Currently, RAF makes all stages of WebGPU presentation pipeline to check for inputs and push things forward. Instead, we need every frame to go through the whole pipeline, and show up together with the DOM changes of other elements.

Talked about this with Andrew. The core piece for implementing this is blocking inside the TextureHost::Lock() function until the relevant frame is done. The only detail to figure out is getting an association of the WR frame with the WebGPU frame, to know what to wait for.

I think there isn't that much value in this issue for MVP (which now means enabling in nightly). A lot of content would run well without this. But we need this for shipping eventually in stable.

Don't we we use readback-to-cpu for presentation? In that case, we get this "dom atomicity" for free, unless we're doing something very unusual here.
Once we do have gpu-accel-present, we'll need this, but it's a solved-problem for e.g. webgl, so I don't think this needs to be tracked separately from accel-present.