Question About Propagation Logic Difference Between Training and Inference

Hi, thank you for open-sourcing this great project! I’ve learned a lot from the code, especially the structured modeling and control design.

When reading pgnn_control.py, I noticed something in the run_propagations function:
```
if not (last_step and module.my_gnn.output_len == 0):
module.aggregate_messages()
module.status_update()
```

This condition seems to skip aggregate_messages() and status_update() for some modules (e.g., leg or wheel modules) during policy inference in simulate_policy.py.

But in the training code mbrl.py, the run_propagations function always performs:
```
for module in modules_list:
module.aggregate_messages()
module.status_update()
```

So every module updates at every step.

My question：

From my understanding, all modules should also run aggregation and update during policy inference, otherwise some modules may not complete the intended propagation.

So I’m confused about:

Why does the control code skip propagation when (last_step and output_len == 0)?

Is this an intentional design difference between training and control?

Thanks for your time! I’d really appreciate any clarification on this behavior. 🙏

Compare run_propagations in pgnn_control.py and mbrl.py, then trace how simulate_policy.py invokes the control path. Determine whether the last_step and output_len condition is intentional; done means documenting the expected propagation behavior or identifying a concrete correction.

Written by the indexing model from the issue text.