BFM-decomp/docs/phase25-frontier-map.md

Phase 25 — Frontier Map (rtu_match-measured, 2026-07-10c)

The systematic map of ALL remaining endgame work, replacing the stale-object-tainted wave-2 leftover list. Built from .run/family_manifest.json (current matched status) + a tools/rtu_match.py measurement pass over every draftable exemplar (.run/frontier_measured.json). Fleet 73.97%.

rtu_match caveat (same blind spot as the §41b object-probe): rtu_match masked-diffs the fn's .text only — it is BLIND to rodata (switch jump tables) and to LINK errors. So an rtu-MATCH on a jump-table or last-referencer fn is a FALSE crackable-now; those are classified into the specialized tiers below via the whole-binary gate / the T7-M1 findings, not by rtu_match alone.

The endgame = 3 buckets (multi-member families, by ov_SC01_077 membership)

bucket	families	byte-MB	action
draftable	42	3.0	crack ONE exemplar → propagate ×134 (the core crack work)
matched-free	235	5.9	exemplar already matched → propagate siblings ×134 (mostly earlier-sweep-failure walls)
absent	2481	2.1	no ov077 member; small fns in other overlays — LAST

Bucket 1 — draftable 42, classified by rtu_match (.text closeness)

sub-bucket	n	fleet-leverage (Σ inst×nins)	tier / tool
reconcile-first (fan-out)	24	456,027	CC1-FAIL raw → strip conflicting externs → rtu_match → crack. Wave-3: reconcile→~70% MATCH.
DIFF-crack (fan-out)	7	123,153	compiles, DIFF 52–349 → the §42c lever-set
near (permuter/fan-out)	6	60,836	DIFF ≤25 (incl. 0x801549f8=3, 0x80164e40=8) → cheap crack / permuter-ILS
F-jumptable (rtu-blind)	4	94,604	_o0 giants (0x8013c414/8013bd74/8013c0f8) + 0x8012ace0 — rodata switch table; §8 jump-table workflow
M-linkwall (rtu-blind)	1	8,040	0x8016d688 — last-referencer of a scratch data symbol; manual undefined-syms

Crack fan-out target = 37 (reconcile-first 24 + DIFF-crack 7 + near 6). Full ranked list: .run/frontier_crack_targets.json. Top by leverage: 0x80133cd4 (399×134), 0x80166994 (369×134), 0x8014fbc0 (22×1996), 0x8014d820 (304×134), 0x801670e4 (279×134), 0x80135480 (258×134), 0x80163ec8 (234×134), 0x8016cbc0 (209×134), 0x80166690 (193×134), 0x80132144 (27×539), 0x8014fe60 (95×268) … down to the ≤100-ins med tail.

Giants caveat: the 200+-ins GIANTS (0x80133cd4, 0x80166994, 0x8014d820, 0x80135480, 0x80163ec8) are the likely Fable5 short-list — the rtu_match+§42c fan-out cracks ≤~180-ins reliably (wave-3 evidence); the giants may need Fable5's compiler-internals reasoning. The map does NOT pre-spend Fable5 — the fan-out attempts them and the ones it can't close become the surgical Fable5 list.

Bucket 2 — matched-free 235 (propagate, no cracking)

Exemplar already matched; siblings unmatched. These are the RESIDUAL of the T7-part-1 matched-free harvest (which banked 16,512) — i.e. the families that FAILED that sweep: the //@EDIT class (file-scope edits not carried per sibling), split-TU-type conflicts, macro-defined bodies. Recovery = propagation enhancements (a family_sweep --edit-remap that carries the exemplar's //@EDIT edits per sibling, symbol-remapped; the split-TU-type reconcile). Deterministic, ~0 agent tokens once the tooling lands. The 266 wave-3-part-2 sweep drops (func_80133AB0/80136824) live here.

Chunked execution plan (smart chunks, biggest-leverage-first)

1. Crack fan-out over the 37 — the core. rtu_match engine (reconcile-first + §42c levers), leverage-ranked waves. Each crack → family_sweep --reconcile ×134. GIANTS that resist → Fable5 short-list.
2. Propagation enhancements → recover matched-free 235 + the wave-3 sweep drops — cheap bulk (5.9 MB).
3. F-jumptable 4 + M-linkwall 1 — specialized workflows (§8 / manual undefined-syms).
4. Fable5 surgical pass — only the giants/new-idioms the fan-out flags.
5. absent 2481 — last, low ROI.

Reproduce

tools/family_manifest.py → .run/frontier_draftable.json (42) → .run/crack3/measure_frontier.py → .run/frontier_measured.json + .run/frontier_crack_targets.json (37).