Hi - sorry I was absent here for so long.
The plug-in is Voxengo's Elephant - the VST 2 version verses the VST 3 version.
I will try it again to see what happens, but the first test definitely yielded a quantifiable difference.

That may be the reason. Elephant has an analog quality to it that might be a little random. I'll bet that if you do the same test with VST 3 vs VST 3 it will also not null. 

I will try it.

Okay - null test with two VST 3 files also yielded a difference, HOWEVER . . .
A visual check on VST 2 vs VST 3 shows a slight perceivable difference, but no visual difference between the two VST3's . . . the mystery deepens . . .

Can you load the same preset into each one?

I'm with Alex on this one.  If you can get just a single pair of VST2/VST3 plugs to null then it is definitely not an issue with Sonar.
 
As others have said, the VST "standard" is interfacing only.  The audio processing should be the same across the board.  If it's not, then it's the plugin writers fault/problem,  not Sonars.

Pdarg well done.  Thanks for putting in the time to get to the bottom on this. 

Have you run this by Alexsey yet? The guy's seriously the DSP guru to the DSP gurus. I'll bet he'll have an answer.

One obvious possibility is that the VST3 version is built from newer code and has some bug fixes that the VST2 version doesn't have :)

I wonder how far is CW with VST3 implementation in X4...