clockula/docs/ROADMAP.md

# Clockula — roadmap

The status view and the work queue. Design rationale for every decision here
lives in [`PLAN.md`](PLAN.md); the shape of the code as it stands will live in
[`ARCHITECTURE.md`](ARCHITECTURE.md) from M2 onward.

Status legend: ✅ done · 🚧 in progress · ⬜ not started

---

## Current state (one line)

✅ **M10 done.** Clockula now has a settings hub (theme, dynamic colour,
language, the alarm/timer/clock defaults every new alarm or timer inherits,
help), a documented, versioned JSON backup (`docs/BACKUP.md`) that exports and
restores alarms/timers/world clocks and settings through SAF (configuration
only, never ring state, never a timer's running anchors) as a whole-state
replace the screen is explicit about before the user confirms, a "why might my
alarm not ring?" self-check that reads real device state (exact-alarm and
notification permissions, full-screen intent, battery optimisation, DND, alarm
volume, a known aggressive-OEM allowlist, and the system's own next alarm
compared against what Clockula computed) and deep-links to the exact settings
page for each, and three Glance home-screen widgets (next alarm, timer,
clock) that mirror existing state and redraw event-driven, on every write,
with no ticker of their own. M11 (release readiness) is next.

---

## How the loop works through this

One milestone per iteration, auto-continuing. Each iteration:

1. Pick the first milestone not marked ✅.
2. Implement it completely: every checklist item under it.
3. Build, and run the tests the milestone names. A milestone is not done until
   they pass.
4. **Commit** (floret-kit submodule first where it changed, then Clockula's
   pointer bump; separate commits, granular, lowercase area prefix).
5. Mark the milestone ✅ here, update **Current state**, and commit that too.
6. Continue to the next milestone.

Rules that hold across every iteration:

- **Never let a Room entity or `@Query` string leak above the data layer**
  (`PLAN.md` §0).
- **Never** add a foojay toolchain resolver, anywhere, not even in a comment.
- Commit each finished unit of work; never accumulate a WIP pile.
- If a milestone turns out to be wrong, say so and amend `PLAN.md`. Do not
  silently build something else.
- `ARCHITECTURE.md` and `CHANGELOG.md` are updated as part of the milestone that
  changes them, not in a later cleanup pass.

---

## v1 milestones

### ✅ M0 — Skeleton & pipeline
The project exists, builds, is themed, and ships itself.
- Gradle project: `settings.gradle.kts`, version catalog copied from Agendula
  (drop `:provider`, add Room + KSP), AGP 9.x conventions, `vcsInfo` off on
  release, no foojay resolver.
- floret-kit as a git submodule + `includeBuild`, with the gitignored
  `floret-kit/local.properties` arrangement documented.
- Hilt, Compose, `MainActivity`, `FloretExpressiveTheme` reseeded to `#6B7A5C`
  (`ui/theme/Color.kt` + `Theme.kt`), hand-tuned light/dark schemes.
- `core-crash` wired; `core-locale` + `res/xml/locales_config.xml`.
- Codeberg repo, canonical from the first commit. Full pipeline per `PLAN.md` §10:
  `ci.yaml`, `translations.yaml`, `release.yaml` (Agendula spine + Calendula `play`
  job), `renovate.yml`, issue/PR templates, F-Droid metadata, fastlane tree.
- `README.md`, `LICENSE` (MIT), `CHANGELOG.md`, `.editorconfig`, `.gitattributes`,
  `.gitignore`, `docs/README.md` index.
- **Done when:** it builds, installs, shows a themed placeholder, and CI is green.

  ✅ `assembleDebug`, `lintDebug` and `testDebugUnitTest` all pass;
  `scripts/check_reproducible_release.sh` reports every invariant holding.
  Two notes for later: the store listing and `fastlane` changelogs are written
  but no screenshots exist yet (M11), and `docs/ARCHITECTURE.md` is still
  unwritten by design. M2 writes it, once there is an architecture to
  describe.

### ✅ M1 — floret-kit: `core-prefs` + `core-di`
Pay the extraction the kit's roadmap has been waiting for a third app to trigger.
- In the kit: `core-prefs` (`ThemeMode`, `dynamicColor`, typed DataStore wrapper,
  `toEnum()`) and `core-di` (`@IoDispatcher` + provider), each with tests, module
  docs, `CHANGELOG` and `ROADMAP`/`ARCHITECTURE` updates.
- Clockula consumes both. Migrating Calendula and Agendula is explicitly out of
  scope; note it in the kit's roadmap as a follow-up.
- **Done when:** the kit's tests pass, Clockula builds against the new modules,
  and both repos are committed (submodule first, then the pointer bump).

