…upations into onsite_proj_overlap.cpp

Step 2 of file split. Move the three overlap/occupation methods to
onsite_proj_overlap.cpp and emit them via method-level instantiation.
Slim the main onsite_proj.cpp includes (drop onsite_proj_print.h,
cell_tools.h, parallel_reduce.h, parameter.h, timer.h, <vector>) that
were only needed by the moved methods.

Verified by 'make -j 30' in build_max_para_test (built abacus_max_para).

…orce_stress.cpp

Step 3 of file split. Move the four cal_{force,stress}_onsite_{dftu,dspin}
wrappers to onsite_proj_force_stress.cpp. Main onsite_proj.cpp now
contains only the singleton accessor and destructor, with method-level
instantiation for those two. Drop the now-unused dftu_base.h and
<complex> includes from the main file.

Verified by 'make -j 30' in build_max_para_test (built abacus_max_para).

…tools_becp.cpp

Move the cal_becp method definition from onsite_proj_tools.cpp to a new
file onsite_proj_tools_becp.cpp with method-level explicit instantiation
for the CPU specialization. Add required includes (math_polyint.h,
math_ylmreal.h) for Nonlocal_maths dependencies. Update CMakeLists.txt.

Build verified: make -j 30 in build_max_para_test succeeds.

…roj_tools.cpp

Move cal_dbecp_s, cal_dbecp_f, save_vkb, revert_vkb, and transfer_gcar
method definitions from onsite_proj_tools.cpp to a new file
onsite_proj_tools_dbecp.cpp with method-level explicit instantiation
for the CPU specialization. Add required includes (math_polyint.h,
math_ylmreal.h, kernels/force_op.h for cal_vkb1_nl_op). Update CMakeLists.

Build verified: make -j 30 in build_max_para_test succeeds.

Move cal_force_dftu and cal_force_dspin to onsite_proj_tools_force.cpp,
and cal_stress_dftu and cal_stress_dspin to onsite_proj_tools_stress.cpp.
Both new files use method-level explicit instantiation for the CPU
specialization. Update CMakeLists.txt.

Build verified: make -j 30 in build_max_para_test succeeds.

Split the 827-line fs_nonlocal_tools.cpp into four functionally-cohesive
files:
- fs_nonlocal_tools.cpp (core): constructor, destructor, allocate_memory,
  delete_memory, whole-class template instantiation
- fs_nonlocal_tools_vkb.cpp: cal_vkb, cal_becp, reduce_pool_becp
- fs_nonlocal_tools_stress.cpp: cal_vkb_deri_s, cal_dbecp_s, cal_stress
- fs_nonlocal_tools_force.cpp: cal_vkb_deri_f, cal_dbecp_f, save_vkb,
  revert_vkb, transfer_gcar, cal_force

Each new file uses method-level explicit instantiation for the CPU
specialization. Update CMakeLists.txt.

Build verified: make -j 30 in build_max_para_test succeeds.
Tests verified: MODULE_PW_pw_test, MODULE_PW_pwdft_soc, onsite_op_test,
dftu_pw/core/operator, deltaspin_pw/core all pass.

Add a bullet to the Required Baseline section: for multi-step refactors
such as splitting a large .cpp into several files, build and commit after
each step rather than batching all changes before verification.

Extract the rho preparation phase (FFT, array allocation, gradient and
laplacian computation) from the 798-line gradcorr method into a separate
gradcorr_prepare_rho function (26 parameters, all by reference).

gradcorr shrinks from 826 to 657 lines. The orchestrator now declares
the shared arrays, calls prepare_rho, then proceeds to the XC kernel
and potential assembly (to be extracted in steps 2-3).

Update xc_functional.h, CMakeLists.txt, and test CMakeLists that
compile xc_grad.cpp directly.

Build verified: make -j 30 in build_max_para_test succeeds.

Extract the main XC computation loop (#pragma omp parallel + #ifdef __LIBXC
nspin branches + OpenMP reduction) from gradcorr into a separate
gradcorr_xc_kernel function (27 parameters, all by reference/pointer).

gradcorr shrinks from 657 to 355 lines. The orchestrator now:
1. Guards + setup
2. Calls prepare_rho (step 1)
3. Calls xc_kernel (this step)
4. [Laplacian stress + assembly — to be extracted in step 3]
5. Cleanup

Update xc_functional.h, CMakeLists.txt, and test CMakeLists.

Build verified: make -j 30 in build_max_para_test succeeds.

