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/src/solvers/incompressible/simple_gpu.rs

@@ -136,7 +136,7 @@ impl SimpleGpuSolver {
     fn copy_to_gpu(&self, flow_field: &FlowField) -> CfdResult<()> {
         let buffers = self
             .gpu_buffers
-            .as_ref()
+            .as_mut()
             .ok_or_else(|| CfdError::gpu_error("GPU buffers not initialized"))?;
 
         // Convert nalgebra matrices to flat vectors
@@ -163,7 +163,7 @@ impl SimpleGpuSolver {
     fn copy_from_gpu(&self, flow_field: &mut FlowField) -> CfdResult<()> {
         let buffers = self
             .gpu_buffers
-            .as_ref()
+            .as_mut()
             .ok_or_else(|| CfdError::gpu_error("GPU buffers not initialized"))?;
 
         // Copy from GPU
@@ -300,7 +300,7 @@ impl SimpleGpuSolver {
     async fn gpu_pressure_update(&self, pressure_relaxation: f32) -> CfdResult<()> {
         let buffers = self
             .gpu_buffers
-            .as_ref()
+            .as_mut()
             .ok_or_else(|| CfdError::gpu_error("GPU buffers not initialized"))?;
 
         // p = p + α_p * p'  (pressure update with relaxation)