Merge branch 'main' into test/release-permission-guard

Declare build.yml's token reach in build.yml

.github/workflows/build.yml

@@ -13,6 +13,17 @@ on:
 # reference amounts to running whatever that repository contains tomorrow. This matters
 # more here than on pull requests: these jobs sign nothing today, but they do publish
 # the artifacts people install.
+# Declared here rather than inherited, for the reason status_check.yml gives for its own
+# block: the token's reach should be readable in the file that uses it, and a repository
+# default that widens later should not silently widen these jobs with it. The repository
+# default is `read` today, so this changes nothing about what runs -- it fixes what a
+# reader can know without leaving the file, and it is what CodeQL alert #1 asked for.
+#
+# The `release` job below overrides this with `contents: write`, which is how job-level
+# permissions work: this is a default, not a ceiling.
+permissions:
+  contents: read
+
 env:
   GRADLE_CACHE_PATHS: |
     ~/.gradle/caches
@@ -81,7 +92,11 @@ jobs:
 
       - name: Verify the released artifacts
         run: |
-          APK=$(ls app/build/outputs/apk/release/*.apk | head -1)
+          # A glob, not `ls | head`: the glob is already here, and parsing ls is what
+          # SC2012 is about. Gradle's names have no spaces today, which is exactly the
+          # kind of assumption that holds until it does not.
+          apks=(app/build/outputs/apk/release/*.apk)
+          APK="${apks[0]}"
           # A release that shipped one ABI, or lost 16 KB alignment, would install
           # fine on a test device and fail for users or at Play submission. Both are
           # cheap to check and expensive to discover later.

.github/workflows/status_check.yml

@@ -182,6 +182,23 @@ jobs:
           docker run --rm "$SHELLCHECK" --version
           git ls-files -z '*.sh' | xargs -0 -r docker run --rm -v "$PWD:/mnt" "$SHELLCHECK"
 
+      # actionlint closes the half shellcheck cannot see. The step above reads .sh files;
+      # a good deal of this repo's bash lives in inline `run:` blocks instead -- the release
+      # verification here, the emulator setup and teardown in this file and in
+      # api37-debug.yml. actionlint parses each workflow and runs shellcheck over every
+      # `run:`, on top of its own checks for expression syntax, `needs:` references, matrix
+      # keys and action input names.
+      #
+      # Pinned by digest for the same reason shellcheck is, and with a second reason of its
+      # own: actionlint's documented install is `bash <(curl -s .../download-actionlint.bash)`
+      # off a moving branch, which would sit badly in a repo that pins every action by SHA.
+      - name: actionlint
+        env:
+          ACTIONLINT: rhysd/actionlint@sha256:9d36088643581e728c969f35141f88139fec77280b2be23c1f66f8e40e1025e7
+        run: |
+          docker run --rm "$ACTIONLINT" -version
+          docker run --rm -v "$PWD:/repo" -w /repo "$ACTIONLINT" -color
+
       # `!cancelled()` rather than a plain sequence: a shellcheck failure above must not
       # cost the ktlint/detekt/lint lists. Same reason this step passes --continue -- one
       # round trip should produce every list, not stop at the first.
@@ -263,8 +280,13 @@ jobs:
           # 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.
+          # api-level must be a POINT release. A bare 37 is not an SDK package and
+          # fails during setup, which cost a run to discover. `37.0` is the choice
+          # here rather than the only option: `37.1` and `37.2-beta*` exist and
+          # abort the same way, and api37-debug.yml's inputs document both, with
+          # the wrinkle that above 37.0 they ship only as google_apis_ps16k.
+          # docs/api-37-emulator-crash.md measures 37.0 rev 6 and 37.1 rev 8 side
+          # by side, so pinning 37.0 is a decision, not a constraint.
           #
           # 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

CLAUDE.md

@@ -185,11 +185,12 @@ install for code that can never run — and on API 37 the full APK does not fit
   `podman run --rm -v "$PWD:/mnt:z" docker.io/koalaman/shellcheck@sha256:61862eba... <files>`
   (the digest is in `status_check.yml`; there is no shellcheck system package on this host).
 
-  **It does not cover inline `run:` blocks in the workflows**, and a good deal of this repo's bash
-  lives there. `actionlint` does cover them — it runs shellcheck over each `run:` — and reports one
-  pre-existing `info` finding in `build.yml`. It is not wired in because every action here is
-  pinned by SHA, and actionlint's usual installer is a `curl | bash` off a moving branch; doing it
-  properly means pinning a container digest. Tracked separately rather than bolted on.
+  **`actionlint` covers the half shellcheck cannot see** — the inline `run:` blocks, where a good
+  deal of this repo's bash lives. It runs shellcheck over each `run:` plus its own checks on
+  expression syntax, `needs:` references, matrix keys and action inputs. It sits in the same job,
+  **pinned by digest** for the reason above and one of its own: its documented installer is a
+  `curl | bash` off a moving branch, which does not belong in a repo that pins every action by SHA.
+  Locally: `podman run --rm -v "$PWD:/repo:z" -w /repo docker.io/rhysd/actionlint@sha256:9d360886... -color`.
 
 ## Dependency versions
 

README.md

@@ -113,7 +113,14 @@ container × codec matrix — including combinations that cannot work, which it
 offers alternatives for rather than hiding.
 
 Conversions run as durable background work, so they survive leaving the app and are
-restored after a restart.
+restored when you reopen it.
+
+One case is not restored, and it is worth knowing about. If Android refuses to let a job
+restart in the background, it is retried on an exponential backoff for about eight and a
+half hours and then given up on — and a job that has been given up on does not come back
+when you reopen the app. Reopening the app is what grants permission to run, so a
+conversion that has stalled this way is best started again from the app rather than waited
+on.
 
 ## Building
 

app/build.gradle.kts

@@ -1,3 +1,4 @@
+import org.gradle.api.tasks.PathSensitivity
 import org.gradle.testing.jacoco.tasks.JacocoReport
 
 plugins {
@@ -206,6 +207,16 @@ detekt {
 // `excludes` is not optional. Without it JaCoCo walks JDK-internal classes that Robolectric has
 // no location for either, and the test JVM dies rather than reporting a number.
 tasks.withType<Test>().configureEach {
+    // ReleasePermissionTest reads .github/workflows/build.yml, and Gradle cannot infer that a
+    // test depends on a file outside the source set. Without this the task stays UP-TO-DATE
+    // when the workflow changes, so the guard goes stale exactly when it matters. Measured:
+    // deleting the release job's `contents: write` and re-running gave "BUILD SUCCESSFUL in
+    // 614ms" with the test never executing; the same mutation under --rerun-tasks failed it.
+    // A guard that does not re-run when its subject changes is not a guard.
+    inputs.file(rootProject.file(".github/workflows/build.yml"))
+        .withPropertyName("releaseWorkflow")
+        .withPathSensitivity(PathSensitivity.RELATIVE)
+
     extensions.configure<JacocoTaskExtension> {
         isIncludeNoLocationClasses = true
         excludes = listOf("jdk.internal.*")

