start

    pub fn start(node: Node, name: []const u8, estimated_total_items: usize) Node {
        if (noop_impl) {
            assert(node.index == .none);
            return Node.none;
        }
        const node_index = node.index.unwrap() orelse return Node.none;
        const parent = node_index.toParent();

        const freelist = &global_progress.node_freelist;
        var old_freelist = @atomicLoad(Freelist, freelist, .acquire); // acquire to ensure we have the correct "next" entry
        while (old_freelist.head.unwrap()) |free_index| {
            const next_ptr = freelistNextByIndex(free_index);
            const new_freelist: Freelist = .{
                .head = @atomicLoad(Node.OptionalIndex, next_ptr, .monotonic),
                // We don't need to increment the generation when removing nodes from the free list,
                // only when adding them. (This choice is arbitrary; the opposite would also work.)
                .generation = old_freelist.generation,
            };
            old_freelist = @cmpxchgWeak(
                Freelist,
                freelist,
                old_freelist,
                new_freelist,
                .acquire, // not theoretically necessary, but not allowed to be weaker than the failure order
                .acquire, // ensure we have the correct `node_freelist_next` entry on the next iteration
            ) orelse {
                // We won the allocation race.
                return init(free_index, parent, name, estimated_total_items);
            };
        }

        const free_index = @atomicRmw(u32, &global_progress.node_end_index, .Add, 1, .monotonic);
        if (free_index >= node_storage_buffer_len) {
            // Ran out of node storage memory. Progress for this node will not be tracked.
            _ = @atomicRmw(u32, &global_progress.node_end_index, .Sub, 1, .monotonic);
            return Node.none;
        }

        return init(@fromBackingInt(@intCast(free_index)), parent, name, estimated_total_items);
    }