The staging uplink is in no VLAN zone and was fully default-deny; admit
TCP 22 on it so the box is reachable from the old LAN before the mgmt
VLAN or the mesh are up. Everything else on the port stays closed.
The pre-cutover DHCP-client uplink into the old LAN was a hand-written
systemd.network block in the gw-cnx-1 machine config. Make it a router
setting next to trunkPorts/accessPorts so every replacement gateway can
stage the same way, with an assertion that the port is not also the WAN,
a trunk or an access port. gw-cnx-1 sets stagingPort = "enp3s0" in the
inventory; the machine-local block is gone.
Move the `router` service instance out of clan.nix into its own clan
module, inventory.nix, imported by clan.nix. Fleet-wide role settings
(mesh subnet, proxy domain/ACME) and the gw-cnx-1 site layout live there;
clan.nix keeps the machine list and the clan-core services. The gw-cnx-1
toplevel is byte-identical before and after the move.
wifi.* settings turn the gateway's own radios into the site AP (hostapd):
SSIDs are defined once in wifi.networks and act as untagged access ports of
their VLAN (each BSS joins br0 with the VLAN's PVID), radios pick what they
broadcast, passphrases are vars prompts. Extra SSIDs on a radio get BSSIDs
derived from its hardware address. A udev rule puts configured radios in AP
mode on appearance, since networkd cannot bridge a station-mode interface
and gives up before hostapd switches it. wifi.enable gates all of it.
The VM test grows a hwsim radio with two SSIDs and a WPA3 station in its
own netns that must get a Kea lease on the SSID's VLAN. Making the client
deterministic (route metrics, loose rp-filter, a guard against a vacuous
negative check) exposed that allowWan was never enforced: networking.nat
opens forward-to-WAN for all of its internalInterfaces, which listed every
VLAN. It now lists only the allowWan VLANs; the duplicate custom rule is
gone. No behavioural change on gw-cnx-1 (all its VLANs allow WAN).
kurogeek