Third site of the same defect. On the restore path DRILL_IP holds a MagicDNS name, so a laptop that cannot complete that name spends all 40 attempts and then dies blaming the ephemeral Tailscale key or MagicDNS propagation — neither of which is wrong, and neither of which is the problem when the node is up and answering on its tailnet address. The name is still tried first, because when it works it proves the resolution path the drill exists to prove. Only after that attempt fails does it ask tailscale for the address, which does not go through DNS. Succeeding that way warns and rewrites DRILL_IP, so drill_ssh() and every later step follow.