Examples:
In my state we have a govt funded web/native app that provides the fuel prices at all the service stations around you. If there was a mapping extension you could turn on, it would show this directly within the mapping app, and also an indicator "cheapest fuel within X km is at Y location" when you're in navigation mode.
In camera view with an extension turned on, provide object detection similar to how QR codes are recognised.
Apps which crash or otherwise stop rendering trigger a fade out to the white-room VR space, whereas frame-skips are handled by hard drops (so the effect is like a teleport to avoid motion sickness).
Like, Valve have really done an incredible job on the user application interface there.
This is already true with even the oldest and most rudimentary AR/VR systems, eg, Gear VR. The interface and head tracking is 90fps all the time. If the app is slow, the viewport will not freeze. Just it will be stuck on one frame.
This is something that I've given some practical thought to. The approach that I considered was to treat each application process as a client to a '3D space' service through which it could add and modify defined geometry.
Failure/slowness of any given application would leave the existing application geometry in-situ within the 3D environment and avoid the most jarring extremes of user experience. I imagined that the service interface itself would mandate the availability of simplified geometries together with meta information. This would allow the service to appropriately degrade the 3D rendered environment to maintain high framerates.
I think this approach could be effective and would not require a new OS.