BFM-decomp/cookbook/C0174.md

§168 — THE COUSIN TIER (P30 S49, 2026-08-12): h_seq's exact-hash brittleness, measured — and the similarity map above it

The finding (Drew's smell, byte-verified). The frontier's "4,513 unique singletons + 1,817 small families" picture was substantially a GROUPING ARTIFACT. h_seq — our loosest tier — is an exact hash of the mnemonic skeleton, so a single inserted instruction makes two functions total strangers. Per-location compilation inserts instructions constantly: a per-location constant crossing the 16-bit boundary turns a 1-instruction li into lui+ori; a bigger jump table adds a case; a dropped flag test deletes one. Probed on the full frontier (120-pair sample per band): in the 0.85–0.99 similarity band, 86/120 near-pairs differ by PURE insertion/deletion (25 with lui inside the indel block — the li-expansion tell); replace-heavy pairs are the minority. Specimen: ov_SC06_010:0x8017bebc (753 ins, "singleton") is 0.987 similar to a MATCHED function in the same binary — 12 small edit blocks, mostly 1–2-instruction insertions. Not a unique function; the B2 walker family wearing a one-instruction disguise per site.

The map (tools/family_cousins.py → .run/family_cousins.json + docs/family-cousins.md). Distinct open skeletons (one per h_seq class, ×N copies collapsed via the family map, R32-checked against an independent sigs+stubs recount), streams via family_remap.stream_words, mnemonic-class tokens (register/imm-blind, h_seq-granularity), K=8 shingle candidate index (posting cap 400), difflib ratio, union-find at ≥0.85. Measured at S49 HEAD (whole-frontier, 11,627 instances / 584,448 ins — totals exact by construction):

category	units	open fns	open ins	lever
A-prop (unit holds a lane-A skeleton)	197	1,286	68,729	family_sweep propagation
seeded (≥0.85 MATCHED body exists)	418	1,652	50,422	crack by EDITING a proven body
cousin-multi (no seed, ≥2 instances)	1,552	5,449	249,799	1 crack seeds the unit's rest
cold (alone at 0.85)	3,240	3,240	215,498	full-price crack

The genuinely cold tail is 215k ins — 37% of the remainder, not 90% — and it is small-function shaped (2,692 of 3,240 cold units < 100 ins; only 43 > 300 ins). Corroborating probes: only 823 of 3,760 overlay singletons are truly alone at their vram address; main's "structurally barren" verdict HOLDS at the similarity tier (94% of main's stub mass < 0.70 to anything anywhere; ~72 tiny fns have seeds — a small correction, not a reversal); call-sequence grouping adds ~94 matches beyond content similarity (redundant — not a tier worth building).

The three laws:

1. A cousin is a SEEDED CRACK, never a remap. Skeleton drift means the member needs its own compile — family_sweep/remap_hseq require an identical skeleton by construction. The seed's value is the draft: the agent edits a proven C body instead of writing one. (Prior-NOTES seeding measured 7/9 and 10/12; a full ≥0.98-similar matched body is a strictly stronger prior.)
2. Rank waves by UNIT weight, not family weight. A crack's real yield is its whole cluster — family members plus the cousins it seeds. The S49 slate: 40 targets = 33,304 unit ins vs 25,010 family-ranked — +33% on the same agent count. (The S29 pricing law, one tier up.)
3. Discount short-function similarity. Shared prologue/epilogue boilerplate inflates ratios on small nins; the targets carry nins and seed-sim separately so wave sizing can discount seeds on <~40-ins functions. A 0.87 on 20 ins is a hint; a 0.95 on 400 ins is a body.

Ops: regen order is sigs → family_hseq.py → family_cousins.py (it fails loud on a stale map, with the regen command in the error). --targets 40 --wave wave7 emits the crack_wave.js slate; seed refs resolve to the matched body's location (engine_core.h DEFINE_ macro vs inline src/<bin>/*.c def). The survey RANKS and SEEDS; only the whole-binary byte-gate banks (G3/P9).