Could be totally off, but given the nature of the PK, shouldn't something like org.hibernate.usertype.UserType#disassemble be provided for actually creating proper CacheKey?
Again, these kinda mappings are beyond what I know, but this looks like a very special compound key to me...