This is sort of a complicated scenario, but I've been able to create a relatively simple test to recreate this. See the attachment for the project. Note that to get this to run using the 3.0.1 Felix framework version, I needed to create a private build of the pax exam trunk with a dependency change to also pick up the latest pax runner snapshot version.
This problem showed up for us with the 3.0.1 release, but also appears to be in the 3.1.0-SNAPSHOT release. The 2.0.5 release does not have the problem and equinox also appears to be handling this correctly. In this scenario, we have a bundle with two embedded jar files that also has a Dynamic-Import: * header and a Require-Bundle header for another bundle. One of the embedded jars uses the same package name as packages defined in the bundle referenced by the Require-Bundle, creating a split package situation. The test is fairly simple, consisting of just a couple of lines:
// if this line is commented out, then the second loadClass() call will succeed. If
// this load occurs first, then a NoClassDefFound error results on the load for TestClass.
Class cls1 = bundle1.loadClass("org.apache.geronimo.harness.EETestDummy");
Class cls = bundle1.loadClass("org.apache.geronimo.embedded.TestClass");
// this will force the fields classes to be resolved.
When this is run, the cls.getDeclaredFields() call will get a NoClassDefFound exception for the class Lorg/apache/geronimo/harness/EETest
That class is defined in one of the embedded jars. If the load of the EETestDummy class is commented out, then the getDeclaredFields() call does not throw the exception. The resolution of the internal EETest class appears to depend on whether the imported package gets resolved first.