For example, section 3.13 ends with the note:
case x of { (a,_) | let b = not a in b :: Bool -> a }
... Programmers are advised, therefore, to avoid guards that end with a type signature.
I would support changing "a type signature" to "an expression type signature" in this note for consistency.
There is no a priori reason why b should depend on a in a pair of bindings such as these:
a = const (\x -> x) b b = const (a :: Int -> Int) (a :: Bool -> Bool)
There is: section 3.16 says that in an expression type signature e::t, the type derived for e must be more specific than t. To derive the type of e, in this case a, before matching it against t, we have no alternative but to use the definition of a.
Even if you don't think the report is ambiguous, it is at least prone to misinterpretation, which is why a couple of examples would really help at the end of 4.5.1.
Fair enough. Perhaps the example on the H' page would help: http://hackage.haskell.org/trac/haskell-prime/wiki/RelaxedDependencyAnalysis