[arekay-nv]

What does this PR do?

Fixes the intermittent CI failure in
test_sigint.py::test_sigint_grace_expiry_abandons_wedged_drain
(FileNotFoundError: .../performance/result_summary.json), seen on 8 runs
across 4 branches; see #558.

Cause: the test waited for metrics/.ready and then slept a fixed 3 s
before SIGSTOPping the aggregator and sending ^C. .ready precedes HTTP
worker startup (_create_issuer) and the session's STARTED, and the
aggregator publishes no live snapshot until STARTED. When worker startup
outlasted the sleep, the wedged-then-SIGKILLed aggregator left neither a
final_snapshot.json nor a live frame, so drain_and_build_report had
nothing to build result_summary.json from. The product behaviour is within
the module's contract ("or no artifacts at all"); the test's timing
assumption was wrong.

Changes

• tests/integration/commands/test_sigint.py
  • New _wait_for_log() helper and _PERF_PHASE_STARTED constant
    ("Starting phase: performance", logged by BenchmarkSession._run_phase
    right after STARTED is published).
  • test_sigint_grace_expiry_abandons_wedged_drain: captures the run log,
    waits for the phase-start line before the 3 s sleep (now ~12 aggregator
    publish ticks past STARTED), and prints the log tail if
    result_summary.json is missing.
  • test_group_sigint_exits_gracefully: same wait, since it uses the same
    .ready + fixed-sleep pattern and expects interrupted perf artifacts.
  • test_sigint_in_drain_window_keeps_summary_authoritative: its inline
    log-polling loop now uses _wait_for_log (same needle and timeout).
  • Corrected the _wait_services_ready docstring, which claimed issuing
    starts right after service readiness.
• src/inference_endpoint/load_generator/session.py: comment-only. Notes
  that test_sigint.py waits on the Starting phase: log line and
  test_benchmark_command.py asserts on it, so the wording must stay in
  sync.

Type of change

• Bug fix

Related issues

Closes #558

Testing

These tests are Linux-only (/proc, CPU affinity), so they were run in a
Linux container (ghcr.io/astral-sh/uv:python3.12-bookworm-slim, aarch64).
Limiting it to 0.5 CPU slows worker startup the way a loaded runner does,
which reproduces the flake:

test_sigint_grace_expiry_abandons_wedged_drain @ 0.5 CPU	Run 1	Run 2	Run 3
main	fail: abort took 6.2 s	fail: FileNotFoundError	fail: FileNotFoundError
This PR	pass	pass	pass

• Full test_sigint.py with no CPU limit: 6 passed, 3 of 3 runs.
• CI's test selection (-m "not slow and not performance") on macOS:
  2599 passed, 55 failed. The failing set is identical to main's; all
  55 are macOS-only (CPU affinity / /proc).
• Pre-commit: all hooks pass except mypy, which on macOS reports the known
  os.sched_setaffinity errors in untouched files.

Known, unchanged: at 0.5 CPU,
test_sigint_in_drain_window_keeps_summary_authoritative fails on both
main and this PR (no final_snapshot.json). It is not seen in CI; details
are in the #558 comment.

Checklist

• Code follows project style
• Pre-commit hooks pass
• Documentation updated (if needed)

🤖 Generated with Claude Code

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅

[codecov-commenter]

Codecov Report

✅ All modified and coverable lines are covered by tests.
⚠️ Please upload report for BASE (main@520eb4d). Learn more about missing BASE report.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #559   +/-   ##
=======================================
  Coverage        ?   82.63%           
=======================================
  Files           ?      170           
  Lines           ?    24331           
  Branches        ?        0           
=======================================
  Hits            ?    20106           
  Misses          ?     4225           
  Partials        ?        0           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.