[All Platforms][Onboard][Regression] v0.0.126 NVIDIA Endpoints default model validation returns HTTP 404

## Description

NemoClaw v0.0.126 non-interactive onboarding with the bundled NVIDIA Endpoints default model `nvidia/nemotron-3-super-120b-a12b` fails provider validation with Chat Completions API HTTP 404. The same signature reproduces on three independent Linux systems before sandbox-specific assertions can run.

This is a regression of #10880, which was closed as completed after the earlier fix was verified.

Platform scope: reproduced on DGX Station, DGX Spark, and Ubuntu 22.04. Other platforms were not tested for this exact provider-validation failure.

Regression: Yes.

OpenShell issue: No. The failure occurs during NemoClaw onboarding provider validation before target sandbox assertions run.

## Environment

```text
NemoClaw: v0.0.126
Provider: build / nvidia-prod
Endpoint: https://integrate.api.nvidia.com/v1
Model: nvidia/nemotron-3-super-120b-a12b
Platforms: DGX Station, DGX Spark, Ubuntu 22.04
Architecture: x86_64 and aarch64 coverage represented
Docker: daemon and container networking preflight passed
OpenShell CLI: 0.0.116
```

## Steps to Reproduce

1. Install NemoClaw v0.0.126 on a supported Linux host.
2. Start non-interactive onboarding for a Deep Agents Code sandbox with provider `build`, valid NVIDIA Endpoints credentials, and the bundled default model.
3. Let onboarding reach `Configuring inference provider`.

## Expected Result

The bundled NVIDIA Endpoints default model passes provider validation and onboarding creates the sandbox.

## Actual Result

Onboarding stops during provider validation. The validation probe reports Chat Completions API HTTP 404. A retry with `--resume` recreates the gateway but receives the same HTTP 404. Downstream test assertions do not execute.

## Logs

```text
[3/8] Configuring inference provider
[non-interactive] Provider: build
NVIDIA Endpoints endpoint validation failed.
Validation probe summary: Chat Completions API: HTTP 404.
Onboard aborted before a sandbox was created.

Global setup: onboard retry failed after 1 attempt(s)
(NVIDIA Endpoint returned retryable HTTP 404; exit 1).
```

The same sanitized signature was captured independently on all three affected platforms. Credentials and provider response bodies are intentionally omitted.

Start by reproducing v0.0.126 non-interactive onboarding with provider build and the bundled NVIDIA model, then trace the provider-validation path at “Configuring inference provider.” Compare the current behavior with the fix described by regression issue #10880 and retry with --resume. Done means validation succeeds and onboarding creates the Deep Agents Code sandbox without the HTTP 404.

Written by the indexing model from the issue text.