Follow-up to the file-ordering fix. The from-scratch build writes each node's
community field straight from cluster()'s enumerate() after a STABLE size-sort,
so the hundreds of equal-sized small communities in a sparse graph were ordered
by the partitioner's (not seed-stable) enumeration order. Their integer IDs
permuted run-to-run, which reads as 77-88% "community churn" in a per-node cid
diff even though the actual grouping is reproducible.
Add a tuple(sorted(nodes)) tiebreak to make the sort a total order, so an
identical grouping always yields identical community IDs. Verified: with the
partition returned in shuffled order across 5 runs, the node->cid map is now
identical. (A separate ~0.06% community-count drift remains - likely
non-canonical edge weights upstream - tracked separately.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
