Tested with WiredTiger which seems to exhibit this more than MMAP.
Covered index query which has to return all the documents from a collection is slower than just a full collection scan even in cases where index is quite small and collection is very large.
Looking at where we spend time, it appears for each "document" we need to "synthesize" from the index we allocate and free some memory. We apparently have some optimization for avoiding that in collection scan but not full index scan. This is to track the same optimization for index scan.