fix: handle BrokenPipeError in LMHandler during parallel execution

[alexzhang13]

Summary

• Fix BrokenPipeError that occurred during parallel evaluation runs (e.g., --parallel 9)
• Add _safe_send() helper method that catches connection errors gracefully
• Handle connection errors separately from other exceptions to prevent cascading failures

Problem

When running parallel evaluations with the RLM eval harness:

python use_cases/example_workflows/eval_harness.py rlm_oolong ... --parallel 9

Multiple workers connect to the LMHandler socket server simultaneously. When workers complete and close their sockets (or the connection is interrupted), the server would raise BrokenPipeError when trying to send responses. The error handler would then try to send an error response on the same broken socket, causing another BrokenPipeError.

Solution

1. Added a _safe_send() helper that wraps socket_send() and catches connection errors
2. Handle BrokenPipeError, ConnectionError, ConnectionResetError, and OSError separately - these are expected during parallel execution and can be silently ignored
3. Use _safe_send() for all socket operations in the handle() method

Test plan

• Verified module imports correctly
• RLM test suite passes (143 tests)
• Manual test with --parallel 9 to confirm BrokenPipeError is no longer logged

🤖 Generated with Claude Code