fix(web): reduce FOT by stripping redundant voice data from transport

Three transport-only optimizations that cut per-session Vercel FOT by ~50-60%:

P0 — Server strips voice.referenceAudioBase64 from already-known characters
in /api/scene and /api/insert-beat responses (defense-in-depth).

P1 — Client strips all voice data from session before sending to
/api/scene, /api/vision, and /api/insert-beat. Voices are retained locally
and re-merged from responses via mergeCharactersPreserveVoice(). The engine
only needs character names + visualDescriptions for scene generation.

P3 — /api/beat-audio returns binary audio (Response with Content-Type)
instead of JSON-wrapped base64, saving ~33% encoding overhead. Client
converts to blob URLs; PlayCanvas accepts a single audioSrc prop.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

app/api/beat-audio/route.ts

@@ -26,7 +26,11 @@ export async function POST(req: Request) {
   try {
     const config = loadEngineConfig(req.headers);
     const result = await requestBeatAudio(config, body);
-    return NextResponse.json(result);
+    if (!result.audio) return new Response(null, { status: 204 });
+    const binary = Buffer.from(result.audio.base64, "base64");
+    return new Response(binary, {
+      headers: { "Content-Type": result.audio.mime },
+    });
   } catch (err) {
     // Engine already swallows synth errors and returns audio:null. Anything
     // that reaches here is config-level — surface so the client can log it.