[Xget7]

Summary

One command records a release benchmark run under Perfetto on a phone and summarizes it, and an agent skill says how to read and compare runs.

scripts/android_perf.py [--seconds N] [--quality P] [--pose x,y,z,yaw,pitch]
  gradle :app:assembleRelease, adb install       # profileable release, timings valid
  perfetto --background                          # sched, cpufreq, thermal, memory, app slices
  am start --ez benchmark true --ez hud false    # the dev app's fixed-pose turn
  wait for "benchmark gpu ms:" in logcat
  pull trace -> output/android-perf/<time>/{trace.perfetto-trace, logcat.txt, report.json}
  scripts/perfetto_analysis.py --pid <app>       # in build/perf-venv; perfetto==0.57.2
    window = "SplatKit benchmark" async slice
    render-thread slice distributions, submit pacing and hitches,
    thread states per CPU cluster, cluster clocks, RSS, load times

 apps/android-dev/app/src/main/AndroidManifest.xml
+  <profileable android:shell="true" />          # release keeps the library's trace sections
 .claude/skills/android-perf/SKILL.md             # record, read, compare; the noise-floor rule
 .gitignore
-  .claude/
+  .claude/*
+  !.claude/skills/

Evidence

Physical Xiaomi Mi 9 (Adreno 640, Vulkan 1.1.128, Android 11), release build, --seconds 20.

• Before: a release build showed no app slices, and a manual run took several adb and Perfetto UI steps.
  After: python3 scripts/android_perf.py --seconds 20 ends with:

  status passed; app slices in trace: True
  benchmark window 19.9 s, 905 frames submitted (45.4 fps)
    SplatKit frame                 mean  12.79 ms  p95  24.92
    SplatKit submit and present    mean  10.12 ms  p95  21.91
  pacing: interval p50 16.9 ms, p95 34.4, max 66.8, 283 hitches over 1.5x median
  SplatKitRender: running 14% (cpus 4-6 9%, cpus 7-7 5%); S 84%
  

  The submit-fps reading agrees with the app's own 45.2 fps.
  The run is GPU-bound (21.8 ms against a 16.7 ms budget), so pacing alternates between one and two display periods.
• One run named the app <pre-initialized> because Perfetto read its /proc entry before the zygote rename, and matching by package name found nothing.
  The analysis now matches the pid the recorder passes, and an earlier trace still analyzes identically by name.
• python3 -m unittest discover -s scripts/tests: 45 tests pass, including the new config, launch, percentile, cluster, pacing and breakdown tests.

Merge Danger

Door: two-way

Blast Radius: tooling

The library is unchanged; the dev app's release build becomes profileable from a shell, which the dev app never ships.
The .gitignore change starts tracking .claude/skills/, while the rest of .claude/ stays ignored.
The first analysis downloads trace_processor through the pinned perfetto package.