### What happened

When a request is routed through the proxy with model-group **fallbacks** and the primary deployment fails (e.g. a `BadRequestError` / `UnsupportedParamsError`), the request correctly falls back to a secondary deployment.

However, the `model` field reported back to the client differs between non-streaming and streaming:

- **Non-streaming** → `response.model` is the **served (fallback) deployment** — correct. - **Streaming** → every chunk's `model` is the **originally requested model group**, not the deployment that actually generated the tokens — the fallback is **silently masked**.

A streaming client therefore cannot tell from the response body that a fallback happened: `model` reports the requested model while a different deployment served the request.

### Reproduction

Config — primary group falls back to a secondary group:

```yaml model_list: - model_name: primary-model litellm_params: model: openai/<provider-model-that-rejects-some-param> api_base: https://.../v1 - model_name: secondary-model litellm_params: model: bedrock/<some-bedrock-model> router_settings: fallbacks: [{"primary-model": ["secondary-model"]}] litellm_se…