fix(site-replication): admit only verifiable peer-edit fences

fix(site-replication): lift a rejoined site's restarted edit counter over stale fence marks

A site removed while unreachable (unilateral removal: the receiver never
dropped it from its peer map, so parse_site_replication_state's load-time
mark pruning never fired) that later rejoins recreates its state object
and restarts edit_generation at zero. The receiver's surviving high-water
mark then silently fences out every stamped delivery from that origin —
peer edits and the add finalize fan-out alike are acked without applying
— until the restarted counter catches up.

Allocate the generation as a hybrid logical clock instead:
max(wall clock in unix nanoseconds, previous + 1), still inside the state
transaction under the distributed state-object lock. Every value a
lifetime hands out is capped by the wall clock at its own allocation, so
a recreated lifetime's first allocation exceeds them all and clears the
stale mark, while a pre-removal delivery still in flight stays below the
new floor and remains correctly fenced. previous+1 keeps allocations
strictly increasing across same-tick allocations and mid-lifetime clock
regressions.

Nothing changes on the wire or in the persisted schema: editGeneration
stays the single fence param and edit_generation the single counter
field, so pre-hybrid receivers get the fix as soon as the sender
upgrades, old binaries preserve the field across rolling up/downgrades,
and marks recorded by plain-counter receivers (small values) are cleared
by any wall-clock allocation. A clock that regresses across a
delete/recreate degrades to a fence that self-heals once real time
passes the previous lifetime's last allocation, and introduces no
rollback window beyond what the plain counter already had.

An epoch-based design (editEpoch wire param + per-origin epoch marks)
was built first and rejected under adversarial review: old binaries
rewriting the state object drop the unknown epoch fields, which both
disarms the fix mid-rolling-upgrade and — because epoch adoption lowers
the generation mark — reopens the pre-restart rollback the fence exists
to prevent; a backwards clock also fences an origin permanently instead
of self-healing. The hybrid clock has none of these modes.

.github/workflows/docker.yml

@@ -94,6 +94,7 @@ jobs:
       short_sha: ${{ steps.check.outputs.short_sha }}
       is_prerelease: ${{ steps.check.outputs.is_prerelease }}
       create_latest: ${{ steps.check.outputs.create_latest }}
+      source_ref: ${{ steps.check.outputs.source_ref }}
     steps:
       - name: Checkout repository
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
@@ -118,6 +119,7 @@ jobs:
           short_sha=""
           is_prerelease=false
           create_latest=false
+          source_ref="$GITHUB_SHA"
 
           if [[ "${{ github.event_name }}" == "workflow_run" ]]; then
             # Triggered by build workflow completion
@@ -137,6 +139,7 @@ jobs:
             # Extract version info from commit message or use commit SHA
             # Use Git to generate consistent short SHA (ensures uniqueness like build.yml)
             short_sha=$(git rev-parse --short "$HEAD_SHA")
+            source_ref="$HEAD_SHA"
 
             # Determine build type based on triggering workflow event and ref
             triggering_event="$TRIGGERING_EVENT"
@@ -261,6 +264,23 @@ jobs:
                 echo "⚠️  Only release versions (latest, v1.0.0, 1.0.0) and prereleases (v1.0.0-alpha1, 1.0.0-beta2) are supported"
                 ;;
             esac
+
+            if [[ "$should_build" == true && "$input_version" != "latest" ]]; then
+              tag_ref="refs/tags/$input_version"
+              if ! git ls-remote --exit-code origin "$tag_ref" >/dev/null 2>&1; then
+                if [[ "$input_version" == v* ]]; then
+                  tag_ref="refs/tags/${input_version#v}"
+                else
+                  tag_ref="refs/tags/v$input_version"
+                fi
+              fi
+
+              if ! git ls-remote --exit-code origin "$tag_ref" >/dev/null 2>&1; then
+                echo "❌ Release tag not found for Docker build: $input_version"
+                exit 1
+              fi
+              source_ref="$tag_ref"
+            fi
           fi
 
