You can create a submachine representing the regular expression transition conditions, and just attach that at the state you wanted, resulting in a finite state machine.
Of course this doesn't hold when you are talking about non-standard regular expressions, and it's probably a nice feature to have when creating the automata, but IMHO it still sounds like a silly idea when CFG-tools like ANTLR and YACC are available.
Wouldn't that still be a finite state automata?
I also hate working with ANTLR/YACC ... heavy, hard to start with, steep learning curve. These tools are designed for industrial-strength compilers, where performance / flexibility matters.
And PEGs are better than CFGs (that's teeshirt material right there :))
The key is the bit about the "recursive call."
If you've got an algorithm that does that, then it cannot be called "iterative".
I.e. backtracking is recursive, no matter how you implement it.
In school textbooks they do differentiate between "recursive" and "iterative" backtracking, to teach you how to get rid of the call-stack and manage your own. But that's another story.