[Ubuntu 26.04][Agent&Skills] OpenClaw tool call fails silently on Ollama-backed sandbox, agent output never written

## Description

On an Ollama-backed sandbox, an OpenClaw agent invocation for a scripted research task fails mid-task with a generic tool-call error, and the expected output file is never written. The sandbox itself remains healthy throughout — only the task invocation fails.

Platform scope: Reproduced on Ubuntu 26.04 (Ollama-local provider) only; other platforms/providers not tested for this specific symptom.
Regression: Unknown — earlier versions not tested.
OpenShell issue: No — the failure is in the OpenClaw agent's tool-call execution, not the OpenShell CLI.

## Environment

```text
Device: Ubuntu 26.04 x86_64 host with NVIDIA GPU
OS: Ubuntu 26.04 LTS, Linux 7.0.0-22-generic
Architecture: x86_64
Node.js: v22.23.1
npm: 10.9.8
Docker: Not captured by installer preflight output (Docker confirmed running)
OpenShell CLI: openshell 0.0.106
NemoClaw: nemoclaw v0.0.114
OpenClaw: v2026.7.1
```

## Steps to Reproduce

1. Onboard a sandbox non-interactively with the Ollama-local inference provider and `agent=openclaw`. Confirm the sandbox reaches Ready and `nemoclaw {sandbox} status` reports the OpenClaw gateway healthy with the Ollama-backed inference route reachable.
2. Upload a small research task description and a text corpus into the sandbox workspace:
```bash
nemoclaw {sandbox} upload
```
3. Invoke the agent non-interactively to complete the task, running `openclaw agent` with a prompt asking it to read the corpus and write a `result.json` summary file to the workspace:
```bash
nemoclaw {sandbox} exec
```
4. Wait for the agent invocation to finish, then attempt to download the expected output file:
```bash
nemoclaw {sandbox} download
```

## Expected Result

The agent completes the task and writes `result.json` to the workspace.

## Actual Result

The agent session printed a generic tool-call failure partway through the task and never produced the expected output file. Attempting to download it afterward reports the path does not exist in the sandbox. The sandbox itself remained healthy throughout (Ready, gateway running) — only this specific task invocation failed.

## Logs

```text
NemoClaw registered
Endpoint: Managed Inference Route (inference.local)
Provider: Local Ollama
Model: qwen3.5:9b
Slash: /nemoclaw

Tool Call failed

Cannot download 'result.json' from sandbox: no such path in the sandbox.
```

Reproduce the failure using the listed Ubuntu, Ollama-local, OpenClaw, and NemoClaw versions, following the upload, exec, and download steps. Start with the OpenClaw tool-call logs and the sandbox gateway status; done means the same task completes and result.json is written and downloadable without affecting sandbox health.

Written by the indexing model from the issue text.