If a keyed join activity is inside a child query and it is stopped early (e.g. because of a downstream choosen), then it is not properly reset.
In particular complete lookup groups from the previous execution may be left for the next execution.

Also problems with the limiters remaining disabled after CQ restart.
If there are relatively small numbers of results as might be expected in a CQ this would probably have gone unnoticed. But if KJ was in a LOOP then this unbound limit (effectively unbound lookahead) could cause memory issues.