It's actually a simple solution.

Have a flamewar detector similar to HN, and then for that thread and those users, have it look like the other people in the thread respond "Ok, I'm sorry." Obviously randomized using machine learning so that people will think that it's a real apology. Everyone on the thread will think the other person apologized and will simply move on. Or, if they respond, no one except the writer will see the responses.