I just ran it and it works fine.
If the problem you're having is simply because you're expecting to see the exact same numbers in the SRT fields then it won't happen because of the different ways XSI and TG2 work with coordinates. That's partly the point of the script.
My pipeline usually involves rendering a LWO mesh of the landscape from a viewpoint within TG2, using the Render Sequence/Output>Micro Exporter. Low rez settings are fine for general reference.
Import that LWO into XSI. The mesh then acts as a properly translated reference point for XSI which you know is positioned, scaled and rotated correctly through translation. What you then do in XSI will translate correctly across to TG2.
Within XSI, child the entire tracking scene under a new model and rotate, translate and scale the track to suit the TG2 mesh. If the camera has an up-vector constraint I initially bake the camera animation to get rid of any possible SRT dependancies which may put the camera roll off kilter.
Select the camera and run xsi2chan6.js making sure the scale is 1:1
Seriously, I just tried it to check the JScript is okay and it worked just as expected. I've used it in production. The key thing I've always found is to not try to fit TG2 to your XSI scene, but do it the other way around. Parenting the tracked scene under a new model means that the camera's relative position to other parts of the track is not changed no matter how you move the model around.
I also avoid exporting a heightfield as a .ter under the Right-Click option, as importing it into XSI usually means you have to do the SRT adjustments yourself and it can go badly wrong. An exported mesh using the Render node means the exported mesh for reference is bang on the money, even though it comes in looking like a LIDAR scan.