> Most probably, any XPath engine has some *initialisation* 
> phase, which is performed once and forever. I think that a 
> not bad definition for the term "initialisation phase" would 
> be: "A set of actions to build the context for subsequent 
> stages of the system".

I'm probably thinking primarily of use cases where the XPath expressions are
written by someone different from the author of the application - for
example, where XPath expressions might occur in some data file to calculate
say the shipping cost of an order: the problem with current mechanisms is
that namespace bindings can be defined by the author of the application that
calls saxon:evaluate() or its Java equivalent, but not by the author of the
XPath expression.

It's true of course that such a data file might contain a number of XPath
expressions and you might want to define the namespace bindings that apply
to all of them. But I don't want to overcomplicate things. One of the
benefits of an interface like saxon:evaluate() is that it is stateless.

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