Distance cues: level reverb sends and facing voices

[ctoth]

Two distance cues for positioned audio. One commit each, tests with each.

1. The reverb send no longer falls with distance

Cacophony takes a send from a voice's output, after its gain (routableSource.ts: outputNode.connect(sendGain), and outputNode is the gain node that follows the panner). The client applies a positioned sound's distance attenuation at that same gain, so the level reaching the room's reverb fell exactly as fast as the direct level: a far sound was quieter but no wetter than a near one.

The send is now given send × min(1 / distanceGain, 32), so the chain receives volume × send wherever the source is, and only the direct level falls with distance.

• Re-gained in place whenever the distance gain changes (listener moves, source moves, spatial profile changes) and whenever send or the chain changes. All of those already pass through applyLevels.
• The feed from a sound's inline effect chain into its named chain is downstream of the same gain and is compensated the same way.
• A change of send on a live aux send now re-gains it in the engine. Before, the send was removed and added again.
• Unchanged: the direct level, the distance model, occlusion (it sits before the tap and still dims direct and reverb alike), non-positional sounds (no distance gain, wire send), ambisonic sounds (no named chains).

The cap. MAX_SEND_DISTANCE_BOOST = 32 (+30 dB), in distanceModel.ts. The engine accepts any finite send gain and does not clamp it, and the distance curves give no usable limit: the default one floors at 1 / 10 000 and a linear one reaches 0, where the tap is silent and nothing can be recovered. So the cap is a choice: under the default curve the reverberant level is even out to 32 m, and beyond that it falls with the direct sound. It caps the boost, not the send, so that distance is the same whatever the send.

2. Voices have a facing, and their reverb send is level with distance too

• A positional voice takes its orientation from the speaking entity's forward, which Client.Spatial.EntityMove already delivers and tweens into the spatial store (in Web Audio axes), with a gentle cone: full level within 60° either side of the facing, falling evenly to −6 dB directly behind (VOICE_CONE_INNER_ANGLE 120, VOICE_CONE_OUTER_ANGLE 360, VOICE_CONE_OUTER_GAIN 10^(−6/20)).
• The facing follows the entity through the store subscription that already moves the voice. A turn ramps like a move.
• An entity with no forward is omnidirectional. A non-positional voice has no panner, so no cone and no distance, as before.
• The distance curve was already the media default: the voice panner took the same three constants DEFAULT_SPATIAL_PROFILE is built from. It is now derived from that profile by name (profilePannerDistance). Web Audio's inverse model is the profile's formula exactly, so the rolloff stays in the panner; there is no second gain stage.
• A voice's rolloff is inside its panner, ahead of the gain node the send is tapped from, so its send fell with distance as a sound's did. It now gets send / distanceGain with the same cap. A panner's gain cannot be read back, so the distance gain is computed from the profile, the voice's position and the engine's listener position.

Not done

• The cone (for voices and for media sounds) is also ahead of the tap, so a source facing away sends up to its cone attenuation less to the room. Not compensated.
• A voice's panner position is smoothed by the engine (30 ms time constant) and its send is re-gained at once, so the two disagree briefly while a voice or the listener is moving.
• No new GMCP message or field, no change to Cacophony, the ambisonic paths or the UI.

Existing tests changed

• Media.test.ts, "rebuilds on an Update to 3D, carrying the state the voice has now": the rebuilt 3D voice is 2 m away, so its send is now 0.6, not the wire 0.3. The non-positional case still asserts 0.3, and both assert the carried wire send.
• LiveKitSpatialAudioBridge.test.ts: the two attachVoice argument assertions gain the third argument (null, no facing), and the mock voice gains setFacing.

Verification

npm run typecheck, npm test (128 files, 1564 tests) and npm run test:audio-cache pass locally. Nothing was listened to: this is verified by unit tests against mocked engine objects and by reading the engine source, not by ear and not in a browser.