  ✅ floret-kit 0.4.0. `core-prefs` (56 kit tests) gives the family a typed
  `Pref<T>`/`PrefStore` over DataStore; reads degrade to the default instead of
  throwing, and `flow()` is distinct-until-changed. The appearance key names are
  frozen at what Calendula and Agendula already write, so the sibling migration
  (a follow-up in the kit's roadmap, not here) needs no data migration.
  `core-di` carries `hilt-compiler` itself, since an `@InstallIn` module only
  reaches a consumer's component if its own library was processed; Clockula's
  `assembleDebug` is the integration test for that binding. Clockula's theme now
  reads the stored preference, with the DataStore built on a corruption handler
  so an unreadable prefs file cannot brick startup.

### ✅ M2 — Data layer
The whole storage stack, headless. No UI.
- Room: `alarms`, `timers`, `world_clocks`, `stopwatch_laps` (`PLAN.md` §5),
  schema exported and committed, DAOs.
- Domain models (plain Kotlin) + mappers. Repositories exposing Flows.
- DataStore prefs (theme, dynamic colour, defaults, stopwatch running state).
- DI modules.
- **Tests:** mappers, repository behaviour, the elapsed-realtime vs wall-clock
  distinction under a simulated clock change.
- Write the first `ARCHITECTURE.md`.
- **Done when:** the schema is exported and committed, no Room type or query
  string appears above `data/`, and the build, lint and unit tests pass.

  ✅ Four tables (`alarms`, `timers`, `world_clocks`, `stopwatch_laps`), schema
  v1 exported to `app/schemas` so migrations can be tested from the beginning.
  The app talks to four repository interfaces speaking domain types and Flows,
  and `ArchitectureRulesTest` fails the build if Room, an entity or a query
  string ever appears above `data/`. Time is a parameter rather than an ambient
  read: `WallClock` and `ElapsedRealtimeClock` are distinct types, so a running
  timer anchors on the monotonic clock and keeps a wall-clock value only as a
  post-reboot fallback, and the stopwatch keeps none at all. That leaves one
  gap: once a new boot's uptime climbs past a stored anchor the reboot goes
  unnoticed, so closing it needs a persisted boot id and M3's `BOOT_COMPLETED`
  receiver. Every read-modify-write runs inside a DAO transaction, and the
  stopwatch run record is written whole in one DataStore edit, so concurrent
  writers cannot swallow each other's edits or persist a torn record. 151 JVM
  tests over fake DAOs and injected fake clocks; the migration test is the only
  thing needing a device.

### ✅ M3 — Alarm engine
The hard part (`PLAN.md` §4), still headless apart from the ring screen's
plumbing.
- Next-fire resolution: local time-of-day + repeat mask → next instant in the
  device zone; `skipNextOccurrence`; snooze.
- `AlarmScheduler` over `setAlarmClock`, single-next-alarm registration,
  `USE_EXACT_ALARM` with a `SCHEDULE_EXACT_ALARM` fallback path.
- Receivers: fire, `BOOT_COMPLETED`, `TIME_SET`, `TIMEZONE_CHANGED`,
  `MY_PACKAGE_REPLACED`.
- Ringing foreground service: audio focus, `USAGE_ALARM`, volume ramp, vibration,
  ringtone URIs; survives process death and reboot mid-ring.
- Full-screen-intent notification + `canUseFullScreenIntent()` handling, degrading
  to a high-priority heads-up notification that still rings, never to silence.
- **Tests:** DST spring-forward and fall-back, every repeat configuration, skip,
  snooze-vs-next-occurrence. These are the tests the app lives or dies on.

  ✅ 326 JVM tests, 172 of them new. The engine is a plain Kotlin class that
  reaches Android through four interfaces it does not implement, so the
  behaviour the app lives or dies on is pinned without an emulator.
  `ArchitectureRulesTest` fails the build if `android.*` ever appears in
  `domain/`, `AlarmEngine.kt` or `system/`. Occurrences are generated by walking
  local dates through `ZonedDateTime.of` rather than by adding 24 hours to an
  instant, which is the whole DST story: a 02:30 alarm rings at 03:30 on a
  spring-forward night instead of vanishing, and once rather than twice on a
  fall-back night. Asserted for Berlin and New York and across all 128 repeat
  masks. Two rules keep constant re-resolution honest: a two-minute grace window
  so a late fire still counts, and a `handled_occurrence` watermark that can
  suppress a re-fire but can never silence the future, because suppression also
  requires the candidate to be at-or-before now. Skip needed a watermark of its
  own, since a bare flag eats a second alarm once the skipped occurrence passes.
  At most one alarm rings and a newly-firing one takes over, so closing a cycle
  touches the shared ring and auto-silence slots only when that alarm is the one
  ringing; every write to `alarm_states` runs under one mutex, because a cold
  start triggered by the fire broadcast otherwise races its own reschedule pass.
  The "never to silence" chain is a testable invariant: full-screen intent, else
  a heads-up notification that still rings, else the default ringtone, else
  forced vibration. Schema v2 adds `alarm_states`, and the boot id closes M2's
  blind spot for running timers and the stopwatch, a scope move from M7
  recorded in `ARCHITECTURE.md` §5. Knowingly open: the migration and the other
  12 instrumentation tests compile but have never run on hardware, as no device
  was attached; an auto-silenced alarm is silently missed, with no missed-alarm
  notification until someone asks for one.

