`minSdk` is 33, because the branches below it cost more than the devices they reach

Context

minSdk 26 was inherited from the template, not chosen for Gloam. It looked free. It is not, and the reason is the test rig rather than the code.

There is one device in the loop — a Redmi amethyst on HyperOS, Android 16, API 36 — and no emulator on the machine at all. That device sits at the top of the supported range, so every platform difference the app depends on lives on the unvalidated side of it:

ShadeService branches three ways on layoutInDisplayCutoutMode — ALWAYS at 30+, SHORT_EDGES at 28+, nothing below. The bottom branch has never executed anywhere.
POST_NOTIFICATIONS is a runtime ask on 33+ and install-granted below it. The shade’s ongoing notification is the documented way out of a very dark screen, so the app’s central safety property has two behaviours and the phone can only show one of them.
TYPE_APPLICATION_OVERLAY on API 26–27 predates three rounds of overlay tightening.

Walking up the ladder, what each step actually deletes from this app:

`minSdk`	What it removes	Devices reachable
26	—	~99%
30	The three-way cutout branch collapses to one line; `<queries>` behaves uniformly	~85–90%
33	`POST_NOTIFICATIONS` becomes one code path; ADR-0004’s below-13 locale backport becomes dead weight	~70%
34	Foreground-service types uniform — already declared, so nothing	~55%
35, 36	Nothing this app uses	~24%

Android 16 was at ~24% of devices in August 2026 (AppBrain; StatCounter agreed within a point).

Decision

minSdk 33. The phone at API 36 stays the primary target and the only place nits are measured; one API-33 AVD covers the rest of the range, as a “does it launch, does the window appear, does the permission flow work” pass at the end of each phase rather than as a test suite.

Alternatives

Hold at 26. Reaches ~99% of devices and keeps every branch. It loses because those branches are not merely untested — they are untestable in this setup, and shipping a safety property whose alternate path has never run once is worse than not reaching the phones it was for.

minSdk 36. Considered because it would collapse the matrix to the one device that exists. It loses badly: it costs roughly three phones in four and buys nothing over 33 that this app uses. Paying 46 points of reach for zero simplification is the worst trade on the table.

minSdk 30. The honest runner-up — ~85–90% reach, and it does kill the cutout branch. It loses because the branch that matters is the notification permission, and 30 leaves that one in place.

Consequences

Roughly 30% of Android devices cannot install Gloam. That is the price, stated plainly.
The direction is asymmetric and this is why the decision is made now: raising minSdk after launch strands existing installs on the last build that fitted them; lowering it later is free. It is one of the few decisions here that is genuinely cheaper to make before users than with them.
The layoutInDisplayCutoutMode branches below R are dead and are removed with this change.
ADR-0004’s below-13 locale backport is now dead weight — the disabled AppLocalesMetadataHolderService manifest entry exists only for devices this app no longer ships to. It is not removed here: that is ADR-0004’s territory and deserves its own reading. Noted in DOD.md so it cannot be forgotten.
AppCompat itself stays regardless: ADR-0006 needs setDefaultNightMode to move the configuration so the window background and the system bars agree.

Amendment, 2026-10-31: `minSdk` is 26 again

The owner’s call, on reach: roughly 30% of phones could not install a free app whose whole mechanism is one overlay window. Lowering is the cheap direction this ADR named — nothing installed is stranded — so the question was only whether the branches below 33 could be made testable, and they now can be:

The CI floor leg is API 26 (default image; aosp_atd starts at 30), so the instrumented window tests run on the oldest platform the app ships to on every pull request, and an API-26 AVD boots locally where the API-33 one segfaults.
The branches are gathered rather than scattered. Every window-geometry difference below Android 11 lives in shade/OverlayLayout.kt, one function per question; the windows carry no SDK_INT checks of their own. Lint’s NewApi at minSdk 26 found everything else, and each is a one-line fallback beside its call.

What older phones get, by level — the platform’s limits rather than choices of ours:

Below	What is different
13	No Add the tile button — `requestAddTileService` does not exist; the row says to use the tile editor. The language choice is kept by AppCompat (ADR-0004), and reaches activities only, so the notification and tile follow the system language. `POST_NOTIFICATIONS` is install-granted
12	The platform does not clamp the shade window’s alpha (measured: 1.0 on API 26), so the window asks for `WINDOW_ALPHA_CLAMP` itself and the shade is exactly as dark as on 12+: ADR-0015’s ceiling holds everywhere. The fused provider is asked for by name, since its constant is not public
11	The shade is an explicit square on the display’s longest edge rather than `MATCH_PARENT`, which the platform keeps clear of the navigation bar for an overlay (measured: 1794 of 1920 px). With the bar on the left that also moves the overlays’ origin past it, so the shade starts a bar’s thickness further left and the bubble’s area starts at 0 (measured: the shade from 126, a tucked bubble off the display). The panel and prompt drop `FLAG_LAYOUT_IN_SCREEN` to sit clear of the bars. Display size and the bubble’s area come from `Display` reads. The backlight override is an integer out of 255, so the floor is one step of it rather than the panel’s (`MIN_INT_BACKLIGHT`); `MIN_BACKLIGHT` would truncate to 0, which 9 and 10 read as no override. Backup reads `@xml/backup_rules`, the twin of the extraction rules
10	The panel reports itself on screen from a post after pre-draw, which can be one frame early
9	No cutout API: a notch strip may stay undimmed. The backlight floor configs do not exist, so the floor is read as unset. The fused provider reports itself disabled to apps, so location comes from the network provider. `canDrawOverlays` reads stale for about a second after the grant, so the ask re-reads while it is on screen. The backlight override is clamped up to the setting’s minimum (measured: 10 of 255)

The table in Context is kept as the record of the argument it was; the cost it priced was untestability, and that is what changed.

minSdk is 33, because the branches below it cost more than the devices they reach