Summary: There was a diagnostic assertion that no hazard pointers were held on pages discarded during an in-memory split, and that's not correct, so the assertion was removed.
I've been running the repro case looking for the failure in
WT-5043 with the new debugging. I was running 10 parallel jobs on ocelot-aws. I started this several days ago, so the tree is running develop as of changeset 359fb2736f.
Over the weekend, after 1100 iterations, one of them hit a hazard check assertion: