/health is probed every minute from outside
the bot's own stack — so a crashed service produces a real
failure-recording, not a silently-correct response. This page is
hydrated from /api/uptime,
which rolls the probe log into windows. The minute-bar timeline
below shows the last 24 hours individually.