Okay, I see the problem.  In the traditional model of credential 
caching, the upper layer checks the cache for a credential, doesn't 
find it, goes out and gets it, and puts it in the cache.  In a 
concurrent scenario a few processes might write duplicate entries to 
the cache, but that's (hopefully) rare and pretty harmless, so under 
this model we don't need the ccache layer to do any duplicate 
suppression.

During a TGS request the KDC might hand us a credential we didn't ask 
for, such as a referral or an alternate TGS cred.  Caching these 
credentials breaks the model--since we never wanted these credentials 
at the upper layer, the upper layer never checks for them.  That also 
calls into question the value of caching those unsolicited responses--
they won't be useful if the same higher-level operation is repeated 
(or even a similar one, such as a request resulting in a refferal to 
the same realm).  They would only save a TGS request if an unrelated 
operation happens to ask for that cross-realm TGT principal.

So I think my preferred solution for this scenario is to change 
get_cred.c not to cache answers it didn't ask for.  (This 
unfortunately isn't quite as simple as removing code due to the 
current design of step_get_tgt(), but it shouldn't be too hard.)  
Handling duplicates inside the ccache layer has a few problems:

* FILE ccaches are essentially append-only, so suppressing duplicates 
by removing the old entry can't be implemented efficiently in the most 
common ccache type.  (Heimdal implements removal by overwriting 
selected parts of the cred entry so that it becomes invisible to 
traversal, but that doesn't prevent the ccache from growing 
unacceptably large.)

* Suppressing duplicates by ignoring the new credential isn't the 
behavior we want for ccache config entries.  We would ideally like to 
be able to change the value of config entries that already exist by 
writing a new credential entry, although that doesn't currently work 
(krb5_cc_get_config() stops when it sees the old entry).

* Copying a moderately-sized ccache would become O(n^2) as we do a 
full read of the ccache for each entry.

In your patch, you noted a krb5_cc_remove_cred() call inside 
krb5_cc_store_cred().  That is an accidental leftover; see
http://krbdev.mit.edu/rt/Ticket/Display.html?id=7906