Merge branch 'main' into docs/instrumented-tests-correction

Merge pull request #72 from JMR-dev/test/r38-2-filecard
Say in tests what the file card says when it does not know
2026-08-24 17:10:27 -05:00 · 2026-08-24 17:10:06 -05:00 · 2026-08-24 17:02:19 -05:00 · 2026-08-24 17:01:33 -05:00 · 2026-08-24 16:54:10 -05:00 · 2026-08-24 16:52:16 -05:00
25 changed files with 3735 additions and 264 deletions
@@ -40,6 +40,104 @@ WEDGE_LOG="$TMP/wedge-diagnostics-api${LABEL}.txt"
 # enough under the job's 60-min cap that a genuine wedge still leaves time to capture it.
 WEDGE_TIMEOUT=1200

+# ---------------------------------------------------------------------------
+# API 37 only, and nothing else sets it, so this is inert everywhere it is not wanted --
+# 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.
+#
+# 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
+# `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.
+#
+# 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.
+# ---------------------------------------------------------------------------
+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'; }
+
+disable_region_sampling() {
+  local round=1 i out before after
+  while [ "$round" -le 3 ]; do
+    echo "--- SystemUI disable, 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: $out"
+      case "$out" in *"new state: disabled"*) break ;; esac
+      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 ~$((i * 5)) s"
+        break
+      fi
+      sleep 5
+    done
+
+    if systemui_disabled; then
+      echo "  verified: com.android.systemui is in pm list packages -d"
+    else
+      echo "  NOT DISABLED after the restart -- the package state did not survive"
+      round=$((round + 1))
+      continue
+    fi
+
+    before="$(count_aborts)"
+    sleep 45
+    after="$(count_aborts)"
+    echo "  abort rate, SystemUI disabled: $((after - before)) new in 45 s (total ${after:-0})"
+    [ "$((after - before))" -eq 0 ] && break
+    echo "  still aborting after round $round"
+    round=$((round + 1))
+  done
+
+  # 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"
+  else
+    echo "::warning::E2E api${LABEL}: SystemUI is still enabled -- expect INSTRUMENTATION_ABORTED"
+  fi
+  return 0
+}
+
+if [ "${E2E_DISABLE_SYSTEM_UI:-}" = "1" ]; then
+  echo "::group::E2E api${LABEL} -- removing the region-sampling listener"
+  disable_region_sampling
+  echo "::endgroup::"
+fi
+
 # Stream logcat from now until the step ends, into a file that survives to the artifact upload.
 # Without this, a failure that happens on-device leaves nothing behind: `adb logcat -d` at the
 # end only has whatever is still in the ring buffer, and a chatty test run evicts the cause.
@@ -0,0 +1,413 @@
+name: API 37 debug
+
+# ---------------------------------------------------------------------------
+# WHAT THIS IS FOR, AND WHY IT IS SEPARATE
+#
+# The android-37.x emulator images abort surfaceflinger inside their own gralloc
+# mapper. That was established locally, under -gpu host and under ANGLE
+# (docs/api-37-emulator-crash.md), but NOT on a GitHub runner: CI runs
+# -gpu swiftshader_indirect, and the one local measurement of that mode was void for a
+# purely local reason (Fedora's SELinux denies execheap to SwiftShader's Reactor JIT --
+# docs/local-emulator.md). What CI actually does at API 37 was an open question, and
+# this workflow is the instrument that answered it.
+#
+# IT IS STILL THE INSTRUMENT. status_check.yml now carries API 37 -- a gating leg that
+# disables SystemUI first, and an advisory one for the two @FailsOnEmulatorApi37 tests
+# -- so this file's job is no longer to decide that, but to test a change to it for one
+# dispatch instead of one commit. The next questions it exists for are written down
+# under "When to revisit" in docs/api-37-emulator-crash.md: a new API 37.x image, or an
+# ATD image for 37, either of which could retire the whole workaround.
+#
+# It is a copy of that E2E job with the matrix replaced by workflow_dispatch inputs,
+# so one hypothesis costs one dispatch rather than one commit. It triggers on nothing
+# else: no push, no pull_request, no schedule. Nothing depends on it and it gates
+# nothing.
+#
+# TWO THINGS THIS DELIBERATELY DOES NOT DO:
+#
+#   - 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.
+#   - It does not change status_check.yml. If a configuration here turns out to work,
+#     the change to the real matrix is proposed separately.
+#
+# THE WATCHDOG IS THE POINT, not a nicety. reactivecircus/android-emulator-runner calls
+# killEmulator() from its own catch block, so a run whose emulator never boots is torn
+# down before a single `script:` line executes -- no probe, no e2e-run.sh, no artifacts,
+# nothing to read afterwards. That is precisely the failure shape API 37 is suspected of.
+# The watchdog therefore starts BEFORE the action, from outside it, and samples the device
+# on its own clock.
+# ---------------------------------------------------------------------------
+
+on:
+  workflow_dispatch:
+    inputs:
+      api_level:
+        description: 'API level, as the SDK spells it. 37.0, 37.1, 37.2-beta3, 36 ... A bare 37 does not exist and fails during SDK setup.'
+        type: string
+        default: '37.0'
+      target:
+        description: 'System image target. android-37.1 and 37.2-beta* ship ONLY as google_apis_ps16k -- there is no plain google_apis above 37.0.'
+        type: string
+        default: 'google_apis'
+      channel:
+        description: 'SDK channel. beta is required for any 37.2-beta* image.'
+        type: choice
+        options: ['stable', 'beta', 'dev', 'canary']
+        default: 'stable'
+      gpu_mode:
+        description: 'The -gpu argument. swiftshader_indirect is what status_check.yml uses today; swangle_indirect is what works locally on API 37.'
+        type: choice
+        options:
+          - swiftshader_indirect
+          - swangle_indirect
+          - angle_indirect
+          - host
+          - auto
+          - guest
+          - 'off'
+        default: 'swiftshader_indirect'
+      disable_system_ui:
+        description: 'Take SystemUI out before the suite runs, which is what stops SurfaceFlinger RegionSamplingThread reaching the mapper bug. Restarts the framework.'
+        type: boolean
+        default: false
+      run_tests:
+        description: 'Run the instrumented suite. false boots, probes and stops -- the cheap loop when the question is only whether it boots and at what abort rate.'
+        type: boolean
+        default: true
+      emulator_boot_timeout:
+        description: 'Seconds the action waits for sys.boot_completed. Do not lower this for a software renderer: a slow boot would be misreported as a failed one.'
+        type: string
+        default: '600'
+      emulator_extra_options:
+        description: 'Appended verbatim to the emulator command line -- e.g. "-verbose", or "-feature -GLDMA,-GLDMA2". The action interpolates it into a sh -c, so "| tee $RUNNER_TEMP/emulator.log" also works and is uploaded.'
+        type: string
+        default: ''
+      disable_animations:
+        description: 'The action settings-puts three animation scales after boot. Each is an adb call that throws if the framework is mid-restart, which would kill the run before the probe. false removes three of those calls.'
+        type: boolean
+        default: true
+      gradle_extra_args:
+        description: 'Passed to e2e-run.sh as E2E_EXTRA_GRADLE_ARGS, its existing hook -- e.g. "--rerun", or -Pandroid.testInstrumentationRunnerArguments.class=... to run one class instead of the suite.'
+        type: string
+        default: ''
+      measure_baseline:
+        description: 'Measure the abort rate for 45 s BEFORE disabling SystemUI. Answers "how fast is it aborting"; costs 45 s of crash-looping first, which is a worse starting point for the disable.'
+        type: boolean
+        default: true
+
+run-name: >-
+  api ${{ inputs.api_level }}/${{ inputs.target }} · gpu ${{ inputs.gpu_mode }} ·
+  systemui ${{ inputs.disable_system_ui && 'disabled' || 'running' }} ·
+  tests ${{ inputs.run_tests && 'yes' || 'no' }}
+
+# No `concurrency` block, unlike status_check.yml. Every dispatch here runs on the same
+# ref (main), so a group keyed on github.ref with cancel-in-progress would make two
+# parallel experiments cancel each other -- which is the opposite of what this is for.
+
+permissions:
+  contents: read
+
+env:
+  GRADLE_CACHE_PATHS: |
+    ~/.gradle/caches
+    ~/.gradle/wrapper
+
+jobs:
+  e2e-api37:
+    name: E2E API ${{ inputs.api_level }} (${{ inputs.gpu_mode }})
+    runs-on: ubuntu-latest
+    timeout-minutes: 60
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - uses: actions/setup-java@b6effb05e454b25005698d916606bdc6ffcbf961 # v5.7.0
+        with:
+          distribution: temurin
+          java-version: '25'   # Matches the daemon JVM pinned in gradle/gradle-daemon-jvm.properties
+
+      - uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
+        with:
+          path: ${{ env.GRADLE_CACHE_PATHS }}
+          key: gradle-${{ runner.os }}-${{ hashFiles('**/*.gradle.kts', 'gradle/libs.versions.toml', 'gradle/wrapper/gradle-wrapper.properties') }}
+          restore-keys: gradle-${{ runner.os }}-
+
+      # Without this the emulator falls back to software rendering and takes minutes
+      # longer to boot, when it boots at all.
+      - name: Enable KVM
+        run: |
+          echo 'KERNEL=="kvm", GROUP="kvm", MODE="0666", OPTIONS+="static_node=kvm"' \
+            | sudo tee /etc/udev/rules.d/99-kvm4all.rules
+          sudo udevadm control --reload-rules
+          sudo udevadm trigger --name-match=kvm
+
+      # Both helpers live in RUNNER_TEMP rather than in the repository: they are debug
+      # instrumentation for this workflow only, and writing them here keeps the whole
+      # experiment in one file that can be read top to bottom.
+      - name: Write the watchdog and the probe
+        env:
+          LABEL: ${{ inputs.api_level }}
+        run: |
+          cat > "$RUNNER_TEMP/watchdog.sh" <<'WATCHDOG'
+          #!/usr/bin/env bash
+          # Samples the device from outside the emulator action, because the action tears the
+          # emulator down on a boot timeout before any script: line runs. Everything here is
+          # `timeout`-wrapped: a wedged adb must not stall the sampler, and no probe may fail.
+          SERIAL="emulator-5554"
+
+          # adb is resolved by path, not by name. The emulator action puts platform-tools on
+          # PATH with core.addPath, which only affects LATER steps -- this one already exists
+          # by then, so a bare `adb` here is not the runner's adb and may be nothing at all.
+          # The first version of this file assumed otherwise and every sample came back
+          # boot=? dma_aborts=0 while the action's own adb was working fine two steps away.
+          # Re-resolved every iteration because platform-tools may be installed after this
+          # starts, and echoed to stdout so a repeat of that failure is visible immediately.
+          ADB=""
+          resolve_adb() {
+            for c in "$ADB" "${ANDROID_HOME:-}/platform-tools/adb" "${ANDROID_SDK_ROOT:-}/platform-tools/adb" "$(command -v adb 2> /dev/null)"; do
+              if [ -n "$c" ] && [ -x "$c" ]; then
+                [ "$c" = "$ADB" ] || echo "watchdog: adb resolved to $c"
+                ADB="$c"
+                return 0
+              fi
+            done
+            return 1
+          }
+          OUT="$RUNNER_TEMP/watchdog-api$LABEL.txt"
+          CRASH="$RUNNER_TEMP/crash-buffer-api$LABEL.txt"
+          GUESTLOG="$RUNNER_TEMP/watchdog-logcat-api$LABEL.txt"
+          STOP="$RUNNER_TEMP/watchdog.stop"
+
+          # A continuous guest logcat, restarted whenever the device goes away. During a
+          # surfaceflinger crash loop the framework restarts every few seconds and adb goes
+          # with it, so a single `adb logcat` would end at the first restart.
+          (
+            while [ ! -f "$STOP" ]; do
+              if resolve_adb; then
+                timeout 120 "$ADB" -s "$SERIAL" wait-for-device > /dev/null 2>&1 \
+                  && timeout 3000 "$ADB" -s "$SERIAL" logcat -v time >> "$GUESTLOG" 2>&1
+              fi
+              sleep 3
+            done
+          ) &
+
+          echo "watchdog started $(date -u +%FT%TZ) -- serial $SERIAL" >> "$OUT"
+          i=0
+          while [ "$i" -lt 300 ]; do
+            i=$((i + 1))
+            [ -f "$STOP" ] && break
+            if ! resolve_adb; then
+              echo "$(date -u +%T) no adb yet" >> "$OUT"
+              sleep 20
+              continue
+            fi
+            boot="$(timeout 20 "$ADB" -s "$SERIAL" shell getprop sys.boot_completed 2> /dev/null | tr -d '\r\n')"
+            sf="$(timeout 20 "$ADB" -s "$SERIAL" shell pidof surfaceflinger 2> /dev/null | tr -d '\r\n')"
+            zy="$(timeout 20 "$ADB" -s "$SERIAL" shell pidof zygote64 2> /dev/null | tr -d '\r\n')"
+            # Kept as a file rather than a variable so the last successful read survives the
+            # action killing the emulator -- which is when it is most worth having.
+            if timeout 30 "$ADB" -s "$SERIAL" logcat -d -b crash > "$CRASH.new" 2> /dev/null; then
+              mv "$CRASH.new" "$CRASH"
+            fi
+            dma="$(grep -c 'hasReadColorBufferDma' "$CRASH" 2> /dev/null || true)"
+            sigabrt="$(grep -c 'signal 6' "$CRASH" 2> /dev/null || true)"
+            printf '%s boot=%-4s surfaceflinger=%-8s zygote64=%-8s dma_aborts=%-5s sigabrt=%s\n' \
+              "$(date -u +%T)" "${boot:-?}" "${sf:-none}" "${zy:-none}" "${dma:-0}" "${sigabrt:-0}" >> "$OUT"
+            # One shot, the first time the device is up: which GLES implementation the guest
+            # actually got. This is the guest-side answer to the same question the emulator's
+            # own gles_mode_selected line answers host-side.
+            if [ "$boot" = "1" ] && [ ! -f "$RUNNER_TEMP/renderer-api$LABEL.txt" ]; then
+              {
+                echo "=== booted at $(date -u +%FT%TZ), watchdog sample $i ==="
+                timeout 30 "$ADB" -s "$SERIAL" shell dumpsys SurfaceFlinger 2>&1 | head -40
+                echo "--- getprop ---"
+                timeout 20 "$ADB" -s "$SERIAL" shell getprop 2>&1 | grep -Ei 'egl|gles|gpu|ranchu|gfxstream' || true
+              } > "$RUNNER_TEMP/renderer-api$LABEL.txt" 2>&1
+            fi
+            sleep 20
+          done
+          echo "watchdog finished $(date -u +%FT%TZ) after $i samples" >> "$OUT"
+          WATCHDOG
+
+          cat > "$RUNNER_TEMP/probe.sh" <<'PROBE'
+          #!/usr/bin/env bash
+          # Runs on the booted device, before the suite. Two jobs: record what the guest got,
+          # and measure the gralloc abort RATE -- which is the number that decides whether a
+          # five-minute test run can survive, and the one comparable with the local figures in
+          # docs/api-37-emulator-crash.md (10-11 per 150 s idle under ANGLE with SystemUI up).
+          #
+          # Never exits non-zero. The action runs script: lines in one try/catch, so a failing
+          # probe would skip e2e-run.sh entirely and the run would measure nothing.
+          exec > >(tee -a "$RUNNER_TEMP/probe-api$LABEL.txt") 2>&1
+          echo "===== probe api$LABEL -- $(date -u +%FT%TZ) ====="
+          adb shell getprop sys.boot_completed
+          adb shell getprop ro.build.fingerprint
+          adb shell getprop ro.build.version.sdk
+          echo "--- SurfaceFlinger (the GLES line names the renderer the guest is on) ---"
+          adb shell dumpsys SurfaceFlinger 2>&1 | head -30
+          echo "--- binder services ---"
+          for s in package activity window; do adb shell service check "$s" 2>&1; done
+
+          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'; }
+
+          if [ "${MEASURE_BASELINE:-true}" = "true" ]; then
+            before="$(count_aborts)"
+            sleep 45
+            after="$(count_aborts)"
+            echo "--- abort rate, SystemUI running: $((after - before)) new in 45 s (total ${after:-0}) ---"
+          else
+            # Skipped on purpose when the question is reliability rather than rate: every
+            # second spent measuring is a second of crash-looping, and the disable is what
+            # has to land. A real CI leg would disable as early as it can, so measure that.
+            echo "--- baseline window skipped (MEASURE_BASELINE=false) ---"
+          fi
+
+          if [ "${DISABLE_SYSTEM_UI:-false}" = "true" ]; then
+            # Three rounds, because ONE round is not reliable and the failure is silent.
+            # Measured: of four runs of the same configuration, three came back with the
+            # suite running and one (32646029143) had SystemUI restarting throughout --
+            # `ActivityManager: Start proc N:com.android.systemui ... GradientColorWallpaper`
+            # eight more times after a `pm disable-user` that had reported
+            # `new state: disabled-user`, and ten more RegionSampling aborts with it. The
+            # framework is being SIGKILLed under this loop, so a package-state change can be
+            # lost with the system_server that accepted it. (An earlier version of this comment
+            # said "every ~20 s". That was the watchdog's SAMPLING interval, not the cadence.
+            # Measured: 20-90 s between aborts, median 60-70 s, 3-5 in a four-minute window --
+            # docs/api-37-emulator-crash.md, "Abort cadence, corrected".)
+            #
+            # Nothing here trusts a command's own report. Each round: disable, take the
+            # framework down and confirm it is DOWN before bringing it back (the previous
+            # version 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), then verify
+            # the package is really disabled and that no abort lands in a quiet window.
+            round=1
+            while [ "$round" -le 3 ]; do
+              echo "--- disable 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: $out"
+                case "$out" in *"new state: disabled"*) break ;; esac
+                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 $((i * 5)) s"
+                  break
+                fi
+                sleep 5
+              done
+
+              if systemui_disabled; then
+                echo "  verified: com.android.systemui is in pm list packages -d"
+              else
+                echo "  NOT DISABLED after the restart -- the package state did not survive"
+                round=$((round + 1))
+                continue
+              fi
+
+              before="$(count_aborts)"
+              sleep 45
+              after="$(count_aborts)"
+              echo "--- abort rate, SystemUI disabled: $((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"
+          fi
+
+          echo "--- crash buffer (tail 60) ---"
+          adb logcat -d -b crash 2>&1 | tail -60
+          echo "===== probe done ====="
+          exit 0
+          PROBE
+
+          chmod +x "$RUNNER_TEMP/watchdog.sh" "$RUNNER_TEMP/probe.sh"
+          echo "helpers written to $RUNNER_TEMP"
+
+      - name: Start the watchdog
+        env:
+          LABEL: ${{ inputs.api_level }}
+        run: |
+          echo "ANDROID_HOME=${ANDROID_HOME:-<unset>}  ANDROID_SDK_ROOT=${ANDROID_SDK_ROOT:-<unset>}"
+          echo "adb on PATH: $(command -v adb || echo '<none -- the watchdog will fall back to ANDROID_HOME>')"
+          nohup bash "$RUNNER_TEMP/watchdog.sh" > "$RUNNER_TEMP/watchdog-stdout.txt" 2>&1 < /dev/null &
+          disown
+          echo "watchdog pid $!"
+
+      - name: Instrumented tests
+        uses: reactivecircus/android-emulator-runner@a421e43855164a8197daf9d8d40fe71c6996bb0d # v2.38.0
+        env:
+          LABEL: ${{ inputs.api_level }}
+          DISABLE_SYSTEM_UI: ${{ inputs.disable_system_ui }}
+          MEASURE_BASELINE: ${{ inputs.measure_baseline }}
+          # e2e-run.sh's own hook, unset in CI's real workflow and therefore inert there.
+          E2E_EXTRA_GRADLE_ARGS: ${{ inputs.gradle_extra_args }}
+        with:
+          api-level: ${{ inputs.api_level }}
+          target: ${{ inputs.target }}
+          channel: ${{ inputs.channel }}
+          arch: x86_64
+          profile: pixel_6
+          emulator-boot-timeout: ${{ inputs.emulator_boot_timeout }}
+          emulator-options: -no-window -gpu ${{ inputs.gpu_mode }} -noaudio -no-boot-anim -camera-back none ${{ inputs.emulator_extra_options }}
+          disable-animations: ${{ inputs.disable_animations }}
+          # disk-size, ram-size: kept exactly as status_check.yml pins them, so this
+          # measures the renderer and not a different device. 8G because the APK plus
+          # FFmpeg does not fit the default userdata partition; 2560M because the
+          # emulator's own RAM floor varies by API level and 2560M is the highest of them.
+          disk-size: 8G
+          ram-size: 2560M
+          # Two lines, because the action splits script: on newlines and runs each as its
+          # own `sh -c`. The first is this workflow's own probe; the second is CI's real
+          # harness, invoked unmodified. run_tests: false replaces it with an echo rather
+          # than a second copy of the job.
+          script: |
+            bash ${{ runner.temp }}/probe.sh
+            ${{ inputs.run_tests && format('bash .github/scripts/e2e-run.sh {0}', inputs.api_level) || 'echo "run_tests=false -- suite skipped, boot and probe only"' }}
+
+      - name: Stop the watchdog
+        if: always()
+        run: |
+          touch "$RUNNER_TEMP/watchdog.stop"
+          echo "----- watchdog samples -----"
+          cat "$RUNNER_TEMP/watchdog-api${{ inputs.api_level }}.txt" 2>/dev/null || echo "(no watchdog output)"
+          echo "----- renderer -----"
+          cat "$RUNNER_TEMP/renderer-api${{ inputs.api_level }}.txt" 2>/dev/null || echo "(never booted, or dumpsys unavailable)"
+          echo "----- crash buffer, last read before teardown (tail 80) -----"
+          tail -80 "$RUNNER_TEMP/crash-buffer-api${{ inputs.api_level }}.txt" 2>/dev/null || echo "(none)"
+
+      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        if: always()
+        with:
+          name: api37-debug-run${{ github.run_number }}
+          path: |
+            ${{ runner.temp }}/watchdog-api*.txt
+            ${{ runner.temp }}/watchdog-logcat-api*.txt
+            ${{ runner.temp }}/watchdog-stdout.txt
+            ${{ runner.temp }}/crash-buffer-api*.txt
+            ${{ runner.temp }}/renderer-api*.txt
+            ${{ runner.temp }}/probe-api*.txt
+            ${{ runner.temp }}/emulator*.log
+            ${{ runner.temp }}/logcat-api*.txt
+            ${{ runner.temp }}/diagnostics-api*.txt
+            ${{ runner.temp }}/wedge-diagnostics-api*.txt
+            app/build/reports/androidTests/
+            app/build/outputs/androidTest-results/
+          if-no-files-found: warn
@@ -156,7 +156,37 @@ jobs:
          key: gradle-${{ runner.os }}-${{ hashFiles('**/*.gradle.kts', 'gradle/libs.versions.toml', 'gradle/wrapper/gradle-wrapper.properties') }}
          restore-keys: gradle-${{ runner.os }}-

+      # Shell is the other language in this repo -- four scripts, one of them the CI
+      # entry point itself -- and nothing was checking it. `git ls-files` rather than a
+      # fixed list, so a script added later is covered without editing this workflow.
+      #
+      # Full severity, `info` included. The findings it raises today are answered with
+      # targeted `disable` directives carrying their reason, the same way
+      # config/detekt/detekt.yml carries only the rules this codebase legitimately
+      # breaks. A blanket --severity=warning would have hidden them and the next real
+      # one alike.
+      #
+      # PINNED BY DIGEST, for the reason CLAUDE.md already gives for pinning ktlint,
+      # detekt and JaCoCo: a new rule in a linter makes files nobody touched stop
+      # passing, so CI goes red on a PR whose diff cannot explain it. That is not
+      # hypothetical here. The first cut of this step used the runner's ambient
+      # shellcheck, which is 0.9.0, and 0.9.0 reports a trap handler as seven
+      # unreachable commands (SC2317) where 0.11.0 reports it once on the declaration
+      # (SC2329) -- same script, same directive, different answer, and a red build on
+      # the PR that introduced the step. The version is printed so a finding that
+      # appears out of nowhere can be tied to a bump of this line.
+      - name: shellcheck
+        env:
+          SHELLCHECK: koalaman/shellcheck@sha256:61862eba1fcf09a484ebcc6feea46f1782532571a34ed51fedf90dd25f925a8d
+        run: |
+          docker run --rm "$SHELLCHECK" --version
+          git ls-files -z '*.sh' | xargs -0 -r docker run --rm -v "$PWD:/mnt" "$SHELLCHECK"
+
+      # `!cancelled()` rather than a plain sequence: a shellcheck failure above must not
+      # cost the ktlint/detekt/lint lists. Same reason this step passes --continue -- one
+      # round trip should produce every list, not stop at the first.
      - name: ktlint, detekt and Android lint
+        if: '!cancelled()'
        run: ./gradlew :app:ktlintCheck :app:detekt :app:lintDebug --continue --stacktrace

      # The XML matters as much as the HTML: it is the one that can be diffed between
@@ -178,8 +208,9 @@ jobs:
  # across it -- none below 34, dataSync at 34, mediaProcessing from 35. Testing a
  # single level would leave two thirds of that branch unexercised.
  #
-  # It stops at 36 rather than targetSdk 37 because the android-37.0 emulator image
-  # is broken, not because 37 does not matter. See docs/api-37-emulator-crash.md.
+  # It reaches targetSdk 37, but the API 37 row is not like the other four and the
+  # comment on it says how. Two tests are excluded there and run in their own
+  # advisory job below. See docs/api-37-emulator-crash.md.
  #
  # FFmpeg is not built here. The AAR is committed under bin/, so a red run means the
  # code is broken rather than that a cross-compile hiccuped.
@@ -204,13 +235,28 @@ jobs:
            api-level: "35"
          - label: "36"
            api-level: "36"
-          # No API 37 row. targetSdk is 37, but the android-37.0 emulator image
-          # crash-loops surfaceflinger inside its own gralloc mapper, so every test
-          # fails there no matter what this app does. Ruling that in took four CI
-          # rounds, so the evidence and the ruled-out fixes are written down rather
-          # than left to be rediscovered: docs/api-37-emulator-crash.md. That file
-          # also records what to try first when re-adding it -- note that the row
-          # needs api-level "37.0", since a bare 37 fails during SDK setup.
+          # 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 in this suite touches system UI -- these are 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.
+          #
+          # api-level must be "37.0". A bare 37 is not an SDK package and fails
+          # during setup, which cost a run to discover.
+          #
+          # notAnnotation removes the two tests that do not pass on this image; they
+          # run in the advisory job below, off the same marker so they cannot end up
+          # in both or neither. docs/api-37-emulator-crash.md has the measurements.
+          - label: "37"
+            api-level: "37.0"
+            disable-system-ui: "1"
+            gradle-args: "-Pandroid.testInstrumentationRunnerArguments.notAnnotation=org.libremediaconverter.FailsOnEmulatorApi37"
    steps:
      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1

@@ -236,6 +282,12 @@ jobs:

