Stop the match line claiming a failure count nobody measured

This PR's own advisory leg caught it. With five markers and a truncated run it printed failed: 4 ... baseline: matches (5 expected, 5 failed) three lines apart. The match line has always printed the baseline twice, which was true while `failed` had to equal it to get there -- and the previous commit removed that requirement for truncated runs without noticing this line depended on it. So the truncated spelling says what happened: `matches (5 expected; 4 of 5 failed, on a run the abort truncated -- not compared)`. Pinned by a third case beside the two from that commit. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Merge pull request #245 from JMR-dev/fix/api37-task-snapshot-crash
2026-09-06 09:03:21 -05:00 · 2026-09-06 09:01:56 -05:00 · 2026-09-06 08:53:59 -05:00 · 2026-09-06 08:50:59 -05:00 · 2026-09-06 08:49:59 -05:00 · 2026-09-06 08:43:20 -05:00
29 changed files with 2301 additions and 203 deletions
@@ -184,6 +184,46 @@ out="$(run_report "$root")"
 assert_contains "same-line annotation removed: counts 2, so it was worth 1" "$out" \
  "  baseline DEVIATION: the tree carries 2 tests marked \`@FailsOnEmulatorApi37\` but the baseline says 3 — update FAILS_ON_EMULATOR_API37_BASELINE"

+# ---------------------------------------------------------------------------
+# 4. A run the abort truncated, with fewer failures than the baseline: NOT a deviation.
+#
+# `expected` comes from `Starting N tests`, printed before anything can abort, so it still
+# answers "is the marked set the size the baseline says". `failed` is a tally of what actually
+# ran, and on a truncated run the tests after the abort never start. Measured on 2026-09-05, two
+# api37-debug dispatches of the same four marked tests: 4/4/4 and then 4/3/3. Announcing the
+# second as "one now passes" is the wrong reading, and #120 is the standing lesson about a notice
+# that is wrong often enough to be skimmed past.
+# ---------------------------------------------------------------------------
+root="$(make_root "$FIXTURE_DIR" 3)"
+cat > "$root/gradle.log" <<'TRUNCATED'
+> Task :app:connectedDebugAndroidTest
+Starting 3 tests on test(AVD) - 16
+There was 2 failure(s).
+Test run failed to complete. Expected 3 tests, received 2. onError: commandError=false message=INSTRUMENTATION_ABORTED: System has crashed.
+TRUNCATED
+out="$(run_report "$root")"
+assert_contains "truncated run: the truncation is reported"  "$out" '  completed cleanly: no'
+assert_absent   "truncated run: the short failure count is not a deviation" "$out" 'tests failed, the baseline is'
+# And the match line has to say what actually happened rather than repeat the baseline: PR #245's
+# advisory leg printed `failed: 4` three lines above `matches (5 expected, 5 failed)`.
+assert_contains "truncated run: the match line does not claim the baseline's failure count" "$out" \
+  '  baseline: matches (3 expected; 2 of 3 failed, on a run the abort truncated — not compared)'
+
+# ---------------------------------------------------------------------------
+# 5. The same short failure count on a run that finished IS a deviation.
+#
+# The pair is the point: case 4 must not have bought its quiet by disabling the check outright.
+# ---------------------------------------------------------------------------
+root="$(make_root "$FIXTURE_DIR" 3)"
+cat > "$root/gradle.log" <<'CLEAN'
+> Task :app:connectedDebugAndroidTest
+Starting 3 tests on test(AVD) - 16
+There was 2 failure(s).
+CLEAN
+out="$(run_report "$root")"
+assert_contains "clean run, short by one: the deviation fires" "$out" \
+  '2 tests failed, the baseline is 3'
+
 echo
 if [ "$failures" -eq 0 ]; then
  echo "e2e-report-shape-test.sh: all checks passed"
@@ -252,8 +252,22 @@ if [ -n "$baseline" ]; then
    if [ "$expected" != "unknown" ] && [ "$expected" != "$baseline" ]; then
      deviations+=("the runner started $expected tests, the baseline is $baseline")
    fi
+    # `expected` is compared on every run and `failed` only on a run that finished, and the
+    # difference is the truncation this file already records rather than compares. `expected`
+    # comes from `Starting N tests`, which is printed before anything can abort, so it answers
+    # "is the marked set the size the baseline says" whatever happens afterwards. `failed` is a
+    # tally of what actually ran: on a truncated run the tests after the abort never start, so
+    # comparing it to the baseline announces a deviation about the framework dying rather than
+    # about the test list. Measured on 2026-09-05, two api37-debug dispatches of the same four
+    # marked tests: 4/4/4 and then 4/3/3, the second having lost the last test to the abort.
+    # Announcing that as "one now passes" is exactly the wrong reading, and #120 is the standing
+    # lesson about a notice that is wrong often enough to be skimmed past.
    if [ "$failed" != "unknown" ] && [ "$failed" != "$baseline" ]; then
-      deviations+=("$failed tests failed, the baseline is $baseline — every test carrying the marker is expected to fail on this image, so fewer means one now passes and more means a new one joined")
+      if [ "$completed" = "**no**" ]; then
+        echo "::debug::$failed of $baseline marked tests failed, on a run the abort truncated — not compared"
+      else
+        deviations+=("$failed tests failed, the baseline is $baseline — every test carrying the marker is expected to fail on this image, so fewer means one now passes and more means a new one joined")
+      fi
    fi
  fi
  if [ -n "$marked" ] && [ "$marked" != "$baseline" ]; then
@@ -283,7 +297,16 @@ if [ -n "$failed_names" ]; then
 fi
 if [ "$advisory" = "yes" ]; then
  if [ "${#deviations[@]}" -eq 0 ]; then
-    echo "  baseline: matches ($baseline expected, $baseline failed)"
+    # Two spellings, because one of them would be a lie half the time. `$baseline expected,
+    # $baseline failed` is only true of a run that finished; on a truncated one `failed` is a
+    # tally of the tests that got to run before the framework died, and printing the baseline in
+    # its place claims a number nobody measured. Seen on PR #245's advisory leg, which reported
+    # `failed: 4` three lines above `matches (5 expected, 5 failed)`.
+    if [ "$failed" != "unknown" ] && [ "$failed" != "$baseline" ]; then
+      echo "  baseline: matches ($baseline expected; $failed of $baseline failed, on a run the abort truncated — not compared)"
+    else
+      echo "  baseline: matches ($baseline expected, $baseline failed)"
+    fi
  else
    printf '  baseline DEVIATION: %s\n' "${deviations[@]}"
  fi
@@ -52,40 +52,72 @@ WEDGE_TIMEOUT=1200
 # the same shape as E2E_EXTRA_GRADLE_ARGS below. The other four E2E legs run byte-identical
 # commands with it unset.
 #
-# WHY IT RUNS HERE, BEFORE THE LOGCAT STREAM: `adb shell stop` ends the `adb logcat` started
-# below, and nothing restarts it, so a disable performed after that point would cost this leg
-# its whole diagnostic story for the part of the run that matters. Everything this function
-# counts comes from `adb logcat -d -b crash`, which is a fresh read each time and independent
-# of the stream.
+# WHY IT RUNS HERE, BEFORE THE LOGCAT STREAM: it is a 45-second wait, and the stream below is
+# meant to cover the suite rather than the wait. Everything this function counts comes from
+# `adb logcat -d -b crash`, a fresh read each time and independent of the stream. (The original
+# reason was stronger and no longer applies: `adb shell stop` would have ended the streamed
+# `adb logcat` and nothing restarts it. There is no `stop` here any more -- see below.)
 #
-# WHAT IT IS FOR: the android-37.x images abort surfaceflinger from RegionSamplingThread inside
-# their own gralloc mapper (docs/api-37-emulator-crash.md). surfaceflinger is a critical service,
-# so init SIGKILLs zygote with it and the framework restarts under the run -- Gradle then reports
+# WHAT IT IS FOR -- AND THE NAME IS NOW WRONG, WHICH IS WHY THIS PARAGRAPH IS LONG.
+# The android-37.x images abort surfaceflinger from RegionSamplingThread inside their own gralloc
+# mapper (docs/api-37-emulator-crash.md). surfaceflinger is a critical service, so init SIGKILLs
+# zygote with it and the framework restarts under the run -- Gradle then reports
 # `cmd: Can't find service: package` and `Starting 0 tests`. RegionSamplingThread exists only
-# because SystemUI registers a nav-bar luma-sampling listener, so removing the package removes
-# the whole chain. Measured cadence of those kills: 20-90 s apart, median 60-70 s, three to five
-# in a four-minute window -- fast enough that install and instrumentation start-up do not fit
-# inside one gap.
+# because SystemUI registers a nav-bar luma-sampling listener, so this was written to remove the
+# package and with it the whole chain. Measured cadence of those kills on `-gpu host`: 20-90 s
+# apart, median 60-70 s, three to five in a four-minute window.
+#
+# **THE DISABLE HALF OF THAT HAS NEVER WORKED, AND THE QUIET WINDOW IS WHAT THE LEG ACTUALLY
+# GETS.** Measured 2026-09-05, two ways that agree:
+#
+#   - On CI, in the gating leg of run 34006456986: `pm disable-user` is accepted at 02:28:37.9 and
+#     `com.android.systemui` really is in `pm list packages -d` at 02:29:33 -- and SystemUI is
+#     started anyway at 02:28:39.5 and again at 02:28:52.3, the second of which (pid 4275) is
+#     alive for the whole instrumentation run, logging `WindowManagerShell ...
+#     app=com.android.systemui` minutes after this function prints its final line.
+#   - Locally on android-37.0, with the package verified disabled before AND after a deliberate
+#     `stop; start`: `com.android.systemui` comes up 3 s after `system_server` regardless.
+#
+# So `pm disable-user --user 0 com.android.systemui` does not stop SystemUI starting on this
+# image, whatever else happens. The name `E2E_DISABLE_SYSTEM_UI` and the name of this function are
+# kept because the matrix row, both workflows and two documents refer to them, and a rename would
+# touch all of that to no benefit -- read this comment, not the name.
+#
+# WHAT IS LEFT IS LOAD-BEARING, so do not delete the function as dead weight. It is the 45-second
+# window with zero new `hasReadColorBufferDma` aborts. The boot-time aborts land close together --
+# 02:28:18 and 02:28:43 in that same run -- and the wait is what puts instrumentation (02:32:42)
+# after them rather than inside one. That is what stops a leg reporting `Starting 0 tests`, and it
+# is why the three-round retry stays.
+#
+# THE `pm disable-user` CALL STAYS TOO, for a narrower reason than it was written for: every green
+# leg and every measurement quoted anywhere about this row was taken with it applied and SystemUI
+# running. Removing it would change the configuration the numbers came from, which is not a change
+# to make while fixing a flake.
+#
+# AND THE FRAMEWORK RESTART IS GONE, having been measured to be worse than nothing. It was written
+# as `adb shell stop; adb shell start`, which are root-only; adbd is not root, so every leg printed
+# `Must be root` twice and restarted nothing. Adding `adb root` made it real, and api37-debug run
+# 34010167885 is what that looks like: `pm disable-user` reports success, the stop lands ~2 s later
+# and kills system_server before PackageManager has flushed its delayed write of package
+# restrictions, so the state is gone on the way back up -- `NOT DISABLED after the restart`, three
+# rounds, `final state: SystemUI STILL ENABLED`, and the leg then reported `expected: 0,
+# received: 0`. A 15 s pause before the stop does make the state survive (bisected locally), and it
+# still does not help, because of the two measurements above. So the restart is removed rather than
+# repaired: it cost the leg every test it had, and there is nothing for it to buy.
 #
 # NOTHING HERE TRUSTS A COMMAND'S OWN REPORT, and that is not paranoia: of four runs of an
 # earlier one-shot version, one (32646029143) reported `new state: disabled-user` and then
-# started SystemUI eight more times, with ten more aborts. `pm disable-user` can be accepted by
-# a system_server that is SIGKILLed before the state is written, and `pm disable-user` does not
-# retract SystemUI's existing region-sampling registration either -- by the time boot completes
-# it has already registered, so only a framework restart brings back a SystemUI-less
-# surfaceflinger. Hence: disable, take the framework DOWN and confirm system_server is really
-# gone (an earlier probe asked `service check` 0.3 s after `stop` and got `found` from the
-# system_server that was still exiting, so its wait was not a wait), bring it back, verify the
-# package against `pm list packages -d`, and require a 45 s window with zero new aborts.
-# Three rounds, because one is not reliable and the failure is silent.
+# started SystemUI eight more times. So this reports what `pm list packages -d` says AND what
+# `pidof` says, side by side, rather than one line implying both.
 # ---------------------------------------------------------------------------
 count_aborts() { adb logcat -d -b crash 2> /dev/null | grep -c 'hasReadColorBufferDma'; }
 systemui_disabled() { adb shell pm list packages -d 2> /dev/null | grep -q 'com.android.systemui'; }
+systemui_pid() { adb shell pidof com.android.systemui 2> /dev/null | tr -d '\r\n'; }

 disable_region_sampling() {
-  local round=1 i out before after
+  local round=1 i out pid before after
  while [ "$round" -le 3 ]; do
-    echo "--- SystemUI disable, round $round ---"
+    echo "--- round $round ---"
    for i in $(seq 1 10); do
      out="$(adb shell pm disable-user --user 0 com.android.systemui 2>&1 | tr -d '\r')"
      echo "  pm attempt $i: $out"
@@ -93,36 +125,20 @@ disable_region_sampling() {
      sleep 5
    done

