Scott asked me to open this bug. We have been seeing CursorNotFound exceptions very infrequently in production. I added some client side logging the result of which is the following:

resultSize=242 queryTime=0 toListTime=108 server=localhost:27021 cursorId=0
resultSize=375 queryTime=0 toListTime=27 server=localhost:27021 cursorId=0
resultSize=227 queryTime=0 toListTime=130 server=localhost:27021 cursorId=0
resultSize=242 queryTime=0 toListTime=179 server=localhost:27021 cursorId=0

The toListTime is the time in ms that it took to read the entire cursor into memory. We do this immediately after the query to get the closest thing to a snapshot view of the world when the query occurred. We were getting odd data races otherwise.