You need to parse the cdata section using an extension such as
saxon:parse(), then you can use the identity template and a no-op
template to filter out the <br/>'s

I think a parse() function almost made it into 2.0...