-    echo "  restarting the framework"
-    adb shell stop
-    for i in $(seq 1 20); do
-      [ -z "$(adb shell pidof system_server 2> /dev/null | tr -d '\r\n')" ] && break
-      sleep 2
-    done
-    echo "  system_server down after ~$((i * 2)) s"
-    adb shell start
-    for i in $(seq 1 30); do
-      if adb shell service check package 2> /dev/null | grep -q ': found' \
-        && adb shell service check activity 2> /dev/null | grep -q ': found' \
-        && [ -n "$(adb shell pidof system_server 2> /dev/null | tr -d '\r\n')" ]; then
-        echo "  services back after ~$((i * 5)) s"
-        break
-      fi
-      sleep 5
-    done
-
    if systemui_disabled; then
-      echo "  verified: com.android.systemui is in pm list packages -d"
+      echo "  pm list packages -d: com.android.systemui is in it"
    else
-      echo "  NOT DISABLED after the restart -- the package state did not survive"
-      round=$((round + 1))
-      continue
+      echo "  pm list packages -d: com.android.systemui is NOT in it"
    fi
+    # Printed next to the line above precisely because the two disagree on this image, and a
+    # reader who sees only the first will believe something that is not true.
+    pid="$(systemui_pid)"
+    echo "  com.android.systemui pid: ${pid:-none} (expected: a pid -- see the header)"

    before="$(count_aborts)"
    sleep 45
    after="$(count_aborts)"
-    echo "  abort rate, SystemUI disabled: $((after - before)) new in 45 s (total ${after:-0})"
+    echo "  aborts: $((after - before)) new in 45 s (total ${after:-0})"
    [ "$((after - before))" -eq 0 ] && break
    echo "  still aborting after round $round"
    round=$((round + 1))
@@ -131,16 +147,16 @@ disable_region_sampling() {
  # A warning rather than an exit. If the disable did not take, the run is about to report
  # `Starting 0 tests` and fail on its own -- and it will do so with the logcat, the crash
  # buffer and the diagnostics attached, which is more useful than dying here with none of it.
-  if systemui_disabled; then
-    echo "  final state: SystemUI disabled"
+  if [ "$((after - before))" -eq 0 ]; then
+    echo "  final state: 45 s with no new aborts -- the suite starts here"
  else
-    echo "::warning::E2E api${LABEL}: SystemUI is still enabled -- expect INSTRUMENTATION_ABORTED"
+    echo "::warning::E2E api${LABEL}: still aborting after three rounds -- expect INSTRUMENTATION_ABORTED"
  fi
  return 0
 }

 if [ "${E2E_DISABLE_SYSTEM_UI:-}" = "1" ]; then
-  echo "::group::E2E api${LABEL} -- removing the region-sampling listener"
+  echo "::group::E2E api${LABEL} -- waiting out the boot-time gralloc aborts"
  disable_region_sampling
  echo "::endgroup::"
 fi
@@ -28,6 +28,15 @@ name: API 37 debug
 #   - It does not fork .github/scripts/e2e-run.sh. That script owns the FAILED-vs-WEDGED
 #     split, the SIGQUIT thread dump and the streamed logcat, and it is the copy CI
 #     exercises every day. This calls it, exactly as status_check.yml does.
+#
+#     The SystemUI disable below is the exception, and it is a real one: this workflow
+#     drives it from its own probe step so `disable_system_ui` can be turned off for a
+#     dispatch, where the real leg gets it through `E2E_DISABLE_SYSTEM_UI`. Two copies of
+#     that logic therefore exist and must be changed together. **This instrument is also
+#     what established that the disable half of it does nothing** -- run 34010167885, in
+#     which making its framework restart real cost the leg every test it had. Read
+#     .github/scripts/e2e-run.sh's header for the measurements; the restart is gone from
+#     both copies and what remains is the 45-second quiet window.
 #   - It does not change status_check.yml. If a configuration here turns out to work,
 #     the change to the real matrix is proposed separately.
 #
@@ -291,45 +300,30 @@ jobs:
                sleep 5
              done

-              # pm disable-user does not retract SystemUI's existing region-sampling
-              # registration -- by the time boot completes it has already registered. Only a
-              # framework restart brings back a SystemUI-less SurfaceFlinger. See
-              # disable_region_sampling in tools/local-emulator/run-e2e.sh.
-              echo "  restarting the framework"
-              adb shell stop
-              for i in $(seq 1 20); do
-                [ -z "$(adb shell pidof system_server 2> /dev/null | tr -d '\r\n')" ] && break
-                sleep 2
-              done
-              echo "  system_server down after $((i * 2)) s"
-              adb shell start
-              for i in $(seq 1 30); do
-                if adb shell service check package 2> /dev/null | grep -q ': found' \
-                  && adb shell service check activity 2> /dev/null | grep -q ': found' \
-                  && [ -n "$(adb shell pidof system_server 2> /dev/null | tr -d '\r\n')" ]; then
-                  echo "  services back after $((i * 5)) s"
-                  break
-                fi
-                sleep 5
-              done
-
+              # NO FRAMEWORK RESTART. There was one here, and making it work (it needed
+              # `adb root`) is what proved the whole disable is ineffective on this image:
+              # SystemUI starts anyway, measured on CI and locally, and the restart itself
+              # loses the package state to PackageManager's delayed write and leaves the leg
+              # reporting `Starting 0 tests`. e2e-run.sh's header carries the measurements.
+              # What is left, and what is load-bearing, is the quiet window below.
              if systemui_disabled; then
-                echo "  verified: com.android.systemui is in pm list packages -d"
+                echo "  pm list packages -d: com.android.systemui is in it"
              else
-                echo "  NOT DISABLED after the restart -- the package state did not survive"
-                round=$((round + 1))
-                continue
+                echo "  pm list packages -d: com.android.systemui is NOT in it"
              fi
+              # Beside it, because the two disagree on this image and the first line alone
+              # reads as a claim about the process that is not true.
+              echo "  com.android.systemui pid: $(adb shell pidof com.android.systemui 2> /dev/null | tr -d '\r\n')"

              before="$(count_aborts)"
              sleep 45
              after="$(count_aborts)"
-              echo "--- abort rate, SystemUI disabled: $((after - before)) new in 45 s (total ${after:-0}) ---"
+              echo "--- aborts: $((after - before)) new in 45 s (total ${after:-0}) ---"
              [ "$((after - before))" -eq 0 ] && break
              echo "  still aborting after round $round"
              round=$((round + 1))
            done
-            systemui_disabled && echo "final state: SystemUI disabled" || echo "final state: SystemUI STILL ENABLED -- expect Starting 0 tests"
+            systemui_disabled && echo "final state: com.android.systemui is disabled in pm (it still runs)" || echo "final state: com.android.systemui is not even disabled in pm"
          fi

          echo "--- crash buffer (tail 60) ---"
@@ -254,31 +254,31 @@ jobs:
            api-level: "36"
          # API 37, and it is NOT the same device as the four rows above it.
          #
-          # CAVEAT, read this before trusting a green here: this leg runs with
-          # SystemUI disabled and the framework restarted under it. No other leg
-          # and no Pixel run uses that configuration. It is defensible only because
-          # nothing THIS LEG RUNS touches system UI -- Media3, FFmpeg and
-          # WorkManager tests -- and because the alternative is no CI coverage of
-          # the level this app targets. **Anything that ever does depend on system
-          # UI must not trust this row.** E2E_DISABLE_SYSTEM_UI is what does it;
-          # .github/scripts/e2e-run.sh explains the mechanism and why every step of
-          # it is verified rather than assumed.
+          # THE CAVEAT THAT USED TO BE HERE IS WITHDRAWN, 2026-09-05, and the
+          # withdrawal is good news. It said this leg "runs with SystemUI disabled
+          # and the framework restarted under it", that no other leg or Pixel run
+          # uses that configuration, and that anything depending on system UI must
+          # not trust this row. **None of that was ever true.** Measured: the
+          # framework restart is two root-only adb commands that answered `Must be
+          # root` on every leg ever run, and `pm disable-user` does not stop SystemUI
+          # starting on this image anyway -- in run 34006456986 the package is
+          # verified disabled at 02:29:33 and SystemUI (pid 4275) is up from 02:28:52
+          # for the whole run. So this row's device configuration is the same as the
+          # other four's, and a green here means what a green on 33-36 means.
          #
-          # "this leg" and not "this suite", since 2026-08-24, and the difference is
-          # now load-bearing: SafPickerRoundTripTest DOES touch system UI. It drives
-          # DocumentsUI and rotates the display, and both reach the gralloc mapper
-          # this image aborts in -- disabling SystemUI removes the IDLE trigger, not
-          # those. Measured per method on android-37.0: the ROTATION test takes the
-          # framework down (INSTRUMENTATION_ABORTED) and carries
-          # @FailsOnEmulatorApi37, so notAnnotation below keeps it off this row; the
-          # PICKER test passes and runs here like anything else. A rotation rebuilds
-          # every surface at once, and starting another app's activity does not.
+          # E2E_DISABLE_SYSTEM_UI still exists and still runs, because what it
+          # actually buys is a 45-second window with no new gralloc aborts before the
+          # suite starts -- the boot-time ones land close together and instrumentation
+          # has to begin after them, not between them. The name is stale and kept:
+          # read .github/scripts/e2e-run.sh's header, which carries the measurements.
          #
-          # So this row does now run one test that depends on system UI, and the
-          # caveat above still applies to it: a green here is not evidence the picker
-          # works on a device with SystemUI running -- the Pixel release check is.
-          # docs/api-37-emulator-crash.md has the per-method measurements, and the
-          # correction that produced them.
+          # notAnnotation below keeps five tests off this row, and one of
+          # them is new. SafPickerRoundTripTest's PICKER test was measured on
+          # 2026-08-24 as passing here and was left on the leg; four gating logcats
+          # read on 2026-09-05 show it aborting system_server from the task-snapshot
+          # path on every single run, pass or fail, which is what had been failing
+          # unrelated PRs (#108). Both of that class's tests now carry the marker.
+          # docs/api-37-emulator-crash.md has the timings and the correction.
          #
          # 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
@@ -76,17 +76,45 @@ days. Read it as the current answer, and see the git history if you need the old
  `angle_indirect` and `swangle_indirect` all boot, while `auto`, `off`, `guest` and
  `swiftshader_indirect` do not. `docs/local-emulator.md` has the evidence and the per-API renderer
  table.
- **CI runs API 37, and it gates.** The matrix is 33/34/35/36/37. **Three** of the 60 instrumented
-  tests cannot pass on that image, for two unrelated reasons: two Media3 hardware transcodes fail
-  inside the emulator's own `c2.goldfish.h264.decoder`, and one SAF test takes the framework down
-  when it rotates the display. All three carry `@FailsOnEmulatorApi37` and run in a separate
-  `continue-on-error` job; the gating leg runs the other 57.
+- **CI runs API 37, and it gates.** The matrix is 33/34/35/36/37. **Five** of the 69 instrumented
+  tests cannot be *run* on that image, for three unrelated reasons: three Media3 tests fail inside
+  the emulator's own `c2.goldfish.h264.decoder`, one SAF test takes the framework down when it
+  rotates the display, and its sibling — the SAF picker round trip — aborts `system_server` from
+  the task-snapshot path whether it passes or not. All five carry `@FailsOnEmulatorApi37` and run
+  in a separate `continue-on-error` job; the gating leg runs the other 64.
+
+  **These two numbers move with the suite and are derived, not remembered.** `grep -cE
+  '^\s*@Test' ` over `app/src/androidTest` is the first; the second is that minus the marker
+  count `.github/scripts/e2e-report-shape.sh` greps. Cross-check against any run's shape rather
+  than trusting the sentence: a leg below 37 reports the first as `expected`, and the API 37
+  gating leg reports the second.
+
+  **That third reason is why "cannot pass" became "cannot be run" on 2026-09-05.** Four gating
+  runs were read logcat-first — 34006456986, 34001744574, 34001377499 and the green 34002313300 —
+  and each carries exactly two `hasReadColorBufferDma` aborts before the suite (surfaceflinger,
+  during boot and the SystemUI disable) and exactly **one** during it: `system_server`, thread
+  `TaskSnapshotPer`, always inside the picker test's window, and nothing else in the gating set
+  reached the mapper at all. Whether the leg went red was luck — one run passed the test and lost
+  the leg anyway with `failed: 0`, another passed it 0.6 s after the abort and went green. That is
+  #108, it cost roughly a third of the gating legs over the wave-4 landings (#190), and a marker
+  is what it needed. `docs/api-37-emulator-crash.md` has the timings.
+
+  **A second thing came out of those logcats, and it withdraws a caveat rather than adding one.**
+  The API 37 row was documented as the one leg running "with SystemUI disabled and the framework
+  restarted under it", which nothing else does. Neither half was ever happening: `adb shell stop`
+  and `start` are root-only and answered `Must be root` on every leg ever run, and `pm
+  disable-user` does not stop SystemUI starting on this image anyway — measured on CI and locally,
+  with and without a real restart. **So this row's device configuration is the same as the other
+  four's, and a green here means what a green at 33–36 means.** `E2E_DISABLE_SYSTEM_UI` is kept
+  under its now-stale name because what it really buys is a 45-second window with no new gralloc
+  aborts before the suite starts, which is load-bearing; `.github/scripts/e2e-run.sh`'s header is
+  where that is written down.

  That job is still called `E2E API 37 Media3 hardware transcode (advisory)`, which no longer
  describes everything in it. The name is kept deliberately — it is not a required context and
  people have learned to look for it — so **read the marker, not the name**, for what it holds.
  **It is red on every PR, by design**: do not read it as your change breaking something, and do
-  not read a green run as evidence those three tests pass.
+  not read a green run as evidence those five tests pass.
  `docs/api-37-emulator-crash.md` has the measurements.

  **That instruction is also why nobody looks, so the job now reports its own shape** — expected,
@@ -106,7 +134,7 @@ days. Read it as the current answer, and see the git history if you need the old
  is gradle never returning, so the log it left says nothing about it.

 Still true, and the reason the advisory job is not simply deleted: **API 37 needs a manual check on
-the Pixel 10 Pro XL before each release.** Those three tests are the one thing CI cannot answer
+the Pixel 10 Pro XL before each release.** Those five tests are the one thing CI cannot answer
 for.

 On a device or emulator, build only the ABI it can execute:
@@ -130,9 +158,9 @@ 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 +311,73 @@ 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.
+
+  **Every number above is `testDebugUnitTest` only, and on 2026-09-05 the instrumented suite got its
+  first read for that reason** — `docs/e2e-read-findings.md`, entries **E1-E6**, tickets
+  **#223-#230**. Four waves had been steered by a figure that **cannot see `app/src/androidTest` at
+  all**, so nothing had ever asked what those 60 device tests pin, only that they were green.
+
+  **It found one test that passes while testing nothing, and it is the one that matters most.**
+  `HardwareFallbackTest` is the only automated check of the hardware→software fallback against a
+  *real* codec failure, and on run `34004304566` the API 33, 34, 35 and 37 legs each log
+  `Routing sample_h264_444.mp4 -> ... via FFMPEG (NO_HARDWARE_ENCODER)` (API 36's logcat artifact on
+  that run is truncated, so it is unread rather than different): emulators expose no
+  hardware encoder, so the job never reaches Media3 and the `catch` it exists to prove is never
+  entered. Its two assertions — succeeded, output non-empty — are true anyway, and it finishes in
+  448 ms. **Deleting that `catch` reddens nothing on any leg** (#223).
+
+  Two things generalise from it. **A test can assert and still not reach**, which no coverage
+  number and no "does it assert something" review would catch — the filter that works is *does this
+  test's premise hold on the machine that runs it?*. And the codebase **already knew**: the sibling
+  `ForcedFailureTest` pins `DeviceCodecs.PERMISSIVE` against exactly this hazard and writes out why,
+  as does `ConversionWorkerTest`. The difference is that their assertions are about the *path*, so
+  without the pin they would fail loudly; `HardwareFallbackTest`'s are about the *output*, so it
+  passes quietly. **Prefer asserting the path over asserting the artefact** where the two differ.
+
+  The read was a triage, not a test push, and six of its seven findings are prose rather than code —
+  the suite itself is in good shape. What had drifted is its self-description.
+
+  **Working the tickets then found the thing the read could not: one production defect.** #238 —
+  joining files picked through the system picker failed outright on the stream-copy path. The
+  concat demuxer whitelists protocols separately from `-safe 0`, and `ffkitsaf` was not on the
+  list; only `STREAM_COPY` feeds it a list file, and every existing join test passed
+  `Uri.fromFile`, so **the one broken combination was the only one a user could reach**. Not a
+  missed line and not an unasserted value — two covered things no test put together, which is the
+  gap shape a coverage number is worst at.
+
+  **E7 is the other reusable result**, because it re-scoped its own ticket. A real
+  `DocumentsProvider` cannot be reached without the picker: an unprotected one is refused at
+  install, instrumentation runs in the app's uid so the test APK's identity is no help, and shell
+  identity is denied too — each denial naming `ACTION_OPEN_DOCUMENT`. So #226 has no cheap headless
+  half. But the *input* bridge needs no documents provider at all, which is what kept #225 headless
+  and is how #238 surfaced.
+
 - **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 +505,29 @@ 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.
@@ -42,6 +42,28 @@
                <action android:name="android.content.action.DOCUMENTS_PROVIDER" />
            </intent-filter>
        </provider>
+
+        <!--
+          A PLAIN provider, for the ffkitsaf bridge on the success path.
+
+          FFmpegKitConfig.getSafParameterForRead is on every real user conversion and was on no
+          passing test: they all pass Uri.fromFile, which takes the other arm. Only its failure
+          side was covered, by UnopenableUriTest naming an authority that does not exist.
+
+          The documents provider above cannot serve this. Any DOCUMENTS_PROVIDER must hold
+          MANAGE_DOCUMENTS or the platform refuses to install it, instrumentation runs in the
+          target app's process and so carries the app's uid, and the resulting denial says what
+          is actually required: access obtained through ACTION_OPEN_DOCUMENT. That means a picker,
+          and the flake it brings. See issue #226.
+
+          The bridge does not need a documents provider. It opens a descriptor through the
+          resolver and hands FFmpeg a saf: path, so any readable content:// URI exercises it, and
+          an ordinary provider is allowed to be exported without a permission.
+        -->
+        <provider
+            android:name="org.libremediaconverter.saf.FixtureContentProvider"
+            android:authorities="org.libremediaconverter.test.content"
+            android:exported="true" />
    </application>

