I expected it's possible to use i.e. Guava-19 in myModuleA and guava-20 in myModuleB, since jigsaw modules have their own classpath.
Let's say myModuleA uses Iterators.emptyIterator(); - which is removed in guava-20 and myModuleB uses the new static method FluentIterable.of(); - which wasn't available in guava-19. Unfortunately, my test is negative. At compile-time, it looks fine. In contrast to runtime the result is a NoSuchMethodError. Means that, the class which was the first on the classloader decides which one fails.
The encapsulation with the underlying coupling? I found a reason for myself. It couldn't be supported because of transitive dependencies would have the same problem as before. If a guava class which has version conflicts occurred in the signature in ModuleA and ModuleB depends on it. Which class should be used?
But why all over the internet we can read "jigsaw - the module system stops the classpath hell"? We have now multiple smaller "similar-to-classpaths" with the same problems. It's more an uncertainty than a question.
Version Conflicts
First a correction: You say that modules have their own class path, which is not correct. The application's class path remains as it is. Parallel to it the module path was introduced but it essentially works in the same way. Particularly, all application classes are loaded by the same class loader (by default at least).
That there is only a single class loader for all application classes also explains why there can't be two versions of the same class: The entire class loading infrastructure is built on the assumption that a fully qualified class name suffices to identify a class with a class loader.
This also opens the path to the solution for multiple versions. Like before you can achieve that by using different class loaders. The module system native way to do that would be to create additional layers (each layer has its own loader).
Module Hell?
So does the module system replace class path hell with module hell? Well, multiple versions of the same library are still not possible without creating new class loaders, so this fundamental problem remains.
On the other hand, now you at least get an error at compile or launch due to split packages. This prevents the program from subtly misbehaving, which is not that bad, either.