Record wave 4's landed coverage, and that the hang watchdog has fired

Numbers re-measured on 6004398 rather than inherited: 94.2% line (2234/2372),
87.5% branch (1171/1338), 628 tests in 96 classes, missed 138 lines and 167
branches.

The entry explains its own branch move, because this file has a history of not
doing that. Wave 4's rise is almost entirely numerator -- up 80, with the
denominator moving -4 -- unlike the 2026-08-29 seam work, where the numerator
rose 37 while the denominator fell 70 and much of the gain was scaffolding
leaving the measurement. The line denominator grew 20, which is the seams the
wave cut rather than untested code.

Also records a methodological result worth more than the percentages: #218 fixed
an intermittent flake whose six-runs-per-arm comparison caught nothing either way
and proved nothing, and was settled by a deterministic mutation instead -- then
confirmed when the race reproduced on #217's leg, below the fix.

The JUnit-timeout trap gains the outcome it was waiting for: the jstack watchdog
caught #125's Room/WorkManager deadlock on CI in 10m57s with the hung test named,
against that ticket's prediction of a 60-minute cap and no cause. #125 is closed
as bounded; the inversion is internal to the two libraries and still live at
work-runtime 2.11.2 / room 2.7.0.

Instrumented counts are untouched: #221 is what changes them, so it carries them.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

CLAUDE.md

@@ -130,9 +130,9 @@ install for code that can never run — and on API 37 the full APK does not fit
 - The `model` package is excluded from `ReturnCount` and `CyclomaticComplexMethod` only. It is the
   decision layer, where one branch is one documented user-visible outcome and the metric counts
   answers rather than complexity. Every other rule still applies there.
-- **Coverage is reported, not gated** — **92.8% of lines (2183/2352), 81.3% of branches
-  (1091/1342)**, measured 2026-09-02 with `./gradlew :app:jacocoTestReport`, against 584 JVM tests
-  in 87 classes.
+- **Coverage is reported, not gated** — **94.2% of lines (2234/2372), 87.5% of branches
+  (1171/1338)**, measured 2026-09-05 with `./gradlew :app:jacocoTestReport`, against 628 JVM tests
+  in 96 classes.
 
