check_sweep.py --paths at c95a339 read 195 entry points / 158 baseline paths /
37 fault(s), rc 1. That leg is not among the fifteen tools/verify.py runs
(grep -c check_sweep tools/verify.py -> 0), so nothing graded the gap.
- 37 rows signed by the lane's own retake.py (run_entry + classify imported,
not re-implemented), --timeout 120 --jobs 6 at HEAD c95a339; read back with
check_sweep.read_baseline: 24 green, 13 red, 0 ungradable. Red rows are
recorded red with their own first fault line; signing is a reading, not a
repair.
- Two of the 37 answered rc 2, "could not grade": CHECK-troupe-gitfile-loc-census
is Python with no shebang and no .py suffix, so check_sweep.interpreter
dispatched it to bash (dies at line 15); CHECK-untracked-classify-2026-11-03
computed REPO_ROOT with one `cd ..` instead of two and exec'd
wizard-work/wizard-work/sol/verify_untracked_classify_2026_11_03.py.
Repaired (shebang; cd ../..), re-taken: rc 2 -> 1 on both, and both 1s are
real MISMATCH verdicts on their own subjects.
- Dispatch class measured narrow: 1 of 195 entry points had a docstring first
line dispatched to bash, 0 of 196 after the repair. (An import/from heuristic
flagged 83 and was wrong; it counts bash wrappers that exec python3.)
- New guard wizard-work/volt/CHECK-q174, 4 legs, 0 faults, rc 0 in 15.5 s:
paths / dispatch / runs / teeth (red control at c95a339 via git show).
- paths leg after: 196 entry points, 196 baseline paths, 0 fault(s), rc 0.
Not fixed, named: the sweep's flat 120 s timeout grades the host as much as the
artifact (the same run read 119 green / 44 fault, 27 of them rc 124; CHECK-q149
alone exceeded 400 s while CHECK-q62-citation reads 0 in 72 s), and the --paths
leg is still not a leg of tools/verify.py.
gate: verify.py --fast 15 legs, 14 pass, 0 fault, 1 skipped (selftest-sweep), rc 0
volt
e7fcea300b