Another idea I have been toying with is the idea of lexing the source into an array of tokens. Then, compiling the token list into a string using a single character to represent each token type. I then could regex against the string (effectivly regexing for token patterns). Then use the indexes found in @- and @+ to return the tokens represented by the captured groups. I wrote a nice object-oriented version, complete with operator overloading, but here is the main idea of the code:

@tokens = ...;
$tokens = join '', map { $tokenToChar{ $_->type } }, @tokens;

if ($tokens =~ / ($IDENT)($DOT $IDENT)* /x) {
     my @captured = map { [ @tokens[ $-[$_], $+[$_] ] } 1..$#-;
     my ($head, $tail) = @catpured;
}
[download]

This worked beautifully. Since a three to four hundered character QL string could be represented by, maybe, 80 tokens, the token string is very small. All the regexes are on single chars so backtracking is minimized. And I could use character classes [$TOK1 $TOK2] to simulate token classes. All in all, it seems like a very efficient approach. I may even be able to apply this technique to P::RD.

Thank you rhesa. I will read through the P::RD FAQ on optimization!

And thanks for all of the feedback so far.