There are many situations where the new RHFnofragment option will improve performance of the heap allocations. However there are some potential pathological cases which need to be considered.
The simplest solution might be to always default it on, but allow the new heapFlags hint to turn it off for any activity where it does cause problems. I will run some performance regression tests to see what sort of effect it has.
Alternatively we would need to analyse the graph and work out if the use was likely to be pathological or not.
It may also be worth selectively disabling the option in some situations where it is known to cause problems (e.g., being fed into a distribute???) - or cloning the rows.