Oracle provide an OSX JVM and will do so for the forseeable future, it's also open source, so the community could carry on if they stopped. The primary problem with the Oracle JVM is lack of retina support for Swing, but if you'd write a Cocoa UI yourself then it doesn't matter of course as Java won't handle any GUI stuff. Retina support for JavaFX2 (the current-gen gui toolkit) is available in Java 8 so it's definitely being actively developed, it's not abandoned or anything.
So the question then becomes, which is better:
a) Take bitcoinj completely out of the Java world via native compilation or transpilation to C++
b) Embed the JVM and link the two worlds together?
(b) is much less ambitious, especially if you're OK with writing a bit of Java code to keep the interface thin. Basically the Java side calls into your app when interesting user-visible things happen, like new transactions appearing, then your GUI can call into the java side to send money. There are auto-translators that make the glue work easy, like
https://code.google.com/p/javacpp/. You probably wouldn't want to expose the entire bitcoinj API that way because it's very large, but the code needed to bring up a wallet app is very small. I knocked one up this weekend in about one evenings worth of coding, completed with nice animations. The interfaces you'd need are basically some Objective-C++ methods that receive information from the Bitcoin side, like the balance having changed, a list of transactions, etc, and then a callback into the Java side to send money. If you look at the javacpp site you can see example code for making calls both ways.
If I were in your shoes, I'd go for (b) because it is the most well trodden path and will let you achieve the best user visible results quickly. The JVM can be bundled with your app and stripped down if you're worried about download size.
If it's unclear how the code would look, let me know and I'll try and knock up a really simple prototype.
There's also (a). I'm investigating transpilation for a few reasons, one of which is to do with a private project. I'm working with the author of j2c:
https://code.google.com/a/eclipselabs.org/p/j2c/. It's a rather sophisticated transpiler that converts Java to clean, readable C++11 that looks much like code a human would write. It's complete enough to transpile the entire standard Java class library, including all the GUI toolkits and other things - so, pretty amazing piece of code. However it's incomplete because where the Java code calls native methods (that would be provided by the JVM) it just spits out stubs you're expected to fill out yourself, for starting threads and so on. As there's no JVM it's just like using a C++ library that is missing a "portability layer".
I'm working on this myself and don't really need much help at the moment, I'm just making steady progress towards getting something up and running. I can let you know once I reach some interesting milestones. The point of this is that whilst you don't need access to most of the API to write a wallet app, I'd like to make every kind of app easy from C++, not just GUI wallets. Then the compile-to-C++ approach is much more appealing, even though it's also more work.