[colbymchenry]

Summary

Lands #1952 by @danusha2345. These calls now link to the method of the class they're written in, or to the one that class inherits:

• in Java and Scala, a call with no receiver;
• in Python, TypeScript and JavaScript, a call on self or this.

Before, they could link to a same-named method of another class that is declared closer to the call. In Java and Scala, a call inside a nested class also reaches the enclosing class's method, and a super call in Java or Python reaches the parent's method.

Root cause

The extractor records render(), this.render(), self.render() and super.render() all under the bare name render, so they reach matchByExactName.

• When several methods of that name are in reach, findBestMatch ranks them by file and line distance, so a sibling class's render declared next to the call beat the caller's own class's, or the one it inherits.
• Above the 500-candidate ceiling, the call was dropped.

From the contributor, and what we changed

• Contributor: matchEnclosingScopeMember, which runs before the ranking.
  • It reads the receiver back from the call's line (bare, self/this, super) and walks the caller's scopes innermost-out: the own class, then what it inherits, then, for a bare call, the enclosing classes.
  • The member it finds is the answer at 0.9, and that member's overloads are ranked among themselves.
  • It leaves a super call in a JS class expression to js-super-calls (JavaScript: aliased constructor receivers and overridden super calls remain unresolved #2497).
  • It comes with __tests__/enclosing-class-member-call.test.ts.
• Ours:
  1. Inherited members through the class's own resolved supertypes.
     • The walk now follows the declaration's own extends/implements edges (getSupertypeNodes), in the order its header names them, and takes a class further up over an interface that only declares the member.
     • The PR's walk by name stays as the fallback when those edges reach nothing. That covers an extends the index resolved to a namesake: Minestom's Player extends LivingEntity points at a nested EntityStatuses.LivingEntity, and Play's doc controllers point at a doc-local BaseController.
     • With the walk by name alone:
       • Django's spatialite, oracle and mysql GIS schema editors reached the postgresql backend's DatabaseSchemaEditor.execute and _delete_index_sql at 0.9. Django has one DatabaseSchemaEditor per backend, and each of these editors inherits both methods from BaseDatabaseSchemaEditor through its own backend's class.
       • A class with two bases took an arbitrary base's member: BaseDeleteView(DeletionMixin, FormMixin) got FormMixin.get_success_url, and AdminPasswordChangeForm(SetUnusablePasswordMixin, SetPasswordMixin, …) got SetPasswordMixin's methods.
       • SystemUI's RotationContextButton extends ContextualButton implements RotationButton got RotationButton's declaration of getCurrentView instead of ButtonDispatcher's implementation.
  2. Only super, this/self and Python's cls name the receiver. The PR also read parent./base. as super and cls./static./Self./$this as self. Java's extractor drops a receiver variable of those names too, so protobuf's parent.getFullName() on an EnumDescriptor went to GenericDescriptor at 0.9.
  3. C++ is left to matchCppImplicitThisCall (fix(cpp): a call a method makes with no receiver links to its class's own or inherited method #2479). That function already finds members in the class, its bases and its enclosing classes. The PR's C++ branch added the caller's own file's free functions at namespace scope, and those beat the overload the arguments fit: fmt's 3-argument copy<Char>(begin, end, out) calls went to format.h's 2-argument copy(string_view, out) at 0.9. It also turned constructor calls into instantiates edges. Some of its namespace picks were right, so C++ namespace lookup with overload fitting would make a good follow-up.
  4. Tests for each of these, and the CHANGELOG bullet in the house format.

Our part of the change is 86 of the 290 lines of the new walk and its helpers, about 30%. We replaced 15 of the PR's lines, all of them C++ handling, the receiver words, the base order and comments.

Validation

• Red to green:
  • enclosing-class-member-call.test.ts has 12 cases. Against main's resolver (31ec908), 6 of them fail: the PR's Scala, Python self., TypeScript this., Java super. and Java nested Builder cases, plus the first-base case.
  • Against the PR as written, all 5 of our new cases fail: namesake base, first base, concrete class before interface, parent variable, and C++ namespace overload.
  • js-class-expressions, js-super-calls and super-call-self-edge pass unchanged. The current head doesn't touch the js-class-expressions expectations, so there is no 0.4 → 0.9 change to justify.
• Full suite on 3df1cc64 (main 31ec908), kernel built from source, CODEGRAPH_KERNEL_EXPECT=1: 768 files passed, 6 skipped; 9,160 tests passed, 0 failed, 85 skipped. Rebased onto 6162d87, which changes only src/sync/watcher.ts (fix(watcher): a deleted link target no longer pins a CPU core on Windows #2938). The patch-id is unchanged, and the PR's test passes there.
• A/B, kernel arms, main 7cceba5 (fix(resolution): library calls on untyped Python, Scala and Java values stay unlinked #2942's library lists included) against this branch:
  • The change touches resolution only, so no kernel parity run is needed.
  • Every changed site was judged by an independent oracle. It reads the receiver from the source line and walks the graph's own contains and extends/implements edges: the caller's class, then its bases depth-first in header order, classes before interfaces, then the enclosing classes for a bare call.
  • Sites the oracle couldn't judge were read by hand.

