I figured it supports UTF-8. I believe the gist of my point is still valid.
Images are content. Layout is content. Typesetting is content. All are components of expression.
The underlying library that the tool uses, click (https://click.palletsprojects.com/en/stable/), turns out to include the idea that the "tweet" strings can contain ECMA-48 control sequences.
This permits not only boldface and italics, but also (on very old terminals and very modern terminal emulators) underline, strikethrough, faint, reverse video, invisible, and even 8 whole colours. (-:
Again, though, I expect that in the 9 years of the system's existence, no-one has actually used this in earnest.
In part, this is because the default mode seems to be to apply a regular expression substitution to attempt to strip out control sequences, because of course ECMA-48 and ECMA-35 are over-expressive permitting things like OSC, NEL, PM, APC, cursor motions ("layout is content"), insert/delete/erase, and code page changes.
Amusingly, the regular expression substitution is not based upon an understanding of ECMA-35 and is faulty.
Is this GPT generated? I find it hard to believe someone actually talks/writes like this.
This gives some irony to your analysis. Because this system thus permits one of the the very same ancient expression-in-images systems that you are alluding to: hieroglyphs are in Unicode.
I expect that in the 9 years of its existence, almost no-one has ever used them in this system.