+    ) -> disk::error::Result<bool> {
+        let Ok(version_id) = Uuid::parse_str(version_id) else {
+            return Ok(false);
+        };
+        let candidates = parts_metadata
+            .iter()
+            .zip(errs.iter())
+            .filter_map(|(file_info, err)| {
+                (err.is_none()
+                    && file_info_is_valid_for_metadata(file_info)
+                    && file_info.version_id == Some(version_id)
+                    && file_info.has_valid_erasure_geometry()
+                    && !file_info.deleted
+                    && !file_info.is_remote()
+                    && file_info.data_dir.is_some()
+                    && !file_info.parts.is_empty()
+                    && file_info.erasure.data_blocks > 0
+                    && file_info
+                        .erasure
+                        .data_blocks
+                        .checked_add(file_info.erasure.parity_blocks)
+                        .is_some_and(|shards| shards == disks.len()))
+                .then_some(file_info)
+            })
+            .collect::<Vec<_>>();
+        let Some(candidate) = candidates.first().copied() else {
+            return Ok(false);
+        };
+        let identity = Self::file_info_quorum_hash(candidate);
+        if candidates
+            .iter()
+            .any(|file_info| Self::file_info_quorum_hash(file_info) != identity)
+        {
+            return Ok(false);
+        }
+
+        let mut available = 0usize;
+        for disk in disks {
+            let Some(disk) = disk else {
+                return Ok(false);
+            };
+            match disk.check_parts(bucket, object, candidate).await {
+                Ok(response)
+                    if !response.results.is_empty() && response.results.iter().all(|result| *result == CHECK_PART_SUCCESS) =>
+                {
+                    available += 1;
+                }
+                Ok(_)
+                | Err(
+                    DiskError::FileNotFound
+                    | DiskError::FileVersionNotFound
+                    | DiskError::PathNotFound
+                    | DiskError::VolumeNotFound,
+                ) => {}
+                Err(_) => return Ok(false),
+            }
+        }
+        if available < candidate.erasure.data_blocks {
+            return Ok(false);
+        }
+
+        let mut wrote = 0usize;
+        for (index, disk) in disks.iter().enumerate() {
+            let Some(disk) = disk else {
+                return Ok(false);
+            };
+            let metadata_absent = matches!(
+                errs.get(index).and_then(Option::as_ref),
+                Some(DiskError::FileNotFound | DiskError::FileVersionNotFound)
+            );
+            if !metadata_absent {
+                continue;
+            }
+            let Some(&shard_index) = candidate.erasure.distribution.get(index) else {
+                return Ok(false);
+            };
+            let mut regenerated = candidate.clone();
+            regenerated.fresh = false;
+            regenerated.erasure.index = shard_index;
+            match disk.write_metadata("", bucket, object, regenerated).await {
+                Ok(()) => wrote += 1,
+                Err(error) => {
+                    warn!(
+                        bucket,
+                        object,
+                        disk_index = index,
+                        error = %error,
+                        "failed to regenerate recoverable xl.meta"
+                    );
+                }
+            }
+        }
+        Ok(wrote > 0)
+    }
+
+    /// Best-effort orphan-data-dir reclaim for an object that is healthy on this
+    /// set. Wraps [`Self::reclaim_orphan_data_dirs`] with the shared logging so
+    /// both `heal_object` exits — the already-healthy early return and the
+    /// post-heal tail — reclaim identically. Never fails the heal: delete errors
+    /// are logged and swallowed. Callers must gate this on `!opts.dry_run`.
+    async fn reclaim_orphan_data_dirs_best_effort(&self, bucket: &str, object: &str) {
+        match self.reclaim_orphan_data_dirs(bucket, object).await {
+            Ok(removed) if removed > 0 => {
+                info!(bucket, object, removed, "heal_object: reclaimed orphaned data directories");
+            }
+            Ok(_) => {}
+            Err(e) => {
+                warn!(bucket, object, error = %e, "heal_object: orphan data-dir reclaim failed");
+            }
+        }
+    }
+
+    /// Prevent dangling cleanup when surviving state cannot prove that deletion
+    /// is safe. Part presence proves only recoverability, never commit: the write
+    /// path can durably rename data before xl.meta is committed.
+    async fn dangling_delete_safety(
         &self,
         bucket: &str,
         object: &str,
         parts_metadata: &[FileInfo],
         errs: &[Option<DiskError>],
         disks: &[Option<DiskStore>],
-    ) -> disk::error::Result<bool> {
-        // A surviving valid, non-deleted, non-remote data FileInfo to rebuild from.
-        let Some(surviving) = parts_metadata
+    ) -> disk::error::Result<DanglingDeleteSafety> {
+        if disks.iter().any(Option::is_none)
+            || errs.iter().flatten().any(|err| {
+                !matches!(
+                    err,
+                    DiskError::FileNotFound
+                        | DiskError::FileVersionNotFound
+                        | DiskError::PathNotFound
+                        | DiskError::VolumeNotFound
+                )
+            })
+        {
+            return Ok(DanglingDeleteSafety::UnsafeToDelete);
+        }
+
+        let mut candidates = Vec::<RecoverableMetaCandidate>::with_capacity(parts_metadata.len());
+        for (fi, err) in parts_metadata.iter().zip(errs.iter()) {
+            if err.is_some() || !file_info_is_valid_for_metadata(fi) {
+                continue;
+            }
+
+            let identity = Self::file_info_quorum_hash(fi);
+            if !candidates.iter().any(|candidate| candidate.identity == identity) {
+                let local_payload = fi.has_valid_erasure_geometry()
+                    && !fi.deleted
+                    && !fi.is_remote()
+                    && fi.data_dir.is_some()
+                    && !fi.parts.is_empty()
+                    && fi.erasure.data_blocks > 0
+                    && fi
+                        .erasure
+                        .data_blocks
+                        .checked_add(fi.erasure.parity_blocks)
+                        .is_some_and(|shards| shards == disks.len());
+                candidates.push(RecoverableMetaCandidate {
+                    identity,
+                    file_info: fi.clone(),
+                    data_count: 0,
+                    local_payload,
+                });
+            }
+        }
+
+        if candidates
             .iter()
-            .find(|fi| fi.has_valid_erasure_geometry() && !fi.deleted && !fi.is_remote())
-            .cloned()
-        else {
-            return Ok(false);
-        };
-
-        // Without a data_dir + parts there is no data to prove recoverable.
-        if surviving.data_dir.is_none() || surviving.parts.is_empty() {
-            return Ok(false);
-        }
-        let data_blocks = surviving.erasure.data_blocks;
-        if data_blocks == 0 {
-            return Ok(false);
+            .any(|candidate| candidate.file_info.deleted || candidate.file_info.is_remote())
+            || candidates.len() > 1
+        {
+            return Ok(DanglingDeleteSafety::UnsafeToDelete);
         }
 
-        // Physically probe part presence on EVERY online disk using the surviving
-        // FileInfo's data_dir/parts. `check_parts` stats `object/<data_dir>/part.N`
-        // directly, so it counts disks that still hold the data even if their
-        // xl.meta was deleted.
-        let mut available = 0usize;
-        for disk in disks.iter().flatten() {
-            if let Ok(resp) = disk.check_parts(bucket, object, &surviving).await
-                && !resp.results.is_empty()
-                && resp.results.iter().all(|r| *r == CHECK_PART_SUCCESS)
-            {
-                available += 1;
-            }
-        }
+        for candidate in candidates.iter_mut().filter(|candidate| candidate.local_payload) {
+            for (disk_index, disk) in disks.iter().enumerate() {
+                let Some(disk) = disk else {
+                    return Ok(DanglingDeleteSafety::UnsafeToDelete);
+                };
+                #[cfg(test)]
+                let check_result = match injected_dangling_check_parts_error(bucket, object, disk_index) {
+                    Some(error) => Err(error),
+                    None => disk.check_parts(bucket, object, &candidate.file_info).await,
+                };
+                #[cfg(not(test))]
+                let check_result = disk.check_parts(bucket, object, &candidate.file_info).await;
 
-        // Torn write: fewer than data_blocks surviving data shards is genuinely
-        // unrecoverable — preserve the current dangling behavior (no resurrection).
-        if available < data_blocks {
-            debug!(
-                bucket,
-                object,
-                available,
-                data_blocks,
-                "heal_object: version not reconstructable (torn write), keeping dangling behavior"
-            );
-            return Ok(false);
-        }
-
-        // Reconstructable: regenerate the surviving xl.meta on every disk whose
-        // metadata is absent so the version regains read-quorum. Each disk gets its
-        // OWN shard index: the disk at physical position `index` holds shard
-        // `distribution[index]` (mirrors `shuffle_disks` + the write path's
-        // `erasure.index = shuffled_pos + 1`). Copying the surviving disk's index
-        // verbatim would write an inconsistent xl.meta that the re-heal then treats
-        // as corrupt.
-        let distribution = &surviving.erasure.distribution;
-        let mut wrote = 0usize;
-        for (index, disk) in disks.iter().enumerate() {
-            let Some(disk) = disk else { continue };
-            let meta_absent = matches!(
-                errs.get(index).and_then(Option::as_ref),
-                Some(DiskError::FileNotFound | DiskError::FileVersionNotFound)
-            ) || !parts_metadata.get(index).map(FileInfo::is_valid).unwrap_or(false);
-            if !meta_absent {
-                continue;
-            }
-            // Without a known shard index for this position we cannot write a
-            // consistent xl.meta; leave it for the normal heal to reconstruct.
-            let Some(&shard_index) = distribution.get(index) else {
-                continue;
-            };
-            let mut regen = surviving.clone();
-            regen.fresh = false; // merge into any existing xl.meta on the disk
-            regen.erasure.index = shard_index;
-            match disk.write_metadata("", bucket, object, regen).await {
-                Ok(()) => wrote += 1,
-                Err(e) => {
-                    warn!(
-                        bucket,
-                        object,
-                        disk_index = index,
-                        error = %e,
-                        "heal_object: failed to regenerate recoverable xl.meta on disk"
-                    );
+                match check_result {
+                    Ok(resp) if !resp.results.is_empty() && resp.results.iter().all(|result| *result == CHECK_PART_SUCCESS) => {
+                        candidate.data_count += 1;
+                    }
+                    Ok(_) => {}
+                    Err(
+                        DiskError::FileNotFound
+                        | DiskError::FileVersionNotFound
+                        | DiskError::PathNotFound
+                        | DiskError::VolumeNotFound,
+                    ) => {}
+                    Err(_) => return Ok(DanglingDeleteSafety::UnsafeToDelete),
                 }
             }
         }
 
-        if wrote == 0 {
-            return Ok(false);
-        }
-
-        info!(
-            bucket,
-            object,
-            available,
-            data_blocks,
-            regenerated_meta_disks = wrote,
-            "heal_object: rescued reconstructable sub-quorum version by regenerating xl.meta"
-        );
-        Ok(true)
+        Ok(
+            if candidates
+                .iter()
+                .any(|candidate| candidate.local_payload && candidate.data_count >= candidate.file_info.erasure.data_blocks)
+            {
+                DanglingDeleteSafety::UnsafeToDelete
+            } else {
+                DanglingDeleteSafety::NoRecoverableCandidate
+            },
+        )
     }
 
     pub(in crate::set_disk) async fn heal_object_dir_locked(
@@ -1393,7 +1537,7 @@ impl crate::storage_api_contracts::heal::HealOperations for SetDisks {
 
 #[cfg(test)]
 mod heal_result_report_tests {
-    use super::SetDisks;
+    use super::{DanglingCheckPartsFailure, DanglingDeleteSafety, SetDisks};
     use super::{HEAL_RENAME_INCOMPLETE, HealRenameFailureScope};
     use crate::disk::endpoint::Endpoint;
     use crate::disk::error::DiskError;
@@ -1405,10 +1549,12 @@ mod heal_result_report_tests {
     use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations as _};
     use crate::{config::storageclass, store::init_format::save_format_file};
     use rustfs_common::heal_channel::{DriveState, HealOpts, HealScanMode};
-    use rustfs_filemeta::{BLOCK_SIZE_V2, FileInfo};
+    use rustfs_filemeta::{BLOCK_SIZE_V2, FileInfo, ObjectPartInfo, TRANSITION_COMPLETE};
     use std::sync::Arc;
     use tempfile::TempDir;
+    use time::OffsetDateTime;
     use tokio::sync::RwLock;
+    use uuid::Uuid;
 
     async fn real_disk() -> (TempDir, Endpoint, DiskStore) {
         let dir = tempfile::tempdir().expect("tempdir should be created");
@@ -1446,6 +1592,68 @@ mod heal_result_report_tests {
         .await
     }
 
+    fn meta_regen_test_fileinfo(object: &str, data_dir: Uuid, mod_time: i64, disk_index: usize) -> FileInfo {
+        let mut fi = FileInfo::new(object, 2, 2);
+        fi.data_dir = Some(data_dir);
+        fi.mod_time = Some(OffsetDateTime::from_unix_timestamp(mod_time).expect("test timestamp should parse"));
+        fi.size = 1;
+        fi.parts = vec![ObjectPartInfo {
+            number: 1,
+            size: 1,
+            actual_size: 1,
+            ..Default::default()
+        }];
+        fi.erasure.index = fi.erasure.distribution[disk_index];
+        fi
+    }
+
+    async fn meta_regen_test_set(
+        bucket: &str,
+        object: &str,
+        data_dirs: &[(Uuid, usize)],
+    ) -> (Vec<TempDir>, Arc<SetDisks>, Vec<Option<DiskStore>>) {
+        let mut temp_dirs = Vec::new();
+        let mut endpoints = Vec::new();
+        let mut disks = Vec::new();
+        for disk_index in 0..4 {
+            let (temp_dir, endpoint, disk) = real_disk().await;
+            disk.make_volume(bucket).await.expect("test bucket should be created");
+            for (data_dir, shard_count) in data_dirs {
+                if disk_index >= *shard_count {
+                    continue;
+                }
+                let part_dir = temp_dir.path().join(bucket).join(object).join(data_dir.to_string());
+                tokio::fs::create_dir_all(&part_dir)
+                    .await
+                    .expect("test data directory should be created");
+                tokio::fs::write(part_dir.join("part.1"), [1u8; 2])
+                    .await
+                    .expect("test data shard should be written");
+            }
+            temp_dirs.push(temp_dir);
+            endpoints.push(endpoint);
+            disks.push(Some(disk));
+        }
+
+        let set = set_disks_with(disks.clone(), endpoints, 2).await;
+        (temp_dirs, set, disks)
+    }
+
+    async fn seed_meta_regen_test_metadata(
+        disks: &[Option<DiskStore>],
+        disk_index: usize,
+        bucket: &str,
+        object: &str,
+        file_info: &FileInfo,
+    ) {
+        disks[disk_index]
+            .as_ref()
+            .expect("metadata test disk should be online")
+            .write_metadata("", bucket, object, file_info.clone())
+            .await
+            .expect("test metadata should be written");
+    }
+
     async fn formatted_single_disk_no_parity_set() -> (TempDir, Arc<SetDisks>) {
         let format = FormatV3::new(1, 1);
         let dir = tempfile::tempdir().expect("tempdir should be created");
@@ -1753,6 +1961,312 @@ mod heal_result_report_tests {
         assert_eq!(result.before.drives[3].state, DriveState::Ok.to_string());
     }
 
+    #[tokio::test]
+    async fn dangling_delete_guard_preserves_conflicting_identities_without_writing_metadata() {
+        let bucket = "bucket-delete-guard-conflict";
+        let object = "object.bin";
+        let old_data_dir = Uuid::parse_str("99999999-9999-9999-9999-999999999999").expect("old data dir should parse");
+        let new_data_dir = Uuid::parse_str("aaaaaaaa-aaaa-aaaa-aaaa-aaaaaaaaaaaa").expect("new data dir should parse");
+        let (_temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[(old_data_dir, 4), (new_data_dir, 2)]).await;
+        let version_id = Uuid::parse_str("bbbbbbbb-bbbb-bbbb-bbbb-bbbbbbbbbbbb").expect("version id should parse");
+        let mut metadata = vec![
+            meta_regen_test_fileinfo(object, old_data_dir, 9, 0),
+            meta_regen_test_fileinfo(object, new_data_dir, 10, 1),
+            FileInfo::default(),
+            FileInfo::default(),
+        ];
+        metadata[0].version_id = Some(version_id);
+        metadata[1].version_id = Some(version_id);
+        assert_eq!(
+            metadata[0].version_id, metadata[1].version_id,
+            "the conflicting candidates must share one version id"
+        );
+        seed_meta_regen_test_metadata(&disks, 0, bucket, object, &metadata[0]).await;
+        seed_meta_regen_test_metadata(&disks, 1, bucket, object, &metadata[1]).await;
+        let errs = vec![None, None, Some(DiskError::FileNotFound), Some(DiskError::FileNotFound)];
+
+        assert!(
+            set.dangling_delete_safety(bucket, object, &metadata, &errs, &disks)
+                .await
+                .expect("conflicting identities should be classified")
+                == DanglingDeleteSafety::UnsafeToDelete
+        );
+        let reversed = vec![
+            metadata[1].clone(),
+            metadata[0].clone(),
+            FileInfo::default(),
+            FileInfo::default(),
+        ];
+        assert!(
+            set.dangling_delete_safety(bucket, object, &reversed, &errs, &disks)
+                .await
+                .expect("reversed identities should be classified")
+                == DanglingDeleteSafety::UnsafeToDelete
+        );
+        let version_id = version_id.to_string();
+        assert!(
+            !set.try_regenerate_explicit_version_meta(bucket, object, &version_id, &metadata, &errs, &disks)
+                .await
+                .expect("conflicting explicit-version candidates should be rejected"),
+            "an explicit version must not select between conflicting metadata identities"
+        );
+        for disk_index in [2, 3] {
+            assert!(
+                matches!(
+                    disks[disk_index]
+                        .as_ref()
+                        .expect("test disk should be online")
+                        .read_version("", bucket, object, "", &ReadOptions::default())
+                        .await,
+                    Err(DiskError::FileNotFound)
+                ),
+                "the delete guard must not manufacture metadata on missing disks"
+            );
+        }
+        let old = disks[0]
+            .as_ref()
+            .expect("first test disk should be online")
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("old metadata should remain readable");
+        let new = disks[1]
+            .as_ref()
+            .expect("second test disk should be online")
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("new metadata should remain readable");
+        assert_eq!(old.data_dir, Some(old_data_dir));
+        assert_eq!(new.data_dir, Some(new_data_dir));
+    }
+
+    #[tokio::test]
+    async fn heal_meta_quorum_failure_preserves_reconstructable_uncommitted_candidate() {
+        let bucket = "bucket-delete-guard-reconstructable";
+        let object = "object.bin";
+        let data_dir = Uuid::parse_str("33333333-3333-3333-3333-333333333333").expect("data dir should parse");
+        let (_temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[(data_dir, 2)]).await;
+        let metadata = [
+            meta_regen_test_fileinfo(object, data_dir, 3, 0),
+            FileInfo::default(),
+            FileInfo::default(),
+            FileInfo::default(),
+        ];
+        seed_meta_regen_test_metadata(&disks, 0, bucket, object, &metadata[0]).await;
+        let (observed_metadata, observed_errs) = SetDisks::read_all_fileinfo(&disks, "", bucket, object, "", true, true, false)
+            .await
+            .expect("test metadata should be readable across the set");
+        assert_eq!(
+            set.dangling_delete_safety(bucket, object, &observed_metadata, &observed_errs, &disks)
+                .await
+                .expect("observed reconstructable candidate should be classified"),
+            DanglingDeleteSafety::UnsafeToDelete
+        );
+
+        let (_, err) = set
+            .heal_object(
+                bucket,
+                object,
+                "",
+                &HealOpts {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("unsafe dangling state should be reported without deletion");
+        assert_eq!(err, Some(DiskError::FileNotFound));
+        let surviving = disks[0]
+            .as_ref()
+            .expect("first test disk should be online")
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("the only metadata copy must be preserved");
+        assert_eq!(surviving.data_dir, Some(data_dir));
+        assert!(
+            matches!(
+                disks[1]
+                    .as_ref()
+                    .expect("second test disk should be online")
+                    .read_version("", bucket, object, "", &ReadOptions::default())
+                    .await,
+                Err(DiskError::FileNotFound)
+            ),
+            "the delete guard must not propagate metadata"
+        );
+    }
+
+    #[tokio::test]
+    async fn heal_meta_quorum_failure_preserves_candidate_when_required_shard_disk_is_offline() {
+        let bucket = "bucket-delete-guard-offline";
+        let object = "object.bin";
+        let data_dir = Uuid::parse_str("44444444-4444-4444-4444-444444444444").expect("data dir should parse");
+        let (temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[(data_dir, 2)]).await;
+        let metadata = meta_regen_test_fileinfo(object, data_dir, 4, 0);
+        seed_meta_regen_test_metadata(&disks, 0, bucket, object, &metadata).await;
+        set.disks.write().await[1] = None;
+
+        let (_, err) = set
+            .heal_object(
+                bucket,
+                object,
+                "",
+                &HealOpts {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("offline shard state should be reported without deletion");
+        assert_eq!(err, Some(DiskError::FileNotFound));
+        let surviving = disks[0]
+            .as_ref()
+            .expect("first test disk should be online")
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("offline uncertainty must preserve the surviving metadata");
+        assert_eq!(surviving.data_dir, Some(data_dir));
+        assert!(
+            temp_dirs[0]
+                .path()
+                .join(bucket)
+                .join(object)
+                .join(data_dir.to_string())
+                .join("part.1")
+                .is_file(),
+            "offline uncertainty must preserve the last online shard"
+        );
+    }
+
+    #[tokio::test]
+    async fn heal_meta_quorum_failure_preserves_candidate_when_part_probe_times_out() {
+        let bucket = "bucket-delete-guard-timeout";
+        let object = "object.bin";
+        let data_dir = Uuid::parse_str("55555555-5555-5555-5555-555555555555").expect("data dir should parse");
+        let (temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[(data_dir, 2)]).await;
+        let metadata = meta_regen_test_fileinfo(object, data_dir, 5, 0);
+        seed_meta_regen_test_metadata(&disks, 0, bucket, object, &metadata).await;
+        let _failure = DanglingCheckPartsFailure::install(bucket, object, 1, DiskError::Timeout);
+
+        let (_, err) = set
+            .heal_object(
+                bucket,
+                object,
+                "",
+                &HealOpts {
+                    no_lock: true,
+                    ..Default::default()
+                },
+            )
+            .await
+            .expect("part probe timeout should be reported without deletion");
+        assert_eq!(err, Some(DiskError::FileNotFound));
+        let surviving = disks[0]
+            .as_ref()
+            .expect("first test disk should be online")
+            .read_version("", bucket, object, "", &ReadOptions::default())
+            .await
+            .expect("probe uncertainty must preserve the surviving metadata");
+        assert_eq!(surviving.data_dir, Some(data_dir));
+        assert!(
+            temp_dirs[0]
+                .path()
+                .join(bucket)
+                .join(object)
+                .join(data_dir.to_string())
+                .join("part.1")
+                .is_file(),
+            "probe uncertainty must preserve the last confirmed shard"
+        );
+    }
+
+    #[tokio::test]
+    async fn dangling_delete_guard_ignores_set_incompatible_geometry() {
+        let bucket = "bucket-delete-guard-short-geometry";
+        let object = "object.bin";
+        let data_dir = Uuid::parse_str("abababab-abab-abab-abab-abababababab").expect("data dir should parse");
+        let (_temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[(data_dir, 1)]).await;
+        let mut candidate = FileInfo::new(object, 1, 0);
+        candidate.data_dir = Some(data_dir);
+        candidate.mod_time = Some(OffsetDateTime::from_unix_timestamp(18).expect("timestamp should parse"));
+        candidate.size = 1;
+        candidate.parts = vec![ObjectPartInfo {
+            number: 1,
+            size: 1,
+            actual_size: 1,
+            ..Default::default()
+        }];
+        candidate.erasure.index = candidate.erasure.distribution[0];
+        seed_meta_regen_test_metadata(&disks, 0, bucket, object, &candidate).await;
+        let metadata = vec![candidate, FileInfo::default(), FileInfo::default(), FileInfo::default()];
+        let errs = vec![
+            None,
+            Some(DiskError::FileNotFound),
+            Some(DiskError::FileNotFound),
+            Some(DiskError::FileNotFound),
+        ];
+
+        assert!(
+            set.dangling_delete_safety(bucket, object, &metadata, &errs, &disks)
+                .await
+                .expect("set-incompatible geometry should be classified")
+                == DanglingDeleteSafety::NoRecoverableCandidate
+        );
+    }
+
+    #[tokio::test]
+    async fn dangling_delete_guard_preserves_delete_marker_and_remote_metadata() {
+        let bucket = "bucket-delete-guard-nonlocal";
+        let object = "object.bin";
+        let (_temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[]).await;
+        let marker = FileInfo {
+            name: object.to_string(),
+            version_id: Some(Uuid::parse_str("eeeeeeee-eeee-eeee-eeee-eeeeeeeeeeee").expect("version id should parse")),
+            deleted: true,
+            mod_time: Some(OffsetDateTime::from_unix_timestamp(14).expect("marker timestamp should parse")),
+            ..Default::default()
+        };
+        let remote_dir = Uuid::parse_str("89898989-8989-8989-8989-898989898989").expect("remote data dir should parse");
+        let mut remote = meta_regen_test_fileinfo(object, remote_dir, 15, 1);
+        remote.transition_status = TRANSITION_COMPLETE.to_string();
+        remote.transition_tier = "WARM".to_string();
+        remote.transitioned_objname = "remote/object.bin".to_string();
+
+        for metadata in [marker, remote] {
+            let candidates = vec![metadata, FileInfo::default(), FileInfo::default(), FileInfo::default()];
+            let errs = vec![
+                None,
+                Some(DiskError::FileNotFound),
+                Some(DiskError::FileNotFound),
+                Some(DiskError::FileNotFound),
+            ];
+            assert_eq!(
+                set.dangling_delete_safety(bucket, object, &candidates, &errs, &disks)
+                    .await
+                    .expect("non-local metadata should be classified"),
+                DanglingDeleteSafety::UnsafeToDelete
+            );
+        }
+    }
+
+    #[tokio::test]
+    async fn dangling_delete_guard_preserves_metadata_read_uncertainty() {
+        let bucket = "bucket-delete-guard-read-error";
+        let object = "object.bin";
+        let (_temp_dirs, set, disks) = meta_regen_test_set(bucket, object, &[]).await;
+        let metadata = vec![FileInfo::default(); disks.len()];
+
+        for read_error in [DiskError::Timeout, DiskError::DiskAccessDenied, DiskError::DiskNotFound] {
+            let mut errs = vec![Some(DiskError::FileNotFound); disks.len()];
+            errs[0] = Some(read_error);
+            assert_eq!(
+                set.dangling_delete_safety(bucket, object, &metadata, &errs, &disks)
+                    .await
+                    .expect("metadata read uncertainty should be classified"),
+                DanglingDeleteSafety::UnsafeToDelete
+            );
+        }
+    }
+
     #[tokio::test]
     async fn heal_no_parity_bitrot_reports_unrecoverable_integrity_failure() {
         let (dir, set) = formatted_single_disk_no_parity_set().await;

