feat(phase-60): vector and topic quality fixtures (60-03)

[RichardHightower]

Summary

Phase 60-03 of v3.2 Prove It (#46). Stacked on #49 (60-01).

A paraphrase set BM25 cannot solve, a --layers bm25|vector|hybrid switch on the custom harness, and a labelled TopicExtractor::cluster artifact behind the README status table.

Honest caveats (do not skip these):

• Mock BM25 is token overlap.
• Mock vector is a committed paraphrase lexicon + TF-IDF cosine, not Candle/HNSW.
• Mock hybrid is RRF (k=60) of those two lists, not live RouteQuery.
• --layers is mock-only; CLI memory search is always hybrid.
• memory-bench all excludes semantic so custom-harness-mock.json is not tanked.
• Topic purity/ARI is a synthetic 80-doc TF-IDF corpus, not live TOC / Candle embeddings.

Committed numbers

Artifact	layers	recall@5	pass
benchmarks/results/semantic-bm25.json	bm25	0.00	0/16
benchmarks/results/semantic-vector.json	vector	1.00	16/16
benchmarks/results/semantic-hybrid.json	hybrid	1.00	16/16
benchmarks/results/topics-quality.json	—	purity 1.00 / ARI 1.00	80 docs, 8 clusters, 0 noise

Closes BENCH-13 / QUAL-01 / QUAL-02 / QUAL-03. Evidence for #40 and #47.

Does not start 60-02 (#39 — maintainer-owned).

Acceptance

• ≥15 semantic tests; hit files share no query tokens (semantic_hits_do_not_contain_query_tokens)
• BM25 recall@5 < 0.4 on the paraphrase set (committed 0.00)
• hybrid recall@5 > bm25 recall@5 (1.00 > 0.00)
• --layers on semantic / run / all
• topics-quality.json from TopicExtractor::cluster
• cargo test -p memory-topics --lib includes hand-computed purity 8/9 and ARI 4.5/7
• README vector/topic rows cite the artifacts; positioning Claims Ledger added

Test plan

• cargo test -p memory-bench --lib (55 passed, including QUAL-01 leak + BM25<0.4 tests)
• cargo test -p memory-topics --lib (167 passed, including metrics + quality)
• cargo clippy -p memory-bench --all-targets -- -D warnings
• cargo clippy -p memory-topics --lib -- -D warnings
• rustfmt --check on the touched crates

Full task pr-precheck is workspace-wide; same targeted bar as #49.