BFM-decomp/phase-ends/current/logs/T2.c3.md

T2.c3 — type_census --check-casts (coder log)

Changed

• tools/type_census.py
  • KEPT_MACROS registry (CAST_ALIAS/SIGN/WIDTH/MISALIGNED/NONSTRUCT → class; LOBU..LOWU → REINTERP), FORM_K regex, DEFLINE_FILE/DEFLINE_RX (after FORM_C).
  • find_sites: form K pass after M; every K site has tu, fn, line, pos, end, name, cls, bclass ("other" default), base, off; CAST_* get parse_base(args[1]) + classify, off += int(args[2]), ctype = args[0], stars=1 (so width/sign fill). raw_counts["K"] counted so coverage per walk stays equal. K not in RAW_FORMS (build_struct_map, deref_total, by_bclass untouched).
  • walk_file: for include/common.h, sites on a #define <KEPT_MACROS|M2C_FIELD> line get defline=True; abs_casts excludes K.
  • check_casts_counts(sites, raw_counts) → ({P,I,X,M} minus defline sites, macro sites minus deflines); used by run_census and selftest.
  • summary: casts.macros {name:n}, casts.check_raw, casts.macro_sites; decls.lying_rows [{callee, defined_in, kind}].
  • check_casts_inputs() (lazy restruct import; rs.ledger_latest(rs.load_ledger()); TSV exceptions set), check_casts_verdict(raw_by_form, macro_sites, latest, lying_rows, exceptions) → (rc, line); --check-casts in main.
  • selftest: +18 checks (per-form positive + near-misses p->f, CAST_ALIAS-as-P/M, (u8)x, q[k]; K sites name/cls/base; common.h define lines dropped vs same text in a .c counted; backed / unbacked / wrong class; lying ledgered vs not; all-zero exact line; raw FAIL).
• config/lying_exceptions.tsv: # contract line + header, no rows.
• Caches: removed .run/P37/census/cache/walk_cache.json; .run/P39/census/cache absent after rm -rf.

Base string

• ledger site base = restruct.site_row f"{bclass}:{base}" from census sites after classify() → "param:a0"; K uses the same classify → match. No mismatch.

Verified

• /usr/bin/python3 tools/type_census.py --selftest → 50/50 checks OK, rc 0.
• live --check-casts --no-cache --out-dir .run/P39/census -j 16 --quiet → check-casts: raw P=408974 I=60666 X=13800 M=19542; macros none (unbacked 0); lying=229 (ledgered 0), exit 1 (expected FAIL). by_form deref_total 502988 (P=408979 incl. the 5 CAST_* define lines in common.h).
• --check-structs --quiet → type_census --check-structs: OK, conflicting_types=0, rc 0.
• make kit-corpus → rc 0, < 2 min.

Notes for the expert

• restruct.py:802/914 iterate ALL tc sites with s.get("bclass") + ":"; K sites therefore always carry a bclass (REINTERP "other"). A CAST_* K site with a ctype on the same base now adds its ctype to restruct's ctypes_by_off (evidence, not a raw site); restruct's RAW_FORMS filters elsewhere exclude K.
• type_census.json grows ~1.2k lines (lying_rows, indent=1).
• Uncommitted side effects of the briefed commands (not in my commit paths): --check-structs (default out-dir) rewrote tracked .run/P37/census/{type_census.json,txt,struct_map_top.json} and docs/struct-map.md (19123 types / 502988 sites); kit-corpus touched decomp-architect/corpus/* and docs/gen3-standards.md diffs present (source of gen3-standards change not verified). Also P37 walk cache was rebuilt by the --check-structs run under the new TOOL_STAMP.