The problem was that the code in replication didn't distinguish between a
setval() failing because the stored sequence number was bigger than the
current (should have been ignored) and a failure from the storage engine.