I think you should, in fact, use the slightly-smaller value. That's not because I've measured, but because I can resolve the apparent contradiction from exiftool: it's showing you a rounded value. Try giving it the -n flag, to disable what exiftool calls "print conversion":

 $ exiftool -n -ScaleFactor35efl sample.jpg
 Scale Factor To 35 mm Equivalent: 0.975938688184433

Which, of course, is ridiculous over precision, but there's the raw value. I agree with Philip's comment that the difference probably has little practical effect, but it seems likely that if you're going for "most correct", the metadata is right.

I suspect that the small difference is due to the edge of the frame being masked out and used for setting black level and similar. (Some of it is for demosaicing edge pixels, but 2% seems excessive for that.)

And I can confirm that this lines up with Canon's specs, too: the spec sheet lists

Effective Pixels Approx 21.1M
Total Pixels     Approx 22.0M

and, a-ha:

22 × 0.975938688184433² ≈ 20.95403911

which is pretty close to the effective number — close enough that I'm going to call the remaining difference rounding error. (This works out to a strip of about 100 pixels around each edge, by the way.)