           {
@@ -271,6 +291,7 @@ jobs:
             echo "short_sha=$short_sha"
             echo "is_prerelease=$is_prerelease"
             echo "create_latest=$create_latest"
+            echo "source_ref=$source_ref"
           } >> "$GITHUB_OUTPUT"
 
           echo "🐳 Docker Build Summary:"
@@ -281,6 +302,7 @@ jobs:
           echo "  - Short SHA: $short_sha"
           echo "  - Is prerelease: $is_prerelease"
           echo "  - Create latest: $create_latest"
+          echo "  - Source ref: $source_ref"
 
   # Build multi-arch Docker images
   # Strategy: Build images using pre-built binaries from dl.rustfs.com
@@ -308,6 +330,7 @@ jobs:
         uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0 # v7
         with:
           persist-credentials: false
+          ref: ${{ needs.build-check.outputs.source_ref }}
 
       - name: Login to Docker Hub
         uses: docker/login-action@c94ce9fb468520275223c153574b00df6fe4bcc9 # v3
@@ -397,7 +420,8 @@ jobs:
           LABELS="org.opencontainers.image.title=RustFS"
           LABELS="$LABELS,org.opencontainers.image.description=RustFS distributed object storage system"
           LABELS="$LABELS,org.opencontainers.image.version=$VERSION"
-          LABELS="$LABELS,org.opencontainers.image.revision=${{ github.sha }}"
+          SOURCE_REVISION="$(git rev-parse HEAD)"
+          LABELS="$LABELS,org.opencontainers.image.revision=$SOURCE_REVISION"
           LABELS="$LABELS,org.opencontainers.image.source=${{ github.server_url }}/${{ github.repository }}"
           LABELS="$LABELS,org.opencontainers.image.created=$(date -u +'%Y-%m-%dT%H:%M:%SZ')"
           LABELS="$LABELS,org.opencontainers.image.build-type=$BUILD_TYPE"

rustfs/src/admin/handlers/site_replication.rs

@@ -1067,9 +1067,13 @@ fn parse_site_replication_state(data: &[u8]) -> S3Result<SiteReplicationState> {
     state.peers = normalize_peer_map_by_identity(state.peers);
     // A peer-edit high-water mark only fences a CURRENT peer. A site that
     // leaves drops below two peers, which clears its own state object and
-    // restarts its generation counter at zero — a mark left over from the
-    // previous membership would then reject every edit it sends after it
-    // rejoins. Dropping departed origins on load also keeps the map bounded.
+    // restarts its generation counter — a mark left over from the previous
+    // membership must not reject the edits it sends after it rejoins. This
+    // pruning covers departures THIS site observed; an origin removed
+    // unilaterally elsewhere stays in this peer map with its mark, and the
+    // wall-clock floor in `next_peer_edit_generation` is what lifts its
+    // restarted counter over that mark. Dropping departed origins on load
+    // also keeps the map bounded.
     state
         .applied_edit_generations
         .retain(|origin, _| state.peers.contains_key(origin));
@@ -5935,11 +5939,51 @@ fn summarize_peer_error_detail(detail: &str) -> String {
     summary
 }
 
