I.e., use the non-normalized name everywhere but inside backends: document references, error messages, ...
If necessary, detect and handle duplicate index names in the backend.

First, because this should make it easier for users to switch backends: the name of the index, from their point of view as developer of the application using Hibernate Search, would be exactly the one they defined, regardless of the underlying technology.

Second, because it's hard as hell to test this whole thing in the TCK without knowing the actual name of indexes we defined ourselves...