docs(llms): the roster count the pages state, and a guard that derives it

[yusuf-gundogdu]

docs/llms/setup.md and docs/llms/methodology.md still described a twenty-two model roster.
Both were written when that was true, and neither moved when five models landed in #572 and a sixth
in #628. Against a document now holding twenty-eight, the pages read:

Page	Said
setup.md	"Twenty-two models are supported"
setup.md	"Twenty-one of the twenty-two run locally"
setup.md	"One of the twenty-two is hosted"
methodology.md	"Twenty-two models, six surfaces, five runs: 660 runs, and all 660 passed."
methodology.md	"The 660 runs above were opened over HTTP"
methodology.md	"Ten, not twenty-two"
tests/unit/lib/agent/tools.test.ts	"which two of twenty-two shipped models carry"

Two pull requests and a review passed over all of it, and that is the part worth fixing rather
than the seven numbers. Nothing anywhere compared the sentence to the file it describes.

The guard

tests/unit/lib/agent/model-roster-docs.test.ts now does that comparison, derived on both sides:
the count comes from modelProfiles() and the total from count * 30, which is the protocol those
pages define — six surfaces, five consecutive runs each.

The first version of it did not bite, and that is recorded in the file. Asserting only that a
page contained the current word left setup.md green when its opening sentence was reverted to
twenty-two, because two later sentences in the same file still said twenty-eight and includes was
satisfied by them. Each claim is now captured by its own sentence and compared; the run total is
checked against every roster size between ten and sixty rather than against the totals this project
happens to have had.

Both halves were confirmed to bite by reverting the two sentences that actually leaked:

Mutation	Result
setup.md opening sentence back to "Twenty-two models are supported"	1 fail
methodology.md back to "660 runs, and all 660 passed"	1 fail
Both restored	2 pass

What it deliberately does not do

A blanket ban on number-words would fail on sentences that are correct. These pages legitimately
say twenty-seven (the models that run locally), twenty-one (the models at the 90-second
ceiling), Ten (the browser sweep) and thirty runs (one model's own sweep). A guard that
failed on those would be deleted within a week, so it pins the claims and the total instead.

The claim patterns are hand-listed because a sentence cannot be derived; the number in each is what
the test derives. A page whose phrasing changes fails loudly rather than silently stopping being
checked — which is the failure mode this file was written after.

Gates

format · lint · typecheck · knip · test (38 groups, 0 fail) · build — all green.
Coverage: 45992/45992 lines, 100.00%.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud