BFM-decomp/phase-ends/CURRENT_PHASE.md

CURRENT PHASE — Phase 28: The Endgame Engine

Started: 2026-07-15 · Generation: Gen2 (20th phase of the arc) · Plan approved: gate 1 ✅ (Drew, 2026-07-15) Baseline: v1.26.0 / PhaseEnd_Phase27 · Roadmap: docs/roadmap-to-100.md §3 P28 Re-scoped from the roadmap — reason below. Owner decisions this session: Fable5 = conditional slot (only if T3 names a NEW wall class); the 39 type-1 modules = deferred, documented, owned by a later phase.

---

Why this phase is re-scoped (the Phase-Start finding — byte-verified, R14)

P28 exists to measure the member-adapt close-rate (roadmap §6: "THE swing number"; all P28/P29 yield projections are withheld until it exists). Phase-Start verification found the evidence behind its presumed value does not survive contact with the bytes — the exact failure R35 was coined to prevent, by the phase that coined it.

Verified this session against the bytes:

1. The 0x8017BEBC refutation is a jump-table artifact, not structural variance. classify_member(ov_SC01_000→ov_SC01_001) → PURE, ndiff=2 (words 343/345 only): 3c01801a/8c224374 vs 3c01801f/8c22c44c = lui $at,%hi(jtbl_801EC44C) / lw $v0,%lo(jtbl_801EC44C)($at) (asm/ov_SC01_001/nonmatchings/ov_SC01_001_jr_8017A4AC/func_8017BEBC.s:380,382) — a switch jump table. config/overlays.mk:112 carves ov_SC01_000_jr_8017BEBC.o for the exemplar; :134 has no such entry for the member. family_remap.symbol_map emits only func_/D_ (family_remap.py:419-422) — it structurally cannot generate jtbl_<ADDR>, so the remap no-ops (2 map entries for a 952-ins fn) and the gate books a BYTE failure. The file's own comment (:403-418) documents this identical bug class for the other two prefixes, calling the result "indistinguishable from a genuine compiler wall." Third instance. The fix (tools/jtbl_family_bank.py, the carve ladder) exists and was never pointed at this family.
2. That family is the least representative in the population. Only 3 of 163 matched-exemplar families carry has_mid_jr (120 of 13,232 members). The roadmap generalized n=1, atypical, to the whole frontier.
3. All three Phase-26 exhaustion probes predate the fix that made such a probe honest. _carry_macros landed in Phase-27 T5 (commit:0637); the probes (tiny-IMM 0/241, PURE 0/134, pinned 0/133) are Phase 26. Phase 27's decision-log calls its own post-fix re-probe "a fourth phantom exhaustion proof" — implicitly numbering those three as the first three, recognizing the mechanism that would have faked them, and never re-running them.
4. The real population was never probed. 163 matched-exemplar families / 13,232 unmatched members / 656,379 templatable ins. Members: PURE 9,955 (75%) · IMM 3,140 (24%) · STRUCT 137 (1.0%). Families: PURE 98 / IMM 35 / MIXED 30. 96 PURE non-jr families / 7,148 members (54%) never honestly probed. The roadmap sizes its swing number on register-drift = STRUCT = 1.0% of the input.

The claim is "unmeasured", not "wrong" (P9). The pre-fix bug produced CC1-FAIL; if those probes failed with DIFF it does not explain them — their failure classification was never recorded, which is itself the defect. R35: a 0% from a broken tool and a 0% from a working one are the same number and opposite facts.

Two further roadmap premises are stale (verified): the parallel gate farm already exists (bulk_harvest.py, 0.4s/fn @ 8 workers ≈ 75× serial; drafting, not the gate, was the documented bottleneck) — what is missing is the same-binary axis; and both named grinder bugs are already fixed (commit:0327, commit:0200) — the real defect is a 22-entry never-retry blacklist built from a broken gate.

