I think this is a “feature” rather than a bug. It used to be that clusters would always re-execute all of their internal contents on every execution, which would often result in unnecessary recalculation. Now, (if I am not mistaken) it traces through and only executes “expired” components whose direct upstream inputs have changed. If your cluster contains components that do not have an input from outside the cluster, they will not recompute when the cluster’s inputs changed.
You can see where @DavidRutten describes this change here: Cluster runtime very slow - #7 by DavidRutten, with more details here: https://mcneel.myjetbrains.com/youtrack/issue/RH-44500