app/src/androidTest/java/org/libremediaconverter/bench/RealMediaBenchmark.kt

@@ -36,8 +36,20 @@ import java.io.File
  *  1. that the hardware path is worth having a second engine for at all, and
  *  2. that x264's CRF is worth the GPL licence the app carries for it.
  *
- * Skips itself when the sample files are absent, so it is harmless in CI. Populate with:
- *   adb push <file>.mp4 /sdcard/Android/data/org.libremediaconverter/files/
+ * Skips itself when the sample files are absent, so it is harmless in CI — every green E2E
+ * leg reports two skips, and these are they.
+ *
+ * The two files it looks for, by exact name:
+ *
+ *  - [H264_SAMPLE] for [hardwareVersusSoftwareOnRealVideo]
+ *  - [AV1_SAMPLE] for [av1InputRoutesAccordingToDeviceDecodeSupport]
+ *
+ * **Where they go, and how, is on [samples] — read it before staging anything.** This used to
+ * carry an `adb push` line naming the external files dir, which [samples] then explains cannot
+ * work: a pushed file stays owned by the shell user and the app reads EACCES, surfacing as an
+ * unparseable input rather than a permission error. The instruction and its own refutation sat
+ * twelve lines apart. It is named in one place now rather than restated here, because restating
+ * it is what let the two drift.
  */
 @UnstableApi
 @RunWith(AndroidJUnit4::class)

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

@@ -1,5 +1,7 @@
 package org.libremediaconverter.saf
 
+import android.app.UiAutomation
+import androidx.compose.ui.test.ComposeTimeoutException
 import androidx.compose.ui.test.assertTextEquals
 import androidx.compose.ui.test.junit4.v2.createAndroidComposeRule
 import androidx.compose.ui.test.onAllNodesWithTag
@@ -10,6 +12,7 @@ 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.Configurator
 import androidx.test.uiautomator.StaleObjectException
 import androidx.test.uiautomator.UiDevice
 import androidx.test.uiautomator.Until
@@ -52,12 +55,138 @@ import org.libremediaconverter.ui.TestTags
  * either ViewModel, and `StateRestorationTester` saves into an in-memory map rather than a
  * `Bundle`.
  *
+ * ### #93: what actually failed was reading the screen, not the picker
+ *
+ * Ninety minutes after this class landed it started failing on gating legs at API 33, 34, 35 and
+ * 37 — on diffs that were two KDoc comments, a MIME lookup table and a README paragraph (#93).
+ * Every failure named the fixture root, so it read as a root-discovery race, and the ticket was
+ * filed on that reading. It was not one, and it was not the `StaleObjectException` #80 had fixed
+ * an hour earlier either.
+ *
+ * **DocumentsUI was fine.** On the API 34 leg of run 32806342548 its own
+ * `ProvidersAccess: Matched roots` names
+ * `content://org.libremediaconverter.test.fixtures/root/lmc-r38-root` five times inside the sixty
+ * seconds the test spent failing, `ActivityTaskManager` logged the `PickActivity` as `Displayed`,
+ * and the provider process started on cue.
+ *
+ * **This process could not read any window at all.** Two counts settle it. Across that whole leg
+ * UiAutomator logged `Retrieving node with selector` 1095 times and `Node not found with selector`
+ * 1095 times — not one selector ever matched, from the first query of the run. The green leg of
+ * the same job asked 7 times and found 5. `UiDevice.getWindowRoots` builds its search set from
+ * `UiAutomation.getWindows()` and, on API 21 and up, from nothing else; an empty list there makes
+ * every selector unfindable and says nothing whatever about SAF. The corroborating detail is that
+ * `By.desc("Show roots")` — the toolbar button, present on that screen whether the roots list is
+ * stale or not — was also not found, 28 s after the picker was displayed.
+ *
+ * **A fresh picker is not the repair, and this was measured rather than assumed.** The same leg
+ * opened a *second* `PickActivity` for the second test, in the same DocumentsUI process
+ * (pid 3299), and read exactly as little from it. So whatever was broken outlived one window.
+ * [requireAReadableScreen] is the part aimed at that: it asks whether this process can see the
+ * app's own window *before* the picker is opened, and [rebuildUiAutomation] tears the connection
+ * down and builds another if it cannot.
+ *
+ * **The check has since caught the real thing, in CI, and the connection rebuild did not repair
+ * it.** Run 32811493607, API 35 and API 37 legs, both tests, 12 s each instead of 60:
+ *
+ * ```
+ * java.lang.AssertionError: UiAutomator cannot see this app's own window, so it could not have
+ *   seen the picker's either. This is not a SAF failure.
+ *     at SafPickerRoundTripTest.requireAReadableScreen
+ * ```
+ *
+ * That is the diagnosis this class could not previously give, and it moves the question off SAF
+ * for good.
+ *
+ * ### What the window list said, and why nothing here can fix it
+ *
+ * [describeWindows] was added to that failure so the next occurrence would close the question
+ * rather than reopen it. It did — on the API 34 leg of run 32812248131 and again, character for
+ * character, on the API 33 leg of run 32812892103:
+ *
+ * ```
+ * ... Waking the device, dismissing the keyguard and rebuilding the UiAutomation connection all
+ * failed to make it readable. What it could see: com.android.systemui[type=3], android[type=3]
+ * ```
+ *
+ * `type=3` is `AccessibilityWindowInfo.TYPE_SYSTEM`. The list is **not** empty — it holds the
+ * system windows and **not one `TYPE_APPLICATION` window**, on a device where the framework had
+ * already logged `Displayed org.libremediaconverter/.MainActivity`. So the application layer
+ * never reaches accessibility on those boots, and every selector in this class, the picker's and
+ * the app's alike, is unfindable for the whole instrumentation run.
+ *
+ * Three CI runs on this branch caught the fault, at API 33, 34, 35 and 37, and every one of them
+ * printed that same list. It is not one level's quirk.
+ *
+ * ### And that list is what identified the occluder
+ *
+ * `android[type=3]` is `system_server`, and what it was holding is in the same logcat, minutes
+ * before this class ever ran:
+ *
+ * ```
+ * ANR in com.google.android.apps.nexuslauncher (com.google.android.apps.nexuslauncher/.NexusLauncherActivity)
+ * Reason: Input dispatching timed out (Application does not have a focused window)
+ * Window{4ed8414 u0 Application Not Responding: com.google.android.apps.nexuslauncher}
+ * ```
+ *
+ * **The launcher ANRs on a loaded runner emulator, and the dialog it leaves behind never goes
+ * away.** It is opaque and fullscreen, so `AccessibilityWindowManager` drops every application
+ * window beneath it — which is how the app can be `Displayed` and unreadable at once, the
+ * contradiction that made #93 look like a SAF bug for six PRs. It is present on both legs
+ * examined, at API 33 and 34, at the failure timestamp.
+ *
+ * So [dismissASystemErrorDialog] is tried first, and it is the remedy with a mechanism behind it.
+ * The other two are kept behind it and are **measured as not the cause**: [unlockTheDevice] (the
+ * keyguard theory, from `KeyguardViewMediator` reporting an unprovisioned device — dismissing it
+ * changed nothing) and [rebuildUiAutomation]. A second `PickActivity` is not a remedy for this
+ * either, and that was measured too: the first failing leg opened one and read as little from it.
+ *
+ * **What is honest about the dialog remedy: it has been shown to do no harm, not to work.** It
+ * was forced on with no dialog present and the suite stayed green, which is the way a blind
+ * `click()` could have broken a healthy run. Dismissing a real ANR dialog has not been observed,
+ * because the fault has never been reproduced locally — not on six warm runs, not on cold
+ * full-suite runs at API 34 and 35 on freshly created AVDs under `swangle_indirect` at two cores,
+ * not under host load. If it recurs, the message now names the dialog and the window list, so the
+ * next step is a measurement rather than another theory.
+ *
+ * ### The whole pick is retried, which is a separate and smaller claim
+ *
+ * [pickTheFixture] also backs out and asks for another picker when the walk comes up short. That
+ * is not the answer to the paragraph above; it is the answer to a picker whose *lists* were built
+ * before their data arrived, which is a real thing DocumentsUI does and which
+ * [tapPickerNode]'s re-find cannot reach either — it re-acquires a handle inside the one picker.
+ *
+ * One API 37 run failed a step deeper than the rest: the root appeared and
+ * `[TEXT='\Qlmc-r38-fixture.mp4\E']` did not. **That shape has not been reproduced or
+ * diagnosed.** It is covered here only because a fresh pick re-walks from Recent, and that is
+ * worth writing down rather than letting the retry read as a fix for something nobody measured.
+ *
  * ### 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.
+ * fixture root entirely, and both tests fail on the assertion that names it. **Re-run after the
+ * #93 retry landed**, because a retry that tolerated an absent root would have made this mutation
+ * vacuous, which is the one thing that must not happen here:
+ *
+ * ```
+ * java.lang.AssertionError: the system picker never showed BySelector [TEXT='\QLMC R38 fixtures\E'],
+ *   in 3 separate pickers (the last one left org.libremediaconverter in front)
+ *     at org.libremediaconverter.saf.SafPickerRoundTripTest.pickTheFixture(SafPickerRoundTripTest.kt:268)
+ * ```
+ *
+ * The root is absent from all three pickers, so all three report it, and the cost of saying so is
+ * bounded: 126 s and 127 s for the two tests, against the 1200 s wrapper timeout in
+ * `.github/scripts/e2e-run.sh`. The clause about what was left in front is not decoration either
+ * — it is what says the retry really did get back to the app between attempts rather than tapping
+ * behind a picker that never closed.
+ *
+ * **That mutation only shows the retry failing correctly.** Showing it *recovering* needs a
+ * failure that goes away, so one was injected: a field making the first
+ * [walkThePickerToTheFixture] of each test return a selector nothing matches. Both tests then
+ * passed, with `ActivityTaskManager` logging four `OPEN_DOCUMENT` starts for the two of them —
+ * two pickers each. That is the run which says the reopened pick completes: that
+ * `pickInput.launch` is not refused from the re-resumed Activity, and that the second test's
+ * reopen, which lands in the last-accessed stack rather than on Recent, still walks to the file.
  * 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
@@ -115,6 +244,10 @@ class SafPickerRoundTripTest {
     private val device: UiDevice =
         UiDevice.getInstance(InstrumentationRegistry.getInstrumentation())
 
