cleanup(fflag): delete dead use_spectral_norm/use_residual in QNetworkConfig — [FFLAG-005]

QNetworkConfig had three write-only fields with zero workspace readers:
use_spectral_norm, spectral_norm_iterations, use_residual. Defaults set
at the single Default impl, never checked in forward/backward. Deleted
all three + their Default init. use_gpu kept — line 138 gates CUDA
construction and errors without it (genuine sanity check).

crates/ml-dqn/src/network.rs

@@ -78,14 +78,8 @@ pub struct QNetworkConfig {
     /// High dropout early prevents overfitting, low dropout late allows fine-tuning
     /// (Wave 26 P1.6)
     pub dropout_schedule: Option<(f64, f64, usize)>,
-    /// Whether to use `GPU` acceleration
+    /// Whether to use `GPU` acceleration (CUDA required — checked at construction).
     pub use_gpu: bool,
-    /// Whether to use spectral normalization for Q-value stability
-    pub use_spectral_norm: bool,
-    /// Number of power iterations for spectral norm estimation
-    pub spectral_norm_iterations: usize,
-    /// Whether to use residual connections (Wave 26 P0.4)
-    pub use_residual: bool,
 }
 
 impl Default for QNetworkConfig {
@@ -99,10 +93,6 @@ impl Default for QNetworkConfig {
             dropout_prob: 0.2,
             dropout_schedule: None, // No adaptive dropout by default
             use_gpu: true,
-            use_spectral_norm: false, // Disabled by default, enable for unstable training
-            spectral_norm_iterations: 1, // 1 iteration is typically sufficient
-            use_residual: false, // Disabled by default, opt-in for deeper networks (Wave 26 P0.4)
-
         }
     }
 }