Interesting question. I've been struggling with it myself, and wonder how people here feel about the issue.

I tend to try and make one of the branches very short, and use that as a guard (something that leaves the current iteration/loop/subroutine as soon as possible), while the long branch is then considered the main code. Most of the time, this agrees with programming logic. The rare case in which it does not, I always discover that a big part of the longer branch should be factored out anyway, because it has too little to do with the function of the subroutine it's in.

This is your third example, although I'd negate COND and swap the branches if the return 2 part was much shorter than the return 1 part.