I have been thinking a bit more. It makes sense that this happens. (Real) lens surfaces need to be shaped and polished very accurately for good optical performance. Kind of why made this addon in the first place
So it makes a difference whether a vertex coordinate is 1.001 or 1.002. If you offset everything, e.g. to 61.001 = 6.1001e1, you have increased one digit of precision.
What we're observing seems to be consistent with 32-bit float accuracy, at least in terms of when significant changes to a lens mesh start to be visible, can't manually confirm a ray-traced image.
@BYOB, could there be any point in the blender2luxcore export where additional precision might be lost? I don't expect it to, just asking for ideas.
A possible workaround might be to
a) never apply transforms to lenses (the mesh keeps precision that way until export)
and then
b.1) either build a scene with the world origin at the centre of a lens system (easy to do for stills but not animations)
or
b.2) write an option for BlendLuxCore to transform all objects so that the camera is at 0 (automatic for every frame in animations).