#3599 Count ambiguous clinical tables as image evidence and attach chunk image_ids

User description

Split out of #3590 (now merged). After a full review, this PR keeps only the two picture-evidence changes. The rest was removed because it lacked evidence of benefit or carried clinical risk.

RAG impact: behaviour change — which images count as clinical evidence (isClinicalImageEvidence) also feeds image-need scoring in clinical-search.ts. The owner has chosen to use the latest green eval-canary (27ef74141f) as the baseline and run the post-merge canary themselves. That baseline predates #3590, #3591 and #3596, so a difference may come from those as well.

What changes

1. Ambiguous clinical images can count as evidence (src/lib/image-filtering.ts). An image whose classifier says "ambiguous" counts only when all of these hold: its type is clinical table, flowchart, medication chart or risk matrix; it has zero admin signal; and it has at least one clinical signal or a relevance score of 0.65 or more. Explicit clinical_evidence and non_clinical classes are unchanged.
2. A chunk's own image_ids are attached during hydration (src/lib/rag/rag-hydration.ts). The pictures attached are the ones the index explicitly linked to that chunk, so attribution stays exact.

Removed in review

• Title-phrase tie-break: its test passes with or without it, so there's no evidence it changes ordering. Per docs/rag-behaviour/refuted-approaches.md, a comparator change needs a discriminating test.
• Community-admission preference: a hard-coded title/file-name match (admissioncommunitypts) that assumed "admission" means community admission unless the question says "inpatient".
• Admission/discharge side evidence in the main answer and model context: it could add back evidence that conflict-aware selection had excluded.
• Document-wide picture fallback: a picture from elsewhere in a document could appear next to text it does not support.
• Bot commits outside search scope: Qodo's operational-alert, roster-dispatch and source-row edits. The source-row edit broke the pinned test answer-source-rail.dom.test.tsx.

Local evidence

• Lint, npm run typecheck, Prettier and npm run check:diff-integrity pass.
• npm run test:focused over the changed files: 4,058 tests passed. The one failure was answer-source-rail.dom.test.tsx, which passes now that the source-row edit is reverted.
• No provider-backed run was made for this PR.

Clinical Governance Preflight

• If this PR changes clinical prompts, answer generation, citation logic, retrieval ranking, source policy, or clinical UI presentation, I reviewed docs/clinical-governance.md and recorded the clinical safety impact above.
• If this PR adds or changes clinical guidance sources, I checked approval metadata and provenance requirements (npm run check:source-governance where applicable).
• Real patient identifiers, clinical notes, and private health information are not included in code, tests, fixtures, screenshots, or PR text.

🤖 Generated with Claude Code

CodeAnt-AI Description

Include relevant clinical charts and chunk-linked images in search evidence

What Changed

• Ambiguous clinical tables, flowcharts, medication charts, and risk matrices now count as evidence when they show clinical relevance and no administrative signals.
• Images explicitly linked to a text passage now appear with that passage, even when they are on a different page. Images from another document are not attached.

Impact

✅ Relevant clinical charts included in search evidence
✅ Chunk-linked images stay with supporting text
✅ No cross-document image attribution

@codeant-ai ask: Your question here

@codeant-ai ask: Can you suggest a safer alternative to storing this secret?

@codeant-ai: Your feedback here

@codeant-ai: Do not flag unused imports.

@codeant-ai: review