Yes, I had a fix like this... but it was completely broken when I tested it with ISO-8859-1 extended characters (aka 'é'), the 1024 figure was not right. So I added a few layers of complex calculations, cursed a lot, wrote more tests directly on XML::Parser... until I realized that the solution was a lot simpler: for CDATA sections, the string passed to the character handler is in the original encoding (I had been staring at such strings for over an hour when it hit me!). So there was no need to do all this, just to use the usual string within a CDATA section... et voilà!

Thanks for looking into it though.