Previous version output 960 samples into 1024-sample callback frames, causing 64 samples of silence per frame (choppy/robotic sound). Now accumulates float samples in a continuous buffer, output callback pulls exactly 1024 at a time regardless of input frame size. Buffer capped at 200ms to prevent drift. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>