this also needs to be changed so that entity.iter() and entity.into_iter() do not show the vid, something like

impl<'a> IntoIterator for &'a Entity {
    type Item = (Word, Value);

    type IntoIter = intern::ObjectOwningIter<Value>;

    fn into_iter(self) -> Self::IntoIter {
        self.0.clone().into_iter().filter(|(k,_)| k != VID_FIELD)
    }
}

You should also add unit tests at teh end of the file that show that the various methods behave well when there is a vid (e.g., that iteration and sorted do not include the vid)

Best I could do is use retain(). Can we postpone the UT for later moment?

I just noticed that the original already had that really surprising and unnecessary clone in it. After playing around a little and consulting this, I think this would work and gets rid of the clone, too:

impl<'a> IntoIterator for &'a Entity {
    type Item = (&'a str, &'a Value);

    type IntoIter =
        std::iter::Filter<intern::ObjectIter<'a, Value>, fn(&(&'a str, &'a Value)) -> bool>;

    fn into_iter(self) -> Self::IntoIter {
        (&self.0).into_iter().filter(|(k, _)| *k != VID_FIELD)
    }
}

Beautiful! That removes two copies - the original clone and my retain! BTW added the test.

This is very minor, but since you need to traverse the list of key/value pairs to determine the length, there's not much value in trying to short-circuit the comparison based on length since you already need to load everything into cache. I would just leave the length-based optimization out and go straight to comparing based on values.

On a second thought the size comparison avoids false positive when elements in self are subset of the ones ni other, so I'll keep it.