BFM-decomp/phase-ends/CURRENT_PHASE.md

CURRENT_PHASE — Phase 32: the last 21 + the parked 5 (short, kill-gated)

Gate 1 approved 2026-09-05 (Drew, plan mode, Max, Fable 5.1). Rules R44–R63 ratified at gate 1 (binding). The approved plan (full text): ~/.claude/plans/plan-mode-enabled-max-iterative-ember.md — its substance is this file. Jumping-off document: docs/frontier-p32.md. Roadmap: docs/roadmap-to-100.md v2 §3 (P32 as written is consumed; P32 is this census) + PhaseEnd_Phase31.md Roadmap delta. Baseline HEAD at open: commit:3900 (Phase-31 close + post-close docs; no src//config/ change since the S80 census — verified git diff commit:3898..HEAD -- src config Makefile empty).

Milestone (gate 2)

Every one of the 21 census rows is BANKED or carries a FINAL verdict (PROVED wall / candidate wall with a gcc-mechanism citation + a bounded-attempt record / deferred with its cost stated); each of the 5 parked payloads is onboarded byte-identical at a byte-proven base (its stubs then join the census and are banked or ledgered the same way) or carries a documented exclusion with the static + runtime evidence; the fleet is byte-identical (R22) at every banked batch; the PhaseEnd carries the final wall ledger, the corrected denominators (main game-code 41,556 — the 22-ins Ghidra-boundary gap), all three metrics, and the roadmap delta (P33 = verify + flip).

Kill gate: a session that banks nothing and produces no new verdict ends the crack work (T3); the parked-5 work is bounded to the static candidate list + one static-RE session; no forced banks, no redefinition (P9). Scale estimate: 3–5 sessions.

Effort / model (R7/R26/R27 — every transition is PROMPTED, never assumed)

• Max through T0–T2 (T2 is non-obvious static RE; T0/T1 are settled-design and would be fine at xHigh).
• Prompt to drop to xHigh for T3 (orchestration) and T4 (ledgering); prompt back to Max for T5 (PhaseEnd).
• Models: T1c one Sonnet agent (123 ins); T3 Opus for all 11 targets (escalation tier — Sonnet plateaued on the ≤120-ins rows in S79); Fable only if an agent names a wall class the codegen map lacks — STOP and prompt Drew first. No Ultracode / Workflow: one-page census → the S80 one-agent-per-function shape (docs/wave-playbook.md §S80 addendum, Agent-tool subagents; tools/agent_verdicts.py for any that outlive a session).

Adjustments vs frontier-p32.md §4 (approved)

1. The parked 5 (T2) run BEFORE the crack pass (T3): the static header probe (2026-09-05) gave every payload a deterministic candidate base, so T2 is a set of ~0-token new_binary.sh byte-gate probes, and any stubs they add ride the same single bounded agent pass.
2. Tooling-first split into three tasks, one commit each (P4/R42/R60).
3. No MCP dependency for the parked 5 (static evidence + our own matched C); MCP only as a fallback (R29 prompt).
4. Walls get a ledger task (T4), not drafting — incl. fixing two dangling best-draft paths (R62).

Tasks (strict order; one commit per task after this file is updated; count banks from the SOURCE)

