Subaru did once specify the 1/4" max difference in rolling circumference, but the general view here is that they have now backed away from that. The original spec did indeed work out to somewhere between 1/32" and 2/32" difference in tread depth, and that was an awfully tight spec. It was probably written by the lawyers.
Quite coincidentally the spec seems to have gone away, and at about the time when Subaru introduced the donut spare, which is a major violator of the 1/4" difference! Subaru is now counseling moderation. Keep them close, if they aren't close it's OK to temporarily drive gently, and so forth.
Yes, the math says that the difference in tread depth is dependent on the difference in circumferences, and it does not vary by actual circumference. So 16", 17", 18", rolling circumference, or measured circumference—it's all the same.
There's a counter-intuitive puzzler that engineers and the like sometimes throw at each other. You're given a very long length of magic string, which has the ability to float in the air when its two ends are tied together. Its length is supposed to be the exact circumference of the earth, along a specified route (plus just a little bit more to account for the knot). You're also given a pair of magic shoes that let you walk along that entire route, never deviating from it, across oceans, through buildings, and over mountains, paying out the string as you go. However, it turns out that the string wasn't measured quite right and is about six feet too long. "It won't matter," you think. "Stretched out over about 25,000 miles the difference will just get lost." Then you get back to your starting point, tie the two ends together, and the string snaps tight, assuming its ability to float in the air—about a foot above the earth, all the way around! Moral of the story: Radius still equals Circumference/(2*pi), even for large values of circumference.