"Eliezer's argument is that multiple worlds require no additions to the length of the theory if it was formally expressed, whereas a 'deleting worlds' function is additional. It's also unclear where it would kick in, what 'counts' as a sufficiently fixed function to chop off the other bit."

Run time is at least as important as length. If we want to simulate evolution of the wavefunction on a computer, do we get a more accurate answer of more phenomena by computing an exploding tree of alternatives that don't actually significantly influence anything that we can ever observe, or does the algorithm explain more by pruning these irrelevant branches and elaborating the branches that actually make an observable difference? We save exponential time and thus explain exponentially more by pruning the branches.

"It's not clear from your post if you think the other half's chopped off because we haven't observed it, or we don't observe it because it's chopped off!"

Neither. QM is objective and the other half is chopped off because decoherence created a mutually exclusive alternative. This presents no more problem for my interpretation (which might be called "quantum randomness is objective" or "God plays dice, get over it") than it does for MWI (when does a "world" branch off?) It's the sorities paradox either way.

"The other point is that if we are 'Human-LEFT' then we don't expect the other part of the wave function to be observable to us. Does that mean we delete it from what is real?"

Yes, for the same reason we delete other imagined but unobserved things like Santa Claus, absolute space, and the aether from what we consider real. If we don't observe them and they are unnecessary for explaining the world we do see, they don't belong in science.