refactor(ml-alpha): move Mamba2 stack 1 from PerceptionTrainer to CfcTrunk (X3)

PerceptionTrainer no longer owns the Mamba2 stack-1 block; the trunk's
mamba2_stack_1 Option is filled in after Mamba2Block::new (which still
runs at the same point in PerceptionTrainer::new so the Mamba2 weight
init order is unchanged). Mamba2AdamW was already constructed against
&mamba2 before the move, so optimizer state is preserved.

CfcTrunk gains mamba2_l1_mut() / mamba2_l1() / mamba2_l2_mut() / mamba2_l2()
accessors that unwrap the Options. Forward, backward, and AdamW step
sites in evaluate_batched / step_batched / evaluate redirect through the
new accessors.

Gradient buffers (mamba2_grads_buffers) and AdamW state (mamba2_adamw)
remain at trainer level — training-only state stays with the trainer.

Verification:
- perception_forward_golden: PASS (max_diff = 0.000000)
- ml-alpha lib tests: 34 pass

Per spec 2026-05-19-ml-alpha-v2-trunk-grows-and-deployability-design.md §1.1, §2.2 (X3).
EOF
)

crates/ml-alpha/src/cfc/trunk.rs

@@ -526,6 +526,25 @@ impl CfcTrunk {
     /// Reads each device tensor back via memcpy_dtoh and packs into a
     /// CheckpointV1 envelope. Currently writes ~22-25k f32s = ~90KB
     /// per CfcTrunk.
+    /// Accessor for Mamba2 stack 1. Panics if X3 migration hasn't run.
+    pub fn mamba2_l1_mut(&mut self) -> &mut crate::mamba2_block::Mamba2Block {
+        self.mamba2_stack_1.as_mut()
+            .expect("Mamba2 stack 1 not populated (X3 migration must have run)")
+    }
+    pub fn mamba2_l1(&self) -> &crate::mamba2_block::Mamba2Block {
+        self.mamba2_stack_1.as_ref()
+            .expect("Mamba2 stack 1 not populated (X3 migration must have run)")
+    }
+    /// Accessor for Mamba2 stack 2. Panics if X5 migration hasn't run.
+    pub fn mamba2_l2_mut(&mut self) -> &mut crate::mamba2_block::Mamba2Block {
+        self.mamba2_stack_2.as_mut()
+            .expect("Mamba2 stack 2 not populated (X5 migration must have run)")
+    }
+    pub fn mamba2_l2(&self) -> &crate::mamba2_block::Mamba2Block {
+        self.mamba2_stack_2.as_ref()
+            .expect("Mamba2 stack 2 not populated (X5 migration must have run)")
+    }
+
     pub fn save_checkpoint(&self, path: &std::path::Path) -> Result<()> {
         use std::io::Write;
         let w_in = {

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

@@ -153,8 +153,9 @@ pub struct PerceptionTrainer {
     heads_grn_bwd_fn: CudaFunction,
     transpose_3d_fn: CudaFunction,
 
-    // Mamba2 encoder block + its optimizer
-    pub mamba2: Mamba2Block,
+    // X3: Mamba2 stack 1 weights moved to `self.trunk.mamba2_stack_1`.
+    // Access via `self.trunk.mamba2_l1()` / `mamba2_l1_mut()`. Gradient
+    // buffers + AdamW state remain at trainer level.
     pub mamba2_adamw: Mamba2AdamW,
     /// Phase 2B: SECOND Mamba2 stack. Sits between the (first-stack-output
     /// + LN_a) and the existing LN_b → CfC. Same hidden_dim as stack 1
@@ -635,6 +636,11 @@ impl PerceptionTrainer {
         let mamba2_grads_buffers = Mamba2BackwardGradsBuffers::new(
             &stream, cfg.n_batch, cfg.seq_len, FEATURE_DIM, HIDDEN_DIM, cfg.mamba2_state_dim,
         ).context("Mamba2BackwardGradsBuffers::new")?;
+        // X3: move Mamba2 stack 1 into the trunk; subsequent access via
+        // self.trunk.mamba2_l1() / mamba2_l1_mut(). Mamba2AdamW was built
+        // above against &mamba2 (read-only borrow that ended before this
+        // move), so the optimizer state is fully constructed.
+        trunk.mamba2_stack_1 = Some(mamba2);
 
         // ── Phase 2B: SECOND Mamba2 stack ──
         // Same hidden_dim + state_dim as stack 1. in_dim = HIDDEN_DIM
@@ -1057,7 +1063,6 @@ impl PerceptionTrainer {
             heads_grn_fwd_fn,
             heads_grn_bwd_fn,
             transpose_3d_fn,
-            mamba2,
             mamba2_adamw,
             mamba2_fwd_scratch,
             mamba2_bwd_scratch,
@@ -1432,7 +1437,8 @@ impl PerceptionTrainer {
 
         // ── 2. Mamba2 stack-1 forward — writes into mamba2_fwd_scratch.
         //       Consumes vsn_out_d (gated features). in_dim = FEATURE_DIM (40).
-        self.mamba2
+        self.trunk
+            .mamba2_l1_mut()
             .forward_train_seq_into(&self.vsn_out_d, &mut self.mamba2_fwd_scratch)
             .context("mamba2 (l1) forward_train_seq_into")?;
 
@@ -2115,7 +2121,8 @@ impl PerceptionTrainer {
         //       Writes:
         //         mamba2_grads_buffers.d_x_from_in = grad on VSN output
         //         (consumed by VSN bwd below).
-        self.mamba2
+        self.trunk
+            .mamba2_l1_mut()
             .backward_from_h_enriched_seq_full_into(
                 &self.vsn_out_d,
                 &self.mamba2_fwd_scratch,
@@ -2293,7 +2300,7 @@ impl PerceptionTrainer {
         // GPU-resident grad-clip + AdamW (zero host roundtrips).
         // Replaces step_from_buffers which did 9× memcpy_dtoh per step.
         self.mamba2_adamw
-            .step_from_buffers_gpu_clip(&mut self.mamba2, &self.mamba2_grads_buffers)
+            .step_from_buffers_gpu_clip(self.trunk.mamba2_l1_mut(), &self.mamba2_grads_buffers)
             .context("mamba2 (l1) AdamW step_from_buffers_gpu_clip")?;
         self.mamba2_l2_adamw
             .step_from_buffers_gpu_clip(&mut self.mamba2_l2, &self.mamba2_l2_grads_buffers)
@@ -2470,7 +2477,8 @@ impl PerceptionTrainer {
         }
 
         // Mamba2 stack-1 fwd → m1.h_enriched_seq.
-        self.mamba2
+        self.trunk
+            .mamba2_l1_mut()
             .forward_train_seq_into(&self.vsn_out_d, &mut self.mamba2_fwd_scratch)
             .context("eval mamba2 (l1) fwd_into")?;