As a result of investigation work, it looks like sampling CE generates variable plans.

Svilen's suggestion on the scope was to make sampling to read only from the beginning of the collection, rather than getting a true $sample (on the performance variants). We could hide this behaviour behind an internal test-only query knob in order to increase testing stability.

If this still does not eliminate the variability, we can consider disabling the most variable workloads, or disabling the sampling variants entirely.