Android: snapshots ~10× slower while a looping Reanimated opacity animation is on screen (follow-up to #3558) #3832

Follow-up to #3558, split out as asked. #3611 (0.21.28) fixed the missing-window capture failure. This issue tracks the remaining latency: Android snapshots are about 10× slower while one small looping animation is on screen.

Setup

• Screen: a full-height @lodev09/react-native-true-sheet 4.0.0-beta.24 sheet (on Android, a separate dialog window). Its header holds an 8×8 Animated.View whose opacity loops forever with Reanimated 4.5.4: withRepeat(withSequence(withTiming(0.4, 800 ms), withTiming(1, 800 ms)), -1).
• React Native 0.86.3 (New Architecture), Expo SDK 57, release build.
• Emulator sdk_gphone64_arm64, Android 16 (API 36), 1080×2400. Host: Apple M5 Max, macOS, shared with other emulators and builds.
• "Animations off" means window_animation_scale, transition_animation_scale and animator_duration_scale set to 0 before launch. Reanimated reads that as Reduce Motion and holds the dot still. Nothing else on screen changes.

0.21.28, 2026-10-17 (the table requested in #3558)

Same session, same sheet. Host load average was 170–240 for both rounds.

On the slow captures, the daemon's durationMs matched the wall time (4.1–5.3 s on the ones checked).

0.21.28 vs 0.21.29-nightly.20261019.75 (includes #3790), 2026-10-19

Same app build, same emulator, one session. Each block is 7 snapshot --json calls; the first call is dropped because it reinstalls the shared helper APK. Blocks run in ABAB order. The numbers are the helper's data.androidSnapshot.elapsedMs. Every capture succeeded and contained the sheet root.

What this shows on our screen:

• perf(android): avoid animation stalls in capture and viewport reads #3790 makes the static capture about 4× faster (164 → 40 ms).
• With the loop running, it changes little at moderate load: 676 → 536 ms at load 29–41, and 443 → 439 ms at load 16–27.
• The loop's cost on the nightly is about 11× (40 ms → 440–540 ms), with an occasional multi-second capture (6.7 s once, as load rose).
• Host load scales the slowdown. The same 0.21.28 measured about 0.5–1.2 s here and 4–15 s at load 170–240.

So the remaining cost seems to come from the animation itself rather than traversal order. That fits your note that Android serves cache misses on the UI thread, which a running animation keeps busy.

Impact and workaround

Anything that snapshots repeatedly (wait, press after a settle, replays) pays this on every capture while a pulsing "live" indicator is on screen. That's common in timers, workouts and live-status UIs. Our workaround is animation scales at 0 on the test emulator, as Espresso suites do. That isn't available when the animation is the thing under test.

Happy to run a build against this screen. I can attach the request logs (requests/*.ndjson) for any of the runs above, or put together a minimal repro app (one TrueSheet sheet with the withRepeat opacity loop).