[macOS][Sandbox] dcode rebuild fails repeatedly during pip install — creates 20+ orphaned sandbox attempts

## Description

When rebuilding a Deep Agents (dcode) sandbox to verify user-authored skills persist across rebuild, the rebuild process fails during the pip install phase (pydantic download). The rebuild is retried automatically, creating 20+ orphaned sandbox instances (`dcode-skills-persist-*`) that are never cleaned up.

Platform scope: Reproduced on macOS (local) and all CI platforms (Ubuntu 22.04, Ubuntu 24.04, DGX Station, DGX Spark, Ubuntu 24.04 GPU, Ubuntu 26.04 GPU)

## Environment

```text
Device: MacBook Pro (Apple M4 Pro)
OS: macOS (Darwin 25.5.0, arm64)
Architecture: arm64
Node.js: v22.22.1
npm: 10.9.2
Docker: Docker Desktop 29.2.1
OpenShell CLI: openshell 0.0.85
NemoClaw: nemoclaw v0.0.95
OpenClaw: N/A (dcode agent)
```

## Steps to Reproduce

1. Install NemoClaw and onboard a Deep Agents sandbox:
```bash
nemoclaw onboard --non-interactive --agent langchain-deepagents-code --name dcode-skills-persist
```
2. Write a custom skill file to the sandbox workspace
3. Run: `nemoclaw dcode-skills-persist rebuild --yes`
4. Observe rebuild fails during pip install phase
5. Run: `nemoclaw list` — observe 20+ orphaned `dcode-skills-persist-*` sandboxes

## Expected Result

Rebuild should complete successfully. User workspace files (skills) should persist across rebuild.

## Actual Result

Rebuild failed during Docker build step 15 — pip install pydantic download hangs or fails. The rebuild retries automatically, creating 20+ orphaned sandbox instances that are never cleaned up.

```text
rebuild failed: Downloading pydantic-2.13.4-py3-none-any.whl (472 kB): expected 1 to be +0
```

## Logs

Not captured

---
[NVB#6520551](https://nvbugspro.nvidia.com/bug/6520551)

Start by reproducing the failure with `nemoclaw dcode-skills-persist rebuild --yes`, then use `nemoclaw list` to confirm whether retry attempts remain orphaned. Trace the rebuild retry and cleanup paths around the Docker pip-install failure, and verify that rebuild completes and user workspace skills persist without leaving extra sandboxes.

Written by the indexing model from the issue text.