Fix clinical dose-verification & safety gaps missed by early-merged #38

[BigSimmo]

Summary

PR #38 (22 RAG fixes) was merged early at ab27de3, before the 6 critical blocking safety fixes were added. Those fixes lived on fix/rag-pipeline-stage3-generation (26a8a6d) and never reached main. PR #39 (security/safety) and PR #40 (RAG/search) since landed but did not address any of B1-B6. This PR re-implements the still-missing blockers against current main.

Gap analysis — all 6 were still missing on main

ID	Issue	Status before this PR	Action
B1	Substring dose match (2.5 mg ⊂ 12.5 mg)	Still broken on main (despacedHaystack.includes(...))	Newly fixed — exact token-set membership, fails closed
B2	Unicode superscript ×10⁹/L mangled to 2.0×10	Still broken on main (no superscript folding)	Newly fixed — fold ⁰¹²…⁹ → ASCII before extract/match
B3	Percentages never extracted (trailing \b after %)	Still broken on main	Newly fixed — % branch no longer ends in \b
B4	Gate ignored answerSections[].body	Still broken on main (scanned only answer.answer)	Newly fixed — verify each section body, scoped to its citation_chunk_ids
B5	safeFallbackAnswer back-fills citations + grounded:true, skips numeric gate	Still broken on main	Newly fixed — fail closed (ungrounded, no citations) + run numeric gate; extractive recovery + strong-retry routing updated
B6	TOCTOU race in retry route (SELECT-then-unguarded-UPDATE)	Still broken on main	Newly fixed — single conditional UPDATE … WHERE status≠processing OR lock stale/null; 0 rows ⇒ 409
N1	Numeric fallback verified against ALL results	Still on main	Fixed — fail closed when no cited chunk maps to a result

None of B1-B6 were already addressed by #39/#40; #39/#40 only hardened the regex units and the warning text. All six were re-applied here, ported to main's diverged code (e.g. retry route keeps main's status:"pending", regex keeps main's mmol/[lL] units).

Empirical B1 verification (the top-priority dose fix)

Run against the new verifyAnswerNumbers:

answer="2.5 mg"  vs source="12.5 mg" => UNVERIFIED
answer="12.5 mg" vs source="12.5 mg" => VERIFIED
answer="500 mg"  vs source="1500 mg" => UNVERIFIED
answer="2.0"     vs source="12.0"    => UNVERIFIED

All four required outcomes confirmed — wrong doses now fail closed; exact doses verify.

Testing

• npm run typecheck — passes
• npm run lint — passes (0 errors; 1 pre-existing unrelated warning in src/lib/chunking.ts)
• npm run test — 381 passed (55 files), including the new B1-B6/N1 regression tests
• Targeted suites (answer-verification, rag-trust, private-access-routes) — 79 passed

New/updated regression tests cover: B1 (2.5↔12.5, 500↔1500, 2.0↔12.0 + exact match), B2 (superscript extract + superscript↔ASCII match), B3 (percentage extract + mismatch flag), B4 (section-body dose flagged / verified), B5 (fail-closed fallback + numeric gate on parse-failure prose), B6 (refuse fresh-processing 409 / allow stale-or-non-processing), N1 (fail closed on unmapped citation).

Do not merge without review. This is a medical app — the dose-faithfulness gate must fail closed.

🤖 Generated with Claude Code