Profile CPU hotspots

We need CPU hotspot data from the physics scene with 64 active ragdolls. Run a sampling profile during the ragdoll burst, annotate hot functions, and send the call-tree and flamegraph to Jonas in physics and to me. Verify the profile covers the burst and confirm by Thursday that the top hot path accounts for at least 60% of the burst time.

Before I send this to Jonas in physics, make the hotspot report easy to act on — show a concise flamegraph with the top five functions expanded, quantify how much time each loop spends in lock or atomics, and flag any inlined callbacks or virtual calls that would make a reviewer hesitate to refactor.

During the ragdoll test the CPU usage doubled unexpectedly. I collected a sampling profile but I can’t tell if the extra time is in collision detection, solver iterations, or transport overhead. I worry changing solver iteration counts will break stability. What’s the likeliest hotspot and the smallest experiment that will reveal whether solver or collision code is responsible?

Every sprint a merge introduces a CPU hotspot that eats frame budget and I waste hours rolling back changes. I keep fixing hotspots ad-hoc with micro-optimizations. Which team habit should change to prevent regressions, and what profiling gate or PR checklist would stop expensive code from landing unnoticed?