I believe the latest version has the fewest bugs in it, so I would not recommend trying with older versions. I was able to reproduce the bug on my system. I was somewhat amused when I read your code, because one of the earlier bug reports against core.rrb-vector was also for a marble game, and probably someone writing code to solve this same 2018 Advent of Code Problem #9 that it appears you are solving. That earlier bug was fixed, but you seem to have unearthed another I have not seen.
Frankly, I am a bit skeptical that I can figure out a way to make core.rrb-vector both (a) correct, and (b) meet the run-time that RRB trees claim to be able to achieve.
I have an idea for writing a similar library based on B-trees instead of RRB trees, but it is primarily an idea I have spent a lot of time thinking and writing about, but I have not written code for, so that approach won't help you unless you like the idea of writing such a library.
In the mean time, there is a Java library containing a class io.lacuna.bifurcan.List that _might_ have fewer bugs than core.rrb-vector, but no promises on that point: https://github.com/lacuna/bifurcan