Fix v2 output quality: stats format, Reddit results, citations, summary structure

- Stats: replace BAD/GOOD examples with strict fill-in-the-blank template
- Reddit: add subreddit-targeted fallback search, soften scoring penalties
  (engagement -10→-3, date confidence -10→-5), add minimum result guarantee
- Citations: limit to 1 per insight, short format, no engagement metrics
- Summary: add bold topic headers template for structured paragraphs

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

scripts/lib/score.py

@@ -21,7 +21,7 @@ WEBSEARCH_NO_DATE_PENALTY = 20  # Heavy penalty for no date signals (low confide
 
 # Default engagement score for unknown
 DEFAULT_ENGAGEMENT = 35
-UNKNOWN_ENGAGEMENT_PENALTY = 10
+UNKNOWN_ENGAGEMENT_PENALTY = 3
 
 
 def log1p_safe(x: Optional[int]) -> float:
@@ -152,9 +152,9 @@ def score_reddit_items(items: List[schema.RedditItem]) -> List[schema.RedditItem
 
         # Apply penalty for low date confidence
         if item.date_confidence == "low":
-            overall -= 10
-        elif item.date_confidence == "med":
             overall -= 5
+        elif item.date_confidence == "med":
+            overall -= 2
 
         item.score = max(0, min(100, int(overall)))
 
@@ -212,9 +212,9 @@ def score_x_items(items: List[schema.XItem]) -> List[schema.XItem]:
 
         # Apply penalty for low date confidence
         if item.date_confidence == "low":
-            overall -= 10
-        elif item.date_confidence == "med":
             overall -= 5
+        elif item.date_confidence == "med":
+            overall -= 2
 
         item.score = max(0, min(100, int(overall)))