Thanks for your quick reply. the node identity comparison helped quite a
bit, although I am still around a minute for a full book of ids. I am
not sure how xsl:number would help here, and what kind of performance
win it would give over count(). I tried something with a nested
transformation, but what should I feed it?

 B B B  <xsl:number select="*[last()]"/>
works (given a set of preceding nodes) but it is slightly slower than a
count() in the xquery. Maybe I should be using xsl:number differently?

It is difficult for me to suggest that without knowing the XML input
structure and whether you want to generate that id based on a count or
numbering only for certain nodes or some particular element type. In
general if I wanted to delegate counting to xsl:number similar to your
function I would define a template in a mode for that e.g.

   <xsl:template match="*" mode="number">
      <xsl:number level="any" from="*[@id]"/>
   </xsl:template>

Both the template or the or the select of the apply-templates can of
course be adapted to more particular needs.

As for being more efficient that using count, that then depends on the
implementation but I would think there is some optimization to be
expected in an XSLT processor for xsl:number.