At least on Linux, there are already a few different GUI toolkits that custom-draw their own widgets. Presumably those projects all implement some sort of standard accessibility interface -- maybe this one could do the same?

Don't blame the GUI, blame the screen reader. It's the screen readers that are stuck in the 70's. We have software that can do facial recognition and run on an raspberry pi but for some reason screen readers still need some ascii text buffer to send to the equivalent of a speak&spell.

The screen reader I'm using includes modern accurate OCR. I'm very interested in this, though. How do you suggest a screen reader should recognise the often subtle differences between control types and states across every application and framework when custom controls are so common?

To add to your point, why require such sophistication if it can be avoided? I mean, we could all send bitmaps back and forth to communicate. We could literally just take pictures of hand written notes. Encoded email seems much saner.