### ✅ M4 — App shell
- Navigation host, four tabs, M3 navigation bar, adaptive rail on wide layouts.
- The **live pill** (`PLAN.md` §9): shared running-state surface, pause + stop
  from any tab.
- The ring screen UI: over-lockscreen, turn-screen-on, snooze/dismiss, optional
  dismiss challenge that cannot strand the user.
- Motion from the kit's `identity` module; predictive back.

  ✅ 480 JVM tests, 154 of them new, and still no Robolectric. Keeping it that
  way meant moving everything that decides something out of the composables:
  `ShellNavigation` (what a tab tap does to the back stack, and that back from a
  tab goes to Alarms while back from Alarms leaves the app),
  `LivePillSelector` (which of several running things the pill is about),
  `ClockFormat`, and `ChallengeGate`. The adaptive part is
  `NavigationSuiteScaffold` taking its own default (the short bar on a phone,
  the wide rail on a tablet) rather than a hand-rolled width branch that would
  miss the tabletop posture. The pill shows for paused and expired timers too,
  not only running ones: pausing from the pill must not make the pill disappear
  under the thumb that pressed it. Its value comes from the snapshot, so a timer
  that reaches zero flips to "finished" before M6's expiry write lands. The
  challenge's escape hatch opens on three wrong answers or sixty seconds of
  ringing, for every challenge kind, asserted as one parameterised invariant;
  snooze is never gated; a completed hold *is* the dismissal, and an
  accessibility click completes the hold, because a challenge TalkBack cannot
  answer is exactly the stranding the requirement forbids. floret-kit went to
  0.5.0 for one change: `CollapsingScaffold`'s `onBack` is nullable, so a
  top-level tab draws no back arrow that would pop nothing.

  Knowingly open: M4 ships no in-app way to *start* a timer or the stopwatch, so
  the live pill cannot be demonstrated by hand until M6/M7. It is proven by its
  unit tests and by an instrumentation test that seeds the repository directly,
  and no debug button was added to paper over it. The nine new instrumentation
  tests compile in the gate but have never run on hardware, as no device was
  attached. `POST_NOTIFICATIONS` is asked for once on first launch; explaining a
  denial, and the exact-alarm and full-screen-intent grant deep links, moved
  wholly to M10's self-check, which is the screen with room to say why.

### ✅ M5 — Alarms screen
List, create, edit, delete, enable/disable. Time picker, repeat-day selector,
label, ringtone picker, vibrate, snooze settings, per-alarm overrides. Next-fire
line ("in 9h 12m") on each row. Built on `GroupedSurface`/`GroupedRow`.

  ✅ 624 JVM tests, 144 of them new, and still no Robolectric, kept for a
  whole *form*, which is where a screen usually reaches for one. Everything a
  tap decides is a pure object (`AlarmListRows`, `NextFireFormat`,
  `RepeatDaysSummary`, `AlarmDefaults`, `AlarmOverrides`, `RingtonePicker`,
  `AudioSourcePolicy`, `AlarmRoutes`) or a ViewModel method tested over the
  real `AlarmEngine`; three new `ArchitectureRulesTest` rules make that a
  build fact rather than a habit.

  The decisions worth knowing: the next-fire line comes from
  `AlarmEngine.upcoming`, which grew a `refresh: Flow<Unit>` cadence parameter
  so the countdown advances without a database write and the screen re-derives
  none of the engine's time arithmetic; the list is ordered by time of day,
  not by next fire, so rows never swap under a thumb at 07:00; every per-alarm
  override is an `OptionPicker` whose first row is "App default (On)", because
  the stored value is ternary and a `Switch` cannot say "I have not chosen";
  silence is an explicit sentinel (`clockula://silent`) that resolves to an
  empty audio-source list and therefore forces vibration, so a "silent" alarm
  is still an alarm; an unreadable sound is reported, never rewritten, because
  an unmounted card comes back; and every editor write is a transactional
  read-modify-write, so a one-shot alarm that disabled itself underneath an open
  editor is not re-enabled by the next unrelated tap.

