[DGX Station][Agent&Skills] Nemotron exhausts iterations before completing the Blender MCP OVRTX render

## Description

With the local Nemotron 3 Ultra model selected, Hermes can complete simple Blender MCP scene inspection but does not complete the demo's OVRTX render workflow. It exhausts the iteration budget while investigating sandbox files, selecting terminal operations, or inspecting runtime details, and no PNG is produced.

Platform scope: Reproduced on DGX Station only; other platforms not tested.
Regression: Unknown — earlier versions not tested.

## Environment

```text
Device: DGX Station
OS: Ubuntu 24.04
Architecture: aarch64
Docker: 29.2.1
OpenShell CLI: 0.0.72
NemoClaw: v0.0.84
Hermes Agent: v0.18.0 (2026.7.1)
Model: NVIDIA Nemotron 3 Ultra 550B A55B NVFP4
vLLM: 0.22.0
```

## Steps to Reproduce

1. Start a visible Blender 5.1 session with the OVRTX add-on and Blender MCP enabled.
2. Confirm the Blender MCP server is enabled in Hermes and simple scene inspection succeeds.
3. Confirm exactly one OVRTX worker is running and the OVRTX control-plane session list is empty.
4. Configure the Hermes sandbox to use the local `nemotron-ultra` vLLM route.
5. Start a fresh Hermes session and run:

nemohermes {sandbox} exec --timeout 1200 -- \
hermes chat -Q --max-turns 15 -q \
"Use blender-mcp-setup and ovrtx-current-scene-workflow. Preserve the current visible Blender scene. Render a low-sample PNG with OVRTX_EXAMPLE to {output-path}. Clear OVRTX state for the visible Blender process through Blender MCP, wait for asynchronous output to finish, and report the host path and file size. Do not use Cycles or Eevee."

6. Check the requested output path from the host and record both GPU utilization streams.
7. Repeat once with `--max-turns 30` as a diagnostic.

## Expected Result

Hermes should use the configured Blender MCP tool, perform the host-side OVRTX cleanup and render through visible Blender, and return a verified non-empty PNG path within the public guide's 15-turn limit.

## Actual Result

```text
The clean 15-turn run exhausts all iterations after 157 seconds. Hermes investigates the sandbox, incorrectly identifies a Hermes process as Blender, and does not start OVRTX. The output PNG is missing.

Increasing the limit to 30 turns takes 476 seconds and still produces no PNG. Hermes eventually uses Blender MCP and finds the correct Blender PID, but then treats the OVRTX control-plane dependency as a sandbox dependency and does not render.

Performance evidence:
- 15 turns: RTX average utilization 0.0%, peak 6%; GB300 average 92.3%, peak 100%.
- 30 turns: RTX average utilization 0.0%, peak 6%; GB300 average 71.9%, peak 100%.

The high GB300 activity and idle RTX show that the run remains in model planning and never reaches OVRTX rendering.
```

## Logs

```text
Reached maximum iterations (15). Requesting summary...

The summary states that it will proceed with Blender MCP, but the proposed operation is not executed and the requested PNG is absent.
```

Reproduce the exact Hermes command with the local nemotron-ultra route, first using --max-turns 15 and then 30, and review the reported iteration and process-identification behavior. Done means the visible Blender MCP workflow performs OVRTX cleanup and rendering and returns a verified non-empty PNG within the public guide's 15-turn limit.

Written by the indexing model from the issue text.