I agree that Braille isn't the best example because there is a BCP-47 token for it that would allow it to be coded as the language of the material. The large-print example still seems to apply though.
Note that I'm not saying this would be a property of the ISBN. I'm saying it would be a property of the thing that the ISBN URI identifies.
It turns out that MARC attaches properties to ISBNs, such as the fact that it's a large-print edition or a Braille edition.
This seems like a bad example. It's hard to imagine a useful model where "braille edition" is a property of an ISBN. I would point to this as a nice example of where we need to change practice as we shift to data with model theoretic semantics.