Thanks for helping identifying the issue. I think it would be nice to clean up those obsoleted files during the run because running out of disk is reality not a good thing to have ;) Of course the bad estimates leading to the resize are the real issue but this at least makes it less bad.
Thanks for the tip about disabling parallel_hash but I also found it in the source. As mentioned before I disable (on production systems) all parallelism, not just for this issue but to prevent the OOM killer from killing Postgres - which happens way more often with parallelism on...
