fix: zen precommit — 2 criticals + 3 lows resolved

CRITICAL: ISV+plan heads wired into submit_forward_ops_main (training).
CRITICAL: backtest single-step 3-branch→4-branch action encoding.
LOW: Remove dead bar_start_ns/bar_duration_ns from MicrostructureState.
LOW: Pre-allocated eps_host_buf replaces Vec allocation in hot path.
LOW: training_sharpe + val_loss added to metrics additional_metrics.

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

crates/ml-features/src/ofi_calculator.rs

@@ -684,11 +684,8 @@ const fn safe_clip(value: f64, min: f64, max: f64) -> f64 {
 /// | 18  | OFI Acceleration          |
 /// | 19  | Toxicity Gradient         |
 #[derive(Debug, Clone)]
-#[allow(dead_code)]
 pub struct MicrostructureState {
-    // --- bar timing ---
-    bar_start_ns: u64,
-    bar_duration_ns: u64,
+    /// Bar midpoint for intra-bar momentum half-split (used by update_snapshot).
     bar_mid_ns: u64,
 
     // --- [8] OFI Trajectory: online linear regression of OFI_L1 over time ---
@@ -758,8 +755,6 @@ impl MicrostructureState {
     /// Create a new state for a bar starting at `bar_start_ns` with given duration.
     pub fn new(bar_start_ns: u64, bar_duration_ns: u64) -> Self {
         Self {
-            bar_start_ns,
-            bar_duration_ns,
             bar_mid_ns: bar_start_ns + bar_duration_ns / 2,
 
             ofi_linreg_n: 0.0,

crates/ml/src/cuda_pipeline/gpu_action_selector.rs

@@ -46,6 +46,8 @@ pub struct GpuActionSelector {
     /// Per-sample epsilon buffer for branching kernel: [max_batch_size] on GPU.
     /// Filled with uniform epsilon before each branching launch.
     epsilon_buf: CudaSlice<f32>,
+    /// Pre-allocated host buffer for epsilon fill — avoids Vec allocation in hot path.
+    eps_host_buf: Vec<f32>,
 }
 
 impl GpuActionSelector {
@@ -71,6 +73,7 @@ impl GpuActionSelector {
             bonus_exposure_ptr: 0, bonus_order_ptr: 0, bonus_urgency_ptr: 0,
             bonus_exposure_buf: None, bonus_order_buf: None, bonus_urgency_buf: None,
             epsilon_buf,
+            eps_host_buf: vec![0.0_f32; max_batch_size],
         })
     }
 
@@ -151,8 +154,9 @@ impl GpuActionSelector {
         if batch_size > self.max_batch_size { return Err(MLError::ModelError(format!("batch_size {} exceeds max {}", batch_size, self.max_batch_size))); }
         // Fill per-sample epsilon buffer with uniform value (kernel reads per_sample_epsilon[idx])
         // Slice to batch_size to match eps_host length (epsilon_buf allocated for max_batch_size)
-        let eps_host = vec![epsilon; batch_size];
-        self.stream.memcpy_htod(&eps_host, &mut self.epsilon_buf.slice_mut(..batch_size))
+        // Fill pre-allocated host buffer (no heap allocation in hot path)
+        self.eps_host_buf[..batch_size].fill(epsilon);
+        self.stream.memcpy_htod(&self.eps_host_buf[..batch_size], &mut self.epsilon_buf.slice_mut(..batch_size))
             .map_err(|e| MLError::ModelError(format!("upload epsilon_buf: {e}")))?;
         let config = launch_config_1d(batch_size);
         let bs_i32 = batch_size as i32;