The Generated Screen That Quietly Jams at AX5 and in German — Putting Layout Resilience Checks on Rork Max SwiftUI

What 23 screens did across 4 conditions

I narrowed things to four conditions and walked all 23 screens by hand. Testing every combination does not fit in a workable afternoon, so I kept only the combinations where breakage was plausible.

Here is what came back.

Condition D turned up problems on 9 of 23 screens, four of them genuinely unusable. Set against 23/23 green at the baseline, it is a fairly blunt demonstration of how little a default-settings pass actually guarantees.

The failures sorted into three types.

The two "no response" cases only surfaced under condition B. Nothing is broken, so they read as green — but to someone running large text, that screen is simply the one part of the app that ignores them. Harder to notice than a break, and for that reason the one I find more troubling.

I fixed those two first, ahead of the other seven. Push-out is visible, so somebody might eventually tell you about it. Nobody tells you about no-response.

Stopping push-out: layoutPriority and ViewThatFits

Starting with the most common failure. The generated settings row looked like this.

// As generated. Correct at the defaults; the switch leaves the screen at AX5.
struct SettingRow: View {
    let title: String
    @Binding var isOn: Bool
 
    var body: some View {
        HStack {
            Text(title)
            Spacer()
            Toggle("", isOn: $isOn)
                .labelsHidden()
        }
        .padding(.horizontal, 16)
    }
}

When an HStack runs short on horizontal space, it compresses its children. Text resists compression, so Toggle loses first. Spacer() merely fills what is left and never enters the tug-of-war. The switch gets pushed out.

The fix came in two parts. First, make the control's priority explicit.

struct SettingRow: View {
    let title: String
    @Binding var isOn: Bool
 
    var body: some View {
        // Give up on the horizontal arrangement rather than break it
        ViewThatFits(in: .horizontal) {
            horizontalLayout
            verticalLayout
        }
        .padding(.horizontal, 16)
        .padding(.vertical, 8)
    }
 
    private var horizontalLayout: some View {
        HStack(spacing: 12) {
            Text(title)
                .layoutPriority(0)          // the label yields
            Spacer(minLength: 8)
            Toggle("", isOn: $isOn)
                .labelsHidden()
                .layoutPriority(1)          // the control holds
                .fixedSize()                // refuse compression outright
        }
    }
 
    private var verticalLayout: some View {
        VStack(alignment: .leading, spacing: 8) {
            Text(title)
                .fixedSize(horizontal: false, vertical: true)
            Toggle("", isOn: $isOn)
                .labelsHidden()
        }
        .frame(maxWidth: .infinity, alignment: .leading)
    }
}

ViewThatFits tries each candidate in order and takes the first one that fits the space it is offered. If the horizontal arrangement will not fit, it drops silently to the stacked version. It holds up better than writing the branch yourself because the decision is made on whether the thing actually fits. Pick a threshold like typeSize > .accessibility1 and that threshold will be wrong the moment the language or the device width changes.

.fixedSize() on the Toggle is what really lands. Without it, raising layoutPriority still leaves the Toggle room to shrink toward its minimum. .fixedSize() declares "I do not go below my ideal size," and that ends the tug-of-war.

Spacer(minLength: 8) is deliberate — Spacer()'s default minimum is not zero. Sometimes you want it crushed to 0 in tight spaces; here I kept 8pt so the two elements stay distinguishable.

Stopping truncation: deciding to abandon fixed widths

For the three fixed-width buttons, the fix was easy and the line-drawing was not.

// As generated
Button("Save Settings") { save() }
    .frame(width: 120, height: 44)
    .background(Color.accentColor)
    .foregroundStyle(.white)
    .clipShape(RoundedRectangle(cornerRadius: 8))

Under Double-Length, "Save SettingsSave Settings" becomes Save Setti.... Nobody can tell what that button does.

Simply dropping the width gives you a button that spans the whole screen, which I wanted to avoid on design grounds. This is where it landed.

Button {
    save()
} label: {
    Text("settings.save")             // String Catalog key
        .lineLimit(2)                  // two lines are allowed
        .multilineTextAlignment(.center)
        .minimumScaleFactor(0.8)       // shrink no further than 80%
        .frame(minWidth: 120)          // keep only the floor
        .frame(maxWidth: .infinity)
        .padding(.vertical, 12)
        .padding(.horizontal, 16)
}
.buttonStyle(.borderedProminent)
.frame(maxWidth: 320)                  // a ceiling to stop it sprawling

Put into words, the reasoning went:

1. Keep only the width floor — minWidth still earns its place as a tap-target guarantee
2. Add a ceiling — maxWidth: 320 stops iPad stretching it across the pane
3. Allow two lines — insisting on one line leaves only truncation or unreadably small type
4. Cap shrinking at 80% — set minimumScaleFactor to something like 0.5 and unreadable text passes as "readable"

I learned the fourth one the hard way. With minimumScaleFactor(0.5), Double-Length text rendered at half size, the layout never broke, and it sailed through as green. Not breaking and being readable are different things. At 0.8, anything that will not fit properly wraps to two lines instead.

