Found while developing a prototype for SPM-4379 because a golden test to optimize $matches would sometimes work and sometimes break on different runs. The use of this unordered map means that consecutive calls to this function may result in a change in $or predicate order. The fix is a one-liner- use an ordered map.

The reason this could be an issue is because different orderings of the $or could result in different plan enumerations in the multiplanner, which would potentially change plans unpredictably between consecutive optimization calls.