Stop the cancel tests losing their race on a loaded runner

Both cancellation tests I added in ad2a75d and d293646 wait for
SessionState.RUNNING and then cancel. That is not enough. The conversion
one passed four consecutive local runs and all five CI legs, then failed
the API 34 and 35 legs of the next PR with state=COMPLETED rc=0, on a diff
that could not reach it. On a loaded runner the thread that observed
RUNNING can be descheduled long enough for a short encode to finish before
it calls cancel. A longer timeout does not help: the wait already
succeeded.

Two changes, because neither is sufficient alone.

A slower encode. The conversion test now targets WEBM_VP9 at BEST, the
slowest thing FFmpegCommandBuilder emits -- libvpx-vp9 -crf 31 -b:v 0,
with -deadline realtime added only on FAST. Probed on an API 34 emulator:
that session is still RUNNING at 1 s and finished by 2 s, against well
under a second for x265 -preset medium.

A bounded retry. An attempt whose session finished before the cancel
landed has not tested anything, so it is a miss rather than a failure and
is retried; only exhausting five attempts fails, and the message reports
every attempt's state and return code so a real breakage is distinguishable
from a slow machine.

The retry does not soften the test. With FFmpegKit.cancel removed from
both engines, every attempt ends COMPLETED, so both still fail -- verified,
each listing five [state=COMPLETED rc=0] outcomes. Two clean runs
beforehand at 64/0/0/3.

The join test gets the same treatment. It has not flaked yet, but it is
the same mechanism and the same fragility, and finding out on CI again is
not worth the round trip.

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

app/src/androidTest/java/org/libremediaconverter/ffmpeg/FFmpegEngineTest.kt

@@ -4,10 +4,20 @@ import android.media.MediaExtractor
 import android.media.MediaFormat
 import androidx.test.ext.junit.runners.AndroidJUnit4
 import androidx.test.platform.app.InstrumentationRegistry
+import com.arthenica.ffmpegkit.FFmpegKit
+import com.arthenica.ffmpegkit.FFmpegSession
+import com.arthenica.ffmpegkit.ReturnCode
+import com.arthenica.ffmpegkit.SessionState
+import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.cancelAndJoin
+import kotlinx.coroutines.delay
+import kotlinx.coroutines.launch
 import kotlinx.coroutines.runBlocking
+import kotlinx.coroutines.withTimeout
 import org.junit.After
 import org.junit.Assert.assertEquals
 import org.junit.Assert.assertTrue
+import org.junit.Assert.fail
 import org.junit.Before
 import org.junit.Test
 import org.junit.runner.RunWith
@@ -139,6 +149,160 @@ class FFmpegEngineTest {
         assertEquals("OggS", magic)
     }
 
