I had a lot of fun playing the game. I also made a local solver to see how far you can theoretically get, and it's definitely infinite if you play perfectly. Working on the local sim / solver was almost as fun as playing the game :)
After that, I tested a few solvers. Single Player Monte Carlo Tree Search seemed promising, but didn't do as well as I expected. A simple beam search does appear to go infinitely pretty easily.
Shockingly, Claude was able to find the exact random generators that the js library uses, list out all of the differences between the languages that might impact the replays (like rounding being treated differently), and include those differences in the simulation. Claud then back-tested on a huge number of games using the original implementation as a golden set generator.
Considering that 2 new pieces are spawned for each captured one, if you keep progressing you will exhaust the real estate, sooner or later. The only way to go on infinitely is not capturing, but that means not scoring points. Hardly satisfying, though.
For example today's board can be played infinitely by moving between a8 and b7. Zero points, infinite game!