Document existing nemotron-stt deployment (custom FastAPI, PoC-tier)

[wangchen615]

Context

While inspecting the cluster for PR #15 (NeMo training pod), I found a nemotron-stt Deployment already exists on NERC (10 days old, replicas: 0, last deployed 2026-04-28). It is not in git — someone built it ad-hoc.

Relevant spec details:

• Image: image-registry.openshift-image-registry.svc:5000/multi-modal-semantic-routing-for-vllm-d266df/nemotron-stt@sha256:3358d8a... — custom internal OpenShift image, not nvcr.io.
• Model: nvidia/nemotron-speech-streaming-en-0.6b loaded from /models/nemotron.nemo on the nemotron-models PVC (8Gi, read-only mount).
• Serving: HTTP on port 8093 with /health probe — likely a FastAPI / Uvicorn wrapper around NeMo's Python inference API.
• Resources: 2-4 CPU, 16-32Gi RAM, 1 H100.
• No Service or Route spec in-cluster (needs verification).

Why keep it (vs. just delete)

This is a valid PoC / experimental STT endpoint. Quick to iterate, no Riva conversion step, no Riva container overhead. Complements the upcoming riva-stt deployment (issue #9) which is the production / live-agent path with native streaming and high concurrency.

Tasks

• Export current live spec: oc get deployment nemotron-stt -o yaml | yq ... and commit as k8s/nemotron-stt-deployment.yaml.
• Check whether a Service exists: oc get svc nemotron-stt; if yes, commit too.
• Check whether a Route exists: oc get route nemotron-stt; if yes, commit too.
• Document the source Dockerfile / build instructions: how was the custom image built? If unknown, note that in a comment.
• Add to CLAUDE.md Model Status table as PoC-tier, distinct from the Riva production-tier entry.

Acceptance

• k8s/nemotron-stt-deployment.yaml (+ Service/Route if present) committed; matches oc get ... -o yaml (minus noisy status/uid fields).
• CLAUDE.md explicitly positions it as PoC / experimental, concurrency ~5-15 streams, appropriate for smoke tests and small-scale experiments.

Part of #5. Surfaced during PR #15 review.