Second, I disagree that the Moorean shift is at work. As you point out, a paradigmatic Moorean shift would require denying that God exists, when all that is needed to vitiate the modal argument is accepting the possibility that God does not exist.

Third, I disagree that the distinction between epistemic possibility and “real” (metaphysical) possibility is, as seems to be implied, a trivial one. Agnostics think that there might be a God and might not be, just as they might think that Goldbach’s conjecture might be true and might be false. If they think about it, they realize that they don’t think that Goldbach’s conjecture metaphysically might be true and also metaphysically might be false; surely Goldbach’s conjecture is either a metaphysically necessary truth or a metaphysically necessary falsehood, so it is only metaphysically possibly true or metaphysically possibly false—not both. Similarly for a God characterized as metaphysically necessarily existent. Either it metaphysically must exist or it metaphysically cannot exist, just as Goldbach’s conjecture metaphysically must be true or metaphysically must be false. I find this one of the most pernicious things about arguments like Plantinga’s “Victorious” argument: It invites the reader to overlook the difference between epistemic possibility and metaphysical possibility and to think, “Gee, surely God’s existence is *possible*”—when the same person would surely also think, “Gee, surely it’s *possible* that God doesn’t exist.” And yet, although he can have both in the epistemic sense of possibility, he can’t have both in the metaphysical sense of possibility—he has to pick one or the other, or remain neutral.

(Addendum: You might think you’d get your “Moorean shift” this way:

(1) Possibly, God exists
(2) The existence or nonexistence of God is a matter of necessity
(3) Therefore, God necessarily exists.

(1′) Possibly, God does not exist (i.e., it is not the case that God necessarily exists)
(2′) The existence or nonexistence of God is a matter of necessity
(3′) Therefore, God necessarily does not exist (i.e., it is not the case that possibly, God exists)

Since we’re leaving (2)=(2′) alone, as simply a matter of definition, the shift is then from (1)—>(3) to not-(3)—>not-(1).

But this is not a true Moorean shift. The nontheist does not state (1′) as a negation of (3)—he does not want to change modus ponens to modus tollens—but, rather, he notices the complementary possibilities “(Metaphysically) possibly, God exists,” and “(Metaphysically) possibly, God doesn’t exist,” and sees no reason to simply ignore the latter of the two—especially when it turns out to be logically contradictory to the former of the two. If one is going to say, “(Metaphysically) possibly, Goldbach’s conjecture is true,” and then derive that Goldbach’s conjecture really is true (as can be done from that assumption), he must first have some reason for denying the complementary option, “(Metaphysically) possibly, Goldbach’s conjecture is false”—from which, after all, could be derived the falsity of Goldbach’s conjecture.)

Of course, a Moorean shift is available, but this is also true of EVERY deductive argument. So we argue:

(1) Possibly, God exists
(2) The existence or non-existence of God is a matter of necessity
:. (3) God exists

And the (former) agnostic turns around and argues:

(1) God does not exist
(2) The existence or non-existence of God is a matter of necessity
:. (3) It is not the case that possibly God exists

(The first premise could also be “possibly, God does not exist”, but that wouldn’t be a paradigm case of a Moorean shift like this argument is.)

Again, you talk about there being an inconsistent set and having to choose which one to reject, but this is just the definition of deductive validity.

So this problem afflicts every deductive argument. Perhaps, though, you think it fails the “epistemic priority” criterion? If so, please explain.