fix: bell mode regression, plus adversarial review and dark-mode follow-ups

[mmazzarolo]

Follow-up to #137, which merged while three adversarial reviewers were still running. They found two problems in that merge, so master currently carries a user-facing regression. This branch fixes it.

bun run validate:static passes (81 tests). The full Maestro suite passes on both platforms against Release builds — Android with R8, iOS re-run after the startup change. The iOS dark-mode fix was verified on device by reproducing the exact failing configuration.

The regression in master

In origin/master today:

bell: { breatheIn: cueBell1, breatheOut: cueBell2, hold: cueBell1 }

buildStepsMetadata gives the audio id hold to both the step after the inhale and the step after the exhale, so any bell assigned to it sounds twice per cycle:

Pattern	Before #137	On master now
Square (the default)	bell1 bell2 bell1 bell2	bell1 bell1 bell2 bell1
Pranayama, custom default	alternating	three identical bells across the loop
Awake, Ujjayi	four identical	fixed

#137 fixed the patterns with no holds and broke the ones with holds — including the default — and added a test asserting the broken mapping.

Bell mode now cues only the two direction changes and leaves the holds silent, so every pattern alternates. Cue players are built independently rather than all-or-nothing, which is what lets a mode cue some steps and not others. The test walks a full Square cycle instead of checking the asset map.

Native views ignored the chosen theme

Reported from a screenshot: with "Use system theme" off and Dark chosen on a light phone, the iOS large title rendered near-black on the app's dark background, and the picker wheel was drawn in light-mode chrome.

Both are UIKit, and UIKit follows the system appearance. Rather than colour each native view by hand, the app now overrides its own appearance with Appearance.setColorScheme, so every native view follows the chosen theme — including any added later. "unspecified" hands control back when the user is following the system.

Worth flagging for reviewers: colors.test.ts asserts the contrast of colours React Native draws and is structurally blind to anything UIKit paints. Three reviewers and I all missed this; it took looking at the screen.

{ stopTogether: false } was doing no work, and cost the loop its failure signal

The reviewer rebuilt the composition without the flag and ran every reachable step length — all reach their target and report finished. The 0.35 ratio alone fixes the short-step bug.

What the flag added was a silent failure: an interrupted circle reported finished: true, so loopAnimations would step on forever with a frozen circle instead of stopping. Removed. The 25 exercise tests still pass, which is the proof it was redundant.

iOS audio interruptions

The answer to "does pausing only on background mean the app talks over a phone call?" is no — expo-audio deactivates the session. But two real consequences fell out:

• during an interruption the exercise, visuals and haptics keep running with no audio at all
• afterwards expo-audio replays the cue it interrupted, so a fragment of "Breathe in" can land mid-exhale — and the same applies on backgrounding, because the JS AppState event arrives after expo-audio's native observers, making the existing stopExerciseAudio() too late

Audio now also stops on the return to active, which covers both. Still uncovered: an interruption that begins and ends without an AppState transition. expo-audio exposes no JS event for it, so that is a known gap rather than an oversight.

Smaller review findings

• customPatternDurationLimits is a four-tuple. normalizePersistedSettingsState maps over it, so its length is what makes the as CustomPatternSteps cast true — as a plain array it was simply untrue, and a fifth entry would have produced a five-element result still typed as four.
• The settings read retried in the same tick, so it could only survive a bridge hiccup, not the locked database its own comment claimed. It pauses first now.
• animate() coalesces its easing, so an explicit undefined no longer falls through to React Native's default curve.
• The accessibility hooks ask at module load rather than first mount, and a change event landing mid-query is no longer overwritten by the stale answer — which, cached at module scope, used to poison every later mount in the process.
• The paused screen announces itself and carries a header role. It was silent to screen readers while the completion screen already announced.
• useKeepAwake no longer covers the completion screen, which never dismisses itself and held the display on until the user came back.
• The iOS settings secondary text used raw slate-500 — 3.07:1 on the dark card, the longest body text in the app.
• hideBottomBorder → hideBottomBorderWeb: only web honours it, and the platform-neutral name promised parity a TypeScript contract cannot enforce.

Known gaps, deliberately not fixed

• Android's stepper buttons still have no accessible name. The current @expo/ui API cannot express it: semantics takes only contentType, FilledTonalButton exposes no content description, and Icon needs a vector drawable. Documented at the call site.
• dark.image does nothing on Android — the plugin only writes a dark drawable when drawable.darkIcon is set. No visible fault, because that splash icon is transparent and the dark colour does apply.
• splash-dark.png ships ~28 KB into the web bundle, where SplashScreenManager returns early and never renders it.