feat(android): target Android 16 (API 36) and handle edge-to-edge

[nseldeib]

Why

Play reports the app's highest non-compliant target API level as Android 15 (API 35). From 2026-10-31, an app whose target API isn't within one year of the latest Android release cannot be updated at all — this is an availability deadline, not a warning.

The toolchain cascade

targetSdk 36 requires compileSdk 36, which the old AGP can't compile against, which forces Gradle up, which forces Paparazzi onto its 2.0.0 alpha line (1.3.5 crashes on compileSdk 36 — cashapp/paparazzi#1877, no 1.3.x backport).

	Before	After
compileSdk / targetSdk	35	36
AGP	8.3.1	8.13.2
Gradle	8.7	8.13
Kotlin	2.0.21	2.3.0
Paparazzi	1.3.5	2.0.0-alpha05
CI / release JDK	17	21

Kotlin wasn't in the original plan's matrix but had to move. Paparazzi alpha05 is compiled against Kotlin 2.3.0, so an older KGP can't read its metadata (the 2.1.20 attempt crashed the compiler outright). Worth reviewing with open eyes: this puts kotlin-stdlib:2.3.0 on releaseRuntimeClasspath. The alpha itself stays test-scope and never enters the AAB, but the compiler and stdlib moved with it. Kotlin 2.3 also made kotlinOptions { jvmTarget } a hard error, hence the compilerOptions DSL migration — the 1.8 bytecode target is deliberately unchanged.

Edge-to-edge

Android 16 makes edge-to-edge unconditional — the windowOptOutEdgeToEdgeEnforcement escape hatch is ignored at target 36. The app had no inset handling at all, and themes.xml was still setting statusBarColor / navigationBarColor, which have been deprecated no-ops since API 35. This was very likely already a live defect on Android 15.

MainActivity now opts in explicitly via enableEdgeToEdge (transparent scrim, light icons); the dead theme attrs are gone.

The padding is applied OUTSIDE BoxWithConstraints, on purpose. CounterScreen derives its whole layout from maxHeight/maxWidth, and CounterBottomBar sizes itself as screenHeight * 0.20f. Insetting inside would leave every derived measurement scaled to a window the content no longer occupies — correct on the device you tested, wrong on one with a taller nav bar.

Test

CounterScreenInsetsTest (4 tests, Robolectric, no emulator) pins that. Insets are injected via a defaulted contentInsets parameter because Robolectric has no real window, so platform insets are all zeroes and the behaviour would otherwise be untestable.

It's mutation-verified. Moving the padding inside BoxWithConstraints leaves every positional assertion passing — the base Column fills what it's given and weight(1f) absorbs the difference. Only the size assertion catches it:

Increment bar height (114.0.dp) must be derived from the USABLE height
(expected ~74.2.dp). Deriving it from the full window would give ~114.2.dp,
which means the inset padding was applied inside BoxWithConstraints.

The first version of this test passed under mutation and was rewritten.

Verification

• 298 tests green across both runners (121 Swift + 177 Android)
• verifyPaparazziDebug green
• bundleRelease → signed AAB, R8 enabled, mapping.txt present, merged manifest reads targetSdkVersion="36"
• Confirmed on a real API 36 device (sdk=36, release=16): status-bar icons light-on-dark with no grey band, header clears the status bar, control row clears the gesture pill, background runs bar to bar

Goldens were re-recorded: Paparazzi 2.0 wraps the canvas to the composable's measured height where 1.3.5 used a full device frame. The numerals themselves render identically — the diff was inspected before accepting.

Reviewer notes

• versionName 1.0 → 1.1. versionCode needs no change; play-release.yml derives it as 110000 + GITHUB_RUN_NUMBER and remains the sole uploader.
• play-release.yml also needed the JDK bump — it carried its own java-version: "17" with a comment that was already stale (it cited Gradle 8.4). Left alone, the Play artifact would build on a different JDK than the gates ran on.
• Suggest a staged rollout (rollout < 1.0) — this changes both the build toolchain and window handling on a live app.
• Predictive back is deliberately out of scope. It defaults on at target 36 and the app has no back handling, so back exits even with a panel open. That's a real UX gap but a separate behavioural feature needing its own scenarios; folding it into a deadline-driven compliance change would widen the diff.
• The 4 new tests are not in .codeyam/test-registry.json. refresh-tests drops the android-tests runner from the merged cache (298 pass, only 121 persist), which turns the reconcile audit into 174 phantom findings; registering them against the stale cache produces REGISTRY_TEST_NOT_WIRED, and the prescribed recovery matches zero tests for this runner even for already-registered ones. Bookkeeping debt in the codeyam tooling — the tests themselves run in CI. session-finalize is owed.

🤖 Generated with Claude Code