Dr. Manhattan's link worked for me, both at home and at the office, on two different machines.

Reading section 3, right column of page 3, bottom of the page, it seems that the stress is applied at the time the behavior is learned. I don't know how long these things linger, so maybe it's still present at the time the learned behavior is performed, since it's performed just after it's learned.

They acknowledge this ambiguity at section 3.1, left side of page 5. They make it clear that the effect (where stress slows the rate at which unrewarded behavior extinguishes) is still present if the stress is applied directly before performing the behavior, after the behavior is already learned. The effect is stronger if stress is present at both the time of learning and the time of performance, though.