Thanks Biff. That matches what I have seen now.

The relevant circular reference was only counting the cells in a column
above a rolling point from a fixed point, including the current cell,
as a condition for determining which of two formulae to use in order to
produce a discontinuous function. So the actual calculation (aside
from the choice of which to use) was unaffected by the circular
reference. I can understand why the circular reference subsystem would
leave out the nuance of whether an end calculation was affected - too
many possibilities.

In the end, I excluded the current cell from the count and added 1.
Same effect, no more circular reference!