fix: raise garbage-transcript threshold to avoid false positives on plate reads

Phonetic run threshold 5 → 12: a plate spellout ("Foxtrot Alpha Uniform Lima
Kilo...") produces 6–8 consecutive phonetic words, triggering false positives
and blocking intelligence extraction on legitimate calls. 12 is safely above
any real spellout (~8 max) while still catching the full-alphabet hallucination
(26 words). Also writes skip_reason="garbage_transcript" to the call doc and
surfaces it in the admin correlation debug endpoint.

drb-c2-core/app/internal/intelligence.py

@@ -94,11 +94,13 @@ def _is_garbage_transcript(transcript: str) -> bool:
     if not words:
         return False
 
+    # Threshold of 12: well above any legitimate plate/name spellout (~6–8 words)
+    # but catches the full-alphabet hallucination (26 words in sequence).
     run = 0
     for w in words:
         if w in _PHONETIC_ALPHA_WORDS:
             run += 1
-            if run >= 5:
+            if run >= 12:
                 return True
         else:
             run = 0
@@ -154,6 +156,10 @@ async def extract_scenes(
             f"Intelligence: call {call_id} — garbage transcript detected "
             f"(Whisper hallucination), skipping extraction"
         )
+        try:
+            await fstore.doc_set("calls", call_id, {"skip_reason": "garbage_transcript"})
+        except Exception:
+            pass
         return []
 
     raw_scenes: list[dict] = await asyncio.to_thread(