• T0 — Phase open (Max) — DONE 2026-09-05: verbatim 5==5 · exclude 8/8 fresh · census 21 (B-CARVE 4 / D-NEAR 10 / F-FAR 7, .run/P32/frontier_p32_open.json) · tools-health OK · check-all 213/213 rc 0 (.run/P32/t0_baseline.log): this file; harness task list (R28); baseline reads — verbatim_check --strict (5==5), exclude_audit --assert-fresh, frontier_classify --json .run/P32/frontier_p32_open.json (21), make tools-health OK, make check-all 213/213 (read of the on-disk S80 artifacts); commit.
• T1a — resident: split the code subseg, bank func_800D128C (243) — DONE 2026-09-05: BANKED byte-identical 8e17e02f… (commits commit:3902 split + commit:3904 bank); R22 after the bank: check-all 213 passed / 0 failed, rc 0 (.run/P32/t1a/r22.log); resident stubs 2 → 1; THREE instrument fixes on the way (cookbook §496 carrier provided-types, §498 carve --pre drop + gate ignoring the extract rc + interleave_check false DRIFT); twin_rescan: 20 open, nothing newly free. Original brief: jr_isolate_all resident --only func_800D128C (dry-run CLEAN 2026-09-05: 3 region files, carve repoint resident → resident_jr_800D00E4, --order + JTBL_PADS) → make extract BINARY=resident && make build BINARY=resident -j8 (EXIT CODE, R53; byte-neutral split) → commit carve state alone (R60) → rtu_match the closeness-0 draft .run/S71_gate14/resident/func_800D128C.c in the new region TU → parallel_gate --r22 (carves jtbl_80113FB8 + jtbl_80114198; pads_audit after) → in-tree build exit code → commit → twin_rescan. TU refusal → §376/§378 chain as a byte-neutral plumbing commit BEFORE re-gating.
• T1b — ov_SC02_017: split + remap the twin, bank func_80186C64 (209) — DONE 2026-09-05: BANKED byte-identical c0253499… (commit:3903 carrier §497 fix — NO source rename needed · commit:3906 split · commit:3907 bank); 0 drafting tokens; four §376 TU spellings + one TU-provided typedef stripped. Original brief (Max/xHigh): byte-neutral tag rename in src/ov_SC02_017/ov_SC02_017_jr_8017DF34.c (:4314/:4319; keep the typedef name) → build sha-unchanged → commit → jr_isolate_all ov_SC02_017 --only func_80186C64 (dry-run REFUSED on the conflict 2026-09-05; expect CLEAN after the rename) → extract/build byte-neutral → commit carve state → family_remap --addr 0x801810C8 --from ov_SC02_016 --to ov_SC02_017 --to-addr 0x80186C64 (d=2 RELOC-ONLY twin; source src/ov_SC02_016/ov_SC02_016_jr_801810C8.c:2839) → strip typedefs the TU already defines (§491 gap 2) → rtu_match → parallel_gate --r22 → commit → twin_rescan.
• T1c — md_MAIN_034: bank func_800CB00C (123) — DONE 2026-09-05: BANKED byte-identical 46153c06… (commit:3908) from the STORED S72 body .run/O21/opus/func_800CB00C.c (rtu MATCH 123/123; the census's best_draft was a name-collided WRONG file, R48/R38) — NO agent; WALL pin dropped (exclude 8→7); backlog 18→17. R22 after the T1b+T1c batch: check-all 213 passed / 0 failed, rc 0 (.run/P32/t1b/r22.log). Original brief: pack via claude_wave_packs (the stored 174-ins draft is WRONG, not a seed; §494 jtbl idioms) → verbatim grep → rtu_match --split md_MAIN_034 --source md_MAIN_034 → parallel_gate (island pads derive at build, §303) → commit → twin_rescan; drop the md_MAIN_034:func_800CB00C pin (exclude_audit --write). Progress report to Drew after T1c.
• T2 — the parked 5 (Max):
  • T2a — DONE 2026-09-05 (commit:3909): controls 7/7 (5 STRONG, 2 CONSISTENT; TEXT_LO estimates == yaml); the five: MAIN/7 STRONG 0x800CEDF8 · MAIN/9 STRONG 0x800CD348 · SC03/53, /54 CONSISTENT top 0x801EF468 (12-way tie) · SC03/56 SPLIT (jal vote 0x80178C8C vs ptrs/lui ~0x801CBB50); memory-map §S45 p7; SETUP row. Brief: tools/payload_base_evidence.py (id word, first prologue → TEXT_LO, abs-pointer window, lui hi-half histogram, jal self-consistency per candidate base, DESTPTR lookup in the window from .run/idxtab_map.json) — R39 controls: must re-derive md_MAIN_008/011 (0x800CEDF8), md_MAIN_042 (0x800CCB1C), md_MAIN_013 (0x800CAE08), md_SC03_073 (0x801EF468) from their payloads alone; refuses a payload with no self-reference (base-independent bytes — say so, never guess). SETUP row (R21); evidence rows → docs/memory-map.md §S45 p7 (G5 static-derived, US). Commit.
  • T2b — DONE 2026-09-05: ALL FIVE onboarded on their first candidate, byte-identical first builds — md_MAIN_007 @0x800CEDF8 (commit:3911, 19 stubs, A4 applied) · md_MAIN_009 @0x800CD348 (commit:3912, 11) · md_SC03_053 + md_SC03_054 @0x801EF468 (commit:3913, 15 + 7) · md_SC03_056 @0x801CBB50 (commit:3914, 4). Fleet 213 → 218; +56 stubs (~2,600 ins). Instrument findings (R34/R35): the all-INCLUDE_ASM first build is a NULL oracle for FINE base errors (+8 builds byte-identical; +0x1000 fails the link) — bases are static-derived STRONG, byte-proven by the first internal-call C bank; a jal-vote base whose targets are shared-engine function starts is OUTWARD-EXPLAINED (SC03/56's false 0x80178C8C). Close chain (sig-modules → tools-health → audit-disc → R22 of 218) running: .run/P32/t2b/close.log. Original brief: onboard probes, one payload per commit, top candidate first (the module-class law: the FIRST build's byte-identity is the arbiter; a miss → next candidate; none → stays parked with the evidence): md_MAIN_007 extracted/retail/MAIN.CD.dir/FILE_007 0x800CEDF8 0x34 (9/9 internal jals on its own prologues at that base; omit symbols.resident.txt, A4 law) · md_MAIN_009 …/FILE_009.dir/0.1 0x800CD348 0x3C (the unique base fitting both in-region jals; then 0x800CAE08, 0x800CCB1C) · md_SC03_053 …/SC03.CD.dir/FILE_053.dir/0.1 0x801EF468 0x4 · md_SC03_054 …/FILE_054.dir/0.1 0x801EF468 0xF0 · md_SC03_056 …/FILE_056.dir/0.1 0x801CBB50 0x4 (then 0x801CBD90, 0x801CC6E0). After each success: make sig-modules, make tools-health (R36), make audit-disc (UNCLAIMED 5 → 5−N), commit.
  • T2c — DONE 2026-09-05: 20 free banks in the five new modules for 0 drafting tokens (commit:3918 8 · commit:3919 5 · commit:3920 7): 7 d=0 hash twins + 3 RELOC-ONLY via family_remap, 9 constant-only NEAR-COUSINs (each constant named by the real-TU diff), 2 hand pass-throughs (main's func_80014128 shape); dedup_extend 0 (no macro-group members); split_indicator population fix (commit:3917, was a stale 213 list — R51). R22 after the batch: check-all 218 passed / 0 failed, rc 0 (.run/P32/t2c/r22.log). Census: 54 stubs / 6,376 ins (.run/P32/frontier_t2c_close.json: main 12 · md_MAIN_007 15 · md_MAIN_009 10 · md_SC03_053 6 · md_SC03_054 4 · md_SC03_056 1 · ov_SC03_105 2 · resident 1 · md_MAIN_003 1 · ov_SC06_022 1 · ov_SC07_002 1). make audit-disc UNCLAIMED 0 of 220; disc-completeness.md P32 section written; cookbook §499 + decision-log S81 (R31). Brief: free banks + census: twin_rescan / dedup_extend for h_exact twins of the new stubs; regenerate the census; leftovers join T3. Update docs/disc-completeness.md (+ ledger regen), decision log (R31).
  • T2d — NOT NEEDED (both MAIN/7 and MAIN/9 onboarded on their first candidate). Brief: (only if MAIN/7 or MAIN/9 refuse every candidate): one bounded static-RE pass over our own C (src/800.c, src/boot.c, src/resident/resident.c: loadDestPtrTable, D_800D3764/D_800D384C, the resourceIdMap sites) for a route producing global index 7 or 9; MCP (R29) only if the C is insufficient. Outcome: a base, or the strengthened exclusion evidence. Progress report after T2.
• T3 — one bounded crack pass, Opus agents, one per function (xHigh orchestration — prompt R27): IN PROGRESS — checkpointed 2026-09-05 09:30 MDT (see the 🛑 block): 31/47 launched, 10 BANKED (1,063 ins, commits commit:3923..commit:3931), 10 more MATCH verified in the real TU awaiting the gate (1,049 ins) + 1 leaf-exact behind a TU decl, 9 NEAR at exact length, 17 Haiku rows still queued; the launching coordinator overflowed its context at 07:07Z. Census 44 stubs / 5,313 ins (.run/P32/frontier_t3_checkpoint.json). Original brief: targets = main:func_80023BF0 11 · func_80038698 11 · func_80039B20 7 · func_80039308 34 · func_80015B6C 40 · func_8002FDE8 35 · func_8001BC6C 28 · ov_SC03_105:func_80185810 37 · md_MAIN_003:func_800CF3E8 54 · ov_SC07_002:func_8017DC80 84 · resident:func_800D06E8 344 (after a byte-neutral fix of the Struct80078E78 typedef clash) + any T2c leftovers. Cheap second looks first: func_80023BF0 (VERIFY the S80 permuter waypoint's semantics before seeding) and func_80039B20. Packs: claude_wave_packs (S79 verdict incl. INERT list, exact-length best draft, matched same-TU neighbour, §494/§495 laws, the fixed permuter recipe permuter_ils … --cycles 8 --secs 150 --j 3); brief = .run/S79w/SYS.md. Per result, never in bulk: verbatim grep → rtu_match (main: --tu src/<TU>.c) → gate (parallel_gate --r22 / main: gate_main <slate> --apply, read the BODY/TABLE/PLUMBING/MIXED line) → commit per bank (R42) → twin_rescan; plateaus → backlog.py log with attempt + cost (R41). Expected 1–4 banks (S79: 3/11 on the Opus F/G tier), ~0.6–1.4M tokens. Progress report after T3.
• T4 — the walls' FINAL verdicts (xHigh; no drafting): the 8 pinned rows — confirm each citation current + instrument exonerated (R40), PROVED vs CANDIDATE, ledger best-draft → an existing real-TU-clean C file (R62: func_80011380 → .run/m3/opus/func_80011380.c; func_801834A4 → .run/S71_gate14/ov_SC03_105*/), exclude_audit --write, backlog.py render; a T3 idiom naming a wall's mechanism re-opens that row (bounded: one attempt, permuter first). Wall table into this file; cookbook §496+; decision log; accelerators.
• T5 — PhaseEnd (Max, Tier 1 — prompt R27): P7 checkbox walk; milestone demo (R22 fleet N/N, tools-health, verbatim_check --strict, final census, wall ledger, parked-5 dispositions via make audit-disc, make report all three metrics + main 143dbb89… with/without SDK dirs, corrected denominators); WAIT for gate 2; PhaseEnd_Phase32.md (+ Roadmap delta + Plain-English Recap); git mv this file → phase-ends/logs/Phase32.md (R19); R23 MCP stop if it ran; leave both uncommitted for Drew (R6); 🛑 (P8).