      - name: Instrumented tests
        uses: reactivecircus/android-emulator-runner@a421e43855164a8197daf9d8d40fe71c6996bb0d # v2.38.0
+        # Both of these are empty on every row but 37, and both are read with a
+        # `:-` default in e2e-run.sh, so the four legs below 37 run the identical
+        # gradle command they always have.
+        env:
+          E2E_DISABLE_SYSTEM_UI: ${{ matrix.disable-system-ui }}
+          E2E_EXTRA_GRADLE_ARGS: ${{ matrix.gradle-args }}
        with:
          api-level: ${{ matrix.api-level }}
          target: google_apis
@@ -297,3 +349,111 @@ jobs:
          name: e2e-wedge-api${{ matrix.label }}
          path: ${{ runner.temp }}/wedge-diagnostics-api${{ matrix.label }}.txt
          if-no-files-found: ignore
+
+  # ---------------------------------------------------------------------------
+  # The two API 37 tests the gating row above excludes, run on their own so they
+  # stay visible instead of disappearing behind a notAnnotation.
+  #
+  # continue-on-error: it reports, it never blocks. That is the whole reason it is
+  # a separate job rather than a sixth matrix row: a row would share the gating
+  # job's `E2E API <label>` name, and a check cannot be both required and advisory
+  # under one name.
+  #
+  # It is named for WHAT IT RUNS, deliberately. Both tests drive a full H.264 ->
+  # H.265 hardware transcode through Media3Engine -- which is exactly what
+  # separates them from the two Media3EngineTest cases that pass here, since those
+  # two never decode video. The current theory about why they fail is in the next
+  # paragraph, where it can be corrected without renaming a check that people have
+  # already learned to look for.
+  #
+  # THEORY, NOT SETTLED: the exception surfaces at `dequeueOutputBuffer` on
+  # `c2.goldfish.h264.decoder`, the emulator's own codec, which gets its frames out
+  # of a host-side colour buffer -- the same readback machinery that aborts
+  # surfaceflinger on this image. What is MEASURED is narrower: these two fail on
+  # the API 37 emulator image; pass at API 36 on this runner under the same renderer
+  # AND the same SystemUI-disable path; pass at API 33-36 without that path at all,
+  # since nothing below 37 needs it; and pass on a physical Pixel 10 Pro XL at 37. That the decoder is the culprit rather than something else
+  # the decode path touches is inference. docs/api-37-emulator-crash.md separates
+  # the two, and the images also differ on the encoder side, which is why "broken
+  # h264 decoder" is not written into this job's name.
+  #
+  # WHEN THIS GOES GREEN, delete the annotation rather than this job: the gating
+  # row picks the tests back up automatically, and this job goes empty and can go
+  # with it.
+  # ---------------------------------------------------------------------------
+  e2e-api37-advisory:
+    name: E2E API 37 Media3 hardware transcode (advisory)
+    runs-on: ubuntu-latest
+    needs: ffmpeg
+    timeout-minutes: 60
+    continue-on-error: true
+    env:
+      E2E_LABEL: "37-media3-transcode"
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - uses: actions/setup-java@b6effb05e454b25005698d916606bdc6ffcbf961 # v5.7.0
+        with:
+          distribution: temurin
+          java-version: '25'
+
+      - uses: actions/cache@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0
+        with:
+          path: ${{ env.GRADLE_CACHE_PATHS }}
+          key: gradle-${{ runner.os }}-${{ hashFiles('**/*.gradle.kts', 'gradle/libs.versions.toml', 'gradle/wrapper/gradle-wrapper.properties') }}
+          restore-keys: gradle-${{ runner.os }}-
+
+      - name: Enable KVM
+        run: |
+          echo 'KERNEL=="kvm", GROUP="kvm", MODE="0666", OPTIONS+="static_node=kvm"' \
+            | sudo tee /etc/udev/rules.d/99-kvm4all.rules
+          sudo udevadm control --reload-rules
+          sudo udevadm trigger --name-match=kvm
+
+      - name: Instrumented tests
+        uses: reactivecircus/android-emulator-runner@a421e43855164a8197daf9d8d40fe71c6996bb0d # v2.38.0
+        env:
+          E2E_DISABLE_SYSTEM_UI: "1"
+          # The complement of the gating row's notAnnotation, off the same marker,
+          # so a test can never be excluded from both jobs or run in both.
+          E2E_EXTRA_GRADLE_ARGS: "-Pandroid.testInstrumentationRunnerArguments.annotation=org.libremediaconverter.FailsOnEmulatorApi37"
+        with:
+          # Every device pin below matches the gating row exactly, so a difference
+          # between the two jobs is the test selection and nothing else.
+          api-level: "37.0"
+          target: google_apis
+          arch: x86_64
+          profile: pixel_6
+          emulator-options: -no-window -gpu swiftshader_indirect -noaudio -no-boot-anim -camera-back none
+          disable-animations: true
+          disk-size: 8G
+          ram-size: 2560M
+          script: bash .github/scripts/e2e-run.sh ${{ env.E2E_LABEL }}
+
+      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        if: always()
+        with:
+          name: e2e-report-api${{ env.E2E_LABEL }}
+          path: |
+            app/build/reports/androidTests/
+            app/build/outputs/androidTest-results/
+          if-no-files-found: warn
+
+      # Uploaded always, and here it matters more than anywhere else in this file:
+      # this job is EXPECTED to be red, so the logcat is the only thing that says
+      # whether it is red for the known reason or for a new one.
+      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        if: always()
+        with:
+          name: e2e-diagnostics-api${{ env.E2E_LABEL }}
+          path: |
+            ${{ runner.temp }}/logcat-api${{ env.E2E_LABEL }}.txt
+            ${{ runner.temp }}/diagnostics-api${{ env.E2E_LABEL }}.txt
+          if-no-files-found: warn
+
+      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
+        if: always()
+        with:
+          name: e2e-wedge-api${{ env.E2E_LABEL }}
+          path: ${{ runner.temp }}/wedge-diagnostics-api${{ env.E2E_LABEL }}.txt
+          if-no-files-found: ignore
@@ -64,16 +64,28 @@ of the first one that fails.
 on this machine (below), so without it an androidTest compile error is not discovered until CI.
 ktlint and detekt also cover the `test`/`androidTest` source sets that `lintDebug` skips.

-## Instrumented tests do not run locally
+## Instrumented tests: where they actually run

-Two independent reasons, so do not spend time on either:
+This section said the opposite until 2026-08-24, and both of its claims had been false for two
+days. Read it as the current answer, and see the git history if you need the old one.

- **Emulators segfault on this host.** qemu dies on every AVD. Instrumented tests run on CI or on
-  the physical Pixel, never in a local emulator.
- **The API 37 image is broken.** `android-37.0` crash-loops surfaceflinger inside its own gralloc
-  mapper, so every test fails there regardless of this app. `docs/api-37-emulator-crash.md` records
-  the evidence and the ruled-out fixes; CI's matrix therefore stops at API 36 even though targetSdk
-  is 37. **API 37 needs a manual check on the Pixel 10 Pro XL before each release.**
+- **Local emulators work, for API 33-36.** `tools/local-emulator/run-e2e.sh` runs them on this
+  host. The segfault that made this look impossible was not a broken machine: SwiftShader's Reactor
+  JIT writes generated shader code onto the heap and executes it, Fedora's SELinux policy denies
+  `execheap`, and qemu dies. Choosing a different renderer avoids it entirely — `-gpu host`,
+  `angle_indirect` and `swangle_indirect` all boot, while `auto`, `off`, `guest` and
+  `swiftshader_indirect` do not. `docs/local-emulator.md` has the evidence and the per-API renderer
+  table.
+- **CI runs API 37, and it gates.** The matrix is 33/34/35/36/37. Two Media3 hardware-transcode
+  tests fail inside the emulator's own `c2.goldfish.h264.decoder` rather than on anything this app
+  does; they carry `@FailsOnEmulatorApi37` and run in a separate `continue-on-error` job,
+  `E2E API 37 Media3 hardware transcode (advisory)`. The gating leg runs the other 55.
+  **That advisory job is red on every PR, by design** — do not read it as your change breaking
+  something, and do not read a green run as evidence those two tests pass.
+  `docs/api-37-emulator-crash.md` has the measurements.
+
+Still true, and the reason the advisory job is not simply deleted: **API 37 needs a manual check on
+the Pixel 10 Pro XL before each release.** The advisory pair is the one thing CI cannot answer for.

 On a device or emulator, build only the ABI it can execute:

@@ -96,10 +108,69 @@ 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** — currently ~31% of lines. A floor needs a baseline that has
-  settled first.
+- **Coverage is reported, not gated** — **29.8% of lines (629/2113), 28.7% of branches**, measured
+  on `main` 2026-08-23 with `./gradlew :app:jacocoTestReport`. A floor needs a baseline that has
+  settled first, and this one has not: the figure **fell** from the ~31% recorded earlier even
+  though the JVM suite went from 11 test files to 43. Main source grew 4,114 -> 5,715 lines over
+  the same period, so the denominator outran the numerator. Re-measure before quoting it; do not
+  assume more tests means a higher percentage here.
+- **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.
+
+  Three things make that a real bar rather than a slogan here:
+
+  - **Unit-testable is broader than it looks.** The pure-seam pattern — `work/FailureOutcome.kt`
+    documents the reasoning — turns "needs a device" into "a pure function plus a thin edge".
+    Robolectric is in the JVM source set, `compose-ui-test-junit4` with it, so Compose screens are
+    unit testable too. Reach for the seam before concluding something cannot be unit tested.
+  - **E2E is runnable locally**, API 33-36, via `tools/local-emulator/run-e2e.sh` — see
+    "Instrumented tests: where they actually run" above. That was believed impossible until the
+    SELinux/renderer cause was found, and it is what makes the e2e half of this norm enforceable.
+  - **A test has to bite.** Revert the line it covers, confirm it goes red, restore. A review of
+    this codebase ran 46 mutations against a 257-test suite and **9 were vacuous** — five of them
+    passing the whole suite over a completely unguarded code path. Green is not evidence.
+
+  Name what you did not cover and why. Genuine exemptions exist; implied coverage is the problem.
+
 - `kotlin.code.style=official`. Gradle stays Kotlin DSL.

+- **File one-off issues with `tools/github/file-issue.sh`, not `gh issue create`.** `gh issue
+  create` does not touch the project board, so the issue exists, carries its labels, and is
+  invisible in the Kanban — indistinguishable from never having been filed. Measured 2026-08-24:
+  eight issues filed as a scripted batch all reached the board; one filed as a one-off minutes
+  later did not. A batch carries the board step in its loop; **one-offs are where it slips**, which
+  is what the script is for. It resolves the project and Status ids by name rather than caching
+  them, and it **reads the item back** — a mutation returning 200 is not evidence the board shows
+  what was asked for. Exit 3 means the issue was created but did not reach the board, and prints
+  the number so it cannot be lost quietly.
+
+  `above-cut` and `backlog` are **labels from the 2026-08-22 triage pass** — "worked autonomously
+  overnight" and "held for manual review". They are not board columns. Status carries board state;
+  do not put a cut label on a newly filed ticket.
+
+- **shellcheck runs in CI**, inside the Static analysis job, over `git ls-files '*.sh'` so a new
+  script is covered without editing the workflow. It runs at full severity, `info` included: the
+  two findings that raises today are answered with targeted `disable` directives carrying their
+  reason, exactly as `config/detekt/detekt.yml` carries only the rules this codebase legitimately
+  breaks. Do not silence it with `--severity=warning` — that hides the next real finding too.
+  **It is pinned by image digest, and joins ktlint/detekt/JaCoCo in the "Dependency versions"
+  rule above** — for exactly the reason stated there, demonstrated the day it was added. The first
+  cut used the runner's ambient shellcheck. That is **0.9.0**, while the container used to check
+  locally was 0.11.0, and the two disagree about how to report a trap handler: 0.11.0 says
+  `SC2329` once on the declaration, 0.9.0 says `SC2317` on each of seven lines in the body. Same
+  script, same directive, one green and one red. Directives that must survive both name both codes.
+
+  Locally, use the same pin rather than whatever is installed:
+  `podman run --rm -v "$PWD:/mnt:z" docker.io/koalaman/shellcheck@sha256:61862eba... <files>`
+  (the digest is in `status_check.yml`; there is no shellcheck system package on this host).
+
+  **It does not cover inline `run:` blocks in the workflows**, and a good deal of this repo's bash
+  lives there. `actionlint` does cover them — it runs shellcheck over each `run:` — and reports one
+  pre-existing `info` finding in `build.yml`. It is not wired in because every action here is
+  pinned by SHA, and actionlint's usual installer is a `curl | bash` off a moving branch; doing it
+  properly means pinning a container digest. Tracked separately rather than bolted on.
+
 ## Dependency versions

 Libraries **float on minor + patch** (`coreKtx = "1.+"`). Three groups deliberately do not:
@@ -306,6 +306,10 @@ dependencies {
    // the tests stay green.
    testImplementation(platform(libs.compose.bom))
    testImplementation(libs.compose.ui.test.junit4)
+    // For `runTest` alone, in EscapedCoroutineErrors.kt. It arrives transitively with the
+    // rule above anyway; declared because a test file imports it directly, and an import of
+    // something nobody asked for breaks the day the library that pulled it in stops.
+    testImplementation(libs.kotlinx.coroutines.test)

    androidTestImplementation(platform(libs.compose.bom))
    androidTestImplementation(libs.androidx.junit)
@@ -0,0 +1,24 @@
+package org.libremediaconverter
+
+/**
+ * Marks an instrumented test that does not pass 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
+ * one — so a test carrying it runs in exactly one of the two and can never fall through both.
+ * 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).
+ *
+ * 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
+ * emulator image cannot currently answer this one". `docs/api-37-emulator-crash.md` has the
+ * measurements and the one bullet in them that is still inference.
+ *
+ * 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.
+ */
+@Retention(AnnotationRetention.RUNTIME)
+@Target(AnnotationTarget.CLASS, AnnotationTarget.FUNCTION)
+annotation class FailsOnEmulatorApi37
@@ -16,6 +16,7 @@ import org.junit.Assert.assertTrue
 import org.junit.Before
 import org.junit.Test
 import org.junit.runner.RunWith
+import org.libremediaconverter.FailsOnEmulatorApi37
 import org.libremediaconverter.model.ConversionRequest
 import org.libremediaconverter.model.OutputFormat
 import java.io.File
@@ -57,6 +58,7 @@ class Media3EngineTest {
    }

    @Test
+    @FailsOnEmulatorApi37
    fun transcodesH264ToH265AndReportsProgress(): Unit = runBlocking {
        val seen = mutableListOf<Int>()

@@ -119,6 +121,7 @@ class Media3EngineTest {
     * HandlerThread indirection holds before any of that lands in Phase 2.
     */
    @Test
+    @FailsOnEmulatorApi37
    fun runsFromAThreadWithNoLooper() {
        val pool = Executors.newSingleThreadExecutor()
        try {
@@ -33,6 +33,7 @@ import androidx.compose.runtime.saveable.rememberSaveable
 import androidx.compose.runtime.setValue
 import androidx.compose.ui.Alignment
 import androidx.compose.ui.Modifier
+import androidx.compose.ui.platform.testTag
 import androidx.compose.ui.text.style.TextAlign
 import androidx.compose.ui.unit.dp
 import androidx.lifecycle.compose.collectAsStateWithLifecycle
@@ -51,6 +52,7 @@ import org.libremediaconverter.model.VideoCodec
 import org.libremediaconverter.ui.PrimaryButtonHeight
 import org.libremediaconverter.ui.ScreenPaddingHorizontal
 import org.libremediaconverter.ui.ScreenPaddingVertical
+import org.libremediaconverter.ui.TestTags
 import java.util.Locale

@UnstableApi
@@ -116,7 +118,8 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                    onClick = { pickInput.launch(arrayOf("*/*")) },
                    modifier = Modifier
                        .fillMaxWidth()
-                        .height(PrimaryButtonHeight),
+                        .height(PrimaryButtonHeight)
+                        .testTag(TestTags.Converter.CHOOSE_FILE),
                ) { Text("Choose file") }
            }

@@ -147,11 +150,16 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                            // The Advanced picker lets an impossible combination be selected on
                            // purpose, so this is what stops it from being run.
                            enabled = validation.isValid,
-                            modifier = Modifier.fillMaxWidth().height(PrimaryButtonHeight),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .height(PrimaryButtonHeight)
+                                .testTag(TestTags.Converter.CONVERT),
                        ) { Text("Convert") }
                        OutlinedButton(
                            onClick = { pickInput.launch(arrayOf("*/*")) },
-                            modifier = Modifier.fillMaxWidth(),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .testTag(TestTags.Converter.CHOOSE_DIFFERENT_FILE),
                        ) { Text("Choose a different file") }
                    }

@@ -160,11 +168,13 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                        Text("Converting… ${s.percent}%")
                        LinearProgressIndicator(
                            progress = { s.percent / 100f },
-                            modifier = Modifier.fillMaxWidth(),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .testTag(TestTags.Converter.PROGRESS),
                        )
                        OutlinedButton(
                            onClick = viewModel::cancel,
-                            modifier = Modifier.fillMaxWidth(),
+                            modifier = Modifier.fillMaxWidth().testTag(TestTags.CANCEL),
                        ) { Text("Cancel") }
                    }

@@ -183,7 +193,7 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                        )
                        OutlinedButton(
                            onClick = viewModel::cancel,
-                            modifier = Modifier.fillMaxWidth(),
+                            modifier = Modifier.fillMaxWidth().testTag(TestTags.CANCEL),
                        ) { Text("Cancel") }
                    }

@@ -202,11 +212,14 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                        }
                        Button(
                            onClick = { chooseDestination.launch(s.suggestedName) },
-                            modifier = Modifier.fillMaxWidth().height(PrimaryButtonHeight),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .height(PrimaryButtonHeight)
+                                .testTag(TestTags.SAVE_FILE),
                        ) { Text("Save file") }
                        OutlinedButton(
                            onClick = viewModel::reset,
-                            modifier = Modifier.fillMaxWidth(),
+                            modifier = Modifier.fillMaxWidth().testTag(TestTags.START_OVER),
                        ) { Text("Start over") }
                    }

@@ -214,7 +227,10 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                        Text("Saved ${s.displayName}.", style = MaterialTheme.typography.bodyLarge)
                        Button(
                            onClick = viewModel::reset,
-                            modifier = Modifier.fillMaxWidth().height(PrimaryButtonHeight),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .height(PrimaryButtonHeight)
+                                .testTag(TestTags.Converter.CONVERT_ANOTHER),
                        ) { Text("Convert another") }
                    }

@@ -226,7 +242,10 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode
                        )
                        Button(
                            onClick = viewModel::reset,
-                            modifier = Modifier.fillMaxWidth().height(PrimaryButtonHeight),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .height(PrimaryButtonHeight)
+                                .testTag(TestTags.START_OVER),
                        ) { Text("Start over") }
                    }
                }
@@ -237,9 +256,12 @@ fun ConverterScreen(modifier: Modifier = Modifier, viewModel: ConversionViewMode

@OptIn(ExperimentalLayoutApi::class)
@Composable
-private fun FormatPicker(selected: OutputFormat?, onSelect: (OutputFormat) -> Unit) {
+internal fun FormatPicker(selected: OutputFormat?, onSelect: (OutputFormat) -> Unit) {
    Text("Output format", style = MaterialTheme.typography.titleSmall)
-    FlowRow(horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+    FlowRow(
+        horizontalArrangement = Arrangement.spacedBy(8.dp),
+        modifier = Modifier.testTag(TestTags.Converter.FORMAT_CHIPS),
+    ) {
        OutputFormat.entries.forEach { format ->
            FilterChip(
                selected = format == selected,
@@ -267,7 +289,7 @@ private fun FormatPicker(selected: OutputFormat?, onSelect: (OutputFormat) -> Un
 */
@OptIn(ExperimentalLayoutApi::class)
@Composable
-private fun AdvancedPicker(
+internal fun AdvancedPicker(
    spec: OutputSpec,
    validation: Validation,
    onContainer: (Container) -> Unit,
@@ -277,14 +299,23 @@ private fun AdvancedPicker(
 ) {
    var expanded by rememberSaveable { mutableStateOf(false) }

-    TextButton(onClick = { expanded = !expanded }) {
+    TextButton(
+        onClick = { expanded = !expanded },
+        modifier = Modifier.testTag(TestTags.Converter.ADVANCED_TOGGLE),
+    ) {
        Text(if (expanded) "Hide advanced" else "Advanced")
    }

    AnimatedVisibility(visible = expanded) {
-        Column(verticalArrangement = Arrangement.spacedBy(12.dp)) {
+        Column(
+            verticalArrangement = Arrangement.spacedBy(12.dp),
+            modifier = Modifier.testTag(TestTags.Converter.ADVANCED_PANEL),
+        ) {
            Text("Container", style = MaterialTheme.typography.titleSmall)
-            FlowRow(horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+            FlowRow(
+                horizontalArrangement = Arrangement.spacedBy(8.dp),
+                modifier = Modifier.testTag(TestTags.Converter.ADVANCED_CONTAINER_CHIPS),
+            ) {
                Container.entries.forEach { container ->
                    FilterChip(
                        selected = container == spec.container,
@@ -295,7 +326,10 @@ private fun AdvancedPicker(
            }

            Text("Video", style = MaterialTheme.typography.titleSmall)
-            FlowRow(horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+            FlowRow(
+                horizontalArrangement = Arrangement.spacedBy(8.dp),
+                modifier = Modifier.testTag(TestTags.Converter.ADVANCED_VIDEO_CHIPS),
+            ) {
                VideoCodec.entries.forEach { codec ->
                    FilterChip(
                        selected = codec == spec.videoCodec,
@@ -306,7 +340,10 @@ private fun AdvancedPicker(
            }

            Text("Audio", style = MaterialTheme.typography.titleSmall)
-            FlowRow(horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+            FlowRow(
+                horizontalArrangement = Arrangement.spacedBy(8.dp),
+                modifier = Modifier.testTag(TestTags.Converter.ADVANCED_AUDIO_CHIPS),
+            ) {
                AudioCodec.entries.forEach { codec ->
                    FilterChip(
                        selected = codec == spec.audioCodec,
@@ -331,9 +368,11 @@ private fun AdvancedPicker(

@OptIn(ExperimentalLayoutApi::class)
@Composable
-private fun ValidationError(invalid: Validation.Invalid, onSuggestion: (OutputSpec) -> Unit) {
+internal fun ValidationError(invalid: Validation.Invalid, onSuggestion: (OutputSpec) -> Unit) {
    Card(
-        modifier = Modifier.fillMaxWidth(),
+        modifier = Modifier
+            .fillMaxWidth()
+            .testTag(TestTags.Converter.VALIDATION_ERROR),
        colors = CardDefaults.cardColors(
            containerColor = MaterialTheme.colorScheme.errorContainer,
            contentColor = MaterialTheme.colorScheme.onErrorContainer,
@@ -347,10 +386,11 @@ private fun ValidationError(invalid: Validation.Invalid, onSuggestion: (OutputSp
            if (invalid.suggestions.isNotEmpty()) {
                Text("Try instead:", style = MaterialTheme.typography.labelMedium)
                FlowRow(horizontalArrangement = Arrangement.spacedBy(8.dp)) {
-                    invalid.suggestions.forEach { suggestion ->
+                    invalid.suggestions.forEachIndexed { index, suggestion ->
                        AssistChip(
                            onClick = { onSuggestion(suggestion) },
                            label = { Text(describe(suggestion)) },
+                            modifier = Modifier.testTag(TestTags.Converter.suggestion(index)),
                        )
                    }
                }
@@ -359,7 +399,7 @@ private fun ValidationError(invalid: Validation.Invalid, onSuggestion: (OutputSp
    }
 }

-private fun describe(spec: OutputSpec): String {
+internal fun describe(spec: OutputSpec): String {
    val video = when (spec.videoCodec) {
        VideoCodec.NONE -> null
        else -> spec.videoCodec.label
@@ -374,9 +414,12 @@ private fun describe(spec: OutputSpec): String {

@OptIn(ExperimentalLayoutApi::class)
@Composable
-private fun QualityPicker(selected: QualityTier, onSelect: (QualityTier) -> Unit) {
+internal fun QualityPicker(selected: QualityTier, onSelect: (QualityTier) -> Unit) {
    Text("Quality", style = MaterialTheme.typography.titleSmall)
-    FlowRow(horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+    FlowRow(
+        horizontalArrangement = Arrangement.spacedBy(8.dp),
+        modifier = Modifier.testTag(TestTags.Converter.QUALITY_CHIPS),
+    ) {
        QualityTier.entries.forEach { tier ->
            FilterChip(
                selected = tier == selected,
@@ -390,9 +433,12 @@ private fun QualityPicker(selected: QualityTier, onSelect: (QualityTier) -> Unit

@OptIn(ExperimentalLayoutApi::class)
@Composable
-private fun EnginePicker(selected: EnginePreference, onSelect: (EnginePreference) -> Unit) {
+internal fun EnginePicker(selected: EnginePreference, onSelect: (EnginePreference) -> Unit) {
    Text("Engine", style = MaterialTheme.typography.titleSmall)
-    FlowRow(horizontalArrangement = Arrangement.spacedBy(8.dp)) {
+    FlowRow(
+        horizontalArrangement = Arrangement.spacedBy(8.dp),
+        modifier = Modifier.testTag(TestTags.Converter.ENGINE_CHIPS),
+    ) {
        EnginePreference.entries.forEach { preference ->
            FilterChip(
                selected = preference == selected,
@@ -403,7 +449,7 @@ private fun EnginePicker(selected: EnginePreference, onSelect: (EnginePreference
    }
 }

-private fun EnginePreference.label(): String = when (this) {
+internal fun EnginePreference.label(): String = when (this) {
    EnginePreference.AUTO -> "Automatic"
    EnginePreference.PREFER_HARDWARE -> "Prefer hardware"
    EnginePreference.FORCE_SOFTWARE -> "Force software"
@@ -418,21 +464,34 @@ private fun EnginePreference.label(): String = when (this) {
 * pretending it has an unknown codec.
 */
@Composable
-private fun FileCard(input: InputFile) {
-    Card(modifier = Modifier.fillMaxWidth()) {
+internal fun FileCard(input: InputFile) {
+    Card(
+        modifier = Modifier
+            .fillMaxWidth()
+            .testTag(TestTags.Converter.FILE_CARD),
+    ) {
        Column(modifier = Modifier.padding(16.dp)) {
-            Text(input.displayName, style = MaterialTheme.typography.titleMedium)
+            Text(
+                input.displayName,
+                style = MaterialTheme.typography.titleMedium,
+                modifier = Modifier.testTag(TestTags.Converter.FILE_CARD_NAME),
+            )
            // The null is handled here rather than inside formatBytes, because "no provider would
            // say" is not a number and a formatter that invented one -- "0 B" -- is the defect
            // this card would be showing. It degrades in words, like the codec rows below it.
            Text(
                input.sizeBytes?.let(::formatBytes) ?: "Size unknown",
                style = MaterialTheme.typography.bodySmall,
+                modifier = Modifier.testTag(TestTags.Converter.FILE_CARD_BYTES),
            )

            val probe = input.probe
            if (probe == null) {
-                Text("Reading…", style = MaterialTheme.typography.bodySmall)
+                Text(
+                    "Reading…",
+                    style = MaterialTheme.typography.bodySmall,
+                    modifier = Modifier.testTag(TestTags.Converter.FILE_CARD_NOTE),
+                )
                return@Column
            }

@@ -442,6 +501,7 @@ private fun FileCard(input: InputFile) {
                InputKind.UNPARSEABLE -> Text(
                    "Could not identify this file. It will be converted with FFmpeg.",
                    style = MaterialTheme.typography.bodySmall,
+                    modifier = Modifier.testTag(TestTags.Converter.FILE_CARD_NOTE),
                )

                InputKind.IMAGE -> {
@@ -481,22 +541,23 @@ private fun FileCard(input: InputFile) {
 }

@Composable
-private fun DetailRow(label: String, value: String) {
+internal fun DetailRow(label: String, value: String) {
    Text(
        "$label: $value",
        style = MaterialTheme.typography.bodySmall,
        color = MaterialTheme.colorScheme.onSurfaceVariant,
+        modifier = Modifier.testTag(TestTags.Converter.detailRow(label)),
    )
 }

