Give the Robolectric choice a reason that is still true

Two test classes justified using Robolectric by asserting that the alternative does not
exist:

  AppRootRestorationTest       "The instrumented tests cannot run on the development
                                host at all (see CLAUDE.md)"
  OutputPublisherStagingTest   "The instrumented suite cannot run on the development
                                host, so this is the only place [it] can be caught"

Both were true when written and stopped being true on 2026-08-22, when the segfault was
traced to SwiftShader's Reactor JIT against SELinux execheap rather than to the machine.
tools/local-emulator/run-e2e.sh has run API 33-36 here since.

The first one cites CLAUDE.md as its authority, and PR #73 corrected CLAUDE.md to say the
opposite. So it was no longer merely stale: a reader who followed the reference found the
contradiction, with the citation making the wrong half look verified. That is the worst
version of this -- R14, R15, R20 and R25 were all the same defect, and this is the fifth.

The choice itself was never wrong, which is why the fix is not to move these tests. Both
belong on the JVM, and the honest reason is cost rather than impossibility: neither needs
anything a device supplies, and both run inside the same ./gradlew invocation as every
other unit test instead of booting an emulator. That argument survives the correction; the
premise did not.

The old line also has a second failure mode worth naming. "Nobody can execute this" invites
a reader to skip the local run and let CI decide, which is the opposite of what the
definition-of-done in #51 asks for.

Verified: the string appears nowhere in app/src now, and testDebugUnitTest, ktlintCheck and
detekt are green.

Closes #46.

.github/workflows/status_check.yml

@@ -240,17 +240,33 @@ jobs:
           # CAVEAT, read this before trusting a green here: this leg runs with
           # SystemUI disabled and the framework restarted under it. No other leg
           # and no Pixel run uses that configuration. It is defensible only because
-          # nothing in this suite touches system UI -- these are Media3, FFmpeg and
+          # nothing THIS LEG RUNS touches system UI -- Media3, FFmpeg and
           # WorkManager tests -- and because the alternative is no CI coverage of
           # the level this app targets. **Anything that ever does depend on system
           # UI must not trust this row.** E2E_DISABLE_SYSTEM_UI is what does it;
           # .github/scripts/e2e-run.sh explains the mechanism and why every step of
           # it is verified rather than assumed.
           #
+          # "this leg" and not "this suite", since 2026-08-24, and the difference is
+          # now load-bearing: SafPickerRoundTripTest DOES touch system UI. It drives
+          # DocumentsUI and rotates the display, and both reach the gralloc mapper
+          # this image aborts in -- disabling SystemUI removes the IDLE trigger, not
+          # those. Measured per method on android-37.0: the ROTATION test takes the
+          # framework down (INSTRUMENTATION_ABORTED) and carries
+          # @FailsOnEmulatorApi37, so notAnnotation below keeps it off this row; the
+          # PICKER test passes and runs here like anything else. A rotation rebuilds
+          # every surface at once, and starting another app's activity does not.
+          #
+          # So this row does now run one test that depends on system UI, and the
+          # caveat above still applies to it: a green here is not evidence the picker
+          # works on a device with SystemUI running -- the Pixel release check is.
+          # docs/api-37-emulator-crash.md has the per-method measurements, and the
+          # correction that produced them.
+          #
           # api-level must be "37.0". A bare 37 is not an SDK package and fails
           # during setup, which cost a run to discover.
           #
-          # notAnnotation removes the two tests that do not pass on this image; they
+          # notAnnotation removes the three tests that do not pass on this image; they
           # run in the advisory job below, off the same marker so they cannot end up
           # in both or neither. docs/api-37-emulator-crash.md has the measurements.
           - label: "37"
@@ -351,7 +367,7 @@ jobs:
           if-no-files-found: ignore
 
   # ---------------------------------------------------------------------------
-  # The two API 37 tests the gating row above excludes, run on their own so they
+  # The three API 37 tests the gating row above excludes, run on their own so they
   # stay visible instead of disappearing behind a notAnnotation.
   #
   # continue-on-error: it reports, it never blocks. That is the whole reason it is
@@ -359,17 +375,29 @@ jobs:
   # job's `E2E API <label>` name, and a check cannot be both required and advisory
   # under one name.
   #
-  # It is named for WHAT IT RUNS, deliberately. Both tests drive a full H.264 ->
-  # H.265 hardware transcode through Media3Engine -- which is exactly what
-  # separates them from the two Media3EngineTest cases that pass here, since those
-  # two never decode video. The current theory about why they fail is in the next
-  # paragraph, where it can be corrected without renaming a check that people have
-  # already learned to look for.
+  # It was named for WHAT IT RUNS, and that is now APPROXIMATE rather than exact.
+  # When this job was created it held two tests, both driving a full H.264 -> H.265
+  # hardware transcode through Media3Engine -- which is exactly what separated them
+  # from the two Media3EngineTest cases that pass here, since those two never decode
+  # video. Since 2026-08-25 it also holds SafPickerRoundTripTest's rotation case,
+  # which drives no transcode at all: a real rotation rebuilds every surface at once
+  # and takes the framework down on this image (INSTRUMENTATION_ABORTED), which is a
+  # different failure from the decoder one below.
+  #
+  # The name is kept anyway, and that is a decision rather than an oversight. This is
+  # not a required context, it is red on every PR by design, and it is one people
+  # have learned to look for -- renaming a check costs more than the imprecision
+  # does. **The marker is the definition, not the name:** what this job holds is the
+  # tests that cannot pass on the API 37 emulator image, whatever their subject. The
+  # theory about the Media3 pair is in the next paragraph, where it can be corrected
+  # without touching the name.
   #
   # THEORY, NOT SETTLED: the exception surfaces at `dequeueOutputBuffer` on
   # `c2.goldfish.h264.decoder`, the emulator's own codec, which gets its frames out
   # of a host-side colour buffer -- the same readback machinery that aborts
-  # surfaceflinger on this image. What is MEASURED is narrower: these two fail on
+  # surfaceflinger on this image. It is about the MEDIA3 PAIR only; the rotation
+  # case above fails for its own reason. What is MEASURED is narrower: those two
+  # fail on
   # the API 37 emulator image; pass at API 36 on this runner under the same renderer
   # AND the same SystemUI-disable path; pass at API 33-36 without that path at all,
   # since nothing below 37 needs it; and pass on a physical Pixel 10 Pro XL at 37. That the decoder is the culprit rather than something else

CLAUDE.md

@@ -76,16 +76,22 @@ days. Read it as the current answer, and see the git history if you need the old
   `angle_indirect` and `swangle_indirect` all boot, while `auto`, `off`, `guest` and
   `swiftshader_indirect` do not. `docs/local-emulator.md` has the evidence and the per-API renderer
   table.
-- **CI runs API 37, and it gates.** The matrix is 33/34/35/36/37. Two Media3 hardware-transcode
-  tests fail inside the emulator's own `c2.goldfish.h264.decoder` rather than on anything this app
-  does; they carry `@FailsOnEmulatorApi37` and run in a separate `continue-on-error` job,
-  `E2E API 37 Media3 hardware transcode (advisory)`. The gating leg runs the other 55.
-  **That advisory job is red on every PR, by design** — do not read it as your change breaking
-  something, and do not read a green run as evidence those two tests pass.
+- **CI runs API 37, and it gates.** The matrix is 33/34/35/36/37. **Three** of the 59 instrumented
+  tests cannot pass on that image, for two unrelated reasons: two Media3 hardware transcodes fail
+  inside the emulator's own `c2.goldfish.h264.decoder`, and one SAF test takes the framework down
+  when it rotates the display. All three carry `@FailsOnEmulatorApi37` and run in a separate
+  `continue-on-error` job; the gating leg runs the other 56.
+
+  That job is still called `E2E API 37 Media3 hardware transcode (advisory)`, which no longer
+  describes everything in it. The name is kept deliberately — it is not a required context and
+  people have learned to look for it — so **read the marker, not the name**, for what it holds.
+  **It is red on every PR, by design**: do not read it as your change breaking something, and do
+  not read a green run as evidence those three tests pass.
   `docs/api-37-emulator-crash.md` has the measurements.
 
 Still true, and the reason the advisory job is not simply deleted: **API 37 needs a manual check on
-the Pixel 10 Pro XL before each release.** The advisory pair is the one thing CI cannot answer for.
+the Pixel 10 Pro XL before each release.** Those three tests are the one thing CI cannot answer
+for.
 
 On a device or emulator, build only the ABI it can execute:
 

app/build.gradle.kts

@@ -235,8 +235,11 @@ val jacocoGeneratedExcludes = listOf(
 )
 
 // AGP 9 compiles Kotlin through its built-in compiler, which writes here rather than to the
-// classic `tmp/kotlin-classes/debug`. All hand-written code in this module is Kotlin, so the
-// javac output (BuildConfig and R only) is not read at all.
+// classic `tmp/kotlin-classes/debug`. All hand-written code in the MAIN source set is Kotlin, so
+// the javac output (BuildConfig and R only) is not read at all. There is now one hand-written
+// Java file in the module -- androidTest's FixtureDocumentsProvider, which cannot be Kotlin
+// because the process it runs in has no Kotlin stdlib; its own header explains why. It is in
+// androidTest, so it is not in this task's classDirectories and this stays accurate.
 val jacocoDebugKotlinClasses = layout.buildDirectory.dir(
     "intermediates/built_in_kotlinc/debug/compileDebugKotlin/classes",
 )
@@ -340,5 +343,11 @@ dependencies {
     androidTestImplementation(libs.androidx.espresso.core)
     androidTestImplementation(libs.compose.ui.test.junit4)
     androidTestImplementation(libs.androidx.work.testing)
+    // androidTest only, and it has to be: UiAutomator drives the whole device, including
+    // windows belonging to other packages. The system file picker is one -- DocumentsUI runs
+    // in its own process, so Compose's matchers cannot see it and Espresso's cannot either
+    // (both are scoped to this process's view hierarchy). Nothing on the JVM has a device to
+    // drive, so there is no unit-test counterpart to add it to.
+    androidTestImplementation(libs.androidx.uiautomator)
     debugImplementation(libs.compose.ui.test.manifest)
 }

app/src/androidTest/AndroidManifest.xml

@@ -0,0 +1,47 @@
+<?xml version="1.0" encoding="utf-8"?>
+<!--
+  The first manifest this source set has ever had, and it exists for one component.
+
+  SafPickerRoundTripTest drives the real system file picker. DocumentsUI only shows what a
+  DocumentsProvider offers it, so a test that picks a file needs a provider to pick from, and
+  that provider has to be declared: a ContentProvider is instantiated by the system from a
+  manifest entry and cannot be registered from test code.
+
+  It is declared HERE rather than in src/debug on purpose. src/debug would put a fake storage
+  root inside the shipped debug APK, where it would show up in every developer's own file
+  picker and in every other app's; this way it is installed only by the instrumentation APK,
+  alongside the test that needs it, and is gone the moment that APK is uninstalled.
+
+  The four attributes are not decoration. Each one is required for the picker to see it:
+
+    exported             DocumentsUI is another app; an unexported provider is invisible to it.
+    permission           MANAGE_DOCUMENTS is held by DocumentsUI and essentially nothing else,
+                         so this is what stops any installed app from reading the fixture. The
+                         provider is exported to the *picker*, not to the world.
+    grantUriPermissions  How the app under test ends up able to read the URI it was handed. The
+                         picker returns the document URI with FLAG_GRANT_READ_URI_PERMISSION,
+                         and that flag does nothing unless the provider allows grants. Without
+                         it the pick "succeeds" and every read of the result fails.
+    DOCUMENTS_PROVIDER   The action DocumentsUI queries the package manager for. No filter, no
+                         root in the drawer.
+
+  The authority carries the .test suffix because this component belongs to the instrumentation
+  package (org.libremediaconverter.test), not to the app. Authorities are global to the device:
+  reusing the app's would collide with the app on any device where both are installed.
+-->
+<manifest xmlns:android="http://schemas.android.com/apk/res/android">
+
+    <application>
+        <provider
+            android:name="org.libremediaconverter.saf.FixtureDocumentsProvider"
+            android:authorities="org.libremediaconverter.test.fixtures"
+            android:exported="true"
+            android:grantUriPermissions="true"
+            android:permission="android.permission.MANAGE_DOCUMENTS">
+            <intent-filter>
+                <action android:name="android.content.action.DOCUMENTS_PROVIDER" />
+            </intent-filter>
+        </provider>
+    </application>
+
+</manifest>

app/src/androidTest/java/org/libremediaconverter/saf/FixtureDocumentsProvider.java