 </manifest>
@@ -1,7 +1,7 @@
 package org.libremediaconverter

 /**
- * Marks an instrumented test that does not pass on the `android-37.x` **emulator** system images.
+ * Marks an instrumented test that cannot be run on the `android-37.x` **emulator** system images.
 *
 * This is a marker, not a skip. Nothing reads it except CI, and CI reads it twice — once with
 * `notAnnotation` to build the gating API 37 leg, and once with `annotation` to build the advisory
@@ -9,6 +9,15 @@ package org.libremediaconverter
 * That is the whole reason there is one annotation rather than a pair of test lists: two lists
 * drift, and the drift is silent in both directions (a test that runs nowhere reads as green).
 *
+ * **"Cannot be run" covers two things, and it said only the first until 2026-09-05.** Four of the
+ * five carriers simply fail: three Media3 tests die in the image's own `c2.goldfish.h264.decoder`,
+ * and the SAF rotation test takes the framework down with it. The fifth —
+ * `SafPickerRoundTripTest.pickingAFileThroughTheSystemPickerFillsInTheFileCard` — **passes about
+ * half the time and aborts `system_server` every time**, which is worse for a gating leg than an
+ * honest failure: it fails the leg from the teardown, with no failing test to point at (#108).
+ * The wording was widened rather than the test excused; that test's own KDoc has the four-run
+ * measurement.
+ *
 * It says only what has been measured: **on the emulator, at API 37.** The same tests pass on a
 * physical Pixel 10 Pro XL at API 37 and at API 33–36 on the same runner under the same renderer,
 * so this must never be read as "this test is allowed to fail at API 37" — only as "the API 37
@@ -17,7 +26,7 @@ package org.libremediaconverter
 *
 * Removing it is the goal, and the trigger is written down: a new API 37.x system image, or an
 * ATD image for 37. Delete the annotation from the tests, and the advisory job goes empty and
- * the gating one grows by two.
+ * the gating one grows by [FAILS_ON_EMULATOR_API37_BASELINE].
 *
 * **How many tests carry it is committed below**, as [FAILS_ON_EMULATOR_API37_BASELINE], and the
 * advisory job checks the run against it. Adding or removing a marker means changing that number
@@ -37,11 +46,28 @@ annotation class FailsOnEmulatorApi37
 * keep printing with nothing to compare to, so it announces that it could not read the baseline
 * rather than falling quiet. If you see that notice, this line is what it means.
 *
- * **One number, both checks, and that is what the marker means.** A test carrying it cannot pass
+ * **One number, both checks, and that is what the marker means.** A test carrying it cannot be run
 * on this image, so the count is simultaneously how many the advisory leg runs and how many fail.
 * A *smaller* failure count is the interesting direction: it means one of them now passes, which
 * is the trigger the KDoc above names for deleting the annotation.
 *
+ * **The picker test is the one to read that sentence carefully for.**
+ * `pickingAFileThroughTheSystemPickerFillsInTheFileCard` was marked on 2026-09-05 for aborting
+ * `system_server` rather than for failing (#108), and on the gating leg it passed two runs of
+ * four. It fails on the advisory leg because the rotation test runs before it and takes the
+ * framework down first — measured, `api37-debug.yml` run 34008889182, which reports
+ * `expected: 4, received: 4, failed: 4` with the four in the order Media3, Media3, rotation,
+ * picker. (Those dispatches predate the third Media3 marker landing on `main`, so their totals
+ * are four rather than five; the ordering they establish is what matters here.)
+ *
+ * **But a second dispatch of the identical configuration reported 4/3/3**, having lost the last
+ * test to the abort rather than to anything about the test list, and that is why
+ * `e2e-report-shape.sh` compares `failed` only on a run that finished. `expected` is compared
+ * always — it comes from `Starting N tests`, which is printed before anything can abort, so it is
+ * the field that answers "is the marked set the size this number says". Read a *clean* run
+ * reporting fewer failures than this as one of them now passing; read a truncated one as the
+ * framework having died, which is this job's normal.
+ *
 * So: adding or removing a [FailsOnEmulatorApi37] means changing this number, in this file, in
 * the same diff. The report says so on the run itself if you forget — it prints the tree's own
 * `grep` count beside this one.
@@ -52,4 +78,4 @@ annotation class FailsOnEmulatorApi37
 * `INSTRUMENTATION_ABORTED`, so the count is a number taken from a partial run. The report
 * records the truncation next to the counts for that reason.
 */
-const val FAILS_ON_EMULATOR_API37_BASELINE = 3
+const val FAILS_ON_EMULATOR_API37_BASELINE = 5
@@ -7,6 +7,10 @@ import androidx.media3.common.MimeTypes
 import androidx.media3.common.util.UnstableApi
 import androidx.test.ext.junit.runners.AndroidJUnit4
 import androidx.test.platform.app.InstrumentationRegistry
+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
@@ -14,6 +18,7 @@ import org.junit.Assert.assertEquals
 import org.junit.Assert.assertFalse
 import org.junit.Assert.assertNull
 import org.junit.Assert.assertTrue
+import org.junit.Assert.fail
 import org.junit.Before
 import org.junit.Test
 import org.junit.runner.RunWith
@@ -262,6 +267,92 @@ class Media3EngineTest {
        }
    }

+    /**
+     * Cancelling a *running* export stops it, completing #224's third engine.
+     *
+     * The two FFmpeg engines were done first (`ad2a75d`, `d293646`); this is
+     * `Media3Engine.transcode`'s `invokeOnCancellation`, which posts `transformer.cancel()` onto the
+     * engine's own `HandlerThread` because `cancel()` has the same single-thread requirement as
+     * `start()`.
+     *
+     * ## Why the assertion is the output file here, and was not for FFmpeg
+     *
+     * The FFmpeg side could not use the file: `invokeOnCancellation` unlinks it, and on POSIX ffmpeg
+     * keeps writing to the unlinked inode, so the path stays gone whether or not the cancel landed.
+     * It asserted the session's return code instead.
+     *
+     * `Media3Engine` deletes nothing — the partial is `ConversionWorker`'s to clean up — so the file
+     * *is* the evidence. An export that was cancelled leaves no moov atom, so `MediaExtractor`
+     * either finds no video track or refuses the file outright with
+     * `IOException: Failed to instantiate extractor` — measured, and both mean interrupted. One
+     * that ran to completion leaves a playable HEVC file, which is the only outcome treated as a
+     * miss. The wait before
+     * reading it is deliberately several times the length of the export, so a *non*-cancelled export
+     * has certainly finished by then: the failure direction is "the file became valid", never "we
+     * did not wait long enough".
+     *
+     * ## Why it retries
+     *
+     * Same reason as the other two, measured there: the committed fixture is 3 s at 320x240 and the
+     * export outruns a naive cancel on a loaded runner. An attempt whose export finished before the
+     * cancel landed has tested nothing, so it is a miss and is retried; only exhausting
+     * [CANCEL_ATTEMPTS] fails. With `transformer.cancel()` removed every attempt produces a playable
+     * file, so the mutation still bites — it just takes five tries to say so.
+     *
+     * Progress having been reported is what proves the export really started, so a miss is
+     * distinguishable from an export that never ran at all — which matters on the API 37 image,
+     * where the decoder is what fails.
+     */
+    @Test
+    @FailsOnEmulatorApi37
+    fun cancellingARunningExportStopsIt(): Unit = runBlocking {
+        val outcomes = mutableListOf<String>()
+
+        repeat(CANCEL_ATTEMPTS) { attempt ->
+            val partial = File(context.cacheDir, "cancelled_export_$attempt.mp4").apply { delete() }
+
+            val job = launch(Dispatchers.IO) {
+                engine.transcode(
+                    input = Uri.fromFile(input),
+                    output = partial,
+                    request = ConversionRequest(OutputFormat.MP4_H265.spec),
+                )
+            }
+
+            // The muxer creating the file is proof the export really started, and it is the
+            // earliest such proof available -- earlier than the first progress tick.
+            withTimeout(TIMEOUT_MS) {
+                while (!partial.exists() && job.isActive) delay(POLL_MS)
+            }
+            val started = partial.exists()
+            job.cancelAndJoin()
+
+            if (!started) {
+                // The export failed before writing anything. That is not a cancellation result
+                // either way, so it is not allowed to pass as one.
+                outcomes += "attempt $attempt never produced an output file to cancel"
+                return@repeat
+            }
+
+            // Several times the export's own length, so a cancel that did not land has certainly
+            // finished. The failure direction is "the file became playable", never "too soon".
+            delay(SETTLE_MS)
+
+            // A cancelled export reports itself two ways and both mean the same thing: no video
+            // track, or MediaExtractor refusing the file outright with "Failed to instantiate
+            // extractor" because there is no moov atom to read. Only a *playable* file is a miss.
+            val video = runCatching { videoMimeTypeOf(partial) }.getOrNull()
+            partial.delete()
+            if (video == null) return@runBlocking
+            outcomes += "attempt $attempt produced a playable $video"
+        }
+
+        fail(
+            "never interrupted a running export in $CANCEL_ATTEMPTS attempts, so either every " +
+                "export finished first or cancellation does not reach the transformer: $outcomes",
+        )
+    }
+
    private fun videoMimeTypeOf(file: File): String? {
        val extractor = MediaExtractor()
        try {
@@ -280,6 +371,19 @@ class Media3EngineTest {
    private companion object {
        const val TIMEOUT_SECONDS = 120L

+        /** Bounds the wait for the muxer to create the file; a hang here is a defect. */
+        const val TIMEOUT_MS = 30_000L
+        const val POLL_MS = 25L
+
+        /**
+         * How long to let a *failed* cancel finish. Several times the export's own length, so
+         * "the file is not playable" cannot mean "not yet".
+         */
+        const val SETTLE_MS = 10_000L
+
+        /** See the KDoc: a miss is the loaded-runner case, not a defect. */
+        const val CANCEL_ATTEMPTS = 5
+
        /**
         * Short on purpose. Nothing is decoded or encoded on this path — the builder refuses the
         * input outright — so anything approaching this is a hang, which is what the test is
@@ -13,6 +13,7 @@ import androidx.work.WorkManager
 import androidx.work.Worker
 import androidx.work.WorkerParameters
 import androidx.work.workDataOf
+import kotlinx.coroutines.CompletableDeferred
 import kotlinx.coroutines.flow.first
 import kotlinx.coroutines.runBlocking
 import kotlinx.coroutines.withTimeout
@@ -27,7 +28,10 @@ import org.junit.runner.RunWith
 import org.libremediaconverter.join.JoinState
 import org.libremediaconverter.join.JoinViewModel
 import org.libremediaconverter.model.ConcatStrategy
+import org.libremediaconverter.model.ConversionRequest
 import org.libremediaconverter.model.Engine
+import org.libremediaconverter.model.OutputFormat
+import org.libremediaconverter.model.QualityTier
 import org.libremediaconverter.work.ConcatWorker
 import org.libremediaconverter.work.ConversionWorker
 import org.libremediaconverter.work.JobTags
@@ -64,6 +68,26 @@ class EchoWorker(context: Context, params: WorkerParameters) : Worker(context, p
 * path, foreground service included — into a synchronous test double, depending on class order.
 */
@UnstableApi
+/**
+ * A [SoftwareTranscoder] that holds the worker in [WorkInfo.State.RUNNING] until released.
+ *
+ * Declared here rather than in `FakeFailures` because it is the only test that needs a job to stay
+ * live on demand, and the shape is specific to that: the others fake a *failure*, this fakes
+ * *duration*.
+ */
+private class BlockingTranscoder(private val released: CompletableDeferred<Unit>) : SoftwareTranscoder {
+    override suspend fun run(
+        request: ConversionRequest,
+        inputPath: String,
+        output: File,
+        durationMs: Long,
+        onProgress: (Int) -> Unit,
+    ) {
+        released.await()
+        output.writeBytes(ByteArray(1_024))
+    }
+}
+
@RunWith(AndroidJUnit4::class)
 class ReattachOnLaunchTest {

@@ -75,7 +99,13 @@ class ReattachOnLaunchTest {
    fun clearTheQueue() = emptyQueueAndStaging()

    @After
-    fun leaveNothingBehind() = emptyQueueAndStaging()
+    fun leaveNothingBehind() {
+        // The suite runs without Android Test Orchestrator, so every class shares one process and
+        // a swapped seam outlives the class that set it. Only one test here swaps one, but a
+        // BlockingTranscoder left in place would hang the next class that converts anything.
+        ConversionDependencies.reset()
+        emptyQueueAndStaging()
+    }

    /**
     * The claim the whole fix rests on, checked against the production request builder rather
@@ -261,6 +291,69 @@ class ReattachOnLaunchTest {
        return request.id
    }

