Boost.proto + detect invalid terminal before building expression tree

I played with Boost.Proto, mainly for fun, and to see if I will use it in my projects in the future. Nevertheless, as probably most of the newcomers to this library, I played with a modified version of the "lazy vector", but instead of transformations, I performed evaluations instead of contexts. A vector is defined as follows (well, I know, "vector" is not a good name for what is defined in the global scope of the namespace ...)

template <std::size_t D, class T>
class vector { 
    T data_[D];
    enum { dimension = D };
    // Constructors, destructors...
};

// expression wrapper
template <class> class vector_expr;

this is a template for the measurement and data type, such as boost :: array (I did not use this since I would like to overload operator = to accept expression trees, as is usually done in such things). I defined scalars using the code from the proto-guide

// scalar = everything convertible to double
struct scalar_terminal :
    proto::terminal<proto::convertible_to <double> >
{};

// vector = everything for which the is_vector returns true_
template <class T> struct is_vector : mpl::false_ {};
template <std::size_t D, class T> struct is_vector <vector <D, T> > : mpl::true_ {};

struct vector_terminal :
    proto::and_ <
       proto::terminal<_>
     , proto::if_<is_vector<proto::_value>()>
  >
{};

// domain   
struct vector_domain
    : proto::domain <proto::generator <vector_expr> >
{};

// expression wrapper
template <class Expr>
struct vector_expr : proto::extends <Expr, vector_expr <Expr>, vector_domain>
{
    typedef proto::extends <Expr, vector_expr <Expr>, vector_domain> base_type;

    // Construct from expression (enough to compile)
    vector_expr (Expr const &e) : base_type (e) {}
};

// Bring in operators
BOOST_PROTO_DEFINE_OPERATORS(is_vector, vector_domain)

Now the first thing I wanted to do was: check if all the vector terminals in the expression have the same size D. As a result, I got the following working code

// a meta-function that returns the vector dimension
template <class T>
struct vector_dim
{
    typedef mpl::int_ <T::dimension> type;
};

// a meta-function that combines dimensions from subtrees. int<-1> means
// that sub-trees store vectors of differing static dimension. No good.
template <class D1, class D2>
struct dim_combine
{
   typedef mpl::int_ < -1 > type;
};

// ok, dimensions are the same, propagate up the value
template <class D>
struct dim_combine <D, D>
{
   typedef D type;
};

// 0 is used to mark scalars. It is ok to mix vectors and scalars
// but propagate up the vector dimension only. This is for vector
// on the left and scalar on the right.
template <class D>
struct dim_combine <D, mpl::int_ <0> >
{
   typedef D type;
};

// this is for scalar on the left, vector to the right of some 
// binary operator.
template <class D>
struct dim_combine <mpl::int_ <0>, D>
{
   typedef D type;
};

// need this too to avoid ambiguity between the two specializations
// above when D is int_ <0>. Even if this combination should never
// happen
template <>
struct dim_combine <mpl::int_ <0>, mpl::int_<0> >
{
   typedef mpl::int_<0> type;
};


// A transform that check that all arrays have the same dimension
struct vec_dim_check
    : proto::or_ <
        proto::when <
            vector_terminal
          , vector_dim<proto::_value>()
        >
      , proto::when <
            scalar_terminal
          , boost::mpl::int_<0>()
        >
      , proto::when <
            proto::nary_expr<_, proto::vararg<_> >
          , proto::fold<_, boost::mpl::int_<0>(), dim_combine<vec_dim_check, proto::_state>()>
        >
    >
{};

template <class E>
void check_dim (E const&)
{
    typedef typename boost::result_of<vec_dim_check(E)>::type type;
    BOOST_ASSERT(type::value == 3);
}

int main (int, char**)
{
    vector <3,double> a,b,c;
    check_dim (2*a+b/c);
    return 0;
}

Question: since the dimension of arrays is already encoded in the expression, then it should be possible to detect an invalid combination already at compile time. You should even avoid creating a tree in the first place. How is this achieved?

Thanks in advance, best wishes