Integrate eval protocol to rllm trainer

[1stprinciple]

Note

Guard signal handler registration to main thread and move LiteLLM policy initialization outside server start for consistent retries and setup.

• Pytest MCP server manager (default_mcp_gym_rollout_processor.py):
  • Guard SIGINT/SIGTERM registration to the main thread using threading.current_thread() checks.
  • Add threading import.
• MCP gym rollout processor:
  • Refactor LiteLLMPolicy creation to occur after server start (and on retries), rather than inside the server start block.
  • Relax retry precondition to require existing server only (no policy check).
  • Keep environment creation and rollout execution unchanged aside from using the newly created policy.

^{Written by Cursor Bugbot for commit 339e8ce. This will update automatically on new commits. Configure here.}

[xzrderek]

thanks!

[cursor]

Bug: Server Cleanup Fails on Policy Creation Error

Policy initialization code was moved outside the exception handling block. When start_server=False (retry scenario), if policy creation fails (lines 242-260), there is no exception handler to clean up the server that was started in a previous call. This can leave the MCP server running indefinitely without cleanup, causing resource leaks. The exception handler at lines 228-233 only catches exceptions during server.start() when start_server=True, not during policy creation when start_server=False.