refactor(evolve): collapse three scanners behind one runScanner core + collectFn seam

[sethvoltz]

What

Collapses the three hand-copied evolve scanner pipelines (scan-friction, scan-preferences, scan-dreaming) into one deep runScanner core fed by per-scanner Taxonomy adapters, and adds a second injectable seam (collectFn) so the full collect → batch → score → bucket pipeline is testable without a DB — closing the real test gap.

Why

The three scanners shared the same four-stage shape with ~35–40 lines of identical scaffolding hand-copied into each file. The model seam (scoreFn) already existed, but turn-collection hit Postgres directly, so no test drove the full pipeline — only the isolated bucket functions were covered.

Changes

• collect.ts (new) — extracted OrchestratorTurn, collectOrchestratorTurns, dbTurnIdToLine to a neutral module (breaks the run-scanner↔scan-friction cycle).
• run-scanner.ts (new) — runScanner<P, T, Ctx> + the Taxonomy interface + RunScannerOptions + SCAN_BATCH_SIZE, plus de-duplicated truncate/clamp/severityFor/severityRank. The core only ever touches the turn_id join key on T (via a ScoredRow constraint); every domain distinction stays inside each bucket.
• scan-*.ts — each is now a Taxonomy + a thin scanX = (opts) => runScanner(taxonomy, opts, ctx) wrapper. Public function names/signatures unchanged. Dreaming's richer bucketing (payload-in-first-pointer, separate note-pointer cap, recall→severity bump, promote/reinforce verb) preserved whole; DreamEvidence threaded through as the opaque ctx.
• run-scanner.test.ts (new) — end-to-end pipeline tests the seam unlocks: cross-session swap, default-30 + override batching, cross-batch merge into one bucket, log-and-continue, per-category caps, dreaming reinforce-merge + payload round-trip + evidence threading.

Scope

Pure internal refactor — no Signal shape change, no propose.ts/rank.ts change, public exports unchanged.

Verification

• Package tsc clean; 140 tests pass (128 preserved verbatim + 12 new); daemon evolve-dreaming consumer green; prettier clean.
• Adversarial multi-agent review confirmed behavior byte-identical (md5-verified) across all three bucket functions, all three scoring prompts, and the collect logic; export surface preserved; core extraction leak-free. The 4 real issues it surfaced (a stale OrchestratorTurn test import, two thin-coverage batching gaps, one redundant assertion) are fixed in this branch.

🤖 Generated with Claude Code

https://claude.ai/code/session_01SeavKpZNAxkrokvKhQF8Ai