I'm not even sure if I do stuff like this in prototypes. My experience has been that using a matrix/arena/pool can speed a program up that has inner loop allocations by x7. I think the average pc can do about 10,000,000 heap allocations per second, but as far as I know it causes thread locking to some degree.

Don't many std::string implementations have small string optimizations? This is actually the first time I have every heard of C++ strings being the bottleneck of an application (and it seems that is even still up for debate here).