+    /**
+     * The percentage itself, which every other test in this class computes and none of them reads.
+     *
+     * `FFmpegEngine` derives progress as `stats.time / durationMs * 100`, and the statistics
+     * callback runs on every conversion here — but every call site omits `onProgress`, so until
+     * this test nothing on any source set had ever looked at the number (#229). #196 covered the
+     * *worker's* progress lambda, and did it with a fake engine that reports whatever the test
+     * tells it to; `ProgressNotificationTest` covers throttling the same way. The arithmetic was
+     * the one part with no reader.
+     *
+     * ## Why the duration is deliberately wrong
+     *
+     * `sample_h264.mp4` is exactly 3.000 s, and this passes **30 s** as the duration. So the
+     * conversion still encodes the whole clip, `stats.time` still climbs to about 3000 ms, and the
+     * reported percentage tops out around **10** rather than 100.
+     *
+     * That is what makes the assertion bite. A range check alone is worthless here: replacing
+     * `percent` with a constant `0` satisfies "every value is in 0..100" and "the values never go
+     * backwards", and so does a list of `[0, 100]`. Pinning the *band* rejects every constant, and
+     * — because the band is a tenth of the way up — it also rejects an implementation that ignores
+     * `durationMs`, which would report ~100 for the same run.
+     *
+     * The bound is deliberately loose (5..25 for an expected 10). The last statistics callback can
+     * land slightly before the final frame, so the peak is "about 3000 ms of a claimed 30 000",
+     * not exactly it.
+     */
+    @Test
+    fun progressIsReportedAsAFractionOfTheDurationItWasGiven() {
+        val seen = mutableListOf<Int>()
+        val out = outputFor("out_progress.mp4")
+        runBlocking {
+            engine.run(
+                request = ConversionRequest(spec = OutputFormat.MP4_H264.spec, quality = QualityTier.BEST),
+                inputPath = input.absolutePath,
+                output = out,
+                // Ten times the fixture's real 3 s. See the KDoc.
+                durationMs = 30_000,
+                onProgress = { percent -> seen += percent },
+            )
+        }
+
+        assertTrue("the statistics callback never reported progress", seen.isNotEmpty())
+        assertTrue("progress out of range: $seen", seen.all { it in 0..100 })
+        assertEquals("progress went backwards: $seen", seen.sorted(), seen)
+        // The band. Rejects any constant, and rejects ignoring durationMs (which would read ~100).
+        val peak = seen.max()
+        assertTrue(
+            "3 s of media against a claimed 30 s should peak near 10%, got $peak from $seen",
+            peak in 5..25,
+        )
+    }
+
+    /**
+     * Cancelling a *running* conversion actually stops the native session.
+     *
+     * Nothing on any source set did this before (#224). Every `cancel` in `app/src/androidTest` is
+     * `WorkManager.cancelWorkById` against work that is **queued or already finished** — the two in
+     * `ReattachOnLaunchTest` cancel a job carrying a one-hour initial delay, and one immediately
+     * after enqueue. On the JVM, `WorkerCancellationTest` and `HardwareFallbackTest`'s cancellation
+     * case drive a `SoftwareTranscoder` double that records the call. No test had ever asked a real
+     * native session to stop. This is `docs/defect-audit.md` **D10**'s forcing condition.
+     *
+     * It is the one path where cancelling wrong is silently expensive rather than loudly broken: a
+     * missed `FFmpegKit.cancel` leaves the native process encoding to completion while the UI says
+     * the job is cancelled, and nothing reports the battery and thermal cost.
+     *
+     * ## Why the assertion is the session's return code, not the output file
+     *
+     * The obvious assertion — the partial output is gone — **cannot fail**, so it would have been a
+     * vacuous test. `invokeOnCancellation` deletes the path, and on POSIX unlinking a file ffmpeg
+     * still holds open leaves ffmpeg writing to the unlinked inode; the path stays gone whether or
+     * not the cancel ever reached the session. Deleting `FFmpegKit.cancel` and keeping
+     * `output.delete()` passes that check every time.
+     *
+     * What distinguishes them is the session's own verdict: a cancelled session ends with the
