Reason: make new MiniMax configurations use MiniMax-M3.1-Flash and preserve reasoning across tool turns.

Summary

MiniMax construction without a model currently inherits the OpenAI default, and the pinned ChatOpenAI adapter drops MiniMax's separate reasoning field from responses and subsequent tool turns. Select MiniMax-M3.1-Flash for new MiniMax configurations and preserve reasoning_content through normal and streamed responses and replayed history.

The current MiniMax catalog now lists Flash first and M3 second, both with a 1M-token context. Legacy M2.7 IDs retain provider routing and token limits outside that catalog; explicit model IDs, endpoint overrides, and M3 settings remain unchanged. Add Flash input/output/cache-read prices and document the default and reasoning settings.

Validation

• uv sync --frozen --python 3.12 with the repository lock (langchain-openai 1.1.6, langchain-core 1.2.3, openai 2.13.0).
• Exact offline CI suite from .github/workflows/test-suite.yml: 134 passed.
• MiniMax suite: 54 passed, including real SDK requests intercepted at the HTTP boundary for graph defaults, explicit/legacy/custom models, global/China endpoints, all five reasoning efforts, 524288-token output requests, sync/async streaming, tool replay, and image/video payload preservation.
• Pinned Ruff, Black, and isort checks pass for all five changed Python files; git diff --check passes.
• make lint reports the same six pre-existing Ruff findings on the untouched pre/beta base and this branch. No lint suppressions added.

Provider responses are mocked; these checks validate serialization and replay without calling the paid API. Browser/live-provider suites are outside the repository's offline CI list.

References: MiniMax OpenAI API, model catalog.