fix(session/runner): force MiniMax XML tool-call mode for threatbook-…

[duguwanglong]

The threatbook-cn-llm gateway exposes MiniMax models over an OpenAI-compatible endpoint but its streaming chunks omit the tool_calls field entirely (observed 2026-04 with minimax-m2.7: first ChoiceDelta only carries content / role, model_extra_keys=[]). With native function-calling the model can never emit a tool call, so every turn ends with finish_reason=stop and tool_calls=0 — users in IM channels just see a short text reply (e.g. a couple of IPs) and no tool execution.

Add threatbook-cn-llm to the MiniMax text-call provider whitelist so _should_use_text_tool_call_mode() returns True for this provider+model pair. The runner then injects the
<minimax:tool_call> XML instructions and the existing text parser picks the calls up from the content stream, restoring tool execution end-to-end.

• Other models routed through the same gateway (qwen, GLM, etc.) remain on the standard OpenAI native function-calling path.
• Comment in _should_use_text_tool_call_mode() records the root cause for future maintainers.

Tests:

• New regression cases in TestMiniMaxTextToolMode:
  • threatbook-cn-llm + minimax-m2.7 → XML mode enabled
  • case-insensitive provider/model id handling
  • threatbook-cn-llm + qwen3.6-plus → XML mode disabled
• tests/session/test_runner_step.py (48 cases) all pass.