Flutter Kotlin test parity matrix

Case-level comparison of every test in the Flutter build (../mobile-app/test/, 392 files) against the Kotlin suite (app/src/test/), produced 2026-08-03. Every Flutter test(...)/testWidgets(...) case was classified by opening both sides and comparing the actual assertions, not just names.

Statuses

• PORTED — a Kotlin test exercises the same logic with equivalent assertions.
• DIVERGED — a Kotlin test covers the scenario but asserts less or differently.
• MISSING — no Kotlin coverage, and the case is portable to a JVM unit test.
• N/A-WIDGET — Flutter widget/golden rendering with no JVM-portable core.
• N/A-FRAMEWORK — Flutter-only plumbing (Riverpod wiring, plugin bridges, Dart re-implementations of things the JVM provides natively, the forward data migration, …).
• N/A-BEHAVIOR / BLOCKED — the case cannot be ported as-is because the Kotlin behavior itself differs; see the divergences at the end of this file.

Two of those labels are easy to over-read, so read them narrowly:

N/A-WIDGET does not mean untestable. It means not portable to a JVM unit test. Compose has an equivalent for most of them, and these are now being ported as instrumentation tests under app/src/androidTest/. They need a device, so CI does not run them and they are run by hand instead — a widget row marked PORTED is therefore compiled and reviewed but not continuously verified, and its note says so. Treat whatever is left in this bucket as unproven behavior rather than covered ground.

Of the original 389, 332 are ported and passing on a Pixel 6 Pro, including 57 goldens whose baselines are recorded and committed. The remaining 57 are reported unportable rather than merely undone, each with its reason on its row: View-based Activities a Compose rule cannot reach, behaviour Kotlin deliberately does not have, device state that cannot be made deterministic, and cases whose nodes cannot be selected without adding test tags to production.

Some Flutter widget cases are pinned once here rather than per screen, because Kotlin shares a composable where Flutter repeats one. Every metric screen's "shows the access gate when permission missing" routes through HealthConnectScreenShell, so it is covered by HealthConnectAccessGateTest and the row says so. What that leaves untested is a screen that forgets to wrap itself at all — a wiring mistake rather than a gate one, and one its own content test catches by rendering content where a gate belongs.

N/A-BEHAVIOR is a hypothesis, not a verdict. It records that the two apps behave differently — not which one is right. Auditing the bucket found four entries that were straightforwardly Kotlin defects: a clipped DST bucket that blanked a day of steps and hydration, a failing history drain that starved the two behind it for the life of the process, and a reminder race that armed the config the user had just moved away from. All four are now fixed and PORTED. Before filing anything else here, check which side is wrong.

N/A-FRAMEWORK was audited in full and is the least trustworthy label. All 407 rows were re-opened against both codebases. The bucket is mostly sound — Riverpod containers, pigeon bridges, Dart re-implementations of Result and Channel, org.mapsforge's tile and label geometry, WorkManager doWork orchestration — but the dismissals were written from the Flutter side, and a dismissal that names the Flutter mechanism ("Kotlin has no RefreshCoordinator") tends to skip the question of whether Kotlin honours the invariant. Eight shipping defects came out of asking that question; they are listed below and fixed. Two patterns account for most of the mislabelling, and are worth checking before trusting any remaining row:

• "Kotlin has no " is about the mechanism, not the behaviour. Every metric screen's LifecycleEventEffect(ON_RESUME) is the refresh coordinator's counterpart, and it is assertable with a TestLifecycleOwner.
• "instrumentation-only" was applied to whole files where only the Glance or AppWidgetManager plumbing needs a device and the decision layer underneath is plain Kotlin over two repository interfaces. Thirty-five widget rows carried that note; six are now JVM tests.

Global tally

Counted from the case rows themselves, not the per-file summary tables (which is where an earlier hand-tally drifted). One row is an exception: the csv_row_converter_test.dart row in features.md collapses 43 cases that map 1:1 by name, and is counted here at its true size.

