feat(sp14-c): aux trunk forward kernel + Rust wrapper + oracle test

3-layer MLP forward (Linear→ELU→Linear→ELU→Linear). Pre-loaded
CudaFunction for graph-capture safety per pearl_no_host_branches_in_captured_graph.
Oracle test verifies bit-for-bit match against numpy reference within
1e-4 tol. Saves h_aux1 and h_aux2 to global memory for backward.

Phase C.3 of SP14 Layer C separate-aux-trunk refactor (plan:
docs/superpowers/plans/2026-05-07-sp14-layer-c-separate-aux-trunk.md).

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

crates/ml/build.rs

@@ -196,6 +196,16 @@ fn main() {
         // Producer-only; ISV[AUX_NEXT_BAR_MSE_EMA_INDEX] +
         // ISV[AUX_REGIME_CE_EMA_INDEX] consumer wires in Commit B.
         "aux_heads_loss_ema_kernel.cu",
+        // SP14 Layer C Phase C.3 (2026-05-08): aux trunk forward kernel —
+        // 3-layer Linear→ELU→Linear→ELU→Linear MLP parallel to Q's GRN
+        // trunk. Reads encoder output (`x_in [B, ENCODER_OUT_DIM]`),
+        // writes `h_s2_aux [B, AUX_HIDDEN_DIM]` plus saved-for-backward
+        // hidden activations `h_aux1` and `h_aux2`. Module is additive
+        // in this commit — wire-up into the collector chain lands in
+        // Phase C.5 (atomic). Backward kernel + Adam updates land in
+        // Phase C.4. See plan §C.3 in
+        // `2026-05-07-sp14-layer-c-separate-aux-trunk.md`.
+        "aux_trunk_forward_kernel.cu",
         // Plan A Phase 0 (Thompson sampling spec 2026-04-26): standalone test
         // kernel exercised only by `distributional_q_tests.rs`. Implements the
         // Thompson direction sampling math (inverse-CDF over C51 atoms,