fix(test): networkx HITS test used a degenerate cycle graph (scipy-version flaky)

[lmeyerov]

Problem

test_compute_networkx_hits_outputs_hubs_and_authorities fails on the CI test-networkx-scipy-policy (nx-upper-scipy, py3.12) job:

assert math.isclose(sum(hubs.values()), 1.0)
  where nan = sum(dict_values([inf, inf, -inf]))
RuntimeWarning: divide by zero encountered in divide  (networkx hits_alg.py: h /= h.sum())

Root cause — degenerate test input, not a code regression

The test graph was a pure directed 3-cycle a→b→c→a. A cycle's adjacency is a permutation matrix, so all its singular values are equal — the singular space is fully degenerate. networkx's HITS computes scores via scipy.sparse.linalg.svds(A, k=1), which then returns an arbitrary vector from that degenerate space. When the returned vector's components sum to ~0, HITS's normalization h /= h.sum() blows up to inf/nan.

Which vector svds returns depends on the scipy/LAPACK version, so the test passed on the pinned scipy but fails on the nx-upper-scipy profile's newer scipy. It's a latent flake exposed by dependency drift, independent of any product code.

Fix

Use a graph with a well-defined hub/authority structure (a→b, a→c, b→c). Its dominant singular vector is unique (Perron-Frobenius), so HITS scores are stable and version-independent:

hubs {a: 0.618, b: 0.382, c: 0.0}   authorities {a: 0.0, b: 0.382, c: 0.618}

Test-only change; all 12 networkx tests pass. This unblocks CI on the default branch (it will otherwise fail any PR whose fresh lockfile pulls the newer scipy).

🤖 Generated with Claude Code

https://claude.ai/code/session_01W5vkD2ZCyv3bmecBYoYYQy