feat: plan conviction scales reward — plan head learns from epoch 1

Plan conviction (ISV-attention-driven) directly scales reward signal.
Low conviction → dampened reward, high → amplified. No hardcoded
constants. The plan head's forward path (h_s2 → ISV feature gate →
Mamba2 temporal → plan MLP → conviction) is fully attention-driven.
Learning flows: conviction → reward → Q-value → C51 loss → gradients.
Plan learns planning, holding, strategic positioning throughout training.

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

crates/ml/src/cuda_pipeline/experience_kernels.cu

@@ -1919,6 +1919,16 @@ extern "C" __global__ void experience_env_step(
         reward *= -1.0f;
     }
 
+    /* Plan conviction as reward scaling (always active, ISV-driven).
+     * The plan head receives ISV-gated h_s2 (temporal attention → feature gate → plan MLP).
+     * Its conviction output scales the reward → Q-value → C51 loss gradient path.
+     * No hardcoded constants — conviction IS the learned scaling factor.
+     * ISV modulates conviction indirectly through the attention-gated trunk. */
+    if (plan_params_ptr != NULL && fabsf(reward) > 1e-8f) {
+        float conviction = plan_params_ptr[i * 6 + 4];  /* [0, 1] */
+        reward *= conviction;  /* low conviction → dampened signal, high → amplified */
+    }
+
     out_rewards[out_off] = reward;
     out_dones[out_off] = (float)done;