  The correctness find: disabling or deleting a ringing alarm used to clear
  `ringingSince` without stopping the ring service (whose auto-silence backstop
  then disarms itself on its own null guard), so the alarm would have sounded
  until the 15-minute wake-lock timeout. Both new ViewModels now dismiss through
  the engine first, conditionally, so a merely snoozed alarm keeps its snooze.

  floret-kit went to 0.6.0 for one change: `CollapsingScaffold` takes a
  `floatingActionButton` slot, forwarded to the underlying `Scaffold`, rather
  than each app re-deriving the inset maths in a `Box`.

  Knowingly open: no hand reordering of alarms (the time order answers it, and
  `alarms` has no `sort_order` column), no multi-select or bulk delete, no undo
  chip after a delete (traded for a confirmation that names the alarm), no
  repeat presets, and no per-alarm auto-silence (still M10 at the earliest).
  Editing the app-wide defaults every override row offers to follow is M10's
  settings screen. The seven new instrumentation tests compile in the gate but
  have never run on hardware, as no device was attached.

### ✅ M6 — Timers
Multiple concurrent timers, presets, labels, add/pause/reset/+1min. Foreground
service with notification controls, expiry ringing reusing M3's audio path.
Elapsed-realtime anchored.

  ✅ 843 JVM tests, 219 of them new, and still no Robolectric, kept for a
  second screen *and* for a foreground service, which is the other place a
  project usually gives in. Every decision is a pure object (`TimerReadings`,
  `TimerExpiry`, `TimerRingPolicy`, `TimerNotificationPolicy`, `TimerPresets`,
  `TimerDurationEntry`, `TimerListRows`, `TimerRoutes`) or a method on the
  engine or one of the two ViewModels, tested over the real
  `TimerRepositoryImpl` and the real `TimerEngine`. `TimerService` holds no
  policy and no state beyond one flag; even its two `delay` amounts come from
  the engine, so the timings are asserted in a JVM test. Three new
  `ArchitectureRulesTest` rules pin it, one of which makes a *second*
  `MediaPlayer` a build failure. The four new instrumentation tests bring the
  compiled total to 32.

  The decisions worth knowing: the expiry slot is one registration on the
  `ELAPSED_REALTIME_WAKEUP` base (so a `TIME_SET` cannot warp a running timer,
  and a timer never populates `getNextAlarmClock()`, which belongs to the user's
  next *alarm*), and its fire is an idempotent sweep carrying no id, so a
  late delivery expires everything due, an early one writes nothing, and the
  anchors *are* the watermark; the service adds a second, independent trigger
  for the same sweep, because neither has to be reliable alone; one
  `systemExempted` foreground service covers the countdown and the ring, alive
  on the live pill's own "active, not running" predicate, and the countdown
  holds no wake lock (the alarm slot is what wakes the device); an alarm
  ringing wins the audio and the timer's ring is deferred, because the
  arbitration is a pure function of current state rather than an event, so the
  reverse case is symmetric and free; two timers expiring
  together share one ring session and one ten-minute window, and each is stopped
  on its own, because silently resetting both destroys the information "which
  one finished"; auto-silence stops the noise and leaves the row EXPIRED, which
  is the load-bearing difference from an alarm; presets are app-wide in
  DataStore with out-of-range values dropped, never clamped; and the
  notification's countdown is the platform chronometer, so the service posts
  once per state change rather than once per second for forty-five minutes.

  Two things the checklist did not name were changed anyway, and are recorded in
  `ARCHITECTURE.md`: "+1 min" on an expired timer now resumes it with
  exactly the extra, in one transaction (M2 left it paused, written before any
  UI existed and pinned by no test; resuming is the only way to avoid
  emitting an intermediate PAUSED frame to the pill and the row); and the live
  pill's timer actions now go through `TimerEngine`, because after M6 there is
  an AlarmManager registration and a service that have to move with the state.
  The pill's contract is otherwise unchanged: M6 *extracted* its precedence into
  `TimerReadings` so the pill, the notification and the ring share one ordering,
  and `LivePillSelectorTest` passing unmodified is the proof.

  No schema change (the database stays at v2; the ring session's one volatile
  fact is a DataStore record, which is `PLAN.md` §5's own record-versus-rows
  split and keeps it out of M10's backup), no new permission (a timer never
  takes over the screen: no full-screen intent, no ring activity), and no
  floret-kit change: the kit earned `CollapsingScaffold`'s
  `floatingActionButton` slot at 0.6.0 for exactly this shape of tab, and M6
  consumes it.

