Simulate retry granularity

Everything follows from one fact: a retry unit of $n$ pages succeeds only if all $n$ pages do, with probability $(1-p)^n$, and every attempt pays for all $n$ pages. The attempts until success are geometric, so a unit costs $n/(1-p)^n$ parses on average, and page-level retry is the case $n = 1$. The simulation's job is to confirm that number by brute force; the interesting output is where each curve crosses "twice the failure-free cost".