Standing procedure (every task)

Fix the TU by a byte-neutral plumbing commit BEFORE any gate (S77 law) · commit banked work before the next command that can touch src/ (R42) · after any bank: twin_rescan, verbatim_check --strict (R62) · after a carve/merge touching a yaml/.mk: make extract BINARY=<b> before building; pads_audit/interleave_check on every touched carve binary (R60) · fleet verify after every banked batch: make clean && make extract-all && make check-all — read EXIT CODES (R53), quote the green count at every close (R58); make tools-health · log each task here + refresh the 🛑 checkpoint before any pause · rules check every 4 tasks (P6) · harvest idioms into the cookbook before the next drafting step (R16/R30) · no Co-Authored-By trailer (R5); keep the Claude-Session: line.

Blockers

• 2026-09-05 09:xx: the successor session's SessionStart hook launched a headless Ghidra MCP server (.run/ghidra-mcp.log); T3/T4 do not use it; no RE writes → never stage the ghidra/ churn (R23).
• Ghidra MCP not running at open (clean stop in the log; the SessionStart hook's server did not persist). Not needed for T0–T2b; if T2d needs it: tools/ghidra_mcp_start.sh → pause → Drew runs /mcp (R29) → G2 ping.

Log

• 2026-09-05 (successor session, later) — Session-start protocol rewritten at Drew's direction (R64 candidate): phase-ends/DIGEST.md created (every phase synopsis + every rule in full + the PROJECT_CONTEXT corrections + the doc map); CLAUDE.md load order = PROJECT_CONTEXT → DIGEST → the three most recent PhaseEnds → CURRENT_PHASE, ≈100k tokens, the 🛑 block replayed VERBATIM; Phase Boundary step 3b maintains the digest (P7); phase-ends/README.md + SETUP §7 updated; the two memories (checkpoint-current-phase-before-pause, session-start-list-rules-in-full) + MEMORY.md updated; .run/P32/t3/PROMPT_TEMPLATE.md saved (verbatim agent prompts). The 🛑 block below was REFRESHED to the verbatim-replay standard and SUPERSEDES the 09:30 block. Ratify R64 at PhaseEnd_Phase32.
• 2026-09-05 09:30 MDT — T3 CHECKPOINT by the successor session (Max, Fable 5.1) after the launching session died. The T3 coordinator (session c6a5fb71) banked 10 functions in 9 commits (T3 (1)–(9), 01:01–01:07 MDT) and then hit "Prompt is too long" at 01:07 MDT; 22 agent completions (07:07Z–08:04Z) arrived into a dead session. Recovered via the recorded route: agent_verdicts.py over the 31 task transcripts → .run/P32/t3/verdicts.jsonl (20 MATCH / 9 NEAR / 2 FAIL); every unbanked draft re-verified with rtu_match in its real TU (10 MATCH confirmed). Instrument incident (R48 class): one agent's tidy-up (find .run/P32/t3/opus -maxdepth 1 -type f ! -name func_800CD674.c -exec mv {} _scratch/, 07:52Z) had swept 11 sibling Opus deliverables — found in _scratch/, moved back, byte-verified at the reported closeness; tools/agent_drafts_restore.py (NEW — transcript replay) written as the fallback; .gitignore allowlist for .run/P32/** so the drafts/ledger/census are committed (R20). Harvest written IN-SESSION (R16/R30): cookbook §500 (10 banked closers, 10 MATCH closers, 9 NEAR classes, two NEW mechanisms — the pinned-base alias basin and #line-equalised ASM_OPERANDS — and the process defects); docs/wave-playbook.md §S80 addendum-2 (per-function work dirs, JSON-only final message, the 20-agent cap, the recovery tools); docs/accelerators.md P32 T3; docs/decision-log.md P32 S82 (R31); SETUP row (R21). Read-only T4 pre-check results from the dead session preserved in the 🛑 block. Effort: this checkpoint at Max; T3 resumes at xHigh (R27).
• 2026-09-05 — gate 1 approved; T0 opened at HEAD commit:3900.
• 2026-09-05 — T3 LAUNCHED (xHigh): 47 targets = census 54 − 7 pinned walls (.run/P32/t3/targets.json, 5,288 ins); packs built (claude_wave_packs, 47/47 neighbours, 10 journals); brief .run/P32/t3/BRIEF.md. 20 Agent-tool drafters running (16 Opus: the 11 old near/far rows + resident func_800D06E8 + the 4 largest new bodies · 3 Sonnet · 1 Haiku); 27 Haiku queued behind the harness's 20-concurrent cap (.run/P32/t3/pending_launch.txt, launched one per freed slot). Recovery + per-result route: .run/P32/t3/LAUNCHED.md.
• 2026-09-05 — T2c DONE (R22 218/218 rc 0) · T2d not needed. Twenty of the 56 new stubs fell to twin remaps and constant flips (0 tokens); the rest (36, all ≤50 ins but one, never drafted) join T3 on the model ladder. Session limit reset here (Drew); effort now xHigh via /model (the T3 orchestration level, R27 satisfied). Settings-file JSON defect fixed (a trailing comma).
• 2026-09-05 — T2b DONE — the parked ledger is EMPTY: all five payloads onboarded byte-identical (MAIN/7 → md_MAIN_007 @0x800CEDF8; MAIN/9 → md_MAIN_009 @0x800CD348; SC03/53, /54 → md_SC03_053/054 @0x801EF468; SC03/56 → md_SC03_056 @0x801CBB50). Every base came from T2a's static alignment (self-calls / fn-ptr-table entries on the module's own function starts; controls 7/7), NOT from the byte check — two negative controls showed the first build is base-lenient (fine offsets) and only catches gross errors. Two tool refinements en route (combined STRONG rule; outward-explained vote bases). The 100% claim loses its last asterisk once the new stubs are banked or ledgered (T2c/T3).
• 2026-09-05 — T1b + T1c DONE (R22 213/213 rc 0 for the batch) · T2a DONE. Both banks were ZERO-token: T1b's body came from family_remap of the d=2 twin (+ four TU spellings), T1c's from a stored S72 body the census had mis-keyed (R48) — the journals (R38) named both. Census 21 → 18 stubs (resident 1 · main 12 · md_MAIN_003 1 · ov_SC03_105 2 · ov_SC06_022 1 · ov_SC07_002 1). T2a's instrument gives every parked payload a bounded probe list; two are STRONG.
• 2026-09-05 — T1a DONE — resident:func_800D128C BANKED (243 ins, byte-identical 8e17e02f, R22 213/213). The stored S71 closeness-0 draft was byte-correct all along; the whole task was three instrument defects the resident (the fleet's one common.h-only, --pre-sandwich binary) exposed in overlay-only assumptions: (1) jr_isolate_all dropped a file-local typedef whose name engine_types.h also defines (§496 — fixed: provided types derived from the TU's own includes); (2) jtbl_carve regenerated JTBL_INTERLEAVE without the --pre hdr.rodata.o clause → extract refused → the gate linked a stale script and booked the byte-correct draft as DIFF; harvest_verify ignored that extract's rc (§498 — both fixed, R49/R61); (3) interleave_check read a --pre line as n=0 (false DRIFT; fixed). R38 then found two more stored MATCH bodies for T1b/T1c (see their rows). Effort stayed Max.
• 2026-09-05 — T0 DONE. Baseline reads all green (.run/P32/t0_baseline.log): verbatim_check --strict 5 bodies == 5 rows; exclude_audit --assert-fresh 8 entries, 8 WALL, 0 stale; frontier_classify → 21 rows = the S80 census exactly; make tools-health OK (sigs fresh, corpus(+resident), cdecl, audit-binaries 213/213, report lint+dedup, cookbook-index, split_indicator 213 OK); make check-all 213 passed / 0 failed, rc 0. Harness task list #1–#11 built (R28). NEXT = T1a.

🛑 SESSION CHECKPOINT — T3 MID-WAVE, REFRESHED (2026-09-05 10:01 MDT; written by the successor session e40106e1; SUPERSEDES the 09:30 block)

0. How to use this block

You are a FRESH SESSION that has read PROJECT_CONTEXT.md, phase-ends/DIGEST.md, PhaseEnd_Phase29/30/31.md and this file, and nothing else (CLAUDE.md protocol, R64 candidate). Replay this block verbatim into your chat, state phase / done / NEXT / effort, list the rules from the digest, then WAIT for Drew. When he confirms, follow §5 in order. Effort: xHigh for T3 and T4 (R27 — Drew sets it; if the status line does not say xHigh, prompt before the first command); Max only for T5. Drafter ladder Haiku ≤50 ins / Sonnet ≤120 / Opus >120; Fable only if an agent names a wall class the codegen map lacks — STOP and prompt Drew first. No Workflow / Ultracode in T3: one Agent-tool subagent per function.

1. Where we are

Phase 32 — the last 21 + the parked 5 (short, kill-gated). Gate 1 approved 2026-09-05 (Drew, plan mode, Max, Fable 5.1); rules R44–R63 ratified then; R64 is a candidate (DIGEST §3). Tasks: T0 ✓ commit:3901 · T1a ✓ commit:3902 commit:3904 · T1b ✓ commit:3903 commit:3906 commit:3907 · T1c ✓ commit:3908 · T2a ✓ commit:3909 · T2b ✓ commit:3911 commit:3912 commit:3913 commit:3914 (fleet 213 → 218) · T2c ✓ commit:3917 commit:3918 commit:3919 commit:3920, close commit:3921 · T2d not needed · T3 IN PROGRESS (launch commit:3922; banks commit:3923 commit:3924 commit:3925 commit:3926 commit:3927 commit:3928 commit:3929 commit:3930 commit:3931; the successor's checkpoint commit:3932 + the protocol commit = HEAD) · T4, T5 pending. Harness task list #1–#11 exists (#9 = T3 in_progress). Fleet 218 binaries (main + resident + 138 ov_* + 78 md_*). Last full R22: 218/218 rc 0 at the T2c close (.run/P32/t2c/r22.log); the 10 T3 banks each built byte-identical in their own binary but no fleet R22 has run since — §5 step 0 runs one before any new bank. Non-ghidra tree clean at HEAD; the ghidra/ churn (db.*.gbf, ~index.*, *.prp) is R23 restart-noise — never stage it. Unpushed commits at this writing: 32 (Drew pushes, R6).

