Thanks for the feedback on the idea, sounds like it might have legs so I'll move onto a physical prototype.
I've noted the concerns about drift. There are a couple of 9-DoF PCB module packages which include magnetometers and are pretty drift resistant (e.g. BNO085). Cheap enough to give it a go and see how it stands up.
For those asking "what's wrong with the primary attitude indicator": Nothing for what it was designed for (deducing roll and pitch when instrument scanning). The issue is that you have to be looking at it, and your brain doesn't automatically turn information presented to your fovea into orientation, it takes conscious processing to do that. Your brain can however turn signals from the peripheral vision into sense of orientation without conscious processing. So this idea is not intended to replace proper instrument scanning or to replace the AI, it's intended to project a sense of the horizon to the edges of your vision.
An example use case of this is where someone diverts their gaze such as looking down at Foreflight on their iPad, or fiddling with the radios. They're more likely to ignore false vestibular cues and retain a sense of their orientation if there's something that looks like a horizon out the edge of their vision.
In case anyone is concerned about the safety aspects of me building and testing this: I'm going to speak to one of the aero club's instrument rated instructors to see if he's game to try this out. I'd suggest he tests first in VFR then only if it feels ok test again in marginal VFR, then finally in IMC. I'll put in an easy to access "off" switch in case it proves to be a distraction or throws off incorrect cues.