+    /**
+     * Reattaching to a conversion that is **running right now**, which nothing had ever driven.
+     *
+     * This class covers a job that finished, one whose staged file is gone, an ambiguous pair, one
+     * still queued, and one the user cancelled. [Reattachment.rank] gives
+     * [WorkInfo.State.RUNNING] the **highest** rank of all — "live work outranks a finished result
+     * because a running job is holding a foreground service" — and no test on either source set
+     * ever produced one. `ReattachmentTest` exercises the ranking as a pure function over
+     * fabricated snapshots; what was missing is a ViewModel meeting a real running job.
+     *
+     * It is also the likeliest reattachment there is: the user starts a conversion, leaves, and
+     * comes back while it is still going.
+     *
+     * ## Why the engine is a fake here, and why that is not a weakening
+     *
+     * The job has to still be running when the ViewModel is built, and every real conversion in
+     * this suite finishes in about a second — racing that is what made the cancellation tests flaky
+     * enough to need retries (#224). A [SoftwareTranscoder] that blocks until released removes the
+     * race outright: the job is `RUNNING` for exactly as long as the test wants.
+     *
+     * Nothing about reattachment depends on which engine is transcoding. What is under test is the
+     * tag query, [Reattachment.choose] over live WorkManager state, and `observe` mapping it to
+     * [ConversionState.Converting] — all of which run identically whatever is doing the work.
+     *
+     * ## What this does not do, and cannot (#230)
+     *
+     * It does not kill the process. `docs/defect-audit.md` D3/D13 record that `am kill` refuses a
+     * process holding a foreground service, and there is a more basic obstacle: **instrumentation
+     * runs in the app's own process**, so any route that really killed it would take the test
+     * runner with it and there would be nothing left to assert with. A relaunch-and-observe test
+     * needs two instrumentation runs, which the runner does not provide.
+     *
+     * So process death stays device-manual, and this is the closest observable analogue: a fresh
+     * ViewModel, with no memory of the work, meeting a job that is genuinely mid-flight.
+     */
+    @Test
+    fun reattachesToAConversionThatIsStillRunning(): Unit = runBlocking {
+        val released = CompletableDeferred<Unit>()
+        ConversionDependencies.software = { BlockingTranscoder(released) }
+
+        val request = ConversionWorker.request(
+            inputUri = Uri.fromFile(stage("running_input.mp3")),
+            displayName = RUNNING_NAME,
+            sizeBytes = RUNNING_SIZE,
+            spec = OutputFormat.MP3.spec,
+            quality = QualityTier.FAST,
+        )
+        workManager.enqueue(request).result.get()
+
+        // Deterministic: the worker cannot finish until this test lets it.
+        withTimeout(TIMEOUT_MS) {
+            workManager.getWorkInfoByIdFlow(request.id).first { it?.state == WorkInfo.State.RUNNING }
+        }
+
+        val reattached = awaitConversion<ConversionState.Converting>()
+
+        assertEquals(RUNNING_NAME, reattached.input.displayName)
+        assertEquals(RUNNING_SIZE, reattached.input.sizeBytes)
+
+        released.complete(Unit)
+        workManager.cancelWorkById(request.id).result.get()
+    }
+
    /**
     * Enqueues a job that stays [WorkInfo.State.ENQUEUED]. The delay is what holds it there: it
     * is long enough that nothing can run it during a test, and it is cancelled either way.
@@ -326,5 +419,9 @@ class ReattachOnLaunchTest {
         * against WorkManager's database, so this is generous rather than tuned.
         */
        const val SETTLE_MS = 5_000L
+
+        /** Read back off the job's tags by the reattaching ViewModel, so both have to survive. */
+        const val RUNNING_NAME = "still_running.mp3"
+        const val RUNNING_SIZE = 4_242L
    }
 }
@@ -12,11 +12,17 @@ import kotlinx.coroutines.withTimeout
 import org.junit.After
 import org.junit.Assert.assertEquals
 import org.junit.Assert.assertTrue
+import org.junit.Assume.assumeTrue
 import org.junit.Before
 import org.junit.Test
 import org.junit.runner.RunWith
+import org.libremediaconverter.codec.AndroidDeviceCodecs
+import org.libremediaconverter.model.ConversionRequest
+import org.libremediaconverter.model.ConversionRouter
+import org.libremediaconverter.model.Engine
 import org.libremediaconverter.model.OutputFormat
 import org.libremediaconverter.model.QualityTier
+import org.libremediaconverter.model.VideoCodec
 import org.libremediaconverter.work.ConversionWorker
 import java.io.File

@@ -34,6 +40,47 @@ import java.io.File
 * hand — a regression test that silently skips is worse than no test, because the count
 * still reads as coverage.
 *
+ * ## Why this skips on emulators, and why that is the honest answer (#223)
+ *
+ * **This test used to pass everywhere while proving nothing.** Two independent facts stop the
+ * fallback happening on an emulator, and both were measured rather than reasoned:
+ *
+ *  1. **The router never sends the job to Media3.** A Fast MP4/H.265 job goes to the hardware path
+ *     only when `device.canEncode(H265)`, and emulators expose no hardware encoder — every leg of
+ *     run `34004304566` logged
+ *     `Routing sample_h264_444.mp4 -> ... via FFMPEG (NO_HARDWARE_ENCODER)`. The whole test
+ *     finished in 448 ms, which is not long enough to fail an export and then re-encode.
+ *  2. **Forcing it to Media3 does not help either, which is the part that settles it.** Pinning
+ *     `ConversionDependencies.deviceCodecs` to [DeviceCodecs.PERMISSIVE] — the trick
+ *     [ForcedFailureTest] uses — makes the router choose Media3, and the export then *succeeds*.
+ *     Measured on a local API 34 emulator: `MediaCodecInfo` logs
+ *     `NoSupport [codec.profileLevel, avc1.F4000C, video/avc]` for **both**
+ *     `c2.goldfish.h264.decoder` and `c2.android.avc.decoder`, and ExoPlayer allocates the
+ *     goldfish decoder anyway, which decodes the file regardless of the profile it declares.
+ *     `c2.android.hevc.encoder` then encodes the result and the job reports `MEDIA3`.
+ *
+ * So the class KDoc above — "Media3 fails partway through the export on every device" — **is not
+ * true of the emulator images**, and no amount of routing pressure makes this fixture force a
+ * fallback there. The emulator cannot answer this question, so the test says so out loud instead
+ * of passing.
+ *
+ * That is why the gate is [assumeTrue] on the *production* premise (`canEncode(H265)`) rather than
+ * a pinned profile: pinning would also swap in software codecs, which is not the path a real
+ * device takes and is what made the forced run succeed. **This is now the third permanent skip**;
+ * the other two are [org.libremediaconverter.bench.RealMediaBenchmark]'s.
+ *
+ * `ForcedFailureTest.hardwareFailureFallsBackToSoftware` still covers the fallback *wiring* on
+ * every leg, with an `ExplodingHardware` double. What only a device with a real hardware encoder
+ * can show is two real engines disagreeing about a real file, and that is what this is for.
+ *
+ * ## Why the assertion is a pair
+ *
+ * `KEY_ENGINE_USED` is `FFMPEG` whether the fallback fired **or** the router went straight there,
+ * so asserting it alone would not have caught any of the above. The premise is asserted
+ * separately: [ConversionRouter.route] chooses `MEDIA3` for this request on this device. Static
+ * routing wanted hardware, the runtime result was software — together, and only together, that is
+ * the fallback.
+ *
 * The fixture was produced with x264, which the host toolchain cannot do (Fedora's
 * ffmpeg ships openh264, which is Constrained Baseline only):
 *
@@ -66,6 +113,15 @@ class HardwareFallbackTest {

    @Test
    fun aFileMedia3CannotDecodeStillConvertsViaFfmpeg(): Unit = runBlocking {
+        // See "Why this skips on emulators" on the class. Without a real hardware encoder the
+        // router never chooses Media3, and forcing it makes the export succeed instead of fail --
+        // so there is no fallback to observe and a green run would mean nothing.
+        assumeTrue(
+            "no hardware HEVC encoder, so the router cannot choose Media3 and there is no " +
+                "fallback to exercise",
+            AndroidDeviceCodecs.get().canEncode(VideoCodec.H265),
+        )
+
        val request = ConversionWorker.request(
            inputUri = Uri.fromFile(input),
            displayName = SAMPLE,
@@ -75,6 +131,19 @@ class HardwareFallbackTest {
            // the tier where the fallback has to rescue the conversion.
            quality = QualityTier.FAST,
        )
+        // The premise, asserted rather than assumed: this request is one the router wants to send
+        // to hardware on this device. Without it the test is green whether the fallback fired or
+        // the job never went near Media3, which is exactly how #223 stayed invisible.
+        val decision = ConversionRouter.route(
+            ConversionRequest(OutputFormat.MP4_H265.spec, quality = QualityTier.FAST),
+            AndroidDeviceCodecs.get(),
+        )
+        assertEquals(
+            "this test only means something if the router sends this job to Media3",
+            Engine.MEDIA3,
+            decision.engine,
+        )
+
        workManager.enqueue(request).result.get()

        val terminal = withTimeout(TIMEOUT_MS) {
@@ -88,6 +157,14 @@ class HardwareFallbackTest {
            terminal?.state,
        )

+        // The outcome. Paired with the routing assertion above this is the fallback and nothing
+        // else: hardware was chosen, software is what ran.
+        assertEquals(
+            "the router chose Media3, so a successful job must have fallen back to FFmpeg",
+            Engine.FFMPEG.name,
+            terminal?.outputData?.getString(ConversionWorker.KEY_ENGINE_USED),
+        )
+
        val out = File(terminal!!.outputData.getString(ConversionWorker.KEY_OUTPUT_PATH)!!)
        assertTrue("no output produced", out.exists() && out.length() > 0)
        out.delete()
@@ -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
@@ -113,6 +123,11 @@ class FFmpegEngineTest {
    fun encodesFlacLosslessAudio() {
        val out = convert(OutputFormat.FLAC)
        assertTrue("no FLAC produced", out.exists() && out.length() > 0)
+        // "fLaC", the native FLAC stream marker. Without this the test passed on any non-empty
+        // file, so a builder arm emitting the wrong encoder into a .flac name shipped green
+        // (#228) -- the same shape the five assertions above already guard against.
+        val magic = out.inputStream().use { String(it.readNBytes(4), Charsets.US_ASCII) }
+        assertEquals("fLaC", magic)
    }

    @Test
@@ -127,6 +142,165 @@ class FFmpegEngineTest {
    fun encodesOpus() {
        val out = convert(OutputFormat.OPUS)
        assertTrue("no Opus produced", out.exists() && out.length() > 0)
+        // OutputFormat.OPUS is Container.OGG, so the file is an Ogg stream: "OggS" (#228).
+        // Deliberately the container marker rather than the codec -- it is what the other
+        // container-level assertions in this class check, and it is four bytes at offset 0.
+        val magic = out.inputStream().use { String(it.readNBytes(4), Charsets.US_ASCII) }
+        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 --------------------
@@ -166,4 +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
+    }
 }
@@ -5,17 +5,29 @@ 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
 import org.libremediaconverter.convert.MediaProbe
 import org.libremediaconverter.convert.StagingNames
 import org.libremediaconverter.ffmpeg.ConcatEngine
+import org.libremediaconverter.ffmpeg.FFmpegEngine
 import org.libremediaconverter.model.ConcatStrategy
+import org.libremediaconverter.work.ConcatWorker
 import java.io.File

 /**
@@ -50,6 +62,82 @@ 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
@@ -149,6 +237,55 @@ class ConcatEngineTest {
        )
    }

