Summary

First of the three E-3c PRs (instrument → short-circuit → figure-aware selection). The run-#58 targeting misses ("no dose figure/regimen", "no schedule/interval") cannot be diagnosed without the actual answer texts — and answers are deliberately not retained at rest (privacy design). This PR adds --dump-answers <path> to eval:answer-quality: per-case {id, question, intent, route, routing_reason, grounded, confidence, citation_count, targeting verdict, answer text, sections} JSON written into the canary artifact directory at eval time.

• Default off; --json summary output byte-identical; confirmation line only in human mode (the --json-out convention).
• parseArgs / buildAnswerDumpRecord exported behind an import.meta main-guard (the eval-quality.ts precedent) so tests import without executing the provider-backed eval.
• The canary's targeting step now passes the flag, so every future answer_quality_eval=true dispatch — including the E-4 pair — yields the texts. Content sensitivity matches the existing tee'd log: all-public corpus, no env values.

The ledger row also records the accepted E-3c design (PR-B: pre-generation validated-extractive short-circuit generalizing the existing LAI pattern, net ~−110 rag.ts lines; PR-C: intent-figure-aware lead promotion with a claim-support nuke-proofing guard).

RAG impact: no retrieval behaviour change — measurement instrument only; no runtime answer path touched.

Verification

• New tests/eval-answer-quality.test.ts — 4/4 (flag parsing incl. default-off + --json independence; dump-record shape incl. missing-text tolerance)
• npm run typecheck, npm run check:github-actions, npm run check:ci-scope, eval-canary workflow test, prettier — all clean
• No provider calls in this PR

Risk and rollout

• Risk: minimal — an opt-in file write in an eval script plus one workflow arg; the eval's exit code and stdout are unchanged.
• Rollback: revert.
• Provider or production effects: none from this diff; the flag only takes effect on explicitly-dispatched canary runs.

Clinical Governance Preflight

• Source-backed claims still require linked source verification before clinical use — no behaviour change
• No patient-identifiable document workflow was introduced or expanded — dump content is eval-fixture questions + answers over the all-public corpus, artifact-scoped (30-day retention)
• Supabase target remains Clinical KB Database (sjrfecxgysukkwxsowpy) — no access changes
• Service-role keys and private document access remain server-only
• Demo/synthetic content remains clearly separated from real clinical sources
• Source metadata, review status, and outdated/unknown-source behavior remain conservative
• Deployment classification/TGA SaMD impact was checked: no clinical decision-support behaviour changed — measurement only

🤖 Generated with Claude Code

https://claude.ai/code/session_01EXsJcLrbZUXwnBeG91cVo9

Generated by Claude Code