transcriber service still runs as root (missed by #489 non-root hardening)

Start with extern/transcriber.yaml and compare its transcriber service with the non-root vllm and reranker definitions described in the issue. Verify the transcriber uses the project-root Hugging Face cache and is covered by init-perms; done means it no longer writes under /root and matches the stack's non-root configuration.

Follow-up to #489 (Run OpenRAG as non-root (OpenShift compatible)).

The non-root hardening was applied to vllm (root compose), reranker (extern/reranker/infinity.yaml), etcd, minio, milvus, rdb, and openrag — but the transcriber service in extern/transcriber.yaml (included via ${TRANSCRIBER_COMPOSE}) was missed. It still runs as root and writes under /root.

Current state vs. its siblings

Impact

• When the transcriber is enabled (TRANSCRIBER_COMPOSE=extern/transcriber.yaml), it runs as root while the rest of the stack runs non-root — defeating the "whole stack non-root" goal of #489.
• It writes to /root/.cache/huggingface, so under OpenShift's restricted SCC (arbitrary UID, no write access to /root) the transcriber would fail to start.
• Its cache bind-mount is never pre-chowned (depends_on: init-perms missing), and it doesn't share the project-root HF cache that vllm/reranker now use.
• It's a vLLM-based image (Whisper), so it has the same getpass.getuser() crash the PR added USER/LOGNAME to vllm for.

Proposed fix (mirrors vllm)

Note the include-relative path: extern/transcriber.yaml lives in extern/, so the project-root cache is ../.cache/huggingface (vllm uses ./, reranker ../../).

services:
  transcriber:
    # Run unprivileged with a non-root user, matching the rest of the stack.
    user: "${PUID:-1000}:${PGID:-1000}"
    build:
      context: .
      dockerfile: ./vllm/Dockerfile.transcriber
    runtime: nvidia
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: all
              capabilities: [gpu]
    environment:
      - HUGGING_FACE_HUB_TOKEN
      - VLLM_MAX_AUDIO_CLIP_FILESIZE_MB=10000
      # Run unprivileged: point HOME and the HF cache at the writable mounted
      # volume so vllm never needs to write under root-owned /root.
      - HOME=/cache
      - HF_HOME=/cache/huggingface
      # vllm calls getpass.getuser(); set USER/LOGNAME so it resolves without
      # an /etc/passwd entry for the running UID.
      - USER=openrag
      - LOGNAME=openrag
    ipc: "host"
    volumes:
      # This file is `include`d, so a relative default resolves relative to
      # extern/. ../ points back at the project root so the cache matches
      # vllm's and the dir init-perms chowns.
      - ${VLLM_CACHE:-../.cache/huggingface}:/cache # HF + framework caches (HOME=/cache)
    depends_on:
      init-perms:
        condition: service_completed_successfully
    command: >
      --model ${TRANSCRIBER_MODEL:-openai/whisper-large-v3-turbo}
      --trust-remote-code
      --gpu_memory_utilization 0.2
    # ports:
    #   - ${TRANSCRIBER_PORT:-8002}:8000

cc @EnjoyBacon7