[sttts]

Summary

The v1beta2 DGDR admission path in ai-dynamo#16661 must default a budget that can actually be scheduled on the node shape selected by shared GPU discovery.

This change consumes gpu.GPUInfo.SelectedShapeGPUs from the source fix in Pranjulcr7/dynamo#1, rather than using total same-SKU inventory. It updates every admission/defaulting fixture and covers the heterogeneous 8+4 case at both the resolver and complete API-server admission boundaries.

For example, discovery of matching H200 nodes with 8 and 4 GPUs records inventory 12 but admission persists the executable selected-shape capacity:

apiVersion: nvidia.com/v1beta2
kind: DynamoGraphDeploymentRequest
spec:
  hardware:
    gpu:
      sku: h200_sxm # ADMISSION-DEFAULT from discovery when omitted
      budget: 8     # ADMISSION-DEFAULT from SelectedShapeGPUs when omitted

The existing API semantics remain unchanged:

• spec.hardware, gpu, sku, and budget remain optional;
• the automatic budget is capped at 32 after selected-shape aggregation;
• an explicit spec.hardware.gpu.budget is authoritative and is never capped;
• SKU filtering still happens in shared discovery;
• total same-SKU inventory remains diagnostic and is not exposed as the default budget.

Part of ai-dynamo#17192.
Depends on ai-dynamo#17262 plus Pranjulcr7/dynamo#1.
Stacked on ai-dynamo#16661.

Validation

Validated locally with the source fix from Pranjulcr7#1 applied as a temporary, uncommitted overlay:

• go test ./internal/webhook/defaulting -count=1
• complete TestDynamoGraphDeploymentRequestValidator_Validate admission table
• admission Envtest with the repository's Kubernetes assets (7.936s)
• repository pre-commit hooks
• git diff --check

This branch intentionally does not duplicate the shared discovery source change; it compiles directly once its declared dependency is present in the base stack.