1) reading from /dev/urandom takes more than 10ms causing an entropy error
2) a malloc fails causing the error to be saved to the fallback ERR_STATE instead of the thread local ERR_STATE
3) java sees failing code and tries to read the last error but now malloc succeeds so it reads it from its local thread ERR_STATE which succeeds or some other thread has clobbered the fallback ERR_STATE
4) java doesn't throw an exception because it sees no error
the other possibility is the openssl random state is shared between processes due to forking. i hear there is a process called the zygote that warms up the vm and forks to create apps. if it initialises openssl then it is possible child processes could get the same random state.
...I'm the packager of SpongyCastle (a task a well-trained monkey could do, I don't alter the code other than package-renaming it), and I confess some relief that the issue is with Android's in-built SecureRandom class, rather than anything in (Bouncy|Spongy)Castle.
But one thing I'm wondering is: would any setSeed on any SecureRandom instance cause a potential problem? Or just on the same instance being used for signing? In other words, do I only need to check that spongycastle -- which I presume is a fork of bouncycastle -- handles things reasonably? Or could any other messing about with SecureRandom mess it up?