Run the PHPUnit suite ~4x faster by splitting test classes across parallel workers that each boot PHP once and run many classes in a single process.

Serial vendor/bin/phpunit vs this runner, measured on the CI run step:

Windows is the slow platform (~3.7x slower than Ubuntu serial); the runner cuts ~49s off it. Both run the full suite and pass.

Full suite: 685 test classes / 5833 fixtures.

Subset — rules-tests/CodeQuality (86 classes / 771 fixtures):

Sweet spot is ~12 workers; more does not help, because the heaviest chunk bounds wall time and 20+ concurrent PHP processes start contending.

Bootstrap dominates a Rector test class, not the assertions:

Average class has ~7 fixtures, so bootstrap is ~56% of an average class's run time. Any runner that spawns a fresh process per class pays that 0.23s boot 685 times — which is why process-per-class parallelism is slower than serial (see subset table).

This runner splits classes into N chunks balanced by fixture count, and each worker runs its whole chunk in one warm process — so the container is built N times, not 685 times.

cd utils-tests-runner && go build -o fast-phpunit .
cd ..
utils-tests-runner/fast-phpunit -p 12                          # whole suite
utils-tests-runner/fast-phpunit -p 8 rules-tests/CodeQuality   # a subtree

Flags: -p workers (default = CPU count), -bin phpunit path.

Two shared-state issues surface when many classes share a process; both are handled so any chunking is safe.

1. Shared temp cache (cross-process). Rector caches parsed files under sys_get_temp_dir()/rector_cached_files; parallel processes racing that directory throw Failed to open directory / Directory not empty. Each worker gets its own TMPDIR.

2. Leaked phpVersion() (in-process) — a latent bug, fixed here. phpVersion(...) is stored in the static SimpleParameterProvider and was never reset between classes, so a version-bound class leaks its version into the next class in the same process — a version-less class then sees, e.g., PHP 8.1 instead of the test default (PhpVersion::PHP_10) and produces wrong output. The serial suite passes only because of its class ordering; any reshuffle (this tool or paratest) can trigger it. Fixed in AbstractRectorTestCase::tearDownAfterClass() by resetting PHP_VERSION_FEATURES to the test default.

Proof of concept. Standalone Go helper that shells out to the existing vendor/bin/phpunit, so it does not change how tests are written or how CI runs. Pure Go, no .php, so it is invisible to ECS / PHPStan / Rector.