fix(hnsw): refresh dead-node set instead of caching it for the index lifetime

[shaunpatterson]

Problem

(*persistentHNSW).removeDeadNodes loaded the tombstoned-vector set (the persisted __vector_dead posting) into ph.deadNodes once, on the first call, and never refreshed it:

// TODO add a path to delete deadNodes
if ph.deadNodes == nil { /* load from vecDead posting, build the map */ }

The index instance lives in the factory's indexMap for the life of the process, so once populated the set was frozen. Any vector deleted after the first removeDeadNodes call was never seen by the neighbour filter, and dead UIDs leaked back into edge lists during subsequent inserts and neighbour updates. The standing // TODO add a path to delete deadNodes acknowledged the gap.

Fix

Cache the dead set as an immutable snapshot tagged with the transaction read timestamp (TxnCache.Ts()), published via atomic.Pointer[deadSnapshot]:

• Correctness — a transaction with a newer StartTs reloads the set and observes deletions committed since the previous load.
• Snapshot isolation — the shared cache only ever advances in time. If a newer snapshot is already cached, an older-ts caller is served its own ts-scoped set and does not install it, so a transaction never observes deletions newer than its own snapshot.
• Performance — an index rebuild streams every key at one StartTs, so the JSON is parsed once and reused across the many removeDeadNodes calls per insert.
• Concurrency — the index instance is shared across the goroutines that drive a rebuild. Publication is lock-free via atomic.Pointer + CompareAndSwap; the snapshot map is immutable after construction, so the filter reads it without synchronization.

The hot path (removeDeadNodes) delegates to loadDeadNodes(tc), which returns the cached snapshot on a timestamp match and otherwise loads, builds, and CAS-publishes a new one.

Test

• TestRemoveDeadNodesRefreshesAcrossTimestamps — cross-timestamp refresh + within-snapshot stability
• TestRemoveDeadNodesSnapshotIsolation — an older-ts caller is unaffected by a newer cached snapshot
• TestLoadDeadNodesConcurrent — concurrent mixed-timestamp loads under -race

go test -race ./tok/..., go vet, gofmt all clean; golangci-lint (repo config) reports no new findings from this change.

🤖 Generated with Claude Code