-/// Allocate the next peer-edit generation. Called inside the state
-/// transaction, so the counter is handed out under the distributed
-/// state-object lock and two nodes of this site can never take the same one.
+/// The wall clock in unix nanoseconds, clamped into u64. A pre-1970 (or
+/// post-2554) clock yields 0, which makes the hybrid allocation below
+/// degrade to the plain `previous + 1` counter — monotone, never panicking.
+fn edit_generation_wall_clock() -> u64 {
+    u64::try_from(OffsetDateTime::now_utc().unix_timestamp_nanos()).unwrap_or(0)
+}
+
+/// Allocate the next peer-edit generation as a hybrid logical clock:
+/// `max(wall clock in unix nanoseconds, previous + 1)`. Called inside the
+/// state transaction, so the value is handed out under the distributed
+/// state-object lock and two nodes of this site can never take the same one
+/// (`previous + 1` keeps the sequence strictly increasing even when two
+/// allocations land in one clock tick, and keeps it monotone on a node
+/// whose clock stepped backwards mid-lifetime).
+///
+/// The wall-clock floor is what survives the counter's death. A site
+/// removed while unreachable — the receiver never dropped it from its peer
+/// map, so the load-time mark pruning in `parse_site_replication_state`
+/// never fired — that later rejoins recreates its state object with the
+/// counter back at zero. A plain counter would then hand out generations
+/// below the receiver's stale high-water mark and every delivery would be
+/// silently fenced until the counter caught up. Jumping to wall time clears
+/// that mark: every value the deleted lifetime handed out was capped by the
+/// wall clock at its own allocation (or by a prior lifetime's cap, applied
+/// inductively), so the recreated lifetime's first allocation exceeds them
+/// all — while a pre-removal delivery still in flight stays below the new
+/// floor and remains correctly fenced. Marks recorded by pre-hybrid
+/// receivers (small plain-counter values) sit far below any wall-clock
+/// value, so a restarted origin passes those too — the fix needs only the
+/// sender upgraded, nothing on the wire or in the receiver changed.
+///
+/// A wall clock that regresses across a delete/recreate (the recreating
+/// node's clock behind the clock that fed the previous lifetime) mints
+/// below the stale mark and the origin stays fenced — but only until real
+/// time passes the previous lifetime's last allocation, because every later
+/// allocation takes the wall-clock floor again. Bounded by the skew,
+/// self-healing, and no rollback window beyond the plain counter's: a
+/// delivery applies only at or above the receiver's mark, so the one
+/// cross-lifetime interleaving that can apply stale content — a
+/// pre-removal delivery whose generation lands above everything the
+/// regressed new lifetime has minted — required the same straggler landing
+/// above the mark under the plain counter, where the recreated counter's
+/// low restart made it strictly easier to hit.
 fn next_peer_edit_generation(state: &mut SiteReplicationState) -> u64 {
-    state.edit_generation = state.edit_generation.saturating_add(1);
+    state.edit_generation = edit_generation_wall_clock().max(state.edit_generation.saturating_add(1));
     state.edit_generation
 }
 
@@ -5970,6 +6014,63 @@ fn peer_edit_fence(queries: &HashMap<String, String>) -> Option<(String, u64)> {
     Some((origin.clone(), generation))
 }
 
