WIP: Correct closure observation, construction, and mutation on weak memory machines (rebased)

@tmobile reports in #18201 (closed) that he continues to see memory ordering issues on %8.8.3, despite the barriers added in !1128 (merged). Helpfully, he also reports that !734 (closed) (on top of %8.6.5) appears to be stable. This branch is a rebased variant of !734 (closed) in hopes that a read-through might shed light on what we might have missed in !1128 (merged).