As you can see from the above images, there is no location on the Scheiber sphere that corresponds to a 4-channel sum. A diagonal split, though, is simply the path between Left Front and Right Back, and another path between Right Front and Left Back. Since the Shadow Vector is agnostic about encode location, they play back with full separation.
I should also note the standard CBS 4 -> 2 encoder has a deviation from ideal as the signal is panned across the sides of the room; I think the pan-locus bends towards the front for the Left Front -> Left Back path, and towards the rear for the Right Front to Right Back path. It was possible to skew the logic control to offset this, but I had to draw the line somewhere ... besides, there was no way to tell which of several possible encodes that CBS was using ... the standard 4 -> 2 unit, the one with selectable front or rear bias, or the Position Encoder.
The biggest advance with an all-digital decoder is, as mentioned by our intrepid Malcolm Lear, is getting rid of annoying transient clicks caused by the finite delay of the logic circuit, which in the analog version is always slightly behind the audio signal by a few milliseconds. The clicks are subtle, appearing only in the suppressed channels with program transients, but better to sync up the audio and control systems. The implementation of logic sensing is also much better, since you don't have to wrestle with fast-acting AGC circuits in the analog domain, and you can do little tricks like velocity sensing (program transients control the speed of the logic circuit).
P.S. Funny side note: these images originally appeared to me as animations as a I came down from 150 micrograms of LSD. They were so vivid (imagine them in dark space with the light chasing around the sphere, and all the vectors trying to hide from the light) that I had to write them down immediately. A few weeks later I ground through some of the math, realized it was indeed possible to built the thing, and spent about a month down at the Los Angeles Patent Library. They told me about filing a Document of Disclosure, which formed the basis of the patent, complete with these drawings, which summarize the action of the Shadow Vector.
P.P.S. Once the prototype was up and running, and being demonstrated at the Consumer Electronics Show in 1974, the CBS Labs design team came over and visited the Audionics suite. I was surprised to find they didn't use Scheiber notation, which would make describing a dynamic decoder incredibly complex and opaque. I was even more surprised that they actually hired an outside mathematician to describe all the weird vectors used in the CBS Paramatrix, while it's intuitively obvious how the Shadow Vector works, thanks to the Scheiber notation. I still don't understand the complex vector math of the TATE decoder designed by Wesley Ruggles, but reading between the lines, it, like the Paramatrix, appeared to be optimized for the six cardinal points of SQ, while the Shadow Vector is agnostic about the encoding localization and linear in its action ... which I describe as "symmetric" in the patent.
It's not clear to me how Dolby Pro-Logic II (music mode) and DTS Neo:6 decoders work. They don't sound as symmetric as the Shadow Vector, nor as quick, and are much less spacious. DPL-II in particular sounds very shut-in and pretty slow compared to what I expect from good dynamic decoding, and DTS Neo:6, while sounding more open and pretty fast, can sound kind of spitty and harsh on some material. Neither sounds anything like Shadow Vector, which is kind of disappointing, considering the original insight of those moving lights goes back to 1973 or so.