[DGX Station][Inference] OpenClaw agent routes to previous model after switching provider from vllm-local to NVIDIA Endpoints

## Description

After switching the inference provider from vllm-local (DeepSeek V4 Flash) to NVIDIA Endpoints (Nemotron 3 Ultra) via `nemoclaw onboard`, `nemoclaw inference get` and `status` report the switch as successful, but the OpenClaw agent inside the sandbox continues routing requests to the previous model. The issue persists after rebuild.

Platform scope: Reproduced on DGX Station (Ubuntu 24.04) only; other platforms not tested.
Regression: Unknown — earlier versions not tested.

## Environment

```text
Device: DGX Station (2x GPU: RTX PRO 6000 Blackwell Max-Q + GB300)
OS: Ubuntu 24.04.4 LTS
Architecture: aarch64
Node.js: v22.23.1
npm: 10.9.8
Docker: 29.2.1
OpenShell CLI: 0.0.72
NemoClaw: v0.0.84
OpenClaw: 2026.6.10
```

## Steps to Reproduce

1. Install NemoClaw v0.0.84 with managed vLLM DeepSeek V4 Flash (vllm-local provider).
2. Confirm: `nemoclaw inference get` shows `vllm-local / deepseek-ai/DeepSeek-V4-Flash`.
3. Run `nemoclaw onboard`, select NVIDIA Endpoints, model `nvidia/nemotron-3-ultra-550b-a55b`. Onboard completes and reports: `Model: nvidia/nemotron-3-ultra-550b-a55b (NVIDIA Endpoints)`.
4. Verify: `nemoclaw inference get` shows `nvidia-prod / nvidia/nemotron-3-ultra-550b-a55b`; `nemoclaw status` shows `Inference: healthy` and `Inference (upstream): healthy`.
5. Run `nemoclaw {sandbox} rebuild --yes`. Rebuild completes successfully.
6. Run `nemoclaw {sandbox} connect`, then `openclaw tui` inside the sandbox.

## Expected Result

TUI routes to `nvidia/nemotron-3-ultra-550b-a55b`; inference requests succeed.

## Actual Result

TUI status bar shows `inference/deepseek-ai/DeepSeek-V4-Flash`, `tokens 0/200k`.

Sending a prompt returns:

```text
run error: LLM request failed: provider rejected the request schema or tool payload.
```

The old model route persists even after rebuild.

## Logs

Not captured.