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

@@ -27,7 +27,7 @@ use crate::multipart_listing::paginate_multipart_listing;
 use futures::{StreamExt, stream};
 use std::future::Future;
 #[cfg(test)]
-use std::sync::atomic::{AtomicBool, Ordering};
+use std::sync::atomic::{AtomicUsize, Ordering};
 use std::time::Duration;
 use tokio::task::JoinSet;
 
@@ -36,6 +36,7 @@ const MULTIPART_LIST_IO_CONCURRENCY: usize = 16;
 #[cfg(test)]
 #[derive(Clone, Copy, PartialEq, Eq)]
 pub(crate) enum MultipartCommitPause {
+    PutPartBeforeLockAcquire,
     PutPartBeforeLockLost,
     PutPartAfterRename,
     BeforeLockLost,
@@ -47,9 +48,10 @@ struct MultipartCommitBarrierState {
     bucket: String,
     object: String,
     pause: MultipartCommitPause,
-    armed: AtomicBool,
+    expected_arrivals: usize,
+    arrivals: AtomicUsize,
     arrived: tokio::sync::Notify,
-    release: tokio::sync::Notify,
+    release: tokio::sync::Semaphore,
 }
 
 #[cfg(test)]
@@ -64,13 +66,24 @@ static MULTIPART_COMMIT_BARRIER: std::sync::OnceLock<std::sync::Mutex<Option<Arc
 #[cfg(test)]
 impl MultipartCommitBarrier {
     pub(crate) fn install(bucket: &str, object: &str, pause: MultipartCommitPause) -> Self {
+        Self::install_for_arrivals(bucket, object, pause, 1)
+    }
+
+    pub(crate) fn install_for_arrivals(
+        bucket: &str,
+        object: &str,
+        pause: MultipartCommitPause,
+        expected_arrivals: usize,
+    ) -> Self {
+        assert!(expected_arrivals > 0, "multipart commit barrier must wait for at least one arrival");
         let state = Arc::new(MultipartCommitBarrierState {
             bucket: bucket.to_string(),
             object: object.to_string(),
             pause,
-            armed: AtomicBool::new(true),
+            expected_arrivals,
+            arrivals: AtomicUsize::new(0),
             arrived: tokio::sync::Notify::new(),
-            release: tokio::sync::Notify::new(),
+            release: tokio::sync::Semaphore::new(0),
         });
         let mut slot = MULTIPART_COMMIT_BARRIER
             .get_or_init(|| std::sync::Mutex::new(None))
@@ -83,20 +96,28 @@ impl MultipartCommitBarrier {
     }
 
     pub(crate) async fn wait_until_paused(&self) {
-        tokio::time::timeout(Duration::from_secs(30), self.state.arrived.notified())
-            .await
-            .expect("multipart completion should reach the deterministic commit barrier");
+        tokio::time::timeout(Duration::from_secs(30), async {
+            loop {
+                let arrived = self.state.arrived.notified();
+                if self.state.arrivals.load(Ordering::Acquire) >= self.state.expected_arrivals {
+                    return;
+                }
+                arrived.await;
+            }
+        })
+        .await
+        .expect("multipart completion should reach the deterministic commit barrier");
     }
 
     pub(crate) fn release(&self) {
-        self.state.release.notify_one();
+        self.state.release.add_permits(self.state.expected_arrivals);
     }
 }
 
 #[cfg(test)]
 impl Drop for MultipartCommitBarrier {
     fn drop(&mut self) {
-        self.state.release.notify_one();
+        self.release();
         let mut slot = MULTIPART_COMMIT_BARRIER
             .get_or_init(|| std::sync::Mutex::new(None))
             .lock()
@@ -117,10 +138,20 @@ async fn pause_multipart_commit(bucket: &str, object: &str, pause: MultipartComm
         .filter(|barrier| barrier.bucket == bucket && barrier.object == object && barrier.pause == pause)
         .cloned();
     if let Some(barrier) = barrier
-        && barrier.armed.swap(false, Ordering::AcqRel)
+        && let Ok(previous) = barrier.arrivals.fetch_update(Ordering::AcqRel, Ordering::Acquire, |current| {
+            (current < barrier.expected_arrivals).then_some(current + 1)
+        })
     {
-        barrier.arrived.notify_one();
-        barrier.release.notified().await;
+        let arrival = previous + 1;
+        if arrival == barrier.expected_arrivals {
+            barrier.arrived.notify_one();
+        }
+        barrier
+            .release
+            .acquire()
+            .await
+            .expect("multipart commit barrier should remain open")
+            .forget();
     }
 }
 
@@ -693,6 +724,8 @@ impl crate::storage_api_contracts::multipart::MultipartOperations for SetDisks {
 
             let part_path = format!("{}/{}/{}", upload_id_path, fi.data_dir.unwrap_or_default(), part_suffix);
 
+            #[cfg(test)]
+            pause_multipart_commit(bucket, object, MultipartCommitPause::PutPartBeforeLockAcquire).await;
             // Serialize only the commit (rename_part), not the whole upload. Each
             // concurrent stream writes to its own unique temp dir (see `tmp_part`
             // above), so the encode/stream phase never conflicts and must stay
@@ -3346,85 +3379,98 @@ mod tests {
     #[tokio::test(flavor = "multi_thread")]
     #[serial]
     async fn complete_holds_object_then_upload_lock_through_commit() {
-        temp_env::async_with_vars([(crate::object_api::ENV_RUSTFS_ENCRYPTED_RANGE_SEEK, Some("true"))], async {
-            let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
-            let signaling = Arc::new(SignalingLockClient::new(Arc::new(LocalClient::with_manager(manager))));
-            let lockers: Vec<Arc<dyn LockClient>> = vec![signaling.clone()];
-            let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks_with_lockers(4, 0, 2, lockers).await;
-            let bucket = "multipart-layout-lock-order-bucket";
-            let object = "object";
-            make_bucket_on_all(&disk_stores, bucket).await;
-            let create_opts = ObjectOptions {
-                user_defined: HashMap::from([(SSEC_ALGORITHM_HEADER.to_string(), "AES256".to_string())]),
-                ..Default::default()
-            };
-            let (upload_id, parts) = stage_upload_with_create_opts(&set_disks, bucket, object, &[0x43; 4096], &create_opts).await;
-            let upload_id_path = SetDisks::get_upload_id_dir(bucket, object, &upload_id);
-            let upload_resource = rustfs_lock::ObjectKey::new(RUSTFS_META_MULTIPART_BUCKET, upload_id_path);
-            let object_resource = rustfs_lock::ObjectKey::new(bucket, object);
-            signaling.set_target(upload_resource.clone());
-            signaling.clear_observed();
-            let _setup_type_guard = SetupTypeGuard::switch_to(SetupType::DistErasure).await;
-            let barrier = MultipartCommitBarrier::install(bucket, object, MultipartCommitPause::AfterRename);
+        temp_env::async_with_vars(
+            [
+                (crate::object_api::ENV_RUSTFS_ENCRYPTED_RANGE_SEEK, Some("true")),
+                (rustfs_config::ENV_OBJECT_LOCK_ACQUIRE_TIMEOUT, Some("60")),
+            ],
+            async {
+                let manager = Arc::new(rustfs_lock::GlobalLockManager::new());
+                let signaling = Arc::new(SignalingLockClient::new(Arc::new(LocalClient::with_manager(manager))));
+                let lockers: Vec<Arc<dyn LockClient>> = vec![signaling.clone()];
+                let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks_with_lockers(4, 0, 2, lockers).await;
+                let bucket = "multipart-layout-lock-order-bucket";
+                let object = "object";
+                make_bucket_on_all(&disk_stores, bucket).await;
+                let create_opts = ObjectOptions {
+                    user_defined: HashMap::from([(SSEC_ALGORITHM_HEADER.to_string(), "AES256".to_string())]),
+                    ..Default::default()
+                };
+                let (upload_id, parts) =
+                    stage_upload_with_create_opts(&set_disks, bucket, object, &[0x43; 4096], &create_opts).await;
+                let upload_id_path = SetDisks::get_upload_id_dir(bucket, object, &upload_id);
+                let upload_resource = rustfs_lock::ObjectKey::new(RUSTFS_META_MULTIPART_BUCKET, upload_id_path);
+                let object_resource = rustfs_lock::ObjectKey::new(bucket, object);
+                signaling.set_target(upload_resource.clone());
+                signaling.clear_observed();
+                let _setup_type_guard = SetupTypeGuard::switch_to(SetupType::DistErasure).await;
+                let barrier = MultipartCommitBarrier::install(bucket, object, MultipartCommitPause::AfterRename);
 
-            let complete_store = set_disks.clone();
-            let complete_upload_id = upload_id.clone();
-            let complete = tokio::spawn(async move {
-                complete_store
-                    .complete_multipart_upload(bucket, object, &complete_upload_id, parts, &ObjectOptions::default())
+                let complete_store = set_disks.clone();
+                let complete_upload_id = upload_id.clone();
+                let complete = tokio::spawn(async move {
+                    complete_store
+                        .complete_multipart_upload(bucket, object, &complete_upload_id, parts, &ObjectOptions::default())
+                        .await
+                });
+                barrier.wait_until_paused().await;
+
+                let observed = signaling.observed();
+                let object_position = observed
+                    .iter()
+                    .position(|resource| resource == &object_resource)
+                    .expect("completion should acquire the object lock");
+                let upload_position = observed
+                    .iter()
+                    .position(|resource| resource == &upload_resource)
+                    .expect("completion should acquire the upload lock");
+                assert!(object_position < upload_position, "completion must acquire object before upload");
+
+                let abort_store = set_disks.clone();
+                let abort_upload_id = upload_id.clone();
+                let abort = tokio::spawn(async move {
+                    abort_store
+                        .abort_multipart_upload(bucket, object, &abort_upload_id, &ObjectOptions::default())
+                        .await
+                });
+                signaling.wait_for_attempts(2).await;
+                tokio::task::yield_now().await;
+                assert!(!abort.is_finished(), "abort must wait until completion releases the upload lock");
+
+                let list_store = set_disks.clone();
+                let list_upload_id = upload_id.clone();
+                let list = tokio::spawn(async move {
+                    list_store
+                        .list_object_parts(bucket, object, &list_upload_id, None, MAX_PARTS_COUNT, &ObjectOptions::default())
+                        .await
+                });
+                signaling.wait_for_attempts(3).await;
+                tokio::task::yield_now().await;
+                assert!(!list.is_finished(), "ListParts must wait until completion releases the upload lock");
+
+                barrier.release();
+                complete
                     .await
-            });
-            barrier.wait_until_paused().await;
-
-            let observed = signaling.observed();
-            let object_position = observed
-                .iter()
-                .position(|resource| resource == &object_resource)
-                .expect("completion should acquire the object lock");
-            let upload_position = observed
-                .iter()
-                .position(|resource| resource == &upload_resource)
-                .expect("completion should acquire the upload lock");
-            assert!(object_position < upload_position, "completion must acquire object before upload");
-
-            let abort_store = set_disks.clone();
-            let abort_upload_id = upload_id.clone();
-            let abort = tokio::spawn(async move {
-                abort_store
-                    .abort_multipart_upload(bucket, object, &abort_upload_id, &ObjectOptions::default())
+                    .expect("completion task should not panic")
+                    .expect("completion should commit after the barrier is released");
+                let abort_err = abort
                     .await
-            });
-            signaling.wait_for_attempts(2).await;
-            tokio::task::yield_now().await;
-            assert!(!abort.is_finished(), "abort must wait until completion releases the upload lock");
-
-            let list_store = set_disks.clone();
-            let list_upload_id = upload_id.clone();
-            let list = tokio::spawn(async move {
-                list_store
-                    .list_object_parts(bucket, object, &list_upload_id, None, MAX_PARTS_COUNT, &ObjectOptions::default())
+                    .expect("abort task should not panic")
+                    .expect_err("the committed upload should no longer exist when abort acquires the lock");
+                assert!(
+                    matches!(abort_err, StorageError::InvalidUploadID(..)),
+                    "abort should return InvalidUploadID after completion, got {abort_err:?}"
+                );
+                let list_err = list
                     .await
-            });
-            signaling.wait_for_attempts(3).await;
-            tokio::task::yield_now().await;
-            assert!(!list.is_finished(), "ListParts must wait until completion releases the upload lock");
-
-            barrier.release();
-            complete
-                .await
-                .expect("completion task should not panic")
-                .expect("completion should commit after the barrier is released");
-            let abort_err = abort
-                .await
-                .expect("abort task should not panic")
-                .expect_err("the committed upload should no longer exist when abort acquires the lock");
-            assert!(matches!(abort_err, StorageError::InvalidUploadID(..)));
-            let list_err = list
-                .await
-                .expect("ListParts task should not panic")
-                .expect_err("the committed upload should no longer exist when ListParts acquires the lock");
-            assert!(matches!(list_err, StorageError::InvalidUploadID(..)));
-        })
+                    .expect("ListParts task should not panic")
+                    .expect_err("the committed upload should no longer exist when ListParts acquires the lock");
+                assert!(
+                    matches!(list_err, StorageError::InvalidUploadID(..)),
+                    "ListParts should return InvalidUploadID after completion, got {list_err:?}"
+                );
+            },
+        )
         .await;
     }
 

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

@@ -25,7 +25,10 @@ use crate::set_disk::read::GetObjectDownstreamWriter;
 
 use crate::bucket::lifecycle::{
     tier_delete_journal::{persist_tier_delete_journal_entry, remove_tier_delete_journal_entry},
-    tier_sweeper::{Jentry, RemoteTierDeleteOutcome, delete_object_from_remote_tier_with_lease_idempotent},
+    tier_sweeper::{
+        Jentry, RemoteTierDeleteOutcome, delete_confirmed_transition_candidate_exact_with_lease_idempotent,
+        delete_object_from_remote_tier_with_lease_idempotent,
+    },
     transition_transaction::{
         TransitionRemoteVersion, TransitionSourceIdentity, TransitionSourceVersionMode, TransitionTransaction,
         TransitionTransactionInit, TransitionTransactionState, delete_transition_transaction_record,
@@ -547,13 +550,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
                 &self.ctx.tier_config_mgr(),
             )
             .await?;
-            return Ok(finish_set_disk_read_lock(
-                gr,
-                read_lock_guard.take(),
-                lock_optimization_enabled,
-                bucket,
-                object,
-            ));
+            return Ok(finish_set_disk_read_lock(gr, read_lock_guard.take(), None, bucket, object));
         }
 
         // App-layer object data cache probe: metadata (etag/size) is resolved
@@ -680,6 +677,18 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
             return Ok(reader);
         }
 
+        let snapshot_lease = if lock_optimization_enabled {
+            match fi.data_dir.filter(|data_dir| !data_dir.is_nil()) {
+                Some(data_dir) => {
+                    let data_dir_path = format!("{object}/{data_dir}");
+                    acquire_snapshot_leases(&disks, bucket, &data_dir_path, fi.erasure.data_blocks).await
+                }
+                None => None,
+            }
+        } else {
+            None
+        };
+
         match codec_streaming_gate.decision {
             GetCodecStreamingDecision::Use => {
                 match Self::get_object_decode_reader_with_fileinfo(
@@ -708,13 +717,7 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
                         // Carry the hook probe result so the app layer skips its
                         // now-redundant lookup on the streaming miss path (ODC-16).
                         reader.body_source = body_source;
-                        return Ok(finish_set_disk_read_lock(
-                            reader,
-                            read_lock_guard.take(),
-                            lock_optimization_enabled,
-                            bucket,
-                            object,
-                        ));
+                        return Ok(finish_set_disk_read_lock(reader, read_lock_guard.take(), snapshot_lease, bucket, object));
                     }
                     core::io_primitives::GetCodecStreamingReaderBuildOutcome::Fallback(reason) => {
                         record_get_codec_streaming_gate_decision(
@@ -753,15 +756,22 @@ impl crate::storage_api_contracts::object::ObjectIO for SetDisks {
         let set_index = self.set_index;
         let pool_index = self.pool_index;
         let skip_verify = opts.skip_verify_bitrot;
-        if lock_optimization_enabled {
+        let producer_snapshot_lease = snapshot_lease.clone();
+        if let Some(lease) = snapshot_lease {
             release_materialized_read_lock(&bucket, &object, read_lock_guard.take());
-            debug!(bucket, object, "Lock optimization: released read lock before streaming read");
+            reader.stream = Box::new(SnapshotLeaseReader {
+                inner: reader.stream,
+                lease: Some(lease),
+                terminal_error: false,
+            });
+            debug!(bucket, object, "Lock optimization: replaced read lock with snapshot leases");
         }
 
-        // When lock optimization is disabled, keep the read-lock guard in the
-        // task so it lives for the duration of the streaming read.
+        // The producer shares the lease lifetime with the body so cancellation
+        // cannot release the snapshot while the duplex task is still unwinding.
         tokio::spawn(async move {
             let _guard = read_lock_guard;
+            let _snapshot_lease = producer_snapshot_lease;
             let mut writer = GetObjectDownstreamWriter::new(wd);
             // Do not wrap the entire read+write pipeline in `disk_read_timeout`.
             // `get_object_with_fileinfo` also waits on `writer`, so an outer timeout
@@ -1663,7 +1673,11 @@ pub(crate) async fn cleanup_uncommitted_transition_upload(
     cleanup_version: &str,
     version_id_exact: bool,
 ) -> std::io::Result<RemoteTierDeleteOutcome> {
-    delete_object_from_remote_tier_with_lease_idempotent(object, cleanup_version, lease, version_id_exact).await
+    if version_id_exact {
+        delete_confirmed_transition_candidate_exact_with_lease_idempotent(object, cleanup_version, lease).await
+    } else {
+        delete_object_from_remote_tier_with_lease_idempotent(object, cleanup_version, lease, false).await
+    }
 }
 
 fn log_transition_upload_cleanup_failure(lease: &TierOperationLease, object: &str, cleanup_version: &str, err: &std::io::Error) {
@@ -1800,7 +1814,7 @@ impl Drop for TransitionUploadCleanup {
     }
 }
 
-async fn cleanup_rejected_transition_upload_durably(
+pub(crate) async fn cleanup_rejected_transition_upload_durably(
     lease: &TierOperationLease,
     object: &str,
     cleanup_version: &str,
@@ -1813,6 +1827,13 @@ async fn cleanup_rejected_transition_upload_durably(
         tier_name: lease.tier_name().to_string(),
         backend_identity: Some(lease.backend_identity()),
         version_id_exact,
+        version_state: if !version_id_exact {
+            rustfs_filemeta::TransitionVersionState::KnownDisabled
+        } else if cleanup_version == "null" {
+            rustfs_filemeta::TransitionVersionState::SuspendedNull
+        } else {
+            rustfs_filemeta::TransitionVersionState::Exact
+        },
     };
 
     let journal_error = if let Some(api) = api.as_ref() {
@@ -1972,12 +1993,83 @@ async fn advance_and_save_transition_transaction(
     next: TransitionTransactionState,
     remote_version: Option<TransitionRemoteVersion>,
 ) -> Result<()> {
+    #[cfg(test)]
+    record_transition_uploaded_save_attempt(transaction, next);
     transaction
         .advance(transaction.fence(), next, remote_version)
         .map_err(Error::other)?;
     save_transition_transaction_if_available(api, transaction).await
 }
 
+#[cfg(test)]
+struct TransitionUploadedSaveProbeState {
+    bucket: String,
+    object: String,
+    attempts: std::sync::atomic::AtomicUsize,
+}
+
+#[cfg(test)]
+struct TransitionUploadedSaveProbe {
+    state: Arc<TransitionUploadedSaveProbeState>,
+}
+
+#[cfg(test)]
+static TRANSITION_UPLOADED_SAVE_PROBE: std::sync::OnceLock<std::sync::Mutex<Option<Arc<TransitionUploadedSaveProbeState>>>> =
+    std::sync::OnceLock::new();
+
+#[cfg(test)]
+impl TransitionUploadedSaveProbe {
+    fn install(bucket: &str, object: &str) -> Self {
+        let state = Arc::new(TransitionUploadedSaveProbeState {
+            bucket: bucket.to_string(),
+            object: object.to_string(),
+            attempts: std::sync::atomic::AtomicUsize::new(0),
+        });
+        let mut slot = TRANSITION_UPLOADED_SAVE_PROBE
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("transition uploaded-save probe mutex should not poison");
+        assert!(slot.is_none(), "transition uploaded-save probe must be installed by one test at a time");
+        *slot = Some(Arc::clone(&state));
+        drop(slot);
+        Self { state }
+    }
+
+    fn attempts(&self) -> usize {
+        self.state.attempts.load(std::sync::atomic::Ordering::Acquire)
+    }
+}
+
+#[cfg(test)]
+impl Drop for TransitionUploadedSaveProbe {
+    fn drop(&mut self) {
+        let mut slot = TRANSITION_UPLOADED_SAVE_PROBE
+            .get_or_init(|| std::sync::Mutex::new(None))
+            .lock()
+            .expect("transition uploaded-save probe mutex should not poison");
+        if slot.as_ref().is_some_and(|state| Arc::ptr_eq(state, &self.state)) {
+            *slot = None;
+        }
+    }
+}
+
+#[cfg(test)]
+fn record_transition_uploaded_save_attempt(transaction: &TransitionTransaction, next: TransitionTransactionState) {
+    if next != TransitionTransactionState::Uploaded {
+        return;
+    }
+    let state = TRANSITION_UPLOADED_SAVE_PROBE
+        .get_or_init(|| std::sync::Mutex::new(None))
+        .lock()
+        .expect("transition uploaded-save probe mutex should not poison")
+        .as_ref()
+        .filter(|state| state.bucket == transaction.source.bucket && state.object == transaction.source.object)
+        .cloned();
+    if let Some(state) = state {
+        state.attempts.fetch_add(1, std::sync::atomic::Ordering::AcqRel);
+    }
+}
+
 async fn delete_transition_transaction_if_available(api: Option<&Arc<ECStore>>, transaction_id: Uuid) -> Result<()> {
     if let Some(api) = api {
         return delete_transition_transaction_record(api.clone(), transaction_id).await;
@@ -2228,11 +2320,78 @@ async fn pause_transition_commit(bucket: &str, object: &str, pause: TransitionCo
     }
 }
 