   **Every figure this file carried before 2026-08-24 was an artifact, roughly half the real one.**
   Robolectric loads classes through its own sandbox classloader with no source location, JaCoCo
@@ -283,6 +283,33 @@ install for code that can never run — and on API 37 the full APK does not fit
   #194 before re-arguing either way — and note the reason it is worth cutting is not coverage but
   that the `runCatching` fallback logs "assuming permissive" while returning empty sets, which makes
   `canEncode` and `canDecode` answer *no* for everything.
+
+  **Wave 4's tests then landed on 2026-09-05**, as #206-#217 for the twelve tickets plus #218
+  (#159) and #219 (#122): 92.8% -> **94.2%** line, 81.3% -> **87.5%** branch, 584 -> 628 tests in 87
+  -> 96 classes. Missed lines 169 -> 138, missed branches 251 -> 167.
+
+  **Its branch move is a different animal from the 2026-08-29 seam work's, and the difference is the
+  point.** That one gained 6.3 branch points with the numerator up 37 (974 -> 1011) while the
+  denominator *fell* 70 (1410 -> 1340) — much of the rise was scaffolding leaving the measurement
+  rather than arms being covered. Here the numerator is up **80** (1091 -> 1171) and
+  the denominator moved **-4** (1342 -> 1338). So this one is almost entirely tests choosing arms
+  nothing had chosen, which is what the entry above warns to check before quoting a branch figure.
+  The line denominator rose the other way, 2352 -> 2372, and that is new production code rather than
+  untested code: the seams the wave cut — `capabilitiesFrom`, `ffprobeInfoFrom`, `sessionOutcome`,
+  and `sweepScope`/`startupSweep`.
+
+  **The two-filter method above is what found the work**, and its second filter earned its place:
+  the largest single gap of the wave (#192, the Cancel button never shown to reach WorkManager) sits
+  on lines that were already green and no line-level filter could see it.
+
+  One result worth carrying forward about *evidence* rather than coverage. #218 fixed a flake whose
+  reproduction is statistical, and running the whole suite six times per arm caught nothing either
+  way — at the observed rate a clean six-run arm is roughly a coin flip, so the comparison was
+  underpowered and proved nothing. What settled it was a deterministic mutation, and then the merge
+  train confirmed it by accident: the race reproduced on #217's Unit tests leg, which sits below
+  #218 and carries the unfixed scope. **Prefer a mutation that must go red to a repetition count**
+  when a fix is for something intermittent.
+
 - **Testable code is not done until it is tested.** If a piece is unit testable, it gets unit
   tests before it counts as done. If it is e2e testable, it gets e2e tests. Both clauses apply —
   a change that is both needs both.
@@ -410,4 +437,29 @@ Because versions float, a build can change without a commit. `./gradlew :app:dep
   run instead is `timeout` on the `Test` tasks plus the jstack watchdog beside it in
   `app/build.gradle.kts`, neither of which moves a thread. `HangBoundTest` guards both numbers,
   and **a timed-out run writes no XML for the class that hung** — the dump is its only
-  attribution, so do not delete the watchdog as stray config.
+  attribution, so do not delete the watchdog as stray config. It has since been exercised in anger:
+  on 2026-09-05 it caught #125's Room/WorkManager deadlock on CI, failing in 10m57s with the hung
+  test named, where that ticket had predicted a 60-minute cap and no cause. #125 is closed as
+  bounded on the strength of it — the inversion itself is internal to the two libraries and still
+  live at `work-runtime` 2.11.2 / `room` 2.7.0.
+- **The JVM suite does not run `LibreMediaConverterApp`.** `app/src/test/resources/robolectric.properties`
+  names `TestLibreMediaConverterApp` for every test, and it differs from the real class in exactly
+  one thing: `sweepScope` is `Dispatchers.Unconfined`, so the startup staging sweep finishes before
+  `onCreate()` returns instead of running on `Dispatchers.IO`.
+
+  **That line is load-bearing — do not delete it as stray config.** Robolectric builds an
+  `Application` per test class that asks for one, and each `onCreate` launched a sweep over the
+  shared `<cacheDir>/conversions/` that nothing joined. So a test asserting about a staged file was
+  racing every sweep the classes before it had left in flight (#159). It was CI-only until wave 4
+  added ten Robolectric classes, at which point `OutputPublisherStagingTest` failed on roughly one
+  local run in six. Per-test opt-in was measured and rejected: **27 of the 58 Robolectric classes
+  touch that directory**. The `SupervisorJob` is kept in the test scope so a throwing sweep is
+  swallowed there exactly as in production — the dispatcher is the only intended difference.
+
+  **It cost one assertion, knowingly.** `AppStartSweepTest` used to open by asserting that the
+  manifest's `android:name` is what Robolectric instantiated, so the sweep is code that actually
+  runs. An `application=` override *replaces* the manifest rather than being checked against it, and
+  `applicationInfo.className` reports the override too — measured — so that claim is not merely
+  unasserted on the JVM now, it is unobservable, and a rewritten version would assert the override
+  against itself. **The manifest link is device-only.** What remains is the `as LibreMediaConverterApp`
+  cast in that class's `setUp`, which catches only the test app ceasing to extend the real one.

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

@@ -10,6 +10,9 @@ 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.runner.lifecycle.ActivityLifecycleCallback
+import androidx.test.runner.lifecycle.ActivityLifecycleMonitorRegistry
+import androidx.test.runner.lifecycle.Stage
 import androidx.test.uiautomator.By
 import androidx.test.uiautomator.BySelector
 import androidx.test.uiautomator.Configurator
@@ -24,6 +27,7 @@ import org.junit.runner.RunWith
 import org.libremediaconverter.FailsOnEmulatorApi37
 import org.libremediaconverter.MainActivity
 import org.libremediaconverter.ui.TestTags
+import java.util.concurrent.atomic.AtomicInteger
 
 /**
  * Choosing a file, through the real system picker, and still having it after a rotation.
@@ -251,6 +255,21 @@ class SafPickerRoundTripTest {
     /** Set by the one test that rotates, read by [restoreOrientation]. See its KDoc. */
     private var rotated = false
 
