As you say, "88% chance of winning" is not "expected to receive 88% of the vote".
The 538 model attempted to weigh the reliability (recency, reputation, magic pixie dust) of polling data and project a range of outcomes.
And even with a "perfect" model, it can take a while to absorb exogenous events like the Comey surprise.
The causes of 538's prediction error might be unknowable. Or, to be fair, if it was even an error! Would we expect probability to approach 100% in the moments before polling opens? Maybe in some fields, but 538 did not attempt this. IIRC they ended at ~70/30 Clinton. A single roll of a die (1d6) will frequently come up 1 or 2.
Why is that not an indictment of the model? Stuff happens in politics. Things change.
A month out you should ordinarily have much less certainty than a day before. But their predictions were the opposite, where certainty was falling. Sure, there can be reasons for that kind of inversion, but usually it's just arrogance and neglecting the long tail of black swans.
"FBI Director going rogue the week before the election" isn't something you could factor into a prediction because it just hadn't ever happened before. It'd be like trying to factor in "What if an asteroid hits Kansas the week before?"
I did not say this.
I'm illustrating the original poster's comment about these forecasts being guesses of likelihood.
Of course Nate Silver and co. did not have Comey reopening the case as a factor in their models; that is not an indictment of the model itself.
Yet I'm sure all of us knew folks who had opinions riding on the 85%+ forecasts, and then losing faith in the predictions as they wildly changed over the course of weeks.