Stopping the no-response case: ScaledMetric and relative fonts

.font(.system(size: 15)) does not follow the text size setting. A text style like .font(.body) does, but generated code tends to write absolute values when the prompt mentions something concrete like "15pt."

// Does not follow
Text(label).font(.system(size: 15))
 
// Follows (relative, text-style based)
Text(label).font(.system(.subheadline, design: .rounded))
 
// Anchored at 15pt but still following, with a ceiling
Text(label).font(.system(size: 15, weight: .medium))
    .dynamicTypeSize(...DynamicTypeSize.accessibility3)

What mattered more was making the non-text dimensions follow too. Leave icons and padding fixed and the text alone balloons, which breaks things in an uglier way.

struct StatCard: View {
    // Declare a base size and it scales with the text size setting
    @ScaledMetric(relativeTo: .headline) private var iconSize: CGFloat = 28
    @ScaledMetric(relativeTo: .body) private var cardPadding: CGFloat = 16
    @ScaledMetric(relativeTo: .body) private var cornerRadius: CGFloat = 12
 
    let title: String
    let value: String
 
    var body: some View {
        HStack(spacing: cardPadding * 0.75) {
            Image(systemName: "chart.bar.fill")
                .font(.system(size: iconSize))
                .frame(width: iconSize, height: iconSize)
 
            VStack(alignment: .leading, spacing: 4) {
                Text(title)
                    .font(.subheadline)
                    .foregroundStyle(.secondary)
                    .fixedSize(horizontal: false, vertical: true)
                Text(value)
                    .font(.headline)
                    .fixedSize(horizontal: false, vertical: true)
            }
            Spacer(minLength: 0)
        }
        .padding(cardPadding)
        .background(.regularMaterial, in: RoundedRectangle(cornerRadius: cornerRadius))
    }
}

@ScaledMetric(relativeTo:) returns a value scaled in proportion to the given text style. At AX5 the icon grows from 28pt to roughly 62pt in practice, and because the padding grows alongside it, the proportions survive the enlargement.

.fixedSize(horizontal: false, vertical: true) on the Text refuses vertical clipping. Without it, HStack height math will sometimes flatten multi-line text into a single line. Choosing where to apply it is unglamorous work with an outsized payoff.

A summary of when to reach for what:

Catching what regeneration undoes, with snapshots

This next part is unavoidable once you are working with Rork Max.

Fixed layouts revert the moment you change a prompt and regenerate the screen. The model has no knowledge of the layoutPriority I added by hand. Regeneration is the whole point of Rork Max, so giving it up was never on the table. Where to draw the boundary is the subject of Regenerable Zones in Rork Max Code, but inside that boundary, "detect it when it breaks" is ultimately the only move available.

So I added snapshot checks.

import XCTest
import SwiftUI
@testable import MyApp
 
final class LayoutResilienceTests: XCTestCase {
 
    /// The four conditions. We deliberately do not run the full matrix.
    struct Condition {
        let name: String
        let locale: Locale
        let typeSize: DynamicTypeSize
    }
 
    static let conditions: [Condition] = [
        .init(name: "base",   locale: Locale(identifier: "en_US"),    typeSize: .large),
        .init(name: "ax5",    locale: Locale(identifier: "en_US"),    typeSize: .accessibility5),
        .init(name: "double", locale: Locale(identifier: "en_DOUBLE"), typeSize: .large),
        .init(name: "worst",  locale: Locale(identifier: "en_DOUBLE"), typeSize: .accessibility5),
    ]
 
    /// Judge overflow and clipping mechanically
    func assertNoOverflow<V: View>(
        _ view: V,
        named name: String,
        containerWidth: CGFloat = 393,      // iPhone 16 logical width
        file: StaticString = #filePath,
        line: UInt = #line
    ) {
        for condition in Self.conditions {
            let hosted = view
                .environment(\.locale, condition.locale)
                .environment(\.dynamicTypeSize, condition.typeSize)
                .frame(width: containerWidth)
 
            let renderer = ImageRenderer(content: hosted)
            renderer.proposedSize = ProposedViewSize(width: containerWidth, height: nil)
 
            guard let image = renderer.uiImage else {
                XCTFail("\(name)/\(condition.name): render failed", file: file, line: line)
                continue
            }
 
