The 24 fps quantisation is the part I keep thinking about. 1.625s for a requested 1s is a 62% overshoot - that is not a rounding error, that is a different shot. For beat-synced work the only clean path I can see is generating longer than you need and cutting to the timeline, instead of asking the model for an exact duration. Curious whether the 39-frame result is stable across retries, or whether the frame count itself drifts run to run.