fix(ace-step): check in resting replicas 0 to stop ArgoCDSyncAlert refires

[igou-hermes]

Proposed by hermes-sre from ArgoCDSyncAlert (warning, firing since 2026-09-09T01:35Z on root-applications, openshift-gitops). No incident issue existed for this firing (prior ArgoCDSyncAlert issues #763/#799–#804 covered other apps and are closed).

Diagnosis

• root-applications sync has been stuck since 2026-09-09T00:22:52Z: waiting for healthy state of argoproj.io/Application/ace-step → app Progressing/OutOfSync → alert.
• ace-step pod ace-step-54585d6f8d-55m2g (ns ace-step): Pending 3d3h, FailedScheduling ×901 — 0/4 nodes are available: 4 node(s) didn't match Pod's node affinity/selector.
• The pod pins node-role.kubernetes.io/burst + nvidia.com/gpu: 2 (PR feat(ace-step): schedule on casval with nvidia.com/gpu: 2 #879). The only burst node, casval (2× 4060 Ti), is provided by MachineSet casval-worker at replicas: 0 at rest (lease model), and the cluster currently advertises only the 2 GiB Quadro P620 on p330.
• PR feat(ace-step): schedule on casval with nvidia.com/gpu: 2 #879 shipped the deployment checked in at replicas: 1, so the moment the casval lease ended the pod recreated and went Pending — holding the root app Progressing/OutOfSync and re-firing ArgoCDSyncAlert indefinitely.
• The repo's own at-rest convention (comfyui, llmkube qwen3-35b) and the ace-step README scale-to-0 workflow both expect replicas: 0 at rest; ArgoCD ignores /spec/replicas so live scale-ups are never reverted.

Change

applications/ace-step/kustomization.yaml: replicas: 1 → replicas: 0 with an explanatory comment. No other field touched.

Validation

• kustomize build --enable-helm applications/ace-step → Deployment ace-step renders replicas: 0
• yamllint -c .yamllint clean

Not merging — human review requested. When merged, the root app should return Synced/Healthy and ArgoCDSyncAlert should resolve on the next reconcile.