[viztor]

Two more layers of the testing system, and both found something.

A language-pair gate

test/docs-pairing.test.ts. The README and docs/quota-meter are published twice, and every count in them is already held to the code — but nothing held the two languages to each other, which is the drift that actually happens: a row added to one and not the other.

It found two, immediately:

• the Chinese README's settings table carried a control that no longer exists — the table listed ten while its own prose said nine;
• the Chinese README's configuration table was missing credentialOrg.

The check compares structure — heading levels, fenced blocks, per-table body rows, list-item count — because prose is the translator's business but a section, a code block or a table row is a fact, and a fact in one language only is invisible to whoever reads the other.

The pair also disagreed about form in two places (one language stated the "three no-number states" as a table and the other as a list; the "two design rules" the other way round). Both now say it the same way, which is what makes the row check fair rather than a demand for literal translation.

The live-run verdicts are unit-tested now

test/e2e/verdict.ts + test/probe-verdict.test.ts. The rules that decide whether a live case skips or fails used to live inside the e2e file, where only a run with a key could exercise them — so a mistake in them showed up as a red or a silently-skipped case, at the vendor's schedule.

They are the rules that decide what a red run means, which makes them worth more than the routing cases they gate: two of the last four rotations turned a live case red for a vendor-side reason, and each time the fix was here. The extraction also puts the picker's rule in one place (usableAsSubject), so a case and its picker cannot disagree about what a usable subject is.

Two coverage gaps closed where they were behavioural

• decoratePlaneAdapters on a model the catalog does not describe: the answer passes through whole — no api/baseUrl invented for a model the plugin knows nothing about. Inventing one would send it to a guessed endpoint, which is worse than the adapter's own answer.
• The hook's "cannot be served" diagnostic: a @ai-sdk/mistral model is filtered out of the picker deliberately, and the warning names the SDK that decided it — once per route/model pair, because the hook runs on every request.

Verification

gate	result
vp check	0 errors, 23 pre-existing warnings
suite	1210 passed / 80 files (+2 files)
coverage (95/90/93/95)	96.32 / 92.05 / 96.20 / 96.31
scripts/check.ts	passes