[Linux][Agent&Skills] openclaw agent second turn exits successfully without an assistant response

## Description

Running two `openclaw agent` commands with the same explicit session ID succeeds on the first turn, but the second command exits successfully without producing an assistant response or recalling the first turn.

Platform scope: Reproduced on DGX Spark (aarch64) and Brev (x86_64); other platforms not tested.

Regression: Unknown — earlier versions were not tested for this exact symptom.

## Environment

```text
Device: DGX Spark / Brev cloud instance
OS: Linux 6.11.0-1014-nvidia / Ubuntu Linux 6.8.0-1064-gcp
Architecture: aarch64 / x86_64
Node.js: v22.22.3 / v22.23.2
npm: 10.9.8
Docker: Docker Engine 28.3.3 / 29.6.2
OpenShell CLI: openshell 0.0.85
NemoClaw: nemoclaw v0.0.103
OpenClaw: 2026.7.1
```

## Steps to Reproduce

1. Install NemoClaw v0.0.103 and onboard an OpenClaw sandbox.
2. Run the first turn:

```bash
openclaw agent --agent main --json -m "Say hello in exactly three words" --session-id context-test
```

3. Confirm that the command exits 0 and returns a three-word assistant response.
4. Run the second turn with the same session ID:

```bash
openclaw agent --agent main --json -m "Now repeat what I just asked you" --session-id context-test
```

5. Inspect the second command output and the session history.

## Expected Result

The second command returns an assistant response that references the first request, and the response is appended to the same session history.

## Actual Result

The first command exits 0 and writes both the user message and assistant response to the session. The second command also exits 0, but its output contains only the submitted prompt, runtime warnings, and the gateway registration banner. It produces no assistant response that references the first turn.

The result reproduces on both DGX Spark and Brev.

## Logs

```text
First turn session history:
{"role":"user","content":"Say hello in exactly three words"}
{"role":"assistant","content":"Hello there, friend!","stopReason":"stop"}

Second command:
openclaw agent --agent main --json -m "Now repeat what I just asked you" --session-id context-test
exit: 0

Second command output contains the submitted prompt and gateway registration banner, but no assistant message. Repeating the second command produces the same result.
```

Start by reproducing the two documented `openclaw agent` commands with the same `context-test` session ID and compare their JSON output and session history. Trace the second-turn session handling around the gateway registration banner, then verify that a response referencing the first request is produced and appended to the same session.

Written by the indexing model from the issue text.