rag.ts decomposition — part 1: quote verification, source block, numeric verification, context selection

[BigSimmo]

Summary

First tranche of the src/lib/rag.ts decomposition (7,874 → ~7,450 lines), applying the same verbatim move-only playbook used for the ClinicalDashboard decomposition. Four leaf families extracted, all cycle-free, one commit per move with per-move gates (typecheck + focused vitest + lint + prettier):

1. Quote-verification family → src/lib/rag-quote-verification.ts — normalizeQuoteVerificationText, tableFactQuoteText, sourceTextForQuoteVerification, isExactSourceQuote, sanitizeQuoteCards, sanitizeConflictsOrGaps, enrichGroundedReviewCitations.
2. Source-block family → src/lib/rag-source-block.ts — truncateForModel, compactContextText, richTableSourceContextEnabled, tableSnippetForFact, formatTableFactForSourceBlock, buildRagSourceBlock.
3. Numeric-verification application → src/lib/answer-verification.ts — applyNumericVerification, unboldUnverifiedNumbers, hasActionableNumericContext now live beside the verifyAnswerNumbers/extractNumericTokens primitives they wrap.
4. Model-context selection → src/lib/rag-context-selection.ts — capPerDocumentCrowding, selectModelContextResults (+ their consts).

Supporting re-homes (all one-way, no new cycles): allowedChunkMap → citations.ts; safeRecord + metadataText → rag-answer-text.ts; appendRoutingReason → rag-routing.ts; and rag.ts's resultCitation was an exact duplicate of citations.citationFromResult, so it now imports that under the old alias (a dedup win). rag.ts re-exports every moved name that tests or routes import, so no consumer changed.

The debt log (docs/process-hardening.md) gained a dated part-1 entry including the measured coupling for the remaining regions: the extractive-answer (L) and answer-quality (M) families are mutually entangled on main and must move together in a dedicated pass; retrieval variants (H) additionally depend on owner-scope helpers; context packing (J) belongs with the cache region (F). The pre-drift region map should not be trusted for part 2.

Verification

• npm run verify:cheap — pass (runtime, sitemap, lint [5 pre-existing warnings, 0 errors], typecheck, 1,181 unit tests)
• npm run verify:ui — 109/109 passed
• npm run build — pass (production build)
• npm run verify:release before release or handoff confidence claims
• npm run format:check — pass (prettier run per move)
• npm run eval:retrieval:quality (must stay 23/23) — could not run here (needs live keys). These moves are answer-side and verbatim (no retrieval selection/ordering/scoring logic changed), but per the standing gate please run it locally before merge and paste the summary.
• npm run eval:rag / eval:quality — no answer-generation logic changed (verbatim moves); same caveat as above if you want belt-and-braces
• npm run check:production-readiness — N/A: no env/privacy/deployment behavior change
• npm run check:deployment-readiness — N/A

Per-move gates that ran green on every commit: tsc --noEmit, focused vitest (rag-trust, rag-content-accuracy, answer-ranking, answer-verification, rag-context-budget — 43–59 tests per move), eslint, prettier. Every moved block was diff-verified byte-identical against its rag.ts source before deletion.

Clinical Governance Preflight

Code-motion only in the answer pipeline; no behavior, prompt, ranking, or data-access change.

• Source-backed claims still require linked source verification before clinical use
• No patient-identifiable document workflow was introduced or expanded without explicit governance approval
• Supabase target remains Clinical KB Database (sjrfecxgysukkwxsowpy)
• 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 when clinical decision-support behavior changed — no clinical decision-support behavior changed

Notes

• Part 2 (extractive answers + quality gates together, then retrieval variants and the cache region) should start from the coupling notes in the debt-log entry, not the original region map.
• Moves 7–8 in the approved sequence are retrieval-touching; when they land, the golden retrieval eval is a hard pre-merge gate.

🤖 Generated with Claude Code

https://claude.ai/code/session_01XPVNBVo4cg9PEYtNhJZBQY

---

Generated by Claude Code

[supabase]

This pull request has been ignored for the connected project sjrfecxgysukkwxsowpy because there are no changes detected in supabase directory. You can change this behaviour in Project Integrations Settings ↗︎.

---

Preview Branches by Supabase.
Learn more about Supabase Branching ↗︎.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.