I personally see your first two arguments the exact opposite way.
The visual display functions a bit like subtitles on TV: I don't really notice it's there, until it can help me. Just like subtitles help when I don't immediately understand something, autocomplete helps when typing something takes a bit longer than I'd like it to or when I'm not entirely sure what I want to write.
The "misuse" of autocomplete always continues to affect my workflow, because I can never ignore it. When it's wrong I have to explicitly opt-out from having it do the wrong thing, instead of quickly opting in to the right thing when it makes sense. It's a lot like auto-correct on phones, correcting words that weren't wrong in the first place. The fact that the popups capture arrow keys makes this effect significantly worse.
After some configuration I now have Emacs set up to handle this perfectly. Enter and movement commands always work like they normally do, never depending on the state of the autocomplete popup. Whenever it's there I can quickly make use of the autocomplete popup using Alt. It never does the wrong thing, because I would have to consciously tell it to do the wrong thing. When it's opt-in, autocomplete for regular English is actually a delight. I notice myself quickly auto-completing long words that take longer to type out.
I fully agree with your third argument. In my experience libraries and language ecosystems that fully expect autocomplete, like Java, tend to become overly verbose. The length of a name or expression affects the reading cost, but autocomplete mostly removes the equivalent writing cost. This then allows programmers to quickly give up on finding better names, or clearer ways of doing things.