  Knowingly open: no hand reordering of timers (`sort_order` and
  `TimerRepository.reorder` stay caller-less, as they do for alarms), no
  per-timer vibrate, volume-ramp or dismiss-challenge overrides (`timers` has
  one nullable settings column and four more would be a schema change), no
  "stop all" when several timers are expired, no undo after a delete (traded for
  a confirmation, as with alarms), no ramp on a timer's ring, no presets
  management screen and no editing of `default_timer_duration_millis` or
  `default_timer_ringtone_uri` (M10's settings screen). `TimerSnapshot.anchorIsStale`
  is still not surfaced in the UI. Of `AlarmClock`'s intent contract only the
  in-process hook `ShellNavigation.tabForAction` exists, wired to one action;
  M9 owns the rest. The four new instrumentation tests compile in the gate but
  have never run on hardware, as no device was attached; M5's seven carry the
  same caveat.

  Two cases in the new suite are left red. Both are defects in the test, not in
  the code, so they are a design conversation rather than a behaviour gap:
  `TimerIntentsTest` §5.17 #3 asserts `containsNoDuplicates` over a list that
  contains the *same* `requestCodeFor(REQUEST_PAUSE, 1L)` call twice by
  construction, which no deterministic function can satisfy; and `TimerPrefsTest`
  §5.10 #11 opens a second DataStore over a live file in the same scope, which
  DataStore refuses by design, as `SettingsPrefsTest` itself documents.

### ✅ M7 — Stopwatch
Start/stop/reset, laps with splits and cumulative times, best/worst lap emphasis.
Foreground service so it survives backgrounding; laps persist across process
death. This is where the big-readout typography gets settled for the whole app.
The post-reboot repair is not here: M3 took it (`ARCHITECTURE.md` §5), and
M7 owns only how the paused result is presented.

  ✅ 968 JVM tests, 125 of them new, and still no Robolectric, kept now for
  a third screen and a third service. Everything the tab decides is a pure
  object (`StopwatchFormat`, `StopwatchReadings`, `LapStats`, `StopwatchLaps`,
  `StopwatchNotificationPolicy`, `StopwatchRows`) or a method on the engine or
  the ViewModel, tested over the real `StopwatchRepositoryImpl` across the
  real `appendLap` transaction and a real DataStore under a `@TempDir`.
  `StopwatchService` holds no policy and no state at all (not even the
  timers' one `alreadyAlerted` flag, because nothing here ever alerts), and it
  has no `delay`. Two new `ArchitectureRulesTest` rules and two extended; the
  four new instrumentation tests bring the compiled total to 36.

  The decisions worth knowing: the foreground service is `specialUse`, not
  `systemExempted`, because a stopwatch is not alarm functionality. It
  schedules nothing, wakes nothing and rings nothing, and claiming the type
  would have the app tell the platform something untrue. It holds no wake
  lock, ever, since the elapsed value is derived from a persisted monotonic
  anchor at read time; the notification is the platform chronometer counting
  up, with its base `now - elapsed` derived at post time and never stored,
  which is how the run record keeps its "no wall-clock field whatsoever" rule;
  its channel is `IMPORTANCE_LOW` and silent, because nothing here ever
  finishes, so there is no alert-once rule and no session window anywhere in the
  path; the readout is hundredths and it is two strings, so the
  fastest-changing glyphs are not the loudest ones; best and worst need three
  laps, ties go to the earliest, and all-equal means neither; laps are capped at
  999, guarded in the engine over one `COUNT(*)` so appending never
  materialises the rows; and the tab ticks at ~60 Hz only while running, so
  a paused or idle stopwatch costs nothing.

  Three things the checklist did not name were changed anyway, and are recorded
  in `ARCHITECTURE.md`: the live pill's stopwatch actions now go through
  `StopwatchEngine` (M6 recorded this as knowingly open), with the pill's own
  contract unchanged and `LivePillSelectorTest` passing unmodified as the
  proof; `LivePillSelector` was re-pointed at a new pure `StopwatchReadings`,
  the same extraction M6 made for timers, so the pill, the shade and the tab
  cannot answer "what does this read right now" three different ways; and
  `ui/theme/Type.kt` stopped being `Typography()`: tabular figures on the
  display, headline and title roles, app-wide, with no call-site change
  anywhere, plus `ClockulaReadoutDefaults` naming the three roles a readout uses.

  The correctness find: a run that met a reboot mid-lap had `accumulated = 0`
  while its laps still recorded forty seconds, so the honest reading is
  floored at the last lap's total as well as at zero. A readout sitting
  below a lap the same run printed is not one anyone can believe, and the laps
  are real measurements that are never deleted to tidy the arithmetic. Recorded
  in `ARCHITECTURE.md` §5. The review found the other half of it: the floor has
  to be banked as well as displayed. Resuming writes
  `accumulated = max(accumulated, lastLapCumulative)`, and pausing and lapping
  bank the same floored reading; otherwise the record ticks from below its own
  last lap, the shade's free-running chronometer counts while the tab sits
  frozen, and the next lap records a cumulative below the previous one. The
  same review pass moved the engine's verbs onto the reading's mode rather than
  the stored row's, so a stale record's Resume button resumes instead of hitting
  a silent no-op. `pauseAfterReboot` itself is still untouched.