2. What happened in the previous session (c6a5fb71, 2026-09-04 22:39 → 2026-09-05 01:07 MDT)

T0–T2c as logged above. T3: census 54 − 7 pinned walls = 47 targets (.run/P32/t3/targets.json); packs built (claude_wave_packs, 47/47 neighbours, 10 journal notes); brief .run/P32/t3/BRIEF.md, laws .run/P32/t3/SYS.md. 20 drafters launched 00:53–00:55 MDT (16 Opus: the 11 old near/far rows + resident func_800D06E8 + the 4 largest new bodies; 3 Sonnet; 1 Haiku), 27 Haiku rows queued behind the harness's 20-concurrent cap; the launch was checkpointed (commit:3922, 00:59). Ten queued rows were launched as slots freed (01:00–01:07). Ten functions banked in nine commits 01:01–01:07 (list in §4). At 01:07:15 MDT the coordinator hit "Prompt is too long" (its context was ~95% full from the whole day BEFORE the wave); the other 22 completions (01:07–02:04 MDT) arrived as notifications into a dead session. At 01:52 MDT one agent (func_800CD674's) tidied the SHARED .run/P32/t3/opus/ with find .run/P32/t3/opus -maxdepth 1 -type f ! -name 'func_800CD674.c' -exec mv {} .run/P32/t3/opus/_scratch/ \;, moving 11 sibling deliverables (the three agents that finished later wrote their finals afterwards and survived). Drew's last message (09:04): "97% context. checkpoint for fresh session, have next session analyze the completed subagents results".

3. What THIS session did (e40106e1, 2026-09-05 09:12 → 10:01 MDT, Max, Fable 5.1)

(a) Dumped the dead session's transcript (~/.claude/projects/-home-musashi-bfm-decomp/c6a5fb71-….jsonl → condensed text with tools/transcript_dump.py, NEW) and saved all 31 agents' full final reports to .run/P32/t3/reports/<fn>__<arm>__<id>.md (tools/agent_reports.py, NEW; tracked) — the levers, inert-lever lists, residual mechanisms and plumbing warnings behind each verdict; read a row's report before banking or routing it. (b) Rebuilt the verdict ledger by the recorded route: .venv/bin/python tools/agent_verdicts.py /tmp/claude-1000/-home-musashi-bfm-decomp/ 7beebba8-05e9-4d18-98ea-ce208f45370f/tasks/a*.output --append .run/P32/t3/verdicts.jsonl → 31 rows: 20 MATCH / 9 NEAR / 2 FAIL (pass ONLY a*.output — the b*.output files there are older sessions' tasks). (c) Found the swept deliverables in _scratch/, moved the 9 missing ones back to .run/P32/t3/opus/<fn>.c, and re-verified EVERY unbanked draft with rtu_match in its real TU — every result equals the agent's verdict (§4). (d) Wrote tools/agent_drafts_restore.py (transcript replay of Write/Edit/cp/mv/heredoc ops; rebuilt 26/30 exactly — the 4 an agent had edited via shell after its last Write were stale, so an on-disk copy always wins). (e) Harvest (R16/R30): cookbook §500 (10 banked closers, 10 MATCH closers, the 9 NEAR classes with their inert-lever lists, two NEW mechanisms — the pinned-base-vs-pseudo-address alias basin, #line-equalised ASM_OPERANDS for cross_jump — and the process defects), docs/wave-playbook.md §S80 addendum-2, docs/accelerators.md P32 T3, docs/decision-log.md P32 S82 (R31), docs/SETUP.md tool row (R21), docs/cookbook-index.md regenerated (tools-health's --check is green), .run/P32/t3/BRIEF.md output contract amended (final message = the JSON line ONLY; scratch only under .run/P32/t3/work/<fn>/; never find/rm/mv in a shared dir). (f) .gitignore allowlist for .run/P32/** (drafts, ledger, census jsons, BRIEF/SYS/LAUNCHED/targets/pending/ harvest_notes/PROMPT_TEMPLATE, restored/sweep_*.py) → committed commit:3932. (g) Drew-directed protocol change: phase-ends/DIGEST.md created (every phase synopsis, every rule in full, the PROJECT_CONTEXT corrections, the doc map); CLAUDE.md session-start rewritten (≈100k-token budget: PROJECT_CONTEXT → DIGEST → last 3 PhaseEnds → CURRENT_PHASE; rules in full from the digest; the 🛑 block replayed VERBATIM); Phase Boundary step 3b (append the phase to the digest, P7); phase-ends/README.md; SETUP §7 pointer; the memories checkpoint-current-phase-before-pause and session-start-list-rules-in-full + MEMORY.md; R64 candidate — ratify at PhaseEnd_Phase32. (h) .run/P32/t3/PROMPT_TEMPLATE.md — the verbatim agent prompts (Haiku plain / Haiku with a twin hint / the Sonnet escalation / Opus / Sonnet). (i) Rebuilt the harness task list; wrote this block. Nothing under src/ or config/ changed this session.