+    /** Counts [MainActivity] creations from the moment [watchForRecreation] is called. */
+    private val recreations = AtomicInteger()
+
+    /**
+     * Counts a rotation's recreation without asking the Activity anything.
+     *
+     * Deliberately not `composeRule.activity`, which resolves through `scenario.onActivity` and so
+     * blocks on the main thread. Polling *that* across a recreation is a plausible reading of the
+     * 20-minute wedges in #122, which would make the obvious barrier the bug it is meant to fix.
+     * The runner's lifecycle monitor is a callback: reading the counter touches no looper.
+     */
+    private val recreationWatcher = ActivityLifecycleCallback { activity, stage ->
+        if (activity is MainActivity && stage == Stage.CREATED) recreations.incrementAndGet()
+    }
+
     /**
      * Leave the device the way it was found — and only if this test moved it.
      *
@@ -270,6 +289,7 @@ class SafPickerRoundTripTest {
      */
     @After
     fun restoreOrientation() {
+        ActivityLifecycleMonitorRegistry.getInstance().removeLifecycleCallback(recreationWatcher)
         if (!rotated) return
         device.setOrientationNatural()
         device.unfreezeRotation()
@@ -303,9 +323,11 @@ class SafPickerRoundTripTest {
         // 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)
+        watchForRecreation()
 
         device.setOrientationLandscape()
         rotated = true
+        awaitRecreation()
         composeRule.waitForIdle()
 
         // Two guards before the assertion that matters, because both of the ways this test could
@@ -675,6 +697,36 @@ class SafPickerRoundTripTest {
      * `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.
      */
+    /** Starts counting [MainActivity] creations, so [awaitRecreation] can wait for the next one. */
+    private fun watchForRecreation() {
+        recreations.set(0)
+        ActivityLifecycleMonitorRegistry.getInstance().addLifecycleCallback(recreationWatcher)
+    }
+
+    /**
+     * Waits for the rotation to actually rebuild [MainActivity], which `waitForIdle` does not.
+     *
+     * **This is #122.** `waitForIdle()` waits for the compose hierarchy to settle. Immediately
+     * after a rotation the window manager has accepted but not yet delivered as a configuration
+     * change, the *old* Activity's composition is already idle — so it returns, `composeRule
+     * .activity` still resolves to the old instance, and the guard below reads an unchanged
+     * identity hash. That is the clean `AssertionError` seen on the API 33 gating leg of #217, and
+     * the wedges on #122 are the same race taken the other way: land while the composition is
+     * being torn down and there is nothing coherent for `waitForIdle` to settle on.
+     *
+     * A bounded wait is worth having even if that second half is wrong. It turns a 20-minute
+     * `WEDGE_TIMEOUT` — which costs the leg and names no test — into a fast failure that says which
+     * test and what it was waiting for.
+     */
+    private fun awaitRecreation() {
+        composeRule.waitUntil(
+            "the rotation did not recreate MainActivity within $RECREATION_TIMEOUT_MS ms",
+            RECREATION_TIMEOUT_MS,
+        ) {
+            recreations.get() > 0
+        }
+    }
+
     private fun awaitNode(tag: String) {
         composeRule.waitUntil("a node tagged $tag exists", APP_TIMEOUT_MS) {
             composeRule.onAllNodesWithTag(tag).fetchSemanticsNodes().isNotEmpty()
@@ -703,6 +755,15 @@ class SafPickerRoundTripTest {
          */
         const val REOPENED_TIMEOUT_MS = 10_000L
 
+        /**
+         * How long a rotation is given to destroy and rebuild the Activity.
+         *
+         * Generous against the API 33 and 34 emulators #122 was measured on, where the rotation is
+         * slow enough for the gap this bound exists to cover to be observable at all — and still
+         * two orders of magnitude inside the 1200 s `WEDGE_TIMEOUT` it replaces.
+         */
+        const val RECREATION_TIMEOUT_MS = 15_000L
+
         /**
          * How long the app is given to take the window focus back after a back press.
          *

app/src/main/java/org/libremediaconverter/LibreMediaConverterApp.kt

@@ -3,6 +3,7 @@ package org.libremediaconverter
 import android.app.Application
 import kotlinx.coroutines.CoroutineScope
 import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.Job
 import kotlinx.coroutines.SupervisorJob
 import kotlinx.coroutines.launch
 import org.libremediaconverter.convert.OutputPublisher
@@ -17,14 +18,36 @@ import org.libremediaconverter.convert.OutputPublisher
  * ever becomes a `Converted` state, or a `reset()`'s delete is cancelled along with the
  * Activity. Process start is the one moment those leftovers are reliably observable.
  */
-class LibreMediaConverterApp : Application() {
+open class LibreMediaConverterApp : Application() {
 