Repo	Language	Edges, main → branch	Wrong → right	Right → wrong	Same target, 0.7 → 0.9	Read by hand
flask	Python	5,288 → 5,288	1	0	7
fastapi	Python	22,326 → 22,327	2	0	14
netbox	Python	77,636 → 77,636	316	0	518	3 cls. calls, right
django	Python	193,809 → 193,809	766	0	1,503	6: 2 cls. right, 4 lookups through a class defined in a test method, plausible
excalidraw	TypeScript	60,663 → 60,663	2	0	15
vscode base + editor	TypeScript	200,924 → 200,925	1,383	0	975
scalachess	Scala	16,586 → 16,586	18	0	28	5 enum-case confidence, right
http4s	Scala	41,435 → 41,435	56	1	113	2
playframework	Scala, Java	98,971 → 99,036	806	0	393	27 doc controllers' Action, right
Minestom	Java	128,852 → 128,903	66	0	127	17, all right (15 super. calls)
spring-ai	Java	168,525 → 168,576	312	0	282
kestra	Java	264,175 → 264,244	90	0	290
SystemUI src (AOSP)	Java	108,683 → 108,982	314	0	33	1
protobuf	Java, Python, C++	390,468 → 390,516	3,208	0	4,256
rocksdb	C++ (its Java, Python)	319,848 → 319,850	8	0	568
leveldb, fmt	C++	identical
gin, serilog (controls)	Go, C#	identical

About 7,400 links now go to the right method, and about 9,100 right links went from 0.7 to 0.9.

• The added edges are mostly super. calls, which main left unresolved: SystemUI 299, kestra 59, spring-ai 51, protobuf 48, Play 35.
• The large groups include:
  • Django test cases' self.settings() reaching SimpleTestCase.settings instead of a mail handler's;
  • vscode's this._register()/this.dispose() reaching Disposable's;
  • spring-ai builders' self() reaching DefaultChatOptionsBuilder.self;
  • protobuf's generated Java builders reaching their own members.
• One site got worse: in http4s, object RespPrelude { case class RespPrelude … RespPrelude(…) } now reaches the object rather than the case class. A bare name equal to its scope is read as a constructor.
• No C++ site changed. Neither did any edge fix(resolution): library calls on untyped Python, Scala and Java values stay unlinked #2942 removed: the library lists act on calls with a receiver, this change on bare names, so they compose. The scalachess, http4s, Play and Minestom deltas are the same on 0174f80 as on 7cceba5.

Against the PR as written (on 0174f80), the branch keeps every Java, Scala, Python and TypeScript improvement. It also fixes:

• 18 Django sites that went from one wrong target to another, 3 that went from right to wrong, and 3 confidence raises on a wrong target;
• 2 SystemUI sites that went from right to wrong (the interface declaration);
• protobuf's parent. sites.

The PR's C++ changes now stay as main has them. That is fmt's 33 retargets and 5 calls → instantiates sites, including the wrong copy/write overloads, and protobuf's 101 C++ retargets and 39 calls → instantiates sites. Leaving C++ out also gives up that branch's right picks: fmt's 6 calls through a local named self, and protobuf pyext's namespace-scoped _GetDescriptor calls.

The name fallback keeps the PR's right answers where the index resolved an extends to a namesake (Minestom's 15 super. calls, Play's 26 doc Action calls). Both of those extends resolutions are bugs of their own, worth a separate fix.

• Cost, median of 3 interleaved runs, CPU seconds:

Repo	Main	Branch
kestra (Java, about 5,000 files)	154.9 s	152.8 s
vscode base + editor (TypeScript)	79.9 s	82.6 s (rounds 79.1-82.8 on main, 75.8-83.3 here)
stress, JS (1,500 widget classes)	12.2 s	14.4 s
stress, Python	11.3 s	13.1 s
stress, Java	28.3 s	27.4 s

The walk runs before the 500-candidate ceiling, so a call on a ubiquitous name pays one pass over its candidates. The stress fixtures are a stand-in for a vendored theme like Metronic (#999): 1,500 widget classes that share init/update/render/destroy and call each other through this./self./a bare call. There, the walk links the 4,500 calls main drops at the ceiling, all to the caller's own class.

Lands #1952 by @danusha2345.

🤖 Generated with Claude Code