What broke, in order:
- The tunnel depended on a VPN client that reserved its own routes. That client was replaced, and the replacement did not pass the virtual machine's traffic at all — the host worked, the guest was isolated.
- Cloudflare's edge is reachable on TCP 7844; from this network that port was blocked on the direct route. With the VPN gone, the tunnel had no path.
- The machine's DNS was rewritten by the VPN client and stopped resolving
argotunnel.com, so even the fallback attempt on the host failed at name resolution. - The watchdog was scheduled every five minutes and had a hard time limit. During the outage its repair branch ran longer than the limit allowed, so the watchdog itself started timing out instead of reporting.
What we changed:
- the connector now resolves the edge by address, not by SRV lookup, so broken DNS cannot stop it
- four connectors run in parallel, so one dropping does not unpublish the site
- the watchdog checks with plain sockets instead of spawning shells, finishes in about four seconds, and only speaks when the state changes
- the local publication stack restarts itself, and does so in the background rather than inside the watchdog's time budget
What is still true and uncomfortable: a site that depends on a residential connection and a consumer VPN has a single point of failure that no amount of scripting removes. The honest fix is a rented host, and that is the next step.