Add Mistral spam detection support

[dtump]

Summary

• add Mistral as a supported AI provider, defaulting to Mistral Large 3
• support global, EU, and future US inference endpoints
• make output token limits configurable and include raw response diagnostics for malformed JSON
• increase the opt-in live benchmark default from one fixture to five
• document Mistral configuration and add provider/dispatcher coverage

Validation

• .venv/bin/ruff check .
• .venv/bin/ruff format --check .
• .venv/bin/pytest -q — 96 passed, 1 skipped
• partial 20-fixture live Mistral run — 15/15 completed fixtures detected as spam before the API returned a 429 rate limit