I don't understand this business about being part of the probability space. Let P_A(\vec{a}, \alpha, \lambda) be the probability that Alice will measure spin-up for her particle, given that she measures spin along axis \vec{a}, and that \alpha represents other details of Alice's detector (above and beyond orientation), and \lambda represents details about the production of the twin pair. Similarly, let P_B(\vec{b}, \beta, \lambda) be the probability that Bob will measure spin-up for his particle, given that he measures along axis \vec{b}, and that \beta represents additional details about Bob's detector. By assuming that the probabilities depend on these particular parameters, where have I made an assumption about the existence of a single joint probability space? What does "contextuality" mean, other than that the outcome might depend both on facts about the particle and facts about the device? The only assumption, it seems to me, is locality, that P_A doesn't depend on \vec{b} and P_B doesn't depend on \vec{b}.
But the predictions of QM for EPR is perfect anti-correlation. Which means that:
If Alice measures spin-up at angle \vec{a}, then Bob will measure spin-down at angle \vec{a}. That seems to me to mean that the probabilities must be 0 or 1:
If P_A(\vec{a}, \alpha, \lambda) is nonzero, then that means that Alice has a chance of measuring spin-up. But if Alice measures spin-up, then Bob has no chance of measuring spin-up at that angle. So Bob's probabilities must be zero whenever Alice's are nonzero, and vice-versa. That's only possible if the probabilities are all zero or one. That means that the outcome is actually deterministic, given \lambda, which in turn implies that the details \alpha and \beta don't matter.
I don't think that the non-contextuality is an assumption, I think it follows from the perfect anti-correlations.