Extract the potential assembly phase (Laplacian stress contribution,
vxc from vtxcgc + dh + laplacian, noncolinear rotation back) from
gradcorr into a separate gradcorr_assemble_vxc function (24 parameters).

gradcorr is now a 183-line orchestrator that:
1. Guards + early returns
2. Setup (nspin0, fac, need_laplacian, stress_gga init)
3. Variable declarations
4. Calls prepare_rho (step 1)
5. Calls xc_kernel (step 2)
6. Calls assemble_vxc (this step)
7. Cleanup (delete arrays)

The original 798-line monolith is now split across:
- xc_grad.cpp (183 lines, orchestrator)
- xc_grad_prepare.cpp (222 lines, rho preparation)
- xc_grad_kernel.cpp (351 lines, XC computation loop)
- xc_grad_assemble.cpp (222 lines, potential assembly)

Build verified: make -j 30 in build_max_para_test succeeds.

- Replace 9 raw pointer allocations (rhotmp1/2, rhogsum1/2, gdr1/2, h1/2, neg) with std::vector, removing all manual delete[] in the orchestrator
- Flatten 2D arrays vsave/vgg (double**) into 1D vectors (vsave[is*nrxx+ir] index pattern), eliminating nested heap allocations
- prepare_rho now resizes vectors instead of new; kernel receives raw pointers via .data() (signature unchanged)
- assemble_vxc signature updated: double** vsave/vgg -> double* (flattened 1D)
- Build verified: make -j 30, 5 surchem tests passed

…iles

After splitting fs_nonlocal_tools.cpp and onsite_proj_tools.cpp into
multiple .cpp files with method-level explicit instantiation, only CPU
instantiations were added to the split files. The whole-class explicit
instantiation in the main file does not instantiate methods defined in
other translation units, causing undefined reference errors for
DEVICE_GPU specializations during GPU linking.

Add #if ((defined __CUDA) || (defined __ROCM)) guarded GPU method-level
explicit instantiations mirroring the existing CPU ones in:
- fs_nonlocal_tools_vkb.cpp (cal_vkb/cal_becp/reduce_pool_becp)
- fs_nonlocal_tools_stress.cpp (cal_vkb_deri_s/cal_dbecp_s/cal_stress)
- fs_nonlocal_tools_force.cpp (cal_vkb_deri_f/cal_dbecp_f/save_vkb/
  revert_vkb/transfer_gcar/cal_force)
- onsite_proj_tools_dbecp.cpp (cal_dbecp_s/cal_dbecp_f/save_vkb/
  revert_vkb/transfer_gcar)
- onsite_proj_tools_force.cpp (cal_force_dftu/cal_force_dspin)
- onsite_proj_tools_stress.cpp (cal_stress_dftu/cal_stress_dspin)

Verified with local GPU build (USE_CUDA=ON): abacus_pw_gpu links cleanly.

- xc_grad_assemble.cpp: replace 'double* dh = new double[]' with
  'std::vector<double> dh', pass .data() to grad_dot
- xc_grad_utils.cpp (grad_rho): replace gdrtmp raw pointer with
  std::vector<std::complex<double>>, use .data() for recip2real
- xc_grad_utils.cpp (grad_dot): replace aux and gaux raw pointers
  with std::vector<std::complex<double>>, use .data() for FFT calls
- xc_functional.cpp: minor comment typo fix (GGA, plane-wave)

All replacements are zero-overhead (vector uses contiguous memory
identical to new[]), exception-safe (auto-dealloc on scope exit), and
consistent with the previous RAII migration in xc_grad.cpp.

Verified: make -j 30 passes; ctest -R 'GRADCORR|VXC|PBE|SCAN|LAPL|HSE' 7/7 passed.

…o to_wannier90_pw_io.cpp

…al/get_trial/integral to to_wannier90_pw_setup.cpp

…o to_wannier90_pw_overlap.cpp, reduce original to 77 lines (ctor/dtor/calculate/set_tpiba)

…ing to symm_basic_order.cpp

…etgroup.cpp

…_pointgroup.cpp

…er include

…rnels.cpp with method-level instantiation

… to atom_pair_output.cpp

…container_insert.cpp, update 8 CMakeLists

…lue_union/shape_synchron to hcontainer_math.cpp

…/Sq/V+cal_flq+cal_ylm+cal_gpow to numerical_basis_overlap.cpp

…overlap_Q/Sq/V to numerical_basis_output.cpp

…ing) to force_pw_driver.cpp; each TU instantiates Forces template class

….cpp; each TU instantiates Forces template class

… assign_dmk_ptr + ctor/dtor to force_stress_lcao_driver.cpp

…ress_lcao_pw.cpp; each TU instantiates Force_Stress_LCAO template class

…lex<double>> to force_stress_lcao_integral.cpp

…metry to force_stress_lcao_utility.cpp; base .cpp now is a thin header-forward file (31 lines includes only). Total across 5 files ~1098 lines (original 1080).