feat(py): box background operation before wrap_model

[huangjeff5]

Summary

• generate() on a background model returns a ModelResponse whose operation is the start handle and whose message is None. It does not poll until done.
• generate_operation() is that same start and returns the Operation.
• await next_fn(...) in wrap_model is always a ModelResponse, so a hook that reads .message is safe.
• The post-model early-return is gated on ActionKind.BACKGROUND_MODEL, not operation is not None. A define_model LRO that also returns a tool request still runs the tool loop.
• A wrap_model that rebuilds ModelResponse without forwarding operation is FAILED_PRECONDITION (pass through response.operation). It does not blame the model.
• A wrap_generate rebuild that drops operation is the same status with a wrap_generate message. We do not split “middleware ate it” from “start never had one.”
• ModelResponse.__eq__ includes operation, so two start handles with different job ids are not equal.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.

[gemini-code-assist]

Code Review

This pull request introduces support for background models (long-running operations) in Genkit. It updates run_one_iteration to handle Operation responses by wrapping them in a ModelResponse before middleware execution, and automatically sets the longRunning capability flag on background models during registration. Additionally, comprehensive tests are added to verify this behavior. The review feedback highlights two robustness improvements: handling cases where the raw response is a dictionary (e.g., over the wire) by deserializing it into an Operation or ModelResponse, and ensuring that if the model's supports field is a Pydantic model rather than a dictionary, it is correctly serialized instead of being overwritten.