Windows IMEs used to emulate the style of edit control to display its candidate text, and the guess was frequently wrong. That's hardly a new problem for East Asian.

What worries me is that there's a general atmosphere that LLMs solved all NLP tasks, despite they don't do so well for kanji pronunciation <-> text mappings, likely for the lack of data - Japanese users especially don't post enough of organic data useful for that, apparently. Discord may be in a "good" spot to be able to (very invasively) collect that data by setting up a keylogger and comparing that to the final user-submitted text, but (aside from all the concerns) I'm not very hopeful about it.