Fix flaky HNSW vector index unit test

[kriszyp]

Summary

Fixes the flaky HierarchicalNavigableSmallWorld indexing > can delete and update and search with vector index with one dimension test seen on unrelated PRs (e.g. run 26469662007).

Purpose

Two independent fragilities surfaced intermittently. HNSW level assignment uses Math.random(), so each run produces a different graph topology and the failure mode varied across runs (~15% local repro rate):

1. Test asserted exact recall. verifySearch used assert.deepEqual on HNSW top‑5 vs brute‑force top‑5. HNSW is approximate; under the test's stress sequence (200 inserts → delete 100 → upsert all 200 with many near‑duplicate modular‑arithmetic vectors), recall@5 can dip below 100%. Replaced with a per‑rank distance tolerance (HNSW distance must be within 0.05 of brute‑force best at that rank).
2. validateConnectivity crashed on stale neighbor refs. HNSW deletes only scrub the deleted node's own connection list, and asymmetric edges can survive (this is logged as asymmetry detected warnings). When BFS traversed a stale neighbor ID, getSync returned undefined and the BFS threw TypeError: Cannot read properties of undefined (reading 'level'). It also inflated visited.size by adding stale IDs before confirming the node existed, which broke isFullyConnected. BFS now uses safeGetSync, skips missing nodes, and tracks queue membership separately so visited reflects real nodes only.

Where to focus review

• Tolerance value (0.05) in unitTests/resources/vectorIndex.test.js: chose this empirically — comfortably covers observed approximation gaps (largest delta seen was ~0.016 cosine) without masking real regressions. Open to a tighter value if reviewers prefer.
• Underlying HNSW characteristic, not a bug per se: asymmetric edges + delete‑only‑scrubs‑own‑list means stale refs are a known tradeoff (full symmetry on delete would require an inverted index). This PR makes the debug utility robust to that; it doesn't change deletion semantics. Filing a follow‑up may be worth considering if we want strong invariants here.
• Integration tests: skipped test:integration:all since no integration tests exercise HNSW (grep confirms zero HNSW/vectorIndex refs under integrationTests/). test:unit:resources (536 tests) all pass.

Verification

• Before fix: 5–7 failures per 30 local mocha runs.
• After fix: 0 failures over 50 consecutive runs.
• Cross‑model review: Codex (no concerns) + Gemini 3.1 Pro (one suggestion incorporated — added assert(results.length >= 5) bounds check).

🤖 Generated with Claude Opus 4.7

[claude]

Reviewed; no blockers found.