test(magnitude): 3 unit tests — epsilon parity, counterfactual weight, micro-reward scaling

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/ml/src/trainers/dqn/smoke_tests/reward_v8.rs

@@ -170,4 +170,38 @@ mod tests {
                 "Flat × mag={mag} must be zero");
         }
     }
+
+    #[test]
+    fn test_magnitude_epsilon_equals_direction() {
+        let eps_start = 0.20_f64;
+        let eps_end = 0.02;
+        let eps_old_mag = eps_end + 0.1 * (eps_start - eps_end);
+        let eps_new_mag = eps_start;
+        assert!(eps_new_mag > eps_old_mag * 4.0,
+            "New magnitude epsilon ({eps_new_mag}) should be >4× old ({eps_old_mag})");
+        assert!((eps_new_mag - eps_start).abs() < 1e-10,
+            "Magnitude epsilon should equal direction epsilon");
+    }
+
+    #[test]
+    fn test_counterfactual_weight() {
+        let cf_weight = 0.3_f64;
+        let flat_fraction = 0.58;
+        let effective = cf_weight * flat_fraction + 1.0 * (1.0 - flat_fraction);
+        assert!(effective > 0.5, "Effective gradient should be >50%: {effective}");
+        assert!(effective < 1.0, "Counterfactual must reduce magnitude gradient: {effective}");
+        let amp = 1.0 / effective;
+        assert!((amp - 1.68).abs() < 0.1, "Expected ~1.68× amplification, got {amp}");
+    }
+
+    #[test]
+    fn test_magnitude_micro_reward_scaling() {
+        let dir_reward = 0.01_f64;
+        let adaptive_scale = 1.0;
+        let micro_vol = 0.005;
+        let small_reward = adaptive_scale * dir_reward * 0.25 / micro_vol;
+        let full_reward = adaptive_scale * dir_reward * 1.00 / micro_vol;
+        assert!((full_reward / small_reward - 4.0).abs() < 1e-10,
+            "Full/Small reward ratio should be 4.0, got {}", full_reward / small_reward);
+    }
 }