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/kernels_tests.rs

@@ -46,17 +46,15 @@ mod cuda_tests {
         d_phi = kernel_manager.copy_to_device(&phi)?;
 
         // Run advection kernel
-        advection_kernel.apply(
-            &d_phi,
-            &mut d_phi_new,
-            velocity as f32,
-            dt as f32,
-            dx as f32,
-        )?;
+        // CFL = v dt / dx ≈ 0.1: march 50 steps so the pulse moves several cells.
+        for _ in 0..50 {
+            advection_kernel.apply(&d_phi, &mut d_phi_new, velocity as f32, dt as f32, dx as f32)?;
+            std::mem::swap(&mut d_phi, &mut d_phi_new);
+        }
 
         // Copy result back
         let mut result = vec![0.0f32; nx];
-        result = kernel_manager.copy_from_device(&d_phi_new)?;
+        result = kernel_manager.copy_from_device(&d_phi)?;
 
         // Verify that the pulse has moved (mass conservation)
         let initial_mass: f32 = phi.iter().sum();
@@ -206,7 +204,7 @@ mod cuda_tests {
         d_source = kernel_manager.copy_to_device(&source)?;
 
         // Solve Poisson equation
-        let max_iterations = 1000;
+        let max_iterations = 20_000; // Jacobi needs O(n²) sweeps on this grid
         let tolerance = 1e-6;
         let iterations = poisson_kernel.solve_2d(
             &mut d_phi,