     /**
      * Deliberately process-lifetime and never cancelled: the work it carries is a single
      * short task that should outlive nothing in particular and be interrupted by nothing.
      * A `SupervisorJob` so a failure here could never take a sibling down with it.
+     *
+     * **`protected open` for #159.** Robolectric builds an `Application` for every test that asks
+     * for one, so on the JVM this is not one background sweep but one *per test* — all of them on
+     * `Dispatchers.IO`, all touching the same `cacheDir`, none of them joined by anything. That is
+     * a race against any test asserting about a file under `conversions/`, and it grew with the
+     * suite: wave 4 added ten Robolectric classes and took it from CI-only to roughly one local run
+     * in six. The JVM suite substitutes a scope that runs the sweep inline — see
+     * `app/src/test/resources/robolectric.properties` and `TestLibreMediaConverterApp`.
+     *
+     * A constructor parameter would be the ordinary way to inject this and is not available: the
+     * framework builds this class, so the seam has to be a member.
      */
-    private val appScope = CoroutineScope(SupervisorJob() + Dispatchers.IO)
+    protected open val sweepScope: CoroutineScope = CoroutineScope(SupervisorJob() + Dispatchers.IO)
+
+    /**
+     * The sweep [onCreate] last started, so a caller that needs it finished can wait for it.
+     *
+     * Nothing in production reads this — process start does not wait for its own housekeeping. It
+     * exists because the alternative for a test is a timed poll, and a poll cannot tell "the sweep
+     * has not run yet" from "the sweep ran and did nothing".
+     */
+    @Volatile
+    var startupSweep: Job? = null
+        private set
 
     override fun onCreate() {
         super.onCreate()
@@ -53,6 +76,6 @@ class LibreMediaConverterApp : Application() {
         //
         // sweepStaging() also re-reads each timestamp immediately before deleting, which
         // closes the window between listing the directory and acting on the listing.
-        appScope.launch { OutputPublisher(this@LibreMediaConverterApp).sweepStaging() }
+        startupSweep = sweepScope.launch { OutputPublisher(this@LibreMediaConverterApp).sweepStaging() }
     }
 }

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

@@ -1,8 +1,8 @@
 package org.libremediaconverter
 
-import org.junit.Assert.assertEquals
+import kotlinx.coroutines.runBlocking
+import org.junit.Assert.assertNotNull
 import org.junit.Assert.assertTrue
-import org.junit.Assert.fail
 import org.junit.Before
 import org.junit.Test
 import org.junit.runner.RunWith
@@ -10,7 +10,6 @@ import org.libremediaconverter.convert.StagingSweep
 import org.robolectric.RobolectricTestRunner
 import org.robolectric.RuntimeEnvironment
 import java.io.File
-import java.util.concurrent.TimeUnit
 
 /**
  * That process start actually sweeps.
@@ -23,8 +22,19 @@ import java.util.concurrent.TimeUnit
  * output ever became a `Converted` state, a `reset()` whose delete was cancelled with the Activity.
  *
  * `onCreate()` is called again rather than a second Application being built: it is what the
- * framework calls at process start, the scope it launches on is already there, and the first test
- * below is what pins that the framework calls it on *this* class.
+ * framework calls at process start, and the scope it launches on is already there.
+ *
+ * **What this class stopped covering in #159, deliberately.** It used to open by asserting that
+ * `RuntimeEnvironment.getApplication()` is a [LibreMediaConverterApp] — that the manifest's
+ * `android:name` points here, so the sweep is code that actually runs. That assertion cannot exist
+ * on the JVM any more: `robolectric.properties` now names [TestLibreMediaConverterApp] for the
+ * whole suite, and an `application=` override replaces the manifest rather than being checked
+ * against it — `applicationInfo.className` reports the override too, measured. So the manifest is
+ * not merely unasserted here, it is unobservable from this source set, and a rewritten version of
+ * that test would have asserted the override against itself. **The manifest link is a device-only
+ * guarantee now**, and it was traded knowingly for the race that override fixes. The cast in
+ * [setUp] still fails if [TestLibreMediaConverterApp] stops extending the real class, which is a
+ * smaller claim than the one withdrawn.
  */
 @RunWith(RobolectricTestRunner::class)
 class AppStartSweepTest {
@@ -34,17 +44,35 @@ class AppStartSweepTest {
 
     @Before
     fun setUp() {
-        // The cast is an assertion in itself: Robolectric builds the Application named in the
-        // merged manifest, so this fails if `android:name` ever stops pointing here -- in which
-        // case the sweep below would be perfectly correct code that never runs.
         app = RuntimeEnvironment.getApplication() as LibreMediaConverterApp
         stagingDir = File(app.cacheDir, "conversions").apply { mkdirs() }
         stagingDir.listFiles()?.forEach { it.delete() }
     }
 
