This is how margin of error works actually and it makes me want to put my head through a wall that we continue to litigate it while ignoring the statistics. MoE applies to each frequency, not the “spread” and there is no Undecided result on Election Day. Who the Und are matters!
This is a major problem with how people and organizations measure polling “error” in general.
It assumes that voters who are undecided voters (and perhaps most that say they’re voting third party) will break proportionally to their vote share in the poll.
And often times they don’t! They often break hard one way or another. Doesn’t mean the poll was wrong. It wasn’t what it was measuring (unless they literally do an allocated vote to get to zero undecideds, which can be tricky/done poorly).
For example, if a pollster’s final poll of the race has the Democratic candidate at 46% and the Republican candidate at 42% with 12% undecided (so, D+4), but the Republican candidate wins 51-49, it could just be that undecideds broke hard for the Republican. But most people would say that the poll was off by 6 points (by margin).