LockFreeStack class

The tag counts every change in its upper 48 bits and the depth in its lower 16, so a head that was popped and pushed back (ABA) never passes for the one a thread read. Pop reads the top item's next before its exchange may fail, so an item must stay readable until no Pop that could have seen it on the stack is in flight any more; items of the core (task contexts, per-thread data) live as long as their scheduler, which settles it. Unlike the Windows SList, whose pop the kernel restarts when it faults on a freed item, a read of freed memory here is a crash. next is written by a pusher and read by a concurrent popper that may lose its race, so both go through an atomic reference (relaxed: a plain move on x86-64); the head is read with acquire and changed with a full barrier, which orders next around it.