As I said, highlighting new comments since your last visit helps a lot in practice, I've seen this in action. There's one thing that has to be done carefully: when the user submits a new comment, the last-visit timestamp should not be reset, so that the user sees new comments since he/she started typing. To avoid "human" race conditions, so to say.
Flat model is good as long as users are responsible enough about quoting, including not quoting a huge comment for making just a short remark. Don't know, maybe a button for replying to a paragraph or selected text might help a bit, also showing/hiding quoted text, having hidden mode the default view.