fix: revert c51/mse to runtime NVRTC — NUM_ATOMS controls loop bounds, not just sizing

c51_loss and mse_loss kernels use NUM_ATOMS for loop bounds AND shared
memory. Precompiling with NUM_ATOMS=51 then running with num_atoms=101
causes wrong loop iterations. These 2 kernels MUST use runtime #define
injection to match the hyperopt config.

35 of 37 kernels remain precompiled. Only c51_loss and mse_loss use
runtime NVRTC (they're the only kernels where a #define controls logic).

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

crates/ml/ml/hyperopt_results/dqn/20260326_072108/best_params.json

@@ -0,0 +1,71 @@
+{
+  "learning_rate": 0.00005994999354431849,
+  "batch_size": 307,
+  "gamma": 0.9572818589229506,
+  "buffer_size": 66246,
+  "max_position_absolute": 1.1030284538648683,
+  "huber_delta": 17.775790174607682,
+  "entropy_coefficient": 0.38184099247597697,
+  "transaction_cost_multiplier": 1.7738774066741243,
+  "per_alpha": 0.45251155566697393,
+  "per_beta_start": 0.20130083854118397,
+  "dueling_hidden_dim": 128,
+  "n_steps": 4,
+  "tau": 0.008986246204051671,
+  "num_atoms": 101,
+  "v_min": -280.91110000212706,
+  "v_max": 280.91110000212706,
+  "noisy_sigma_init": 0.24583814763225417,
+  "minimum_profit_factor": 1.5028035348681215,
+  "weight_decay": 0.00010597074820791314,
+  "kelly_fractional": 0.5096243658522589,
+  "kelly_max_fraction": 0.12010947414322902,
+  "kelly_min_trades": 20,
+  "volatility_window": 23,
+  "use_ensemble_uncertainty": false,
+  "ensemble_size": 5.0,
+  "beta_variance": 0.5,
+  "beta_disagreement": 0.5,
+  "beta_entropy": 0.2,
+  "warmup_ratio": 0.0,
+  "curiosity_weight": 0.0,
+  "td_error_clamp_max": 10.0,
+  "batch_diversity_cooldown": 50.0,
+  "lr_decay_type": 0.0,
+  "sharpe_weight": 0.10176748614555509,
+  "gae_lambda": 0.95,
+  "noisy_sigma_initial": 0.5,
+  "noisy_sigma_final": 0.3,
+  "use_spectral_norm": true,
+  "use_attention": true,
+  "use_residual": true,
+  "norm_type": 1.0,
+  "activation_type": 1.0,
+  "num_quantiles": 32,
+  "qr_kappa": 1.0,
+  "iqn_lambda": 0.8479734309322158,
+  "spectral_norm_sigma_max": 4.8543579475657985,
+  "hidden_dim_base": 128,
+  "noisy_epsilon_floor": 0.1,
+  "count_bonus_coefficient": 0.23831546372779777,
+  "cql_alpha": 0.5059776371295834,
+  "eval_softmax_temp": 1.0,
+  "dsr_eta": 0.003947815597664894,
+  "branch_hidden_dim": 64,
+  "gradient_accumulation_steps": 1,
+  "w_dsr": 1.0,
+  "w_pnl": 0.3,
+  "w_dd": 1.0,
+  "w_idle": 0.01,
+  "dd_threshold": 0.01,
+  "loss_aversion": 1.5,
+  "time_decay_rate": 0.0005,
+  "q_gap_threshold": 0.1,
+  "c51_warmup_epochs": 10,
+  "dt_pretrain_epochs": 0,
+  "her_ratio": 0.48591734800375336,
+  "curiosity_weight_tunable": 0.09052455736721342,
+  "use_cvar_action_selection": 0.028233223186665324,
+  "cvar_alpha": 0.03929750758139713,
+  "min_hold_bars": 9
+}