            // Wider than the container means horizontal push-out
            let renderedWidth = image.size.width
            XCTAssertLessThanOrEqual(
                renderedWidth, containerWidth + 0.5,
                "\(name)/\(condition.name): width \(renderedWidth)pt exceeds container \(containerWidth)pt",
                file: file, line: line
            )
 
            // Compare against the reference image (records on first run)
            assertSnapshot(image, named: "\(name)_\(condition.name)", file: file, line: line)
        }
    }
 
    func testSettingRow() {
        assertNoOverflow(
            SettingRow(title: "settings.background_refresh", isOn: .constant(true)),
            named: "SettingRow"
        )
    }
 
    func testStatCard() {
        assertNoOverflow(
            StatCard(title: "stats.total_views", value: "128,400"),
            named: "StatCard"
        )
    }
}

Passing nil as the height in proposedSize is the key detail. Fix the height and content that ought to grow vertically gets crammed in, and the overflow stops being detectable. Fix only the width, let the height run free, and you get back the height actually required — with any width excess falling out as a plain number.

The en_DOUBLE pseudolocale is defined on the test target. Xcode's scheme setting has no effect at test runtime, so I ship a .strings file holding the doubled values.

Worth noting the CI time. My first version ran every screen × all 12 DynamicTypeSize steps × 3 locales, and took 11 minutes 12 seconds. That is not a wait you accept on every run.

The case for narrowing came out of the measured distribution.

Running all 12 steps found nothing the 4 conditions missed. Things that break, break at the worst case. Cutting to 2 conditions, though, dropped two cases of the "shows up at AX5 alone but is masked by a different failure in the combined case" variety. Four conditions is where I settled.

The push-out that image width will never catch

Here is one I missed, written down.

Right after the snapshot checks went in, I was pleased with myself. Width overflow now fell out as a number, so push-out could not slip past. Except — the very settings screen that started all of this passed assertNoOverflow green.

The reason is plain enough. ImageRenderer renders at the width you propose. With a parent carrying .frame(width: 393), a child shoved beyond that edge is simply not drawn, and the image comes back 393pt wide. So the check only ever caught push-out where the parent itself grew. Push-out that gets clipped away looks, numerically, like a perfect fit.

Recounting the nine condition-D cases by failure shape:

The reference diff catches all of it. What a diff tells you, though, stops at "something changed" — whether that something is the fix or the breakage is left for you to judge by eye. Re-record the references after every regeneration and you buy that judgement call over and over. The only way to spend less of it is to give each failure shape a name and a condition.

Count whether a thumb can reach it, not whether it looks present. That, in one line, is what the original screen was telling me.

So I built a marker that reports the rectangle of anything tappable.

import SwiftUI
 
struct ControlFrame: Equatable {
    let id: String
    let rect: CGRect
}
 
struct ControlFrameKey: PreferenceKey {
    static var defaultValue: [ControlFrame] = []
    static func reduce(value: inout [ControlFrame], nextValue: () -> [ControlFrame]) {
        value.append(contentsOf: nextValue())
    }
}
 
extension View {
    /// Attach only to tappable elements; reports its rect in the probe coordinate space
    func trackHitArea(_ id: String) -> some View {
        background(
            GeometryReader { proxy in
                Color.clear.preference(
                    key: ControlFrameKey.self,
                    value: [ControlFrame(id: id, rect: proxy.frame(in: .named("probe")))]
                )
            }
        )
    }
}

Only Toggles, Buttons and tappable rows get the marker — 61 sites across 23 screens. Measuring decorative views as well would grow the numbers without growing what you can decide from them.

On the measuring side, run a layout pass through a UIHostingController and collect what comes back. ImageRenderer hands you an image but leaves the preference values out of reach, so this one step goes through hosting.

import SwiftUI
import UIKit
 
@MainActor
func measureHitAreas<V: View>(
    of view: V,
    locale: Locale,
    typeSize: DynamicTypeSize,
    containerWidth: CGFloat
) -> [ControlFrame] {
    final class Box { var frames: [ControlFrame] = [] }
    let box = Box()
 
    let probe = view
        .environment(\.locale, locale)
        .environment(\.dynamicTypeSize, typeSize)
        .frame(width: containerWidth)
        .coordinateSpace(name: "probe")
        .onPreferenceChange(ControlFrameKey.self) { box.frames = $0 }
 
    let host = UIHostingController(rootView: probe)
    let fitted = host.sizeThatFits(
        in: CGSize(width: containerWidth, height: .greatestFiniteMagnitude)
    )
    host.view.frame = CGRect(origin: .zero, size: fitted)
    host.view.setNeedsLayout()
    host.view.layoutIfNeeded()
 
    return box.frames
}

I kept the assertions down to three: nothing hangs off either edge, width is not under 44pt, height is not under 44pt.

