Some theory:

For a minute I thought this can be trivially solved by counting the normalization of all words (<=8) from the dictionary in a hash... e.g.

  DB<211> join "",sort split //,"electron"
 => "ceelnort"

  DB<212> join "",sort split //,"elector"
 => "ceelort"
[download]

As next step successively the count from all smaller words had to be added to covering words, e.g striking the "n" from "ceelnort" leads to "ceelort", so $count{ceelnort}+=$count{ceelort}

But than I realized that the best covering word from the dictionary is not necessarily the best solution.

take this counterexample for 3 out of 4 letters, the number showing the coverage-count

1 a 
1   b
3 a b
2 a   c
2   b c 
4 a b    d
[download]

so the word (a,b,d) is the maximum with a count 4, but the set (a,b,c) would cover 5 words!!!

(yes this also works with repeated letters)

IMHO this problem belongs to the family of Maximum coverage problem and Set_cover_problem, so finding a guarantied best solution shouldn't be trivial w/o brute force.

OTOH adapting the sort order of letters might already lead to very good solutions...

update

Maybe you can use the above count hash to solve the dual problem:

"which of the n-8 letters cover the minimum of words" (n including repetition)

E.g. "d" is in only one out of 6 words with 4 letters => the remaining 3 letters cover 5 words.

"c" is only in 2 remaining words => (a,b) cover a maximum of 3 words and so on.

Not sure if this leads to the guarantied best solution, sounds to easy... =)