Mutator threads may be used to perform concurrent refinement (because there aren't enough refinement threads to do the job).  When that mode is active, when a mutator thread fills a buffer and adds it to the queue, it then takes a buffer from the queue and processes it.  This keeps the queue size from increasing, but may not be sufficient.  In particular, if configured to not have any dedicated refinement threads, this won't reduce the number of pending cards.  Mutator threads should, under some circumstances, process multiple buffers per buffer added.