@@ -11,6 +11,17 @@ slow-timeout = { period = "60s" }
 filter = 'binary_id(cap-rendering) and test(/^decoder::model::/)'
 threads-required = 3
 
+# Wall-clock tests run first and alone. crash_safety hands frames to a
+# real-time VideoToolbox session on a 30 fps schedule, and realtime_loop
+# counts the frames playback presents against the clock. Sharing the runner
+# with the seek storms, chaos and export tests starves them. On macOS,
+# crash/equality lost 3 to 7 frames in 6 shared runs and passed in all 6
+# solo reruns; realtime30 presented 65% of its frames shared and passed solo.
+[[profile.default.overrides]]
+filter = '(binary_id(cap-verify::crash) and test(=crash_safety)) or (binary_id(cap-verify::playback) and test(=realtime_loop))'
+threads-required = "num-test-threads"
+priority = 100
+
 [profile.ci]
 # No retries: a flaky case is a bug. run.sh reruns failures once afterwards
 # only to label them flaky in the job summary; the job still fails.
@@ -31,6 +42,12 @@ store-failure-output = true
 filter = 'binary_id(cap-rendering) and test(/^decoder::model::/)'
 threads-required = 3
 
+# Wall-clock tests run first and alone, as in the default profile.
+[[profile.ci.overrides]]
+filter = '(binary_id(cap-verify::crash) and test(=crash_safety)) or (binary_id(cap-verify::playback) and test(=realtime_loop))'
+threads-required = "num-test-threads"
+priority = 100
+
 
 # The FULL-tier seek storms (3000 seeks per decoder) are the slowest tests:
 # on the macos-15 runner one storm fixture takes 240 to 570 s per decoder,
 
@@ -24,6 +24,7 @@ jobs:
       rust: ${{ steps.filter.outputs.rust }}
       gpui: ${{ steps.filter.outputs.rust == 'true' || steps.filter.outputs.gpui == 'true' || steps.desktop-resources.outputs.desktop-resources == 'true' }}
       tauri-plugins: ${{ steps.filter.outputs.tauri-plugins }}
+      gpui-patches: ${{ steps.filter.outputs.gpui-patches }}
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -42,7 +43,7 @@ jobs:
               - 'crates/**'
               - 'vendor/**'
               - 'scripts/diagnostic-build.rs'
-              - 'scripts/setup.js'
+              - 'scripts/setup*'
               - 'Cargo.toml'
               - 'Cargo.lock'
             gpui:
@@ -65,6 +66,9 @@ jobs:
               - 'scripts/sync-desktop-versions.mjs'
               - 'scripts/verify-gpui-release-inputs.mjs'
               - 'scripts/verify-windows-stack.mjs'
+            gpui-patches:
+              - 'apps/desktop-gpui/patches/**'
+              - 'scripts/prepare-gpui-dependency.sh'
             tauri-plugins:
               - 'Cargo.lock'
               - 'bun.lock'
@@ -123,6 +127,7 @@ jobs:
           __tests__/unit/linkified-text.test.ts
           __tests__/unit/loom-import-ui.test.ts
           __tests__/unit/media-player-store-refs.test.ts
+          __tests__/unit/playback-store.test.ts
           __tests__/unit/seo-demo-videos.test.ts
           __tests__/unit/signed-baa-card.test.ts
           __tests__/unit/summary-editor.test.ts
@@ -273,6 +278,9 @@ jobs:
           prefix-key: v1-rust-tests
           save-cache: ${{ github.ref == 'refs/heads/main' }}
 
+      - name: Test native dependency setup
+        run: node --test scripts/setup.test.mjs
+
       - name: Run setup
         run: node scripts/setup.js
 
@@ -370,6 +378,70 @@ jobs:
           fi
           cargo +1.95.0 test --manifest-path apps/desktop-gpui/Cargo.toml --locked --target ${{ matrix.settings.target }} -p cap-gpui-kernels
 
