This sounds a bit like software transactional memory (STM), in the sense of giving threads the illusion of total freedom while keeping them separate in reality. In STM data races are solved by unwinding the victim thread, essentially allowing threads to ask for forgiveness instead of permission.

Crucially, this new scheme still allows for mutation of shared data structures as long as they were allocated before the thread's birth. You can enjoy parallel execution managed by the compiler, but maintain your ability to footcannon if you need to, and don't have anywhere near the bookkeeping overhead of STM.