BFM-decomp/phase-ends/current/tasks/T4.md

T4 — fold the head types fleet-wide (rung L)

Status: done | expert: expert-opus55 | ctx-at-completion: n/a | commit: see COMMIT | coder runs: c1 opus55 done (d87c35435), c2 opus55 done (5ae5fe94c), c3 opus55 done (61c49eda0) Done: 162 of 200 local head-type definitions folded onto Unkstruct_800B5CB8 through judged rung-L head-fold batches (accesses renamed by offset), each batch committed after its R22; 38 kept with cause; the three canonical names sit in no dup class, layout twin or variant camp. Files:

• tools/restruct.py — rung L head-fold mode (--only <canonical names>), type-aware offset renames incl. includers, sibling headers, macro bodies; --status --rung L --only buckets; atomic flocked inflight, apply.lock, TOOL-ERROR restore; +selftests (76)
• tools/restruct_cycle.sh — $TASK in commit msg, ledger committed with add -f, no CURRENT_PHASE.md append
• include/struct_types.h — player block pad44[3] → u8 unk44..46 (md_MAIN_011 widths)
• src/** (≈145 TUs), src/shared/engine_types.h, src/shared/ov/*.h — folds (batches t38_4probe, L1, L2, HL1, HRL1)
• docs/ops/p38-t4-2026-09-29-restruct-head-fold-mode.md + docs/ops/INDEX.md — the mode, ledger, cycle gotcha
• docs/struct-map.md — regenerated stamp Decisions:
• binding: a head fold is keyed by the canonical NAME and layout SIZE, not by layout hash; non-head objects of the same size are excluded by name glob with a cause (HF_EXCLUDE: Blk152* copy blob, Fr_* stack frame); width/sign mismatches are left to the byte judge, never refused up front.
• Local definitions become typedef <Canon> <Local>; aliases (uses keep the local name); renaming uses to the canonical name is left for a later readability pass. Deviations:
• 0x24C (Unkstruct_80126B58) has no local definition anywhere (scan .run/P38/t4/headsized.json); nothing to fold.
• Player block: 0 of 6 md_MAIN_011 copies folded (all DIFFERS) even after the pad44 split.
• Overall --check-structs still FAILs (fleet-wide gates for later tasks); the done-when's clause (the three in no dup class/camp) holds.
• Batch t38_4L1 ran twice concurrently (my relaunch); 3 edited-unjudged files reverted by hand; the tree committed only after a green R22. Findings:
• Kept by cause: DIFFERS 29 (Ent_8017FD58 family 20 + Ent_801202A0 in engine_types.h; md_MAIN_011 player copies 6; ov_SC03_094 Ent_8017D898; ov_SC03_102 Ent), indexed array spanning fields 4 (Cnt_8018E8A0 ×3, Actor_80178D40 unkDC[7]), canonical pad at 0x3C 3 (Ent_8018D654 ×2, Ent_8018D3A4), COMPILE-ERROR 2 (Obj8017EBC4 ov_SC06_029_jr_8017C954.c:4027, Obj1CDC ov_SC05_010_jr_80181CDC.c:2873). Full list logs/T4.md.
• Remaining head-sized dup classes are all kept/excluded locals: Ent_8017FD58×20, Cnt_8018E8A0×3, Ent_8018D3A4/654, Ent/EntD820, Blk152×6, Fr_×2.
• The 20 Ent_8017FD58 DIFFERS + pad3C refusals are the max-site split's losers: evidence for T5's Unkstruct_800B5CB8 split/union call.
• Counts after: dup_classes=205 outside_canon=5917 dead=142 variants=40 pad_names=25 parse_error_decls=1106. Gotchas:
• workflow: a detached restruct_cycle.sh log stays 0 bytes until its first batch line; ps for restruct_cycl[e] before any relaunch (a double launch raced inflight.json).
• harness: restruct_cycle.sh stops on final X≠Y and leaves edited files as restruct left them; revert unjudged files before R22.
• harness: .run/P37/restruct/ledger.jsonl is tracked under an ignored dir; commit_task.sh cannot stage it (the cycle uses add -f).
• generalizable: an opaque-tier struct fold must rename accesses in every file the definition reaches (includers, their sibling headers, macro bodies), type-aware by base, or the judge sees compile errors, not layout verdicts.
• harness: a header head-fold judges every includer (~55 s/unit, 29 units = 28 min); budget header batches. Research: none (one retriever-code answer inline, no report) Next task needs: T5 decides the 800B5CB8 split/unions using the kept DIFFERS/pad3C units above; md_MAIN_011 player copies need width reconciliation; --rung L --only <canon names> is the fold tool for any new canonical type. Verified: .venv/bin/python tools/type_census.py --out-dir .run/P38/census --check-structs; .venv/bin/python tools/restruct.py --check-ledger && make clean && make extract-all JOBS=16 && make check-all JOBS=16 → census FAIL (fleet counts above; head names in 0 dup classes/twins/variants), check-ledger 0 violations, extract-all 217 extracted 0 failed, check-all 218 passed 0 failed, rc 0 (.run/logs/t4verify.log) Full log: phase-ends/current/logs/T4.md