Add minimal reproduction of compositor bug

external/bugs/composite-transform/README.md

@@ -0,0 +1,46 @@
+# WebKit nested composite-transform frame pacing reproduction
+
+This is a dependency-free reproduction for an iOS WebKit frame-pacing issue involving two nested animation clocks:
+
+- A route-sized parent surface whose `translate3d()` progress is updated by `requestAnimationFrame`.
+- A child with an infinite CSS `rotate()` animation, drawn as a rounded border spinner.
+
+On the originating device, spinner durations within part of the `0.5s`–`1.1s` range produced regular 33–34 ms rAF gaps, while durations outside that band were smooth. The workload per frame is effectively unchanged, making the duration-dependent behavior unexpected.
+
+## Run
+
+Open `index.html` directly. The reproduction is a single, self-contained HTML file and works from a `file://` URL without a server.
+
+To test it through a local HTTP server instead, run this from the repository root:
+
+```sh
+npx vite external/bugs/composite-transform
+```
+
+This directory is independent of the Vue demo and has no build step or runtime dependencies. The CSS, test scene, controls, and measurement code are all in `index.html`, so that file can be attached directly to a browser bug report.
+
+## Protocol
+
+1. Do not scroll or touch the page while a sample is running.
+2. Run **Sweep 0.5–1.1 s** with **rAF-driven** selected and **Spinner enabled** checked.
+3. Repeat with **Static translate3d parent**.
+4. Repeat with **No parent transform**.
+5. Clear **Spinner enabled** and run the sweep as a control.
+
+Each duration runs for 1.8 seconds. In the rAF-parent case, the incoming route moves for 700 ms and then remains stationary for the rest of the measurement.
+
+## Expected
+
+Changing only `animation-duration` should alter angular velocity, not cause a stable 60 Hz page to begin missing display deadlines. The rAF stream should remain close to 16.7 ms throughout the sweep on a 60 Hz device.
+
+## Actual observation in the source application
+
+The problematic sample showed:
+
+- A 16 ms baseline.
+- Repeated 33–34 ms gaps during the route transition.
+- The same gaps continuing while the parent was stationary and the spinner remained visible.
+- Normal pacing after the spinner was removed from the DOM.
+- No route-loading delay, JavaScript long task, or layout shift correlated with the gaps.
+
+The reproduction intentionally contains no Vue, Vue Router, Suspense, application store, network request, or service-worker caching logic in its measured path.