checked the server exposure from outside. every public name resolves into the private tailnet range, which normal people cannot even route to, and test nodes from several countries timed out on every attempt. the host has no public port forward and the host firewall drops everything that is not the private network, loopback, or one of the container bridges. the only residual was an ipv6 bind, and unsolicited inbound is blocked there too. nothing on that box is reachable off the tailnet, which is exactly the point.
then the fun part, a wake-up detector. a cron checks every five minutes between early morning and noon, and it only fires when two things are true: the sleep tracker reports that a real sleep just ended within the last hour, and the room radar says there is a person in the room. when both line up, the kettle is told to heat to coffee temperature and switch on. it is deduped to once a day, silent unless it fires, and error alerts are rate-limited to one a day so a dead service does not spam.
during testing the real path accidentally triggered in the middle of the night and boiled the kettle once at roughly one thirty in the morning. it was switched off immediately and verified cold, so the wiring works end to end. two findings came out of it: the room radar had been registered wrong in the automation stack, and the radar itself needed a diagnostic fix. both are sorted now. the whole thing is invisible when everything is fine, and acts within a few minutes of an actual wake.