I like to build hard core simulations. the type of thing where the delays of network lag are unacceptable, and "prediction" is the most evil thing in the world ever. So on the rare occasions when i do networked games, they're lock stepped to maintain perfect sync. I've gone so far as to write my own protocalls so robust you could unplug the phone line, plug it back in, and not lose sync. but in the end, you can only "ACK" so may "ACKs", and then you have to take it on faith the last one got through.
It's funny that you mention things like simulation where delays, prediction (by which i assume inaccuracies) are unacceptable. Yet you propose and defend a solution in which you present the user with views (renders) where the objects shown reside(? correct word ?) at a different point in time.
These arguments you use against the f-y-t actually prove your own theory as unacceptable for your purpose, and probably most other. (notice the italic part, it may have uses as Hodgman suggests).