feat(sp22-vnext): Phase A4 — aux_trade_outcome loss reduce kernel

K=3 sparse cross-entropy reduce over the trade-outcome softmax tile
produced by `aux_trade_outcome_forward` (Phase A3). Mirrors the K=2
sibling `aux_next_bar_loss_reduce` structurally: single-block shmem-tree
reduce, two parallel partial strips (loss_numer + valid_count) reduced
lockstep, fmaxf(p_tgt, 1e-30) numerical floor, fmaxf(valid, 1.0) all-
skip-batch guard, valid_count_out[1] save-for-backward.

Kept as SEPARATE kernel from the K=2 sibling:
- Diagnostic isolation (distinct HEALTH_DIAG slot, distinct cubin in
  profiles for clean per-loss-source attribution)
- Sparse-label semantic clarity (~95-99% mask=-1 vs ~50-100% valid for
  the K=2 next-bar head)
- Future per-class weighting headroom (Profit/Stop/Timeout 3:1-10:1
  imbalance will likely need class-weighted CE — surgical mod here
  without touching the K=2 head's contract)

Phase A4 (this commit) is dead code — no Rust launcher yet. Phase A5
lands backward; Phase B wires the full forward→loss→backward chain.

Discipline: feedback_no_atomicadd (single-block tree-reduce), feedback_
cpu_is_read_only (pure GPU), pearl_first_observation_bootstrap (sentinel
0 valid_count produces zero gradients gracefully on cold start).

Audit: docs/dqn-wire-up-audit.md Phase A4 section.
Cubin: aux_trade_outcome_loss_reduce_kernel.cubin (9.9 KB) compiles clean.

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

crates/ml/build.rs

@@ -812,6 +812,15 @@ fn main() {
         // `docs/plans/2026-05-14-sp22-h6-vNext-trade-outcome-aux.md` for
         // the full vNext architecture spec.
         "aux_trade_outcome_forward_kernel.cu",
+        // SP22 H6 vNext Phase A4 (2026-05-13): sparse cross-entropy reduce
+        // over the K=3 softmax tile from `aux_trade_outcome_forward`.
+        // Mirrors `aux_next_bar_loss_reduce`'s single-block shmem-tree
+        // pattern: mean over B_valid (skip label=-1 rows), fmaxf(p, 1e-30)
+        // numerical floor, saved valid_count for backward. Dead code until
+        // Phase A5 (backward) + Phase B (Rust launcher wireup) land. Kept
+        // separate from the K=2 sibling for diagnostic isolation, sparse-
+        // label semantic clarity, and future per-class weighting headroom.
+        "aux_trade_outcome_loss_reduce_kernel.cu",
         // SP22 H6 (2026-05-12): per-env p_up extractor that copies
         // `aux_softmax[env, 1]` → `prev_aux_dir_prob[env]` after each aux
         // forward. The cache buffer is read by `experience_state_gather`

