This is what I don't get: if you have an accessibility layer that is fully usable, that means you have the capability to describe your user interface using only speech/text.

So what are you doing that couldn't be performant in the DOM? Because any kind of fast updates that you're doing that's too fast to render in an XML tree is also going to be too happening too fast to describe with a screen reader?

The DOM is asking you to be able to describe your interface -- performantly -- in entirely pure text. If you can't do that, then there's no way you can be fully accessible, because the screen reader needs to be able to describe your interface -- performantly -- in entirely pure text.