Simulation suffers too much from the barnes-hut approximation. Body position deviates too much from the brute-force O(n²) approach.
I'm unsure whether this is being caused by bad implementation or if this much loss of accuracy is expected. When collisions are disabled some bodies can also slingshot sometimes which indicates there's something odd going on with the implementation itself.