Hmm, I can't reproduce this using Pd-extended 0.42.5-20091104 on Mac OS X 10.5.8/Intel. That bug does ring a bell, but I thought it was fixed.
.hc
It is definitely there on my end (svn checked out approx. a week ago). The fix I suggested definitely fixes it. I've tested it on the last l2ork rehearsal on 9 machines running concurrently and it was rock solid for over an hour and a half running a complex patch and obviously using the fix I sent out in my previous email. Speaking of that fix, it seems to me that the additional "check_geometry" call is indeed superfluous as the debugger output shows that last check_geomtery shifts the canvas a tiny bit (unless this is compiz's doing on Linux), introducing inconsistency to the window/canvas position (if I understand correctly what check_geometry does).
Best wishes,
Ico