> Or is it that an attribute node is on the self:: axis when it 
> is the context node but you can't select it using a name test?

Correct.
> 
> Another small but significant wrinkle is that 'self::change' 
> works in xsltproc.

I assume you mean that it selects the attribute, i.e. it's broken.
> 
> However, updating to use XSLT 2.0, this gets the blessing of Saxon:
> 'self::attribute(change)'.

Yes, that one is worth remembering. Though I prefer (@* except @change).

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