4. Census — 44 stubs / 5,313 ins (.run/P32/frontier_t3_checkpoint.json), every row with its state and draft path

binary	stubs · ins	rows
main	9 · 1,643	MATCH-unbanked: func_80015B6C 120 (.run/P32/t3/opus/func_80015B6C.c, TU src/800.c) · func_8002FDE8 73 (.run/P32/t3/opus/func_8002FDE8.c, TU src/800_b_2.c) · NEAR: func_80039308 518 → 17 (…/opus/func_80039308.c, TU src/800_c.c) · func_8001BC6C 69 → 6 (…/opus/func_8001BC6C.c, TU src/800.c) · WALLS (pinned in config/wave_exclude.txt, T4): func_80011380 192 (boot −O0; .run/m3/opus/func_80011380.c 6-off) · func_80032A74 (800_b_2) · func_80020DA4 100 (800; .run/S79w/sonnet/func_80020DA4.c 2-off) · func_80039DEC (800_c) · func_800391D4 (800_c)
md_MAIN_003	1 · 469	NEAR func_800CF3E8 469 → 27 (…/opus/func_800CF3E8.c)
md_MAIN_007	14 · 701	MATCH-unbanked (Haiku, .run/P32/t3/haiku/<fn>.c): func_800CF148 33 · func_800CF2BC 32 · func_800CEEFC 25 · func_800CEF94 25 · func_800CF068 21 · leaf-exact behind the TU decl: func_800CF3B0 22 (src/md_MAIN_007/md_MAIN_007.c:91 = extern void func_800CF3B0(void *a0, void *a1, void *a2);, the BANKED func_800CF390 at :94 calls it with 3 args; the real signature is s32 (void)) · NEAR: func_800CF6D0 249 → 137 · func_800CF408 178 → 49 (…/opus/) · QUEUED Haiku: func_800CF0F4 21 · func_800CF1CC 21 · func_800CF268 21 · func_800CF33C 21 · func_800CF220 19 (twin ov_SC03_001:0x80182860 d=4) · func_800CEFF8 13 (twin ov_SC02_005:0x801865d0 d=3)
md_MAIN_009	10 · 603	NEAR: func_800CD92C 247 → 15 · func_800CD674 174 → 2 (…/opus/) · QUEUED Haiku: func_800CD464 25 · func_800CD558 24 · func_800CD408 23 · func_800CD5B8 23 · func_800CD4C8 22 · func_800CD520 22 (twin ov_SC03_011:0x8017dbd8 d=3) · func_800CD614 22 · func_800CD3B8 21
md_SC03_053	2 · 116	MATCH-unbanked (Sonnet): func_801EF734 44 (.run/P32/t3/sonnet/func_801EF734.c) · func_801EF7E4 72 (…/sonnet/func_801EF7E4.c) → the module reaches 100% C
md_SC03_054	3 · 698	MATCH-unbanked: func_801EF6D8 — targets.json nins 663 = 604 code ins + six jump tables (B-CARVE; …/opus/func_801EF6D8.c) · QUEUED Haiku: func_801F0048 20 · func_801F00C8 15 (twin md_SC03_053:0x801ef9dc d=5 — family_remap refused, reloc-count 1≠2, hand-adapt)
md_SC03_056	1 · 23	QUEUED Haiku func_801CBB80 23 (twin ov_SC06_025:0x8017ded8 d=5)
ov_SC03_105	2 · 595	NEAR func_80185810 489 → 35 (…/opus/func_80185810.c) · WALL (pinned) func_801834A4 106 (.run/S71_gate14/ov_SC03_105*/func_801834A4.c 6-off)
ov_SC06_022	1 · 119	WALL (pinned) func_8017DF28 119 (.run/S79w/sonnet/func_8017DF28.c 2-off)
ov_SC07_002	1 · 346	NEAR func_8017DC80 346 → 46 (…/opus/func_8017DC80.c)

FAIL rows: haiku func_801EF734 (47 vs 44 — superseded by the Sonnet MATCH); haiku func_800CF3B0 (leaf-exact 22/22; the TU decl). The 10 BANKED in T3 so far (1,063 ins): md_SC03_054 func_801EF558 96 (commit:3923, sha 06bd73df…) · md_SC03_053 func_801EF49C 49 (commit:3924), func_801EF6B0 33 + func_801EF95C 20 (commit:3927), func_801EF624 35 (commit:3928) (all c0848f30…) · resident func_800D06E8 344 (commit:3925, 8e17e02f… — the resident is 145/145 C; its jtbl_80113FA4 carved by jtbl_carve --func) · main func_80039B20 79 (commit:3926), func_80023BF0 281 (commit:3930), func_80038698 74 (commit:3931) (all 143dbb89… via gate_main) · md_MAIN_007 func_800CEE2C 52 (commit:3929, 2ff702b6…).

