This makes it possible to use these steps independent from each
other. For example, a future patch can get rid of the re-validation
that's done over and over again when the API is called.
A significant change is that this gets rid of an expensive deep
clone. It was necessary before exactly because validation and
normalization was intertwined. Normalized properties would mess with
the later inheritance.
Strictly splitting validation and normalization (and executing them
in this order) solved this. The only downside of this is that
inherited properties are validated twice. But this is much less of a
problem, compared to the deep clone, I would like to argue.
This was always covered by tests. You can still see the tests fail
when you flip the execution order of inheritance and parameter
validation.
Bug: T301337
Change-Id: Ie5728094f9ed813f53b709d8b5283c4b99dc7d63