-fn parse_transition_version_id(remote_version: &str) -> std::result::Result<Option<Uuid>, uuid::Error> {
+#[cfg(test)]
+fn persisted_transition_version(
+    remote_version: &str,
+) -> std::io::Result<(Option<String>, rustfs_filemeta::TransitionVersionState)> {
+    persisted_transition_version_with_gate(remote_version, remote_version_state_writer_enabled())
+}
+
+#[cfg(test)]
+fn remote_version_state_writer_enabled() -> bool {
+    remote_version_state_writer_fleet_proof().is_some()
+}
+
+fn remote_version_state_writer_fleet_proof() -> Option<crate::services::notification_sys::RemoteVersionStateFleetProofToken> {
+    remote_version_state_writer_requested()
+        .then(crate::services::notification_sys::acquire_remote_version_state_fleet_proof)
+        .flatten()
+}
+
+fn remote_version_state_writer_requested() -> bool {
+    remote_version_state_writer_enabled_for(
+        rustfs_utils::get_env_bool(
+            rustfs_config::ENV_TIER_REMOTE_VERSION_STATE_WRITE,
+            rustfs_config::DEFAULT_TIER_REMOTE_VERSION_STATE_WRITE,
+        ),
+        rustfs_utils::get_env_bool(
+            rustfs_config::ENV_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED,
+            rustfs_config::DEFAULT_TIER_REMOTE_VERSION_STATE_FLEET_CONFIRMED,
+        ),
+        true,
+    )
+}
+
+fn remote_version_state_writer_fleet_proof_matches(
+    proof: &crate::services::notification_sys::RemoteVersionStateFleetProofToken,
+) -> bool {
+    remote_version_state_writer_fleet_proof_matches_for(
+        remote_version_state_writer_requested(),
+        crate::services::notification_sys::remote_version_state_fleet_proof_matches(proof),
+    )
+}
+
+fn remote_version_state_writer_fleet_proof_matches_for(requested: bool, fleet_proof_matches: bool) -> bool {
+    requested && fleet_proof_matches
+}
+
+fn remote_version_state_writer_enabled_for(requested: bool, fleet_confirmed: bool, fleet_proof_valid: bool) -> bool {
+    requested && fleet_confirmed && fleet_proof_valid
+}
+
+fn persisted_transition_version_with_gate(
+    remote_version: &str,
+    remote_version_state_writer_enabled: bool,
+) -> std::io::Result<(Option<String>, rustfs_filemeta::TransitionVersionState)> {
     if remote_version.is_empty() {
-        return Ok(None);
+        return Ok((None, rustfs_filemeta::TransitionVersionState::KnownDisabled));
+    }
+
+    match Uuid::parse_str(remote_version) {
+        Ok(version_id) if version_id.is_nil() => Err(std::io::Error::new(
+            std::io::ErrorKind::InvalidData,
+            "remote tier returned a nil object version ID",
+        )),
+        Ok(_) => Ok((Some(remote_version.to_string()), rustfs_filemeta::TransitionVersionState::Exact)),
+        Err(_) if !remote_version_state_writer_enabled => Err(std::io::Error::new(
+            std::io::ErrorKind::Unsupported,
+            "opaque remote tier versions require the operator-attested live fleet capability gate",
+        )),
+        Err(_) if remote_version == "null" => {
+            Ok((Some(remote_version.to_string()), rustfs_filemeta::TransitionVersionState::SuspendedNull))
+        }
+        Err(_) => Ok((Some(remote_version.to_string()), rustfs_filemeta::TransitionVersionState::Exact)),
     }
-    Uuid::parse_str(remote_version).map(|version_id| (!version_id.is_nil()).then_some(version_id))
 }
 
 #[cfg(test)]
@@ -2470,16 +2629,20 @@ mod transition_upload_completion_tests {
 
 #[cfg(test)]
 mod transition_version_id_tests {
-    use super::{TransitionUploadCandidate, parse_transition_version_id};
+    use super::{
+        TransitionUploadCandidate, persisted_transition_version, persisted_transition_version_with_gate,
+        remote_version_state_writer_enabled_for, remote_version_state_writer_fleet_proof_matches_for,
+    };
+    use rustfs_filemeta::TransitionVersionState;
     use uuid::Uuid;
 
     #[test]
     fn normalizes_persisted_unversioned_ids_and_preserves_put_constraints() {
-        assert_eq!(parse_transition_version_id("").expect("empty remote version should be valid"), None);
         assert_eq!(
-            parse_transition_version_id(&Uuid::nil().to_string()).expect("nil remote version should be valid"),
-            None
+            persisted_transition_version("").expect("empty remote version identifies an unversioned tier"),
+            (None, TransitionVersionState::KnownDisabled)
         );
+        assert!(persisted_transition_version(&Uuid::nil().to_string()).is_err());
         let nil_put_response = Uuid::nil().to_string();
         let nil_candidate = TransitionUploadCandidate::from_put_response(nil_put_response.clone());
         assert_eq!(nil_candidate.cleanup_version(), nil_put_response);
@@ -2491,12 +2654,14 @@ mod transition_version_id_tests {
     }
 
     #[test]
-    fn preserves_valid_remote_id_and_rejects_invalid_text() {
+    fn preserves_uuid_and_gates_opaque_remote_ids() {
         let version_id = Uuid::new_v4();
         assert_eq!(
-            parse_transition_version_id(&version_id.to_string()).expect("UUID remote version should be valid"),
-            Some(version_id)
+            persisted_transition_version(&version_id.to_string()).expect("UUID remote version"),
+            (Some(version_id.to_string()), TransitionVersionState::Exact)
         );
+        assert!(persisted_transition_version("null").is_err());
+        assert!(persisted_transition_version("opaque-version-token").is_err());
         assert_eq!(
             TransitionUploadCandidate::from_put_response(version_id.to_string()).cleanup_version(),
             version_id.to_string()
@@ -2505,7 +2670,64 @@ mod transition_version_id_tests {
             TransitionUploadCandidate::from_put_response("opaque-version-token".to_string()).cleanup_version(),
             "opaque-version-token"
         );
-        assert!(parse_transition_version_id("not-a-uuid").is_err());
+    }
+
+    #[test]
+    fn remote_version_state_writer_requires_request_and_fleet_confirmation() {
+        for (case, requested, fleet_confirmed, fleet_proof_valid, expected) in [
+            ("old defaults", false, false, false, false),
+            ("missing fleet confirmation", true, false, true, false),
+            ("missing local opt-in", false, true, true, false),
+            ("missing fleet proof", true, true, false, false),
+            ("explicitly unconfirmed fleet", true, false, true, false),
+            ("rolled-back writer", false, true, true, false),
+            ("fully upgraded fleet", true, true, true, true),
+        ] {
+            assert_eq!(
+                remote_version_state_writer_enabled_for(requested, fleet_confirmed, fleet_proof_valid),
+                expected,
+                "{case}"
+            );
+        }
+    }
+
+    #[test]
+    fn remote_version_state_commit_rechecks_operator_gate_and_live_proof() {
+        for (case, requested, fleet_proof_matches, expected) in [
+            ("operator gate closed", false, true, false),
+            ("fleet proof changed", true, false, false),
+            ("current authorization", true, true, true),
+        ] {
+            assert_eq!(
+                remote_version_state_writer_fleet_proof_matches_for(requested, fleet_proof_matches),
+                expected,
+                "{case}"
+            );
+        }
+    }
+
+    #[test]
+    fn fleet_gate_enables_null_and_opaque_remote_version_states() {
+        for (remote_version, expected) in [
+            ("null", (Some("null".to_string()), TransitionVersionState::SuspendedNull)),
+            (
+                "opaque-version-token",
+                (Some("opaque-version-token".to_string()), TransitionVersionState::Exact),
+            ),
+        ] {
+            assert!(
+                persisted_transition_version_with_gate(remote_version, false).is_err(),
+                "missing fleet confirmation must reject {remote_version:?}"
+            );
+            assert_eq!(
+                persisted_transition_version_with_gate(remote_version, true).expect("fleet-confirmed state must be persisted"),
+                expected
+            );
+        }
+        assert_eq!(
+            persisted_transition_version_with_gate("", true).expect("empty remote version identifies an unversioned tier"),
+            (None, TransitionVersionState::KnownDisabled)
+        );
     }
 }
 
@@ -3665,6 +3887,16 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
         let dest_obj = transaction.remote_object.clone();
         let mut transition_meta = (*oi.user_defined).clone();
         transition_meta.insert("name".to_string(), object.to_string());
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut transition_meta,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TRANSACTION_ID,
+            transaction.transaction_id.to_string(),
+        );
+        rustfs_utils::http::metadata_compat::insert_str(
+            &mut transition_meta,
+            rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
+            rustfs_utils::crypto::hex(transaction.backend_fingerprint),
+        );
 
         if let Some(content_type) = oi.content_type.as_ref().filter(|value| !value.is_empty()) {
             transition_meta.insert(CONTENT_TYPE.to_ascii_lowercase(), content_type.clone());
@@ -3775,6 +4007,24 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
             delete_transition_transaction_after_remote_cleanup(transaction_api.as_ref(), transaction_id, bucket, object).await;
             return Err(err.into());
         }
+        let fleet_proof = remote_version_state_writer_fleet_proof();
+        let remote_version_requires_fleet_proof =
+            !candidate.remote_version().is_empty() && Uuid::parse_str(candidate.remote_version()).is_err();
+        let (transition_version_id, transition_version_state) =
+            match persisted_transition_version_with_gate(candidate.remote_version(), fleet_proof.is_some()) {
+                Ok(version) => version,
+                Err(err) => {
+                    let cleanup_api = transition_cleanup_store(&self.ctx).await;
+                    if let Err(cleanup_err) = upload_cleanup.cleanup_rejected_upload(cleanup_api).await {
+                        return Err(StorageError::Io(std::io::Error::other(format!(
+                            "{err}; rejected remote upload cleanup failed: {cleanup_err}"
+                        ))));
+                    }
+                    delete_transition_transaction_after_remote_cleanup(transaction_api.as_ref(), transaction_id, bucket, object)
+                        .await;
+                    return Err(err.into());
+                }
+            };
         if let Err(err) = advance_and_save_transition_transaction(
             transaction_api.as_ref(),
             &mut transaction,
@@ -3792,16 +4042,6 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
             delete_transition_transaction_after_remote_cleanup(transaction_api.as_ref(), transaction_id, bucket, object).await;
             return Err(err);
         }
-        let transition_version_id = match parse_transition_version_id(candidate.remote_version()) {
-            Ok(version_id) => version_id,
-            Err(err) => {
-                if upload_cleanup.cleanup().await.is_ok() {
-                    delete_transition_transaction_after_remote_cleanup(transaction_api.as_ref(), transaction_id, bucket, object)
-                        .await;
-                }
-                return Err(err.into());
-            }
-        };
 
         let mut commit_opts = opts.clone();
         commit_opts.no_lock = true;
@@ -3859,7 +4099,11 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
         current_fi.transition_status = TRANSITION_COMPLETE.to_string();
         current_fi.transitioned_objname = dest_obj;
         current_fi.transition_tier = opts.transition.tier.clone();
-        current_fi.transition_version_id = transition_version_id;
+        current_fi.transition_version_id = transition_version_id
+            .as_deref()
+            .and_then(|version_id| Uuid::parse_str(version_id).ok());
+        current_fi.transition_version = transition_version_id;
+        current_fi.transition_version_state = transition_version_state;
         rustfs_utils::http::metadata_compat::insert_str(
             &mut current_fi.metadata,
             rustfs_utils::http::metadata_compat::SUFFIX_TRANSITION_TIER_DESTINATION_ID,
@@ -3896,6 +4140,21 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
         }
         #[cfg(test)]
         pause_transition_commit(bucket, object, TransitionCommitPause::AfterLeaseValidation).await;
+        // This check is the fleet-proof lease linearization point. Revocation
+        // blocks later commits; an already-authorized local quorum commit is
+        // allowed to finish without holding a synchronous lock across I/O.
+        if remote_version_requires_fleet_proof
+            && !fleet_proof
+                .as_ref()
+                .is_some_and(remote_version_state_writer_fleet_proof_matches)
+        {
+            drop(transition_lock_guard);
+            if upload_cleanup.cleanup().await.is_ok() {
+                delete_transition_transaction_after_remote_cleanup(transaction_api.as_ref(), transaction_id, bucket, object)
+                    .await;
+            }
+            return Err(Error::other("remote version state fleet capability changed during transition"));
+        }
         if let Err(err) = advance_and_save_transition_transaction(
             transaction_api.as_ref(),
             &mut transaction,
@@ -4076,6 +4335,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
             let mut p_reader = PutObjReader::new(hash_reader);
             return match self_.clone().put_object(bucket, object, &mut p_reader, &ropts).await {
                 Ok(restored_info) => {
+                    let restored_info = self_.finalize_restore_metadata(bucket, object, &restored_info, &opts).await?;
                     send_event(EventArgs {
                         event_name: EventName::ObjectRestoreCompleted.as_str().to_string(),
                         bucket_name: bucket.to_string(),
@@ -4210,6 +4470,7 @@ impl crate::storage_api_contracts::object::ObjectOperations for SetDisks {
                 return set_restore_header_fn(&mut oi, Some(err)).await;
             }
         };
+        let restored_info = self_.finalize_restore_metadata(bucket, object, &restored_info, opts).await?;
         send_event(EventArgs {
             event_name: EventName::ObjectRestoreCompleted.as_str().to_string(),
             bucket_name: bucket.to_string(),
@@ -4668,6 +4929,33 @@ mod transition_commit_failure_tests {
     }
 
     #[tokio::test]
+    async fn rejected_unsupported_remote_versions_are_cleaned_up() {
+        for remote_version in ["null", "opaque-version-token"] {
+            let manager = TierConfigMgr::new();
+            let backend = register_mock_tier(&manager, "WARM").await;
+            let lease = TierConfigMgr::acquire_operation_lease(&manager, "WARM")
+                .await
+                .expect("mock tier lease should be available");
+            let candidate = TransitionUploadCandidate::from_put_response(remote_version.to_string());
+
+            persisted_transition_version(candidate.remote_version()).expect_err("unsupported writer version must fail closed");
+            cleanup_rejected_transition_upload_durably(
+                &lease,
+                "remote/object",
+                candidate.cleanup_version(),
+                candidate.cleanup_version_is_exact(),
+                None,
+            )
+            .await
+            .expect("rejected remote upload must be cleaned up");
+
+            assert_eq!(
+                backend.remove_versions().await,
+                vec![("remote/object".to_string(), candidate.cleanup_version().to_string())]
+            );
+        }
+    }
+
     #[serial_test::serial(restore_multipart_failure_point)]
     async fn multipart_restore_aborts_every_post_create_failure() {
         let (temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
@@ -6615,6 +6903,45 @@ mod transition_upload_integrity_tests {
         );
     }
 
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn unversioned_remote_version_is_persisted_without_version_id() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "transition-unversioned-tier-bucket";
+        let object = "object.bin";
+        let payload = b"unversioned remote tier must commit without a version id".repeat(1024);
+        let original = write_source(&set_disks, &disk_stores, bucket, object, &payload).await;
+        let tier_name = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let backend = register_mock_tier(&runtime_sources::global_tier_config_mgr(), &tier_name).await;
+        backend.set_put_remote_version(Some(String::new())).await;
+        let save_probe = TransitionUploadedSaveProbe::install(bucket, object);
+
+        set_disks
+            .transition_object(bucket, object, &transition_options(&original, tier_name))
+            .await
+            .expect("an unversioned remote version must commit");
+        let (fi, _, _) = set_disks
+            .get_object_fileinfo(
+                bucket,
+                object,
+                &ObjectOptions {
+                    no_lock: true,
+                    metadata_cache_safe: false,
+                    ..Default::default()
+                },
+                true,
+                false,
+            )
+            .await
+            .expect("committed unversioned transition metadata should be readable");
+        assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version, None);
+        assert_eq!(fi.transition_version_state, rustfs_filemeta::TransitionVersionState::KnownDisabled);
+        assert_eq!(save_probe.attempts(), 1);
+        assert_eq!(backend.remove_count().await, 0);
+        assert_eq!(backend.object_count().await, 1);
+    }
+
     #[tokio::test]
     #[serial_test::serial]
     async fn opaque_remote_version_is_cleaned_before_parse_failure() {
@@ -6630,7 +6957,7 @@ mod transition_upload_integrity_tests {
         set_disks
             .transition_object(bucket, object, &transition_options(&original, tier_name))
             .await
-            .expect_err("an unparseable remote version must fail closed");
+            .expect_err("an opaque remote version must fail closed until the capability gate is active");
         let removed_versions = backend.remove_versions().await;
         assert_eq!(removed_versions.len(), 1);
         assert_eq!(removed_versions[0].1, "opaque-version-token");
@@ -6638,6 +6965,38 @@ mod transition_upload_integrity_tests {
         assert_local_source_intact(&set_disks, bucket, object, &payload).await;
     }
 
+    #[tokio::test]
+    #[serial_test::serial]
+    async fn nil_remote_version_is_cleaned_exactly_before_transaction_persistence() {
+        let (_temp_dirs, disk_stores, set_disks) = hermetic_set_disks(4).await;
+        let bucket = "transition-nil-version-bucket";
+        let object = "object.bin";
+        let payload = b"nil remote version must retain local data".repeat(1024);
+        let original = write_source(&set_disks, &disk_stores, bucket, object, &payload).await;
+        let tier_name = format!("COLDTIER{}", &Uuid::new_v4().simple().to_string()[..8]).to_uppercase();
+        let remote_version = Uuid::nil().to_string();
+        let backend = register_mock_tier(&runtime_sources::global_tier_config_mgr(), &tier_name).await;
+        backend.set_put_remote_version(Some(remote_version.clone())).await;
+        let save_probe = TransitionUploadedSaveProbe::install(bucket, object);
+
+        set_disks
+            .transition_object(bucket, object, &transition_options(&original, tier_name))
+            .await
+            .expect_err("a nil remote version must fail closed before transaction persistence");
+        let put_versions = backend.put_versions().await;
+        let removed_versions = backend.remove_versions().await;
+        assert_eq!(removed_versions, put_versions);
+        assert_eq!(removed_versions.len(), 1);
+        assert_eq!(
+            removed_versions.first().map(|(_, version)| version.as_str()),
+            Some(remote_version.as_str())
+        );
+        assert_eq!(save_probe.attempts(), 0, "nil remote version must be rejected before saving Uploaded");
+        assert_eq!(backend.exact_remove_count(), 1);
+        assert_eq!(backend.object_count().await, 0);
+        assert_local_source_intact(&set_disks, bucket, object, &payload).await;
+    }
+
     #[tokio::test]
     #[serial_test::serial]
     async fn authoritative_read_failure_after_upload_cleans_exact_candidate_and_preserves_source() {
@@ -6957,7 +7316,6 @@ mod transition_source_identity_matrix_tests {
     async fn transition_source_identity_field_matrix_rejects_single_field_drift() {
         #[derive(Clone, Copy, Debug)]
         enum IdentityField {
-            VersionId,
             DataDir,
             ModTime,
             Size,
@@ -6973,7 +7331,6 @@ mod transition_source_identity_matrix_tests {
         let backend = register_mock_tier(&runtime_sources::global_tier_config_mgr(), &tier_name).await;
 
         for (index, field) in [
-            IdentityField::VersionId,
             IdentityField::DataDir,
             IdentityField::ModTime,
             IdentityField::Size,
@@ -6985,23 +7342,36 @@ mod transition_source_identity_matrix_tests {
             let object = format!("identity-{index}.bin");
             let payload = vec![u8::try_from(index + 1).expect("matrix index should fit u8"); 1024 * 1024];
             let mut reader = PutObjReader::from_vec(payload);
+            let source_version_id = Uuid::new_v4();
+            let source_opts = ObjectOptions {
+                version_id: Some(source_version_id.to_string()),
+                versioned: true,
+                ..Default::default()
+            };
             let original = set_disks
-                .put_object(bucket, &object, &mut reader, &ObjectOptions::default())
+                .put_object(bucket, &object, &mut reader, &source_opts)
                 .await
                 .expect("source object should be written");
             let (source, _, _) = set_disks
-                .get_object_fileinfo(bucket, &object, &ObjectOptions::default(), true, false)
+                .get_object_fileinfo(bucket, &object, &source_opts, true, false)
                 .await
                 .expect("source metadata should resolve");
+            assert_eq!(source.version_id, Some(source_version_id));
+            assert_eq!(
+                transition_source_identity(bucket, &object, &source, &source_opts, &get_raw_etag(&source.metadata))
+                    .expect("persisted versioned source identity should build")
+                    .version_mode,
+                TransitionSourceVersionMode::Versioned
+            );
             let opts = ObjectOptions {
                 no_lock: true,
+                versioned: true,
                 transition: TransitionOptions {
                     status: TRANSITION_PENDING.to_string(),
                     tier: tier_name.clone(),
                     etag: original.etag.clone().unwrap_or_default(),
                     ..Default::default()
                 },
-                version_id: original.version_id.map(|version| version.to_string()),
                 mod_time: original.mod_time,
                 ..Default::default()
             };
@@ -7014,7 +7384,6 @@ mod transition_source_identity_matrix_tests {
 
             let mut changed = source.clone();
             match field {
-                IdentityField::VersionId => changed.version_id = Some(Uuid::new_v4()),
                 IdentityField::DataDir => changed.data_dir = Some(Uuid::new_v4()),
                 IdentityField::ModTime => {
                     changed.mod_time = changed.mod_time.map(|value| value + time::Duration::nanoseconds(1));
@@ -7032,12 +7401,19 @@ mod transition_source_identity_matrix_tests {
                     .await
                     .expect("single-field metadata drift should be written");
             }
+            let persisted_opts = ObjectOptions {
+                version_id: changed.version_id.map(|version_id| version_id.to_string()),
+                versioned: true,
+                ..Default::default()
+            };
+            let (persisted, _, _) = set_disks
+                .get_object_fileinfo(bucket, &object, &persisted_opts, true, false)
+                .await
+                .expect("drifted source metadata should resolve");
             put_barrier.release();
 
-            transition
-                .await
-                .expect("transition task should not panic")
-                .expect_err("transition must reject a source whose identity changed after upload");
+            let result = transition.await.expect("transition task should not panic");
+            assert!(result.is_err(), "transition must reject {field:?} drift");
             let expected_attempts = index + 1;
             assert_eq!(backend.put_count().await, expected_attempts);
             assert_eq!(backend.remove_count().await, expected_attempts);
@@ -7048,26 +7424,23 @@ mod transition_source_identity_matrix_tests {
             );
 
             match field {
-                IdentityField::VersionId => assert_ne!(source.version_id, changed.version_id),
-                IdentityField::DataDir => assert_ne!(source.data_dir, changed.data_dir),
-                IdentityField::ModTime => assert_ne!(source.mod_time, changed.mod_time),
-                IdentityField::Size => assert_ne!(source.size, changed.size),
-                IdentityField::Etag => assert_ne!(get_raw_etag(&source.metadata), get_raw_etag(&changed.metadata)),
-            }
-            if !matches!(field, IdentityField::VersionId) {
-                assert_eq!(source.version_id, changed.version_id);
+                IdentityField::DataDir => assert_ne!(source.data_dir, persisted.data_dir),
+                IdentityField::ModTime => assert_ne!(source.mod_time, persisted.mod_time),
+                IdentityField::Size => assert_ne!(source.size, persisted.size),
+                IdentityField::Etag => assert_ne!(get_raw_etag(&source.metadata), get_raw_etag(&persisted.metadata)),
             }
+            assert_eq!(source.version_id, persisted.version_id);
             if !matches!(field, IdentityField::DataDir) {
-                assert_eq!(source.data_dir, changed.data_dir);
+                assert_eq!(source.data_dir, persisted.data_dir);
             }
             if !matches!(field, IdentityField::ModTime) {
-                assert_eq!(source.mod_time, changed.mod_time);
+                assert_eq!(source.mod_time, persisted.mod_time);
             }
             if !matches!(field, IdentityField::Size) {
-                assert_eq!(source.size, changed.size);
+                assert_eq!(source.size, persisted.size);
             }
             if !matches!(field, IdentityField::Etag) {
-                assert_eq!(get_raw_etag(&source.metadata), get_raw_etag(&changed.metadata));
+                assert_eq!(get_raw_etag(&source.metadata), get_raw_etag(&persisted.metadata));
             }
         }
     }

crates/ecstore/src/set_disk/replication.rs

@@ -13,7 +13,10 @@
 // limitations under the License.
 
