Maybe this question is so obvious it doesn't bear asking, but I haven't been able to answer it from the internet.
In the case of Galera multi-master replication, is this statement still guaranteed to be atomic?
UPDATE table
SET col = new_value
WHERE
primary_key = ? and col = old_value;
in that if two clients issued the same request for a given primary_key and old_value, only one would successfully update a row, is that correct?
Yes. Atomicity is correct even if both clients are updating on different cluster members.
Upon commit, (even if explicit), like locally, one of the client's will get a DEADLOCK error.
Reference: certification based replication