[derek73]

Summary

• Parallel by default. -n auto --dist worksteal in pyproject's addopts. On a 12-core Mac the full suite went from 110s serial to 35s (measured 2026-10-23). worksteal because the property grids are a few long tests among 23,000 short ones. With --dist loadfile, test_properties.py alone set the wall time (66s).
• Benchmark guards stay serial. The clock-timed scaling tests can't share the CPU: under -n auto, maiden_clause read 6.4× against its 6.0 bound on a tree that passes serially. tests/v2/test_benchmark.py is marked serial. tests/conftest.py deselects serial tests in a parallel run's workers and prints a summary line saying how to run them: uv run pytest -n0 -m serial. -n0 still runs everything in one process.
• CI runs two passes. The parallel suite, then -n0 -m serial. The coverage job adds --cov-append to the second pass, and the ja-extra job keeps NAMEPARSER_REQUIRE_CLOCK_GUARDS on both passes.
• Fixes a hidden order dependence that parallel runs exposed. tools/vocab/build.py read _lexicon.py from ROOT at first use. Three test_vocab.py tests repoint ROOT at a fixture tree, so each passed only after an earlier test had warmed the cache, and failed when run alone. The path is now bound at import, like VOCAB and OUTPUT. A new test fails against the old loader.
• Docs. AGENTS.md and CONTRIBUTING.md give both commands. AGENTS.md gains a gotcha on cache-warming order dependence, and drops the stale "runs every test twice" description of the conftest fixture.

Verification

• uv run pytest: 23,162 passed, 125 skipped, 3 xfailed in 35s.
• uv run pytest -n0 -m serial: 39 passed.
• uv run pytest -n0: 23,201 passed, which is the two passes combined.
• Also clean under -n 3 --dist load and -n 11, with every test module run on its own.
• Coverage: parallel plus the appended serial pass covers 4,625 of 4,669 lines, identical to one serial --cov run. That serial run took 4:45 under the 3.11 line tracer; the parallel one took 1:20.
• ruff check and mypy are clean.

Not verified locally: the ja-extra job's parallel pass, since namedivider isn't installed here, and CI's 4-vCPU runners, where the gain will be smaller than on 12 cores.

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.05%. Comparing base (04f806d) to head (15337b9).
⚠️ Report is 1 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #861   +/-   ##
=======================================
  Coverage   99.05%   99.05%           
=======================================
  Files          47       47           
  Lines        4669     4669           
=======================================
  Hits         4625     4625           
  Misses         44       44           

☔ 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.