[WSL2][DCode] custom OpenAI-compatible onboarding smoke passes but agent requests fail before reaching endpoint

## Description

On NemoClaw v0.0.124 under WSL2 x86_64, LangChain Deep Agents Code onboarding against a custom OpenAI-compatible endpoint succeeds, and the product's inference smoke probe passes. A real non-interactive DCode request then fails with `RemoteError` category `agent_remote_failure` before any chat-completions request reaches the endpoint.

The behavior reproduces through both a generic custom OpenAI-compatible provider and the compatible-endpoint control path of the native OpenRouter selection flow.

Platform scope: Reproduced on WSL2 x86_64. Other platforms were not evaluated with the same endpoint.

Related: #11806 reports the same terminal `agent_remote_failure` with the NVIDIA provider on Ubuntu 24.04, but this report covers a custom OpenAI-compatible endpoint on WSL2 where the endpoint receives no request.

## Environment

```text
Platform: WSL2 x86_64
NemoClaw: v0.0.124
OpenShell: v0.0.116
LangChain Deep Agents Code app: v0.1.55
Provider: custom OpenAI-compatible endpoint
Endpoint capabilities: authenticated GET /v1/models and POST /v1/chat/completions
```

## Steps to Reproduce

1. Start a reachable OpenAI-compatible endpoint with model-listing and chat-completions responses.
2. Onboard a DCode sandbox with the custom provider, endpoint URL, model, and API key.
3. Confirm onboarding reports that Chat Completions is available and that the inference smoke passed.
4. Run a non-interactive DCode prompt inside the sandbox.

## Expected Result

DCode sends the request through the configured compatible endpoint and returns the endpoint response.

## Actual Result

The LangGraph server starts, then DCode reports `RemoteError` with category `agent_remote_failure`. The endpoint receives no chat-completions request even though the onboarding smoke passed.

## Impact

DCode custom OpenAI-compatible onboarding can report success while the resulting sandbox cannot perform inference.

## Logs

```text
Onboarding provider validation: passed.
Inference smoke probe: passed.
Agent request: RemoteError, category agent_remote_failure.
Endpoint chat-completions request count after agent request: zero.
```

Start by comparing the custom-provider onboarding validation and inference smoke probe with the non-interactive DCode request path after the LangGraph server starts. Reproduce the listed WSL2 flow and trace where the request becomes agent_remote_failure; done means the configured endpoint receives the chat-completions request and DCode returns its response.

Written by the indexing model from the issue text.