@@ -0,0 +1,234 @@
+package org.libremediaconverter.saf;
+
+import android.database.Cursor;
+import android.database.MatrixCursor;
+import android.os.CancellationSignal;
+import android.os.ParcelFileDescriptor;
+import android.provider.DocumentsContract.Document;
+import android.provider.DocumentsContract.Root;
+import android.provider.DocumentsProvider;
+
+import java.io.File;
+import java.io.FileNotFoundException;
+import java.io.FileOutputStream;
+import java.io.IOException;
+import java.io.InputStream;
+import java.io.OutputStream;
+
+/**
+ * One file, offered to the system file picker, so that picking one can be tested at all.
+ *
+ * <p>DocumentsUI does not browse a filesystem: it lists what {@link DocumentsProvider}s hand it.
+ * So a test that drives the real picker has to supply the thing being picked, and it has to
+ * supply it as a manifest-declared component, because a {@code ContentProvider} is instantiated
+ * by the system and cannot be registered from test code. {@code
+ * app/src/androidTest/AndroidManifest.xml} is that declaration and says why each of its
+ * attributes is load-bearing.
+ *
+ * <h2>The only Java file in this module, and it has to be</h2>
+ *
+ * <p>Everything else here is Kotlin. This cannot be: <b>the Kotlin standard library is not on
+ * this class's classpath at runtime.</b>
+ *
+ * <p>Instrumentation code normally never notices. The test APK's dex is loaded into the app's
+ * process, where the app APK supplies {@code kotlin.jvm.internal.Intrinsics} — so the test APK is
+ * built without it, deliberately, since packaging a second copy is what {@code
+ * checkDebugAndroidTestDuplicateClasses} exists to prevent. A provider is different. It is a
+ * component of the instrumentation <i>package</i>, so when DocumentsUI queries it the system
+ * starts a plain {@code org.libremediaconverter.test} process with only the test APK on its dex
+ * path, and no app APK anywhere. The Kotlin version of this file crashed there on its first
+ * query, before returning a single row:
+ *
+ * <pre>
+ * FATAL EXCEPTION: binder:6369_2
+ * Process: org.libremediaconverter.test
+ * java.lang.NoClassDefFoundError: Failed resolution of: Lkotlin/jvm/internal/Intrinsics;
+ *     at org.libremediaconverter.saf.FixtureDocumentsProvider.queryDocument
+ * </pre>
+ *
+ * <p>The compiler emits that reference for the null checks on almost every function, so there is
+ * no Kotlin dialect that avoids it. For the same reason nothing here imports {@code androidx.*}:
+ * those classes are absent from this process for exactly the same reason. Framework and JDK only.
+ *
+ * <h2>Why a provider rather than a file in Downloads</h2>
+ *
+ * <p>That would have worked, and it would have tested less. Two properties are what {@code
+ * SafPickerRoundTripTest} actually needs:
+ *
+ * <ul>
+ *   <li><b>The root declares {@link Root#COLUMN_MIME_TYPES}, and DocumentsUI filters by it.</b>
+ *       That is what gives the screen's MIME filter a mutation with a shape: ask for a type this
+ *       root does not offer and the root itself is not in the picker, so the failure reads as
+ *       "the fixture root is not there" rather than "one file among the hundreds in Downloads was
+ *       not listed".
+ *   <li><b>The contents are exactly this and nothing else.</b> A shared directory accumulates
+ *       whatever earlier runs and other tests left in it, and a picker test that finds the wrong
+ *       file passes.
+ * </ul>
+ *
+ * <p>The descriptor is opened on a real file rather than served through a pipe, deliberately.
+ * {@code InputQuery.sizeOf} falls back to {@code ParcelFileDescriptor.statSize} when a provider
+ * omits {@code OpenableColumns.SIZE}, and a pipe's {@code statSize} is {@code -1} — an unknown
+ * size, which is a different case with a screen of its own. This fixture is meant to be an
+ * ordinary, fully described file, so that the one thing under test is the round trip.
+ */
+public final class FixtureDocumentsProvider extends DocumentsProvider {
+
+    /**
+     * What the picker calls this root.
+     *
+     * <p>Deliberately not a word any other root uses. The picker's own landing screen already
+     * offers "Images", "Audio", "Videos" and "Documents", and a UiAutomator selector that could
+     * match two things is not a selector.
+     */
+    public static final String ROOT_TITLE = "LMC R38 fixtures";
+
+    /**
+     * What the file card has to end up showing.
+     *
+     * <p>The same string reaches the assertion two ways — as the picker row UiAutomator taps, and
+     * as {@code OpenableColumns.DISPLAY_NAME} on the URI the app is handed — which is exactly the
+     * round trip under test.
+     */
+    public static final String FIXTURE_DISPLAY_NAME = "lmc-r38-fixture.mp4";
+
+    /**
+     * The type the root advertises, and the one the MIME mutation has to stop matching.
+     *
+     * <p>A real type rather than something invented, so the wildcard filter the screen passes
+     * today is not the only filter under which this test could pass.
+     */
+    public static final String FIXTURE_MIME_TYPE = "video/mp4";
+
+    private static final String ROOT_ID = "lmc-r38-root";
+    private static final String ROOT_DOCUMENT_ID = "root";
+    private static final String FIXTURE_DOCUMENT_ID = "root/" + FIXTURE_DISPLAY_NAME;
+
+    /** Already in this source set, and already a real H.264 MP4 the engines can open. */
+    private static final String FIXTURE_ASSET = "sample_h264.mp4";
+
+    private static final String[] DEFAULT_ROOT_PROJECTION = {
+        Root.COLUMN_ROOT_ID,
+        Root.COLUMN_DOCUMENT_ID,
+        Root.COLUMN_TITLE,
+        Root.COLUMN_SUMMARY,
+        Root.COLUMN_MIME_TYPES,
+        Root.COLUMN_FLAGS,
+        Root.COLUMN_ICON,
+    };
+
+    private static final String[] DEFAULT_DOCUMENT_PROJECTION = {
+        Document.COLUMN_DOCUMENT_ID,
+        Document.COLUMN_DISPLAY_NAME,
+        Document.COLUMN_MIME_TYPE,
+        Document.COLUMN_FLAGS,
+        Document.COLUMN_SIZE,
+        Document.COLUMN_LAST_MODIFIED,
+    };
+
+    @Override
+    public boolean onCreate() {
+        return true;
+    }
+
+    /**
+     * The single root.
+     *
+     * <p>{@link Root#COLUMN_MIME_TYPES} is the important column. Left null it would mean "this
+     * root supports everything", the picker would list it whatever was asked for, and the MIME
+     * mutation would have nothing to bite on.
+     */
+    @Override
+    public Cursor queryRoots(String[] projection) {
+        MatrixCursor cursor = new MatrixCursor(projection != null ? projection : DEFAULT_ROOT_PROJECTION);
+        cursor.newRow()
+            .add(Root.COLUMN_ROOT_ID, ROOT_ID)
+            .add(Root.COLUMN_DOCUMENT_ID, ROOT_DOCUMENT_ID)
+            .add(Root.COLUMN_TITLE, ROOT_TITLE)
+            .add(Root.COLUMN_SUMMARY, "Instrumentation fixture")
+            .add(Root.COLUMN_MIME_TYPES, FIXTURE_MIME_TYPE)
+            .add(Root.COLUMN_FLAGS, Root.FLAG_LOCAL_ONLY)
+            .add(Root.COLUMN_ICON, android.R.drawable.ic_menu_gallery);
+        return cursor;
+    }
+
+    @Override
+    public Cursor queryDocument(String documentId, String[] projection) throws FileNotFoundException {
+        MatrixCursor cursor = new MatrixCursor(projection != null ? projection : DEFAULT_DOCUMENT_PROJECTION);
+        if (ROOT_DOCUMENT_ID.equals(documentId)) {
+            addDirectoryRow(cursor);
+        } else if (FIXTURE_DOCUMENT_ID.equals(documentId)) {
+            addFixtureRow(cursor);
+        } else {
+            throw new FileNotFoundException("no such document: " + documentId);
+        }
+        return cursor;
+    }
+
+    @Override
+    public Cursor queryChildDocuments(String parentDocumentId, String[] projection, String sortOrder)
+            throws FileNotFoundException {
+        MatrixCursor cursor = new MatrixCursor(projection != null ? projection : DEFAULT_DOCUMENT_PROJECTION);
+        if (ROOT_DOCUMENT_ID.equals(parentDocumentId)) {
+            addFixtureRow(cursor);
+        }
+        return cursor;
+    }
+
+    @Override
+    public ParcelFileDescriptor openDocument(String documentId, String mode, CancellationSignal signal)
+            throws FileNotFoundException {
+        if (!FIXTURE_DOCUMENT_ID.equals(documentId)) {
+            throw new FileNotFoundException("no such document: " + documentId);
+        }
+        return ParcelFileDescriptor.open(fixtureFile(), ParcelFileDescriptor.MODE_READ_ONLY);
+    }
+
+    private void addDirectoryRow(MatrixCursor cursor) {
+        cursor.newRow()
+            .add(Document.COLUMN_DOCUMENT_ID, ROOT_DOCUMENT_ID)
+            .add(Document.COLUMN_DISPLAY_NAME, ROOT_TITLE)
+            .add(Document.COLUMN_MIME_TYPE, Document.MIME_TYPE_DIR)
+            .add(Document.COLUMN_FLAGS, 0)
+            .add(Document.COLUMN_SIZE, null);
+    }
+
+    private void addFixtureRow(MatrixCursor cursor) throws FileNotFoundException {
+        File file = fixtureFile();
+        cursor.newRow()
+            .add(Document.COLUMN_DOCUMENT_ID, FIXTURE_DOCUMENT_ID)
+            .add(Document.COLUMN_DISPLAY_NAME, FIXTURE_DISPLAY_NAME)
+            .add(Document.COLUMN_MIME_TYPE, FIXTURE_MIME_TYPE)
+            .add(Document.COLUMN_FLAGS, 0)
+            .add(Document.COLUMN_SIZE, file.length())
+            .add(Document.COLUMN_LAST_MODIFIED, file.lastModified());
+    }
+
+    /**
+     * The fixture on disk, unpacked from this APK's own assets the first time anything asks.
+     *
+     * <p>On demand rather than seeded once in {@link #onCreate()}, because this process is started
+     * by whoever queries the provider and can be killed between two queries of the same test.
+     *
+     * <p>A failure here is reported as {@link FileNotFoundException} rather than swallowed. A
+     * provider that answers with a zero-byte file would put the test on the "Size unknown" screen
+     * with nothing saying why.
+     */
+    private File fixtureFile() throws FileNotFoundException {
+        File file = new File(getContext().getFilesDir(), FIXTURE_DISPLAY_NAME);
+        if (file.length() > 0L) {
+            return file;
+        }
+        try (InputStream source = getContext().getAssets().open(FIXTURE_ASSET);
+                OutputStream sink = new FileOutputStream(file)) {
+            byte[] buffer = new byte[8192];
+            int read;
+            while ((read = source.read(buffer)) != -1) {
+                sink.write(buffer, 0, read);
+            }
+        } catch (IOException e) {
+            throw new FileNotFoundException("could not unpack " + FIXTURE_ASSET + ": " + e);
+        }
+        return file;
+    }
+}

