Partial functions and exceptions are a compromise solution for the fact that you sometimes do know more than the compiler does. I think it's fine to throw an exception in the case of "programmer error". It's the equivalent of assertions in other languages. Yes, it can blow up, but at least the error is a bit more localised.
Having head return a Maybe means that you'll have to awkwardly handly a Nothing case even in situations where there is no sane behaviour to be added because it just simply would make no sense for a particular list to be empty unless you've introduced a bug somewhere else. It's hard to "recover" from such an error.
The same goes for e.g. division, which is partial too (can't divide by 0), but having it return Maybe would make arithmetic incredibly awkward. You could instead define e.g. x/0=0 or any other value—some languages like Coq or Pony do that, but I think that has the drawback that this makes it rather easy to mask some ugly errors.
In many such cases, the Haskell type system (without advanced extensions) is not expressive enough to encode everything you know about your values. In a language with dependent types, such as Idris, you can specify the length of the list in your type; then you can have a type-safe, total head function that doesn't return Maybe. You can also write a division function that requires a proof (possibly implicit) that the denominator is not zero. But dependently typed languages are much more niche than Haskell.
Having partial functions in the Prelude is, as far as I know, widely regarded as a mistake and they are only kept around for backwards compatibility. Anyone writing code nowadays should be using safeHead or non-empty lists.
Or pattern matching that takes account of the empty list case. Orrrrr using a fold! I usually find when I start matching on list values that the function could be better expressed with a fold instead.
Having head return a Maybe means that you'll have to awkwardly handly a Nothing case even in situations where there is no sane behaviour to be added because it just simply would make no sense for a particular list to be empty unless you've introduced a bug somewhere else. It's hard to "recover" from such an error.
The same goes for e.g. division, which is partial too (can't divide by 0), but having it return Maybe would make arithmetic incredibly awkward. You could instead define e.g. x/0=0 or any other value—some languages like Coq or Pony do that, but I think that has the drawback that this makes it rather easy to mask some ugly errors.
In many such cases, the Haskell type system (without advanced extensions) is not expressive enough to encode everything you know about your values. In a language with dependent types, such as Idris, you can specify the length of the list in your type; then you can have a type-safe, total head function that doesn't return Maybe. You can also write a division function that requires a proof (possibly implicit) that the denominator is not zero. But dependently typed languages are much more niche than Haskell.