In his example you have several missing value indicator that each contain some sort of reason as to why it's missing. Collapsing that to NA would lose data, which is why the post argue against it.
If it's important and your algorithm can handle it, then it's data, not missing data, and there's no need to be concerned with the underlying representation.
If it's not important or your algorithm can't handle it, then you don't lose anything by collapsing it to NA.