app/src/androidTest/java/org/libremediaconverter/saf/SafPickerRoundTripTest.kt

@@ -0,0 +1,328 @@
+package org.libremediaconverter.saf
+
+import androidx.compose.ui.test.assertTextEquals
+import androidx.compose.ui.test.junit4.v2.createAndroidComposeRule
+import androidx.compose.ui.test.onAllNodesWithTag
+import androidx.compose.ui.test.onNodeWithTag
+import androidx.compose.ui.test.performClick
+import androidx.media3.common.util.UnstableApi
+import androidx.test.ext.junit.runners.AndroidJUnit4
+import androidx.test.platform.app.InstrumentationRegistry
+import androidx.test.uiautomator.By
+import androidx.test.uiautomator.BySelector
+import androidx.test.uiautomator.StaleObjectException
+import androidx.test.uiautomator.UiDevice
+import androidx.test.uiautomator.Until
+import org.junit.After
+import org.junit.Assert.assertNotEquals
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.FailsOnEmulatorApi37
+import org.libremediaconverter.MainActivity
+import org.libremediaconverter.ui.TestTags
+
+/**
+ * Choosing a file, through the real system picker, and still having it after a rotation.
+ *
+ * Two defects, and neither is reachable from anywhere else in this repo.
+ *
+ * **The picker is opened with a filter, and a filter can hide the user's file.** `ConverterScreen`
+ * launches `ActivityResultContracts.OpenDocument` with a MIME array; DocumentsUI hides every root
+ * and every document that array does not match. Narrow it and the app still compiles, still
+ * renders, still passes every JVM test — and the user taps "Choose file" and is shown an empty
+ * picker. Nothing in either source set drove SAF **as a picker** before this: the only SAF coverage
+ * is the publish side, in `OutputPublisherPublishTest`, against hand-written `ContentProvider`
+ * fakes. The launcher wiring, the filter, and the read grant that comes back had never been
+ * executed by a test.
+ *
+ * **The picked file has to survive a rotation.** `MainActivity` declares no `configChanges`, so
+ * every rotation destroys and recreates it, and `ConversionViewModel` holds the picked file in a
+ * plain `MutableStateFlow` with no `SavedStateHandle` behind it. The only thing that carries it
+ * across is the retained `ViewModelStore` the Activity gets from resolving the ViewModel through
+ * `LocalViewModelStoreOwner`. Scope it to the composition instead and the file is gone.
+ *
+ * ### Why these two are one test class
+ *
+ * A rotation test alone has no bite of its own. `AppRootRestorationTest` already catches
+ * `rememberSaveable` -> `remember` on the JVM, and a second test whose only mutation is one an
+ * existing test catches is the vacuous test this whole decomposition exists to prevent. So the
+ * rotation here runs **from a real picked input**, which is a state no JVM test can produce:
+ * `AppRootRestorationTest` injects a stub `content` lambda specifically to avoid standing up
+ * either ViewModel, and `StateRestorationTester` saves into an in-memory map rather than a
+ * `Bundle`.
+ *
+ * ### The mutations, and what they printed
+ *
+ * Both were run, not asserted. Narrowing the wildcard array `ConverterScreen.kt` passes to
+ * `pickInput.launch` — to `arrayOf("application/x-lmc-no-such-type")` — empties the picker of the
+ * fixture root entirely, and [pickingAFileThroughTheSystemPickerFillsInTheFileCard] fails on the
+ * assertion that names it.
+ * Making the ViewModel composition-scoped leaves the picker test alone and fails
+ * [thePickedInputSurvivesARealRotation], with `:app:testDebugUnitTest` still BUILD SUCCESSFUL —
+ * which is the divergence this ticket was filed to establish, and which was doubted on it. It is
+ * `viewModel()` -> `viewModel(viewModelStoreOwner = remember { <a plain ViewModelStoreOwner> })`,
+ * **plus** `factory = ViewModelProvider.AndroidViewModelFactory()` and a `MutableCreationExtras`
+ * carrying `APPLICATION_KEY`. The factory half is not decoration: an owner that is not a
+ * `HasDefaultViewModelProviderFactory` contributes no creation extras, and the default factory
+ * cannot construct an `AndroidViewModel` without them — so the owner swap alone crashes on
+ * construction instead of demonstrating the scope. The PR body quotes both failures verbatim.
+ *
+ * ### It has to be an unlocked emulator
+ *
+ * The Pixel 10 Pro XL is secure-locked and cannot be unlocked from a shell, so the picker cannot be
+ * driven there at all. That is why this gap survived as long as it did.
+ * `tools/local-emulator/run-e2e.sh` runs API 33-36 on the development host, and both tests pass
+ * there: **59 / 0 / 0 / 2 at API 33 and again at API 36**, whole suite, 2026-08-24.
+ *
+ * ### Why only the rotation test carries [FailsOnEmulatorApi37]
+ *
+ * This class is the first thing in the suite that touches system UI, and the android-37.x images
+ * are where that stops being free: surfaceflinger aborts inside the guest's Gralloc5 mapper, init
+ * SIGKILLs zygote with it, and the framework restarts underneath the run. Disabling SystemUI --
+ * the deviation the API 37 leg already makes -- removes the *idle* trigger, not this one.
+ *
+ * The marker is on one method and not on the class, because that is what was measured, one method
+ * per fresh emulator, on `android-37.0` under `swangle_indirect`:
+ *
+ * ```
+ * thePickedInputSurvivesARealRotation            INSTRUMENTATION_ABORTED: System has crashed.
+ *                                               Expected 1 tests, received 0
+ * pickingAFileThroughTheSystemPickerFillsInTheFileCard                              PASSED
+ * ```
+ *
+ * A rotation rebuilds every surface on screen at once, which the mapper does not survive; merely
+ * starting DocumentsUI does not.
+ *
+ * **The first version of this said the class, and it was wrong.** The picker test had failed at
+ * API 37 too -- with a `StaleObjectException` that turned out to be this file's own bug rather
+ * than the image's, and which CI then reproduced deterministically at API 33, 34 and 35. Fixing
+ * it ([tapPickerNode]) and re-measuring is what separated the two. An annotation is a claim about
+ * an image, and a broken test makes every image look broken; **re-measure after fixing a test
+ * before deciding what the platform did.**
+ *
+ * The annotation says only that, and CI reads it twice, so the rotation test runs on the advisory
+ * API 37 leg and not the gating one. **Do not read it as "a rotation is allowed to lose the
+ * file".** That is what API 33 through 36 are for, and they answer it.
+ */
+@UnstableApi
+@RunWith(AndroidJUnit4::class)
+class SafPickerRoundTripTest {
+
+    @get:Rule
+    val composeRule = createAndroidComposeRule<MainActivity>()
+
+    private val device: UiDevice =
+        UiDevice.getInstance(InstrumentationRegistry.getInstrumentation())
+
+    /** Set by the one test that rotates, read by [restoreOrientation]. See its KDoc. */
+    private var rotated = false
+
+    /**
+     * Leave the device the way it was found — and only if this test moved it.
+     *
+     * Two things are deliberate here, and both are about the *other* tests on the device rather
+     * than about these two.
+     *
+     * The flag, because this runs after every test in the class, not only the one that rotated. An
+     * unconditional restore issues a WindowManager rotation request after the picker test as well,
+     * which has nothing to undo; JUnit does not promise method order, so that is an interaction
+     * between two tests that no single-class run would ever show. Tracked as a flag rather than
+     * read back off `isNaturalOrientation`, because a device whose *natural* orientation is
+     * landscape would answer that question the wrong way round.
+     *
+     * And `unfreezeRotation`, because `setOrientationNatural` does not merely rotate: it freezes
+     * the rotation there. A run that stopped after it would hand the next test a device that
+     * cannot rotate at all.
+     */
+    @After
+    fun restoreOrientation() {
+        if (!rotated) return
+        device.setOrientationNatural()
+        device.unfreezeRotation()
+        device.waitForIdle()
+    }
+
+    @Test
+    fun pickingAFileThroughTheSystemPickerFillsInTheFileCard() {
+        pickTheFixture()
+
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD_NAME)
+            .assertTextEquals(FixtureDocumentsProvider.FIXTURE_DISPLAY_NAME)
+
+        // Not the same assertion twice. The name above comes from a metadata query, which a URI
+        // with no read grant answers just as well; this line only appears once something has
+        // opened the file and read its header. It is what says the picker handed back a URI the
+        // app can actually USE -- delete grantUriPermissions from the fixture's manifest entry and
+        // the name still arrives while this goes red.
+        //
+        // The whole "Container: MP4" and not "MP4": DetailRow renders the label and the value as
+        // one semantics node.
+        awaitNode(TestTags.Converter.detailRow(CONTAINER_LABEL))
+        composeRule.onNodeWithTag(TestTags.Converter.detailRow(CONTAINER_LABEL))
+            .assertTextEquals("$CONTAINER_LABEL: MP4")
+    }
+
+    @Test
+    @FailsOnEmulatorApi37
+    fun thePickedInputSurvivesARealRotation() {
+        pickTheFixture()
+        // The identity hash rather than the Activity itself, so nothing here keeps a destroyed
+        // Activity reachable across the recreation it is being used to detect.
+        val before = System.identityHashCode(composeRule.activity)
+
+        device.setOrientationLandscape()
+        rotated = true
+        composeRule.waitForIdle()
+
+        // Two guards before the assertion that matters, because both of the ways this test could
+        // pass while proving nothing are silent ones.
+        //
+        // A device that ignored the rotation request would leave the app exactly as it was, and
+        // "the file is still there" would then be a statement about a screen nothing happened to.
+        assertNotEquals(
+            "the device did not actually rotate, so nothing below is about a rotation",
+            NATURAL_ROTATION,
+            device.displayRotation,
+        )
+        // And a rotation that did NOT recreate the Activity -- a configChanges attribute added to
+        // the manifest, an aspect-ratio or orientation lock -- would make this a recomposition
+        // test. The retained ViewModelStore is only interesting because the Activity around it
+        // really was destroyed and rebuilt.
+        assertNotEquals(
+            "the rotation did not recreate MainActivity, so the retained ViewModelStore was never used",
+            before,
+            System.identityHashCode(composeRule.activity),
+        )
+
+        awaitNode(TestTags.Converter.FILE_CARD_NAME)
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD_NAME)
+            .assertTextEquals(FixtureDocumentsProvider.FIXTURE_DISPLAY_NAME)
+    }
+
+    // --- driving the picker ---------------------------------------------------------------
+
+    /**
+     * Taps "Choose file", walks the system picker to the fixture, and returns once the app has it.
+     *
+     * Everything between the first tap and the last belongs to `com.google.android.documentsui`,
+     * which is why UiAutomator is here at all: Compose's matchers stop at this process's
+     * composition and Espresso's at its view hierarchy, and the picker is neither.
+     */
+    private fun pickTheFixture() {
+        composeRule.onNodeWithTag(TestTags.Converter.CHOOSE_FILE).performClick()
+
+        // THIS is the line the MIME filter mutation fails on. DocumentsUI matches the requested
+        // types against Root.COLUMN_MIME_TYPES and drops the roots that cannot answer, so a filter
+        // the fixture root does not satisfy takes the root out of the picker altogether -- along
+        // with "Images", "Audio", "Videos" and "Documents", measured on API 34.
+        tapPickerNode(By.text(FixtureDocumentsProvider.ROOT_TITLE)) {
+            // Which screen the picker opens on is its own business: it lands on Recent, where the
+            // roots are a strip at the bottom, but a device with a populated Recent may need the
+            // drawer. Looking in the second place widens where the root is searched for; it does
+            // not weaken what has to be found, which is still this root.
+            device.findObject(By.desc(SHOW_ROOTS_DESCRIPTION))?.click()
+        }
+
+        tapPickerNode(By.text(FixtureDocumentsProvider.FIXTURE_DISPLAY_NAME))
+
+        awaitNode(TestTags.Converter.FILE_CARD_NAME)
+    }
+
+    /**
+     * Finds the picker node [selector] names and taps it, re-finding it if it goes stale.
+     *
+     * **The re-finding is not padding, and this is not a retry of the assertion.** A `UiObject2`
+     * holds an `AccessibilityNodeInfo` captured when it was found, and DocumentsUI is still
+     * settling when the node first appears — its list rebinds, the roots strip lays out, a window
+     * animates. If the node is replaced in that gap, `click()` throws `StaleObjectException`
+     * against the handle rather than missing the target. Measured on a cold API 34 emulator:
+     *
+     * ```
+     * androidx.test.uiautomator.StaleObjectException
+     *   at androidx.test.uiautomator.UiObject2.getAccessibilityNodeInfo(UiObject2.java:1042)
+     *   at androidx.test.uiautomator.UiObject2.click(UiObject2.java:526)
+     * ```
+     *
+     * So what is retried is *acquiring a handle to a node that has to be there anyway* — every
+     * attempt still goes through [awaitPickerNode], which fails outright if the node is absent.
+     * The MIME mutation's bite is untouched: a root that is not in the picker is not found on any
+     * attempt, and the failure is still "the system picker never showed" rather than a stale one.
+     */
+    private fun tapPickerNode(selector: BySelector, ifAbsent: () -> Unit = {}) {
+        var stale: StaleObjectException? = null
+        repeat(TAP_ATTEMPTS) { attempt ->
+            // ifAbsent only on the first attempt: it navigates, and re-navigating from a screen it
+            // already reached would walk away from the node.
+            val node = awaitPickerNode(selector, if (attempt == 0) ifAbsent else ({}))
+            device.waitForIdle()
+            try {
+                node.click()
+                return
+            } catch (e: StaleObjectException) {
+                stale = e
+            }
+        }
+        throw AssertionError("$selector kept going stale between finding it and tapping it", stale)
+    }
+
+    /**
+     * The picker node [selector] names, or a failure that says which one was missing.
+     *
+     * [ifAbsent] runs once, after the first wait comes up empty, and then the wait is repeated. A
+     * null return from `findObject` is deliberately not an error there: it is the "already on the
+     * right screen" case.
+     */
+    private fun awaitPickerNode(selector: BySelector, ifAbsent: () -> Unit = {}) =
+        device.wait(Until.findObject(selector), PICKER_TIMEOUT_MS)
+            ?: run {
+                ifAbsent()
+                requireNotNull(device.wait(Until.findObject(selector), PICKER_TIMEOUT_MS)) {
+                    "the system picker never showed $selector"
+                }
+            }
+
+    /**
+     * Blocks until [tag] is in the composition, so an assertion cannot race the picker's result.
+     *
+     * The described overload of `waitUntil`, not the bare one. A timeout is how both of this
+     * class's mutations report themselves, and the bare overload's message is
+     * `Condition still not satisfied after 30000 ms` — which names neither the node nor the test.
+     * With the description it says which affordance never arrived, which is the whole finding.
+     */
+    private fun awaitNode(tag: String) {
+        composeRule.waitUntil("a node tagged $tag exists", APP_TIMEOUT_MS) {
+            composeRule.onAllNodesWithTag(tag).fetchSemanticsNodes().isNotEmpty()
+        }
+    }
+
+    private companion object {
+
+        /**
+         * Generous on purpose. This waits on another app being started, and on FFprobe spawning a
+         * native process over a `content://` URI; a timeout that merely usually passes is a flaky
+         * gating leg on five API levels, which costs far more than the seconds it saves.
+         */
+        const val PICKER_TIMEOUT_MS = 30_000L
+        const val APP_TIMEOUT_MS = 30_000L
+
+        /** `Surface.ROTATION_0`, named rather than `0` so the comparison reads. */
+        const val NATURAL_ROTATION = 0
+
+        /**
+         * How many times a picker node may be re-found before its staleness is the finding.
+         *
+         * Three, not "until the timeout". Each attempt already waits up to [PICKER_TIMEOUT_MS] for
+         * the node to exist, so this bounds only the settling window after it does; a node that is
+         * still being replaced after three of those is telling you something about the device, and
+         * a loop that hid it would be the flake rather than the fix.
+         */
+        const val TAP_ATTEMPTS = 3
+
+        /** DocumentsUI's drawer button. It carries no text, only this description. */
+        const val SHOW_ROOTS_DESCRIPTION = "Show roots"
+
+        /** The detail row `MediaProbe` fills in for anything it could open and identify. */
+        const val CONTAINER_LABEL = "Container"
+    }
+}

app/src/main/java/org/libremediaconverter/convert/ConverterScreen.kt

@@ -87,6 +87,89 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
         ActivityResultContracts.RequestPermission(),
     ) { viewModel.convert() }
 