+    /** The app under test, whose own window is what [requireAReadableScreen] asks for. */
+    private val appPackage: String =
+        InstrumentationRegistry.getInstrumentation().targetContext.packageName
+
     /** Set by the one test that rotates, read by [restoreOrientation]. See its KDoc. */
     private var rotated = false
 
@@ -208,25 +341,276 @@ class SafPickerRoundTripTest {
      * 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.
+     *
+     * **What is retried here is the whole pick.** [tapPickerNode]'s re-find re-acquires a handle
+     * to a node inside the picker that is already open, so it cannot reach a list that was built
+     * before its data arrived. Backing out and tapping "Choose file" again gets a *second*
+     * `PickActivity`, which rebuilds every list in it — and is what a user does when a picker
+     * comes up wrong. It is **not** the answer to the unreadable-screen failure in the class
+     * KDoc; [requireAReadableScreen], one line above, is the part aimed at that.
+     *
+     * The first attempt keeps the full [PICKER_TIMEOUT_MS]; the later ones use
+     * [REOPENED_TIMEOUT_MS], because by then the picker's process, its provider and its root cache
+     * are all warm and the only thing being waited on is one screen. That is what keeps the cost
+     * of a genuinely absent root bounded — see the class KDoc.
      */
     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()
+        var missing: BySelector? = null
+        repeat(PICK_ATTEMPTS) { attempt ->
+            requireAReadableScreen()
+            openThePicker()
+            missing = walkThePickerToTheFixture(
+                if (attempt == 0) PICKER_TIMEOUT_MS else REOPENED_TIMEOUT_MS,
+            )
+            if (missing == null) {
+                awaitNode(TestTags.Converter.FILE_CARD_NAME)
+                return
+            }
+            dismissThePicker()
         }
+        throw AssertionError(
+            "the system picker never showed $missing, in $PICK_ATTEMPTS separate pickers " +
+                "(the last one left ${device.currentPackageName} in front)",
+        )
+    }
 
