I realized I gave a poor example of the semantic issue with removing upper bound for partition outputs, but the crux is that in that modeling the stats on partition outputs would be treated in a special way where upper bound being null means it's partitioned rather than "unknown", which is inconsistent with the other stats.
>