-private fun formatDuration(ms: Long): String {
+internal fun formatDuration(ms: Long): String {
    val totalSeconds = ms / 1000
    val minutes = totalSeconds / 60
    val seconds = totalSeconds % 60
    return String.format(Locale.US, "%d:%02d", minutes, seconds)
 }

-private fun formatBytes(bytes: Long): String = when {
+internal fun formatBytes(bytes: Long): String = when {
    bytes >= 1_000_000_000 -> String.format(Locale.US, "%.1f GB", bytes / 1e9)
    bytes >= 1_000_000 -> String.format(Locale.US, "%.1f MB", bytes / 1e6)
    bytes >= 1_000 -> String.format(Locale.US, "%.0f kB", bytes / 1e3)
@@ -21,6 +21,7 @@ import androidx.compose.runtime.getValue
 import androidx.compose.runtime.remember
 import androidx.compose.ui.Alignment
 import androidx.compose.ui.Modifier
+import androidx.compose.ui.platform.testTag
 import androidx.compose.ui.text.style.TextAlign
 import androidx.compose.ui.unit.dp
 import androidx.lifecycle.compose.collectAsStateWithLifecycle
@@ -31,6 +32,7 @@ import org.libremediaconverter.model.ConcatStrategy
 import org.libremediaconverter.ui.PrimaryButtonHeight
 import org.libremediaconverter.ui.ScreenPaddingHorizontal
 import org.libremediaconverter.ui.ScreenPaddingVertical
+import org.libremediaconverter.ui.TestTags
 import org.libremediaconverter.work.ConcatWorker

@UnstableApi
@@ -82,7 +84,8 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                    onClick = { pickInputs.launch(arrayOf("video/*")) },
                    modifier = Modifier
                        .fillMaxWidth()
-                        .height(PrimaryButtonHeight),
+                        .height(PrimaryButtonHeight)
+                        .testTag(TestTags.Join.CHOOSE_FILES),
                ) { Text("Choose files") }
            }

@@ -97,11 +100,16 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                        s.inputs.forEach { FileRow(it) }
                        Button(
                            onClick = viewModel::join,
-                            modifier = Modifier.fillMaxWidth().height(PrimaryButtonHeight),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .height(PrimaryButtonHeight)
+                                .testTag(TestTags.Join.JOIN),
                        ) { Text("Join ${s.inputs.size} files") }
                        OutlinedButton(
                            onClick = { pickInputs.launch(arrayOf("video/*")) },
-                            modifier = Modifier.fillMaxWidth(),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .testTag(TestTags.Join.CHOOSE_DIFFERENT_FILES),
                        ) { Text("Choose different files") }
                    }

@@ -110,10 +118,14 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                        // Indeterminate on purpose: FFmpeg reports progress against a
                        // single input's duration, which means nothing across a
                        // concatenation. A fabricated percentage would be worse than none.
-                        LinearProgressIndicator(modifier = Modifier.fillMaxWidth())
+                        LinearProgressIndicator(
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .testTag(TestTags.Join.PROGRESS),
+                        )
                        OutlinedButton(
                            onClick = viewModel::cancel,
-                            modifier = Modifier.fillMaxWidth(),
+                            modifier = Modifier.fillMaxWidth().testTag(TestTags.CANCEL),
                        ) { Text("Cancel") }
                    }

@@ -127,7 +139,7 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                        )
                        OutlinedButton(
                            onClick = viewModel::cancel,
-                            modifier = Modifier.fillMaxWidth(),
+                            modifier = Modifier.fillMaxWidth().testTag(TestTags.CANCEL),
                        ) { Text("Cancel") }
                    }

@@ -146,11 +158,14 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                        )
                        Button(
                            onClick = { chooseDestination.launch(s.suggestedName) },
-                            modifier = Modifier.fillMaxWidth().height(PrimaryButtonHeight),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .height(PrimaryButtonHeight)
+                                .testTag(TestTags.SAVE_FILE),
                        ) { Text("Save file") }
                        OutlinedButton(
                            onClick = viewModel::reset,
-                            modifier = Modifier.fillMaxWidth(),
+                            modifier = Modifier.fillMaxWidth().testTag(TestTags.START_OVER),
                        ) { Text("Start over") }
                    }

@@ -158,7 +173,10 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                        Text("Saved ${s.displayName}.", style = MaterialTheme.typography.bodyLarge)
                        Button(
                            onClick = viewModel::reset,
-                            modifier = Modifier.fillMaxWidth().height(PrimaryButtonHeight),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .height(PrimaryButtonHeight)
+                                .testTag(TestTags.Join.JOIN_MORE),
                        ) { Text("Join more") }
                    }

@@ -170,7 +188,10 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
                        )
                        Button(
                            onClick = viewModel::reset,
-                            modifier = Modifier.fillMaxWidth().height(PrimaryButtonHeight),
+                            modifier = Modifier
+                                .fillMaxWidth()
+                                .height(PrimaryButtonHeight)
+                                .testTag(TestTags.START_OVER),
                        ) { Text("Start over") }
                    }
                }
@@ -180,8 +201,12 @@ fun JoinScreen(modifier: Modifier = Modifier, viewModel: JoinViewModel = viewMod
 }

@Composable
-private fun FileRow(input: InputFile) {
-    Card(modifier = Modifier.fillMaxWidth()) {
+internal fun FileRow(input: InputFile) {
+    Card(
+        modifier = Modifier
+            .fillMaxWidth()
+            .testTag(TestTags.Join.fileRow(input.displayName)),
+    ) {
        Column(modifier = Modifier.padding(12.dp)) {
            Text(input.displayName, style = MaterialTheme.typography.bodyMedium)
        }
@@ -0,0 +1,142 @@
+package org.libremediaconverter.ui
+
+/**
+ * Where a test finds each affordance on the two screens.
+ *
+ * Every button, picker and card in `ConverterScreen` and `JoinScreen` carries one of these through
+ * `Modifier.testTag`, so a test names a symbol and never a literal. That is the whole reason the
+ * table exists: `"Cancel"`, `"Start over"` and `"Save file"` are each rendered by both screens and
+ * by more than one state branch, so rewording one of them would otherwise redden several
+ * independent test files at once, and none of those diffs would explain why.
+ *
+ * Tags are applied inside `main`, never handed in by the caller. A tag a test passes down as a
+ * `Modifier` proves only that the test set it -- it would stay green with the affordance's own tag
+ * deleted, which is exactly the vacuous test `CLAUDE.md` records nine of.
+ *
+ * ### Public rather than `internal`, deliberately
+ *
+ * `androidTest` **is** a friend source set of `main` here: an `androidTest` file referencing the
+ * `internal` `Destination.CONVERT` compiles clean through `:app:compileDebugAndroidTestKotlin`
+ * under AGP 9.3.1 (measured 2026-08-24 -- nothing in the repo referenced a main `internal` from
+ * `androidTest`, so the question had no in-tree answer until then). `internal` would compile today.
+ *
+ * It is public anyway. That friendship is AGP wiring rather than something this project states, and
+ * this table is a contract read from three source sets: `main` applies the tags, `src/test` and
+ * `src/androidTest` name them. Public buys no external exposure in an application module -- nothing
+ * consumes it from outside -- so the durable answer costs nothing here.
+ *
+ * ### Invariants
+ *
+ * Values are distinct, which `TagTableUniquenessTest` asserts. Two affordances sharing a tag would
+ * break the "resolves to exactly one node" assertion in a file nobody had touched.
+ */
+object TestTags {
+
+    /**
+     * Affordances both screens render, under one name each.
+     *
+     * Shared rather than per-screen because only one screen is composed at a time -- the shell
+     * swaps them -- so a tag can only ever resolve within the screen under test.
+     */
+    const val CANCEL: String = "action.cancel"
+
+    /** Rendered by `Converted`/`Joined` and again by `Failed` on both screens. */
+    const val START_OVER: String = "action.startOver"
+
+    const val SAVE_FILE: String = "action.saveFile"
+
+    /** `ConverterScreen`. */
+    object Converter {
+        const val CHOOSE_FILE: String = "converter.chooseFile"
+        const val CONVERT: String = "converter.convert"
+        const val CHOOSE_DIFFERENT_FILE: String = "converter.chooseDifferentFile"
+        const val CONVERT_ANOTHER: String = "converter.convertAnother"
+
+        /** The determinate bar in `Converting`. It carries no text, so nothing else can find it. */
+        const val PROGRESS: String = "converter.progress"
+
+        const val FILE_CARD: String = "converter.fileCard"
+        const val FILE_CARD_NAME: String = "converter.fileCard.name"
+
+        /**
+         * The byte size, or `"Size unknown"`.
+         *
+         * Named for bytes rather than "size" because the `IMAGE` branch also renders a row labelled
+         * `Size` -- pixel dimensions -- through [detailRow], and the two mean different things.
+         */
+        const val FILE_CARD_BYTES: String = "converter.fileCard.bytes"
+
+        /**
+         * The one-line explanation that stands in for the detail rows: `"Reading…"` while the probe
+         * is still running, or the unreadable-file line once it has finished and found nothing.
+         * The two are mutually exclusive, so one tag covers both.
+         */
+        const val FILE_CARD_NOTE: String = "converter.fileCard.note"
+
+        /**
+         * The chip rows, not the pickers around them.
+         *
+         * Each tag sits on the `FlowRow` of chips, so the prose a picker renders beside it -- the
+         * `"Custom — set below."` line under the formats, the tier description under the quality
+         * chips -- is outside the tagged node. Tagging the picker as a whole would mean wrapping
+         * three sibling emissions in a layout that does not exist today.
+         */
+        const val FORMAT_CHIPS: String = "converter.formatChips"
+
+        const val QUALITY_CHIPS: String = "converter.qualityChips"
+        const val ENGINE_CHIPS: String = "converter.engineChips"
+
+        /** The `Advanced` / `Hide advanced` toggle. Present whether or not the panel is open. */
+        const val ADVANCED_TOGGLE: String = "converter.advanced.toggle"
+
+        /** The panel the toggle gates. Absent from the tree while collapsed. */
+        const val ADVANCED_PANEL: String = "converter.advanced.panel"
+
+        /**
+         * The three chip rows inside the panel, separately.
+         *
+         * Separately because their labels collide: `"Copy"` and `"None"` are both a `VideoCodec`
+         * and an `AudioCodec`, and `"MP3"` and `"FLAC"` are both a `Container` and an `AudioCodec`,
+         * so a text matcher over the open panel is ambiguous for four of the chips.
+         */
+        const val ADVANCED_CONTAINER_CHIPS: String = "converter.advanced.containerChips"
+
+        const val ADVANCED_VIDEO_CHIPS: String = "converter.advanced.videoChips"
+        const val ADVANCED_AUDIO_CHIPS: String = "converter.advanced.audioChips"
+
+        /** The error card. Rendered outside the panel, so it is reachable while collapsed. */
+        const val VALIDATION_ERROR: String = "converter.validationError"
+
+        /** One detail line of the file card, by the label it renders: `Container`, `Video`, ... */
+        fun detailRow(label: String): String = "converter.fileCard.row:$label"
+
+        /**
+         * One suggested output on the validation card, by position.
+         *
+         * By position rather than by the text of the suggestion, because that text comes from
+         * `describe`, which is itself under test -- a tag derived from it would move whenever the
+         * thing it is meant to locate changed.
+         */
+        fun suggestion(index: Int): String = "converter.validationError.suggestion:$index"
+    }
+
+    /** `JoinScreen`. */
+    object Join {
+        const val CHOOSE_FILES: String = "join.chooseFiles"
+        const val JOIN: String = "join.join"
+        const val CHOOSE_DIFFERENT_FILES: String = "join.chooseDifferentFiles"
+        const val JOIN_MORE: String = "join.joinMore"
+
+        /** The indeterminate bar in `Joining`. */
+        const val PROGRESS: String = "join.progress"
+
+        /**
+         * One picked input, by the name it displays.
+         *
+         * By name rather than by position, so the tag is derived from data the row already holds
+         * and can stay inside `FileRow`. Passing an index down would mean the call site owned the
+         * tag, and a test that supplies its own tag asserts nothing about the screen.
+         */
+        fun fileRow(displayName: String): String = "join.fileRow:$displayName"
+    }
+}
@@ -8,7 +8,6 @@ import androidx.compose.runtime.setValue
 import androidx.compose.ui.platform.testTag
 import androidx.compose.ui.test.assertIsSelected
 import androidx.compose.ui.test.junit4.StateRestorationTester
-import androidx.compose.ui.test.junit4.v2.createComposeRule
 import androidx.compose.ui.test.onNodeWithTag
 import androidx.compose.ui.test.onNodeWithText
 import androidx.compose.ui.test.performClick
@@ -42,8 +41,11 @@ import org.robolectric.RobolectricTestRunner
@RunWith(RobolectricTestRunner::class)
 class AppRootRestorationTest {

+    // Not `createComposeRule()` directly: see [drainEscapedCoroutineErrors]. Every Compose test
+    // class in this source set starts there, whether or not it is the one that happens to be
+    // running when another test's escaped coroutine error is delivered.
    @get:Rule
-    val composeRule = createComposeRule()
+    val composeRule = createDrainedComposeRule()

    private val restoration = StateRestorationTester(composeRule)