-        tapPickerNode(By.text(FixtureDocumentsProvider.FIXTURE_DISPLAY_NAME))
+    /**
+     * Refuses to go near the picker until this process can read a window it already knows is there.
+     *
+     * **This is the check that would have answered #93 outright**, instead of leaving six PRs to
+     * infer a SAF fault from a picker that was never the problem. It is here because of what the
+     * failing logcat counts. Across the whole API 34 leg UiAutomator
+     * asked for a node 1095 times and logged `Node not found` 1095 times — it never read anything,
+     * from the first query of the run onwards. The green leg of the same job asked 7 times and
+     * found 5. So the window list `UiDevice` searches, `UiAutomation.getWindows()`, was empty for
+     * that entire instrumentation run; on API 21 and up that list is the *only* place
+     * `getWindowRoots` looks, so an empty one makes every selector unfindable and says nothing
+     * about the app, the picker or the fixture.
+     *
+     * The probe is deliberately the app's **own** window, asked while the app is in front and
+     * before anything is tapped. It is the one window that must be readable for any of the rest to
+     * mean anything, so a failure here is unambiguous — where "the picker never showed the root"
+     * was not, and is what sent #93 looking at package installation and root caches.
+     *
+     * The repair is [rebuildUiAutomation]. It has been forced on and measured — a rebuilt
+     * connection still reads windows, which is the way it could have been worse than nothing —
+     * but it has **never been run against the real fault**, because the fault has never been
+     * reproduced on demand. See the class KDoc. What is certain is that a fresh picker is *not*
+     * the repair: the failing leg opened a second `PickActivity` for the second test, in the
+     * same DocumentsUI process, and read exactly as little from it.
+     */
+    private fun requireAReadableScreen() {
+        val app = By.pkg(appPackage)
+        if (device.wait(Until.hasObject(app), READABLE_TIMEOUT_MS) == true) return
+        dismissASystemErrorDialog()
+        if (device.wait(Until.hasObject(app), READABLE_TIMEOUT_MS) == true) return
+        unlockTheDevice()
+        if (device.wait(Until.hasObject(app), READABLE_TIMEOUT_MS) == true) return
+        rebuildUiAutomation()
+        if (device.wait(Until.hasObject(app), READABLE_TIMEOUT_MS) != true) {
+            throw AssertionError(
+                "UiAutomator cannot see this app's own window, so it could not have seen the " +
+                    "picker's either. This is not a SAF failure. Closing a system error dialog, " +
+                    "waking the device, dismissing the keyguard and rebuilding the UiAutomation " +
+                    "connection all failed to make it readable. What it could see: " +
+                    describeWindows(),
+            )
+        }
+    }
 
-        awaitNode(TestTags.Converter.FILE_CARD_NAME)
+    /**
+     * Closes a system "isn't responding" dialog, if that is what is on top of the app.
+     *
+     * **This is the occluder #93 turned out to have**, and it took the window list in the failure
+     * message to find it. `AppNotRespondingDialog` belongs to `system_server`, so it is the
+     * `android[type=3]` in `com.android.systemui[type=3], android[type=3]` — and it is opaque and
+     * fullscreen, so `AccessibilityWindowManager` drops every application window beneath it. The
+     * app is `Displayed` and unreadable at the same time, which is exactly the contradiction this
+     * class spent #93 failing to explain. It is not even this app's dialog:
+     *
+     * ```
+     * ANR in com.google.android.apps.nexuslauncher (com.google.android.apps.nexuslauncher/.NexusLauncherActivity)
+     * Reason: Input dispatching timed out (Application does not have a focused window)
+     * Window{4ed8414 u0 Application Not Responding: com.google.android.apps.nexuslauncher}
+     * ```
+     *
+     * The launcher ANRs on a loaded runner emulator minutes before this class runs, and the dialog
+     * it leaves behind never goes away on its own.
+     *
+     * Dismissed by resource id rather than by button text, because the text is localised and the
+     * ids are not, and by id rather than by "the first button in the system window", because that
+     * would click whatever system window happened to be there. `aerr_wait` first: it dismisses the
+     * dialog and leaves the offending app alone, which is the polite answer when the app is not
+     * ours. Back is not tried — `BaseErrorDialog` swallows key events.
+     */
+    private fun dismissASystemErrorDialog() {
+        for (id in ERROR_DIALOG_BUTTONS) {
+            val button = device.findObject(By.res(id)) ?: continue
+            button.click()
+            device.waitForIdle()
+            return
+        }
+    }
+
+    /**
+     * Wakes the display and asks the keyguard to go away.
+     *
+     * The cheapest explanation for "this process cannot see the app's own window" is that
+     * something is in front of it, and on a runner emulator that something is the lock screen:
+     * these images come up unprovisioned, and `KeyguardViewMediator` says so in as many words --
+     * `we need to show the keyguard since the device isn't provisioned yet`. An occluded window is
+     * not in the accessibility window list, which is the same symptom as a broken connection and
+     * has a far more ordinary cause.
+     *
+     * `wm dismiss-keyguard` rather than a swipe, because it is a request to the window manager
+     * rather than a gesture that has to land somewhere this process cannot see. It is only
+     * attempted on the failure path -- a device that was readable never reaches here -- so a run
+     * where the keyguard was never up pays nothing and is not altered.
+     */
+    private fun unlockTheDevice() {
+        device.wakeUp()
+        device.executeShellCommand("wm dismiss-keyguard")
+        device.waitForIdle()
+    }
+
+    /** The accessibility window list, for a failure message that says what was actually there. */
+    private fun describeWindows(): String {
+        val windows = InstrumentationRegistry.getInstrumentation().uiAutomation.windows
+        if (windows.isEmpty()) return "no windows at all (UiAutomation.getWindows() is empty)"
+        return windows.joinToString(", ") { "${it.root?.packageName ?: "?"}[type=${it.type}]" }
+    }
+
+    /**
+     * Tears down this run's `UiAutomation` connection and establishes a new one.
+     *
+     * `Instrumentation.getUiAutomation` hands back the existing connection unless the flags differ
+     * from the ones it was created with, in which case it destroys it and builds another — so
+     * asking for different flags and then for the original ones back is how a test reaches the
+     * connection at all. `UiDevice` re-reads the flags from `Configurator` on every call rather
+     * than caching an instance, so the next selector goes through the new connection.
+     *
+     * `FLAG_DONT_SUPPRESS_ACCESSIBILITY_SERVICES` is toggled rather than chosen: it is only being
+     * used as a value that differs from whatever is configured, and it is put back.
+     *
+     * **Forced on and measured, because the obvious way for this to be worse than nothing is
+     * silent.** `UiDevice` puts `FLAG_RETRIEVE_INTERACTIVE_WINDOWS` on the service info during its
+     * own initialisation, and `getWindows()` is empty without it — so a rebuilt connection that
+     * did not get the flag back would cause exactly the emptiness this is meant to cure, on the
+     * one path where it is the last hope. Run unconditionally on every attempt, on a cold API 34
+     * emulator, both tests passed, and logcat shows the connection really being replaced rather
+     * than handed back: `Init UiAutomation[id=2, flags=0]`, then `id=4, flags=1`, then
+     * `id=6, flags=0`, with `Registering UiTestAutomationService` between each.
+     */
+    private fun rebuildUiAutomation() {
+        val configurator = Configurator.getInstance()
+        val flags = configurator.uiAutomationFlags
+        val instrumentation = InstrumentationRegistry.getInstrumentation()
+        configurator.uiAutomationFlags = flags xor UiAutomation.FLAG_DONT_SUPPRESS_ACCESSIBILITY_SERVICES
+        instrumentation.getUiAutomation(configurator.uiAutomationFlags)
+        configurator.uiAutomationFlags = flags
+        instrumentation.getUiAutomation(flags)
+    }
+
+    /** Waits for the app to be showing its own screen again, then asks for a picker. */
+    private fun openThePicker() {
+        awaitNode(TestTags.Converter.CHOOSE_FILE)
+        composeRule.onNodeWithTag(TestTags.Converter.CHOOSE_FILE).performClick()
+    }
+
+    /**
+     * Null once the fixture URI is with the app, or the selector whose list never carried it.
+     *
+     * Three things have to be there, in order, and the `when` names them in that order so that a
+     * failure says which one was missing rather than "the picker did not work".
+     *
+     * **The first branch is what tells an unreadable picker from an absent root.** In #93 neither
+     * the root *nor the toolbar's "Show roots" button* could be found for sixty seconds, and a
+     * stale roots list would have left the toolbar findable. Both arrived as one message. Asking
+     * for the picker's package on its own separates them: `never showed BySelector [PKG=...]`
+     * means the picker was not readable, and the root selector means the root was not offered.
+     *
+     * The second 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.
+     *
+     * **The third takes no recovery action of its own, and that is deliberate rather than an
+     * oversight.** [openTheRootsDrawer] exists because a root has a *second* place it can be
+     * shown; a document in a directory listing has no second place, so there is nothing an
+     * in-picker action could do. Its recovery is the outer loop: a fresh picker re-walks from
+     * Recent into the root, which rebuilds the directory listing as well as the roots strip.
+     */
+    private fun walkThePickerToTheFixture(timeoutMs: Long): BySelector? {
+        val picker = By.pkg(DOCUMENTS_UI_PACKAGE)
+        val root = By.text(FixtureDocumentsProvider.ROOT_TITLE)
+        val fixture = By.text(FixtureDocumentsProvider.FIXTURE_DISPLAY_NAME)
+        return when {
+            device.wait(Until.hasObject(picker), timeoutMs) != true -> picker
+            !tapPickerNode(root, timeoutMs, ifAbsent = ::openTheRootsDrawer) -> root
+            !tapPickerNode(fixture, timeoutMs) -> fixture
+            else -> null
+        }
+    }
+
+    /**
+     * The picker's own drawer, opened only when the root was not on the screen it landed on.
+     *
+     * **In practice it never runs, and #80 was right to say so.** A hierarchy dump taken on a
+     * cold API 34 emulator while this test was passing has the fixture root on the landing
+     * screen — `text="LMC R38 fixtures"` at `android:id/title`, under a `BROWSE FILES IN OTHER
+     * APPS` header — with the drawer shut (`Show roots` present, `Hide roots` absent). So the
+     * roots strip is the normal path and the drawer is a widening, kept because a device with a
+     * populated Recent may push the strip off screen. Looking in a second place widens where the
+     * root is searched for; it does not weaken what has to be found, which is still this root.
+     */
+    private fun openTheRootsDrawer() {
+        device.findObject(By.desc(SHOW_ROOTS_DESCRIPTION))?.click()
+    }
+
+    /**
+     * Backs out of the picker until the app has the window focus again.
+     *
+     * **The focus is asked of the Activity, not of UiAutomator, and that is not a stylistic
+     * choice.** The failure this retry exists for is a picker window UiAutomator cannot see, so a
+     * probe that went through the same accessibility window list would cheerfully report "the
+     * picker is gone" about the window that is still in front — and the reopened pick would then
+     * tap "Choose file" behind it. `Activity.hasWindowFocus` comes from the framework instead, and
+     * answers about the app rather than about the picker.
+     *
+     * It is also why this counts backs rather than pressing a fixed number of them. One back is
+     * enough from Recent and two are needed from inside the root, but a third from Recent would
+     * finish `MainActivity` and take the rest of the test with it.
+     */
+    private fun dismissThePicker() {
+        repeat(BACK_PRESSES) {
+            if (awaitAppFocus()) return
+            // Before the back press, not instead of it: an app-error dialog swallows key events,
+            // so a back aimed at the picker lands on the dialog and nothing moves. Measured --
+            // API 34 of run 32813885120 exhausted all four presses with `android` in front, which
+            // is that dialog, while the launcher it belonged to went on ANRing behind everything.
+            dismissASystemErrorDialog()
+            device.pressBack()
+        }
+        // The check after the last press, and not a spare one: `repeat` presses on its final
+        // iteration too, so without this a dismissal that worked on the last press would still be
+        // reported as a failure to close.
+        if (!awaitAppFocus()) {
+            throw AssertionError(
+                "the system picker would not close: after $BACK_PRESSES back presses the app " +
+                    "still does not have the window focus, and ${device.currentPackageName} is " +
+                    "in front. What could be seen: " + describeWindows(),
+            )
+        }
+    }
+
+    /** True once [MainActivity] has the window focus, false if it does not take it in time. */
+    private fun awaitAppFocus(): Boolean = try {
+        composeRule.waitUntil("the app has the window focus back", FOCUS_TIMEOUT_MS) {
+            composeRule.activity.hasWindowFocus()
+        }
+        true
+    } catch (_: ComposeTimeoutException) {
+        false
     }
 
