Cool Tweedie model is clearly better. You can improve density overlay plots by adding the following after the plot function call
+ scale_x_sqrt()
and for the interval plots add
+ scale_y_sqrt()
These will apply square root transformation for axes, so that it is easier to see details both near 0 and in the tail.
It seems you are using an old version of bayesplot. You would get better ppc_loo_pit plots with the latest CRAN release, or at least add the following to get the correct diagonal line.
+ geom_abline()
Although the Tweedie model is better the LOO-PIT plot shows it’s not perfect. I did guess correctly that the Posterior PIT-ECDF was too good due to double use of data and flexible model. LOO-PIT plot shows that the Tweedie model is slightly overdispersed compared to the data.