+/// The largest generation an incoming fence may carry: this site's clock in
+/// unix nanoseconds plus a day of cross-site skew. Every site authenticates
+/// peer traffic with the shared site-replicator service account, so the
+/// receiver cannot tell WHICH site stamped a fence — a compromised peer can
+/// claim any origin, and a u64::MAX-scale generation would raise that
+/// origin's high-water mark past anything the genuine site ever allocates,
+/// silently fencing out its every future edit. A genuine generation is the
+/// hybrid clock of [`next_peer_edit_generation`], floored by the sender's
+/// wall time, so it exceeds this site's clock only by cross-site skew — and
+/// the mark a ceiling-level forgery can still plant decays as real time
+/// passes it, within the allowance. Only an origin still allocating with the
+/// pre-hybrid plain counter cannot outrun such a mark until it upgrades: its
+/// genuine generations never approach nanosecond scale.
+fn peer_edit_fence_generation_ceiling(now_unix_nanos: u64) -> u64 {
+    const CROSS_SITE_SKEW_ALLOWANCE_NANOS: u64 = 24 * 60 * 60 * 1_000_000_000;
+    now_unix_nanos.saturating_add(CROSS_SITE_SKEW_ALLOWANCE_NANOS)
+}
+
+/// Whether an incoming fence may be honoured, as far as this site can vouch
+/// for it. The sender's identity is unverifiable (shared service account),
+/// so the check runs over what the receiving state knows: the claimed origin
+/// must be a site this state currently replicates with — the same membership
+/// rule the load-time mark pruning applies, so every mark recorded behind
+/// this check is one a reload would keep — and not this site itself, which
+/// never delivers edits to itself; and the generation must sit under
+/// [`peer_edit_fence_generation_ceiling`]. The caller IGNORES an
+/// inadmissible fence rather than failing the request: the delivery applies
+/// exactly as an unstamped (pre-fence) delivery would, no high-water mark is
+/// read or written, and the worst a forged fence achieves is forfeiting an
+/// ordering guarantee its sender was never owed.
+fn peer_edit_fence_is_admissible(
+    state: &SiteReplicationState,
+    local_deployment_id: &str,
+    fence: &(String, u64),
+    generation_ceiling: u64,
+) -> bool {
+    let (origin, generation) = fence;
+    let result = if origin == local_deployment_id || !state.peers.contains_key(origin) {
+        "fence_origin_not_a_remote_peer"
+    } else if *generation > generation_ceiling {
+        "fence_generation_beyond_ceiling"
+    } else {
+        return true;
+    };
+    warn!(
+        event = EVENT_ADMIN_SITE_REPLICATION_STATE,
+        component = LOG_COMPONENT_ADMIN,
+        subsystem = LOG_SUBSYSTEM_SITE_REPLICATION,
+        result,
+        origin = %origin,
+        generation = *generation,
+        generation_ceiling,
+        "ignoring inadmissible peer-edit fence"
+    );
+    false
+}
+
 /// True when a strictly newer edit from the same origin site already landed
 /// here. No lock on the sending side can order deliveries issued by two
 /// nodes of that site, so ordering is decided here, on the generation the
@@ -9515,6 +9616,7 @@ impl Operation for SRPeerEditHandler {
         let ilm_expiry_override = sr_edit_ilm_expiry_override(&req.uri);
         let endpoint_refresh_requested = queries.get("refresh-targets").is_some_and(|value| value == "true");
         let commit_fence = peer_edit_fence(&queries);
+        let fence_generation_ceiling = peer_edit_fence_generation_ceiling(edit_generation_wall_clock());
         let local_endpoint = site_replication_local_endpoint(&req.uri, &req.headers);
         let (refresh_id, incoming) = if endpoint_refresh_requested {
             let refresh: EndpointRefreshRequest = read_site_replication_json(req, "", false).await?;
@@ -9532,6 +9634,11 @@ impl Operation for SRPeerEditHandler {
         let outcome = update_site_replication_state_when_changed(move |state| {
             let mut incoming = incoming;
             let local_peer = local_peer_at_endpoint(commit_endpoint, state);
+            // The fence is self-reported — the shared service account means
+            // the sender cannot be identified — so it is honoured only after
+            // the admissibility check, against the same state it will gate.
+            let commit_fence = commit_fence
+                .filter(|fence| peer_edit_fence_is_admissible(state, &local_peer.deployment_id, fence, fence_generation_ceiling));
             // Ordering fence: the sending site allocates the generation under
             // its state-object lock, so a delivery that lost the race carries
             // a generation this site has already passed. Applying it would
@@ -11913,6 +12020,17 @@ mod tests {
             handler_block.contains("record_applied_peer_edit_generation(state, origin, *generation);"),
             "SRPeerEditHandler must record the applied generation so later stale deliveries are recognised"
         );
+        // Fence hardening: origin and generation are self-reported by a
+        // caller the shared service account cannot identify, so the handler
+        // must pass the fence through the admissibility check — against the
+        // same state the fence gates, i.e. inside the transaction — before
+        // reading or raising any high-water mark.
+        assert!(
+            handler_block.contains(
+                ".filter(|fence| peer_edit_fence_is_admissible(state, &local_peer.deployment_id, fence, fence_generation_ceiling))"
+            ),
+            "SRPeerEditHandler must admit a fence only through peer_edit_fence_is_admissible inside the state transaction"
+        );
         // P1-15 PR2: both halves of the fence and the edit they fence share
         // ONE transaction. Checking the fence against a state read outside the
         // lock would let the check pass on one snapshot and the write land on
@@ -13207,6 +13325,94 @@ mod tests {
         assert!(peer_edit_delivery_is_stale(&state, origin, generation - 1));
     }
 