@MainActor
func assertHitAreasReachable<V: View>(
    _ view: V,
    named name: String,
    containerWidth: CGFloat = 393,
    minimumSide: CGFloat = 44,
    file: StaticString = #filePath,
    line: UInt = #line
) {
    for condition in LayoutResilienceTests.conditions {
        let frames = measureHitAreas(
            of: view,
            locale: condition.locale,
            typeSize: condition.typeSize,
            containerWidth: containerWidth
        )
 
        // Do not let a forgotten marker pass as success
        XCTAssertFalse(
            frames.isEmpty,
            "\(name)/\(condition.name): trackHitArea reported nothing",
            file: file, line: line
        )
 
        for frame in frames {
            XCTAssertGreaterThanOrEqual(
                frame.rect.minX, -0.5,
                "\(name)/\(condition.name): \(frame.id) hangs \(-frame.rect.minX)pt off the left edge",
                file: file, line: line
            )
            XCTAssertLessThanOrEqual(
                frame.rect.maxX, containerWidth + 0.5,
                "\(name)/\(condition.name): \(frame.id) ends at \(frame.rect.maxX)pt, past the edge",
                file: file, line: line
            )
            XCTAssertGreaterThanOrEqual(
                frame.rect.width, minimumSide,
                "\(name)/\(condition.name): \(frame.id) is \(frame.rect.width)pt wide, under 44pt",
                file: file, line: line
            )
            XCTAssertGreaterThanOrEqual(
                frame.rect.height, minimumSide,
                "\(name)/\(condition.name): \(frame.id) is \(frame.rect.height)pt tall, under 44pt",
                file: file, line: line
            )
        }
    }
}

44pt is a floor and nothing more. On screens where @ScaledMetric drives the padding, the hit areas grow along with the text at AX5. Compare for equality and a correct expansion gets reported as a failure.

Run time came to +38 seconds on top of the 2m40s snapshot pass, for 3m18s total. For four layout passes over 61 sites, lighter than I had braced for.

One caveat. onPreferenceChange only fires after a layout pass, so drop sizeThatFits and layoutIfNeeded() and you get back zero rectangles. That is why the empty check sits first in the assertions above — I once spent a couple of days looking at a green test that was measuring nothing at all.

Changing the input: putting the constraints in the prompt

With detection in place, it was worth trying to have less broken code generated in the first place. I added this to the prompt.

Layout constraints:
- Must not become unusable at AX5 (310%) text size
- Label/control HStacks must fall back to a vertical stack via ViewThatFits
- Controls must carry layoutPriority(1) and fixedSize()
- Do not use absolute .font(.system(size:)); write against text styles
- Declare icon and padding dimensions with @ScaledMetric(relativeTo:)
- Do not use .frame(width:) on buttons; bracket with minWidth and maxWidth

The effect was, honestly, partial. Generating the same screen five times each way:

One-in-five to four-in-five is a large improvement, and it is not five in five. The more complex the screen, the more likely a couple of the constraints drop out. Which means the prompt is a tool for reducing checks, not a tool for replacing them.

There is a small disappointment in that conclusion. If instructions were honored every time, no tests would be needed. But asking a human to spot, by eye, the one run in five that missed is asking for something that will not happen. What a machine produces, a machine should verify. It feels like a reasonable place to have landed.

The order to adopt this in, and the one thing to do first

For anyone starting the same work, here is the sequence.

1. Run pseudolocalization once, by hand (about 30 minutes) — set the scheme's App Language to Double-Length and open every screen. You will come away with a feel for your app's weak spots
2. Do the same at AX5 (about 30 minutes) — Settings → Accessibility → Display & Text Size → Larger Text
3. List only the screens that jammed — do not try to fix everything; restrict yourself to unusable and meaning-lost cases
4. Apply the ViewThatFits + layoutPriority + fixedSize trio — this stops most push-out
5. Add snapshot checks to the fixed screens only — start with all of them and you will abandon the effort. Add trackHitArea measurements alongside them so clipped push-out does not slip through
6. Add the constraints to your prompt — this move only means anything once the checks exist

If you pick one, pick the first. Thirty minutes, not a line of code, and you know where your app actually stands. In my case those thirty minutes handed me nine screens' worth of homework. It was not an enjoyable half hour, but it beat hearing about it in a review.

An indie developer's hours are finite. Which is exactly why a check that surfaces nine issues in thirty minutes sits, for me, on the favourable side of the ledger.

The broader production accessibility work, VoiceOver included, is written up in Production-Quality VoiceOver and Dynamic Type for Rork Apps. That one is on the React Native side, but the thinking on label design carries over to Rork Max unchanged.

That screen at maximum text size has become the place that quietly tells me who I had been overlooking. Next time you open your app, it is worth touching that setting once.