+     * cancel return code, a completed one ends successfully. That is a fact about the session
+     * rather than about timing, so it is read *after* waiting for the session to leave
+     * [SessionState.RUNNING] rather than at a fixed delay.
+     *
+     * ## Why it cancels on RUNNING rather than on the first progress callback
+     *
+     * Measured. Cancelling from the first `onProgress` was tried first and **failed on a local API
+     * 34 emulator with `state=COMPLETED rc=0`** — every committed fixture is 2-3 s at 320x240, and
+     * the encode finishes before the first statistics callback has been delivered and acted on. The
+     * progress callback proves the session is running, but arrives too late to interrupt anything.
+     * `FFmpegKit.listSessions` shows the session [SessionState.RUNNING] far earlier.
+     *
+     * ## Why it retries, which is the part that took two attempts to get right
+     *
+     * Waiting for `RUNNING` is not on its own enough. With `MP4_H265` at [QualityTier.BEST] this
+     * passed four consecutive local runs and all five CI legs, then failed on the API 34 and 35 legs
+     * of the next PR with `state=COMPLETED rc=0`. Nothing had changed: on a loaded runner the thread
+     * that observed `RUNNING` can be descheduled long enough for a short encode to finish before it
+     * calls `cancel`. A longer timeout does not help — the wait already succeeded.
+     *
+     * Two changes together, because neither is sufficient:
+     *
+     *  - **A slower encode.** `WEBM_VP9` at `BEST` is the slowest thing this builder emits:
+     *    `libvpx-vp9 -crf 31 -b:v 0`, with `-deadline realtime` added **only** on
+     *    [QualityTier.FAST]. Probed on an API 34 emulator, that session is still `RUNNING` at 1 s
+     *    and finished by 2 s, against well under a second for x265 `-preset medium`.
+     *  - **Retrying the attempt.** An attempt whose session finished before the cancel landed has
+     *    not tested anything, so it is not a failure — it is a miss, and it is retried. Only
+     *    exhausting [CANCEL_ATTEMPTS] is a failure, and its message says which case it hit.
+     *
+     * That keeps the mutation honest: with `FFmpegKit.cancel` removed **every** attempt ends
+     * `COMPLETED`, so the test still fails — it just takes [CANCEL_ATTEMPTS] tries to say so.
+     *
+     * The session is identified by diffing against the ids present before each attempt, because
+     * this class has already produced eight of them by the time this executes.
+     */
+    @Test
+    fun cancellingARunningConversionCancelsTheNativeSession(): Unit = runBlocking {
+        val outcomes = mutableListOf<String>()
+
+        repeat(CANCEL_ATTEMPTS) { attempt ->
+            val before = FFmpegKit.listSessions().map { it.getSessionId() }.toSet()
+            val out = outputFor("out_cancelled_$attempt.webm")
+
+            val job = launch(Dispatchers.IO) {
+                engine.run(
+                    // The slowest target this builder emits -- see the KDoc. Not decoration:
+                    // with a faster one this loses the race on a loaded CI runner.
+                    request = ConversionRequest(spec = OutputFormat.WEBM_VP9.spec, quality = QualityTier.BEST),
+                    inputPath = input.absolutePath,
+                    output = out,
+                    durationMs = 3_000,
+                )
+            }
+
+            val ours = withTimeout(TIMEOUT_MS) {
+                var found: FFmpegSession? = null
+                while (found == null) {
+                    found = FFmpegKit.listSessions().firstOrNull { it.getSessionId() !in before }
+                    if (found == null) delay(POLL_MS)
+                }
+                found
+            }
+            job.cancelAndJoin()
+            withTimeout(TIMEOUT_MS) {
+                while (ours.getState() == SessionState.RUNNING) delay(POLL_MS)
+            }
+
+            if (ReturnCode.isCancel(ours.getReturnCode())) return@runBlocking
+            // The encode beat us to it. That attempt proved nothing either way, so try again.
+            outcomes += "state=${ours.getState()} rc=${ours.getReturnCode()}"
+        }
+
+        fail(
+            "never interrupted a running session in $CANCEL_ATTEMPTS attempts, so either every " +
+                "encode finished first or cancellation does not reach it: $outcomes",
+        )
+    }
+
     // --- the quality tier the GPL licence was taken for --------------------
 
