[pawell67]

A static<...> return type whose arguments reference no template type is not resolved on a generic receiver:

/**
 * @template TKey of array-key
 * @template TValue
 */
class Box
{
    /** @return static<int, string> */
    public function concrete() { return new static(); }

    /** @return static<int, TValue> */
    public function withTemplate() { return new static(); }

    /** @return array<TKey, TValue> */
    public function all(): array { return []; }
}

/** @param Box<string, bool> $box */
function test(Box $box): void {
    \PHPStan\dumpType($box->concrete()->all());     // array<TKey of (int|string), TValue>   (expected array<int, string>)
    \PHPStan\dumpType($box->withTemplate()->all()); // array<int, bool>
}

CalledOnTypeUnresolvedMethodPrototypeReflection::transformStaticType() turns the GenericStaticType into getStaticObjectType(), which hands the declaring class' ClassReflection (without the type arguments) to the GenericObjectType. When an argument contains a template, traverse() recreates the object and the reflection is resolved lazily, which is why only the fully concrete case breaks. I reproduced this on 2.2.14 and 2.3.0, so it is not a regression.

It is visible in Laravel's Collection (@return static<int, mixed> on flatten(), collapse(), values(), …): $collection->flatten()->merge($x) leaks TKey/TValue, which Larastan has to baseline in its monica E2E.

The fix builds the object without the stale reflection in transformStaticType() only. I first changed GenericStaticType::getStaticObjectType(), but that also changes new static(...) (NewHandler) and drops the expected error of ReturnTypeRuleTest::testBug4590, so I narrowed it to the method return type resolution.

The turbo-ext mirror of CalledOnTypeUnresolvedMethodPrototypeReflection is ported in the same change, and the new nsrt files pass with and without the extension loaded.
Closes phpstan/phpstan#12680

🤖 Generated with Claude Code

[pawell67]

The last commit is a cherry-pick of #6697 (author kept): ResultCacheInfoCommandTest::testWarmAnalysisDoesNotRewriteUnchangedResultCache fails in the ubuntu "Integration tests" job regardless of this change, so it is included here only to get a green run. It drops out of the diff once #6697 is merged and this branch is rebased.

[pawell67]

Rebased on the current 2.3.x (the #6697 cherry-pick is gone, it landed upstream) and bumped the expected turbo version for the ported commit.

The last commit is a cherry-pick of #6729 (author kept): the result-cache-parse-error E2E step fails on 2.3.x since the BetterReflection update, independently of this change. It is here only to get a green run and drops out once #6729 is merged and this branch is rebased.

[pawell67]

@staabm would you mind taking a look? With phpstan/phpstan#15468 merged, CI here is fully green now. The fix is in CalledOnTypeUnresolvedMethodPrototypeReflection::transformStaticType() plus its turbo-ext mirror: a fully concrete static<int, string> return type kept the declaring class reflection, so calls on the result saw unresolved TKey/TValue. The last commit is the cherry-pick of #6729 and drops out once that is merged.

[staabm]

why do we need to adjust this tests expectations?

[pawell67]

This is not part of the fix: it comes from the last commit, a cherry-pick of #6729 (author kept), because the result-cache-parse-error E2E step fails on 2.3.x since the BetterReflection update, independently of this PR. It drops out of the diff once #6729 is merged and this branch is rebased, so please review it there.

[pawell67]

@staabm fixed. Removed not related changes

[staabm]

@pawell67 please fix

1) PHPStan\Build\RequiredPhpVersionCommentTest::testFixtureHasRequiredLintComment@/home/runner/work/phpstan-src/phpstan-src/tests/PHPStan/Build/../Rules/Functions/data/bug-11507.php with data ('/home/runner/work/phpstan-src...07.php')
Fixture uses promoted properties on line 14 which requires PHP 8.0. Add a `<?php // lint >= 8.0` comment on the first line so the fixture is skipped on older PHP versions in CI.
Failed asserting that 0 is equal to 80000 or is greater than 80000.

---

//cc @SanderMuller please review

[SanderMuller]

I checked 2846823 against 2.3.x (5fb9346).

• The example from the description gives array<int, string> with this PR, and the base gives array<TKey of (int|string), TValue>. I also probed a subclass that binds the templates, a generic subclass, a union receiver and new Box(). Two more probes used a nested static<int, Box<int, TValue>> and a template method that returns static<TKey|int, TValue|TOther>. Every result that changes is now the resolved type, and nothing changes for the worse. For GSub<float>, $g->concrete() is GSub<string> on both sides. But ->all() on it gave array<int, float> on the base, and it now gives array<int, string>.
• Both new nsrt files fail with the change in CalledOnTypeUnresolvedMethodPrototypeReflection.php reverted. bug-12680.php matches Ondřej's playground snippet, and the reporter's snippet in phpstan/phpstan#12680 now gives mixed for both lines instead of TValue.
• Turbo: I built the extension from this head (macOS arm64, PHP 8.5). smoke.php and signature-parity.php pass, and the two nsrt files pass with the extension loaded. With the extension active and the PHP change reverted, my probe still gives the fixed types, so the C++ port carries the fix by itself. I did not build it for other platforms or PHP versions.
• The full paratest suite passes with and without the extension (22615 tests). make phpstan is clean, and so is phpcs on the changed source file.
• Performance: cold self-analysis without the extension, 3 interleaved rounds. The median user CPU is 120.9 s on the base and 121.2 s with this PR. The runs took 119.7 to 128.9 s and 119.0 to 130.2 s, so there is no measurable difference.
• CI: the red jobs also fail on #6782, #6783 and #6785. The Turbo compile jobs fail in zend_globals_macros.h ('ZEND_API' does not name a type), and d78c82f on 2.3.x targets that. Its CI was still running when I checked. result-cache-parse-error fails on every current 2.3.x PR. 87bbf77 on 2.3.x adds the same lint comment to bug-11507.php, so after a rebase that line drops out of the diff.

Two small things in the description. It still says the C++ mirror is not ported and the suite did not run with the extension, but c6d2009 ports it. And phpstan/phpstan#12680 is still open, so a Closes line could go in if this is the fix for it.

[staabm]

thank you!