[petlenz]

CI took 45–55 minutes per PR, almost all of it compiling: the tests themselves run in about 30 s. This keeps ASan+UBSan on every PR for both compilers, and makes it cheaper.

Where the time went

• numsim_add_test builds each of the ~78 test files twice: plain, and an ASan+UBSan copy. The sanitizer copy compiles about 1.9× slower.
• For one test file, the optimisation level makes no difference: -O0, -O1 and -O3 each take about 29 s under ASan.
• A precompiled header saves about 1.5 s of those 29 s.
• -ftime-report: 66 % of the time is code generation for instantiated templates (callgraph functions expansion alone is 48 %). Every test file instantiates the same engine, tmech and gtest templates again.
• Compiling 4 test files as one unit took 56 s, against 117 s for the 4 separately.

Change

1. Batched tests. add_numsim_test collects the single-file tests. Six at a time go into one generated translation unit (tests/batch_tests.cmake):

   • each file sits in its own namespace;
   • every #include of the batch is hoisted above the namespaces, so the inner ones are no-ops;
   • no test file changes, and the add_numsim_test(...) lines stay as they are, so open PRs that add tests merge without conflicts.

   Built standalone, as before:

   • tests with several sources (test_read_check_layout);
   • tests that need their own flags or check include hygiene (test_material_conformance, test_host_neutral_includes);
   • tests that define C entry points (NUMSIM_MATERIALS_DEFINE_* or extern "C"), whose symbols ignore namespaces. This is detected from the source.

   -DNUMSIM_MATERIALS_BATCH_TESTS=OFF gives one executable per test again, for TDD on one file. NUMSIM_MATERIALS_TEST_BATCH_SIZE sets the batch size (default 6).

2. CI.

   • The matrix is {g++-14, clang++-19} × {plain, asan}: four parallel jobs.
   • The ASan jobs build only the sanitizer executables (numsim_materials_asan_tests) and run ctest -R _asan$.
   • The plain jobs build with NUMSIM_SANITIZER_TESTS=OFF and also run the report tests.
   • ctest runs with -j$(nproc).

Checks

• Local gate, g++-14 -O1, sanitizers on, -j4: 1486 tests, 0 warnings, the same count as main. Wall time 25 min → 11 min.
• No two test files share a full test name, and no fixture suite spans two files, so the batched executables register the same tests as before.
• Peak compiler memory: the heaviest ASan batch (with test_mean_field_local) takes 3.6 GB. The runners have 16 GB for 4 parallel compiles. If CI shows memory pressure, the batch size goes down.

Merge order

Independent of the epic stack. #228 also edits ci.yml: whichever merges second needs a small conflict resolution, which I will do.

[petlenz]

First CI run:

Job	Time	Tests
clang++-19	5 min	751
g++-14	8 min	751
clang++-19, asan	13 min	735
g++-14, asan	16 min	735

Wall clock 16 min, down from 45–55 min. 751 + 735 = 1486: the same tests as before. The plain jobs also run the report tests. No memory problems with batch size 6.