  No schema change (the database stays at v2; `LapDao` gains one `COUNT(*)` and
  nothing else), one new permission (`FOREGROUND_SERVICE_SPECIAL_USE`, the
  one that names what the service actually does), and no floret-kit change:
  the kit stays at 0.6.0, untouched.

  Knowingly open: no lap export, sharing, deletion or editing, and no "previous
  run" history; no Lap button on the live pill (it has two controls by design);
  no user-configurable lap cap or readout precision; no reordering, grouping or
  filtering of laps; no countdown-to-start, split-lap alarm, sound or vibration
  from the stopwatch (a build rule enforces the silence); no analog stopwatch
  face and no `MaterialShapes` showpiece, which `PLAN.md` §8 reserves for the
  world-clock face, the ring screen and timer progress. `StopwatchReading.anchorIsStale`
  is still not surfaced, the same answer M6 gave for timers. Of `AlarmClock`'s
  intent contract only the in-process hook `ShellNavigation.tabForAction` grew a
  second action; M9 owns the rest. The four new instrumentation tests compile in
  the gate but have never run on hardware, as no device was attached; M5's
  seven and M6's four carry the same caveat.

### ✅ M8 — World clock
IANA zone list with search, ICU-localised city and zone display names, offset and
day-difference relative to home, reorder, home-zone handling. The analog face
built from `MaterialShapes` / `androidx.graphics.shapes` is the app's one
deliberate showpiece.

  ✅ 1 139 JVM tests, 171 of them new, and still no Robolectric, kept now
  for a fourth screen. This was the milestone most tempted to break it,
  because ICU and `Canvas` both look like they need a device. They do not:
  everything ICU answers is behind `ZoneNames` (faked, with per-method call
  counters and a `failEverything` flag), everything the face draws is
  `DrawScope` calls over `AnalogFace`'s angles and `DayNight`'s fraction (two
  pure objects asserted at every minute of the day without a pixel), and
  everything the list and the picker decide is `WorldClockRows`,
  `ZonePickerRows`, `ZoneSearch`, `ZoneComparisons`, `ZoneCatalog` and
  `WorldClocks`. Every write is a `WorldClockViewModel` method over the real
  `WorldClockRepositoryImpl` across `FakeWorldClockDao` (so `addIfAbsent` and
  `reorder`'s real transactions run) and a real DataStore under a `@TempDir`.
  Three new `ArchitectureRulesTest` rules; the four new instrumentation tests
  bring the compiled total to 40.

  The decisions worth knowing: there is no engine and must not be one, because a
  world clock owns no AlarmManager slot, no service and no ring session, so a
  fourth engine would be three empty seams and a `Mutex` guarding nothing; the
  tab reads the wall clock, and re-reads both the instant and the device
  zone on every tick, which is what makes `TIME_SET` and `TIMEZONE_CHANGED`
  free (no new receiver branch); the zone ids stay pure (through a new
  `ZoneProvider.available()`) while only the names cross the ICU seam, which
  is one file, `runCatching`-guarded end to end and pinned there by a build
  rule; the catalog keeps only region-based ids that are their own canonical
  id (so the tzdata's own canonicalisation drops Asia/Calcutta in favour of
  Asia/Kolkata and no table of ours can rot), unless the canonical id is not on
  the device at all, since ICU and tzdata are separately updatable APEX modules
  that can disagree, plus `UTC` by name; search is
  word-prefix over folded text, so "erl" deliberately does not find Berlin;
  the offset and day are computed at the instant from both zones' rules, the
  same discipline `AlarmOccurrences` uses, so Berlin↔Sydney is ten hours in
  January and eight in July; labels are data, not strings, so Kathmandu is
  `Ahead(4, 45)` rather than a signed 285 for a translator to interpret; the
  cadence is 200 ms with the state truncated to the second (face) and the
  minute (rows) and `distinctUntilChanged` over it; the list is capped at
  24, with the cap spelled out on the screen once it is reached, because
  M3's `ExtendedFloatingActionButton` has no `enabled` to dim; and
  reduce-motion snaps the morph but never stops the second hand,
  because a clock whose second hand stops is a broken clock.

