feat(ml-alpha): PerceptionTrainer.save_checkpoint + alpha_train wiring (X13+X14)

X13: Adds PerceptionTrainer::save_checkpoint as a thin delegate to
self.trunk.save_checkpoint. Inference-only serialization — grads + AdamW
state aren't included.

X14: Inside the existing auc_h6000_improved block in alpha_train.rs,
calls trainer.save_checkpoint(out_dir / 'trunk_best_h6000.bin') so the
trained trunk lands alongside alpha_train_summary.json. Extends
AlphaTrainSummary with best_h6000_ckpt_path (Option<String>) so
downstream tooling (fxt-backtest --checkpoint) can locate the file
without re-deriving the path.

After this commit, every alpha-perception Argo workflow run produces
a CheckpointV2 file at every new-best-h6000 epoch, ready for backtest
consumption.

Verification:
- ml-alpha lib tests: 34 pass
- alpha_train example builds clean (release)

Per spec §1.1 (X13+X14).

crates/ml-alpha/examples/alpha_train.rs

@@ -164,7 +164,7 @@ struct Cli {
     decision_stride: usize,
 }
 
-#[derive(Serialize)]
+#[derive(Serialize, serde::Deserialize, Default)]
 struct AlphaTrainSummary {
     epochs: usize,
     seq_len: usize,
@@ -200,6 +200,11 @@ struct AlphaTrainSummary {
     best_auc_h6000_per_horizon: [f32; N_HORIZONS],
     /// True if training stopped early (patience exceeded).
     early_stopped: bool,
+    /// Relative path (within `out_dir`) to the best-h6000 trunk
+    /// checkpoint file, or None if no h6000 improvement was ever recorded.
+    /// 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>,
 }
 
 /// Linear warmup then cosine decay to `lr_min`. `step_idx` is the
@@ -310,6 +315,7 @@ fn main() -> Result<()> {
     let mut best_auc_h6000 = f32::NEG_INFINITY;
     let mut best_auc_h6000_epoch = 0usize;
     let mut best_auc_h6000_per_horizon = [0.5_f32; N_HORIZONS];
+    let mut best_h6000_ckpt_path: Option<String> = None;
     let mut auc_h6000_no_improvement = 0usize;
 
     let lr_min_cfc = cli.lr_cfc * cli.lr_min_factor;
@@ -606,7 +612,16 @@ fn main() -> Result<()> {
             best_auc_h6000_epoch = epoch;
             best_auc_h6000_per_horizon = per_horizon_auc;
             auc_h6000_no_improvement = 0;
-            tracing::info!(epoch, auc_h6000, "new best auc_h6000");
+            // X14: emit a CheckpointV2 file alongside summary.json so
+            // fxt-backtest --checkpoint can load the trained trunk.
+            let ckpt_path = cli.out.join("trunk_best_h6000.bin");
+            trainer.save_checkpoint(&ckpt_path)
+                .context("save_checkpoint(trunk_best_h6000.bin)")?;
+            best_h6000_ckpt_path = Some("trunk_best_h6000.bin".to_string());
+            tracing::info!(
+                epoch, auc_h6000, path = %ckpt_path.display(),
+                "new best auc_h6000 (saved checkpoint)"
+            );
         } else {
             auc_h6000_no_improvement += 1;
         }
@@ -650,6 +665,7 @@ fn main() -> Result<()> {
         best_auc_h6000,
         best_auc_h6000_per_horizon,
         early_stopped,
+        best_h6000_ckpt_path,
     };
     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

@@ -2306,6 +2306,14 @@ impl PerceptionTrainer {
 
     /// Single-sequence forward-only eval — thin wrapper around
     /// [`evaluate_batched`] for cfg.n_batch == 1.
+    /// X13: Persist the trunk's v2 weights to a CheckpointV2 bincode file.
+    /// Delegates to `self.trunk.save_checkpoint` — gradient buffers and
+    /// AdamW state are NOT serialized (training-only, not needed for
+    /// inference).
+    pub fn save_checkpoint(&self, path: &std::path::Path) -> Result<()> {
+        self.trunk.save_checkpoint(path).context("trunk save_checkpoint")
+    }
+
     pub fn evaluate(
         &mut self,
         snapshots: &[Mbp10RawInput],