fix: gate traceback disclosure behind DEBUG_TRACEBACKS flag, run container as non-root

[man4ish]

What

1. Gates traceback disclosure in /rag/query's error response behind DEBUG_TRACEBACKS (default off).
2. Splits the traceback assertion test into two explicit states.
3. Runs the container as a non-root user, with ownership fixed on every path the app/nginx actually writes to.
4. Checked whether ghcr.io/omnibioai/omnibioai-base:latest has a pinnable version — it doesn't (see below).
5. Also fixed a build-blocking bug found along the way: Dockerfile still had COPY utils/ ./utils/, but utils/ was deleted in a prior cleanup PR (chore: remove zero-byte stub modules with no references (duplicated inline in rag/engine.py) #8) — this broke docker build outright. Removed the dead COPY line so the image builds again.

---

1. /rag/query traceback gating (api/routes/rag.py)

New _debug_tracebacks_enabled() (same on/off convention as api/auth.py's _auth_enabled(): os.getenv("DEBUG_TRACEBACKS", "").strip().lower() == "true"). The except block in /query now only adds "trace" to the response detail when that's true.

Before/after (captured via TestClient against a mocked engine raising Exception("Query Failed")):

DEBUG_TRACEBACKS unset (new default):

{
  "detail": {
    "error": "Query Failed"
  }
}

DEBUG_TRACEBACKS=true (preserves prior behavior exactly):

{
  "detail": {
    "error": "Query Failed",
    "trace": "Traceback (most recent call last):\n  File \".../api/routes/rag.py\", line 63, in query\n    result = engine.query(...)\n  ..."
  }
}

/rag/stream (api/routes/rag.py event_stream error handler) already only ever returned {"type": "error", "message": str(e)} — no traceback, no change needed. This PR makes that the consistent behavior everywhere: neither endpoint leaks a stack trace by default now.

2. Test split (tests/test_rag_routes.py)

test_query_endpoint_failure (which asserted "trace" in data["detail"] unconditionally) is now two tests:

• test_query_endpoint_failure_traceback_enabled — sets DEBUG_TRACEBACKS=true via monkeypatch.setenv, asserts "trace" in data["detail"].
• test_query_endpoint_failure_traceback_disabled — clears the env var via monkeypatch.delenv, asserts "trace" not in data["detail"].

Both states are explicitly tested — nothing was deleted or weakened.

3. Non-root container (Dockerfile)

Checked every path the app/nginx actually write to at runtime before touching permissions:

Path	Written by	Fix
/app/data/faiss_index/	VectorStore.save() via scripts/build_index.py on first boot	pre-created, chown -R appuser:appuser /app
/app/logs/, /app/cache/	reserved for future app use (currently unwritten by any code)	pre-created + owned, so a future write doesn't silently hit a permission error
/var/log/nginx/{access,error}.log	nginx, opened at startup (not lazily)	chown -R appuser:appuser /var/log/nginx
/var/lib/nginx/{body,proxy,fastcgi,scgi,uwsgi}	nginx worker temp dirs, created on first request	chown -R appuser:appuser /var/lib/nginx
/etc/nginx/conf.d/devhub.conf	CMD script, generated at container start (embeds runtime JWT_SECRET)	chown -R appuser:appuser /etc/nginx/conf.d

Also: stock nginx.conf sets user www-data; (assumes a root master process able to setuid) and pid /run/nginx.pid; (/run is root:root here) — both adjusted: the user directive is dropped (master already runs as appuser, so no setuid needed) and the pid file is redirected to /tmp (world-writable).

USER appuser (uid/gid 10001, no home, nologin shell) is the last setup instruction before CMD.

4. Base image pinning

Checked the GHCR registry directly (gh api /orgs/OmniBioAI/packages/container/omnibioai-base/versions): one version exists, tagged only "latest" — no semver/release tags, no dedicated base-image repo with GitHub releases. There's nothing to pin to that wouldn't be guessing. Left as ghcr.io/omnibioai/omnibioai-base:latest, unchanged — this remains a follow-up blocked on the base image publishing versioned tags. (Pinning to the current image digest was considered and rejected: that's a materially different decision — freezing to today's exact build rather than tracking intentional base-image updates — and the task asked for a pinnable version, not a digest.)

---

Verification

pytest: pytest -q → 175 passed (174-test baseline + 1 net-new test from the traceback split). Both new states individually confirmed:

tests/test_rag_routes.py::test_query_endpoint_failure_traceback_enabled PASSED
tests/test_rag_routes.py::test_query_endpoint_failure_traceback_disabled PASSED

Docker build + non-root run (fresh build, no cache reused beyond unrelated earlier layers):

$ docker build --platform linux/arm64 -t omnibioai-dev-hub:nonroot-test .
...
#28 [backend 17/17] RUN groupadd --gid 10001 appuser && useradd ... && chown -R appuser:appuser /app ...
#28 DONE 0.7s
...
naming to docker.io/library/omnibioai-dev-hub:nonroot-test done

Ran it (Ollama reachable via --add-host=ollama:host-gateway, pre-built FAISS index bind-mounted read-only so startup skips a full 19-repo rebuild):

$ docker run -d --add-host=ollama:host-gateway -p 18082:8082 -p 15173:5173 \
    -v "$(pwd)/data/faiss_index:/app/data/faiss_index:ro" omnibioai-dev-hub:nonroot-test

$ docker exec <container> id
uid=10001(appuser) gid=10001(appuser) groups=10001(appuser)

$ docker logs <container>
⏳ Waiting for Ollama...
✅ Ollama is ready
✅ Index already exists, skipping build
🌐 nginx started on port 5173
INFO:     Started server process [1]
INFO:     Waiting for application startup.
INFO:     Application startup complete.
INFO:     Uvicorn running on http://0.0.0.0:8082 (Press CTRL+C to quit)
INFO:     172.17.0.1:46372 - "POST /rag/query HTTP/1.1" 200 OK
INFO:     127.0.0.1:43804 - "POST /rag/query HTTP/1.0" 200 OK

Confirmed all three checks:

• Container starts without permission errors — clean startup log above, nginx.conf rewrite and devhub.conf generation both succeeded.
• /rag/query responds successfully from inside the running container — tested both directly against uvicorn (port 8082, HTTP 200 with a full grounded answer) and through nginx's reverse proxy (port 5173 → /rag/, also HTTP 200).
• Logs and cache writes succeed without permission errors:
  • docker exec <container> ps aux — uvicorn, nginx master, and all nginx workers run as appuser, not root.
  • /var/log/nginx/error.log — empty (no errors logged).
  • /var/lib/nginx/{body,fastcgi,proxy,scgi,uwsgi} — auto-created by nginx workers on first request, all owned appuser:appuser.
  • /etc/nginx/conf.d/devhub.conf — written successfully by the CMD script as appuser.
  • /app/logs, /app/cache — confirmed owned by appuser:appuser (no code path currently writes there during a query, so this is "ready and correctly owned" rather than "observed under load" — noted for transparency).

Scope

3 files: Dockerfile, api/routes/rag.py, tests/test_rag_routes.py.

⚠️ Do not auto-merge

Per instructions, this PR is not to be merged automatically — same as PR #9 and #10. Results reported above for review; merge only after explicit confirmation.