As part of WT-15743, a workaround was merged for the fact that opening a stable cursor on a layered table can return both ENOENT and WT_NOTFOUND. wt_open_cursor has code specifically mapping a WT_NOTFOUND to ENOENT, but it only does it in one spot. So there are clearly multiple error paths.

The other, possibly bigger, problem is that our testing doesn't seem to hit the WT_NOTFOUND scenario.

The scope for this ticket is to (1) figure out where the WT_NOTFOUND came from, and (2) add a test for the scenario that raised it.