[DGX Spark][Onboard] nemoclaw onboard exits 0 when NEMOCLAW_VLLM_MODEL set to unrecognised slug

[wangericnv]

Description

When NEMOCLAW_VLLM_MODEL is set to an unrecognised slug and nemoclaw onboard is run in non-interactive vLLM mode, the command correctly fast-fails (prints an error listing valid slugs, no docker pull is triggered), but exits with code 0 instead of non-zero. Any script or CI job checking the exit code to detect failure will incorrectly treat this as success.

Related: NVBug 6282062 fixed the same exit-code-0 symptom for the nemoclaw connect + gated-model path. This bug covers the onboard + invalid-slug path, which is a separate code location.

Environment

Device:        DGX Spark (GB10)
OS:            Ubuntu 22.04
Architecture:  x86_64
Node.js:       v22.22.1
npm:           10.9.4
Docker:        28.3.3
OpenShell CLI: embedded in NemoClaw (no standalone openshell binary)
NemoClaw:      v0.0.63
OpenClaw:      N/A (issue occurs at onboard time, sandbox not created)

Steps to Reproduce

1. Confirm no sandbox exists: nemoclaw list
2. Run:

NEMOCLAW_EXPERIMENTAL=1 \
NEMOCLAW_PROVIDER=install-vllm \
NEMOCLAW_VLLM_MODEL=not-a-real-model \
NEMOCLAW_NON_INTERACTIVE=1 \
NEMOCLAW_ACCEPT_THIRD_PARTY_SOFTWARE=1 \
nemoclaw onboard --fresh --name vllm-bad-slug 2>&1 | tee /tmp/vllm-bad.log

3. Check exit code: echo "EXIT:$?"
4. Check log: grep -i 'slug\|valid\|recognised\|unrecognised' /tmp/vllm-bad.log
5. Confirm no docker pull started: docker ps --filter name=nemoclaw-vllm

Expected Result

• Step 3: EXIT:1 (or any non-zero value)
• Step 4: Log contains an actionable error listing valid slugs, e.g. Unrecognised vLLM model 'not-a-real-model'. Valid values: qwen3.6-35b-a3b-nvfp4, qwen3.6-27b, ...
• Step 5: No nemoclaw-vllm container running (fast-fail before docker pull — correct)

Actual Result

• Step 3: EXIT:0 ← incorrect; should be non-zero
• Step 4: Error message IS printed correctly (unrecognised slug listed with valid values)
• Step 5: No docker pull triggered — fast-fail behavior is correct

The error message is present but the process exits 0, making the failure undetectable by exit-code checks.

Logs

Not captured (exit code observed interactively during v0.0.63 PR #5038 test run; DevTest folder 1197336, task 698753 marked FAIL for this bug).

---

NVB#6303374

[kagura-chen]

I'd like to work on this. The fix mirrors what #4567 did for connect — add an early preflightVllmModelEnv() check in the onboard path so an unrecognised NEMOCLAW_VLLM_MODEL slug fails fast with exit 1 before any Docker/OpenShell preflight runs, rather than falling through resolveVllmInstallModel's catch-and-return-null path.