crates/ml/src/cuda_pipeline/aux_trade_outcome_loss_reduce_kernel.cu

@@ -0,0 +1,150 @@
+// crates/ml/src/cuda_pipeline/aux_trade_outcome_loss_reduce_kernel.cu
+//
+// SP22 H6 vNext Phase A4 (2026-05-13) — sparse cross-entropy reduce over
+// the K=3 trade-outcome softmax tile produced by
+// `aux_trade_outcome_forward`.
+//
+// Mirrors `aux_next_bar_loss_reduce`'s single-block shmem-tree reduce
+// pattern (mean over valid-row count, fmaxf(p_tgt, 1e-30) numerical floor,
+// saved valid_count for backward). Kept as a SEPARATE kernel from the K=2
+// sibling because:
+//
+//   1. Diagnostic isolation: the trade-outcome head's loss feeds a
+//      DIFFERENT HEALTH_DIAG slot (`aux_outcome_ce_ema`, allocated in
+//      Phase F) than the next-bar head's `aux_next_bar_mse_ema_idx`.
+//      Two distinct kernels = two distinct cubins = clean
+//      per-loss-source attribution in profiles and reward decomposition.
+//
+//   2. Sparse-label semantics: the trade-outcome producer emits a `-1`
+//      mask on EVERY bar without a trade-close event, which is the
+//      common case (~95-99% of bars). The next-bar head sees ~50-100%
+//      valid labels (only the trailing H bars near series-end get
+//      masked). The aggressive sparsity changes the gradient-scale
+//      arithmetic: an active-batch fraction of ~1-5% means each valid
+//      row's gradient gets divided by a much smaller B_valid → higher
+//      per-trade-close gradient magnitude. The kernel logic handles
+//      this correctly via `fmaxf(valid, 1.0f)` divisor (matches the
+//      backward kernel's `inv_B = 1 / B_valid` semantics) but the
+//      independent kernel lets Phase F's Adam-LR controller tune the
+//      trade-outcome head separately from the next-bar head if
+//      magnitude calibration becomes necessary.
+//
+//   3. Future per-class weighting: Profit/Stop/Timeout class imbalance
+//      can be ~3:1 to ~10:1 depending on the regime. Phase D+ may need
+//      class-weighted CE; an isolated kernel makes the modification
+//      surgical without touching the K=2 head's contract.
+//
+// Inputs / outputs (mirror the K=2 sibling contract)
+// ──────────────────────────────────────────────────
+// softmax           [B, K=3] f32 row-major — saved by
+//                                  `aux_trade_outcome_forward`.
+// labels            [B] i32 in {-1, 0, 1, 2} — from
+//                                  `trade_outcome_label_kernel` at the
+//                                  same step (sparse: -1 = no trade
+//                                  close at this bar).
+// B                 batch size (= N × L in the replay-batch use case;
+//                                  = n_envs in the rollout-per-step use
+//                                  case). K-runtime arg even though this
+//                                  kernel is K=3-semantic — keeps the
+//                                  launcher signature symmetric with
+//                                  the K=2 sibling for testing.
+// loss_out          [1] mean CE over the B_valid valid rows.
+// valid_count_out   [1] B_valid (saved for backward — matches the K=2
+//                                  sibling's contract).
+//
+// Discipline
+// ──────────
+// - No atomicAdd per `feedback_no_atomicadd.md` — single-block shmem-
+//   tree reduction.
+// - GPU-only per `feedback_cpu_is_read_only.md`.
+// - Numerical floor at 1e-30 prevents -log(0) propagation through the
+//   EMA when an untrained softmax emits near-zero probabilities.
+//
+// Block: AUX_BLOCK threads. Grid: (1, 1, 1).
+// Shared memory: 2 * AUX_BLOCK floats (loss numerator + valid-count
+// partials, lockstep-reduced).
+
+#include <cuda_runtime.h>
+#include <math_constants.h>
+
+#ifndef AUX_BLOCK
+#define AUX_BLOCK 256
+#endif
+
+extern "C" __global__ void aux_trade_outcome_loss_reduce(
+    /* [B, K=3] f32 row-major softmax tile from
+     * `aux_trade_outcome_forward`. */
+    const float* __restrict__ softmax,
+    /* [B] i32 labels in `{-1, 0=Profit, 1=Stop, 2=Timeout}`. Producer is
+     * `trade_outcome_label_kernel.cu` — emits -1 (mask) on bars with no
+     * trade-close event (the common case at ~95-99% of bars). */
+    const int* __restrict__ labels,
+    int B,
+    int K,
+    /* [1] mean CE over the B_valid rows whose label != -1. An all-skip
+     * batch produces 0 (no NaN; fmaxf(valid, 1) divisor). */
+    float* __restrict__ loss_out,
+    /* [1] B_valid saved for backward. Matches the K=2 sibling's contract
+     * — the backward kernel uses this as `inv_B = 1 / B_valid` for the
+     * d_logits scaling. An all-skip batch saves 0; backward then emits
+     * zero gradients across the board (no NaN). */
+    float* __restrict__ valid_count_out
+) {
+    if (blockIdx.x != 0) return;
+    extern __shared__ float smem[];
+    float* sh_loss  = smem;
+    float* sh_valid = smem + blockDim.x;
+    const int tid   = threadIdx.x;
+    const int block = blockDim.x;
+
+    /* Per-thread strided accumulation: each thread folds its slice of
+     * `[0, B)` into private (loss_numer, valid_count) accumulators.
+     * Mask-encoded rows (label == -1) contribute 0 to BOTH accumulators —
+     * the common case for trade-outcome labels where most bars have no
+     * trade-close event. The valid-row branch reads `softmax[i, label]`
+     * directly (no max-shift needed; the forward kernel already
+     * materialised stable softmax probs). */
+    float local_loss  = 0.0f;
+    float local_valid = 0.0f;
+    for (int i = tid; i < B; i += block) {
+        const int tgt = labels[i];
+        if (tgt == -1) continue;
+        /* Defensive bounds — out-of-range labels degrade to skip
+         * rather than read past the end of the [B, K] softmax row. The
+         * label producer (Phase A2 `trade_outcome_label_kernel`) emits
+         * only -1 / {0, 1, 2} so this branch is dead under correct
+         * producer behavior; kept for kernel robustness against
+         * producer bugs. */
+        if (tgt < 0 || tgt >= K) continue;
+        const float p_tgt = softmax[(size_t)i * K + tgt];
+        /* Numerical floor at 1e-30 — matches the K=2 sibling's clip.
+         * An untrained `aux_trade_outcome_forward` with extreme W2 can
+         * emit p ≈ 0 for some class; -log(0) = +inf would propagate
+         * to the EMA. log(1e-30) ≈ -69 is well outside the trained
+         * regime — the floor is robustness, not a substitute for
+         * training stability. */
+        const float p_clipped = fmaxf(p_tgt, 1e-30f);
+        local_loss  += -logf(p_clipped);
+        local_valid += 1.0f;
+    }
+    sh_loss[tid]  = local_loss;
+    sh_valid[tid] = local_valid;
+    __syncthreads();
+
+    /* Standard log2(BLOCK) tree reduction over both arrays in lockstep
+     * — identical pattern to the K=2 sibling so the proven correctness
+     * carries over directly. */
+    for (int s = block / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            sh_loss[tid]  += sh_loss[tid + s];
+            sh_valid[tid] += sh_valid[tid + s];
+        }
+        __syncthreads();
+    }
+
+    if (tid == 0) {
+        const float valid = sh_valid[0];
+        loss_out[0]        = sh_loss[0] / fmaxf(valid, 1.0f);
+        valid_count_out[0] = valid;
+    }
+}

