Re-register the DI container in setUp() of tests with PHP-version-dependent reflection assertions - fixes racy/flaky tests

[phpstan-bot]

Summary

HasPropertyTypeTest::testIsSuperTypeOf was intermittently failing in CI on the
hasProperty(foo) -> isSuperTypeOf(Closure) data set with Expected 'Maybe', Actual 'No'.

The data set's expected value is PHP_VERSION_ID < 80200 ? Maybe : No, while the actual
value comes from ObjectType(Closure)->hasInstanceProperty('foo'), which for a final class
returns Maybe/No depending on ClassReflection::allowsDynamicProperties() — and that is
version-gated (PhpVersion::deprecatesDynamicProperties(), i.e. >= 8.2).

These tests construct Type objects and call their methods directly, so reflection is
resolved through the global ReflectionProviderStaticAccessor. When another test leaves a
container configured with a different PhpVersion registered there, the Closure
ClassReflection is built with that foreign version and the result no longer matches
PHP_VERSION_ID. An existing PHPUnit event subscriber re-initializes the container before
each test, but it does not fire reliably under the actual test runner (paratest), so the
flake persisted. setUp() is a runner-agnostic guarantee.

Changes

• tests/PHPStan/Type/Accessory/HasPropertyTypeTest.php
  • Add setUp() calling self::getContainer() to re-register the default (runtime)
    container before every test.
  • Add a deterministic regression test testIsSuperTypeOfClosureRespectsActivePhpVersion
    that asserts Maybe under PHP 8.1 and No under PHP 8.2 by building containers with
    explicit PhpVersions, restoring the default container in a finally block so no
    global state leaks.
• tests/PHPStan/Type/ObjectTypeTest.php, tests/PHPStan/Type/TypeCombinatorTest.php,
  tests/PHPStan/Type/Generic/GenericObjectTypeTest.php
  • Add the same setUp() re-registration. These are the sibling type unit tests that call
    Type methods directly and have version-dependent assertions (dynamic-property handling
    of final classes / reflected variance of built-in generics), so they share the same
    latent flakiness.

Root cause

A family of type unit tests assert PHP-version-dependent reflection results (computed from
the active container's PhpVersion) against expectations derived from the global
PHP_VERSION_ID constant. These two sources diverge whenever the globally-registered
reflection provider belongs to a container whose PhpVersion differs from the runtime —
which happens when another test registered such a container and nothing re-registers the
default container before the test executes. Calling self::getContainer() in setUp()
pins the runtime container before each test, eliminating the coupling.

Tests that go through the analyser (RuleTestCase, TypeInferenceTestCase) already call
self::getContainer() during their analyse/assert flow, so they are not affected.

Test

• testIsSuperTypeOfClosureRespectsActivePhpVersion deterministically verifies the
  version-dependent behavior on every runtime by switching the active container's
  PhpVersion.
• Verified the flake reproduces: with the container-init subscriber disabled and setUp()
  removed, the Closure data set fails (Expected 'No', Actual 'Maybe') once a foreign
  container is registered; re-enabling setUp() fixes it without the subscriber.
• Probed HasMethodTypeTest (the closest sibling): its Closure case is No regardless of
  version, so it is not affected and was left unchanged.

Fixes phpstan/phpstan#14860

[VincentLanglet]

Should the fix be in the base TestCase then @staabm ?

[staabm]

my understanding is, that the lazy construction of the DI container is a performance optimization.
if we add the suggested workaround to the base TestCase we will kill the optimization in all tests.

[VincentLanglet]

What about mooving this logic in the PHPStanTestCase setup inside a condition (set to false by default). This way it's easy to enable on the right files

[phpstan-bot]

Done. I addressed VincentLanglet's review by moving the container re-registration logic into the base PHPStanTestCase, gated behind an opt-in flag.

What changed

src/Testing/PHPStanTestCase.php — added a protected bool $reinitializeContainerBeforeEachTest = false; property and a setUp() that re-registers the runtime container only when the flag is enabled. The default (false) preserves the lazy-container performance optimization staabm flagged in the discussion — the work only happens in tests that explicitly opt in, not in every test.

The four affected type unit tests (HasPropertyTypeTest, ObjectTypeTest, GenericObjectTypeTest, TypeCombinatorTest) — replaced their duplicated setUp() methods (and the now-unused Override imports) with a single line: protected bool $reinitializeContainerBeforeEachTest = true;. The explanatory comment stays so it's obvious why each one opts in.

Seven other subclasses that override setUp() — added parent::setUp(). This was a required follow-up: once the base class defines setUp(), PHPStan's own phpunit.callParent rule (correctly) flags any override that doesn't call the parent.

Verification

• make tests — 17446 tests pass (96 skipped, as usual)
• make phpstan — no errors
• make cs — no errors

Committed as two logical commits (the refactor, then the parent::setUp() follow-ups), rebased onto the force-updated remote branch, and pushed.

One trade-off worth noting for the reviewer: making this opt-in via the base class means every existing setUp() override in the suite now needs parent::setUp() to stay correct — that's why the second commit touches seven unrelated test files. That's the cost of centralizing it versus the previous per-file approach.

[VincentLanglet]

WDYT of this way @staabm ?

Broken build

[VincentLanglet]

Broken build

Which one ?

Extensions need to be updated after