Two assertion gaps, not coverage gaps, which is why they lasted. AppRootRestorationTest
already drives AppRoot at Compact and Expanded, so JaCoCo is green on useRail -- but it
asserts only that the selected tab survives recreation, through a stub `content`
composable. Nothing anywhere queried for a rail or a bar, and nothing composed the real
screens. Measured before this file existed:
- transposing the NavigationRail and NavigationBar bodies passed the entire suite
- transposing Content's two arms passed it too
A tablet showing phone chrome, or the Convert tab opening the Join screen, and 546 tests
with nothing to say about either. AppRoot's own KDoc is why that matters more than it
looks: from targetSdk 37 the app is resized and rotated whether or not it is ready, so the
width class is not a preference.
WindowWidthSizeClass.Medium appears in no test in either source set today. useRail is
`!= Compact`, so Medium takes the rail; narrowing it to `== Expanded` is one character and
breaks every tablet and unfolded foldable. That mutation is red now, and it is red only
because of the Medium test -- the Compact and Expanded ones both survive it.
Two things this needed:
**createAndroidComposeRule rather than createComposeRule.** Rendering AppRoot with its
default content reaches ConverterScreen's `viewModel = viewModel()`, which needs a
ViewModelStoreOwner. It works because both ViewModels are
`@JvmOverloads constructor(app: Application, ...)` so AndroidViewModelFactory can build
them, and because ui-test-manifest's debugImplementation entry already puts a
ComponentActivity in the merged manifest the unit tests build against -- which
app/build.gradle.kts says in terms. Checked with a throwaway spike before the ticket was
filed, rather than discovered here.
**Two tags, applied inside main.** The only production change: TestTags.Shell, set on the
rail and the bar. There is no other way to tell the two apart -- both render the same two
destinations with the same labels and the same selection state, so any assertion writable
without them is satisfied by either layout. In TestTags and applied by the shell rather
than handed down by the test, for the reason that file's KDoc gives: a tag the test
supplies proves only that the test set it. TagTableUniquenessTest covers the new group.
564 -> 568 JVM tests, 0 failures.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
audioArgs has six arms. Five are named codecs with tests; AAC arrives through the `else`,
so nothing named it -- neither "aac" nor "192k" appeared anywhere in
FFmpegCommandBuilderTest. It is the audio MP4 and M4A get, which is to say the audio the
picker offers first and most conversions produce.
Both halves are asserted, and the bitrate is the half worth arguing for: an -b:a that
quietly changed would fail nothing, look wrong in no command line, and surface only as
files that sound different from the ones the app produced last month. Both mutations
confirmed red -- 192k -> 128k and aac -> libfdk_aac.
Asserted through MP4_H264 and M4A_AAC rather than one of them, so an AAC arm added above
the `else` later has to keep answering the same way for both.
**Deliberately not added here: an ENCODABLE_AUDIO-vs-audioArgs agreement test**, the
obvious companion to VideoCodecMimeAgreementTest. It would freeze the answer to F1, which
is open: ContainerCapabilities.kt:84 says "nothing here emits a Vorbis encoder" and
FFmpegCommandBuilder.kt:188 does. docs/coverage-read-findings.md says in terms that the
tempting fix there locks in the wrong answer and that the decision comes first. This is
the AAC arm only.
563 -> 564 JVM tests, 0 failures. No production code changed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
An assertion gap rather than a coverage one, which is the reason it survived. JaCoCo is
green on build()'s `if (indeterminate)` because ProgressNotificationTest drives it through
a real worker -- but that test reads the notification id and EXTRA_PROGRESS and nothing
else. Nothing had ever read the text. Swapping the two branches passed the whole suite;
so did replacing the caller's title with a constant. Both are red now.
What it costs to get wrong is small and permanent: a conversion four minutes in still
saying "Preparing", or one that has not started reporting yet claiming 0%. Neither is a
crash, and nothing else here would have found it.
Nothing in the suite had constructed ConversionNotifications directly, and the reason
turned out to be mechanical rather than an oversight: build() reaches
WorkManager.getInstance for the Cancel action's PendingIntent, so the notification cannot
be built without one. installTestWorkManager in setUp is the whole fixture, and the KDoc
records the coupling so the next person does not rediscover it.
areEnabled() in the same file is deliberately still untested. It has no caller anywhere in
app/src/main, so a test would assert that a function nobody calls returns what the platform
told it -- and would imply the app handles the disabled-notification case, which it does
not. That is F5 in docs/coverage-read-findings.md, and it asks for a decision rather than a
test.
561 -> 563 JVM tests, 0 failures. No production code changed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
JMR-devandClaude Opus 5