You may want to look at my solution as it will likely be the heat death of the universe before your code finishes if you wrote it in Perl. I too considered divide and conquer approach but abandoned it for simplicity and the fact that I can do millions of iterations per second in C.
you're right with all. It is the result of putting code together when it's too late. I let the program run with Devel::NYTProf on a small subset of the dict. The result is what I assumed: Answering the question if a combination of letters is able to produce a word is the most expensive task. I tried to make it faster with a C-like implementation in perl, but it was slower than the hash-approach.
The program is still running. Currently at iteration 879000. So, a long time to go... ;-)
Anyway, a very very intersting puzzle. And it was fun looking at the different approaches and seeing that some of the "old" monks took that challenge.