fix(php): bind namespaced receivers in ::class and constructor position

[mwebber92]

Problem

PHP writes every namespaced name as a qualified_name node. Two extractor paths accept only the simple forms, so a class referenced through a namespace produces no edge and no unresolved_refs row — there is no trace of it anywhere in the index, and impact truthfully reports the class has no consumers.

extractStaticMemberRef accepts a receiver only when its node kind is identifier | type_identifier | simple_identifier | name | scoped_type_identifier. None of those is what a namespaced receiver produces, so all of these are dropped:

use App\SoapTypes as Type;

Type\Bankverbindung::class;                 // alias
\App\SoapTypes\Bankverbindung::class;       // fully qualified
SoapTypes\Bankverbindung::class;            // namespace-relative
\App\Models\User::TABLE;                    // any ::CONST, same shape

A bare Bankverbindung::class is a name and was always handled — which makes this easy to miss.

extractInstantiation strips a . or :: qualifier but not PHP's \, so new App\SoapTypes\Bankverbindung() pushes the unresolvable literal App\SoapTypes\Bankverbindung.

The diagnostic tell is that a type hint binds while ::class on the next line does not:

public function hint(Type\Bankverbindung $x) {}                // ✅ references edge
public function make() { return Type\Bankverbindung::class; }  // ❌ nothing

Fix

walkPhpTypePosition already handles qualified_name correctly, with the comment "match on the trailing simple name — what the class node is stored as, and what a use import brings into scope." This applies the same rule in the two places that lack it.

Both walkers, in two commits. codegraph-kernel/src/php.rs carried the identical pair of defects, and since PHP is in DEFAULT_ROUTED the kernel is what actually runs — fixing only the TypeScript side would leave a default build's output unchanged. The Rust commit needs no language gating: php.rs is PHP-only and carries its own private copy of strip_generic_and_qualifier rather than sharing one.

The TypeScript changes are gated to PHP: C# also has a qualified_name node and is in STATIC_MEMBER_LANGS, so an ungated change would add unrelated C# edges, and a backslash carries no qualifier meaning in the other languages sharing the instantiation path.

kernel-php-parity needed no fixture change — it compares the two walkers against each other at runtime rather than against goldens, so it stays green once both agree. Its docstring described the old instantiation shape and is updated to match.

This makes the alias moot rather than resolved — Type\Bankverbindung and \App\SoapTypes\Bankverbindung both reduce to Bankverbindung — so a same-named class in another namespace stays ambiguous here, exactly as it is for a type hint today. Resolving aliases properly is a separate change; extractPHPImports records use X\Y as Z with exportedName set to the last segment as though it were a class, so it cannot expand a namespace alias.

Evidence

Measured on real PHP codebases (Symfony-style SOAP clients and WordPress plugins), 12 indexes:

index	references + instantiates edges
A	206 → 1,013
B	426 → 728
C	1,107 → 1,628
D	3,122 → 4,151
E–H (4 similar)	~4,400 → ~6,900 each
total	22,485 → 34,336 (+53%)

Node count and every other edge kind are unchanged — the change is purely additive. On one repo, impact for a class used through an alias went from 3 of 6 consumer files to 6 of 6.

Re-indexing all twelve a second time through the kernel rather than the WASM walker reproduced 34,336 exactly, every index identical — a broader parity check than the three torture fixtures.

Tests

__tests__/php-qualified-static-member-refs.test.ts — five cases: aliased, fully-qualified and namespace-relative ::class, a qualified constructor, and a negative case pinning the capitalization guard. Each block names the mutation it catches.

Verified they can fail: with the test file in place and the tree-sitter.ts changes reverted, 4 go red and the negative case stays green.

Full suite with the kernel staged, so every kernel-*-parity suite actually runs rather than skipping: 3,234 passed, 0 failed on the v1.6.0-based branch. On main it is 4,336 passed / 3 failed, and those three (object-literal-methods, ui-steps-api ×2) fail identically on clean main — pre-existing and unrelated.

Cargo release build of the kernel is ~60 s warm, ~110 s cold.