fix(ml): correct sign error in continuous policy log_probs

The tensor-based log_probs() had an inverted sign on the squared
difference term: .sub(&(x * -0.5)) = +0.5*x² instead of -0.5*x².
This caused actions far from the mean to get higher log probabilities.

Also fix test assertions: continuous log probability densities CAN
be positive (when σ is small and action is near mean), unlike
discrete log probabilities which are always <= 0.

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

crates/ml/src/ppo/continuous_policy.rs

@@ -330,7 +330,7 @@ impl ContinuousPolicyNetwork {
 
         let log_prob = (log_2pi_tensor * (-0.5))?
             .sub(&log_stds)?
-            .sub(&(squared_diff * (-0.5))?)?;
+            .sub(&(squared_diff * 0.5)?)?;
 
         Ok(log_prob.squeeze(1)?) // Remove extra dimension if present
     }
@@ -567,9 +567,8 @@ mod tests {
         // Action should be in bounds
         assert!(action >= 0.0 && action <= 1.0);
 
-        // Log prob should be finite and negative
+        // Log prob should be finite (can be positive for continuous distributions with small σ)
         assert!(log_prob.is_finite());
-        assert!(log_prob <= 0.0);
     }
 
     #[test]
@@ -592,7 +591,7 @@ mod tests {
         assert_eq!(log_probs.dims(), &[2]);
 
         let log_probs_vec = log_probs.to_vec1::<f32>().unwrap();
-        assert!(log_probs_vec.iter().all(|&lp| lp.is_finite() && lp <= 0.0));
+        assert!(log_probs_vec.iter().all(|&lp| lp.is_finite()));
     }
 
     #[test]