The agent-class column will show if there are discrepancies between
what a hostmanager reports and the local scf that ocsbow sees.  Also
tolerates broader range of mismatch.

This just factors out each docker inspect call so failures can be
trapped more easily.  Behavior should be altered.

Instead, the 'prot' elements for docker-managed things are long-lived
and will be immediately updated when we learn new service info.

The logic handling the agent list and reconciling it against docker
inspection results is rewritten (and probably improved).

Docstring in site_config explains them all.