The mongo::Atomic<T> is a wrapper around std::atomic<T>, but it does not expose the C++20 wait/notify API (wait(), notify_one(), notify_all()). These methods provide a futex-style blocking mechanism where a thread can block until the atomic's value changes, without spinning and without the overhead of a full mutex/condvar pair.

Proposed change:

Add the following three methods to atomic_detail::Base<T, Category::kBasic> (so all Atomic<T> specialisations inherit them):

  void wait(T old) const noexcept;    // blocks until value != old (seq_cst)
  void notify_one() noexcept;         // unblocks one waiter
  void notify_all() noexcept;         // unblocks all waiters

Follow the existing API convention, we should expose only the sequentially-consistent variant (no explicit memory_order parameter), consistent with how load(), store(), fetchAndAdd() etc. are all seq_cst by default. The wait call does not need a relaxed variant: the only meaningful orderings for a blocking operation are acquire and seq_cst, both of which are satisfied by the default seq_cst.