ID can appear in many places in the genome, and makes "gaps" that can be
filled by other operators when mutating. This allows programs of 20
genes to produce short code, with many identities, and learn stably on
SimpleGoal

"rank" simply does not work nor learns anything...

Randomize programs in a fixed way: values close to x.5 get cast to int(x), values further away may slip to x-1 or x+1

Add negative versions of functions that have a positive image, so, on
average on random DNAs, the output of a program is not biased towards
positive numbers.