Narrow the match subject when its assigned result is later narrowed

[phpstan-bot]

Summary

A narrowing applied to a variable assigned from an if/elseif/else chain is projected back onto the condition subject (e.g. after $class = ...; if ($class === null) return;, the subject $x is narrowed to the values that produced a non-null $class). The same did not happen when the value was produced by a match expression — the subject stayed mixed. This change makes $var = match (...) build the same subject↔result relationship as the if form.

Changes

• src/Analyser/ExprHandler/MatchHandler.php: extracted getArmScopesAndTypes() from resolveType(). It returns, per reachable arm, [scope, bodyType] where scope has the match subject narrowed to that arm's condition. resolveType() now unions the body types from this method. The enum fast path is preserved, so huge-enum matches keep their performance.
• src/Analyser/ExprHandler/AssignHandler.php: when the assigned expression of a $var = ... is a Match_, call processMatchForConditionalExpressionsAfterAssign(). It assigns each arm's body type to $var inside that arm's subject-narrowed scope, then folds the per-arm scopes together with MutatingScope::mergeWith($scope, true) in the same order an if/elseif/else chain would. The conditional-expression holders that fall out of the merge are lifted onto the real scope. Added a small mergeConditionalExpressions() helper. MatchHandler is now injected into AssignHandler.

Root cause

PHPStan records "conditional expression holders" that capture relationships like "when $class is 'some_class', $x is 'aa'". For if/elseif/else, these are produced automatically by MutatingScope::mergeWith() / createConditionalExpressions() because the assignment $class = … happens inside each branch scope where the subject is already narrowed. A match expression resolves to a single union before the assignment, so the assignment to $class happens outside the per-arm narrowed scopes and the relationship was lost. The fix reconstructs the per-arm narrowed scopes, performs the assignment inside them, and reuses the exact same merge machinery — guaranteeing match and if produce identical holders.

Test

tests/PHPStan/Analyser/nsrt/bug-14772.php covers, with assertType():

• the reported case (mixed subject, narrowing the result to non-null narrows the subject to 'aa'|'bb'|'cc');
• narrowing the result to a single arm value ('some_class' → 'aa');
• match (true) with === comparison arms;
• int subject;
• a non-null default arm (narrowing to the default value yields mixed~('aa'|'bb'));
• two arms sharing the same body type (1|2);
• an enum subject (Suit::Hearts).
  Each assertion fails before the fix (the subject stays mixed/int) and passes after, matching the if equivalent.

Fixes phpstan/phpstan#14772
Fixes phpstan/phpstan#10843

[ondrejmirtes]

Add regression test for phpstan/phpstan#10843

[phpstan-bot]

Done. I addressed the review by adding the requested regression test for issue #10843.

What I did

Verified #10843 is actually fixed by this PR. Issue #10843 is the reverse direction of the PR's main case: instead of narrowing the result from the subject, here narrowing the subject ($date) should narrow the result ($a). I confirmed the PR's conditional-expression machinery handles both directions:

• Before the fix (parent commit's source): $a inside if ($date instanceof HelloWorld) is 1|false
• After the fix: $a narrows to 1

So this is a genuine regression test — it fails before the fix and passes after.

Added tests/PHPStan/Analyser/nsrt/bug-10843.php — the issue author's verbatim reproducer (the match (true) { $date instanceof HelloWorld => 1, default => false } case), with an assertType('1', $a) inside the if ($date instanceof HelloWorld) branch. nsrt files are auto-discovered by NodeScopeResolverTest.

Verified: NodeScopeResolverTest passes (1634 tests, including the new file). The change is a test-data file only — no src/ changes — so PHPStan self-analysis is unaffected.

Committed as a separate logical change ("Add regression test for #10843") and pushed to the PR branch.