feat(extraction): instantiates + decorates graph edges

[andreinknv]

Summary

Two new structural edges that fill significant gaps in the call graph for modern JS/TS / Java / C# / Python / Kotlin codebases.

instantiates edges from new Foo(...)

Previously the extractor only recognised call_expression; new_expression (and equivalents in other grammars) was silently ignored, so constructor invocations produced zero graph edges.

Adds an INSTANTIATION_KINDS set covering new_expression, object_creation_expression, and instance_creation_expression. Wired into both the top-level visitNode dispatcher AND the per-function-body visitForCallsAndStructure walker — new calls inside function bodies were being missed by the body walker even after the top-level dispatch was hooked up.

Handles three subtleties:

• Generic types: new Map<K, V>() strips the angle-bracket suffix to produce 'Map', so resolution can match the class node.
• Qualified constructors: new ns.Foo() keeps the trailing identifier ('Foo') — same shape as the existing extractCall does for member-call.
• Nested calls: children are still walked, so new Foo(bar()) produces both an instantiates ref and a calls ref.

decorates edges from @Decorator annotations

Tree-sitter places decorator nodes BEFORE the symbol they apply to, so a naive walk-time dispatch saw the wrong nodeStack head (file/class instead of class/method). The fix runs decorator extraction from inside extractClass / extractFunction / extractMethod / extractProperty / extractField, after the symbol's node id is known.

Looks for decorator nodes in two places:

• Direct named children of the declaration (method/property style).
• Preceding siblings in the parent (TypeScript class style: @Foo class X {} parses as parent { decorator, class_decl }).

Handles two subtleties:

• Sibling boundary: walks BACKWARD from the declaration and stops at the first non-decorator separator. Without that stop, @A class Foo {} @B class Bar {} would attribute @A to Bar. Caught by reviewer; covered by a regression test.
• Tree-sitter wrapper identity: parent/namedChild returns fresh JS wrapper objects, so sibling === declNode is unreliable — uses startIndex comparison instead.

Also recognises Java's marker_annotation (no-args annotations like @Override/@Deprecated) alongside decorator/annotation.

Test plan

Verified live on a synthetic NestJS-shape fixture:

src              kind         tgt
UserController   decorates    Controller   (class decorator)
list             decorates    Get          (method decorator)
bootstrap        instantiates UserService  (new ...())
bootstrap        instantiates UserController

• 6 new extraction tests covering new Foo() / generic-stripping / qualified-new / @Foo class X / two-adjacent-decorated-classes regression / @Foo method()
• npx vitest run — 386 passed (was 380, +6 new)
• npx tsc --noEmit clean
• npm run build succeeds

🤖 Generated with Claude Code

[colbymchenry]

Reviewed and approved. Pushed two follow-ups before merging (commit 63a4bbe):

1. Kind-aware scoring in name-matcher.ts

The existing scorer only had a kind bonus for calls (preferring function/method). With the two new ref kinds, an instantiates ref like new Logger() could tie or pick a same-named function over the actual class when both exist across modules. Added:

• instantiates → +25 for class/struct/interface
• decorates → +25 for function/method, +15 for class/interface (Python class decorators, Java annotation interfaces)

2. Python instantiation promotion in createEdges

The PR description listed Python, but Python (and Ruby) have no new keyword — Foo() is the standard instantiation syntax and looks identical to a function call at extraction time. Resolution can tell the difference once the target is known, so I mirrored the existing extends→implements promotion: when a calls ref resolves to a class/struct, kind is promoted to instantiates. This covers Python without any extractor changes.

Known behavior change (left as-is, intentionally):

Dart's new Widget() previously emitted a calls edge via extractBareCall. The new INSTANTIATION_KINDS branch wins ahead of that path in visitForCallsAndStructure, so it now emits instantiates instead. Semantically more accurate, but a breaking change for any consumer querying "what calls Widget" on Dart code. Worth a release-note line.

Verified: 386 → 389 passing (+3 tests covering the kind biases and the Python promotion). tsc --noEmit clean.

Merging.