Add a lower bound for TTL on well known results.

It costs both us and the remote server for us to fetch the well known
for every single request we send, so we add a minimum cache period. This
is set to 5m so that we still honour the basic premise of "refetch
frequently".