Fix ranking_ci zero-division when comparing a feature to itself

Two fixes from dependency audit:

1. _inference.py: guard in ranking_ci() when both arguments name the same
   feature. diff==0 and SE==0 (rho_a - rho_b is the zero vector), so
   0/0 produced z_stat=inf. Return z_stat=0.0, p_value=1.0 in this case —
   the null of equal importance trivially holds when comparing a feature
   against itself.

2. test_inference.py: test_wrong_feature_names_length_raises was matching
   against "n_features" which no longer appears in the ValueError message.
   Updated regex to match the actual message "len(feature_names)=...".

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

Author: burningcost

src/shap_relativities/_inference.py

@@ -579,10 +579,15 @@ def ranking_ci(self, feature_a: str, feature_b: str) -> dict[str, float]:
         se_diff = float(np.sqrt(max(var_diff, 0.0)))
 
         diff = theta_a - theta_b
-        z_stat = diff / se_diff if se_diff > 0 else float("inf")
-
-        # One-sided p-value for H1: theta_a > theta_b
-        p_value = float(1.0 - stats.norm.cdf(z_stat))
+        # Guard: when both feature arguments are identical, diff==0 and SE==0.
+        # 0/0 is indeterminate; the correct result is z_stat=0, p_value=1.
+        if diff == 0.0 and se_diff == 0.0:
+            z_stat = 0.0
+            p_value = 1.0
+        else:
+            z_stat = diff / se_diff if se_diff > 0 else float("inf")
+            # One-sided p-value for H1: theta_a > theta_b
+            p_value = float(1.0 - stats.norm.cdf(z_stat))
 
         # Two-sided CI on the difference
         z_ci = float(stats.norm.ppf((1 + self.ci_level) / 2))

tests/test_inference.py

@@ -336,7 +336,7 @@ def test_mismatched_shapes_raises(self):
             SHAPInference(np.ones((100, 2)), np.ones(99), ["a", "b"])
 
     def test_wrong_feature_names_length_raises(self):
-        with pytest.raises(ValueError, match="n_features"):
+        with pytest.raises(ValueError, match="len\(feature_names\)"):
             SHAPInference(np.ones((100, 2)), np.ones(100), ["a"])
 
     def test_1d_shap_raises(self):