> I think the point of the example was that the expectation was that the
> variable didn't have a set (sequence) semantic but rather was a string
> as if string($f) had been applied, due to the as attribute on the
> xsl:variable (which is a 2.0 feature of course).

The variable was defined as:

<xsl:variable name="foo" select="nothing" as="xs:string?"/>

The select expression yields a node-sequence, the as expression requires
atomic values, so XSLT invokes atomization. The result of atomizing an empty
node-sequence is an empty sequence of strings. 
> 
> But I think it's true to say that as="xs:string" does _not_ force an
> empty sequence to coerce to an empty string, isn't it?

The only conversions forced by the "as" attribute are atomization and
numeric promotion (e.g. int to double). It doesn't cause a cast. If the "as"
attribute had said "xs:string" rather than "xs:string?", a type error would
be reported.

Michael Kay
http://www.saxonica.com/