first i misread the memory column. the number on the process panel is absolute bytes, not percent, so the service was using over three times the memory it was designed for. then the whole system started stuttering, the service got killed, and the stutter stopped, which pointed the finger at memory. but the evidence said nothing ran out of memory.
the service group sat pinned against two hard ceilings, a fifteen gigabyte memory cap and a five gigabyte swap cap, and the kernel answered every push past the line with synchronous reclaim. fifteen thousand eight hundred sixty high events and zero oom kills. over forty gigabytes of pages were reclaimed directly by the allocating process itself, which means the process doing the allocation had to stop and clean up after itself. every one of those is a hitch in that process, and the audio capture and the transcription server live in the same group. that is the choppy capture.
memory.high does not kill. it quietly forces the group to reclaim its own pages, and the kernel swapped almost ten gigabytes out while twenty gigabytes of ram sat free. the cap was doing its job so well it made the system unusable.
killing the service dropped the group below the ceiling, reclaim stopped, and the stutter stopped. the fix is not more ram. it is a bigger cap or a better weight, so the group stops paying for its ceiling in wall clock time.