So yes a 14 bit ADC can adequately represent a 14 stop photocell, and the accuracy is going to be limited by the photocell's noise floor (unless it is actually more than 14 stops, measured in dB from the noise floor to clipping. HDR raises the point of clipping but can't do anything about the noise floor which is often based on physical limits like Brownian motion etc.
I know what dithering is when it pertains to display. I've taken enough computer graphics courses to know the basics of dithering an image.
Where I'm drawing a complete blank is where you suggest that the sensor is somehow able to dither the incoming image. I haven't designed a semiconductor more complex than a single fet or npn, but I honestly can't imagine how you'd dither an image at capture time. Moreover I can't imagine why you'd do it at that point, it seems like the digic would be the place to do it.