Display-size layout collapse

Font scale is only half the accessibility geometry. Display size (density) is a separate slider, and users who raise one usually raise both. The failure is a control that is not merely ugly but absent.

Companion to android-font-scale-verification, which covers the text-scale axis and the sentinel-string method. This one covers density and the layout traps that appear when space genuinely runs out.

Drive both axes from adb

adb shell wm density 540            # stock is often 420; 540 ≈ largest "Display size"
adb shell settings put system font_scale 2.0
adb shell wm size                   # px never changes; only dp does
# restore
adb shell wm density 420 && adb shell settings put system font_scale 1.0

Usable width in dp = px / (density / 160). 1080px at 540 → 320dp, the narrowest configuration a real user can produce on a phone.

Force-stop and relaunch after every change; a configuration change mid-process can leave a stale layout.

The test matrix that matters

Three points, not a grid: stock, stock + 2.0x text, largest display + 2.0x text. The third is where things break, and each fix must be re-checked at the first so the normal case is untouched.

Trap 1 — screenHeightDp never fires as a threshold

The largest display size still reports ~693dp of height, because raising density shrinks everything, including the elements you are trying to fit. A threshold like screenHeightDp >= 560 silently never triggers.

What actually overflows is text lines, so divide:

val room = windowHeightDp / LocalDensity.current.fontScale
// stock 891 · largest display + 2x text 347 → threshold 400 fires only there

In an app that pins its text scale (android-pin-single-font-scale), fontScale never changes: divide by the pinned constant instead (Ironvellum uses FIXED_FONT_SCALE).

Derive the threshold from three real measurements; do not guess a round number.

Trap 2 — measure the window, not the display

LocalConfiguration.current.screenHeightDp describes the whole display, so it is wrong in split screen, and lint flags it (ConfigurationScreenWidthHeight).

val density = LocalDensity.current
val windowHeightDp = with(density) { LocalWindowInfo.current.containerSize.height.toDp().value }

Take the lint hint here rather than suppressing it — the replacement is more correct, not merely quieter.

Trap 3 — scrolling the page collapses a weighted panel

The instinct when content overflows is to wrap the screen in verticalScroll. If any child uses Modifier.weight(...), that child now measures against an infinite height constraint and collapses to nothing — usually the very panel holding the missing control. Scrolling and weighting are mutually exclusive in one column.

Trap 4 — unweighted children win the measure pass

In a Column, unweighted children are measured first, in declaration order. A header declared before the button takes all the height it wants, and a button declared last is measured at zero height: present in the tree, zero pixels, unreachable.

Fix by making the thing that must survive the only unweighted child:

Column(Modifier.weight(1f, fill = false).verticalScroll(rememberScrollState())) {
    Title(); Note(); Rows()      // everything that may shrink
}
PrimaryButton()                   // unweighted → measured first → always laid out

When space is genuinely gone, choose what yields

If the screen still cannot fit, drop the largest purely informative element and say so in a comment: which element, why it was picked, and where that information still exists. Never let the primary action be the thing that goes. Surface the tradeoff as an owner decision with costed options rather than choosing silently.

Trap 5 — the calendar breaks instrumented tests, not the app

A flow test that drives "today's" content fails on the days your seed data does not cover — three mornings a year, on a machine nobody is watching. Walk the UI to a day that has content instead of trusting the clock:

if (allText().any { it == "REST DAY" }) { /* pick a day from the week rail */ }

Same shape for anything keyed on today: streaks, weekly rails, "this month".

Verification

Assert the control by sentinel string plus bounds at each of the three configurations, and on a scrolling screen confirm reachability by swiping before calling absence a defect. Then mutation-test: rebuild the pre-fix layout, confirm the control disappears, restore. A layout fix nobody watched fail is not proven.