Rules engine performance

The rules engine is a bottleneck. Can we make it faster?

Profile with valgrind using NT format.
We have * for max_length (and related +/-). Could we benefit from having similar variables for min_length? I'm thinking rule rejection, for a start.
Try to improve dupe rejection, perhaps by "unifying" rules a little (unified rule would only be used internally to the dupe rejection stuff).
Note to self: The --rules:NT case-toggler (especially when used with max length less than 14) can be 50-60% faster if we add ->N and >N to it, eg.

-c T8QT[z0]T[z1]T[z2]T[z3]T[z4]T[z5]T[z6]T[z7]

could be

-c ->9 >8 T8QT[z0]T[z1]T[z2]T[z3]T[z4]T[z5]T[z6]T[z7]

Enhance the rules optimizer - and perhaps add a unifier step: If changing things like $1$2$3 to Az"123" we not only gain performance but also help the dupe-rule suppression.
Also the optimizer: Is it possible to sometimes do implicit rule rejection for rules that use TN, DN, xNM, iNX, oNX, XNMI and so on, when max-len is < N (or max-len < I for XNMI)? Would there be side effects? Theoretically we might change something beyond max length and then truncate the beginning of the word. We could however add a mnemonic that either enables or disables (depending on that we'd use as default) such optimizations per line. Or the optimizer (rejector) could possibly be clever enough to make some safe decisions.
...add here

Description