Assessing the properties of patient-specific treatment effect estimates from causal forest algorithms under essential heterogeneity