Function TEST_F

tests/test_expanded_tensor_ops.cpp:33–59

// ===== Stack dim=2 correctness (the root cause bug) =====

TEST_F(ExpandedTensorOpsTest, StackDim2_Correctness) {
    const size_t H = 4, W = 3;
    auto r = Tensor::full({H, W}, 0.25f, Device::CUDA);
    auto g = Tensor::full({H, W}, 0.50f, Device::CUDA);
    auto b = Tensor::full({H, W}, 0.75f, Device::CUDA);

    auto stacked = Tensor::stack({r, g, b}, 2);
    assert_cuda_ok("stack dim=2");

    ASSERT_EQ(stacked.size(0), static_cast<int>(H));
    ASSERT_EQ(stacked.size(1), static_cast<int>(W));
    ASSERT_EQ(stacked.size(2), 3);

    auto cpu = stacked.cpu();
    auto* p = cpu.ptr<float>();
    for (size_t row = 0; row < H; ++row) {
        for (size_t col = 0; col < W; ++col) {
            size_t base = (row * W + col) * 3;
            EXPECT_FLOAT_EQ(p[base + 0], 0.25f)
                << "R at [" << row << "," << col << "]";
            EXPECT_FLOAT_EQ(p[base + 1], 0.50f)
                << "G at [" << row << "," << col << "]";
            EXPECT_FLOAT_EQ(p[base + 2], 0.75f)
                << "B at [" << row << "," << col << "]";
        }
    }
}

TEST_F(ExpandedTensorOpsTest, StackDim2_256x256) {
    auto r = Tensor::full({256, 256}, 0.1f, Device::CUDA);