 use super::*;
+use crate::bucket::lifecycle::lifecycle;
+use rustfs_filemeta::RestoreStatusOps;
 use rustfs_utils::http::headers::{AMZ_RESTORE_EXPIRY_DAYS, AMZ_RESTORE_REQUEST_DATE};
+use s3s::dto::{RestoreStatus, Timestamp};
 
 #[derive(Clone, Copy, Debug, Eq, PartialEq)]
 struct RestoreCleanupIdentity {
@@ -43,6 +46,69 @@ impl RestoreCleanupIdentity {
 }
 
 impl SetDisks {
+    pub(super) async fn finalize_restore_metadata(
+        &self,
+        bucket: &str,
+        object: &str,
+        obj_info: &ObjectInfo,
+        opts: &ObjectOptions,
+    ) -> Result<ObjectInfo> {
+        let expected = RestoreCleanupIdentity::from_object_info(obj_info);
+        let expected_operation_id = restore_operation_id_from_metadata(&opts.user_defined)?;
+        let expected_etag = obj_info
+            .etag
+            .clone()
+            .unwrap_or_else(|| get_raw_etag(obj_info.user_defined.as_ref()));
+        let version_id = expected.version_id.map(|v| v.to_string());
+        let _lock_guard = if !opts.no_lock {
+            Some(
+                self.acquire_write_lock_diag("restore_finalize_metadata", bucket, object)
+                    .await?,
+            )
+        } else {
+            None
+        };
+        let read_opts = ObjectOptions {
+            version_id,
+            versioned: opts.versioned,
+            version_suspended: opts.version_suspended,
+            ..Default::default()
+        };
+        let (mut fi, _, disks) = self
+            .get_object_fileinfo_gated(bucket, object, &read_opts, false, false)
+            .await?;
+        if let Some(expected_operation_id) = expected_operation_id {
+            require_restore_operation_id(&fi.metadata, expected_operation_id)?;
+        }
+        if !expected.matches_file_info(&fi, &expected_etag) {
+            return Err(Error::other("restored object changed before restore metadata finalization"));
+        }
+        let restore_expiry =
+            lifecycle::expected_expiry_time(OffsetDateTime::now_utc(), opts.transition.restore_request.days.unwrap_or(1));
+        fi.metadata.insert(
+            X_AMZ_RESTORE.as_str().to_string(),
+            RestoreStatus {
+                is_restore_in_progress: Some(false),
+                restore_expiry_date: Some(Timestamp::from(restore_expiry)),
+            }
+            .to_string(),
+        );
+        self.invalidate_get_object_metadata_cache(bucket, object).await;
+        self.update_object_meta_with_opts(
+            bucket,
+            object,
+            fi.clone(),
+            disks.as_slice(),
+            &UpdateMetadataOpts {
+                replace_user_metadata: true,
+                ..Default::default()
+            },
+        )
+        .await?;
+        self.invalidate_get_object_metadata_cache(bucket, object).await;
+        Ok(ObjectInfo::from_file_info(&fi, bucket, object, opts.versioned || opts.version_suspended))
+    }
+
     pub async fn update_restore_metadata(
         &self,
         bucket: &str,

crates/ecstore/src/set_disk/transition_matrix_tests.rs

@@ -13,10 +13,11 @@
 // limitations under the License.
 
 use super::*;
-use crate::bucket::lifecycle::lifecycle::{TRANSITION_COMPLETE, TRANSITION_PENDING, TransitionOptions};
+use crate::bucket::lifecycle::lifecycle::{TRANSITION_COMPLETE, TRANSITION_PENDING, TransitionOptions, expected_expiry_time};
 use crate::ecstore_validation_blackbox::make_local_set_disks;
 use crate::services::tier::test_util::register_mock_tier;
 use crate::storage_api_contracts::object::{ObjectIO as _, ObjectOperations as _};
+use rustfs_filemeta::{RestoreStatusOps as _, parse_restore_obj_status};
 use tokio::io::AsyncReadExt;
 
 async fn prime_metadata_generation(set_disks: &SetDisks, bucket: &str, object: &str) -> GetObjectMetadataCacheKey {
@@ -83,13 +84,57 @@ async fn transition_and_restore_reclaim_prior_metadata_generations() {
     let transitioned_generation = prime_metadata_generation(&set_disks, bucket, object).await;
     let mut restore_opts = ObjectOptions::default();
     restore_opts.transition.restore_request.days = Some(1);
-    Arc::clone(&set_disks)
-        .restore_transitioned_object(bucket, object, &restore_opts)
-        .await
-        .expect("restore should succeed");
+    let restore_started = OffsetDateTime::now_utc();
+    let expiry_from_restore_start = temp_env::async_with_vars(
+        [
+            ("RUSTFS_ILM_DEBUG_DAY_SECS", Some("1")),
+            ("RUSTFS_ILM_PROCESS_TIME", Some("1")),
+        ],
+        async {
+            let expiry_from_restore_start = expected_expiry_time(restore_started, 1);
+            let get_barrier = backend.arm_get_barrier().await;
+            let restore_set = Arc::clone(&set_disks);
+            let restore =
+                tokio::spawn(async move { restore_set.restore_transitioned_object(bucket, object, &restore_opts).await });
+            get_barrier.wait_until_paused().await;
+            tokio::time::timeout(Duration::from_secs(5), async {
+                loop {
+                    if expected_expiry_time(OffsetDateTime::now_utc(), 1) > expiry_from_restore_start {
+                        break;
+                    }
+                    tokio::time::sleep(Duration::from_millis(10)).await;
+                }
+            })
+            .await
+            .expect("test clock should cross the next accelerated lifecycle boundary");
+            get_barrier.release();
+            restore
+                .await
+                .expect("restore task should join")
+                .expect("restore should succeed");
+            expiry_from_restore_start
+        },
+    )
+    .await;
     assert_generation_reclaimed(&set_disks, &transitioned_generation).await;
     assert_eq!(backend.get_count().await, 1, "restore should read the remote candidate exactly once");
 
+    let restored_info = set_disks
+        .get_object_info(bucket, object, &ObjectOptions::default())
+        .await
+        .expect("restored object metadata should be readable");
+    let restore_status = parse_restore_obj_status(
+        restored_info
+            .user_defined
+            .get(s3s::header::X_AMZ_RESTORE.as_str())
+            .expect("completed restore header should be present"),
+    )
+    .expect("completed restore header should parse");
+    assert!(
+        restore_status.expiry().expect("completed restore should have an expiry") > expiry_from_restore_start,
+        "restore expiry must be based on completion, not the time the remote copy started"
+    );
+
     let mut restored = Vec::new();
     set_disks
         .get_object_reader(bucket, object, None, HeaderMap::new(), &ObjectOptions::default())

crates/ecstore/src/store/bucket.rs

@@ -87,7 +87,7 @@ fn bucket_deleted_marker_volume(bucket: &str) -> String {
     format!("{RUSTFS_META_BUCKET}/{}", bucket_deleted_marker_prefix(bucket))
 }
 
-async fn await_bucket_namespace_operation<T, F>(
+pub(crate) async fn await_bucket_namespace_operation<T, F>(
     guard: Option<&rustfs_lock::NamespaceLockGuard>,
     bucket: &str,
     operation: &'static str,
@@ -568,6 +568,7 @@ mod tests {
     };
     use crate::bucket::metadata::table_bucket_catalog_metadata_prefix;
     use crate::bucket::metadata_sys;
+    use crate::cluster::rpc::peer_s3_client::install_delete_bucket_empty_scan_barrier;
     use crate::disk::{BUCKET_META_PREFIX, RUSTFS_META_BUCKET};
     use crate::error::StorageError;
     use crate::object_api::{ObjectOptions, PutObjReader};
@@ -589,6 +590,7 @@ mod tests {
     use std::sync::atomic::{AtomicBool, Ordering};
     use std::time::{Duration, SystemTime};
     use time::OffsetDateTime;
+    use tokio::io::AsyncReadExt;
     use tokio::sync::{Notify, OnceCell};
     use tokio_util::sync::CancellationToken;
     use uuid::Uuid;
@@ -1048,7 +1050,7 @@ mod tests {
             .await
             .expect("bucket should be created");
         for disk_path in &disk_paths {
-            tokio::fs::create_dir_all(disk_path.join(&bucket).join("empty-directory"))
+            tokio::fs::create_dir_all(disk_path.join(&bucket).join("empty-directory/nested/leaf"))
                 .await
                 .expect("empty directory remnant should be created");
         }
@@ -1257,6 +1259,55 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    #[serial]
+    async fn bucket_delete_preserves_put_committed_after_empty_scan() {
+        let (_, ecstore) = setup_bucket_delete_test_env().await;
+        let bucket = format!("bucket-delete-empty-scan-race-{}", Uuid::new_v4().simple());
+        let object = "committed-after-empty-scan";
+        let payload = b"object committed after DeleteBucket empty scan".to_vec();
+
+        ecstore
+            .make_bucket(&bucket, &MakeBucketOptions::default())
+            .await
+            .expect("bucket should be created");
+
+        let barrier = install_delete_bucket_empty_scan_barrier();
+        let delete_store = ecstore.clone();
+        let delete_bucket = bucket.clone();
+        let delete = tokio::spawn(async move {
+            delete_store
+                .delete_bucket(&delete_bucket, &DeleteBucketOptions::default())
+                .await
+        });
+        barrier.wait_until_paused().await;
+
+        let mut put_reader = PutObjReader::from_vec(payload.clone());
+        ecstore
+            .put_object(&bucket, object, &mut put_reader, &ObjectOptions::default())
+            .await
+            .expect("PUT should commit while DeleteBucket is paused after its empty scan");
+
+        barrier.release();
+        let err = delete
+            .await
+            .expect("DeleteBucket task should join")
+            .expect_err("DeleteBucket must reject a PUT committed after its empty scan");
+        assert!(matches!(err, StorageError::BucketNotEmpty(name) if name == bucket));
+
+        let mut reader = ecstore
+            .get_object_reader(&bucket, object, None, http::HeaderMap::new(), &ObjectOptions::default())
+            .await
+            .expect("committed object should remain readable after DeleteBucket fails");
+        let mut restored = Vec::new();
+        reader
+            .stream
+            .read_to_end(&mut restored)
+            .await
+            .expect("object body should remain readable");
+        assert_eq!(restored, payload);
+    }
+
     #[tokio::test]
     #[serial]
     async fn bucket_recreation_does_not_publish_unverified_usage() {

crates/ecstore/src/store/init.rs

@@ -562,10 +562,12 @@ mod tests {
             },
             tier_sweeper::Jentry,
             transition_transaction::{
-                TRANSITION_TRANSACTION_RECORD_PREFIX, TransitionCleanupDecision, TransitionCleanupProof, TransitionRemoteVersion,
-                TransitionSourceIdentity, TransitionSourceVersionMode, TransitionTransaction, TransitionTransactionInit,
-                TransitionTransactionState, load_transition_transaction_record, recover_transition_transaction_records,
-                save_transition_transaction_record,
+                TRANSITION_TRANSACTION_RECORD_PREFIX, TransitionCleanupDecision, TransitionCleanupProof, TransitionOperatorError,
+                TransitionOperatorProbe, TransitionRemoteVersion, TransitionSourceIdentity, TransitionSourceVersionMode,
+                TransitionTransaction, TransitionTransactionInit, TransitionTransactionState,
+                delete_transition_candidate_for_operator, finalize_missing_transition_transaction_for_operator,
+                inspect_transition_transaction_for_operator, load_transition_transaction_record,
+                recover_transition_transaction_records, save_transition_transaction_record,
             },
         },
         client::transition_api::ReaderImpl,
@@ -575,6 +577,7 @@ mod tests {
         services::tier::{
             test_util::{MockWarmBackend, MockWarmOp, TransitionCleanupStoreBarrier, register_mock_tier},
             tier::{TIER_CONFIG_FILE, TierConfigMgr},
+            tier_config::{TierConfig, TierType, TierWasabi},
             tier_mutation_intent::{
                 TIER_MUTATION_INTENT_RECORD_PREFIX, TierMutationIntent, TierMutationIntentKind, TierMutationIntentState,
                 TierMutationIntentTarget, advance_tier_mutation_intent_record_idempotent, delete_tier_mutation_intent_record,
@@ -1163,6 +1166,37 @@ mod tests {
             .len()
     }
 
+    #[cfg(feature = "test-util")]
+    async fn register_transition_reconcile_test_tier(
+        handle: &Arc<tokio::sync::RwLock<TierConfigMgr>>,
+        tier_name: &str,
+    ) -> MockWarmBackend {
+        let backend = MockWarmBackend::new();
+        let mut manager = handle.write().await;
+        manager.tiers.insert(
+            tier_name.to_string(),
+            TierConfig {
+                version: "v1".to_string(),
+                tier_type: TierType::Wasabi,
+                name: tier_name.to_string(),
+                wasabi: Some(TierWasabi {
+                    name: tier_name.to_string(),
+                    endpoint: "https://s3.wasabisys.com".to_string(),
+                    access_key: "test-access-key".to_string(),
+                    secret_key: "test-secret-key".to_string(),
+                    bucket: "mock-tier".to_string(),
+                    prefix: format!("mock/{}/", uuid::Uuid::new_v4()),
+                    region: "us-east-1".to_string(),
+                }),
+                ..Default::default()
+            },
+        );
+        manager
+            .install_test_driver(tier_name, Box::new(backend.clone()))
+            .expect("mock fallback tier driver should install");
+        backend
+    }
+
     #[cfg(feature = "test-util")]
     async fn wait_for_tier_delete_journal_recovery(
         store: Arc<crate::store::ECStore>,
@@ -1311,14 +1345,16 @@ mod tests {
             version_id: "version-a".to_string(),
             tier_name: tier_a.to_string(),
             backend_identity: Some(identity_a),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         };
         let entry_b = Jentry {
             obj_name: "remote-b".to_string(),
             version_id: "version-b".to_string(),
             tier_name: tier_b.to_string(),
             backend_identity: Some(identity_b),
-            version_id_exact: false,
+            version_id_exact: true,
+            version_state: rustfs_filemeta::TransitionVersionState::Exact,
         };
         let remove_a = backend_a.arm_failing_remove_barrier().await;
         persist_tier_delete_journal_entry(store_a.clone(), &entry_a)
@@ -2720,10 +2756,12 @@ mod tests {
     #[serial_test::serial(storage_class_env)]
     async fn transition_transaction_recovery_deletes_provider_recovered_unknown_upload() {
         let versioned_remote = uuid::Uuid::new_v4().to_string();
+        let nil_remote = uuid::Uuid::nil().to_string();
         for (case, tier_name, remote_version) in [
             ("missing", "TXPROBEMISSING", None),
             ("unversioned", "TXPROBEUNVERSIONED", Some(String::new())),
             ("versioned", "TXPROBEVERSIONED", Some(versioned_remote)),
+            ("nil-version", "TXPROBENILVERSION", Some(nil_remote)),
         ] {
             let temp_dir = tempfile::tempdir().expect("create temp store dir");
             let (ctx, store, _shutdown) = without_storage_class_env(build_isolated_test_store(
@@ -2734,7 +2772,7 @@ mod tests {
             .await;
             crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
 
-            let backend = register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
+            let backend = register_transition_reconcile_test_tier(&ctx.tier_config_mgr(), tier_name).await;
             let backend_identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
                 .await
                 .expect("tier lease should resolve")
@@ -2805,6 +2843,157 @@ mod tests {
         }
     }
 
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn operator_reconcile_deletes_exact_candidate_before_finalizing_record() {
+        let temp_dir = tempfile::tempdir().expect("create temp store dir");
+        let (ctx, store, _shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "transition-operator-reconcile", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+        let tier_name = "TXOPERATOR";
+        let backend = register_transition_reconcile_test_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let backend_identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
+            .await
+            .expect("tier lease should resolve")
+            .backend_identity();
+        let mut transaction = TransitionTransaction::new(TransitionTransactionInit {
+            deployment_id: ctx.deployment_id().expect("test store should initialize deployment id"),
+            transaction_id: uuid::Uuid::new_v4(),
+            owner_epoch: uuid::Uuid::new_v4(),
+            write_id: uuid::Uuid::new_v4(),
+            source: TransitionSourceIdentity {
+                bucket: "source-bucket".to_string(),
+                object: "source-object".to_string(),
+                version_id: None,
+                data_dir: uuid::Uuid::new_v4(),
+                mod_time_unix_nanos: 1_770_000_000_000_000_000,
+                size: 42,
+                etag: "source-etag".to_string(),
+                version_mode: TransitionSourceVersionMode::Unversioned,
+            },
+            tier_name: tier_name.to_string(),
+            backend_fingerprint: backend_identity,
+            not_after_unix_nanos: 1,
+        })
+        .expect("transaction should build");
+        transaction
+            .advance(transaction.fence(), TransitionTransactionState::UploadOutcomeUnknown, None)
+            .expect("transaction should enter unknown upload outcome state");
+
+        let remote_version = uuid::Uuid::new_v4().to_string();
+        backend.set_put_remote_version(Some(remote_version.clone())).await;
+        let candidate = bytes::Bytes::from_static(b"operator-confirmed transition candidate");
+        backend
+            .put(
+                &transaction.remote_object,
+                ReaderImpl::Body(candidate.clone()),
+                i64::try_from(candidate.len()).expect("test candidate length should fit i64"),
+            )
+            .await
+            .expect("mock backend should accept candidate");
+        save_transition_transaction_record(store.clone(), &transaction)
+            .await
+            .expect("transaction record should persist");
+
+        let status = inspect_transition_transaction_for_operator(store.clone(), transaction.transaction_id)
+            .await
+            .expect("operator inspection should probe the candidate");
+        assert_eq!(status.probe, TransitionOperatorProbe::VersionedPresent(remote_version.clone()));
+
+        let wrong_version = uuid::Uuid::new_v4().to_string();
+        let err = delete_transition_candidate_for_operator(store.clone(), transaction.transaction_id, &wrong_version)
+            .await
+            .expect_err("a mismatched exact version must fail before deleting a candidate");
+        assert!(matches!(
+            err,
+            TransitionOperatorError::CandidateVersionMismatch {
+                expected,
+                actual: TransitionOperatorProbe::VersionedPresent(ref observed),
+            } if expected == wrong_version && observed == &remote_version
+        ));
+        assert!(backend.contains(&transaction.remote_object).await);
+        assert_eq!(backend.exact_remove_count(), 0);
+        load_transition_transaction_record(store.clone(), transaction.transaction_id)
+            .await
+            .expect("an incorrect exact version must retain the transaction journal");
+
+        let result = delete_transition_candidate_for_operator(store.clone(), transaction.transaction_id, &remote_version)
+            .await
+            .expect("operator-confirmed exact candidate should be deleted");
+        assert_eq!(result.status.probe, TransitionOperatorProbe::Missing);
+        assert!(result.journal_observed_after_delete);
+        assert_eq!(backend.exact_remove_count(), 1);
+        assert_eq!(backend.remove_versions().await, vec![(transaction.remote_object.clone(), remote_version)]);
+        load_transition_transaction_record(store.clone(), transaction.transaction_id)
+            .await
+            .expect("candidate deletion must retain the transaction journal");
+
+        finalize_missing_transition_transaction_for_operator(store.clone(), transaction.transaction_id)
+            .await
+            .expect("a separately confirmed missing candidate should permit finalization");
+        assert!(matches!(
+            load_transition_transaction_record(store, transaction.transaction_id).await,
+            Err(Error::ConfigNotFound)
+        ));
+    }
+
+    #[cfg(feature = "test-util")]
+    #[tokio::test]
+    #[serial_test::serial(storage_class_env)]
+    async fn operator_finalize_retains_record_without_missing_proof() {
+        let temp_dir = tempfile::tempdir().expect("create temp store dir");
+        let (ctx, store, _shutdown) =
+            without_storage_class_env(build_isolated_test_store(temp_dir.path(), "transition-operator-fail-closed", &[4])).await;
+        crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
+
+        let tier_name = "TXOPERATORFAIL";
+        let backend = register_transition_reconcile_test_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let backend_identity = TierConfigMgr::acquire_operation_lease(&ctx.tier_config_mgr(), tier_name)
+            .await
+            .expect("tier lease should resolve")
+            .backend_identity();
+        let mut transaction = TransitionTransaction::new(TransitionTransactionInit {
+            deployment_id: ctx.deployment_id().expect("test store should initialize deployment id"),
+            transaction_id: uuid::Uuid::new_v4(),
+            owner_epoch: uuid::Uuid::new_v4(),
+            write_id: uuid::Uuid::new_v4(),
+            source: TransitionSourceIdentity {
+                bucket: "source-bucket".to_string(),
+                object: "source-object".to_string(),
+                version_id: None,
+                data_dir: uuid::Uuid::new_v4(),
+                mod_time_unix_nanos: 1_770_000_000_000_000_000,
+                size: 42,
+                etag: "source-etag".to_string(),
+                version_mode: TransitionSourceVersionMode::Unversioned,
+            },
+            tier_name: tier_name.to_string(),
+            backend_fingerprint: backend_identity,
+            not_after_unix_nanos: 1,
+        })
+        .expect("transaction should build");
+        transaction
+            .advance(transaction.fence(), TransitionTransactionState::UploadOutcomeUnknown, None)
+            .expect("transaction should enter unknown upload outcome state");
+        save_transition_transaction_record(store.clone(), &transaction)
+            .await
+            .expect("transaction record should persist");
+        backend
+            .set_transition_candidate_probe_override(Some(TransitionCandidateProbe::Unsupported))
+            .await;
+
+        assert!(matches!(
+            finalize_missing_transition_transaction_for_operator(store.clone(), transaction.transaction_id).await,
+            Err(TransitionOperatorError::CandidateNotMissing(TransitionOperatorProbe::Unsupported))
+        ));
+        load_transition_transaction_record(store, transaction.transaction_id)
+            .await
+            .expect("an unsupported probe must retain the transaction journal");
+        assert_eq!(backend.remove_count().await, 0);
+    }
+
     #[cfg(feature = "test-util")]
     #[tokio::test]
     #[serial_test::serial(storage_class_env)]
@@ -2815,7 +3004,7 @@ mod tests {
         crate::bucket::metadata_sys::init_bucket_metadata_sys(store.clone(), Vec::new()).await;
 
         let tier_name = "TXRESPONSELOSS";
-        let backend = register_mock_tier(&ctx.tier_config_mgr(), tier_name).await;
+        let backend = register_transition_reconcile_test_tier(&ctx.tier_config_mgr(), tier_name).await;
         let bucket = "transition-response-loss-bucket";
         let object = "source.bin";
         store

crates/ecstore/src/store/list_objects.rs

@@ -6895,6 +6895,7 @@ mod test {
         )
         .await
         .expect("local disk should be created");
+        disk.make_volume("bucket").await.expect("test bucket should be created");
         let mut fi = FileInfo::new("object", 1, 1);
         fi.volume = "bucket".to_owned();
         fi.name = "object".to_owned();

crates/ecstore/src/store/mod.rs

@@ -141,6 +141,7 @@ fn should_enqueue_transition_immediately(oi: &ObjectInfo) -> bool {
 const MAX_UPLOADS_LIST: usize = 10000;
 
 mod bucket;
+pub(crate) use bucket::await_bucket_namespace_operation;
 mod heal;
 mod heal_walk;
 pub use heal_walk::HealWalkVersion;

