The searchable digital record is the key difference. Statements out loud to one or more people are often: A) tailored to that audience, B) carry much more context than a piece of text on a screen, and C) are a way of growing and exploring new ideas without the commitment of recording them to an easily searchable record for the rest of time. The long-term ramifications of saying something stupid in a conversation that isn't recorded are far less severe than they would be if every word you said was recorded and searchable forever.

A tweet is really more like submitting an article to a publication of record with your name and face attached to it. But, the ease with which twitter allows people to post makes it seem more like an ephemeral conversation. Anyone who has been burned by a stupid joke tweet from 10 years ago learned that lesson the hard way.

Easy to find "bad tweets", easy to immediately publicize the "bad tweets" to the entire planet almost instantly, easy to have it trusted that the bad thing was said as the twitter timestamp exists and a bunch of trusted people talking about a screenshot if you delete it confirms it's genuine, not to mention most people don't use twitter imaging themselves giving a rehearsed speech in a public square, merely as a way to share their amusing thoughts to the planet for some immediate validation, the dynamic is pretty different.

your first point (tweeting is as safe as talking) is entirely invalidated by your second (long searchable record helps find statements to criticize people for)

as for your third point, undocumented public speech (no audio or video recordings and no effective note taking) is vastly different from creating public documents (anything on the open internet).