If those types aren't enough for your use case, then you will be forced to roll your own types in some binary/string data anyway, so it seems like it's strictly more work if you just always force everyone to roll their own.

And even then you still end up with the possibility of people using the wrong syntax for your hand-rolled types, like not quoting values that are supposed to be strings or quoting values that are supposed to be numbers.

Besides, wouldn't it be easier to fix this just by adding some more types, or deprecating everything except REG_SZ or something? What's the advantage of moving to a directory based model?

File-based configs on Linux have the same problem anyway. The semantics of any config file you find are defined by the application that's consuming it. Any two config files that superficially seem to be using the same format, may in fact use a completely different one - and you'll never know, up until you edit one and it blows up in your face, because it cannot parse empty lines or #-comments, or escape characters, or negative values, or values larger than 2^16, or...

Whether altering Windows registry or a Linux config file, you cannot make a correct modification without knowing what the owners of the modified settings expect.