Well, that's better news than I was suspecting. Once I got to the notes about not favoring automated features, I was concerned that the whole idea would be a wash.
On the plus side, if I'm never going to release it, I'm basically feature complete. All I need to do is add some additional error checking. (It should have some error checking now, but it could use a little more.)
If you're curious, the thing is currently a 140 line BASH script using curl to make the Twitter calls. It took me ~2-3 hours to get all the features I wanted and to verify they all worked. (This would be how I got a fully functional proof-of-concept before I thought to ask about whether such a program would be allowed.) I may try out a show using the current script before I do much more to it. It would be cool to integrate location tracking and timed uploads of photos, but both of those can wait until I rewrite the thing in Python -- and I won't be doing that until I can prove the core idea is moderately successful.
You are correct that there will be no automated account creation. I clearly read the section about scraping the site will cause your account to be terminated. Automating web browser actions in such a manner is functionally the same, and sites can be programmatically scraped in a manner that they look like they're having automated browser actions, so each is really just as bad. The plan has always been for each character to get the account created manually. I am glad that I don't have to worry about limits there. It should be easy enough to stay within reason.
re:other limits. The FAQS or API docs or other site information mention specific limit numbers. That's where I found out such things existed and were enforced. It is easier to just plan for reason and check for error, though, so I'm happy with the answers. To pull this off, the timing on everything needs to look like normal human tweets, so there shouldn't be any problems over-stretching the bounds here.
re: Replying to @people. This wouldn't be at all random. This would be very specific feedback *initiated* by the human. This is really to help keep up the illusion that it is all really happening. If someone asks a specific question about the story, the answer can be presented as if it is just one person replying to another via Twitter. I would expect most people to just watch silently and that would be it, but if someone asks a question directed to a character it only makes sense for that character to respond. In general, the focus will be on the other characters. Randomly including names of followers would -- quite rightly -- be seen as spam.
Since this isn't going to be a distributed project, I'll sketch out an idea, then create a website for the particular show before requesting a "source" identifier. This should make it even easier to identify the tweets as coming from an automated source.
I hope that also clarifies my personal stance on some of the issues you raised.