perf: also warm STT during the popup so the first hold-to-ask is fast (Whisper is CPU, no quota)

frontend/app.js

@@ -144,7 +144,9 @@ async function warmup() {
     if (!client) { warmedSession = false; return; }   // let it retry if the connection failed
     const c = document.createElement("canvas"); c.width = 32; c.height = 32;
     const blob = await new Promise((r) => c.toBlob(r, "image/jpeg", 0.6));
-    if (blob) client.predict("/describe", { image: handle_file(blob), lang }).catch(() => {});  // load the model on the GPU; discard the result
+    if (blob) client.predict("/describe", { image: handle_file(blob), lang }).catch(() => {});  // warm the VLM on the GPU; discard the result
+    // also warm speech-to-text (Whisper, CPU) so the first "hold to ask" is fast too
+    try { const wav = await (await fetch(SILENT)).blob(); client.predict("/detect_lang", { audio: handle_file(wav) }).catch(() => {}); } catch (e) {}
   } catch (e) { warmedSession = false; }
 }