2016 Postmortem

Furthermore, the sample in an exit poll (or any poll) is NEVER perfectly random. In fact, it isn't even close.

Exit pollsters do not sample voters at every precinct. Instead, they pick precincts in advance (trying to get a representative subset of precincts that are representative of the state as a whole). But this is just guesswork. Precincts that look representative based off of previous voting data might be extremely unrepresentative of the electorate in the current election. For example, in past elections, voting was not as polarized by education levels as it was in this election. So a good subset of precincts in the past might be skewed towards lower or higher education levels voters, in a way that didn't matter in 2012 but did matter in 2016. And this is before we start talking about the incredibly low response rate (of all polls, including exit polls), which significantly increases the probability of differential non-response bias.

You are taking a supposed anamoly in the data (that actually has very credible benign explanations), saying that it would be statistically improbable under perfect assumptions that don't exist in reality, and then immediately jumping to the conclusion that systemic fraud is the explanation. But you should apply the same rigor to the alternative explanation (fraud) as you are applying to analyzing the anamoly.

Due to the decentralized nature of our elections, a systemic conspiracy would require tens of thousands of people to succeed at flipping votes in some way, and not get caught. Let's assume for the sake of argument that tens of thousands of people DID attempt to pull off a systemic conspiracy. Each one of them has some probability of getting caught. Let's further assume for the sake of argument that the probability of getting caught is small, say 1%. What are the odds that (say) 10000 people succeed in not getting caught with a 1% detection rate? Approximately 1 over 10 to the 44th power.