fix(ci): the audit called a pricing table a retired model

[catomean]

fix(ci): the audit called a pricing table a retired model

OrangeCat has nothing wrong with it. It renders a tier comparison —

models: ['Claude 3.5 Sonnet', 'GPT-4o', 'Gemini 2.0 Flash']

— human-readable copy that never reaches an API. Every entry but one carries
a space and was filtered out; GPT-4o was not, so it slipped the shape filter,
landed nearest an OpenRouter marker, and got announced as a retired pin.

That is the one failure a daily gate cannot afford. Six repos really are broken
by dead pins right now, and a report that puts a healthy repo alongside them
teaches the reader to skim — taking the real six with it. This fleet already has
that scar: gates that fail red on good code are what trained --no-verify.

The rule is measured, not guessed. On 2026-08-26 Groq listed 14 model ids and
OpenRouter 416, and not one of those 430 contained a capital letter. So an id
with a capital cannot be a RETIRED id at those vendors — it was never one of
their ids at all. "Retired" is the wrong diagnosis, not just the wrong target.

Such an id is reported as unattributed — listed, not judged. Never dropped.
For GPT-4o that bucket is also literally right: it IS an OpenAI product name,
and OpenAI is a vendor this audit does not query.

The rule stays where the evidence is. Together AI really does serve ids with
capitals (meta-llama/Llama-3.3-70B-Instruct-Turbo-Free is in this fleet), so
it applies only to the two catalogues that were counted, and a test pins that.

Live sweep before: 12 retired across 6 repos. After: 11 across 5, with OrangeCat
correctly absent and every other finding unchanged.

Self-test: 36 checks, no network, no key, no checkout. The new block asserts
both halves — the label is not called retired, AND the genuinely retired pin
sitting in the same file is still judged. Suppressing a false positive is only
worth doing if the true positive beside it survives.

🤖 Generated with Claude Code