diag(policy-quality): Task 0.7 — eval-mode action distribution (H10 signal)

Wires the eval_dist HEALTH_DIAG group with per-magnitude action distribution
read from the validation backtest. H10 detection signal — training-mode
entropy may look uniform while eval-mode argmax collapses to Quarter.

Implementation:
  * GpuBacktestEvaluator::read_eval_action_distribution_per_magnitude()
    reads actions_history_buf to host, decodes magnitude bin per sample
    (action layout: dir*27 + mag*9 + ord*3 + urg → mag = (a/9) % 3),
    returns [Quarter, Half, Full] normalized over non-skipped entries.
  * DQNTrainer.last_eval_magnitude_dist field, populated in metrics.rs
    after each validation backtest. Non-fatal on readback error.
  * HEALTH_DIAG eval_dist group [eq, eh, ef] now shows real values.

Per plan Task 0.7. Complete — no stubs.

crates/ml/src/cuda_pipeline/gpu_backtest_evaluator.rs

@@ -857,6 +857,43 @@ impl GpuBacktestEvaluator {
         self.dqn_graph = None;
     }
 
+    /// Task 0.7 — read back the actions_history buffer and compute the
+    /// per-magnitude action distribution observed during the most recent
+    /// evaluation. Action encoding is `dir*27 + mag*9 + ord*3 + urg`, so
+    /// magnitude = `(action / 9) % 3` (0=Quarter, 1=Half, 2=Full).
+    ///
+    /// Returns `[quarter_frac, half_frac, full_frac]` summing to 1.0 over
+    /// non-negative action entries (negative entries indicate uninitialized
+    /// or skipped steps and are excluded).
+    pub fn read_eval_action_distribution_per_magnitude(&self) -> Result<[f32; 3], MLError> {
+        let len = self.n_windows * self.max_len;
+        let mut host = vec![0_i32; len];
+        self.stream
+            .memcpy_dtoh(&self.actions_history_buf, &mut host)
+            .map_err(|e| MLError::ModelError(format!("actions_history dtoh: {e}")))?;
+        let mut counts = [0_u64; 3];
+        let mut total = 0_u64;
+        for &a in &host {
+            if a < 0 {
+                continue;
+            }
+            let mag = ((a / 9) % 3) as usize;
+            if mag < 3 {
+                counts[mag] += 1;
+                total += 1;
+            }
+        }
+        if total == 0 {
+            return Ok([0.0; 3]);
+        }
+        let t = total as f32;
+        Ok([
+            counts[0] as f32 / t,
+            counts[1] as f32 / t,
+            counts[2] as f32 / t,
+        ])
+    }
+
     /// Reset mutable evaluation state: portfolio, done flags, step returns, actions history.
     /// Must be called before each `evaluate_dqn_graphed` when reusing the evaluator
     /// across epochs — otherwise done_flags=1 from the previous evaluation causes

crates/ml/src/trainers/dqn/trainer/constructor.rs

@@ -581,6 +581,7 @@ impl DQNTrainer {
             prev_controller_values: super::ControllerPrevValues::default(),
             controller_fire_counts: super::ControllerFireCounts::default(),
             controller_total_epochs: 0,
+            last_eval_magnitude_dist: [0.0_f32; 3],
 
             // Wave 16 Portfolio Features (action masking always active)
             max_position,

crates/ml/src/trainers/dqn/trainer/metrics.rs

@@ -628,6 +628,16 @@ impl DQNTrainer {
             0.0
         };
 
+        // Task 0.7 — capture eval-mode per-magnitude action distribution from
+        // the actions_history buffer for H10 detection. Failure is non-fatal:
+        // distribution stays at last value if readback errors (won't break val).
+        if let Some(ref ev) = self.gpu_evaluator {
+            match ev.read_eval_action_distribution_per_magnitude() {
+                Ok(dist) => self.last_eval_magnitude_dist = dist,
+                Err(e) => tracing::warn!("eval magnitude dist readback failed (non-fatal): {e}"),
+            }
+        }
+
         Ok(-val_sharpe)
     }
 

crates/ml/src/trainers/dqn/trainer/mod.rs

@@ -285,6 +285,13 @@ pub struct DQNTrainer {
     /// doesn't collapse to a single bin too fast.
     pub(crate) explore_entropy_mag_history: Vec<(u32, f32)>,
 
+    /// Task 0.7 — per-magnitude action distribution observed during the
+    /// most recent VALIDATION backtest (eval-mode rollout, distinct from
+    /// training-mode action_dist). H10 detection signal: training-mode
+    /// entropy may look uniform while eval-mode argmax collapses to Quarter.
+    /// Layout: [Quarter, Half, Full] in [0, 1].
+    pub(crate) last_eval_magnitude_dist: [f32; 3],
+
     /// Task 0.9 — prior-epoch values of adaptive-controller outputs, used to
     /// detect "fire" = value changed from the prior epoch. Controllers audited:
     /// anti-LR, adaptive tau, adaptive gamma, adaptive grad-clip, CQL alpha

crates/ml/src/trainers/dqn/trainer/training_loop.rs

@@ -2307,8 +2307,11 @@ impl DQNTrainer {
                 0.0_f32, 0.0_f32, 0.0_f32, 0.0_f32, 0.0_f32, 0.0_f32,
                 // Track 1 — noisy/VSN/drift (6 f32)
                 0.0_f32, 0.0_f32, 0.0_f32, 0.0_f32, 0.0_f32, 0.0_f32,
-                // Track 1 — eval dist (3 f32)
-                0.0_f32, 0.0_f32, 0.0_f32,
+                // Track 1 — eval dist (3 f32): per-magnitude action distribution
+                // observed in the most recent validation backtest. H10 signal.
+                self.last_eval_magnitude_dist[0],
+                self.last_eval_magnitude_dist[1],
+                self.last_eval_magnitude_dist[2],
                 // Track 2 — reward contrib (6 f32)
                 0.0_f32, 0.0_f32, 0.0_f32, 0.0_f32, 0.0_f32, 0.0_f32,
                 // Track 3 — controllers (6 bool per-epoch fire + 1 f32 max running rate)