feat(ai-import): use OpenRouter supported_parameters for temperature

[silasg]

Summary

Builds on #4. OpenRouter advertises each model's accepted parameters via supported_parameters. When a model is picked from that provider, we now record whether it supports temperature and pass it into the request config — so a model known not to accept it skips even the first (doomed) attempt instead of learning by failure.

Stacked on #4 — base branch is claude/issue-1151-fix-proposal-4a0dkf. Review/merge #4 first; the diff here is only the OpenRouter capability wiring. Independent of #5.

Changes

• Parse supported_parameters into CloudModel.supportsTemperature (true/false; undefined when the provider advertises nothing).
• Carry the flag from the model picker into a new cloud_ai_model_supports_temperature setting (null = unknown; cleared when the provider changes).
• Feed it into the cloud LLM request config as supportsTemperature.

Testing

Added a model-list spec asserting supportsTemperature is derived correctly from supported_parameters (present → true, absent → false, no metadata → undefined).

Same environment caveat as #4: the pure cloud-model-list mapping is verified via an esbuild bundle of the real module and tsc; the picker/settings wiring is syntax-checked but the full Karma run is blocked locally by the xlsx egress policy. Please confirm via pnpm test in CI.

🤖 Generated with Claude Code

---

Generated by Claude Code