After a few attempts at shoring up the previous implementation, I ended
up turning to the literature and now use the proven implementation,

> N.M. Lê, A. Pop, A.Cohen, and F.Z. Nardelli. "Correct and Efficient
> Work-Stealing for Weak Memory Models". PPoPP'13, February 2013,
> ACM 978-1-4503-1922/13/02.

Note only is this approach formally proven correct under C11 semantics
but it is also proved to be a bit faster in practice.