Thanks for the suggestion! I did try as you mentioned and now get posteriors like this:
I’d say they look better now, or at least closer to what the “shotgun” method produced. Unfortunately traces and energy_plot are still worrying:
(The reason why I did the sum of failures in my original model, was that I didn’t want to “force” particular parts to be the ones that fail. But now I believe you’re right and “forcing” a subset to be the failures without setting that subset apart otherwise doesn’t change the idea behind the model.)