Thanks for your answers guys. The math is indeed a bit odd. Since I have 33 projections surfaces which all have differents ratios, I can tell that the z translate value is "ratio dependant", but it is still unclear in which ways. And it clearly depend of the 3D projector style.

In the mean time, I progressed quite a bit. Here is what I came up with, as a reference and in the hope it might help others :

- First, I ended up using 3D rect project because it was more suited for the tilting effect I was trying to achieve. (you can easily change the horizontal and vertical reference, which allows to rotate around the border of the image)

- As the 3D rect is more of a 3D square (doesn't respect the texture ratio by default), the ratio has to be adapted afterwards. And this is a really strange input range. In case, the javascript function that does the ttrick :

function main()
{
    var x = arguments[0] // Size in px of input stream
    var y = arguments[1] // Size in px of output stream

    // The output goes in "aspect mod" of 3D Rect projector

    if (x < y){
        return -100 * ( y/x ) +100
    }else if (x > y) {
        return 100 * ( x/y ) -100
    }else {
        return 0
    }
}   

- For the Z values, I brute forced it also, with something like that (It's inside a user actor, which I have 33 instances of, so each get his own little z value)

For the brute forcing itself, I did this : (this is quite a mess, but basically I generate two background colors with one pixel of difference, I send it to the corresponding virtual stage. Then the virtual stage is grabbed, and displayed on the "preview stage" with a normal projector. This way I can see clearly when the yellow border touches the blue border, and using a midi controller (with a few knob, each doing a different factor 10 increment) I managed to achieve excellent precision quite smoothly.

- So, quite a mess I have to say. A "match stage target size" actor that would have been nice, clearly. But each of the 33 virtual stage is re-projected on the main stage (called Real VP here) using a similar mechanism and a 3D Quad Distort projector to match then angle of the scenery (the scenery is 33 actual frames) so at the end I was able to re-use 80% of my calibration mechanism" for this other purpose, so it was less of a waste of time than I originally belived.

On the bright side, the result is pretty cool ...