wrong hole order in the presence of where clauses

[lclem]

The holes in the following fragment

module code.hole-test where

f : ∀ {A : Set} → A → A
f a = ? where
    g : ∀ {B : Set} → B → B
    g b = ?

are reported by Agda as

?0 : B
?1 : A

while the kernel expects them the other way around. Fixing this requires analysing the nesting structure of the where clauses.