[colbymchenry]

Summary

A TypeScript or JavaScript method called on an element of a class field, like this.handlers[type].handle() or this.items[0].save(), now links to the method of the type the field is declared to hold (Item[], Array<Item>, ReadonlyArray<Item>, Record<string, Handler>), or to nothing. On main it kept the bare method name, which exact-matched whichever project method shared it, often another class's: this.listeners[type].push(h) reached an unrelated UndoStack.push.

Root cause

A member chain rooted at this collapses to the bare method name, so the resolver can read the owner off the enclosing class (keepsBareTsJsReceiver, and the kernel's keeps_bare_receiver). A subscript in that chain leaves the resolver nothing to read the receiver from, so the bare name went to name matching.

The fix (from the contributor)

• Extraction, both engines. this.<field>[k].<m>() is emitted as this.<field>[].<m>, and this['svc'].m() as this.svc.m, the field itself (tsThisSubscriptReceiver / this_subscript_of). Any other subscript in a this-rooted chain (this.a.b[k].m(), this[k].m()), and one under window (window['App'].init()), now emits nothing, as arr[0].run() already did.
• Resolution. matchTsThisFieldCall reads the field's declaration on the calling class or a base, as it does for this.<field>.<m>(). The field reader also records the element type of X[], Array<X>, ReadonlyArray<X> and Record<K, X>, and the method resolves on that type or nowhere. There is no element type for a type parameter of the declaring class, a union or intersection with another type (| null and | undefined keep it, also when the union goes on below), an array of arrays, or a non-plain element.
• tsClassTypeParameter reads a class's type-parameter list from its keyword to the list's own >, over several lines and through braces in constraints (<T extends { id: string }>).

What we changed on top

• Conflict with fix(cpp): code that calls operator++ by its name no longer stops an index or sync #2769: kept escapeRegExp in tsClassTypeParameter.
• tsClassTypeParameter also accepts the type keyword. The PR's keyword match (class|interface) stopped seeing a generic type alias's parameters, and the T['p'] path from fix(ts): a local annotated as one of a type's properties has that property's type #2640 passes type-alias owners to it. So State<Item = unknown>['current'] read Item as the project class Item, where main reads it as the type parameter. New case in ts-array-annotated-receivers.test.ts: it fails on the PR as written (the alias half) and on main (the brace-constraint half), and passes here.
• CHANGELOG: the bullet now covers JavaScript, says to re-index, and credits the author.

Validation

Tests. ts-this-field-element.test.ts (the contributor's) fails 10 of 11 on main with both the kernel and wasm (CODEGRAPH_KERNEL=0), each case linking to the decoy class, and passes here on both. The new ts-array-annotated-receivers case fails on the PR as written (the type-alias half) and on main (the brace-constraint half), and passes here. The contributor's extraction cases in expression-receiver-calls and kernel-tsjs-parity pass.

A/B against main 9c4794d (graph dumps by natural keys, kernel built per arm; the main commits after it touch only the viewer and tests):

Repo	Lang	Files	Added	Retargeted	Typed (same target)	Removed
flask (control)	Python	91	0	0	0	0, byte-identical
gin (control)	Go	110	0	0	0	0, byte-identical
vscode src/vs/base + src/vs/editor	TS	1,395	19	24	35	45
pixijs	TS	1,629	7	5	5	15
typeorm	TS	3,652	0	2	2	4
excalidraw	TS	731	0	0	0	1
three.js	JS	1,763	0	0	0	116
playcanvas engine	JS	2,326	0	0	0	119
pdf.js	JS	595	0	0	0	53

Every changed edge was read at its call site:

• Added and retargeted, all right. They land on the element type the field declares. vscode's this.modelLineProjections[i].getViewLineCount() moves from the class's own method to IModelLineProjection (17 sites), this._lenses[i].dispose() from CodeLensContribution::dispose to CodeLensWidget::dispose, and this._chunks[0].slice() from a project slice function to VSBuffer::slice. pixijs's this.buffers[0].update() moves from RopeGeometry::update to Buffer::update. typeorm's this.connectedQueryRunners[0].release() now reaches the declared QueryRunner interface rather than one driver's runner.
• TS removals: 50 of 65 were wrong. These are bare-name guesses on strings, arrays, Sets, Maps and DOM nodes: vscode's this._lines[i].substring() → OffsetRange::substring ×13, .focus() on DOM elements, and pixijs's this._gpuData[key]?.destroy() → the class's own destroy. The 15 right ones lost were right only by name, in shapes the PR leaves unresolved by design. Those are a getter (this.viewItems), an index-signature field, a cast ((this.filters[0] as MaskFilter)), a mixin object's this, a T[] type parameter, a union type alias, and deep chains like this.replacements[0].range.getStartPosition().
• JS removals: about 230 of 288 were wrong. JS fields have no written type, so every element call now goes unlinked. Examples: three.js .set() → Capsule::set ×41 and .copy() → NURBSCurve::copy ×22, playcanvas SVG setAttribute → StandardMaterial::setAttribute ×27, pdf.js classList.add → Word64::add. About 58 were right by luck (Color::setHex, GraphNode::setLocalPosition, PDFThumbnailView::toggleSelected). This is the deliberate trade the PR describes, the same one arr[0].run() already makes.
• window['App'].init(): no such call in any of the 9 repos. The one window-rooted subscript chain changed was excalidraw's window.visualDebug.data[…].push(element), which had linked to BinaryHeap::push (wrong, removed).

Kernel parity (scripts/kernel-parity.mjs) holds on every TS/JS repo above and on this repository's src/ and __tests__/, with the same deferral counts as main.

Cost: the index CPU time is within noise. Medians of 3 interleaved runs: pixijs 32.4 s → 33.3 s and three.js 64.7 s → 67.9 s, against a spread of 28.6–33.8 s and 60.1–72.6 s between runs of the same arm.

Suite (CODEGRAPH_KERNEL_EXPECT=1 npx vitest run --maxWorkers=4 --minWorkers=1, on this commit): 756 files, 9,028 tests, 8,921 passed, 82 skipped. The 25 that failed were all in the four daemon suites (mcp-daemon, daemon-older-version, daemon-silent-holder, daemon-pid-reuse), because the run's environment had CODEGRAPH_NO_DAEMON=1, which their spawned servers inherit. Run again with the daemon allowed, those four files pass: 58 passed, 1 skipped. Main then moved by one PHP-only commit (5c5bfd4, import-resolver.ts). After rebasing onto it, with the changed lines identical, this PR's test files, the PHP alias test and the resolution suites pass: 8 files, 401 tests.

Lands #2766 by @inth3shadows.

🤖 Generated with Claude Code