Fix stale install dirs of fetched dependencies; split the Gaussian-nucleus quadrature at the nuclear scale

[susilehtola]

Two independent fixes, one commit each.

1. Fetched dependencies' install locations are no longer frozen by an old build tree

A build tree first configured before the private dependency layout (#354) kept installing OpenOrbitalOptimizer's package config to lib64/cmake/OpenOrbitalOptimizer. helfemConfig.cmake only looks under lib/helfem-deps, so find_package(helfem) failed for every consumer of the install.

The cause: OpenOrbitalOptimizer computes OpenOrbitalOptimizer_INSTALL_CMAKEDIR from CMAKE_INSTALL_LIBDIR once, caches it, and from then on reads the cached value back. Eigen does the same with INCLUDE_INSTALL_DIR, CMAKEPACKAGE_INSTALL_DIR and PKGCONFIG_INSTALL_DIR. Those are generic names, so they also landed in an including project's cache. libxc does the same with FORTRAN_MODULE_INSTALL_DIR. helfem_fetch_privately now takes these variables after an INSTALL_DIRS keyword and clears them both before and after the fetch.

To verify, I seeded the stale values on the command line and hid the system Eigen, wignernj and libxc, so all five dependencies were fetched:

• master still uses lib64/cmake/... and keeps all four variables cached;
• with this change, every package config installs under lib/helfem-deps and nothing installs outside the expected directories;
• two consumers of the install build and run.

2. Gaussian nucleus: the quadrature is split at the nuclear scale

-Z erf(μr)/r turns over at r ~ 1/μ, about 1e-4 bohr, which is a thousand times smaller than the first element. The order refinement only resolved that by brute force. For Ne with the Visscher-Dyall Rrms (20-node LIP, 6 elements) it hit the order cap (n=512, relative change still 2.1e-12).

GaussianNucleusT now reports breakpoints at 2^k/μ, doubling until erfc(2^k) < eps(T). Beyond the last one the integrand is polynomial again. The same calculation now prints no warning and converges to the same energy, -128.5470607586 Eh. The spherical and hollow nuclei already split at their radius.

ctest -L unit passes (11/11).

🤖 Generated with Claude Code