refactor(rl): pre-allocate 56 replay-step gradient buffers for Graph C

Move all step_synthetic/dqn_replay_step alloc_zeros to persistent
trainer fields (ss_* prefix). Enables CUDA Graph capture of the replay
training step — all device pointers are now stable across steps.

Introduces reduce_axis0_free() to resolve borrow-checker E0502 when
both source (per-batch scratch) and destination (reduced grad) are
self fields passed to the same function.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

crates/ml-alpha/build.rs

@@ -99,6 +99,7 @@ const KERNELS: &[&str] = &[
     "rl_noisy_linear_backward",      // NoisyNet: factored noisy linear backward — grad_mu_w/sigma_w/mu_b/sigma_b per-batch scratch for reduce_axis0
     "rl_sample_tau",                 // CUDA graph prereq: device-side xorshift32 tau ~ U(0,1) for IQN; replaces host ChaCha8 + mapped-pinned upload
     "rl_sample_noise",               // CUDA graph prereq: device-side factored noise f(rand) for NoisyLinear; replaces host ChaCha8 + mapped-pinned upload
+    "rl_write_u64",                  // CUDA graph prereq: single-thread u64 scalar write for device-resident ts_ns (graph-captured kernels read via pointer)
     "rl_increment_step",             // device-resident step counter bump (ISV[548] += 1.0); graph-safe prereq — removes scalar current_step from all downstream kernel args
     "rl_fused_controllers",          // fused kernel: 10 RL ISV controllers in one launch (gamma, tau, ppo_clip, entropy_coef, rollout_steps, per_alpha, reward_scale, ppo_ratio_clamp, gate_threshold, q_distill_lambda) — saves 9 kernel launch overheads (~40-80μs/step)
 ];

crates/ml-alpha/cuda/rl_multires_features_update.cu

@@ -40,7 +40,7 @@ extern "C" __global__ void rl_multires_features_update(
     const float*   __restrict__ bid_sz,           // [BOOK_LEVELS]
     const float*   __restrict__ ask_sz,           // [BOOK_LEVELS]
     const float*   __restrict__ isv,
-    unsigned long long            current_ts_ns,
+    const unsigned long long* __restrict__ current_ts_ns_ptr,
     int                           b_size
 ) {
     const int b = blockIdx.x * blockDim.x + threadIdx.x;
@@ -48,6 +48,7 @@ extern "C" __global__ void rl_multires_features_update(
 
     const float mid = 0.5f * (bid_px[0] + ask_px[0]);
     const float old_mid = prev_mid[b];
+    const unsigned long long current_ts_ns = current_ts_ns_ptr[0];
     const unsigned long long old_ts = prev_ts_ns[b];
 
     // dt in seconds (protect against zero/backward timestamps).

crates/ml-alpha/cuda/rl_write_u64.cu

@@ -0,0 +1,13 @@
+// rl_write_u64.cu — Single-thread scalar write for device-resident u64.
+// Launched OUTSIDE graph capture to stage changing scalar values into
+// device buffers that graph-captured kernels read via pointer.
+// Grid=(1,1,1), Block=(1,1,1).
+
+#include <stdint.h>
+
+extern "C" __global__ void rl_write_u64(
+    uint64_t* __restrict__ dst,
+    uint64_t               val
+) {
+    if (threadIdx.x == 0) dst[0] = val;
+}