+    ConverterScreenContent(
+        state = state,
+        settings = settings,
+        validation = validation,
+        actions = ConverterActions(
+            onPickInput = { pickInput.launch(arrayOf("*/*")) },
+            onPreset = viewModel::setPreset,
+            onContainer = viewModel::setContainer,
+            onVideoCodec = viewModel::setVideoCodec,
+            onAudioCodec = viewModel::setAudioCodec,
+            onSuggestion = viewModel::applySuggestion,
+            onQuality = viewModel::setQuality,
+            onEnginePreference = viewModel::setEnginePreference,
+            onConvert = { requestNotifications.launch(Manifest.permission.POST_NOTIFICATIONS) },
+            onCancel = viewModel::cancel,
+            onSave = { suggestedName -> chooseDestination.launch(suggestedName) },
+            onReset = viewModel::reset,
+        ),
+        modifier = modifier,
+    )
+}
+
+/**
+ * Everything [ConverterScreenContent] can ask for, in one value.
+ *
+ * A holder rather than twelve parameters because detekt's `LongParameterList` sits at its default
+ * threshold of six and `config/detekt/detekt.yml` does not relax it for `@Composable` the way it
+ * relaxes `LongMethod` and `CyclomaticComplexMethod` -- `AdvancedPicker` already sits exactly on
+ * that threshold. The rule exempts data classes, so the callbacks travel together.
+ *
+ * In production every one of these is a launcher or a `ConversionViewModel` call. Naming them here
+ * instead of handing the content a ViewModel is the whole point of the seam: a test can render a
+ * [ConversionState] no ViewModel can be driven into, since `Waiting` needs a denied foreground
+ * start and `Converted` needs a worker run that has already succeeded.
+ */
+internal data class ConverterActions(
+    /** Open the document picker. The `Idle` and `Ready` branches both offer it. */
+    val onPickInput: () -> Unit,
+    val onPreset: (OutputFormat) -> Unit,
+    val onContainer: (Container) -> Unit,
+    val onVideoCodec: (VideoCodec) -> Unit,
+    val onAudioCodec: (AudioCodec) -> Unit,
+    val onSuggestion: (OutputSpec) -> Unit,
+    val onQuality: (QualityTier) -> Unit,
+    val onEnginePreference: (EnginePreference) -> Unit,
+    /**
+     * Start the job. It asks for the notification permission first, which is why the screen never
+     * calls `convert` directly -- the launcher's result callback does, whichever way it went.
+     */
+    val onConvert: () -> Unit,
+    val onCancel: () -> Unit,
+    /**
+     * Open the save dialog for the finished output.
+     *
+     * Takes the suggested name rather than reading it back off the state, because the name comes
+     * from the job -- see `ConversionWorker.KEY_SUGGESTED_NAME` -- and the branch that renders the
+     * button is the only place that has it.
+     */
+    val onSave: (suggestedName: String) -> Unit,
+    val onReset: () -> Unit,
+)
+
+/**
+ * The converter screen, with its state handed in.
+ *
+ * Split from [ConverterScreen] so that state has somewhere to come from other than a live
+ * `ConversionViewModel`. Driving the screen through a real one needs a `WorkManager` and a media
+ * probe in the constructor, and even then two of the six states are unreachable: `Waiting` follows
+ * a denied foreground start and `Converted` follows a completed worker.
+ *
+ * `internal` rather than private, because `src/test` is a friend of `main` and this is what the
+ * state tests compose. The leaves below stay exactly where they were -- this function is a move,
+ * not a redesign, and the tests that already pin those leaves are what says so.
+ */
+@UnstableApi
+@Composable
+internal fun ConverterScreenContent(
+    state: ConversionState,
+    settings: ConversionSettings,
+    validation: Validation,
+    actions: ConverterActions,
+    modifier: Modifier = Modifier,
+) {
     Column(
         modifier = modifier
             .fillMaxSize()
@@ -115,7 +198,7 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                     modifier = Modifier.padding(bottom = 16.dp),
                 )
                 Button(
-                    onClick = { pickInput.launch(arrayOf("*/*")) },
+                    onClick = actions.onPickInput,
                     modifier = Modifier
                         .fillMaxWidth()
                         .height(PrimaryButtonHeight)
@@ -132,21 +215,19 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
 
                     is ConversionState.Ready -> {
                         FileCard(s.input)
-                        FormatPicker(settings.matchingPreset, viewModel::setPreset)
+                        FormatPicker(settings.matchingPreset, actions.onPreset)
                         AdvancedPicker(
                             spec = settings.spec,
                             validation = validation,
-                            onContainer = viewModel::setContainer,
-                            onVideoCodec = viewModel::setVideoCodec,
-                            onAudioCodec = viewModel::setAudioCodec,
-                            onSuggestion = viewModel::applySuggestion,
+                            onContainer = actions.onContainer,
+                            onVideoCodec = actions.onVideoCodec,
+                            onAudioCodec = actions.onAudioCodec,
+                            onSuggestion = actions.onSuggestion,
                         )
-                        QualityPicker(settings.quality, viewModel::setQuality)
-                        EnginePicker(settings.enginePreference, viewModel::setEnginePreference)
+                        QualityPicker(settings.quality, actions.onQuality)
+                        EnginePicker(settings.enginePreference, actions.onEnginePreference)
                         Button(
-                            onClick = {
-                                requestNotifications.launch(Manifest.permission.POST_NOTIFICATIONS)
-                            },
+                            onClick = actions.onConvert,
                             // The Advanced picker lets an impossible combination be selected on
                             // purpose, so this is what stops it from being run.
                             enabled = validation.isValid,
@@ -156,7 +237,7 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                                 .testTag(TestTags.Converter.CONVERT),
                         ) { Text("Convert") }
                         OutlinedButton(
-                            onClick = { pickInput.launch(arrayOf("*/*")) },
+                            onClick = actions.onPickInput,
                             modifier = Modifier
                                 .fillMaxWidth()
                                 .testTag(TestTags.Converter.CHOOSE_DIFFERENT_FILE),
@@ -173,7 +254,7 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                                 .testTag(TestTags.Converter.PROGRESS),
                         )
                         OutlinedButton(
-                            onClick = viewModel::cancel,
+                            onClick = actions.onCancel,
                             modifier = Modifier.fillMaxWidth().testTag(TestTags.CANCEL),
                         ) { Text("Cancel") }
                     }
@@ -192,7 +273,7 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                             style = MaterialTheme.typography.bodyMedium,
                         )
                         OutlinedButton(
-                            onClick = viewModel::cancel,
+                            onClick = actions.onCancel,
                             modifier = Modifier.fillMaxWidth().testTag(TestTags.CANCEL),
                         ) { Text("Cancel") }
                     }
@@ -208,17 +289,21 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                             // explains why a job was slow, makes the software fallback
                             // visible, and is how the user learns a remux happened rather
                             // than a re-encode.
-                            AssistChip(onClick = {}, label = { Text(s.routeReason) })
+                            AssistChip(
+                                onClick = {},
+                                label = { Text(s.routeReason) },
+                                modifier = Modifier.testTag(TestTags.Converter.ROUTE_REASON),
+                            )
                         }
                         Button(
-                            onClick = { chooseDestination.launch(s.suggestedName) },
+                            onClick = { actions.onSave(s.suggestedName) },
                             modifier = Modifier
                                 .fillMaxWidth()
                                 .height(PrimaryButtonHeight)
                                 .testTag(TestTags.SAVE_FILE),
                         ) { Text("Save file") }
                         OutlinedButton(
-                            onClick = viewModel::reset,
+                            onClick = actions.onReset,
                             modifier = Modifier.fillMaxWidth().testTag(TestTags.START_OVER),
                         ) { Text("Start over") }
                     }
@@ -226,7 +311,7 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                     is ConversionState.Saved -> {
                         Text("Saved ${s.displayName}.", style = MaterialTheme.typography.bodyLarge)
                         Button(
-                            onClick = viewModel::reset,
+                            onClick = actions.onReset,
                             modifier = Modifier
                                 .fillMaxWidth()
                                 .height(PrimaryButtonHeight)
@@ -241,7 +326,7 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                             style = MaterialTheme.typography.bodyMedium,
                         )
                         Button(
-                            onClick = viewModel::reset,
+                            onClick = actions.onReset,
                             modifier = Modifier
                                 .fillMaxWidth()
                                 .height(PrimaryButtonHeight)

app/src/main/java/org/libremediaconverter/join/JoinScreen.kt

@@ -53,6 +53,46 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
         remember(destinationMime) { ActivityResultContracts.CreateDocument(destinationMime) },
     ) { uri -> uri?.let(viewModel::save) }
 
+    JoinScreenContent(
+        state = state,
+        actions = JoinActions(
+            onPickInputs = { pickInputs.launch(arrayOf("video/*")) },
+            onJoin = viewModel::join,
+            onCancel = viewModel::cancel,
+            onSave = { suggestedName -> chooseDestination.launch(suggestedName) },
+            onReset = viewModel::reset,
+        ),
+        modifier = modifier,
+    )
+}
+
+/**
+ * Everything [JoinScreenContent] can ask for, in one value.
+ *
+ * Five callbacks would fit under detekt's `LongParameterList` threshold, unlike the converter's
+ * twelve. It is a holder anyway, so both screens present the same shape to the state tests and
+ * neither one has to be reworked the first time a branch grows a button.
+ */
+internal data class JoinActions(
+    /** Open the multi-document picker. The `Idle` and `Ready` branches both offer it. */
+    val onPickInputs: () -> Unit,
+    val onJoin: () -> Unit,
+    val onCancel: () -> Unit,
+    /** Open the save dialog. Takes the name the job chose -- see `ConcatWorker.KEY_SUGGESTED_NAME`. */
+    val onSave: (suggestedName: String) -> Unit,
+    val onReset: () -> Unit,
+)
+
+/**
+ * The join screen, with its state handed in.
+ *
+ * The same split as [org.libremediaconverter.convert.ConverterScreenContent], for the same reason:
+ * `JoinState.Waiting` follows a denied foreground start and `JoinState.Joined` follows a completed
+ * concatenation, so neither is reachable by driving a real `JoinViewModel`.
+ */
+@UnstableApi
+@Composable
+internal fun JoinScreenContent(state: JoinState, actions: JoinActions, modifier: Modifier = Modifier) {
     Column(
         modifier = modifier
             .fillMaxSize()
@@ -81,7 +121,7 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                     modifier = Modifier.padding(bottom = 16.dp),
                 )
                 Button(
-                    onClick = { pickInputs.launch(arrayOf("video/*")) },
+                    onClick = actions.onPickInputs,
                     modifier = Modifier
                         .fillMaxWidth()
                         .height(PrimaryButtonHeight)
@@ -99,14 +139,14 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                     is JoinState.Ready -> {
                         s.inputs.forEach { FileRow(it) }
                         Button(
-                            onClick = viewModel::join,
+                            onClick = actions.onJoin,
                             modifier = Modifier
                                 .fillMaxWidth()
                                 .height(PrimaryButtonHeight)
                                 .testTag(TestTags.Join.JOIN),
                         ) { Text("Join ${s.inputs.size} files") }
                         OutlinedButton(
-                            onClick = { pickInputs.launch(arrayOf("video/*")) },
+                            onClick = actions.onPickInputs,
                             modifier = Modifier
                                 .fillMaxWidth()
                                 .testTag(TestTags.Join.CHOOSE_DIFFERENT_FILES),
@@ -124,7 +164,7 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                                 .testTag(TestTags.Join.PROGRESS),
                         )
                         OutlinedButton(
-                            onClick = viewModel::cancel,
+                            onClick = actions.onCancel,
                             modifier = Modifier.fillMaxWidth().testTag(TestTags.CANCEL),
                         ) { Text("Cancel") }
                     }
@@ -138,7 +178,7 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                             style = MaterialTheme.typography.bodyMedium,
                         )
                         OutlinedButton(
-                            onClick = viewModel::cancel,
+                            onClick = actions.onCancel,
                             modifier = Modifier.fillMaxWidth().testTag(TestTags.CANCEL),
                         ) { Text("Cancel") }
                     }
@@ -157,14 +197,14 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                             style = MaterialTheme.typography.bodySmall,
                         )
                         Button(
-                            onClick = { chooseDestination.launch(s.suggestedName) },
+                            onClick = { actions.onSave(s.suggestedName) },
                             modifier = Modifier
                                 .fillMaxWidth()
                                 .height(PrimaryButtonHeight)
                                 .testTag(TestTags.SAVE_FILE),
                         ) { Text("Save file") }
                         OutlinedButton(
-                            onClick = viewModel::reset,
+                            onClick = actions.onReset,
                             modifier = Modifier.fillMaxWidth().testTag(TestTags.START_OVER),
                         ) { Text("Start over") }
                     }
@@ -172,7 +212,7 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                     is JoinState.Saved -> {
                         Text("Saved ${s.displayName}.", style = MaterialTheme.typography.bodyLarge)
                         Button(
-                            onClick = viewModel::reset,
+                            onClick = actions.onReset,
                             modifier = Modifier
                                 .fillMaxWidth()
                                 .height(PrimaryButtonHeight)
@@ -187,7 +227,7 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                             style = MaterialTheme.typography.bodyMedium,
                         )
                         Button(
-                            onClick = viewModel::reset,
+                            onClick = actions.onReset,
                             modifier = Modifier
                                 .fillMaxWidth()
                                 .height(PrimaryButtonHeight)

app/src/main/java/org/libremediaconverter/ui/TestTags.kt

@@ -55,6 +55,15 @@ object TestTags {
         /** The determinate bar in `Converting`. It carries no text, so nothing else can find it. */
         const val PROGRESS: String = "converter.progress"
 
