[colbymchenry]

Summary

A method that hands its call on to another object's method of the same name, like a wrapper's def acceptsURL(url) = proxied.acceptsURL(url), no longer links that call back to itself. Before, the wrapper looked recursive in callers and impact. Recursive calls through this/self, bare calls and calls in Go keep their link.

Root cause

matchMethodCall's strategy 3 guesses a method by its name when it can't type the receiver. With only one method of that name in the project, the single-candidate branch took it outright. A delegating wrapper is often the only method of that name in the repository, because the method it delegates to belongs to a library, so the guess was the calling method itself. (The word-overlap branch already refuses the caller.)

What came from the contributor, and what we changed

From @danusha2345 (#1939): isForeignReceiverSelfLoop in the single-candidate branch (no edge when the one candidate is the caller and the receiver is not this, self, Self, super or cls), the Scala delegating-wrapper test, and the recursion test (Java this.depth() and bare size(), Go named receiver).

Ours, on top:

1. Go is left out. The A/B showed every Go case was real recursion (8 of 8): cobra's sub.checkCommandGroups() over c.commands and cmd.Traverse(…), etcd's child.Repr(…), child.recoverAndclean() and tOp.isWrite() over op.thenOps, prometheus's subParameter.toLabels(). A Go method's own receiver and its struct fields are typed before strategy 3, so a Go call that still arrives there is on a local the resolver couldn't type, mostly a range variable over the type's own children. In every other language the rule is kept.
2. Tests. The PR description lists a TypeScript case and a Java case the diff didn't have. The TypeScript one is real (a wrapper handing its call to an untyped parameter self-loops on main) and is added. The Java one no longer reproduces: main now gives a receiver of an unindexed or imported type no guess at all. Also added a Go range-variable recursion test, which fails on the PR as written.
3. A CHANGELOG bullet (the PR had none), next to the VB.NET bullet about name-only guesses on untyped receivers.

Validation

Tests (__tests__/resolution.test.ts, "Name-only method match never binds a delegating call to its own caller"): 2 of 4 fail on main (the Scala and TypeScript wrappers), 1 fails on the PR as written (Go range-variable recursion), all 4 pass here.

A/B, main (3a93f57) vs this change, fresh indexes compared site by site, 24 repos in 11 languages (flask, click, fastapi, gin, cobra, etcd, prometheus, hugo, ripgrep, jq, Slim, serilog, eShopOnWeb, nowinandroid, okhttp, mihon, http4s, playframework, spring-modulith, express, zod, bulletproof-react, excalidraw, vben). Nothing is added. Every removed edge is a calls self-loop that strategy 3 had guessed at 0.7:

	removed self-loops	wrong (delegation)	real recursion lost
PR as written	71	56	15
this change	63	56	7

• Wrong self-loops removed (56), all by reading the source: playframework 19 (proxied.acceptsURL, underlying.withUrl, JDBC ps.executeQuery, Selenium element.click(), Java calling a Scala object's same-named method), vben 9 (form.validate() and friends in FormApi), http4s 8 (pairs.drop(n) on a Vector, segments.splitAt), mihon 8 (ClassLoader parent.loadClass, flow.collectAsState()), okhttp 5, ripgrep 3, flask 2 (jinja loader.list_templates()), spring-modulith 2. Two of playframework's were sites that also had the right edge, which they keep.
• Real recursion through another instance of the caller's type now gets no edge (7): okhttp's Kotlin extension functions parentFile.isDescendentOf(…) and cause?.isNetworkUnusable() (3), fastapi's child_branch.effective_low_priority_routes() on a just-built _IncludedRouter (2), a serilog C# extension method (1), spring-modulith's it.containsTypeInAnyParent(…) (1). Nothing cheap tells them apart from the delegations: mihon's extension functions StateFlow.collectAsLazyPagingItems() and Preference.collectAsState() are delegations with the same shape. This is the trade-off the PR describes, measured.
• fix(play): a Class.method reference reaches a member of the class it names #2875 changed Play's resolver after this A/B, so playframework was indexed again on that base (main at 4c572e0 vs this change on it): 17 self-loops removed (15 sites, plus 2 that keep their right edge), the same set less Java's two Multipart.multipartParser / Multipart.randomBoundary calls, which main itself now links to the Scala Multipart object (0.9). Nothing added.
• Rebased onto current main, which since added fix(resolution): prefer the caller's own copy of a duplicated module #1936 (matchByExactName and findBestMatch) and fix(scala): resolve Outer.Inner(...) to the Scala type, never another language's constructor #1951 (Scala Outer.Inner(...) in matchMethodCall, just before strategy 3, plus a fall-through when a typed Scala match lands on another language's constructor). playframework indexed once more, main (b6379e6) vs this change on it: the same 17 self-loops removed, nothing else, and no other strategy-3 self-loop left (instance-method 0.7 self-loops 17 → 0). All 249 tests of resolution.test.ts pass, fix(scala): resolve Outer.Inner(...) to the Scala type, never another language's constructor #1951's Scala cases and the Scala wrapper case among them.
• Kernel parity doesn't apply: TypeScript resolution only, same kernel binary in both arms.

Cost: okhttp index process CPU time, 8 interleaved rounds with the order alternated (other agents' jobs were running): median 32.8 s on main, 32.3 s with this change. The change adds one id comparison to a branch that already returns, so the difference is noise.

Suite on this commit (CODEGRAPH_KERNEL_EXPECT=1 npx vitest run --maxWorkers=4 --minWorkers=1, Windows): 760 files passed, 6 skipped; 9,041 tests passed, 83 skipped, 0 failed.

Lands #1939 by @danusha2345.

🤖 Generated with Claude Code