People concerned about "weird bokeh" need not be. Just think about a (non-Xtrans) bayer array, there's only red and blue in alternate lines. Does the bokeh turn out with weird colours? No, it's all interpolated and corrected in software.
The main difference is everything that has gone before (be it monochrome sensors, bayer, x-trans, foveon) has had each and every photosite capturing the same angle of view. This dual pixel AF sensor is different. Each photosite (sub pixel) is only able to see half of the phase.
Imagine for a moment that you could capture an image independently from the left and the right hand phase of this dual pixel sensor (which Canon will obviously not let you do natively). If you've got everything in focus, the two halves will look identical, but if the something is out of focus, it will appear in a different place on the frame in the two images (think of a split image focusing screen).
Instead of a full sized square photodiode behind the microlens, there are two photodiodes - according to the canon marketing material, they are rectangular in shape, together forming a square of the size/shape of a normal photodiode, with a vertical division. This will allow one side to see one phase, and the other side, the other phase. Theoretically if you have a perfect point source of light (bright, about the size of one pixel should it be in focus) out of focus, together the two phases will see a perfect circle of blur. But on its own, the one side would theoretically see just a semi-circle of blur, with a vertical cut off. The other side will see the other half.
This is not going on over the width of just one sub pixel, but could in extreme cases of blur (think 85L wide open at close focusing distances) cover half the frame. Unlike with a bayer sensor and a simple AA filter to blur out neighbouring pixels, there is no way you can re-assemble this data meaningfully if each half is exposed differently, and one half is blown out.
The only way I can picture it working is to do something along the lines of staggering it - one dual pixel having left underexposed, right overexposed, and then the next dual pixel being the opposite way around - left overexposed, right underexposed. That way, some post processing could look at them in pairs of left phase, and check first of all to see if the overexposed channel has blown out (if so, use the neighbours underexposed left phase value). And secondly, do the same for the underexposed channel if it gets down to the noise floor. Without doing that, bokeh would look odd.