feat: add len_var scorer (B-0)

metrics/len_var/README.md

@@ -0,0 +1,8 @@
+# Length-Variance (len_var) Scorer
+
+| Key | Value |
+|-----|-------|
+| **Purpose** | Detects internal length fluctuation in a single answer |
+| **Formula** | σ(window_lengths) / μ(window_lengths) |
+| **Threshold (tentative)** | <0.20 : ✅ / 0.20-0.39 : ⚠️ / ≥0.40 : 🔥 |
+| **Dependencies** | None (pure Python stdlib) |

metrics/len_var/scorer.py

@@ -0,0 +1,17 @@
+import statistics as st
+from typing import List
+
+def _window_lengths(tokens: List[str], window: int = 5) -> List[int]:
+    return [len(" ".join(tokens[i:i+window]))
+            for i in range(len(tokens) - window + 1)]
+
+def score(prediction: str, reference: str = "", window: int = 5) -> float:
+    """
+    Length-variance score (0 = 安定, 1 ≒ 大揺れ)
+    定義: sliding-window の文字長 標準偏差 ÷ 平均長
+    """
+    toks = prediction.split()
+    if len(toks) < window:
+        return 0.0
+    lens = _window_lengths(toks, window)
+    return st.pstdev(lens) / max(st.mean(lens), 1)

metrics/len_var/tests/test_len_var.py

@@ -0,0 +1,9 @@
+from metrics.len_var.scorer import score
+
+def test_constant_length():
+    s = "word " * 10
+    assert score(s) < 0.05
+
+def test_varied_length():
+    s = "a a a a a " + "this sentence is considerably longer "
+    assert score(s) > 0.2