The important point is that package types (the ... in a type expression (module ...) or in an expression (module M : ...)) are not module types. Syntactically, they are indeed a subset of module types, but it is a strict subset, and they obey different rules than module types.
Package types live at the boundary between types and module types. In a packing expression (module M : ...), the package type must produce a well-typed module type. This is why you cannot write (module M : S with type t = 'a), because "S with type t = 'a" is not a proper module type. But (module S with type t = 'a) is a correct type.
Package types must also support all operations on types, including equality, unification, etc. For instance, unifying (module S1 with type t = 'a) and (module S2 with type t = 'b) is defined as checking than S1 and S2 are equivalent paths and unifying 'a and 'b. The fact that a package type can also be seen as a module type is nowhere used in this definition. (Of course, one must be careful when defining equality of package types that it does not allow to cast a module from one module type to an incompatible one.)
There is some flexibility in both the definition of admissible package types and the definition of equality between package types.
For instance, I don't see any problem or difficulty in your proposal of allowing constraints on types nested in sub-modules (but this should be checked carefully). Feel free to open a feature request for this!
Will do.
Similarly, one could allow "with type t :=" constraints in addition to "with type t =". Is there a need for that?
I haven't seen one yet, but I'll keep my eyes out.
One could also relax equality of package types (module S1 with ...) and (module S2 with ...) by checking that S1 and S2 expand to structurally equivalent module types (with the exact same ordering between value components!).
Would that replace the dependence on paths that you describe above? The use of paths has turned out to be pretty fragile in my experience, with module types that should have been the same showing up as different in the presence of rebinding of module signatures, which we would normally do as a way of improving concision and readability.
y