+        /**
+         * The chip on `Converted` that says which engine ran the job and why.
+         *
+         * Conditional on `routeReason` being non-blank, and that condition is what the tag is for:
+         * its text comes from the finished job, so a text matcher looking for it would have to
+         * name a routing explanation the screen does not own.
+         */
+        const val ROUTE_REASON: String = "converter.routeReason"
+
         const val FILE_CARD: String = "converter.fileCard"
         const val FILE_CARD_NAME: String = "converter.fileCard.name"
 

app/src/test/java/org/libremediaconverter/AppRootRestorationTest.kt

@@ -33,9 +33,12 @@ import org.robolectric.RobolectricTestRunner
  * representation survives a `Bundle` round trip. A JVM round-trip test on the
  * saver covers the representation.
  *
- * Robolectric rather than the instrumented suite, deliberately. The instrumented tests
- * cannot run on the development host at all (see CLAUDE.md), and a red test nobody can
- * execute is not a loop anyone can work in.
+ * Robolectric rather than the instrumented suite, deliberately -- but not because the
+ * instrumented suite is unavailable. It runs on this host for API 33-36
+ * (`tools/local-emulator/run-e2e.sh`), and CI runs 33-37. The reason is cost: this test
+ * needs a composition and a saved-state round trip, nothing a device supplies, and it runs
+ * in the same `./gradlew` invocation as every other JVM test instead of booting an
+ * emulator. A loop measured in seconds is a loop people stay inside.
  */
 @UnstableApi
 @RunWith(RobolectricTestRunner::class)

app/src/test/java/org/libremediaconverter/convert/ConverterScreenContentTest.kt

@@ -0,0 +1,115 @@
+package org.libremediaconverter.convert
+
+import android.net.Uri
+import androidx.compose.ui.test.onNodeWithTag
+import androidx.compose.ui.test.performClick
+import androidx.compose.ui.test.performScrollTo
+import androidx.media3.common.util.UnstableApi
+import org.junit.Assert.assertEquals
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.createDrainedComposeRule
+import org.libremediaconverter.model.Validation
+import org.libremediaconverter.ui.TestTags
+import org.robolectric.RobolectricTestRunner
+import java.io.File
+
+/**
+ * The seam carries a `ConversionState` in and an action back out.
+ *
+ * The defect this bites on is the extraction having quietly stopped being an extraction: a
+ * `ConverterScreenContent` that ignores the `state` it was handed, or renders the finished job's
+ * affordances without wiring them to the callbacks the entry point supplies. Neither shows up at
+ * compile time -- an unread parameter compiles, and a `Button` whose `onClick` does nothing is a
+ * valid `Button` -- and neither is visible from the leaf tests, which compose `FileCard`,
+ * `AdvancedPicker` and the pickers directly and never see a state at all.
+ *
+ * **Both assertions were unreachable before R38.5**, which is the point of the ticket rather than
+ * a remark about it. `ConversionState.Converted` is produced only by a `ConversionWorker` run that
+ * has already succeeded, so no test can drive a real `ConversionViewModel` into it: it would need
+ * a `WorkManager`, a media probe, a staged output file and a completed job. Handing the state in
+ * is the only way to ask what the screen does with it.
+ *
+ * Deliberately not the state matrix. Which affordances each of the six `ConversionState`s renders
+ * is R38.6 (#62); this file asserts only that the injection point exists and works in both
+ * directions, so the two PRs cannot collide over the same cases.
+ */
+@UnstableApi
+@RunWith(RobolectricTestRunner::class)
+class ConverterScreenContentTest {
+
+    // Not `createComposeRule()` directly: see [org.libremediaconverter.drainEscapedCoroutineErrors].
+    @get:Rule
+    val composeRule = createDrainedComposeRule()
+
+    /** What the screen asked to save, in the order it asked. Empty until Save is tapped. */
+    private val savedAs = mutableListOf<String>()
+
+    @Test
+    fun `a converted job renders the save button`() {
+        setContent(converted())
+
+        composeRule.onNodeWithTag(TestTags.SAVE_FILE).assertExists()
+    }
+
+    /**
+     * The direction that did not exist before this change.
+     *
+     * Asserting the *name* rather than just that something was called: the suggested name comes
+     * from the job -- `ConversionWorker.KEY_SUGGESTED_NAME` -- and is what the save dialog opens
+     * with, so a Save button wired to the wrong branch's state would hand over the wrong one and
+     * a bare "was called" check would stay green.
+     */
+    @Test
+    fun `tapping save hands back the name the finished job chose`() {
+        setContent(converted())
+
+        composeRule.onNodeWithTag(TestTags.SAVE_FILE).performScrollTo().performClick()
+
+        assertEquals(listOf("holiday.mp4"), savedAs)
+    }
+
+    /**
+     * `staged` names a file that does not exist, on purpose.
+     *
+     * The branch renders `formatBytes(s.staged.length())`, and `length()` answers `0L` for a
+     * missing path rather than throwing, so the size line reads `0 B` and no temporary folder is
+     * needed. `routeReason` stays blank, which is what keeps the routing chip out of the tree --
+     * that chip is R38.6's case, not this file's.
+     */
+    private fun converted() = ConversionState.Converted(
+        input = InputFile(
+            uri = Uri.parse("content://test/holiday.mkv"),
+            displayName = "holiday.mkv",
+            sizeBytes = 12_345_678L,
+        ),
+        staged = File("no-such-staged-output.mp4"),
+        suggestedName = "holiday.mp4",
+        mimeType = "video/mp4",
+    )
+
+    private fun setContent(state: ConversionState) {
+        composeRule.setContent {
+            ConverterScreenContent(
+                state = state,
+                settings = ConversionSettings(),
+                validation = Validation.Valid,
+                actions = ConverterActions(
+                    onPickInput = {},
+                    onPreset = {},
+                    onContainer = {},
+                    onVideoCodec = {},
+                    onAudioCodec = {},
+                    onSuggestion = {},
+                    onQuality = {},
+                    onEnginePreference = {},
+                    onConvert = {},
+                    onCancel = {},
+                    onSave = { suggestedName -> savedAs += suggestedName },
+                    onReset = {},
+                ),
+            )
+        }
+    }
+}

app/src/test/java/org/libremediaconverter/convert/ConverterStateAffordancesTest.kt