Section	Cases	PORTED	DIVERGED	MISSING	N/A-WIDGET	N/A-FRAMEWORK	N/A-BEHAVIOR	BLOCKED
Manual entry / hydration / mindfulness / caffeine	537	453	39	0	19	9	17	0
Other features (settings, activity, sleep, heart, imports/csv+route, bodyenergy, readiness, widgets, onboarding, achievements)	871	619	73	26	23	123	7	0
Dashboard + device sync	89	70	9	0	4	6	0	0
Apple Health import	97	70	4	0	0	23	0	0
Body / cycle / recovery	48	44	2	0	0	2	0	0
Domain / data / core	1,055	729	85	9	5	222	5	0
UI / goldens / integration (the device rows left with the device layer)	744	646	65	9	6	10	0	8
Total	3,441	2,631	277	44	57	395	29	8

Of the 2,952 portable cases (PORTED + DIVERGED + MISSING), 2,631 are ported 1:1 — up from 1,365 when the matrix was first built. Each section ends with a ### Portable gaps list of what it still has outstanding and a "blocked on a behavior decision" section for rows no test can fix. The 44 remaining MISSING are 26 features rows (Body Energy diagnostics, never ported, and the features/readiness cases whose seams are still private to their screens), 9 domain/data rows, and 9 devices/UI rows.

The Kotlin suite is at 3,126 unit tests, up from 2,103 before this work, plus 402 instrumentation tests that run on a device rather than in CI.

Behavior divergences surfaced by the comparison

These are cases where a 1:1 test CANNOT be ported because the Kotlin behavior itself differs from Flutter — they need a product decision, not a test:

• Dashboard supported-metrics gating — Kotlin has no device-support concept for tiles at all (drives ~9 dashboard MISSING rows).
• Dashboard sensor status — BLE registry is sensors only; no watch-kind filtering filtering: a paired watch is counted like a sensor.
• Dashboard resume guard — Kotlin reloads on every resume; Flutter has a guard interval.
• Dashboard tile layout model — Kotlin persists an authoritative saved list (absent = hidden); Flutter persists order + hidden-set (several DIVERGED rows follow from this, not from missing tests).
• Home-widget metric catalog — Kotlin excludes CAFFEINE + CARDIO_LOAD; Flutter excludes caffeine + intensityMinutes.
• Mid-load display — Kotlin deliberately keeps the previous display while a new range loads; Flutter clears it.

Fixed during the parity work rather than deferred:

• All six daily aggregate reads now fold same-date buckets through byLocalDate. Health Connect clips a range's final bucket, and a 24h slicer is instant-aligned, so a range containing a 25-hour local day returned the leftover hour as a second bucket on the final date — which the readers either emitted twice or let overwrite the whole day.
• HistorySyncScheduler guards each drain individually. Its once-per-open latch is claimed before the first drain runs, so an escaping throw starved the other two for the life of the process.
• HydrationReminderController holds a mutex across the read and the arm, so two rapid config changes can no longer leave the earlier one scheduled.
• HcFixture.allRecords seeds the fixture's hrv, restingHeartRate and basalMetabolicRate arrays, which it had never parsed — 777 records that no corpus test had ever seen.
• Three stress-factor strings escaped their literal percent. HRV is %1$d% above your usual baseline. parsed the bare % as an %a hex-float conversion against the Int it was given, so the stress details screen threw whenever HRV sat above baseline; its neighbours printed %b as "true" and %o as octal. StringFormatSpecifierTest now checks every source string, and the test task declares strings.xml as an input so an edit cannot leave the guard stale.
• The readiness confidence line resolves through strings.xml. The daily panel and the details screen each carried their own copy of the mapping, both built from English literals, so the line stayed English in all sixteen languages.
• The route backfill now runs altitudes through RouteElevation's smoothing + hysteresis filter, like a live recording and an imported route already did, so Flutter's fixtures pass with their exact expectations.
• ActivityDetailViewModel.load() guards the speed / cadence / marker / recovery reads individually, so one failing read costs its card rather than the screen.
• HydrationReminderController.handleQuickAdd swallows a failing re-anchor: a reminder is a nicety and must never undo a drink that already landed.
• Float.dragSteps short-circuits a zero step instead of dividing by it.
• BleDeviceRepository.updateBatteryLevel early-returns on an unchanged percent, so an identical reading neither re-persists nor advances batteryUpdatedAt.
• The skin-temperature card reads the newest entry that actually CARRIES a delta — the same population its chart draws — so a delta-less newest reading no longer blanks it.
• DailyGoalProgress.currentStreakDays(today) skips an unmet day that has not finished yet; only a PAST unmet day breaks the run.
• PeriodTitles.kt renders a past rolling window as the dated span it covers ("12 Apr – 11 May 2026"), with both years when it straddles one.
• The widget launch-route allow-list parses the body_energy segment strictly as an ISO date, so yesterday and 2026-13-45 are dropped.

