Harry-found silent data loss issue

Harry has discovered a silent data loss bug in trunk, but it goes all the way back to early 5.0.

Some rows are not visble after flush. Compared Harry repro running a memtable instead of a regular Harry model, and it still reproduces (in other words, it is almost certainly not a Harry issue).

Simple schema, and only one flush is required for repro, so not an unlikely bug. Max partition size is 1k rows, so not an unlikely setup here, either. No concurrency involved; reproduces stably. Good chance this is related to the fact the schema has DESC clusterings.

I am working on posting a Harry branch that stably reproduces it.