fix: support GPT-5.6 Responses reasoning

[jmdots]

Summary

• support GPT-5.6 OpenAI aliases through the Responses API
• configure native reasoning mode/context and encrypted reasoning replay
• preserve signed empty reasoning parts needed before tool calls
• preserve complete streamed output items, including encrypted reasoning
• add regression coverage for model settings, compaction, and response reconstruction

Closes #746

Compatibility note

Existing conversations created before this change may contain incomplete or incompatible Responses API reasoning history. Resuming those conversations may not work reliably; start a fresh conversation if resume fails.

Testing

• uv run pytest tests/test_compaction_reasoning.py tests/test_model_factory_reasoning.py tests/test_chatgpt_codex_client.py -q (61 passed)

[jmdots]

This isn't right, yet. I think it has something to do with "all_turns" reasoning context.

"/REDACTED/REDACTED/src/code_puppy/.venv/lib/python3.13/site-packages/pydantic_ai/models/ope
nai.py", line 221, in _map_api_errors
    raise ModelHTTPError(
        status_code=status_code, model_name=model_name, body=e.body,
headers=dict(e.response.headers)
    ) from e
pydantic_ai.exceptions.ModelHTTPError: status_code: 400, model_name: gpt-5.6-luna,
body: {'message': "Item 'REDACTED' of
type 'message' was provided without its required 'reasoning' item:
'REDACTED'.", 'type':
'invalid_request_error', 'param': 'input', 'code': None}