Retrain context_model with SumContextEncoder-based network (time+label folded into one context vector)

Retrain context_model: coordinates-only modality, label as context (drop index head)

Add conditioning-tutorial checkpoint: context_model