crates/ecstore/src/test_tracing.rs

@@ -0,0 +1,43 @@
+//! Test-only guard against tracing's process-global callsite-interest cache.
+//!
+//! Any test that installs its own subscriber and then asserts on span or event
+//! context is asserting on process-global state. See
+//! [`pin_callsite_interest_for_test`] for what goes wrong and why holding its
+//! guard fixes it.
+
+/// Stop a sibling test thread from poisoning tracing's process-global callsite
+/// interest cache while this test asserts on span or event context.
+///
+/// `tracing` caches every callsite's `Interest` in **process-global** state, and
+/// the first thread to reach a callsite fixes that value. While at most one
+/// dispatcher is registered, tracing-core takes a fast path
+/// (`Dispatchers::rebuilder` -> `Rebuilder::JustOne`) that derives a
+/// newly-registered callsite's interest from whichever subscriber is current
+/// *on the registering thread*, and registration happens exactly once (guarded
+/// by a CAS in `DefaultCallsite::register`).
+///
+/// In a multi-threaded `cargo test` binary a sibling test therefore routinely
+/// reaches a production callsite first, from a thread with no subscriber
+/// installed: the interest is derived from that thread's `NoSubscriber` and
+/// cached as `Interest::never()` for the whole process. From then on the
+/// callsite is dead for *every* later caller, including a test that installed
+/// its own subscriber — an `info_span!` silently evaluates to `Span::none()`, and
+/// a `warn!`/`info!` event never fires at all.
+///
+/// Registering a second, inert dispatcher closes both halves of that race:
+///
+/// * constructing it rebuilds every *already-registered* callsite's interest
+///   against the live dispatcher set — which includes the caller's subscriber —
+///   repairing whatever a sibling may already have poisoned; and
+/// * holding it alive keeps tracing-core off the single-dispatcher fast path, so
+///   a callsite registered *later* by any thread is resolved against that live
+///   set instead of the registering thread's `NoSubscriber`.
+///
+/// Call this **after** installing the test's subscriber, and hold the returned
+/// guard for the rest of the test. Only the `cargo test` fallback needs it:
+/// nextest runs each test in its own process, where there is no sibling thread
+/// to lose the race to (see `docs/testing/README.md`).
+#[must_use = "callsite interest is only pinned while the returned guard is held"]
+pub(crate) fn pin_callsite_interest_for_test() -> tracing::Dispatch {
+    tracing::Dispatch::new(tracing_core::subscriber::NoSubscriber::default())
+}

crates/filemeta/src/fileinfo.rs

@@ -219,6 +219,16 @@ impl ErasureInfo {
 }
 
 // #[derive(Debug, Clone)]
+#[derive(Serialize, Deserialize, Debug, PartialEq, Eq, Clone, Copy, Default)]
+#[serde(rename_all = "kebab-case")]
+pub enum TransitionVersionState {
+    #[default]
+    Unknown,
+    KnownDisabled,
+    SuspendedNull,
+    Exact,
+}
+
 #[derive(Serialize, Deserialize, Debug, PartialEq, Clone, Default)]
 pub struct FileInfo {
     pub volume: String,
@@ -230,6 +240,10 @@ pub struct FileInfo {
     pub transitioned_objname: String,
     pub transition_tier: String,
     pub transition_version_id: Option<Uuid>,
+    #[serde(default)]
+    pub transition_version: Option<String>,
+    #[serde(default)]
+    pub transition_version_state: TransitionVersionState,
     pub expire_restored: bool,
     pub data_dir: Option<Uuid>,
     pub mod_time: Option<OffsetDateTime>,
@@ -459,6 +473,10 @@ impl FileInfo {
         if self.mod_time.is_none_or(|mod_time| mod_time <= OffsetDateTime::UNIX_EPOCH)
             || (!allow_nil_version_id && self.version_id.is_some_and(|version_id| version_id.is_nil()))
             || self.transition_version_id.is_some_and(|version_id| version_id.is_nil())
+            || self
+                .transition_version
+                .as_ref()
+                .is_some_and(|version_id| version_id.is_empty())
             || self.size != 0
             || self.data_dir.is_some()
             || self.mode.is_some()
@@ -492,6 +510,7 @@ impl FileInfo {
             || !self.transitioned_objname.is_empty()
             || !self.transition_tier.is_empty()
             || self.transition_version_id.is_some()
+            || self.transition_version.is_some()
             || self.expire_restored
             || self.size != 0
             || self.data_dir.is_some()
@@ -536,6 +555,25 @@ impl FileInfo {
     /// return `None`.
     pub fn validate(&self, mode: ValidationMode) -> Result<Option<ValidatedErasureLayout>> {
         self.validate_collection_bounds()?;
+        if let (Some(version), Some(version_id)) = (&self.transition_version, self.transition_version_id)
+            && Uuid::parse_str(version).ok() != Some(version_id)
+        {
+            return Err(Error::FileCorrupt);
+        }
+        let transition_state_valid = match self.transition_version_state {
+            TransitionVersionState::Unknown => true,
+            TransitionVersionState::KnownDisabled => self.transition_version.is_none() && self.transition_version_id.is_none(),
+            TransitionVersionState::SuspendedNull => {
+                self.transition_version.as_deref() == Some("null") && self.transition_version_id.is_none()
+            }
+            TransitionVersionState::Exact => self
+                .transition_version
+                .as_deref()
+                .is_some_and(|version| version != "null" && !version.is_empty()),
+        };
+        if !transition_state_valid {
+            return Err(Error::FileCorrupt);
+        }
 
         let erasure_layout = match mode {
             ValidationMode::RequireErasure => Some(self.validate_erasure_geometry()?),
@@ -832,6 +870,8 @@ impl FileInfo {
             && self.transition_tier == other.transition_tier
             && self.transitioned_objname == other.transitioned_objname
             && self.transition_version_id == other.transition_version_id
+            && self.transition_version == other.transition_version
+            && self.transition_version_state == other.transition_version_state
     }
 
     /// Check if metadata maps are equal
@@ -1351,6 +1391,15 @@ mod tests {
         assert_file_corrupt(&fi, ValidationMode::DeleteOnly);
     }
 
+    #[test]
+    fn metadata_read_validation_rejects_conflicting_transition_versions() {
+        let mut fi = one_shard_validation_fileinfo(1);
+        fi.transition_version_id = Some(Uuid::new_v4());
+        fi.transition_version = Some(Uuid::new_v4().to_string());
+
+        assert_file_corrupt(&fi, ValidationMode::RequireErasure);
+    }
+
     #[test]
     fn metadata_read_validation_requires_canonical_delete_marker_shape() {
         let marker = FileInfo {
@@ -1722,6 +1771,12 @@ mod tests {
                     transitioned_objname,
                     transition_tier,
                     transition_version_id,
+                    transition_version: transition_version_id.map(|version_id| version_id.to_string()),
+                    transition_version_state: if transition_version_id.is_some() {
+                        TransitionVersionState::Exact
+                    } else {
+                        TransitionVersionState::Unknown
+                    },
                     expire_restored,
                     data_dir,
                     mod_time,

crates/filemeta/src/filemeta.rs

@@ -356,6 +356,14 @@ impl FileMeta {
             .versions
             .partition_point(|existing| existing.header.sorts_before(&new_shallow.header));
         self.versions.insert(insert_pos, new_shallow);
+        // `partition_point` only returns the canonical slot when `versions` is
+        // already canonically ordered, and nothing establishes that: `FileMeta::load`
+        // replays the on-disk order verbatim, and metadata written before the
+        // canonical-order fix ordered equal-`mod_time` ties by insertion rather than
+        // by `sorts_before`. Re-sort so an insert leaves the list canonical either
+        // way, matching what `set_idx` already does on the replace path. Cheap: after
+        // a correctly placed insert the slice is a single sorted run.
+        self.sort_by_mod_time();
         Ok(())
 
         // if !ver.valid() {
@@ -1215,6 +1223,57 @@ mod test {
         assert!(fm.versions[0].header.sorts_before(&fm.versions[1].header));
     }
 
+    /// `add_version_filemata` positions an inserted version with
+    /// `partition_point(sorts_before)`, which only yields the canonical slot when
+    /// `versions` is already canonically ordered. Nothing establishes that
+    /// precondition on load: `FileMeta::unmarshal_msg` pushes versions in file
+    /// order, and metadata written before the canonical-order fix ordered
+    /// equal-`mod_time` ties by insertion instead of by `sorts_before`. An insert
+    /// into such a list must still leave it canonical, the same way `set_idx`
+    /// already re-sorts on the replace path.
+    #[test]
+    fn add_version_filemata_canonicalizes_versions_loaded_out_of_order() {
+        let mod_time = OffsetDateTime::from_unix_timestamp(1_700_000_000).expect("valid test timestamp");
+        let first = version_for_ordering(VersionType::Object, Uuid::from_u128(70), mod_time, 70);
+        let second = version_for_ordering(VersionType::Object, Uuid::from_u128(80), mod_time, 80);
+        let (early, late) = if first.header().sorts_before(&second.header()) {
+            (first, second)
+        } else {
+            (second, first)
+        };
+
+        // Persist the pair the wrong way round to emulate a pre-canonical-order
+        // xl.meta, bypassing add_version_filemata so the bad order really lands on
+        // the wire.
+        let mut legacy = FileMeta::new();
+        legacy
+            .versions
+            .push(FileMetaShallowVersion::try_from(late).expect("shallow later version"));
+        legacy
+            .versions
+            .push(FileMetaShallowVersion::try_from(early).expect("shallow earlier version"));
+
+        let mut loaded =
+            FileMeta::load(&legacy.marshal_msg().expect("serialize legacy metadata")).expect("reload legacy metadata");
+        assert!(
+            !loaded.versions[0].header.sorts_before(&loaded.versions[1].header),
+            "fixture must reach add_version_filemata non-canonically ordered"
+        );
+
+        loaded
+            .add_version_filemata(version_for_ordering(VersionType::Delete, Uuid::from_u128(90), mod_time, 90))
+            .expect("insert equal-time delete marker");
+
+        assert_eq!(loaded.versions.len(), 3);
+        assert!(
+            loaded
+                .versions
+                .windows(2)
+                .all(|pair| pair[0].header.sorts_before(&pair[1].header)),
+            "insert must leave the version list canonically ordered"
+        );
+    }
+
     /// Regression for backlog#799 B16: replication reset state must be
     /// persisted under both internal prefixes, never as a bare ARN. A bare-ARN
     /// key (produced by `ObjectInfo::replication_state`) has no internal prefix,

crates/filemeta/src/filemeta/version.rs

@@ -26,13 +26,14 @@ use super::msgp_decode::{
     PrependByteReader, prealloc_hint, read_exact_vec, read_nil_or_array_len, read_nil_or_map_len, skip_msgp_value,
 };
 use super::*;
-use crate::ChecksumInfo;
+use crate::{ChecksumInfo, TransitionVersionState};
 use rustfs_utils::HashAlgorithm;
 use rustfs_utils::http::{
     RUSTFS_INTERNAL_PREFIX, SUFFIX_CRC, SUFFIX_FREE_VERSION, SUFFIX_INLINE_DATA, SUFFIX_PURGESTATUS, SUFFIX_TIER_FV_ID,
     SUFFIX_TIER_FV_MARKER, SUFFIX_TRANSITION_STATUS, SUFFIX_TRANSITION_TIER, SUFFIX_TRANSITION_TIER_DESTINATION_ID,
-    SUFFIX_TRANSITIONED_OBJECTNAME, SUFFIX_TRANSITIONED_VERSION_ID, contains_key_bytes, get_bytes, get_consistent_bytes, get_str,
-    has_internal_suffix, insert_bytes, is_internal_key, remove_bytes, strip_internal_prefix,
+    SUFFIX_TRANSITIONED_OBJECTNAME, SUFFIX_TRANSITIONED_VERSION_ID, SUFFIX_TRANSITIONED_VERSION_STATE, contains_key_bytes,
+    get_bytes, get_consistent_bytes, get_str, has_internal_suffix, insert_bytes, is_internal_key, remove_bytes,
+    strip_internal_prefix,
 };
 
 const MSGPACK_EXT8: u8 = 0xc7;
@@ -43,6 +44,7 @@ const MSGPACK_FIXEXT8: u8 = 0xd7;
 const MSGPACK_TIME_EXT_LEGACY: i8 = 5;
 const MSGPACK_TIME_EXT_OFFICIAL: i8 = -1;
 const MSGPACK_TIME_LEN: u8 = 12;
+const MAX_TRANSITION_VERSION_LEN: usize = 1024;
 
 /// Sentinel signature returned when a version has no computable body (invalid /
 /// missing inner object). Mirrors MinIO's `signatureErr` so such versions never
@@ -251,23 +253,93 @@ fn parse_legacy_uuid_bytes(bytes: &[u8], field: &str) -> Result<Option<Uuid>> {
 
 /// Decode a stored transitioned-version-id from a version's `meta_sys`.
 ///
-/// RustFS writes it as 16 raw UUID bytes; MinIO-migrated tiered objects store
-/// the remote tier's version id as a UUID *string*. Accept both, and treat any
-/// absent / nil / otherwise-unparseable value as "no tier version" (matching the
-/// tolerant pre-hardening behavior) rather than failing the whole object read —
-/// a malformed tier id must not make an otherwise-readable object unreadable.
-fn transitioned_version_id_from_meta_sys(meta_sys: &HashMap<String, Vec<u8>>) -> Option<Uuid> {
-    let value = get_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_ID)?;
+/// Legacy RustFS writes used 16 raw UUID bytes. New writes and MinIO-migrated
+/// records use the provider's exact UTF-8 version text. Empty, nil UUID, and
+/// malformed bytes are not usable remote versions.
+fn transitioned_version_from_meta_sys(meta_sys: &HashMap<String, Vec<u8>>) -> Result<Option<String>> {
+    if !contains_key_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_ID) {
+        return Ok(None);
+    }
+    let Some(value) = get_consistent_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_ID) else {
+        return Ok(None);
+    };
+    let value = value.to_vec();
     if value.is_empty() {
-        return None;
+        return Ok(None);
     }
     if let Ok(id) = Uuid::from_slice(&value) {
-        return (!id.is_nil()).then_some(id);
+        return Ok((!id.is_nil()).then(|| id.to_string()));
     }
-    std::str::from_utf8(&value)
+    let Ok(value) = String::from_utf8(value) else {
+        return Ok(None);
+    };
+    if value.is_empty()
+        || value.len() > MAX_TRANSITION_VERSION_LEN
+        || value.chars().any(char::is_control)
+        || Uuid::parse_str(&value).is_ok_and(|id| id.is_nil())
+    {
+        Ok(None)
+    } else {
+        Ok(Some(value))
+    }
+}
+
+fn transition_version_state_from_meta_sys(
+    meta_sys: &HashMap<String, Vec<u8>>,
+    version: Option<&str>,
+) -> Result<TransitionVersionState> {
+    if !contains_key_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE) {
+        return Ok(TransitionVersionState::Unknown);
+    }
+    let value = get_consistent_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE).ok_or(Error::FileCorrupt)?;
+    let state = match value {
+        b"known-disabled" => TransitionVersionState::KnownDisabled,
+        b"suspended-null" => TransitionVersionState::SuspendedNull,
+        b"exact" => TransitionVersionState::Exact,
+        b"unknown" => TransitionVersionState::Unknown,
+        _ => return Err(Error::FileCorrupt),
+    };
+    let valid = match state {
+        TransitionVersionState::Unknown | TransitionVersionState::KnownDisabled => version.is_none(),
+        TransitionVersionState::SuspendedNull => version == Some("null"),
+        TransitionVersionState::Exact => version.is_some_and(|value| value != "null"),
+    };
+    valid.then_some(state).ok_or(Error::FileCorrupt)
+}
+
+fn transition_version_state_bytes(state: TransitionVersionState) -> &'static [u8] {
+    match state {
+        TransitionVersionState::Unknown => b"unknown",
+        TransitionVersionState::KnownDisabled => b"known-disabled",
+        TransitionVersionState::SuspendedNull => b"suspended-null",
+        TransitionVersionState::Exact => b"exact",
+    }
+}
+
+fn set_transition_version_state(meta_sys: &mut HashMap<String, Vec<u8>>, state: TransitionVersionState) {
+    if state == TransitionVersionState::Unknown {
+        remove_bytes(meta_sys, SUFFIX_TRANSITIONED_VERSION_STATE);
+    } else {
+        insert_bytes(
+            meta_sys,
+            SUFFIX_TRANSITIONED_VERSION_STATE,
+            transition_version_state_bytes(state).to_vec(),
+        );
+    }
+}
+
+fn legacy_transitioned_version_id_from_meta_sys(meta_sys: &HashMap<String, Vec<u8>>) -> Option<Uuid> {
+    transitioned_version_from_meta_sys(meta_sys)
         .ok()
-        .and_then(|s| Uuid::parse_str(s.trim()).ok())
-        .filter(|id| !id.is_nil())
+        .flatten()
+        .and_then(|value| Uuid::parse_str(&value).ok())
+}
+
+fn transitioned_version_bytes(fi: &FileInfo) -> Option<Vec<u8>> {
+    fi.transition_version
+        .as_ref()
+        .map(|version| version.as_bytes().to_vec())
+        .or_else(|| fi.transition_version_id.map(|version_id| version_id.as_bytes().to_vec()))
 }
 
 fn parse_legacy_erasure_algo(value: &str) -> ErasureAlgo {
@@ -2398,7 +2470,9 @@ impl MetaObject {
         let transitioned_objname = get_bytes(&self.meta_sys, SUFFIX_TRANSITIONED_OBJECTNAME)
             .map(|v| String::from_utf8_lossy(&v).to_string())
             .unwrap_or_default();
-        let transition_version_id = transitioned_version_id_from_meta_sys(&self.meta_sys);
+        let transition_version = transitioned_version_from_meta_sys(&self.meta_sys)?;
+        let transition_version_state = transition_version_state_from_meta_sys(&self.meta_sys, transition_version.as_deref())?;
+        let transition_version_id = transition_version.as_deref().and_then(|value| Uuid::parse_str(value).ok());
         let transition_tier = get_bytes(&self.meta_sys, SUFFIX_TRANSITION_TIER)
             .map(|v| String::from_utf8_lossy(&v).to_string())
             .unwrap_or_default();
@@ -2419,6 +2493,8 @@ impl MetaObject {
             transition_status,
             transitioned_objname,
             transition_version_id,
+            transition_version,
+            transition_version_state,
             transition_tier,
             ..Default::default()
         })
@@ -2431,13 +2507,12 @@ impl MetaObject {
             SUFFIX_TRANSITIONED_OBJECTNAME,
             fi.transitioned_objname.as_bytes().to_vec(),
         );
-        if let Some(transition_version_id) = fi.transition_version_id.as_ref() {
-            insert_bytes(
-                &mut self.meta_sys,
-                SUFFIX_TRANSITIONED_VERSION_ID,
-                transition_version_id.as_bytes().to_vec(),
-            );
+        if let Some(transition_version) = transitioned_version_bytes(fi) {
+            insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
+        } else {
+            remove_bytes(&mut self.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID);
         }
+        set_transition_version_state(&mut self.meta_sys, fi.transition_version_state);
         insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITION_TIER, fi.transition_tier.as_bytes().to_vec());
         if let Some(destination_id) = get_str(&fi.metadata, SUFFIX_TRANSITION_TIER_DESTINATION_ID) {
             insert_bytes(&mut self.meta_sys, SUFFIX_TRANSITION_TIER_DESTINATION_ID, destination_id.into_bytes());
@@ -2501,6 +2576,7 @@ impl MetaObject {
                 SUFFIX_TRANSITION_TIER,
                 SUFFIX_TRANSITIONED_OBJECTNAME,
                 SUFFIX_TRANSITIONED_VERSION_ID,
+                SUFFIX_TRANSITIONED_VERSION_STATE,
             ] {
                 if let Some(v) = get_bytes(&self.meta_sys, suffix) {
                     insert_bytes(&mut delete_marker.meta_sys, suffix, v);
@@ -2562,8 +2638,11 @@ impl From<FileInfo> for MetaObject {
             );
         }
 
-        if let Some(vid) = &value.transition_version_id {
-            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, vid.as_bytes().to_vec());
+        if let Some(transition_version) = transitioned_version_bytes(&value) {
+            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
+        }
+        if !value.transition_status.is_empty() {
+            set_transition_version_state(&mut meta_sys, value.transition_version_state);
         }
 
         if !value.transition_tier.is_empty() {
@@ -2706,7 +2785,11 @@ impl MetaDeleteMarker {
                 .map(|v| String::from_utf8_lossy(&v).to_string())
                 .unwrap_or_default();
 
-            fi.transition_version_id = transitioned_version_id_from_meta_sys(&self.meta_sys);
+            fi.transition_version = transitioned_version_from_meta_sys(&self.meta_sys).ok().flatten();
+            fi.transition_version_id = legacy_transitioned_version_id_from_meta_sys(&self.meta_sys);
+            fi.transition_version_state =
+                transition_version_state_from_meta_sys(&self.meta_sys, fi.transition_version.as_deref())
+                    .unwrap_or(TransitionVersionState::Unknown);
         }
 
         fi
@@ -2859,8 +2942,11 @@ impl From<FileInfo> for MetaDeleteMarker {
                 value.transitioned_objname.as_bytes().to_vec(),
             );
         }
-        if let Some(version_id) = value.transition_version_id {
-            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, version_id.as_bytes().to_vec());
+        if let Some(transition_version) = transitioned_version_bytes(&value) {
+            insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, transition_version);
+        }
+        if !value.transition_status.is_empty() || value.tier_free_version() {
+            set_transition_version_state(&mut meta_sys, value.transition_version_state);
         }
         if !value.transition_tier.is_empty() {
             insert_bytes(&mut meta_sys, SUFFIX_TRANSITION_TIER, value.transition_tier.as_bytes().to_vec());
@@ -3412,7 +3498,7 @@ mod tests {
             .insert("x-rustfs-internal-healing".to_string(), "true".to_string());
         marker.metadata.insert("content-type".to_string(), "text/plain".to_string());
         let remote_version_id = Uuid::new_v4();
-        marker.transition_version_id = Some(remote_version_id);
+        marker.transition_version = Some(remote_version_id.to_string());
 
         let converted = MetaDeleteMarker::from(marker);
 
@@ -3420,7 +3506,19 @@ mod tests {
         assert_eq!(converted.meta_sys.get("x-minio-internal-purgestatus"), Some(&b"pending".to_vec()));
         assert_eq!(
             get_bytes(&converted.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID),
-            Some(remote_version_id.as_bytes().to_vec())
+            Some(remote_version_id.to_string().into_bytes())
+        );
+        assert_eq!(
+            converted
+                .meta_sys
+                .get(&format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_ID}")),
+            Some(&remote_version_id.to_string().into_bytes())
+        );
+        assert_eq!(
+            converted
+                .meta_sys
+                .get(&format!("{}{SUFFIX_TRANSITIONED_VERSION_ID}", rustfs_utils::http::MINIO_INTERNAL_PREFIX)),
+            Some(&remote_version_id.to_string().into_bytes())
         );
         assert!(!converted.meta_sys.contains_key("x-rustfs-internal-healing"));
         assert!(!converted.meta_sys.contains_key("content-type"));
@@ -4097,19 +4195,129 @@ mod tests {
             .into_fileinfo("b", "k", false)
             .expect("into_fileinfo");
         assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
+        assert_eq!(fi.transition_version_state, TransitionVersionState::Unknown);
     }
 
