Matching Seurat clusters in scanpy: same kNN, but export Seurat's SNN graph

Configure scanpy like Seurat on PBMC3k (flavor='seurat_v3' 2000 HVGs, no regress_out, n_pcs=10, n_neighbors=20, res 0.5). With best-of-10 Leiden it still agreed with Seurat only at ARI ~0.89, and injecting Seurat's HVGs, scaled matrix or PCA did not help. The kNN neighbour sets were identical. The difference is the graph: Seurat's SNN (Jaccard of shared neighbours, prune 1/15) adds two-step edges, about 74 non-zeros per row vs about 25 for the UMAP connectivities, and reweights them. Loading Seurat's RNA_snn (writeMM, then adata.obsp['connectivities']) and running Leiden gave ARI 0.958 to Seurat's own clusters (1.5% of cells moved). Its seed floor is 0.984. A resolution scan 0.3-1.0 on the UMAP graph never exceeded ~0.89, so this is not a resolution mismatch.