My Grafana dashboard said zero container restarts in the last 24 hours. Four containers had restarted in that window.

That panel could never have shown anything but zero. The query ran changes() over container_start_time_seconds, which assumes a restart updates an existing series. It doesn't. Docker Compose recreates the container with a new ID, cAdvisor starts a new series holding one constant value, and changes() over a constant is always 0. Every recreate was invisible. The fix counts distinct container IDs per name instead.

I only caught it by checking the panel against something it should have seen. A query for containers younger than a day showed four that the restart panel said never restarted.

The error panel next to it had the opposite problem. It matched about 770 lines in six hours. 719 were iperf3 logging a harmless client disconnect after every speed test, and 44 were Loki logging the panel's own query, which contains the word error. On a 30-second refresh the panel was generating its own errors. The new filter takes it from 770 to 9, and the real Kometa failures still come through.