Let echoes ring on with a Delay node in sound FX #324 #327

A voice used to end the moment its sources did, and Stop sound cut it
dead. An echo has to outlive its sound, and a stop must not click.

- Delay node: a delay line with a damped feedback loop, added at mix
  next to the untouched sound. Time 0.01 to 2 s, feedback up to 0.95,
  damping 20 kHz to 500 Hz evenly in octaves, all fed by parameters
  and moved by Set sound FX parameters like any other number.
- Ping-pong is a setting of the node. It takes the sound into the left
  of the line and swaps the sides every round.
- A voice with an echo rings on for its tail: until the last echo is
  60 dB below the sound, from time, feedback and mix (fxTailOf). It
  stays reachable meanwhile, and a change counts the tail again.
- The end of a tail is timed on the audio clock with a silent
  ConstantSourceNode, so a paused run holds it. Then the voice is
  disconnected, since an echo is a loop of nodes.
- Stop sound fades a voice out in 10 ms instead of cutting it. Let
  ring fades only its sources and keeps the voice for its echoes.
  Where the context is suspended it cuts at once.
- Stop sound takes a voice. Connected, it stops only that play, and a
  0 from a Play sound that played nothing stops nothing.
- Play sound hands out tail: length plus the echoes at its start
  values.
- Destroying a behavior and ending a run still cut, echoes included.
- No limit on voices. The emit of a sound FX node now also gets the
  node's settings.

Finding: in Chrome every round of a feedback loop comes one render
quantum late (128 samples, 2.7 ms at 48 kHz), measured as 0, 128, 256,
384 for the first four echoes. Left so, since a browser without it
would then hear the echoes early. Written in the wiki.

Tests:
- sound-library, soundfx, node-catalog, sounds, keep-on-scene-load:
  489 tests green. New: fade on stop, cut while suspended, echo and
  ping-pong wiring, ringing until the tail, tail counted again, let
  ring, stop by voice, echo knobs, delay compile, tail sums.
- Counter-checks: no tail, a stop without fade, a connected voice 0
  read as unconnected. Each fails its tests, files back by cmp.
- e2e 34 (melektaus-atx): ALL PASSED, 37 checks. A blip through a
  ping-pong Delay (0.3 s, 0.5, 0.5) rang 2.97 s, 2.95 s worked out.
  Let ring kept the voice, a stop after it cut it. Rendered offline
  with the player's own stageNodes the echoes are 0.25 left, 0.125
  right, 0.063 left. 64 stereo voices of it took about 0.3 s for 2 s,
  so about 400 play at once.
- Gate (melektaus-atx): CHECK EXIT 0, 242 test files, 3713 tests.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/helpers.ts

@@ -272,7 +272,11 @@ export function fakeContext(overrides: Partial<NodeContext> = {}): FakeContext {
       ctx.played.push({ assetId, volume, loop, detached, fx, fxParams }),
     // Every sound lasts a quarter of a second, as if it had been loaded.
     soundLength: () => 0.25,
-    stopSound: (assetId) => ctx.silenced.push(assetId),
+    // The tail of a voice is its number in tenths of a second, so a test can tell one from another.
+    soundTail: (voice) => voice / 10,
+    // A voice by its number, and letting ring, are written into the line, so one list says it all.
+    stopSound: (assetId, letRing) => ctx.silenced.push(letRing ? `${assetId} (let ring)` : assetId),
+    stopVoice: (voice, letRing) => ctx.silenced.push(`voice ${voice}${letRing ? ' (let ring)' : ''}`),
     tuned: [],
     tuneSound: (voice, fxParams, time, ease) => ctx.tuned.push({ voice, fxParams, time, ease }),
     loadScene: (sceneId) => ctx.loaded.push(sceneId),