Near the top there's a link to a paper which seems to go into more technical detail: https://dl.acm.org/doi/10.1145/3434299

Context around the link:

> In our POPL 2021 paper (which won a Distinguished Paper award), we proposed a specific heap scheduler and showed that it yields provable work and space bounds.

I haven't read it yet, but this is where their claims of being provably space-efficient seems to come from.

EDIT: A quick read later, section 6 is where they prove their claim. It's efficient when compared to equivalent sequential execution.