3-9 replicas (HPA) previously relied on implicit scheduler spreading.
ScheduleAnyway (not DoNotSchedule) so pods still get scheduled if a node
is briefly unavailable, just less evenly.
Forgejo registry unreachable from worker nodes (network isolation +
host-to-ClusterIP routing gaps). Both management-service and queue-operator
now ship from the same public GHCR image, with queue-operator selected via
command override.