docs: explain save vs save_weights for TFQ layers

docs/design.md

@@ -87,6 +87,33 @@ operations. Additionally, you can create a layer with a high degree of
 complexity supporting batch circuit, batch control parameter value, and perform
 batch readout operations. See `tfq.layers.Sample` for an example.
 
+### Saving Keras models
+
+Saving a complete model with `tf.keras.Model.save()` can fail when a layer
+was created with arguments that Keras cannot serialize and the layer does not
+override `get_config()`. For example, a model containing
+`tfq.layers.ControlledPQC` fails to save in TFQ 0.7.2. Depending on the
+format, the error is a `NotEncodableError` or a `NotImplementedError` saying
+the layer must override `get_config()`. The message does not always make
+clear that the layer is the cause.
+
+As an alternative, `tf.keras.Model.save_weights()` saves only the model
+weights, without the complete model configuration. Recreate the model
+architecture before loading the saved weights:
+
+```python
+model.save_weights("my_model.weights.h5")
+
+new_model = build_model()  # Recreate the same model architecture.
+new_model.load_weights("my_model.weights.h5")
+```
+
+For details about the project's Keras setup, see the
+[installation guide](install.md). TensorFlow Quantum's installation
+instructions use the legacy `tf-keras` package via
+`TF_USE_LEGACY_KERAS=1`. Saving and loading behavior may depend on the
+TensorFlow and Keras versions in use.
+
 ### Differentiators
 
 Unlike many TensorFlow operations, observables in quantum circuits do not have