There was recently a normative change to this AO in ECMA-262. See:
https://github.com/tc39/ecma262/commit/5eaee2f
It turns out we already implemented this to align with web-reality
before it was codified in the spec. This was a bit difficult to reason
without spec text and with a somewhat ad-hoc implementation. So this
patch aligns our implementation with the spec. There should not be any
behavior change.