T3.c4: restruct t3c4_rejudge: 31 bodies re-walked, 122 macro rows carried; check-casts unbacked 0

phase-ends/current/logs/T3.c4.md

@@ -0,0 +1,29 @@
+# T3.c4 — check-casts unbacked 96 → 0
+
+## Causes (96 unbacked K sites, 31 bodies, 18 files; scratch .run/T3c4/diag.py, diag2.py)
+- 60 base spelling `local:param_1` (restruct) vs `param:param_1` (census): restruct `file_view`/`body_sites` params_of matched only `d["name"](`; an alias-named def (`func_X_impl(…) __asm__("func_X")`) found no params → every param base classed local. Census `find_decls_and_flows` has the alias fallback (type_census.py:488-490). Rows t3_S1 32, S8 8, S10 6, S2 7, S3 4, S18 3. Side effect: those bodies' bbt bases (`param:…`) never met their sites → sites classed NON-STRUCT.
+- 23 superseded: a later S row re-judged a body already respelled; macro (form K) sites are not RAW_FORMS → the latest row has no KEPT entry for them (17: MEMBERS/KEPT rows, an earlier row backs exactly; 6: latest NO-SITE, outside BACKED_S_VERDICTS).
+- 13 nested base spelling: outer KEPT row keyed on pre-respell text (`nested:*(s32 **)(a0 + 0x20)`); after the inner site became `CAST_SIGN(…)` (or a member) the census parses the outer macro's base as `other:None` / another nested text. Same off, same line.
+- None from off format; no respell without a KEPT row.
+
+## Fix (tools/restruct.py)
+- params_of alias fallback in `body_sites` and `file_view` (census rule).
+- `s_macro_rows(tu, text, rows)` (called in work_file_S before the UNCHANGED return): every K site of a drawn body in the final text gets a `KEPT(<cls>)` row in the census key (`site_row` of the K site, `macro_of`=backing label), backed only by an entry KEPT(<cls>) of this row or an earlier rung-S row of the body (`s_prior_rows`, lazy ledger index `S_PRIOR`) with same off and (same base or same line); REINTERP by line. No-edit verdict (NO-SITE/TYPE-NOT-*) carrying one → KEPT. Unbacked macros get nothing (gate stays red).
+- `--redraw` for rung S in --apply (was D only; key (rung,tu,unit)); lifted bodies not s_finished. `--only TU:FN` names one body (overlay fn names repeat). A lifted named TU:FN body with no drawable base (plan_S skips: bbt has only type-not-canonical bases) is re-walked with `{}` bases.
+- selftest: alias-def param control; real-case control (func_8017DD0C shape: `CAST_WIDTH(s32, CAST_SIGN(s32 *, a0, 0x20), 8)`, prior t3_S1 row nested key, re-judge NO-SITE) negative unbacked 2 without prior row, positive KEPT unbacked 0 with it; TU:FN only. 93/93.
+- type_census matcher unchanged.
+
+## Backing
+- R22 `make clean && make extract-all JOBS=16 && make check-all JOBS=16` → 218 passed, 0 failed of 218 (on 0938e511ed+); `delever_oracle --snapshot-baseline` (7428 objects); `--calibrate ov_SC04_011 ov_SC03_015 md_SC07_004 main -j 16` OK (re-run at a83e31047f after the 3rd tool commit).
+- `restruct --apply --rung S --batch 50 --label t3c4_rejudge --redraw KEPT MEMBERS NO-SITE -j 12 --only <31 TU:FN>` (.run/T3c4/only.txt) → 18 files UNCHANGED, written 0, 31 rows all KEPT, 122 macro rows carried (t3_S1 57, S4 11, S2 16, S3 15, S8 8, S10 6, S5 4, S18 3, S9 1, S15 1). src untouched → no make check needed (git status src clean).
+- First apply attempt: plan_S drew 0 (all 18 bases type-not-canonical) → the named-body re-walk added.
+
+## Verified
+- `type_census.py --sites --check-casts -j 16` → `… (unbacked 0); lying=229 (ledgered 0)` (rc 1 for raw, expected).
+- `restruct.py --selftest` 93/93; `type_census.py --selftest` 50/50; `make kit-corpus` ok; `make tools-health` OK.
+
+## For the expert
+- Bodies re-judged here now have `bases=[]` rows (not s_judged): not done, but plan_S never draws them (no usable base).
+- Latent: a REFUSED / COMBINATION-FAILED row on a body holding macros leaves them unbacked (not in the 96).
+- `--plan` path does not apply --redraw for S (apply only).
+- Census outputs under .run/P37/census and .run/P36/delever/calibration.json were already dirty at start; not committed here.