fix(namespace): noun-record updates preserve legacy inline HNSW adjacency — the placeholder-adjacency write stamped out pre-codec records' stored connections (crash-window unreachability); codec-era records were never at risk (empty field is the blob marker); pin covers the legacy shape
Some checks failed
CI / Node 22 (push) Failing after 7m32s
CI / Node 24 (push) Failing after 7m24s
CI / Bun (latest) (push) Successful in 12m17s

This commit is contained in:
David Snelling 2026-08-03 15:37:04 -07:00
parent 7a28a94639
commit 4679c89458
2 changed files with 68 additions and 2 deletions

View file

@ -77,8 +77,27 @@ export class SaveNounOperation implements Operation {
? null ? null
: await this.storage.getNoun(this.noun.id) : await this.storage.getNoun(this.noun.id)
// Save new noun // PRESERVE stored graph state on updates. Callers stage this op with
await this.storage.saveNoun(this.noun) // placeholder adjacency ({connections: empty, level: 0}) because the
// vector index owns those values and persists them at flush. Codec-era
// records (2.4.0+) carry an empty connections field by design (adjacency
// lives in a separate compressed blob — the placeholder is harmless), but
// LEGACY pre-codec records store adjacency INLINE: writing the
// placeholder over one stamped out its stored connections, leaving a
// crash window (until the next flush) where a reload found the node
// unreachable. Stale adjacency in that window is tolerable — HNSW
// self-corrects at the reindex flush; EMPTY adjacency is silent recall
// loss. The read above is already paid for rollback; preservation is free.
const toSave: HNSWNoun =
previousNoun && this.noun.connections.size === 0
? {
...this.noun,
connections: previousNoun.connections || this.noun.connections,
level: previousNoun.level ?? this.noun.level
}
: this.noun
await this.storage.saveNoun(toSave)
// Return rollback action // Return rollback action
return async () => { return async () => {

View file

@ -145,3 +145,50 @@ describe('level field shadow — user metadata named level is a real field', ()
expect(EXPECTED_INDEX_EPOCH).toBe(2) expect(EXPECTED_INDEX_EPOCH).toBe(2)
}) })
}) })
describe('noun-record writes never stamp over stored graph state', () => {
let brain: Brainy
beforeEach(async () => {
brain = new Brainy({
requireSubtype: false,
storage: { type: 'memory' as const },
embeddingFunction: stubEmbedding
})
await brain.init()
})
afterEach(async () => {
await brain.close()
})
it('a data-changing update preserves LEGACY inline connections in the record', async () => {
// Codec-era records carry an EMPTY connections field by design (the
// adjacency lives in a separate compressed blob) — the clobber window
// exists only for legacy pre-codec records whose adjacency is inline.
// Simulate one: write the record with inline connections directly.
const id = await brain.add({
data: 'legacy-shaped node',
type: NounType.Concept,
metadata: { n: 1 }
})
const storage = (brain as any).storage
const rec = await storage.getNoun(id)
const legacy = {
...rec,
connections: new Map([[0, new Set(['00000000-0000-4000-8000-00000000aaaa'])]]),
level: 1
}
await storage.saveNoun(legacy)
const before = await storage.getNoun(id)
expect(before.connections.size).toBeGreaterThan(0)
// A data-changing update stages SaveNounOperation with placeholder
// adjacency — the legacy inline connections must survive the write.
await brain.update({ id, data: 'completely re-embedded text' })
const after = await storage.getNoun(id)
expect(after.connections.size).toBeGreaterThan(0)
expect(after.level).toBe(1)
})
})