I'm having some issues with the interpretability of models drawn with bayesnet. This is more a suggestion for improvements, than an error report, but anyway:

Having identical shapes for latent and observed variables, makes distinguishing between them very challenging, and only doing so with different shadings does not help when you have suboptimal print or when people have problems with their eyes.

Is it possible to change the default observed variable shape to a square? Or at least provide an option for it in the package that does not involve making all sorts of customizations? This would also make the library more compatible with how DAGs are usually presented in for example structural equation modelling.

As a bonus, this would then free up shadings to be used in other ways, to communicate emphasis on certain parts of the graph.

Issues aside though, i really like your tikzlibrary. :)