This has been seen under somewhat different circumstances by a couple of customers.
Initial sync of a large db (multi TB, couple billion documents). Foreground build of _id index starts at A, and cache usage rises steadily to 6x configured maximum before being terminated by OOM.
Mongorestore of 100GB db. Multiple parallel background index builds begin at A; at B cache usage begins to grow until it reaches about 1.5x configured max and is terminated by OOM.
No complete repro yet, although may have got a partial repro: during initial sync of a 500 GB db cache usage briefly rose to about 120% of configured max.