Music in sections and layers, switched on bar boundaries and mixed per listener.

A game declares a piece — its tempo, its layer names, and for each section what every layer plays — and then cues it from its own state. The names are the game's; nothing here knows what a section or a layer is for.

Cauldron2D.Audio.Music.piece(audio, "orbit", %{
  bpm: 112,
  beats_per_bar: 4,
  layers: [:drums, :bass, :pad],
  sections: %{
    calm: %{drums: calm_drums, bass: calm_bass, pad: pad},
    tense: %{drums: tense_drums, bass: tense_bass, pad: pad}
  },
  fingerprint: "3"
})

Cauldron2D.Audio.Music.cue(audio, %{piece: "orbit", section: :tense, layers: %{drums: 1.0, bass: 1.0, pad: 0.4}})

What a layer plays is either performed live or rendered once:

• a Strudel chain as a string — ~s|s("bd*4").bank("RolandTR909")| — is performed live on the listener's stage. A section's chains play as one stack, each layer's level applied to its chain as its velocity — multiplying whatever gain the chain sets, never replacing it — and the music level to the whole, sounding notes included, at the piece's bpm / beats_per_bar / 60 cycles per second. Nothing is rendered ahead, so a cue takes effect at once: a new section at the next cycle line, new gains on the next block. Every layer of a section must be a chain for the section to play this way
• a TuningFork.Score, a zero-arity function returning PCM, or PCM as {:pcm, binary} is rendered once, two stems at a time in tasks, and kept through TuningFork.Cache in the game's music directory under piece/section/layer and a fingerprint of the piece's own, what was written, the tempo, the rate and channels — so a changed layer is rendered again. A cue moves its section to the front of what is still to render and takes effect when the section is ready — at the next bar boundary of the playing section, with a cross-fade; new gains at once

The music level from Cauldron2D.Audio.levels/3 scales either.

A director made static (Cauldron2D.Audio started with static_music: true) plays the first section it is cued and no other: the section's chains are rendered once, as one pattern, over the piece's :loop cycles (default 8) — whatever still sounds at the end folded back over the start, so it loops without a seam — kept through TuningFork.Cache under piece/section/static and the fingerprint, and looped as a bed. Every later cue is ignored until stop/1. Rendering costs once a section and rate on a node; playing it back costs a copy a chunk.