Summary

Fix delegation observability so terminal tool traces are classified by structured exit_code when present, instead of substring matching on the word error.

Previously a successful terminal result such as:

{"output":"error count: 0", "exit_code": 0, "error": null}

was marked as status: "error" in tool_trace because the serialized result contained the string error near the beginning.

Changes

• Teach _looks_like_error_output() to treat JSON results with exit_code as terminal-style results.
• exit_code == 0 => not an error.
• nonzero exit_code => error.
• Reuse _looks_like_error_output() for _run_single_child() tool_trace status.
• Add regression test for terminal output containing the word error with exit_code: 0.

Test plan

source venv/bin/activate
pytest -q tests/tools/test_delegate.py -q

Result: passed.