fix(ai): the platform LLM was calling a model that no longer exists

[catomean]

Found by asking the Cat a real question in production and reading why it answered with its fallback:

platform-llm: model call failed  status: 404
model: llama-3.3-70b-versatile

Groq stopped serving that id. Confirmed against the live API — 14 models available, and it is not among them.

This was never the Cat's problem

Every callPlatformJson caller returned null:

feature	file
offer engine	services/cat/offer-engine.ts
writing engine	services/cat/writing-engine.ts
writing revise	services/cat/writing-revise.ts
prompt suggestions	services/cat/prompt-suggestions.ts
platform feedback	services/cat/platform-feedback.ts
image suggestions	services/images/image-suggest.ts
voice intent router	services/voice/intent-router.ts
the Cat's replies	services/mentions/cat-reply.ts

Each is written to "degrade gracefully" — so eight features degraded gracefully into doing nothing, and the log line for it was a warn.

Three changes, and the model id is the least important

Failover. resolveProvider returned the first provider whose key existed and stopped. So a dead Groq id took everything down while OpenRouter sat configured and unused. It now returns a list and the call walks it. The ordering is unchanged and isn't the point — the point is that there's a second entry at all.

A 404 is now an error, not a warn, and says what it means: the pinned id has rotted. A model that no longer exists is a configuration fault, not a hiccup — it will fail identically until a human changes a constant, so it must not share a log level with a timeout.

check:ai-models asks each provider whether it still serves what we pinned, reading the ids out of source so the check can't drift from the constants.

Deliberately not in verify: it needs the network and a provider key, and a gate that goes red when an API hiccups is a gate that gets switched off. It runs where the keys live, and a provider it cannot ask is reported as SKIPPED on its own line — because the failure mode of this entire class of check is a skip being mistaken for a pass.

Proven both ways, against the live APIs

[check-ai-models] groq: openai/gpt-oss-120b — served (14 models)
[check-ai-models] openrouter: nvidia/nemotron-3-super-120b-a12b:free — served (417 models)
[check-ai-models] OK — every pinned model is still served.

and with the old id restored:

[check-ai-models] FAIL: 1 pinned model(s) no longer exist:
  groq: 'llama-3.3-70b-versatile' is NOT served (14 models available)

Verification

• npm run verify — exit 0 (2369 tests)
• gate proven red-then-green against the real provider APIs

🤖 Generated with Claude Code