I noticed, while testing cleanupOrphaned command, that both cleanupOrphaned and moveChunk delete documents from shards with w="majority". The code path is in RangeDeleter:

https://github.com/mongodb/mongo/blob/master/src/mongo/db/range_deleter_db_env.cpp#L127

The deleter waits up to an hour for a majority to catch up, then errors out. I think a kernel dev should verify my understanding of this, and the manual should mention this behavior when discussing sharding internals.