It’s likely due to the lower number of joints and lack of any external constraints - it’s basically a simple case where a single solver iteration already gets the exact required answer.
When we have cases like a balancing quadruped, there are multiple contacts and joints involved, so a single solver iteration will be unlikely to reach a global solution for this system. Therefore, we end up in a state where the solver is continuously revising the forces being applied by the contacts and joints in a feedback loop. As these forces propagate through the system and the system is integrated, small positional or velocity errors accumulating, which result in additional forces being applied by the constraints next iteration to correct them.
If the solver is given more iterations, it should result in less error both in terms of positional and velocity error.