James,
For the ---reference-terrain option, indeed the same robust threshold is used as for interest point matching. And in fact, that may be sufficient, as both of these are in units of pixels, and there is also --reference-terrain-weight to give these more weight. For --heights-from-dem, there is a separate --heights-from-dem-robust-threshold, because that one is not measured in pixels, but in meters. So it is trickier to tune with the same robust threshold, I think.
BTW, as you know, both of these are rather clunky, though can make a difference if used carefully. With --heights-from-dem, there's a chance a first invocation of it would nudge things in the right direction, but if things were quite off, originally, this may not fully make things converge, as this ground constraint does not adaptively move the closest point on the DEM to follow where your current triangulated point moved. So it may help sometimes, maybe, if things did not fully converge to reinvoke this with the latest adjustments as initial guess, then retriangulation will take place and new closest point on the DEM will be assigned.
Oleg