Hmm, I went back and looked, and I was wrong. The bug is
not symmetric. It is still unimportant because it only represents a fraction of 1% of the interval.
Basically, every 2016 blocks, we look at the previous 2015 blocks for the difficulty adjustment. This is off by one, because we had intended to look at the previous 2016 blocks.
I was confused about the symmetry part because this was mostly discussed in threads about time-warp attacks. Some alt-coins allow difficulty to be skewed more easily in one direction than the other, so Art Forz showed that he could game them by cranking difficulty way up, or way down. Bitcoin's difficulty algorithm is perfectly symmetric*, so the best we can do is tweak the final block's timestamp (within the very liberal timestamp rules) to shift a few hours of actual time into or out of the apparent time of the current period.
Say you are a miner that wanted a lower difficulty in the next period. You can make that happen if you find the appropriate block by setting the timestamp in that block as far forward as the rest of the network will accept, about 3 hours. This makes the next difficulty about 1% lower than it would have been if you'd put the actual time into that block. If not for the bug, the next retarget period would start at that timestamp, so the apparent interval of that period would be 3 hours shorter, which would bump the difficulty up by about 1%. You could again manipulate the current timestamp, but the best you can do is 3 more hours, which compensates for the 3 hours borrowed before. So, the worst that can happen is that if a lot of miners get together, they can keep the difficulty about 1% too low for a long time. They can't ever get a second 1%.
But, because of the bug, the timestamp of that block is never actually checked. The period is calculated based on the timestamp of the
following block. That block can be as early as 1 second later than the
median timestamp of the 11 blocks before it. This slight asymmetry means that a large group of miners
could repeatedly keep the difficulty a fraction of a percent lower than they could with the liberal timestamp rules alone. It also means that borrowed time doesn't necessarily need to be paid back, if they can get a 51% cabal working. Not like there aren't worse things they could do at that point...
See
this thread for some discussion on both issues.
*
Well, not perfectly symmetric, because of the bug. But we use 2016 blocks at the retarget interval, come hell or high water. Art had ASICs several years ago, so he would "test" alt-coins by throwing a lot of hashing power at them, then leaving. If he represented 90% of the hashing power on some alt coin, it could take months for them to retarget back down to their normal speeds. Some of them responded by adding code to allow difficulty to drop faster if hashing power left. This is the sort of asymmetry that really matters, and Art showed them that it opened them up to serious manipulation. 
