The script checks for stability by storing hashes of each generation and checks for hash collisions. When a hash collision is found, it marks the pattern as stable and records the generation number of the generation that it collided with. Unfortunately, the script was not properly clearing out the hash list after each new pattern was generated (this is what is fixed now). Thus, if a pattern ever stabilized with an empty field in generation 2 at 98%, that generation 2 would also get marked down for 99%, even though it almost surely stabilized at generation 1.

The good news is that this problem would only ever come up if two consecutive patterns stabilized to the exact same pattern, which almost always would be the empty field, meaning the middle densities (i.e., the interesting ones) were unaffected by the bug. I ran the bug-free script for 1000 trials just to make sure that the resulting graphs look pretty much the same as the ones included with this post, and they do.

Good catch!