fix: zero clippy errors + remove dead cfg gate + hex literals

- Fixed 5 clippy warnings: unused variables, empty else, hex literals
- Removed dead #[cfg(not(feature = "cuda"))] gate in PPO
- Monitoring reduce error properly handled (no let _ on must_use)
- Stack size constants use hex (0x4000, 0x10000) per clippy

Workspace status:
  cargo check --workspace --lib: 0 errors
  cargo clippy --workspace --lib -D warnings: 0 errors
  ml-core: 302/302 passed
  ml-dqn: 359/359 passed
  ml-ppo: 168/168 passed
  smoke_e2e_dqn_training_loop: PASSED (0.59s)

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

crates/ml-ppo/src/ppo.rs

@@ -622,11 +622,6 @@ impl PPO {
                 Ok((0.0, 0.0))
             }
         }
-        #[cfg(not(feature = "cuda"))]
-        {
-            let _ = batch;
-            Ok((0.0, 0.0))
-        }
     }
 
     /// Update PPO networks from GPU-resident experience data.

crates/ml/src/cuda_pipeline/gpu_curiosity_trainer.rs

@@ -186,10 +186,10 @@ impl GpuCuriosityTrainer {
             use cudarc::driver::sys::{cuCtxSetLimit, CUlimit, CUresult};
             // 16KB: kernel uses ~3KB/thread (513 floats) + call frames + nvcc spills.
             // 8KB was insufficient on some GPUs with deep call chains.
-            let result = unsafe { cuCtxSetLimit(CUlimit::CU_LIMIT_STACK_SIZE, 16384) };
+            let result = unsafe { cuCtxSetLimit(CUlimit::CU_LIMIT_STACK_SIZE, 0x4000) };
             if result != CUresult::CUDA_SUCCESS {
                 return Err(MLError::ModelError(format!(
-                    "cuCtxSetLimit(STACK_SIZE, 16384) failed: {result:?}"
+                    "cuCtxSetLimit(STACK_SIZE, 0x4000) failed: {result:?}"
                 )));
             }
         }

crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs

@@ -412,7 +412,7 @@ impl GpuDqnTrainer {
         // per array. Default 1KB causes ILLEGAL_ADDRESS stack overflow.
         {
             use cudarc::driver::sys::{cuCtxSetLimit, CUlimit, CUresult};
-            let stack_bytes = 65536_usize;
+            let stack_bytes = 0x10000_usize; // 64KB
             let result = unsafe { cuCtxSetLimit(CUlimit::CU_LIMIT_STACK_SIZE, stack_bytes) };
             if result != CUresult::CUDA_SUCCESS {
                 return Err(MLError::ModelError(format!(

crates/ml/src/trainers/dqn/trainer/training_loop.rs

@@ -791,11 +791,13 @@ impl DQNTrainer {
 
         if let Some(ref mut mon) = self.gpu_monitoring {
             if let Err(e) = mon.reduce(collector.rewards_gpu(), collector.actions_gpu(), count) {
+                debug!("GPU monitoring reduce failed: {e}");
             }
+            // Catch async monitoring kernel crashes before they poison the stream
             if let Some(ref stream) = self.cuda_stream {
-                if let Err(e) = stream.synchronize() {
+                if stream.synchronize().is_err() {
+                    warn!("GPU monitoring kernel crashed — disabling monitoring");
                     self.gpu_monitoring = None;
-                } else {
                 }
             }
         }