Footnote 4 says a two-hour grace period instead of ten minutes "only reduces the expected block interval time under attack by ~2.2 seconds", and points to the grace period summary post. That post gives one block every 599.9997 seconds with a ten-minute grace and 596.7211 seconds with two hours, from Pieter Wuille's simulation. The difference is 3.28 seconds. The post's ~0.1% and ~0.65% block rate figures also come to about 3.3 seconds.
I got the same result independently. A simulation of the attack using Bitcoin Core's retarget arithmetic gives 599.71 and 596.43 seconds (3.27 s), and the closed form (7200 - 600) / 2016 gives 3.27 s.
This changes ~2.2 to ~3.3. The rationale is unaffected: the two-hour grace still only trivially worsens the block rate under attack.
