feat: Add openai-responses-api support and filter out unsupported models.

[tinysnake]

Summary

callVisionModel currently hardcodes one REST shape (POST {baseUrl}/chat/completions with a messages body and choices[].message.content parsing). Every model in pi-ai's registry carries a declared api type, and the delegate path already holds the full Model<Api> from modelRegistry.find() — so we can dispatch the request/response shape on that authority instead of guessing.

This PR makes vision delegation speak both openai-completions and openai-responses:

model.api	Endpoint	Request body	Response parsing
openai-completions	POST {baseUrl}/chat/completions	messages[] with image_url data URL (unchanged)	choices[].message.content || reasoning_content
openai-responses	POST {baseUrl}/responses	input[] with input_image / input_text, max_output_tokens (≥16), nested reasoning: { effort }	output[].message.content[].output_text
anything else	—	no network call	throws a clear "unsupported API type" error

Motivation

With a openai-responses model (e.g. gpt-5.6-luna on opencode-go) configured as the vision model, delegation sent a well-formed request to the wrong endpoint semantics and failed with a 400 — the model registry already declared the correct API type, but the extension ignored it. Fallback (openai-codex) was also broken in that environment.

Behavior changes

• Unsupported API types (anthropic-messages, google-generative-ai, …) now fail with an explicit error and flow into the existing retry/fallback path — classified as a client error, so no retry is wasted and the configured fallback model still fires. Previously they silently hit /chat/completions and failed confusingly.
• Model pickers + auto-detect filter to supported types: visionCapableModels() and autoDetectDefaults() now use the new lib/supported.ts (isUsableVisionModel()), so users only see vision models the delegate path can actually run. Adding a future API type means relaxing one filter + adding one dispatch branch.

Testing

• pnpm typecheck clean, pnpm test:run — 361 tests pass (was 353).
• New tests:
  • responses request shape: /responses URL, input[] body, input_image/input_text, max_output_tokens ≥ 16, system role, nested reasoning.effort (and no flat reasoning_effort)
  • responses response parsing: output[].message.content[].output_text, empty-output → "no content" error
  • unsupported API type: clear error, zero network calls
  • integration: unsupported primary → no retry → fallback fires
  • auto-detect: anthropic-messages models filtered out; openai-responses models still picked
• Verified end-to-end against opencode-go/gpt-5.6-luna (responses): image description returned correctly.