We uncovered a bug on the timeseries insert path. When we have an index with key pattern {loc: "2dsphere", metaField: 1} (where metaField must be the timeseries meta field), the second time we insert a document into a given bucket, we treat it as an "update" of the bucket rather than a new "insert". Because we go the update path, we skip some logic that updates the underlying indexes.

In the geo-specific case, this does cause a bug where we may miss results in a query if the query should match an index key from the second document inserted to the bucket but not the first document.