@@ -0,0 +1,53 @@
+package org.libremediaconverter
+
+import androidx.compose.ui.test.junit4.v2.createComposeRule
+import kotlinx.coroutines.test.runTest
+
+/**
+ * Clears coroutine errors this module's tests deliberately let escape, so they land on the test
+ * that caused them instead of on the next one to start.
+ *
+ * **Every Compose test class in `src/test` has to start here.** `createComposeRule` runs the
+ * composition inside `runTest`, and `runTest` opens by throwing `UncaughtExceptionsBeforeTest` for
+ * anything already sitting in kotlinx-coroutines-test's collector -- a process-wide
+ * `CoroutineExceptionHandler` it installs once and never removes.
+ *
+ * There is one deposit into that collector here, and it is not a mistake:
+ * `ConversionViewModelProbeFailureTest.an OutOfMemoryError is not swallowed` proves an OOM raised
+ * inside the probe is rethrown rather than reported as an unreadable file. `onInputPicked` runs it
+ * in `viewModelScope.launch`, which has no exception handler by design -- the ViewModel's own KDoc
+ * says a real OOM should reach the thread's handler and take the process down. On the JVM the
+ * collector takes it instead, holds it, and hands it to whichever `runTest` starts next.
+ *
+ * It surfaced as two *different* Compose test classes failing on two consecutive runs of the same,
+ * green, code, with a message naming neither the test nor the error's origin. Which class catches
+ * it moves because the throw happens on a real `Dispatchers.IO` thread, after the state assertion
+ * that ends the test that caused it -- so it can be delivered long after that class is done.
+ *
+ * A `@Before` method cannot do this: the compose rule's `runTest` wraps the statement that calls
+ * `@Before`, so it has already thrown. `@BeforeClass` cannot either -- Robolectric runs it outside
+ * the sandbox classloader, where the collector is a different object. Draining while the rule is
+ * being *constructed* is early enough, because JUnit builds a fresh test-class instance, and with
+ * it every `@get:Rule` field, before evaluating any rule.
+ *
+ * The real fix is a seam: give the probe hop an injectable dispatcher the way
+ * `ConversionViewModel`'s constructor already does for `cleanupDispatcher`, and the error would
+ * have somewhere to land. That is a production change, so it belongs in its own commit.
+ */
+fun drainEscapedCoroutineErrors() {
+    // Entering a test scope is what flushes the collector; the flush is reported as this
+    // throwing, and there is nothing to assert about an error another test already asserted on.
+    runCatching { runTest {} }
+}
+
+/**
+ * [createComposeRule], with [drainEscapedCoroutineErrors] run first. Use this rather than
+ * `createComposeRule` directly in `src/test`.
+ *
+ * It also keeps the one mixed import in one place: the rule comes from the **v2** package
+ * (`androidx.compose.ui.test.junit4.v2`) while `StateRestorationTester`, which takes it, does not.
+ */
+fun createDrainedComposeRule() = run {
+    drainEscapedCoroutineErrors()
+    createComposeRule()
+}
@@ -0,0 +1,159 @@
+package org.libremediaconverter.convert
+
+import android.os.Bundle
+import android.os.Parcel
+import android.os.Parcelable
+import androidx.compose.runtime.CompositionLocalProvider
+import androidx.compose.runtime.MutableState
+import androidx.compose.runtime.saveable.LocalSaveableStateRegistry
+import androidx.compose.runtime.saveable.SaveableStateRegistry
+import androidx.compose.ui.test.onNodeWithTag
+import androidx.compose.ui.test.performClick
+import androidx.media3.common.util.UnstableApi
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.createDrainedComposeRule
+import org.libremediaconverter.model.AudioCodec
+import org.libremediaconverter.model.Container
+import org.libremediaconverter.model.OutputSpec
+import org.libremediaconverter.model.Validation
+import org.libremediaconverter.model.VideoCodec
+import org.libremediaconverter.ui.TestTags
+import org.robolectric.RobolectricTestRunner
+
+/**
+ * What `AdvancedPickerTest`'s restoration test cannot see.
+ *
+ * `StateRestorationTester` saves into an **in-memory map**, never a `Bundle`. That is enough to
+ * discriminate `rememberSaveable` from `remember`, and it is where it stops: the map holds object
+ * references, so a value the platform could never parcel goes in and comes back out looking green.
+ * `AppRootRestorationTest` has the same blind spot and `DestinationSaverTest` is the split it
+ * prompted; this is that split for `expanded`, the only `rememberSaveable` on either screen's
+ * leaves.
+ *
+ * ### The saved representation is not the Boolean
+ *
+ * `var expanded by rememberSaveable { mutableStateOf(false) }` passes no `stateSaver`, so
+ * `autoSaver` saves **the `MutableState` itself**, not the `false` inside it. That works only
+ * because `mutableStateOf` on Android returns a `Parcelable` implementation -- the same call on a
+ * plain JVM returns one that is not. So what stands between an open panel and a rotation that
+ * closes it is a platform-specific detail of a factory function nothing here names directly, and
+ * an in-memory map cannot tell the two apart.
+ *
+ * Pinning it is the move `DestinationSaverTest` makes about names versus ordinals. Passing an
+ * explicit `stateSaver` would save a bare `Boolean` instead and is a perfectly reasonable edit --
+ * it is just not the one in the tree, and it should be made on purpose rather than discovered
+ * after a rotation.
+ *
+ * ### Shared bite, stated rather than implied
+ *
+ * `rememberSaveable` -> `remember` empties the registry, so it reddens this file *and* the
+ * restoration test in `AdvancedPickerTest`. Both failures belong in any report of that mutation.
+ */
+@UnstableApi
+@RunWith(RobolectricTestRunner::class)
+class AdvancedPanelSavedStateTest {
+
+    // Not `createComposeRule()` directly: see [drainEscapedCoroutineErrors].
+    @get:Rule
+    val composeRule = createDrainedComposeRule()
+
+    /**
+     * `canBeSaved = { true }` deliberately.
+     *
+     * A predicate mirroring what a `Bundle` accepts would be a hand-written copy of the thing
+     * under test, and a `false` from it *drops* the entry silently -- so the test would fail by
+     * finding nothing saved, which is also how a `remember` regression fails. Two causes, one
+     * symptom, is not a test. The type is checked on the way out instead.
+     */
+    private val registry = SaveableStateRegistry(restoredValues = null, canBeSaved = { true })
+
+    @Test
+    fun `the panel registers its open state with the registry, and nothing else`() {
+        setPicker()
+
+        // Collapsed is a saved value, not an absent one: `rememberSaveable` registers its provider
+        // on first composition, whatever the state happens to be. Exactly one, because `expanded`
+        // is the only saveable in the subtree -- a second would mean something else began saving.
+        assertEquals(1, savedValues().size)
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).performClick()
+
+        val saved = theOneSavedValue()
+        assertTrue("saved as ${saved?.javaClass?.name}", saved is MutableState<*>)
+        assertEquals(true, (saved as MutableState<*>).value)
+    }
+
+    @Test
+    fun `the open panel survives a real Parcel, not just an in-memory map`() {
+        setPicker()
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).performClick()
+
+        val saved = theOneSavedValue()
+
+        // The claim the restoration test cannot make. A `MutableState` that was not `Parcelable`
+        // would satisfy `StateRestorationTester` and then be dropped by the platform.
+        assertTrue("saved as ${saved?.javaClass?.name}", saved is Parcelable)
+
+        val restored = throughARealBundle(saved as Parcelable)
+
+        assertTrue("restored as ${restored.javaClass.name}", restored is MutableState<*>)
+        assertEquals(true, (restored as MutableState<*>).value)
+    }
+
+    private fun setPicker() {
+        composeRule.setContent {
+            CompositionLocalProvider(LocalSaveableStateRegistry provides registry) {
+                AdvancedPicker(
+                    spec = OutputSpec(Container.MP4, VideoCodec.H264, AudioCodec.AAC),
+                    validation = Validation.Valid,
+                    onContainer = {},
+                    onVideoCodec = {},
+                    onAudioCodec = {},
+                    onSuggestion = {},
+                )
+            }
+        }
+    }
+
+    /** Every value the picker hands the host to persist, keys dropped -- they are positional. */
+    private fun savedValues(): List<Any?> = composeRule.runOnIdle { registry.performSave().values.flatten() }
+
+    /**
+     * The single saved value, asserted rather than assumed.
+     *
+     * `single()` on an empty list throws `NoSuchElementException: List is empty`, which names
+     * neither the panel nor the registry -- and an empty registry is exactly how the
+     * `rememberSaveable` -> `remember` regression shows up here.
+     */
+    private fun theOneSavedValue(): Any? {
+        val values = savedValues()
+        assertEquals("the panel should register exactly one saved value", 1, values.size)
+        return values.first()
+    }
+
+    /** A write and a read through a real `Parcel`, which is what the tester's map stands in for. */
+    private fun throughARealBundle(value: Parcelable): Parcelable {
+        val bundle = Bundle().apply { putParcelable(KEY, value) }
+        val parcel = Parcel.obtain()
+        return try {
+            parcel.writeBundle(bundle)
+            parcel.setDataPosition(0)
+            val restored = requireNotNull(parcel.readBundle(javaClass.classLoader)) {
+                "the Bundle did not survive the Parcel"
+            }
+            requireNotNull(restored.getParcelable(KEY, Parcelable::class.java)) {
+                "the saved state did not survive the Parcel"
+            }
+        } finally {
+            parcel.recycle()
+        }
+    }
+
+    private companion object {
+        const val KEY = "expanded"
+    }
+}
@@ -0,0 +1,321 @@
+package org.libremediaconverter.convert
+
+import androidx.compose.ui.test.assertIsDisplayed
+import androidx.compose.ui.test.assertTextEquals
+import androidx.compose.ui.test.hasAnyAncestor
+import androidx.compose.ui.test.hasTestTag
+import androidx.compose.ui.test.hasText
+import androidx.compose.ui.test.junit4.StateRestorationTester
+import androidx.compose.ui.test.onAllNodesWithTag
+import androidx.compose.ui.test.onNodeWithTag
+import androidx.compose.ui.test.onNodeWithText
+import androidx.compose.ui.test.performClick
+import androidx.media3.common.util.UnstableApi
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.createDrainedComposeRule
+import org.libremediaconverter.model.AudioCodec
+import org.libremediaconverter.model.Container
+import org.libremediaconverter.model.ContainerCapabilities
+import org.libremediaconverter.model.InputProbe
+import org.libremediaconverter.model.OutputSpec
+import org.libremediaconverter.model.Validation
+import org.libremediaconverter.model.VideoCodec
+import org.libremediaconverter.ui.TestTags
+import org.robolectric.RobolectricTestRunner
+
+/**
+ * The gate over the Advanced chips, and the error card that deliberately sits outside it.
+ *
+ * Two defects, and they pull in opposite directions.
+ *
+ * The first is the chips escaping the gate, or never being reachable through it. `AdvancedPicker`
+ * is the one leaf on this screen that is not stateless -- `expanded` is its own `rememberSaveable`
+ * -- and Container, Video and Audio live inside `AnimatedVisibility(visible = expanded)`. Nothing
+ * else on the screen hides anything, so a refactor that flattened the panel, or wired the toggle to
+ * a state nobody reads, would render an app that looks reasonable in a screenshot and is wrong.
+ *
+ * The second is the opposite mistake, and it is the one this file exists for: **moving the
+ * `ValidationError` call inside the `AnimatedVisibility`**. It is invoked after that block, so an
+ * invalid spec explains itself and offers one-tap fixes *while the section is collapsed*. That is
+ * the only route out of an invalid spec for a user who never opened Advanced -- and since the only
+ * way to reach an invalid spec is through Advanced, hiding the way out behind the same toggle looks
+ * locally sensible and is a trap. Tidying the two `if` blocks into one is a plausible edit, it
+ * compiles, and until this file existed nothing went red. Every assertion about the error card here
+ * therefore runs with the toggle untouched, and asserts the panel is absent in the same test, so a
+ * future `expanded = true` default cannot quietly satisfy it either.
+ *
+ * The invalid specs come from [ContainerCapabilities.validate] rather than from a hand-built
+ * [Validation.Invalid], so the messages and the suggestions are the real pairing. A hand-built one
+ * would keep passing after `validate` stopped producing anything like it.
+ *
+ * Node location is by the three separate chip-row tags, never by text. `"Copy"` and `"None"` are
+ * each both a [VideoCodec] and an [AudioCodec], and `"MP3"` and `"FLAC"` are each both a
+ * [Container] and an [AudioCodec], so a text matcher over the open panel is ambiguous for four
+ * chips -- which is what the separate tags are for.
+ */
+@UnstableApi
+@RunWith(RobolectricTestRunner::class)
+class AdvancedPickerTest {
+
+    // Not `createComposeRule()` directly: see [drainEscapedCoroutineErrors].
+    @get:Rule
+    val composeRule = createDrainedComposeRule()
+
+    private val restoration = StateRestorationTester(composeRule)
+
+    private val containers = mutableListOf<Container>()
+    private val videoCodecs = mutableListOf<VideoCodec>()
+    private val audioCodecs = mutableListOf<AudioCodec>()
+    private val applied = mutableListOf<OutputSpec>()
+
+    // --- the expand gate ----------------------------------------------------
+
+    @Test
+    fun `the three chip rows appear only while the panel is expanded`() {
+        setPicker()
+
+        assertPanelHidden()
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).performClick()
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_PANEL).assertExists()
+        ROW_TAGS.forEach { composeRule.onNodeWithTag(it).assertExists() }
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).performClick()
+
+        // The exit transition outlives the click, so absence has to be waited for rather than
+        // asserted straight away -- unlike the initial collapsed state, which has no animation
+        // in flight.
+        composeRule.waitUntil { nodeCount(TestTags.Converter.ADVANCED_PANEL) == 0 }
+        assertPanelHidden()
+    }
+
+    /** The toggle is the only affordance the collapsed picker offers, so it has to say so. */
+    @Test
+    fun `the toggle names the direction it will move in`() {
+        setPicker()
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).assertTextEquals("Advanced")
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).performClick()
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE)
+            .assertTextEquals("Hide advanced")
+    }
+
+    /**
+     * The four colliding labels, one per row.
+     *
+     * `"Copy"` is a video codec *and* an audio codec; `"MP3"` is a container *and* an audio codec.
+     * Clicking each through its own row is what proves the rows are wired to different callbacks
+     * -- a picker that handed every chip to `onAudioCodec` would look identical on screen.
+     */
+    @Test
+    fun `each chip row reports to its own callback, including the labels that collide`() {
+        setPicker()
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).performClick()
+
+        chipIn(TestTags.Converter.ADVANCED_VIDEO_CHIPS, "Copy").performClick()
+
+        assertEquals(listOf(VideoCodec.COPY), videoCodecs)
+        assertEquals(emptyList<AudioCodec>(), audioCodecs)
+
+        chipIn(TestTags.Converter.ADVANCED_AUDIO_CHIPS, "Copy").performClick()
+
+        assertEquals(listOf(AudioCodec.COPY), audioCodecs)
+
+        chipIn(TestTags.Converter.ADVANCED_CONTAINER_CHIPS, "MP3").performClick()
+
+        assertEquals(listOf(Container.MP3), containers)
+        // Still only the one audio click. `MP3` is an AudioCodec label too, and the container row
+        // must not be reporting through that callback.
+        assertEquals(listOf(AudioCodec.COPY), audioCodecs)
+    }
+
+    // --- the error card, which is outside the gate --------------------------
+
+    /**
+     * The headline case. Dropping both tracks is reachable from the collapsed screen -- the
+     * `None`/`None` pair is set inside Advanced, but the user can close it again -- and the
+     * explanation has to still be there.
+     */
+    @Test
+    fun `an empty output explains itself while the section is collapsed`() {
+        val spec = OutputSpec(Container.MP4, VideoCodec.NONE, AudioCodec.NONE)
+        val invalid = invalidFor(spec)
+
+        assertEquals("This would produce an empty file — keep at least one track.", invalid.message)
+
+        setPicker(spec, invalid)
+
+        assertPanelHidden()
+        composeRule.onNodeWithTag(TestTags.Converter.VALIDATION_ERROR).assertExists()
+        composeRule.onNodeWithText(invalid.message).assertIsDisplayed()
+    }
+
+    @Test
+    fun `a codec the container cannot hold explains itself while the section is collapsed`() {
+        val spec = OutputSpec(Container.WEBM, VideoCodec.H264, AudioCodec.OPUS)
+        val invalid = invalidFor(spec)
+
+        assertEquals("WebM cannot hold H.264 video.", invalid.message)
+
+        setPicker(spec, invalid)
+
+        assertPanelHidden()
+        composeRule.onNodeWithText(invalid.message).assertIsDisplayed()
+    }
+
+    /**
+     * Clicking a suggestion, with the toggle never touched.
+     *
+     * The second suggestion rather than the first, and its count pinned first: with one suggestion
+     * a picker that handed every chip `suggestions[0]` would pass, and `onNodeWithTag` on a
+     * suggestion index that no longer exists reports an unhelpful matcher failure rather than
+     * saying the list shrank.
+     */
+    @Test
+    fun `a suggestion chip applies its own spec without the section ever being opened`() {
+        val spec = OutputSpec(Container.WEBM, VideoCodec.H264, AudioCodec.OPUS)
+        val invalid = invalidFor(spec)
+
+        assertEquals(2, invalid.suggestions.size)
+        val second = invalid.suggestions[1]
+
+        setPicker(spec, invalid)
+
+        assertPanelHidden()
+        composeRule.onNodeWithTag(TestTags.Converter.suggestion(1)).assertTextEquals(describe(second))
+        composeRule.onNodeWithTag(TestTags.Converter.suggestion(1)).performClick()
+
+        assertEquals(listOf(second), applied)
+        // What the chips offer is what `validate` said would work, not a repair of the test's own.
+        assertTrue(
+            "suggestion $second should itself validate",
+            ContainerCapabilities.validate(second, PROBE).isValid,
+        )
+    }
+
+    /** A valid spec has nothing to say, collapsed or not. */
+    @Test
+    fun `a valid spec renders no error card`() {
+        setPicker()
+
+        composeRule.onNodeWithTag(TestTags.Converter.VALIDATION_ERROR).assertDoesNotExist()
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).performClick()
+
+        composeRule.onNodeWithTag(TestTags.Converter.VALIDATION_ERROR).assertDoesNotExist()
+    }
+
+    // --- recreation ---------------------------------------------------------
+
+    /**
+     * `expanded` is the only `rememberSaveable` on either screen's leaves.
+     *
+     * `MainActivity` declares no `configChanges`, so a rotation destroys and rebuilds the whole
+     * composition. A panel the user opened, set three chips in, and left open must not close
+     * itself on the way back. `remember` would.
+     *
+     * What this cannot see is the saved *representation* -- `StateRestorationTester` saves into an
+     * in-memory map rather than a `Bundle`. `AdvancedPanelSavedStateTest` covers that half.
+     */
+    @Test
+    fun `an open panel is still open after recreation`() {
+        restoration.setContent {
+            AdvancedPicker(
+                spec = VALID_SPEC,
+                validation = Validation.Valid,
+                onContainer = {},
+                onVideoCodec = {},
+                onAudioCodec = {},
+                onSuggestion = {},
+            )
+        }
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).performClick()
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_PANEL).assertExists()
+
+        restoration.emulateSavedInstanceStateRestore()
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_PANEL).assertExists()
+        ROW_TAGS.forEach { composeRule.onNodeWithTag(it).assertExists() }
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE)
+            .assertTextEquals("Hide advanced")
+    }
+
+    /** The default has to survive too, or the panel would spring open on every rotation. */
+    @Test
+    fun `a collapsed panel is still collapsed after recreation`() {
+        restoration.setContent {
+            AdvancedPicker(
+                spec = VALID_SPEC,
+                validation = Validation.Valid,
+                onContainer = {},
+                onVideoCodec = {},
+                onAudioCodec = {},
+                onSuggestion = {},
+            )
+        }
+
+        assertPanelHidden()
+
+        restoration.emulateSavedInstanceStateRestore()
+
+        assertPanelHidden()
+    }
+
+    // --- helpers ------------------------------------------------------------
+
+    private fun setPicker(spec: OutputSpec = VALID_SPEC, validation: Validation = Validation.Valid) {
+        composeRule.setContent {
+            AdvancedPicker(
+                spec = spec,
+                validation = validation,
+                onContainer = { containers += it },
+                onVideoCodec = { videoCodecs += it },
+                onAudioCodec = { audioCodecs += it },
+                onSuggestion = { applied += it },
+            )
+        }
+    }
+
+    /** The whole panel, by every tag it owns, so a partial escape counts as a failure. */
+    private fun assertPanelHidden() {
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_PANEL).assertDoesNotExist()
+        ROW_TAGS.forEach { composeRule.onNodeWithTag(it).assertDoesNotExist() }
+    }
+
+    private fun nodeCount(tag: String) = composeRule.onAllNodesWithTag(tag).fetchSemanticsNodes().size
+
+    private fun chipIn(rowTag: String, label: String) =
+        composeRule.onNode(hasText(label) and hasAnyAncestor(hasTestTag(rowTag)))
+
+    private fun invalidFor(spec: OutputSpec): Validation.Invalid {
+        val validation = ContainerCapabilities.validate(spec, PROBE)
+        return validation as? Validation.Invalid
+            ?: throw AssertionError("$spec was expected to be invalid, but validate said $validation")
+    }
+
+    private companion object {
+        val ROW_TAGS = listOf(
+            TestTags.Converter.ADVANCED_CONTAINER_CHIPS,
+            TestTags.Converter.ADVANCED_VIDEO_CHIPS,
+            TestTags.Converter.ADVANCED_AUDIO_CHIPS,
+        )
+
+        val VALID_SPEC = OutputSpec(Container.MP4, VideoCodec.H264, AudioCodec.AAC)
+
+        /** An ordinary H.264/AAC MP4, so the suggestions have a real source to repair towards. */
+        val PROBE = InputProbe(
+            videoCodec = "h264",
+            audioCodec = "aac",
+            durationMs = 90_000,
+            container = Container.MP4,
+        )
+    }
+}
@@ -0,0 +1,152 @@
+package org.libremediaconverter.convert
+
+import org.junit.Assert.assertEquals
+import org.junit.Test
+import org.libremediaconverter.model.AudioCodec
+import org.libremediaconverter.model.Container
+import org.libremediaconverter.model.EnginePreference
+import org.libremediaconverter.model.OutputSpec
+import org.libremediaconverter.model.VideoCodec
+
+/**
+ * The four pure helpers behind the converter screen's prose, pinned at the points where they
+ * change what they say.
+ *
+ * No Compose rule and no Robolectric: these are `String` in, `String` out, and running them under a
+ * device sandbox would buy nothing while hiding the boundaries in a rendered tree.
+ *
+ * The defect each group bites on:
+ *
+ * - **[formatBytes] picks a unit by comparing against three thresholds.** Every one of them is a
+ *   `>=`, and a `>` would move a file sitting exactly on a boundary into the unit below -- `1 GB`
+ *   shown as `1000.0 MB`. Only a value *on* the threshold can tell the two apart, so each of the
+ *   three is asserted at the boundary and one below it. The unit prefixes are decimal, matching
+ *   what the file manager and the provider report, not powers of two.
+ * - **[formatDuration] has no hours field.** An hour-long recording reads `60:00`, and that is the
+ *   contract rather than an oversight -- the row is a length, not a clock. Pinned so that adding
+ *   hours is a deliberate change with a red test in front of it instead of a silent reformat.
+ * - **[describe] builds the suggestion-chip label out of up to three parts**, and the parts are
+ *   conditional: [VideoCodec.NONE] and [AudioCodec.NONE] drop out entirely, so an image output
+ *   with neither track has to render as the container alone rather than as a container followed
+ *   by a dangling separator.
+ * - **[EnginePreference] carries no `label` property**, unlike every other enum the screen
+ *   renders; its three display strings live in a `when` in the screen file. Adding a constant is
+ *   caught by the compiler because that `when` is exhaustive, but nothing stops two constants
+ *   being given the same string, which is what the distinctness assertion is for.
+ */
+class ConverterFormattersTest {
+
+    @Test
+    fun `bytes below a kilobyte are counted exactly`() {
+        assertEquals("0 B", formatBytes(0))
+        assertEquals("1 B", formatBytes(1))
+        assertEquals("999 B", formatBytes(999))
+    }
+
+    @Test
+    fun `each unit starts exactly on its threshold rather than one byte past it`() {
+        assertEquals("1 kB", formatBytes(1_000))
+        assertEquals("1.0 MB", formatBytes(1_000_000))
+        assertEquals("1.0 GB", formatBytes(1_000_000_000))
+    }
+
+    /**
+     * One byte below each threshold, which is the half a `>=` to `>` change leaves alone. Both
+     * halves are needed: the boundary values alone would still pass if the comparison let
+     * everything through.
+     */
+    @Test
+    fun `a value just below a threshold stays in the smaller unit`() {
+        assertEquals("999 B", formatBytes(999))
+        assertEquals("1000 kB", formatBytes(999_999))
+        assertEquals("1000.0 MB", formatBytes(999_999_999))
+    }
+
+    @Test
+    fun `a real file size reads as one decimal place`() {
+        assertEquals("12.3 MB", formatBytes(12_345_678))
+        assertEquals("1.5 GB", formatBytes(1_500_000_000))
+    }
+
+    @Test
+    fun `a duration is minutes and zero-padded seconds`() {
+        assertEquals("0:00", formatDuration(0))
+        assertEquals("0:01", formatDuration(1_000))
+        assertEquals("0:59", formatDuration(59_000))
+        assertEquals("1:00", formatDuration(60_000))
+        assertEquals("1:30", formatDuration(90_000))
+    }
+
+    /** Sub-second remainders are dropped rather than rounded up into the next second. */
+    @Test
+    fun `a partial second does not become a whole one`() {
+        assertEquals("0:00", formatDuration(999))
+        assertEquals("0:59", formatDuration(59_999))
+    }
+
+    /** No hours field, deliberately: an hour is `60:00` and two hours are `120:00`. */
+    @Test
+    fun `an hour and beyond keeps counting in minutes`() {
+        assertEquals("60:00", formatDuration(3_600_000))
+        assertEquals("61:01", formatDuration(3_661_000))
+        assertEquals("120:00", formatDuration(7_200_000))
+    }
+
+    @Test
+    fun `a spec with both tracks names the container and joins the two codecs`() {
+        assertEquals(
+            "MP4 · H.264 + AAC",
+            describe(OutputSpec(Container.MP4, VideoCodec.H264, AudioCodec.AAC)),
+        )
+    }
+
+    @Test
+    fun `a track set to none is left out instead of being named none`() {
+        assertEquals(
+            "MP3 · MP3",
+            describe(OutputSpec(Container.MP3, VideoCodec.NONE, AudioCodec.MP3)),
+        )
+        assertEquals(
+            "MP4 · H.264",
+            describe(OutputSpec(Container.MP4, VideoCodec.H264, AudioCodec.NONE)),
+        )
+    }
+
+    /** An image output has neither track, so there is nothing for the separator to separate. */
+    @Test
+    fun `a spec with no tracks at all is the container alone, with no trailing separator`() {
+        assertEquals("GIF", describe(OutputSpec(Container.GIF, VideoCodec.NONE, AudioCodec.NONE)))
+        assertEquals(
+            "PNG frames",
+            describe(OutputSpec(Container.IMAGE_SEQUENCE, VideoCodec.NONE, AudioCodec.NONE)),
+        )
+    }
+
+    /** `Copy` is a codec here, not the absence of one, so a remux describes both tracks. */
+    @Test
+    fun `a remux names copy on both tracks rather than dropping them`() {
+        assertEquals(
+            "Matroska · Copy + Copy",
+            describe(OutputSpec(Container.MKV, VideoCodec.COPY, AudioCodec.COPY)),
+        )
+    }
+
+    @Test
+    fun `each engine preference has the wording the chips show`() {
+        assertEquals("Automatic", EnginePreference.AUTO.label())
+        assertEquals("Prefer hardware", EnginePreference.PREFER_HARDWARE.label())
+        assertEquals("Force software", EnginePreference.FORCE_SOFTWARE.label())
+    }
+
+    /**
+     * Two constants sharing a label would render as two identical chips, one of which the user
+     * could not choose deliberately. The exhaustive `when` cannot catch that; this does.
+     */
+    @Test
+    fun `no two engine preferences render the same chip`() {
+        val labels = EnginePreference.entries.map { it.label() }
+
+        assertEquals(EnginePreference.entries.size, labels.toSet().size)
+        assertEquals(emptyList<String>(), labels.filter { it.isBlank() })
+    }
+}
@@ -0,0 +1,196 @@
+package org.libremediaconverter.convert
+
+import android.net.Uri
+import androidx.compose.ui.test.assertCountEquals
+import androidx.compose.ui.test.onAllNodesWithTag
+import androidx.compose.ui.test.onNodeWithTag
+import androidx.compose.ui.test.performClick
+import androidx.media3.common.util.UnstableApi
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.createDrainedComposeRule
+import org.libremediaconverter.model.AudioCodec
+import org.libremediaconverter.model.Container
+import org.libremediaconverter.model.EnginePreference
+import org.libremediaconverter.model.InputKind
+import org.libremediaconverter.model.InputProbe
+import org.libremediaconverter.model.OutputFormat
+import org.libremediaconverter.model.OutputSpec
+import org.libremediaconverter.model.QualityTier
+import org.libremediaconverter.model.Validation
+import org.libremediaconverter.model.VideoCodec
+import org.libremediaconverter.ui.TestTags
+import org.robolectric.RobolectricTestRunner
+
+/**
+ * Each leaf of the converter screen renders, and each tag it claims resolves to exactly one node.
+ *
+ * The defect this bites on is a tag that is not where the table says it is: dropped by a refactor
+ * that rewrote a `Modifier` chain, applied to the wrong one of two siblings, or duplicated onto a
+ * leaf that is rendered twice. None of that is visible at compile time -- a `testTag` is a string
+ * handed to a modifier -- and none of it shows up in the app either, because nothing but a test
+ * ever reads one.
+ *
+ * It has to be caught here rather than by the children that consume the tags. R38.2, R38.3 and
+ * R38.4 all *begin* by locating a node through one of these, so a tag that had quietly moved would
+ * surface as three unrelated PRs failing on a line their own diffs do not touch. Counting the nodes
+ * rather than asserting existence is deliberate: `onNodeWithTag` on two matches throws about
+ * ambiguity in one place and passes in another, so "exactly one" is the property worth pinning.
+ *
+ * Deliberately *not* the state matrix. Which affordances each `ConversionState` renders is R38.6,
+ * and it needs the state seam R38.5 extracts -- the branch buttons tagged in this change (Convert,
+ * Cancel, Save file, Start over, ...) therefore have no bite yet, which the PR body records.
+ */
+@UnstableApi
+@RunWith(RobolectricTestRunner::class)
+class ConverterLeafTagsTest {
+
+    @get:Rule
+    val composeRule = createDrainedComposeRule()
+
+    private fun assertResolvesToOneNode(tag: String) {
+        composeRule.onAllNodesWithTag(tag).assertCountEquals(1)
+    }
+
+    private fun input(sizeBytes: Long? = 12_345_678L, probe: InputProbe? = VIDEO_PROBE) = InputFile(
+        uri = Uri.parse("content://test/clip.mkv"),
+        displayName = "clip.mkv",
+        sizeBytes = sizeBytes,
+        probe = probe,
+    )
+
+    @Test
+    fun `the format picker tags its chip row`() {
+        composeRule.setContent { FormatPicker(OutputFormat.MP4_H264) {} }
+
+        assertResolvesToOneNode(TestTags.Converter.FORMAT_CHIPS)
+    }
+
+    @Test
+    fun `the quality picker tags its chip row`() {
+        composeRule.setContent { QualityPicker(QualityTier.FAST) {} }
+
+        assertResolvesToOneNode(TestTags.Converter.QUALITY_CHIPS)
+    }
+
+    @Test
+    fun `the engine picker tags its chip row`() {
+        composeRule.setContent { EnginePicker(EnginePreference.AUTO) {} }
+
+        assertResolvesToOneNode(TestTags.Converter.ENGINE_CHIPS)
+    }
+
+    @Test
+    fun `the advanced picker tags its toggle, which is all it renders while collapsed`() {
+        setAdvancedPicker()
+
+        assertResolvesToOneNode(TestTags.Converter.ADVANCED_TOGGLE)
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_PANEL).assertDoesNotExist()
+    }
+
+    /**
+     * The panel and its three rows only exist once the toggle has been clicked, which is R38.4's
+     * subject. Expanding is the only way to reach the tags at all, so the smoke test has to do it.
+     */
+    @Test
+    fun `expanding the advanced picker tags the panel and each of its three chip rows`() {
+        setAdvancedPicker()
+
+        composeRule.onNodeWithTag(TestTags.Converter.ADVANCED_TOGGLE).performClick()
+
+        assertResolvesToOneNode(TestTags.Converter.ADVANCED_PANEL)
+        assertResolvesToOneNode(TestTags.Converter.ADVANCED_CONTAINER_CHIPS)
+        assertResolvesToOneNode(TestTags.Converter.ADVANCED_VIDEO_CHIPS)
+        assertResolvesToOneNode(TestTags.Converter.ADVANCED_AUDIO_CHIPS)
+    }
+
+    @Test
+    fun `the validation card tags itself and every suggestion on it`() {
+        composeRule.setContent {
+            ValidationError(
+                Validation.Invalid(
+                    message = "WebM cannot hold H.264 video.",
+                    suggestions = listOf(
+                        OutputSpec(Container.MKV, VideoCodec.H264, AudioCodec.AAC),
+                        OutputSpec(Container.WEBM, VideoCodec.VP9, AudioCodec.OPUS),
+                    ),
+                ),
+            ) {}
+        }
+
+        assertResolvesToOneNode(TestTags.Converter.VALIDATION_ERROR)
+        assertResolvesToOneNode(TestTags.Converter.suggestion(0))
+        assertResolvesToOneNode(TestTags.Converter.suggestion(1))
+    }
+
+    @Test
+    fun `the file card tags itself, its name and its size line`() {
+        composeRule.setContent { FileCard(input()) }
+
+        assertResolvesToOneNode(TestTags.Converter.FILE_CARD)
+        assertResolvesToOneNode(TestTags.Converter.FILE_CARD_NAME)
+        assertResolvesToOneNode(TestTags.Converter.FILE_CARD_BYTES)
+    }
+
+    /**
+     * Both writers of the note line get their own case. They are two separate `Text` calls in two
+     * branches that share one tag, so a test of either alone would leave the other unguarded.
+     */
+    @Test
+    fun `the file card tags the note it shows while the probe is still running`() {
+        composeRule.setContent { FileCard(input(probe = null)) }
+
+        assertResolvesToOneNode(TestTags.Converter.FILE_CARD_NOTE)
+    }
+
+    @Test
+    fun `the file card tags the note it shows when nothing could read the file`() {
+        composeRule.setContent { FileCard(input(probe = InputProbe(kind = InputKind.UNPARSEABLE))) }
+
+        assertResolvesToOneNode(TestTags.Converter.FILE_CARD_NOTE)
+    }
+
+    @Test
+    fun `a detail row tags itself with the label it renders`() {
+        composeRule.setContent { DetailRow("Container", "Matroska") }
+
+        assertResolvesToOneNode(TestTags.Converter.detailRow("Container"))
+    }
+
+    /** The rows the file card builds carry the same per-label tags, one per row it renders. */
+    @Test
+    fun `the file card's detail rows are each tagged by their own label`() {
+        composeRule.setContent { FileCard(input()) }
+
+        assertResolvesToOneNode(TestTags.Converter.detailRow("Container"))
+        assertResolvesToOneNode(TestTags.Converter.detailRow("Video"))
+        assertResolvesToOneNode(TestTags.Converter.detailRow("Audio"))
+        assertResolvesToOneNode(TestTags.Converter.detailRow("Length"))
+    }
+
+    private fun setAdvancedPicker() {
+        composeRule.setContent {
+            AdvancedPicker(
+                spec = OutputSpec(Container.MP4, VideoCodec.H264, AudioCodec.AAC),
+                validation = Validation.Valid,
+                onContainer = {},
+                onVideoCodec = {},
+                onAudioCodec = {},
+                onSuggestion = {},
+            )
+        }
+    }
+
+    private companion object {
+        val VIDEO_PROBE = InputProbe(
+            videoCodec = "video/avc",
+            audioCodec = "audio/mp4a-latm",
+            durationMs = 90_000,
+            kind = InputKind.VIDEO,
+            container = Container.MKV,
+            width = 1920,
+            height = 1080,
+        )
+    }
+}
@@ -0,0 +1,169 @@
+package org.libremediaconverter.convert
+
+import androidx.compose.ui.test.SemanticsNodeInteraction
+import androidx.compose.ui.test.assertIsNotSelected
+import androidx.compose.ui.test.assertIsSelected
+import androidx.compose.ui.test.hasAnyAncestor
+import androidx.compose.ui.test.hasTestTag
+import androidx.compose.ui.test.hasText
+import androidx.compose.ui.test.onNodeWithText
+import androidx.compose.ui.test.performClick
+import androidx.media3.common.util.UnstableApi
+import org.junit.Assert.assertEquals
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.createDrainedComposeRule
+import org.libremediaconverter.model.EnginePreference
+import org.libremediaconverter.model.OutputFormat
+import org.libremediaconverter.model.QualityTier
+import org.libremediaconverter.ui.TestTags
+import org.robolectric.RobolectricTestRunner
+
+/**
+ * Each picker lights the chip it was handed and reports the constant that was pressed.
+ *
+ * The defect this bites on is a picker that renders perfectly and answers wrongly. All three are
+ * the same dozen lines with a different enum substituted, so the failure mode is a copy-paste that
+ * survives review: an `onClick` that closes over the picker's `selected` parameter instead of the
+ * chip's own entry hands back one constant no matter which chip was tapped, and an inverted
+ * `entry == selected` lights every chip except the right one. Neither throws, neither changes the
+ * set of labels on screen, and a test that only asserted "the callback ran" would pass over both.
+ *
+ * Clicking every chip in turn and comparing the whole recorded list against `entries` is what makes
+ * the constant load-bearing rather than the click count -- a hardcoded `onSelect` fires the same
+ * number of times as a correct one. Selection is asserted over every chip for the same reason: the
+ * one that should be lit proves nothing on its own, because `!=` lights it too whenever the enum
+ * has exactly one entry, and lights all its siblings whenever it has more.
+ *
+ * Labels come from `OutputFormat.label` and `QualityTier.label`; [label], which the screen owns
+ * because `EnginePreference` carries no label of its own, supplies the third set. Retyping any of
+ * them here would turn a rename into a red test that named the wrong cause.
+ *
+ * Not covered, deliberately: the `"Output format"`, `"Quality"` and `"Engine"` headings, which are
+ * untagged `Text` calls with no enum behind them and no behaviour to bite on.
+ */
+@UnstableApi
+@RunWith(RobolectricTestRunner::class)
+class ConverterPickerSelectionTest {
+
+    @get:Rule
+    val composeRule = createDrainedComposeRule()
+
+    /**
+     * The chip carrying [label] inside the row tagged [rowTag].
+     *
+     * By ancestor rather than by direct child: how many semantics nodes Material 3 puts between a
+     * `FlowRow` and its chips is that library's business, and a matcher that assumed "one" would
+     * break on an upgrade that changed nothing this test is about.
+     */
+    private fun chipIn(rowTag: String, label: String): SemanticsNodeInteraction =
+        composeRule.onNode(hasAnyAncestor(hasTestTag(rowTag)) and hasText(label))
+
+    private fun assertOnlySelected(rowTag: String, labels: List<String>, selected: String?) {
+        labels.forEach { label ->
+            val chip = chipIn(rowTag, label)
+            if (label == selected) chip.assertIsSelected() else chip.assertIsNotSelected()
+        }
+    }
+
+    @Test
+    fun `the format picker lights the selected format and no other`() {
+        composeRule.setContent { FormatPicker(OutputFormat.WEBM_VP9) {} }
+
+        assertOnlySelected(
+            rowTag = TestTags.Converter.FORMAT_CHIPS,
+            labels = OutputFormat.entries.map { it.label },
+            selected = OutputFormat.WEBM_VP9.label,
+        )
+    }
+
+    /** A spec no preset can express lights nothing, which is what the custom line stands in for. */
+    @Test
+    fun `the format picker lights nothing when the spec is custom`() {
+        composeRule.setContent { FormatPicker(null) {} }
+
+        assertOnlySelected(
+            rowTag = TestTags.Converter.FORMAT_CHIPS,
+            labels = OutputFormat.entries.map { it.label },
+            selected = null,
+        )
+        composeRule.onNodeWithText(CUSTOM_SPEC_NOTE).assertExists()
+    }
+
+    @Test
+    fun `a selected format hides the custom line`() {
+        composeRule.setContent { FormatPicker(OutputFormat.MP3) {} }
+
+        composeRule.onNodeWithText(CUSTOM_SPEC_NOTE).assertDoesNotExist()
+    }
+
+    @Test
+    fun `clicking a format chip reports that format`() {
+        val picked = mutableListOf<OutputFormat>()
+        composeRule.setContent { FormatPicker(null) { picked += it } }
+
+        OutputFormat.entries.forEach { chipIn(TestTags.Converter.FORMAT_CHIPS, it.label).performClick() }
+
+        assertEquals(OutputFormat.entries.toList(), picked)
+    }
+
+    @Test
+    fun `the quality picker lights the selected tier and no other`() {
+        composeRule.setContent { QualityPicker(QualityTier.BEST) {} }
+
+        assertOnlySelected(
+            rowTag = TestTags.Converter.QUALITY_CHIPS,
+            labels = QualityTier.entries.map { it.label },
+            selected = QualityTier.BEST.label,
+        )
+    }
+
+    /** The line under the chips describes what was chosen, not whichever tier was written first. */
+    @Test
+    fun `the quality picker explains the tier that is selected`() {
+        composeRule.setContent { QualityPicker(QualityTier.BEST) {} }
+
+        composeRule.onNodeWithText(QualityTier.BEST.description).assertExists()
+        composeRule.onNodeWithText(QualityTier.FAST.description).assertDoesNotExist()
+    }
+
+    @Test
+    fun `clicking a quality chip reports that tier`() {
+        val picked = mutableListOf<QualityTier>()
+        composeRule.setContent { QualityPicker(QualityTier.FAST) { picked += it } }
+
+        QualityTier.entries.forEach { chipIn(TestTags.Converter.QUALITY_CHIPS, it.label).performClick() }
+
+        assertEquals(QualityTier.entries.toList(), picked)
+    }
+
+    @Test
+    fun `the engine picker lights the selected preference and no other`() {
+        composeRule.setContent { EnginePicker(EnginePreference.FORCE_SOFTWARE) {} }
+
+        assertOnlySelected(
+            rowTag = TestTags.Converter.ENGINE_CHIPS,
+            labels = EnginePreference.entries.map { it.label() },
+            selected = EnginePreference.FORCE_SOFTWARE.label(),
+        )
+    }
+
+    @Test
+    fun `clicking an engine chip reports that preference`() {
+        val picked = mutableListOf<EnginePreference>()
+        composeRule.setContent { EnginePicker(EnginePreference.AUTO) { picked += it } }
+
+        EnginePreference.entries.forEach { chipIn(TestTags.Converter.ENGINE_CHIPS, it.label()).performClick() }
+
+        assertEquals(EnginePreference.entries.toList(), picked)
+    }
+
+    private companion object {
+        /**
+         * Copied byte for byte out of `ConverterScreen.kt` -- it holds a U+2014 em dash, which
+         * retyped as ASCII would match nothing and fail as "no node found" rather than as a reword.
+         */
+        const val CUSTOM_SPEC_NOTE: String = "Custom — set below."
+    }
+}
@@ -0,0 +1,254 @@
+package org.libremediaconverter.convert
+
+import android.net.Uri
+import androidx.compose.ui.test.assertCountEquals
+import androidx.compose.ui.test.assertTextEquals
+import androidx.compose.ui.test.onChildren
+import androidx.compose.ui.test.onNodeWithTag
+import androidx.media3.common.util.UnstableApi
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.createDrainedComposeRule
+import org.libremediaconverter.model.AudioCodec
+import org.libremediaconverter.model.Container
+import org.libremediaconverter.model.InputKind
+import org.libremediaconverter.model.InputProbe
+import org.libremediaconverter.model.VideoCodec
+import org.libremediaconverter.ui.TestTags
+import org.robolectric.RobolectricTestRunner
+
+/**
+ * What the source-info card says when it does not know something.
+ *
+ * The defect is a card that invents an answer instead of admitting it has none. Two of them are
+ * live here and neither had a test before this file:
+ *
+ * - **`InputFile.sizeBytes` is nullable and the card is the reader that has to say so in words.**
+ *   `sizeBytes` used to be `0L` for "nobody told me", and [UnknownInputSizeTest] records what that
+ *   cost at the space check. The card is the other reader, and its failure mode is the mirror
+ *   image: hand the null to `formatBytes` and it renders `"0 B"` -- a measurement, shown to the
+ *   user, that no provider ever made. It renders **independently of the probe**, which is why the
+ *   same assertion appears twice below, with the probe present and absent. That independence is
+ *   the contract; a test covering only the probed case would leave the branch a user actually hits
+ *   first -- the card is on screen before the probe finishes -- unguarded.
+ * - **The codec rows degrade in words too.** `CodecNames.describeVideo`/`describeAudio` answer
+ *   `"Unknown"` for a codec nothing named, the `VIDEO` branch answers `"No audio track"` for a file
+ *   with no audio, and the two `> 0` guards drop the dimension and length rows rather than printing
+ *   `0` and `0:00`. Each of those has a case below on **both** sides of the guard, because a test
+ *   of the present side alone stays green with the guard deleted.
+ *
+ * ### What cannot be asserted here, so that it is a decision rather than an omission
+ *
+ * The `probe == null` branch exits before `HorizontalDivider`, and **the divider's absence is not
+ * observable from a test**: Material 3 renders it as a `Box` with no semantics modifier, so it
+ * contributes no node to the semantics tree at all. What is asserted instead is everything the
+ * divider precedes -- no detail row for any label the four kind branches can emit -- plus the
+ * card's child count, which pins "these three texts and nothing else" without having to enumerate.
+ *
+ * The early exit itself is enforced by the compiler rather than by this file, which the PR body
+ * records: deleting `return@Column` un-smart-casts `probe`, and the `probe.kind` below it stops
+ * compiling. The mutation that reddens the test here is the compilable form of that regression --
+ * defaulting the null away with `?: InputProbe()` and letting the kind rows render.
+ */
+@UnstableApi
+@RunWith(RobolectricTestRunner::class)
+class FileCardTest {
+
+    @get:Rule
+    val composeRule = createDrainedComposeRule()
+
+    @Test
+    fun `a file no provider could measure says so in words rather than showing a zero`() {
+        setFileCard(input(sizeBytes = null, probe = VIDEO_PROBE))
+
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD_BYTES)
+            .assertTextEquals("Size unknown")
+    }
+
+    /**
+     * The same line, with no probe at all. Separate from the case above rather than folded into
+     * it because `setContent` may only be called once per rule, and because two independent reds
+     * are the evidence that the size line does not depend on the probe.
+     */
+    @Test
+    fun `the size line says the same thing while the probe is still running`() {
+        setFileCard(input(sizeBytes = null, probe = null))
+
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD_BYTES)
+            .assertTextEquals("Size unknown")
+    }
+
+    @Test
+    fun `a size that was reported is formatted rather than replaced by the unknown line`() {
+        setFileCard(input(sizeBytes = 12_345_678L, probe = VIDEO_PROBE))
+
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD_NAME).assertTextEquals("clip.mkv")
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD_BYTES).assertTextEquals("12.3 MB")
+    }
+
+    /**
+     * The note and the emptiness are one behaviour, so they are one test: a regression that keeps
+     * the note but renders the rows anyway would leave a note-only test green.
+     */
+    @Test
+    fun `while the probe is still running the card shows the reading note and nothing else`() {
+        setFileCard(input(probe = null))
+
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD_NOTE)
+            .assertTextEquals("Reading…")
+        assertNoDetailRows()
+        // Name, size, note. Catches a row whose label is not in EVERY_ROW_LABEL as well.
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD).onChildren().assertCountEquals(3)
+    }
+
+    @Test
+    fun `a file nothing could read gets the explanatory line instead of unknown codecs`() {
+        setFileCard(input(probe = InputProbe(kind = InputKind.UNPARSEABLE)))
+
+        composeRule.onNodeWithTag(TestTags.Converter.FILE_CARD_NOTE)
+            .assertTextEquals("Could not identify this file. It will be converted with FFmpeg.")
+        assertNoDetailRows()
+    }
+
+    @Test
+    fun `an image gets its type and its pixel dimensions`() {
+        setFileCard(input(probe = InputProbe(kind = InputKind.IMAGE, width = 1920, height = 1080)))
+
+        assertRow("Type", "Image")
+        assertRow("Size", "1920×1080")
+    }
+
+    /** The `width > 0` guard, from the side that would print `0×0` if it were dropped. */
+    @Test
+    fun `an image whose dimensions nothing reported gets the type row alone`() {
+        setFileCard(input(probe = InputProbe(kind = InputKind.IMAGE)))
+
+        assertRow("Type", "Image")
+        assertNoRow("Size")
+    }
+
+    @Test
+    fun `an audio-only file says it has no video track rather than leaving the row blank`() {
+        setFileCard(
+            input(
+                probe = InputProbe(
+                    audioCodec = "aac",
+                    hasVideo = false,
+                    durationMs = 90_000,
+                    kind = InputKind.AUDIO_ONLY,
+                    container = Container.MP3,
+                ),
+            ),
+        )
+
+        assertRow("Container", Container.MP3.label)
+        assertRow("Video", "No video track")
+        assertRow("Audio", AudioCodec.AAC.label)
+        assertRow("Length", "1:30")
+        assertNoRow("Type")
+        assertNoRow("Size")
+    }
+
+    /**
+     * Everything the audio branch can fail to know, at once: no container, no codec name, no
+     * duration. Each degrades in its own words, and the length row disappears rather than
+     * claiming `0:00`.
+     */
+    @Test
+    fun `an audio-only file nothing else could describe degrades one row at a time`() {
+        setFileCard(input(probe = InputProbe(hasVideo = false, kind = InputKind.AUDIO_ONLY)))
+
+        assertRow("Container", "Unknown")
+        assertRow("Video", "No video track")
+        assertRow("Audio", "Unknown")
+        assertNoRow("Length")
+    }
+
+    @Test
+    fun `a video file composes its codec with its dimensions on one row`() {
+        setFileCard(input(probe = VIDEO_PROBE))
+
+        assertRow("Container", Container.MP4.label)
+        assertRow("Video", "${VideoCodec.H264.label} · 1920×1080")
+        assertRow("Audio", AudioCodec.AAC.label)
+        assertRow("Length", "1:30")
+    }
+
+    /**
+     * `"No audio track"` rather than `describeAudio(null)`'s `"Unknown"`. The video branch knows
+     * the difference between a track it could not name and a track that is not there; the audio
+     * branch above cannot, because a file with no audio is not audio-only.
+     */
+    @Test
+    fun `a video file with no audio track says so instead of naming an unknown codec`() {
+        setFileCard(input(probe = VIDEO_PROBE.copy(audioCodec = null)))
+
+        assertRow("Audio", "No audio track")
+    }
+
+    /** Both `> 0` guards on the video branch, plus the codec name nothing supplied. */
+    @Test
+    fun `a video file missing its codec, dimensions and duration omits them rather than faking them`() {
+        setFileCard(
+            input(
+                probe = VIDEO_PROBE.copy(
+                    videoCodec = null,
+                    width = 0,
+                    height = 0,
+                    durationMs = 0,
+                ),
+            ),
+        )
+
+        assertRow("Video", "Unknown")
+        assertNoRow("Length")
+    }
+
+    /**
+     * The row is one node, not a label node beside a value node. A test matching on `"Container"`
+     * alone would pass against either shape.
+     */
+    @Test
+    fun `a detail row renders its label and its value as a single node`() {
+        composeRule.setContent { DetailRow("Container", "Matroska") }
+
+        composeRule.onNodeWithTag(TestTags.Converter.detailRow("Container"))
+            .assertTextEquals("Container: Matroska")
+    }
+
+    private fun setFileCard(input: InputFile) = composeRule.setContent { FileCard(input) }
+
+    private fun input(sizeBytes: Long? = 12_345_678L, probe: InputProbe? = VIDEO_PROBE) = InputFile(
+        uri = Uri.parse("content://test/clip.mkv"),
+        displayName = "clip.mkv",
+        sizeBytes = sizeBytes,
+        probe = probe,
+    )
+
+    private fun assertRow(label: String, value: String) {
+        composeRule.onNodeWithTag(TestTags.Converter.detailRow(label))
+            .assertTextEquals("$label: $value")
+    }
+
+    private fun assertNoRow(label: String) {
+        composeRule.onNodeWithTag(TestTags.Converter.detailRow(label)).assertDoesNotExist()
+    }
+
+    private fun assertNoDetailRows() = EVERY_ROW_LABEL.forEach(::assertNoRow)
+
+    private companion object {
+        /** Every label the four kind branches can emit, so absence can be asserted exhaustively. */
+        val EVERY_ROW_LABEL = listOf("Container", "Video", "Audio", "Length", "Type", "Size")
+
+        val VIDEO_PROBE = InputProbe(
+            videoCodec = "h264",
+            audioCodec = "aac",
+            durationMs = 90_000,
+            kind = InputKind.VIDEO,
+            container = Container.MP4,
+            width = 1920,
+            height = 1080,
+        )
+    }
+}
@@ -0,0 +1,53 @@
+package org.libremediaconverter.join
+
+import android.net.Uri
+import androidx.compose.ui.test.assertCountEquals
+import androidx.compose.ui.test.onAllNodesWithTag
+import androidx.media3.common.util.UnstableApi
+import org.junit.Rule
+import org.junit.Test
+import org.junit.runner.RunWith
+import org.libremediaconverter.convert.InputFile
+import org.libremediaconverter.createDrainedComposeRule
+import org.libremediaconverter.ui.TestTags
+import org.robolectric.RobolectricTestRunner
+
+/**
+ * The join screen's one leaf renders, and tags itself with the file it is showing.
+ *
+ * `FileRow` is the only place on either screen where the same leaf is rendered more than once at a
+ * time -- one row per picked input -- so it is the only tag that cannot be a constant. It is
+ * derived from `displayName`, inside `FileRow` itself, and that is the part worth a test: a row
+ * that took its tag from the call site would let R38.7 pass a tag in and assert nothing, which is
+ * the vacuous shape `CLAUDE.md` records nine of in one review.
+ *
+ * Two rows are rendered here rather than one, because a tag derived from the wrong thing -- a
+ * constant, an index the row does not have -- would still resolve to one node with a single input
+ * on screen.
+ *
+ * Deliberately not the state matrix: which affordances each `JoinState` renders is R38.7.
+ */
+@UnstableApi
+@RunWith(RobolectricTestRunner::class)
+class JoinLeafTagsTest {
+
+    @get:Rule
+    val composeRule = createDrainedComposeRule()
+
+    private fun input(displayName: String) = InputFile(
+        uri = Uri.parse("content://test/$displayName"),
+        displayName = displayName,
+        sizeBytes = 4_000_000L,
+    )
+
+    @Test
+    fun `each file row is tagged with the name it displays`() {
+        composeRule.setContent {
+            FileRow(input("first.mp4"))
+            FileRow(input("second.mp4"))
+        }
+
+        composeRule.onAllNodesWithTag(TestTags.Join.fileRow("first.mp4")).assertCountEquals(1)
+        composeRule.onAllNodesWithTag(TestTags.Join.fileRow("second.mp4")).assertCountEquals(1)
+    }
+}
@@ -0,0 +1,40 @@
+package org.libremediaconverter.ui
+
+import org.junit.Assert.assertEquals
+import org.junit.Assert.assertTrue
+import org.junit.Test
+
+/**
+ * No two entries of [TestTags] may share a value.
+ *
+ * A duplicated value is the one mistake this table invites -- the constants are added in blocks of
+ * near-identical lines, and a copy-paste that keeps the old string still compiles, still reads
+ * correctly at the call site, and still passes every test in the file that placed it. It surfaces
+ * later, in someone else's PR, as an affordance that "resolves to exactly one node" finding two,
+ * with nothing in that diff to explain it.
+ *
+ * Read by reflection rather than from a hand-written list, because a hand-written list would be a
+ * second copy of the table with the same copy-paste failure in it.
+ */
+class TagTableUniquenessTest {
+
+    private fun tagsIn(vararg holders: Class<*>): List<String> = holders.flatMap { holder ->
+        holder.declaredFields
+            .filter { it.type == String::class.java }
+            .map { it.get(null) as String }
+    }
+
+    @Test
+    fun `every tag constant has its own value`() {
+        val tags = tagsIn(
+            TestTags::class.java,
+            TestTags.Converter::class.java,
+            TestTags.Join::class.java,
+        )
+
+        // Without this the check would pass on an empty list, which is what a reflection call
+        // that stopped finding the constants would hand it.
+        assertTrue("reflection found only ${tags.size} tag constants, so it is not reading the table", tags.size > 20)
+        assertEquals(emptyList<String>(), tags.groupBy { it }.filterValues { it.size > 1 }.keys.toList())
+    }
+}
@@ -1,127 +1,536 @@
-# API 37 is not tested in CI: a crash in Google's `android-37.0` emulator image
+# API 37 on the emulator: a guest gralloc bug that only the host GL renderer triggers