+    /**
+     * A failed join tells the user the return code and what FFmpeg said.
+     *
+     * **This is the device half of #203/#217**, whose PR closed by noting the join legs had not
+     * been run. Running them would not have answered it: nothing on either source set drove a real
+     * join *failure*, so the unified message was asserted only against values a JVM test hands to
+     * `sessionOutcome` directly.
+     *
+     * What is device-only here is that the three reads behind that message work against a real
+     * native session at all — `getReturnCode`, `getFailStackTrace` and `getAllLogsAsString`. If
+     * the log tail came back null or empty on a device, the user would get `Joining failed (1): `
+     * with nothing after the colon and every JVM test would still pass.
+     *
+     * **What this deliberately does not pin is the preference between the two detail sources.** On
+     * an ordinary non-zero return code FFmpegKit reports no fail stack trace, so the stack-trace-
+     * first rule and the log-tail-first rule produce the same text and no assertion here can tell
+     * them apart. That ordering is [SessionOutcomeTest][org.libremediaconverter.ffmpeg.SessionOutcomeTest]'s
+     * job, where both sources can be non-blank at once. Asserting it here would be a test whose
+     * KDoc claims more than it checks — the `probeForConcat` mistake wave 3 caught.
+     *
+     * The failure is forced with an input that does not exist, which the concat demuxer rejects
+     * the same way on every FFmpeg build, rather than with malformed media whose handling varies.
+     */
+    @Test
+    fun aFailedJoinReportsTheReturnCodeAndWhatFFmpegSaid(): Unit = runBlocking {
+        val missing = File(context.cacheDir, "no_such_clip.mp4").also { it.delete() }
+        val out = output("joined_failure.mp4")
+
+        val failure = runCatching {
+            engine.join(listOf(Uri.fromFile(clipA), Uri.fromFile(missing)), out)
+        }.exceptionOrNull()
+
+        assertTrue(
+            "a join over a missing input must fail, got $failure",
+            failure is FFmpegEngine.FFmpegException,
+        )
+        val message = failure?.message.orEmpty()
+        assertTrue(
+            "the message must name the operation and carry the return code, was: '$message'",
+            message.startsWith("Joining failed ("),
+        )
+        // The half a JVM test cannot reach: a real session actually produced detail to show.
+        val detail = message.substringAfter("): ", "")
+        assertTrue(
+            "the message stopped at the return code and told the user nothing, was: '$message'",
+            detail.isNotBlank(),
+        )
+    }
+
    @Test
    fun theListFileIsCleanedUpAfterJoining(): Unit = runBlocking {
        val out = output("joined_cleanup.mp4")
@@ -180,4 +317,13 @@ 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
+    }
 }
@@ -0,0 +1,123 @@
+package org.libremediaconverter.saf
+
+import androidx.media3.common.util.UnstableApi
+import androidx.test.ext.junit.runners.AndroidJUnit4
+import androidx.test.platform.app.InstrumentationRegistry
+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.assertTrue
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.ffmpeg.ConcatEngine
+import org.libremediaconverter.model.Engine
+import org.libremediaconverter.model.OutputFormat
+import org.libremediaconverter.model.QualityTier
+import org.libremediaconverter.work.ConversionWorker
+import java.io.File
+
+/**
+ * A `content://` input reaching FFmpeg successfully, which nothing had ever driven (#225).
+ *
+ * `FFmpegKitConfig.getSafParameterForRead` stands between a SAF grant and the native process, and
+ * it is on **every real user conversion**. Every passing convert and join test in this suite hands
+ * the worker a `Uri.fromFile(...)`, which takes the `uri.path` arm instead — so the bridge was
+ * exercised only on its failure side, by `UnopenableUriTest` naming an authority that does not
+ * exist. That proves the error message, not the bridge.
+ *
+ * ## Why a plain provider rather than the documents one
+ *
+ * [FixtureDocumentsProvider] cannot be reached from the app, measured three ways on an API 34
+ * emulator (#226): a `DOCUMENTS_PROVIDER` declared without `MANAGE_DOCUMENTS` is refused at install
+ * — *"Provider must be protected by MANAGE_DOCUMENTS"*; instrumentation runs in the **target app's
+ * process**, so `Instrumentation.getContext()` still carries the app's uid and is denied; and
+ * `adoptShellPermissionIdentity(MANAGE_DOCUMENTS)` is denied identically. The denial names the only
+ * way in: *"you obtain access using ACTION_OPEN_DOCUMENT or related APIs"*.
+ *
+ * The bridge does not need one. It opens a descriptor through the resolver and hands FFmpeg a
+ * `saf:` path, so any readable `content://` URI exercises it — and [FixtureContentProvider] is an
+ * ordinary provider, which may be exported without a permission. The whole class is headless: no
+ * DocumentsUI, and none of the flake #190 records.
+ *
+ * ## Why MP3
+ *
+ * The bridge lives on the FFmpeg arm, and MP3 is the format the router sends there unconditionally
+ * — no platform encoder exists at any API level, so `ConversionWorkerTest.routesAnMp3JobToFfmpeg…`
+ * relies on the same fact. Choosing a video target would make the engine depend on the device's
+ * codecs, and #223 is what that costs.
+ *
+ * *Mutation:* make `getSafParameterForRead` return `uri.toString()`. FFmpeg cannot open it and both
+ * tests fail; nothing else in either suite notices.
+ */
+@UnstableApi
+@RunWith(AndroidJUnit4::class)
+class ContentUriInputTest {
+
+    private val context = InstrumentationRegistry.getInstrumentation().targetContext
+    private val workManager = WorkManager.getInstance(context)
+
+    @After
+    fun tearDown() {
+        File(context.cacheDir, "conversions").listFiles()?.forEach { it.delete() }
+    }
+
+    @Test
+    fun aContentUriInputConvertsThroughTheSafBridge(): Unit = runBlocking {
+        val input = FixtureContentProvider.uriFor(SAMPLE)
+        val request = ConversionWorker.request(
+            inputUri = input,
+            displayName = SAMPLE,
+            sizeBytes = 0L,
+            spec = OutputFormat.MP3.spec,
+            quality = QualityTier.FAST,
+        )
+        workManager.enqueue(request).result.get()
+
+        val terminal = withTimeout(TIMEOUT_MS) {
+            workManager.getWorkInfoByIdFlow(request.id).first { it != null && it.state.isFinished }
+        }
+
+        val error = terminal?.outputData?.getString(ConversionWorker.KEY_ERROR)
+        assertEquals(
+            "a content:// input must convert, but failed with: $error",
+            WorkInfo.State.SUCCEEDED,
+            terminal?.state,
+        )
+        // The bridge is on the FFmpeg arm only, so this is part of the claim rather than colour.
+        assertEquals(Engine.FFMPEG.name, terminal?.outputData?.getString(ConversionWorker.KEY_ENGINE_USED))
+
+        val out = File(terminal!!.outputData.getString(ConversionWorker.KEY_OUTPUT_PATH)!!)
+        assertTrue("no output produced from a content:// input", out.exists() && out.length() > 0)
+        out.delete()
+    }
+
+    /**
+     * The same bridge on the join path, which has its own copy of the call (`ConcatEngine:36`).
+     *
+     * Driven through the engine rather than `ConcatWorker` because the engine is where the branch
+     * is; the worker adds a foreground service and nothing else this is about.
+     */
+    @Test
+    fun contentUriInputsJoinThroughTheSafBridge(): Unit = runBlocking {
+        val out = File(context.cacheDir, "joined_from_content.mp4").apply { delete() }
+        val result = ConcatEngine(context).join(
+            listOf(FixtureContentProvider.uriFor(CLIP_A), FixtureContentProvider.uriFor(CLIP_B)),
+            out,
+            OutputFormat.MP4_H264,
+        )
+
+        assertTrue("no output produced from content:// inputs", result.output.length() > 0)
+        out.delete()
+    }
+
+    private companion object {
+        const val SAMPLE = "sample_h264.mp4"
+        const val CLIP_A = "clip_a.mp4"
+        const val CLIP_B = "clip_b.mp4"
+        const val TIMEOUT_MS = 300_000L
+    }
+}
@@ -0,0 +1,135 @@
+package org.libremediaconverter.saf;
+
+import android.content.ContentProvider;
+import android.content.ContentValues;
+import android.database.Cursor;
+import android.database.MatrixCursor;
+import android.net.Uri;
+import android.os.ParcelFileDescriptor;
+import android.provider.OpenableColumns;
+
+import java.io.File;
+import java.io.FileNotFoundException;
+import java.io.FileOutputStream;
+import java.io.IOException;
+import java.io.InputStream;
+import java.io.OutputStream;
+
+/**
+ * A plain {@link ContentProvider} serving the committed media fixtures over {@code content://}.
+ *
+ * <p><b>Why this exists alongside {@link FixtureDocumentsProvider}.</b> Every passing convert and
+ * join test hands the worker a {@code Uri.fromFile(...)}, which takes the {@code uri.path} arm and
+ * never touches {@code FFmpegKitConfig.getSafParameterForRead}. That bridge is on 100% of real user
+ * conversions and was on 0% of tested ones; only its failure side was covered, by
+ * {@code UnopenableUriTest} pointing at an authority that does not exist.
+ *
+ * <p><b>Why not the documents provider.</b> It cannot be reached. Measured three ways on an API 34
+ * emulator: a {@code DOCUMENTS_PROVIDER} declared without {@code MANAGE_DOCUMENTS} is refused at
+ * install ("Provider must be protected by MANAGE_DOCUMENTS"); instrumentation runs in the target
+ * app's process, so {@code Instrumentation.getContext()} still carries the app's uid and is denied;
+ * and {@code adoptShellPermissionIdentity(MANAGE_DOCUMENTS)} is denied identically. The denial says
+ * what is required — <i>"you obtain access using ACTION_OPEN_DOCUMENT or related APIs"</i> — so a
+ * documents provider is reachable only through a picker-issued grant. See issue #226.
+ *
+ * <p>The bridge does not need one. {@code getSafParameterForRead} opens a file descriptor through
+ * the resolver and hands FFmpeg a {@code saf:} path; any readable {@code content://} URI exercises
+ * it. An ordinary provider may be exported without a permission, so this one is, and the whole test
+ * stays headless — no DocumentsUI, and none of the flake #190 records.
+ *
+ * <p>Unlike {@link FixtureDocumentsProvider} this may use {@code androidx} and Kotlin freely — it is
+ * loaded into the app process like any other provider, not into the bare test process. It is kept
+ * in Java anyway, next to its sibling, so the two read alike.
+ */
+public final class FixtureContentProvider extends ContentProvider {
+
+    /** Authority. Distinct from the documents provider's, and from anything the app declares. */
+    public static final String AUTHORITY = "org.libremediaconverter.test.content";
+
+    /** Builds a URI for one of this source set's committed assets, e.g. {@code sample_h264.mp4}. */
+    public static Uri uriFor(String assetName) {
+        return new Uri.Builder().scheme("content").authority(AUTHORITY).appendPath(assetName).build();
+    }
+
+    @Override
+    public boolean onCreate() {
+        return true;
+    }
+
+    @Override
+    public ParcelFileDescriptor openFile(Uri uri, String mode) throws FileNotFoundException {
+        if (!"r".equals(mode)) {
+            throw new FileNotFoundException("this provider is read-only: " + mode);
+        }
+        return ParcelFileDescriptor.open(unpack(assetOf(uri)), ParcelFileDescriptor.MODE_READ_ONLY);
+    }
+
+    /**
+     * Enough of {@link OpenableColumns} for {@code InputQuery.describe} to name and size the input.
+     *
+     * <p>Without these the app reaches the "Size unknown" screen, which is a different test.
+     */
+    @Override
+    public Cursor query(Uri uri, String[] projection, String selection, String[] args, String sort) {
+        String asset = assetOf(uri);
+        File file;
+        try {
+            file = unpack(asset);
+        } catch (FileNotFoundException e) {
+            return null;
+        }
+        MatrixCursor cursor = new MatrixCursor(
+            new String[] {OpenableColumns.DISPLAY_NAME, OpenableColumns.SIZE});
+        cursor.newRow().add(OpenableColumns.DISPLAY_NAME, asset).add(OpenableColumns.SIZE, file.length());
+        return cursor;
+    }
+
+    @Override
+    public String getType(Uri uri) {
+        return assetOf(uri).endsWith(".m4a") ? "audio/mp4" : "video/mp4";
+    }
+
+    @Override
+    public Uri insert(Uri uri, ContentValues values) {
+        throw new UnsupportedOperationException("read-only fixture provider");
+    }
+
+    @Override
+    public int delete(Uri uri, String selection, String[] args) {
+        throw new UnsupportedOperationException("read-only fixture provider");
+    }
+
+    @Override
+    public int update(Uri uri, ContentValues values, String selection, String[] args) {
+        throw new UnsupportedOperationException("read-only fixture provider");
+    }
+
+    private static String assetOf(Uri uri) {
+        String asset = uri.getLastPathSegment();
+        return asset == null ? "" : asset;
+    }
+
+    /**
+     * The asset on disk, unpacked the first time anything asks.
+     *
+     * <p>Reported as {@link FileNotFoundException} rather than swallowed: a provider answering with
+     * a zero-byte file would fail the conversion for a reason nothing states.
+     */
+    private File unpack(String asset) throws FileNotFoundException {
+        File file = new File(getContext().getCacheDir(), "provided_" + asset);
+        if (file.length() > 0L) {
+            return file;
+        }
+        try (InputStream source = getContext().getAssets().open(asset);
+                OutputStream sink = new FileOutputStream(file)) {
+            byte[] buffer = new byte[8192];
+            int read;
+            while ((read = source.read(buffer)) != -1) {
+                sink.write(buffer, 0, read);
+            }
+        } catch (IOException e) {
+            throw new FileNotFoundException("could not unpack " + asset + ": " + e);
+        }
+        return file;
+    }
+}
@@ -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.
@@ -203,6 +207,8 @@ import org.libremediaconverter.ui.TestTags
 * driven there at all. That is why this gap survived as long as it did.
 * `tools/local-emulator/run-e2e.sh` runs API 33-36 on the development host, and both tests pass
 * there: **59 / 0 / 0 / 2 at API 33 and again at API 36**, whole suite, 2026-08-24.
+ * (Since #223 the skip column reads 3 on an emulator — `HardwareFallbackTest` now announces
+ * that it cannot run without a hardware HEVC encoder rather than passing vacuously.)
 *
 * ### Why only the rotation test carries [FailsOnEmulatorApi37]
 *
@@ -251,6 +257,21 @@ 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.
     *
@@ -270,13 +291,42 @@ class SafPickerRoundTripTest {
     */
    @After
    fun restoreOrientation() {
+        ActivityLifecycleMonitorRegistry.getInstance().removeLifecycleCallback(recreationWatcher)
        if (!rotated) return
        device.setOrientationNatural()
        device.unfreezeRotation()
        device.waitForIdle()
    }

+    /**
+     * **Marked for API 37 because of what it does to the image, not because it fails there.**
+     *
+     * This is the one place the marker's KDoc phrase "cannot pass on this image" does not fit, and
+     * the distinction is worth keeping rather than smoothing over. Across the four gating API 37
+     * runs whose logcats were read on 2026-09-05 — 34006456986, 34001744574, 34001377499 and the
+     * green 34002313300 — the leg carries exactly two `hasReadColorBufferDma` aborts before the
+     * suite starts (both `surfaceflinger`, during boot and the SystemUI disable) and then exactly
+     * **one** during it. Every time, that one is `system_server` on the `TaskSnapshotPer` thread,
+     * and every time it lands inside this test's window. No other test in the gating set reaches
+     * the mapper at all.
+     *
+     * So this test kills the framework on that image whether it passes or not, and whether the leg
+     * goes red is luck: 34001377499 passed it and lost the leg anyway (`failed: 0`, teardown
+     * broken), 34002313300 passed it 0.6 s after the abort and went green. That is #108, and it is
+     * why the leg was failing on unrelated PRs.
+     *
+     * `docs/api-37-emulator-crash.md` measured this test on 2026-08-24, recorded "passes, 4 aborts
+     * in the window", and concluded that a rotation reaches the mapper where starting DocumentsUI
+     * does not. The aborts were seen; what was not drawn out is that they are this test's own and
+     * are not intermittent.
+     *
+     * The marker is what routes it off the gating leg and into the advisory job beside its
+     * rotation sibling. **It is not a statement about the picker**: the same test passes on API
+     * 33–36 on the same runner and on the Pixel 10 Pro XL, which is where API 37's answer comes
+     * from.
+     */
    @Test
+    @FailsOnEmulatorApi37
    fun pickingAFileThroughTheSystemPickerFillsInTheFileCard() {
        pickTheFixture()

@@ -303,9 +353,11 @@ 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
@@ -580,6 +632,9 @@ class SafPickerRoundTripTest {
     * 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.
+     *
+     * **[forceStopThePicker] is the escalation after the presses, and it exists because a back
+     * press is not always deliverable.** See its own KDoc for the measurement.
     */
    private fun dismissThePicker() {
        repeat(BACK_PRESSES) {
@@ -594,15 +649,50 @@ class SafPickerRoundTripTest {
        // 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()) return
+        forceStopThePicker()
        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(),
+                "the system picker would not close: after $BACK_PRESSES back presses and a " +
+                    "force-stop of $DOCUMENTS_UI_PACKAGE the app still does not have the window " +
+                    "focus, and ${device.currentPackageName} is in front. What could be seen: " +
+                    describeWindows(),
            )
        }
    }

