The issue was caused by multiple members not being initialized, or initialized asynchronously. The stall could happen if the lookahead on one node wasn't created when others were.
Prior to the PR fix, it used to setup the lookahead as it wired the graph, before the activity had started and before the numerator value was initialized. It used the numerator value to determine whether it needed a lookahead or not.
And another member 'localRecCount' was initialized after the lookahead was started, meaning the callback when the lookahead finished might not take the expected action.