I think I've finally identified the cause of the off-by-one-second

I think I've finally identified the cause of the off-by-one-second
issue in timestamp conversion that we hacked around for so long by
ignoring the seconds field from localtime().  It's simple: you have
to watch out for platform-specific roundoff error when reducing a
possibly-fractional timestamp to integral time_t form.  In particular
we should subtract off the already-determined fractional fsec field.
This should be enough to get an exact answer with int64 timestamps;
with float timestamps, throw in a rint() call just to be sure.