The alternative is a slew of loosely-coupled and opaque attributes (IDs and classes), especially where you're not sure if a particular class or ID is used for p

The alternative is a slew of loosely-coupled and opaque attributes (IDs and classes), especially where you're not sure if a particular class or ID is used for presentation (CSS) or functionality (JS). How do you handle refactoring when you use some classes for JS and some for CSS? With custom attributes there is no confusion, and although it goes completely against the separation of concerns and now your mark-up isn't completely declarative -- it's a necessary drawback.