The fast initial split algorithm operates by iterating the set of zones for a collection and writing chunks for each zone and for each "hole" between zones. However, the main assumption of this logic is that the set of zones comes in sorted order. Otherwise, the results are unpredictable.
The current caller of the algorhtm doesn't pass a sort based on the min key of the tags, which means tags could come in the order in which they were inserted, which is not necessarily sorted.
This is not a data-loss bug, but it renders the Geo sharding feature unusable in random cases.