fix(agentx): allow sparse low-concurrency profile tails

[cquil11]

Summary

Lower the AgentX profile-liveness cutoff from 98% to 95%. On a one-hour profile, the allowed quiet tail changes from 72 seconds to 180 seconds; the run still fails when neither TTFT nor inter-token latency reaches that cutoff.

False-positive runs

• MI355X MiniMax-M3 TP2 c1: 167 valid requests, 0 errors, complete server metrics; both latency signals reached 97.2%.
• MI355X MiniMax-M3 TP4 c1: successful low-concurrency profile with the same sparse-tail pattern.

Error and fix

At low concurrency, a long request can remain in flight through the end of an otherwise healthy profile, leaving no new latency sample in the final 72 seconds. The 95% cutoff tolerates that sparse tail while still detecting the intended failure mode: a server that produces neither latency signal for the final three minutes.

Validation

• 103 focused tests passed
• Ruff and format checks passed
• all pre-commit hooks passed

[github-actions]

Try out this PR

Quick install:

pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@c5f353f0ae120bc4d2531122dbcec90194ffcaa9

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@c5f353f0ae120bc4d2531122dbcec90194ffcaa9

---

Last updated for commit: c5f353f • 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.