While looking for a reason why 29-07C.PS in PS3CET hinder consistent result, I found a problem in 
current setpagedevice.

The difference was caused by i7 machine, on page 2 and 4, where it generated undefined errors where 
they should be a configrationerror.

'undefined' error was happening because .selectmedia procedure in gs_setpd.ps was building an error 
information dictionary ('<failed>' in its comment) with many keys those doesn't exist in <request> 
dictionary, such as /LeadinEdge or /InsertSheet.  This draws the policy 0 procedure (line 544) to 'get' 
non-existant keys from <merged> dictionary.

I have done above observations oh trunk rev. 10609.

I have not confirmed why this didn't happen on i7a and macpro machines, but I guess on those 
machines, /PageSize happend to be the first key pulled out from <failed> dictionary so it didn't end up 
with 'undeinfed'

Reproducing procedure:

---begin test.ps
%!
% Remove all but 'letter' page size
<<
  /InputAttributes currentpagedevice /InputAttributes get
  1 1 2 index length 1 sub {
    1 index exch undef
  } for
>> setpagedevice
<< /Policies << /PolicyNotFound 0 >> /PageSize [ 500 700 ] >> setpagedevice
---end test.ps

$ bin/gs -q -IResource/Init -sDEVICE=pbmraw -o /dev/null -dBATCH test.ps