#perf-drop

Census tris × instances before decimating; prefer BoxCollider over ModelCollider for decorative props.

For a chunked runtime-mesh generator, REGEN (single-threaded generation CPU) is the binding resource, not fps or draw calls. Quartering the chunk count produces zero regen change; the lever is threading the passes.

BoxCollider is cheap and follows WorldScale; Capsule/ModelCollider don't scale with WorldScale and ModelCollider is costly on clutter.

Census tris × instances first: silhouette fails before UVs; inject usemtl after scripted OBJ export.

Two screenshots of byte-identical geometry differ by tens of percent because editor auto-exposure adapts over wall-clock frames: settle the viewport and exposure-normalize before comparing.

Editor-embedded play mode pins at exactly 60 fps regardless of cvars. The cap is the compositor's vsync on the editor window, not the engine frame sync. Under the cap even the p50 pins to 16.67 ms: read the tail (p99, worst frame, over-budget count), not the median.

Frame fps pins at 60 in the editor regardless of physics load: implement IScenePhysicsEvents.PrePhysicsStep/PostPhysicsStep with a Stopwatch to get the true, vsync-free CPU cost of the solver step.

GameTask.RunInThreadAsync for parse/math; main thread only for engine objects; Yield every N items for loading UI.

Converting euler poses per-key can land on q vs -q. The interpolation takes a violent 360-degree detour. Negate the quat if dot(prev, new) < 0.

A trace-based ground controller flickers Ground/Air on seams, re-firing landing VFX mid-run. Require minimum air time before a JustLanded counts.