(frame-selected-window (posn-window position)) might not be the window
under the mouse cursor.You are correct. Though in testing 5 different mouse key bindings it was correct 4 out of 5 times because the depress part of the cross-frame drag event selected the window of the depress which is probably more accurate than existing results. I agree though that there may be uses of drag events where one does not want to select the window of the depress.
This speaks to my point in my most recent prior message, "If we just had a way to get a window from a set of coordinates within a frame, then I think this would help solve a lot of this." If the event-end of Emacs mouse drag events included a window, rather than a frame, when the endpoint of the drag is at a position unique to a window (considering Z-frame order), I think that would solve all these issues and simplify parts of the posn code.