The specification of fn:idref() (in 15.5.3) says:

A node $N has an IDREF value equal to V if both of the following conditions are
true:

    * The is-idrefs property (See Section 5.6 is-idrefs AccessorDM.)of $N is
true
    * The sequence fn:tokenize(fn:normalize-space($N), ' ') contains a string
that is equal to V ...

Now if $N is a node that has acquired the is-idrefs property by virtue of
having the schema type xs:IDREFS, then the call to normalize-space() will
implicitly atomize $N to produce a sequence of atomic values of type xs:IDREF;
and if there is more than one such atomic value, the call will fail with a type
error, because its signature expects an argument of type xs:string? (that is,
it does not accept a sequence of more than one string).

I think that what is intended is 

    fn:tokenize(fn:normalize-space(fn:string($N)), ' ')