Realistically if you take into account the possibility of using up to four works, with some but not all vowels replaced with numbers, a mixture of upper/lower, and some other random numbers and special characters in an undisclosed place... that creates such a huge number of possibilities that it might as well not be dictionary based from the cracker's point of view.
Write your scheme down, compute the number of permutations it makes available, then divide by the total number of possibilities allowed by your password set (a straight [possible_chars]^[length] computation).
The results will surprise you. The space of things you describe is far smaller relative to the available space than you realize.
You can indeed make this big enough to work, and it's easy to make it big enough that only a dedicated cracker could get through it, if that. My point isn't that you're wrong, but that you may be less right than you realize.
The thing is, even if it's true for the one scheme I suggested, unless you know the format a password will be in you can't aim directly for that scheme, meaning you can't limit the number of possibilities that much.
I could pick a password that is "coRInStaNdr3ws19@90" which is my first name, my first school, my year of birth, and a few capitals/numbers/special characters. If an attacker knows that's my type of password then it might make it easier for him, but 'corin' isn't in dictionaries (sure they might think to find out my name and add it as a dictionary word), and I doubt "st" and "andrews" would be. Realistically, unless I give away my rough password format, they won't be able to narrow it down and will be left with a password that's just as hard to crack as if it was completely random.
You're making a false assumption that the hacker must somehow divine your pattern, but that is not necessary. An attacker could mount a Markov-chain based attack based on common phonetic patterns (and even common numbers, since some will show up more often than others) and radically cut down the search space even without assuming any aspect of your "pattern", and what you cited will get caught in that orders of magnitude before a truly random password. Along with enormous numbers of patterns that you could cite, as well, this is pretty much a superset of them all.
When you cut the possibility space down, you have cut the possibility space down. You can't fool entropy. The math is quite vicious that way.
You're encountering the "everybody can create an encryption scheme that they themselves can't crack" problem. You may not be able to think of how to abuse low-entropy passwords to crack something far longer than you "ought" to be able to, but that doesn't make it impossible, or even necessarily hard for an attacker.
