From the predictor rankings, we can see that apart from the two multilevel terms (which are merged into a single one here, but that’s not the point), the search goes through 16 non-multilevel terms. In the reference model, however, there are 17 non-multilevel terms. So the search is terminated before the full model is reached. Of course, as I said above, this is due to CV Varsel Error: Infinite or missing values in 'x' - #8 by fweber144, but currently, I don’t see a way to avoid that, apart from providing a custom divergence_minimizer function where you play around with lme4’s tuning parameters until the full model projection converges as it should. This is a tedious work, though (when I played around with your reprex the last time, I already tried to adjust lme4’s tuning parameters, but unsuccessfully). In theory, when running up to the full-model size (and not only until full-model size minus 1), the predictive performance curve could jump up to the reference model’s red dashed line at the last (full-model) size, so the custom search_terms in combination with the search being terminated before the full model is reached could explain the non-convergence. I have to admit, however, that such a jumpy behavior would be strange, too. So when I have the time, I will take a look at your (updated) reprex.