The first of these two phases can happen in one of there ways:
By pre-processing the stylesheet to make a copy
without any css:rule r other CSS Within
elements that would cause problems;
By post-processing the generated HTML to remove the
css:rule elements that were spuriously
generated;
By modifying the XSLT processor so that it ignores the
content of css:rule elements.
The first option is easily done using
fn:transform() and wrapper XSLT, although care
must be taken to pass on any stylesheet parameters. However, it
is not possible in general, because of stylesheet imports that
might be conditional (depending on use-when
expressions or XSLT 3 shadow attributes, for example), and
because package import resolution happens dynamically.
The second option can however be done with
fn:transform(), since the output documents
generated by any xsl:result-document elements in
the transform are not written out but are included in the map
returned by fn:transform() and can be
post-processed.
The author of this paper wrote both XSLT for option (2) and
also a Java class that implements XSLT extension elements for
Saxon which make css:rule and
css:media simply return the empty sequence when
the stylesheet is executed.