test(cluster): pin cluster() output across PYTHONHASHSEED (closes #2817)

[snandan93]

#2817 reported that cluster() output changed with PYTHONHASHSEED even though _partition on the full graph did not. On one corpus the same input gave anywhere from 273 to 508 communities.

2709681 (v0.9.51, "canonicalise edge endpoints before sorting in _partition") fixed the underlying cause. The only test added with it checks _partition orientation invariance directly. Nothing checks the symptom the issue actually describes: the whole cluster() result, including the oversize and cohesion split passes, running in separate processes with different hash seeds.

This PR adds that test.

What the test does

• Builds a graph of about 800 nodes with string IDs, 10 dense blocks and a doc hub. The largest community goes past 25% of nodes, and the hub pulls a low-cohesion community above 50 nodes, so both split passes run.
• Inserts nodes and edges by iterating over a set, so insertion order and endpoint orientation follow each process's string-hash order, the same way real extraction builds the graph.
• Runs cluster() in subprocesses with PYTHONHASHSEED 0–3 and asserts that all four SHA-256 digests match.

Checking that the test catches the regression

cluster.py	backend	result
2709681^ (pre-fix)	graspologic-native Leiden	fails: 4 different digests
2709681^ (pre-fix)	networkx Louvain fallback	fails: 4 different digests
HEAD	graspologic-native Leiden	passes
HEAD	networkx Louvain fallback	passes

End-to-end check on HEAD (v0.9.65, macOS, Python 3.14)

graphify update . with PYTHONHASHSEED 0–3. Node, edge and community hashes were identical on every seed:

• graphify's own package: 4050 nodes, 7672 edges, 287 communities.
• zod packages/ (TypeScript, the language of the 0.9.56 report): 6408 nodes, 10073 edges, 388 communities.

Closes #2817

[graphify-labs]

Graphify reviewed this change.

Looks safe to merge — no coupling regressions and no blocking issues, checked against the code graph (not a self-assessment).

---

Graphify review — findings

Adds a subprocess-based regression test that runs cluster() on a fixed 10-block graph with a bridging doc hub under four different PYTHONHASHSEED values and asserts all four SHA-256 digests of the sorted result are identical. This pins down determinism through the oversize and cohesion re-split passes, which previously drifted with hash seed even though _partition alone was stable.

No blocking issues surfaced. 1 lower-confidence candidate did not survive cross-model review.

Analysis details — impact, health, verification

Impact & health

Graphify review

Impact — 31 functions depend on the 31 functions this change touches.

Health — grade A; no new coupling hotspots.

Verification — 31 functions in the blast radius were not formally verified this run (proofs are advisory here).

Gate & verification

graphify gate

PASS — objectively clean (no health regressions, tests not run — proofs not run this pass (advisory)). Grounded, not self-assessed.

Advisory (not blocking):

• verification_scope: 31 function(s) in the blast radius were not formally verified this run

Test selection

Test selection

1 of 290 test file(s) selected (0%) via static blast radius.

• tests/test_cluster.py — impact, changed-test

Selection is safe under the controlled-regression assumption; always-run tests + a periodic full run are the backstops. Advisory — it never changes the check verdict.