+    /**
+     * Kills the picker's process, for when no back press can reach it.
+     *
+     * **The failure this exists for cannot be answered with input, and that is the whole point.**
+     * Measured on the gating API 37 legs of runs 34006456986 and 34001744574, which fail this way
+     * and whose logcats say the same thing in the same order. `UiObject2.click()` on the fixture's
+     * root is injected at the node's centre and the framework discards it —
+     * `InputDispatcher: No new touched window at (539.0, 525.0) in display 0` — because
+     * `PickActivity` has published accessibility nodes but has no touchable window there yet.
+     * `click()` cannot see that and returns normally, so the walk goes on to wait out
+     * [PICKER_TIMEOUT_MS] for a fixture that was never navigated to. By the time this function's
+     * caller starts pressing back, WindowManager is still saying
+     * `no window has focus but ...PickActivity may eventually add a window when it finishes
+     * starting up` — and goes on saying it for another 63 s. Every one of the four presses is
+     * dropped, and DocumentsUI ANRs on `Input dispatching timed out`.
+     *
+     * So the picker is in front, unreachable by key or by touch, and [pickTheFixture]'s whole
+     * point — that a second `PickActivity` rebuilds every window and list in it — is unreachable
+     * with it. `am force-stop` goes around input entirely: `UiAutomation` runs shell commands as
+     * uid 2000, which holds `FORCE_STOP_PACKAGES`, so the picker's process is killed, its
+     * activity leaves the task it was launched into, and `MainActivity` — the activity below it in
+     * that same task — is resumed with the focus.
+     *
+     * **Only on the failure path**, after every back press has been spent, so a picker that closes
+     * the ordinary way never reaches this and is not altered by it. If the framework itself is
+     * gone, this cannot help either, and the caller still reports what it could see.
+     */
+    private fun forceStopThePicker() {
+        device.executeShellCommand("am force-stop $DOCUMENTS_UI_PACKAGE")
+        device.waitForIdle()
+    }
+
    /** 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) {
@@ -675,6 +765,36 @@ 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 +823,15 @@ 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.
         *
@@ -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
+    }
+}
@@ -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() }
    }
 }
@@ -35,6 +35,21 @@ object FFmpegConcatCommand {
            add("concat")
            add("-safe")
            add("0")
+            // And -protocol_whitelist permits the *scheme* those paths carry, which is a
+            // separate gate (#238). Every input the user actually picks is a content:// URI --
+            // JoinScreen uses OpenMultipleDocuments -- so ConcatEngine maps it through
+            // FFmpegKitConfig.getSafParameterForRead and writes an `ffkitsaf:` path into the
+            // list file. The concat demuxer applies its own whitelist, defaulting to
+            // "file,crypto,data", and refused every one of them:
+            //
+            //   [ffkitsaf @ ...] Protocol 'ffkitsaf' not on whitelist 'file,crypto,data'!
+            //
+            // This only widens that default. It is on the stream-copy branch alone because it
+            // is the only one that feeds the demuxer a list file -- REENCODE passes each input
+            // with its own -i, where the whitelist does not apply, which is why joining over SAF
+            // worked for mismatched clips and failed for matching ones.
+            add("-protocol_whitelist")
+            add(PROTOCOL_WHITELIST)
            add("-i")
            add(listFile.absolutePath)
            add("-c")
@@ -84,4 +99,12 @@ object FFmpegConcatCommand {
            add(output.absolutePath)
        }
    }
+
+    /**
+     * The concat demuxer's protocol whitelist: FFmpeg's own default, plus ffmpeg-kit's SAF scheme.
+     *
+     * Spelled out rather than appended to an unknown default, because the default is FFmpeg's and
+     * could change under us; naming all four keeps the command self-describing. See #238.
+     */
+    private const val PROTOCOL_WHITELIST = "file,crypto,data,ffkitsaf"
 }
@@ -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
    }
 }
@@ -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)
+}
@@ -6,6 +6,7 @@ import android.net.Uri
 import androidx.activity.ComponentActivity
 import androidx.compose.ui.test.assertIsDisplayed
 import androidx.compose.ui.test.junit4.v2.createAndroidComposeRule
+import androidx.compose.ui.test.onAllNodesWithTag
 import androidx.compose.ui.test.onNodeWithTag
 import androidx.compose.ui.test.performClick
 import androidx.media3.common.util.UnstableApi
@@ -77,6 +78,28 @@ class LauncherWiringTest {
     * The transposition guard. A picked file has to reach `onInputPicked`, which is observable as
     * the screen arriving at `Ready` with the file card showing — `save()` from `Idle` returns at
     * its own guard and leaves nothing behind.
+     *
+     * ## Why this waits rather than asserting straight away (#220)
+     *
+     * `onInputPicked` does not reach `Ready` on the calling thread. It hops twice —
+     * `withContext(pickDispatcher) { InputQuery.describe(...) }` and then the probe — and
+     * `pickDispatcher` defaults to `Dispatchers.IO`, a real background thread that Compose's
+     * idling does not know about. `deliver` therefore returns with the state still `Idle` more
+     * often than not, and asserting immediately was a race the test usually won.
+     *
+     * It lost five times on CI in one day, on PRs whose diffs were instrumented tests and
+     * documentation, which is what #220 was filed for. `waitUntil` polls through
+     * `waitForIdle`, so it drains the main looper each time round and sees the recomposition that
+     * the IO hop eventually posts back.
+     *
+     * **Injecting the dispatcher would be better and is not available here.** `pickDispatcher` is
+     * a constructor parameter precisely so a test can pin it, but this test composes the real
+     * `ConverterScreen`, which resolves its own ViewModel through `viewModel()` — the seam exists
+     * one layer below the thing under test. Pinning it would mean not testing the launcher edge,
+     * which is the whole point of this class.
+     *
+     * The wait does not weaken the assertion: transposing the two callbacks leaves the screen in
+     * `Idle` forever, so it fails on the timeout with the same meaning it failed with before.
     */
    @Test
    fun `a picked document is loaded as input rather than saved to`() {
@@ -85,6 +108,11 @@ class LauncherWiringTest {
        composeRule.onNodeWithTag(TestTags.Converter.CHOOSE_FILE).performClick()
        deliver(Uri.parse("content://test/holiday.mkv"))

+        composeRule.waitUntil(PICK_TIMEOUT_MS) {
+            composeRule.onAllNodesWithTag(TestTags.Converter.FILE_CARD_NAME)
+                .fetchSemanticsNodes()
+                .isNotEmpty()
+        }
        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD_NAME).assertIsDisplayed()
    }

@@ -138,4 +166,13 @@ class LauncherWiringTest {
        )
        composeRule.waitForIdle()
    }
+
+    private companion object {
+        /**
+         * Long enough that a slow CI runner is not the reason this fails, short enough that a
+         * genuinely transposed callback does not stall the suite. The pick normally lands in
+         * single-digit milliseconds.
+         */
+        const val PICK_TIMEOUT_MS = 10_000L
+    }
 }
@@ -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")
@@ -56,6 +56,33 @@ class FFmpegConcatCommandTest {
        assertEquals("0", args[args.indexOf("-safe") + 1])
    }

+    /**
+     * The gate that `-safe 0` does not open, and the one every real join needs (#238).
+     *
+     * `-safe 0` permits absolute *paths*; the concat demuxer separately whitelists the *protocol*,
+     * defaulting to `file,crypto,data`. `JoinScreen` picks with `OpenMultipleDocuments`, so real
+     * inputs are `content://` and `ConcatEngine` writes `ffkitsaf:` paths into the list file — which
+     * the demuxer refused outright, failing every stream-copy join a user could actually start.
+     *
+     * The re-encode strategy has no equivalent assertion because it needs none: it passes each
+     * input with its own `-i` and never feeds the demuxer a list file. That asymmetry is exactly
+     * why the defect survived — joining mismatched clips over SAF worked.
+     */
+    @Test
+    fun `stream copy whitelists the protocol its list file entries actually use`() {
+        val args = FFmpegConcatCommand.build(
+            ConcatStrategy.STREAM_COPY,
+            inputs,
+            listFile,
+            output,
+            OutputFormat.MP4_H264,
+        )
+        val whitelist = args[args.indexOf("-protocol_whitelist") + 1].split(",")
+        assertTrue("ffmpeg-kit's SAF scheme must be permitted, got $whitelist", "ffkitsaf" in whitelist)
+        // The defaults have to survive too: the list file itself is opened over `file`.
+        assertTrue("the demuxer still reads the list file itself, got $whitelist", "file" in whitelist)
+    }
+
    @Test
    fun `re-encode passes every input separately and builds a filter graph`() {
        val args = FFmpegConcatCommand.build(
@@ -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
@@ -315,25 +315,90 @@ clean zero. Its own post-disable check on the run recorded below printed

 So what is reliably achieved is a **rate collapse** — from roughly one abort every fourteen
 seconds to one every forty-five — which a 47-second Gradle run survives and a five-minute one
-might not. The 180-second zero above is one measurement on a device that had been up for twelve
-minutes and had already cycled its framework several times. The harness prints the quiet-check
-delta on every run precisely so this is visible rather than assumed.
+might not.

-One ordering detail cost a whole run and is now encoded in `disable_region_sampling`: by the time
-`sys.boot_completed` flips, SystemUI has **already registered**, and `pm disable-user` does not
-retract an existing registration — it only stops the package being started again. Disabling it
-and proceeding straight to the tests fails exactly as before. The harness therefore does
-`stop; start` afterwards, so the framework that comes back never starts SystemUI at all.
+**And that restart has never happened — which is how the disable turned out not to work either.**
+Corrected 2026-09-05; this replaces the two paragraphs above rather than qualifying them.
+
+`adb shell stop` and `start` are root-only, adbd is not root on a booted emulator, and all three
+copies of this logic called them without `adb root`. On CI both printed `Must be root`, between
+lines that read as if the restart had happened; `run-e2e.sh` sent them to `/dev/null`, so its
+`Must be root` was never even visible. Neither number in those logs was an observation either —
+the `pidof` loop breaks when the process is gone and otherwise falls out at its last iteration,
+and the old code printed the iteration count either way, so `system_server down after ~40 s` is
+what a stop that did nothing looks like.
+
+Adding `adb root` made the restart real, and **that is what proved the disable ineffective**.
+`api37-debug` run 34010167885, `disable_system_ui=true`:
+
+```
+--- disable round 1 ---
+  pm attempt 1: Package com.android.systemui new state: disabled-user
+  restarting the framework
+  adbd is running as root
+  system_server down after 2 s
+  services back after 10 s
+  NOT DISABLED after the restart -- the package state did not survive
+```
+
+Three rounds of that, then `final state: SystemUI STILL ENABLED`, and the leg reported
+`expected: 0, received: 0` — `Starting 0 tests`, the exact failure this function exists to
+prevent.
+
+Bisected locally on `android-37.0`, which explains the lost state and nothing else:
+
+| arm | sequence | disabled after the restart? |
+|---|---|---|
+| A | `pm disable-user`, then `stop` at once | **no** |
+| B | `pm disable-user`, wait 15 s, then `stop` | **yes** |
+
+That is PackageManager's delayed write of package restrictions: the stop kills `system_server`
+before the settings are flushed, and arm A is what CI did. **Arm B does not help either**, which
+is the measurement that matters. With the package verified `disabled-user` before *and* after a
+further clean restart:
+
+```
+  package still disabled? YES
+  processes:
+   9275  00:17 system_server
+   9695  00:14 com.android.systemui     <- started 3 s after system_server
+```
+
+CI's own logcat says the same without any restart at all. In the gating leg of run 34006456986,
+`pm disable-user` is accepted at 02:28:37.9 and the package really is in `pm list packages -d` at
+02:29:33 — and SystemUI is started at 02:28:39.5 and again at 02:28:52.3, the second of which
+(pid 4275) is alive for the whole instrumentation run.
+
+**So `pm disable-user --user 0 com.android.systemui` does not stop SystemUI starting on this
+image**, with or without a framework restart, on CI or locally. The premise this section was
+built on — "the framework that comes back never starts SystemUI at all" — is false.
+
+Two things follow, pointing in opposite directions.
+
+- **The restart is removed rather than repaired**, in all three copies. It cost a leg every test
+  it had and there is nothing for it to buy. What is kept is the 45-second window with zero new
+  aborts, which was always the part doing the work: in that same run the boot aborts land at
+  02:28:18 and 02:28:43, and the wait is what puts instrumentation at 02:32:42 — after them
+  rather than inside one. The `pm disable-user` call is kept too, for a narrower reason than it
+  was written for: every green leg and every number quoted about this row was measured with it
+  applied, and changing the configuration while fixing a flake is not a trade worth making.
+- **The rate collapse recorded above is not evidence of what it says.** Both arms of that
+  comparison had SystemUI running. What it measured is a device twelve minutes into its uptime
+  against one that had just booted — a real difference, and a different claim. The quiet gate is
+  still worth having on exactly that reading.

 ### The two deviations, stated plainly

 1. **The renderer is ANGLE, not the host GPU.** Shared with nothing else in the matrix — API
   33–36 run `-gpu host` locally, and CI runs `swiftshader_indirect`.
-2. **SystemUI is disabled.** The API 37 leg does not run the same device configuration as any
-   other leg or as the Pixel. It was defensible here because nothing in this suite touched
-   system UI — Media3, FFmpeg and WorkManager tests — and because the alternative is no local
-   API 37 coverage at all. **Anything that ever does depend on system UI must not trust this
-   leg.** Something now does; see the section below.
+2. **SystemUI is asked to be disabled, and runs anyway.** This was written as the deviation that
+   mattered — "anything that ever does depend on system UI must not trust this leg" — and the
+   measurements above say the deviation does not exist: the package is marked `disabled-user` and
+   `com.android.systemui` is up for the whole leg regardless. **The correction is good news
+   rather than bad.** This row is *more* comparable to API 33–36 and to the Pixel than it has
+   been claiming, not less, and the test that depends on system UI (see the section below) was
+   never running in the exotic configuration this bullet describes. What `pm disable-user` leaves
+   behind is a package-manager flag nothing acts on.

 ### Something does depend on system UI now, and half of it is excluded

@@ -341,12 +406,13 @@ Added 2026-08-24, and the first entry on this page that is not a codec.

 `SafPickerRoundTripTest` drives the real system file picker and rotates the display. Both reach
 the gralloc mapper — DocumentsUI is another app's windows, and a rotation rebuilds every surface
-on screen — and **disabling SystemUI does not help**, because it removes the *idle* trigger
-(RegionSamplingThread's nav-bar luma sampling) and not this one.
+on screen — and **disabling SystemUI does not help**. Two reasons now, and only the first was
+known when this was written: it removes the *idle* trigger (RegionSamplingThread's nav-bar luma
+sampling) and not this one, and — see the section above — it does not remove SystemUI either.

-Measured one method per fresh emulator, `android-37.0`, `swangle_indirect`, SystemUI disabled and
-verified quiet — separately, because inferring the second from the first is the mistake this
-page's opening correction is about:
+Measured one method per fresh emulator, `android-37.0`, `swangle_indirect`, with the disable
+applied and verified quiet — separately, because inferring the second from the first is the
+mistake this page's opening correction is about:

 | test | result on android-37.0 | `hasReadColorBufferDma` aborts in the window |
 |---|---|---|
@@ -357,6 +423,111 @@ So a rotation, which rebuilds every surface at once, is what the mapper does not
 starting DocumentsUI is not. Only the rotation test carries `@FailsOnEmulatorApi37`; the picker
 test runs on the gating leg like anything else.