+    /// A fence is self-reported: every site authenticates peer traffic with
+    /// the same site-replicator credential, so a compromised peer can stamp
+    /// ANY origin with ANY generation. The receiver must refuse to let such
+    /// a stamp touch the high-water marks — an origin it does not replicate
+    /// with, its own deployment id, and a generation no genuine counter
+    /// could have reached are all ignored, and ignoring one plants no mark.
+    #[test]
+    fn forged_peer_edit_fences_cannot_poison_the_high_water_marks() {
+        let mut state = SiteReplicationState {
+            peers: BTreeMap::from([
+                (
+                    "site-local".to_string(),
+                    PeerInfo {
+                        deployment_id: "site-local".to_string(),
+                        ..peer("local", "https://local.example:9000")
+                    },
+                ),
+                (
+                    "site-victim".to_string(),
+                    PeerInfo {
+                        deployment_id: "site-victim".to_string(),
+                        ..peer("victim", "https://victim.example:9000")
+                    },
+                ),
+            ]),
+            ..Default::default()
+        };
+        let ceiling = peer_edit_fence_generation_ceiling(edit_generation_wall_clock());
+
+        // An origin outside the current membership is refused outright...
+        let unknown = ("site-unknown".to_string(), 4u64);
+        assert!(!peer_edit_fence_is_admissible(&state, "site-local", &unknown, ceiling));
+
+        // No site delivers edits to itself: a fence claiming the receiver as
+        // its origin is forged by construction, current peer or not.
+        let own = ("site-local".to_string(), 4u64);
+        assert!(!peer_edit_fence_is_admissible(&state, "site-local", &own, ceiling));
+
+        // A generation past the ceiling (the u64::MAX poisoning) is refused
+        // even when it names a current peer.
+        let poisoned = ("site-victim".to_string(), u64::MAX);
+        assert!(!peer_edit_fence_is_admissible(&state, "site-local", &poisoned, ceiling));
+        let barely_over = ("site-victim".to_string(), ceiling + 1);
+        assert!(!peer_edit_fence_is_admissible(&state, "site-local", &barely_over, ceiling));
+
+        // With no mark planted, the victim's genuine deliveries keep
+        // applying, and its admitted fence works end to end.
+        let genuine = ("site-victim".to_string(), 1u64);
+        assert!(peer_edit_fence_is_admissible(&state, "site-local", &genuine, ceiling));
+        assert!(!peer_edit_delivery_is_stale(&state, &genuine.0, genuine.1));
+        record_applied_peer_edit_generation(&mut state, &genuine.0, genuine.1);
+        assert_eq!(state.applied_edit_generations.get("site-victim"), Some(&1));
+    }
+
+    /// The generation ceiling must admit both live generation shapes: the
+    /// small plain counter pre-hybrid sites still allocate, and the
+    /// wall-clock-floored hybrid of `next_peer_edit_generation` — including
+    /// one running AHEAD of the receiver's clock by cross-site skew, which
+    /// is the whole reason the allowance exists, and the ceiling itself,
+    /// the last admissible value.
+    #[test]
+    fn peer_edit_fence_ceiling_admits_counter_and_clock_seeded_generations() {
+        let state = SiteReplicationState {
+            peers: BTreeMap::from([(
+                "site-origin".to_string(),
+                PeerInfo {
+                    deployment_id: "site-origin".to_string(),
+                    ..peer("origin", "https://origin.example:9000")
+                },
+            )]),
+            ..Default::default()
+        };
+        let now = edit_generation_wall_clock();
+        let ceiling = peer_edit_fence_generation_ceiling(now);
+
+        let counter = ("site-origin".to_string(), 42u64);
+        assert!(peer_edit_fence_is_admissible(&state, "site-local", &counter, ceiling));
+
+        // One hour ahead of the receiver's clock: within the skew allowance.
+        // Shrinking the allowance to zero must turn this fence away.
+        let clock_ahead = ("site-origin".to_string(), now + 60 * 60 * 1_000_000_000);
+        assert!(peer_edit_fence_is_admissible(&state, "site-local", &clock_ahead, ceiling));
+
+        // The boundary is inclusive: rejection starts strictly past it.
+        let at_ceiling = ("site-origin".to_string(), ceiling);
+        assert!(peer_edit_fence_is_admissible(&state, "site-local", &at_ceiling, ceiling));
+    }
+
     /// P1-15 review follow-up: a site that leaves the mesh drops below two
     /// peers, which clears its state object and restarts its generation
     /// counter at zero. A mark left over from its previous membership would
@@ -13244,6 +13450,104 @@ mod tests {
         assert!(!peer_edit_delivery_is_stale(&reloaded, "origin-site", 1));
     }
 
