It's actually much weirder than that: banking changes the axis of rotation and thus kills the rotational inertia. The tracks bank super aggressively in order to prevent the ball from accelerating too much and hopping the track. This is part of why the descent is so smooth and all the balls move at more or less the same speed.
Also to be fair the final system does lose a ball every 30ish minutes. The tuning was largely me staring at the run or taking a video trying to catch where they get lost. Instead of hand tuning I would just update the generator and print another one. I'm considering closing the loop with a camera but that would be a whole new project.
First, I thought about Ansys or CATIA software but I couldn’t find any module specialized for simulation of balls.
But I think that people from those companies could help as well and participate in simulation as an interesting usecase. (These software are expensive for personal projects.)
My point was that these software could help to find weak parts in trajectory - so instead of trying to figure it out by looking where balls are too quick to fall from the ride - you can simulate it. I saw real tramway simulation done in Ansys.
I think the physics are different, a ball is basically a car without a differential, so it's going to behave differently on the tracks. I'd imagine the ball is harder to simulate because of that.
One of the results for hilbert curve marble tracks, mentioned elsewhere in the thread, was a video showing how to make one in blender, which has a physics engine so it can simulate it pretty well.
But for hobby purposes I would suggest to contact some university, they have such software, and they could find simulation of balls motion at marble fountain interesting for research (and educational) purposes.
I haven't actually measured it but that's a good thought, I may borrow a thermal camera and do some testing! It's not noticeably warm to the touch but this functionally a system that converts potential energy into heat and sound so there's probably a measurable change.
Also to be fair the final system does lose a ball every 30ish minutes. The tuning was largely me staring at the run or taking a video trying to catch where they get lost. Instead of hand tuning I would just update the generator and print another one. I'm considering closing the loop with a camera but that would be a whole new project.