fix(m12p6): 7th-edge — correct reseed seqno + skip suspect HNSW graph on reseed-pending close (rc11)

Fixes the rc9 over-correction: forcing baseline for ALL nodes (including
caught-up ones) caused needless reseed cascades. Now only divergent nodes
(frontier > baseline) use baseline as the reseed seqno; at/below-baseline
nodes use frontier+1 so the leader picks cheap catch-up vs snapshot.

Also skips the HNSW graph checkpoint on SIGTERM when the shard is reseed-
pending: the in-memory index reflects suspect/divergent data the next boot
discards, so saving it risks a "Failed to read vectors" failure on the
post-reseed open. Durable checkpoints and WAL flush still run.

close_shared() gains a save_graphs bool; shutdown_inner_impl() is the
shared implementation; node.rs passes !reseed_pending.

k8s/cluster/statefulset.yaml

@@ -84,7 +84,7 @@ spec:
               mountPath: /data
       containers:
         - name: tidaldb
-          image: registry.threesix.ai/tidal/server@sha256:818c923368ecca6e433c42a13672132e9b467c449fa23c035201fe7189f1f52b  # m12-rc9 (= rc8 + Bug 6: reseed latch re-baselines from the stream baseline, not frontier+1, so a divergent-rejoin node force-installs the snapshot instead of looping)
+          image: registry.threesix.ai/tidal/server@sha256:2e1f7a0e2603fc135f6e79a8c376ca265f8ad74724904b1922df310e8e6da306  # m12-rc11 (= rc10 + corrected Bug 6: only force the re-baselining snapshot when frontier > baseline (divergent); at/below baseline keep frontier+1 so a caught-up node does NOT needlessly reseed-cascade)
           imagePullPolicy: IfNotPresent
           # The image ENTRYPOINT is the bare binary. We override the command with
           # a tiny /bin/sh wrapper (the bookworm-slim runtime HAS a shell) so we

tidal-server/src/cluster/election_driver.rs

@@ -293,14 +293,23 @@ impl ElectionRuntime {
                 // node is missing in the PREVIOUS stream's numbering
                 // (`own.frontier + 1`); the marker boot re-baselines onto the
                 // leader's stream via the snapshot regardless.
-                // m12p6: re-baseline from the new term's STREAM BASELINE, not
-                // `frontier + 1`. When the node's frontier sits at/above the
-                // baseline (post-baseline old-term data), `frontier + 1` lands
-                // above the leader's tail → `wal_covers` answers needed=false → the
-                // reseed install no-ops → re-detect → self-restart loop. A
-                // `from_seqno <= baseline` forces the snapshot that re-baselines the
-                // node onto the leader's committed history.
-                node.latch_reseed_marker(ReseedReason::SnapshotRequired, baseline);
+                // m12p6: pick the resume seqno so the leader's `wal_covers` makes
+                // the RIGHT needed-decision. When the node's frontier sits ABOVE the
+                // baseline it has post-baseline old-term (DIVERGENT) data: `frontier
+                // + 1` would land above the leader's tail → `wal_covers` answers
+                // needed=false → the install no-ops → re-detect → self-restart loop.
+                // Re-baseline from `baseline` (`<= baseline` forces the snapshot that
+                // discards the divergent suffix). When the frontier is AT/BELOW the
+                // baseline the node is genuinely BEHIND (or exactly caught up):
+                // `frontier + 1` is correct and lets the leader's needed-decision
+                // pick a cheap catch-up vs snapshot — forcing `baseline` here would
+                // reseed an already-caught-up node and cascade restarts.
+                let from_seqno = if position.frontier > baseline {
+                    baseline
+                } else {
+                    position.frontier.saturating_add(1)
+                };
+                node.latch_reseed_marker(ReseedReason::SnapshotRequired, from_seqno);
                 self.joined_term.store(term, Ordering::Release);
                 true
             }