     #[test]
-    fn meta_object_transition_version_id_unparseable_stays_readable_as_none() {
-        // A non-UUID / non-16-byte tier version id must NOT make the object
-        // unreadable; it is tolerated as "no tier version" (compat with
-        // pre-hardening behavior and foreign/edge metadata).
+    fn meta_object_transition_version_id_opaque_text_is_preserved() {
         let mut sys = HashMap::new();
-        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"not-a-uuid".to_vec());
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"opaque-generation-42".to_vec());
         let fi = make_meta_object_with_sys(sys)
             .into_fileinfo("b", "k", false)
-            .expect("unparseable transition version id must not fail the object read");
+            .expect("opaque transition version id must decode");
         assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version.as_deref(), Some("opaque-generation-42"));
+        assert_eq!(fi.transition_version_state, TransitionVersionState::Unknown);
+    }
+
+    #[test]
+    fn meta_object_transition_version_state_exact_round_trips_dual_keys() {
+        let id = sample_version_id();
+        let expected_version = id.to_string();
+        let fi = FileInfo {
+            transition_status: "complete".to_string(),
+            transition_version: Some(expected_version.clone()),
+            transition_version_state: TransitionVersionState::Exact,
+            ..Default::default()
+        };
+
+        let object = MetaObject::from(fi);
+        assert_eq!(
+            object
+                .meta_sys
+                .get(&format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_STATE}"))
+                .map(Vec::as_slice),
+            Some(b"exact".as_slice())
+        );
+        assert_eq!(
+            object
+                .meta_sys
+                .get(&format!(
+                    "{}{SUFFIX_TRANSITIONED_VERSION_STATE}",
+                    rustfs_utils::http::MINIO_INTERNAL_PREFIX
+                ))
+                .map(Vec::as_slice),
+            Some(b"exact".as_slice())
+        );
+        assert_eq!(
+            legacy_transitioned_version_id_from_meta_sys(&object.meta_sys),
+            Some(id),
+            "UUID exact writes must remain readable by the legacy UUID consumer"
+        );
+        let decoded = object.into_fileinfo("b", "k", false).expect("exact state should round trip");
+        assert_eq!(decoded.transition_version_state, TransitionVersionState::Exact);
+        assert_eq!(decoded.transition_version.as_deref(), Some(expected_version.as_str()));
+    }
+
+    #[test]
+    fn set_transition_known_disabled_removes_stale_version_dual_keys() {
+        let mut meta_sys = HashMap::new();
+        insert_bytes(&mut meta_sys, SUFFIX_TRANSITIONED_VERSION_ID, b"stale-legacy-version".to_vec());
+        let mut object = make_meta_object_with_sys(meta_sys);
+        object.set_transition(&FileInfo {
+            transition_status: TRANSITION_COMPLETE.to_string(),
+            transitioned_objname: "remote/object".to_string(),
+            transition_version_state: TransitionVersionState::KnownDisabled,
+            transition_tier: "WARM".to_string(),
+            ..Default::default()
+        });
+
+        assert_eq!(get_bytes(&object.meta_sys, SUFFIX_TRANSITIONED_VERSION_ID), None);
+        assert!(
+            !object
+                .meta_sys
+                .contains_key(&format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_ID}"))
+        );
+        assert!(
+            !object
+                .meta_sys
+                .contains_key(&format!("{}{SUFFIX_TRANSITIONED_VERSION_ID}", rustfs_utils::http::MINIO_INTERNAL_PREFIX))
+        );
+        let decoded = object
+            .into_fileinfo("b", "k", false)
+            .expect("known-disabled transition must remain readable after replacing stale metadata");
+        assert_eq!(decoded.transition_version, None);
+        assert_eq!(decoded.transition_version_state, TransitionVersionState::KnownDisabled);
+    }
+
+    #[test]
+    fn meta_object_transition_version_state_conflict_fails_closed() {
+        let mut sys = HashMap::new();
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, sample_version_id().as_bytes().to_vec());
+        sys.insert(format!("{RUSTFS_INTERNAL_PREFIX}{SUFFIX_TRANSITIONED_VERSION_STATE}"), b"exact".to_vec());
+        sys.insert(
+            format!("{}{SUFFIX_TRANSITIONED_VERSION_STATE}", rustfs_utils::http::MINIO_INTERNAL_PREFIX),
+            b"known-disabled".to_vec(),
+        );
+
+        make_meta_object_with_sys(sys)
+            .into_fileinfo("b", "k", false)
+            .expect_err("conflicting state keys must fail closed");
+    }
+
+    #[test]
+    fn meta_object_transition_version_id_invalid_utf8_yields_none() {
+        let mut sys = HashMap::new();
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, vec![0xff]);
+        let fi = make_meta_object_with_sys(sys)
+            .into_fileinfo("b", "k", false)
+            .expect("invalid transition version bytes must not fail the object read");
+        assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version, None);
+    }
+
+    #[test]
+    fn meta_object_transition_version_id_unsafe_text_yields_none() {
+        for value in [b"opaque\0version".to_vec(), vec![b'x'; MAX_TRANSITION_VERSION_LEN + 1]] {
+            let mut sys = HashMap::new();
+            insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, value);
+            let fi = make_meta_object_with_sys(sys)
+                .into_fileinfo("b", "k", false)
+                .expect("unsafe transition version text must not fail the object read");
+            assert_eq!(fi.transition_version_id, None);
+            assert_eq!(fi.transition_version, None);
+        }
     }
 
     #[test]
@@ -4123,6 +4331,7 @@ mod tests {
             .into_fileinfo("b", "k", false)
             .expect("string-form transition version id must decode");
         assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
     }
 
     #[test]
@@ -4152,16 +4361,14 @@ mod tests {
         }
         .into_fileinfo("b", "k", false);
         assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
     }
 
     #[test]
-    fn delete_marker_free_version_transition_version_id_unparseable_stays_readable() {
-        // A malformed tier version id must not make a free-version record corrupt:
-        // it decodes to None and stays readable. Otherwise free-version expiry
-        // fails and the remote-tier object leaks.
+    fn delete_marker_free_version_transition_version_id_opaque_text_is_preserved() {
         let mut sys = HashMap::new();
         insert_bytes(&mut sys, SUFFIX_FREE_VERSION, vec![]);
-        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"not-a-uuid".to_vec());
+        insert_bytes(&mut sys, SUFFIX_TRANSITIONED_VERSION_ID, b"opaque-generation-42".to_vec());
         insert_bytes(&mut sys, SUFFIX_TRANSITION_TIER, b"WARM".to_vec());
         insert_bytes(&mut sys, SUFFIX_TRANSITIONED_OBJECTNAME, b"remote-object".to_vec());
         let fi = MetaDeleteMarker {
@@ -4172,8 +4379,9 @@ mod tests {
         .into_fileinfo("b", "k", false);
 
         assert_eq!(fi.transition_version_id, None);
+        assert_eq!(fi.transition_version.as_deref(), Some("opaque-generation-42"));
         fi.validate_for_metadata_read()
-            .expect("free-version record with an unparseable tier id must remain readable");
+            .expect("free-version record with an opaque tier id must remain readable");
     }
 
     #[test]
@@ -4193,6 +4401,7 @@ mod tests {
         .into_fileinfo("b", "k", false);
 
         assert_eq!(fi.transition_version_id, Some(id));
+        assert_eq!(fi.transition_version, Some(id.to_string()));
     }
 
     #[test]

crates/heal/Cargo.toml

@@ -26,6 +26,9 @@ documentation = "https://docs.rs/rustfs-heal/latest/rustfs_heal/"
 keywords = ["RustFS", "heal", "erasure-coding", "Minio"]
 categories = ["web-programming", "development-tools", "filesystem"]
 
+[lints]
+workspace = true
+
 [dependencies]
 rustfs-config = { workspace = true }
 rustfs-concurrency = { workspace = true }
@@ -53,7 +56,7 @@ serial_test = { workspace = true }
 tempfile = { workspace = true }
 walkdir = { workspace = true }
 http = { workspace = true }
-temp-env = { workspace = true }
+temp-env = { workspace = true, features = ["async_closure"] }
 tokio = { workspace = true, features = ["test-util", "fs", "rt-multi-thread"] }
 
 [lib]

crates/heal/src/heal/channel.rs

@@ -482,7 +482,7 @@ impl HealChannelProcessor {
                     request_id: client_token,
                     success: false,
                     data: None,
-                    error: Some(error_text.clone()),
+                    error: Some(error_text),
                 };
                 let _ = response_tx.send(Ok(response.clone()));
                 self.publish_response(response);

crates/heal/src/heal/event.rs

@@ -576,7 +576,7 @@ mod tests {
         assert_eq!(handler.event_count(), 1);
 
         handler.add_event(event.clone());
-        handler.add_event(event.clone());
+        handler.add_event(event);
         assert_eq!(handler.event_count(), 3);
     }
 
@@ -593,7 +593,7 @@ mod tests {
 
         handler.add_event(event.clone());
         handler.add_event(event.clone());
-        handler.add_event(event.clone()); // Should remove oldest
+        handler.add_event(event); // Should remove oldest
 
         assert_eq!(handler.event_count(), 2);
     }
@@ -610,7 +610,7 @@ mod tests {
         };
 
         handler.add_event(event.clone());
-        handler.add_event(event.clone());
+        handler.add_event(event);
 
         let events = handler.get_events();
         assert_eq!(events.len(), 2);

crates/heal/src/heal/manager.rs

@@ -2892,7 +2892,7 @@ fn update_task_running_metric_for_task(active_heals: &HashMap<String, Arc<HealTa
     gauge!(
         "rustfs_heal_task_running",
         "type" => type_label.to_string(),
-        "set" => set_label.clone()
+        "set" => set_label
     )
     .set(count as f64);
 }
@@ -3482,7 +3482,7 @@ mod tests {
         );
 
         assert_eq!(queue.push(blocked), QueuePushOutcome::Accepted);
-        assert_eq!(queue.push(runnable.clone()), QueuePushOutcome::Accepted);
+        assert_eq!(queue.push(runnable), QueuePushOutcome::Accepted);
 
         let mut running = HashMap::new();
         running.insert("pool_0_set_1".to_string(), 1);

crates/heal/tests/heal_b920_subquorum_union_test.rs

@@ -28,6 +28,7 @@ use rustfs_heal::heal::storage::{
 };
 use serial_test::serial;
 use std::{
+    future::Future,
     path::{Path, PathBuf},
     sync::Arc,
 };
@@ -52,15 +53,9 @@ fn versioned_test_data(seed: u8) -> Vec<u8> {
         .collect()
 }
 
-/// Disable the dangling-delete grace window so the destructive path is genuinely
-/// LIVE in these tests: without the decision-1 guard, a recoverable version WOULD
-/// be dangling-deleted here. With grace at its 1h default the delete path would be
-/// masked and the test would prove nothing.
-fn disable_dangling_grace() {
-    // Safe under nextest: each test runs in its own process and is `#[serial]`.
-    unsafe {
-        std::env::set_var(GRACE_ENV, "0");
-    }
+/// Disable the grace window while the destructive heal decision is evaluated.
+async fn with_dangling_grace_disabled<T>(future: impl Future<Output = T>) -> T {
+    temp_env::async_with_vars([(GRACE_ENV, Some("0"))], future).await
 }
 
 /// Build a real N-disk single-set `ECStore` + `ECStoreHealStorage` via the
@@ -250,7 +245,6 @@ mod serial_tests {
     #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
     #[serial]
     async fn union_meta_lost_data_present_is_repaired_not_destroyed() {
-        disable_dangling_grace();
         let (disk_paths, ecstore, heal_storage) = heal_env_n(8).await;
         let bucket = "b920-meta-lost";
         let object = "obj.bin";
@@ -268,10 +262,10 @@ mod serial_tests {
         }
 
         // Heal the version through the real heal storage (Deep).
-        let (_result, error) = heal_storage
-            .heal_object(bucket, object, Some(&v1), &deep_heal_opts())
-            .await
-            .expect("heal_object call must not itself error");
+        let (_result, error) =
+            with_dangling_grace_disabled(heal_storage.heal_object(bucket, object, Some(&v1), &deep_heal_opts()))
+                .await
+                .expect("heal_object call must not itself error");
         assert!(
             error.is_none(),
             "recoverable version must heal without error (must NOT be dangling-deleted): {error:?}"
@@ -300,7 +294,6 @@ mod serial_tests {
     #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
     #[serial]
     async fn deep_heal_torn_minority_is_dangling_deleted_with_grace_zero() {
-        disable_dangling_grace();
         let (disk_paths, ecstore, heal_storage) = heal_env_n(4).await;
         let bucket = "b920-torn";
         let object = "obj.bin";
@@ -318,10 +311,10 @@ mod serial_tests {
         // is a real destructive action, not a no-op).
         assert_eq!(count_part_files(&object_dir(&disk_paths[0], bucket, object)), 1);
 
-        let (_result, error) = heal_storage
-            .heal_object(bucket, object, Some(&v1), &deep_heal_opts())
-            .await
-            .expect("heal_object call must not itself error");
+        let (_result, error) =
+            with_dangling_grace_disabled(heal_storage.heal_object(bucket, object, Some(&v1), &deep_heal_opts()))
+                .await
+                .expect("heal_object call must not itself error");
         // A dangling delete reports the version as gone (FileVersionNotFound),
         // proving the destructive path fired for a genuinely torn write.
         assert!(error.is_some(), "a torn (< data_blocks) version must NOT be silently treated as healed");
@@ -349,7 +342,6 @@ mod serial_tests {
     #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
     #[serial]
     async fn deep_heal_restores_subquorum_but_reconstructable_version_wider_set() {
-        disable_dangling_grace();
         let (disk_paths, ecstore, heal_storage) = heal_env_n(8).await;
         let bucket = "b920-reconstruct";
         let object = "obj.bin";
@@ -373,10 +365,10 @@ mod serial_tests {
             "disk-walk must enumerate the reconstructable sub-quorum version"
         );
 
-        let (_result, error) = heal_storage
-            .heal_object(bucket, object, Some(&v1), &deep_heal_opts())
-            .await
-            .expect("heal_object call must not itself error");
+        let (_result, error) =
+            with_dangling_grace_disabled(heal_storage.heal_object(bucket, object, Some(&v1), &deep_heal_opts()))
+                .await
+                .expect("heal_object call must not itself error");
         assert!(error.is_none(), "reconstructable version must heal cleanly: {error:?}");
 
         // part.* + xl.meta physically restored on the 4 wiped disks.
@@ -463,7 +455,6 @@ mod serial_tests {
     #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
     #[serial]
     async fn offline_disk_during_walk_does_not_dangling_delete() {
-        disable_dangling_grace();
         let (disk_paths, ecstore, heal_storage) = heal_env_n(4).await;
         let bucket = "b920-offline";
         let object = "obj.bin";
@@ -491,8 +482,7 @@ mod serial_tests {
             remove: false,
             ..Default::default()
         };
-        let (_result, error) = heal_storage
-            .heal_object(bucket, object, Some(&v1), &normal_opts)
+        let (_result, error) = with_dangling_grace_disabled(heal_storage.heal_object(bucket, object, Some(&v1), &normal_opts))
             .await
             .expect("heal_object call must not itself error");
         assert!(error.is_none(), "quorum-present object must not be destroyed: {error:?}");
@@ -510,7 +500,6 @@ mod serial_tests {
     #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
     #[serial]
     async fn deep_heal_keeps_present_ec2_plus_2_shards_healthy() {
-        disable_dangling_grace();
         let (disk_paths, ecstore, heal_storage) = heal_env_n(4).await;
         let bucket = "b1044-deep-verify";
         let object = "obj.bin";
@@ -524,10 +513,10 @@ mod serial_tests {
             assert_eq!(count_part_files(&object_dir(disk, bucket, object)), 1, "intact shard must remain present");
         }
 
-        let (_result, error) = heal_storage
-            .heal_object(bucket, object, Some(&v1), &deep_heal_opts())
-            .await
-            .expect("deep heal_object call must not itself error");
+        let (_result, error) =
+            with_dangling_grace_disabled(heal_storage.heal_object(bucket, object, Some(&v1), &deep_heal_opts()))
+                .await
+                .expect("deep heal_object call must not itself error");
         assert!(error.is_none(), "Deep heal must retain the three intact EC2+2 shards: {error:?}");
 
         assert!(

crates/heal/tests/heal_integration_test.rs

@@ -390,8 +390,10 @@ mod serial_tests {
 
         // ─── 1️⃣ delete format.json on one disk ──────────────
         let format_path = disk_paths[0].join(".rustfs.sys").join("format.json");
-        std::fs::remove_dir_all(&disk_paths[0]).expect("failed to delete all contents under disk_paths[0]");
+        let failed_disk_path = disk_paths[0].with_extension("failed");
+        std::fs::rename(&disk_paths[0], &failed_disk_path).expect("failed to detach disk_paths[0]");
         std::fs::create_dir_all(&disk_paths[0]).expect("failed to recreate disk_paths[0] directory");
+        assert!(!format_path.exists(), "replacement disk already contains format.json before heal");
         println!("✅ Deleted format.json on disk: {:?}", disk_paths[0]);
 
         let (_format_result, format_error) = heal_storage.heal_format(false).await.expect("failed to run heal_format");

crates/io-metrics/src/internode_metrics.rs

@@ -67,6 +67,7 @@ const INTERNODE_MSGPACK_JSON_FALLBACK_TOTAL: &str = "rustfs_system_network_inter
 const INTERNODE_MSGPACK_JSON_DECODE_ERROR_TOTAL: &str = "rustfs_system_network_internode_msgpack_json_decode_error_total";
 const INTERNODE_SIGNATURE_V1_FALLBACK_TOTAL: &str = "rustfs_system_network_internode_signature_v1_fallback_total";
 const INTERNODE_BODY_DIGEST_FALLBACK_TOTAL: &str = "rustfs_system_network_internode_body_digest_fallback_total";
+const INTERNODE_REPLAY_SCOPE_FALLBACK_TOTAL: &str = "rustfs_system_network_internode_replay_scope_fallback_total";
 const INTERNODE_REPLAY_CACHE_OVERFLOW_TOTAL: &str = "rustfs_system_network_internode_replay_cache_overflow_total";
 const ERASURE_WRITE_QUORUM_FAILURES_TOTAL: &str = "rustfs_system_storage_erasure_write_quorum_failures_total";
 
@@ -159,6 +160,7 @@ pub struct InternodeMetricsSnapshot {
     pub operation_write_shutdown_errors_total: u64,
     pub signature_v1_fallback_total: u64,
     pub body_digest_fallback_total: u64,
+    pub replay_scope_fallback_total: u64,
     pub replay_cache_overflow_total: u64,
 }
 
@@ -180,6 +182,7 @@ pub struct InternodeMetrics {
     msgpack_json_decode_error_total: AtomicU64,
     signature_v1_fallback_total: AtomicU64,
     body_digest_fallback_total: AtomicU64,
+    replay_scope_fallback_total: AtomicU64,
     replay_cache_overflow_total: AtomicU64,
 }
 
@@ -431,6 +434,13 @@ impl InternodeMetrics {
         counter!(INTERNODE_BODY_DIGEST_FALLBACK_TOTAL).increment(1);
     }
 
+    /// Count an accepted v1/v2 request that does not carry the replay-scoped signature. This is
+    /// the convergence signal for `RUSTFS_INTERNODE_RPC_REPLAY_SCOPE_STRICT`.
+    pub fn record_replay_scope_fallback(&self) {
+        self.replay_scope_fallback_total.fetch_add(1, Ordering::Relaxed);
+        counter!(INTERNODE_REPLAY_SCOPE_FALLBACK_TOTAL).increment(1);
+    }
+
     /// Count a body-bound internode RPC rejected because the replay-protection nonce cache was
     /// full. Overflow fails closed, so a sustained non-zero rate means
     /// `RUSTFS_INTERNODE_RPC_REPLAY_CACHE_CAPACITY` is undersized for this node's peak legitimate
@@ -488,6 +498,7 @@ impl InternodeMetrics {
             operation_write_shutdown_errors_total: self.operation_write_shutdown_errors_total.load(Ordering::Relaxed),
             signature_v1_fallback_total: self.signature_v1_fallback_total.load(Ordering::Relaxed),
             body_digest_fallback_total: self.body_digest_fallback_total.load(Ordering::Relaxed),
+            replay_scope_fallback_total: self.replay_scope_fallback_total.load(Ordering::Relaxed),
             replay_cache_overflow_total: self.replay_cache_overflow_total.load(Ordering::Relaxed),
         }
     }
@@ -510,6 +521,7 @@ impl InternodeMetrics {
         self.msgpack_json_decode_error_total.store(0, Ordering::Relaxed);
         self.signature_v1_fallback_total.store(0, Ordering::Relaxed);
         self.body_digest_fallback_total.store(0, Ordering::Relaxed);
+        self.replay_scope_fallback_total.store(0, Ordering::Relaxed);
         self.replay_cache_overflow_total.store(0, Ordering::Relaxed);
     }
 }
@@ -887,6 +899,19 @@ mod tests {
         assert_eq!(metrics.snapshot().signature_v1_fallback_total, 0);
     }
 
+    #[test]
+    fn replay_scope_fallback_counter_updates_snapshot_and_resets() {
+        let metrics = InternodeMetrics::default();
+        assert_eq!(metrics.snapshot().replay_scope_fallback_total, 0);
+
+        metrics.record_replay_scope_fallback();
+        metrics.record_replay_scope_fallback();
+        assert_eq!(metrics.snapshot().replay_scope_fallback_total, 2);
+
+        metrics.reset_for_test();
+        assert_eq!(metrics.snapshot().replay_scope_fallback_total, 0);
+    }
+
     #[test]
     fn cluster_peer_flips_offline_after_threshold_and_back_online() {
         // Unique addr keeps this independent of the process-global registry / other tests.

crates/keystone/Cargo.toml

@@ -25,6 +25,9 @@ keywords = ["rustfs", "openstack", "keystone", "authentication", "s3"]
 categories = ["authentication", "web-programming"]
 authors.workspace = true
 
+[lints]
+workspace = true
+
 [dependencies]
 tokio = { workspace = true, features = ["rt", "sync"] }
 reqwest = { workspace = true, features = ["json"] }

crates/kms/Cargo.toml

@@ -44,7 +44,9 @@ argon2 = { workspace = true }
 chacha20poly1305 = { workspace = true }
 rand = { workspace = true, features = ["serde"] }
 base64 = { workspace = true }
+hex = { workspace = true }
 sha2 = { workspace = true }
+subtle = { workspace = true }
 zeroize = { workspace = true, features = ["derive"] }
 
 # Configuration and storage
@@ -55,7 +57,7 @@ tempfile = { workspace = true }
 moka = { workspace = true, features = ["future"] }
 
 # Additional dependencies
-md5 = { workspace = true }
+md-5 = { workspace = true }
 arc-swap = { workspace = true }
 rustfs-utils = { workspace = true }
 rustfs-security-governance = { workspace = true }

crates/kms/src/backends/local.rs

@@ -35,7 +35,6 @@ use std::collections::HashMap;
 use std::path::{Component, Path, PathBuf};
 use std::time::Duration;
 use tokio::fs;
-use tokio::sync::RwLock;
 use tracing::{debug, warn};
 
 /// Reject key identifiers that would not name a single file directly inside the key
@@ -77,8 +76,6 @@ const LOCAL_KMS_ARGON2_P_COST: u32 = 1;
 /// Local KMS client that stores keys in local files
 pub struct LocalKmsClient {
     config: LocalConfig,
-    /// In-memory cache of loaded keys for performance
-    key_cache: RwLock<HashMap<String, MasterKeyInfo>>,
     /// Master encryption key for encrypting stored keys
     master_cipher: Option<Aes256Gcm>,
     /// Legacy pre-beta.9 master cipher for reading pre-Argon2 key files
@@ -139,13 +136,14 @@ impl LocalKmsClient {
             (None, None)
         };
 
-        Ok(Self {
+        let client = Self {
             config,
-            key_cache: RwLock::new(HashMap::new()),
             master_cipher,
             legacy_master_cipher,
             dek_crypto: AesDekCrypto::new(),
-        })
+        };
+        client.validate_existing_keys().await?;
+        Ok(client)
     }
 
     /// Derive a 256-bit key from the master key string using a persistent Argon2id salt.
@@ -193,10 +191,35 @@ impl LocalKmsClient {
 
         let mut salt = [0u8; LOCAL_KMS_MASTER_KEY_SALT_LEN];
         rand::rng().fill(&mut salt[..]);
-        fs::write(&salt_path, salt).await?;
-        Self::set_file_permissions(&salt_path, config.file_permissions).await?;
-        debug!(path = ?salt_path, "Local KMS master key salt created");
-        Ok(salt)
+        let temp_path = config
+            .key_dir
+            .join(format!("{LOCAL_KMS_MASTER_KEY_SALT_FILE}.tmp-{}", uuid::Uuid::new_v4()));
+        fs::write(&temp_path, salt).await?;
+        Self::set_file_permissions(&temp_path, config.file_permissions).await?;
+        match fs::hard_link(&temp_path, &salt_path).await {
+            Ok(()) => {
+                if let Err(error) = fs::remove_file(&temp_path).await {
+                    warn!(path = ?temp_path, %error, "Failed to remove Local KMS salt temporary file");
+                }
+                debug!(path = ?salt_path, "Local KMS master key salt created");
+                Ok(salt)
+            }
+            Err(error) if error.kind() == std::io::ErrorKind::AlreadyExists => {
+                let _ = fs::remove_file(&temp_path).await;
+                let bytes = fs::read(&salt_path).await?;
+                bytes.try_into().map_err(|_| {
+                    KmsError::configuration_error(format!(
+                        "Local KMS master key salt at {} must be exactly {} bytes",
+                        salt_path.display(),
+                        LOCAL_KMS_MASTER_KEY_SALT_LEN
+                    ))
+                })
+            }
+            Err(error) => {
+                let _ = fs::remove_file(&temp_path).await;
+                Err(error.into())
+            }
+        }
     }
 
