feat(sp14/sp22): enlarge aux head + trunk capacity (8x head, 2x trunk-H2)

After Path C confirmed aux head's 28% accuracy at H=60 is the H6 Phase 3
bottleneck (not the mechanism itself), enlarge aux capacity:
- AUX_HIDDEN_DIM: 32 -> 256 (8x, matches input dim, removes bottleneck)
- AUX_TRUNK_H2: 128 -> 256 (2x, uniform trunk width)

Architecture changes:
- Aux head: 256 -> Linear -> 256 -> ELU -> Linear -> 2 (was 256->32->2)
- Aux trunk: 256 -> 256 -> 256 -> 256 (was 256 -> 256 -> 128 -> 256)
- +160K params total (mostly aux_nb_w1 + aux_rg_w1: [256, 256] each)

Side effects:
- Checkpoint fingerprint change (intentional)
- Thread utilisation improves: AUX_BLOCK=256 threads x H=256 = 1:1
  (vs 8:1 at H=32 — most threads idle previously)

Phase 3 mechanism stays DORMANT (W=0, beta=0) for this validation
smoke. Verdict criteria: aux_dir_acc improves from 0.28 toward 0.50+
with the larger capacity. If yes, re-activate Phase 3 priors. If no,
the bottleneck is signal/horizon, not capacity.

Cargo build clean (full nvcc rebuild).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

crates/ml/src/cuda_pipeline/aux_heads_kernel.cu

@@ -64,7 +64,9 @@
 /* AUX_HIDDEN_DIM kept as a compile-time literal so kernel launch configs
  * don't need to thread `H` through. Both heads share the same hidden width. */
 #ifndef AUX_HIDDEN_DIM
-#define AUX_HIDDEN_DIM 32
+/* 2026-05-13: enlarged 32 → 256 (matches input h_s2_aux dim). See
+ * `gpu_aux_heads.rs::AUX_HIDDEN_DIM` for rationale and side effects. */
+#define AUX_HIDDEN_DIM 256
 #endif
 
 /* Block dim used by reduction-shaped launches (per-tensor backward, label

crates/ml/src/cuda_pipeline/gpu_aux_heads.rs

@@ -49,7 +49,24 @@ use super::gpu_dqn_trainer::{AUX_HEADS_CUBIN, AUX_HEADS_LOSS_EMA_CUBIN};
 /// Hidden dim shared by both auxiliary heads (Linear_1 output width).
 /// Kept in sync with the `AUX_HIDDEN_DIM` macro at the top of
 /// `aux_heads_kernel.cu`.
-pub(crate) const AUX_HIDDEN_DIM: usize = 32;
+///
+/// **Enlarged 32 → 256 (2026-05-13)** as part of the aux-head capacity
+/// upgrade after Path C investigation revealed the aux head's 28%
+/// accuracy at H=60 bars. The prior 32-unit bottleneck was severely
+/// under-capacity for a 256-dim input → 2-way classification task. The
+/// new 256 hidden dim matches the input dimension exactly (no
+/// dimensionality reduction in Layer 1) and gives the head full
+/// capacity to learn an arbitrary mapping. Architecture:
+///   256 (h_s2_aux input) → Linear → 256 (hidden) → ELU → Linear → 2 (logits)
+///
+/// Side effects:
+/// - Param buffer grows: aux_nb_w1/aux_rg_w1 went from [32, 256] to
+///   [256, 256] each → +64K params per head → +128K params total.
+/// - Fingerprint changes → checkpoint incompatibility (intentional,
+///   fresh-start training).
+/// - Thread utilisation improves: AUX_BLOCK=256 threads × H=256 hidden
+///   = 1 thread per hidden unit (vs 8 threads/unit at H=32).
+pub(crate) const AUX_HIDDEN_DIM: usize = 256;
 
 /// Output cardinality of the next-bar direction classification head
 /// (SP13 B1.1a, 2026-05-05). Flipped 1 → 2: the head was a K=1 MSE

crates/ml/src/cuda_pipeline/gpu_aux_trunk.rs

@@ -29,11 +29,12 @@
 //!   - `h_aux2 [B, H2]`       — Layer-2 post-ELU activation
 //!   - `h_s2_aux [B, AUX_HIDDEN_DIM]` — final linear output (no activation)
 //!
-//! Production topology (verified C.2):
+//! Production topology (verified C.2; widths uplifted 2026-05-13):
 //!   - `ENCODER_OUT_DIM = config.shared_h1 = 256`
 //!   - `H1 = AUX_TRUNK_H1 = 256`
-//!   - `H2 = AUX_TRUNK_H2 = 128`
-//!   - `AUX_HIDDEN_DIM = config.shared_h2 = 256` (matches existing aux head input dim)
+//!   - `H2 = AUX_TRUNK_H2 = 256` (was 128 pre-2026-05-13 — bottleneck removed)
+//!   - `AUX_HIDDEN_DIM = config.shared_h2 = 256` (matches existing aux head input dim;
+//!     internal aux head hidden layer ALSO enlarged 32 → 256 in same commit)
 //!
 //! # Pearls applied
 //!
@@ -65,10 +66,16 @@ use super::gpu_dqn_trainer::{
 pub(crate) const AUX_TRUNK_H1: usize = 256;
 
 /// Hidden width of the aux trunk's second internal layer (Linear_2 → ELU
-/// output). Smaller than H1 for a mild bottleneck — encourages the aux
-/// trunk to compress its representation before lifting back up to
-/// `AUX_HIDDEN_DIM` in Layer 3.
-pub(crate) const AUX_TRUNK_H2: usize = 128;
+/// output).
+///
+/// **Enlarged 128 → 256 (2026-05-13)** alongside `AUX_HIDDEN_DIM 32 → 256`
+/// to remove the aux trunk's internal bottleneck. After the Path C
+/// investigation found the aux head's 28% accuracy at H=60 bars, the
+/// hypothesis is that capacity (not signal) was the bottleneck. The
+/// trunk now has uniform 256 width throughout: 256 (encoder) → 256 (L1)
+/// → 256 (L2) → 256 (L3 = h_s2_aux). This gives the trunk maximum
+/// expressive power before the aux head's classification layer.
+pub(crate) const AUX_TRUNK_H2: usize = 256;
 
 /// Block dim used by the per-sample forward kernel. Mirrors
 /// `aux_heads_kernel.cu`'s `AUX_BLOCK = 256`.