On the other other, there is a movement to remove such late binding/rebinding features from the Java platform. However, I think JVMTI is expected to remain supported if the agents are loaded ahead of time. Only on-demand loading of agents will be removed from OpenJDK.
That way AOT compilation becomes possible/efficient and dependencies can't hack into other packages in an unmaintainable way without explicit permission from the user, so that updating the JVM and dependencies will be even more streamlined.
This doesn't concern features that have been traditionally abused (such as reflection on core OpenJDK classes), but also harmless (from an encapsulation perspective) uses of JNI, and access to future features such as FFI/Panama. Justification for restricting those as well is not so much OpenJDK updates, I think, but that it could cause crashes that might be blamed on OpenJDK.
Java does have "agents", which enable arbitrary control of bytecode at class loading time -- bytecode in existing .class files can be transformed, or completely new bytecode generated on-the-fly. But generating bytecode just to replace a class with a test double would be insane.
From another comment I learned there's also a separate "Tool Interface" that can do this bytecode-replacement job, but again, it's a huge amount of tricksiness and bother for something more easily and clearly accomplished in source code with a boring old anonymous subclass. (Though it won't work if the class is sealed, or final, or your code uses reflection to check the class name, or...)
TIL Go has monkey-patching.
But since it has monkey-patching, how is it statically typed? Or does Go monkey-patching amount to creating an instance of a defined-on-the-fly subclass?
The latter would be interesting, because Java lets you do this too -- conveniently "new up" an instance of a subclass that you define inline of a given class or interface (this used to be the easiest way to define callbacks, before lambdas came along), so this testing strategy is available to Java too, but seems not to be preferred.
Separate question: IIUC, monkey-patching is convenient if your test code directly needs to create a TestX instead of a real X, but what if you need the test double "deeper down" -- that is, you need an X that creates a Y that creates a TestZ instead of a real Z?