fix(hyperopt): max_steps_per_epoch by GPU tier, adam_epsilon 1e-3→1e-8

- Hyperopt adapter now sets max_training_steps_per_epoch:
  RTX 3050 (≤40GB) = 200 steps, H100 (≥40GB) = 2000 steps.
  Without this, each trial trained the full dataset (2917 steps/epoch)
  making hyperopt 11x slower than necessary on local GPU.

- adam_epsilon default 1e-3→1e-8 everywhere (conservative(), DQNConfig).
  The old 1e-3 was a BF16 workaround (bf16(1e-8)=0 → div-by-zero).
  Adam is now f32, so standard 1e-8 is correct.

- Early stopping enabled in dqn-localdev.toml (patience=20).

Hyperopt: 2 trials × 5 epochs in 132s (was ~20min). Zero NaN.

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

config/training/dqn-localdev.toml

@@ -32,7 +32,7 @@ buffer_size = 10000
 min_replay_size = 100
 
 [early_stopping]
-enabled = false
+enabled = true
 patience = 20
 min_epochs_before_stopping = 50
 

crates/ml-dqn/src/dqn.rs

@@ -325,7 +325,7 @@ impl Default for DQNConfig {
             // Agent-level trading parameters
             minimum_profit_factor: 1.5, // BUG #7 FIX: 50% margin above breakeven
             weight_decay: 1e-4, // WAVE 30: Standard L2 regularization strength
-            adam_epsilon: 1e-3, // BF16-safe: bf16(1e-8)=0
+            adam_epsilon: 1e-8, // f32 Adam (master weights are f32)
             dropout_rate: 0.0, // Disabled by default; hyperopt wires from dropout_initial
 
             // Mixed precision: disabled by default (auto-detected at runtime)
@@ -610,7 +610,7 @@ impl DQNConfig {
 
             minimum_profit_factor: 1.5,
             weight_decay: 1e-4,
-            adam_epsilon: 1e-3, // BF16-safe: bf16(1e-8)=0
+            adam_epsilon: 1e-8, // f32 Adam (master weights are f32)
 
             ..Default::default()
         }
@@ -677,7 +677,7 @@ impl DQNConfig {
 
             minimum_profit_factor: 2.0, // Higher safety margin for conservative config
             weight_decay: 1e-4,
-            adam_epsilon: 1e-3, // BF16-safe: bf16(1e-8)=0
+            adam_epsilon: 1e-8, // f32 Adam (master weights are f32)
 
             ..Default::default()
         }
@@ -769,7 +769,7 @@ impl DQNConfig {
 
             minimum_profit_factor: 1.5,
             weight_decay: 1e-4,
-            adam_epsilon: 1e-3, // BF16-safe: bf16(1e-8)=0
+            adam_epsilon: 1e-8, // f32 Adam (master weights are f32)
             dropout_rate: 0.0,
 
         }

crates/ml/src/hyperopt/adapters/dqn.rs

@@ -2700,6 +2700,10 @@ impl HyperparameterOptimizable for DQNTrainer {
             // medium GPUs (16-24GB) where optimal_n_episodes might not kick in.
             gpu_n_episodes: if budget.gpu_memory_mb >= 40960 { 512 } else { 128 },
             gpu_timesteps_per_episode: 500,
+            // Cap steps/epoch by GPU tier: RTX 3050 = fast iteration, H100 = full quality.
+            // Without this, the default (0=unlimited) trains on the entire dataset per epoch,
+            // making each hyperopt trial ~10min on RTX 3050 instead of ~30s.
+            max_training_steps_per_epoch: if budget.gpu_memory_mb >= 40960 { 2000 } else { 200 },
             avg_spread: 0.0001,
 
             // GPU-dynamic network sizing (from hyperopt search space)

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

@@ -1503,7 +1503,7 @@ impl DQNHyperparameters {
 
             // WAVE 30: L2 Weight Decay
             weight_decay: 1e-4,                   // Default: 0.0001 (standard regularization strength)
-            adam_epsilon: 1e-3, // BF16-safe
+            adam_epsilon: 1e-8, // f32 Adam (master weights are f32)
 
             // Conservative Q-Learning (CQL)
             cql_alpha: 0.1,                       // Default: mild conservatism (0.1 of 0.0-1.0 range)
@@ -1713,7 +1713,7 @@ pub(crate) fn dqn_default_config() -> DQNConfig {
 
         minimum_profit_factor: 1.5,
         weight_decay: 1e-4,
-            adam_epsilon: 1e-3, // BF16-safe
+            adam_epsilon: 1e-8, // f32 Adam (master weights are f32)
         ..Default::default()
     }
 }