kubernetes-nodes: raising minReplicas does not raise the MachineDeployment's replicas on upgrade #4786

Raising minReplicas on an existing node pool does not raise the MachineDeployment's replicas, although the comment above spec in nodegroup.yaml says CAPI clamps the live value into [min, max] on update.

The clamp in CAPI's calculateMachineDeploymentReplicas (v1.10.1, internal/webhooks/machinedeployment.go) runs only when the incoming object has spec.replicas unset. helm-controller applies the chart through Helm's three-way merge, so the object the webhook sees still carries the live value, and the function returns it unchanged at its first check.

I saw this in the computeplane e2e suite. Its pool is first rendered with minReplicas: 0, then upgraded to min and max 2. After the upgrade the MachineDeployment had the min-size annotation at "2" and spec.replicas: 0. What then lifts such a pool is the cluster-autoscaler, through --enforce-node-group-min-size or a scale-up for pending pods. That path was deadlocked by a separate autoscaler defect, fixed in #4768.

With the autoscaler fixed the pool does reach its min, so this is mostly a wrong comment: the guarantee it describes holds for create, not for update. It also means a pool raised above its live size waits on the autoscaler's next loop rather than on the upgrade itself.