Like, I don't think we necessarily have to go all the way back to the 2G-era(??) "plain text and menus" model. I actually think it would be quite cool and useful to support things like LaTeX formulas and images, and you can do that even on the lower-power chips these days, in a way too simple to confuse screen readers (although whether they can keep up with audio synthesis is another question - I remember seeing @freya once rail against ESP32s as an accessibility pothole, and I suspected it was a synthesis performance problem but was too intimidated to ask!).
But like. "Dumb phone made in the last 5 years" is probably a good reference point for an intentionally pruned feature set for basic respectful web browsing. And then other more "behavioral" transformers can vary from simple to advanced to cover a wide range of use cases.