feat(models): upgrade MiniMax provider to M3 as default

[octo-patch]

OpenJudge Version

0.2.0

Description

Upgrade the MiniMaxChatModel provider to use the new MiniMax-M3 as the default model, while keeping M2.7 / M2.7-highspeed available as alternatives. The older M2.5 / M2.5-highspeed entries are removed from the supported list.

Changes

• openjudge/models/minimax_chat_model.py
  • MINIMAX_MODELS now lists MiniMax-M3, MiniMax-M2.7, MiniMax-M2.7-highspeed (M3 first).
  • Default model parameter changed from MiniMax-M2.7 to MiniMax-M3.
  • Docstring example and supported-models block updated accordingly.
• tests/models/test_minimax_chat_model.py
  • test_default_model now expects MiniMax-M3.
  • test_custom_model uses MiniMax-M2.7 (was M2.5).
  • New test_minimax_models_contain_m3; list-length assertion relaxed to >= 3.
  • Integration tests run against MiniMax-M3.
• skills/mmx-cli/SKILL.md
  • Text-generation entry bumped to MiniMax-M3.
• cookbooks/auto_arena/examples/config.yaml
  • Example MiniMax candidate model upgraded to MiniMax-M3.

The OpenAI-compatible base URL (https://api.minimax.io/v1), temperature clamping, <think> stripping, and the rest of the model logic are unchanged.

How to test

pytest tests/models/test_minimax_chat_model.py -m unit -v

All 26 unit tests in this file pass locally; the wider tests/models/ unit suite (54 tests) also passes.

Checklist

• Code has been formatted with pre-commit (black/isort/flake8 on touched files clean)
• All unit tests are passing
• Docstrings are in Google style
• Related documentation has been updated (SKILL.md, cookbook example)
• Code is ready for review