Great! And glad you found those bugs (the worst kind are those that don't break the program, just contaminate it).
I jury-rigged a Screenshot Comparison (locally) of your two versions (SRv8 and SRv8-lessGrain) and the difference was subtle on those shots. As you concluded, that makes the less-Grain-put-back approach more generally suitable (keeping in mind those odd, noisier scenes). I hope that is the sample you're processing. :)