BFM-decomp/phase-ends/current/logs/T5.c3.md

T5.c3 — calibrate, pilot, fleet sweep of the rung-S I/M recipe ladder (done)

Calibrate

• delever_oracle.py --snapshot-baseline → 7428 objects at 9d2c7e9dc, 218/218 link inputs.
• --calibrate ov_SC04_011 ov_SC03_015 md_SC07_004 main -j 16 → 146/146 identical, twin 32/0, positive control DIFFERS, linked-control OK.
• restruct.py --selftest --real → 112/112 OK (brief expected 110/110; the selftest now has 112 cases).

Pilot t5_imS1 (ov_SC07_006, commit 1134b7a12)

• 17 files · 58 bodies · final 17/17 · written 9 · R22 218/218 (130.6 s) · restruct wall 25.1 s.
• im_by_recipe {R2-extern-array: 14}; im_by_cause {R2-NO-SCALAR-EXTERN 140, R2-WIDTH 91, R2-DIFFERS 26, R2-REKEYS 12, R2-COMPILE-ERROR 9, R3-NOT-CONTIGUOUS 1}.
• ledger im.seconds over 58 bodies: mean 0.31 s, max 3.31 s.

Sweep (BATCH=200, LABEL_PREFIX=t5_im, TASK=T5, all detached, one at a time)

• ov_* (.run/P37/restruct/cycle_t5_im_ov.log): S2–S13 green, each final N/N + R22 218/218, committed 6908e68e4 … 05b01fb54; S14 nothing drawable.
  • restruct wall per batch: S2 679 s (heaviest files first), then 97–398 s.
• md_* (cycle_t5_im_md.log): S15 46/46, drawn 9, R22 218/218, 8a09f83c7; S16 nothing drawable.
• main (cycle_t5_im_main.log): S18 6/6, drawn 0, b06a6794e; S19 nothing drawable.
• src/shared/*: --plan --only 'src/shared/*' → 0 drawable files; not run.
• after: restruct.py --plan --rung S --forms I,M → 0 of 0 drawable files.
• totals over 15 batch records (S1–S13, S15, S18): drawn 1671 (all R2-extern-array); residual 35947 {R2-NO-SCALAR-EXTERN 18879, R2-WIDTH 10126, R2-DIFFERS 3983, R2-REKEYS 1755, R2-COMPILE-ERROR 1182, R3-NOT-CONTIGUOUS 15, R3-COMPILE-ERROR 4, R2-HEADER-EXTERN 3}.

Final gate

• type_census.py --check-casts --residue --quiet → exit 1: raw P=257464 I=49836 X=13790 M=15128 (uncovered P=0 I=12157 X=0 M=13704).
• I/M uncovered vs T5.c1 baseline I=35625 M=14762 → I=12157 M=13704. The rest are non-global-base I/M sites (outside the --forms I,M scope); the exit is non-zero for that reason.

Notes for the expert

• S2 restruct ran 679 s, over the 10-min target; later batches ran 97–398 s, so BATCH stayed at 200. Fixed cost per batch is about 8 min (R22 about 130 s plus census).
• Each batch printed "readability snapshot refused (a stale census?)"; the cycle continues past it. Not investigated.
• Draw rate is low (1671 of 37618 sites, 4.4 %). R2-NO-SCALAR-EXTERN and R2-WIDTH are 80 % of the residual.
• No tool edits. No restore needed.