-**Status:** open upstream, worked around by removing API 37 from the E2E matrix.
-The app itself is verified good on real API 37 hardware — this is an emulator bug only.
-**Last verified:** 2026-08-21, against emulator `37.1.11.0` and system image revision 6
+**Status:** the bug is real and still open upstream, but the previous diagnosis in this file was
+wrong about its most important detail. **The renderer decides whether API 37 boots**, and once it
+boots, disabling SystemUI collapses the crash rate far enough to run a suite —
+`tools/local-emulator/run-e2e.sh 37` gets through the whole instrumented suite and comes back with
+**2 failures, 0 errors and the two by-design skips** (measured 49 / 2 / 0 / 2 at `22c7914`, where
+the suite was 49 tests — [Reading these totals](#reading-these-totals) before comparing any total
+with another). The crashes do not stop outright, and the two failures are real; both are quantified
+below. CI now takes API 37 as two jobs — a gating leg and an advisory one for those two
+failures — see [So should CI take API 37?](#so-should-ci-take-api-37).
+**Last verified:** 2026-08-22, emulator `37.1.11.0` (build 15917651), Fedora 44,
+against system images `android-37.0` rev 6 **and** `android-37.1` rev 8.

-`minSdk` is 33 and `targetSdk` is 37, and the E2E matrix in
-[`status_check.yml`](../.github/workflows/status_check.yml) runs API 33 through 36.
-API 37 is deliberately absent. This is why.
+## The correction

-## Summary
+This file previously said, under "What was ruled out":

-The `android-37.0` emulator system image crashes `surfaceflinger` in a loop. The app
-under test never gets a working framework, so every instrumented test fails regardless
-of what the app does. The bug is in the emulator image, not in this project.
+> **GPU mode.** Both `swiftshader_indirect` and `host` crash, with the same assertion and
+> the same frames. The crash is in the gralloc mapper, below the renderer.

-The crash is an assertion inside the emulator's own gralloc implementation:
+**That is wrong.** The mapper is below the renderer, but *whether the mapper's bad path is
+reached* is not. Re-measured on 2026-08-22, seven runs, one variable at a time:
+
+| # | system image | `-gpu` | GLES the emulator chose | booted? | surfaceflinger aborts |
+|---|---|---|---|---|---|
+| r01 | `android-37.0` rev 6 | `host` | host (Mesa Iris Xe) | **no**, 422 s | 71, looping |
+| r02 | `android-37.1` rev 8 | `host` | host (Mesa Iris Xe) | **no**, 362 s | 65, looping |
+| r03 | `android-37.0` rev 6 | `swangle_indirect` | ANGLE | **yes, 85 s** | 1 |
+| r04 | `android-37.0` rev 6 | `host` + `-feature -GLDMA,-GLDMA2,-GLDirectMem` | host | **no**, 363 s | 57, looping |
+| r05 | `android-37.0` rev 6 | `angle_indirect` | ANGLE | **yes, 112 s** | 2 |
+| r06 | `android-37.1` rev 8 | `swangle_indirect` | ANGLE | **yes, 285 s** | 23 |
+| r07 | `android-37.0` rev 6 | `host` + `-feature -HostComposition` | host | **no**, wedged adb at 208 s | not readable |
+
+The discriminator is exact across all seven: **a run boots if and only if the emulator log says
+something other than `gles_mode_selected:host`.**
+
+One caveat about how independent those rows are, because the table flatters itself. `-gpu
+angle_indirect` (r05) and `-gpu swangle_indirect` (r03) both logged `gles_mode_selected:swangle`
+and both reported the same adapter, differing only in the Vulkan backend beneath
+(`vulkan_mode_selected:lavapipe` against `swiftshader`). So they are closer to one GLES path
+reached two ways than to two renderers agreeing — note that at API 33–36
+[`docs/local-emulator.md`](local-emulator.md) records `angle_indirect` resolving to ANGLE on
+*llvmpipe*, a genuinely different adapter, which it did not do here. What is 7-for-7 is the
+host-GLES-versus-not split, not "two independent renderers both work".
+
+```
+# r01, r02, r04, r07 -- never boots
+INFO | emuglConfig_init: vulkan_mode_selected:host gles_mode_selected:host
+INFO | Graphics Adapter Android Emulator OpenGL ES Translator (Mesa Intel(R) Iris(R) Xe Graphics (TGL GT2))
+
+# r03, r05, r06 -- boots
+INFO | emuglConfig_init: vulkan_mode_selected:swiftshader gles_mode_selected:swangle
+INFO | Graphics Adapter Android Emulator OpenGL ES Translator (ANGLE (Google, Vulkan 1.2.0
+     | (SwiftShader Device (Subzero) (0x0000C0DE)), SwiftShader driver-5.0.0))
+```
+
+### Why the wrong claim looked right
+
+It rested on two samples of two different things, and neither of them was ANGLE.
+
+- The **local** `swiftshader_indirect` sample was void. On this workstation *every*
+  SwiftShader-GLES launch segfaults the host emulator before the guest matters at all —
+  SELinux denies `execheap` to SwiftShader's Reactor JIT. That is
+  [`docs/local-emulator.md`](local-emulator.md), and it was not yet understood when this file
+  was written. So "`swiftshader_indirect` crashes" was true, for an entirely unrelated reason,
+  and told you nothing about the gralloc assertion.
+- The **CI** sample was one `swiftshader_indirect` run on a GPU-less `ubuntu-latest`, and the
+  **local** sample was one `-gpu host` run. Two renderers, one measurement each, and the pair
+  written up as "both GPU modes".
+
+`angle_indirect` and `swangle_indirect` — the two modes that work — had never been tried on
+API 37. Neither had a second system image.
+
+The lesson is the same one `docs/local-emulator.md` ends on, which makes it worth repeating:
+"both backends fail" is a claim about a matrix, and a matrix needs cells, not inference. Two
+observations of two different configurations do not establish anything about a third.
+
+## What the bug actually is
+
+`surfaceflinger` aborts inside the emulator's own gralloc mapper:

 ```
 Executable:    /system/bin/surfaceflinger
 signal 6 (SIGABRT), code -1 (SI_QUEUE), tid: RegionSampling
 Abort message: 'Assertion failed: !rcEnc->featureInfo()->hasReadColorBufferDma'

-  #03  mapper.ranchu.so   GoldfishMapper::readFromHost(cb_handle_t const&) const
-  #04  mapper.ranchu.so   GoldfishMapper::GoldfishMapper()::'lambda'(...)::__invoke
-  #05  libui.so           android::Gralloc5Mapper::lock(...)
-  #06  libui.so           android::GraphicBufferMapper::lock(...)
-  #07  libui.so           android::GraphicBuffer::lockAsync(...)
-  #08  libui.so           android::GraphicBuffer::lock(...)
-  #09  surfaceflinger     android::RegionSamplingThread::threadMain()
+  #03  /vendor/lib64/hw/mapper.ranchu.so  GoldfishMapper::readFromHost(cb_handle_t const&) const+543
+  #04  /vendor/lib64/hw/mapper.ranchu.so  GoldfishMapper::GoldfishMapper()::'lambda'(...)::__invoke+704
+  #05  /system/lib64/libui.so             android::Gralloc5Mapper::lock(...)+63
+  #06  /system/lib64/libui.so             android::GraphicBufferMapper::lock(...)+198
+  #07  /system/lib64/libui.so             android::GraphicBuffer::lockAsync(...)+545
+  #08  /system/lib64/libui.so             android::GraphicBuffer::lock(...)+67
+  #09  /system/bin/surfaceflinger         android::RegionSamplingThread::threadMain()+2571
 ```

-`RegionSamplingThread` is SystemUI's navigation-bar luma sampling. It calls
-`GraphicBuffer::lock`, which routes into `GoldfishMapper::readFromHost`, which asserts
-that the host has *not* negotiated the `ReadColorBufferDma` capability. On this image
-the host has, so the assertion fails and `surfaceflinger` aborts. It restarts and
-aborts again.
+`RegionSamplingThread` is SystemUI's nav-bar luma sampling. It locks a `GraphicBuffer` for CPU
+read; that routes through the Gralloc5 mapper into `GoldfishMapper::readFromHost`, which is the
+*non-DMA* readback path and asserts that the host has not negotiated `ReadColorBufferDma`. The
+host always has, so the assert fires whenever that path is taken.

-## Impact
+Two facts pin down what "always" means:

-The failure surfaces in two different ways depending on how far the job gets, which is
-why it took several rounds to identify:
+- **The capability is negotiated regardless of renderer.** The evidence is the aborts
+  themselves: the assertion that fires is `!hasReadColorBufferDma`, and it fires under ANGLE
+  (r03/r05/r06) as well as under the host translator — just far less often. That is a direct
+  observation of the guest having negotiated DMA readback under both, and it stands alone.
+  (Supporting only, and weaker than it first looks: `ANDROID_EMU_read_color_buffer_dma` appears
+  in exactly one file in the SDK, `emulator/lib64/libgfxstream_backend.so`, which every `-gpu`
+  mode goes through. A string search establishes where the extension is implemented, not that
+  it is negotiated on every path.)
+- **It is not gated by any feature flag the emulator exposes.** See the ruled-out list below.

-| Guest RAM | Where it dies | What CI reports |
-|---|---|---|
-| 1536 MB | during APK install | `Unknown failure: cmd: Can't find service: package` |
-| 2560 MB | during the test run | `There were failing tests` — all of them |
+So the renderer does not decide whether the guest *believes* DMA readback exists. It decides how
+often `RegionSamplingThread` ends up in `readFromHost` — which under the host GL translator is
+constantly, and under ANGLE is occasionally.

-At 2560 MB the install succeeds and the tests actually execute, then fail wholesale.
-The first failure in the report is misleading:
+### Why one abort takes down the whole device
+
+`surfaceflinger` is a critical service. When it dies, `init` kills the framework with it:

 ```
-kotlin.UninitializedPropertyAccessException: lateinit property output has not
-    been initialized
-    at Media3EngineTest.tearDown(Media3EngineTest.kt:53)
-
-java.lang.IllegalStateException: WorkManager is not initialized properly.
-    You have explicitly disabled WorkManagerInitializer in your manifest, ...
+08-22 21:40:28.253 I/init: Sending SIGKILL to service 'zygote' (pid 470) process group...
+08-22 21:40:28.260 I/init: Service 'zygote' (pid 470) received SIGKILL
 ```

-Neither is a real defect in this project. `tearDown` throws because `setUp` never got
-far enough to assign `output`, and WorkManager's `InitializationProvider` never runs
-because content-provider installation fails on a framework whose `surfaceflinger` is
-crash-looping. The same tests pass at API 33, 34, 35, and 36 in the same CI run, and
-the first `surfaceflinger` abort is timestamped *before* the test results are reported.
-
-This is not inference. The full suite was run against a physical API 37 device and
-passed — see [Verified on real API 37 hardware](#verified-on-real-api-37-hardware)
-below. `ConversionWorkerTest` and `ConcatWorkerTest`, which drive a real WorkManager
-round trip and are among the tests that failed this way in CI, both pass there.
+Everything above zygote goes with it, which is why the symptoms look nothing like a graphics
+bug. Under `-gpu host` the cycle repeats every five to seven seconds forever and
+`sys.boot_completed` is never set. Under ANGLE the aborts are sparse enough that the boot
+usually completes between them — but they do not stop, and each one is a framework restart.
+That is the difference between "boots" and "is usable", and it is the reason this is not simply
+fixed by changing the renderer. See [Can the suite run on it?](#can-the-suite-run-on-it) below.