tidal-server/src/cluster/node.rs

@@ -1113,7 +1113,15 @@ impl ShardReplica {
             // because no `{data_dir}/vector` file was written on SIGTERM. `close_shared`
             // takes `&self` and is idempotent (the `closed` CAS), so the trailing
             // `drop(db)` and any later `Drop` from a lingering clone are no-ops.
-            if let Err(e) = db.close_shared() {
+            // m12p6 7th-edge: when this shard is RESEED-PENDING (a divergent /
+            // quarantined node about to be re-baselined — its marker latched), skip
+            // the HNSW-graph save. The in-memory index reflects suspect data the
+            // next boot DISCARDS via the snapshot install, so saving it is
+            // pointless and risks committing a graph the next open fails to read
+            // ("Failed to read vectors" → a slow rebuild). The durable checkpoints
+            // + WAL flush still run.
+            let reseed_pending = matches!(self.reseed_marker_store.load(), Ok(Some(_)));
+            if let Err(e) = db.close_shared(!reseed_pending) {
                 tracing::error!(
                     region = %self.region_name,
                     error = %e,
@@ -1124,7 +1132,8 @@ impl ShardReplica {
             drop(db);
             tracing::info!(
                 region = %self.region_name,
-                "region cluster node shutdown: database closed (checkpoint + WAL fsync + HNSW graph)"
+                reseed_pending,
+                "region cluster node shutdown: database closed (checkpoint + WAL fsync + HNSW graph unless reseed-pending)"
             );
         }
     }
@@ -2698,15 +2707,19 @@ impl ShardReplica {
         baseline: u64,
     ) {
         self.cluster_metrics.set_divergence_quarantined(true);
-        // m12p6: the reseed must re-baseline from the new term's STREAM BASELINE,
-        // NOT `frontier + 1`. A divergent node's frontier sits AT/ABOVE the
-        // baseline, so `frontier + 1` lands above the leader's WAL tail → the
-        // leader answers `needed=false` (nothing to ship above its tail) → the
-        // reseed install no-ops, the divergent suffix is re-detected, and the node
-        // self-restart-loops. A `from_seqno <= baseline` makes `wal_covers` return
-        // needed=true → the snapshot installs and re-baselines onto the leader,
-        // discarding the divergent suffix.
-        self.latch_reseed_marker(ReseedReason::Quarantine, baseline);
+        // m12p6: re-baseline a DIVERGENT node (frontier ABOVE the new term's
+        // baseline) from `baseline` — `frontier + 1` would land above the leader's
+        // WAL tail → `wal_covers` answers needed=false → the install no-ops and the
+        // divergent suffix self-restart-loops. A `from_seqno <= baseline` forces the
+        // snapshot that discards the suffix. (When the frontier is at/below the
+        // baseline — not the divergent case — `frontier + 1` lets the leader pick a
+        // cheap catch-up vs snapshot, avoiding a needless reseed cascade.)
+        let from_seqno = if frontier > baseline {
+            baseline
+        } else {
+            frontier.saturating_add(1)
+        };
+        self.latch_reseed_marker(ReseedReason::Quarantine, from_seqno);
         tracing::error!(
             term,
             tail_term,

tidal/src/db/lifecycle.rs

@@ -56,16 +56,31 @@ impl TidalDb {
     /// Callers MUST have quiesced the write/apply path first (the cluster node
     /// stops the ship queue and signals the segment receiver before calling this).
     ///
+    /// `save_graphs` (m12p6 7th-edge): when `false`, SKIP the HNSW-graph
+    /// checkpoint. The cluster node passes `false` when the shard is RESEED-PENDING
+    /// (a divergent/quarantined node about to be re-baselined): its in-memory
+    /// index reflects suspect/divergent data the next boot will DISCARD via the
+    /// snapshot install, so persisting it is pointless AND risks committing a
+    /// graph the next open then fails to read ("Failed to read vectors") — forcing
+    /// a slow rebuild. Everything else in the close (durable checkpoints, WAL
+    /// flush/marker) still runs.
+    ///
     /// # Errors
     ///
     /// Surfaces the first durable-flush failure, exactly like [`close`](Self::close).
-    pub fn close_shared(&self) -> crate::Result<()> {
-        self.shutdown_inner()
+    pub fn close_shared(&self, save_graphs: bool) -> crate::Result<()> {
+        self.shutdown_inner_impl(save_graphs)
     }
 
-    /// Internal shutdown logic shared by `close()` and `Drop`.
-    #[allow(clippy::too_many_lines)]
+    /// Internal shutdown logic shared by `close()` and `Drop` — always saves the
+    /// HNSW graph (the clean-shutdown path; only a reseed-pending close skips it,
+    /// via [`close_shared`](Self::close_shared)).
     pub(crate) fn shutdown_inner(&self) -> crate::Result<()> {
+        self.shutdown_inner_impl(true)
+    }
+
+    #[allow(clippy::too_many_lines)]
+    fn shutdown_inner_impl(&self, save_graphs: bool) -> crate::Result<()> {
         // CAS: first caller to flip false -> true executes the shutdown body.
         if self
             .closed
@@ -228,10 +243,17 @@ impl TidalDb {
             // logged but never fails close()/Drop — a missing graph simply rebuilds
             // on the next open, which is correct, just slow. The checkpoint thread
             // is already joined above, so the registry is quiescent here.
-            crate::db::state_rebuild::checkpoint_embedding_graphs(
-                &self.embedding_registry,
-                crate::db::state_rebuild::vector_graph_dir(&self.config).as_deref(),
-            );
+            if save_graphs {
+                crate::db::state_rebuild::checkpoint_embedding_graphs(
+                    &self.embedding_registry,
+                    crate::db::state_rebuild::vector_graph_dir(&self.config).as_deref(),
+                );
+            } else {
+                tracing::info!(
+                    "skipping HNSW graph checkpoint on close: reseed-pending shard, \
+                     its in-memory index is suspect/divergent and will be re-baselined"
+                );
+            }
             if let Err(e) = storage.flush() {
                 tracing::error!(error = %e, "storage flush failed during shutdown");
                 first_err.get_or_insert(e);