I think you can start with a simple regex that pulls those words from the main dictionary that contain only the allowed letters and are not too long. Make the regex a subset of the full test. But that will cull the list of a quarter million words down to a much smaller number, in an in-memory @list.

Then apply the full analysis to only the entries on that list. You might employ subsequent passes with patters generated for the specific case: if you have 1 letter E, look for anything containing two E's and scratch them off, etc. You could generate such a case for each unique letter and then join them together with '|' to make a rejection test. Final pattern would be /e.*e|f.*f.*f|…/ to tell you if the candidate word has more than 1 E, or more than 2 Fs, etc.

So you see the test for more than N of the same letter in your rack is N+1 copies separated by ".*".