+    /// The unilateral-removal rejoin gap the hybrid clock closes. The origin
+    /// was removed while unreachable, but THIS site never dropped it from
+    /// its peer map, so the load-time mark pruning never fired and the mark
+    /// from the previous membership survives. The origin's recreated state
+    /// object restarts its counter, and with a plain `previous + 1` counter
+    /// every delivery it sent — generations 1, 2, … below the stale mark —
+    /// would be silently acked-and-dropped until the counter caught up. The
+    /// wall-clock floor in `next_peer_edit_generation` lifts the restarted
+    /// counter over every value the deleted lifetime handed out. Reverting
+    /// the allocation to the plain counter (dropping the wall-clock max)
+    /// turns the not-stale assertion red.
+    #[test]
+    fn hybrid_generation_unfences_a_rejoined_origin_whose_counter_restarted() {
+        // First lifetime of the origin's state object: two allocations, both
+        // capped by the wall clock at their own allocation.
+        let mut first_life = SiteReplicationState::default();
+        let straggler = next_peer_edit_generation(&mut first_life);
+        let last_applied = next_peer_edit_generation(&mut first_life);
+        assert!(last_applied > straggler, "allocations must be strictly increasing");
+
+        // The receiver applied up to `last_applied` and keeps the origin in
+        // its peer map across the unilateral removal — reloading must keep
+        // the mark, which is exactly why pruning cannot cover this case.
+        let mut receiver = SiteReplicationState::default();
+        receiver.peers.insert(
+            "origin-site".to_string(),
+            PeerInfo {
+                deployment_id: "origin-site".to_string(),
+                ..peer("origin", "https://origin.example:9000")
+            },
+        );
+        record_applied_peer_edit_generation(&mut receiver, "origin-site", last_applied);
+        let mut receiver = parse_site_replication_state(&serde_json::to_vec(&receiver).expect("serialize")).expect("reload");
+        assert_eq!(receiver.applied_edit_generations.get("origin-site"), Some(&last_applied));
+
+        // The origin rejoins with a RECREATED state object: counter back at
+        // zero. The wall-clock floor must lift its first allocation over the
+        // previous lifetime's mark…
+        let mut second_life = SiteReplicationState::default();
+        let restarted = next_peer_edit_generation(&mut second_life);
+        assert!(
+            !peer_edit_delivery_is_stale(&receiver, "origin-site", restarted),
+            "the recreated lifetime's first allocation ({restarted}) must not be fenced by the previous lifetime's mark ({last_applied})"
+        );
+        record_applied_peer_edit_generation(&mut receiver, "origin-site", restarted);
+
+        // …while a pre-removal delivery still in flight stays below the new
+        // floor and remains correctly fenced — the rollback the fence exists
+        // to reject.
+        assert!(
+            peer_edit_delivery_is_stale(&receiver, "origin-site", straggler),
+            "a pre-removal in-flight delivery ({straggler}) must stay fenced after the rejoin"
+        );
+    }
+
+    /// Marks recorded before the hybrid clock existed are small plain-counter
+    /// values, far below any wall-clock allocation: a restarted origin passes
+    /// them as soon as the SENDER runs the hybrid clock — nothing changes on
+    /// the wire or in the receiver, so pre-hybrid receivers get the fix too.
+    /// The other direction is unchanged: among plain-counter values the
+    /// generation order still fences the delivery that lost the race.
+    #[test]
+    fn hybrid_generation_passes_marks_recorded_by_plain_counter_receivers() {
+        let mut receiver = SiteReplicationState::default();
+        record_applied_peer_edit_generation(&mut receiver, "origin-site", 57);
+        assert!(peer_edit_delivery_is_stale(&receiver, "origin-site", 56));
+        assert!(!peer_edit_delivery_is_stale(&receiver, "origin-site", 57));
+
+        let mut rejoined = SiteReplicationState::default();
+        let restarted = next_peer_edit_generation(&mut rejoined);
+        assert!(
+            !peer_edit_delivery_is_stale(&receiver, "origin-site", restarted),
+            "a wall-clock allocation ({restarted}) must clear a plain-counter mark (57)"
+        );
+    }
+
+    /// The `previous + 1` half of the hybrid clock: allocations stay strictly
+    /// increasing even when the wall clock cannot move them forward — two
+    /// allocations inside one clock tick, or a clock that stepped backwards
+    /// mid-lifetime (a counter already ahead of the wall clock advances by
+    /// exactly one per allocation instead of jumping back). Dropping the
+    /// `previous + 1` half (allocating bare wall time) turns this red.
+    #[test]
+    fn hybrid_generation_is_strictly_increasing_when_the_clock_stalls() {
+        let mut state = SiteReplicationState {
