Fix headless thinking replay and benchmark efficiency

[ericleepi314]

Summary

• preserve Anthropic signed thinking and redacted-thinking blocks across tool turns
• emit thinking blocks in headless stream-json logs
• preserve captured Bash output when detached descendants hold pipes open
• apply the existing deferred-tool filter to actual model requests

Validation

• 114 focused regression tests pass
• 88/89 broader query/headless tests pass; the remaining unrelated test requires PyYAML in the host environment
• Harbor pypi-server benchmark: reward 1.0, 7 model calls, 8 tool calls, 75.9s agent execution
• benchmark artifact: eval/harbor/jobs/2026-07-23__01-38-54/pypi-server__UFBsWUn