I always believed that IBM utilities worked as a set of phases chained logically and that they, either completed the tasks(or phases) or not at all.
With this in mind LOAD’s SHRLEVEL CHANGE support somehow changed this concept in that it resembles me more a program issuing SQL INSERTs than a batch utility that inserts rows massively.
Stupid as this might sound. Have you tried with a differente DISP option for the SYSDISC DD defining a new dataset on the restarted job?
Just my humble hint.