Upgrade default MiniMax model to M3

[octo-patch]

Summary

Bump the default model for the MiniMax recaption provider from MiniMax-M2.7 to MiniMax-M3, the latest generation released by MiniMax. M3 ships notable improvements over M2.7 (long-context understanding and instruction following), so it's a better fit for the recaption step that rewrites editing instructions.

The previous default and its highspeed variant remain selectable for users who want to pin the older generation:

export RECAPTION_PROVIDER="minimax"
export OPENAI_COMPATIBLE_MODEL="MiniMax-M2.7"           # previous default
# or
export OPENAI_COMPATIBLE_MODEL="MiniMax-M2.7-highspeed" # previous highspeed variant

Changes

• agent/llm_provider.py — OpenAICompatibleProvider's default model now MiniMax-M3; docstrings updated.
• agent/config.py — comment block reflects new default.
• README.md — Option 2 (MiniMax) describes M3 as default; M2.7 and M2.7-highspeed listed as previous-generation alternatives.
• tests/test_llm_provider.py — test_defaults_to_minimax asserts the new default.
• tests/test_integration.py — explicit model= arguments in the live API tests bumped to MiniMax-M3.

Unchanged

• Base URL: still https://api.minimax.io/v1.
• Env vars: MINIMAX_API_KEY / OPENAI_COMPATIBLE_API_KEY, OPENAI_COMPATIBLE_BASE_URL, OPENAI_COMPATIBLE_MODEL, RECAPTION_PROVIDER.
• Temperature clamping (0.01, 1.0] — required by MiniMax, applies to M3 the same way.
• <think>...</think> tag stripping — M3 still emits these in some configurations, so the existing post-processing is kept.
• ROI detection (multimodal) still uses Gemini directly.

Test plan

• pytest tests/test_llm_provider.py -v — 33/33 pass locally.
• Manual factory check: get_recaption_provider() with RECAPTION_PROVIDER=minimax returns an OpenAICompatibleProvider whose _model is MiniMax-M3 and _base_url is https://api.minimax.io/v1.
• Live tests/test_integration.py (requires RUN_INTEGRATION_TESTS=1 + a valid MINIMAX_API_KEY) — not run in this PR; unit tests cover the wiring.