chore(ml): file-level allow(unsafe_code) on 12 CUDA-launch files

CUDA kernel launches via cudarc::launch_builder and MappedF32::new
(cuMemHostAlloc DEVICEMAP FFI) inherently require unsafe blocks. The
workspace-wide '-W unsafe-code' lint produced ~80 warnings across these
files, all structurally identical. Match the established pattern from
ml-backtesting/src/harness.rs and ml-backtesting/src/sim/mod.rs: single
file-level allow with a comment explaining the rationale.

Files: alpha_dqn_h600_smoke, alpha_baseline, cublaslt_debug,
gpu_walk_forward, cuda_pipeline/mod, hyperopt adapters (mamba2, ppo),
DQN smoke_tests (helpers, td_propagation), trainers/ppo, and two
ml/tests files. Documented in dqn-wire-up-audit.md.

crates/ml/examples/alpha_baseline.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 //! Phase E.3 Task 23 — Composition backtest with cost sweep.
 //!
 //! Trains the Phase E execution-policy DQN (linear Q + Phase 1d.3

crates/ml/examples/alpha_dqn_h600_smoke.rs

@@ -38,6 +38,12 @@
 //!   --n-episodes 1000
 //! ```
 
+// CUDA kernel launches via `cudarc::launch_builder.launch(...)` and
+// `MappedF32::new` (`cuMemHostAlloc(DEVICEMAP)` FFI) require `unsafe`. The
+// workspace-wide `-W unsafe-code` lint is therefore noise here. Same rationale
+// as `ml-backtesting/src/harness.rs` and `sim/mod.rs`'s file-level allows.
+#![allow(unsafe_code)]
+
 use std::fs::File;
 use std::io::Write;
 use std::mem::MaybeUninit;

crates/ml/src/cuda_pipeline/cublaslt_debug.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 //! Standalone cublasLtMatmul debug test.
 //! Tests the EXACT same API call that fails in the experience collector on H100.
 //! Run: SQLX_OFFLINE=true cargo test -p ml --lib -- cuda_pipeline::cublaslt_debug --nocapture

crates/ml/src/cuda_pipeline/gpu_walk_forward.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 //! GPU-resident walk-forward data manager.
 //!
 //! Uploads the ENTIRE dataset to GPU VRAM once, then provides per-fold

crates/ml/src/cuda_pipeline/mod.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 //! GPU Pipeline: Pre-uploaded training data for DQN/PPO trainers.
 //!
 //! Eliminates per-bar heap allocations by uploading all training data

crates/ml/src/hyperopt/adapters/mamba2.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 //! MAMBA-2 Hyperparameter Optimization Adapter
 //!
 //! This module provides a production-ready adapter for optimizing MAMBA-2

crates/ml/src/hyperopt/adapters/ppo.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 //! PPO Hyperparameter Optimization Adapter
 //!
 //! This module provides a production-ready adapter for optimizing PPO

crates/ml/src/trainers/dqn/smoke_tests/helpers.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 use crate::trainers::dqn::{DQNHyperparameters, DQNTrainer};
 use ml_core::device::MlDevice;
 

crates/ml/src/trainers/dqn/smoke_tests/td_propagation.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 //! TD propagation diagnostic: does sparse-reward Q-learning actually learn?
 //!
 //! After Phase 2's reward-shaping purge, training reward is essentially pure

crates/ml/src/trainers/ppo.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 //! PPO Trainer with gRPC Integration
 //!
 //! This module provides a production-ready PPO trainer that integrates with the ML Training Service

crates/ml/tests/regime_wr_oracle_tests.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 //! Per-regime WR GPU oracle tests (audit follow-up 2026-05-09).
 //!
 //! Validates that `compute_backtest_metrics` correctly bucketises

crates/ml/tests/smoke_test_real_data.rs

@@ -1,3 +1,7 @@
+// CUDA kernel launches via cudarc + MappedF32::new (cuMemHostAlloc DEVICEMAP) require
+// unsafe FFI. Workspace -W unsafe-code is structural noise here. Same rationale as
+// ml-backtesting/src/harness.rs file-level allow.
+#![allow(unsafe_code)]
 #![allow(
     clippy::assertions_on_constants,
     clippy::assertions_on_result_states,