Loss is automatically converted?

From the bayesian_neural_network.py example-

# Model compilation.
  optimizer = tf.keras.optimizers.Adam(lr=FLAGS.learning_rate)
  # We use the categorical_crossentropy loss since the MNIST dataset contains
  # ten labels. The Keras API will then automatically add the
  # Kullback-Leibler divergence (contained on the individual layers of
  # the model), to the cross entropy loss, effectively
  # calcuating the (negated) Evidence Lower Bound Loss (ELBO)
  model.compile(optimizer, loss='categorical_crossentropy',
                metrics=['accuracy'], experimental_run_tf_function=False)
  return model

How does TF know to automatically convert the loss to probability and is converting a loss to probability compatible just a matter of adding KL to whatever loss you want to use?

Start with the bayesian_neural_network.py example and its model.compile block. Trace how the categorical_crossentropy loss and layer KL terms are handled, then document whether the example converts a loss to a probability and whether adding KL is generally valid; done means the explanation answers both questions clearly.