possible (page, offset) and then choosing top 4 offsets for all
content bytes.
i.e. this should do better than the previous "greedy" approach in
which we picked a single high-priority offset to resolve, and then
tried to find 3 more than didn't introduce too much additional error.
Major outstanding issue is that we need to also evaluate whether it's
worth storing 1, 2 or 3 offsets as well, since sometimes this will
have lower total error.