ppmx v0.04 is here!

[osmanturan]

But if you want some collisions, then masking out context bits is the way to do it. For example, mask bit 5 in text (upper/lower case) and low bits in images and audio. It's hard to do that by choosing magic multipliers.

Yep. Current context masks are full order-n now. After implementing a data detection routine, I'll optimize context masks for certain data types.

[encode]

Added SEE to PPMX:
calgary.tar -> 763 KB

[encode]

Done some funny SEE experiments.
Experimented with a nice scheme:
The escape count provided by the classic PPM estimator translated to a probability. After such probability is adjusted by the three independent dynamic mixers. And the escape is coded as a flag with this adjusted probability.
After all, I made a new scheme which is kind of an approximation of a scheme described above. I do not translate the escape into a probability, instead I adjust the escape count on the fly. This makes PPMX faster not just due to such SEE scheme is far more simpler, but also because we able to code a symbol or escape in just one range coder call. Now, I'm running new SEE thru my optimizer...