     @Test
@@ -176,4 +340,19 @@ class FFmpegEngineTest {
         }.exceptionOrNull()
         assertTrue("expected an FFmpegException, got $failure", failure is FFmpegEngine.FFmpegException)
     }
+
+    private companion object {
+        /** Generous: it bounds a hang, and every wait here normally settles in well under a second. */
+        const val TIMEOUT_MS = 30_000L
+        const val POLL_MS = 50L
+
+        /**
+         * How many times to try to catch the session mid-encode.
+         *
+         * Each miss costs about the length of one VP9 encode -- a second or two -- and a miss is
+         * the loaded-runner case rather than a defect. Five is enough that exhausting them means
+         * cancellation is not reaching the session, which is what the failure message says.
+         */
+        const val CANCEL_ATTEMPTS = 5
+    }
 }

app/src/androidTest/java/org/libremediaconverter/join/ConcatEngineTest.kt

@@ -5,10 +5,20 @@ import android.media.MediaFormat
 import android.net.Uri
 import androidx.test.ext.junit.runners.AndroidJUnit4
 import androidx.test.platform.app.InstrumentationRegistry
+import com.arthenica.ffmpegkit.FFmpegKit
+import com.arthenica.ffmpegkit.FFmpegSession
+import com.arthenica.ffmpegkit.ReturnCode
+import com.arthenica.ffmpegkit.SessionState
+import kotlinx.coroutines.Dispatchers
+import kotlinx.coroutines.cancelAndJoin
+import kotlinx.coroutines.delay
+import kotlinx.coroutines.launch
 import kotlinx.coroutines.runBlocking
+import kotlinx.coroutines.withTimeout
 import org.junit.After
 import org.junit.Assert.assertEquals
 import org.junit.Assert.assertTrue
+import org.junit.Assert.fail
 import org.junit.Before
 import org.junit.Test
 import org.junit.runner.RunWith
@@ -16,6 +26,7 @@ import org.libremediaconverter.convert.MediaProbe
 import org.libremediaconverter.convert.StagingNames
 import org.libremediaconverter.ffmpeg.ConcatEngine
 import org.libremediaconverter.model.ConcatStrategy
+import org.libremediaconverter.work.ConcatWorker
 import java.io.File
 
 /**
@@ -50,6 +61,82 @@ class ConcatEngineTest {
         (staged + listOf(clipA, clipB, clipMismatched)).forEach { it.delete() }
     }
 
