[Chore] Add regression coverage for incomplete inference setup

[roomote-roomote]

​Opened on behalf of @mrubens. Follow up by mentioning @roomote-roomote, in the web UI, or in Slack.

What changed

• Added server-side coverage for an OpenAI-compatible /v1/models request returning HTTP 400 with no models.
• Added onboarding coverage proving that this discovery failure shows the provider error without saving model configuration or advancing past inference setup.

Why this change was made

OpenAI-compatible endpoint failures can leave operators repeatedly attempting to continue setup. This coverage pins the intended incomplete state so later setup or provider-validation changes cannot accidentally mark the inference step complete.

Impact

There is no production behavior change. The setup completion contract now has focused regression protection across model discovery and the onboarding interaction.

[roomote-community]

No code issues found. See task

_{Reviewed ff0b50a}