OCP4: UDP NodePort traffic from a single sender is silently dropped on an OpenShift node while other senders work

Issue

• One specific sender's UDP traffic to a NodePort never reaches the backend pod. Other senders to the same NodePort work. The same sender works on other NodePorts.
• The problem often starts after the backend pod on that node restarts, and it does not recover on its own.
• A tcpdump -i any capture on the node shows the sender's packets arriving on the physical interface and bond, but the chain stops there: the packets never appear on br-ex or ovn-k8s-mp0 and never reach the pod.
• The OVN-KUBE-ETP DNAT rule counters for the NodePort do not increase with the sender's traffic.
• The node's UdpNoPorts counter (nstat -z UdpNoPorts) increases at the sender's packet rate.