> Two designs come into mind:
> 
> Design I:
>    <style name="A">
>        <style-props prop-1="x" prop-2="y" prop-3="z"/>
>    </style>
>    <style name="B" parent="A">
>        <style-props prop-1="v" prop-22="yy" prop-33="zz"/>
>    </style>
> 
> Design II:
>    <style name="A">
>        <style-props prop-1="x" prop-2="y" prop-3="z"/>
>        <style name="B">
>            <style-props prop-1="v" prop-22="yy" prop-33="zz"/>
>        </style>
>    </style>
> 
> 
> In both design the style properties should be interpreted as:
> style 'A': prop-1=x, prop-2=y, prop-3=z
> style 'B': prop-1=v, prop-2=y, prop-3=z, prop-22=yy, prop-33=zz
> 
> What do you imagine is a better design?

I would tend to prefer II, but either is quite viable. The advantage of II
is that you can easily search for the nearest attribute using 

(ancestor-or-self::style/@prop-1)[last()]

whereas design I needs a recursive function (or template) to do the
navigation.

> 
> How can I map it most efficiently to absolute styles?
> Creating a large variable for all mapped styles right at the 
> beginning or map them just-in-time when found accepting 
> redundancy in multiple mapping of the same style?

With Saxon you could get the best of both worlds with a memo-function.

Otherwise it essentially depends on the average number of times each style
is "computed". If the average is much less than one, do it on-demand, if
it's much greater than one, do it eagerly; otherwise, you'll have to measure
to see where the cut-off comes.
> 
> In case the only way is the empiric approach to implement 
> both ways and test which is faster in an XSLT engine, any 
> links to such performance harness to reuse (preferable Java)?
> 

Remember that performance measurements will vary from one XSLT processor to
another. With Java, remember that warm-up time can dominate the measurement
- any measurement run should execute for at least 30 seconds to get
reproducible results.

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