     /**
@@ -244,21 +628,24 @@ class SafPickerRoundTripTest {
      *   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.
+     * So what is retried is *acquiring a handle to a node that has to be there anyway*. **A node
+     * that is simply not in this picker is reported rather than retried here** — it comes back as
+     * `false`, and [pickTheFixture] answers it with a whole new picker, which is the only thing
+     * that rebuilds a list or a window. The MIME mutation's bite is untouched either way: a root
+     * that is not in the picker is not found on any attempt or in any picker, and the failure is
+     * still "the system picker never showed" rather than a stale one.
      */
-    private fun tapPickerNode(selector: BySelector, ifAbsent: () -> Unit = {}) {
+    private fun tapPickerNode(selector: BySelector, timeoutMs: Long, ifAbsent: () -> Unit = {}): Boolean {
         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 ({}))
+            val node = awaitPickerNode(selector, timeoutMs, if (attempt == 0) ifAbsent else ({}))
+                ?: return false
             device.waitForIdle()
             try {
                 node.click()
-                return
+                return true
             } catch (e: StaleObjectException) {
                 stale = e
             }
@@ -267,19 +654,17 @@ class SafPickerRoundTripTest {
     }
 
     /**
-     * The picker node [selector] names, or a failure that says which one was missing.
+     * The picker node [selector] names, or null if this picker never showed it.
      *
      * [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)
+    private fun awaitPickerNode(selector: BySelector, timeoutMs: Long, ifAbsent: () -> Unit) =
+        device.wait(Until.findObject(selector), timeoutMs)
             ?: run {
                 ifAbsent()
-                requireNotNull(device.wait(Until.findObject(selector), PICKER_TIMEOUT_MS)) {
-                    "the system picker never showed $selector"
-                }
+                device.wait(Until.findObject(selector), timeoutMs)
             }
 
