_record_survival needs to iterate through all job records, but job records from the generations before the checkpoint would have been removed if removes_subdirs=True, by _record_survival itself:

This is an issue when the checkpoint is from a completed run, where _record_survival is executed and batch records are generated. A typical occurrence is on HPC with a request limit on resources.

Two possible remedies:

• Comment out this two lines to leave job records untouched
• Somehow detect from which generation the optimisation resumes, only iterate the newly generated job records, and reuse the previous batch records.