test(uat): document GPU Frontend pin vs demo and proving-run status

The serve graph matches the demo topology but not its cpu-worker
Frontend placement. Spell that out, and note that the first GCP
inference nightly is the proving run for #1644.

Signed-off-by: Rohit Rajani <rorajani@nvidia.com>

Author: rorajani

.github/workflows/uat-gcp.yaml

@@ -698,8 +698,10 @@ jobs:
       # Frontend cold-pulled the ~12GB vllm-runtime image on a CPU-pool node and
       # wedged in ContainerCreating past the readiness budget; phase_serve now
       # selects the GPU pool for both graph components, whose a3-megagpu-8g class
-      # pulls that image in a fraction of the budget. AWS/Azure serve steps stay
-      # commented until this lane demonstrates a served completion.
+      # pulls that image in a fraction of the budget. The first GCP inference
+      # nightly (or manual `run serve`) after this is the proving run — serving
+      # was never verified end-to-end. AWS/Azure serve steps stay commented until
+      # this lane demonstrates a served completion.
       - name: UAT - serve (DynamoGraphDeployment + endpoint)
         id: serve
         if: steps.conformance.outcome == 'success' && inputs.intent == 'inference'

tests/uat/lib/phases.sh

@@ -1398,10 +1398,13 @@ EOF
 
 phase_serve() {
   # Deploy the served inference graph (Dynamo) — the intent=inference CUJ, the
-  # DC3 counterpart of phase_train. Mirrors the DynamoGraphDeployment in
-  # demos/cuj2-inference.md (demos/workloads/inference/vllm-agg.yaml): the KAI
-  # queue and a two-component (Frontend + decode Worker) graph serving an
-  # OpenAI-compatible endpoint. The worker requests its GPU as a scalar
+  # DC3 counterpart of phase_train. Graph topology mirrors the
+  # DynamoGraphDeployment in demos/cuj2-inference.md
+  # (demos/workloads/inference/vllm-agg.yaml): the KAI queue and a
+  # two-component (Frontend + decode Worker) graph serving an OpenAI-compatible
+  # endpoint. Frontend placement intentionally diverges — the demo pins
+  # nodeGroup=cpu-worker; this graph selects the GPU pool for both components
+  # (pool-selection note below, #1644). The worker requests its GPU as a scalar
   # nvidia.com/gpu limit — the device-plugin production default (#1327).
   #
   # Tolerations are a portable SUPERSET of the taints across all UAT clusters