     #[cfg(unix)]
@@ -242,6 +265,12 @@ impl LocalKmsClient {
 
         let content = fs::read(&key_path).await?;
         let stored_key: StoredMasterKey = serde_json::from_slice(&content)?;
+        if stored_key.key_id != key_id {
+            return Err(KmsError::invalid_key(format!(
+                "Local KMS key file identity mismatch: expected {key_id:?}, found {:?}",
+                stored_key.key_id
+            )));
+        }
 
         let encrypted_bytes = BASE64
             .decode(&stored_key.encrypted_key_material)
@@ -326,8 +355,43 @@ impl LocalKmsClient {
     /// Save a master key to disk
     async fn save_master_key(&self, master_key: &MasterKeyInfo, key_material: &[u8]) -> Result<()> {
         let key_path = self.master_key_path(&master_key.key_id)?;
+        let content = self.encode_master_key(master_key, key_material)?;
+        let temp_path = key_path.with_extension(format!("tmp-{}", uuid::Uuid::new_v4()));
+        fs::write(&temp_path, &content).await?;
+        Self::set_file_permissions(&temp_path, self.config.file_permissions).await?;
+        fs::rename(&temp_path, &key_path).await?;
 
-        // Encrypt key material if master cipher is available
+        debug!(key_id = %master_key.key_id, path = ?key_path, "Local KMS master key saved");
+        Ok(())
+    }
+
+    async fn save_new_master_key(&self, master_key: &MasterKeyInfo, key_material: &[u8]) -> Result<()> {
+        let key_path = self.master_key_path(&master_key.key_id)?;
+        let content = self.encode_master_key(master_key, key_material)?;
+        let temp_path = key_path.with_extension(format!("tmp-{}", uuid::Uuid::new_v4()));
+        fs::write(&temp_path, &content).await?;
+        Self::set_file_permissions(&temp_path, self.config.file_permissions).await?;
+
+        match fs::hard_link(&temp_path, &key_path).await {
+            Ok(()) => {
+                if let Err(error) = fs::remove_file(&temp_path).await {
+                    warn!(path = ?temp_path, %error, "Failed to remove Local KMS key temporary file");
+                }
+                debug!(key_id = %master_key.key_id, path = ?key_path, "Local KMS master key created");
+                Ok(())
+            }
+            Err(error) if error.kind() == std::io::ErrorKind::AlreadyExists => {
+                let _ = fs::remove_file(&temp_path).await;
+                Err(KmsError::key_already_exists(&master_key.key_id))
+            }
+            Err(error) => {
+                let _ = fs::remove_file(&temp_path).await;
+                Err(error.into())
+            }
+        }
+    }
+
+    fn encode_master_key(&self, master_key: &MasterKeyInfo, key_material: &[u8]) -> Result<Vec<u8>> {
         let (encrypted_key_material, nonce, at_rest_protection) = if let Some(ref cipher) = self.master_cipher {
             let mut nonce_bytes = [0u8; 12];
             rand::rng().fill(&mut nonce_bytes[..]);
@@ -362,17 +426,7 @@ impl LocalKmsClient {
             at_rest_protection,
         };
 
-        let content = serde_json::to_vec_pretty(&stored_key)?;
-
-        // Write to temporary file first, then rename for atomicity
-        let temp_path = key_path.with_extension("tmp");
-        fs::write(&temp_path, &content).await?;
-        Self::set_file_permissions(&temp_path, self.config.file_permissions).await?;
-
-        fs::rename(&temp_path, &key_path).await?;
-
-        debug!(key_id = %master_key.key_id, path = ?key_path, "Local KMS master key saved");
-        Ok(())
+        serde_json::to_vec_pretty(&stored_key).map_err(Into::into)
     }
 
     /// Get the actual key material for a master key
@@ -381,6 +435,23 @@ impl LocalKmsClient {
         Ok(key_material)
     }
 
+    async fn validate_existing_keys(&self) -> Result<()> {
+        let mut entries = fs::read_dir(&self.config.key_dir).await?;
+        while let Some(entry) = entries.next_entry().await? {
+            let path = entry.path();
+            if path.extension().is_none_or(|extension| extension != "key") {
+                continue;
+            }
+
+            let key_id = path
+                .file_stem()
+                .and_then(|stem| stem.to_str())
+                .ok_or_else(|| KmsError::configuration_error("Local KMS key file name must be valid UTF-8"))?;
+            self.decode_stored_key(key_id).await?;
+        }
+        Ok(())
+    }
+
     /// Encrypt data using a master key
     async fn encrypt_with_master_key(&self, key_id: &str, plaintext: &[u8]) -> Result<(Vec<u8>, Vec<u8>)> {
         // Load the actual master key material
@@ -513,11 +584,7 @@ impl KmsClient for LocalKmsClient {
         let master_key = MasterKeyInfo::new_with_description(key_id.to_string(), algorithm.to_string(), Some(created_by), None);
 
         // Save to disk
-        self.save_master_key(&master_key, &key_material).await?;
-
-        // Cache the key
-        let mut cache = self.key_cache.write().await;
-        cache.insert(key_id.to_string(), master_key.clone());
+        self.save_new_master_key(&master_key, &key_material).await?;
 
         debug!(key_id, "Local KMS master key created");
         Ok(master_key)
@@ -526,23 +593,7 @@ impl KmsClient for LocalKmsClient {
     async fn describe_key(&self, key_id: &str, _context: Option<&OperationContext>) -> Result<KeyInfo> {
         debug!("Describing key: {}", key_id);
 
-        // Check cache first
-        {
-            let cache = self.key_cache.read().await;
-            if let Some(master_key) = cache.get(key_id) {
-                return Ok(master_key.clone().into());
-            }
-        }
-
-        // Load from disk
         let master_key = self.load_master_key(key_id).await?;
-
-        // Update cache
-        {
-            let mut cache = self.key_cache.write().await;
-            cache.insert(key_id.to_string(), master_key.clone());
-        }
-
         Ok(master_key.into())
     }
 
@@ -602,10 +653,6 @@ impl KmsClient for LocalKmsClient {
         let key_material = self.get_key_material(key_id).await?;
         self.save_master_key(&master_key, &key_material).await?;
 
-        // Update cache
-        let mut cache = self.key_cache.write().await;
-        cache.insert(key_id.to_string(), master_key);
-
         debug!(key_id, "Local KMS key enabled");
         Ok(())
     }
@@ -621,10 +668,6 @@ impl KmsClient for LocalKmsClient {
         let key_material = self.get_key_material(key_id).await?;
         self.save_master_key(&master_key, &key_material).await?;
 
-        // Update cache
-        let mut cache = self.key_cache.write().await;
-        cache.insert(key_id.to_string(), master_key);
-
         debug!(key_id, "Local KMS key disabled");
         Ok(())
     }
@@ -646,10 +689,6 @@ impl KmsClient for LocalKmsClient {
         let key_material = self.get_key_material(key_id).await?;
         self.save_master_key(&master_key, &key_material).await?;
 
-        // Update cache
-        let mut cache = self.key_cache.write().await;
-        cache.insert(key_id.to_string(), master_key);
-
         debug!(key_id, "Local KMS key deletion scheduled");
         Ok(())
     }
@@ -665,31 +704,17 @@ impl KmsClient for LocalKmsClient {
         let key_material = self.get_key_material(key_id).await?;
         self.save_master_key(&master_key, &key_material).await?;
 
-        // Update cache
-        let mut cache = self.key_cache.write().await;
-        cache.insert(key_id.to_string(), master_key);
-
         debug!(key_id, "Local KMS key deletion canceled");
         Ok(())
     }
 
     async fn rotate_key(&self, key_id: &str, _context: Option<&OperationContext>) -> Result<MasterKeyInfo> {
-        debug!("Rotating key: {}", key_id);
-
-        let mut master_key = self.load_master_key(key_id).await?;
-        master_key.version += 1;
-        master_key.rotated_at = Some(Zoned::now());
-
-        // Generate new key material
-        let key_material = generate_key_material(&master_key.algorithm)?;
-        self.save_master_key(&master_key, &key_material).await?;
-
-        // Update cache
-        let mut cache = self.key_cache.write().await;
-        cache.insert(key_id.to_string(), master_key.clone());
-
-        debug!(key_id, "Local KMS key rotated");
-        Ok(master_key)
+        if !fs::try_exists(self.master_key_path(key_id)?).await? {
+            return Err(KmsError::key_not_found(key_id));
+        }
+        Err(KmsError::invalid_operation(
+            "Local KMS key rotation is unavailable until historical key versions can be retained",
+        ))
     }
 
     async fn health_check(&self) -> Result<()> {
@@ -745,11 +770,6 @@ impl KmsBackend for LocalKmsBackend {
     async fn create_key(&self, request: CreateKeyRequest) -> Result<CreateKeyResponse> {
         let key_id = request.key_name.unwrap_or_else(|| uuid::Uuid::new_v4().to_string());
 
-        // `save_master_key` writes through a temp file and renames over the destination, so
-        // creating a key under an existing name would replace its material and silently
-        // destroy the ability to decrypt everything wrapped under it. The sibling
-        // `KmsClient::create_key` has always refused this; the backend path did not, and
-        // this is the path the admin API uses.
         if self.client.master_key_path(&key_id)?.exists() {
             return Err(KmsError::key_already_exists(&key_id));
         }
@@ -767,11 +787,8 @@ impl KmsBackend for LocalKmsBackend {
                 request.description.clone(),
             );
 
-            // Save to disk and cache
-            self.client.save_master_key(&master_key, &key_material).await?;
-
-            let mut cache = self.client.key_cache.write().await;
-            cache.insert(key_id.clone(), master_key.clone());
+            // Save to disk
+            self.client.save_new_master_key(&master_key, &key_material).await?;
 
             master_key
         };
@@ -888,10 +905,6 @@ impl KmsBackend for LocalKmsBackend {
                 .await
                 .map_err(|e| KmsError::internal_error(format!("Failed to delete key file: {e}")))?;
 
-            // Remove from cache
-            let mut cache = self.client.key_cache.write().await;
-            cache.remove(key_id);
-
             debug!(key_id, "Local KMS key deleted immediately");
 
             // Return success response for immediate deletion
@@ -935,10 +948,6 @@ impl KmsBackend for LocalKmsBackend {
 
         self.client.save_master_key(&master_key, &existing_key_material).await?;
 
-        // Update cache
-        let mut cache = self.client.key_cache.write().await;
-        cache.insert(key_id.to_string(), master_key.clone());
-
         // Convert master_key to KeyMetadata for response
         let key_metadata = KeyMetadata {
             key_id: master_key.key_id.clone(),
@@ -986,10 +995,6 @@ impl KmsBackend for LocalKmsBackend {
 
         self.client.save_master_key(&master_key, &existing_key_material).await?;
 
-        // Update cache
-        let mut cache = self.client.key_cache.write().await;
-        cache.insert(key_id.to_string(), master_key.clone());
-
         // Convert master_key to KeyMetadata for response
         let key_metadata = KeyMetadata {
             key_id: master_key.key_id.clone(),
@@ -1189,6 +1194,18 @@ mod tests {
         .expect("stored encrypted key should deserialize");
         assert_eq!(stored.at_rest_protection, StoredKeyProtection::EncryptedMasterKey);
         assert_eq!(stored.nonce.len(), 12);
+
+        let wrong_master_error = match LocalKmsClient::new(LocalConfig {
+            key_dir: client.config.key_dir.clone(),
+            master_key: Some("wrong-master-key".to_string()),
+            file_permissions: Some(0o600),
+        })
+        .await
+        {
+            Ok(_) => panic!("wrong master key must fail initialization"),
+            Err(error) => error,
+        };
+        assert!(matches!(wrong_master_error, KmsError::CryptographicError { .. }));
     }
 
     #[tokio::test]
@@ -1228,17 +1245,57 @@ mod tests {
             master_key: None,
             file_permissions: Some(0o600),
         };
-        let client_without_master = LocalKmsClient::new(config)
-            .await
-            .expect("client without master key should still initialize in dev-mode tests");
-
-        let err = client_without_master
-            .describe_key("encrypted-key", None)
-            .await
-            .expect_err("encrypted key should require a master key to read");
+        let err = match LocalKmsClient::new(config).await {
+            Ok(_) => panic!("initialization must reject an unreadable encrypted key"),
+            Err(error) => error,
+        };
         assert!(err.to_string().contains("requires a configured master key"));
     }
 
+    #[tokio::test]
+    async fn local_key_rotation_is_rejected_without_overwriting_key_material() {
+        let (client, _temp_dir) = create_test_client().await;
+        let key_id = "rotation-key";
+        client.create_key(key_id, "AES_256", None).await.expect("create key");
+        let original_material = client.get_key_material(key_id).await.expect("load original material");
+
+        let error = client
+            .rotate_key(key_id, None)
+            .await
+            .expect_err("rotation must remain unavailable without historical key versions");
+
+        assert!(matches!(error, KmsError::InvalidOperation { .. }));
+        assert_eq!(
+            client.get_key_material(key_id).await.expect("reload original material"),
+            original_material
+        );
+    }
+
+    #[tokio::test]
+    async fn startup_rejects_key_file_with_mismatched_embedded_id() {
+        let (client, temp_dir) = create_dev_mode_client().await;
+        client.create_key("file-name", "AES_256", None).await.expect("create key");
+        let key_path = client.master_key_path("file-name").expect("valid key id");
+        let mut stored: serde_json::Value =
+            serde_json::from_slice(&fs::read(&key_path).await.expect("read key file")).expect("decode key file");
+        stored["key_id"] = serde_json::json!("embedded-name");
+        fs::write(&key_path, serde_json::to_vec_pretty(&stored).expect("encode mismatched key"))
+            .await
+            .expect("write mismatched key");
+
+        let error = match LocalKmsClient::new(LocalConfig {
+            key_dir: temp_dir.path().to_path_buf(),
+            master_key: None,
+            file_permissions: Some(0o600),
+        })
+        .await
+        {
+            Ok(_) => panic!("mismatched key identity must fail initialization"),
+            Err(error) => error,
+        };
+        assert!(matches!(error, KmsError::InvalidKey { .. }));
+    }
+
     #[tokio::test]
     async fn test_load_master_key_accepts_legacy_rfc3339_timestamp() {
         let (client, _temp_dir) = create_dev_mode_client().await;
@@ -1329,17 +1386,6 @@ mod tests {
         )
         .await
         .expect("write beta.5 fixture");
-        let mut explicit_protection = stored_key.clone();
-        let explicit_object = explicit_protection.as_object_mut().expect("beta.5 fixture is a JSON object");
-        explicit_object.insert("key_id".to_string(), serde_json::json!("beta5-explicit-key"));
-        explicit_object.insert("at_rest_protection".to_string(), serde_json::json!("encrypted-master-key"));
-        fs::write(
-            temp_dir.path().join("beta5-explicit-key.key"),
-            serde_json::to_vec_pretty(&explicit_protection).expect("serialize explicit-protection fixture"),
-        )
-        .await
-        .expect("write explicit-protection fixture");
-
         let client = LocalKmsClient::new(LocalConfig {
             key_dir: temp_dir.path().to_path_buf(),
             master_key: Some("beta5-test-master-key".to_string()),
@@ -1353,24 +1399,34 @@ mod tests {
             .await
             .expect("decrypt beta.5 SHA-256 protected key");
         assert_eq!(material, vec![0x42; 32]);
+
+        let mut explicit_protection = stored_key.clone();
+        let explicit_object = explicit_protection.as_object_mut().expect("beta.5 fixture is a JSON object");
+        explicit_object.insert("key_id".to_string(), serde_json::json!("beta5-explicit-key"));
+        explicit_object.insert("at_rest_protection".to_string(), serde_json::json!("encrypted-master-key"));
+        fs::write(
+            temp_dir.path().join("beta5-explicit-key.key"),
+            serde_json::to_vec_pretty(&explicit_protection).expect("serialize explicit-protection fixture"),
+        )
+        .await
+        .expect("write explicit-protection fixture");
         let explicit_error = client
             .get_key_material("beta5-explicit-key")
             .await
             .expect_err("explicit current protection must not fall back to the beta.5 KDF");
         assert!(matches!(explicit_error, KmsError::CryptographicError { .. }));
 
-        let wrong_key_client = LocalKmsClient::new(LocalConfig {
+        let wrong_key_error = match LocalKmsClient::new(LocalConfig {
             key_dir: temp_dir.path().to_path_buf(),
             master_key: Some("wrong-beta5-master-key".to_string()),
             file_permissions: Some(0o600),
         })
         .await
-        .expect("initialize local KMS with wrong beta.5 master key");
-        let error = wrong_key_client
-            .get_key_material("beta5-key")
-            .await
-            .expect_err("wrong beta.5 master key must not decrypt the fixture");
-        assert!(matches!(error, KmsError::CryptographicError { .. }));
+        {
+            Ok(_) => panic!("wrong beta.5 master key must fail initialization"),
+            Err(error) => error,
+        };
+        assert!(matches!(wrong_key_error, KmsError::CryptographicError { .. }));
     }
 
     /// R03-CAN-072 / R03-CAN-073: key identifiers arrive from request input, so every path
@@ -1429,9 +1485,7 @@ mod tests {
         }
     }
 
-    /// R07-CAN-103: `save_master_key` writes a temp file and renames over the destination,
-    /// so creating a key under an existing name would replace its material and silently
-    /// destroy the ability to decrypt anything wrapped under it.
+    /// R07-CAN-103: creating a duplicate key must preserve its original material.
     #[tokio::test]
     async fn backend_create_key_refuses_to_replace_existing_key_material() {
         let temp_dir = TempDir::new().expect("Failed to create temp dir");
@@ -1469,4 +1523,33 @@ mod tests {
             .expect("original key material must survive the refused create");
         assert_eq!(original, after, "existing key material must not be replaced");
     }
+
+    #[tokio::test]
+    async fn concurrent_backend_create_allows_only_one_writer() {
+        let temp_dir = TempDir::new().expect("create key directory");
+        let config = || LocalConfig {
+            key_dir: temp_dir.path().to_path_buf(),
+            master_key: Some("test-master-key".to_string()),
+            file_permissions: Some(0o600),
+        };
+        let first = LocalKmsBackend {
+            client: LocalKmsClient::new(config()).await.expect("create first client"),
+        };
+        let second = LocalKmsBackend {
+            client: LocalKmsClient::new(config()).await.expect("create second client"),
+        };
+        let request = || CreateKeyRequest {
+            key_name: Some("concurrent-key".to_string()),
+            ..Default::default()
+        };
+
+        let (first_result, second_result) = tokio::join!(first.create_key(request()), second.create_key(request()));
+        assert_ne!(first_result.is_ok(), second_result.is_ok(), "exactly one create must succeed");
+        let error = first_result
+            .err()
+            .or_else(|| second_result.err())
+            .expect("one create must fail");
+        assert!(matches!(error, KmsError::KeyAlreadyExists { .. }));
+        assert_eq!(first.client.get_key_material("concurrent-key").await.expect("load key").len(), 32);
+    }
 }

crates/kms/src/lib.rs

@@ -177,8 +177,8 @@ mod tests {
         let service1 = manager.get_encryption_service().await.expect("Service should be available");
 
         // Reconfigure to new service (zero-downtime)
-        let temp_dir2 = TempDir::new().expect("Failed to create temp dir");
-        let config2 = KmsConfig::local(temp_dir2.path().to_path_buf()).with_insecure_development_defaults();
+        let mut config2 = config1;
+        config2.timeout = std::time::Duration::from_secs(45);
         manager.reconfigure(config2).await.expect("Reconfiguration should succeed");
 
         // Verify version 2

crates/kms/src/service.rs

@@ -20,6 +20,7 @@ use crate::manager::KmsManager;
 use crate::types::*;
 use base64::Engine;
 use jiff::Zoned;
+use md5::{Digest as Md5Digest, Md5};
 use rand::random;
 use std::collections::HashMap;
 use std::io::Cursor;
@@ -27,6 +28,12 @@ use tokio::io::{AsyncRead, AsyncReadExt};
 use tracing::debug;
 use zeroize::Zeroize;
 
+fn md5_hex(input: impl AsRef<[u8]>) -> String {
+    let mut hasher = Md5::new();
+    hasher.update(input.as_ref());
+    hex::encode(hasher.finalize())
+}
+
 /// Data key for object encryption
 /// SECURITY: This struct automatically zeros sensitive key material when dropped
 #[derive(Debug, Clone)]
@@ -486,8 +493,7 @@ impl ObjectEncryptionService {
 
         // Validate key MD5 if provided
         if let Some(expected_md5) = customer_key_md5 {
-            let actual_md5 = md5::compute(customer_key);
-            let actual_md5_hex = format!("{actual_md5:x}");
+            let actual_md5_hex = md5_hex(customer_key);
             if actual_md5_hex != expected_md5.to_lowercase() {
                 return Err(KmsError::validation_error("Customer key MD5 mismatch"));
             }

crates/kms/src/service_manager.rs

@@ -20,10 +20,12 @@ use crate::error::{KmsError, Result};
 use crate::manager::KmsManager;
 use crate::service::ObjectEncryptionService;
 use arc_swap::ArcSwap;
+use sha2::{Digest, Sha256};
 use std::sync::{
     Arc, OnceLock,
     atomic::{AtomicU64, Ordering},
 };
+use subtle::ConstantTimeEq;
 use tokio::sync::{Mutex, RwLock};
 use tracing::{debug, error, info, warn};
 
@@ -31,6 +33,53 @@ const LOG_COMPONENT_KMS: &str = "kms";
 const LOG_SUBSYSTEM_SERVICE: &str = "service";
 const EVENT_KMS_SERVICE_STATE: &str = "kms_service_state";
 
+fn local_master_key_fingerprint(master_key: Option<&str>) -> [u8; 32] {
+    let mut digest = Sha256::new();
+    digest.update([u8::from(master_key.is_some())]);
+    if let Some(master_key) = master_key {
+        digest.update(master_key.as_bytes());
+    }
+    digest.finalize().into()
+}
+
+fn validate_local_transition(current: Option<&KmsConfig>, new: &KmsConfig) -> Result<()> {
+    let Some(current) = current else {
+        return Ok(());
+    };
+    let BackendConfig::Local(current_local) = &current.backend_config else {
+        return Ok(());
+    };
+    let BackendConfig::Local(new_local) = &new.backend_config else {
+        return Err(KmsError::configuration_error("Local KMS backend cannot be changed after configuration"));
+    };
+
+    if current_local.key_dir != new_local.key_dir {
+        return Err(KmsError::configuration_error(
+            "Local KMS key directory cannot be changed after configuration",
+        ));
+    }
+    if current_local.file_permissions != new_local.file_permissions {
+        return Err(KmsError::configuration_error(
+            "Local KMS file permissions cannot be changed after configuration",
+        ));
+    }
+    if current.allow_insecure_dev_defaults != new.allow_insecure_dev_defaults {
+        return Err(KmsError::configuration_error(
+            "Local KMS development mode cannot be changed after configuration",
+        ));
+    }
+
+    let current_master_key = local_master_key_fingerprint(current_local.master_key.as_deref());
+    let new_master_key = local_master_key_fingerprint(new_local.master_key.as_deref());
+    if !bool::from(current_master_key.ct_eq(&new_master_key)) {
+        return Err(KmsError::configuration_error(
+            "Local KMS master key cannot be changed after configuration",
+        ));
+    }
+
+    Ok(())
+}
+
 /// KMS service status
 #[derive(Debug, Clone, PartialEq, serde::Serialize, serde::Deserialize)]
 pub enum KmsServiceStatus {
@@ -106,7 +155,12 @@ impl KmsServiceManager {
 
     /// Configure KMS with new configuration
     pub async fn configure(&self, new_config: KmsConfig) -> Result<()> {
+        let _guard = self.lifecycle_mutex.lock().await;
         new_config.validate()?;
+        {
+            let config = self.config.read().await;
+            validate_local_transition(config.as_ref(), &new_config)?;
+        }
 
         // Update configuration
         {
@@ -254,11 +308,9 @@ impl KmsServiceManager {
             "KMS service reconfiguring"
         );
         new_config.validate()?;
-
-        // Configure with new config
         {
-            let mut config = self.config.write().await;
-            *config = Some(new_config.clone());
+            let config = self.config.read().await;
+            validate_local_transition(config.as_ref(), &new_config)?;
         }
 
         // Create new service version without stopping old one
@@ -268,6 +320,11 @@ impl KmsServiceManager {
                 // Get old version for logging (lock-free read)
                 let old_version = self.current_service.load().as_ref().as_ref().map(|sv| sv.version);
 