+    /**
+     * The property the whole substitution exists for, asserted directly rather than waited on.
+     *
+     * #159 is not "the sweep is slow", it is "the sweep is still running while some later test
+     * reads the directory". [TestLibreMediaConverterApp] answers that by finishing the sweep before
+     * `onCreate()` returns, and this is the only place that claim is checked -- every other test in
+     * the suite benefits from it silently and would go back to racing without saying why.
+     *
+     * Deterministic in the direction that matters: `Dispatchers.Unconfined` runs a `launch` whose
+     * body never suspends to completion inline, so this cannot flake green-to-red. Putting the test
+     * app back on `Dispatchers.IO` makes it a race that the assertion loses essentially every time,
+     * which is what a six-run suite comparison could not show -- at the rate #159 was observed at,
+     * a clean six-run arm is a coin flip.
+     */
     @Test
-    fun `the application the manifest starts is the one that sweeps`() {
-        assertEquals(LibreMediaConverterApp::class.java, RuntimeEnvironment.getApplication().javaClass)
+    fun `the sweep is finished before onCreate returns`() {
+        app.onCreate()
+
+        val sweep = app.startupSweep
+        assertNotNull("onCreate() started no sweep", sweep)
+        assertTrue(
+            "the JVM suite's sweep outlived onCreate(), so it is in flight during test bodies again",
+            sweep?.isCompleted == true,
+        )
     }
 
     @Test
@@ -64,35 +92,23 @@ class AppStartSweepTest {
 
         app.onCreate()
 
-        awaitGone(abandoned)
+        // Joined rather than polled. `onCreate` publishes the sweep it started, so this waits for
+        // that exact sweep -- where a timed poll could not tell "swept" from "not started yet", and
+        // answered the second case by failing after ten seconds.
+        val sweep = app.startupSweep
+        assertNotNull("onCreate() started no sweep to wait for", sweep)
+        runBlocking { sweep?.join() }
+
+        assertTrue("process start left ${abandoned.name} in staging; nothing swept it", !abandoned.exists())
         // The other half, and the one that says the sweep is a sweep rather than a
         // `clearStaging()`: the directory is shared by the convert tab, the join tab and
         // ConcatEngine's list file, so deleting everything could take a file from a running job.
         assertTrue("a file written moments ago belongs to a live job", live.exists())
     }
 
-    /**
-     * Waits for [file] to be deleted.
-     *
-     * The sweep runs on `Dispatchers.IO`, deliberately: it lists a directory and stats every entry
-     * on the path that decides how long the launcher icon stays unresponsive. So there is nothing
-     * to join, and the wait is a bounded poll — long enough for a directory listing, short enough
-     * that a sweep which never happens fails rather than hangs.
-     */
-    private fun awaitGone(file: File) {
-        val deadline = System.nanoTime() + TimeUnit.SECONDS.toNanos(AWAIT_TIMEOUT_SECONDS)
-        while (System.nanoTime() < deadline) {
-            if (!file.exists()) return
-            Thread.sleep(POLL_INTERVAL_MS)
-        }
-        fail("process start left ${file.name} in staging; nothing swept it")
-    }
-
     private fun stagedFile(name: String): File = File(stagingDir, name).apply { writeBytes(ByteArray(4096)) }
 
     private companion object {
         const val ONE_MINUTE_MS = 60L * 1000
-        const val AWAIT_TIMEOUT_SECONDS = 10L
-        const val POLL_INTERVAL_MS = 5L
     }
 }

