fix: DQNTrainer hard-errors on CUDA init failure, no silent CPU fallback

cuda_if_available() silently fell back to CPU when CUDA init failed,
causing DQNTrainer to crash later with "cuda_stream() called on CPU
device". This was the root cause of the CI "hang" — the test failed
immediately but the error was invisible due to buffered output.

- DQNTrainer::new() now uses MlDevice::cuda(0)? with explicit error
- cuda_if_available() now logs the CUDA error before falling back

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

crates/ml-core/src/device.rs

@@ -64,7 +64,10 @@ impl MlDevice {
     pub fn cuda_if_available(ordinal: usize) -> Self {
         match Self::cuda(ordinal) {
             Ok(dev) => dev,
-            Err(_) => MlDevice::Cpu,
+            Err(e) => {
+                tracing::warn!("CUDA device {ordinal} unavailable ({e}), falling back to CPU");
+                MlDevice::Cpu
+            }
         }
     }
 

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

@@ -118,11 +118,19 @@ impl DQNTrainer {
         }
 
         // Use override device if provided (hyperopt shares one CUDA context),
-        // otherwise auto-detect GPU
+        // otherwise require CUDA GPU — no CPU fallback in CUDA builds.
         let device = if let Some(dev) = override_device {
             dev
         } else {
-            MlDevice::cuda_if_available(0)
+            match MlDevice::cuda(0) {
+                Ok(dev) => dev,
+                Err(e) => {
+                    tracing::error!("CUDA device init failed: {e}");
+                    return Err(anyhow::anyhow!(
+                        "DQNTrainer requires CUDA GPU. CUDA init failed: {e}"
+                    ));
+                }
+            }
         };
 
         // Fork a dedicated CudaStream once. All GPU components (experience collector,