Accessibility is a giant problem with almost every open source GUI library too.

It seems like an absolute slam dunk for AI. Create an integrated AI screen reader app that observes the screen and talks to you. It wouldn't even need to be a huge model. It could be a small local model that can just read and interpret UI elements and feed the results as tool use calls into an accessibility layer. It'd probably be a custom trained accessibility model, could probably be based on an existing small computer use model fine tuned for the purpose.

This seems like the "accessibility solution to end all accessibility solutions" that totally takes the burden out of the code and puts it in one place that can be uniformly improved.

This is the way. It just wasn't possible 5 years ago.