ci: fix nightly quality cross-platform regressions

[antoneri]

Summary

This PR fixes the three failing job classes in the current Nightly Quality run on master.

It keeps the scope intentionally small:

• fix macOS C++ correctness configuration on Homebrew/OpenMP runners
• fix Windows C++ correctness when cmake/ctest resolve to paths containing spaces
• fix the sanitizer failures that become visible once the build-path issues are out of the way

Root Causes

1. macOS C++ correctness

mk/common.mk called scripts/build_config.py through $(shell ...) and passed CPPFLAGS/CXXFLAGS/LDFLAGS as CLI arguments. On the macOS nightly runner, that quoting path was unstable and repeatedly produced:

build_config.py: error: argument --cppflags: expected one argument

The result was an incomplete CMake configure on macos-15.

2. Windows C++ correctness

mk/test.mk invoked $(CMAKE) and $(CTEST) without quoting. On windows-2025, those variables resolved to paths under C:\Program Files\..., so Git Bash split the command at the space and failed before configure/build.

3. Sanitizers

The Linux sanitizer log showed a leak rooted in generateSubNetwork(Network&), and local sanitizer verification also exposed a second issue in the OpenMP move-commit path:

• InfoNode::deleteChildren() only deleted active children and ignored collapsedFirstChild / collapsedLastChild
• tryMoveEachNodeIntoBestModuleInParallel() read m_emptyModules.back() before checking that the vector was non-empty

Changes

• mk/common.mk
  • stop piping build flags through fragile CLI quoting
  • pass CPPFLAGS / CXXFLAGS / LDFLAGS / MACOSX_DEPLOYMENT_TARGET via environment instead
• scripts/build_config.py
  • read those values from the environment by default
• mk/test.mk
  • quote $(CMAKE) and $(CTEST) invocations so paths with spaces remain executable on Windows Git Bash
• src/core/InfoNode.cpp
  • make deleteChildren() also clean up collapsed child chains
• src/core/InfomapBase.cpp
  • treat collapsed root children as existing state when reinitializing from a Network
• src/core/InfomapOptimizer.h
  • guard the OpenMP empty-module check before reading m_emptyModules.back()
• test/cpp/test_partition.cpp
  • add a regression test that proves deleteChildren() clears collapsed children too

Verification

Local verification run after the final patch set:

• PATH="/opt/homebrew/bin:$PATH" make doctor MODE=release OPENMP=1 CXX=clang++ CPPFLAGS='-I/opt/homebrew/opt/libomp/include' CXXFLAGS='-I/opt/homebrew/opt/libomp/include' LDFLAGS='-L/opt/homebrew/opt/libomp/lib' MACOSX_DEPLOYMENT_TARGET=15.0
• PATH="/opt/homebrew/bin:$PATH" make test-native JOBS=1
• PATH="/opt/homebrew/bin:$PATH" make test-sanitizers JOBS=1
• PATH="/opt/homebrew/bin:$PATH" make test-native JOBS=1 CMAKE='/tmp/cmake space/cmake' CTEST='/tmp/cmake space/ctest' CTEST_ARGS='-R infomap_cpp_partition_tests'

That last command uses symlinks under a path containing a space to prove the quoted cmake / ctest invocations behave correctly.

Risk

This is still a CI-fix PR, so I’m keeping it draft.

The behavioral changes are intentionally narrow:

• Make/build-config plumbing only
• one bounded cleanup fix for collapsed children
• one bounded OpenMP safety guard in the existing parallel move path