My suspicion is that the kernel density estimate of the mean is using a bandwidth that is too wide to show you any evidence of the non-zero values within that very narrow window. You can already see that the blue “curves” are basically just a couple of straight lines, suggesting that even when plotting the predictions for an individual sample it is struggling to illustrate this tiny bump in the predictions. If you want to check this out, you can use calculate the means of studentt_ppc by hand and then use az.plot_kde() to plot them, tweaking the bw argument to see if you can get the bump to show. Or you can just plot the raw means.