I'd much rather see Canon bake the continuous gyroscope data into the EXIF tags and leave the processing to computers that actually have the CPU power to handle it. Ideally, they'd downsample the video footage and compress it at somewhere around 1280p instead of 1080p, leaving a nice margin so that when you process the footage, you'd still have 1080p, and they would show you the center portion of that footage so that in the absence of massive correction, what you see would be the final, cropped output.
That scheme would also have the advantage of being able to correct more robustly by virtue of being able to look at the entire set of data rather than trying to guess whether that was a shake or a deliberate pan. And when compensating for a particular shake would cause a black border, if you're working with the data all at once, you can retroactively adjust the center of previous frames to avoid the problem, or at least smooth it out in ways that you cannot accomplish in real time.
And just to clarify, lens IS does do this, but it isn't really designed for video; it is designed for holding the image dead still. For video, it is suboptimal, because that usually isn't what you want. You want motion to be smoothed out, not eliminated. Otherwise, when you hit the limit of its range, you get a nasty jerk in the picture.
As for "no downsampling", you're downsampling anyway. Just about nobody shoots RAW video. And with the 16:9 aspect ratio, you're throwing away pixels on the top and bottom no matter what, even after downsampling, because you don't use the full height of the sensor. Those pixels could be made available for post-processing... essentially for free, without even needing to crop, though you would probably want to crop a little bit to accommodate horizontal shake. And because vertical shake usually has the highest amplitude, that's a real win.