fix(anthropic): keep MiniMax thinking blocks on replay

[ahinks]

Problem

MiniMax's /anthropic endpoints were classified as generic third-party in _manage_thinking_signatures, so ALL thinking blocks were stripped from every replayed assistant turn. The model's prior chain of thought silently disappeared between turns and inside tool loops, breaking interleaved reasoning on multi-turn agentic work after the 2nd or 3rd turn. Same bug class as the Kimi fix in #13848 / ddd81e9, which MiniMax was left out of.

Evidence (live probe, 2026-08-01, MiniMax-M3 via api.minimax.io/anthropic)

• Turn 1 returns a signed thinking block (64-char signature)
• Replay verbatim (signed block included): HTTP 200
• Replay stripped: HTTP 200
• Replay with content-mutated thinking, stale signature: HTTP 200

MiniMax does not enforce thinking signatures at all, so the strip was pure loss, no 400 protection.

Fix

Replay MiniMax thinking blocks as-is, mirroring the Kimi family rule. Two files:

• agent/anthropic_adapter.py: add _is_minimax_anthropic_endpoint branch to _manage_thinking_signatures (pass-through), update the docstring
• tests/agent/test_anthropic_minimax_thinking_replay.py: new, pins both MiniMax endpoints (io + cn) and the interleaved tool-loop turn

Verification

12/12 tests pass: new MiniMax suite (3), Kimi replay suite (3), DeepSeek thinking (3), thinking block order (3).

[alt-glitch]

This was generated by AI during triage.

Related: #75748 and #66694 address the same MiniMax replay family. This patch preserves signed thinking blocks verbatim, whereas #75748 normalizes them unsigned; this is a replay-contract decision, not a duplicate.

[teknium1]

Thanks for tracing the current replay path. The current-main premise is verified: agent/anthropic_adapter.py:2490-2493 classifies MiniMax as third-party, and agent/anthropic_adapter.py:2521-2528 strips its replayed thinking blocks. The proposed branch correctly reuses the existing MiniMax-only predicate.

Problems

• The new pass branch applies to redacted_thinking too: _THINKING_TYPES includes it at agent/anthropic_adapter.py:2487, but tests/agent/test_anthropic_minimax_thinking_replay.py covers only normal thinking blocks. Please establish the MiniMax contract for replaying redacted_thinking.data, or handle it conservatively.
• The signed-block policy needs maintainer consolidation. Open #75748 changes the same branch to demote signatures, while this PR preserves them verbatim; the MEMBER discussion identifies this as a replay-contract decision.

Suggested changes

• Select one signed-thinking replay policy and consolidate the related MiniMax PRs around it.
• Add a redacted-thinking regression case if verbatim replay is intended.

Automated hermes-sweeper review.

[teknium1]

This pass-through also preserves redacted_thinking blocks and their opaque data payloads because _THINKING_TYPES includes that subtype, but the new tests only exercise ordinary thinking blocks. Please add a MiniMax replay contract test for redacted_thinking, or limit the carve-out until that payload behavior is verified.