This is probably on the borderline of what should be attempted with just regexen. The approach I'd take would be:

• Use HTML::TreeBuilder to handle parsing the HTML
• look_down the tree for p elements
• Get the text contents of each of those elements and tokenize it on whitespace (i.e. using split)
• Walk over the list of tokens, ignoring simple text, remembering the last mentioned book for wikirefs, and replacing textual references with the corresponding wikirefs
• join the resulting items back together and replace the element contents

When the regex going gets tough, the tough get writing simple parsers and it's not so tough any more.