compress: better handle `dictBuilder overflow` case

If we producing many good pattren-candidates: they will come to dict-builder.
And dict-builder will sort them `first by score then alphabetically` (and evict tail).
it means high chance to see in dictBuilder `00` and `001` words (begin of alphabet) and don't see `ff` (end of alphabet).

Need: or better filter candidates before dictBuilder, or improve dictBuilder sorting logic, or do any other tricks. Or add one more limit: "soft limit of dict builder", "hard limit of dict builder", "hard limit of final dictionary".