@@ -0,0 +1,403 @@
+package org.libremediaconverter.convert
+
+import android.net.Uri
+import androidx.compose.ui.semantics.ProgressBarRangeInfo
+import androidx.compose.ui.test.assertIsEnabled
+import androidx.compose.ui.test.assertIsNotEnabled
+import androidx.compose.ui.test.assertRangeInfoEquals
+import androidx.compose.ui.test.assertTextEquals
+import androidx.compose.ui.test.onNodeWithTag
+import androidx.compose.ui.test.onNodeWithText
+import androidx.compose.ui.test.performClick
+import androidx.compose.ui.test.performScrollTo
+import androidx.media3.common.util.UnstableApi
+import org.junit.Assert.assertEquals
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.createDrainedComposeRule
+import org.libremediaconverter.model.AudioCodec
+import org.libremediaconverter.model.Container
+import org.libremediaconverter.model.OutputSpec
+import org.libremediaconverter.model.Validation
+import org.libremediaconverter.model.VideoCodec
+import org.libremediaconverter.ui.TestTags
+import org.robolectric.RobolectricTestRunner
+import java.io.File
+
+/**
+ * Every `ConversionState` renders its own affordances, and only its own.
+ *
+ * The defect this bites on is a `when` arm that has drifted from the state it names: a button
+ * offered in a state where it cannot work, a state's own data never reaching the node that is
+ * supposed to display it, or an affordance wired to the wrong callback. None of that is a compile
+ * error -- every arm of the `when` returns `Unit`, so an arm can render anything at all -- and none
+ * of it is visible from the leaf tests, which compose `FileCard`, `AdvancedPicker` and the three
+ * pickers directly and never see a `ConversionState`.
+ *
+ * The arm most worth guarding is `Ready`'s `enabled = validation.isValid`. The Advanced picker
+ * deliberately lets an impossible container / codec combination be selected -- `AdvancedPicker`'s
+ * KDoc says teaching the constraint beats hiding it -- so that single expression is the only thing
+ * standing between an invalid spec and a job that cannot succeed. `enabled = true` compiles, renders
+ * an identical screen apart from one colour, and passes every other test in this suite.
+ *
+ * Callbacks are asserted by **identity, over the whole log**: [fired] records all twelve of them and
+ * each assertion compares the complete list against one expected entry. A bare "the callback ran"
+ * check stays green when an arm fires the right callback for the wrong reason, and a check on one
+ * callback alone stays green when an arm fires two.
+ *
+ * ### Not asserted here, so that each is a decision rather than an omission
+ *
+ * - **`Failed`'s error colour.** #62's table asks for the message "in the error colour". Compose
+ *   publishes no text colour to the semantics tree -- there is no `SemanticsProperties` entry for
+ *   it -- so it is unobservable from a JVM test, the same limit `FileCardTest` records for
+ *   `HorizontalDivider`. The message text itself is asserted; the colour would need a screenshot.
+ * - **The three `assertDoesNotExist` checks on [TestTags.Converter.FILE_CARD] are compile-guarded,
+ *   not guarded by this file.** `Idle` is a `data object`, and `Saved` and `Failed` carry only a
+ *   `displayName` and a `message`; none of the three has an `input`, so `FileCard(s.input)` does not
+ *   compile in those arms. The lines stay because they state the intent cheaply, but they are not
+ *   what stops a `FileCard` appearing there and this file does not claim they are.
+ * - **Which constant each chip hands back** belongs to `ConverterPickerSelectionTest`, and **what
+ *   the file card says about an unknown size** to `FileCardTest`. This file asserts that `Ready`
+ *   puts those leaves on screen at all, not what they then do.
+ * - **The suggested name `Converted` hands to the save dialog** is pinned by
+ *   `ConverterScreenContentTest`; repeating it here would be a second copy of one assertion.
+ * - **`ConverterScreen`'s permission dance.** `requestNotifications` calls `convert()` on both grant
+ *   and deny, deliberately -- the KDoc explains that the foreground service runs either way -- and
+ *   it lives in the entry point, above the seam this file composes.
+ * - **`is ConversionState.Idle -> Unit` in the nested `when`.** The outer `when` peels `Idle` off
+ *   first, so that arm is permanently unreachable and no test can reach it.
+ */
+@UnstableApi
+@RunWith(RobolectricTestRunner::class)
+class ConverterStateAffordancesTest {
+
+    // Not `createComposeRule()` directly: see [org.libremediaconverter.drainEscapedCoroutineErrors].
+    @get:Rule
+    val composeRule = createDrainedComposeRule()
+
+    /**
+     * Every callback the screen fired, in order, tagged with the value it carried.
+     *
+     * All twelve are recorded rather than only the one under test, so an assertion can be
+     * `assertEquals(listOf("cancel"), fired)` -- which says "this one and nothing else".
+     */
+    private val fired = mutableListOf<String>()
+
+    // -------------------------------------------------------------------- Idle
+
+    @Test
+    fun `an idle screen offers the prompt and the picker, and nothing to act on yet`() {
+        setContent(ConversionState.Idle)
+
+        composeRule.onNodeWithText("Pick a file to convert.").assertExists()
+        composeRule.onNodeWithTag(TestTags.Converter.CHOOSE_FILE).assertExists()
+        composeRule.onNodeWithTag(TestTags.Converter.CONVERT).assertDoesNotExist()
+        composeRule.onNodeWithTag(TestTags.CANCEL).assertDoesNotExist()
+        // Compile-guarded rather than guarded here -- `Idle` has no `input`. See the class KDoc.
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD).assertDoesNotExist()
+    }
+
+    /**
+     * No [performScrollTo] on this one, unlike every other click below. `Idle` is the centred
+     * branch outside the `verticalScroll` column, so it has no scrollable ancestor to scroll in.
+     */
+    @Test
+    fun `tapping choose file on an idle screen asks for a file and does nothing else`() {
+        setContent(ConversionState.Idle)
+
+        composeRule.onNodeWithTag(TestTags.Converter.CHOOSE_FILE).performClick()
+
+        assertEquals(listOf("pickInput"), fired)
+    }
+
+    // ------------------------------------------------------------------- Ready
+
+    /**
+     * All four pickers, the card above them and both buttons below, in one assertion each.
+     *
+     * A superset of #62's "all five pickers": which four or five of these count as a picker is not
+     * worth arguing about, so the case names everything the arm emits.
+     */
+    @Test
+    fun `a picked file offers its card, all four pickers and both buttons`() {
+        setContent(ConversionState.Ready(input()))
+
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD).assertExists()
+        composeRule.onNodeWithTag(TestTags.Converter.FORMAT_CHIPS).assertExists()
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).assertExists()
+        composeRule.onNodeWithTag(TestTags.Converter.QUALITY_CHIPS).assertExists()
+        composeRule.onNodeWithTag(TestTags.Converter.ENGINE_CHIPS).assertExists()
+        composeRule.onNodeWithTag(TestTags.Converter.CONVERT).assertExists()
+        composeRule.onNodeWithTag(TestTags.Converter.CHOOSE_DIFFERENT_FILE).assertExists()
+    }
+
+    /** The card is handed `s.input`, so the name on it is how the state is shown to have arrived. */
+    @Test
+    fun `the file card on a picked file names the file that was picked`() {
+        setContent(ConversionState.Ready(input()))
+
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD_NAME).assertTextEquals("holiday.mkv")
+    }
+
+    @Test
+    fun `convert is offered for a spec that can be produced`() {
+        setContent(ConversionState.Ready(input()), validation = Validation.Valid)
+
+        composeRule.onNodeWithTag(TestTags.Converter.CONVERT).assertIsEnabled()
+    }
+
+    @Test
+    fun `tapping convert starts the job and does nothing else`() {
+        setContent(ConversionState.Ready(input()), validation = Validation.Valid)
+
+        composeRule.onNodeWithTag(TestTags.Converter.CONVERT).performScrollTo().performClick()
+
+        assertEquals(listOf("convert"), fired)
+    }
+
+    /**
+     * The bite named in #62. Reverting `enabled = validation.isValid` to `enabled = true` reddens
+     * exactly this case, and nothing else in the repository.
+     */
+    @Test
+    fun `convert is withheld for a spec that cannot be produced`() {
+        setContent(ConversionState.Ready(input()), validation = INVALID)
+
+        composeRule.onNodeWithTag(TestTags.Converter.CONVERT).assertIsNotEnabled()
+    }
+
+    /** The other button on the arm goes back to the picker rather than starting anything. */
+    @Test
+    fun `tapping choose a different file asks for a file rather than converting`() {
+        setContent(ConversionState.Ready(input()))
+
+        composeRule
+            .onNodeWithTag(TestTags.Converter.CHOOSE_DIFFERENT_FILE)
+            .performScrollTo()
+            .performClick()
+
+        assertEquals(listOf("pickInput"), fired)
+    }
+
+    // -------------------------------------------------------------- Converting
+
+    /**
+     * Two independent readings of the same `percent`, on purpose.
+     *
+     * The heading is a string and the bar is a float, and the arm computes them from the state
+     * separately -- `"${s.percent}%"` against `s.percent / 100f`. A hardcoded bar and a hardcoded
+     * heading are different mistakes, so neither assertion covers the other.
+     */
+    @Test
+    fun `a running job reports how far it has got, in words and on the bar`() {
+        setContent(ConversionState.Converting(input(), percent = 42))
+
+        composeRule.onNodeWithText("Converting… 42%").assertExists()
+        composeRule
+            .onNodeWithTag(TestTags.Converter.PROGRESS)
+            .assertRangeInfoEquals(ProgressBarRangeInfo(0.42f, 0f..1f))
+    }
+
+    @Test
+    fun `a running job offers cancel and not start over`() {
+        setContent(ConversionState.Converting(input(), percent = 42))
+
+        composeRule.onNodeWithTag(TestTags.CANCEL).assertExists()
+        composeRule.onNodeWithTag(TestTags.START_OVER).assertDoesNotExist()
+    }
+
+    @Test
+    fun `tapping cancel on a running job cancels it and does nothing else`() {
+        setContent(ConversionState.Converting(input(), percent = 42))
+
+        composeRule.onNodeWithTag(TestTags.CANCEL).performScrollTo().performClick()
+
+        assertEquals(listOf("cancel"), fired)
+    }
+
+    // ----------------------------------------------------------------- Waiting
+
+    /**
+     * The second bite named in #62. Deleting the `Cancel` button from the `Waiting` arm reddens
+     * this case and the one below it.
+     *
+     * The paragraph is asserted in full rather than by a fragment because it is the only thing the
+     * arm renders besides the card and the button, and because its wording is the arm's whole
+     * job -- `FailureOutcome` records that two different causes land here and the state cannot tell
+     * them apart, so the text has to cover both. A reword should redden one test, and this is it.
+     */
+    @Test
+    fun `a paused job explains why and still offers cancel`() {
+        setContent(ConversionState.Waiting(input()))
+
+        composeRule.onNodeWithText(PAUSED_PARAGRAPH).assertExists()
+        composeRule.onNodeWithTag(TestTags.CANCEL).assertExists()
+    }
+
+    @Test
+    fun `tapping cancel on a paused job cancels it and does nothing else`() {
+        setContent(ConversionState.Waiting(input()))
+
+        composeRule.onNodeWithTag(TestTags.CANCEL).performScrollTo().performClick()
+
+        assertEquals(listOf("cancel"), fired)
+    }
+
+    // --------------------------------------------------------------- Converted
+
+    @Test
+    fun `a finished job offers save and start over, and no longer offers cancel`() {
+        setContent(converted())
+
+        composeRule.onNodeWithTag(TestTags.SAVE_FILE).assertExists()
+        composeRule.onNodeWithTag(TestTags.START_OVER).assertExists()
+        composeRule.onNodeWithTag(TestTags.CANCEL).assertDoesNotExist()
+    }
+
+    @Test
+    fun `tapping start over on a finished job resets and does not save`() {
+        setContent(converted())
+
+        composeRule.onNodeWithTag(TestTags.START_OVER).performScrollTo().performClick()
+
+        assertEquals(listOf("reset"), fired)
+    }
+
+    /**
+     * The chip carries the job's own explanation, so its text is the assertion rather than its
+     * presence: a chip showing the engine name, or the previous job's reason, would still exist.
+     */
+    @Test
+    fun `a finished job shows the routing decision the job reported`() {
+        setContent(converted(routeReason = "Software — the MKV input needed a re-encode"))
+
+        composeRule
+            .onNodeWithTag(TestTags.Converter.ROUTE_REASON)
+            .assertTextEquals("Software — the MKV input needed a re-encode")
+    }
+
+    /** The other side of the `isNotBlank` guard, which is unguarded without a case of its own. */
+    @Test
+    fun `a finished job that reported no routing decision shows no chip`() {
+        setContent(converted(routeReason = ""))
+
+        composeRule.onNodeWithTag(TestTags.Converter.ROUTE_REASON).assertDoesNotExist()
+    }
+
+    // ------------------------------------------------------------------- Saved
+
+    @Test
+    fun `a saved file names itself and offers another conversion`() {
+        setContent(ConversionState.Saved(displayName = "holiday.mp4"))
+
+        composeRule.onNodeWithText("Saved holiday.mp4.").assertExists()
+        composeRule.onNodeWithTag(TestTags.Converter.CONVERT_ANOTHER).assertExists()
+        // Compile-guarded rather than guarded here -- `Saved` has no `input`. See the class KDoc.
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD).assertDoesNotExist()
+    }
+
+    @Test
+    fun `tapping convert another after a save resets and does nothing else`() {
+        setContent(ConversionState.Saved(displayName = "holiday.mp4"))
+
+        composeRule
+            .onNodeWithTag(TestTags.Converter.CONVERT_ANOTHER)
+            .performScrollTo()
+            .performClick()
+
+        assertEquals(listOf("reset"), fired)
+    }
+
+    // ------------------------------------------------------------------ Failed
+
+    /**
+     * The message is the arm's only output that carries information, and it comes from the state.
+     * An arm rendering a fixed apology would look right and say nothing, which is why the assertion
+     * is on the text handed in rather than on a node existing.
+     */
+    @Test
+    fun `a failed job renders the reason it was given and offers a restart`() {
+        setContent(ConversionState.Failed(message = "Ran out of space while writing the output."))
+
+        composeRule.onNodeWithText("Ran out of space while writing the output.").assertExists()
+        composeRule.onNodeWithTag(TestTags.START_OVER).assertExists()
+        composeRule.onNodeWithTag(TestTags.SAVE_FILE).assertDoesNotExist()
+        // Compile-guarded rather than guarded here -- `Failed` has no `input`. See the class KDoc.
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD).assertDoesNotExist()
+    }
+
+    @Test
+    fun `tapping start over after a failure resets and does nothing else`() {
+        setContent(ConversionState.Failed(message = "Ran out of space while writing the output."))
+
+        composeRule.onNodeWithTag(TestTags.START_OVER).performScrollTo().performClick()
+
+        assertEquals(listOf("reset"), fired)
+    }
+
+    // ------------------------------------------------------------------ Harness
+
+    private fun input() = InputFile(
+        uri = Uri.parse("content://test/holiday.mkv"),
+        displayName = "holiday.mkv",
+        sizeBytes = 12_345_678L,
+    )
+
+    /**
+     * `staged` names a path that does not exist, deliberately: `File.length()` answers `0L` for a
+     * missing file rather than throwing, so the size line reads `0 B` and no temporary folder is
+     * needed to render the arm.
+     */
+    private fun converted(routeReason: String = "") = ConversionState.Converted(
+        input = input(),
+        staged = File("no-such-staged-output.mp4"),
+        routeReason = routeReason,
+        suggestedName = "holiday.mp4",
+        mimeType = "video/mp4",
+    )
+
+    private fun setContent(state: ConversionState, validation: Validation = Validation.Valid) {
+        composeRule.setContent {
+            ConverterScreenContent(
+                state = state,
+                settings = ConversionSettings(),
+                validation = validation,
+                actions = ConverterActions(
+                    onPickInput = { fired += "pickInput" },
+                    onPreset = { fired += "preset:$it" },
+                    onContainer = { fired += "container:$it" },
+                    onVideoCodec = { fired += "videoCodec:$it" },
+                    onAudioCodec = { fired += "audioCodec:$it" },
+                    onSuggestion = { fired += "suggestion:$it" },
+                    onQuality = { fired += "quality:$it" },
+                    onEnginePreference = { fired += "engine:$it" },
+                    onConvert = { fired += "convert" },
+                    onCancel = { fired += "cancel" },
+                    onSave = { fired += "save:$it" },
+                    onReset = { fired += "reset" },
+                ),
+            )
+        }
+    }
+
+    private companion object {
+
+        /**
+         * A spec no container can hold, with somewhere to go instead.
+         *
+         * Built here rather than run through `ContainerCapabilities` because what makes a spec
+         * invalid is that class's subject; all this arm needs is a `Validation` that answers
+         * `isValid == false`.
+         */
+        val INVALID = Validation.Invalid(
+            message = "WebM cannot hold H.264 video.",
+            suggestions = listOf(OutputSpec(Container.MKV, VideoCodec.H264, AudioCodec.AAC)),
+        )
+
+        /** Copied from the `Waiting` arm, where it is written as two concatenated fragments. */
+        const val PAUSED_PARAGRAPH =
+            "Paused. Android limits background media processing, so this will " +
+                "resume automatically — keeping the app open helps it along."
+    }
+}

app/src/test/java/org/libremediaconverter/convert/OutputPublisherStagingTest.kt

@@ -18,8 +18,10 @@ import java.util.UUID
  * the actual filesystem — the same calls `reset()` makes, without needing a ViewModel (both
  * of those construct a `WorkManager`, which is not initialised on the JVM classpath).
  *
- * The instrumented suite cannot run on the development host, so this is the only place the
- * "Start over leaks a full-size copy" defect can be caught before CI.
+ * The instrumented suite could also catch the "Start over leaks a full-size copy" defect --
+ * it runs on this host for API 33-36 (`tools/local-emulator/run-e2e.sh`) and on CI for
+ * 33-37. Here rather than there because a real `cacheDir` is all the defect needs, and
+ * finding it costs an emulator boot there and a few seconds here.
  */
 @RunWith(RobolectricTestRunner::class)
 class OutputPublisherStagingTest {

app/src/test/java/org/libremediaconverter/join/JoinScreenContentTest.kt

@@ -0,0 +1,83 @@
+package org.libremediaconverter.join
+
+import androidx.compose.ui.test.onNodeWithTag
+import androidx.compose.ui.test.performClick
+import androidx.compose.ui.test.performScrollTo
+import androidx.media3.common.util.UnstableApi
+import org.junit.Assert.assertEquals
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.createDrainedComposeRule
+import org.libremediaconverter.model.ConcatStrategy
+import org.libremediaconverter.ui.TestTags
+import org.robolectric.RobolectricTestRunner
+import java.io.File
+
+/**
+ * The join screen's half of the same seam, and the same two directions.
+ *
+ * The defect is the one `ConverterScreenContentTest` describes -- a content composable that
+ * ignores the state handed to it, or renders the finished job's affordances unwired -- and it has
+ * to be asked separately here because the two screens share no code. `JoinScreen` and
+ * `ConverterScreen` were extracted in the same commit by the same hand, which is exactly the
+ * circumstance in which one of them gets the wiring right and the other does not.
+ *
+ * `JoinState.Joined` is unreachable through a real `JoinViewModel` for the same reason
+ * `ConversionState.Converted` is: only a `ConcatWorker` run that has already succeeded produces
+ * one, carrying the strategy it chose and the name it picked.
+ *
+ * `JoinScreenKt` is the honest remaining coverage gap on this repo, and closing it is R38.7 (#63),
+ * not this file. Which affordances each `JoinState` renders belongs there; this asserts only that
+ * the injection point exists.
+ */
+@UnstableApi
+@RunWith(RobolectricTestRunner::class)
+class JoinScreenContentTest {
+
+    // Not `createComposeRule()` directly: see [org.libremediaconverter.drainEscapedCoroutineErrors].
+    @get:Rule
+    val composeRule = createDrainedComposeRule()
+
+    /** What the screen asked to save, in the order it asked. Empty until Save is tapped. */
+    private val savedAs = mutableListOf<String>()
+
+    @Test
+    fun `a finished join renders the save button`() {
+        setContent(joined())
+
+        composeRule.onNodeWithTag(TestTags.SAVE_FILE).assertExists()
+    }
+
+    @Test
+    fun `tapping save hands back the name the finished join chose`() {
+        setContent(joined())
+
+        composeRule.onNodeWithTag(TestTags.SAVE_FILE).performScrollTo().performClick()
+
+        assertEquals(listOf("joined.mp4"), savedAs)
+    }
+
+    /** `staged` names a missing file deliberately -- see the same helper on the converter side. */
+    private fun joined() = JoinState.Joined(
+        staged = File("no-such-staged-output.mp4"),
+        strategy = ConcatStrategy.STREAM_COPY,
+        suggestedName = "joined.mp4",
+        mimeType = "video/mp4",
+    )
+
+    private fun setContent(state: JoinState) {
+        composeRule.setContent {
+            JoinScreenContent(
+                state = state,
+                actions = JoinActions(
+                    onPickInputs = {},
+                    onJoin = {},
+                    onCancel = {},
+                    onSave = { suggestedName -> savedAs += suggestedName },
+                    onReset = {},
+                ),
+            )
+        }
+    }
+}