+    /**
+     * Cancelling a *running* join actually stops the native session.
+     *
+     * The `FFmpegEngine` half of #224 landed first (PR #236); this is the same gap in
+     * [ConcatEngine]. Before these two, no test on any source set had ever asked a real native
+     * session to stop — every `cancel` in `app/src/androidTest` targets WorkManager entries that
+     * are queued or already finished.
+     *
+     * ## Two things carried over from the conversion side, both measured there
+     *
+     * **The assertion is the session's return code.** A cancelled session ends with the cancel
+     * code, a completed one does not. The alternative — checking the output file — is even less
+     * available here than it was for conversions: [ConcatEngine] does not delete its output on
+     * cancellation at all. Its `invokeOnCancellation` is `FFmpegKit.cancel(...)` and nothing else,
+     * where [org.libremediaconverter.ffmpeg.FFmpegEngine]'s also deletes the partial. Whether that
+     * asymmetry is deliberate is a separate question from this test, which is why this asserts the
+     * thing that is true of both.
+     *
+     * **The cancel is triggered on [SessionState.RUNNING], not on progress.** `ConcatWorker`
+     * publishes no progress at all, so there is no callback to hang it on even in principle — but
+     * the conversion side established the deeper reason: the committed clips are 2 s at 320x240 and
+     * the encode outruns a callback-triggered cancel.
+     *
+     * **And the attempt is retried**, for the reason the conversion side measured the hard way: on
+     * a loaded runner the thread that observed `RUNNING` can be descheduled long enough for a short
+     * encode to finish before it calls `cancel`, which failed two CI legs there. An attempt whose
+     * session finished first has tested nothing, so it is a miss rather than a failure; only
+     * exhausting [CANCEL_ATTEMPTS] fails, and with `FFmpegKit.cancel` removed every attempt misses,
+     * so the mutation still bites.
+     *
+     * The inputs are deliberately the **mismatched** pair, so [ConcatStrategy.REENCODE] is chosen.
+     * A stream copy of two short clips is close to instantaneous and would leave nothing to
+     * interrupt; re-encoding is the case where a user would actually reach for Cancel.
+     *
+     * *Mutation:* drop `FFmpegKit.cancel(session.getSessionId())` from `ConcatEngine`'s
+     * `invokeOnCancellation` — the session runs to completion and this fails.
+     */
+    @Test
+    fun cancellingARunningJoinCancelsTheNativeSession(): Unit = runBlocking {
+        val outcomes = mutableListOf<String>()
+
+        repeat(CANCEL_ATTEMPTS) { attempt ->
+            val before = FFmpegKit.listSessions().map { it.getSessionId() }.toSet()
+            val out = output("cancelled_join_$attempt.mp4")
+
+            val job = launch(Dispatchers.IO) {
+                engine.join(
+                    listOf(Uri.fromFile(clipA), Uri.fromFile(clipMismatched)),
+                    out,
+                    ConcatWorker.DEFAULT_FORMAT,
+                )
+            }
+
+            val ours = withTimeout(TIMEOUT_MS) {
+                var found: FFmpegSession? = null
+                while (found == null) {
+                    found = FFmpegKit.listSessions().firstOrNull { it.getSessionId() !in before }
+                    if (found == null) delay(POLL_MS)
+                }
+                found
+            }
+            job.cancelAndJoin()
+            withTimeout(TIMEOUT_MS) {
+                while (ours.getState() == SessionState.RUNNING) delay(POLL_MS)
+            }
+
+            if (ReturnCode.isCancel(ours.getReturnCode())) return@runBlocking
+            outcomes += "state=${ours.getState()} rc=${ours.getReturnCode()}"
+        }
+
+        fail(
+            "never interrupted a running join in $CANCEL_ATTEMPTS attempts, so either every " +
+                "encode finished first or cancellation does not reach it: $outcomes",
+        )
+    }
+
     private fun copyAsset(name: String): File {
         val out = File(context.cacheDir, name)
         InstrumentationRegistry.getInstrumentation().context.assets
@@ -180,4 +267,13 @@ class ConcatEngineTest {
             a.width != mismatched.width || a.height != mismatched.height,
         )
     }
+
+    private companion object {
+        /** Generous: it bounds a hang, and both waits here normally settle in well under a second. */
+        const val TIMEOUT_MS = 30_000L
+        const val POLL_MS = 50L
+
+        /** See the conversion side: a miss is the loaded-runner case, not a defect. */
+        const val CANCEL_ATTEMPTS = 5
+    }
 }

app/src/androidTest/java/org/libremediaconverter/work/NotificationCancelActionTest.kt

