Keras 2.8 Model.save default shifts from HDF5 to TensorFlow SavedModel, affecting cross‑version loading

Starting with Keras 2.8, the default `Model.save` output switched from the HDF5 file format to the TensorFlow SavedModel directory structure, changing how models are persisted and loaded.

When loading an HDF5‑saved model created with an earlier Keras release into a newer environment, it is unclear whether explicit declaration of any custom objects is required and whether the SavedModel format guarantees compatibility across Keras versions.

What custom object arguments must be supplied to `load_model` for an HDF5 model saved with Keras 2.7? Can models saved in the SavedModel format be loaded reliably across multiple Keras releases without additional configuration?