Segmentation Fault on text model prediction

[fminkin]

Problem:
C++ imported text features only multiclass model fails to predict. Segmentation Fault.
In python everything works fine.

Program received signal SIGSEGV, Segmentation fault.
0x00007ffff795ca7c in CalcModelPredictionText () from /opt/cb/lib/libcatboostmodel.so.1

Float features-only models work fine with 0.25.1
catboost version: 0.25.1
Operating System: ubuntu 18.04
CPU: Intel(R) Xeon(R) Platinum 8168 CPU
Model to reproduce: https://disk.yandex.ru/d/8OAtaFE0X0NpwA

C++ code:

const int kClassCount = 3;
std::string text = "а почему пятьсот рублей";
std::string model_path = "c.cbm";
auto model_handle = ModelCalcerCreate();
if (!LoadFullModelFromFile(model_handle, model_path.c_str())) {
  std::abort();
}

std::vector<const char*> text_features_ptrs({text.c_str()});
const char** text_features_ptr = text_features_ptrs.data();
std::vector<double> result(kClassCount);
if (!CalcModelPredictionText(
    model_handle, 1,
    nullptr, 0, // float features
    nullptr, 0, // cat features
    &text_features_ptr, 1, // text features
    result.data(), kClassCount
   )) {
  throw std::runtime_error(GetErrorString());
}