BFM-decomp/phase-ends/current/logs/T3.md

T3 log — rung S sweep, overlays (P, X)

Timeline

• start: context loaded (card slice, plan sections, T3 entry, T1/T2 summaries).
• restruct --status (.run/logs/t3_status.log): 173,459 rows, rung S 0.
• restruct --plan --rung S (.run/logs/t3_planS.log): 50 of 2157 drawable files · 3929 bodies · 5642 typed bases · 35087 sites; skipped bases type-not-canonical 87,074; plan includes md_* → sweep needs an ov_* filter.
• retriever-code (inline, no report): rung S mechanics. Unit = file; plan_S :3119 from CENSUS/body_base_type.json; S1 MEMBERS / S2 greedy_kept :753 / KEPT-ALL / REFUSED; site verdict KEPT(<body hint>) (one hint per body, :1117-1122); kept sites are left raw (no respell); no-field skips never add fields; --apply allow_local=False; no code writes macro spellings.
• restruct_cycle.sh read: per batch calibrate-if-stale → apply -j 12 → R22 → census → commit; S body ≈ 2 s.
• T3.c1 briefed (opus55): per-site class, KEPT() / RESIDUAL(), CAST_* respell of kept P sites, --only glob, calibrate, exemplar ov_SC05_010_jr_80180F84.c.
• T3.c1 done (3ff15ee462, 3507628e40, fe2e814776): selftest 85/85, --real 90/90, exemplar apply 1/1 IDENTICAL 25.5 s, R22 218/218, check-casts macros CAST_ALIAS=3 CAST_SIGN=10 CAST_WIDTH=6 unbacked 0. Deviation (coder): BACKED_S_VERDICTS += MEMBERS, S+A; new body verdict KEPT. Exemplar +0x34 store = RESIDUAL(TYPE-NOT-CANONICAL) in real apply (KEPT(SCHED-ALIAS) only under try-file local types).
• cycle batch 1 (LABEL_PREFIX=t3_ TASK=T3 restruct_cycle.sh 1 1 50 S "ov_*", log .run/P39/restruct/cycle_t3_a.log): 08b5b03dd0, 50 files/2088 drawable, 3803 bodies, 24546 members / 2835 kept / 9346 skipped, 1707 s apply, R22 218/218 112 s. Batch kept at 50 files though > 10 min: detached (R115), heaviest first; deviation.
• cycle 2..60 launched (.run/P39/restruct/cycle_t3_b.log).
• retriever-code (inline): --write-types is preview-only; canonical = struct defined in engine_types.h / struct_types.h; no-field = no leaf at exact off; head_types.py regenerates only its marked region; no tool adds fields to hand-authored structs.
• cycle 2..60 (cycle_t3_b.log): t3_S2..S5 green (R22 218/218 each, last 69f9da8c0f); members 12384/7786/7878/6694; then t3_S6 0 of 0 drawable … type-not-canonical 103469 → cycle exit 0 "nothing drawable". Drawable count never fell (2088→2089). Probe: current body_base_type.json 109,359 bases, top type Unkstruct_800D3888 (absent from canonical headers). Hypothesis: census rerun renames struct-map types as sites become members; plan reads drifting map; plan also re-draws finished files.
• T3.c2 briefed: confirm, pin base→type (pre-sweep .run/P39/census/body_base_type.json), exclude finished files.
• T3.c2 done (d476f3c92e): cause confirmed (pinned 31,408/116,484 ov bases canonical vs current 0/103,469; ov_MAIN_012 func_8012E9C0 param_1 800B5CB8 → 800D3888); --bbt/RESTRUCT_BBT pin + s_finished; plan 1901 drawable.
• cycle 6..60 with pin (cycle_t3_c.log): t3_S6..S14, S16..S18 drew ~350 bodies each but 0 sites (written 7-9 files each); only t3_S15 real (5136 members, 37 kept, 1774 skipped). All R22 green. Stopped (TERM after INT ignored by bash during R22); restruct --restore dropped t3_S19 (8 files, 310 rows); HEAD 2f4594dce3.
• Hypothesis: apply path reads unpinned map; s_finished marks zero-site files done. T3.c3 briefed.
• T3.c3 done (568de6c713, 357451cbb9, e88040d15a): real cause = s_body_edits checked the drifting live struct_map before the canonical branch → every pinned base NO-TYPE-IN-MAP (t3_S6: 340 rows NO-SITE 332 / KEPT 8); written files in zero-site batches = MISALIGNED/NON-STRUCT respells. Fix: apply uses pin; s_judged (≥1 CANONICAL/LOCAL base) gates s_finished and ledger_index. Probe ov_MAIN_012_jr_801549F8.c IDENTICAL members 1. Plan 1872 drawable. check-casts unbacked 96 (t3_S1..S10) → T3.c4 briefed.