BFM-decomp/phase-ends/current/logs/T7.1.md

T7.1 log — GTE to one spelling

Timeline:

• Start HEAD 32f83ea79a; no restruct_cycle running. Read T7/T1/T2/T6 summaries, plan sections, task entry.
• Baseline lever_census --check --strict --residue config/lever_residue.tsv --sites -j 16 (.run/logs/t71_strict0.log): gte/direct 207, gte/via-macro 6256, gte-unsigned/direct 270, gte-lever 434 (346 direct, 88 via variant), per-TU asm defs 314 {gte 150, launder 154, instruction 9, barrier 1} (GTE variants 64).
• Scratch analysis (.run/P39/t71/gte_sites*.py over .run/P36/census/lever_sites.jsonl): 207 inline gte in 172 files / 189 bodies (ov 190, src/800.c 9, shared/ov 8), 159 distinct texts; 270 gte-unsigned = 2 statements × 134 copies of func_8013D9B0 + 2; via non-header macros 15 (gte_stsxy3_f4 10, gte_rt 3, gte_mvmva0 2).
• retriever-code (inline answer, no report): lever_census kind/via mechanics (classify_direct_gte :119-137, strict_failures :1125-1146, pt :954); gte_consolidate inventory/canonical/variants/apply/ledger mechanics; render_header emitted canonicals only; gte-unsigned = statement not in sig_cache.
• c1 (55ba77f9d8): lever_census strict direct GTE filter (via header macros pass), split output, by-kind defs, selftest v-vii. Live: direct 492 = via-per-TU 15 + inline 477.
• c2 (8c3937da22): gte_consolidate signs all inline (823/823), canonical table gains inline-only signatures, variants/aliases emitted in header under // !FAKE: lines, file_edits deletes per-TU defs; header 51 -> 82 #defines (68 canonical + 11 variants + 2 aliases + guard). Coder stopped at context threshold before R22.
• Expert ran R22 on 8c3937da22: 218/218 exit 0 (.run/logs/t71c2_r22.log).
• c3 (e672acfa6d, 6d61877b30, 72e4c48c89): recalibrated 146/146; added --keep-levers; batches c3b1 40 files, c3b2 395 files, all CONSOLIDATED IDENTICAL, 0 REFUSED/DEFS-ONLY; R22 218/218 (.run/logs/t71c3_r22.log); kit-corpus, tools-health OK.
• Verify (expert): .run/logs/t71_verify_a.log (selftests, --check rc 0, strict counts), .run/logs/t71_verify_r22.log 218/218 exit 0 (src unchanged since 6d61877b30).
• D1: type_census refreshed into .run/P39/census (.run/logs/t71_tc.log; first snapshot attempt refused STALE, .run/logs/t71_d1.log), then snapshot + timeline + README (.run/logs/t71_d1b.log).

Hypotheses rejected:

• "gte-lever direct statements must be respelled": the tool's default respelled them into variant calls; plan leaves them to T8, so c3 added --keep-levers.

Coder briefs: c1 lever_census filter; c2 gte_consolidate header variants + inventory + header regen; c3 recalibrate + apply + R22 + census. Retriever questions: one retriever-code (GTE census/consolidate mechanics), answer inline, no report id. Open: 55 files keep a stale /* GTE VARIANT gte_rt_m … */ comment and // !FAKE: gte via gte_rt_m marker text although the call now reads gte_rtv0tr_m (e.g. src/ov_SC06_008/ov_SC06_008_jr_801380E0.c:1006,1028); comment-only, left for T8 which re-marks the gte-lever sites.