[PoC] Parallel PHPUnit runner in Go (~4x faster) + fix phpVersion leak between test classes

[TomasVotruba]

Parallel PHPUnit runner in Go — measured ~4x local

Proof of concept: run the suite faster by splitting test classes across parallel workers that each boot PHP once and run many classes in a single process.

CI-measured (GitHub runners, 4 vCPU, full suite)

Same suite, serial vendor/bin/phpunit vs the Go runner, measured on the run step of this PR:

Platform	Serial	Go runner	Speedup
ubuntu-latest	31s	16s	~1.9x

Full suite (685 classes / 5833 fixtures) runs and passes. A fast_tests CI job (added here) runs the runner so these numbers are reproducible on every push.

Local speed (24-core host, full suite)

Mode	Wall time	vs serial
Serial vendor/bin/phpunit (1 process)	38.3s	1.0x
fast-phpunit -p 8	~11s	~3.5x
fast-phpunit -p 12	~9s	~4.2x
fast-phpunit -p 24	~9.8s	~3.9x

Speedup scales with cores; it flattens once the heaviest chunk bounds wall time.

Why it is faster

Bootstrap dominates a Rector test class, not the assertions:

Step	Time
Boot only (container build, 0 tests)	~0.23s — fixed, per process
One class, 27 fixtures (warm)	0.92s → ~26ms/fixture

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 actually slower than serial (5.71s serial vs 8.55s xargs -P8 on the CodeQuality subset). 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.

Isolation — required to make it correct

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

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

2. Leaked phpVersion() (in-process) — a latent bug, fixed here. phpVersion(...) is stored in the static SimpleParameterProvider and was never reset between classes. A version-bound class leaked its version into the next class in the same process:

 // NullToStrictIntPregSlitFuncCallLimitArgRector fixture — expected UNCHANGED,
 // but ran under a leaked PHP 8.1 instead of the test default (PhpVersion::PHP_10)
-preg_split('/\s/', $output, 0, PREG_SPLIT_NO_EMPTY);
+preg_split('/\s/', $output, NULL, PREG_SPLIT_NO_EMPTY);

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. Useful on its own, independent of the runner.

Scope

• utils-tests-runner/ is a standalone Go helper that shells out to php vendor/phpunit/phpunit/phpunit. It does not change how tests are written; it is invoked via one extra CI job. Pure Go, no .php, so it is invisible to ECS / PHPStan / Rector.
• Only production change: the phpVersion reset in AbstractRectorTestCase. ECS + PHPStan level 8 pass; full suite passes under -p 24 locally and on CI.

Open questions

• Keep the Go tool in-repo, or split the phpVersion reset into its own PR and drop the tool?
• Replace the serial tests job with the runner once trusted, or keep both for a while?