Autocut: score-discontinuity result-sizing on the rerank separatrix (v0.42.3.0)

Fixes one wave of #1663 (the floor/ceiling retrieval redesign): recommendation #2, "the direct fix for the 20-vs-1 problem, no LLM call."

What it does. Search returns the confident handful instead of a fixed top-K. When the cross-encoder rerank scores show a clear cliff, autocut cuts there — one obvious answer comes back as one result, a real cluster as that cluster, a broad ambiguous query still returns the full set. Default-ON in balanced/tokenmax; documented no-op in conservative (no reranker → no trustworthy signal).

Why it's trustworthy. gbrain measured (documented in return-policy.ts) that the raw RRF/cosine rank1→rank2 gap is ~identical whether rank-1 is right or wrong — not a separatrix. The cross-encoder rerank score is. So autocut cuts on rerank_score only, and only where the reranker ran.

Reviews run on this branch

• /plan-eng-review — CLEAR (rerank-gating, default-ON + eval gate, DRY parallel-module).
• /codex (plan) — caught the unscored-tail recall risk → fixed by D4 (reranker_top_n_in = searchLimit in reranked modes, no unscored tail).
• /codex (pre-landing review on the diff) — caught a P1 introduced by the master merge: applyAliasHop() (master's recent retrieval work) injects an exact alias-match page after reranking with no rerank_score, and autocut would drop it when cutting. Fixed: applyAutocut gains a preserve predicate; hybrid passes r => r.alias_hit === true. Also a P2: cache-HIT meta now carries autocut/adaptive_return/mode/embedding_column. 3 new regression tests.

Eval gate (in-repo, runs in CI)

bun run eval:autocut (test/search/autocut-eval.test.ts) measures precision-lift-without-recall-regression over labeled qrels fixtures with modeled cross-encoder distributions — no API key, no sibling repo. Result: precision 0.33 → 0.94, recall 1.00 → 0.95, zero recall regression on enumeration queries. Live-corpus PrecisionMemBench remains an optional empirical confirmation, not a blocker.

Diff

• NEW src/core/search/autocut.ts — pure algorithm + resolve ladder + preserve predicate.
• mode.ts — autocut/autocut_jump knobs → knobsHash (KNOBS_HASH_VERSION 7→8, stacked on master's title_boost v=7); reranker_top_n_in = searchLimit for reranked modes.
• hybrid.ts — wired after adaptive-return + alias-hop, before the limit slice, first page only; cache-miss AND cache-HIT meta both carry the trimmed-set decision fields.
• rerank_score first-class on SearchResult; query op autocut boolean; --explain per-result rerank score + formatAutocutSummary; gbrain search modes + metric-glossary.
• Tests: pure-fn, agent-surface, IRON-RULE behavioral (PGLite via rerankerFn seam), preserve regression, eval gate, v=8 knobsHash + bundle pins.

Verification

• bun run verify — 29/29 green; bun run typecheck — clean; full search suite (377 tests) green.
• Full unit suite: the local machine was under heavy load and the parallel runner SIGTERM-killed shards on the wall-clock cap across attempts — zero actual test failures observed in any shard (shard 4 + serial completed clean; the killed shards showed no (fail) markers). CI runs the authoritative full suite on a clean machine. One known pre-existing, environment-dependent failure lives in master's own test/audit/batch-retry-audit.test.ts (its "ENOENT no-op" case assumes an empty default ~/.gbrain/audit; a real brain has audit files) — not in this diff, fails on master independently, passes in clean CI.

Not closing #1663

One wave. Remaining: query-shape router, structural exact-lookup tier, CRAG-style auto-escalation to think.

🤖 Generated with Claude Code