[Ubuntu 24.04][Onboard] Hermes onboarding reports API port 8642 forward not reachable

> [!WARNING]
> **Regression** — the same Ubuntu 24.04 GPU 8642-forward contract was previously closed as fixed in public issues #5206 and #9290; it is broken again in v0.0.114. Last directly verified good release is unknown.

## Description

Hermes onboarding on Ubuntu 24.04 GPU creates the sandbox but fails deployment verification because the OpenAI-compatible API forward on port 8642 never becomes reachable. The same Hermes managed-startup recovery validation passes on Jetson Thor, Ubuntu 22.04, Ubuntu 24.04 CPU, and Ubuntu 26 GPU, so this blocks only the Ubuntu 24.04 GPU execution path in the current run.

Platform scope: Platform-specific in this verification: fails on Ubuntu 24.04 GPU; verified working on Jetson Thor, Ubuntu 22.04, Ubuntu 24.04 CPU, and Ubuntu 26 GPU for the same release and test flow.
Regression: Yes — the same Ubuntu 24.04 GPU 8642-forward contract was previously closed as fixed in public issues #5206 and #9290; it is broken again in v0.0.114. Last directly verified good release is unknown.
OpenShell issue: Unknown. The failing surface is NemoClaw Hermes onboarding and the managed OpenShell forward startup.

## Environment

```text
Device: Ubuntu 24.04 GPU runner class
OS: Ubuntu 24.04
Architecture: x86_64
Node.js: v22.23.2
npm: 10.9.x
Docker: Docker Engine 29.5.2
OpenShell CLI: 0.0.106
NemoClaw: v0.0.114
OpenClaw: Hermes Agent v0.19.0
```

## Steps to Reproduce

1. Install NemoClaw v0.0.114 on an Ubuntu 24.04 x86_64 host with an NVIDIA GPU.
2. Onboard a Hermes sandbox with NVIDIA Endpoints and the default managed startup path.
3. During deployment verification, wait for NemoClaw to start the OpenAI-compatible API forward on port 8642.
4. Check whether the local API forward is reachable on 127.0.0.1:8642.

## Expected Result

Hermes onboarding should complete successfully. The OpenAI-compatible API forward on port 8642 should be started and reachable from the host while the Hermes sandbox is Ready.

## Actual Result

Onboarding reports that deployment verification found issues. The sandbox is created, but the API port forward is not reachable from the host, and the final onboarding result is unsuccessful.

## Logs

```text
Deployment verification found issues:

api: port forward not working (connection refused)
The OpenAI-compatible API on port 8642 is not reachable from the host.

Hermes is not ready

Port 8642 forward did not start:
ssh process started but local forward listener was not reachable
local forward listener did not open on 127.0.0.1:8642 within 10000ms
last probe failed with Connection refused (os error 111)
```

Start by reproducing Hermes onboarding on the Ubuntu 24.04 GPU environment described, focusing on deployment verification, managed startup, and the port-forward path for 127.0.0.1:8642. Compare the failing flow with the listed working platforms and inspect the resulting SSH and listener checks. Done means onboarding completes and the OpenAI-compatible API forward is reachable while the sandbox is Ready.

Written by the indexing model from the issue text.