VM memory must be static, not ballooned

xcpng-vm sets memory_min = memory_max. Leave memory_min unset and the provider writes dynamic = [0, memory_max], which switches on XCP-ng Dynamic Memory Control: when the host is overcommitted, its scheduler squeezes every guest toward its dynamic minimum and the balloon driver returns the pages from inside the guest. MemTotal drops and Kubernetes just sees a smaller node.

That is backwards for a cluster node. Ballooning reacts to host pressure and is blind to in-guest demand, so a node that is OOM-killing pods gets nothing back. One node sat at 2 GiB of its nominal 8 GiB this way: kubelet missed heartbeats, the node flapped NotReady, containerd tore down sandboxes, and the visible symptom was Longhorn pods crash-looping — the storage layer was fine. Static memory reserves the full amount on the host and cannot be squeezed.