5. NEXT — the T3 resume order (each step's evidence quoted; R42: commit before the next command that can touch src/)

0. Preflight: git status --short | grep -v ghidra/ (empty) · .venv/bin/python tools/verbatim_check.py --strict (5 == 5) · make check-all → 218 passed, 0 failed of 218, exit 0 (R22 baseline before any bank; if any binary is red, fix the baseline first — R56) · read cookbook §500-B for each MATCH's plumbing note.
1. md_SC03_053 ×2 (Sonnet MATCH): for func_801EF734 then func_801EF7E4: .venv/bin/python tools/rtu_match.py <fn> --split md_SC03_053 --source md_SC03_053 --c .run/P32/t3/sonnet/<fn>.c --asm-subdir asm/md_SC03_053/nonmatchings/md_SC03_053 --work .run/P32/t3/verify/<fn> (expect MATCH (N ins)) → splice the draft over the one INCLUDE_ASM("asm/md_SC03_053/ nonmatchings/md_SC03_053", <fn>); line (python: assert the line occurs once, replace with the draft text) → make build BINARY=md_SC03_053 -j8 (read ${PIPESTATUS[0]}, R53) → sha1sum build/md_SC03_053/md_SC03_053 == config/check.md_SC03_053.sha (c0848f30…) → git add src/md_SC03_053/md_SC03_053.c + commit ("T3 (10) …") → md_SC03_053 is 100% C.
2. main ×2 (Opus MATCH): write .run/P32/t3/gate/slate_main2.json = [{"fn":"func_80015B6C","draft":"/home/musashi/ bfm-decomp/.run/P32/t3/opus/func_80015B6C.c"},{"fn":"func_8002FDE8","draft":"/home/musashi/bfm-decomp/.run/P32/t3/opus/ func_8002FDE8.c"}] (ABSOLUTE paths) → .venv/bin/python tools/gate_main.py .run/P32/t3/gate/slate_main2.json --apply (a clean main rebuild; read the "slate 2 -> N compatible" and the "BANKED n main functions -- 143dbb89… BYTE-IDENTICAL" lines; on a drop, read the BODY/TABLE/PLUMBING/MIXED verdict) → git add src/800.c src/800_b_2.c + commit. Keep func_8002FDE8's block-scope extern s16 D_800A46D2; — the scalar spelling IS the lever (§500-B).
3. md_SC03_054 func_801EF6D8 (604 + six jtbls): .venv/bin/python tools/jtbl_carve.py md_SC03_054 --func func_801EF6D8 (as T3 (3) did for the resident) → make extract BINARY=md_SC03_054 (EXIT CODE — a failed extract leaves a stale linker script, §498) → splice → make build BINARY=md_SC03_054 -j8 → sha 06bd73df… → .venv/bin/python tools/pads_audit.py md_SC03_054 + tools/interleave_check.py md_SC03_054 (R60) → commit the carve state (ONLY this binary's config/overlays.mk lines + config/splat.md_SC03_054.yaml) + the bank. Alternative: tools/parallel_gate.py … --r22 (it carves). The draft's rtu_match MATCH was re-confirmed against the CURRENT TU after func_801EF558 landed.
4. md_MAIN_007 (six): FIRST the byte-neutral plumbing commit (the S77 law): src/md_MAIN_007/md_MAIN_007.c:91 extern void func_800CF3B0(void *a0, void *a1, void *a2); → extern s32 func_800CF3B0(); (no-proto, the §376 fix_arity_callers class — keeps func_800CF390's 3-arg call legal and admits the s32 (void) definition) → make build BINARY=md_MAIN_007 -j8 → sha 2ff702b6… UNCHANGED → commit. (If the sha moves, keep the 3-param decl with s32 return and define func_800CF3B0 with three unused pointer params instead — the gate decides.) Then rtu_match each of func_800CEEFC, func_800CEF94, func_800CF068, func_800CF148 (drop its block-scope s32 func_800CF3B0(void) redeclaration only if cc1 objects), func_800CF2BC, func_800CF3B0 → splice → one build → sha → commit (per bank, or per verified same-TU batch).
5. After 1–4: .venv/bin/python tools/twin_rescan.py (a bank changes the twin graph — the queued md_MAIN_009 rows may gain twins; it prints ready family_remap command lines) · verbatim_check --strict · R22 make clean && make extract-all && make check-all → quote 218 passed / 0 failed, exit 0 (R58) · make tools-health → OK · refresh this block + commit. Expected census after 1–4: 33 stubs / ~4,250 ins.
6. Ledger the 9 NEAR in the backlog (grep -n add_argument tools/backlog.py first — the dead session's query on keys addr/fn returned nothing for the wall rows, so confirm the row schema; then tools/backlog.py log … per row with closeness · class · best draft · cost "1 Opus agent, 30–70 min" (R41) · tools/backlog.py render): func_800CD674 2 REGALLOC-PERM $a3↔$t1 · func_8001BC6C 6 REGALLOC-PERM $v0↔$v1 (qty_compare one span unit; pinned seed → §494 permuter_ils … --cycles 8 --secs 150 --j 3) · func_800CD92C 15 §S7 prologue weave · func_80039308 17 sched2 + cross-block regalloc · func_800CF3E8 27 pinned-base alias basin (§500-D1) · func_80185810 35 [permuter] 4 emission windows · func_8017DC80 46 reload-slot frame + la $a0 slot · func_800CF408 49 [permuter] 3 hunks · func_800CF6D0 137 sched1 rank_for_schedule class tie.
7. Launch the 17 queued Haiku rows from .run/P32/t3/pending_launch.txt with the prompts in .run/P32/t3/PROMPT_TEMPLATE.md (substitute fn / binary / nins / sub / asm / tu / twin; keep the arm; the packs are at .run/P32/t3/packs/<fn>.md — regenerate with claude_wave_packs if missing): dequeue the row (sed -i '/^<binary> <fn> /d' pending_launch.txt) BEFORE launching; ≤20 concurrent ("Concurrent subagent limit reached. … Do not retry" = wait for a completion); each agent's final message is now the JSON line only (BRIEF amended). Per result, never in bulk: grep -c '\.ent\|\.word' <draft> (0) → your OWN rtu_match → splice/build/sha → commit → twin_rescan. A Haiku FAIL whose note names a C-structure residual at near-exact length → escalate to Sonnet with the shape hint (the PROMPT_TEMPLATE has the escalation prompt); a TU conflicting types → the §376 plumbing commit first.
8. The NEAR tail (bounded): permuter on the two REGALLOC-PERM seeds (func_800CD674, func_8001BC6C) via the repaired permuter_ils; ONE Opus second look only where §500 names an untested lever (func_800CF3E8: an unpinned alias of p6 for the tag load alone). Plateaus stay in the backlog. Kill gate stands: a session with no bank and no new verdict ends T3.
9. T3 close: any new idiom from steps 7–8 → cookbook §500 addendum (R16/R30) + harvest_notes.md; regenerate docs/progress.fleet.md (make report); then T4.

6. Files, tools, exact invocations, gotchas

• .run/P32/t3/ (git-tracked unless noted): reports/<fn>__<arm>__<id>.md (the 31 agents' full final reports — start here for any row) · BRIEF.md (agent brief, amended) · SYS.md (the laws) · targets.json (47 rows: name, binary, nins, sub, asm, tu, klass, closeness, model, twin, twin_d) · pending_launch.txt (17 rows binary fn nins [twin=…]) · verdicts.jsonl (31) · LAUNCHED.md (recovery route + status) · harvest_notes.md · PROMPT_TEMPLATE.md · {opus,sonnet,haiku}/func_*.c (the deliverables: 12 + 4 + 11) · restored/sweep_func_80039308.py (the dead-local sweeper an agent wrote — a tool candidate) · untracked: packs/<fn>.md (47; regenerable), verify/ + verify_*.txt + gate/slate_*.json (the coordinator's own checks), work/ rtu/ (agent scratch), opus/_scratch/ opus/x/ opus/var/ opus/_h/ opus/_wk_8017DC80/ opus/sweep/ (agent leftovers — ignorable), restored/*.c (transcript replays — inferior to the on-disk finals). Census jsons: .run/P32/frontier_*.json (tracked).
• Transcripts: ~/.claude/projects/-home-musashi-bfm-decomp/c6a5fb71-f2ae-4846-a4a9-a43a4c1c3496/subagents/agent-<id>.jsonl (31; /tmp/claude-1000/-home-musashi-bfm-decomp/7beebba8-05e9-4d18-98ea-ce208f45370f/tasks/<id>.output are symlinks to them; the id→function map is the file field of each verdicts.jsonl row). Never cat one; use agent_verdicts.py (JSON verdicts) / agent_reports.py (full final reports) / agent_drafts_restore.py (rebuild a missing draft) / transcript_dump.py (a session's own transcript).
• rtu_match: --split <sub> --source <binary> --c <draft> --asm-subdir <asmdir> --work <dir>; main adds --tu src/<sub>.c; <sub>/<asmdir>/<tu> come from targets.json or corpus.stubs(<binary>) (asm_path, path). The verdict line starts MATCH/DIFF/CC1. It compiles the WHOLE TU with the draft spliced — the gate predictor.
• Build check: make build BINARY=<b> -j8 2>&1 | grep -v "warning\|In function" | tail -2; rc=${PIPESTATUS[0]} then sha1sum build/<b>/<b> vs config/check.<b>.sha. R53: exit code, never the file (a failed build leaves the old binary).
• gate_main: the ONLY way to gate main (incremental builds false-pass/false-diff on main, R43/§S52); slate = JSON list of {fn, draft(ABSOLUTE)}; --apply keeps the splice on success; then git add the TU + commit.
• Carves: jtbl_carve.py <b> --func <fn> → extract (rc) → build; pads_audit.py <b>, interleave_check.py <b>; a DIFF that starts at the first code byte is a stale linker script, never the draft (§498).
• Gotchas that bit this week: set -o pipefail + grep -c exits 1 on zero matches (a "no ghidra staged" guard aborted a good commit) · never cd inside a compound command (the cwd persists into later calls) · masked_diff.py leaves src/.masked_diff_probe.<pid>.c while running — they are other agents' live probes, never delete them · the .run/P32/t3/<arm>/ dirs are SHARED (agents are now told not to touch siblings) · the SessionStart hook launches a headless Ghidra MCP server (.run/ghidra-mcp.log) — T3/T4 never use it · agent_verdicts.py over a whole tasks/ dir picks up older sessions' outputs — pass a*.output.

