On, unlocked and showing your app are three different things -- an injected gesture drives whatever is actually on top, and every frame statistic then describes that

scope: generic · severity: trap · confidence: proven · subsystem: display

taimen 2026-09-09. Three scroll arms in a row reported ‘NOT MEASURED: 102 frames in 6.2s. The screen never changed’ and scrollY 0 -> 0 with a healthy Epiphany showing the article – grim showed phosh’s lockscreen (‘Slide up to unlock’) on top. org.gnome.ScreenSaver.SetActive(false) had returned success and the output WAS on. ‘sudo loginctl unlock-sessions’ dismissed it; the same arm then scrolled 0 -> 17614 and reported 638 client commits. org.gnome.desktop.session idle-delay is 60 s on this image, so any arm whose page load takes a minute starts locked. Separately, and after that was fixed, four more arms reported the same void with the screen ON and UNLOCKED: grim showed phosh’s app grid over the browser, wlrctl toplevel focus did not dismiss it, and one injected KEY_ESC did.

Symptom — a gesture arm that runs, produces a full 60 Hz of frames, and moves the page zero pixels:

drag         on org.gnome.Epiphany
  NOT MEASURED: 102 frames in 6.2s. The screen never changed during the
  run -- the gesture hit nothing that animates.
[arm] scrollY 0 -> 0

The frames are real. They are phosh’s, animating its own lockscreen while the injected drag swipes at “Slide up to unlock”. The browser is up, mapped, focused and one surface underneath, and it commits nothing all run.

Why unblank is not enough — an-injected-touch-does-not-wake-a-blanked-screen covers the output being off; org.gnome.ScreenSaver.SetActive(false) fixes that, returns success, and leaves the session locked. The screensaver interface has no Unlock. The lever is logind:

sudo -n loginctl unlock-sessions

ph-ui.py unblank now does both and reports failure if either half refused, so anything routed through ph-gesture-bench.py is covered.

Why it bites unattended arms particularly — org.gnome.desktop.session idle-delay is 60 s on this image. Any arm that waits for a page to load, or for a build, or for a video to buffer, spends longer than that with no input and starts its measurement locked. It does not matter that the arm unblanked at the top: the lock arrives during the wait.

The instrument that finds it in one shot — grim -t png /tmp/pre.png immediately before the gesture, and look at it. Three arms were spent reasoning about touch injection, focus and coordinates for a symptom a single screenshot named instantly. A gesture arm should capture the frame it is about to drive.

The second surface: phosh’s overview

Fixing the lock exposed the same failure one layer up. phosh’s overview and app grid are shell layer surfaces: they sit over every toplevel, and activating a toplevel does not dismiss them. The session lands there whenever the last window closes – which is the first thing most arms do, since they stop the app under test before relaunching it with the environment they want to measure.

So the sequence every arm runs is exactly the one that breaks it:

stop the app        -> no toplevels left, phosh shows the overview
relaunch it         -> a window maps UNDER the overview
unblank, unlock     -> both succeed, and change nothing
drag                -> six flings into the app grid

and the arm reports 396 frames at 60 Hz, zero dropped, and scrollY 0 -> 0. wlrctl toplevel focus app_id:... does not help: the toplevel is activated and still covered. One KEY_ESC does, immediately.

ph-ui.py unblank() now does all three – on, unlocked, overview dismissed – and ph-key.py esc is the verb. Everything routed through ph-gesture-bench.py gets it, and that tool additionally raises the app when exactly one is open.

The instrument that finds any of these in one shot is a screenshot of the frame you are about to drive. Four arms and two sessions have now been spent reasoning about touch injection, coordinates and focus for symptoms a single grim named instantly. tools/repro/scroll-blank/ writes /tmp/pre-<label>.png before every drag for exactly this reason; look at it first when an arm goes void.