diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index 60536f5..bf2b0f6 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -1,9 +1,13 @@
 name: Build
 
+# Pull requests are covered by status_check.yml, which builds the real FFmpeg AAR and
+# runs the instrumented suite across API 33-37. This workflow keeps the post-merge and
+# release duties, and deliberately does not duplicate PR validation: its unit job falls
+# back to a stub AAR, which is a weaker check than the one status_check.yml performs.
 on:
   push:
     branches: [main]
-  pull_request:
+    tags: ['v*']
 
 jobs:
   test:
diff --git a/.github/workflows/status_check.yml b/.github/workflows/status_check.yml
new file mode 100644
index 0000000..e400e26
--- /dev/null
+++ b/.github/workflows/status_check.yml
@@ -0,0 +1,184 @@
+name: Status check
+
+on:
+  pull_request:
+    branches: [main]
+
+# A newer push to the same PR makes the in-flight run obsolete. Emulator matrices are
+# expensive, so cancel rather than let them pile up.
+concurrency:
+  group: status-check-${{ github.ref }}
+  cancel-in-progress: true
+
+env:
+  # Must match the coordinate app/build.gradle.kts loads from app/libs/.
+  FFMPEG_AAR: ffmpeg-kit-next-8.1.1.aar
+
+jobs:
+  # ---------------------------------------------------------------------------
+  # FFmpeg is not committed: the AAR is ~35 MB of native code, and F-Droid strips
+  # checked-in binaries. Every other job needs it, because the app module compiles
+  # against it and the instrumented tests exercise it for real.
+  #
+  # Building it is a full cross-compile of FFmpeg, x264, x265 and SVT-AV1, so it is
+  # cached on the contents of tools/ffmpeg. That directory pins the upstream tag and
+  # the configure flags, which is exactly what determines the output.
+  # ---------------------------------------------------------------------------
+  ffmpeg:
+    name: FFmpeg AAR
+    runs-on: ubuntu-latest
+    timeout-minutes: 120
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Restore cached AAR
+        id: cache
+        uses: actions/cache@v4
+        with:
+          path: tools/ffmpeg/out
+          key: ffmpeg-aar-${{ hashFiles('tools/ffmpeg/Containerfile', 'tools/ffmpeg/build-ffmpeg.sh') }}
+
+      # The Nix store plus the build tree runs to several gigabytes, which does not fit
+      # alongside the runner's preinstalled toolchains.
+      - name: Free disk space
+        if: steps.cache.outputs.cache-hit != 'true'
+        run: |
+          sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /usr/local/share/boost
+          sudo docker image prune -af || true
+          df -h /
+
+      - name: Build the AAR
+        if: steps.cache.outputs.cache-hit != 'true'
+        working-directory: tools/ffmpeg
+        run: |
+          mkdir -p out
+          docker build -t ffmpeg-kit-builder:ci -f Containerfile .
+          docker run --rm -v "$PWD/out":/work/out ffmpeg-kit-builder:ci full
+
+      - name: Check the AAR is present and plausibly complete
+        run: |
+          AAR=$(find tools/ffmpeg/out -name 'ffmpeg-kit-next*.aar' | head -1)
+          test -n "$AAR" || { echo "::error::no AAR produced"; exit 1; }
+          # A truncated or stub AAR would still satisfy `test -f`, so check it carries
+          # native libraries for both ABIs before letting the matrix depend on it.
+          for abi in arm64-v8a x86_64; do
+            n=$(unzip -l "$AAR" | grep -c "jni/$abi/.*\.so$" || true)
+            echo "  $abi: $n shared libraries"
+            test "$n" -gt 0 || { echo "::error::AAR has no $abi libraries"; exit 1; }
+          done
+          cp "$AAR" "${{ env.FFMPEG_AAR }}"
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: ffmpeg-aar
+          path: ${{ env.FFMPEG_AAR }}
+          retention-days: 1
+
+  # ---------------------------------------------------------------------------
+  # JVM tests: the routing matrix, the FFmpeg argument builder, the concat planner
+  # and the retry rule. These need no device and are the fastest signal on a PR.
+  # ---------------------------------------------------------------------------
+  unit:
+    name: Unit tests
+    runs-on: ubuntu-latest
+    needs: ffmpeg
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/download-artifact@v4
+        with:
+          name: ffmpeg-aar
+          path: app/libs/
+
+      - uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: '17'   # AGP 9 will not run on anything older
+
+      - uses: gradle/actions/setup-gradle@v4
+
+      - run: ./gradlew :app:testDebugUnitTest
+
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: unit-test-report
+          path: app/build/reports/tests/
+
+  # ---------------------------------------------------------------------------
+  # Instrumented tests across the whole supported range. minSdk is 33 and targetSdk
+  # is 37, and the foreground-service type differs across that range -- none below
+  # 34, dataSync at 34, mediaProcessing from 35 -- so a single API level would leave
+  # two thirds of that branch unexercised.
+  # ---------------------------------------------------------------------------
+  instrumented:
+    name: E2E API ${{ matrix.api-level }}
+    runs-on: ubuntu-latest
+    needs: ffmpeg
+    timeout-minutes: 60
+    strategy:
+      # Report every API level rather than stopping at the first red one: knowing
+      # whether a failure is universal or specific to one level is most of the
+      # diagnosis.
+      fail-fast: false
+      matrix:
+        include:
+          - api-level: 33
+            system-image-api-level: 33
+          - api-level: 34
+            system-image-api-level: 34
+          - api-level: 35
+            system-image-api-level: 35
+          - api-level: 36
+            system-image-api-level: 36
+          # API 37 is published as android-37.0, not android-37, so the image level
+          # has to be given separately or the download resolves to nothing.
+          - api-level: 37
+            system-image-api-level: "37.0"
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/download-artifact@v4
+        with:
+          name: ffmpeg-aar
+          path: app/libs/
+
+      - uses: actions/setup-java@v4
+        with:
+          distribution: temurin
+          java-version: '17'
+
+      - uses: gradle/actions/setup-gradle@v4
+
+      # 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
+
+      - name: Instrumented tests
+        uses: reactivecircus/android-emulator-runner@v2
+        with:
+          api-level: ${{ matrix.api-level }}
+          system-image-api-level: ${{ matrix.system-image-api-level }}
+          target: google_apis
+          arch: x86_64
+          profile: pixel_6
+          # -gpu swiftshader_indirect is the usual CI choice. It is correct here only
+          # because runners have no GPU to pass through; on a workstation the same
+          # setting routes through SwiftShader's JIT, which is a known crash source.
+          emulator-options: -no-window -gpu swiftshader_indirect -noaudio -no-boot-anim -camera-back none
+          disable-animations: true
+          script: ./gradlew :app:connectedDebugAndroidTest
+
+      - uses: actions/upload-artifact@v4
+        if: always()
+        with:
+          name: e2e-report-api${{ matrix.api-level }}
+          path: |
+            app/build/reports/androidTests/
+            app/build/outputs/androidTest-results/