Instead of keeping the last N entries, keep the last N entries for each
channel. This ensures that lots of audit records don't age out the
cluster records (or vice versa).
The overall approach does not change, and that approach is overall pretty
lame. We're still rewriting a big summary blob on every log commit. We
still should refactor this later. This solves the immediate problem at a
small cost of increasing the log summary structure by ~2x.