I don’t think there will be any debate about the design and implementation of the type, but I strongly disagree with the modeling motivation (not to mention the heavy use of the royal “we” when discussing that motivation).
Imposing a zero-sum constraint on individual parameters addresses identifiability issues only by pretty substantially changing the assumed population model.
For example from the mathematical perspective it breaks infinite exchangeability/conditional independence of the model, making it fundamentally different from classic hierarchal models. One consequence is that the zero-sum model cannot consistently generalize beyond the given contexts and so inferences and predictions can no longer be made for new contexts.
From a more statistical perspective the zero-sum constraint fixes the population mean to exactly the average of the individual parameters which is extremely unlikely to happen if the population model is actually conditionally independently normal.
As with so many popular hacks the results are achieved only at the expense of fundamentally changing the model which is, in my opinion, very poor practice. That said separating the design/implementation from these (at least what I consider) controversial modeling applications would avoid these issues in the software discussion entirely.