Phase 5: optional [graph] analysis layer + core scalars

[zolizoli]

Phase 5 — analysis

The optional analysis layer that produces real encoding channels, plus the dependency-free core scalars. Stacked on #4. The dependency arrow points inward: the core never imports analyze.

Core (no extra)

• lexograph.scalars — lengths, positions, frequencies: the built-in channel sources, so a walk/spiral/dotplot works with nothing installed beyond the core.

lexograph[graph] extra

• analyze/embeddings.py — embed_sentences (sentence-transformers, L2-normalised). The heavy import is deferred to call time, so importing analyze is cheap and the whole pipeline is testable with injected vectors.
• analyze/graph.py — knn_graph (cosine kNN), embedding_distances (→ semantic recurrence dotplot), pagerank_scores (size), community_labels (colour; Louvain default, KMeans option).
• analyze/backbone.py — disparity-filter backbone, ported from sibling kenon (Serrano et al. 2009).
• analyze_text → Analysis — aligned size / community / distances arrays from a text. embeddings are injectable.

a = analyze_text(load_demo_text())               # PageRank + Louvain + distances
text_walk(load_demo_text(), colour=a.community, colour_kind="categorical", size=a.size)
recurrence_plot(load_demo_text(), distances=a.distances, threshold=0.35)

Verification

• make ci green: ruff + ty + 180 tests / 1 skipped. The graph pipeline (kNN, PageRank, Louvain/KMeans, disparity backbone, analyze_text) is covered with synthetic embeddings — no model download. The real embedder is an opt-in smoke test (LEXOGRAPH_RUN_MODEL=1).
• mkdocs build --strict green — mkdocstrings/griffe static-analyses analyze without importing the heavy deps (same pattern the family's RTD uses, which installs only the docs extra).
• pyproject gains the [graph] extra; uv.lock regenerated.

🤖 Generated with Claude Code