     /**
@@ -306,9 +691,69 @@ class SafPickerRoundTripTest {
         const val PICKER_TIMEOUT_MS = 30_000L
         const val APP_TIMEOUT_MS = 30_000L
 
+        /**
+         * The same wait once a picker has already come and gone, and shorter for a reason.
+         *
+         * What [PICKER_TIMEOUT_MS] is generous about is a cold start: DocumentsUI's process, the
+         * fixture's provider process, the root cache. By the second attempt all three are warm and
+         * the only thing left to wait on is one screen being laid out — measured at 2.7 to 3.4 s
+         * from the picker starting, on cold CI emulators at API 33, 34 and 35. Ten seconds is
+         * three times the worst of those, and it is what keeps a genuinely absent root — the MIME
+         * mutation — from costing three full-length attempts.
+         */
+        const val REOPENED_TIMEOUT_MS = 10_000L
+
+        /**
+         * How long the app is given to take the window focus back after a back press.
+         *
+         * Short, because this is asked once per back press and the first one is always asked while
+         * the picker is still in front, where it is *expected* to time out.
+         */
+        const val FOCUS_TIMEOUT_MS = 3_000L
+
+        /**
+         * How long this process is given to be able to read the screen at all.
+         *
+         * Short, and it is not waiting on anything being drawn: the app is already in front
+         * when this is asked. It is waiting only on the accessibility window list existing,
+         * which either does within a poll or two or -- as in #93 -- not at all.
+         */
+        const val READABLE_TIMEOUT_MS = 5_000L
+
         /** `Surface.ROTATION_0`, named rather than `0` so the comparison reads. */
         const val NATURAL_ROTATION = 0
 
+        /**
+         * How many pickers the fixture may fail to appear in before that is the finding.
+         *
+         * Three. Each one is a fresh `PickActivity` -- a fresh window, a fresh accessibility
+         * registration, a fresh roots query and a fresh directory load -- so this bounds the thing
+         * #93 measured, which is a picker that came up unreadable *once*. A root that is genuinely
+         * not offered is absent from all three, which is what keeps #64's MIME mutation red.
+         */
+        const val PICK_ATTEMPTS = 3
+
+        /**
+         * How many back presses may be spent getting out of a picker.
+         *
+         * One is enough from Recent, two from inside the fixture's own directory. Four leaves room
+         * for a picker that has been navigated deeper than this test ever navigates it, and stops
+         * well short of the count that would start finishing `MainActivity` instead.
+         */
+        const val BACK_PRESSES = 4
+
+        /**
+         * The package the system picker runs in.
+         *
+         * Named rather than resolved: `PackageManager.resolveActivity` is deprecated from API 33
+         * and its replacement is a lint argument this test does not need to have. A wrong value
+         * here cannot pass silently -- it is the first thing [walkThePickerToTheFixture] looks
+         * for, so the failure would read `never showed BySelector [PKG='...']` on every device.
+         * It is `com.google.android.documentsui` on every `google_apis` emulator image the CI
+         * matrix uses and on the Pixel 10 Pro XL.
+         */
+        const val DOCUMENTS_UI_PACKAGE = "com.google.android.documentsui"
+
         /**
          * How many times a picker node may be re-found before its staleness is the finding.
          *
@@ -319,6 +764,19 @@ class SafPickerRoundTripTest {
          */
         const val TAP_ATTEMPTS = 3
 
+        /**
+         * The buttons on the framework's app-error dialogs, by resource id.
+         *
+         * `aerr_wait` is first because it dismisses the dialog without killing the app under it,
+         * and the app under it is usually the launcher rather than anything this suite owns.
+         * `button1` catches the plainer `BaseErrorDialog` shapes that have no `aerr_` ids.
+         */
+        val ERROR_DIALOG_BUTTONS = listOf(
+            "android:id/aerr_wait",
+            "android:id/aerr_close",
+            "android:id/button1",
+        )
+
         /** DocumentsUI's drawer button. It carries no text, only this description. */
         const val SHOW_ROOTS_DESCRIPTION = "Show roots"
 

app/src/test/java/org/libremediaconverter/ci/ReleasePermissionTest.kt

@@ -0,0 +1,67 @@
+package org.libremediaconverter.ci
+
+import org.junit.Assert.assertTrue
+import org.junit.Test
+import java.io.File
+
+/**
+ * That the release job still holds the one permission it needs to publish.
+ *
+ * `build.yml`'s `release` job declares `contents: write`, and nothing was checking it. Deleting
+ * those two lines leaves actionlint clean and CodeQL silent — a *narrower* permission is not an
+ * alert — and the job is `if: startsWith(github.ref, 'refs/tags/v')`, so no pull request and no
+ * merge to `main` can exercise it. Measured: with the declaration removed, every gating check
+ * still passes. The first thing that would notice is a release failing to publish, at the moment
+ * someone is trying to cut one.
+ *
+ * The deletion also looks like tidying. A top-level `permissions: contents: read` now sits
+ * directly above it, so a reader could reasonably take the job-level block for a duplicate. It is
+ * an override, not a duplicate, and a comment saying so is not a check.
+ *
+ * `BackupExclusionsTest` is the precedent: a file that is configuration rather than code, load
+ * bearing, and unguarded because nothing compiles it.
+ *
+ * **What this pins, and what it does not.** It asserts the declaration exists in the `release`
+ * job's block. It cannot assert that a release actually publishes — that needs a tag push, which
+ * is the thing no PR can do. So this is a tripwire against silent removal, not proof the release
+ * path works.
+ */
+class ReleasePermissionTest {
+
+    @Test
+    fun `the release job declares the write permission it needs to publish`() {
+        val release = jobBlock("release")
+        assertTrue(
+            "build.yml's `release` job no longer declares `contents: write`. It is the only " +
+                "permission that lets the job create a release, the top-level block above it is " +
+                "`contents: read`, and nothing else in CI would catch this until a tag failed to " +
+                "publish. If the release moved elsewhere, delete this test deliberately.",
+            release.any { it.trimStart().startsWith("contents: write") },
+        )
+    }
+
+    /**
+     * The lines of one top-level job, from its `  <name>:` header to the next job at that indent.
+     *
+     * Line-based rather than parsed: the module has no YAML dependency, and adding one to read two
+     * lines would be a worse trade than a scan that fails loudly when the shape changes.
+     */
+    private fun jobBlock(name: String): List<String> {
+        val lines = workflow.readLines()
+        val start = lines.indexOfFirst { it == "  $name:" }
+        check(start >= 0) { "no `  $name:` job in ${workflow.path} — has the file been restructured?" }
+        val rest = lines.drop(start + 1)
+        val end = rest.indexOfFirst { it.matches(Regex("^ {2}[A-Za-z0-9_-]+:.*")) }
+        return if (end < 0) rest else rest.take(end)
+    }
+
+    /**
+     * Found by walking up rather than by a fixed relative path: Gradle's working directory for the
+     * unit tests is the module, but that is a default rather than a promise.
+     */
+    private val workflow: File
+        get() = generateSequence(File(".").absoluteFile) { it.parentFile }
+            .map { File(it, ".github/workflows/build.yml") }
+            .firstOrNull { it.isFile }
+            ?: error("could not find .github/workflows/build.yml above ${File(".").absolutePath}")
+}

