PR NVIDIA#502 added a Phase 1 "hardware" subtype before the existing "smi"
subtype in GPU measurements. The snapshot validation action used
subtypes[0] to read gpu.model, which now hits "hardware" (no model
field) instead of "smi", causing GPU model: null on all H100 runners.

Fix: query by subtype name (select(.name == "smi")) instead of index.

Signed-off-by: Yuan Chen <[email protected]>

PR NVIDIA#502 added a Phase 1 "hardware" subtype before the existing "smi"
subtype in GPU measurements. The snapshot validation action used
subtypes[0] to read gpu.model, which now hits "hardware" (no model
field) instead of "smi", causing GPU model: null on all H100 runners.

Fix: query by subtype name (select(.name == "smi")) instead of index.

Signed-off-by: Yuan Chen <[email protected]>

Adds an nfd componentRefs entry to the 12 production GPU leaf overlays
(H100, GB200, RTX Pro 6000 on EKS / AKS / GKE / OKE / LKE) flipping
overrides.topologyUpdater.enable to true. Each cluster running these
recipes now publishes per-node NodeResourceTopology (NRT) CRDs
describing NUMA zones, GPU-to-NUMA affinity, and NIC-to-NUMA affinity
— the fact base downstream NUMA-aware schedulers and recipe
auto-resolution flows can consume.

Overlay change: each leaf gains a fresh componentRefs entry with
name=nfd, type=Helm, and overrides.topologyUpdater.enable=true.
mergeComponentRef (pkg/recipe/metadata.go:532) deep-merges the
overrides onto the base nfd entry from recipes/overlays/base.yaml
without replacing inherited source/version/valuesFile.

Chart configuration (recipes/components/nfd/values.yaml):

* topologyUpdater.createCRDs: true ensures the noderesourcetopologies
  CRD is installed by the chart on overlays that flip enable=true.
  Managed K8s control planes (EKS / AKS / GKE / OKE / LKE) do not
  preinstall it. The chart guards the CRD template on
  (enable && createCRDs), so this is a no-op when TU is off.
* topologyUpdater.kubeletStateDir: "" disables the broad
  /var/lib/kubelet hostPath mount. TU only needs the pod-resources
  gRPC socket, which is mounted via a dedicated hostPath; the empty
  state-dir stops the chart from exposing kubelet state files (TLS
  material, pod manifests, checkpoint data) into the TU container.

Scheduling (recipes/registry.yaml): nfd.nodeScheduling.{system,
accelerated}.tolerationPaths now include topologyUpdater.tolerations
so the bundler injects --accelerated-node-toleration values into TU
pods. Without this, on every targeted GPU cluster (which carry
nvidia.com/gpu=present:NoSchedule), the TU DaemonSet would have been
unschedulable. nodeSelector deliberately not added — TU runs on all
nodes (same rationale as worker), and topology data on system nodes
is needed for cross-NUMA scheduling decisions.

Kind-chain overlays (h100-kind-*, kind-*) are intentionally excluded:
KWOK and kind clusters lack a real kubelet pod-resources gRPC socket,
so TU would CrashLoopBackOff. The KubeletPodResources feature gate
has been Beta-default since Kubernetes 1.15 and reached GA in 1.28
(KEP-606); AICR's affected leaves require K8s >= 1.30, so the
prerequisite is satisfied in practice.

Coverage:

* New chainsaw step assert-bundle-topology-nfd in CUJ1-training
  asserts the rendered NFD bundle contains topologyUpdater.enable=
  true, createCRDs=true, the GPU-taint toleration, master.enable=
  true, gc.enable=true, and enableNodeFeatureApi=true.
* New TestNFDTopologyUpdater_OverlayCoverage in pkg/recipe/
  metadata_test.go covers all 12 GPU platform+intent overlays
  (expected ON) and both kind-chain leaves (expected OFF).
  Type-assertion failures on ON cases promote to t.Fatal so a
  malformed override produces a loud regression rather than a
  silent false-negative.
* docs/user/component-catalog.md gains the missing NFD row and a
  Topology Updater section documenting which recipes run TU and
  the kubelet pod-resources prerequisite.

Closes the local NFD adoption initiative (Tracks A and B previously
shipped via PRs NVIDIA#482, NVIDIA#494, NVIDIA#495, NVIDIA#502, NVIDIA#511, NVIDIA#518, NVIDIA#688). Track C
(recipe auto-resolution from NRT data) and scheduler-integration
work remain open for future PRs.

Signed-off-by: Carlos Eduardo Arango Gutierrez <[email protected]>