A whole lot of the "parse, don't validate" ethos is potentially about handling incoming data with complicated formats by writing a minimalist interpreter that reads the input and outputs either guaranteed-valid domain objects or error messages. Most the time you can get away with just a parser, but one occasionally runs into more complex formats that are easier to handle if you think of them as a declarative language that you handle using general-purpose interpreter implementation patterns. And it's just not a wheel worth reinventing.

The "very clever JSON" example the parent poster gives isn't particularly farfetched. The description reminds me of, for example, every Jetbrains REST API I've ever had to interact with. Also the data file formats for some commercial applications I've had to interact with. I don't like formats like this, but when I'm stuck with them, choosing the most maintainable option from the choices I've been left with is not a code smell; it's just sound engineering.