I have a Python test that reproduces the failure seen in WT-9440. (That is, it fails on the original WT-9439 branch, and not on the new one.) Since otherwise this problem happens only under load and not often even then, this seems like a plus.

The failure occurred when opening the history store while reading a checkpoint cursor, if the metadata had been evicted and needed to be reread. There's a path in the row-store inmem code that does global visibility tests, which can interact with the checkpoint cursor. This can be reached by forcibly evicting the metadata; otherwise it doesn't happen much.

I'm putting the test in its own ticket because we don't want to hold up merging WT-9439 as it's in turn holding up WT-9252 :-|