As a shortcut, the two-body problem is solved by time-sampling rather than polynomially. Time-sampling is limited by processing speed on the computer running the simulation; typically to 60Hz but potentially less on lower-powered devices and especially if your browser de-prioritises the thread (e.g. if you switch tabs mid-descent for a bit). As a result, the simulation varies slightly from system to system, probably by enough in real-world cases to make a few % difference in optimal score. I could calculate the optimal score for ideal conditions (and I think it’d be close to 50%), but players on slower devices might be able to get a little higher with perfect play owing to the calculation shortcut.

An earlier version didn’t have this limitation, but performance was choppy on my mobile so I simplified the simulation.

tl;dr: it varies slightly by device.