---

Tasks

• T0 — Tool-health + the 2 instrument fixes that gate an honest probe · xHigh · ⏳ IN PROGRESS
  • make tools-health green (audit-corpus + audit-cdecl + report, fail-closed).
  • family_remap.img_path (:32-35) hardcodes 0.4.dec → None for the 4 SC07 overlays → stream_words→None → classify_member → ("LEN",[]) → member silently dropped as not-templatable. Twin of the P27 T7 new_overlay.sh bug, left in a second tool. Fix + assert coverage (R32): a missing image fails loud.
  • Regenerate .run/family_hseq.json (covers 134 overlays; 138 sig files exist; carries pre-T7 68.9 vs the honest 67.0).
  • Verify: 138 overlays in the map; SC07 members classify to a real class, not LEN.
• T1 — jtbl_ prefix + the honest 0x8017BEBC re-probe (roadmap B2) · Max · blocked by T0
  • Add jtbl_<ADDR> to symbol_map's keys, belt-and-braces, on the file's own documented rationale.
  • Re-probe via tools/jtbl_family_bank.py (carve ladder raw→scoped→recovered→reconciled, :116-193) — not family_sweep (no carve step). --raw first (a reconciled body is TU-specific and poisons the template).
  • Bounded ~8 members, byte-gated, R22-clean. Either outcome is the deliverable. Correct calibration.md + decision-log.md (R31) in-session (R30).
• T2 — Purge the poisoned grinder blacklist · xHigh (small)
  • 22 permanent never-retry entries built from a known-broken gate (tooling-audit.md:927-931: 16/22 never compiled, 5 absent); worklist.md:50,56 marks two of them "none — MATCH". Purge; fix harvest_verify's --src/--asm-subdir split coupling that poisoned it; re-derive only from the fixed gate.
• T3 — The stratified templatability re-probe: THE swing number · Max · 🚩 milestone report to Drew · blocked by T0,T1
  • Strata: PURE non-jr (96 fam / 7,148 members) · IMM (35 fam, imm_map_tier1) · MIXED (30 fam). ~8–12 members/class, byte-gated, R22-clean.
  • Classify every failure (harvest_verify.classify_fail → DIFF/PLUMBING/CC1-FAIL/SKIP) — the distinction Phase 26 never recorded.
  • Do not score with masked_diff — mask_for drops HI16/LO16 (:92-124), blind to this delta class (would score all 112 0x8017BEBC members a perfect 0). The gate decides.
  • Deliverable: honest per-class close-rate + taxonomy → rewrite calibration.md's decisive table; decision-log (R31); Roadmap delta.
  • Decision point (P5d): high → T4 + a bounded first harvest (the gate's "measured engine throughput"); ~0 → thesis retired on evidence that holds, pivot to T5/T6.
• T4 — member_adapt.py, scoped to what T3 measured · xHigh · CONDITIONAL on T3
  • Reuse: classify_member (per-delta + index attribution), imm_map_tier1, remap's imm_map hook, jtbl_family_bank's ladder — the "delta engine" is largely written. The novel part is the post-flight diff (exemplar compiled in the member's own TU vs member target); pre- and post-flight can disagree for reasons outside the C body. On failure emit the per-member delta brief.
• T5 — Resident flag-plant → 100% or an honest wall dossier · xHigh→Max · independent
  • Fix progress.py first (R35): #if 0-blind (:418-485) → resident.c:868-925 double-counts func_800D00E4 as REAL and stub; :512 sums len()s instead of unioning sets. Reported 123/146=85.62%; honest 122/145=85.5%; third opinion sig-resident=144. Reconcile before planting a flag on the denominator.
  • Prefetch: resident already imported (~index.dat id 3) → one headless DecompileFunctions run over the 21 addrs (R23 stop-MCP; R29 /mcp prompt). None of the 21 cached (.run/ghidra_c/=861, all ov_SC01_077).
  • Targets (difficulty.resident.md): 7 non-jtbl leaves · 9 other · 5 jtbl. Ledger is ov_SC01_077-hardcoded (backlog.py:37,94) → scope to --binary.
• T6 — Gate throughput: the free win, then the missing axis · xHigh
  • Unconditional: the --chunk 1 double-build (harvest_verify.py:199-213) — the bisect re-runs attempt() on the same single element vs an unchanged baseline = a duplicate build. 1.35→1.0 builds/draft = ~26% fewer builds on the hot path, one line.
  • Conditional (T3 volume): the same-binary shard farm — shard by directory (the build is CWD-relative → no Makefile change); a shard must own its build/; cp -al hardlink asm/(815M)+extracted/(760M); 8–16 shards (15 GB RAM / WSL2 bound, not cores).
• T7 — Burn-down tracker + the stale-label sweep · xHigh (cheap)
  • Burn-down over progress --weighted (roadmap §1 velocity = a standing obligation with no instrument today).
  • worklist.py:194 reads source; the key is source_overlay → the R32 assertion rides a hardcoded default.
  • Stale 134→138 labels: fuel_manifest.reach_buckets, progress.fleet.md:7, second-oracle.md:60, backlog.py:22, worklist.py:14, roadmap:161-165.

Conditional — Fable5 slot (window ~7/19, Agent(model: fable), parallel-isolated): only if T3's taxonomy names a new wall class cheap-Opus has no idiom for. Distill same-session (R30). Fable5 discovers, cheap-Opus applies. Never behemoths.

