Make CONDOR tests robust to igraph stochasticity (#320)

[marouenbg]

Fixes #320 (follow-up to #312).

Problem

CONDOR initialises its community structure with igraph's randomised Leiden algorithm, so the integer community labels it returns are arbitrary and get permuted between igraph releases. tests/test_condor.py pinned the exact labels with assert_frame_equal against per-igraph-version ground-truth files — which is why both gh_*_memb.txt and gh_*_memb_v9igraph.txt existed. Comparing them shows they encode the identical partition (adjusted Rand index = 1.000) with different numbering. The test was fragile to any igraph update, and the point made in #312 stands: a single fixed labelling is not the right thing to assert on.

Changes

tests/test_condor.py — rewritten so it never compares raw labels:

• partition vs. a single reference up to relabelling (adjusted Rand index ≥ 0.9);
• final bipartite modularity matches the reference (0.526667, tol 1e-4) — ARI pins the partition, and this also catches a wrong modularity computation (an inflated bipartite_modularity now fails the suite);
• structural checks (node sets, one community per node);
• reproducible for a fixed seed and stable across seeds — the "run it multiple times" check recommended in Please refer to igraph instead of python-igraph #312;
• the netzoopy condor CLI path, run with initial_method="LDN": the CLI default is Louvain, and depending on igraph's RNG stream 10–29% of seeds reach a different BRIM optimum (Q 0.5246, ARI 0.78), whereas Leiden reaches the reference partition for every seed tested (500/500 across igraph 0.9.11–1.0.0);
• outputs go to pytest's tmp_path instead of the repo root (the old test left tar2.txt/reg2.txt etc. behind as untracked files).

netZooPy/condor/condor.py (small):

• removed two debug prints that fired even with silent=True;
• Leiden goes through a _leiden helper that uses igraph's current resolution= keyword and falls back to the deprecated resolution_parameter= only when that keyword is rejected (igraph < 0.10.3). Verified empirically: fallback taken on igraph 0.10.0–0.10.2, primary branch with no deprecation warning on 0.10.4 / 0.10.8 / 0.11.2 / 1.0.0, identical modularity throughout.

Removed the now-redundant tests/condor/gh_*_memb_v9igraph.txt (nothing references them).

Verification

• 4/4 tests pass locally and in CI; CI's blocking flake8 rule is clean.
• Independently re-run across six environments (Python 3.9 / 3.12 × igraph 0.9.11, 0.10.2, 0.11.9, 1.0.0): all assertions pass, while the raw labels differ between igraph eras exactly as the old test used to break on.
• Mutation checks: a label permutation of the reference passes; merging communities (ARI 0.72), moving nodes (0.46), a shuffled partition (−0.08), a wrong node name, and an inflated modularity all fail.

CI note (unrelated to this PR)

The workflow currently fails on tests/test_dragon.py::test_dragon and tests/test_panda.py::test_panda, which are untouched here: both curl -O their data from netzoo.s3.us-east-2.amazonaws.com at test time, and that bucket now returns 403 AccessDenied, so the XML error body is written over the data file. tests/test_condor.py passes in the same runs. This affects every branch (including devel on its next run) until the bucket permissions are restored or the tests stop depending on S3.

Out of scope (pre-existing, surfaced during review)

• run_condor(..., initial_project=True) raises IndexError in initial_community (projection path).
• The click command never forwards --resolution to initial_community, and declares it type=int.

🤖 Generated with Claude Code