 ## Environment

-Reproduced identically in two unrelated environments, so it is not specific to a host
-GPU, driver, or CI runner.
-
 | | GitHub Actions | Local workstation |
 |---|---|---|
-| Host | `ubuntu-latest`, no GPU | Fedora, Intel Iris Xe (TGL GT2) |
+| Host | `ubuntu-latest`, no GPU | Fedora 44, Intel Iris Xe (TGL GT2), kernel `7.1.8-200.fc44` |
 | Emulator | `37.1.11.0` (build 15917651) | `37.1.11.0` (build 15917651) |
-| GPU mode | `swiftshader_indirect` | `host` |
-| Result | boots, aborts during tests | aborts before boot completes |
+| GPU mode measured | `swiftshader_indirect` | `host`, `angle_indirect`, `swangle_indirect` |

-System image: `system-images;android-37.0;google_apis;x86_64`, `Pkg.Revision=6`,
-`AndroidVersion.ApiLevel=37.0`, `AndroidVersion.ExtensionLevel=22`
+Images, both reproducing it:

 ```
-Build fingerprint: google/sdk_gphone64_x86_64/emu64xa:17/CE2A.260420.019/15611780:userdebug/dev-keys
-Kernel Release:    6.12.58-android16-6-gccafb60de224-ab14828483
+system-images;android-37.0;google_apis;x86_64          Pkg.Revision=6   ApiLevel=37.0  ExtensionLevel=22
+  fingerprint google/sdk_gphone64_x86_64/emu64xa:17/CE2A.260420.019/15611780:userdebug/dev-keys
+system-images;android-37.1;google_apis_ps16k;x86_64    Pkg.Revision=8   ApiLevel=37.1  ExtensionLevel=23
+  ro.build.version.codename=REL   (a release image, not a preview)
 ```

+**Note the `ps16k` in the second one — it is not optional, and it is why the 37.1 result is
+interpretable.** From API 37.1 onward Google ships *only* 16 KB-page x86_64 images; there is no
+plain `google_apis` variant to pick. `sdkmanager --list` for 37.1 and 37.2-beta* offers nothing
+but `google_apis_ps16k` and `google_apis_playstore_ps16k`. That makes page-size alignment a
+prerequisite rather than a detail: a `.so` that is not 16 KB aligned will not load on such a
+guest, and the resulting failure looks like an app bug. Checked before the first `ps16k` boot,
+using the same test `build.yml` applies to release APKs — all 20 libraries in the committed
+`bin/ffmpeg-kit-next-8.1.1.aar`, both ABIs, report `0x4000`:
+
+```
+$ for f in jni/*/*.so; do readelf -lW "$f" | awk '$1=="LOAD"{print $NF}' | sort -u; done
+0x4000   (x20: libavcodec, libavdevice, libavfilter, libavformat, libavutil,
+          libc++_shared, libffmpegkit, libffmpegkit_abidetect, libswresample, libswscale
+          -- arm64-v8a and x86_64)
+```
+
+So when `android-37.1` reproduced the abort, that was the gralloc bug and not a page-size
+mismatch. `image_pkg_for_api` in `tools/local-emulator/run-e2e.sh` encodes the `ps16k` tag for
+37.1; if this ever fails after an FFmpeg rebuild, re-run the alignment check first.
+
 ## What was ruled out, and how

-Each of these was tested rather than reasoned about, because the first three attempts
-at this bug were plausible fixes that turned out to address earlier, unrelated failures.
+**A newer system image.** This file's own revisit trigger was "a new `android-37.0` system image
+revision ships (this was revision 6)". That trigger was written too narrowly and would never have
+fired: `android-37.0` is *still* revision 6, but Google shipped a whole new minor level.
+`android-37.1` `google_apis_ps16k` revision 8 — a `REL` build, not a beta — was installed and
+tested (r02, r06) and **behaves identically**: same assertion, same frames, never boots under
+`-gpu host`, and *worse* under ANGLE (23 aborts to `37.0`'s 1). `android-37.2-beta3` exists too
+but was not needed; two independent images agreeing settles it, and a beta could not be used by
+CI anyway.

-**Guest memory.** The emulator raises an undersized guest to a minimum on its own, but
-only for API levels it recognises, and it does not recognise `"37.0"`. API 33 bumps to
-2048 MB and 34/35/36 to 2560 MB, while API 37 logged no bump at all and ran at the
-`pixel_6` default of 1536 MB. Setting `ram-size: 2560M` explicitly fixed that asymmetry
-and did change the outcome — the job got past install and into the test run — but it is
-not the underlying bug. At the moment of failure the guest reported `MemTotal 2527392
-kB` with `MemAvailable 1507104 kB`: 1.5 GB free, and no OOM kills.
+**An ATD image.** Still does not exist for API 37. `sdkmanager --list` offers `aosp_atd` and
+`google_atd` for API 30 through 36 and nothing above:

-**GPU mode.** Both `swiftshader_indirect` and `host` crash, with the same assertion and
-the same frames. The crash is in the gralloc mapper, below the renderer.
+```
+system-images;android-36;google_atd;x86_64 | 1 | Google APIs ATD Intel x86_64 Atom System Image
+(no android-37 ATD of any kind)
+```

-**Disabling the DMA feature.** `GLDMA` is the host feature that most plausibly backs the
-guest's `hasReadColorBufferDma`. Launching with `-feature -GLDMA` was accepted by the
-emulator — the log confirms `Feature 'GLDMA' (51) is overridden to 'disabled'` — and
-`surfaceflinger` still aborted 13 times and the device never finished booting. Whatever
-sets that guest capability, it is not this flag.
+For API 37 the only x86_64 images are `google_apis`, `google_apis_playstore`, their `ps16k`
+16 KB-page variants, and Wear OS. Check again when revisiting.

-**An ATD image.** `google_atd` / `aosp_atd` images are built for automated testing and
-ship without the SystemUI package set, which is what drives `RegionSamplingThread` in
-the first place. That would likely sidestep the bug class entirely, but **no ATD image
-exists for `android-37.0`** — only `google_apis`, `google_apis_playstore`, the `ps16k`
-16 KB-page variants, and Wear OS. Check again when revisiting; if an ATD image appears,
-try it before anything else here.
+**The DMA feature flags.** `GLDMA` alone was ruled out previously; `GLDMA2` and `GLDirectMem`
+were not, and the per-image `advancedFeatures.ini` turns all three on. Disabling all three
+together (r04) is accepted by the emulator and changes nothing:
+
+```
+INFO | Feature 'GLDMA' (51) is overridden to 'disabled'
+INFO | Feature 'GLDMA2' (52) is overridden to 'disabled'
+INFO | Feature 'GLDirectMem' (53) is overridden to 'disabled'
+... 57 surfaceflinger aborts, device never boots
+```
+
+**Host composition.** `-feature -HostComposition` (r07) was the best remaining guess at what
+forces the readback. It did not help; it made things worse, wedging adb entirely at 208 s so the
+crash buffer could not even be read. Recorded as inconclusive rather than ruled out, because no
+evidence came back from it.
+
+**Guest feature negotiation differing from API 36.** It does not. The image-level
+`advancedFeatures.ini` for `android-37.0` is byte-identical to `android-36`'s except for one
+unrelated line:
+
+```
+$ diff android-36/google_apis/x86_64/advancedFeatures.ini android-37.0/google_apis/x86_64/advancedFeatures.ini
+QemuCameraSensorOrientation = on
+```
+
+`GLDMA`, `GLDMA2`, `GLDirectMem`, `GrallocSync`, `HostComposition` and `YUVCache` are on in
+both. API 36 boots and passes. So nothing about the host/guest feature handshake changed — the
+regression is in the guest's Gralloc5 mapper or in what API 37's `RegionSamplingThread` asks of
+it, not in what the emulator advertises.
+
+**Guest memory.** Ruled out previously and not revisited; every run above used
+`hw.ramSize=2560`, the same value the E2E matrix pins, and none of them OOMed.
+
+**A host-side crash.** Not this bug, and worth stating because the other emulator failure on this
+workstation *is* host-side. Every run above left `coredumpctl` empty and produced zero
+`avc: denied` lines, and the qemu process was still alive at the end of the ones that never
+booted (`emulator_alive=yes`). The host emulator is fine; the guest is not.
+
+## Can the suite run on it?
+
+**Almost.** `tools/local-emulator/run-e2e.sh 37` now runs the whole suite locally, and all of it
+passes except two tests. Measured at `22c7914`: **49 tests, 2 failures, 0 errors, 2 skipped** — 45
+passed, the two `Media3EngineTest` failures dissected below, and the two `assumeTrue` skips every
+level has. It costs two deviations from how every other level is run, and both are worth
+understanding before trusting the leg.
+
+Two things about that total before it is compared with anything. It is the size of the suite on
+the checkout that ran, not a property of API 37 — `app/src/androidTest` held 49 `@Test` methods at
+`22c7914`, and a newer checkout reports its own count; see
+[Reading these totals](#reading-these-totals). And **the Pixel has never run 49**: its green run
+was 40 / 0 / 0 / 2 at `edd6385`, the same suite nine tests earlier. What compares across the two
+is two failures against none, and the same two skips — not the totals.
+
+The same numbers and the same two test names came back twice, which is real corroboration — but
+by two different routes, and only one of them is the harness. The first was driven by hand
+(`pm disable-user`, then several minutes of incidental framework restarts, then `e2e-run.sh`
+directly); the second went through `disable_region_sampling`'s `stop; start`. **The harness path
+itself has one green measurement.** What would make this routine is a second consecutive
+`run-e2e.sh 37` whose only failures are the same two.
+
+### Booting is not the same as being usable
+
+Changing the renderer gets the device to `sys.boot_completed=1`, and that is all it gets you. The
+aborts do not stop, and each one is a framework restart. A five-minute test run does not survive
+that. What it looks like from Gradle:
+
+```
+Shell command failed (1): rm -rf "/sdcard/Android/media/org.libremediaconverter/..."
+    rm: ...: Transport endpoint is not connected
+Starting 0 tests on lmc_e2e_api37(AVD) - 17
+Shell command failed (20): am get-current-user
+    cmd: Can't find service: activity
+Device emulator-5572 failed to uninstall test APK org.libremediaconverter.
+    [cmd: Can't find service: package]
+Test run failed to complete. No test results.
+    onError: commandError=false message=INSTRUMENTATION_ABORTED: System has crashed.
+```
+
+Measured idle rate on `android-37.0` under `swangle_indirect`: **10 aborts in 150 s, then 11 more
+in the next 150 s**. Steady, not a start-up transient.
+
+### The fix is to remove the region-sampling listener, not to survive it
+
+`RegionSamplingThread` exists only because SystemUI registers a nav-bar luma-sampling listener.
+Take SystemUI away and the thread is never started, so the mapper's bad path is never called:
+
+```
+$ adb shell pm disable-user --user 0 com.android.systemui
+Package com.android.systemui new state: disabled-user
+
+=== aborts at start of measurement: 36
+=== idle 180s with SystemUI disabled ===
+=== aborts after: 36   NEW IN WINDOW: 0
+--- services still up? ---
+  activity   Service activity: found
+  package    Service package: found
+  window     Service window: found
+```
+
+**Zero in 180 s, against 10–11 per 150 s.** That is the strongest evidence that region sampling
+is the dominant trigger, and it is worth recording even by someone who never wants the workaround.
+It does not establish it as the *only* trigger: the paragraph below has an abort surviving the
+disable, and nothing measured here says whether that residue is a second caller of the readback
+path or a disable that did not fully take.
+
+Do not read that as "the crashes stop", though, because the harness path does not reproduce a
+clean zero. Its own post-disable check on the run recorded below printed
+
+```
+  quiet check: 1 new surfaceflinger aborts in 45 s (want 0)
+  surfaceflinger hasReadColorBufferDma aborts: 4     (whole run)
+```
+
+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.
+
+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.
+
+### 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 is defensible here only because nothing in this suite touches
+   system UI — these are Media3, FFmpeg and WorkManager tests — and because the alternative is no
+   local API 37 coverage at all. **Anything that ever does depend on system UI must not trust
+   this leg.**
+
+### The two remaining failures are the same bug, one layer down
+
+```
+org.libremediaconverter.convert.Media3EngineTest > runsFromAThreadWithNoLooper          FAILED
+org.libremediaconverter.convert.Media3EngineTest > transcodesH264ToH265AndReportsProgress FAILED
+
+androidx.media3.transformer.ExportException: Codec exception:
+    CodecInfo{type=VideoDecoder, ..., mime=video/avc, name=c2.goldfish.h264.decoder}
+  at androidx.media3.transformer.DefaultCodec.maybeDequeueOutputBuffer(DefaultCodec.java:398)
+Caused by: android.media.MediaCodec$CodecException:
+  at android.media.MediaCodec.native_dequeueOutputBuffer(Native Method)
+```
+
+Three measurements say this is the emulator image and not this app, and not the software
+renderer. A fourth bullet offers a mechanism, and is inference rather than measurement:
+
+- **Control at API 35 under the identical renderer.** `GPU_MODE=swangle_indirect
+  tools/local-emulator/run-e2e.sh 35` → **49 / 0 / 0 / 2** at `22c7914`, green.
+  `c2.goldfish.h264.decoder` is perfectly happy under ANGLE one API level down, so the renderer is
+  not what breaks it.
+- **Real API 37 hardware passes**, see below. There is no `c2.goldfish.*` codec on a Pixel.
+- **API 36 against API 37 on CI, back to back, everything else held.** Same two tests, same
+  `-gpu swiftshader_indirect`, same SystemUI-disable path — `pm disable-user`, `stop`, wait for
+  `system_server` to actually be gone, `start`, then verify against `pm list packages -d`. Both
+  runs were narrowed to the two failing tests:
+
+  ```
+  -Pandroid.testInstrumentationRunnerArguments.class=\
+    org.libremediaconverter.convert.Media3EngineTest#transcodesH264ToH265AndReportsProgress,\
+    org.libremediaconverter.convert.Media3EngineTest#runsFromAThreadWithNoLooper
+  ```
+
+  and the filter is confirmed three independent ways: `tests="2"` in the XML, `Expected 2 tests`
+  in the abort message, and `run started: 2 tests` in the guest logcat.
+
+  | run | api | result XML |
+  |---|---|---|
+  | [32660148155](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32660148155) | 37.0 | `tests="2" failures="2" errors="0" skipped="0"` |
+  | [32660152961](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32660152961) | 36 | `tests="2" failures="0" errors="0" skipped="0" time="4.603"` |
+
+  API 37 fails with the signature above — `name=c2.goldfish.h264.decoder`,
+  `MediaCodec$CodecException` at `dequeueOutputBuffer(MediaCodec.java:4274)`. API 36 passes both in
+  4.603 s, and `c2.goldfish.h264.decoder` is in *its* logcat too (44 mentions), so the two runs are
+  not being served by different decoder names. **What this falsifies is "the stripped
+  configuration is what breaks these tests"** — a reading none of the other measurements
+  addresses, because they all compare against a device that still had SystemUI. Here SystemUI is
+  absent and the framework has been restarted on both sides, and the healthy image is green anyway.
+
+  Two things it does **not** control, which is why it narrows the claim rather than closing it:
+
+  - **The restarts were not performed under equal conditions.** API 36 did its `stop`/`start` with
+    `dma_aborts=0`; API 37's did the same restart with two aborts already logged. "A framework
+    restart performed while the abort loop is running" therefore remains uncontrolled.
+  - **The images differ on the encoder side.** These tests transcode H.264 → H.265. The API 37
+    logcat carries `c2.goldfish.hevc.decoder` (16 mentions in the control run) where API 36 carries
+    `c2.android.hevc.encoder` (32). The pipeline is not identical end to end, which is a second
+    reason "the image ships a broken h264 decoder" is the wrong *shape* of claim: what is measured
+    is that these two tests fail on the API 37 image, pass at API 36 under the same renderer *and*
+    the same disable path, and pass at 33–36 without needing that path at all — because nothing
+    below 37 has the bug it works around.
+- The failing call is `dequeueOutputBuffer` on the *goldfish* decoder — the emulator's own codec,
+  which like `RegionSamplingThread` gets its frames out of a host-side colour buffer. Same
+  readback machinery, one layer down. This is inference rather than a measurement, and is flagged
+  as such; what is measured is the first three bullets.
+
+**Do not try `-feature -HardwareDecoder`.** It is the obvious next idea and it is much worse:
+forcing the guest onto software decoders took the run from 2 failures to **46**, across
+`RemuxTest`, `ForcedFailureTest`, `HardwareFallbackTest` and `UnopenableUriTest` as well. The
+suite depends on those decoders existing.
+
+### The intact-SystemUI counterfactual cannot be measured on CI
+
+The control the block above still lacks is the obvious one: run those same two tests at API 37
+with SystemUI **left running**. Passing would put the failure on the disable rather than on the
+image; failing on the decoder would make the decoder attribution direct instead of inferred.
+
+**Seven dispatches of `api37-debug.yml`, zero verdicts.** Not bad luck — a mechanism, which is why
+this is written down rather than left as a gap for the next person to spend seven runs on:
+
+| arm | run | result XML | what actually happened |
+|---|---|---|---|
+| E1 | [32660528355](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32660528355) | `tests="1" failures="1"`, `<failure>` body empty | `Expected 2 tests, received 0. INSTRUMENTATION_ABORTED: System has crashed.` |
+| E2 | [32660533845](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32660533845) | `tests="0"` | never installed: `Failed to commit install session ... Failure calling service package: Broken pipe (32)` |
+| E3 | [32660539259](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32660539259) | `tests="0"` | `Test run failed to complete. No test results.` |
+| E4 | [32661117237](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32661117237) | `tests="2" failures="2"` | both failed in `@Before`, never reached MediaCodec |
+| E5 | [32661121972](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32661121972) | `tests="2" failures="2"` | same |
+| S1 | [32661127224](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32661127224) | `tests="1" failures="1"` | same, single-test arm |
+| S2 | [32661132024](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32661132024) | `tests="1" failures="1"` | same |
+
+While the framework is crash-looping, the guest cannot reliably create per-user private
+directories. An app installed during the loop has no cache directory — and `Media3EngineTest`
+copies its H.264 fixture into `context.cacheDir` in `@Before`, so it dies there, **before any
+MediaCodec exists**:
+
+```
+W/ContextImpl( 8216): Failed to ensure /data/user/0/org.libremediaconverter/cache
+I/TestRunner( 8216): run started: 1 tests
+E/TestRunner( 8216): failed: transcodesH264ToH265AndReportsProgress(...)
+E/TestRunner( 8216): java.io.FileNotFoundException:
+  /data/user/0/org.libremediaconverter/cache/sample_h264.mp4: open failed: ENOENT
+  at org.libremediaconverter.convert.Media3EngineTest.setUp(Media3EngineTest.kt:47)
+```
+
+Not app-specific: `com.google.android.googlesdksetup` and `com.google.android.apps.nexuslauncher`
+hit the same `Failed to ensure /data/user/0/<pkg>/cache` in the same logcats.
+
+**The result XML masks this, and reading only the report gets you the wrong bug.** What E4, E5, S1
+and S2 report is
+
+```
+<failure>kotlin.UninitializedPropertyAccessException: lateinit property output has not been initialized
+at org.libremediaconverter.convert.Media3EngineTest.tearDown(Media3EngineTest.kt:56)
+```
+
+— `tearDown` failing because `setUp` threw before it assigned `output`. That looks like a
+teardown defect in this repository and is not one; the cause is only in the guest logcat.
+
+So the obstacle is structural: install, data-directory creation and instrumentation start-up do
+not fit between framework kills, and four of the seven runs show the directory creation itself is
+broken during the loop. More dispatches of this shape would repeat these outcomes. The
+counterfactual is still open on the **Pixel 10 Pro XL**, the one API 37 device here that is not an
+emulator — but a Pixel has no `c2.goldfish.*` codec at all, so it answers "does the app work at
+API 37", not "is that codec broken".
+
+#### Abort cadence, corrected
+
+`.github/workflows/api37-debug.yml` carried "roughly every 20 s" for the kill cycle in its own
+comments. That number was the watchdog's **sampling** interval, not the cadence, and the two got
+conflated. Measured across the seven runs above, gaps between successive `hasReadColorBufferDma`
+aborts run **20 s to 90 s, median 60–70 s — three to five aborts in a four-minute window**.
+Slower than assumed, and still not slow enough: install, data-directory creation and
+instrumentation start-up do not fit inside one gap.
+
+`sys.boot_completed` held at `1` throughout every one of those test windows. The device reports
+itself booted while zygote is being killed under it, which is why no boot-state check catches
+this and why `stop`/`start` waits must poll `pidof system_server` and `service check` instead
+(see `disable_region_sampling` in `tools/local-emulator/run-e2e.sh`).
+
+### So should CI take API 37?
+
+**Yes, as two jobs: a gating `E2E API 37` and an advisory leg carrying the two tests that do not
+pass.** That reverses the answer this section gave, and the reversal is measured rather than
+argued — two of its three reasons were claims *about CI*, and CI had never been measured. The
+instrument that measured it is [`.github/workflows/api37-debug.yml`](../.github/workflows/api37-debug.yml),
+dispatch-only, a copy of the E2E job with the matrix replaced by inputs.
+
+Every run below is `ubuntu-latest`, KVM on, `pixel_6`, x86_64, disk 8G, RAM 2560M, emulator
+`37.1.11.0` build 15917651 — the same emulator build the local investigation used. Every **API
+37** row is `system-images;android-37.0;google_apis;x86_64`; c2 is the API 36 control and runs
+that level's own image, which is the whole point of it.
+
+| # | run | api | `-gpu` | SystemUI | suite | verdict |
+|---|---|---|---|---|---|---|
+| c1 | [32644947334](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32644947334) | 37.0 | swiftshader_indirect | running | `Starting 0 tests` | FAIL |
+| c2 | [32644965828](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32644965828) | **36** | swiftshader_indirect | running | 57 tests, BUILD SUCCESSFUL | green control |
+| c3 | [32644970240](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32644970240) | 37.0 | swangle_indirect | running | `Starting 0 tests` | FAIL |
+| c5 | [32645543238](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32645543238) | 37.0 | swangle_indirect | disabled | 57 / 2 / 0 / 2 | suite ran |
+| c6 | [32646029143](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32646029143) | 37.0 | swiftshader_indirect | one-shot disable, **did not hold** | `Starting 0 tests` | FAIL |
+| c8 | [32646611485](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32646611485) | 37.0 | swiftshader_indirect | disabled, verified | 57 / 2 / 0 / 2 | suite ran |
+| c9 | [32646615706](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32646615706) | 37.0 | swiftshader_indirect | disabled, verified | 57 / 2 / 0 / 2 | suite ran |
+| c10 | [32646619472](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32646619472) | 37.0 | swiftshader_indirect | disabled, verified | 57 / 2 / 0 / 2 | suite ran |
+| c11 | [32647138060](https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32647138060) | 37.0 | swiftshader_indirect | disabled, verified | 57 / 2 / 0 / 2 | suite ran |
+
+57 is that checkout's own `@Test` count at `acc71bc`, so those are whole-suite runs and not
+truncated ones — see [Reading these totals](#reading-these-totals). Taking the three old reasons
+in turn:
+
+1. **"Nothing says a runner would be stable" — measured, and it is.** `-gpu swiftshader_indirect`
+   on a GPU-less runner resolves to `gles_mode_selected:swiftshader`, a third renderer that
+   locally never survives to say anything (Fedora's SELinux denies `execheap` to SwiftShader's
+   Reactor JIT — see [`local-emulator.md`](local-emulator.md)). It boots `android-37.0` in about
+   60 s. The local discriminator — fatal iff `gles_mode_selected:host` — holds, and a runner with
+   no GPU can never select host, so CI was never in the fatal class. Switching CI's `-gpu` changes
+   nothing either way: c1 and c3 both fail with SystemUI up, under swiftshader and swangle
+   respectively, and c5 and c8–c11 show the suite running under either once SystemUI is gone.
+2. **"Bespoke device surgery" — still true, and now a written caveat rather than a reason to skip
+   the level.** It is one env flag, `E2E_DISABLE_SYSTEM_UI`, read by `.github/scripts/e2e-run.sh`
+   and unset on every other leg. What it costs is stated where it can be read from the failing
+   check: the API 37 row runs a device configuration no other leg and no Pixel run uses. What
+   makes it dependable is verification, not repetition — c6 is the counter-case, a one-shot
+   `pm disable-user` that reported `new state: disabled-user` and then started SystemUI eight more
+   times. The verified form is 4/4; the unverified form was 3/4.
+3. **"Permanently red or permanently allow-listed" — this was the real objection, and it is the
+   one the split answers.** The two failures are marked `@FailsOnEmulatorApi37` in
+   `app/src/androidTest`. The gating job runs `notAnnotation` on that marker and must be green;
+   the advisory job runs `annotation` on the *same* marker, reports, and never blocks. One marker
+   rather than two lists, so a test cannot silently end up in neither job — which would read as
+   green.
+
+The cost is about three minutes on the API 37 leg — the `stop`/`start` plus a 45 s quiet window,
+and another round when the first does not verify. Measured wall clock for the whole job, boot
+included: 6–7 minutes at API 37 against ~6 at API 36.
+
+Two things this does **not** buy. The advisory job is expected red, so a *third* failure there is
+the signal and the run's logcat is the only thing that distinguishes it — which is why that job
+uploads diagnostics unconditionally. And a green `E2E API 37` still does not replace the release
+check on the Pixel: the emulator leg runs without SystemUI, and the Pixel does not.
+
+The *local* story changed at the same time and independently: API 37 is no longer a level nobody
+can look at. A regression that shows up at 37 and not at 36 can be reproduced on this workstation
+in about four minutes.

 ## Verified on real API 37 hardware

-The bug is confined to the emulator image. On 2026-08-21 the whole instrumented suite
-was run against a physical device and passed:
+Unchanged and still true. On 2026-08-21 the whole instrumented suite ran green on a physical
+device:

 ```
 Device:  Pixel 10 Pro XL (mustang), arm64-v8a
@@ -132,65 +541,94 @@ API:     37 (Android 17, codename REL -- a release build, not a preview)
  40 tests, 0 failures, 0 errors, 2 skipped        BUILD SUCCESSFUL
 ```

+**That "40" is a measurement of the tree it ran on, not a baseline for today**, and it is not a
+contradiction of the totals in [`docs/local-emulator.md`](local-emulator.md) either.
+
+### Reading these totals
+
+Every total in this file and in [`docs/local-emulator.md`](local-emulator.md) is the size of
+`app/src/androidTest` on the checkout that produced it, and nothing else. The reported total has
+equalled that checkout's `@Test` count everywhere it has been checked:
+
+| checkout | `@Test` methods | total the run reported |
+|---|---|---|
+| `edd6385` | 40 | 40 — the Pixel run above |
+| `22c7914` | 49 | 49 — the four local levels, and API 37 |
+| `18c53a3` | 57 | not run |
+
+So the number to expect is not written down here. It is derived from the checkout in front of
+you, which is the only thing that cannot go stale:
+
+```bash
+grep -rho '@Test' app/src/androidTest | wc -l
+```
+
+**Before each release, run the suite on the Pixel 10 Pro XL and expect that many tests, 0
+failures, 0 errors, 2 skipped.** The failure, error and skip counts are the invariant; the total
+is not. A total that disagrees with your own checkout's count is the signal — an old checkout, a
+stale build, or tests that never ran — and it is worth stopping on either way.
+
 The two skips are `RealMediaBenchmark.hardwareVersusSoftwareOnRealVideo` and
-`av1InputRoutesAccordingToDeviceDecodeSupport`, which `assumeTrue` their sample files
-are present and skip when they are not. That is by design and unrelated to API level.
+`av1InputRoutesAccordingToDeviceDecodeSupport`, which `assumeTrue` their sample files are present
+and skip when they are not. That is by design and unrelated to API level.

 One harmless warning appears during the run and can be ignored:
-`No UID for androidx.test.services in user 0`, from an `appops` call the test services
-package makes before it is fully registered.
-
-So the app is correct on Android 17. What is missing is only *automated* coverage in
-CI. Until the image is fixed, run the suite on a physical API 37 device before release;
-that is the substitute for the missing matrix row.
+`No UID for androidx.test.services in user 0`, from an `appops` call the test services package
+makes before it is fully registered.

