It'd be nice to have a runner-only variant. But we don't have it.
There where a few attempts at it in the past, but nothing that was worth keeping. I don't know if you're familiar with the source code somewhat, but basically you just need to make a product without the 'org.csstudio.opibuilder.editor' plugin in it. At that point you probably also don't need the 'workspace' concept. But once you start taking that away, it becomes somewhat challenging to keep all the existing features (such as restoring windows on restart). In the absence of a workspace, the runner could then fetch display resources directly from the server. I don't see this landing soon though.
A parallel strategy that we've been exploring as time allows, is to implement web-based support for OPI displays (just upload the file to the Yamcs Web UI under 'Telemetry > Displays'). AS-IS, this is quite beta so I would not recommend using it with end-users unless you can stick to simple displays (not all bells and whistles supported).
For the final question: there's no absolute numbers. Being built on top of Netty, the HTTP server is about as efficient as we can make it. You can do the test, for example by using our Python client package and just make a huge amount of client connections, each doing a parameter subscription. You could also consider ensuring a predefined minimum target, rather than establishing a maximum. Easier to verify.