[cekees]

Two pieces of test-harness work, one PR.

1. Hardening for the macOS Intel runner

The runner is shared and slow, and the near-term hardware and OS picture is likely to stay unstable; the decision is to chase its failures down and harden the harness rather than retire the platform. Three failures there today were the harness, not the code under test:

• test_parallel_space_multigrid: the helper printed its complete RESULT line and then died on signal 11 at interpreter exit (PETSc/MPI teardown).
• test_without_petsc4py: a 4-rank mpiexec hit its 300 s budget inside a job that took 2 h 15 min for a suite that takes 15 min on Linux.
• test_launch_slots: its driver's own 120 s budget, same cause.

The wrapper that test/conftest.py already puts on every subprocess.run from a test (the relaunch of an mpiexec that never started, #95) now also:

• relaunches a command that died by a signal, once, and reports it as a RuntimeWarning naming the command, the signal and the tail of stderr, so the count stays visible in the summary instead of vanishing into a green run; a repeat fails as before;
• scales a test's timeout by PROTEUS_TEST_TIMEOUT_SCALE: a number, or auto for the one-minute load average per core clamped to [1, 4], read at launch;
• runs every launched interpreter with PYTHONFAULTHANDLER=1, so a crash in Python prints the stack it happened in. That, with the warning's stderr tail, is the material for chasing the exit crash down.

The workflow sets the scale to 3 on the Intel leg and auto elsewhere, and PYTHONFAULTHANDLER=1 for the xdist workers. Six tests with a fake launcher that finishes its work and then kills itself. The full suite through the new wrapper: 1,806 passed, no relaunch warnings.

2. Coverage of the proteus package

[tool.coverage] in pyproject.toml measures proteus/ and nothing else: not test/, not scripts/, not the decks under proteus/tests. Branch coverage. The point is to make code nothing runs visible, so the measurement has to reach where the suite's work actually happens: most of it is in processes the tests start (parun, the helpers under mpiexec). coverage 7.10's patch = subprocess does that without the old .pth trick, and it reaches grandchildren through mpiexec: 40 of the 44 lines of Archiver.allGatherIncremental are reached by the rank processes of the parallel-metadata tests. pytest-cov combines the xdist workers' and the children's data files, and [tool.coverage.paths] reports the installed package that CI runs as the checkout's files.

The linux-64 / py3.12 leg runs under coverage (one leg; measuring slows the run and the answer is the same everywhere) and keeps coverage.xml and the HTML report as job artifacts. Locally:

python -m pytest --cov=proteus --cov-report=term-missing:skip-covered ../test

Environment files: pytest-cov >= 7.0.0, coverage >= 7.10.

Tool choice, briefly. coverage.py with pytest-cov remains the one that handles this suite's shape: xdist workers, subprocesses, MPI grandchildren, and a combine step with path mapping. slipcover is faster but has no subprocess or combine story; Codecov or a badge can sit on top of the coverage.xml artifact later if wanted. Not measured: the Cython extensions (cfemIntegrals, cmeshTools, the mprans kernels). That needs a build with linetrace on and CYTHON_TRACE defined, which is a separate, slower build and a Cython.Coverage plugin line in the config; worth doing once the Python-level numbers have settled, since C-level dead code is a different hunt.

🤖 Generated with Claude Code