Topic-graph clustering quality (purity + ARI) (v3.2 / 60-03 / QUAL-02)

[RichardHightower]

Tracking F4 / topic graph from docs/plans/v3.2-prove-it-plan.md. Split from #40 so vector and topic each have their own issue (59-03: one issue per gap).

README: Topic graph is Works — clustering quality is not benchmarked.

Work (60-03 / QUAL-02)

• Labelled fixture of 60–100 short documents in 6–8 known clusters.
• memory-topics::metrics::{purity, adjusted_rand_index} over cluster() output.
• Commit benchmarks/results/topics-quality.json.
• README topic row cites the artifact (ARI) or stays Works with the number.

Success

cargo test -p memory-topics metrics covers purity/ARI on a hand-computed 3-cluster example. topics-quality.json committed.

See also #40 (vector / QUAL-01).

[RichardHightower]

60-03 evidence is in #50.

benchmarks/results/topics-quality.json: purity 1.00 / ARI 1.00, 80 docs, 8 predicted clusters, 0 noise, embedding tfidf-l2-top32.

This is TopicExtractor::cluster on a synthetic labelled TF-IDF corpus, not Candle embeddings and not live TOC summaries. README topic row now cites the artifact.