+#### That last sentence was wrong for twelve days, and the aborts in the table said so
+
+**Corrected 2026-09-05.** Read the second row again: the picker test passes *and takes four
+`hasReadColorBufferDma` aborts with it*. This section counted them, put them in the table, and then
+drew the conclusion from the pass/fail column alone. The right question is not "does the test
+pass" but "does the image survive it", and the answer had been printed in the right-hand column
+from the day it was written.
+
+Four gating API 37 runs read logcat-first — 34006456986, 34001744574, 34001377499, and the **green**
+34002313300 — say it without ambiguity. Each carries exactly two aborts before the suite starts
+(both `surfaceflinger`, during boot and the SystemUI disable) and then exactly **one** during it:
+
+| run | picker test window | the run's only in-suite abort | leg |
+|---|---|---|---|
+| 34006456986 | 02:33:04.2 → 02:34:46.9, **failed** | 02:34:46.845 | red, `failed: 1` |
+| 34001744574 | 00:55:41.4 → 00:57:23.9, **failed** | 00:57:23.794 | red, `failed: 1` |
+| 34001377499 | 00:35:53.3 → 00:36:00.6, passed | 00:35:59.662 | red, `failed: 0` |
+| 34002313300 | 00:58:12.7 → 00:58:19.8, passed | 00:58:19.218 | green |
+
+Every one is `system_server`, thread `TaskSnapshotPer`, and every one lands inside that test's
+window. Nothing else in the gating set reached the mapper at all. So the picker test is
+**deterministic** in what it does to the image and a coin flip in what the leg reports: 34001377499
+passed it and lost the leg from teardown with no failing test to name, and 34002313300 passed it
+0.6 s after the abort and went green.
+
+That is #108, which had been filed against this behaviour in August and left open because the
+trigger was unknown. The trigger is this test. It now carries `@FailsOnEmulatorApi37` too, and the
+marker's KDoc had to widen from "does not pass on this image" to "cannot be run on this image" to
+say so honestly.
+
+The stack, for the record, is a different caller from either of the two above:
+
+```
+Cmdline: system_server        name: TaskSnapshotPer
+Abort message: 'Assertion failed: !rcEnc->featureInfo()->hasReadColorBufferDma'
+
+  #04  mapper.ranchu.so   GoldfishMapper::readFromHost(cb_handle_t const&) const+543
+  #06  libui.so           android::Gralloc5Mapper::lock(...)+63
+  #10  libandroid_runtime.so  android::lockImageFromBuffer(...)+374
+  #15  framework.jar      android.media.ImageReader$SurfaceImage.getPlanes+50
+  #17  services.jar       com.android.server.wm.TaskSnapshotConvertUtil.copyToSwBitmapDirect+56
+  #28  services.jar       com.android.server.wm.SnapshotPersistQueue$StoreWriteQueueItem.writeBuffer+66
+  #32  services.jar       com.android.server.wm.SnapshotPersistQueue$1.run+186
+```
+
+WindowManager writing a task snapshot to disk, which needs the buffer as a software bitmap, which
+is the non-DMA readback path. `PickActivity` is started **into the app's own task** (`Task #11
+A=10234:org.libremediaconverter` in the logcat), so the snapshot being persisted is that task's,
+and the churn at the end of the pick is what schedules it.
+
+#### There is no shell knob for task snapshots, and that was checked rather than assumed
+
+#108 asks whether `TaskSnapshotPersister` is suppressible the way the region-sampling listener was.
+Probed on a local `android-37.0 google_apis x86_64` AVD, 2026-09-05:
+
+```
+getprop | grep -i snapshot                          # nothing but apexd-snapshotde
+settings list global | grep -iE 'snapshot|recents'  # empty
+device_config list window_manager | grep -i snapshot # empty
+cmd window help                                     # no snapshot or screenshot command
+dumpsys window | grep -i snapshot                   # mSnapshotEnabled=true, for Task and Activity
+```
+
+`mSnapshotEnabled` is real state and there is nothing that sets it from outside. The only
+`device_config` hits anywhere in the tree are aconfig flags — e.g.
+`windowing_frontend/com.android.window.flags.respect_requested_task_snapshot_resolution` — which
+tune the snapshot rather than disable it. So the marker is the available answer, not the lazy one.
+
+#### When the picker test does fail, the abort is the coda and not the cause
+
+Worth separating, because the failure message points the wrong way. In both runs where the test
+itself went red, it had been broken for 98 seconds before the abort landed. The discriminator is
+one line, present in both reds and absent from the green:
+
+```
+I/InputDispatcher: No new touched window at (539.0, 525.0) in display 0
+```
+
+(539, 525) is the centre of the fixture's root row — the same coordinates the green run clicks.
+The touch reaches no window and is discarded; `UiObject2.click()` cannot see that and returns
+normally. DocumentsUI then logs nothing at all, where the green run logs `DocumentStack` and
+`Creating new directory loader` 40 ms after its click. The walk waits out its timeout twice for a
+fixture it never navigated to, and by the time the back presses start, WindowManager is still
+saying `no window has focus but ...PickActivity may eventually add a window when it finishes
+starting up` — for another 63 s. All four presses are dropped, DocumentsUI ANRs on
+`Input dispatching timed out`, and only *then* does the abort fire and make the failure message
+read `no windows at all`.
+
+`SafPickerRoundTripTest.forceStopThePicker` is the answer to that half: `am force-stop` goes around
+input entirely, so the picker's process can be removed from a task no key press can reach and
+`pickTheFixture`'s whole-picker retry — which exists for exactly this — becomes reachable again.
+That is a fix to the test on every level, not to API 37.
+
+**It was made to bite before it was believed.** On a local API 36 emulator, with the walk cut short
+so the picker is left open and in front and with `device.pressBack()` removed, so that nothing but
+the force-stop can close it:
+
+| | result |
+|---|---|
+| with `forceStopThePicker()` | **passes** — `ActivityManager: Force stopping com.google.android.documentsui ... from pid 5334`, `Killing 5269:com.google.android.documentsui (adj 0)`, a second `PickActivity` opens, the retry completes the pick |
+| with the one call removed | **fails** — `the system picker would not close: after 4 back presses ... com.google.android.documentsui is in front`, which is the API 37 failure verbatim |
+
+The unmutated class passes on that emulator either way, which is the point of running the mutation
+at all: the recovery path is unreachable on a healthy device, so a green suite says nothing about it.
+
 #### The correction that produced that table