  Three things outside the checklist were changed anyway, and are recorded in
  `ARCHITECTURE.md`: `ZoneProvider` gained `available()`, because
  `ZoneId.getAvailableZoneIds()` at a call site is the same ambient read
  `PLAN.md` §5 bans for `now()`; `ui/common/AlarmTimeFormatter.kt` gained
  `rememberLocalTimeFormatter` with the existing function delegating to it, so
  M5's "one 12/24-hour formatter" survives a fourth caller instead of growing a
  second answer; and `ShellDefaults.EmptyTabScreen` was deleted, its own
  KDoc having said it existed only until each of M5–M8 replaced its own tab;
  the world clock was its last caller.

  No schema change at all (the database stays at v2; `WorldClockEntity`, its
  DAO, its mapper and its repository are not edited), no new permission and no
  manifest entry, one new dependency coordinate
  (`androidx.graphics:graphics-shapes:1.0.1`, pinned to the version that already
  resolved transitively, because M8 imports that package directly), and no
  floret-kit change: the kit stays at 0.6.0, untouched, with
  `ReorderableColumn`, `GroupedSurface`, `FullScreenPicker`, `InlineTextField`
  and `CollapsingScaffold`'s FAB slot already covering the screen.

  Knowingly open: no rename UI for `world_clocks.label` (the column is honoured
  on read and M10's backup import will carry labels); no automatic "home" row
  while travelling (a separate feature with its own preference, and the hero
  face already follows the device); no per-city detail screen, map,
  sunrise/sunset or day-length bar (`DayNight` is a stated convention for a
  shape morph, not an ephemeris); no live times or long zone names in the picker;
  no seeded data, so a fresh install really has an empty list; no reordering by
  anything but `sort_order`; and no `MaterialShapes` on the ring screen or timer
  progress, which `PLAN.md` §8 also sanctions; the build rule makes widening
  that allowlist a deliberate edit. The four new instrumentation tests compile
  in the gate but have never run on hardware, as no device was attached;
  M5's seven, M6's four and M7's four carry the same caveat.

### ✅ M9 — System interop
The full `android.provider.AlarmClock` contract (`PLAN.md` §6): `SET_ALARM`,
`SET_TIMER`, `SHOW_ALARMS`, `SHOW_TIMERS`, `DISMISS_ALARM`, `SNOOZE_ALARM`, plus
next-alarm publishing. Hostile-input validation at the intent boundary.
- **Tests:** extra validation, including malformed and out-of-range input.

  ✅ 1 420 JVM tests, 281 of them new, and still no Robolectric, kept now
  for a *boundary*, which is the last place a project usually gives in, because
  "it is an `Intent`" sounds like it needs a device. It does not: the door is
  split so that everything which decides is pure (`AlarmClockRequests`,
  `AlarmMatching`, `IntentExtras`, `InteropText`, `InteropDeepLinks`) and the
  Android half is two policy-free files: a `Bundle` reader and an `Intent`
  builder. `AlarmClockHandler` orchestrates with no `android.*` import, as
  all three engines do, and is tested over both real engines through a new
  `InteropHarness`. It is deliberately not a composition of the existing
  harnesses, since each builds a DataStore over the same file name and DataStore
  refuses a second instance over a live one. Four new `ArchitectureRulesTest`
  rules and a new `ManifestRulesTest`, because the manifest is where this
  milestone could be silently wrong and none of it fails a compiler. The five
  new instrumentation tests bring the compiled total to 45.

  `ACTION_DISMISS_TIMER`, the seventh action, named neither in this checklist
  nor in `PLAN.md` §6's table, was built anyway: "the full contract" is the
  milestone's own first sentence, and shipping six of seven and calling it full
  would be the kind of quiet gap this loop exists to avoid.

  The decisions worth knowing: one range rule settles every extra. A value
  that decides when something will ring in the *future* is dropped out of
  range, and a value that modifies what the user is doing *now* is clamped, so
  hour 25 never becomes an enabled 23:00 alarm while a 1 000-minute snooze
  becomes the 60 minutes `ClockPrefs` would have allowed; a `Bundle` is
  modelled honestly as `Map<String, Any?>` with four *total* accessors,
  because hostile input is an `Int` where a `String` was documented far more
  often than it is an out-of-range number, and the platform's typed getters
  cannot tell "absent" from "wrong type"; `SKIP_UI` never means skip
  validation, so an incomplete `SET_ALARM` creates the row from the extras that
  *were* valid, at the next whole hour, and lands the user in its editor, since
  M5 locked that the editor edits a persisted row and nothing is ever written
  and hidden; `ALARM_SEARCH_MODE_TIME` matches exactly, never "nearest", and
  an ambiguous 0..12 hour with no `IS_PM` carries *both* readings as candidates
  so the question only reaches the user when the data cannot settle it;
  `ALARM_SEARCH_MODE_ALL` never asks (taken literally, the contract's "show
  the results" would make the mode useless), while every other search with two
  or more matches opens the milestone's one new surface, an M3 basic dialog with
  a `ListItem` list where tapping a row *is* the answer.

