fix(minimax): ignore M3 models.dev context underreport

[mark202]

Summary

• reject models.dev's 512K underreport for MiniMax-M3
• fall through to the hardcoded 1M MiniMax-M3 context default after stale cache invalidation

Test Plan

• python -m pytest tests/agent/test_minimax_provider.py::TestMinimaxM3StaleCacheGuard::test_stale_m3_cache_dropped_and_reresolves_to_1m -q
• python -m pytest tests/agent/test_minimax_provider.py -q
• python -m pytest tests/agent/test_model_metadata.py tests/agent/test_models_dev.py tests/agent/test_minimax_provider.py -q
• python -m ruff check agent/model_metadata.py tests/agent/test_minimax_provider.py
• python -m ty check agent/model_metadata.py

[teknium1]

Thanks for the focused MiniMax-M3 investigation. This is an automated hermes-sweeper review; current main already provides the requested guarantee.

• 5a4297a11a83c38ac24eec7df0e4e41d6b3dbb9f (fix(model_metadata): prefer hardcoded 1M for MiniMax M3 over stale models.dev probe) is on main and shipped in v2026.6.19.
• agent/model_metadata.py:2375-2384 detects MiniMax-M3, rejects a models.dev value below the hardcoded catalog value, and returns that 1M catalog value.
• agent/model_metadata.py:281-287 defines minimax-m3 as 1_000_000, ahead of the generic MiniMax fallback.

Closing as implemented on main.