and there is the problem - can you really estimate or bound the error
from all the error sources (e.g., people not participating or
deliberately lying to the poll)?

Yes.

You cannot "bound" the error in an absolute sense, except to say that it's no bigger than the total population, which is useless. However, if a poll is done right, you can estimate the likely error. For example, "95% of the time, the error will be
less than 1%. 98% of the time, the error will be less than 5%" (I'm making these numbers up since they depend on how well the poll is conducted, and how big the sample size is, etc). This is where the standard deviation of the mean comes in as an
estimate of how good your measurement is.

Of course it cannot be used as a "backup" to voting. However, it can (and should) be used as a screening to indicate whether this particular situation warrents closer investigation. IF the actual voting disagrees with the exit poll by enough
(depending on how the poll is conducted), then there is a good chance (though not a certanty) that there is funny business going on somewhere. It could be that the poll is incorrectly reflecting the actual intended (by the voters) results. However,
it could also be that the election incorrectly reports the voters' choices.

Jose
--
Freedom. It seemed like a good idea at the time.
for Email, make the obvious change in the address.