feat: adapted rank norm (skip zeros) + remove Q-drift penalty + aligned Sharpe

Rank normalization: holding bars (reward≈0) passthrough as zero,
only actual trade P&L is ranked. Preserves magnitude for trades.
Q-drift: hardcoded penalty removed, E1 enrichment handles adaptively.
Metrics: sharpe_raw (un-annualized per-trade) for direct comparison
between training and validation.

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

crates/ml/src/cuda_pipeline/c51_grad_kernel.cu

@@ -133,12 +133,10 @@ extern "C" __global__ void c51_grad_kernel(
         branch_base += batch_size * A_d * num_atoms;
     }
 
-    /* Q-mean drift regularization: ADAPTIVE quadratic penalty.
-     * Small drift → tiny penalty (don't interfere). Large drift → hard correction.
-     * lambda=0.1 (scaled up from 0.01 for trade-level reward's smaller Q-values). */
-    float q_mean_ema = q_mean_ema_ptr[0];
-    float drift_penalty = 0.1f * q_mean_ema * fabsf(q_mean_ema);
-    d_val_sum += drift_penalty;
+    /* Q-mean drift: handled adaptively by E1 enrichment (Q-value reality check).
+     * Hardcoded penalty removed — E1 computes bias correction from actual eval
+     * performance and adapts to changing reward scales automatically. */
+    (void)q_mean_ema_ptr; /* parameter kept for ABI compatibility */
 
     /* Single deterministic write — no atomicAdd */
     d_value_logits[b * num_atoms + j] = d_val_sum;

crates/ml/src/cuda_pipeline/reward_shaping_kernel.cu

@@ -34,6 +34,10 @@ extern "C" __global__ void reward_rank_normalize(
      * so high-return trades during volatile periods get lower
      * rank weight than the same return during calm periods. */
     float r_i = (std_ema > 1e-6f) ? (raw_r - return_mean_ema) / std_ema : raw_r;
+
+    /* Skip near-zero rewards (holding bars). Only rank actual trade P&L. */
+    int is_trade = (fabsf(r_i) > 0.001f) ? 1 : 0;
+
     float abs_i = fabsf(r_i);
 
     __shared__ float tile[TILE_SIZE];
@@ -65,6 +69,13 @@ extern "C" __global__ void reward_rank_normalize(
     /* Rank in [0, 1], sign of Sharpe contribution preserved */
     float rank = (float)count / (float)N;
     float sign = (r_i > 0.0f) ? 1.0f : (r_i < 0.0f) ? -1.0f : 0.0f;
+
+    /* Holding bars: passthrough zero, don't rank */
+    if (!is_trade) {
+        rewards_out[i] = 0.0f;
+        return;
+    }
+
     /* G13: input is already in Sharpe units — no rescale by std_ema needed */
     rewards_out[i] = sign * rank;
 }

crates/ml/src/trainers/dqn/financials.rs

@@ -7,6 +7,7 @@ use crate::cuda_pipeline::gpu_experience_collector::TradeStats;
 #[derive(Debug, Clone, Default)]
 pub(crate) struct EpochFinancials {
     pub sharpe: f64,
+    pub sharpe_raw: f64,
     pub sortino: f64,
     pub win_rate: f64,
     pub max_drawdown: f64,
@@ -91,7 +92,7 @@ pub(crate) fn compute_epoch_financials(
     };
     let trade_annualization = trades_per_year.max(1.0).sqrt();
 
-    let (sharpe, sortino) = if total_trades > 1 {
+    let (sharpe, sharpe_raw, sortino) = if total_trades > 1 {
         // Per-trade mean and variance from GPU-aggregated sums
         let mean = trade_stats.sum_returns / n_trades;
         let var = (trade_stats.sum_sq_returns / n_trades) - mean * mean;
@@ -103,6 +104,8 @@ pub(crate) fn compute_epoch_financials(
             0.0
         };
 
+        let sharpe_raw = if std > 1e-10 { mean / std } else { 0.0 };
+
         // Sortino: downside deviation from losing trades only.
         // sum_losses is sum of |returns| for losers. We need sum of squared
         // returns for losers — approximate from loss_count * (avg_loss)^2.
@@ -120,9 +123,9 @@ pub(crate) fn compute_epoch_financials(
             0.0
         };
 
-        (s, sort)
+        (s, sharpe_raw, sort)
     } else {
-        (0.0, 0.0)
+        (0.0, 0.0, 0.0)
     };
 
     // MaxDD from step_returns equity curve over last 10K bars, capped at 100%.
@@ -182,6 +185,7 @@ pub(crate) fn compute_epoch_financials(
 
     EpochFinancials {
         sharpe,
+        sharpe_raw,
         sortino,
         win_rate,
         max_drawdown: max_dd,

crates/ml/src/trainers/dqn/trainer/training_loop.rs

@@ -2017,6 +2017,8 @@ impl DQNTrainer {
             "Epoch {}/{}: val_Sharpe={:.2} (deterministic backtest on fixed validation window)",
             epoch + 1, self.hyperparams.epochs, val_sharpe,
         );
+        let val_sharpe_raw = val_sharpe / (self.hyperparams.bars_per_day * 252.0).sqrt();
+        info!("  val_Sharpe_raw={:.6} (un-annualized per-bar)", val_sharpe_raw);
 
         // Prometheus metrics
         training_metrics::set_epoch("dqn", "current", (epoch + 1) as f64);
@@ -2196,9 +2198,10 @@ impl DQNTrainer {
                 financials.buy_pct, financials.sell_pct, financials.hold_pct,
             );
             info!(
-                "Epoch {}/{}: Sharpe={:.2} WinRate={:.1}% MaxDD={:.3}% PF={:.2} Return={:+.2}% Trades={}",
+                "Epoch {}/{}: Sharpe={:.2} Sharpe_raw={:.6} WinRate={:.1}% MaxDD={:.3}% PF={:.2} Return={:+.2}% Trades={}",
                 epoch + 1, self.hyperparams.epochs,
-                financials.sharpe, financials.win_rate * 100.0,
+                financials.sharpe, financials.sharpe_raw,
+                financials.win_rate * 100.0,
                 financials.max_drawdown * 100.0, financials.profit_factor,
                 financials.total_return * 100.0, financials.total_trades,
             );