7. Carried context for T4 and T5

T4 (xHigh, no drafting) — the dead session's read-only pre-check (07:00Z): exclude_audit --assert-fresh: 7 entries, 7 WALL, 0 stale. Best drafts re-run in their REAL TU: main:func_80011380 .run/m3/opus/func_80011380.c DIFF 6 (the ledger's .run/backlog_drafts/func_80011380.c is MISSING — R62 path fix) · main:func_80032A74 .run/S79w/opus/ func_80032A74.c CC1 FAIL (redefinition of struct Slot/Slot, conflicting types for A12/B12 vs src/800_shared.h:68/95 — §376 plumbing, re-probe before a final verdict, R40) · main:func_80020DA4 .run/S79w/sonnet/ func_80020DA4.c DIFF 2 · main:func_80039DEC both .run/S79w/{permuter,sonnet}/func_80039DEC.c CC1 FAIL (conflicting types for func_80039DEC — a def-side declaration wall, §495: fix the TU's own decl first) · main:func_800391D4 .run/S79w/sonnet/func_800391D4.c CC1 FAIL (conflicting types for D_80073140 at t.c:3310 vs :2622) · ov_SC06_022:func_8017DF28 DIFF 2 · ov_SC03_105:func_801834A4 S71_gate14 variants DIFF 6 (the ledger's .run/backlog_drafts/func_801834A4.c is MISSING). Deliverables: each of the 7 rows PROVED vs CANDIDATE with a current citation + an instrument-exonerated bounded attempt (R40), exclude_audit --write, backlog.py render, the wall table into this file, cookbook §501 if a verdict changes, decision-log, accelerators. A T3 idiom naming a wall's mechanism re-opens that row (one bounded attempt, permuter first). T5 (Max, Tier 1 — prompt R27; WAIT for gate 2): P7 checkbox walk over every task above; milestone demo = R22 fleet 218/218 · make tools-health OK · verbatim_check --strict 5==5 · the final census (tools/frontier_classify.py --json) · the wall ledger · parked-5 dispositions via make audit-disc (UNCLAIMED 0 of 220 at T2c) · make report all three metrics + main 143dbb89… WITH and WITHOUT the SDK object dirs (.run/obj40, .run/obj42; fresh-extract fallback) · corrected denominators (main game-code 41,556 — the 22-ins Ghidra-boundary gap; the fleet denominator grew by the five modules' ~2,600 ins at T2b). Then PhaseEnd_Phase32.md (+ Roadmap delta: P33 = verify + flip; + Plain-English Recap; + the rules table incl. R64) → append the P32 synopsis + rules to phase-ends/DIGEST.md (step 3b) → git mv CURRENT_PHASE.md phase-ends/logs/Phase32.md (R19) → R23 MCP stop if it ran → leave both uncommitted for Drew (R6) → 🛑.

8. Environment

Fleet 218; shas of the binaries T3 touches: main 143dbb89f34491258bbc27810d0a12ec8b43a8dd · resident 8e17e02ff8954d07c979449198f7e1645046b353 · md_SC03_053 c0848f308342e499e2bd8c2db86f6225a29e1592 · md_SC03_054 06bd73dfbdb10748adcbac7093ead5d76131206a · md_MAIN_007 2ff702b605ab5cfc18474c464c4c07e5f8ffd48c · md_MAIN_009 d270f695… · md_SC03_056 bc768a6b… (config/check.<b>.sha is authoritative). Toolchain unchanged (SETUP §5.4). Harness: 20 concurrent Agent-tool subagents max; an Agent call beyond that returns "Concurrent subagent limit reached" (do not retry — wait). Claude commits, Drew pushes. Last published metrics (P31 close, 213 binaries): instr 100.0% · distinct 99.9% · fn 99.99% — regenerate with make report at the T3 close (218 binaries).

9. Standing reminders (the rules that bite in T3)

Every bank commits before the next command that can touch src/ (R42) · count banks from the SOURCE and quote the fleet's green count at every close (R58) · verify a build from its exit code (R53) · a masked score is not a closeness until its diff is read (R63) · an __asm__ body is a verbatim, not a bank (R62) · fix the TU by a byte-neutral plumbing commit BEFORE any gate (S77 law) · after a carve/merge touching a yaml/.mk: make extract before building, then pads_audit/interleave_check (R60) · rules check every 4 tasks (P6) · harvest idioms into the cookbook before the next drafting step (R16/R30) · no Co-Authored-By (R5) · never stage ghidra/ (R23).

Plain-English recap (R18). Last night thirty-one AI agents were set on the leftover functions; twenty produced exact matches, but the coordinator ran out of memory after banking the first ten and one agent accidentally swept the others' finished files into a side folder. This session read the whole record, pulled every result back out, confirmed each one byte-for-byte, put the swept files back, wrote the lessons into the cookbook and the playbook, and — at Drew's direction — made session start cheaper (a one-page digest of all 32 phases instead of reading them all) and made this checkpoint the complete hand-off the next session replays verbatim. The next session's first jobs: re-verify the fleet, bank the ten ready matches, then launch the seventeen queued small functions.