[rambohe-ch]

Reason for Change:

The modeldeployment chart named the EPP Service <name>-inferencepool-epp and used the same value for the EPP pods' inferencepool label. Both are limited to 63 characters, so any deployment name over 45 characters failed to install. The AIManager smoke test with nvidia/NVIDIA-Nemotron-3-Nano-4B-BF16 hit this. The InferencePool name <name>-inferencepool had the same problem: Istio copies it into the istio.io/inferencepool-name label on the pool's shadow Service.

This PR names the InferencePool and the EPP Service after the deployment, and uses the deployment name as the inferencepool label value. The chart's name value is now capped at 63 characters.

 InferencePool                          
-  <name>-inferencepool
+  <name>
 EPP Service + EPP pod label `inferencepool`
-  <name>-inferencepool-epp
+  <name>
 EPP Deployment / ServiceAccount / Role / RoleBinding / ConfigMap
   <name>-inferencepool-epp              # unchanged
 values.schema.json: name
+  maxLength: 63

Name length	Before	After
≤ 45	works	works
46 – 63	EPP Service / label value > 63 characters, install fails	works
≥ 64	fails at the API server	rejected by the chart schema

E2E:

• Split EPPServiceName into EPPDeploymentName and EPPSelector, following the new naming.
• The Helm backend shortens release names over 53 characters with a hash suffix, so it can install deployment names of up to 63 characters.
• Added chart unit tests checking that every generated name and label value fits its Kubernetes limit for a 63-character name, and that a 64-character name is rejected. The assertion accepts both Helm 3 and Helm 4 error text.
• When a readiness wait times out, the error now shows the last real failure instead of a client rate-limiter error.
• The end-to-end 63-character e2e case is deferred: it needs fix(BREAKING): bound resource name lengths to fit Kubernetes limits kaito#2450, which is not yet in KAITO's nightly-latest image. Older KAITO creates a Workspace StatefulSet whose pod label exceeds 63 characters, so the inference pod is never created.

⚠️ Breaking changes

1. InferencePool renamed from <name>-inferencepool to <name>. Anything that refers to the pool by name must be updated: external HTTPRoutes, scripts, kubectl commands.
2. EPP Service renamed from <name>-inferencepool-epp to <name>.
3. EPP pod label inferencepool changes from <name>-inferencepool-epp to <name>. This label is in the EPP Deployment's spec.selector, which Kubernetes does not allow to change, so helm upgrade of an existing release fails with spec.selector: ... field is immutable. Before upgrading, delete the EPP Deployment and let Helm recreate it:

   kubectl delete deployment <name>-inferencepool-epp -n <namespace>
   helm upgrade <release> oci://ghcr.io/kaito-project/helm/modeldeployment -n <namespace> ...

   The model does not route traffic until the new EPP pod is Ready. Alternatively, run helm uninstall followed by helm install.
4. Metric label values change. The modelharness Azure PodMonitor copies the inferencepool pod label onto EPP metrics, and EPP labels its pool metrics with the InferencePool name. Update any dashboard, alert or recording rule that filters on inferencepool="<name>-inferencepool-epp" or on the pool name <name>-inferencepool.
5. name is limited to 63 characters. The chart schema now rejects longer names at helm install time. Such names never deployed successfully anyway.

Unchanged: the EPP Deployment, ServiceAccount, Role, RoleBinding and ConfigMap keep <name>-inferencepool-epp, so the status reporter and the benchmark workflow need no changes.

Requirements

• added unit tests and e2e tests (if applicable).

Issue Fixed:

Notes for Reviewers:

[zhuangqh]

LGTM. please add "BREAKING" to pr title.
backward compatibility: Delete existing deployment will result as object leaking.