I think you're making it too difficult. The way we use to do it before this became too difficult was stand outside polling places and ask where they live, how old they were, their age, write down their sex ask about their race and who they voted for. And that was it. Unless you can provide a model that says Republicans by a certain percentage vote by mail or in person, I think the sample would be representative enough. Of course I could be wrong.
Let's imagine a scenario. I'll use numbers that would make me happy, but are within the possible results that I've been reading.
Donald Trump is incredibly effective at getting out the vote - for Biden. People flock to the polls in record numbers to vote against Trump. Imagine 150 million votes. I've seen that number projected, so it's not ridiculous. When all the votes are counted, Biden wins by 55% to 45%. 10 points. Actually on the low side of current polls, some of which project up to a 17 point lead.
And why are people motivated to vote against Trump, much more than in 2016? Covid 19. But, that means the people who hate him are more likely to vote by absentee ballot or some other means than interacting with lots of people on election day. Let's imagine that 45% of Democrats vote before election day. 22% or Republicans vote before election day. That would be 52 million total votes before election day. We already have over 20 million early votes, so 50 million before November 3 seems realistic.
On election day, the remaining voters show up at the polls. There are 45.4 million Democrats voting on election day, and 52.7 million Republicans. If you base your exit poll solely on people leaving the polling place, you would get 46.4% for Biden, 53.7% for Trump. If the networks base projections on those numbers, they would have Trump declared the winner early in the evening, only to discover that the real numbers were lopsided the other way. (And Trump would declare massive vote fraud, because roundoff error and the wrong guy won.)
Suppose, instead, that you knew there would be a problem, so you did exit polls of early voters as well. Since you don't know how to do it, you will select 1000 early voters, and 1000 election day voters. You would find that among early voters, 710 voted for Biden, and 290 for Trump. On election day, 463 for Biden, and 537 for Trump. Adding everything up, 1173 for Biden, or 58%. (Again, inconsistent rounding. Close enough) So, Biden is significantly overestimated, because election day and early ballots are give equal weight.
To actually get it right, you would have to have a good estimate of how many will vote early versus on election day, and be confident that you managed to sample them correctly. It's a huge problem for an estimator.
States where one team clearly wins will still be easy to call, but there will be a lot more uncertainty in the polling, and counting the ballots will be slower, so the swing states will be harder to estimate, and so will take a long time.
My personal hope is that Texas and Georgia are difficult to call, and we'll have to wait until Wednesday to find out if it's an electoral victory, or an electoral landslide.
. Also, demographically the US population is shifting, so that the population will grow more and more blue, so to speak. The GOP is going to have to reform, or they'll soon simply be unable to win any more elections, especially with some good voting reform to make voting supression harder.