Grouping is used specifically to avoid spatial leakage, i.e., to avoid training on a point close to the test point. If you assign each sample its own group, you're defeating the purpose of using GroupKFold. It becomes regular KFold, and spatial bias re-enters. So convergence should happen as clusters approach singleton groups, but that's not a desired outcome if your goal is spatial generalization. GroupKFold isn't meant to approximate KFold. Instead, it's meant to avoid the illusion that your model is better than it really is. So if GroupKFold gives you lower performance, that's a sign of a well-done validation for spatial tasks.