app/src/test/java/org/libremediaconverter/codec/CodecVocabularyTest.kt

@@ -140,4 +140,34 @@ class CodecVocabularyTest {
         assertFalse("this device has no AVC decoder, and x264 is AVC", hevcOnly.canDecode("x264"))
         assertTrue("a name nobody knows keeps the permissive answer", hevcOnly.canDecode("cinepak"))
     }
+
+    /**
+     * The other half of the null policy, at the seam it exists for — #86.
+     *
+     * `mimeFor`'s `COPY, NONE -> null` arm carries its consequence in a comment: "Returning null
+     * makes canEncode answer true, which is the right answer: a copied or absent track places no
+     * demand on the hardware." That is a product decision, and until this test nothing held it. A
+     * MIME appearing in that arm would make a device with no matching encoder refuse a stream copy
+     * — a job that never encodes anything — and the router would send it to FFmpeg to re-mux what
+     * Media3 could have re-muxed.
+     *
+     * The `H264` line is what makes the other two mean something: without it, a `canEncode` that
+     * simply returned `true` would satisfy this test. `NONE` is asserted separately from `COPY`
+     * because they are one arm today and two answers, and splitting the arm must not silently
+     * halve the coverage.
+     */
+    @Test
+    fun `a device with no video encoder at all still permits a copied or absent track`() {
+        val noEncoders = AndroidDeviceCodecs.forTesting(encoders = emptySet(), decoders = setOf("video/avc"))
+        assertTrue(
+            "a copied track is re-muxed, not encoded, so no encoder is required",
+            noEncoders.canEncode(VideoCodec.COPY),
+        )
+        assertTrue("an absent track places no demand on the hardware", noEncoders.canEncode(VideoCodec.NONE))
+        assertFalse(
+            "this device has no AVC encoder, so an H.264 target has to be refused — without this, " +
+                "a canEncode that always answered true would satisfy the two assertions above",
+            noEncoders.canEncode(VideoCodec.H264),
+        )
+    }
 }

app/src/test/java/org/libremediaconverter/codec/VideoCodecMimeAgreementTest.kt

@@ -0,0 +1,162 @@
+package org.libremediaconverter.codec
+
+import androidx.media3.common.util.UnstableApi
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertNotNull
+import org.junit.Assert.assertNull
+import org.junit.Test
+import org.libremediaconverter.convert.Media3Engine
+import org.libremediaconverter.model.VideoCodec
+
+/**
+ * Bites on #86: a fifth `VideoCodec -> MIME` table, and nothing checking it agrees with the fourth.
+ *
+ * [AndroidDeviceCodecs.mimeFor] and [Media3Engine.videoMimeTypeFor] take the same enum and return a
+ * MIME string, from opposite ends of one export. The first asks the device *"have you an encoder
+ * for this?"*; the second tells Transformer *"produce this."* If they name different MIME types for
+ * the same codec, the app checks for one encoder and then requests another — the check passes, the
+ * export succeeds, and the user's H.265 file contains H.264. Both were `private` until #85 and #87
+ * widened them, so this assertion could not be written before; each table had per-arm tests that
+ * pinned its own answers and could not see the other side.
+ *
+ * **They do not agree everywhere, and must not be forced to.** Three buckets, all pinned below:
+ *
+ * - **H.264 and H.265** — both tables name a MIME, and it has to be the same one. This is the
+ *   bucket the defect lives in.
+ * - **VP8, VP9 and AV1** — the device table names a real MIME, Transformer's returns null. That is
+ *   correct, not drift: `Transformer.setVideoMimeType` will not accept them, so the router sends
+ *   them to FFmpeg before Media3 is asked anything, while a device may still genuinely own a VP9
+ *   encoder and `canEncode` has to give a truthful answer about it. Flattening `mimeFor` to null
+ *   here to "make the tables agree" would make `canEncode(VP9)` answer true on hardware that has
+ *   no VP9 encoder. The routing half of that claim is proved in
+ *   `Media3EngineMimeTypesTest.the router sends exactly H264 and H265 video encodes to Media3`,
+ *   which drives the real router; it is not repeated here.
+ * - **COPY and NONE** — neither names a MIME, because neither is encoded at all.
+ *
+ * The fourth bucket is asserted empty: a codec Transformer names and the device check cannot ask
+ * about would mean `canEncode` waving through a target the app then really does encode.
+ *
+ * **Audio has no partner, and that is a gap rather than a decision.** [Media3Engine.audioMimeTypeFor]
+ * is the same shape one enum over — `AudioCodec -> MIME` — but [AndroidDeviceCodecs] enumerates
+ * `video/` MIME types only, so there is no device-side audio table to cross-check it against. An
+ * audio encoder this device lacks is therefore not caught up front the way a video one is; the job
+ * reaches Media3 and falls back after failing. Named here so the asymmetry reads as unfinished
+ * rather than intended.
+ */
+@UnstableApi
+class VideoCodecMimeAgreementTest {
+
+    /** Both tables name a MIME. The pair has to match; this is the whole point of the file. */
+    private val bothNameAMime = setOf(VideoCodec.H264, VideoCodec.H265)
+
+    /** Only the device table names one, because Transformer is never asked for these. */
+    private val deviceOnly = setOf(VideoCodec.VP8, VideoCodec.VP9, VideoCodec.AV1)
+
+    /** Neither names one: nothing is encoded, so there is no encoder to name. */
+    private val neitherNamesOne = setOf(VideoCodec.COPY, VideoCodec.NONE)
+
+    /**
+     * Sorts every [VideoCodec] by what the two tables actually answer, then compares the sorting
+     * with the buckets documented above.
+     *
+     * This is what makes the agreement test below non-vacuous, and it is deliberately an exact
+     * comparison in all four directions. A codec added to the enum lands in some bucket and fails
+     * here rather than arriving unclassified. A table that starts returning null for everything —
+     * the shape a filtered loop would pass on — empties two buckets and fails here. And a
+     * *convergence* fails too: giving `videoMimeTypeFor(VP9)` a real MIME moves VP9 out of
+     * `deviceOnly`, which is the point. The divergence should be deliberate and visible, so
+     * changing it should require saying so in this file.
+     */
+    @Test
+    fun `each video codec is in the bucket the two tables actually put it in`() {
+        assertEquals(
+            "codecs both tables name a MIME for",
+            bothNameAMime,
+            VideoCodec.entries.filter { device(it) != null && transformer(it) != null }.toSet(),
+        )
+        assertEquals(
+            "codecs only the device check names a MIME for, because Transformer will not encode them",
+            deviceOnly,
+            VideoCodec.entries.filter { device(it) != null && transformer(it) == null }.toSet(),
+        )
+        assertEquals(
+            "codecs neither table names a MIME for, because nothing is encoded",
+            neitherNamesOne,
+            VideoCodec.entries.filter { device(it) == null && transformer(it) == null }.toSet(),
+        )
+        assertEquals(
+            "codecs Transformer names a MIME for that the device check cannot ask about — canEncode " +
+                "would answer true without looking, for a codec Media3 really is told to produce",
+            emptySet<VideoCodec>(),
+            VideoCodec.entries.filter { device(it) == null && transformer(it) != null }.toSet(),
+        )
+    }
+
+    /**
+     * The cross-check itself.
+     *
+     * Per-arm tests in either file cannot catch this: each pins its own table's answers, so a pair
+     * changed in lockstep with its own expectations stays green on both sides while the two tables
+     * describe different codecs.
+     */
+    @Test
+    fun `where both tables name a MIME they name the same one`() {
+        bothNameAMime.forEach { codec ->
+            val asked = device(codec)
+            val requested = transformer(codec)
+            assertNotNull("AndroidDeviceCodecs has no MIME to ask the device about for ${codec.label}", asked)
+            assertNotNull("Media3Engine has no MIME to give Transformer for ${codec.label}", requested)
+            assertEquals(
+                "${codec.label}: the device is asked about $asked and Transformer is then told to " +
+                    "produce $requested, so the capability check answers about a codec that is not the output",
+                asked,
+                requested,
+            )
+        }
+    }
+
+    /**
+     * The documented divergence, asserted rather than described.
+     *
+     * Both halves matter. The null side is Media3's refusal; the non-null side is the device
+     * check's genuine question, and it is the half a reader "tidying up" the disagreement would
+     * delete.
+     */
+    @Test
+    fun `the codecs Transformer will not encode are still codecs this device may or may not have`() {
+        deviceOnly.forEach { codec ->
+            assertNotNull(
+                "${codec.label} goes to FFmpeg, but canEncode still has to answer truthfully about " +
+                    "this device's encoder — a null here makes it answer true without looking",
+                device(codec),
+            )
+            assertNull(
+                "Transformer rejects ${codec.label}, so naming a MIME for it would request an export " +
+                    "Media3 cannot perform",
+                transformer(codec),
+            )
+        }
+    }
+
+    /**
+     * Guards every comparison above against passing as `null == null`.
+     *
+     * `MediaFormat`'s MIME types are Java compile-time constants and are inlined, so the unit-test
+     * classpath's stubbed `android.jar` never supplies them; `MimeTypes`' come from a real
+     * `media3-common` jar. If either stopped holding, the buckets would collapse and this fails
+     * first, with the reason. Same guard, and the same reason, as
+     * `CodecVocabularyTest.the MIME constants are real strings rather than stubs`.
+     */
+    @Test
+    fun `both tables return real MIME strings rather than stubs`() {
+        assertEquals("video/avc", AndroidDeviceCodecs.mimeFor(VideoCodec.H264))
+        assertEquals("video/hevc", AndroidDeviceCodecs.mimeFor(VideoCodec.H265))
+        assertEquals("video/x-vnd.on2.vp9", AndroidDeviceCodecs.mimeFor(VideoCodec.VP9))
+        assertEquals("video/avc", Media3Engine.videoMimeTypeFor(VideoCodec.H264))
+        assertEquals("video/hevc", Media3Engine.videoMimeTypeFor(VideoCodec.H265))
+    }
+
+    private fun device(codec: VideoCodec): String? = AndroidDeviceCodecs.mimeFor(codec)
+
+    private fun transformer(codec: VideoCodec): String? = Media3Engine.videoMimeTypeFor(codec)
+}

