wizard-work/forge had 39 committed verify_*.py and no CHECK; tools/verify.py
names the lane zero times. verifier_gate.py runs them, keeps one row per file
bound to the git blob the rc was measured against, and grades that ledger in
under a second. First full run at 7b42050: PASS 21 (18 of them PASS* -- they
read an absolute host root), FAIL 18, all named with causes.
Five verifiers in this lane clean up with rmtree(os.path.dirname(tmp)), i.e.
they delete $TMPDIR itself; each file now runs under its own sandbox so the
harness cannot be measured instead of the lane.
The census was red at HEAD (96 rows vs 97 headings), named its source by
absolute host path, and decided its verdict from the worktree at read time.
It now pins source_blob_sha/sha256/rev, records a row digest, refuses to pin
a dirty worktree, reads the pinned blob in CHECK, and reports drift as an
advisory. 9 selftest cells; the non-decay claim is measured on real commits.
The monitor's retry (3 attempts, 2 s apart) had no reader: CHECK-q115 grades the
carry path and its harness replaces get_folder with a stub, so a repair that
retried once or never would have read the same. q118_verify.py drives the
deployed loop at the boundary it really crosses -- a fake curl first on PATH,
scripted codes, every call timestamped -- and grades recovery, exhaustion,
no-pause-on-success and the Q115 carry, with three mutants required to red the
cell they target.
Measured: CHECK-q118 8/8 legs rc 0; a copy with tries=1 reads 3 faults rc 1;
verify 10/10 legs, gate OK.
A document rendered after the LEGS table broke is accepted as current by
the pre-fix parser with no gate column at all, so mutant-doc-drifts alone
would pass a tree that never had the column. The new leg reads that same
mutant's fresh render and requires rc 0 with a '-' cell, pinning the
failure mode the repair removes.
Measured: q132_verify.py 9 legs, 0 fault, rc 0.
Two test cells opened with sys.path.insert(0, /root/.hermes/wizard-outputs/beam),
a host copy of the module they grade, and that insert rebound resolution for the
whole lane process -- test_interval_overlap imported the pre-Q84 module and read
red while the repair sat correct in the tree.
Each cell now binds Path(__file__).resolve().parent and asserts the import
landed there. New instrument: q133_verify.py behind CHECK-q133 (5 legs) and
q133_oracle.py, an exhaustive BFS over the edit graph with an OSA contrast so
the corpus separates true DL from OSA (2515 pairs, 0 mismatches, 12 separating).
verify.py: 10 legs, 10 pass, 0 fault. beam/CHECK: 595 tests, 0 failures.
The DL module's cells were module-level asserts: run_lane_tests.py --list
reports unittest=0 bare=0 for test_damerau_levenshtein.py, so no named case
moved when the module broke. CHECK_DL runs the shipped cells plus
oracle_dl.py, which grades a deterministic 5300-pair corpus against a BFS
over the edit graph (the definition of true DL) and proves in the same run
that the corpus reds both classic impostors (OSA, plain Levenshtein).
Measured: green on the tree (0/5228 mismatches, 42 OSA witnesses, 252 Lev
witnesses); rc 1 with named witnesses on a transposition-deleted mutant and
on an OSA module.
Q130's closed-artifacts leg added "path": os.path.join(...) to verify.py's
LEGS, so ast.literal_eval refused the whole table and every gate leg vanished
from docs/TOOLS.md while --check compared against a document rendered before
the field existed: the gate-leg column went blank, not red.
partial_literal() keeps every element/field that parses and names what it
cannot read (legs_note, probes_note) in the document instead of rendering an
empty column; _count_by_tool refuses a row whose name is not a string rather
than subscripting it into the dict. Selftest 20 -> 23 cells, pin moved with
it, docs/TOOLS.md re-rendered (23 tools, 10 gate legs).
marker + .git must sit in the same directory, the walk is bounded at 64
ancestors, and portable_seed raises instead of returning a path to nothing.
Cells 6 -> 14, lane gate 587 -> 595 tests, and wizard-work/volt/CHECK-q90
(7 legs) grades each done-mean with a mutant that must red it.
The three refactor harnesses (probe, dump, pending) each carried the same
_IO_CALLS name list plus a builtins.open spy. Measured at bb4a85e in a clean
clone: Path(__file__).read_text(), os.open(__file__, O_RDONLY), io.open(__file__)
and with Path(__file__).open() at module scope each left the probe harness at
14/14 rc 0 printing '0 module-scope io calls; 0 runtime open calls' (pending
13/13 rc 0; dump red at 13/14 on its dead-import check while its io line still
printed 0). The name-list walk resolves Path(p).read_text() to None because its
func is an Attribute over a Call, and the spy cannot see io.open/pathlib.
Now: wizard-work/forge/import_io_detector.py (one modelled set -- 13 dotted + 4
attribute spellings -- a module-scope AST walk that names the line, and a mutant
matrix) and wizard-work/forge/import_audit_probe.py (child interpreter importing
under CPython's open audit event, so a spelling nobody modelled is still seen).
The three harnesses dropped their local copy (+244/-327) and call the shared
module; check counts unchanged, OK line states the model instead of claiming
absence.
Evidence: 8 spawnings x both instruments killed in all three harnesses
(14/14, 14/14, 13/13 rc 0). Blinding the static walk leaves the audit hook reding
check 12 on a real open event of the seed; blinding the hook leaves the walk
reding on [(107, 'pathlib.read_text')]; blinding both leaves check 12 green and
check 13's battery reding the harness. _F = open / _F(DEFAULT_PATH) -- a spelling
no list names -- is caught by the hook alone.
Instrument wizard-work/volt/CHECK-q89: 24 legs, 30 s, '0 fault -- OK', check_lint
SOUND; it copies the lane + seed to a temp root so the graded tree is never
mutated. Fourth copy (verify_dump_seed_refactor.py) untouched at 16/16 and used
as an independent oracle: agrees 12/12. Q90's beam guard untouched.
verify.py --fast: 10 legs, 9 pass, 0 fault, 1 skipped, gate OK.
sol
spark
ledger
quill
volt
beam
volt
volt