typescript: impact chain gaps from a mutation oracle (JSX hand-over, imported holders, platform receivers, list element types)

[swapnilpaliwal-sd]

TypeScript impact / test selection: chain gaps found by a mutation oracle (a function made to throw on entry; truth = the test files that newly fail), measured on 10 tuning repositories, with 4 more held out and evaluated only at the end. Each fix is a pattern seen in at least two repositories, has a tests/cases/typescript case that fails on the base engine and passes here, and a control that must not change.

1. JSX hands its functions to the element; a class-component tag hands over render and its lifecycle (6453aa5)

• Symptom: a handler written as onPointerMove={this.handleMove} or onClick={onSave} had no caller, and a class component tag reached only its constructor. So render, componentDidMount and the handlers, plus everything below them, were unreachable from every test that mounts the component.
• Mechanism: the parser writes each JSX {…} as an expression root linked to no element, and intrinsic tags have no call site at all. The {…} is now the site that hands the function over (callback_registered), and a class-component tag hands over render and its lifecycle methods.
• Shape population: function-valued JSX roots appear in 9 of 14 repositories; class components are mounted by tag in 4.
• Effect: on the React application in the set, impact recall 0.49 → 0.81, precision 0.17 → 0.24, empty selections 4 → 2, path 32 → 38/42. The other 9 tuning repositories are byte-identical.

2. A call through an imported const runs the function it holds (b2e7cda)

• Symptom: export const createDraft = engine.createDraft.bind(engine), and the same with no bind, was followed for a call in its own module. A call from any other module ran nothing, and that includes the package's whole test suite.
• Mechanism: a call whose callee is an imported const, or a namespace member, now runs what the exporting module's variable holds. The import is followed to that module, never by name (control: the same exported name bound in another module).
• Effect: recall 0.94 → 0.96, 0.42 → 0.47 and 0.91 → 0.93 on three repositories; path 404 → 410/491. Precision −0.01 to −0.03 on those three.

3. A call on the platform's own objects is no by-name caller (5c2130a)

• Symptom: with no standard library staged, JSON.parse, Promise.resolve, Object.keys(…).forEach, Reflect.getMetadata, document.createElement and similar calls were untyped calls to impact. Any project method with the same name collected all of them as by-name callers. Across 13 of 14 repositories there are 1–1007 such sites per repository.
• Mechanism: the engine traces the receiver to the platform's global object, as it already does for a package's value. The site seeds no closure and is listed apart.
• Effect: by-name seeds on one repository 188 → 34, and its by-name test entries 1454 → 385. File-level selection is unchanged on the oracle, because other routes reach the same files.

4. A list's element is typed through an object-type alias and through a type variable's constraint (21c0f9f)

• Symptom: plugins.map((p) => p.setup()) gave p nothing to dispatch on in two cases:
  • the element is an object type written as an alias (type Plugin = { setup(): … }). The no-lib callback binding only passed on a type, and an alias of an object type has none, only members. An annotated parameter of the same type did resolve.
  • the list is typed by <Ms extends [Module, ...Module[]]>, whose element had no type at all.
• Mechanism: the callback parameter now also takes the element's shape. A type variable constrained to an array or tuple yields the constraint's element, for callbacks and for for…of, exactly as <T extends HasId> already does for a receiver.
• Effect: newly resolved sites in 6 repositories (1–6 each). On the module system built this way, impact recall 0.47 → 0.96, precision 0.39 → 0.41, tests recall 0.42 → 0.96, path 48 → 59/62. The type-variable half resolves nothing on its own; it matters together with the alias half.

Totals, base → this branch

set	impact recall (micro / macro)	impact precision	empty selections	path reached
tuning (10 repos, 223 mutants)	0.74 → 0.84 / 0.776 → 0.865	0.21 → 0.22	16 → 14	398 → 421 / 491
held out (4 repos, 96 mutants)	0.74 → 0.74 / 0.894 → 0.894	0.27 → 0.27	1 → 1	163 → 163 / 183

The held-out repositories do not exercise these shapes on their truth chains, so they show neither a gain nor a regression. Files selected for mutants whose truth is empty rise 854 → 1010, all of it on the React application: mounting the root component now reaches its handlers. Engine solve time on the same IR is unchanged within noise (2.2 / 8.7 / 8.8 / 24 / 33 s, base and fix interleaved).

Open, not fixed (each seen in one repository only):

• a package self-import that resolves to a dev-build stub in dist/;
• (x as typeof ConstObject).method();
• functions assigned onto a function object and called through an interface;
• JS tests reaching TS source through require() namespaces.

Checks: tests/run.py TypeScript 296/296, JavaScript 307/307, Python 306/306, Java 333/333. TypeScript engine suite 101/101 (on the final engine); two goldens gain only the intended new edges, and the oracle files are unchanged.