No one knows "all Kanji" much like no one knows "all English words" even in native English-speaking countries.

Do kanji handwriting-recognition models even “know” (i.e. are trained on) all in-use kanji, or do they just not bother with some of the rarest ones because their presence in the dataset would decrease the likelihood of correctly matching more-common characters, which it’s overwhelmingly-likely are what you’re writing?

Presumably it's the same as words in English, where you'd never make a word suggestion model that just blindly considers all possible words. You swipe jow or gow and it will change it to "how" despite the other two existing in arguably correct English in some very obscure case.

Indeed. Which means if your goal is fluency, it's silly to strive for 100.0%.