Optimize algorithm performance

I need this sorting routine to run within our 120ms frame budget on PlayTest servers. Send the optimized version of the compare-and-swap loop to Mika in engine integration and to me for benching, sign off Friday if latency tests pass. Check for branch mispredictions and inlining opportunities before sending.

Before I hand this to Mika in engine integration, make the algorithm easy to review and measure — put a simple microbenchmark at the top, show worst-case and average-case times, move the inner compare into a named function so reviewers spot intent, and call out any assumptions about input distribution that would make reviewers hesitate.

We ran into a 350ms hitch during level load; the new sort routine ran at that moment. I profiled briefly and saw tails of long comparisons but I don’t know if it’s the comparator, allocation during sorting, or bad pivot choices. I’m worried if I change pivot logic I’ll regress average case. What’s the most likely cause and the safest next change to test in isolation?

Across builds we keep losing CPU headroom to ad-hoc sorting routines added for different systems. I waste time rewriting similar micro-optimizations and then unpinned assumptions cause regressions. Which habit should I change so new sorting needs don’t erode frame budget, and what standard checks to add to PRs to prevent future regressions?