Summary

Salvage of PR #7096 by @kshitijk4poor. Cherry-picked onto current main with contributor authorship preserved.

Aligns MiniMax provider with official API documentation (source). Fixes 6 bugs:

1. Transport mismatch — providers.py had openai_chat for minimax, should be anthropic_messages
2. Credential leak in switch_model() — fell back to resolve_anthropic_token() for all anthropic_messages providers, sending Anthropic creds to MiniMax
3. Prompt caching sent to MiniMax — is_native_anthropic was set from api_mode alone, now requires provider == "anthropic"
4. Dot-to-hyphen corruption — MiniMax-M2.7 → MiniMax-M2-7 (model-not-found). Added minimax to preserve-dots set
5. Trajectory compressor 404 — raw /anthropic URL → OpenAI SDK appended /chat/completions (404). Now uses _to_openai_base_url()
6. Doctor health check — MiniMax entries had None URL. Now uses /v1/models

Also corrects:

• Context window: 204,800 (was 1M/1.048M)
• Model catalog: M2.7/M2.5/M2.1/M2 only (M1 not on /anthropic endpoint)
• Thinking: fully supported in manual mode (was blocked)
• Max output: 131,072 tokens

Salvage notes

• Dropped test_setup_model_selection.py change (file was deleted from main in dead code cleanup)
• 39 tests added/updated across 9 test classes
• 2562 passed, 6 pre-existing flaky failures (pass in isolation)

Closes #7096