rtx-cfd legacy GPU: kernel outputs were device-buffer CLONES (every write lost) — the buffers are borrowed mutably now; the Poisson kernels keep boundary Dirichlet values; tests: the advection pulse marched 50 steps, the Jacobi budget 20k, the reduction reference in f64

crates/specialized/rtx-cfd/tests/gpu_reduction_tests.rs

@@ -37,7 +37,7 @@ mod gpu_reduction_tests {
 
         // Create test data
         let test_data: Vec<f32> = (1..=100).map(|i| i as f32).collect();
-        let expected_sum = test_data.iter().sum::<f32>();
+        let expected_sum = test_data.iter().map(|&v| v as f64).sum::<f64>() as f32; // f64 reference: a sequential f32 sum of a million terms carries ~1e-3 of rounding
 
         // Copy to GPU
         let gpu_data = manager.copy_to_device(&test_data)?;
@@ -100,7 +100,7 @@ mod gpu_reduction_tests {
         // Create large test array
         let n = 1_000_000;
         let test_data: Vec<f32> = (0..n).map(|i| (i % 100) as f32).collect();
-        let expected_sum = test_data.iter().sum::<f32>();
+        let expected_sum = test_data.iter().map(|&v| v as f64).sum::<f64>() as f32; // f64 reference: a sequential f32 sum of a million terms carries ~1e-3 of rounding
 
         // Copy to GPU
         let gpu_data = manager.copy_to_device(&test_data)?;