[DGX Spark][Onboard] onboard --resume shows a provider/model selection prompt after a canonical unpinned intent, regression of GH#7231

> [!WARNING]
> **Regression** — worked in v0.0.96 (GH#7231 fix verified and closed), broken again in v0.0.118.

## Description

DGX Spark managed-vLLM onboard `--resume` shows an interactive provider/model selection prompt instead of continuing the saved Express intent, contradicting the resume-without-reselection contract this exact scenario had previously fixed.

Platform scope: Reproduced on DGX Spark only; other platforms not tested.

## Environment

```text
Device: DGX Spark
OS: Linux 6.17.0-1022-nvidia
Architecture: aarch64
Node.js: v22.23.2
npm: 10.9.8
Docker: Docker Engine 29.2.1
OpenShell CLI: openshell 0.0.106
NemoClaw: v0.0.118
OpenClaw: N/A (onboard did not complete before resume)
```

## Steps to Reproduce

1. Start a non-interactive Express (managed vLLM) onboard on DGX Spark with `NEMOCLAW_PROVIDER=install-vllm`:
```bash
nemoclaw onboard --name {sandbox} --non-interactive --yes --yes-i-accept-third-party-software
```
2. Force/hit a failure at the provider-selection step so onboard aborts before sandbox creation.
3. Confirm the saved onboard session under `~/.nemoclaw/onboard-session.json` records a canonical unpinned Spark intent (only `version`, `kind=spark`, `sandboxName` — no provider/model pinned).
4. Run:
```bash
nemoclaw onboard --resume
```

## Expected Result

Resume continues the saved Express/managed-vLLM intent without any provider or model selection prompt.

## Actual Result

Resume output contains an interactive provider/model selection prompt instead of silently continuing the saved intent:

```text
matches /select (?:an? )?(?:provider|model)|choose (?:an? )?(?:provider|model)/i
```

## Logs

```text
Not captured — CI failure summary truncated the full resume stdout; only the matched prompt pattern was retained.
```

Start by reproducing `nemoclaw onboard --resume` after creating the canonical unpinned session at `~/.nemoclaw/onboard-session.json`, using `NEMOCLAW_PROVIDER=install-vllm` on DGX Spark. Inspect the resume path and its handling of the saved session; done means the saved Express/managed-vLLM intent resumes without a provider or model selection prompt.

Written by the indexing model from the issue text.