A possibility off the top of my head is a load failure from external storage caused
by replication lag, followed by storing of the bogus diff in the diff cache.

Tim poked in some fallback code which should prevent this from happening.
Resolving FIXED.

I've committed three changes to address this:

1. external storage checks the master if a blob is not available on the slave
2. Revision::getText() does not deserialize or inflate if the external storage

engine returned false

3. DifferenceEngine::getDiffBody() checks for errors from Revision::getText()

and handles them appropriately.

That should fix the problem one way or another.