refactor(ml): extract shared items from dqn/ to top-level modules

Move mixed_precision, xavier_init, portfolio_tracker, action_space,
order_router from dqn/ to ml/src/ root. Extract TradingAction enum
from dqn/agent.rs to standalone trading_action module. Re-export
from dqn/mod.rs for backward compatibility.

These modules are shared infrastructure used by PPO, trainers,
hyperopt, and ensemble — not DQN-specific.

Test results: 2757 passed, 0 failed, 25 ignored (unchanged).

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

crates/ml/src/dqn/agent.rs

@@ -24,38 +24,7 @@ use super::network::{QNetwork, QNetworkConfig};
 use super::{Experience, ReplayBuffer, ReplayBufferConfig};
 use crate::MLError;
 
-/// Trading actions available to the `DQN` agent
-#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
-pub enum TradingAction {
-    /// Buy signal
-    Buy = 0,
-    /// Sell signal  
-    Sell = 1,
-    /// Hold/Do nothing
-    Hold = 2,
-}
-
-impl TradingAction {
-    /// Convert action to integer index
-    pub fn to_int(self) -> u8 {
-        self as u8
-    }
-
-    /// Convert integer index to action
-    pub fn from_int(val: u8) -> Option<Self> {
-        match val {
-            0 => Some(TradingAction::Buy),
-            1 => Some(TradingAction::Sell),
-            2 => Some(TradingAction::Hold),
-            _ => None,
-        }
-    }
-
-    /// Get all possible actions
-    pub fn all() -> [TradingAction; 3] {
-        [TradingAction::Buy, TradingAction::Sell, TradingAction::Hold]
-    }
-}
+pub use crate::trading_action::TradingAction;
 
 /// Trading state representation for `DQN`
 #[derive(Debug, Clone, Serialize, Deserialize)]

crates/ml/src/dqn/mod.rs

@@ -4,9 +4,15 @@
 //! with all 6 components: Double Q-learning, Dueling Networks, Prioritized Experience Replay,
 //! Multi-step Learning, Distributional RL (C51), and Noisy Networks.
 
+// Shared modules moved to crate root — re-export for backward compatibility
+pub use crate::action_space;
+pub use crate::order_router;
+pub use crate::mixed_precision;
+pub use crate::xavier_init;
+pub use crate::portfolio_tracker;
+pub use crate::trading_action::TradingAction;
+
 // Original DQN components
-pub mod action_space; // Factored action space (5 DQN exposure actions + deterministic order routing)
-pub mod order_router; // Deterministic order routing: exposure → (order_type, urgency) from microstructure
 pub mod count_bonus; // UCB count-based exploration bonus for action diversity
 pub mod agent;
 pub mod attention; // Multi-head self-attention for temporal pattern recognition (Wave 26 P1.2)
@@ -19,7 +25,6 @@ pub mod hindsight_replay; // Hindsight Experience Replay for 5-10x data efficien
 pub mod logging; // Optimized logging utilities for ML training (Wave 30)
 pub mod network;
 pub mod nstep_buffer; // N-step experience buffer for multi-step returns (Wave 2.2)
-pub mod portfolio_tracker;
 pub mod regime_conditional; // Regime-conditional Q-networks (3 heads: Trending, Ranging, Volatile)
 pub mod replay_buffer;
 pub mod residual; // Residual/skip connections for better gradient flow (Wave 26 P0.4)
@@ -29,7 +34,6 @@ pub mod logit_clipping; // Bug #12: Clip logits before softmax to prevent satura
 pub mod target_update; // Target network update strategies (Polyak averaging, hard updates)
 pub mod trade_executor; // Risk controls and execution simulation
 pub mod trainable_adapter; // UnifiedTrainable trait implementation
-pub mod xavier_init; // Xavier/Glorot weight initialization
 
