What was happening here is that because we were detecting window clicks
inside _input, when you clicked overlapping windows, all of them would
receive the click, and all of them would set themselves to active and
move to the foreground. The solution to this was to move the
click-detection into _gui_input instead, which stops the event from
traveling to the back elements. There is, however, a gotcha. Because
_gui_input only fires on the topmost element clicked, the window itself
doesn't receive the event because its children are covering it up and
receive the event instead. To work around this I iterated all the
children of the window, connecting to and forwarding their _gui_input
events to the window root. Thus, buttons still work when you click on
them, but the window also gets to respond as it needs to.
Before this, maximizing a window functioned by reparenting it to a
margincontainer. Now, maximizing a window is a matter of setting its
anchors and offsets. It will require more code but the reparenting
strategy was causing problems because maximized windows were always
displayed below windowed ones. This could technically be solved by
modifying z-indexes of windows individually but this seems like a less
annoying solution than that in the long run.