feat(sections): label by acoustic repetition, not fixed position

[seonghobae]

What

assign_section_labels named sections by fixed position — index 1 was always verse, index 2 always chorus, regardless of the audio. So a song that opens intro→chorus got the chorus mislabeled, and two acoustically identical segments could receive different labels.

Now labels come from actual acoustic repetition: reuse the chroma the boundary detector already computes, take each segment's mean chroma, greedily group segments whose cosine similarity clears a threshold (a repeated section), and name from repetition — most-repeated group → chorus, other repeats → verse, unique edges → intro/outro, unique middles → bridge.

Why

Section labels drive the rehearsal roadmap; position-based labels quietly mislabel any song whose form doesn't match the assumed intro/verse/chorus order. Repetition is the actual signal that distinguishes verse/chorus.

Scope / notes

• librosa-only, no new dependency. Boundary detection is unchanged (already real SSM/novelty).
• Positional labeling stays as a fallback when no repetition groups are supplied.
• New tests assert repeats get a consistent label (an A-B-A-B-A pattern the old positional labeler would have split) and that chroma grouping separates repeated vs distinct segments.
• 100% coverage on the module; full engine suite 439 passed; ruff + mypy clean.
• Still a heuristic (most-repeated = chorus); documented, with meter/ML as the upgrade path.

🤖 Generated with Claude Code