 ## Reproducing it

-Locally, with `-gpu host` so the emulator itself does not segfault on Intel graphics:
+Both halves, so the renderer claim can be checked rather than taken on trust:

 ```bash
+export ANDROID_HOME="$HOME/Android/Sdk"
+export PATH="$ANDROID_HOME/platform-tools:$ANDROID_HOME/emulator:$ANDROID_HOME/cmdline-tools/latest/bin:$PATH"
+
 sdkmanager --install "system-images;android-37.0;google_apis;x86_64"
 echo no | avdmanager create avd -n api37_repro \
    -k "system-images;android-37.0;google_apis;x86_64" -d pixel_6 --force

-$ANDROID_HOME/emulator/emulator -avd api37_repro \
-    -no-window -gpu host -noaudio -no-boot-anim -camera-back none -no-snapshot &
+# never boots -- surfaceflinger aborts every ~6 s, forever
+emulator -avd api37_repro -no-window -gpu host \
+    -noaudio -no-boot-anim -camera-back none -no-snapshot &

-# Boot never completes. Count the aborts:
-adb logcat -d -b crash | grep -c hasReadColorBufferDma
+# boots in ~85 s, having aborted once or twice on the way
+emulator -avd api37_repro -no-window -gpu swangle_indirect \
+    -noaudio -no-boot-anim -camera-back none -no-snapshot &
 ```

-`sys.boot_completed` never reaches `1`, `pgrep -f system_server` stays empty, and
-`keystore2`'s watchdog logs `await_boot_completed ... Overdue` indefinitely.
+Count the aborts either way:

-To see the CI-side form instead, restore the API 37 row in the E2E matrix of
-`status_check.yml` (`api-level: "37.0"` — a bare `37` fails earlier still, during SDK
-setup, because there is no `platforms;android-37`).
+```bash
+adb -s emulator-5554 logcat -d -b crash | grep -c hasReadColorBufferDma
+```
+
+Under `-gpu host`, `sys.boot_completed` never reaches `1`, `pgrep -f system_server` stays empty,
+and `keystore2`'s watchdog logs `await_boot_completed ... Overdue` indefinitely.
+
+`tools/local-emulator/run-e2e.sh 37` does all of this, with the working renderer picked
+automatically — see `gpu_for_api` in that file.

 ## Filing this upstream

-Not yet filed. To file it:
+Not yet filed. The report is stronger than it was, because the renderer dependency narrows it:

-1. Go to <https://issuetracker.google.com/> and sign in with a Google account.
-2. Choose **Report an issue**, then pick the component for the Android emulator — search
-   the component picker for "Emulator"; it sits under the Android Studio component tree.
-   If the picker is unclear, Android Studio's **Help → Submit Feedback** opens the same
-   tracker with the component preselected, and the emulator's own **Extended controls →
-   Help → File a bug** does likewise.
-3. Title it for the mechanism, not the symptom, so it is searchable — for example:
-   `surfaceflinger aborts in GoldfishMapper::readFromHost (hasReadColorBufferDma) on
-   android-37.0 google_apis x86_64`.
-4. Paste the assertion and backtrace from the top of this document, the environment
-   table, and the reproduction steps above. State explicitly that it reproduces on two
-   unrelated hosts under both GPU modes — that is the detail that stops it being closed
-   as a local graphics problem.
-5. List what was ruled out. Bugs that arrive with `-feature -GLDMA` already eliminated
-   tend not to bounce back asking for it.
-6. Attach:
-   - the guest tombstone, via `adb pull /data/tombstones` (or the `pbtombstone` output
-     the crash log names)
+1. Go to <https://issuetracker.google.com/>, **Report an issue**, and pick the Android emulator
+   component (search the component picker for "Emulator"; Android Studio's **Help → Submit
+   Feedback** opens the same tracker with it preselected).
+2. Title it for the mechanism: `surfaceflinger aborts in GoldfishMapper::readFromHost
+   (hasReadColorBufferDma) on android-37.0 and android-37.1 x86_64 -- fatal under -gpu host,
+   intermittent under ANGLE`.
+3. Paste the assertion and backtrace, the environment block, and the seven-row matrix. The
+   matrix is the valuable part: it shows the abort is not renderer-specific but its *frequency*
+   is, which points at the readback path rather than at any one GL implementation.
+4. State that it reproduces on two independent system images (`37.0` rev 6 and `37.1` rev 8) and
+   on two unrelated hosts, and that `-feature -GLDMA,-GLDMA2,-GLDirectMem` does not suppress it.
+5. Attach:
+   - the guest tombstone, via `adb pull /data/tombstones` (or the `pbtombstone` output the crash
+     log names)
   - `adb logcat -d -b crash > crash.txt`
-   - the emulator's own stdout log, captured by redirecting the launch command
+   - the emulator's own stdout log (`-verbose -debug all`, redirected)
   - the AVD's `config.ini`
   - a link to a failing CI job, which shows it on hardware you do not control:
     <https://github.com/JMR-dev/LibreMediaConverter/actions/runs/32545625459/job/96963461184>
@@ -199,16 +637,46 @@ Record the issue number here once filed.

 ## When to revisit

-Re-add the API 37 row when any of these happens:
+The old trigger list named "a new `android-37.0` revision", which is why nothing ever fired even
+though a new API level shipped. Watch for these instead:

- a new `android-37.0` system image revision ships (this was revision 6)
- an ATD image appears for API 37
- the upstream issue is marked fixed
+- **any new API 37.x system image**, not just a new revision of `37.0` — `37.1` rev 8 and
+  `37.2-beta*` already exist, and more will. Test with `-gpu host`: if it boots, the guest mapper
+  is fixed.
+- **an ATD image for API 37.** Still none as of 2026-08-22. ATD images ship without SystemUI,
+  which is what drives `RegionSamplingThread`, so one would very likely sidestep the bug
+  entirely. Try it before anything else here.
+- **the upstream issue being marked fixed.**
+- **`E2E API 37 Media3 hardware transcode (advisory)` going green.** Nothing announces this: the
+  job is `continue-on-error`, so it fixing itself looks exactly like a check nobody reads
+  quietly ceasing to be red. It is listed here because that makes it the *least* likely of these
+  triggers to be noticed, not the most. When it happens, delete `@FailsOnEmulatorApi37` from the
+  two tests rather than the job — the gating leg picks them back up on its own, and the advisory
+  job then runs nothing and can go.

-Until then the gap is narrower than the missing row suggests. `targetSdk` is 37, so the
-app is compiled and unit-tested against it; the API-dependent behaviour this matrix
-exists to exercise — the foreground service type, absent below 34, `dataSync` at 34,
-`mediaProcessing` from 35 — is covered at 35 and 36; and the full instrumented suite has
-been run green on real API 37 hardware. What is missing is *automated* API 37 coverage,
-so a regression there would not be caught by a pull request. Run the suite on a physical
-API 37 device before each release for as long as this row is absent.
+## Correction owed to `CLAUDE.md`
+
+`CLAUDE.md` currently says:
+
+> - **The API 37 image is broken.** `android-37.0` crash-loops surfaceflinger inside its own
+>   gralloc mapper, so every test fails there regardless of this app.
+>   `docs/api-37-emulator-crash.md` records the evidence and the ruled-out fixes; CI's matrix
+>   therefore stops at API 36 even though targetSdk is 37.
+
+The first sentence is right, and now under-specified in one direction and over-specified in the
+other: it is not only `android-37.0` (it is `37.1` too), and it does not crash-loop under every
+renderer. **The last clause is now simply false: CI's matrix does not stop at API 36 any more.**
+Proposed replacement, offered for review rather than applied here — `CLAUDE.md` is left alone
+deliberately, because several branches touch it:
+
+> - **The API 37 images crash-loop surfaceflinger under the host GL renderer.** Both
+>   `android-37.0` and `android-37.1` abort inside their own gralloc mapper
+>   (`RegionSamplingThread` → `GoldfishMapper::readFromHost`), and when surfaceflinger dies init
+>   SIGKILLs zygote, so the framework restarts under the test run. Under `-gpu host` it never
+>   boots at all; under `-gpu swangle_indirect` it boots and the aborts merely become
+>   intermittent. `docs/api-37-emulator-crash.md` has the seven-run matrix and the ruled-out
+>   list, and `tools/local-emulator/run-e2e.sh` picks the working renderer per API level.
+>   CI takes API 37 as two jobs: a gating `E2E API 37` that disables SystemUI first, and an
+>   advisory leg carrying the two `@FailsOnEmulatorApi37` tests. The gating leg therefore runs a
+>   device configuration nothing else does. **API 37 still needs a manual check on the Pixel 10
+>   Pro XL before each release** — it is the only API 37 run with SystemUI intact.
@@ -1,8 +1,10 @@
 # Emulators do run on this host: the segfault is SwiftShader's JIT against SELinux

-**Status:** solved. Local instrumented runs work with `-gpu host`, and the suite is green
-on API 33–36 — 49 tests, 0 failures, 0 errors, 2 skipped on every level. See
-[The sweep, run](#the-sweep-run).
+**Status:** solved. Local instrumented runs work with `-gpu host`, and the whole suite is green
+on API 33–36 — 0 failures, 0 errors and the two by-design skips on every level, measured as
+49 / 0 / 0 / 2 at `22c7914`, where the suite was 49 tests. See [The sweep, run](#the-sweep-run),
+and [Reading these totals](api-37-emulator-crash.md#reading-these-totals) before comparing any
+total with another checkout's.
 **Last verified:** 2026-08-22, emulator `37.1.11.0` (build 15917651), Fedora 44,
 kernel `7.1.8-200.fc44`, `selinux-policy-44.6-1.fc44`

@@ -190,11 +192,28 @@ emulator -avd <name> -no-window -gpu host \
 `.github/scripts/e2e-run.sh` for the run itself. Use it rather than the raw command:

 ```bash
-tools/local-emulator/run-e2e.sh          # API 33 34 35 36
+tools/local-emulator/run-e2e.sh          # API 33 34 35 36 37
 tools/local-emulator/run-e2e.sh 35       # one level
+tools/local-emulator/run-e2e.sh 37 37.1  # both API 37 images
 GPU_MODE=swangle_indirect tools/local-emulator/run-e2e.sh 35
 ```

+Levels are the labels above, not SDK ints: API 37's SDK directories are dotted
+(`android-37.0`, `android-37.1`) and there is no `android-37`, so `37` is accepted as a
+spelling of `37.0`. Setting `GPU_MODE` forces one renderer on every level, which is what
+you want when measuring a mode; leaving it unset lets `gpu_for_api` pick, which is what
+you want when running the suite — 33–36 need `host` and 37 must not have it.
+
+**A bare `run-e2e.sh` exits 1, and that is the design.** API 37 is in the default list
+deliberately — leaving it out is what left the level unlooked-at for as long as it was — and
+it is permanently two failures short of green: `Media3EngineTest` cannot drive the emulator's
+`c2.goldfish.h264.decoder` on those images, which
+[`api-37-emulator-crash.md`](api-37-emulator-crash.md) pins on the image and not on this app
+(API 35 under the same renderer is green). The summary row names the two expected failures so
+that a third is visibly new, and the script repeats the point on the way out. Anything that
+treats a non-zero exit as breakage — a wrapper, a hook, a habit — should name the levels it
+wants: `run-e2e.sh 33 34 35 36` is the sweep that can be green.
+
 `swangle_indirect` is the fallback worth knowing about. It is entirely software, so it
 does not depend on reaching the session's GPU — useful over plain SSH, where `-gpu host`
 has not been tested and may not find a device. It is also the closest local analogue to
@@ -236,8 +255,9 @@ after an AGP upgrade.
 ## The sweep, run

 `tools/local-emulator/run-e2e.sh`, one invocation per level so each got a freshly created
-AVD, `-gpu host` throughout, 2026-08-22 19:42–19:56. Every level matches the physical
-Pixel 10 Pro XL (API 37) baseline of 49 / 0 / 0 / 2 exactly:
+AVD, `-gpu host` throughout, 2026-08-22 19:42–19:56, on `22c7914`. All four levels agree exactly,
+and 49 is that checkout's whole suite — every `@Test` in `app/src/androidTest`, two of which skip
+by design everywhere:

 | API | Android | AVD | Boot | `connectedDebugAndroidTest` | Tests | Failures | Errors | Skipped |
 |---|---|---|---|---|---|---|---|---|
@@ -246,6 +266,12 @@ Pixel 10 Pro XL (API 37) baseline of 49 / 0 / 0 / 2 exactly:
 | 35 | 15 | `lmc_e2e_api35` | 40 s | 3 m 46 s | 49 | 0 | 0 | 2 |
 | 36 | 16 | `lmc_e2e_api36` | 90 s | 2 m 18 s | 49 | 0 | 0 | 2 |

+The physical Pixel has never reported 49, and an earlier version of this paragraph said the
+sweep matched it exactly. Its green API 37 run was 40 / 0 / 0 / 2, at `edd6385` — the same suite
+nine tests earlier. What matches is 0 failures, 0 errors and the same two skips; totals only ever
+match between runs of one checkout, which
+[`api-37-emulator-crash.md`](api-37-emulator-crash.md#reading-these-totals) sets out.
+
 Thirteen and a half minutes for the four levels, AVD creation and cold boots included;
 fifteen with the pre-warm build in front of them. Nothing needed a retry, and no level
 produced a `diagnostics-api*.txt` — `e2e-run.sh` writes that only on the failure path, so
@@ -352,9 +378,11 @@ and the same binaries against the same kernel boot fine under `-gpu host`.
 before `sys.boot_completed` is ever set. Nothing in the guest — system image variant,
 RAM, disk size, ATD versus `google_apis` — can influence a host-side `mprotect` denial,
 so none of those axes was varied. (The API 37 failure in
-[`api-37-emulator-crash.md`](api-37-emulator-crash.md) is genuinely guest-side and
-genuinely unrelated: there the host emulator survives and the guest's `surfaceflinger`
-aborts.)
+[`api-37-emulator-crash.md`](api-37-emulator-crash.md) is genuinely guest-side — there the host
+emulator survives and the guest's `surfaceflinger` aborts — but it is **not** unrelated, as this
+paragraph originally claimed. Both are decided by the renderer, in opposite directions: below 37
+you must avoid SwiftShader GLES and `-gpu host` is the answer; at 37 you must avoid the *host* GL
+translator and `-gpu host` is the thing that never boots.)

 **Turning the SELinux boolean on** — deliberately *not* done, though it would almost
 certainly work:
@@ -401,8 +429,13 @@ are easy to forget to look at.
  "SwiftShader 4.0.0.1" as reported by the GLES translator.
 - **If `-gpu host` regresses** after a Mesa or kernel update, fall back to
  `GPU_MODE=swangle_indirect`, which needs no GPU at all.
- **This changes nothing about API 37.** That image is broken for a different reason and
-  still must be checked on the physical Pixel 10 Pro XL before each release.
+- **API 37 needs the opposite renderer, and this file used to say it needed nothing.** The
+  original bullet here read "This changes nothing about API 37"; that turned out to be wrong.
+  The API 37 images abort `surfaceflinger` under the *host* GL translator and boot under ANGLE —
+  the exact mirror of the rule above — and `run-e2e.sh` therefore picks the renderer per API
+  level. See [`api-37-emulator-crash.md`](api-37-emulator-crash.md), which was rewritten on
+  2026-08-22 with the seven-run matrix. API 37 still must be checked on the physical Pixel 10 Pro
+  XL before each release.

 ## Correction owed to `CLAUDE.md`

@@ -432,12 +465,14 @@ Proposed replacement for the section, offered for review rather than applied her
 >   `auto` (the default), `off` and `guest` do when headless. `-gpu host` works, and the
 >   harness both picks it and refuses the others. `docs/local-emulator.md` has the
 >   backtrace and the mode matrix.
-> - **The API 37 image is broken.** `android-37.0` crash-loops surfaceflinger inside its
->   own gralloc mapper, so every test fails there regardless of this app —
->   `docs/api-37-emulator-crash.md` records the evidence and the ruled-out fixes. This is
->   unrelated to the renderer above: it is a guest-side bug that CI hits too, which is why
->   the matrix stops at API 36 even though targetSdk is 37. **API 37 needs a manual check
->   on the Pixel 10 Pro XL before each release.**
+> - **API 37 needs the opposite renderer, and SystemUI turned off.** Both `android-37.0` and
+>   `android-37.1` abort surfaceflinger inside their own gralloc mapper, and init SIGKILLs
+>   zygote each time. Under `-gpu host` they never boot; under `-gpu swangle_indirect` they
+>   boot, and disabling SystemUI removes the trigger. `run-e2e.sh` does all of that per level,
+>   and the local API 37 result is two failures and the two usual skips, not a clean run. CI
+>   takes API 37 as a gating leg plus an advisory one carrying those two tests.
+>   `docs/api-37-emulator-crash.md` has the matrix and the reasoning. **API 37 needs a manual
+>   check on the Pixel 10 Pro XL before each release.**

 The wording is worth getting right rather than merely correcting, because the original was
 not a careless sentence — it was a reasonable inference from three crashes, written down
@@ -80,6 +80,17 @@ jacoco = "0.8.15"
 # 4.16.1 is the newest RELEASED version; the 4.17 line is beta-only at the time of writing.
 robolectric = "4.16.1"

+# kotlinx-coroutines-test. Already on the unit-test classpath transitively, through
+# compose-ui-test-junit4 -- declared here because a source file now imports it, and a direct
+# import of a transitive is a dependency nobody chose.
+#
+# PINNED, for the same reason as robolectric above: org.jetbrains.kotlinx is not one of the
+# groups in the prerelease guard's `floatedGroupPrefixes`, so a "1.+" here would be free to
+# resolve to a milestone build. This value is what the Compose BOM already resolves it to, so
+# stating it changes nothing in the graph today; if the BOM moves ahead, Gradle takes the
+# higher version and this stays a floor rather than a conflict.
+coroutinesTest = "1.9.0"
+
 [libraries]
 androidx-core-ktx = { group = "androidx.core", name = "core-ktx", version.ref = "coreKtx" }
 androidx-activity-compose = { group = "androidx.activity", name = "activity-compose", version.ref = "activityCompose" }
@@ -134,6 +145,10 @@ androidx-espresso-core = { group = "androidx.test.espresso", name = "espresso-co
 # a TDD loop anyone here can execute.
 robolectric = { group = "org.robolectric", name = "robolectric", version.ref = "robolectric" }

+# Only for its `runTest`, and only to drain the collector kotlinx-coroutines-test installs
+# process-wide. See EscapedCoroutineErrors.kt in the JVM test source set.
+kotlinx-coroutines-test = { group = "org.jetbrains.kotlinx", name = "kotlinx-coroutines-test", version.ref = "coroutinesTest" }
+
 [plugins]
 # com.android.application and org.jetbrains.kotlin.plugin.compose are deliberately absent.
 # They come from the root buildscript classpath (see build.gradle.kts) so that a newer KGP
@@ -0,0 +1,253 @@
+#!/usr/bin/env bash
+#
+# Files a GitHub issue AND puts it on the project board, as one operation.
+#
+# Usage:  tools/github/file-issue.sh --title TITLE (--body TEXT | --body-file PATH) [options]
+#
+#   --status NAME     board column, matched case-insensitively against the board's own
+#                     options; a miss lists what is available. Default: Backlog
+#   --label NAME      repeatable. Passed through to `gh issue create` unchanged.
+#   --project N       project number. Default: $ISSUE_PROJECT_NUMBER, else 6
+#   --repo OWNER/NAME default: whatever `gh repo view` resolves in the working directory
+#   --dry-run         resolve and validate everything, create nothing
+#
+# EXIT CODE: 0 only when the issue exists, is on the board, AND reads back carrying the
+# Status that was asked for. 2 for a usage or validation error, before anything is created.
+# **3 means the issue was created but did not reach the board** -- the number is printed on
+# a line of its own, because that combination is the entire failure this script exists to
+# prevent and it must never be quiet.
+#
+# WHY THIS EXISTS
+#
+# `gh issue create` does not touch the project board. The issue is created, carries its
+# labels, and is invisible in the Kanban -- which looks exactly like a ticket nobody filed.
+# Measured 2026-08-24: eight issues filed as a scripted batch all reached the board; one
+# filed as a one-off a few minutes later did not, and was caught only because someone went
+# looking. A batch carries the board step inside its loop. One-offs are where it slips, so
+# one-offs are what this is for.
+#
+# Adding an item and setting a field value are GraphQL-only. REST can list project items
+# and field definitions, but the `fields` array it returns on an item carries Title and
+# nothing else -- a REST-only check reports every item's Status as unset, which is why the
+# read-back at the end is a GraphQL query rather than the cheaper REST one.
+#
+# WHAT IT DELIBERATELY DOES NOT DO
+#
+# It does not cache the project, field or option ids. Resolving them by name costs one
+# GraphQL query per run, and it means a renamed or reordered column cannot make this write
+# a stale id. The ids are the fragile part; the names are what people actually use.
+#
+# It does not apply triage labels for you. `above-cut` and `backlog` are labels from one
+# specific 2026-08-22 triage pass -- they mean "worked autonomously overnight" and "held for
+# manual review", not "this is in the Backlog column". Status carries board state. Pass
+# --label only for things that are true about the issue itself.
+#
+# It does not create the project, the Status field, or a missing option. Anything absent is
+# an error to report, not to invent.
+
+set -euo pipefail
+
+readonly EXIT_USAGE=2
+readonly EXIT_ORPHANED=3
+
+die() {
+    printf 'file-issue: %s\n' "$1" >&2
+    exit "${2:-$EXIT_USAGE}"
+}
+
+title=""
+body=""
+body_file=""
+status="Backlog"
+project="${ISSUE_PROJECT_NUMBER:-6}"
+repo=""
+dry_run=0
+labels=()
+
+while [ $# -gt 0 ]; do
+    case "$1" in
+        --title)     [ $# -ge 2 ] || die "--title needs a value"; title="$2"; shift 2 ;;
+        --body)      [ $# -ge 2 ] || die "--body needs a value"; body="$2"; shift 2 ;;
+        --body-file) [ $# -ge 2 ] || die "--body-file needs a path"; body_file="$2"; shift 2 ;;
+        --status)    [ $# -ge 2 ] || die "--status needs a value"; status="$2"; shift 2 ;;
+        --label)     [ $# -ge 2 ] || die "--label needs a value"; labels+=("$2"); shift 2 ;;
+        --project)   [ $# -ge 2 ] || die "--project needs a number"; project="$2"; shift 2 ;;
+        --repo)      [ $# -ge 2 ] || die "--repo needs OWNER/NAME"; repo="$2"; shift 2 ;;
+        --dry-run)   dry_run=1; shift ;;
+        -h|--help)   awk 'NR > 1 && /^#/ { sub(/^# ?/, ""); print; next } NR > 1 { exit }' "$0"
+                     exit 0 ;;
+        *)           die "unknown argument: $1" ;;
+    esac
+done
+
+[ -n "$title" ] || die "--title is required"
+if [ -n "$body" ] && [ -n "$body_file" ]; then
+    die "pass --body or --body-file, not both"
+fi
+[ -n "$body" ] || [ -n "$body_file" ] || die "one of --body or --body-file is required"
+if [ -n "$body_file" ] && [ ! -r "$body_file" ]; then
+    die "--body-file is not readable: $body_file"
+fi
+case "$project" in
+    ''|*[!0-9]*) die "--project must be a number, got: $project" ;;
+esac
+
+command -v gh >/dev/null 2>&1 || die "gh is not on PATH"
+
+if [ -z "$repo" ]; then
+    repo=$(gh repo view --json nameWithOwner --jq '.nameWithOwner') \
+        || die "could not resolve the repository; pass --repo OWNER/NAME"
+fi
+owner="${repo%%/*}"
+[ -n "$owner" ] || die "could not read an owner out of: $repo"
+
+# ---------------------------------------------------------------------------
+# Resolve the board by NAME. Every id below is read fresh; none is hardcoded.
+# ---------------------------------------------------------------------------
+
+# The $names in the query are GraphQL variables, declared by the query and bound by the
+# -f flags. Expanding them in the shell would send this shell's idea of $owner to the
+# API instead of declaring a parameter -- which is why every query here is single-quoted.
+# shellcheck disable=SC2016
+board=$(gh api graphql \
+    -f query='
+        query($owner: String!, $number: Int!) {
+          user(login: $owner) {
+            projectV2(number: $number) {
+              id
+              title
+              field(name: "Status") {
+                ... on ProjectV2SingleSelectField { id options { id name } }
+              }
+            }
+          }
+        }' \
+    -f owner="$owner" -F number="$project" 2>&1) \
+    || die "could not read project $project for $owner. A 403 naming scopes means gh is
+missing 'project'; a 403 naming a rate limit is the GraphQL budget, not permissions. The
+API said: $board"
+
+project_id=$(printf '%s' "$board" | jq -r '.data.user.projectV2.id // empty')
+field_id=$(printf '%s' "$board" | jq -r '.data.user.projectV2.field.id // empty')
+project_title=$(printf '%s' "$board" | jq -r '.data.user.projectV2.title // empty')
+
+[ -n "$project_id" ] || die "no project number $project under user $owner"
+[ -n "$field_id" ] || die "project $project has no single-select field named 'Status'"
+
+# Case-insensitive match, so "backlog" and "Backlog" both work. The canonical name is
+# what gets reported back, so a sloppy argument still produces an exact log line.
+option=$(printf '%s' "$board" | jq -r --arg want "$status" '
+    .data.user.projectV2.field.options[]
+    | select((.name | ascii_downcase) == ($want | ascii_downcase))
+    | "\(.id)\t\(.name)"' | head -n 1)
+
+if [ -z "$option" ]; then
+    printf 'file-issue: no Status option named %s. Available:\n' "$status" >&2
+    printf '%s' "$board" | jq -r '.data.user.projectV2.field.options[] | "  " + .name' >&2
+    exit "$EXIT_USAGE"
+fi
+option_id="${option%%$'\t'*}"
+status_canonical="${option#*$'\t'}"
+
+printf 'repo     %s\n' "$repo"
+printf 'board    %s (project %s)\n' "$project_title" "$project"
+printf 'status   %s\n' "$status_canonical"
+printf 'labels   %s\n' "${labels[*]:-(none)}"
+printf 'title    %s\n' "$title"
+
+if [ "$dry_run" -eq 1 ]; then
+    printf '\ndry run: everything above resolved; nothing was created.\n'
+    exit 0
+fi
+
+# ---------------------------------------------------------------------------
+# Create. Past this line a failure can leave an issue off the board, so every
+# error path prints the number.
+# ---------------------------------------------------------------------------
+
+create_args=(--repo "$repo" --title "$title")
+if [ -n "$body_file" ]; then
+    create_args+=(--body-file "$body_file")
+else
+    create_args+=(--body "$body")
+fi
+for label in ${labels[@]+"${labels[@]}"}; do
+    create_args+=(--label "$label")
+done
+
+issue_url=$(gh issue create "${create_args[@]}") || die "gh issue create failed; nothing was filed"
+issue_number="${issue_url##*/}"
+case "$issue_number" in
+    ''|*[!0-9]*) die "could not read an issue number out of: $issue_url" ;;
+esac
+
+orphaned() {
+    printf 'file-issue: %s\n' "$1" >&2
+    printf 'file-issue: THE ISSUE EXISTS BUT IS NOT ON THE BOARD. Fix it by hand:\n' >&2
+    printf '%s\n' "$issue_url" >&2
+    exit "$EXIT_ORPHANED"
+}
+
+content_id=$(gh api "/repos/$repo/issues/$issue_number" --jq '.node_id') \
+    || orphaned "could not read the node id for #$issue_number"
+
+# shellcheck disable=SC2016  # GraphQL variables, as above
+item_id=$(gh api graphql \
+    -f query='
+        mutation($project: ID!, $content: ID!) {
+          addProjectV2ItemById(input: {projectId: $project, contentId: $content}) {
+            item { id }
+          }
+        }' \
+    -f project="$project_id" -f content="$content_id" \
+    --jq '.data.addProjectV2ItemById.item.id') \
+    || orphaned "could not add #$issue_number to the board"
+[ -n "$item_id" ] || orphaned "the board add returned no item id for #$issue_number"
+
+# shellcheck disable=SC2016  # GraphQL variables, as above
+gh api graphql \
+    -f query='
+        mutation($project: ID!, $item: ID!, $field: ID!, $option: String!) {
+          updateProjectV2ItemFieldValue(input: {
+            projectId: $project, itemId: $item, fieldId: $field,
+            value: {singleSelectOptionId: $option}
+          }) { projectV2Item { id } }
+        }' \
+    -f project="$project_id" -f item="$item_id" -f field="$field_id" -f option="$option_id" \
+    >/dev/null \
+    || orphaned "#$issue_number is on the board but its Status could not be set"
+
+# ---------------------------------------------------------------------------
+# Read back. A mutation returning 200 is not evidence the board shows what was
+# asked for -- this is the only check that is.
+# ---------------------------------------------------------------------------
+
+# shellcheck disable=SC2016  # GraphQL variables, as above
+readback=$(gh api graphql \
+    -f query='
+        query($item: ID!) {
+          node(id: $item) {
+            ... on ProjectV2Item {
+              content { ... on Issue { number } }
+              fieldValueByName(name: "Status") {
+                ... on ProjectV2ItemFieldSingleSelectValue { name }
+              }
+            }
+          }
+        }' \
+    -f item="$item_id") \
+    || orphaned "#$issue_number was written but could not be read back"
+
+seen_number=$(printf '%s' "$readback" | jq -r '.data.node.content.number // empty')
+seen_status=$(printf '%s' "$readback" | jq -r '.data.node.fieldValueByName.name // empty')
+
+if [ "$seen_number" != "$issue_number" ]; then
+    orphaned "read-back names issue #${seen_number:-<none>}, expected #$issue_number"
+fi
+if [ "$seen_status" != "$status_canonical" ]; then
+    orphaned "read-back Status is ${seen_status:-<unset>}, expected $status_canonical"
+fi
+
+printf '\n#%s on %s as %s -- verified by read-back\n' \
+    "$issue_number" "$project_title" "$seen_status"
+printf '%s\n' "$issue_url"
@@ -3,13 +3,24 @@
 # Runs the instrumented suite on a local emulator, on this workstation, for one or more
 # API levels.
 #
-# Usage:  tools/local-emulator/run-e2e.sh [API ...]        # default: 33 34 35 36
+# Usage:  tools/local-emulator/run-e2e.sh [API ...]        # default: 33 34 35 36 37
 #
-#   GPU_MODE=host          renderer to use; see the refusal list below
+# API levels are the labels below, not SDK ints: 33-36, plus `37` (= `37.0`) and `37.1`.
+#
+#   GPU_MODE=              force one renderer on every level; unset means per-API (gpu_for_api)
 #   EMULATOR_PORT=5560     console port, so the serial is deterministic
 #   BOOT_TIMEOUT=300       seconds to wait for sys.boot_completed
 #   KEEP_AVD=1             do not delete an AVD this script created
 #
+# EXIT CODE: 0 only if every level was green; 1 if any level failed, wedged or could not be
+# set up; 2 if it refused to start at all. **A bare `run-e2e.sh` therefore exits 1 by design.**
+# API 37 is in the default list on purpose -- leaving it out is what left the level unlooked-at
+# for as long as it was -- and it is permanently two failures short of green, on the emulator's
+# own c2.goldfish.h264.decoder rather than on anything this app does. The summary names the two,
+# so a third is visibly new, and the last line printed says the same thing. Anything that reads a
+# non-zero exit as breakage should name the levels it wants: `run-e2e.sh 33 34 35 36` is the
+# sweep that can be green. docs/api-37-emulator-crash.md has the measurements.
+#
 # WHY THIS EXISTS, AND WHAT IT DELIBERATELY DOES NOT DO
 #
 # It is a *launcher*, not a second test harness. The diagnostics -- the FAILED-vs-WEDGED
@@ -34,6 +45,14 @@
 # those modes was measured crashing. docs/local-emulator.md has the backtrace, the faulting
 # page's RW-without-E segment flags, and the full mode matrix.
 #
+# AND THE ONE THING API 37 NEEDS THAT 33-36 DO NOT: the opposite renderer. On the API 37
+# images the guest's Gralloc5 mapper aborts surfaceflinger from RegionSamplingThread
+# (`Assertion failed: !rcEnc->featureInfo()->hasReadColorBufferDma`). Under `-gpu host` that
+# repeats every few seconds and the device never boots; under ANGLE it fires a handful of
+# times and the boot survives. So `host` is required below 37 and forbidden at 37, which is
+# why the renderer is chosen per level in gpu_for_api rather than set once.
+# docs/api-37-emulator-crash.md has that matrix.
+#
 # THE OTHER LOCAL-ONLY HAZARD: a physical Pixel is usually plugged into this machine, so
 # `adb` is ambiguous in a way it never is on a runner, and an unpinned run would install
 # and execute this suite on the phone. Every path below pins the emulator serial.
@@ -65,12 +84,14 @@ export ANDROID_HOME="${ANDROID_HOME:-$HOME/Android/Sdk}"
 export ANDROID_SDK_ROOT="$ANDROID_HOME"
 export PATH="$ANDROID_HOME/platform-tools:$ANDROID_HOME/emulator:$ANDROID_HOME/cmdline-tools/latest/bin:$PATH"

