The search space for a length n is: 2 x 26^(n/3) x 10^(n/6) x 30 x 26^(n/2) = 60 x 26^(5/6 x n) x 10^(n/6) = 60 x 10^(5/6 x n x log10(26)) x 10^(n/6) =~ 6 x 10^(1.17914n) x 10^(0.16666n) = 6 x 10^(1.34581n)
So that's about a 6 followed by 1.5n zeroes of search space for length n, approximately. Using only 100 characters of the ASCII code tables, you have a search space of 10^(2n) possible passwords, which is about n/2 orders of magnitude larger than Haddock's search space.
Disclaimer: I'm a HS student, no crypto folk, so the above might be wrong, but it's basic combinatorial stuff. And that's just theory, in practice, with just "password1", "letmein" and "123456" for a dictionary attack, you can hack into like 3% of the accounts of a site that doesn't use Haddock.
Edit: Oh, I didn't even think of this: THE WORDS ARE ALL VALID ENGLISH WORDS! That makes hacking in a lot easier.
In the end, though, these passwords should be "secure enough" for most purposes, and should be generally more secure than the password of your average user. Using haddock to generate temporary passwords will hopefully just encourage users to be better about passwords in general.
Also, it may be that giving users memorable passwords beats letting them make their own, whereupon many of them choose weak ones.
First attempt: quick typing, conscious mistake.
Second attempt: quick typing, wrong password again.
Third attempt: okay, I'll type it in carefully now. Wrong password.
Fourth attempt: oh, yes, that was my Yahoo! password. Remembers password and enters it cautiously, because he/she knows that he has a maximum of 5 login attempts.
Five are enough for me. Or at most 10. 100 seem ridiculous. How often does it happen that you mistype your password 99 times and get it right on the one-hundredth, before you're tired of being shown the "wrong password" message and click the "forgot password" link?
Haddock uses the UNIX wordlist with a known algorithm for password generation: words = %W(#{random_word} #{random_symbol}#{random_word}) if I'm reading the source right. It seems like it's reducing the search space drastically for a gain in memorability that is difficult to quantify. Might be a win, but might be a loss.
But I don't really know this stuff so well-- any real crypto folks care to comment?