If I'm correct it would take an attacker ~10 years to go through the whole keyspace with a brute force attack. Would this be correct or am I missing something here
First let's consider simple fact: The more power you have the shorter attack will take. There is no such thing as "it would take attacker 10 years". Also, 10 years is very, very short compared to proper key, which is very far from measurement in life-times.
Also, you seem to imply that fact that attacker will have to try multiple files will make it hard. Imagine that I have botnet with 40 000 machines. I can assign one machine for each of your files. So it will just take me as long as it took you to decrypt it (if not better). I can then try most common choices of A and B which would me take maybe day if i'm unlucky (considering that your decryption takes you minute, which is massive amount of time for password-based encryption).
While when you had used password with just 6 characters, I would have somewhat harder time recovering those.
If you assume that attacker knows your file, your whole process is basically useless. There simply isn't enough entropy in A, B and file selection. If you assume attacker doesn't know your file then all you do is still useless, because you would be just as secure in getting first 256-bytes of that file (since you assume that this file is high-entropy, if it isn't consistently high-entropy then you simply hash that file and end in same thing).
Also, one advantage of using passwords is that user himself choose how secure he is. If he uses 4-character password, it will be broken. If he uses 20-character password then there is good chance it will never be broken. With your scheme everyone who doesn't have 20TB worth of storage for keys is limited to 4 characters of security (and with 20TB you won't be much better).
TL;DR: Use established best practices. With your scheme you are arguably worse than with md5.
Also speaking of your scheme, it seems to come from world of "security by obscurity". There is no point in using crc-16 to shorten your key. Scrypt returns high-entropy mixed key, which you can simply cut. But your scheme doesn't spread entropy trough your returned key (which is terrible - you can know part of key with part of file). Also, you use scrypt to convert key to key, which is something pointless, since your key generation will take longer with longer file. So long in fact that user will get bored. Just imagine 60bytes file with A=0, B=1. It will take one minute to get key. You would be far better with just scrypting hash of whole file (you would still be able to configure scrypt time to something reasonable).