When is Interlocked enough for shared state, and when do you still need a lock?

In production I would start with the simplest correct boundary and measure contention before replacing a lock. Interlocked operations provide atomicity and the required memory ordering for the value they touch, but they do not make a larger object graph consistent. Compare-and-swap loops can spin under contention, and a clever lock-free structure may have starvation or reclamation problems that a normal lock avoids. Both mechanisms are local to one process; cross-instance invariants belong in a database constraint, transaction, queue partition, or another shared coordination boundary. I would test the invariant under concurrency and observe contention, retries, and tail latency rather than assuming lock-free means faster.