[acgxv]

Changes

Recover original sources when extraction produced no note or the first vector candidates missed the evidence. Add a bounded supplemental HTTP/MCP recall path that combines original text, derived text, and graph incidence, while preserving source bytes, permission checks, and native hash/range citations.

Classify historical graph results at the caller's as_of. Add alias declaration authority so extracted or legacy aliases cannot silently cross entity boundaries. Existing aliases remain stored as unknown; alias-only identity lookup now requires a caller declaration. Keep PG, Qdrant, and full 4096-dimensional vectors.

Evidence and limits

• Same 36 frozen questions, corpus, embeddings, and initial candidates. Final test factual score: 32/32 (previous 30/32); strict claim-supported score: 30/32 (unchanged). Original strict citations regress by two cases while contrast cases improve by two. The four development cases are reported separately.
• Both trials and all failures are retained. Six semantically identical requests reuse prior responses; 30 changed final requests run once. No answer-driven resampling. Assistant semantic reviews are not an independent blinded judge.
• Remove eight off-topic graph-only candidate additions caused by ambiguous legacy aliases. This does not demonstrate independent graph scoring uplift or repair all graph fragmentation.
• Explicit lexical recovery can scan the visible corpus; large-scale query cost is not measured.
• New cost: $0.403979. Cumulative settled: $8.503227; conservative exposure including old unresolved reservations: $9.587028, below the $10 ceiling.

Validation

462 Rust unit tests, 105 integration tests, 183 Python contracts, four focused acceptance tests, formatting, docs, Clippy, and vstyle passed. One existing integration test reported a leaky pipe; a focused rerun passed without it. Complete logs are retained.

All 36 final API evidence packs, supplemental recalls, and claim checks match the recorded trial. All eleven original-column data fingerprints match the restored fixture; new alias authority is audited separately. Current and previous evidence verifiers pass offline. Owned API/Docker resources are removed; backups, unrelated containers, and shared images are preserved. No production migration was run.

Report: docs/evidence/benchmarking/2026-10-12-recall-time-graph.md.