You found a bug. It shouldn't be possible to train the learner with TuneToken present in the parameter set. This has nothing to do with the train-test split. If you are really worried by this, you can check the resampling splits in instance$archive$benchmark_result$resamplings after the optimization.