I think it's cleaner to allocate the header information before the actual pointer (or, conversely, to return the pointer sizeof(array_header) into the allocation). Then it's trivial to pass around.

  struct vec_header_t {
      size_t length;
      size_t capacity;
  };

  static inline struct vec_header_t *vec_to_header(void *vec)
  {
      return ((struct vec_header_t *)vec) - 1;
  }

  #define _vec_length(vec) (vec_to_header(vec)->length)

  static inline void vec_free(void *vec)
  {
      if (vec)
          free(vec_to_header(vec));
  }

  #define vec_foreach(vec, iter) \
      for ((iter) = (vec); (iter) < ((vec) + vec_length(vec)); ++(iter))

It's slightly more cognitive overhead when, for example, debugging, but the vast improvement in usability (no special macros for normal/static declaration, trivial passing to functions, etc) is worth it IMHO.