> Just a shot in the dark here.  The only way I can think of this
 > occurring (Vwindow_list being too small), is that something modified
 > Vwindow_list when we don't expect it to be modified.  I looked through
 > all the uses of Vwindow_list and window_list(), and they all look
 > kosher.  The only modification seems to be setting Vwindow_list to nil.
 > But window_list() does call into lisp (nconc2, which calls Fnconc) when
 > constructing Vwindow_list, and it uses Vwindow_list as a temporary in
 > the process.  Is it possible that during that call something happens
 > asynchronously (via gc?) that calls something that sets Vwindow_list
 > to nil again, while Vwindow_list is being constructed?

Improbable but I attached some code that could detect such an
incidence.  Put a breakpoint at the

	Vwindow_list = window_list_2;

line.

 > I don't think
 > so, off hand, but I'm grasping at straws.

I also added a variable called `window-list-lengths' which records the
last 20 or so last changes in the number of windows seen by window_list.
Maybe you can detect some strange 2 ~> 6 or 6 ~> 2 change while running
Gnus.  I'm still not sure whether your daemon setup could be involved
somehow.

martin