Sound

Everything First Light is heard through. One instance, on the scene's sound entity.

It builds its emitters under a single sound_root child and rebuilds them whole if a scene rebuild takes them away. It makes no gameplay decision and writes nothing into FL.state: it listens to the bus and it sounds.

What is where

The three fires and the well are spatial, so where a fire is and how it is doing are things the player hears from the direction of the altar. A fire's source volume follows FL.state.fires[id].heat and drops to almost nothing when it goes out; its pitch rises with the heat.

The bus it answers

say · ui.action · ui.focus · tend · flare · fire.died · fire.relit · scorch · stagger · daeva.spawned · daeva.banished · dawn · result · reset, plus FL.state.phase, which it reads every frame rather than taking from the phase event — so a hot reload lands on whatever the night is already doing.

ui.focus is offered rather than required: a screen that moves a controller focus can FL.emit("ui.focus") and get the quiet tick, and nothing breaks if nobody ever does.

The Old Keeper

He is the only voice. The WORDS are FL.lines, written and keyed by the loop slice, which fires them at their moments with FL.speak(key). The READINGS are .soundClip assets under game/audio/clips/, one per line, named vo_<key> — eleven of them, generated once through audio_gen.speech by game.audio.bake and cached, so a hot reload costs nothing.

game.audio.keeper holds the mapping and a verbatim snapshot of what each cached reading says. Keeper.audit() compares that snapshot against FL.lines and reports which readings are fresh, which are of older words, and which are missing — so a line rewritten after its clip was made shows up as a mismatch instead of playing the wrong reading under the new caption.

A line with no cached reading is captioned but not spoken, and the component logs which one. The bed still ducks for as long as the caption holds, so a missing reading costs the delivery and not the line.

The mix

Channel levels are set at build: ambient 0.75, music 0.55, sfx 0.78, ui 0.55, voice 1.0, and master 0.72 for headroom. The bed and the fires duck to 38% while the Keeper speaks and come back over about half a second, and the whole mix sits at 40% while the game is paused.

Measured against the master meter with audio.peakSince: the night itself runs about 0.20 at the median, the Keeper's whole eleven-line script peaked at 0.80 with the bed and score under it, and nothing reached 1.0. A stagger over a full mix reached 1.16 before the master came down — which is what the 0.72 is for.

Two things that go quiet without an error

A source told to autoplay can come up neverStarted or noDevice — a browser will not start audio before the page has been touched — so the bed, the well and the three fires are asked twice a second whether they are sounding and started again if they are not.

The event bus lives on the game.FL module table, and editing that file hot-reloads it into a NEW table with a NEW bus: every subscription taken on the old one is left listening to something nobody emits on, and all of this component's cues go silent with nothing in the log. The module is therefore compared rather than remembered — twice a second — and the subscriptions are retaken when it changes.

Re-baking

Nothing here creates a clip: asset.create on the gameplay path files into the ephemeral runtime store, which would re-buy the Keeper's voice on every load. The library is made once from a tool —

local B = require("game.audio.bake")
B.synth()              -- free, deterministic, always works
B.speech()             -- 2 credits a line, skips what is already on disk
B.effects({ "flare" }) -- always spends; name what to buy