is the 'server' node the element that contains server instances?

Yes; that's where you access the actual running server.

The operations to actually start/stop/restart the server and check their
status are under "server-config" which I'm not happy about. Not
intuitive. It's that way because of a kind of chicken-and-egg problem.

Emanuel, WDYT about a more sophisticated ProxyController for the
servers? Installed by the ServerAddHandler. It handles the lifecycle and
status ops directly; proxies the rest. Where it gets complicated is
things like read-resource, read-resource-description, read-attribute,
all of which would require custom handling. It also make a server
resource an odd duck, since it's resource description varies depending
on whether the server is started.

Yes, we could think of doing something like that - maybe we should just
create a ManagedServer for all configured servers and make it the proxy.
It should be aware of the lifecyle/state of the server anyway and also
holds the associated managed connection.

Yes, that sounds like the right way to implement it.

Heiko/Ike/Emanuel/anyone – any concerns about the "resource description
varies" issue?

What I mean is, for all other resources, read-resource-description will
return fixed content. For host=x,server=y it will vary depending on
whether the server is actually started. If not running, it will just
provide the operation to start the server and a "status" attribute,
which will tell you it's stopped. Once you start the server, then
read-resource-description will provide that, plus ops to stop and
restart the server, plus all the normal attributes and operations
exposed by the root resource of a server.