+            // A counter far ahead of any wall clock this test will see.
+            edit_generation: u64::MAX / 2,
+            ..Default::default()
+        };
+        assert_eq!(next_peer_edit_generation(&mut state), u64::MAX / 2 + 1);
+        assert_eq!(next_peer_edit_generation(&mut state), u64::MAX / 2 + 2);
+        // Saturation pins at the ceiling instead of wrapping; the equal-value
+        // escape (`applied > generation` is false for equal) keeps deliveries
+        // applying rather than fencing the origin out.
+        state.edit_generation = u64::MAX;
+        assert_eq!(next_peer_edit_generation(&mut state), u64::MAX);
+    }
+
     #[test]
     fn test_retry_stats_for_state_counts_pending_and_failed() {
         let state = SiteReplicationState {
@@ -16044,10 +16348,77 @@ mod tests {
             generations.len(),
             "two nodes took the same edit generation, so their deliveries cannot be ordered: {generations:?}"
         );
+        // The hybrid clock allocates `max(wall nanos, previous + 1)` — the
+        // persisted counter is the largest allocation, and the `+ 1` half
+        // keeps allocations distinct even inside one clock tick.
+        assert_eq!(
+            Some(&load_site_replication_state().await.expect("reload").edit_generation),
+            unique.last(),
+            "the persisted counter must be the largest allocation handed out"
+        );
+    }
+
+    /// The unilateral-removal rejoin, end to end across the state object's
+    /// real lifecycle: dropping below two peers clears the object (the
+    /// counter dies with it), and the recreated object's first allocation —
+    /// raced by two nodes — must clear the previous lifetime's values via
+    /// the wall-clock floor, so a receiver still holding the old mark
+    /// accepts the restarted counter instead of fencing it.
+    #[tokio::test(flavor = "multi_thread", worker_threads = 4)]
+    #[serial]
+    async fn test_recreated_state_object_allocates_over_the_previous_lifetimes_mark() {
+        publish_ready_iam_context().await;
+        let seed = || SiteReplicationState {
+            peers: ["site-a", "site-b"]
+                .into_iter()
+                .map(|name| (name.to_string(), peer(name, &format!("https://{name}.example:9000"))))
+                .collect(),
+            ..Default::default()
+        };
+
+        save_site_replication_state(&seed()).await.expect("seed state");
+        let straggler = update_site_replication_state(|state| Ok(next_peer_edit_generation(state)))
+            .await
+            .expect("first-life allocation");
+        let last_applied = update_site_replication_state(|state| Ok(next_peer_edit_generation(state)))
+            .await
+            .expect("first-life allocation");
+        // A receiver that never dropped this site from its peer map holds
+        // this mark across the removal.
+        let mut receiver = SiteReplicationState::default();
+        record_applied_peer_edit_generation(&mut receiver, "origin-site", last_applied);
+
+        // Unilateral removal: the site drops below two peers, which clears
+        // its state object and the counter with it.
+        let mut departed = seed();
+        departed.peers.remove("site-b");
+        save_site_replication_state(&departed).await.expect("clear state");
         assert_eq!(
             load_site_replication_state().await.expect("reload").edit_generation,
-            generations.len() as u64,
-            "the persisted counter must account for every allocation"
+            0,
+            "clearing the state object must take the counter with it"
+        );
+
+        // Rejoin recreates the state object; two nodes race the first
+        // allocation of the new life.
+        save_site_replication_state(&seed()).await.expect("recreate state");
+        let node_a = tokio::spawn(update_site_replication_state(|state| Ok(next_peer_edit_generation(state))));
+        let node_b = tokio::spawn(update_site_replication_state(|state| Ok(next_peer_edit_generation(state))));
+        let generation_a = node_a.await.expect("node a task").expect("node a allocation");
+        let generation_b = node_b.await.expect("node b task").expect("node b allocation");
+        assert_ne!(generation_a, generation_b, "racing allocations must stay distinct");
+
+        // The receiver's stale mark must not fence the restarted counter…
+        let restarted = generation_a.min(generation_b);
+        assert!(
+            !peer_edit_delivery_is_stale(&receiver, "origin-site", restarted),
+            "the recreated life's first allocation ({restarted}) must clear the previous life's mark ({last_applied})"
+        );
+        record_applied_peer_edit_generation(&mut receiver, "origin-site", restarted);
+        // …while the cleared life's in-flight leftovers stay fenced.
+        assert!(
+            peer_edit_delivery_is_stale(&receiver, "origin-site", straggler),
+            "a pre-removal in-flight delivery ({straggler}) must stay fenced after the rejoin"
         );
     }
 

