feat(sp20): c51_loss_batched aux_conf_at_state reward gate

Adds the Phase 5 consumer kernel-side gate. New kernel arg
`const float* __restrict__ aux_conf_at_state` appended to
`c51_loss_batched`'s signature. Gate computation runs once per sample
at the kernel-entry reward-setup site (after the #27 ensemble-
disagreement adjustment), then the gated `reward` propagates through
every branch's `block_bellman_project_f` call without per-branch changes.

Formula:
    gate    = sigmoid((aux_conf - threshold) / temp)
    reward  = gate * reward
where:
    threshold = ISV[AUX_CONF_THRESHOLD_INDEX=518]
    temp      = max(ISV[AUX_GATE_TEMP_INDEX=519], 1e-3)

Mathematical interpretation: at low aux confidence (gate→0),
`r_used → 0`, so the Bellman target becomes `gamma * Q(s', a')`. The
Q value at the current state collapses toward `gamma * Q(s', a')` —
model gets no reward feedback on uncertain transitions. Effectively
"don't update Q on uncertain transitions" — the "uncertain-state
neutralizer" semantic from the Phase 3 Task 3.4 audit doc spec §4.4.

NULL-tolerant: `aux_conf_at_state == NULL` OR `isv_signals == NULL`
⇒ gate skipped (identity, no-op = pre-Phase-5 behaviour). Test
scaffolds without a wired aux head still work.

Out of scope: `iqn_dual_head_kernel.cu` — IQN is the auxiliary loss,
C51 is production. Gating IQN is more complexity for marginal gain.

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

crates/ml/src/cuda_pipeline/c51_loss_kernel.cu

@@ -584,7 +584,21 @@ extern "C" __global__ void c51_loss_batched(
     const float* __restrict__ cvar_alpha_buf,      /* [B] per-sample CVaR alpha from risk branch. NULL = use iqn_readiness. */
 
     /* ── B3/G4: ISV signals for health-scaled Expected SARSA temperature ── */
-    const float* __restrict__ isv_signals          /* [ISV_DIM=23] pinned device-mapped. isv_signals[12] = learning_health. NULL = 0.5 fallback. */
+    const float* __restrict__ isv_signals,          /* [ISV_DIM=23] pinned device-mapped. isv_signals[12] = learning_health. NULL = 0.5 fallback. */
+
+    /* ── SP20 Phase 5: per-sample aux confidence at sampled state ──
+     * `[B]` f32 written by PER's direct-to-trainer `gather_f32_scalar`
+     * from the replay buffer's `aux_conf` ring (filled by the
+     * SP20 Phase 3 Task 3.4 `experience_env_step` producer per (i, t)).
+     * Range `[0, 0.5]` (K=2 peak-softmax-above-uniform); 0.0 = uniform /
+     * no info sentinel. The reward gate at the Bellman target reads
+     * `aux_conf[sample_id]` and computes
+     * `gate = sigmoid((aux_conf - ISV[AUX_CONF_THRESHOLD_INDEX=518])
+     *                 / max(ISV[AUX_GATE_TEMP_INDEX=519], 1e-3))`,
+     * then `r_used = gate * reward` before the Bellman projection.
+     * NULL-tolerant: when NULL (test scaffolds without a wired aux head),
+     * gate = 1.0 → behaviour is identical to pre-Phase-5 (no-op). */
+    const float* __restrict__ aux_conf_at_state      /* [B] f32 per-sample aux confidence at sampled state. NULL = no-op (gate = 1.0). */
 ) {
     extern __shared__ float shmem_f[];
 
@@ -665,6 +679,49 @@ extern "C" __global__ void c51_loss_batched(
         reward -= ensemble_disagreement_weight * ensemble_std[sample_id];
     }
 
+    /* ── SP20 Phase 5: aux→Q reward gate ────────────────────────────────
+     * Gate the reward (NOT the Bellman target) by aux head confidence at
+     * the sampled state. At low aux confidence (gate→0), `r_used → 0`,
+     * so the Bellman target collapses toward `gamma * Q(s', a')` and the
+     * model gets no reward feedback on uncertain transitions —
+     * effectively "don't update Q on uncertain transitions" (the
+     * "uncertain-state neutralizer" semantic per Phase 3 Task 3.4 audit
+     * doc spec §4.4).
+     *
+     * Formula:
+     *   gate = sigmoid((aux_conf - threshold) / temp)
+     *   r_gated = gate * reward
+     *
+     * where threshold = ISV[AUX_CONF_THRESHOLD_INDEX=518]
+     *                 = clamp(aux_dir_acc_ema - 0.50, 0.01, 0.20)
+     *           temp  = max(ISV[AUX_GATE_TEMP_INDEX=519], 1e-3)
+     *                 = max(aux_conf_std_ema, 0.01)
+     *
+     * (both produced by `sp20_controllers_compute_kernel` per Phase 1.3).
+     *
+     * NULL-tolerance:
+     *   - aux_conf_at_state == NULL ⇒ gate = 1.0 (identity, no-op).
+     *     Test scaffolds without a wired aux head get pre-Phase-5
+     *     behaviour.
+     *   - isv_signals == NULL ⇒ gate = 1.0 (cannot read threshold/temp,
+     *     so degrade to identity rather than producing garbage).
+     *
+     * Applied AFTER the #27 ensemble-disagreement adjustment so the
+     * cascade stays well-defined: ensemble subtracts a magnitude penalty
+     * from raw reward (semantic: "reduce effective reward magnitude
+     * when uncertain"); then this gate scales the resulting reward
+     * toward zero when the aux head can't predict the next bar
+     * (semantic: "don't trust the reward signal at all when aux is
+     * lost"). The two mechanisms compose: ensemble damps magnitude,
+     * aux damps the entire signal. */
+    if (aux_conf_at_state != NULL && isv_signals != NULL) {
+        float aux_conf  = aux_conf_at_state[sample_id];
+        float threshold = isv_signals[518];                 /* AUX_CONF_THRESHOLD_INDEX */
+        float temp      = fmaxf(isv_signals[519], 1e-3f);   /* AUX_GATE_TEMP_INDEX */
+        float gate      = 1.0f / (1.0f + __expf(-(aux_conf - threshold) / temp));
+        reward          = gate * reward;
+    }
+
     /* ── Curiosity Q-penalty base (per-sample; factored per-branch below) ──
      * Shrinks gamma for novel/uncertain states where the curiosity forward model
      * has high MSE. Applied as a multiplicative factor per branch below. */