People just don't want to maintain two completely different stacks (one on the server, one on the client).
It's kind of neat that it works. It's also a bit fidgety: the cannibalized code can cause issues (which, e.g. prevented C++11 adoption for a while in some experiments), and now CERN depends on bits of an old C++ compiler to read their data. Some may question the wisdom of making a multi-billion dollar dataset without a spec and dependent on internals of C++ classes (indeed experiments are slowly moving to formats with a clear spec), but for sure having a standard for reflection is better than the home-grown solution they rely on now.
[1]: https://indico.cern.ch/event/408139/contributions/979831/att...
Since then, a lot has changed, and now it is all based on cling ( https://root.cern/cling/ ), that originates from clang and llvm. cling is responsible generates the serialization / reflection of the classes needed within the ROOT framework.