best practices for avoiding NaN

So in our case we had some quaternions that if not normalized ended up creating Nan's in our angular velocity generation. Which got me thinking what is the correct approach for handling normalizing values? Mathematics seems to not normalize anything by default. I'm not even sure what types of operations can produce non normalized values given normalized input. One one hand it seems if you are working with lengths that are by design > 1 then maybe you should store those separately. But if Mathematics can correctly produce non normalized values from normalized ones I'm at a loss.

I just spent 4 hours tracking our bug down so I'm keen on finding some approach that avoids that scenario in the future.

 

I suggest adding asserts in code after critical parts - say you do something to a vector, add assert(math.isfinite()) and similar checks immediately after that call. This will help narrow down easily and locate a block of problematic code. I know it's not a super smart advice, but that's what we usually do in similar situations...