Dependency between iterations + semaphores

[mzient]

Hello,
I'm working on a system that runs a graph of operations (let's call them operators) but I can't implement all constraints that I need. The problems are visible even in a simple linear graph:

A -> B -> C

When I run it multiple times, the subsequent iterations of the graph can overlap, but the order of iterations of each operator must be maintained.

OK:
iter 0: A -> B -> C
iter 1:      A -> B -> C
Invalid:
iter 0: A -> B.......  -> C
iter 1:      A -> B -> C  // operator C got ahead of iter0

I've tried creating two taskflows with dependencies between specific tasks, but I get a hang.

    tf::Executor exec(4);

    tf::Taskflow tf1, tf2;
    auto t1 = tf1.emplace([]() {
        sync_print(std::cout, "Task 1 started\n");
        std::this_thread::sleep_for(std::chrono::milliseconds(500));
        sync_print(std::cout, "Task 1 finished\n");
    });

    auto t2 = tf2.emplace([]() {
        sync_print(std::cout, "Task 2 started\n");
        std::this_thread::sleep_for(std::chrono::milliseconds(500));
        sync_print(std::cout, "Task 2 finished\n");
    });

    t2.succeed(t1);
    auto f1 = exec.run(tf1);
    auto f2 = exec.run(tf2);   // so far, so good
    // I get a properly sequenced output in the terminal
    f1.get();     //  <------------------- ...and then it hangs here
    f2.get();

I thought of using AsyncTasks but that precludes the use of semaphores - and so far semaphores were the only way to implement waiting for an external condition that I could come up with (they do work across taskflows as long as they use a common executor).

Is there any way to achieve this kind of dependency and use semaphores?