I suspect that any attempt to produce a hierarchal top-down style tree of mathematical subjects is bound to be problematic. You might make the majority of the people happy at the time you create the tree (like the arXiv now) but long-term it's likely to have the same problems that people see with the MSC classification.
The history of mathematics is that deep connections are found between fields that are at earlier times perceived as distant. So fields glom together. Similarly, fields drift apart (like math.GT and math.AT now, in the 50's there was just topology) for various reasons, evolution of techniques being one of them.
My initial guess would be that the best long-term solution would be to have an MSC classification that is as "flat" as possible. So that when you pick your MSC classification, it's like choosing flavours at an ice-cream shop.