BFM-decomp/cookbook/C0176.md

§170 — THE A-PROP WORD-DIFF CARD (P30 S49): the lane that had no owner

The gap. family_cousins.py --adapt-cards builds cards only for seeded units, so lane A — 1,700 open functions / 76,419 ins, every one with a byte-proven matched sibling — was skipped by construction. A lane-A member shares its family's h_seq with a MATCHED sibling, so the cousin (mnemonic) diff is EMPTY: the differences live in WORDS — a struct field offset, a data symbol, a register. The mechanical remap refuses exactly these (unresolved immediates / regalloc drift), which is why the ≥16-reach head swept 0 banked / 245 failed / 188 refused.

The card (--aprop-cards → .run/aprop_cards.json): positional WORD diff member-vs-matched sibling (same length by construction, so no alignment needed), each site classified IMM / REG / OTHER and disassembled both sides, grouped BY FAMILY. One agent then learns the parameterization once and emits N drafts — a 54-member family is ONE card, not 54. Measured on the head: 13 families / 433 members, median 2 differing words per member; 5 members were byte-IDENTICAL to the seed (pure dedup_propagate work, not agent work — check for those first).

Measured (S49 calibration, 9 batches / 108 members, 4.5M tok): 98 agent-MATCH (91%) — the best agent rate of any wave type — → 56 banked (57%), ≈ 80k tok per banked function vs the per-member cousin card's 157k and a crack wave's 400k+. The agent notes are substitutions, not decompilations ("src struct = D_801D0358"; "cb=func_8017FBA4; 0x12C, 0x4B0, 0x38E…").

⚠ The conversion gap. 91% agent → 57% gate is the WORST conversion measured (the per-member cousin wave ran 81% → 92%). 14 of ~50 groups banked zero.

Hypothesis, not yet proven: family-batched cards CONCENTRATE members into one destination TU by construction — precisely the §169 collision the spread law names, which held the 7a wave to 64%. Re-gate the unbanked drafts ONE PER TU before scaling a batched A-prop wave; if the spread law is the cause the same drafts bank, and the fix is to interleave families per gate batch rather than to redraft.

STRUCK 2026-08-13 (S50) — REFUTED. Batching was never the discriminator (5-draft groups banked 5/5; the two "concentrated" groups banked 12/12 and 10/10 once fixed). The cause was a STALE SEED SYMBOL: match_one is blind to a relocation's target NAME, so a carried-over D_8018xxxx scores MATCH standalone and dies at link. 23 of 24 banked after a mechanical rebase; A-prop's true conversion is 87%, not 57%. See §171 — and note the test itself was never needed: the answer was already sitting in .run/harvest_failed.<binary>.classified.txt.

Three ops notes paid for on the head: (1) family_sweep --hseq defaults to --band substantial, so tiny/mid families are silently OUT OF SCOPE — a "never attempted" verdict can be a band default, not a difficulty (5 of 13 heads); (2) dedup_extend extends MACRO-backed groups only — for a matched INLINE def use dedup_propagate --addr; (3) that tool names its own blocker precisely (missing file-scope extern (CARRY-FIXABLE): D_…) — a data symbol the matched body's own TU never declares at file scope blocks every sibling until it is carried.