[WSL2][Inference] Hermes chat remains idle after submitting a prompt with GPT-OSS 20B

## Description

On NemoClaw v0.0.121, a Hermes sandbox on WSL2 completes onboarding with the NVIDIA Endpoints `openai/gpt-oss-20b` model and reports the expected inference route. The Hermes TUI starts successfully, but submitting a chat prompt produces no response and remains idle until the client-side wait times out.

Platform scope: Platform-specific: reproduced on both WSL2 x86_64 and WSL2 arm64; verified working with the same model and Hermes flow on Ubuntu, DGX, and Jetson platforms.

Regression: Unknown — earlier versions were not tested with this exact model and flow.

OpenShell issue: No.

## Environment

```text
Device: Windows PCs running WSL2
OS: Windows 11 / WSL2 Ubuntu 24.04
Architecture: x86_64 and arm64
Node.js: v22.23.1 (x86_64), v22.23.2 (arm64)
npm: 10.9.8
Docker: Docker Desktop
OpenShell CLI: openshell 0.0.106
NemoClaw: nemoclaw v0.0.121
Agent: Hermes
Model: openai/gpt-oss-20b through NVIDIA Endpoints
```

## Steps to Reproduce

1. On a Windows 11 machine with WSL2 Ubuntu 24.04 and Docker Desktop, install NemoClaw v0.0.121 with the Hermes agent selected.
2. Onboard a new Hermes sandbox using NVIDIA Endpoints and model `openai/gpt-oss-20b`.
3. Confirm the sandbox is Ready and the configured inference route reports `openai/gpt-oss-20b`.
4. Connect to the sandbox, start the Hermes TUI, and wait for the ready prompt.
5. Submit: `What is 12 * 11? Answer with just the number.`
6. Wait for a response.

## Expected Result

Hermes submits the request through the configured inference route and displays `132`.

## Actual Result

The TUI accepts the input but never displays a response. It remains at the ready screen until the 180-second wait expires. The same symptom reproduces on independent WSL2 x86_64 and arm64 machines. The environment preflight, Docker daemon, Windows interop, required credential, onboarding, sandbox readiness, and inference-route checks all pass before the chat attempt.

## Logs

```text
Inference route set: nvidia-prod / openai/gpt-oss-20b
Model: openai/gpt-oss-20b (NVIDIA Endpoints)
Welcome to Hermes Agent! Type your message or /help for commands.
Prompt submitted: What is 12 * 11? Answer with just the number.
Timed out after 180 seconds waiting for a response containing 132.
```

Reproduce the Hermes TUI timeout on Windows 11 with WSL2 Ubuntu 24.04 using the listed NemoClaw, OpenShell, Docker, and Node.js versions, then compare the request flow with Ubuntu, DGX, or Jetson where it works. Use the inference-route and timeout logs as the initial checkpoints; done means the prompt returns 132 instead of timing out.

Written by the indexing model from the issue text.