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

@@ -0,0 +1,28 @@
+package org.libremediaconverter
+
+import kotlinx.coroutines.CoroutineScope
+import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.SupervisorJob
+
+/**
+ * The [LibreMediaConverterApp] the JVM suite runs, differing from it in exactly one thing: the
+ * startup sweep runs inline on the thread that builds the Application instead of on
+ * `Dispatchers.Unconfined`.
+ *
+ * **This is #159.** Robolectric builds an `Application` per test class that asks for one, and each
+ * one launches a sweep over the shared `<cacheDir>/conversions/`. Nothing joins them, so a test
+ * asserting about a staged file is racing however many sweeps the classes before it left in
+ * flight — `OutputPublisherStagingTest` being the one that lost, at roughly one local run in six
+ * once wave 4 added ten more Robolectric classes. Making the sweep finish before `onCreate()`
+ * returns removes the race for every test at once rather than asking each to opt in; 27 of the
+ * suite's 58 Robolectric classes touch that directory, so opting in was not a real option.
+ *
+ * `Dispatchers.Unconfined` is what makes it inline: `sweepStaging()` is a plain function, so an
+ * `Unconfined` `launch` runs it to completion before returning. The `SupervisorJob` is kept so this
+ * differs from production in the dispatcher alone — a sweep that throws is logged and swallowed
+ * here exactly as it is there, rather than taking Application construction down with it and failing
+ * every test in the class for an unrelated reason.
+ */
+class TestLibreMediaConverterApp : LibreMediaConverterApp() {
+    override val sweepScope: CoroutineScope = CoroutineScope(SupervisorJob() + Dispatchers.Unconfined)
+}

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

@@ -122,24 +122,25 @@ class OutputPublisherStagingTest {
 
     /**
      * Makes `cacheDir/conversions` a regular file, which is the whole precondition of the test
-     * above -- and does it in a loop, because a single delete-then-write loses a race that CI
-     * caught and this machine does not reproduce.
+     * above -- and does it in a loop, because a single delete-then-write once lost a race that CI
+     * caught and this machine did not reproduce.
      *
-     * `LibreMediaConverterApp.onCreate` ends with
-     * `appScope.launch { OutputPublisher(...).sweepStaging() }` on `Dispatchers.IO`, and
-     * `sweepStaging` reads `stagingDir`, whose getter calls `mkdirs()`. Robolectric instantiates
-     * the application for every test that asks for one, so that background `mkdirs()` is in flight
-     * across the whole suite, on a thread the paused main looper does not control. Between deleting
-     * this path and writing it there is a window where the path does not exist and that `mkdirs()`
-     * can win, which is `FileNotFoundException: ... (Is a directory)` out of `writeBytes` -- run
-     * 33069641674 on #149, once, against 468 tests that pass here.
+     * **That race is closed at the source as of #159, and the loop is kept anyway.**
+     * `LibreMediaConverterApp.onCreate` launched its staging sweep on `Dispatchers.IO`, and
+     * `sweepStaging` reads `stagingDir`, whose getter calls `mkdirs()`. Robolectric builds an
+     * application for every test class that asks for one, so that background `mkdirs()` was in
+     * flight across the whole suite, on a thread the paused main looper does not control. Between
+     * deleting this path and writing it there is a window where the path does not exist and that
+     * `mkdirs()` could win -- `FileNotFoundException: ... (Is a directory)` out of `writeBytes`,
+     * run 33069641674 on #149, once, against 468 tests that passed here. The JVM suite now runs
+     * `TestLibreMediaConverterApp`, whose sweep finishes before `onCreate()` returns, so nothing is
+     * sweeping while a test body runs.
      *
-     * Retrying closes it rather than narrowing it, because the race is not symmetric: `mkdirs()`
-     * fails on an existing regular file, so the invariant only has to survive being *established*.
-     * Once a write lands, nothing in the suite can turn this back into a directory.
-     *
-     * The wider problem -- application-scope IO work racing every Robolectric test that shares
-     * `cacheDir` -- is #159, and is deliberately not fixed here.
+     * The loop stays because it is what would catch that substitution being undone. Without it the
+     * regression returns as this one class failing rarely on CI -- the exact shape that took #159
+     * from a single run on #149 to a wave-4 flake before anyone chased it. Retrying closes the
+     * window rather than narrowing it, because the race is not symmetric: `mkdirs()` fails on an
+     * existing regular file, so the invariant only has to survive being *established*.
      */
     private fun stagingPathAsRegularFile(): File {
         val stagingPath = File(cacheDir, "conversions")

app/src/test/resources/robolectric.properties

@@ -10,3 +10,9 @@
 # Set here rather than in a @Config on each class so a later Robolectric test does not have
 # to rediscover it. Remove it once Robolectric ships an android-all jar for 37.
 sdk=36
+
+# Every test gets TestLibreMediaConverterApp, whose only difference from the real one is that the
+# startup sweep runs inline rather than on Dispatchers.IO. Set suite-wide because the race it fixes
+# (#159) is suite-wide: any class that builds an Application leaves a sweep of the shared staging
+# directory in flight for whatever runs next. TestLibreMediaConverterApp explains the choice.
+application=org.libremediaconverter.TestLibreMediaConverterApp