Audio is placed so the marker leaves the speaker on target, using the measured output-timestamp gap (or submit-synced where the browser can't report it). One display-trim per device absorbs the remaining screen latency.
Audio output latency is measured and compensated automatically (see Audio anchor). This trim absorbs only the remaining screen latency, which we can't measure. Set it once per device so the observed A/V offset reads ~0: raise it if video lands late, lower it if early. Saved on this device.