---

Milestone (gate 2)

1. The swing number measured through working instruments — stratified, every failure classified, calibration.md rewritten on it, Roadmap delta stated. (A byte-grounded ~0% is as valid as a high rate — P9.)
2. Resident at 100% or an honest wall dossier — on a denominator progress.py and sig-resident agree on.
3. 140/140 byte-identical from a genuinely clean tree (R22), 0 NON_MATCHING (G4), all three gates green + fail-closed.
4. All three metrics + velocity reported (roadmap §1).

Standing invariants (roadmap §5)

Whole-binary byte-gate = sole arbiter (G3/P9) · R22 clean-fleet per banked batch · tool-health green before matching · one commit per task after this file is updated (Drew pushes, R6/R20) · idioms → cookbook in the producing session (R16/R30) · pivots → decision-log.md (R31) · effort/model toggles prompted, never assumed (R26/R27) · every tool fix carries a negative control (it must change an answer the old tool gave).

Blockers

(none yet)

Progress log

• 2026-07-15 — Phase Start. Load order read (PROJECT_CONTEXT + 27 PhaseEnds); no CURRENT_PHASE (P27 closed clean). Plan-mode + Max confirmed. 3 Explore agents surveyed the member_adapt substrate / the gate substrate / the frontier+resident. Phase-Start finding (above) byte-verified by me, not taken on an agent's word (R14). Plan approved (gate 1). Task list built (R28). Starting T0.

• 2026-07-15 — T0 (in progress). family_remap.img_path fixed: derives the payload from config/splat.<bin>.yaml's target_path (R33 — the file the BUILD reads, so it cannot drift from the bytes) and raises on a missing payload (R32 — a silent None was the defect).

  • Negative control (the fix must change an answer the old tool gave): ov_SC01_001 → 0.4.dec unchanged (no regression) · ov_SC07_006 None → …/1.4.dec · resident → …/FILE_010.dir/1.1 (works for free — useful for T5) · ov_SC99_999 → raises (was a silent None).

  • Downstream: all 233 shared substantial fns between ov_SC07_006 and ov_SC01_001 classify PURE (reloc-only — the most templatable class). Under the old tool every one returned LEN = "not templatable" and poisoned its family's diff_class to MIXED (family_hseq.py:141-143). Fourth instance of the project's dominant defect class, and it sat directly under the number P28 was chartered to measure.

  • .run/family_hseq.json regenerated (pre-image kept at .run/family_hseq.json.pre-T0): 134 → 138 overlays, newly visible = the 4 P27 SC07 overlays; metrics re-baselined 68.9 → 67.0% instr (now agreeing with the committed docs/progress.fleet.md); LEN across the whole frontier = 0 (a phantom-LEN from a missing image is now structurally impossible).

  • 🔑 T0 FINDING — a large, doubly-hidden target pool. Regenerating exposed 1,255 families / 6,268 members / 230,612 ins whose ONLY unmatched members are in the 4 new SC07 overlays (0 elsewhere — a perfectly clean partition), each behind an already-matched, byte-proven ov_SC01_077 exemplar. Classes PURE 5,575 (89%) / IMM 633 (10%) / STRUCT 60 (1%). It was hidden twice over: P27's disc audit created it by onboarding the overlays but never regenerated the map — and had it, the img_path hardcode would have classified every member LEN and it would have looked like nothing. Prediction, not a bank (h_seq predicts; the gate decides — Phase 26 predicted PURE too and the gate refused). → T3's headline stratum, and a better probe than planned: the exemplar is already byte-proven, so a failure isolates the templating mechanism with no drafting variable.

  • ↺ SELF-CORRECTION (R14). The plan's own population figures (163 families / 13,232 members) came from the stale map — my own numbers were an instance of the defect this phase is about. Honest, from the fixed map: 1,418 matched-exemplar families / 21,889 unmatched members (PURE 17,024 = 78% · IMM 4,473 = 20% · STRUCT 392 = 1.8% — the roadmap's "register-drift" swing class remains ~2% of the input, so that framing is unchanged). The legacy pool is unchanged at 163 families (a clean consistency check: the fix and the 4 overlays are purely additive); it grew 13,232 → 15,621 members because legacy families also gained SC07 members.

  • T3 strata (honest, from the fixed map): SC07-only 1,255 fam / 6,268 mem / 230,612 ins · legacy PURE non-jr 95 / 7,993 / 478,379 · legacy IMM 36 / 6,644 / 212,707 · legacy MIXED 30 / 968 / 10,462 · legacy PURE w/ jr 2 / 16 / 5,088 (T1's jtbl territory — note B2's family is far smaller here than the roadmap's "×112 ≈ 106k ins"; T1 re-derives it). Total addressable = 937,248 ins = 21.7% of all remaining weight = 7.16pp of fleet instr if it all banked. This is the prize the roadmap declared dead. It stays a prediction until T3's gate.