fix(voice): gate the PTT/Sarvam send-paths on uploadStatus + extractionInFlight too

The voiceSubmitRef path (browser SpeechRecognition end-of-speech
auto-fire) was guarded in dfaa4d6 — but two OTHER voice send-paths
in the PTT recorder.onstop callback (L1289 srFallback + L1311 Sarvam
authoritative text) had no guard. If the user accidentally pressed
PTT during the upload + extraction window, or if a stray TTS
playback bled into the mic and the recorder picked it up, the
Sarvam transcription would still drop the resulting chat turn into
the middle of the wait — producing the "I can help with that!
Please tell me what you'd like to know" unprompted message reported
in screenshot #53.

Single early-return at the top of the PTT post-transcribe callback
now blocks ALL voice send-paths while uploadStatus || extractionInFlight.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

frontend/src/app/page.tsx

@@ -1283,6 +1283,18 @@ export default function Page() {
           // unknown). Prefer the SR fallback transcript when present so the
           // turn still goes through; otherwise surface the friendly
           // user_message and DO NOT call send() with empty text.
+          // ADR-044 — defensive: suppress voice-driven send() during the
+          // upload-index window AND while the background LLM extraction
+          // is still running. Even an explicit PTT press shouldn't drop
+          // a chat turn into the middle of the wait — the user can
+          // re-press once the card lands. Mirrors the voiceSubmitRef
+          // guard for the browser SpeechRecognition path.
+          const __voiceBlocked = uploadStatus || extractionInFlight;
+          if (__voiceBlocked) {
+            setInput("");
+            maybeResumeLive();
+            return;
+          }
           if (tr.error_code) {
             if (srFallback) {
               setInputFromTranscript(srFallback);