feat: Implement score-aware time decay in gravity score calculation, allowing high-scoring items to decay slower.

Author: tosaki

ranking.py

@@ -2,15 +2,38 @@
 
 def calculate_gravity_score(base_score: float, published_at_ts: float, gravity: float = 1.8) -> float:
     """
-    Calculate the gravity score based on: Score / (Time + 2)^G
-    Time is in hours since publication.
+    Calculate the gravity score with score-aware time decay.
+    
+    High-scoring items decay slower than low-scoring items.
+    Formula: Score / (Time + base_offset + score_bonus)^effective_gravity
+    
+    - base_offset: 0.5 (reduced from 2 to make decay faster overall)
+    - score_bonus: 0 to 3 hours extra buffer based on score (0-100)
+    - effective_gravity: slightly reduced for high scores
+    
+    This ensures truly important news (high L2 score) stays relevant longer,
+    while regular items decay more quickly.
     """
     now = time.time()
 
     # Calculate age in hours
-    # Ensure age is at least 0 to avoid issues with future timestamps (clock skew)
     age_hours = max(0, (now - published_at_ts) / 3600)
 
-    # Standard HNews Gravity Formula
-    # score / (age_hours + 2) ^ gravity
-    return base_score / pow((age_hours + 2), gravity)
+    # Base offset reduced from 2 to 0.5 for faster overall decay
+    base_offset = 0.5
+    
+    # Score bonus: high scores get up to 3 hours of extra buffer
+    # Normalized assuming L2 scores range 0-100
+    # score 0 -> 0h bonus, score 50 -> 1.5h bonus, score 100 -> 3h bonus
+    score_bonus = (base_score / 100) * 3.0
+    
+    # Effective gravity: high scores decay slightly slower
+    # score 0 -> gravity, score 100 -> gravity * 0.85
+    gravity_reduction = (base_score / 100) * 0.15
+    effective_gravity = gravity - gravity_reduction
+    
+    # Combined offset
+    total_offset = base_offset + score_bonus
+    
+    # Final formula
+    return base_score / pow((age_hours + total_offset), effective_gravity)