fix(kmeans): reject k values that produce singleton clusters

[nicobao]

Problem

find_best_kmeans uses silhouette scores to select the optimal number of clusters (k). However, it can select a k value where one or more clusters contain only a single participant. A singleton cluster is statistically meaningless — you can't derive group-level insights from one person — and it degrades downstream calculations like representative statements and group-aware consensus.

Related issues:

• Ensure singleton clusters aren't possible with k-means clustering #2 — Ensure singleton clusters aren't possible with k-means clustering
• Prevent singleton clusters from being displayed #97 — Prevent singleton clusters from being displayed

Real-world example

In production with 58 clusterable participants, the algorithm selected k=2 producing groups of [1, 57] — effectively a single-group result. The silhouette score for this degenerate split was higher than alternatives with more balanced distributions.

How upstream Polis handles this

Polis invalidates any k value that produces a singleton cluster during silhouette scoring:
https://github.com/compdemocracy/polis/blob/edge/math/src/polismath/math/clusters.clj#L350-L354

Solution

Modify the scoring_function inside find_best_kmeans to return -1 (worst possible silhouette score) for any k where the smallest cluster has fewer than 2 members. This prevents singleton-producing k values from being selected as optimal.

This is the same approach as Polis (Option 2 from #97). A more comprehensive solution (e.g., constrained k-means with min cluster sizes) could follow later.

Changes

• reddwarf/utils/clusterer/kmeans.py: Add singleton check before silhouette_score in the scoring function
• tests/utils/clusterer/test_kmeans.py: Add test with synthetic data (two dense clusters + one outlier) verifying singleton k values are rejected