From the N/A-FRAMEWORK audit

• A home widget no longer overwrites itself with -- and "No data" when its read fails or times out. The 30-minute tick regularly lands while Health Connect is mid-update or Doze has slowed the read past its budget, and the failure snapshot was byte-identical to a genuinely empty day — so the home screen asserted the user had slept nothing and walked nowhere, and held that until a later tick happened to succeed. A read that did not happen now leaves the last good snapshot alone. Timeouts had been falling out of the loaders as the same null an empty day produces, which is why the two were drawn the same.
• Phone-to-phone syncs and finished Apple Health imports tell every placed widget to redraw. updatePeriodMillis is honoured at the system's convenience and not at all in Doze, so a sync at 08:05 left the morning's tiles on their pre-sync numbers while the user watched it succeed.
• Both reminder schedules are re-planned whenever the app comes to the foreground. A reminder that fires without POST_NOTIFICATIONS cancels its own alarm and returns before rescheduling, and that permission is auto-revoked for unused apps by default — so re-granting it in system settings armed nothing and the toggle went on reading "on" forever. The boot receivers only covered reboot, reinstall and clock changes, none of which is that sequence.
• MindfulnessReminderController holds the same scheduling mutex its hydration twin already had. Its HydrationReminderControllerTest counterpart was also vacuous: it probed for the interleaving with yield(), which does not suspend on an unconfined dispatcher with an empty queue, so the overlap it asserted against could never have been observed.
• The quick-beverage widget re-anchors the hydration reminder after logging. It hid the notification but left the alarm armed from the previous drink, so the nag arrived minutes after the user drank — the in-app save and the notification quick-add both re-anchored, and only the tile did not.
• findMatchingImportedClientRecordIds no longer filters candidates through a hardcoded apple_health_ prefix. CSV ids are minted csv_<slug>_<hex>, so every one was discarded before the wanted-set check and the CSV importer could never find a duplicate: re-importing an identical file reported "already present 0" for records Health Connect was holding. The fake in CsvImportServiceTest returned the intersection, so it passed against behaviour production could not produce.
• The phone-to-phone sync codec reads back the blood-pressure body position and cuff site it has always written. Both were encoded and neither decoded, so a reading taken sitting down on the left upper arm arrived on the other phone with both fields erased — into a store other apps read.
• The language picker names every language in that language. It resolved each entry through a translatable resource, so a German UI offered "Estnisch" and an Estonian UI offered "Hispaania" — exonyms, shown to the one user who by definition cannot read the current UI language. The five settings_language_* name strings are retired; only "System default" is still translated.
• verify-translations.py sees specifiers that carry flags, width or precision. Its pattern matched only %2$s in %1$.1f C · %2$s, so dropping the temperature or retyping it %1$.1d — which throws when the card draws — both passed the gate. Plural branches a locale adds beyond the base (Spanish's many, Polish's few) were compared against nothing at all and are now checked against the base's other; values-es already ships such a branch. Both holes were confirmed by sabotaging a scratch copy of res/ and watching the old script pass it.

Found while verifying the mapsforge rows

The 19 mapsforge rows are correctly N/A — tile indexing, label collision and the render transform are all org.mapsforge's. Checking that turned up three defects in the ~200 lines of Kotlin around it, which no parity row tracked:

• AndroidGraphicFactory.createInstance ran inside the composable's remember(context, mapPacksKey), so importing or deleting a map pack re-ran it. It writes a public static INSTANCE with no null check and no synchronization, on a non-volatile field read from the MapWorkerPool render threads — it now runs once per process. The clearResourceMemoryCache() on dispose went with it: it is global, so one map going away stripped the symbols from another still on screen.
• The tile-cache directory was named from the pack ids while the MapView is rebuilt on ids and paths. Re-importing a pack under the same id at a new path gave the new map the old map's cache name — and the cache is created with persistent = false, so the outgoing map's destroyAll() deleted the directory the incoming one had just opened. String.hashCode could also collide two genuinely different pack sets onto one cache. It is now a digest of the same string the map is keyed on.
• zoomForBounds was a hand-rolled ladder over the lat/lon span in degrees. It ignored the viewport, so a 540px preview and a 2160px tablet map got the same zoom when they are three levels apart; and it ignored latitude, when Mercator stretches latitude into y by roughly 1 / cos(latitude) — the same 0.2° box is 583px tall at the equator, 1169px at 60°N and 2266px at 75°N. A northern route was drawn one to two levels too close, with its ends off screen. It now delegates to LatLongUtils.zoomForBounds and clamps to the zoom range the packs actually carry, rather than a fixed 7..16 that stopped a city extract reaching the detail it was imported for.

---

Generated from docs/engineering/test-parity/README.md. Edit it there; changes made here are overwritten.