There is a case where I initialize a mode in one window, show it in another,
move down several pages, and then switch to the buffer in the first window (and
the first page will be shown instead of the current).
If my understanding is correct, initialization is the only place in which the
behavior of "window t" is evident and beneficial.
I'm not sure what I dislike the least. "window t" and the need to track
window-buffer, a dedicated initialization mechanism, or switching to "fully
initialized mode" the first time a real window is processed.
I suppose I could try finding what goes wrong with the current patch in the
meanwhile. I think that my previous experiment suggests a confusion regarding
the value of (window-buffer) is a likely cause of this bug.