Stop retired models from failing API key checks; lock off unavailable models

[ProfSynapse]

Summary

Adding an OpenRouter key failed with "No endpoints found for liquid/lfm-2.5-1.2b-thinking:free": the key check sent a test message to a free model OpenRouter has removed. A new provider is only turned on after the check passes, so the key was saved but OpenRouter stayed off. It looked connected but didn't work.

Rule applied everywhere: a retired model never fails a provider. It only can't be turned on.

Key checks

Provider	Before	Now
OpenRouter	chat to liquid/lfm-2.5-1.2b-thinking:free (404)	GET /api/v1/key
OpenAI	chat to gpt-5-nano	GET /v1/models (403 = restricted key, accepted)
Anthropic	chat to claude-haiku-4-5-20251001	GET /v1/models
Google	gemini-2.5-flash:generateContent	GET /v1beta/models
Mistral	chat to mistral-tiny	GET /v1/models
Groq	chat to llama3-8b-8192 (retired, always failed)	GET /openai/v1/models
DeepSeek	chat to deepseek-v4-flash	GET /models
Perplexity, Requesty	chat test message	still a chat test message, but a model error (400/403/404 about the model, not the key) accepts the key

Unavailable models

The provider dialog fetches the provider's model list and locks off chat models it no longer lists, labelled "No longer available from ". Saved toggles are not changed, so a model that comes back keeps its setting. Dots and hyphens compare equal, because OpenRouter serves anthropic/claude-haiku-4-5 but lists anthropic/claude-haiku-4.5. If the list can't be fetched or comes back empty, every model stays available.

Live today, this flags only OpenRouter's xiaomi/mimo-v2-omni and xiaomi/mimo-v2-pro (0 endpoints). Nothing is flagged for OpenAI, Anthropic or Google.

Test plan

• New unit tests: LLMValidationService.test.ts (23), ProviderModelAvailability.test.ts (12). The validator tests fail on the old code (7 of the original 14).
• Full Jest: 556 suites / 6,663 tests passing; npm run build passes
• Live against real endpoints:
  • valid keys pass for OpenRouter, OpenAI, Anthropic and Google;
  • bogus keys are rejected for all nine providers, with the provider's own message.
• In the running app (test vault), the OpenRouter dialog shows 58 models, with the two MiMo models locked off and labelled
• Mistral, Groq, DeepSeek, Perplexity and Requesty valid-key path: no keys available. Their bad-key responses were verified live.

🤖 Generated with Claude Code