fix(network/shape): bandwidth-shaper unit runs before netplan-created egress devices exist and never retries

[brunodam]

Problem

solo-provisioner-bandwidth-shaper.service is ordered After=network-pre.target Before=network.target. Physical NICs exist by then (udev settles during sysinit.target), but devices created by netplan / systemd-networkd — a bond, bridge, or VLAN — do not: systemd-networkd.service is itself ordered After=network-pre.target, so it has not yet created them when the shaper unit runs.

On such a host the boot replay fails:

Cannot find device "bond0"

solo-provisioner-bandwidth-shaper.sh runs under set -e, and while the teardown line is guarded (tc qdisc del dev "$NIC" root 2>/dev/null || true) the build lines are not:

tc qdisc add dev "$NIC" root handle 1: htb default {{.Device.DefaultMinor}}
tc class  add dev "$NIC" parent 1:   classid 1:1  htb rate "{{.Device.Rate}}" ceil "{{.Device.Rate}}"

The unit is Type=oneshot with no Restart=, so nothing retries. The egress NIC stays completely unshaped for the life of the boot, and because Before= is ordering-only (not a requirement) solo-provisioner-daemon.service starts anyway and reports healthy — the daemon only reconciles the per-pod $VETH hierarchy and nft set membership, never $EGRESS. The failure is visible only in systemctl status solo-provisioner-bandwidth-shaper.service.

Related, smaller: the unit sets DefaultDependencies=no without Wants=network-pre.target, so the After=network-pre.target is inert unless something else pulls that target into the boot transaction.

Proposed fix

Render the unit with a dependency on the configured egress device rather than on a network target — After=sys-subsystem-net-devices-<nic>.device plus Wants=/BindsTo= as appropriate — or move it to After=network-online.target and accept the later start. Either way add a bounded retry (Restart=on-failure with RestartSec=, or a wait-for-device loop in the script) so a transient miss self-heals, and keep Before=solo-provisioner-daemon.service.

The device name is already known at render time (--egress-interface), so the unit template can name it directly.

Acceptance

• On a host whose egress interface is a netplan-created bond, bridge, or VLAN, the $EGRESS HTB hierarchy is present after a reboot (tc class show dev <nic> shows the trunk and leaf classes).
• A transient device-not-ready failure is retried rather than leaving the NIC unshaped for the whole boot.
• The rendered unit pulls its ordering target into the transaction (no inert After=).
• Unit-render test covering a virtual device name.