feat(crt-train): add --smoothness-base-lambda CLI flag + telemetry

crates/ml-alpha/examples/alpha_train.rs

@@ -153,6 +153,14 @@ struct Cli {
     #[arg(long, default_value_t = 0)]
     cv_train_window: usize,
 
+    /// CRT.train intervention A — output-smoothness regularizer.
+    ///
+    /// λ[h] = base × (HORIZONS[h] / HORIZONS[0]). Default 0.0 disables
+    /// the regularizer (kernel still launches, produces zero gradient).
+    /// Recommended sweep: 0.001 / 0.01 / 0.1. See
+    /// `docs/superpowers/specs/2026-05-20-crt-train-output-smoothness-design.md`.
+    #[arg(long, default_value_t = 0.0)]
+    smoothness_base_lambda: f32,
 }
 
 #[derive(Serialize, serde::Deserialize, Default)]
@@ -196,6 +204,11 @@ struct AlphaTrainSummary {
     /// Written by the X14 wiring inside the `auc_h6000_improved` block.
     #[serde(default, skip_serializing_if = "Option::is_none")]
     best_h6000_ckpt_path: Option<String>,
+    /// Per-horizon raw mean-squared adjacent-prob diff at the final epoch
+    /// (CRT.train intervention A telemetry). All zeros when
+    /// `--smoothness-base-lambda` is unset (default).
+    #[serde(default)]
+    final_smooth_loss_per_horizon: [f32; N_HORIZONS],
 }
 
 /// Linear warmup then cosine decay to `lr_min`. `step_idx` is the
@@ -261,7 +274,7 @@ fn main() -> Result<()> {
         seed: cli.seed,
         horizon_weights,
         n_batch: cli.batch_size,
-        smoothness_base_lambda: 0.0,
+        smoothness_base_lambda: cli.smoothness_base_lambda,
     };
     let mut trainer = PerceptionTrainer::new(&dev, &trainer_cfg).context("trainer init")?;
 
@@ -655,6 +668,7 @@ fn main() -> Result<()> {
         best_auc_h6000_per_horizon,
         early_stopped,
         best_h6000_ckpt_path,
+        final_smooth_loss_per_horizon: trainer.last_smoothness_loss_per_horizon(),
     };
     let summary_path = cli.out.join("alpha_train_summary.json");
     std::fs::write(&summary_path, serde_json::to_vec_pretty(&summary).context("serialize summary")?)

crates/ml-alpha/src/trainer/perception.rs

@@ -3928,6 +3928,19 @@ impl PerceptionTrainer {
         download(&self.stream, &self.probs_per_k_d)
     }
 
+    /// Last-step per-horizon raw mean-squared-diff (pre-λ). Reads from
+    /// the mapped-pinned host shadow populated by `step_batched`. Returns
+    /// all zeros when the smoothness regularizer is disabled
+    /// (`cfg.smoothness_base_lambda == 0.0`).
+    pub fn last_smoothness_loss_per_horizon(&self) -> [f32; N_HORIZONS] {
+        let raw = self.smoothness_loss_per_horizon_host_d.read_all();
+        let mut out = [0.0f32; N_HORIZONS];
+        for (i, v) in raw.into_iter().take(N_HORIZONS).enumerate() {
+            out[i] = v;
+        }
+        out
+    }
+
     /// Convenience for callers that only need the LAST position's probs
     /// (legacy API — final-position AUC scoring).
     pub fn last_probs(&self) -> Result<[f32; N_HORIZONS]> {