createCollection on mongos does not check for the collection's existence in the cluster; it simply forwards the create command to (what it believes is) the database's primary shard.

If movePrimary has been called from another mongos, this mongos will cause the collection to be created on the wrong (old primary) shard, leaving it unreachable.

I believe once SERVER-30375 is done, this means we will be unable to migrate chunks for this collection onto this shard until the unreachable collection is manually deleted directly from the shard.