I don't know how this is done, but I have a genaerall idea of how to quickly
solve this.

1. Render text to 0,0 cords.
2. Translate the _picture_ to the cords given by computed values. By computed I

mean from all dependecies that it is done now plus from dx,dy in some next version.

OK. There is a hack for this one, so I say it works for me ;).

The problem is that if one makes a picture so small that has to use fonts like
1px size or something like that, then it won't work. The simplest solution is to
multiply all text position and width values by 10 or 100 (pretty easy to
recalculate by hand). All other elements should scaled by the same value which
might be done by hand or they could be grouped and scaled like this:
<g transform="scale(10)">

non-text elements here

</g>
viewBox also has to be change (multiplying it's values by the same value)

For a hack in works see new version of:
http://test.wikipedia.org/wiki/Image:Text_positioning_test.svg

(resizing whole SVG to 400x240 was only for fitting the test picture in my
screen and not really an issue here, viewBox also has to be changed)