Let’s say c_i is the actual time at which block i is found, and t_i is its timestamp.
If the attacker (who is assumed to have 100% hashrate) follows this strategy:
- The first block in every period p uses as timestamp the previous block’s timestamp minus the grace period: t_{2016p} = t_{2016p-1} - G (the minimum legal time according to the proposed rule).
- The 2014 middle blocks use the minimum legal time, which doesn’t really matter as long as it’s low enough: t_{2016p+k} = 0 for 0 < k < 2015.
- The final block in the period uses the current time t_{2016p+2015} = c_{2016p+2015}.
Then the observed duration of the 2015 blocks in period p (relevant for difficulty adjustment) is t_{2016p+2015} - t_{2016p} = c_{2016p+2015} - c_{2016p-1} + G, i.e. the time it took to mine 2016 blocks plus G.
The difficulty multiplier m will thus be m = \frac{P}{X + G}, where X is Erlang distributed with k=2016 and \lambda=\operatorname{hashrate} \cdot \frac{\operatorname{target}}{2^{256}}.
In a simulation with G=600 I get an effective average block interval under constant hashrate of 599.9997… seconds. With G=7200 I get 596.7211… seconds.