Technical follow-up: It seems to me that implementing today's platform accessibility APIs (e.g. UI Automation for Windows, AT-SPI for Unix) would be difficult if not impractical for an immediate-mode toolkit like this one.

These accessibility APIs expose a tree of UI objects, which the client (e.g. screen reader) can freely navigate and query. This basically assumes that there's a tree of UI objects in memory, which is the case for all mainstream toolkits as far as I know. But that's not the case for an immediate-mode toolkit. At least with Nuklear, the content and state of the UI (e.g. the label and current checked state of a checkbox) aren't stored anywhere in the toolkit's data structures. So I guess applications would have to play a very active role in implementing accessibility APIs, much more than they would with, say, Qt or even Win32.