I assume tptacek's objection is that salts need to be unique (at most, unpredictable enough to discourage precomputation), and don't need to be produced by a CSPRNG.
Edit: or maybe something else from the editorial history of the document?
There's that, but more generally and importantly, application developers who take special measures to generate salts tend not to be using secure password hashing algorithms --- the libraries for things like bcrypt tend to handle this for you.
