feat(synthesis): representative claim digest + map/reduce version split

[bryanmatthewsimonson]

Real-use problem

After #191 attached ~2,335 claims across ~109 articles, an Opus brief came back deeper but narrower than the earlier 2-claim brief: position holders collapsed to 1–3 each, and central cruxes (furin cleavage site, DEFUSE proposal) vanished. The brief said why itself — "13 with extracted claims in the digest."

Cause: DIGEST_CLAIM_CAP = 150 filled the reduce's claim index first-come by article order, so ~13 claim-dense articles consumed all 150 slots and the rest of the corpus never reached the reduce — starving cross-article proposals and claim-anchored cruxes of the whole tail.

Three moves (all reduce-side — the map cache is untouched)

1. Representative claim index (selectDigestClaims): keep every is_key claim, then round-robin one-per-article so the 150 span the whole corpus instead of clustering in its densest few. claimIndex now carries is_key.
2. Map/reduce version split — the important structural fix. corpusExtractKey (the map-extract cache key) now keys on MAP_PROMPT_VERSION (held at corpus-v2), not the overall CORPUS_PROMPT_VERSION (bumped to corpus-v3). So this reduce change and all future reduce-prompt tuning stale briefs via corpusInputHash without orphaning the expensive cached extracts. The two start equal and diverge exactly when a reduce-only change lands. (Without this, every reduce tweak would blow away the map cache from feat(synthesis): cache per-article map extracts to skip redundant LLM calls #192 — caching and tuning were at war.)
3. Reduce-prompt completeness nudge: list EVERY holder (the list is the corpus map, not a sample) and enumerate ALL major cruxes including scientific disputes surfaced in the extracts, not only those with a claim in the index.

Why this is safe

Reduce-side only — the map prompt is unchanged, so your 100 cached extracts survive (verified: corpusExtractKey keys on MAP_PROMPT_VERSION, pinned at corpus-v2). Briefs go stale and re-run cheaply. No wire kind, no schema change, no fused score.

Verification

• Tests: representative selection (breadth over 200 articles + is_key priority + dense-article can't monopolize); the map/overall version decoupling (a reduce bump doesn't move the cache key); the prompt breadth lines.
• npm test → 1976 pass · npm run build clean · npm run lint 0 errors.

After merge + reload: stay on Opus, re-run — the 100 extracts reuse for free, and the reduce should produce a brief with full holder lists and the furin/DEFUSE cruxes restored. Compare against the current brief; if a crux still reads thin, the remaining lever is a modest DIGEST_CLAIM_CAP raise.

🤖 Generated with Claude Code