T2: --check-structs instrument, actor-struct 4-byte fix; log and summary

docs/struct-map.md

@@ -1,6 +1,6 @@
 # The struct map — the types the fleet's cast sites need (Phase 37 T1)
 
-> **Generated by `tools/type_census.py` at `09b00a443` (2026-09-30T02:32:09Z); never typed (R75).** Every raw pointer-cast dereference in `src/` was assigned to a base (a parameter, a local, a global by value or by address, an absolute address) and the bases were clustered by EVIDENCE only — the same global symbol, an argument passed to a callee's parameter, an assignment or return, a pointer field read, a shared body's fleet-wide identity — never by a similar offset signature. Each cluster is one type the tree needs. Names are placeholders (`Unkstruct_<addr>`: the lowest global instance's address, else the lowest member function's); fields are placeholders (`unk<HEX>`) until Phase 38 cites evidence.
+> **Generated by `tools/type_census.py` at `3ac7abee7` (2026-09-30T02:38:06Z); never typed (R75).** Every raw pointer-cast dereference in `src/` was assigned to a base (a parameter, a local, a global by value or by address, an absolute address) and the bases were clustered by EVIDENCE only — the same global symbol, an argument passed to a callee's parameter, an assignment or return, a pointer field read, a shared body's fleet-wide identity — never by a similar offset signature. Each cluster is one type the tree needs. Names are placeholders (`Unkstruct_<addr>`: the lowest global instance's address, else the lowest member function's); fields are placeholders (`unk<HEX>`) until Phase 38 cites evidence.
 
 **18739 types explain 498866 of 502985 dereference sites (99.2 %); 4119 sites have a base the census could not classify; 727 types carry a width conflict.** Merges by evidence: param 36454, assign 14664, arg 10168, field 9613, ret-of 3373, return 1524, assign& 1050, arg& 270.
 

phase-ends/current/logs/T2.md

@@ -0,0 +1,9 @@
+# T2 log — --check-structs instrument and actor-struct fix
+- Start: read card slice, plan Context/Interfaces/Cookbook/Research, T2 entry, tasks/T1.md, docs/actor-struct.md.
+- Hook denied a sed read of tools/restruct.py source (expert learns tools from --help); design delegated whole to coder.
+- Coder brief T2.c1 (opus55): add --check-structs (gating dup_classes, outside_canon, dead, variants, pad_names via restruct's own audit, parse_error_decls via restruct's latest-verdict function, controls); non-gating conflicting_types/tu_conflict/types_floor_lying; CANON_HEADERS single source; ≥2 selftests; actor-struct +0x108/+0x10C s16→s32; --check unchanged.
+- T2.c1 returned done, commit 3ac7abee7: restruct gains ledger_latest (one latest-row function; ledger_index, settled-set and apply_batch --redraw use it) and audit_type_violations; type_census --check-structs + 3 selftests (24/24); restruct selftest 54/54; --check line identical before/after; control list was hard-coded (not doc-parsed) so D_80078F08/F0C moved into PLAYER_BLOCK at width 4, doc_disputed=0; audit-types 26 violations all pad-name kind.
+- Verify (run.sh t2verify, .run/logs/t2verify.log): selftest OK; --check-structs FAIL dup_classes=206 outside_canon=6076 dead=141 variants=40 pad_names=26 parse_error_decls=1106 controls=4/4; doc_disputed=0; non-gating conflicting_types=160 tu_conflict=0 types_floor_lying=1796; rc=1.
+- Known-true vs T1 baseline: 206/6,076 (4,085+1,904+63+24)/141/40/26/4-4 all equal. parse_error_decls 1,106 latest-row (1,107 all-rows); P37's 1,120 was a pre-latest-row figure, T1 recorded none.
+- docs/struct-map.md regenerated by the verify census (stamp line) committed with the task.
+- Retrievers: none. Report ids: none.

phase-ends/current/tasks/INDEX.md

@@ -1,3 +1,4 @@
 <!-- Task index for one phase. One line per task, written by tools/task_log.py finish; the cumulative copy is phase-ends/TASK_INDEX.md, prefixed with the phase. Grep to find which task touched a subject. -->
 <!-- T<n> | <status> | <title> | <tags> | tasks/T<n>.md | logs/T<n>.md | <research ids, comma-separated> -->
 T1 | done | phase-start health and baselines | - | tasks/T1.md | logs/T1.md | -
+T2 | done | --check-structs instrument and actor-struct fix | - | tasks/T2.md | logs/T2.md | -

phase-ends/current/tasks/T2.md

@@ -0,0 +1,25 @@
+# T2 — --check-structs instrument and actor-struct fix
+
+Status: done | expert: expert-opus55 | ctx-at-completion: n/a | commit: see COMMIT | coder runs: c1 opus55 done (3ac7abee7)
+Done: `type_census.py --check-structs` gates the struct milestone's counts; on the T1 tree it FAILS with T1's baseline; actor-struct doc fixed so controls show 0 doc-disputed.
+Files:
+- tools/type_census.py — `--check-structs` mode, counts/controls/non-gating lines, 3 new selftests (24 total); CANON_HEADERS the single canonical list
+- tools/restruct.py — `ledger_latest` (the one latest-row-per-unit function; ledger_index, settled-set, `--redraw` use it); `audit_type_violations` (the --audit-types rule, imported by census)
+- docs/actor-struct.md — +0x108/+0x10C s16 → s32
+- docs/struct-map.md — regenerated stamp
+Decisions:
+- binding: every "latest ledger verdict" reader goes through `restruct.ledger_latest`; no second filter.
+- pad_names gates all restruct --audit-types violations; today all 26 are pad-name kind (`audit_other=0` reported non-gating to expose any other kind).
+Deviations:
+- The census control's doc-disputed list was hard-coded in type_census, not parsed from the doc; D_80078F08/F0C moved into PLAYER_BLOCK at width 4 alongside the doc edit.
+- `tu_conflict=0`: the restruct ledger has no TU-CONFLICT verdicts; P37's 45 TU-CONFLICT rows live elsewhere (census/logs). Non-gating, left for T6.
+Findings:
+- Baseline counts: dup_classes=206 outside_canon=6076 dead=141 variants=40 pad_names=26 parse_error_decls=1106 controls=4/4; non-gating conflicting_types=160 tu_conflict=0 types_floor_lying=1796.
+- parse_error_decls 1,106 latest-row vs 1,107 all-rows; the P37 "1,120" predates latest-row counting. conflicting_types 160 latest-row vs P37's "417".
+- `--check` output identical before/after.
+Gotchas:
+- harness: the expert's Bash hook denies reading tool source even by line range (sed); code facts must come via coder or retriever-code.
+- generalizable: a census "control" that compares against a doc may be a hard-coded list, not a parse; fixing the doc alone does not move it.
+Next task needs: `--check-structs` is the phase gate; baseline above; census at .run/P38/census.
+Verified: .venv/bin/python tools/type_census.py --selftest && .venv/bin/python tools/type_census.py --out-dir .run/P38/census --check-structs; test $? -eq 1 → selftest OK, rc=1 with the baseline counts, doc_disputed=0 (.run/logs/t2verify.log)
+Full log: phase-ends/current/logs/T2.md