But where is the ultimate limit to this idea?

I didn't find an answer at first, but for me you would want the anchor to be at most the repeating pattern (because a longer anchor will have the same ratio in the data). So you could keep as an anchor either the longest non repeating sequence so far, or as big a sequence as memory will allow.

Actually if your anchor isn't too uniform, rather than compute checksums at each successive position you could then use Boyer Moore to "jump" to the next occurence of the pattern.

If the candidate pattern is too small, let's say N values, just read N values at a time in a loop until you find something that's different from your candidate. That something is probably a good choice of anchor (or as you said /($pattern)*($something)/).