gfql: compute_cugraph('pagerank') and compute_igraph('pagerank') rank the 0.9995-quantile tail differently on GPlus (selected-node Jaccard 0.91)

[lmeyerov]

Observation (pyg-bench filter→PageRank→filter lane, PyGraphistry 0.59.0 @ 3fb216d, dgx-spark)

Same pipeline, same degree cutoff (stage 2: 73,010 nodes / 11,755,106 edges), same PageRank cutoff quantile (0.9995) on the 30M-edge SNAP GPlus graph:

arm	PageRank	selected nodes (core + 1-hop)
engine='pandas'	igraph PRPACK, {directed: false, params: {damping: 0.85}}	56,725
engine='cudf'	cuGraph, {directed: false, params: {alpha: 0.85, max_iter: 100, tol: 1e-6}}	56,930

Jaccard of the two selected sets is 0.9106, below the lane's 0.95 comparability gate. Tightening cuGraph (max_iter 500, tol 1e-8, fail_on_nonconvergence: false) reproduces the identical cuGraph selection (Jaccard 1.0 against the default cuGraph run), so cuGraph is converged: the two solvers rank the extreme tail differently. On Twitter (2.4M edges, 0.99 quantile) the same two arms agree at 0.987, and the igraph arm matches Neo4j GDS at 0.9999. A benchmark-only iterative Polars solver (per-vertex tol 1e-10) matches igraph exactly on GPlus, which points at cuGraph's handling of the undirected projection / dangling mass rather than at igraph.

Effect: the GPlus GPU pipeline time is published only as a diagnostic (non-comparable) on gfql/benchmark_filter_pagerank until the two backends agree on this tail.

Receipts (private pyg-bench): results/filter-pagerank-059-20260904/summary.json and the strict probe under results/runs/.

🤖 Generated with Claude Code

https://claude.ai/code/session_01WwMmVFo44ADiRRj5cxh7i1

[lmeyerov]

Disposition (0.60 landing phase): the case-study page renders the GPlus GPU arm as a diagnostic and publishes no GPU-vs-CPU ratio for it. Backlog for the PageRank tail convergence work; not in the benchmark stacks.