+                {
+                    let mut config = self.config.write().await;
+                    *config = Some(new_config);
+                }
+
                 // Atomically switch to new service version (lock-free, instant CAS operation)
                 // This is a true atomic operation - no waiting for locks, instant switch
                 // Old service will be dropped when no more Arc references exist
@@ -304,8 +361,6 @@ impl KmsServiceManager {
             Err(e) => {
                 let err_msg = format!("Failed to reconfigure KMS: {e}");
                 error!("{}", err_msg);
-                let mut status = self.status.write().await;
-                *status = KmsServiceStatus::Error(err_msg.clone());
                 Err(KmsError::backend_error(&err_msg))
             }
         }
@@ -477,4 +532,105 @@ mod tests {
         };
         assert!(static_config.secret_key.is_empty());
     }
+
+    #[tokio::test]
+    async fn forbidden_local_master_key_change_preserves_running_config_and_service() {
+        use crate::types::{CreateKeyRequest, KeyUsage};
+        use std::collections::HashMap;
+        use tempfile::TempDir;
+
+        let key_dir = TempDir::new().expect("create local KMS directory");
+        let config = |master_key: &str| {
+            let mut config = KmsConfig::local(key_dir.path().to_path_buf());
+            let BackendConfig::Local(local) = &mut config.backend_config else {
+                panic!("local constructor must create local backend config");
+            };
+            local.master_key = Some(master_key.to_string());
+            config.allow_insecure_dev_defaults = true;
+            config
+        };
+        let manager = KmsServiceManager::new();
+        manager
+            .configure(config("working-master-key"))
+            .await
+            .expect("configure local KMS");
+        manager.start().await.expect("start local KMS");
+        manager
+            .get_manager()
+            .await
+            .expect("running KMS manager")
+            .create_key(CreateKeyRequest {
+                key_name: Some("existing-key".to_string()),
+                key_usage: KeyUsage::EncryptDecrypt,
+                description: None,
+                policy: None,
+                tags: HashMap::new(),
+                origin: None,
+            })
+            .await
+            .expect("create encrypted key");
+        let service_version = manager.get_service_version().await;
+
+        let error = manager
+            .reconfigure(config("wrong-master-key"))
+            .await
+            .expect_err("local master key change must be rejected");
+
+        assert!(error.to_string().contains("master key cannot be changed"));
+        assert_eq!(manager.get_status().await, KmsServiceStatus::Running);
+        assert_eq!(manager.get_service_version().await, service_version);
+        let current = manager.get_config().await.expect("working config must remain");
+        let BackendConfig::Local(local) = current.backend_config else {
+            panic!("working config must remain local");
+        };
+        assert_eq!(local.master_key.as_deref(), Some("working-master-key"));
+    }
+
+    #[tokio::test]
+    async fn configure_cannot_replace_existing_local_backend() {
+        use base64::Engine as _;
+        use tempfile::TempDir;
+
+        let key_dir = TempDir::new().expect("create local KMS directory");
+        let mut local = KmsConfig::local(key_dir.path().to_path_buf());
+        local.allow_insecure_dev_defaults = true;
+        let manager = KmsServiceManager::new();
+        manager.configure(local.clone()).await.expect("configure local KMS");
+
+        let encoded_key = base64::engine::general_purpose::STANDARD.encode([0x5au8; 32]);
+        let error = manager
+            .configure(KmsConfig::static_kms("static-key".to_string(), encoded_key))
+            .await
+            .expect_err("existing local backend must be immutable");
+
+        assert!(error.to_string().contains("backend cannot be changed"));
+        let current = manager.get_config().await.expect("local config must remain");
+        assert!(matches!(current.backend_config, BackendConfig::Local(_)));
+    }
+
+    #[tokio::test]
+    async fn reconfigure_allows_safe_local_runtime_settings_only() {
+        use tempfile::TempDir;
+
+        let key_dir = TempDir::new().expect("create local KMS directory");
+        let mut initial = KmsConfig::local(key_dir.path().to_path_buf());
+        initial.allow_insecure_dev_defaults = true;
+        let manager = KmsServiceManager::new();
+        manager.configure(initial.clone()).await.expect("configure local KMS");
+        manager.start().await.expect("start local KMS");
+
+        let mut updated = initial;
+        updated.default_key_id = Some("evaluation-key".to_string());
+        updated.timeout = std::time::Duration::from_secs(45);
+        updated.enable_cache = false;
+        manager
+            .reconfigure(updated.clone())
+            .await
+            .expect("update safe local settings");
+
+        let current = manager.get_config().await.expect("updated config");
+        assert_eq!(current.default_key_id, updated.default_key_id);
+        assert_eq!(current.timeout, updated.timeout);
+        assert!(!current.enable_cache);
+    }
 }

crates/madmin/src/metrics.rs

@@ -545,6 +545,8 @@ pub struct ScannerMetrics {
     pub collected_at: DateTime<Utc>,
     #[serde(rename = "current_cycle")]
     pub current_cycle: u64,
+    #[serde(rename = "current_cycle_active", default, skip_serializing_if = "Option::is_none")]
+    pub current_cycle_active: Option<bool>,
     #[serde(rename = "current_started")]
     pub current_started: DateTime<Utc>,
     #[serde(rename = "cycle_complete_times")]
@@ -718,7 +720,31 @@ pub struct ScannerMetrics {
 }
 
 impl ScannerMetrics {
+    pub fn is_current_cycle_active(&self) -> bool {
+        self.current_cycle_active.unwrap_or(self.current_cycle > 0)
+    }
+
     pub fn merge(&mut self, other: &Self) {
+        // Legacy nodes omit the activity field and use a non-zero cycle as
+        // their active signal. New nodes publish explicit first-cycle and idle
+        // states, including cycle zero.
+        let self_cycle_active = self.is_current_cycle_active();
+        let other_cycle_active = other.is_current_cycle_active();
+        let self_cycle_authority = (
+            self_cycle_active,
+            self.current_cycle,
+            self.cycles_completed_at.len(),
+            self.cycles_completed_at.as_slice(),
+            self.current_started,
+        );
+        let other_cycle_authority = (
+            other_cycle_active,
+            other.current_cycle,
+            other.cycles_completed_at.len(),
+            other.cycles_completed_at.as_slice(),
+            other.current_started,
+        );
+        let other_cycle_is_authoritative = self_cycle_authority < other_cycle_authority;
         let other_is_newer = self.collected_at < other.collected_at;
         if other_is_newer {
             self.collected_at = other.collected_at;
@@ -854,15 +880,12 @@ impl ScannerMetrics {
             self.ongoing_buckets = other.ongoing_buckets;
         }
 
-        if self.current_cycle < other.current_cycle {
+        if other_cycle_is_authoritative {
             self.current_cycle = other.current_cycle;
             self.cycles_completed_at = other.cycles_completed_at.clone();
             self.current_started = other.current_started;
         }
-
-        if other.cycles_completed_at.len() > self.cycles_completed_at.len() {
-            self.cycles_completed_at = other.cycles_completed_at.clone();
-        }
+        self.current_cycle_active = Some(self_cycle_active || other_cycle_active);
 
         if !other.life_time_ops.is_empty() && self.life_time_ops.is_empty() {
             self.life_time_ops = other.life_time_ops.clone();
@@ -931,7 +954,13 @@ impl Metrics {
         if let Some(scanner) = other.scanner.as_ref() {
             match self.scanner {
                 Some(ref mut s_scanner) => s_scanner.merge(scanner),
-                None => self.scanner = Some(scanner.clone()),
+                None => {
+                    let mut scanner = scanner.clone();
+                    if scanner.current_cycle_active.is_none() {
+                        scanner.current_cycle_active = Some(scanner.is_current_cycle_active());
+                    }
+                    self.scanner = Some(scanner);
+                }
             }
         }
 
@@ -1389,6 +1418,280 @@ pub struct Operations {
 mod tests {
     use super::*;
 
+    #[test]
+    fn scanner_metrics_serializes_cycle_active_presence() {
+        let missing_value = serde_json::to_value(ScannerMetrics::default()).expect("scanner metrics should serialize");
+        assert!(missing_value.get("current_cycle_active").is_none());
+        let missing: ScannerMetrics =
+            serde_json::from_value(missing_value).expect("older scanner metrics without cycle-active should decode");
+        assert_eq!(missing.current_cycle_active, None);
+
+        let explicit_false_value = serde_json::to_value(ScannerMetrics {
+            current_cycle_active: Some(false),
+            ..Default::default()
+        })
+        .expect("scanner metrics with explicit cycle-active should serialize");
+        assert_eq!(explicit_false_value["current_cycle_active"], serde_json::Value::Bool(false));
+        let explicit_false: ScannerMetrics =
+            serde_json::from_value(explicit_false_value).expect("explicit cycle-active should decode");
+        assert_eq!(explicit_false.current_cycle_active, Some(false));
+    }
+
+    #[test]
+    fn scanner_metrics_merge_prefers_an_active_first_cycle() {
+        let collected_at = Utc::now();
+        let idle_started = collected_at - chrono::Duration::hours(1);
+        let active_started = collected_at - chrono::Duration::seconds(5);
+        let mut scanner = ScannerMetrics {
+            collected_at,
+            current_cycle: 0,
+            current_cycle_active: Some(false),
+            current_started: idle_started,
+            ..Default::default()
+        };
+
+        scanner.merge(&ScannerMetrics {
+            collected_at: collected_at + chrono::Duration::seconds(1),
+            current_cycle: 0,
+            current_cycle_active: Some(true),
+            current_started: active_started,
+            ..Default::default()
+        });
+
+        assert_eq!(scanner.current_cycle_active, Some(true));
+        assert_eq!(scanner.current_cycle, 0);
+        assert_eq!(scanner.current_started, active_started);
+    }
+
+    #[test]
+    fn metrics_merge_preserves_explicit_active_first_cycle() {
+        let mut aggregated = Metrics::default();
+        aggregated.merge(&Metrics {
+            scanner: Some(ScannerMetrics {
+                current_cycle_active: Some(true),
+                ..Default::default()
+            }),
+            ..Default::default()
+        });
+
+        let scanner = aggregated.scanner.expect("aggregated scanner metrics");
+        assert_eq!(scanner.current_cycle_active, Some(true));
+        assert_eq!(scanner.current_cycle, 0);
+    }
+
+    #[test]
+    fn scanner_metrics_merge_preserves_legacy_nonzero_active_signal() {
+        let collected_at = Utc::now();
+        let mut scanner = ScannerMetrics {
+            collected_at,
+            ..Default::default()
+        };
+
+        scanner.merge(&ScannerMetrics {
+            collected_at: collected_at + chrono::Duration::seconds(1),
+            current_cycle: 7,
+            ..Default::default()
+        });
+
+        assert_eq!(scanner.current_cycle_active, Some(true));
+        assert_eq!(scanner.current_cycle, 7);
+    }
+
+    #[test]
+    fn metrics_merge_normalizes_first_legacy_scanner_snapshot() {
+        let legacy = Metrics {
+            scanner: Some(ScannerMetrics {
+                current_cycle: 7,
+                ..Default::default()
+            }),
+            ..Default::default()
+        };
+        let mut aggregated = Metrics::default();
+
+        aggregated.merge(&legacy);
+
+        let scanner = aggregated.scanner.expect("aggregated scanner metrics");
+        assert_eq!(scanner.current_cycle_active, Some(true));
+        assert_eq!(scanner.current_cycle, 7);
+    }
+
+    #[test]
+    fn scanner_metrics_merge_preserves_explicit_inactive_nonzero_cycle() {
+        let collected_at = Utc::now();
+        let mut scanner = ScannerMetrics::default();
+
+        scanner.merge(&ScannerMetrics {
+            collected_at,
+            current_cycle: 7,
+            current_cycle_active: Some(false),
+            ..Default::default()
+        });
+
+        assert_eq!(scanner.current_cycle_active, Some(false));
+        assert_eq!(scanner.current_cycle, 7);
+    }
+
+    #[test]
+    fn scanner_metrics_merge_cycle_active_is_order_independent() {
+        let collected_at = Utc::now();
+        let legacy_active = ScannerMetrics {
+            collected_at,
+            current_cycle: 7,
+            current_started: collected_at - chrono::Duration::seconds(10),
+            ..Default::default()
+        };
+        let explicit_idle = ScannerMetrics {
+            collected_at,
+            current_cycle: 0,
+            current_cycle_active: Some(false),
+            current_started: collected_at - chrono::Duration::hours(1),
+            ..Default::default()
+        };
+
+        let mut legacy_first = legacy_active.clone();
+        legacy_first.merge(&explicit_idle);
+        let mut legacy_second = explicit_idle.clone();
+        legacy_second.merge(&legacy_active);
+        assert_eq!(legacy_first.current_cycle_active, Some(true));
+        assert_eq!(legacy_second.current_cycle_active, Some(true));
+        assert_eq!(legacy_first.current_cycle, 7);
+        assert_eq!(legacy_second.current_cycle, 7);
+
+        let explicit_inactive_nonzero = ScannerMetrics {
+            collected_at,
+            current_cycle: 7,
+            current_cycle_active: Some(false),
+            ..Default::default()
+        };
+        let mut inactive_first = explicit_inactive_nonzero.clone();
+        inactive_first.merge(&explicit_idle);
+        let mut inactive_second = explicit_idle;
+        inactive_second.merge(&explicit_inactive_nonzero);
+        assert_eq!(inactive_first.current_cycle_active, Some(false));
+        assert_eq!(inactive_second.current_cycle_active, Some(false));
+        assert_eq!(inactive_first.current_cycle, 7);
+        assert_eq!(inactive_second.current_cycle, 7);
+    }
+
+    #[test]
+    fn scanner_metrics_merge_cycle_authority_is_order_independent() {
+        let collected_at = Utc::now();
+        let completion = collected_at - chrono::Duration::minutes(1);
+        let earlier_active = ScannerMetrics {
+            collected_at,
+            current_cycle: 7,
+            current_cycle_active: Some(true),
+            current_started: collected_at - chrono::Duration::seconds(10),
+            cycles_completed_at: vec![completion],
+            ..Default::default()
+        };
+        let later_active = ScannerMetrics {
+            collected_at: collected_at + chrono::Duration::seconds(1),
+            current_cycle: 7,
+            current_cycle_active: Some(true),
+            current_started: collected_at - chrono::Duration::seconds(5),
+            cycles_completed_at: vec![completion],
+            ..Default::default()
+        };
+
+        let mut earlier_first = earlier_active.clone();
+        earlier_first.merge(&later_active);
+        let mut later_first = later_active.clone();
+        later_first.merge(&earlier_active);
+
+        assert_eq!(earlier_first.current_started, later_active.current_started);
+        assert_eq!(later_first.current_started, later_active.current_started);
+        assert_eq!(earlier_first.cycles_completed_at, later_active.cycles_completed_at);
+        assert_eq!(later_first.cycles_completed_at, later_active.cycles_completed_at);
+
+        let stale_idle = ScannerMetrics {
+            collected_at,
+            current_cycle_active: Some(false),
+            current_started: collected_at - chrono::Duration::hours(1),
+            ..Default::default()
+        };
+        let completed_idle = ScannerMetrics {
+            collected_at: collected_at + chrono::Duration::seconds(1),
+            current_cycle_active: Some(false),
+            current_started: collected_at - chrono::Duration::seconds(5),
+            cycles_completed_at: vec![completion],
+            ..Default::default()
+        };
+
+        let mut stale_first = stale_idle.clone();
+        stale_first.merge(&completed_idle);
+        let mut completed_first = completed_idle.clone();
+        completed_first.merge(&stale_idle);
+
+        assert_eq!(stale_first.current_started, completed_idle.current_started);
+        assert_eq!(completed_first.current_started, completed_idle.current_started);
+        assert_eq!(stale_first.cycles_completed_at, completed_idle.cycles_completed_at);
+        assert_eq!(completed_first.cycles_completed_at, completed_idle.cycles_completed_at);
+    }
+
+    #[test]
+    fn scanner_metrics_merge_cycle_authority_is_associative() {
+        let collected_at = Utc::now();
+        let older_completion = collected_at - chrono::Duration::minutes(3);
+        let last_completion = collected_at - chrono::Duration::minutes(1);
+        let cycle_seven = ScannerMetrics {
+            collected_at,
+            current_cycle: 7,
+            current_cycle_active: Some(true),
+            current_started: collected_at - chrono::Duration::seconds(10),
+            cycles_completed_at: vec![older_completion, last_completion],
+            ..Default::default()
+        };
+        let cycle_eight = ScannerMetrics {
+            collected_at: collected_at + chrono::Duration::seconds(1),
+            current_cycle: 8,
+            current_cycle_active: Some(true),
+            current_started: collected_at - chrono::Duration::seconds(5),
+            cycles_completed_at: vec![last_completion],
+            ..Default::default()
+        };
+        let newer_idle = ScannerMetrics {
+            collected_at: collected_at + chrono::Duration::hours(1),
+            current_cycle_active: Some(false),
+            current_started: collected_at - chrono::Duration::hours(1),
+            ..Default::default()
+        };
+
+        let mut left_associative = cycle_seven.clone();
+        left_associative.merge(&cycle_eight);
+        left_associative.merge(&newer_idle);
+
+        let mut right_group = cycle_eight.clone();
+        right_group.merge(&newer_idle);
+        let mut right_associative = cycle_seven.clone();
+        right_associative.merge(&right_group);
+
+        assert_eq!(left_associative.current_cycle, 8);
+        assert_eq!(right_associative.current_cycle, 8);
+        assert_eq!(left_associative.current_started, cycle_eight.current_started);
+        assert_eq!(right_associative.current_started, cycle_eight.current_started);
+        assert_eq!(left_associative.cycles_completed_at, cycle_eight.cycles_completed_at);
+        assert_eq!(right_associative.cycles_completed_at, cycle_eight.cycles_completed_at);
+
+        for order in [
+            [&cycle_seven, &cycle_eight, &newer_idle],
+            [&cycle_seven, &newer_idle, &cycle_eight],
+            [&cycle_eight, &cycle_seven, &newer_idle],
+            [&cycle_eight, &newer_idle, &cycle_seven],
+            [&newer_idle, &cycle_seven, &cycle_eight],
+            [&newer_idle, &cycle_eight, &cycle_seven],
+        ] {
+            let mut merged = ScannerMetrics::default();
+            for scanner in order {
+                merged.merge(scanner);
+            }
+            assert_eq!(merged.current_cycle_active, Some(true));
+            assert_eq!(merged.current_cycle, 8);
+            assert_eq!(merged.current_started, cycle_eight.current_started);
+            assert_eq!(merged.cycles_completed_at, cycle_eight.cycles_completed_at);
+        }
+    }
+
     #[test]
     fn scanner_metrics_merge_aggregates_partial_cycles_by_source() {
         let collected_at = Utc::now();

crates/obs/src/metrics/stats_collector.rs

@@ -262,11 +262,11 @@ async fn obs_site_replication_stats() -> ReplicationStats {
 }
 
 fn current_scanner_cycle_age_seconds(
-    current_cycle: u64,
+    current_cycle_active: bool,
     current_started: chrono::DateTime<Utc>,
     now: chrono::DateTime<Utc>,
 ) -> u64 {
-    if current_cycle == 0 {
+    if !current_cycle_active {
         0
     } else {
         now.signed_duration_since(current_started).num_seconds().max(0) as u64
@@ -1090,7 +1090,7 @@ pub async fn collect_scanner_metric_stats() -> Option<ScannerStats> {
     let reference_time = metrics.cycles_completed_at.last().copied().unwrap_or(metrics.current_started);
     let last_activity_seconds = now.signed_duration_since(reference_time).num_seconds().max(0) as u64;
     let active_paths = metrics.active_scan_paths as u64;
-    let current_cycle_age_seconds = current_scanner_cycle_age_seconds(metrics.current_cycle, metrics.current_started, now);
+    let current_cycle_age_seconds = current_scanner_cycle_age_seconds(metrics.current_cycle_active, metrics.current_started, now);
     let current_scan_mode = scanner_scan_mode_code(&metrics.current_scan_mode);
     let current_cycle_age = current_cycle_age_seconds as f64;
     let last_cycle_duration = metrics.last_cycle_duration_seconds;
@@ -1464,21 +1464,21 @@ mod tests {
     fn current_scanner_cycle_age_seconds_returns_zero_when_idle() {
         let now = Utc::now();
 
-        assert_eq!(current_scanner_cycle_age_seconds(0, now - chrono::Duration::seconds(30), now), 0);
+        assert_eq!(current_scanner_cycle_age_seconds(false, now - chrono::Duration::seconds(30), now), 0);
     }
 
     #[test]
     fn current_scanner_cycle_age_seconds_clamps_future_start() {
         let now = Utc::now();
 
-        assert_eq!(current_scanner_cycle_age_seconds(4, now + chrono::Duration::seconds(30), now), 0);
+        assert_eq!(current_scanner_cycle_age_seconds(true, now + chrono::Duration::seconds(30), now), 0);
     }
 
     #[test]
-    fn current_scanner_cycle_age_seconds_reports_active_elapsed_time() {
+    fn current_scanner_cycle_age_seconds_reports_active_first_cycle_elapsed_time() {
         let now = Utc::now();
 
-        assert_eq!(current_scanner_cycle_age_seconds(4, now - chrono::Duration::seconds(45), now), 45);
+        assert_eq!(current_scanner_cycle_age_seconds(true, now - chrono::Duration::seconds(45), now), 45);
     }
 
     #[test]

crates/protocols/Cargo.toml

@@ -25,6 +25,9 @@ description = "Protocol implementations for RustFS (FTPS, SFTP, etc.)"
 keywords = ["ftp", "sftp", "protocol", "storage", "rustfs"]
 categories = ["network-programming", "filesystem"]
 
+[lints]
+workspace = true
+
 [features]
 default = []
 ftps = ["dep:libunftp", "dep:unftp-core", "dep:rustls", "dep:rustfs-tls-runtime", "dep:subtle"]
@@ -44,7 +47,7 @@ swift = [
     "dep:tokio-util",
     "dep:serde",
     "dep:urlencoding",
-    "dep:md5",
+    "dep:md-5",
     "dep:quick-xml",
     "dep:hmac",
     "dep:sha1",
@@ -108,7 +111,7 @@ http-body-util = { workspace = true, optional = true }
 tokio-util = { workspace = true, optional = true, features = ["rt", "io", "compat"] }
 serde = { workspace = true, optional = true, features = ["derive"] }
 urlencoding = { workspace = true, optional = true }
-md5 = { workspace = true, optional = true }
+md-5 = { workspace = true, optional = true }
 quick-xml = { workspace = true, optional = true, features = ["serialize"] }
 hmac = { workspace = true, optional = true }
 sha1 = { workspace = true, optional = true }

crates/protocols/src/common/dummy_storage.rs

@@ -608,7 +608,7 @@ impl StorageBackend for DummyBackend {
             inner.put_object_calls.push(PutObjectCall {
                 bucket: input.bucket.to_string(),
                 key: input.key.to_string(),
-                metadata: input.metadata.clone(),
+                metadata: input.metadata,
             });
             let stall = inner.stall_put_object;
             let entered = inner.put_object_entered.clone();
@@ -731,7 +731,7 @@ impl StorageBackend for DummyBackend {
             inner.create_multipart_calls.push(CreateMultipartCall {
                 bucket: input.bucket.to_string(),
                 key: input.key.to_string(),
-                metadata: input.metadata.clone(),
+                metadata: input.metadata,
             });
         }
         match self.inner.lock().expect("lock").create_multipart_upload.pop_front() {
@@ -749,7 +749,7 @@ impl StorageBackend for DummyBackend {
             inner.upload_part_calls.push(UploadPartCall {
                 bucket: input.bucket.to_string(),
                 key: input.key.to_string(),
-                upload_id: input.upload_id.to_string(),
+                upload_id: input.upload_id,
                 part_number: input.part_number,
                 content_length: input.content_length,
             });
@@ -787,7 +787,7 @@ impl StorageBackend for DummyBackend {
             inner.complete_multipart_calls.push(CompleteCall {
                 bucket: input.bucket.to_string(),
                 key: input.key.to_string(),
-                upload_id: input.upload_id.to_string(),
+                upload_id: input.upload_id,
                 part_count,
             });
         }
@@ -808,7 +808,7 @@ impl StorageBackend for DummyBackend {
             inner.abort_multipart_calls.push(AbortCall {
                 bucket: input.bucket.to_string(),
                 key: input.key.to_string(),
-                upload_id: input.upload_id.to_string(),
+                upload_id: input.upload_id,
             });
         }
         match self.inner.lock().expect("lock").abort_multipart_upload.pop_front() {

crates/protocols/src/ftps/driver.rs

@@ -379,9 +379,7 @@ where
             .await
             .map_err(|_| Error::new(ErrorKind::PermanentFileNotAvailable, "Access denied"))?;
 
-        let prefix_with_slash = prefix
-            .clone()
-            .map(|p| if p.ends_with('/') { p.to_string() } else { format!("{}/", p) });
+        let prefix_with_slash = prefix.clone().map(|p| if p.ends_with('/') { p } else { format!("{}/", p) });
 
         let list_input = ListObjectsV2Input::builder()
             .bucket(bucket)