[hermes] sandbox stop/start wipes data

### Investigation Summary

_No response_

### Description

From the documentation:
>Stop a sandbox’s container to free CPU, memory, and GPU resources without losing anything:
`nemohermes stop`

Unfortunately, if you do this, then the files in the sandbox are reset to a default state. Custom skills are gone, memory is gone, achievements db is gone, session DB is gone, etc.
At least. the config, env and policies are not modified.

### Reproduction Steps

1. `nemohermes stop`
2. `nemohermes start`

### Environment

- Hardware: GB10 platform
- OS: Ubuntu 24.04 LTS
- OpenShell 0.0.72
- NemoClaw 0.0.88
- Hermes Agent 2026.7.1

### Debug Output

N/A

### Logs

```shell
$ nemoclaw hermesbox stop
Hermes Agent gateway is managed by the sandbox; leaving it running while host forwards stop.
Stopping container 'openshell-hermesbox-1234'…
openshell-hermesbox-1234
Sandbox 'hermesbox' stopped. Workspace state is preserved.
Start it again with 'nemoclaw hermesbox start'.

$ nemoclaw hermesbox start
openshell-hermesbox-1234
Container 'openshell-hermesbox-1234' started.
Checking gateway health and host forwards…
✓ Active gateway set to 'nemoclaw'
Probe complete: recovered Hermes Agent gateway in 'hermesbox'.
```

### Checklist

- [x] I confirmed this bug is reproducible
- [x] I searched existing issues and this is not a duplicate

Reproduce the issue with the documented `nemoclaw hermesbox stop` and `nemoclaw hermesbox start` commands, then trace the sandbox stop/start entry points that manage the container and workspace. Confirm which files or volumes are reset and verify that custom skills, memory, achievement data, and session data survive a stop/start cycle while config, environment, and policies remain unchanged.

Written by the indexing model from the issue text.