fix: min_hold_bars 10→50 + adaptive hold scales with base

min_hold=10 allowed ~1200 trades/day — costs destroyed the edge.
min_hold=50 (~6.5 min) limits to ~230 trades/day max.

Adaptive extension now scales 0-2× base (was hardcoded 0-8 bars).
With base=50: range is 50-150 bars depending on ISV stability.
Losing positions (negative reward_ema) get base only (50 bars).

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

config/training/dqn-production.toml

@@ -67,7 +67,7 @@ min_epochs_before_stopping = 80
 gpu_n_episodes = 4096
 initial_capital = 35000.0
 tx_cost_multiplier = 0.18  # bps: IBKR ES RT = $4.50/contract ($0.85 comm + $1.38 exch + $0.02 reg × 2 sides). At ES=$5100: 0.18 bps × 5100 × 0.0001 = 0.09 pts = $4.50
-min_hold_bars = 10  # ~26s hold at 23 volume bars/min
+min_hold_bars = 50  # ~6.5 min at ~11.5K imbalance bars/day. ISV extends to 50-150 bars adaptively.
 
 [experience.fill_simulation]
 ioc_fill_prob = 0.85

crates/ml/src/cuda_pipeline/trade_physics.cuh

@@ -157,8 +157,9 @@ __device__ __forceinline__ float enforce_hold(
     if (isv_signals != NULL) {
         float stability = 1.0f / (1.0f + isv_signals[1]);  /* inv grad_norm: high norm → low stability */
         float confidence = 1.0f - fminf(isv_signals[3], 1.0f);  /* inv ensemble_var: high var → low confidence */
-        /* Extension: 0-8 bars based on stability × confidence */
-        int extension = (int)(stability * confidence * 8.0f);
+        /* Extension: 0-2× base hold based on stability × confidence.
+         * Scales with min_hold_bars: base=50 → extension up to 100 bars. */
+        int extension = (int)(stability * confidence * (float)min_hold_bars * 2.0f);
         /* If losing (negative reward_ema), reduce hold to base only */
         if (isv_signals[5] < -0.001f) extension = 0;
         adaptive_hold = min_hold_bars + extension;