[Ubuntu 24.04][Agent&Skills] DCode non-interactive agent request fails after sandbox is ready

## Description

A DCode non-interactive agent request fails after onboarding reports a ready sandbox and a healthy inference route. The agent runtime starts successfully, then the model request fails with a non-retryable remote agent error.

Platform scope: Reproduced on Ubuntu 24.04 GPU host only; other platforms not tested
Regression: Unknown — earlier versions not tested
OpenShell issue: No

## Environment

```text
Device: Ubuntu 24.04 GPU host
OS: Ubuntu 24.04, Linux 6.17.0-23-generic
Architecture: x86_64
Node.js: v22.23.2
npm: 10.9.8
Docker: Docker version 29.5.2, build 79eb04c
OpenShell CLI: openshell 0.0.116
NemoClaw: nemoclaw v0.0.124
OpenClaw: N/A — DCode LangChain Deep Agents Code runtime
```

## Steps to Reproduce

1. Install NemoClaw v0.0.124 on an Ubuntu 24.04 GPU host.
2. Onboard a DCode sandbox using the NVIDIA provider and a supported Nemotron model.
3. Confirm the sandbox is Ready and the inference route is configured for inference.local.
4. Run a non-interactive DCode task with dcode -n and ask it to modify files and run tests.

## Expected Result

The DCode agent request should complete normally through the configured inference route and return a successful result when the task is completed.

## Actual Result

The sandbox reaches Ready and the inference smoke check passes, but the non-interactive DCode request exits nonzero. The failure is reported as a non-retryable remote agent failure after the LangGraph server starts.

## Logs

```text
NemoClaw version: nemoclaw v0.0.124
OpenShell version: openshell 0.0.116
Node.js: v22.23.2
npm: 10.9.8
Docker: Docker version 29.5.2, build 79eb04c

Inference configured:
Workspace: default
Route: inference.local
Provider: nvidia-prod
Model: nvidia/nemotron-3-ultra-550b-a55b
Timeout: 60s default
Inference smoke passed: nvidia-prod / nvidia/nemotron-3-ultra-550b-a55b
Inference route set: nvidia-prod / nvidia/nemotron-3-ultra-550b-a55b

Running task non-interactively...
App: v0.1.55 / Agent: agent default / Model: nvidia/nemotron-3-ultra-550b-a55b
Starting LangGraph server...
Server ready
managed non-interactive error: error_class=RemoteError category=agent_remote_failure retryable=false
Model request failed: RemoteError category=agent_remote_failure retryable=false
AssertionError: command failed: dcode -n Work in sandbox work deepagent-journey. Read task.md and complete it exactly. Run the requested tests before finishing.
```

Start by reproducing the non-interactive dcode -n task on the stated Ubuntu 24.04 environment, confirming that the sandbox and inference.local smoke check pass before the LangGraph server request fails. Trace the managed non-interactive error and RemoteError agent_remote_failure, then verify that the task completes successfully and the requested tests run; no source file or failing test is identified in the report.

Written by the indexing model from the issue text.