Android: going back re-attaches and redraws the whole covered screen (hide instead of remove halves the worst back frame)

[enzomanuelmangano]

Note

Written by Claude Opus 5.5 on behalf of Enzo (@enzomanuelmangano).

The problem

On Android, going back in a native stack drops frames when the screen underneath is heavy. When a screen is pushed, ScreenStack removes the fragment of the screen it covers (transaction.remove, android/.../legacy/ScreenStack.kt). Its views leave the window and Android drops their display lists. Going back adds the fragment again, so the first frame of the pop re-attaches and redraws the whole screen underneath.

Measurements

Reproduction: https://github.com/enzomanuelmangano/repro-lab/tree/main/demos/screens-android-back-reattach. The covered screen has ~200 translucent views and a Skia canvas, and the pushed screen is light. Redmi Note 8 Pro, release build, 8 pops per run, two interleaved runs per build:

	4.30.0	covered screen hidden instead of removed
worst frame of each pop	47–74 ms (median ~56)	22–35 ms (median ~27)
frames over 50 ms per run	7, 8	0, 0
PSS after the runs	~235 MB	~216 MB

Before / after

Android's "Profile GPU rendering" bars right after the back. Each bar is one frame, and the green line is 16 ms. On the left, the tall blue/green bars are the covered screen being re-recorded and redrawn during the pop.

Video of the back, 3× slower: https://github.com/enzomanuelmangano/repro-lab/blob/main/media/screens-android-back-reattach.mp4

The idea

Hide the covered fragment (transaction.hide) instead of removing it, and show it again on the way back. Preloaded (INACTIVE) screens and fragments that were never added are still removed. The change is about 30 lines in ScreenStack.kt, plus two lines in ScreenFragment/ScreensAnimation. A hidden fragment stays resumed, so those two places treat isHidden as disappearing.
Diff against 4.30.0: enzomanuelmangano/repro-lab@main...fix/screens-android-back-reattach

Trade-offs I know of:

• The covered screen's views stay attached. Its GPU surfaces (Skia, WebGPU, video) and running animations are kept while it's covered. That is the point for the redraw, and it costs memory on deep stacks (although PSS went down in this test).
• Lifecycle: a hidden fragment does not get onPause. The appear/disappear events are kept correct by the isHidden checks above, but code that relies on fragment lifecycle callbacks would see a difference.
• It needs checking with transparent modals, deep stacks popped to root, and process restore.

We've shipped this as a patch in a production app (v4 stack) for a few days. Given the focus on the v5 stack: is this something you'd consider for v4, maybe behind an option? Or does the v5 Android stack already avoid the re-attach? If it's of interest, I'm happy to open a PR in whatever shape you prefer.