 // Wave 16 Portfolio Features
 pub mod entropy_regularization; // Entropy-based policy diversity
@@ -72,7 +76,7 @@ pub mod performance_validation;
 // Re-export core DQN types for public usage
 pub use action_space::{ExposureLevel, FactoredAction, OrderType, Urgency};
 pub use order_router::OrderRouter;
-pub use agent::{AgentMetrics, DQNAgent, TradingAction, TradingState};
+pub use agent::{AgentMetrics, DQNAgent, TradingState};
 pub use dqn::{GradientResult, DQN, DQNConfig};
 pub use experience::{Experience, ExperienceBatch};
 pub use nstep_buffer::NStepBuffer;
@@ -112,9 +116,6 @@ pub use logit_clipping::{clip_logits, clip_logits_default, softmax_with_clipping
 // Re-export GAE components (Wave 26 P1.9)
 pub use gae::{GAECalculator, GAEConfig};
 
-// Wave 26 P2.1: Mixed precision (AMP) for 2x speedup
-pub mod mixed_precision;
-
 // Wave 26 P2.3: Ensemble Q-network for uncertainty estimation
 pub mod ensemble_network;
 

crates/ml/src/lib.rs

@@ -780,6 +780,14 @@ pub const PRECISION_FACTOR: i64 = 100_000_000;
 /// Maximum inference latency target in microseconds
 pub const MAX_INFERENCE_LATENCY_US: u64 = 100;
 
+// ========== SHARED INFRASTRUCTURE (extracted from dqn/) ==========
+pub mod action_space; // Factored action space (5 DQN exposure actions + deterministic order routing)
+pub mod mixed_precision; // AMP utilities for 2x speedup (FP16/BF16)
+pub mod order_router; // Deterministic order routing: exposure -> (order_type, urgency) from microstructure
+pub mod portfolio_tracker; // Portfolio state tracking for training
+pub mod trading_action; // TradingAction enum (Buy/Sell/Hold)
+pub mod xavier_init; // Xavier/Glorot weight initialization
+
 // ========== CORE ML MODULES ==========
 // Core ML modules
 pub mod backtesting; // Backtesting framework for barrier optimization

crates/ml/src/order_router.rs

@@ -5,7 +5,7 @@
 //! spread and volatility conditions. This separates the alpha model
 //! (what exposure?) from execution (how to route?).
 
-use super::action_space::{ExposureLevel, FactoredAction, OrderType, Urgency};
+use crate::action_space::{ExposureLevel, FactoredAction, OrderType, Urgency};
 
 /// Deterministic order routing based on market microstructure.
 /// DQN selects exposure level (0-4), OrderRouter selects order type + urgency.

crates/ml/src/portfolio_tracker.rs

@@ -8,8 +8,8 @@
 //! The PortfolioTracker is used by DQNTrainer to provide portfolio features
 //! to the reward function, enabling P&L-based reward calculations.
 
-use super::action_space::FactoredAction;
-use super::agent::TradingAction;
+use crate::action_space::FactoredAction;
+use crate::trading_action::TradingAction;
 use tracing::{debug, warn};
 
 /// Trade action enum with quantities for TradeExecutor

crates/ml/src/trading_action.rs

@@ -0,0 +1,36 @@
+//! Trading action enum shared across DQN, PPO, ensemble, and other ML modules.
+
+use serde::{Deserialize, Serialize};
+
+/// Trading actions available to the `DQN` agent
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+pub enum TradingAction {
+    /// Buy signal
+    Buy = 0,
+    /// Sell signal
+    Sell = 1,
+    /// Hold/Do nothing
+    Hold = 2,
+}
+
+impl TradingAction {
+    /// Convert action to integer index
+    pub fn to_int(self) -> u8 {
+        self as u8
+    }
+
+    /// Convert integer index to action
+    pub fn from_int(val: u8) -> Option<Self> {
+        match val {
+            0 => Some(TradingAction::Buy),
+            1 => Some(TradingAction::Sell),
+            2 => Some(TradingAction::Hold),
+            _ => None,
+        }
+    }
+
+    /// Get all possible actions
+    pub fn all() -> [TradingAction; 3] {
+        [TradingAction::Buy, TradingAction::Sell, TradingAction::Hold]
+    }
+}