fix(scanner): enforce the verdict schema where the rows are written

[elfrost]

Follow-up to #154 (ADR-016), driven by its first production run.

What the first run showed

Scan #109 (overwirehq/claude-code-telegram, 2026-09-22) produced verdicts.json with 7 scanner rows and 6 curation rows — the plumbing worked end to end, and the published post carried both new blocks.

Every one of the 6 curation rows was invalid:

Defect	Rows
rule empty	6/6
verdict free prose ("by-design not-reachable", "FP - parameterized")	6/6
confirmed-real used for 13 dependency CVEs under a "0 real" headline	1

The test suite was green throughout. It exercises the constructor; production does not — the curation half is hand-written JSON, so VerdictRecord.__post_init__ never runs. A closed vocabulary is only closed if something closes it.

A fourth problem is structural rather than drift: reports/ is gitignored, so the corpus lived on one machine and would not survive a clone. ADR-014's measurement only worked because the archived reports happened to still be on disk.

Fixes

• scanner/verdict_corpus.py + scanner/run_verdict_check.py — --check reports every problem in every row in one pass and exits 2. /daily Phase 4 must see exit 0 before publishing. A checker that reveals one defect per run costs as many runs as there are defects, so fields are checked independently rather than through the first-raising constructor.
• rule required on every record. ADR-014 needed sqlalchemy-execute-raw-query fired 157 times, not "some rules were dismissed".
• dependency-currency added; confirmed-real documented as first-party only.
• corpus/verdicts/ committed — the durable half. reports/ stays ignored (it holds raw dumps and unsent disclosure drafts); a negation rule there would need a fragile reports/* + !reports/*/ ladder that risks exposing reports/disclosures/. --summary counts the corpus.

scanner/verdicts.py hit 296 of the 300-line ceiling, so validation and corpus loading moved out.

Not fixed, on purpose

Scan #109's six invalid rows are left as they are. Repairing them would mean inventing rule identities that were never recorded, and a corpus with one fabricated entry is worth less than one with a known hole. Scan #110 is the first valid entry.

Verification

• ruff clean · black 77 files unchanged · pytest 416 passed (26 new)
• Every module under the 300-line limit
• --check run against the real scan docs: seventy-ninth public scan (liaohch3/claude-tap) — 1 real finding withheld #109 file: 12 problems across 6 rows, in one pass

🤖 Generated with Claude Code