Ok, here is the bug I have been looking for. Kim, not Jan...

[David Kastrup <dak@gnu.org>]

With -fno-crossjumping, the real assert came to light.

It is xdisp.c in line 6122:

  /* The function move_it_vertically_backward may move over more
     than the specified y-distance.  If it->w is small, e.g. a
     mini-buffer window, we may end up in front of the window's
     display area.  Start displaying at the start of the line
     containing PT in this case.  */
  if (it.current_y <= 0)
    {
      init_iterator (&it, w, PT, PT_BYTE, NULL, DEFAULT_FACE_ID);
      move_it_vertically_backward (&it, 0);
      xassert (IT_CHARPOS (it) <= PT);
      it.current_y = 0;
    }

I suppose that the reseat_1 in move_it_vertically_backward will move
forward again.  Something like that.

-- 
David Kastrup, Kriemhildstr. 15, 44793 Bochum