Improve song quality: backing humanization + Watts mode

[leoneperdigao]

Summary

Two rounds of quality work to make generated tracks sound less robotic.

1. Backing humanization (groove)

The procedural backing was too perfect — every bar a byte-identical repeat with flat random velocity and dead-static synthesis. This adds human imperfection (all seeded, so a given seed still reproduces exactly):

• Accents & groove — musical accent curve (down/back-beat emphasis, ghosted off-beats); correlated per-bar timing drift instead of white-noise jitter; bass sits slightly behind the beat.
• Variation — occasional dropped/added ghost hits, phrase-end drum fills, bass pickup notes (no two bars identical).
• Articulation — per-note ADSR jitter; a pool of drum one-shot variants (kills the machine-gun identical-sample effect); bass per-note detune, slow pitch drift, optional glide, velocity-tracked filter.
• Movement — pads get a real time-varying filter (fixes a bug that averaged the LFO into a static cutoff).

2. Watts mode (spoken word)

A contemplative mode pairing original Alan-Watts-style lyrics with a warm, measured delivery:

• Lyrics — new WattsLyricGenerator: original, non-verbatim lines (the talks are copyrighted) across six theme groups, assembled into flowing passages. Selectable via lyrics_style / --lyrics watts.
• Voice — new watts preset (no synthetic tremor, slow, spacious); deliberate_pacing makes the pipeline synthesize sentence-by-sentence so breaths/pauses fall naturally between thoughts. Configurable edge-tts voice/rate (en-GB-ThomasNeural, slower). Default headache vibrato dialed back to de-robotize all presets.
• Space — configurable vocal reverb room for a lecture-hall feel.
• Bundle — config/moods/watts.yaml ties it together (vocal-led mix over a gentle lo-fi bed). --mood watts just works; --lyrics/--voice also compose with other moods.

Design spec and implementation plan included under docs/superpowers/.

Note: this branch also carries the earlier release-packaging workflow already present on feature/release.

Test Plan

• pytest -m "not slow" — 193 passed (added tests for lyric/voice/mixing/config + backing determinism & bar variation)
• ruff check + ruff format --check clean
• --mood watts renders end-to-end: quality report 9 passed · 0 warnings · 0 failures
• Composition verified: --mood melancholic --lyrics watts --voice watts
• Existing moods unchanged
• Reviewer: listen to a --mood watts render for warmth, pacing, and hall space