@@ -0,0 +1,129 @@
+package org.libremediaconverter.work
+
+import android.net.Uri
+import androidx.media3.common.util.UnstableApi
+import androidx.test.ext.junit.runners.AndroidJUnit4
+import androidx.test.platform.app.InstrumentationRegistry
+import androidx.work.OneTimeWorkRequestBuilder
+import androidx.work.WorkInfo
+import androidx.work.WorkManager
+import kotlinx.coroutines.flow.first
+import kotlinx.coroutines.runBlocking
+import kotlinx.coroutines.withTimeout
+import org.junit.After
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertNotNull
+import org.junit.Before
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.model.OutputFormat
+import org.libremediaconverter.model.QualityTier
+import java.io.File
+import java.util.concurrent.TimeUnit
+
+/**
+ * The Cancel button in the notification shade actually cancels the job.
+ *
+ * `ConversionNotifications.build` attaches one action, wired to
+ * `WorkManager.createCancelPendingIntent(id)`. Before this test `createCancelPendingIntent` had
+ * **no references anywhere outside its own declaration** — no JVM test, no instrumented test
+ * (#227).
+ *
+ * That matters more than an ordinary uncovered line. A conversion runs in a foreground service and
+ * the user is invited to leave the app; once they do, this action is the only way to stop it. If
+ * the `PendingIntent` carries the wrong id, the button does nothing, the notification stays, and
+ * the job runs to completion — with no error, no log, and no screen to look at.
+ *
+ * ## Why this fires the intent rather than reading the shade
+ *
+ * The obvious version asks `NotificationManager.getActiveNotifications()` for id 1001 and taps what
+ * it finds. That was rejected: the instrumented suite grants no runtime permissions, so
+ * `POST_NOTIFICATIONS` is denied throughout, and whether a suppressed foreground-service
+ * notification is returned there is a platform detail that varies — the test would be asserting
+ * something about notification *visibility* rather than about cancellation.
+ *
+ * The `PendingIntent` is the subject; where it is read from is incidental. Building the
+ * notification for a real, live work id and firing its action exercises exactly the thing that can
+ * be wrong — a real `PendingIntent` dispatch reaching real `WorkManager` — and does it the same way
+ * on every API level.
+ *
+ * ## Why the job is delayed rather than running
+ *
+ * A conversion of the committed 3 s fixture finishes in well under a second on an emulator
+ * (`HardwareFallbackTest` completed one in 448 ms), so racing a cancel against a running job would
+ * be flaky in the direction that fails. An initial delay keeps the job reliably `ENQUEUED`, which
+ * is a state `cancelWorkById` acts on identically — what is under test is whether firing the action
+ * reaches WorkManager with the right id, not which state it interrupts.
+ *
+ * *Mutation:* build the `PendingIntent` from `UUID.randomUUID()` instead of the request's id. The
+ * notification looks identical and the job is never cancelled.
+ */
+@UnstableApi
+@RunWith(AndroidJUnit4::class)
+class NotificationCancelActionTest {
+
+    private val context = InstrumentationRegistry.getInstrumentation().targetContext
+    private val workManager = WorkManager.getInstance(context)
+    private lateinit var input: File
+
+    @Before
+    fun setUp() {
+        input = File(context.cacheDir, "cancel_action_sample.mp4")
+        InstrumentationRegistry.getInstrumentation().context.assets
+            .open("sample_h264.mp4")
+            .use { asset -> input.outputStream().use { asset.copyTo(it) } }
+    }
+
+    @After
+    fun tearDown() {
+        input.delete()
+        File(context.cacheDir, "conversions").listFiles()?.forEach { it.delete() }
+    }
+
+    @Test
+    fun theNotificationsCancelActionCancelsThatJob(): Unit = runBlocking {
+        val request = ConversionWorker.request(
+            inputUri = Uri.fromFile(input),
+            displayName = input.name,
+            sizeBytes = input.length(),
+            spec = OutputFormat.MP4_H264.spec,
+            quality = QualityTier.FAST,
+        ).let { base ->
+            // Rebuild with a delay so the job stays ENQUEUED for the whole test. See the KDoc.
+            OneTimeWorkRequestBuilder<ConversionWorker>()
+                .setInputData(base.workSpec.input)
+                .setInitialDelay(1, TimeUnit.HOURS)
+                .build()
+        }
+        workManager.enqueue(request).result.get()
+
+        // The job is queued and waiting, which is the state the cancel has to interrupt.
+        assertEquals(
+            WorkInfo.State.ENQUEUED,
+            withTimeout(TIMEOUT_MS) {
+                workManager.getWorkInfoByIdFlow(request.id).first { it != null }
+            }?.state,
+        )
+
+        val notification = ConversionNotifications(context)
+            .build(request.id, title = input.name, percent = 0, indeterminate = true)
+        val action = notification.actions?.firstOrNull()
+        assertNotNull("the progress notification carries no action to cancel with", action)
+
+        // The whole point: fire it the way the shade would, and see the job stop.
+        action!!.actionIntent.send()
+
+        val terminal = withTimeout(TIMEOUT_MS) {
+            workManager.getWorkInfoByIdFlow(request.id).first { it != null && it.state.isFinished }
+        }
+        assertEquals(
+            "firing the notification's Cancel action must cancel the job it was built for",
+            WorkInfo.State.CANCELLED,
+            terminal?.state,
+        )
+    }
+
+    private companion object {
+        const val TIMEOUT_MS = 30_000L
+    }
+}