feat(stt/voice): better recognition, sustained barge-in, fragment discard

STT (whisper_worker): add VAD tuning (speech_pad_ms so soft first/last words
aren't clipped), condition_on_previous_text=False + temperature fallback +
no_speech/logprob/compression thresholds to reject noisy/quiet decodes, drop
per-segment non-speech, and a hallucination guard that blanks Whisper's classic
Korean silence/noise boilerplate ("감사합니다" 등) when no_speech_prob is high.

Barge-in (bot.mjs): stop the bot's TTS only when the speaking (green ring) stays
on for >= WSAI_BARGE_IN_MS (default 700ms), not on the first blip — cancelled if
speaking stops in time. AfterSilence 800->1000ms so trailing soft words finish.

Fragment gate (dashboard): a lone connective filler ("그러면") carries no
answerable intent -> discard as [대기] instead of replying; hidden by the noise
filter like [잡음].

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

wsai/dashboard.py

@@ -49,6 +49,20 @@ def _thought_summary(reply: str) -> str:
     return "감정 태그 없음 · 기본 톤으로 답변"
 
 
+# Lone connective/fillers that clearly expect more to come ("그러면…"). On their
+# own they carry no answerable intent, so we discard them instead of blurting a
+# confused reply. (A fuller version would buffer and wait for the continuation.)
+_FRAGMENT_FILLERS = {
+    "그러면", "그래서", "그런데", "근데", "그리고", "그러니까", "그럼", "그게",
+    "저기", "있잖아", "그", "음", "어", "저", "그러", "그니까",
+}
+
+
+def _is_fragment(text: str) -> bool:
+    t = (text or "").strip().rstrip(" .,!?…~").strip()
+    return t in _FRAGMENT_FILLERS
+
+
 def _make_handler(dash: "Dashboard"):
     monitor = dash.monitor
 
@@ -722,6 +736,13 @@ class Dashboard:
                 turn.replied("[잡음]")
                 turn.finish()
                 return {"heard": heard, "reply": "[잡음]", "wav": b""}
+            if _is_fragment(heard):
+                # A lone connective ("그러면") — no answerable intent yet. Discard
+                # rather than reply to a fragment.
+                turn.thought("불완전한 말(연결어) — 대답하지 않고 폐기")
+                turn.replied("[대기]")
+                turn.finish()
+                return {"heard": heard, "reply": "[대기]", "wav": b""}
             t_llm = time.monotonic()
             reply_text = self._think(heard)
             self._step(turn, "LLM" if self.brain else "echo", (time.monotonic() - t_llm) * 1000)
@@ -1305,7 +1326,7 @@ function turnFilter(){
     text:$('tfText').value.trim().toLowerCase() };
 }
 function isNoiseTurn(t){
-  return (t.heard==='(빈 결과)') || (t.reply==='[잡음]');
+  return (t.heard==='(빈 결과)') || (t.reply==='[잡음]') || (t.reply==='[대기]');
 }
 function turnMatches(t, f){
   if(hideNoise && isNoiseTurn(t)) return false;   // 잡음 로그 표시 안 함 토글