My first version advanced the solver once per frame with dt = 1/60. For stiff or fast systems, that step was too coarse and the energy drifted or blew up. So, I split each frame into 20 substeps of dt / steps (about 1/1200 s each). Each simulation can override the count with custom_steps, so lightweight systems stay cheap and touchy ones get more precision. Here we go, our new simulation_runner.py.