pure subroutine fill_u_centre_impl(u_face, buf)
! assumed-shape-ok: diag fill — fires once per output frame (cadence-bounded);
! face-sized u_face(:,:,:) has nx+1 first dim, incompatible with cell-sized buf.
real(wp), intent(in) :: u_face(:, :, :)
real(wp), intent(inout) :: buf(:, :, :) ! assumed-shape-ok: diag fill — cadence-bounded
integer :: i, j, k, nx, ny, nz
nx = min(size(buf, 1), size(u_face, 1) - 1)
ny = min(size(buf, 2), size(u_face, 2))
nz = min(size(buf, 3), size(u_face, 3))
do concurrent(k=1:nz, j=1:ny, i=1:nx)
buf(i, j, k) = 0.5_wp*(u_face(i, j, k) + u_face(i + 1, j, k))
end do
end subroutine fill_u_centre_impl