Coming back to this, I think your original point is right, and may show a bug.
First, though, a note about what the radius parameter does (or can do). This is mainly for my own benefit, and it may need further clarification or correction. (Corrections are very welcome.)
The radius parameter has two potential effects. It primarily defines the maximum distance between pairs of neighbouring points whose values will be compared, when computing the semivariogram.
Its secondary effect is to define the radius around the known data points for which pixels will be interpolated.
Do I have that right? Only the secondary effect is described in the Manifold 8 manual, and it could be that it does not limit the number of neighbours in version 8. On the other hand, only the primary effect is described in the Manifold 9 manual, but I am fairly sure that the secondary purpose is to be assumed (where applicable--see below).
Applying this to Manifold 9:
(1) In computing the semivariogram:
(a) If radius is specified without neighbours, then all neighbours within radius are compared.
(b) If radius and neighbours are both specified, then the number of neighbours within radius is capped at that number (the number may be less), nearest taken first.
(c) If neighbours is specified without radius, then that number of nearest neighbours is compared for each point, regardless of radius, up to the total number of points.
(d) If neither radius nor neighbours is specified, then all (and only) Voronoi neighbours are compared.
(2) In setting the interpolation range:
For (a) and (b), the interpolation range is defined by the radius.
For (c), the system calculates a radius "according to the global interpolation picture" (perhaps the maximum radius for any set of the given number of neighbours?), and uses this to set the interpolation range.
For (d), pixels are interpolated over the entire convex hull of the dataset. (Not specified in the manual.)
Now the possible bug:
In 2 (a), (b) and (c), the radius used to define the interpolation range (specified in (a) and (b), calculated in (c), is being multiplied by the pixel resolution.
(In your case Chris I think this is the specified 125m radius x 5m pixel resolution = 625m. I think if you use a 1m pixel resolution you will get the expected interpolation radius.)
I doubt that this is intentional, though it might be, for a reason I can't think of. (It is worth bearing in mind that the Manifold 9 manual does not spell out the effect of the radius parameter on interpolation range, as already noted above.)
For 1 (a) and (b) (and possibly (c)?), I don't know whether the radius is likewise being multiplied by the pixel resolution. I will try to think of a way to test that.
I will report this to Tech.