What changed

• Allow LangChain StructuredTool instances to register in ToolManager by accepting tools that expose invoke.
• Stop the research-agent ReAct loop once max_react_step is reached, even after tool-selection errors.
• Record tool-selection errors as tool steps so repeated invalid selections can make progress toward the configured limit.
• Add unittest coverage for tool registration and loop termination behavior.

Why

The CLI could appear stuck during the research phase: one research agent completed, while another kept cycling through tool_selection -> call_tool and never wrote a report. The root causes were an empty/under-registered toolset for LangChain tools plus termination logic that let tool-selection errors bypass the max-step cutoff.

Validation

• .venv/bin/python -m unittest tests/test_research_agent_loop.py
• .venv/bin/python -m py_compile contest_trade/agents/research_agent.py contest_trade/tools/tool_utils.py tests/test_research_agent_loop.py
• .venv/bin/python -m cli.main config
• git diff --cached --check