chore: Upgrade OVN chart/images to 2026.1.3 and fix chassis-claiming regression

[skrobul]

• Upgrades the OVN chart to 2026.1.3+a659ab8a2 and points it at the renamed openstack-helm repo (was openstack-helm-infra, which no longer publishes this chart).

• Updates the OVN images (ovn_ovsdb_nb, ovn_ovsdb_sb, ovn_northd, ovn_controller) from docker.io/openstackhelm/ovn:ubuntu_jammy-20250111 to quay.io/airshipit/ovn:ubuntu_noble-20260711.

• While validating the upgrade, found that a baremetal/external port on a est network wasn't being claimed by any chassis. Root cause turned out to be a pre-existing (not upgrade-related) bug in neutron_understack.routers.link_vxlan_network_ha_chassis_group: it infers the gateway chassis by requiring every HA_Chassis row in the whole NB database to share one chassis_name. A single stale row left behind by a decommissioned/replaced chassis made that check ambiguous, silently blocking HA_Chassis_Group population for every vxlan network in the fleet, not just the one being tested.

• Fixed the check to ignore HA_Chassis rows whose chassis no longer exists in the Southbound DB, so a stale row elsewhere doesn't block inference.

• Added scripts/cleanup_dead_ovn_ha_chassis.py: a dry-run-by-default operator script that (1) removes stale HA_Chassis rows pointing at chassis no longer in the SB DB, and (2) repopulates any per-network HA_Chassis_Group left empty as a result, deriving the chassis from the router's own group — no more manually detaching/reattaching subnets per router to work around it.

Warning

Deploying this requires coordinated updated of the Network Node's OVS.

Deployment

Backup both databases

SB Database:

kubectl exec -i ovn-ovsdb-sb-0 -c ovsdb -- ovsdb-client backup unix:/var/run/ovn/ovnsb_db.sock > ovnsb_backup_$(date +%s).db

NB Database:

kubectl exec -i ovn-ovsdb-nb-0 -c ovsdb -- ovsdb-client backup unix:/var/run/ovn/ovnnb_db.sock > ovnnb_backup_$(date +%s).db

If you are not copying the commands verbatim, note that missing -t from kubectl exec is intentional, otherwise the backup does not work.

Steps

Specifically, the ovn-controller must use the compatible version of OVS libraries to talk to the openvswitch on the host.

On each network node, upgrade OVS from 2.17 to 3.3.4:

apt-mark unhold openvswitch-common openvswitch-switch python3-openvswitch
apt-get install openvswitch-common openvswitch-switch python3-openvswitch
apt-mark hold openvswitch-common openvswitch-switch python3-openvswitch

Then restart all of the OVN related pods.

Manual schema migration might be needed too (check the northd logs):

ovsdb-client convert tcp:ovn-ovsdb-nb-0.ovn-ovsdb-nb.openstack.svc.cluster.local:6641 /usr/share/ovn/ovn-nb.ovsschema
ovsdb-client convert tcp:ovn-ovsdb-sb-0.ovn-ovsdb-sb.openstack.svc.cluster.local:6642 /usr/share/ovn/ovn-sb.ovsschema

Following the restart, monitor the logs for any unusual errors (especially "OVNSB commit failed"). You may need to delete records in Chassis, Chassis_Private tables. Ideally start with ovn-sbctl --db=tcp:ovn-ovsdb-sb-0.ovn-ovsdb-sb.openstack.svc.cluster.local:6642 chassis-del <stale-chassis-name> if stale chassis is found.

If after that the routing still does not work, check the cleanup script included in this PR.