Sorry Stuart, we only just noticed your response, thanks to the publicity around http://ashtonkemerling.com/blog/2016/06/11/my-increasing-frustration-with-clojure/
I'm going to try and explain our use-case a bit for context, but please understand that this issue was simply a question about what appears to be inconsistent behaviour from a function that looks and smells a lot like a public API function (who would have thought 'eval' would be private
Our company uses Clojure to build a cloud platform that does computation in response to user requests using modules loaded from a database. The modules are trusted code, but they are independent from our main platform so there might be multiple versions of the same module running at the same time (we want to avoid namespace collisions). We've looked at lots of approaches to keep modules from interfering with each other including containers and micro services running separate JVMs, but in order to have acceptable response times for simple queries like "is this input valid?" we want to run simple queries in the same JVM as the web server.
The general approach we use to answer a query from a user is to build a namespace for our computation (which might require loading other namespaces from the computation module), then eval the expression in the context we've built. We have a LRU cache for module namespaces but we still end up with a lot of metaspace churn for the eval, which we mitigate by using a clojure interpreter to handle simple queries (eval is too slow).
When we implemented our LRU modules namespace cache we wanted to experiment with loading module namespaces into their own class loader to help track class lifecycle and GC, and hopefully to allow multiple namespace versions to coexist. We've since concluded that this is impractical because clojure has so many global lookups related to namespaces, so now we preprocess module namespaces and perform name mangling on load, and explicitly unregister loaded namespaces when the cache expires so that their classes can be collected. We avoid using language features like multimethods and protocols that use globals in modules.
Once again, we're not looking for the Clojure team to implement containers for us (though that would certainly be a nice feature to have!), this was simply an inconsistency we noticed between API and implementation. What is the expected entry point for Java-interop eval?