+  # Cap only builds gpui's crates as dependencies, so their own tests never run
+  # in the app jobs above. Run the ones covering Cap's patches against real
+  # devices: Metal on macOS, WARP on Windows, and Mesa's lavapipe (Vulkan) and
+  # llvmpipe (GL) on Linux. Windows renders with D3D11 and its runners expose
+  # no wgpu adapter, so wgpu's tests run on macOS and Linux only. Zed's lockfile predates the patches' dependency
+  # changes, so these builds cannot use --locked.
+  gpui-patch-tests:
+    name: GPUI patch tests (${{ matrix.settings.name }})
+    needs: changes
+    if: needs.changes.outputs.gpui-patches == 'true'
+    timeout-minutes: 90
+    strategy:
+      fail-fast: false
+      matrix:
+        settings:
+          - name: macOS
+            runner: blacksmith-6vcpu-macos-15
+          - name: Windows
+            runner: blacksmith-8vcpu-windows-2025
+          - name: Linux
+            runner: blacksmith-8vcpu-ubuntu-2404
+    runs-on: ${{ matrix.settings.runner }}
+    env:
+      CARGO_PROFILE_DEV_DEBUG: "0"
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+
+      - name: Install Linux desktop dependencies
+        if: ${{ runner.os == 'Linux' }}
+        uses: ./.github/actions/install-desktop-deps
+
+      - name: Install software GPU drivers
+        if: ${{ runner.os == 'Linux' }}
+        run: sudo apt-get install -y --no-install-recommends mesa-vulkan-drivers libvulkan1 libegl1 libegl-mesa0 libgl1-mesa-dri
+
+      - name: Allow long paths in Zed's git dependencies
+        if: ${{ runner.os == 'Windows' }}
+        run: git config --global core.longpaths true
+
+      - name: Rust setup
+        uses: dtolnay/rust-toolchain@1.95.0
+
+      - name: Prepare GPUI dependency
+        shell: bash
+        run: ./scripts/prepare-gpui-dependency.sh
+
+      - name: Test GPUI patches
+        shell: bash
+        working-directory: ../zed-cap
+        env:
+          GPUI_REQUIRE_TEST_DEVICE: "1"
+        run: |
+          cargo +1.95.0 test -p gpui --lib stream
+          if [[ "$RUNNER_OS" == "Linux" ]]; then
+            WGPU_BACKEND=vulkan cargo +1.95.0 test -p gpui_wgpu --lib wgpu_atlas -- --nocapture
+            WGPU_BACKEND=gl cargo +1.95.0 test -p gpui_wgpu --lib wgpu_atlas -- --nocapture
+          elif [[ "$RUNNER_OS" == "macOS" ]]; then
+            cargo +1.95.0 test -p gpui_wgpu --lib wgpu_atlas -- --nocapture
+            cargo +1.95.0 test -p gpui_macos --lib -- metal_atlas metal_renderer
+          else
+            cargo +1.95.0 test -p gpui_windows --lib directx_atlas
+          fi
+
   tauri-plugins:
     name: Verify Tauri plugin versions
     runs-on: blacksmith-8vcpu-ubuntu-2404
 
@@ -29,6 +29,8 @@ jobs:
       - uses: actions/checkout@v4
         with:
           persist-credentials: false
+      - name: Test performance report isolation
+        run: python3 -B scripts/verify/test-perf-ab.py
       - name: Scope
         id: scope
         env:
 
@@ -46,6 +46,7 @@ jobs:
             __tests__/unit/recording-storage-lifecycle.test.ts \
             __tests__/unit/desktop-recording-*.test.ts \
             __tests__/unit/desktop-segments-*.test.ts \
+            __tests__/unit/segments-audio-download.test.ts \
             __tests__/unit/finalize-desktop-recording.test.ts \
             __tests__/unit/media-server-progress.test.ts \
             __tests__/unit/media-processing-budget.test.ts \
@@ -56,4 +57,7 @@ jobs:
             __tests__/unit/video-edit-published-output.test.ts \
             __tests__/unit/upload-progress-playback.test.ts \
             __tests__/unit/recording-prepare.test.ts \
-            __tests__/unit/s3-bucket-connections.test.ts
+            __tests__/unit/s3-bucket-connections.test.ts \
+            __tests__/unit/media-processing-workflows.test.ts \
+            __tests__/unit/new-media-comment-thumbnail.test.ts \
+            __tests__/unit/comment-reply-parent.test.ts
@@ -146,7 +146,7 @@ jobs:
       - name: Timestamp pipeline unit + property tests
         shell: bash
         run: |
-          cargo test --locked -p cap-timestamp -p cap-enc-ffmpeg
+          cargo test --locked -p cap-timestamp -p cap-enc-ffmpeg -p ffmpeg-hw-device
           cargo test --locked -p cap-camera-effects --lib
           cargo test --locked -p cap-recording --lib -- --test-threads=1
           # --nocapture so a WARP-adapter notch skip prints instead of
 
@@ -251,6 +251,15 @@ jobs:
           done
           SwitchAudioSource -s "BlackHole 2ch" -t output
           SwitchAudioSource -s "BlackHole 2ch" -t input
+          # macOS 15 plays alert and interface sounds through the default
+          # output, BlackHole, so ScreenCaptureKit records them as system
+          # audio and Instant's mixed track and every export carry them over
+          # the microphone markers. Hosted runners and macOS 26 capture
+          # silence there; silence these too so every runner records the same.
+          osascript -e 'set volume alert volume 0' || true
+          defaults write -g com.apple.sound.uiaudio.enabled -int 0
+          defaults write -g com.apple.sound.beep.volume -float 0
+          defaults write -g com.apple.sound.beep.feedback -int 0
           echo "outputs:"; SwitchAudioSource -a -t output
           echo "default output: $(SwitchAudioSource -c -t output)"
 
 
@@ -38,6 +38,7 @@ outputs
 .zed
 .output
 .vinxi
+*.config.timestamp_*.js
 native-deps*
 .tinyb