A parser which consumes the string would be an example application.
I agree that in many cases the caching is going to hide or eliminate the problem, and in fact, it was a little tricky to compose the original example.
However, in code that doesn't need to know the actual length of the string in characters, it's poor practice to use length(), whose cost scales with the size of the string on SVf_UTF8 scalars. The advantage of bytes::length() is that it's basically a lookup on a member of the SV struct. There's a little bit of extra math to deal with offsets, but it's still O(1) rather than O(n).
The ne "" idiom is also O(1) because it's just checking to see whether the total byte size of the string is 0, not counting characters and seeing if the count is 0. Which makes it an effective replacement.