Test debt from #217: at-scale ordering parity, real-chain self-heal integration, consolidated coverage gaps

[edheltzel]

Deferred test debt from the #217 review cycle (reviewer report findings 225-4 + consolidated gaps; RedTeam correctness pass; see .agents/atlas/artifacts/2026-07-13-pr224-225-gate-decision.md locally).

1. At-scale ordering parity (RedTeam correctness): the committed KNN-vs-brute-force parity test uses 3–8 well-separated vectors — it structurally cannot exercise near-tie reordering or scale. Add a parity test at O(1k+) vectors including near-tie clusters (expectation: top-K membership identical; near-ties ≤~1e-6 separation may permute).
2. Real-chain self-heal integration test (225-4): production heals via getDb → ensureSchema → createVecTable on reopen + first-query ensureVecIndexSynced count-mismatch rebuild. No test drives that chain against a legacy on-disk cosine index (current tests call the pieces directly / stub the sync).
3. Consolidated coverage gaps (reviewer): bruteforce-over-empty-embeddings backend value; embed-throws-inside-stale-liveness-window reporting; formatHybridModeNote(true, "none"); dimension-mismatch exclusion at reindex; marked-duplicate over-fetch × converted distance values.

Refs #217, #146.