Of course TMTOWTDI. I just don't see the advantage of this code over just inlining the sub construct_record code directly in the while loop. Plus, you've increased the number of global variables you're using. There are a couple other things I could nitpick, like that you've got five different regexes all checking for the string START.
This format has some of the nastiest things to deal with. They normally do not occur all at once!
I disagree - I don't find this format nasty and there are plenty of data formats this complicated. Which was exactly my point - a state machine type approach can handle them all. Anyway, as I said, as long as it works you're free to write code like this - I personally still disagree with it ;-)