tool arg validation dumps injected args

Realized this was an issue due to https://github.com/langchain-ai/deepagents/issues/491 This is an issue bc we shouldn't be serializing context Root Cause In langchain_core/tools/base.py, the _parse_input method validates tool args and then calls model_dump(): ```py result = input_args.model_validate(tool_input) result_dict = result.model_dump() # <-- triggers serialization of ToolRuntime ``` When ToolRuntime is injected into tool args, Pydantic serializes it according to the type hints in the tool function. If the tool is typed as ToolRuntime[None, SomeState] but the actual runtime has a non-None context (e.g., {'num': 1}), Pydantic warns about the type mismatch during serialization. The injected args are handled separately afterwards (around line 745-747), but model_dump() still attempts to serialize them first. Potential Solutions 1. Exclude injected args from serialization - pass exclude parameter to model_dump() for fields that are in _injected_args_keys 2. Restructure validation flow to not serialize args, this might be breaking... Fix would be in core.