[colbymchenry]

Summary

Fixes #2589. react-dom's root.render(<Popup />) used to link to the project's one class-component render(), an error boundary's in the reporter's app, in two situations: when root was declared outside the calling function, and when it held what createRoot(el) returned. After this change:

• A JS/TS receiver the calling function doesn't type is read by its binding in scope at the call.
• A class component's render() and lifecycle methods are never a by-name guess.

Root cause

matchMethodCall asks inferLocalReceiverType for the receiver's type, and that scan only covers the calling function. It reads nothing from these places:

• a module's let root: Root | null
• a let popupRoot: Root in the function around the one rendering
• an outer function's manager: RunnerManager parameter

So #2169's rule (a type the project doesn't declare is an outside type) never ran, and Strategy 3 handed the call to the only same-language method named render. An untyped const r = createRoot(el) reached the same guess even inside the calling function.

Fix

A receiver's binding in scope (jsBindingAt, jsBindingReading in name-matcher.ts):

• A per-file brace-scope reader (comments, strings and regex literals blanked) finds the innermost binding of the receiver's name at the call. That can be a const/let/var, a parameter list, a destructuring, a for head, a function/class, a lone or destructured arrow parameter, or a catch clause.
• A declaration or parameter written with a type, or a declaration whose whole value is a new of a class, types the call. From there the existing typed pipeline applies (imported-type file, outside types, resolveMethodOnType), exactly as for the calling function's own declaration.
• Two cases say nothing and are left to the existing strategies:
  • An annotated object literal stays its own members': vite's const server: ViteDevServer = { … } is the close it declares, not the interface's.
  • A new of a parameter or variable (new Backend(root)) names no class.
• A nearer binding of the same name always wins. That covers a parameter, an arrow's or a destructured one, and the calling function's own untyped declaration. Another function's binding holds only in that function.
• It runs only where the binding may lie outside the calling function: in a function nested in another, or for a name the module scope declares. A call outside any function was already read from the file's start. With that pre-filter, the per-reference cost dropped from about 340 µs to about 80 µs on nitro.

React's own methods (isReactLifecycleMethod): a lifecycle method of a class extending Component/PureComponent (render, componentDidMount, …) is ruled out of Strategy 3's by-name candidates for a receiver other than this/super, the same way a Vue component's own methods are. It also counts as narrowing, so a lone other render isn't then guessed instead. That covers the untyped const r = createRoot(el) and ReactDOM.createRoot(el) shapes.

Calls I made

• I dropped a broader rule I tried first. It treated any receiver whose value a package from outside the project returned as having no project method. The A/B showed it unlinking correct calls through values a hook hands back, such as const parentContext = useContext(AnalyticsContext) in social-app (AnalyticsContextType::metric) and outline's drag-and-drop manager, so only the React-specific rule covers untyped react-dom roots.
• Some cases are left as they were:
  • An x.render() on a component instance named after its class (Strategy 2).
  • Classes that extend a project base component rather than Component directly.
  • A declaration inside an anonymous callback that no function node covers, which the existing scan already reads from the file's start.

Validation

• New test __tests__/js-receiver-binding-scope.test.ts (3 tests): all 3 fail on main and pass on the fix.
  • Every shape in the issue, plus ReactDOM.createRoot / hydrateRoot through the default import, and a .jsx class component.
  • Project types from a module binding, an outer function's declaration and an outer parameter.
  • Shadowing: a typed parameter, an arrow's own, a destructured one, another function's declaration, and a new of a parameter.
• A/B against main 89cd08f, each arm indexed fresh. "Same target" means the same edge with its confidence raised to 0.9.

repo	removed	added	retargeted	same target
jest	15	1	1	19
nitro	7
nuxt	5			4
webdriverio	3	4	1	10
vite	2	4	4	10
social-app	1	7		2
outline	1	2		10
vben	1
memfs			1	1
insomnia				36
t3-turbo, vue3-antdv-admin, pathe, c12
flask, gin (controls)	byte-identical dumps

• Triage of every changed edge:
  • The 35 removals were all wrong guesses. They are calls on values whose binding outside the calling function has a type from outside the project: rolldown's RolldownWatcher, a ViteDevServer parameter, env-runner's RunnerManager, node's Server, ChildProcess and AsyncLocalStorage, NodeJS.WriteStream (jest's outputStream.write going to BufferedConsole::write), playwright's BrowserServer, unstorage's Storage, HTMLElement, a zod schema and Omit<…>. One more is nitro's const root = ReactDOM.createRoot(el); root.render(…), the issue's own shape, which went to an unrelated example's ErrorBoundary::render.
  • The 25 added or retargeted edges were all precise. Examples: memfs's onAbort = () => watcher.close() reaching FsaNodeFsWatcher::close instead of CoreWatcher::close, jest's currentSpec.cancel() reaching Spec::cancel instead of PCancelable::cancel, and webdriverio's server.close() reaching SessionServer::close instead of ViteServer::close. Two of them are vite's new WebSocket(…) in an outer function linking to its vendored ws.d.ts class, which is what the same declaration in the calling function already gets.
  • The 92 same-target changes are module-level singletons like insomnia's const repoFileWatcherRegistry = new RepoFileWatcherRegistry(), now typed rather than guessed.
• Kernel parity: this is a resolution-only change and no extraction changed, so kernel parity is unaffected.
• Index time: +1.7% total, as the median of 8 interleaved runs on vite, nuxt, webdriverio and insomnia. That is within this machine's noise: per repo the change ranged from -3.6% to +5.2%, and the last 5 rounds alone came to +2.2%. A resolver profile of insomnia puts the new reading at about 0.4 s of roughly 21 s of resolver time, summed over its workers.
• Base: the branch sits on main f3fdba2. The A/B baseline was 89cd08f. The four commits in between touch neither TS/JS resolution nor TS/JS extraction: fix(bundle): bundle Node 24.21.0 so Windows commands no longer abort as they exit #2806 bundle, fix(add-lang): check-grammar fails a grammar that crashes or hangs its process #2808 add-lang script, fix(extraction): drop the extractVariables hook, which the core never called #2809 an extractor hook the core never called, and fix(pascal): calls inside nested Delphi routines are indexed, each under its routine #2810 Pascal nested routines.
• Sync: a CLI sync that retypes a module-level binding, adds an untyped createRoot local and adds a file with a module-level new Store() matches a fresh index.
• Full suite on this commit (CODEGRAPH_KERNEL_EXPECT=1): 726 test files passed and 6 skipped; 8,664 tests passed and 82 skipped.

🤖 Generated with Claude Code