[ProfSynapse]

Summary

This follows #420 and makes sure no key check can break again because a model was retired.

• Providers with a model-free key check use it. This was done in Stop retired models from failing API key checks; lock off unavailable models #420: OpenRouter /key, plus the model lists for OpenAI, Anthropic, Google, Mistral, Groq and DeepSeek.
• Providers without one (Perplexity, Requesty) now send the test message to the cheapest models in their own registry:
  • priced models come first, up to three attempts;
  • an error about the model moves on to the next one;
  • an error about the key stops immediately;
  • if every attempt is refused only as a model, the key is accepted.
• Today's picks:
  • Perplexity: sonar → sonar-reasoning-pro → sonar-deep-research
  • Requesty: openai/gpt-5.4-nano → mistral/mistral-large-4 → moonshot/kimi-k2.7-code
• Requesty's old hard-coded test model, openai/gpt-5-nano, was not in its registry at all.
• Guard test: fails if ValidationService.ts names a model id again, either as model: '...' or as a /models/<id>:generateContent URL.
• Removed Gemini 2.5 Flash and Gemini 2.5 Pro from Requesty. Google Cloud retires both on 2026-10-20, and new Gemini API projects already get "no longer available to new users".

Test plan

• LLMValidationService.test.ts, 29 tests:
  • cheapest model first;
  • fallback to the next model;
  • every model gone → key accepted;
  • a key refusal stops at once;
  • ordering, deduplication and the attempt cap;
  • the guard test.
• Full Jest: 6,669 passing; npm run build; check_model_registry.py requesty perplexity clean
• Live with bad keys: Perplexity stopped after sonar and Requesty after openai/gpt-5.4-nano, each with the provider's own error. No valid keys were available to test a successful check.

🤖 Generated with Claude Code