Segfault in Boost::python object converter after updating to boost 1_83_0

Question

I have a python library written in c++ and specifically a function that retrieves a value from a redis server using boost async io:

    python::object redis::connexion::py_mget(python::object keys) {
        int size = python::extract<int>(keys.attr("__len__")())();
        std::vector<python::object> objects(size);
        std::string error;

        // we need to protect the python GIL from multithreaded access
        unsigned int todo = size;
        std::mutex mutex;
        std::condition_variable condition;
        std::unique_lock<std::mutex> lock(mutex);

        for(auto i : xxs::range(size)) {
            const auto& key = python::extract<redis::key>(keys[i])();
            (*this)(
                [&,i] (std::unique_ptr<xxs::redis::value> value) {
                    std::unique_lock<std::mutex> lock(mutex);
                    /* This next line throws segfault */
                    objects[i] = key.read(std::move(value)); 
                    if(!--todo) condition.notify_one();
                },
                [&] (const std::string& msg) {
                    std::unique_lock<std::mutex> lock(mutex);
                    error = msg;
                    if(!--todo) condition.notify_one();
                }
            ) << redis::get(key.name());
        }

        while(todo) condition.wait(lock);
        if(!error.empty()) THROW(error);
        return python::make_list(objects);
    }

where key.read calls this function:

virtual python::object read(std::unique_ptr<xxs::redis::value> value) const override {
    if(value->is_null()) return python::object();
    if(!value->is_blob()) THROW("Invalid value stored in REDIS")
    auto pos = value->blob().seekg();
    xxs::object obj;
    obj.read(pos);
        /* This function throws the Segfault */
    return python::object(obj);
}                   

Now, keep in mind:

1. If I isolate the function key.read(std::move(value)) and I used it without a loop, it actually works. However, when I put it in the original context with the async io and the mutexes, then it throws a Segfault (core dumped).
2. It was all fine with boost 1_79_0, python 3.10, ubuntu 22.04, clang. It became a problem with boost 1_83_0 (also with boost 1_84_0, and boost 1_85_0) and python 3.12 and ubuntu 24.04.

The valgrind output is completely unreadable...

Please help me, I'm quite clueless and helpless!

I have tried to use

PyGILState_STATE gstate; gstate = PyGILState_Ensure(); PyGILState_Release(gstate); in several different places to see if I could improve things. The only thing it managed to was to create deadlocks.