Robbert describes a scenario involving Windows Server KDCs.  Were you 
actually able to reproduce this using the steps followed, and did you 
use a Windows KDC?  I do not see the problem using a current MIT krb5 
KDC.

It does sound like there might be a relevant difference in behavior 
between the MIT krb5 1.10 and 1.12 client code, but absent a time 
differential between client and KDC (which Robbert specifically claims 
does not exist), I cannot imagine what it would be.