Fix CI ccache SIGILL by building Linux for x86-64-v3

[zfergus]

Description

Builds the Linux CI jobs for x86-64-v3 instead of -march=native, so one ccache serves every runner CPU. Also fixes CI comments that named MeshFEMCore as a source of -march=native.

• Pinned baseline: continuous.yml, cuda.yml, python.yml, and coverage.yml (Linux) pass -DSIMD_CXX_FLAGS=-march=x86-64-v3 (AVX2, FMA, BMI1/2, LZCNT, F16C). The flag string ccache hashes now fully describes the generated code.
• One cache: the continuous, cuda, python, and coverage cache keys have no CPU component on any OS, so one cache per OS and config serves every runner CPU (for example <os>-<config>-cache, with a cuda- prefix for CUDA). The CPU fingerprint steps are gone, apart from a CPU log line in the Linux guard.
• Linux wheels: pypi.yml passes CMAKE_ARGS=-DSIMD_CXX_FLAGS=-march=x86-64-v3 through CIBW_ENVIRONMENT_LINUX. Without it, wheels take -march=native from the runner that built them. The 1.6.0 Linux wheels contain no AVX-512 but already need AVX2, FMA, BMI and LZCNT. macOS and Windows wheels are unchanged.
• Docs: docs.yml gets its ccache back (reverting Fix docs SIGILL by dropping ccache from the docs build #270's removal) with the same x86-64-v3 baseline through CMAKE_ARGS and the same guard. It uses a new key, docs-ccache, because entries are immutable and the old docs-cache holds -march=native objects.
• Debug without -g: the Debug configures in continuous.yml (Linux and macOS) and cuda.yml pass -DCMAKE_C_FLAGS_DEBUG=-O0 -DCMAKE_CXX_FLAGS_DEBUG=-O0, so asserts stay on (NDEBUG undefined) but debug info is gone. In the first Fix CI ccache SIGILL by keying on the resolved ISA #269 run the Debug ccache was 0.41 GB against 0.02 GB for Release on Linux, and 0.53 GB against 0.14 GB on CUDA. The CUDA job empties CMAKE_CUDA_FLAGS_DEBUG instead, because nvcc would apply -O0 to device code. Windows is unchanged, since MSVC's Debug flags differ.
• SIMD flag validation: when SIMD_CXX_FLAGS is already set, CMakeLists.txt skips find_package(SIMD) and only checks that the compiler accepts the flags, through the new ipc_toolkit_check_cxx_flags() in ipc_toolkit_filter_flags.cmake. Rejected flags are a FATAL_ERROR instead of a silent fallback to a non-SIMD build, so a typo can't pass as a working build. Auto-detection is unchanged when the flags are not set.
• Guard: the Linux jobs fail with a clear error if the runner CPU lacks an x86-64-v3 feature, instead of a SIGILL later.
• Coverage: coverage.yml gets the same baseline, guard, and plain coverage-cache key. Its stray trailing backslashes are gone.
• Comments: the comments in all five workflows now name FindSIMD and Tight-Inclusion. MeshFEMCore's CMake never runs, since meshfem_sparse.cmake uses DOWNLOAD_ONLY.
• Python on Linux: python.yml no longer sets IPCTK_WITH_SIMD=0 on Linux (added in Add env var to disable SIMD in setup.py #231 with no stated reason). It builds with SIMD and the same x86-64-v3 pin and guard, so the Linux Python tests now exercise the SIMD code through the bindings.

Impact:

• The first run after merge is cold on every OS, because every key changes. After that the cache is shared across CPUs. Only 16 of 40 recent Linux jobs restored a cache, because keys were split across six CPU models.

• No Linux CI job compiles the -march=native default or the AVX-512 xsimd path (batch width 8 instead of 4) any more. Users on AVX-512 hardware run code CI no longer builds.

• No source or API changes.

• Windows already passes explicit /arch:AVX2, so its objects do not depend on the runner.

• macOS still builds with -march=native. It shares one cache because 27 of 29 recent macOS jobs ran on the same Apple M1. A cache seeded by an M2 Pro runner could in principle reach M1 runners.

Why

• SIGILL after restoring a cache on EPYC 9V74 runners: Extremum Sum Potential #259, Fix CMake options being forced in the cache #268, and the docs build (Fix docs SIGILL by dropping ccache from the docs build #270).
• ccache hashes -march=native literally. On an Ice Lake runner it expands to -march=icelake-server -mavx512f …, and the cached objects are only safe on a runner with the same ISA.
• Tight-Inclusion's own flags are -march=native -mavx2 -mno-avx512f …, which caps its objects at AVX2. I left them as they are. A switch there would make them strictly deterministic.
• With -DSIMD_CXX_FLAGS=-march=x86-64-v3, all 94 ipc_toolkit translation units compile with it and none with -march=native.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Enhancement (non-breaking change which improves existing functionality)

How Has This Been Tested?

• All five workflows parse as YAML, and every non-Windows run block passes bash -n.
• The guard was simulated against fake /proc/cpuinfo files: it passes on a v3-capable CPU and fails with a clear message when bmi2, fma, or avx2 is missing.
• Configured locally with -DSIMD_CXX_FLAGS=-march=x86-64-v3 and checked compile_commands.json (see above). Configure only, no build.
• Not tested: the wheel build. pypi.yml only runs on workflow_dispatch, and I haven't checked that cibuildwheel passes CMAKE_ARGS into the manylinux container.
• Not tested yet: the Linux jobs themselves, including the nvcc host pass accepting -march=x86-64-v3 through -Xcompiler. CI on this PR is the test, since each workflow triggers on its own file. A re-run after the first shows the cross-CPU cache hits.

Test Configuration:

• OS and Version: macOS 26.6.2
• Compiler and Version: Apple clang 21.0.0 (configure only)

Checklist

• I have followed the project style guide
• My code follows the clang-format style guidelines of this project
• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• Any dependent changes have been merged and published in downstream modules

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 96.92%. Comparing base (101332c) to head (93b414e).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #272      +/-   ##
==========================================
- Coverage   96.92%   96.92%   -0.01%     
==========================================
  Files         194      194              
  Lines       17348    17349       +1     
  Branches      944      943       -1     
==========================================
  Hits        16815    16815              
- Misses        533      534       +1     

Flag	Coverage Δ
unittests	96.92% <ø> (-0.01%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.