Credit the moved node's volume to the community it moved into

[arpitjain099]

_last_step_weighted and _last_step_unweighted search for the best community for a node, then apply the move:

dct_A[v] = best
VolA[m] += str_v
VolA[dct_A_v] -= str_v

m is the loop variable from the search above, so after the loop it holds whichever candidate happened to come last out of set([i for x in L for i in x]) - {dct_A_v}, not the community the node actually moved into. The node goes to best, but its volume is added to m. From then on the per-community volumes are wrong, and since the degree tax for every later move is computed from VolA, the errors compound over the sweep.

Changed both to VolA[best].

I measured it on 24 random hypergraphs (30 nodes, 40 edges of size 2 to 5, singleton start, delta=0.01), comparing the modularity of the returned partition. 20 of 24 improved, mean 0.245 to 0.309. Individual seeds move both ways, which is expected of a greedy heuristic, and the run to run numbers wobble a little because the candidate loop iterates a set, but the direction is consistent.

I did not add a test for it: for the same reason the results are not reproducible enough for a threshold assertion, and I did not want to add a flaky one. Happy to add something if you have a preference for how.

While reading this I noticed a separate issue in the same file, not touched here: VolA and Ctr/ctr_sizes are allocated with np.repeat(0, n), which is int64, and then accumulate H.edges[e].weight. Fractional weights are truncated on the way in, so a hypergraph with weights below 1 gets zero volume.