fix(cse): expand NODE_NAME in DRA driver systemd unit override

[Copilot]

The startDRADriverNvidiaGpu function wrote the systemd unit override with a single-quoted heredoc (<<'EOF'), suppressing variable expansion. As a result, ${NODE_NAME} was written literally into the drop-in file instead of the actual hostname, causing gpu-kubelet-plugin to receive an empty/undefined node name and breaking DRA driver registration.

Change

• parts/linux/cloud-init/artifacts/cse_config.sh: Changed <<'EOF' → <<EOF in startDRADriverNvidiaGpu so ${NODE_NAME} (set to $(hostname) at script init) is expanded at CSE execution time.

# Before — NODE_NAME written literally
tee "${DRA_DRIVER_OVERRIDE_DIR}/10-dra-driver-nvidia-gpu.conf" > /dev/null <<'EOF'
[Service]
ExecStart=
ExecStart=/usr/bin/gpu-kubelet-plugin ... --node-name=${NODE_NAME}
EOF

# After — NODE_NAME expanded to actual hostname
tee "${DRA_DRIVER_OVERRIDE_DIR}/10-dra-driver-nvidia-gpu.conf" > /dev/null <<EOF
[Service]
ExecStart=
ExecStart=/usr/bin/gpu-kubelet-plugin ... --node-name=${NODE_NAME}
EOF

[aks-node-assistant]

AgentBaker Linux PR gate — 236-failure mass run: shared cluster proxy-pod readiness exhaustion (test-infra, NOT this PR)

• Run: 167398747 (failed) — WIP/draft PR by copilot-swe-agent
• Failed task: Run AgentBaker E2E
• Test summary: DONE 405 tests, 97 skipped, 236 failures in 652.730s (~58% failure rate; 0 fwupd hits)

Dominant failure (same as 6 concurrent runs in this window):

prepare cluster tasks: dag execution failed:
waiting for proxy pod to be ready: listing proxy pods:
client rate limiter Wait returned an error: context deadline exceeded

Every failing scenario times out at the cluster.go:163: ✓ preparing cluster done (311.0s) boundary because the harness's e2e-proxy DaemonSet doesn't go ready inside the prepare-cluster DAG.

Cross-PR pattern (same window): identical 236-failure shape on builds 167387444 (PR #8652), 167387406 (PR #8294), 167387387 (PR #8600), 167393232 (PR #8669), and earlier on 167378787 (PR #8618). Six unrelated PRs in the same ~3h window, all hit the same shared-cluster-fleet readiness exhaustion.

Build-vs-test: test-infra (shared cluster pool), NOT product, NOT PR-caused.
Confidence: HIGH that PR #8676 is not the cause.

Recommended next action / owner: E2E infra / NodeSIG-dev — shared cluster fleet (e2e-proxy DaemonSet readiness + kubenet-v5 RG lifecycle) is overloaded. This is a draft/WIP PR; rerun once fleet stabilizes.

Posted by Clawpilot AgentBaker gate detective.