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

@@ -140,10 +140,16 @@ class ConversionViewModelProbeFailureTest {
     private fun pickedProbe(): InputProbe? {
         val viewModel = ConversionViewModel(app, Dispatchers.Unconfined)
         viewModel.onInputPicked(INPUT)
+        // The predicate is the guard, and it is the only one needed. It requires `Ready`, so a
+        // pick that ended in `Failed` never satisfies it and `awaitState` fails on its timeout
+        // naming what it was waiting for -- "Ready with a probe" -- which says more than a
+        // separate assertion could. A `ready as? ConversionState.Failed` check used to sit here
+        // and was dead: `Ready` and `Failed` are sibling subtypes of one sealed interface, so
+        // the cast was always null and the assertNull could never fire. Measured, not assumed --
+        // flipping it to assertNotNull failed all three callers of this helper.
         val ready = awaitState(viewModel.state, "Ready with a probe") {
             it is ConversionState.Ready && it.input.probe != null
         }
-        assertNull("nothing here should reach a terminal failure", (ready as? ConversionState.Failed))
         return (ready as ConversionState.Ready).input.probe
     }
 

tools/local-emulator/run-e2e.sh

@@ -236,10 +236,24 @@ ensure_avd() {
   else
     if [ ! -d "$img_dir" ]; then
       echo "  installing $pkg"
-      yes | sdkmanager --install "$pkg" > /dev/null 2>&1 || {
+      # Read sdkmanager's own status, not the pipeline's. `yes` never ends, so the moment
+      # sdkmanager exits and closes the pipe, `yes` dies of SIGPIPE with 141 -- and this
+      # script runs under `pipefail`, which takes the rightmost NON-ZERO status. A package
+      # that installed perfectly therefore reported "FAILED to install".
+      #
+      # Measured rather than reasoned: under `set -o pipefail`, `yes | true` exits 141 on
+      # every run, and `yes | sh -c 'exit 3'` exits 3 -- so the pipeline status cannot tell
+      # a clean install from a broken one, while ${PIPESTATUS[1]} reports 0 and 3.
+      #
+      # The `echo no | avdmanager` below is deliberately NOT changed. One line fits the pipe
+      # buffer, so echo has already exited before the close and there is no signal to
+      # receive; `echo no | true` measured 0 on every run. Only an unbounded producer is
+      # exposed to this.
+      yes | sdkmanager --install "$pkg" > /dev/null 2>&1
+      if [ "${PIPESTATUS[1]}" -ne 0 ]; then
         echo "  FAILED to install $pkg"
         return 1
-      }
+      fi
     fi
     echo "  creating AVD $avd from $pkg"
     echo no | avdmanager create avd -n "$avd" -k "$pkg" -d pixel_6 --force > /dev/null 2>&1 || {