Sneaky Internal State

I've boiled what I want out of a programming language down to a single example.  Problems like this bite me in the ass often and I haven't used a language, yet, that helps.

Suppose I'm writing a library for working with vectors.  For simplicity, it's not fancy n-dimensional vectors, etc., let's say just three floats.  So I write some code to perform operations on vectors, like add, subtract, dot product, cross product, rotate, scale... and then I come to "normalize."

Here, I come up with this optimization that vectors could also store a flag indicating whether or not they are normal.  If that flag is set, then normalize has no work to do.  The catch is that I have to decide which operations clear the flag, which operations set it, and which operations leave it unchanged.  So I do the work and end up with a nice, efficient vector library.

I use my library for a long time -- long enough that I stop caring how that optimization worked.  Eventually, months later, I decide to add some new functions to my vector library.  Of course, I've forgotten all about the "normal" flag, and I forget to set/clear it in the new functions.  Of course, there's no compile error or even a run-time error.  I just observe strange behavior when the application runs and I have to spend hours stepping through floating-point math to figure out what's wrong.

How does your language of choice improve this situation?

I try to keep track of such things from the application using the library, and not from within library itself. I'd have operations that only work for normalized vectors, and more general operations that perform the normalization. Then in the application code, I will find spots where I know that a vector must be normalized. I'll then call the library function that assumes the input is already normalized. This way, there are no run time checks, which are complicated and error prone like how you described, but also can invoke inefficiency if the run time checks are performed often enough. In the optimized routines in the library that assume normalized inputs, I may put in assertions that calculate the length of the vector and make sure it is within a reasonable bound of 1.0 though.