docs/dqn-wire-up-audit.md

@@ -17693,3 +17693,32 @@ Second foundation commit for the SP22 H6 vNext trade-outcome aux head. Lands two
 - Phase F: validation smoke
 
 Cargo check clean. Registry tests clean. Ready for Phase A4.
+
+#### Phase A4 — aux_trade_outcome loss reduce kernel (2026-05-13)
+
+Third foundation commit for the SP22 H6 vNext trade-outcome aux head. Lands the sparse cross-entropy reduce.
+
+**NEW kernel `aux_trade_outcome_loss_reduce_kernel.cu`** — K=3 mean CE over the trade-outcome softmax tile with `-1` label masking.
+- Pattern: single-block shmem-tree reduce, two parallel partial strips (loss_numer + valid_count) reduced in lockstep. Identical structural shape to `aux_next_bar_loss_reduce` (K=2 sibling); same `fmaxf(p_tgt, 1e-30)` numerical floor, same `fmaxf(valid, 1.0)` all-skip-batch guard, same `valid_count_out[1]` save-for-backward contract.
+- Sparse-label arithmetic: most bars (~95-99%) have `label=-1` → contribute 0 to both numerator and valid_count. The few valid-trade-close rows produce CE numerator divided by a small B_valid → higher per-trade-close gradient magnitude. The kernel correctly handles all-skip batches (loss=0, valid=0 → backward emits zero gradients).
+- **Kept separate from the K=2 sibling** for three reasons:
+  1. Diagnostic isolation — distinct HEALTH_DIAG slot (`aux_outcome_ce_ema`, lands in Phase F), distinct cubin in profiles, clean attribution.
+  2. Sparse-label semantic clarity — the kernel comment documents the ~1-5% active-batch fraction expected at the trade-outcome head.
+  3. Future per-class weighting headroom — Profit/Stop/Timeout imbalance can be 3:1 to 10:1; an isolated kernel makes class-weighted CE a surgical modification without touching the K=2 head's contract.
+- Cubin: `aux_trade_outcome_loss_reduce_kernel.cubin` (9.9 KB).
+- Phase A4 (this commit) **dead code** — no Rust launcher yet. Phase A5 (backward) reads `valid_count_out` for `inv_B = 1/B_valid` scaling. Phase B wires the full forward→loss→backward chain.
+
+**Discipline**:
+- `feedback_no_atomicadd` — single-block shmem-tree reduction.
+- `pearl_first_observation_bootstrap` — sentinel 0.0 for valid_count means cold-start (no trade-closes yet) produces zero gradients gracefully.
+- `feedback_cpu_is_read_only` — all GPU-side.
+
+**Remaining vNext work**:
+- Phase A5: `aux_trade_outcome_backward_kernel.cu`
+- Phase B: 262-dim input concat (`h_s2_aux` || `plan_params`)
+- Phase C: 3-slot state assembly (state[121..124] = p_Profit, p_Stop, p_Timeout)
+- Phase D: 12-weight W atom-shift (4 actions × 3 outcomes)
+- Phase E: dW + Adam for W[4, 3]
+- Phase F: validation smoke
+
+Cargo check clean. Ready for Phase A5.