I was thinking earlier about the difference between named tamplates and functions in xslt 2.0 and have not found satisfactory criterion for a decision of what to use in each case. I was not first one who has troubled with this, see stylesheet functions or named templates.

To feel easy I deliberately have decided to use functions whenever possible, avoid named tamplates completely, and use matching templates to apply logic depending on context (something like virtual function). I've forgot about the issue until yesterday. To realize the difference one should stop thinking of it, quite opposite she must start solving practical xslt tasks, and if there is any difference, except syntactic, it will manifest itself somehow.

To make things obvious to those whose programming roots are in a language like C++ I shall compare xsl:function with free standing (or static) C++ function, and named xsl:template with C++ member function. In C++ you can use both free standing and member functions interchangeably, however if there is only one argument (among others) whose state transition this function represents then it's preferrable to define it as a member function. The most important difference between these two type of functions is that a member function has hidden argument "this", and is able to access its private state.

Please, do not try to think I'm going to compare template context item in xslt 2.0 with "this" in C++, quite opposite I consider context item as a part of a state. I'm arguing however, of private state that can be passed through template call chain with tunnel parameters. Think of a call tunneling some state (like options, flags, values), and that state accessed several levels deep in call hierarchy, whenever one needs to. You cannot do it with xsl:function, you cannot pass all private state through the function call, you just do not know of it.

This way my answer to the tacit question is:

•  use xsl:function to perform independent unit of logic;
•  use named xsl:template when a functionality is achieved cooperatively, and when you will possibly need to share the state between different implementation blocks;

After thinking through this, I've noticed that such distinction does not exist in XQuery 1.0. There is no tunneling there.