BFM-decomp/cookbook/C0175.md

§169 — THE MICRO-ADAPT LANE (P30 S49): edit a proven body, don't crack a new one

The lane. For a cousin (§168) whose drift vs an already-matched seed is ≤3 blocks / ≤6 mnemonic tokens, the work is an EDIT, not a crack: copy the seed's byte-proven C, re-point every symbol at the TARGET's .s, apply the one semantic change the diff implies (an indel of lui/ori/addiu is almost always a changed CONSTANT — the value is readable in the card's member words), gate. Tools: family_cousins.py --adapt-cards → .run/adapt_cards.json; tools/wave/adapt_wave.js.

Measured, pilot of 30 (S49): 25 agent-MATCH (83%, 0 refuted by the adversarial verifier) → 16 banked (64% MATCH→bank; 53% end-to-end) at 2.7M tokens, 28 haiku / 2 sonnet ≈ 30k tok per banked instance vs a crack wave's ~75k. The edit is small even when the body is not, so route by body size only to cover the COPY cost (haiku ≤60 ins), not the reasoning.

Three laws the pilot paid for:

1. A banked cousin usually does NOT propagate. Cousins are byte-VARIANT by construction, so there is typically no byte-identical sibling to stamp: 2 of 8 propagated (+4). The card's reach counts COUSIN members (future adapt fuel), NOT dedup copies. Price a card at ~1 instance plus optionality, never at ×reach. (R14 — I framed it as ×reach before measuring.)
2. The residual failure class is INTEGRATION, not codegen. 9 of 25 MATCHes died at the whole-binary gate, clustered per-TU (5 in one binary, 2 in another): standalone-MATCH, host-TU-rejected — the §52b/§161c declaration surface, which the reconcile ladder exists for. A per-TU recovery pass belongs between the wave and the scale-up.
3. Never wrap a self-timing tool in a tighter outer timeout, and kill process GROUPS. gate_stage scales its own propagation timeout (1800 + 1800×banks); a 7200s outer cap killed a healthy 5-bank group mid-fleet-write AND orphaned its dedup_propagate child, which kept rewriting src/ through a subsequent git checkout --. Symptom: a later gate opens onto a dirty tree it did not create. Diagnosis: pgrep -af dedup_propagate. The same 5 drafts banked 5/5 when re-run untimed. Generalizes S48's "guard the CAMPAIGN, not the process."

MEASURED TWICE (S49 waves 7a/7b) — and the spread law. 7a: 30 cards → 25 MATCH (83%) → 16 banked (64%). 7b (relaxed thresholds, 59 cards): 48 MATCH (81%) → 44 banked (92%), 75% end-to-end, ~157k tok/banked fn. The bank-rate jump is NOT the thresholds — it is TU SPREAD: 7a concentrated drafts in few destination TUs (5 in one binary) and its 9 gate failures were per-TU declaration collisions between sibling drafts; 7b spread 48 drafts over 35 TUs, mostly one apiece, and only 2 groups lost anything. Build slates that spread across destination TUs; when several drafts must share one TU, gate them together and expect the §161c reconcile ladder.

Sizing the lane (S49, seeded pool = 828 skeletons / 1,652 members / 50,422 ins): today's ≤3/≤6 cut captures 47% of the pool's instructions; ≤6 blocks/≤16 tokens captures 78% (+247 skeletons, +15,576 ins) and is still a holdable edit; past ≤8/≤24 the curve flattens (+4 skeletons) — the remainder is genuinely different code, i.e. a seeded crack. Absolute thresholds mis-sort large bodies (MIXED median is 35 ins with a 29% edit fraction — small bodies rewritten heavily — while only 25 skeletons are big-body/small-edit), so union the absolute rule with edit_fraction ≤ 0.20 rather than replacing it. Note the tier boundary: a pure IMMEDIATE change leaves the mnemonic stream identical ⇒ same h_seq ⇒ that member is family_sweep/imm_map work, never a cousin. Every replace block in a cousin card is a genuinely different MNEMONIC.