fix(llm): gate MiniMax reasoning_split by model capability (#826)

MinimaxChatOpenAI unconditionally set reasoning_split=True, but the
kwarg is only valid on M2.x reasoning models. The openai SDK's strict
kwarg validation raised TypeError for Coding Plan and any other non-
reasoning MiniMax model.

Adds requires_reasoning_split to ModelCapabilities, gates the payload
injection on it, and only sets True for _MINIMAX_THINKING (M2.x exact
IDs and the ^MiniMax-M\d forward-compat pattern). Same shape as the
existing supports_tool_choice gate.

Regression tests cover both halves: M2.x models still receive the flag,
non-reasoning MiniMax models do not.

tradingagents/llm_clients/capabilities.py

@@ -38,6 +38,12 @@ class ModelCapabilities:
     # DeepSeek thinking-mode models 400 if reasoning_content from prior
     # assistant turns is not echoed back on the next request.
     requires_reasoning_content_roundtrip: bool = False
+    # MiniMax M2.x reasoning models need ``reasoning_split=True`` so the
+    # <think> block lands in ``reasoning_details`` instead of polluting
+    # ``content``. The flag is rejected by non-reasoning MiniMax models
+    # (Coding Plan, MiniMax-Text-01, etc.), so we only set it where the
+    # model actually consumes it. (#826)
+    requires_reasoning_split: bool = False
 
 
 # DeepSeek's thinking models accept the ``tools`` array but reject the
@@ -74,6 +80,7 @@ _MINIMAX_THINKING = ModelCapabilities(
     supports_json_mode=False,
     supports_json_schema=False,
     preferred_structured_method="function_calling",
+    requires_reasoning_split=True,
 )
 
 _DEFAULT = ModelCapabilities(