And transformations of unimodal distributions can be multimodal. (Take the inverse logit of a diffuse Gaussian)
- They require density estimates, which are much much noisier than CDF estimates especially in the tails (which is the only region we care about).
Basically HPDIs are a computationally burdensome, minimally interpretable, and highly contextual bad solution to a non-existent problem and should be left in the darkest ditch of history.