Take SymbolExtractor's declarations from the scanner

[Firehed]

SymbolExtractor walked the AST itself to decide what a file declares, while every other consumer asked DeclarationScanner. Two opinions about the same question is how a function_exists-guarded polyfill came to resolve on hover while being invisible to completion (SC.5), and how its class_exists twin dropped out of open-document lookup (SC.9). This is the last of those in src/, so no consumer can now disagree about what a file declares.

Slice SC.3. Plan step: 0002 Step 3a teardown. RFC 1 §4.2, §8.1.

What changed

SymbolExtractor no longer extends NodeVisitorAbstract or builds a NodeTraverser. It reads DeclarationScanner's class-likes and functions, and takes each class-like's methods from ClassLike::getMethods() on the node the scanner already hands back.

Deleted with it: the hand-tracked $namespace field, the Stmt\Namespace_ branch, the leaveNode reset, and the three places an FQN was rebuilt by string concatenation against a namespacedName the parser had already computed.

The one thing added is a sort. The scanner returns class-likes and functions as separate lists, and callers read $symbols[0] as the file's first declaration, so the output is ordered by position rather than by kind.

Behaviour

Preserved, and proven by tests that already existed rather than new ones: WritePathParityTest and PrefixSearchParityTest both drive a real SymbolExtractor and passed without regenerating a golden, alongside SymbolExtractorTest's per-kind cases.

One latent bug goes away as a side effect. leaveNode reset the current class to null rather than to the enclosing one, so after a nested class-like closed, the outer class's remaining methods stopped being indexed. Methods now come from each class-like's own node, so nesting cannot drop them.

Baseline

phpstan-baseline.neon 43 → 39. The four drained are exactly this file's frozen AST-traversal violations; bin/check-baseline-shrink records the shrink. The confinement rule in phpstan.neon was already enforcing this — the entries were what kept it quiet.

Checklist

• SymbolExtractor no longer references NodeVisitorAbstract, NodeTraverser or NodeFinder
• Class-like kind mapping and Class::method FQN strings unchanged, pinned by the existing tests
• Its four baseline entries drained, not relocated
• Parity goldens pass unregenerated
• No new fixtures, no new enforcement — the rule for this already exists

Notes

Two manifest rows are now wrong and neither is edited here, to keep this diff to one file. SC.3's own note says Class::method FQNs "are its own and stay", which read as blocking — getMethods() is the route it missed. SC.10 is filed as "the invariant has no mechanism"; phpstan.neon has enforced it all along, so that row has no work in it.

Candidate closes (pending review verification): none.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.61%. Comparing base (d101d21) to head (2aea286).
✅ All tests successful. No failed tests found.

Additional details and impacted files

@@             Coverage Diff              @@
##               main     #418      +/-   ##
============================================
- Coverage     98.62%   98.61%   -0.02%     
+ Complexity     1814     1805       -9     
============================================
  Files           128      128              
  Lines          4571     4534      -37     
============================================
- Hits           4508     4471      -37     
  Misses           63       63              

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.