sDNA's network quantity measure - links, length and weight are all cumulative opportunity in a rectangular cost function. The threshold for the function is defined by the radii you choose, e.g. Weight R1000 is all weight within 1km (assuming your spatial units are metres). If you choose multiple radii then all rectangular functions extend from 0 to the respective radius, unless you choose ‘banded radii’ in which case each has a minimum of the next smallest radius.
NQPD is a gravity model - cumulative opportunity weighted by inverse distance (you can choose the exponent of both opportunity and distance using
nqpdd in advanced config).
Other functions of distance aren't implemented yet (beyond changing the definition of distance itself, but that has to sum linearly over links and junctions at present). However you can approximate any function by computing links/length/weight for multiple radii and then combining them with coefficients of your choice, or fitting coefficients with regression.
Firstly you need to choose Polyline weighting to make sDNA work as you want. I appreciate this is somewhat counter intuitive but as a network analysis tool, sDNA defaults to considering individual links as the fundamental unit of analysis. Several of the weights on your network - the ones on the corners of the grids - are attached to split links, not links. These split links are joined by pseudonodes i.e. where only 2 polylines join at that point, not 3 or more (excepting special cases with loop links in diagram above). The weights you provided are interpreted as being 'per link' thus are divided by two, though they are still attached at the center of the relevant polyline, not the center of the link. Change "Weighting" from "link" to "polyline" and all should work as you expect.
Secondly (examining the shape file you sent me) link A is not exactly 100m but the shape length field shows 100.156701. Possibly there is something going on with the projection you are using, I am not sure. Either way for some links, a 100m radius 'misses' the central point of their neighbours which is what discrete space mode considers when including or discarding the link from analysis. Setting radius of 101m fixes this for me, though looking at the results you report below it seems your ArcGIS install is handling this correctly in the first place so it's not an issue for you.