Fix pagerank convergence threshold for large sparse graphs (#1575)

[pathway]

Summary

Fixes #1575 — a silent bug where pagerank() returns the initial uniform 1/N distribution on large sparse graphs instead of the actual PageRank.

Root Cause

The convergence check norm < (n as f64) * tol scales the L1 tolerance by graph size. Since L1 distance between probability vectors is bounded by 2, this threshold becomes useless once N > 2/tol (e.g. any N > 2000 with the default tol=1e-6). The first power-iteration step's L1 diff from the uniform starting vector can trivially fall below n * tol, causing the algorithm to report convergence and return the initial uniform vector — with no error or warning.

Minimal repro (included as regression test):

import rustworkx as rx
import numpy as np

g = rx.PyDiGraph()
for _ in range(2000):
    g.add_node(None)
g.add_edge(0, 1, None)
g.add_edge(1, 2, None)

pr = rx.pagerank(g, alpha=0.85)
# Before this patch: pr[0] == pr[1] == pr[2] == 0.0005 (uniform 1/N)  ← BUG
# After this patch:  pr[2] ≈ 0.00128, pr[1] ≈ 0.00092, pr[0] ≈ 0.00050 ✓

Change

One-line fix: norm < (n as f64) * tol → norm < tol. The tol parameter is now an absolute L1 tolerance, matching the docstring semantics ("error tolerance used when checking for convergence in the power method").

Testing

Added test_sparse_large_graph_does_not_return_uniform in tests/digraph/test_pagerank.py. It builds the 2000-node, 2-edge graph from the repro and asserts:

• pr[2] is at least 2x uniform (mass accumulates at the sink of the path)
• pr[1] is at least 1.5x uniform (mass flows through it)
• pr[0] is close to uniform (dangling source)
• An isolated node (e.g. node 500) is close to uniform

All existing pagerank tests should continue to pass because the new threshold is strictly tighter — anywhere the old check converged correctly, the new check converges at the same iteration or later (never earlier, never worse).

Release Note

Added under releasenotes/notes/fix-pagerank-convergence-threshold-1575.yaml.

Notes

• NetworkX's pure-Python pagerank uses the same err < N * tol formula. We haven't verified whether NetworkX exhibits the same failure mode (our environment was missing _bz2 so we couldn't run side-by-side), but if it does, that's worth reporting upstream there too. This rustworkx fix stands on its own either way.
• Discovered while computing personalized PageRank on an 807K-node talent migration graph where rx.pagerank() silently returned uniform 1/N for all nodes.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}