app/src/test/java/org/libremediaconverter/join/JoinStateAffordancesTest.kt

@@ -0,0 +1,271 @@
+package org.libremediaconverter.join
+
+import android.net.Uri
+import androidx.compose.ui.semantics.ProgressBarRangeInfo
+import androidx.compose.ui.semantics.SemanticsProperties
+import androidx.compose.ui.semantics.getOrNull
+import androidx.compose.ui.test.SemanticsMatcher
+import androidx.compose.ui.test.assertRangeInfoEquals
+import androidx.compose.ui.test.assertTextEquals
+import androidx.compose.ui.test.onNodeWithTag
+import androidx.compose.ui.test.onNodeWithText
+import androidx.compose.ui.test.performClick
+import androidx.compose.ui.test.performScrollTo
+import androidx.media3.common.util.UnstableApi
+import org.junit.Assert.assertEquals
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.convert.InputFile
+import org.libremediaconverter.createDrainedComposeRule
+import org.libremediaconverter.model.ConcatStrategy
+import org.libremediaconverter.ui.TestTags
+import org.robolectric.RobolectricTestRunner
+import java.io.File
+
+/**
+ * Every `JoinState` renders its own affordances, wired to its own callback.
+ *
+ * The defect is a branch of `JoinScreenContent`'s `when` that reads the wrong thing: a count taken
+ * from a literal rather than from `inputs`, a strategy line that describes the other strategy, a
+ * button wired to the neighbouring branch's callback, a `Failed` that drops the message it carries.
+ * None of that is visible at compile time -- every branch of the `when` type-checks against the
+ * same `JoinScreenContent` signature -- and none of it is visible from the leaf tests either, which
+ * compose `FileRow` on its own and never see a state.
+ *
+ * `JoinScreenContentTest` deliberately asks only whether the seam exists, using `Joined`. This is
+ * the matrix behind it: seven states, each pinned to what it lets the user do next.
+ *
+ * ### Two assertions here that nothing else in the suite makes
+ *
+ * **Order.** A join is the one flow where the order of the inputs is the content of the output --
+ * the empty state promises "in the order you want them" -- so the rows are read back sorted by
+ * their position on screen and compared as a list, not as a set. `JoinLeafTagsTest` proves a row
+ * tags itself with the file it shows; nothing proved the rows come out in the order they went in.
+ *
+ * **Indeterminate.** The join progress bar carries no percentage, on purpose: FFmpeg reports
+ * progress against one input's duration, which means nothing across a concatenation. The converter
+ * screen's bar is determinate, so "it has a progress bar" is the assertion that would not notice a
+ * fabricated percentage arriving here.
+ *
+ * ### Not asserted here, deliberately
+ *
+ * `JoinState.Joined.mimeType` is not rendered by this composable at all -- it is read by the entry
+ * point, to open the save dialog with a type that matches the finished job. The colour of the
+ * `Failed` message is `MaterialTheme.colorScheme.error`, which is theme lookup rather than state
+ * logic, so it is left to the eye. The `is JoinState.Idle -> Unit` arm inside the scrolling branch
+ * is unreachable by construction: the outer `when` peels `Idle` off first.
+ */
+@UnstableApi
+@RunWith(RobolectricTestRunner::class)
+class JoinStateAffordancesTest {
+
+    // Not `createComposeRule()` directly: see [org.libremediaconverter.drainEscapedCoroutineErrors].
+    @get:Rule
+    val composeRule = createDrainedComposeRule()
+
+    /** Which callback the screen invoked, in order, with what it passed. Empty until one fires. */
+    private val events = mutableListOf<String>()
+
+    @Test
+    fun `the empty state asks for files in order and offers the picker`() {
+        setContent(JoinState.Idle)
+
+        composeRule.onNodeWithText("Pick two or more files to join, in the order you want them.").assertExists()
+        // No `performScrollTo` on this one: `Idle` is the centred branch, outside the scrolling
+        // column every other state renders into, so there is nothing to scroll.
+        composeRule.onNodeWithTag(TestTags.Join.CHOOSE_FILES).performClick()
+
+        assertEquals(listOf("pickInputs"), events)
+    }
+
+    /**
+     * The rows come out in the order the inputs went in.
+     *
+     * Sorted by position rather than trusting the order `fetchSemanticsNodes` happens to return, so
+     * the assertion is about what the user sees down the screen. Three inputs, with names whose
+     * alphabetical order is not their picked order, so a list that had been sorted anywhere on the
+     * way through would not be able to pass this.
+     */
+    @Test
+    fun `the picked inputs are listed in the order they were picked`() {
+        val picked = listOf("intro.mp4", "middle.mp4", "outro.mp4")
+        setContent(JoinState.Ready(inputs = picked.map(::input)))
+
+        val topToBottom = composeRule.onAllNodes(isFileRow)
+            .fetchSemanticsNodes()
+            .sortedBy { it.positionInRoot.y }
+            .map { it.config[SemanticsProperties.TestTag] }
+
+        assertEquals(picked.map(TestTags.Join::fileRow), topToBottom)
+    }
+
+    /**
+     * Three inputs, not two: two is the minimum a join accepts, so a button that had been
+     * hardcoded to the smallest legal join would still read correctly with two on screen.
+     */
+    @Test
+    fun `the join button counts the files it will join`() {
+        setContent(JoinState.Ready(inputs = listOf(input("intro.mp4"), input("middle.mp4"), input("outro.mp4"))))
+
+        composeRule.onNodeWithTag(TestTags.Join.JOIN).assertTextEquals("Join 3 files")
+        composeRule.onNodeWithTag(TestTags.Join.JOIN).performScrollTo().performClick()
+
+        assertEquals(listOf("join"), events)
+    }
+
+    /** `Ready` is the one working state that still offers the picker, to replace the selection. */
+    @Test
+    fun `a ready join can be repicked`() {
+        setContent(JoinState.Ready(inputs = listOf(input("intro.mp4"), input("outro.mp4"))))
+
+        composeRule.onNodeWithTag(TestTags.Join.CHOOSE_DIFFERENT_FILES).performScrollTo().performClick()
+
+        assertEquals(listOf("pickInputs"), events)
+    }
+
+    @Test
+    fun `a running join names the count and shows a bar with no percentage`() {
+        setContent(JoinState.Joining(inputs = listOf(input("intro.mp4"), input("outro.mp4"))))
+
+        composeRule.onNodeWithText("Joining 2 files…").assertExists()
+        composeRule.onNodeWithTag(TestTags.Join.PROGRESS).assertRangeInfoEquals(ProgressBarRangeInfo.Indeterminate)
+        composeRule.onNodeWithTag(TestTags.CANCEL).performScrollTo().performClick()
+
+        assertEquals(listOf("cancel"), events)
+    }
+
+    /**
+     * The paragraph is byte-identical to the converter screen's, which is the point of asserting
+     * the whole of it rather than a fragment: the two branches were worded together, and a reword
+     * that lands on one screen only is the failure this notices.
+     */
+    @Test
+    fun `a paused join explains itself and still offers cancel`() {
+        setContent(JoinState.Waiting(inputs = listOf(input("intro.mp4"), input("outro.mp4"))))
+
+        composeRule.onNodeWithText(PAUSED_PARAGRAPH).assertExists()
+        composeRule.onNodeWithTag(TestTags.CANCEL).performScrollTo().performClick()
+
+        assertEquals(listOf("cancel"), events)
+    }
+
+    @Test
+    fun `a stream copied join says nothing was re-encoded`() {
+        setContent(joined(ConcatStrategy.STREAM_COPY))
+
+        composeRule.onNodeWithText(STREAM_COPY_EXPLANATION).assertExists()
+        composeRule.onNodeWithText(REENCODE_EXPLANATION).assertDoesNotExist()
+    }
+
+    /**
+     * The other half of the pair. Asserting the absence of the stream-copy line as well, because a
+     * branch that had collapsed to one answer would still render *an* explanation.
+     */
+    @Test
+    fun `a re-encoded join says the files differed`() {
+        setContent(joined(ConcatStrategy.REENCODE))
+
+        composeRule.onNodeWithText(REENCODE_EXPLANATION).assertExists()
+        composeRule.onNodeWithText(STREAM_COPY_EXPLANATION).assertDoesNotExist()
+    }
+
+    /** The size comes from the staged file, which is missing here, so `length()` answers `0L`. */
+    @Test
+    fun `a finished join reports the size of what it produced`() {
+        setContent(joined(ConcatStrategy.STREAM_COPY))
+
+        composeRule.onNodeWithText("Joined — 0 MB.").assertExists()
+    }
+
+    @Test
+    fun `a finished join offers save and start over, and they are not the same button`() {
+        setContent(joined(ConcatStrategy.STREAM_COPY))
+
+        composeRule.onNodeWithTag(TestTags.SAVE_FILE).performScrollTo().performClick()
+        composeRule.onNodeWithTag(TestTags.START_OVER).performScrollTo().performClick()
+
+        assertEquals(listOf("save:joined.mp4", "reset"), events)
+    }
+
+    @Test
+    fun `a saved join names the file and offers to join more`() {
+        setContent(JoinState.Saved(displayName = "holiday-joined.mp4"))
+
+        composeRule.onNodeWithText("Saved holiday-joined.mp4.").assertExists()
+        composeRule.onNodeWithTag(TestTags.Join.JOIN_MORE).assertTextEquals("Join more")
+        composeRule.onNodeWithTag(TestTags.Join.JOIN_MORE).performScrollTo().performClick()
+
+        assertEquals(listOf("reset"), events)
+    }
+
+    /**
+     * The message is the whole content of this state -- it is the only thing that says why the job
+     * stopped -- and it arrives as a string the failure produced, so a branch that rendered a fixed
+     * apology instead would look correct on screen.
+     */
+    @Test
+    fun `a failed join renders the message it carries`() {
+        setContent(JoinState.Failed(message = "The second file has no audio track, so joining stopped."))
+
+        composeRule.onNodeWithText("The second file has no audio track, so joining stopped.").assertExists()
+    }
+
+    @Test
+    fun `a failed join offers start over`() {
+        setContent(JoinState.Failed(message = "The second file has no audio track, so joining stopped."))
+
+        composeRule.onNodeWithTag(TestTags.START_OVER).performScrollTo().performClick()
+
+        assertEquals(listOf("reset"), events)
+    }
+
+    /** Anything `FileRow` tagged, whichever file it is showing. The prefix comes from the table. */
+    private val isFileRow = SemanticsMatcher("is a join file row") { node ->
+        node.config.getOrNull(SemanticsProperties.TestTag)?.startsWith(TestTags.Join.fileRow("")) == true
+    }
+
+    private fun input(displayName: String) = InputFile(
+        uri = Uri.parse("content://test/$displayName"),
+        displayName = displayName,
+        sizeBytes = 4_000_000L,
+    )
+
+    /** `staged` names a missing file deliberately -- see the same helper in `JoinScreenContentTest`. */
+    private fun joined(strategy: ConcatStrategy) = JoinState.Joined(
+        staged = File("no-such-staged-output.mp4"),
+        strategy = strategy,
+        suggestedName = "joined.mp4",
+        mimeType = "video/mp4",
+    )
+
+    private fun setContent(state: JoinState) {
+        composeRule.setContent {
+            JoinScreenContent(
+                state = state,
+                actions = JoinActions(
+                    onPickInputs = { events += "pickInputs" },
+                    onJoin = { events += "join" },
+                    onCancel = { events += "cancel" },
+                    onSave = { suggestedName -> events += "save:$suggestedName" },
+                    onReset = { events += "reset" },
+                ),
+            )
+        }
+    }
+
+    private companion object {
+        /** Byte-identical to the converter screen's, and split the same way `main` splits it. */
+        const val PAUSED_PARAGRAPH =
+            "Paused. Android limits background media processing, so this will " +
+                "resume automatically — keeping the app open helps it along."
+
+        const val STREAM_COPY_EXPLANATION =
+            "Files matched, so they were joined without " +
+                "re-encoding — no quality loss."
+
+        const val REENCODE_EXPLANATION =
+            "Files differed in format, so they were re-encoded " +
+                "to match."
+    }
+}

docs/api-37-emulator-crash.md

@@ -230,11 +230,17 @@ booted (`emulator_alive=yes`). The host emulator is fine; the guest is not.
 
 ## Can the suite run on it?
 
-**Almost.** `tools/local-emulator/run-e2e.sh 37` now runs the whole suite locally, and all of it
-passes except two tests. Measured at `22c7914`: **49 tests, 2 failures, 0 errors, 2 skipped** — 45
-passed, the two `Media3EngineTest` failures dissected below, and the two `assumeTrue` skips every
-level has. It costs two deviations from how every other level is run, and both are worth
-understanding before trusting the leg.
+**Almost, and less so than it was.** `tools/local-emulator/run-e2e.sh 37` runs the whole suite
+locally. Measured at `22c7914`: **49 tests, 2 failures, 0 errors, 2 skipped** — 45 passed, the two
+`Media3EngineTest` failures dissected below, and the two `assumeTrue` skips every level has. It
+costs two deviations from how every other level is run, and both are worth understanding before
+trusting the leg.
+
+**That was the high-water mark.** On 2026-08-24 a test that touches system UI joined the suite,
+and the level stopped *finishing* rather than merely failing two —
+[see below](#something-does-depend-on-system-ui-now-and-it-is-excluded-rather-than-trusted).
+Two `Media3EngineTest` failures is what **CI's gating leg** expects, because it filters on
+`notAnnotation`; a local `run-e2e.sh 37` does not filter and sees more.
 