-GPU_MODE="${GPU_MODE:-host}"
+# Empty means "let each level pick" -- see gpu_for_api. Setting GPU_MODE forces one renderer
+# on every level, which is what you want when measuring a mode, not when running the suite.
+GPU_MODE="${GPU_MODE:-}"
 EMULATOR_PORT="${EMULATOR_PORT:-5560}"
 BOOT_TIMEOUT="${BOOT_TIMEOUT:-300}"
 SERIAL="emulator-${EMULATOR_PORT}"
 APIS=("$@")
-[ "${#APIS[@]}" -eq 0 ] && APIS=(33 34 35 36)
+[ "${#APIS[@]}" -eq 0 ] && APIS=(33 34 35 36 37)

 # Matches CI. `disk-size: 8G` because the FFmpeg libraries do not fit the default userdata
 # partition; `ram-size: 2560M` because the emulator's own floor varies by API level and
@@ -83,21 +104,28 @@ RESULTS_DIR="app/build/outputs/androidTest-results"
 LOG_DIR="${TMPDIR:-/tmp}/lmc-local-e2e"
 mkdir -p "$LOG_DIR"

+# The two things that outlive a level, declared here rather than where they are first
+# assigned, because the cleanup trap below can fire before either has been reached.
+EMU_PID=""
+CREATED_AVDS=()
+
 # ---------------------------------------------------------------- renderer preflight ---
-case "$GPU_MODE" in
-  swiftshader_indirect | auto | off | guest)
-    echo "REFUSING to launch with -gpu $GPU_MODE."
-    echo "On this host that resolves to SwiftShader's GLES, whose JIT is denied execheap by"
-    echo "SELinux; the emulator segfaults (exit 139) before boot. See docs/local-emulator.md."
-    echo "Working modes: host (default), angle_indirect, swangle_indirect."
-    exit 2
-    ;;
-  host | angle_indirect | swangle_indirect) ;;
-  *)
-    echo "Unrecognised GPU_MODE '$GPU_MODE'. Known-good: host, angle_indirect, swangle_indirect."
-    exit 2
-    ;;
-esac
+if [ -n "$GPU_MODE" ]; then
+  case "$GPU_MODE" in
+    swiftshader_indirect | auto | off | guest)
+      echo "REFUSING to launch with -gpu $GPU_MODE."
+      echo "On this host that resolves to SwiftShader's GLES, whose JIT is denied execheap by"
+      echo "SELinux; the emulator segfaults (exit 139) before boot. See docs/local-emulator.md."
+      echo "Working modes: host, angle_indirect, swangle_indirect."
+      exit 2
+      ;;
+    host | angle_indirect | swangle_indirect) ;;
+    *)
+      echo "Unrecognised GPU_MODE '$GPU_MODE'. Known-good: host, angle_indirect, swangle_indirect."
+      exit 2
+      ;;
+  esac
+fi

 # A courtesy, not a gate: the boolean being on means SwiftShader would work too, and the
 # refusal list above could be relaxed. It is off on a stock Fedora.
@@ -121,14 +149,90 @@ host_forensics() {
  journalctl --since "$since" --no-pager 2> /dev/null | grep -E 'avc: .*denied' | tail -10 || echo "  (none)"
 }

+# The API 37 counterpart of host_forensics. `-gpu host` there aborts surfaceflinger in a loop
+# and the device never boots; the working renderers abort it a few times and survive. Either way
+# the count is the number to look at, and the crash buffer is where it lives -- so print it on
+# every 37 level, not only on the failure path, because a level that passed with 40 aborts is
+# telling you something a level that passed with 1 is not.
+guest_forensics() {
+  local api="$1" n
+  case "$api" in 37 | 37.*) ;; *) return 0 ;; esac
+  n="$(emu_adb logcat -d -b crash 2> /dev/null | grep -c 'hasReadColorBufferDma')"
+  echo "  surfaceflinger hasReadColorBufferDma aborts: ${n:-?}  (docs/api-37-emulator-crash.md)"
+}
+
+# API label -> system image. API 33-36 are plain integers with a `google_apis` image. API 37
+# is not: its SDK directories are dotted minor versions (`android-37.0`, `android-37.1`), there
+# is no `android-37`, and from 37.1 onwards Google ships only 16 KB-page (`ps16k`) images for
+# x86_64. `37` is accepted as a spelling of `37.0` because that is what people type.
+image_pkg_for_api() {
+  case "$1" in
+    37 | 37.0) echo "system-images;android-37.0;google_apis;x86_64" ;;
+    37.1) echo "system-images;android-37.1;google_apis_ps16k;x86_64" ;;
+    *) echo "system-images;android-$1;google_apis;x86_64" ;;
+  esac
+}
+
+# The renderer requirement is per-API and the two levels want OPPOSITE things, which is why this
+# is a function and not a constant.
+#
+#   33-36: must NOT be SwiftShader GLES (host-side SELinux/execheap segfault) -- `host` is right.
+#   37.x:  must NOT be the host GL translator. With `-gpu host` the guest's Gralloc5 mapper
+#          aborts surfaceflinger in a loop and the device never boots; under ANGLE the same
+#          assertion fires a handful of times and the boot survives it. Measured, not guessed --
+#          docs/api-37-emulator-crash.md has the matrix.
+#
+# `swangle_indirect` rather than `angle_indirect` for 37: both boot, and swangle names its
+# renderer outright instead of resolving through `auto`'s path.
+gpu_for_api() {
+  if [ -n "$GPU_MODE" ]; then
+    echo "$GPU_MODE"
+    return
+  fi
+  case "$1" in
+    37 | 37.*) echo "swangle_indirect" ;;
+    *) echo "host" ;;
+  esac
+}
+
+# `lmc_e2e_api37.0` would be a legal AVD name but an awkward one to type and to grep for.
+# `37` and `37.0` therefore give two AVD names (`lmc_e2e_api37`, `lmc_e2e_api37_0`) for the one
+# image. Harmless -- two AVDs off the same system image cost only disk -- and deliberately not
+# normalised, so that `run-e2e.sh 37 37.0` does not have both levels fight over one AVD.
+avd_for_api() { echo "lmc_e2e_api${1//./_}"; }
+
+# Where avdmanager actually put the AVD. `$HOME/.android/avd` is only the default:
+# ANDROID_AVD_HOME, ANDROID_USER_HOME and ANDROID_SDK_HOME each move it, and hardcoding the
+# default meant a machine that sets any of them silently ran every level at stock RAM and
+# userdata size. Rather than encode a precedence that cannot be verified from here, look in
+# every location avdmanager honours and let the existence check pick.
+avd_config_path() {
+  local avd="$1" base cfg
+  for base in "${ANDROID_AVD_HOME:-}" \
+    "${ANDROID_USER_HOME:+$ANDROID_USER_HOME/avd}" \
+    "${ANDROID_SDK_HOME:+$ANDROID_SDK_HOME/.android/avd}" \
+    "$HOME/.android/avd"; do
+    [ -n "$base" ] || continue
+    cfg="$base/${avd}.avd/config.ini"
+    if [ -f "$cfg" ]; then
+      echo "$cfg"
+      return 0
+    fi
+  done
+  return 1
+}
+
 ensure_avd() {
  local api="$1" avd="$2"
-  local pkg="system-images;android-${api};google_apis;x86_64"
+  local pkg
+  pkg="$(image_pkg_for_api "$api")"
+  local img_dir="$ANDROID_HOME/system-images/${pkg#system-images;}"
+  img_dir="${img_dir//;//}"

  if avdmanager list avd -c 2> /dev/null | grep -qx "$avd"; then
    echo "  reusing existing AVD $avd"
  else
-    if [ ! -d "$ANDROID_HOME/system-images/android-${api}/google_apis/x86_64" ]; then
+    if [ ! -d "$img_dir" ]; then
      echo "  installing $pkg"
      yes | sdkmanager --install "$pkg" > /dev/null 2>&1 || {
        echo "  FAILED to install $pkg"
@@ -144,18 +248,37 @@ ensure_avd() {
  fi

  # Written into config.ini rather than passed on the command line, which is how
-  # reactivecircus/android-emulator-runner applies the same two settings in CI.
-  local cfg="$HOME/.android/avd/${avd}.avd/config.ini"
-  sed -i -e '/^disk\.dataPartition\.size=/d' -e '/^hw\.ramSize=/d' "$cfg"
-  printf 'disk.dataPartition.size=%s\nhw.ramSize=%s\n' "$DISK_SIZE_BYTES" "$RAM_SIZE_MB" >> "$cfg"
+  # reactivecircus/android-emulator-runner applies the same two settings in CI. On the reuse
+  # path too, so an AVD left over from an older run gets today's pins.
+  #
+  # A level that cannot be pinned FAILS rather than running at the defaults. Unpinned, it
+  # dies much later with "not enough space", which reads as a device problem -- CI's own
+  # history is where that lesson comes from -- and nothing points back to a `sed` that
+  # edited a path this script guessed wrong.
+  local cfg
+  if ! cfg="$(avd_config_path "$avd")"; then
+    echo "  FAILED: no config.ini for $avd in any directory avdmanager uses"
+    echo "  (ANDROID_AVD_HOME=${ANDROID_AVD_HOME:-unset}, ANDROID_USER_HOME=${ANDROID_USER_HOME:-unset},"
+    echo "   ANDROID_SDK_HOME=${ANDROID_SDK_HOME:-unset}, HOME=$HOME)"
+    return 1
+  fi
+  if ! sed -i -e '/^disk\.dataPartition\.size=/d' -e '/^hw\.ramSize=/d' "$cfg"; then
+    echo "  FAILED to rewrite $cfg"
+    return 1
+  fi
+  if ! printf 'disk.dataPartition.size=%s\nhw.ramSize=%s\n' \
+    "$DISK_SIZE_BYTES" "$RAM_SIZE_MB" >> "$cfg"; then
+    echo "  FAILED to write the RAM/disk pins into $cfg"
+    return 1
+  fi
 }

 boot_emulator() {
-  local avd="$1" api="$2"
+  local avd="$1" api="$2" gpu="$3"
  local boot_log="$LOG_DIR/emulator-api${api}.log"

  emulator -avd "$avd" -port "$EMULATOR_PORT" \
-    -no-window -gpu "$GPU_MODE" -noaudio -no-boot-anim -camera-back none -no-snapshot \
+    -no-window -gpu "$gpu" -noaudio -no-boot-anim -camera-back none -no-snapshot \
    > "$boot_log" 2>&1 &
  EMU_PID=$!

@@ -181,6 +304,92 @@ boot_emulator() {
  return 1
 }

+# API 37 only, and the reason API 37 can be run at all.
+#
+# The abort that breaks these images is reached from SurfaceFlinger's RegionSamplingThread,
+# which exists only because SystemUI registers a nav-bar luma-sampling listener. Each abort
+# kills surfaceflinger, and init responds by SIGKILLing zygote -- so the whole framework
+# restarts underneath the test run, which arrives as `Can't find service: package` and
+# `INSTRUMENTATION_ABORTED: System has crashed`. Under the host GL renderer that repeats
+# forever; under ANGLE it is roughly one every fifteen seconds, which a five-minute suite does
+# not survive either.
+#
+# Removing the listener removes the whole chain. Measured on android-37.0 under
+# swangle_indirect: 10-11 aborts per 150 s idle with SystemUI running, and 0 in 180 s with it
+# disabled, framework services up throughout.
+#
+# THIS IS A DEVIATION, and it is deliberately loud rather than silent. The API 37 leg does not
+# run the same device configuration as API 33-36 or as the Pixel. It is defensible only
+# because nothing in this suite touches SystemUI -- these are Media3, FFmpeg and WorkManager
+# tests -- and because the alternative is no API 37 coverage at all. Anything that ever does
+# depend on system UI must not trust this leg. docs/api-37-emulator-crash.md explains why.
+#
+# The retry loop is not defensive padding: at the moment boot_completed flips, the framework
+# may be in one of its restarts and `pm` is simply not published yet. The first attempt at this
+# 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.
+disable_region_sampling() {
+  local api="$1" out i before after ready
+  case "$api" in 37 | 37.*) ;; *) return 0 ;; esac
+
+  out=""
+  for i in $(seq 1 20); do
+    out="$(emu_adb shell pm disable-user --user 0 com.android.systemui 2>&1 | tr -d '\r')"
+    case "$out" in
+      *"new state: disabled"*)
+        echo "  SystemUI disabled on attempt $i"
+        break
+        ;;
+    esac
+    out=""
+    sleep 5
+  done
+  if [ -z "$out" ]; then
+    echo "  WARNING: could not disable SystemUI after 20 attempts."
+    echo "  Expect INSTRUMENTATION_ABORTED -- docs/api-37-emulator-crash.md"
+    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.
+  ready=0
+  for i in $(seq 1 30); do
+    if emu_adb shell service check package 2> /dev/null | grep -q ': found' \
+      && emu_adb shell service check activity 2> /dev/null | grep -q ': found'; then
+      ready=1
+      break
+    fi
+    sleep 5
+  done
+  if [ "$ready" -ne 1 ]; then
+    echo "  WARNING: package and activity services still absent 150 s after the restart."
+    echo "  Expect INSTRUMENTATION_ABORTED -- docs/api-37-emulator-crash.md"
+  fi
+
+  # Prove it worked rather than assume it. Zero new aborts over this window is what makes the
+  # difference between a run that completes and one that reports `Starting 0 tests`.
+  before="$(emu_adb logcat -d -b crash 2> /dev/null | grep -c 'hasReadColorBufferDma')"
+  emu_adb shell 'sleep 45' > /dev/null 2>&1
+  after="$(emu_adb logcat -d -b crash 2> /dev/null | grep -c 'hasReadColorBufferDma')"
+  echo "  quiet check: $((after - before)) new surfaceflinger aborts in 45 s (want 0)"
+  if [ "$((after - before))" -ne 0 ]; then
+    echo "  WARNING: region sampling is still live; the run may not survive."
+  fi
+  return 0
+}
+
 # CI gets this from the action's `disable-animations: true`.
 disable_animations() {
  local s
@@ -189,17 +398,81 @@ disable_animations() {
  done
 }

+# `${EMU_PID:-0}` used to guard these three calls, and it guarded the wrong thing: EMU_PID
+# is *empty*, not unset, if the background launch never produced a job, and `kill` reads pid
+# 0 as "the sender's whole process group" -- this script and, on a terminal, everything else
+# in the foreground group with it. The `kill -0` wait loop had the same shape and would have
+# spent its full grace period testing the group. Nothing to stop is now a return, never a
+# guess. (boot_emulator's own `kill -0 "$EMU_PID"` is unguarded and cannot reach that form:
+# it runs only after the assignment.)
+#
+# max_wait is a parameter so the interrupt path need not sit through the full grace period.
 stop_emulator() {
+  local max_wait="${1:-30}" waited=0
+  [ -n "${EMU_PID:-}" ] || return 0
  emu_adb emu kill > /dev/null 2>&1
-  local waited=0
-  while kill -0 "${EMU_PID:-0}" 2> /dev/null && [ "$waited" -lt 30 ]; do
+  while kill -0 "$EMU_PID" 2> /dev/null && [ "$waited" -lt "$max_wait" ]; do
    sleep 2
    waited=$((waited + 2))
  done
-  kill -9 "${EMU_PID:-0}" 2> /dev/null
-  wait "${EMU_PID:-0}" 2> /dev/null
+  kill -9 "$EMU_PID" 2> /dev/null
+  wait "$EMU_PID" 2> /dev/null
+  EMU_PID=""
 }

+delete_created_avds() {
+  local avd
+  [ "${KEEP_AVD:-0}" = "1" ] && return 0
+  for avd in ${CREATED_AVDS[@]+"${CREATED_AVDS[@]}"}; do
+    # A SIGKILLed emulator does not get to remove its own lock files, and avdmanager can
+    # refuse over them. Staying silent there would leak the very thing this exists to clean.
+    avdmanager delete avd -n "$avd" > /dev/null 2>&1 \
+      || echo "  WARNING: could not delete AVD $avd -- 'avdmanager delete avd -n $avd' by hand"
+  done
+  CREATED_AVDS=()
+}
+
+# What an interrupted sweep used to leave behind: a headless emulator holding console port
+# $EMULATOR_PORT, and an lmc_e2e_apiNN AVD. The next run's `emulator -port` then collides
+# with the orphan, and `emu_adb` can resolve to it -- on a workstation that also has the
+# Pixel plugged in, exactly the ambiguity the ANDROID_SERIAL pinning exists to prevent. A
+# sweep is up to five boots long, so the window for one Ctrl-C is not small.
+#
+# Idempotent, and called explicitly on the normal path so its output cannot land after the
+# summary; the EXIT trap then finds nothing left to do. The emulator logs are deliberately
+# NOT removed -- they live in $LOG_DIR and are the only evidence a failed boot leaves.
+CLEANED=0
+cleanup() {
+  [ "$CLEANED" = "1" ] && return 0
+  CLEANED=1
+  stop_emulator "${1:-30}"
+  delete_created_avds
+}
+
+# 6 s, not 30: Ctrl-C has already reached the emulator through the foreground process group,
+# so this is only waiting for it to finish writing, and `kill -9` follows regardless. The
+# EXIT trap is disarmed before exiting so the status below is the one that survives.
+#
+# bash runs a trap only between commands, so this starts when whatever was in the foreground
+# returns -- which for Ctrl-C is immediately, because the same interrupt reached that command
+# too. `kill -INT` aimed at this script alone waits for the foreground command to finish.
+# Invoked indirectly -- installed as the INT and TERM trap a few lines below. Both codes,
+# because shellcheck 0.9.0 reports this as unreachable commands (SC2317) and 0.11.0 as an
+# uninvoked function (SC2329); CI pins 0.11.0 but a local install may be either.
+# shellcheck disable=SC2317,SC2329
+on_signal() {
+  echo
+  echo "interrupted (SIG$1) -- stopping the emulator and removing the AVDs this run created"
+  echo "  emulator logs kept in $LOG_DIR"
+  cleanup 6
+  trap - EXIT
+  exit "$2"
+}
+
+trap 'on_signal INT 130' INT
+trap 'on_signal TERM 143' TERM
+trap cleanup EXIT
+
 # The XML is authoritative. The console counter double-counts skips, so a run that reports
 # "42 tests" on stdout can be 40 in the report.
 #
@@ -236,37 +509,44 @@ PY
 }

 # ------------------------------------------------------------------------------ main ---
-CREATED_AVDS=()
 SUMMARY=()
 overall=0

-for api in "${APIS[@]}"; do
-  if [ "$api" = "37" ] || [ "$api" = "37.0" ]; then
-    echo "SKIPPING API $api: the android-37.0 image crash-loops surfaceflinger."
-    echo "  See docs/api-37-emulator-crash.md. Test API 37 on the physical Pixel."
-    continue
-  fi
+# Whether the red exit is the expected one depends on which level produced it, and only the
+# loop knows that -- so it is recorded where `overall` is set rather than guessed from the
+# summary afterwards. A note at the end claiming a genuine API 34 failure was "by design"
+# would be the same defect it is there to prevent, one layer up.
+NON37_RED=0
+mark_red() {
+  overall=1
+  case "$1" in 37 | 37.*) ;; *) NON37_RED=1 ;; esac
+}

-  avd="lmc_e2e_api${api}"
+for api in "${APIS[@]}"; do
+  avd="$(avd_for_api "$api")"
+  gpu="$(gpu_for_api "$api")"
  started="$(date '+%Y-%m-%d %H:%M:%S')"
  echo "=============================================================="
-  echo "API $api  (avd=$avd gpu=$GPU_MODE serial=$SERIAL)"
+  echo "API $api  (avd=$avd gpu=$gpu serial=$SERIAL)"
+  echo "  image: $(image_pkg_for_api "$api")"
  echo "=============================================================="

  if ! ensure_avd "$api" "$avd"; then
    SUMMARY+=("API $api: AVD SETUP FAILED")
-    overall=1
+    mark_red "$api"
    continue
  fi

-  if ! boot_emulator "$avd" "$api"; then
+  if ! boot_emulator "$avd" "$api" "$gpu"; then
    host_forensics "$started"
+    guest_forensics "$api"
    SUMMARY+=("API $api: BOOT FAILED")
-    overall=1
+    mark_red "$api"
    stop_emulator
    continue
  fi

+  disable_region_sampling "$api"
  disable_animations
  rm -rf "$RESULTS_DIR"

@@ -281,23 +561,43 @@ for api in "${APIS[@]}"; do
  unset ANDROID_SERIAL E2E_EXTRA_GRADLE_ARGS

  line="$(summarise_results "$api")"
+  guest_forensics "$api"
+  # API 37 is in the default list on purpose, and it is expected to be red. Leaving it out would
+  # put the level back where this whole exercise found it -- untested and unlooked-at -- but a
+  # summary that just says "2 failures" with no explanation trains people to ignore the exit
+  # code. So the row says which two, and a THIRD failure is then obviously new.
+  case "$api" in
+    37 | 37.*)
+      line="$line
+    expected here: 2 failures, both Media3EngineTest, on c2.goldfish.h264.decoder.
+    A third is new -- docs/api-37-emulator-crash.md"
+      ;;
+  esac
  if [ "$rc" -ne 0 ]; then
    line="$line  [gradle exit $rc]"
-    overall=1
+    mark_red "$api"
    host_forensics "$started"
  fi
  SUMMARY+=("$line")
  stop_emulator
 done

-if [ "${KEEP_AVD:-0}" != "1" ]; then
-  for avd in ${CREATED_AVDS[@]+"${CREATED_AVDS[@]}"}; do
-    avdmanager delete avd -n "$avd" > /dev/null 2>&1
-  done
-fi
+cleanup

 echo
 echo "===================== LOCAL E2E SUMMARY ======================"
 printf '%s\n' ${SUMMARY[@]+"${SUMMARY[@]}"}
 echo "=============================================================="
+
+# An unexplained red exit trains people to stop reading exit codes, and this one is expected
+# whenever API 37 is in the sweep -- which the default list makes the common case. Said here
+# rather than only in the docs, because this is where it is actually read. Only when 37.x is
+# the ONLY thing that went red: a note calling a real failure elsewhere "by design" would be
+# worse than no note at all.
+if [ "$overall" -ne 0 ] && [ "$NON37_RED" -eq 0 ]; then
+  echo "note: the only level that went red is API 37, which exits non-zero by design -- it is"
+  echo "      permanently 2 failures short of green. Confirm its row above shows exactly those"
+  echo "      two and nothing else; docs/api-37-emulator-crash.md says why they are the image."
+fi
+
 exit "$overall"