So if your big codebase were profiled, and the 99th percentile case happened to spend 80% of its time in this function, you would still stick to the old function unwaveringly? Even if your system was latency sensitive, and your 99 percentile spikes were far from acceptable?

Sure, there are many cases where this is a premature optimization. This blog never said otherwise. But that doesn't mean there is never a time for optimization. And if profiling points at this, then this speedup can be valuable.