  A correctness find, fixed here: the `AlarmClockInfo` `showIntent` (the
  thing the system opens when the user taps the status-bar alarm icon or the
  lockscreen's next-alarm line) pointed at the *ring screen* with
  `CLEAR_TASK`, so tapping "my next alarm" opened a ring screen for an alarm
  that was not ringing, which resolved to `Finished` and closed itself. It now
  opens `MainActivity` on the Alarms tab, and a build rule fails the build if
  `AndroidAlarmScheduler` ever names `AlarmRingActivity` again. That intent is
  what "next-alarm publishing" means to a user.

  The schema went to v3: `delete_after_use` on `alarms` and on `timers`, two
  `ALTER TABLE`s in `MIGRATION_2_3`, `3.json` exported and committed, `1.json`
  and `2.json` untouched. The contract says in two places that a `SKIP_UI`
  alarm or timer should be removed once dismissed, and without the column a
  daily "wake me at 6:30" turns the Alarms tab into a graveyard.
  The flag is set only on the `SKIP_UI` create path and honoured inside the two
  engines, under their locks.

  Four things outside the checklist were changed anyway, and are recorded in
  `ARCHITECTURE.md`: `AlarmEngine` gained `dismissUpcoming(id)` and a
  `minutesOverride` on `snooze`, because both move the ring slot, the
  auto-silence backstop and the next-alarm registration, and a handler calling
  three repositories in a row would race a fire broadcast; `TimerEngine` gained
  `dismissAllExpired()` and `dismissExpired(id)` (the first is the "stop all"
  the UI still deliberately does not offer, because that rule was about a thumb
  on a ring and a programmatic request has stated exactly what it is discarding);
  `domain/text/TextFolding` was extracted and `ZoneSearch` delegates to it,
  with `ZoneSearchTest` passing unmodified as the proof; and the shell's
  `openTab` became `openRequest: ShellRequest?`, since three of the contract's
  outcomes are not "select a tab".

  No new `<uses-permission>`: the one new permission string is
  `android:permission` *on the door*, a requirement on the caller. No new
  dependency coordinate and no floret-kit change; the kit stays at 0.6.0,
  untouched, because the one new surface is material3's own `AlertDialog`.

  Knowingly open: no voice-interaction follow-on flows
  (`VoiceInteractor.CompleteVoiceRequest`/`PickOptionRequest`), so
  `android.intent.category.VOICE` is deliberately not declared and Clockula
  accepts a `clockula://` deeplink while publishing none; no "most closely
  matched" time search; no toast on a `SKIP_UI` write (the status-bar icon is
  the platform's own receipt); no `delete_after_use` surface in either editor, so
  an assistant-set alarm the user edits stays transient, as it does in Google
  Clock; no `EXTRA_MESSAGE` carried into the timer setup panel when no length
  was given; and `AlarmScheduler.systemNextAlarm()` is still caller-less, since it
  belongs to M10's self-check screen. The five new instrumentation tests and the
  two new migration cases compile in the gate but have never run on
  hardware, as no device was attached; M5's seven, M6's four, M7's four and
  M8's four carry the same caveat.

### ✅ M10 — Settings, backup, the self-check, and widgets
- Settings composed from kit components: theme, dynamic colour, language,
  alarm/timer/clock defaults, about + crash reporting.
- JSON backup export/import via SAF (`PLAN.md` §7), schema documented in
  `docs/BACKUP.md`.
- The "why might my alarm not ring?" self-check screen (`PLAN.md` §4):
  real device state, deep links to the exact settings pages.
- Three Glance home-screen widgets: next alarm, timer, clock (`PLAN.md` §9).
- **Tests:** backup round-trip and unknown-field tolerance; self-check rule
  logic per permission/state combination; widget model selection/formatting.

### ⬜ M11 — Release readiness
Translations wired to Weblate, fastlane metadata and screenshots, F-Droid
reproducibility guard verified against the submodule, instrumentation smoke
tests, accessibility pass (TalkBack on the ring screen especially), on-device
reliability soak across a few real nights. Then 1.0.0.

---

## Post-v1

Deferred deliberately, in rough priority order:

- **Dock / screensaver mode**: a `DreamService` full-screen clock. Cheapest big
  win, and the surface people look at most.
- **Quick Settings tile**: timer control and next alarm from the shade.
- **Bedtime / sleep schedule**: wind-down and wake, with DND handoff. A feature
  in its own right, realistically v2.
- **`core-notification` in floret-kit**: revisit once Clockula's real
  notification surface is known (`PLAN.md` §3).
- **Migrate Calendula and Agendula onto `core-prefs` + `core-di`**: tracked in
  the kit's roadmap, not here.
- **Wear OS companion**: unscoped.