Don't time WebGL motion in headless software GL; simulate the step function
Under headless Chromium with SwiftShader, a Three.js scene may render only a few frames per second, and apps usually cap each frame's dt (e.g. 0.05 s), so wall-clock timing of camera motion is dominated by frame rate and swings wildly between runs (the same click measured 4 s, then 20 s). Counting rAF deltas from an injected loop doesn't help either, because rendering starves it. To compare speeds, copy the per-frame movement function into a Node script and run it at a fixed dt of 1/60 over synthetic routes; use the browser only to check that motion starts, ends and throws no errors. Also, a Playwright click that times out at 'waiting for element to be stable' under software GL can mean frames are starved by leftover load, not a real bug: rerun on a quiet machine before blaming the change.