scripts/security/check_preview_release_workflow.sh

@@ -195,6 +195,13 @@ IFS= read -r -d '' expected_docker_automatic_guard <<'EOF' || true
 EOF
 expected_docker_automatic_guard=${expected_docker_automatic_guard%$'\n'}
 require_job_if "$docker_workflow" "build-check" "$expected_docker_automatic_guard"
+require_line "$docker_workflow" '      source_ref: ${{ steps.check.outputs.source_ref }}' "Docker source ref output"
+require_line "$docker_workflow" '            source_ref="$HEAD_SHA"' "automatic Docker source ref"
+require_line "$docker_workflow" '              source_ref="$tag_ref"' "manual Docker source ref"
+require_line "$docker_workflow" '          ref: ${{ needs.build-check.outputs.source_ref }}' "Docker release source checkout"
+require_line "$docker_workflow" '          SOURCE_REVISION="$(git rev-parse HEAD)"' "Docker source revision resolution"
+require_line "$docker_workflow" '          LABELS="$LABELS,org.opencontainers.image.revision=$SOURCE_REVISION"' "Docker revision label"
+require_absent "$docker_workflow" 'org.opencontainers.image.revision=${{ github.sha }}' "Docker revision must not use the workflow branch SHA"
 
 docker_manual_guard=$(awk '
   $0 == "              *-preview*)" { in_preview = 1 }