fix(audio): preserve audioAsVoice propagation

- Keep audioAsVoice-only payloads from being filtered out
- Allow empty payloads through when they carry the flag
- Remove temporary debug logs around audioAsVoice buffering

Co-authored-by: Manuel Hettich <17690367+ManuelHettich@users.noreply.github.com>

src/agents/pi-embedded-subscribe.ts

@@ -437,7 +437,8 @@ export function subscribeEmbeddedPiSession(params: {
     lastBlockReplyText = chunk;
     assistantTexts.push(chunk);
     if (!params.onBlockReply) return;
-    const { text: cleanedText, mediaUrls, audioAsVoice } = splitMediaFromOutput(chunk);
+    const splitResult = splitMediaFromOutput(chunk);
+    const { text: cleanedText, mediaUrls, audioAsVoice } = splitResult;
     // Skip empty payloads, but always emit if audioAsVoice is set (to propagate the flag)
     if (!cleanedText && (!mediaUrls || mediaUrls.length === 0) && !audioAsVoice) return;
     void params.onBlockReply({

src/telegram/bot.ts

@@ -1388,7 +1388,6 @@ async function deliverReplies(params: {
         });
       } else if (kind === "audio") {
         const useVoice = reply.audioAsVoice === true; // default false (backward compatible)
-        log.warn(`[DEBUG] Audio media: audioAsVoice=${reply.audioAsVoice}, useVoice=${useVoice}`);
         if (useVoice) {
           // Voice message - displays as round playable bubble (opt-in via [[audio_as_voice]])
           await bot.api.sendVoice(chatId, file, {