This is why MD5 is not a good password-based key derivation function. :-)
Using punctuation does not make a password "more random", it only makes the random choice set slightly larger. Bumping the length is much more significant than bumping the set as it's exponential.
The debate is whether it's easier for a user to both pick and remember a "random-enough" - i.e not easily bruteforce-able, regardless of hashing technology - [A-Za-z]{10} password or a [A-Za-z0-9!@#$%^&()_-+=;':",.<>/?\|[\]{}]{8} password.
For the record, Google's 2-factor auth generated app-specific passwords are [a-z]{16}, which gives 1E22, or 230 billion years with bcrypt at 0.3s.
For instance, a plain alphanum password of length 9 is stronger than a password of length 8 that allows for special characters. (because 62^9 > 94^8)
In other words, adding one character to your password is as good as choosing from a bigger character set.
Okay, beginning at length 10, you'd need to add two characters, but that's really all you need to do up to length 20. (which is an insane length for a random password anyway)
So the problem is that passwords tend not to be random. This greatly decreases the problem-space. But b/scrypt solves the problem of length, within reason.