fix(copilot_auth): route Responses-only GPT models through /responses

[snair20-1984]

Why

Every Copilot model was built as an OpenAIChatModel, so every request went to /chat/completions. Copilot's newer GPT models (gpt-5.5, gpt-5.6-luna/sol/terra, gpt-5.4-mini, gpt-5.3-codex, plus grok and mai-code) are Responses-only and reject that with:

status_code: 400, model_name: gpt-5.6-terra, body: {'message': 'model "gpt-5.6-terra" is not accessible via the /chat/completions endpoint', 'code': 'unsupported_api_for_model'}

The Copilot /models catalogue already says which endpoints each model accepts (supported_endpoints), but fetch_copilot_models threw that away and add_models_to_config persisted nothing the model factory could key off.

Routing those models through OpenAIResponsesModel clears the 400 but exposes a second problem: Copilot's /responses proxy re-encrypts every identifier on every SSE event. The id in output_item.added, the item_id on each function_call_arguments.delta / output_text.delta, and the item.id in output_item.done are all different opaque strings for the same item. pydantic-ai keys streamed parts by that id, so tool calls came through with empty arguments (read_file ... file_path Field required, retried to death) and text was split into one part per delta.

What

• utils.py — fetch_copilot_models now returns {"id", "api", "context_length"} entries derived from supported_endpoints and capabilities.limits.max_context_window_tokens. add_models_to_config persists a new copilot_api: "chat" | "responses" field and uses the catalogue context length (terra was registered at 128k; the API says 400k). Chat stays the default for dual-endpoint models (gpt-5.4, gpt-5-mini) and for Claude/Gemini, so the reasoning_opaque path is untouched. Bare-string ids and the DEFAULT_COPILOT_MODELS fallback still work.
• register_callbacks.py — _create_copilot_model picks OpenAIResponsesModel when copilot_api == "responses". Entries written before the field existed fall back to a name heuristic (gpt-5* / codex), so existing copilot_models.json files are fixed without a re-login. Provider, headers, _CopilotAuth and base_url handling are unchanged; the chat branch still returns the same class as before, so this rebases cleanly onto fix(copilot_auth): tolerate missing object/index in chat completions #23's CopilotChatModel.
• responses_stream.py (new) — httpx byte-stream shim, same shape as the reasoning_client.py interceptor, that rewrites id / item_id per output_index so every event for an item shares the first id seen. Installed only on the Responses path; /chat/completions, non-streaming JSON and error responses pass through send untouched.

Tests

• tests/test_copilot_auth_utils.py (new): preferred_api_for_endpoints truth table, catalogue parsing with the real response shape, registration writes copilot_api and context length, normalised entries round-trip.
• tests/test_copilot_responses_stream.py (new): id normalisation across the exact event sequence captured from gpt-5.6-terra, chunk-boundary buffering, non-JSON/[DONE] passthrough, MockTransport coverage that only streaming /responses bodies are wrapped.
• tests/test_copilot_auth_model.py: Responses vs chat class selection, heuristic fallback, shim installed only on the Responses path, Claude keeps its reasoning profile.

Full suite: 2181 passed, 2 skipped. ruff check / ruff format --check clean.

Verified live against an enterprise Copilot host: copilot-gpt-5.6-terra answers and completes a multi-tool round trip (list_files → read_file → answer), copilot-claude-sonnet-5 (chat + reasoning_opaque) and copilot-gpt-5.4 (dual-endpoint, chat) still work, and a legacy entry without copilot_api selects the Responses model via the heuristic.

Follow-ups (not in this PR)

• code_puppy's model_factory still strips reasoning_effort for all Copilot GPT models because /chat/completions rejected it; the /responses catalogue entries advertise reasoning_effort support, so that can now be enabled behind copilot_api == "responses".
• The plugin still builds a legacy httpx.AsyncClient, which triggers pydantic-ai's httpx2 deprecation warning; separate change.

Fixes #57