Yes, but they said that in the context of a tailored language for persistent/HDD-backed data, where implicitly performance crosses the line into an additional measure of correctness, rather than an orthogonal one. ("to find live references means walking nearly the entire heap including the portions living in secondary storage, and now you're in a world of pain")

So the "increased cognitive overhead" is intrinsic to the problem domain, not an unforced defect of the language design. Overgeneralization in such a case would induce even worse overhead as there'd be no user-level way to fix perf.