### Check for existing issues

- [x] I have searched the existing issues and checked that my issue is not a duplicate.

### What happened?

With a router fallback configured from an OpenAI model group to an Azure one, an invalid `batches.create` request against the OpenAI group is retried against the Azure group, and the caller receives **Azure's** error instead of OpenAI's.

Sending `completion_window="5m"` to the OpenAI group should surface OpenAI's validation error:

```json {"error": {"type": "invalid_request_error", "param": "completion_window", "message": "... must be one of: 24h"}} ```

Instead the response carries an Azure-shaped body with `code=quotaExceeded`, which says nothing about the actual mistake. The caller has no way to work out what was wrong with their request.

Beyond the misleading error, the same fallback means batch creation can land on a deployment other than the one that owns `input_file_id`. Batch and file resources are owned by the provider that created them, so creating batch state under a different provider's credentials than the input file's is not a safe retry target.

**Expected:** file and batch operations stay within the model group tha…