fix(agentx): accept either latency signal as liveness

[cquil11]

Summary

• keep the AgentX profiling-coverage threshold at 98%
• accept either late TTFT or late ITL as proof of global server activity
• continue failing when both signals stop before the cutoff

False-positive runs

• MI325X c2: 130/130 requests succeeded; TTFT coverage was 93.0%, ITL coverage was 100%
• MI325X TP8 c4: 295/295 requests succeeded; TTFT coverage was 97.9504%, ITL coverage was 100%

Error and fix

The liveness gate required both TTFT and ITL to reach 98%. Healthy long-running responses could keep streaming through the profile tail without starting a new request, causing ProfileMetricCoverageError despite continued server activity.

This changes the verdict from TTFT and ITL to TTFT or ITL. A new first token or continued streaming proves liveness; a genuinely hung server still fails because neither signal reaches the cutoff.

Validation

• 103 focused unit tests passed
• all applicable pre-commit hooks passed
• regression coverage includes TTFT=97.9504% with ITL=100%
• runs where both signals stop early remain invalid

[github-actions]

Try out this PR

Quick install:

pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@0cf63afada56a7deda5f48348cf011d517fd1a34

Recommended with virtual environment (using uv):

uv venv --python 3.12 && source .venv/bin/activate
uv pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@0cf63afada56a7deda5f48348cf011d517fd1a34

---

Last updated for commit: 0cf63af • Browse code

[github-actions]

Fern Docs Preview: generation failed — see the Actions log for details. This does not block merge; ask a maintainer to retry if needed.