 Two things about that total before it is compared with anything. It is the size of the suite on
 the checkout that ran, not a property of API 37 — `app/src/androidTest` held 49 `@Test` methods at
@@ -320,10 +326,60 @@ and proceeding straight to the tests fails exactly as before. The harness theref
 1. **The renderer is ANGLE, not the host GPU.** Shared with nothing else in the matrix — API
    33–36 run `-gpu host` locally, and CI runs `swiftshader_indirect`.
 2. **SystemUI is disabled.** The API 37 leg does not run the same device configuration as any
-   other leg or as the Pixel. It is defensible here only because nothing in this suite touches
-   system UI — these are Media3, FFmpeg and WorkManager tests — and because the alternative is no
-   local API 37 coverage at all. **Anything that ever does depend on system UI must not trust
-   this leg.**
+   other leg or as the Pixel. It was defensible here because nothing in this suite touched
+   system UI — Media3, FFmpeg and WorkManager tests — and because the alternative is no local
+   API 37 coverage at all. **Anything that ever does depend on system UI must not trust this
+   leg.** Something now does; see the section below.
+
+### Something does depend on system UI now, and half of it is excluded
+
+Added 2026-08-24, and the first entry on this page that is not a codec.
+
+`SafPickerRoundTripTest` drives the real system file picker and rotates the display. Both reach
+the gralloc mapper — DocumentsUI is another app's windows, and a rotation rebuilds every surface
+on screen — and **disabling SystemUI does not help**, because it removes the *idle* trigger
+(RegionSamplingThread's nav-bar luma sampling) and not this one.
+
+Measured one method per fresh emulator, `android-37.0`, `swangle_indirect`, SystemUI disabled and
+verified quiet — separately, because inferring the second from the first is the mistake this
+page's opening correction is about:
+
+| test | result on android-37.0 | `hasReadColorBufferDma` aborts in the window |
+|---|---|---|
+| `thePickedInputSurvivesARealRotation` | **fails**: `INSTRUMENTATION_ABORTED: System has crashed.`, `Expected 1 tests, received 0`. The framework dies **during** it, so the JUnit XML carries a failure with no text at all. | 3 |
+| `pickingAFileThroughTheSystemPickerFillsInTheFileCard` | **passes** | 4 |
+
+So a rotation, which rebuilds every surface at once, is what the mapper does not survive. Merely
+starting DocumentsUI is not. Only the rotation test carries `@FailsOnEmulatorApi37`; the picker
+test runs on the gating leg like anything else.
+
+#### The correction that produced that table
+
+**The first version of this section said both tests failed, and put the marker on the class.** The
+picker test had indeed failed at API 37 — with `androidx.test.uiautomator.StaleObjectException`,
+which looked like a framework restart invalidating an accessibility node, because that is exactly
+what it looks like.
+
+It was the test's own bug. `UiObject2` caches the `AccessibilityNodeInfo` it was found with, and
+DocumentsUI is still settling when a node first appears; the handle went stale before `click()`.
+CI then reproduced it **deterministically** at API 33, 34 and 35 — every cold runner emulator, not
+intermittently — which is what made it obviously not an API 37 property. It had passed locally
+only because the emulator was warm.
+
+The lesson is worth more than the measurement: **an annotation is a claim about an image, and a
+broken test makes every image look broken.** Re-measure after fixing a test before deciding what
+the platform did. Both the abort and the stale node produce "the run fell over", and only one of
+them was the image.
+
+#### Two consequences worth stating rather than discovering
+
+- **`run-e2e.sh 37` applies no annotation filter**, unlike CI, so a local API 37 run includes the
+  rotation test and therefore **does not finish**: its totals come back short and which later
+  tests ran is arbitrary. The summary row says so.
+- **The advisory job is still named `E2E API 37 Media3 hardware transcode (advisory)`** and now
+  carries a test that is neither Media3 nor a transcode. Renaming a check is a branch-protection
+  change and was deliberately not made in the same PR; the name is stale, the behaviour is
+  correct.
 
 ### The two remaining failures are the same bug, one layer down
 
@@ -650,9 +706,15 @@ though a new API level shipped. Watch for these instead:
 - **`E2E API 37 Media3 hardware transcode (advisory)` going green.** Nothing announces this: the
   job is `continue-on-error`, so it fixing itself looks exactly like a check nobody reads
   quietly ceasing to be red. It is listed here because that makes it the *least* likely of these
-  triggers to be noticed, not the most. When it happens, delete `@FailsOnEmulatorApi37` from the
-  two tests rather than the job — the gating leg picks them back up on its own, and the advisory
-  job then runs nothing and can go.
+  triggers to be noticed, not the most. When it happens, delete `@FailsOnEmulatorApi37` from
+  everything carrying it rather than deleting the job — the gating leg picks them back up on its
+  own, and the advisory job then runs nothing and can go.
+
+  **It is not two tests any more.** As of 2026-08-24 the marker is on `Media3EngineTest`'s two
+  methods *and* on `SafPickerRoundTripTest` as a class, and the two groups fail for unrelated
+  reasons — a codec and the gralloc mapper. They can go green independently, so check both before
+  concluding the marker is done; and the job's name still says "Media3 hardware transcode", which
+  half of what it runs is not.
 
 ## Correction owed to `CLAUDE.md`
 

gradle/libs.versions.toml

@@ -42,6 +42,18 @@ annotation = "1.+"
 junit = "4.+"
 androidxJunit = "1.+"
 espressoCore = "3.+"
+# UiAutomator. FLOATING, and the argument for it is the one the guard already makes:
+# androidx.test.uiautomator is inside `floatedGroupPrefixes` ("androidx."), so `2.+` reads
+# as "the newest RELEASED 2.x" exactly the way `work = "2.+"` does -- and this library does
+# publish alphas above its stable, so without the guard it would be a pin.
+#
+# Not pinned like ktlint/detekt/robolectric, because it is not that kind of dependency. Those
+# are pinned because a new *rule* or a new *runtime* makes untouched files fail -- the tool
+# changes its verdict on code nobody edited. UiAutomator has no verdict: it clicks what a
+# selector names, and a selector that stops matching is this repo's test to fix, in a diff
+# that says so. `2.` and not bare `+` because 3.x does not exist yet and a major is where the
+# selector API would be free to change under exactly that assumption.
+uiautomator = "2.+"
 # PINNED, unlike its neighbours. Under semver a 0.x minor is allowed to break, and
 # this library is load-bearing exactly where breakage is hardest to see: the wrapper
 # reaches for smartexception.java.Exceptions only when an FFmpeg call FAILS, so a
@@ -139,6 +151,11 @@ junit = { group = "junit", name = "junit", version.ref = "junit" }
 androidx-junit = { group = "androidx.test.ext", name = "junit", version.ref = "androidxJunit" }
 androidx-espresso-core = { group = "androidx.test.espresso", name = "espresso-core", version.ref = "espressoCore" }
 
+# The only way to touch UI this app does not own. Compose's own matchers stop at this
+# process's composition, and the system file picker is a DocumentsUI activity in another
+# process -- so a SAF round trip is unreachable without it.
+androidx-uiautomator = { group = "androidx.test.uiautomator", name = "uiautomator", version.ref = "uiautomator" }
+
 # Robolectric — an Android runtime for the JVM test source set, so file-lifecycle behaviour
 # that needs a real Context can be verified without a device. The instrumented suite cannot
 # run on the development host at all (see CLAUDE.md), so an androidTest-only red test is not

tools/local-emulator/run-e2e.sh

@@ -15,11 +15,13 @@
 # EXIT CODE: 0 only if every level was green; 1 if any level failed, wedged or could not be
 # set up; 2 if it refused to start at all. **A bare `run-e2e.sh` therefore exits 1 by design.**
 # API 37 is in the default list on purpose -- leaving it out is what left the level unlooked-at
-# for as long as it was -- and it is permanently two failures short of green, on the emulator's
-# own c2.goldfish.h264.decoder rather than on anything this app does. The summary names the two,
-# so a third is visibly new, and the last line printed says the same thing. Anything that reads a
-# non-zero exit as breakage should name the levels it wants: `run-e2e.sh 33 34 35 36` is the
-# sweep that can be green. docs/api-37-emulator-crash.md has the measurements.
+# for as long as it was -- and it is permanently short of green, on the emulator image rather
+# than on anything this app does. Since 2026-08-24 it does not even FINISH: one of its expected
+# failures kills the framework, so the totals come back short with an arbitrary tail. The summary
+# names every failure it expects, so an unnamed one is visibly new, and the last line printed
+# says the same thing. Anything that reads a non-zero exit as breakage should name the levels it
+# wants: `run-e2e.sh 33 34 35 36` is the sweep that can be green.
+# docs/api-37-emulator-crash.md has the measurements.
 #
 # WHY THIS EXISTS, AND WHAT IT DELIBERATELY DOES NOT DO
 #
@@ -320,9 +322,20 @@ boot_emulator() {
 #
 # THIS IS A DEVIATION, and it is deliberately loud rather than silent. The API 37 leg does not
 # run the same device configuration as API 33-36 or as the Pixel. It is defensible only
-# because nothing in this suite touches SystemUI -- these are Media3, FFmpeg and WorkManager
-# tests -- and because the alternative is no API 37 coverage at all. Anything that ever does
-# depend on system UI must not trust this leg. docs/api-37-emulator-crash.md explains why.
+# because nothing in this suite touched SystemUI -- Media3, FFmpeg and WorkManager tests --
+# and because the alternative is no API 37 coverage at all. Anything that ever does depend on
+# system UI must not trust this leg. docs/api-37-emulator-crash.md explains why.
+#
+# "Touched", past tense, since 2026-08-24. SafPickerRoundTripTest drives DocumentsUI and rotates
+# the display, and both reach the gralloc mapper these images abort in -- disabling SystemUI
+# removes the IDLE trigger, not those. Measured per method on android-37.0: the ROTATION test
+# takes the framework down (INSTRUMENTATION_ABORTED) and carries @FailsOnEmulatorApi37; the
+# picker test passes.
+#
+# THIS SCRIPT APPLIES NO ANNOTATION FILTER, unlike CI, so a local `run-e2e.sh 37` runs the
+# rotation test anyway -- and because that test kills the framework rather than merely failing,
+# THE LEVEL DOES NOT FINISH. Its totals come back short and which later tests ran is arbitrary.
+# CI's gating leg never sees it.
 #
 # The retry loop is not defensive padding: at the moment boot_completed flips, the framework
 # may be in one of its restarts and `pm` is simply not published yet. The first attempt at this
@@ -564,13 +577,25 @@ for api in "${APIS[@]}"; do
   guest_forensics "$api"
   # API 37 is in the default list on purpose, and it is expected to be red. Leaving it out would
   # put the level back where this whole exercise found it -- untested and unlooked-at -- but a
-  # summary that just says "2 failures" with no explanation trains people to ignore the exit
-  # code. So the row says which two, and a THIRD failure is then obviously new.
+  # summary that just says "N failures" with no explanation trains people to ignore the exit
+  # code. So the row NAMES the expected ones, and anything else is then obviously new.
+  #
+  # The list grew on 2026-08-24 and the shape of the row changed with it. The two
+  # Media3EngineTest failures are a codec; the third is the gralloc bug reached through system
+  # UI, and it takes the framework DOWN rather than merely failing -- so the level does not
+  # finish, and the totals come back SHORT (50 of 59 when this was written) with the later
+  # tests never run. A run whose totals do not add up is expected here now, which it never
+  # was before.
   case "$api" in
     37 | 37.*)
       line="$line
-    expected here: 2 failures, both Media3EngineTest, on c2.goldfish.h264.decoder.
-    A third is new -- docs/api-37-emulator-crash.md"
+    expected here: 2 Media3EngineTest failures on c2.goldfish.h264.decoder, plus
+    SafPickerRoundTripTest.thePickedInputSurvivesARealRotation -- which kills the framework
+    rather than merely failing, so the run ABORTS partway and the total comes back SHORT with
+    an arbitrary tail. That is expected here too, and never was before. Anything else is new.
+    CI's gating leg sees only the first two: the rotation test carries @FailsOnEmulatorApi37
+    and this script, unlike CI, applies no annotation filter.
+    docs/api-37-emulator-crash.md"
       ;;
   esac
   if [ "$rc" -ne 0 ]; then
@@ -596,8 +621,9 @@ echo "=============================================================="
 # worse than no note at all.
 if [ "$overall" -ne 0 ] && [ "$NON37_RED" -eq 0 ]; then
   echo "note: the only level that went red is API 37, which exits non-zero by design -- it is"
-  echo "      permanently 2 failures short of green. Confirm its row above shows exactly those"
-  echo "      two and nothing else; docs/api-37-emulator-crash.md says why they are the image."
+  echo "      permanently short of green, and since 2026-08-24 it does not even finish. Confirm"
+  echo "      its row above names every failure it shows; docs/api-37-emulator-crash.md says why"
+  echo "      each of them is the image rather than this app."
 fi
 
 exit "$overall"