The container size is irrelevant to this, and so are all of the other specifics of the question for that matter. If your passwords are brute-forceable then all bets are off.

Storing a bit array of $6n\pm1$ would take only 170 MB and lookup would be faster, but it's still pointless.

The answer to "is it safe" is yes, but if you're implicitly asking if it's a good solution to the problem of the message being larger than $2^{127}-1$, the answer is no. It's more efficient to divide the message into pieces, make shares of each piece, and concatenate them.

I understand the first paragraph of this answer, but I can't make sense of any part of the second. There is no black box in the halting problem, so I don't see how it's relevant, and that goes double for "hardware/environmental data dependent interrupts" etc.

AES-GCM-SIV isn't strongly nonce-misuse resistant according to this.

@JAAAY I'm not sure I understand either question. I think they're based on a false premise, since they both ask about 3-argument operations ($a+b+c$), but djb was talking about 3-operand instructions where one operand is an output ($a\leftarrow b+c$). x86 has $a\leftarrow b+c$, but for xor it only has $a\leftarrow a\oplus b$ where $a$ appears on both sides.