[WSL x86][Agent] Long streaming request ends with an empty agent response after more than 60 seconds

## Description

On NemoClaw v0.0.125 with OpenClaw and the managed NVIDIA Nemotron endpoint, a long reasoning request remains active for more than 60 seconds but does not produce a valid completed response. The command eventually returns the user-visible error that the agent could not generate a response.

Platform scope: Reproduced on WSL2 x86_64 only; other platforms were not tested for this symptom.

Regression: Unknown.

## Environment

- Device: WSL2 x86_64 test host
- OS: Linux 6.6.87.2-microsoft-standard-WSL2
- Architecture: x86_64
- Node.js: v22.23.1
- npm: 10.9.8
- Docker: Docker Engine 29.5.3
- OpenShell CLI: 0.0.116
- NemoClaw: v0.0.125
- OpenClaw: 2026.7.1
- Inference: managed NVIDIA endpoint using `nvidia/nemotron-3-super-120b-a12b`

## Steps to reproduce

1. Install NemoClaw v0.0.125 on a WSL2 x86_64 host with Docker Desktop integration.
2. Onboard an OpenClaw sandbox using the managed NVIDIA endpoint and `nvidia/nemotron-3-super-120b-a12b`.
3. Confirm the provider smoke check and sandbox readiness succeed.
4. Send a reasoning prompt intended to stream for longer than 60 seconds.
5. Wait for the command to finish and inspect the returned response.

## Expected result

The request should continue streaming beyond 60 seconds and finish with a non-empty, substantive response. A long but active stream should not end as an empty agent response.

## Actual result

The request runs for about 70 seconds and then returns no substantive model content. The visible response is `Agent could not generate a response`. Provider preflight, installation, onboarding, sandbox readiness, and SSH connectivity all succeeded before the request.

## Logs

```text
Provider: NVIDIA Endpoints
Model: nvidia/nemotron-3-super-120b-a12b
Inference smoke passed
Sandbox ready

Request duration: approximately 70 seconds
Final response:
Agent could not generate a response. Please try again.
```

Start by reproducing the request on WSL2 x86_64 with the managed NVIDIA endpoint and nvidia/nemotron-3-super-120b-a12b, then trace the streaming request path around the 60-second boundary. Compare the active stream with the final empty response and verify that long requests finish with substantive model content rather than the visible agent error.

Written by the indexing model from the issue text.