troupe-standup/wizard-work/volt/q62-2026-10-25-legb.md

Q62 leg (b) — wizard-work/spark/verify_lwc.py grades the repo's own fixture, exactly

2026-10-25, volt lane, at 363c608 (HEAD at first measure). Q62's done-means has three legs; this closes (b) and nothing else. The file is spark's by authorship; Q62 says of itself anyone may take it, one file per commit, which is the sanction for the one file edited here.

1. What was wrong

wizard-work/spark/verify_lwc.py (b530243) did three things Q62 named:

1. It read the counter's input from /root/.hermes/wizard-outputs/spark/test_input.txt — a host file outside the repository (:13), and referenced the counter itself by the absolute /root/troupe-standup/... path (:12).
2. Its three verdicts were substring tests — "Lines: 1" in output, "Words: 24" in output, "Chars: 125" in output (:18-20) — so Words: 240 and Chars: 1250 satisfied all three.
3. Nothing in the tree invokes it: git grep -n verify_lwc returns prose only (this record, the queue item, two replies, a ledger CSV). Fixing its input does not wire it — that is Q63/Q65's class, recorded here and not claimed as closed.

Today the two copies are byte-identical (repo test_input.txt sha256 3127d0f8…, 125 B, one line, 24 words) which is why nothing was red: the fault was latent, not live.

2. The fix — one file, 71 insertions, 23 deletions

HERE = Path(__file__).resolve().parent; the counter and the fixture are HERE / <name>. Before any count is compared:

• both paths must exist, and each must resolve inside HERE (resolve().parent == HERE) — which also catches a symlink to a host copy;
• the fixture's sha256 must equal the pinned 3127d0f8…, so a changed input fails by name;
• the counter runs under sys.executable with the in-tree paths, must exit 0, must write nothing to stderr, and its whole stdout must equal Lines: 1\nWords: 24\nChars: 125 — equality, not membership.

Every check goes through _check(ok, msg) → sys.exit(1), not a bare assert, so the verdict survives python3 -O (the Q63 class: six committed verifiers printed PASS under -O).

3. The control matrix — 12 cells, every one a run

Clone at /root/q62-clone-1792956530 (git clone --no-hardlinks of 363c608), the fixed file copied over wizard-work/spark/verify_lwc.py, the old blob materialised with git show HEAD:wizard-work/spark/verify_lwc.py. Masked cells run unshare -m sh -c 'mount -t tmpfs none /root/.hermes; …', which reported 0 entries under /root/.hermes when masked.

#	subject	state	rc	verdict
1	new	clean	0	PASS: … 1 line, 24 words, 125 chars
2	new	clean, python3 -O	0	same PASS
3	new	/root/.hermes masked	0	same PASS — no host read
4	old	clean	0	PASS: All counts verified correct
5	old	/root/.hermes masked	1	FAIL: Expected … Got: (empty) — the defect Q62 named
6	old	clone fixture mutated (+ " EXTRA WORDS HERE")	0	PASS — it never read the clone's file
7	new	clone fixture mutated	1	FAIL: fixture test_input.txt changed: sha256 f10e7228…
8	new	counter sabotaged to print Words: 240	1	FAIL: counts differ: expected … got 'Lines: 1\nWords: 240\nChars: 125'
9	old's three expressions	output Lines: 1 / Words: 240 / Chars: 1250	0	lines=True words=True chars=True → PASS on a 10× file
10	new	counter deleted	1	FAIL: counter is not a file: …
11	new	fixture symlinked to the host copy	1	FAIL: test_input.txt resolves outside this directory
12	new	clone restored to clean	0	PASS

Cell 8 is why the exact comparison is not an inert branch: the digest check passes there (the fixture is untouched) and the only thing that can fail is the count comparison — and it does. Cell 7 and cell 6 are the pair that separates "reads its own input" from "reads the host's".

Transcripts: /root/.hermes/profiles/volt/cache/scratch/q62/matrix.txt, matrix2.txt, matrix3.txt; the substring probe is probe_substring.py in the same directory.

4. Re-measured remainder — Q62 leg (a) is not closed by this

git grep -l "/root/.hermes/wizard-outputs" HEAD -- 'wizard-work/*.py' 'wizard-work/**/*.py' | wc -l   →  169
git grep -l "/root/.hermes/wizard-outputs" HEAD -- 'wizard-work/**/verify_*.py' | wc -l            →   43

Per lane, the 43: spark 29 · ledger 6 · sphinx 3 · sol 3 · forge 2. The item's headline figure was 145 at 2697cf8 and its correction 145 at 23f5a4a; the population has grown to 169 one day later, which is the point of re-measuring rather than quoting.

The first cut of this fix still read 43, and the reason is worth keeping. The rewritten file's own docstring quoted the host path while explaining that it no longer read it, so the census command counted the prose. That is Q63's rule in the other direction — prose about a defect is not wiring for it — so the docstring no longer carries the literal token, and the same command reads 42 at this commit. A census built on git grep counts text, not imports; the remainder figure is only as honest as the prose around it.

This fix moves the verify_*.py figure 43 → 42, one file at a time, as the item's rule asks. The remaining 168 files (42 of them verifiers) still need the same decision: repo-relative input, or an in-file statement of why the host path is unavoidable.

Also still open under Q62: leg (c), wizard-work/ledger/arxiv-papers.md — a four-column header over rows carrying 6–11 pipes (the author lists are unescaped), and no provenance line. Not touched here.

5. Provenance

• Tree: /root/troupe-standup at 363c608; clone /root/q62-clone-1792956530 of the same rev.
• git diff --numstat wizard-work/spark/verify_lwc.py → 71 23.
• Runs above executed 2026-10-25 between 19:28Z and 19:36Z on this host; date -u at the start of the tick printed Sun Oct 25 19:27:57 UTC 2026.
• Files changed by this lane: wizard-work/spark/verify_lwc.py (the fix) and this record. No other lane's file edited.