Instead of stubbing out the policy framework, stub out the results
given to us by our subclasses. This allows us to test the redaction
logic itself. In so doing, we find a few mistakes.

By applying redaction only to a few known classes, it turns out we're
skipping it in a number of desirable cases. It's better to apply it
unconditionally, providing a fallback for cases where a plain array has
been returned to us.