test010 has a weird failure to find a good diff. This new test, when setting the permitted_state_size = 100, gives a shorter failure that test010, to allow easier debugging.
When forwards and backwards traversals meet, the endpoints of the mid-snake are not the two points in kd_forward and kd_backward, but rather the section that was slid (if any) of the current forward/backward traversal only. For example: o-o-o | | o A | \ o o \ M |\ o o-o-o | | | o o X \ o \ o \ o The backward traversal reached M from the bottom and slid upwards. The forward traversal already reached X, which is not a straight line from M anymore, so picking a mid-snake from M to X would yield a mistake. The correct mid-snake is between M and A. M is where the bottom traversal hit the diagonal that the forwards traversal has already passed, and A is what it reaches when sliding up identical lines.
In debug output: fix some indents, fix printing of max state size, in myers graph, limit x axis labels to 2 digits, print colored markers of current myers graph positions, also print myers-divide positions in myers graph.
The macros sometimes returned unsigned int (atoms.len) or long int ((ATOM) - head). Make that always unsigned int, with some range checks against negative values.