The primary point of my argument is that we don't have things in mind that current rendering engines can't handle, precisely because our current engines are so hell-bent on replicating reality. Why dream of what your tools can't make?
An example would be good, but I think hypertele-Xii's point isn't about what engines can handle, but about where engine developers are (and aren't) spending their money and time, and what they're making easy for users to do --- that if they make realism easy, then that's where users will go.
Not OP, but I think rendering scenarios where light doesn't follow straight lines would be problematic for current rendering engines. Also, some hypothetical portal or other dimension scenarios.