 **The first version of this section said both tests failed, and put the marker on the class.** The
@@ -0,0 +1,382 @@
+# E2E-read findings
+
+**Status:** seven findings; E4 fixed, the rest standing, none urgent — **plus one confirmed vacuous test, which is a
+ticket rather than an entry here** (see [Not covered here](#not-covered-here)). `E1`–`E6` came from
+the 2026-09-05 read of the instrumented suite. Every entry here is a *test-suite* observation —
+something a new test would not fix, because the test already exists and the problem is what it
+claims rather than what it runs.
+**Scope:** what reading all 60 instrumented tests turned up that writing a 61st would not fix.
+**Last verified:** `main` at `4b02294`, 2026-09-05. **60 `@Test` methods in 12 classes**, three
+carrying `@FailsOnEmulatorApi37`, gating API 37 leg 57.
+
+## Why this document exists, and why it is separate from the other two
+
+`docs/coverage-read-findings.md` (`F1`–`F10`) came from reading a **JaCoCo report**, and JaCoCo
+measures `testDebugUnitTest` only. So four waves of coverage work have been shaped by a number that
+**cannot see `app/src/androidTest` at all**. The instrumented suite has never had the equivalent
+read: nothing has asked what those 60 tests actually pin, only that they are green.
+
+That is the gap this read is in. It is a **triage, not a test push** — the same shape as wave 4's
+read, which "moved no number at all, and that is its result".
+
+`docs/defect-audit.md` (`D1`–`D16`) is the record of things *wrong at runtime*. Nothing here is
+wrong at runtime. These are tests whose names, KDoc or reputation overstate what they execute.
+
+Entry ids are `E1`–`E6` so they cannot be confused with `F1`–`F10` or `D1`–`D16`.
+
+## How to read the confidence labels
+
+Same vocabulary as the other two documents, deliberately:
+
+- **Confirmed by inspection** — the control flow is fully readable and the finding follows from it.
+- **Confirmed by measurement** — observed in a CI artifact, with the run id recorded.
+- **No action** — recorded because it looks like a finding and is not.
+
+## The method, and the one filter that found everything
+
+A coverage number is useless here by construction, so the read used a different question, applied
+to every one of the 60 tests:
+
+> **If the behaviour this test is named for stopped working, would it go red?**
+
+Three answers, and only the third is a gap:
+
+- **yes** — the test bites. Most of the suite.
+- **no, and that is deliberate and written down** — `RealMediaBenchmark` asserts nothing on purpose
+  (E2); `transcodesH264ToH265AndReportsProgress` declines to assert progress for a stated reason
+  (E3). These are entries here, not tickets.
+- **no, and nothing says so** — the gap. One test, and it is the most important one in the suite.
+
+**The reusable part is the second filter**, because "does it assert something?" would have cleared
+the vacuous test — it asserts two things. What it does not do is *reach the code it names*:
+
+> **Does the test's own premise hold on the machine that runs it?**
+
+`HardwareFallbackTest` asserts `SUCCEEDED` and a non-empty output, and both are true of a
+conversion that never went near the path it exists to prove (**#223**). See
+[Not covered here](#not-covered-here); it is filed rather than recorded here because a test fixes it.
+
+---
+
+## E1 — `RemuxTest`'s class KDoc argues for engine assertions three of its tests do not make, and they are right not to
+
+**Severity: low · Confirmed by inspection · the KDoc is what is wrong, not the tests**
+
+```
+app/src/androidTest/java/org/libremediaconverter/convert/RemuxTest.kt:31-42
+```
+
+The class KDoc is headed **"Why these assert the engine, not just the file"** and makes a specific
+argument:
+
+> A remux routed to FFmpeg produces a perfectly correct file — `-c copy` moves the same samples
+> into the same container. So an output-only assertion passes whether the hardware transmux path
+> ran or never executed at all […] which makes "silently always FFmpeg" the most likely way for
+> this feature to regress.
+
+Five of its seven tests run a conversion. **Three assert no engine at all:**
+
+| test | output container | asserts engine? |
+|---|---|---|
+| `mkvToMp4RemuxesOnHardware` | MP4 | **yes** — `MEDIA3` |
+| `mp4ToMkvRemuxesOnFFmpeg` | MKV | **yes** — `FFMPEG` |
+| `webmToMkvKeepsVp9WithoutReencoding` | MKV | no |
+| `audioOnlySourceRemuxesIntoMka` | MKV (`.mka`) | no |
+| `mp4ToMpegTsAndAviProduceTheirOwnContainers` | MPEG-TS, then AVI | **TS only**; the AVI half does not |
+
+### Why this is not a gap
+
+`ConversionRouter.MEDIA3_CONTAINERS = setOf(Container.MP4)` (`ConversionRouter.kt:37`), and every
+one of the three produces MKV or AVI. **They can only ever be FFmpeg**, so the regression the KDoc
+names — "silently always FFmpeg" — is not a thing that can happen to them. The two tests where the
+hardware path is genuinely at risk are exactly the two that assert it.
+
+An engine assertion on the other three would be near-tautological given today's router. It would
+catch one thing: somebody adding MKV or AVI to `MEDIA3_CONTAINERS` without a muxer to match — which
+is what `Media3MuxersTest` is for, on the JVM, where it does not need a device.
+
+### Why it is recorded rather than dropped
+
+**This was the strongest-looking candidate of the whole read and it dissolved on tracing**, which
+is the same shape as `F5` in the coverage document (filed as a test gap, and only stopped being one
+when someone went looking for its callers). Recorded so the next read does not re-file it.
+
+**The fix is one line of KDoc**, not three tests: the class asserts the engine *where the engine is
+in doubt*, which is a better rule than the one it currently states.
+
+---
+
+## E2 — three of the 60 instrumented tests assert nothing, and two of them never run
+
+**Severity: n/a · No action — deliberate, documented, and load-bearing as documentation**
+
+```
+app/src/androidTest/java/org/libremediaconverter/bench/RealMediaBenchmark.kt:25-53
+```
+
+`reportDeviceEncoderCapabilities` logs and asserts nothing. `hardwareVersusSoftwareOnRealVideo` and
+`av1InputRoutesAccordingToDeviceDecodeSupport` are `assumeTrue`-guarded on media that is **not
+committed** and must be staged by hand into the app's internal `filesDir`, so they skip in every
+automated run — they are the "2 skipped" every green leg reports, and `docs/local-emulator.md:305`
+says so.
+
+The class KDoc is unambiguous: *"This is a benchmark, not part of the automated suite […] Not a
+correctness test — the assertions are deliberately loose."*
+
+**No action.** Recorded for one reason: **the suite's headline number is 60, and three of those 60
+are not tests.** Any future statement of the form "60 instrumented tests cover X" is off by three,
+and two of the three have never executed on CI at all.
+
+**It is the opposite of E-nothing, though** — `reportDeviceEncoderCapabilities` runs on every leg
+and logs `BENCH can-encode:`, and **that log line is what confirmed the vacuous test this read
+found** (**#223**). An assertion-free test that prints the machine's capabilities turned out to be
+the only oracle in the suite. See [Not covered here](#not-covered-here).
+
+---
+
+## E3 — `transcodesH264ToH265AndReportsProgress` does not assert that progress was reported
+
+**Severity: low · No action on the test; the name is the inaccurate part**
+
+```
+app/src/androidTest/java/org/libremediaconverter/convert/Media3EngineTest.kt:73, :90-93
+```
+
+```kotlin
+// Deliberately NOT asserting that progress fired. Polling is on a 250 ms tick,
+// and a 3 s 320x240 clip can finish inside one tick on fast hardware, which
+// would make the assertion fail intermittently for no real defect.
+seen.forEach { assertTrue("progress out of range: $it", it in 0..100) }
+```
+
+`seen` is empty-safe: `forEach` on an empty list asserts nothing, so replacing `onProgress` with a
+no-op reddens nothing here. The reasoning is sound and the alternative really is a flaky test.
+
+**No action on the body.** The name says `AndReportsProgress` and the body says it does not check
+that, which is the `probeForConcat` shape from `CLAUDE.md` — *a passing test with a wrong
+explanation is its own failure mode* — in its mildest form, since here the KDoc immediately corrects
+the name.
+
+**Contrast the FFmpeg side, which is a real gap and is filed as #229**: `FFmpegEngine`'s percentage
+arithmetic is executed by every FFmpeg test and observed by none, because every call site omits
+`onProgress` entirely. Media3's is unasserted; FFmpeg's is unobserved. Only the second is a ticket.
+
+---
+
+## E4 — the marker's KDoc says removing it grows the gating leg by two; three tests carry it
+
+**Severity: low · Confirmed by inspection · one line**
+
+```
+app/src/androidTest/java/org/libremediaconverter/FailsOnEmulatorApi37.kt:20
+```
+
+> Delete the annotation from the tests, and the advisory job goes empty and the gating one grows by
+> **two**.
+
+Three tests carry it — `Media3EngineTest:72`, `Media3EngineTest:135`, `SafPickerRoundTripTest:320` —
+and `FAILS_ON_EMULATOR_API37_BASELINE = 3` eleven lines further down the same file, where the count
+is machine-checked by `.github/scripts/e2e-report-shape.sh`.
+
+The third marker was added when the SAF rotation test was excluded; the sentence was not updated
+with it. **Everything that is checked is consistent at three**; only the prose says two, which is
+exactly why it drifted — and a good argument for the baseline const being a const.
+
+---
+
+## E5 — `coverage-read-findings.md`'s F7 calls covered code uncovered
+
+**Severity: low · Confirmed by inspection · half of F7 is stale**
+
+F7 says `probeWithExtractor`'s catch (`MediaProbe.kt:180-182`) is unreachable on Robolectric and
+"stays device-only", measured across four URI shapes. **The unreachability claim is correct and
+stands.** The implication readers take from it — that nothing exercises it — does not:
+
+```
+app/src/androidTest/java/org/libremediaconverter/convert/RemuxTest.kt:111
+```
+
+`probeDistinguishesAudioFromImagesFromRubbish` feeds it a file of random bytes and asserts
+`InputKind.UNPARSEABLE`, on a device, on every gating leg.
+
+**"Device-only" holds; "uncovered" does not** — and the difference matters, because F7 is one of the
+six entries that document calls "no action", on the grounds that a test would not help. A test
+already exists. The entry should say so.
+
+**This is the failure mode the split between the two documents was meant to prevent**, and it caught
+this repo out: a JaCoCo-derived document cannot see `androidTest`, so it will keep re-deriving
+"uncovered" for anything the instrumented suite covers. That is a structural reason for this
+document to exist, not a one-off correction.
+
+---
+
+## E6 — the suite's one device-capability assertion derives its expectation from the call it is testing
+
+**Severity: low · Confirmed by inspection · no independent oracle exists**
+
+```
+app/src/androidTest/java/org/libremediaconverter/work/ConversionWorkerTest.kt:151-152
+```
+
+```kotlin
+val hasHardwareHevc = AndroidDeviceCodecs.get().canEncode(VideoCodec.H265)
+```
+
+and then the expectation is `if (hasHardwareHevc) MEDIA3 else FFMPEG`. The test asks
+`AndroidDeviceCodecs` what to expect and then checks that the router agreed with
+`AndroidDeviceCodecs`. **If the whole enumeration returned empty, this would still pass** — and
+empty is precisely what the `runCatching` fallback returns (the reason `#194` was worth cutting;
+it logs "assuming permissive" while making `canEncode` answer *no* for everything).
+
+Its KDoc defends the choice, and the defence is good:
+
+> Asserting MEDIA3 unconditionally tests the test machine, not the router.
+
+That is true, and there is no third source of truth on a device: `MediaCodecList` is what
+`AndroidDeviceCodecs` reads, so any oracle built from it is the same oracle.
+
+**No action, but read it with #223.** It is the same missing oracle that makes the
+vacuous-test fix a judgement call rather than a one-liner — you cannot assert "this device has
+hardware HEVC" from inside the suite without asking the class under test. The honest options are a
+visible skip or a red test, and that decision is the ticket's.
+
+---
+
+## E7 — a real `DocumentsProvider` cannot be reached without the picker, so there is no cheap SAF test
+
+**Severity: n/a · Confirmed by measurement · this is a platform rule, not a gap**
+
+Added 2026-09-06, from doing #225 and #226 rather than from reading.
+
+`OutputPublisher.publish`'s destination side is asserted only against Robolectric fakes —
+`FakeSafProvider`, registered with `asDocumentsProvider = true`, which is the flag that *makes*
+`DocumentsContract.isDocumentUri` answer true. #226 split that into a cheap headless half (drive a
+real `DocumentsProvider` directly) and an expensive picker-driven half.
+
+**The cheap half does not exist.** Three approaches, all measured on an API 34 emulator:
+
+| approach | result |
+|---|---|
+| a second `DOCUMENTS_PROVIDER` declared **without** `MANAGE_DOCUMENTS` | refused at install: `SecurityException: Provider must be protected by MANAGE_DOCUMENTS` |
+| create the document as the **test APK**, which owns the provider | denied — instrumentation runs *in the target app's process*, so it carries the app's uid whatever `Context` is asked |
+| `uiAutomation.adoptShellPermissionIdentity(MANAGE_DOCUMENTS)` | denied identically |
+
+The denial names the only way in:
+
+> `Permission Denial: opening provider …FixtureDocumentsProvider from
+> ProcessRecord{… org.libremediaconverter/u0a192} requires that you obtain access using
+> ACTION_OPEN_DOCUMENT or related APIs`
+
+And the intent filter is not optional: without it `isDocumentUri` returns false, which is exactly
+the branch guarding `deletePartialOutput` — so a provider without the filter tests nothing the
+ticket is about.
+
+**So any test of `publish` against a real `DocumentsProvider` must drive DocumentsUI**, and pays
+#190's flake tax. The work is one item at that cost, not two, and #226 was updated to say so.
+
+### What this does *not* block, which is the useful half
+
+`FFmpegKitConfig.getSafParameterForRead` — the bridge on every real conversion and join — needs no
+documents provider. It opens a descriptor through the resolver, so **any readable `content://` URI
+exercises it**, and an ordinary `ContentProvider` may be exported without a permission. That is what
+`FixtureContentProvider` is, and it made #225 headless.
+
+**That distinction was worth the trouble**: the first test ever to hand the join path a real
+`content://` input found #238, a defect that broke joining for every user who picks matched files.
+The expensive gate protects the *destination* side; the *input* side never needed it.
+
+## Summary
+
+| ID | Finding | Severity | Evidence | Action |
+|---|---|---|---|---|
+| E1 | `RemuxTest`'s KDoc claims engine assertions three of its tests correctly omit | low | confirmed by inspection; traced through `MEDIA3_CONTAINERS` | **one line of KDoc** — the tests are right |
+| E2 | Three of the 60 instrumented tests assert nothing; two never run | n/a | confirmed by inspection; `docs/local-emulator.md:305` | **no action** — deliberate; but 60 ≠ 60 |
+| E3 | `…AndReportsProgress` does not assert progress fired | low | confirmed by inspection; reason inline | **no action** — the name overstates, the KDoc corrects it |
+| E4 | The API 37 marker's KDoc says "two"; three tests carry it | low | confirmed by inspection; baseline const says 3 | **fixed** in #243 — it names the constant now |
+| E5 | `coverage-read-findings.md` F7's "uncovered" half is stale | low | confirmed by inspection; `RemuxTest.kt:111` drives it | **amend F7** — "device-only" stands, "uncovered" does not |
+| E6 | The device-capability assertion asks the class under test what to expect | low | confirmed by inspection; no third oracle exists on a device | **no action** — read with **#223** |
+| E7 | A real `DocumentsProvider` is unreachable without the picker, so #226 has no cheap half | n/a | measured three ways on API 34; each denial names `ACTION_OPEN_DOCUMENT` | **no action** — it re-scoped #226 |
+
+**Six of the seven are prose, not code**, and that is the shape of this read. The instrumented suite
+is in good condition: 57 of its 60 tests bite, the fixtures are committed with their generation
+recipes, and the one class that asserts nothing says so in its first line. What this read found is
+that **the suite's self-description has drifted from the suite** in five small places and one large
+one.
+
+**The large one is not in this table**, because a test fixes it: **#223**.
+
+## Not covered here
+
+**The vacuous test.** `HardwareFallbackTest.aFileMedia3CannotDecodeStillConvertsViaFfmpeg` passes on
+every CI leg without ever entering the fallback it exists to prove. It is **#223**, not an entry
+here, because a test fixes it — and it is the reason this read happened rather than an aside from it.
+
+Measured, not inferred, on run **`34004304566`** (all legs green), from each leg's own
+`e2e-diagnostics-api*` logcat:
+
+```
+I/AndroidDeviceCodecs: Hardware video encoders: []
+I/RealMediaBenchmark: BENCH can-encode: COPY=true, H264=false, H265=false, VP9=false, VP8=false, AV1=false
+I/ConversionWorker: Routing sample_h264_444.mp4 -> OutputSpec(container=MP4, videoCodec=H265,
+                    audioCodec=AAC) via FFMPEG (NO_HARDWARE_ENCODER)
+```
+
+Identical on **API 33, 34, 35 and 37**. (API 36's logcat artifact on that run is truncated to 838 KB
+and carries no test output at all, so it is unread rather than different.) The job is routed
+**straight to FFmpeg before Media3 is attempted**, the `catch` in `runMedia3OrFallBack` is never
+entered, and the test's two assertions — `SUCCEEDED`, output non-empty — are true anyway. It ran in
+448 ms.
+
+**The repository already knew.** `ForcedFailureTest.hardwareFailureFallsBackToSoftware`, in the same
+package, pins `ConversionDependencies.deviceCodecs = { DeviceCodecs.PERMISSIVE }` and says why:
+
+> most emulators expose no hardware video encoder at all -- so the router would legitimately send
+> the job straight to FFmpeg and the hardware path would never be attempted. Without this the test
+> passes on a Pixel and fails on every emulator, which says nothing about the code under test.
+
+`ConversionWorkerTest.routesAFastMp4JobByDeviceCapability` records the same fact a third time. The
+knowledge is in two sibling files; `HardwareFallbackTest` is the one that walked into it — and
+because its assertions are about the *output* rather than the *path*, it passes where
+`ForcedFailureTest` would have failed. **That asymmetry is why nobody noticed.**
+
+**State it precisely.** The fallback *wiring* is covered on every leg by `ForcedFailureTest`, with
+fakes. What has never run on any emulator is a fallback triggered by a **real** mid-export codec
+failure — which is the case `HardwareFallbackTest` exists for, and the only reason
+`sample_h264_444.mp4` is committed at all. That fixture, generated with x264 because Fedora's
+ffmpeg ships openh264 and cannot produce High 4:4:4, does nothing on any CI leg today.
+
+The fix is not one assertion. `KEY_ENGINE_USED` is `FFMPEG` **whether the fallback fired or the
+router went straight there** — asserting it changes nothing. The vacuity guard is two facts
+together: the router chose `MEDIA3` for this request on this device, *and* the worker reported
+`FFMPEG`. Whether to reach that with `assumeTrue` (a visible skip on emulators, and the "2 skipped"
+becomes 3) or with an assertion (red on emulators, announcing it cannot test what it claims) is a
+decision, not a detail — see **E6** for why no third option exists — and **#223** leaves it open.
+
+**The other e2e gaps this read found are tickets too**, and are not repeated here:
+
+| # | Gap |
+|---|---|
+| # | Gap | Outcome |
+|---|---|---|
+| **#223** | `HardwareFallbackTest` never attempts the hardware path on any emulator leg | closed — it skips instead of passing vacuously |
+| **#224** | Cancelling a *running* native session, in any of the three engines | closed — all three engines |
+| **#225** | No `content://` input has reached a *successful* conversion — the ffkitsaf bridge | closed, and it found **#238** |
+| **#226** | `OutputPublisher.publish` against a real `DocumentsProvider` | **open** — re-scoped by E7; one picker-driven item, not two |
+| **#227** | The notification's Cancel action has never been fired | closed |
+| **#228** | `encodesFlacLosslessAudio` and `encodesOpus` pass on any non-empty file | closed |
+| **#229** | FFmpeg's progress percentage is computed everywhere and asserted nowhere | closed |
+| **#230** | *(spike)* whether a running conversion's process can be killed | closed — it cannot; the runner shares the app's process |
+
+**The read's own result, once the tickets were worked: one production defect.** #238 — joining files
+picked through the system picker failed outright on the stream-copy path, because the concat demuxer
+whitelists protocols separately from `-safe 0` and `ffkitsaf` was not on the list. Only `STREAM_COPY`
+feeds the demuxer a list file, and every existing join test passed `Uri.fromFile`, so the one broken
+combination was the only one a user could reach.
+
+That is the argument for this kind of read in one line: the gap was not a missed line or an
+unasserted value, it was **a combination of two covered things that no test put together**.
+
+**Nothing here was filed as a coverage delta.** Each names the mutation that has to go red, which is
+the acceptance criterion wave 4 established and which caught two vacuous tests in that wave before
+they shipped. #223 is the one that shows why the criterion matters: it has two passing assertions and
+still tests nothing.
@@ -312,6 +312,18 @@ on sample media that is deliberately not committed. Its third test,
 `reportDeviceEncoderCapabilities`, has no such guard and runs. A level reporting 0 skipped
 would mean someone had staged sample files, not that something improved.

+**Since #223 there is a third, and it is the interesting one.**
+`HardwareFallbackTest.aFileMedia3CannotDecodeStillConvertsViaFfmpeg` is `assumeTrue`-guarded on
+`AndroidDeviceCodecs.get().canEncode(H265)`, which is false on every emulator image — so it now
+skips here and runs only on the Pixel. It used to *pass* on emulators without ever attempting the
+hardware path, which is worse. **Expect `skipped="3"` locally**, and note the guard is a property
+of the machine rather than of staged files: a level reporting 2 would mean an emulator image had
+gained a hardware HEVC encoder, which is worth knowing.
+
+That test's KDoc carries the measurement, including the part that decides it: forcing the route to
+Media3 anyway does *not* produce a fallback, because the goldfish decoder decodes the High 4:4:4
+fixture despite declaring `NoSupport` for its profile.
+
 ### What the sweep adds, and what it does not

 **The renderer rule held four more times.** No boot log contains the string
@@ -355,12 +355,10 @@ boot_emulator() {
 # may be in one of its restarts and `pm` is simply not published yet. The first attempt at this
 # failed exactly that way, with `cmd: Can't find service: package`.
 #
-# The framework restart at the end is not optional, and finding that out cost a run. By the
-# time `sys.boot_completed` flips, SystemUI has already registered its region-sampling listener,
-# and `pm disable-user` does not retract a registration that already happened -- it only stops
-# the package being started again. So the first attempt disabled SystemUI, reported success, and
-# then died exactly as before with `Starting 0 tests` and four more aborts. `stop; start` cycles
-# zygote deliberately, and the framework that comes back up does not start SystemUI at all.
+# This used to end with a framework restart, described here as "not optional". It was neither
+# optional nor happening -- see the block inside the function. What the first attempt's
+# `Starting 0 tests` and four more aborts actually showed is that a `pm disable-user` on its own
+# buys nothing, which is still true; what was wrong is the conclusion that a restart would.
 disable_region_sampling() {
  local api="$1" out i before after ready
  case "$api" in 37 | 37.*) ;; *) return 0 ;; esac
@@ -383,14 +381,18 @@ disable_region_sampling() {
    return 0
  fi

-  echo "  restarting the framework so the region-sampling listener goes with it"
-  emu_adb shell stop > /dev/null 2>&1
-  emu_adb shell start > /dev/null 2>&1
-  # There is no property worth waiting on here, and an earlier version of this only looked
-  # like it was waiting on one: `stop` does not clear sys.boot_completed, so it still reads
-  # `1` throughout the restart and any loop over it returns at once. The loop below is the
-  # wait -- and it polls the better thing anyway, since `Can't find service: package` is the
-  # failure it exists to prevent.
+  # NO FRAMEWORK RESTART, and the two lines that used to be here are why this comment is long.
+  # They were `emu_adb shell stop` and `emu_adb shell start`, both redirected to /dev/null, and
+  # both root-only -- so what they printed there was `Must be root` and what they did was nothing,
+  # here and in the two CI copies alike. Making them real (2026-09-05) is what established that
+  # the disable never worked in the first place: with the package verified `disabled-user` before
+  # AND after a clean restart on android-37.0, `com.android.systemui` comes up 3 s after
+  # `system_server` regardless, and the same is visible in CI's own logcat. The restart also loses
+  # the package state to PackageManager's delayed write if it lands too soon after the `pm` call,
+  # which cost api37-debug run 34010167885 every test in the leg.
+  #
+  # So the useful part of this function is the quiet window below, not the disable. See
+  # .github/scripts/e2e-run.sh's header, and docs/api-37-emulator-crash.md.
  ready=0
  for i in $(seq 1 30); do
    if emu_adb shell service check package 2> /dev/null | grep -q ': found' \