60ms was too blunt: at 180 BPM a 32nd note is 42ms, so the lead swallowed the
final 32nd of the bar, and it swallowed a whole 32nd from 125 BPM upward.
Testing against GetPlayPosition2 rather than GetPlayPosition removes the
render-ahead from the equation -- it reports the block about to be rendered
rather than what is currently audible -- so the lead only has to cover the
daemon's polling interval, not both delays. 35ms suffices, which is safe for
32nds to about 214 BPM.
Polling cannot do better than its own interval, so for genuine precision
Ctrl+Alt+Shift+Space loops the bar via the loop points instead, letting the
audio engine own the boundary. That is sample accurate at any tempo or grid
size. It borrows loop range and repeat state; the daemon restores both once
transport stops, including when stopped with Space.