crates/ml/ml/hyperopt_results/dqn/20260326_072108/campaign_summary.json

@@ -0,0 +1,7 @@
+{
+  "trials_completed": 1,
+  "best_loss": 1000000.0,
+  "best_params_json": "{\n  \"learning_rate\": 0.00005994999354431849,\n  \"batch_size\": 307,\n  \"gamma\": 0.9572818589229506,\n  \"buffer_size\": 66246,\n  \"max_position_absolute\": 1.1030284538648683,\n  \"huber_delta\": 17.775790174607682,\n  \"entropy_coefficient\": 0.38184099247597697,\n  \"transaction_cost_multiplier\": 1.7738774066741243,\n  \"per_alpha\": 0.45251155566697393,\n  \"per_beta_start\": 0.20130083854118397,\n  \"dueling_hidden_dim\": 128,\n  \"n_steps\": 4,\n  \"tau\": 0.008986246204051671,\n  \"num_atoms\": 101,\n  \"v_min\": -280.91110000212706,\n  \"v_max\": 280.91110000212706,\n  \"noisy_sigma_init\": 0.24583814763225417,\n  \"minimum_profit_factor\": 1.5028035348681215,\n  \"weight_decay\": 0.00010597074820791314,\n  \"kelly_fractional\": 0.5096243658522589,\n  \"kelly_max_fraction\": 0.12010947414322902,\n  \"kelly_min_trades\": 20,\n  \"volatility_window\": 23,\n  \"use_ensemble_uncertainty\": false,\n  \"ensemble_size\": 5.0,\n  \"beta_variance\": 0.5,\n  \"beta_disagreement\": 0.5,\n  \"beta_entropy\": 0.2,\n  \"warmup_ratio\": 0.0,\n  \"curiosity_weight\": 0.0,\n  \"td_error_clamp_max\": 10.0,\n  \"batch_diversity_cooldown\": 50.0,\n  \"lr_decay_type\": 0.0,\n  \"sharpe_weight\": 0.10176748614555509,\n  \"gae_lambda\": 0.95,\n  \"noisy_sigma_initial\": 0.5,\n  \"noisy_sigma_final\": 0.3,\n  \"use_spectral_norm\": true,\n  \"use_attention\": true,\n  \"use_residual\": true,\n  \"norm_type\": 1.0,\n  \"activation_type\": 1.0,\n  \"num_quantiles\": 32,\n  \"qr_kappa\": 1.0,\n  \"iqn_lambda\": 0.8479734309322158,\n  \"spectral_norm_sigma_max\": 4.8543579475657985,\n  \"hidden_dim_base\": 128,\n  \"noisy_epsilon_floor\": 0.1,\n  \"count_bonus_coefficient\": 0.23831546372779777,\n  \"cql_alpha\": 0.5059776371295834,\n  \"eval_softmax_temp\": 1.0,\n  \"dsr_eta\": 0.003947815597664894,\n  \"branch_hidden_dim\": 64,\n  \"gradient_accumulation_steps\": 1,\n  \"w_dsr\": 1.0,\n  \"w_pnl\": 0.3,\n  \"w_dd\": 1.0,\n  \"w_idle\": 0.01,\n  \"dd_threshold\": 0.01,\n  \"loss_aversion\": 1.5,\n  \"time_decay_rate\": 0.0005,\n  \"q_gap_threshold\": 0.1,\n  \"c51_warmup_epochs\": 10,\n  \"dt_pretrain_epochs\": 0,\n  \"her_ratio\": 0.48591734800375336,\n  \"curiosity_weight_tunable\": 0.09052455736721342,\n  \"use_cvar_action_selection\": 0.028233223186665324,\n  \"cvar_alpha\": 0.03929750758139713,\n  \"min_hold_bars\": 9\n}",
+  "best_checkpoint_path": null,
+  "total_time_seconds": 4.693175485
+}

crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs

@@ -3779,12 +3779,31 @@ fn compile_per_update_kernel(stream: &Arc<CudaStream>) -> Result<CudaFunction, M
 /// which match the production configuration. These are hardcoded at build time.
 fn compile_c51_loss_kernel(
     stream: &Arc<CudaStream>,
-    _config: &GpuDqnTrainConfig,
+    config: &GpuDqnTrainConfig,
 ) -> Result<CudaFunction, MLError> {
+    // c51_loss uses NUM_ATOMS for BOTH loop bounds AND array sizing —
+    // MUST be runtime-compiled with the correct #define injected.
+    let defines = format!(
+        "#define NUM_ATOMS {na}\n\
+         #define V_MIN ({v_min:.1}f)\n\
+         #define V_MAX ({v_max:.1}f)\n\
+         #define BRANCH_0_SIZE {b0}\n\
+         #define BRANCH_1_SIZE {b1}\n\
+         #define BRANCH_2_SIZE {b2}\n",
+        na = config.num_atoms,
+        v_min = config.v_min,
+        v_max = config.v_max,
+        b0 = config.branch_0_size,
+        b1 = config.branch_1_size,
+        b2 = config.branch_2_size,
+    );
+    let kernel_src = include_str!("c51_loss_kernel.cu");
+    let full_source = format!("{defines}\n{kernel_src}");
     let context = stream.context();
-    let ptx = Ptx::from_binary(C51_LOSS_CUBIN.to_vec());
+    let ptx = crate::cuda_pipeline::compile_ptx_for_device(&full_source, &context)
+        .map_err(|e| MLError::ModelError(format!("c51_loss_kernel compilation: {e}")))?;
     let module = context.load_module(ptx)
-        .map_err(|e| MLError::ModelError(format!("c51_loss cubin load: {e}")))?;
+        .map_err(|e| MLError::ModelError(format!("c51_loss module load: {e}")))?;
     module.load_function("c51_loss_batched")
         .map_err(|e| MLError::ModelError(format!("c51_loss_batched load: {e}")))
 }
@@ -3804,13 +3823,31 @@ fn compile_c51_grad_kernel(
         .map_err(|e| MLError::ModelError(format!("c51_grad_kernel load: {e}")))
 }
 
-/// Load the MSE loss kernel from precompiled cubin (ZERO runtime nvcc).
+/// Compile MSE loss kernel at runtime with correct NUM_ATOMS injected.
+/// MSE uses NUM_ATOMS for loop bounds — MUST match runtime config.
 fn compile_mse_loss_kernel(
     stream: &Arc<CudaStream>,
-    _config: &GpuDqnTrainConfig,
+    config: &GpuDqnTrainConfig,
 ) -> Result<CudaFunction, MLError> {
+    let defines = format!(
+        "#define NUM_ATOMS {na}\n\
+         #define V_MIN ({v_min:.1}f)\n\
+         #define V_MAX ({v_max:.1}f)\n\
+         #define BRANCH_0_SIZE {b0}\n\
+         #define BRANCH_1_SIZE {b1}\n\
+         #define BRANCH_2_SIZE {b2}\n",
+        na = config.num_atoms,
+        v_min = config.v_min,
+        v_max = config.v_max,
+        b0 = config.branch_0_size,
+        b1 = config.branch_1_size,
+        b2 = config.branch_2_size,
+    );
+    let kernel_src = include_str!("mse_loss_kernel.cu");
+    let full_source = format!("{defines}\n{kernel_src}");
     let context = stream.context();
-    let ptx = Ptx::from_binary(MSE_LOSS_CUBIN.to_vec());
+    let ptx = crate::cuda_pipeline::compile_ptx_for_device(&full_source, &context)
+        .map_err(|e| MLError::ModelError(format!("mse_loss_kernel compilation: {e}")))?;
     let module = context.load_module(ptx)
         .map_err(|e| MLError::ModelError(format!("mse_loss cubin load: {e}")))?;
     module.load_function("mse_loss_batched")